diff --git a/.clang-tidy b/.clang-tidy index 3078beacc..952c0cca8 100644 --- a/.clang-tidy +++ b/.clang-tidy @@ -12,6 +12,7 @@ Checks: > -readability-implicit-bool-conversion, -readability-magic-numbers, -readability-uppercase-literal-suffix, + -readability-simplify-boolean-expr, clang-analyzer-*, -clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling, performance-*, diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index c7fa2203a..76d96e63c 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -10,7 +10,7 @@ git, python3, mpi, - openblas, # TODO: Use the generic `blas` so users could switch between alternative implementations + blas, cudaPackages, darwin, rocmPackages, @@ -181,6 +181,7 @@ effectiveStdenv.mkDerivation ( ++ optionals useMpi [ mpi ] ++ optionals useOpenCL [ clblast ] ++ optionals useRocm rocmBuildInputs + ++ optionals useBlas [ blas ] ++ optionals useVulkan vulkanBuildInputs; cmakeFlags = @@ -216,8 +217,7 @@ effectiveStdenv.mkDerivation ( # Should likely use `rocmPackages.clr.gpuTargets`. "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102" ] - ++ optionals useMetalKit [ (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1") ] - ++ optionals useBlas [ (lib.cmakeFeature "LLAMA_BLAS_VENDOR" "OpenBLAS") ]; + ++ optionals useMetalKit [ (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1") ]; # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level, # if they haven't been added yet. diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 992c34a03..0e7643bba 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -15,14 +15,133 @@ on: types: [opened, synchronize, reopened] paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m'] +concurrency: + group: ${{ github.workflow }}-${{ github.ref }} + cancel-in-progress: true + env: BRANCH_NAME: ${{ github.head_ref || github.ref_name }} GGML_NLOOP: 3 GGML_N_THREADS: 1 jobs: + macOS-latest-cmake-arm64: + runs-on: macos-14 + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v3 + + - name: Dependencies + id: depends + continue-on-error: true + run: | + brew update + + - name: Build + id: cmake_build + run: | + sysctl -a + mkdir build + cd build + cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON .. + cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) + + - name: Test + id: cmake_test + run: | + cd build + ctest -L main --verbose --timeout 900 + + - name: Determine tag name + id: tag + shell: bash + run: | + BUILD_NUMBER="$(git rev-list --count HEAD)" + SHORT_HASH="$(git rev-parse --short=7 HEAD)" + if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then + echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT + else + SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') + echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT + fi + + - name: Pack artifacts + id: pack_artifacts + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + run: | + cp LICENSE ./build/bin/ + zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/* + + - name: Upload artifacts + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + uses: actions/upload-artifact@v3 + with: + path: | + llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip + + macOS-latest-cmake-x64: + runs-on: macos-latest + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v3 + + - name: Dependencies + id: depends + continue-on-error: true + run: | + brew update + + - name: Build + id: cmake_build + run: | + sysctl -a + mkdir build + cd build + cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON .. + cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) + + - name: Test + id: cmake_test + run: | + cd build + ctest -L main --verbose --timeout 900 + + - name: Determine tag name + id: tag + shell: bash + run: | + BUILD_NUMBER="$(git rev-list --count HEAD)" + SHORT_HASH="$(git rev-parse --short=7 HEAD)" + if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then + echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT + else + SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') + echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT + fi + + - name: Pack artifacts + id: pack_artifacts + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + run: | + cp LICENSE ./build/bin/ + zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/* + + - name: Upload artifacts + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + uses: actions/upload-artifact@v3 + with: + path: | + llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip + ubuntu-focal-make: runs-on: ubuntu-20.04 + env: + LLAMA_NODE_AVAILABLE: true + LLAMA_PYTHON_AVAILABLE: true steps: - name: Clone @@ -35,6 +154,14 @@ jobs: sudo apt-get update sudo apt-get install build-essential gcc-8 + - uses: actions/setup-node@v4 + with: + node-version: "20" + + - uses: actions/setup-python@v4 + with: + python-version: "3.11" + - name: Build id: make_build env: @@ -98,6 +225,17 @@ jobs: cd build ctest -L main --verbose --timeout 900 + - name: Test llama2c conversion + id: llama2c_test + run: | + cd build + echo "Fetch tokenizer" + wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin + echo "Fetch llama2c model" + wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin + ./bin/convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf + ./bin/main -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256 + # ubuntu-latest-cmake-sanitizer: # runs-on: ubuntu-latest # @@ -662,6 +800,7 @@ jobs: windows-latest-cmake-sycl: runs-on: windows-latest + defaults: run: shell: bash @@ -670,7 +809,6 @@ jobs: WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/62641e01-1e8d-4ace-91d6-ae03f7f8a71f/w_BaseKit_p_2024.0.0.49563_offline.exe WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel - steps: - name: Clone id: checkout @@ -685,6 +823,32 @@ jobs: id: cmake_build run: examples/sycl/win-build-sycl.bat + - name: Determine tag name + id: tag + shell: bash + run: | + BUILD_NUMBER="$(git rev-list --count HEAD)" + SHORT_HASH="$(git rev-parse --short=7 HEAD)" + if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then + echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT + else + SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') + echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT + fi + + - name: Pack artifacts + id: pack_artifacts + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + run: | + 7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip .\build\bin\* + + - name: Upload artifacts + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + uses: actions/upload-artifact@v3 + with: + path: | + llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip + ios-xcode-build: runs-on: macos-latest @@ -748,6 +912,8 @@ jobs: - macOS-latest-cmake - windows-latest-cmake - windows-latest-cmake-cublas + - macOS-latest-cmake-arm64 + - macOS-latest-cmake-x64 steps: - name: Clone diff --git a/.github/workflows/close-issue.yml b/.github/workflows/close-issue.yml index a151c6780..7f21daec0 100644 --- a/.github/workflows/close-issue.yml +++ b/.github/workflows/close-issue.yml @@ -19,5 +19,5 @@ jobs: close-issue-message: "This issue was closed because it has been inactive for 14 days since being marked as stale." days-before-pr-stale: -1 days-before-pr-close: -1 - operations-per-run: 1000 + operations-per-run: 10000 repo-token: ${{ secrets.GITHUB_TOKEN }} diff --git a/.github/workflows/code-coverage.yml b/.github/workflows/code-coverage.yml index 392db8a08..4112518bb 100644 --- a/.github/workflows/code-coverage.yml +++ b/.github/workflows/code-coverage.yml @@ -5,6 +5,10 @@ env: GGML_NLOOP: 3 GGML_N_THREADS: 1 +concurrency: + group: ${{ github.workflow }}-${{ github.ref }} + cancel-in-progress: true + jobs: run: runs-on: ubuntu-20.04 diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml index 94f9161fc..9591bfc2a 100644 --- a/.github/workflows/docker.yml +++ b/.github/workflows/docker.yml @@ -15,6 +15,10 @@ on: branches: - master +concurrency: + group: ${{ github.workflow }}-${{ github.ref }} + cancel-in-progress: true + jobs: push_to_registry: name: Push Docker image to Docker Hub diff --git a/.github/workflows/editorconfig.yml b/.github/workflows/editorconfig.yml index 0e0993cd4..7b2a00c90 100644 --- a/.github/workflows/editorconfig.yml +++ b/.github/workflows/editorconfig.yml @@ -14,6 +14,10 @@ on: branches: - master +concurrency: + group: ${{ github.workflow }}-${{ github.ref }} + cancel-in-progress: true + jobs: editorconfig: runs-on: ubuntu-latest diff --git a/.github/workflows/nix-ci-aarch64.yml b/.github/workflows/nix-ci-aarch64.yml index 8d0a3fd7f..109a793ea 100644 --- a/.github/workflows/nix-ci-aarch64.yml +++ b/.github/workflows/nix-ci-aarch64.yml @@ -17,6 +17,10 @@ on: types: [opened, synchronize, reopened] paths: ['**/*.nix', 'flake.lock'] +concurrency: + group: ${{ github.workflow }}-${{ github.ref }} + cancel-in-progress: true + jobs: nix-build-aarch64: runs-on: ubuntu-latest diff --git a/.github/workflows/nix-ci.yml b/.github/workflows/nix-ci.yml index 01c5a9d5a..8b5b99c8f 100644 --- a/.github/workflows/nix-ci.yml +++ b/.github/workflows/nix-ci.yml @@ -8,6 +8,10 @@ on: pull_request: types: [opened, synchronize, reopened] +concurrency: + group: ${{ github.workflow }}-${{ github.ref }} + cancel-in-progress: true + jobs: nix-eval: strategy: diff --git a/.github/workflows/python-check-requirements.yml b/.github/workflows/python-check-requirements.yml index b82205992..4092b12fa 100644 --- a/.github/workflows/python-check-requirements.yml +++ b/.github/workflows/python-check-requirements.yml @@ -16,6 +16,10 @@ on: - 'requirements.txt' - 'requirements/*.txt' +concurrency: + group: ${{ github.workflow }}-${{ github.ref }} + cancel-in-progress: true + jobs: python-check-requirements: runs-on: ubuntu-latest diff --git a/.github/workflows/python-lint.yml b/.github/workflows/python-lint.yml index ea0a05ea1..4bdd79c4a 100644 --- a/.github/workflows/python-lint.yml +++ b/.github/workflows/python-lint.yml @@ -2,6 +2,10 @@ name: flake8 Lint on: [push, pull_request] +concurrency: + group: ${{ github.workflow }}-${{ github.ref }} + cancel-in-progress: true + jobs: flake8-lint: runs-on: ubuntu-latest diff --git a/.github/workflows/server.yml b/.github/workflows/server.yml index 65ca7d9ca..f07d25536 100644 --- a/.github/workflows/server.yml +++ b/.github/workflows/server.yml @@ -18,6 +18,10 @@ on: schedule: - cron: '0 0 * * *' +concurrency: + group: ${{ github.workflow }}-${{ github.ref }} + cancel-in-progress: true + jobs: server: runs-on: ubuntu-latest @@ -31,7 +35,6 @@ jobs: include: - build_type: Release sanitizer: "" - disabled_on_pr: true fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken container: diff --git a/.github/workflows/zig-build.yml b/.github/workflows/zig-build.yml index 68a698ab9..cb43954eb 100644 --- a/.github/workflows/zig-build.yml +++ b/.github/workflows/zig-build.yml @@ -6,6 +6,10 @@ on: branches: - master +concurrency: + group: ${{ github.workflow }}-${{ github.ref }} + cancel-in-progress: true + jobs: build: strategy: diff --git a/.gitignore b/.gitignore index 1ad8d929b..9fb5b80c3 100644 --- a/.gitignore +++ b/.gitignore @@ -11,7 +11,10 @@ *.gcda *.dot *.bat +*.tmp *.metallib +*.etag +*.lastModified .DS_Store .build/ .cache/ @@ -47,6 +50,7 @@ models-mnt /embedding /gguf /gguf-llama-simple +/gguf-split /gritlm /imatrix /infill @@ -55,6 +59,9 @@ models-mnt /llava-cli /lookahead /lookup +/lookup-create +/lookup-merge +/lookup-stats /main /metal /passkey @@ -70,6 +77,7 @@ models-mnt /batched-bench /export-lora /finetune +/retrieval /speculative /parallel /train-text-from-scratch diff --git a/CMakeLists.txt b/CMakeLists.txt index fc4cff28f..b25cfd2fc 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -99,6 +99,7 @@ option(LLAMA_CUDA_F16 "llama: use 16 bit floats for some set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K") set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING "llama: max. batch size for using peer access") +option(LLAMA_CUDA_NO_PEER_COPY "llama: do not use peer to peer copies" OFF) option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF) option(LLAMA_HIPBLAS "llama: use hipBLAS" OFF) option(LLAMA_HIP_UMA "llama: use HIP unified memory architecture" OFF) @@ -368,7 +369,9 @@ if (LLAMA_CUBLAS) enable_language(CUDA) set(GGML_HEADERS_CUDA ggml-cuda.h) - set(GGML_SOURCES_CUDA ggml-cuda.cu) + + file(GLOB GGML_SOURCES_CUDA "ggml-cuda/*.cu") + list(APPEND GGML_SOURCES_CUDA "ggml-cuda.cu") add_compile_definitions(GGML_USE_CUBLAS) if (LLAMA_CUDA_FORCE_DMMV) @@ -387,6 +390,9 @@ if (LLAMA_CUBLAS) endif() add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER}) add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE}) + if (LLAMA_CUDA_NO_PEER_COPY) + add_compile_definitions(GGML_CUDA_NO_PEER_COPY) + endif() if (LLAMA_STATIC) if (WIN32) @@ -515,7 +521,9 @@ if (LLAMA_HIPBLAS) message(STATUS "HIP and hipBLAS found") set(GGML_HEADERS_ROCM ggml-cuda.h) - set(GGML_SOURCES_ROCM ggml-cuda.cu) + + file(GLOB GGML_SOURCES_ROCM "ggml-cuda/*.cu") + list(APPEND GGML_SOURCES_ROCM "ggml-cuda.cu") add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS) @@ -531,11 +539,15 @@ if (LLAMA_HIPBLAS) add_compile_definitions(GGML_CUDA_FORCE_MMQ) endif() + if (LLAMA_CUDA_NO_PEER_COPY) + add_compile_definitions(GGML_CUDA_NO_PEER_COPY) + endif() + add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X}) add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y}) add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER}) - set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX) + set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX) if (LLAMA_STATIC) message(FATAL_ERROR "Static linking not supported for HIP/ROCm") diff --git a/Makefile b/Makefile index 1daad45ed..08eafb1e7 100644 --- a/Makefile +++ b/Makefile @@ -1,15 +1,16 @@ # Define the default target now so that it is always the first target BUILD_TARGETS = \ main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \ - simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search \ - speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o + simple batched batched-bench save-load-state server gguf gguf-split llama-bench libllava.a llava-cli baby-llama beam-search \ + retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o # Binaries only useful for tests TEST_TARGETS = \ tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \ tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama \ tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope \ - tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease + tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease \ + tests/test-json-schema-to-grammar # Code coverage output files COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report @@ -397,6 +398,7 @@ ifdef LLAMA_CUBLAS MK_CPPFLAGS += -DGGML_USE_CUBLAS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include MK_LDFLAGS += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib OBJS += ggml-cuda.o + OBJS += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu)) MK_NVCCFLAGS += -use_fast_math ifdef LLAMA_FATAL_WARNINGS MK_NVCCFLAGS += -Werror all-warnings @@ -451,18 +453,29 @@ ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE else MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE -#ifdef LLAMA_CUDA_CUBLAS -# MK_NVCCFLAGS += -DGGML_CUDA_CUBLAS -#endif # LLAMA_CUDA_CUBLAS +ifdef LLAMA_CUDA_NO_PEER_COPY + MK_NVCCFLAGS += -DGGML_CUDA_NO_PEER_COPY +endif # LLAMA_CUDA_NO_PEER_COPY ifdef LLAMA_CUDA_CCBIN MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN) endif -ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml-common.h + ifdef JETSON_EOL_MODULE_DETECT +define NVCC_COMPILE $(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@ +endef # NVCC_COMPILE else +define NVCC_COMPILE $(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@ +endef # NVCC_COMPILE endif # JETSON_EOL_MODULE_DETECT + +ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh + $(NVCC_COMPILE) + +ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh) + $(NVCC_COMPILE) + endif # LLAMA_CUBLAS ifdef LLAMA_CLBLAST @@ -509,7 +522,6 @@ ggml-vulkan.o: ggml-vulkan.cpp ggml-vulkan.h endif # LLAMA_VULKAN ifdef LLAMA_HIPBLAS - ifeq ($(wildcard /opt/rocm),) ROCM_PATH ?= /usr GPU_TARGETS ?= $(shell $(shell which amdgpu-arch)) @@ -534,9 +546,17 @@ endif # LLAMA_HIP_UMA ifdef LLAMA_CUDA_FORCE_DMMV HIPFLAGS += -DGGML_CUDA_FORCE_DMMV endif # LLAMA_CUDA_FORCE_DMMV +ifdef LLAMA_CUDA_NO_PEER_COPY + HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY +endif # LLAMA_CUDA_NO_PEER_COPY OBJS += ggml-cuda.o -ggml-cuda.o: ggml-cuda.cu ggml-cuda.h + OBJS += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu)) +ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh) $(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $< + +ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh + $(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $< + endif # LLAMA_HIPBLAS ifdef LLAMA_METAL @@ -666,9 +686,15 @@ console.o: common/console.cpp common/console.h grammar-parser.o: common/grammar-parser.cpp common/grammar-parser.h $(CXX) $(CXXFLAGS) -c $< -o $@ +json-schema-to-grammar.o: common/json-schema-to-grammar.cpp common/json-schema-to-grammar.h + $(CXX) $(CXXFLAGS) -c $< -o $@ + train.o: common/train.cpp common/train.h $(CXX) $(CXXFLAGS) -c $< -o $@ +ngram-cache.o: common/ngram-cache.cpp common/ngram-cache.h + $(CXX) $(CXXFLAGS) -c $< -o $@ + libllama.so: llama.o ggml.o $(OBJS) $(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) @@ -676,7 +702,8 @@ libllama.a: llama.o ggml.o $(OBJS) $(COMMON_DEPS) ar rcs libllama.a llama.o ggml.o $(OBJS) $(COMMON_DEPS) clean: - rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS) + rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult lookup-create lookup-merge lookup-stats common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS) + rm -vrf ggml-cuda/*.o find examples pocs -type f -name "*.o" -delete # @@ -745,7 +772,7 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) +server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h common/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp json-schema-to-grammar.o common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2) @@ -794,6 +821,10 @@ export-lora: examples/export-lora/export-lora.cpp ggml.o common/common.h $(OBJS) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) +retrieval: examples/retrieval/retrieval.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) @@ -806,9 +837,15 @@ lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) +lookup: examples/lookup/lookup.cpp ggml.o llama.o ngram-cache.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + $(CXX) $(CXXFLAGS) -c examples/lookup/lookup-create.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp) -o lookup-create $(LDFLAGS) + $(CXX) $(CXXFLAGS) -c examples/lookup/lookup-merge.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp) -o lookup-merge $(LDFLAGS) + $(CXX) $(CXXFLAGS) -c examples/lookup/lookup-stats.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp) -o lookup-stats $(LDFLAGS) passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) @@ -865,6 +902,10 @@ tests/test-double-float: tests/test-double-float.cpp ggml.o $(OBJS) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) +tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp json-schema-to-grammar.o ggml.o llama.o grammar-parser.o $(OBJS) + $(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + tests/test-grad0: tests/test-grad0.cpp ggml.o $(OBJS) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) diff --git a/README-sycl.md b/README-sycl.md index 32adfda47..cbf14f2da 100644 --- a/README-sycl.md +++ b/README-sycl.md @@ -29,6 +29,7 @@ For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building). ## News - 2024.3 + - A blog is published: **Run LLM on all Intel GPUs Using llama.cpp**: [intel.com](https://www.intel.com/content/www/us/en/developer/articles/technical/run-llm-on-all-gpus-using-llama-cpp-artical.html) or [medium.com](https://medium.com/@jianyu_neo/run-llm-on-all-intel-gpus-using-llama-cpp-fd2e2dcbd9bd). - New base line is ready: [tag b2437](https://github.com/ggerganov/llama.cpp/tree/b2437). - Support multiple cards: **--split-mode**: [none|layer]; not support [row], it's on developing. - Support to assign main GPU by **--main-gpu**, replace $GGML_SYCL_DEVICE. @@ -115,7 +116,7 @@ You can choose between **F16** and **F32** build. F16 is faster for long-prompt # Or, for F32: docker build -t llama-cpp-sycl -f .devops/main-intel.Dockerfile . -# Note: you can also use the ".devops/main-server.Dockerfile", which compiles the "server" example +# Note: you can also use the ".devops/server-intel.Dockerfile", which compiles the "server" example ``` ### Run diff --git a/README.md b/README.md index c2f3342f0..f9cf19616 100644 --- a/README.md +++ b/README.md @@ -17,10 +17,12 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) ### Hot topics +- Fix major bug in Metal batched inference https://github.com/ggerganov/llama.cpp/pull/6225 - Multi-GPU pipeline parallelizm support https://github.com/ggerganov/llama.cpp/pull/6017 - Looking for contributions to add Deepseek support: https://github.com/ggerganov/llama.cpp/issues/5981 - Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962 - Initial Mamba support has been added: https://github.com/ggerganov/llama.cpp/pull/5328 +- Support loading sharded model, using `gguf-split` CLI https://github.com/ggerganov/llama.cpp/pull/6187 ---- @@ -165,6 +167,7 @@ Unless otherwise noted these projects are open-source with permissive licensing: - [cztomsik/ava](https://github.com/cztomsik/ava) (MIT) - [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal) - [pythops/tenere](https://github.com/pythops/tenere) (AGPL) +- [RecurseChat](https://recurse.chat/) (proprietary) - [semperai/amica](https://github.com/semperai/amica) - [withcatai/catai](https://github.com/withcatai/catai) - [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT) diff --git a/build.zig b/build.zig index 90609359b..a1de7083a 100644 --- a/build.zig +++ b/build.zig @@ -122,6 +122,7 @@ pub fn build(b: *std.build.Builder) !void { const console = make.obj("console", "common/console.cpp"); const sampling = make.obj("sampling", "common/sampling.cpp"); const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp"); + const json_schema_to_grammar = make.obj("json-schema-to-grammar", "common/json-schema-to-grammar.cpp"); const train = make.obj("train", "common/train.cpp"); const clip = make.obj("clip", "examples/llava/clip.cpp"); const llava = make.obj("llava", "examples/llava/llava.cpp"); @@ -133,7 +134,7 @@ pub fn build(b: *std.build.Builder) !void { _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, train }); _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, train }); - const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, sampling, grammar_parser, clip, llava }); + const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, sampling, grammar_parser, json_schema_to_grammar, clip, llava }); if (server.target.isWindows()) { server.linkSystemLibrary("ws2_32"); } diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt index af2629a46..1d840e5f7 100644 --- a/common/CMakeLists.txt +++ b/common/CMakeLists.txt @@ -47,6 +47,8 @@ if (BUILD_SHARED_LIBS) set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON) endif() +set(TARGET json-schema-to-grammar) +add_library(${TARGET} OBJECT json-schema-to-grammar.cpp json-schema-to-grammar.h) set(TARGET common) @@ -60,8 +62,11 @@ add_library(${TARGET} STATIC console.cpp grammar-parser.h grammar-parser.cpp + json.hpp train.h train.cpp + ngram-cache.h + ngram-cache.cpp ) if (BUILD_SHARED_LIBS) diff --git a/common/common.cpp b/common/common.cpp index 5f10718ec..9dec08430 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -39,6 +39,9 @@ #endif #if defined(LLAMA_USE_CURL) #include +#include +#include +#include #endif #if defined(_MSC_VER) @@ -61,7 +64,7 @@ #else #include #endif -#define LLAMA_CURL_MAX_PATH_LENGTH PATH_MAX +#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083 #define LLAMA_CURL_MAX_HEADER_LENGTH 256 #endif // LLAMA_USE_CURL @@ -101,7 +104,7 @@ int32_t get_num_physical_cores() { return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4; } -void process_escapes(std::string& input) { +void process_escapes(std::string & input) { std::size_t input_len = input.length(); std::size_t output_idx = 0; @@ -154,8 +157,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { return result; } -static bool gpt_params_find_arg(int argc, char ** argv, gpt_params & params, int & i, bool & invalid_param) { - std::string arg = argv[i]; +bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param) { llama_sampling_params& sparams = params.sparams; if (arg == "-s" || arg == "--seed") { @@ -648,14 +650,6 @@ static bool gpt_params_find_arg(int argc, char ** argv, gpt_params & params, int params.model = argv[i]; return true; } - if (arg == "-mu" || arg == "--model-url") { - if (++i >= argc) { - invalid_param = true; - return true; - } - params.model_url = argv[i]; - return true; - } if (arg == "-md" || arg == "--model-draft") { if (++i >= argc) { invalid_param = true; @@ -672,6 +666,30 @@ static bool gpt_params_find_arg(int argc, char ** argv, gpt_params & params, int params.model_alias = argv[i]; return true; } + if (arg == "-mu" || arg == "--model-url") { + if (++i >= argc) { + invalid_param = true; + return true; + } + params.model_url = argv[i]; + return true; + } + if (arg == "-hfr" || arg == "--hf-repo") { + if (++i >= argc) { + invalid_param = true; + return true; + } + params.hf_repo = argv[i]; + return true; + } + if (arg == "-hff" || arg == "--hf-file") { + if (++i >= argc) { + invalid_param = true; + return true; + } + params.hf_file = argv[i]; + return true; + } if (arg == "--lora") { if (++i >= argc) { invalid_param = true; @@ -948,6 +966,22 @@ static bool gpt_params_find_arg(int argc, char ** argv, gpt_params & params, int } return true; } + if (arg == "-lcs" || arg == "--lookup-cache-static") { + if (++i >= argc) { + invalid_param = true; + return true; + } + params.lookup_cache_static = argv[i]; + return true; + } + if (arg == "-lcd" || arg == "--lookup-cache-dynamic") { + if (++i >= argc) { + invalid_param = true; + return true; + } + params.lookup_cache_dynamic = argv[i]; + return true; + } if (arg == "--save-all-logits" || arg == "--kl-divergence-base") { if (++i >= argc) { invalid_param = true; @@ -1201,13 +1235,15 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { std::replace(arg.begin(), arg.end(), '_', '-'); } - if (!gpt_params_find_arg(argc, argv, params, i, invalid_param)) { + if (!gpt_params_find_arg(argc, argv, arg, params, i, invalid_param)) { throw std::invalid_argument("error: unknown argument: " + arg); } } + if (invalid_param) { throw std::invalid_argument("error: invalid parameter for argument: " + arg); } + if (params.prompt_cache_all && (params.interactive || params.interactive_first || params.instruct)) { @@ -1215,6 +1251,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n"); } + // short-hand to avoid specifying --hf-file -> default it to --model + if (!params.hf_repo.empty() && params.hf_file.empty()) { + params.hf_file = params.model; + } + if (params.escape) { process_escapes(params.prompt); process_escapes(params.input_prefix); @@ -1404,12 +1445,20 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" layer range to apply the control vector(s) to, start and end inclusive\n"); printf(" -m FNAME, --model FNAME\n"); printf(" model path (default: %s)\n", params.model.c_str()); - printf(" -mu MODEL_URL, --model-url MODEL_URL\n"); - printf(" model download url (default: %s)\n", params.model_url.c_str()); printf(" -md FNAME, --model-draft FNAME\n"); - printf(" draft model for speculative decoding\n"); + printf(" draft model for speculative decoding (default: unused)\n"); + printf(" -mu MODEL_URL, --model-url MODEL_URL\n"); + printf(" model download url (default: unused)\n"); + printf(" -hfr REPO, --hf-repo REPO\n"); + printf(" Hugging Face model repository (default: unused)\n"); + printf(" -hff FILE, --hf-file FILE\n"); + printf(" Hugging Face model file (default: unused)\n"); printf(" -ld LOGDIR, --logdir LOGDIR\n"); printf(" path under which to save YAML logs (no logging if unset)\n"); + printf(" -lcs FNAME, --lookup-cache-static FNAME\n"); + printf(" path to static lookup cache to use for lookup decoding (not updated by generation)\n"); + printf(" -lcd FNAME, --lookup-cache-dynamic FNAME\n"); + printf(" path to dynamic lookup cache to use for lookup decoding (updated by generation)\n"); printf(" --override-kv KEY=TYPE:VALUE\n"); printf(" advanced option to override model metadata by key. may be specified multiple times.\n"); printf(" types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n"); @@ -1590,6 +1639,9 @@ static ggml_type kv_cache_type_from_str(const std::string & s) { if (s == "q4_1") { return GGML_TYPE_Q4_1; } + if (s == "iq4_nl") { + return GGML_TYPE_IQ4_NL; + } if (s == "q5_0") { return GGML_TYPE_Q5_0; } @@ -1653,25 +1705,13 @@ void llama_batch_add( #ifdef LLAMA_USE_CURL -struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, - struct llama_model_params params) { - // Basic validation of the model_url - if (!model_url || strlen(model_url) == 0) { - fprintf(stderr, "%s: invalid model_url\n", __func__); - return NULL; - } - - // Initialize libcurl globally - auto curl = curl_easy_init(); - - if (!curl) { - fprintf(stderr, "%s: error initializing libcurl\n", __func__); - return NULL; - } +static bool llama_download_file(CURL * curl, const char * url, const char * path) { + bool force_download = false; // Set the URL, allow to follow http redirection - curl_easy_setopt(curl, CURLOPT_URL, model_url); + curl_easy_setopt(curl, CURLOPT_URL, url); curl_easy_setopt(curl, CURLOPT_FOLLOWLOCATION, 1L); + #if defined(_WIN32) // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of // operating system. Currently implemented under MS-Windows. @@ -1680,16 +1720,16 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha // Check if the file already exists locally struct stat model_file_info; - auto file_exists = (stat(path_model, &model_file_info) == 0); + auto file_exists = (stat(path, &model_file_info) == 0); // If the file exists, check for ${path_model}.etag or ${path_model}.lastModified files char etag[LLAMA_CURL_MAX_HEADER_LENGTH] = {0}; - char etag_path[LLAMA_CURL_MAX_PATH_LENGTH] = {0}; - snprintf(etag_path, sizeof(etag_path), "%s.etag", path_model); + char etag_path[PATH_MAX] = {0}; + snprintf(etag_path, sizeof(etag_path), "%s.etag", path); char last_modified[LLAMA_CURL_MAX_HEADER_LENGTH] = {0}; - char last_modified_path[LLAMA_CURL_MAX_PATH_LENGTH] = {0}; - snprintf(last_modified_path, sizeof(last_modified_path), "%s.lastModified", path_model); + char last_modified_path[PATH_MAX] = {0}; + snprintf(last_modified_path, sizeof(last_modified_path), "%s.lastModified", path); if (file_exists) { auto * f_etag = fopen(etag_path, "r"); @@ -1697,7 +1737,7 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha if (!fgets(etag, sizeof(etag), f_etag)) { fprintf(stderr, "%s: unable to read file %s\n", __func__, etag_path); } else { - fprintf(stderr, "%s: previous model file found %s: %s\n", __func__, etag_path, etag); + fprintf(stderr, "%s: previous file found %s: %s\n", __func__, etag_path, etag); } fclose(f_etag); } @@ -1707,7 +1747,7 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha if (!fgets(last_modified, sizeof(last_modified), f_last_modified)) { fprintf(stderr, "%s: unable to read file %s\n", __func__, last_modified_path); } else { - fprintf(stderr, "%s: previous model file found %s: %s\n", __func__, last_modified_path, + fprintf(stderr, "%s: previous file found %s: %s\n", __func__, last_modified_path, last_modified); } fclose(f_last_modified); @@ -1725,6 +1765,11 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t { llama_load_model_from_url_headers *headers = (llama_load_model_from_url_headers *) userdata; + // Convert header field name to lowercase + for (size_t i = 0; i < n_items && buffer[i] != ':'; ++i) { + buffer[i] = tolower(buffer[i]); + } + const char * etag_prefix = "etag: "; if (strncmp(buffer, etag_prefix, strlen(etag_prefix)) == 0) { strncpy(headers->etag, buffer + strlen(etag_prefix), n_items - strlen(etag_prefix) - 2); // Remove CRLF @@ -1747,7 +1792,7 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha if (res != CURLE_OK) { curl_easy_cleanup(curl); fprintf(stderr, "%s: curl_easy_perform() failed: %s\n", __func__, curl_easy_strerror(res)); - return NULL; + return false; } long http_code = 0; @@ -1755,30 +1800,34 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha if (http_code != 200) { // HEAD not supported, we don't know if the file has changed // force trigger downloading - file_exists = false; + force_download = true; fprintf(stderr, "%s: HEAD invalid http status code received: %ld\n", __func__, http_code); } } // If the ETag or the Last-Modified headers are different: trigger a new download - if (!file_exists || strcmp(etag, headers.etag) != 0 || strcmp(last_modified, headers.last_modified) != 0) { - char path_model_temporary[LLAMA_CURL_MAX_PATH_LENGTH] = {0}; - snprintf(path_model_temporary, sizeof(path_model_temporary), "%s.downloadInProgress", path_model); + bool should_download = !file_exists + || force_download + || (strlen(headers.etag) > 0 && strcmp(etag, headers.etag) != 0) + || (strlen(headers.last_modified) > 0 && strcmp(last_modified, headers.last_modified) != 0); + if (should_download) { + char path_temporary[PATH_MAX] = {0}; + snprintf(path_temporary, sizeof(path_temporary), "%s.downloadInProgress", path); if (file_exists) { - fprintf(stderr, "%s: deleting previous downloaded model file: %s\n", __func__, path_model); - if (remove(path_model) != 0) { + fprintf(stderr, "%s: deleting previous downloaded file: %s\n", __func__, path); + if (remove(path) != 0) { curl_easy_cleanup(curl); - fprintf(stderr, "%s: unable to delete file: %s\n", __func__, path_model); - return NULL; + fprintf(stderr, "%s: unable to delete file: %s\n", __func__, path); + return false; } } // Set the output file - auto * outfile = fopen(path_model_temporary, "wb"); + auto * outfile = fopen(path_temporary, "wb"); if (!outfile) { curl_easy_cleanup(curl); - fprintf(stderr, "%s: error opening local file for writing: %s\n", __func__, path_model); - return NULL; + fprintf(stderr, "%s: error opening local file for writing: %s\n", __func__, path); + return false; } typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd); @@ -1792,15 +1841,30 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha // display download progress curl_easy_setopt(curl, CURLOPT_NOPROGRESS, 0L); + // helper function to hide password in URL + auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string { + std::size_t protocol_pos = url.find("://"); + if (protocol_pos == std::string::npos) { + return url; // Malformed URL + } + + std::size_t at_pos = url.find('@', protocol_pos + 3); + if (at_pos == std::string::npos) { + return url; // No password in URL + } + + return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos); + }; + // start the download - fprintf(stderr, "%s: downloading model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__, - model_url, path_model, headers.etag, headers.last_modified); + fprintf(stderr, "%s: downloading from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__, + llama_download_hide_password_in_url(url).c_str(), path, headers.etag, headers.last_modified); auto res = curl_easy_perform(curl); if (res != CURLE_OK) { fclose(outfile); curl_easy_cleanup(curl); fprintf(stderr, "%s: curl_easy_perform() failed: %s\n", __func__, curl_easy_strerror(res)); - return NULL; + return false; } long http_code = 0; @@ -1809,7 +1873,7 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha fclose(outfile); curl_easy_cleanup(curl); fprintf(stderr, "%s: invalid http status code received: %ld\n", __func__, http_code); - return NULL; + return false; } // Clean up @@ -1821,7 +1885,7 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha if (etag_file) { fputs(headers.etag, etag_file); fclose(etag_file); - fprintf(stderr, "%s: model etag saved %s: %s\n", __func__, etag_path, headers.etag); + fprintf(stderr, "%s: file etag saved %s: %s\n", __func__, etag_path, headers.etag); } } @@ -1831,42 +1895,177 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha if (last_modified_file) { fputs(headers.last_modified, last_modified_file); fclose(last_modified_file); - fprintf(stderr, "%s: model last modified saved %s: %s\n", __func__, last_modified_path, + fprintf(stderr, "%s: file last modified saved %s: %s\n", __func__, last_modified_path, headers.last_modified); } } - if (rename(path_model_temporary, path_model) != 0) { + if (rename(path_temporary, path) != 0) { + curl_easy_cleanup(curl); + fprintf(stderr, "%s: unable to rename file: %s to %s\n", __func__, path_temporary, path); + return false; + } + } + + return true; +} + +struct llama_model * llama_load_model_from_url( + const char * model_url, + const char * path_model, + const struct llama_model_params & params) { + // Basic validation of the model_url + if (!model_url || strlen(model_url) == 0) { + fprintf(stderr, "%s: invalid model_url\n", __func__); + return NULL; + } + + // Initialize libcurl + auto * curl = curl_easy_init(); + + if (!curl) { + fprintf(stderr, "%s: error initializing libcurl\n", __func__); + return NULL; + } + + if (!curl) { + fprintf(stderr, "%s: error initializing libcurl\n", __func__); + return NULL; + } + + if (!llama_download_file(curl, model_url, path_model)) { + return NULL; + } + + // check for additional GGUFs split to download + int n_split = 0; + { + struct gguf_init_params gguf_params = { + /*.no_alloc = */ true, + /*.ctx = */ NULL, + }; + auto * ctx_gguf = gguf_init_from_file(path_model, gguf_params); + if (!ctx_gguf) { + fprintf(stderr, "\n%s: failed to load input GGUF from %s\n", __func__, path_model); curl_easy_cleanup(curl); - fprintf(stderr, "%s: unable to rename file: %s to %s\n", __func__, path_model_temporary, path_model); return NULL; } + + auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT); + if (key_n_split >= 0) { + n_split = gguf_get_val_u16(ctx_gguf, key_n_split); + } + + gguf_free(ctx_gguf); } curl_easy_cleanup(curl); + if (n_split > 1) { + char split_prefix[PATH_MAX] = {0}; + char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0}; + + // Verify the first split file format + // and extract split URL and PATH prefixes + { + if (!llama_split_prefix(split_prefix, sizeof(split_prefix), path_model, 0, n_split)) { + fprintf(stderr, "\n%s: unexpected model file name: %s" + " n_split=%d\n", __func__, path_model, n_split); + return NULL; + } + + if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model_url, 0, n_split)) { + fprintf(stderr, "\n%s: unexpected model url: %s" + " n_split=%d\n", __func__, model_url, n_split); + return NULL; + } + } + + // Prepare download in parallel + std::vector> futures_download; + for (int idx = 1; idx < n_split; idx++) { + futures_download.push_back(std::async(std::launch::async, [&split_prefix, &split_url_prefix, &n_split](int download_idx) -> bool { + char split_path[PATH_MAX] = {0}; + llama_split_path(split_path, sizeof(split_path), split_prefix, download_idx, n_split); + + char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0}; + llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split); + + auto * curl = curl_easy_init(); + bool res = llama_download_file(curl, split_url, split_path); + curl_easy_cleanup(curl); + + return res; + }, idx)); + } + + // Wait for all downloads to complete + for (auto & f : futures_download) { + if (!f.get()) { + return NULL; + } + } + } + return llama_load_model_from_file(path_model, params); } +struct llama_model * llama_load_model_from_hf( + const char * repo, + const char * model, + const char * path_model, + const struct llama_model_params & params) { + // construct hugging face model url: + // + // --repo ggml-org/models --file tinyllama-1.1b/ggml-model-f16.gguf + // https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf + // + // --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf + // https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf + // + + std::string model_url = "https://huggingface.co/"; + model_url += repo; + model_url += "/resolve/main/"; + model_url += model; + + return llama_load_model_from_url(model_url.c_str(), path_model, params); +} + #else -struct llama_model * llama_load_model_from_url(const char * /*model_url*/, const char * /*path_model*/, - struct llama_model_params /*params*/) { +struct llama_model * llama_load_model_from_url( + const char * /*model_url*/, + const char * /*path_model*/, + const struct llama_model_params & /*params*/) { fprintf(stderr, "%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__); return nullptr; } +struct llama_model * llama_load_model_from_hf( + const char * /*repo*/, + const char * /*model*/, + const char * /*path_model*/, + const struct llama_model_params & /*params*/) { + fprintf(stderr, "%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__); + return nullptr; +} + #endif // LLAMA_USE_CURL std::tuple llama_init_from_gpt_params(gpt_params & params) { auto mparams = llama_model_params_from_gpt_params(params); llama_model * model = nullptr; - if (!params.model_url.empty()) { + + if (!params.hf_repo.empty() && !params.hf_file.empty()) { + model = llama_load_model_from_hf(params.hf_repo.c_str(), params.hf_file.c_str(), params.model.c_str(), mparams); + } else if (!params.model_url.empty()) { model = llama_load_model_from_url(params.model_url.c_str(), params.model.c_str(), mparams); } else { model = llama_load_model_from_file(params.model.c_str(), mparams); } + if (model == NULL) { fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str()); return std::make_tuple(nullptr, nullptr); @@ -1906,7 +2105,7 @@ std::tuple llama_init_from_gpt_par } for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) { - const std::string& lora_adapter = std::get<0>(params.lora_adapter[i]); + const std::string & lora_adapter = std::get<0>(params.lora_adapter[i]); float lora_scale = std::get<1>(params.lora_adapter[i]); int err = llama_model_apply_lora_from_file(model, lora_adapter.c_str(), diff --git a/common/common.h b/common/common.h index 8dd8a3edc..99ee90bc3 100644 --- a/common/common.h +++ b/common/common.h @@ -88,18 +88,22 @@ struct gpt_params { // // sampling parameters struct llama_sampling_params sparams; - std::string model = "models/7B/ggml-model-f16.gguf"; // model path - std::string model_url = ""; // model url to download - std::string model_draft = ""; // draft model for speculative decoding - std::string model_alias = "unknown"; // model alias - std::string prompt = ""; - std::string prompt_file = ""; // store the external prompt file name - std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state - std::string input_prefix = ""; // string to prefix user inputs with - std::string input_suffix = ""; // string to suffix user inputs with + std::string model = "models/7B/ggml-model-f16.gguf"; // model path + std::string model_draft = ""; // draft model for speculative decoding + std::string model_alias = "unknown"; // model alias + std::string model_url = ""; // model url to download + std::string hf_repo = ""; // HF repo + std::string hf_file = ""; // HF file + std::string prompt = ""; + std::string prompt_file = ""; // store the external prompt file name + std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state + std::string input_prefix = ""; // string to prefix user inputs with + std::string input_suffix = ""; // string to suffix user inputs with std::vector antiprompt; // string upon seeing which more user input is prompted - std::string logdir = ""; // directory in which to save YAML log files - std::string logits_file = ""; // file for saving *all* logits + std::string logdir = ""; // directory in which to save YAML log files + std::string lookup_cache_static = ""; // path of static ngram cache file for lookup decoding + std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding + std::string logits_file = ""; // file for saving *all* logits std::vector kv_overrides; @@ -139,7 +143,7 @@ struct gpt_params { bool interactive_first = false; // wait for user input immediately bool multiline_input = false; // reverse the usage of `\` bool simple_io = false; // improves compatibility with subprocesses and limited consoles - bool cont_batching = false; // insert new sequences for decoding on-the-fly + bool cont_batching = true; // insert new sequences for decoding on-the-fly bool input_prefix_bos = false; // prefix BOS to user inputs, preceding input_prefix bool ignore_eos = false; // ignore generated EOS tokens @@ -167,6 +171,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params); void gpt_print_usage(int argc, char ** argv, const gpt_params & params); +bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param); + std::string get_system_info(const gpt_params & params); std::string gpt_random_prompt(std::mt19937 & rng); @@ -192,8 +198,8 @@ std::tuple llama_init_from_gpt_par struct llama_model_params llama_model_params_from_gpt_params (const gpt_params & params); struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params); -struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, - struct llama_model_params params); +struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const struct llama_model_params & params); +struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const struct llama_model_params & params); // Batch utils @@ -302,3 +308,10 @@ struct llama_control_vector_load_info { // Load control vectors, scale each by strength, and add them together. // On error, returns {-1, empty} llama_control_vector_data llama_control_vector_load(const std::vector & load_infos); + +// +// Split utils +// +static const char * const LLM_KV_SPLIT_NO = "split.no"; +static const char * const LLM_KV_SPLIT_COUNT = "split.count"; +static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count"; diff --git a/common/json-schema-to-grammar.cpp b/common/json-schema-to-grammar.cpp new file mode 100644 index 000000000..0e4680346 --- /dev/null +++ b/common/json-schema-to-grammar.cpp @@ -0,0 +1,721 @@ +#include "json-schema-to-grammar.h" +#include +#include +#include +#include +#include +#include +#include +#include +#include + +using json = nlohmann::ordered_json; + +const std::string SPACE_RULE = "\" \"?"; + +std::unordered_map PRIMITIVE_RULES = { + {"boolean", "(\"true\" | \"false\") space"}, + {"number", "(\"-\"? ([0-9] | [1-9] [0-9]*)) (\".\" [0-9]+)? ([eE] [-+]? [0-9]+)? space"}, + {"integer", "(\"-\"? ([0-9] | [1-9] [0-9]*)) space"}, + {"value", "object | array | string | number | boolean"}, + {"object", "\"{\" space ( string \":\" space value (\",\" space string \":\" space value)* )? \"}\" space"}, + {"array", "\"[\" space ( value (\",\" space value)* )? \"]\" space"}, + {"uuid", "\"\\\"\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] " + "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] " + "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] " + "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] " + "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] \"\\\"\" space"}, + {"string", " \"\\\"\" (\n" + " [^\"\\\\] |\n" + " \"\\\\\" ([\"\\\\/bfnrt] | \"u\" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])\n" + " )* \"\\\"\" space"}, + {"null", "\"null\" space"} +}; +std::vector OBJECT_RULE_NAMES = {"object", "array", "string", "number", "boolean", "null", "value"}; + +std::unordered_map DATE_RULES = { + {"date", "[0-9] [0-9] [0-9] [0-9] \"-\" ( \"0\" [1-9] | \"1\" [0-2] ) \"-\" ( \"0\" [1-9] | [1-2] [0-9] | \"3\" [0-1] )"}, + {"time", "([01] [0-9] | \"2\" [0-3]) \":\" [0-5] [0-9] \":\" [0-5] [0-9] ( \".\" [0-9] [0-9] [0-9] )? ( \"Z\" | ( \"+\" | \"-\" ) ( [01] [0-9] | \"2\" [0-3] ) \":\" [0-5] [0-9] )"}, + {"date-time", "date \"T\" time"}, + {"date-string", "\"\\\"\" date \"\\\"\" space"}, + {"time-string", "\"\\\"\" time \"\\\"\" space"}, + {"date-time-string", "\"\\\"\" date-time \"\\\"\" space"} +}; + +static bool is_reserved_name(const std::string & name) { + static std::unordered_set RESERVED_NAMES; + if (RESERVED_NAMES.empty()) { + RESERVED_NAMES.insert("root"); + for (const auto &p : PRIMITIVE_RULES) RESERVED_NAMES.insert(p.first); + for (const auto &p : DATE_RULES) RESERVED_NAMES.insert(p.first); + } + return RESERVED_NAMES.find(name) != RESERVED_NAMES.end(); +} + +std::regex INVALID_RULE_CHARS_RE("[^a-zA-Z0-9-]+"); +std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"]"); +std::regex GRAMMAR_RANGE_LITERAL_ESCAPE_RE("[\r\n\"\\]\\-\\\\]"); +std::unordered_map GRAMMAR_LITERAL_ESCAPES = { + {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"} +}; + +std::unordered_set NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'}; +std::unordered_set ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'[', ']', '(', ')', '|', '{', '}', '*', '+', '?'}; + +template +std::string join(Iterator begin, Iterator end, const std::string & separator) { + std::ostringstream result; + if (begin != end) { + result << *begin; + for (Iterator it = begin + 1; it != end; ++it) { + result << separator << *it; + } + } + return result.str(); +} + +static std::vector split(const std::string & str, const std::string & delimiter) { + std::vector tokens; + size_t start = 0; + size_t end = str.find(delimiter); + + while (end != std::string::npos) { + tokens.push_back(str.substr(start, end - start)); + start = end + delimiter.length(); + end = str.find(delimiter, start); + } + + tokens.push_back(str.substr(start)); + + return tokens; +} + +static std::string repeat(const std::string & str, size_t n) { + if (n == 0) { + return ""; + } + + std::string result; + result.reserve(str.length() * n); + + for (size_t i = 0; i < n; ++i) { + result += str; + } + + return result; +} + +static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function & replacement) { + std::smatch match; + std::string result; + + std::string::const_iterator searchStart(input.cbegin()); + std::string::const_iterator searchEnd(input.cend()); + + while (std::regex_search(searchStart, searchEnd, match, regex)) { + result.append(searchStart, searchStart + match.position()); + result.append(replacement(match)); + searchStart = match.suffix().first; + } + + result.append(searchStart, searchEnd); + + return result; +} + +static std::string format_literal(const std::string & literal) { + std::string escaped = replacePattern(literal, GRAMMAR_LITERAL_ESCAPE_RE, [&](const std::smatch & match) { + char c = match.str()[0]; + return GRAMMAR_LITERAL_ESCAPES.at(c); + }); + return "\"" + escaped + "\""; +} + + +class SchemaConverter { +private: + std::function _fetch_json; + bool _dotall; + std::map _rules; + std::unordered_map _refs; + std::unordered_set _refs_being_resolved; + std::vector _errors; + std::vector _warnings; + + std::string _add_rule(const std::string & name, const std::string & rule) { + std::string esc_name = regex_replace(name, INVALID_RULE_CHARS_RE, "-"); + if (_rules.find(esc_name) == _rules.end() || _rules[esc_name] == rule) { + _rules[esc_name] = rule; + return esc_name; + } else { + int i = 0; + while (_rules.find(esc_name + std::to_string(i)) != _rules.end() && _rules[esc_name + std::to_string(i)] != rule) { + i++; + } + std::string key = esc_name + std::to_string(i); + _rules[key] = rule; + return key; + } + } + + std::string _generate_union_rule(const std::string & name, const std::vector & alt_schemas) { + std::vector rules; + for (size_t i = 0; i < alt_schemas.size(); i++) { + rules.push_back(visit(alt_schemas[i], name + (name.empty() ? "alternative-" : "-") + std::to_string(i))); + } + return join(rules.begin(), rules.end(), " | "); + } + + std::string _visit_pattern(const std::string & pattern, const std::string & name) { + if (!(pattern.front() == '^' && pattern.back() == '$')) { + _errors.push_back("Pattern must start with '^' and end with '$'"); + return ""; + } + std::string sub_pattern = pattern.substr(1, pattern.length() - 2); + std::unordered_map sub_rule_ids; + + size_t i = 0; + size_t length = sub_pattern.length(); + + using literal_or_rule = std::pair; + auto to_rule = [&](const literal_or_rule & ls) { + auto is_literal = ls.second; + auto s = ls.first; + return is_literal ? "\"" + s + "\"" : s; + }; + std::function transform = [&]() -> literal_or_rule { + size_t start = i; + std::vector seq; + + auto get_dot = [&]() { + std::string rule; + if (_dotall) { + rule = "[\\U00000000-\\U0010FFFF]"; + } else { + rule = "[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]"; + } + return _add_rule("dot", rule); + }; + + // Joins the sequence, merging consecutive literals together. + auto join_seq = [&]() { + std::vector ret; + + std::string literal; + auto flush_literal = [&]() { + if (literal.empty()) { + return false; + } + ret.push_back(std::make_pair(literal, true)); + literal.clear(); + return true; + }; + + for (const auto & item : seq) { + auto is_literal = item.second; + if (is_literal) { + literal += item.first; + } else { + flush_literal(); + ret.push_back(item); + } + } + flush_literal(); + + std::vector results; + for (const auto & item : ret) { + results.push_back(to_rule(item)); + } + return std::make_pair(join(results.begin(), results.end(), " "), false); + }; + + while (i < length) { + char c = sub_pattern[i]; + if (c == '.') { + seq.push_back(std::make_pair(get_dot(), false)); + i++; + } else if (c == '(') { + i++; + if (i < length) { + if (sub_pattern[i] == '?') { + _warnings.push_back("Unsupported pattern syntax"); + } + } + seq.push_back(std::make_pair("(" + to_rule(transform()) + ")", false)); + } else if (c == ')') { + i++; + if (start > 0 && sub_pattern[start - 1] != '(') { + _errors.push_back("Unbalanced parentheses"); + } + return join_seq(); + } else if (c == '[') { + std::string square_brackets = std::string(1, c); + i++; + while (i < length && sub_pattern[i] != ']') { + if (sub_pattern[i] == '\\') { + square_brackets += sub_pattern.substr(i, 2); + i += 2; + } else { + square_brackets += sub_pattern[i]; + i++; + } + } + if (i >= length) { + _errors.push_back("Unbalanced square brackets"); + } + square_brackets += ']'; + i++; + seq.push_back(std::make_pair(square_brackets, false)); + } else if (c == '|') { + seq.push_back(std::make_pair("|", false)); + i++; + } else if (c == '*' || c == '+' || c == '?') { + seq.back() = std::make_pair(to_rule(seq.back()) + c, false); + i++; + } else if (c == '{') { + std::string curly_brackets = std::string(1, c); + i++; + while (i < length && sub_pattern[i] != '}') { + curly_brackets += sub_pattern[i]; + i++; + } + if (i >= length) { + _errors.push_back("Unbalanced curly brackets"); + } + curly_brackets += '}'; + i++; + auto nums = split(curly_brackets.substr(1, curly_brackets.length() - 2), ","); + int min_times = 0; + int max_times = std::numeric_limits::max(); + try { + if (nums.size() == 1) { + min_times = max_times = std::stoi(nums[0]); + } else if (nums.size() != 2) { + _errors.push_back("Wrong number of values in curly brackets"); + } else { + if (!nums[0].empty()) { + min_times = std::stoi(nums[0]); + } + if (!nums[1].empty()) { + max_times = std::stoi(nums[1]); + } + } + } catch (const std::invalid_argument & e) { + _errors.push_back("Invalid number in curly brackets"); + return std::make_pair("", false); + } + auto &last = seq.back(); + auto &sub = last.first; + auto sub_is_literal = last.second; + + if (min_times == 0 && max_times == std::numeric_limits::max()) { + sub += "*"; + } else if (min_times == 0 && max_times == 1) { + sub += "?"; + } else if (min_times == 1 && max_times == std::numeric_limits::max()) { + sub += "+"; + } else { + if (!sub_is_literal) { + std::string & sub_id = sub_rule_ids[sub]; + if (sub_id.empty()) { + sub_id = _add_rule(name + "-" + std::to_string(sub_rule_ids.size()), sub); + } + sub = sub_id; + } + std::string result; + if (sub_is_literal && min_times > 0) { + result = "\"" + repeat(sub.substr(1, sub.length() - 2), min_times) + "\""; + } else { + for (int j = 0; j < min_times; j++) { + if (j > 0) { + result += " "; + } + result += sub; + } + } + if (min_times > 0 && min_times < max_times) { + result += " "; + } + if (max_times == std::numeric_limits::max()) { + result += sub + "*"; + } else { + for (int j = min_times; j < max_times; j++) { + if (j > min_times) { + result += " "; + } + result += sub + "?"; + } + } + seq.back().first = result; + seq.back().second = false; + } + } else { + std::string literal; + auto is_non_literal = [&](char c) { + return NON_LITERAL_SET.find(c) != NON_LITERAL_SET.end(); + }; + while (i < length) { + if (sub_pattern[i] == '\\' && i < length - 1) { + char next = sub_pattern[i + 1]; + if (ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.find(next) != ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.end()) { + i++; + literal += sub_pattern[i]; + i++; + } else { + literal += sub_pattern.substr(i, 2); + i += 2; + } + } else if (sub_pattern[i] == '"') { + literal += "\\\""; + i++; + } else if (!is_non_literal(sub_pattern[i]) && + (i == length - 1 || literal.empty() || sub_pattern[i + 1] == '.' || !is_non_literal(sub_pattern[i + 1]))) { + literal += sub_pattern[i]; + i++; + } else { + break; + } + } + if (!literal.empty()) { + seq.push_back(std::make_pair(literal, true)); + } + } + } + return join_seq(); + }; + return _add_rule(name, "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space"); + } + + std::string _resolve_ref(const std::string & ref) { + std::string ref_name = ref.substr(ref.find_last_of('/') + 1); + if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) { + _refs_being_resolved.insert(ref); + json resolved = _refs[ref]; + ref_name = visit(resolved, ref_name); + _refs_being_resolved.erase(ref); + } + return ref_name; + } + + std::string _build_object_rule( + const std::vector> & properties, + const std::unordered_set & required, + const std::string & name, + const json & additional_properties) + { + std::vector required_props; + std::vector optional_props; + std::unordered_map prop_kv_rule_names; + for (const auto & kv : properties) { + const auto &prop_name = kv.first; + const auto &prop_schema = kv.second; + + std::string prop_rule_name = visit(prop_schema, name + (name.empty() ? "" : "-") + prop_name); + prop_kv_rule_names[prop_name] = _add_rule( + name + (name.empty() ? "" : "-") + prop_name + "-kv", + format_literal(json(prop_name).dump()) + " space \":\" space " + prop_rule_name + ); + if (required.find(prop_name) != required.end()) { + required_props.push_back(prop_name); + } else { + optional_props.push_back(prop_name); + } + } + if (additional_properties.is_object() || (additional_properties.is_boolean() && additional_properties.get())) { + std::string sub_name = name + (name.empty() ? "" : "-") + "additional"; + std::string value_rule = visit(additional_properties.is_object() ? additional_properties : json::object(), sub_name + "-value"); + std::string kv_rule = _add_rule(sub_name + "-kv", _add_rule("string", PRIMITIVE_RULES.at("string")) + " \":\" space " + value_rule); + prop_kv_rule_names["*"] = kv_rule; + optional_props.push_back("*"); + } + + std::string rule = "\"{\" space "; + for (size_t i = 0; i < required_props.size(); i++) { + if (i > 0) { + rule += " \",\" space "; + } + rule += prop_kv_rule_names[required_props[i]]; + } + + if (!optional_props.empty()) { + rule += " ("; + if (!required_props.empty()) { + rule += " \",\" space ( "; + } + + std::function &, bool)> get_recursive_refs = [&](const std::vector & ks, bool first_is_optional) { + std::string res; + if (ks.empty()) { + return res; + } + std::string k = ks[0]; + std::string kv_rule_name = prop_kv_rule_names[k]; + if (k == "*") { + res = _add_rule( + name + (name.empty() ? "" : "-") + "additional-kvs", + kv_rule_name + " ( \",\" space " + kv_rule_name + " )*" + ); + } else if (first_is_optional) { + res = "( \",\" space " + kv_rule_name + " )?"; + } else { + res = kv_rule_name; + } + if (ks.size() > 1) { + res += " " + _add_rule( + name + (name.empty() ? "" : "-") + k + "-rest", + get_recursive_refs(std::vector(ks.begin() + 1, ks.end()), true) + ); + } + return res; + }; + + for (size_t i = 0; i < optional_props.size(); i++) { + if (i > 0) { + rule += " | "; + } + rule += get_recursive_refs(std::vector(optional_props.begin() + i, optional_props.end()), false); + } + if (!required_props.empty()) { + rule += " )"; + } + rule += " )?"; + } + + rule += " \"}\" space"; + + return rule; + } + +public: + SchemaConverter( + const std::function & fetch_json, + bool dotall) + : _fetch_json(fetch_json), _dotall(dotall) + { + _rules["space"] = SPACE_RULE; + } + + void resolve_refs(json & schema, const std::string & url) { + /* + * Resolves all $ref fields in the given schema, fetching any remote schemas, + * replacing each $ref with absolute reference URL and populates _refs with the + * respective referenced (sub)schema dictionaries. + */ + std::function visit_refs = [&](json & n) { + if (n.is_array()) { + for (auto & x : n) { + visit_refs(x); + } + } else if (n.is_object()) { + if (n.contains("$ref")) { + std::string ref = n["$ref"]; + if (_refs.find(ref) == _refs.end()) { + json target; + if (ref.find("https://") == 0) { + std::string base_url = ref.substr(0, ref.find('#')); + auto it = _refs.find(base_url); + if (it != _refs.end()) { + target = it->second; + } else { + // Fetch the referenced schema and resolve its refs + auto referenced = _fetch_json(ref); + resolve_refs(referenced, base_url); + _refs[base_url] = referenced; + } + if (ref.find('#') == std::string::npos || ref.substr(ref.find('#') + 1).empty()) { + return; + } + } else if (ref.find("#/") == 0) { + target = schema; + n["$ref"] = url + ref; + ref = url + ref; + } else { + _errors.push_back("Unsupported ref: " + ref); + return; + } + std::string pointer = ref.substr(ref.find('#') + 1); + std::vector tokens = split(pointer, "/"); + for (size_t i = 1; i < tokens.size(); ++i) { + std::string sel = tokens[i]; + if (target.is_null() || !target.contains(sel)) { + _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump()); + return; + } + target = target[sel]; + } + _refs[ref] = target; + } + } else { + for (auto & kv : n.items()) { + visit_refs(kv.value()); + } + } + } + }; + + visit_refs(schema); + } + + std::string _generate_constant_rule(const json & value) { + return format_literal(value.dump()); + } + + std::string visit(const json & schema, const std::string & name) { + json schema_type = schema.contains("type") ? schema["type"] : json(); + std::string schema_format = schema.contains("format") ? schema["format"].get() : ""; + std::string rule_name = is_reserved_name(name) ? name + "-" : name.empty() ? "root" : name; + + if (schema.contains("$ref")) { + return _add_rule(rule_name, _resolve_ref(schema["$ref"])); + } else if (schema.contains("oneOf") || schema.contains("anyOf")) { + std::vector alt_schemas = schema.contains("oneOf") ? schema["oneOf"].get>() : schema["anyOf"].get>(); + return _add_rule(rule_name, _generate_union_rule(name, alt_schemas)); + } else if (schema_type.is_array()) { + std::vector schema_types; + for (const auto & t : schema_type) { + schema_types.push_back({{"type", t}}); + } + return _add_rule(rule_name, _generate_union_rule(name, schema_types)); + } else if (schema.contains("const")) { + return _add_rule(rule_name, _generate_constant_rule(schema["const"])); + } else if (schema.contains("enum")) { + std::vector enum_values; + for (const auto & v : schema["enum"]) { + enum_values.push_back(_generate_constant_rule(v)); + } + return _add_rule(rule_name, join(enum_values.begin(), enum_values.end(), " | ")); + } else if ((schema_type.is_null() || schema_type == "object") + && (schema.contains("properties") || + (schema.contains("additionalProperties") && schema["additionalProperties"] != true))) { + std::unordered_set required; + if (schema.contains("required") && schema["required"].is_array()) { + for (const auto & item : schema["required"]) { + if (item.is_string()) { + required.insert(item.get()); + } + } + } + std::vector> properties; + if (schema.contains("properties")) { + for (const auto & prop : schema["properties"].items()) { + properties.emplace_back(prop.key(), prop.value()); + } + } + return _add_rule(rule_name, + _build_object_rule( + properties, required, name, + schema.contains("additionalProperties") ? schema["additionalProperties"] : json())); + } else if ((schema_type.is_null() || schema_type == "object") && schema.contains("allOf")) { + std::unordered_set required; + std::vector> properties; + std::string hybrid_name = name; + std::function add_component = [&](const json & comp_schema, bool is_required) { + if (comp_schema.contains("$ref")) { + add_component(_refs[comp_schema["$ref"]], is_required); + } else if (comp_schema.contains("properties")) { + for (const auto & prop : comp_schema["properties"].items()) { + properties.emplace_back(prop.key(), prop.value()); + if (is_required) { + required.insert(prop.key()); + } + } + } else { + // todo warning + } + }; + for (auto & t : schema["allOf"]) { + if (t.contains("anyOf")) { + for (auto & tt : t["anyOf"]) { + add_component(tt, false); + } + } else { + add_component(t, true); + } + } + return _add_rule(rule_name, _build_object_rule(properties, required, hybrid_name, json())); + } else if ((schema_type.is_null() || schema_type == "array") && (schema.contains("items") || schema.contains("prefixItems"))) { + json items = schema.contains("items") ? schema["items"] : schema["prefixItems"]; + if (items.is_array()) { + std::string rule = "\"[\" space "; + for (size_t i = 0; i < items.size(); i++) { + if (i > 0) { + rule += " \",\" space "; + } + rule += visit(items[i], name + (name.empty() ? "" : "-") + "tuple-" + std::to_string(i)); + } + rule += " \"]\" space"; + return _add_rule(rule_name, rule); + } else { + std::string item_rule_name = visit(items, name + (name.empty() ? "" : "-") + "item"); + std::string list_item_operator = "( \",\" space " + item_rule_name + " )"; + std::string successive_items; + int min_items = schema.contains("minItems") ? schema["minItems"].get() : 0; + json max_items_json = schema.contains("maxItems") ? schema["maxItems"] : json(); + int max_items = max_items_json.is_number_integer() ? max_items_json.get() : -1; + if (min_items > 0) { + successive_items += repeat(list_item_operator, min_items - 1); + min_items--; + } + if (max_items >= 0 && max_items > min_items) { + successive_items += repeat(list_item_operator + "?", max_items - min_items - 1); + } else { + successive_items += list_item_operator + "*"; + } + std::string rule; + if (min_items == 0) { + rule = "\"[\" space ( " + item_rule_name + " " + successive_items + " )? \"]\" space"; + } else { + rule = "\"[\" space " + item_rule_name + " " + successive_items + " \"]\" space"; + } + return _add_rule(rule_name, rule); + } + } else if ((schema_type.is_null() || schema_type == "string") && schema.contains("pattern")) { + return _visit_pattern(schema["pattern"], rule_name); + } else if ((schema_type.is_null() || schema_type == "string") && std::regex_match(schema_format, std::regex("^uuid[1-5]?$"))) { + return _add_rule(rule_name == "root" ? "root" : schema_format, PRIMITIVE_RULES.at("uuid")); + } else if ((schema_type.is_null() || schema_type == "string") && DATE_RULES.find(schema_format) != DATE_RULES.end()) { + for (const auto & kv : DATE_RULES) { + _add_rule(kv.first, kv.second); + } + return schema_format + "-string"; + } else if (schema.empty() || schema_type == "object") { + for (const auto & n : OBJECT_RULE_NAMES) { + _add_rule(n, PRIMITIVE_RULES.at(n)); + } + return _add_rule(rule_name, "object"); + } else { + if (!schema_type.is_string() || PRIMITIVE_RULES.find(schema_type.get()) == PRIMITIVE_RULES.end()) { + _errors.push_back("Unrecognized schema: " + schema.dump()); + return ""; + } + // TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero + return _add_rule(rule_name == "root" ? "root" : schema_type.get(), PRIMITIVE_RULES.at(schema_type.get())); + } + } + + void check_errors() { + if (!_errors.empty()) { + throw std::runtime_error("JSON schema conversion failed:\n" + join(_errors.begin(), _errors.end(), "\n")); + } + if (!_warnings.empty()) { + fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", join(_warnings.begin(), _warnings.end(), "; ").c_str()); + } + } + + std::string format_grammar() { + std::stringstream ss; + for (const auto & kv : _rules) { + ss << kv.first << " ::= " << kv.second << std::endl; + } + return ss.str(); + } +}; + +std::string json_schema_to_grammar(const json & schema) { + SchemaConverter converter([](const std::string &) { return json::object(); }, /* dotall= */ false); + auto copy = schema; + converter.resolve_refs(copy, "input"); + converter.visit(copy, ""); + converter.check_errors(); + return converter.format_grammar(); +} diff --git a/common/json-schema-to-grammar.h b/common/json-schema-to-grammar.h new file mode 100644 index 000000000..e1abed303 --- /dev/null +++ b/common/json-schema-to-grammar.h @@ -0,0 +1,4 @@ +#pragma once +#include "json.hpp" + +std::string json_schema_to_grammar(const nlohmann::ordered_json& schema); diff --git a/examples/server/json.hpp b/common/json.hpp similarity index 93% rename from examples/server/json.hpp rename to common/json.hpp index ea945f346..a858728c4 100644 --- a/examples/server/json.hpp +++ b/common/json.hpp @@ -1,9 +1,9 @@ // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT /****************************************************************************\ @@ -27,7 +27,6 @@ #endif // JSON_NO_IO #include // random_access_iterator_tag #include // unique_ptr -#include // accumulate #include // string, stoi, to_string #include // declval, forward, move, pair, swap #include // vector @@ -35,10 +34,10 @@ // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -48,10 +47,10 @@ // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -60,7 +59,7 @@ #ifndef JSON_SKIP_LIBRARY_VERSION_CHECK #if defined(NLOHMANN_JSON_VERSION_MAJOR) && defined(NLOHMANN_JSON_VERSION_MINOR) && defined(NLOHMANN_JSON_VERSION_PATCH) - #if NLOHMANN_JSON_VERSION_MAJOR != 3 || NLOHMANN_JSON_VERSION_MINOR != 11 || NLOHMANN_JSON_VERSION_PATCH != 2 + #if NLOHMANN_JSON_VERSION_MAJOR != 3 || NLOHMANN_JSON_VERSION_MINOR != 11 || NLOHMANN_JSON_VERSION_PATCH != 3 #warning "Already included a different version of the library!" #endif #endif @@ -68,7 +67,7 @@ #define NLOHMANN_JSON_VERSION_MAJOR 3 // NOLINT(modernize-macro-to-enum) #define NLOHMANN_JSON_VERSION_MINOR 11 // NOLINT(modernize-macro-to-enum) -#define NLOHMANN_JSON_VERSION_PATCH 2 // NOLINT(modernize-macro-to-enum) +#define NLOHMANN_JSON_VERSION_PATCH 3 // NOLINT(modernize-macro-to-enum) #ifndef JSON_DIAGNOSTICS #define JSON_DIAGNOSTICS 0 @@ -150,10 +149,10 @@ // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -173,16 +172,19 @@ // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT #include // nullptr_t #include // exception +#if JSON_DIAGNOSTICS + #include // accumulate +#endif #include // runtime_error #include // to_string #include // vector @@ -190,10 +192,10 @@ // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -206,10 +208,10 @@ // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -218,10 +220,10 @@ // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -231,10 +233,10 @@ // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -318,10 +320,10 @@ NLOHMANN_JSON_NAMESPACE_END // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-FileCopyrightText: 2016-2021 Evan Nemerson // SPDX-License-Identifier: MIT @@ -2483,6 +2485,14 @@ JSON_HEDLEY_DIAGNOSTIC_POP #endif #endif +#ifndef JSON_HAS_STATIC_RTTI + #if !defined(_HAS_STATIC_RTTI) || _HAS_STATIC_RTTI != 0 + #define JSON_HAS_STATIC_RTTI 1 + #else + #define JSON_HAS_STATIC_RTTI 0 + #endif +#endif + #ifdef JSON_HAS_CPP_17 #define JSON_INLINE_VARIABLE inline #else @@ -2590,12 +2600,13 @@ JSON_HEDLEY_DIAGNOSTIC_POP class NumberUnsignedType, class NumberFloatType, \ template class AllocatorType, \ template class JSONSerializer, \ - class BinaryType> + class BinaryType, \ + class CustomBaseClass> #define NLOHMANN_BASIC_JSON_TPL \ basic_json + AllocatorType, JSONSerializer, BinaryType, CustomBaseClass> // Macros to simplify conversion from/to types @@ -2745,7 +2756,10 @@ JSON_HEDLEY_DIAGNOSTIC_POP #define NLOHMANN_DEFINE_TYPE_INTRUSIVE_WITH_DEFAULT(Type, ...) \ friend void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \ - friend void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { Type nlohmann_json_default_obj; NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM_WITH_DEFAULT, __VA_ARGS__)) } + friend void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { const Type nlohmann_json_default_obj{}; NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM_WITH_DEFAULT, __VA_ARGS__)) } + +#define NLOHMANN_DEFINE_TYPE_INTRUSIVE_ONLY_SERIALIZE(Type, ...) \ + friend void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } /*! @brief macro @@ -2756,10 +2770,12 @@ JSON_HEDLEY_DIAGNOSTIC_POP inline void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \ inline void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM, __VA_ARGS__)) } +#define NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE_ONLY_SERIALIZE(Type, ...) \ + inline void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } + #define NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE_WITH_DEFAULT(Type, ...) \ inline void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \ - inline void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { Type nlohmann_json_default_obj; NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM_WITH_DEFAULT, __VA_ARGS__)) } - + inline void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { const Type nlohmann_json_default_obj{}; NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM_WITH_DEFAULT, __VA_ARGS__)) } // inspired from https://stackoverflow.com/a/26745591 // allows to call any std function as if (e.g. with begin): @@ -2923,10 +2939,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -2998,10 +3014,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -3040,10 +3056,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-FileCopyrightText: 2018 The Abseil Authors // SPDX-License-Identifier: MIT @@ -3214,10 +3230,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -3226,14 +3242,15 @@ NLOHMANN_JSON_NAMESPACE_END #include // false_type, is_constructible, is_integral, is_same, true_type #include // declval #include // tuple +#include // char_traits // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -3298,10 +3315,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -3318,10 +3335,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -3342,10 +3359,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT #ifndef INCLUDE_NLOHMANN_JSON_FWD_HPP_ @@ -3389,7 +3406,8 @@ NLOHMANN_JSON_NAMESPACE_END template class AllocatorType = std::allocator, template class JSONSerializer = adl_serializer, - class BinaryType = std::vector> + class BinaryType = std::vector, // cppcheck-suppress syntaxError + class CustomBaseClass = void> class basic_json; /// @brief JSON Pointer defines a string syntax for identifying a specific value within a JSON document @@ -3577,6 +3595,63 @@ struct actual_object_comparator template using actual_object_comparator_t = typename actual_object_comparator::type; +///////////////// +// char_traits // +///////////////// + +// Primary template of char_traits calls std char_traits +template +struct char_traits : std::char_traits +{}; + +// Explicitly define char traits for unsigned char since it is not standard +template<> +struct char_traits : std::char_traits +{ + using char_type = unsigned char; + using int_type = uint64_t; + + // Redefine to_int_type function + static int_type to_int_type(char_type c) noexcept + { + return static_cast(c); + } + + static char_type to_char_type(int_type i) noexcept + { + return static_cast(i); + } + + static constexpr int_type eof() noexcept + { + return static_cast(EOF); + } +}; + +// Explicitly define char traits for signed char since it is not standard +template<> +struct char_traits : std::char_traits +{ + using char_type = signed char; + using int_type = uint64_t; + + // Redefine to_int_type function + static int_type to_int_type(char_type c) noexcept + { + return static_cast(c); + } + + static char_type to_char_type(int_type i) noexcept + { + return static_cast(i); + } + + static constexpr int_type eof() noexcept + { + return static_cast(EOF); + } +}; + /////////////////// // is_ functions // /////////////////// @@ -3613,7 +3688,6 @@ template struct is_default_constructible> : conjunction...> {}; - template struct is_constructible : std::is_constructible {}; @@ -3629,7 +3703,6 @@ struct is_constructible> : is_default_constructible struct is_constructible> : is_default_constructible> {}; - template struct is_iterator_traits : std::false_type {}; @@ -4039,7 +4112,6 @@ struct value_in_range_of_impl2 } }; - template struct value_in_range_of_impl2 { @@ -4138,10 +4210,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -4165,28 +4237,28 @@ inline std::size_t concat_length() } template -inline std::size_t concat_length(const char* cstr, Args&& ... rest); +inline std::size_t concat_length(const char* cstr, const Args& ... rest); template -inline std::size_t concat_length(const StringType& str, Args&& ... rest); +inline std::size_t concat_length(const StringType& str, const Args& ... rest); template -inline std::size_t concat_length(const char /*c*/, Args&& ... rest) +inline std::size_t concat_length(const char /*c*/, const Args& ... rest) { - return 1 + concat_length(std::forward(rest)...); + return 1 + concat_length(rest...); } template -inline std::size_t concat_length(const char* cstr, Args&& ... rest) +inline std::size_t concat_length(const char* cstr, const Args& ... rest) { // cppcheck-suppress ignoredReturnValue - return ::strlen(cstr) + concat_length(std::forward(rest)...); + return ::strlen(cstr) + concat_length(rest...); } template -inline std::size_t concat_length(const StringType& str, Args&& ... rest) +inline std::size_t concat_length(const StringType& str, const Args& ... rest) { - return str.size() + concat_length(std::forward(rest)...); + return str.size() + concat_length(rest...); } template @@ -4277,7 +4349,7 @@ template inline OutStringType concat(Args && ... args) { OutStringType str; - str.reserve(concat_length(std::forward(args)...)); + str.reserve(concat_length(args...)); concat_into(str, std::forward(args)...); return str; } @@ -4286,7 +4358,6 @@ inline OutStringType concat(Args && ... args) NLOHMANN_JSON_NAMESPACE_END - NLOHMANN_JSON_NAMESPACE_BEGIN namespace detail { @@ -4334,9 +4405,9 @@ class exception : public std::exception { case value_t::array: { - for (std::size_t i = 0; i < current->m_parent->m_value.array->size(); ++i) + for (std::size_t i = 0; i < current->m_parent->m_data.m_value.array->size(); ++i) { - if (¤t->m_parent->m_value.array->operator[](i) == current) + if (¤t->m_parent->m_data.m_value.array->operator[](i) == current) { tokens.emplace_back(std::to_string(i)); break; @@ -4347,7 +4418,7 @@ class exception : public std::exception case value_t::object: { - for (const auto& element : *current->m_parent->m_value.object) + for (const auto& element : *current->m_parent->m_data.m_value.object) { if (&element.second == current) { @@ -4410,17 +4481,17 @@ class parse_error : public exception template::value, int> = 0> static parse_error create(int id_, const position_t& pos, const std::string& what_arg, BasicJsonContext context) { - std::string w = concat(exception::name("parse_error", id_), "parse error", - position_string(pos), ": ", exception::diagnostics(context), what_arg); + const std::string w = concat(exception::name("parse_error", id_), "parse error", + position_string(pos), ": ", exception::diagnostics(context), what_arg); return {id_, pos.chars_read_total, w.c_str()}; } template::value, int> = 0> static parse_error create(int id_, std::size_t byte_, const std::string& what_arg, BasicJsonContext context) { - std::string w = concat(exception::name("parse_error", id_), "parse error", - (byte_ != 0 ? (concat(" at byte ", std::to_string(byte_))) : ""), - ": ", exception::diagnostics(context), what_arg); + const std::string w = concat(exception::name("parse_error", id_), "parse error", + (byte_ != 0 ? (concat(" at byte ", std::to_string(byte_))) : ""), + ": ", exception::diagnostics(context), what_arg); return {id_, byte_, w.c_str()}; } @@ -4454,7 +4525,7 @@ class invalid_iterator : public exception template::value, int> = 0> static invalid_iterator create(int id_, const std::string& what_arg, BasicJsonContext context) { - std::string w = concat(exception::name("invalid_iterator", id_), exception::diagnostics(context), what_arg); + const std::string w = concat(exception::name("invalid_iterator", id_), exception::diagnostics(context), what_arg); return {id_, w.c_str()}; } @@ -4472,7 +4543,7 @@ class type_error : public exception template::value, int> = 0> static type_error create(int id_, const std::string& what_arg, BasicJsonContext context) { - std::string w = concat(exception::name("type_error", id_), exception::diagnostics(context), what_arg); + const std::string w = concat(exception::name("type_error", id_), exception::diagnostics(context), what_arg); return {id_, w.c_str()}; } @@ -4489,7 +4560,7 @@ class out_of_range : public exception template::value, int> = 0> static out_of_range create(int id_, const std::string& what_arg, BasicJsonContext context) { - std::string w = concat(exception::name("out_of_range", id_), exception::diagnostics(context), what_arg); + const std::string w = concat(exception::name("out_of_range", id_), exception::diagnostics(context), what_arg); return {id_, w.c_str()}; } @@ -4506,7 +4577,7 @@ class other_error : public exception template::value, int> = 0> static other_error create(int id_, const std::string& what_arg, BasicJsonContext context) { - std::string w = concat(exception::name("other_error", id_), exception::diagnostics(context), what_arg); + const std::string w = concat(exception::name("other_error", id_), exception::diagnostics(context), what_arg); return {id_, w.c_str()}; } @@ -4525,10 +4596,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -4549,10 +4620,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -5055,10 +5126,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -5075,10 +5146,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -5147,10 +5218,10 @@ template class iteration_proxy_value // older GCCs are a bit fussy and require explicit noexcept specifiers on defaulted functions iteration_proxy_value(iteration_proxy_value&&) noexcept(std::is_nothrow_move_constructible::value - && std::is_nothrow_move_constructible::value) = default; + && std::is_nothrow_move_constructible::value) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor,cppcoreguidelines-noexcept-move-operations) iteration_proxy_value& operator=(iteration_proxy_value&&) noexcept(std::is_nothrow_move_assignable::value - && std::is_nothrow_move_assignable::value) = default; + && std::is_nothrow_move_assignable::value) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor,cppcoreguidelines-noexcept-move-operations) ~iteration_proxy_value() = default; /// dereference operator (needed for range-based for) @@ -5297,11 +5368,11 @@ namespace std #pragma clang diagnostic ignored "-Wmismatched-tags" #endif template -class tuple_size<::nlohmann::detail::iteration_proxy_value> +class tuple_size<::nlohmann::detail::iteration_proxy_value> // NOLINT(cert-dcl58-cpp) : public std::integral_constant {}; template -class tuple_element> +class tuple_element> // NOLINT(cert-dcl58-cpp) { public: using type = decltype( @@ -5340,7 +5411,7 @@ namespace detail /* * Note all external_constructor<>::construct functions need to call - * j.m_value.destroy(j.m_type) to avoid a memory leak in case j contains an + * j.m_data.m_value.destroy(j.m_data.m_type) to avoid a memory leak in case j contains an * allocated value (e.g., a string). See bug issue * https://github.com/nlohmann/json/issues/2865 for more information. */ @@ -5353,9 +5424,9 @@ struct external_constructor template static void construct(BasicJsonType& j, typename BasicJsonType::boolean_t b) noexcept { - j.m_value.destroy(j.m_type); - j.m_type = value_t::boolean; - j.m_value = b; + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::boolean; + j.m_data.m_value = b; j.assert_invariant(); } }; @@ -5366,18 +5437,18 @@ struct external_constructor template static void construct(BasicJsonType& j, const typename BasicJsonType::string_t& s) { - j.m_value.destroy(j.m_type); - j.m_type = value_t::string; - j.m_value = s; + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::string; + j.m_data.m_value = s; j.assert_invariant(); } template static void construct(BasicJsonType& j, typename BasicJsonType::string_t&& s) { - j.m_value.destroy(j.m_type); - j.m_type = value_t::string; - j.m_value = std::move(s); + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::string; + j.m_data.m_value = std::move(s); j.assert_invariant(); } @@ -5386,9 +5457,9 @@ struct external_constructor int > = 0 > static void construct(BasicJsonType& j, const CompatibleStringType& str) { - j.m_value.destroy(j.m_type); - j.m_type = value_t::string; - j.m_value.string = j.template create(str); + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::string; + j.m_data.m_value.string = j.template create(str); j.assert_invariant(); } }; @@ -5399,18 +5470,18 @@ struct external_constructor template static void construct(BasicJsonType& j, const typename BasicJsonType::binary_t& b) { - j.m_value.destroy(j.m_type); - j.m_type = value_t::binary; - j.m_value = typename BasicJsonType::binary_t(b); + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::binary; + j.m_data.m_value = typename BasicJsonType::binary_t(b); j.assert_invariant(); } template static void construct(BasicJsonType& j, typename BasicJsonType::binary_t&& b) { - j.m_value.destroy(j.m_type); - j.m_type = value_t::binary; - j.m_value = typename BasicJsonType::binary_t(std::move(b)); + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::binary; + j.m_data.m_value = typename BasicJsonType::binary_t(std::move(b)); j.assert_invariant(); } }; @@ -5421,9 +5492,9 @@ struct external_constructor template static void construct(BasicJsonType& j, typename BasicJsonType::number_float_t val) noexcept { - j.m_value.destroy(j.m_type); - j.m_type = value_t::number_float; - j.m_value = val; + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::number_float; + j.m_data.m_value = val; j.assert_invariant(); } }; @@ -5434,9 +5505,9 @@ struct external_constructor template static void construct(BasicJsonType& j, typename BasicJsonType::number_unsigned_t val) noexcept { - j.m_value.destroy(j.m_type); - j.m_type = value_t::number_unsigned; - j.m_value = val; + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::number_unsigned; + j.m_data.m_value = val; j.assert_invariant(); } }; @@ -5447,9 +5518,9 @@ struct external_constructor template static void construct(BasicJsonType& j, typename BasicJsonType::number_integer_t val) noexcept { - j.m_value.destroy(j.m_type); - j.m_type = value_t::number_integer; - j.m_value = val; + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::number_integer; + j.m_data.m_value = val; j.assert_invariant(); } }; @@ -5460,9 +5531,9 @@ struct external_constructor template static void construct(BasicJsonType& j, const typename BasicJsonType::array_t& arr) { - j.m_value.destroy(j.m_type); - j.m_type = value_t::array; - j.m_value = arr; + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::array; + j.m_data.m_value = arr; j.set_parents(); j.assert_invariant(); } @@ -5470,9 +5541,9 @@ struct external_constructor template static void construct(BasicJsonType& j, typename BasicJsonType::array_t&& arr) { - j.m_value.destroy(j.m_type); - j.m_type = value_t::array; - j.m_value = std::move(arr); + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::array; + j.m_data.m_value = std::move(arr); j.set_parents(); j.assert_invariant(); } @@ -5485,9 +5556,9 @@ struct external_constructor using std::begin; using std::end; - j.m_value.destroy(j.m_type); - j.m_type = value_t::array; - j.m_value.array = j.template create(begin(arr), end(arr)); + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::array; + j.m_data.m_value.array = j.template create(begin(arr), end(arr)); j.set_parents(); j.assert_invariant(); } @@ -5495,14 +5566,14 @@ struct external_constructor template static void construct(BasicJsonType& j, const std::vector& arr) { - j.m_value.destroy(j.m_type); - j.m_type = value_t::array; - j.m_value = value_t::array; - j.m_value.array->reserve(arr.size()); + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::array; + j.m_data.m_value = value_t::array; + j.m_data.m_value.array->reserve(arr.size()); for (const bool x : arr) { - j.m_value.array->push_back(x); - j.set_parent(j.m_value.array->back()); + j.m_data.m_value.array->push_back(x); + j.set_parent(j.m_data.m_value.array->back()); } j.assert_invariant(); } @@ -5511,13 +5582,13 @@ struct external_constructor enable_if_t::value, int> = 0> static void construct(BasicJsonType& j, const std::valarray& arr) { - j.m_value.destroy(j.m_type); - j.m_type = value_t::array; - j.m_value = value_t::array; - j.m_value.array->resize(arr.size()); + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::array; + j.m_data.m_value = value_t::array; + j.m_data.m_value.array->resize(arr.size()); if (arr.size() > 0) { - std::copy(std::begin(arr), std::end(arr), j.m_value.array->begin()); + std::copy(std::begin(arr), std::end(arr), j.m_data.m_value.array->begin()); } j.set_parents(); j.assert_invariant(); @@ -5530,9 +5601,9 @@ struct external_constructor template static void construct(BasicJsonType& j, const typename BasicJsonType::object_t& obj) { - j.m_value.destroy(j.m_type); - j.m_type = value_t::object; - j.m_value = obj; + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::object; + j.m_data.m_value = obj; j.set_parents(); j.assert_invariant(); } @@ -5540,9 +5611,9 @@ struct external_constructor template static void construct(BasicJsonType& j, typename BasicJsonType::object_t&& obj) { - j.m_value.destroy(j.m_type); - j.m_type = value_t::object; - j.m_value = std::move(obj); + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::object; + j.m_data.m_value = std::move(obj); j.set_parents(); j.assert_invariant(); } @@ -5554,9 +5625,9 @@ struct external_constructor using std::begin; using std::end; - j.m_value.destroy(j.m_type); - j.m_type = value_t::object; - j.m_value.object = j.template create(begin(obj), end(obj)); + j.m_data.m_value.destroy(j.m_data.m_type); + j.m_data.m_type = value_t::object; + j.m_data.m_value.object = j.template create(begin(obj), end(obj)); j.set_parents(); j.assert_invariant(); } @@ -5626,7 +5697,8 @@ template::type; - external_constructor::construct(j, static_cast(e)); + static constexpr value_t integral_value_t = std::is_unsigned::value ? value_t::number_unsigned : value_t::number_integer; + external_constructor::construct(j, static_cast(e)); } #endif // JSON_DISABLE_ENUM_SERIALIZATION @@ -5796,10 +5868,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -5908,10 +5980,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -6041,10 +6113,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -6067,10 +6139,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -6094,6 +6166,8 @@ NLOHMANN_JSON_NAMESPACE_END // #include +// #include + NLOHMANN_JSON_NAMESPACE_BEGIN namespace detail @@ -6140,7 +6214,6 @@ class file_input_adapter std::FILE* m_file; }; - /*! Input adapter for a (caching) istream. Ignores a UFT Byte Order Mark at beginning of input. Does not support changing the underlying std::streambuf @@ -6214,16 +6287,16 @@ class iterator_input_adapter : current(std::move(first)), end(std::move(last)) {} - typename std::char_traits::int_type get_character() + typename char_traits::int_type get_character() { if (JSON_HEDLEY_LIKELY(current != end)) { - auto result = std::char_traits::to_int_type(*current); + auto result = char_traits::to_int_type(*current); std::advance(current, 1); return result; } - return std::char_traits::eof(); + return char_traits::eof(); } private: @@ -6239,7 +6312,6 @@ class iterator_input_adapter } }; - template struct wide_string_input_helper; @@ -6363,7 +6435,7 @@ struct wide_string_input_helper } }; -// Wraps another input apdater to convert wide character types into individual bytes. +// Wraps another input adapter to convert wide character types into individual bytes. template class wide_string_input_adapter { @@ -6408,7 +6480,6 @@ class wide_string_input_adapter std::size_t utf8_bytes_filled = 0; }; - template struct iterator_input_adapter_factory { @@ -6565,10 +6636,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -6710,7 +6781,6 @@ struct json_sax virtual ~json_sax() = default; }; - namespace detail { /*! @@ -6812,7 +6882,7 @@ class json_sax_dom_parser JSON_ASSERT(ref_stack.back()->is_object()); // add null at given key and store the reference for later - object_element = &(ref_stack.back()->m_value.object->operator[](val)); + object_element = &(ref_stack.back()->m_data.m_value.object->operator[](val)); return true; } @@ -6887,8 +6957,8 @@ class json_sax_dom_parser if (ref_stack.back()->is_array()) { - ref_stack.back()->m_value.array->emplace_back(std::forward(v)); - return &(ref_stack.back()->m_value.array->back()); + ref_stack.back()->m_data.m_value.array->emplace_back(std::forward(v)); + return &(ref_stack.back()->m_data.m_value.array->back()); } JSON_ASSERT(ref_stack.back()->is_object()); @@ -7007,7 +7077,7 @@ class json_sax_dom_callback_parser // add discarded value at given key and store the reference for later if (keep && ref_stack.back()) { - object_element = &(ref_stack.back()->m_value.object->operator[](val) = discarded); + object_element = &(ref_stack.back()->m_data.m_value.object->operator[](val) = discarded); } return true; @@ -7092,7 +7162,7 @@ class json_sax_dom_callback_parser // remove discarded value if (!keep && !ref_stack.empty() && ref_stack.back()->is_array()) { - ref_stack.back()->m_value.array->pop_back(); + ref_stack.back()->m_data.m_value.array->pop_back(); } return true; @@ -7159,7 +7229,7 @@ class json_sax_dom_callback_parser if (ref_stack.empty()) { root = std::move(value); - return {true, &root}; + return {true, & root}; } // skip this value if we already decided to skip the parent @@ -7175,8 +7245,8 @@ class json_sax_dom_callback_parser // array if (ref_stack.back()->is_array()) { - ref_stack.back()->m_value.array->emplace_back(std::move(value)); - return {true, &(ref_stack.back()->m_value.array->back())}; + ref_stack.back()->m_data.m_value.array->emplace_back(std::move(value)); + return {true, & (ref_stack.back()->m_data.m_value.array->back())}; } // object @@ -7201,9 +7271,9 @@ class json_sax_dom_callback_parser /// stack to model hierarchy of values std::vector ref_stack {}; /// stack to manage which values to keep - std::vector keep_stack {}; + std::vector keep_stack {}; // NOLINT(readability-redundant-member-init) /// stack to manage which object keys to keep - std::vector key_keep_stack {}; + std::vector key_keep_stack {}; // NOLINT(readability-redundant-member-init) /// helper to hold the reference for the next object element BasicJsonType* object_element = nullptr; /// whether a syntax error occurred @@ -7298,10 +7368,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -7322,6 +7392,8 @@ NLOHMANN_JSON_NAMESPACE_END // #include +// #include + NLOHMANN_JSON_NAMESPACE_BEGIN namespace detail @@ -7416,7 +7488,7 @@ class lexer : public lexer_base using number_float_t = typename BasicJsonType::number_float_t; using string_t = typename BasicJsonType::string_t; using char_type = typename InputAdapterType::char_type; - using char_int_type = typename std::char_traits::int_type; + using char_int_type = typename char_traits::int_type; public: using token_type = typename lexer_base::token_type; @@ -7523,7 +7595,7 @@ class lexer : public lexer_base for (auto range = ranges.begin(); range != ranges.end(); ++range) { get(); - if (JSON_HEDLEY_LIKELY(*range <= current && current <= *(++range))) + if (JSON_HEDLEY_LIKELY(*range <= current && current <= *(++range))) // NOLINT(bugprone-inc-dec-in-conditions) { add(current); } @@ -7566,7 +7638,7 @@ class lexer : public lexer_base switch (get()) { // end of file while parsing string - case std::char_traits::eof(): + case char_traits::eof(): { error_message = "invalid string: missing closing quote"; return token_type::parse_error; @@ -8155,7 +8227,7 @@ class lexer : public lexer_base { case '\n': case '\r': - case std::char_traits::eof(): + case char_traits::eof(): case '\0': return true; @@ -8172,7 +8244,7 @@ class lexer : public lexer_base { switch (get()) { - case std::char_traits::eof(): + case char_traits::eof(): case '\0': { error_message = "invalid comment; missing closing '*/'"; @@ -8601,10 +8673,10 @@ scan_number_done: token_type scan_literal(const char_type* literal_text, const std::size_t length, token_type return_type) { - JSON_ASSERT(std::char_traits::to_char_type(current) == literal_text[0]); + JSON_ASSERT(char_traits::to_char_type(current) == literal_text[0]); for (std::size_t i = 1; i < length; ++i) { - if (JSON_HEDLEY_UNLIKELY(std::char_traits::to_char_type(get()) != literal_text[i])) + if (JSON_HEDLEY_UNLIKELY(char_traits::to_char_type(get()) != literal_text[i])) { error_message = "invalid literal"; return token_type::parse_error; @@ -8622,7 +8694,7 @@ scan_number_done: { token_buffer.clear(); token_string.clear(); - token_string.push_back(std::char_traits::to_char_type(current)); + token_string.push_back(char_traits::to_char_type(current)); } /* @@ -8630,7 +8702,7 @@ scan_number_done: This function provides the interface to the used input adapter. It does not throw in case the input reached EOF, but returns a - `std::char_traits::eof()` in that case. Stores the scanned characters + `char_traits::eof()` in that case. Stores the scanned characters for use in error messages. @return character read from the input @@ -8650,9 +8722,9 @@ scan_number_done: current = ia.get_character(); } - if (JSON_HEDLEY_LIKELY(current != std::char_traits::eof())) + if (JSON_HEDLEY_LIKELY(current != char_traits::eof())) { - token_string.push_back(std::char_traits::to_char_type(current)); + token_string.push_back(char_traits::to_char_type(current)); } if (current == '\n') @@ -8691,7 +8763,7 @@ scan_number_done: --position.chars_read_current_line; } - if (JSON_HEDLEY_LIKELY(current != std::char_traits::eof())) + if (JSON_HEDLEY_LIKELY(current != char_traits::eof())) { JSON_ASSERT(!token_string.empty()); token_string.pop_back(); @@ -8885,7 +8957,7 @@ scan_number_done: // end of input (the null byte is needed when parsing from // string literals) case '\0': - case std::char_traits::eof(): + case char_traits::eof(): return token_type::end_of_input; // error @@ -8903,7 +8975,7 @@ scan_number_done: const bool ignore_comments = false; /// the current character - char_int_type current = std::char_traits::eof(); + char_int_type current = char_traits::eof(); /// whether the next get() call should just return current bool next_unget = false; @@ -8937,10 +9009,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -9129,7 +9201,6 @@ static inline bool little_endianness(int num = 1) noexcept return *reinterpret_cast(&num) == 1; } - /////////////////// // binary reader // /////////////////// @@ -9147,7 +9218,7 @@ class binary_reader using binary_t = typename BasicJsonType::binary_t; using json_sax_t = SAX; using char_type = typename InputAdapterType::char_type; - using char_int_type = typename std::char_traits::int_type; + using char_int_type = typename char_traits::int_type; public: /*! @@ -9220,7 +9291,7 @@ class binary_reader get(); } - if (JSON_HEDLEY_UNLIKELY(current != std::char_traits::eof())) + if (JSON_HEDLEY_UNLIKELY(current != char_traits::eof())) { return sax->parse_error(chars_read, get_token_string(), parse_error::create(110, chars_read, exception_message(input_format, concat("expected end of input; last byte: 0x", get_token_string()), "value"), nullptr)); @@ -9303,7 +9374,7 @@ class binary_reader exception_message(input_format_t::bson, concat("string length must be at least 1, is ", std::to_string(len)), "string"), nullptr)); } - return get_string(input_format_t::bson, len - static_cast(1), result) && get() != std::char_traits::eof(); + return get_string(input_format_t::bson, len - static_cast(1), result) && get() != char_traits::eof(); } /*! @@ -9404,7 +9475,7 @@ class binary_reader { std::array cr{{}}; static_cast((std::snprintf)(cr.data(), cr.size(), "%.2hhX", static_cast(element_type))); // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg) - std::string cr_str{cr.data()}; + const std::string cr_str{cr.data()}; return sax->parse_error(element_type_parse_position, cr_str, parse_error::create(114, element_type_parse_position, concat("Unsupported BSON record type 0x", cr_str), nullptr)); } @@ -9497,7 +9568,7 @@ class binary_reader switch (get_char ? get() : current) { // EOF - case std::char_traits::eof(): + case char_traits::eof(): return unexpect_eof(input_format_t::cbor, "value"); // Integer 0x00..0x17 (0..23) @@ -10272,7 +10343,7 @@ class binary_reader switch (get()) { // EOF - case std::char_traits::eof(): + case char_traits::eof(): return unexpect_eof(input_format_t::msgpack, "value"); // positive fixint @@ -11339,7 +11410,7 @@ class binary_reader exception_message(input_format, concat("expected '#' after type information; last byte: 0x", last_token), "size"), nullptr)); } - bool is_error = get_ubjson_size_value(result.first, is_ndarray); + const bool is_error = get_ubjson_size_value(result.first, is_ndarray); if (input_format == input_format_t::bjdata && is_ndarray) { if (inside_ndarray) @@ -11354,7 +11425,7 @@ class binary_reader if (current == '#') { - bool is_error = get_ubjson_size_value(result.first, is_ndarray); + const bool is_error = get_ubjson_size_value(result.first, is_ndarray); if (input_format == input_format_t::bjdata && is_ndarray) { return sax->parse_error(chars_read, get_token_string(), parse_error::create(112, chars_read, @@ -11374,7 +11445,7 @@ class binary_reader { switch (prefix) { - case std::char_traits::eof(): // EOF + case char_traits::eof(): // EOF return unexpect_eof(input_format, "value"); case 'T': // true @@ -11819,7 +11890,7 @@ class binary_reader This function provides the interface to the used input adapter. It does not throw in case the input reached EOF, but returns a -'ve valued - `std::char_traits::eof()` in that case. + `char_traits::eof()` in that case. @return character read from the input */ @@ -11961,7 +12032,7 @@ class binary_reader JSON_HEDLEY_NON_NULL(3) bool unexpect_eof(const input_format_t format, const char* context) const { - if (JSON_HEDLEY_UNLIKELY(current == std::char_traits::eof())) + if (JSON_HEDLEY_UNLIKELY(current == char_traits::eof())) { return sax->parse_error(chars_read, "", parse_error::create(110, chars_read, exception_message(format, "unexpected end of input", context), nullptr)); @@ -12028,7 +12099,7 @@ class binary_reader InputAdapterType ia; /// the current character - char_int_type current = std::char_traits::eof(); + char_int_type current = char_traits::eof(); /// the number of characters read std::size_t chars_read = 0; @@ -12090,10 +12161,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -12439,13 +12510,25 @@ class parser m_lexer.get_token_string(), parse_error::create(101, m_lexer.get_position(), exception_message(token_type::uninitialized, "value"), nullptr)); } + case token_type::end_of_input: + { + if (JSON_HEDLEY_UNLIKELY(m_lexer.get_position().chars_read_total == 1)) + { + return sax->parse_error(m_lexer.get_position(), + m_lexer.get_token_string(), + parse_error::create(101, m_lexer.get_position(), + "attempting to parse an empty input; check that your input string or stream contains the expected JSON", nullptr)); + } + return sax->parse_error(m_lexer.get_position(), + m_lexer.get_token_string(), + parse_error::create(101, m_lexer.get_position(), exception_message(token_type::literal_or_value, "value"), nullptr)); + } case token_type::uninitialized: case token_type::end_array: case token_type::end_object: case token_type::name_separator: case token_type::value_separator: - case token_type::end_of_input: case token_type::literal_or_value: default: // the last token was unexpected { @@ -12607,10 +12690,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -12620,10 +12703,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -12779,10 +12862,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -12887,7 +12970,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci { JSON_ASSERT(m_object != nullptr); - switch (m_object->m_type) + switch (m_object->m_data.m_type) { case value_t::object: { @@ -12984,17 +13067,17 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci { JSON_ASSERT(m_object != nullptr); - switch (m_object->m_type) + switch (m_object->m_data.m_type) { case value_t::object: { - m_it.object_iterator = m_object->m_value.object->begin(); + m_it.object_iterator = m_object->m_data.m_value.object->begin(); break; } case value_t::array: { - m_it.array_iterator = m_object->m_value.array->begin(); + m_it.array_iterator = m_object->m_data.m_value.array->begin(); break; } @@ -13028,17 +13111,17 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci { JSON_ASSERT(m_object != nullptr); - switch (m_object->m_type) + switch (m_object->m_data.m_type) { case value_t::object: { - m_it.object_iterator = m_object->m_value.object->end(); + m_it.object_iterator = m_object->m_data.m_value.object->end(); break; } case value_t::array: { - m_it.array_iterator = m_object->m_value.array->end(); + m_it.array_iterator = m_object->m_data.m_value.array->end(); break; } @@ -13067,17 +13150,17 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci { JSON_ASSERT(m_object != nullptr); - switch (m_object->m_type) + switch (m_object->m_data.m_type) { case value_t::object: { - JSON_ASSERT(m_it.object_iterator != m_object->m_value.object->end()); + JSON_ASSERT(m_it.object_iterator != m_object->m_data.m_value.object->end()); return m_it.object_iterator->second; } case value_t::array: { - JSON_ASSERT(m_it.array_iterator != m_object->m_value.array->end()); + JSON_ASSERT(m_it.array_iterator != m_object->m_data.m_value.array->end()); return *m_it.array_iterator; } @@ -13111,17 +13194,17 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci { JSON_ASSERT(m_object != nullptr); - switch (m_object->m_type) + switch (m_object->m_data.m_type) { case value_t::object: { - JSON_ASSERT(m_it.object_iterator != m_object->m_value.object->end()); + JSON_ASSERT(m_it.object_iterator != m_object->m_data.m_value.object->end()); return &(m_it.object_iterator->second); } case value_t::array: { - JSON_ASSERT(m_it.array_iterator != m_object->m_value.array->end()); + JSON_ASSERT(m_it.array_iterator != m_object->m_data.m_value.array->end()); return &*m_it.array_iterator; } @@ -13164,7 +13247,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci { JSON_ASSERT(m_object != nullptr); - switch (m_object->m_type) + switch (m_object->m_data.m_type) { case value_t::object: { @@ -13215,7 +13298,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci { JSON_ASSERT(m_object != nullptr); - switch (m_object->m_type) + switch (m_object->m_data.m_type) { case value_t::object: { @@ -13262,7 +13345,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci JSON_ASSERT(m_object != nullptr); - switch (m_object->m_type) + switch (m_object->m_data.m_type) { case value_t::object: return (m_it.object_iterator == other.m_it.object_iterator); @@ -13307,7 +13390,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci JSON_ASSERT(m_object != nullptr); - switch (m_object->m_type) + switch (m_object->m_data.m_type) { case value_t::object: JSON_THROW(invalid_iterator::create(213, "cannot compare order of object iterators", m_object)); @@ -13363,7 +13446,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci { JSON_ASSERT(m_object != nullptr); - switch (m_object->m_type) + switch (m_object->m_data.m_type) { case value_t::object: JSON_THROW(invalid_iterator::create(209, "cannot use offsets with object iterators", m_object)); @@ -13442,7 +13525,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci { JSON_ASSERT(m_object != nullptr); - switch (m_object->m_type) + switch (m_object->m_data.m_type) { case value_t::object: JSON_THROW(invalid_iterator::create(209, "cannot use offsets with object iterators", m_object)); @@ -13471,7 +13554,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci { JSON_ASSERT(m_object != nullptr); - switch (m_object->m_type) + switch (m_object->m_data.m_type) { case value_t::object: JSON_THROW(invalid_iterator::create(208, "cannot use operator[] for object iterators", m_object)); @@ -13541,10 +13624,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -13673,13 +13756,55 @@ NLOHMANN_JSON_NAMESPACE_END // #include +// #include +// __ _____ _____ _____ +// __| | __| | | | JSON for Modern C++ +// | | |__ | | | | | | version 3.11.3 +// |_____|_____|_____|_|___| https://github.com/nlohmann/json +// +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann +// SPDX-License-Identifier: MIT + + + +#include // conditional, is_same + +// #include + + +NLOHMANN_JSON_NAMESPACE_BEGIN +namespace detail +{ + +/*! +@brief Default base class of the @ref basic_json class. + +So that the correct implementations of the copy / move ctors / assign operators +of @ref basic_json do not require complex case distinctions +(no base class / custom base class used as customization point), +@ref basic_json always has a base class. +By default, this class is used because it is empty and thus has no effect +on the behavior of @ref basic_json. +*/ +struct json_default_base {}; + +template +using json_base_class = typename std::conditional < + std::is_same::value, + json_default_base, + T + >::type; + +} // namespace detail +NLOHMANN_JSON_NAMESPACE_END + // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -13911,7 +14036,7 @@ class json_pointer const char* p = s.c_str(); char* p_end = nullptr; errno = 0; // strtoull doesn't reset errno - unsigned long long res = std::strtoull(p, &p_end, 10); // NOLINT(runtime/int) + const unsigned long long res = std::strtoull(p, &p_end, 10); // NOLINT(runtime/int) if (p == p_end // invalid input or empty string || errno == ERANGE // out of range || JSON_HEDLEY_UNLIKELY(static_cast(p_end - p) != s.size())) // incomplete read @@ -14067,7 +14192,7 @@ class json_pointer if (reference_token == "-") { // explicitly treat "-" as index beyond the end - ptr = &ptr->operator[](ptr->m_value.array->size()); + ptr = &ptr->operator[](ptr->m_data.m_value.array->size()); } else { @@ -14119,7 +14244,7 @@ class json_pointer { // "-" always fails the range check JSON_THROW(detail::out_of_range::create(402, detail::concat( - "array index '-' (", std::to_string(ptr->m_value.array->size()), + "array index '-' (", std::to_string(ptr->m_data.m_value.array->size()), ") is out of range"), ptr)); } @@ -14176,7 +14301,7 @@ class json_pointer if (JSON_HEDLEY_UNLIKELY(reference_token == "-")) { // "-" cannot be used for const access - JSON_THROW(detail::out_of_range::create(402, detail::concat("array index '-' (", std::to_string(ptr->m_value.array->size()), ") is out of range"), ptr)); + JSON_THROW(detail::out_of_range::create(402, detail::concat("array index '-' (", std::to_string(ptr->m_data.m_value.array->size()), ") is out of range"), ptr)); } // use unchecked array access @@ -14226,7 +14351,7 @@ class json_pointer { // "-" always fails the range check JSON_THROW(detail::out_of_range::create(402, detail::concat( - "array index '-' (", std::to_string(ptr->m_value.array->size()), + "array index '-' (", std::to_string(ptr->m_data.m_value.array->size()), ") is out of range"), ptr)); } @@ -14421,7 +14546,7 @@ class json_pointer { case detail::value_t::array: { - if (value.m_value.array->empty()) + if (value.m_data.m_value.array->empty()) { // flatten empty array as null result[reference_string] = nullptr; @@ -14429,10 +14554,10 @@ class json_pointer else { // iterate array and use index as reference string - for (std::size_t i = 0; i < value.m_value.array->size(); ++i) + for (std::size_t i = 0; i < value.m_data.m_value.array->size(); ++i) { flatten(detail::concat(reference_string, '/', std::to_string(i)), - value.m_value.array->operator[](i), result); + value.m_data.m_value.array->operator[](i), result); } } break; @@ -14440,7 +14565,7 @@ class json_pointer case detail::value_t::object: { - if (value.m_value.object->empty()) + if (value.m_data.m_value.object->empty()) { // flatten empty object as null result[reference_string] = nullptr; @@ -14448,7 +14573,7 @@ class json_pointer else { // iterate object and use keys as reference string - for (const auto& element : *value.m_value.object) + for (const auto& element : *value.m_data.m_value.object) { flatten(detail::concat(reference_string, '/', detail::escape(element.first)), element.second, result); } @@ -14495,7 +14620,7 @@ class json_pointer BasicJsonType result; // iterate the JSON object values - for (const auto& element : *value.m_value.object) + for (const auto& element : *value.m_data.m_value.object) { if (JSON_HEDLEY_UNLIKELY(!element.second.is_primitive())) { @@ -14671,10 +14796,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -14763,10 +14888,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -14789,10 +14914,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -14978,7 +15103,7 @@ class binary_writer { case value_t::object: { - write_bson_object(*j.m_value.object); + write_bson_object(*j.m_data.m_value.object); break; } @@ -15013,7 +15138,7 @@ class binary_writer case value_t::boolean: { - oa->write_character(j.m_value.boolean + oa->write_character(j.m_data.m_value.boolean ? to_char_type(0xF5) : to_char_type(0xF4)); break; @@ -15021,42 +15146,42 @@ class binary_writer case value_t::number_integer: { - if (j.m_value.number_integer >= 0) + if (j.m_data.m_value.number_integer >= 0) { // CBOR does not differentiate between positive signed // integers and unsigned integers. Therefore, we used the // code from the value_t::number_unsigned case here. - if (j.m_value.number_integer <= 0x17) + if (j.m_data.m_value.number_integer <= 0x17) { - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_integer <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_integer <= (std::numeric_limits::max)()) { oa->write_character(to_char_type(0x18)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_integer <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_integer <= (std::numeric_limits::max)()) { oa->write_character(to_char_type(0x19)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_integer <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_integer <= (std::numeric_limits::max)()) { oa->write_character(to_char_type(0x1A)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } else { oa->write_character(to_char_type(0x1B)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } } else { // The conversions below encode the sign in the first // byte, and the value is converted to a positive number. - const auto positive_number = -1 - j.m_value.number_integer; - if (j.m_value.number_integer >= -24) + const auto positive_number = -1 - j.m_data.m_value.number_integer; + if (j.m_data.m_value.number_integer >= -24) { write_number(static_cast(0x20 + positive_number)); } @@ -15086,52 +15211,52 @@ class binary_writer case value_t::number_unsigned: { - if (j.m_value.number_unsigned <= 0x17) + if (j.m_data.m_value.number_unsigned <= 0x17) { - write_number(static_cast(j.m_value.number_unsigned)); + write_number(static_cast(j.m_data.m_value.number_unsigned)); } - else if (j.m_value.number_unsigned <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits::max)()) { oa->write_character(to_char_type(0x18)); - write_number(static_cast(j.m_value.number_unsigned)); + write_number(static_cast(j.m_data.m_value.number_unsigned)); } - else if (j.m_value.number_unsigned <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits::max)()) { oa->write_character(to_char_type(0x19)); - write_number(static_cast(j.m_value.number_unsigned)); + write_number(static_cast(j.m_data.m_value.number_unsigned)); } - else if (j.m_value.number_unsigned <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits::max)()) { oa->write_character(to_char_type(0x1A)); - write_number(static_cast(j.m_value.number_unsigned)); + write_number(static_cast(j.m_data.m_value.number_unsigned)); } else { oa->write_character(to_char_type(0x1B)); - write_number(static_cast(j.m_value.number_unsigned)); + write_number(static_cast(j.m_data.m_value.number_unsigned)); } break; } case value_t::number_float: { - if (std::isnan(j.m_value.number_float)) + if (std::isnan(j.m_data.m_value.number_float)) { // NaN is 0xf97e00 in CBOR oa->write_character(to_char_type(0xF9)); oa->write_character(to_char_type(0x7E)); oa->write_character(to_char_type(0x00)); } - else if (std::isinf(j.m_value.number_float)) + else if (std::isinf(j.m_data.m_value.number_float)) { // Infinity is 0xf97c00, -Infinity is 0xf9fc00 oa->write_character(to_char_type(0xf9)); - oa->write_character(j.m_value.number_float > 0 ? to_char_type(0x7C) : to_char_type(0xFC)); + oa->write_character(j.m_data.m_value.number_float > 0 ? to_char_type(0x7C) : to_char_type(0xFC)); oa->write_character(to_char_type(0x00)); } else { - write_compact_float(j.m_value.number_float, detail::input_format_t::cbor); + write_compact_float(j.m_data.m_value.number_float, detail::input_format_t::cbor); } break; } @@ -15139,7 +15264,7 @@ class binary_writer case value_t::string: { // step 1: write control byte and the string length - const auto N = j.m_value.string->size(); + const auto N = j.m_data.m_value.string->size(); if (N <= 0x17) { write_number(static_cast(0x60 + N)); @@ -15169,15 +15294,15 @@ class binary_writer // step 2: write the string oa->write_characters( - reinterpret_cast(j.m_value.string->c_str()), - j.m_value.string->size()); + reinterpret_cast(j.m_data.m_value.string->c_str()), + j.m_data.m_value.string->size()); break; } case value_t::array: { // step 1: write control byte and the array size - const auto N = j.m_value.array->size(); + const auto N = j.m_data.m_value.array->size(); if (N <= 0x17) { write_number(static_cast(0x80 + N)); @@ -15206,7 +15331,7 @@ class binary_writer // LCOV_EXCL_STOP // step 2: write each element - for (const auto& el : *j.m_value.array) + for (const auto& el : *j.m_data.m_value.array) { write_cbor(el); } @@ -15215,32 +15340,32 @@ class binary_writer case value_t::binary: { - if (j.m_value.binary->has_subtype()) + if (j.m_data.m_value.binary->has_subtype()) { - if (j.m_value.binary->subtype() <= (std::numeric_limits::max)()) + if (j.m_data.m_value.binary->subtype() <= (std::numeric_limits::max)()) { write_number(static_cast(0xd8)); - write_number(static_cast(j.m_value.binary->subtype())); + write_number(static_cast(j.m_data.m_value.binary->subtype())); } - else if (j.m_value.binary->subtype() <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.binary->subtype() <= (std::numeric_limits::max)()) { write_number(static_cast(0xd9)); - write_number(static_cast(j.m_value.binary->subtype())); + write_number(static_cast(j.m_data.m_value.binary->subtype())); } - else if (j.m_value.binary->subtype() <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.binary->subtype() <= (std::numeric_limits::max)()) { write_number(static_cast(0xda)); - write_number(static_cast(j.m_value.binary->subtype())); + write_number(static_cast(j.m_data.m_value.binary->subtype())); } - else if (j.m_value.binary->subtype() <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.binary->subtype() <= (std::numeric_limits::max)()) { write_number(static_cast(0xdb)); - write_number(static_cast(j.m_value.binary->subtype())); + write_number(static_cast(j.m_data.m_value.binary->subtype())); } } // step 1: write control byte and the binary array size - const auto N = j.m_value.binary->size(); + const auto N = j.m_data.m_value.binary->size(); if (N <= 0x17) { write_number(static_cast(0x40 + N)); @@ -15270,7 +15395,7 @@ class binary_writer // step 2: write each element oa->write_characters( - reinterpret_cast(j.m_value.binary->data()), + reinterpret_cast(j.m_data.m_value.binary->data()), N); break; @@ -15279,7 +15404,7 @@ class binary_writer case value_t::object: { // step 1: write control byte and the object size - const auto N = j.m_value.object->size(); + const auto N = j.m_data.m_value.object->size(); if (N <= 0x17) { write_number(static_cast(0xA0 + N)); @@ -15308,7 +15433,7 @@ class binary_writer // LCOV_EXCL_STOP // step 2: write each element - for (const auto& el : *j.m_value.object) + for (const auto& el : *j.m_data.m_value.object) { write_cbor(el.first); write_cbor(el.second); @@ -15337,7 +15462,7 @@ class binary_writer case value_t::boolean: // true and false { - oa->write_character(j.m_value.boolean + oa->write_character(j.m_data.m_value.boolean ? to_char_type(0xC3) : to_char_type(0xC2)); break; @@ -15345,75 +15470,75 @@ class binary_writer case value_t::number_integer: { - if (j.m_value.number_integer >= 0) + if (j.m_data.m_value.number_integer >= 0) { // MessagePack does not differentiate between positive // signed integers and unsigned integers. Therefore, we used // the code from the value_t::number_unsigned case here. - if (j.m_value.number_unsigned < 128) + if (j.m_data.m_value.number_unsigned < 128) { // positive fixnum - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_unsigned <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits::max)()) { // uint 8 oa->write_character(to_char_type(0xCC)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_unsigned <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits::max)()) { // uint 16 oa->write_character(to_char_type(0xCD)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_unsigned <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits::max)()) { // uint 32 oa->write_character(to_char_type(0xCE)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_unsigned <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits::max)()) { // uint 64 oa->write_character(to_char_type(0xCF)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } } else { - if (j.m_value.number_integer >= -32) + if (j.m_data.m_value.number_integer >= -32) { // negative fixnum - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_integer >= (std::numeric_limits::min)() && - j.m_value.number_integer <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_integer >= (std::numeric_limits::min)() && + j.m_data.m_value.number_integer <= (std::numeric_limits::max)()) { // int 8 oa->write_character(to_char_type(0xD0)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_integer >= (std::numeric_limits::min)() && - j.m_value.number_integer <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_integer >= (std::numeric_limits::min)() && + j.m_data.m_value.number_integer <= (std::numeric_limits::max)()) { // int 16 oa->write_character(to_char_type(0xD1)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_integer >= (std::numeric_limits::min)() && - j.m_value.number_integer <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_integer >= (std::numeric_limits::min)() && + j.m_data.m_value.number_integer <= (std::numeric_limits::max)()) { // int 32 oa->write_character(to_char_type(0xD2)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_integer >= (std::numeric_limits::min)() && - j.m_value.number_integer <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_integer >= (std::numeric_limits::min)() && + j.m_data.m_value.number_integer <= (std::numeric_limits::max)()) { // int 64 oa->write_character(to_char_type(0xD3)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } } break; @@ -15421,48 +15546,48 @@ class binary_writer case value_t::number_unsigned: { - if (j.m_value.number_unsigned < 128) + if (j.m_data.m_value.number_unsigned < 128) { // positive fixnum - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_unsigned <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits::max)()) { // uint 8 oa->write_character(to_char_type(0xCC)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_unsigned <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits::max)()) { // uint 16 oa->write_character(to_char_type(0xCD)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_unsigned <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits::max)()) { // uint 32 oa->write_character(to_char_type(0xCE)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } - else if (j.m_value.number_unsigned <= (std::numeric_limits::max)()) + else if (j.m_data.m_value.number_unsigned <= (std::numeric_limits::max)()) { // uint 64 oa->write_character(to_char_type(0xCF)); - write_number(static_cast(j.m_value.number_integer)); + write_number(static_cast(j.m_data.m_value.number_integer)); } break; } case value_t::number_float: { - write_compact_float(j.m_value.number_float, detail::input_format_t::msgpack); + write_compact_float(j.m_data.m_value.number_float, detail::input_format_t::msgpack); break; } case value_t::string: { // step 1: write control byte and the string length - const auto N = j.m_value.string->size(); + const auto N = j.m_data.m_value.string->size(); if (N <= 31) { // fixstr @@ -15489,15 +15614,15 @@ class binary_writer // step 2: write the string oa->write_characters( - reinterpret_cast(j.m_value.string->c_str()), - j.m_value.string->size()); + reinterpret_cast(j.m_data.m_value.string->c_str()), + j.m_data.m_value.string->size()); break; } case value_t::array: { // step 1: write control byte and the array size - const auto N = j.m_value.array->size(); + const auto N = j.m_data.m_value.array->size(); if (N <= 15) { // fixarray @@ -15517,7 +15642,7 @@ class binary_writer } // step 2: write each element - for (const auto& el : *j.m_value.array) + for (const auto& el : *j.m_data.m_value.array) { write_msgpack(el); } @@ -15528,10 +15653,10 @@ class binary_writer { // step 0: determine if the binary type has a set subtype to // determine whether or not to use the ext or fixext types - const bool use_ext = j.m_value.binary->has_subtype(); + const bool use_ext = j.m_data.m_value.binary->has_subtype(); // step 1: write control byte and the byte string length - const auto N = j.m_value.binary->size(); + const auto N = j.m_data.m_value.binary->size(); if (N <= (std::numeric_limits::max)()) { std::uint8_t output_type{}; @@ -15576,18 +15701,18 @@ class binary_writer } else if (N <= (std::numeric_limits::max)()) { - std::uint8_t output_type = use_ext - ? 0xC8 // ext 16 - : 0xC5; // bin 16 + const std::uint8_t output_type = use_ext + ? 0xC8 // ext 16 + : 0xC5; // bin 16 oa->write_character(to_char_type(output_type)); write_number(static_cast(N)); } else if (N <= (std::numeric_limits::max)()) { - std::uint8_t output_type = use_ext - ? 0xC9 // ext 32 - : 0xC6; // bin 32 + const std::uint8_t output_type = use_ext + ? 0xC9 // ext 32 + : 0xC6; // bin 32 oa->write_character(to_char_type(output_type)); write_number(static_cast(N)); @@ -15596,12 +15721,12 @@ class binary_writer // step 1.5: if this is an ext type, write the subtype if (use_ext) { - write_number(static_cast(j.m_value.binary->subtype())); + write_number(static_cast(j.m_data.m_value.binary->subtype())); } // step 2: write the byte string oa->write_characters( - reinterpret_cast(j.m_value.binary->data()), + reinterpret_cast(j.m_data.m_value.binary->data()), N); break; @@ -15610,7 +15735,7 @@ class binary_writer case value_t::object: { // step 1: write control byte and the object size - const auto N = j.m_value.object->size(); + const auto N = j.m_data.m_value.object->size(); if (N <= 15) { // fixmap @@ -15630,7 +15755,7 @@ class binary_writer } // step 2: write each element - for (const auto& el : *j.m_value.object) + for (const auto& el : *j.m_data.m_value.object) { write_msgpack(el.first); write_msgpack(el.second); @@ -15670,7 +15795,7 @@ class binary_writer { if (add_prefix) { - oa->write_character(j.m_value.boolean + oa->write_character(j.m_data.m_value.boolean ? to_char_type('T') : to_char_type('F')); } @@ -15679,19 +15804,19 @@ class binary_writer case value_t::number_integer: { - write_number_with_ubjson_prefix(j.m_value.number_integer, add_prefix, use_bjdata); + write_number_with_ubjson_prefix(j.m_data.m_value.number_integer, add_prefix, use_bjdata); break; } case value_t::number_unsigned: { - write_number_with_ubjson_prefix(j.m_value.number_unsigned, add_prefix, use_bjdata); + write_number_with_ubjson_prefix(j.m_data.m_value.number_unsigned, add_prefix, use_bjdata); break; } case value_t::number_float: { - write_number_with_ubjson_prefix(j.m_value.number_float, add_prefix, use_bjdata); + write_number_with_ubjson_prefix(j.m_data.m_value.number_float, add_prefix, use_bjdata); break; } @@ -15701,10 +15826,10 @@ class binary_writer { oa->write_character(to_char_type('S')); } - write_number_with_ubjson_prefix(j.m_value.string->size(), true, use_bjdata); + write_number_with_ubjson_prefix(j.m_data.m_value.string->size(), true, use_bjdata); oa->write_characters( - reinterpret_cast(j.m_value.string->c_str()), - j.m_value.string->size()); + reinterpret_cast(j.m_data.m_value.string->c_str()), + j.m_data.m_value.string->size()); break; } @@ -15716,7 +15841,7 @@ class binary_writer } bool prefix_required = true; - if (use_type && !j.m_value.array->empty()) + if (use_type && !j.m_data.m_value.array->empty()) { JSON_ASSERT(use_count); const CharType first_prefix = ubjson_prefix(j.front(), use_bjdata); @@ -15739,10 +15864,10 @@ class binary_writer if (use_count) { oa->write_character(to_char_type('#')); - write_number_with_ubjson_prefix(j.m_value.array->size(), true, use_bjdata); + write_number_with_ubjson_prefix(j.m_data.m_value.array->size(), true, use_bjdata); } - for (const auto& el : *j.m_value.array) + for (const auto& el : *j.m_data.m_value.array) { write_ubjson(el, use_count, use_type, prefix_required, use_bjdata); } @@ -15762,7 +15887,7 @@ class binary_writer oa->write_character(to_char_type('[')); } - if (use_type && !j.m_value.binary->empty()) + if (use_type && !j.m_data.m_value.binary->empty()) { JSON_ASSERT(use_count); oa->write_character(to_char_type('$')); @@ -15772,21 +15897,21 @@ class binary_writer if (use_count) { oa->write_character(to_char_type('#')); - write_number_with_ubjson_prefix(j.m_value.binary->size(), true, use_bjdata); + write_number_with_ubjson_prefix(j.m_data.m_value.binary->size(), true, use_bjdata); } if (use_type) { oa->write_characters( - reinterpret_cast(j.m_value.binary->data()), - j.m_value.binary->size()); + reinterpret_cast(j.m_data.m_value.binary->data()), + j.m_data.m_value.binary->size()); } else { - for (size_t i = 0; i < j.m_value.binary->size(); ++i) + for (size_t i = 0; i < j.m_data.m_value.binary->size(); ++i) { oa->write_character(to_char_type('U')); - oa->write_character(j.m_value.binary->data()[i]); + oa->write_character(j.m_data.m_value.binary->data()[i]); } } @@ -15800,9 +15925,9 @@ class binary_writer case value_t::object: { - if (use_bjdata && j.m_value.object->size() == 3 && j.m_value.object->find("_ArrayType_") != j.m_value.object->end() && j.m_value.object->find("_ArraySize_") != j.m_value.object->end() && j.m_value.object->find("_ArrayData_") != j.m_value.object->end()) + if (use_bjdata && j.m_data.m_value.object->size() == 3 && j.m_data.m_value.object->find("_ArrayType_") != j.m_data.m_value.object->end() && j.m_data.m_value.object->find("_ArraySize_") != j.m_data.m_value.object->end() && j.m_data.m_value.object->find("_ArrayData_") != j.m_data.m_value.object->end()) { - if (!write_bjdata_ndarray(*j.m_value.object, use_count, use_type)) // decode bjdata ndarray in the JData format (https://github.com/NeuroJSON/jdata) + if (!write_bjdata_ndarray(*j.m_data.m_value.object, use_count, use_type)) // decode bjdata ndarray in the JData format (https://github.com/NeuroJSON/jdata) { break; } @@ -15814,7 +15939,7 @@ class binary_writer } bool prefix_required = true; - if (use_type && !j.m_value.object->empty()) + if (use_type && !j.m_data.m_value.object->empty()) { JSON_ASSERT(use_count); const CharType first_prefix = ubjson_prefix(j.front(), use_bjdata); @@ -15837,10 +15962,10 @@ class binary_writer if (use_count) { oa->write_character(to_char_type('#')); - write_number_with_ubjson_prefix(j.m_value.object->size(), true, use_bjdata); + write_number_with_ubjson_prefix(j.m_data.m_value.object->size(), true, use_bjdata); } - for (const auto& el : *j.m_value.object) + for (const auto& el : *j.m_data.m_value.object) { write_number_with_ubjson_prefix(el.first.size(), true, use_bjdata); oa->write_characters( @@ -15990,19 +16115,19 @@ class binary_writer void write_bson_unsigned(const string_t& name, const BasicJsonType& j) { - if (j.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) + if (j.m_data.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) { write_bson_entry_header(name, 0x10 /* int32 */); - write_number(static_cast(j.m_value.number_unsigned), true); + write_number(static_cast(j.m_data.m_value.number_unsigned), true); } - else if (j.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) + else if (j.m_data.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) { write_bson_entry_header(name, 0x12 /* int64 */); - write_number(static_cast(j.m_value.number_unsigned), true); + write_number(static_cast(j.m_data.m_value.number_unsigned), true); } else { - JSON_THROW(out_of_range::create(407, concat("integer number ", std::to_string(j.m_value.number_unsigned), " cannot be represented by BSON as it does not fit int64"), &j)); + JSON_THROW(out_of_range::create(407, concat("integer number ", std::to_string(j.m_data.m_value.number_unsigned), " cannot be represented by BSON as it does not fit int64"), &j)); } } @@ -16083,13 +16208,13 @@ class binary_writer switch (j.type()) { case value_t::object: - return header_size + calc_bson_object_size(*j.m_value.object); + return header_size + calc_bson_object_size(*j.m_data.m_value.object); case value_t::array: - return header_size + calc_bson_array_size(*j.m_value.array); + return header_size + calc_bson_array_size(*j.m_data.m_value.array); case value_t::binary: - return header_size + calc_bson_binary_size(*j.m_value.binary); + return header_size + calc_bson_binary_size(*j.m_data.m_value.binary); case value_t::boolean: return header_size + 1ul; @@ -16098,13 +16223,13 @@ class binary_writer return header_size + 8ul; case value_t::number_integer: - return header_size + calc_bson_integer_size(j.m_value.number_integer); + return header_size + calc_bson_integer_size(j.m_data.m_value.number_integer); case value_t::number_unsigned: - return header_size + calc_bson_unsigned_size(j.m_value.number_unsigned); + return header_size + calc_bson_unsigned_size(j.m_data.m_value.number_unsigned); case value_t::string: - return header_size + calc_bson_string_size(*j.m_value.string); + return header_size + calc_bson_string_size(*j.m_data.m_value.string); case value_t::null: return header_size + 0ul; @@ -16130,28 +16255,28 @@ class binary_writer switch (j.type()) { case value_t::object: - return write_bson_object_entry(name, *j.m_value.object); + return write_bson_object_entry(name, *j.m_data.m_value.object); case value_t::array: - return write_bson_array(name, *j.m_value.array); + return write_bson_array(name, *j.m_data.m_value.array); case value_t::binary: - return write_bson_binary(name, *j.m_value.binary); + return write_bson_binary(name, *j.m_data.m_value.binary); case value_t::boolean: - return write_bson_boolean(name, j.m_value.boolean); + return write_bson_boolean(name, j.m_data.m_value.boolean); case value_t::number_float: - return write_bson_double(name, j.m_value.number_float); + return write_bson_double(name, j.m_data.m_value.number_float); case value_t::number_integer: - return write_bson_integer(name, j.m_value.number_integer); + return write_bson_integer(name, j.m_data.m_value.number_integer); case value_t::number_unsigned: return write_bson_unsigned(name, j); case value_t::string: - return write_bson_string(name, *j.m_value.string); + return write_bson_string(name, *j.m_data.m_value.string); case value_t::null: return write_bson_null(name); @@ -16173,8 +16298,8 @@ class binary_writer */ static std::size_t calc_bson_object_size(const typename BasicJsonType::object_t& value) { - std::size_t document_size = std::accumulate(value.begin(), value.end(), static_cast(0), - [](size_t result, const typename BasicJsonType::object_t::value_type & el) + const std::size_t document_size = std::accumulate(value.begin(), value.end(), static_cast(0), + [](size_t result, const typename BasicJsonType::object_t::value_type & el) { return result += calc_bson_element_size(el.first, el.second); }); @@ -16424,35 +16549,35 @@ class binary_writer return 'Z'; case value_t::boolean: - return j.m_value.boolean ? 'T' : 'F'; + return j.m_data.m_value.boolean ? 'T' : 'F'; case value_t::number_integer: { - if ((std::numeric_limits::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits::max)()) + if ((std::numeric_limits::min)() <= j.m_data.m_value.number_integer && j.m_data.m_value.number_integer <= (std::numeric_limits::max)()) { return 'i'; } - if ((std::numeric_limits::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits::max)()) + if ((std::numeric_limits::min)() <= j.m_data.m_value.number_integer && j.m_data.m_value.number_integer <= (std::numeric_limits::max)()) { return 'U'; } - if ((std::numeric_limits::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits::max)()) + if ((std::numeric_limits::min)() <= j.m_data.m_value.number_integer && j.m_data.m_value.number_integer <= (std::numeric_limits::max)()) { return 'I'; } - if (use_bjdata && ((std::numeric_limits::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits::max)())) + if (use_bjdata && ((std::numeric_limits::min)() <= j.m_data.m_value.number_integer && j.m_data.m_value.number_integer <= (std::numeric_limits::max)())) { return 'u'; } - if ((std::numeric_limits::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits::max)()) + if ((std::numeric_limits::min)() <= j.m_data.m_value.number_integer && j.m_data.m_value.number_integer <= (std::numeric_limits::max)()) { return 'l'; } - if (use_bjdata && ((std::numeric_limits::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits::max)())) + if (use_bjdata && ((std::numeric_limits::min)() <= j.m_data.m_value.number_integer && j.m_data.m_value.number_integer <= (std::numeric_limits::max)())) { return 'm'; } - if ((std::numeric_limits::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits::max)()) + if ((std::numeric_limits::min)() <= j.m_data.m_value.number_integer && j.m_data.m_value.number_integer <= (std::numeric_limits::max)()) { return 'L'; } @@ -16462,35 +16587,35 @@ class binary_writer case value_t::number_unsigned: { - if (j.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) + if (j.m_data.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) { return 'i'; } - if (j.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) + if (j.m_data.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) { return 'U'; } - if (j.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) + if (j.m_data.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) { return 'I'; } - if (use_bjdata && j.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) + if (use_bjdata && j.m_data.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) { return 'u'; } - if (j.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) + if (j.m_data.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) { return 'l'; } - if (use_bjdata && j.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) + if (use_bjdata && j.m_data.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) { return 'm'; } - if (j.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) + if (j.m_data.m_value.number_unsigned <= static_cast((std::numeric_limits::max)())) { return 'L'; } - if (use_bjdata && j.m_value.number_unsigned <= (std::numeric_limits::max)()) + if (use_bjdata && j.m_data.m_value.number_unsigned <= (std::numeric_limits::max)()) { return 'M'; } @@ -16499,7 +16624,7 @@ class binary_writer } case value_t::number_float: - return get_ubjson_float_prefix(j.m_value.number_float); + return get_ubjson_float_prefix(j.m_data.m_value.number_float); case value_t::string: return 'S'; @@ -16548,7 +16673,7 @@ class binary_writer std::size_t len = (value.at(key).empty() ? 0 : 1); for (const auto& el : value.at(key)) { - len *= static_cast(el.m_value.number_unsigned); + len *= static_cast(el.m_data.m_value.number_unsigned); } key = "_ArrayData_"; @@ -16570,70 +16695,70 @@ class binary_writer { for (const auto& el : value.at(key)) { - write_number(static_cast(el.m_value.number_unsigned), true); + write_number(static_cast(el.m_data.m_value.number_unsigned), true); } } else if (dtype == 'i') { for (const auto& el : value.at(key)) { - write_number(static_cast(el.m_value.number_integer), true); + write_number(static_cast(el.m_data.m_value.number_integer), true); } } else if (dtype == 'u') { for (const auto& el : value.at(key)) { - write_number(static_cast(el.m_value.number_unsigned), true); + write_number(static_cast(el.m_data.m_value.number_unsigned), true); } } else if (dtype == 'I') { for (const auto& el : value.at(key)) { - write_number(static_cast(el.m_value.number_integer), true); + write_number(static_cast(el.m_data.m_value.number_integer), true); } } else if (dtype == 'm') { for (const auto& el : value.at(key)) { - write_number(static_cast(el.m_value.number_unsigned), true); + write_number(static_cast(el.m_data.m_value.number_unsigned), true); } } else if (dtype == 'l') { for (const auto& el : value.at(key)) { - write_number(static_cast(el.m_value.number_integer), true); + write_number(static_cast(el.m_data.m_value.number_integer), true); } } else if (dtype == 'M') { for (const auto& el : value.at(key)) { - write_number(static_cast(el.m_value.number_unsigned), true); + write_number(static_cast(el.m_data.m_value.number_unsigned), true); } } else if (dtype == 'L') { for (const auto& el : value.at(key)) { - write_number(static_cast(el.m_value.number_integer), true); + write_number(static_cast(el.m_data.m_value.number_integer), true); } } else if (dtype == 'd') { for (const auto& el : value.at(key)) { - write_number(static_cast(el.m_value.number_float), true); + write_number(static_cast(el.m_data.m_value.number_float), true); } } else if (dtype == 'D') { for (const auto& el : value.at(key)) { - write_number(static_cast(el.m_value.number_float), true); + write_number(static_cast(el.m_data.m_value.number_float), true); } } return false; @@ -16757,11 +16882,11 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // // SPDX-FileCopyrightText: 2008-2009 Björn Hoehrmann -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -16782,11 +16907,11 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // // SPDX-FileCopyrightText: 2009 Florian Loitsch -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -17692,7 +17817,7 @@ void grisu2(char* buf, int& len, int& decimal_exponent, FloatType value) // NB: If the neighbors are computed for single-precision numbers, there is a single float // (7.0385307e-26f) which can't be recovered using strtod. The resulting double precision // value is off by 1 ulp. -#if 0 +#if 0 // NOLINT(readability-avoid-unconditional-preprocessor-if) const boundaries w = compute_boundaries(static_cast(value)); #else const boundaries w = compute_boundaries(value); @@ -17994,11 +18119,11 @@ class serializer const unsigned int indent_step, const unsigned int current_indent = 0) { - switch (val.m_type) + switch (val.m_data.m_type) { case value_t::object: { - if (val.m_value.object->empty()) + if (val.m_data.m_value.object->empty()) { o->write_characters("{}", 2); return; @@ -18016,8 +18141,8 @@ class serializer } // first n-1 elements - auto i = val.m_value.object->cbegin(); - for (std::size_t cnt = 0; cnt < val.m_value.object->size() - 1; ++cnt, ++i) + auto i = val.m_data.m_value.object->cbegin(); + for (std::size_t cnt = 0; cnt < val.m_data.m_value.object->size() - 1; ++cnt, ++i) { o->write_characters(indent_string.c_str(), new_indent); o->write_character('\"'); @@ -18028,8 +18153,8 @@ class serializer } // last element - JSON_ASSERT(i != val.m_value.object->cend()); - JSON_ASSERT(std::next(i) == val.m_value.object->cend()); + JSON_ASSERT(i != val.m_data.m_value.object->cend()); + JSON_ASSERT(std::next(i) == val.m_data.m_value.object->cend()); o->write_characters(indent_string.c_str(), new_indent); o->write_character('\"'); dump_escaped(i->first, ensure_ascii); @@ -18045,8 +18170,8 @@ class serializer o->write_character('{'); // first n-1 elements - auto i = val.m_value.object->cbegin(); - for (std::size_t cnt = 0; cnt < val.m_value.object->size() - 1; ++cnt, ++i) + auto i = val.m_data.m_value.object->cbegin(); + for (std::size_t cnt = 0; cnt < val.m_data.m_value.object->size() - 1; ++cnt, ++i) { o->write_character('\"'); dump_escaped(i->first, ensure_ascii); @@ -18056,8 +18181,8 @@ class serializer } // last element - JSON_ASSERT(i != val.m_value.object->cend()); - JSON_ASSERT(std::next(i) == val.m_value.object->cend()); + JSON_ASSERT(i != val.m_data.m_value.object->cend()); + JSON_ASSERT(std::next(i) == val.m_data.m_value.object->cend()); o->write_character('\"'); dump_escaped(i->first, ensure_ascii); o->write_characters("\":", 2); @@ -18071,7 +18196,7 @@ class serializer case value_t::array: { - if (val.m_value.array->empty()) + if (val.m_data.m_value.array->empty()) { o->write_characters("[]", 2); return; @@ -18089,8 +18214,8 @@ class serializer } // first n-1 elements - for (auto i = val.m_value.array->cbegin(); - i != val.m_value.array->cend() - 1; ++i) + for (auto i = val.m_data.m_value.array->cbegin(); + i != val.m_data.m_value.array->cend() - 1; ++i) { o->write_characters(indent_string.c_str(), new_indent); dump(*i, true, ensure_ascii, indent_step, new_indent); @@ -18098,9 +18223,9 @@ class serializer } // last element - JSON_ASSERT(!val.m_value.array->empty()); + JSON_ASSERT(!val.m_data.m_value.array->empty()); o->write_characters(indent_string.c_str(), new_indent); - dump(val.m_value.array->back(), true, ensure_ascii, indent_step, new_indent); + dump(val.m_data.m_value.array->back(), true, ensure_ascii, indent_step, new_indent); o->write_character('\n'); o->write_characters(indent_string.c_str(), current_indent); @@ -18111,16 +18236,16 @@ class serializer o->write_character('['); // first n-1 elements - for (auto i = val.m_value.array->cbegin(); - i != val.m_value.array->cend() - 1; ++i) + for (auto i = val.m_data.m_value.array->cbegin(); + i != val.m_data.m_value.array->cend() - 1; ++i) { dump(*i, false, ensure_ascii, indent_step, current_indent); o->write_character(','); } // last element - JSON_ASSERT(!val.m_value.array->empty()); - dump(val.m_value.array->back(), false, ensure_ascii, indent_step, current_indent); + JSON_ASSERT(!val.m_data.m_value.array->empty()); + dump(val.m_data.m_value.array->back(), false, ensure_ascii, indent_step, current_indent); o->write_character(']'); } @@ -18131,7 +18256,7 @@ class serializer case value_t::string: { o->write_character('\"'); - dump_escaped(*val.m_value.string, ensure_ascii); + dump_escaped(*val.m_data.m_value.string, ensure_ascii); o->write_character('\"'); return; } @@ -18153,24 +18278,24 @@ class serializer o->write_characters("\"bytes\": [", 10); - if (!val.m_value.binary->empty()) + if (!val.m_data.m_value.binary->empty()) { - for (auto i = val.m_value.binary->cbegin(); - i != val.m_value.binary->cend() - 1; ++i) + for (auto i = val.m_data.m_value.binary->cbegin(); + i != val.m_data.m_value.binary->cend() - 1; ++i) { dump_integer(*i); o->write_characters(", ", 2); } - dump_integer(val.m_value.binary->back()); + dump_integer(val.m_data.m_value.binary->back()); } o->write_characters("],\n", 3); o->write_characters(indent_string.c_str(), new_indent); o->write_characters("\"subtype\": ", 11); - if (val.m_value.binary->has_subtype()) + if (val.m_data.m_value.binary->has_subtype()) { - dump_integer(val.m_value.binary->subtype()); + dump_integer(val.m_data.m_value.binary->subtype()); } else { @@ -18184,21 +18309,21 @@ class serializer { o->write_characters("{\"bytes\":[", 10); - if (!val.m_value.binary->empty()) + if (!val.m_data.m_value.binary->empty()) { - for (auto i = val.m_value.binary->cbegin(); - i != val.m_value.binary->cend() - 1; ++i) + for (auto i = val.m_data.m_value.binary->cbegin(); + i != val.m_data.m_value.binary->cend() - 1; ++i) { dump_integer(*i); o->write_character(','); } - dump_integer(val.m_value.binary->back()); + dump_integer(val.m_data.m_value.binary->back()); } o->write_characters("],\"subtype\":", 12); - if (val.m_value.binary->has_subtype()) + if (val.m_data.m_value.binary->has_subtype()) { - dump_integer(val.m_value.binary->subtype()); + dump_integer(val.m_data.m_value.binary->subtype()); o->write_character('}'); } else @@ -18211,7 +18336,7 @@ class serializer case value_t::boolean: { - if (val.m_value.boolean) + if (val.m_data.m_value.boolean) { o->write_characters("true", 4); } @@ -18224,19 +18349,19 @@ class serializer case value_t::number_integer: { - dump_integer(val.m_value.number_integer); + dump_integer(val.m_data.m_value.number_integer); return; } case value_t::number_unsigned: { - dump_integer(val.m_value.number_unsigned); + dump_integer(val.m_data.m_value.number_unsigned); return; } case value_t::number_float: { - dump_float(val.m_value.number_float); + dump_float(val.m_data.m_value.number_float); return; } @@ -18810,8 +18935,8 @@ class serializer ? (byte & 0x3fu) | (codep << 6u) : (0xFFu >> type) & (byte); - std::size_t index = 256u + static_cast(state) * 16u + static_cast(type); - JSON_ASSERT(index < 400); + const std::size_t index = 256u + static_cast(state) * 16u + static_cast(type); + JSON_ASSERT(index < utf8d.size()); state = utf8d[index]; return state; } @@ -18878,10 +19003,10 @@ NLOHMANN_JSON_NAMESPACE_END // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -18998,7 +19123,7 @@ template , template::value, int> = 0> - T & at(KeyType && key) + T & at(KeyType && key) // NOLINT(cppcoreguidelines-missing-std-forward) { for (auto it = this->begin(); it != this->end(); ++it) { @@ -19026,7 +19151,7 @@ template , template::value, int> = 0> - const T & at(KeyType && key) const + const T & at(KeyType && key) const // NOLINT(cppcoreguidelines-missing-std-forward) { for (auto it = this->begin(); it != this->end(); ++it) { @@ -19060,7 +19185,7 @@ template , template::value, int> = 0> - size_type erase(KeyType && key) + size_type erase(KeyType && key) // NOLINT(cppcoreguidelines-missing-std-forward) { for (auto it = this->begin(); it != this->end(); ++it) { @@ -19151,7 +19276,7 @@ template , template::value, int> = 0> - size_type count(KeyType && key) const + size_type count(KeyType && key) const // NOLINT(cppcoreguidelines-missing-std-forward) { for (auto it = this->begin(); it != this->end(); ++it) { @@ -19177,7 +19302,7 @@ template , template::value, int> = 0> - iterator find(KeyType && key) + iterator find(KeyType && key) // NOLINT(cppcoreguidelines-missing-std-forward) { for (auto it = this->begin(); it != this->end(); ++it) { @@ -19240,7 +19365,9 @@ NLOHMANN_JSON_NAMESPACE_END #if defined(JSON_HAS_CPP_17) - #include + #if JSON_HAS_STATIC_RTTI + #include + #endif #include #endif @@ -19271,6 +19398,7 @@ The invariants are checked by member function assert_invariant(). */ NLOHMANN_BASIC_JSON_TPL_DECLARATION class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-special-member-functions) + : public ::nlohmann::detail::json_base_class { private: template friend struct detail::external_constructor; @@ -19297,6 +19425,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// workaround type for MSVC using basic_json_t = NLOHMANN_BASIC_JSON_TPL; + using json_base_class_t = ::nlohmann::detail::json_base_class; JSON_PRIVATE_UNLESS_TESTED: // convenience aliases for types residing in namespace detail; @@ -19368,7 +19497,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @} - ///////////////////// // container types // ///////////////////// @@ -19410,7 +19538,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @} - /// @brief returns the allocator associated with the container /// @sa https://json.nlohmann.me/api/basic_json/get_allocator/ static allocator_type get_allocator() @@ -19425,7 +19552,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec { basic_json result; - result["copyright"] = "(C) 2013-2022 Niels Lohmann"; + result["copyright"] = "(C) 2013-2023 Niels Lohmann"; result["name"] = "JSON for Modern C++"; result["url"] = "https://github.com/nlohmann/json"; result["version"]["string"] = @@ -19473,7 +19600,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec result["compiler"] = {{"family", "unknown"}, {"version", "unknown"}}; #endif - #if defined(_MSVC_LANG) result["compiler"]["c++"] = std::to_string(_MSVC_LANG); #elif defined(__cplusplus) @@ -19484,7 +19610,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec return result; } - /////////////////////////// // JSON value data types // /////////////////////////// @@ -19692,7 +19817,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec object = nullptr; // silence warning, see #821 if (JSON_HEDLEY_UNLIKELY(t == value_t::null)) { - JSON_THROW(other_error::create(500, "961c151d2e87f2686a955a9be24d316f1362bf21 3.11.2", nullptr)); // LCOV_EXCL_LINE + JSON_THROW(other_error::create(500, "961c151d2e87f2686a955a9be24d316f1362bf21 3.11.3", nullptr)); // LCOV_EXCL_LINE } break; } @@ -19731,6 +19856,16 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec void destroy(value_t t) { + if ( + (t == value_t::object && object == nullptr) || + (t == value_t::array && array == nullptr) || + (t == value_t::string && string == nullptr) || + (t == value_t::binary && binary == nullptr) + ) + { + //not initialized (e.g. due to exception in the ctor) + return; + } if (t == value_t::array || t == value_t::object) { // flatten the current json_value to a heap-allocated stack @@ -19761,18 +19896,18 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // its children to the stack to be processed later if (current_item.is_array()) { - std::move(current_item.m_value.array->begin(), current_item.m_value.array->end(), std::back_inserter(stack)); + std::move(current_item.m_data.m_value.array->begin(), current_item.m_data.m_value.array->end(), std::back_inserter(stack)); - current_item.m_value.array->clear(); + current_item.m_data.m_value.array->clear(); } else if (current_item.is_object()) { - for (auto&& it : *current_item.m_value.object) + for (auto&& it : *current_item.m_data.m_value.object) { stack.push_back(std::move(it.second)); } - current_item.m_value.object->clear(); + current_item.m_data.m_value.object->clear(); } // it's now safe that current_item get destructed @@ -19849,10 +19984,10 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec */ void assert_invariant(bool check_parents = true) const noexcept { - JSON_ASSERT(m_type != value_t::object || m_value.object != nullptr); - JSON_ASSERT(m_type != value_t::array || m_value.array != nullptr); - JSON_ASSERT(m_type != value_t::string || m_value.string != nullptr); - JSON_ASSERT(m_type != value_t::binary || m_value.binary != nullptr); + JSON_ASSERT(m_data.m_type != value_t::object || m_data.m_value.object != nullptr); + JSON_ASSERT(m_data.m_type != value_t::array || m_data.m_value.array != nullptr); + JSON_ASSERT(m_data.m_type != value_t::string || m_data.m_value.string != nullptr); + JSON_ASSERT(m_data.m_type != value_t::binary || m_data.m_value.binary != nullptr); #if JSON_DIAGNOSTICS JSON_TRY @@ -19871,11 +20006,11 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec void set_parents() { #if JSON_DIAGNOSTICS - switch (m_type) + switch (m_data.m_type) { case value_t::array: { - for (auto& element : *m_value.array) + for (auto& element : *m_data.m_value.array) { element.m_parent = this; } @@ -19884,7 +20019,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec case value_t::object: { - for (auto& element : *m_value.object) + for (auto& element : *m_data.m_value.object) { element.second.m_parent = this; } @@ -19925,7 +20060,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec { // see https://github.com/nlohmann/json/issues/2838 JSON_ASSERT(type() == value_t::array); - if (JSON_HEDLEY_UNLIKELY(m_value.array->capacity() != old_capacity)) + if (JSON_HEDLEY_UNLIKELY(m_data.m_value.array->capacity() != old_capacity)) { // capacity has changed: update all parents set_parents(); @@ -19981,7 +20116,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @brief create an empty value with a given type /// @sa https://json.nlohmann.me/api/basic_json/basic_json/ basic_json(const value_t v) - : m_type(v), m_value(v) + : m_data(v) { assert_invariant(); } @@ -20055,12 +20190,12 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec *this = nullptr; break; case value_t::discarded: - m_type = value_t::discarded; + m_data.m_type = value_t::discarded; break; default: // LCOV_EXCL_LINE JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE } - JSON_ASSERT(m_type == val.type()); + JSON_ASSERT(m_data.m_type == val.type()); set_parents(); assert_invariant(); } @@ -20076,7 +20211,10 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec bool is_an_object = std::all_of(init.begin(), init.end(), [](const detail::json_ref& element_ref) { - return element_ref->is_array() && element_ref->size() == 2 && (*element_ref)[0].is_string(); + // The cast is to ensure op[size_type] is called, bearing in mind size_type may not be int; + // (many string types can be constructed from 0 via its null-pointer guise, so we get a + // broken call to op[key_type], the wrong semantics and a 4804 warning on Windows) + return element_ref->is_array() && element_ref->size() == 2 && (*element_ref)[static_cast(0)].is_string(); }); // adjust type if type deduction is not wanted @@ -20098,22 +20236,22 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec if (is_an_object) { // the initializer list is a list of pairs -> create object - m_type = value_t::object; - m_value = value_t::object; + m_data.m_type = value_t::object; + m_data.m_value = value_t::object; for (auto& element_ref : init) { auto element = element_ref.moved_or_copied(); - m_value.object->emplace( - std::move(*((*element.m_value.array)[0].m_value.string)), - std::move((*element.m_value.array)[1])); + m_data.m_value.object->emplace( + std::move(*((*element.m_data.m_value.array)[0].m_data.m_value.string)), + std::move((*element.m_data.m_value.array)[1])); } } else { // the initializer list describes an array -> create array - m_type = value_t::array; - m_value.array = create(init.begin(), init.end()); + m_data.m_type = value_t::array; + m_data.m_value.array = create(init.begin(), init.end()); } set_parents(); @@ -20126,8 +20264,8 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec static basic_json binary(const typename binary_t::container_type& init) { auto res = basic_json(); - res.m_type = value_t::binary; - res.m_value = init; + res.m_data.m_type = value_t::binary; + res.m_data.m_value = init; return res; } @@ -20137,8 +20275,8 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec static basic_json binary(const typename binary_t::container_type& init, typename binary_t::subtype_type subtype) { auto res = basic_json(); - res.m_type = value_t::binary; - res.m_value = binary_t(init, subtype); + res.m_data.m_type = value_t::binary; + res.m_data.m_value = binary_t(init, subtype); return res; } @@ -20148,8 +20286,8 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec static basic_json binary(typename binary_t::container_type&& init) { auto res = basic_json(); - res.m_type = value_t::binary; - res.m_value = std::move(init); + res.m_data.m_type = value_t::binary; + res.m_data.m_value = std::move(init); return res; } @@ -20159,8 +20297,8 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec static basic_json binary(typename binary_t::container_type&& init, typename binary_t::subtype_type subtype) { auto res = basic_json(); - res.m_type = value_t::binary; - res.m_value = binary_t(std::move(init), subtype); + res.m_data.m_type = value_t::binary; + res.m_data.m_value = binary_t(std::move(init), subtype); return res; } @@ -20182,10 +20320,9 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @brief construct an array with count copies of given value /// @sa https://json.nlohmann.me/api/basic_json/basic_json/ - basic_json(size_type cnt, const basic_json& val) - : m_type(value_t::array) + basic_json(size_type cnt, const basic_json& val): + m_data{cnt, val} { - m_value.array = create(cnt, val); set_parents(); assert_invariant(); } @@ -20207,10 +20344,10 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec } // copy type from first iterator - m_type = first.m_object->m_type; + m_data.m_type = first.m_object->m_data.m_type; // check if iterator range is complete for primitive values - switch (m_type) + switch (m_data.m_type) { case value_t::boolean: case value_t::number_float: @@ -20235,55 +20372,55 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec break; } - switch (m_type) + switch (m_data.m_type) { case value_t::number_integer: { - m_value.number_integer = first.m_object->m_value.number_integer; + m_data.m_value.number_integer = first.m_object->m_data.m_value.number_integer; break; } case value_t::number_unsigned: { - m_value.number_unsigned = first.m_object->m_value.number_unsigned; + m_data.m_value.number_unsigned = first.m_object->m_data.m_value.number_unsigned; break; } case value_t::number_float: { - m_value.number_float = first.m_object->m_value.number_float; + m_data.m_value.number_float = first.m_object->m_data.m_value.number_float; break; } case value_t::boolean: { - m_value.boolean = first.m_object->m_value.boolean; + m_data.m_value.boolean = first.m_object->m_data.m_value.boolean; break; } case value_t::string: { - m_value = *first.m_object->m_value.string; + m_data.m_value = *first.m_object->m_data.m_value.string; break; } case value_t::object: { - m_value.object = create(first.m_it.object_iterator, - last.m_it.object_iterator); + m_data.m_value.object = create(first.m_it.object_iterator, + last.m_it.object_iterator); break; } case value_t::array: { - m_value.array = create(first.m_it.array_iterator, - last.m_it.array_iterator); + m_data.m_value.array = create(first.m_it.array_iterator, + last.m_it.array_iterator); break; } case value_t::binary: { - m_value = *first.m_object->m_value.binary; + m_data.m_value = *first.m_object->m_data.m_value.binary; break; } @@ -20297,7 +20434,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec assert_invariant(); } - /////////////////////////////////////// // other constructors and destructor // /////////////////////////////////////// @@ -20310,58 +20446,59 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @brief copy constructor /// @sa https://json.nlohmann.me/api/basic_json/basic_json/ basic_json(const basic_json& other) - : m_type(other.m_type) + : json_base_class_t(other) { + m_data.m_type = other.m_data.m_type; // check of passed value is valid other.assert_invariant(); - switch (m_type) + switch (m_data.m_type) { case value_t::object: { - m_value = *other.m_value.object; + m_data.m_value = *other.m_data.m_value.object; break; } case value_t::array: { - m_value = *other.m_value.array; + m_data.m_value = *other.m_data.m_value.array; break; } case value_t::string: { - m_value = *other.m_value.string; + m_data.m_value = *other.m_data.m_value.string; break; } case value_t::boolean: { - m_value = other.m_value.boolean; + m_data.m_value = other.m_data.m_value.boolean; break; } case value_t::number_integer: { - m_value = other.m_value.number_integer; + m_data.m_value = other.m_data.m_value.number_integer; break; } case value_t::number_unsigned: { - m_value = other.m_value.number_unsigned; + m_data.m_value = other.m_data.m_value.number_unsigned; break; } case value_t::number_float: { - m_value = other.m_value.number_float; + m_data.m_value = other.m_data.m_value.number_float; break; } case value_t::binary: { - m_value = *other.m_value.binary; + m_data.m_value = *other.m_data.m_value.binary; break; } @@ -20378,15 +20515,15 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @brief move constructor /// @sa https://json.nlohmann.me/api/basic_json/basic_json/ basic_json(basic_json&& other) noexcept - : m_type(std::move(other.m_type)), - m_value(std::move(other.m_value)) + : json_base_class_t(std::forward(other)), + m_data(std::move(other.m_data)) { // check that passed value is valid other.assert_invariant(false); // invalidate payload - other.m_type = value_t::null; - other.m_value = {}; + other.m_data.m_type = value_t::null; + other.m_data.m_value = {}; set_parents(); assert_invariant(); @@ -20398,15 +20535,17 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec std::is_nothrow_move_constructible::value&& std::is_nothrow_move_assignable::value&& std::is_nothrow_move_constructible::value&& - std::is_nothrow_move_assignable::value + std::is_nothrow_move_assignable::value&& + std::is_nothrow_move_assignable::value ) { // check that passed value is valid other.assert_invariant(); using std::swap; - swap(m_type, other.m_type); - swap(m_value, other.m_value); + swap(m_data.m_type, other.m_data.m_type); + swap(m_data.m_value, other.m_data.m_value); + json_base_class_t::operator=(std::move(other)); set_parents(); assert_invariant(); @@ -20418,7 +20557,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec ~basic_json() noexcept { assert_invariant(false); - m_value.destroy(m_type); } /// @} @@ -20458,7 +20596,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @sa https://json.nlohmann.me/api/basic_json/type/ constexpr value_t type() const noexcept { - return m_type; + return m_data.m_type; } /// @brief return whether type is primitive @@ -20479,14 +20617,14 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @sa https://json.nlohmann.me/api/basic_json/is_null/ constexpr bool is_null() const noexcept { - return m_type == value_t::null; + return m_data.m_type == value_t::null; } /// @brief return whether value is a boolean /// @sa https://json.nlohmann.me/api/basic_json/is_boolean/ constexpr bool is_boolean() const noexcept { - return m_type == value_t::boolean; + return m_data.m_type == value_t::boolean; } /// @brief return whether value is a number @@ -20500,63 +20638,63 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @sa https://json.nlohmann.me/api/basic_json/is_number_integer/ constexpr bool is_number_integer() const noexcept { - return m_type == value_t::number_integer || m_type == value_t::number_unsigned; + return m_data.m_type == value_t::number_integer || m_data.m_type == value_t::number_unsigned; } /// @brief return whether value is an unsigned integer number /// @sa https://json.nlohmann.me/api/basic_json/is_number_unsigned/ constexpr bool is_number_unsigned() const noexcept { - return m_type == value_t::number_unsigned; + return m_data.m_type == value_t::number_unsigned; } /// @brief return whether value is a floating-point number /// @sa https://json.nlohmann.me/api/basic_json/is_number_float/ constexpr bool is_number_float() const noexcept { - return m_type == value_t::number_float; + return m_data.m_type == value_t::number_float; } /// @brief return whether value is an object /// @sa https://json.nlohmann.me/api/basic_json/is_object/ constexpr bool is_object() const noexcept { - return m_type == value_t::object; + return m_data.m_type == value_t::object; } /// @brief return whether value is an array /// @sa https://json.nlohmann.me/api/basic_json/is_array/ constexpr bool is_array() const noexcept { - return m_type == value_t::array; + return m_data.m_type == value_t::array; } /// @brief return whether value is a string /// @sa https://json.nlohmann.me/api/basic_json/is_string/ constexpr bool is_string() const noexcept { - return m_type == value_t::string; + return m_data.m_type == value_t::string; } /// @brief return whether value is a binary array /// @sa https://json.nlohmann.me/api/basic_json/is_binary/ constexpr bool is_binary() const noexcept { - return m_type == value_t::binary; + return m_data.m_type == value_t::binary; } /// @brief return whether value is discarded /// @sa https://json.nlohmann.me/api/basic_json/is_discarded/ constexpr bool is_discarded() const noexcept { - return m_type == value_t::discarded; + return m_data.m_type == value_t::discarded; } /// @brief return the type of the JSON value (implicit) /// @sa https://json.nlohmann.me/api/basic_json/operator_value_t/ constexpr operator value_t() const noexcept { - return m_type; + return m_data.m_type; } /// @} @@ -20571,7 +20709,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec { if (JSON_HEDLEY_LIKELY(is_boolean())) { - return m_value.boolean; + return m_data.m_value.boolean; } JSON_THROW(type_error::create(302, detail::concat("type must be boolean, but is ", type_name()), this)); @@ -20580,97 +20718,97 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// get a pointer to the value (object) object_t* get_impl_ptr(object_t* /*unused*/) noexcept { - return is_object() ? m_value.object : nullptr; + return is_object() ? m_data.m_value.object : nullptr; } /// get a pointer to the value (object) constexpr const object_t* get_impl_ptr(const object_t* /*unused*/) const noexcept { - return is_object() ? m_value.object : nullptr; + return is_object() ? m_data.m_value.object : nullptr; } /// get a pointer to the value (array) array_t* get_impl_ptr(array_t* /*unused*/) noexcept { - return is_array() ? m_value.array : nullptr; + return is_array() ? m_data.m_value.array : nullptr; } /// get a pointer to the value (array) constexpr const array_t* get_impl_ptr(const array_t* /*unused*/) const noexcept { - return is_array() ? m_value.array : nullptr; + return is_array() ? m_data.m_value.array : nullptr; } /// get a pointer to the value (string) string_t* get_impl_ptr(string_t* /*unused*/) noexcept { - return is_string() ? m_value.string : nullptr; + return is_string() ? m_data.m_value.string : nullptr; } /// get a pointer to the value (string) constexpr const string_t* get_impl_ptr(const string_t* /*unused*/) const noexcept { - return is_string() ? m_value.string : nullptr; + return is_string() ? m_data.m_value.string : nullptr; } /// get a pointer to the value (boolean) boolean_t* get_impl_ptr(boolean_t* /*unused*/) noexcept { - return is_boolean() ? &m_value.boolean : nullptr; + return is_boolean() ? &m_data.m_value.boolean : nullptr; } /// get a pointer to the value (boolean) constexpr const boolean_t* get_impl_ptr(const boolean_t* /*unused*/) const noexcept { - return is_boolean() ? &m_value.boolean : nullptr; + return is_boolean() ? &m_data.m_value.boolean : nullptr; } /// get a pointer to the value (integer number) number_integer_t* get_impl_ptr(number_integer_t* /*unused*/) noexcept { - return is_number_integer() ? &m_value.number_integer : nullptr; + return is_number_integer() ? &m_data.m_value.number_integer : nullptr; } /// get a pointer to the value (integer number) constexpr const number_integer_t* get_impl_ptr(const number_integer_t* /*unused*/) const noexcept { - return is_number_integer() ? &m_value.number_integer : nullptr; + return is_number_integer() ? &m_data.m_value.number_integer : nullptr; } /// get a pointer to the value (unsigned number) number_unsigned_t* get_impl_ptr(number_unsigned_t* /*unused*/) noexcept { - return is_number_unsigned() ? &m_value.number_unsigned : nullptr; + return is_number_unsigned() ? &m_data.m_value.number_unsigned : nullptr; } /// get a pointer to the value (unsigned number) constexpr const number_unsigned_t* get_impl_ptr(const number_unsigned_t* /*unused*/) const noexcept { - return is_number_unsigned() ? &m_value.number_unsigned : nullptr; + return is_number_unsigned() ? &m_data.m_value.number_unsigned : nullptr; } /// get a pointer to the value (floating-point number) number_float_t* get_impl_ptr(number_float_t* /*unused*/) noexcept { - return is_number_float() ? &m_value.number_float : nullptr; + return is_number_float() ? &m_data.m_value.number_float : nullptr; } /// get a pointer to the value (floating-point number) constexpr const number_float_t* get_impl_ptr(const number_float_t* /*unused*/) const noexcept { - return is_number_float() ? &m_value.number_float : nullptr; + return is_number_float() ? &m_data.m_value.number_float : nullptr; } /// get a pointer to the value (binary) binary_t* get_impl_ptr(binary_t* /*unused*/) noexcept { - return is_binary() ? m_value.binary : nullptr; + return is_binary() ? m_data.m_value.binary : nullptr; } /// get a pointer to the value (binary) constexpr const binary_t* get_impl_ptr(const binary_t* /*unused*/) const noexcept { - return is_binary() ? m_value.binary : nullptr; + return is_binary() ? m_data.m_value.binary : nullptr; } /*! @@ -21053,7 +21191,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec #if defined(JSON_HAS_CPP_17) && (defined(__GNUC__) || (defined(_MSC_VER) && _MSC_VER >= 1910 && _MSC_VER <= 1914)) detail::negation>, #endif -#if defined(JSON_HAS_CPP_17) +#if defined(JSON_HAS_CPP_17) && JSON_HAS_STATIC_RTTI detail::negation>, #endif detail::is_detected_lazy @@ -21090,7 +21228,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @} - //////////////////// // element access // //////////////////// @@ -21108,7 +21245,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec { JSON_TRY { - return set_parent(m_value.array->at(idx)); + return set_parent(m_data.m_value.array->at(idx)); } JSON_CATCH (std::out_of_range&) { @@ -21131,7 +21268,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec { JSON_TRY { - return m_value.array->at(idx); + return m_data.m_value.array->at(idx); } JSON_CATCH (std::out_of_range&) { @@ -21155,8 +21292,8 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this)); } - auto it = m_value.object->find(key); - if (it == m_value.object->end()) + auto it = m_data.m_value.object->find(key); + if (it == m_data.m_value.object->end()) { JSON_THROW(out_of_range::create(403, detail::concat("key '", key, "' not found"), this)); } @@ -21175,8 +21312,8 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this)); } - auto it = m_value.object->find(std::forward(key)); - if (it == m_value.object->end()) + auto it = m_data.m_value.object->find(std::forward(key)); + if (it == m_data.m_value.object->end()) { JSON_THROW(out_of_range::create(403, detail::concat("key '", string_t(std::forward(key)), "' not found"), this)); } @@ -21193,8 +21330,8 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this)); } - auto it = m_value.object->find(key); - if (it == m_value.object->end()) + auto it = m_data.m_value.object->find(key); + if (it == m_data.m_value.object->end()) { JSON_THROW(out_of_range::create(403, detail::concat("key '", key, "' not found"), this)); } @@ -21213,8 +21350,8 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this)); } - auto it = m_value.object->find(std::forward(key)); - if (it == m_value.object->end()) + auto it = m_data.m_value.object->find(std::forward(key)); + if (it == m_data.m_value.object->end()) { JSON_THROW(out_of_range::create(403, detail::concat("key '", string_t(std::forward(key)), "' not found"), this)); } @@ -21228,8 +21365,8 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // implicitly convert null value to an empty array if (is_null()) { - m_type = value_t::array; - m_value.array = create(); + m_data.m_type = value_t::array; + m_data.m_value.array = create(); assert_invariant(); } @@ -21237,17 +21374,17 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec if (JSON_HEDLEY_LIKELY(is_array())) { // fill up array with null values if given idx is outside range - if (idx >= m_value.array->size()) + if (idx >= m_data.m_value.array->size()) { #if JSON_DIAGNOSTICS // remember array size & capacity before resizing - const auto old_size = m_value.array->size(); - const auto old_capacity = m_value.array->capacity(); + const auto old_size = m_data.m_value.array->size(); + const auto old_capacity = m_data.m_value.array->capacity(); #endif - m_value.array->resize(idx + 1); + m_data.m_value.array->resize(idx + 1); #if JSON_DIAGNOSTICS - if (JSON_HEDLEY_UNLIKELY(m_value.array->capacity() != old_capacity)) + if (JSON_HEDLEY_UNLIKELY(m_data.m_value.array->capacity() != old_capacity)) { // capacity has changed: update all parents set_parents(); @@ -21261,7 +21398,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec assert_invariant(); } - return m_value.array->operator[](idx); + return m_data.m_value.array->operator[](idx); } JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a numeric argument with ", type_name()), this)); @@ -21274,7 +21411,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // const operator[] only works for arrays if (JSON_HEDLEY_LIKELY(is_array())) { - return m_value.array->operator[](idx); + return m_data.m_value.array->operator[](idx); } JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a numeric argument with ", type_name()), this)); @@ -21287,15 +21424,15 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // implicitly convert null value to an empty object if (is_null()) { - m_type = value_t::object; - m_value.object = create(); + m_data.m_type = value_t::object; + m_data.m_value.object = create(); assert_invariant(); } // operator[] only works for objects if (JSON_HEDLEY_LIKELY(is_object())) { - auto result = m_value.object->emplace(std::move(key), nullptr); + auto result = m_data.m_value.object->emplace(std::move(key), nullptr); return set_parent(result.first->second); } @@ -21309,8 +21446,8 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // const operator[] only works for objects if (JSON_HEDLEY_LIKELY(is_object())) { - auto it = m_value.object->find(key); - JSON_ASSERT(it != m_value.object->end()); + auto it = m_data.m_value.object->find(key); + JSON_ASSERT(it != m_data.m_value.object->end()); return it->second; } @@ -21340,15 +21477,15 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // implicitly convert null value to an empty object if (is_null()) { - m_type = value_t::object; - m_value.object = create(); + m_data.m_type = value_t::object; + m_data.m_value.object = create(); assert_invariant(); } // operator[] only works for objects if (JSON_HEDLEY_LIKELY(is_object())) { - auto result = m_value.object->emplace(std::forward(key), nullptr); + auto result = m_data.m_value.object->emplace(std::forward(key), nullptr); return set_parent(result.first->second); } @@ -21364,8 +21501,8 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // const operator[] only works for objects if (JSON_HEDLEY_LIKELY(is_object())) { - auto it = m_value.object->find(std::forward(key)); - JSON_ASSERT(it != m_value.object->end()); + auto it = m_data.m_value.object->find(std::forward(key)); + JSON_ASSERT(it != m_data.m_value.object->end()); return it->second; } @@ -21602,7 +21739,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec IteratorType result = end(); - switch (m_type) + switch (m_data.m_type) { case value_t::boolean: case value_t::number_float: @@ -21619,32 +21756,32 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec if (is_string()) { AllocatorType alloc; - std::allocator_traits::destroy(alloc, m_value.string); - std::allocator_traits::deallocate(alloc, m_value.string, 1); - m_value.string = nullptr; + std::allocator_traits::destroy(alloc, m_data.m_value.string); + std::allocator_traits::deallocate(alloc, m_data.m_value.string, 1); + m_data.m_value.string = nullptr; } else if (is_binary()) { AllocatorType alloc; - std::allocator_traits::destroy(alloc, m_value.binary); - std::allocator_traits::deallocate(alloc, m_value.binary, 1); - m_value.binary = nullptr; + std::allocator_traits::destroy(alloc, m_data.m_value.binary); + std::allocator_traits::deallocate(alloc, m_data.m_value.binary, 1); + m_data.m_value.binary = nullptr; } - m_type = value_t::null; + m_data.m_type = value_t::null; assert_invariant(); break; } case value_t::object: { - result.m_it.object_iterator = m_value.object->erase(pos.m_it.object_iterator); + result.m_it.object_iterator = m_data.m_value.object->erase(pos.m_it.object_iterator); break; } case value_t::array: { - result.m_it.array_iterator = m_value.array->erase(pos.m_it.array_iterator); + result.m_it.array_iterator = m_data.m_value.array->erase(pos.m_it.array_iterator); break; } @@ -21672,7 +21809,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec IteratorType result = end(); - switch (m_type) + switch (m_data.m_type) { case value_t::boolean: case value_t::number_float: @@ -21690,33 +21827,33 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec if (is_string()) { AllocatorType alloc; - std::allocator_traits::destroy(alloc, m_value.string); - std::allocator_traits::deallocate(alloc, m_value.string, 1); - m_value.string = nullptr; + std::allocator_traits::destroy(alloc, m_data.m_value.string); + std::allocator_traits::deallocate(alloc, m_data.m_value.string, 1); + m_data.m_value.string = nullptr; } else if (is_binary()) { AllocatorType alloc; - std::allocator_traits::destroy(alloc, m_value.binary); - std::allocator_traits::deallocate(alloc, m_value.binary, 1); - m_value.binary = nullptr; + std::allocator_traits::destroy(alloc, m_data.m_value.binary); + std::allocator_traits::deallocate(alloc, m_data.m_value.binary, 1); + m_data.m_value.binary = nullptr; } - m_type = value_t::null; + m_data.m_type = value_t::null; assert_invariant(); break; } case value_t::object: { - result.m_it.object_iterator = m_value.object->erase(first.m_it.object_iterator, + result.m_it.object_iterator = m_data.m_value.object->erase(first.m_it.object_iterator, last.m_it.object_iterator); break; } case value_t::array: { - result.m_it.array_iterator = m_value.array->erase(first.m_it.array_iterator, + result.m_it.array_iterator = m_data.m_value.array->erase(first.m_it.array_iterator, last.m_it.array_iterator); break; } @@ -21741,7 +21878,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this)); } - return m_value.object->erase(std::forward(key)); + return m_data.m_value.object->erase(std::forward(key)); } template < typename KeyType, detail::enable_if_t < @@ -21754,10 +21891,10 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this)); } - const auto it = m_value.object->find(std::forward(key)); - if (it != m_value.object->end()) + const auto it = m_data.m_value.object->find(std::forward(key)); + if (it != m_data.m_value.object->end()) { - m_value.object->erase(it); + m_data.m_value.object->erase(it); return 1; } return 0; @@ -21795,7 +21932,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec JSON_THROW(out_of_range::create(401, detail::concat("array index ", std::to_string(idx), " is out of range"), this)); } - m_value.array->erase(m_value.array->begin() + static_cast(idx)); + m_data.m_value.array->erase(m_data.m_value.array->begin() + static_cast(idx)); } else { @@ -21805,7 +21942,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @} - //////////// // lookup // //////////// @@ -21821,7 +21957,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec if (is_object()) { - result.m_it.object_iterator = m_value.object->find(key); + result.m_it.object_iterator = m_data.m_value.object->find(key); } return result; @@ -21835,7 +21971,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec if (is_object()) { - result.m_it.object_iterator = m_value.object->find(key); + result.m_it.object_iterator = m_data.m_value.object->find(key); } return result; @@ -21851,7 +21987,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec if (is_object()) { - result.m_it.object_iterator = m_value.object->find(std::forward(key)); + result.m_it.object_iterator = m_data.m_value.object->find(std::forward(key)); } return result; @@ -21867,7 +22003,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec if (is_object()) { - result.m_it.object_iterator = m_value.object->find(std::forward(key)); + result.m_it.object_iterator = m_data.m_value.object->find(std::forward(key)); } return result; @@ -21878,7 +22014,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec size_type count(const typename object_t::key_type& key) const { // return 0 for all nonobject types - return is_object() ? m_value.object->count(key) : 0; + return is_object() ? m_data.m_value.object->count(key) : 0; } /// @brief returns the number of occurrences of a key in a JSON object @@ -21888,14 +22024,14 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec size_type count(KeyType && key) const { // return 0 for all nonobject types - return is_object() ? m_value.object->count(std::forward(key)) : 0; + return is_object() ? m_data.m_value.object->count(std::forward(key)) : 0; } /// @brief check the existence of an element in a JSON object /// @sa https://json.nlohmann.me/api/basic_json/contains/ bool contains(const typename object_t::key_type& key) const { - return is_object() && m_value.object->find(key) != m_value.object->end(); + return is_object() && m_data.m_value.object->find(key) != m_data.m_value.object->end(); } /// @brief check the existence of an element in a JSON object @@ -21904,7 +22040,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec detail::is_usable_as_basic_json_key_type::value, int> = 0> bool contains(KeyType && key) const { - return is_object() && m_value.object->find(std::forward(key)) != m_value.object->end(); + return is_object() && m_data.m_value.object->find(std::forward(key)) != m_data.m_value.object->end(); } /// @brief check the existence of an element in a JSON object given a JSON pointer @@ -21923,7 +22059,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @} - /////////////// // iterators // /////////////// @@ -22062,7 +22197,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @} - ////////////// // capacity // ////////////// @@ -22074,7 +22208,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @sa https://json.nlohmann.me/api/basic_json/empty/ bool empty() const noexcept { - switch (m_type) + switch (m_data.m_type) { case value_t::null: { @@ -22085,13 +22219,13 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec case value_t::array: { // delegate call to array_t::empty() - return m_value.array->empty(); + return m_data.m_value.array->empty(); } case value_t::object: { // delegate call to object_t::empty() - return m_value.object->empty(); + return m_data.m_value.object->empty(); } case value_t::string: @@ -22113,7 +22247,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @sa https://json.nlohmann.me/api/basic_json/size/ size_type size() const noexcept { - switch (m_type) + switch (m_data.m_type) { case value_t::null: { @@ -22124,13 +22258,13 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec case value_t::array: { // delegate call to array_t::size() - return m_value.array->size(); + return m_data.m_value.array->size(); } case value_t::object: { // delegate call to object_t::size() - return m_value.object->size(); + return m_data.m_value.object->size(); } case value_t::string: @@ -22152,18 +22286,18 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @sa https://json.nlohmann.me/api/basic_json/max_size/ size_type max_size() const noexcept { - switch (m_type) + switch (m_data.m_type) { case value_t::array: { // delegate call to array_t::max_size() - return m_value.array->max_size(); + return m_data.m_value.array->max_size(); } case value_t::object: { // delegate call to object_t::max_size() - return m_value.object->max_size(); + return m_data.m_value.object->max_size(); } case value_t::null: @@ -22184,7 +22318,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @} - /////////////// // modifiers // /////////////// @@ -22196,53 +22329,53 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @sa https://json.nlohmann.me/api/basic_json/clear/ void clear() noexcept { - switch (m_type) + switch (m_data.m_type) { case value_t::number_integer: { - m_value.number_integer = 0; + m_data.m_value.number_integer = 0; break; } case value_t::number_unsigned: { - m_value.number_unsigned = 0; + m_data.m_value.number_unsigned = 0; break; } case value_t::number_float: { - m_value.number_float = 0.0; + m_data.m_value.number_float = 0.0; break; } case value_t::boolean: { - m_value.boolean = false; + m_data.m_value.boolean = false; break; } case value_t::string: { - m_value.string->clear(); + m_data.m_value.string->clear(); break; } case value_t::binary: { - m_value.binary->clear(); + m_data.m_value.binary->clear(); break; } case value_t::array: { - m_value.array->clear(); + m_data.m_value.array->clear(); break; } case value_t::object: { - m_value.object->clear(); + m_data.m_value.object->clear(); break; } @@ -22266,15 +22399,15 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // transform null object into an array if (is_null()) { - m_type = value_t::array; - m_value = value_t::array; + m_data.m_type = value_t::array; + m_data.m_value = value_t::array; assert_invariant(); } // add element to array (move semantics) - const auto old_capacity = m_value.array->capacity(); - m_value.array->push_back(std::move(val)); - set_parent(m_value.array->back(), old_capacity); + const auto old_capacity = m_data.m_value.array->capacity(); + m_data.m_value.array->push_back(std::move(val)); + set_parent(m_data.m_value.array->back(), old_capacity); // if val is moved from, basic_json move constructor marks it null, so we do not call the destructor } @@ -22299,15 +22432,15 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // transform null object into an array if (is_null()) { - m_type = value_t::array; - m_value = value_t::array; + m_data.m_type = value_t::array; + m_data.m_value = value_t::array; assert_invariant(); } // add element to array - const auto old_capacity = m_value.array->capacity(); - m_value.array->push_back(val); - set_parent(m_value.array->back(), old_capacity); + const auto old_capacity = m_data.m_value.array->capacity(); + m_data.m_value.array->push_back(val); + set_parent(m_data.m_value.array->back(), old_capacity); } /// @brief add an object to an array @@ -22331,13 +22464,13 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // transform null object into an object if (is_null()) { - m_type = value_t::object; - m_value = value_t::object; + m_data.m_type = value_t::object; + m_data.m_value = value_t::object; assert_invariant(); } // add element to object - auto res = m_value.object->insert(val); + auto res = m_data.m_value.object->insert(val); set_parent(res.first->second); } @@ -22387,15 +22520,15 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // transform null object into an array if (is_null()) { - m_type = value_t::array; - m_value = value_t::array; + m_data.m_type = value_t::array; + m_data.m_value = value_t::array; assert_invariant(); } // add element to array (perfect forwarding) - const auto old_capacity = m_value.array->capacity(); - m_value.array->emplace_back(std::forward(args)...); - return set_parent(m_value.array->back(), old_capacity); + const auto old_capacity = m_data.m_value.array->capacity(); + m_data.m_value.array->emplace_back(std::forward(args)...); + return set_parent(m_data.m_value.array->back(), old_capacity); } /// @brief add an object to an object if key does not exist @@ -22412,13 +22545,13 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // transform null object into an object if (is_null()) { - m_type = value_t::object; - m_value = value_t::object; + m_data.m_type = value_t::object; + m_data.m_value = value_t::object; assert_invariant(); } // add element to array (perfect forwarding) - auto res = m_value.object->emplace(std::forward(args)...); + auto res = m_data.m_value.object->emplace(std::forward(args)...); set_parent(res.first->second); // create result iterator and set iterator to the result of emplace @@ -22436,14 +22569,14 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec iterator insert_iterator(const_iterator pos, Args&& ... args) { iterator result(this); - JSON_ASSERT(m_value.array != nullptr); + JSON_ASSERT(m_data.m_value.array != nullptr); - auto insert_pos = std::distance(m_value.array->begin(), pos.m_it.array_iterator); - m_value.array->insert(pos.m_it.array_iterator, std::forward(args)...); - result.m_it.array_iterator = m_value.array->begin() + insert_pos; + auto insert_pos = std::distance(m_data.m_value.array->begin(), pos.m_it.array_iterator); + m_data.m_value.array->insert(pos.m_it.array_iterator, std::forward(args)...); + result.m_it.array_iterator = m_data.m_value.array->begin() + insert_pos; // This could have been written as: - // result.m_it.array_iterator = m_value.array->insert(pos.m_it.array_iterator, cnt, val); + // result.m_it.array_iterator = m_data.m_value.array->insert(pos.m_it.array_iterator, cnt, val); // but the return value of insert is missing in GCC 4.8, so it is written this way instead. set_parents(); @@ -22570,7 +22703,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec JSON_THROW(invalid_iterator::create(202, "iterators first and last must point to objects", this)); } - m_value.object->insert(first.m_it.object_iterator, last.m_it.object_iterator); + m_data.m_value.object->insert(first.m_it.object_iterator, last.m_it.object_iterator); } /// @brief updates a JSON object from another object, overwriting existing keys @@ -22587,8 +22720,8 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // implicitly convert null value to an empty object if (is_null()) { - m_type = value_t::object; - m_value.object = create(); + m_data.m_type = value_t::object; + m_data.m_value.object = create(); assert_invariant(); } @@ -22613,16 +22746,16 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec { if (merge_objects && it.value().is_object()) { - auto it2 = m_value.object->find(it.key()); - if (it2 != m_value.object->end()) + auto it2 = m_data.m_value.object->find(it.key()); + if (it2 != m_data.m_value.object->end()) { it2->second.update(it.value(), true); continue; } } - m_value.object->operator[](it.key()) = it.value(); + m_data.m_value.object->operator[](it.key()) = it.value(); #if JSON_DIAGNOSTICS - m_value.object->operator[](it.key()).m_parent = this; + m_data.m_value.object->operator[](it.key()).m_parent = this; #endif } } @@ -22632,12 +22765,12 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec void swap(reference other) noexcept ( std::is_nothrow_move_constructible::value&& std::is_nothrow_move_assignable::value&& - std::is_nothrow_move_constructible::value&& + std::is_nothrow_move_constructible::value&& // NOLINT(cppcoreguidelines-noexcept-swap,performance-noexcept-swap) std::is_nothrow_move_assignable::value ) { - std::swap(m_type, other.m_type); - std::swap(m_value, other.m_value); + std::swap(m_data.m_type, other.m_data.m_type); + std::swap(m_data.m_value, other.m_data.m_value); set_parents(); other.set_parents(); @@ -22649,7 +22782,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec friend void swap(reference left, reference right) noexcept ( std::is_nothrow_move_constructible::value&& std::is_nothrow_move_assignable::value&& - std::is_nothrow_move_constructible::value&& + std::is_nothrow_move_constructible::value&& // NOLINT(cppcoreguidelines-noexcept-swap,performance-noexcept-swap) std::is_nothrow_move_assignable::value ) { @@ -22658,13 +22791,13 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @brief exchanges the values /// @sa https://json.nlohmann.me/api/basic_json/swap/ - void swap(array_t& other) // NOLINT(bugprone-exception-escape) + void swap(array_t& other) // NOLINT(bugprone-exception-escape,cppcoreguidelines-noexcept-swap,performance-noexcept-swap) { // swap only works for arrays if (JSON_HEDLEY_LIKELY(is_array())) { using std::swap; - swap(*(m_value.array), other); + swap(*(m_data.m_value.array), other); } else { @@ -22674,13 +22807,13 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @brief exchanges the values /// @sa https://json.nlohmann.me/api/basic_json/swap/ - void swap(object_t& other) // NOLINT(bugprone-exception-escape) + void swap(object_t& other) // NOLINT(bugprone-exception-escape,cppcoreguidelines-noexcept-swap,performance-noexcept-swap) { // swap only works for objects if (JSON_HEDLEY_LIKELY(is_object())) { using std::swap; - swap(*(m_value.object), other); + swap(*(m_data.m_value.object), other); } else { @@ -22690,13 +22823,13 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @brief exchanges the values /// @sa https://json.nlohmann.me/api/basic_json/swap/ - void swap(string_t& other) // NOLINT(bugprone-exception-escape) + void swap(string_t& other) // NOLINT(bugprone-exception-escape,cppcoreguidelines-noexcept-swap,performance-noexcept-swap) { // swap only works for strings if (JSON_HEDLEY_LIKELY(is_string())) { using std::swap; - swap(*(m_value.string), other); + swap(*(m_data.m_value.string), other); } else { @@ -22706,13 +22839,13 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec /// @brief exchanges the values /// @sa https://json.nlohmann.me/api/basic_json/swap/ - void swap(binary_t& other) // NOLINT(bugprone-exception-escape) + void swap(binary_t& other) // NOLINT(bugprone-exception-escape,cppcoreguidelines-noexcept-swap,performance-noexcept-swap) { // swap only works for strings if (JSON_HEDLEY_LIKELY(is_binary())) { using std::swap; - swap(*(m_value.binary), other); + swap(*(m_data.m_value.binary), other); } else { @@ -22728,7 +22861,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec if (JSON_HEDLEY_LIKELY(is_binary())) { using std::swap; - swap(*(m_value.binary), other); + swap(*(m_data.m_value.binary), other); } else { @@ -22756,31 +22889,31 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec switch (lhs_type) \ { \ case value_t::array: \ - return (*lhs.m_value.array) op (*rhs.m_value.array); \ + return (*lhs.m_data.m_value.array) op (*rhs.m_data.m_value.array); \ \ case value_t::object: \ - return (*lhs.m_value.object) op (*rhs.m_value.object); \ + return (*lhs.m_data.m_value.object) op (*rhs.m_data.m_value.object); \ \ case value_t::null: \ return (null_result); \ \ case value_t::string: \ - return (*lhs.m_value.string) op (*rhs.m_value.string); \ + return (*lhs.m_data.m_value.string) op (*rhs.m_data.m_value.string); \ \ case value_t::boolean: \ - return (lhs.m_value.boolean) op (rhs.m_value.boolean); \ + return (lhs.m_data.m_value.boolean) op (rhs.m_data.m_value.boolean); \ \ case value_t::number_integer: \ - return (lhs.m_value.number_integer) op (rhs.m_value.number_integer); \ + return (lhs.m_data.m_value.number_integer) op (rhs.m_data.m_value.number_integer); \ \ case value_t::number_unsigned: \ - return (lhs.m_value.number_unsigned) op (rhs.m_value.number_unsigned); \ + return (lhs.m_data.m_value.number_unsigned) op (rhs.m_data.m_value.number_unsigned); \ \ case value_t::number_float: \ - return (lhs.m_value.number_float) op (rhs.m_value.number_float); \ + return (lhs.m_data.m_value.number_float) op (rhs.m_data.m_value.number_float); \ \ case value_t::binary: \ - return (*lhs.m_value.binary) op (*rhs.m_value.binary); \ + return (*lhs.m_data.m_value.binary) op (*rhs.m_data.m_value.binary); \ \ case value_t::discarded: \ default: \ @@ -22789,27 +22922,27 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec } \ else if (lhs_type == value_t::number_integer && rhs_type == value_t::number_float) \ { \ - return static_cast(lhs.m_value.number_integer) op rhs.m_value.number_float; \ + return static_cast(lhs.m_data.m_value.number_integer) op rhs.m_data.m_value.number_float; \ } \ else if (lhs_type == value_t::number_float && rhs_type == value_t::number_integer) \ { \ - return lhs.m_value.number_float op static_cast(rhs.m_value.number_integer); \ + return lhs.m_data.m_value.number_float op static_cast(rhs.m_data.m_value.number_integer); \ } \ else if (lhs_type == value_t::number_unsigned && rhs_type == value_t::number_float) \ { \ - return static_cast(lhs.m_value.number_unsigned) op rhs.m_value.number_float; \ + return static_cast(lhs.m_data.m_value.number_unsigned) op rhs.m_data.m_value.number_float; \ } \ else if (lhs_type == value_t::number_float && rhs_type == value_t::number_unsigned) \ { \ - return lhs.m_value.number_float op static_cast(rhs.m_value.number_unsigned); \ + return lhs.m_data.m_value.number_float op static_cast(rhs.m_data.m_value.number_unsigned); \ } \ else if (lhs_type == value_t::number_unsigned && rhs_type == value_t::number_integer) \ { \ - return static_cast(lhs.m_value.number_unsigned) op rhs.m_value.number_integer; \ + return static_cast(lhs.m_data.m_value.number_unsigned) op rhs.m_data.m_value.number_integer; \ } \ else if (lhs_type == value_t::number_integer && rhs_type == value_t::number_unsigned) \ { \ - return lhs.m_value.number_integer op static_cast(rhs.m_value.number_unsigned); \ + return lhs.m_data.m_value.number_integer op static_cast(rhs.m_data.m_value.number_unsigned); \ } \ else if(compares_unordered(lhs, rhs))\ {\ @@ -22826,8 +22959,8 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // an operation is computed as an odd number of inverses of others static bool compares_unordered(const_reference lhs, const_reference rhs, bool inverse = false) noexcept { - if ((lhs.is_number_float() && std::isnan(lhs.m_value.number_float) && rhs.is_number()) - || (rhs.is_number_float() && std::isnan(rhs.m_value.number_float) && lhs.is_number())) + if ((lhs.is_number_float() && std::isnan(lhs.m_data.m_value.number_float) && rhs.is_number()) + || (rhs.is_number_float() && std::isnan(rhs.m_data.m_value.number_float) && lhs.is_number())) { return true; } @@ -23171,7 +23304,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec #endif // JSON_NO_IO /// @} - ///////////////////// // deserialization // ///////////////////// @@ -23328,7 +23460,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec JSON_HEDLEY_RETURNS_NON_NULL const char* type_name() const noexcept { - switch (m_type) + switch (m_data.m_type) { case value_t::null: return "null"; @@ -23352,17 +23484,43 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec } } - JSON_PRIVATE_UNLESS_TESTED: ////////////////////// // member variables // ////////////////////// - /// the type of the current element - value_t m_type = value_t::null; + struct data + { + /// the type of the current element + value_t m_type = value_t::null; - /// the value of the current element - json_value m_value = {}; + /// the value of the current element + json_value m_value = {}; + + data(const value_t v) + : m_type(v), m_value(v) + { + } + + data(size_type cnt, const basic_json& val) + : m_type(value_t::array) + { + m_value.array = create(cnt, val); + } + + data() noexcept = default; + data(data&&) noexcept = default; + data(const data&) noexcept = delete; + data& operator=(data&&) noexcept = delete; + data& operator=(const data&) noexcept = delete; + + ~data() noexcept + { + m_value.destroy(m_type); + } + }; + + data m_data = {}; #if JSON_DIAGNOSTICS /// a pointer to a parent value (for debugging purposes) @@ -23543,7 +23701,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec return from_cbor(ptr, ptr + len, strict, allow_exceptions, tag_handler); } - JSON_HEDLEY_WARN_UNUSED_RESULT JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_cbor(ptr, ptr + len)) static basic_json from_cbor(detail::span_input_adapter&& i, @@ -23667,7 +23824,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec return res ? result : basic_json(value_t::discarded); } - /// @brief create a JSON value from an input in BJData format /// @sa https://json.nlohmann.me/api/basic_json/from_bjdata/ template @@ -23890,7 +24046,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec } // make sure the top element of the pointer exists - json_pointer top_pointer = ptr.top(); + json_pointer const top_pointer = ptr.top(); if (top_pointer != ptr) { result.at(top_pointer); @@ -23902,7 +24058,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec // parent must exist when performing patch add per RFC6902 specs basic_json& parent = result.at(ptr); - switch (parent.m_type) + switch (parent.m_data.m_type) { case value_t::null: case value_t::object: @@ -23948,7 +24104,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec }; // wrapper for "remove" operation; remove value at ptr - const auto operation_remove = [this, &result](json_pointer & ptr) + const auto operation_remove = [this, & result](json_pointer & ptr) { // get reference to parent of JSON pointer ptr const auto last_path = ptr.back(); @@ -23991,13 +24147,13 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec bool string_type) -> basic_json & { // find value - auto it = val.m_value.object->find(member); + auto it = val.m_data.m_value.object->find(member); // context-sensitive error message - const auto error_msg = (op == "op") ? "operation" : detail::concat("operation '", op, '\''); + const auto error_msg = (op == "op") ? "operation" : detail::concat("operation '", op, '\''); // NOLINT(bugprone-unused-local-non-trivial-variable) // check if desired value is present - if (JSON_HEDLEY_UNLIKELY(it == val.m_value.object->end())) + if (JSON_HEDLEY_UNLIKELY(it == val.m_data.m_value.object->end())) { // NOLINTNEXTLINE(performance-inefficient-string-concatenation) JSON_THROW(parse_error::create(105, 0, detail::concat(error_msg, " must have member '", member, "'"), &val)); @@ -24052,7 +24208,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec json_pointer from_ptr(from_path); // the "from" location must exist - use at() - basic_json v = result.at(from_ptr); + basic_json const v = result.at(from_ptr); // The move operation is functionally identical to a // "remove" operation on the "from" location, followed @@ -24069,7 +24225,7 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec const json_pointer from_ptr(from_path); // the "from" location must exist - use at() - basic_json v = result.at(from_ptr); + basic_json const v = result.at(from_ptr); // The copy is functionally identical to an "add" // operation at the target location using the value @@ -24311,7 +24467,11 @@ inline namespace json_literals /// @brief user-defined string literal for JSON values /// @sa https://json.nlohmann.me/api/basic_json/operator_literal_json/ JSON_HEDLEY_NON_NULL(1) -inline nlohmann::json operator "" _json(const char* s, std::size_t n) +#if !defined(JSON_HEDLEY_GCC_VERSION) || JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0) + inline nlohmann::json operator ""_json(const char* s, std::size_t n) +#else + inline nlohmann::json operator "" _json(const char* s, std::size_t n) +#endif { return nlohmann::json::parse(s, s + n); } @@ -24319,7 +24479,11 @@ inline nlohmann::json operator "" _json(const char* s, std::size_t n) /// @brief user-defined string literal for JSON pointer /// @sa https://json.nlohmann.me/api/basic_json/operator_literal_json_pointer/ JSON_HEDLEY_NON_NULL(1) -inline nlohmann::json::json_pointer operator "" _json_pointer(const char* s, std::size_t n) +#if !defined(JSON_HEDLEY_GCC_VERSION) || JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0) + inline nlohmann::json::json_pointer operator ""_json_pointer(const char* s, std::size_t n) +#else + inline nlohmann::json::json_pointer operator "" _json_pointer(const char* s, std::size_t n) +#endif { return nlohmann::json::json_pointer(std::string(s, n)); } @@ -24338,7 +24502,7 @@ namespace std // NOLINT(cert-dcl58-cpp) /// @brief hash value for JSON objects /// @sa https://json.nlohmann.me/api/basic_json/std_hash/ NLOHMANN_BASIC_JSON_TPL_DECLARATION -struct hash +struct hash // NOLINT(cert-dcl58-cpp) { std::size_t operator()(const nlohmann::NLOHMANN_BASIC_JSON_TPL& j) const { @@ -24371,8 +24535,8 @@ struct less< ::nlohmann::detail::value_t> // do not remove the space after '<', /// @brief exchanges the values of two JSON objects /// @sa https://json.nlohmann.me/api/basic_json/std_swap/ NLOHMANN_BASIC_JSON_TPL_DECLARATION -inline void swap(nlohmann::NLOHMANN_BASIC_JSON_TPL& j1, nlohmann::NLOHMANN_BASIC_JSON_TPL& j2) noexcept( // NOLINT(readability-inconsistent-declaration-parameter-name) - is_nothrow_move_constructible::value&& // NOLINT(misc-redundant-expression) +inline void swap(nlohmann::NLOHMANN_BASIC_JSON_TPL& j1, nlohmann::NLOHMANN_BASIC_JSON_TPL& j2) noexcept( // NOLINT(readability-inconsistent-declaration-parameter-name, cert-dcl58-cpp) + is_nothrow_move_constructible::value&& // NOLINT(misc-redundant-expression,cppcoreguidelines-noexcept-swap,performance-noexcept-swap) is_nothrow_move_assignable::value) { j1.swap(j2); @@ -24383,17 +24547,22 @@ inline void swap(nlohmann::NLOHMANN_BASIC_JSON_TPL& j1, nlohmann::NLOHMANN_BASIC } // namespace std #if JSON_USE_GLOBAL_UDLS - using nlohmann::literals::json_literals::operator "" _json; // NOLINT(misc-unused-using-decls,google-global-names-in-headers) - using nlohmann::literals::json_literals::operator "" _json_pointer; //NOLINT(misc-unused-using-decls,google-global-names-in-headers) + #if !defined(JSON_HEDLEY_GCC_VERSION) || JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0) + using nlohmann::literals::json_literals::operator ""_json; // NOLINT(misc-unused-using-decls,google-global-names-in-headers) + using nlohmann::literals::json_literals::operator ""_json_pointer; //NOLINT(misc-unused-using-decls,google-global-names-in-headers) + #else + using nlohmann::literals::json_literals::operator "" _json; // NOLINT(misc-unused-using-decls,google-global-names-in-headers) + using nlohmann::literals::json_literals::operator "" _json_pointer; //NOLINT(misc-unused-using-decls,google-global-names-in-headers) + #endif #endif // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT @@ -24428,16 +24597,17 @@ inline void swap(nlohmann::NLOHMANN_BASIC_JSON_TPL& j1, nlohmann::NLOHMANN_BASIC #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM #undef JSON_HAS_THREE_WAY_COMPARISON #undef JSON_HAS_RANGES + #undef JSON_HAS_STATIC_RTTI #undef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON #endif // #include // __ _____ _____ _____ // __| | __| | | | JSON for Modern C++ -// | | |__ | | | | | | version 3.11.2 +// | | |__ | | | | | | version 3.11.3 // |_____|_____|_____|_|___| https://github.com/nlohmann/json // -// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann +// SPDX-FileCopyrightText: 2013-2023 Niels Lohmann // SPDX-License-Identifier: MIT diff --git a/common/log.h b/common/log.h index eb111e784..e4edcac7d 100644 --- a/common/log.h +++ b/common/log.h @@ -234,7 +234,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std:: // INTERNAL, DO NOT USE // USE LOG() INSTEAD // -#ifndef _MSC_VER +#if !defined(_MSC_VER) or defined(__INTEL_LLVM_COMPILER) #define LOG_IMPL(str, ...) \ do { \ if (LOG_TARGET != nullptr) \ @@ -257,7 +257,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std:: // INTERNAL, DO NOT USE // USE LOG_TEE() INSTEAD // -#ifndef _MSC_VER +#if !defined(_MSC_VER) or defined(__INTEL_LLVM_COMPILER) #define LOG_TEE_IMPL(str, ...) \ do { \ if (LOG_TARGET != nullptr) \ @@ -566,6 +566,7 @@ inline void log_print_usage() printf(" --log-new Create a separate new log file on start. " "Each log file will have unique name: \"..log\"\n"); printf(" --log-append Don't truncate the old log file.\n"); + printf("\n"); } #define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv) diff --git a/common/ngram-cache.cpp b/common/ngram-cache.cpp new file mode 100644 index 000000000..3ca112ef1 --- /dev/null +++ b/common/ngram-cache.cpp @@ -0,0 +1,282 @@ +#include "ngram-cache.h" +#include "common.h" +#include "log.h" + +#include +#include + +void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max, + std::vector & inp, int nnew, bool print_progress) { + const int64_t t_start_ms = ggml_time_ms(); + const int64_t inp_size = inp.size(); + + const int64_t n_todo = inp_size * (ngram_max - ngram_min + 1); + int64_t n_done = 0; + + for (int64_t ngram_size = ngram_min; ngram_size <= ngram_max; ++ngram_size) { + const int64_t i_start = std::max(inp_size - nnew, ngram_size); + for (int64_t i = i_start; i < inp_size; ++i) { + const int64_t ngram_start = i - ngram_size; + llama_ngram ngram(&inp[ngram_start], ngram_size); + const llama_token token = inp[i]; + + llama_ngram_cache::iterator part_it = ngram_cache.find(ngram); + if (part_it == ngram_cache.end()) { + llama_ngram_cache_part part; + part.emplace(token, 1); + ngram_cache.emplace(ngram, part); + } else { + llama_ngram_cache_part::iterator token_count_it = part_it->second.find(token); + if (token_count_it == part_it->second.end()) { + part_it->second.emplace(token, 1); + } else { + token_count_it->second++; + } + } + ++n_done; + + if (print_progress && n_done % 10000000 == 0) { + const int64_t t_now_ms = ggml_time_ms(); + const int64_t eta_ms = (inp_size*(ngram_max-ngram_min+1) - n_done) * (t_now_ms - t_start_ms) / n_done; + const int64_t eta_min = eta_ms / (60*1000); + const int64_t eta_s = (eta_ms - 60*1000*eta_min) / 1000; + + fprintf(stderr, "%s: %" PRId64 "/%" PRId64 " done, ETA: %02" PRId64 ":%02" PRId64 "\n", __func__, n_done, n_todo, eta_min, eta_s); + } + } + } +} + +// Helper function to get a token from the combined, speculative sequence of inp and draft. +static llama_token get_token(const std::vector & inp, const std::vector & draft, const size_t i) { + return i < inp.size() ? inp[i] : draft[1 + i - inp.size()]; +} + +// If sample size or percentage are below these thresholds the draft is aborted early: +constexpr int draft_min_sample_size_lax[LLAMA_NGRAM_MAX] = { 2, 2, 1, 1}; +constexpr int draft_min_percent_lax[LLAMA_NGRAM_MAX] = {66, 50, 50, 50}; +constexpr int draft_min_sample_size_strict[LLAMA_NGRAM_MAX] = { 4, 3, 2, 2}; +constexpr int draft_min_percent_strict[LLAMA_NGRAM_MAX] = {75, 66, 66, 66}; + +// Helper function that tries to draft a token from only the static ngram cache: +static llama_token try_draft(llama_ngram_cache & nc_static, const llama_ngram ngram_static) { + llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static); + if (part_static_it == nc_static.end()) { + return -1; + } + const llama_ngram_cache_part part_static = part_static_it->second; + + int max_count_static = 0; + int sum_count_static = 0; + llama_token max_token = -1; + + for (std::pair token_count_static : part_static) { + const llama_token token = token_count_static.first; + const int32_t count_static = token_count_static.second; + + if (count_static > max_count_static) { + max_token = token; + max_count_static = count_static; + } + sum_count_static += count_static; + } + + if (sum_count_static < draft_min_sample_size_lax[LLAMA_NGRAM_STATIC-1]) { + return -1; + } + if (100*max_count_static < draft_min_percent_lax[LLAMA_NGRAM_STATIC-1]*sum_count_static) { + return -1; + } + return max_token; +} + +// Try to draft a token from primary cache (context/dynamic), validate with static cache: +static llama_token try_draft( + llama_ngram_cache & nc_primary, const std::vector & ngrams_primary, llama_ngram_cache_part & part_static, + const int * min_sample_size, const int * min_percent) { + + llama_token drafted_token = -1; + + for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == -1; --i) { + const llama_ngram ngram_primary = ngrams_primary[i]; + + llama_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary); + if (part_primary_it == nc_primary.end()) { + continue; + } + const llama_ngram_cache_part part_primary = part_primary_it->second; + + int max_count_primary = 0; + int max_count_static = 0; + int sum_count_primary = 0; + llama_token max_token = -1; + + for (std::pair token_count_primary : part_primary) { + const llama_token token = token_count_primary.first; + + llama_ngram_cache_part::iterator token_count_static_it = part_static.find(token); + + const int32_t count_primary = token_count_primary.second; + const int32_t count_static = token_count_static_it != part_static.end() ? 100*token_count_static_it->second : 1; + + if (count_primary*count_static > max_count_primary*max_count_static) { + max_token = token; + max_count_primary = count_primary; + max_count_static = count_static; + } + sum_count_primary += count_primary; + } + + if (sum_count_primary < min_sample_size[i]) { + continue; + } + if (100*max_count_primary < min_percent[i]*sum_count_primary) { + continue;; + } + drafted_token = max_token; + } + + return drafted_token; +} + +void llama_ngram_cache_draft( + std::vector & inp, std::vector & draft, int n_draft, int ngram_min, int ngram_max, + llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static +) { + GGML_ASSERT(draft.size() == 1); + const int inp_size = inp.size(); + + if (inp_size < LLAMA_NGRAM_STATIC) { + return; + } + + while ((int) draft.size()-1 < n_draft) { + llama_token drafted_token = -1; + + const int ngram_start_static = inp_size-LLAMA_NGRAM_STATIC + draft.size()-1; + llama_ngram ngram_static; + for (int j = ngram_start_static; j < ngram_start_static + LLAMA_NGRAM_STATIC; ++j) { + ngram_static.tokens[j-ngram_start_static] = get_token(inp, draft, j); + } + llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static); + llama_ngram_cache_part part_static; + if (part_static_it != nc_static.end()) { + part_static = part_static_it->second; + } + + // cd = context + dynamic + std::vector ngrams_cd; + for (int ngram_size_cd = ngram_min; ngram_size_cd <= ngram_max; ++ngram_size_cd) { + const int ngram_start_cd = inp_size-ngram_size_cd + draft.size()-1; + llama_ngram ngram_cd; + for (int j = ngram_start_cd; j < ngram_start_cd + ngram_size_cd; ++j) { + ngram_cd.tokens[j-ngram_start_cd] = get_token(inp, draft, j); + } + ngrams_cd.push_back(ngram_cd); + } + if (drafted_token == -1) { + drafted_token = try_draft(nc_context, ngrams_cd, part_static, draft_min_sample_size_lax, draft_min_percent_lax); + } + if (drafted_token == -1) { + drafted_token = try_draft(nc_dynamic, ngrams_cd, part_static, draft_min_sample_size_strict, draft_min_percent_strict); + } + if (drafted_token == -1) { + drafted_token = try_draft(nc_static, ngram_static); + } + + if (drafted_token == -1) { + break; + } + + LOG(" - draft candidate: token=%d\n", drafted_token); + draft.push_back(drafted_token); + } +} + +void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename) { + std::ofstream file_out(filename, std::ios::binary); + for (std::pair item : ngram_cache) { + const llama_ngram ngram = item.first; + llama_ngram_cache_part token_counts = item.second; + GGML_ASSERT(!token_counts.empty()); + const int32_t ntokens = token_counts.size(); + GGML_ASSERT(ntokens > 0); + + file_out.write(reinterpret_cast(&ngram), sizeof(llama_ngram)); + file_out.write(reinterpret_cast(&ntokens), sizeof(int32_t)); + for (std::pair item2 : token_counts) { + const llama_token token = item2.first; + const int32_t count = item2.second; + GGML_ASSERT(count > 0); + + file_out.write(reinterpret_cast(&token), sizeof(llama_token)); + file_out.write(reinterpret_cast(&count), sizeof(int32_t)); + } + } + +} + +llama_ngram_cache llama_ngram_cache_load(std::string & filename) { + std::ifstream hashmap_file(filename, std::ios::binary); + if (!hashmap_file) { + throw std::ifstream::failure("Unable to open file " + filename); + } + llama_ngram_cache ngram_cache; + + llama_ngram ngram; + int32_t ntokens; + llama_token token; + int32_t count; + + char * ngramc = reinterpret_cast(&ngram); + char * ntokensc = reinterpret_cast(&ntokens); + char * tokenc = reinterpret_cast(&token); + char * countc = reinterpret_cast(&count); + while(hashmap_file.read(ngramc, sizeof(llama_ngram))) { + GGML_ASSERT(!hashmap_file.eof()); + GGML_ASSERT(hashmap_file.read(ntokensc, sizeof(int32_t))); + GGML_ASSERT(ntokens > 0); + llama_ngram_cache_part token_counts; + + for (int i = 0; i < ntokens; ++i) { + GGML_ASSERT(!hashmap_file.eof()); + GGML_ASSERT(hashmap_file.read(tokenc, sizeof(llama_token))); + GGML_ASSERT(!hashmap_file.eof()); + GGML_ASSERT(hashmap_file.read(countc, sizeof(int32_t))); + GGML_ASSERT(count > 0); + token_counts.emplace(token, count); + } + + ngram_cache.emplace(ngram, token_counts); + } + GGML_ASSERT(hashmap_file.eof()); + + return ngram_cache; +} + +void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add) { + for (std::pair ngram_part : ngram_cache_add) { + const llama_ngram ngram = ngram_part.first; + llama_ngram_cache_part part = ngram_part.second; + + llama_ngram_cache::iterator part_merged_it = ngram_cache_target.find(ngram); + if (part_merged_it == ngram_cache_target.end()) { + ngram_cache_target.emplace(ngram, part); + continue; + } + + for (std::pair token_count : part) { + const llama_token token = token_count.first; + const int32_t count = token_count.second; + GGML_ASSERT(count > 0); + + llama_ngram_cache_part::iterator token_count_merged_it = part_merged_it->second.find(token); + if (token_count_merged_it == part_merged_it->second.end()) { + part_merged_it->second.emplace(token, count); + continue; + } + + token_count_merged_it->second += count; + } + } +} diff --git a/common/ngram-cache.h b/common/ngram-cache.h new file mode 100644 index 000000000..e4fa4cbd1 --- /dev/null +++ b/common/ngram-cache.h @@ -0,0 +1,94 @@ +#pragma once + +#include "llama.h" + +#include +#include +#include + +#define LLAMA_NGRAM_MIN 1 +#define LLAMA_NGRAM_MAX 4 +#define LLAMA_NGRAM_STATIC 2 + +// Data structures to map n-grams to empirical token probabilities: + +struct llama_ngram { + llama_token tokens[LLAMA_NGRAM_MAX]; + + llama_ngram() { + for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) { + tokens[i] = -1; + } + } + + llama_ngram(const llama_token * input, const int ngram_size) { + for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) { + tokens[i] = i < ngram_size ? input[i] : -1; + } + } + + bool operator==(const llama_ngram & other) const { + for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) { + if (tokens[i] != other.tokens[i]) { + return false; + } + } + return true; + } +}; + +struct llama_ngram_hash_function { + size_t operator()(const llama_ngram & ngram) const { + size_t hash = 0; + for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) { + hash ^= std::hash{}(ngram.tokens[i]); + } + return hash; + } +}; + +// token -> number of times token has been seen +typedef std::unordered_map llama_ngram_cache_part; + +// n-gram -> empirical distribution of following tokens +typedef std::unordered_map llama_ngram_cache; + + +// Update an ngram cache with tokens. +// ngram_cache: the cache to modify. +// ngram_min/ngram_max: the min/max size of the ngrams to extract from inp_data. +// inp_data: the token sequence with which to update ngram_cache. +// nnew: how many new tokens have been appended to inp_data since the last call to this function. +// print_progress: whether to print progress to stderr. +// +// In order to get correct results inp_data can ONLY BE APPENDED TO. +// Changes in the middle need a complete rebuild. +void llama_ngram_cache_update( + llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max, std::vector & inp_data, int nnew, bool print_progress); + +// Try to draft tokens from ngram caches. +// inp: the tokens generated so far. +// draft: the token sequence to draft. Expected to initially contain the previously sampled token. +// n_draft: maximum number of tokens to add to draft. +// ngram_min/gram_max: the min/max size of the ngrams in nc_context and nc_dynamic. +// nc_context: ngram cache based on current context. +// nc_dynamic: ngram cache based on previous user generations. +// nc_static: ngram cache generated from a large text corpus, used for validation. +void llama_ngram_cache_draft( + std::vector & inp, std::vector & draft, int n_draft, int ngram_min, int ngram_max, + llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static); + +// Save an ngram cache to a file. +// ngram_cache: the ngram cache to save. +// filename: the path under which to save the ngram cache. +void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename); + +// Load an ngram cache saved with llama_ngram_cache_save. +// filename: the path from which to load the ngram cache. +// returns: an ngram cache containing the information saved to filename. +llama_ngram_cache llama_ngram_cache_load(std::string & filename); + +// Merge two ngram caches. +// ngram_cache_target: the ngram cache to which to add the information from ngram_cache_add. +// ngram_cache_add: the ngram cache to add to ngram_cache_target. +void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add); diff --git a/common/sampling.cpp b/common/sampling.cpp index 5a5450982..45d68b26c 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -168,77 +168,20 @@ static llama_token llama_sampling_sample_impl( bool is_resampling) { // Add a parameter to indicate if we are resampling const llama_sampling_params & params = ctx_sampling->params; - const int n_vocab = llama_n_vocab(llama_get_model(ctx_main)); - const float temp = params.temp; - const int32_t penalty_last_n = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n; - const float penalty_repeat = params.penalty_repeat; - const float penalty_freq = params.penalty_freq; - const float penalty_present = params.penalty_present; const int mirostat = params.mirostat; const float mirostat_tau = params.mirostat_tau; const float mirostat_eta = params.mirostat_eta; - const bool penalize_nl = params.penalize_nl; - - auto & prev = ctx_sampling->prev; - auto & cur = ctx_sampling->cur; + std::vector original_logits; + auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, !is_resampling, &original_logits); + if (!is_resampling) { + GGML_ASSERT(!original_logits.empty()); + } llama_token id = 0; - // Get a pointer to the logits float * logits = llama_get_logits_ith(ctx_main, idx); - // Declare original_logits at the beginning of the function scope - std::vector original_logits; - - if (!is_resampling) { - // Only make a copy of the original logits if we are not in the resampling phase, not sure if I actually have to do this. - original_logits = std::vector(logits, logits + llama_n_vocab(llama_get_model(ctx_main))); - } - - // apply params.logit_bias map - for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) { - logits[it->first] += it->second; - } - - if (ctx_cfg) { - float * logits_guidance = llama_get_logits_ith(ctx_cfg, idx); - llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale); - } - - cur.clear(); - - for (llama_token token_id = 0; token_id < n_vocab; token_id++) { - cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f}); - } - - llama_token_data_array cur_p = { cur.data(), cur.size(), false }; - - // apply penalties - const auto& penalty_tokens = params.use_penalty_prompt_tokens ? params.penalty_prompt_tokens : prev; - const int penalty_tokens_used_size = std::min((int)penalty_tokens.size(), penalty_last_n); - if (penalty_tokens_used_size) { - const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))]; - - llama_sample_repetition_penalties(ctx_main, &cur_p, - penalty_tokens.data() + penalty_tokens.size() - penalty_tokens_used_size, - penalty_tokens_used_size, penalty_repeat, penalty_freq, penalty_present); - - if (!penalize_nl) { - for (size_t idx = 0; idx < cur_p.size; idx++) { - if (cur_p.data[idx].id == llama_token_nl(llama_get_model(ctx_main))) { - cur_p.data[idx].logit = nl_logit; - break; - } - } - } - } - - // If we are in the resampling phase, apply grammar checks before sampling logic - if (is_resampling && ctx_sampling->grammar != NULL) { - llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar); - } - if (temp < 0.0) { // greedy sampling, with probs llama_sample_softmax(ctx_main, &cur_p); @@ -302,11 +245,13 @@ static llama_token llama_sampling_sample_impl( return id; } -static llama_token_data_array llama_sample_probability_distribution_impl( +static llama_token_data_array llama_sampling_prepare_impl( struct llama_sampling_context * ctx_sampling, struct llama_context * ctx_main, struct llama_context * ctx_cfg, - const int idx) { + const int idx, + bool apply_grammar, + std::vector * original_logits) { const llama_sampling_params & params = ctx_sampling->params; const int n_vocab = llama_n_vocab(llama_get_model(ctx_main)); @@ -315,6 +260,7 @@ static llama_token_data_array llama_sample_probability_distribution_impl( const float penalty_repeat = params.penalty_repeat; const float penalty_freq = params.penalty_freq; const float penalty_present = params.penalty_present; + const bool penalize_nl = params.penalize_nl; auto & prev = ctx_sampling->prev; @@ -323,8 +269,10 @@ static llama_token_data_array llama_sample_probability_distribution_impl( // Get a pointer to the logits float * logits = llama_get_logits_ith(ctx_main, idx); - // Declare original_logits at the beginning of the function scope - std::vector original_logits; + if (apply_grammar && original_logits != NULL) { + // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this. + *original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))}; + } // apply params.logit_bias map for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) { @@ -364,12 +312,11 @@ static llama_token_data_array llama_sample_probability_distribution_impl( } } - // apply grammar checks - if (ctx_sampling->grammar != NULL) { + // apply grammar checks before sampling logic + if (apply_grammar && ctx_sampling->grammar != NULL) { llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar); } - llama_sample_softmax(ctx_main, &cur_p); return cur_p; } @@ -382,12 +329,14 @@ llama_token llama_sampling_sample( return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false); } -llama_token_data_array llama_sampling_probability_distribution( +llama_token_data_array llama_sampling_prepare( struct llama_sampling_context * ctx_sampling, struct llama_context * ctx_main, struct llama_context * ctx_cfg, - const int idx) { - return llama_sample_probability_distribution_impl(ctx_sampling,ctx_main, ctx_cfg, idx); + const int idx, + bool apply_grammar, + std::vector * original_logits) { + return llama_sampling_prepare_impl(ctx_sampling,ctx_main, ctx_cfg, idx, apply_grammar, original_logits); } void llama_sampling_accept( diff --git a/common/sampling.h b/common/sampling.h index 79a998be8..56ed991b8 100644 --- a/common/sampling.h +++ b/common/sampling.h @@ -131,12 +131,14 @@ llama_token llama_sampling_sample( struct llama_context * ctx_cfg, int idx = 0); -// returns the probability that token of given id will be sampled -llama_token_data_array llama_sampling_probability_distribution( +// Prepares and adjusts the set of token candidates for sampling based on penalties, biases, and sampling parameters. +llama_token_data_array llama_sampling_prepare( struct llama_sampling_context * ctx_sampling, struct llama_context * ctx_main, struct llama_context * ctx_cfg, - int idx = 0); + int idx = 0, + bool apply_grammar = true, + std::vector * original_logits = nullptr); void llama_sampling_accept( struct llama_sampling_context * ctx_sampling, diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 1e49d56c1..723ea18e3 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -93,31 +93,42 @@ class Model(ABC): if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None: self.gguf_writer.add_context_length(n_ctx) + print(f"gguf: context length = {n_ctx}") n_embd = self.find_hparam(["hidden_size", "n_embd"]) self.gguf_writer.add_embedding_length(n_embd) + print(f"gguf: embedding length = {n_embd}") if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None: self.gguf_writer.add_feed_forward_length(n_ff) + print(f"gguf: feed forward length = {n_ff}") n_head = self.find_hparam(["num_attention_heads", "n_head"]) self.gguf_writer.add_head_count(n_head) + print(f"gguf: head count = {n_head}") if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None: self.gguf_writer.add_head_count_kv(n_head_kv) + print(f"gguf: key-value head count = {n_head_kv}") if (rope_theta := self.hparams.get("rope_theta")) is not None: self.gguf_writer.add_rope_freq_base(rope_theta) + print(f"gguf: rope theta = {rope_theta}") if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None: self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps) + print(f"gguf: rms norm epsilon = {f_rms_eps}") if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None: self.gguf_writer.add_layer_norm_eps(f_norm_eps) + print(f"gguf: layer norm epsilon = {f_norm_eps}") if (n_experts := self.hparams.get("num_local_experts")) is not None: self.gguf_writer.add_expert_count(n_experts) + print(f"gguf: expert count = {n_experts}") if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None: self.gguf_writer.add_expert_used_count(n_experts_used) + print(f"gguf: experts used count = {n_experts_used}") self.gguf_writer.add_file_type(self.ftype) + print(f"gguf: file type = {self.ftype}") def write_tensors(self): block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer"))) @@ -1051,6 +1062,21 @@ class MixtralModel(Model): self._set_vocab_sentencepiece() +@Model.register("GrokForCausalLM") +class GrokModel(Model): + model_arch = gguf.MODEL_ARCH.GROK + + def set_vocab(self): + self._set_vocab_sentencepiece() + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def set_gguf_parameters(self): + super().set_gguf_parameters() + self.gguf_writer.add_name("Grok") + + @Model.register("MiniCPMForCausalLM") class MiniCPMModel(Model): model_arch = gguf.MODEL_ARCH.MINICPM diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index b59cc65bf..76496bf06 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -34,6 +34,7 @@ else() add_subdirectory(perplexity) add_subdirectory(quantize) add_subdirectory(quantize-stats) + add_subdirectory(retrieval) add_subdirectory(save-load-state) add_subdirectory(simple) add_subdirectory(passkey) diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp index ee1f8f1bf..7aaf63ceb 100644 --- a/examples/batched/batched.cpp +++ b/examples/batched/batched.cpp @@ -48,6 +48,8 @@ int main(int argc, char ** argv) { params.prompt = "Hello my name is"; } + process_escapes(params.prompt); + // init LLM llama_backend_init(); @@ -78,7 +80,7 @@ int main(int argc, char ** argv) { llama_context_params ctx_params = llama_context_default_params(); ctx_params.seed = 1234; - ctx_params.n_ctx = n_kv_req; + ctx_params.n_ctx = n_kv_req; ctx_params.n_batch = std::max(n_len, n_parallel); ctx_params.n_seq_max = n_parallel; ctx_params.n_threads = params.n_threads; diff --git a/examples/convert-llama2c-to-ggml/README.md b/examples/convert-llama2c-to-ggml/README.md index 0f37d295b..6da2d7e18 100644 --- a/examples/convert-llama2c-to-ggml/README.md +++ b/examples/convert-llama2c-to-ggml/README.md @@ -21,6 +21,8 @@ An example command using a model from [karpathy/tinyllamas](https://huggingface. `$ ./convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin` +Note: The vocabulary for `stories260K.bin` should be its own tokenizer `tok512.bin` found in [karpathy/tinyllamas/stories260K](https://huggingface.co/karpathy/tinyllamas/tree/main/stories260K). + Now you can use the model with a command like: `$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256` diff --git a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp index 8209dcb64..6b5c66530 100644 --- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp +++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp @@ -1,6 +1,7 @@ #include "ggml.h" #include "llama.h" #include "common.h" +#include "log.h" #include #include @@ -78,111 +79,101 @@ typedef struct { struct TransformerWeights { // token embedding table - float* token_embedding_table; // (vocab_size, dim) + std::vector token_embedding_table; // (vocab_size, dim) // weights for rmsnorms - float* rms_att_weight; // (layer, dim) rmsnorm weights - float* rms_ffn_weight; // (layer, dim) + std::vector rms_att_weight; // (layer, dim) rmsnorm weights + std::vector rms_ffn_weight; // (layer, dim) // weights for matmuls - float* wq; // (layer, dim, dim) - float* wk; // (layer, dim, dim) - float* wv; // (layer, dim, dim) - float* wo; // (layer, dim, dim) + std::vector wq; // (layer, dim, dim) + std::vector wk; // (layer, dim, dim) + std::vector wv; // (layer, dim, dim) + std::vector wo; // (layer, dim, dim) // weights for ffn - float* w1; // (layer, hidden_dim, dim) - float* w2; // (layer, dim, hidden_dim) - float* w3; // (layer, hidden_dim, dim) + std::vector w1; // (layer, hidden_dim, dim) + std::vector w2; // (layer, dim, hidden_dim) + std::vector w3; // (layer, hidden_dim, dim) // final rmsnorm - float* rms_final_weight; // (dim,) + std::vector rms_final_weight; // (dim,) // freq_cis for RoPE relatively positional embeddings - // float* freq_cis_real; // (seq_len, dim/2) - // float* freq_cis_imag; // (seq_len, dim/2) + // std::vector freq_cis_real; // (seq_len, dim/2) + // std::vector freq_cis_imag; // (seq_len, dim/2) // (optional) classifier weights for the logits, on the last layer - float* wcls; - - ~TransformerWeights() { - delete[] token_embedding_table; - delete[] rms_att_weight; - delete[] rms_ffn_weight; - delete[] wq; - delete[] wk; - delete[] wv; - delete[] wo; - delete[] w1; - delete[] w2; - delete[] w3; - delete[] rms_final_weight; - delete[] wcls; - } + std::vector wcls; }; -static void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) { - // we calloc instead of malloc to keep valgrind happy - w->token_embedding_table = new float[p->vocab_size * p->dim](); - printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim); +static void alloc_weights(TransformerWeights * w, const Config * p, bool shared_weights) { + const int n_multiqueries = p->n_kv_heads <= 0 || p->n_kv_heads >= p->n_heads ? 1 : p->n_heads / p->n_kv_heads; + try { + w->token_embedding_table.resize(p->vocab_size * p->dim); + LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim); - w->rms_att_weight = new float[p->n_layers * p->dim](); - printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim); + w->rms_att_weight.resize(p->n_layers * p->dim); + LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim); - w->rms_ffn_weight = new float[p->n_layers * p->dim](); - printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim); + w->rms_ffn_weight.resize(p->n_layers * p->dim); + LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim); - w->wq = new float[p->n_layers * p->dim * p->dim](); - printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim); + w->wq.resize(p->n_layers * p->dim * p->dim); + LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim); - w->wk = new float[p->n_layers * p->dim * p->dim](); - printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim); + w->wk.resize(p->n_layers * p->dim * p->dim / n_multiqueries); + LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries); - w->wv = new float[p->n_layers * p->dim * p->dim](); - printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim); + w->wv.resize(p->n_layers * p->dim * p->dim / n_multiqueries); + LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries); - w->wo = new float[p->n_layers * p->dim * p->dim](); - printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim); + w->wo.resize(p->n_layers * p->dim * p->dim); + LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim); - w->w1 = new float[p->n_layers * p->hidden_dim * p->dim](); - printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim); + w->w1.resize(p->n_layers * p->hidden_dim * p->dim); + LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim); - w->w2 = new float[p->n_layers * p->hidden_dim * p->dim](); - printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim); + w->w2.resize(p->n_layers * p->hidden_dim * p->dim); + LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim); - w->w3 = new float[p->n_layers * p->hidden_dim * p->dim](); - printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim); + w->w3.resize(p->n_layers * p->hidden_dim * p->dim); + LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim); - w->rms_final_weight = new float[p->dim](); - printf("[%s:AK] Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim); + w->rms_final_weight.resize(p->dim); + LOG("%s: Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim); - if (shared_weights) { - w->wcls = NULL; - } else { - w->wcls = new float[p->vocab_size * p->dim](); - printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim); + if (shared_weights) { + w->wcls = {}; + } else { + w->wcls.resize(p->vocab_size * p->dim); + LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim); + } + } + catch (std::length_error &) { + die("Invalid configuration. Failed to allocate memory for weights"); } } -static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) { - if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast(p->vocab_size * p->dim)) return 1; - if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast(p->n_layers * p->dim)) return 1; - if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast(p->n_layers * p->dim * p->dim)) return 1; - if (fread(w->wk, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast(p->n_layers * p->dim * p->dim)) return 1; - if (fread(w->wv, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast(p->n_layers * p->dim * p->dim)) return 1; - if (fread(w->wo, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast(p->n_layers * p->dim * p->dim)) return 1; - if (fread(w->rms_ffn_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast(p->n_layers * p->dim)) return 1; - if (fread(w->w1, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast(p->n_layers * p->dim * p->hidden_dim)) return 1; - if (fread(w->w2, sizeof(float), p->n_layers * p->hidden_dim * p->dim, f) != static_cast(p->n_layers * p->hidden_dim * p->dim)) return 1; - if (fread(w->w3, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast(p->n_layers * p->dim * p->hidden_dim)) return 1; - if (fread(w->rms_final_weight, sizeof(float), p->dim, f) != static_cast(p->dim)) return 1; +static int checkpoint_init_weights(TransformerWeights * w, const Config * p, FILE * f, bool shared_weights) { + if (fread(w->token_embedding_table.data(), sizeof(float), w->token_embedding_table.size(), f) != w->token_embedding_table.size()) return 1; + if (fread(w->rms_att_weight.data(), sizeof(float), w->rms_att_weight.size(), f) != w->rms_att_weight.size()) return 1; + if (fread(w->wq.data(), sizeof(float), w->wq.size(), f) != w->wq.size()) return 1; + if (fread(w->wk.data(), sizeof(float), w->wk.size(), f) != w->wk.size()) return 1; + if (fread(w->wv.data(), sizeof(float), w->wv.size(), f) != w->wv.size()) return 1; + if (fread(w->wo.data(), sizeof(float), w->wo.size(), f) != w->wo.size()) return 1; + if (fread(w->rms_ffn_weight.data(), sizeof(float), w->rms_ffn_weight.size(), f) != w->rms_ffn_weight.size()) return 1; + if (fread(w->w1.data(), sizeof(float), w->w1.size(), f) != w->w1.size()) return 1; + if (fread(w->w2.data(), sizeof(float), w->w2.size(), f) != w->w2.size()) return 1; + if (fread(w->w3.data(), sizeof(float), w->w3.size(), f) != w->w3.size()) return 1; + if (fread(w->rms_final_weight.data(), sizeof(float), w->rms_final_weight.size(), f) != w->rms_final_weight.size()) return 1; // Skip freq_cis_real & freq_cis_imag int head_size = p->dim / p->n_heads; fseek(f, p->seq_len * head_size * sizeof(float), SEEK_CUR); - if (!shared_weights && fread(w->wcls, sizeof(float), p->vocab_size * p->dim, f) != static_cast(p->vocab_size * p->dim)) return 1; + if (!shared_weights && fread(w->wcls.data(), sizeof(float), w->wcls.size(), f) != w->wcls.size()) return 1; // Check we didn't forget to read anything auto curr = ftell(f); fseek(f, 0, SEEK_END); auto end = ftell(f); if (curr != end) { - printf("Error: failed to read the checkpoint file to the end (curr = %ld, end = %ld)\n", curr, end); + LOG("%s: Error: failed to read the checkpoint file to the end (curr = %ld, end = %ld)\n", __func__, curr, end); return 1; } @@ -190,20 +181,20 @@ static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bo } static void print_sample_weights(TransformerWeights *w){ - printf("----- Quick print of first of the weight vales of all the variables\n"); - printf("%f\n", w->token_embedding_table[0]); - printf("%f\n", w->rms_att_weight[0]); - printf("%f\n", w->rms_ffn_weight[0]); + LOG("----- Quick print of first of the weight vales of all the variables\n"); + LOG("%f\n", w->token_embedding_table[0]); + LOG("%f\n", w->rms_att_weight[0]); + LOG("%f\n", w->rms_ffn_weight[0]); - printf("%f\n", w->wq[0]); - printf("%f\n", w->wk[0]); - printf("%f\n", w->wv[0]); - printf("%f\n", w->wo[0]); - printf("%f\n", w->w1[0]); - printf("%f\n", w->w2[0]); - printf("%f\n", w->w3[0]); - printf("%f\n", w->rms_att_weight[0]); - if (w->wcls) printf("%f\n", w->wcls[0]); + LOG("%f\n", w->wq[0]); + LOG("%f\n", w->wk[0]); + LOG("%f\n", w->wv[0]); + LOG("%f\n", w->wo[0]); + LOG("%f\n", w->w1[0]); + LOG("%f\n", w->w2[0]); + LOG("%f\n", w->w3[0]); + LOG("%f\n", w->rms_att_weight[0]); + if (!w->wcls.empty()) LOG("%f\n", w->wcls[0]); } //////////////////////////////////////////////////////////////////////////////////////////////////////////// @@ -225,14 +216,16 @@ struct llama_vocab { }; struct my_llama_hparams { - uint32_t n_vocab = 32000; - uint32_t n_ctx = 512; // this is provided as user input? - uint32_t n_embd = 4096; - uint32_t n_ff = 11008; - uint32_t n_mult = 4; - uint32_t n_head = 32; - uint32_t n_layer = 32; - uint32_t n_rot = 64; + uint32_t n_vocab = 32000; + uint32_t n_ctx = 512; // this is provided as user input? + uint32_t n_embd = 4096; + uint32_t n_ff = 11008; + uint32_t n_mult = 4; + uint32_t n_head = 32; + uint32_t n_head_kv = 32; + uint32_t n_layer = 32; + uint32_t n_rot = 64; + bool operator!=(const my_llama_hparams& other) const { return memcmp(this, &other, sizeof(my_llama_hparams)); } @@ -325,14 +318,30 @@ struct train_params { }; static void print_params(struct my_llama_hparams * params) { - printf("%s: n_vocab: %u\n", __func__, params->n_vocab); - printf("%s: n_ctx: %u\n", __func__, params->n_ctx); - printf("%s: n_embd: %u\n", __func__, params->n_embd); - printf("%s: n_mult: %u\n", __func__, params->n_mult); - printf("%s: n_head: %u\n", __func__, params->n_head); - printf("%s: n_ff: %u\n", __func__, params->n_ff); - printf("%s: n_layer: %u\n", __func__, params->n_layer); - printf("%s: n_rot: %u\n", __func__, params->n_rot); + LOG("%s: n_vocab: %u\n", __func__, params->n_vocab); + LOG("%s: n_ctx: %u\n", __func__, params->n_ctx); + LOG("%s: n_embd: %u\n", __func__, params->n_embd); + LOG("%s: n_mult: %u\n", __func__, params->n_mult); + LOG("%s: n_head: %u\n", __func__, params->n_head); + LOG("%s: n_head_kv: %u\n", __func__, params->n_head_kv); + LOG("%s: n_ff: %u\n", __func__, params->n_ff); + LOG("%s: n_layer: %u\n", __func__, params->n_layer); + LOG("%s: n_rot: %u\n", __func__, params->n_rot); +} + +static void print_tensor_info(const struct ggml_context * ctx) { + for (auto t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) { + LOG("%s: Allocating ", __func__); + int64_t total = 1; + int i = 0; + for (; i < ggml_n_dims(t); ++i) { + if (i > 0) LOG("x "); + LOG("[%" PRId64 "] ", t->ne[i]); + total *= t->ne[i]; + } + if (i > 1) LOG("= [%" PRId64 "] ", total); + LOG("float space for %s\n", ggml_get_name(t)); + } } static void init_model(struct my_llama_model * model) { @@ -342,6 +351,8 @@ static void init_model(struct my_llama_model * model) { const uint32_t n_layer = hparams.n_layer; const uint32_t n_vocab = hparams.n_vocab; + const uint32_t n_multiqueries = hparams.n_head_kv <= 0 || hparams.n_head_kv >= hparams.n_head ? 1 : hparams.n_head / hparams.n_head_kv; + const uint32_t n_ff = hparams.n_ff; struct ggml_context * ctx = model->ctx; @@ -350,25 +361,8 @@ static void init_model(struct my_llama_model * model) { model->train_tokens = 0; model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab); - printf("[%s:GG] Allocating [%u] x [%u] = [%u] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab); - model->norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); - printf("[%s:GG] Allocating [%u] float space for model->norm\n",__func__,n_embd); - model->output = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab); - printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab); - - // printing the per-layer allocations here so we dont print in the for loop. - printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wq for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer); - printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wk for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer); - printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wv for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer); - printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wo for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer); - - printf("[%s:GG] Allocating [%u] float space for layer.ffn_norm for [%u] layers\n",__func__,n_embd, n_layer); - - printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w1 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer); - printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w2 for [%u] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer); - printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w3 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer); ggml_set_name(model->tok_embeddings, "tok_embeddings.weight"); ggml_set_name(model->norm, "norm.weight"); @@ -383,8 +377,8 @@ static void init_model(struct my_llama_model * model) { layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); layer.wq = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd); - layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd); - layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd); + layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd / n_multiqueries); + layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd / n_multiqueries); layer.wo = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd); layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); @@ -406,6 +400,8 @@ static void init_model(struct my_llama_model * model) { ggml_format_name(layer.w2, "%s.feed_forward.w2.weight", layers_i.c_str()); ggml_format_name(layer.w3, "%s.feed_forward.w3.weight", layers_i.c_str()); } + + print_tensor_info(ctx); } static float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) { @@ -421,9 +417,9 @@ static int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) { static void print_row(struct ggml_tensor * probs, int i) { for (int k = 0; k < probs->ne[0]; ++k) { float p = get_f32_2d(probs, k, i); - printf(" %f", p); + LOG(" %f", p); } - printf("\n"); + LOG("\n"); } static void print_matrix(struct ggml_tensor * probs) { @@ -431,33 +427,12 @@ static void print_matrix(struct ggml_tensor * probs) { for (int i = 0; i < probs->ne[1]; ++i) { for (int k = 0; k < probs->ne[0]; ++k) { float p = get_f32_2d(probs, k, i); - printf(" %.2f", p); + LOG(" %.2f", p); } - printf("\n"); + LOG("\n"); } } -#ifdef __GNUC__ -#ifdef __MINGW32__ -__attribute__((format(gnu_printf, 1, 2))) -#else -__attribute__((format(printf, 1, 2))) -#endif -#endif -static std::string format(const char * fmt, ...) { - va_list ap, ap2; - va_start(ap, fmt); - va_copy(ap2, ap); - int size = vsnprintf(NULL, 0, fmt, ap); - GGML_ASSERT(size >= 0 && size < INT_MAX); - std::vector buf(size + 1); - int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2); - GGML_ASSERT(size2 == size); - va_end(ap2); - va_end(ap); - return std::string(buf.data(), size); -} - struct llama_file { // use FILE * so we don't have to re-open the file to mmap FILE * fp; @@ -549,8 +524,9 @@ static std::string llama_escape_whitespaces(const std::string & text) { return out.str(); } -static void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) { +static void load_vocab(const char * filename, const Config * config, struct llama_vocab * vocab) { if (is_ggml_file(filename)) { + LOG("%s: Loading vocabulary from gguf file %s\n", __func__, filename); struct ggml_context * ctx_data = NULL; struct gguf_init_params params = { @@ -578,6 +554,9 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx); const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx); + if (n_vocab != static_cast(config->vocab_size)) { + die_fmt("vocab size mismatch: (gguf) %u != (llama2c) %d", n_vocab, config->vocab_size); + } vocab->id_to_token.resize(n_vocab); @@ -595,7 +574,7 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab gguf_free(ctx); } else { // assume llama2.c vocabulary - printf("Assuming llama2.c vocabulary since %s is not a gguf file\n", filename); + LOG("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename); llama_file file(filename, "rb"); if (!file.fp) { die_fmt("%s: %s", strerror(errno), filename); @@ -638,38 +617,15 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab } static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) { - int ct; - switch (ggml_n_dims(gg_weights)) { - case 1: - ct = 0; - for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){ - float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0]); - *ptr = karpathy_weights[ct]; - ct++; - } - break; - case 2: - ct = 0; - for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) { - for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) { - float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1]); - *ptr = karpathy_weights[ct]; - ct++; - } - } - break; - case 3: - ct = 0; - for (int i2 = 0; i2 < gg_weights->ne[2]; i2++) { - for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) { - for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) { - float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1] + i2*gg_weights->nb[2]); - *ptr = karpathy_weights[ct]; - ct++; - } - } - } - break; + int size = 1; + for (int dim = 0; dim < ggml_n_dims(gg_weights); ++dim) { + size *= gg_weights->ne[dim]; + } + for (int ct = 0; ct < size; ++ct) { + int64_t i0 = 0; int64_t i1 = 0; + int64_t i2 = 0; int64_t i3 = 0; + ggml_unravel_index(gg_weights, ct, &i0, &i1, &i2, &i3); + ggml_set_f32_nd(gg_weights, i0, i1, i2, i3, karpathy_weights[ct]); } } @@ -679,16 +635,18 @@ static void save_as_llama_model( // convert AK weights into GG weights one by one. // w->token_embedding_table -> model->tok_embeddings // float* -> struct ggml_tensor - convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table); - convert_weights_ak_to_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table); + convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table.data()); + convert_weights_ak_to_gg(model->output, !w->wcls.empty() ? w->wcls.data() : w->token_embedding_table.data()); - convert_weights_ak_to_gg(model->norm, w->rms_final_weight); + convert_weights_ak_to_gg(model->norm, w->rms_final_weight.data()); //print_row(model->norm, 0); // for rms-att-weight int row_length = model->hparams.n_embd; int n_ff = model->hparams.n_ff; + const uint32_t n_multiqueries = model->hparams.n_head_kv <= 0 || model->hparams.n_head_kv >= model->hparams.n_head ? 1 : model->hparams.n_head / model->hparams.n_head_kv; + for (uint32_t i = 0; i < model->hparams.n_layer; ++i){ auto & layer = model->layers[i]; // 1d @@ -697,9 +655,10 @@ static void save_as_llama_model( // from 3d matrix layer x dim x dim to 2d matrix dim x dim convert_weights_ak_to_gg(layer.wq , &w->wq[i*row_length*row_length]); - convert_weights_ak_to_gg(layer.wk , &w->wk[i*row_length*row_length]); - convert_weights_ak_to_gg(layer.wv , &w->wv[i*row_length*row_length]); convert_weights_ak_to_gg(layer.wo , &w->wo[i*row_length*row_length]); + // from 3d matrix layer x dim x dim to 2d matrix dim x dim / n_multiqueries + convert_weights_ak_to_gg(layer.wk , &w->wk[i*row_length*row_length/n_multiqueries]); + convert_weights_ak_to_gg(layer.wv , &w->wv[i*row_length*row_length/n_multiqueries]); convert_weights_ak_to_gg(layer.w1 , &w->w1[i*row_length*n_ff]); convert_weights_ak_to_gg(layer.w2 , &w->w2[i*n_ff*row_length]); @@ -736,8 +695,8 @@ static void save_as_llama_model( gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd); gguf_set_val_u32(ctx, KV_FEED_FORWARD_LENGTH, model->hparams.n_ff); gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head); - // n_head_kv is optional, default to n_head - // gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, ...); + gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head); + gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, model->hparams.n_head_kv); gguf_set_val_u32(ctx, KV_BLOCK_COUNT, model->hparams.n_layer); gguf_set_val_u32(ctx, KV_ROPE_DIMENSION_COUNT, model->hparams.n_rot); gguf_set_val_f32(ctx, KV_ATTENTION_LAYERNORM_RMS_EPS, 1e-5f); @@ -789,12 +748,12 @@ static void save_as_llama_model( static struct train_params get_default_train_params() { struct train_params params; - params.fn_vocab_model = "models/7B/ggml-model-f16.gguf"; + params.fn_vocab_model = "models/7B/ggml-model-f16.gguf"; params.fn_llama2c_output_model = "ak_llama_model.bin"; - params.fn_train_data = "shakespeare.txt"; - params.fn_checkpoint_in = "checkpoint.bin"; - params.fn_checkpoint_out = "checkpoint.bin"; - params.fn_model_out = "ggml-checkpoint-f32.bin"; + params.fn_train_data = "shakespeare.txt"; + params.fn_checkpoint_in = "checkpoint.bin"; + params.fn_checkpoint_out = "checkpoint.bin"; + params.fn_model_out = "ggml-checkpoint-f32.bin"; params.seed = -1; @@ -829,8 +788,8 @@ static struct train_params get_default_train_params() { params.adam_alpha = 1e-3f; params.adam_decay = 1e-3f; - params.mem_model_gb = 2; - params.mem_compute_gb = 24; + params.mem_model_gb = 2; + params.mem_compute_gb = 24; params.mem_compute0_gb = 8; params.mem_compute1_gb = 2; @@ -916,19 +875,30 @@ int main(int argc, char ** argv) { if (!params_parse(argc, argv, ¶ms)) { return 1; } + log_set_target(stdout); Config config; TransformerWeights weights = {}; { - FILE *file = fopen(params.fn_llama2c_model, "rb"); - if (!file) { printf("Unable to open the checkpoint file %s!\n", params.fn_llama2c_model); return 1; } + LOG("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model); + FILE *file = fopen(params.fn_llama2c_model, "r"); + if (!file) { + LOG("%s: Unable to open the checkpoint file %s!\n", __func__, params.fn_llama2c_model); + return 1; + } // read in the config header - if(fread(&config, sizeof(Config), 1, file) != 1) { return 1; } + if (fread(&config, sizeof(Config), 1, file) != 1) { + LOG("%s: Unable to read llama2c config from %s!\n",__func__,params.fn_llama2c_model); + return 1; + } auto shared_weights = config.vocab_size > 0; config.vocab_size = abs(config.vocab_size); // read in the Transformer weights - malloc_weights(&weights, &config, shared_weights); - if(checkpoint_init_weights(&weights, &config, file, shared_weights)) { return 1; } + alloc_weights(&weights, &config, shared_weights); + if (checkpoint_init_weights(&weights, &config, file, shared_weights)) { + LOG("%s: Unable to initialize transformer weights from %s!",__func__,params.fn_llama2c_model); + return 1; + } fclose(file); } @@ -936,15 +906,18 @@ int main(int argc, char ** argv) { load_vocab(params.fn_vocab_model, &config, &vocab); struct my_llama_model model; - model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx); - model.hparams.n_ctx = params.n_ctx; - model.hparams.n_embd = config.dim; //params.n_embd; - model.hparams.n_ff = config.hidden_dim; - model.hparams.n_mult = 32;//params.n_mult; - model.hparams.n_head = config.n_heads; //params.n_head; - model.hparams.n_layer = config.n_layers; //params.n_layer; - model.hparams.n_rot = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head); + model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx); + model.hparams.n_ctx = params.n_ctx; + model.hparams.n_embd = config.dim; //params.n_embd; + model.hparams.n_ff = config.hidden_dim; + model.hparams.n_mult = 32;//params.n_mult; + model.hparams.n_head = config.n_heads; //params.n_head; + model.hparams.n_head_kv = config.n_kv_heads; + model.hparams.n_layer = config.n_layers; //params.n_layer; + model.hparams.n_rot = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head); + print_params(&model.hparams); + struct ggml_init_params lcparams; lcparams.mem_size = 1024ll*1024ll*1024ll*((size_t) params.mem_model_gb); lcparams.mem_buffer = NULL; @@ -956,7 +929,7 @@ int main(int argc, char ** argv) { model.name = basename(params.fn_llama2c_model); save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model); - printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model); + LOG("%s: Saving llama.c model file %s in ggml format at %s\n", __func__, params.fn_llama2c_model, params.fn_llama2c_output_model); ggml_free(model.ctx); return 0; diff --git a/examples/gguf-split/gguf-split.cpp b/examples/gguf-split/gguf-split.cpp index 8e12e6493..b1af59992 100644 --- a/examples/gguf-split/gguf-split.cpp +++ b/examples/gguf-split/gguf-split.cpp @@ -1,32 +1,31 @@ #include "llama.h" -#include "ggml.h" #include "common.h" #include #include -#include #include #include -#include #include #include #include -#include #include +#include +#include + +#if defined(_WIN32) + #include + #ifndef PATH_MAX + #define PATH_MAX MAX_PATH + #endif + #include +#endif enum split_operation : uint8_t { SPLIT_OP_SPLIT, SPLIT_OP_MERGE, }; -static const char * const LLM_KV_GENERAL_SPLIT_I_SPLIT = "general.split"; -static const char * const LLM_KV_GENERAL_SPLIT_N_SPLIT = "general.split_count"; - -static const int SPLIT_FILENAME_MAX = 256; - -static const char * const SPLIT_FILENAME_FORMAT = "%s-%05d-of-%05d.gguf"; - struct split_params { split_operation operation = SPLIT_OP_SPLIT; int n_split_tensors = 128; @@ -116,13 +115,13 @@ static bool split_params_parse(int argc, const char ** argv, split_params & para try { if (!split_params_parse_ex(argc, argv, params)) { split_print_usage(argv[0]); - exit(1); + exit(EXIT_FAILURE); } } catch (const std::invalid_argument & ex) { fprintf(stderr, "%s\n", ex.what()); split_print_usage(argv[0]); - exit(1); + exit(EXIT_FAILURE); } return result; } @@ -134,12 +133,6 @@ static void zeros(std::ofstream & file, size_t n) { } } -static std::string split_file_name(const std::string & path, int i_split, int n_split) { - char f_split[SPLIT_FILENAME_MAX] = {0}; - snprintf(f_split, sizeof(f_split), SPLIT_FILENAME_FORMAT, path.c_str(), i_split + 1, n_split); - return std::string(f_split); -} - struct split_strategy { const split_params params; std::ifstream & f_input; @@ -180,8 +173,9 @@ struct split_strategy { if (i_split == 0) { gguf_set_kv(ctx_out, ctx_gguf); } - gguf_set_val_u8(ctx_out, LLM_KV_GENERAL_SPLIT_I_SPLIT, i_split); - gguf_set_val_u8(ctx_out, LLM_KV_GENERAL_SPLIT_N_SPLIT, n_split); + gguf_set_val_u16(ctx_out, LLM_KV_SPLIT_NO, i_split); + gguf_set_val_u16(ctx_out, LLM_KV_SPLIT_COUNT, n_split); + gguf_set_val_i32(ctx_out, LLM_KV_SPLIT_TENSORS_COUNT, n_tensors); // populate the original tensors, so we get an initial metadata for (int i = i_split * params.n_split_tensors; i < n_tensors && i < (i_split + 1) * params.n_split_tensors; ++i) { @@ -189,10 +183,11 @@ struct split_strategy { gguf_add_tensor(ctx_out, meta); } - auto split_name = split_file_name(params.output, i_split, n_split); + char split_path[PATH_MAX] = {0}; + llama_split_path(split_path, sizeof(split_path), params.output.c_str(), i_split, n_split); - fprintf(stderr, "%s: %s ...", __func__, split_name.c_str()); - fout = std::ofstream(split_name, std::ios::binary); + fprintf(stderr, "%s: %s ...", __func__, split_path); + fout = std::ofstream(split_path, std::ios::binary); fout.exceptions(std::ofstream::failbit); // fail fast on write errors auto meta_size = gguf_get_meta_size(ctx_out); @@ -250,19 +245,23 @@ static void gguf_split(const split_params & split_params) { std::ifstream f_input(split_params.input.c_str(), std::ios::binary); if (!f_input.is_open()) { fprintf(stderr, "%s: failed to open input GGUF from %s\n", __func__, split_params.input.c_str()); - exit(1); + exit(EXIT_FAILURE); } auto * ctx_gguf = gguf_init_from_file(split_params.input.c_str(), params); if (!ctx_gguf) { fprintf(stderr, "%s: failed to load input GGUF from %s\n", __func__, split_params.input.c_str()); - exit(1); + exit(EXIT_FAILURE); } split_strategy strategy(split_params, f_input, ctx_gguf, ctx_meta); + + char first_split_path[PATH_MAX] = {0}; + llama_split_path(first_split_path, sizeof(first_split_path), + split_params.output.c_str(), strategy.i_split, strategy.n_split); fprintf(stderr, "%s: %s -> %s (%d tensors per file)\n", __func__, split_params.input.c_str(), - split_file_name(split_params.output, strategy.i_split, strategy.n_split).c_str(), + first_split_path, split_params.n_split_tensors); strategy.split_start(); @@ -298,7 +297,9 @@ static void gguf_merge(const split_params & split_params) { std::vector ctx_metas; std::vector ctx_ggufs; - std::string split_prefix; + char split_path[PATH_MAX] = {0}; + strncpy(split_path, split_params.input.c_str(), sizeof(split_path) - 1); + char split_prefix[PATH_MAX] = {0}; // First pass to find KV and tensors metadata for (int i_split = 0; i_split < n_split; i_split++) { @@ -309,89 +310,66 @@ static void gguf_merge(const split_params & split_params) { /*.ctx = */ &ctx_meta, }; - auto split_name = split_params.input; if (i_split > 0) { - split_name = split_file_name(split_prefix, i_split, n_split); + llama_split_path(split_path, sizeof(split_path), split_prefix, i_split, n_split); } - fprintf(stderr, "%s: reading metadata %s ...", __func__, split_name.c_str()); + fprintf(stderr, "%s: reading metadata %s ...", __func__, split_path); - auto * ctx_gguf = gguf_init_from_file(split_name.c_str(), params); + auto * ctx_gguf = gguf_init_from_file(split_path, params); if (!ctx_gguf) { fprintf(stderr, "\n%s: failed to load input GGUF from %s\n", __func__, split_params.input.c_str()); - exit(1); + exit(EXIT_FAILURE); } ctx_ggufs.push_back(ctx_gguf); ctx_metas.push_back(ctx_meta); if (i_split == 0) { - auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_GENERAL_SPLIT_N_SPLIT); + auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT); if (key_n_split < 0) { fprintf(stderr, "\n%s: input file does not contain %s metadata\n", __func__, - LLM_KV_GENERAL_SPLIT_N_SPLIT); + LLM_KV_SPLIT_COUNT); gguf_free(ctx_gguf); + ggml_free(ctx_meta); gguf_free(ctx_out); fout.close(); - exit(1); + exit(EXIT_FAILURE); } - n_split = gguf_get_val_u8(ctx_gguf, key_n_split); + n_split = gguf_get_val_u16(ctx_gguf, key_n_split); if (n_split < 1) { fprintf(stderr, "\n%s: input file does not contain a valid split count %d\n", __func__, n_split); gguf_free(ctx_gguf); + ggml_free(ctx_meta); gguf_free(ctx_out); fout.close(); - exit(1); + exit(EXIT_FAILURE); + } + + // Verify the file naming and extract split_prefix + if (!llama_split_prefix(split_prefix, sizeof (split_prefix), split_path, i_split, n_split)) { + fprintf(stderr, "\n%s: unexpected input file name: %s" + " i_split=%d" + " n_split=%d\n", __func__, + split_path, i_split, n_split); + gguf_free(ctx_gguf); + ggml_free(ctx_meta); + gguf_free(ctx_out); + fout.close(); + exit(EXIT_FAILURE); } // Do not trigger merge if we try to merge again the output - gguf_set_val_u8(ctx_out, LLM_KV_GENERAL_SPLIT_N_SPLIT, 0); + gguf_set_val_u16(ctx_gguf, LLM_KV_SPLIT_COUNT, 0); // Set metadata from the first split gguf_set_kv(ctx_out, ctx_gguf); } - // Verify the file naming - { - int i_split_file = 0; - int n_split_file = 0; - const char * i_split_format = "-00000-of-00000.gguf"; - - if (split_name.size() < strlen(i_split_format)) { - fprintf(stderr, "\n%s: unexpected input file name: %s\n", __func__, split_params.input.c_str()); - for (auto * _ctx_gguf : ctx_ggufs) { - gguf_free(_ctx_gguf); - } - gguf_free(ctx_out); - fout.close(); - exit(1); - } - - split_prefix = split_name.substr(0, split_name.size() - strlen(i_split_format)); - - const char * split_name_c_str = split_name.c_str(); - int n_part = sscanf(&split_name_c_str[0] + split_prefix.size(), "-%d-of-%d", &i_split_file, &n_split_file); - - if (n_part != 2 || i_split_file - 1 != i_split || n_split_file != n_split) { - fprintf(stderr, "\n%s: unexpected input file name: %s" - " i_split=%d i_split_file=%d" - " n_split=%d n_split_file=%d\n", __func__, - split_params.input.c_str(), - i_split, i_split_file, - n_split, n_split_file); - for (auto * _ctx_gguf : ctx_ggufs) { - gguf_free(_ctx_gguf); - } - gguf_free(ctx_out); - fout.close(); - exit(1); - } - } - auto n_tensors = gguf_get_n_tensors(ctx_gguf); for (int i_tensor = 0; i_tensor < n_tensors; i_tensor++) { const char * t_name = gguf_get_tensor_name(ctx_gguf, i_tensor); @@ -411,18 +389,19 @@ static void gguf_merge(const split_params & split_params) { // Write tensors data for (int i_split = 0; i_split < n_split; i_split++) { - auto split_name = split_file_name(split_prefix, i_split, n_split); - std::ifstream f_input(split_name.c_str(), std::ios::binary); + llama_split_path(split_path, sizeof(split_path), split_prefix, i_split, n_split); + std::ifstream f_input(split_path, std::ios::binary); if (!f_input.is_open()) { - fprintf(stderr, "%s: failed to open input GGUF from %s\n", __func__, split_name.c_str()); - for (auto * _ctx_gguf : ctx_ggufs) { - gguf_free(_ctx_gguf); + fprintf(stderr, "%s: failed to open input GGUF from %s\n", __func__, split_path); + for (uint32_t i = 0; i < ctx_ggufs.size(); i++) { + gguf_free(ctx_ggufs[i]); + ggml_free(ctx_metas[i]); } gguf_free(ctx_out); fout.close(); - exit(1); + exit(EXIT_FAILURE); } - fprintf(stderr, "%s: writing tensors %s ...", __func__, split_name.c_str()); + fprintf(stderr, "%s: writing tensors %s ...", __func__, split_path); auto * ctx_gguf = ctx_ggufs[i_split]; auto * ctx_meta = ctx_metas[i_split]; @@ -481,8 +460,8 @@ int main(int argc, const char ** argv) { break; case SPLIT_OP_MERGE: gguf_merge(params); break; - default:split_print_usage(argv[0]); - exit(1); + default: split_print_usage(argv[0]); + exit(EXIT_FAILURE); } return 0; diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp index ea79b9062..264e73f4e 100644 --- a/examples/imatrix/imatrix.cpp +++ b/examples/imatrix/imatrix.cpp @@ -50,29 +50,31 @@ private: void keep_imatrix(int ncall) const; }; +// remove any prefix and suffixes from the name +// CUDA0#blk.0.attn_k.weight#0 => blk.0.attn_k.weight +static std::string filter_tensor_name(const char * name) { + std::string wname; + const char * p = strchr(name, '#'); + if (p != NULL) { + p = p + 1; + const char * q = strchr(p, '#'); + if (q != NULL) { + wname = std::string(p, q - p); + } else { + wname = p; + } + } else { + wname = name; + } + return wname; +} + bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) { GGML_UNUSED(user_data); const struct ggml_tensor * src0 = t->src[0]; const struct ggml_tensor * src1 = t->src[1]; - - std::string wname; - { - // remove any prefix and suffixes from the name - // CUDA0#blk.0.attn_k.weight#0 => blk.0.attn_k.weight - const char * p = strchr(src0->name, '#'); - if (p != NULL) { - p = p + 1; - const char * q = strchr(p, '#'); - if (q != NULL) { - wname = std::string(p, q - p); - } else { - wname = p; - } - } else { - wname = src0->name; - } - } + std::string wname = filter_tensor_name(src0->name); // when ask is true, the scheduler wants to know if we are interested in data from this tensor // if we return true, a follow-up call will be made with ask=false in which we can do the actual collection @@ -112,6 +114,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * // this is necessary to guarantee equal number of "ncall" for each tensor for (int ex = 0; ex < n_as; ++ex) { src0 = t->src[2 + ex]; + wname = filter_tensor_name(src0->name); auto& e = m_stats[wname]; if (e.values.empty()) { e.values.resize(src1->ne[0], 0); diff --git a/examples/json-schema-pydantic-example.py b/examples/json-schema-pydantic-example.py new file mode 100644 index 000000000..69ebfd409 --- /dev/null +++ b/examples/json-schema-pydantic-example.py @@ -0,0 +1,74 @@ +# Usage: +#! ./server -m some-model.gguf & +#! pip install pydantic +#! python json-schema-pydantic-example.py + +from pydantic import BaseModel, TypeAdapter +from annotated_types import MinLen +from typing import Annotated, List, Optional +import json, requests + +if True: + + def create_completion(*, response_model=None, endpoint="http://localhost:8080/v1/chat/completions", messages, **kwargs): + ''' + Creates a chat completion using an OpenAI-compatible endpoint w/ JSON schema support + (llama.cpp server, llama-cpp-python, Anyscale / Together...) + + The response_model param takes a type (+ supports Pydantic) and behaves just as w/ Instructor (see below) + ''' + if response_model: + type_adapter = TypeAdapter(response_model) + schema = type_adapter.json_schema() + messages = [{ + "role": "system", + "content": f"You respond in JSON format with the following schema: {json.dumps(schema, indent=2)}" + }] + messages + response_format={"type": "json_object", "schema": schema} + + data = requests.post(endpoint, headers={"Content-Type": "application/json"}, + json=dict(messages=messages, response_format=response_format, **kwargs)).json() + if 'error' in data: + raise Exception(data['error']['message']) + + content = data["choices"][0]["message"]["content"] + return type_adapter.validate_json(content) if type_adapter else content + +else: + + # This alternative branch uses Instructor + OpenAI client lib. + # Instructor support streamed iterable responses, retry & more. + # (see https://python.useinstructor.com/) + #! pip install instructor openai + import instructor, openai + client = instructor.patch( + openai.OpenAI(api_key="123", base_url="http://localhost:8080"), + mode=instructor.Mode.JSON_SCHEMA) + create_completion = client.chat.completions.create + + +if __name__ == '__main__': + + class QAPair(BaseModel): + question: str + concise_answer: str + justification: str + + class PyramidalSummary(BaseModel): + title: str + summary: str + question_answers: Annotated[List[QAPair], MinLen(2)] + sub_sections: Optional[Annotated[List['PyramidalSummary'], MinLen(2)]] + + print("# Summary\n", create_completion( + model="...", + response_model=PyramidalSummary, + messages=[{ + "role": "user", + "content": f""" + You are a highly efficient corporate document summarizer. + Create a pyramidal summary of an imaginary internal document about our company processes + (starting high-level, going down to each sub sections). + Keep questions short, and answers even shorter (trivia / quizz style). + """ + }])) diff --git a/examples/json-schema-to-grammar.py b/examples/json-schema-to-grammar.py index 6a977f031..91dd734cc 100755 --- a/examples/json-schema-to-grammar.py +++ b/examples/json-schema-to-grammar.py @@ -1,8 +1,10 @@ #!/usr/bin/env python3 import argparse +import itertools import json import re import sys +from typing import Any, Dict, List, Set, Tuple, Union # whitespace is constrained to a single space char to prevent model "running away" in # whitespace. Also maybe improves generation quality? @@ -12,26 +14,54 @@ PRIMITIVE_RULES = { 'boolean': '("true" | "false") space', 'number': '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space', 'integer': '("-"? ([0-9] | [1-9] [0-9]*)) space', + 'value' : 'object | array | string | number | boolean', + 'object' : '"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', + 'array' : '"[" space ( value ("," space value)* )? "]" space', + 'uuid' : '"\\"" ' + ' "-" '.join('[0-9a-fA-F]' * n for n in [8, 4, 4, 4, 12]) + ' "\\"" space', 'string': r''' "\"" ( [^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) - )* "\"" space ''', + )* "\"" space''', 'null': '"null" space', } +OBJECT_RULE_NAMES = ['object', 'array', 'string', 'number', 'boolean', 'null', 'value'] + +# TODO: support "uri", "email" string formats +DATE_RULES = { + 'date' : '[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', + 'time' : '([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', + 'date-time': 'date "T" time', + 'date-string': '"\\"" date "\\"" space', + 'time-string': '"\\"" time "\\"" space', + 'date-time-string': '"\\"" date-time "\\"" space', +} + +RESERVED_NAMES = set(["root", *PRIMITIVE_RULES.keys(), *DATE_RULES.keys()]) INVALID_RULE_CHARS_RE = re.compile(r'[^a-zA-Z0-9-]+') GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]') -GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"'} +GRAMMAR_RANGE_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"\]\-\\]') +GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]'} +NON_LITERAL_SET = set('|.()[]{}*+?') +ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('[]()|{}*+?') + +DATE_PATTERN = '[0-9]{4}-(0[1-9]|1[0-2])-([0-2][0-9]|3[0-1])' +TIME_PATTERN = '([01][0-9]|2[0-3])(:[0-5][0-9]){2}(\\.[0-9]{1,3})?(Z|[+-](([01][0-9]|2[0-3]):[0-5][0-9]))' # Cap millisecond precision w/ 3 digits class SchemaConverter: - def __init__(self, prop_order): + def __init__(self, *, prop_order, allow_fetch, dotall, raw_pattern): self._prop_order = prop_order + self._allow_fetch = allow_fetch + self._dotall = dotall + self._raw_pattern = raw_pattern self._rules = {'space': SPACE_RULE} + self._refs = {} + self._refs_being_resolved = set() def _format_literal(self, literal): escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub( - lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), json.dumps(literal) + lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), literal ) return f'"{escaped}"' @@ -41,78 +71,420 @@ class SchemaConverter: key = esc_name else: i = 0 - while f'{esc_name}{i}' in self._rules: + while f'{esc_name}{i}' in self._rules and self._rules[f'{esc_name}{i}'] != rule: i += 1 key = f'{esc_name}{i}' self._rules[key] = rule return key + def resolve_refs(self, schema: dict, url: str): + ''' + Resolves all $ref fields in the given schema, fetching any remote schemas, + replacing $ref with absolute reference URL and populating self._refs with the + respective referenced (sub)schema dictionaries. + ''' + def visit(n: dict): + if isinstance(n, list): + return [visit(x) for x in n] + elif isinstance(n, dict): + ref = n.get('$ref') + if ref is not None and ref not in self._refs: + if ref.startswith('https://'): + assert self._allow_fetch, 'Fetching remote schemas is not allowed (use --allow-fetch for force)' + import requests + + frag_split = ref.split('#') + base_url = frag_split[0] + + target = self._refs.get(base_url) + if target is None: + target = self.resolve_refs(requests.get(ref).json(), base_url) + self._refs[base_url] = target + + if len(frag_split) == 1 or frag_split[-1] == '': + return target + elif ref.startswith('#/'): + target = schema + ref = f'{url}{ref}' + n['$ref'] = ref + else: + raise ValueError(f'Unsupported ref {ref}') + + for sel in ref.split('#')[-1].split('/')[1:]: + assert target is not None and sel in target, f'Error resolving ref {ref}: {sel} not in {target}' + target = target[sel] + + self._refs[ref] = target + else: + for v in n.values(): + visit(v) + + return n + return visit(schema) + + def _generate_union_rule(self, name, alt_schemas): + return ' | '.join(( + self.visit(alt_schema, f'{name}{"-" if name else "alternative-"}{i}') + for i, alt_schema in enumerate(alt_schemas) + )) + + def _visit_pattern(self, pattern, name): + ''' + Transforms a regular expression pattern into a GBNF rule. + + Input: https://json-schema.org/understanding-json-schema/reference/regular_expressions + Output: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md + + Unsupported features: negative/positive lookaheads, greedy/non-greedy modifiers. + + Mostly a 1:1 translation, except for {x} / {x,} / {x,y} quantifiers for which + we define sub-rules to keep the output lean. + ''' + + assert pattern.startswith('^') and pattern.endswith('$'), 'Pattern must start with "^" and end with "$"' + pattern = pattern[1:-1] + sub_rule_ids = {} + + i = 0 + length = len(pattern) + + def to_rule(s: Tuple[str, bool]) -> str: + (txt, is_literal) = s + return "\"" + txt + "\"" if is_literal else txt + + def transform() -> Tuple[str, bool]: + ''' + Parse a unit at index i (advancing it), and return its string representation + whether it's a literal. + ''' + nonlocal i + nonlocal pattern + nonlocal sub_rule_ids + + start = i + # For each component of this sequence, store its string representation and whether it's a literal. + # We only need a flat structure here to apply repetition operators to the last item, and + # to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially + # (GBNF's syntax is luckily very close to regular expressions!) + seq: list[Tuple[str, bool]] = [] + + def get_dot(): + if self._dotall: + rule = '[\\U00000000-\\U0010FFFF]' + else: + # Accept any character... except \n and \r line break chars (\x0A and \xOD) + rule = '[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]' + return self._add_rule(f'dot', rule) + + def join_seq(): + nonlocal seq + ret = [] + for is_literal, g in itertools.groupby(seq, lambda x: x[1]): + if is_literal: + ret.append((''.join(x[0] for x in g), True)) + else: + ret.extend(g) + if len(ret) == 1: + return ret[0] + return (' '.join(to_rule(x) for x in seq), False) + + while i < length: + c = pattern[i] + if c == '.': + seq.append((get_dot(), False)) + i += 1 + elif c == '(': + i += 1 + if i < length: + assert pattern[i] != '?', f'Unsupported pattern syntax "{pattern[i]}" at index {i} of /{pattern}/' + seq.append((f'({to_rule(transform())})', False)) + elif c == ')': + i += 1 + assert start > 0 and pattern[start-1] == '(', f'Unbalanced parentheses; start = {start}, i = {i}, pattern = {pattern}' + return join_seq() + elif c == '[': + square_brackets = c + i += 1 + while i < length and pattern[i] != ']': + if pattern[i] == '\\': + square_brackets += pattern[i:i+2] + i += 2 + else: + square_brackets += pattern[i] + i += 1 + assert i < length, f'Unbalanced square brackets; start = {start}, i = {i}, pattern = {pattern}' + square_brackets += ']' + i += 1 + seq.append((square_brackets, False)) + elif c == '|': + seq.append(('|', False)) + i += 1 + elif c in ('*', '+', '?'): + seq[-1] = (to_rule(seq[-1]) + c, False) + i += 1 + elif c == '{': + curly_brackets = c + i += 1 + while i < length and pattern[i] != '}': + curly_brackets += pattern[i] + i += 1 + assert i < length, f'Unbalanced curly brackets; start = {start}, i = {i}, pattern = {pattern}' + curly_brackets += '}' + i += 1 + nums = [s.strip() for s in curly_brackets[1:-1].split(',')] + min_times = 0 + max_times = None + try: + if len(nums) == 1: + min_times = int(nums[0]) + max_times = min_times + else: + assert len(nums) == 2 + min_times = int(nums[0]) if nums[0] else 0 + max_times = int(nums[1]) if nums[1] else None + except ValueError: + raise ValueError(f'Invalid quantifier {curly_brackets} in /{pattern}/') + + (sub, sub_is_literal) = seq[-1] + + if min_times == 0 and max_times is None: + seq[-1] = (f'{sub}*', False) + elif min_times == 0 and max_times == 1: + seq[-1] = (f'{sub}?', False) + elif min_times == 1 and max_times is None: + seq[-1] = (f'{sub}+', False) + else: + if not sub_is_literal: + id = sub_rule_ids.get(sub) + if id is None: + id = self._add_rule(f'{name}-{len(sub_rule_ids) + 1}', sub) + sub_rule_ids[sub] = id + sub = id + + seq[-1] = ( + ' '.join( + ([f'"{sub[1:-1] * min_times}"'] if sub_is_literal else [sub] * min_times) + + ([f'{sub}?'] * (max_times - min_times) if max_times is not None else [f'{sub}*'])), + False + ) + else: + literal = '' + while i < length: + if pattern[i] == '\\' and i < length - 1: + next = pattern[i + 1] + if next in ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS: + i += 1 + literal += pattern[i] + i += 1 + else: + literal += pattern[i:i+2] + i += 2 + elif pattern[i] == '"' and not self._raw_pattern: + literal += '\\"' + i += 1 + elif pattern[i] not in NON_LITERAL_SET and \ + (i == length - 1 or literal == '' or pattern[i+1] == '.' or pattern[i+1] not in NON_LITERAL_SET): + literal += pattern[i] + i += 1 + else: + break + if literal: + seq.append((literal, True)) + + return join_seq() + + return self._add_rule( + name, + to_rule(transform()) if self._raw_pattern \ + else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space") + + + def _resolve_ref(self, ref): + ref_name = ref.split('/')[-1] + if ref_name not in self._rules and ref not in self._refs_being_resolved: + self._refs_being_resolved.add(ref) + resolved = self._refs[ref] + ref_name = self.visit(resolved, ref_name) + self._refs_being_resolved.remove(ref) + return ref_name + + def _generate_constant_rule(self, value): + return self._format_literal(json.dumps(value)) + def visit(self, schema, name): schema_type = schema.get('type') - rule_name = name or 'root' + schema_format = schema.get('format') + rule_name = name + '-' if name in RESERVED_NAMES else name or 'root' - if 'oneOf' in schema or 'anyOf' in schema: - rule = ' | '.join(( - self.visit(alt_schema, f'{name}{"-" if name else ""}{i}') - for i, alt_schema in enumerate(schema.get('oneOf') or schema['anyOf']) - )) - return self._add_rule(rule_name, rule) + if (ref := schema.get('$ref')) is not None: + return self._add_rule(rule_name, self._resolve_ref(ref)) + + elif 'oneOf' in schema or 'anyOf' in schema: + return self._add_rule(rule_name, self._generate_union_rule(name, schema.get('oneOf') or schema['anyOf'])) + + elif isinstance(schema_type, list): + return self._add_rule(rule_name, self._generate_union_rule(name, [{'type': t} for t in schema_type])) elif 'const' in schema: - return self._add_rule(rule_name, self._format_literal(schema['const'])) + return self._add_rule(rule_name, self._generate_constant_rule(schema['const'])) elif 'enum' in schema: - rule = ' | '.join((self._format_literal(v) for v in schema['enum'])) + rule = ' | '.join((self._generate_constant_rule(v) for v in schema['enum'])) return self._add_rule(rule_name, rule) - elif schema_type == 'object' and 'properties' in schema: - # TODO: `required` keyword - prop_order = self._prop_order - prop_pairs = sorted( - schema['properties'].items(), - # sort by position in prop_order (if specified) then by key - key=lambda kv: (prop_order.get(kv[0], len(prop_order)), kv[0]), + elif schema_type in (None, 'object') and \ + ('properties' in schema or \ + ('additionalProperties' in schema and schema['additionalProperties'] is not True)): + required = set(schema.get('required', [])) + properties = list(schema.get('properties', {}).items()) + return self._add_rule(rule_name, self._build_object_rule(properties, required, name, schema.get('additionalProperties'))) + + elif schema_type in (None, 'object') and 'allOf' in schema: + required = set() + properties = [] + hybrid_name = name + def add_component(comp_schema, is_required): + if (ref := comp_schema.get('$ref')) is not None: + comp_schema = self._refs[ref] + + if 'properties' in comp_schema: + for prop_name, prop_schema in comp_schema['properties'].items(): + properties.append((prop_name, prop_schema)) + if is_required: + required.add(prop_name) + + for t in schema['allOf']: + if 'anyOf' in t: + for tt in t['anyOf']: + add_component(tt, is_required=False) + else: + add_component(t, is_required=True) + + return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=[])) + + elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema): + items = schema.get('items') or schema['prefixItems'] + if isinstance(items, list): + return self._add_rule( + rule_name, + '"[" space ' + + ' "," space '.join( + self.visit(item, f'{name}{"-" if name else ""}tuple-{i}') + for i, item in enumerate(items)) + + ' "]" space') + else: + item_rule_name = self.visit(items, f'{name}{"-" if name else ""}item') + list_item_operator = f'( "," space {item_rule_name} )' + successive_items = "" + min_items = schema.get("minItems", 0) + max_items = schema.get("maxItems") + if min_items > 0: + successive_items = list_item_operator * (min_items - 1) + min_items -= 1 + if max_items is not None and max_items > min_items: + successive_items += (list_item_operator + "?") * (max_items - min_items - 1) + else: + successive_items += list_item_operator + "*" + if min_items == 0: + rule = f'"[" space ( {item_rule_name} {successive_items} )? "]" space' + else: + rule = f'"[" space {item_rule_name} {successive_items} "]" space' + return self._add_rule(rule_name, rule) + + elif schema_type in (None, 'string') and 'pattern' in schema: + return self._visit_pattern(schema['pattern'], rule_name) + + elif schema_type in (None, 'string') and re.match(r'^uuid[1-5]?$', schema_format or ''): + return self._add_rule( + 'root' if rule_name == 'root' else schema_format, + PRIMITIVE_RULES['uuid'] ) - rule = '"{" space' - for i, (prop_name, prop_schema) in enumerate(prop_pairs): - prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}') - if i > 0: - rule += ' "," space' - rule += fr' {self._format_literal(prop_name)} space ":" space {prop_rule_name}' - rule += ' "}" space' + elif schema_type in (None, 'string') and schema_format in DATE_RULES: + for t, r in DATE_RULES.items(): + self._add_rule(t, r) + return schema_format + '-string' - return self._add_rule(rule_name, rule) - - elif schema_type == 'array' and 'items' in schema: - # TODO `prefixItems` keyword - item_rule_name = self.visit(schema['items'], f'{name}{"-" if name else ""}item') - list_item_operator = f'("," space {item_rule_name})' - successive_items = "" - min_items = schema.get("minItems", 0) - if min_items > 0: - first_item = f"({item_rule_name})" - successive_items = list_item_operator * (min_items - 1) - min_items -= 1 - else: - first_item = f"({item_rule_name})?" - max_items = schema.get("maxItems") - if max_items is not None and max_items > min_items: - successive_items += (list_item_operator + "?") * (max_items - min_items - 1) - else: - successive_items += list_item_operator + "*" - rule = f'"[" space {first_item} {successive_items} "]" space' - return self._add_rule(rule_name, rule) + elif (schema_type == 'object') or (len(schema) == 0): + for n in OBJECT_RULE_NAMES: + self._add_rule(n, PRIMITIVE_RULES[n]) + return self._add_rule(rule_name, 'object') else: assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}' + # TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero return self._add_rule( 'root' if rule_name == 'root' else schema_type, PRIMITIVE_RULES[schema_type] ) + def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[str], name: str, additional_properties: Union[bool, Any]): + prop_order = self._prop_order + # sort by position in prop_order (if specified) then by original order + sorted_props = [kv[0] for _, kv in sorted(enumerate(properties), key=lambda ikv: (prop_order.get(ikv[1][0], len(prop_order)), ikv[0]))] + + prop_kv_rule_names = {} + for prop_name, prop_schema in properties: + prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}') + prop_kv_rule_names[prop_name] = self._add_rule( + f'{name}{"-" if name else ""}{prop_name}-kv', + fr'{self._format_literal(json.dumps(prop_name))} space ":" space {prop_rule_name}' + ) + required_props = [k for k in sorted_props if k in required] + optional_props = [k for k in sorted_props if k not in required] + + if additional_properties == True or isinstance(additional_properties, dict): + sub_name = f'{name}{"-" if name else ""}additional' + value_rule = self.visit({} if additional_properties == True else additional_properties, f'{sub_name}-value') + prop_kv_rule_names["*"] = self._add_rule( + f'{sub_name}-kv', + self._add_rule('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}' + ) + optional_props.append("*") + + rule = '"{" space ' + rule += ' "," space '.join(prop_kv_rule_names[k] for k in required_props) + + if optional_props: + rule += ' (' + if required_props: + rule += ' "," space ( ' + + def get_recursive_refs(ks, first_is_optional): + [k, *rest] = ks + kv_rule_name = prop_kv_rule_names[k] + if k == '*': + res = self._add_rule( + f'{name}{"-" if name else ""}additional-kvs', + f'{kv_rule_name} ( "," space ' + kv_rule_name + ' )*' + ) + elif first_is_optional: + res = f'( "," space {kv_rule_name} )?' + else: + res = kv_rule_name + if len(rest) > 0: + res += ' ' + self._add_rule( + f'{name}{"-" if name else ""}{k}-rest', + get_recursive_refs(rest, first_is_optional=True) + ) + return res + + rule += ' | '.join( + get_recursive_refs(optional_props[i:], first_is_optional=False) + for i in range(len(optional_props)) + ) + if required_props: + rule += ' )' + rule += ' )?' + + rule += ' "}" space' + + return rule + def format_grammar(self): - return '\n'.join((f'{name} ::= {rule}' for name, rule in self._rules.items())) + return '\n'.join( + f'{name} ::= {rule}' + for name, rule in sorted(self._rules.items(), key=lambda kv: kv[0]) + ) def main(args_in = None): @@ -129,16 +501,47 @@ def main(args_in = None): type=lambda s: s.split(','), help=''' comma-separated property names defining the order of precedence for object properties; - properties not specified here are given lower precedence than those that are, and are - sorted alphabetically + properties not specified here are given lower precedence than those that are, and + are kept in their original order from the schema. Required properties are always + given precedence over optional properties. ''' ) + parser.add_argument( + '--allow-fetch', + action='store_true', + default=False, + help='Whether to allow fetching referenced schemas over HTTPS') + parser.add_argument( + '--dotall', + action='store_true', + default=False, + help='Whether to treat dot (".") as matching all chars including line breaks in regular expression patterns') + parser.add_argument( + '--raw-pattern', + action='store_true', + default=False, + help='Treats string patterns as raw patterns w/o quotes (or quote escapes)') + parser.add_argument('schema', help='file containing JSON schema ("-" for stdin)') args = parser.parse_args(args_in) - schema = json.load(sys.stdin if args.schema == '-' else open(args.schema)) - prop_order = {name: idx for idx, name in enumerate(args.prop_order)} - converter = SchemaConverter(prop_order) + if args.schema.startswith('https://'): + url = args.schema + import requests + schema = requests.get(url).json() + elif args.schema == '-': + url = 'stdin' + schema = json.load(sys.stdin) + else: + url = f'file://{args.schema}' + with open(args.schema) as f: + schema = json.load(f) + converter = SchemaConverter( + prop_order={name: idx for idx, name in enumerate(args.prop_order)}, + allow_fetch=args.allow_fetch, + dotall=args.dotall, + raw_pattern=args.raw_pattern) + schema = converter.resolve_refs(schema, url) converter.visit(schema, '') print(converter.format_grammar()) diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index 4cb230804..82413b79d 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -249,6 +249,9 @@ static ggml_type ggml_type_from_name(const std::string & s) { if (s == "q5_1") { return GGML_TYPE_Q5_1; } + if (s == "iq4_nl") { + return GGML_TYPE_IQ4_NL; + } return GGML_TYPE_COUNT; } diff --git a/examples/llava/MobileVLM-README.md b/examples/llava/MobileVLM-README.md index 9eba791da..4d5fef020 100644 --- a/examples/llava/MobileVLM-README.md +++ b/examples/llava/MobileVLM-README.md @@ -1,11 +1,13 @@ # MobileVLM -Currently this implementation supports [MobileVLM-v1.7](https://huggingface.co/mtgv/MobileVLM-1.7B) variants. +Currently this implementation supports [MobileVLM-1.7B](https://huggingface.co/mtgv/MobileVLM-1.7B) / [MobileVLM_V2-1.7B](https://huggingface.co/mtgv/MobileVLM_V2-1.7B) variants. for more information, please go to [Meituan-AutoML/MobileVLM](https://github.com/Meituan-AutoML/MobileVLM) The implementation is based on llava, and is compatible with llava and mobileVLM. The usage is basically same as llava. +Notice: The overall process of model inference for both **MobileVLM** and **MobileVLM_V2** models is the same, but the process of model conversion is a little different. Therefore, using MobiVLM as an example, the different conversion step will be shown. + ## Usage Build with cmake or run `make llava-cli` to build it. @@ -34,7 +36,7 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336 python ./examples/llava/llava-surgery.py -m path/to/MobileVLM-1.7B ``` -3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` to convert the LLaVA image encoder to GGUF: +3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` (for **V2** the arg is `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF: ```sh python ./examples/llava/convert-image-encoder-to-gguf \ @@ -44,6 +46,14 @@ python ./examples/llava/convert-image-encoder-to-gguf \ --projector-type ldp ``` +```sh +python ./examples/llava/convert-image-encoder-to-gguf \ + -m path/to/clip-vit-large-patch14-336 \ + --llava-projector path/to/MobileVLM-1.7B_V2/llava.projector \ + --output-dir path/to/MobileVLM-1.7B_V2 \ + --projector-type ldpv2 +``` + 4. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF: ```sh diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index 690bca2eb..48caafa87 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -119,6 +119,7 @@ static std::string format(const char * fmt, ...) { #define TN_LLAVA_PROJ "mm.%d.%s" #define TN_MVLM_PROJ_MLP "mm.model.mlp.%d.%s" #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s" +#define TN_MVLM_PROJ_PEG "mm.model.peg.%d.%s" #define TN_IMAGE_NEWLINE "model.image_newline" @@ -126,12 +127,14 @@ enum projector_type { PROJECTOR_TYPE_MLP, PROJECTOR_TYPE_MLP_NORM, PROJECTOR_TYPE_LDP, + PROJECTOR_TYPE_LDPV2, PROJECTOR_TYPE_UNKNOWN, }; static std::map PROJECTOR_TYPE_NAMES = { { PROJECTOR_TYPE_MLP, "mlp" }, { PROJECTOR_TYPE_LDP, "ldp" }, + { PROJECTOR_TYPE_LDPV2, "ldpv2"}, }; @@ -475,6 +478,14 @@ struct clip_vision_model { struct ggml_tensor * mm_model_block_2_block_2_0_w; struct ggml_tensor * mm_model_block_2_block_2_1_w; struct ggml_tensor * mm_model_block_2_block_2_1_b; + + // MobileVLM_V2 projection + struct ggml_tensor * mm_model_mlp_0_w; + struct ggml_tensor * mm_model_mlp_0_b; + struct ggml_tensor * mm_model_mlp_2_w; + struct ggml_tensor * mm_model_mlp_2_b; + struct ggml_tensor * mm_model_peg_0_w; + struct ggml_tensor * mm_model_peg_0_b; }; struct clip_ctx { @@ -807,6 +818,29 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 } embeddings = block_1; } + else if (ctx->proj_type == PROJECTOR_TYPE_LDPV2) + { + int n_patch = 24; + struct ggml_tensor * mlp_0 = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings); + mlp_0 = ggml_add(ctx0, mlp_0, model.mm_model_mlp_0_b); + mlp_0 = ggml_gelu(ctx0, mlp_0); + struct ggml_tensor * mlp_2 = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, mlp_0); + mlp_2 = ggml_add(ctx0, mlp_2, model.mm_model_mlp_2_b); + // mlp_2 ne = [2048, 576, 1, 1] + // // AVG Pool Layer 2*2, strides = 2 + mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 0, 2, 3)); + // mlp_2 ne = [576, 2048, 1, 1] + mlp_2 = ggml_reshape_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]); + // mlp_2 ne [24, 24, 2048, 1] + mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0); + // weight ne = [3, 3, 2048, 1] + struct ggml_tensor * peg_0 = ggml_conv_depthwise_2d(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1); + peg_0 = ggml_add(ctx0, peg_0, mlp_2); + peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3)); + peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b); + peg_0 = ggml_reshape_3d(ctx0, peg_0, peg_0->ne[0], peg_0->ne[1] * peg_0->ne[2], peg_0->ne[3]); + embeddings = peg_0; + } else { GGML_ASSERT(false); } @@ -1177,7 +1211,18 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { vision_model.mm_model_block_2_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight")); vision_model.mm_model_block_2_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight")); vision_model.mm_model_block_2_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias")); - } else { + } + else if (new_clip->proj_type == PROJECTOR_TYPE_LDPV2) + { + // MobilVLM_V2 projection + vision_model.mm_model_mlp_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 0, "weight")); + vision_model.mm_model_mlp_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 0, "bias")); + vision_model.mm_model_mlp_2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 2, "weight")); + vision_model.mm_model_mlp_2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 2, "bias")); + vision_model.mm_model_peg_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "weight")); + vision_model.mm_model_peg_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "bias")); + } + else { std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type]; throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); } @@ -1966,6 +2011,9 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) { if (ctx->proj_type == PROJECTOR_TYPE_LDP) { return ctx->vision_model.mm_model_block_1_block_2_1_b->ne[0]; } + if (ctx->proj_type == PROJECTOR_TYPE_LDPV2) { + return ctx->vision_model.mm_model_peg_0_b->ne[0]; + } if (ctx->proj_type == PROJECTOR_TYPE_MLP) { return ctx->vision_model.mm_2_b->ne[0]; } diff --git a/examples/llava/convert-image-encoder-to-gguf.py b/examples/llava/convert-image-encoder-to-gguf.py index c69f89ac2..b00bf7c6d 100644 --- a/examples/llava/convert-image-encoder-to-gguf.py +++ b/examples/llava/convert-image-encoder-to-gguf.py @@ -1,6 +1,7 @@ import argparse import os import json +import re import torch import numpy as np @@ -38,9 +39,11 @@ def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: b def get_tensor_name(name: str) -> str: if "projection" in name: return name - if "mm_projector" in name: - return name.replace("model.mm_projector", "mm") + name = name.replace("model.mm_projector", "mm") + name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1) + name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1) + return name return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln") @@ -83,7 +86,7 @@ ap.add_argument("--clip-model-is-vision", action="store_true", required=False, ap.add_argument("--clip-model-is-openclip", action="store_true", required=False, help="The clip model is from openclip (for ViT-SO400M type))") ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.") -ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp", choices=["mlp", "ldp"], default="mlp") +ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp") ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None) # Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711 # Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5 diff --git a/examples/lookup/CMakeLists.txt b/examples/lookup/CMakeLists.txt index c060b8f56..b91633f63 100644 --- a/examples/lookup/CMakeLists.txt +++ b/examples/lookup/CMakeLists.txt @@ -3,3 +3,21 @@ add_executable(${TARGET} lookup.cpp) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) + +set(TARGET lookup-create) +add_executable(${TARGET} lookup-create.cpp) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) + +set(TARGET lookup-merge) +add_executable(${TARGET} lookup-merge.cpp) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) + +set(TARGET lookup-stats) +add_executable(${TARGET} lookup-stats.cpp) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/lookup/lookup-create.cpp b/examples/lookup/lookup-create.cpp new file mode 100644 index 000000000..46a6bed07 --- /dev/null +++ b/examples/lookup/lookup-create.cpp @@ -0,0 +1,43 @@ +#include "ggml.h" +#include "llama.h" +#include "common.h" +#include "ngram-cache.h" + +#include +#include +#include +#include +#include +#include + +int main(int argc, char ** argv){ + gpt_params params; + + if (!gpt_params_parse(argc, argv, params)) { + return 1; + } + // init llama.cpp + llama_backend_init(); + llama_numa_init(params.numa); + + llama_model * model = NULL; + llama_context * ctx = NULL; + + // load the model + std::tie(model, ctx) = llama_init_from_gpt_params(params); + GGML_ASSERT(model != nullptr); + + // tokenize the prompt + const bool add_bos = llama_should_add_bos_token(model); + + std::vector inp; + inp = ::llama_tokenize(ctx, params.prompt, add_bos, true); + fprintf(stderr, "%s: tokenization done\n", __func__); + + + llama_ngram_cache ngram_cache; + llama_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true); + fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str()); + + llama_ngram_cache_save(ngram_cache, params.lookup_cache_static); +} diff --git a/examples/lookup/lookup-merge.cpp b/examples/lookup/lookup-merge.cpp new file mode 100644 index 000000000..07c93eb8d --- /dev/null +++ b/examples/lookup/lookup-merge.cpp @@ -0,0 +1,47 @@ +#include "ggml.h" +#include "llama.h" +#include "common.h" +#include "ngram-cache.h" + +#include +#include +#include +#include +#include +#include +#include + +static void print_usage() { + fprintf(stderr, "Merges multiple lookup cache files into a single one.\n"); + fprintf(stderr, "Usage: lookup-merge [--help] lookup_part_1.bin lookup_part_2.bin ... lookup_merged.bin\n"); +} + +int main(int argc, char ** argv){ + if (argc < 3) { + print_usage(); + exit(1); + } + + std::vector args; + args.resize(argc-1); + for (int i = 0; i < argc-1; ++i) { + args[i] = argv[i+1]; + if (args[i] == "-h" || args[i] == "--help") { + print_usage(); + exit(0); + } + } + + fprintf(stderr, "lookup-merge: loading file %s\n", args[0].c_str()); + llama_ngram_cache ngram_cache_merged = llama_ngram_cache_load(args[0]); + + for (size_t i = 1; i < args.size()-1; ++i) { + fprintf(stderr, "lookup-merge: loading file %s\n", args[i].c_str()); + llama_ngram_cache ngram_cache = llama_ngram_cache_load(args[i]); + + llama_ngram_cache_merge(ngram_cache_merged, ngram_cache); + } + + fprintf(stderr, "lookup-merge: saving file %s\n", args.back().c_str()); + llama_ngram_cache_save(ngram_cache_merged, args.back()); +} diff --git a/examples/lookup/lookup-stats.cpp b/examples/lookup/lookup-stats.cpp new file mode 100644 index 000000000..31f227773 --- /dev/null +++ b/examples/lookup/lookup-stats.cpp @@ -0,0 +1,163 @@ +#include "ggml.h" +#include "common.h" +#include "llama.h" +#include "log.h" +#include "ngram-cache.h" + +#include +#include +#include +#include +#include +#include +#include + +int main(int argc, char ** argv){ + gpt_params params; + + if (!gpt_params_parse(argc, argv, params)) { + return 1; + } + + const int n_draft = params.n_draft; + + // init llama.cpp + llama_backend_init(); + llama_numa_init(params.numa); + + llama_model * model = NULL; + llama_context * ctx = NULL; + + // load the model + std::tie(model, ctx) = llama_init_from_gpt_params(params); + llama_set_rng_seed(ctx, params.seed); + GGML_ASSERT(llama_n_vocab(model) < (1 << 16)); + + // tokenize the prompt + const bool add_bos = llama_should_add_bos_token(model); + LOG("add_bos tgt: %d\n", add_bos); + + std::vector inp; + inp = ::llama_tokenize(ctx, params.prompt, add_bos, true); + + llama_ngram_cache ngram_cache_context; + llama_ngram_cache ngram_cache_dynamic; + llama_ngram_cache ngram_cache_static; + int64_t t_draft_flat_us = 0; + int64_t t_draft_us = 0; + + { + const int64_t t_start_draft_us = ggml_time_us(); + + if (!params.lookup_cache_static.empty()) { + try { + ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static); + } catch (std::ifstream::failure const &) { + fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str()); + exit(1); + } + } + + if (!params.lookup_cache_dynamic.empty()) { + try { + ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic); + } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program + } + + t_draft_flat_us += ggml_time_us() - t_start_draft_us; + } + + const int n_input = inp.size(); + const int n_ctx = params.n_ctx; + + int n_drafted = 0; + int n_accept = 0; + + const int64_t t_start_ms = ggml_time_ms(); + + // Iterate over input tokens in chunks of size n_ctx. + // Each chunk is treated as if a sequential generation but with pre-determined tokens to ensure reproducibility. + for (int i_start = 0; i_start + n_ctx < n_input; i_start += n_ctx) { + const std::vector inp_slice(inp.begin() + i_start, inp.begin() + i_start + n_ctx); + std::vector pseudo_output; + pseudo_output.push_back(inp_slice[0]); + + while ((int) pseudo_output.size() < n_ctx) { + // Simulate drafting and decoding from draft: + std::vector draft; + draft.push_back(pseudo_output.back()); + + { + const int64_t t_start_draft_us = ggml_time_us(); + llama_ngram_cache_draft(pseudo_output, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static); + t_draft_us += ggml_time_us() - t_start_draft_us; + } + + n_drafted += draft.size() - 1; + + for (size_t j = 1; j < draft.size() && (int) pseudo_output.size() < n_ctx; ++j) { + const llama_token ground_truth = inp_slice[pseudo_output.size()]; + const llama_token drafted = draft[j]; + + if (ground_truth != drafted) { + break; + } + + ++n_accept; + pseudo_output.push_back(ground_truth); + + { + const int64_t t_start_draft_us = ggml_time_us(); + llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false); + t_draft_us += ggml_time_us() - t_start_draft_us; + } + } + + // After each simulated batch decoding simulate the sampling of a single token: + if ((int) pseudo_output.size() < n_ctx) { + pseudo_output.push_back(inp_slice[pseudo_output.size()]); + { + const int64_t t_start_draft_us = ggml_time_us(); + llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false); + t_draft_us += ggml_time_us() - t_start_draft_us; + } + } + + draft.erase(draft.begin()); + + } + if (i_start > 0 && i_start / 100000 != (i_start - n_ctx) / 100000) { + const int64_t t_now_ms = ggml_time_ms(); + const int64_t eta_ms = (n_input - i_start) * (t_now_ms - t_start_ms) / i_start; + const int64_t eta_min = eta_ms / (60*1000); + const int64_t eta_s = (eta_ms - 60*1000*eta_min) / 1000; + + LOG_TEE("lookup-stats: %d/%d done, ETA: %02" PRId64 ":%02" PRId64 "\n", i_start, n_input, eta_min, eta_s); + } + + // After each chunk, update the dynamic ngram cache with the context ngram cache: + llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context); + ngram_cache_context.clear(); + } + + LOG_TEE("\n"); + + LOG_TEE("\n"); + LOG_TEE("n_draft = %d\n", n_draft); + LOG_TEE("n_predict = %d\n", n_input - n_input % n_ctx); + LOG_TEE("n_drafted = %d\n", n_drafted); + LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3); + LOG_TEE("t_draft = %.2f ms, %.2f us per token, %.2f tokens per second\n", + t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us)); + LOG_TEE("n_accept = %d\n", n_accept); + LOG_TEE("accept = %.3f%%\n", 100.0f * n_accept / n_drafted); + + llama_free(ctx); + llama_free_model(model); + + llama_backend_free(); + + fprintf(stderr, "\n\n"); + + return 0; +} diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp index b53fae110..2e8c35de3 100644 --- a/examples/lookup/lookup.cpp +++ b/examples/lookup/lookup.cpp @@ -1,12 +1,15 @@ -#include "common.h" #include "ggml.h" #include "llama.h" +#include "common.h" +#include "ngram-cache.h" #include #include #include +#include #include #include +#include int main(int argc, char ** argv){ gpt_params params; @@ -15,11 +18,7 @@ int main(int argc, char ** argv){ return 1; } - // max/min n-grams size to search for in prompt - const int ngram_max = 4; - const int ngram_min = 1; - - // length of the candidate / draft sequence, if match is found + // max. number of additional tokens to draft if match is found const int n_draft = params.n_draft; const bool dump_kv_cache = params.dump_kv_cache; @@ -39,6 +38,8 @@ int main(int argc, char ** argv){ // load the model std::tie(model, ctx) = llama_init_from_gpt_params(params); + llama_set_rng_seed(ctx, params.seed); + GGML_ASSERT(llama_n_vocab(model) < (1 << 16)); // tokenize the prompt const bool add_bos = llama_should_add_bos_token(model); @@ -47,6 +48,35 @@ int main(int argc, char ** argv){ std::vector inp; inp = ::llama_tokenize(ctx, params.prompt, add_bos, true); + llama_ngram_cache ngram_cache_context; + llama_ngram_cache ngram_cache_dynamic; + llama_ngram_cache ngram_cache_static; + int64_t t_draft_flat_us = 0; + int64_t t_draft_us = 0; + + { + // Fill up context ngram cache with tokens from user input: + const int64_t t_start_draft_us = ggml_time_us(); + llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false); + + if (!params.lookup_cache_static.empty()) { + try { + ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static); + } catch (std::ifstream::failure const &) { + fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str()); + exit(1); + } + } + + if (!params.lookup_cache_dynamic.empty()) { + try { + ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic); + } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program + } + + t_draft_flat_us += ggml_time_us() - t_start_draft_us; + } + const int max_context_size = llama_n_ctx(ctx); const int max_tokens_list_size = max_context_size - 4; @@ -76,8 +106,6 @@ int main(int argc, char ** argv){ int n_drafted = 0; int n_accept = 0; - int64_t t_draft_us = 0; - int n_past = inp.size(); bool has_eos = false; @@ -129,6 +157,12 @@ int main(int argc, char ** argv){ ++n_past; ++i_dft; inp.push_back(id); + { + // Update context ngram cache with the newly accepted token: + const int64_t t_start_draft_us = ggml_time_us(); + llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false); + t_draft_us += ggml_time_us() - t_start_draft_us; + } if (params.use_color) { // color accepted draft token @@ -149,6 +183,12 @@ int main(int argc, char ** argv){ draft.clear(); draft.push_back(id); inp.push_back(id); + { + // Update context ngram cache with the newly accepted token: + const int64_t t_start_draft_us = ggml_time_us(); + llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false); + t_draft_us += ggml_time_us() - t_start_draft_us; + } break; } @@ -163,44 +203,19 @@ int main(int argc, char ** argv){ llama_batch_clear(batch_tgt); llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true); - // generate n_pred tokens through prompt lookup - auto prompt_lookup = [&]() -> void { - const int inp_size = inp.size(); - for (int ngram_size = ngram_max ; ngram_size > ngram_min; --ngram_size){ - const llama_token * ngram = &inp[inp_size - ngram_size]; - - for (int i = 0; i <= (int) inp_size - (ngram_size * 2); ++i) { - bool match = true; - for (int j = 0; j < ngram_size; ++j) { - if (inp[i + j] != ngram[j]) { - match = false; - break; - } - } - - if (match) { - const int startIdx = i + ngram_size; - const int endIdx = startIdx + n_draft; - if (endIdx < inp_size) { - for (int j = startIdx; j < endIdx; ++j) { - LOG(" - draft candidate %d: %d\n", j, inp[j]); - draft.push_back(inp[j]); - llama_batch_add(batch_tgt, inp[j], n_past + (j - startIdx) + 1, { 0 }, true); - ++n_drafted; - } - return; - } - } - } - } - return; - }; - + // Draft already contains a single token sampled from the model: + GGML_ASSERT(draft.size() == 1); + GGML_ASSERT(draft[0] == inp.back()); const int64_t t_start_draft_us = ggml_time_us(); - prompt_lookup(); + llama_ngram_cache_draft(inp, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static); + + for (size_t i = 1; i < draft.size(); ++i) { + llama_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true); + } t_draft_us += ggml_time_us() - t_start_draft_us; + n_drafted += draft.size() - 1; llama_decode(ctx, batch_tgt); ++n_past; @@ -210,19 +225,24 @@ int main(int argc, char ** argv){ auto t_dec_end = ggml_time_us(); + // Update dynamic ngram cache with context ngram cache and save it to disk: + llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context); + llama_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic); + LOG_TEE("\n\n"); LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f)); LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f)); LOG_TEE("\n"); - LOG_TEE("n_draft = %d\n", n_draft); - LOG_TEE("n_predict = %d\n", n_predict); - LOG_TEE("n_drafted = %d\n", n_drafted); - LOG_TEE("t_draft = %.2f ms, %.2f us per token, %.2f tokens per second\n", + LOG_TEE("n_draft = %d\n", n_draft); + LOG_TEE("n_predict = %d\n", n_predict); + LOG_TEE("n_drafted = %d\n", n_drafted); + LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3); + LOG_TEE("t_draft = %.2f ms, %.2f us per token, %.2f tokens per second\n", t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us)); - LOG_TEE("n_accept = %d\n", n_accept); - LOG_TEE("accept = %.3f%%\n", 100.0f * n_accept / n_drafted); + LOG_TEE("n_accept = %d\n", n_accept); + LOG_TEE("accept = %.3f%%\n", 100.0f * n_accept / n_drafted); LOG_TEE("\ntarget:\n"); llama_print_timings(ctx); diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 7662ec80c..79e60ea7b 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -189,6 +189,18 @@ static void prepare_imatrix(const std::string& imatrix_file, } } +static ggml_type parse_ggml_type(const char * arg) { + ggml_type result = GGML_TYPE_COUNT; + for (int j = 0; j < GGML_TYPE_COUNT; ++j) { + auto type = ggml_type(j); + const auto * name = ggml_type_name(type); + if (name && strcmp(arg, name) == 0) { + result = type; break; + } + } + return result; +} + int main(int argc, char ** argv) { if (argc < 3) { usage(argv[0]); @@ -203,6 +215,18 @@ int main(int argc, char ** argv) { for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) { if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) { params.quantize_output_tensor = false; + } else if (strcmp(argv[arg_idx], "--output-tensor-type") == 0) { + if (arg_idx < argc-1) { + params.output_tensor_type = parse_ggml_type(argv[++arg_idx]); + } else { + usage(argv[0]); + } + } else if (strcmp(argv[arg_idx], "--token-embedding-type") == 0) { + if (arg_idx < argc-1) { + params.token_embedding_type = parse_ggml_type(argv[++arg_idx]); + } else { + usage(argv[0]); + } } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) { params.allow_requantize = true; } else if (strcmp(argv[arg_idx], "--pure") == 0) { diff --git a/examples/regex-to-grammar.py b/examples/regex-to-grammar.py new file mode 100644 index 000000000..c98118e95 --- /dev/null +++ b/examples/regex-to-grammar.py @@ -0,0 +1,20 @@ +import json, subprocess, sys, os + +assert len(sys.argv) >= 2 +[_, pattern, *rest] = sys.argv + +print(subprocess.check_output( + [ + "python", + os.path.join( + os.path.dirname(os.path.realpath(__file__)), + "json-schema-to-grammar.py"), + *rest, + "-", + "--raw-pattern", + ], + text=True, + input=json.dumps({ + "type": "string", + "pattern": pattern, + }, indent=2))) diff --git a/examples/retrieval/CMakeLists.txt b/examples/retrieval/CMakeLists.txt new file mode 100644 index 000000000..eaabae08d --- /dev/null +++ b/examples/retrieval/CMakeLists.txt @@ -0,0 +1,5 @@ +set(TARGET retrieval) +add_executable(${TARGET} retrieval.cpp) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/retrieval/README.md b/examples/retrieval/README.md new file mode 100644 index 000000000..2b2595c46 --- /dev/null +++ b/examples/retrieval/README.md @@ -0,0 +1,69 @@ +# llama.cpp/examples/retrieval + +Demonstration of simple retrieval technique based on cosine similarity + +More info: +https://github.com/ggerganov/llama.cpp/pull/6193 + +### How to use + +`retieval.cpp` has parameters of its own: +- `--context-file`: file to be embedded - state this option multiple times to embed multiple files +- `--chunk-size`: minimum size of each text chunk to be embedded +- `--chunk-separator`: STRING to divide chunks by. newline by default + +`retrieval` example can be tested as follows: + +```bash +make -j && ./retrieval --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator . +``` + +This chunks and embeds all given files and starts a loop requesting query inputs: + +``` +Enter query: +``` + +On each query input, top k chunks are shown along with file name, chunk position within file and original text: + +``` +Enter query: describe the mit license +batch_decode: n_tokens = 6, n_seq = 1 +Top 3 similar chunks: +filename: README.md +filepos: 119 +similarity: 0.762334 +textdata: +png) + +[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT) + +[Roadmap](https://github. +-------------------- +filename: License +filepos: 0 +similarity: 0.725146 +textdata: +MIT License + +Copyright (c) 2023 Georgi Gerganov + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. +-------------------- +filename: README.md +filepos: 9178 +similarity: 0.621722 +textdata: +com/cztomsik/ava) (MIT) +- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal) +- [pythops/tenere](https://github. +-------------------- +``` diff --git a/examples/retrieval/retrieval.cpp b/examples/retrieval/retrieval.cpp new file mode 100644 index 000000000..5ba71e76a --- /dev/null +++ b/examples/retrieval/retrieval.cpp @@ -0,0 +1,350 @@ +#include "common.h" +#include "llama.h" + +#include +#include + +struct retrieval_params { + std::vector context_files; // context files to embed + int32_t chunk_size = 64; // chunk size for context embedding + std::string chunk_separator = "\n"; // chunk separator for context embedding +}; + +static void retrieval_params_print_usage(int argc, char ** argv, gpt_params & gpt_params, retrieval_params & params) { + gpt_print_usage(argc, argv, gpt_params); + printf("retrieval options:\n"); + printf(" --context-file FNAME file containing context to embed.\n"); + printf(" specify multiple files by providing --context-file option multiple times.\n"); + printf(" --chunk-size N minimum length of embedded text chunk (default:%d)\n", params.chunk_size); + printf(" --chunk-separator STRING\n"); + printf(" string to separate chunks (default: \"\\n\")\n"); + printf("\n"); +} + +static void retrieval_params_parse(int argc, char ** argv, gpt_params & gpt_params, retrieval_params & retrieval_params) { + int i = 1; + std::string arg; + while (i < argc) { + arg = argv[i]; + bool invalid_gpt_param = false; + if(gpt_params_find_arg(argc, argv, argv[i], gpt_params, i, invalid_gpt_param)) { + if (invalid_gpt_param) { + fprintf(stderr, "error: invalid argument: %s\n", arg.c_str()); + retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params); + exit(1); + } + // option was parsed by gpt_params_find_arg + } else if (arg == "--context-file") { + if (++i >= argc) { + fprintf(stderr, "error: missing argument for --context-file\n"); + retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params); + exit(1); + } + std::ifstream file(argv[i]); + if (!file) { + fprintf(stderr, "error: failed to open file '%s'\n", argv[i]); + retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params); + exit(1); + } + // store the external file name in params + retrieval_params.context_files.push_back(argv[i]); + } else if (arg == "--chunk-size") { + if (++i >= argc) { + fprintf(stderr, "error: missing argument for --chunk-size\n"); + retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params); + exit(1); + } + retrieval_params.chunk_size = std::stoi(argv[i]); + } else if (arg == "--chunk-separator") { + if (++i >= argc) { + fprintf(stderr, "error: missing argument for --chunk-separator\n"); + retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params); + exit(1); + } + retrieval_params.chunk_separator = argv[i]; + } else { + // unknown argument + fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); + retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params); + exit(1); + } + i++; + } +} + +struct chunk { + // filename + std::string filename; + // original file position + size_t filepos; + // original text data + std::string textdata = ""; + // tokenized text data + std::vector tokens; + // embedding + std::vector embedding; +}; + +// chunk file data to chunks of size >= chunk_size +// chunk_separator is the separator between chunks +static std::vector chunk_file(const std::string & filename, int chunk_size, const std::string & chunk_separator) { + std::vector chunks; + std::ifstream f(filename.c_str()); + + if (!f.is_open()) { + fprintf(stderr, "Error: could not open file %s\n", filename.c_str()); + return chunks; + } + + chunk current_chunk; + char buffer[1024]; + int64_t filepos = 0; + std::string current = ""; + while (f.read(buffer, 1024)) { + current += std::string(buffer, f.gcount()); + size_t pos; + while ((pos = current.find(chunk_separator)) != std::string::npos) { + current_chunk.textdata += current.substr(0, pos + chunk_separator.size()); + if ((int) current_chunk.textdata.size() > chunk_size) { + // save chunk + current_chunk.filepos = filepos; + current_chunk.filename = filename; + chunks.push_back(current_chunk); + // update filepos + filepos += (int) current_chunk.textdata.size(); + // reset current_chunk + current_chunk = chunk(); + } + current = current.substr(pos + chunk_separator.size()); + } + + } + // add leftover data to last chunk + if (current_chunk.textdata.size() > 0) { + if (chunks.empty()) { + current_chunk.filepos = filepos; + current_chunk.filename = filename; + chunks.push_back(current_chunk); + } else { + chunks.back().textdata += current_chunk.textdata; + } + } + f.close(); + return chunks; +} + +static void batch_add_seq(llama_batch & batch, const std::vector & tokens, int seq_id) { + for (size_t i = 0; i < tokens.size(); i++) { + llama_batch_add(batch, tokens[i], i, { seq_id }, i == tokens.size() - 1); + } +} + +static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) { + // clear previous kv_cache values (irrelevant for embeddings) + llama_kv_cache_clear(ctx); + + // run model + fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq); + if (llama_decode(ctx, batch) < 0) { + fprintf(stderr, "%s : failed to decode\n", __func__); + } + + for (int i = 0; i < batch.n_tokens; i++) { + if (!batch.logits[i]) { + continue; + } + + // try to get sequence embeddings - supported only when pooling_type is not NONE + const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]); + if (embd == NULL) { + embd = llama_get_embeddings_ith(ctx, i); + if (embd == NULL) { + fprintf(stderr, "%s: failed to get embeddings for token %d\n", __func__, i); + continue; + } + } + + float * out = output + batch.seq_id[i][0] * n_embd; + llama_embd_normalize(embd, out, n_embd); + } +} + +int main(int argc, char ** argv) { + gpt_params params; + retrieval_params retrieval_params; + + retrieval_params_parse(argc, argv, params, retrieval_params); + + // For BERT models, batch size must be equal to ubatch size + params.n_ubatch = params.n_batch; + + if (retrieval_params.chunk_size <= 0) { + fprintf(stderr, "chunk_size must be positive\n"); + return 1; + } + if (retrieval_params.context_files.empty()) { + fprintf(stderr, "context_files must be specified\n"); + return 1; + } + params.embedding = true; + + print_build_info(); + + printf("processing files:\n"); + for (auto & context_file : retrieval_params.context_files) { + printf("%s\n", context_file.c_str()); + } + + std::vector chunks; + for (auto & context_file : retrieval_params.context_files) { + std::vector file_chunk = chunk_file(context_file, retrieval_params.chunk_size, retrieval_params.chunk_separator); + chunks.insert(chunks.end(), file_chunk.begin(), file_chunk.end()); + } + printf("Number of chunks: %ld\n", chunks.size()); + + llama_backend_init(); + llama_numa_init(params.numa); + + llama_model * model; + llama_context * ctx; + + // load the model + std::tie(model, ctx) = llama_init_from_gpt_params(params); + if (model == NULL) { + fprintf(stderr, "%s: error: unable to load model\n", __func__); + return 1; + } + + const int n_ctx_train = llama_n_ctx_train(model); + const int n_ctx = llama_n_ctx(ctx); + + if (n_ctx > n_ctx_train) { + fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n", + __func__, n_ctx_train, n_ctx); + } + + // print system information + { + fprintf(stderr, "\n"); + fprintf(stderr, "%s\n", get_system_info(params).c_str()); + } + + // max batch size + const uint64_t n_batch = params.n_batch; + GGML_ASSERT(params.n_batch >= params.n_ctx); + + // tokenize the prompts and trim + for (auto & chunk : chunks) { + auto inp = ::llama_tokenize(ctx, chunk.textdata, true, false); + if (inp.size() > n_batch) { + fprintf(stderr, "%s: error: chunk size (%lld) exceeds batch size (%lld), increase batch size and re-run\n", + __func__, (long long int) inp.size(), (long long int) n_batch); + return 1; + } + // add eos if not present + if (inp.empty() || inp.back() != llama_token_eos(model)) { + inp.push_back(llama_token_eos(model)); + } + chunk.tokens = inp; + } + + // tokenization stats + if (params.verbose_prompt) { + for (int i = 0; i < (int) chunks.size(); i++) { + fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, chunks[i].textdata.c_str()); + fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, chunks[i].tokens.size()); + for (int j = 0; j < (int) chunks[i].tokens.size(); j++) { + fprintf(stderr, "%6d -> '%s'\n", chunks[i].tokens[j], llama_token_to_piece(ctx, chunks[i].tokens[j]).c_str()); + } + fprintf(stderr, "\n\n"); + } + } + + // initialize batch + const int n_chunks = chunks.size(); + struct llama_batch batch = llama_batch_init(n_batch, 0, 1); + + // allocate output + const int n_embd = llama_n_embd(model); + std::vector embeddings(n_chunks * n_embd, 0); + float * emb = embeddings.data(); + + // break into batches + int p = 0; // number of prompts processed already + int s = 0; // number of prompts in current batch + for (int k = 0; k < n_chunks; k++) { + // clamp to n_batch tokens + auto & inp = chunks[k].tokens; + + const uint64_t n_toks = inp.size(); + + // encode if at capacity + if (batch.n_tokens + n_toks > n_batch) { + float * out = emb + p * n_embd; + batch_decode(ctx, batch, out, s, n_embd); + llama_batch_clear(batch); + p += s; + s = 0; + } + + // add to batch + batch_add_seq(batch, inp, s); + s += 1; + } + + // final batch + float * out = emb + p * n_embd; + batch_decode(ctx, batch, out, s, n_embd); + + // save embeddings to chunks + for (int i = 0; i < n_chunks; i++) { + chunks[i].embedding = std::vector(emb + i * n_embd, emb + (i + 1) * n_embd); + // clear tokens as they are no longer needed + chunks[i].tokens.clear(); + } + + // start loop, receive query and return top k similar chunks based on cosine similarity + std::string query; + while (true) { + printf("Enter query: "); + std::getline(std::cin, query); + std::vector query_tokens = llama_tokenize(ctx, query, true); + + struct llama_batch query_batch = llama_batch_init(n_batch, 0, 1); + batch_add_seq(query_batch, query_tokens, 0); + + std::vector query_emb(n_embd, 0); + batch_decode(ctx, query_batch, query_emb.data(), 1, n_embd); + + llama_batch_clear(query_batch); + + // compute cosine similarities + { + std::vector> similarities; + for (int i = 0; i < n_chunks; i++) { + float sim = llama_embd_similarity_cos(chunks[i].embedding.data(), query_emb.data(), n_embd); + similarities.push_back(std::make_pair(i, sim)); + } + + // sort similarities + std::sort(similarities.begin(), similarities.end(), [](const std::pair & a, const std::pair & b) { + return a.second > b.second; + }); + + printf("Top %d similar chunks:\n", params.sparams.top_k); + for (int i = 0; i < std::min(params.sparams.top_k, (int) chunks.size()); i++) { + printf("filename: %s\n", chunks[similarities[i].first].filename.c_str()); + printf("filepos: %lld\n", (long long int) chunks[similarities[i].first].filepos); + printf("similarity: %f\n", similarities[i].second); + printf("textdata:\n%s\n", chunks[similarities[i].first].textdata.c_str()); + printf("--------------------\n"); + } + } + } + + // clean up + llama_print_timings(ctx); + llama_free(ctx); + llama_free_model(model); + llama_backend_free(); +} diff --git a/examples/server/CMakeLists.txt b/examples/server/CMakeLists.txt index f94de1e99..d2ee47d01 100644 --- a/examples/server/CMakeLists.txt +++ b/examples/server/CMakeLists.txt @@ -2,12 +2,16 @@ set(TARGET server) option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON) option(LLAMA_SERVER_SSL "Build SSL support for the server" OFF) include_directories(${CMAKE_CURRENT_SOURCE_DIR}) -add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h) +add_executable(${TARGET} + server.cpp + utils.hpp + httplib.h +) install(TARGETS ${TARGET} RUNTIME) target_compile_definitions(${TARGET} PRIVATE SERVER_VERBOSE=$ ) -target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT}) +target_link_libraries(${TARGET} PRIVATE common json-schema-to-grammar ${CMAKE_THREAD_LIBS_INIT}) if (LLAMA_SERVER_SSL) find_package(OpenSSL REQUIRED) target_link_libraries(${TARGET} PRIVATE OpenSSL::SSL OpenSSL::Crypto) diff --git a/examples/server/README.md b/examples/server/README.md index 755e1d538..49121a460 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -16,17 +16,20 @@ The project is under active development, and we are [looking for feedback and co **Command line options:** -- `--threads N`, `-t N`: Set the number of threads to use during generation. -- `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation. +- `--threads N`, `-t N`: Set the number of threads to use during generation. Not used if model layers are offloaded to GPU. The server is using batching, this parameter is used only if one token is to be processed on CPU backend. +- `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation. Not used if model layers are offloaded to GPU. - `--threads-http N`: number of threads in the http server pool to process requests (default: `max(std::thread::hardware_concurrency() - 1, --parallel N + 2)`) - `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`). -- `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf). +- `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (default: unused). +- `-hfr REPO, --hf-repo REPO`: Hugging Face model repository (default: unused). +- `-hff FILE, --hf-file FILE`: Hugging Face model file (default: unused). - `-a ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses. - `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of 4096. - `-ngl N`, `--n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance. - `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS. - `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS. -- `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `512`. +- `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `2048`. +- `-ub N`, `--ubatch-size N`: physical maximum batch size. Default: `512`. - `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended. - `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped. - `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed. @@ -57,7 +60,7 @@ see https://github.com/ggerganov/llama.cpp/issues/1437 - `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included. - `--metrics`: enable prometheus `/metrics` compatible endpoint (default: disabled) - `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name (default: template taken from model's metadata). We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) -- `--log-disable`: Output logs to stdout only, default: enabled. +- `--log-disable`: Output logs to stdout only, not to `llama.log`. default: enabled. - `--log-format FORMAT`: Define the log output to FORMAT: json or text (default: json) **If compiled with `LLAMA_SERVER_SSL=ON`** @@ -260,7 +263,7 @@ node index.js `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:`. In this case, `[img-12]` will be replaced by the embeddings of the image with id `12` in the following `image_data` array: `{..., "image_data": [{"data": "", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA. - `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1) + `id_slot`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1) `cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch. (default: false) @@ -357,7 +360,7 @@ Notice that each `probs` is an array of length `n_probs`. - `default_generation_settings` - the default generation settings for the `/completion` endpoint, has the same fields as the `generation_settings` response object from the `/completion` endpoint. - `total_slots` - the total number of slots for process requests (defined by `--parallel` option) -- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only ChatML-tuned models, such as Dolphin, OpenOrca, OpenHermes, OpenChat-3.5, etc can be used with this endpoint. +- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only model with [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, ChatML template will be used. *Options:* diff --git a/examples/server/chat.mjs b/examples/server/chat.mjs index 219ebb51a..a79c8a3cd 100644 --- a/examples/server/chat.mjs +++ b/examples/server/chat.mjs @@ -26,8 +26,9 @@ const propOrder = grammarJsonSchemaPropOrder let grammar = null if (grammarJsonSchemaFile) { - const schema = JSON.parse(readFileSync(grammarJsonSchemaFile, 'utf-8')) - const converter = new SchemaConverter(propOrder) + let schema = JSON.parse(readFileSync(grammarJsonSchemaFile, 'utf-8')) + const converter = new SchemaConverter({prop_order: propOrder, allow_fetch: true}) + schema = await converter.resolveRefs(schema, grammarJsonSchemaFile) converter.visit(schema, '') grammar = converter.formatGrammar() } diff --git a/examples/server/completion.js.hpp b/examples/server/completion.js.hpp index 10734091e..f8bb4b0f2 100644 --- a/examples/server/completion.js.hpp +++ b/examples/server/completion.js.hpp @@ -483,4 +483,4 @@ unsigned char completion_js[] = { 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x7d, 0x0a }; -unsigned int completion_js_len = 5796; +size_t completion_js_len = 5796; diff --git a/examples/server/httplib.h b/examples/server/httplib.h index 28746000c..f360bd93e 100644 --- a/examples/server/httplib.h +++ b/examples/server/httplib.h @@ -1,14 +1,14 @@ // // httplib.h // -// Copyright (c) 2023 Yuji Hirose. All rights reserved. +// Copyright (c) 2024 Yuji Hirose. All rights reserved. // MIT License // #ifndef CPPHTTPLIB_HTTPLIB_H #define CPPHTTPLIB_HTTPLIB_H -#define CPPHTTPLIB_VERSION "0.12.2" +#define CPPHTTPLIB_VERSION "0.15.3" /* * Configuration @@ -82,6 +82,10 @@ #define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 8192 #endif +#ifndef CPPHTTPLIB_RANGE_MAX_COUNT +#define CPPHTTPLIB_RANGE_MAX_COUNT 1024 +#endif + #ifndef CPPHTTPLIB_TCP_NODELAY #define CPPHTTPLIB_TCP_NODELAY false #endif @@ -160,10 +164,6 @@ using ssize_t = long; #define WSA_FLAG_NO_HANDLE_INHERIT 0x80 #endif -#ifndef strcasecmp -#define strcasecmp _stricmp -#endif // strcasecmp - using socket_t = SOCKET; #ifdef CPPHTTPLIB_USE_POLL #define poll(fds, nfds, timeout) WSAPoll(fds, nfds, timeout) @@ -172,9 +172,15 @@ using socket_t = SOCKET; #else // not _WIN32 #include -#ifndef _AIX +#if !defined(_AIX) && !defined(__MVS__) #include #endif +#ifdef __MVS__ +#include +#ifndef NI_MAXHOST +#define NI_MAXHOST 1025 +#endif +#endif #include #include #include @@ -187,6 +193,7 @@ using socket_t = SOCKET; #endif #include #include +#include #include #include #include @@ -207,6 +214,7 @@ using socket_t = int; #include #include #include +#include #include #include #include @@ -223,6 +231,9 @@ using socket_t = int; #include #include #include +#include +#include +#include #ifdef CPPHTTPLIB_OPENSSL_SUPPORT #ifdef _WIN32 @@ -237,7 +248,6 @@ using socket_t = int; #ifdef _MSC_VER #pragma comment(lib, "crypt32.lib") -#pragma comment(lib, "cryptui.lib") #endif #elif defined(CPPHTTPLIB_USE_CERTS_FROM_MACOSX_KEYCHAIN) && defined(__APPLE__) #include @@ -259,10 +269,8 @@ using socket_t = int; #include #include -#if OPENSSL_VERSION_NUMBER < 0x1010100fL -#error Sorry, OpenSSL versions prior to 1.1.1 are not supported -#elif OPENSSL_VERSION_NUMBER < 0x30000000L -#define SSL_get1_peer_certificate SSL_get_peer_certificate +#if OPENSSL_VERSION_NUMBER < 0x30000000L +#error Sorry, OpenSSL versions prior to 3.0.0 are not supported #endif #endif @@ -321,7 +329,7 @@ struct scope_exit { explicit scope_exit(std::function &&f) : exit_function(std::move(f)), execute_on_destruction{true} {} - scope_exit(scope_exit &&rhs) + scope_exit(scope_exit &&rhs) noexcept : exit_function(std::move(rhs.exit_function)), execute_on_destruction{rhs.execute_on_destruction} { rhs.release(); @@ -344,6 +352,81 @@ private: } // namespace detail +enum StatusCode { + // Information responses + Continue_100 = 100, + SwitchingProtocol_101 = 101, + Processing_102 = 102, + EarlyHints_103 = 103, + + // Successful responses + OK_200 = 200, + Created_201 = 201, + Accepted_202 = 202, + NonAuthoritativeInformation_203 = 203, + NoContent_204 = 204, + ResetContent_205 = 205, + PartialContent_206 = 206, + MultiStatus_207 = 207, + AlreadyReported_208 = 208, + IMUsed_226 = 226, + + // Redirection messages + MultipleChoices_300 = 300, + MovedPermanently_301 = 301, + Found_302 = 302, + SeeOther_303 = 303, + NotModified_304 = 304, + UseProxy_305 = 305, + unused_306 = 306, + TemporaryRedirect_307 = 307, + PermanentRedirect_308 = 308, + + // Client error responses + BadRequest_400 = 400, + Unauthorized_401 = 401, + PaymentRequired_402 = 402, + Forbidden_403 = 403, + NotFound_404 = 404, + MethodNotAllowed_405 = 405, + NotAcceptable_406 = 406, + ProxyAuthenticationRequired_407 = 407, + RequestTimeout_408 = 408, + Conflict_409 = 409, + Gone_410 = 410, + LengthRequired_411 = 411, + PreconditionFailed_412 = 412, + PayloadTooLarge_413 = 413, + UriTooLong_414 = 414, + UnsupportedMediaType_415 = 415, + RangeNotSatisfiable_416 = 416, + ExpectationFailed_417 = 417, + ImATeapot_418 = 418, + MisdirectedRequest_421 = 421, + UnprocessableContent_422 = 422, + Locked_423 = 423, + FailedDependency_424 = 424, + TooEarly_425 = 425, + UpgradeRequired_426 = 426, + PreconditionRequired_428 = 428, + TooManyRequests_429 = 429, + RequestHeaderFieldsTooLarge_431 = 431, + UnavailableForLegalReasons_451 = 451, + + // Server error responses + InternalServerError_500 = 500, + NotImplemented_501 = 501, + BadGateway_502 = 502, + ServiceUnavailable_503 = 503, + GatewayTimeout_504 = 504, + HttpVersionNotSupported_505 = 505, + VariantAlsoNegotiates_506 = 506, + InsufficientStorage_507 = 507, + LoopDetected_508 = 508, + NotExtended_510 = 510, + NetworkAuthenticationRequired_511 = 511, +}; + using Headers = std::multimap; using Params = std::multimap; @@ -373,17 +456,18 @@ public: DataSink &operator=(DataSink &&) = delete; std::function write; + std::function is_writable; std::function done; std::function done_with_trailer; std::ostream os; private: - class data_sink_streambuf : public std::streambuf { + class data_sink_streambuf final : public std::streambuf { public: explicit data_sink_streambuf(DataSink &sink) : sink_(sink) {} protected: - std::streamsize xsputn(const char *s, std::streamsize n) { + std::streamsize xsputn(const char *s, std::streamsize n) override { sink_.write(s, static_cast(n)); return n; } @@ -465,6 +549,7 @@ struct Request { MultipartFormDataMap files; Ranges ranges; Match matches; + std::unordered_map path_params; // for client ResponseHandler response_handler; @@ -476,8 +561,7 @@ struct Request { bool has_header(const std::string &key) const; std::string get_header_value(const std::string &key, size_t id = 0) const; - template - T get_header_value(const std::string &key, size_t id = 0) const; + uint64_t get_header_value_u64(const std::string &key, size_t id = 0) const; size_t get_header_value_count(const std::string &key) const; void set_header(const std::string &key, const std::string &val); @@ -509,14 +593,14 @@ struct Response { bool has_header(const std::string &key) const; std::string get_header_value(const std::string &key, size_t id = 0) const; - template - T get_header_value(const std::string &key, size_t id = 0) const; + uint64_t get_header_value_u64(const std::string &key, size_t id = 0) const; size_t get_header_value_count(const std::string &key) const; void set_header(const std::string &key, const std::string &val); - void set_redirect(const std::string &url, int status = 302); + void set_redirect(const std::string &url, int status = StatusCode::Found_302); void set_content(const char *s, size_t n, const std::string &content_type); void set_content(const std::string &s, const std::string &content_type); + void set_content(std::string &&s, const std::string &content_type); void set_content_provider( size_t length, const std::string &content_type, ContentProvider provider, @@ -573,15 +657,16 @@ public: TaskQueue() = default; virtual ~TaskQueue() = default; - virtual void enqueue(std::function fn) = 0; + virtual bool enqueue(std::function fn) = 0; virtual void shutdown() = 0; virtual void on_idle() {} }; -class ThreadPool : public TaskQueue { +class ThreadPool final : public TaskQueue { public: - explicit ThreadPool(size_t n) : shutdown_(false) { + explicit ThreadPool(size_t n, size_t mqr = 0) + : shutdown_(false), max_queued_requests_(mqr) { while (n) { threads_.emplace_back(worker(*this)); n--; @@ -591,13 +676,17 @@ public: ThreadPool(const ThreadPool &) = delete; ~ThreadPool() override = default; - void enqueue(std::function fn) override { + bool enqueue(std::function fn) override { { std::unique_lock lock(mutex_); + if (max_queued_requests_ > 0 && jobs_.size() >= max_queued_requests_) { + return false; + } jobs_.push_back(std::move(fn)); } cond_.notify_one(); + return true; } void shutdown() override { @@ -647,6 +736,7 @@ private: std::list> jobs_; bool shutdown_; + size_t max_queued_requests_ = 0; std::condition_variable cond_; std::mutex mutex_; @@ -658,6 +748,82 @@ using SocketOptions = std::function; void default_socket_options(socket_t sock); +const char *status_message(int status); + +std::string get_bearer_token_auth(const Request &req); + +namespace detail { + +class MatcherBase { +public: + virtual ~MatcherBase() = default; + + // Match request path and populate its matches and + virtual bool match(Request &request) const = 0; +}; + +/** + * Captures parameters in request path and stores them in Request::path_params + * + * Capture name is a substring of a pattern from : to /. + * The rest of the pattern is matched agains the request path directly + * Parameters are captured starting from the next character after + * the end of the last matched static pattern fragment until the next /. + * + * Example pattern: + * "/path/fragments/:capture/more/fragments/:second_capture" + * Static fragments: + * "/path/fragments/", "more/fragments/" + * + * Given the following request path: + * "/path/fragments/:1/more/fragments/:2" + * the resulting capture will be + * {{"capture", "1"}, {"second_capture", "2"}} + */ +class PathParamsMatcher final : public MatcherBase { +public: + PathParamsMatcher(const std::string &pattern); + + bool match(Request &request) const override; + +private: + static constexpr char marker = ':'; + // Treat segment separators as the end of path parameter capture + // Does not need to handle query parameters as they are parsed before path + // matching + static constexpr char separator = '/'; + + // Contains static path fragments to match against, excluding the '/' after + // path params + // Fragments are separated by path params + std::vector static_fragments_; + // Stores the names of the path parameters to be used as keys in the + // Request::path_params map + std::vector param_names_; +}; + +/** + * Performs std::regex_match on request path + * and stores the result in Request::matches + * + * Note that regex match is performed directly on the whole request. + * This means that wildcard patterns may match multiple path segments with /: + * "/begin/(.*)/end" will match both "/begin/middle/end" and "/begin/1/2/end". + */ +class RegexMatcher final : public MatcherBase { +public: + RegexMatcher(const std::string &pattern) : regex_(pattern) {} + + bool match(Request &request) const override; + +private: + std::regex regex_; +}; + +ssize_t write_headers(Stream &strm, const Headers &headers); + +} // namespace detail + class Server { public: using Handler = std::function; @@ -702,6 +868,7 @@ public: bool remove_mount_point(const std::string &mount_point); Server &set_file_extension_and_mimetype_mapping(const std::string &ext, const std::string &mime); + Server &set_default_file_mimetype(const std::string &mime); Server &set_file_request_handler(Handler handler); Server &set_error_handler(HandlerWithResponse handler); @@ -718,6 +885,8 @@ public: Server &set_socket_options(SocketOptions socket_options); Server &set_default_headers(Headers headers); + Server & + set_header_writer(std::function const &writer); Server &set_keep_alive_max_count(size_t count); Server &set_keep_alive_timeout(time_t sec); @@ -765,9 +934,14 @@ protected: size_t payload_max_length_ = CPPHTTPLIB_PAYLOAD_MAX_LENGTH; private: - using Handlers = std::vector>; + using Handlers = + std::vector, Handler>>; using HandlersForContentReader = - std::vector>; + std::vector, + HandlerWithContentReader>>; + + static std::unique_ptr + make_matcher(const std::string &pattern); socket_t create_server_socket(const std::string &host, int port, int socket_flags, @@ -778,16 +952,16 @@ private: bool routing(Request &req, Response &res, Stream &strm); bool handle_file_request(const Request &req, Response &res, bool head = false); - bool dispatch_request(Request &req, Response &res, const Handlers &handlers); - bool - dispatch_request_for_content_reader(Request &req, Response &res, - ContentReader content_reader, - const HandlersForContentReader &handlers); + bool dispatch_request(Request &req, Response &res, + const Handlers &handlers) const; + bool dispatch_request_for_content_reader( + Request &req, Response &res, ContentReader content_reader, + const HandlersForContentReader &handlers) const; - bool parse_request_line(const char *s, Request &req); + bool parse_request_line(const char *s, Request &req) const; void apply_ranges(const Request &req, Response &res, - std::string &content_type, std::string &boundary); - bool write_response(Stream &strm, bool close_connection, const Request &req, + std::string &content_type, std::string &boundary) const; + bool write_response(Stream &strm, bool close_connection, Request &req, Response &res); bool write_response_with_content(Stream &strm, bool close_connection, const Request &req, Response &res); @@ -806,21 +980,23 @@ private: bool read_content_core(Stream &strm, Request &req, Response &res, ContentReceiver receiver, MultipartContentHeader multipart_header, - ContentReceiver multipart_receiver); + ContentReceiver multipart_receiver) const; virtual bool process_and_close_socket(socket_t sock); + std::atomic is_running_{false}; + std::atomic done_{false}; + struct MountPointEntry { std::string mount_point; std::string base_dir; Headers headers; }; std::vector base_dirs_; - - std::atomic is_running_{false}; - std::atomic done_{false}; std::map file_extension_and_mimetype_map_; + std::string default_file_mimetype_ = "application/octet-stream"; Handler file_request_handler_; + Handlers get_handlers_; Handlers post_handlers_; HandlersForContentReader post_handlers_for_content_reader_; @@ -831,18 +1007,22 @@ private: Handlers delete_handlers_; HandlersForContentReader delete_handlers_for_content_reader_; Handlers options_handlers_; + HandlerWithResponse error_handler_; ExceptionHandler exception_handler_; HandlerWithResponse pre_routing_handler_; Handler post_routing_handler_; - Logger logger_; Expect100ContinueHandler expect_100_continue_handler_; + Logger logger_; + int address_family_ = AF_UNSPEC; bool tcp_nodelay_ = CPPHTTPLIB_TCP_NODELAY; SocketOptions socket_options_ = default_socket_options; Headers default_headers_; + std::function header_writer_ = + detail::write_headers; }; enum class Error { @@ -860,17 +1040,19 @@ enum class Error { UnsupportedMultipartBoundaryChars, Compression, ConnectionTimeout, + ProxyConnection, // For internal use only SSLPeerCouldBeClosed_, }; -std::string to_string(const Error error); +std::string to_string(Error error); std::ostream &operator<<(std::ostream &os, const Error &obj); class Result { public: + Result() = default; Result(std::unique_ptr &&res, Error err, Headers &&request_headers = Headers{}) : res_(std::move(res)), err_(err), @@ -893,13 +1075,13 @@ public: bool has_request_header(const std::string &key) const; std::string get_request_header_value(const std::string &key, size_t id = 0) const; - template - T get_request_header_value(const std::string &key, size_t id = 0) const; + uint64_t get_request_header_value_u64(const std::string &key, + size_t id = 0) const; size_t get_request_header_value_count(const std::string &key) const; private: std::unique_ptr res_; - Error err_; + Error err_ = Error::Unknown; Headers request_headers_; }; @@ -1059,16 +1241,21 @@ public: bool send(Request &req, Response &res, Error &error); Result send(const Request &req); - size_t is_socket_open() const; - - socket_t socket() const; - void stop(); + std::string host() const; + int port() const; + + size_t is_socket_open() const; + socket_t socket() const; + void set_hostname_addr_map(std::map addr_map); void set_default_headers(Headers headers); + void + set_header_writer(std::function const &writer); + void set_address_family(int family); void set_tcp_nodelay(bool on); void set_socket_options(SocketOptions socket_options); @@ -1117,6 +1304,7 @@ public: void set_ca_cert_path(const std::string &ca_cert_file_path, const std::string &ca_cert_dir_path = std::string()); void set_ca_cert_store(X509_STORE *ca_cert_store); + X509_STORE *create_ca_cert_store(const char *ca_cert, std::size_t size) const; #endif #ifdef CPPHTTPLIB_OPENSSL_SUPPORT @@ -1145,14 +1333,14 @@ protected: // Also, shutdown_ssl and close_socket should also NOT be called concurrently // with a DIFFERENT thread sending requests using that socket. virtual void shutdown_ssl(Socket &socket, bool shutdown_gracefully); - void shutdown_socket(Socket &socket); + void shutdown_socket(Socket &socket) const; void close_socket(Socket &socket); bool process_request(Stream &strm, Request &req, Response &res, bool close_connection, Error &error); bool write_content_with_provider(Stream &strm, const Request &req, - Error &error); + Error &error) const; void copy_settings(const ClientImpl &rhs); @@ -1177,6 +1365,10 @@ protected: // Default headers Headers default_headers_; + // Header writer + std::function header_writer_ = + detail::write_headers; + // Settings std::string client_cert_path_; std::string client_key_path_; @@ -1239,7 +1431,8 @@ private: Result send_(Request &&req); socket_t create_client_socket(Error &error) const; - bool read_response_line(Stream &strm, const Request &req, Response &res); + bool read_response_line(Stream &strm, const Request &req, + Response &res) const; bool write_request(Stream &strm, Request &req, bool close_connection, Error &error); bool redirect(Request &req, Response &res, Error &error); @@ -1258,7 +1451,7 @@ private: const std::string &content_type); ContentProviderWithoutLength get_multipart_content_provider( const std::string &boundary, const MultipartFormDataItems &items, - const MultipartFormDataProviderItems &provider_items); + const MultipartFormDataProviderItems &provider_items) const; std::string adjust_host_string(const std::string &host) const; @@ -1431,16 +1624,21 @@ public: bool send(Request &req, Response &res, Error &error); Result send(const Request &req); - size_t is_socket_open() const; - - socket_t socket() const; - void stop(); + std::string host() const; + int port() const; + + size_t is_socket_open() const; + socket_t socket() const; + void set_hostname_addr_map(std::map addr_map); void set_default_headers(Headers headers); + void + set_header_writer(std::function const &writer); + void set_address_family(int family); void set_tcp_nodelay(bool on); void set_socket_options(SocketOptions socket_options); @@ -1497,6 +1695,7 @@ public: const std::string &ca_cert_dir_path = std::string()); void set_ca_cert_store(X509_STORE *ca_cert_store); + void load_ca_cert_store(const char *ca_cert, std::size_t size); long get_openssl_verify_result() const; @@ -1538,7 +1737,7 @@ private: std::mutex ctx_mutex_; }; -class SSLClient : public ClientImpl { +class SSLClient final : public ClientImpl { public: explicit SSLClient(const std::string &host); @@ -1546,16 +1745,19 @@ public: explicit SSLClient(const std::string &host, int port, const std::string &client_cert_path, - const std::string &client_key_path); + const std::string &client_key_path, + const std::string &private_key_password = std::string()); explicit SSLClient(const std::string &host, int port, X509 *client_cert, - EVP_PKEY *client_key); + EVP_PKEY *client_key, + const std::string &private_key_password = std::string()); ~SSLClient() override; bool is_valid() const override; void set_ca_cert_store(X509_STORE *ca_cert_store); + void load_ca_cert_store(const char *ca_cert, std::size_t size); long get_openssl_verify_result() const; @@ -1564,7 +1766,7 @@ public: private: bool create_and_connect_socket(Socket &socket, Error &error) override; void shutdown_ssl(Socket &socket, bool shutdown_gracefully) override; - void shutdown_ssl_impl(Socket &socket, bool shutdown_socket); + void shutdown_ssl_impl(Socket &socket, bool shutdown_gracefully); bool process_socket(const Socket &socket, std::function callback) override; @@ -1608,15 +1810,9 @@ inline void duration_to_sec_and_usec(const T &duration, U callback) { callback(static_cast(sec), static_cast(usec)); } -template -inline T get_header_value(const Headers & /*headers*/, - const std::string & /*key*/, size_t /*id*/ = 0, - uint64_t /*def*/ = 0) {} - -template <> -inline uint64_t get_header_value(const Headers &headers, - const std::string &key, size_t id, - uint64_t def) { +inline uint64_t get_header_value_u64(const Headers &headers, + const std::string &key, size_t id, + uint64_t def) { auto rng = headers.equal_range(key); auto it = rng.first; std::advance(it, static_cast(id)); @@ -1628,14 +1824,14 @@ inline uint64_t get_header_value(const Headers &headers, } // namespace detail -template -inline T Request::get_header_value(const std::string &key, size_t id) const { - return detail::get_header_value(headers, key, id, 0); +inline uint64_t Request::get_header_value_u64(const std::string &key, + size_t id) const { + return detail::get_header_value_u64(headers, key, id, 0); } -template -inline T Response::get_header_value(const std::string &key, size_t id) const { - return detail::get_header_value(headers, key, id, 0); +inline uint64_t Response::get_header_value_u64(const std::string &key, + size_t id) const { + return detail::get_header_value_u64(headers, key, id, 0); } template @@ -1665,21 +1861,106 @@ inline ssize_t Stream::write_format(const char *fmt, const Args &...args) { inline void default_socket_options(socket_t sock) { int yes = 1; #ifdef _WIN32 - setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, reinterpret_cast(&yes), - sizeof(yes)); + setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, + reinterpret_cast(&yes), sizeof(yes)); setsockopt(sock, SOL_SOCKET, SO_EXCLUSIVEADDRUSE, - reinterpret_cast(&yes), sizeof(yes)); + reinterpret_cast(&yes), sizeof(yes)); #else #ifdef SO_REUSEPORT - setsockopt(sock, SOL_SOCKET, SO_REUSEPORT, reinterpret_cast(&yes), - sizeof(yes)); + setsockopt(sock, SOL_SOCKET, SO_REUSEPORT, + reinterpret_cast(&yes), sizeof(yes)); #else - setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, reinterpret_cast(&yes), - sizeof(yes)); + setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, + reinterpret_cast(&yes), sizeof(yes)); #endif #endif } +inline const char *status_message(int status) { + switch (status) { + case StatusCode::Continue_100: return "Continue"; + case StatusCode::SwitchingProtocol_101: return "Switching Protocol"; + case StatusCode::Processing_102: return "Processing"; + case StatusCode::EarlyHints_103: return "Early Hints"; + case StatusCode::OK_200: return "OK"; + case StatusCode::Created_201: return "Created"; + case StatusCode::Accepted_202: return "Accepted"; + case StatusCode::NonAuthoritativeInformation_203: + return "Non-Authoritative Information"; + case StatusCode::NoContent_204: return "No Content"; + case StatusCode::ResetContent_205: return "Reset Content"; + case StatusCode::PartialContent_206: return "Partial Content"; + case StatusCode::MultiStatus_207: return "Multi-Status"; + case StatusCode::AlreadyReported_208: return "Already Reported"; + case StatusCode::IMUsed_226: return "IM Used"; + case StatusCode::MultipleChoices_300: return "Multiple Choices"; + case StatusCode::MovedPermanently_301: return "Moved Permanently"; + case StatusCode::Found_302: return "Found"; + case StatusCode::SeeOther_303: return "See Other"; + case StatusCode::NotModified_304: return "Not Modified"; + case StatusCode::UseProxy_305: return "Use Proxy"; + case StatusCode::unused_306: return "unused"; + case StatusCode::TemporaryRedirect_307: return "Temporary Redirect"; + case StatusCode::PermanentRedirect_308: return "Permanent Redirect"; + case StatusCode::BadRequest_400: return "Bad Request"; + case StatusCode::Unauthorized_401: return "Unauthorized"; + case StatusCode::PaymentRequired_402: return "Payment Required"; + case StatusCode::Forbidden_403: return "Forbidden"; + case StatusCode::NotFound_404: return "Not Found"; + case StatusCode::MethodNotAllowed_405: return "Method Not Allowed"; + case StatusCode::NotAcceptable_406: return "Not Acceptable"; + case StatusCode::ProxyAuthenticationRequired_407: + return "Proxy Authentication Required"; + case StatusCode::RequestTimeout_408: return "Request Timeout"; + case StatusCode::Conflict_409: return "Conflict"; + case StatusCode::Gone_410: return "Gone"; + case StatusCode::LengthRequired_411: return "Length Required"; + case StatusCode::PreconditionFailed_412: return "Precondition Failed"; + case StatusCode::PayloadTooLarge_413: return "Payload Too Large"; + case StatusCode::UriTooLong_414: return "URI Too Long"; + case StatusCode::UnsupportedMediaType_415: return "Unsupported Media Type"; + case StatusCode::RangeNotSatisfiable_416: return "Range Not Satisfiable"; + case StatusCode::ExpectationFailed_417: return "Expectation Failed"; + case StatusCode::ImATeapot_418: return "I'm a teapot"; + case StatusCode::MisdirectedRequest_421: return "Misdirected Request"; + case StatusCode::UnprocessableContent_422: return "Unprocessable Content"; + case StatusCode::Locked_423: return "Locked"; + case StatusCode::FailedDependency_424: return "Failed Dependency"; + case StatusCode::TooEarly_425: return "Too Early"; + case StatusCode::UpgradeRequired_426: return "Upgrade Required"; + case StatusCode::PreconditionRequired_428: return "Precondition Required"; + case StatusCode::TooManyRequests_429: return "Too Many Requests"; + case StatusCode::RequestHeaderFieldsTooLarge_431: + return "Request Header Fields Too Large"; + case StatusCode::UnavailableForLegalReasons_451: + return "Unavailable For Legal Reasons"; + case StatusCode::NotImplemented_501: return "Not Implemented"; + case StatusCode::BadGateway_502: return "Bad Gateway"; + case StatusCode::ServiceUnavailable_503: return "Service Unavailable"; + case StatusCode::GatewayTimeout_504: return "Gateway Timeout"; + case StatusCode::HttpVersionNotSupported_505: + return "HTTP Version Not Supported"; + case StatusCode::VariantAlsoNegotiates_506: return "Variant Also Negotiates"; + case StatusCode::InsufficientStorage_507: return "Insufficient Storage"; + case StatusCode::LoopDetected_508: return "Loop Detected"; + case StatusCode::NotExtended_510: return "Not Extended"; + case StatusCode::NetworkAuthenticationRequired_511: + return "Network Authentication Required"; + + default: + case StatusCode::InternalServerError_500: return "Internal Server Error"; + } +} + +inline std::string get_bearer_token_auth(const Request &req) { + if (req.has_header("Authorization")) { + static std::string BearerHeaderPrefix = "Bearer "; + return req.get_header_value("Authorization") + .substr(BearerHeaderPrefix.length()); + } + return ""; +} + template inline Server & Server::set_read_timeout(const std::chrono::duration &duration) { @@ -1720,6 +2001,7 @@ inline std::string to_string(const Error error) { return "Unsupported HTTP multipart boundary characters"; case Error::Compression: return "Compression failed"; case Error::ConnectionTimeout: return "Connection timed out"; + case Error::ProxyConnection: return "Proxy connection failed"; case Error::Unknown: return "Unknown"; default: break; } @@ -1733,10 +2015,9 @@ inline std::ostream &operator<<(std::ostream &os, const Error &obj) { return os; } -template -inline T Result::get_request_header_value(const std::string &key, - size_t id) const { - return detail::get_header_value(request_headers_, key, id, 0); +inline uint64_t Result::get_request_header_value_u64(const std::string &key, + size_t id) const { + return detail::get_header_value_u64(request_headers_, key, id, 0); } template @@ -1790,7 +2071,7 @@ void hosted_at(const std::string &hostname, std::vector &addrs); std::string append_query_params(const std::string &path, const Params ¶ms); -std::pair make_range_header(Ranges ranges); +std::pair make_range_header(const Ranges &ranges); std::pair make_basic_authentication_header(const std::string &username, @@ -1810,6 +2091,9 @@ std::string trim_copy(const std::string &s); void split(const char *b, const char *e, char d, std::function fn); +void split(const char *b, const char *e, char d, size_t m, + std::function fn); + bool process_client_socket(socket_t sock, time_t read_timeout_sec, time_t read_timeout_usec, time_t write_timeout_sec, time_t write_timeout_usec, @@ -1844,7 +2128,7 @@ enum class EncodingType { None = 0, Gzip, Brotli }; EncodingType encoding_type(const Request &req, const Response &res); -class BufferStream : public Stream { +class BufferStream final : public Stream { public: BufferStream() = default; ~BufferStream() override = default; @@ -1884,19 +2168,19 @@ public: Callback callback) = 0; }; -class nocompressor : public compressor { +class nocompressor final : public compressor { public: - virtual ~nocompressor() = default; + ~nocompressor() override = default; bool compress(const char *data, size_t data_length, bool /*last*/, Callback callback) override; }; #ifdef CPPHTTPLIB_ZLIB_SUPPORT -class gzip_compressor : public compressor { +class gzip_compressor final : public compressor { public: gzip_compressor(); - ~gzip_compressor(); + ~gzip_compressor() override; bool compress(const char *data, size_t data_length, bool last, Callback callback) override; @@ -1906,10 +2190,10 @@ private: z_stream strm_; }; -class gzip_decompressor : public decompressor { +class gzip_decompressor final : public decompressor { public: gzip_decompressor(); - ~gzip_decompressor(); + ~gzip_decompressor() override; bool is_valid() const override; @@ -1923,7 +2207,7 @@ private: #endif #ifdef CPPHTTPLIB_BROTLI_SUPPORT -class brotli_compressor : public compressor { +class brotli_compressor final : public compressor { public: brotli_compressor(); ~brotli_compressor(); @@ -1935,7 +2219,7 @@ private: BrotliEncoderState *state_ = nullptr; }; -class brotli_decompressor : public decompressor { +class brotli_decompressor final : public decompressor { public: brotli_decompressor(); ~brotli_decompressor(); @@ -1972,6 +2256,29 @@ private: std::string glowable_buffer_; }; +class mmap { +public: + mmap(const char *path); + ~mmap(); + + bool open(const char *path); + void close(); + + bool is_open() const; + size_t size() const; + const char *data() const; + +private: +#if defined(_WIN32) + HANDLE hFile_; + HANDLE hMapping_; +#else + int fd_; +#endif + size_t size_; + void *addr_; +}; + } // namespace detail // ---------------------------------------------------------------------------- @@ -2003,7 +2310,7 @@ inline bool from_hex_to_i(const std::string &s, size_t i, size_t cnt, val = 0; for (; cnt; i++, cnt--) { if (!s[i]) { return false; } - int v = 0; + auto v = 0; if (is_hex(s[i], v)) { val = val * 16 + v; } else { @@ -2014,7 +2321,7 @@ inline bool from_hex_to_i(const std::string &s, size_t i, size_t cnt, } inline std::string from_i_to_hex(size_t n) { - const char *charset = "0123456789abcdef"; + static const auto charset = "0123456789abcdef"; std::string ret; do { ret = charset[n & 15] + ret; @@ -2025,7 +2332,7 @@ inline std::string from_i_to_hex(size_t n) { inline size_t to_utf8(int code, char *buff) { if (code < 0x0080) { - buff[0] = (code & 0x7F); + buff[0] = static_cast(code & 0x7F); return 1; } else if (code < 0x0800) { buff[0] = static_cast(0xC0 | ((code >> 6) & 0x1F)); @@ -2064,8 +2371,8 @@ inline std::string base64_encode(const std::string &in) { std::string out; out.reserve(in.size()); - int val = 0; - int valb = -6; + auto val = 0; + auto valb = -6; for (auto c : in) { val = (val << 8) + static_cast(c); @@ -2112,6 +2419,11 @@ inline bool is_valid_path(const std::string &path) { // Read component auto beg = i; while (i < path.size() && path[i] != '/') { + if (path[i] == '\0') { + return false; + } else if (path[i] == '\\') { + return false; + } i++; } @@ -2196,7 +2508,7 @@ inline std::string decode_url(const std::string &s, for (size_t i = 0; i < s.size(); i++) { if (s[i] == '%' && i + 1 < s.size()) { if (s[i + 1] == 'u') { - int val = 0; + auto val = 0; if (from_hex_to_i(s, i + 2, 4, val)) { // 4 digits Unicode codes char buff[4]; @@ -2207,7 +2519,7 @@ inline std::string decode_url(const std::string &s, result += s[i]; } } else { - int val = 0; + auto val = 0; if (from_hex_to_i(s, i + 1, 2, val)) { // 2 digits hex codes result += static_cast(val); @@ -2260,16 +2572,30 @@ inline std::string trim_copy(const std::string &s) { return s.substr(r.first, r.second - r.first); } +inline std::string trim_double_quotes_copy(const std::string &s) { + if (s.length() >= 2 && s.front() == '"' && s.back() == '"') { + return s.substr(1, s.size() - 2); + } + return s; +} + inline void split(const char *b, const char *e, char d, std::function fn) { + return split(b, e, d, (std::numeric_limits::max)(), std::move(fn)); +} + +inline void split(const char *b, const char *e, char d, size_t m, + std::function fn) { size_t i = 0; size_t beg = 0; + size_t count = 1; while (e ? (b + i < e) : (b[i] != '\0')) { - if (b[i] == d) { + if (b[i] == d && count < m) { auto r = trim(b, e, beg, i); if (r.first < r.second) { fn(&b[r.first], &b[r.second]); } beg = i + 1; + count++; } i++; } @@ -2345,6 +2671,105 @@ inline void stream_line_reader::append(char c) { } } +inline mmap::mmap(const char *path) +#if defined(_WIN32) + : hFile_(NULL), hMapping_(NULL) +#else + : fd_(-1) +#endif + , + size_(0), addr_(nullptr) { + open(path); +} + +inline mmap::~mmap() { close(); } + +inline bool mmap::open(const char *path) { + close(); + +#if defined(_WIN32) + std::wstring wpath; + for (size_t i = 0; i < strlen(path); i++) { + wpath += path[i]; + } + + hFile_ = ::CreateFile2(wpath.c_str(), GENERIC_READ, FILE_SHARE_READ, + OPEN_EXISTING, NULL); + + if (hFile_ == INVALID_HANDLE_VALUE) { return false; } + + LARGE_INTEGER size{}; + if (!::GetFileSizeEx(hFile_, &size)) { return false; } + size_ = static_cast(size.QuadPart); + + hMapping_ = + ::CreateFileMappingFromApp(hFile_, NULL, PAGE_READONLY, size_, NULL); + + if (hMapping_ == NULL) { + close(); + return false; + } + + addr_ = ::MapViewOfFileFromApp(hMapping_, FILE_MAP_READ, 0, 0); +#else + fd_ = ::open(path, O_RDONLY); + if (fd_ == -1) { return false; } + + struct stat sb; + if (fstat(fd_, &sb) == -1) { + close(); + return false; + } + size_ = static_cast(sb.st_size); + + addr_ = ::mmap(NULL, size_, PROT_READ, MAP_PRIVATE, fd_, 0); +#endif + + if (addr_ == nullptr) { + close(); + return false; + } + + return true; +} + +inline bool mmap::is_open() const { return addr_ != nullptr; } + +inline size_t mmap::size() const { return size_; } + +inline const char *mmap::data() const { + return static_cast(addr_); +} + +inline void mmap::close() { +#if defined(_WIN32) + if (addr_) { + ::UnmapViewOfFile(addr_); + addr_ = nullptr; + } + + if (hMapping_) { + ::CloseHandle(hMapping_); + hMapping_ = NULL; + } + + if (hFile_ != INVALID_HANDLE_VALUE) { + ::CloseHandle(hFile_); + hFile_ = INVALID_HANDLE_VALUE; + } +#else + if (addr_ != nullptr) { + munmap(addr_, size_); + addr_ = nullptr; + } + + if (fd_ != -1) { + ::close(fd_); + fd_ = -1; + } +#endif + size_ = 0; +} inline int close_socket(socket_t sock) { #ifdef _WIN32 return closesocket(sock); @@ -2354,7 +2779,7 @@ inline int close_socket(socket_t sock) { } template inline ssize_t handle_EINTR(T fn) { - ssize_t res = false; + ssize_t res = 0; while (true) { res = fn(); if (res < 0 && errno == EINTR) { continue; } @@ -2399,7 +2824,7 @@ inline ssize_t select_read(socket_t sock, time_t sec, time_t usec) { return handle_EINTR([&]() { return poll(&pfd_read, 1, timeout); }); #else #ifndef _WIN32 - if (sock >= FD_SETSIZE) { return 1; } + if (sock >= FD_SETSIZE) { return -1; } #endif fd_set fds; @@ -2427,7 +2852,7 @@ inline ssize_t select_write(socket_t sock, time_t sec, time_t usec) { return handle_EINTR([&]() { return poll(&pfd_read, 1, timeout); }); #else #ifndef _WIN32 - if (sock >= FD_SETSIZE) { return 1; } + if (sock >= FD_SETSIZE) { return -1; } #endif fd_set fds; @@ -2458,7 +2883,7 @@ inline Error wait_until_socket_is_ready(socket_t sock, time_t sec, if (poll_res == 0) { return Error::ConnectionTimeout; } if (poll_res > 0 && pfd_read.revents & (POLLIN | POLLOUT)) { - int error = 0; + auto error = 0; socklen_t len = sizeof(error); auto res = getsockopt(sock, SOL_SOCKET, SO_ERROR, reinterpret_cast(&error), &len); @@ -2490,7 +2915,7 @@ inline Error wait_until_socket_is_ready(socket_t sock, time_t sec, if (ret == 0) { return Error::ConnectionTimeout; } if (ret > 0 && (FD_ISSET(sock, &fdsr) || FD_ISSET(sock, &fdsw))) { - int error = 0; + auto error = 0; socklen_t len = sizeof(error); auto res = getsockopt(sock, SOL_SOCKET, SO_ERROR, reinterpret_cast(&error), &len); @@ -2512,7 +2937,7 @@ inline bool is_socket_alive(socket_t sock) { return detail::read_socket(sock, &buf[0], sizeof(buf), MSG_PEEK) > 0; } -class SocketStream : public Stream { +class SocketStream final : public Stream { public: SocketStream(socket_t sock, time_t read_timeout_sec, time_t read_timeout_usec, time_t write_timeout_sec, time_t write_timeout_usec); @@ -2537,11 +2962,11 @@ private: size_t read_buff_off_ = 0; size_t read_buff_content_size_ = 0; - static const size_t read_buff_size_ = 1024 * 4; + static const size_t read_buff_size_ = 1024l * 4; }; #ifdef CPPHTTPLIB_OPENSSL_SUPPORT -class SSLSocketStream : public Stream { +class SSLSocketStream final : public Stream { public: SSLSocketStream(socket_t sock, SSL *ssl, time_t read_timeout_sec, time_t read_timeout_usec, time_t write_timeout_sec, @@ -2666,7 +3091,7 @@ socket_t create_socket(const std::string &host, const std::string &ip, int port, #ifndef _WIN32 if (hints.ai_family == AF_UNIX) { const auto addrlen = host.length(); - if (addrlen > sizeof(sockaddr_un::sun_path)) return INVALID_SOCKET; + if (addrlen > sizeof(sockaddr_un::sun_path)) { return INVALID_SOCKET; } auto sock = socket(hints.ai_family, hints.ai_socktype, hints.ai_protocol); if (sock != INVALID_SOCKET) { @@ -2735,17 +3160,27 @@ socket_t create_socket(const std::string &host, const std::string &ip, int port, #endif if (tcp_nodelay) { - int yes = 1; - setsockopt(sock, IPPROTO_TCP, TCP_NODELAY, reinterpret_cast(&yes), - sizeof(yes)); + auto yes = 1; +#ifdef _WIN32 + setsockopt(sock, IPPROTO_TCP, TCP_NODELAY, + reinterpret_cast(&yes), sizeof(yes)); +#else + setsockopt(sock, IPPROTO_TCP, TCP_NODELAY, + reinterpret_cast(&yes), sizeof(yes)); +#endif } if (socket_options) { socket_options(sock); } if (rp->ai_family == AF_INET6) { - int no = 0; - setsockopt(sock, IPPROTO_IPV6, IPV6_V6ONLY, reinterpret_cast(&no), - sizeof(no)); + auto no = 0; +#ifdef _WIN32 + setsockopt(sock, IPPROTO_IPV6, IPV6_V6ONLY, + reinterpret_cast(&no), sizeof(no)); +#else + setsockopt(sock, IPPROTO_IPV6, IPV6_V6ONLY, + reinterpret_cast(&no), sizeof(no)); +#endif } // bind or connect @@ -2804,7 +3239,7 @@ inline bool bind_ip_address(socket_t sock, const std::string &host) { return ret; } -#if !defined _WIN32 && !defined ANDROID && !defined _AIX +#if !defined _WIN32 && !defined ANDROID && !defined _AIX && !defined __MVS__ #define USE_IF2IP #endif @@ -2860,7 +3295,7 @@ inline socket_t create_client_socket( #ifdef USE_IF2IP auto ip_from_if = if2ip(address_family, intf); if (ip_from_if.empty()) { ip_from_if = intf; } - if (!bind_ip_address(sock2, ip_from_if.c_str())) { + if (!bind_ip_address(sock2, ip_from_if)) { error = Error::BindIPAddress; return false; } @@ -2888,13 +3323,14 @@ inline socket_t create_client_socket( #ifdef _WIN32 auto timeout = static_cast(read_timeout_sec * 1000 + read_timeout_usec / 1000); - setsockopt(sock2, SOL_SOCKET, SO_RCVTIMEO, (char *)&timeout, - sizeof(timeout)); + setsockopt(sock2, SOL_SOCKET, SO_RCVTIMEO, + reinterpret_cast(&timeout), sizeof(timeout)); #else timeval tv; tv.tv_sec = static_cast(read_timeout_sec); tv.tv_usec = static_cast(read_timeout_usec); - setsockopt(sock2, SOL_SOCKET, SO_RCVTIMEO, (char *)&tv, sizeof(tv)); + setsockopt(sock2, SOL_SOCKET, SO_RCVTIMEO, + reinterpret_cast(&tv), sizeof(tv)); #endif } { @@ -2902,13 +3338,14 @@ inline socket_t create_client_socket( #ifdef _WIN32 auto timeout = static_cast(write_timeout_sec * 1000 + write_timeout_usec / 1000); - setsockopt(sock2, SOL_SOCKET, SO_SNDTIMEO, (char *)&timeout, - sizeof(timeout)); + setsockopt(sock2, SOL_SOCKET, SO_SNDTIMEO, + reinterpret_cast(&timeout), sizeof(timeout)); #else timeval tv; tv.tv_sec = static_cast(write_timeout_sec); tv.tv_usec = static_cast(write_timeout_usec); - setsockopt(sock2, SOL_SOCKET, SO_SNDTIMEO, (char *)&tv, sizeof(tv)); + setsockopt(sock2, SOL_SOCKET, SO_SNDTIMEO, + reinterpret_cast(&tv), sizeof(tv)); #endif } @@ -3008,18 +3445,20 @@ inline constexpr unsigned int operator"" _t(const char *s, size_t l) { } // namespace udl -inline const char * +inline std::string find_content_type(const std::string &path, - const std::map &user_data) { + const std::map &user_data, + const std::string &default_content_type) { auto ext = file_extension(path); auto it = user_data.find(ext); - if (it != user_data.end()) { return it->second.c_str(); } + if (it != user_data.end()) { return it->second; } using udl::operator""_t; switch (str2tag(ext)) { - default: return nullptr; + default: return default_content_type; + case "css"_t: return "text/css"; case "csv"_t: return "text/csv"; case "htm"_t: @@ -3072,76 +3511,6 @@ find_content_type(const std::string &path, } } -inline const char *status_message(int status) { - switch (status) { - case 100: return "Continue"; - case 101: return "Switching Protocol"; - case 102: return "Processing"; - case 103: return "Early Hints"; - case 200: return "OK"; - case 201: return "Created"; - case 202: return "Accepted"; - case 203: return "Non-Authoritative Information"; - case 204: return "No Content"; - case 205: return "Reset Content"; - case 206: return "Partial Content"; - case 207: return "Multi-Status"; - case 208: return "Already Reported"; - case 226: return "IM Used"; - case 300: return "Multiple Choice"; - case 301: return "Moved Permanently"; - case 302: return "Found"; - case 303: return "See Other"; - case 304: return "Not Modified"; - case 305: return "Use Proxy"; - case 306: return "unused"; - case 307: return "Temporary Redirect"; - case 308: return "Permanent Redirect"; - case 400: return "Bad Request"; - case 401: return "Unauthorized"; - case 402: return "Payment Required"; - case 403: return "Forbidden"; - case 404: return "Not Found"; - case 405: return "Method Not Allowed"; - case 406: return "Not Acceptable"; - case 407: return "Proxy Authentication Required"; - case 408: return "Request Timeout"; - case 409: return "Conflict"; - case 410: return "Gone"; - case 411: return "Length Required"; - case 412: return "Precondition Failed"; - case 413: return "Payload Too Large"; - case 414: return "URI Too Long"; - case 415: return "Unsupported Media Type"; - case 416: return "Range Not Satisfiable"; - case 417: return "Expectation Failed"; - case 418: return "I'm a teapot"; - case 421: return "Misdirected Request"; - case 422: return "Unprocessable Entity"; - case 423: return "Locked"; - case 424: return "Failed Dependency"; - case 425: return "Too Early"; - case 426: return "Upgrade Required"; - case 428: return "Precondition Required"; - case 429: return "Too Many Requests"; - case 431: return "Request Header Fields Too Large"; - case 451: return "Unavailable For Legal Reasons"; - case 501: return "Not Implemented"; - case 502: return "Bad Gateway"; - case 503: return "Service Unavailable"; - case 504: return "Gateway Timeout"; - case 505: return "HTTP Version Not Supported"; - case 506: return "Variant Also Negotiates"; - case 507: return "Insufficient Storage"; - case 508: return "Loop Detected"; - case 510: return "Not Extended"; - case 511: return "Network Authentication Required"; - - default: - case 500: return "Internal Server Error"; - } -} - inline bool can_compress_content_type(const std::string &content_type) { using udl::operator""_t; @@ -3218,7 +3587,7 @@ inline bool gzip_compressor::compress(const char *data, size_t data_length, data += strm_.avail_in; auto flush = (last && data_length == 0) ? Z_FINISH : Z_NO_FLUSH; - int ret = Z_OK; + auto ret = Z_OK; std::array buff{}; do { @@ -3262,7 +3631,7 @@ inline bool gzip_decompressor::decompress(const char *data, size_t data_length, Callback callback) { assert(is_valid_); - int ret = Z_OK; + auto ret = Z_OK; do { constexpr size_t max_avail_in = @@ -3276,16 +3645,12 @@ inline bool gzip_decompressor::decompress(const char *data, size_t data_length, data += strm_.avail_in; std::array buff{}; - while (strm_.avail_in > 0) { + while (strm_.avail_in > 0 && ret == Z_OK) { strm_.avail_out = static_cast(buff.size()); strm_.next_out = reinterpret_cast(buff.data()); - auto prev_avail_in = strm_.avail_in; - ret = inflate(&strm_, Z_NO_FLUSH); - if (prev_avail_in - strm_.avail_in == 0) { return false; } - assert(ret != Z_STREAM_ERROR); switch (ret) { case Z_NEED_DICT: @@ -3298,7 +3663,7 @@ inline bool gzip_decompressor::decompress(const char *data, size_t data_length, } } - if (ret != Z_OK && ret != Z_STREAM_END) return false; + if (ret != Z_OK && ret != Z_STREAM_END) { return false; } } while (data_length > 0); @@ -3367,7 +3732,7 @@ inline bool brotli_decompressor::decompress(const char *data, return 0; } - const uint8_t *next_in = (const uint8_t *)data; + auto next_in = reinterpret_cast(data); size_t avail_in = data_length; size_t total_out; @@ -3437,6 +3802,9 @@ inline bool parse_header(const char *beg, const char *end, T fn) { } if (p < end) { + auto key_len = key_end - beg; + if (!key_len) { return false; } + auto key = std::string(beg, key_end); auto val = compare_case_ignore(key, "Location") ? std::string(p, end) @@ -3526,11 +3894,7 @@ inline bool read_content_without_length(Stream &strm, uint64_t r = 0; for (;;) { auto n = strm.read(buf, CPPHTTPLIB_RECV_BUFSIZ); - if (n < 0) { - return false; - } else if (n == 0) { - return true; - } + if (n <= 0) { return true; } if (!out(buf, static_cast(n), r, 0)) { return false; } r += static_cast(n); @@ -3566,7 +3930,7 @@ inline bool read_content_chunked(Stream &strm, T &x, if (!line_reader.getline()) { return false; } - if (strcmp(line_reader.ptr(), "\r\n")) { return false; } + if (strcmp(line_reader.ptr(), "\r\n") != 0) { return false; } if (!line_reader.getline()) { return false; } } @@ -3576,7 +3940,7 @@ inline bool read_content_chunked(Stream &strm, T &x, // Trailer if (!line_reader.getline()) { return false; } - while (strcmp(line_reader.ptr(), "\r\n")) { + while (strcmp(line_reader.ptr(), "\r\n") != 0) { if (line_reader.size() > CPPHTTPLIB_HEADER_MAX_LENGTH) { return false; } // Exclude line terminator @@ -3595,8 +3959,8 @@ inline bool read_content_chunked(Stream &strm, T &x, } inline bool is_chunked_transfer_encoding(const Headers &headers) { - return !strcasecmp(get_header_value(headers, "Transfer-Encoding", 0, ""), - "chunked"); + return compare_case_ignore( + get_header_value(headers, "Transfer-Encoding", 0, ""), "chunked"); } template @@ -3611,14 +3975,14 @@ bool prepare_content_receiver(T &x, int &status, #ifdef CPPHTTPLIB_ZLIB_SUPPORT decompressor = detail::make_unique(); #else - status = 415; + status = StatusCode::UnsupportedMediaType_415; return false; #endif } else if (encoding.find("br") != std::string::npos) { #ifdef CPPHTTPLIB_BROTLI_SUPPORT decompressor = detail::make_unique(); #else - status = 415; + status = StatusCode::UnsupportedMediaType_415; return false; #endif } @@ -3634,7 +3998,7 @@ bool prepare_content_receiver(T &x, int &status, }; return callback(std::move(out)); } else { - status = 500; + status = StatusCode::InternalServerError_500; return false; } } @@ -3662,7 +4026,7 @@ bool read_content(Stream &strm, T &x, size_t payload_max_length, int &status, } else if (!has_header(x.headers, "Content-Length")) { ret = read_content_without_length(strm, out); } else { - auto len = get_header_value(x.headers, "Content-Length"); + auto len = get_header_value_u64(x.headers, "Content-Length", 0, 0); if (len > payload_max_length) { exceed_payload_max_length = true; skip_content_with_length(strm, len); @@ -3672,7 +4036,10 @@ bool read_content(Stream &strm, T &x, size_t payload_max_length, int &status, } } - if (!ret) { status = exceed_payload_max_length ? 413 : 400; } + if (!ret) { + status = exceed_payload_max_length ? StatusCode::PayloadTooLarge_413 + : StatusCode::BadRequest_400; + } return ret; }); } // namespace detail @@ -3720,6 +4087,8 @@ inline bool write_content(Stream &strm, const ContentProvider &content_provider, return ok; }; + data_sink.is_writable = [&]() -> bool { return strm.is_writable(); }; + while (offset < end_offset && !is_shutting_down()) { if (!strm.is_writable()) { error = Error::Write; @@ -3764,6 +4133,8 @@ write_content_without_length(Stream &strm, return ok; }; + data_sink.is_writable = [&]() -> bool { return strm.is_writable(); }; + data_sink.done = [&](void) { data_available = false; }; while (data_available && !is_shutting_down()) { @@ -3814,6 +4185,8 @@ write_content_chunked(Stream &strm, const ContentProvider &content_provider, return ok; }; + data_sink.is_writable = [&]() -> bool { return strm.is_writable(); }; + auto done_with_trailer = [&](const Headers *trailer) { if (!ok) { return; } @@ -3898,7 +4271,8 @@ inline bool redirect(T &cli, Request &req, Response &res, new_req.path = path; new_req.redirect_count_ -= 1; - if (res.status == 303 && (req.method != "GET" && req.method != "HEAD")) { + if (res.status == StatusCode::SeeOther_303 && + (req.method != "GET" && req.method != "HEAD")) { new_req.method = "GET"; new_req.body.clear(); new_req.headers.clear(); @@ -3910,7 +4284,8 @@ inline bool redirect(T &cli, Request &req, Response &res, if (ret) { req = new_req; res = new_res; - res.location = location; + + if (res.location.empty()) { res.location = location; } } return ret; } @@ -3957,14 +4332,34 @@ inline bool parse_multipart_boundary(const std::string &content_type, if (pos == std::string::npos) { return false; } auto end = content_type.find(';', pos); auto beg = pos + strlen(boundary_keyword); - boundary = content_type.substr(beg, end - beg); - if (boundary.length() >= 2 && boundary.front() == '"' && - boundary.back() == '"') { - boundary = boundary.substr(1, boundary.size() - 2); - } + boundary = trim_double_quotes_copy(content_type.substr(beg, end - beg)); return !boundary.empty(); } +inline void parse_disposition_params(const std::string &s, Params ¶ms) { + std::set cache; + split(s.data(), s.data() + s.size(), ';', [&](const char *b, const char *e) { + std::string kv(b, e); + if (cache.find(kv) != cache.end()) { return; } + cache.insert(kv); + + std::string key; + std::string val; + split(b, e, '=', [&](const char *b2, const char *e2) { + if (key.empty()) { + key.assign(b2, e2); + } else { + val.assign(b2, e2); + } + }); + + if (!key.empty()) { + params.emplace(trim_double_quotes_copy((key)), + trim_double_quotes_copy((val))); + } + }); +} + #ifdef CPPHTTPLIB_NO_EXCEPTIONS inline bool parse_range_header(const std::string &s, Ranges &ranges) { #else @@ -3975,9 +4370,9 @@ inline bool parse_range_header(const std::string &s, Ranges &ranges) try { if (std::regex_match(s, m, re_first_range)) { auto pos = static_cast(m.position(1)); auto len = static_cast(m.length(1)); - bool all_valid_ranges = true; + auto all_valid_ranges = true; split(&s[pos], &s[pos + len], ',', [&](const char *b, const char *e) { - if (!all_valid_ranges) return; + if (!all_valid_ranges) { return; } static auto re_another_range = std::regex(R"(\s*(\d*)-(\d*))"); std::cmatch cm; if (std::regex_match(b, e, cm, re_another_range)) { @@ -4022,11 +4417,6 @@ public: bool parse(const char *buf, size_t n, const ContentReceiver &content_callback, const MultipartContentHeader &header_callback) { - // TODO: support 'filename*' - static const std::regex re_content_disposition( - R"~(^Content-Disposition:\s*form-data;\s*name="(.*?)"(?:;\s*filename="(.*?)")?(?:;\s*filename\*=\S+)?\s*$)~", - std::regex_constants::icase); - buf_append(buf, n); while (buf_size() > 0) { @@ -4059,18 +4449,54 @@ public: break; } - static const std::string header_name = "content-type:"; const auto header = buf_head(pos); - if (start_with_case_ignore(header, header_name)) { - file_.content_type = trim_copy(header.substr(header_name.size())); + + if (!parse_header(header.data(), header.data() + header.size(), + [&](std::string &&, std::string &&) {})) { + is_valid_ = false; + return false; + } + + static const std::string header_content_type = "Content-Type:"; + + if (start_with_case_ignore(header, header_content_type)) { + file_.content_type = + trim_copy(header.substr(header_content_type.size())); } else { + static const std::regex re_content_disposition( + R"~(^Content-Disposition:\s*form-data;\s*(.*)$)~", + std::regex_constants::icase); + std::smatch m; if (std::regex_match(header, m, re_content_disposition)) { - file_.name = m[1]; - file_.filename = m[2]; - } else { - is_valid_ = false; - return false; + Params params; + parse_disposition_params(m[1], params); + + auto it = params.find("name"); + if (it != params.end()) { + file_.name = it->second; + } else { + is_valid_ = false; + return false; + } + + it = params.find("filename"); + if (it != params.end()) { file_.filename = it->second; } + + it = params.find("filename*"); + if (it != params.end()) { + // Only allow UTF-8 enconnding... + static const std::regex re_rfc5987_encoding( + R"~(^UTF-8''(.+?)$)~", std::regex_constants::icase); + + std::smatch m2; + if (std::regex_match(it->second, m2, re_rfc5987_encoding)) { + file_.filename = decode_url(m2[1], false); // override... + } else { + is_valid_ = false; + return false; + } + } } } buf_erase(pos + crlf_.size()); @@ -4108,9 +4534,9 @@ public: buf_erase(crlf_.size()); state_ = 1; } else { - if (dash_crlf_.size() > buf_size()) { return true; } - if (buf_start_with(dash_crlf_)) { - buf_erase(dash_crlf_.size()); + if (dash_.size() > buf_size()) { return true; } + if (buf_start_with(dash_)) { + buf_erase(dash_.size()); is_valid_ = true; buf_erase(buf_size()); // Remove epilogue } else { @@ -4143,7 +4569,6 @@ private: const std::string dash_ = "--"; const std::string crlf_ = "\r\n"; - const std::string dash_crlf_ = "--\r\n"; std::string boundary_; std::string dash_boundary_crlf_; std::string crlf_dash_boundary_; @@ -4230,28 +4655,32 @@ inline std::string to_lower(const char *beg, const char *end) { return out; } -inline std::string make_multipart_data_boundary() { +inline std::string random_string(size_t length) { static const char data[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"; // std::random_device might actually be deterministic on some // platforms, but due to lack of support in the c++ standard library, // doing better requires either some ugly hacks or breaking portability. - std::random_device seed_gen; + static std::random_device seed_gen; // Request 128 bits of entropy for initialization - std::seed_seq seed_sequence{seed_gen(), seed_gen(), seed_gen(), seed_gen()}; - std::mt19937 engine(seed_sequence); + static std::seed_seq seed_sequence{seed_gen(), seed_gen(), seed_gen(), + seed_gen()}; - std::string result = "--cpp-httplib-multipart-data-"; + static std::mt19937 engine(seed_sequence); - for (auto i = 0; i < 16; i++) { + std::string result; + for (size_t i = 0; i < length; i++) { result += data[engine() % (sizeof(data) - 1)]; } - return result; } +inline std::string make_multipart_data_boundary() { + return "--cpp-httplib-multipart-data-" + detail::random_string(16); +} + inline bool is_multipart_boundary_chars_valid(const std::string &boundary) { auto valid = true; for (size_t i = 0; i < boundary.size(); i++) { @@ -4304,48 +4733,97 @@ serialize_multipart_formdata(const MultipartFormDataItems &items, body += item.content + serialize_multipart_formdata_item_end(); } - if (finish) body += serialize_multipart_formdata_finish(boundary); + if (finish) { body += serialize_multipart_formdata_finish(boundary); } return body; } +inline bool range_error(Request &req, Response &res) { + if (!req.ranges.empty() && 200 <= res.status && res.status < 300) { + ssize_t contant_len = static_cast( + res.content_length_ ? res.content_length_ : res.body.size()); + + ssize_t prev_first_pos = -1; + ssize_t prev_last_pos = -1; + size_t overwrapping_count = 0; + + // NOTE: The following Range check is based on '14.2. Range' in RFC 9110 + // 'HTTP Semantics' to avoid potential denial-of-service attacks. + // https://www.rfc-editor.org/rfc/rfc9110#section-14.2 + + // Too many ranges + if (req.ranges.size() > CPPHTTPLIB_RANGE_MAX_COUNT) { return true; } + + for (auto &r : req.ranges) { + auto &first_pos = r.first; + auto &last_pos = r.second; + + if (first_pos == -1 && last_pos == -1) { + first_pos = 0; + last_pos = contant_len; + } + + if (first_pos == -1) { + first_pos = contant_len - last_pos; + last_pos = contant_len - 1; + } + + if (last_pos == -1) { last_pos = contant_len - 1; } + + // Range must be within content length + if (!(0 <= first_pos && first_pos <= last_pos && + last_pos <= contant_len - 1)) { + return true; + } + + // Ranges must be in ascending order + if (first_pos <= prev_first_pos) { return true; } + + // Request must not have more than two overlapping ranges + if (first_pos <= prev_last_pos) { + overwrapping_count++; + if (overwrapping_count > 2) { return true; } + } + + prev_first_pos = (std::max)(prev_first_pos, first_pos); + prev_last_pos = (std::max)(prev_last_pos, last_pos); + } + } + + return false; +} + inline std::pair -get_range_offset_and_length(const Request &req, size_t content_length, - size_t index) { - auto r = req.ranges[index]; +get_range_offset_and_length(Range r, size_t content_length) { + (void)(content_length); // patch to get rid of "unused parameter" on release build + assert(r.first != -1 && r.second != -1); + assert(0 <= r.first && r.first < static_cast(content_length)); + assert(r.first <= r.second && + r.second < static_cast(content_length)); - if (r.first == -1 && r.second == -1) { - return std::make_pair(0, content_length); - } - - auto slen = static_cast(content_length); - - if (r.first == -1) { - r.first = (std::max)(static_cast(0), slen - r.second); - r.second = slen - 1; - } - - if (r.second == -1) { r.second = slen - 1; } return std::make_pair(r.first, static_cast(r.second - r.first) + 1); } -inline std::string make_content_range_header_field(size_t offset, size_t length, - size_t content_length) { +inline std::string make_content_range_header_field( + const std::pair &offset_and_length, size_t content_length) { + auto st = offset_and_length.first; + auto ed = st + offset_and_length.second - 1; + std::string field = "bytes "; - field += std::to_string(offset); + field += std::to_string(st); field += "-"; - field += std::to_string(offset + length - 1); + field += std::to_string(ed); field += "/"; field += std::to_string(content_length); return field; } template -bool process_multipart_ranges_data(const Request &req, Response &res, +bool process_multipart_ranges_data(const Request &req, const std::string &boundary, const std::string &content_type, - SToken stoken, CToken ctoken, - Content content) { + size_t content_length, SToken stoken, + CToken ctoken, Content content) { for (size_t i = 0; i < req.ranges.size(); i++) { ctoken("--"); stoken(boundary); @@ -4356,50 +4834,51 @@ bool process_multipart_ranges_data(const Request &req, Response &res, ctoken("\r\n"); } - auto offsets = get_range_offset_and_length(req, res.body.size(), i); - auto offset = offsets.first; - auto length = offsets.second; + auto offset_and_length = + get_range_offset_and_length(req.ranges[i], content_length); ctoken("Content-Range: "); - stoken(make_content_range_header_field(offset, length, res.body.size())); + stoken(make_content_range_header_field(offset_and_length, content_length)); ctoken("\r\n"); ctoken("\r\n"); - if (!content(offset, length)) { return false; } + + if (!content(offset_and_length.first, offset_and_length.second)) { + return false; + } ctoken("\r\n"); } ctoken("--"); stoken(boundary); - ctoken("--\r\n"); + ctoken("--"); return true; } -inline bool make_multipart_ranges_data(const Request &req, Response &res, +inline void make_multipart_ranges_data(const Request &req, Response &res, const std::string &boundary, const std::string &content_type, + size_t content_length, std::string &data) { - return process_multipart_ranges_data( - req, res, boundary, content_type, + process_multipart_ranges_data( + req, boundary, content_type, content_length, [&](const std::string &token) { data += token; }, [&](const std::string &token) { data += token; }, [&](size_t offset, size_t length) { - if (offset < res.body.size()) { - data += res.body.substr(offset, length); - return true; - } - return false; + assert(offset + length <= content_length); + data += res.body.substr(offset, length); + return true; }); } -inline size_t -get_multipart_ranges_data_length(const Request &req, Response &res, - const std::string &boundary, - const std::string &content_type) { +inline size_t get_multipart_ranges_data_length(const Request &req, + const std::string &boundary, + const std::string &content_type, + size_t content_length) { size_t data_length = 0; process_multipart_ranges_data( - req, res, boundary, content_type, + req, boundary, content_type, content_length, [&](const std::string &token) { data_length += token.size(); }, [&](const std::string &token) { data_length += token.size(); }, [&](size_t /*offset*/, size_t length) { @@ -4411,13 +4890,13 @@ get_multipart_ranges_data_length(const Request &req, Response &res, } template -inline bool write_multipart_ranges_data(Stream &strm, const Request &req, - Response &res, - const std::string &boundary, - const std::string &content_type, - const T &is_shutting_down) { +inline bool +write_multipart_ranges_data(Stream &strm, const Request &req, Response &res, + const std::string &boundary, + const std::string &content_type, + size_t content_length, const T &is_shutting_down) { return process_multipart_ranges_data( - req, res, boundary, content_type, + req, boundary, content_type, content_length, [&](const std::string &token) { strm.write(token); }, [&](const std::string &token) { strm.write(token); }, [&](size_t offset, size_t length) { @@ -4426,18 +4905,6 @@ inline bool write_multipart_ranges_data(Stream &strm, const Request &req, }); } -inline std::pair -get_range_offset_and_length(const Request &req, const Response &res, - size_t index) { - auto r = req.ranges[index]; - - if (r.second == -1) { - r.second = static_cast(res.content_length_) - 1; - } - - return std::make_pair(r.first, r.second - r.first + 1); -} - inline bool expect_content(const Request &req) { if (req.method == "POST" || req.method == "PUT" || req.method == "PATCH" || req.method == "PRI" || req.method == "DELETE") { @@ -4471,7 +4938,7 @@ inline std::string message_digest(const std::string &s, const EVP_MD *algo) { std::stringstream ss; for (auto i = 0u; i < hash_length; ++i) { ss << std::hex << std::setw(2) << std::setfill('0') - << (unsigned int)hash[i]; + << static_cast(hash[i]); } return ss.str(); @@ -4564,7 +5031,7 @@ inline bool retrieve_root_certs_from_keychain(CFObjectPtr &certs) { inline bool add_certs_to_x509_store(CFArrayRef certs, X509_STORE *store) { auto result = false; - for (int i = 0; i < CFArrayGetCount(certs); ++i) { + for (auto i = 0; i < CFArrayGetCount(certs); ++i) { const auto cert = reinterpret_cast( CFArrayGetValueAtIndex(certs, i)); @@ -4707,7 +5174,7 @@ inline bool parse_www_authenticate(const Response &res, s = s.substr(pos + 1); auto beg = std::sregex_iterator(s.begin(), s.end(), re); for (auto i = beg; i != std::sregex_iterator(); ++i) { - auto m = *i; + const auto &m = *i; auto key = s.substr(static_cast(m.position(1)), static_cast(m.length(1))); auto val = m.length(2) > 0 @@ -4724,20 +5191,6 @@ inline bool parse_www_authenticate(const Response &res, return false; } -// https://stackoverflow.com/questions/440133/how-do-i-create-a-random-alpha-numeric-string-in-c/440240#answer-440240 -inline std::string random_string(size_t length) { - auto randchar = []() -> char { - const char charset[] = "0123456789" - "ABCDEFGHIJKLMNOPQRSTUVWXYZ" - "abcdefghijklmnopqrstuvwxyz"; - const size_t max_index = (sizeof(charset) - 1); - return charset[static_cast(std::rand()) % max_index]; - }; - std::string str(length, 0); - std::generate_n(str.begin(), length, randchar); - return str; -} - class ContentProviderAdapter { public: explicit ContentProviderAdapter( @@ -4782,7 +5235,7 @@ inline void hosted_at(const std::string &hostname, const auto &addr = *reinterpret_cast(rp->ai_addr); std::string ip; - int dummy = -1; + auto dummy = -1; if (detail::get_ip_and_port(addr, sizeof(struct sockaddr_storage), ip, dummy)) { addrs.push_back(ip); @@ -4802,10 +5255,11 @@ inline std::string append_query_params(const std::string &path, } // Header utilities -inline std::pair make_range_header(Ranges ranges) { +inline std::pair +make_range_header(const Ranges &ranges) { std::string field = "bytes="; auto i = 0; - for (auto r : ranges) { + for (const auto &r : ranges) { if (i != 0) { field += ", "; } if (r.first != -1) { field += std::to_string(r.first); } field += '-'; @@ -4924,7 +5378,7 @@ inline void Response::set_redirect(const std::string &url, int stat) { if (300 <= stat && stat < 400) { this->status = stat; } else { - this->status = 302; + this->status = StatusCode::Found_302; } } } @@ -4943,13 +5397,22 @@ inline void Response::set_content(const std::string &s, set_content(s.data(), s.size(), content_type); } +inline void Response::set_content(std::string &&s, + const std::string &content_type) { + body = std::move(s); + + auto rng = headers.equal_range("Content-Type"); + headers.erase(rng.first, rng.second); + set_header("Content-Type", content_type); +} + inline void Response::set_content_provider( size_t in_length, const std::string &content_type, ContentProvider provider, ContentProviderResourceReleaser resource_releaser) { set_header("Content-Type", content_type); content_length_ = in_length; if (in_length > 0) { content_provider_ = std::move(provider); } - content_provider_resource_releaser_ = resource_releaser; + content_provider_resource_releaser_ = std::move(resource_releaser); is_chunked_content_provider_ = false; } @@ -4959,7 +5422,7 @@ inline void Response::set_content_provider( set_header("Content-Type", content_type); content_length_ = 0; content_provider_ = detail::ContentProviderAdapter(std::move(provider)); - content_provider_resource_releaser_ = resource_releaser; + content_provider_resource_releaser_ = std::move(resource_releaser); is_chunked_content_provider_ = false; } @@ -4969,7 +5432,7 @@ inline void Response::set_chunked_content_provider( set_header("Content-Type", content_type); content_length_ = 0; content_provider_ = detail::ContentProviderAdapter(std::move(provider)); - content_provider_resource_releaser_ = resource_releaser; + content_provider_resource_releaser_ = std::move(resource_releaser); is_chunked_content_provider_ = true; } @@ -5010,7 +5473,7 @@ inline SocketStream::SocketStream(socket_t sock, time_t read_timeout_sec, write_timeout_sec_(write_timeout_sec), write_timeout_usec_(write_timeout_usec), read_buff_(read_buff_size_, 0) {} -inline SocketStream::~SocketStream() {} +inline SocketStream::~SocketStream() = default; inline bool SocketStream::is_readable() const { return select_read(sock_, read_timeout_sec_, read_timeout_usec_) > 0; @@ -5120,6 +5583,99 @@ inline socket_t BufferStream::socket() const { return 0; } inline const std::string &BufferStream::get_buffer() const { return buffer; } +inline PathParamsMatcher::PathParamsMatcher(const std::string &pattern) { + // One past the last ending position of a path param substring + std::size_t last_param_end = 0; + +#ifndef CPPHTTPLIB_NO_EXCEPTIONS + // Needed to ensure that parameter names are unique during matcher + // construction + // If exceptions are disabled, only last duplicate path + // parameter will be set + std::unordered_set param_name_set; +#endif + + while (true) { + const auto marker_pos = pattern.find(marker, last_param_end); + if (marker_pos == std::string::npos) { break; } + + static_fragments_.push_back( + pattern.substr(last_param_end, marker_pos - last_param_end)); + + const auto param_name_start = marker_pos + 1; + + auto sep_pos = pattern.find(separator, param_name_start); + if (sep_pos == std::string::npos) { sep_pos = pattern.length(); } + + auto param_name = + pattern.substr(param_name_start, sep_pos - param_name_start); + +#ifndef CPPHTTPLIB_NO_EXCEPTIONS + if (param_name_set.find(param_name) != param_name_set.cend()) { + std::string msg = "Encountered path parameter '" + param_name + + "' multiple times in route pattern '" + pattern + "'."; + throw std::invalid_argument(msg); + } +#endif + + param_names_.push_back(std::move(param_name)); + + last_param_end = sep_pos + 1; + } + + if (last_param_end < pattern.length()) { + static_fragments_.push_back(pattern.substr(last_param_end)); + } +} + +inline bool PathParamsMatcher::match(Request &request) const { + request.matches = std::smatch(); + request.path_params.clear(); + request.path_params.reserve(param_names_.size()); + + // One past the position at which the path matched the pattern last time + std::size_t starting_pos = 0; + for (size_t i = 0; i < static_fragments_.size(); ++i) { + const auto &fragment = static_fragments_[i]; + + if (starting_pos + fragment.length() > request.path.length()) { + return false; + } + + // Avoid unnecessary allocation by using strncmp instead of substr + + // comparison + if (std::strncmp(request.path.c_str() + starting_pos, fragment.c_str(), + fragment.length()) != 0) { + return false; + } + + starting_pos += fragment.length(); + + // Should only happen when we have a static fragment after a param + // Example: '/users/:id/subscriptions' + // The 'subscriptions' fragment here does not have a corresponding param + if (i >= param_names_.size()) { continue; } + + auto sep_pos = request.path.find(separator, starting_pos); + if (sep_pos == std::string::npos) { sep_pos = request.path.length(); } + + const auto ¶m_name = param_names_[i]; + + request.path_params.emplace( + param_name, request.path.substr(starting_pos, sep_pos - starting_pos)); + + // Mark everythin up to '/' as matched + starting_pos = sep_pos + 1; + } + // Returns false if the path is longer than the pattern + return starting_pos >= request.path.length(); +} + +inline bool RegexMatcher::match(Request &request) const { + request.path_params.clear(); + return std::regex_match(request.path, request.matches, regex_); +} + } // namespace detail // HTTP server implementation @@ -5131,69 +5687,72 @@ inline Server::Server() #endif } -inline Server::~Server() {} +inline Server::~Server() = default; + +inline std::unique_ptr +Server::make_matcher(const std::string &pattern) { + if (pattern.find("/:") != std::string::npos) { + return detail::make_unique(pattern); + } else { + return detail::make_unique(pattern); + } +} inline Server &Server::Get(const std::string &pattern, Handler handler) { - get_handlers_.push_back( - std::make_pair(std::regex(pattern), std::move(handler))); + get_handlers_.emplace_back(make_matcher(pattern), std::move(handler)); return *this; } inline Server &Server::Post(const std::string &pattern, Handler handler) { - post_handlers_.push_back( - std::make_pair(std::regex(pattern), std::move(handler))); + post_handlers_.emplace_back(make_matcher(pattern), std::move(handler)); return *this; } inline Server &Server::Post(const std::string &pattern, HandlerWithContentReader handler) { - post_handlers_for_content_reader_.push_back( - std::make_pair(std::regex(pattern), std::move(handler))); + post_handlers_for_content_reader_.emplace_back(make_matcher(pattern), + std::move(handler)); return *this; } inline Server &Server::Put(const std::string &pattern, Handler handler) { - put_handlers_.push_back( - std::make_pair(std::regex(pattern), std::move(handler))); + put_handlers_.emplace_back(make_matcher(pattern), std::move(handler)); return *this; } inline Server &Server::Put(const std::string &pattern, HandlerWithContentReader handler) { - put_handlers_for_content_reader_.push_back( - std::make_pair(std::regex(pattern), std::move(handler))); + put_handlers_for_content_reader_.emplace_back(make_matcher(pattern), + std::move(handler)); return *this; } inline Server &Server::Patch(const std::string &pattern, Handler handler) { - patch_handlers_.push_back( - std::make_pair(std::regex(pattern), std::move(handler))); + patch_handlers_.emplace_back(make_matcher(pattern), std::move(handler)); return *this; } inline Server &Server::Patch(const std::string &pattern, HandlerWithContentReader handler) { - patch_handlers_for_content_reader_.push_back( - std::make_pair(std::regex(pattern), std::move(handler))); + patch_handlers_for_content_reader_.emplace_back(make_matcher(pattern), + std::move(handler)); return *this; } inline Server &Server::Delete(const std::string &pattern, Handler handler) { - delete_handlers_.push_back( - std::make_pair(std::regex(pattern), std::move(handler))); + delete_handlers_.emplace_back(make_matcher(pattern), std::move(handler)); return *this; } inline Server &Server::Delete(const std::string &pattern, HandlerWithContentReader handler) { - delete_handlers_for_content_reader_.push_back( - std::make_pair(std::regex(pattern), std::move(handler))); + delete_handlers_for_content_reader_.emplace_back(make_matcher(pattern), + std::move(handler)); return *this; } inline Server &Server::Options(const std::string &pattern, Handler handler) { - options_handlers_.push_back( - std::make_pair(std::regex(pattern), std::move(handler))); + options_handlers_.emplace_back(make_matcher(pattern), std::move(handler)); return *this; } @@ -5231,6 +5790,11 @@ Server::set_file_extension_and_mimetype_mapping(const std::string &ext, return *this; } +inline Server &Server::set_default_file_mimetype(const std::string &mime) { + default_file_mimetype_ = mime; + return *this; +} + inline Server &Server::set_file_request_handler(Handler handler) { file_request_handler_ = std::move(handler); return *this; @@ -5272,7 +5836,6 @@ inline Server &Server::set_logger(Logger logger) { inline Server & Server::set_expect_100_continue_handler(Expect100ContinueHandler handler) { expect_100_continue_handler_ = std::move(handler); - return *this; } @@ -5296,6 +5859,12 @@ inline Server &Server::set_default_headers(Headers headers) { return *this; } +inline Server &Server::set_header_writer( + std::function const &writer) { + header_writer_ = writer; + return *this; +} + inline Server &Server::set_keep_alive_max_count(size_t count) { keep_alive_max_count_ = count; return *this; @@ -5331,8 +5900,7 @@ inline Server &Server::set_payload_max_length(size_t length) { inline bool Server::bind_to_port(const std::string &host, int port, int socket_flags) { - if (bind_internal(host, port, socket_flags) < 0) return false; - return true; + return bind_internal(host, port, socket_flags) >= 0; } inline int Server::bind_to_any_port(const std::string &host, int socket_flags) { return bind_internal(host, 0, socket_flags); @@ -5366,7 +5934,7 @@ inline void Server::stop() { } } -inline bool Server::parse_request_line(const char *s, Request &req) { +inline bool Server::parse_request_line(const char *s, Request &req) const { auto len = strlen(s); if (len < 2 || s[len - 2] != '\r' || s[len - 1] != '\n') { return false; } len -= 2; @@ -5407,7 +5975,7 @@ inline bool Server::parse_request_line(const char *s, Request &req) { size_t count = 0; detail::split(req.target.data(), req.target.data() + req.target.size(), '?', - [&](const char *b, const char *e) { + 2, [&](const char *b, const char *e) { switch (count) { case 0: req.path = detail::decode_url(std::string(b, e), false); @@ -5430,7 +5998,10 @@ inline bool Server::parse_request_line(const char *s, Request &req) { } inline bool Server::write_response(Stream &strm, bool close_connection, - const Request &req, Response &res) { + Request &req, Response &res) { + // NOTE: `req.ranges` should be empty, otherwise it will be applied + // incorrectly to the error content. + req.ranges.clear(); return write_response_core(strm, close_connection, req, res, false); } @@ -5486,11 +6057,11 @@ inline bool Server::write_response_core(Stream &strm, bool close_connection, detail::BufferStream bstrm; if (!bstrm.write_format("HTTP/1.1 %d %s\r\n", res.status, - detail::status_message(res.status))) { + status_message(res.status))) { return false; } - if (!detail::write_headers(bstrm, res.headers)) { return false; } + if (!header_writer_(bstrm, res.headers)) { return false; } // Flush buffer auto &data = bstrm.get_buffer(); @@ -5508,7 +6079,6 @@ inline bool Server::write_response_core(Stream &strm, bool close_connection, if (write_content_with_provider(strm, req, res, boundary, content_type)) { res.content_provider_success_ = true; } else { - res.content_provider_success_ = false; ret = false; } } @@ -5533,15 +6103,16 @@ Server::write_content_with_provider(Stream &strm, const Request &req, return detail::write_content(strm, res.content_provider_, 0, res.content_length_, is_shutting_down); } else if (req.ranges.size() == 1) { - auto offsets = - detail::get_range_offset_and_length(req, res.content_length_, 0); - auto offset = offsets.first; - auto length = offsets.second; - return detail::write_content(strm, res.content_provider_, offset, length, - is_shutting_down); + auto offset_and_length = detail::get_range_offset_and_length( + req.ranges[0], res.content_length_); + + return detail::write_content(strm, res.content_provider_, + offset_and_length.first, + offset_and_length.second, is_shutting_down); } else { return detail::write_multipart_ranges_data( - strm, req, res, boundary, content_type, is_shutting_down); + strm, req, res, boundary, content_type, res.content_length_, + is_shutting_down); } } else { if (res.is_chunked_content_provider_) { @@ -5598,7 +6169,7 @@ inline bool Server::read_content(Stream &strm, Request &req, Response &res) { const auto &content_type = req.get_header_value("Content-Type"); if (!content_type.find("application/x-www-form-urlencoded")) { if (req.body.size() > CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH) { - res.status = 413; // NOTE: should be 414? + res.status = StatusCode::PayloadTooLarge_413; // NOTE: should be 414? return false; } detail::parse_query_text(req.body, req.params); @@ -5617,10 +6188,11 @@ inline bool Server::read_content_with_content_receiver( std::move(multipart_receiver)); } -inline bool Server::read_content_core(Stream &strm, Request &req, Response &res, - ContentReceiver receiver, - MultipartContentHeader multipart_header, - ContentReceiver multipart_receiver) { +inline bool +Server::read_content_core(Stream &strm, Request &req, Response &res, + ContentReceiver receiver, + MultipartContentHeader multipart_header, + ContentReceiver multipart_receiver) const { detail::MultipartFormDataParser multipart_form_data_parser; ContentReceiverWithProgress out; @@ -5628,7 +6200,7 @@ inline bool Server::read_content_core(Stream &strm, Request &req, Response &res, const auto &content_type = req.get_header_value("Content-Type"); std::string boundary; if (!detail::parse_multipart_boundary(content_type, boundary)) { - res.status = 400; + res.status = StatusCode::BadRequest_400; return false; } @@ -5664,7 +6236,7 @@ inline bool Server::read_content_core(Stream &strm, Request &req, Response &res, if (req.is_multipart_form_data()) { if (!multipart_form_data_parser.is_valid()) { - res.status = 400; + res.status = StatusCode::BadRequest_400; return false; } } @@ -5683,17 +6255,26 @@ inline bool Server::handle_file_request(const Request &req, Response &res, if (path.back() == '/') { path += "index.html"; } if (detail::is_file(path)) { - detail::read_file(path, res.body); - auto type = - detail::find_content_type(path, file_extension_and_mimetype_map_); - if (type) { res.set_header("Content-Type", type); } for (const auto &kv : entry.headers) { - res.set_header(kv.first.c_str(), kv.second); + res.set_header(kv.first, kv.second); } - res.status = req.has_header("Range") ? 206 : 200; + + auto mm = std::make_shared(path.c_str()); + if (!mm->is_open()) { return false; } + + res.set_content_provider( + mm->size(), + detail::find_content_type(path, file_extension_and_mimetype_map_, + default_file_mimetype_), + [mm](size_t offset, size_t length, DataSink &sink) -> bool { + sink.write(mm->data() + offset, length); + return true; + }); + if (!head && file_request_handler_) { file_request_handler_(req, res); } + return true; } } @@ -5789,13 +6370,14 @@ inline bool Server::listen_internal() { #ifdef _WIN32 auto timeout = static_cast(read_timeout_sec_ * 1000 + read_timeout_usec_ / 1000); - setsockopt(sock, SOL_SOCKET, SO_RCVTIMEO, (char *)&timeout, - sizeof(timeout)); + setsockopt(sock, SOL_SOCKET, SO_RCVTIMEO, + reinterpret_cast(&timeout), sizeof(timeout)); #else timeval tv; tv.tv_sec = static_cast(read_timeout_sec_); tv.tv_usec = static_cast(read_timeout_usec_); - setsockopt(sock, SOL_SOCKET, SO_RCVTIMEO, (char *)&tv, sizeof(tv)); + setsockopt(sock, SOL_SOCKET, SO_RCVTIMEO, + reinterpret_cast(&tv), sizeof(tv)); #endif } { @@ -5803,17 +6385,22 @@ inline bool Server::listen_internal() { #ifdef _WIN32 auto timeout = static_cast(write_timeout_sec_ * 1000 + write_timeout_usec_ / 1000); - setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO, (char *)&timeout, - sizeof(timeout)); + setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO, + reinterpret_cast(&timeout), sizeof(timeout)); #else timeval tv; tv.tv_sec = static_cast(write_timeout_sec_); tv.tv_usec = static_cast(write_timeout_usec_); - setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO, (char *)&tv, sizeof(tv)); + setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO, + reinterpret_cast(&tv), sizeof(tv)); #endif } - task_queue->enqueue([this, sock]() { process_and_close_socket(sock); }); + if (!task_queue->enqueue( + [this, sock]() { process_and_close_socket(sock); })) { + detail::shutdown_socket(sock); + detail::close_socket(sock); + } } task_queue->shutdown(); @@ -5829,7 +6416,7 @@ inline bool Server::routing(Request &req, Response &res, Stream &strm) { } // File handler - bool is_head_request = req.method == "HEAD"; + auto is_head_request = req.method == "HEAD"; if ((req.method == "GET" || is_head_request) && handle_file_request(req, res, is_head_request)) { return true; @@ -5895,17 +6482,17 @@ inline bool Server::routing(Request &req, Response &res, Stream &strm) { return dispatch_request(req, res, patch_handlers_); } - res.status = 400; + res.status = StatusCode::BadRequest_400; return false; } inline bool Server::dispatch_request(Request &req, Response &res, - const Handlers &handlers) { + const Handlers &handlers) const { for (const auto &x : handlers) { - const auto &pattern = x.first; + const auto &matcher = x.first; const auto &handler = x.second; - if (std::regex_match(req.path, req.matches, pattern)) { + if (matcher->match(req)) { handler(req, res); return true; } @@ -5915,18 +6502,18 @@ inline bool Server::dispatch_request(Request &req, Response &res, inline void Server::apply_ranges(const Request &req, Response &res, std::string &content_type, - std::string &boundary) { + std::string &boundary) const { if (req.ranges.size() > 1) { - boundary = detail::make_multipart_data_boundary(); - auto it = res.headers.find("Content-Type"); if (it != res.headers.end()) { content_type = it->second; res.headers.erase(it); } - res.headers.emplace("Content-Type", - "multipart/byteranges; boundary=" + boundary); + boundary = detail::make_multipart_data_boundary(); + + res.set_header("Content-Type", + "multipart/byteranges; boundary=" + boundary); } auto type = detail::encoding_type(req, res); @@ -5937,16 +6524,17 @@ inline void Server::apply_ranges(const Request &req, Response &res, if (req.ranges.empty()) { length = res.content_length_; } else if (req.ranges.size() == 1) { - auto offsets = - detail::get_range_offset_and_length(req, res.content_length_, 0); - auto offset = offsets.first; - length = offsets.second; + auto offset_and_length = detail::get_range_offset_and_length( + req.ranges[0], res.content_length_); + + length = offset_and_length.second; + auto content_range = detail::make_content_range_header_field( - offset, length, res.content_length_); + offset_and_length, res.content_length_); res.set_header("Content-Range", content_range); } else { - length = detail::get_multipart_ranges_data_length(req, res, boundary, - content_type); + length = detail::get_multipart_ranges_data_length( + req, boundary, content_type, res.content_length_); } res.set_header("Content-Length", std::to_string(length)); } else { @@ -5965,28 +6553,22 @@ inline void Server::apply_ranges(const Request &req, Response &res, if (req.ranges.empty()) { ; } else if (req.ranges.size() == 1) { - auto offsets = - detail::get_range_offset_and_length(req, res.body.size(), 0); - auto offset = offsets.first; - auto length = offsets.second; + auto offset_and_length = + detail::get_range_offset_and_length(req.ranges[0], res.body.size()); + auto offset = offset_and_length.first; + auto length = offset_and_length.second; + auto content_range = detail::make_content_range_header_field( - offset, length, res.body.size()); + offset_and_length, res.body.size()); res.set_header("Content-Range", content_range); - if (offset < res.body.size()) { - res.body = res.body.substr(offset, length); - } else { - res.body.clear(); - res.status = 416; - } + + assert(offset + length <= res.body.size()); + res.body = res.body.substr(offset, length); } else { std::string data; - if (detail::make_multipart_ranges_data(req, res, boundary, content_type, - data)) { - res.body.swap(data); - } else { - res.body.clear(); - res.status = 416; - } + detail::make_multipart_ranges_data(req, res, boundary, content_type, + res.body.size(), data); + res.body.swap(data); } if (type != detail::EncodingType::None) { @@ -6025,12 +6607,12 @@ inline void Server::apply_ranges(const Request &req, Response &res, inline bool Server::dispatch_request_for_content_reader( Request &req, Response &res, ContentReader content_reader, - const HandlersForContentReader &handlers) { + const HandlersForContentReader &handlers) const { for (const auto &x : handlers) { - const auto &pattern = x.first; + const auto &matcher = x.first; const auto &handler = x.second; - if (std::regex_match(req.path, req.matches, pattern)) { + if (matcher->match(req)) { handler(req, res, content_reader); return true; } @@ -6050,15 +6632,10 @@ Server::process_request(Stream &strm, bool close_connection, if (!line_reader.getline()) { return false; } Request req; + Response res; - res.version = "HTTP/1.1"; - - for (const auto &header : default_headers_) { - if (res.headers.find(header.first) == res.headers.end()) { - res.headers.insert(header); - } - } + res.headers = default_headers_; #ifdef _WIN32 // TODO: Increase FD_SETSIZE statically (libzmq), dynamically (MySQL). @@ -6068,7 +6645,7 @@ Server::process_request(Stream &strm, bool close_connection, if (strm.socket() >= FD_SETSIZE) { Headers dummy; detail::read_headers(strm, dummy); - res.status = 500; + res.status = StatusCode::InternalServerError_500; return write_response(strm, close_connection, req, res); } #endif @@ -6078,14 +6655,14 @@ Server::process_request(Stream &strm, bool close_connection, if (line_reader.size() > CPPHTTPLIB_REQUEST_URI_MAX_LENGTH) { Headers dummy; detail::read_headers(strm, dummy); - res.status = 414; + res.status = StatusCode::UriTooLong_414; return write_response(strm, close_connection, req, res); } // Request line and headers if (!parse_request_line(line_reader.ptr(), req) || !detail::read_headers(strm, req.headers)) { - res.status = 400; + res.status = StatusCode::BadRequest_400; return write_response(strm, close_connection, req, res); } @@ -6109,7 +6686,7 @@ Server::process_request(Stream &strm, bool close_connection, if (req.has_header("Range")) { const auto &range_header_value = req.get_header_value("Range"); if (!detail::parse_range_header(range_header_value, req.ranges)) { - res.status = 416; + res.status = StatusCode::RangeNotSatisfiable_416; return write_response(strm, close_connection, req, res); } } @@ -6117,22 +6694,22 @@ Server::process_request(Stream &strm, bool close_connection, if (setup_request) { setup_request(req); } if (req.get_header_value("Expect") == "100-continue") { - auto status = 100; + int status = StatusCode::Continue_100; if (expect_100_continue_handler_) { status = expect_100_continue_handler_(req, res); } switch (status) { - case 100: - case 417: + case StatusCode::Continue_100: + case StatusCode::ExpectationFailed_417: strm.write_format("HTTP/1.1 %d %s\r\n\r\n", status, - detail::status_message(status)); + status_message(status)); break; default: return write_response(strm, close_connection, req, res); } } - // Rounting - bool routed = false; + // Routing + auto routed = false; #ifdef CPPHTTPLIB_NO_EXCEPTIONS routed = routing(req, res, strm); #else @@ -6144,7 +6721,7 @@ Server::process_request(Stream &strm, bool close_connection, exception_handler_(req, res, ep); routed = true; } else { - res.status = 500; + res.status = StatusCode::InternalServerError_500; std::string val; auto s = e.what(); for (size_t i = 0; s[i]; i++) { @@ -6162,17 +6739,29 @@ Server::process_request(Stream &strm, bool close_connection, exception_handler_(req, res, ep); routed = true; } else { - res.status = 500; + res.status = StatusCode::InternalServerError_500; res.set_header("EXCEPTION_WHAT", "UNKNOWN"); } } #endif - if (routed) { - if (res.status == -1) { res.status = req.ranges.empty() ? 200 : 206; } + if (res.status == -1) { + res.status = req.ranges.empty() ? StatusCode::OK_200 + : StatusCode::PartialContent_206; + } + + if (detail::range_error(req, res)) { + res.body.clear(); + res.content_length_ = 0; + res.content_provider_ = nullptr; + res.status = StatusCode::RangeNotSatisfiable_416; + return write_response(strm, close_connection, req, res); + } + return write_response_with_content(strm, close_connection, req, res); } else { - if (res.status == -1) { res.status = 404; } + if (res.status == -1) { res.status = StatusCode::NotFound_404; } + return write_response(strm, close_connection, req, res); } } @@ -6272,7 +6861,7 @@ inline socket_t ClientImpl::create_client_socket(Error &error) const { // Check is custom IP specified for host_ std::string ip; auto it = addr_map_.find(host_); - if (it != addr_map_.end()) ip = it->second; + if (it != addr_map_.end()) { ip = it->second; } return detail::create_client_socket( host_, ip, port_, address_family_, tcp_nodelay_, socket_options_, @@ -6297,7 +6886,7 @@ inline void ClientImpl::shutdown_ssl(Socket & /*socket*/, socket_requests_are_from_thread_ == std::this_thread::get_id()); } -inline void ClientImpl::shutdown_socket(Socket &socket) { +inline void ClientImpl::shutdown_socket(Socket &socket) const { if (socket.sock == INVALID_SOCKET) { return; } detail::shutdown_socket(socket.sock); } @@ -6322,7 +6911,7 @@ inline void ClientImpl::close_socket(Socket &socket) { } inline bool ClientImpl::read_response_line(Stream &strm, const Request &req, - Response &res) { + Response &res) const { std::array buf{}; detail::stream_line_reader line_reader(strm, buf.data(), buf.size()); @@ -6330,9 +6919,9 @@ inline bool ClientImpl::read_response_line(Stream &strm, const Request &req, if (!line_reader.getline()) { return false; } #ifdef CPPHTTPLIB_ALLOW_LF_AS_LINE_TERMINATOR - const static std::regex re("(HTTP/1\\.[01]) (\\d{3})(?: (.*?))?\r\n"); -#else const static std::regex re("(HTTP/1\\.[01]) (\\d{3})(?: (.*?))?\r?\n"); +#else + const static std::regex re("(HTTP/1\\.[01]) (\\d{3})(?: (.*?))?\r\n"); #endif std::cmatch m; @@ -6344,7 +6933,7 @@ inline bool ClientImpl::read_response_line(Stream &strm, const Request &req, res.reason = std::string(m[3]); // Ignore '100 Continue' - while (res.status == 100) { + while (res.status == StatusCode::Continue_100) { if (!line_reader.getline()) { return false; } // CRLF if (!line_reader.getline()) { return false; } // next response line @@ -6492,15 +7081,31 @@ inline bool ClientImpl::handle_request(Stream &strm, Request &req, if (!ret) { return false; } + if (res.get_header_value("Connection") == "close" || + (res.version == "HTTP/1.0" && res.reason != "Connection established")) { + // TODO this requires a not-entirely-obvious chain of calls to be correct + // for this to be safe. + + // This is safe to call because handle_request is only called by send_ + // which locks the request mutex during the process. It would be a bug + // to call it from a different thread since it's a thread-safety issue + // to do these things to the socket if another thread is using the socket. + std::lock_guard guard(socket_mutex_); + shutdown_ssl(socket_, true); + shutdown_socket(socket_); + close_socket(socket_); + } + if (300 < res.status && res.status < 400 && follow_location_) { req = req_save; ret = redirect(req, res, error); } #ifdef CPPHTTPLIB_OPENSSL_SUPPORT - if ((res.status == 401 || res.status == 407) && + if ((res.status == StatusCode::Unauthorized_401 || + res.status == StatusCode::ProxyAuthenticationRequired_407) && req.authorization_count_ < 5) { - auto is_proxy = res.status == 407; + auto is_proxy = res.status == StatusCode::ProxyAuthenticationRequired_407; const auto &username = is_proxy ? proxy_digest_auth_username_ : digest_auth_username_; const auto &password = @@ -6571,7 +7176,7 @@ inline bool ClientImpl::redirect(Request &req, Response &res, Error &error) { } else { if (next_scheme == "https") { #ifdef CPPHTTPLIB_OPENSSL_SUPPORT - SSLClient cli(next_host.c_str(), next_port); + SSLClient cli(next_host, next_port); cli.copy_settings(*this); if (ca_cert_store_) { cli.set_ca_cert_store(ca_cert_store_); } return detail::redirect(cli, req, res, path, location, error); @@ -6579,7 +7184,7 @@ inline bool ClientImpl::redirect(Request &req, Response &res, Error &error) { return false; #endif } else { - ClientImpl cli(next_host.c_str(), next_port); + ClientImpl cli(next_host, next_port); cli.copy_settings(*this); return detail::redirect(cli, req, res, path, location, error); } @@ -6588,7 +7193,7 @@ inline bool ClientImpl::redirect(Request &req, Response &res, Error &error) { inline bool ClientImpl::write_content_with_provider(Stream &strm, const Request &req, - Error &error) { + Error &error) const { auto is_shutting_down = []() { return false; }; if (req.is_chunked_content_provider_) { @@ -6616,32 +7221,32 @@ inline bool ClientImpl::write_request(Stream &strm, Request &req, // Prepare additional headers if (close_connection) { if (!req.has_header("Connection")) { - req.headers.emplace("Connection", "close"); + req.set_header("Connection", "close"); } } if (!req.has_header("Host")) { if (is_ssl()) { if (port_ == 443) { - req.headers.emplace("Host", host_); + req.set_header("Host", host_); } else { - req.headers.emplace("Host", host_and_port_); + req.set_header("Host", host_and_port_); } } else { if (port_ == 80) { - req.headers.emplace("Host", host_); + req.set_header("Host", host_); } else { - req.headers.emplace("Host", host_and_port_); + req.set_header("Host", host_and_port_); } } } - if (!req.has_header("Accept")) { req.headers.emplace("Accept", "*/*"); } + if (!req.has_header("Accept")) { req.set_header("Accept", "*/*"); } #ifndef CPPHTTPLIB_NO_DEFAULT_USER_AGENT if (!req.has_header("User-Agent")) { auto agent = std::string("cpp-httplib/") + CPPHTTPLIB_VERSION; - req.headers.emplace("User-Agent", agent); + req.set_header("User-Agent", agent); } #endif @@ -6650,23 +7255,23 @@ inline bool ClientImpl::write_request(Stream &strm, Request &req, if (!req.is_chunked_content_provider_) { if (!req.has_header("Content-Length")) { auto length = std::to_string(req.content_length_); - req.headers.emplace("Content-Length", length); + req.set_header("Content-Length", length); } } } else { if (req.method == "POST" || req.method == "PUT" || req.method == "PATCH") { - req.headers.emplace("Content-Length", "0"); + req.set_header("Content-Length", "0"); } } } else { if (!req.has_header("Content-Type")) { - req.headers.emplace("Content-Type", "text/plain"); + req.set_header("Content-Type", "text/plain"); } if (!req.has_header("Content-Length")) { auto length = std::to_string(req.body.size()); - req.headers.emplace("Content-Length", length); + req.set_header("Content-Length", length); } } @@ -6706,7 +7311,7 @@ inline bool ClientImpl::write_request(Stream &strm, Request &req, const auto &path = url_encode_ ? detail::encode_url(req.path) : req.path; bstrm.write_format("%s %s HTTP/1.1\r\n", req.method.c_str(), path.c_str()); - detail::write_headers(bstrm, req.headers); + header_writer_(bstrm, req.headers); // Flush buffer auto &data = bstrm.get_buffer(); @@ -6734,12 +7339,10 @@ inline std::unique_ptr ClientImpl::send_with_content_provider( ContentProvider content_provider, ContentProviderWithoutLength content_provider_without_length, const std::string &content_type, Error &error) { - if (!content_type.empty()) { - req.headers.emplace("Content-Type", content_type); - } + if (!content_type.empty()) { req.set_header("Content-Type", content_type); } #ifdef CPPHTTPLIB_ZLIB_SUPPORT - if (compress_) { req.headers.emplace("Content-Encoding", "gzip"); } + if (compress_) { req.set_header("Content-Encoding", "gzip"); } #endif #ifdef CPPHTTPLIB_ZLIB_SUPPORT @@ -6800,10 +7403,9 @@ inline std::unique_ptr ClientImpl::send_with_content_provider( req.content_provider_ = detail::ContentProviderAdapter( std::move(content_provider_without_length)); req.is_chunked_content_provider_ = true; - req.headers.emplace("Transfer-Encoding", "chunked"); + req.set_header("Transfer-Encoding", "chunked"); } else { req.body.assign(body, content_length); - ; } } @@ -6864,7 +7466,8 @@ inline bool ClientImpl::process_request(Stream &strm, Request &req, } // Body - if ((res.status != 204) && req.method != "HEAD" && req.method != "CONNECT") { + if ((res.status != StatusCode::NoContent_204) && req.method != "HEAD" && + req.method != "CONNECT") { auto redirect = 300 < res.status && res.status < 400 && follow_location_; if (req.response_handler && !redirect) { @@ -6909,24 +7512,6 @@ inline bool ClientImpl::process_request(Stream &strm, Request &req, } } - if (res.get_header_value("Connection") == "close" || - (res.version == "HTTP/1.0" && res.reason != "Connection established")) { - // TODO this requires a not-entirely-obvious chain of calls to be correct - // for this to be safe. Maybe a code refactor (such as moving this out to - // the send function and getting rid of the recursiveness of the mutex) - // could make this more obvious. - - // This is safe to call because process_request is only called by - // handle_request which is only called by send, which locks the request - // mutex during the process. It would be a bug to call it from a different - // thread since it's a thread-safety issue to do these things to the socket - // if another thread is using the socket. - std::lock_guard guard(socket_mutex_); - shutdown_ssl(socket_, true); - shutdown_socket(socket_); - close_socket(socket_); - } - // Log if (logger_) { logger_(req, res); } @@ -6935,13 +7520,14 @@ inline bool ClientImpl::process_request(Stream &strm, Request &req, inline ContentProviderWithoutLength ClientImpl::get_multipart_content_provider( const std::string &boundary, const MultipartFormDataItems &items, - const MultipartFormDataProviderItems &provider_items) { - size_t cur_item = 0, cur_start = 0; + const MultipartFormDataProviderItems &provider_items) const { + size_t cur_item = 0; + size_t cur_start = 0; // cur_item and cur_start are copied to within the std::function and maintain // state between successive calls return [&, cur_item, cur_start](size_t offset, DataSink &sink) mutable -> bool { - if (!offset && items.size()) { + if (!offset && !items.empty()) { sink.os << detail::serialize_multipart_formdata(items, boundary, false); return true; } else if (cur_item < provider_items.size()) { @@ -6954,12 +7540,13 @@ inline ContentProviderWithoutLength ClientImpl::get_multipart_content_provider( } DataSink cur_sink; - bool has_data = true; + auto has_data = true; cur_sink.write = sink.write; cur_sink.done = [&]() { has_data = false; }; - if (!provider_items[cur_item].provider(offset - cur_start, cur_sink)) + if (!provider_items[cur_item].provider(offset - cur_start, cur_sink)) { return false; + } if (!has_data) { sink.os << detail::serialize_multipart_formdata_item_end(); @@ -7078,14 +7665,15 @@ inline Result ClientImpl::Get(const std::string &path, const Params ¶ms, if (params.empty()) { return Get(path, headers); } std::string path_with_query = append_query_params(path, params); - return Get(path_with_query.c_str(), headers, progress); + return Get(path_with_query, headers, std::move(progress)); } inline Result ClientImpl::Get(const std::string &path, const Params ¶ms, const Headers &headers, ContentReceiver content_receiver, Progress progress) { - return Get(path, params, headers, nullptr, content_receiver, progress); + return Get(path, params, headers, nullptr, std::move(content_receiver), + std::move(progress)); } inline Result ClientImpl::Get(const std::string &path, const Params ¶ms, @@ -7094,12 +7682,13 @@ inline Result ClientImpl::Get(const std::string &path, const Params ¶ms, ContentReceiver content_receiver, Progress progress) { if (params.empty()) { - return Get(path, headers, response_handler, content_receiver, progress); + return Get(path, headers, std::move(response_handler), + std::move(content_receiver), std::move(progress)); } std::string path_with_query = append_query_params(path, params); - return Get(path_with_query.c_str(), headers, response_handler, - content_receiver, progress); + return Get(path_with_query, headers, std::move(response_handler), + std::move(content_receiver), std::move(progress)); } inline Result ClientImpl::Head(const std::string &path) { @@ -7201,7 +7790,7 @@ inline Result ClientImpl::Post(const std::string &path, const Headers &headers, const auto &content_type = detail::serialize_multipart_formdata_get_content_type(boundary); const auto &body = detail::serialize_multipart_formdata(items, boundary); - return Post(path, headers, body, content_type.c_str()); + return Post(path, headers, body, content_type); } inline Result ClientImpl::Post(const std::string &path, const Headers &headers, @@ -7214,7 +7803,7 @@ inline Result ClientImpl::Post(const std::string &path, const Headers &headers, const auto &content_type = detail::serialize_multipart_formdata_get_content_type(boundary); const auto &body = detail::serialize_multipart_formdata(items, boundary); - return Post(path, headers, body, content_type.c_str()); + return Post(path, headers, body, content_type); } inline Result @@ -7423,9 +8012,7 @@ inline Result ClientImpl::Delete(const std::string &path, req.headers = headers; req.path = path; - if (!content_type.empty()) { - req.headers.emplace("Content-Type", content_type); - } + if (!content_type.empty()) { req.set_header("Content-Type", content_type); } req.body.assign(body, content_length); return send_(std::move(req)); @@ -7458,13 +8045,6 @@ inline Result ClientImpl::Options(const std::string &path, return send_(std::move(req)); } -inline size_t ClientImpl::is_socket_open() const { - std::lock_guard guard(socket_mutex_); - return socket_.is_open(); -} - -inline socket_t ClientImpl::socket() const { return socket_.sock; } - inline void ClientImpl::stop() { std::lock_guard guard(socket_mutex_); @@ -7488,6 +8068,17 @@ inline void ClientImpl::stop() { close_socket(socket_); } +inline std::string ClientImpl::host() const { return host_; } + +inline int ClientImpl::port() const { return port_; } + +inline size_t ClientImpl::is_socket_open() const { + std::lock_guard guard(socket_mutex_); + return socket_.is_open(); +} + +inline socket_t ClientImpl::socket() const { return socket_.sock; } + inline void ClientImpl::set_connection_timeout(time_t sec, time_t usec) { connection_timeout_sec_ = sec; connection_timeout_usec_ = usec; @@ -7536,6 +8127,11 @@ inline void ClientImpl::set_default_headers(Headers headers) { default_headers_ = std::move(headers); } +inline void ClientImpl::set_header_writer( + std::function const &writer) { + header_writer_ = writer; +} + inline void ClientImpl::set_address_family(int family) { address_family_ = family; } @@ -7575,9 +8171,7 @@ inline void ClientImpl::set_proxy_digest_auth(const std::string &username, proxy_digest_auth_username_ = username; proxy_digest_auth_password_ = password; } -#endif -#ifdef CPPHTTPLIB_OPENSSL_SUPPORT inline void ClientImpl::set_ca_cert_path(const std::string &ca_cert_file_path, const std::string &ca_cert_dir_path) { ca_cert_file_path_ = ca_cert_file_path; @@ -7589,9 +8183,34 @@ inline void ClientImpl::set_ca_cert_store(X509_STORE *ca_cert_store) { ca_cert_store_ = ca_cert_store; } } -#endif -#ifdef CPPHTTPLIB_OPENSSL_SUPPORT +inline X509_STORE *ClientImpl::create_ca_cert_store(const char *ca_cert, + std::size_t size) const { + auto mem = BIO_new_mem_buf(ca_cert, static_cast(size)); + if (!mem) { return nullptr; } + + auto inf = PEM_X509_INFO_read_bio(mem, nullptr, nullptr, nullptr); + if (!inf) { + BIO_free_all(mem); + return nullptr; + } + + auto cts = X509_STORE_new(); + if (cts) { + for (auto i = 0; i < static_cast(sk_X509_INFO_num(inf)); i++) { + auto itmp = sk_X509_INFO_value(inf, i); + if (!itmp) { continue; } + + if (itmp->x509) { X509_STORE_add_cert(cts, itmp->x509); } + if (itmp->crl) { X509_STORE_add_crl(cts, itmp->crl); } + } + } + + sk_X509_INFO_pop_free(inf, X509_INFO_free); + BIO_free_all(mem); + return cts; +} + inline void ClientImpl::enable_server_certificate_verification(bool enabled) { server_certificate_verification_ = enabled; } @@ -7655,7 +8274,7 @@ bool ssl_connect_or_accept_nonblocking(socket_t sock, SSL *ssl, U ssl_connect_or_accept, time_t timeout_sec, time_t timeout_usec) { - int res = 0; + auto res = 0; while ((res = ssl_connect_or_accept(ssl)) != 1) { auto err = SSL_get_error(ssl, res); switch (err) { @@ -7718,7 +8337,7 @@ inline SSLSocketStream::SSLSocketStream(socket_t sock, SSL *ssl, SSL_clear_mode(ssl, SSL_MODE_AUTO_RETRY); } -inline SSLSocketStream::~SSLSocketStream() {} +inline SSLSocketStream::~SSLSocketStream() = default; inline bool SSLSocketStream::is_readable() const { return detail::select_read(sock_, read_timeout_sec_, read_timeout_usec_) > 0; @@ -7736,7 +8355,7 @@ inline ssize_t SSLSocketStream::read(char *ptr, size_t size) { auto ret = SSL_read(ssl_, ptr, static_cast(size)); if (ret < 0) { auto err = SSL_get_error(ssl_, ret); - int n = 1000; + auto n = 1000; #ifdef _WIN32 while (--n >= 0 && (err == SSL_ERROR_WANT_READ || (err == SSL_ERROR_SYSCALL && @@ -7769,7 +8388,7 @@ inline ssize_t SSLSocketStream::write(const char *ptr, size_t size) { auto ret = SSL_write(ssl_, ptr, static_cast(handle_size)); if (ret < 0) { auto err = SSL_get_error(ssl_, ret); - int n = 1000; + auto n = 1000; #ifdef _WIN32 while (--n >= 0 && (err == SSL_ERROR_WANT_WRITE || (err == SSL_ERROR_SYSCALL && @@ -7822,10 +8441,10 @@ inline SSLServer::SSLServer(const char *cert_path, const char *private_key_path, SSL_CTX_set_min_proto_version(ctx_, TLS1_1_VERSION); - // add default password callback before opening encrypted private key if (private_key_password != nullptr && (private_key_password[0] != '\0')) { - SSL_CTX_set_default_passwd_cb_userdata(ctx_, - (char *)private_key_password); + SSL_CTX_set_default_passwd_cb_userdata( + ctx_, + reinterpret_cast(const_cast(private_key_password))); } if (SSL_CTX_use_certificate_chain_file(ctx_, cert_path) != 1 || @@ -7927,16 +8546,23 @@ inline SSLClient::SSLClient(const std::string &host, int port) inline SSLClient::SSLClient(const std::string &host, int port, const std::string &client_cert_path, - const std::string &client_key_path) + const std::string &client_key_path, + const std::string &private_key_password) : ClientImpl(host, port, client_cert_path, client_key_path) { ctx_ = SSL_CTX_new(TLS_client_method()); detail::split(&host_[0], &host_[host_.size()], '.', [&](const char *b, const char *e) { - host_components_.emplace_back(std::string(b, e)); + host_components_.emplace_back(b, e); }); if (!client_cert_path.empty() && !client_key_path.empty()) { + if (!private_key_password.empty()) { + SSL_CTX_set_default_passwd_cb_userdata( + ctx_, reinterpret_cast( + const_cast(private_key_password.c_str()))); + } + if (SSL_CTX_use_certificate_file(ctx_, client_cert_path.c_str(), SSL_FILETYPE_PEM) != 1 || SSL_CTX_use_PrivateKey_file(ctx_, client_key_path.c_str(), @@ -7948,16 +8574,23 @@ inline SSLClient::SSLClient(const std::string &host, int port, } inline SSLClient::SSLClient(const std::string &host, int port, - X509 *client_cert, EVP_PKEY *client_key) + X509 *client_cert, EVP_PKEY *client_key, + const std::string &private_key_password) : ClientImpl(host, port) { ctx_ = SSL_CTX_new(TLS_client_method()); detail::split(&host_[0], &host_[host_.size()], '.', [&](const char *b, const char *e) { - host_components_.emplace_back(std::string(b, e)); + host_components_.emplace_back(b, e); }); if (client_cert != nullptr && client_key != nullptr) { + if (!private_key_password.empty()) { + SSL_CTX_set_default_passwd_cb_userdata( + ctx_, reinterpret_cast( + const_cast(private_key_password.c_str()))); + } + if (SSL_CTX_use_certificate(ctx_, client_cert) != 1 || SSL_CTX_use_PrivateKey(ctx_, client_key) != 1) { SSL_CTX_free(ctx_); @@ -7989,6 +8622,11 @@ inline void SSLClient::set_ca_cert_store(X509_STORE *ca_cert_store) { } } +inline void SSLClient::load_ca_cert_store(const char *ca_cert, + std::size_t size) { + set_ca_cert_store(ClientImpl::create_ca_cert_store(ca_cert, size)); +} + inline long SSLClient::get_openssl_verify_result() const { return verify_result_; } @@ -8003,14 +8641,14 @@ inline bool SSLClient::create_and_connect_socket(Socket &socket, Error &error) { inline bool SSLClient::connect_with_proxy(Socket &socket, Response &res, bool &success, Error &error) { success = true; - Response res2; + Response proxy_res; if (!detail::process_client_socket( socket.sock, read_timeout_sec_, read_timeout_usec_, write_timeout_sec_, write_timeout_usec_, [&](Stream &strm) { Request req2; req2.method = "CONNECT"; req2.path = host_and_port_; - return process_request(strm, req2, res2, false, error); + return process_request(strm, req2, proxy_res, false, error); })) { // Thread-safe to close everything because we are assuming there are no // requests in flight @@ -8021,12 +8659,12 @@ inline bool SSLClient::connect_with_proxy(Socket &socket, Response &res, return false; } - if (res2.status == 407) { + if (proxy_res.status == StatusCode::ProxyAuthenticationRequired_407) { if (!proxy_digest_auth_username_.empty() && !proxy_digest_auth_password_.empty()) { std::map auth; - if (detail::parse_www_authenticate(res2, auth, true)) { - Response res3; + if (detail::parse_www_authenticate(proxy_res, auth, true)) { + proxy_res = Response(); if (!detail::process_client_socket( socket.sock, read_timeout_sec_, read_timeout_usec_, write_timeout_sec_, write_timeout_usec_, [&](Stream &strm) { @@ -8037,7 +8675,7 @@ inline bool SSLClient::connect_with_proxy(Socket &socket, Response &res, req3, auth, 1, detail::random_string(10), proxy_digest_auth_username_, proxy_digest_auth_password_, true)); - return process_request(strm, req3, res3, false, error); + return process_request(strm, req3, proxy_res, false, error); })) { // Thread-safe to close everything because we are assuming there are // no requests in flight @@ -8048,17 +8686,28 @@ inline bool SSLClient::connect_with_proxy(Socket &socket, Response &res, return false; } } - } else { - res = res2; - return false; } } + // If status code is not 200, proxy request is failed. + // Set error to ProxyConnection and return proxy response + // as the response of the request + if (proxy_res.status != StatusCode::OK_200) { + error = Error::ProxyConnection; + res = std::move(proxy_res); + // Thread-safe to close everything because we are assuming there are + // no requests in flight + shutdown_ssl(socket, true); + shutdown_socket(socket); + close_socket(socket); + return false; + } + return true; } inline bool SSLClient::load_certs() { - bool ret = true; + auto ret = true; std::call_once(initialize_cert_, [&]() { std::lock_guard guard(ctx_mutex_); @@ -8134,7 +8783,11 @@ inline bool SSLClient::initialize_ssl(Socket &socket, Error &error) { return true; }, [&](SSL *ssl2) { - SSL_set_tlsext_host_name(ssl2, host_.c_str()); + // NOTE: Direct call instead of using the OpenSSL macro to suppress + // -Wold-style-cast warning + // SSL_set_tlsext_host_name(ssl2, host_.c_str()); + SSL_ctrl(ssl2, SSL_CTRL_SET_TLSEXT_HOSTNAME, TLSEXT_NAMETYPE_host_name, + static_cast(const_cast(host_.c_str()))); return true; }); @@ -8208,8 +8861,8 @@ SSLClient::verify_host_with_subject_alt_name(X509 *server_cert) const { auto type = GEN_DNS; - struct in6_addr addr6; - struct in_addr addr; + struct in6_addr addr6 {}; + struct in_addr addr {}; size_t addr_len = 0; #ifndef __MINGW32__ @@ -8234,8 +8887,9 @@ SSLClient::verify_host_with_subject_alt_name(X509 *server_cert) const { for (decltype(count) i = 0; i < count && !dsn_matched; i++) { auto val = sk_GENERAL_NAME_value(alt_names, i); if (val->type == type) { - auto name = (const char *)ASN1_STRING_get0_data(val->d.ia5); - auto name_len = (size_t)ASN1_STRING_length(val->d.ia5); + auto name = + reinterpret_cast(ASN1_STRING_get0_data(val->d.ia5)); + auto name_len = static_cast(ASN1_STRING_length(val->d.ia5)); switch (type) { case GEN_DNS: dsn_matched = check_host_name(name, name_len); break; @@ -8253,7 +8907,8 @@ SSLClient::verify_host_with_subject_alt_name(X509 *server_cert) const { if (dsn_matched || ip_matched) { ret = true; } } - GENERAL_NAMES_free((STACK_OF(GENERAL_NAME) *)alt_names); + GENERAL_NAMES_free(const_cast( + reinterpret_cast(alt_names))); return ret; } @@ -8282,7 +8937,7 @@ inline bool SSLClient::check_host_name(const char *pattern, std::vector pattern_components; detail::split(&pattern[0], &pattern[pattern_len], '.', [&](const char *b, const char *e) { - pattern_components.emplace_back(std::string(b, e)); + pattern_components.emplace_back(b, e); }); if (host_components_.size() != pattern_components.size()) { return false; } @@ -8361,7 +9016,7 @@ inline Client::Client(const std::string &host, int port, : cli_(detail::make_unique(host, port, client_cert_path, client_key_path)) {} -inline Client::~Client() {} +inline Client::~Client() = default; inline bool Client::is_valid() const { return cli_ != nullptr && cli_->is_valid(); @@ -8421,19 +9076,20 @@ inline Result Client::Get(const std::string &path, const Headers &headers, } inline Result Client::Get(const std::string &path, const Params ¶ms, const Headers &headers, Progress progress) { - return cli_->Get(path, params, headers, progress); + return cli_->Get(path, params, headers, std::move(progress)); } inline Result Client::Get(const std::string &path, const Params ¶ms, const Headers &headers, ContentReceiver content_receiver, Progress progress) { - return cli_->Get(path, params, headers, content_receiver, progress); + return cli_->Get(path, params, headers, std::move(content_receiver), + std::move(progress)); } inline Result Client::Get(const std::string &path, const Params ¶ms, const Headers &headers, ResponseHandler response_handler, ContentReceiver content_receiver, Progress progress) { - return cli_->Get(path, params, headers, response_handler, content_receiver, - progress); + return cli_->Get(path, params, headers, std::move(response_handler), + std::move(content_receiver), std::move(progress)); } inline Result Client::Head(const std::string &path) { return cli_->Head(path); } @@ -8665,12 +9321,16 @@ inline bool Client::send(Request &req, Response &res, Error &error) { inline Result Client::send(const Request &req) { return cli_->send(req); } +inline void Client::stop() { cli_->stop(); } + +inline std::string Client::host() const { return cli_->host(); } + +inline int Client::port() const { return cli_->port(); } + inline size_t Client::is_socket_open() const { return cli_->is_socket_open(); } inline socket_t Client::socket() const { return cli_->socket(); } -inline void Client::stop() { cli_->stop(); } - inline void Client::set_hostname_addr_map(std::map addr_map) { cli_->set_hostname_addr_map(std::move(addr_map)); @@ -8680,6 +9340,11 @@ inline void Client::set_default_headers(Headers headers) { cli_->set_default_headers(std::move(headers)); } +inline void Client::set_header_writer( + std::function const &writer) { + cli_->set_header_writer(writer); +} + inline void Client::set_address_family(int family) { cli_->set_address_family(family); } @@ -8754,7 +9419,9 @@ inline void Client::enable_server_certificate_verification(bool enabled) { } #endif -inline void Client::set_logger(Logger logger) { cli_->set_logger(logger); } +inline void Client::set_logger(Logger logger) { + cli_->set_logger(std::move(logger)); +} #ifdef CPPHTTPLIB_OPENSSL_SUPPORT inline void Client::set_ca_cert_path(const std::string &ca_cert_file_path, @@ -8770,6 +9437,10 @@ inline void Client::set_ca_cert_store(X509_STORE *ca_cert_store) { } } +inline void Client::load_ca_cert_store(const char *ca_cert, std::size_t size) { + set_ca_cert_store(cli_->create_ca_cert_store(ca_cert, size)); +} + inline long Client::get_openssl_verify_result() const { if (is_ssl_) { return static_cast(*cli_).get_openssl_verify_result(); diff --git a/examples/server/index.html.hpp b/examples/server/index.html.hpp index 20551520e..ae8c88980 100644 --- a/examples/server/index.html.hpp +++ b/examples/server/index.html.hpp @@ -417,2375 +417,2405 @@ unsigned char index_html[] = { 0x20, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x6e, 0x6f, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x2c, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, - 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x61, 0x63, 0x68, 0x65, 0x5f, 0x70, 0x72, - 0x6f, 0x6d, 0x70, 0x74, 0x3a, 0x20, 0x74, 0x72, 0x75, 0x65, 0x2c, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, - 0x79, 0x3a, 0x20, 0x27, 0x27, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x20, 0x53, 0x54, 0x41, - 0x52, 0x54, 0x3a, 0x20, 0x53, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, - 0x66, 0x6f, 0x72, 0x20, 0x73, 0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67, 0x20, - 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x70, 0x61, 0x72, - 0x61, 0x6d, 0x65, 0x74, 0x65, 0x72, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x62, - 0x72, 0x6f, 0x77, 0x73, 0x65, 0x72, 0x73, 0x20, 0x4c, 0x6f, 0x63, 0x61, - 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x20, 0x2a, 0x2f, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, - 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, - 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, - 0x3d, 0x20, 0x22, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x63, 0x70, 0x70, 0x5f, - 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x5f, 0x6c, 0x6f, 0x63, 0x61, 0x6c, - 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x22, 0x3b, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, - 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, - 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, - 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x74, 0x61, 0x67, - 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, - 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x73, 0x65, 0x74, 0x49, - 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, - 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, - 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, - 0x20, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x73, - 0x74, 0x72, 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, 0x63, 0x6f, 0x6e, - 0x74, 0x65, 0x6e, 0x74, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x6e, 0x5f, 0x6b, 0x65, + 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x6d, 0x69, + 0x6e, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x20, 0x66, 0x72, 0x6f, 0x6d, + 0x20, 0x65, 0x61, 0x63, 0x68, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x65, + 0x72, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x61, + 0x67, 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x61, 0x63, 0x68, 0x65, + 0x5f, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x3a, 0x20, 0x74, 0x72, 0x75, + 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x70, 0x69, + 0x5f, 0x6b, 0x65, 0x79, 0x3a, 0x20, 0x27, 0x27, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x20, + 0x53, 0x54, 0x41, 0x52, 0x54, 0x3a, 0x20, 0x53, 0x75, 0x70, 0x70, 0x6f, + 0x72, 0x74, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x73, 0x74, 0x6f, 0x72, 0x69, + 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x65, 0x74, 0x65, 0x72, 0x73, 0x20, 0x69, + 0x6e, 0x20, 0x62, 0x72, 0x6f, 0x77, 0x73, 0x65, 0x72, 0x73, 0x20, 0x4c, + 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x20, + 0x2a, 0x2f, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, + 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, + 0x65, 0x79, 0x20, 0x3d, 0x20, 0x22, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x63, + 0x70, 0x70, 0x5f, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x5f, 0x6c, 0x6f, + 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x22, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, - 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x52, 0x61, 0x77, 0x54, 0x65, 0x78, 0x74, - 0x28, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, - 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, - 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, - 0x73, 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, - 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, - 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, - 0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, - 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, - 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, - 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, - 0x61, 0x41, 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x74, 0x61, - 0x67, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x20, 0x3d, 0x20, - 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, - 0x2e, 0x67, 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, - 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, - 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, - 0x27, 0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x69, - 0x74, 0x65, 0x6d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, - 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, - 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x4a, 0x53, - 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x69, 0x74, 0x65, - 0x6d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, - 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, - 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, - 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x52, 0x61, 0x77, 0x54, 0x65, - 0x78, 0x74, 0x28, 0x74, 0x61, 0x67, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x74, - 0x65, 0x6d, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, - 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x67, 0x65, 0x74, 0x49, 0x74, 0x65, - 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, - 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, - 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, - 0x61, 0x67, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, - 0x66, 0x20, 0x28, 0x21, 0x69, 0x74, 0x65, 0x6d, 0x29, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, - 0x72, 0x6e, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, - 0x72, 0x6e, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, 0x72, 0x65, 0x61, 0x74, - 0x65, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, - 0x72, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x75, 0x73, 0x65, 0x72, 0x20, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, - 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x61, 0x76, + 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, + 0x74, 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, + 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, + 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x73, + 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, + 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, + 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, + 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x4a, 0x53, 0x4f, + 0x4e, 0x2e, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, + 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, + 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, + 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, + 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x52, 0x61, 0x77, 0x54, + 0x65, 0x78, 0x74, 0x28, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, 0x6e, + 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, + 0x67, 0x65, 0x2e, 0x73, 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, + 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, + 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, + 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x2c, + 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, + 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, + 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, + 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, + 0x28, 0x74, 0x61, 0x67, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x74, 0x65, 0x6d, + 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, + 0x61, 0x67, 0x65, 0x2e, 0x67, 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, + 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, + 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, + 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, + 0x28, 0x21, 0x69, 0x74, 0x65, 0x6d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, + 0x69, 0x74, 0x65, 0x6d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, + 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, + 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x52, 0x61, + 0x77, 0x54, 0x65, 0x78, 0x74, 0x28, 0x74, 0x61, 0x67, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x69, 0x74, 0x65, 0x6d, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, + 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x67, 0x65, 0x74, + 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, + 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, + 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, + 0x2b, 0x20, 0x74, 0x61, 0x67, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x69, 0x74, 0x65, 0x6d, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, 0x72, + 0x65, 0x61, 0x74, 0x65, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, + 0x69, 0x6e, 0x65, 0x72, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x75, 0x73, 0x65, + 0x72, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, + 0x61, 0x6e, 0x64, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, + 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, - 0x28, 0x7b, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, - 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, - 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x20, 0x6e, - 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x27, 0x27, 0x2c, 0x20, 0x74, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, 0x7b, 0x20, 0x73, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x70, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x3a, 0x20, 0x7b, 0x7d, 0x20, 0x7d, 0x20, 0x7d, - 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x6c, 0x65, - 0x74, 0x27, 0x73, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x6c, - 0x6f, 0x63, 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, - 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, - 0x6e, 0x64, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, - 0x69, 0x66, 0x20, 0x74, 0x68, 0x65, 0x72, 0x65, 0x20, 0x61, 0x72, 0x65, - 0x20, 0x61, 0x6e, 0x79, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, - 0x75, 0x73, 0x65, 0x72, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, - 0x6e, 0x67, 0x73, 0x20, 0x61, 0x72, 0x65, 0x20, 0x73, 0x74, 0x6f, 0x72, - 0x65, 0x64, 0x20, 0x69, 0x6e, 0x20, 0x6f, 0x6e, 0x65, 0x20, 0x6f, 0x62, - 0x6a, 0x65, 0x63, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, - 0x69, 0x6e, 0x20, 0x66, 0x6f, 0x72, 0x6d, 0x20, 0x6f, 0x66, 0x20, 0x7b, - 0x20, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x6e, 0x61, - 0x6d, 0x65, 0x22, 0x3a, 0x20, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x64, 0x61, 0x74, 0x61, 0x22, 0x20, 0x7d, 0x20, 0x61, 0x6e, - 0x64, 0x20, 0x7b, 0x20, 0x22, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, - 0x73, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x6e, 0x61, 0x6d, - 0x65, 0x22, 0x3a, 0x22, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, - 0x64, 0x61, 0x74, 0x61, 0x22, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, - 0x28, 0x27, 0x49, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x69, 0x6e, 0x67, 0x20, - 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x73, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, - 0x65, 0x74, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x6c, - 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, - 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x4f, 0x62, - 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, 0x29, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x69, 0x6d, 0x70, 0x6f, + 0x74, 0x65, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, + 0x7b, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x27, 0x27, 0x2c, 0x20, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, 0x7b, 0x20, + 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x7b, 0x7d, 0x2c, + 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x3a, 0x20, 0x7b, 0x7d, 0x20, + 0x7d, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, + 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, + 0x74, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x73, 0x61, + 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, + 0x67, 0x73, 0x20, 0x69, 0x66, 0x20, 0x74, 0x68, 0x65, 0x72, 0x65, 0x20, + 0x61, 0x72, 0x65, 0x20, 0x61, 0x6e, 0x79, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x2f, 0x2f, 0x20, 0x75, 0x73, 0x65, 0x72, 0x20, 0x74, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x65, + 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x61, 0x72, 0x65, 0x20, 0x73, + 0x74, 0x6f, 0x72, 0x65, 0x64, 0x20, 0x69, 0x6e, 0x20, 0x6f, 0x6e, 0x65, + 0x20, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x2f, 0x2f, 0x20, 0x69, 0x6e, 0x20, 0x66, 0x6f, 0x72, 0x6d, 0x20, 0x6f, + 0x66, 0x20, 0x7b, 0x20, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x6e, 0x61, 0x6d, 0x65, 0x22, 0x3a, 0x20, 0x22, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x64, 0x61, 0x74, 0x61, 0x22, 0x20, 0x7d, + 0x20, 0x61, 0x6e, 0x64, 0x20, 0x7b, 0x20, 0x22, 0x73, 0x65, 0x74, 0x74, + 0x69, 0x6e, 0x67, 0x73, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x6e, 0x61, 0x6d, 0x65, 0x22, 0x3a, 0x22, 0x73, 0x65, 0x74, 0x74, 0x69, + 0x6e, 0x67, 0x73, 0x64, 0x61, 0x74, 0x61, 0x22, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, + 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x49, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x69, + 0x6e, 0x67, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, + 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, + 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, + 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, + 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, + 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, + 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x69, + 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x77, 0x65, 0x72, + 0x65, 0x20, 0x73, 0x75, 0x63, 0x63, 0x65, 0x73, 0x73, 0x66, 0x75, 0x6c, + 0x6c, 0x79, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x2e, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x50, 0x72, 0x6f, + 0x63, 0x65, 0x73, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x73, 0x61, 0x76, 0x65, + 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, + 0x61, 0x6e, 0x64, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x69, 0x6e, 0x67, + 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, + 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, + 0x5b, 0x5d, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, + 0x6f, 0x67, 0x28, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, + 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, - 0x2f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x77, 0x65, 0x72, 0x65, 0x20, 0x73, - 0x75, 0x63, 0x63, 0x65, 0x73, 0x73, 0x66, 0x75, 0x6c, 0x6c, 0x79, 0x20, - 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x2e, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, - 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x50, 0x72, 0x6f, 0x63, 0x65, 0x73, - 0x73, 0x69, 0x6e, 0x67, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, - 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, - 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x69, 0x6d, 0x61, - 0x67, 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, 0x5b, 0x5d, 0x20, - 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, - 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, - 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x73, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, + 0x73, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, + 0x6f, 0x76, 0x65, 0x72, 0x72, 0x69, 0x64, 0x65, 0x20, 0x64, 0x65, 0x66, + 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x61, 0x76, 0x65, + 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x64, 0x65, 0x66, + 0x61, 0x75, 0x6c, 0x74, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x73, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, + 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, + 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, + 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, + 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x73, 0x27, 0x2c, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, + 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, + 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x2f, 0x2f, 0x20, 0x6e, 0x6f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, + 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x64, + 0x65, 0x74, 0x65, 0x63, 0x74, 0x65, 0x64, 0x2e, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, + 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x49, 0x6e, 0x69, 0x74, 0x69, 0x61, 0x6c, + 0x69, 0x7a, 0x69, 0x6e, 0x67, 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, + 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, + 0x61, 0x76, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, + 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x61, 0x76, 0x65, + 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, + 0x20, 0x22, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x22, 0x3a, 0x20, + 0x7b, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x73, + 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x3a, 0x20, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, + 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, + 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, + 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, 0x2c, 0x20, 0x73, + 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, + 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, + 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, + 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x52, + 0x65, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x74, 0x6f, 0x20, 0x64, 0x65, 0x66, + 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, + 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x27, + 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, + 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, + 0x3d, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x20, 0x3d, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, - 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x3b, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x6f, 0x76, 0x65, - 0x72, 0x72, 0x69, 0x64, 0x65, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, - 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, - 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, - 0x74, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, - 0x6e, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, - 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, - 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, - 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, + 0x75, 0x65, 0x5b, 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, + 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, + 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, + 0x70, 0x70, 0x6c, 0x79, 0x28, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x74, 0x2e, 0x64, 0x61, + 0x74, 0x61, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, + 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, + 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, + 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x3a, 0x20, 0x27, 0x27, + 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, + 0x20, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, + 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x69, 0x6d, 0x61, 0x67, + 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, 0x5b, 0x5d, 0x20, 0x7d, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, + 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, + 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, + 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, + 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, + 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, + 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, + 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x4c, 0x6f, 0x61, 0x64, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, + 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x28, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x67, + 0x65, 0x74, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, + 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, + 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, + 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, + 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, - 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, 0x2c, - 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, - 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, - 0x20, 0x6e, 0x6f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x64, 0x65, 0x74, 0x65, - 0x63, 0x74, 0x65, 0x64, 0x2e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, - 0x28, 0x27, 0x49, 0x6e, 0x69, 0x74, 0x69, 0x61, 0x6c, 0x69, 0x7a, 0x69, - 0x6e, 0x67, 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, - 0x61, 0x67, 0x65, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x61, 0x76, 0x69, - 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, - 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x22, 0x64, - 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x22, 0x3a, 0x20, 0x7b, 0x20, 0x73, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, - 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x20, 0x7d, 0x0a, + 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c, + 0x61, 0x73, 0x74, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x66, 0x20, 0x28, 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, 0x65, + 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x29, 0x20, 0x7b, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x41, + 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c, + 0x20, 0x72, 0x65, 0x73, 0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x29, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, + 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x20, 0x3d, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, 0x65, 0x64, 0x54, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x6c, + 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, + 0x67, 0x28, 0x27, 0x4e, 0x6f, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, + 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c, 0x20, 0x75, 0x73, 0x69, 0x6e, + 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x6e, 0x6f, 0x20, 0x61, + 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x6c, 0x61, 0x73, + 0x74, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x20, 0x77, 0x61, 0x73, 0x20, 0x66, 0x6f, 0x75, 0x6e, + 0x64, 0x2c, 0x20, 0x73, 0x6f, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x66, + 0x72, 0x6f, 0x6d, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x2e, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, + 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, + 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, + 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, + 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x41, 0x70, 0x70, 0x6c, 0x79, + 0x69, 0x6e, 0x67, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, + 0x61, 0x6e, 0x64, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x20, 0x69, + 0x6e, 0x74, 0x65, 0x72, 0x6e, 0x61, 0x6c, 0x20, 0x64, 0x61, 0x74, 0x61, + 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x73, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, + 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, + 0x70, 0x70, 0x6c, 0x79, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, + 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, + 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x73, 0x61, + 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, + 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, + 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, + 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x28, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, + 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x54, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x20, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, + 0x65, 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, + 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, + 0x20, 0x3d, 0x3d, 0x20, 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, + 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x2f, 0x2f, 0x20, 0x77, 0x65, 0x20, 0x64, 0x6f, 0x6e, 0x27, 0x74, + 0x20, 0x77, 0x61, 0x6e, 0x74, 0x20, 0x74, 0x6f, 0x20, 0x73, 0x61, 0x76, + 0x65, 0x20, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, + 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, + 0x20, 0x73, 0x6f, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, 0x63, 0x72, + 0x65, 0x61, 0x74, 0x65, 0x20, 0x61, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x6f, + 0x6e, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, + 0x65, 0x74, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x27, 0x55, 0x73, + 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2d, 0x27, + 0x20, 0x2b, 0x20, 0x44, 0x61, 0x74, 0x65, 0x2e, 0x6e, 0x6f, 0x77, 0x28, + 0x29, 0x2e, 0x74, 0x6f, 0x53, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x28, 0x29, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, + 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x27, 0x6e, 0x61, 0x6d, 0x65, 0x27, 0x3a, + 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x27, 0x64, 0x61, 0x74, 0x61, 0x27, + 0x3a, 0x20, 0x7b, 0x20, 0x27, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x27, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x27, 0x70, 0x61, 0x72, 0x61, 0x6d, + 0x73, 0x27, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, + 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x53, 0x61, 0x76, 0x69, 0x6e, + 0x67, 0x20, 0x61, 0x73, 0x20, 0x27, 0x20, 0x2b, 0x20, 0x6e, 0x65, 0x77, + 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, + 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, + 0x2f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x20, 0x69, 0x6e, 0x20, 0x74, 0x68, + 0x65, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x20, 0x73, + 0x6c, 0x6f, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, + 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, + 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, + 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, + 0x6c, 0x61, 0x73, 0x74, 0x27, 0x2c, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x6e, 0x64, 0x20, + 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x69, 0x74, 0x20, 0x62, 0x61, 0x63, 0x6b, + 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x70, 0x70, 0x6c, 0x79, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4c, 0x6f, 0x61, 0x64, 0x41, + 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x41, 0x75, 0x74, 0x6f, 0x73, + 0x61, 0x76, 0x65, 0x64, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, 0x2c, 0x20, 0x73, 0x61, 0x76, 0x65, - 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, - 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, - 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x52, 0x65, 0x73, 0x65, - 0x74, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x20, 0x74, 0x6f, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, - 0x74, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, - 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x27, 0x64, 0x65, 0x66, - 0x61, 0x75, 0x6c, 0x74, 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, - 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x73, - 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x5b, - 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x5d, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, - 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, - 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x70, 0x70, 0x6c, - 0x79, 0x28, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x27, 0x2c, + 0x20, 0x7b, 0x20, 0x27, 0x6e, 0x61, 0x6d, 0x65, 0x27, 0x3a, 0x20, 0x73, + 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x27, 0x64, 0x61, 0x74, + 0x61, 0x27, 0x3a, 0x20, 0x7b, 0x20, 0x27, 0x73, 0x65, 0x73, 0x73, 0x69, + 0x6f, 0x6e, 0x27, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x27, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x27, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x20, 0x7d, 0x29, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, + 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x43, 0x68, 0x65, 0x63, + 0x6b, 0x69, 0x6e, 0x67, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x75, 0x74, + 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, + 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, + 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4c, 0x6f, 0x61, 0x64, + 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x41, 0x75, 0x74, 0x6f, + 0x73, 0x61, 0x76, 0x65, 0x64, 0x28, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x2f, 0x2a, 0x20, 0x45, 0x4e, 0x44, 0x3a, 0x20, 0x53, 0x75, 0x70, + 0x70, 0x6f, 0x72, 0x74, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x73, 0x74, 0x6f, + 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, + 0x64, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x65, 0x74, 0x65, 0x72, 0x73, + 0x20, 0x69, 0x6e, 0x20, 0x62, 0x72, 0x6f, 0x77, 0x73, 0x65, 0x72, 0x73, + 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, + 0x65, 0x20, 0x2a, 0x2f, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, + 0x74, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, + 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, + 0x65, 0x72, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, + 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, + 0x2f, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x6c, 0x79, 0x20, + 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x61, + 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x3f, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, + 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x3d, 0x20, + 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x28, 0x28, 0x29, 0x20, + 0x3d, 0x3e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, + 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x21, 0x3d, 0x20, 0x6e, + 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, + 0x20, 0x68, 0x61, 0x73, 0x20, 0x74, 0x68, 0x65, 0x20, 0x75, 0x73, 0x65, + 0x72, 0x20, 0x73, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, 0x20, 0x61, 0x20, + 0x63, 0x68, 0x61, 0x74, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x74, 0x53, 0x74, 0x61, 0x72, + 0x74, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, + 0x65, 0x64, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, + 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2e, 0x6c, 0x65, + 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, 0x20, 0x30, 0x29, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x72, 0x61, + 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, + 0x65, 0x20, 0x3d, 0x20, 0x28, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, + 0x69, 0x70, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, + 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x72, 0x65, 0x70, + 0x6c, 0x61, 0x63, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, + 0x3d, 0x20, 0x28, 0x73, 0x74, 0x72, 0x2c, 0x20, 0x65, 0x78, 0x74, 0x72, + 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, + 0x74, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x20, 0x3d, 0x20, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, - 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, - 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2c, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x6c, - 0x65, 0x63, 0x74, 0x65, 0x64, 0x3a, 0x20, 0x27, 0x27, 0x20, 0x7d, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x74, 0x2e, - 0x64, 0x61, 0x74, 0x61, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, - 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x64, - 0x61, 0x74, 0x61, 0x3a, 0x20, 0x5b, 0x5d, 0x20, 0x7d, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, - 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, - 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x41, 0x6e, 0x64, - 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, - 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x73, 0x65, 0x6c, - 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, - 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4c, 0x6f, 0x61, - 0x64, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x41, 0x75, 0x74, - 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x67, 0x65, 0x74, 0x20, - 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x6c, 0x61, - 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, - 0x65, 0x74, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, 0x65, 0x64, 0x54, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x6c, 0x6f, - 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, - 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x4f, 0x62, 0x6a, - 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c, 0x61, 0x73, 0x74, - 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, - 0x20, 0x28, 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, 0x65, 0x64, 0x54, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, - 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x41, 0x75, 0x74, 0x6f, - 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c, 0x20, 0x72, 0x65, - 0x73, 0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x29, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, - 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, - 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, - 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, - 0x4e, 0x6f, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, - 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x66, 0x6f, - 0x75, 0x6e, 0x64, 0x2c, 0x20, 0x75, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x64, - 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x6e, 0x6f, 0x20, 0x61, 0x75, 0x74, 0x6f, - 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x75, - 0x73, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x20, 0x77, 0x61, 0x73, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c, 0x20, - 0x73, 0x6f, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x66, 0x72, 0x6f, 0x6d, - 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x2e, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, - 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, - 0x6f, 0x67, 0x28, 0x27, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x69, 0x6e, 0x67, - 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x6e, 0x64, - 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x20, 0x69, 0x6e, 0x74, 0x65, - 0x72, 0x6e, 0x61, 0x6c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x66, 0x72, - 0x6f, 0x6d, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, - 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x70, 0x70, 0x6c, - 0x79, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, - 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x6f, - 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x73, 0x61, 0x76, 0x65, 0x64, - 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, - 0x6f, 0x67, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, - 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, - 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x75, 0x74, - 0x6f, 0x73, 0x61, 0x76, 0x65, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, - 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x20, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x2e, 0x2e, - 0x2e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, - 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, - 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x3d, - 0x20, 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, - 0x20, 0x77, 0x65, 0x20, 0x64, 0x6f, 0x6e, 0x27, 0x74, 0x20, 0x77, 0x61, - 0x6e, 0x74, 0x20, 0x74, 0x6f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x20, 0x6f, - 0x76, 0x65, 0x72, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x20, 0x73, 0x6f, - 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, 0x63, 0x72, 0x65, 0x61, 0x74, - 0x65, 0x20, 0x61, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x6f, 0x6e, 0x65, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, - 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4e, - 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x27, 0x55, 0x73, 0x65, 0x72, 0x54, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2d, 0x27, 0x20, 0x2b, 0x20, - 0x44, 0x61, 0x74, 0x65, 0x2e, 0x6e, 0x6f, 0x77, 0x28, 0x29, 0x2e, 0x74, - 0x6f, 0x53, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x28, 0x29, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6e, 0x65, - 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, - 0x7b, 0x20, 0x27, 0x6e, 0x61, 0x6d, 0x65, 0x27, 0x3a, 0x20, 0x6e, 0x65, - 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4e, 0x61, 0x6d, - 0x65, 0x2c, 0x20, 0x27, 0x64, 0x61, 0x74, 0x61, 0x27, 0x3a, 0x20, 0x7b, - 0x20, 0x27, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x27, 0x3a, 0x20, - 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2c, 0x20, 0x27, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x27, 0x3a, - 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x20, 0x7d, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, - 0x6f, 0x67, 0x28, 0x27, 0x53, 0x61, 0x76, 0x69, 0x6e, 0x67, 0x20, 0x61, - 0x73, 0x20, 0x27, 0x20, 0x2b, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x29, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, - 0x61, 0x76, 0x65, 0x20, 0x69, 0x6e, 0x20, 0x74, 0x68, 0x65, 0x20, 0x61, - 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x20, 0x73, 0x6c, 0x6f, 0x74, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, - 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, - 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, - 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c, 0x61, 0x73, - 0x74, 0x27, 0x2c, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x6c, 0x6f, 0x61, - 0x64, 0x20, 0x69, 0x74, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x20, 0x61, 0x6e, - 0x64, 0x20, 0x61, 0x70, 0x70, 0x6c, 0x79, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x4c, 0x6f, 0x61, 0x64, 0x41, 0x6e, 0x64, 0x41, - 0x70, 0x70, 0x6c, 0x79, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, - 0x64, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, - 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, - 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, - 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, - 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x73, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x27, 0x2c, 0x20, 0x7b, 0x20, - 0x27, 0x6e, 0x61, 0x6d, 0x65, 0x27, 0x3a, 0x20, 0x73, 0x65, 0x6c, 0x65, - 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, - 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x27, 0x64, 0x61, 0x74, 0x61, 0x27, 0x3a, - 0x20, 0x7b, 0x20, 0x27, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x27, - 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x27, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, - 0x27, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x20, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, - 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x43, 0x68, 0x65, 0x63, 0x6b, 0x69, 0x6e, - 0x67, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, - 0x76, 0x65, 0x64, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, - 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4c, 0x6f, 0x61, 0x64, 0x41, 0x6e, 0x64, - 0x41, 0x70, 0x70, 0x6c, 0x79, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, - 0x65, 0x64, 0x28, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2a, - 0x20, 0x45, 0x4e, 0x44, 0x3a, 0x20, 0x53, 0x75, 0x70, 0x70, 0x6f, 0x72, - 0x74, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x73, 0x74, 0x6f, 0x72, 0x69, 0x6e, - 0x67, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x65, 0x74, 0x65, 0x72, 0x73, 0x20, 0x69, 0x6e, - 0x20, 0x62, 0x72, 0x6f, 0x77, 0x73, 0x65, 0x72, 0x73, 0x20, 0x4c, 0x6f, - 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x20, 0x2a, - 0x2f, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x20, - 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x6e, 0x75, 0x6c, - 0x6c, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, - 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x6e, 0x75, 0x6c, - 0x6c, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, - 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x6c, 0x79, 0x20, 0x67, 0x65, 0x6e, - 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x61, 0x20, 0x63, 0x6f, - 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x3f, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x65, 0x6e, 0x65, - 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, - 0x70, 0x75, 0x74, 0x65, 0x64, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x20, 0x21, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, - 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x68, 0x61, - 0x73, 0x20, 0x74, 0x68, 0x65, 0x20, 0x75, 0x73, 0x65, 0x72, 0x20, 0x73, - 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, 0x20, 0x61, 0x20, 0x63, 0x68, 0x61, - 0x74, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x63, 0x68, 0x61, 0x74, 0x53, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, - 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x28, - 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, - 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, - 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, - 0x68, 0x20, 0x3e, 0x20, 0x30, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, - 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x20, 0x3d, - 0x20, 0x28, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, - 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, - 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, - 0x69, 0x70, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, - 0x2f, 0x20, 0x73, 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x74, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, - 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x28, - 0x73, 0x74, 0x72, 0x2c, 0x20, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, - 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x73, - 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x65, - 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, - 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, - 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x7b, - 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, - 0x2c, 0x20, 0x2e, 0x2e, 0x2e, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, - 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x53, 0x74, 0x72, 0x69, 0x6e, - 0x67, 0x28, 0x73, 0x74, 0x72, 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, - 0x63, 0x65, 0x41, 0x6c, 0x6c, 0x28, 0x2f, 0x5c, 0x7b, 0x5c, 0x7b, 0x28, - 0x2e, 0x2a, 0x3f, 0x29, 0x5c, 0x7d, 0x5c, 0x7d, 0x2f, 0x67, 0x2c, 0x20, - 0x28, 0x5f, 0x2c, 0x20, 0x6b, 0x65, 0x79, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x74, - 0x74, 0x69, 0x6e, 0x67, 0x73, 0x5b, 0x6b, 0x65, 0x79, 0x5d, 0x29, 0x29, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, - 0x69, 0x6f, 0x6e, 0x20, 0x72, 0x75, 0x6e, 0x4c, 0x6c, 0x61, 0x6d, 0x61, - 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x6c, 0x6c, 0x61, - 0x6d, 0x61, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x68, - 0x61, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, - 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x20, 0x3d, 0x20, - 0x5b, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x74, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, - 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, - 0x70, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, - 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, - 0x6e, 0x65, 0x77, 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x22, 0x61, - 0x6c, 0x72, 0x65, 0x61, 0x64, 0x79, 0x20, 0x72, 0x75, 0x6e, 0x6e, 0x69, - 0x6e, 0x67, 0x22, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, - 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x41, 0x62, 0x6f, 0x72, 0x74, - 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x28, 0x29, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, - 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, + 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, + 0x20, 0x28, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, + 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, + 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x74, 0x74, 0x69, + 0x6e, 0x67, 0x73, 0x2c, 0x20, 0x2e, 0x2e, 0x2e, 0x65, 0x78, 0x74, 0x72, + 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x53, 0x74, + 0x72, 0x69, 0x6e, 0x67, 0x28, 0x73, 0x74, 0x72, 0x29, 0x2e, 0x72, 0x65, + 0x70, 0x6c, 0x61, 0x63, 0x65, 0x41, 0x6c, 0x6c, 0x28, 0x2f, 0x5c, 0x7b, + 0x5c, 0x7b, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5c, 0x7d, 0x5c, 0x7d, 0x2f, + 0x67, 0x2c, 0x20, 0x28, 0x5f, 0x2c, 0x20, 0x6b, 0x65, 0x79, 0x29, 0x20, + 0x3d, 0x3e, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, + 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x5b, 0x6b, 0x65, 0x79, + 0x5d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x66, 0x75, + 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x72, 0x75, 0x6e, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, - 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, - 0x72, 0x3a, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, - 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x29, 0x29, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x74, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x63, - 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x3b, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, - 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x74, 0x6f, 0x70, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x77, - 0x68, 0x69, 0x6c, 0x65, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, - 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x6c, - 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, 0x20, 0x30, 0x20, 0x26, 0x26, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, - 0x61, 0x67, 0x65, 0x73, 0x5b, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, - 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, - 0x67, 0x74, 0x68, 0x20, 0x2d, 0x20, 0x31, 0x5d, 0x2e, 0x63, 0x6f, 0x6e, - 0x74, 0x65, 0x6e, 0x74, 0x2e, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x28, 0x2f, - 0x5c, 0x6e, 0x24, 0x2f, 0x29, 0x20, 0x21, 0x3d, 0x20, 0x6e, 0x75, 0x6c, - 0x6c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x68, 0x61, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x75, 0x72, + 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, + 0x20, 0x3d, 0x20, 0x5b, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, + 0x72, 0x79, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, + 0x63, 0x72, 0x69, 0x70, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, + 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, + 0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72, + 0x28, 0x22, 0x61, 0x6c, 0x72, 0x65, 0x61, 0x64, 0x79, 0x20, 0x72, 0x75, + 0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x22, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x41, 0x62, + 0x6f, 0x72, 0x74, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, + 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, + 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, + 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, + 0x74, 0x2c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x2c, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, + 0x6c, 0x6c, 0x65, 0x72, 0x3a, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, + 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, + 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, + 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, + 0x66, 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x74, 0x6f, 0x70, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c, 0x65, 0x20, 0x28, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, + 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, + 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, 0x20, 0x30, + 0x20, 0x26, 0x26, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, - 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x70, 0x6f, 0x70, 0x28, - 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5b, 0x63, 0x75, 0x72, 0x72, + 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, + 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x2d, 0x20, 0x31, 0x5d, 0x2e, + 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2e, 0x6d, 0x61, 0x74, 0x63, + 0x68, 0x28, 0x2f, 0x5c, 0x6e, 0x24, 0x2f, 0x29, 0x20, 0x21, 0x3d, 0x20, + 0x6e, 0x75, 0x6c, 0x6c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, + 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x70, + 0x6f, 0x70, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, + 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, + 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x2c, 0x20, 0x5b, 0x63, + 0x68, 0x61, 0x72, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, + 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5d, 0x5d, 0x29, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x22, 0x43, + 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x66, 0x69, + 0x6e, 0x69, 0x73, 0x68, 0x65, 0x64, 0x3a, 0x20, 0x27, 0x22, 0x2c, 0x20, + 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, + 0x67, 0x65, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, + 0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, + 0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, + 0x2c, 0x20, 0x22, 0x27, 0x2c, 0x20, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, + 0x79, 0x3a, 0x20, 0x22, 0x2c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, + 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, + 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x70, 0x75, 0x73, 0x68, + 0x28, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, + 0x64, 0x20, 0x3d, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x6c, 0x6f, + 0x74, 0x5f, 0x69, 0x64, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, + 0x63, 0x74, 0x65, 0x64, 0x5f, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x20, 0x26, + 0x26, 0x20, 0x21, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, 0x75, 0x6c, 0x74, + 0x69, 0x6d, 0x6f, 0x64, 0x61, 0x6c, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x6c, + 0x65, 0x72, 0x74, 0x28, 0x22, 0x54, 0x68, 0x65, 0x20, 0x73, 0x65, 0x72, + 0x76, 0x65, 0x72, 0x20, 0x77, 0x61, 0x73, 0x20, 0x6e, 0x6f, 0x74, 0x20, + 0x63, 0x6f, 0x6d, 0x70, 0x69, 0x6c, 0x65, 0x64, 0x20, 0x66, 0x6f, 0x72, + 0x20, 0x6d, 0x75, 0x6c, 0x74, 0x69, 0x6d, 0x6f, 0x64, 0x61, 0x6c, 0x20, + 0x6f, 0x72, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6d, 0x6f, 0x64, 0x65, 0x6c, + 0x20, 0x70, 0x72, 0x6f, 0x6a, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x20, 0x63, + 0x61, 0x6e, 0x27, 0x74, 0x20, 0x62, 0x65, 0x20, 0x6c, 0x6f, 0x61, 0x64, + 0x65, 0x64, 0x2e, 0x22, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, + 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x2c, 0x20, 0x5b, 0x63, 0x68, 0x61, 0x72, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5d, 0x5d, 0x29, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, - 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x22, 0x43, 0x6f, 0x6d, 0x70, - 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x66, 0x69, 0x6e, 0x69, 0x73, - 0x68, 0x65, 0x64, 0x3a, 0x20, 0x27, 0x22, 0x2c, 0x20, 0x63, 0x75, 0x72, - 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, - 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, - 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, - 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, 0x2c, 0x20, 0x22, - 0x27, 0x2c, 0x20, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3a, 0x20, - 0x22, 0x2c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, - 0x61, 0x67, 0x65, 0x73, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x64, 0x61, - 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x20, 0x3d, - 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, - 0x64, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, - 0x64, 0x5f, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x20, 0x26, 0x26, 0x20, 0x21, - 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, 0x75, 0x6c, 0x74, 0x69, 0x6d, 0x6f, - 0x64, 0x61, 0x6c, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x6c, 0x65, 0x72, 0x74, - 0x28, 0x22, 0x54, 0x68, 0x65, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, - 0x20, 0x77, 0x61, 0x73, 0x20, 0x6e, 0x6f, 0x74, 0x20, 0x63, 0x6f, 0x6d, - 0x70, 0x69, 0x6c, 0x65, 0x64, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x6d, 0x75, - 0x6c, 0x74, 0x69, 0x6d, 0x6f, 0x64, 0x61, 0x6c, 0x20, 0x6f, 0x72, 0x20, - 0x74, 0x68, 0x65, 0x20, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x20, 0x70, 0x72, - 0x6f, 0x6a, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x20, 0x63, 0x61, 0x6e, 0x27, - 0x74, 0x20, 0x62, 0x65, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x65, 0x64, 0x2e, - 0x22, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, - 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, - 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, - 0x72, 0x79, 0x2c, 0x20, 0x5b, 0x63, 0x68, 0x61, 0x72, 0x2c, 0x20, 0x63, - 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, - 0x65, 0x73, 0x5d, 0x5d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x69, 0x66, 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, - 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, - 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x64, 0x61, 0x74, 0x61, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, - 0x73, 0x65, 0x6e, 0x64, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, - 0x20, 0x74, 0x6f, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, - 0x74, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x6d, - 0x73, 0x67, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, - 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, - 0x61, 0x6c, 0x72, 0x65, 0x61, 0x64, 0x79, 0x20, 0x72, 0x75, 0x6e, 0x6e, - 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, - 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, - 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, - 0x70, 0x74, 0x2c, 0x20, 0x5b, 0x22, 0x7b, 0x7b, 0x75, 0x73, 0x65, 0x72, - 0x7d, 0x7d, 0x22, 0x2c, 0x20, 0x6d, 0x73, 0x67, 0x5d, 0x5d, 0x29, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x70, - 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x2c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x3a, 0x20, - 0x6d, 0x73, 0x67, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x3a, 0x20, 0x73, 0x65, - 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, - 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2e, 0x66, - 0x6c, 0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x28, 0x5b, 0x6e, 0x61, 0x6d, 0x65, - 0x2c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x5d, 0x29, 0x20, 0x3d, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x3a, 0x20, 0x41, 0x72, - 0x72, 0x61, 0x79, 0x2e, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x28, - 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, - 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, - 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, - 0x27, 0x27, 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, - 0x2f, 0x5e, 0x5c, 0x73, 0x2f, 0x2c, 0x20, 0x27, 0x27, 0x29, 0x20, 0x3a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2c, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x22, 0x5c, - 0x6e, 0x22, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, - 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x69, 0x6d, - 0x61, 0x67, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x3d, 0x20, - 0x60, 0x41, 0x20, 0x63, 0x68, 0x61, 0x74, 0x20, 0x62, 0x65, 0x74, 0x77, - 0x65, 0x65, 0x6e, 0x20, 0x61, 0x20, 0x63, 0x75, 0x72, 0x69, 0x6f, 0x75, - 0x73, 0x20, 0x68, 0x75, 0x6d, 0x61, 0x6e, 0x20, 0x61, 0x6e, 0x64, 0x20, - 0x61, 0x6e, 0x20, 0x61, 0x72, 0x74, 0x69, 0x66, 0x69, 0x63, 0x69, 0x61, - 0x6c, 0x20, 0x69, 0x6e, 0x74, 0x65, 0x6c, 0x6c, 0x69, 0x67, 0x65, 0x6e, - 0x63, 0x65, 0x20, 0x61, 0x73, 0x73, 0x69, 0x73, 0x74, 0x61, 0x6e, 0x74, - 0x2e, 0x20, 0x54, 0x68, 0x65, 0x20, 0x61, 0x73, 0x73, 0x69, 0x73, 0x74, - 0x61, 0x6e, 0x74, 0x20, 0x67, 0x69, 0x76, 0x65, 0x73, 0x20, 0x68, 0x65, - 0x6c, 0x70, 0x66, 0x75, 0x6c, 0x2c, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, - 0x6c, 0x65, 0x64, 0x2c, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x70, 0x6f, 0x6c, - 0x69, 0x74, 0x65, 0x20, 0x61, 0x6e, 0x73, 0x77, 0x65, 0x72, 0x73, 0x20, - 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x68, 0x75, 0x6d, 0x61, 0x6e, - 0x27, 0x73, 0x20, 0x71, 0x75, 0x65, 0x73, 0x74, 0x69, 0x6f, 0x6e, 0x73, - 0x2e, 0x5c, 0x6e, 0x55, 0x53, 0x45, 0x52, 0x3a, 0x5b, 0x69, 0x6d, 0x67, - 0x2d, 0x31, 0x30, 0x5d, 0x24, 0x7b, 0x6d, 0x73, 0x67, 0x7d, 0x5c, 0x6e, - 0x41, 0x53, 0x53, 0x49, 0x53, 0x54, 0x41, 0x4e, 0x54, 0x3a, 0x60, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x72, 0x75, 0x6e, - 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, - 0x2c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x3a, 0x20, 0x73, 0x6c, - 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x5b, 0x22, 0x3c, - 0x2f, 0x73, 0x3e, 0x22, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x28, 0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, - 0x3a, 0x22, 0x29, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x28, 0x22, 0x7b, 0x7b, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x7d, 0x3a, - 0x22, 0x29, 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x2c, 0x20, 0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x22, - 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x75, 0x6e, 0x43, - 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, - 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, - 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x61, - 0x6c, 0x72, 0x65, 0x61, 0x64, 0x79, 0x20, 0x72, 0x75, 0x6e, 0x6e, 0x69, - 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x70, - 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x7d, 0x20, 0x3d, 0x20, 0x73, 0x65, - 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, - 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, - 0x5b, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, - 0x72, 0x69, 0x70, 0x74, 0x2c, 0x20, 0x5b, 0x22, 0x22, 0x2c, 0x20, 0x70, - 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x5d, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6e, 0x4c, 0x6c, 0x61, 0x6d, 0x61, - 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x6c, 0x6f, 0x74, - 0x5f, 0x69, 0x64, 0x3a, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, - 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, - 0x6f, 0x70, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x22, 0x22, 0x29, 0x2e, 0x66, 0x69, 0x6e, - 0x61, 0x6c, 0x6c, 0x79, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, - 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, - 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, - 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2e, 0x6d, 0x61, 0x70, - 0x28, 0x28, 0x5b, 0x5f, 0x2c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x5d, 0x29, - 0x20, 0x3d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x41, 0x72, 0x72, 0x61, 0x79, 0x2e, 0x69, 0x73, 0x41, 0x72, - 0x72, 0x61, 0x79, 0x28, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x3f, 0x20, - 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, - 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, - 0x65, 0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, - 0x29, 0x20, 0x3a, 0x20, 0x64, 0x61, 0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, - 0x27, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, + 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6c, + 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x20, 0x3d, 0x20, 0x64, 0x61, 0x74, 0x61, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x2f, 0x2f, 0x20, 0x73, 0x65, 0x6e, 0x64, 0x20, 0x6d, 0x65, 0x73, 0x73, + 0x61, 0x67, 0x65, 0x20, 0x74, 0x6f, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, + 0x72, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x63, 0x68, 0x61, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, + 0x20, 0x28, 0x6d, 0x73, 0x67, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, + 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, + 0x67, 0x28, 0x27, 0x61, 0x6c, 0x72, 0x65, 0x61, 0x64, 0x79, 0x20, 0x72, + 0x75, 0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, + 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, + 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, + 0x63, 0x72, 0x69, 0x70, 0x74, 0x2c, 0x20, 0x5b, 0x22, 0x7b, 0x7b, 0x75, + 0x73, 0x65, 0x72, 0x7d, 0x7d, 0x22, 0x2c, 0x20, 0x6d, 0x73, 0x67, 0x5d, + 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, + 0x74, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x3d, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, + 0x65, 0x3a, 0x20, 0x6d, 0x73, 0x67, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, - 0x74, 0x20, 0x3d, 0x20, 0x5b, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x74, - 0x6f, 0x70, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x2e, 0x70, 0x72, - 0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, - 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, - 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, - 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x61, - 0x62, 0x6f, 0x72, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, - 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6e, - 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x73, 0x65, 0x74, 0x20, - 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x28, 0x65, 0x29, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, - 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, - 0x28, 0x5b, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, - 0x70, 0x6c, 0x6f, 0x61, 0x64, 0x49, 0x6d, 0x61, 0x67, 0x65, 0x20, 0x3d, - 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x76, 0x65, 0x6e, - 0x74, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, - 0x6e, 0x74, 0x2e, 0x67, 0x65, 0x74, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, - 0x74, 0x42, 0x79, 0x49, 0x64, 0x28, 0x22, 0x66, 0x69, 0x6c, 0x65, 0x49, - 0x6e, 0x70, 0x75, 0x74, 0x22, 0x29, 0x2e, 0x63, 0x6c, 0x69, 0x63, 0x6b, - 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, - 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x67, 0x65, 0x74, 0x45, 0x6c, - 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x42, 0x79, 0x49, 0x64, 0x28, 0x22, 0x66, - 0x69, 0x6c, 0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, 0x22, 0x29, 0x2e, 0x61, - 0x64, 0x64, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, - 0x6e, 0x65, 0x72, 0x28, 0x22, 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x22, - 0x2c, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x28, - 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, - 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x46, 0x69, 0x6c, 0x65, 0x20, - 0x3d, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x74, 0x61, 0x72, 0x67, - 0x65, 0x74, 0x2e, 0x66, 0x69, 0x6c, 0x65, 0x73, 0x5b, 0x30, 0x5d, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, - 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x46, 0x69, 0x6c, - 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x61, - 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x46, 0x69, - 0x6c, 0x65, 0x52, 0x65, 0x61, 0x64, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, - 0x61, 0x64, 0x65, 0x72, 0x2e, 0x6f, 0x6e, 0x6c, 0x6f, 0x61, 0x64, 0x20, - 0x3d, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x28, - 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x6d, - 0x61, 0x67, 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x72, - 0x65, 0x61, 0x64, 0x65, 0x72, 0x2e, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2c, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x6c, 0x65, - 0x63, 0x74, 0x65, 0x64, 0x3a, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, - 0x64, 0x61, 0x74, 0x61, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, + 0x74, 0x2e, 0x66, 0x6c, 0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x28, 0x5b, 0x6e, + 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x5d, 0x29, 0x20, + 0x3d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, + 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x3a, + 0x20, 0x41, 0x72, 0x72, 0x61, 0x79, 0x2e, 0x69, 0x73, 0x41, 0x72, 0x72, + 0x61, 0x79, 0x28, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x3f, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, 0x61, + 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, + 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, + 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, + 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x5c, 0x73, 0x2f, 0x2c, 0x20, 0x27, 0x27, + 0x29, 0x20, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x61, + 0x74, 0x61, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, + 0x28, 0x22, 0x5c, 0x6e, 0x22, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, + 0x5f, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, + 0x20, 0x3d, 0x20, 0x60, 0x41, 0x20, 0x63, 0x68, 0x61, 0x74, 0x20, 0x62, + 0x65, 0x74, 0x77, 0x65, 0x65, 0x6e, 0x20, 0x61, 0x20, 0x63, 0x75, 0x72, + 0x69, 0x6f, 0x75, 0x73, 0x20, 0x68, 0x75, 0x6d, 0x61, 0x6e, 0x20, 0x61, + 0x6e, 0x64, 0x20, 0x61, 0x6e, 0x20, 0x61, 0x72, 0x74, 0x69, 0x66, 0x69, + 0x63, 0x69, 0x61, 0x6c, 0x20, 0x69, 0x6e, 0x74, 0x65, 0x6c, 0x6c, 0x69, + 0x67, 0x65, 0x6e, 0x63, 0x65, 0x20, 0x61, 0x73, 0x73, 0x69, 0x73, 0x74, + 0x61, 0x6e, 0x74, 0x2e, 0x20, 0x54, 0x68, 0x65, 0x20, 0x61, 0x73, 0x73, + 0x69, 0x73, 0x74, 0x61, 0x6e, 0x74, 0x20, 0x67, 0x69, 0x76, 0x65, 0x73, + 0x20, 0x68, 0x65, 0x6c, 0x70, 0x66, 0x75, 0x6c, 0x2c, 0x20, 0x64, 0x65, + 0x74, 0x61, 0x69, 0x6c, 0x65, 0x64, 0x2c, 0x20, 0x61, 0x6e, 0x64, 0x20, + 0x70, 0x6f, 0x6c, 0x69, 0x74, 0x65, 0x20, 0x61, 0x6e, 0x73, 0x77, 0x65, + 0x72, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x68, 0x75, + 0x6d, 0x61, 0x6e, 0x27, 0x73, 0x20, 0x71, 0x75, 0x65, 0x73, 0x74, 0x69, + 0x6f, 0x6e, 0x73, 0x2e, 0x5c, 0x6e, 0x55, 0x53, 0x45, 0x52, 0x3a, 0x5b, + 0x69, 0x6d, 0x67, 0x2d, 0x31, 0x30, 0x5d, 0x24, 0x7b, 0x6d, 0x73, 0x67, + 0x7d, 0x5c, 0x6e, 0x41, 0x53, 0x53, 0x49, 0x53, 0x54, 0x41, 0x4e, 0x54, + 0x3a, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, + 0x72, 0x75, 0x6e, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, + 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x69, 0x6d, 0x61, 0x67, - 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, 0x5b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7b, 0x20, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, 0x69, 0x6d, - 0x61, 0x67, 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x72, 0x65, 0x70, - 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x69, - 0x6d, 0x61, 0x67, 0x65, 0x5c, 0x2f, 0x5b, 0x5e, 0x3b, 0x5d, 0x2b, 0x3b, - 0x62, 0x61, 0x73, 0x65, 0x36, 0x34, 0x2c, 0x2f, 0x2c, 0x20, 0x27, 0x27, - 0x29, 0x2c, 0x20, 0x69, 0x64, 0x3a, 0x20, 0x31, 0x30, 0x20, 0x7d, 0x5d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x69, - 0x6d, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, - 0x65, 0x61, 0x64, 0x65, 0x72, 0x2e, 0x72, 0x65, 0x61, 0x64, 0x41, 0x73, - 0x44, 0x61, 0x74, 0x61, 0x55, 0x52, 0x4c, 0x28, 0x73, 0x65, 0x6c, 0x65, - 0x63, 0x74, 0x65, 0x64, 0x46, 0x69, 0x6c, 0x65, 0x29, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, - 0x6f, 0x6e, 0x20, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x49, 0x6e, - 0x70, 0x75, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x65, 0x73, 0x73, - 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, - 0x6e, 0x61, 0x6c, 0x28, 0x22, 0x22, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x75, 0x62, - 0x6d, 0x69, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, - 0x74, 0x6f, 0x70, 0x28, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x68, 0x61, 0x74, 0x28, 0x6d, 0x65, 0x73, - 0x73, 0x61, 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, - 0x73, 0x61, 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, - 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x53, 0x75, 0x62, 0x6d, 0x69, - 0x74, 0x73, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, - 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, - 0x77, 0x68, 0x69, 0x63, 0x68, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x31, 0x33, - 0x20, 0x26, 0x26, 0x20, 0x21, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x73, - 0x68, 0x69, 0x66, 0x74, 0x4b, 0x65, 0x79, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x75, 0x62, - 0x6d, 0x69, 0x74, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, - 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, - 0x6f, 0x72, 0x6d, 0x20, 0x6f, 0x6e, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, - 0x3d, 0x24, 0x7b, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x7d, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, - 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, - 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x4e, 0x61, - 0x6d, 0x65, 0x3d, 0x24, 0x7b, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, - 0x69, 0x6e, 0x67, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3f, 0x20, - 0x22, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x22, 0x20, 0x3a, 0x20, - 0x6e, 0x75, 0x6c, 0x6c, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, 0x6e, 0x69, - 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x28, 0x65, 0x29, 0x20, 0x3d, - 0x3e, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x65, 0x2e, 0x74, 0x61, 0x72, 0x67, - 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x6f, 0x6e, 0x6b, 0x65, 0x79, 0x70, 0x72, 0x65, 0x73, 0x73, 0x3d, - 0x24, 0x7b, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x53, 0x75, 0x62, 0x6d, 0x69, - 0x74, 0x73, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, - 0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x53, 0x61, 0x79, 0x20, - 0x73, 0x6f, 0x6d, 0x65, 0x74, 0x68, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, - 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x32, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, - 0x74, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, - 0x22, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x22, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, - 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x72, 0x69, 0x67, 0x68, 0x74, - 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x74, - 0x79, 0x70, 0x65, 0x3d, 0x22, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x22, - 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x3d, 0x24, 0x7b, - 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3e, 0x53, 0x65, 0x6e, 0x64, 0x3c, 0x2f, - 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, - 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, - 0x24, 0x7b, 0x75, 0x70, 0x6c, 0x6f, 0x61, 0x64, 0x49, 0x6d, 0x61, 0x67, - 0x65, 0x7d, 0x3e, 0x55, 0x70, 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x49, 0x6d, - 0x61, 0x67, 0x65, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, - 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x73, 0x74, 0x6f, 0x70, 0x7d, - 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x3d, 0x24, 0x7b, - 0x21, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3e, 0x53, 0x74, 0x6f, 0x70, 0x3c, - 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, - 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, - 0x3d, 0x24, 0x7b, 0x72, 0x65, 0x73, 0x65, 0x74, 0x7d, 0x3e, 0x52, 0x65, - 0x73, 0x65, 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, - 0x6e, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, - 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x73, 0x28, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, - 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x28, 0x65, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6e, 0x43, - 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x29, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, - 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, - 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x73, 0x75, - 0x62, 0x6d, 0x69, 0x74, 0x7d, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, - 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x22, 0x20, 0x64, 0x69, 0x73, 0x61, - 0x62, 0x6c, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x67, 0x65, 0x6e, 0x65, 0x72, - 0x61, 0x74, 0x69, 0x6e, 0x67, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, - 0x3e, 0x53, 0x74, 0x61, 0x72, 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, - 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, - 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x73, 0x74, 0x6f, 0x70, - 0x7d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x3d, 0x24, - 0x7b, 0x21, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3e, 0x53, 0x74, 0x6f, 0x70, - 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, - 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, - 0x24, 0x7b, 0x72, 0x65, 0x73, 0x65, 0x74, 0x7d, 0x3e, 0x52, 0x65, 0x73, - 0x65, 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, - 0x76, 0x3e, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x68, - 0x61, 0x74, 0x4c, 0x6f, 0x67, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, - 0x70, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x65, 0x73, - 0x73, 0x61, 0x67, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, - 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, - 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, - 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x75, 0x73, - 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, - 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, - 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x20, 0x74, 0x6f, 0x20, 0x62, 0x6f, 0x74, - 0x74, 0x6f, 0x6d, 0x20, 0x28, 0x69, 0x66, 0x20, 0x6e, 0x65, 0x65, 0x64, - 0x65, 0x64, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, - 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, - 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x70, 0x61, 0x72, - 0x65, 0x6e, 0x74, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, - 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x20, 0x26, 0x26, 0x20, 0x70, 0x61, - 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x48, - 0x65, 0x69, 0x67, 0x68, 0x74, 0x20, 0x3c, 0x3d, 0x20, 0x70, 0x61, 0x72, - 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x54, 0x6f, - 0x70, 0x20, 0x2b, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x6f, - 0x66, 0x66, 0x73, 0x65, 0x74, 0x48, 0x65, 0x69, 0x67, 0x68, 0x74, 0x20, - 0x2b, 0x20, 0x33, 0x30, 0x30, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, - 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x54, 0x6f, 0x28, 0x30, - 0x2c, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, - 0x6f, 0x6c, 0x6c, 0x48, 0x65, 0x69, 0x67, 0x68, 0x74, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x6d, 0x65, 0x73, 0x73, 0x61, - 0x67, 0x65, 0x73, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x73, 0x43, 0x6f, 0x6d, - 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x4d, 0x6f, 0x64, 0x65, 0x20, - 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, - 0x20, 0x27, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, - 0x27, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x63, 0x68, 0x61, 0x74, 0x4c, 0x69, 0x6e, 0x65, 0x20, 0x3d, - 0x20, 0x28, 0x5b, 0x75, 0x73, 0x65, 0x72, 0x2c, 0x20, 0x64, 0x61, 0x74, - 0x61, 0x5d, 0x2c, 0x20, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x29, 0x20, 0x3d, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x3a, + 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x2c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x3a, 0x20, + 0x5b, 0x22, 0x3c, 0x2f, 0x73, 0x3e, 0x22, 0x2c, 0x20, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, + 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x29, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x28, 0x22, 0x7b, 0x7b, 0x75, 0x73, 0x65, 0x72, + 0x7d, 0x7d, 0x3a, 0x22, 0x29, 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, + 0x7d, 0x7d, 0x22, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, + 0x75, 0x6e, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, + 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, + 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, + 0x28, 0x27, 0x61, 0x6c, 0x72, 0x65, 0x61, 0x64, 0x79, 0x20, 0x72, 0x75, + 0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x7b, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x7d, 0x20, 0x3d, + 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, + 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, + 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, + 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, + 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2c, 0x20, 0x5b, 0x22, 0x22, + 0x2c, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x5d, 0x5d, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6e, 0x4c, 0x6c, + 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, + 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, + 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x3a, 0x20, 0x73, 0x6c, 0x6f, 0x74, + 0x5f, 0x69, 0x64, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x73, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x22, 0x22, 0x29, 0x2e, + 0x66, 0x69, 0x6e, 0x61, 0x6c, 0x6c, 0x79, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x6c, 0x65, 0x74, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x4d, 0x65, 0x73, - 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x41, 0x72, 0x72, 0x61, 0x79, - 0x2e, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x28, 0x64, 0x61, 0x74, - 0x61, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, - 0x66, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x20, - 0x3e, 0x20, 0x30, 0x20, 0x26, 0x26, 0x20, 0x69, 0x73, 0x41, 0x72, 0x72, - 0x61, 0x79, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, - 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x68, 0x74, 0x6d, - 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, - 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x7d, 0x20, 0x64, 0x61, 0x74, 0x61, - 0x3d, 0x24, 0x7b, 0x64, 0x61, 0x74, 0x61, 0x7d, 0x20, 0x2f, 0x3e, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, - 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x65, - 0x78, 0x74, 0x20, 0x3d, 0x20, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, - 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3f, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x61, - 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, - 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, - 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, 0x2e, - 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x5c, 0x73, - 0x2b, 0x2f, 0x2c, 0x20, 0x27, 0x27, 0x29, 0x20, 0x3a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x61, - 0x74, 0x61, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, - 0x69, 0x73, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, - 0x4d, 0x6f, 0x64, 0x65, 0x20, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, - 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x4d, 0x61, - 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x69, 0x73, 0x68, 0x7d, 0x20, 0x74, - 0x65, 0x78, 0x74, 0x3d, 0x24, 0x7b, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x28, 0x74, 0x65, 0x78, 0x74, 0x29, 0x7d, 0x20, 0x2f, 0x3e, - 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, - 0x75, 0x73, 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, - 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x70, 0x20, 0x6b, 0x65, 0x79, - 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x3e, 0x3c, 0x73, - 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x7b, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x28, 0x75, 0x73, 0x65, 0x72, 0x29, 0x7d, 0x3a, - 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x20, 0x24, 0x7b, - 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x3c, 0x2f, 0x70, 0x3e, - 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, - 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, - 0x69, 0x73, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, - 0x4d, 0x6f, 0x64, 0x65, 0x20, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, - 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, - 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x3e, 0x24, 0x7b, 0x6d, 0x65, 0x73, - 0x73, 0x61, 0x67, 0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, - 0x60, 0x20, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x70, 0x20, - 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, - 0x3e, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x3c, - 0x2f, 0x70, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6f, 0x6d, 0x70, 0x6c, - 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x45, 0x64, 0x69, 0x74, 0x20, 0x3d, 0x20, - 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, - 0x74, 0x20, 0x3d, 0x20, 0x65, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, - 0x2e, 0x69, 0x6e, 0x6e, 0x65, 0x72, 0x54, 0x65, 0x78, 0x74, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, - 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, - 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x20, 0x3d, 0x20, 0x5b, - 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, - 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x69, 0x64, 0x3d, 0x22, - 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x72, 0x65, 0x66, 0x3d, 0x24, 0x7b, - 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x7d, 0x20, 0x6b, - 0x65, 0x79, 0x3d, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, - 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x7d, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6d, - 0x67, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x22, 0x77, 0x69, 0x64, - 0x74, 0x68, 0x3a, 0x20, 0x36, 0x30, 0x25, 0x3b, 0x24, 0x7b, 0x21, 0x73, + 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x6c, 0x65, 0x63, - 0x74, 0x65, 0x64, 0x20, 0x3f, 0x20, 0x60, 0x64, 0x69, 0x73, 0x70, 0x6c, - 0x61, 0x79, 0x3a, 0x20, 0x6e, 0x6f, 0x6e, 0x65, 0x3b, 0x60, 0x20, 0x3a, - 0x20, 0x60, 0x60, 0x7d, 0x22, 0x20, 0x73, 0x72, 0x63, 0x3d, 0x22, 0x24, - 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2e, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x6c, - 0x65, 0x63, 0x74, 0x65, 0x64, 0x7d, 0x22, 0x2f, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, - 0x6e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x65, 0x64, 0x69, - 0x74, 0x61, 0x62, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x69, 0x73, 0x43, 0x6f, - 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x4d, 0x6f, 0x64, 0x65, - 0x7d, 0x20, 0x72, 0x65, 0x66, 0x3d, 0x24, 0x7b, 0x63, 0x6f, 0x6e, 0x74, - 0x61, 0x69, 0x6e, 0x65, 0x72, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, - 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, - 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x45, 0x64, 0x69, - 0x74, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, - 0x65, 0x73, 0x2e, 0x66, 0x6c, 0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, 0x63, - 0x68, 0x61, 0x74, 0x4c, 0x69, 0x6e, 0x65, 0x29, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x70, - 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, - 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, - 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, - 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2e, + 0x6d, 0x61, 0x70, 0x28, 0x28, 0x5b, 0x5f, 0x2c, 0x20, 0x64, 0x61, 0x74, + 0x61, 0x5d, 0x29, 0x20, 0x3d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x41, 0x72, 0x72, 0x61, 0x79, 0x2e, 0x69, + 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x28, 0x64, 0x61, 0x74, 0x61, 0x29, + 0x20, 0x3f, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, + 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, + 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, + 0x28, 0x27, 0x27, 0x29, 0x20, 0x3a, 0x20, 0x64, 0x61, 0x74, 0x61, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x2e, 0x6a, 0x6f, + 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, + 0x72, 0x69, 0x70, 0x74, 0x20, 0x3d, 0x20, 0x5b, 0x5d, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x73, 0x74, 0x6f, 0x70, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, + 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, + 0x2e, 0x70, 0x72, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, 0x61, + 0x75, 0x6c, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, + 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, + 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, + 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x73, + 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, + 0x28, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, + 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, + 0x61, 0x74, 0x65, 0x28, 0x5b, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x75, 0x70, 0x6c, 0x6f, 0x61, 0x64, 0x49, 0x6d, 0x61, 0x67, + 0x65, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x2e, 0x70, 0x72, 0x65, + 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, + 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x67, 0x65, 0x74, 0x45, 0x6c, 0x65, + 0x6d, 0x65, 0x6e, 0x74, 0x42, 0x79, 0x49, 0x64, 0x28, 0x22, 0x66, 0x69, + 0x6c, 0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, 0x22, 0x29, 0x2e, 0x63, 0x6c, + 0x69, 0x63, 0x6b, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x67, 0x65, + 0x74, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x42, 0x79, 0x49, 0x64, + 0x28, 0x22, 0x66, 0x69, 0x6c, 0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, 0x22, + 0x29, 0x2e, 0x61, 0x64, 0x64, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, + 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28, 0x22, 0x63, 0x68, 0x61, 0x6e, + 0x67, 0x65, 0x22, 0x2c, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, + 0x6e, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x46, 0x69, + 0x6c, 0x65, 0x20, 0x3d, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x74, + 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x66, 0x69, 0x6c, 0x65, 0x73, 0x5b, + 0x30, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, + 0x46, 0x69, 0x6c, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x72, 0x65, 0x61, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, + 0x20, 0x46, 0x69, 0x6c, 0x65, 0x52, 0x65, 0x61, 0x64, 0x65, 0x72, 0x28, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x72, 0x65, 0x61, 0x64, 0x65, 0x72, 0x2e, 0x6f, 0x6e, 0x6c, 0x6f, + 0x61, 0x64, 0x20, 0x3d, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, + 0x6e, 0x20, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x20, + 0x3d, 0x20, 0x72, 0x65, 0x61, 0x64, 0x65, 0x72, 0x2e, 0x72, 0x65, 0x73, + 0x75, 0x6c, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, + 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x3a, 0x20, 0x69, 0x6d, 0x61, + 0x67, 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x20, 0x7d, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, + 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x69, + 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, 0x5b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7b, 0x20, 0x64, 0x61, 0x74, 0x61, 0x3a, + 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x2e, + 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x64, 0x61, 0x74, + 0x61, 0x3a, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5c, 0x2f, 0x5b, 0x5e, 0x3b, + 0x5d, 0x2b, 0x3b, 0x62, 0x61, 0x73, 0x65, 0x36, 0x34, 0x2c, 0x2f, 0x2c, + 0x20, 0x27, 0x27, 0x29, 0x2c, 0x20, 0x69, 0x64, 0x3a, 0x20, 0x31, 0x30, + 0x20, 0x7d, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, + 0x64, 0x5f, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x74, 0x72, + 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x61, 0x64, 0x65, 0x72, 0x2e, 0x72, 0x65, 0x61, + 0x64, 0x41, 0x73, 0x44, 0x61, 0x74, 0x61, 0x55, 0x52, 0x4c, 0x28, 0x73, + 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x46, 0x69, 0x6c, 0x65, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, + 0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, + 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, + 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x22, 0x22, 0x29, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, + 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x28, 0x65, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x68, 0x61, 0x74, 0x28, + 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x53, 0x75, + 0x62, 0x6d, 0x69, 0x74, 0x73, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x76, 0x65, + 0x6e, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x76, 0x65, + 0x6e, 0x74, 0x2e, 0x77, 0x68, 0x69, 0x63, 0x68, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x31, 0x33, 0x20, 0x26, 0x26, 0x20, 0x21, 0x65, 0x76, 0x65, 0x6e, + 0x74, 0x2e, 0x73, 0x68, 0x69, 0x66, 0x74, 0x4b, 0x65, 0x79, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, + 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x66, 0x6f, 0x72, 0x6d, 0x20, 0x6f, 0x6e, 0x73, 0x75, 0x62, + 0x6d, 0x69, 0x74, 0x3d, 0x24, 0x7b, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, + 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, + 0x61, 0x72, 0x65, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, + 0x73, 0x4e, 0x61, 0x6d, 0x65, 0x3d, 0x24, 0x7b, 0x67, 0x65, 0x6e, 0x65, + 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x20, 0x3f, 0x20, 0x22, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x22, + 0x20, 0x3a, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x28, 0x65, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x65, 0x2e, 0x74, + 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x6f, 0x6e, 0x6b, 0x65, 0x79, 0x70, 0x72, 0x65, + 0x73, 0x73, 0x3d, 0x24, 0x7b, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x53, 0x75, + 0x62, 0x6d, 0x69, 0x74, 0x73, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x6c, + 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x53, + 0x61, 0x79, 0x20, 0x73, 0x6f, 0x6d, 0x65, 0x74, 0x68, 0x69, 0x6e, 0x67, + 0x2e, 0x2e, 0x2e, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x6f, 0x77, 0x73, + 0x3d, 0x32, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, + 0x74, 0x65, 0x78, 0x74, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, + 0x65, 0x7d, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, + 0x69, 0x76, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x72, 0x69, + 0x67, 0x68, 0x74, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, + 0x6e, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x73, 0x75, 0x62, 0x6d, + 0x69, 0x74, 0x22, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, + 0x3d, 0x24, 0x7b, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, + 0x67, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3e, 0x53, 0x65, 0x6e, + 0x64, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, + 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x6c, 0x6f, 0x61, 0x64, 0x49, + 0x6d, 0x61, 0x67, 0x65, 0x7d, 0x3e, 0x55, 0x70, 0x6c, 0x6f, 0x61, 0x64, + 0x20, 0x49, 0x6d, 0x61, 0x67, 0x65, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, + 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, + 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x73, 0x74, + 0x6f, 0x70, 0x7d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, + 0x3d, 0x24, 0x7b, 0x21, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, + 0x6e, 0x67, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3e, 0x53, 0x74, + 0x6f, 0x70, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, + 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x72, 0x65, 0x73, 0x65, 0x74, 0x7d, + 0x3e, 0x52, 0x65, 0x73, 0x65, 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, + 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x72, 0x6d, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, + 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, + 0x69, 0x6f, 0x6e, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x73, 0x28, + 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x20, 0x3d, + 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x28, 0x65, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x75, 0x6e, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, + 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, + 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, + 0x7b, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x7d, 0x20, 0x74, 0x79, 0x70, + 0x65, 0x3d, 0x22, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x22, 0x20, 0x64, + 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x67, 0x65, + 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x7d, 0x3e, 0x53, 0x74, 0x61, 0x72, 0x74, 0x3c, 0x2f, 0x62, + 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, + 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x73, + 0x74, 0x6f, 0x70, 0x7d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, + 0x64, 0x3d, 0x24, 0x7b, 0x21, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, + 0x69, 0x6e, 0x67, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3e, 0x53, + 0x74, 0x6f, 0x70, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, + 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x72, 0x65, 0x73, 0x65, 0x74, 0x7d, 0x3e, + 0x52, 0x65, 0x73, 0x65, 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, + 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x43, 0x68, 0x61, 0x74, 0x4c, 0x6f, 0x67, 0x20, 0x3d, 0x20, 0x28, + 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x73, + 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x20, 0x3d, + 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, + 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, + 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, + 0x2f, 0x20, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x20, 0x74, 0x6f, 0x20, + 0x62, 0x6f, 0x74, 0x74, 0x6f, 0x6d, 0x20, 0x28, 0x69, 0x66, 0x20, 0x6e, + 0x65, 0x65, 0x64, 0x65, 0x64, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, 0x72, + 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, + 0x6e, 0x65, 0x72, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, + 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, + 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, + 0x66, 0x20, 0x28, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x20, 0x26, 0x26, + 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, + 0x6c, 0x6c, 0x48, 0x65, 0x69, 0x67, 0x68, 0x74, 0x20, 0x3c, 0x3d, 0x20, + 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, + 0x6c, 0x54, 0x6f, 0x70, 0x20, 0x2b, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, + 0x74, 0x2e, 0x6f, 0x66, 0x66, 0x73, 0x65, 0x74, 0x48, 0x65, 0x69, 0x67, + 0x68, 0x74, 0x20, 0x2b, 0x20, 0x33, 0x30, 0x30, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, + 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x54, + 0x6f, 0x28, 0x30, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, + 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x48, 0x65, 0x69, 0x67, 0x68, 0x74, + 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x6d, 0x65, + 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x73, + 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x4d, 0x6f, + 0x64, 0x65, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, + 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, + 0x69, 0x6f, 0x6e, 0x27, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x74, 0x4c, 0x69, 0x6e, + 0x65, 0x20, 0x3d, 0x20, 0x28, 0x5b, 0x75, 0x73, 0x65, 0x72, 0x2c, 0x20, + 0x64, 0x61, 0x74, 0x61, 0x5d, 0x2c, 0x20, 0x69, 0x6e, 0x64, 0x65, 0x78, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, + 0x67, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, + 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x41, 0x72, + 0x72, 0x61, 0x79, 0x2e, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x28, + 0x64, 0x61, 0x74, 0x61, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x6f, + 0x62, 0x73, 0x20, 0x3e, 0x20, 0x30, 0x20, 0x26, 0x26, 0x20, 0x69, 0x73, + 0x41, 0x72, 0x72, 0x61, 0x79, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, + 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, + 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x62, + 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x7d, 0x20, 0x64, + 0x61, 0x74, 0x61, 0x3d, 0x24, 0x7b, 0x64, 0x61, 0x74, 0x61, 0x7d, 0x20, + 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x3d, 0x20, 0x69, 0x73, 0x41, 0x72, + 0x72, 0x61, 0x79, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3f, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, + 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, + 0x74, 0x65, 0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, + 0x27, 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, + 0x5e, 0x5c, 0x73, 0x2b, 0x2f, 0x2c, 0x20, 0x27, 0x27, 0x29, 0x20, 0x3a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x64, 0x61, 0x74, 0x61, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, + 0x20, 0x3d, 0x20, 0x69, 0x73, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, + 0x69, 0x6f, 0x6e, 0x4d, 0x6f, 0x64, 0x65, 0x20, 0x3f, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, + 0x78, 0x74, 0x20, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, + 0x7b, 0x4d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x69, 0x73, 0x68, + 0x7d, 0x20, 0x74, 0x65, 0x78, 0x74, 0x3d, 0x24, 0x7b, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x74, 0x65, 0x78, 0x74, 0x29, 0x7d, + 0x20, 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, + 0x66, 0x20, 0x28, 0x75, 0x73, 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x70, 0x20, + 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, + 0x3e, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x7b, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x75, 0x73, 0x65, 0x72, + 0x29, 0x7d, 0x3a, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, + 0x20, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x3c, + 0x2f, 0x70, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x20, 0x69, 0x73, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, + 0x69, 0x6f, 0x6e, 0x4d, 0x6f, 0x64, 0x65, 0x20, 0x3f, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, + 0x6d, 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, 0x6b, 0x65, 0x79, + 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x3e, 0x24, 0x7b, + 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, + 0x61, 0x6e, 0x3e, 0x60, 0x20, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, + 0x3c, 0x70, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, + 0x65, 0x78, 0x7d, 0x3e, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, + 0x65, 0x7d, 0x3c, 0x2f, 0x70, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6f, + 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x45, 0x64, 0x69, 0x74, + 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, + 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x3d, 0x20, 0x65, 0x2e, 0x74, 0x61, 0x72, + 0x67, 0x65, 0x74, 0x2e, 0x69, 0x6e, 0x6e, 0x65, 0x72, 0x54, 0x65, 0x78, + 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, + 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x20, + 0x3d, 0x20, 0x5b, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x69, + 0x64, 0x3d, 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x72, 0x65, 0x66, + 0x3d, 0x24, 0x7b, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, + 0x7d, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, + 0x61, 0x67, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x7d, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x69, 0x6d, 0x67, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x22, + 0x77, 0x69, 0x64, 0x74, 0x68, 0x3a, 0x20, 0x36, 0x30, 0x25, 0x3b, 0x24, + 0x7b, 0x21, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, + 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x20, 0x3f, 0x20, 0x60, 0x64, 0x69, + 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a, 0x20, 0x6e, 0x6f, 0x6e, 0x65, 0x3b, + 0x60, 0x20, 0x3a, 0x20, 0x60, 0x60, 0x7d, 0x22, 0x20, 0x73, 0x72, 0x63, + 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, + 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x7d, 0x22, 0x2f, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x73, 0x70, 0x61, 0x6e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, + 0x65, 0x64, 0x69, 0x74, 0x61, 0x62, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x69, + 0x73, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x4d, + 0x6f, 0x64, 0x65, 0x7d, 0x20, 0x72, 0x65, 0x66, 0x3d, 0x24, 0x7b, 0x63, + 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x7d, 0x20, 0x6f, 0x6e, + 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x68, 0x61, 0x6e, 0x64, + 0x6c, 0x65, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, + 0x45, 0x64, 0x69, 0x74, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6d, 0x65, 0x73, + 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x66, 0x6c, 0x61, 0x74, 0x4d, 0x61, + 0x70, 0x28, 0x63, 0x68, 0x61, 0x74, 0x4c, 0x69, 0x6e, 0x65, 0x29, 0x7d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x60, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, + 0x46, 0x6f, 0x72, 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, + 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, + 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, + 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, + 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, + 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, + 0x3a, 0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, + 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x28, + 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, + 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, + 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, - 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, - 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, - 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x65, 0x6c, 0x2e, 0x74, - 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, - 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, - 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, - 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x70, 0x61, - 0x72, 0x73, 0x65, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x28, 0x65, 0x6c, 0x2e, - 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x29, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, - 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, - 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, - 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x4d, 0x61, - 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x70, 0x61, 0x72, - 0x73, 0x65, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x28, 0x65, 0x6c, 0x2e, 0x74, - 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, - 0x29, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, - 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, - 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x73, 0x69, - 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x27, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, - 0x61, 0x74, 0x65, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, - 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, - 0x4f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, - 0x20, 0x3d, 0x3e, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, - 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, - 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x20, 0x3d, 0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, - 0x72, 0x74, 0x4a, 0x53, 0x4f, 0x4e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, - 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x29, - 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x3d, + 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, + 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, + 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, + 0x20, 0x70, 0x61, 0x72, 0x73, 0x65, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x28, + 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, + 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x20, 0x3d, + 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, + 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, + 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, + 0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, + 0x70, 0x61, 0x72, 0x73, 0x65, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x28, 0x65, + 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x29, 0x29, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x72, 0x61, 0x6d, + 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, + 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x3d, + 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x27, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x20, 0x3d, 0x20, 0x4a, 0x53, 0x4f, - 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x67, 0x72, 0x61, - 0x6d, 0x6d, 0x61, 0x72, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, - 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6e, 0x65, - 0x77, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x43, 0x6f, 0x6e, 0x76, - 0x65, 0x72, 0x74, 0x65, 0x72, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, + 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, + 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, + 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x28, + 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x73, 0x70, 0x6c, 0x69, 0x74, - 0x28, 0x27, 0x2c, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x64, - 0x75, 0x63, 0x65, 0x28, 0x28, 0x61, 0x63, 0x63, 0x2c, 0x20, 0x63, 0x75, - 0x72, 0x2c, 0x20, 0x69, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x7b, 0x20, - 0x2e, 0x2e, 0x2e, 0x61, 0x63, 0x63, 0x2c, 0x20, 0x5b, 0x63, 0x75, 0x72, - 0x2e, 0x74, 0x72, 0x69, 0x6d, 0x28, 0x29, 0x5d, 0x3a, 0x20, 0x69, 0x20, - 0x7d, 0x29, 0x2c, 0x20, 0x7b, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, - 0x74, 0x65, 0x72, 0x2e, 0x76, 0x69, 0x73, 0x69, 0x74, 0x28, 0x73, 0x63, - 0x68, 0x65, 0x6d, 0x61, 0x2c, 0x20, 0x27, 0x27, 0x29, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, - 0x72, 0x3a, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, - 0x2e, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x47, 0x72, 0x61, 0x6d, 0x6d, - 0x61, 0x72, 0x28, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, - 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x61, 0x6c, 0x65, 0x72, 0x74, 0x28, 0x60, 0x43, 0x6f, 0x6e, - 0x76, 0x65, 0x72, 0x74, 0x20, 0x66, 0x61, 0x69, 0x6c, 0x65, 0x64, 0x3a, - 0x20, 0x24, 0x7b, 0x65, 0x2e, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, - 0x7d, 0x60, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x46, - 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x20, 0x3d, 0x20, - 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2c, 0x20, 0x6d, 0x61, - 0x78, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, - 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x20, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, + 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, + 0x6e, 0x76, 0x65, 0x72, 0x74, 0x4a, 0x53, 0x4f, 0x4e, 0x53, 0x63, 0x68, + 0x65, 0x6d, 0x61, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x20, 0x3d, + 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, + 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, + 0x61, 0x20, 0x3d, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, + 0x73, 0x65, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x29, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, + 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x53, 0x63, 0x68, + 0x65, 0x6d, 0x61, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, + 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x5f, 0x6f, 0x72, 0x64, 0x65, + 0x72, 0x3a, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, + 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, + 0x4f, 0x72, 0x64, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x2e, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, 0x2c, 0x27, + 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x64, 0x75, 0x63, 0x65, 0x28, + 0x28, 0x61, 0x63, 0x63, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x2c, 0x20, 0x69, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x61, + 0x63, 0x63, 0x2c, 0x20, 0x5b, 0x63, 0x75, 0x72, 0x2e, 0x74, 0x72, 0x69, + 0x6d, 0x28, 0x29, 0x5d, 0x3a, 0x20, 0x69, 0x20, 0x7d, 0x29, 0x2c, 0x20, + 0x7b, 0x7d, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x6c, 0x6c, 0x6f, 0x77, 0x5f, 0x66, + 0x65, 0x74, 0x63, 0x68, 0x3a, 0x20, 0x74, 0x72, 0x75, 0x65, 0x2c, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, + 0x63, 0x68, 0x65, 0x6d, 0x61, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, + 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x2e, + 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x52, 0x65, 0x66, 0x73, 0x28, + 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2c, 0x20, 0x27, 0x69, 0x6e, 0x70, + 0x75, 0x74, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, + 0x2e, 0x76, 0x69, 0x73, 0x69, 0x74, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, + 0x61, 0x2c, 0x20, 0x27, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, + 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3a, 0x20, + 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x2e, 0x66, 0x6f, + 0x72, 0x6d, 0x61, 0x74, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x28, + 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, + 0x6c, 0x65, 0x72, 0x74, 0x28, 0x60, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, + 0x74, 0x20, 0x66, 0x61, 0x69, 0x6c, 0x65, 0x64, 0x3a, 0x20, 0x24, 0x7b, + 0x65, 0x2e, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x60, 0x29, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x46, 0x6c, 0x6f, 0x61, + 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x20, 0x3d, 0x20, 0x28, 0x7b, 0x20, + 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x2c, 0x20, + 0x6d, 0x69, 0x6e, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x73, + 0x74, 0x65, 0x70, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, + 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, + 0x6d, 0x65, 0x7d, 0x22, 0x3e, 0x24, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, + 0x7d, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, + 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, + 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x24, 0x7b, + 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, 0x22, + 0x24, 0x7b, 0x6d, 0x69, 0x6e, 0x7d, 0x22, 0x20, 0x6d, 0x61, 0x78, 0x3d, + 0x22, 0x24, 0x7b, 0x6d, 0x61, 0x78, 0x7d, 0x22, 0x20, 0x73, 0x74, 0x65, + 0x70, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x74, 0x65, 0x70, 0x7d, 0x22, 0x20, + 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, + 0x7d, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, + 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, + 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x7d, + 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x20, + 0x3d, 0x20, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2c, 0x20, + 0x6d, 0x61, 0x78, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x2c, 0x20, 0x6e, 0x61, + 0x6d, 0x65, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x29, + 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, + 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, + 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, + 0x65, 0x7d, 0x22, 0x3e, 0x24, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x7d, + 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, + 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, + 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x24, 0x7b, 0x6e, + 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, 0x22, 0x24, + 0x7b, 0x6d, 0x69, 0x6e, 0x7d, 0x22, 0x20, 0x6d, 0x61, 0x78, 0x3d, 0x22, + 0x24, 0x7b, 0x6d, 0x61, 0x78, 0x7d, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, + 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, + 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, + 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, + 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, + 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, + 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x2e, + 0x70, 0x72, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, 0x61, 0x75, + 0x6c, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, + 0x61, 0x75, 0x6c, 0x74, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, + 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x52, 0x65, 0x73, 0x65, 0x74, 0x42, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, + 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, + 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x27, 0x64, 0x65, + 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, + 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, + 0x64, 0x3e, 0x55, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, + 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, + 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, + 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x75, 0x73, + 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, + 0x73, 0x65, 0x74, 0x7d, 0x3e, 0x52, 0x65, 0x73, 0x65, 0x74, 0x20, 0x61, + 0x6c, 0x6c, 0x20, 0x74, 0x6f, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, + 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x28, 0x28, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, + 0x76, 0x65, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, + 0x6f, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x72, 0x79, 0x20, 0x63, 0x68, 0x61, + 0x6e, 0x67, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x28, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x73, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x43, + 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, + 0x3d, 0x3e, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x24, - 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x3e, 0x24, 0x7b, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x7d, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, + 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x47, 0x72, 0x61, + 0x6d, 0x6d, 0x61, 0x72, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, - 0x3d, 0x22, 0x72, 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, - 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x6d, 0x69, - 0x6e, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x69, 0x6e, 0x7d, 0x22, 0x20, 0x6d, - 0x61, 0x78, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x61, 0x78, 0x7d, 0x22, 0x20, - 0x73, 0x74, 0x65, 0x70, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x74, 0x65, 0x70, - 0x7d, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6e, - 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, - 0x22, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x22, 0x20, 0x6f, - 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, - 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46, 0x6c, 0x6f, - 0x61, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, - 0x3e, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3c, 0x2f, 0x73, - 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, - 0x6c, 0x64, 0x20, 0x3d, 0x20, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, - 0x6c, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x2c, - 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x20, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, - 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x24, 0x7b, - 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x3e, 0x24, 0x7b, 0x6c, 0x61, 0x62, - 0x65, 0x6c, 0x7d, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, - 0x22, 0x72, 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, - 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x6d, 0x69, 0x6e, - 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x69, 0x6e, 0x7d, 0x22, 0x20, 0x6d, 0x61, - 0x78, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x61, 0x78, 0x7d, 0x22, 0x20, 0x6e, - 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, - 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, - 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, - 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x52, 0x65, 0x73, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, - 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, - 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, - 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x41, 0x6e, 0x64, 0x41, 0x70, - 0x70, 0x6c, 0x79, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x42, 0x75, 0x74, 0x74, - 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, - 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, - 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x3d, 0x20, - 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, - 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x64, 0x69, 0x73, 0x61, - 0x62, 0x6c, 0x65, 0x64, 0x3e, 0x55, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x64, - 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, - 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, - 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, - 0x7b, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x7d, 0x3e, 0x52, 0x65, 0x73, 0x65, - 0x74, 0x20, 0x61, 0x6c, 0x6c, 0x20, 0x74, 0x6f, 0x20, 0x64, 0x65, 0x66, - 0x61, 0x75, 0x6c, 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, - 0x74, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x75, 0x74, - 0x6f, 0x73, 0x61, 0x76, 0x65, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x20, 0x6f, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x72, 0x79, 0x20, - 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x28, - 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, - 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x47, 0x72, 0x61, 0x6d, 0x6d, - 0x61, 0x72, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x20, 0x3d, 0x20, - 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, - 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, - 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3c, 0x2f, 0x6c, 0x61, 0x62, - 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, - 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, - 0x72, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x67, 0x72, 0x61, - 0x6d, 0x6d, 0x61, 0x72, 0x22, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, - 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x55, 0x73, 0x65, 0x20, 0x67, - 0x62, 0x6e, 0x66, 0x20, 0x6f, 0x72, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x20, - 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2b, 0x63, 0x6f, 0x6e, 0x76, 0x65, - 0x72, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, - 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x7d, 0x22, 0x20, - 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, - 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, - 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, - 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x70, 0x72, - 0x6f, 0x70, 0x2d, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x22, 0x20, 0x70, 0x6c, - 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x6f, - 0x72, 0x64, 0x65, 0x72, 0x3a, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x31, 0x2c, - 0x70, 0x72, 0x6f, 0x70, 0x32, 0x2c, 0x70, 0x72, 0x6f, 0x70, 0x33, 0x22, - 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, - 0x70, 0x64, 0x61, 0x74, 0x65, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, - 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, - 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x74, 0x79, 0x70, 0x65, - 0x3d, 0x22, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x22, 0x20, 0x6f, 0x6e, - 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x63, 0x6f, 0x6e, 0x76, - 0x65, 0x72, 0x74, 0x4a, 0x53, 0x4f, 0x4e, 0x53, 0x63, 0x68, 0x65, 0x6d, - 0x61, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x7d, 0x3e, 0x43, 0x6f, - 0x6e, 0x76, 0x65, 0x72, 0x74, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x20, 0x53, - 0x63, 0x68, 0x65, 0x6d, 0x61, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, - 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, - 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x46, 0x69, 0x65, 0x6c, 0x64, - 0x53, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, - 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, - 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x68, 0x74, - 0x6d, 0x6c, 0x46, 0x6f, 0x72, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x6d, 0x70, - 0x74, 0x22, 0x3e, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x3c, 0x2f, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, - 0x72, 0x65, 0x61, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, - 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x70, 0x72, - 0x6f, 0x6d, 0x70, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, - 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x7d, - 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, - 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, - 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, - 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x43, 0x68, 0x61, 0x74, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, - 0x46, 0x6f, 0x72, 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, - 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, - 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, - 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x53, - 0x65, 0x74, 0x28, 0x29, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, - 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, - 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, - 0x75, 0x73, 0x65, 0x72, 0x22, 0x3e, 0x55, 0x73, 0x65, 0x72, 0x20, 0x6e, - 0x61, 0x6d, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, - 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, - 0x65, 0x3d, 0x22, 0x75, 0x73, 0x65, 0x72, 0x22, 0x20, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, - 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x75, 0x73, 0x65, 0x72, - 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, - 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, - 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, - 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x62, - 0x6f, 0x74, 0x22, 0x3e, 0x42, 0x6f, 0x74, 0x20, 0x6e, 0x61, 0x6d, 0x65, - 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, - 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, - 0x63, 0x68, 0x61, 0x72, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, - 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x22, 0x20, - 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, - 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, - 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, - 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, - 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, - 0x3e, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, - 0x20, 0x69, 0x64, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x34, 0x20, - 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, - 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, - 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x22, 0x3e, 0x43, 0x68, 0x61, 0x74, 0x20, 0x68, 0x69, - 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, - 0x64, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, - 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x68, 0x69, 0x73, 0x74, 0x6f, - 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x68, - 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x31, 0x20, - 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, - 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, - 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x43, 0x6f, 0x6e, 0x74, 0x72, - 0x6f, 0x6c, 0x28, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, - 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x6f, - 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x43, 0x6f, 0x6e, 0x66, - 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, - 0x3d, 0x3e, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x6d, 0x70, - 0x74, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x46, 0x69, 0x65, 0x6c, - 0x64, 0x53, 0x65, 0x74, 0x28, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, - 0x73, 0x65, 0x74, 0x3e, 0x24, 0x7b, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, - 0x72, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x28, 0x29, 0x7d, 0x3c, - 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, - 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, - 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, - 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x42, 0x75, - 0x74, 0x74, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x63, - 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x73, 0x6c, 0x69, 0x6d, 0x22, 0x3e, - 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, - 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, - 0x3d, 0x22, 0x74, 0x79, 0x70, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x3d, 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x63, 0x68, 0x65, - 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, - 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, - 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, - 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, - 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, - 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x43, 0x68, 0x61, 0x74, 0x3c, 0x2f, - 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x73, - 0x6c, 0x69, 0x6d, 0x22, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, - 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, - 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x74, 0x79, 0x70, 0x65, 0x22, - 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x63, 0x6f, 0x6d, 0x70, - 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, - 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, - 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, - 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x22, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, - 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, - 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x43, - 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x3c, 0x2f, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, + 0x64, 0x3d, 0x22, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x22, 0x20, + 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, + 0x72, 0x22, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, 0x64, + 0x65, 0x72, 0x3d, 0x22, 0x55, 0x73, 0x65, 0x20, 0x67, 0x62, 0x6e, 0x66, + 0x20, 0x6f, 0x72, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x20, 0x53, 0x63, 0x68, + 0x65, 0x6d, 0x61, 0x2b, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x22, + 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x67, + 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, + 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, + 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, + 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x70, 0x2d, + 0x6f, 0x72, 0x64, 0x65, 0x72, 0x22, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, + 0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x6f, 0x72, 0x64, 0x65, + 0x72, 0x3a, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x31, 0x2c, 0x70, 0x72, 0x6f, + 0x70, 0x32, 0x2c, 0x70, 0x72, 0x6f, 0x70, 0x33, 0x22, 0x20, 0x6f, 0x6e, + 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, + 0x74, 0x65, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, + 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, + 0x72, 0x64, 0x65, 0x72, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, + 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x62, + 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x22, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, + 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, + 0x4a, 0x53, 0x4f, 0x4e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x47, 0x72, + 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x7d, 0x3e, 0x43, 0x6f, 0x6e, 0x76, 0x65, + 0x72, 0x74, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x20, 0x53, 0x63, 0x68, 0x65, + 0x6d, 0x61, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, + 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x6f, 0x6e, + 0x74, 0x72, 0x6f, 0x6c, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x53, 0x65, 0x74, + 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, + 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x46, + 0x6f, 0x72, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x22, 0x3e, + 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, + 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, + 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x6d, 0x70, + 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, - 0x63, 0x68, 0x61, 0x74, 0x27, 0x20, 0x3f, 0x20, 0x43, 0x68, 0x61, 0x74, - 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x28, 0x29, - 0x20, 0x3a, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, - 0x6e, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x28, + 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x22, 0x20, 0x6f, + 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, + 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, + 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, + 0x68, 0x61, 0x74, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, + 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, + 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x6f, 0x6e, 0x74, + 0x72, 0x6f, 0x6c, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x53, 0x65, 0x74, 0x28, 0x29, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, - 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x72, - 0x65, 0x64, 0x69, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x22, 0x2c, 0x20, - 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x30, 0x34, 0x38, 0x2c, 0x20, 0x6d, - 0x69, 0x6e, 0x3a, 0x20, 0x2d, 0x31, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, - 0x3a, 0x20, 0x22, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, - 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, - 0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x20, 0x7d, 0x29, 0x7d, + 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, + 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x75, 0x73, 0x65, + 0x72, 0x22, 0x3e, 0x55, 0x73, 0x65, 0x72, 0x20, 0x6e, 0x61, 0x6d, 0x65, + 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, + 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, + 0x75, 0x73, 0x65, 0x72, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, + 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x22, 0x20, + 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, + 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, + 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x62, 0x6f, 0x74, 0x22, + 0x3e, 0x42, 0x6f, 0x74, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3c, 0x2f, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, + 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, + 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x63, 0x68, 0x61, + 0x72, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, + 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, + 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, + 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, - 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, - 0x54, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x22, - 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x2e, 0x30, 0x2c, 0x20, - 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, - 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, - 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, - 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, - 0x72, 0x65, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, - 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x65, 0x6e, 0x61, 0x6c, 0x69, - 0x7a, 0x65, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x20, 0x73, 0x65, - 0x71, 0x75, 0x65, 0x6e, 0x63, 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, - 0x3a, 0x20, 0x32, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, - 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, - 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, - 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, - 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, - 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, - 0x6c, 0x74, 0x79, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, - 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, - 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x43, 0x6f, 0x6e, 0x73, 0x69, 0x64, 0x65, - 0x72, 0x20, 0x4e, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x20, 0x66, - 0x6f, 0x72, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x69, 0x7a, 0x65, 0x22, - 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x30, 0x34, 0x38, 0x2c, - 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, - 0x65, 0x3a, 0x20, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c, - 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c, - 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, + 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, + 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x50, 0x72, + 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, + 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, + 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, + 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x7d, + 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, + 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, + 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, + 0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x22, 0x3e, 0x43, 0x68, 0x61, 0x74, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, + 0x72, 0x79, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, + 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, + 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x6e, 0x61, + 0x6d, 0x65, 0x3d, 0x22, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, + 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74, + 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x7d, + 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x31, 0x20, 0x6f, 0x6e, 0x69, + 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, + 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x47, 0x72, 0x61, + 0x6d, 0x6d, 0x61, 0x72, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x28, + 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, + 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, + 0x6f, 0x72, 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, + 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, + 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x6f, + 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x53, 0x65, + 0x74, 0x28, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, + 0x3e, 0x24, 0x7b, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x43, 0x6f, + 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x28, 0x29, 0x7d, 0x3c, 0x2f, 0x66, 0x69, + 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, + 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x72, 0x6d, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, + 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x24, 0x7b, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x42, 0x75, 0x74, 0x74, 0x6f, + 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x63, 0x6c, 0x61, 0x73, + 0x73, 0x3d, 0x22, 0x73, 0x6c, 0x69, 0x6d, 0x22, 0x3e, 0x3c, 0x69, 0x6e, + 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, + 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x74, + 0x79, 0x70, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, + 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, + 0x64, 0x3d, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, + 0x3d, 0x3d, 0x20, 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x7d, 0x20, 0x6f, + 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, + 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, + 0x2f, 0x3e, 0x20, 0x43, 0x68, 0x61, 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, + 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x73, 0x6c, 0x69, 0x6d, + 0x22, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, + 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, + 0x6d, 0x65, 0x3d, 0x22, 0x74, 0x79, 0x70, 0x65, 0x22, 0x20, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, + 0x69, 0x6f, 0x6e, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, + 0x3d, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, + 0x3d, 0x20, 0x22, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, + 0x6e, 0x22, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, + 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x43, 0x6f, 0x6d, 0x70, + 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, + 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, + 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x73, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, + 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x63, 0x68, 0x61, + 0x74, 0x27, 0x20, 0x3f, 0x20, 0x43, 0x68, 0x61, 0x74, 0x43, 0x6f, 0x6e, + 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x28, 0x29, 0x20, 0x3a, 0x20, + 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x43, 0x6f, + 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x28, 0x29, 0x7d, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, + 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x6f, 0x70, 0x2d, 0x4b, - 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x69, 0x6e, 0x67, 0x22, 0x2c, 0x20, - 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, 0x30, 0x2c, 0x20, 0x6d, 0x69, - 0x6e, 0x3a, 0x20, 0x2d, 0x31, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, - 0x20, 0x22, 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x22, 0x2c, 0x20, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x20, + 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x72, 0x65, 0x64, 0x69, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, + 0x3a, 0x20, 0x32, 0x30, 0x34, 0x38, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, + 0x20, 0x2d, 0x31, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, + 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x22, 0x2c, 0x20, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, + 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, + 0x65, 0x64, 0x69, 0x63, 0x74, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, + 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, + 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x65, 0x6d, + 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c, 0x20, 0x6d, + 0x61, 0x78, 0x3a, 0x20, 0x32, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, + 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, + 0x20, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, + 0x65, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, + 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, + 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, - 0x3a, 0x20, 0x22, 0x54, 0x6f, 0x70, 0x2d, 0x50, 0x20, 0x73, 0x61, 0x6d, - 0x70, 0x6c, 0x69, 0x6e, 0x67, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, - 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, - 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, - 0x6f, 0x70, 0x5f, 0x70, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, - 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x3a, 0x20, 0x22, 0x50, 0x65, 0x6e, 0x61, 0x6c, 0x69, 0x7a, 0x65, 0x20, + 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x20, 0x73, 0x65, 0x71, 0x75, 0x65, + 0x6e, 0x63, 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, + 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, + 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x72, 0x65, 0x70, + 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, + 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, + 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, + 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, + 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, + 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, + 0x20, 0x22, 0x43, 0x6f, 0x6e, 0x73, 0x69, 0x64, 0x65, 0x72, 0x20, 0x4e, + 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x20, 0x66, 0x6f, 0x72, 0x20, + 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x69, 0x7a, 0x65, 0x22, 0x2c, 0x20, 0x6d, + 0x61, 0x78, 0x3a, 0x20, 0x32, 0x30, 0x34, 0x38, 0x2c, 0x20, 0x6d, 0x69, + 0x6e, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, + 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, 0x74, + 0x5f, 0x6e, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, 0x74, + 0x5f, 0x6e, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, + 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, + 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x6f, 0x70, 0x2d, 0x4b, 0x20, 0x73, 0x61, + 0x6d, 0x70, 0x6c, 0x69, 0x6e, 0x67, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, + 0x3a, 0x20, 0x31, 0x30, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, + 0x2d, 0x31, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, + 0x6f, 0x70, 0x5f, 0x6b, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x20, 0x7d, 0x29, 0x7d, + 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, - 0x4d, 0x69, 0x6e, 0x2d, 0x50, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x69, + 0x54, 0x6f, 0x70, 0x2d, 0x50, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x69, 0x6e, 0x67, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, - 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6d, 0x69, 0x6e, 0x5f, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, - 0x6d, 0x69, 0x6e, 0x5f, 0x70, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, - 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x65, 0x74, 0x61, 0x69, - 0x6c, 0x73, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, - 0x3e, 0x4d, 0x6f, 0x72, 0x65, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, - 0x73, 0x3c, 0x2f, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3e, 0x0a, + 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, + 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, + 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x6e, + 0x2d, 0x50, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x69, 0x6e, 0x67, 0x22, + 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, + 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, + 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6d, 0x69, 0x6e, 0x5f, 0x70, 0x22, 0x2c, + 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, + 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x6e, + 0x5f, 0x70, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, + 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x73, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3e, 0x4d, 0x6f, + 0x72, 0x65, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x3c, 0x2f, + 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, + 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, + 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, + 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, + 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x46, 0x53, + 0x2d, 0x5a, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, + 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x66, 0x73, 0x5f, + 0x7a, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, + 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, + 0x74, 0x66, 0x73, 0x5f, 0x7a, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, - 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, + 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, + 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, + 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x20, 0x50, 0x22, 0x2c, 0x20, 0x6d, + 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, + 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, + 0x20, 0x22, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x22, + 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, + 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, + 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x20, 0x7d, 0x29, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, + 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, + 0x22, 0x50, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x20, 0x70, 0x65, + 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, + 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, + 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x70, + 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, + 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, + 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, + 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, + 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, - 0x54, 0x46, 0x53, 0x2d, 0x5a, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, + 0x46, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x20, 0x70, 0x65, + 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, - 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, - 0x66, 0x73, 0x5f, 0x7a, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, + 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x66, + 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, + 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2e, 0x74, 0x66, 0x73, 0x5f, 0x7a, 0x20, 0x7d, 0x29, 0x7d, + 0x75, 0x65, 0x2e, 0x66, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, + 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, - 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, - 0x20, 0x22, 0x54, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x20, 0x50, 0x22, - 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, - 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, - 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, - 0x5f, 0x70, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, - 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, + 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x68, 0x72, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, + 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, + 0x22, 0x74, 0x68, 0x72, 0x65, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, + 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, + 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, + 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, + 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x30, + 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x20, 0x7d, - 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, - 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, - 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, - 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, - 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, - 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, - 0x20, 0x22, 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, - 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, - 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, - 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x20, 0x7d, 0x29, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, - 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, - 0x3a, 0x20, 0x22, 0x46, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, - 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, - 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, - 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, - 0x20, 0x22, 0x66, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, - 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, - 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x66, 0x72, 0x65, 0x71, 0x75, 0x65, - 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x20, + 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, + 0x20, 0x30, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, + 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x6e, 0x6f, + 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x3c, 0x2f, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, + 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, + 0x74, 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, + 0x31, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, + 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, + 0x3d, 0x20, 0x31, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, + 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x4d, + 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x76, 0x31, 0x3c, 0x2f, + 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, + 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, + 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, + 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, + 0x22, 0x32, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, + 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, + 0x3d, 0x3d, 0x20, 0x32, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, + 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, + 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x76, 0x32, 0x3c, + 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, + 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, + 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, + 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, + 0x61, 0x74, 0x20, 0x74, 0x61, 0x75, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, + 0x3a, 0x20, 0x31, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, + 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, + 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, + 0x75, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, + 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, + 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, + 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, + 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, + 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, + 0x61, 0x74, 0x20, 0x65, 0x74, 0x61, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, + 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, + 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, + 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, + 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, + 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, + 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x72, 0x20, 0x2f, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, - 0x73, 0x73, 0x3d, 0x22, 0x74, 0x68, 0x72, 0x65, 0x65, 0x22, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, - 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, - 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, - 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x3d, 0x22, 0x30, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, - 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, - 0x20, 0x3d, 0x3d, 0x20, 0x30, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, - 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, - 0x20, 0x6e, 0x6f, 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, - 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, - 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, - 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, - 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x3d, 0x22, 0x31, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, - 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, - 0x74, 0x20, 0x3d, 0x3d, 0x20, 0x31, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, - 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, - 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, - 0x3e, 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x76, - 0x31, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, - 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, - 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, - 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x3d, 0x22, 0x32, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, - 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, - 0x61, 0x74, 0x20, 0x3d, 0x3d, 0x20, 0x32, 0x7d, 0x20, 0x6f, 0x6e, 0x69, - 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, - 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, - 0x2f, 0x3e, 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, - 0x76, 0x32, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, - 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, - 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x72, - 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x74, 0x61, 0x75, 0x22, 0x2c, 0x20, - 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6d, - 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, - 0x65, 0x3a, 0x20, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, - 0x5f, 0x74, 0x61, 0x75, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, - 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, - 0x74, 0x61, 0x75, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, - 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, - 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x72, - 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x65, 0x74, 0x61, 0x22, 0x2c, 0x20, - 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, - 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, - 0x3a, 0x20, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, - 0x65, 0x74, 0x61, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, - 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, - 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, - 0x74, 0x61, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, - 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, - 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, - 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, - 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x53, 0x68, 0x6f, 0x77, 0x20, 0x50, 0x72, - 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x22, - 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, 0x2c, 0x20, 0x6d, - 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, - 0x20, 0x22, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x22, 0x2c, 0x20, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, - 0x6f, 0x62, 0x73, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, - 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, - 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x61, 0x70, 0x69, - 0x5f, 0x6b, 0x65, 0x79, 0x22, 0x3e, 0x41, 0x50, 0x49, 0x20, 0x4b, 0x65, - 0x79, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, - 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, - 0x22, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, 0x79, 0x22, 0x20, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x61, 0x70, 0x69, 0x5f, - 0x6b, 0x65, 0x79, 0x7d, 0x22, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, - 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x45, 0x6e, 0x74, 0x65, 0x72, - 0x20, 0x41, 0x50, 0x49, 0x20, 0x6b, 0x65, 0x79, 0x22, 0x20, 0x6f, 0x6e, - 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, - 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x7d, 0x20, 0x2f, 0x3e, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, + 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, + 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, + 0x20, 0x22, 0x53, 0x68, 0x6f, 0x77, 0x20, 0x50, 0x72, 0x6f, 0x62, 0x61, + 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x22, 0x2c, 0x20, 0x6d, + 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, + 0x20, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6e, + 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73, + 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, + 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, + 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, + 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, + 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x6e, 0x20, 0x50, 0x72, 0x6f, 0x62, 0x61, + 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x66, 0x72, 0x6f, + 0x6d, 0x20, 0x65, 0x61, 0x63, 0x68, 0x20, 0x53, 0x61, 0x6d, 0x70, 0x6c, + 0x65, 0x72, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, + 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x6e, 0x61, + 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6d, 0x69, 0x6e, 0x5f, 0x6b, 0x65, 0x65, + 0x70, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, + 0x6d, 0x69, 0x6e, 0x5f, 0x6b, 0x65, 0x65, 0x70, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x2f, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x73, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x72, 0x6d, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x43, 0x6f, 0x6c, 0x6f, 0x72, - 0x20, 0x3d, 0x20, 0x28, 0x70, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x72, 0x20, 0x3d, 0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, - 0x6f, 0x72, 0x28, 0x31, 0x39, 0x32, 0x20, 0x2a, 0x20, 0x28, 0x31, 0x20, - 0x2d, 0x20, 0x70, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x20, 0x3d, 0x20, 0x4d, - 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x31, 0x39, - 0x32, 0x20, 0x2a, 0x20, 0x70, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x60, 0x72, 0x67, - 0x62, 0x61, 0x28, 0x24, 0x7b, 0x72, 0x7d, 0x2c, 0x24, 0x7b, 0x67, 0x7d, - 0x2c, 0x30, 0x2c, 0x30, 0x2e, 0x33, 0x29, 0x60, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, - 0x74, 0x69, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x70, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, 0x61, - 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, - 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, - 0x69, 0x65, 0x73, 0x20, 0x7d, 0x20, 0x3d, 0x20, 0x6d, 0x73, 0x67, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, - 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x21, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, - 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, - 0x73, 0x20, 0x7c, 0x7c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, - 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, - 0x69, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3d, - 0x3d, 0x3d, 0x20, 0x30, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6d, 0x73, - 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, - 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, - 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x2e, - 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, 0x20, 0x31, 0x29, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x2f, 0x2f, 0x20, 0x4e, 0x6f, 0x74, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x62, - 0x79, 0x74, 0x65, 0x20, 0x70, 0x61, 0x69, 0x72, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, - 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, - 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x5b, - 0x30, 0x5d, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2e, 0x73, - 0x74, 0x61, 0x72, 0x74, 0x73, 0x57, 0x69, 0x74, 0x68, 0x28, 0x27, 0x62, - 0x79, 0x74, 0x65, 0x3a, 0x20, 0x5c, 0x5c, 0x27, 0x29, 0x29, 0x20, 0x72, - 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, - 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, - 0x70, 0x6c, 0x69, 0x74, 0x44, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x63, - 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, - 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x2e, - 0x6d, 0x61, 0x70, 0x28, 0x70, 0x72, 0x6f, 0x62, 0x20, 0x3d, 0x3e, 0x20, - 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3a, 0x20, - 0x70, 0x72, 0x6f, 0x62, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, - 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, - 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, - 0x65, 0x73, 0x3a, 0x20, 0x5b, 0x70, 0x72, 0x6f, 0x62, 0x5d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, - 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, - 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, - 0x69, 0x65, 0x73, 0x7d, 0x20, 0x64, 0x61, 0x74, 0x61, 0x3d, 0x24, 0x7b, - 0x73, 0x70, 0x6c, 0x69, 0x74, 0x44, 0x61, 0x74, 0x61, 0x7d, 0x20, 0x2f, - 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x2c, - 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x7d, 0x20, 0x3d, - 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, - 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, - 0x73, 0x5b, 0x30, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, - 0x20, 0x3d, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x2e, 0x66, 0x69, 0x6e, - 0x64, 0x28, 0x70, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x2e, 0x74, 0x6f, 0x6b, - 0x5f, 0x73, 0x74, 0x72, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x6d, 0x73, 0x67, - 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x70, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x20, 0x3d, 0x20, 0x66, 0x6f, 0x75, - 0x6e, 0x64, 0x20, 0x3f, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x43, 0x6f, 0x6c, - 0x6f, 0x72, 0x28, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2e, 0x70, 0x72, 0x6f, - 0x62, 0x29, 0x20, 0x3a, 0x20, 0x27, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x70, - 0x61, 0x72, 0x65, 0x6e, 0x74, 0x27, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x6f, - 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, - 0x6e, 0x20, 0x3d, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, - 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x62, - 0x2d, 0x73, 0x65, 0x74, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x70, 0x72, 0x6f, - 0x62, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x28, 0x70, 0x2c, 0x20, 0x69, - 0x6e, 0x64, 0x65, 0x78, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, - 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x69, 0x74, 0x6c, - 0x65, 0x3d, 0x24, 0x7b, 0x60, 0x70, 0x72, 0x6f, 0x62, 0x3a, 0x20, 0x24, - 0x7b, 0x70, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x7d, 0x60, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, - 0x7b, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, - 0x3a, 0x20, 0x27, 0x30, 0x2e, 0x33, 0x65, 0x6d, 0x27, 0x2c, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x43, - 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x70, 0x2e, 0x74, 0x6f, 0x6b, 0x5f, - 0x73, 0x74, 0x72, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x74, - 0x65, 0x6e, 0x74, 0x20, 0x3f, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x43, 0x6f, - 0x6c, 0x6f, 0x72, 0x28, 0x70, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x29, 0x20, - 0x3a, 0x20, 0x27, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x70, 0x61, 0x72, 0x65, - 0x6e, 0x74, 0x27, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, - 0x3e, 0x24, 0x7b, 0x70, 0x2e, 0x74, 0x6f, 0x6b, 0x5f, 0x73, 0x74, 0x72, - 0x7d, 0x3a, 0x20, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, + 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, - 0x7b, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, - 0x70, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x20, 0x2a, 0x20, 0x31, 0x30, 0x30, - 0x29, 0x7d, 0x25, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, + 0x3d, 0x22, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, 0x79, 0x22, 0x3e, 0x41, + 0x50, 0x49, 0x20, 0x4b, 0x65, 0x79, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, + 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, + 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, + 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, + 0x79, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, 0x79, 0x7d, 0x22, 0x20, 0x70, + 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, + 0x45, 0x6e, 0x74, 0x65, 0x72, 0x20, 0x41, 0x50, 0x49, 0x20, 0x6b, 0x65, + 0x79, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, + 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, + 0x73, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, + 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, + 0x73, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x2f, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x62, + 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x29, 0x20, + 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x20, 0x3d, 0x20, 0x4d, 0x61, 0x74, + 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x31, 0x39, 0x32, 0x20, + 0x2a, 0x20, 0x28, 0x31, 0x20, 0x2d, 0x20, 0x70, 0x29, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x67, 0x20, 0x3d, 0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, + 0x6f, 0x72, 0x28, 0x31, 0x39, 0x32, 0x20, 0x2a, 0x20, 0x70, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, + 0x6e, 0x20, 0x60, 0x72, 0x67, 0x62, 0x61, 0x28, 0x24, 0x7b, 0x72, 0x7d, + 0x2c, 0x24, 0x7b, 0x67, 0x7d, 0x2c, 0x30, 0x2c, 0x30, 0x2e, 0x33, 0x29, + 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x72, 0x6f, 0x62, + 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x3d, 0x20, + 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x64, 0x61, + 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6d, 0x70, + 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, + 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x7d, 0x20, 0x3d, + 0x20, 0x6d, 0x73, 0x67, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x21, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, + 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, + 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x7c, 0x7c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, + 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, + 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, + 0x67, 0x74, 0x68, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x30, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, + 0x6e, 0x74, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, + 0x74, 0x69, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, + 0x3e, 0x20, 0x31, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x4e, 0x6f, 0x74, 0x20, + 0x66, 0x6f, 0x72, 0x20, 0x62, 0x79, 0x74, 0x65, 0x20, 0x70, 0x61, 0x69, + 0x72, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, + 0x74, 0x69, 0x65, 0x73, 0x5b, 0x30, 0x5d, 0x2e, 0x63, 0x6f, 0x6e, 0x74, + 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x74, 0x61, 0x72, 0x74, 0x73, 0x57, 0x69, + 0x74, 0x68, 0x28, 0x27, 0x62, 0x79, 0x74, 0x65, 0x3a, 0x20, 0x5c, 0x5c, + 0x27, 0x29, 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6d, + 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x44, 0x61, 0x74, + 0x61, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, + 0x74, 0x69, 0x65, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x70, 0x72, 0x6f, + 0x62, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, + 0x65, 0x6e, 0x74, 0x3a, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x2e, 0x63, 0x6f, + 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, + 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, + 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x3a, 0x20, 0x5b, 0x70, 0x72, + 0x6f, 0x62, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x29, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, + 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x62, 0x61, + 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x7d, 0x20, 0x64, 0x61, + 0x74, 0x61, 0x3d, 0x24, 0x7b, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x44, 0x61, + 0x74, 0x61, 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x70, + 0x72, 0x6f, 0x62, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, + 0x74, 0x20, 0x7d, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, + 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, + 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x5b, 0x30, 0x5d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x20, 0x3d, 0x20, 0x70, 0x72, 0x6f, 0x62, + 0x73, 0x2e, 0x66, 0x69, 0x6e, 0x64, 0x28, 0x70, 0x20, 0x3d, 0x3e, 0x20, + 0x70, 0x2e, 0x74, 0x6f, 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x20, 0x3d, 0x3d, + 0x3d, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, + 0x74, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x20, + 0x3d, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x20, 0x3f, 0x20, 0x70, 0x72, + 0x6f, 0x62, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x28, 0x66, 0x6f, 0x75, 0x6e, + 0x64, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x29, 0x20, 0x3a, 0x20, 0x27, 0x74, + 0x72, 0x61, 0x6e, 0x73, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x27, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, + 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x3d, 0x20, 0x68, 0x74, 0x6d, + 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, + 0x22, 0x70, 0x72, 0x6f, 0x62, 0x2d, 0x73, 0x65, 0x74, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x60, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, + 0x28, 0x70, 0x2c, 0x20, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x7d, - 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x7b, 0x20, 0x62, - 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x43, 0x6f, 0x6c, - 0x6f, 0x72, 0x3a, 0x20, 0x70, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x20, 0x7d, - 0x7d, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, - 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x3d, 0x24, 0x7b, 0x70, 0x6f, 0x70, 0x6f, - 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x7d, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x24, 0x7b, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, - 0x65, 0x6e, 0x74, 0x2e, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x28, 0x2f, 0x5c, - 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x29, 0x20, 0x3f, 0x20, 0x68, 0x74, 0x6d, - 0x6c, 0x60, 0x3c, 0x62, 0x72, 0x20, 0x2f, 0x3e, 0x60, 0x20, 0x3a, 0x20, - 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, - 0x20, 0x70, 0x6f, 0x6f, 0x72, 0x20, 0x6d, 0x61, 0x6e, 0x73, 0x20, 0x6d, - 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x20, 0x72, 0x65, 0x70, 0x6c, - 0x61, 0x63, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x4d, 0x61, 0x72, 0x6b, 0x64, 0x6f, - 0x77, 0x6e, 0x69, 0x73, 0x68, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x61, 0x72, - 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x64, - 0x20, 0x3d, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x74, 0x65, - 0x78, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, - 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x26, 0x2f, 0x67, - 0x2c, 0x20, 0x27, 0x26, 0x61, 0x6d, 0x70, 0x3b, 0x27, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, - 0x61, 0x63, 0x65, 0x28, 0x2f, 0x3c, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, - 0x6c, 0x74, 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, - 0x3e, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x67, 0x74, 0x3b, 0x27, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, - 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x23, 0x7b, 0x31, 0x2c, - 0x36, 0x7d, 0x20, 0x28, 0x2e, 0x2a, 0x29, 0x24, 0x2f, 0x67, 0x69, 0x6d, - 0x2c, 0x20, 0x27, 0x3c, 0x68, 0x33, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x68, - 0x33, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5c, - 0x2a, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5c, 0x2a, 0x5c, 0x2a, - 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, - 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, - 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, - 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5f, 0x5f, 0x28, - 0x2e, 0x2a, 0x3f, 0x29, 0x5f, 0x5f, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, - 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x73, - 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, - 0x65, 0x28, 0x2f, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5c, 0x2a, - 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31, 0x3c, - 0x2f, 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, - 0x2f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, 0x2f, 0x67, 0x2c, 0x20, - 0x27, 0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x65, 0x6d, 0x3e, - 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, - 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x60, 0x60, 0x60, - 0x2e, 0x2a, 0x3f, 0x5c, 0x6e, 0x28, 0x5b, 0x5c, 0x73, 0x5c, 0x53, 0x5d, - 0x2a, 0x3f, 0x29, 0x60, 0x60, 0x60, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, - 0x70, 0x72, 0x65, 0x3e, 0x3c, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x24, 0x31, - 0x3c, 0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x3c, 0x2f, 0x70, 0x72, 0x65, - 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x60, 0x28, - 0x2e, 0x2a, 0x3f, 0x29, 0x60, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x63, - 0x6f, 0x64, 0x65, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x63, 0x6f, 0x64, 0x65, - 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x6e, - 0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c, 0x62, 0x72, 0x20, 0x2f, - 0x3e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, - 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, - 0x73, 0x70, 0x61, 0x6e, 0x20, 0x64, 0x61, 0x6e, 0x67, 0x65, 0x72, 0x6f, - 0x75, 0x73, 0x6c, 0x79, 0x53, 0x65, 0x74, 0x49, 0x6e, 0x6e, 0x65, 0x72, - 0x48, 0x54, 0x4d, 0x4c, 0x3d, 0x24, 0x7b, 0x7b, 0x20, 0x5f, 0x5f, 0x68, - 0x74, 0x6d, 0x6c, 0x3a, 0x20, 0x6d, 0x64, 0x20, 0x7d, 0x7d, 0x20, 0x2f, - 0x3e, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x4d, 0x6f, - 0x64, 0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, - 0x6e, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x61, 0x72, - 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x6c, 0x6c, 0x61, - 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, - 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, - 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, - 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x6b, - 0x65, 0x6e, 0x73, 0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x65, - 0x64, 0x7d, 0x20, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x65, 0x64, - 0x2c, 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, - 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x6b, - 0x65, 0x6e, 0x73, 0x5f, 0x63, 0x61, 0x63, 0x68, 0x65, 0x64, 0x7d, 0x20, - 0x63, 0x61, 0x63, 0x68, 0x65, 0x64, 0x2c, 0x20, 0x24, 0x7b, 0x6c, 0x6c, - 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x2e, 0x70, - 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x70, 0x65, 0x72, - 0x5f, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x5f, 0x6d, 0x73, 0x2e, 0x74, 0x6f, - 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x29, 0x7d, 0x6d, 0x73, 0x20, 0x70, - 0x65, 0x72, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x2c, 0x20, 0x24, 0x7b, - 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, - 0x2e, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x70, - 0x65, 0x72, 0x5f, 0x73, 0x65, 0x63, 0x6f, 0x6e, 0x64, 0x2e, 0x74, 0x6f, - 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x32, 0x29, 0x7d, 0x20, 0x74, 0x6f, - 0x6b, 0x65, 0x6e, 0x73, 0x20, 0x70, 0x65, 0x72, 0x20, 0x73, 0x65, 0x63, - 0x6f, 0x6e, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69, 0x6d, 0x70, 0x6c, 0x65, - 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x69, 0x6d, 0x70, - 0x6c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x70, - 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, - 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, - 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x29, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x3d, - 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, - 0x20, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x27, 0x30, 0x70, 0x78, 0x27, 0x2c, - 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, 0x27, 0x30, 0x70, 0x78, 0x27, - 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, - 0x65, 0x66, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x28, - 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, - 0x65, 0x72, 0x52, 0x65, 0x66, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x52, - 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x3b, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, - 0x6f, 0x67, 0x67, 0x6c, 0x65, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, - 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x62, - 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, - 0x72, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x72, 0x65, 0x63, 0x74, 0x20, 0x3d, 0x20, 0x62, 0x75, 0x74, 0x74, 0x6f, - 0x6e, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, - 0x2e, 0x67, 0x65, 0x74, 0x42, 0x6f, 0x75, 0x6e, 0x64, 0x69, 0x6e, 0x67, - 0x43, 0x6c, 0x69, 0x65, 0x6e, 0x74, 0x52, 0x65, 0x63, 0x74, 0x28, 0x29, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x3a, 0x20, - 0x60, 0x24, 0x7b, 0x72, 0x65, 0x63, 0x74, 0x2e, 0x62, 0x6f, 0x74, 0x74, - 0x6f, 0x6d, 0x20, 0x2b, 0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x2e, - 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x59, 0x7d, 0x70, 0x78, 0x60, 0x2c, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, 0x60, 0x24, 0x7b, 0x72, 0x65, - 0x63, 0x74, 0x2e, 0x6c, 0x65, 0x66, 0x74, 0x20, 0x2b, 0x20, 0x77, 0x69, - 0x6e, 0x64, 0x6f, 0x77, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x58, - 0x7d, 0x70, 0x78, 0x60, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x20, 0x3d, 0x20, 0x21, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, - 0x6c, 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, 0x64, 0x65, 0x20, - 0x3d, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d, 0x3e, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, - 0x66, 0x20, 0x28, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, - 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x20, 0x26, 0x26, - 0x20, 0x21, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, - 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x63, 0x6f, 0x6e, - 0x74, 0x61, 0x69, 0x6e, 0x73, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, - 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x29, 0x20, 0x26, 0x26, 0x20, 0x21, - 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, - 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, - 0x6e, 0x73, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x74, 0x61, 0x72, - 0x67, 0x65, 0x74, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x66, 0x61, 0x6c, - 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, - 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, - 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x61, 0x64, 0x64, 0x45, 0x76, 0x65, 0x6e, - 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28, 0x27, 0x6d, - 0x6f, 0x75, 0x73, 0x65, 0x64, 0x6f, 0x77, 0x6e, 0x27, 0x2c, 0x20, 0x68, - 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x4f, 0x75, - 0x74, 0x73, 0x69, 0x64, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x28, - 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, - 0x74, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x76, 0x65, 0x45, 0x76, 0x65, 0x6e, - 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28, 0x27, 0x6d, - 0x6f, 0x75, 0x73, 0x65, 0x64, 0x6f, 0x77, 0x6e, 0x27, 0x2c, 0x20, 0x68, - 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x4f, 0x75, - 0x74, 0x73, 0x69, 0x64, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x5d, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, - 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, - 0x3d, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x73, 0x74, 0x79, - 0x6c, 0x65, 0x7d, 0x20, 0x72, 0x65, 0x66, 0x3d, 0x24, 0x7b, 0x62, 0x75, - 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x7d, 0x20, 0x6f, 0x6e, 0x43, - 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x74, 0x6f, 0x67, 0x67, 0x6c, - 0x65, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x7d, 0x3e, 0x24, 0x7b, - 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, - 0x65, 0x6e, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x69, 0x73, 0x4f, - 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x26, 0x26, - 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, - 0x61, 0x6c, 0x7d, 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x3d, 0x22, 0x23, 0x70, - 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x72, 0x65, 0x66, 0x3d, 0x24, 0x7b, 0x70, 0x6f, 0x70, - 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, - 0x72, 0x2d, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x22, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, - 0x6f, 0x70, 0x3a, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x2c, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, + 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x74, 0x69, 0x74, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x60, 0x70, 0x72, + 0x6f, 0x62, 0x3a, 0x20, 0x24, 0x7b, 0x70, 0x2e, 0x70, 0x72, 0x6f, 0x62, + 0x7d, 0x60, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, + 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, + 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x27, 0x30, 0x2e, 0x33, 0x65, + 0x6d, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, + 0x6f, 0x75, 0x6e, 0x64, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x70, + 0x2e, 0x74, 0x6f, 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3f, 0x20, 0x70, + 0x72, 0x6f, 0x62, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x28, 0x70, 0x2e, 0x70, + 0x72, 0x6f, 0x62, 0x29, 0x20, 0x3a, 0x20, 0x27, 0x74, 0x72, 0x61, 0x6e, + 0x73, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x27, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, - 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6c, 0x65, 0x66, - 0x74, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x70, 0x72, - 0x6f, 0x70, 0x73, 0x2e, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, - 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, - 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x2f, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, - 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, - 0x2f, 0x20, 0x53, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x3a, 0x20, 0x70, 0x72, - 0x65, 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x20, - 0x28, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x69, 0x74, - 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x64, 0x65, 0x76, 0x65, - 0x6c, 0x6f, 0x70, 0x69, 0x74, 0x2f, 0x70, 0x72, 0x65, 0x61, 0x63, 0x74, - 0x2d, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x2f, 0x62, 0x6c, 0x6f, 0x62, - 0x2f, 0x6d, 0x61, 0x73, 0x74, 0x65, 0x72, 0x2f, 0x73, 0x72, 0x63, 0x2f, - 0x70, 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, 0x72, 0x74, 0x61, - 0x6c, 0x2e, 0x6a, 0x73, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2a, - 0x2a, 0x20, 0x52, 0x65, 0x64, 0x69, 0x72, 0x65, 0x63, 0x74, 0x20, 0x72, - 0x65, 0x6e, 0x64, 0x65, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x6f, 0x66, 0x20, - 0x64, 0x65, 0x73, 0x63, 0x65, 0x6e, 0x64, 0x61, 0x6e, 0x74, 0x73, 0x20, - 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x67, 0x69, 0x76, - 0x65, 0x6e, 0x20, 0x43, 0x53, 0x53, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, - 0x74, 0x6f, 0x72, 0x20, 0x2a, 0x2f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6c, 0x61, 0x73, 0x73, 0x20, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x20, - 0x65, 0x78, 0x74, 0x65, 0x6e, 0x64, 0x73, 0x20, 0x43, 0x6f, 0x6d, 0x70, - 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x44, - 0x69, 0x64, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x70, 0x72, 0x6f, - 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, 0x6c, 0x65, 0x74, 0x20, 0x69, - 0x20, 0x69, 0x6e, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x70, 0x2e, 0x74, 0x6f, + 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x7d, 0x3a, 0x20, 0x3c, 0x2f, 0x73, 0x70, + 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, + 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, + 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x70, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x20, + 0x2a, 0x20, 0x31, 0x30, 0x30, 0x29, 0x7d, 0x25, 0x3c, 0x2f, 0x73, 0x70, + 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, + 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, + 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x70, + 0x6f, 0x76, 0x65, 0x72, 0x7d, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, + 0x24, 0x7b, 0x7b, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, + 0x6e, 0x64, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x70, 0x43, 0x6f, + 0x6c, 0x6f, 0x72, 0x20, 0x7d, 0x7d, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, + 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x3d, 0x24, + 0x7b, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, + 0x64, 0x72, 0x65, 0x6e, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6d, 0x73, 0x67, + 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2e, 0x6d, 0x61, 0x74, + 0x63, 0x68, 0x28, 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x29, 0x20, + 0x3f, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x62, 0x72, 0x20, 0x2f, + 0x3e, 0x60, 0x20, 0x3a, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, + 0x74, 0x65, 0x6e, 0x74, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x70, 0x6f, 0x6f, 0x72, 0x20, 0x6d, + 0x61, 0x6e, 0x73, 0x20, 0x6d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, + 0x20, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x6d, 0x65, 0x6e, 0x74, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x4d, + 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x69, 0x73, 0x68, 0x20, 0x3d, + 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x6d, 0x64, 0x20, 0x3d, 0x20, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x2e, 0x74, 0x65, 0x78, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, + 0x28, 0x2f, 0x26, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x61, 0x6d, 0x70, + 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x3c, 0x2f, + 0x67, 0x2c, 0x20, 0x27, 0x26, 0x6c, 0x74, 0x3b, 0x27, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, + 0x61, 0x63, 0x65, 0x28, 0x2f, 0x3e, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, + 0x67, 0x74, 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, + 0x5e, 0x23, 0x7b, 0x31, 0x2c, 0x36, 0x7d, 0x20, 0x28, 0x2e, 0x2a, 0x29, + 0x24, 0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c, 0x68, 0x33, 0x3e, + 0x24, 0x31, 0x3c, 0x2f, 0x68, 0x33, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, + 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x2a, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, + 0x29, 0x5c, 0x2a, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x73, + 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x73, 0x74, + 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, + 0x28, 0x2f, 0x5f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, 0x5f, 0x2f, + 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, + 0x24, 0x31, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27, + 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, + 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x2a, 0x28, 0x2e, + 0x2a, 0x3f, 0x29, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, + 0x6d, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, + 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, + 0x5f, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31, + 0x3c, 0x2f, 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, + 0x28, 0x2f, 0x60, 0x60, 0x60, 0x2e, 0x2a, 0x3f, 0x5c, 0x6e, 0x28, 0x5b, + 0x5c, 0x73, 0x5c, 0x53, 0x5d, 0x2a, 0x3f, 0x29, 0x60, 0x60, 0x60, 0x2f, + 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x70, 0x72, 0x65, 0x3e, 0x3c, 0x63, 0x6f, + 0x64, 0x65, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e, + 0x3c, 0x2f, 0x70, 0x72, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, + 0x65, 0x28, 0x2f, 0x60, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x60, 0x2f, 0x67, + 0x2c, 0x20, 0x27, 0x3c, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x24, 0x31, 0x3c, + 0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, + 0x65, 0x28, 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, + 0x3c, 0x62, 0x72, 0x20, 0x2f, 0x3e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, + 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, 0x64, 0x61, + 0x6e, 0x67, 0x65, 0x72, 0x6f, 0x75, 0x73, 0x6c, 0x79, 0x53, 0x65, 0x74, + 0x49, 0x6e, 0x6e, 0x65, 0x72, 0x48, 0x54, 0x4d, 0x4c, 0x3d, 0x24, 0x7b, + 0x7b, 0x20, 0x5f, 0x5f, 0x68, 0x74, 0x6d, 0x6c, 0x3a, 0x20, 0x6d, 0x64, + 0x20, 0x7d, 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, + 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d, + 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, + 0x28, 0x21, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x2f, + 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, + 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, + 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x5f, 0x70, 0x72, 0x65, + 0x64, 0x69, 0x63, 0x74, 0x65, 0x64, 0x7d, 0x20, 0x70, 0x72, 0x65, 0x64, + 0x69, 0x63, 0x74, 0x65, 0x64, 0x2c, 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, + 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x5f, 0x63, 0x61, 0x63, + 0x68, 0x65, 0x64, 0x7d, 0x20, 0x63, 0x61, 0x63, 0x68, 0x65, 0x64, 0x2c, + 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, + 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x69, 0x6d, 0x69, + 0x6e, 0x67, 0x73, 0x2e, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x65, + 0x64, 0x5f, 0x70, 0x65, 0x72, 0x5f, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x5f, + 0x6d, 0x73, 0x2e, 0x74, 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x29, + 0x7d, 0x6d, 0x73, 0x20, 0x70, 0x65, 0x72, 0x20, 0x74, 0x6f, 0x6b, 0x65, + 0x6e, 0x2c, 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, + 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x69, + 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x2e, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, + 0x74, 0x65, 0x64, 0x5f, 0x70, 0x65, 0x72, 0x5f, 0x73, 0x65, 0x63, 0x6f, + 0x6e, 0x64, 0x2e, 0x74, 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x32, + 0x29, 0x7d, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x20, 0x70, 0x65, + 0x72, 0x20, 0x73, 0x65, 0x63, 0x6f, 0x6e, 0x64, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, + 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, + 0x72, 0x20, 0x69, 0x6d, 0x70, 0x6c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, + 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x20, 0x3d, + 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x66, + 0x61, 0x6c, 0x73, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, + 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, + 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x20, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x27, + 0x30, 0x70, 0x78, 0x27, 0x2c, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, + 0x27, 0x30, 0x70, 0x78, 0x27, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x62, 0x75, + 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x20, 0x3d, 0x20, 0x75, 0x73, + 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x20, 0x3d, + 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, + 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x74, 0x6f, 0x67, 0x67, 0x6c, 0x65, 0x50, 0x6f, + 0x70, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, + 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x63, 0x74, 0x20, 0x3d, 0x20, + 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, + 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x67, 0x65, 0x74, 0x42, 0x6f, 0x75, + 0x6e, 0x64, 0x69, 0x6e, 0x67, 0x43, 0x6c, 0x69, 0x65, 0x6e, 0x74, 0x52, + 0x65, 0x63, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, + 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x60, 0x24, 0x7b, 0x72, 0x65, 0x63, 0x74, + 0x2e, 0x62, 0x6f, 0x74, 0x74, 0x6f, 0x6d, 0x20, 0x2b, 0x20, 0x77, 0x69, + 0x6e, 0x64, 0x6f, 0x77, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x59, + 0x7d, 0x70, 0x78, 0x60, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, + 0x60, 0x24, 0x7b, 0x72, 0x65, 0x63, 0x74, 0x2e, 0x6c, 0x65, 0x66, 0x74, + 0x20, 0x2b, 0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x2e, 0x73, 0x63, + 0x72, 0x6f, 0x6c, 0x6c, 0x58, 0x7d, 0x70, 0x78, 0x60, 0x2c, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x21, 0x69, 0x73, + 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x68, 0x61, + 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, + 0x73, 0x69, 0x64, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, + 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x6f, 0x70, 0x6f, + 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, + 0x6e, 0x74, 0x20, 0x26, 0x26, 0x20, 0x21, 0x70, 0x6f, 0x70, 0x6f, 0x76, + 0x65, 0x72, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, + 0x74, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x73, 0x28, 0x65, + 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x29, + 0x20, 0x26, 0x26, 0x20, 0x21, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, + 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x63, + 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x73, 0x28, 0x65, 0x76, 0x65, 0x6e, + 0x74, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, - 0x66, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x5b, 0x69, 0x5d, 0x20, - 0x21, 0x3d, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, - 0x70, 0x73, 0x5b, 0x69, 0x5d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, - 0x75, 0x72, 0x6e, 0x20, 0x73, 0x65, 0x74, 0x54, 0x69, 0x6d, 0x65, 0x6f, - 0x75, 0x74, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, - 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x44, 0x69, 0x64, 0x4d, - 0x6f, 0x75, 0x6e, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x73, - 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x74, 0x72, - 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, - 0x61, 0x79, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, - 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x2e, - 0x62, 0x69, 0x6e, 0x64, 0x28, 0x74, 0x68, 0x69, 0x73, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, - 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, - 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, - 0x6e, 0x65, 0x6e, 0x74, 0x57, 0x69, 0x6c, 0x6c, 0x55, 0x6e, 0x6d, 0x6f, - 0x75, 0x6e, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, - 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x28, 0x66, 0x61, 0x6c, - 0x73, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75, 0x6e, - 0x74, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, - 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, - 0x20, 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, - 0x6f, 0x74, 0x65, 0x2e, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x4e, 0x6f, - 0x64, 0x65, 0x29, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, - 0x6f, 0x74, 0x65, 0x2e, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x4e, 0x6f, - 0x64, 0x65, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x76, 0x65, 0x43, 0x68, 0x69, - 0x6c, 0x64, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, - 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x69, 0x6e, 0x64, - 0x4e, 0x6f, 0x64, 0x65, 0x28, 0x6e, 0x6f, 0x64, 0x65, 0x29, 0x20, 0x7b, + 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, + 0x3d, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, + 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, + 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x61, 0x64, + 0x64, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, + 0x65, 0x72, 0x28, 0x27, 0x6d, 0x6f, 0x75, 0x73, 0x65, 0x64, 0x6f, 0x77, + 0x6e, 0x27, 0x2c, 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, + 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, 0x64, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, - 0x75, 0x72, 0x6e, 0x20, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x6e, - 0x6f, 0x64, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x73, 0x74, 0x72, - 0x69, 0x6e, 0x67, 0x27, 0x20, 0x3f, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, - 0x65, 0x6e, 0x74, 0x2e, 0x71, 0x75, 0x65, 0x72, 0x79, 0x53, 0x65, 0x6c, - 0x65, 0x63, 0x74, 0x6f, 0x72, 0x28, 0x6e, 0x6f, 0x64, 0x65, 0x29, 0x20, - 0x3a, 0x20, 0x6e, 0x6f, 0x64, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, - 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x28, 0x73, - 0x68, 0x6f, 0x77, 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x29, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, - 0x20, 0x28, 0x21, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x73, 0x4d, 0x6f, - 0x75, 0x6e, 0x74, 0x65, 0x64, 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x2f, 0x2f, 0x20, 0x63, 0x6c, 0x65, 0x61, 0x6e, 0x20, 0x75, 0x70, 0x20, - 0x6f, 0x6c, 0x64, 0x20, 0x6e, 0x6f, 0x64, 0x65, 0x20, 0x69, 0x66, 0x20, - 0x6d, 0x6f, 0x76, 0x69, 0x6e, 0x67, 0x20, 0x62, 0x61, 0x73, 0x65, 0x73, - 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, - 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, - 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x74, 0x68, - 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x50, 0x6f, 0x69, 0x6e, 0x74, - 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, - 0x6f, 0x50, 0x6f, 0x69, 0x6e, 0x74, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x74, - 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x69, 0x6e, - 0x74, 0x6f, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, - 0x6e, 0x74, 0x6f, 0x20, 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, - 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, - 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x20, 0x3d, 0x20, - 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x68, 0x74, 0x6d, 0x6c, 0x60, - 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, - 0x78, 0x79, 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x2c, 0x20, 0x74, 0x68, 0x69, - 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, - 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, - 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, - 0x2e, 0x66, 0x69, 0x6e, 0x64, 0x4e, 0x6f, 0x64, 0x65, 0x28, 0x74, 0x68, - 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x69, 0x6e, 0x74, - 0x6f, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, - 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x20, 0x3d, - 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x68, 0x74, 0x6d, 0x6c, - 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, - 0x78, 0x79, 0x7d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x3d, - 0x24, 0x7b, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, - 0x78, 0x74, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x73, 0x68, 0x6f, 0x77, 0x20, - 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, - 0x73, 0x2e, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7c, - 0x7c, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x24, 0x7b, 0x50, 0x6f, - 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x2c, 0x20, 0x74, - 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x2c, 0x20, 0x74, 0x68, - 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x29, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, - 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x68, 0x69, 0x67, 0x68, - 0x2d, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, - 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20, 0x72, 0x65, - 0x6e, 0x64, 0x65, 0x72, 0x73, 0x20, 0x69, 0x74, 0x73, 0x20, 0x66, 0x69, - 0x72, 0x73, 0x74, 0x20, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x20, 0x69, 0x66, - 0x20, 0x69, 0x74, 0x20, 0x65, 0x78, 0x69, 0x73, 0x74, 0x73, 0x2e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, - 0x61, 0x73, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x64, 0x69, 0x74, 0x69, - 0x6f, 0x6e, 0x61, 0x6c, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x69, - 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x78, 0x79, 0x2e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x20, 0x50, 0x6f, 0x72, 0x74, - 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x20, 0x65, 0x78, 0x74, 0x65, - 0x6e, 0x64, 0x73, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, - 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x65, - 0x74, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x43, 0x6f, 0x6e, 0x74, 0x65, 0x78, - 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, - 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, 0x6f, 0x6e, 0x74, - 0x65, 0x78, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, - 0x72, 0x28, 0x7b, 0x20, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, - 0x20, 0x7d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x63, 0x68, 0x69, - 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7c, 0x7c, 0x20, 0x6e, 0x75, 0x6c, - 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, - 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x41, 0x70, 0x70, 0x28, 0x70, - 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, - 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x64, 0x69, 0x76, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, - 0x6d, 0x6f, 0x64, 0x65, 0x2d, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, - 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, - 0x65, 0x7d, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x68, 0x65, 0x61, 0x64, 0x65, 0x72, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x68, 0x31, 0x3e, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, - 0x70, 0x3c, 0x2f, 0x68, 0x31, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x68, 0x65, 0x61, 0x64, 0x65, - 0x72, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x6d, 0x61, 0x69, 0x6e, 0x20, 0x69, 0x64, 0x3d, 0x22, - 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x22, 0x3e, 0x0a, 0x20, 0x20, + 0x75, 0x72, 0x6e, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, + 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x76, + 0x65, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, + 0x65, 0x72, 0x28, 0x27, 0x6d, 0x6f, 0x75, 0x73, 0x65, 0x64, 0x6f, 0x77, + 0x6e, 0x27, 0x2c, 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, + 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, 0x64, 0x65, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x5d, 0x29, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, + 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, + 0x73, 0x2e, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x7d, 0x20, 0x72, 0x65, 0x66, + 0x3d, 0x24, 0x7b, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, + 0x7d, 0x20, 0x6f, 0x6e, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, + 0x74, 0x6f, 0x67, 0x67, 0x6c, 0x65, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, + 0x72, 0x7d, 0x3e, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, + 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x7d, 0x3c, 0x2f, 0x73, 0x70, + 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x24, 0x7b, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x20, 0x26, 0x26, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, - 0x7b, 0x63, 0x68, 0x61, 0x74, 0x53, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3f, 0x20, 0x43, 0x68, 0x61, - 0x74, 0x4c, 0x6f, 0x67, 0x20, 0x3a, 0x20, 0x43, 0x6f, 0x6e, 0x66, 0x69, - 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x6d, 0x61, - 0x69, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, - 0x69, 0x64, 0x3d, 0x22, 0x77, 0x72, 0x69, 0x74, 0x65, 0x22, 0x3e, 0x0a, + 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x7d, 0x20, 0x69, 0x6e, 0x74, + 0x6f, 0x3d, 0x22, 0x23, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x22, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x64, 0x69, 0x76, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x66, 0x3d, + 0x24, 0x7b, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x70, + 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x2d, 0x63, 0x6f, 0x6e, 0x74, 0x65, + 0x6e, 0x74, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, + 0x24, 0x7b, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x70, 0x6f, 0x73, + 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, + 0x74, 0x6f, 0x70, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, 0x70, + 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x6c, 0x65, 0x66, 0x74, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, - 0x3d, 0x20, 0x27, 0x63, 0x68, 0x61, 0x74, 0x27, 0x20, 0x3f, 0x20, 0x4d, - 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, 0x20, - 0x3a, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, - 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x73, 0x7d, 0x20, 0x2f, 0x3e, + 0x20, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x70, 0x6f, 0x70, + 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x24, 0x7b, 0x50, + 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x60, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x53, 0x6f, 0x75, 0x72, 0x63, + 0x65, 0x3a, 0x20, 0x70, 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, + 0x72, 0x74, 0x61, 0x6c, 0x20, 0x28, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, + 0x2f, 0x2f, 0x67, 0x69, 0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, + 0x2f, 0x64, 0x65, 0x76, 0x65, 0x6c, 0x6f, 0x70, 0x69, 0x74, 0x2f, 0x70, + 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, + 0x2f, 0x62, 0x6c, 0x6f, 0x62, 0x2f, 0x6d, 0x61, 0x73, 0x74, 0x65, 0x72, + 0x2f, 0x73, 0x72, 0x63, 0x2f, 0x70, 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, + 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x2e, 0x6a, 0x73, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x2a, 0x20, 0x52, 0x65, 0x64, 0x69, 0x72, + 0x65, 0x63, 0x74, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x69, 0x6e, + 0x67, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x65, 0x73, 0x63, 0x65, 0x6e, 0x64, + 0x61, 0x6e, 0x74, 0x73, 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x74, 0x68, + 0x65, 0x20, 0x67, 0x69, 0x76, 0x65, 0x6e, 0x20, 0x43, 0x53, 0x53, 0x20, + 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x20, 0x2a, 0x2f, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x20, 0x50, 0x6f, + 0x72, 0x74, 0x61, 0x6c, 0x20, 0x65, 0x78, 0x74, 0x65, 0x6e, 0x64, 0x73, + 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, + 0x6e, 0x65, 0x6e, 0x74, 0x44, 0x69, 0x64, 0x55, 0x70, 0x64, 0x61, 0x74, + 0x65, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, + 0x6c, 0x65, 0x74, 0x20, 0x69, 0x20, 0x69, 0x6e, 0x20, 0x70, 0x72, 0x6f, + 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, + 0x73, 0x5b, 0x69, 0x5d, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x5b, 0x69, 0x5d, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x73, 0x65, 0x74, + 0x54, 0x69, 0x6d, 0x65, 0x6f, 0x75, 0x74, 0x28, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, + 0x74, 0x44, 0x69, 0x64, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x28, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, + 0x69, 0x73, 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, + 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, + 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x20, 0x3d, 0x20, + 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, + 0x61, 0x79, 0x65, 0x72, 0x2e, 0x62, 0x69, 0x6e, 0x64, 0x28, 0x74, 0x68, + 0x69, 0x73, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, + 0x4c, 0x61, 0x79, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x57, 0x69, 0x6c, + 0x6c, 0x55, 0x6e, 0x6d, 0x6f, 0x75, 0x6e, 0x74, 0x28, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, + 0x72, 0x28, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, + 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x66, + 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, + 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x20, 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x2e, 0x70, 0x61, 0x72, + 0x65, 0x6e, 0x74, 0x4e, 0x6f, 0x64, 0x65, 0x29, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x2e, 0x70, 0x61, 0x72, + 0x65, 0x6e, 0x74, 0x4e, 0x6f, 0x64, 0x65, 0x2e, 0x72, 0x65, 0x6d, 0x6f, + 0x76, 0x65, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x28, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x66, 0x69, 0x6e, 0x64, 0x4e, 0x6f, 0x64, 0x65, 0x28, 0x6e, 0x6f, + 0x64, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x79, 0x70, + 0x65, 0x6f, 0x66, 0x20, 0x6e, 0x6f, 0x64, 0x65, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x27, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x20, 0x3f, 0x20, + 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x71, 0x75, 0x65, + 0x72, 0x79, 0x53, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x28, 0x6e, + 0x6f, 0x64, 0x65, 0x29, 0x20, 0x3a, 0x20, 0x6e, 0x6f, 0x64, 0x65, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, + 0x79, 0x65, 0x72, 0x28, 0x73, 0x68, 0x6f, 0x77, 0x20, 0x3d, 0x20, 0x74, + 0x72, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, 0x29, 0x20, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, 0x6c, 0x65, 0x61, + 0x6e, 0x20, 0x75, 0x70, 0x20, 0x6f, 0x6c, 0x64, 0x20, 0x6e, 0x6f, 0x64, + 0x65, 0x20, 0x69, 0x66, 0x20, 0x6d, 0x6f, 0x76, 0x69, 0x6e, 0x67, 0x20, + 0x62, 0x61, 0x73, 0x65, 0x73, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, + 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x21, + 0x3d, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, + 0x50, 0x6f, 0x69, 0x6e, 0x74, 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x50, 0x6f, 0x69, 0x6e, 0x74, 0x65, + 0x72, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, + 0x70, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, + 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x26, 0x26, 0x20, + 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, + 0x74, 0x65, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, + 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, + 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, 0x20, 0x2f, 0x3e, 0x60, + 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x2c, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x3d, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x66, 0x69, 0x6e, 0x64, 0x4e, 0x6f, + 0x64, 0x65, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, + 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, + 0x6f, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, + 0x28, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, + 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, 0x20, 0x63, 0x6f, 0x6e, + 0x74, 0x65, 0x78, 0x74, 0x3d, 0x24, 0x7b, 0x74, 0x68, 0x69, 0x73, 0x2e, + 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x7d, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, + 0x73, 0x68, 0x6f, 0x77, 0x20, 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, 0x68, 0x69, 0x6c, 0x64, + 0x72, 0x65, 0x6e, 0x20, 0x7c, 0x7c, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x2f, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, - 0x6f, 0x74, 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x70, 0x3e, 0x3c, 0x24, 0x7b, - 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, - 0x69, 0x6f, 0x6e, 0x49, 0x6e, 0x66, 0x6f, 0x7d, 0x20, 0x2f, 0x3e, 0x3c, - 0x2f, 0x70, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x70, 0x3e, 0x50, 0x6f, 0x77, 0x65, 0x72, - 0x65, 0x64, 0x20, 0x62, 0x79, 0x20, 0x3c, 0x61, 0x20, 0x68, 0x72, 0x65, + 0x2f, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, + 0x78, 0x79, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x60, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, + 0x6f, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, + 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, + 0x65, 0x72, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, + 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, + 0x20, 0x68, 0x69, 0x67, 0x68, 0x2d, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x20, + 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x74, 0x68, + 0x61, 0x74, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x73, 0x20, 0x69, + 0x74, 0x73, 0x20, 0x66, 0x69, 0x72, 0x73, 0x74, 0x20, 0x63, 0x68, 0x69, + 0x6c, 0x64, 0x20, 0x69, 0x66, 0x20, 0x69, 0x74, 0x20, 0x65, 0x78, 0x69, + 0x73, 0x74, 0x73, 0x2e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, + 0x75, 0x73, 0x65, 0x64, 0x20, 0x61, 0x73, 0x20, 0x61, 0x20, 0x63, 0x6f, + 0x6e, 0x64, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x20, 0x72, 0x65, + 0x6e, 0x64, 0x65, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x78, + 0x79, 0x2e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, + 0x20, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, + 0x20, 0x65, 0x78, 0x74, 0x65, 0x6e, 0x64, 0x73, 0x20, 0x43, 0x6f, 0x6d, + 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x67, 0x65, 0x74, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x43, + 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, + 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, + 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x7b, 0x20, 0x63, 0x68, 0x69, + 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7d, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, + 0x6e, 0x20, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7c, + 0x7c, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, + 0x41, 0x70, 0x70, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x63, 0x6c, + 0x61, 0x73, 0x73, 0x3d, 0x22, 0x6d, 0x6f, 0x64, 0x65, 0x2d, 0x24, 0x7b, + 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x7d, 0x22, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x65, 0x61, + 0x64, 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x31, 0x3e, 0x6c, 0x6c, 0x61, + 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x3c, 0x2f, 0x68, 0x31, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, + 0x68, 0x65, 0x61, 0x64, 0x65, 0x72, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6d, 0x61, 0x69, 0x6e, + 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, + 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x63, 0x68, 0x61, 0x74, 0x53, 0x74, + 0x61, 0x72, 0x74, 0x65, 0x64, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, + 0x3f, 0x20, 0x43, 0x68, 0x61, 0x74, 0x4c, 0x6f, 0x67, 0x20, 0x3a, 0x20, + 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x7d, 0x20, + 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x2f, 0x6d, 0x61, 0x69, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x65, 0x63, + 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x77, 0x72, 0x69, + 0x74, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, + 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x63, 0x68, 0x61, 0x74, + 0x27, 0x20, 0x3f, 0x20, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x49, + 0x6e, 0x70, 0x75, 0x74, 0x20, 0x3a, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, + 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, + 0x73, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, + 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x70, 0x3e, 0x3c, 0x24, 0x7b, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x47, 0x65, + 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x49, 0x6e, 0x66, 0x6f, + 0x7d, 0x20, 0x2f, 0x3e, 0x3c, 0x2f, 0x70, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x70, 0x3e, + 0x50, 0x6f, 0x77, 0x65, 0x72, 0x65, 0x64, 0x20, 0x62, 0x79, 0x20, 0x3c, + 0x61, 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, 0x22, 0x68, 0x74, 0x74, 0x70, + 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x69, 0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, + 0x6f, 0x6d, 0x2f, 0x67, 0x67, 0x65, 0x72, 0x67, 0x61, 0x6e, 0x6f, 0x76, + 0x2f, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x22, 0x3e, + 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x3c, 0x2f, 0x61, + 0x3e, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x3c, 0x61, 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, 0x22, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, - 0x69, 0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x67, 0x67, - 0x65, 0x72, 0x67, 0x61, 0x6e, 0x6f, 0x76, 0x2f, 0x6c, 0x6c, 0x61, 0x6d, - 0x61, 0x2e, 0x63, 0x70, 0x70, 0x22, 0x3e, 0x6c, 0x6c, 0x61, 0x6d, 0x61, - 0x2e, 0x63, 0x70, 0x70, 0x3c, 0x2f, 0x61, 0x3e, 0x20, 0x61, 0x6e, 0x64, - 0x20, 0x3c, 0x61, 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, 0x22, 0x68, 0x74, - 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x67, 0x6d, 0x6c, 0x2e, 0x61, - 0x69, 0x22, 0x3e, 0x67, 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x3c, 0x2f, - 0x61, 0x3e, 0x2e, 0x3c, 0x2f, 0x70, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x6f, 0x74, - 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x68, 0x28, - 0x41, 0x70, 0x70, 0x29, 0x2c, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, - 0x6e, 0x74, 0x2e, 0x71, 0x75, 0x65, 0x72, 0x79, 0x53, 0x65, 0x6c, 0x65, - 0x63, 0x74, 0x6f, 0x72, 0x28, 0x27, 0x23, 0x63, 0x6f, 0x6e, 0x74, 0x61, - 0x69, 0x6e, 0x65, 0x72, 0x27, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x3c, - 0x2f, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3e, 0x0a, 0x3c, 0x2f, 0x68, - 0x65, 0x61, 0x64, 0x3e, 0x0a, 0x0a, 0x3c, 0x62, 0x6f, 0x64, 0x79, 0x3e, - 0x0a, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x69, 0x64, 0x3d, 0x22, - 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x22, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, - 0x79, 0x70, 0x65, 0x3d, 0x22, 0x66, 0x69, 0x6c, 0x65, 0x22, 0x20, 0x69, - 0x64, 0x3d, 0x22, 0x66, 0x69, 0x6c, 0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, - 0x22, 0x20, 0x61, 0x63, 0x63, 0x65, 0x70, 0x74, 0x3d, 0x22, 0x69, 0x6d, - 0x61, 0x67, 0x65, 0x2f, 0x2a, 0x22, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, - 0x3d, 0x22, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a, 0x20, 0x6e, - 0x6f, 0x6e, 0x65, 0x3b, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x3c, 0x2f, 0x64, - 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x69, - 0x64, 0x3d, 0x22, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x22, 0x3e, 0x3c, - 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x3c, 0x2f, 0x62, 0x6f, 0x64, 0x79, - 0x3e, 0x0a, 0x0a, 0x3c, 0x2f, 0x68, 0x74, 0x6d, 0x6c, 0x3e, 0x0a, 0x0a + 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x22, 0x3e, 0x67, 0x67, 0x6d, 0x6c, + 0x2e, 0x61, 0x69, 0x3c, 0x2f, 0x61, 0x3e, 0x2e, 0x3c, 0x2f, 0x70, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x2f, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, + 0x65, 0x72, 0x28, 0x68, 0x28, 0x41, 0x70, 0x70, 0x29, 0x2c, 0x20, 0x64, + 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x71, 0x75, 0x65, 0x72, + 0x79, 0x53, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x28, 0x27, 0x23, + 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x27, 0x29, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, + 0x3e, 0x0a, 0x3c, 0x2f, 0x68, 0x65, 0x61, 0x64, 0x3e, 0x0a, 0x0a, 0x3c, + 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, + 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, + 0x65, 0x72, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, + 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x66, 0x69, + 0x6c, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x66, 0x69, 0x6c, 0x65, + 0x49, 0x6e, 0x70, 0x75, 0x74, 0x22, 0x20, 0x61, 0x63, 0x63, 0x65, 0x70, + 0x74, 0x3d, 0x22, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x2f, 0x2a, 0x22, 0x20, + 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x22, 0x64, 0x69, 0x73, 0x70, 0x6c, + 0x61, 0x79, 0x3a, 0x20, 0x6e, 0x6f, 0x6e, 0x65, 0x3b, 0x22, 0x3e, 0x0a, + 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x3c, + 0x64, 0x69, 0x76, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x70, 0x6f, 0x72, 0x74, + 0x61, 0x6c, 0x22, 0x3e, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x3c, + 0x2f, 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x0a, 0x3c, 0x2f, 0x68, 0x74, + 0x6d, 0x6c, 0x3e, 0x0a, 0x0a }; -unsigned int index_html_len = 33456; +size_t index_html_len = 33809; diff --git a/examples/server/index.js.hpp b/examples/server/index.js.hpp index e09b3c8c5..24d0b33fa 100644 --- a/examples/server/index.js.hpp +++ b/examples/server/index.js.hpp @@ -12,7 +12,7 @@ unsigned char index_js[] = { 0x74, 0x2c, 0x6e, 0x3d, 0x21, 0x31, 0x3b, 0x77, 0x68, 0x69, 0x6c, 0x65, 0x28, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x21, 0x3d, 0x3d, 0x6f, 0x29, 0x7b, 0x6c, 0x65, 0x74, 0x20, 0x5f, 0x3d, 0x6f, 0x3b, 0x6f, 0x3d, 0x76, - 0x6f, 0x69, 0x64, 0x20, 0x30, 0x3b, 0x73, 0x2b, 0x2b, 0x3b, 0x77, 0x68, + 0x6f, 0x69, 0x64, 0x20, 0x30, 0x3b, 0x6c, 0x2b, 0x2b, 0x3b, 0x77, 0x68, 0x69, 0x6c, 0x65, 0x28, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x21, 0x3d, 0x3d, 0x5f, 0x29, 0x7b, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x3d, 0x5f, 0x2e, 0x6f, 0x3b, 0x5f, 0x2e, 0x6f, 0x3d, 0x76, 0x6f, 0x69, 0x64, @@ -21,7 +21,7 @@ unsigned char index_js[] = { 0x70, 0x28, 0x5f, 0x29, 0x29, 0x74, 0x72, 0x79, 0x7b, 0x5f, 0x2e, 0x63, 0x28, 0x29, 0x7d, 0x63, 0x61, 0x74, 0x63, 0x68, 0x28, 0x65, 0x29, 0x7b, 0x69, 0x66, 0x28, 0x21, 0x6e, 0x29, 0x7b, 0x74, 0x3d, 0x65, 0x3b, 0x6e, - 0x3d, 0x21, 0x30, 0x7d, 0x7d, 0x5f, 0x3d, 0x69, 0x7d, 0x7d, 0x73, 0x3d, + 0x3d, 0x21, 0x30, 0x7d, 0x7d, 0x5f, 0x3d, 0x69, 0x7d, 0x7d, 0x6c, 0x3d, 0x30, 0x3b, 0x66, 0x2d, 0x2d, 0x3b, 0x69, 0x66, 0x28, 0x6e, 0x29, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x74, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x5f, 0x28, 0x74, 0x29, 0x7b, 0x69, 0x66, 0x28, @@ -37,8 +37,8 @@ unsigned char index_js[] = { 0x30, 0x3b, 0x72, 0x2b, 0x2b, 0x3b, 0x74, 0x72, 0x79, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x28, 0x29, 0x7d, 0x66, 0x69, 0x6e, 0x61, 0x6c, 0x6c, 0x79, 0x7b, 0x72, 0x2d, 0x2d, 0x3b, 0x69, 0x3d, 0x6e, - 0x7d, 0x7d, 0x6c, 0x65, 0x74, 0x20, 0x66, 0x3d, 0x30, 0x2c, 0x73, 0x3d, - 0x30, 0x2c, 0x6c, 0x3d, 0x30, 0x3b, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, + 0x7d, 0x7d, 0x6c, 0x65, 0x74, 0x20, 0x66, 0x3d, 0x30, 0x2c, 0x6c, 0x3d, + 0x30, 0x2c, 0x73, 0x3d, 0x30, 0x3b, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x63, 0x28, 0x74, 0x29, 0x7b, 0x69, 0x66, 0x28, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x3d, 0x3d, 0x3d, 0x69, 0x29, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x6c, 0x65, 0x74, 0x20, 0x6e, 0x3d, 0x74, @@ -141,9 +141,9 @@ unsigned char index_js[] = { 0x20, 0x73, 0x69, 0x64, 0x65, 0x2d, 0x65, 0x66, 0x66, 0x65, 0x63, 0x74, 0x73, 0x22, 0x29, 0x7d, 0x28, 0x29, 0x3b, 0x69, 0x66, 0x28, 0x6e, 0x21, 0x3d, 0x3d, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x76, 0x29, 0x7b, 0x69, 0x66, - 0x28, 0x73, 0x3e, 0x31, 0x30, 0x30, 0x29, 0x74, 0x28, 0x29, 0x3b, 0x74, + 0x28, 0x6c, 0x3e, 0x31, 0x30, 0x30, 0x29, 0x74, 0x28, 0x29, 0x3b, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x76, 0x3d, 0x6e, 0x3b, 0x74, 0x68, 0x69, 0x73, - 0x2e, 0x69, 0x2b, 0x2b, 0x3b, 0x6c, 0x2b, 0x2b, 0x3b, 0x66, 0x2b, 0x2b, + 0x2e, 0x69, 0x2b, 0x2b, 0x3b, 0x73, 0x2b, 0x2b, 0x3b, 0x66, 0x2b, 0x2b, 0x3b, 0x74, 0x72, 0x79, 0x7b, 0x66, 0x6f, 0x72, 0x28, 0x6c, 0x65, 0x74, 0x20, 0x74, 0x3d, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x74, 0x3b, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x21, 0x3d, 0x3d, 0x74, 0x3b, 0x74, 0x3d, 0x74, @@ -191,7 +191,7 @@ unsigned char index_js[] = { 0x6f, 0x69, 0x64, 0x20, 0x30, 0x29, 0x3b, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x78, 0x3d, 0x74, 0x3b, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x73, 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x3b, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x67, - 0x3d, 0x6c, 0x2d, 0x31, 0x3b, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x66, 0x3d, + 0x3d, 0x73, 0x2d, 0x31, 0x3b, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x66, 0x3d, 0x34, 0x7d, 0x28, 0x79, 0x2e, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x6e, 0x65, 0x77, 0x20, 0x68, 0x29, 0x2e, 0x68, 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x29, 0x7b, 0x74, @@ -201,9 +201,9 @@ unsigned char index_js[] = { 0x3d, 0x3d, 0x28, 0x33, 0x36, 0x26, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x66, 0x29, 0x29, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x21, 0x30, 0x3b, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x66, 0x26, 0x3d, 0x2d, 0x35, 0x3b, 0x69, 0x66, - 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x67, 0x3d, 0x3d, 0x3d, 0x6c, 0x29, + 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x67, 0x3d, 0x3d, 0x3d, 0x73, 0x29, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x21, 0x30, 0x3b, 0x74, 0x68, 0x69, - 0x73, 0x2e, 0x67, 0x3d, 0x6c, 0x3b, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x66, + 0x73, 0x2e, 0x67, 0x3d, 0x73, 0x3b, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x66, 0x7c, 0x3d, 0x31, 0x3b, 0x69, 0x66, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x3e, 0x30, 0x26, 0x26, 0x21, 0x70, 0x28, 0x74, 0x68, 0x69, 0x73, 0x29, 0x29, 0x7b, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x66, 0x26, 0x3d, 0x2d, @@ -437,321 +437,340 @@ unsigned char index_js[] = { 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x74, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x3f, 0x71, 0x28, 0x74, 0x29, 0x3a, 0x6e, 0x75, 0x6c, 0x6c, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, - 0x69, 0x6f, 0x6e, 0x20, 0x42, 0x28, 0x74, 0x29, 0x7b, 0x76, 0x61, 0x72, - 0x20, 0x6e, 0x2c, 0x65, 0x3b, 0x69, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, - 0x21, 0x3d, 0x28, 0x74, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x29, 0x26, 0x26, - 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x29, - 0x7b, 0x66, 0x6f, 0x72, 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x65, 0x3d, 0x74, - 0x2e, 0x5f, 0x5f, 0x63, 0x2e, 0x62, 0x61, 0x73, 0x65, 0x3d, 0x6e, 0x75, - 0x6c, 0x6c, 0x2c, 0x6e, 0x3d, 0x30, 0x3b, 0x6e, 0x3c, 0x74, 0x2e, 0x5f, - 0x5f, 0x6b, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3b, 0x6e, 0x2b, - 0x2b, 0x29, 0x69, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x28, - 0x65, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x6b, 0x5b, 0x6e, 0x5d, 0x29, 0x26, - 0x26, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x65, 0x2e, 0x5f, 0x5f, 0x65, - 0x29, 0x7b, 0x74, 0x2e, 0x5f, 0x5f, 0x65, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, - 0x63, 0x2e, 0x62, 0x61, 0x73, 0x65, 0x3d, 0x65, 0x2e, 0x5f, 0x5f, 0x65, - 0x3b, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x7d, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x20, 0x42, 0x28, 0x74, 0x29, 0x7d, 0x7d, 0x66, 0x75, 0x6e, 0x63, - 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x47, 0x28, 0x74, 0x29, 0x7b, 0x28, 0x21, - 0x74, 0x2e, 0x5f, 0x5f, 0x64, 0x26, 0x26, 0x28, 0x74, 0x2e, 0x5f, 0x5f, - 0x64, 0x3d, 0x21, 0x30, 0x29, 0x26, 0x26, 0x48, 0x2e, 0x70, 0x75, 0x73, - 0x68, 0x28, 0x74, 0x29, 0x26, 0x26, 0x21, 0x7a, 0x2e, 0x5f, 0x5f, 0x72, - 0x2b, 0x2b, 0x7c, 0x7c, 0x50, 0x21, 0x3d, 0x3d, 0x43, 0x2e, 0x64, 0x65, - 0x62, 0x6f, 0x75, 0x6e, 0x63, 0x65, 0x52, 0x65, 0x6e, 0x64, 0x65, 0x72, - 0x69, 0x6e, 0x67, 0x29, 0x26, 0x26, 0x28, 0x28, 0x50, 0x3d, 0x43, 0x2e, - 0x64, 0x65, 0x62, 0x6f, 0x75, 0x6e, 0x63, 0x65, 0x52, 0x65, 0x6e, 0x64, - 0x65, 0x72, 0x69, 0x6e, 0x67, 0x29, 0x7c, 0x7c, 0x4e, 0x29, 0x28, 0x7a, - 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x7a, - 0x28, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x74, 0x2c, 0x6e, 0x2c, 0x65, - 0x2c, 0x5f, 0x2c, 0x69, 0x2c, 0x6f, 0x2c, 0x72, 0x2c, 0x75, 0x2c, 0x66, - 0x3b, 0x66, 0x6f, 0x72, 0x28, 0x48, 0x2e, 0x73, 0x6f, 0x72, 0x74, 0x28, - 0x24, 0x29, 0x3b, 0x74, 0x3d, 0x48, 0x2e, 0x73, 0x68, 0x69, 0x66, 0x74, - 0x28, 0x29, 0x3b, 0x29, 0x74, 0x2e, 0x5f, 0x5f, 0x64, 0x26, 0x26, 0x28, - 0x6e, 0x3d, 0x48, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x2c, 0x5f, - 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x2c, 0x6f, 0x3d, 0x28, 0x69, - 0x3d, 0x28, 0x65, 0x3d, 0x74, 0x29, 0x2e, 0x5f, 0x5f, 0x76, 0x29, 0x2e, - 0x5f, 0x5f, 0x65, 0x2c, 0x75, 0x3d, 0x5b, 0x5d, 0x2c, 0x66, 0x3d, 0x5b, - 0x5d, 0x2c, 0x28, 0x72, 0x3d, 0x65, 0x2e, 0x5f, 0x5f, 0x50, 0x29, 0x26, - 0x26, 0x28, 0x28, 0x5f, 0x3d, 0x4d, 0x28, 0x7b, 0x7d, 0x2c, 0x69, 0x29, - 0x29, 0x2e, 0x5f, 0x5f, 0x76, 0x3d, 0x69, 0x2e, 0x5f, 0x5f, 0x76, 0x2b, - 0x31, 0x2c, 0x43, 0x2e, 0x76, 0x6e, 0x6f, 0x64, 0x65, 0x26, 0x26, 0x43, - 0x2e, 0x76, 0x6e, 0x6f, 0x64, 0x65, 0x28, 0x5f, 0x29, 0x2c, 0x5f, 0x74, - 0x28, 0x72, 0x2c, 0x5f, 0x2c, 0x69, 0x2c, 0x65, 0x2e, 0x5f, 0x5f, 0x6e, - 0x2c, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x21, 0x3d, 0x3d, 0x72, 0x2e, - 0x6f, 0x77, 0x6e, 0x65, 0x72, 0x53, 0x56, 0x47, 0x45, 0x6c, 0x65, 0x6d, - 0x65, 0x6e, 0x74, 0x2c, 0x33, 0x32, 0x26, 0x69, 0x2e, 0x5f, 0x5f, 0x75, - 0x3f, 0x5b, 0x6f, 0x5d, 0x3a, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x75, 0x2c, - 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x6f, 0x3f, 0x71, 0x28, 0x69, 0x29, - 0x3a, 0x6f, 0x2c, 0x21, 0x21, 0x28, 0x33, 0x32, 0x26, 0x69, 0x2e, 0x5f, - 0x5f, 0x75, 0x29, 0x2c, 0x66, 0x29, 0x2c, 0x5f, 0x2e, 0x5f, 0x5f, 0x2e, - 0x5f, 0x5f, 0x6b, 0x5b, 0x5f, 0x2e, 0x5f, 0x5f, 0x69, 0x5d, 0x3d, 0x5f, - 0x2c, 0x69, 0x74, 0x28, 0x75, 0x2c, 0x5f, 0x2c, 0x66, 0x29, 0x2c, 0x5f, - 0x2e, 0x5f, 0x5f, 0x65, 0x21, 0x3d, 0x6f, 0x26, 0x26, 0x42, 0x28, 0x5f, - 0x29, 0x29, 0x2c, 0x48, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3e, - 0x6e, 0x26, 0x26, 0x48, 0x2e, 0x73, 0x6f, 0x72, 0x74, 0x28, 0x24, 0x29, - 0x29, 0x3b, 0x7a, 0x2e, 0x5f, 0x5f, 0x72, 0x3d, 0x30, 0x7d, 0x66, 0x75, - 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x4a, 0x28, 0x74, 0x2c, 0x6e, - 0x2c, 0x65, 0x2c, 0x5f, 0x2c, 0x69, 0x2c, 0x6f, 0x2c, 0x72, 0x2c, 0x75, - 0x2c, 0x66, 0x2c, 0x73, 0x2c, 0x6c, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, - 0x63, 0x2c, 0x68, 0x2c, 0x61, 0x2c, 0x70, 0x2c, 0x64, 0x2c, 0x76, 0x3d, - 0x5f, 0x26, 0x26, 0x5f, 0x2e, 0x5f, 0x5f, 0x6b, 0x7c, 0x7c, 0x56, 0x2c, - 0x79, 0x3d, 0x6e, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3b, 0x66, - 0x6f, 0x72, 0x28, 0x65, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x66, 0x2c, 0x4b, - 0x28, 0x65, 0x2c, 0x6e, 0x2c, 0x76, 0x29, 0x2c, 0x66, 0x3d, 0x65, 0x2e, - 0x5f, 0x5f, 0x64, 0x2c, 0x63, 0x3d, 0x30, 0x3b, 0x63, 0x3c, 0x79, 0x3b, - 0x63, 0x2b, 0x2b, 0x29, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x28, 0x61, - 0x3d, 0x65, 0x2e, 0x5f, 0x5f, 0x6b, 0x5b, 0x63, 0x5d, 0x29, 0x26, 0x26, - 0x22, 0x62, 0x6f, 0x6f, 0x6c, 0x65, 0x61, 0x6e, 0x22, 0x21, 0x3d, 0x74, - 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x61, 0x26, 0x26, 0x22, 0x66, 0x75, - 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x21, 0x3d, 0x74, 0x79, 0x70, - 0x65, 0x6f, 0x66, 0x20, 0x61, 0x26, 0x26, 0x28, 0x68, 0x3d, 0x2d, 0x31, - 0x3d, 0x3d, 0x3d, 0x61, 0x2e, 0x5f, 0x5f, 0x69, 0x3f, 0x54, 0x3a, 0x76, - 0x5b, 0x61, 0x2e, 0x5f, 0x5f, 0x69, 0x5d, 0x7c, 0x7c, 0x54, 0x2c, 0x61, - 0x2e, 0x5f, 0x5f, 0x69, 0x3d, 0x63, 0x2c, 0x5f, 0x74, 0x28, 0x74, 0x2c, - 0x61, 0x2c, 0x68, 0x2c, 0x69, 0x2c, 0x6f, 0x2c, 0x72, 0x2c, 0x75, 0x2c, - 0x66, 0x2c, 0x73, 0x2c, 0x6c, 0x29, 0x2c, 0x70, 0x3d, 0x61, 0x2e, 0x5f, - 0x5f, 0x65, 0x2c, 0x61, 0x2e, 0x72, 0x65, 0x66, 0x26, 0x26, 0x68, 0x2e, - 0x72, 0x65, 0x66, 0x21, 0x3d, 0x61, 0x2e, 0x72, 0x65, 0x66, 0x26, 0x26, - 0x28, 0x68, 0x2e, 0x72, 0x65, 0x66, 0x26, 0x26, 0x72, 0x74, 0x28, 0x68, - 0x2e, 0x72, 0x65, 0x66, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x61, 0x29, - 0x2c, 0x6c, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x61, 0x2e, 0x72, 0x65, - 0x66, 0x2c, 0x61, 0x2e, 0x5f, 0x5f, 0x63, 0x7c, 0x7c, 0x70, 0x2c, 0x61, - 0x29, 0x29, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x64, 0x26, 0x26, - 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x70, 0x26, 0x26, 0x28, 0x64, 0x3d, - 0x70, 0x29, 0x2c, 0x36, 0x35, 0x35, 0x33, 0x36, 0x26, 0x61, 0x2e, 0x5f, - 0x5f, 0x75, 0x7c, 0x7c, 0x68, 0x2e, 0x5f, 0x5f, 0x6b, 0x3d, 0x3d, 0x3d, - 0x61, 0x2e, 0x5f, 0x5f, 0x6b, 0x3f, 0x66, 0x3d, 0x51, 0x28, 0x61, 0x2c, - 0x66, 0x2c, 0x74, 0x29, 0x3a, 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, - 0x6f, 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, - 0x61, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x26, 0x26, 0x76, 0x6f, 0x69, 0x64, - 0x20, 0x30, 0x21, 0x3d, 0x3d, 0x61, 0x2e, 0x5f, 0x5f, 0x64, 0x3f, 0x66, - 0x3d, 0x61, 0x2e, 0x5f, 0x5f, 0x64, 0x3a, 0x70, 0x26, 0x26, 0x28, 0x66, - 0x3d, 0x70, 0x2e, 0x6e, 0x65, 0x78, 0x74, 0x53, 0x69, 0x62, 0x6c, 0x69, - 0x6e, 0x67, 0x29, 0x2c, 0x61, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x76, 0x6f, - 0x69, 0x64, 0x20, 0x30, 0x2c, 0x61, 0x2e, 0x5f, 0x5f, 0x75, 0x26, 0x3d, - 0x2d, 0x31, 0x39, 0x36, 0x36, 0x30, 0x39, 0x29, 0x3b, 0x65, 0x2e, 0x5f, - 0x5f, 0x64, 0x3d, 0x66, 0x2c, 0x65, 0x2e, 0x5f, 0x5f, 0x65, 0x3d, 0x64, - 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x4b, 0x28, - 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x5f, - 0x2c, 0x69, 0x2c, 0x6f, 0x2c, 0x72, 0x2c, 0x75, 0x2c, 0x66, 0x3d, 0x6e, - 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x2c, 0x73, 0x3d, 0x65, 0x2e, - 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x2c, 0x6c, 0x3d, 0x73, 0x2c, 0x63, - 0x3d, 0x30, 0x3b, 0x66, 0x6f, 0x72, 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x6b, - 0x3d, 0x5b, 0x5d, 0x2c, 0x5f, 0x3d, 0x30, 0x3b, 0x5f, 0x3c, 0x66, 0x3b, - 0x5f, 0x2b, 0x2b, 0x29, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x28, 0x69, - 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x6b, 0x5b, 0x5f, 0x5d, 0x3d, 0x6e, 0x75, - 0x6c, 0x6c, 0x3d, 0x3d, 0x28, 0x69, 0x3d, 0x6e, 0x5b, 0x5f, 0x5d, 0x29, - 0x7c, 0x7c, 0x22, 0x62, 0x6f, 0x6f, 0x6c, 0x65, 0x61, 0x6e, 0x22, 0x3d, - 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x69, 0x7c, 0x7c, 0x22, - 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x3d, 0x3d, 0x74, - 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x69, 0x3f, 0x6e, 0x75, 0x6c, 0x6c, - 0x3a, 0x22, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x22, 0x3d, 0x3d, 0x74, - 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x69, 0x7c, 0x7c, 0x22, 0x6e, 0x75, - 0x6d, 0x62, 0x65, 0x72, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, - 0x66, 0x20, 0x69, 0x7c, 0x7c, 0x22, 0x62, 0x69, 0x67, 0x69, 0x6e, 0x74, - 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x69, 0x7c, - 0x7c, 0x69, 0x2e, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x72, 0x75, 0x63, 0x74, - 0x6f, 0x72, 0x3d, 0x3d, 0x53, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x3f, 0x4f, - 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x69, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, - 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x69, 0x29, 0x3a, 0x46, 0x28, 0x69, - 0x29, 0x3f, 0x4f, 0x28, 0x6a, 0x2c, 0x7b, 0x63, 0x68, 0x69, 0x6c, 0x64, - 0x72, 0x65, 0x6e, 0x3a, 0x69, 0x7d, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, - 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x3a, 0x76, - 0x6f, 0x69, 0x64, 0x20, 0x30, 0x3d, 0x3d, 0x3d, 0x69, 0x2e, 0x63, 0x6f, - 0x6e, 0x73, 0x74, 0x72, 0x75, 0x63, 0x74, 0x6f, 0x72, 0x26, 0x26, 0x69, - 0x2e, 0x5f, 0x5f, 0x62, 0x3e, 0x30, 0x3f, 0x4f, 0x28, 0x69, 0x2e, 0x74, - 0x79, 0x70, 0x65, 0x2c, 0x69, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2c, - 0x69, 0x2e, 0x6b, 0x65, 0x79, 0x2c, 0x69, 0x2e, 0x72, 0x65, 0x66, 0x3f, - 0x69, 0x2e, 0x72, 0x65, 0x66, 0x3a, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x69, - 0x2e, 0x5f, 0x5f, 0x76, 0x29, 0x3a, 0x69, 0x29, 0x3f, 0x28, 0x69, 0x2e, - 0x5f, 0x5f, 0x3d, 0x74, 0x2c, 0x69, 0x2e, 0x5f, 0x5f, 0x62, 0x3d, 0x74, - 0x2e, 0x5f, 0x5f, 0x62, 0x2b, 0x31, 0x2c, 0x75, 0x3d, 0x59, 0x28, 0x69, - 0x2c, 0x65, 0x2c, 0x72, 0x3d, 0x5f, 0x2b, 0x63, 0x2c, 0x6c, 0x29, 0x2c, - 0x69, 0x2e, 0x5f, 0x5f, 0x69, 0x3d, 0x75, 0x2c, 0x6f, 0x3d, 0x6e, 0x75, - 0x6c, 0x6c, 0x2c, 0x2d, 0x31, 0x21, 0x3d, 0x3d, 0x75, 0x26, 0x26, 0x28, - 0x6c, 0x2d, 0x2d, 0x2c, 0x28, 0x6f, 0x3d, 0x65, 0x5b, 0x75, 0x5d, 0x29, - 0x26, 0x26, 0x28, 0x6f, 0x2e, 0x5f, 0x5f, 0x75, 0x7c, 0x3d, 0x31, 0x33, - 0x31, 0x30, 0x37, 0x32, 0x29, 0x29, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, - 0x3d, 0x6f, 0x7c, 0x7c, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x3d, 0x6f, - 0x2e, 0x5f, 0x5f, 0x76, 0x3f, 0x28, 0x2d, 0x31, 0x3d, 0x3d, 0x75, 0x26, - 0x26, 0x63, 0x2d, 0x2d, 0x2c, 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, - 0x6f, 0x6e, 0x22, 0x21, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, - 0x69, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x26, 0x26, 0x28, 0x69, 0x2e, 0x5f, - 0x5f, 0x75, 0x7c, 0x3d, 0x36, 0x35, 0x35, 0x33, 0x36, 0x29, 0x29, 0x3a, - 0x75, 0x21, 0x3d, 0x3d, 0x72, 0x26, 0x26, 0x28, 0x75, 0x3d, 0x3d, 0x3d, - 0x72, 0x2b, 0x31, 0x3f, 0x63, 0x2b, 0x2b, 0x3a, 0x75, 0x3e, 0x72, 0x3f, - 0x6c, 0x3e, 0x66, 0x2d, 0x72, 0x3f, 0x63, 0x2b, 0x3d, 0x75, 0x2d, 0x72, - 0x3a, 0x63, 0x2d, 0x2d, 0x3a, 0x63, 0x3d, 0x75, 0x3c, 0x72, 0x26, 0x26, - 0x75, 0x3d, 0x3d, 0x72, 0x2d, 0x31, 0x3f, 0x75, 0x2d, 0x72, 0x3a, 0x30, - 0x2c, 0x75, 0x21, 0x3d, 0x3d, 0x5f, 0x2b, 0x63, 0x26, 0x26, 0x28, 0x69, + 0x69, 0x6f, 0x6e, 0x20, 0x42, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, + 0x7b, 0x76, 0x61, 0x72, 0x20, 0x5f, 0x2c, 0x69, 0x3d, 0x74, 0x2e, 0x5f, + 0x5f, 0x76, 0x2c, 0x6f, 0x3d, 0x69, 0x2e, 0x5f, 0x5f, 0x65, 0x2c, 0x72, + 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x50, 0x3b, 0x69, 0x66, 0x28, 0x72, 0x29, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x28, 0x5f, 0x3d, 0x4d, 0x28, 0x7b, + 0x7d, 0x2c, 0x69, 0x29, 0x29, 0x2e, 0x5f, 0x5f, 0x76, 0x3d, 0x69, 0x2e, + 0x5f, 0x5f, 0x76, 0x2b, 0x31, 0x2c, 0x43, 0x2e, 0x76, 0x6e, 0x6f, 0x64, + 0x65, 0x26, 0x26, 0x43, 0x2e, 0x76, 0x6e, 0x6f, 0x64, 0x65, 0x28, 0x5f, + 0x29, 0x2c, 0x69, 0x74, 0x28, 0x72, 0x2c, 0x5f, 0x2c, 0x69, 0x2c, 0x74, + 0x2e, 0x5f, 0x5f, 0x6e, 0x2c, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x21, + 0x3d, 0x3d, 0x72, 0x2e, 0x6f, 0x77, 0x6e, 0x65, 0x72, 0x53, 0x56, 0x47, + 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x2c, 0x33, 0x32, 0x26, 0x69, + 0x2e, 0x5f, 0x5f, 0x75, 0x3f, 0x5b, 0x6f, 0x5d, 0x3a, 0x6e, 0x75, 0x6c, + 0x6c, 0x2c, 0x6e, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x6f, 0x3f, + 0x71, 0x28, 0x69, 0x29, 0x3a, 0x6f, 0x2c, 0x21, 0x21, 0x28, 0x33, 0x32, + 0x26, 0x69, 0x2e, 0x5f, 0x5f, 0x75, 0x29, 0x2c, 0x65, 0x29, 0x2c, 0x5f, + 0x2e, 0x5f, 0x5f, 0x76, 0x3d, 0x69, 0x2e, 0x5f, 0x5f, 0x76, 0x2c, 0x5f, + 0x2e, 0x5f, 0x5f, 0x2e, 0x5f, 0x5f, 0x6b, 0x5b, 0x5f, 0x2e, 0x5f, 0x5f, + 0x69, 0x5d, 0x3d, 0x5f, 0x2c, 0x5f, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x76, + 0x6f, 0x69, 0x64, 0x20, 0x30, 0x2c, 0x5f, 0x2e, 0x5f, 0x5f, 0x65, 0x21, + 0x3d, 0x6f, 0x26, 0x26, 0x47, 0x28, 0x5f, 0x29, 0x2c, 0x5f, 0x7d, 0x66, + 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x47, 0x28, 0x74, 0x29, + 0x7b, 0x76, 0x61, 0x72, 0x20, 0x6e, 0x2c, 0x65, 0x3b, 0x69, 0x66, 0x28, + 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x28, 0x74, 0x3d, 0x74, 0x2e, 0x5f, + 0x5f, 0x29, 0x26, 0x26, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x74, 0x2e, + 0x5f, 0x5f, 0x63, 0x29, 0x7b, 0x66, 0x6f, 0x72, 0x28, 0x74, 0x2e, 0x5f, + 0x5f, 0x65, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x2e, 0x62, 0x61, 0x73, + 0x65, 0x3d, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x6e, 0x3d, 0x30, 0x3b, 0x6e, + 0x3c, 0x74, 0x2e, 0x5f, 0x5f, 0x6b, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, + 0x68, 0x3b, 0x6e, 0x2b, 0x2b, 0x29, 0x69, 0x66, 0x28, 0x6e, 0x75, 0x6c, + 0x6c, 0x21, 0x3d, 0x28, 0x65, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x6b, 0x5b, + 0x6e, 0x5d, 0x29, 0x26, 0x26, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x65, + 0x2e, 0x5f, 0x5f, 0x65, 0x29, 0x7b, 0x74, 0x2e, 0x5f, 0x5f, 0x65, 0x3d, + 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x2e, 0x62, 0x61, 0x73, 0x65, 0x3d, 0x65, + 0x2e, 0x5f, 0x5f, 0x65, 0x3b, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x7d, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x47, 0x28, 0x74, 0x29, 0x7d, 0x7d, + 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x7a, 0x28, 0x74, + 0x29, 0x7b, 0x28, 0x21, 0x74, 0x2e, 0x5f, 0x5f, 0x64, 0x26, 0x26, 0x28, + 0x74, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x21, 0x30, 0x29, 0x26, 0x26, 0x48, + 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x74, 0x29, 0x26, 0x26, 0x21, 0x4a, + 0x2e, 0x5f, 0x5f, 0x72, 0x2b, 0x2b, 0x7c, 0x7c, 0x50, 0x21, 0x3d, 0x3d, + 0x43, 0x2e, 0x64, 0x65, 0x62, 0x6f, 0x75, 0x6e, 0x63, 0x65, 0x52, 0x65, + 0x6e, 0x64, 0x65, 0x72, 0x69, 0x6e, 0x67, 0x29, 0x26, 0x26, 0x28, 0x28, + 0x50, 0x3d, 0x43, 0x2e, 0x64, 0x65, 0x62, 0x6f, 0x75, 0x6e, 0x63, 0x65, + 0x52, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x69, 0x6e, 0x67, 0x29, 0x7c, 0x7c, + 0x4e, 0x29, 0x28, 0x4a, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, + 0x6f, 0x6e, 0x20, 0x4a, 0x28, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x74, + 0x2c, 0x6e, 0x2c, 0x65, 0x2c, 0x5f, 0x3d, 0x5b, 0x5d, 0x2c, 0x69, 0x3d, + 0x5b, 0x5d, 0x3b, 0x66, 0x6f, 0x72, 0x28, 0x48, 0x2e, 0x73, 0x6f, 0x72, + 0x74, 0x28, 0x24, 0x29, 0x3b, 0x74, 0x3d, 0x48, 0x2e, 0x73, 0x68, 0x69, + 0x66, 0x74, 0x28, 0x29, 0x3b, 0x29, 0x74, 0x2e, 0x5f, 0x5f, 0x64, 0x26, + 0x26, 0x28, 0x65, 0x3d, 0x48, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, + 0x2c, 0x6e, 0x3d, 0x42, 0x28, 0x74, 0x2c, 0x5f, 0x2c, 0x69, 0x29, 0x7c, + 0x7c, 0x6e, 0x2c, 0x30, 0x3d, 0x3d, 0x3d, 0x65, 0x7c, 0x7c, 0x48, 0x2e, + 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3e, 0x65, 0x3f, 0x28, 0x6f, 0x74, + 0x28, 0x5f, 0x2c, 0x6e, 0x2c, 0x69, 0x29, 0x2c, 0x69, 0x2e, 0x6c, 0x65, + 0x6e, 0x67, 0x74, 0x68, 0x3d, 0x5f, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, + 0x68, 0x3d, 0x30, 0x2c, 0x6e, 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, + 0x2c, 0x48, 0x2e, 0x73, 0x6f, 0x72, 0x74, 0x28, 0x24, 0x29, 0x29, 0x3a, + 0x6e, 0x26, 0x26, 0x43, 0x2e, 0x5f, 0x5f, 0x63, 0x26, 0x26, 0x43, 0x2e, + 0x5f, 0x5f, 0x63, 0x28, 0x6e, 0x2c, 0x56, 0x29, 0x29, 0x3b, 0x6e, 0x26, + 0x26, 0x6f, 0x74, 0x28, 0x5f, 0x2c, 0x6e, 0x2c, 0x69, 0x29, 0x2c, 0x4a, + 0x2e, 0x5f, 0x5f, 0x72, 0x3d, 0x30, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, + 0x69, 0x6f, 0x6e, 0x20, 0x4b, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x2c, + 0x5f, 0x2c, 0x69, 0x2c, 0x6f, 0x2c, 0x72, 0x2c, 0x75, 0x2c, 0x66, 0x2c, + 0x6c, 0x2c, 0x73, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x63, 0x2c, 0x68, + 0x2c, 0x61, 0x2c, 0x70, 0x2c, 0x64, 0x2c, 0x76, 0x3d, 0x5f, 0x26, 0x26, + 0x5f, 0x2e, 0x5f, 0x5f, 0x6b, 0x7c, 0x7c, 0x56, 0x2c, 0x79, 0x3d, 0x6e, + 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3b, 0x66, 0x6f, 0x72, 0x28, + 0x65, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x66, 0x2c, 0x51, 0x28, 0x65, 0x2c, + 0x6e, 0x2c, 0x76, 0x29, 0x2c, 0x66, 0x3d, 0x65, 0x2e, 0x5f, 0x5f, 0x64, + 0x2c, 0x63, 0x3d, 0x30, 0x3b, 0x63, 0x3c, 0x79, 0x3b, 0x63, 0x2b, 0x2b, + 0x29, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x28, 0x61, 0x3d, 0x65, 0x2e, + 0x5f, 0x5f, 0x6b, 0x5b, 0x63, 0x5d, 0x29, 0x26, 0x26, 0x22, 0x62, 0x6f, + 0x6f, 0x6c, 0x65, 0x61, 0x6e, 0x22, 0x21, 0x3d, 0x74, 0x79, 0x70, 0x65, + 0x6f, 0x66, 0x20, 0x61, 0x26, 0x26, 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, + 0x69, 0x6f, 0x6e, 0x22, 0x21, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, + 0x20, 0x61, 0x26, 0x26, 0x28, 0x68, 0x3d, 0x2d, 0x31, 0x3d, 0x3d, 0x3d, + 0x61, 0x2e, 0x5f, 0x5f, 0x69, 0x3f, 0x54, 0x3a, 0x76, 0x5b, 0x61, 0x2e, + 0x5f, 0x5f, 0x69, 0x5d, 0x7c, 0x7c, 0x54, 0x2c, 0x61, 0x2e, 0x5f, 0x5f, + 0x69, 0x3d, 0x63, 0x2c, 0x69, 0x74, 0x28, 0x74, 0x2c, 0x61, 0x2c, 0x68, + 0x2c, 0x69, 0x2c, 0x6f, 0x2c, 0x72, 0x2c, 0x75, 0x2c, 0x66, 0x2c, 0x6c, + 0x2c, 0x73, 0x29, 0x2c, 0x70, 0x3d, 0x61, 0x2e, 0x5f, 0x5f, 0x65, 0x2c, + 0x61, 0x2e, 0x72, 0x65, 0x66, 0x26, 0x26, 0x68, 0x2e, 0x72, 0x65, 0x66, + 0x21, 0x3d, 0x61, 0x2e, 0x72, 0x65, 0x66, 0x26, 0x26, 0x28, 0x68, 0x2e, + 0x72, 0x65, 0x66, 0x26, 0x26, 0x75, 0x74, 0x28, 0x68, 0x2e, 0x72, 0x65, + 0x66, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x61, 0x29, 0x2c, 0x73, 0x2e, + 0x70, 0x75, 0x73, 0x68, 0x28, 0x61, 0x2e, 0x72, 0x65, 0x66, 0x2c, 0x61, + 0x2e, 0x5f, 0x5f, 0x63, 0x7c, 0x7c, 0x70, 0x2c, 0x61, 0x29, 0x29, 0x2c, + 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x64, 0x26, 0x26, 0x6e, 0x75, 0x6c, + 0x6c, 0x21, 0x3d, 0x70, 0x26, 0x26, 0x28, 0x64, 0x3d, 0x70, 0x29, 0x2c, + 0x36, 0x35, 0x35, 0x33, 0x36, 0x26, 0x61, 0x2e, 0x5f, 0x5f, 0x75, 0x7c, + 0x7c, 0x68, 0x2e, 0x5f, 0x5f, 0x6b, 0x3d, 0x3d, 0x3d, 0x61, 0x2e, 0x5f, + 0x5f, 0x6b, 0x3f, 0x66, 0x3d, 0x58, 0x28, 0x61, 0x2c, 0x66, 0x2c, 0x74, + 0x29, 0x3a, 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, + 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x61, 0x2e, 0x74, + 0x79, 0x70, 0x65, 0x26, 0x26, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x21, + 0x3d, 0x3d, 0x61, 0x2e, 0x5f, 0x5f, 0x64, 0x3f, 0x66, 0x3d, 0x61, 0x2e, + 0x5f, 0x5f, 0x64, 0x3a, 0x70, 0x26, 0x26, 0x28, 0x66, 0x3d, 0x70, 0x2e, + 0x6e, 0x65, 0x78, 0x74, 0x53, 0x69, 0x62, 0x6c, 0x69, 0x6e, 0x67, 0x29, + 0x2c, 0x61, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, + 0x30, 0x2c, 0x61, 0x2e, 0x5f, 0x5f, 0x75, 0x26, 0x3d, 0x2d, 0x31, 0x39, + 0x36, 0x36, 0x30, 0x39, 0x29, 0x3b, 0x65, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, + 0x66, 0x2c, 0x65, 0x2e, 0x5f, 0x5f, 0x65, 0x3d, 0x64, 0x7d, 0x66, 0x75, + 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x51, 0x28, 0x74, 0x2c, 0x6e, + 0x2c, 0x65, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x5f, 0x2c, 0x69, 0x2c, + 0x6f, 0x2c, 0x72, 0x2c, 0x75, 0x2c, 0x66, 0x3d, 0x6e, 0x2e, 0x6c, 0x65, + 0x6e, 0x67, 0x74, 0x68, 0x2c, 0x6c, 0x3d, 0x65, 0x2e, 0x6c, 0x65, 0x6e, + 0x67, 0x74, 0x68, 0x2c, 0x73, 0x3d, 0x6c, 0x2c, 0x63, 0x3d, 0x30, 0x3b, + 0x66, 0x6f, 0x72, 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x6b, 0x3d, 0x5b, 0x5d, + 0x2c, 0x5f, 0x3d, 0x30, 0x3b, 0x5f, 0x3c, 0x66, 0x3b, 0x5f, 0x2b, 0x2b, + 0x29, 0x72, 0x3d, 0x5f, 0x2b, 0x63, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x21, + 0x3d, 0x28, 0x69, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x6b, 0x5b, 0x5f, 0x5d, + 0x3d, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x28, 0x69, 0x3d, 0x6e, 0x5b, + 0x5f, 0x5d, 0x29, 0x7c, 0x7c, 0x22, 0x62, 0x6f, 0x6f, 0x6c, 0x65, 0x61, + 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x69, + 0x7c, 0x7c, 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, + 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x69, 0x3f, 0x6e, + 0x75, 0x6c, 0x6c, 0x3a, 0x22, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x22, + 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x69, 0x7c, 0x7c, + 0x22, 0x6e, 0x75, 0x6d, 0x62, 0x65, 0x72, 0x22, 0x3d, 0x3d, 0x74, 0x79, + 0x70, 0x65, 0x6f, 0x66, 0x20, 0x69, 0x7c, 0x7c, 0x22, 0x62, 0x69, 0x67, + 0x69, 0x6e, 0x74, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, + 0x20, 0x69, 0x7c, 0x7c, 0x69, 0x2e, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x72, + 0x75, 0x63, 0x74, 0x6f, 0x72, 0x3d, 0x3d, 0x53, 0x74, 0x72, 0x69, 0x6e, + 0x67, 0x3f, 0x4f, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x69, 0x2c, 0x6e, + 0x75, 0x6c, 0x6c, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x6e, 0x75, 0x6c, + 0x6c, 0x29, 0x3a, 0x46, 0x28, 0x69, 0x29, 0x3f, 0x4f, 0x28, 0x6a, 0x2c, + 0x7b, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x3a, 0x69, 0x7d, + 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x6e, + 0x75, 0x6c, 0x6c, 0x29, 0x3a, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x3d, + 0x3d, 0x3d, 0x69, 0x2e, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x72, 0x75, 0x63, + 0x74, 0x6f, 0x72, 0x26, 0x26, 0x69, 0x2e, 0x5f, 0x5f, 0x62, 0x3e, 0x30, + 0x3f, 0x4f, 0x28, 0x69, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x2c, 0x69, 0x2e, + 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2c, 0x69, 0x2e, 0x6b, 0x65, 0x79, 0x2c, + 0x69, 0x2e, 0x72, 0x65, 0x66, 0x3f, 0x69, 0x2e, 0x72, 0x65, 0x66, 0x3a, + 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x69, 0x2e, 0x5f, 0x5f, 0x76, 0x29, 0x3a, + 0x69, 0x29, 0x3f, 0x28, 0x69, 0x2e, 0x5f, 0x5f, 0x3d, 0x74, 0x2c, 0x69, + 0x2e, 0x5f, 0x5f, 0x62, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x62, 0x2b, 0x31, + 0x2c, 0x75, 0x3d, 0x5a, 0x28, 0x69, 0x2c, 0x65, 0x2c, 0x72, 0x2c, 0x73, + 0x29, 0x2c, 0x69, 0x2e, 0x5f, 0x5f, 0x69, 0x3d, 0x75, 0x2c, 0x6f, 0x3d, + 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x2d, 0x31, 0x21, 0x3d, 0x3d, 0x75, 0x26, + 0x26, 0x28, 0x73, 0x2d, 0x2d, 0x2c, 0x28, 0x6f, 0x3d, 0x65, 0x5b, 0x75, + 0x5d, 0x29, 0x26, 0x26, 0x28, 0x6f, 0x2e, 0x5f, 0x5f, 0x75, 0x7c, 0x3d, + 0x31, 0x33, 0x31, 0x30, 0x37, 0x32, 0x29, 0x29, 0x2c, 0x6e, 0x75, 0x6c, + 0x6c, 0x3d, 0x3d, 0x6f, 0x7c, 0x7c, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, + 0x3d, 0x6f, 0x2e, 0x5f, 0x5f, 0x76, 0x3f, 0x28, 0x2d, 0x31, 0x3d, 0x3d, + 0x75, 0x26, 0x26, 0x63, 0x2d, 0x2d, 0x2c, 0x22, 0x66, 0x75, 0x6e, 0x63, + 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x21, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, + 0x66, 0x20, 0x69, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x26, 0x26, 0x28, 0x69, 0x2e, 0x5f, 0x5f, 0x75, 0x7c, 0x3d, 0x36, 0x35, 0x35, 0x33, 0x36, 0x29, - 0x29, 0x29, 0x3a, 0x28, 0x6f, 0x3d, 0x65, 0x5b, 0x5f, 0x5d, 0x29, 0x26, - 0x26, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x6f, 0x2e, 0x6b, 0x65, 0x79, - 0x26, 0x26, 0x6f, 0x2e, 0x5f, 0x5f, 0x65, 0x26, 0x26, 0x28, 0x6f, 0x2e, - 0x5f, 0x5f, 0x65, 0x3d, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x64, 0x26, 0x26, - 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x71, 0x28, 0x6f, 0x29, 0x29, - 0x2c, 0x75, 0x74, 0x28, 0x6f, 0x2c, 0x6f, 0x2c, 0x21, 0x31, 0x29, 0x2c, - 0x65, 0x5b, 0x5f, 0x5d, 0x3d, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x6c, 0x2d, - 0x2d, 0x29, 0x3b, 0x69, 0x66, 0x28, 0x6c, 0x29, 0x66, 0x6f, 0x72, 0x28, - 0x5f, 0x3d, 0x30, 0x3b, 0x5f, 0x3c, 0x73, 0x3b, 0x5f, 0x2b, 0x2b, 0x29, - 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x28, 0x6f, 0x3d, 0x65, 0x5b, 0x5f, - 0x5d, 0x29, 0x26, 0x26, 0x30, 0x3d, 0x3d, 0x28, 0x31, 0x33, 0x31, 0x30, - 0x37, 0x32, 0x26, 0x6f, 0x2e, 0x5f, 0x5f, 0x75, 0x29, 0x26, 0x26, 0x28, - 0x6f, 0x2e, 0x5f, 0x5f, 0x65, 0x3d, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x64, - 0x26, 0x26, 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x71, 0x28, 0x6f, - 0x29, 0x29, 0x2c, 0x75, 0x74, 0x28, 0x6f, 0x2c, 0x6f, 0x29, 0x29, 0x7d, - 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x51, 0x28, 0x74, - 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x5f, 0x2c, - 0x69, 0x3b, 0x69, 0x66, 0x28, 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, - 0x6f, 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, - 0x74, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x29, 0x7b, 0x66, 0x6f, 0x72, 0x28, - 0x5f, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x6b, 0x2c, 0x69, 0x3d, 0x30, 0x3b, - 0x5f, 0x26, 0x26, 0x69, 0x3c, 0x5f, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, - 0x68, 0x3b, 0x69, 0x2b, 0x2b, 0x29, 0x5f, 0x5b, 0x69, 0x5d, 0x26, 0x26, - 0x28, 0x5f, 0x5b, 0x69, 0x5d, 0x2e, 0x5f, 0x5f, 0x3d, 0x74, 0x2c, 0x6e, - 0x3d, 0x51, 0x28, 0x5f, 0x5b, 0x69, 0x5d, 0x2c, 0x6e, 0x2c, 0x65, 0x29, - 0x29, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x7d, 0x72, - 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x2e, 0x5f, 0x5f, 0x65, 0x21, - 0x3d, 0x6e, 0x26, 0x26, 0x28, 0x65, 0x2e, 0x69, 0x6e, 0x73, 0x65, 0x72, - 0x74, 0x42, 0x65, 0x66, 0x6f, 0x72, 0x65, 0x28, 0x74, 0x2e, 0x5f, 0x5f, - 0x65, 0x2c, 0x6e, 0x7c, 0x7c, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x2c, 0x6e, - 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x65, 0x29, 0x2c, 0x6e, 0x26, 0x26, 0x6e, - 0x2e, 0x6e, 0x65, 0x78, 0x74, 0x53, 0x69, 0x62, 0x6c, 0x69, 0x6e, 0x67, + 0x29, 0x3a, 0x75, 0x21, 0x3d, 0x3d, 0x72, 0x26, 0x26, 0x28, 0x75, 0x3d, + 0x3d, 0x3d, 0x72, 0x2b, 0x31, 0x3f, 0x63, 0x2b, 0x2b, 0x3a, 0x75, 0x3e, + 0x72, 0x3f, 0x73, 0x3e, 0x66, 0x2d, 0x72, 0x3f, 0x63, 0x2b, 0x3d, 0x75, + 0x2d, 0x72, 0x3a, 0x63, 0x2d, 0x2d, 0x3a, 0x75, 0x3c, 0x72, 0x3f, 0x75, + 0x3d, 0x3d, 0x72, 0x2d, 0x31, 0x26, 0x26, 0x28, 0x63, 0x3d, 0x75, 0x2d, + 0x72, 0x29, 0x3a, 0x63, 0x3d, 0x30, 0x2c, 0x75, 0x21, 0x3d, 0x3d, 0x5f, + 0x2b, 0x63, 0x26, 0x26, 0x28, 0x69, 0x2e, 0x5f, 0x5f, 0x75, 0x7c, 0x3d, + 0x36, 0x35, 0x35, 0x33, 0x36, 0x29, 0x29, 0x29, 0x3a, 0x28, 0x6f, 0x3d, + 0x65, 0x5b, 0x72, 0x5d, 0x29, 0x26, 0x26, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, + 0x3d, 0x6f, 0x2e, 0x6b, 0x65, 0x79, 0x26, 0x26, 0x6f, 0x2e, 0x5f, 0x5f, + 0x65, 0x26, 0x26, 0x30, 0x3d, 0x3d, 0x28, 0x31, 0x33, 0x31, 0x30, 0x37, + 0x32, 0x26, 0x6f, 0x2e, 0x5f, 0x5f, 0x75, 0x29, 0x26, 0x26, 0x28, 0x6f, + 0x2e, 0x5f, 0x5f, 0x65, 0x3d, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x64, 0x26, + 0x26, 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x71, 0x28, 0x6f, 0x29, + 0x29, 0x2c, 0x66, 0x74, 0x28, 0x6f, 0x2c, 0x6f, 0x2c, 0x21, 0x31, 0x29, + 0x2c, 0x65, 0x5b, 0x72, 0x5d, 0x3d, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x73, + 0x2d, 0x2d, 0x29, 0x3b, 0x69, 0x66, 0x28, 0x73, 0x29, 0x66, 0x6f, 0x72, + 0x28, 0x5f, 0x3d, 0x30, 0x3b, 0x5f, 0x3c, 0x6c, 0x3b, 0x5f, 0x2b, 0x2b, + 0x29, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x28, 0x6f, 0x3d, 0x65, 0x5b, + 0x5f, 0x5d, 0x29, 0x26, 0x26, 0x30, 0x3d, 0x3d, 0x28, 0x31, 0x33, 0x31, + 0x30, 0x37, 0x32, 0x26, 0x6f, 0x2e, 0x5f, 0x5f, 0x75, 0x29, 0x26, 0x26, + 0x28, 0x6f, 0x2e, 0x5f, 0x5f, 0x65, 0x3d, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, + 0x64, 0x26, 0x26, 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x71, 0x28, + 0x6f, 0x29, 0x29, 0x2c, 0x66, 0x74, 0x28, 0x6f, 0x2c, 0x6f, 0x29, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x58, 0x28, - 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, - 0x6e, 0x3d, 0x6e, 0x7c, 0x7c, 0x5b, 0x5d, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, - 0x3d, 0x3d, 0x74, 0x7c, 0x7c, 0x22, 0x62, 0x6f, 0x6f, 0x6c, 0x65, 0x61, - 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x74, - 0x7c, 0x7c, 0x28, 0x46, 0x28, 0x74, 0x29, 0x3f, 0x74, 0x2e, 0x73, 0x6f, - 0x6d, 0x65, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, - 0x28, 0x74, 0x29, 0x7b, 0x58, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7d, 0x29, - 0x29, 0x3a, 0x6e, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x74, 0x29, 0x29, - 0x2c, 0x6e, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, - 0x59, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x2c, 0x5f, 0x29, 0x7b, 0x76, - 0x61, 0x72, 0x20, 0x69, 0x3d, 0x74, 0x2e, 0x6b, 0x65, 0x79, 0x2c, 0x6f, - 0x3d, 0x74, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x2c, 0x72, 0x3d, 0x65, 0x2d, - 0x31, 0x2c, 0x75, 0x3d, 0x65, 0x2b, 0x31, 0x2c, 0x66, 0x3d, 0x6e, 0x5b, - 0x65, 0x5d, 0x3b, 0x69, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, - 0x3d, 0x66, 0x7c, 0x7c, 0x66, 0x26, 0x26, 0x69, 0x3d, 0x3d, 0x66, 0x2e, - 0x6b, 0x65, 0x79, 0x26, 0x26, 0x6f, 0x3d, 0x3d, 0x3d, 0x66, 0x2e, 0x74, - 0x79, 0x70, 0x65, 0x29, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x65, - 0x3b, 0x69, 0x66, 0x28, 0x5f, 0x3e, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x21, - 0x3d, 0x66, 0x26, 0x26, 0x30, 0x3d, 0x3d, 0x28, 0x31, 0x33, 0x31, 0x30, - 0x37, 0x32, 0x26, 0x66, 0x2e, 0x5f, 0x5f, 0x75, 0x29, 0x3f, 0x31, 0x3a, - 0x30, 0x29, 0x29, 0x66, 0x6f, 0x72, 0x28, 0x3b, 0x72, 0x3e, 0x3d, 0x30, - 0x7c, 0x7c, 0x75, 0x3c, 0x6e, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, - 0x3b, 0x29, 0x7b, 0x69, 0x66, 0x28, 0x72, 0x3e, 0x3d, 0x30, 0x29, 0x7b, - 0x69, 0x66, 0x28, 0x28, 0x66, 0x3d, 0x6e, 0x5b, 0x72, 0x5d, 0x29, 0x26, - 0x26, 0x30, 0x3d, 0x3d, 0x28, 0x31, 0x33, 0x31, 0x30, 0x37, 0x32, 0x26, - 0x66, 0x2e, 0x5f, 0x5f, 0x75, 0x29, 0x26, 0x26, 0x69, 0x3d, 0x3d, 0x66, - 0x2e, 0x6b, 0x65, 0x79, 0x26, 0x26, 0x6f, 0x3d, 0x3d, 0x3d, 0x66, 0x2e, - 0x74, 0x79, 0x70, 0x65, 0x29, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, - 0x72, 0x3b, 0x72, 0x2d, 0x2d, 0x7d, 0x69, 0x66, 0x28, 0x75, 0x3c, 0x6e, - 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x29, 0x7b, 0x69, 0x66, 0x28, - 0x28, 0x66, 0x3d, 0x6e, 0x5b, 0x75, 0x5d, 0x29, 0x26, 0x26, 0x30, 0x3d, - 0x3d, 0x28, 0x31, 0x33, 0x31, 0x30, 0x37, 0x32, 0x26, 0x66, 0x2e, 0x5f, - 0x5f, 0x75, 0x29, 0x26, 0x26, 0x69, 0x3d, 0x3d, 0x66, 0x2e, 0x6b, 0x65, - 0x79, 0x26, 0x26, 0x6f, 0x3d, 0x3d, 0x3d, 0x66, 0x2e, 0x74, 0x79, 0x70, - 0x65, 0x29, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x75, 0x3b, 0x75, - 0x2b, 0x2b, 0x7d, 0x7d, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x2d, 0x31, - 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x5a, 0x28, - 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x22, 0x2d, 0x22, 0x3d, 0x3d, - 0x3d, 0x6e, 0x5b, 0x30, 0x5d, 0x3f, 0x74, 0x2e, 0x73, 0x65, 0x74, 0x50, - 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x79, 0x28, 0x6e, 0x2c, 0x6e, 0x75, - 0x6c, 0x6c, 0x3d, 0x3d, 0x65, 0x3f, 0x22, 0x22, 0x3a, 0x65, 0x29, 0x3a, - 0x74, 0x5b, 0x6e, 0x5d, 0x3d, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x65, - 0x3f, 0x22, 0x22, 0x3a, 0x22, 0x6e, 0x75, 0x6d, 0x62, 0x65, 0x72, 0x22, - 0x21, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x65, 0x7c, 0x7c, - 0x41, 0x2e, 0x74, 0x65, 0x73, 0x74, 0x28, 0x6e, 0x29, 0x3f, 0x65, 0x3a, - 0x65, 0x2b, 0x22, 0x70, 0x78, 0x22, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, - 0x69, 0x6f, 0x6e, 0x20, 0x74, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, - 0x2c, 0x5f, 0x2c, 0x69, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x6f, 0x3b, - 0x74, 0x3a, 0x69, 0x66, 0x28, 0x22, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x22, - 0x3d, 0x3d, 0x3d, 0x6e, 0x29, 0x69, 0x66, 0x28, 0x22, 0x73, 0x74, 0x72, - 0x69, 0x6e, 0x67, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, - 0x20, 0x65, 0x29, 0x74, 0x2e, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x2e, 0x63, - 0x73, 0x73, 0x54, 0x65, 0x78, 0x74, 0x3d, 0x65, 0x3b, 0x65, 0x6c, 0x73, - 0x65, 0x7b, 0x69, 0x66, 0x28, 0x22, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, - 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x5f, 0x26, - 0x26, 0x28, 0x74, 0x2e, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x2e, 0x63, 0x73, - 0x73, 0x54, 0x65, 0x78, 0x74, 0x3d, 0x5f, 0x3d, 0x22, 0x22, 0x29, 0x2c, - 0x5f, 0x29, 0x66, 0x6f, 0x72, 0x28, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x5f, - 0x29, 0x65, 0x26, 0x26, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x65, 0x7c, 0x7c, - 0x5a, 0x28, 0x74, 0x2e, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x2c, 0x6e, 0x2c, + 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x5f, + 0x2c, 0x69, 0x3b, 0x69, 0x66, 0x28, 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, + 0x69, 0x6f, 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, + 0x20, 0x74, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x29, 0x7b, 0x66, 0x6f, 0x72, + 0x28, 0x5f, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x6b, 0x2c, 0x69, 0x3d, 0x30, + 0x3b, 0x5f, 0x26, 0x26, 0x69, 0x3c, 0x5f, 0x2e, 0x6c, 0x65, 0x6e, 0x67, + 0x74, 0x68, 0x3b, 0x69, 0x2b, 0x2b, 0x29, 0x5f, 0x5b, 0x69, 0x5d, 0x26, + 0x26, 0x28, 0x5f, 0x5b, 0x69, 0x5d, 0x2e, 0x5f, 0x5f, 0x3d, 0x74, 0x2c, + 0x6e, 0x3d, 0x58, 0x28, 0x5f, 0x5b, 0x69, 0x5d, 0x2c, 0x6e, 0x2c, 0x65, + 0x29, 0x29, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x7d, + 0x74, 0x2e, 0x5f, 0x5f, 0x65, 0x21, 0x3d, 0x6e, 0x26, 0x26, 0x28, 0x65, + 0x2e, 0x69, 0x6e, 0x73, 0x65, 0x72, 0x74, 0x42, 0x65, 0x66, 0x6f, 0x72, + 0x65, 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x65, 0x2c, 0x6e, 0x7c, 0x7c, 0x6e, + 0x75, 0x6c, 0x6c, 0x29, 0x2c, 0x6e, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x65, + 0x29, 0x3b, 0x64, 0x6f, 0x7b, 0x6e, 0x3d, 0x6e, 0x26, 0x26, 0x6e, 0x2e, + 0x6e, 0x65, 0x78, 0x74, 0x53, 0x69, 0x62, 0x6c, 0x69, 0x6e, 0x67, 0x7d, + 0x77, 0x68, 0x69, 0x6c, 0x65, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, + 0x6e, 0x26, 0x26, 0x38, 0x3d, 0x3d, 0x3d, 0x6e, 0x2e, 0x6e, 0x6f, 0x64, + 0x65, 0x54, 0x79, 0x70, 0x65, 0x29, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, + 0x6e, 0x20, 0x6e, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, + 0x20, 0x59, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x20, 0x6e, 0x3d, 0x6e, 0x7c, 0x7c, 0x5b, 0x5d, 0x2c, 0x6e, + 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x74, 0x7c, 0x7c, 0x22, 0x62, 0x6f, 0x6f, + 0x6c, 0x65, 0x61, 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, + 0x66, 0x20, 0x74, 0x7c, 0x7c, 0x28, 0x46, 0x28, 0x74, 0x29, 0x3f, 0x74, + 0x2e, 0x73, 0x6f, 0x6d, 0x65, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, + 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x59, 0x28, 0x74, 0x2c, 0x6e, + 0x29, 0x7d, 0x29, 0x29, 0x3a, 0x6e, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, + 0x74, 0x29, 0x29, 0x2c, 0x6e, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, + 0x6f, 0x6e, 0x20, 0x5a, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x2c, 0x5f, + 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x69, 0x3d, 0x74, 0x2e, 0x6b, 0x65, + 0x79, 0x2c, 0x6f, 0x3d, 0x74, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x2c, 0x72, + 0x3d, 0x65, 0x2d, 0x31, 0x2c, 0x75, 0x3d, 0x65, 0x2b, 0x31, 0x2c, 0x66, + 0x3d, 0x6e, 0x5b, 0x65, 0x5d, 0x3b, 0x69, 0x66, 0x28, 0x6e, 0x75, 0x6c, + 0x6c, 0x3d, 0x3d, 0x3d, 0x66, 0x7c, 0x7c, 0x66, 0x26, 0x26, 0x69, 0x3d, + 0x3d, 0x66, 0x2e, 0x6b, 0x65, 0x79, 0x26, 0x26, 0x6f, 0x3d, 0x3d, 0x3d, + 0x66, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x26, 0x26, 0x30, 0x3d, 0x3d, 0x28, + 0x31, 0x33, 0x31, 0x30, 0x37, 0x32, 0x26, 0x66, 0x2e, 0x5f, 0x5f, 0x75, + 0x29, 0x29, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x65, 0x3b, 0x69, + 0x66, 0x28, 0x5f, 0x3e, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x66, + 0x26, 0x26, 0x30, 0x3d, 0x3d, 0x28, 0x31, 0x33, 0x31, 0x30, 0x37, 0x32, + 0x26, 0x66, 0x2e, 0x5f, 0x5f, 0x75, 0x29, 0x3f, 0x31, 0x3a, 0x30, 0x29, + 0x29, 0x66, 0x6f, 0x72, 0x28, 0x3b, 0x72, 0x3e, 0x3d, 0x30, 0x7c, 0x7c, + 0x75, 0x3c, 0x6e, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3b, 0x29, + 0x7b, 0x69, 0x66, 0x28, 0x72, 0x3e, 0x3d, 0x30, 0x29, 0x7b, 0x69, 0x66, + 0x28, 0x28, 0x66, 0x3d, 0x6e, 0x5b, 0x72, 0x5d, 0x29, 0x26, 0x26, 0x30, + 0x3d, 0x3d, 0x28, 0x31, 0x33, 0x31, 0x30, 0x37, 0x32, 0x26, 0x66, 0x2e, + 0x5f, 0x5f, 0x75, 0x29, 0x26, 0x26, 0x69, 0x3d, 0x3d, 0x66, 0x2e, 0x6b, + 0x65, 0x79, 0x26, 0x26, 0x6f, 0x3d, 0x3d, 0x3d, 0x66, 0x2e, 0x74, 0x79, + 0x70, 0x65, 0x29, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x72, 0x3b, + 0x72, 0x2d, 0x2d, 0x7d, 0x69, 0x66, 0x28, 0x75, 0x3c, 0x6e, 0x2e, 0x6c, + 0x65, 0x6e, 0x67, 0x74, 0x68, 0x29, 0x7b, 0x69, 0x66, 0x28, 0x28, 0x66, + 0x3d, 0x6e, 0x5b, 0x75, 0x5d, 0x29, 0x26, 0x26, 0x30, 0x3d, 0x3d, 0x28, + 0x31, 0x33, 0x31, 0x30, 0x37, 0x32, 0x26, 0x66, 0x2e, 0x5f, 0x5f, 0x75, + 0x29, 0x26, 0x26, 0x69, 0x3d, 0x3d, 0x66, 0x2e, 0x6b, 0x65, 0x79, 0x26, + 0x26, 0x6f, 0x3d, 0x3d, 0x3d, 0x66, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x29, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x75, 0x3b, 0x75, 0x2b, 0x2b, + 0x7d, 0x7d, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x2d, 0x31, 0x7d, 0x66, + 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x74, 0x74, 0x28, 0x74, + 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x22, 0x2d, 0x22, 0x3d, 0x3d, 0x3d, + 0x6e, 0x5b, 0x30, 0x5d, 0x3f, 0x74, 0x2e, 0x73, 0x65, 0x74, 0x50, 0x72, + 0x6f, 0x70, 0x65, 0x72, 0x74, 0x79, 0x28, 0x6e, 0x2c, 0x6e, 0x75, 0x6c, + 0x6c, 0x3d, 0x3d, 0x65, 0x3f, 0x22, 0x22, 0x3a, 0x65, 0x29, 0x3a, 0x74, + 0x5b, 0x6e, 0x5d, 0x3d, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x65, 0x3f, + 0x22, 0x22, 0x3a, 0x22, 0x6e, 0x75, 0x6d, 0x62, 0x65, 0x72, 0x22, 0x21, + 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x65, 0x7c, 0x7c, 0x41, + 0x2e, 0x74, 0x65, 0x73, 0x74, 0x28, 0x6e, 0x29, 0x3f, 0x65, 0x3a, 0x65, + 0x2b, 0x22, 0x70, 0x78, 0x22, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, + 0x6f, 0x6e, 0x20, 0x6e, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x2c, + 0x5f, 0x2c, 0x69, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x6f, 0x3b, 0x74, + 0x3a, 0x69, 0x66, 0x28, 0x22, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x22, 0x3d, + 0x3d, 0x3d, 0x6e, 0x29, 0x69, 0x66, 0x28, 0x22, 0x73, 0x74, 0x72, 0x69, + 0x6e, 0x67, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, + 0x65, 0x29, 0x74, 0x2e, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x2e, 0x63, 0x73, + 0x73, 0x54, 0x65, 0x78, 0x74, 0x3d, 0x65, 0x3b, 0x65, 0x6c, 0x73, 0x65, + 0x7b, 0x69, 0x66, 0x28, 0x22, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x22, + 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x5f, 0x26, 0x26, + 0x28, 0x74, 0x2e, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x2e, 0x63, 0x73, 0x73, + 0x54, 0x65, 0x78, 0x74, 0x3d, 0x5f, 0x3d, 0x22, 0x22, 0x29, 0x2c, 0x5f, + 0x29, 0x66, 0x6f, 0x72, 0x28, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x5f, 0x29, + 0x65, 0x26, 0x26, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x65, 0x7c, 0x7c, 0x74, + 0x74, 0x28, 0x74, 0x2e, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x2c, 0x6e, 0x2c, 0x22, 0x22, 0x29, 0x3b, 0x69, 0x66, 0x28, 0x65, 0x29, 0x66, 0x6f, 0x72, 0x28, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x65, 0x29, 0x5f, 0x26, 0x26, 0x65, 0x5b, 0x6e, 0x5d, 0x3d, 0x3d, 0x3d, 0x5f, 0x5b, 0x6e, 0x5d, 0x7c, 0x7c, - 0x5a, 0x28, 0x74, 0x2e, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x2c, 0x6e, 0x2c, - 0x65, 0x5b, 0x6e, 0x5d, 0x29, 0x7d, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, - 0x66, 0x28, 0x22, 0x6f, 0x22, 0x3d, 0x3d, 0x3d, 0x6e, 0x5b, 0x30, 0x5d, - 0x26, 0x26, 0x22, 0x6e, 0x22, 0x3d, 0x3d, 0x3d, 0x6e, 0x5b, 0x31, 0x5d, - 0x29, 0x6f, 0x3d, 0x6e, 0x21, 0x3d, 0x3d, 0x28, 0x6e, 0x3d, 0x6e, 0x2e, - 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x28, 0x50, 0x6f, - 0x69, 0x6e, 0x74, 0x65, 0x72, 0x43, 0x61, 0x70, 0x74, 0x75, 0x72, 0x65, - 0x29, 0x24, 0x7c, 0x43, 0x61, 0x70, 0x74, 0x75, 0x72, 0x65, 0x24, 0x2f, - 0x2c, 0x22, 0x24, 0x31, 0x22, 0x29, 0x29, 0x2c, 0x6e, 0x3d, 0x6e, 0x2e, - 0x74, 0x6f, 0x4c, 0x6f, 0x77, 0x65, 0x72, 0x43, 0x61, 0x73, 0x65, 0x28, - 0x29, 0x69, 0x6e, 0x20, 0x74, 0x3f, 0x6e, 0x2e, 0x74, 0x6f, 0x4c, 0x6f, - 0x77, 0x65, 0x72, 0x43, 0x61, 0x73, 0x65, 0x28, 0x29, 0x2e, 0x73, 0x6c, - 0x69, 0x63, 0x65, 0x28, 0x32, 0x29, 0x3a, 0x6e, 0x2e, 0x73, 0x6c, 0x69, - 0x63, 0x65, 0x28, 0x32, 0x29, 0x2c, 0x74, 0x2e, 0x6c, 0x7c, 0x7c, 0x28, - 0x74, 0x2e, 0x6c, 0x3d, 0x7b, 0x7d, 0x29, 0x2c, 0x74, 0x2e, 0x6c, 0x5b, - 0x6e, 0x2b, 0x6f, 0x5d, 0x3d, 0x65, 0x2c, 0x65, 0x3f, 0x5f, 0x3f, 0x65, - 0x2e, 0x75, 0x3d, 0x5f, 0x2e, 0x75, 0x3a, 0x28, 0x65, 0x2e, 0x75, 0x3d, - 0x44, 0x61, 0x74, 0x65, 0x2e, 0x6e, 0x6f, 0x77, 0x28, 0x29, 0x2c, 0x74, - 0x2e, 0x61, 0x64, 0x64, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, - 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28, 0x6e, 0x2c, 0x6f, 0x3f, 0x65, 0x74, - 0x3a, 0x6e, 0x74, 0x2c, 0x6f, 0x29, 0x29, 0x3a, 0x74, 0x2e, 0x72, 0x65, - 0x6d, 0x6f, 0x76, 0x65, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, - 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28, 0x6e, 0x2c, 0x6f, 0x3f, 0x65, 0x74, - 0x3a, 0x6e, 0x74, 0x2c, 0x6f, 0x29, 0x3b, 0x65, 0x6c, 0x73, 0x65, 0x7b, - 0x69, 0x66, 0x28, 0x69, 0x29, 0x6e, 0x3d, 0x6e, 0x2e, 0x72, 0x65, 0x70, - 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x78, 0x6c, 0x69, 0x6e, 0x6b, 0x28, - 0x48, 0x7c, 0x3a, 0x68, 0x29, 0x2f, 0x2c, 0x22, 0x68, 0x22, 0x29, 0x2e, - 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x73, 0x4e, 0x61, - 0x6d, 0x65, 0x24, 0x2f, 0x2c, 0x22, 0x73, 0x22, 0x29, 0x3b, 0x65, 0x6c, - 0x73, 0x65, 0x20, 0x69, 0x66, 0x28, 0x22, 0x77, 0x69, 0x64, 0x74, 0x68, - 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, 0x68, 0x65, 0x69, 0x67, - 0x68, 0x74, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, 0x68, 0x72, - 0x65, 0x66, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, 0x6c, 0x69, - 0x73, 0x74, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, 0x66, 0x6f, - 0x72, 0x6d, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, 0x74, 0x61, - 0x62, 0x49, 0x6e, 0x64, 0x65, 0x78, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, - 0x26, 0x22, 0x64, 0x6f, 0x77, 0x6e, 0x6c, 0x6f, 0x61, 0x64, 0x22, 0x21, - 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, 0x72, 0x6f, 0x77, 0x53, 0x70, 0x61, - 0x6e, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, 0x63, 0x6f, 0x6c, - 0x53, 0x70, 0x61, 0x6e, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, - 0x72, 0x6f, 0x6c, 0x65, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x6e, - 0x20, 0x69, 0x6e, 0x20, 0x74, 0x29, 0x74, 0x72, 0x79, 0x7b, 0x74, 0x5b, - 0x6e, 0x5d, 0x3d, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x65, 0x3f, 0x22, - 0x22, 0x3a, 0x65, 0x3b, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x20, 0x74, 0x7d, - 0x63, 0x61, 0x74, 0x63, 0x68, 0x28, 0x74, 0x29, 0x7b, 0x7d, 0x22, 0x66, - 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, - 0x70, 0x65, 0x6f, 0x66, 0x20, 0x65, 0x7c, 0x7c, 0x28, 0x6e, 0x75, 0x6c, - 0x6c, 0x3d, 0x3d, 0x65, 0x7c, 0x7c, 0x21, 0x31, 0x3d, 0x3d, 0x3d, 0x65, - 0x26, 0x26, 0x22, 0x2d, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x5b, 0x34, 0x5d, - 0x3f, 0x74, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x76, 0x65, 0x41, 0x74, 0x74, - 0x72, 0x69, 0x62, 0x75, 0x74, 0x65, 0x28, 0x6e, 0x29, 0x3a, 0x74, 0x2e, - 0x73, 0x65, 0x74, 0x41, 0x74, 0x74, 0x72, 0x69, 0x62, 0x75, 0x74, 0x65, - 0x28, 0x6e, 0x2c, 0x65, 0x29, 0x29, 0x7d, 0x7d, 0x66, 0x75, 0x6e, 0x63, - 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6e, 0x74, 0x28, 0x74, 0x29, 0x7b, 0x76, - 0x61, 0x72, 0x20, 0x6e, 0x3d, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x6c, 0x5b, - 0x74, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x2b, 0x21, 0x31, 0x5d, 0x3b, 0x69, - 0x66, 0x28, 0x74, 0x2e, 0x74, 0x29, 0x7b, 0x69, 0x66, 0x28, 0x74, 0x2e, - 0x74, 0x3c, 0x3d, 0x6e, 0x2e, 0x75, 0x29, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x7d, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x74, 0x2e, 0x74, 0x3d, 0x44, - 0x61, 0x74, 0x65, 0x2e, 0x6e, 0x6f, 0x77, 0x28, 0x29, 0x3b, 0x72, 0x65, - 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x28, 0x43, 0x2e, 0x65, 0x76, 0x65, - 0x6e, 0x74, 0x3f, 0x43, 0x2e, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x74, - 0x29, 0x3a, 0x74, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, - 0x6e, 0x20, 0x65, 0x74, 0x28, 0x74, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, + 0x74, 0x74, 0x28, 0x74, 0x2e, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x2c, 0x6e, + 0x2c, 0x65, 0x5b, 0x6e, 0x5d, 0x29, 0x7d, 0x65, 0x6c, 0x73, 0x65, 0x20, + 0x69, 0x66, 0x28, 0x22, 0x6f, 0x22, 0x3d, 0x3d, 0x3d, 0x6e, 0x5b, 0x30, + 0x5d, 0x26, 0x26, 0x22, 0x6e, 0x22, 0x3d, 0x3d, 0x3d, 0x6e, 0x5b, 0x31, + 0x5d, 0x29, 0x6f, 0x3d, 0x6e, 0x21, 0x3d, 0x3d, 0x28, 0x6e, 0x3d, 0x6e, + 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x28, 0x50, + 0x6f, 0x69, 0x6e, 0x74, 0x65, 0x72, 0x43, 0x61, 0x70, 0x74, 0x75, 0x72, + 0x65, 0x29, 0x24, 0x7c, 0x43, 0x61, 0x70, 0x74, 0x75, 0x72, 0x65, 0x24, + 0x2f, 0x69, 0x2c, 0x22, 0x24, 0x31, 0x22, 0x29, 0x29, 0x2c, 0x6e, 0x3d, + 0x6e, 0x2e, 0x74, 0x6f, 0x4c, 0x6f, 0x77, 0x65, 0x72, 0x43, 0x61, 0x73, + 0x65, 0x28, 0x29, 0x69, 0x6e, 0x20, 0x74, 0x3f, 0x6e, 0x2e, 0x74, 0x6f, + 0x4c, 0x6f, 0x77, 0x65, 0x72, 0x43, 0x61, 0x73, 0x65, 0x28, 0x29, 0x2e, + 0x73, 0x6c, 0x69, 0x63, 0x65, 0x28, 0x32, 0x29, 0x3a, 0x6e, 0x2e, 0x73, + 0x6c, 0x69, 0x63, 0x65, 0x28, 0x32, 0x29, 0x2c, 0x74, 0x2e, 0x6c, 0x7c, + 0x7c, 0x28, 0x74, 0x2e, 0x6c, 0x3d, 0x7b, 0x7d, 0x29, 0x2c, 0x74, 0x2e, + 0x6c, 0x5b, 0x6e, 0x2b, 0x6f, 0x5d, 0x3d, 0x65, 0x2c, 0x65, 0x3f, 0x5f, + 0x3f, 0x65, 0x2e, 0x75, 0x3d, 0x5f, 0x2e, 0x75, 0x3a, 0x28, 0x65, 0x2e, + 0x75, 0x3d, 0x44, 0x61, 0x74, 0x65, 0x2e, 0x6e, 0x6f, 0x77, 0x28, 0x29, + 0x2c, 0x74, 0x2e, 0x61, 0x64, 0x64, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, + 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28, 0x6e, 0x2c, 0x6f, 0x3f, + 0x5f, 0x74, 0x3a, 0x65, 0x74, 0x2c, 0x6f, 0x29, 0x29, 0x3a, 0x74, 0x2e, + 0x72, 0x65, 0x6d, 0x6f, 0x76, 0x65, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, + 0x69, 0x73, 0x74, 0x65, 0x6e, 0x65, 0x72, 0x28, 0x6e, 0x2c, 0x6f, 0x3f, + 0x5f, 0x74, 0x3a, 0x65, 0x74, 0x2c, 0x6f, 0x29, 0x3b, 0x65, 0x6c, 0x73, + 0x65, 0x7b, 0x69, 0x66, 0x28, 0x69, 0x29, 0x6e, 0x3d, 0x6e, 0x2e, 0x72, + 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x78, 0x6c, 0x69, 0x6e, + 0x6b, 0x28, 0x48, 0x7c, 0x3a, 0x68, 0x29, 0x2f, 0x2c, 0x22, 0x68, 0x22, + 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x73, + 0x4e, 0x61, 0x6d, 0x65, 0x24, 0x2f, 0x2c, 0x22, 0x73, 0x22, 0x29, 0x3b, + 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x28, 0x22, 0x77, 0x69, 0x64, + 0x74, 0x68, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, 0x68, 0x65, + 0x69, 0x67, 0x68, 0x74, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, + 0x68, 0x72, 0x65, 0x66, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, + 0x6c, 0x69, 0x73, 0x74, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, + 0x66, 0x6f, 0x72, 0x6d, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, + 0x74, 0x61, 0x62, 0x49, 0x6e, 0x64, 0x65, 0x78, 0x22, 0x21, 0x3d, 0x3d, + 0x6e, 0x26, 0x26, 0x22, 0x64, 0x6f, 0x77, 0x6e, 0x6c, 0x6f, 0x61, 0x64, + 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, 0x72, 0x6f, 0x77, 0x53, + 0x70, 0x61, 0x6e, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x22, 0x63, + 0x6f, 0x6c, 0x53, 0x70, 0x61, 0x6e, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, + 0x26, 0x22, 0x72, 0x6f, 0x6c, 0x65, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x26, + 0x26, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x74, 0x29, 0x74, 0x72, 0x79, 0x7b, + 0x74, 0x5b, 0x6e, 0x5d, 0x3d, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x65, + 0x3f, 0x22, 0x22, 0x3a, 0x65, 0x3b, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x20, + 0x74, 0x7d, 0x63, 0x61, 0x74, 0x63, 0x68, 0x28, 0x74, 0x29, 0x7b, 0x7d, + 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x3d, 0x3d, + 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x65, 0x7c, 0x7c, 0x28, 0x6e, + 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x65, 0x7c, 0x7c, 0x21, 0x31, 0x3d, 0x3d, + 0x3d, 0x65, 0x26, 0x26, 0x22, 0x2d, 0x22, 0x21, 0x3d, 0x3d, 0x6e, 0x5b, + 0x34, 0x5d, 0x3f, 0x74, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x76, 0x65, 0x41, + 0x74, 0x74, 0x72, 0x69, 0x62, 0x75, 0x74, 0x65, 0x28, 0x6e, 0x29, 0x3a, + 0x74, 0x2e, 0x73, 0x65, 0x74, 0x41, 0x74, 0x74, 0x72, 0x69, 0x62, 0x75, + 0x74, 0x65, 0x28, 0x6e, 0x2c, 0x65, 0x29, 0x29, 0x7d, 0x7d, 0x66, 0x75, + 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x65, 0x74, 0x28, 0x74, 0x29, + 0x7b, 0x69, 0x66, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x6c, 0x29, 0x7b, + 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x6c, + 0x5b, 0x74, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x2b, 0x21, 0x31, 0x5d, 0x3b, + 0x69, 0x66, 0x28, 0x74, 0x2e, 0x74, 0x29, 0x7b, 0x69, 0x66, 0x28, 0x74, + 0x2e, 0x74, 0x3c, 0x3d, 0x6e, 0x2e, 0x75, 0x29, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x7d, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x74, 0x2e, 0x74, 0x3d, + 0x44, 0x61, 0x74, 0x65, 0x2e, 0x6e, 0x6f, 0x77, 0x28, 0x29, 0x3b, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x28, 0x43, 0x2e, 0x65, 0x76, + 0x65, 0x6e, 0x74, 0x3f, 0x43, 0x2e, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x28, + 0x74, 0x29, 0x3a, 0x74, 0x29, 0x7d, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, + 0x69, 0x6f, 0x6e, 0x20, 0x5f, 0x74, 0x28, 0x74, 0x29, 0x7b, 0x69, 0x66, + 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x6c, 0x29, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x6c, 0x5b, 0x74, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x2b, 0x21, 0x30, 0x5d, 0x28, 0x43, 0x2e, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x3f, 0x43, 0x2e, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x74, 0x29, 0x3a, 0x74, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, - 0x69, 0x6f, 0x6e, 0x20, 0x5f, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, + 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x2c, 0x5f, 0x2c, 0x69, 0x2c, 0x6f, 0x2c, 0x72, 0x2c, 0x75, 0x2c, 0x66, - 0x2c, 0x73, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x6c, 0x2c, 0x63, 0x2c, + 0x2c, 0x6c, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x73, 0x2c, 0x63, 0x2c, 0x68, 0x2c, 0x61, 0x2c, 0x70, 0x2c, 0x64, 0x2c, 0x76, 0x2c, 0x79, 0x2c, 0x6d, 0x2c, 0x67, 0x2c, 0x62, 0x2c, 0x6b, 0x2c, 0x53, 0x2c, 0x77, 0x2c, 0x78, 0x2c, 0x45, 0x3d, 0x6e, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x3b, 0x69, @@ -762,16 +781,16 @@ unsigned char index_js[] = { 0x28, 0x66, 0x3d, 0x21, 0x21, 0x28, 0x33, 0x32, 0x26, 0x65, 0x2e, 0x5f, 0x5f, 0x75, 0x29, 0x2c, 0x6f, 0x3d, 0x5b, 0x75, 0x3d, 0x6e, 0x2e, 0x5f, 0x5f, 0x65, 0x3d, 0x65, 0x2e, 0x5f, 0x5f, 0x65, 0x5d, 0x29, 0x2c, 0x28, - 0x6c, 0x3d, 0x43, 0x2e, 0x5f, 0x5f, 0x62, 0x29, 0x26, 0x26, 0x6c, 0x28, + 0x73, 0x3d, 0x43, 0x2e, 0x5f, 0x5f, 0x62, 0x29, 0x26, 0x26, 0x73, 0x28, 0x6e, 0x29, 0x3b, 0x74, 0x3a, 0x69, 0x66, 0x28, 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x45, 0x29, 0x74, 0x72, 0x79, 0x7b, 0x69, 0x66, 0x28, 0x79, 0x3d, 0x6e, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2c, 0x6d, 0x3d, - 0x28, 0x6c, 0x3d, 0x45, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, - 0x54, 0x79, 0x70, 0x65, 0x29, 0x26, 0x26, 0x5f, 0x5b, 0x6c, 0x2e, 0x5f, - 0x5f, 0x63, 0x5d, 0x2c, 0x67, 0x3d, 0x6c, 0x3f, 0x6d, 0x3f, 0x6d, 0x2e, + 0x28, 0x73, 0x3d, 0x45, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, + 0x54, 0x79, 0x70, 0x65, 0x29, 0x26, 0x26, 0x5f, 0x5b, 0x73, 0x2e, 0x5f, + 0x5f, 0x63, 0x5d, 0x2c, 0x67, 0x3d, 0x73, 0x3f, 0x6d, 0x3f, 0x6d, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, - 0x6c, 0x2e, 0x5f, 0x5f, 0x3a, 0x5f, 0x2c, 0x65, 0x2e, 0x5f, 0x5f, 0x63, + 0x73, 0x2e, 0x5f, 0x5f, 0x3a, 0x5f, 0x2c, 0x65, 0x2e, 0x5f, 0x5f, 0x63, 0x3f, 0x76, 0x3d, 0x28, 0x63, 0x3d, 0x6e, 0x2e, 0x5f, 0x5f, 0x63, 0x3d, 0x65, 0x2e, 0x5f, 0x5f, 0x63, 0x29, 0x2e, 0x5f, 0x5f, 0x3d, 0x63, 0x2e, 0x5f, 0x5f, 0x45, 0x3a, 0x28, 0x22, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x74, @@ -782,7 +801,7 @@ unsigned char index_js[] = { 0x28, 0x6e, 0x2e, 0x5f, 0x5f, 0x63, 0x3d, 0x63, 0x3d, 0x6e, 0x65, 0x77, 0x20, 0x49, 0x28, 0x79, 0x2c, 0x67, 0x29, 0x2c, 0x63, 0x2e, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x72, 0x75, 0x63, 0x74, 0x6f, 0x72, 0x3d, 0x45, 0x2c, - 0x63, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x3d, 0x66, 0x74, 0x29, + 0x63, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x3d, 0x6c, 0x74, 0x29, 0x2c, 0x6d, 0x26, 0x26, 0x6d, 0x2e, 0x73, 0x75, 0x62, 0x28, 0x63, 0x29, 0x2c, 0x63, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x3d, 0x79, 0x2c, 0x63, 0x2e, 0x73, 0x74, 0x61, 0x74, 0x65, 0x7c, 0x7c, 0x28, 0x63, 0x2e, 0x73, @@ -875,7 +894,7 @@ unsigned char index_js[] = { 0x7b, 0x66, 0x6f, 0x72, 0x28, 0x63, 0x2e, 0x73, 0x74, 0x61, 0x74, 0x65, 0x3d, 0x63, 0x2e, 0x5f, 0x5f, 0x73, 0x2c, 0x63, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x21, 0x31, 0x2c, 0x6b, 0x26, 0x26, 0x6b, 0x28, 0x6e, 0x29, 0x2c, - 0x6c, 0x3d, 0x63, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x63, + 0x73, 0x3d, 0x63, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x63, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2c, 0x63, 0x2e, 0x73, 0x74, 0x61, 0x74, 0x65, 0x2c, 0x63, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x29, 0x2c, 0x77, 0x3d, 0x30, 0x3b, 0x77, 0x3c, 0x63, 0x2e, 0x5f, 0x73, @@ -884,7 +903,7 @@ unsigned char index_js[] = { 0x63, 0x2e, 0x5f, 0x73, 0x62, 0x5b, 0x77, 0x5d, 0x29, 0x3b, 0x63, 0x2e, 0x5f, 0x73, 0x62, 0x3d, 0x5b, 0x5d, 0x7d, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x64, 0x6f, 0x7b, 0x63, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x21, 0x31, 0x2c, - 0x6b, 0x26, 0x26, 0x6b, 0x28, 0x6e, 0x29, 0x2c, 0x6c, 0x3d, 0x63, 0x2e, + 0x6b, 0x26, 0x26, 0x6b, 0x28, 0x6e, 0x29, 0x2c, 0x73, 0x3d, 0x63, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x63, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2c, 0x63, 0x2e, 0x73, 0x74, 0x61, 0x74, 0x65, 0x2c, 0x63, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x29, 0x2c, 0x63, 0x2e, @@ -902,14 +921,14 @@ unsigned char index_js[] = { 0x6f, 0x72, 0x65, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x7c, 0x7c, 0x28, 0x64, 0x3d, 0x63, 0x2e, 0x67, 0x65, 0x74, 0x53, 0x6e, 0x61, 0x70, 0x73, 0x68, 0x6f, 0x74, 0x42, 0x65, 0x66, 0x6f, 0x72, 0x65, 0x55, 0x70, 0x64, - 0x61, 0x74, 0x65, 0x28, 0x61, 0x2c, 0x70, 0x29, 0x29, 0x2c, 0x4a, 0x28, + 0x61, 0x74, 0x65, 0x28, 0x61, 0x2c, 0x70, 0x29, 0x29, 0x2c, 0x4b, 0x28, 0x74, 0x2c, 0x46, 0x28, 0x78, 0x3d, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, - 0x6c, 0x26, 0x26, 0x6c, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x3d, 0x3d, - 0x6a, 0x26, 0x26, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x6c, 0x2e, 0x6b, - 0x65, 0x79, 0x3f, 0x6c, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, - 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x3a, 0x6c, 0x29, 0x3f, 0x78, + 0x73, 0x26, 0x26, 0x73, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x3d, 0x3d, + 0x6a, 0x26, 0x26, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x73, 0x2e, 0x6b, + 0x65, 0x79, 0x3f, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, + 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x3a, 0x73, 0x29, 0x3f, 0x78, 0x3a, 0x5b, 0x78, 0x5d, 0x2c, 0x6e, 0x2c, 0x65, 0x2c, 0x5f, 0x2c, 0x69, - 0x2c, 0x6f, 0x2c, 0x72, 0x2c, 0x75, 0x2c, 0x66, 0x2c, 0x73, 0x29, 0x2c, + 0x2c, 0x6f, 0x2c, 0x72, 0x2c, 0x75, 0x2c, 0x66, 0x2c, 0x6c, 0x29, 0x2c, 0x63, 0x2e, 0x62, 0x61, 0x73, 0x65, 0x3d, 0x6e, 0x2e, 0x5f, 0x5f, 0x65, 0x2c, 0x6e, 0x2e, 0x5f, 0x5f, 0x75, 0x26, 0x3d, 0x2d, 0x31, 0x36, 0x31, 0x2c, 0x63, 0x2e, 0x5f, 0x5f, 0x68, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, @@ -930,198 +949,198 @@ unsigned char index_js[] = { 0x5f, 0x76, 0x3d, 0x3d, 0x3d, 0x65, 0x2e, 0x5f, 0x5f, 0x76, 0x3f, 0x28, 0x6e, 0x2e, 0x5f, 0x5f, 0x6b, 0x3d, 0x65, 0x2e, 0x5f, 0x5f, 0x6b, 0x2c, 0x6e, 0x2e, 0x5f, 0x5f, 0x65, 0x3d, 0x65, 0x2e, 0x5f, 0x5f, 0x65, 0x29, - 0x3a, 0x6e, 0x2e, 0x5f, 0x5f, 0x65, 0x3d, 0x6f, 0x74, 0x28, 0x65, 0x2e, + 0x3a, 0x6e, 0x2e, 0x5f, 0x5f, 0x65, 0x3d, 0x72, 0x74, 0x28, 0x65, 0x2e, 0x5f, 0x5f, 0x65, 0x2c, 0x6e, 0x2c, 0x65, 0x2c, 0x5f, 0x2c, 0x69, 0x2c, - 0x6f, 0x2c, 0x72, 0x2c, 0x66, 0x2c, 0x73, 0x29, 0x3b, 0x28, 0x6c, 0x3d, - 0x43, 0x2e, 0x64, 0x69, 0x66, 0x66, 0x65, 0x64, 0x29, 0x26, 0x26, 0x6c, + 0x6f, 0x2c, 0x72, 0x2c, 0x66, 0x2c, 0x6c, 0x29, 0x3b, 0x28, 0x73, 0x3d, + 0x43, 0x2e, 0x64, 0x69, 0x66, 0x66, 0x65, 0x64, 0x29, 0x26, 0x26, 0x73, 0x28, 0x6e, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, - 0x20, 0x69, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x6e, - 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x3b, - 0x66, 0x6f, 0x72, 0x28, 0x76, 0x61, 0x72, 0x20, 0x5f, 0x3d, 0x30, 0x3b, - 0x5f, 0x3c, 0x65, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3b, 0x5f, - 0x2b, 0x2b, 0x29, 0x72, 0x74, 0x28, 0x65, 0x5b, 0x5f, 0x5d, 0x2c, 0x65, - 0x5b, 0x2b, 0x2b, 0x5f, 0x5d, 0x2c, 0x65, 0x5b, 0x2b, 0x2b, 0x5f, 0x5d, - 0x29, 0x3b, 0x43, 0x2e, 0x5f, 0x5f, 0x63, 0x26, 0x26, 0x43, 0x2e, 0x5f, - 0x5f, 0x63, 0x28, 0x6e, 0x2c, 0x74, 0x29, 0x2c, 0x74, 0x2e, 0x73, 0x6f, - 0x6d, 0x65, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, - 0x28, 0x6e, 0x29, 0x7b, 0x74, 0x72, 0x79, 0x7b, 0x74, 0x3d, 0x6e, 0x2e, - 0x5f, 0x5f, 0x68, 0x2c, 0x6e, 0x2e, 0x5f, 0x5f, 0x68, 0x3d, 0x5b, 0x5d, - 0x2c, 0x74, 0x2e, 0x73, 0x6f, 0x6d, 0x65, 0x28, 0x28, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x74, 0x2e, 0x63, - 0x61, 0x6c, 0x6c, 0x28, 0x6e, 0x29, 0x7d, 0x29, 0x29, 0x7d, 0x63, 0x61, - 0x74, 0x63, 0x68, 0x28, 0x74, 0x29, 0x7b, 0x43, 0x2e, 0x5f, 0x5f, 0x65, - 0x28, 0x74, 0x2c, 0x6e, 0x2e, 0x5f, 0x5f, 0x76, 0x29, 0x7d, 0x7d, 0x29, - 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6f, - 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x2c, 0x5f, 0x2c, 0x69, 0x2c, - 0x6f, 0x2c, 0x72, 0x2c, 0x75, 0x2c, 0x66, 0x29, 0x7b, 0x76, 0x61, 0x72, - 0x20, 0x73, 0x2c, 0x6c, 0x2c, 0x63, 0x2c, 0x68, 0x2c, 0x61, 0x2c, 0x70, - 0x2c, 0x64, 0x2c, 0x76, 0x3d, 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, - 0x2c, 0x79, 0x3d, 0x6e, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2c, 0x6d, - 0x3d, 0x6e, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x3b, 0x69, 0x66, 0x28, 0x22, - 0x73, 0x76, 0x67, 0x22, 0x3d, 0x3d, 0x3d, 0x6d, 0x26, 0x26, 0x28, 0x69, - 0x3d, 0x21, 0x30, 0x29, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x6f, - 0x29, 0x66, 0x6f, 0x72, 0x28, 0x73, 0x3d, 0x30, 0x3b, 0x73, 0x3c, 0x6f, - 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3b, 0x73, 0x2b, 0x2b, 0x29, - 0x69, 0x66, 0x28, 0x28, 0x61, 0x3d, 0x6f, 0x5b, 0x73, 0x5d, 0x29, 0x26, - 0x26, 0x22, 0x73, 0x65, 0x74, 0x41, 0x74, 0x74, 0x72, 0x69, 0x62, 0x75, - 0x74, 0x65, 0x22, 0x69, 0x6e, 0x20, 0x61, 0x3d, 0x3d, 0x21, 0x21, 0x6d, - 0x26, 0x26, 0x28, 0x6d, 0x3f, 0x61, 0x2e, 0x6c, 0x6f, 0x63, 0x61, 0x6c, - 0x4e, 0x61, 0x6d, 0x65, 0x3d, 0x3d, 0x3d, 0x6d, 0x3a, 0x33, 0x3d, 0x3d, - 0x3d, 0x61, 0x2e, 0x6e, 0x6f, 0x64, 0x65, 0x54, 0x79, 0x70, 0x65, 0x29, - 0x29, 0x7b, 0x74, 0x3d, 0x61, 0x2c, 0x6f, 0x5b, 0x73, 0x5d, 0x3d, 0x6e, - 0x75, 0x6c, 0x6c, 0x3b, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x7d, 0x69, 0x66, - 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x74, 0x29, 0x7b, 0x69, 0x66, - 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x3d, 0x6d, 0x29, 0x72, 0x65, - 0x74, 0x75, 0x72, 0x6e, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, - 0x74, 0x2e, 0x63, 0x72, 0x65, 0x61, 0x74, 0x65, 0x54, 0x65, 0x78, 0x74, - 0x4e, 0x6f, 0x64, 0x65, 0x28, 0x79, 0x29, 0x3b, 0x74, 0x3d, 0x69, 0x3f, - 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x63, 0x72, 0x65, - 0x61, 0x74, 0x65, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x4e, 0x53, - 0x28, 0x22, 0x68, 0x74, 0x74, 0x70, 0x3a, 0x2f, 0x2f, 0x77, 0x77, 0x77, - 0x2e, 0x77, 0x33, 0x2e, 0x6f, 0x72, 0x67, 0x2f, 0x32, 0x30, 0x30, 0x30, - 0x2f, 0x73, 0x76, 0x67, 0x22, 0x2c, 0x6d, 0x29, 0x3a, 0x64, 0x6f, 0x63, - 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x63, 0x72, 0x65, 0x61, 0x74, 0x65, - 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x28, 0x6d, 0x2c, 0x79, 0x2e, - 0x69, 0x73, 0x26, 0x26, 0x79, 0x29, 0x2c, 0x6f, 0x3d, 0x6e, 0x75, 0x6c, - 0x6c, 0x2c, 0x75, 0x3d, 0x21, 0x31, 0x7d, 0x69, 0x66, 0x28, 0x6e, 0x75, - 0x6c, 0x6c, 0x3d, 0x3d, 0x3d, 0x6d, 0x29, 0x76, 0x3d, 0x3d, 0x3d, 0x79, - 0x7c, 0x7c, 0x75, 0x26, 0x26, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x3d, - 0x3d, 0x3d, 0x79, 0x7c, 0x7c, 0x28, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, - 0x3d, 0x79, 0x29, 0x3b, 0x65, 0x6c, 0x73, 0x65, 0x7b, 0x69, 0x66, 0x28, - 0x6f, 0x3d, 0x6f, 0x26, 0x26, 0x78, 0x2e, 0x63, 0x61, 0x6c, 0x6c, 0x28, - 0x74, 0x2e, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x4e, 0x6f, 0x64, 0x65, 0x73, - 0x29, 0x2c, 0x76, 0x3d, 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x7c, - 0x7c, 0x54, 0x2c, 0x21, 0x75, 0x26, 0x26, 0x6e, 0x75, 0x6c, 0x6c, 0x21, - 0x3d, 0x6f, 0x29, 0x66, 0x6f, 0x72, 0x28, 0x76, 0x3d, 0x7b, 0x7d, 0x2c, - 0x73, 0x3d, 0x30, 0x3b, 0x73, 0x3c, 0x74, 0x2e, 0x61, 0x74, 0x74, 0x72, - 0x69, 0x62, 0x75, 0x74, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, - 0x68, 0x3b, 0x73, 0x2b, 0x2b, 0x29, 0x76, 0x5b, 0x28, 0x61, 0x3d, 0x74, - 0x2e, 0x61, 0x74, 0x74, 0x72, 0x69, 0x62, 0x75, 0x74, 0x65, 0x73, 0x5b, - 0x73, 0x5d, 0x29, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3d, 0x61, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x66, 0x6f, 0x72, 0x28, 0x73, 0x20, - 0x69, 0x6e, 0x20, 0x76, 0x29, 0x61, 0x3d, 0x76, 0x5b, 0x73, 0x5d, 0x2c, - 0x22, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x22, 0x3d, 0x3d, - 0x73, 0x7c, 0x7c, 0x28, 0x22, 0x64, 0x61, 0x6e, 0x67, 0x65, 0x72, 0x6f, - 0x75, 0x73, 0x6c, 0x79, 0x53, 0x65, 0x74, 0x49, 0x6e, 0x6e, 0x65, 0x72, - 0x48, 0x54, 0x4d, 0x4c, 0x22, 0x3d, 0x3d, 0x73, 0x3f, 0x63, 0x3d, 0x61, - 0x3a, 0x22, 0x6b, 0x65, 0x79, 0x22, 0x3d, 0x3d, 0x3d, 0x73, 0x7c, 0x7c, - 0x73, 0x20, 0x69, 0x6e, 0x20, 0x79, 0x7c, 0x7c, 0x74, 0x74, 0x28, 0x74, - 0x2c, 0x73, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x61, 0x2c, 0x69, 0x29, - 0x29, 0x3b, 0x66, 0x6f, 0x72, 0x28, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x79, - 0x29, 0x61, 0x3d, 0x79, 0x5b, 0x73, 0x5d, 0x2c, 0x22, 0x63, 0x68, 0x69, - 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x22, 0x3d, 0x3d, 0x73, 0x3f, 0x68, 0x3d, - 0x61, 0x3a, 0x22, 0x64, 0x61, 0x6e, 0x67, 0x65, 0x72, 0x6f, 0x75, 0x73, - 0x6c, 0x79, 0x53, 0x65, 0x74, 0x49, 0x6e, 0x6e, 0x65, 0x72, 0x48, 0x54, - 0x4d, 0x4c, 0x22, 0x3d, 0x3d, 0x73, 0x3f, 0x6c, 0x3d, 0x61, 0x3a, 0x22, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x22, 0x3d, 0x3d, 0x73, 0x3f, 0x70, 0x3d, - 0x61, 0x3a, 0x22, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x22, 0x3d, - 0x3d, 0x73, 0x3f, 0x64, 0x3d, 0x61, 0x3a, 0x22, 0x6b, 0x65, 0x79, 0x22, - 0x3d, 0x3d, 0x3d, 0x73, 0x7c, 0x7c, 0x75, 0x26, 0x26, 0x22, 0x66, 0x75, - 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x21, 0x3d, 0x74, 0x79, 0x70, - 0x65, 0x6f, 0x66, 0x20, 0x61, 0x7c, 0x7c, 0x76, 0x5b, 0x73, 0x5d, 0x3d, - 0x3d, 0x3d, 0x61, 0x7c, 0x7c, 0x74, 0x74, 0x28, 0x74, 0x2c, 0x73, 0x2c, - 0x61, 0x2c, 0x76, 0x5b, 0x73, 0x5d, 0x2c, 0x69, 0x29, 0x3b, 0x69, 0x66, - 0x28, 0x6c, 0x29, 0x75, 0x7c, 0x7c, 0x63, 0x26, 0x26, 0x28, 0x6c, 0x2e, - 0x5f, 0x5f, 0x68, 0x74, 0x6d, 0x6c, 0x3d, 0x3d, 0x3d, 0x63, 0x2e, 0x5f, - 0x5f, 0x68, 0x74, 0x6d, 0x6c, 0x7c, 0x7c, 0x6c, 0x2e, 0x5f, 0x5f, 0x68, - 0x74, 0x6d, 0x6c, 0x3d, 0x3d, 0x3d, 0x74, 0x2e, 0x69, 0x6e, 0x6e, 0x65, - 0x72, 0x48, 0x54, 0x4d, 0x4c, 0x29, 0x7c, 0x7c, 0x28, 0x74, 0x2e, 0x69, - 0x6e, 0x6e, 0x65, 0x72, 0x48, 0x54, 0x4d, 0x4c, 0x3d, 0x6c, 0x2e, 0x5f, - 0x5f, 0x68, 0x74, 0x6d, 0x6c, 0x29, 0x2c, 0x6e, 0x2e, 0x5f, 0x5f, 0x6b, - 0x3d, 0x5b, 0x5d, 0x3b, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x28, - 0x63, 0x26, 0x26, 0x28, 0x74, 0x2e, 0x69, 0x6e, 0x6e, 0x65, 0x72, 0x48, - 0x54, 0x4d, 0x4c, 0x3d, 0x22, 0x22, 0x29, 0x2c, 0x4a, 0x28, 0x74, 0x2c, - 0x46, 0x28, 0x68, 0x29, 0x3f, 0x68, 0x3a, 0x5b, 0x68, 0x5d, 0x2c, 0x6e, - 0x2c, 0x65, 0x2c, 0x5f, 0x2c, 0x69, 0x26, 0x26, 0x22, 0x66, 0x6f, 0x72, - 0x65, 0x69, 0x67, 0x6e, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x22, 0x21, - 0x3d, 0x3d, 0x6d, 0x2c, 0x6f, 0x2c, 0x72, 0x2c, 0x6f, 0x3f, 0x6f, 0x5b, - 0x30, 0x5d, 0x3a, 0x65, 0x2e, 0x5f, 0x5f, 0x6b, 0x26, 0x26, 0x71, 0x28, - 0x65, 0x2c, 0x30, 0x29, 0x2c, 0x75, 0x2c, 0x66, 0x29, 0x2c, 0x6e, 0x75, - 0x6c, 0x6c, 0x21, 0x3d, 0x6f, 0x29, 0x66, 0x6f, 0x72, 0x28, 0x73, 0x3d, - 0x6f, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3b, 0x73, 0x2d, 0x2d, - 0x3b, 0x29, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x6f, 0x5b, 0x73, 0x5d, - 0x26, 0x26, 0x57, 0x28, 0x6f, 0x5b, 0x73, 0x5d, 0x29, 0x3b, 0x75, 0x7c, - 0x7c, 0x28, 0x73, 0x3d, 0x22, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x22, 0x2c, - 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x21, 0x3d, 0x3d, 0x70, 0x26, 0x26, - 0x28, 0x70, 0x21, 0x3d, 0x3d, 0x74, 0x5b, 0x73, 0x5d, 0x7c, 0x7c, 0x22, - 0x70, 0x72, 0x6f, 0x67, 0x72, 0x65, 0x73, 0x73, 0x22, 0x3d, 0x3d, 0x3d, - 0x6d, 0x26, 0x26, 0x21, 0x70, 0x7c, 0x7c, 0x22, 0x6f, 0x70, 0x74, 0x69, - 0x6f, 0x6e, 0x22, 0x3d, 0x3d, 0x3d, 0x6d, 0x26, 0x26, 0x70, 0x21, 0x3d, - 0x3d, 0x76, 0x5b, 0x73, 0x5d, 0x29, 0x26, 0x26, 0x74, 0x74, 0x28, 0x74, - 0x2c, 0x73, 0x2c, 0x70, 0x2c, 0x76, 0x5b, 0x73, 0x5d, 0x2c, 0x21, 0x31, - 0x29, 0x2c, 0x73, 0x3d, 0x22, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, - 0x22, 0x2c, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x21, 0x3d, 0x3d, 0x64, - 0x26, 0x26, 0x64, 0x21, 0x3d, 0x3d, 0x74, 0x5b, 0x73, 0x5d, 0x26, 0x26, - 0x74, 0x74, 0x28, 0x74, 0x2c, 0x73, 0x2c, 0x64, 0x2c, 0x76, 0x5b, 0x73, - 0x5d, 0x2c, 0x21, 0x31, 0x29, 0x29, 0x7d, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x20, 0x74, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, - 0x20, 0x72, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x74, - 0x72, 0x79, 0x7b, 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, - 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x74, 0x3f, - 0x74, 0x28, 0x6e, 0x29, 0x3a, 0x74, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, - 0x6e, 0x74, 0x3d, 0x6e, 0x7d, 0x63, 0x61, 0x74, 0x63, 0x68, 0x28, 0x74, - 0x29, 0x7b, 0x43, 0x2e, 0x5f, 0x5f, 0x65, 0x28, 0x74, 0x2c, 0x65, 0x29, - 0x7d, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, - 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x76, 0x61, 0x72, - 0x20, 0x5f, 0x2c, 0x69, 0x3b, 0x69, 0x66, 0x28, 0x43, 0x2e, 0x75, 0x6e, - 0x6d, 0x6f, 0x75, 0x6e, 0x74, 0x26, 0x26, 0x43, 0x2e, 0x75, 0x6e, 0x6d, - 0x6f, 0x75, 0x6e, 0x74, 0x28, 0x74, 0x29, 0x2c, 0x28, 0x5f, 0x3d, 0x74, - 0x2e, 0x72, 0x65, 0x66, 0x29, 0x26, 0x26, 0x28, 0x5f, 0x2e, 0x63, 0x75, - 0x72, 0x72, 0x65, 0x6e, 0x74, 0x26, 0x26, 0x5f, 0x2e, 0x63, 0x75, 0x72, - 0x72, 0x65, 0x6e, 0x74, 0x21, 0x3d, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x65, - 0x7c, 0x7c, 0x72, 0x74, 0x28, 0x5f, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, - 0x6e, 0x29, 0x29, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x28, 0x5f, - 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x29, 0x29, 0x7b, 0x69, 0x66, 0x28, - 0x5f, 0x2e, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x57, - 0x69, 0x6c, 0x6c, 0x55, 0x6e, 0x6d, 0x6f, 0x75, 0x6e, 0x74, 0x29, 0x74, - 0x72, 0x79, 0x7b, 0x5f, 0x2e, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, - 0x6e, 0x74, 0x57, 0x69, 0x6c, 0x6c, 0x55, 0x6e, 0x6d, 0x6f, 0x75, 0x6e, - 0x74, 0x28, 0x29, 0x7d, 0x63, 0x61, 0x74, 0x63, 0x68, 0x28, 0x74, 0x29, - 0x7b, 0x43, 0x2e, 0x5f, 0x5f, 0x65, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7d, - 0x5f, 0x2e, 0x62, 0x61, 0x73, 0x65, 0x3d, 0x5f, 0x2e, 0x5f, 0x5f, 0x50, - 0x3d, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x3d, - 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x7d, 0x69, 0x66, 0x28, 0x5f, 0x3d, - 0x74, 0x2e, 0x5f, 0x5f, 0x6b, 0x29, 0x66, 0x6f, 0x72, 0x28, 0x69, 0x3d, - 0x30, 0x3b, 0x69, 0x3c, 0x5f, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, - 0x3b, 0x69, 0x2b, 0x2b, 0x29, 0x5f, 0x5b, 0x69, 0x5d, 0x26, 0x26, 0x75, - 0x74, 0x28, 0x5f, 0x5b, 0x69, 0x5d, 0x2c, 0x6e, 0x2c, 0x65, 0x7c, 0x7c, - 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x21, 0x3d, - 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x74, 0x2e, 0x74, 0x79, 0x70, - 0x65, 0x29, 0x3b, 0x65, 0x7c, 0x7c, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, - 0x74, 0x2e, 0x5f, 0x5f, 0x65, 0x7c, 0x7c, 0x57, 0x28, 0x74, 0x2e, 0x5f, - 0x5f, 0x65, 0x29, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, 0x3d, 0x74, 0x2e, 0x5f, - 0x5f, 0x65, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x76, 0x6f, 0x69, - 0x64, 0x20, 0x30, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, - 0x20, 0x66, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x72, - 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x72, 0x75, 0x63, 0x74, 0x6f, 0x72, 0x28, 0x74, - 0x2c, 0x65, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, - 0x20, 0x73, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x76, - 0x61, 0x72, 0x20, 0x5f, 0x2c, 0x69, 0x2c, 0x6f, 0x2c, 0x72, 0x3b, 0x43, - 0x2e, 0x5f, 0x5f, 0x26, 0x26, 0x43, 0x2e, 0x5f, 0x5f, 0x28, 0x74, 0x2c, - 0x6e, 0x29, 0x2c, 0x69, 0x3d, 0x28, 0x5f, 0x3d, 0x22, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, - 0x6f, 0x66, 0x20, 0x65, 0x29, 0x3f, 0x6e, 0x75, 0x6c, 0x6c, 0x3a, 0x65, - 0x26, 0x26, 0x65, 0x2e, 0x5f, 0x5f, 0x6b, 0x7c, 0x7c, 0x6e, 0x2e, 0x5f, - 0x5f, 0x6b, 0x2c, 0x6f, 0x3d, 0x5b, 0x5d, 0x2c, 0x72, 0x3d, 0x5b, 0x5d, - 0x2c, 0x5f, 0x74, 0x28, 0x6e, 0x2c, 0x74, 0x3d, 0x28, 0x21, 0x5f, 0x26, - 0x26, 0x65, 0x7c, 0x7c, 0x6e, 0x29, 0x2e, 0x5f, 0x5f, 0x6b, 0x3d, 0x4c, - 0x28, 0x6a, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x5b, 0x74, 0x5d, 0x29, - 0x2c, 0x69, 0x7c, 0x7c, 0x54, 0x2c, 0x54, 0x2c, 0x76, 0x6f, 0x69, 0x64, - 0x20, 0x30, 0x21, 0x3d, 0x3d, 0x6e, 0x2e, 0x6f, 0x77, 0x6e, 0x65, 0x72, - 0x53, 0x56, 0x47, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x2c, 0x21, - 0x5f, 0x26, 0x26, 0x65, 0x3f, 0x5b, 0x65, 0x5d, 0x3a, 0x69, 0x3f, 0x6e, - 0x75, 0x6c, 0x6c, 0x3a, 0x6e, 0x2e, 0x66, 0x69, 0x72, 0x73, 0x74, 0x43, - 0x68, 0x69, 0x6c, 0x64, 0x3f, 0x78, 0x2e, 0x63, 0x61, 0x6c, 0x6c, 0x28, - 0x6e, 0x2e, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x4e, 0x6f, 0x64, 0x65, 0x73, - 0x29, 0x3a, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x6f, 0x2c, 0x21, 0x5f, 0x26, - 0x26, 0x65, 0x3f, 0x65, 0x3a, 0x69, 0x3f, 0x69, 0x2e, 0x5f, 0x5f, 0x65, - 0x3a, 0x6e, 0x2e, 0x66, 0x69, 0x72, 0x73, 0x74, 0x43, 0x68, 0x69, 0x6c, - 0x64, 0x2c, 0x5f, 0x2c, 0x72, 0x29, 0x2c, 0x69, 0x74, 0x28, 0x6f, 0x2c, + 0x20, 0x6f, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x66, + 0x6f, 0x72, 0x28, 0x76, 0x61, 0x72, 0x20, 0x5f, 0x3d, 0x30, 0x3b, 0x5f, + 0x3c, 0x65, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3b, 0x5f, 0x2b, + 0x2b, 0x29, 0x75, 0x74, 0x28, 0x65, 0x5b, 0x5f, 0x5d, 0x2c, 0x65, 0x5b, + 0x2b, 0x2b, 0x5f, 0x5d, 0x2c, 0x65, 0x5b, 0x2b, 0x2b, 0x5f, 0x5d, 0x29, + 0x3b, 0x43, 0x2e, 0x5f, 0x5f, 0x63, 0x26, 0x26, 0x43, 0x2e, 0x5f, 0x5f, + 0x63, 0x28, 0x6e, 0x2c, 0x74, 0x29, 0x2c, 0x74, 0x2e, 0x73, 0x6f, 0x6d, + 0x65, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, + 0x6e, 0x29, 0x7b, 0x74, 0x72, 0x79, 0x7b, 0x74, 0x3d, 0x6e, 0x2e, 0x5f, + 0x5f, 0x68, 0x2c, 0x6e, 0x2e, 0x5f, 0x5f, 0x68, 0x3d, 0x5b, 0x5d, 0x2c, + 0x74, 0x2e, 0x73, 0x6f, 0x6d, 0x65, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, + 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x74, 0x2e, 0x63, 0x61, + 0x6c, 0x6c, 0x28, 0x6e, 0x29, 0x7d, 0x29, 0x29, 0x7d, 0x63, 0x61, 0x74, + 0x63, 0x68, 0x28, 0x74, 0x29, 0x7b, 0x43, 0x2e, 0x5f, 0x5f, 0x65, 0x28, + 0x74, 0x2c, 0x6e, 0x2e, 0x5f, 0x5f, 0x76, 0x29, 0x7d, 0x7d, 0x29, 0x29, + 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x72, 0x74, + 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x2c, 0x5f, 0x2c, 0x69, 0x2c, 0x6f, + 0x2c, 0x72, 0x2c, 0x75, 0x2c, 0x66, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, + 0x6c, 0x2c, 0x73, 0x2c, 0x63, 0x2c, 0x68, 0x2c, 0x61, 0x2c, 0x70, 0x2c, + 0x64, 0x2c, 0x76, 0x3d, 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2c, + 0x79, 0x3d, 0x6e, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2c, 0x6d, 0x3d, + 0x6e, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x3b, 0x69, 0x66, 0x28, 0x22, 0x73, + 0x76, 0x67, 0x22, 0x3d, 0x3d, 0x3d, 0x6d, 0x26, 0x26, 0x28, 0x69, 0x3d, + 0x21, 0x30, 0x29, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x6f, 0x29, + 0x66, 0x6f, 0x72, 0x28, 0x6c, 0x3d, 0x30, 0x3b, 0x6c, 0x3c, 0x6f, 0x2e, + 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3b, 0x6c, 0x2b, 0x2b, 0x29, 0x69, + 0x66, 0x28, 0x28, 0x61, 0x3d, 0x6f, 0x5b, 0x6c, 0x5d, 0x29, 0x26, 0x26, + 0x22, 0x73, 0x65, 0x74, 0x41, 0x74, 0x74, 0x72, 0x69, 0x62, 0x75, 0x74, + 0x65, 0x22, 0x69, 0x6e, 0x20, 0x61, 0x3d, 0x3d, 0x21, 0x21, 0x6d, 0x26, + 0x26, 0x28, 0x6d, 0x3f, 0x61, 0x2e, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x4e, + 0x61, 0x6d, 0x65, 0x3d, 0x3d, 0x3d, 0x6d, 0x3a, 0x33, 0x3d, 0x3d, 0x3d, + 0x61, 0x2e, 0x6e, 0x6f, 0x64, 0x65, 0x54, 0x79, 0x70, 0x65, 0x29, 0x29, + 0x7b, 0x74, 0x3d, 0x61, 0x2c, 0x6f, 0x5b, 0x6c, 0x5d, 0x3d, 0x6e, 0x75, + 0x6c, 0x6c, 0x3b, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x7d, 0x69, 0x66, 0x28, + 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x74, 0x29, 0x7b, 0x69, 0x66, 0x28, + 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x3d, 0x6d, 0x29, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, + 0x2e, 0x63, 0x72, 0x65, 0x61, 0x74, 0x65, 0x54, 0x65, 0x78, 0x74, 0x4e, + 0x6f, 0x64, 0x65, 0x28, 0x79, 0x29, 0x3b, 0x74, 0x3d, 0x69, 0x3f, 0x64, + 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x63, 0x72, 0x65, 0x61, + 0x74, 0x65, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x4e, 0x53, 0x28, + 0x22, 0x68, 0x74, 0x74, 0x70, 0x3a, 0x2f, 0x2f, 0x77, 0x77, 0x77, 0x2e, + 0x77, 0x33, 0x2e, 0x6f, 0x72, 0x67, 0x2f, 0x32, 0x30, 0x30, 0x30, 0x2f, + 0x73, 0x76, 0x67, 0x22, 0x2c, 0x6d, 0x29, 0x3a, 0x64, 0x6f, 0x63, 0x75, + 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x63, 0x72, 0x65, 0x61, 0x74, 0x65, 0x45, + 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x28, 0x6d, 0x2c, 0x79, 0x2e, 0x69, + 0x73, 0x26, 0x26, 0x79, 0x29, 0x2c, 0x6f, 0x3d, 0x6e, 0x75, 0x6c, 0x6c, + 0x2c, 0x75, 0x3d, 0x21, 0x31, 0x7d, 0x69, 0x66, 0x28, 0x6e, 0x75, 0x6c, + 0x6c, 0x3d, 0x3d, 0x3d, 0x6d, 0x29, 0x76, 0x3d, 0x3d, 0x3d, 0x79, 0x7c, + 0x7c, 0x75, 0x26, 0x26, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x3d, 0x3d, + 0x3d, 0x79, 0x7c, 0x7c, 0x28, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x3d, + 0x79, 0x29, 0x3b, 0x65, 0x6c, 0x73, 0x65, 0x7b, 0x69, 0x66, 0x28, 0x6f, + 0x3d, 0x6f, 0x26, 0x26, 0x78, 0x2e, 0x63, 0x61, 0x6c, 0x6c, 0x28, 0x74, + 0x2e, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x4e, 0x6f, 0x64, 0x65, 0x73, 0x29, + 0x2c, 0x76, 0x3d, 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x7c, 0x7c, + 0x54, 0x2c, 0x21, 0x75, 0x26, 0x26, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, + 0x6f, 0x29, 0x66, 0x6f, 0x72, 0x28, 0x76, 0x3d, 0x7b, 0x7d, 0x2c, 0x6c, + 0x3d, 0x30, 0x3b, 0x6c, 0x3c, 0x74, 0x2e, 0x61, 0x74, 0x74, 0x72, 0x69, + 0x62, 0x75, 0x74, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, + 0x3b, 0x6c, 0x2b, 0x2b, 0x29, 0x76, 0x5b, 0x28, 0x61, 0x3d, 0x74, 0x2e, + 0x61, 0x74, 0x74, 0x72, 0x69, 0x62, 0x75, 0x74, 0x65, 0x73, 0x5b, 0x6c, + 0x5d, 0x29, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3d, 0x61, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x66, 0x6f, 0x72, 0x28, 0x6c, 0x20, 0x69, + 0x6e, 0x20, 0x76, 0x29, 0x61, 0x3d, 0x76, 0x5b, 0x6c, 0x5d, 0x2c, 0x22, + 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x22, 0x3d, 0x3d, 0x6c, + 0x7c, 0x7c, 0x28, 0x22, 0x64, 0x61, 0x6e, 0x67, 0x65, 0x72, 0x6f, 0x75, + 0x73, 0x6c, 0x79, 0x53, 0x65, 0x74, 0x49, 0x6e, 0x6e, 0x65, 0x72, 0x48, + 0x54, 0x4d, 0x4c, 0x22, 0x3d, 0x3d, 0x6c, 0x3f, 0x63, 0x3d, 0x61, 0x3a, + 0x22, 0x6b, 0x65, 0x79, 0x22, 0x3d, 0x3d, 0x3d, 0x6c, 0x7c, 0x7c, 0x6c, + 0x20, 0x69, 0x6e, 0x20, 0x79, 0x7c, 0x7c, 0x6e, 0x74, 0x28, 0x74, 0x2c, + 0x6c, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x61, 0x2c, 0x69, 0x29, 0x29, + 0x3b, 0x66, 0x6f, 0x72, 0x28, 0x6c, 0x20, 0x69, 0x6e, 0x20, 0x79, 0x29, + 0x61, 0x3d, 0x79, 0x5b, 0x6c, 0x5d, 0x2c, 0x22, 0x63, 0x68, 0x69, 0x6c, + 0x64, 0x72, 0x65, 0x6e, 0x22, 0x3d, 0x3d, 0x6c, 0x3f, 0x68, 0x3d, 0x61, + 0x3a, 0x22, 0x64, 0x61, 0x6e, 0x67, 0x65, 0x72, 0x6f, 0x75, 0x73, 0x6c, + 0x79, 0x53, 0x65, 0x74, 0x49, 0x6e, 0x6e, 0x65, 0x72, 0x48, 0x54, 0x4d, + 0x4c, 0x22, 0x3d, 0x3d, 0x6c, 0x3f, 0x73, 0x3d, 0x61, 0x3a, 0x22, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x22, 0x3d, 0x3d, 0x6c, 0x3f, 0x70, 0x3d, 0x61, + 0x3a, 0x22, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x22, 0x3d, 0x3d, + 0x6c, 0x3f, 0x64, 0x3d, 0x61, 0x3a, 0x22, 0x6b, 0x65, 0x79, 0x22, 0x3d, + 0x3d, 0x3d, 0x6c, 0x7c, 0x7c, 0x75, 0x26, 0x26, 0x22, 0x66, 0x75, 0x6e, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x21, 0x3d, 0x74, 0x79, 0x70, 0x65, + 0x6f, 0x66, 0x20, 0x61, 0x7c, 0x7c, 0x76, 0x5b, 0x6c, 0x5d, 0x3d, 0x3d, + 0x3d, 0x61, 0x7c, 0x7c, 0x6e, 0x74, 0x28, 0x74, 0x2c, 0x6c, 0x2c, 0x61, + 0x2c, 0x76, 0x5b, 0x6c, 0x5d, 0x2c, 0x69, 0x29, 0x3b, 0x69, 0x66, 0x28, + 0x73, 0x29, 0x75, 0x7c, 0x7c, 0x63, 0x26, 0x26, 0x28, 0x73, 0x2e, 0x5f, + 0x5f, 0x68, 0x74, 0x6d, 0x6c, 0x3d, 0x3d, 0x3d, 0x63, 0x2e, 0x5f, 0x5f, + 0x68, 0x74, 0x6d, 0x6c, 0x7c, 0x7c, 0x73, 0x2e, 0x5f, 0x5f, 0x68, 0x74, + 0x6d, 0x6c, 0x3d, 0x3d, 0x3d, 0x74, 0x2e, 0x69, 0x6e, 0x6e, 0x65, 0x72, + 0x48, 0x54, 0x4d, 0x4c, 0x29, 0x7c, 0x7c, 0x28, 0x74, 0x2e, 0x69, 0x6e, + 0x6e, 0x65, 0x72, 0x48, 0x54, 0x4d, 0x4c, 0x3d, 0x73, 0x2e, 0x5f, 0x5f, + 0x68, 0x74, 0x6d, 0x6c, 0x29, 0x2c, 0x6e, 0x2e, 0x5f, 0x5f, 0x6b, 0x3d, + 0x5b, 0x5d, 0x3b, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x28, 0x63, + 0x26, 0x26, 0x28, 0x74, 0x2e, 0x69, 0x6e, 0x6e, 0x65, 0x72, 0x48, 0x54, + 0x4d, 0x4c, 0x3d, 0x22, 0x22, 0x29, 0x2c, 0x4b, 0x28, 0x74, 0x2c, 0x46, + 0x28, 0x68, 0x29, 0x3f, 0x68, 0x3a, 0x5b, 0x68, 0x5d, 0x2c, 0x6e, 0x2c, + 0x65, 0x2c, 0x5f, 0x2c, 0x69, 0x26, 0x26, 0x22, 0x66, 0x6f, 0x72, 0x65, + 0x69, 0x67, 0x6e, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x22, 0x21, 0x3d, + 0x3d, 0x6d, 0x2c, 0x6f, 0x2c, 0x72, 0x2c, 0x6f, 0x3f, 0x6f, 0x5b, 0x30, + 0x5d, 0x3a, 0x65, 0x2e, 0x5f, 0x5f, 0x6b, 0x26, 0x26, 0x71, 0x28, 0x65, + 0x2c, 0x30, 0x29, 0x2c, 0x75, 0x2c, 0x66, 0x29, 0x2c, 0x6e, 0x75, 0x6c, + 0x6c, 0x21, 0x3d, 0x6f, 0x29, 0x66, 0x6f, 0x72, 0x28, 0x6c, 0x3d, 0x6f, + 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3b, 0x6c, 0x2d, 0x2d, 0x3b, + 0x29, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x6f, 0x5b, 0x6c, 0x5d, 0x26, + 0x26, 0x57, 0x28, 0x6f, 0x5b, 0x6c, 0x5d, 0x29, 0x3b, 0x75, 0x7c, 0x7c, + 0x28, 0x6c, 0x3d, 0x22, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x22, 0x2c, 0x76, + 0x6f, 0x69, 0x64, 0x20, 0x30, 0x21, 0x3d, 0x3d, 0x70, 0x26, 0x26, 0x28, + 0x70, 0x21, 0x3d, 0x3d, 0x74, 0x5b, 0x6c, 0x5d, 0x7c, 0x7c, 0x22, 0x70, + 0x72, 0x6f, 0x67, 0x72, 0x65, 0x73, 0x73, 0x22, 0x3d, 0x3d, 0x3d, 0x6d, + 0x26, 0x26, 0x21, 0x70, 0x7c, 0x7c, 0x22, 0x6f, 0x70, 0x74, 0x69, 0x6f, + 0x6e, 0x22, 0x3d, 0x3d, 0x3d, 0x6d, 0x26, 0x26, 0x70, 0x21, 0x3d, 0x3d, + 0x76, 0x5b, 0x6c, 0x5d, 0x29, 0x26, 0x26, 0x6e, 0x74, 0x28, 0x74, 0x2c, + 0x6c, 0x2c, 0x70, 0x2c, 0x76, 0x5b, 0x6c, 0x5d, 0x2c, 0x21, 0x31, 0x29, + 0x2c, 0x6c, 0x3d, 0x22, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x22, + 0x2c, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x21, 0x3d, 0x3d, 0x64, 0x26, + 0x26, 0x64, 0x21, 0x3d, 0x3d, 0x74, 0x5b, 0x6c, 0x5d, 0x26, 0x26, 0x6e, + 0x74, 0x28, 0x74, 0x2c, 0x6c, 0x2c, 0x64, 0x2c, 0x76, 0x5b, 0x6c, 0x5d, + 0x2c, 0x21, 0x31, 0x29, 0x29, 0x7d, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x20, 0x74, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, + 0x75, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x74, 0x72, + 0x79, 0x7b, 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, + 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x74, 0x3f, 0x74, + 0x28, 0x6e, 0x29, 0x3a, 0x74, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, + 0x74, 0x3d, 0x6e, 0x7d, 0x63, 0x61, 0x74, 0x63, 0x68, 0x28, 0x74, 0x29, + 0x7b, 0x43, 0x2e, 0x5f, 0x5f, 0x65, 0x28, 0x74, 0x2c, 0x65, 0x29, 0x7d, + 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x66, 0x74, + 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, + 0x5f, 0x2c, 0x69, 0x3b, 0x69, 0x66, 0x28, 0x43, 0x2e, 0x75, 0x6e, 0x6d, + 0x6f, 0x75, 0x6e, 0x74, 0x26, 0x26, 0x43, 0x2e, 0x75, 0x6e, 0x6d, 0x6f, + 0x75, 0x6e, 0x74, 0x28, 0x74, 0x29, 0x2c, 0x28, 0x5f, 0x3d, 0x74, 0x2e, + 0x72, 0x65, 0x66, 0x29, 0x26, 0x26, 0x28, 0x5f, 0x2e, 0x63, 0x75, 0x72, + 0x72, 0x65, 0x6e, 0x74, 0x26, 0x26, 0x5f, 0x2e, 0x63, 0x75, 0x72, 0x72, + 0x65, 0x6e, 0x74, 0x21, 0x3d, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x65, 0x7c, + 0x7c, 0x75, 0x74, 0x28, 0x5f, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x6e, + 0x29, 0x29, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x28, 0x5f, 0x3d, + 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x29, 0x29, 0x7b, 0x69, 0x66, 0x28, 0x5f, + 0x2e, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x57, 0x69, + 0x6c, 0x6c, 0x55, 0x6e, 0x6d, 0x6f, 0x75, 0x6e, 0x74, 0x29, 0x74, 0x72, + 0x79, 0x7b, 0x5f, 0x2e, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, + 0x74, 0x57, 0x69, 0x6c, 0x6c, 0x55, 0x6e, 0x6d, 0x6f, 0x75, 0x6e, 0x74, + 0x28, 0x29, 0x7d, 0x63, 0x61, 0x74, 0x63, 0x68, 0x28, 0x74, 0x29, 0x7b, + 0x43, 0x2e, 0x5f, 0x5f, 0x65, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7d, 0x5f, + 0x2e, 0x62, 0x61, 0x73, 0x65, 0x3d, 0x5f, 0x2e, 0x5f, 0x5f, 0x50, 0x3d, + 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x3d, 0x76, + 0x6f, 0x69, 0x64, 0x20, 0x30, 0x7d, 0x69, 0x66, 0x28, 0x5f, 0x3d, 0x74, + 0x2e, 0x5f, 0x5f, 0x6b, 0x29, 0x66, 0x6f, 0x72, 0x28, 0x69, 0x3d, 0x30, + 0x3b, 0x69, 0x3c, 0x5f, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3b, + 0x69, 0x2b, 0x2b, 0x29, 0x5f, 0x5b, 0x69, 0x5d, 0x26, 0x26, 0x66, 0x74, + 0x28, 0x5f, 0x5b, 0x69, 0x5d, 0x2c, 0x6e, 0x2c, 0x65, 0x7c, 0x7c, 0x22, + 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x21, 0x3d, 0x74, + 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x74, 0x2e, 0x74, 0x79, 0x70, 0x65, + 0x29, 0x3b, 0x65, 0x7c, 0x7c, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x74, + 0x2e, 0x5f, 0x5f, 0x65, 0x7c, 0x7c, 0x57, 0x28, 0x74, 0x2e, 0x5f, 0x5f, + 0x65, 0x29, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, + 0x65, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, 0x76, 0x6f, 0x69, 0x64, + 0x20, 0x30, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, + 0x6c, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x72, 0x65, + 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x72, 0x75, 0x63, 0x74, 0x6f, 0x72, 0x28, 0x74, 0x2c, + 0x65, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, + 0x73, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x76, 0x61, + 0x72, 0x20, 0x5f, 0x2c, 0x69, 0x2c, 0x6f, 0x2c, 0x72, 0x3b, 0x43, 0x2e, + 0x5f, 0x5f, 0x26, 0x26, 0x43, 0x2e, 0x5f, 0x5f, 0x28, 0x74, 0x2c, 0x6e, + 0x29, 0x2c, 0x69, 0x3d, 0x28, 0x5f, 0x3d, 0x22, 0x66, 0x75, 0x6e, 0x63, + 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, + 0x66, 0x20, 0x65, 0x29, 0x3f, 0x6e, 0x75, 0x6c, 0x6c, 0x3a, 0x65, 0x26, + 0x26, 0x65, 0x2e, 0x5f, 0x5f, 0x6b, 0x7c, 0x7c, 0x6e, 0x2e, 0x5f, 0x5f, + 0x6b, 0x2c, 0x6f, 0x3d, 0x5b, 0x5d, 0x2c, 0x72, 0x3d, 0x5b, 0x5d, 0x2c, + 0x69, 0x74, 0x28, 0x6e, 0x2c, 0x74, 0x3d, 0x28, 0x21, 0x5f, 0x26, 0x26, + 0x65, 0x7c, 0x7c, 0x6e, 0x29, 0x2e, 0x5f, 0x5f, 0x6b, 0x3d, 0x4c, 0x28, + 0x6a, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x5b, 0x74, 0x5d, 0x29, 0x2c, + 0x69, 0x7c, 0x7c, 0x54, 0x2c, 0x54, 0x2c, 0x76, 0x6f, 0x69, 0x64, 0x20, + 0x30, 0x21, 0x3d, 0x3d, 0x6e, 0x2e, 0x6f, 0x77, 0x6e, 0x65, 0x72, 0x53, + 0x56, 0x47, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x2c, 0x21, 0x5f, + 0x26, 0x26, 0x65, 0x3f, 0x5b, 0x65, 0x5d, 0x3a, 0x69, 0x3f, 0x6e, 0x75, + 0x6c, 0x6c, 0x3a, 0x6e, 0x2e, 0x66, 0x69, 0x72, 0x73, 0x74, 0x43, 0x68, + 0x69, 0x6c, 0x64, 0x3f, 0x78, 0x2e, 0x63, 0x61, 0x6c, 0x6c, 0x28, 0x6e, + 0x2e, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x4e, 0x6f, 0x64, 0x65, 0x73, 0x29, + 0x3a, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x6f, 0x2c, 0x21, 0x5f, 0x26, 0x26, + 0x65, 0x3f, 0x65, 0x3a, 0x69, 0x3f, 0x69, 0x2e, 0x5f, 0x5f, 0x65, 0x3a, + 0x6e, 0x2e, 0x66, 0x69, 0x72, 0x73, 0x74, 0x43, 0x68, 0x69, 0x6c, 0x64, + 0x2c, 0x5f, 0x2c, 0x72, 0x29, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, 0x64, 0x3d, + 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x2c, 0x6f, 0x74, 0x28, 0x6f, 0x2c, 0x74, 0x2c, 0x72, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, - 0x6e, 0x20, 0x6c, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x73, 0x74, - 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x6c, 0x74, 0x29, 0x7d, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x63, 0x74, 0x28, 0x74, 0x2c, 0x6e, + 0x6e, 0x20, 0x63, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x73, 0x74, + 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x63, 0x74, 0x29, 0x7d, 0x66, 0x75, 0x6e, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x68, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x5f, 0x2c, 0x69, 0x2c, 0x6f, 0x2c, 0x72, 0x2c, 0x75, 0x3d, 0x4d, 0x28, 0x7b, 0x7d, 0x2c, 0x74, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x3b, 0x66, 0x6f, 0x72, 0x28, @@ -1146,7 +1165,7 @@ unsigned char index_js[] = { 0x74, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x2c, 0x75, 0x2c, 0x5f, 0x7c, 0x7c, 0x74, 0x2e, 0x6b, 0x65, 0x79, 0x2c, 0x69, 0x7c, 0x7c, 0x74, 0x2e, 0x72, 0x65, 0x66, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x7d, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x68, 0x74, 0x28, 0x74, 0x2c, 0x6e, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x61, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x65, 0x3d, 0x7b, 0x5f, 0x5f, 0x63, 0x3a, 0x6e, 0x3d, 0x22, 0x5f, 0x5f, 0x63, 0x43, 0x22, 0x2b, 0x44, 0x2b, 0x2b, 0x2c, 0x5f, 0x5f, 0x3a, 0x74, 0x2c, 0x43, 0x6f, 0x6e, 0x73, 0x75, @@ -1171,7 +1190,7 @@ unsigned char index_js[] = { 0x21, 0x3d, 0x3d, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x26, 0x26, 0x65, 0x2e, 0x73, 0x6f, 0x6d, 0x65, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x74, 0x2e, 0x5f, 0x5f, - 0x65, 0x3d, 0x21, 0x30, 0x2c, 0x47, 0x28, 0x74, 0x29, 0x7d, 0x29, 0x29, + 0x65, 0x3d, 0x21, 0x30, 0x2c, 0x7a, 0x28, 0x74, 0x29, 0x7d, 0x29, 0x29, 0x7d, 0x2c, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x73, 0x75, 0x62, 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x65, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x74, 0x29, 0x3b, 0x76, 0x61, 0x72, @@ -1235,14 +1254,14 @@ unsigned char index_js[] = { 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x74, 0x26, 0x26, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x5f, 0x76, 0x26, 0x26, 0x28, 0x6e, 0x26, 0x26, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x73, 0x62, 0x2e, 0x70, 0x75, 0x73, 0x68, - 0x28, 0x6e, 0x29, 0x2c, 0x47, 0x28, 0x74, 0x68, 0x69, 0x73, 0x29, 0x29, + 0x28, 0x6e, 0x29, 0x2c, 0x7a, 0x28, 0x74, 0x68, 0x69, 0x73, 0x29, 0x29, 0x7d, 0x2c, 0x49, 0x2e, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x74, 0x79, 0x70, 0x65, 0x2e, 0x66, 0x6f, 0x72, 0x63, 0x65, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x5f, 0x76, 0x26, 0x26, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x5f, 0x65, 0x3d, 0x21, 0x30, 0x2c, 0x74, 0x26, 0x26, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x5f, 0x68, - 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x74, 0x29, 0x2c, 0x47, 0x28, 0x74, + 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x74, 0x29, 0x2c, 0x7a, 0x28, 0x74, 0x68, 0x69, 0x73, 0x29, 0x29, 0x7d, 0x2c, 0x49, 0x2e, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x74, 0x79, 0x70, 0x65, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x3d, 0x6a, 0x2c, 0x48, 0x3d, 0x5b, 0x5d, 0x2c, 0x4e, 0x3d, 0x22, @@ -1256,113 +1275,115 @@ unsigned char index_js[] = { 0x24, 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x2e, 0x5f, 0x5f, 0x76, 0x2e, 0x5f, 0x5f, 0x62, 0x2d, 0x6e, 0x2e, 0x5f, - 0x5f, 0x76, 0x2e, 0x5f, 0x5f, 0x62, 0x7d, 0x2c, 0x7a, 0x2e, 0x5f, 0x5f, + 0x5f, 0x76, 0x2e, 0x5f, 0x5f, 0x62, 0x7d, 0x2c, 0x4a, 0x2e, 0x5f, 0x5f, 0x72, 0x3d, 0x30, 0x2c, 0x44, 0x3d, 0x30, 0x3b, 0x76, 0x61, 0x72, 0x20, - 0x61, 0x74, 0x2c, 0x70, 0x74, 0x2c, 0x64, 0x74, 0x2c, 0x76, 0x74, 0x2c, - 0x79, 0x74, 0x3d, 0x30, 0x2c, 0x6d, 0x74, 0x3d, 0x5b, 0x5d, 0x2c, 0x67, - 0x74, 0x3d, 0x5b, 0x5d, 0x2c, 0x62, 0x74, 0x3d, 0x43, 0x2e, 0x5f, 0x5f, - 0x62, 0x2c, 0x6b, 0x74, 0x3d, 0x43, 0x2e, 0x5f, 0x5f, 0x72, 0x2c, 0x53, - 0x74, 0x3d, 0x43, 0x2e, 0x64, 0x69, 0x66, 0x66, 0x65, 0x64, 0x2c, 0x77, - 0x74, 0x3d, 0x43, 0x2e, 0x5f, 0x5f, 0x63, 0x2c, 0x78, 0x74, 0x3d, 0x43, - 0x2e, 0x75, 0x6e, 0x6d, 0x6f, 0x75, 0x6e, 0x74, 0x3b, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x43, 0x74, 0x28, 0x74, 0x2c, 0x6e, - 0x29, 0x7b, 0x43, 0x2e, 0x5f, 0x5f, 0x68, 0x26, 0x26, 0x43, 0x2e, 0x5f, - 0x5f, 0x68, 0x28, 0x70, 0x74, 0x2c, 0x74, 0x2c, 0x79, 0x74, 0x7c, 0x7c, - 0x6e, 0x29, 0x2c, 0x79, 0x74, 0x3d, 0x30, 0x3b, 0x76, 0x61, 0x72, 0x20, - 0x65, 0x3d, 0x70, 0x74, 0x2e, 0x5f, 0x5f, 0x48, 0x7c, 0x7c, 0x28, 0x70, - 0x74, 0x2e, 0x5f, 0x5f, 0x48, 0x3d, 0x7b, 0x5f, 0x5f, 0x3a, 0x5b, 0x5d, - 0x2c, 0x5f, 0x5f, 0x68, 0x3a, 0x5b, 0x5d, 0x7d, 0x29, 0x3b, 0x72, 0x65, - 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x3e, 0x3d, 0x65, 0x2e, 0x5f, 0x5f, - 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x26, 0x26, 0x65, 0x2e, 0x5f, - 0x5f, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x7b, 0x5f, 0x5f, 0x56, 0x3a, - 0x67, 0x74, 0x7d, 0x29, 0x2c, 0x65, 0x2e, 0x5f, 0x5f, 0x5b, 0x74, 0x5d, - 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x45, 0x74, - 0x28, 0x74, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x79, - 0x74, 0x3d, 0x31, 0x2c, 0x55, 0x74, 0x28, 0x71, 0x74, 0x2c, 0x74, 0x29, - 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x55, 0x74, - 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, - 0x5f, 0x3d, 0x43, 0x74, 0x28, 0x61, 0x74, 0x2b, 0x2b, 0x2c, 0x32, 0x29, - 0x3b, 0x69, 0x66, 0x28, 0x5f, 0x2e, 0x74, 0x3d, 0x74, 0x2c, 0x21, 0x5f, - 0x2e, 0x5f, 0x5f, 0x63, 0x26, 0x26, 0x28, 0x5f, 0x2e, 0x5f, 0x5f, 0x3d, - 0x5b, 0x65, 0x3f, 0x65, 0x28, 0x6e, 0x29, 0x3a, 0x71, 0x74, 0x28, 0x76, - 0x6f, 0x69, 0x64, 0x20, 0x30, 0x2c, 0x6e, 0x29, 0x2c, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x76, 0x61, 0x72, - 0x20, 0x6e, 0x3d, 0x5f, 0x2e, 0x5f, 0x5f, 0x4e, 0x3f, 0x5f, 0x2e, 0x5f, - 0x5f, 0x4e, 0x5b, 0x30, 0x5d, 0x3a, 0x5f, 0x2e, 0x5f, 0x5f, 0x5b, 0x30, - 0x5d, 0x2c, 0x65, 0x3d, 0x5f, 0x2e, 0x74, 0x28, 0x6e, 0x2c, 0x74, 0x29, - 0x3b, 0x6e, 0x21, 0x3d, 0x3d, 0x65, 0x26, 0x26, 0x28, 0x5f, 0x2e, 0x5f, - 0x5f, 0x4e, 0x3d, 0x5b, 0x65, 0x2c, 0x5f, 0x2e, 0x5f, 0x5f, 0x5b, 0x31, - 0x5d, 0x5d, 0x2c, 0x5f, 0x2e, 0x5f, 0x5f, 0x63, 0x2e, 0x73, 0x65, 0x74, - 0x53, 0x74, 0x61, 0x74, 0x65, 0x28, 0x7b, 0x7d, 0x29, 0x29, 0x7d, 0x5d, - 0x2c, 0x5f, 0x2e, 0x5f, 0x5f, 0x63, 0x3d, 0x70, 0x74, 0x2c, 0x21, 0x70, - 0x74, 0x2e, 0x75, 0x29, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x69, 0x3d, - 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x2c, 0x6e, - 0x2c, 0x65, 0x29, 0x7b, 0x69, 0x66, 0x28, 0x21, 0x5f, 0x2e, 0x5f, 0x5f, - 0x63, 0x2e, 0x5f, 0x5f, 0x48, 0x29, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, - 0x21, 0x30, 0x3b, 0x76, 0x61, 0x72, 0x20, 0x69, 0x3d, 0x5f, 0x2e, 0x5f, - 0x5f, 0x63, 0x2e, 0x5f, 0x5f, 0x48, 0x2e, 0x5f, 0x5f, 0x2e, 0x66, 0x69, - 0x6c, 0x74, 0x65, 0x72, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, - 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, - 0x20, 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x7d, 0x29, 0x29, 0x3b, 0x69, 0x66, - 0x28, 0x69, 0x2e, 0x65, 0x76, 0x65, 0x72, 0x79, 0x28, 0x28, 0x66, 0x75, - 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x72, 0x65, - 0x74, 0x75, 0x72, 0x6e, 0x21, 0x74, 0x2e, 0x5f, 0x5f, 0x4e, 0x7d, 0x29, - 0x29, 0x29, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x21, 0x6f, 0x7c, 0x7c, - 0x6f, 0x2e, 0x63, 0x61, 0x6c, 0x6c, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2c, - 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x3b, 0x76, 0x61, 0x72, 0x20, 0x72, - 0x3d, 0x21, 0x31, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x69, - 0x2e, 0x66, 0x6f, 0x72, 0x45, 0x61, 0x63, 0x68, 0x28, 0x28, 0x66, 0x75, - 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x69, 0x66, - 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x4e, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, - 0x6e, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x5b, 0x30, 0x5d, 0x3b, 0x74, 0x2e, - 0x5f, 0x5f, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x4e, 0x2c, 0x74, 0x2e, 0x5f, - 0x5f, 0x4e, 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x2c, 0x6e, 0x21, - 0x3d, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x5b, 0x30, 0x5d, 0x26, 0x26, 0x28, - 0x72, 0x3d, 0x21, 0x30, 0x29, 0x7d, 0x7d, 0x29, 0x29, 0x2c, 0x21, 0x28, - 0x21, 0x72, 0x26, 0x26, 0x5f, 0x2e, 0x5f, 0x5f, 0x63, 0x2e, 0x70, 0x72, - 0x6f, 0x70, 0x73, 0x3d, 0x3d, 0x3d, 0x74, 0x29, 0x26, 0x26, 0x28, 0x21, - 0x6f, 0x7c, 0x7c, 0x6f, 0x2e, 0x63, 0x61, 0x6c, 0x6c, 0x28, 0x74, 0x68, - 0x69, 0x73, 0x2c, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x29, 0x7d, 0x3b, - 0x70, 0x74, 0x2e, 0x75, 0x3d, 0x21, 0x30, 0x3b, 0x76, 0x61, 0x72, 0x20, - 0x6f, 0x3d, 0x70, 0x74, 0x2e, 0x73, 0x68, 0x6f, 0x75, 0x6c, 0x64, 0x43, - 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x55, 0x70, 0x64, 0x61, - 0x74, 0x65, 0x2c, 0x72, 0x3d, 0x70, 0x74, 0x2e, 0x63, 0x6f, 0x6d, 0x70, - 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x57, 0x69, 0x6c, 0x6c, 0x55, 0x70, 0x64, - 0x61, 0x74, 0x65, 0x3b, 0x70, 0x74, 0x2e, 0x63, 0x6f, 0x6d, 0x70, 0x6f, - 0x6e, 0x65, 0x6e, 0x74, 0x57, 0x69, 0x6c, 0x6c, 0x55, 0x70, 0x64, 0x61, - 0x74, 0x65, 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, - 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x69, 0x66, 0x28, 0x74, 0x68, - 0x69, 0x73, 0x2e, 0x5f, 0x5f, 0x65, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, - 0x5f, 0x3d, 0x6f, 0x3b, 0x6f, 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, - 0x2c, 0x69, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x2c, 0x6f, 0x3d, - 0x5f, 0x7d, 0x72, 0x26, 0x26, 0x72, 0x2e, 0x63, 0x61, 0x6c, 0x6c, 0x28, - 0x74, 0x68, 0x69, 0x73, 0x2c, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7d, - 0x2c, 0x70, 0x74, 0x2e, 0x73, 0x68, 0x6f, 0x75, 0x6c, 0x64, 0x43, 0x6f, - 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, - 0x65, 0x3d, 0x69, 0x7d, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x5f, - 0x2e, 0x5f, 0x5f, 0x4e, 0x7c, 0x7c, 0x5f, 0x2e, 0x5f, 0x5f, 0x7d, 0x66, - 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x48, 0x74, 0x28, 0x74, - 0x2c, 0x6e, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x65, 0x3d, 0x43, 0x74, - 0x28, 0x61, 0x74, 0x2b, 0x2b, 0x2c, 0x33, 0x29, 0x3b, 0x21, 0x43, 0x2e, - 0x5f, 0x5f, 0x73, 0x26, 0x26, 0x49, 0x74, 0x28, 0x65, 0x2e, 0x5f, 0x5f, - 0x48, 0x2c, 0x6e, 0x29, 0x26, 0x26, 0x28, 0x65, 0x2e, 0x5f, 0x5f, 0x3d, - 0x74, 0x2c, 0x65, 0x2e, 0x69, 0x3d, 0x6e, 0x2c, 0x70, 0x74, 0x2e, 0x5f, - 0x5f, 0x48, 0x2e, 0x5f, 0x5f, 0x68, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, - 0x65, 0x29, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, - 0x20, 0x50, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x76, 0x61, 0x72, - 0x20, 0x65, 0x3d, 0x43, 0x74, 0x28, 0x61, 0x74, 0x2b, 0x2b, 0x2c, 0x34, - 0x29, 0x3b, 0x21, 0x43, 0x2e, 0x5f, 0x5f, 0x73, 0x26, 0x26, 0x49, 0x74, + 0x70, 0x74, 0x2c, 0x64, 0x74, 0x2c, 0x76, 0x74, 0x2c, 0x79, 0x74, 0x2c, + 0x6d, 0x74, 0x3d, 0x30, 0x2c, 0x67, 0x74, 0x3d, 0x5b, 0x5d, 0x2c, 0x62, + 0x74, 0x3d, 0x5b, 0x5d, 0x2c, 0x6b, 0x74, 0x3d, 0x43, 0x2c, 0x53, 0x74, + 0x3d, 0x6b, 0x74, 0x2e, 0x5f, 0x5f, 0x62, 0x2c, 0x77, 0x74, 0x3d, 0x6b, + 0x74, 0x2e, 0x5f, 0x5f, 0x72, 0x2c, 0x78, 0x74, 0x3d, 0x6b, 0x74, 0x2e, + 0x64, 0x69, 0x66, 0x66, 0x65, 0x64, 0x2c, 0x43, 0x74, 0x3d, 0x6b, 0x74, + 0x2e, 0x5f, 0x5f, 0x63, 0x2c, 0x45, 0x74, 0x3d, 0x6b, 0x74, 0x2e, 0x75, + 0x6e, 0x6d, 0x6f, 0x75, 0x6e, 0x74, 0x2c, 0x55, 0x74, 0x3d, 0x6b, 0x74, + 0x2e, 0x5f, 0x5f, 0x3b, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, + 0x20, 0x48, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x6b, 0x74, 0x2e, + 0x5f, 0x5f, 0x68, 0x26, 0x26, 0x6b, 0x74, 0x2e, 0x5f, 0x5f, 0x68, 0x28, + 0x64, 0x74, 0x2c, 0x74, 0x2c, 0x6d, 0x74, 0x7c, 0x7c, 0x6e, 0x29, 0x2c, + 0x6d, 0x74, 0x3d, 0x30, 0x3b, 0x76, 0x61, 0x72, 0x20, 0x65, 0x3d, 0x64, + 0x74, 0x2e, 0x5f, 0x5f, 0x48, 0x7c, 0x7c, 0x28, 0x64, 0x74, 0x2e, 0x5f, + 0x5f, 0x48, 0x3d, 0x7b, 0x5f, 0x5f, 0x3a, 0x5b, 0x5d, 0x2c, 0x5f, 0x5f, + 0x68, 0x3a, 0x5b, 0x5d, 0x7d, 0x29, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, + 0x6e, 0x20, 0x74, 0x3e, 0x3d, 0x65, 0x2e, 0x5f, 0x5f, 0x2e, 0x6c, 0x65, + 0x6e, 0x67, 0x74, 0x68, 0x26, 0x26, 0x65, 0x2e, 0x5f, 0x5f, 0x2e, 0x70, + 0x75, 0x73, 0x68, 0x28, 0x7b, 0x5f, 0x5f, 0x56, 0x3a, 0x62, 0x74, 0x7d, + 0x29, 0x2c, 0x65, 0x2e, 0x5f, 0x5f, 0x5b, 0x74, 0x5d, 0x7d, 0x66, 0x75, + 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x50, 0x74, 0x28, 0x74, 0x29, + 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6d, 0x74, 0x3d, 0x31, + 0x2c, 0x4e, 0x74, 0x28, 0x7a, 0x74, 0x2c, 0x74, 0x29, 0x7d, 0x66, 0x75, + 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x4e, 0x74, 0x28, 0x74, 0x2c, + 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x5f, 0x3d, 0x48, + 0x74, 0x28, 0x70, 0x74, 0x2b, 0x2b, 0x2c, 0x32, 0x29, 0x3b, 0x69, 0x66, + 0x28, 0x5f, 0x2e, 0x74, 0x3d, 0x74, 0x2c, 0x21, 0x5f, 0x2e, 0x5f, 0x5f, + 0x63, 0x26, 0x26, 0x28, 0x5f, 0x2e, 0x5f, 0x5f, 0x3d, 0x5b, 0x65, 0x3f, + 0x65, 0x28, 0x6e, 0x29, 0x3a, 0x7a, 0x74, 0x28, 0x76, 0x6f, 0x69, 0x64, + 0x20, 0x30, 0x2c, 0x6e, 0x29, 0x2c, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, + 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, + 0x5f, 0x2e, 0x5f, 0x5f, 0x4e, 0x3f, 0x5f, 0x2e, 0x5f, 0x5f, 0x4e, 0x5b, + 0x30, 0x5d, 0x3a, 0x5f, 0x2e, 0x5f, 0x5f, 0x5b, 0x30, 0x5d, 0x2c, 0x65, + 0x3d, 0x5f, 0x2e, 0x74, 0x28, 0x6e, 0x2c, 0x74, 0x29, 0x3b, 0x6e, 0x21, + 0x3d, 0x3d, 0x65, 0x26, 0x26, 0x28, 0x5f, 0x2e, 0x5f, 0x5f, 0x4e, 0x3d, + 0x5b, 0x65, 0x2c, 0x5f, 0x2e, 0x5f, 0x5f, 0x5b, 0x31, 0x5d, 0x5d, 0x2c, + 0x5f, 0x2e, 0x5f, 0x5f, 0x63, 0x2e, 0x73, 0x65, 0x74, 0x53, 0x74, 0x61, + 0x74, 0x65, 0x28, 0x7b, 0x7d, 0x29, 0x29, 0x7d, 0x5d, 0x2c, 0x5f, 0x2e, + 0x5f, 0x5f, 0x63, 0x3d, 0x64, 0x74, 0x2c, 0x21, 0x64, 0x74, 0x2e, 0x75, + 0x29, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x69, 0x3d, 0x66, 0x75, 0x6e, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, + 0x7b, 0x69, 0x66, 0x28, 0x21, 0x5f, 0x2e, 0x5f, 0x5f, 0x63, 0x2e, 0x5f, + 0x5f, 0x48, 0x29, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x21, 0x30, 0x3b, + 0x76, 0x61, 0x72, 0x20, 0x69, 0x3d, 0x5f, 0x2e, 0x5f, 0x5f, 0x63, 0x2e, + 0x5f, 0x5f, 0x48, 0x2e, 0x5f, 0x5f, 0x2e, 0x66, 0x69, 0x6c, 0x74, 0x65, + 0x72, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, + 0x74, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x21, 0x21, 0x74, + 0x2e, 0x5f, 0x5f, 0x63, 0x7d, 0x29, 0x29, 0x3b, 0x69, 0x66, 0x28, 0x69, + 0x2e, 0x65, 0x76, 0x65, 0x72, 0x79, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, + 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x21, 0x74, 0x2e, 0x5f, 0x5f, 0x4e, 0x7d, 0x29, 0x29, 0x29, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x21, 0x6f, 0x7c, 0x7c, 0x6f, 0x2e, + 0x63, 0x61, 0x6c, 0x6c, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2c, 0x74, 0x2c, + 0x6e, 0x2c, 0x65, 0x29, 0x3b, 0x76, 0x61, 0x72, 0x20, 0x72, 0x3d, 0x21, + 0x31, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x69, 0x2e, 0x66, + 0x6f, 0x72, 0x45, 0x61, 0x63, 0x68, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, + 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x69, 0x66, 0x28, 0x74, + 0x2e, 0x5f, 0x5f, 0x4e, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, + 0x74, 0x2e, 0x5f, 0x5f, 0x5b, 0x30, 0x5d, 0x3b, 0x74, 0x2e, 0x5f, 0x5f, + 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x4e, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, 0x4e, + 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x2c, 0x6e, 0x21, 0x3d, 0x3d, + 0x74, 0x2e, 0x5f, 0x5f, 0x5b, 0x30, 0x5d, 0x26, 0x26, 0x28, 0x72, 0x3d, + 0x21, 0x30, 0x29, 0x7d, 0x7d, 0x29, 0x29, 0x2c, 0x21, 0x28, 0x21, 0x72, + 0x26, 0x26, 0x5f, 0x2e, 0x5f, 0x5f, 0x63, 0x2e, 0x70, 0x72, 0x6f, 0x70, + 0x73, 0x3d, 0x3d, 0x3d, 0x74, 0x29, 0x26, 0x26, 0x28, 0x21, 0x6f, 0x7c, + 0x7c, 0x6f, 0x2e, 0x63, 0x61, 0x6c, 0x6c, 0x28, 0x74, 0x68, 0x69, 0x73, + 0x2c, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x29, 0x7d, 0x3b, 0x64, 0x74, + 0x2e, 0x75, 0x3d, 0x21, 0x30, 0x3b, 0x76, 0x61, 0x72, 0x20, 0x6f, 0x3d, + 0x64, 0x74, 0x2e, 0x73, 0x68, 0x6f, 0x75, 0x6c, 0x64, 0x43, 0x6f, 0x6d, + 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, + 0x2c, 0x72, 0x3d, 0x64, 0x74, 0x2e, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, + 0x65, 0x6e, 0x74, 0x57, 0x69, 0x6c, 0x6c, 0x55, 0x70, 0x64, 0x61, 0x74, + 0x65, 0x3b, 0x64, 0x74, 0x2e, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, + 0x6e, 0x74, 0x57, 0x69, 0x6c, 0x6c, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, + 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x2c, + 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x69, 0x66, 0x28, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x5f, 0x5f, 0x65, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x5f, 0x3d, + 0x6f, 0x3b, 0x6f, 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x2c, 0x69, + 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x2c, 0x6f, 0x3d, 0x5f, 0x7d, + 0x72, 0x26, 0x26, 0x72, 0x2e, 0x63, 0x61, 0x6c, 0x6c, 0x28, 0x74, 0x68, + 0x69, 0x73, 0x2c, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7d, 0x2c, 0x64, + 0x74, 0x2e, 0x73, 0x68, 0x6f, 0x75, 0x6c, 0x64, 0x43, 0x6f, 0x6d, 0x70, + 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x3d, + 0x69, 0x7d, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x5f, 0x2e, 0x5f, + 0x5f, 0x4e, 0x7c, 0x7c, 0x5f, 0x2e, 0x5f, 0x5f, 0x7d, 0x66, 0x75, 0x6e, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x24, 0x74, 0x28, 0x74, 0x2c, 0x6e, + 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x65, 0x3d, 0x48, 0x74, 0x28, 0x70, + 0x74, 0x2b, 0x2b, 0x2c, 0x33, 0x29, 0x3b, 0x21, 0x6b, 0x74, 0x2e, 0x5f, + 0x5f, 0x73, 0x26, 0x26, 0x47, 0x74, 0x28, 0x65, 0x2e, 0x5f, 0x5f, 0x48, + 0x2c, 0x6e, 0x29, 0x26, 0x26, 0x28, 0x65, 0x2e, 0x5f, 0x5f, 0x3d, 0x74, + 0x2c, 0x65, 0x2e, 0x69, 0x3d, 0x6e, 0x2c, 0x64, 0x74, 0x2e, 0x5f, 0x5f, + 0x48, 0x2e, 0x5f, 0x5f, 0x68, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x65, + 0x29, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, + 0x44, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, + 0x65, 0x3d, 0x48, 0x74, 0x28, 0x70, 0x74, 0x2b, 0x2b, 0x2c, 0x34, 0x29, + 0x3b, 0x21, 0x6b, 0x74, 0x2e, 0x5f, 0x5f, 0x73, 0x26, 0x26, 0x47, 0x74, 0x28, 0x65, 0x2e, 0x5f, 0x5f, 0x48, 0x2c, 0x6e, 0x29, 0x26, 0x26, 0x28, 0x65, 0x2e, 0x5f, 0x5f, 0x3d, 0x74, 0x2c, 0x65, 0x2e, 0x69, 0x3d, 0x6e, - 0x2c, 0x70, 0x74, 0x2e, 0x5f, 0x5f, 0x68, 0x2e, 0x70, 0x75, 0x73, 0x68, + 0x2c, 0x64, 0x74, 0x2e, 0x5f, 0x5f, 0x68, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x65, 0x29, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, - 0x6e, 0x20, 0x4e, 0x74, 0x28, 0x74, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, - 0x72, 0x6e, 0x20, 0x79, 0x74, 0x3d, 0x35, 0x2c, 0x44, 0x74, 0x28, 0x28, + 0x6e, 0x20, 0x54, 0x74, 0x28, 0x74, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x20, 0x6d, 0x74, 0x3d, 0x35, 0x2c, 0x41, 0x74, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x7b, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x3a, 0x74, 0x7d, 0x7d, 0x29, 0x2c, 0x5b, 0x5d, 0x29, 0x7d, 0x66, - 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x24, 0x74, 0x28, 0x74, - 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x79, 0x74, 0x3d, 0x36, 0x2c, 0x50, + 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x56, 0x74, 0x28, 0x74, + 0x2c, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x6d, 0x74, 0x3d, 0x36, 0x2c, 0x44, 0x74, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, @@ -1377,197 +1398,204 @@ unsigned char index_js[] = { 0x64, 0x20, 0x30, 0x7d, 0x29, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x65, 0x3f, 0x65, 0x3a, 0x65, 0x2e, 0x63, 0x6f, 0x6e, 0x63, 0x61, 0x74, 0x28, 0x74, 0x29, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, - 0x6e, 0x20, 0x44, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x76, 0x61, - 0x72, 0x20, 0x65, 0x3d, 0x43, 0x74, 0x28, 0x61, 0x74, 0x2b, 0x2b, 0x2c, - 0x37, 0x29, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x49, 0x74, + 0x6e, 0x20, 0x41, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x76, 0x61, + 0x72, 0x20, 0x65, 0x3d, 0x48, 0x74, 0x28, 0x70, 0x74, 0x2b, 0x2b, 0x2c, + 0x37, 0x29, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x47, 0x74, 0x28, 0x65, 0x2e, 0x5f, 0x5f, 0x48, 0x2c, 0x6e, 0x29, 0x3f, 0x28, 0x65, 0x2e, 0x5f, 0x5f, 0x56, 0x3d, 0x74, 0x28, 0x29, 0x2c, 0x65, 0x2e, 0x69, 0x3d, 0x6e, 0x2c, 0x65, 0x2e, 0x5f, 0x5f, 0x68, 0x3d, 0x74, 0x2c, 0x65, 0x2e, 0x5f, 0x5f, 0x56, 0x29, 0x3a, 0x65, 0x2e, 0x5f, 0x5f, 0x7d, 0x66, - 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x54, 0x74, 0x28, 0x74, - 0x2c, 0x6e, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x79, - 0x74, 0x3d, 0x38, 0x2c, 0x44, 0x74, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, + 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x46, 0x74, 0x28, 0x74, + 0x2c, 0x6e, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6d, + 0x74, 0x3d, 0x38, 0x2c, 0x41, 0x74, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x7d, 0x29, 0x2c, 0x6e, 0x29, 0x7d, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x56, 0x74, 0x28, 0x74, 0x29, 0x7b, - 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, 0x70, 0x74, 0x2e, 0x63, 0x6f, 0x6e, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x4d, 0x74, 0x28, 0x74, 0x29, 0x7b, + 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, 0x64, 0x74, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x5b, 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x5d, 0x2c, - 0x65, 0x3d, 0x43, 0x74, 0x28, 0x61, 0x74, 0x2b, 0x2b, 0x2c, 0x39, 0x29, + 0x65, 0x3d, 0x48, 0x74, 0x28, 0x70, 0x74, 0x2b, 0x2b, 0x2c, 0x39, 0x29, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x65, 0x2e, 0x63, 0x3d, 0x74, 0x2c, 0x6e, 0x3f, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x3d, 0x3d, 0x65, 0x2e, 0x5f, 0x5f, 0x26, 0x26, 0x28, 0x65, 0x2e, 0x5f, 0x5f, 0x3d, 0x21, - 0x30, 0x2c, 0x6e, 0x2e, 0x73, 0x75, 0x62, 0x28, 0x70, 0x74, 0x29, 0x29, + 0x30, 0x2c, 0x6e, 0x2e, 0x73, 0x75, 0x62, 0x28, 0x64, 0x74, 0x29, 0x29, 0x2c, 0x6e, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x3a, 0x74, 0x2e, 0x5f, 0x5f, 0x7d, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x41, 0x74, 0x28, 0x74, 0x2c, 0x6e, - 0x29, 0x7b, 0x43, 0x2e, 0x75, 0x73, 0x65, 0x44, 0x65, 0x62, 0x75, 0x67, - 0x56, 0x61, 0x6c, 0x75, 0x65, 0x26, 0x26, 0x43, 0x2e, 0x75, 0x73, 0x65, - 0x44, 0x65, 0x62, 0x75, 0x67, 0x56, 0x61, 0x6c, 0x75, 0x65, 0x28, 0x6e, - 0x3f, 0x6e, 0x28, 0x74, 0x29, 0x3a, 0x74, 0x29, 0x7d, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x46, 0x74, 0x28, 0x74, 0x29, 0x7b, - 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, 0x43, 0x74, 0x28, 0x61, 0x74, 0x2b, - 0x2b, 0x2c, 0x31, 0x30, 0x29, 0x2c, 0x65, 0x3d, 0x45, 0x74, 0x28, 0x29, - 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x2e, 0x5f, 0x5f, - 0x3d, 0x74, 0x2c, 0x70, 0x74, 0x2e, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, - 0x65, 0x6e, 0x74, 0x44, 0x69, 0x64, 0x43, 0x61, 0x74, 0x63, 0x68, 0x7c, - 0x7c, 0x28, 0x70, 0x74, 0x2e, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, - 0x6e, 0x74, 0x44, 0x69, 0x64, 0x43, 0x61, 0x74, 0x63, 0x68, 0x3d, 0x66, - 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x2c, 0x5f, 0x29, - 0x7b, 0x6e, 0x2e, 0x5f, 0x5f, 0x26, 0x26, 0x6e, 0x2e, 0x5f, 0x5f, 0x28, - 0x74, 0x2c, 0x5f, 0x29, 0x2c, 0x65, 0x5b, 0x31, 0x5d, 0x28, 0x74, 0x29, - 0x7d, 0x29, 0x2c, 0x5b, 0x65, 0x5b, 0x30, 0x5d, 0x2c, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x29, 0x7b, 0x65, 0x5b, 0x31, 0x5d, - 0x28, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x29, 0x7d, 0x5d, 0x7d, 0x66, - 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x4d, 0x74, 0x28, 0x29, - 0x7b, 0x76, 0x61, 0x72, 0x20, 0x74, 0x3d, 0x43, 0x74, 0x28, 0x61, 0x74, - 0x2b, 0x2b, 0x2c, 0x31, 0x31, 0x29, 0x3b, 0x69, 0x66, 0x28, 0x21, 0x74, - 0x2e, 0x5f, 0x5f, 0x29, 0x7b, 0x66, 0x6f, 0x72, 0x28, 0x76, 0x61, 0x72, - 0x20, 0x6e, 0x3d, 0x70, 0x74, 0x2e, 0x5f, 0x5f, 0x76, 0x3b, 0x6e, 0x75, - 0x6c, 0x6c, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x21, 0x6e, 0x2e, 0x5f, - 0x5f, 0x6d, 0x26, 0x26, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x3d, 0x6e, - 0x2e, 0x5f, 0x5f, 0x3b, 0x29, 0x6e, 0x3d, 0x6e, 0x2e, 0x5f, 0x5f, 0x3b, - 0x76, 0x61, 0x72, 0x20, 0x65, 0x3d, 0x6e, 0x2e, 0x5f, 0x5f, 0x6d, 0x7c, - 0x7c, 0x28, 0x6e, 0x2e, 0x5f, 0x5f, 0x6d, 0x3d, 0x5b, 0x30, 0x2c, 0x30, - 0x5d, 0x29, 0x3b, 0x74, 0x2e, 0x5f, 0x5f, 0x3d, 0x22, 0x50, 0x22, 0x2b, - 0x65, 0x5b, 0x30, 0x5d, 0x2b, 0x22, 0x2d, 0x22, 0x2b, 0x65, 0x5b, 0x31, - 0x5d, 0x2b, 0x2b, 0x7d, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, - 0x2e, 0x5f, 0x5f, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, - 0x20, 0x57, 0x74, 0x28, 0x29, 0x7b, 0x66, 0x6f, 0x72, 0x28, 0x76, 0x61, - 0x72, 0x20, 0x74, 0x3b, 0x74, 0x3d, 0x6d, 0x74, 0x2e, 0x73, 0x68, 0x69, - 0x66, 0x74, 0x28, 0x29, 0x3b, 0x29, 0x69, 0x66, 0x28, 0x74, 0x2e, 0x5f, - 0x5f, 0x50, 0x26, 0x26, 0x74, 0x2e, 0x5f, 0x5f, 0x48, 0x29, 0x74, 0x72, - 0x79, 0x7b, 0x74, 0x2e, 0x5f, 0x5f, 0x48, 0x2e, 0x5f, 0x5f, 0x68, 0x2e, - 0x66, 0x6f, 0x72, 0x45, 0x61, 0x63, 0x68, 0x28, 0x52, 0x74, 0x29, 0x2c, - 0x74, 0x2e, 0x5f, 0x5f, 0x48, 0x2e, 0x5f, 0x5f, 0x68, 0x2e, 0x66, 0x6f, - 0x72, 0x45, 0x61, 0x63, 0x68, 0x28, 0x6a, 0x74, 0x29, 0x2c, 0x74, 0x2e, - 0x5f, 0x5f, 0x48, 0x2e, 0x5f, 0x5f, 0x68, 0x3d, 0x5b, 0x5d, 0x7d, 0x63, - 0x61, 0x74, 0x63, 0x68, 0x28, 0x75, 0x29, 0x7b, 0x74, 0x2e, 0x5f, 0x5f, - 0x48, 0x2e, 0x5f, 0x5f, 0x68, 0x3d, 0x5b, 0x5d, 0x2c, 0x43, 0x2e, 0x5f, - 0x5f, 0x65, 0x28, 0x75, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, 0x76, 0x29, 0x7d, - 0x7d, 0x43, 0x2e, 0x5f, 0x5f, 0x62, 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, - 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x70, 0x74, 0x3d, 0x6e, 0x75, - 0x6c, 0x6c, 0x2c, 0x62, 0x74, 0x26, 0x26, 0x62, 0x74, 0x28, 0x74, 0x29, - 0x7d, 0x2c, 0x43, 0x2e, 0x5f, 0x5f, 0x72, 0x3d, 0x66, 0x75, 0x6e, 0x63, - 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x6b, 0x74, 0x26, 0x26, - 0x6b, 0x74, 0x28, 0x74, 0x29, 0x2c, 0x61, 0x74, 0x3d, 0x30, 0x3b, 0x76, - 0x61, 0x72, 0x20, 0x6e, 0x3d, 0x28, 0x70, 0x74, 0x3d, 0x74, 0x2e, 0x5f, - 0x5f, 0x63, 0x29, 0x2e, 0x5f, 0x5f, 0x48, 0x3b, 0x6e, 0x26, 0x26, 0x28, - 0x64, 0x74, 0x3d, 0x3d, 0x3d, 0x70, 0x74, 0x3f, 0x28, 0x6e, 0x2e, 0x5f, - 0x5f, 0x68, 0x3d, 0x5b, 0x5d, 0x2c, 0x70, 0x74, 0x2e, 0x5f, 0x5f, 0x68, - 0x3d, 0x5b, 0x5d, 0x2c, 0x6e, 0x2e, 0x5f, 0x5f, 0x2e, 0x66, 0x6f, 0x72, - 0x45, 0x61, 0x63, 0x68, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, - 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x74, 0x2e, 0x5f, 0x5f, 0x4e, 0x26, - 0x26, 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x4e, - 0x29, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, 0x56, 0x3d, 0x67, 0x74, 0x2c, 0x74, - 0x2e, 0x5f, 0x5f, 0x4e, 0x3d, 0x74, 0x2e, 0x69, 0x3d, 0x76, 0x6f, 0x69, - 0x64, 0x20, 0x30, 0x7d, 0x29, 0x29, 0x29, 0x3a, 0x28, 0x6e, 0x2e, 0x5f, - 0x5f, 0x68, 0x2e, 0x66, 0x6f, 0x72, 0x45, 0x61, 0x63, 0x68, 0x28, 0x52, - 0x74, 0x29, 0x2c, 0x6e, 0x2e, 0x5f, 0x5f, 0x68, 0x2e, 0x66, 0x6f, 0x72, - 0x45, 0x61, 0x63, 0x68, 0x28, 0x6a, 0x74, 0x29, 0x2c, 0x6e, 0x2e, 0x5f, - 0x5f, 0x68, 0x3d, 0x5b, 0x5d, 0x2c, 0x61, 0x74, 0x3d, 0x30, 0x29, 0x29, - 0x2c, 0x64, 0x74, 0x3d, 0x70, 0x74, 0x7d, 0x2c, 0x43, 0x2e, 0x64, 0x69, - 0x66, 0x66, 0x65, 0x64, 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, - 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x53, 0x74, 0x26, 0x26, 0x53, 0x74, 0x28, - 0x74, 0x29, 0x3b, 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, 0x74, 0x2e, 0x5f, - 0x5f, 0x63, 0x3b, 0x6e, 0x26, 0x26, 0x6e, 0x2e, 0x5f, 0x5f, 0x48, 0x26, - 0x26, 0x28, 0x6e, 0x2e, 0x5f, 0x5f, 0x48, 0x2e, 0x5f, 0x5f, 0x68, 0x2e, - 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x26, 0x26, 0x28, 0x31, 0x21, 0x3d, - 0x3d, 0x6d, 0x74, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x6e, 0x29, 0x26, - 0x26, 0x76, 0x74, 0x3d, 0x3d, 0x3d, 0x43, 0x2e, 0x72, 0x65, 0x71, 0x75, - 0x65, 0x73, 0x74, 0x41, 0x6e, 0x69, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, - 0x46, 0x72, 0x61, 0x6d, 0x65, 0x7c, 0x7c, 0x28, 0x28, 0x76, 0x74, 0x3d, - 0x43, 0x2e, 0x72, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x41, 0x6e, 0x69, - 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x46, 0x72, 0x61, 0x6d, 0x65, 0x29, - 0x7c, 0x7c, 0x4f, 0x74, 0x29, 0x28, 0x57, 0x74, 0x29, 0x29, 0x2c, 0x6e, - 0x2e, 0x5f, 0x5f, 0x48, 0x2e, 0x5f, 0x5f, 0x2e, 0x66, 0x6f, 0x72, 0x45, - 0x61, 0x63, 0x68, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, - 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x74, 0x2e, 0x69, 0x26, 0x26, 0x28, 0x74, - 0x2e, 0x5f, 0x5f, 0x48, 0x3d, 0x74, 0x2e, 0x69, 0x29, 0x2c, 0x74, 0x2e, - 0x5f, 0x5f, 0x56, 0x21, 0x3d, 0x3d, 0x67, 0x74, 0x26, 0x26, 0x28, 0x74, - 0x2e, 0x5f, 0x5f, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x56, 0x29, 0x2c, 0x74, - 0x2e, 0x69, 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x2c, 0x74, 0x2e, - 0x5f, 0x5f, 0x56, 0x3d, 0x67, 0x74, 0x7d, 0x29, 0x29, 0x29, 0x2c, 0x64, - 0x74, 0x3d, 0x70, 0x74, 0x3d, 0x6e, 0x75, 0x6c, 0x6c, 0x7d, 0x2c, 0x43, - 0x2e, 0x5f, 0x5f, 0x63, 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, - 0x6e, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x6e, 0x2e, 0x73, 0x6f, 0x6d, - 0x65, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, - 0x74, 0x29, 0x7b, 0x74, 0x72, 0x79, 0x7b, 0x74, 0x2e, 0x5f, 0x5f, 0x68, - 0x2e, 0x66, 0x6f, 0x72, 0x45, 0x61, 0x63, 0x68, 0x28, 0x52, 0x74, 0x29, - 0x2c, 0x74, 0x2e, 0x5f, 0x5f, 0x68, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x68, - 0x2e, 0x66, 0x69, 0x6c, 0x74, 0x65, 0x72, 0x28, 0x28, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x72, 0x65, 0x74, - 0x75, 0x72, 0x6e, 0x21, 0x74, 0x2e, 0x5f, 0x5f, 0x7c, 0x7c, 0x6a, 0x74, - 0x28, 0x74, 0x29, 0x7d, 0x29, 0x29, 0x7d, 0x63, 0x61, 0x74, 0x63, 0x68, - 0x28, 0x6c, 0x29, 0x7b, 0x6e, 0x2e, 0x73, 0x6f, 0x6d, 0x65, 0x28, 0x28, - 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, - 0x74, 0x2e, 0x5f, 0x5f, 0x68, 0x26, 0x26, 0x28, 0x74, 0x2e, 0x5f, 0x5f, - 0x68, 0x3d, 0x5b, 0x5d, 0x29, 0x7d, 0x29, 0x29, 0x2c, 0x6e, 0x3d, 0x5b, - 0x5d, 0x2c, 0x43, 0x2e, 0x5f, 0x5f, 0x65, 0x28, 0x6c, 0x2c, 0x74, 0x2e, - 0x5f, 0x5f, 0x76, 0x29, 0x7d, 0x7d, 0x29, 0x29, 0x2c, 0x77, 0x74, 0x26, - 0x26, 0x77, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7d, 0x2c, 0x43, 0x2e, - 0x75, 0x6e, 0x6d, 0x6f, 0x75, 0x6e, 0x74, 0x3d, 0x66, 0x75, 0x6e, 0x63, - 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x78, 0x74, 0x26, 0x26, - 0x78, 0x74, 0x28, 0x74, 0x29, 0x3b, 0x76, 0x61, 0x72, 0x20, 0x6e, 0x2c, - 0x65, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x3b, 0x65, 0x26, 0x26, 0x65, - 0x2e, 0x5f, 0x5f, 0x48, 0x26, 0x26, 0x28, 0x65, 0x2e, 0x5f, 0x5f, 0x48, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x57, 0x74, 0x28, 0x74, 0x2c, 0x6e, + 0x29, 0x7b, 0x6b, 0x74, 0x2e, 0x75, 0x73, 0x65, 0x44, 0x65, 0x62, 0x75, + 0x67, 0x56, 0x61, 0x6c, 0x75, 0x65, 0x26, 0x26, 0x6b, 0x74, 0x2e, 0x75, + 0x73, 0x65, 0x44, 0x65, 0x62, 0x75, 0x67, 0x56, 0x61, 0x6c, 0x75, 0x65, + 0x28, 0x6e, 0x3f, 0x6e, 0x28, 0x74, 0x29, 0x3a, 0x74, 0x29, 0x7d, 0x66, + 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x4c, 0x74, 0x28, 0x74, + 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, 0x48, 0x74, 0x28, 0x70, + 0x74, 0x2b, 0x2b, 0x2c, 0x31, 0x30, 0x29, 0x2c, 0x65, 0x3d, 0x50, 0x74, + 0x28, 0x29, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x2e, + 0x5f, 0x5f, 0x3d, 0x74, 0x2c, 0x64, 0x74, 0x2e, 0x63, 0x6f, 0x6d, 0x70, + 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x44, 0x69, 0x64, 0x43, 0x61, 0x74, 0x63, + 0x68, 0x7c, 0x7c, 0x28, 0x64, 0x74, 0x2e, 0x63, 0x6f, 0x6d, 0x70, 0x6f, + 0x6e, 0x65, 0x6e, 0x74, 0x44, 0x69, 0x64, 0x43, 0x61, 0x74, 0x63, 0x68, + 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x2c, + 0x5f, 0x29, 0x7b, 0x6e, 0x2e, 0x5f, 0x5f, 0x26, 0x26, 0x6e, 0x2e, 0x5f, + 0x5f, 0x28, 0x74, 0x2c, 0x5f, 0x29, 0x2c, 0x65, 0x5b, 0x31, 0x5d, 0x28, + 0x74, 0x29, 0x7d, 0x29, 0x2c, 0x5b, 0x65, 0x5b, 0x30, 0x5d, 0x2c, 0x66, + 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x29, 0x7b, 0x65, 0x5b, + 0x31, 0x5d, 0x28, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x29, 0x7d, 0x5d, + 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x4f, 0x74, + 0x28, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x74, 0x3d, 0x48, 0x74, 0x28, + 0x70, 0x74, 0x2b, 0x2b, 0x2c, 0x31, 0x31, 0x29, 0x3b, 0x69, 0x66, 0x28, + 0x21, 0x74, 0x2e, 0x5f, 0x5f, 0x29, 0x7b, 0x66, 0x6f, 0x72, 0x28, 0x76, + 0x61, 0x72, 0x20, 0x6e, 0x3d, 0x64, 0x74, 0x2e, 0x5f, 0x5f, 0x76, 0x3b, + 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x21, 0x6e, + 0x2e, 0x5f, 0x5f, 0x6d, 0x26, 0x26, 0x6e, 0x75, 0x6c, 0x6c, 0x21, 0x3d, + 0x3d, 0x6e, 0x2e, 0x5f, 0x5f, 0x3b, 0x29, 0x6e, 0x3d, 0x6e, 0x2e, 0x5f, + 0x5f, 0x3b, 0x76, 0x61, 0x72, 0x20, 0x65, 0x3d, 0x6e, 0x2e, 0x5f, 0x5f, + 0x6d, 0x7c, 0x7c, 0x28, 0x6e, 0x2e, 0x5f, 0x5f, 0x6d, 0x3d, 0x5b, 0x30, + 0x2c, 0x30, 0x5d, 0x29, 0x3b, 0x74, 0x2e, 0x5f, 0x5f, 0x3d, 0x22, 0x50, + 0x22, 0x2b, 0x65, 0x5b, 0x30, 0x5d, 0x2b, 0x22, 0x2d, 0x22, 0x2b, 0x65, + 0x5b, 0x31, 0x5d, 0x2b, 0x2b, 0x7d, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x20, 0x74, 0x2e, 0x5f, 0x5f, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, + 0x6f, 0x6e, 0x20, 0x52, 0x74, 0x28, 0x29, 0x7b, 0x66, 0x6f, 0x72, 0x28, + 0x76, 0x61, 0x72, 0x20, 0x74, 0x3b, 0x74, 0x3d, 0x67, 0x74, 0x2e, 0x73, + 0x68, 0x69, 0x66, 0x74, 0x28, 0x29, 0x3b, 0x29, 0x69, 0x66, 0x28, 0x74, + 0x2e, 0x5f, 0x5f, 0x50, 0x26, 0x26, 0x74, 0x2e, 0x5f, 0x5f, 0x48, 0x29, + 0x74, 0x72, 0x79, 0x7b, 0x74, 0x2e, 0x5f, 0x5f, 0x48, 0x2e, 0x5f, 0x5f, + 0x68, 0x2e, 0x66, 0x6f, 0x72, 0x45, 0x61, 0x63, 0x68, 0x28, 0x71, 0x74, + 0x29, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, 0x48, 0x2e, 0x5f, 0x5f, 0x68, 0x2e, + 0x66, 0x6f, 0x72, 0x45, 0x61, 0x63, 0x68, 0x28, 0x42, 0x74, 0x29, 0x2c, + 0x74, 0x2e, 0x5f, 0x5f, 0x48, 0x2e, 0x5f, 0x5f, 0x68, 0x3d, 0x5b, 0x5d, + 0x7d, 0x63, 0x61, 0x74, 0x63, 0x68, 0x28, 0x6e, 0x29, 0x7b, 0x74, 0x2e, + 0x5f, 0x5f, 0x48, 0x2e, 0x5f, 0x5f, 0x68, 0x3d, 0x5b, 0x5d, 0x2c, 0x6b, + 0x74, 0x2e, 0x5f, 0x5f, 0x65, 0x28, 0x6e, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, + 0x76, 0x29, 0x7d, 0x7d, 0x6b, 0x74, 0x2e, 0x5f, 0x5f, 0x62, 0x3d, 0x66, + 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x64, + 0x74, 0x3d, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x53, 0x74, 0x26, 0x26, 0x53, + 0x74, 0x28, 0x74, 0x29, 0x7d, 0x2c, 0x6b, 0x74, 0x2e, 0x5f, 0x5f, 0x3d, + 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x2c, 0x6e, + 0x29, 0x7b, 0x74, 0x26, 0x26, 0x6e, 0x2e, 0x5f, 0x5f, 0x6b, 0x26, 0x26, + 0x6e, 0x2e, 0x5f, 0x5f, 0x6b, 0x2e, 0x5f, 0x5f, 0x6d, 0x26, 0x26, 0x28, + 0x74, 0x2e, 0x5f, 0x5f, 0x6d, 0x3d, 0x6e, 0x2e, 0x5f, 0x5f, 0x6b, 0x2e, + 0x5f, 0x5f, 0x6d, 0x29, 0x2c, 0x55, 0x74, 0x26, 0x26, 0x55, 0x74, 0x28, + 0x74, 0x2c, 0x6e, 0x29, 0x7d, 0x2c, 0x6b, 0x74, 0x2e, 0x5f, 0x5f, 0x72, + 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, + 0x7b, 0x77, 0x74, 0x26, 0x26, 0x77, 0x74, 0x28, 0x74, 0x29, 0x2c, 0x70, + 0x74, 0x3d, 0x30, 0x3b, 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, 0x28, 0x64, + 0x74, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x29, 0x2e, 0x5f, 0x5f, 0x48, + 0x3b, 0x6e, 0x26, 0x26, 0x28, 0x76, 0x74, 0x3d, 0x3d, 0x3d, 0x64, 0x74, + 0x3f, 0x28, 0x6e, 0x2e, 0x5f, 0x5f, 0x68, 0x3d, 0x5b, 0x5d, 0x2c, 0x64, + 0x74, 0x2e, 0x5f, 0x5f, 0x68, 0x3d, 0x5b, 0x5d, 0x2c, 0x6e, 0x2e, 0x5f, + 0x5f, 0x2e, 0x66, 0x6f, 0x72, 0x45, 0x61, 0x63, 0x68, 0x28, 0x28, 0x66, + 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x74, + 0x2e, 0x5f, 0x5f, 0x4e, 0x26, 0x26, 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x3d, + 0x74, 0x2e, 0x5f, 0x5f, 0x4e, 0x29, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, 0x56, + 0x3d, 0x62, 0x74, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, 0x4e, 0x3d, 0x74, 0x2e, + 0x69, 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x7d, 0x29, 0x29, 0x29, + 0x3a, 0x28, 0x6e, 0x2e, 0x5f, 0x5f, 0x68, 0x2e, 0x66, 0x6f, 0x72, 0x45, + 0x61, 0x63, 0x68, 0x28, 0x71, 0x74, 0x29, 0x2c, 0x6e, 0x2e, 0x5f, 0x5f, + 0x68, 0x2e, 0x66, 0x6f, 0x72, 0x45, 0x61, 0x63, 0x68, 0x28, 0x42, 0x74, + 0x29, 0x2c, 0x6e, 0x2e, 0x5f, 0x5f, 0x68, 0x3d, 0x5b, 0x5d, 0x2c, 0x70, + 0x74, 0x3d, 0x30, 0x29, 0x29, 0x2c, 0x76, 0x74, 0x3d, 0x64, 0x74, 0x7d, + 0x2c, 0x6b, 0x74, 0x2e, 0x64, 0x69, 0x66, 0x66, 0x65, 0x64, 0x3d, 0x66, + 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x78, + 0x74, 0x26, 0x26, 0x78, 0x74, 0x28, 0x74, 0x29, 0x3b, 0x76, 0x61, 0x72, + 0x20, 0x6e, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x3b, 0x6e, 0x26, 0x26, + 0x6e, 0x2e, 0x5f, 0x5f, 0x48, 0x26, 0x26, 0x28, 0x6e, 0x2e, 0x5f, 0x5f, + 0x48, 0x2e, 0x5f, 0x5f, 0x68, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, + 0x26, 0x26, 0x28, 0x31, 0x21, 0x3d, 0x3d, 0x67, 0x74, 0x2e, 0x70, 0x75, + 0x73, 0x68, 0x28, 0x6e, 0x29, 0x26, 0x26, 0x79, 0x74, 0x3d, 0x3d, 0x3d, + 0x6b, 0x74, 0x2e, 0x72, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x41, 0x6e, + 0x69, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x46, 0x72, 0x61, 0x6d, 0x65, + 0x7c, 0x7c, 0x28, 0x28, 0x79, 0x74, 0x3d, 0x6b, 0x74, 0x2e, 0x72, 0x65, + 0x71, 0x75, 0x65, 0x73, 0x74, 0x41, 0x6e, 0x69, 0x6d, 0x61, 0x74, 0x69, + 0x6f, 0x6e, 0x46, 0x72, 0x61, 0x6d, 0x65, 0x29, 0x7c, 0x7c, 0x49, 0x74, + 0x29, 0x28, 0x52, 0x74, 0x29, 0x29, 0x2c, 0x6e, 0x2e, 0x5f, 0x5f, 0x48, 0x2e, 0x5f, 0x5f, 0x2e, 0x66, 0x6f, 0x72, 0x45, 0x61, 0x63, 0x68, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, - 0x7b, 0x74, 0x72, 0x79, 0x7b, 0x52, 0x74, 0x28, 0x74, 0x29, 0x7d, 0x63, - 0x61, 0x74, 0x63, 0x68, 0x28, 0x74, 0x29, 0x7b, 0x6e, 0x3d, 0x74, 0x7d, - 0x7d, 0x29, 0x29, 0x2c, 0x65, 0x2e, 0x5f, 0x5f, 0x48, 0x3d, 0x76, 0x6f, - 0x69, 0x64, 0x20, 0x30, 0x2c, 0x6e, 0x26, 0x26, 0x43, 0x2e, 0x5f, 0x5f, + 0x7b, 0x74, 0x2e, 0x69, 0x26, 0x26, 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x48, + 0x3d, 0x74, 0x2e, 0x69, 0x29, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, 0x56, 0x21, + 0x3d, 0x3d, 0x62, 0x74, 0x26, 0x26, 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x3d, + 0x74, 0x2e, 0x5f, 0x5f, 0x56, 0x29, 0x2c, 0x74, 0x2e, 0x69, 0x3d, 0x76, + 0x6f, 0x69, 0x64, 0x20, 0x30, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, 0x56, 0x3d, + 0x62, 0x74, 0x7d, 0x29, 0x29, 0x29, 0x2c, 0x76, 0x74, 0x3d, 0x64, 0x74, + 0x3d, 0x6e, 0x75, 0x6c, 0x6c, 0x7d, 0x2c, 0x6b, 0x74, 0x2e, 0x5f, 0x5f, + 0x63, 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, + 0x2c, 0x6e, 0x29, 0x7b, 0x6e, 0x2e, 0x73, 0x6f, 0x6d, 0x65, 0x28, 0x28, + 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, + 0x74, 0x72, 0x79, 0x7b, 0x74, 0x2e, 0x5f, 0x5f, 0x68, 0x2e, 0x66, 0x6f, + 0x72, 0x45, 0x61, 0x63, 0x68, 0x28, 0x71, 0x74, 0x29, 0x2c, 0x74, 0x2e, + 0x5f, 0x5f, 0x68, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x68, 0x2e, 0x66, 0x69, + 0x6c, 0x74, 0x65, 0x72, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, + 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x21, 0x74, 0x2e, 0x5f, 0x5f, 0x7c, 0x7c, 0x42, 0x74, 0x28, 0x74, 0x29, + 0x7d, 0x29, 0x29, 0x7d, 0x63, 0x61, 0x74, 0x63, 0x68, 0x28, 0x75, 0x29, + 0x7b, 0x6e, 0x2e, 0x73, 0x6f, 0x6d, 0x65, 0x28, 0x28, 0x66, 0x75, 0x6e, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x74, 0x2e, 0x5f, + 0x5f, 0x68, 0x26, 0x26, 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x68, 0x3d, 0x5b, + 0x5d, 0x29, 0x7d, 0x29, 0x29, 0x2c, 0x6e, 0x3d, 0x5b, 0x5d, 0x2c, 0x6b, + 0x74, 0x2e, 0x5f, 0x5f, 0x65, 0x28, 0x75, 0x2c, 0x74, 0x2e, 0x5f, 0x5f, + 0x76, 0x29, 0x7d, 0x7d, 0x29, 0x29, 0x2c, 0x43, 0x74, 0x26, 0x26, 0x43, + 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7d, 0x2c, 0x6b, 0x74, 0x2e, 0x75, + 0x6e, 0x6d, 0x6f, 0x75, 0x6e, 0x74, 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, + 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x45, 0x74, 0x26, 0x26, 0x45, + 0x74, 0x28, 0x74, 0x29, 0x3b, 0x76, 0x61, 0x72, 0x20, 0x6e, 0x2c, 0x65, + 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x3b, 0x65, 0x26, 0x26, 0x65, 0x2e, + 0x5f, 0x5f, 0x48, 0x26, 0x26, 0x28, 0x65, 0x2e, 0x5f, 0x5f, 0x48, 0x2e, + 0x5f, 0x5f, 0x2e, 0x66, 0x6f, 0x72, 0x45, 0x61, 0x63, 0x68, 0x28, 0x28, + 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, + 0x74, 0x72, 0x79, 0x7b, 0x71, 0x74, 0x28, 0x74, 0x29, 0x7d, 0x63, 0x61, + 0x74, 0x63, 0x68, 0x28, 0x74, 0x29, 0x7b, 0x6e, 0x3d, 0x74, 0x7d, 0x7d, + 0x29, 0x29, 0x2c, 0x65, 0x2e, 0x5f, 0x5f, 0x48, 0x3d, 0x76, 0x6f, 0x69, + 0x64, 0x20, 0x30, 0x2c, 0x6e, 0x26, 0x26, 0x6b, 0x74, 0x2e, 0x5f, 0x5f, 0x65, 0x28, 0x6e, 0x2c, 0x65, 0x2e, 0x5f, 0x5f, 0x76, 0x29, 0x29, 0x7d, - 0x3b, 0x76, 0x61, 0x72, 0x20, 0x4c, 0x74, 0x3d, 0x22, 0x66, 0x75, 0x6e, + 0x3b, 0x76, 0x61, 0x72, 0x20, 0x6a, 0x74, 0x3d, 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x72, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x41, 0x6e, 0x69, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x46, 0x72, 0x61, 0x6d, 0x65, - 0x3b, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x4f, 0x74, + 0x3b, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x49, 0x74, 0x28, 0x74, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x6e, 0x2c, 0x65, 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x29, 0x7b, 0x63, 0x6c, 0x65, 0x61, 0x72, 0x54, 0x69, 0x6d, 0x65, 0x6f, 0x75, 0x74, 0x28, - 0x5f, 0x29, 0x2c, 0x4c, 0x74, 0x26, 0x26, 0x63, 0x61, 0x6e, 0x63, 0x65, + 0x5f, 0x29, 0x2c, 0x6a, 0x74, 0x26, 0x26, 0x63, 0x61, 0x6e, 0x63, 0x65, 0x6c, 0x41, 0x6e, 0x69, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x46, 0x72, 0x61, 0x6d, 0x65, 0x28, 0x6e, 0x29, 0x2c, 0x73, 0x65, 0x74, 0x54, 0x69, 0x6d, 0x65, 0x6f, 0x75, 0x74, 0x28, 0x74, 0x29, 0x7d, 0x2c, 0x5f, 0x3d, 0x73, 0x65, 0x74, 0x54, 0x69, 0x6d, 0x65, 0x6f, 0x75, 0x74, 0x28, 0x65, - 0x2c, 0x31, 0x30, 0x30, 0x29, 0x3b, 0x4c, 0x74, 0x26, 0x26, 0x28, 0x6e, + 0x2c, 0x31, 0x30, 0x30, 0x29, 0x3b, 0x6a, 0x74, 0x26, 0x26, 0x28, 0x6e, 0x3d, 0x72, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x41, 0x6e, 0x69, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x46, 0x72, 0x61, 0x6d, 0x65, 0x28, 0x65, 0x29, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, - 0x52, 0x74, 0x28, 0x74, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, - 0x70, 0x74, 0x2c, 0x65, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x3b, 0x22, + 0x71, 0x74, 0x28, 0x74, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, + 0x64, 0x74, 0x2c, 0x65, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x3b, 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x65, 0x26, 0x26, 0x28, 0x74, 0x2e, 0x5f, 0x5f, 0x63, 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x2c, 0x65, - 0x28, 0x29, 0x29, 0x2c, 0x70, 0x74, 0x3d, 0x6e, 0x7d, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6a, 0x74, 0x28, 0x74, 0x29, 0x7b, - 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, 0x70, 0x74, 0x3b, 0x74, 0x2e, 0x5f, - 0x5f, 0x63, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x28, 0x29, 0x2c, 0x70, 0x74, + 0x28, 0x29, 0x29, 0x2c, 0x64, 0x74, 0x3d, 0x6e, 0x7d, 0x66, 0x75, 0x6e, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x42, 0x74, 0x28, 0x74, 0x29, 0x7b, + 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, 0x64, 0x74, 0x3b, 0x74, 0x2e, 0x5f, + 0x5f, 0x63, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x28, 0x29, 0x2c, 0x64, 0x74, 0x3d, 0x6e, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, - 0x49, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, + 0x47, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x21, 0x74, 0x7c, 0x7c, 0x74, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x21, 0x3d, 0x3d, 0x6e, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x7c, 0x7c, 0x6e, 0x2e, 0x73, 0x6f, 0x6d, 0x65, 0x28, 0x28, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x6e, 0x2c, 0x65, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x21, 0x3d, 0x3d, 0x74, 0x5b, 0x65, 0x5d, 0x7d, 0x29, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, - 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x71, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, + 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x7a, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x22, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x6e, 0x3f, 0x6e, 0x28, 0x74, 0x29, 0x3a, 0x6e, 0x7d, 0x66, - 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x42, 0x74, 0x28, 0x74, + 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x4a, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x7b, 0x43, 0x5b, 0x74, 0x5d, 0x3d, 0x6e, 0x2e, 0x62, 0x69, 0x6e, 0x64, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x2c, 0x43, 0x5b, 0x74, 0x5d, 0x7c, 0x7c, 0x28, 0x28, 0x29, 0x3d, 0x3e, 0x7b, 0x7d, 0x29, 0x29, - 0x7d, 0x6c, 0x65, 0x74, 0x20, 0x47, 0x74, 0x2c, 0x7a, 0x74, 0x3b, 0x66, - 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x4a, 0x74, 0x28, 0x74, - 0x29, 0x7b, 0x69, 0x66, 0x28, 0x7a, 0x74, 0x29, 0x7a, 0x74, 0x28, 0x29, - 0x3b, 0x7a, 0x74, 0x3d, 0x74, 0x26, 0x26, 0x74, 0x2e, 0x53, 0x28, 0x29, - 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x4b, 0x74, + 0x7d, 0x6c, 0x65, 0x74, 0x20, 0x4b, 0x74, 0x2c, 0x51, 0x74, 0x3b, 0x66, + 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x58, 0x74, 0x28, 0x74, + 0x29, 0x7b, 0x69, 0x66, 0x28, 0x51, 0x74, 0x29, 0x51, 0x74, 0x28, 0x29, + 0x3b, 0x51, 0x74, 0x3d, 0x74, 0x26, 0x26, 0x74, 0x2e, 0x53, 0x28, 0x29, + 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x59, 0x74, 0x28, 0x7b, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x74, 0x7d, 0x29, 0x7b, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6e, 0x3d, 0x58, 0x74, 0x28, 0x74, 0x29, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6e, 0x3d, 0x74, 0x6e, 0x28, 0x74, 0x29, 0x3b, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x74, 0x3b, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x3d, 0x44, 0x74, 0x28, 0x28, 0x29, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x3d, 0x41, 0x74, 0x28, 0x28, 0x29, 0x3d, 0x3e, 0x7b, 0x6c, 0x65, 0x74, 0x20, 0x74, 0x3d, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x5f, 0x76, 0x3b, 0x77, 0x68, 0x69, 0x6c, 0x65, 0x28, 0x74, 0x3d, 0x74, 0x2e, 0x5f, 0x5f, 0x29, 0x69, 0x66, 0x28, 0x74, 0x2e, @@ -1592,7 +1620,7 @@ unsigned char index_js[] = { 0x74, 0x3f, 0x30, 0x3a, 0x21, 0x30, 0x3d, 0x3d, 0x3d, 0x74, 0x3f, 0x22, 0x22, 0x3a, 0x74, 0x7c, 0x7c, 0x22, 0x22, 0x7d, 0x29, 0x7d, 0x2c, 0x5b, 0x5d, 0x29, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x65, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x4b, 0x74, 0x2e, 0x64, 0x69, 0x73, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x59, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x4e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x5f, 0x73, 0x74, 0x22, 0x3b, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x2e, 0x64, 0x65, 0x66, 0x69, 0x6e, 0x65, 0x50, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, @@ -1603,14 +1631,14 @@ unsigned char index_js[] = { 0x75, 0x65, 0x3a, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x7d, 0x2c, 0x74, 0x79, 0x70, 0x65, 0x3a, 0x7b, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x75, 0x72, 0x61, 0x62, 0x6c, 0x65, 0x3a, 0x21, 0x30, 0x2c, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x3a, 0x4b, 0x74, 0x7d, 0x2c, 0x70, 0x72, 0x6f, 0x70, 0x73, + 0x75, 0x65, 0x3a, 0x59, 0x74, 0x7d, 0x2c, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x3a, 0x7b, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x75, 0x72, 0x61, 0x62, 0x6c, 0x65, 0x3a, 0x21, 0x30, 0x2c, 0x67, 0x65, 0x74, 0x28, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x7b, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x74, 0x68, 0x69, 0x73, 0x7d, 0x7d, 0x7d, 0x2c, 0x5f, 0x5f, 0x62, 0x3a, 0x7b, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x75, 0x72, 0x61, 0x62, 0x6c, 0x65, 0x3a, 0x21, 0x30, 0x2c, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x31, - 0x7d, 0x7d, 0x29, 0x3b, 0x42, 0x74, 0x28, 0x22, 0x5f, 0x5f, 0x62, 0x22, + 0x7d, 0x7d, 0x29, 0x3b, 0x4a, 0x74, 0x28, 0x22, 0x5f, 0x5f, 0x62, 0x22, 0x2c, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x3d, 0x3e, 0x7b, 0x69, 0x66, 0x28, 0x22, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x6e, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x29, @@ -1625,8 +1653,8 @@ unsigned char index_js[] = { 0x6e, 0x2e, 0x5f, 0x5f, 0x6e, 0x70, 0x3d, 0x74, 0x3d, 0x7b, 0x7d, 0x3b, 0x74, 0x5b, 0x5f, 0x5d, 0x3d, 0x69, 0x3b, 0x65, 0x5b, 0x5f, 0x5d, 0x3d, 0x69, 0x2e, 0x70, 0x65, 0x65, 0x6b, 0x28, 0x29, 0x7d, 0x7d, 0x7d, 0x74, - 0x28, 0x6e, 0x29, 0x7d, 0x29, 0x3b, 0x42, 0x74, 0x28, 0x22, 0x5f, 0x5f, - 0x72, 0x22, 0x2c, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x3d, 0x3e, 0x7b, 0x4a, + 0x28, 0x6e, 0x29, 0x7d, 0x29, 0x3b, 0x4a, 0x74, 0x28, 0x22, 0x5f, 0x5f, + 0x72, 0x22, 0x2c, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x3d, 0x3e, 0x7b, 0x58, 0x74, 0x28, 0x29, 0x3b, 0x6c, 0x65, 0x74, 0x20, 0x65, 0x2c, 0x5f, 0x3d, 0x6e, 0x2e, 0x5f, 0x5f, 0x63, 0x3b, 0x69, 0x66, 0x28, 0x5f, 0x29, 0x7b, 0x5f, 0x2e, 0x5f, 0x5f, 0x24, 0x66, 0x26, 0x3d, 0x2d, 0x32, 0x3b, 0x65, @@ -1640,14 +1668,14 @@ unsigned char index_js[] = { 0x5f, 0x5f, 0x24, 0x66, 0x7c, 0x3d, 0x31, 0x3b, 0x5f, 0x2e, 0x73, 0x65, 0x74, 0x53, 0x74, 0x61, 0x74, 0x65, 0x28, 0x7b, 0x7d, 0x29, 0x7d, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x7d, 0x28, 0x29, 0x7d, - 0x47, 0x74, 0x3d, 0x5f, 0x3b, 0x4a, 0x74, 0x28, 0x65, 0x29, 0x3b, 0x74, - 0x28, 0x6e, 0x29, 0x7d, 0x29, 0x3b, 0x42, 0x74, 0x28, 0x22, 0x5f, 0x5f, + 0x4b, 0x74, 0x3d, 0x5f, 0x3b, 0x58, 0x74, 0x28, 0x65, 0x29, 0x3b, 0x74, + 0x28, 0x6e, 0x29, 0x7d, 0x29, 0x3b, 0x4a, 0x74, 0x28, 0x22, 0x5f, 0x5f, 0x65, 0x22, 0x2c, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x2c, 0x5f, 0x29, - 0x3d, 0x3e, 0x7b, 0x4a, 0x74, 0x28, 0x29, 0x3b, 0x47, 0x74, 0x3d, 0x76, + 0x3d, 0x3e, 0x7b, 0x58, 0x74, 0x28, 0x29, 0x3b, 0x4b, 0x74, 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x3b, 0x74, 0x28, 0x6e, 0x2c, 0x65, 0x2c, - 0x5f, 0x29, 0x7d, 0x29, 0x3b, 0x42, 0x74, 0x28, 0x22, 0x64, 0x69, 0x66, + 0x5f, 0x29, 0x7d, 0x29, 0x3b, 0x4a, 0x74, 0x28, 0x22, 0x64, 0x69, 0x66, 0x66, 0x65, 0x64, 0x22, 0x2c, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x3d, 0x3e, - 0x7b, 0x4a, 0x74, 0x28, 0x29, 0x3b, 0x47, 0x74, 0x3d, 0x76, 0x6f, 0x69, + 0x7b, 0x58, 0x74, 0x28, 0x29, 0x3b, 0x4b, 0x74, 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x3b, 0x6c, 0x65, 0x74, 0x20, 0x65, 0x3b, 0x69, 0x66, 0x28, 0x22, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x6e, 0x2e, 0x74, 0x79, 0x70, 0x65, @@ -1667,11 +1695,11 @@ unsigned char index_js[] = { 0x6e, 0x20, 0x74, 0x29, 0x7b, 0x6c, 0x65, 0x74, 0x20, 0x6f, 0x3d, 0x6e, 0x5b, 0x69, 0x5d, 0x2c, 0x72, 0x3d, 0x74, 0x5b, 0x69, 0x5d, 0x3b, 0x69, 0x66, 0x28, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x3d, 0x3d, 0x3d, 0x6f, - 0x29, 0x7b, 0x6f, 0x3d, 0x51, 0x74, 0x28, 0x65, 0x2c, 0x69, 0x2c, 0x72, + 0x29, 0x7b, 0x6f, 0x3d, 0x5a, 0x74, 0x28, 0x65, 0x2c, 0x69, 0x2c, 0x72, 0x2c, 0x5f, 0x29, 0x3b, 0x6e, 0x5b, 0x69, 0x5d, 0x3d, 0x6f, 0x7d, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x6f, 0x2e, 0x6f, 0x28, 0x72, 0x2c, 0x5f, 0x29, 0x7d, 0x7d, 0x7d, 0x74, 0x28, 0x6e, 0x29, 0x7d, 0x29, 0x3b, 0x66, 0x75, - 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x51, 0x74, 0x28, 0x74, 0x2c, + 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x5a, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x2c, 0x5f, 0x29, 0x7b, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x3d, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x74, 0x26, 0x26, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x3d, 0x3d, 0x3d, 0x74, 0x2e, 0x6f, 0x77, @@ -1689,7 +1717,7 @@ unsigned char index_js[] = { 0x74, 0x74, 0x72, 0x69, 0x62, 0x75, 0x74, 0x65, 0x28, 0x6e, 0x2c, 0x65, 0x29, 0x3b, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x74, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x76, 0x65, 0x41, 0x74, 0x74, 0x72, 0x69, 0x62, 0x75, 0x74, 0x65, - 0x28, 0x6e, 0x29, 0x7d, 0x7d, 0x29, 0x7d, 0x7d, 0x42, 0x74, 0x28, 0x22, + 0x28, 0x6e, 0x29, 0x7d, 0x7d, 0x29, 0x7d, 0x7d, 0x4a, 0x74, 0x28, 0x22, 0x75, 0x6e, 0x6d, 0x6f, 0x75, 0x6e, 0x74, 0x22, 0x2c, 0x28, 0x74, 0x2c, 0x6e, 0x29, 0x3d, 0x3e, 0x7b, 0x69, 0x66, 0x28, 0x22, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x22, 0x3d, 0x3d, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, @@ -1707,7 +1735,7 @@ unsigned char index_js[] = { 0x24, 0x75, 0x3b, 0x69, 0x66, 0x28, 0x6e, 0x29, 0x7b, 0x74, 0x2e, 0x5f, 0x5f, 0x24, 0x75, 0x3d, 0x76, 0x6f, 0x69, 0x64, 0x20, 0x30, 0x3b, 0x6e, 0x2e, 0x64, 0x28, 0x29, 0x7d, 0x7d, 0x7d, 0x74, 0x28, 0x6e, 0x29, 0x7d, - 0x29, 0x3b, 0x42, 0x74, 0x28, 0x22, 0x5f, 0x5f, 0x68, 0x22, 0x2c, 0x28, + 0x29, 0x3b, 0x4a, 0x74, 0x28, 0x22, 0x5f, 0x5f, 0x68, 0x22, 0x2c, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x2c, 0x5f, 0x29, 0x3d, 0x3e, 0x7b, 0x69, 0x66, 0x28, 0x5f, 0x3c, 0x33, 0x7c, 0x7c, 0x39, 0x3d, 0x3d, 0x3d, 0x5f, 0x29, 0x6e, 0x2e, 0x5f, 0x5f, 0x24, 0x66, 0x7c, 0x3d, 0x32, 0x3b, 0x74, @@ -1735,23 +1763,23 @@ unsigned char index_js[] = { 0x6f, 0x70, 0x73, 0x29, 0x69, 0x66, 0x28, 0x21, 0x28, 0x5f, 0x20, 0x69, 0x6e, 0x20, 0x74, 0x29, 0x29, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x21, 0x30, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x21, 0x31, 0x7d, 0x3b, - 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x58, 0x74, 0x28, - 0x74, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x44, 0x74, + 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x74, 0x6e, 0x28, + 0x74, 0x29, 0x7b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x41, 0x74, 0x28, 0x28, 0x29, 0x3d, 0x3e, 0x61, 0x28, 0x74, 0x29, 0x2c, 0x5b, 0x5d, - 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x59, - 0x74, 0x28, 0x74, 0x29, 0x7b, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6e, - 0x3d, 0x4e, 0x74, 0x28, 0x74, 0x29, 0x3b, 0x6e, 0x2e, 0x63, 0x75, 0x72, - 0x72, 0x65, 0x6e, 0x74, 0x3d, 0x74, 0x3b, 0x47, 0x74, 0x2e, 0x5f, 0x5f, + 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6e, + 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6e, + 0x3d, 0x54, 0x74, 0x28, 0x74, 0x29, 0x3b, 0x6e, 0x2e, 0x63, 0x75, 0x72, + 0x72, 0x65, 0x6e, 0x74, 0x3d, 0x74, 0x3b, 0x4b, 0x74, 0x2e, 0x5f, 0x5f, 0x24, 0x66, 0x7c, 0x3d, 0x34, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, - 0x20, 0x44, 0x74, 0x28, 0x28, 0x29, 0x3d, 0x3e, 0x6d, 0x28, 0x28, 0x29, + 0x20, 0x41, 0x74, 0x28, 0x28, 0x29, 0x3d, 0x3e, 0x6d, 0x28, 0x28, 0x29, 0x3d, 0x3e, 0x6e, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x28, 0x29, 0x29, 0x2c, 0x5b, 0x5d, 0x29, 0x7d, 0x66, 0x75, 0x6e, 0x63, 0x74, - 0x69, 0x6f, 0x6e, 0x20, 0x5a, 0x74, 0x28, 0x74, 0x29, 0x7b, 0x63, 0x6f, - 0x6e, 0x73, 0x74, 0x20, 0x6e, 0x3d, 0x4e, 0x74, 0x28, 0x74, 0x29, 0x3b, + 0x69, 0x6f, 0x6e, 0x20, 0x65, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x6e, 0x3d, 0x54, 0x74, 0x28, 0x74, 0x29, 0x3b, 0x6e, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x3d, 0x74, 0x3b, - 0x48, 0x74, 0x28, 0x28, 0x29, 0x3d, 0x3e, 0x77, 0x28, 0x28, 0x29, 0x3d, + 0x24, 0x74, 0x28, 0x28, 0x29, 0x3d, 0x3e, 0x77, 0x28, 0x28, 0x29, 0x3d, 0x3e, 0x6e, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x28, 0x29, - 0x29, 0x2c, 0x5b, 0x5d, 0x29, 0x7d, 0x76, 0x61, 0x72, 0x20, 0x74, 0x6e, + 0x29, 0x2c, 0x5b, 0x5d, 0x29, 0x7d, 0x76, 0x61, 0x72, 0x20, 0x5f, 0x6e, 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x2c, 0x6e, 0x2c, 0x65, 0x2c, 0x5f, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x69, 0x3b, 0x6e, 0x5b, 0x30, 0x5d, 0x3d, 0x30, 0x3b, 0x66, 0x6f, 0x72, 0x28, @@ -1771,21 +1799,21 @@ unsigned char index_js[] = { 0x3a, 0x36, 0x3d, 0x3d, 0x3d, 0x72, 0x3f, 0x5f, 0x5b, 0x31, 0x5d, 0x5b, 0x6e, 0x5b, 0x2b, 0x2b, 0x6f, 0x5d, 0x5d, 0x2b, 0x3d, 0x75, 0x2b, 0x22, 0x22, 0x3a, 0x72, 0x3f, 0x28, 0x69, 0x3d, 0x74, 0x2e, 0x61, 0x70, 0x70, - 0x6c, 0x79, 0x28, 0x75, 0x2c, 0x74, 0x6e, 0x28, 0x74, 0x2c, 0x75, 0x2c, + 0x6c, 0x79, 0x28, 0x75, 0x2c, 0x5f, 0x6e, 0x28, 0x74, 0x2c, 0x75, 0x2c, 0x65, 0x2c, 0x5b, 0x22, 0x22, 0x2c, 0x6e, 0x75, 0x6c, 0x6c, 0x5d, 0x29, 0x29, 0x2c, 0x5f, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x69, 0x29, 0x2c, 0x75, 0x5b, 0x30, 0x5d, 0x3f, 0x6e, 0x5b, 0x30, 0x5d, 0x7c, 0x3d, 0x32, 0x3a, 0x28, 0x6e, 0x5b, 0x6f, 0x2d, 0x32, 0x5d, 0x3d, 0x30, 0x2c, 0x6e, 0x5b, 0x6f, 0x5d, 0x3d, 0x69, 0x29, 0x29, 0x3a, 0x5f, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x75, 0x29, 0x7d, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, - 0x20, 0x5f, 0x7d, 0x2c, 0x6e, 0x6e, 0x3d, 0x6e, 0x65, 0x77, 0x20, 0x4d, + 0x20, 0x5f, 0x7d, 0x2c, 0x6f, 0x6e, 0x3d, 0x6e, 0x65, 0x77, 0x20, 0x4d, 0x61, 0x70, 0x3b, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, - 0x65, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, - 0x6e, 0x6e, 0x2e, 0x67, 0x65, 0x74, 0x28, 0x74, 0x68, 0x69, 0x73, 0x29, + 0x72, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x76, 0x61, 0x72, 0x20, 0x6e, 0x3d, + 0x6f, 0x6e, 0x2e, 0x67, 0x65, 0x74, 0x28, 0x74, 0x68, 0x69, 0x73, 0x29, 0x3b, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x7c, 0x7c, 0x28, - 0x6e, 0x3d, 0x6e, 0x65, 0x77, 0x20, 0x4d, 0x61, 0x70, 0x2c, 0x6e, 0x6e, + 0x6e, 0x3d, 0x6e, 0x65, 0x77, 0x20, 0x4d, 0x61, 0x70, 0x2c, 0x6f, 0x6e, 0x2e, 0x73, 0x65, 0x74, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2c, 0x6e, 0x29, - 0x29, 0x2c, 0x28, 0x6e, 0x3d, 0x74, 0x6e, 0x28, 0x74, 0x68, 0x69, 0x73, + 0x29, 0x2c, 0x28, 0x6e, 0x3d, 0x5f, 0x6e, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2c, 0x6e, 0x2e, 0x67, 0x65, 0x74, 0x28, 0x74, 0x29, 0x7c, 0x7c, 0x28, 0x6e, 0x2e, 0x73, 0x65, 0x74, 0x28, 0x74, 0x2c, 0x6e, 0x3d, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x28, 0x74, 0x29, 0x7b, 0x66, 0x6f, @@ -1816,10 +1844,10 @@ unsigned char index_js[] = { 0x30, 0x3b, 0x66, 0x3c, 0x74, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3b, 0x66, 0x2b, 0x2b, 0x29, 0x7b, 0x66, 0x26, 0x26, 0x28, 0x31, 0x3d, 0x3d, 0x3d, 0x5f, 0x26, 0x26, 0x75, 0x28, 0x29, 0x2c, 0x75, 0x28, 0x66, - 0x29, 0x29, 0x3b, 0x66, 0x6f, 0x72, 0x28, 0x76, 0x61, 0x72, 0x20, 0x73, - 0x3d, 0x30, 0x3b, 0x73, 0x3c, 0x74, 0x5b, 0x66, 0x5d, 0x2e, 0x6c, 0x65, - 0x6e, 0x67, 0x74, 0x68, 0x3b, 0x73, 0x2b, 0x2b, 0x29, 0x6e, 0x3d, 0x74, - 0x5b, 0x66, 0x5d, 0x5b, 0x73, 0x5d, 0x2c, 0x31, 0x3d, 0x3d, 0x3d, 0x5f, + 0x29, 0x29, 0x3b, 0x66, 0x6f, 0x72, 0x28, 0x76, 0x61, 0x72, 0x20, 0x6c, + 0x3d, 0x30, 0x3b, 0x6c, 0x3c, 0x74, 0x5b, 0x66, 0x5d, 0x2e, 0x6c, 0x65, + 0x6e, 0x67, 0x74, 0x68, 0x3b, 0x6c, 0x2b, 0x2b, 0x29, 0x6e, 0x3d, 0x74, + 0x5b, 0x66, 0x5d, 0x5b, 0x6c, 0x5d, 0x2c, 0x31, 0x3d, 0x3d, 0x3d, 0x5f, 0x3f, 0x22, 0x3c, 0x22, 0x3d, 0x3d, 0x3d, 0x6e, 0x3f, 0x28, 0x75, 0x28, 0x29, 0x2c, 0x72, 0x3d, 0x5b, 0x72, 0x5d, 0x2c, 0x5f, 0x3d, 0x33, 0x29, 0x3a, 0x69, 0x2b, 0x3d, 0x6e, 0x3a, 0x34, 0x3d, 0x3d, 0x3d, 0x5f, 0x3f, @@ -1835,7 +1863,7 @@ unsigned char index_js[] = { 0x5f, 0x3d, 0x35, 0x2c, 0x65, 0x3d, 0x69, 0x2c, 0x69, 0x3d, 0x22, 0x22, 0x29, 0x3a, 0x22, 0x2f, 0x22, 0x3d, 0x3d, 0x3d, 0x6e, 0x26, 0x26, 0x28, 0x5f, 0x3c, 0x35, 0x7c, 0x7c, 0x22, 0x3e, 0x22, 0x3d, 0x3d, 0x3d, 0x74, - 0x5b, 0x66, 0x5d, 0x5b, 0x73, 0x2b, 0x31, 0x5d, 0x29, 0x3f, 0x28, 0x75, + 0x5b, 0x66, 0x5d, 0x5b, 0x6c, 0x2b, 0x31, 0x5d, 0x29, 0x3f, 0x28, 0x75, 0x28, 0x29, 0x2c, 0x33, 0x3d, 0x3d, 0x3d, 0x5f, 0x26, 0x26, 0x28, 0x72, 0x3d, 0x72, 0x5b, 0x30, 0x5d, 0x29, 0x2c, 0x5f, 0x3d, 0x72, 0x2c, 0x28, 0x72, 0x3d, 0x72, 0x5b, 0x30, 0x5d, 0x29, 0x2e, 0x70, 0x75, 0x73, 0x68, @@ -1851,53 +1879,53 @@ unsigned char index_js[] = { 0x2c, 0x72, 0x7d, 0x28, 0x74, 0x29, 0x29, 0x2c, 0x6e, 0x29, 0x2c, 0x61, 0x72, 0x67, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x73, 0x2c, 0x5b, 0x5d, 0x29, 0x29, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3e, 0x31, 0x3f, 0x6e, - 0x3a, 0x6e, 0x5b, 0x30, 0x5d, 0x7d, 0x76, 0x61, 0x72, 0x20, 0x5f, 0x6e, - 0x3d, 0x65, 0x6e, 0x2e, 0x62, 0x69, 0x6e, 0x64, 0x28, 0x4c, 0x29, 0x3b, + 0x3a, 0x6e, 0x5b, 0x30, 0x5d, 0x7d, 0x76, 0x61, 0x72, 0x20, 0x75, 0x6e, + 0x3d, 0x72, 0x6e, 0x2e, 0x62, 0x69, 0x6e, 0x64, 0x28, 0x4c, 0x29, 0x3b, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x7b, 0x49, 0x20, 0x61, 0x73, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x2c, 0x6a, 0x20, 0x61, 0x73, 0x20, 0x46, 0x72, 0x61, 0x67, 0x6d, 0x65, 0x6e, 0x74, 0x2c, 0x68, 0x20, 0x61, 0x73, 0x20, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, - 0x5f, 0x20, 0x61, 0x73, 0x20, 0x62, 0x61, 0x74, 0x63, 0x68, 0x2c, 0x63, + 0x5f, 0x20, 0x61, 0x73, 0x20, 0x62, 0x61, 0x74, 0x63, 0x68, 0x2c, 0x68, 0x74, 0x20, 0x61, 0x73, 0x20, 0x63, 0x6c, 0x6f, 0x6e, 0x65, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x2c, 0x6d, 0x20, 0x61, 0x73, 0x20, 0x63, - 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x2c, 0x68, 0x74, 0x20, 0x61, + 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x2c, 0x61, 0x74, 0x20, 0x61, 0x73, 0x20, 0x63, 0x72, 0x65, 0x61, 0x74, 0x65, 0x43, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x2c, 0x4c, 0x20, 0x61, 0x73, 0x20, 0x63, 0x72, 0x65, 0x61, 0x74, 0x65, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x2c, 0x52, 0x20, 0x61, 0x73, 0x20, 0x63, 0x72, 0x65, 0x61, 0x74, 0x65, 0x52, 0x65, 0x66, 0x2c, 0x77, 0x20, 0x61, 0x73, 0x20, 0x65, 0x66, 0x66, 0x65, 0x63, - 0x74, 0x2c, 0x4c, 0x20, 0x61, 0x73, 0x20, 0x68, 0x2c, 0x5f, 0x6e, 0x20, - 0x61, 0x73, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x2c, 0x6c, 0x74, 0x20, 0x61, + 0x74, 0x2c, 0x4c, 0x20, 0x61, 0x73, 0x20, 0x68, 0x2c, 0x75, 0x6e, 0x20, + 0x61, 0x73, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x2c, 0x63, 0x74, 0x20, 0x61, 0x73, 0x20, 0x68, 0x79, 0x64, 0x72, 0x61, 0x74, 0x65, 0x2c, 0x55, 0x20, 0x61, 0x73, 0x20, 0x69, 0x73, 0x56, 0x61, 0x6c, 0x69, 0x64, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x2c, 0x43, 0x20, 0x61, 0x73, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x2c, 0x73, 0x74, 0x20, 0x61, 0x73, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x2c, 0x61, 0x20, 0x61, 0x73, - 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x58, 0x20, 0x61, 0x73, + 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x59, 0x20, 0x61, 0x73, 0x20, 0x74, 0x6f, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x41, 0x72, 0x72, 0x61, 0x79, 0x2c, 0x75, 0x20, 0x61, 0x73, 0x20, 0x75, 0x6e, 0x74, 0x72, 0x61, - 0x63, 0x6b, 0x65, 0x64, 0x2c, 0x54, 0x74, 0x20, 0x61, 0x73, 0x20, 0x75, - 0x73, 0x65, 0x43, 0x61, 0x6c, 0x6c, 0x62, 0x61, 0x63, 0x6b, 0x2c, 0x59, - 0x74, 0x20, 0x61, 0x73, 0x20, 0x75, 0x73, 0x65, 0x43, 0x6f, 0x6d, 0x70, - 0x75, 0x74, 0x65, 0x64, 0x2c, 0x56, 0x74, 0x20, 0x61, 0x73, 0x20, 0x75, - 0x73, 0x65, 0x43, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x2c, 0x41, 0x74, + 0x63, 0x6b, 0x65, 0x64, 0x2c, 0x46, 0x74, 0x20, 0x61, 0x73, 0x20, 0x75, + 0x73, 0x65, 0x43, 0x61, 0x6c, 0x6c, 0x62, 0x61, 0x63, 0x6b, 0x2c, 0x6e, + 0x6e, 0x20, 0x61, 0x73, 0x20, 0x75, 0x73, 0x65, 0x43, 0x6f, 0x6d, 0x70, + 0x75, 0x74, 0x65, 0x64, 0x2c, 0x4d, 0x74, 0x20, 0x61, 0x73, 0x20, 0x75, + 0x73, 0x65, 0x43, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x2c, 0x57, 0x74, 0x20, 0x61, 0x73, 0x20, 0x75, 0x73, 0x65, 0x44, 0x65, 0x62, 0x75, 0x67, - 0x56, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x48, 0x74, 0x20, 0x61, 0x73, 0x20, - 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x46, 0x74, + 0x56, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x24, 0x74, 0x20, 0x61, 0x73, 0x20, + 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x4c, 0x74, 0x20, 0x61, 0x73, 0x20, 0x75, 0x73, 0x65, 0x45, 0x72, 0x72, 0x6f, 0x72, - 0x42, 0x6f, 0x75, 0x6e, 0x64, 0x61, 0x72, 0x79, 0x2c, 0x4d, 0x74, 0x20, - 0x61, 0x73, 0x20, 0x75, 0x73, 0x65, 0x49, 0x64, 0x2c, 0x24, 0x74, 0x20, + 0x42, 0x6f, 0x75, 0x6e, 0x64, 0x61, 0x72, 0x79, 0x2c, 0x4f, 0x74, 0x20, + 0x61, 0x73, 0x20, 0x75, 0x73, 0x65, 0x49, 0x64, 0x2c, 0x56, 0x74, 0x20, 0x61, 0x73, 0x20, 0x75, 0x73, 0x65, 0x49, 0x6d, 0x70, 0x65, 0x72, 0x61, - 0x74, 0x69, 0x76, 0x65, 0x48, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x2c, 0x50, + 0x74, 0x69, 0x76, 0x65, 0x48, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x2c, 0x44, 0x74, 0x20, 0x61, 0x73, 0x20, 0x75, 0x73, 0x65, 0x4c, 0x61, 0x79, 0x6f, - 0x75, 0x74, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x44, 0x74, 0x20, - 0x61, 0x73, 0x20, 0x75, 0x73, 0x65, 0x4d, 0x65, 0x6d, 0x6f, 0x2c, 0x55, + 0x75, 0x74, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x41, 0x74, 0x20, + 0x61, 0x73, 0x20, 0x75, 0x73, 0x65, 0x4d, 0x65, 0x6d, 0x6f, 0x2c, 0x4e, 0x74, 0x20, 0x61, 0x73, 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x64, 0x75, - 0x63, 0x65, 0x72, 0x2c, 0x4e, 0x74, 0x20, 0x61, 0x73, 0x20, 0x75, 0x73, - 0x65, 0x52, 0x65, 0x66, 0x2c, 0x58, 0x74, 0x20, 0x61, 0x73, 0x20, 0x75, - 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x5a, 0x74, 0x20, + 0x63, 0x65, 0x72, 0x2c, 0x54, 0x74, 0x20, 0x61, 0x73, 0x20, 0x75, 0x73, + 0x65, 0x52, 0x65, 0x66, 0x2c, 0x74, 0x6e, 0x20, 0x61, 0x73, 0x20, 0x75, + 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x65, 0x6e, 0x20, 0x61, 0x73, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, - 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x45, 0x74, 0x20, 0x61, 0x73, + 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x50, 0x74, 0x20, 0x61, 0x73, 0x20, 0x75, 0x73, 0x65, 0x53, 0x74, 0x61, 0x74, 0x65, 0x7d, 0x3b, 0x0a }; -unsigned int index_js_len = 22800; +size_t index_js_len = 23136; diff --git a/examples/server/json-schema-to-grammar.mjs.hpp b/examples/server/json-schema-to-grammar.mjs.hpp index 0a05c369d..ef1d5dcf3 100644 --- a/examples/server/json-schema-to-grammar.mjs.hpp +++ b/examples/server/json-schema-to-grammar.mjs.hpp @@ -1,152 +1,1027 @@ unsigned char json_schema_to_grammar_mjs[] = { - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x53, 0x50, 0x41, 0x43, 0x45, 0x5f, - 0x52, 0x55, 0x4c, 0x45, 0x20, 0x3d, 0x20, 0x27, 0x22, 0x20, 0x22, 0x3f, - 0x27, 0x3b, 0x0a, 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x52, - 0x49, 0x4d, 0x49, 0x54, 0x49, 0x56, 0x45, 0x5f, 0x52, 0x55, 0x4c, 0x45, - 0x53, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x62, 0x6f, 0x6f, 0x6c, - 0x65, 0x61, 0x6e, 0x3a, 0x20, 0x27, 0x28, 0x22, 0x74, 0x72, 0x75, 0x65, - 0x22, 0x20, 0x7c, 0x20, 0x22, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x22, 0x29, - 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x6e, - 0x75, 0x6d, 0x62, 0x65, 0x72, 0x3a, 0x20, 0x27, 0x28, 0x22, 0x2d, 0x22, - 0x3f, 0x20, 0x28, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x20, 0x7c, 0x20, 0x5b, - 0x31, 0x2d, 0x39, 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x2a, 0x29, - 0x29, 0x20, 0x28, 0x22, 0x2e, 0x22, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, - 0x2b, 0x29, 0x3f, 0x20, 0x28, 0x5b, 0x65, 0x45, 0x5d, 0x20, 0x5b, 0x2d, - 0x2b, 0x5d, 0x3f, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x2b, 0x29, 0x3f, - 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x69, - 0x6e, 0x74, 0x65, 0x67, 0x65, 0x72, 0x3a, 0x20, 0x27, 0x28, 0x22, 0x2d, + 0x2f, 0x2f, 0x20, 0x57, 0x41, 0x52, 0x4e, 0x49, 0x4e, 0x47, 0x3a, 0x20, + 0x54, 0x68, 0x69, 0x73, 0x20, 0x66, 0x69, 0x6c, 0x65, 0x20, 0x77, 0x61, + 0x73, 0x20, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x20, 0x66, 0x72, 0x6f, + 0x6d, 0x20, 0x6a, 0x73, 0x6f, 0x6e, 0x2d, 0x73, 0x63, 0x68, 0x65, 0x6d, + 0x61, 0x2d, 0x74, 0x6f, 0x2d, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, + 0x2e, 0x70, 0x79, 0x2c, 0x20, 0x70, 0x6c, 0x65, 0x61, 0x73, 0x65, 0x20, + 0x66, 0x69, 0x78, 0x20, 0x62, 0x75, 0x67, 0x73, 0x20, 0x2f, 0x20, 0x61, + 0x64, 0x64, 0x20, 0x66, 0x65, 0x61, 0x74, 0x75, 0x72, 0x65, 0x73, 0x20, + 0x74, 0x68, 0x65, 0x72, 0x65, 0x20, 0x66, 0x69, 0x72, 0x73, 0x74, 0x2e, + 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x53, 0x50, 0x41, 0x43, 0x45, + 0x5f, 0x52, 0x55, 0x4c, 0x45, 0x20, 0x3d, 0x20, 0x27, 0x22, 0x20, 0x22, + 0x3f, 0x27, 0x3b, 0x0a, 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, + 0x52, 0x49, 0x4d, 0x49, 0x54, 0x49, 0x56, 0x45, 0x5f, 0x52, 0x55, 0x4c, + 0x45, 0x53, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x62, 0x6f, 0x6f, + 0x6c, 0x65, 0x61, 0x6e, 0x3a, 0x20, 0x27, 0x28, 0x22, 0x74, 0x72, 0x75, + 0x65, 0x22, 0x20, 0x7c, 0x20, 0x22, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x22, + 0x29, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, 0x20, 0x20, + 0x6e, 0x75, 0x6d, 0x62, 0x65, 0x72, 0x3a, 0x20, 0x27, 0x28, 0x22, 0x2d, 0x22, 0x3f, 0x20, 0x28, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x20, 0x7c, 0x20, 0x5b, 0x31, 0x2d, 0x39, 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x2a, - 0x29, 0x29, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, 0x20, - 0x20, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x60, 0x20, 0x22, - 0x5c, 0x5c, 0x22, 0x22, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x5b, 0x5e, 0x22, 0x5c, 0x5c, 0x5c, 0x5c, 0x5d, 0x20, - 0x7c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x22, 0x5c, - 0x5c, 0x5c, 0x5c, 0x22, 0x20, 0x28, 0x5b, 0x22, 0x5c, 0x5c, 0x5c, 0x5c, - 0x2f, 0x62, 0x66, 0x6e, 0x72, 0x74, 0x5d, 0x20, 0x7c, 0x20, 0x22, 0x75, - 0x22, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x61, 0x2d, 0x66, 0x41, 0x2d, 0x46, - 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x61, 0x2d, 0x66, 0x41, 0x2d, 0x46, - 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x61, 0x2d, 0x66, 0x41, 0x2d, 0x46, - 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x61, 0x2d, 0x66, 0x41, 0x2d, 0x46, - 0x5d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x2a, 0x20, - 0x22, 0x5c, 0x5c, 0x22, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x60, - 0x2c, 0x0a, 0x20, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3a, 0x20, 0x27, 0x22, - 0x6e, 0x75, 0x6c, 0x6c, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, - 0x2c, 0x0a, 0x7d, 0x3b, 0x0a, 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x49, 0x4e, 0x56, 0x41, 0x4c, 0x49, 0x44, 0x5f, 0x52, 0x55, 0x4c, 0x45, - 0x5f, 0x43, 0x48, 0x41, 0x52, 0x53, 0x5f, 0x52, 0x45, 0x20, 0x3d, 0x20, - 0x2f, 0x5b, 0x5e, 0x5c, 0x64, 0x41, 0x2d, 0x5a, 0x61, 0x2d, 0x7a, 0x2d, - 0x5d, 0x2b, 0x2f, 0x67, 0x3b, 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x47, 0x52, 0x41, 0x4d, 0x4d, 0x41, 0x52, 0x5f, 0x4c, 0x49, 0x54, 0x45, + 0x29, 0x29, 0x20, 0x28, 0x22, 0x2e, 0x22, 0x20, 0x5b, 0x30, 0x2d, 0x39, + 0x5d, 0x2b, 0x29, 0x3f, 0x20, 0x28, 0x5b, 0x65, 0x45, 0x5d, 0x20, 0x5b, + 0x2d, 0x2b, 0x5d, 0x3f, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x2b, 0x29, + 0x3f, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, 0x20, 0x20, + 0x69, 0x6e, 0x74, 0x65, 0x67, 0x65, 0x72, 0x3a, 0x20, 0x27, 0x28, 0x22, + 0x2d, 0x22, 0x3f, 0x20, 0x28, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x20, 0x7c, + 0x20, 0x5b, 0x31, 0x2d, 0x39, 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, + 0x2a, 0x29, 0x29, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, + 0x20, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x27, 0x6f, 0x62, + 0x6a, 0x65, 0x63, 0x74, 0x20, 0x7c, 0x20, 0x61, 0x72, 0x72, 0x61, 0x79, + 0x20, 0x7c, 0x20, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x7c, 0x20, + 0x6e, 0x75, 0x6d, 0x62, 0x65, 0x72, 0x20, 0x7c, 0x20, 0x62, 0x6f, 0x6f, + 0x6c, 0x65, 0x61, 0x6e, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x6f, 0x62, 0x6a, + 0x65, 0x63, 0x74, 0x3a, 0x20, 0x27, 0x22, 0x7b, 0x22, 0x20, 0x73, 0x70, + 0x61, 0x63, 0x65, 0x20, 0x28, 0x20, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, + 0x20, 0x22, 0x3a, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x20, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x20, 0x28, 0x22, 0x2c, 0x22, 0x20, 0x73, 0x70, + 0x61, 0x63, 0x65, 0x20, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x22, + 0x3a, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x20, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x29, 0x2a, 0x20, 0x29, 0x3f, 0x20, 0x22, 0x7d, 0x22, 0x20, + 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x61, 0x72, + 0x72, 0x61, 0x79, 0x3a, 0x20, 0x27, 0x22, 0x5b, 0x22, 0x20, 0x73, 0x70, + 0x61, 0x63, 0x65, 0x20, 0x28, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, + 0x28, 0x22, 0x2c, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x20, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x29, 0x2a, 0x20, 0x29, 0x3f, 0x20, 0x22, 0x5d, + 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, 0x20, 0x20, + 0x75, 0x75, 0x69, 0x64, 0x3a, 0x20, 0x27, 0x22, 0x5c, 0x5c, 0x22, 0x22, + 0x20, 0x27, 0x20, 0x2b, 0x20, 0x5b, 0x38, 0x2c, 0x20, 0x34, 0x2c, 0x20, + 0x34, 0x2c, 0x20, 0x34, 0x2c, 0x20, 0x31, 0x32, 0x5d, 0x2e, 0x6d, 0x61, + 0x70, 0x28, 0x6e, 0x20, 0x3d, 0x3e, 0x20, 0x5b, 0x2e, 0x2e, 0x2e, 0x6e, + 0x65, 0x77, 0x20, 0x41, 0x72, 0x72, 0x61, 0x79, 0x28, 0x6e, 0x29, 0x5d, + 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x5f, 0x20, 0x3d, 0x3e, 0x20, 0x27, 0x5b, + 0x30, 0x2d, 0x39, 0x61, 0x2d, 0x66, 0x41, 0x2d, 0x46, 0x5d, 0x27, 0x29, + 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, 0x29, 0x2e, 0x6a, + 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x20, 0x22, 0x2d, 0x22, 0x20, 0x27, 0x29, + 0x20, 0x2b, 0x20, 0x27, 0x20, 0x22, 0x5c, 0x5c, 0x22, 0x22, 0x20, 0x73, + 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x73, 0x74, 0x72, + 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x60, 0x20, 0x22, 0x5c, 0x5c, 0x22, 0x22, + 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x5b, + 0x5e, 0x22, 0x5c, 0x5c, 0x5c, 0x5c, 0x5d, 0x20, 0x7c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x22, 0x5c, 0x5c, 0x5c, 0x5c, 0x22, + 0x20, 0x28, 0x5b, 0x22, 0x5c, 0x5c, 0x5c, 0x5c, 0x2f, 0x62, 0x66, 0x6e, + 0x72, 0x74, 0x5d, 0x20, 0x7c, 0x20, 0x22, 0x75, 0x22, 0x20, 0x5b, 0x30, + 0x2d, 0x39, 0x61, 0x2d, 0x66, 0x41, 0x2d, 0x46, 0x5d, 0x20, 0x5b, 0x30, + 0x2d, 0x39, 0x61, 0x2d, 0x66, 0x41, 0x2d, 0x46, 0x5d, 0x20, 0x5b, 0x30, + 0x2d, 0x39, 0x61, 0x2d, 0x66, 0x41, 0x2d, 0x46, 0x5d, 0x20, 0x5b, 0x30, + 0x2d, 0x39, 0x61, 0x2d, 0x66, 0x41, 0x2d, 0x46, 0x5d, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x2a, 0x20, 0x22, 0x5c, 0x5c, 0x22, + 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x60, 0x2c, 0x0a, 0x20, 0x20, + 0x6e, 0x75, 0x6c, 0x6c, 0x3a, 0x20, 0x27, 0x22, 0x6e, 0x75, 0x6c, 0x6c, + 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, 0x7d, 0x3b, + 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x4f, 0x42, 0x4a, 0x45, 0x43, + 0x54, 0x5f, 0x52, 0x55, 0x4c, 0x45, 0x5f, 0x4e, 0x41, 0x4d, 0x45, 0x53, + 0x20, 0x3d, 0x20, 0x5b, 0x27, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x27, + 0x2c, 0x20, 0x27, 0x61, 0x72, 0x72, 0x61, 0x79, 0x27, 0x2c, 0x20, 0x27, + 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x2c, 0x20, 0x27, 0x6e, 0x75, + 0x6d, 0x62, 0x65, 0x72, 0x27, 0x2c, 0x20, 0x27, 0x62, 0x6f, 0x6f, 0x6c, + 0x65, 0x61, 0x6e, 0x27, 0x2c, 0x20, 0x27, 0x6e, 0x75, 0x6c, 0x6c, 0x27, + 0x2c, 0x20, 0x27, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x27, 0x5d, 0x3b, 0x0a, + 0x0a, 0x2f, 0x2f, 0x20, 0x54, 0x4f, 0x44, 0x4f, 0x3a, 0x20, 0x73, 0x75, + 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x22, 0x75, 0x72, 0x69, 0x22, 0x2c, + 0x20, 0x22, 0x65, 0x6d, 0x61, 0x69, 0x6c, 0x22, 0x20, 0x73, 0x74, 0x72, + 0x69, 0x6e, 0x67, 0x20, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x73, 0x0a, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x44, 0x41, 0x54, 0x45, 0x5f, 0x52, + 0x55, 0x4c, 0x45, 0x53, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x27, 0x64, 0x61, 0x74, 0x65, 0x27, 0x20, 0x20, 0x20, 0x3a, 0x20, + 0x27, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, + 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, + 0x20, 0x22, 0x2d, 0x22, 0x20, 0x28, 0x20, 0x22, 0x30, 0x22, 0x20, 0x5b, + 0x31, 0x2d, 0x39, 0x5d, 0x20, 0x7c, 0x20, 0x22, 0x31, 0x22, 0x20, 0x5b, + 0x30, 0x2d, 0x32, 0x5d, 0x20, 0x29, 0x20, 0x22, 0x2d, 0x22, 0x20, 0x28, + 0x20, 0x5c, 0x22, 0x30, 0x5c, 0x22, 0x20, 0x5b, 0x31, 0x2d, 0x39, 0x5d, + 0x20, 0x7c, 0x20, 0x5b, 0x31, 0x2d, 0x32, 0x5d, 0x20, 0x5b, 0x30, 0x2d, + 0x39, 0x5d, 0x20, 0x7c, 0x20, 0x22, 0x33, 0x22, 0x20, 0x5b, 0x30, 0x2d, + 0x31, 0x5d, 0x20, 0x29, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x27, + 0x74, 0x69, 0x6d, 0x65, 0x27, 0x20, 0x20, 0x20, 0x3a, 0x20, 0x27, 0x28, + 0x5b, 0x30, 0x31, 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x20, 0x7c, + 0x20, 0x22, 0x32, 0x22, 0x20, 0x5b, 0x30, 0x2d, 0x33, 0x5d, 0x29, 0x20, + 0x22, 0x3a, 0x22, 0x20, 0x5b, 0x30, 0x2d, 0x35, 0x5d, 0x20, 0x5b, 0x30, + 0x2d, 0x39, 0x5d, 0x20, 0x22, 0x3a, 0x22, 0x20, 0x5b, 0x30, 0x2d, 0x35, + 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x20, 0x28, 0x20, 0x22, 0x2e, + 0x22, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, + 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x20, 0x29, 0x3f, 0x20, 0x28, + 0x20, 0x22, 0x5a, 0x22, 0x20, 0x7c, 0x20, 0x28, 0x20, 0x22, 0x2b, 0x22, + 0x20, 0x7c, 0x20, 0x22, 0x2d, 0x22, 0x20, 0x29, 0x20, 0x28, 0x20, 0x5b, + 0x30, 0x31, 0x5d, 0x20, 0x5b, 0x30, 0x2d, 0x39, 0x5d, 0x20, 0x7c, 0x20, + 0x22, 0x32, 0x22, 0x20, 0x5b, 0x30, 0x2d, 0x33, 0x5d, 0x20, 0x29, 0x20, + 0x22, 0x3a, 0x22, 0x20, 0x5b, 0x30, 0x2d, 0x35, 0x5d, 0x20, 0x5b, 0x30, + 0x2d, 0x39, 0x5d, 0x20, 0x29, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x27, 0x64, 0x61, 0x74, 0x65, 0x2d, 0x74, 0x69, 0x6d, 0x65, 0x27, 0x3a, + 0x20, 0x27, 0x64, 0x61, 0x74, 0x65, 0x20, 0x22, 0x54, 0x22, 0x20, 0x74, + 0x69, 0x6d, 0x65, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x27, 0x64, + 0x61, 0x74, 0x65, 0x2d, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x3a, + 0x20, 0x27, 0x22, 0x5c, 0x5c, 0x22, 0x22, 0x20, 0x64, 0x61, 0x74, 0x65, + 0x20, 0x22, 0x5c, 0x5c, 0x22, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, + 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x27, 0x74, 0x69, 0x6d, 0x65, + 0x2d, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x3a, 0x20, 0x27, 0x22, + 0x5c, 0x5c, 0x22, 0x22, 0x20, 0x74, 0x69, 0x6d, 0x65, 0x20, 0x22, 0x5c, + 0x5c, 0x22, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x27, 0x64, 0x61, 0x74, 0x65, 0x2d, 0x74, 0x69, + 0x6d, 0x65, 0x2d, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x3a, 0x20, + 0x27, 0x22, 0x5c, 0x5c, 0x22, 0x22, 0x20, 0x64, 0x61, 0x74, 0x65, 0x2d, + 0x74, 0x69, 0x6d, 0x65, 0x20, 0x22, 0x5c, 0x5c, 0x22, 0x22, 0x20, 0x73, + 0x70, 0x61, 0x63, 0x65, 0x27, 0x2c, 0x0a, 0x7d, 0x3b, 0x0a, 0x0a, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x52, 0x45, 0x53, 0x45, 0x52, 0x56, 0x45, + 0x44, 0x5f, 0x4e, 0x41, 0x4d, 0x45, 0x53, 0x20, 0x3d, 0x20, 0x7b, 0x27, + 0x72, 0x6f, 0x6f, 0x74, 0x27, 0x3a, 0x20, 0x74, 0x72, 0x75, 0x65, 0x2c, + 0x20, 0x2e, 0x2e, 0x2e, 0x50, 0x52, 0x49, 0x4d, 0x49, 0x54, 0x49, 0x56, + 0x45, 0x5f, 0x52, 0x55, 0x4c, 0x45, 0x53, 0x2c, 0x20, 0x2e, 0x2e, 0x2e, + 0x44, 0x41, 0x54, 0x45, 0x5f, 0x52, 0x55, 0x4c, 0x45, 0x53, 0x7d, 0x3b, + 0x0a, 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x49, 0x4e, 0x56, 0x41, + 0x4c, 0x49, 0x44, 0x5f, 0x52, 0x55, 0x4c, 0x45, 0x5f, 0x43, 0x48, 0x41, + 0x52, 0x53, 0x5f, 0x52, 0x45, 0x20, 0x3d, 0x20, 0x2f, 0x5b, 0x5e, 0x5c, + 0x64, 0x41, 0x2d, 0x5a, 0x61, 0x2d, 0x7a, 0x2d, 0x5d, 0x2b, 0x2f, 0x67, + 0x3b, 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x47, 0x52, 0x41, 0x4d, + 0x4d, 0x41, 0x52, 0x5f, 0x4c, 0x49, 0x54, 0x45, 0x52, 0x41, 0x4c, 0x5f, + 0x45, 0x53, 0x43, 0x41, 0x50, 0x45, 0x5f, 0x52, 0x45, 0x20, 0x3d, 0x20, + 0x2f, 0x5b, 0x5c, 0x6e, 0x5c, 0x72, 0x22, 0x5d, 0x2f, 0x67, 0x3b, 0x0a, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x47, 0x52, 0x41, 0x4d, 0x4d, 0x41, + 0x52, 0x5f, 0x52, 0x41, 0x4e, 0x47, 0x45, 0x5f, 0x4c, 0x49, 0x54, 0x45, 0x52, 0x41, 0x4c, 0x5f, 0x45, 0x53, 0x43, 0x41, 0x50, 0x45, 0x5f, 0x52, - 0x45, 0x20, 0x3d, 0x20, 0x2f, 0x5b, 0x5c, 0x6e, 0x5c, 0x72, 0x22, 0x5d, - 0x2f, 0x67, 0x3b, 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x47, 0x52, + 0x45, 0x20, 0x3d, 0x20, 0x2f, 0x5b, 0x5c, 0x6e, 0x5c, 0x72, 0x22, 0x5c, + 0x5d, 0x5c, 0x2d, 0x5c, 0x5c, 0x5d, 0x2f, 0x67, 0x3b, 0x0a, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x47, 0x52, 0x41, 0x4d, 0x4d, 0x41, 0x52, 0x5f, + 0x4c, 0x49, 0x54, 0x45, 0x52, 0x41, 0x4c, 0x5f, 0x45, 0x53, 0x43, 0x41, + 0x50, 0x45, 0x53, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x27, 0x5c, 0x72, 0x27, + 0x3a, 0x20, 0x27, 0x5c, 0x5c, 0x72, 0x27, 0x2c, 0x20, 0x27, 0x5c, 0x6e, + 0x27, 0x3a, 0x20, 0x27, 0x5c, 0x5c, 0x6e, 0x27, 0x2c, 0x20, 0x27, 0x22, + 0x27, 0x3a, 0x20, 0x27, 0x5c, 0x5c, 0x22, 0x27, 0x2c, 0x20, 0x27, 0x2d, + 0x27, 0x3a, 0x20, 0x27, 0x5c, 0x5c, 0x2d, 0x27, 0x2c, 0x20, 0x27, 0x5d, + 0x27, 0x3a, 0x20, 0x27, 0x5c, 0x5c, 0x5d, 0x27, 0x20, 0x7d, 0x3b, 0x0a, + 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x4e, 0x4f, 0x4e, 0x5f, 0x4c, + 0x49, 0x54, 0x45, 0x52, 0x41, 0x4c, 0x5f, 0x53, 0x45, 0x54, 0x20, 0x3d, + 0x20, 0x6e, 0x65, 0x77, 0x20, 0x53, 0x65, 0x74, 0x28, 0x27, 0x7c, 0x2e, + 0x28, 0x29, 0x5b, 0x5d, 0x7b, 0x7d, 0x2a, 0x2b, 0x3f, 0x27, 0x29, 0x3b, + 0x0a, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x45, 0x53, 0x43, 0x41, 0x50, + 0x45, 0x44, 0x5f, 0x49, 0x4e, 0x5f, 0x52, 0x45, 0x47, 0x45, 0x58, 0x50, + 0x53, 0x5f, 0x42, 0x55, 0x54, 0x5f, 0x4e, 0x4f, 0x54, 0x5f, 0x49, 0x4e, + 0x5f, 0x4c, 0x49, 0x54, 0x45, 0x52, 0x41, 0x4c, 0x53, 0x20, 0x3d, 0x20, + 0x6e, 0x65, 0x77, 0x20, 0x53, 0x65, 0x74, 0x28, 0x27, 0x5b, 0x5d, 0x28, + 0x29, 0x7c, 0x7b, 0x7d, 0x2a, 0x2b, 0x3f, 0x27, 0x29, 0x3b, 0x0a, 0x0a, + 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, + 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x43, 0x6f, 0x6e, 0x76, 0x65, + 0x72, 0x74, 0x65, 0x72, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x72, 0x75, 0x63, 0x74, 0x6f, 0x72, 0x28, 0x6f, 0x70, 0x74, + 0x69, 0x6f, 0x6e, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, + 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, + 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x5f, 0x6f, 0x72, 0x64, 0x65, 0x72, + 0x20, 0x7c, 0x7c, 0x20, 0x7b, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x6c, 0x6c, 0x6f, 0x77, 0x46, + 0x65, 0x74, 0x63, 0x68, 0x20, 0x3d, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, + 0x6e, 0x73, 0x2e, 0x61, 0x6c, 0x6c, 0x6f, 0x77, 0x5f, 0x66, 0x65, 0x74, + 0x63, 0x68, 0x20, 0x7c, 0x7c, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x64, + 0x6f, 0x74, 0x61, 0x6c, 0x6c, 0x20, 0x3d, 0x20, 0x6f, 0x70, 0x74, 0x69, + 0x6f, 0x6e, 0x73, 0x2e, 0x64, 0x6f, 0x74, 0x61, 0x6c, 0x6c, 0x20, 0x7c, + 0x7c, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x75, 0x6c, 0x65, 0x73, + 0x20, 0x3d, 0x20, 0x7b, 0x27, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x3a, + 0x20, 0x53, 0x50, 0x41, 0x43, 0x45, 0x5f, 0x52, 0x55, 0x4c, 0x45, 0x7d, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, + 0x72, 0x65, 0x66, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x65, 0x66, + 0x73, 0x42, 0x65, 0x69, 0x6e, 0x67, 0x52, 0x65, 0x73, 0x6f, 0x6c, 0x76, + 0x65, 0x64, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x53, 0x65, 0x74, + 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x5f, + 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x4c, 0x69, 0x74, 0x65, 0x72, 0x61, + 0x6c, 0x28, 0x6c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, + 0x73, 0x63, 0x61, 0x70, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x6c, 0x69, 0x74, + 0x65, 0x72, 0x61, 0x6c, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, + 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x47, 0x52, 0x41, 0x4d, + 0x4d, 0x41, 0x52, 0x5f, 0x4c, 0x49, 0x54, 0x45, 0x52, 0x41, 0x4c, 0x5f, + 0x45, 0x53, 0x43, 0x41, 0x50, 0x45, 0x5f, 0x52, 0x45, 0x2c, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x20, 0x3d, 0x3e, 0x20, 0x47, 0x52, 0x41, 0x4d, 0x4d, 0x41, 0x52, 0x5f, 0x4c, 0x49, 0x54, 0x45, 0x52, 0x41, - 0x4c, 0x5f, 0x45, 0x53, 0x43, 0x41, 0x50, 0x45, 0x53, 0x20, 0x3d, 0x20, - 0x7b, 0x27, 0x5c, 0x72, 0x27, 0x3a, 0x20, 0x27, 0x5c, 0x5c, 0x72, 0x27, - 0x2c, 0x20, 0x27, 0x5c, 0x6e, 0x27, 0x3a, 0x20, 0x27, 0x5c, 0x5c, 0x6e, - 0x27, 0x2c, 0x20, 0x27, 0x22, 0x27, 0x3a, 0x20, 0x27, 0x5c, 0x5c, 0x22, - 0x27, 0x7d, 0x3b, 0x0a, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, - 0x63, 0x6c, 0x61, 0x73, 0x73, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, - 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x72, 0x75, 0x63, 0x74, 0x6f, - 0x72, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x29, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, - 0x5f, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x3d, - 0x20, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x7c, - 0x7c, 0x20, 0x7b, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, - 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x75, 0x6c, 0x65, 0x73, 0x20, 0x3d, 0x20, - 0x6e, 0x65, 0x77, 0x20, 0x4d, 0x61, 0x70, 0x28, 0x29, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x75, 0x6c, - 0x65, 0x73, 0x2e, 0x73, 0x65, 0x74, 0x28, 0x27, 0x73, 0x70, 0x61, 0x63, - 0x65, 0x27, 0x2c, 0x20, 0x53, 0x50, 0x41, 0x43, 0x45, 0x5f, 0x52, 0x55, - 0x4c, 0x45, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, - 0x5f, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x4c, 0x69, 0x74, 0x65, 0x72, - 0x61, 0x6c, 0x28, 0x6c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x29, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x65, 0x73, 0x63, 0x61, 0x70, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x4a, 0x53, - 0x4f, 0x4e, 0x2e, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, - 0x28, 0x6c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x29, 0x2e, 0x72, 0x65, + 0x4c, 0x5f, 0x45, 0x53, 0x43, 0x41, 0x50, 0x45, 0x53, 0x5b, 0x6d, 0x5d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x60, 0x22, 0x24, 0x7b, 0x65, + 0x73, 0x63, 0x61, 0x70, 0x65, 0x64, 0x7d, 0x22, 0x60, 0x3b, 0x0a, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x5f, 0x66, 0x6f, 0x72, 0x6d, 0x61, + 0x74, 0x52, 0x61, 0x6e, 0x67, 0x65, 0x43, 0x68, 0x61, 0x72, 0x28, 0x6c, + 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x4a, 0x53, 0x4f, + 0x4e, 0x2e, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, + 0x6c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x29, 0x2e, 0x73, 0x6c, 0x69, + 0x63, 0x65, 0x28, 0x31, 0x2c, 0x20, 0x2d, 0x31, 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x47, 0x52, 0x41, 0x4d, 0x4d, 0x41, 0x52, 0x5f, 0x4c, 0x49, 0x54, - 0x45, 0x52, 0x41, 0x4c, 0x5f, 0x45, 0x53, 0x43, 0x41, 0x50, 0x45, 0x5f, - 0x52, 0x45, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x20, - 0x3d, 0x3e, 0x20, 0x47, 0x52, 0x41, 0x4d, 0x4d, 0x41, 0x52, 0x5f, 0x4c, - 0x49, 0x54, 0x45, 0x52, 0x41, 0x4c, 0x5f, 0x45, 0x53, 0x43, 0x41, 0x50, - 0x45, 0x53, 0x5b, 0x6d, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, - 0x60, 0x22, 0x24, 0x7b, 0x65, 0x73, 0x63, 0x61, 0x70, 0x65, 0x64, 0x7d, - 0x22, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x5f, - 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x6e, 0x61, 0x6d, 0x65, - 0x2c, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x65, 0x73, 0x63, 0x4e, 0x61, 0x6d, - 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2e, 0x72, 0x65, 0x70, - 0x6c, 0x61, 0x63, 0x65, 0x28, 0x49, 0x4e, 0x56, 0x41, 0x4c, 0x49, 0x44, - 0x5f, 0x52, 0x55, 0x4c, 0x45, 0x5f, 0x43, 0x48, 0x41, 0x52, 0x53, 0x5f, - 0x52, 0x45, 0x2c, 0x20, 0x27, 0x2d, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6b, 0x65, 0x79, 0x20, 0x3d, 0x20, - 0x65, 0x73, 0x63, 0x4e, 0x61, 0x6d, 0x65, 0x3b, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, - 0x72, 0x75, 0x6c, 0x65, 0x73, 0x2e, 0x68, 0x61, 0x73, 0x28, 0x65, 0x73, - 0x63, 0x4e, 0x61, 0x6d, 0x65, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, - 0x2e, 0x5f, 0x72, 0x75, 0x6c, 0x65, 0x73, 0x2e, 0x67, 0x65, 0x74, 0x28, - 0x65, 0x73, 0x63, 0x4e, 0x61, 0x6d, 0x65, 0x29, 0x20, 0x3d, 0x3d, 0x3d, - 0x20, 0x72, 0x75, 0x6c, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, - 0x6b, 0x65, 0x79, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, - 0x69, 0x20, 0x3d, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x77, 0x68, 0x69, 0x6c, 0x65, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, - 0x2e, 0x5f, 0x72, 0x75, 0x6c, 0x65, 0x73, 0x2e, 0x68, 0x61, 0x73, 0x28, - 0x60, 0x24, 0x7b, 0x65, 0x73, 0x63, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x24, - 0x7b, 0x69, 0x7d, 0x60, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x20, 0x2b, 0x3d, 0x20, 0x31, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x6b, 0x65, 0x79, 0x20, 0x3d, 0x20, 0x60, 0x24, 0x7b, - 0x65, 0x73, 0x63, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x24, 0x7b, 0x69, 0x7d, - 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x75, 0x6c, 0x65, - 0x73, 0x2e, 0x73, 0x65, 0x74, 0x28, 0x6b, 0x65, 0x79, 0x2c, 0x20, 0x72, - 0x75, 0x6c, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, - 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6b, 0x65, 0x79, 0x3b, 0x0a, 0x20, 0x20, - 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x76, 0x69, 0x73, 0x69, 0x74, 0x28, 0x73, - 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x29, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x20, - 0x3d, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x74, 0x79, 0x70, - 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20, - 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x7c, 0x7c, 0x20, 0x27, 0x72, 0x6f, 0x6f, - 0x74, 0x27, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, - 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x6f, 0x6e, 0x65, 0x4f, - 0x66, 0x20, 0x7c, 0x7c, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, - 0x61, 0x6e, 0x79, 0x4f, 0x66, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x75, 0x6c, - 0x65, 0x20, 0x3d, 0x20, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, - 0x6f, 0x6e, 0x65, 0x4f, 0x66, 0x20, 0x7c, 0x7c, 0x20, 0x73, 0x63, 0x68, - 0x65, 0x6d, 0x61, 0x2e, 0x61, 0x6e, 0x79, 0x4f, 0x66, 0x29, 0x2e, 0x6d, - 0x61, 0x70, 0x28, 0x28, 0x61, 0x6c, 0x74, 0x53, 0x63, 0x68, 0x65, 0x6d, - 0x61, 0x2c, 0x20, 0x69, 0x29, 0x20, 0x3d, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x76, 0x69, - 0x73, 0x69, 0x74, 0x28, 0x61, 0x6c, 0x74, 0x53, 0x63, 0x68, 0x65, 0x6d, - 0x61, 0x2c, 0x20, 0x60, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x24, - 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x20, 0x22, 0x2d, 0x22, 0x20, - 0x3a, 0x20, 0x22, 0x22, 0x7d, 0x24, 0x7b, 0x69, 0x7d, 0x60, 0x29, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, - 0x28, 0x27, 0x20, 0x7c, 0x20, 0x27, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x47, 0x52, 0x41, 0x4d, 0x4d, 0x41, 0x52, 0x5f, 0x52, 0x41, 0x4e, + 0x47, 0x45, 0x5f, 0x4c, 0x49, 0x54, 0x45, 0x52, 0x41, 0x4c, 0x5f, 0x45, + 0x53, 0x43, 0x41, 0x50, 0x45, 0x5f, 0x52, 0x45, 0x2c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x6d, 0x20, 0x3d, 0x3e, 0x20, 0x47, 0x52, 0x41, + 0x4d, 0x4d, 0x41, 0x52, 0x5f, 0x4c, 0x49, 0x54, 0x45, 0x52, 0x41, 0x4c, + 0x5f, 0x45, 0x53, 0x43, 0x41, 0x50, 0x45, 0x53, 0x5b, 0x6d, 0x5d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x6e, + 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x65, 0x73, 0x63, + 0x4e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2e, + 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x49, 0x4e, 0x56, 0x41, + 0x4c, 0x49, 0x44, 0x5f, 0x52, 0x55, 0x4c, 0x45, 0x5f, 0x43, 0x48, 0x41, + 0x52, 0x53, 0x5f, 0x52, 0x45, 0x2c, 0x20, 0x27, 0x2d, 0x27, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6b, 0x65, 0x79, + 0x20, 0x3d, 0x20, 0x65, 0x73, 0x63, 0x4e, 0x61, 0x6d, 0x65, 0x3b, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x73, 0x63, + 0x4e, 0x61, 0x6d, 0x65, 0x20, 0x69, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x5f, 0x72, 0x75, 0x6c, 0x65, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x5f, 0x72, 0x75, 0x6c, 0x65, 0x73, 0x5b, 0x65, 0x73, 0x63, + 0x4e, 0x61, 0x6d, 0x65, 0x5d, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x72, 0x75, + 0x6c, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6b, 0x65, 0x79, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x69, 0x20, 0x3d, + 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, + 0x69, 0x6c, 0x65, 0x20, 0x28, 0x28, 0x60, 0x24, 0x7b, 0x65, 0x73, 0x63, + 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x24, 0x7b, 0x69, 0x7d, 0x60, 0x20, 0x69, + 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x75, 0x6c, 0x65, + 0x73, 0x29, 0x20, 0x26, 0x26, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, + 0x5f, 0x72, 0x75, 0x6c, 0x65, 0x73, 0x5b, 0x60, 0x24, 0x7b, 0x65, 0x73, + 0x63, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x24, 0x7b, 0x69, 0x7d, 0x60, 0x5d, + 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x29, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x20, + 0x2b, 0x3d, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6b, 0x65, 0x79, 0x20, + 0x3d, 0x20, 0x60, 0x24, 0x7b, 0x65, 0x73, 0x63, 0x4e, 0x61, 0x6d, 0x65, + 0x7d, 0x24, 0x7b, 0x69, 0x7d, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, + 0x5f, 0x72, 0x75, 0x6c, 0x65, 0x73, 0x5b, 0x6b, 0x65, 0x79, 0x5d, 0x20, + 0x3d, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6b, 0x65, 0x79, 0x3b, 0x0a, + 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, + 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x52, 0x65, 0x66, 0x73, + 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2c, 0x20, 0x75, 0x72, 0x6c, + 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x76, 0x69, 0x73, 0x69, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x73, + 0x79, 0x6e, 0x63, 0x20, 0x28, 0x6e, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x41, + 0x72, 0x72, 0x61, 0x79, 0x2e, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, + 0x28, 0x6e, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x50, 0x72, + 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x2e, 0x61, 0x6c, 0x6c, 0x28, 0x6e, 0x2e, + 0x6d, 0x61, 0x70, 0x28, 0x76, 0x69, 0x73, 0x69, 0x74, 0x29, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, + 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, + 0x20, 0x6e, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x6f, 0x62, 0x6a, 0x65, + 0x63, 0x74, 0x27, 0x20, 0x26, 0x26, 0x20, 0x6e, 0x20, 0x21, 0x3d, 0x3d, + 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x72, 0x65, 0x66, + 0x20, 0x3d, 0x20, 0x6e, 0x2e, 0x24, 0x72, 0x65, 0x66, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x74, + 0x61, 0x72, 0x67, 0x65, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x66, 0x20, 0x21, + 0x3d, 0x3d, 0x20, 0x75, 0x6e, 0x64, 0x65, 0x66, 0x69, 0x6e, 0x65, 0x64, + 0x20, 0x26, 0x26, 0x20, 0x21, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, + 0x65, 0x66, 0x73, 0x5b, 0x72, 0x65, 0x66, 0x5d, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, + 0x20, 0x28, 0x72, 0x65, 0x66, 0x2e, 0x73, 0x74, 0x61, 0x72, 0x74, 0x73, + 0x57, 0x69, 0x74, 0x68, 0x28, 0x27, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, + 0x2f, 0x2f, 0x27, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, + 0x21, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x6c, 0x6c, 0x6f, 0x77, + 0x46, 0x65, 0x74, 0x63, 0x68, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, + 0x68, 0x72, 0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45, 0x72, 0x72, + 0x6f, 0x72, 0x28, 0x27, 0x46, 0x65, 0x74, 0x63, 0x68, 0x69, 0x6e, 0x67, + 0x20, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x20, 0x73, 0x63, 0x68, 0x65, + 0x6d, 0x61, 0x73, 0x20, 0x69, 0x73, 0x20, 0x6e, 0x6f, 0x74, 0x20, 0x61, + 0x6c, 0x6c, 0x6f, 0x77, 0x65, 0x64, 0x20, 0x28, 0x75, 0x73, 0x65, 0x20, + 0x2d, 0x2d, 0x61, 0x6c, 0x6c, 0x6f, 0x77, 0x2d, 0x66, 0x65, 0x74, 0x63, + 0x68, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x66, 0x6f, 0x72, 0x63, 0x65, 0x29, + 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x66, 0x65, 0x74, 0x63, 0x68, 0x20, 0x3d, 0x20, 0x28, 0x61, 0x77, 0x61, + 0x69, 0x74, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x28, 0x27, 0x6e, + 0x6f, 0x64, 0x65, 0x2d, 0x66, 0x65, 0x74, 0x63, 0x68, 0x27, 0x29, 0x29, + 0x2e, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x3b, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x72, 0x61, 0x67, 0x53, 0x70, 0x6c, + 0x69, 0x74, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x66, 0x2e, 0x73, 0x70, 0x6c, + 0x69, 0x74, 0x28, 0x27, 0x23, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x62, 0x61, 0x73, 0x65, 0x55, 0x72, 0x6c, 0x20, 0x3d, + 0x20, 0x66, 0x72, 0x61, 0x67, 0x53, 0x70, 0x6c, 0x69, 0x74, 0x5b, 0x30, + 0x5d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x65, 0x66, 0x73, 0x5b, + 0x62, 0x61, 0x73, 0x65, 0x55, 0x72, 0x6c, 0x5d, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, + 0x20, 0x28, 0x21, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, + 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, + 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x52, 0x65, 0x66, 0x73, 0x28, 0x61, + 0x77, 0x61, 0x69, 0x74, 0x20, 0x66, 0x65, 0x74, 0x63, 0x68, 0x28, 0x72, + 0x65, 0x66, 0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e, 0x28, 0x72, 0x65, 0x73, + 0x20, 0x3d, 0x3e, 0x20, 0x72, 0x65, 0x73, 0x2e, 0x6a, 0x73, 0x6f, 0x6e, + 0x28, 0x29, 0x29, 0x2c, 0x20, 0x62, 0x61, 0x73, 0x65, 0x55, 0x72, 0x6c, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, + 0x65, 0x66, 0x73, 0x5b, 0x62, 0x61, 0x73, 0x65, 0x55, 0x72, 0x6c, 0x5d, + 0x20, 0x3d, 0x20, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x66, 0x72, 0x61, 0x67, 0x53, 0x70, + 0x6c, 0x69, 0x74, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3d, + 0x3d, 0x3d, 0x20, 0x31, 0x20, 0x7c, 0x7c, 0x20, 0x66, 0x72, 0x61, 0x67, + 0x53, 0x70, 0x6c, 0x69, 0x74, 0x5b, 0x66, 0x72, 0x61, 0x67, 0x53, 0x70, + 0x6c, 0x69, 0x74, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x2d, + 0x20, 0x31, 0x5d, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x27, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, - 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, - 0x28, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x72, - 0x75, 0x6c, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, - 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x27, 0x63, 0x6f, - 0x6e, 0x73, 0x74, 0x27, 0x20, 0x69, 0x6e, 0x20, 0x73, 0x63, 0x68, 0x65, - 0x6d, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, - 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, 0x75, 0x6c, - 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, - 0x5f, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x4c, 0x69, 0x74, 0x65, 0x72, - 0x61, 0x6c, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x63, 0x6f, + 0x61, 0x72, 0x67, 0x65, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, + 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x66, 0x2e, 0x73, 0x74, + 0x61, 0x72, 0x74, 0x73, 0x57, 0x69, 0x74, 0x68, 0x28, 0x27, 0x23, 0x2f, + 0x27, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, + 0x20, 0x3d, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x66, 0x20, 0x3d, 0x20, 0x60, 0x24, 0x7b, 0x75, 0x72, 0x6c, 0x7d, + 0x24, 0x7b, 0x72, 0x65, 0x66, 0x7d, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x2e, 0x24, + 0x72, 0x65, 0x66, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x66, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, + 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, + 0x6e, 0x65, 0x77, 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x55, + 0x6e, 0x73, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x20, 0x72, + 0x65, 0x66, 0x20, 0x24, 0x7b, 0x72, 0x65, 0x66, 0x7d, 0x60, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, + 0x6f, 0x72, 0x73, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x66, 0x2e, 0x73, 0x70, + 0x6c, 0x69, 0x74, 0x28, 0x27, 0x23, 0x27, 0x29, 0x5b, 0x31, 0x5d, 0x2e, + 0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, 0x2f, 0x27, 0x29, 0x2e, 0x73, + 0x6c, 0x69, 0x63, 0x65, 0x28, 0x31, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x65, 0x6c, 0x20, 0x6f, 0x66, + 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x73, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x74, 0x61, 0x72, 0x67, 0x65, + 0x74, 0x20, 0x7c, 0x7c, 0x20, 0x21, 0x28, 0x73, 0x65, 0x6c, 0x20, 0x69, + 0x6e, 0x20, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x29, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, + 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x45, 0x72, 0x72, 0x6f, + 0x72, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x69, 0x6e, 0x67, 0x20, + 0x72, 0x65, 0x66, 0x20, 0x24, 0x7b, 0x72, 0x65, 0x66, 0x7d, 0x3a, 0x20, + 0x24, 0x7b, 0x73, 0x65, 0x6c, 0x7d, 0x20, 0x6e, 0x6f, 0x74, 0x20, 0x69, + 0x6e, 0x20, 0x24, 0x7b, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x73, 0x74, 0x72, + 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, 0x74, 0x61, 0x72, 0x67, 0x65, + 0x74, 0x29, 0x7d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x61, 0x72, + 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, + 0x5b, 0x73, 0x65, 0x6c, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, + 0x72, 0x65, 0x66, 0x73, 0x5b, 0x72, 0x65, 0x66, 0x5d, 0x20, 0x3d, 0x20, + 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, + 0x77, 0x61, 0x69, 0x74, 0x20, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, + 0x2e, 0x61, 0x6c, 0x6c, 0x28, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x73, 0x28, 0x6e, 0x29, 0x2e, 0x6d, 0x61, + 0x70, 0x28, 0x76, 0x69, 0x73, 0x69, 0x74, 0x29, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, + 0x74, 0x75, 0x72, 0x6e, 0x20, 0x76, 0x69, 0x73, 0x69, 0x74, 0x28, 0x73, + 0x63, 0x68, 0x65, 0x6d, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, + 0x0a, 0x20, 0x20, 0x5f, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x65, + 0x55, 0x6e, 0x69, 0x6f, 0x6e, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x6e, 0x61, + 0x6d, 0x65, 0x2c, 0x20, 0x61, 0x6c, 0x74, 0x53, 0x63, 0x68, 0x65, 0x6d, + 0x61, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, + 0x74, 0x75, 0x72, 0x6e, 0x20, 0x61, 0x6c, 0x74, 0x53, 0x63, 0x68, 0x65, + 0x6d, 0x61, 0x73, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x6d, + 0x61, 0x70, 0x28, 0x28, 0x61, 0x6c, 0x74, 0x53, 0x63, 0x68, 0x65, 0x6d, + 0x61, 0x2c, 0x20, 0x69, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x76, 0x69, 0x73, 0x69, 0x74, 0x28, 0x61, 0x6c, 0x74, 0x53, + 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2c, 0x20, 0x60, 0x24, 0x7b, 0x6e, 0x61, + 0x6d, 0x65, 0x20, 0x3f, 0x3f, 0x20, 0x27, 0x27, 0x7d, 0x24, 0x7b, 0x6e, + 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x20, 0x27, 0x2d, 0x27, 0x20, 0x3a, 0x20, + 0x27, 0x61, 0x6c, 0x74, 0x65, 0x72, 0x6e, 0x61, 0x74, 0x69, 0x76, 0x65, + 0x2d, 0x27, 0x7d, 0x24, 0x7b, 0x69, 0x7d, 0x60, 0x29, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, + 0x20, 0x7c, 0x20, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x5f, 0x76, 0x69, 0x73, 0x69, 0x74, 0x50, 0x61, 0x74, 0x74, + 0x65, 0x72, 0x6e, 0x28, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, 0x2c, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, + 0x6e, 0x2e, 0x73, 0x74, 0x61, 0x72, 0x74, 0x73, 0x57, 0x69, 0x74, 0x68, + 0x28, 0x27, 0x5e, 0x27, 0x29, 0x20, 0x7c, 0x7c, 0x20, 0x21, 0x70, 0x61, + 0x74, 0x74, 0x65, 0x72, 0x6e, 0x2e, 0x65, 0x6e, 0x64, 0x73, 0x57, 0x69, + 0x74, 0x68, 0x28, 0x27, 0x24, 0x27, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x6e, + 0x65, 0x77, 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x27, 0x50, 0x61, + 0x74, 0x74, 0x65, 0x72, 0x6e, 0x20, 0x6d, 0x75, 0x73, 0x74, 0x20, 0x73, + 0x74, 0x61, 0x72, 0x74, 0x20, 0x77, 0x69, 0x74, 0x68, 0x20, 0x22, 0x5e, + 0x22, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x65, 0x6e, 0x64, 0x20, 0x77, 0x69, + 0x74, 0x68, 0x20, 0x22, 0x24, 0x22, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x74, 0x74, + 0x65, 0x72, 0x6e, 0x20, 0x3d, 0x20, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, + 0x6e, 0x2e, 0x73, 0x6c, 0x69, 0x63, 0x65, 0x28, 0x31, 0x2c, 0x20, 0x2d, + 0x31, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x73, 0x75, 0x62, 0x52, 0x75, 0x6c, 0x65, 0x49, 0x64, 0x73, + 0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x6c, 0x65, 0x74, 0x20, 0x69, 0x20, 0x3d, 0x20, 0x30, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x65, 0x6e, + 0x67, 0x74, 0x68, 0x20, 0x3d, 0x20, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, + 0x6e, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3b, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x65, 0x74, + 0x44, 0x6f, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, + 0x72, 0x75, 0x6c, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x64, 0x6f, + 0x74, 0x61, 0x6c, 0x6c, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x3d, 0x20, 0x27, + 0x5b, 0x5c, 0x5c, 0x55, 0x30, 0x30, 0x30, 0x30, 0x30, 0x30, 0x30, 0x30, + 0x2d, 0x5c, 0x5c, 0x55, 0x30, 0x30, 0x31, 0x30, 0x46, 0x46, 0x46, 0x46, + 0x5d, 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, + 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x41, 0x63, 0x63, 0x65, 0x70, 0x74, + 0x20, 0x61, 0x6e, 0x79, 0x20, 0x63, 0x68, 0x61, 0x72, 0x61, 0x63, 0x74, + 0x65, 0x72, 0x2e, 0x2e, 0x2e, 0x20, 0x65, 0x78, 0x63, 0x65, 0x70, 0x74, + 0x20, 0x5c, 0x6e, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x5c, 0x72, 0x20, 0x6c, + 0x69, 0x6e, 0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x20, 0x63, 0x68, + 0x61, 0x72, 0x73, 0x20, 0x28, 0x5c, 0x78, 0x30, 0x41, 0x20, 0x61, 0x6e, + 0x64, 0x20, 0x5c, 0x78, 0x4f, 0x44, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x3d, 0x20, 0x27, + 0x5b, 0x5c, 0x5c, 0x55, 0x30, 0x30, 0x30, 0x30, 0x30, 0x30, 0x30, 0x30, + 0x2d, 0x5c, 0x5c, 0x78, 0x30, 0x39, 0x5c, 0x5c, 0x78, 0x30, 0x42, 0x5c, + 0x5c, 0x78, 0x30, 0x43, 0x5c, 0x5c, 0x78, 0x30, 0x45, 0x2d, 0x5c, 0x5c, + 0x55, 0x30, 0x30, 0x31, 0x30, 0x46, 0x46, 0x46, 0x46, 0x5d, 0x27, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, + 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, + 0x27, 0x64, 0x6f, 0x74, 0x27, 0x2c, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x6f, 0x52, + 0x75, 0x6c, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x5b, 0x73, 0x2c, 0x20, 0x69, + 0x73, 0x4c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x5d, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x69, 0x73, 0x4c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x20, + 0x3f, 0x20, 0x22, 0x5c, 0x22, 0x22, 0x20, 0x2b, 0x20, 0x73, 0x20, 0x2b, + 0x20, 0x22, 0x5c, 0x22, 0x22, 0x20, 0x3a, 0x20, 0x73, 0x3b, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x72, + 0x61, 0x6e, 0x73, 0x66, 0x6f, 0x72, 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x29, + 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x74, 0x61, 0x72, 0x74, 0x20, + 0x3d, 0x20, 0x69, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, + 0x2f, 0x20, 0x46, 0x6f, 0x72, 0x20, 0x65, 0x61, 0x63, 0x68, 0x20, 0x63, + 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x6f, 0x66, 0x20, + 0x74, 0x68, 0x69, 0x73, 0x20, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, + 0x65, 0x2c, 0x20, 0x73, 0x74, 0x6f, 0x72, 0x65, 0x20, 0x69, 0x74, 0x73, + 0x20, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x72, 0x65, 0x70, 0x72, + 0x65, 0x73, 0x65, 0x6e, 0x74, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x61, + 0x6e, 0x64, 0x20, 0x77, 0x68, 0x65, 0x74, 0x68, 0x65, 0x72, 0x20, 0x69, + 0x74, 0x27, 0x73, 0x20, 0x61, 0x20, 0x6c, 0x69, 0x74, 0x65, 0x72, 0x61, + 0x6c, 0x2e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, + 0x57, 0x65, 0x20, 0x6f, 0x6e, 0x6c, 0x79, 0x20, 0x6e, 0x65, 0x65, 0x64, + 0x20, 0x61, 0x20, 0x66, 0x6c, 0x61, 0x74, 0x20, 0x73, 0x74, 0x72, 0x75, + 0x63, 0x74, 0x75, 0x72, 0x65, 0x20, 0x68, 0x65, 0x72, 0x65, 0x20, 0x74, + 0x6f, 0x20, 0x61, 0x70, 0x70, 0x6c, 0x79, 0x20, 0x72, 0x65, 0x70, 0x65, + 0x74, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6f, 0x70, 0x65, 0x72, 0x61, + 0x74, 0x6f, 0x72, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, + 0x6c, 0x61, 0x73, 0x74, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x2c, 0x20, 0x61, + 0x6e, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, + 0x74, 0x6f, 0x20, 0x6d, 0x65, 0x72, 0x67, 0x65, 0x20, 0x6c, 0x69, 0x74, + 0x65, 0x72, 0x61, 0x6c, 0x73, 0x20, 0x61, 0x74, 0x20, 0x74, 0x68, 0x65, + 0x20, 0x61, 0x6e, 0x64, 0x20, 0x28, 0x77, 0x65, 0x27, 0x72, 0x65, 0x20, + 0x70, 0x61, 0x72, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x67, 0x72, 0x6f, 0x75, + 0x70, 0x65, 0x64, 0x20, 0x28, 0x20, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, + 0x63, 0x65, 0x73, 0x20, 0x29, 0x20, 0x72, 0x65, 0x63, 0x75, 0x72, 0x73, + 0x69, 0x76, 0x65, 0x6c, 0x79, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x64, 0x6f, + 0x6e, 0x27, 0x74, 0x20, 0x74, 0x72, 0x65, 0x61, 0x74, 0x20, 0x27, 0x7c, + 0x27, 0x20, 0x73, 0x70, 0x65, 0x63, 0x69, 0x61, 0x6c, 0x6c, 0x79, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x28, 0x47, 0x42, + 0x4e, 0x46, 0x27, 0x73, 0x20, 0x73, 0x79, 0x6e, 0x74, 0x61, 0x78, 0x20, + 0x69, 0x73, 0x20, 0x6c, 0x75, 0x63, 0x6b, 0x69, 0x6c, 0x79, 0x20, 0x76, + 0x65, 0x72, 0x79, 0x20, 0x63, 0x6c, 0x6f, 0x73, 0x65, 0x20, 0x74, 0x6f, + 0x20, 0x72, 0x65, 0x67, 0x75, 0x6c, 0x61, 0x72, 0x20, 0x65, 0x78, 0x70, + 0x72, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x73, 0x21, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, + 0x65, 0x71, 0x20, 0x3d, 0x20, 0x5b, 0x5d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6a, 0x6f, + 0x69, 0x6e, 0x53, 0x65, 0x71, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x74, 0x20, 0x3d, 0x20, + 0x5b, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x66, 0x6f, 0x72, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x5b, + 0x69, 0x73, 0x4c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x2c, 0x20, 0x67, + 0x5d, 0x20, 0x6f, 0x66, 0x20, 0x67, 0x72, 0x6f, 0x75, 0x70, 0x42, 0x79, + 0x28, 0x73, 0x65, 0x71, 0x2c, 0x20, 0x78, 0x20, 0x3d, 0x3e, 0x20, 0x78, + 0x5b, 0x31, 0x5d, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x69, 0x73, + 0x4c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x5b, 0x5b, 0x2e, 0x2e, + 0x2e, 0x67, 0x5d, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x78, 0x20, 0x3d, 0x3e, + 0x20, 0x78, 0x5b, 0x30, 0x5d, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, + 0x27, 0x27, 0x29, 0x2c, 0x20, 0x74, 0x72, 0x75, 0x65, 0x5d, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x2e, + 0x70, 0x75, 0x73, 0x68, 0x28, 0x2e, 0x2e, 0x2e, 0x67, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, + 0x74, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x31, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x72, + 0x65, 0x74, 0x5b, 0x30, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x5b, 0x72, 0x65, 0x74, + 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x78, 0x20, 0x3d, 0x3e, 0x20, 0x74, 0x6f, + 0x52, 0x75, 0x6c, 0x65, 0x28, 0x78, 0x29, 0x29, 0x2e, 0x6a, 0x6f, 0x69, + 0x6e, 0x28, 0x27, 0x20, 0x27, 0x29, 0x2c, 0x20, 0x66, 0x61, 0x6c, 0x73, + 0x65, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c, + 0x65, 0x20, 0x28, 0x69, 0x20, 0x3c, 0x20, 0x6c, 0x65, 0x6e, 0x67, 0x74, + 0x68, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x20, 0x3d, 0x20, 0x70, + 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, 0x5b, 0x69, 0x5d, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, + 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x2e, 0x27, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, + 0x71, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x5b, 0x67, 0x65, 0x74, 0x44, + 0x6f, 0x74, 0x28, 0x29, 0x2c, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x5d, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x20, 0x2b, 0x3d, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x63, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x28, + 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x20, 0x2b, 0x3d, 0x20, 0x31, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, + 0x28, 0x69, 0x20, 0x3c, 0x20, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x61, 0x74, 0x74, 0x65, + 0x72, 0x6e, 0x5b, 0x69, 0x5d, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x3f, + 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, + 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, + 0x55, 0x6e, 0x73, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x20, + 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, 0x20, 0x73, 0x79, 0x6e, 0x74, + 0x61, 0x78, 0x20, 0x22, 0x24, 0x7b, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, + 0x6e, 0x5b, 0x69, 0x5d, 0x7d, 0x22, 0x20, 0x61, 0x74, 0x20, 0x69, 0x6e, + 0x64, 0x65, 0x78, 0x20, 0x24, 0x7b, 0x69, 0x7d, 0x20, 0x6f, 0x66, 0x20, + 0x2f, 0x24, 0x7b, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, 0x7d, 0x2f, + 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x71, 0x2e, 0x70, 0x75, 0x73, 0x68, + 0x28, 0x5b, 0x60, 0x28, 0x24, 0x7b, 0x74, 0x6f, 0x52, 0x75, 0x6c, 0x65, + 0x28, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x66, 0x6f, 0x72, 0x6d, 0x28, 0x29, + 0x29, 0x7d, 0x29, 0x60, 0x2c, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x5d, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x20, + 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x29, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x20, 0x2b, + 0x3d, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x74, 0x61, 0x72, 0x74, + 0x20, 0x3c, 0x3d, 0x20, 0x30, 0x20, 0x7c, 0x7c, 0x20, 0x70, 0x61, 0x74, + 0x74, 0x65, 0x72, 0x6e, 0x5b, 0x73, 0x74, 0x61, 0x72, 0x74, 0x20, 0x2d, + 0x20, 0x31, 0x5d, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x27, 0x28, 0x27, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, + 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x55, 0x6e, 0x62, 0x61, + 0x6c, 0x61, 0x6e, 0x63, 0x65, 0x64, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, + 0x74, 0x68, 0x65, 0x73, 0x65, 0x73, 0x3b, 0x20, 0x73, 0x74, 0x61, 0x72, + 0x74, 0x20, 0x3d, 0x20, 0x24, 0x7b, 0x73, 0x74, 0x61, 0x72, 0x74, 0x7d, + 0x2c, 0x20, 0x69, 0x20, 0x3d, 0x20, 0x24, 0x7b, 0x69, 0x7d, 0x2c, 0x20, + 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, 0x20, 0x3d, 0x20, 0x24, 0x7b, + 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, 0x7d, 0x60, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, + 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6a, 0x6f, 0x69, 0x6e, 0x53, 0x65, 0x71, + 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, + 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x5b, 0x27, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, + 0x74, 0x20, 0x73, 0x71, 0x75, 0x61, 0x72, 0x65, 0x42, 0x72, 0x61, 0x63, + 0x6b, 0x65, 0x74, 0x73, 0x20, 0x3d, 0x20, 0x63, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x20, 0x2b, 0x3d, + 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c, 0x65, 0x20, 0x28, 0x69, 0x20, 0x3c, + 0x20, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x26, 0x26, 0x20, 0x70, + 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, 0x5b, 0x69, 0x5d, 0x20, 0x21, 0x3d, + 0x3d, 0x20, 0x27, 0x5d, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, + 0x28, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, 0x5b, 0x69, 0x5d, 0x20, + 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x5c, 0x5c, 0x27, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x73, 0x71, 0x75, 0x61, 0x72, 0x65, 0x42, 0x72, 0x61, 0x63, + 0x6b, 0x65, 0x74, 0x73, 0x20, 0x2b, 0x3d, 0x20, 0x70, 0x61, 0x74, 0x74, + 0x65, 0x72, 0x6e, 0x2e, 0x73, 0x6c, 0x69, 0x63, 0x65, 0x28, 0x69, 0x2c, + 0x20, 0x69, 0x20, 0x2b, 0x20, 0x32, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, + 0x20, 0x2b, 0x3d, 0x20, 0x32, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, + 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x71, 0x75, 0x61, 0x72, 0x65, + 0x42, 0x72, 0x61, 0x63, 0x6b, 0x65, 0x74, 0x73, 0x20, 0x2b, 0x3d, 0x20, + 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, 0x5b, 0x69, 0x5d, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x20, 0x2b, 0x3d, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, + 0x20, 0x28, 0x69, 0x20, 0x3e, 0x3d, 0x20, 0x6c, 0x65, 0x6e, 0x67, 0x74, + 0x68, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x6e, + 0x65, 0x77, 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x55, 0x6e, + 0x62, 0x61, 0x6c, 0x61, 0x6e, 0x63, 0x65, 0x64, 0x20, 0x73, 0x71, 0x75, + 0x61, 0x72, 0x65, 0x20, 0x62, 0x72, 0x61, 0x63, 0x6b, 0x65, 0x74, 0x73, + 0x3b, 0x20, 0x73, 0x74, 0x61, 0x72, 0x74, 0x20, 0x3d, 0x20, 0x24, 0x7b, + 0x73, 0x74, 0x61, 0x72, 0x74, 0x7d, 0x2c, 0x20, 0x69, 0x20, 0x3d, 0x20, + 0x24, 0x7b, 0x69, 0x7d, 0x2c, 0x20, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, + 0x6e, 0x20, 0x3d, 0x20, 0x24, 0x7b, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, + 0x6e, 0x7d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x73, 0x71, 0x75, 0x61, 0x72, 0x65, 0x42, 0x72, + 0x61, 0x63, 0x6b, 0x65, 0x74, 0x73, 0x20, 0x2b, 0x3d, 0x20, 0x27, 0x5d, + 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x20, 0x2b, 0x3d, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x71, 0x2e, 0x70, + 0x75, 0x73, 0x68, 0x28, 0x5b, 0x73, 0x71, 0x75, 0x61, 0x72, 0x65, 0x42, + 0x72, 0x61, 0x63, 0x6b, 0x65, 0x74, 0x73, 0x2c, 0x20, 0x66, 0x61, 0x6c, + 0x73, 0x65, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, + 0x28, 0x63, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x7c, 0x27, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x73, 0x65, 0x71, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x5b, 0x27, 0x7c, + 0x27, 0x2c, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x5d, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x20, + 0x2b, 0x3d, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, + 0x28, 0x63, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x2a, 0x27, 0x20, 0x7c, + 0x7c, 0x20, 0x63, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x2b, 0x27, 0x20, + 0x7c, 0x7c, 0x20, 0x63, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x3f, 0x27, + 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x73, 0x65, 0x71, 0x5b, 0x73, 0x65, 0x71, 0x2e, 0x6c, 0x65, + 0x6e, 0x67, 0x74, 0x68, 0x20, 0x2d, 0x20, 0x31, 0x5d, 0x20, 0x3d, 0x20, + 0x5b, 0x74, 0x6f, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x73, 0x65, 0x71, 0x5b, + 0x73, 0x65, 0x71, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x2d, + 0x20, 0x31, 0x5d, 0x29, 0x20, 0x2b, 0x20, 0x63, 0x2c, 0x20, 0x66, 0x61, + 0x6c, 0x73, 0x65, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x20, 0x2b, 0x3d, 0x20, 0x31, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, + 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x27, 0x7b, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x75, + 0x72, 0x6c, 0x79, 0x42, 0x72, 0x61, 0x63, 0x6b, 0x65, 0x74, 0x73, 0x20, + 0x3d, 0x20, 0x63, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x20, 0x2b, 0x3d, 0x20, 0x31, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, + 0x6c, 0x65, 0x20, 0x28, 0x69, 0x20, 0x3c, 0x20, 0x6c, 0x65, 0x6e, 0x67, + 0x74, 0x68, 0x20, 0x26, 0x26, 0x20, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, + 0x6e, 0x5b, 0x69, 0x5d, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x27, 0x7d, 0x27, + 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x6c, 0x79, 0x42, 0x72, 0x61, + 0x63, 0x6b, 0x65, 0x74, 0x73, 0x20, 0x2b, 0x3d, 0x20, 0x70, 0x61, 0x74, + 0x74, 0x65, 0x72, 0x6e, 0x5b, 0x69, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x20, 0x2b, + 0x3d, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x69, 0x20, 0x3e, 0x3d, 0x20, + 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, + 0x72, 0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45, 0x72, 0x72, 0x6f, + 0x72, 0x28, 0x60, 0x55, 0x6e, 0x62, 0x61, 0x6c, 0x61, 0x6e, 0x63, 0x65, + 0x64, 0x20, 0x63, 0x75, 0x72, 0x6c, 0x79, 0x20, 0x62, 0x72, 0x61, 0x63, + 0x6b, 0x65, 0x74, 0x73, 0x3b, 0x20, 0x73, 0x74, 0x61, 0x72, 0x74, 0x20, + 0x3d, 0x20, 0x24, 0x7b, 0x73, 0x74, 0x61, 0x72, 0x74, 0x7d, 0x2c, 0x20, + 0x69, 0x20, 0x3d, 0x20, 0x24, 0x7b, 0x69, 0x7d, 0x2c, 0x20, 0x70, 0x61, + 0x74, 0x74, 0x65, 0x72, 0x6e, 0x20, 0x3d, 0x20, 0x24, 0x7b, 0x70, 0x61, + 0x74, 0x74, 0x65, 0x72, 0x6e, 0x7d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x6c, + 0x79, 0x42, 0x72, 0x61, 0x63, 0x6b, 0x65, 0x74, 0x73, 0x20, 0x2b, 0x3d, + 0x20, 0x27, 0x7d, 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x20, 0x2b, 0x3d, 0x20, 0x31, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x6e, 0x75, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x63, + 0x75, 0x72, 0x6c, 0x79, 0x42, 0x72, 0x61, 0x63, 0x6b, 0x65, 0x74, 0x73, + 0x2e, 0x73, 0x6c, 0x69, 0x63, 0x65, 0x28, 0x31, 0x2c, 0x20, 0x2d, 0x31, + 0x29, 0x2e, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, 0x2c, 0x27, 0x29, + 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x73, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x2e, + 0x74, 0x72, 0x69, 0x6d, 0x28, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6d, + 0x69, 0x6e, 0x54, 0x69, 0x6d, 0x65, 0x73, 0x2c, 0x20, 0x6d, 0x61, 0x78, + 0x54, 0x69, 0x6d, 0x65, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6e, 0x75, 0x6d, + 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x31, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x6e, 0x54, 0x69, 0x6d, + 0x65, 0x73, 0x20, 0x3d, 0x20, 0x70, 0x61, 0x72, 0x73, 0x65, 0x49, 0x6e, + 0x74, 0x28, 0x6e, 0x75, 0x6d, 0x73, 0x5b, 0x30, 0x5d, 0x2c, 0x20, 0x31, + 0x30, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x6d, 0x61, 0x78, 0x54, 0x69, 0x6d, 0x65, 0x73, + 0x20, 0x3d, 0x20, 0x6d, 0x69, 0x6e, 0x54, 0x69, 0x6d, 0x65, 0x73, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, + 0x6e, 0x75, 0x6d, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, + 0x21, 0x3d, 0x3d, 0x20, 0x32, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, + 0x68, 0x72, 0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45, 0x72, 0x72, + 0x6f, 0x72, 0x28, 0x60, 0x49, 0x6e, 0x76, 0x61, 0x6c, 0x69, 0x64, 0x20, + 0x71, 0x75, 0x61, 0x6e, 0x74, 0x69, 0x66, 0x69, 0x65, 0x72, 0x20, 0x24, + 0x7b, 0x63, 0x75, 0x72, 0x6c, 0x79, 0x42, 0x72, 0x61, 0x63, 0x6b, 0x65, + 0x74, 0x73, 0x7d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x6e, + 0x54, 0x69, 0x6d, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6d, 0x73, + 0x5b, 0x30, 0x5d, 0x20, 0x3f, 0x20, 0x70, 0x61, 0x72, 0x73, 0x65, 0x49, + 0x6e, 0x74, 0x28, 0x6e, 0x75, 0x6d, 0x73, 0x5b, 0x30, 0x5d, 0x2c, 0x20, + 0x31, 0x30, 0x29, 0x20, 0x3a, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x61, 0x78, + 0x54, 0x69, 0x6d, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6d, 0x73, + 0x5b, 0x31, 0x5d, 0x20, 0x3f, 0x20, 0x70, 0x61, 0x72, 0x73, 0x65, 0x49, + 0x6e, 0x74, 0x28, 0x6e, 0x75, 0x6d, 0x73, 0x5b, 0x31, 0x5d, 0x2c, 0x20, + 0x31, 0x30, 0x29, 0x20, 0x3a, 0x20, 0x49, 0x6e, 0x66, 0x69, 0x6e, 0x69, + 0x74, 0x79, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x5b, 0x73, 0x75, 0x62, 0x2c, + 0x20, 0x73, 0x75, 0x62, 0x49, 0x73, 0x4c, 0x69, 0x74, 0x65, 0x72, 0x61, + 0x6c, 0x5d, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x71, 0x5b, 0x73, 0x65, 0x71, + 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x2d, 0x20, 0x31, 0x5d, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x66, 0x20, 0x28, 0x6d, 0x69, 0x6e, 0x54, 0x69, 0x6d, 0x65, + 0x73, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x30, 0x20, 0x26, 0x26, 0x20, 0x6d, + 0x61, 0x78, 0x54, 0x69, 0x6d, 0x65, 0x73, 0x20, 0x3d, 0x3d, 0x3d, 0x20, + 0x49, 0x6e, 0x66, 0x69, 0x6e, 0x69, 0x74, 0x79, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x73, 0x65, 0x71, 0x5b, 0x73, 0x65, 0x71, 0x2e, 0x6c, 0x65, 0x6e, 0x67, + 0x74, 0x68, 0x20, 0x2d, 0x20, 0x31, 0x5d, 0x20, 0x3d, 0x20, 0x5b, 0x60, + 0x24, 0x7b, 0x73, 0x75, 0x62, 0x7d, 0x2a, 0x60, 0x2c, 0x20, 0x66, 0x61, + 0x6c, 0x73, 0x65, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, + 0x66, 0x20, 0x28, 0x6d, 0x69, 0x6e, 0x54, 0x69, 0x6d, 0x65, 0x73, 0x20, + 0x3d, 0x3d, 0x3d, 0x20, 0x30, 0x20, 0x26, 0x26, 0x20, 0x6d, 0x61, 0x78, + 0x54, 0x69, 0x6d, 0x65, 0x73, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x31, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x73, 0x65, 0x71, 0x5b, 0x73, 0x65, 0x71, 0x2e, 0x6c, + 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x2d, 0x20, 0x31, 0x5d, 0x20, 0x3d, + 0x20, 0x5b, 0x60, 0x24, 0x7b, 0x73, 0x75, 0x62, 0x7d, 0x3f, 0x60, 0x2c, + 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, + 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6d, 0x69, 0x6e, 0x54, 0x69, 0x6d, + 0x65, 0x73, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x31, 0x20, 0x26, 0x26, 0x20, + 0x6d, 0x61, 0x78, 0x54, 0x69, 0x6d, 0x65, 0x73, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x49, 0x6e, 0x66, 0x69, 0x6e, 0x69, 0x74, 0x79, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x73, 0x65, 0x71, 0x5b, 0x73, 0x65, 0x71, 0x2e, 0x6c, 0x65, 0x6e, + 0x67, 0x74, 0x68, 0x20, 0x2d, 0x20, 0x31, 0x5d, 0x20, 0x3d, 0x20, 0x5b, + 0x60, 0x24, 0x7b, 0x73, 0x75, 0x62, 0x7d, 0x2b, 0x60, 0x2c, 0x20, 0x66, + 0x61, 0x6c, 0x73, 0x65, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x73, 0x75, 0x62, 0x49, 0x73, + 0x4c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x6c, 0x65, 0x74, 0x20, 0x69, 0x64, 0x20, 0x3d, 0x20, 0x73, 0x75, + 0x62, 0x52, 0x75, 0x6c, 0x65, 0x49, 0x64, 0x73, 0x5b, 0x73, 0x75, 0x62, + 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x69, 0x64, 0x20, + 0x3d, 0x3d, 0x3d, 0x20, 0x75, 0x6e, 0x64, 0x65, 0x66, 0x69, 0x6e, 0x65, + 0x64, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x64, 0x20, + 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, + 0x75, 0x6c, 0x65, 0x28, 0x60, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, + 0x2d, 0x24, 0x7b, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x2e, 0x6b, 0x65, + 0x79, 0x73, 0x28, 0x73, 0x75, 0x62, 0x52, 0x75, 0x6c, 0x65, 0x49, 0x64, + 0x73, 0x29, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x2b, 0x20, + 0x31, 0x7d, 0x60, 0x2c, 0x20, 0x73, 0x75, 0x62, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x73, 0x75, 0x62, 0x52, 0x75, 0x6c, 0x65, 0x49, 0x64, + 0x73, 0x5b, 0x73, 0x75, 0x62, 0x5d, 0x20, 0x3d, 0x20, 0x69, 0x64, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x75, 0x62, 0x20, 0x3d, + 0x20, 0x69, 0x64, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x65, 0x64, 0x53, 0x75, + 0x62, 0x20, 0x3d, 0x20, 0x41, 0x72, 0x72, 0x61, 0x79, 0x2e, 0x66, 0x72, + 0x6f, 0x6d, 0x28, 0x7b, 0x20, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x3a, + 0x20, 0x6d, 0x69, 0x6e, 0x54, 0x69, 0x6d, 0x65, 0x73, 0x20, 0x7d, 0x2c, + 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x75, 0x62, 0x49, 0x73, + 0x4c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x20, 0x3f, 0x20, 0x60, 0x22, + 0x24, 0x7b, 0x73, 0x75, 0x62, 0x2e, 0x73, 0x6c, 0x69, 0x63, 0x65, 0x28, + 0x31, 0x2c, 0x20, 0x2d, 0x31, 0x29, 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, + 0x74, 0x28, 0x6d, 0x69, 0x6e, 0x54, 0x69, 0x6d, 0x65, 0x73, 0x29, 0x7d, + 0x22, 0x60, 0x20, 0x3a, 0x20, 0x73, 0x75, 0x62, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x61, + 0x6c, 0x53, 0x75, 0x62, 0x20, 0x3d, 0x20, 0x6d, 0x61, 0x78, 0x54, 0x69, + 0x6d, 0x65, 0x73, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x75, 0x6e, 0x64, 0x65, + 0x66, 0x69, 0x6e, 0x65, 0x64, 0x20, 0x3f, 0x20, 0x41, 0x72, 0x72, 0x61, + 0x79, 0x2e, 0x66, 0x72, 0x6f, 0x6d, 0x28, 0x7b, 0x20, 0x6c, 0x65, 0x6e, + 0x67, 0x74, 0x68, 0x3a, 0x20, 0x6d, 0x61, 0x78, 0x54, 0x69, 0x6d, 0x65, + 0x73, 0x20, 0x2d, 0x20, 0x6d, 0x69, 0x6e, 0x54, 0x69, 0x6d, 0x65, 0x73, + 0x20, 0x7d, 0x2c, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x60, 0x24, + 0x7b, 0x73, 0x75, 0x62, 0x7d, 0x3f, 0x60, 0x29, 0x20, 0x3a, 0x20, 0x5b, + 0x60, 0x24, 0x7b, 0x73, 0x75, 0x62, 0x7d, 0x2a, 0x60, 0x5d, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x73, 0x65, 0x71, 0x5b, 0x73, 0x65, 0x71, 0x2e, 0x6c, 0x65, 0x6e, 0x67, + 0x74, 0x68, 0x20, 0x2d, 0x20, 0x31, 0x5d, 0x20, 0x3d, 0x20, 0x5b, 0x72, + 0x65, 0x70, 0x65, 0x61, 0x74, 0x65, 0x64, 0x53, 0x75, 0x62, 0x2e, 0x63, + 0x6f, 0x6e, 0x63, 0x61, 0x74, 0x28, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, + 0x61, 0x6c, 0x53, 0x75, 0x62, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, + 0x27, 0x20, 0x27, 0x29, 0x2c, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x5d, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, + 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6c, 0x69, 0x74, + 0x65, 0x72, 0x61, 0x6c, 0x20, 0x3d, 0x20, 0x27, 0x27, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, + 0x6c, 0x65, 0x20, 0x28, 0x69, 0x20, 0x3c, 0x20, 0x6c, 0x65, 0x6e, 0x67, + 0x74, 0x68, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x61, + 0x74, 0x74, 0x65, 0x72, 0x6e, 0x5b, 0x69, 0x5d, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x27, 0x5c, 0x5c, 0x27, 0x20, 0x26, 0x26, 0x20, 0x69, 0x20, 0x3c, + 0x20, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x2d, 0x20, 0x31, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6e, + 0x65, 0x78, 0x74, 0x20, 0x3d, 0x20, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, + 0x6e, 0x5b, 0x69, 0x20, 0x2b, 0x20, 0x31, 0x5d, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x45, 0x53, 0x43, 0x41, 0x50, 0x45, 0x44, 0x5f, + 0x49, 0x4e, 0x5f, 0x52, 0x45, 0x47, 0x45, 0x58, 0x50, 0x53, 0x5f, 0x42, + 0x55, 0x54, 0x5f, 0x4e, 0x4f, 0x54, 0x5f, 0x49, 0x4e, 0x5f, 0x4c, 0x49, + 0x54, 0x45, 0x52, 0x41, 0x4c, 0x53, 0x2e, 0x68, 0x61, 0x73, 0x28, 0x6e, + 0x65, 0x78, 0x74, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x20, 0x2b, 0x3d, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x6c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x20, 0x2b, 0x3d, 0x20, 0x70, + 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, 0x5b, 0x69, 0x5d, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x20, 0x2b, 0x3d, 0x20, 0x31, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x6c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, 0x20, 0x2b, 0x3d, + 0x20, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, 0x2e, 0x73, 0x6c, 0x69, + 0x63, 0x65, 0x28, 0x69, 0x2c, 0x20, 0x69, 0x20, 0x2b, 0x20, 0x32, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x20, 0x2b, 0x3d, 0x20, 0x32, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, + 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, + 0x5b, 0x69, 0x5d, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x22, 0x27, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, + 0x20, 0x2b, 0x3d, 0x20, 0x27, 0x5c, 0x5c, 0x22, 0x27, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x20, 0x2b, 0x3d, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, + 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x4e, 0x4f, 0x4e, + 0x5f, 0x4c, 0x49, 0x54, 0x45, 0x52, 0x41, 0x4c, 0x5f, 0x53, 0x45, 0x54, + 0x2e, 0x68, 0x61, 0x73, 0x28, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, + 0x5b, 0x69, 0x5d, 0x29, 0x20, 0x26, 0x26, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x28, 0x69, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x6c, 0x65, 0x6e, 0x67, 0x74, + 0x68, 0x20, 0x2d, 0x20, 0x31, 0x20, 0x7c, 0x7c, 0x20, 0x6c, 0x69, 0x74, + 0x65, 0x72, 0x61, 0x6c, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x27, 0x20, + 0x7c, 0x7c, 0x20, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, 0x5b, 0x69, + 0x20, 0x2b, 0x20, 0x31, 0x5d, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x2e, + 0x27, 0x20, 0x7c, 0x7c, 0x20, 0x21, 0x4e, 0x4f, 0x4e, 0x5f, 0x4c, 0x49, + 0x54, 0x45, 0x52, 0x41, 0x4c, 0x5f, 0x53, 0x45, 0x54, 0x2e, 0x68, 0x61, + 0x73, 0x28, 0x70, 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, 0x5b, 0x69, 0x2b, + 0x31, 0x5d, 0x29, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x69, + 0x74, 0x65, 0x72, 0x61, 0x6c, 0x20, 0x2b, 0x3d, 0x20, 0x70, 0x61, 0x74, + 0x74, 0x65, 0x72, 0x6e, 0x5b, 0x69, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, + 0x20, 0x2b, 0x3d, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, + 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x66, 0x20, 0x28, 0x6c, 0x69, 0x74, 0x65, 0x72, 0x61, 0x6c, + 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x27, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, + 0x65, 0x71, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x5b, 0x6c, 0x69, 0x74, + 0x65, 0x72, 0x61, 0x6c, 0x2c, 0x20, 0x74, 0x72, 0x75, 0x65, 0x5d, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6a, 0x6f, + 0x69, 0x6e, 0x53, 0x65, 0x71, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, + 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, + 0x22, 0x5c, 0x22, 0x5c, 0x5c, 0x5c, 0x22, 0x5c, 0x22, 0x20, 0x22, 0x20, + 0x2b, 0x20, 0x74, 0x6f, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x74, 0x72, 0x61, + 0x6e, 0x73, 0x66, 0x6f, 0x72, 0x6d, 0x28, 0x29, 0x29, 0x20, 0x2b, 0x20, + 0x22, 0x20, 0x5c, 0x22, 0x5c, 0x5c, 0x5c, 0x22, 0x5c, 0x22, 0x20, 0x73, + 0x70, 0x61, 0x63, 0x65, 0x22, 0x29, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x5f, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x52, 0x65, + 0x66, 0x28, 0x72, 0x65, 0x66, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x6c, 0x65, 0x74, 0x20, 0x72, 0x65, 0x66, 0x4e, 0x61, 0x6d, 0x65, + 0x20, 0x3d, 0x20, 0x72, 0x65, 0x66, 0x2e, 0x73, 0x70, 0x6c, 0x69, 0x74, + 0x28, 0x27, 0x2f, 0x27, 0x29, 0x2e, 0x70, 0x6f, 0x70, 0x28, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x28, 0x72, + 0x65, 0x66, 0x4e, 0x61, 0x6d, 0x65, 0x20, 0x69, 0x6e, 0x20, 0x74, 0x68, + 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x75, 0x6c, 0x65, 0x73, 0x29, 0x20, 0x26, + 0x26, 0x20, 0x21, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x65, 0x66, + 0x73, 0x42, 0x65, 0x69, 0x6e, 0x67, 0x52, 0x65, 0x73, 0x6f, 0x6c, 0x76, + 0x65, 0x64, 0x2e, 0x68, 0x61, 0x73, 0x28, 0x72, 0x65, 0x66, 0x29, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x5f, 0x72, 0x65, 0x66, 0x73, 0x42, 0x65, 0x69, 0x6e, 0x67, + 0x52, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x64, 0x2e, 0x61, 0x64, 0x64, + 0x28, 0x72, 0x65, 0x66, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x73, 0x6f, 0x6c, + 0x76, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, + 0x72, 0x65, 0x66, 0x73, 0x5b, 0x72, 0x65, 0x66, 0x5d, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x66, 0x4e, 0x61, 0x6d, 0x65, + 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x76, 0x69, 0x73, 0x69, + 0x74, 0x28, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x64, 0x2c, 0x20, + 0x72, 0x65, 0x66, 0x4e, 0x61, 0x6d, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x65, + 0x66, 0x73, 0x42, 0x65, 0x69, 0x6e, 0x67, 0x52, 0x65, 0x73, 0x6f, 0x6c, + 0x76, 0x65, 0x64, 0x2e, 0x64, 0x65, 0x6c, 0x65, 0x74, 0x65, 0x28, 0x72, + 0x65, 0x66, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x72, 0x65, + 0x66, 0x4e, 0x61, 0x6d, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x5f, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x65, 0x43, + 0x6f, 0x6e, 0x73, 0x74, 0x61, 0x6e, 0x74, 0x52, 0x75, 0x6c, 0x65, 0x28, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x5f, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x4c, 0x69, 0x74, 0x65, + 0x72, 0x61, 0x6c, 0x28, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x73, 0x74, 0x72, + 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x76, + 0x69, 0x73, 0x69, 0x74, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2c, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, + 0x61, 0x54, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x20, 0x73, 0x63, 0x68, 0x65, + 0x6d, 0x61, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, + 0x61, 0x46, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x20, 0x3d, 0x20, 0x73, 0x63, + 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, + 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x61, + 0x6d, 0x65, 0x20, 0x69, 0x6e, 0x20, 0x52, 0x45, 0x53, 0x45, 0x52, 0x56, + 0x45, 0x44, 0x5f, 0x4e, 0x41, 0x4d, 0x45, 0x53, 0x20, 0x3f, 0x20, 0x6e, + 0x61, 0x6d, 0x65, 0x20, 0x2b, 0x20, 0x27, 0x2d, 0x27, 0x20, 0x3a, 0x20, + 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x27, 0x27, 0x20, 0x3f, + 0x20, 0x27, 0x72, 0x6f, 0x6f, 0x74, 0x27, 0x20, 0x3a, 0x20, 0x6e, 0x61, + 0x6d, 0x65, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x72, 0x65, 0x66, 0x20, 0x3d, 0x20, 0x73, 0x63, 0x68, + 0x65, 0x6d, 0x61, 0x2e, 0x24, 0x72, 0x65, 0x66, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x66, 0x20, 0x21, 0x3d, + 0x3d, 0x20, 0x75, 0x6e, 0x64, 0x65, 0x66, 0x69, 0x6e, 0x65, 0x64, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, + 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, + 0x6d, 0x65, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x65, + 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x52, 0x65, 0x66, 0x28, 0x72, 0x65, 0x66, + 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, + 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, + 0x61, 0x2e, 0x6f, 0x6e, 0x65, 0x4f, 0x66, 0x20, 0x7c, 0x7c, 0x20, 0x73, + 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x61, 0x6e, 0x79, 0x4f, 0x66, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, + 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, + 0x6d, 0x65, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x67, 0x65, + 0x6e, 0x65, 0x72, 0x61, 0x74, 0x65, 0x55, 0x6e, 0x69, 0x6f, 0x6e, 0x52, + 0x75, 0x6c, 0x65, 0x28, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x73, 0x63, + 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x6f, 0x6e, 0x65, 0x4f, 0x66, 0x20, 0x7c, + 0x7c, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x61, 0x6e, 0x79, + 0x4f, 0x66, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, + 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x41, 0x72, 0x72, + 0x61, 0x79, 0x2e, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x28, 0x73, + 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x29, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, + 0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, + 0x65, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x67, 0x65, 0x6e, + 0x65, 0x72, 0x61, 0x74, 0x65, 0x55, 0x6e, 0x69, 0x6f, 0x6e, 0x52, 0x75, + 0x6c, 0x65, 0x28, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x73, 0x63, 0x68, + 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x2e, 0x6d, 0x61, 0x70, 0x28, + 0x74, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x7b, 0x20, 0x74, 0x79, 0x70, 0x65, + 0x3a, 0x20, 0x74, 0x20, 0x7d, 0x29, 0x29, 0x29, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, + 0x20, 0x28, 0x27, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x27, 0x20, 0x69, 0x6e, + 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, + 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, + 0x65, 0x28, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, + 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, + 0x74, 0x65, 0x43, 0x6f, 0x6e, 0x73, 0x74, 0x61, 0x6e, 0x74, 0x52, 0x75, + 0x6c, 0x65, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x27, 0x65, 0x6e, 0x75, 0x6d, 0x27, 0x20, 0x69, 0x6e, 0x20, 0x73, 0x63, 0x68, 0x65, @@ -154,158 +1029,616 @@ unsigned char json_schema_to_grammar_mjs[] = { 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x3d, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x65, 0x6e, 0x75, 0x6d, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x76, 0x20, 0x3d, 0x3e, 0x20, 0x74, 0x68, - 0x69, 0x73, 0x2e, 0x5f, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x4c, 0x69, - 0x74, 0x65, 0x72, 0x61, 0x6c, 0x28, 0x76, 0x29, 0x29, 0x2e, 0x6a, 0x6f, - 0x69, 0x6e, 0x28, 0x27, 0x20, 0x7c, 0x20, 0x27, 0x29, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, - 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, - 0x65, 0x28, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, - 0x72, 0x75, 0x6c, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x63, + 0x69, 0x73, 0x2e, 0x5f, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x65, + 0x43, 0x6f, 0x6e, 0x73, 0x74, 0x61, 0x6e, 0x74, 0x52, 0x75, 0x6c, 0x65, + 0x28, 0x76, 0x29, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x20, + 0x7c, 0x20, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, + 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, 0x75, 0x6c, + 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, + 0x20, 0x69, 0x66, 0x20, 0x28, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, + 0x54, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x75, 0x6e, 0x64, + 0x65, 0x66, 0x69, 0x6e, 0x65, 0x64, 0x20, 0x7c, 0x7c, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, - 0x20, 0x27, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x27, 0x20, 0x26, 0x26, - 0x20, 0x27, 0x70, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, + 0x20, 0x27, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x27, 0x29, 0x20, 0x26, + 0x26, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x28, 0x27, 0x70, 0x72, 0x6f, 0x70, 0x65, + 0x72, 0x74, 0x69, 0x65, 0x73, 0x27, 0x20, 0x69, 0x6e, 0x20, 0x73, 0x63, + 0x68, 0x65, 0x6d, 0x61, 0x20, 0x7c, 0x7c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x28, 0x27, 0x61, 0x64, 0x64, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, + 0x50, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, 0x27, 0x20, + 0x69, 0x6e, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x20, 0x26, 0x26, + 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x61, 0x64, 0x64, 0x69, + 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x70, 0x65, 0x72, + 0x74, 0x69, 0x65, 0x73, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x74, 0x72, 0x75, + 0x65, 0x29, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x71, 0x75, 0x69, + 0x72, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x53, 0x65, + 0x74, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x72, 0x65, 0x71, + 0x75, 0x69, 0x72, 0x65, 0x64, 0x20, 0x7c, 0x7c, 0x20, 0x5b, 0x5d, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, + 0x20, 0x3d, 0x20, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x2e, 0x65, 0x6e, + 0x74, 0x72, 0x69, 0x65, 0x73, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, + 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, 0x20, + 0x3f, 0x3f, 0x20, 0x7b, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, + 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x5f, 0x62, 0x75, 0x69, 0x6c, 0x64, 0x4f, 0x62, 0x6a, 0x65, + 0x63, 0x74, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x65, + 0x72, 0x74, 0x69, 0x65, 0x73, 0x2c, 0x20, 0x72, 0x65, 0x71, 0x75, 0x69, + 0x72, 0x65, 0x64, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x73, + 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x61, 0x64, 0x64, 0x69, 0x74, 0x69, + 0x6f, 0x6e, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, + 0x65, 0x73, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, + 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x28, 0x73, 0x63, + 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x75, 0x6e, 0x64, 0x65, 0x66, 0x69, 0x6e, 0x65, 0x64, 0x20, 0x7c, + 0x7c, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, + 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, + 0x27, 0x29, 0x20, 0x26, 0x26, 0x20, 0x27, 0x61, 0x6c, 0x6c, 0x4f, 0x66, 0x27, 0x20, 0x69, 0x6e, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x29, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, - 0x54, 0x4f, 0x44, 0x4f, 0x3a, 0x20, 0x60, 0x72, 0x65, 0x71, 0x75, 0x69, - 0x72, 0x65, 0x64, 0x60, 0x20, 0x6b, 0x65, 0x79, 0x77, 0x6f, 0x72, 0x64, - 0x20, 0x28, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x70, 0x79, 0x74, 0x68, 0x6f, - 0x6e, 0x20, 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x61, - 0x74, 0x69, 0x6f, 0x6e, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, - 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, - 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, - 0x72, 0x6f, 0x70, 0x50, 0x61, 0x69, 0x72, 0x73, 0x20, 0x3d, 0x20, 0x4f, - 0x62, 0x6a, 0x65, 0x63, 0x74, 0x2e, 0x65, 0x6e, 0x74, 0x72, 0x69, 0x65, - 0x73, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x70, 0x72, 0x6f, - 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, 0x29, 0x2e, 0x73, 0x6f, 0x72, - 0x74, 0x28, 0x28, 0x61, 0x2c, 0x20, 0x62, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, - 0x20, 0x73, 0x6f, 0x72, 0x74, 0x20, 0x62, 0x79, 0x20, 0x70, 0x6f, 0x73, - 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x70, 0x72, 0x6f, - 0x70, 0x5f, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x28, 0x69, 0x66, 0x20, - 0x73, 0x70, 0x65, 0x63, 0x69, 0x66, 0x69, 0x65, 0x64, 0x29, 0x20, 0x74, - 0x68, 0x65, 0x6e, 0x20, 0x62, 0x79, 0x20, 0x6b, 0x65, 0x79, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x41, 0x20, 0x3d, 0x20, 0x74, 0x79, - 0x70, 0x65, 0x6f, 0x66, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, - 0x65, 0x72, 0x5b, 0x61, 0x5b, 0x30, 0x5d, 0x5d, 0x20, 0x3d, 0x3d, 0x3d, - 0x20, 0x27, 0x6e, 0x75, 0x6d, 0x62, 0x65, 0x72, 0x27, 0x20, 0x3f, 0x20, - 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x5b, 0x61, 0x5b, - 0x30, 0x5d, 0x5d, 0x20, 0x3a, 0x20, 0x49, 0x6e, 0x66, 0x69, 0x6e, 0x69, - 0x74, 0x79, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x42, - 0x20, 0x3d, 0x20, 0x74, 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x70, 0x72, - 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x5b, 0x62, 0x5b, 0x30, 0x5d, - 0x5d, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x6e, 0x75, 0x6d, 0x62, 0x65, - 0x72, 0x27, 0x20, 0x3f, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, - 0x65, 0x72, 0x5b, 0x62, 0x5b, 0x30, 0x5d, 0x5d, 0x20, 0x3a, 0x20, 0x49, - 0x6e, 0x66, 0x69, 0x6e, 0x69, 0x74, 0x79, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, - 0x6f, 0x72, 0x64, 0x65, 0x72, 0x41, 0x20, 0x2d, 0x20, 0x6f, 0x72, 0x64, - 0x65, 0x72, 0x42, 0x20, 0x7c, 0x7c, 0x20, 0x61, 0x5b, 0x30, 0x5d, 0x2e, - 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x65, 0x43, 0x6f, 0x6d, 0x70, 0x61, 0x72, - 0x65, 0x28, 0x62, 0x5b, 0x30, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x3d, - 0x20, 0x27, 0x22, 0x7b, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72, 0x6f, 0x70, - 0x50, 0x61, 0x69, 0x72, 0x73, 0x2e, 0x66, 0x6f, 0x72, 0x45, 0x61, 0x63, - 0x68, 0x28, 0x28, 0x5b, 0x70, 0x72, 0x6f, 0x70, 0x4e, 0x61, 0x6d, 0x65, - 0x2c, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, - 0x5d, 0x2c, 0x20, 0x69, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x70, 0x72, 0x6f, 0x70, 0x52, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, - 0x65, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x76, 0x69, 0x73, - 0x69, 0x74, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x53, 0x63, 0x68, 0x65, 0x6d, - 0x61, 0x2c, 0x20, 0x60, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x24, - 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x20, 0x22, 0x2d, 0x22, 0x20, - 0x3a, 0x20, 0x22, 0x22, 0x7d, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x4e, - 0x61, 0x6d, 0x65, 0x7d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x69, 0x20, 0x3e, 0x20, - 0x30, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x2b, 0x3d, 0x20, 0x27, - 0x20, 0x22, 0x2c, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, - 0x2b, 0x3d, 0x20, 0x60, 0x20, 0x24, 0x7b, 0x74, 0x68, 0x69, 0x73, 0x2e, - 0x5f, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x4c, 0x69, 0x74, 0x65, 0x72, - 0x61, 0x6c, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x4e, 0x61, 0x6d, 0x65, 0x29, - 0x7d, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x20, 0x22, 0x3a, 0x22, 0x20, - 0x73, 0x70, 0x61, 0x63, 0x65, 0x20, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, - 0x52, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x60, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x2b, 0x3d, 0x20, - 0x27, 0x20, 0x22, 0x7d, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, - 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, - 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, - 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, - 0x6d, 0x65, 0x2c, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x72, 0x65, 0x71, 0x75, 0x69, 0x72, 0x65, 0x64, 0x20, + 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x53, 0x65, 0x74, 0x28, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x70, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, 0x20, + 0x3d, 0x20, 0x5b, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x61, 0x64, 0x64, 0x43, 0x6f, 0x6d, + 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x63, 0x6f, + 0x6d, 0x70, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2c, 0x20, 0x69, 0x73, + 0x52, 0x65, 0x71, 0x75, 0x69, 0x72, 0x65, 0x64, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x66, 0x20, 0x3d, 0x20, 0x63, + 0x6f, 0x6d, 0x70, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x24, 0x72, + 0x65, 0x66, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x66, 0x20, 0x21, 0x3d, 0x3d, 0x20, + 0x75, 0x6e, 0x64, 0x65, 0x66, 0x69, 0x6e, 0x65, 0x64, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6d, 0x70, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x20, 0x3d, 0x20, + 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x65, 0x66, 0x73, 0x5b, 0x72, + 0x65, 0x66, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x27, 0x70, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, + 0x69, 0x65, 0x73, 0x27, 0x20, 0x69, 0x6e, 0x20, 0x63, 0x6f, 0x6d, 0x70, + 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, + 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x5b, 0x70, 0x72, 0x6f, 0x70, + 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x53, 0x63, + 0x68, 0x65, 0x6d, 0x61, 0x5d, 0x20, 0x6f, 0x66, 0x20, 0x4f, 0x62, 0x6a, + 0x65, 0x63, 0x74, 0x2e, 0x65, 0x6e, 0x74, 0x72, 0x69, 0x65, 0x73, 0x28, + 0x63, 0x6f, 0x6d, 0x70, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x70, + 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, 0x29, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, + 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x5b, 0x70, 0x72, 0x6f, 0x70, 0x4e, + 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x53, 0x63, 0x68, + 0x65, 0x6d, 0x61, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x69, + 0x73, 0x52, 0x65, 0x71, 0x75, 0x69, 0x72, 0x65, 0x64, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x71, 0x75, 0x69, 0x72, 0x65, 0x64, 0x2e, + 0x61, 0x64, 0x64, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x4e, 0x61, 0x6d, 0x65, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x20, 0x6f, 0x66, 0x20, 0x73, + 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x61, 0x6c, 0x6c, 0x4f, 0x66, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, + 0x66, 0x20, 0x28, 0x27, 0x61, 0x6e, 0x79, 0x4f, 0x66, 0x27, 0x20, 0x69, + 0x6e, 0x20, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x74, 0x74, 0x20, 0x6f, 0x66, 0x20, 0x74, 0x2e, + 0x61, 0x6e, 0x79, 0x4f, 0x66, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x64, 0x64, + 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x28, 0x74, 0x74, + 0x2c, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x61, 0x64, 0x64, 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, + 0x74, 0x28, 0x74, 0x2c, 0x20, 0x74, 0x72, 0x75, 0x65, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, 0x75, + 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x5f, 0x62, 0x75, 0x69, 0x6c, 0x64, 0x4f, 0x62, 0x6a, 0x65, 0x63, + 0x74, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x65, 0x72, + 0x74, 0x69, 0x65, 0x73, 0x2c, 0x20, 0x72, 0x65, 0x71, 0x75, 0x69, 0x72, + 0x65, 0x64, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x2f, 0x2a, + 0x20, 0x61, 0x64, 0x64, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x50, + 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, 0x3d, 0x20, 0x2a, + 0x2f, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, - 0x20, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, - 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x61, 0x72, 0x72, 0x61, 0x79, 0x27, - 0x20, 0x26, 0x26, 0x20, 0x27, 0x69, 0x74, 0x65, 0x6d, 0x73, 0x27, 0x20, - 0x69, 0x6e, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x54, 0x4f, - 0x44, 0x4f, 0x20, 0x60, 0x70, 0x72, 0x65, 0x66, 0x69, 0x78, 0x49, 0x74, - 0x65, 0x6d, 0x73, 0x60, 0x20, 0x6b, 0x65, 0x79, 0x77, 0x6f, 0x72, 0x64, - 0x20, 0x28, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x70, 0x79, 0x74, 0x68, 0x6f, - 0x6e, 0x20, 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x61, - 0x74, 0x69, 0x6f, 0x6e, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x52, 0x75, - 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, - 0x73, 0x2e, 0x76, 0x69, 0x73, 0x69, 0x74, 0x28, 0x73, 0x63, 0x68, 0x65, - 0x6d, 0x61, 0x2e, 0x69, 0x74, 0x65, 0x6d, 0x73, 0x2c, 0x20, 0x60, 0x24, - 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, - 0x20, 0x3f, 0x20, 0x22, 0x2d, 0x22, 0x20, 0x3a, 0x20, 0x22, 0x22, 0x7d, - 0x69, 0x74, 0x65, 0x6d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x75, 0x6c, 0x65, - 0x20, 0x3d, 0x20, 0x60, 0x22, 0x5b, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, - 0x65, 0x20, 0x28, 0x24, 0x7b, 0x69, 0x74, 0x65, 0x6d, 0x52, 0x75, 0x6c, - 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x20, 0x28, 0x22, 0x2c, 0x22, 0x20, - 0x73, 0x70, 0x61, 0x63, 0x65, 0x20, 0x24, 0x7b, 0x69, 0x74, 0x65, 0x6d, - 0x52, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x29, 0x2a, 0x29, - 0x3f, 0x20, 0x22, 0x5d, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x60, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, - 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, - 0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, - 0x65, 0x2c, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x50, 0x52, - 0x49, 0x4d, 0x49, 0x54, 0x49, 0x56, 0x45, 0x5f, 0x52, 0x55, 0x4c, 0x45, - 0x53, 0x5b, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, - 0x5d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45, - 0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x55, 0x6e, 0x72, 0x65, 0x63, 0x6f, - 0x67, 0x6e, 0x69, 0x7a, 0x65, 0x64, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, - 0x61, 0x3a, 0x20, 0x24, 0x7b, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x73, 0x74, - 0x72, 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, 0x73, 0x63, 0x68, 0x65, - 0x6d, 0x61, 0x29, 0x7d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, + 0x20, 0x28, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, + 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x75, 0x6e, 0x64, 0x65, 0x66, 0x69, + 0x6e, 0x65, 0x64, 0x20, 0x7c, 0x7c, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, + 0x61, 0x54, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x61, + 0x72, 0x72, 0x61, 0x79, 0x27, 0x29, 0x20, 0x26, 0x26, 0x20, 0x28, 0x27, + 0x69, 0x74, 0x65, 0x6d, 0x73, 0x27, 0x20, 0x69, 0x6e, 0x20, 0x73, 0x63, + 0x68, 0x65, 0x6d, 0x61, 0x20, 0x7c, 0x7c, 0x20, 0x27, 0x70, 0x72, 0x65, + 0x66, 0x69, 0x78, 0x49, 0x74, 0x65, 0x6d, 0x73, 0x27, 0x20, 0x69, 0x6e, + 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x29, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x69, 0x74, 0x65, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x63, 0x68, 0x65, + 0x6d, 0x61, 0x2e, 0x69, 0x74, 0x65, 0x6d, 0x73, 0x20, 0x3f, 0x3f, 0x20, + 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x70, 0x72, 0x65, 0x66, 0x69, + 0x78, 0x49, 0x74, 0x65, 0x6d, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x41, 0x72, 0x72, 0x61, 0x79, 0x2e, + 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x28, 0x69, 0x74, 0x65, 0x6d, + 0x73, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, + 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x27, 0x22, 0x5b, 0x22, 0x20, 0x73, + 0x70, 0x61, 0x63, 0x65, 0x20, 0x27, 0x20, 0x2b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x74, 0x65, + 0x6d, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x28, 0x69, 0x74, 0x65, 0x6d, + 0x2c, 0x20, 0x69, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x76, 0x69, 0x73, 0x69, 0x74, 0x28, 0x69, 0x74, 0x65, 0x6d, 0x2c, + 0x20, 0x60, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x3f, 0x20, + 0x27, 0x27, 0x7d, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x20, + 0x27, 0x2d, 0x27, 0x20, 0x3a, 0x20, 0x27, 0x27, 0x7d, 0x74, 0x75, 0x70, + 0x6c, 0x65, 0x2d, 0x24, 0x7b, 0x69, 0x7d, 0x60, 0x29, 0x29, 0x2e, 0x6a, + 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x20, 0x22, 0x2c, 0x22, 0x20, 0x73, 0x70, + 0x61, 0x63, 0x65, 0x20, 0x27, 0x29, 0x20, 0x2b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x27, 0x20, 0x22, + 0x5d, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x52, 0x75, 0x6c, 0x65, 0x4e, 0x61, + 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x76, 0x69, + 0x73, 0x69, 0x74, 0x28, 0x69, 0x74, 0x65, 0x6d, 0x73, 0x2c, 0x20, 0x60, + 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x3f, 0x20, 0x27, 0x27, + 0x7d, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x20, 0x27, 0x2d, + 0x27, 0x20, 0x3a, 0x20, 0x27, 0x27, 0x7d, 0x69, 0x74, 0x65, 0x6d, 0x60, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x69, 0x73, 0x74, 0x49, 0x74, 0x65, + 0x6d, 0x4f, 0x70, 0x65, 0x72, 0x61, 0x74, 0x6f, 0x72, 0x20, 0x3d, 0x20, + 0x60, 0x28, 0x20, 0x22, 0x2c, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, + 0x20, 0x24, 0x7b, 0x69, 0x74, 0x65, 0x6d, 0x52, 0x75, 0x6c, 0x65, 0x4e, + 0x61, 0x6d, 0x65, 0x7d, 0x20, 0x29, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x73, 0x75, 0x63, + 0x63, 0x65, 0x73, 0x73, 0x69, 0x76, 0x65, 0x49, 0x74, 0x65, 0x6d, 0x73, + 0x20, 0x3d, 0x20, 0x27, 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6d, 0x69, 0x6e, 0x49, 0x74, + 0x65, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, + 0x2e, 0x6d, 0x69, 0x6e, 0x49, 0x74, 0x65, 0x6d, 0x73, 0x20, 0x7c, 0x7c, + 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x61, 0x78, 0x49, 0x74, 0x65, + 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, + 0x6d, 0x61, 0x78, 0x49, 0x74, 0x65, 0x6d, 0x73, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6d, 0x69, + 0x6e, 0x49, 0x74, 0x65, 0x6d, 0x73, 0x20, 0x3e, 0x20, 0x30, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x73, 0x75, 0x63, 0x63, 0x65, 0x73, 0x73, 0x69, 0x76, 0x65, 0x49, 0x74, + 0x65, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x6c, 0x69, 0x73, 0x74, 0x49, 0x74, + 0x65, 0x6d, 0x4f, 0x70, 0x65, 0x72, 0x61, 0x74, 0x6f, 0x72, 0x2e, 0x72, + 0x65, 0x70, 0x65, 0x61, 0x74, 0x28, 0x6d, 0x69, 0x6e, 0x49, 0x74, 0x65, + 0x6d, 0x73, 0x20, 0x2d, 0x20, 0x31, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x6e, 0x49, 0x74, + 0x65, 0x6d, 0x73, 0x2d, 0x2d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x66, 0x20, 0x28, 0x6d, 0x61, 0x78, 0x49, 0x74, 0x65, 0x6d, + 0x73, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x75, 0x6e, 0x64, 0x65, 0x66, 0x69, + 0x6e, 0x65, 0x64, 0x20, 0x26, 0x26, 0x20, 0x6d, 0x61, 0x78, 0x49, 0x74, + 0x65, 0x6d, 0x73, 0x20, 0x3e, 0x20, 0x6d, 0x69, 0x6e, 0x49, 0x74, 0x65, + 0x6d, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x73, 0x75, 0x63, 0x63, 0x65, 0x73, 0x73, 0x69, + 0x76, 0x65, 0x49, 0x74, 0x65, 0x6d, 0x73, 0x20, 0x2b, 0x3d, 0x20, 0x60, + 0x24, 0x7b, 0x6c, 0x69, 0x73, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x4f, 0x70, + 0x65, 0x72, 0x61, 0x74, 0x6f, 0x72, 0x7d, 0x3f, 0x60, 0x2e, 0x72, 0x65, + 0x70, 0x65, 0x61, 0x74, 0x28, 0x6d, 0x61, 0x78, 0x49, 0x74, 0x65, 0x6d, + 0x73, 0x20, 0x2d, 0x20, 0x6d, 0x69, 0x6e, 0x49, 0x74, 0x65, 0x6d, 0x73, + 0x20, 0x2d, 0x20, 0x31, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x75, + 0x63, 0x63, 0x65, 0x73, 0x73, 0x69, 0x76, 0x65, 0x49, 0x74, 0x65, 0x6d, + 0x73, 0x20, 0x2b, 0x3d, 0x20, 0x60, 0x24, 0x7b, 0x6c, 0x69, 0x73, 0x74, + 0x49, 0x74, 0x65, 0x6d, 0x4f, 0x70, 0x65, 0x72, 0x61, 0x74, 0x6f, 0x72, + 0x7d, 0x2a, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x3d, 0x20, + 0x6d, 0x69, 0x6e, 0x49, 0x74, 0x65, 0x6d, 0x73, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x30, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3f, 0x20, 0x60, 0x22, 0x5b, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, + 0x65, 0x20, 0x28, 0x20, 0x24, 0x7b, 0x69, 0x74, 0x65, 0x6d, 0x52, 0x75, + 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x20, 0x24, 0x7b, 0x73, 0x75, + 0x63, 0x63, 0x65, 0x73, 0x73, 0x69, 0x76, 0x65, 0x49, 0x74, 0x65, 0x6d, + 0x73, 0x7d, 0x20, 0x29, 0x3f, 0x20, 0x22, 0x5d, 0x22, 0x20, 0x73, 0x70, + 0x61, 0x63, 0x65, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3a, 0x20, 0x60, 0x22, 0x5b, 0x22, 0x20, 0x73, 0x70, + 0x61, 0x63, 0x65, 0x20, 0x24, 0x7b, 0x69, 0x74, 0x65, 0x6d, 0x52, 0x75, + 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x20, 0x24, 0x7b, 0x73, 0x75, + 0x63, 0x63, 0x65, 0x73, 0x73, 0x69, 0x76, 0x65, 0x49, 0x74, 0x65, 0x6d, + 0x73, 0x7d, 0x20, 0x22, 0x5d, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, + 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, + 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, 0x75, 0x6c, 0x65, + 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, + 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x20, + 0x3d, 0x3d, 0x3d, 0x20, 0x75, 0x6e, 0x64, 0x65, 0x66, 0x69, 0x6e, 0x65, + 0x64, 0x20, 0x7c, 0x7c, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, + 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x73, 0x74, 0x72, + 0x69, 0x6e, 0x67, 0x27, 0x29, 0x20, 0x26, 0x26, 0x20, 0x27, 0x70, 0x61, + 0x74, 0x74, 0x65, 0x72, 0x6e, 0x27, 0x20, 0x69, 0x6e, 0x20, 0x73, 0x63, + 0x68, 0x65, 0x6d, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x5f, 0x76, 0x69, 0x73, 0x69, 0x74, 0x50, 0x61, 0x74, 0x74, + 0x65, 0x72, 0x6e, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x70, + 0x61, 0x74, 0x74, 0x65, 0x72, 0x6e, 0x2c, 0x20, 0x72, 0x75, 0x6c, 0x65, + 0x4e, 0x61, 0x6d, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, 0x28, 0x73, + 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, + 0x3d, 0x20, 0x75, 0x6e, 0x64, 0x65, 0x66, 0x69, 0x6e, 0x65, 0x64, 0x20, + 0x7c, 0x7c, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, + 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x73, 0x74, 0x72, 0x69, 0x6e, + 0x67, 0x27, 0x29, 0x20, 0x26, 0x26, 0x20, 0x2f, 0x5e, 0x75, 0x75, 0x69, + 0x64, 0x5b, 0x31, 0x2d, 0x35, 0x5d, 0x3f, 0x24, 0x2f, 0x2e, 0x74, 0x65, + 0x73, 0x74, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x66, 0x6f, + 0x72, 0x6d, 0x61, 0x74, 0x20, 0x7c, 0x7c, 0x20, 0x27, 0x27, 0x29, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, + 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, + 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x72, 0x6f, 0x6f, 0x74, 0x27, + 0x20, 0x3f, 0x20, 0x27, 0x72, 0x6f, 0x6f, 0x74, 0x27, 0x20, 0x3a, 0x20, + 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x46, 0x6f, 0x72, 0x6d, 0x61, 0x74, + 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x50, 0x52, 0x49, 0x4d, 0x49, 0x54, 0x49, 0x56, 0x45, 0x5f, 0x52, 0x55, + 0x4c, 0x45, 0x53, 0x5b, 0x27, 0x75, 0x75, 0x69, 0x64, 0x27, 0x5d, 0x29, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, + 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x75, 0x6e, 0x64, 0x65, + 0x66, 0x69, 0x6e, 0x65, 0x64, 0x20, 0x7c, 0x7c, 0x20, 0x73, 0x63, 0x68, + 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, + 0x27, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x29, 0x20, 0x26, 0x26, + 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2e, 0x66, 0x6f, 0x72, 0x6d, + 0x61, 0x74, 0x20, 0x69, 0x6e, 0x20, 0x44, 0x41, 0x54, 0x45, 0x5f, 0x52, + 0x55, 0x4c, 0x45, 0x53, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x5b, 0x74, 0x2c, 0x20, 0x72, 0x5d, 0x20, 0x6f, 0x66, 0x20, 0x4f, + 0x62, 0x6a, 0x65, 0x63, 0x74, 0x2e, 0x65, 0x6e, 0x74, 0x72, 0x69, 0x65, + 0x73, 0x28, 0x44, 0x41, 0x54, 0x45, 0x5f, 0x52, 0x55, 0x4c, 0x45, 0x53, + 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, + 0x6c, 0x65, 0x28, 0x74, 0x2c, 0x20, 0x72, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, + 0x61, 0x46, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x20, 0x2b, 0x20, 0x27, 0x2d, + 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x69, 0x66, 0x20, 0x28, + 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x20, + 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x27, + 0x29, 0x20, 0x7c, 0x7c, 0x20, 0x28, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, + 0x2e, 0x6b, 0x65, 0x79, 0x73, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, + 0x29, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x30, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x6e, 0x20, 0x6f, 0x66, 0x20, 0x4f, 0x42, 0x4a, 0x45, 0x43, 0x54, 0x5f, + 0x52, 0x55, 0x4c, 0x45, 0x5f, 0x4e, 0x41, 0x4d, 0x45, 0x53, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, + 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, + 0x6e, 0x2c, 0x20, 0x50, 0x52, 0x49, 0x4d, 0x49, 0x54, 0x49, 0x56, 0x45, + 0x5f, 0x52, 0x55, 0x4c, 0x45, 0x53, 0x5b, 0x6e, 0x5d, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, + 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x27, 0x6f, 0x62, + 0x6a, 0x65, 0x63, 0x74, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x28, 0x73, 0x63, 0x68, + 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x20, 0x69, 0x6e, 0x20, 0x50, + 0x52, 0x49, 0x4d, 0x49, 0x54, 0x49, 0x56, 0x45, 0x5f, 0x52, 0x55, 0x4c, + 0x45, 0x53, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, + 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x55, 0x6e, 0x72, 0x65, + 0x63, 0x6f, 0x67, 0x6e, 0x69, 0x7a, 0x65, 0x64, 0x20, 0x73, 0x63, 0x68, + 0x65, 0x6d, 0x61, 0x3a, 0x20, 0x24, 0x7b, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, + 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, 0x73, 0x63, + 0x68, 0x65, 0x6d, 0x61, 0x29, 0x7d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x2f, 0x2f, 0x20, 0x54, 0x4f, 0x44, 0x4f, 0x3a, 0x20, 0x73, 0x75, 0x70, + 0x70, 0x6f, 0x72, 0x74, 0x20, 0x6d, 0x69, 0x6e, 0x69, 0x6d, 0x75, 0x6d, + 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x69, 0x6d, 0x75, 0x6d, 0x2c, 0x20, 0x65, + 0x78, 0x63, 0x6c, 0x75, 0x73, 0x69, 0x76, 0x65, 0x4d, 0x69, 0x6e, 0x69, + 0x6d, 0x75, 0x6d, 0x2c, 0x20, 0x65, 0x78, 0x63, 0x6c, 0x75, 0x73, 0x69, + 0x76, 0x65, 0x4d, 0x61, 0x78, 0x69, 0x6d, 0x75, 0x6d, 0x20, 0x61, 0x74, + 0x20, 0x6c, 0x65, 0x61, 0x73, 0x74, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x7a, + 0x65, 0x72, 0x6f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, - 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, - 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x72, 0x6f, 0x6f, 0x74, 0x27, 0x20, - 0x3f, 0x20, 0x27, 0x72, 0x6f, 0x6f, 0x74, 0x27, 0x20, 0x3a, 0x20, 0x73, - 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x2c, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x50, 0x52, 0x49, 0x4d, 0x49, - 0x54, 0x49, 0x56, 0x45, 0x5f, 0x52, 0x55, 0x4c, 0x45, 0x53, 0x5b, 0x73, - 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, 0x5d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x66, 0x6f, 0x72, + 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x72, 0x75, 0x6c, 0x65, 0x4e, + 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x72, 0x6f, 0x6f, + 0x74, 0x27, 0x20, 0x3f, 0x20, 0x27, 0x72, 0x6f, 0x6f, 0x74, 0x27, 0x20, + 0x3a, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x54, 0x79, 0x70, 0x65, + 0x2c, 0x20, 0x50, 0x52, 0x49, 0x4d, 0x49, 0x54, 0x49, 0x56, 0x45, 0x5f, + 0x52, 0x55, 0x4c, 0x45, 0x53, 0x5b, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, + 0x54, 0x79, 0x70, 0x65, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x5f, 0x62, 0x75, + 0x69, 0x6c, 0x64, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x52, 0x75, 0x6c, + 0x65, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, + 0x2c, 0x20, 0x72, 0x65, 0x71, 0x75, 0x69, 0x72, 0x65, 0x64, 0x2c, 0x20, + 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x61, 0x64, 0x64, 0x69, 0x74, 0x69, + 0x6f, 0x6e, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, + 0x65, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, + 0x72, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x70, 0x72, + 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x6f, 0x72, 0x74, 0x20, 0x62, 0x79, 0x20, + 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x6e, 0x20, + 0x70, 0x72, 0x6f, 0x70, 0x5f, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x28, + 0x69, 0x66, 0x20, 0x73, 0x70, 0x65, 0x63, 0x69, 0x66, 0x69, 0x65, 0x64, + 0x29, 0x20, 0x74, 0x68, 0x65, 0x6e, 0x20, 0x62, 0x79, 0x20, 0x6f, 0x72, + 0x69, 0x67, 0x69, 0x6e, 0x61, 0x6c, 0x20, 0x6f, 0x72, 0x64, 0x65, 0x72, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, + 0x6f, 0x72, 0x74, 0x65, 0x64, 0x50, 0x72, 0x6f, 0x70, 0x73, 0x20, 0x3d, + 0x20, 0x70, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, 0x2e, + 0x6d, 0x61, 0x70, 0x28, 0x28, 0x5b, 0x6b, 0x5d, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x6b, 0x29, 0x2e, 0x73, 0x6f, 0x72, 0x74, 0x28, 0x28, 0x61, 0x2c, + 0x20, 0x62, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6f, 0x72, 0x64, + 0x65, 0x72, 0x41, 0x20, 0x3d, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, + 0x64, 0x65, 0x72, 0x5b, 0x61, 0x5d, 0x20, 0x7c, 0x7c, 0x20, 0x49, 0x6e, + 0x66, 0x69, 0x6e, 0x69, 0x74, 0x79, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6f, 0x72, 0x64, 0x65, + 0x72, 0x42, 0x20, 0x3d, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, + 0x65, 0x72, 0x5b, 0x62, 0x5d, 0x20, 0x7c, 0x7c, 0x20, 0x49, 0x6e, 0x66, + 0x69, 0x6e, 0x69, 0x74, 0x79, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6f, 0x72, 0x64, 0x65, + 0x72, 0x41, 0x20, 0x2d, 0x20, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x42, 0x20, + 0x7c, 0x7c, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, + 0x73, 0x2e, 0x66, 0x69, 0x6e, 0x64, 0x49, 0x6e, 0x64, 0x65, 0x78, 0x28, + 0x28, 0x5b, 0x6b, 0x5d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x6b, 0x20, 0x3d, + 0x3d, 0x3d, 0x20, 0x61, 0x29, 0x20, 0x2d, 0x20, 0x70, 0x72, 0x6f, 0x70, + 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, 0x2e, 0x66, 0x69, 0x6e, 0x64, 0x49, + 0x6e, 0x64, 0x65, 0x78, 0x28, 0x28, 0x5b, 0x6b, 0x5d, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x6b, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x62, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x4b, + 0x76, 0x52, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x73, 0x20, 0x3d, + 0x20, 0x7b, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, + 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x5b, 0x70, 0x72, 0x6f, + 0x70, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x53, + 0x63, 0x68, 0x65, 0x6d, 0x61, 0x5d, 0x20, 0x6f, 0x66, 0x20, 0x70, 0x72, + 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x70, 0x72, 0x6f, 0x70, 0x52, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, + 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x76, 0x69, 0x73, 0x69, + 0x74, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, + 0x2c, 0x20, 0x60, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x3f, + 0x20, 0x27, 0x27, 0x7d, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, + 0x20, 0x27, 0x2d, 0x27, 0x20, 0x3a, 0x20, 0x27, 0x27, 0x7d, 0x24, 0x7b, + 0x70, 0x72, 0x6f, 0x70, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x60, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x4b, + 0x76, 0x52, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x73, 0x5b, 0x70, + 0x72, 0x6f, 0x70, 0x4e, 0x61, 0x6d, 0x65, 0x5d, 0x20, 0x3d, 0x20, 0x74, + 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, + 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x24, + 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x3f, 0x20, 0x27, 0x27, 0x7d, + 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x20, 0x27, 0x2d, 0x27, + 0x20, 0x3a, 0x20, 0x27, 0x27, 0x7d, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, + 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x2d, 0x6b, 0x76, 0x60, 0x2c, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x24, 0x7b, 0x74, 0x68, + 0x69, 0x73, 0x2e, 0x5f, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x4c, 0x69, + 0x74, 0x65, 0x72, 0x61, 0x6c, 0x28, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x73, + 0x74, 0x72, 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, 0x70, 0x72, 0x6f, + 0x70, 0x4e, 0x61, 0x6d, 0x65, 0x29, 0x29, 0x7d, 0x20, 0x73, 0x70, 0x61, + 0x63, 0x65, 0x20, 0x22, 0x3a, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, + 0x20, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x52, 0x75, 0x6c, 0x65, 0x4e, + 0x61, 0x6d, 0x65, 0x7d, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x71, 0x75, 0x69, + 0x72, 0x65, 0x64, 0x50, 0x72, 0x6f, 0x70, 0x73, 0x20, 0x3d, 0x20, 0x73, + 0x6f, 0x72, 0x74, 0x65, 0x64, 0x50, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x66, + 0x69, 0x6c, 0x74, 0x65, 0x72, 0x28, 0x6b, 0x20, 0x3d, 0x3e, 0x20, 0x72, + 0x65, 0x71, 0x75, 0x69, 0x72, 0x65, 0x64, 0x2e, 0x68, 0x61, 0x73, 0x28, + 0x6b, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x50, + 0x72, 0x6f, 0x70, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x6f, 0x72, 0x74, 0x65, + 0x64, 0x50, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x66, 0x69, 0x6c, 0x74, 0x65, + 0x72, 0x28, 0x6b, 0x20, 0x3d, 0x3e, 0x20, 0x21, 0x72, 0x65, 0x71, 0x75, + 0x69, 0x72, 0x65, 0x64, 0x2e, 0x68, 0x61, 0x73, 0x28, 0x6b, 0x29, 0x29, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, + 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x61, 0x64, 0x64, 0x69, 0x74, 0x69, + 0x6f, 0x6e, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, 0x69, + 0x65, 0x73, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x6f, 0x62, 0x6a, 0x65, + 0x63, 0x74, 0x27, 0x20, 0x7c, 0x7c, 0x20, 0x61, 0x64, 0x64, 0x69, 0x74, + 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x70, 0x65, 0x72, 0x74, + 0x69, 0x65, 0x73, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, + 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x73, 0x75, 0x62, 0x4e, 0x61, 0x6d, 0x65, 0x20, + 0x3d, 0x20, 0x60, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x3f, + 0x20, 0x27, 0x27, 0x7d, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, + 0x20, 0x27, 0x2d, 0x27, 0x20, 0x3a, 0x20, 0x27, 0x27, 0x7d, 0x61, 0x64, + 0x64, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x60, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x52, 0x75, 0x6c, 0x65, 0x20, 0x3d, 0x20, 0x74, + 0x68, 0x69, 0x73, 0x2e, 0x76, 0x69, 0x73, 0x69, 0x74, 0x28, 0x61, 0x64, + 0x64, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x70, + 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x74, + 0x72, 0x75, 0x65, 0x20, 0x3f, 0x20, 0x7b, 0x7d, 0x20, 0x3a, 0x20, 0x61, + 0x64, 0x64, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x50, 0x72, 0x6f, + 0x70, 0x65, 0x72, 0x74, 0x69, 0x65, 0x73, 0x2c, 0x20, 0x60, 0x24, 0x7b, + 0x73, 0x75, 0x62, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x2d, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x70, 0x72, 0x6f, 0x70, 0x4b, 0x76, 0x52, 0x75, 0x6c, 0x65, 0x4e, 0x61, + 0x6d, 0x65, 0x73, 0x5b, 0x27, 0x2a, 0x27, 0x5d, 0x20, 0x3d, 0x20, 0x74, + 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, + 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x24, + 0x7b, 0x73, 0x75, 0x62, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x2d, 0x6b, 0x76, + 0x60, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, + 0x24, 0x7b, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, + 0x75, 0x6c, 0x65, 0x28, 0x27, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x27, + 0x2c, 0x20, 0x50, 0x52, 0x49, 0x4d, 0x49, 0x54, 0x49, 0x56, 0x45, 0x5f, + 0x52, 0x55, 0x4c, 0x45, 0x53, 0x5b, 0x27, 0x73, 0x74, 0x72, 0x69, 0x6e, + 0x67, 0x27, 0x5d, 0x29, 0x7d, 0x20, 0x22, 0x3a, 0x22, 0x20, 0x73, 0x70, + 0x61, 0x63, 0x65, 0x20, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x52, + 0x75, 0x6c, 0x65, 0x7d, 0x60, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x50, 0x72, + 0x6f, 0x70, 0x73, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x27, 0x2a, 0x27, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x3d, + 0x20, 0x27, 0x22, 0x7b, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x20, + 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, + 0x2b, 0x3d, 0x20, 0x72, 0x65, 0x71, 0x75, 0x69, 0x72, 0x65, 0x64, 0x50, + 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6b, 0x20, 0x3d, + 0x3e, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x4b, 0x76, 0x52, 0x75, 0x6c, 0x65, + 0x4e, 0x61, 0x6d, 0x65, 0x73, 0x5b, 0x6b, 0x5d, 0x29, 0x2e, 0x6a, 0x6f, + 0x69, 0x6e, 0x28, 0x27, 0x20, 0x22, 0x2c, 0x22, 0x20, 0x73, 0x70, 0x61, + 0x63, 0x65, 0x20, 0x27, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, + 0x50, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, + 0x20, 0x3e, 0x20, 0x30, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x2b, 0x3d, 0x20, 0x27, 0x20, + 0x28, 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, + 0x20, 0x28, 0x72, 0x65, 0x71, 0x75, 0x69, 0x72, 0x65, 0x64, 0x50, 0x72, + 0x6f, 0x70, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, + 0x20, 0x30, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x2b, 0x3d, 0x20, 0x27, 0x20, + 0x22, 0x2c, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x20, 0x28, 0x20, + 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x67, 0x65, 0x74, 0x52, 0x65, 0x63, 0x75, 0x72, 0x73, 0x69, 0x76, 0x65, + 0x52, 0x65, 0x66, 0x73, 0x20, 0x3d, 0x20, 0x28, 0x6b, 0x73, 0x2c, 0x20, + 0x66, 0x69, 0x72, 0x73, 0x74, 0x49, 0x73, 0x4f, 0x70, 0x74, 0x69, 0x6f, + 0x6e, 0x61, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x5b, 0x6b, 0x2c, 0x20, 0x2e, 0x2e, 0x2e, 0x72, 0x65, 0x73, 0x74, 0x5d, + 0x20, 0x3d, 0x20, 0x6b, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6b, 0x76, 0x52, + 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x70, 0x72, + 0x6f, 0x70, 0x4b, 0x76, 0x52, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, + 0x73, 0x5b, 0x6b, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x72, 0x65, 0x73, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6b, + 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x2a, 0x27, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, + 0x61, 0x64, 0x64, 0x52, 0x75, 0x6c, 0x65, 0x28, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x60, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x3f, 0x20, + 0x27, 0x27, 0x7d, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3f, 0x20, + 0x27, 0x2d, 0x27, 0x20, 0x3a, 0x20, 0x27, 0x27, 0x7d, 0x61, 0x64, 0x64, + 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x2d, 0x6b, 0x76, 0x73, 0x60, + 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x24, 0x7b, 0x6b, 0x76, 0x52, + 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x20, 0x28, 0x20, 0x22, + 0x2c, 0x22, 0x20, 0x73, 0x70, 0x61, 0x63, 0x65, 0x20, 0x60, 0x20, 0x2b, + 0x20, 0x6b, 0x76, 0x52, 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x20, + 0x2b, 0x20, 0x60, 0x20, 0x29, 0x2a, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, + 0x20, 0x69, 0x66, 0x20, 0x28, 0x66, 0x69, 0x72, 0x73, 0x74, 0x49, 0x73, + 0x4f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, + 0x73, 0x20, 0x3d, 0x20, 0x60, 0x28, 0x20, 0x22, 0x2c, 0x22, 0x20, 0x73, + 0x70, 0x61, 0x63, 0x65, 0x20, 0x24, 0x7b, 0x6b, 0x76, 0x52, 0x75, 0x6c, + 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x7d, 0x20, 0x29, 0x3f, 0x60, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, + 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x6b, 0x76, 0x52, + 0x75, 0x6c, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x74, 0x2e, + 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, 0x20, 0x30, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x73, 0x20, 0x2b, 0x3d, 0x20, 0x27, 0x20, 0x27, 0x20, 0x2b, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x61, 0x64, 0x64, 0x52, 0x75, + 0x6c, 0x65, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x60, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x20, + 0x3f, 0x3f, 0x20, 0x27, 0x27, 0x7d, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, + 0x20, 0x3f, 0x20, 0x27, 0x2d, 0x27, 0x20, 0x3a, 0x20, 0x27, 0x27, 0x7d, + 0x24, 0x7b, 0x6b, 0x7d, 0x2d, 0x72, 0x65, 0x73, 0x74, 0x60, 0x2c, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x67, 0x65, 0x74, 0x52, 0x65, 0x63, 0x75, 0x72, 0x73, 0x69, 0x76, 0x65, + 0x52, 0x65, 0x66, 0x73, 0x28, 0x72, 0x65, 0x73, 0x74, 0x2c, 0x20, 0x74, + 0x72, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x72, 0x65, 0x73, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x2b, 0x3d, 0x20, + 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x70, + 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x28, 0x5f, 0x2c, 0x20, 0x69, 0x29, + 0x20, 0x3d, 0x3e, 0x20, 0x67, 0x65, 0x74, 0x52, 0x65, 0x63, 0x75, 0x72, + 0x73, 0x69, 0x76, 0x65, 0x52, 0x65, 0x66, 0x73, 0x28, 0x6f, 0x70, 0x74, + 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x73, + 0x6c, 0x69, 0x63, 0x65, 0x28, 0x69, 0x29, 0x2c, 0x20, 0x66, 0x61, 0x6c, + 0x73, 0x65, 0x29, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x20, + 0x7c, 0x20, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x71, 0x75, 0x69, 0x72, 0x65, 0x64, + 0x50, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, + 0x20, 0x3e, 0x20, 0x30, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, 0x20, 0x2b, 0x3d, 0x20, + 0x27, 0x20, 0x29, 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, 0x6c, 0x65, + 0x20, 0x2b, 0x3d, 0x20, 0x27, 0x20, 0x29, 0x3f, 0x27, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x72, 0x75, + 0x6c, 0x65, 0x20, 0x2b, 0x3d, 0x20, 0x27, 0x20, 0x22, 0x7d, 0x22, 0x20, + 0x73, 0x70, 0x61, 0x63, 0x65, 0x27, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x72, 0x75, 0x6c, 0x65, + 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x20, 0x3d, 0x20, 0x27, 0x27, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, - 0x75, 0x6c, 0x65, 0x73, 0x2e, 0x66, 0x6f, 0x72, 0x45, 0x61, 0x63, 0x68, - 0x28, 0x28, 0x72, 0x75, 0x6c, 0x65, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, - 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x20, 0x2b, 0x3d, 0x20, - 0x60, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x20, 0x3a, 0x3a, 0x3d, - 0x20, 0x24, 0x7b, 0x72, 0x75, 0x6c, 0x65, 0x7d, 0x5c, 0x6e, 0x60, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x67, 0x72, 0x61, 0x6d, - 0x6d, 0x61, 0x72, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x7d, 0x0a + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x5b, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x72, + 0x75, 0x6c, 0x65, 0x5d, 0x20, 0x6f, 0x66, 0x20, 0x4f, 0x62, 0x6a, 0x65, + 0x63, 0x74, 0x2e, 0x65, 0x6e, 0x74, 0x72, 0x69, 0x65, 0x73, 0x28, 0x74, + 0x68, 0x69, 0x73, 0x2e, 0x5f, 0x72, 0x75, 0x6c, 0x65, 0x73, 0x29, 0x2e, + 0x73, 0x6f, 0x72, 0x74, 0x28, 0x28, 0x5b, 0x61, 0x5d, 0x2c, 0x20, 0x5b, + 0x62, 0x5d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x61, 0x2e, 0x6c, 0x6f, 0x63, + 0x61, 0x6c, 0x65, 0x43, 0x6f, 0x6d, 0x70, 0x61, 0x72, 0x65, 0x28, 0x62, + 0x29, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x20, 0x2b, 0x3d, 0x20, 0x60, + 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x20, 0x3a, 0x3a, 0x3d, 0x20, + 0x24, 0x7b, 0x72, 0x75, 0x6c, 0x65, 0x7d, 0x5c, 0x6e, 0x60, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, + 0x74, 0x75, 0x72, 0x6e, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, + 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, + 0x48, 0x65, 0x6c, 0x70, 0x65, 0x72, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, + 0x69, 0x6f, 0x6e, 0x20, 0x74, 0x6f, 0x20, 0x67, 0x72, 0x6f, 0x75, 0x70, + 0x20, 0x65, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x73, 0x20, 0x62, 0x79, + 0x20, 0x61, 0x20, 0x6b, 0x65, 0x79, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, + 0x69, 0x6f, 0x6e, 0x0a, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, + 0x2a, 0x20, 0x67, 0x72, 0x6f, 0x75, 0x70, 0x42, 0x79, 0x28, 0x69, 0x74, + 0x65, 0x72, 0x61, 0x62, 0x6c, 0x65, 0x2c, 0x20, 0x6b, 0x65, 0x79, 0x46, + 0x6e, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6c, + 0x61, 0x73, 0x74, 0x4b, 0x65, 0x79, 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6c, + 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x67, 0x72, 0x6f, + 0x75, 0x70, 0x20, 0x3d, 0x20, 0x5b, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x66, + 0x6f, 0x72, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x6c, + 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x20, 0x6f, 0x66, 0x20, 0x69, 0x74, 0x65, + 0x72, 0x61, 0x62, 0x6c, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6b, 0x65, 0x79, 0x20, 0x3d, + 0x20, 0x6b, 0x65, 0x79, 0x46, 0x6e, 0x28, 0x65, 0x6c, 0x65, 0x6d, 0x65, + 0x6e, 0x74, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, + 0x28, 0x6c, 0x61, 0x73, 0x74, 0x4b, 0x65, 0x79, 0x20, 0x21, 0x3d, 0x3d, + 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x20, 0x26, 0x26, 0x20, 0x6b, 0x65, 0x79, + 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x4b, 0x65, 0x79, + 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x79, 0x69, + 0x65, 0x6c, 0x64, 0x20, 0x5b, 0x6c, 0x61, 0x73, 0x74, 0x4b, 0x65, 0x79, + 0x2c, 0x20, 0x67, 0x72, 0x6f, 0x75, 0x70, 0x5d, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x67, 0x72, 0x6f, 0x75, 0x70, 0x20, 0x3d, 0x20, + 0x5b, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x67, 0x72, 0x6f, 0x75, 0x70, 0x2e, 0x70, 0x75, 0x73, 0x68, + 0x28, 0x65, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x4b, 0x65, 0x79, 0x20, 0x3d, + 0x20, 0x6b, 0x65, 0x79, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x67, 0x72, 0x6f, 0x75, 0x70, 0x2e, 0x6c, 0x65, + 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, 0x20, 0x30, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x79, 0x69, 0x65, 0x6c, 0x64, 0x20, 0x5b, 0x6c, + 0x61, 0x73, 0x74, 0x4b, 0x65, 0x79, 0x2c, 0x20, 0x67, 0x72, 0x6f, 0x75, + 0x70, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x7d, 0x0a }; -unsigned int json_schema_to_grammar_mjs_len = 3695; +unsigned int json_schema_to_grammar_mjs_len = 19690; diff --git a/examples/server/public/index.html b/examples/server/public/index.html index 84038ddce..bbc5c2f68 100644 --- a/examples/server/public/index.html +++ b/examples/server/public/index.html @@ -630,14 +630,16 @@ const grammarJsonSchemaPropOrder = signal('') const updateGrammarJsonSchemaPropOrder = (el) => grammarJsonSchemaPropOrder.value = el.target.value - const convertJSONSchemaGrammar = () => { + const convertJSONSchemaGrammar = async () => { try { - const schema = JSON.parse(params.value.grammar) - const converter = new SchemaConverter( - grammarJsonSchemaPropOrder.value + let schema = JSON.parse(params.value.grammar) + const converter = new SchemaConverter({ + prop_order: grammarJsonSchemaPropOrder.value .split(',') - .reduce((acc, cur, i) => ({ ...acc, [cur.trim()]: i }), {}) - ) + .reduce((acc, cur, i) => ({ ...acc, [cur.trim()]: i }), {}), + allow_fetch: true, + }) + schema = await converter.resolveRefs(schema, 'input') converter.visit(schema, '') params.value = { ...params.value, diff --git a/examples/server/public/index.js b/examples/server/public/index.js index 9db5a9d9f..aa3ffb5e0 100644 --- a/examples/server/public/index.js +++ b/examples/server/public/index.js @@ -1 +1 @@ -function t(){throw new Error("Cycle detected")}function n(){if(u>1){u--;return}let t,n=!1;while(void 0!==_){let i=_;_=void 0;f++;while(void 0!==i){const _=i.o;i.o=void 0;i.f&=-3;if(!(8&i.f)&&a(i))try{i.c()}catch(e){if(!n){t=e;n=!0}}i=_}}f=0;u--;if(n)throw t}function e(t){if(u>0)return t();u++;try{return t()}finally{n()}}let i,_,o=0;function r(t){if(o>0)return t();const n=i;i=void 0;o++;try{return t()}finally{o--;i=n}}let u=0,f=0,l=0;function s(t){if(void 0===i)return;let n=t.n;if(void 0===n||n.t!==i){n={i:0,S:t,p:i.s,n:void 0,t:i,e:void 0,x:void 0,r:n};if(void 0!==i.s)i.s.n=n;i.s=n;t.n=n;if(32&i.f)t.S(n);return n}else if(-1===n.i){n.i=0;if(void 0!==n.n){n.n.p=n.p;if(void 0!==n.p)n.p.n=n.n;n.p=i.s;n.n=void 0;i.s.n=n;i.s=n}return n}}function c(t){this.v=t;this.i=0;this.n=void 0;this.t=void 0}c.prototype.h=function(){return!0};c.prototype.S=function(t){if(this.t!==t&&void 0===t.e){t.x=this.t;if(void 0!==this.t)this.t.e=t;this.t=t}};c.prototype.U=function(t){if(void 0!==this.t){const n=t.e,e=t.x;if(void 0!==n){n.x=e;t.e=void 0}if(void 0!==e){e.e=n;t.x=void 0}if(t===this.t)this.t=e}};c.prototype.subscribe=function(t){const n=this;return S((function(){const e=n.value,i=32&this.f;this.f&=-33;try{t(e)}finally{this.f|=i}}))};c.prototype.valueOf=function(){return this.value};c.prototype.toString=function(){return this.value+""};c.prototype.toJSON=function(){return this.value};c.prototype.peek=function(){return this.v};Object.defineProperty(c.prototype,"value",{get(){const t=s(this);if(void 0!==t)t.i=this.i;return this.v},set(e){if(i instanceof v)!function(){throw new Error("Computed cannot have side-effects")}();if(e!==this.v){if(f>100)t();this.v=e;this.i++;l++;u++;try{for(let t=this.t;void 0!==t;t=t.x)t.t.N()}finally{n()}}}});function h(t){return new c(t)}function a(t){for(let n=t.s;void 0!==n;n=n.n)if(n.S.i!==n.i||!n.S.h()||n.S.i!==n.i)return!0;return!1}function p(t){for(let n=t.s;void 0!==n;n=n.n){const e=n.S.n;if(void 0!==e)n.r=e;n.S.n=n;n.i=-1;if(void 0===n.n){t.s=n;break}}}function d(t){let n,e=t.s;while(void 0!==e){const t=e.p;if(-1===e.i){e.S.U(e);if(void 0!==t)t.n=e.n;if(void 0!==e.n)e.n.p=t}else n=e;e.S.n=e.r;if(void 0!==e.r)e.r=void 0;e=t}t.s=n}function v(t){c.call(this,void 0);this.x=t;this.s=void 0;this.g=l-1;this.f=4}(v.prototype=new c).h=function(){this.f&=-3;if(1&this.f)return!1;if(32==(36&this.f))return!0;this.f&=-5;if(this.g===l)return!0;this.g=l;this.f|=1;if(this.i>0&&!a(this)){this.f&=-2;return!0}const t=i;try{p(this);i=this;const t=this.x();if(16&this.f||this.v!==t||0===this.i){this.v=t;this.f&=-17;this.i++}}catch(t){this.v=t;this.f|=16;this.i++}i=t;d(this);this.f&=-2;return!0};v.prototype.S=function(t){if(void 0===this.t){this.f|=36;for(let t=this.s;void 0!==t;t=t.n)t.S.S(t)}c.prototype.S.call(this,t)};v.prototype.U=function(t){if(void 0!==this.t){c.prototype.U.call(this,t);if(void 0===this.t){this.f&=-33;for(let t=this.s;void 0!==t;t=t.n)t.S.U(t)}}};v.prototype.N=function(){if(!(2&this.f)){this.f|=6;for(let t=this.t;void 0!==t;t=t.x)t.t.N()}};v.prototype.peek=function(){if(!this.h())t();if(16&this.f)throw this.v;return this.v};Object.defineProperty(v.prototype,"value",{get(){if(1&this.f)t();const n=s(this);this.h();if(void 0!==n)n.i=this.i;if(16&this.f)throw this.v;return this.v}});function y(t){return new v(t)}function m(t){const e=t.u;t.u=void 0;if("function"==typeof e){u++;const _=i;i=void 0;try{e()}catch(n){t.f&=-2;t.f|=8;g(t);throw n}finally{i=_;n()}}}function g(t){for(let n=t.s;void 0!==n;n=n.n)n.S.U(n);t.x=void 0;t.s=void 0;m(t)}function b(t){if(i!==this)throw new Error("Out-of-order effect");d(this);i=t;this.f&=-2;if(8&this.f)g(this);n()}function k(t){this.x=t;this.u=void 0;this.s=void 0;this.o=void 0;this.f=32}k.prototype.c=function(){const t=this.S();try{if(8&this.f)return;if(void 0===this.x)return;const n=this.x();if("function"==typeof n)this.u=n}finally{t()}};k.prototype.S=function(){if(1&this.f)t();this.f|=1;this.f&=-9;m(this);p(this);u++;const n=i;i=this;return b.bind(this,n)};k.prototype.N=function(){if(!(2&this.f)){this.f|=2;this.o=_;_=this}};k.prototype.d=function(){this.f|=8;if(!(1&this.f))g(this)};function S(t){const n=new k(t);try{n.c()}catch(t){n.d();throw t}return n.d.bind(n)}var x,w,C,E,U,H,N,P,$,D={},T=[],V=/acit|ex(?:s|g|n|p|$)|rph|grid|ows|mnc|ntw|ine[ch]|zoo|^ord|itera/i,A=Array.isArray;function F(t,n){for(var e in n)t[e]=n[e];return t}function M(t){var n=t.parentNode;n&&n.removeChild(t)}function W(t,n,e){var i,_,o,r={};for(o in n)"key"==o?i=n[o]:"ref"==o?_=n[o]:r[o]=n[o];if(arguments.length>2&&(r.children=arguments.length>3?x.call(arguments,2):e),"function"==typeof t&&null!=t.defaultProps)for(o in t.defaultProps)void 0===r[o]&&(r[o]=t.defaultProps[o]);return O(t,r,i,_,null)}function O(t,n,e,i,_){var o={type:t,props:n,key:e,ref:i,__k:null,__:null,__b:0,__e:null,__d:void 0,__c:null,__h:null,constructor:void 0,__v:null==_?++C:_};return null==_&&null!=w.vnode&&w.vnode(o),o}function L(){return{current:null}}function R(t){return t.children}function I(t,n){this.props=t,this.context=n}function j(t,n){if(null==n)return t.__?j(t.__,t.__.__k.indexOf(t)+1):null;for(var e;nn&&U.sort(P));G.__r=0}function z(t,n,e,i,_,o,r,u,f,l,s){var c,h,a,p,d,v,y,m,g,b,k=0,S=i&&i.__k||T,x=S.length,w=x,C=n.length;for(e.__k=[],c=0;c0?O(p.type,p.props,p.key,p.ref?p.ref:null,p.__v):p)&&(p.__=e,p.__b=e.__b+1,-1===(m=X(p,S,y=c+k,w))?a=D:(a=S[m]||D,S[m]=void 0,w--),it(t,p,a,_,o,r,u,f,l,s),d=p.__e,(h=p.ref)&&a.ref!=h&&(a.ref&&rt(a.ref,null,p),s.push(h,p.__c||d,p)),null!=d&&(null==v&&(v=d),b=!(g=a===D||null===a.__v)&&m===y,g?-1==m&&k--:m!==y&&(m===y+1?(k++,b=!0):m>y?w>C-y?(k+=m-y,b=!0):k--:k=m(null!=f?1:0))for(;r>=0||u=0){if((f=n[r])&&_==f.key&&o===f.type)return r;r--}if(u2&&(u.children=arguments.length>3?x.call(arguments,2):e),O(t.type,u,i||t.key,_||t.ref,null)}function ht(t,n){var e={__c:n="__cC"+$++,__:t,Consumer:function(t,n){return t.children(n)},Provider:function(t){var e,i;return this.getChildContext||(e=[],(i={})[n]=this,this.getChildContext=function(){return i},this.shouldComponentUpdate=function(t){this.props.value!==t.value&&e.some((function(t){t.__e=!0,q(t)}))},this.sub=function(t){e.push(t);var n=t.componentWillUnmount;t.componentWillUnmount=function(){e.splice(e.indexOf(t),1),n&&n.call(t)}}),t.children}};return e.Provider.__=e.Consumer.contextType=e}x=T.slice,w={__e:function(t,n,e,i){for(var _,o,r;n=n.__;)if((_=n.__c)&&!_.__)try{if((o=_.constructor)&&null!=o.getDerivedStateFromError&&(_.setState(o.getDerivedStateFromError(t)),r=_.__d),null!=_.componentDidCatch&&(_.componentDidCatch(t,i||{}),r=_.__d),r)return _.__E=_}catch(n){t=n}throw t}},C=0,E=function(t){return null!=t&&void 0===t.constructor},I.prototype.setState=function(t,n){var e;e=null!=this.__s&&this.__s!==this.state?this.__s:this.__s=F({},this.state),"function"==typeof t&&(t=t(F({},e),this.props)),t&&F(e,t),null!=t&&this.__v&&(n&&this._sb.push(n),q(this))},I.prototype.forceUpdate=function(t){this.__v&&(this.__e=!0,t&&this.__h.push(t),q(this))},I.prototype.render=R,U=[],N="function"==typeof Promise?Promise.prototype.then.bind(Promise.resolve()):setTimeout,P=function(t,n){return t.__v.__b-n.__v.__b},G.__r=0,$=0;var at,pt,dt,vt,yt=0,mt=[],gt=[],bt=w.__b,kt=w.__r,St=w.diffed,xt=w.__c,wt=w.unmount;function Ct(t,n){w.__h&&w.__h(pt,t,yt||n),yt=0;var e=pt.__H||(pt.__H={__:[],__h:[]});return t>=e.__.length&&e.__.push({__V:gt}),e.__[t]}function Et(t){return yt=1,Ut(Bt,t)}function Ut(t,n,e){var i=Ct(at++,2);if(i.t=t,!i.__c&&(i.__=[e?e(n):Bt(void 0,n),function(t){var n=i.__N?i.__N[0]:i.__[0],e=i.t(n,t);n!==e&&(i.__N=[e,i.__[1]],i.__c.setState({}))}],i.__c=pt,!pt.u)){var _=function(t,n,e){if(!i.__c.__H)return!0;var _=i.__c.__H.__.filter((function(t){return t.__c}));if(_.every((function(t){return!t.__N})))return!o||o.call(this,t,n,e);var r=!1;return _.forEach((function(t){if(t.__N){var n=t.__[0];t.__=t.__N,t.__N=void 0,n!==t.__[0]&&(r=!0)}})),!(!r&&i.__c.props===t)&&(!o||o.call(this,t,n,e))};pt.u=!0;var o=pt.shouldComponentUpdate,r=pt.componentWillUpdate;pt.componentWillUpdate=function(t,n,e){if(this.__e){var i=o;o=void 0,_(t,n,e),o=i}r&&r.call(this,t,n,e)},pt.shouldComponentUpdate=_}return i.__N||i.__}function Ht(t,n){var e=Ct(at++,3);!w.__s&&jt(e.__H,n)&&(e.__=t,e.i=n,pt.__H.__h.push(e))}function Nt(t,n){var e=Ct(at++,4);!w.__s&&jt(e.__H,n)&&(e.__=t,e.i=n,pt.__h.push(e))}function Pt(t){return yt=5,Dt((function(){return{current:t}}),[])}function $t(t,n,e){yt=6,Nt((function(){return"function"==typeof t?(t(n()),function(){return t(null)}):t?(t.current=n(),function(){return t.current=null}):void 0}),null==e?e:e.concat(t))}function Dt(t,n){var e=Ct(at++,7);return jt(e.__H,n)?(e.__V=t(),e.i=n,e.__h=t,e.__V):e.__}function Tt(t,n){return yt=8,Dt((function(){return t}),n)}function Vt(t){var n=pt.context[t.__c],e=Ct(at++,9);return e.c=t,n?(null==e.__&&(e.__=!0,n.sub(pt)),n.props.value):t.__}function At(t,n){w.useDebugValue&&w.useDebugValue(n?n(t):t)}function Ft(t){var n=Ct(at++,10),e=Et();return n.__=t,pt.componentDidCatch||(pt.componentDidCatch=function(t,i){n.__&&n.__(t,i),e[1](t)}),[e[0],function(){e[1](void 0)}]}function Mt(){var t=Ct(at++,11);if(!t.__){for(var n=pt.__v;null!==n&&!n.__m&&null!==n.__;)n=n.__;var e=n.__m||(n.__m=[0,0]);t.__="P"+e[0]+"-"+e[1]++}return t.__}function Wt(){for(var t;t=mt.shift();)if(t.__P&&t.__H)try{t.__H.__h.forEach(Rt),t.__H.__h.forEach(It),t.__H.__h=[]}catch(u){t.__H.__h=[],w.__e(u,t.__v)}}w.__b=function(t){pt=null,bt&&bt(t)},w.__r=function(t){kt&&kt(t),at=0;var n=(pt=t.__c).__H;n&&(dt===pt?(n.__h=[],pt.__h=[],n.__.forEach((function(t){t.__N&&(t.__=t.__N),t.__V=gt,t.__N=t.i=void 0}))):(n.__h.forEach(Rt),n.__h.forEach(It),n.__h=[],at=0)),dt=pt},w.diffed=function(t){St&&St(t);var n=t.__c;n&&n.__H&&(n.__H.__h.length&&(1!==mt.push(n)&&vt===w.requestAnimationFrame||((vt=w.requestAnimationFrame)||Lt)(Wt)),n.__H.__.forEach((function(t){t.i&&(t.__H=t.i),t.__V!==gt&&(t.__=t.__V),t.i=void 0,t.__V=gt}))),dt=pt=null},w.__c=function(t,n){n.some((function(t){try{t.__h.forEach(Rt),t.__h=t.__h.filter((function(t){return!t.__||It(t)}))}catch(s){n.some((function(t){t.__h&&(t.__h=[])})),n=[],w.__e(s,t.__v)}})),xt&&xt(t,n)},w.unmount=function(t){wt&&wt(t);var n,e=t.__c;e&&e.__H&&(e.__H.__.forEach((function(t){try{Rt(t)}catch(t){n=t}})),e.__H=void 0,n&&w.__e(n,e.__v))};var Ot="function"==typeof requestAnimationFrame;function Lt(t){var n,e=function(){clearTimeout(i),Ot&&cancelAnimationFrame(n),setTimeout(t)},i=setTimeout(e,100);Ot&&(n=requestAnimationFrame(e))}function Rt(t){var n=pt,e=t.__c;"function"==typeof e&&(t.__c=void 0,e()),pt=n}function It(t){var n=pt;t.__c=t.__(),pt=n}function jt(t,n){return!t||t.length!==n.length||n.some((function(n,e){return n!==t[e]}))}function Bt(t,n){return"function"==typeof n?n(t):n}function qt(t,n){w[t]=n.bind(null,w[t]||(()=>{}))}let Gt,zt;function Jt(t){if(zt)zt();zt=t&&t.S()}function Kt({data:t}){const n=Xt(t);n.value=t;const e=Dt(()=>{let t=this.__v;while(t=t.__)if(t.__c){t.__c.__$f|=4;break}this.__$u.c=()=>{var t;if(!E(e.peek())&&3===(null==(t=this.base)?void 0:t.nodeType))this.base.data=e.peek();else{this.__$f|=1;this.setState({})}};return y(()=>{let t=n.value.value;return 0===t?0:!0===t?"":t||""})},[]);return e.value}Kt.displayName="_st";Object.defineProperties(c.prototype,{constructor:{configurable:!0,value:void 0},type:{configurable:!0,value:Kt},props:{configurable:!0,get(){return{data:this}}},__b:{configurable:!0,value:1}});qt("__b",(t,n)=>{if("string"==typeof n.type){let t,e=n.props;for(let i in e){if("children"===i)continue;let _=e[i];if(_ instanceof c){if(!t)n.__np=t={};t[i]=_;e[i]=_.peek()}}}t(n)});qt("__r",(t,n)=>{Jt();let e,i=n.__c;if(i){i.__$f&=-2;e=i.__$u;if(void 0===e)i.__$u=e=function(t){let n;S((function(){n=this}));n.c=()=>{i.__$f|=1;i.setState({})};return n}()}Gt=i;Jt(e);t(n)});qt("__e",(t,n,e,i)=>{Jt();Gt=void 0;t(n,e,i)});qt("diffed",(t,n)=>{Jt();Gt=void 0;let e;if("string"==typeof n.type&&(e=n.__e)){let t=n.__np,i=n.props;if(t){let n=e.U;if(n)for(let e in n){let i=n[e];if(void 0!==i&&!(e in t)){i.d();n[e]=void 0}}else{n={};e.U=n}for(let _ in t){let o=n[_],r=t[_];if(void 0===o){o=Qt(e,_,r,i);n[_]=o}else o.o(r,i)}}}t(n)});function Qt(t,n,e,i){const _=n in t&&void 0===t.ownerSVGElement,o=h(e);return{o:(t,n)=>{o.value=t;i=n},d:S(()=>{const e=o.value.value;if(i[n]!==e){i[n]=e;if(_)t[n]=e;else if(e)t.setAttribute(n,e);else t.removeAttribute(n)}})}}qt("unmount",(t,n)=>{if("string"==typeof n.type){let t=n.__e;if(t){const n=t.U;if(n){t.U=void 0;for(let t in n){let e=n[t];if(e)e.d()}}}}else{let t=n.__c;if(t){const n=t.__$u;if(n){t.__$u=void 0;n.d()}}}t(n)});qt("__h",(t,n,e,i)=>{if(i<3||9===i)n.__$f|=2;t(n,e,i)});I.prototype.shouldComponentUpdate=function(t,n){const e=this.__$u;if(!(e&&void 0!==e.s||4&this.__$f))return!0;if(3&this.__$f)return!0;for(let i in n)return!0;for(let i in t)if("__source"!==i&&t[i]!==this.props[i])return!0;for(let i in this.props)if(!(i in t))return!0;return!1};function Xt(t){return Dt(()=>h(t),[])}function Yt(t){const n=Pt(t);n.current=t;Gt.__$f|=4;return Dt(()=>y(()=>n.current()),[])}function Zt(t){const n=Pt(t);n.current=t;Ht(()=>S(()=>n.current()),[])}var tn=function(t,n,e,i){var _;n[0]=0;for(var o=1;o=5&&((_||!t&&5===i)&&(r.push(i,0,_,e),i=6),t&&(r.push(i,t,0,e),i=6)),_=""},f=0;f"===n?(i=1,_=""):_=n+_[0]:o?n===o?o="":_+=n:'"'===n||"'"===n?o=n:">"===n?(u(),i=1):i&&("="===n?(i=5,e=_,_=""):"/"===n&&(i<5||">"===t[f][l+1])?(u(),3===i&&(r=r[0]),i=r,(r=r[0]).push(2,0,i),i=0):" "===n||"\t"===n||"\n"===n||"\r"===n?(u(),i=2):_+=n),3===i&&"!--"===_&&(i=4,r=r[0])}return u(),r}(t)),n),arguments,[])).length>1?n:n[0]}var _n=en.bind(W);export{I as Component,R as Fragment,c as Signal,e as batch,ct as cloneElement,y as computed,ht as createContext,W as createElement,L as createRef,S as effect,W as h,_n as html,st as hydrate,E as isValidElement,w as options,lt as render,h as signal,K as toChildArray,r as untracked,Tt as useCallback,Yt as useComputed,Vt as useContext,At as useDebugValue,Ht as useEffect,Ft as useErrorBoundary,Mt as useId,$t as useImperativeHandle,Nt as useLayoutEffect,Dt as useMemo,Ut as useReducer,Pt as useRef,Xt as useSignal,Zt as useSignalEffect,Et as useState}; +function t(){throw new Error("Cycle detected")}const n=Symbol.for("preact-signals");function e(){if(f>1){f--;return}let t,n=!1;while(void 0!==o){let _=o;o=void 0;l++;while(void 0!==_){const i=_.o;_.o=void 0;_.f&=-3;if(!(8&_.f)&&p(_))try{_.c()}catch(e){if(!n){t=e;n=!0}}_=i}}l=0;f--;if(n)throw t}function _(t){if(f>0)return t();f++;try{return t()}finally{e()}}let i,o,r=0;function u(t){if(r>0)return t();const n=i;i=void 0;r++;try{return t()}finally{r--;i=n}}let f=0,l=0,s=0;function c(t){if(void 0===i)return;let n=t.n;if(void 0===n||n.t!==i){n={i:0,S:t,p:i.s,n:void 0,t:i,e:void 0,x:void 0,r:n};if(void 0!==i.s)i.s.n=n;i.s=n;t.n=n;if(32&i.f)t.S(n);return n}else if(-1===n.i){n.i=0;if(void 0!==n.n){n.n.p=n.p;if(void 0!==n.p)n.p.n=n.n;n.p=i.s;n.n=void 0;i.s.n=n;i.s=n}return n}}function h(t){this.v=t;this.i=0;this.n=void 0;this.t=void 0}h.prototype.brand=n;h.prototype.h=function(){return!0};h.prototype.S=function(t){if(this.t!==t&&void 0===t.e){t.x=this.t;if(void 0!==this.t)this.t.e=t;this.t=t}};h.prototype.U=function(t){if(void 0!==this.t){const n=t.e,e=t.x;if(void 0!==n){n.x=e;t.e=void 0}if(void 0!==e){e.e=n;t.x=void 0}if(t===this.t)this.t=e}};h.prototype.subscribe=function(t){const n=this;return w((function(){const e=n.value,_=32&this.f;this.f&=-33;try{t(e)}finally{this.f|=_}}))};h.prototype.valueOf=function(){return this.value};h.prototype.toString=function(){return this.value+""};h.prototype.toJSON=function(){return this.value};h.prototype.peek=function(){return this.v};Object.defineProperty(h.prototype,"value",{get(){const t=c(this);if(void 0!==t)t.i=this.i;return this.v},set(n){if(i instanceof y)!function(){throw new Error("Computed cannot have side-effects")}();if(n!==this.v){if(l>100)t();this.v=n;this.i++;s++;f++;try{for(let t=this.t;void 0!==t;t=t.x)t.t.N()}finally{e()}}}});function a(t){return new h(t)}function p(t){for(let n=t.s;void 0!==n;n=n.n)if(n.S.i!==n.i||!n.S.h()||n.S.i!==n.i)return!0;return!1}function d(t){for(let n=t.s;void 0!==n;n=n.n){const e=n.S.n;if(void 0!==e)n.r=e;n.S.n=n;n.i=-1;if(void 0===n.n){t.s=n;break}}}function v(t){let n,e=t.s;while(void 0!==e){const t=e.p;if(-1===e.i){e.S.U(e);if(void 0!==t)t.n=e.n;if(void 0!==e.n)e.n.p=t}else n=e;e.S.n=e.r;if(void 0!==e.r)e.r=void 0;e=t}t.s=n}function y(t){h.call(this,void 0);this.x=t;this.s=void 0;this.g=s-1;this.f=4}(y.prototype=new h).h=function(){this.f&=-3;if(1&this.f)return!1;if(32==(36&this.f))return!0;this.f&=-5;if(this.g===s)return!0;this.g=s;this.f|=1;if(this.i>0&&!p(this)){this.f&=-2;return!0}const t=i;try{d(this);i=this;const t=this.x();if(16&this.f||this.v!==t||0===this.i){this.v=t;this.f&=-17;this.i++}}catch(t){this.v=t;this.f|=16;this.i++}i=t;v(this);this.f&=-2;return!0};y.prototype.S=function(t){if(void 0===this.t){this.f|=36;for(let t=this.s;void 0!==t;t=t.n)t.S.S(t)}h.prototype.S.call(this,t)};y.prototype.U=function(t){if(void 0!==this.t){h.prototype.U.call(this,t);if(void 0===this.t){this.f&=-33;for(let t=this.s;void 0!==t;t=t.n)t.S.U(t)}}};y.prototype.N=function(){if(!(2&this.f)){this.f|=6;for(let t=this.t;void 0!==t;t=t.x)t.t.N()}};y.prototype.peek=function(){if(!this.h())t();if(16&this.f)throw this.v;return this.v};Object.defineProperty(y.prototype,"value",{get(){if(1&this.f)t();const n=c(this);this.h();if(void 0!==n)n.i=this.i;if(16&this.f)throw this.v;return this.v}});function m(t){return new y(t)}function g(t){const n=t.u;t.u=void 0;if("function"==typeof n){f++;const _=i;i=void 0;try{n()}catch(n){t.f&=-2;t.f|=8;b(t);throw n}finally{i=_;e()}}}function b(t){for(let n=t.s;void 0!==n;n=n.n)n.S.U(n);t.x=void 0;t.s=void 0;g(t)}function k(t){if(i!==this)throw new Error("Out-of-order effect");v(this);i=t;this.f&=-2;if(8&this.f)b(this);e()}function S(t){this.x=t;this.u=void 0;this.s=void 0;this.o=void 0;this.f=32}S.prototype.c=function(){const t=this.S();try{if(8&this.f)return;if(void 0===this.x)return;const n=this.x();if("function"==typeof n)this.u=n}finally{t()}};S.prototype.S=function(){if(1&this.f)t();this.f|=1;this.f&=-9;g(this);d(this);f++;const n=i;i=this;return k.bind(this,n)};S.prototype.N=function(){if(!(2&this.f)){this.f|=2;this.o=o;o=this}};S.prototype.d=function(){this.f|=8;if(!(1&this.f))b(this)};function w(t){const n=new S(t);try{n.c()}catch(t){n.d();throw t}return n.d.bind(n)}var x,C,E,U,H,P,N,$,D,T={},V=[],A=/acit|ex(?:s|g|n|p|$)|rph|grid|ows|mnc|ntw|ine[ch]|zoo|^ord|itera/i,F=Array.isArray;function M(t,n){for(var e in n)t[e]=n[e];return t}function W(t){var n=t.parentNode;n&&n.removeChild(t)}function L(t,n,e){var _,i,o,r={};for(o in n)"key"==o?_=n[o]:"ref"==o?i=n[o]:r[o]=n[o];if(arguments.length>2&&(r.children=arguments.length>3?x.call(arguments,2):e),"function"==typeof t&&null!=t.defaultProps)for(o in t.defaultProps)void 0===r[o]&&(r[o]=t.defaultProps[o]);return O(t,r,_,i,null)}function O(t,n,e,_,i){var o={type:t,props:n,key:e,ref:_,__k:null,__:null,__b:0,__e:null,__d:void 0,__c:null,constructor:void 0,__v:null==i?++E:i,__i:-1,__u:0};return null==i&&null!=C.vnode&&C.vnode(o),o}function R(){return{current:null}}function j(t){return t.children}function I(t,n){this.props=t,this.context=n}function q(t,n){if(null==n)return t.__?q(t.__,t.__i+1):null;for(var e;ne?(ot(_,n,i),i.length=_.length=0,n=void 0,H.sort($)):n&&C.__c&&C.__c(n,V));n&&ot(_,n,i),J.__r=0}function K(t,n,e,_,i,o,r,u,f,l,s){var c,h,a,p,d,v=_&&_.__k||V,y=n.length;for(e.__d=f,Q(e,n,v),f=e.__d,c=0;c0?O(i.type,i.props,i.key,i.ref?i.ref:null,i.__v):i)?(i.__=t,i.__b=t.__b+1,u=Z(i,e,r,s),i.__i=u,o=null,-1!==u&&(s--,(o=e[u])&&(o.__u|=131072)),null==o||null===o.__v?(-1==u&&c--,"function"!=typeof i.type&&(i.__u|=65536)):u!==r&&(u===r+1?c++:u>r?s>f-r?c+=u-r:c--:u(null!=f&&0==(131072&f.__u)?1:0))for(;r>=0||u=0){if((f=n[r])&&0==(131072&f.__u)&&i==f.key&&o===f.type)return r;r--}if(u2&&(u.children=arguments.length>3?x.call(arguments,2):e),O(t.type,u,_||t.key,i||t.ref,null)}function at(t,n){var e={__c:n="__cC"+D++,__:t,Consumer:function(t,n){return t.children(n)},Provider:function(t){var e,_;return this.getChildContext||(e=[],(_={})[n]=this,this.getChildContext=function(){return _},this.shouldComponentUpdate=function(t){this.props.value!==t.value&&e.some((function(t){t.__e=!0,z(t)}))},this.sub=function(t){e.push(t);var n=t.componentWillUnmount;t.componentWillUnmount=function(){e.splice(e.indexOf(t),1),n&&n.call(t)}}),t.children}};return e.Provider.__=e.Consumer.contextType=e}x=V.slice,C={__e:function(t,n,e,_){for(var i,o,r;n=n.__;)if((i=n.__c)&&!i.__)try{if((o=i.constructor)&&null!=o.getDerivedStateFromError&&(i.setState(o.getDerivedStateFromError(t)),r=i.__d),null!=i.componentDidCatch&&(i.componentDidCatch(t,_||{}),r=i.__d),r)return i.__E=i}catch(n){t=n}throw t}},E=0,U=function(t){return null!=t&&null==t.constructor},I.prototype.setState=function(t,n){var e;e=null!=this.__s&&this.__s!==this.state?this.__s:this.__s=M({},this.state),"function"==typeof t&&(t=t(M({},e),this.props)),t&&M(e,t),null!=t&&this.__v&&(n&&this._sb.push(n),z(this))},I.prototype.forceUpdate=function(t){this.__v&&(this.__e=!0,t&&this.__h.push(t),z(this))},I.prototype.render=j,H=[],N="function"==typeof Promise?Promise.prototype.then.bind(Promise.resolve()):setTimeout,$=function(t,n){return t.__v.__b-n.__v.__b},J.__r=0,D=0;var pt,dt,vt,yt,mt=0,gt=[],bt=[],kt=C,St=kt.__b,wt=kt.__r,xt=kt.diffed,Ct=kt.__c,Et=kt.unmount,Ut=kt.__;function Ht(t,n){kt.__h&&kt.__h(dt,t,mt||n),mt=0;var e=dt.__H||(dt.__H={__:[],__h:[]});return t>=e.__.length&&e.__.push({__V:bt}),e.__[t]}function Pt(t){return mt=1,Nt(zt,t)}function Nt(t,n,e){var _=Ht(pt++,2);if(_.t=t,!_.__c&&(_.__=[e?e(n):zt(void 0,n),function(t){var n=_.__N?_.__N[0]:_.__[0],e=_.t(n,t);n!==e&&(_.__N=[e,_.__[1]],_.__c.setState({}))}],_.__c=dt,!dt.u)){var i=function(t,n,e){if(!_.__c.__H)return!0;var i=_.__c.__H.__.filter((function(t){return!!t.__c}));if(i.every((function(t){return!t.__N})))return!o||o.call(this,t,n,e);var r=!1;return i.forEach((function(t){if(t.__N){var n=t.__[0];t.__=t.__N,t.__N=void 0,n!==t.__[0]&&(r=!0)}})),!(!r&&_.__c.props===t)&&(!o||o.call(this,t,n,e))};dt.u=!0;var o=dt.shouldComponentUpdate,r=dt.componentWillUpdate;dt.componentWillUpdate=function(t,n,e){if(this.__e){var _=o;o=void 0,i(t,n,e),o=_}r&&r.call(this,t,n,e)},dt.shouldComponentUpdate=i}return _.__N||_.__}function $t(t,n){var e=Ht(pt++,3);!kt.__s&&Gt(e.__H,n)&&(e.__=t,e.i=n,dt.__H.__h.push(e))}function Dt(t,n){var e=Ht(pt++,4);!kt.__s&&Gt(e.__H,n)&&(e.__=t,e.i=n,dt.__h.push(e))}function Tt(t){return mt=5,At((function(){return{current:t}}),[])}function Vt(t,n,e){mt=6,Dt((function(){return"function"==typeof t?(t(n()),function(){return t(null)}):t?(t.current=n(),function(){return t.current=null}):void 0}),null==e?e:e.concat(t))}function At(t,n){var e=Ht(pt++,7);return Gt(e.__H,n)?(e.__V=t(),e.i=n,e.__h=t,e.__V):e.__}function Ft(t,n){return mt=8,At((function(){return t}),n)}function Mt(t){var n=dt.context[t.__c],e=Ht(pt++,9);return e.c=t,n?(null==e.__&&(e.__=!0,n.sub(dt)),n.props.value):t.__}function Wt(t,n){kt.useDebugValue&&kt.useDebugValue(n?n(t):t)}function Lt(t){var n=Ht(pt++,10),e=Pt();return n.__=t,dt.componentDidCatch||(dt.componentDidCatch=function(t,_){n.__&&n.__(t,_),e[1](t)}),[e[0],function(){e[1](void 0)}]}function Ot(){var t=Ht(pt++,11);if(!t.__){for(var n=dt.__v;null!==n&&!n.__m&&null!==n.__;)n=n.__;var e=n.__m||(n.__m=[0,0]);t.__="P"+e[0]+"-"+e[1]++}return t.__}function Rt(){for(var t;t=gt.shift();)if(t.__P&&t.__H)try{t.__H.__h.forEach(qt),t.__H.__h.forEach(Bt),t.__H.__h=[]}catch(n){t.__H.__h=[],kt.__e(n,t.__v)}}kt.__b=function(t){dt=null,St&&St(t)},kt.__=function(t,n){t&&n.__k&&n.__k.__m&&(t.__m=n.__k.__m),Ut&&Ut(t,n)},kt.__r=function(t){wt&&wt(t),pt=0;var n=(dt=t.__c).__H;n&&(vt===dt?(n.__h=[],dt.__h=[],n.__.forEach((function(t){t.__N&&(t.__=t.__N),t.__V=bt,t.__N=t.i=void 0}))):(n.__h.forEach(qt),n.__h.forEach(Bt),n.__h=[],pt=0)),vt=dt},kt.diffed=function(t){xt&&xt(t);var n=t.__c;n&&n.__H&&(n.__H.__h.length&&(1!==gt.push(n)&&yt===kt.requestAnimationFrame||((yt=kt.requestAnimationFrame)||It)(Rt)),n.__H.__.forEach((function(t){t.i&&(t.__H=t.i),t.__V!==bt&&(t.__=t.__V),t.i=void 0,t.__V=bt}))),vt=dt=null},kt.__c=function(t,n){n.some((function(t){try{t.__h.forEach(qt),t.__h=t.__h.filter((function(t){return!t.__||Bt(t)}))}catch(u){n.some((function(t){t.__h&&(t.__h=[])})),n=[],kt.__e(u,t.__v)}})),Ct&&Ct(t,n)},kt.unmount=function(t){Et&&Et(t);var n,e=t.__c;e&&e.__H&&(e.__H.__.forEach((function(t){try{qt(t)}catch(t){n=t}})),e.__H=void 0,n&&kt.__e(n,e.__v))};var jt="function"==typeof requestAnimationFrame;function It(t){var n,e=function(){clearTimeout(_),jt&&cancelAnimationFrame(n),setTimeout(t)},_=setTimeout(e,100);jt&&(n=requestAnimationFrame(e))}function qt(t){var n=dt,e=t.__c;"function"==typeof e&&(t.__c=void 0,e()),dt=n}function Bt(t){var n=dt;t.__c=t.__(),dt=n}function Gt(t,n){return!t||t.length!==n.length||n.some((function(n,e){return n!==t[e]}))}function zt(t,n){return"function"==typeof n?n(t):n}function Jt(t,n){C[t]=n.bind(null,C[t]||(()=>{}))}let Kt,Qt;function Xt(t){if(Qt)Qt();Qt=t&&t.S()}function Yt({data:t}){const n=tn(t);n.value=t;const e=At(()=>{let t=this.__v;while(t=t.__)if(t.__c){t.__c.__$f|=4;break}this.__$u.c=()=>{var t;if(!U(e.peek())&&3===(null==(t=this.base)?void 0:t.nodeType))this.base.data=e.peek();else{this.__$f|=1;this.setState({})}};return m(()=>{let t=n.value.value;return 0===t?0:!0===t?"":t||""})},[]);return e.value}Yt.displayName="_st";Object.defineProperties(h.prototype,{constructor:{configurable:!0,value:void 0},type:{configurable:!0,value:Yt},props:{configurable:!0,get(){return{data:this}}},__b:{configurable:!0,value:1}});Jt("__b",(t,n)=>{if("string"==typeof n.type){let t,e=n.props;for(let _ in e){if("children"===_)continue;let i=e[_];if(i instanceof h){if(!t)n.__np=t={};t[_]=i;e[_]=i.peek()}}}t(n)});Jt("__r",(t,n)=>{Xt();let e,_=n.__c;if(_){_.__$f&=-2;e=_.__$u;if(void 0===e)_.__$u=e=function(t){let n;w((function(){n=this}));n.c=()=>{_.__$f|=1;_.setState({})};return n}()}Kt=_;Xt(e);t(n)});Jt("__e",(t,n,e,_)=>{Xt();Kt=void 0;t(n,e,_)});Jt("diffed",(t,n)=>{Xt();Kt=void 0;let e;if("string"==typeof n.type&&(e=n.__e)){let t=n.__np,_=n.props;if(t){let n=e.U;if(n)for(let e in n){let _=n[e];if(void 0!==_&&!(e in t)){_.d();n[e]=void 0}}else{n={};e.U=n}for(let i in t){let o=n[i],r=t[i];if(void 0===o){o=Zt(e,i,r,_);n[i]=o}else o.o(r,_)}}}t(n)});function Zt(t,n,e,_){const i=n in t&&void 0===t.ownerSVGElement,o=a(e);return{o:(t,n)=>{o.value=t;_=n},d:w(()=>{const e=o.value.value;if(_[n]!==e){_[n]=e;if(i)t[n]=e;else if(e)t.setAttribute(n,e);else t.removeAttribute(n)}})}}Jt("unmount",(t,n)=>{if("string"==typeof n.type){let t=n.__e;if(t){const n=t.U;if(n){t.U=void 0;for(let t in n){let e=n[t];if(e)e.d()}}}}else{let t=n.__c;if(t){const n=t.__$u;if(n){t.__$u=void 0;n.d()}}}t(n)});Jt("__h",(t,n,e,_)=>{if(_<3||9===_)n.__$f|=2;t(n,e,_)});I.prototype.shouldComponentUpdate=function(t,n){const e=this.__$u;if(!(e&&void 0!==e.s||4&this.__$f))return!0;if(3&this.__$f)return!0;for(let _ in n)return!0;for(let _ in t)if("__source"!==_&&t[_]!==this.props[_])return!0;for(let _ in this.props)if(!(_ in t))return!0;return!1};function tn(t){return At(()=>a(t),[])}function nn(t){const n=Tt(t);n.current=t;Kt.__$f|=4;return At(()=>m(()=>n.current()),[])}function en(t){const n=Tt(t);n.current=t;$t(()=>w(()=>n.current()),[])}var _n=function(t,n,e,_){var i;n[0]=0;for(var o=1;o=5&&((i||!t&&5===_)&&(r.push(_,0,i,e),_=6),t&&(r.push(_,t,0,e),_=6)),i=""},f=0;f"===n?(_=1,i=""):i=n+i[0]:o?n===o?o="":i+=n:'"'===n||"'"===n?o=n:">"===n?(u(),_=1):_&&("="===n?(_=5,e=i,i=""):"/"===n&&(_<5||">"===t[f][l+1])?(u(),3===_&&(r=r[0]),_=r,(r=r[0]).push(2,0,_),_=0):" "===n||"\t"===n||"\n"===n||"\r"===n?(u(),_=2):i+=n),3===_&&"!--"===i&&(_=4,r=r[0])}return u(),r}(t)),n),arguments,[])).length>1?n:n[0]}var un=rn.bind(L);export{I as Component,j as Fragment,h as Signal,_ as batch,ht as cloneElement,m as computed,at as createContext,L as createElement,R as createRef,w as effect,L as h,un as html,ct as hydrate,U as isValidElement,C as options,st as render,a as signal,Y as toChildArray,u as untracked,Ft as useCallback,nn as useComputed,Mt as useContext,Wt as useDebugValue,$t as useEffect,Lt as useErrorBoundary,Ot as useId,Vt as useImperativeHandle,Dt as useLayoutEffect,At as useMemo,Nt as useReducer,Tt as useRef,tn as useSignal,en as useSignalEffect,Pt as useState}; diff --git a/examples/server/public/json-schema-to-grammar.mjs b/examples/server/public/json-schema-to-grammar.mjs index 3f1b255c2..17bd60fde 100644 --- a/examples/server/public/json-schema-to-grammar.mjs +++ b/examples/server/public/json-schema-to-grammar.mjs @@ -1,112 +1,538 @@ +// WARNING: This file was ported from json-schema-to-grammar.py, please fix bugs / add features there first. const SPACE_RULE = '" "?'; const PRIMITIVE_RULES = { boolean: '("true" | "false") space', number: '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space', integer: '("-"? ([0-9] | [1-9] [0-9]*)) space', + value: 'object | array | string | number | boolean', + object: '"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', + array: '"[" space ( value ("," space value)* )? "]" space', + uuid: '"\\"" ' + [8, 4, 4, 4, 12].map(n => [...new Array(n)].map(_ => '[0-9a-fA-F]').join('')).join(' "-" ') + ' "\\"" space', string: ` "\\"" ( [^"\\\\] | "\\\\" (["\\\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) )* "\\"" space`, null: '"null" space', }; +const OBJECT_RULE_NAMES = ['object', 'array', 'string', 'number', 'boolean', 'null', 'value']; + +// TODO: support "uri", "email" string formats +const DATE_RULES = { + 'date' : '[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', + 'time' : '([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', + 'date-time': 'date "T" time', + 'date-string': '"\\"" date "\\"" space', + 'time-string': '"\\"" time "\\"" space', + 'date-time-string': '"\\"" date-time "\\"" space', +}; + +const RESERVED_NAMES = {'root': true, ...PRIMITIVE_RULES, ...DATE_RULES}; const INVALID_RULE_CHARS_RE = /[^\dA-Za-z-]+/g; const GRAMMAR_LITERAL_ESCAPE_RE = /[\n\r"]/g; -const GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"'}; +const GRAMMAR_RANGE_LITERAL_ESCAPE_RE = /[\n\r"\]\-\\]/g; +const GRAMMAR_LITERAL_ESCAPES = { '\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]' }; + +const NON_LITERAL_SET = new Set('|.()[]{}*+?'); +const ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = new Set('[]()|{}*+?'); export class SchemaConverter { - constructor(propOrder) { - this._propOrder = propOrder || {}; - this._rules = new Map(); - this._rules.set('space', SPACE_RULE); + constructor(options) { + this._propOrder = options.prop_order || {}; + this._allowFetch = options.allow_fetch || false; + this._dotall = options.dotall || false; + this._rules = {'space': SPACE_RULE}; + this._refs = {}; + this._refsBeingResolved = new Set(); } _formatLiteral(literal) { - const escaped = JSON.stringify(literal).replace( + const escaped = literal.replace( GRAMMAR_LITERAL_ESCAPE_RE, m => GRAMMAR_LITERAL_ESCAPES[m] ); return `"${escaped}"`; } + _formatRangeChar(literal) { + return JSON.stringify(literal).slice(1, -1).replace( + GRAMMAR_RANGE_LITERAL_ESCAPE_RE, + m => GRAMMAR_LITERAL_ESCAPES[m] + ); + } + _addRule(name, rule) { let escName = name.replace(INVALID_RULE_CHARS_RE, '-'); let key = escName; - if (this._rules.has(escName)) { - if (this._rules.get(escName) === rule) { + if (escName in this._rules) { + if (this._rules[escName] === rule) { return key; } let i = 0; - while (this._rules.has(`${escName}${i}`)) { + while ((`${escName}${i}` in this._rules) && (this._rules[`${escName}${i}`] !== rule)) { i += 1; } key = `${escName}${i}`; } - this._rules.set(key, rule); + this._rules[key] = rule; return key; } + async resolveRefs(schema, url) { + const visit = async (n) => { + if (Array.isArray(n)) { + return Promise.all(n.map(visit)); + } else if (typeof n === 'object' && n !== null) { + let ref = n.$ref; + let target; + if (ref !== undefined && !this._refs[ref]) { + if (ref.startsWith('https://')) { + if (!this._allowFetch) { + throw new Error('Fetching remote schemas is not allowed (use --allow-fetch for force)'); + } + const fetch = (await import('node-fetch')).default; + + const fragSplit = ref.split('#'); + const baseUrl = fragSplit[0]; + + target = this._refs[baseUrl]; + if (!target) { + target = await this.resolveRefs(await fetch(ref).then(res => res.json()), baseUrl); + this._refs[baseUrl] = target; + } + + if (fragSplit.length === 1 || fragSplit[fragSplit.length - 1] === '') { + return target; + } + } else if (ref.startsWith('#/')) { + target = schema; + ref = `${url}${ref}`; + n.$ref = ref; + } else { + throw new Error(`Unsupported ref ${ref}`); + } + + const selectors = ref.split('#')[1].split('/').slice(1); + for (const sel of selectors) { + if (!target || !(sel in target)) { + throw new Error(`Error resolving ref ${ref}: ${sel} not in ${JSON.stringify(target)}`); + } + target = target[sel]; + } + + this._refs[ref] = target; + } else { + await Promise.all(Object.values(n).map(visit)); + } + } + + return n; + }; + + return visit(schema); + } + + _generateUnionRule(name, altSchemas) { + return altSchemas + .map((altSchema, i) => this.visit(altSchema, `${name ?? ''}${name ? '-' : 'alternative-'}${i}`)) + .join(' | '); + } + + _visitPattern(pattern, name) { + if (!pattern.startsWith('^') || !pattern.endsWith('$')) { + throw new Error('Pattern must start with "^" and end with "$"'); + } + pattern = pattern.slice(1, -1); + const subRuleIds = {}; + + let i = 0; + const length = pattern.length; + + const getDot = () => { + let rule; + if (this._dotall) { + rule = '[\\U00000000-\\U0010FFFF]'; + } else { + // Accept any character... except \n and \r line break chars (\x0A and \xOD) + rule = '[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]'; + } + return this._addRule('dot', rule); + }; + + + const toRule = ([s, isLiteral]) => isLiteral ? "\"" + s + "\"" : s; + + const transform = () => { + const start = i; + // For each component of this sequence, store its string representation and whether it's a literal. + // We only need a flat structure here to apply repetition operators to the last item, and + // to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially + // (GBNF's syntax is luckily very close to regular expressions!) + const seq = []; + + const joinSeq = () => { + const ret = []; + for (const [isLiteral, g] of groupBy(seq, x => x[1])) { + if (isLiteral) { + ret.push([[...g].map(x => x[0]).join(''), true]); + } else { + ret.push(...g); + } + } + if (ret.length === 1) { + return ret[0]; + } + return [ret.map(x => toRule(x)).join(' '), false]; + }; + + while (i < length) { + const c = pattern[i]; + if (c === '.') { + seq.push([getDot(), false]); + i += 1; + } else if (c === '(') { + i += 1; + if (i < length) { + if (pattern[i] === '?') { + throw new Error(`Unsupported pattern syntax "${pattern[i]}" at index ${i} of /${pattern}/`); + } + } + seq.push([`(${toRule(transform())})`, false]); + } else if (c === ')') { + i += 1; + if (start <= 0 || pattern[start - 1] !== '(') { + throw new Error(`Unbalanced parentheses; start = ${start}, i = ${i}, pattern = ${pattern}`); + } + return joinSeq(); + } else if (c === '[') { + let squareBrackets = c; + i += 1; + while (i < length && pattern[i] !== ']') { + if (pattern[i] === '\\') { + squareBrackets += pattern.slice(i, i + 2); + i += 2; + } else { + squareBrackets += pattern[i]; + i += 1; + } + } + if (i >= length) { + throw new Error(`Unbalanced square brackets; start = ${start}, i = ${i}, pattern = ${pattern}`); + } + squareBrackets += ']'; + i += 1; + seq.push([squareBrackets, false]); + } else if (c === '|') { + seq.push(['|', false]); + i += 1; + } else if (c === '*' || c === '+' || c === '?') { + seq[seq.length - 1] = [toRule(seq[seq.length - 1]) + c, false]; + i += 1; + } else if (c === '{') { + let curlyBrackets = c; + i += 1; + while (i < length && pattern[i] !== '}') { + curlyBrackets += pattern[i]; + i += 1; + } + if (i >= length) { + throw new Error(`Unbalanced curly brackets; start = ${start}, i = ${i}, pattern = ${pattern}`); + } + curlyBrackets += '}'; + i += 1; + const nums = curlyBrackets.slice(1, -1).split(',').map(s => s.trim()); + let minTimes, maxTimes; + if (nums.length === 1) { + minTimes = parseInt(nums[0], 10); + maxTimes = minTimes; + } else { + if (nums.length !== 2) { + throw new Error(`Invalid quantifier ${curlyBrackets}`); + } + minTimes = nums[0] ? parseInt(nums[0], 10) : 0; + maxTimes = nums[1] ? parseInt(nums[1], 10) : Infinity; + } + + let [sub, subIsLiteral] = seq[seq.length - 1]; + + if (minTimes === 0 && maxTimes === Infinity) { + seq[seq.length - 1] = [`${sub}*`, false]; + } else if (minTimes === 0 && maxTimes === 1) { + seq[seq.length - 1] = [`${sub}?`, false]; + } else if (minTimes === 1 && maxTimes === Infinity) { + seq[seq.length - 1] = [`${sub}+`, false]; + } else { + if (!subIsLiteral) { + let id = subRuleIds[sub]; + if (id === undefined) { + id = this._addRule(`${name}-${Object.keys(subRuleIds).length + 1}`, sub); + subRuleIds[sub] = id; + } + sub = id; + } + + const repeatedSub = Array.from({ length: minTimes }, () => subIsLiteral ? `"${sub.slice(1, -1).repeat(minTimes)}"` : sub); + const optionalSub = maxTimes !== undefined ? Array.from({ length: maxTimes - minTimes }, () => `${sub}?`) : [`${sub}*`]; + seq[seq.length - 1] = [repeatedSub.concat(optionalSub).join(' '), false]; + } + } else { + let literal = ''; + while (i < length) { + if (pattern[i] === '\\' && i < length - 1) { + const next = pattern[i + 1]; + if (ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.has(next)) { + i += 1; + literal += pattern[i]; + i += 1; + } else { + literal += pattern.slice(i, i + 2); + i += 2; + } + } else if (pattern[i] === '"') { + literal += '\\"'; + i += 1; + } else if (!NON_LITERAL_SET.has(pattern[i]) && + (i === length - 1 || literal === '' || pattern[i + 1] === '.' || !NON_LITERAL_SET.has(pattern[i+1]))) { + literal += pattern[i]; + i += 1; + } else { + break; + } + } + if (literal !== '') { + seq.push([literal, true]); + } + } + } + + return joinSeq(); + }; + + return this._addRule(name, "\"\\\"\" " + toRule(transform()) + " \"\\\"\" space") + } + + _resolveRef(ref) { + let refName = ref.split('/').pop(); + if (!(refName in this._rules) && !this._refsBeingResolved.has(ref)) { + this._refsBeingResolved.add(ref); + const resolved = this._refs[ref]; + refName = this.visit(resolved, refName); + this._refsBeingResolved.delete(ref); + } + return refName; + } + + _generateConstantRule(value) { + return this._formatLiteral(JSON.stringify(value)); + } + visit(schema, name) { const schemaType = schema.type; - const ruleName = name || 'root'; + const schemaFormat = schema.format; + const ruleName = name in RESERVED_NAMES ? name + '-' : name == '' ? 'root' : name; - if (schema.oneOf || schema.anyOf) { - const rule = (schema.oneOf || schema.anyOf).map((altSchema, i) => - this.visit(altSchema, `${name}${name ? "-" : ""}${i}`) - ).join(' | '); - - return this._addRule(ruleName, rule); + const ref = schema.$ref; + if (ref !== undefined) { + return this._addRule(ruleName, this._resolveRef(ref)); + } else if (schema.oneOf || schema.anyOf) { + return this._addRule(ruleName, this._generateUnionRule(name, schema.oneOf || schema.anyOf)); + } else if (Array.isArray(schemaType)) { + return this._addRule(ruleName, this._generateUnionRule(name, schemaType.map(t => ({ type: t })))); } else if ('const' in schema) { - return this._addRule(ruleName, this._formatLiteral(schema.const)); + return this._addRule(ruleName, this._generateConstantRule(schema.const)); } else if ('enum' in schema) { - const rule = schema.enum.map(v => this._formatLiteral(v)).join(' | '); + const rule = schema.enum.map(v => this._generateConstantRule(v)).join(' | '); return this._addRule(ruleName, rule); - } else if (schemaType === 'object' && 'properties' in schema) { - // TODO: `required` keyword (from python implementation) - const propOrder = this._propOrder; - const propPairs = Object.entries(schema.properties).sort((a, b) => { - // sort by position in prop_order (if specified) then by key - const orderA = typeof propOrder[a[0]] === 'number' ? propOrder[a[0]] : Infinity; - const orderB = typeof propOrder[b[0]] === 'number' ? propOrder[b[0]] : Infinity; - return orderA - orderB || a[0].localeCompare(b[0]); - }); - - let rule = '"{" space'; - propPairs.forEach(([propName, propSchema], i) => { - const propRuleName = this.visit(propSchema, `${name}${name ? "-" : ""}${propName}`); - if (i > 0) { - rule += ' "," space'; + } else if ((schemaType === undefined || schemaType === 'object') && + ('properties' in schema || + ('additionalProperties' in schema && schema.additionalProperties !== true))) { + const required = new Set(schema.required || []); + const properties = Object.entries(schema.properties ?? {}); + return this._addRule(ruleName, this._buildObjectRule(properties, required, name, schema.additionalProperties)); + } else if ((schemaType === undefined || schemaType === 'object') && 'allOf' in schema) { + const required = new Set(); + const properties = []; + const addComponent = (compSchema, isRequired) => { + const ref = compSchema.$ref; + if (ref !== undefined) { + compSchema = this._refs[ref]; } - rule += ` ${this._formatLiteral(propName)} space ":" space ${propRuleName}`; - }); - rule += ' "}" space'; - return this._addRule(ruleName, rule); - } else if (schemaType === 'array' && 'items' in schema) { - // TODO `prefixItems` keyword (from python implementation) - const itemRuleName = this.visit(schema.items, `${name}${name ? "-" : ""}item`); - const rule = `"[" space (${itemRuleName} ("," space ${itemRuleName})*)? "]" space`; - return this._addRule(ruleName, rule); + if ('properties' in compSchema) { + for (const [propName, propSchema] of Object.entries(compSchema.properties)) { + properties.push([propName, propSchema]); + if (isRequired) { + required.add(propName); + } + } + } + }; + + for (const t of schema.allOf) { + if ('anyOf' in t) { + for (const tt of t.anyOf) { + addComponent(tt, false); + } + } else { + addComponent(t, true); + } + } + + return this._addRule(ruleName, this._buildObjectRule(properties, required, name, /* additionalProperties= */ false)); + } else if ((schemaType === undefined || schemaType === 'array') && ('items' in schema || 'prefixItems' in schema)) { + const items = schema.items ?? schema.prefixItems; + if (Array.isArray(items)) { + return this._addRule( + ruleName, + '"[" space ' + + items.map((item, i) => this.visit(item, `${name ?? ''}${name ? '-' : ''}tuple-${i}`)).join(' "," space ') + + ' "]" space' + ); + } else { + const itemRuleName = this.visit(items, `${name ?? ''}${name ? '-' : ''}item`); + const listItemOperator = `( "," space ${itemRuleName} )`; + let successiveItems = ''; + let minItems = schema.minItems || 0; + const maxItems = schema.maxItems; + if (minItems > 0) { + successiveItems = listItemOperator.repeat(minItems - 1); + minItems--; + } + if (maxItems !== undefined && maxItems > minItems) { + successiveItems += `${listItemOperator}?`.repeat(maxItems - minItems - 1); + } else { + successiveItems += `${listItemOperator}*`; + } + const rule = minItems === 0 + ? `"[" space ( ${itemRuleName} ${successiveItems} )? "]" space` + : `"[" space ${itemRuleName} ${successiveItems} "]" space`; + return this._addRule(ruleName, rule); + } + } else if ((schemaType === undefined || schemaType === 'string') && 'pattern' in schema) { + return this._visitPattern(schema.pattern, ruleName); + } else if ((schemaType === undefined || schemaType === 'string') && /^uuid[1-5]?$/.test(schema.format || '')) { + return this._addRule( + ruleName === 'root' ? 'root' : schemaFormat, + PRIMITIVE_RULES['uuid']) + } else if ((schemaType === undefined || schemaType === 'string') && schema.format in DATE_RULES) { + for (const [t, r] of Object.entries(DATE_RULES)) { + this._addRule(t, r); + } + return schemaFormat + '-string'; + } else if ((schemaType === 'object') || (Object.keys(schema).length === 0)) { + for (const n of OBJECT_RULE_NAMES) { + this._addRule(n, PRIMITIVE_RULES[n]); + } + return this._addRule(ruleName, 'object'); } else { - if (!PRIMITIVE_RULES[schemaType]) { + if (!(schemaType in PRIMITIVE_RULES)) { throw new Error(`Unrecognized schema: ${JSON.stringify(schema)}`); } - return this._addRule( - ruleName === 'root' ? 'root' : schemaType, - PRIMITIVE_RULES[schemaType] + // TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero + return this._addRule(ruleName === 'root' ? 'root' : schemaType, PRIMITIVE_RULES[schemaType]); + } + } + + _buildObjectRule(properties, required, name, additionalProperties) { + const propOrder = this._propOrder; + // sort by position in prop_order (if specified) then by original order + const sortedProps = properties.map(([k]) => k).sort((a, b) => { + const orderA = propOrder[a] || Infinity; + const orderB = propOrder[b] || Infinity; + return orderA - orderB || properties.findIndex(([k]) => k === a) - properties.findIndex(([k]) => k === b); + }); + + const propKvRuleNames = {}; + for (const [propName, propSchema] of properties) { + const propRuleName = this.visit(propSchema, `${name ?? ''}${name ? '-' : ''}${propName}`); + propKvRuleNames[propName] = this._addRule( + `${name ?? ''}${name ? '-' : ''}${propName}-kv`, + `${this._formatLiteral(JSON.stringify(propName))} space ":" space ${propRuleName}` ); } + const requiredProps = sortedProps.filter(k => required.has(k)); + const optionalProps = sortedProps.filter(k => !required.has(k)); + + if (typeof additionalProperties === 'object' || additionalProperties === true) { + const subName = `${name ?? ''}${name ? '-' : ''}additional`; + const valueRule = this.visit(additionalProperties === true ? {} : additionalProperties, `${subName}-value`); + propKvRuleNames['*'] = this._addRule( + `${subName}-kv`, + `${this._addRule('string', PRIMITIVE_RULES['string'])} ":" space ${valueRule}`); + optionalProps.push('*'); + } + + let rule = '"{" space '; + rule += requiredProps.map(k => propKvRuleNames[k]).join(' "," space '); + + if (optionalProps.length > 0) { + rule += ' ('; + if (requiredProps.length > 0) { + rule += ' "," space ( '; + } + + const getRecursiveRefs = (ks, firstIsOptional) => { + const [k, ...rest] = ks; + const kvRuleName = propKvRuleNames[k]; + let res; + if (k === '*') { + res = this._addRule( + `${name ?? ''}${name ? '-' : ''}additional-kvs`, + `${kvRuleName} ( "," space ` + kvRuleName + ` )*` + ) + } else if (firstIsOptional) { + res = `( "," space ${kvRuleName} )?`; + } else { + res = kvRuleName; + } + if (rest.length > 0) { + res += ' ' + this._addRule( + `${name ?? ''}${name ? '-' : ''}${k}-rest`, + getRecursiveRefs(rest, true) + ); + } + return res; + }; + + rule += optionalProps.map((_, i) => getRecursiveRefs(optionalProps.slice(i), false)).join(' | '); + if (requiredProps.length > 0) { + rule += ' )'; + } + rule += ' )?'; + } + + rule += ' "}" space'; + + return rule; } formatGrammar() { let grammar = ''; - this._rules.forEach((rule, name) => { + for (const [name, rule] of Object.entries(this._rules).sort(([a], [b]) => a.localeCompare(b))) { grammar += `${name} ::= ${rule}\n`; - }); + } return grammar; } } + +// Helper function to group elements by a key function +function* groupBy(iterable, keyFn) { + let lastKey = null; + let group = []; + for (const element of iterable) { + const key = keyFn(element); + if (lastKey !== null && key !== lastKey) { + yield [lastKey, group]; + group = []; + } + group.push(element); + lastKey = key; + } + if (group.length > 0) { + yield [lastKey, group]; + } +} diff --git a/examples/server/server.cpp b/examples/server/server.cpp index d2a8e541d..338e60f28 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1,6 +1,7 @@ #include "utils.hpp" #include "common.h" +#include "json-schema-to-grammar.h" #include "llama.h" #include "grammar-parser.h" @@ -29,7 +30,7 @@ #include #include -using json = nlohmann::json; +using json = nlohmann::ordered_json; bool server_verbose = false; bool server_log_json = true; @@ -178,6 +179,7 @@ struct server_slot { llama_token sampled; struct llama_sampling_params sparams; llama_sampling_context * ctx_sampling = nullptr; + json json_schema; int32_t ga_i = 0; // group-attention state int32_t ga_n = 1; // group-attention factor @@ -845,10 +847,25 @@ struct server_context { slot.sparams.penalize_nl = json_value(data, "penalize_nl", default_sparams.penalize_nl); slot.params.n_keep = json_value(data, "n_keep", slot.params.n_keep); slot.params.seed = json_value(data, "seed", default_params.seed); - slot.sparams.grammar = json_value(data, "grammar", default_sparams.grammar); slot.sparams.n_probs = json_value(data, "n_probs", default_sparams.n_probs); slot.sparams.min_keep = json_value(data, "min_keep", default_sparams.min_keep); + // process "json_schema" and "grammar" + if (data.contains("json_schema") && data.contains("grammar")) { + send_error(task, "Either \"json_schema\" or \"grammar\" can be specified, but not both", ERROR_TYPE_INVALID_REQUEST); + return false; + } else if (data.contains("json_schema") && !data.contains("grammar")) { + try { + auto schema = json_value(data, "json_schema", json::object()); + slot.sparams.grammar = json_schema_to_grammar(schema); + } catch (const std::exception & e) { + send_error(task, std::string("\"json_schema\": ") + e.what(), ERROR_TYPE_INVALID_REQUEST); + return false; + } + } else { + slot.sparams.grammar = json_value(data, "grammar", default_sparams.grammar); + } + if (slot.params.cache_prompt && slot.ga_n != 1) { LOG_WARNING("cache_prompt is not supported with group-attention", {}); slot.params.cache_prompt = false; @@ -1235,7 +1252,7 @@ struct server_context { {"penalize_nl", slot.sparams.penalize_nl}, {"stop", slot.params.antiprompt}, {"n_predict", slot.params.n_predict}, // TODO: fix duplicate key n_predict - {"n_keep", params.n_keep}, + {"n_keep", slot.params.n_keep}, {"ignore_eos", ignore_eos}, {"stream", slot.params.stream}, {"logit_bias", slot.sparams.logit_bias}, @@ -1746,7 +1763,7 @@ struct server_context { } // process in chunks of params.n_batch - int32_t n_batch = llama_n_batch(ctx); + int32_t n_batch = llama_n_batch(ctx); int32_t n_ubatch = llama_n_ubatch(ctx); // next, batch any pending prompts without exceeding n_batch @@ -2196,7 +2213,11 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co printf(" -m FNAME, --model FNAME\n"); printf(" model path (default: %s)\n", params.model.c_str()); printf(" -mu MODEL_URL, --model-url MODEL_URL\n"); - printf(" model download url (default: %s)\n", params.model_url.c_str()); + printf(" model download url (default: unused)\n"); + printf(" -hfr REPO, --hf-repo REPO\n"); + printf(" Hugging Face model repository (default: unused)\n"); + printf(" -hff FILE, --hf-file FILE\n"); + printf(" Hugging Face model file (default: unused)\n"); printf(" -a ALIAS, --alias ALIAS\n"); printf(" set an alias for the model, will be added as `model` field in completion response\n"); printf(" --lora FNAME apply LoRA adapter (implies --no-mmap)\n"); @@ -2213,7 +2234,7 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co printf(" -to N, --timeout N server read/write timeout in seconds (default: %d)\n", sparams.read_timeout); printf(" --embeddings enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled"); printf(" -np N, --parallel N number of slots for process requests (default: %d)\n", params.n_parallel); - printf(" -cb, --cont-batching enable continuous batching (a.k.a dynamic batching) (default: disabled)\n"); + printf(" -cb, --cont-batching enable continuous batching (a.k.a dynamic batching) (default: enabled)\n"); printf(" -spf FNAME, --system-prompt-file FNAME\n"); printf(" set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n"); printf(" -ctk TYPE, --cache-type-k TYPE\n"); @@ -2325,6 +2346,18 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams, break; } params.model_url = argv[i]; + } else if (arg == "-hfr" || arg == "--hf-repo") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.hf_repo = argv[i]; + } else if (arg == "-hff" || arg == "--hf-file") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.hf_file = argv[i]; } else if (arg == "-a" || arg == "--alias") { if (++i >= argc) { invalid_param = true; diff --git a/examples/server/tests/features/parallel.feature b/examples/server/tests/features/parallel.feature index a66fed626..6cd306a2b 100644 --- a/examples/server/tests/features/parallel.feature +++ b/examples/server/tests/features/parallel.feature @@ -4,7 +4,8 @@ Feature: Parallel Background: Server startup Given a server listening on localhost:8080 - And a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models + And a model file tinyllamas/split/stories15M-00001-of-00003.gguf from HF repo ggml-org/models + And a model file test-model-00001-of-00003.gguf And 42 as server seed And 128 as batch size And 256 KV cache size diff --git a/examples/server/tests/features/security.feature b/examples/server/tests/features/security.feature index 1d6aa40ea..eb82e7aca 100644 --- a/examples/server/tests/features/security.feature +++ b/examples/server/tests/features/security.feature @@ -37,6 +37,22 @@ Feature: Security | llama.cpp | no | | hackme | raised | + Scenario Outline: OAI Compatibility (invalid response formats) + Given a system prompt test + And a user prompt test + And a response format + And a model test + And 2 max tokens to predict + And streaming is disabled + Given an OAI compatible chat completions request with raised api error + + Examples: Prompts + | response_format | + | {"type": "sound"} | + | {"type": "json_object", "schema": 123} | + | {"type": "json_object", "schema": {"type": 123}} | + | {"type": "json_object", "schema": {"type": "hiccup"}} | + Scenario Outline: CORS Options Given a user api key llama.cpp diff --git a/examples/server/tests/features/server.feature b/examples/server/tests/features/server.feature index 45a988db6..646a4e49d 100644 --- a/examples/server/tests/features/server.feature +++ b/examples/server/tests/features/server.feature @@ -4,8 +4,8 @@ Feature: llama.cpp server Background: Server startup Given a server listening on localhost:8080 - And a model url https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K.gguf - And a model file stories260K.gguf + And a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models + And a model file test-model.gguf And a model alias tinyllama-2 And 42 as server seed # KV Cache corresponds to the total amount of tokens @@ -70,6 +70,22 @@ Feature: llama.cpp server | codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 128 | (thanks\|happy\|bird\|Annabyear)+ | -1 | 64 | enabled | | + Scenario Outline: OAI Compatibility w/ response format + Given a model test + And a system prompt test + And a user prompt test + And a response format + And 10 max tokens to predict + Given an OAI compatible chat completions request with no api error + Then tokens are predicted matching + + Examples: Prompts + | response_format | n_predicted | re_content | + | {"type": "json_object", "schema": {"const": "42"}} | 5 | "42" | + | {"type": "json_object", "schema": {"items": [{"type": "integer"}]}} | 10 | \[ -300 \] | + | {"type": "json_object"} | 10 | \{ " Jacky. | + + Scenario: Tokenize / Detokenize When tokenizing: """ diff --git a/examples/server/tests/features/steps/steps.py b/examples/server/tests/features/steps/steps.py index 40c97001a..86c3339dc 100644 --- a/examples/server/tests/features/steps/steps.py +++ b/examples/server/tests/features/steps/steps.py @@ -16,7 +16,6 @@ import numpy as np import openai from behave import step from behave.api.async_step import async_run_until_complete -from huggingface_hub import hf_hub_download from prometheus_client import parser @@ -24,24 +23,29 @@ from prometheus_client import parser def step_server_config(context, server_fqdn, server_port): context.server_fqdn = server_fqdn context.server_port = int(server_port) + context.n_gpu_layer = None if 'PORT' in os.environ: context.server_port = int(os.environ['PORT']) print(f"$PORT set, overriding server port with to {context.server_port}") if 'FQDN' in os.environ: context.server_fqdn = os.environ['FQDN'] print(f"$FQDN set, overriding server fqdn with to {context.server_fqdn}") + if 'N_GPU_LAYERS' in os.environ: + context.n_gpu_layer = int(os.environ['N_GPU_LAYERS']) + print(f"$N_GPU_LAYERS set, overriding n_gpu_layer with to {context.n_gpu_layer}") context.base_url = f'http://{context.server_fqdn}:{context.server_port}' context.model_alias = None context.model_file = None + context.model_hf_repo = None + context.model_hf_file = None context.model_url = None context.n_batch = None context.n_ubatch = None context.n_ctx = None context.n_ga = None context.n_ga_w = None - context.n_gpu_layer = None context.n_predict = None context.n_prompts = 0 context.n_server_predict = None @@ -56,6 +60,7 @@ def step_server_config(context, server_fqdn, server_port): context.seed = None context.server_seed = None context.user_api_key = None + context.response_format = None context.tasks_result = [] context.concurrent_tasks = [] @@ -64,9 +69,9 @@ def step_server_config(context, server_fqdn, server_port): @step('a model file {hf_file} from HF repo {hf_repo}') def step_download_hf_model(context, hf_file, hf_repo): - context.model_file = hf_hub_download(repo_id=hf_repo, filename=hf_file) - if context.debug: - print(f"model file: {context.model_file}") + context.model_hf_repo = hf_repo + context.model_hf_file = hf_file + context.model_file = os.path.basename(hf_file) @step('a model file {model_file}') @@ -266,6 +271,11 @@ def step_max_tokens(context, max_tokens): context.n_predict = max_tokens +@step('a response format {response_format}') +def step_response_format(context, response_format): + context.response_format = json.loads(response_format) + + @step('streaming is {enable_streaming}') def step_streaming(context, enable_streaming): context.enable_streaming = enable_streaming == 'enabled' @@ -381,6 +391,9 @@ async def step_oai_chat_completions(context, api_error): enable_streaming=context.enable_streaming if hasattr(context, 'enable_streaming') else None, + response_format=context.response_format + if hasattr(context, 'response_format') else None, + seed=await completions_seed(context), user_api_key=context.user_api_key @@ -440,6 +453,8 @@ async def step_oai_chat_completions(context): if hasattr(context, 'n_predict') else None, enable_streaming=context.enable_streaming if hasattr(context, 'enable_streaming') else None, + response_format=context.response_format + if hasattr(context, 'response_format') else None, seed=await completions_seed(context), user_api_key=context.user_api_key if hasattr(context, 'user_api_key') else None) @@ -460,6 +475,8 @@ async def step_oai_chat_completions(context): if hasattr(context, 'n_predict') else None, enable_streaming=context.enable_streaming if hasattr(context, 'enable_streaming') else None, + response_format=context.response_format + if hasattr(context, 'response_format') else None, seed=context.seed if hasattr(context, 'seed') else context.server_seed @@ -742,6 +759,7 @@ async def oai_chat_completions(user_prompt, model=None, n_predict=None, enable_streaming=None, + response_format=None, seed=None, user_api_key=None, expect_api_error=None): @@ -767,6 +785,8 @@ async def oai_chat_completions(user_prompt, "stream": enable_streaming, "seed": seed } + if response_format is not None: + payload['response_format'] = response_format completion_response = { 'content': '', 'timings': { @@ -827,6 +847,7 @@ async def oai_chat_completions(user_prompt, model=model, max_tokens=n_predict, stream=enable_streaming, + response_format=payload.get('response_format'), seed=seed ) except openai.error.AuthenticationError as e: @@ -1059,6 +1080,10 @@ def start_server_background(context): server_args.extend(['--model', context.model_file]) if context.model_url: server_args.extend(['--model-url', context.model_url]) + if context.model_hf_repo: + server_args.extend(['--hf-repo', context.model_hf_repo]) + if context.model_hf_file: + server_args.extend(['--hf-file', context.model_hf_file]) if context.n_batch: server_args.extend(['--batch-size', context.n_batch]) if context.n_ubatch: diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp index 2ddb2cd21..7d9ab622b 100644 --- a/examples/server/utils.hpp +++ b/examples/server/utils.hpp @@ -12,7 +12,7 @@ #define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613" -using json = nlohmann::json; +using json = nlohmann::ordered_json; // https://community.openai.com/t/openai-chat-list-of-error-codes-and-types/357791/11 enum error_type { @@ -95,8 +95,8 @@ static inline void server_log(const char *level, const char *function, int line, const std::string str = ss.str(); printf("%.*s\n", (int)str.size(), str.data()); - fflush(stdout); } + fflush(stdout); } // @@ -352,39 +352,71 @@ static json oaicompat_completion_params_parse( // https://platform.openai.com/docs/api-reference/chat/create llama_sampling_params default_sparams; llama_params["model"] = json_value(body, "model", std::string("unknown")); - llama_params["prompt"] = format_chat(model, chat_template, body["messages"]); - llama_params["cache_prompt"] = json_value(body, "cache_prompt", false); - llama_params["temperature"] = json_value(body, "temperature", 0.0); - llama_params["top_k"] = json_value(body, "top_k", default_sparams.top_k); - llama_params["top_p"] = json_value(body, "top_p", 1.0); - llama_params["n_predict"] = json_value(body, "max_tokens", -1); - llama_params["logit_bias"] = json_value(body, "logit_bias", json::object()); llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0); + llama_params["logit_bias"] = json_value(body, "logit_bias", json::object()); + llama_params["n_predict"] = json_value(body, "max_tokens", -1); llama_params["presence_penalty"] = json_value(body, "presence_penalty", 0.0); llama_params["seed"] = json_value(body, "seed", LLAMA_DEFAULT_SEED); llama_params["stream"] = json_value(body, "stream", false); - llama_params["mirostat"] = json_value(body, "mirostat", default_sparams.mirostat); - llama_params["mirostat_tau"] = json_value(body, "mirostat_tau", default_sparams.mirostat_tau); - llama_params["mirostat_eta"] = json_value(body, "mirostat_eta", default_sparams.mirostat_eta); - llama_params["penalize_nl"] = json_value(body, "penalize_nl", default_sparams.penalize_nl); - llama_params["typical_p"] = json_value(body, "typical_p", default_sparams.typical_p); - llama_params["repeat_last_n"] = json_value(body, "repeat_last_n", default_sparams.penalty_last_n); - llama_params["ignore_eos"] = json_value(body, "ignore_eos", false); - llama_params["tfs_z"] = json_value(body, "tfs_z", default_sparams.tfs_z); + llama_params["temperature"] = json_value(body, "temperature", 0.0); + llama_params["top_p"] = json_value(body, "top_p", 1.0); - if (body.count("grammar") != 0) { - llama_params["grammar"] = json_value(body, "grammar", json::object()); - } + // Apply chat template to the list of messages + llama_params["prompt"] = format_chat(model, chat_template, body["messages"]); - // Handle 'stop' field + // Handle "stop" field if (body.contains("stop") && body["stop"].is_string()) { llama_params["stop"] = json::array({body["stop"].get()}); } else { llama_params["stop"] = json_value(body, "stop", json::array()); } + // Some chat templates don't use EOS token to stop generation + // We must add their end sequences to list of stop words + llama_params["stop"].push_back("<|im_end|>"); // chatml + llama_params["stop"].push_back(""); // gemma - // Ensure there is ChatML-specific end sequence among stop words - llama_params["stop"].push_back("<|im_end|>"); + // Handle "response_format" field + if (body.contains("response_format")) { + json response_format = json_value(body, "response_format", json::object()); + std::string response_type = json_value(response_format, "type", std::string()); + if (response_type == "json_object") { + llama_params["json_schema"] = json_value(response_format, "schema", json::object()); + } else if (!response_type.empty() && response_type != "text") { + throw std::runtime_error("response_format type must be one of \"text\" or \"json_object\", but got: " + response_type); + } + } + + // Handle "n" field + int n_choices = json_value(body, "n", 1); + if (n_choices != 1) { + throw std::runtime_error("Only one completion choice is allowed"); + } + + // Handle "logprobs" field + // TODO: The response format of this option is not yet OAI-compatible, but seems like no one really using it; We may need to fix it in the future + if (body.contains("logprobs")) { + llama_params["n_probs"] = json_value(body, "top_logprobs", 20); + } else if (body.contains("top_logprobs")) { + throw std::runtime_error("top_logprobs requires logprobs to be set to true"); + } + + // Params supported by OAI but unsupported by llama.cpp + static const std::vector unsupported_params { "tools", "tool_choice" }; + for (auto & param : unsupported_params) { + if (body.contains(param)) { + throw std::runtime_error("Unsupported param: " + param); + } + } + + // Copy remaining properties to llama_params + // This allows user to use llama.cpp-specific params like "mirostat", "tfs_z",... via OAI endpoint. + // See "launch_slot_with_task()" for a complete list of params supported by llama.cpp + for (const auto & item : body.items()) { + // Exception: if "n_predict" is present, we overwrite the value specified earlier by "max_tokens" + if (!llama_params.contains(item.key()) || item.key() == "n_predict") { + llama_params[item.key()] = item.value(); + } + } return llama_params; } diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp index e991b8846..8b31b678a 100644 --- a/examples/speculative/speculative.cpp +++ b/examples/speculative/speculative.cpp @@ -219,7 +219,8 @@ int main(int argc, char ** argv) { if (params.sparams.temp > 0) { // stochastic verification - llama_token_data_array dist_tgt = llama_sampling_probability_distribution(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft]); + llama_token_data_array dist_tgt = llama_sampling_prepare(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft], true, NULL); + llama_sample_softmax(ctx_tgt, &dist_tgt); float p_tgt = 0, p_dft = 0; // GGML_ASSERT(dist_tgt.size() == dist_dft.size()); diff --git a/examples/sycl/win-build-sycl.bat b/examples/sycl/win-build-sycl.bat index f9d43f8ed..1b0dc41ba 100644 --- a/examples/sycl/win-build-sycl.bat +++ b/examples/sycl/win-build-sycl.bat @@ -3,9 +3,13 @@ :: Copyright (C) 2024 Intel Corporation :: SPDX-License-Identifier: MIT -mkdir -p build + +IF not exist build (mkdir build) cd build +if %errorlevel% neq 0 goto ERROR + @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force +if %errorlevel% neq 0 goto ERROR :: for FP16 :: faster for long-prompt inference @@ -13,11 +17,18 @@ cd build :: for FP32 cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release - - +if %errorlevel% neq 0 goto ERROR :: build example/main only :: make main :: build all binary make -j +if %errorlevel% neq 0 goto ERROR + cd .. +exit /B 0 + +:ERROR +echo comomand error: %errorlevel% +exit /B %errorlevel% + diff --git a/examples/ts-type-to-grammar.sh b/examples/ts-type-to-grammar.sh new file mode 100755 index 000000000..a7ad00173 --- /dev/null +++ b/examples/ts-type-to-grammar.sh @@ -0,0 +1,28 @@ +#!/bin/bash +# +# ./examples/ts-type-to-grammar.sh "{a:string,b:string,c?:string}" +# python examples/json-schema-to-grammar.py https://json.schemastore.org/tsconfig.json +# +set -euo pipefail + +readonly type="$1" + +# Create a temporary directory +TMPDIR="" +trap 'rm -fR "$TMPDIR"' EXIT +TMPDIR=$(mktemp -d) + +DTS_FILE="$TMPDIR/type.d.ts" +SCHEMA_FILE="$TMPDIR/schema.json" + +echo "export type MyType = $type" > "$DTS_FILE" + +# This is a fork of typescript-json-schema, actively maintained as of March 2024: +# https://github.com/vega/ts-json-schema-generator +npx ts-json-schema-generator --unstable --no-top-ref --path "$DTS_FILE" --type MyType -e none > "$SCHEMA_FILE" + +# Alternative, not actively maintained as of March 2024: +# https://github.com/YousefED/typescript-json-schema +# npx typescript-json-schema --defaultProps --required "$DTS_FILE" MyType | tee "$SCHEMA_FILE" >&2 + +./examples/json-schema-to-grammar.py "$SCHEMA_FILE" diff --git a/flake.lock b/flake.lock index 80de76dbf..f21694b0d 100644 --- a/flake.lock +++ b/flake.lock @@ -20,11 +20,11 @@ }, "nixpkgs": { "locked": { - "lastModified": 1710451336, - "narHash": "sha256-pP86Pcfu3BrAvRO7R64x7hs+GaQrjFes+mEPowCfkxY=", + "lastModified": 1711163522, + "narHash": "sha256-YN/Ciidm+A0fmJPWlHBGvVkcarYWSC+s3NTPk/P+q3c=", "owner": "NixOS", "repo": "nixpkgs", - "rev": "d691274a972b3165335d261cc4671335f5c67de9", + "rev": "44d0940ea560dee511026a53f0e2e2cde489b4d4", "type": "github" }, "original": { diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 139025588..4f50c9f9f 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -2,265 +2,54 @@ #include "ggml.h" #include "ggml-backend-impl.h" -#if defined(GGML_USE_HIPBLAS) -#define GGML_COMMON_DECL_HIP -#define GGML_COMMON_IMPL_HIP -#else -#define GGML_COMMON_DECL_CUDA -#define GGML_COMMON_IMPL_CUDA -#endif -#include "ggml-common.h" +#include "ggml-cuda/common.cuh" +#include "ggml-cuda/acc.cuh" +#include "ggml-cuda/alibi.cuh" +#include "ggml-cuda/arange.cuh" +#include "ggml-cuda/argsort.cuh" +#include "ggml-cuda/binbcast.cuh" +#include "ggml-cuda/clamp.cuh" +#include "ggml-cuda/concat.cuh" +#include "ggml-cuda/convert.cuh" +#include "ggml-cuda/cpy.cuh" +#include "ggml-cuda/diagmask.cuh" +#include "ggml-cuda/dmmv.cuh" +#include "ggml-cuda/getrows.cuh" +#include "ggml-cuda/im2col.cuh" +#include "ggml-cuda/mmq.cuh" +#include "ggml-cuda/mmvq.cuh" +#include "ggml-cuda/norm.cuh" +#include "ggml-cuda/pad.cuh" +#include "ggml-cuda/pool2d.cuh" +#include "ggml-cuda/quantize.cuh" +#include "ggml-cuda/rope.cuh" +#include "ggml-cuda/scale.cuh" +#include "ggml-cuda/softmax.cuh" +#include "ggml-cuda/sumrows.cuh" +#include "ggml-cuda/tsembd.cuh" +#include "ggml-cuda/unary.cuh" +#include "ggml-cuda/upscale.cuh" #include -#include +#include #include #include #include #include #include #include +#include +#include +#include #include #include #include #include -#include -#include - -// stringize macro for converting __CUDA_ARCH_LIST__ (list of integers) to string -#define STRINGIZE_IMPL(...) #__VA_ARGS__ -#define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__) - -#if defined(GGML_USE_HIPBLAS) -#include -#include -#include -#ifdef __HIP_PLATFORM_AMD__ -// for rocblas_initialize() -#include "rocblas/rocblas.h" -#endif // __HIP_PLATFORM_AMD__ -#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F -#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F -#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F -#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT -#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT -#define CUBLAS_OP_N HIPBLAS_OP_N -#define CUBLAS_OP_T HIPBLAS_OP_T -#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS -#define CUBLAS_TF32_TENSOR_OP_MATH 0 -#define CUDA_R_16F HIPBLAS_R_16F -#define CUDA_R_32F HIPBLAS_R_32F -#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width) -#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6 -#define cublasCreate hipblasCreate -#define cublasGemmEx hipblasGemmEx -#define cublasGemmBatchedEx hipblasGemmBatchedEx -#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx -#define cublasHandle_t hipblasHandle_t -#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS -#define cublasSetStream hipblasSetStream -#define cublasSgemm hipblasSgemm -#define cublasStatus_t hipblasStatus_t -#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6 -#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer -#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess -#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess -#define cudaDeviceProp hipDeviceProp_t -#define cudaDeviceSynchronize hipDeviceSynchronize -#define cudaError_t hipError_t -#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled -#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled -#define cudaEventCreateWithFlags hipEventCreateWithFlags -#define cudaEventDisableTiming hipEventDisableTiming -#define cudaEventRecord hipEventRecord -#define cudaEventSynchronize hipEventSynchronize -#define cudaEvent_t hipEvent_t -#define cudaEventDestroy hipEventDestroy -#define cudaFree hipFree -#define cudaFreeHost hipHostFree -#define cudaGetDevice hipGetDevice -#define cudaGetDeviceCount hipGetDeviceCount -#define cudaGetDeviceProperties hipGetDeviceProperties -#define cudaGetErrorString hipGetErrorString -#define cudaGetLastError hipGetLastError -#define cudaHostRegister hipHostRegister -#define cudaHostRegisterPortable hipHostRegisterPortable -#define cudaHostRegisterReadOnly hipHostRegisterReadOnly -#define cudaHostUnregister hipHostUnregister -#define cudaLaunchHostFunc hipLaunchHostFunc -#ifdef GGML_HIP_UMA -#define cudaMalloc hipMallocManaged -#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size) -#else -#define cudaMalloc hipMalloc -#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault) -#endif -#define cudaMemcpy hipMemcpy -#define cudaMemcpyAsync hipMemcpyAsync -#define cudaMemcpyPeerAsync hipMemcpyPeerAsync -#define cudaMemcpy2DAsync hipMemcpy2DAsync -#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice -#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost -#define cudaMemcpyHostToDevice hipMemcpyHostToDevice -#define cudaMemcpyKind hipMemcpyKind -#define cudaMemset hipMemset -#define cudaMemsetAsync hipMemsetAsync -#define cudaMemGetInfo hipMemGetInfo -#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize -#define cudaSetDevice hipSetDevice -#define cudaStreamCreateWithFlags hipStreamCreateWithFlags -#define cudaStreamFireAndForget hipStreamFireAndForget -#define cudaStreamNonBlocking hipStreamNonBlocking -#define cudaStreamPerThread hipStreamPerThread -#define cudaStreamSynchronize hipStreamSynchronize -#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags) -#define cudaStream_t hipStream_t -#define cudaSuccess hipSuccess -#define __trap abort -#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS -#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED -#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED -#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE -#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH -#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR -#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED -#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR -#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED -#else -#include -#include -#include -#include - -#if CUDART_VERSION < 11020 -#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED -#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH -#define CUBLAS_COMPUTE_16F CUDA_R_16F -#define CUBLAS_COMPUTE_32F CUDA_R_32F -#define cublasComputeType_t cudaDataType_t -#endif // CUDART_VERSION < 11020 - -#endif // defined(GGML_USE_HIPBLAS) - -#define CUDART_HMAX 11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed) - -#define CC_PASCAL 600 -#define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products -#define CC_VOLTA 700 -#define CC_OFFSET_AMD 1000000 -#define CC_RDNA1 (CC_OFFSET_AMD + 1010) -#define CC_RDNA2 (CC_OFFSET_AMD + 1030) -#define CC_RDNA3 (CC_OFFSET_AMD + 1100) - -#define GGML_CUDA_MAX_NODES 8192 - -// define this if you want to always fallback to MMQ kernels and not use cuBLAS for matrix multiplication -// on modern hardware, using cuBLAS is recommended as it utilizes F16 tensor cores which are very performant -// for large computational tasks. the drawback is that this requires some extra amount of VRAM: -// - 7B quantum model: +100-200 MB -// - 13B quantum model: +200-400 MB -// -//#define GGML_CUDA_FORCE_MMQ - -// TODO: improve this to be correct for more hardware -// for example, currently fails for GeForce GTX 1660 which is TURING arch (> VOLTA) but does not have tensor cores -#if !defined(GGML_CUDA_FORCE_MMQ) -#define CUDA_USE_TENSOR_CORES -#endif - -#define MMVQ_MAX_BATCH_SIZE 8 // max batch size to use MMVQ kernels -#define MMQ_MAX_BATCH_SIZE 32 // max batch size to use MMQ kernels when tensor cores are available - -#if defined(GGML_USE_HIPBLAS) -#define __CUDA_ARCH__ 1300 - -#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \ - defined(__gfx1150__) || defined(__gfx1151__) -#define RDNA3 -#endif - -#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \ - defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__) -#define RDNA2 -#endif - -#ifndef __has_builtin - #define __has_builtin(x) 0 -#endif - -typedef int8_t int8x4_t __attribute__((ext_vector_type(4))); -typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4))); -static __device__ __forceinline__ int __vsubss4(const int a, const int b) { - const int8x4_t va = reinterpret_cast(a); - const int8x4_t vb = reinterpret_cast(b); -#if __has_builtin(__builtin_elementwise_sub_sat) - const int8x4_t c = __builtin_elementwise_sub_sat(va, vb); - return reinterpret_cast(c); -#else - int8x4_t c; - int16_t tmp; -#pragma unroll - for (int i = 0; i < 4; i++) { - tmp = va[i] - vb[i]; - if(tmp > std::numeric_limits::max()) tmp = std::numeric_limits::max(); - if(tmp < std::numeric_limits::min()) tmp = std::numeric_limits::min(); - c[i] = tmp; - } - return reinterpret_cast(c); -#endif // __has_builtin(__builtin_elementwise_sub_sat) -} - -static __device__ __forceinline__ int __vsub4(const int a, const int b) { - return __vsubss4(a, b); -} - -static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigned int b) { - const uint8x4_t& va = reinterpret_cast(a); - const uint8x4_t& vb = reinterpret_cast(b); - unsigned int c; - uint8x4_t& vc = reinterpret_cast(c); -#pragma unroll - for (int i = 0; i < 4; ++i) { - vc[i] = va[i] == vb[i] ? 0xff : 0x00; - } - return c; -} - -static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) { -#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) - c = __builtin_amdgcn_sdot4(a, b, c, false); -#elif defined(RDNA3) - c = __builtin_amdgcn_sudot4( true, a, true, b, c, false); -#elif defined(__gfx1010__) || defined(__gfx900__) - int tmp1; - int tmp2; - asm("\n \ - v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \ - v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \ - v_add3_u32 %0, %1, %2, %0 \n \ - v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \ - v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \ - v_add3_u32 %0, %1, %2, %0 \n \ - " - : "+v"(c), "=&v"(tmp1), "=&v"(tmp2) - : "v"(a), "v"(b) - ); -#else - const int8x4_t va = reinterpret_cast(a); - const int8x4_t vb = reinterpret_cast(b); - c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3]; -#endif - return c; -} -#endif // defined(GGML_USE_HIPBLAS) - -#if defined(_MSC_VER) -#pragma warning(disable: 4244 4267) // possible loss of data -#endif static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size"); [[noreturn]] -static void ggml_cuda_error(const char * stmt, const char * func, const char * file, const int line, const char * msg) { +void ggml_cuda_error(const char * stmt, const char * func, const char * file, int line, const char * msg) { int id = -1; // in case cudaGetDevice fails cudaGetDevice(&id); @@ -271,178 +60,9 @@ static void ggml_cuda_error(const char * stmt, const char * func, const char * f GGML_ASSERT(!"CUDA error"); } -#define CUDA_CHECK_GEN(err, success, error_fn) \ - do { \ - auto err_ = (err); \ - if (err_ != (success)) { \ - ggml_cuda_error(#err, __func__, __FILE__, __LINE__, error_fn(err_)); \ - } \ - } while (0) - -#define CUDA_CHECK(err) CUDA_CHECK_GEN(err, cudaSuccess, cudaGetErrorString) - -#if CUDART_VERSION >= 12000 - static const char * cublas_get_error_str(const cublasStatus_t err) { - return cublasGetStatusString(err); - } -#else - static const char * cublas_get_error_str(const cublasStatus_t err) { - switch (err) { - case CUBLAS_STATUS_SUCCESS: return "CUBLAS_STATUS_SUCCESS"; - case CUBLAS_STATUS_NOT_INITIALIZED: return "CUBLAS_STATUS_NOT_INITIALIZED"; - case CUBLAS_STATUS_ALLOC_FAILED: return "CUBLAS_STATUS_ALLOC_FAILED"; - case CUBLAS_STATUS_INVALID_VALUE: return "CUBLAS_STATUS_INVALID_VALUE"; - case CUBLAS_STATUS_ARCH_MISMATCH: return "CUBLAS_STATUS_ARCH_MISMATCH"; - case CUBLAS_STATUS_MAPPING_ERROR: return "CUBLAS_STATUS_MAPPING_ERROR"; - case CUBLAS_STATUS_EXECUTION_FAILED: return "CUBLAS_STATUS_EXECUTION_FAILED"; - case CUBLAS_STATUS_INTERNAL_ERROR: return "CUBLAS_STATUS_INTERNAL_ERROR"; - case CUBLAS_STATUS_NOT_SUPPORTED: return "CUBLAS_STATUS_NOT_SUPPORTED"; - default: return "unknown error"; - } - } -#endif // CUDART_VERSION >= 12000 - -#define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str) - -#if !defined(GGML_USE_HIPBLAS) -static const char * cu_get_error_str(CUresult err) { - const char * err_str; - cuGetErrorString(err, &err_str); - return err_str; -} -#define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str) -#endif - -#if CUDART_VERSION >= 11100 -#define GGML_CUDA_ASSUME(x) __builtin_assume(x) -#else -#define GGML_CUDA_ASSUME(x) -#endif // CUDART_VERSION >= 11100 - -#ifdef GGML_CUDA_F16 -typedef half dfloat; // dequantize float -typedef half2 dfloat2; -#else -typedef float dfloat; // dequantize float -typedef float2 dfloat2; -#endif //GGML_CUDA_F16 - -static __device__ __forceinline__ int get_int_from_int8(const int8_t * x8, const int & i32) { - const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment - - int x32 = 0; - x32 |= x16[0] << 0; - x32 |= x16[1] << 16; - - return x32; -} - -static __device__ __forceinline__ int get_int_from_uint8(const uint8_t * x8, const int & i32) { - const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment - - int x32 = 0; - x32 |= x16[0] << 0; - x32 |= x16[1] << 16; - - return x32; -} - -static __device__ __forceinline__ int get_int_from_int8_aligned(const int8_t * x8, const int & i32) { - return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment -} - -static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t * x8, const int & i32) { - return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment -} - -template -using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int k, cudaStream_t stream); -typedef to_t_cuda_t to_fp32_cuda_t; -typedef to_t_cuda_t to_fp16_cuda_t; - -typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v); -typedef void (*dot_kernel_k_t)(const void * __restrict__ vx, const int ib, const int iqs, const float * __restrict__ y, float & v); -typedef void (*cpy_kernel_t)(const char * cx, char * cdst); -typedef void (*ggml_cuda_func_t)(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst); -typedef void (*ggml_cuda_op_mul_mat_t)( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, - const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, - const int64_t src1_padded_row_size, cudaStream_t stream); -typedef void (*ggml_cuda_op_flatten_t)( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream); - -typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs); -typedef void (*allocate_tiles_cuda_t)(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc); -typedef void (*load_tiles_cuda_t)( - const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, - int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row); -typedef float (*vec_dot_q_mul_mat_cuda_t)( - const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, - const int * __restrict__ y_qs, const half2 * __restrict__ y_ms, const int & i, const int & j, const int & k); - -#define WARP_SIZE 32 -#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses - -#define CUDA_GELU_BLOCK_SIZE 256 -#define CUDA_SILU_BLOCK_SIZE 256 -#define CUDA_TANH_BLOCK_SIZE 256 -#define CUDA_RELU_BLOCK_SIZE 256 -#define CUDA_HARDSIGMOID_BLOCK_SIZE 256 -#define CUDA_HARDSWISH_BLOCK_SIZE 256 -#define CUDA_SQR_BLOCK_SIZE 256 -#define CUDA_CPY_BLOCK_SIZE 32 -#define CUDA_SCALE_BLOCK_SIZE 256 -#define CUDA_CLAMP_BLOCK_SIZE 256 -#define CUDA_ROPE_BLOCK_SIZE 256 -#define CUDA_SOFT_MAX_BLOCK_SIZE 1024 -#define CUDA_ALIBI_BLOCK_SIZE 32 -#define CUDA_DIAG_MASK_INF_BLOCK_SIZE 32 -#define CUDA_QUANTIZE_BLOCK_SIZE 256 -#define CUDA_DEQUANTIZE_BLOCK_SIZE 256 -#define CUDA_GET_ROWS_BLOCK_SIZE 256 -#define CUDA_UPSCALE_BLOCK_SIZE 256 -#define CUDA_CONCAT_BLOCK_SIZE 256 -#define CUDA_PAD_BLOCK_SIZE 256 -#define CUDA_ARANGE_BLOCK_SIZE 256 -#define CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE 256 -#define CUDA_ACC_BLOCK_SIZE 256 -#define CUDA_IM2COL_BLOCK_SIZE 256 -#define CUDA_POOL2D_BLOCK_SIZE 256 - -#define CUDA_Q8_0_NE_ALIGN 2048 - -// dmmv = dequantize_mul_mat_vec -#ifndef GGML_CUDA_DMMV_X -#define GGML_CUDA_DMMV_X 32 -#endif -#ifndef GGML_CUDA_MMV_Y -#define GGML_CUDA_MMV_Y 1 -#endif - -#ifndef K_QUANTS_PER_ITERATION -#define K_QUANTS_PER_ITERATION 2 -#else -static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2"); -#endif - -#ifndef GGML_CUDA_PEER_MAX_BATCH_SIZE -#define GGML_CUDA_PEER_MAX_BATCH_SIZE 128 -#endif // GGML_CUDA_PEER_MAX_BATCH_SIZE - -#define MUL_MAT_SRC1_COL_STRIDE 128 - -#define MAX_STREAMS 8 -static cudaStream_t g_cudaStreams[GGML_CUDA_MAX_DEVICES][MAX_STREAMS] = { { nullptr } }; - -struct ggml_tensor_extra_gpu { - void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors - cudaEvent_t events[GGML_CUDA_MAX_DEVICES][MAX_STREAMS]; // events for synchronizing multiple GPUs -}; - // this is faster on Windows // probably because the Windows CUDA libraries forget to make this check before invoking the drivers -static void ggml_cuda_set_device(const int device) { +void ggml_cuda_set_device(int device) { int current_device; CUDA_CHECK(cudaGetDevice(¤t_device)); @@ -453,4966 +73,993 @@ static void ggml_cuda_set_device(const int device) { CUDA_CHECK(cudaSetDevice(device)); } -static int g_device_count = -1; -static int g_main_device = 0; -static std::array g_default_tensor_split = {}; +int ggml_cuda_get_device() { + int id; + CUDA_CHECK(cudaGetDevice(&id)); + return id; +} -struct cuda_device_capabilities { - int cc; // compute capability - size_t smpb; // max. shared memory per block - bool vmm; // virtual memory support - size_t vmm_granularity; // granularity of virtual memory -}; +static ggml_cuda_device_info ggml_cuda_init() { +#ifdef __HIP_PLATFORM_AMD__ + // Workaround for a rocBLAS bug when using multiple graphics cards: + // https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346 + rocblas_initialize(); + CUDA_CHECK(cudaDeviceSynchronize()); +#endif -static cuda_device_capabilities g_device_caps[GGML_CUDA_MAX_DEVICES] = { {0, 0, false, 0} }; + ggml_cuda_device_info info = {}; -static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr}; + cudaError_t err = cudaGetDeviceCount(&info.device_count); + if (err != cudaSuccess) { + fprintf(stderr, "%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err)); + return info; + } -[[noreturn]] -static __device__ void no_device_code( - const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) { + GGML_ASSERT(info.device_count <= GGML_CUDA_MAX_DEVICES); + + int64_t total_vram = 0; +#if defined(GGML_CUDA_FORCE_MMQ) + fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ: yes\n", __func__); +#else + fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ: no\n", __func__); +#endif +#if defined(CUDA_USE_TENSOR_CORES) + fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: yes\n", __func__); +#else + fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: no\n", __func__); +#endif + fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count); + for (int id = 0; id < info.device_count; ++id) { + int device_vmm = 0; + +#if !defined(GGML_USE_HIPBLAS) + CUdevice device; + CU_CHECK(cuDeviceGet(&device, id)); + CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device)); + + if (device_vmm) { + CUmemAllocationProp alloc_prop = {}; + alloc_prop.type = CU_MEM_ALLOCATION_TYPE_PINNED; + alloc_prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE; + alloc_prop.location.id = id; + CU_CHECK(cuMemGetAllocationGranularity(&info.devices[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED)); + } +#endif // !defined(GGML_USE_HIPBLAS) + info.devices[id].vmm = !!device_vmm; + + cudaDeviceProp prop; + CUDA_CHECK(cudaGetDeviceProperties(&prop, id)); + fprintf(stderr, " Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no"); + + info.default_tensor_split[id] = total_vram; + total_vram += prop.totalGlobalMem; #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) - printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n", - file_name, line, function_name, arch); - (void) arch_list; + info.devices[id].cc = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD; #else - printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n", - file_name, line, function_name, arch, arch_list); + info.devices[id].cc = 100*prop.major + 10*prop.minor; #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) - __trap(); - - (void) no_device_code; // suppress unused function warning -} - -#ifdef __CUDA_ARCH__ -#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__)) -#else -#define NO_DEVICE_CODE GGML_ASSERT(false && "NO_DEVICE_CODE not valid in host code.") -#endif // __CUDA_ARCH__ - -static __device__ __forceinline__ float warp_reduce_sum(float x) { -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - x += __shfl_xor_sync(0xffffffff, x, mask, 32); - } - return x; -} - -static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) { -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32); - a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32); - } - return a; -} - -#ifdef GGML_CUDA_F16 -static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) { -#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32)); - } - return a; -#else - (void) a; - NO_DEVICE_CODE; -#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL -} -#endif // GGML_CUDA_F16 - -static __device__ __forceinline__ float warp_reduce_max(float x) { -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32)); - } - return x; -} - -//static __device__ __forceinline__ half2 warp_reduce_max(half2 x) { -//#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX -//#pragma unroll -// for (int mask = 16; mask > 0; mask >>= 1) { -// x = __hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32)); -// } -// return x; -//#else -// (void) x; -// NO_DEVICE_CODE; -//#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX -//} - -static __device__ __forceinline__ float op_repeat(const float a, const float b) { - return b; - GGML_UNUSED(a); -} - -static __device__ __forceinline__ float op_add(const float a, const float b) { - return a + b; -} - -static __device__ __forceinline__ float op_mul(const float a, const float b) { - return a * b; -} - -static __device__ __forceinline__ float op_div(const float a, const float b) { - return a / b; -} - -template -static __global__ void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst, - int ne0, int ne1, int ne2, int ne3, - int ne10, int ne11, int ne12, int ne13, - /*int s0, */ int s1, int s2, int s3, - /*int s10,*/ int s11, int s12, int s13) { - const int i0s = blockDim.x*blockIdx.x + threadIdx.x; - const int i1 = (blockDim.y*blockIdx.y + threadIdx.y); - const int i2 = (blockDim.z*blockIdx.z + threadIdx.z) / ne3; - const int i3 = (blockDim.z*blockIdx.z + threadIdx.z) % ne3; - - if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) { - return; + info.devices[id].smpb = prop.sharedMemPerBlock; } - const int i11 = i1 % ne11; - const int i12 = i2 % ne12; - const int i13 = i3 % ne13; - - const size_t i_src0 = i3*s3 + i2*s2 + i1*s1; - const size_t i_src1 = i13*s13 + i12*s12 + i11*s11; - const size_t i_dst = i_src0; - - const src0_t * src0_row = src0 + i_src0; - const src1_t * src1_row = src1 + i_src1; - dst_t * dst_row = dst + i_dst; - - for (int i0 = i0s; i0 < ne0; i0 += blockDim.x*gridDim.x) { - const int i10 = i0 % ne10; - dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]); + for (int id = 0; id < info.device_count; ++id) { + info.default_tensor_split[id] /= total_vram; } + + // configure logging to stdout + // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr)); + + return info; } -template -static __global__ void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst, - int ne0, int ne1, int ne2, int ne3, - int ne10, int ne11, int ne12, int ne13, - /*int s0, */ int s1, int s2, int s3, - /*int s10,*/ int s11, int s12, int s13) { - - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - const int i3 = i/(ne2*ne1*ne0); - const int i2 = (i/(ne1*ne0)) % ne2; - const int i1 = (i/ne0) % ne1; - const int i0 = i % ne0; - - if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) { - return; - } - - const int i11 = i1 % ne11; - const int i12 = i2 % ne12; - const int i13 = i3 % ne13; - - const size_t i_src0 = i3*s3 + i2*s2 + i1*s1; - const size_t i_src1 = i13*s13 + i12*s12 + i11*s11; - const size_t i_dst = i_src0; - - const src0_t * src0_row = src0 + i_src0; - const src1_t * src1_row = src1 + i_src1; - dst_t * dst_row = dst + i_dst; - - const int i10 = i0 % ne10; - dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]); +const ggml_cuda_device_info & ggml_cuda_info() { + static ggml_cuda_device_info info = ggml_cuda_init(); + return info; } -static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne, - const int ne10, const int ne11, const int ne12, - const int nb1, const int nb2, int offset) { - const int i = blockDim.x * blockIdx.x + threadIdx.x; - if (i >= ne) { - return; - } - int src1_idx = i - offset; - int oz = src1_idx / nb2; - int oy = (src1_idx - (oz * nb2)) / nb1; - int ox = src1_idx % nb1; - if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) { - dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11]; - } else { - dst[i] = x[i]; - } -} +// #define DEBUG_CUDA_MALLOC -static __global__ void gelu_f32(const float * x, float * dst, const int k) { - const float GELU_COEF_A = 0.044715f; - const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f; - const int i = blockDim.x*blockIdx.x + threadIdx.x; +// buffer pool for cuda (legacy) +struct ggml_cuda_pool_leg : public ggml_cuda_pool { + static const int MAX_BUFFERS = 256; - if (i >= k) { - return; + int device; + struct ggml_cuda_buffer { + void * ptr = nullptr; + size_t size = 0; + }; + + ggml_cuda_buffer buffer_pool[MAX_BUFFERS] = {}; + size_t pool_size = 0; + + explicit ggml_cuda_pool_leg(int device) : + device(device) { } - float xi = x[i]; - dst[i] = 0.5f*xi*(1.0f + tanhf(SQRT_2_OVER_PI*xi*(1.0f + GELU_COEF_A*xi*xi))); -} - -static __global__ void silu_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = x[i] / (1.0f + expf(-x[i])); -} - -static __global__ void gelu_quick_f32(const float * x, float * dst, int k) { - const float GELU_QUICK_COEF = -1.702f; - const int i = blockDim.x*blockIdx.x + threadIdx.x; - if (i >= k) { - return; - } - dst[i] = x[i] * (1.0f / (1.0f + expf(GELU_QUICK_COEF * x[i]))); -} - -static __global__ void tanh_f32(const float * x, float * dst, int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - if (i >= k) { - return; - } - dst[i] = tanhf(x[i]); -} - -static __global__ void relu_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = fmaxf(x[i], 0); -} - -static __global__ void hardsigmoid_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); -} - -static __global__ void hardswish_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); -} - -static __global__ void leaky_relu_f32(const float * x, float * dst, const int k, const float negative_slope) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - if (i >= k) { - return; - } - dst[i] = fmaxf(x[i], 0) + fminf(x[i], 0.0f) * negative_slope; -} - -static __global__ void sqr_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = x[i] * x[i]; -} - -template -static __global__ void norm_f32(const float * x, float * dst, const int ncols, const float eps) { - const int row = blockIdx.x*blockDim.y + threadIdx.y; - const int tid = threadIdx.x; - - float2 mean_var = make_float2(0.f, 0.f); - - for (int col = tid; col < ncols; col += block_size) { - const float xi = x[row*ncols + col]; - mean_var.x += xi; - mean_var.y += xi * xi; - } - - // sum up partial sums - mean_var = warp_reduce_sum(mean_var); - if (block_size > WARP_SIZE) { - __shared__ float2 s_sum[32]; - int warp_id = threadIdx.x / WARP_SIZE; - int lane_id = threadIdx.x % WARP_SIZE; - if (lane_id == 0) { - s_sum[warp_id] = mean_var; + ~ggml_cuda_pool_leg() { + ggml_cuda_set_device(device); + for (int i = 0; i < MAX_BUFFERS; ++i) { + ggml_cuda_buffer & b = buffer_pool[i]; + if (b.ptr != nullptr) { + CUDA_CHECK(cudaFree(b.ptr)); + pool_size -= b.size; + } } - __syncthreads(); - mean_var = s_sum[lane_id]; - mean_var = warp_reduce_sum(mean_var); + GGML_ASSERT(pool_size == 0); } - const float mean = mean_var.x / ncols; - const float var = mean_var.y / ncols - mean * mean; - const float inv_std = rsqrtf(var + eps); - - for (int col = tid; col < ncols; col += block_size) { - dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_std; - } -} - -static __global__ void concat_f32(const float * x,const float * y, float * dst, const int ne0, const int ne02) { - int nidx = threadIdx.x + blockIdx.x * blockDim.x; - if (nidx >= ne0) { - return; - } - // operation - int offset_dst = - nidx + - blockIdx.y * ne0 + - blockIdx.z * ne0 * gridDim.y; - if (blockIdx.z < ne02) { // src0 - int offset_src = - nidx + - blockIdx.y * ne0 + - blockIdx.z * ne0 * gridDim.y; - dst[offset_dst] = x[offset_src]; - } else { - int offset_src = - nidx + - blockIdx.y * ne0 + - (blockIdx.z - ne02) * ne0 * gridDim.y; - dst[offset_dst] = y[offset_src]; - } -} - -static __global__ void upscale_f32(const float * x, float * dst, const int ne00, const int ne00xne01, const int scale_factor) { - // blockIdx.z: idx of ne02*ne03 - // blockIdx.y: idx of ne01*scale_factor， aka ne1 - // blockIDx.x: idx of ne00*scale_factor / BLOCK_SIZE - // ne00xne01: ne00 * ne01 - int ne0 = ne00 * scale_factor; - int nidx = threadIdx.x + blockIdx.x * blockDim.x; - if (nidx >= ne0) { - return; - } - // operation - int i00 = nidx / scale_factor; - int i01 = blockIdx.y / scale_factor; - int offset_src = - i00 + - i01 * ne00 + - blockIdx.z * ne00xne01; - int offset_dst = - nidx + - blockIdx.y * ne0 + - blockIdx.z * ne0 * gridDim.y; - dst[offset_dst] = x[offset_src]; -} - -static __global__ void pad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02, const int ne03) { - // blockIdx.z: idx of ne2*ne3, aka ne02*ne03 - // blockIdx.y: idx of ne1 - // blockIDx.x: idx of ne0 / BLOCK_SIZE - int nidx = threadIdx.x + blockIdx.x * blockDim.x; - if (nidx >= ne0) { - return; - } - - // operation - int offset_dst = - nidx + - blockIdx.y * ne0 + - blockIdx.z * ne0 * gridDim.y; - if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) { - int offset_src = - nidx + - blockIdx.y * ne00 + - blockIdx.z * ne00 * ne01; - dst[offset_dst] = x[offset_src]; - } else { - dst[offset_dst] = 0.0f; - } -} - -static __global__ void arange_f32(float * dst, const int ne0, const float start, const float step) { - // blockIDx.x: idx of ne0 / BLOCK_SIZE - int nidx = threadIdx.x + blockIdx.x * blockDim.x; - if (nidx >= ne0) { - return; - } - dst[nidx] = start + step * nidx; -} - -static __global__ void timestep_embedding_f32(const float * timesteps, float * dst, const int nb1, const int dim, const int max_period) { - // blockIDx.y: idx of timesteps->ne[0] - // blockIDx.x: idx of ((dim + 1) / 2) / BLOCK_SIZE - int i = blockIdx.y; - int j = threadIdx.x + blockIdx.x * blockDim.x; - float * embed_data = (float *)((char *)dst + i*nb1); - - if (dim % 2 != 0 && j == ((dim + 1) / 2)) { - embed_data[dim] = 0.f; - } - - int half = dim / 2; - if (j >= half) { - return; - } - - float timestep = timesteps[i]; - float freq = (float)expf(-logf(max_period) * j / half); - float arg = timestep * freq; - embed_data[j] = cosf(arg); - embed_data[j + half] = sinf(arg); -} - -template -static __global__ void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps) { - // blockIdx.x: num_groups idx - // threadIdx.x: block_size idx - int start = blockIdx.x * group_size; - int end = start + group_size; - - start += threadIdx.x; - - if (end >= ne_elements) { - end = ne_elements; - } - - float tmp = 0.0f; // partial sum for thread in warp - - for (int j = start; j < end; j += block_size) { - tmp += x[j]; - } - - tmp = warp_reduce_sum(tmp); - if (block_size > WARP_SIZE) { - __shared__ float s_sum[32]; - int warp_id = threadIdx.x / WARP_SIZE; - int lane_id = threadIdx.x % WARP_SIZE; - if (lane_id == 0) { - s_sum[warp_id] = tmp; + void * alloc(size_t size, size_t * actual_size) override { +#ifdef DEBUG_CUDA_MALLOC + int nnz = 0; + size_t max_size = 0; +#endif + size_t best_diff = 1ull << 36; + int ibest = -1; + for (int i = 0; i < MAX_BUFFERS; ++i) { + ggml_cuda_buffer& b = buffer_pool[i]; + if (b.ptr != nullptr) { +#ifdef DEBUG_CUDA_MALLOC + ++nnz; + if (b.size > max_size) max_size = b.size; +#endif + if (b.size >= size) { + size_t diff = b.size - size; + if (diff < best_diff) { + best_diff = diff; + ibest = i; + if (!best_diff) { + void * ptr = b.ptr; + *actual_size = b.size; + b.ptr = nullptr; + b.size = 0; + return ptr; + } + } + } + } } - __syncthreads(); - tmp = s_sum[lane_id]; - tmp = warp_reduce_sum(tmp); - } - - float mean = tmp / group_size; - tmp = 0.0f; - - for (int j = start; j < end; j += block_size) { - float xi = x[j] - mean; - dst[j] = xi; - tmp += xi * xi; - } - - tmp = warp_reduce_sum(tmp); - if (block_size > WARP_SIZE) { - __shared__ float s_sum[32]; - int warp_id = threadIdx.x / WARP_SIZE; - int lane_id = threadIdx.x % WARP_SIZE; - if (lane_id == 0) { - s_sum[warp_id] = tmp; + if (ibest >= 0) { + ggml_cuda_buffer& b = buffer_pool[ibest]; + void * ptr = b.ptr; + *actual_size = b.size; + b.ptr = nullptr; + b.size = 0; + return ptr; } - __syncthreads(); - tmp = s_sum[lane_id]; - tmp = warp_reduce_sum(tmp); + void * ptr; + size_t look_ahead_size = (size_t) (1.05 * size); + look_ahead_size = 256 * ((look_ahead_size + 255)/256); + ggml_cuda_set_device(device); + CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size)); + *actual_size = look_ahead_size; + pool_size += look_ahead_size; +#ifdef DEBUG_CUDA_MALLOC + fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz, + (uint32_t)(max_size/1024/1024), (uint32_t)(pool_size/1024/1024), (uint32_t)(size/1024/1024)); +#endif + return ptr; } - float variance = tmp / group_size; - float scale = rsqrtf(variance + eps); - for (int j = start; j < end; j += block_size) { - dst[j] *= scale; - } -} - -template -static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps) { - const int row = blockIdx.x*blockDim.y + threadIdx.y; - const int tid = threadIdx.x; - - float tmp = 0.0f; // partial sum for thread in warp - - for (int col = tid; col < ncols; col += block_size) { - const float xi = x[row*ncols + col]; - tmp += xi * xi; - } - - // sum up partial sums - tmp = warp_reduce_sum(tmp); - if (block_size > WARP_SIZE) { - __shared__ float s_sum[32]; - int warp_id = threadIdx.x / WARP_SIZE; - int lane_id = threadIdx.x % WARP_SIZE; - if (lane_id == 0) { - s_sum[warp_id] = tmp; + void free(void * ptr, size_t size) override { + for (int i = 0; i < MAX_BUFFERS; ++i) { + ggml_cuda_buffer& b = buffer_pool[i]; + if (b.ptr == nullptr) { + b.ptr = ptr; + b.size = size; + return; + } } - __syncthreads(); - tmp = s_sum[lane_id]; - tmp = warp_reduce_sum(tmp); + fprintf(stderr, "WARNING: cuda buffer pool full, increase MAX_CUDA_BUFFERS\n"); + ggml_cuda_set_device(device); + CUDA_CHECK(cudaFree(ptr)); + pool_size -= size; + } +}; + +// pool with virtual memory +#if !defined(GGML_USE_HIPBLAS) +struct ggml_cuda_pool_vmm : public ggml_cuda_pool { + static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB + + int device; + CUdeviceptr pool_addr = 0; + size_t pool_used = 0; + size_t pool_size = 0; + size_t granularity; + + explicit ggml_cuda_pool_vmm(int device) : + device(device), + granularity(ggml_cuda_info().devices[device].vmm_granularity) { } - const float mean = tmp / ncols; - const float scale = rsqrtf(mean + eps); - - for (int col = tid; col < ncols; col += block_size) { - dst[row*ncols + col] = scale * x[row*ncols + col]; + ~ggml_cuda_pool_vmm() { + if (pool_addr != 0) { + CU_CHECK(cuMemUnmap(pool_addr, pool_size)); + CU_CHECK(cuMemAddressFree(pool_addr, CUDA_POOL_VMM_MAX_SIZE)); + } } + + void * alloc(size_t size, size_t * actual_size) override { + // round up the allocation size to the alignment to ensure that all allocations are aligned for all data types + const size_t alignment = 128; + size = alignment * ((size + alignment - 1) / alignment); + + size_t avail = pool_size - pool_used; + + if (size > avail) { + // round up to the next multiple of the granularity + size_t reserve_size = size - avail; + reserve_size = granularity * ((reserve_size + granularity - 1) / granularity); + + GGML_ASSERT(pool_size + reserve_size <= CUDA_POOL_VMM_MAX_SIZE); + + // allocate more physical memory + CUmemAllocationProp prop = {}; + prop.type = CU_MEM_ALLOCATION_TYPE_PINNED; + prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE; + prop.location.id = device; + CUmemGenericAllocationHandle handle; + CU_CHECK(cuMemCreate(&handle, reserve_size, &prop, 0)); + + // reserve virtual address space (if not already reserved) + if (pool_addr == 0) { + CU_CHECK(cuMemAddressReserve(&pool_addr, CUDA_POOL_VMM_MAX_SIZE, 0, 0, 0)); + } + + // map at the end of the pool + CU_CHECK(cuMemMap(pool_addr + pool_size, reserve_size, 0, handle, 0)); + + // the memory allocation handle is no longer needed after mapping + CU_CHECK(cuMemRelease(handle)); + + // set access + CUmemAccessDesc access = {}; + access.location.type = CU_MEM_LOCATION_TYPE_DEVICE; + access.location.id = device; + access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE; + CU_CHECK(cuMemSetAccess(pool_addr + pool_size, reserve_size, &access, 1)); + + // add to the pool + pool_size += reserve_size; + + //printf("cuda pool[%d]: size increased to %llu MB (reserved %llu MB)\n", + // device, (unsigned long long) (pool_size/1024/1024), + // (unsigned long long) (reserve_size/1024/1024)); + } + + GGML_ASSERT(pool_addr != 0); + + void * ptr = (void *) (pool_addr + pool_used); + *actual_size = size; + pool_used += size; + +#ifdef DEBUG_CUDA_MALLOC + printf("cuda pool[%d]: allocated %llu bytes at %llx\n", device, (unsigned long long) size, ptr); +#endif + + return ptr; + } + + void free(void * ptr, size_t size) override { +#ifdef DEBUG_CUDA_MALLOC + printf("cuda pool[%d]: freed %llu bytes at %llx\n", device, (unsigned long long) size, ptr); +#endif + + pool_used -= size; + + // all deallocations must be in reverse order of the allocations + GGML_ASSERT(ptr == (void *) (pool_addr + pool_used)); + } +}; +#endif // !defined(GGML_USE_HIPBLAS) + +std::unique_ptr ggml_backend_cuda_context::new_pool_for_device(int device) { +#if !defined(GGML_USE_HIPBLAS) + if (ggml_cuda_info().devices[device].vmm) { + return std::unique_ptr(new ggml_cuda_pool_vmm(device)); + } +#endif + return std::unique_ptr(new ggml_cuda_pool_leg(device)); } -static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, dfloat2 & v){ - const block_q4_0 * x = (const block_q4_0 *) vx; +// cuda buffer - const dfloat d = x[ib].d; +struct ggml_backend_cuda_buffer_context { + int device; + void * dev_ptr = nullptr; + std::string name; - const int vui = x[ib].qs[iqs]; + ggml_backend_cuda_buffer_context(int device, void * dev_ptr) : + device(device), dev_ptr(dev_ptr), + name(GGML_CUDA_NAME + std::to_string(device)) { + } - v.x = vui & 0xF; - v.y = vui >> 4; + ~ggml_backend_cuda_buffer_context() { + CUDA_CHECK(cudaFree(dev_ptr)); + } +}; -#ifdef GGML_CUDA_F16 - v = __hsub2(v, {8.0f, 8.0f}); - v = __hmul2(v, {d, d}); -#else - v.x = (v.x - 8.0f) * d; - v.y = (v.y - 8.0f) * d; -#endif // GGML_CUDA_F16 +GGML_CALL static const char * ggml_backend_cuda_buffer_get_name(ggml_backend_buffer_t buffer) { + ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; + return ctx->name.c_str(); } -static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, dfloat2 & v){ - const block_q4_1 * x = (const block_q4_1 *) vx; - - const dfloat d = __low2half(x[ib].dm); - const dfloat m = __high2half(x[ib].dm); - - const int vui = x[ib].qs[iqs]; - - v.x = vui & 0xF; - v.y = vui >> 4; - -#ifdef GGML_CUDA_F16 - v = __hmul2(v, {d, d}); - v = __hadd2(v, {m, m}); -#else - v.x = (v.x * d) + m; - v.y = (v.y * d) + m; -#endif // GGML_CUDA_F16 +GGML_CALL static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) { + return buffer->iface.get_name == ggml_backend_cuda_buffer_get_name; } -static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, dfloat2 & v){ - const block_q5_0 * x = (const block_q5_0 *) vx; - - const dfloat d = x[ib].d; - - uint32_t qh; - memcpy(&qh, x[ib].qh, sizeof(qh)); - - const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; - const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10; - - v.x = ((x[ib].qs[iqs] & 0xf) | xh_0); - v.y = ((x[ib].qs[iqs] >> 4) | xh_1); - -#ifdef GGML_CUDA_F16 - v = __hsub2(v, {16.0f, 16.0f}); - v = __hmul2(v, {d, d}); -#else - v.x = (v.x - 16.0f) * d; - v.y = (v.y - 16.0f) * d; -#endif // GGML_CUDA_F16 +GGML_CALL static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) { + ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; + delete ctx; } -static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, dfloat2 & v){ - const block_q5_1 * x = (const block_q5_1 *) vx; - - const dfloat d = __low2half(x[ib].dm); - const dfloat m = __high2half(x[ib].dm); - - uint32_t qh; - memcpy(&qh, x[ib].qh, sizeof(qh)); - - const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; - const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10; - - v.x = ((x[ib].qs[iqs] & 0xf) | xh_0); - v.y = ((x[ib].qs[iqs] >> 4) | xh_1); - -#ifdef GGML_CUDA_F16 - v = __hmul2(v, {d, d}); - v = __hadd2(v, {m, m}); -#else - v.x = (v.x * d) + m; - v.y = (v.y * d) + m; -#endif // GGML_CUDA_F16 +GGML_CALL static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) { + ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; + return ctx->dev_ptr; } -static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, dfloat2 & v){ - const block_q8_0 * x = (const block_q8_0 *) vx; +GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { + ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; - const dfloat d = x[ib].d; - - v.x = x[ib].qs[iqs + 0]; - v.y = x[ib].qs[iqs + 1]; - -#ifdef GGML_CUDA_F16 - v = __hmul2(v, {d, d}); -#else - v.x *= d; - v.y *= d; -#endif // GGML_CUDA_F16 -} - -template -static __global__ void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) { - - const int i = blockIdx.x; - - // assume 32 threads - const int tid = threadIdx.x; - const int il = tid/8; - const int ir = tid%8; - const int ib = 8*i + ir; - if (ib >= nb32) { + if (tensor->view_src != NULL && tensor->view_offs == 0) { + assert(tensor->view_src->buffer->buft == buffer->buft); + tensor->backend = tensor->view_src->backend; + tensor->extra = tensor->view_src->extra; return; } - dst_t * y = yy + 256*i + 32*ir + 4*il; + if (ggml_is_quantized(tensor->type)) { + // initialize padding to 0 to avoid possible NaN values + size_t original_size = ggml_nbytes(tensor); + size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor); - const block_q4_0 * x = (const block_q4_0 *)vx + ib; - const float d = __half2float(x->d); - const float dm = -8*d; - - const uint8_t * q = x->qs + 4*il; - - for (int l = 0; l < 4; ++l) { - y[l+ 0] = d * (q[l] & 0xF) + dm; - y[l+16] = d * (q[l] >> 4) + dm; + if (padded_size > original_size && tensor->view_src == nullptr) { + ggml_cuda_set_device(ctx->device); + CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size)); + } } } -template -static __global__ void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) { +GGML_CALL static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { + ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; - const int i = blockIdx.x; - - // assume 32 threads - const int tid = threadIdx.x; - const int il = tid/8; - const int ir = tid%8; - const int ib = 8*i + ir; - if (ib >= nb32) { - return; - } - - dst_t * y = yy + 256*i + 32*ir + 4*il; - - const block_q4_1 * x = (const block_q4_1 *)vx + ib; - const float2 d = __half22float2(x->dm); - - const uint8_t * q = x->qs + 4*il; - - for (int l = 0; l < 4; ++l) { - y[l+ 0] = d.x * (q[l] & 0xF) + d.y; - y[l+16] = d.x * (q[l] >> 4) + d.y; - } + ggml_cuda_set_device(ctx->device); + CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cudaStreamPerThread)); + CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); } -//================================== k-quants +GGML_CALL static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { + ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; -template -static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { + ggml_cuda_set_device(ctx->device); + CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cudaStreamPerThread)); + CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); +} - const int i = blockIdx.x; - const block_q2_K * x = (const block_q2_K *) vx; - - const int tid = threadIdx.x; -#if QK_K == 256 - const int n = tid/32; - const int l = tid - 32*n; - const int is = 8*n + l/16; - - const uint8_t q = x[i].qs[32*n + l]; - dst_t * y = yy + i*QK_K + 128*n; - - float dall = __low2half(x[i].dm); - float dmin = __high2half(x[i].dm); - y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4); - y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4); - y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4); - y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4); +GGML_CALL static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) { + if (ggml_backend_buffer_is_cuda(src->buffer)) { + ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src->buffer->context; + ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)dst->buffer->context; + if (src_ctx->device == dst_ctx->device) { + CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(src), cudaMemcpyDeviceToDevice, cudaStreamPerThread)); + } else { +#ifdef GGML_CUDA_NO_PEER_COPY + return false; #else - const int is = tid/16; // 0 or 1 - const int il = tid%16; // 0...15 - const uint8_t q = x[i].qs[il] >> (2*is); - dst_t * y = yy + i*QK_K + 16*is + il; - float dall = __low2half(x[i].dm); - float dmin = __high2half(x[i].dm); - y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4); - y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4); + CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, dst_ctx->device, src->data, src_ctx->device, ggml_nbytes(src), cudaStreamPerThread)); #endif - -} - -template -static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { - - const int i = blockIdx.x; - const block_q3_K * x = (const block_q3_K *) vx; - -#if QK_K == 256 - const int r = threadIdx.x/4; - const int tid = r/2; - const int is0 = r%2; - const int l0 = 16*is0 + 4*(threadIdx.x%4); - const int n = tid / 4; - const int j = tid - 4*n; - - uint8_t m = 1 << (4*n + j); - int is = 8*n + 2*j + is0; - int shift = 2*j; - - int8_t us = is < 4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) : - is < 8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) : - is < 12 ? (x[i].scales[is-8] >> 4) | (((x[i].scales[is+0] >> 4) & 3) << 4) : - (x[i].scales[is-8] >> 4) | (((x[i].scales[is-4] >> 6) & 3) << 4); - float d_all = x[i].d; - float dl = d_all * (us - 32); - - dst_t * y = yy + i*QK_K + 128*n + 32*j; - const uint8_t * q = x[i].qs + 32*n; - const uint8_t * hm = x[i].hmask; - - for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)); -#else - const int tid = threadIdx.x; - const int is = tid/16; // 0 or 1 - const int il = tid%16; // 0...15 - const int im = il/8; // 0...1 - const int in = il%8; // 0...7 - - dst_t * y = yy + i*QK_K + 16*is + il; - - const uint8_t q = x[i].qs[il] >> (2*is); - const uint8_t h = x[i].hmask[in] >> (2*is + im); - const float d = (float)x[i].d; - - if (is == 0) { - y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4)); - y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4)); - } else { - y[ 0] = d * ((x[i].scales[0] >> 4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4)); - y[32] = d * ((x[i].scales[1] >> 4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4)); + } + CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); + return true; } -#endif + return false; + GGML_UNUSED(buffer); } -#if QK_K == 256 -static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) { - if (j < 4) { - d = q[j] & 63; m = q[j + 4] & 63; - } else { - d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4); - m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4); +GGML_CALL static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { + ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; + + ggml_cuda_set_device(ctx->device); + CUDA_CHECK(cudaDeviceSynchronize()); + CUDA_CHECK(cudaMemset(ctx->dev_ptr, value, buffer->size)); + CUDA_CHECK(cudaDeviceSynchronize()); +} + +static ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = { + /* .get_name = */ ggml_backend_cuda_buffer_get_name, + /* .free_buffer = */ ggml_backend_cuda_buffer_free_buffer, + /* .get_base = */ ggml_backend_cuda_buffer_get_base, + /* .init_tensor = */ ggml_backend_cuda_buffer_init_tensor, + /* .set_tensor = */ ggml_backend_cuda_buffer_set_tensor, + /* .get_tensor = */ ggml_backend_cuda_buffer_get_tensor, + /* .cpy_tensor = */ ggml_backend_cuda_buffer_cpy_tensor, + /* .clear = */ ggml_backend_cuda_buffer_clear, + /* .reset = */ NULL, +}; + +// cuda buffer type +struct ggml_backend_cuda_buffer_type_context { + int device; + std::string name; +}; + +GGML_CALL static const char * ggml_backend_cuda_buffer_type_name(ggml_backend_buffer_type_t buft) { + ggml_backend_cuda_buffer_type_context * ctx = (ggml_backend_cuda_buffer_type_context *)buft->context; + + return ctx->name.c_str(); +} + +GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { + ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context; + + ggml_cuda_set_device(buft_ctx->device); + + size = std::max(size, (size_t)1); // cudaMalloc returns null for size 0 + + void * dev_ptr; + cudaError_t err = cudaMalloc(&dev_ptr, size); + if (err != cudaSuccess) { + fprintf(stderr, "%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size/1024.0/1024.0, buft_ctx->device, cudaGetErrorString(err)); + return nullptr; } + + ggml_backend_cuda_buffer_context * ctx = new ggml_backend_cuda_buffer_context(buft_ctx->device, dev_ptr); + + return ggml_backend_buffer_init(buft, ggml_backend_cuda_buffer_interface, ctx, size); } -#endif -template -static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const block_q4_K * x = (const block_q4_K *) vx; +GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { + return 128; - const int i = blockIdx.x; + GGML_UNUSED(buft); +} -#if QK_K == 256 - // assume 32 threads - const int tid = threadIdx.x; - const int il = tid/8; - const int ir = tid%8; - const int is = 2*il; - const int n = 4; +GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { + size_t size = ggml_nbytes(tensor); + int64_t ne0 = tensor->ne[0]; - dst_t * y = yy + i*QK_K + 64*il + n*ir; - - const float dall = __low2half(x[i].dm); - const float dmin = __high2half(x[i].dm); - - const uint8_t * q = x[i].qs + 32*il + n*ir; - - uint8_t sc, m; - get_scale_min_k4(is + 0, x[i].scales, sc, m); - const float d1 = dall * sc; const float m1 = dmin * m; - get_scale_min_k4(is + 1, x[i].scales, sc, m); - const float d2 = dall * sc; const float m2 = dmin * m; - for (int l = 0; l < n; ++l) { - y[l + 0] = d1 * (q[l] & 0xF) - m1; - y[l +32] = d2 * (q[l] >> 4) - m2; + if (ggml_is_quantized(tensor->type)) { + if (ne0 % MATRIX_ROW_PADDING != 0) { + size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); + } } -#else - const int tid = threadIdx.x; - const uint8_t * q = x[i].qs; - dst_t * y = yy + i*QK_K; - const float d = (float)x[i].dm[0]; - const float m = (float)x[i].dm[1]; - y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4); - y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >> 4) - m * (x[i].scales[1] >> 4); -#endif + + return size; + + GGML_UNUSED(buft); } -template -static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const block_q5_K * x = (const block_q5_K *) vx; +GGML_CALL static bool ggml_backend_cuda_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) { + if (!ggml_backend_is_cuda(backend)) { + return false; + } - const int i = blockIdx.x; + ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context; + ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; -#if QK_K == 256 - // assume 64 threads - this is very slightly better than the one below - const int tid = threadIdx.x; - const int il = tid/16; // il is in 0...3 - const int ir = tid%16; // ir is in 0...15 - const int is = 2*il; // is is in 0...6 - - dst_t * y = yy + i*QK_K + 64*il + 2*ir; - - const float dall = __low2half(x[i].dm); - const float dmin = __high2half(x[i].dm); - - const uint8_t * ql = x[i].qs + 32*il + 2*ir; - const uint8_t * qh = x[i].qh + 2*ir; - - uint8_t sc, m; - get_scale_min_k4(is + 0, x[i].scales, sc, m); - const float d1 = dall * sc; const float m1 = dmin * m; - get_scale_min_k4(is + 1, x[i].scales, sc, m); - const float d2 = dall * sc; const float m2 = dmin * m; - - uint8_t hm = 1 << (2*il); - y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1; - y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1; - hm <<= 1; - y[32] = d2 * ((ql[ 0] >> 4) + (qh[ 0] & hm ? 16 : 0)) - m2; - y[33] = d2 * ((ql[ 1] >> 4) + (qh[ 1] & hm ? 16 : 0)) - m2; -#else - const int tid = threadIdx.x; - const uint8_t q = x[i].qs[tid]; - const int im = tid/8; // 0...3 - const int in = tid%8; // 0...7 - const int is = tid/16; // 0 or 1 - const uint8_t h = x[i].qh[in] >> im; - const float d = x[i].d; - dst_t * y = yy + i*QK_K + tid; - y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16)); - y[32] = d * x[i].scales[is+2] * ((q >> 4) - ((h >> 4) & 1 ? 0 : 16)); -#endif + return buft_ctx->device == cuda_ctx->device; } -template -static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const block_q6_K * x = (const block_q6_K *) vx; +static ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = { + /* .get_name = */ ggml_backend_cuda_buffer_type_name, + /* .alloc_buffer = */ ggml_backend_cuda_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_cuda_buffer_type_get_alignment, + /* .get_max_size = */ NULL, // defaults to SIZE_MAX + /* .get_alloc_size = */ ggml_backend_cuda_buffer_type_get_alloc_size, + /* .supports_backend = */ ggml_backend_cuda_buffer_type_supports_backend, + /* .is_host = */ NULL, +}; - const int i = blockIdx.x; -#if QK_K == 256 +GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) { + static std::mutex mutex; + std::lock_guard lock(mutex); - // assume 64 threads - this is very slightly better than the one below - const int tid = threadIdx.x; - const int ip = tid/32; // ip is 0 or 1 - const int il = tid - 32*ip; // 0...32 - const int is = 8*ip + il/16; + if (device >= ggml_backend_cuda_get_device_count()) { + return nullptr; + } - dst_t * y = yy + i*QK_K + 128*ip + il; + static ggml_backend_buffer_type ggml_backend_cuda_buffer_types[GGML_CUDA_MAX_DEVICES]; - const float d = x[i].d; + static bool ggml_backend_cuda_buffer_type_initialized = false; - const uint8_t * ql = x[i].ql + 64*ip + il; - const uint8_t qh = x[i].qh[32*ip + il]; - const int8_t * sc = x[i].scales + is; + if (!ggml_backend_cuda_buffer_type_initialized) { + for (int i = 0; i < GGML_CUDA_MAX_DEVICES; i++) { + ggml_backend_cuda_buffer_types[i] = { + /* .iface = */ ggml_backend_cuda_buffer_type_interface, + /* .context = */ new ggml_backend_cuda_buffer_type_context{i, GGML_CUDA_NAME + std::to_string(i)}, + }; + } + ggml_backend_cuda_buffer_type_initialized = true; + } - y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32); - y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32); - y[64] = d * sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32); - y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32); -#else - - // assume 32 threads - const int tid = threadIdx.x; - const int ip = tid/16; // 0 or 1 - const int il = tid - 16*ip; // 0...15 - - dst_t * y = yy + i*QK_K + 16*ip + il; - - const float d = x[i].d; - - const uint8_t ql = x[i].ql[16*ip + il]; - const uint8_t qh = x[i].qh[il] >> (2*ip); - const int8_t * sc = x[i].scales; - - y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32); - y[32] = d * sc[ip+2] * ((int8_t)((ql >> 4) | (((qh >> 4) & 3) << 4)) - 32); -#endif + return &ggml_backend_cuda_buffer_types[device]; } -inline bool ggml_cuda_supports_mmq(enum ggml_type type) { - switch (type) { +// cuda split buffer + +static int64_t get_row_rounding(ggml_type type, const std::array & tensor_split) { + int64_t min_compute_capability = INT_MAX; + int64_t max_compute_capability = INT_MIN; + for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) { + if (tensor_split[id] < (id + 1 < ggml_backend_cuda_get_device_count() ? tensor_split[id + 1] : 1.0f)) { + if (min_compute_capability > ggml_cuda_info().devices[id].cc) { + min_compute_capability = ggml_cuda_info().devices[id].cc; + } + if (max_compute_capability < ggml_cuda_info().devices[id].cc) { + max_compute_capability = ggml_cuda_info().devices[id].cc; + } + } + } + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + switch(type) { case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: case GGML_TYPE_Q5_1: case GGML_TYPE_Q8_0: + return max_compute_capability >= CC_RDNA2 ? 128 : 64; + case GGML_TYPE_F16: + case GGML_TYPE_F32: + return 1; + case GGML_TYPE_Q2_K: + return max_compute_capability >= CC_RDNA2 ? 128 : 32; + case GGML_TYPE_Q3_K: + return min_compute_capability < CC_RDNA2 ? 128 : 64; + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + return max_compute_capability >= CC_RDNA2 ? 128 : 64; + default: + GGML_ASSERT(false); + } +#else + switch(type) { + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + return max_compute_capability >= CC_VOLTA ? 128 : 64; + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + return 64; + case GGML_TYPE_F16: + case GGML_TYPE_F32: + return 1; case GGML_TYPE_Q2_K: case GGML_TYPE_Q3_K: case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + return max_compute_capability >= CC_VOLTA ? 128 : 64; case GGML_TYPE_Q6_K: - return true; + return 64; default: - return false; + GGML_ASSERT(false); } +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) } -template -static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) { +static void get_row_split(int64_t * row_low, int64_t * row_high, const ggml_tensor * tensor, const std::array & tensor_split, int id) { + const int64_t nrows = ggml_nrows(tensor); + const int64_t rounding = get_row_rounding(tensor->type, tensor_split); - const int i = blockIdx.x; - const block_iq2_xxs * x = (const block_iq2_xxs *) vx; + *row_low = id == 0 ? 0 : nrows*tensor_split[id]; + *row_low -= *row_low % rounding; - const int tid = threadIdx.x; -#if QK_K == 256 - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 8*il; - const uint16_t * q2 = x[i].qs + 4*ib; - const uint8_t * aux8 = (const uint8_t *)q2; - const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[il]); - const uint32_t aux32 = q2[2] | (q2[3] << 16); - const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f; - const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127]; - for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); -#else - assert(false); -#endif - -} - -template -static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) { - - const int i = blockIdx.x; - const block_iq2_xs * x = (const block_iq2_xs *) vx; - - const int tid = threadIdx.x; -#if QK_K == 256 - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 8*il; - const uint16_t * q2 = x[i].qs + 4*ib; - const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511)); - const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; - const uint8_t signs = ksigns_iq2xs[q2[il] >> 9]; - for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); -#else - assert(false); -#endif - -} - -template -static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_t * __restrict__ yy) { - - const int i = blockIdx.x; - const block_iq2_s * x = (const block_iq2_s *) vx; - - const int tid = threadIdx.x; -#if QK_K == 256 - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 8*il; - const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300))); - const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; - const uint8_t signs = x[i].qs[QK_K/8+4*ib+il]; - for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); -#else - assert(false); -#endif - -} - -template -static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) { - - const int i = blockIdx.x; - const block_iq3_xxs * x = (const block_iq3_xxs *) vx; - - const int tid = threadIdx.x; -#if QK_K == 256 - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 8*il; - const uint8_t * q3 = x[i].qs + 8*ib; - const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib; - const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]); - const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]); - const uint32_t aux32 = gas[0] | (gas[1] << 16); - const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f; - const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127]; - for (int j = 0; j < 4; ++j) { - y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); - y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); - } -#else - assert(false); -#endif - -} - -template -static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) { - - const int i = blockIdx.x; - const block_iq3_s * x = (const block_iq3_s *) vx; - - const int tid = threadIdx.x; -#if QK_K == 256 - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 8*il; - const uint8_t * qs = x[i].qs + 8*ib; - const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256))); - const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf)); - const uint8_t signs = x[i].signs[4*ib + il]; - for (int j = 0; j < 4; ++j) { - y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); - y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); - } -#else - assert(false); -#endif - -} - -template -static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) { - - const int i = blockIdx.x; - const block_iq1_s * x = (const block_iq1_s *) vx; - - const int tid = threadIdx.x; -#if QK_K == 256 - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 8*il; - const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA; - const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 7) + 1); - uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32; - grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)]; - grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f; - grid32[0] &= 0x0f0f0f0f; - for (int j = 0; j < 8; ++j) { - y[j] = d * (q[j] + delta); - } -#else - assert(false); -#endif - -} - -static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113}; - -template -static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) { - - const int i = blockIdx.x; - const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL); - - const int tid = threadIdx.x; - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 4*il; - const uint8_t * q4 = x[ib].qs + 4*il; - const float d = (float)x[ib].d; - for (int j = 0; j < 4; ++j) { - y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf]; - y[j+16] = d * kvalues_iq4nl[q4[j] >> 4]; - } - -} - -#if QK_K != 64 -template -static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const int i = blockIdx.x; - const block_iq4_xs * x = (const block_iq4_xs *)vx; - - const int tid = threadIdx.x; - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 4*il; - const uint8_t * q4 = x[i].qs + 16*ib + 4*il; - const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32); - for (int j = 0; j < 4; ++j) { - y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf]; - y[j+16] = d * kvalues_iq4nl[q4[j] >> 4]; - } -} -#endif - -static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) { - - static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION"); - - const int row = blockIdx.x*blockDim.y + threadIdx.y; - if (row > nrows) return; - - const int num_blocks_per_row = ncols / QK_K; - const int ib0 = row*num_blocks_per_row; - - const block_q2_K * x = (const block_q2_K *)vx + ib0; - - float tmp = 0; // partial sum for thread in warp - -#if QK_K == 256 - const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...15 - const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0,1 - - const int step = 16/K_QUANTS_PER_ITERATION; - - const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... - const int in = tid - step*im; // 0...15 or 0...7 - - const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 or 0...14 in steps of 2 - const int q_offset = 32*im + l0; - const int s_offset = 8*im; - const int y_offset = 128*im + l0; - - uint32_t aux[4]; - const uint8_t * d = (const uint8_t *)aux; - const uint8_t * m = (const uint8_t *)(aux + 2); - - for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { - - const float * y = yy + i * QK_K + y_offset; - const uint8_t * q = x[i].qs + q_offset; - - const float dall = __low2half(x[i].dm); - const float dmin = __high2half(x[i].dm); - - const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset); - aux[0] = a[0] & 0x0f0f0f0f; - aux[1] = a[1] & 0x0f0f0f0f; - aux[2] = (a[0] >> 4) & 0x0f0f0f0f; - aux[3] = (a[1] >> 4) & 0x0f0f0f0f; - - float sum1 = 0, sum2 = 0; - for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { - sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3) - + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3) - + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3) - + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3) - + y[l+16] * d[1] * ((q[l+16] >> 0) & 3) - + y[l+48] * d[3] * ((q[l+16] >> 2) & 3) - + y[l+80] * d[5] * ((q[l+16] >> 4) & 3) - +y[l+112] * d[7] * ((q[l+16] >> 6) & 3); - sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6] - + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7]; - - } - tmp += dall * sum1 - dmin * sum2; - - } -#else - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3 - const int offset = tid * K_QUANTS_PER_ITERATION; - - uint32_t uaux[2]; - const uint8_t * d = (const uint8_t *)uaux; - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - - const float * y = yy + i * QK_K + offset; - const uint8_t * q = x[i].qs + offset; - const uint32_t * s = (const uint32_t *)x[i].scales; - - uaux[0] = s[0] & 0x0f0f0f0f; - uaux[1] = (s[0] >> 4) & 0x0f0f0f0f; - - const float2 dall = __half22float2(x[i].dm); - - float sum1 = 0, sum2 = 0; - for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { - const uint8_t ql = q[l]; - sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3) - + y[l+16] * d[1] * ((ql >> 2) & 3) - + y[l+32] * d[2] * ((ql >> 4) & 3) - + y[l+48] * d[3] * ((ql >> 6) & 3); - sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7]; - } - tmp += dall.x * sum1 - dall.y * sum2; - } -#endif - - // sum up partial sums and write back result - tmp = warp_reduce_sum(tmp); - - if (threadIdx.x == 0) { - dst[row] = tmp; - } -} - -static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) { - - const int row = blockIdx.x*blockDim.y + threadIdx.y; - if (row > nrows) return; - - const int num_blocks_per_row = ncols / QK_K; - const int ib0 = row*num_blocks_per_row; - - const block_q3_K * x = (const block_q3_K *)vx + ib0; - - float tmp = 0; // partial sum for thread in warp - -#if QK_K == 256 - - const uint16_t kmask1 = 0x0303; - const uint16_t kmask2 = 0x0f0f; - - const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16 - const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0,1 - - const int n = K_QUANTS_PER_ITERATION; // iterations in the inner loop - const int step = 16/K_QUANTS_PER_ITERATION; - const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... - const int in = tid - step*im; // 0....15 or 0...7 - - const uint8_t m = 1 << (4*im); - - const int l0 = n*in; // 0...15 or 0...14 in steps of 2 - const int q_offset = 32*im + l0; - const int y_offset = 128*im + l0; - - uint16_t utmp[4]; - const int8_t * s = (const int8_t *)utmp; - - const uint16_t s_shift = 4*im; - - for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { - - const float * y = yy + i * QK_K + y_offset; - const uint8_t * q = x[i].qs + q_offset; - const uint8_t * h = x[i].hmask + l0; - - const uint16_t * a = (const uint16_t *)x[i].scales; - utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4); - utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4); - utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4); - utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4); - - const float d = x[i].d; - - float sum = 0; - for (int l = 0; l < n; ++l) { - sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4)) - + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4)) - + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4)) - + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4)); - sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4)) - + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4)) - + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4)) - + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4)); - } - tmp += d * sum; - - } -#else - - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3 - const int offset = tid * K_QUANTS_PER_ITERATION; // 0...15 or 0...14 - const int in = offset/8; // 0 or 1 - const int im = offset%8; // 0...7 - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - - const float * y = yy + i * QK_K + offset; - const uint8_t * q = x[i].qs + offset; - const uint8_t * s = x[i].scales; - - const float dall = (float)x[i].d; - - float sum = 0; - for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { - const uint8_t hl = x[i].hmask[im+l] >> in; - const uint8_t ql = q[l]; - sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4)) - + y[l+16] * dall * ((s[0] >> 4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4)) - + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4)) - + y[l+48] * dall * ((s[1] >> 4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4)); - } - tmp += sum; - } -#endif - - // sum up partial sums and write back result - tmp = warp_reduce_sum(tmp); - - if (threadIdx.x == 0) { - dst[row] = tmp; - } -} - -static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) { - - const int row = blockIdx.x*blockDim.y + threadIdx.y; - if (row > nrows) return; - const int num_blocks_per_row = ncols / QK_K; - const int ib0 = row*num_blocks_per_row; - - const block_q4_K * x = (const block_q4_K *)vx + ib0; - -#if QK_K == 256 - const uint16_t kmask1 = 0x3f3f; - const uint16_t kmask2 = 0x0f0f; - const uint16_t kmask3 = 0xc0c0; - - const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16 - const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0,1 - - const int step = 8/K_QUANTS_PER_ITERATION; // 8 or 4 - - const int il = tid/step; // 0...3 - const int ir = tid - step*il; // 0...7 or 0...3 - const int n = 2 * K_QUANTS_PER_ITERATION; // 2 or 4 - - const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 - const int in = il%2; - - const int l0 = n*(2*ir + in); - const int q_offset = 32*im + l0; - const int y_offset = 64*im + l0; - - uint16_t aux[4]; - const uint8_t * sc = (const uint8_t *)aux; - -#if K_QUANTS_PER_ITERATION == 2 - uint32_t q32[4]; - const uint8_t * q4 = (const uint8_t *)q32; -#else - uint16_t q16[4]; - const uint8_t * q4 = (const uint8_t *)q16; -#endif - - float tmp = 0; // partial sum for thread in warp - - for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { - - const float * y1 = yy + i*QK_K + y_offset; - const float * y2 = y1 + 128; - - const float dall = __low2half(x[i].dm); - const float dmin = __high2half(x[i].dm); - - const uint16_t * a = (const uint16_t *)x[i].scales; - aux[0] = a[im+0] & kmask1; - aux[1] = a[im+2] & kmask1; - aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2); - aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2); - -#if K_QUANTS_PER_ITERATION == 2 - const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset); - const uint32_t * q2 = q1 + 16; - - q32[0] = q1[0] & 0x0f0f0f0f; - q32[1] = q1[0] & 0xf0f0f0f0; - q32[2] = q2[0] & 0x0f0f0f0f; - q32[3] = q2[0] & 0xf0f0f0f0; - - float4 s = {0.f, 0.f, 0.f, 0.f}; - float smin = 0; - for (int l = 0; l < 4; ++l) { - s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+ 4]; - s.z += y2[l] * q4[l+8]; s.w += y2[l+32] * q4[l+12]; - smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7]; - } - tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin; -#else - const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset); - const uint16_t * q2 = q1 + 32; - - q16[0] = q1[0] & 0x0f0f; - q16[1] = q1[0] & 0xf0f0; - q16[2] = q2[0] & 0x0f0f; - q16[3] = q2[0] & 0xf0f0; - - float4 s = {0.f, 0.f, 0.f, 0.f}; - float smin = 0; - for (int l = 0; l < 2; ++l) { - s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2]; - s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6]; - smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7]; - } - tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin; -#endif - - } -#else - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); - - const int step = tid * K_QUANTS_PER_ITERATION; - - uint16_t aux16[2]; - const uint8_t * s = (const uint8_t *)aux16; - - float tmp = 0; - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - const uint8_t * q = x[i].qs + step; - const float * y = yy + i*QK_K + step; - const uint16_t * a = (const uint16_t *)x[i].scales; - aux16[0] = a[0] & 0x0f0f; - aux16[1] = (a[0] >> 4) & 0x0f0f; - const float d = (float)x[i].dm[0]; - const float m = (float)x[i].dm[1]; - float sum = 0.f; - for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { - sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2]) - + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2]) - + y[j+32] * (d * s[1] * (q[j+ 0] >> 4) - m * s[3]) - + y[j+48] * (d * s[1] * (q[j+16] >> 4) - m * s[3]); - } - tmp += sum; - } - -#endif - - // sum up partial sums and write back result - tmp = warp_reduce_sum(tmp); - - if (tid == 0) { - dst[row] = tmp; - } -} - -static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols) { - - const int row = blockIdx.x; - const int num_blocks_per_row = ncols / QK_K; - const int ib0 = row*num_blocks_per_row; - - const block_q5_K * x = (const block_q5_K *)vx + ib0; - - float tmp = 0; // partial sum for thread in warp - -#if QK_K == 256 - const uint16_t kmask1 = 0x3f3f; - const uint16_t kmask2 = 0x0f0f; - const uint16_t kmask3 = 0xc0c0; - - const int tid = threadIdx.x/2; // 0...15 - const int ix = threadIdx.x%2; - - const int il = tid/4; // 0...3 - const int ir = tid - 4*il;// 0...3 - const int n = 2; - - const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 - const int in = il%2; - - const int l0 = n*(2*ir + in); - const int q_offset = 32*im + l0; - const int y_offset = 64*im + l0; - - const uint8_t hm1 = 1 << (2*im); - const uint8_t hm2 = hm1 << 4; - - uint16_t aux[4]; - const uint8_t * sc = (const uint8_t *)aux; - - uint16_t q16[8]; - const uint8_t * q4 = (const uint8_t *)q16; - - for (int i = ix; i < num_blocks_per_row; i += 2) { - - const uint8_t * ql1 = x[i].qs + q_offset; - const uint8_t * qh = x[i].qh + l0; - const float * y1 = yy + i*QK_K + y_offset; - const float * y2 = y1 + 128; - - const float dall = __low2half(x[i].dm); - const float dmin = __high2half(x[i].dm); - - const uint16_t * a = (const uint16_t *)x[i].scales; - aux[0] = a[im+0] & kmask1; - aux[1] = a[im+2] & kmask1; - aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2); - aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2); - - float4 sum = {0.f, 0.f, 0.f, 0.f}; - float smin = 0; - const uint16_t * q1 = (const uint16_t *)ql1; - const uint16_t * q2 = q1 + 32; - q16[0] = q1[0] & 0x0f0f; - q16[1] = q1[8] & 0x0f0f; - q16[2] = (q1[0] >> 4) & 0x0f0f; - q16[3] = (q1[8] >> 4) & 0x0f0f; - q16[4] = q2[0] & 0x0f0f; - q16[5] = q2[8] & 0x0f0f; - q16[6] = (q2[0] >> 4) & 0x0f0f; - q16[7] = (q2[8] >> 4) & 0x0f0f; - for (int l = 0; l < n; ++l) { - sum.x += y1[l+ 0] * (q4[l +0] + (qh[l+ 0] & (hm1 << 0) ? 16 : 0)) - + y1[l+16] * (q4[l +2] + (qh[l+16] & (hm1 << 0) ? 16 : 0)); - sum.y += y1[l+32] * (q4[l +4] + (qh[l+ 0] & (hm1 << 1) ? 16 : 0)) - + y1[l+48] * (q4[l +6] + (qh[l+16] & (hm1 << 1) ? 16 : 0)); - sum.z += y2[l+ 0] * (q4[l +8] + (qh[l+ 0] & (hm2 << 0) ? 16 : 0)) - + y2[l+16] * (q4[l+10] + (qh[l+16] & (hm2 << 0) ? 16 : 0)); - sum.w += y2[l+32] * (q4[l+12] + (qh[l+ 0] & (hm2 << 1) ? 16 : 0)) - + y2[l+48] * (q4[l+14] + (qh[l+16] & (hm2 << 1) ? 16 : 0)); - smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3] - + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7]; - } - tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin; - } - -#else - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); - const int step = tid * K_QUANTS_PER_ITERATION; - const int im = step/8; - const int in = step%8; - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - const uint8_t * q = x[i].qs + step; - const int8_t * s = x[i].scales; - const float * y = yy + i*QK_K + step; - const float d = x[i].d; - float sum = 0.f; - for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { - const uint8_t h = x[i].qh[in+j] >> im; - sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16)) - + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16)) - + y[j+32] * d * s[2] * ((q[j+ 0] >> 4) - ((h >> 4) & 1 ? 0 : 16)) - + y[j+48] * d * s[3] * ((q[j+16] >> 4) - ((h >> 6) & 1 ? 0 : 16)); - } - tmp += sum; - } -#endif - - // sum up partial sums and write back result - tmp = warp_reduce_sum(tmp); - - if (threadIdx.x == 0) { - dst[row] = tmp; - } -} - -static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) { - - static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION"); - - const int row = blockIdx.x*blockDim.y + threadIdx.y; - if (row > nrows) return; - - const int num_blocks_per_row = ncols / QK_K; - const int ib0 = row*num_blocks_per_row; - - const block_q6_K * x = (const block_q6_K *)vx + ib0; - -#if QK_K == 256 - - const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16 - const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1 - - const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8 - - const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... - const int in = tid - step*im; // 0...15 or 0...7 - -#if K_QUANTS_PER_ITERATION == 1 - const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 - const int is = 0; -#else - const int l0 = 4 * in; // 0, 4, 8, ..., 28 - const int is = in / 4; -#endif - const int ql_offset = 64*im + l0; - const int qh_offset = 32*im + l0; - const int s_offset = 8*im + is; - const int y_offset = 128*im + l0; - - float tmp = 0; // partial sum for thread in warp - - for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { - - const float * y = yy + i * QK_K + y_offset; - const uint8_t * ql = x[i].ql + ql_offset; - const uint8_t * qh = x[i].qh + qh_offset; - const int8_t * s = x[i].scales + s_offset; - - const float d = x[i].d; - -#if K_QUANTS_PER_ITERATION == 1 - float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32) - + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32) - + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32) - + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32) - + y[64] * s[4] * d * ((int8_t)((ql[ 0] >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32) - + y[80] * s[5] * d * ((int8_t)((ql[16] >> 4) | ((qh[16] & 0x30) >> 0)) - 32) - + y[96] * s[6] * d * ((int8_t)((ql[32] >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32) - +y[112] * s[7] * d * ((int8_t)((ql[48] >> 4) | ((qh[16] & 0xc0) >> 2)) - 32); - tmp += sum; -#else - float sum = 0; - for (int l = 0; l < 4; ++l) { - sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32) - + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32) - + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32) - + y[l+96] * s[6] * d * ((int8_t)((ql[l+32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32); - } - tmp += sum; -#endif - - } - -#else - - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...7 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0...3 - - const int step = tid * K_QUANTS_PER_ITERATION; - - float tmp = 0; // partial sum for thread in warp - - for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { - - const float * y = yy + i * QK_K + step; - const uint8_t * ql = x[i].ql + step; - const uint8_t * qh = x[i].qh + step; - const int8_t * s = x[i].scales; - - const float d = x[i+0].d; - - float sum = 0; - for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { - sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32) - + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32) - + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >> 4) | ((qh[j] & 0x30) >> 0)) - 32) - + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >> 4) | ((qh[j] & 0xc0) >> 2)) - 32); - } - tmp += sum; - - } - -#endif - - // sum up partial sums and write back result - tmp = warp_reduce_sum(tmp); - - if (tid == 0) { - dst[row] = tmp; - } -} - -static __device__ void convert_f16(const void * vx, const int ib, const int iqs, dfloat2 & v){ - const half * x = (const half *) vx; - - // automatic half -> float type cast if dfloat == float - v.x = x[ib + iqs + 0]; - v.y = x[ib + iqs + 1]; -} - -static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded) { - const int ix = blockDim.x*blockIdx.x + threadIdx.x; - - if (ix >= kx_padded) { - return; - } - - const int iy = blockDim.y*blockIdx.y + threadIdx.y; - - const int i_padded = iy*kx_padded + ix; - - block_q8_1 * y = (block_q8_1 *) vy; - - const int ib = i_padded / QK8_1; // block index - const int iqs = i_padded % QK8_1; // quant index - - const float xi = ix < kx ? x[iy*kx + ix] : 0.0f; - float amax = fabsf(xi); - float sum = xi; - - amax = warp_reduce_max(amax); - sum = warp_reduce_sum(sum); - - const float d = amax / 127; - const int8_t q = amax == 0.0f ? 0 : roundf(xi / d); - - y[ib].qs[iqs] = q; - - if (iqs > 0) { - return; - } - - reinterpret_cast(y[ib].ds.x) = d; - reinterpret_cast(y[ib].ds.y) = sum; -} - -template -static __global__ void k_get_rows( - const void * src0, const int32_t * src1, dst_t * dst, - int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ - /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ - /*size_t s0,*/ size_t s1, size_t s2, size_t s3, - /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, - size_t s10, size_t s11, size_t s12/*, size_t s13*/) { - - const int i00 = (blockIdx.x*blockDim.x + threadIdx.x)*2; - const int i10 = blockDim.y*blockIdx.y + threadIdx.y; - const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12; - const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12; - - if (i00 >= ne00) { - return; - } - - const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; - - dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; - const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03; - - const int ib = i00/qk; // block index - const int iqs = (i00%qk)/qr; // quant index - const int iybs = i00 - i00%qk; // dst block start index - const int y_offset = qr == 1 ? 1 : qk/2; - - // dequantize - dfloat2 v; - dequantize_kernel(src0_row, ib, iqs, v); - - dst_row[iybs + iqs + 0] = v.x; - dst_row[iybs + iqs + y_offset] = v.y; -} - -template -static __global__ void k_get_rows_float( - const src0_t * src0, const int32_t * src1, dst_t * dst, - int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ - /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ - /*size_t s0,*/ size_t s1, size_t s2, size_t s3, - /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, - size_t s10, size_t s11, size_t s12/*, size_t s13*/) { - - const int i00 = blockIdx.x*blockDim.x + threadIdx.x; - const int i10 = blockDim.y*blockIdx.y + threadIdx.y; - const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12; - const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12; - - if (i00 >= ne00) { - return; - } - - const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; - - dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; - const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03); - - dst_row[i00] = src0_row[i00]; -} - -template -static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k) { - const int i = 2*(blockDim.x*blockIdx.x + threadIdx.x); - - if (i >= k) { - return; - } - - const int ib = i/qk; // block index - const int iqs = (i%qk)/qr; // quant index - const int iybs = i - i%qk; // y block start index - const int y_offset = qr == 1 ? 1 : qk/2; - - // dequantize - dfloat2 v; - dequantize_kernel(vx, ib, iqs, v); - - y[iybs + iqs + 0] = v.x; - y[iybs + iqs + y_offset] = v.y; -} - -template -static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - - const src_t * x = (src_t *) vx; - - y[i] = x[i]; -} - -template -static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, half * __restrict__ y, const int k) { -#if __CUDA_ARCH__ >= CC_PASCAL - constexpr int nint = CUDA_Q8_0_NE_ALIGN/sizeof(int) + WARP_SIZE; - - const int i0 = CUDA_Q8_0_NE_ALIGN*blockIdx.x; - const int * x0 = ((int *) vx) + blockIdx.x * nint; - half2 * y2 = (half2 *) (y + i0); - - __shared__ int vals[nint]; - -#pragma unroll - for (int ix0 = 0; ix0 < nint; ix0 += WARP_SIZE) { - if (need_check && i0*sizeof(block_q8_0)/QK8_0 + sizeof(int)*(ix0 + threadIdx.x) >= k*sizeof(block_q8_0)/QK8_0) { - break; - } - - const int ix = ix0 + threadIdx.x; - vals[ix] = x0[ix]; - } - -#pragma unroll - for (int iy = 0; iy < CUDA_Q8_0_NE_ALIGN; iy += 2*WARP_SIZE) { - if (need_check && i0 + iy + 2*threadIdx.x >= k) { - return; - } - - const half * b0 = ((const half *) vals) + (sizeof(block_q8_0)/sizeof(half)) * ((iy + 2*threadIdx.x)/QK8_0); - const half d = *b0; - const char2 qs = ((const char2 *) (b0 + 1))[threadIdx.x % (QK8_0/2)]; - - y2[iy/2 + threadIdx.x] = __hmul2(make_half2(qs.x, qs.y), __half2half2(d)); - } -#else - (void) vx; (void) y; (void) k; - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= CC_PASCAL -} - -// VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called -// MMVQ = mul_mat_vec_q, MMQ = mul_mat_q - -#define VDR_Q4_0_Q8_1_MMVQ 2 -#define VDR_Q4_0_Q8_1_MMQ 4 - -template static __device__ __forceinline__ float vec_dot_q4_0_q8_1_impl( - const int * v, const int * u, const float & d4, const half2 & ds8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - int sumi = 0; - -#pragma unroll - for (int i = 0; i < vdr; ++i) { - const int vi0 = (v[i] >> 0) & 0x0F0F0F0F; - const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; - - // SIMD dot product of quantized values - sumi = __dp4a(vi0, u[2*i+0], sumi); - sumi = __dp4a(vi1, u[2*i+1], sumi); - } - - const float2 ds8f = __half22float2(ds8); - - // second part effectively subtracts 8 from each quant value - return d4 * (sumi * ds8f.x - (8*vdr/QI4_0) * ds8f.y); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -#define VDR_Q4_1_Q8_1_MMVQ 2 -#define VDR_Q4_1_Q8_1_MMQ 4 - -template static __device__ __forceinline__ float vec_dot_q4_1_q8_1_impl( - const int * v, const int * u, const half2 & dm4, const half2 & ds8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - int sumi = 0; - -#pragma unroll - for (int i = 0; i < vdr; ++i) { - const int vi0 = (v[i] >> 0) & 0x0F0F0F0F; - const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; - - // SIMD dot product of quantized values - sumi = __dp4a(vi0, u[2*i+0], sumi); - sumi = __dp4a(vi1, u[2*i+1], sumi); - } - -#ifdef GGML_CUDA_F16 - const float2 tmp = __half22float2(__hmul2(dm4, ds8)); - const float d4d8 = tmp.x; - const float m4s8 = tmp.y; -#else - const float2 dm4f = __half22float2(dm4); - const float2 ds8f = __half22float2(ds8); - const float d4d8 = dm4f.x * ds8f.x; - const float m4s8 = dm4f.y * ds8f.y; -#endif // GGML_CUDA_F16 - - // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it - return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1)); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -#define VDR_Q5_0_Q8_1_MMVQ 2 -#define VDR_Q5_0_Q8_1_MMQ 4 - -template static __device__ __forceinline__ float vec_dot_q5_0_q8_1_impl( - const int * vl, const int * vh, const int * u, const float & d5, const half2 & ds8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - int sumi = 0; - -#pragma unroll - for (int i = 0; i < vdr; ++i) { - int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits - vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4 - vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12 - vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20 - vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28 - sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values - - int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits - vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4 - vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12 - vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20 - vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28 - sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values - } - - const float2 ds8f = __half22float2(ds8); - - // second part effectively subtracts 16 from each quant value - return d5 * (sumi * ds8f.x - (16*vdr/QI5_0) * ds8f.y); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -#define VDR_Q5_1_Q8_1_MMVQ 2 -#define VDR_Q5_1_Q8_1_MMQ 4 - -template static __device__ __forceinline__ float vec_dot_q5_1_q8_1_impl( - const int * vl, const int * vh, const int * u, const half2 & dm5, const half2 & ds8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - int sumi = 0; - -#pragma unroll - for (int i = 0; i < vdr; ++i) { - int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits - vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4 - vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12 - vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20 - vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28 - sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values - - int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits - vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4 - vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12 - vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20 - vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28 - sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values - } - -#ifdef GGML_CUDA_F16 - const float2 tmp = __half22float2(__hmul2(dm5, ds8)); - const float d5d8 = tmp.x; - const float m5s8 = tmp.y; -#else - const float2 dm5f = __half22float2(dm5); - const float2 ds8f = __half22float2(ds8); - const float d5d8 = dm5f.x * ds8f.x; - const float m5s8 = dm5f.y * ds8f.y; -#endif // GGML_CUDA_F16 - - // scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it - return sumi*d5d8 + m5s8 / (QI5_1 / vdr); - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -#define VDR_Q8_0_Q8_1_MMVQ 2 -#define VDR_Q8_0_Q8_1_MMQ 8 - -template static __device__ __forceinline__ float vec_dot_q8_0_q8_1_impl( - const int * v, const int * u, const float & d8_0, const float & d8_1) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - int sumi = 0; - -#pragma unroll - for (int i = 0; i < vdr; ++i) { - // SIMD dot product of quantized values - sumi = __dp4a(v[i], u[i], sumi); - } - - return d8_0*d8_1 * sumi; -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -template static __device__ __forceinline__ float vec_dot_q8_1_q8_1_impl( - const int * v, const int * u, const half2 & dm8, const half2 & ds8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - int sumi = 0; - -#pragma unroll - for (int i = 0; i < vdr; ++i) { - // SIMD dot product of quantized values - sumi = __dp4a(v[i], u[i], sumi); - } - -#ifdef GGML_CUDA_F16 - const float2 tmp = __half22float2(__hmul2(dm8, ds8)); - const float d8d8 = tmp.x; - const float m8s8 = tmp.y; -#else - const float2 dm8f = __half22float2(dm8); - const float2 ds8f = __half22float2(ds8); - const float d8d8 = dm8f.x * ds8f.x; - const float m8s8 = dm8f.y * ds8f.y; -#endif // GGML_CUDA_F16 - - // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it - return sumi*d8d8 + m8s8 / (QI8_1 / vdr); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -#define VDR_Q2_K_Q8_1_MMVQ 1 -#define VDR_Q2_K_Q8_1_MMQ 2 - -// contiguous v/x values -static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq( - const int & v, const int * __restrict__ u, const uint8_t * __restrict__ scales, - const half2 & dm2, const float * __restrict__ d8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - float sumf_d = 0.0f; - float sumf_m = 0.0f; - -#pragma unroll - for (int i = 0; i < QR2_K; ++i) { - const int sc = scales[2*i]; - - const int vi = (v >> (2*i)) & 0x03030303; - - sumf_d += d8[i] * (__dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product - - // fill int with 4x m - int m = sc >> 4; - m |= m << 8; - m |= m << 16; - sumf_m += d8[i] * __dp4a(m, u[i], 0); // multiply constant q2_K part with sum of q8_1 values - } - - const float2 dm2f = __half22float2(dm2); - - return dm2f.x*sumf_d - dm2f.y*sumf_m; -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -// contiguous u/y values -static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq( - const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ scales, - const half2 & dm2, const float & d8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - int sumi_d = 0; - int sumi_m = 0; - -#pragma unroll - for (int i0 = 0; i0 < QI8_1; i0 += QI8_1/2) { - int sumi_d_sc = 0; - - const int sc = scales[i0 / (QI8_1/2)]; - - // fill int with 4x m - int m = sc >> 4; - m |= m << 8; - m |= m << 16; - -#pragma unroll - for (int i = i0; i < i0 + QI8_1/2; ++i) { - sumi_d_sc = __dp4a(v[i], u[i], sumi_d_sc); // SIMD dot product - sumi_m = __dp4a(m, u[i], sumi_m); // multiply sum of q8_1 values with m - } - - sumi_d += sumi_d_sc * (sc & 0xF); - } - - const float2 dm2f = __half22float2(dm2); - - return d8 * (dm2f.x*sumi_d - dm2f.y*sumi_m); -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -#define VDR_Q3_K_Q8_1_MMVQ 1 -#define VDR_Q3_K_Q8_1_MMQ 2 - -// contiguous v/x values -static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq( - const int & vl, const int & vh, const int * __restrict__ u, const uint8_t * __restrict__ scales, - const int & scale_offset, const float & d3, const float * __restrict__ d8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - float sumf = 0.0f; - -#pragma unroll - for (int i = 0; i < QR3_K; ++i) { - const int isc = scale_offset + 2*i; - - const int isc_low = isc % (QK_K/32); - const int sc_shift_low = 4 * (isc / (QK_K/32)); - const int sc_low = (scales[isc_low] >> sc_shift_low) & 0xF; - - const int isc_high = isc % (QK_K/64); - const int sc_shift_high = 2 * (isc / (QK_K/64)); - const int sc_high = ((scales[(QK_K/32) + isc_high] >> sc_shift_high) & 3) << 4; - - const int sc = (sc_low | sc_high) - 32; - - const int vil = (vl >> (2*i)) & 0x03030303; - - const int vih = ((vh >> i) << 2) & 0x04040404; - - const int vi = __vsubss4(vil, vih); - - sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product - } - - return d3 * sumf; -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -// contiguous u/y values -static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq( - const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ scales, - const float & d3, const float & d8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - int sumi = 0; - -#pragma unroll - for (int i0 = 0; i0 < QR3_K*VDR_Q3_K_Q8_1_MMQ; i0 += QI8_1/2) { - int sumi_sc = 0; - - for (int i = i0; i < i0 + QI8_1/2; ++i) { - sumi_sc = __dp4a(v[i], u[i], sumi_sc); // SIMD dot product - } - - sumi += sumi_sc * scales[i0 / (QI8_1/2)]; - } - - return d3*d8 * sumi; -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -#define VDR_Q4_K_Q8_1_MMVQ 2 -#define VDR_Q4_K_Q8_1_MMQ 8 - -// contiguous v/x values -static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq( - const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc, - const uint8_t * __restrict__ m, const half2 & dm4, const float * __restrict__ d8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - float sumf_d = 0.0f; - float sumf_m = 0.0f; - -#pragma unroll - for (int i = 0; i < QR4_K; ++i) { - const int v0i = (v[0] >> (4*i)) & 0x0F0F0F0F; - const int v1i = (v[1] >> (4*i)) & 0x0F0F0F0F; - - const int dot1 = __dp4a(v1i, u[2*i+1], __dp4a(v0i, u[2*i+0], 0)); // SIMD dot product - const int dot2 = __dp4a(0x01010101, u[2*i+1], __dp4a(0x01010101, u[2*i+0], 0)); // sum of u - - sumf_d += d8[i] * (dot1 * sc[i]); - sumf_m += d8[i] * (dot2 * m[i]); // multiply constant part of q4_K with sum of q8_1 values - } - - const float2 dm4f = __half22float2(dm4); - - return dm4f.x*sumf_d - dm4f.y*sumf_m; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -// contiguous u/y values -static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq( - const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc, - const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - float sumf_d = 0.0f; - float sumf_m = 0.0f; - -#pragma unroll - for (int i = 0; i < QR4_K*VDR_Q4_K_Q8_1_MMQ/QI8_1; ++i) { - int sumi_d = 0; - -#pragma unroll - for (int j = 0; j < QI8_1; ++j) { - sumi_d = __dp4a((v[j] >> (4*i)) & 0x0F0F0F0F, u[i*QI8_1 + j], sumi_d); // SIMD dot product - } - - const float2 ds8f = __half22float2(ds8[i]); - - sumf_d += ds8f.x * (sc[i] * sumi_d); - sumf_m += ds8f.y * m[i]; // sum of q8_1 block * q4_K min val - } - - const float2 dm4f = __half22float2(dm4); - - return dm4f.x*sumf_d - dm4f.y*sumf_m; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -#define VDR_Q5_K_Q8_1_MMVQ 2 -#define VDR_Q5_K_Q8_1_MMQ 8 - -// contiguous v/x values -static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq( - const int * __restrict__ vl, const int * __restrict__ vh, const int * __restrict__ u, const uint8_t * __restrict__ sc, - const uint8_t * __restrict__ m, const half2 & dm5, const float * __restrict__ d8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - float sumf_d = 0.0f; - float sumf_m = 0.0f; - -#pragma unroll - for (int i = 0; i < QR5_K; ++i) { - const int vl0i = (vl[0] >> (4*i)) & 0x0F0F0F0F; - const int vl1i = (vl[1] >> (4*i)) & 0x0F0F0F0F; - - const int vh0i = ((vh[0] >> i) << 4) & 0x10101010; - const int vh1i = ((vh[1] >> i) << 4) & 0x10101010; - - const int v0i = vl0i | vh0i; - const int v1i = vl1i | vh1i; - - const int dot1 = __dp4a(v0i, u[2*i+0], __dp4a(v1i, u[2*i+1], 0)); // SIMD dot product - const int dot2 = __dp4a(0x01010101, u[2*i+0], __dp4a(0x01010101, u[2*i+1], 0)); // sum of u - - sumf_d += d8[i] * (dot1 * sc[i]); - sumf_m += d8[i] * (dot2 * m[i]); - - } - - const float2 dm5f = __half22float2(dm5); - - return dm5f.x*sumf_d - dm5f.y*sumf_m; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -// contiguous u/y values -static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq( - const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc, - const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - float sumf_d = 0.0f; - float sumf_m = 0.0f; - -#pragma unroll - for (int i = 0; i < QR5_K*VDR_Q5_K_Q8_1_MMQ/QI8_1; ++i) { - int sumi_d = 0; - -#pragma unroll - for (int j = 0; j < QI8_1; ++j) { - sumi_d = __dp4a(v[i*QI8_1 + j], u[i*QI8_1 + j], sumi_d); // SIMD dot product - } - - const float2 ds8f = __half22float2(ds8[i]); - - sumf_d += ds8f.x * (sc[i] * sumi_d); - sumf_m += ds8f.y * m[i]; // sum of q8_1 block * q4_K min val - } - - const float2 dm4f = __half22float2(dm4); - - return dm4f.x*sumf_d - dm4f.y*sumf_m; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -#define VDR_Q6_K_Q8_1_MMVQ 1 -#define VDR_Q6_K_Q8_1_MMQ 8 - -// contiguous v/x values -static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq( - const int & vl, const int & vh, const int * __restrict__ u, const int8_t * __restrict__ scales, - const float & d, const float * __restrict__ d8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - float sumf = 0.0f; - -#pragma unroll - for (int i = 0; i < QR6_K; ++i) { - const int sc = scales[4*i]; - - const int vil = (vl >> (4*i)) & 0x0F0F0F0F; - - const int vih = ((vh >> (4*i)) << 4) & 0x30303030; - - const int vi = __vsubss4((vil | vih), 0x20202020); // vi = (vil | vih) - 32 - - sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product - } - - return d*sumf; -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -// contiguous u/y values -static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq( - const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ sc, - const float & d6, const float * __restrict__ d8) { - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - float sumf_d = 0.0f; - -#pragma unroll - for (int i0 = 0; i0 < VDR_Q6_K_Q8_1_MMQ; i0 += 4) { - int2 sumi_d = {0, 0}; // 2 q6_K scales per q8_1 scale - -#pragma unroll - for (int i = i0; i < i0 + 2; ++i) { - sumi_d.x = __dp4a(v[2*i+0], u[2*i+0], sumi_d.x); // SIMD dot product - sumi_d.x = __dp4a(v[2*i+1], u[2*i+1], sumi_d.x); // SIMD dot product - - sumi_d.y = __dp4a(v[2*i+4], u[2*i+4], sumi_d.y); // SIMD dot product - sumi_d.y = __dp4a(v[2*i+5], u[2*i+5], sumi_d.y); // SIMD dot product - } - - sumf_d += d8[i0/4] * (sc[i0/2+0]*sumi_d.x + sc[i0/2+1]*sumi_d.y); - } - - return d6 * sumf_d; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A -} - -static __device__ __forceinline__ float vec_dot_q4_0_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { - - const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq; - - int v[VDR_Q4_0_Q8_1_MMVQ]; - int u[2*VDR_Q4_0_Q8_1_MMVQ]; - -#pragma unroll - for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) { - v[i] = get_int_from_uint8(bq4_0->qs, iqs + i); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0); - } - - return vec_dot_q4_0_q8_1_impl(v, u, bq4_0->d, bq8_1->ds); -} - -template static __device__ __forceinline__ void allocate_tiles_q4_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { - (void)x_qh; (void)x_sc; - - __shared__ int tile_x_qs[mmq_y * (WARP_SIZE) + mmq_y]; - __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI4_0) + mmq_y/QI4_0]; - - *x_ql = tile_x_qs; - *x_dm = (half2 *) tile_x_d; -} - -template static __device__ __forceinline__ void load_tiles_q4_0( - const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, - int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { - (void)x_qh; (void)x_sc; - GGML_CUDA_ASSUME(i_offset >= 0); - GGML_CUDA_ASSUME(i_offset < nwarps); - GGML_CUDA_ASSUME(k >= 0); - GGML_CUDA_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI4_0; - const int kqsx = k % QI4_0; - - const block_q4_0 * bx0 = (const block_q4_0 *) vx; - - float * x_dmf = (float *) x_dm; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbx; - - x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx); - // x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbx] = bxi->d; - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI4_0; - const int kbxd = k % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_0) { - int i = i0 + i_offset * QI4_0 + k / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbxd] = bxi->d; - } -} - -static __device__ __forceinline__ float vec_dot_q4_0_q8_1_mul_mat( - const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, - const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { - (void)x_qh; (void)x_sc; - - const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); - const float * x_dmf = (const float *) x_dm; - - int u[2*VDR_Q4_0_Q8_1_MMQ]; - -#pragma unroll - for (int l = 0; l < VDR_Q4_0_Q8_1_MMQ; ++l) { - u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; - u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_0) % WARP_SIZE]; - } - - return vec_dot_q4_0_q8_1_impl - (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dmf[i * (WARP_SIZE/QI4_0) + i/QI4_0 + k/QI4_0], - y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); -} - -static __device__ __forceinline__ float vec_dot_q4_1_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { - - const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq; - - int v[VDR_Q4_1_Q8_1_MMVQ]; - int u[2*VDR_Q4_1_Q8_1_MMVQ]; - -#pragma unroll - for (int i = 0; i < VDR_Q4_1_Q8_1_MMVQ; ++i) { - v[i] = get_int_from_uint8_aligned(bq4_1->qs, iqs + i); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_1); - } - - return vec_dot_q4_1_q8_1_impl(v, u, bq4_1->dm, bq8_1->ds); -} - -template static __device__ __forceinline__ void allocate_tiles_q4_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { - (void)x_qh; (void)x_sc; - - __shared__ int tile_x_qs[mmq_y * (WARP_SIZE) + + mmq_y]; - __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI4_1) + mmq_y/QI4_1]; - - *x_ql = tile_x_qs; - *x_dm = tile_x_dm; -} - -template static __device__ __forceinline__ void load_tiles_q4_1( - const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, - int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { - (void)x_qh; (void)x_sc; - - GGML_CUDA_ASSUME(i_offset >= 0); - GGML_CUDA_ASSUME(i_offset < nwarps); - GGML_CUDA_ASSUME(k >= 0); - GGML_CUDA_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI4_1; - const int kqsx = k % QI4_1; - - const block_q4_1 * bx0 = (const block_q4_1 *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbx; - - x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI4_1; - const int kbxd = k % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_1) { - int i = i0 + i_offset * QI4_1 + k / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dm[i * (WARP_SIZE/QI4_1) + i / QI4_1 + kbxd] = bxi->dm; - } -} - -static __device__ __forceinline__ float vec_dot_q4_1_q8_1_mul_mat( - const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, - const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { - (void)x_qh; (void)x_sc; - - const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); - - int u[2*VDR_Q4_1_Q8_1_MMQ]; - -#pragma unroll - for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) { - u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; - u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_1) % WARP_SIZE]; - } - - return vec_dot_q4_1_q8_1_impl - (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dm[i * (WARP_SIZE/QI4_1) + i/QI4_1 + k/QI4_1], - y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); -} - -static __device__ __forceinline__ float vec_dot_q5_0_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { - - const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq; - - int vl[VDR_Q5_0_Q8_1_MMVQ]; - int vh[VDR_Q5_0_Q8_1_MMVQ]; - int u[2*VDR_Q5_0_Q8_1_MMVQ]; - -#pragma unroll - for (int i = 0; i < VDR_Q5_0_Q8_1_MMVQ; ++i) { - vl[i] = get_int_from_uint8(bq5_0->qs, iqs + i); - vh[i] = get_int_from_uint8(bq5_0->qh, 0) >> (4 * (iqs + i)); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_0); - } - - return vec_dot_q5_0_q8_1_impl(vl, vh, u, bq5_0->d, bq8_1->ds); -} - -template static __device__ __forceinline__ void allocate_tiles_q5_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { - (void)x_qh; (void)x_sc; - - __shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y]; - __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI5_0) + mmq_y/QI5_0]; - - *x_ql = tile_x_ql; - *x_dm = (half2 *) tile_x_d; -} - -template static __device__ __forceinline__ void load_tiles_q5_0( - const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, - int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { - (void)x_qh; (void)x_sc; - - GGML_CUDA_ASSUME(i_offset >= 0); - GGML_CUDA_ASSUME(i_offset < nwarps); - GGML_CUDA_ASSUME(k >= 0); - GGML_CUDA_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI5_0; - const int kqsx = k % QI5_0; - - const block_q5_0 * bx0 = (const block_q5_0 *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbx; - - const int ql = get_int_from_uint8(bxi->qs, kqsx); - const int qh = get_int_from_uint8(bxi->qh, 0) >> (4 * (k % QI5_0)); - - int qs0 = (ql >> 0) & 0x0F0F0F0F; - qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 - qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 - qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 - qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 - qs0 = __vsubss4(qs0, 0x10101010); // subtract 16 - - x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0; - - int qs1 = (ql >> 4) & 0x0F0F0F0F; - qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 - qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 - qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 - qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 - qs1 = __vsubss4(qs1, 0x10101010); // subtract 16 - - x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1; - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI5_0; - const int kbxd = k % blocks_per_tile_x_row; - float * x_dmf = (float *) x_dm; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_0) { - int i = i0 + i_offset * QI5_0 + k / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dmf[i * (WARP_SIZE/QI5_0) + i / QI5_0 + kbxd] = bxi->d; - } -} - -static __device__ __forceinline__ float vec_dot_q5_0_q8_1_mul_mat( - const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, - const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { - (void)x_qh; (void)x_sc; - - const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); - const int index_bx = i * (WARP_SIZE/QI5_0) + i/QI5_0 + k/QI5_0; - const float * x_dmf = (const float *) x_dm; - const float * y_df = (const float *) y_ds; - - int u[2*VDR_Q5_0_Q8_1_MMQ]; - -#pragma unroll - for (int l = 0; l < VDR_Q5_0_Q8_1_MMQ; ++l) { - u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; - u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_0) % WARP_SIZE]; - } - - return vec_dot_q8_0_q8_1_impl - (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dmf[index_bx], y_df[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); -} - -static __device__ __forceinline__ float vec_dot_q5_1_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { - - const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq; - - int vl[VDR_Q5_1_Q8_1_MMVQ]; - int vh[VDR_Q5_1_Q8_1_MMVQ]; - int u[2*VDR_Q5_1_Q8_1_MMVQ]; - -#pragma unroll - for (int i = 0; i < VDR_Q5_1_Q8_1_MMVQ; ++i) { - vl[i] = get_int_from_uint8_aligned(bq5_1->qs, iqs + i); - vh[i] = get_int_from_uint8_aligned(bq5_1->qh, 0) >> (4 * (iqs + i)); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_1); - } - - return vec_dot_q5_1_q8_1_impl(vl, vh, u, bq5_1->dm, bq8_1->ds); -} - -template static __device__ __forceinline__ void allocate_tiles_q5_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { - (void)x_qh; (void)x_sc; - - __shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y]; - __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI5_1) + mmq_y/QI5_1]; - - *x_ql = tile_x_ql; - *x_dm = tile_x_dm; -} - -template static __device__ __forceinline__ void load_tiles_q5_1( - const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, - int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { - (void)x_qh; (void)x_sc; - - GGML_CUDA_ASSUME(i_offset >= 0); - GGML_CUDA_ASSUME(i_offset < nwarps); - GGML_CUDA_ASSUME(k >= 0); - GGML_CUDA_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI5_1; - const int kqsx = k % QI5_1; - - const block_q5_1 * bx0 = (const block_q5_1 *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbx; - - const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx); - const int qh = get_int_from_uint8_aligned(bxi->qh, 0) >> (4 * (k % QI5_1)); - - int qs0 = (ql >> 0) & 0x0F0F0F0F; - qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 - qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 - qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 - qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 - - x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0; - - int qs1 = (ql >> 4) & 0x0F0F0F0F; - qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 - qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 - qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 - qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 - - x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1; - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI5_1; - const int kbxd = k % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_1) { - int i = i0 + i_offset * QI5_1 + k / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dm[i * (WARP_SIZE/QI5_1) + i / QI5_1 + kbxd] = bxi->dm; - } -} - -static __device__ __forceinline__ float vec_dot_q5_1_q8_1_mul_mat( - const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, - const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { - (void)x_qh; (void)x_sc; - - const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); - const int index_bx = i * (WARP_SIZE/QI5_1) + + i/QI5_1 + k/QI5_1; - - int u[2*VDR_Q5_1_Q8_1_MMQ]; - -#pragma unroll - for (int l = 0; l < VDR_Q5_1_Q8_1_MMQ; ++l) { - u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; - u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_1) % WARP_SIZE]; - } - - return vec_dot_q8_1_q8_1_impl - (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dm[index_bx], y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); -} - -static __device__ __forceinline__ float vec_dot_q8_0_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { - - const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq; - - int v[VDR_Q8_0_Q8_1_MMVQ]; - int u[VDR_Q8_0_Q8_1_MMVQ]; - -#pragma unroll - for (int i = 0; i < VDR_Q8_0_Q8_1_MMVQ; ++i) { - v[i] = get_int_from_int8(bq8_0->qs, iqs + i); - u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - } - - return vec_dot_q8_0_q8_1_impl(v, u, bq8_0->d, __low2half(bq8_1->ds)); -} - -template static __device__ __forceinline__ void allocate_tiles_q8_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { - (void)x_qh; (void)x_sc; - - __shared__ int tile_x_qs[mmq_y * (WARP_SIZE) + mmq_y]; - __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI8_0) + mmq_y/QI8_0]; - - *x_ql = tile_x_qs; - *x_dm = (half2 *) tile_x_d; -} - -template static __device__ __forceinline__ void load_tiles_q8_0( - const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, - int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { - (void)x_qh; (void)x_sc; - - GGML_CUDA_ASSUME(i_offset >= 0); - GGML_CUDA_ASSUME(i_offset < nwarps); - GGML_CUDA_ASSUME(k >= 0); - GGML_CUDA_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI8_0; - const int kqsx = k % QI8_0; - float * x_dmf = (float *) x_dm; - - const block_q8_0 * bx0 = (const block_q8_0 *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = min(i, i_max); - } - - const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbx; - - x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_int8(bxi->qs, kqsx); - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI8_0; - const int kbxd = k % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI8_0) { - int i = i0 + i_offset * QI8_0 + k / blocks_per_tile_x_row; - - if (need_check) { - i = min(i, i_max); - } - - const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dmf[i * (WARP_SIZE/QI8_0) + i / QI8_0 + kbxd] = bxi->d; - } -} - -static __device__ __forceinline__ float vec_dot_q8_0_q8_1_mul_mat( - const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, - const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { - (void)x_qh; (void)x_sc; - - const float * x_dmf = (const float *) x_dm; - const float * y_df = (const float *) y_ds; - - return vec_dot_q8_0_q8_1_impl - (&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[j * WARP_SIZE + k], x_dmf[i * (WARP_SIZE/QI8_0) + i/QI8_0 + k/QI8_0], - y_df[j * (WARP_SIZE/QI8_1) + k/QI8_1]); -} - -static __device__ __forceinline__ float vec_dot_q2_K_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { - - const block_q2_K * bq2_K = (const block_q2_K *) vbq; - - const int bq8_offset = QR2_K * (iqs / QI8_1); - const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2); - - const uint8_t * scales = bq2_K->scales + scale_offset; - - const int v = get_int_from_uint8_aligned(bq2_K->qs, iqs); - int u[QR2_K]; - float d8[QR2_K]; - -#pragma unroll - for (int i = 0; i < QR2_K; ++ i) { - u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); - d8[i] = __low2float(bq8_1[bq8_offset + i].ds); - } - - return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8); -} - -template static __device__ __forceinline__ void allocate_tiles_q2_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { - (void)x_qh; - - __shared__ int tile_x_ql[mmq_y * (WARP_SIZE) + mmq_y]; - __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI2_K) + mmq_y/QI2_K]; - __shared__ int tile_x_sc[mmq_y * (WARP_SIZE/4) + mmq_y/4]; - - *x_ql = tile_x_ql; - *x_dm = tile_x_dm; - *x_sc = tile_x_sc; -} - -template static __device__ __forceinline__ void load_tiles_q2_K( - const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, - int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { - (void)x_qh; - - GGML_CUDA_ASSUME(i_offset >= 0); - GGML_CUDA_ASSUME(i_offset < nwarps); - GGML_CUDA_ASSUME(k >= 0); - GGML_CUDA_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI2_K; - const int kqsx = k % QI2_K; - - const block_q2_K * bx0 = (const block_q2_K *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = min(i, i_max); - } - - const block_q2_K * bxi = bx0 + i*blocks_per_row + kbx; - - x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI2_K; - const int kbxd = k % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI2_K) { - int i = (i0 + i_offset * QI2_K + k / blocks_per_tile_x_row) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q2_K * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dm[i * (WARP_SIZE/QI2_K) + i / QI2_K + kbxd] = bxi->dm; - } - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) { - int i = i0 + i_offset * 4 + k / (WARP_SIZE/4); - - if (need_check) { - i = min(i, i_max); - } - - const block_q2_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI2_K/4); - - x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = get_int_from_uint8_aligned(bxi->scales, k % (QI2_K/4)); - } -} - -static __device__ __forceinline__ float vec_dot_q2_K_q8_1_mul_mat( - const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, - const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { - (void)x_qh; - - const int kbx = k / QI2_K; - const int ky = (k % QI2_K) * QR2_K; - const float * y_df = (const float *) y_ds; - - int v[QR2_K*VDR_Q2_K_Q8_1_MMQ]; - - const int kqsx = i * (WARP_SIZE + 1) + kbx*QI2_K + (QI2_K/2) * (ky/(2*QI2_K)) + ky % (QI2_K/2); - const int shift = 2 * ((ky % (2*QI2_K)) / (QI2_K/2)); - -#pragma unroll - for (int l = 0; l < QR2_K*VDR_Q2_K_Q8_1_MMQ; ++l) { - v[l] = (x_ql[kqsx + l] >> shift) & 0x03030303; - } - - const uint8_t * scales = ((const uint8_t *) &x_sc[i * (WARP_SIZE/4) + i/4 + kbx*4]) + ky/4; - - const int index_y = j * WARP_SIZE + (QR2_K*k) % WARP_SIZE; - return vec_dot_q2_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dm[i * (WARP_SIZE/QI2_K) + i/QI2_K + kbx], y_df[index_y/QI8_1]); -} - -static __device__ __forceinline__ float vec_dot_q3_K_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { - - const block_q3_K * bq3_K = (const block_q3_K *) vbq; - - const int bq8_offset = QR3_K * (iqs / (QI3_K/2)); - const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2); - - const float d = bq3_K->d; - - const int vl = get_int_from_uint8(bq3_K->qs, iqs); - - // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted - const int vh = ~get_int_from_uint8(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset; - - int u[QR3_K]; - float d8[QR3_K]; - -#pragma unroll - for (int i = 0; i < QR3_K; ++i) { - u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); - d8[i] = __low2float(bq8_1[bq8_offset + i].ds); - } - - return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8); -} - -template static __device__ __forceinline__ void allocate_tiles_q3_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { - - __shared__ int tile_x_ql[mmq_y * (WARP_SIZE) + mmq_y]; - __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI3_K) + mmq_y/QI3_K]; - __shared__ int tile_x_qh[mmq_y * (WARP_SIZE/2) + mmq_y/2]; - __shared__ int tile_x_sc[mmq_y * (WARP_SIZE/4) + mmq_y/4]; - - *x_ql = tile_x_ql; - *x_dm = tile_x_dm; - *x_qh = tile_x_qh; - *x_sc = tile_x_sc; -} - -template static __device__ __forceinline__ void load_tiles_q3_K( - const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, - int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { - - GGML_CUDA_ASSUME(i_offset >= 0); - GGML_CUDA_ASSUME(i_offset < nwarps); - GGML_CUDA_ASSUME(k >= 0); - GGML_CUDA_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI3_K; - const int kqsx = k % QI3_K; - - const block_q3_K * bx0 = (const block_q3_K *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = min(i, i_max); - } - - const block_q3_K * bxi = bx0 + i*blocks_per_row + kbx; - - x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx); - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI3_K; - const int kbxd = k % blocks_per_tile_x_row; - float * x_dmf = (float *) x_dm; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI3_K) { - int i = (i0 + i_offset * QI3_K + k / blocks_per_tile_x_row) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q3_K * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dmf[i * (WARP_SIZE/QI3_K) + i / QI3_K + kbxd] = bxi->d; - } - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 2) { - int i = i0 + i_offset * 2 + k / (WARP_SIZE/2); - - if (need_check) { - i = min(i, i_max); - } - - const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/2)) / (QI3_K/2); - - // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted - x_qh[i * (WARP_SIZE/2) + i / 2 + k % (WARP_SIZE/2)] = ~get_int_from_uint8(bxi->hmask, k % (QI3_K/2)); - } - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) { - int i = i0 + i_offset * 4 + k / (WARP_SIZE/4); - - if (need_check) { - i = min(i, i_max); - } - - const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI3_K/4); - - const int ksc = k % (QI3_K/4); - - const int ksc_low = ksc % (QI3_K/8); - const int shift_low = 4 * (ksc / (QI3_K/8)); - const int sc_low = (get_int_from_uint8(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F; - - const int ksc_high = QI3_K/8; - const int shift_high = 2 * ksc; - const int sc_high = ((get_int_from_uint8(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030; - - const int sc = __vsubss4(sc_low | sc_high, 0x20202020); - - x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = sc; - } -} - -static __device__ __forceinline__ float vec_dot_q3_K_q8_1_mul_mat( - const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, - const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { - - const int kbx = k / QI3_K; - const int ky = (k % QI3_K) * QR3_K; - const float * x_dmf = (const float *) x_dm; - const float * y_df = (const float *) y_ds; - - const int8_t * scales = ((const int8_t *) (x_sc + i * (WARP_SIZE/4) + i/4 + kbx*4)) + ky/4; - - int v[QR3_K*VDR_Q3_K_Q8_1_MMQ]; - -#pragma unroll - for (int l = 0; l < QR3_K*VDR_Q3_K_Q8_1_MMQ; ++l) { - const int kqsx = i * (WARP_SIZE + 1) + kbx*QI3_K + (QI3_K/2) * (ky/(2*QI3_K)) + ky % (QI3_K/2); - const int shift = 2 * ((ky % 32) / 8); - const int vll = (x_ql[kqsx + l] >> shift) & 0x03030303; - - const int vh = x_qh[i * (WARP_SIZE/2) + i/2 + kbx * (QI3_K/2) + (ky+l)%8] >> ((ky+l) / 8); - const int vlh = (vh << 2) & 0x04040404; - - v[l] = __vsubss4(vll, vlh); - } - - const int index_y = j * WARP_SIZE + (k*QR3_K) % WARP_SIZE; - return vec_dot_q3_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dmf[i * (WARP_SIZE/QI3_K) + i/QI3_K + kbx], y_df[index_y/QI8_1]); -} - -static __device__ __forceinline__ float vec_dot_q4_K_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { - -#ifndef GGML_QKK_64 - const block_q4_K * bq4_K = (const block_q4_K *) vbq; - - int v[2]; - int u[2*QR4_K]; - float d8[QR4_K]; - - // iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6 - const int bq8_offset = QR4_K * ((iqs/2) / (QI8_1/2)); - - // iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12 - // iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44 - // iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76 - // iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108 - - const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4)); - v[0] = q4[0]; - v[1] = q4[4]; - - const uint16_t * scales = (const uint16_t *)bq4_K->scales; - uint16_t aux[2]; - const int j = bq8_offset/2; - if (j < 2) { - aux[0] = scales[j+0] & 0x3f3f; - aux[1] = scales[j+2] & 0x3f3f; + if (id == ggml_backend_cuda_get_device_count() - 1) { + *row_high = nrows; } else { - aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2); - aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2); - } - const uint8_t * sc = (const uint8_t *)aux; - const uint8_t * m = sc + 2; - - for (int i = 0; i < QR4_K; ++i) { - const block_q8_1 * bq8i = bq8_1 + bq8_offset + i; - d8[i] = __low2float(bq8i->ds); - - const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4); - u[2*i+0] = q8[0]; - u[2*i+1] = q8[4]; - } - - return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8); - -#else - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const block_q4_K * bq4_K = (const block_q4_K *) vbq; - - float sumf_d = 0.0f; - float sumf_m = 0.0f; - - uint16_t aux16[2]; - const uint8_t * s = (const uint8_t *)aux16; - - const uint16_t * a = (const uint16_t *)bq4_K->scales; - aux16[0] = a[0] & 0x0f0f; - aux16[1] = (a[0] >> 4) & 0x0f0f; - - const float dall = bq4_K->dm[0]; - const float dmin = bq4_K->dm[1]; - - const float d8_1 = __low2float(bq8_1[0].ds); - const float d8_2 = __low2float(bq8_1[1].ds); - - const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); - const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); - const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2)); - const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4); - - const int * q4 = (const int *)bq4_K->qs + (iqs/2); - const int v1 = q4[0]; - const int v2 = q4[4]; - - const int dot1 = __dp4a(ui2, v2 & 0x0f0f0f0f, __dp4a(ui1, v1 & 0x0f0f0f0f, 0)); - const int dot2 = __dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, __dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0)); - const int dot3 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0)); - const int dot4 = __dp4a(0x01010101, ui4, __dp4a(0x01010101, ui3, 0)); - - sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]); - sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]); - - return dall * sumf_d - dmin * sumf_m; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A - -#endif -} - -template static __device__ __forceinline__ void allocate_tiles_q4_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { - (void)x_qh; - - __shared__ int tile_x_ql[mmq_y * (WARP_SIZE) + mmq_y]; - __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI4_K) + mmq_y/QI4_K]; - __shared__ int tile_x_sc[mmq_y * (WARP_SIZE/8) + mmq_y/8]; - - *x_ql = tile_x_ql; - *x_dm = tile_x_dm; - *x_sc = tile_x_sc; -} - -template static __device__ __forceinline__ void load_tiles_q4_K( - const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, - int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { - (void)x_qh; - - GGML_CUDA_ASSUME(i_offset >= 0); - GGML_CUDA_ASSUME(i_offset < nwarps); - GGML_CUDA_ASSUME(k >= 0); - GGML_CUDA_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI4_K; // == 0 if QK_K == 256 - const int kqsx = k % QI4_K; // == k if QK_K == 256 - - const block_q4_K * bx0 = (const block_q4_K *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_K * bxi = bx0 + i*blocks_per_row + kbx; - - x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI4_K; // == 1 if QK_K == 256 - const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_K) { - int i = (i0 + i_offset * QI4_K + k / blocks_per_tile_x_row) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd; - -#if QK_K == 256 - x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm; -#else - x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]}; -#endif - } - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { - int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q4_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI4_K/8); - - const int * scales = (const int *) bxi->scales; - - const int ksc = k % (WARP_SIZE/8); - - // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8 - int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits - scales8 |= (scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030; // upper 2 bits - - x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8; + *row_high = nrows*tensor_split[id + 1]; + *row_high -= *row_high % rounding; } } -static __device__ __forceinline__ float vec_dot_q4_K_q8_1_mul_mat( - const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, - const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { - (void)x_qh; +static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) { + static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); - const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2*((k % 16) / 8); - - const int index_y = j * WARP_SIZE + (QR4_K*k) % WARP_SIZE; - return vec_dot_q4_K_q8_1_impl_mmq(&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[index_y], sc, sc+8, - x_dm[i * (WARP_SIZE/QI4_K) + i/QI4_K], &y_ds[index_y/QI8_1]); + return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]); } -static __device__ __forceinline__ float vec_dot_q5_K_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { - -#ifndef GGML_QKK_64 - const block_q5_K * bq5_K = (const block_q5_K *) vbq; - - int vl[2]; - int vh[2]; - int u[2*QR5_K]; - float d8[QR5_K]; - - const int bq8_offset = QR5_K * ((iqs/2) / (QI8_1/2)); - const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4)); - const int * qh = (const int *)(bq5_K->qh + 4 * ((iqs/2)%4)); - - vl[0] = ql[0]; - vl[1] = ql[4]; - - vh[0] = qh[0] >> bq8_offset; - vh[1] = qh[4] >> bq8_offset; - - const uint16_t * scales = (const uint16_t *)bq5_K->scales; - uint16_t aux[2]; - const int j = bq8_offset/2; - if (j < 2) { - aux[0] = scales[j+0] & 0x3f3f; - aux[1] = scales[j+2] & 0x3f3f; - } else { - aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2); - aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2); - } - const uint8_t * sc = (const uint8_t *)aux; - const uint8_t * m = sc + 2; - -#pragma unroll - for (int i = 0; i < QR5_K; ++i) { - const block_q8_1 * bq8i = bq8_1 + bq8_offset + i; - d8[i] = __low2float(bq8i->ds); - - const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4); - u[2*i+0] = q8[0]; - u[2*i+1] = q8[4]; - } - - return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8); - -#else - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const block_q5_K * bq5_K = (const block_q5_K *) vbq; - - const int8_t * s = bq5_K->scales; - - const float d = bq5_K->d; - - const float d8_1 = __low2half(bq8_1[0].ds); - const float d8_2 = __low2half(bq8_1[1].ds); - - const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); - const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); - const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2)); - const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4); - - const int * ql = (const int *)bq5_K->qs + (iqs/2); - const int vl1 = ql[0]; - const int vl2 = ql[4]; - - const int step = 4 * (iqs/2); // 0, 4, 8, 12 - const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6 - const int in = step%8; // 0, 4, 0, 4 - const int vh = (*((const int *)(bq5_K->qh + in))) >> im; - - const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f); - const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f); - const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f); - const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f); - - const float sumf_d = d8_1 * (__dp4a(ui1, v1, 0) * s[0] + __dp4a(ui2, v2, 0) * s[1]) - + d8_2 * (__dp4a(ui3, v3, 0) * s[2] + __dp4a(ui4, v4, 0) * s[3]); - - return d * sumf_d; - -#else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= MIN_CC_DP4A - -#endif -} - -template static __device__ __forceinline__ void allocate_tiles_q5_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { - (void)x_qh; - - __shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y]; - __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI5_K) + mmq_y/QI5_K]; - __shared__ int tile_x_sc[mmq_y * (WARP_SIZE/8) + mmq_y/8]; - - *x_ql = tile_x_ql; - *x_dm = tile_x_dm; - *x_sc = tile_x_sc; -} - -template static __device__ __forceinline__ void load_tiles_q5_K( - const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, - int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { - (void)x_qh; - - GGML_CUDA_ASSUME(i_offset >= 0); - GGML_CUDA_ASSUME(i_offset < nwarps); - GGML_CUDA_ASSUME(k >= 0); - GGML_CUDA_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI5_K; // == 0 if QK_K == 256 - const int kqsx = k % QI5_K; // == k if QK_K == 256 - - const block_q5_K * bx0 = (const block_q5_K *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_K * bxi = bx0 + i*blocks_per_row + kbx; - const int ky = QR5_K*kqsx; - - const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx); - const int ql0 = (ql >> 0) & 0x0F0F0F0F; - const int ql1 = (ql >> 4) & 0x0F0F0F0F; - - const int qh = get_int_from_uint8_aligned(bxi->qh, kqsx % (QI5_K/4)); - const int qh0 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 0)) << 4) & 0x10101010; - const int qh1 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 1)) << 4) & 0x10101010; - - const int kq0 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + 0; - const int kq1 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + (QI5_K/4); - - x_ql[i * (2*WARP_SIZE + 1) + kq0] = ql0 | qh0; - x_ql[i * (2*WARP_SIZE + 1) + kq1] = ql1 | qh1; - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI5_K; // == 1 if QK_K == 256 - const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_K) { - int i = (i0 + i_offset * QI5_K + k / blocks_per_tile_x_row) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd; - -#if QK_K == 256 - x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm; -#endif - } - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { - int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q5_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI5_K/8); - - const int * scales = (const int *) bxi->scales; - - const int ksc = k % (WARP_SIZE/8); - - // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8 - int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits - scales8 |= (scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030; // upper 2 bits - - x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8; - } -} - -static __device__ __forceinline__ float vec_dot_q5_K_q8_1_mul_mat( - const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, - const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { - (void)x_qh; - - const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2 * ((k % 16) / 8); - - const int index_x = i * (QR5_K*WARP_SIZE + 1) + QR5_K*k; - const int index_y = j * WARP_SIZE + (QR5_K*k) % WARP_SIZE; - return vec_dot_q5_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, sc+8, - x_dm[i * (WARP_SIZE/QI5_K) + i/QI5_K], &y_ds[index_y/QI8_1]); -} - -static __device__ __forceinline__ float vec_dot_q6_K_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { - - const block_q6_K * bq6_K = (const block_q6_K *) vbq; - - const int bq8_offset = 2 * QR6_K * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/4); - const int scale_offset = (QI6_K/4) * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/8); - const int vh_shift = 2 * ((iqs % (QI6_K/2)) / (QI6_K/4)); - - const int vl = get_int_from_uint8(bq6_K->ql, iqs); - const int vh = get_int_from_uint8(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift; - - const int8_t * scales = bq6_K->scales + scale_offset; - - int u[QR6_K]; - float d8[QR6_K]; - -#pragma unroll - for (int i = 0; i < QR6_K; ++i) { - u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1); - d8[i] = __low2float(bq8_1[bq8_offset + 2*i].ds); - } - - return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8); -} - -template static __device__ __forceinline__ void allocate_tiles_q6_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { - (void)x_qh; - - __shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y]; - __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI6_K) + mmq_y/QI6_K]; - __shared__ int tile_x_sc[mmq_y * (WARP_SIZE/8) + mmq_y/8]; - - *x_ql = tile_x_ql; - *x_dm = tile_x_dm; - *x_sc = tile_x_sc; -} - -template static __device__ __forceinline__ void load_tiles_q6_K( - const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, - int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { - (void)x_qh; - - GGML_CUDA_ASSUME(i_offset >= 0); - GGML_CUDA_ASSUME(i_offset < nwarps); - GGML_CUDA_ASSUME(k >= 0); - GGML_CUDA_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI6_K; // == 0 if QK_K == 256 - const int kqsx = k % QI6_K; // == k if QK_K == 256 - - const block_q6_K * bx0 = (const block_q6_K *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = min(i, i_max); - } - - const block_q6_K * bxi = bx0 + i*blocks_per_row + kbx; - const int ky = QR6_K*kqsx; - - const int ql = get_int_from_uint8(bxi->ql, kqsx); - const int ql0 = (ql >> 0) & 0x0F0F0F0F; - const int ql1 = (ql >> 4) & 0x0F0F0F0F; - - const int qh = get_int_from_uint8(bxi->qh, (QI6_K/4) * (kqsx / (QI6_K/2)) + kqsx % (QI6_K/4)); - const int qh0 = ((qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) << 4) & 0x30303030; - const int qh1 = (qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) & 0x30303030; - - const int kq0 = ky - ky % QI6_K + k % (QI6_K/2) + 0; - const int kq1 = ky - ky % QI6_K + k % (QI6_K/2) + (QI6_K/2); - - x_ql[i * (2*WARP_SIZE + 1) + kq0] = __vsubss4(ql0 | qh0, 0x20202020); - x_ql[i * (2*WARP_SIZE + 1) + kq1] = __vsubss4(ql1 | qh1, 0x20202020); - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI6_K; // == 1 if QK_K == 256 - const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 - float * x_dmf = (float *) x_dm; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI6_K) { - int i = (i0 + i_offset * QI6_K + k / blocks_per_tile_x_row) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q6_K * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dmf[i * (WARP_SIZE/QI6_K) + i / QI6_K + kbxd] = bxi->d; - } - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { - int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; - - if (need_check) { - i = min(i, i_max); - } - - const block_q6_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / 4; - - x_sc[i * (WARP_SIZE/8) + i / 8 + k % (WARP_SIZE/8)] = get_int_from_int8(bxi->scales, k % (QI6_K/8)); - } -} - -static __device__ __forceinline__ float vec_dot_q6_K_q8_1_mul_mat( - const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, - const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { - (void)x_qh; - - const float * x_dmf = (const float *) x_dm; - const float * y_df = (const float *) y_ds; - - const int8_t * sc = ((const int8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/8]); - - const int index_x = i * (QR6_K*WARP_SIZE + 1) + QR6_K*k; - const int index_y = j * WARP_SIZE + (QR6_K*k) % WARP_SIZE; - return vec_dot_q6_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, x_dmf[i * (WARP_SIZE/QI6_K) + i/QI6_K], &y_df[index_y/QI8_1]); -} - -static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { -#if QK_K == 256 - const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq; - -#if QR2_XXS == 8 - const int ib32 = iqs; - const uint16_t * q2 = bq2->qs + 4*ib32; - const uint8_t * aux8 = (const uint8_t *)q2; - const int8_t * q8 = bq8_1[ib32].qs; - uint32_t aux32 = q2[2] | (q2[3] << 16); - int sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); - const uint8_t signs = ksigns_iq2xs[aux32 & 127]; - for (int j = 0; j < 8; ++j) { - sumi += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - aux32 >>= 7; - } - const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.25f; - return d * sumi; -#else - // iqs is 0...15 - const int ib32 = iqs/2; - const int il = iqs%2; - const uint16_t * q2 = bq2->qs + 4*ib32; - const uint8_t * aux8 = (const uint8_t *)q2; - const uint8_t * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]); - const uint8_t * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]); - const uint32_t aux32 = q2[2] | (q2[3] << 16); - const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * __low2float(bq8_1[ib32].ds) * 0.25f; - const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127]; - const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127]; - const int8_t * q8 = bq8_1[ib32].qs + 16*il; - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < 8; ++j) { - sumi1 += q8[j+0] * grid1[j] * (signs1 & kmask_iq2xs[j] ? -1 : 1); - sumi2 += q8[j+8] * grid2[j] * (signs2 & kmask_iq2xs[j] ? -1 : 1); - } - return d * (sumi1 + sumi2); -#endif -#else - assert(false); - return 0.f; -#endif -} - -static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics -#if QK_K == 256 - const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq; - - const int ib32 = iqs; - const uint16_t * q2 = bq2->qs + 4*ib32; - const int8_t * q8 = bq8_1[ib32].qs; - const uint8_t ls1 = bq2->scales[ib32] & 0xf; - const uint8_t ls2 = bq2->scales[ib32] >> 4; - int sumi1 = 0; - for (int l = 0; l < 2; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511)); - const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9)); - const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]); - const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]); - sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1); - sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1); - q8 += 8; - } - int sumi2 = 0; - for (int l = 2; l < 4; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511)); - const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9)); - const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]); - const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]); - sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2); - sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2); - q8 += 8; - } - const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f; - return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); -#else - (void) ksigns64; - assert(false); - return 0.f; -#endif -#else - (void) ksigns64; - assert(false); - return 0.f; -#endif -} - -// TODO -static __device__ __forceinline__ float vec_dot_iq2_s_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics -#if QK_K == 256 - const block_iq2_s * bq2 = (const block_iq2_s *) vbq; - - const int ib32 = iqs; - const int8_t * q8 = bq8_1[ib32].qs; - const uint8_t * signs = bq2->qs + QK_K/8 + 4*ib32; - const uint8_t ls1 = bq2->scales[ib32] & 0xf; - const uint8_t ls2 = bq2->scales[ib32] >> 4; - int sumi1 = 0; - for (int l = 0; l < 2; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300))); - const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201); - const uint32_t signs1 = __vcmpeq4(((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201); - const int grid_l = __vsub4(grid[0] ^ signs0, signs0); - const int grid_h = __vsub4(grid[1] ^ signs1, signs1); - sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1); - sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1); - q8 += 8; - } - int sumi2 = 0; - for (int l = 2; l < 4; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300))); - const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201); - const uint32_t signs1 = __vcmpeq4(((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201); - const int grid_l = __vsub4(grid[0] ^ signs0, signs0); - const int grid_h = __vsub4(grid[1] ^ signs1, signs1); - sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2); - sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2); - q8 += 8; - } - const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f; - return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); -#else - (void) ksigns64; - assert(false); - return 0.f; -#endif -#else - (void) ksigns64; - assert(false); - return 0.f; -#endif -} - -static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics -#if QK_K == 256 - const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq; - - const int ib32 = iqs; - const uint8_t * q3 = bq2->qs + 8*ib32; - const uint16_t * gas = (const uint16_t *)(bq2->qs + QK_K/4) + 2*ib32; - const int8_t * q8 = bq8_1[ib32].qs; - uint32_t aux32 = gas[0] | (gas[1] << 16); - int sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint32_t * grid1 = iq3xxs_grid + q3[2*l+0]; - const uint32_t * grid2 = iq3xxs_grid + q3[2*l+1]; - const uint32_t * signs = (const uint32_t *)(ksigns64 + (aux32 & 127)); - const int grid_l = __vsub4(grid1[0] ^ signs[0], signs[0]); - const int grid_h = __vsub4(grid2[0] ^ signs[1], signs[1]); - sumi = __dp4a(grid_l, *((int *)q8+0), sumi); - sumi = __dp4a(grid_h, *((int *)q8+1), sumi); - q8 += 8; - aux32 >>= 7; - } - const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.5f; - return d * sumi; -#else - assert(false); - return 0.f; -#endif -#else - assert(false); - return 0.f; -#endif -} - -// TODO: don't use lookup table for signs -static __device__ __forceinline__ float vec_dot_iq3_s_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics -#if QK_K == 256 - const block_iq3_s * bq2 = (const block_iq3_s *) vbq; - - const int ib32 = iqs; - const uint8_t * qs = bq2->qs + 8*ib32; - const int8_t * q8 = bq8_1[ib32].qs; - int sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint32_t * grid1 = iq3s_grid + (qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256)); - const uint32_t * grid2 = iq3s_grid + (qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256)); - uint32_t signs0 = __vcmpeq4(((bq2->signs[4*ib32+l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201); - uint32_t signs1 = __vcmpeq4(((bq2->signs[4*ib32+l] >> 4) * 0x01010101) & 0x08040201, 0x08040201); - const int grid_l = __vsub4(grid1[0] ^ signs0, signs0); - const int grid_h = __vsub4(grid2[0] ^ signs1, signs1); - sumi = __dp4a(grid_l, *((int *)q8+0), sumi); - sumi = __dp4a(grid_h, *((int *)q8+1), sumi); - q8 += 8; - } - const float d = (float)bq2->d * (1 + 2*((bq2->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * __low2float(bq8_1[ib32].ds); - return d * sumi; -#else - assert(false); - return 0.f; -#endif -#else - assert(false); - return 0.f; -#endif -} - -static __device__ __forceinline__ float vec_dot_iq1_s_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { -#if QK_K == 256 - const block_iq1_s * bq1 = (const block_iq1_s *) vbq; - - const int ib32 = iqs; - int sumi = 0; -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const int * q8 = (const int *)bq8_1[ib32].qs; - for (int l = 0; l < 4; ++l) { - const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8))); - int grid0 = grid[0] & 0x0f0f0f0f; - int grid1 = (grid[0] >> 4) & 0x0f0f0f0f; - sumi = __dp4a(q8[2*l+1], grid1, __dp4a(q8[2*l+0], grid0, sumi)); - } -#else - const int8_t * q8 = bq8_1[ib32].qs; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8))); - for (int j = 0; j < 4; ++j) { - sumi += q8[j] * (grid[j] & 0xf) + q8[j+4] * (grid[j] >> 4); - } - q8 += 8; - } -#endif - const float delta = bq1->qh[ib32] & 0x8000 ? -1-IQ1S_DELTA : -1+IQ1S_DELTA; - const float d1q = (float)bq1->d * (2*((bq1->qh[ib32] >> 12) & 7) + 1); - const float d = d1q * __low2float (bq8_1[ib32].ds); - const float m = d1q * __high2float(bq8_1[ib32].ds); - return d * sumi + m * delta; -#else - assert(false); - return 0.f; -#endif -} - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics -static __device__ __forceinline__ void get_int_from_table_16(const uint32_t & q4, const uint8_t * values, - int & val1, int & val2) { - - uint32_t aux32; const uint8_t * q8 = (const uint8_t *)&aux32; - aux32 = q4 & 0x0f0f0f0f; - uint16_t v1 = values[q8[0]] | (values[q8[1]] << 8); - uint16_t v2 = values[q8[2]] | (values[q8[3]] << 8); - val1 = v1 | (v2 << 16); - aux32 = (q4 >> 4) & 0x0f0f0f0f; - v1 = values[q8[0]] | (values[q8[1]] << 8); - v2 = values[q8[2]] | (values[q8[3]] << 8); - val2 = v1 | (v2 << 16); -} -#endif - -static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { - - const block_iq4_nl * bq = (const block_iq4_nl *) vbq; - -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - const uint16_t * q4 = (const uint16_t *)bq->qs + 2*iqs; - const int32_t * q8 = (const int32_t *)bq8_1->qs + iqs; - - const uint8_t * values = (const uint8_t *)kvalues_iq4nl; - - int v1, v2; - int sumi1 = 0, sumi2 = 0; - for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) { - const uint32_t aux = q4[2*l] | (q4[2*l+1] << 16); - get_int_from_table_16(aux, values, v1, v2); - sumi1 = __dp4a(v1, q8[l+0], sumi1); - sumi2 = __dp4a(v2, q8[l+4], sumi2); - } - -#else - const uint8_t * q4 = bq->qs + 4*iqs; - const int8_t * q8 = bq8_1->qs + 4*iqs; - - int sumi1 = 0, sumi2 = 0; - for (int l = 0; l < 4*VDR_Q4_0_Q8_1_MMVQ; ++l) { - sumi1 += q8[l+ 0] * kvalues_iq4nl[q4[l] & 0xf]; - sumi2 += q8[l+16] * kvalues_iq4nl[q4[l] >> 4]; - } -#endif - const float d = (float)bq->d * __low2float(bq8_1->ds); - return d * (sumi1 + sumi2); -} - -static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1( - const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { - -#if QK_K == 256 -#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics - - const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq; - const uint8_t * values = (const uint8_t *)kvalues_iq4nl; - - //// iqs is 0...7 - //const int ib64 = iqs/2; - //const int il = iqs%2; - //const int32_t * q8_1 = (const int *)bq8_1[2*ib64+0].qs + 2*il; - //const int32_t * q8_2 = (const int *)bq8_1[2*ib64+1].qs + 2*il; - //const uint32_t * q4_1 = (const uint32_t *)bq4->qs + 8*ib64 + 2*il; - //const uint32_t * q4_2 = q4_1 + 4; - //const int8_t ls1 = (bq4->scales_l[ib64] & 0xf) | (((bq4->scales_h >> (4*ib64+0)) & 3) << 4); - //const int8_t ls2 = (bq4->scales_l[ib64] >> 4) | (((bq4->scales_h >> (4*ib64+2)) & 3) << 4); - //const float d1 = (float)bq4->d * (ls1 - 32) * __low2float(bq8_1[2*ib64+0].ds); - //const float d2 = (float)bq4->d * (ls2 - 32) * __low2float(bq8_1[2*ib64+1].ds); - //int v1, v2; - //int sumi1 = 0, sumi2 = 0; - //for (int j = 0; j < 2; ++j) { - // get_int_from_table_16(q4_1[j], values, v1, v2); - // sumi1 = __dp4a(v2, q8_1[j+4], __dp4a(v1, q8_1[j+0], sumi1)); - // get_int_from_table_16(q4_2[j], values, v1, v2); - // sumi2 = __dp4a(v2, q8_2[j+4], __dp4a(v1, q8_2[j+0], sumi2)); - //} - //return d1 * sumi1 + d2 * sumi2; - - // iqs is 0...7 - const int ib32 = iqs; - const int32_t * q8 = (const int *)bq8_1[ib32].qs; - const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32; - const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4); - const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds); - int v1, v2; - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < 4; ++j) { - get_int_from_table_16(q4[j], values, v1, v2); - sumi1 = __dp4a(v1, q8[j+0], sumi1); - sumi2 = __dp4a(v2, q8[j+4], sumi2); - } - return d * (sumi1 + sumi2); - - //// iqs is 0...15 - //const int ib32 = iqs/2; - //const int il = iqs%2; - //const int32_t * q8 = (const int *)bq8_1[ib32].qs + 2*il; - //const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32 + 2*il; - //const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4); - //const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds); - //int v1, v2; - //int sumi1 = 0, sumi2 = 0; - //for (int j = 0; j < 2; ++j) { - // get_int_from_table_16(q4[j], values, v1, v2); - // sumi1 = __dp4a(v1, q8[j+0], sumi1); - // sumi2 = __dp4a(v2, q8[j+4], sumi2); - //} - //return d * (sumi1 + sumi2); -#else - assert(false); - return 0.f; -#endif -#else - return vec_dot_iq4_xs_q8_1(vbq, bq8_1, iqs); -#endif -} - -template -static __device__ __forceinline__ void mul_mat_q( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { - - const block_q_t * x = (const block_q_t *) vx; - const block_q8_1 * y = (const block_q8_1 *) vy; - - const int blocks_per_row_x = ncols_x / qk; - const int blocks_per_col_y = nrows_y / QK8_1; - const int blocks_per_warp = WARP_SIZE / qi; - - const int & ncols_dst = ncols_y; - - const int row_dst_0 = blockIdx.x*mmq_y; - const int & row_x_0 = row_dst_0; - - const int col_dst_0 = blockIdx.y*mmq_x; - const int & col_y_0 = col_dst_0; - - int * tile_x_ql = nullptr; - half2 * tile_x_dm = nullptr; - int * tile_x_qh = nullptr; - int * tile_x_sc = nullptr; - - allocate_tiles(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc); - - __shared__ int tile_y_qs[mmq_x * WARP_SIZE]; - __shared__ half2 tile_y_ds[mmq_x * WARP_SIZE/QI8_1]; - - float sum[mmq_y/WARP_SIZE][mmq_x/nwarps] = {{0.0f}}; - - for (int ib0 = 0; ib0 < blocks_per_row_x; ib0 += blocks_per_warp) { - - load_tiles(x + row_x_0*blocks_per_row_x + ib0, tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc, - threadIdx.y, nrows_x-row_x_0-1, threadIdx.x, blocks_per_row_x); - -#pragma unroll - for (int ir = 0; ir < qr; ++ir) { - const int kqs = ir*WARP_SIZE + threadIdx.x; - const int kbxd = kqs / QI8_1; - -#pragma unroll - for (int i = 0; i < mmq_x; i += nwarps) { - const int col_y_eff = min(col_y_0 + threadIdx.y + i, ncols_y-1); // to prevent out-of-bounds memory accesses - - const block_q8_1 * by0 = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + kbxd]; - - const int index_y = (threadIdx.y + i) * WARP_SIZE + kqs % WARP_SIZE; - tile_y_qs[index_y] = get_int_from_int8_aligned(by0->qs, threadIdx.x % QI8_1); - } - -#pragma unroll - for (int ids0 = 0; ids0 < mmq_x; ids0 += nwarps * QI8_1) { - const int ids = (ids0 + threadIdx.y * QI8_1 + threadIdx.x / (WARP_SIZE/QI8_1)) % mmq_x; - const int kby = threadIdx.x % (WARP_SIZE/QI8_1); - const int col_y_eff = min(col_y_0 + ids, ncols_y-1); - - // if the sum is not needed it's faster to transform the scale to f32 ahead of time - const half2 * dsi_src = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + ir*(WARP_SIZE/QI8_1) + kby].ds; - half2 * dsi_dst = &tile_y_ds[ids * (WARP_SIZE/QI8_1) + kby]; - if (need_sum) { - *dsi_dst = *dsi_src; - } else { - float * dfi_dst = (float *) dsi_dst; - *dfi_dst = __low2float(*dsi_src); - } - } - - __syncthreads(); - -// #pragma unroll // unrolling this loop causes too much register pressure - for (int k = ir*WARP_SIZE/qr; k < (ir+1)*WARP_SIZE/qr; k += vdr) { -#pragma unroll - for (int j = 0; j < mmq_x; j += nwarps) { -#pragma unroll - for (int i = 0; i < mmq_y; i += WARP_SIZE) { - sum[i/WARP_SIZE][j/nwarps] += vec_dot( - tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc, tile_y_qs, tile_y_ds, - threadIdx.x + i, threadIdx.y + j, k); +struct ggml_backend_cuda_split_buffer_type_context { + std::array tensor_split; +}; + +struct ggml_backend_cuda_split_buffer_context { + ~ggml_backend_cuda_split_buffer_context() { + for (ggml_tensor_extra_gpu * extra : tensor_extras) { + for (int id = 0; id < GGML_CUDA_MAX_DEVICES; ++id) { + for (int64_t is = 0; is < GGML_CUDA_MAX_STREAMS; ++is) { + if (extra->events[id][is] != nullptr) { + CUDA_CHECK(cudaEventDestroy(extra->events[id][is])); } } + if (extra->data_device[id] != nullptr) { + CUDA_CHECK(cudaFree(extra->data_device[id])); + } } - - __syncthreads(); + delete extra; } } -#pragma unroll - for (int j = 0; j < mmq_x; j += nwarps) { - const int col_dst = col_dst_0 + j + threadIdx.y; + std::vector tensor_extras; +}; - if (col_dst >= ncols_dst) { - return; +GGML_CALL static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backend_buffer_t buffer) { + return GGML_CUDA_NAME "_Split"; + + GGML_UNUSED(buffer); +} + +static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) { + return buffer->iface.get_name == ggml_backend_cuda_split_buffer_get_name; + GGML_UNUSED(ggml_backend_buffer_is_cuda_split); // only used in debug builds currently, avoid unused function warning in release builds +} + +GGML_CALL static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) { + ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context; + delete ctx; +} + +GGML_CALL static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) { + // the pointers are stored in the tensor extras, this is just a dummy address and never dereferenced + return (void *)0x1000; + + GGML_UNUSED(buffer); +} + +GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { + GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported + + ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context; + ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context; + + const int64_t ne0 = tensor->ne[0]; + + ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{}; + ctx->tensor_extras.push_back(extra); + + for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) { + int64_t row_low, row_high; + get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id); + + int64_t nrows_split = row_high - row_low; + if (nrows_split == 0) { + continue; } -#pragma unroll - for (int i = 0; i < mmq_y; i += WARP_SIZE) { - const int row_dst = row_dst_0 + threadIdx.x + i; + size_t size = ggml_nbytes_split(tensor, nrows_split); + const size_t original_size = size; - if (row_dst >= nrows_dst) { - continue; - } - - dst[col_dst*nrows_dst + row_dst] = sum[i/WARP_SIZE][j/nwarps]; + // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses + if (ne0 % MATRIX_ROW_PADDING != 0) { + size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); } + + // FIXME: do not crash if cudaMalloc fails + // currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first + ggml_cuda_set_device(id); + char * buf; + CUDA_CHECK(cudaMalloc(&buf, size)); + + // set padding to 0 to avoid possible NaN values + if (size > original_size) { + CUDA_CHECK(cudaMemset(buf + original_size, 0, size - original_size)); + } + + extra->data_device[id] = buf; + + for (int64_t is = 0; is < GGML_CUDA_MAX_STREAMS; ++is) { + CUDA_CHECK(cudaEventCreateWithFlags(&extra->events[id][is], cudaEventDisableTiming)); + } + } + tensor->extra = extra; +} + +GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { + // split tensors must always be set in their entirety at once + GGML_ASSERT(offset == 0); + GGML_ASSERT(size == ggml_nbytes(tensor)); + + ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context; + + const int64_t ne0 = tensor->ne[0]; + const size_t nb1 = tensor->nb[1]; + ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra; + + for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) { + int64_t row_low, row_high; + get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id); + + int64_t nrows_split = row_high - row_low; + if (nrows_split == 0) { + continue; + } + + const size_t offset_split = row_low*nb1; + size_t size = ggml_nbytes_split(tensor, nrows_split); + const size_t original_size = size; + + // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses + if (ne0 % MATRIX_ROW_PADDING != 0) { + size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); + } + + const char * buf_host = (const char *)data + offset_split; + CUDA_CHECK(cudaMemcpyAsync(extra->data_device[id], buf_host, original_size, cudaMemcpyHostToDevice, cudaStreamPerThread)); + } + + for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) { + CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); } } -#define MMQ_X_Q4_0_RDNA2 64 -#define MMQ_Y_Q4_0_RDNA2 128 -#define NWARPS_Q4_0_RDNA2 8 -#define MMQ_X_Q4_0_RDNA1 64 -#define MMQ_Y_Q4_0_RDNA1 64 -#define NWARPS_Q4_0_RDNA1 8 -#if defined(CUDA_USE_TENSOR_CORES) -#define MMQ_X_Q4_0_AMPERE 4 -#define MMQ_Y_Q4_0_AMPERE 32 -#define NWARPS_Q4_0_AMPERE 4 -#else -#define MMQ_X_Q4_0_AMPERE 64 -#define MMQ_Y_Q4_0_AMPERE 128 -#define NWARPS_Q4_0_AMPERE 4 -#endif -#define MMQ_X_Q4_0_PASCAL 64 -#define MMQ_Y_Q4_0_PASCAL 64 -#define NWARPS_Q4_0_PASCAL 8 +GGML_CALL static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { + // split tensors must always be set in their entirety at once + GGML_ASSERT(offset == 0); + GGML_ASSERT(size == ggml_nbytes(tensor)); -template static __global__ void -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - __launch_bounds__(WARP_SIZE*NWARPS_Q4_0_RDNA2, 2) -#endif // defined(RDNA3) || defined(RDNA2) -#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) - mul_mat_q4_0( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { + ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context; -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - const int mmq_x = MMQ_X_Q4_0_RDNA2; - const int mmq_y = MMQ_Y_Q4_0_RDNA2; - const int nwarps = NWARPS_Q4_0_RDNA2; -#else - const int mmq_x = MMQ_X_Q4_0_RDNA1; - const int mmq_y = MMQ_Y_Q4_0_RDNA1; - const int nwarps = NWARPS_Q4_0_RDNA1; -#endif // defined(RDNA3) || defined(RDNA2) + const int64_t ne0 = tensor->ne[0]; + const size_t nb1 = tensor->nb[1]; + ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra; - mul_mat_q, - load_tiles_q4_0, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) { + int64_t row_low, row_high; + get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id); -#elif __CUDA_ARCH__ >= CC_VOLTA - const int mmq_x = MMQ_X_Q4_0_AMPERE; - const int mmq_y = MMQ_Y_Q4_0_AMPERE; - const int nwarps = NWARPS_Q4_0_AMPERE; + int64_t nrows_split = row_high - row_low; + if (nrows_split == 0) { + continue; + } - mul_mat_q, - load_tiles_q4_0, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + const size_t offset_split = row_low*nb1; + size_t size = ggml_nbytes_split(tensor, nrows_split); + const size_t original_size = size; -#elif __CUDA_ARCH__ >= MIN_CC_DP4A - const int mmq_x = MMQ_X_Q4_0_PASCAL; - const int mmq_y = MMQ_Y_Q4_0_PASCAL; - const int nwarps = NWARPS_Q4_0_PASCAL; + // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses + if (ne0 % MATRIX_ROW_PADDING != 0) { + size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); + } - mul_mat_q, - load_tiles_q4_0, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); -#else - (void) vec_dot_q4_0_q8_1_mul_mat; - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= CC_VOLTA + char * buf_host = (char *)data + offset_split; + CUDA_CHECK(cudaMemcpyAsync(buf_host, extra->data_device[id], original_size, cudaMemcpyDeviceToHost, cudaStreamPerThread)); + } + + for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) { + CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); + } } -#define MMQ_X_Q4_1_RDNA2 64 -#define MMQ_Y_Q4_1_RDNA2 128 -#define NWARPS_Q4_1_RDNA2 8 -#define MMQ_X_Q4_1_RDNA1 64 -#define MMQ_Y_Q4_1_RDNA1 64 -#define NWARPS_Q4_1_RDNA1 8 -#if defined(CUDA_USE_TENSOR_CORES) -#define MMQ_X_Q4_1_AMPERE 4 -#define MMQ_Y_Q4_1_AMPERE 32 -#define NWARPS_Q4_1_AMPERE 4 -#else -#define MMQ_X_Q4_1_AMPERE 64 -#define MMQ_Y_Q4_1_AMPERE 128 -#define NWARPS_Q4_1_AMPERE 4 -#endif -#define MMQ_X_Q4_1_PASCAL 64 -#define MMQ_Y_Q4_1_PASCAL 64 -#define NWARPS_Q4_1_PASCAL 8 - -template static __global__ void -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - __launch_bounds__(WARP_SIZE*NWARPS_Q4_1_RDNA2, 2) -#endif // defined(RDNA3) || defined(RDNA2) -#elif __CUDA_ARCH__ < CC_VOLTA - __launch_bounds__(WARP_SIZE*NWARPS_Q4_1_PASCAL, 2) -#endif // __CUDA_ARCH__ < CC_VOLTA - mul_mat_q4_1( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { - -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - const int mmq_x = MMQ_X_Q4_1_RDNA2; - const int mmq_y = MMQ_Y_Q4_1_RDNA2; - const int nwarps = NWARPS_Q4_1_RDNA2; -#else - const int mmq_x = MMQ_X_Q4_1_RDNA1; - const int mmq_y = MMQ_Y_Q4_1_RDNA1; - const int nwarps = NWARPS_Q4_1_RDNA1; -#endif // defined(RDNA3) || defined(RDNA2) - - mul_mat_q, - load_tiles_q4_1, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= CC_VOLTA - const int mmq_x = MMQ_X_Q4_1_AMPERE; - const int mmq_y = MMQ_Y_Q4_1_AMPERE; - const int nwarps = NWARPS_Q4_1_AMPERE; - - mul_mat_q, - load_tiles_q4_1, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= MIN_CC_DP4A - const int mmq_x = MMQ_X_Q4_1_PASCAL; - const int mmq_y = MMQ_Y_Q4_1_PASCAL; - const int nwarps = NWARPS_Q4_1_PASCAL; - - mul_mat_q, - load_tiles_q4_1, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); -#else - (void) vec_dot_q4_1_q8_1_mul_mat; - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= CC_VOLTA +GGML_CALL static void ggml_backend_cuda_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { + GGML_UNUSED(buffer); + GGML_UNUSED(value); } -#define MMQ_X_Q5_0_RDNA2 64 -#define MMQ_Y_Q5_0_RDNA2 128 -#define NWARPS_Q5_0_RDNA2 8 -#define MMQ_X_Q5_0_RDNA1 64 -#define MMQ_Y_Q5_0_RDNA1 64 -#define NWARPS_Q5_0_RDNA1 8 -#if defined(CUDA_USE_TENSOR_CORES) -#define MMQ_X_Q5_0_AMPERE 4 -#define MMQ_Y_Q5_0_AMPERE 32 -#define NWARPS_Q5_0_AMPERE 4 -#else -#define MMQ_X_Q5_0_AMPERE 128 -#define MMQ_Y_Q5_0_AMPERE 64 -#define NWARPS_Q5_0_AMPERE 4 -#endif -#define MMQ_X_Q5_0_PASCAL 64 -#define MMQ_Y_Q5_0_PASCAL 64 -#define NWARPS_Q5_0_PASCAL 8 +static struct ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = { + /* .get_name = */ ggml_backend_cuda_split_buffer_get_name, + /* .free_buffer = */ ggml_backend_cuda_split_buffer_free_buffer, + /* .get_base = */ ggml_backend_cuda_split_buffer_get_base, + /* .init_tensor = */ ggml_backend_cuda_split_buffer_init_tensor, + /* .set_tensor = */ ggml_backend_cuda_split_buffer_set_tensor, + /* .get_tensor = */ ggml_backend_cuda_split_buffer_get_tensor, + /* .cpy_tensor = */ NULL, + /* .clear = */ ggml_backend_cuda_split_buffer_clear, + /* .reset = */ NULL, +}; -template static __global__ void -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - __launch_bounds__(WARP_SIZE*NWARPS_Q5_0_RDNA2, 2) -#endif // defined(RDNA3) || defined(RDNA2) -#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) - mul_mat_q5_0( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { +// cuda split buffer type -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - const int mmq_x = MMQ_X_Q5_0_RDNA2; - const int mmq_y = MMQ_Y_Q5_0_RDNA2; - const int nwarps = NWARPS_Q5_0_RDNA2; -#else - const int mmq_x = MMQ_X_Q5_0_RDNA1; - const int mmq_y = MMQ_Y_Q5_0_RDNA1; - const int nwarps = NWARPS_Q5_0_RDNA1; -#endif // defined(RDNA3) || defined(RDNA2) +GGML_CALL static const char * ggml_backend_cuda_split_buffer_type_name(ggml_backend_buffer_type_t buft) { + return GGML_CUDA_NAME "_Split"; - mul_mat_q, - load_tiles_q5_0, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= CC_VOLTA - const int mmq_x = MMQ_X_Q5_0_AMPERE; - const int mmq_y = MMQ_Y_Q5_0_AMPERE; - const int nwarps = NWARPS_Q5_0_AMPERE; - - mul_mat_q, - load_tiles_q5_0, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= MIN_CC_DP4A - const int mmq_x = MMQ_X_Q5_0_PASCAL; - const int mmq_y = MMQ_Y_Q5_0_PASCAL; - const int nwarps = NWARPS_Q5_0_PASCAL; - - mul_mat_q, - load_tiles_q5_0, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); -#else - (void) vec_dot_q5_0_q8_1_mul_mat; - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= CC_VOLTA + GGML_UNUSED(buft); } -#define MMQ_X_Q5_1_RDNA2 64 -#define MMQ_Y_Q5_1_RDNA2 128 -#define NWARPS_Q5_1_RDNA2 8 -#define MMQ_X_Q5_1_RDNA1 64 -#define MMQ_Y_Q5_1_RDNA1 64 -#define NWARPS_Q5_1_RDNA1 8 -#if defined(CUDA_USE_TENSOR_CORES) -#define MMQ_X_Q5_1_AMPERE 4 -#define MMQ_Y_Q5_1_AMPERE 32 -#define NWARPS_Q5_1_AMPERE 4 -#else -#define MMQ_X_Q5_1_AMPERE 128 -#define MMQ_Y_Q5_1_AMPERE 64 -#define NWARPS_Q5_1_AMPERE 4 -#endif -#define MMQ_X_Q5_1_PASCAL 64 -#define MMQ_Y_Q5_1_PASCAL 64 -#define NWARPS_Q5_1_PASCAL 8 +GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { + // since we don't know the exact split after rounding, we cannot allocate the device buffers at this point + // instead, we allocate them for each tensor separately in init_tensor + // however, the size still represents the maximum cumulative size of all the device buffers after the tensors are allocated, + // as returned by get_alloc_size. this limit is enforced during tensor allocation by ggml-alloc, so it must be correct. + ggml_backend_cuda_split_buffer_context * ctx = new ggml_backend_cuda_split_buffer_context(); -template static __global__ void -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - __launch_bounds__(WARP_SIZE*NWARPS_Q5_1_RDNA2, 2) -#endif // defined(RDNA3) || defined(RDNA2) -#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -mul_mat_q5_1( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { - -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - const int mmq_x = MMQ_X_Q5_1_RDNA2; - const int mmq_y = MMQ_Y_Q5_1_RDNA2; - const int nwarps = NWARPS_Q5_1_RDNA2; -#else - const int mmq_x = MMQ_X_Q5_1_RDNA1; - const int mmq_y = MMQ_Y_Q5_1_RDNA1; - const int nwarps = NWARPS_Q5_1_RDNA1; -#endif // defined(RDNA3) || defined(RDNA2) - - mul_mat_q, - load_tiles_q5_1, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= CC_VOLTA - const int mmq_x = MMQ_X_Q5_1_AMPERE; - const int mmq_y = MMQ_Y_Q5_1_AMPERE; - const int nwarps = NWARPS_Q5_1_AMPERE; - - mul_mat_q, - load_tiles_q5_1, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= MIN_CC_DP4A - const int mmq_x = MMQ_X_Q5_1_PASCAL; - const int mmq_y = MMQ_Y_Q5_1_PASCAL; - const int nwarps = NWARPS_Q5_1_PASCAL; - - mul_mat_q, - load_tiles_q5_1, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); -#else - (void) vec_dot_q5_1_q8_1_mul_mat; - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= CC_VOLTA + return ggml_backend_buffer_init(buft, ggml_backend_cuda_split_buffer_interface, ctx, size); } -#define MMQ_X_Q8_0_RDNA2 64 -#define MMQ_Y_Q8_0_RDNA2 128 -#define NWARPS_Q8_0_RDNA2 8 -#define MMQ_X_Q8_0_RDNA1 64 -#define MMQ_Y_Q8_0_RDNA1 64 -#define NWARPS_Q8_0_RDNA1 8 -#if defined(CUDA_USE_TENSOR_CORES) -#define MMQ_X_Q8_0_AMPERE 4 -#define MMQ_Y_Q8_0_AMPERE 32 -#define NWARPS_Q8_0_AMPERE 4 -#else -#define MMQ_X_Q8_0_AMPERE 128 -#define MMQ_Y_Q8_0_AMPERE 64 -#define NWARPS_Q8_0_AMPERE 4 -#endif -#define MMQ_X_Q8_0_PASCAL 64 -#define MMQ_Y_Q8_0_PASCAL 64 -#define NWARPS_Q8_0_PASCAL 8 +GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { + return 128; -template static __global__ void -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - __launch_bounds__(WARP_SIZE*NWARPS_Q8_0_RDNA2, 2) -#endif // defined(RDNA3) || defined(RDNA2) -#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) - mul_mat_q8_0( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { - -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - const int mmq_x = MMQ_X_Q8_0_RDNA2; - const int mmq_y = MMQ_Y_Q8_0_RDNA2; - const int nwarps = NWARPS_Q8_0_RDNA2; -#else - const int mmq_x = MMQ_X_Q8_0_RDNA1; - const int mmq_y = MMQ_Y_Q8_0_RDNA1; - const int nwarps = NWARPS_Q8_0_RDNA1; -#endif // defined(RDNA3) || defined(RDNA2) - - mul_mat_q, - load_tiles_q8_0, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= CC_VOLTA - const int mmq_x = MMQ_X_Q8_0_AMPERE; - const int mmq_y = MMQ_Y_Q8_0_AMPERE; - const int nwarps = NWARPS_Q8_0_AMPERE; - - mul_mat_q, - load_tiles_q8_0, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= MIN_CC_DP4A - const int mmq_x = MMQ_X_Q8_0_PASCAL; - const int mmq_y = MMQ_Y_Q8_0_PASCAL; - const int nwarps = NWARPS_Q8_0_PASCAL; - - mul_mat_q, - load_tiles_q8_0, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); -#else - (void) vec_dot_q8_0_q8_1_mul_mat; - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= CC_VOLTA + GGML_UNUSED(buft); } -#define MMQ_X_Q2_K_RDNA2 64 -#define MMQ_Y_Q2_K_RDNA2 128 -#define NWARPS_Q2_K_RDNA2 8 -#define MMQ_X_Q2_K_RDNA1 128 -#define MMQ_Y_Q2_K_RDNA1 32 -#define NWARPS_Q2_K_RDNA1 8 -#if defined(CUDA_USE_TENSOR_CORES) -#define MMQ_X_Q2_K_AMPERE 4 -#define MMQ_Y_Q2_K_AMPERE 32 -#define NWARPS_Q2_K_AMPERE 4 -#else -#define MMQ_X_Q2_K_AMPERE 64 -#define MMQ_Y_Q2_K_AMPERE 128 -#define NWARPS_Q2_K_AMPERE 4 -#endif -#define MMQ_X_Q2_K_PASCAL 64 -#define MMQ_Y_Q2_K_PASCAL 64 -#define NWARPS_Q2_K_PASCAL 8 +GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { + ggml_backend_cuda_split_buffer_type_context * ctx = (ggml_backend_cuda_split_buffer_type_context *)buft->context; -template static __global__ void -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - __launch_bounds__(WARP_SIZE*NWARPS_Q2_K_RDNA2, 2) -#endif // defined(RDNA3) || defined(RDNA2) -#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -mul_mat_q2_K( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { + size_t total_size = 0; -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - const int mmq_x = MMQ_X_Q2_K_RDNA2; - const int mmq_y = MMQ_Y_Q2_K_RDNA2; - const int nwarps = NWARPS_Q2_K_RDNA2; -#else - const int mmq_x = MMQ_X_Q2_K_RDNA1; - const int mmq_y = MMQ_Y_Q2_K_RDNA1; - const int nwarps = NWARPS_Q2_K_RDNA1; -#endif // defined(RDNA3) || defined(RDNA2) + const int64_t ne0 = tensor->ne[0]; - mul_mat_q, - load_tiles_q2_K, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) { + int64_t row_low, row_high; + get_row_split(&row_low, &row_high, tensor, ctx->tensor_split, id); -#elif __CUDA_ARCH__ >= CC_VOLTA - const int mmq_x = MMQ_X_Q2_K_AMPERE; - const int mmq_y = MMQ_Y_Q2_K_AMPERE; - const int nwarps = NWARPS_Q2_K_AMPERE; + int64_t nrows_split = row_high - row_low; + if (nrows_split == 0) { + continue; + } - mul_mat_q, - load_tiles_q2_K, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + total_size += ggml_nbytes_split(tensor, nrows_split); -#elif __CUDA_ARCH__ >= MIN_CC_DP4A - const int mmq_x = MMQ_X_Q2_K_PASCAL; - const int mmq_y = MMQ_Y_Q2_K_PASCAL; - const int nwarps = NWARPS_Q2_K_PASCAL; - - mul_mat_q, - load_tiles_q2_K, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); -#else - (void) vec_dot_q2_K_q8_1_mul_mat; - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= CC_VOLTA -} - -#define MMQ_X_Q3_K_RDNA2 128 -#define MMQ_Y_Q3_K_RDNA2 64 -#define NWARPS_Q3_K_RDNA2 8 -#define MMQ_X_Q3_K_RDNA1 32 -#define MMQ_Y_Q3_K_RDNA1 128 -#define NWARPS_Q3_K_RDNA1 8 -#if defined(CUDA_USE_TENSOR_CORES) -#define MMQ_X_Q3_K_AMPERE 4 -#define MMQ_Y_Q3_K_AMPERE 32 -#define NWARPS_Q3_K_AMPERE 4 -#else -#define MMQ_X_Q3_K_AMPERE 128 -#define MMQ_Y_Q3_K_AMPERE 128 -#define NWARPS_Q3_K_AMPERE 4 -#endif -#define MMQ_X_Q3_K_PASCAL 64 -#define MMQ_Y_Q3_K_PASCAL 64 -#define NWARPS_Q3_K_PASCAL 8 - -template static __global__ void -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - __launch_bounds__(WARP_SIZE*NWARPS_Q3_K_RDNA2, 2) -#endif // defined(RDNA3) || defined(RDNA2) -#elif __CUDA_ARCH__ < CC_VOLTA - __launch_bounds__(WARP_SIZE*NWARPS_Q3_K_PASCAL, 2) -#endif // __CUDA_ARCH__ < CC_VOLTA - mul_mat_q3_K( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { - -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - const int mmq_x = MMQ_X_Q3_K_RDNA2; - const int mmq_y = MMQ_Y_Q3_K_RDNA2; - const int nwarps = NWARPS_Q3_K_RDNA2; -#else - const int mmq_x = MMQ_X_Q3_K_RDNA1; - const int mmq_y = MMQ_Y_Q3_K_RDNA1; - const int nwarps = NWARPS_Q3_K_RDNA1; -#endif // defined(RDNA3) || defined(RDNA2) - - mul_mat_q, - load_tiles_q3_K, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= CC_VOLTA - const int mmq_x = MMQ_X_Q3_K_AMPERE; - const int mmq_y = MMQ_Y_Q3_K_AMPERE; - const int nwarps = NWARPS_Q3_K_AMPERE; - - mul_mat_q, - load_tiles_q3_K, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= MIN_CC_DP4A - const int mmq_x = MMQ_X_Q3_K_PASCAL; - const int mmq_y = MMQ_Y_Q3_K_PASCAL; - const int nwarps = NWARPS_Q3_K_PASCAL; - - mul_mat_q, - load_tiles_q3_K, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); -#else - (void) vec_dot_q3_K_q8_1_mul_mat; - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= CC_VOLTA -} - -#define MMQ_X_Q4_K_RDNA2 64 -#define MMQ_Y_Q4_K_RDNA2 128 -#define NWARPS_Q4_K_RDNA2 8 -#define MMQ_X_Q4_K_RDNA1 32 -#define MMQ_Y_Q4_K_RDNA1 64 -#define NWARPS_Q4_K_RDNA1 8 -#if defined(CUDA_USE_TENSOR_CORES) -#define MMQ_X_Q4_K_AMPERE 4 -#define MMQ_Y_Q4_K_AMPERE 32 -#define NWARPS_Q4_K_AMPERE 4 -#else -#define MMQ_X_Q4_K_AMPERE 64 -#define MMQ_Y_Q4_K_AMPERE 128 -#define NWARPS_Q4_K_AMPERE 4 -#endif -#define MMQ_X_Q4_K_PASCAL 64 -#define MMQ_Y_Q4_K_PASCAL 64 -#define NWARPS_Q4_K_PASCAL 8 - -template static __global__ void -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - __launch_bounds__(WARP_SIZE*NWARPS_Q4_K_RDNA2, 2) -#endif // defined(RDNA3) || defined(RDNA2) -#elif __CUDA_ARCH__ < CC_VOLTA - __launch_bounds__(WARP_SIZE*NWARPS_Q4_K_PASCAL, 2) -#endif // __CUDA_ARCH__ < CC_VOLTA - mul_mat_q4_K( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { - -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - const int mmq_x = MMQ_X_Q4_K_RDNA2; - const int mmq_y = MMQ_Y_Q4_K_RDNA2; - const int nwarps = NWARPS_Q4_K_RDNA2; -#else - const int mmq_x = MMQ_X_Q4_K_RDNA1; - const int mmq_y = MMQ_Y_Q4_K_RDNA1; - const int nwarps = NWARPS_Q4_K_RDNA1; -#endif // defined(RDNA3) || defined(RDNA2) - - mul_mat_q, - load_tiles_q4_K, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= CC_VOLTA - const int mmq_x = MMQ_X_Q4_K_AMPERE; - const int mmq_y = MMQ_Y_Q4_K_AMPERE; - const int nwarps = NWARPS_Q4_K_AMPERE; - - mul_mat_q, - load_tiles_q4_K, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= MIN_CC_DP4A - const int mmq_x = MMQ_X_Q4_K_PASCAL; - const int mmq_y = MMQ_Y_Q4_K_PASCAL; - const int nwarps = NWARPS_Q4_K_PASCAL; - - mul_mat_q, - load_tiles_q4_K, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); -#else - (void) vec_dot_q4_K_q8_1_mul_mat; - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= CC_VOLTA -} - -#define MMQ_X_Q5_K_RDNA2 64 -#define MMQ_Y_Q5_K_RDNA2 128 -#define NWARPS_Q5_K_RDNA2 8 -#define MMQ_X_Q5_K_RDNA1 32 -#define MMQ_Y_Q5_K_RDNA1 64 -#define NWARPS_Q5_K_RDNA1 8 -#if defined(CUDA_USE_TENSOR_CORES) -#define MMQ_X_Q5_K_AMPERE 4 -#define MMQ_Y_Q5_K_AMPERE 32 -#define NWARPS_Q5_K_AMPERE 4 -#else -#define MMQ_X_Q5_K_AMPERE 64 -#define MMQ_Y_Q5_K_AMPERE 128 -#define NWARPS_Q5_K_AMPERE 4 -#endif -#define MMQ_X_Q5_K_PASCAL 64 -#define MMQ_Y_Q5_K_PASCAL 64 -#define NWARPS_Q5_K_PASCAL 8 - -template static __global__ void -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - __launch_bounds__(WARP_SIZE*NWARPS_Q5_K_RDNA2, 2) -#endif // defined(RDNA3) || defined(RDNA2) -#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -mul_mat_q5_K( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { - -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - const int mmq_x = MMQ_X_Q5_K_RDNA2; - const int mmq_y = MMQ_Y_Q5_K_RDNA2; - const int nwarps = NWARPS_Q5_K_RDNA2; -#else - const int mmq_x = MMQ_X_Q5_K_RDNA1; - const int mmq_y = MMQ_Y_Q5_K_RDNA1; - const int nwarps = NWARPS_Q5_K_RDNA1; -#endif // defined(RDNA3) || defined(RDNA2) - - mul_mat_q, - load_tiles_q5_K, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= CC_VOLTA - const int mmq_x = MMQ_X_Q5_K_AMPERE; - const int mmq_y = MMQ_Y_Q5_K_AMPERE; - const int nwarps = NWARPS_Q5_K_AMPERE; - - mul_mat_q, - load_tiles_q5_K, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= MIN_CC_DP4A - const int mmq_x = MMQ_X_Q5_K_PASCAL; - const int mmq_y = MMQ_Y_Q5_K_PASCAL; - const int nwarps = NWARPS_Q5_K_PASCAL; - - mul_mat_q, - load_tiles_q5_K, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); -#else - (void) vec_dot_q5_K_q8_1_mul_mat; - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= CC_VOLTA -} - -#define MMQ_X_Q6_K_RDNA2 64 -#define MMQ_Y_Q6_K_RDNA2 128 -#define NWARPS_Q6_K_RDNA2 8 -#define MMQ_X_Q6_K_RDNA1 32 -#define MMQ_Y_Q6_K_RDNA1 64 -#define NWARPS_Q6_K_RDNA1 8 -#if defined(CUDA_USE_TENSOR_CORES) -#define MMQ_X_Q6_K_AMPERE 4 -#define MMQ_Y_Q6_K_AMPERE 32 -#define NWARPS_Q6_K_AMPERE 4 -#else -#define MMQ_X_Q6_K_AMPERE 64 -#define MMQ_Y_Q6_K_AMPERE 64 -#define NWARPS_Q6_K_AMPERE 4 -#endif -#define MMQ_X_Q6_K_PASCAL 64 -#define MMQ_Y_Q6_K_PASCAL 64 -#define NWARPS_Q6_K_PASCAL 8 - -template static __global__ void -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - __launch_bounds__(WARP_SIZE*NWARPS_Q6_K_RDNA2, 2) -#endif // defined(RDNA3) || defined(RDNA2) -#elif __CUDA_ARCH__ < CC_VOLTA - __launch_bounds__(WARP_SIZE*NWARPS_Q6_K_PASCAL, 2) -#endif // __CUDA_ARCH__ < CC_VOLTA - mul_mat_q6_K( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { - -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) - const int mmq_x = MMQ_X_Q6_K_RDNA2; - const int mmq_y = MMQ_Y_Q6_K_RDNA2; - const int nwarps = NWARPS_Q6_K_RDNA2; -#else - const int mmq_x = MMQ_X_Q6_K_RDNA1; - const int mmq_y = MMQ_Y_Q6_K_RDNA1; - const int nwarps = NWARPS_Q6_K_RDNA1; -#endif // defined(RDNA3) || defined(RDNA2) - - mul_mat_q, - load_tiles_q6_K, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= CC_VOLTA - const int mmq_x = MMQ_X_Q6_K_AMPERE; - const int mmq_y = MMQ_Y_Q6_K_AMPERE; - const int nwarps = NWARPS_Q6_K_AMPERE; - - mul_mat_q, - load_tiles_q6_K, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - -#elif __CUDA_ARCH__ >= MIN_CC_DP4A - const int mmq_x = MMQ_X_Q6_K_PASCAL; - const int mmq_y = MMQ_Y_Q6_K_PASCAL; - const int nwarps = NWARPS_Q6_K_PASCAL; - - mul_mat_q, - load_tiles_q6_K, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); -#else - (void) vec_dot_q6_K_q8_1_mul_mat; - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ >= CC_VOLTA -} - -template -#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) -// tell the compiler to use as many registers as it wants, see nwarps definition below -__launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1) -#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) -static __global__ void mul_mat_vec_q( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) { - -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3)) - constexpr int nwarps = 1; - constexpr int rows_per_cuda_block = 1; -#else - constexpr int nwarps = ncols_y <= 4 ? 4 : 2; - constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2; -#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3) - - const int tid = WARP_SIZE*threadIdx.y + threadIdx.x; - const int row0 = rows_per_cuda_block*blockIdx.x; - const int blocks_per_row_x = ncols_x / qk; - const int blocks_per_col_y = nrows_y / QK8_1; - constexpr int blocks_per_iter = vdr * nwarps*WARP_SIZE / qi; - -// partial sum for each thread - float tmp[ncols_y][rows_per_cuda_block] = {0.0f}; - - const block_q_t * x = (const block_q_t *) vx; - const block_q8_1 * y = (const block_q8_1 *) vy; - - for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) { - const int kby = kbx * (qk/QK8_1); // y block index that aligns with kbx - - // x block quant index when casting the quants to int - const int kqs = vdr * (tid % (qi/vdr)); - -#pragma unroll - for (int j = 0; j < ncols_y; ++j) { -#pragma unroll - for (int i = 0; i < rows_per_cuda_block; ++i) { - tmp[j][i] += vec_dot_q_cuda( - &x[kbx + (row0 + i)*blocks_per_row_x], &y[j*blocks_per_col_y + kby], kqs); - } + // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses + if (ne0 % MATRIX_ROW_PADDING != 0) { + total_size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); } } - __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][WARP_SIZE]; - if (threadIdx.y > 0) { -#pragma unroll - for (int j = 0; j < ncols_y; ++j) { -#pragma unroll - for (int i = 0; i < rows_per_cuda_block; ++i) { - tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i]; - } - } - } - __syncthreads(); - if (threadIdx.y > 0) { - return; - } - - // sum up partial sums and write back result -#pragma unroll - for (int j = 0; j < ncols_y; ++j) { -#pragma unroll - for (int i = 0; i < rows_per_cuda_block; ++i) { -#pragma unroll - for (int l = 0; l < nwarps-1; ++l) { - tmp[j][i] += tmp_shared[l][j][i][threadIdx.x]; - } - tmp[j][i] = warp_reduce_sum(tmp[j][i]); - } - - if (threadIdx.x < rows_per_cuda_block) { - dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x]; - } - } + return total_size; } -template -static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) { - // qk = quantized weights per x block - // qr = number of quantized weights per data value in x block - const int row = blockIdx.x*blockDim.y + threadIdx.y; +GGML_CALL static bool ggml_backend_cuda_split_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) { + return ggml_backend_is_cuda(backend); - if (row >= nrows) { - return; - } + GGML_UNUSED(buft); +} - const int tid = threadIdx.x; +GGML_CALL static bool ggml_backend_cuda_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) { + return false; - const int iter_stride = 2*GGML_CUDA_DMMV_X; - const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter - const int y_offset = qr == 1 ? 1 : qk/2; + GGML_UNUSED(buft); +} -// partial sum for each thread -#ifdef GGML_CUDA_F16 - half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics -#else - float tmp = 0.0f; -#endif // GGML_CUDA_F16 +static ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface = { + /* .get_name = */ ggml_backend_cuda_split_buffer_type_name, + /* .alloc_buffer = */ ggml_backend_cuda_split_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_cuda_split_buffer_type_get_alignment, + /* .get_max_size = */ NULL, // defaults to SIZE_MAX + /* .get_alloc_size = */ ggml_backend_cuda_split_buffer_type_get_alloc_size, + /* .supports_backend = */ ggml_backend_cuda_split_buffer_type_supports_backend, + /* .is_host = */ ggml_backend_cuda_split_buffer_type_is_host, +}; - for (int i = 0; i < ncols; i += iter_stride) { - const int col = i + vals_per_iter*tid; - const int ib = (row*ncols + col)/qk; // x block index - const int iqs = (col%qk)/qr; // x quant index - const int iybs = col - col%qk; // y block start index +GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split) { + static std::mutex mutex; + std::lock_guard lock(mutex); -// processing >2 values per i iter is faster for fast GPUs -#pragma unroll - for (int j = 0; j < vals_per_iter; j += 2) { - // process 2 vals per j iter + static std::map, struct ggml_backend_buffer_type> buft_map; - // dequantize - // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val - dfloat2 v; - dequantize_kernel(vx, ib, iqs + j/qr, v); + std::array tensor_split_arr = {}; - // matrix multiplication - // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2 -#ifdef GGML_CUDA_F16 - tmp += __hmul2(v, { - y[iybs + iqs + j/qr + 0], - y[iybs + iqs + j/qr + y_offset] - }); -#else - tmp += v.x * y[iybs + iqs + j/qr + 0]; - tmp += v.y * y[iybs + iqs + j/qr + y_offset]; -#endif // GGML_CUDA_F16 + bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + GGML_CUDA_MAX_DEVICES, [](float x) { return x == 0.0f; }); + if (all_zero) { + tensor_split_arr = ggml_cuda_info().default_tensor_split; + } else { + float split_sum = 0.0f; + for (int i = 0; i < ggml_backend_cuda_get_device_count(); ++i) { + tensor_split_arr[i] = split_sum; + split_sum += tensor_split[i]; + } + for (int i = 0; i < ggml_backend_cuda_get_device_count(); ++i) { + tensor_split_arr[i] /= split_sum; } } - // sum up partial sums and write back result - tmp = warp_reduce_sum(tmp); - - if (tid == 0) { -#ifdef GGML_CUDA_F16 - dst[row] = tmp.x + tmp.y; -#else - dst[row] = tmp; -#endif // GGML_CUDA_F16 + auto it = buft_map.find(tensor_split_arr); + if (it != buft_map.end()) { + return &it->second; } + + struct ggml_backend_buffer_type buft { + /* .iface = */ ggml_backend_cuda_split_buffer_type_interface, + /* .context = */ new ggml_backend_cuda_split_buffer_type_context{tensor_split_arr}, + }; + + auto result = buft_map.emplace(tensor_split_arr, buft); + return &result.first->second; } +// host buffer type + +GGML_CALL static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_type_t buft) { + return GGML_CUDA_NAME "_Host"; + + GGML_UNUSED(buft); +} + +GGML_CALL static const char * ggml_backend_cuda_host_buffer_name(ggml_backend_buffer_t buffer) { + return GGML_CUDA_NAME "_Host"; + + GGML_UNUSED(buffer); +} + +GGML_CALL static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) { + CUDA_CHECK(cudaFreeHost(buffer->context)); +} + +static void * ggml_cuda_host_malloc(size_t size) { + if (getenv("GGML_CUDA_NO_PINNED") != nullptr) { + return nullptr; + } + + void * ptr = nullptr; + cudaError_t err = cudaMallocHost((void **) &ptr, size); + if (err != cudaSuccess) { + // clear the error + cudaGetLastError(); + fprintf(stderr, "%s: warning: failed to allocate %.2f MiB of pinned memory: %s\n", __func__, + size/1024.0/1024.0, cudaGetErrorString(err)); + return nullptr; + } + + return ptr; +} + +GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { + void * ptr = ggml_cuda_host_malloc(size); + + if (ptr == nullptr) { + // fallback to cpu buffer + return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size); + } + + ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size); + buffer->buft = buft; + buffer->iface.get_name = ggml_backend_cuda_host_buffer_name; + buffer->iface.free_buffer = ggml_backend_cuda_host_buffer_free_buffer; + + return buffer; +} + +GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() { + static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_type_host = { + /* .iface = */ { + /* .get_name = */ ggml_backend_cuda_host_buffer_type_name, + /* .alloc_buffer = */ ggml_backend_cuda_host_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_cpu_buffer_type()->iface.get_alignment, + /* .get_max_size = */ NULL, // defaults to SIZE_MAX + /* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size, + /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend, + /* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host, + }, + /* .context = */ nullptr, + }; + + return &ggml_backend_cuda_buffer_type_host; +} + +//static bool ggml_backend_buffer_is_cuda_host(ggml_backend_buffer_t buffer) { +// return buffer->buft->iface.get_name == ggml_backend_cuda_host_buffer_type_name; +//} + +/// kernels + +typedef void (*ggml_cuda_op_mul_mat_t)( + ggml_backend_cuda_context & ctx, + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, + const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, + const int64_t src1_padded_row_size, cudaStream_t stream); + +#ifndef GGML_CUDA_PEER_MAX_BATCH_SIZE +#define GGML_CUDA_PEER_MAX_BATCH_SIZE 128 +#endif // GGML_CUDA_PEER_MAX_BATCH_SIZE + +#define MUL_MAT_SRC1_COL_STRIDE 128 + static __global__ void mul_mat_p021_f16_f32( const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y) { @@ -5502,1797 +1149,6 @@ static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous } } -static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - float * dsti = (float *) cdsti; - - *dsti = *xi; -} - -static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - half * dsti = (half *) cdsti; - - *dsti = __float2half(*xi); -} - -static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) { - const half * xi = (const half *) cxi; - half * dsti = (half *) cdsti; - - *dsti = *xi; -} - -static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) { - const half * xi = (const half *) cxi; - float * dsti = (float *) cdsti; - - *dsti = *xi; -} - -template -static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, - const int nb12, const int nb13) { - const int64_t i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= ne) { - return; - } - - // determine indices i03/i13, i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor - // then combine those indices with the corresponding byte offsets to get the total offsets - const int64_t i03 = i/(ne00 * ne01 * ne02); - const int64_t i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); - const int64_t i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; - const int64_t i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; - const int64_t x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; - - const int64_t i13 = i/(ne10 * ne11 * ne12); - const int64_t i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); - const int64_t i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; - const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; - const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13; - - cpy_1(cx + x_offset, cdst + dst_offset); -} - -static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - block_q8_0 * dsti = (block_q8_0 *) cdsti; - - float amax = 0.0f; // absolute max - - for (int j = 0; j < QK8_0; j++) { - const float v = xi[j]; - amax = fmaxf(amax, fabsf(v)); - } - - const float d = amax / ((1 << 7) - 1); - const float id = d ? 1.0f/d : 0.0f; - - dsti->d = d; - - for (int j = 0; j < QK8_0; ++j) { - const float x0 = xi[j]*id; - - dsti->qs[j] = roundf(x0); - } -} - -static __device__ void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - block_q4_0 * dsti = (block_q4_0 *) cdsti; - - float amax = 0.0f; - float vmax = 0.0f; - - for (int j = 0; j < QK4_0; ++j) { - const float v = xi[j]; - if (amax < fabsf(v)) { - amax = fabsf(v); - vmax = v; - } - } - - const float d = vmax / -8; - const float id = d ? 1.0f/d : 0.0f; - - dsti->d = d; - - for (int j = 0; j < QK4_0/2; ++j) { - const float x0 = xi[0 + j]*id; - const float x1 = xi[QK4_0/2 + j]*id; - - const uint8_t xi0 = min(15, (int8_t)(x0 + 8.5f)); - const uint8_t xi1 = min(15, (int8_t)(x1 + 8.5f)); - - dsti->qs[j] = xi0; - dsti->qs[j] |= xi1 << 4; - } -} - -static __device__ void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - block_q4_1 * dsti = (block_q4_1 *) cdsti; - - float vmin = FLT_MAX; - float vmax = -FLT_MAX; - - for (int j = 0; j < QK4_1; ++j) { - const float v = xi[j]; - - if (v < vmin) vmin = v; - if (v > vmax) vmax = v; - } - - const float d = (vmax - vmin) / ((1 << 4) - 1); - const float id = d ? 1.0f/d : 0.0f; - - dsti->dm.x = d; - dsti->dm.y = vmin; - - for (int j = 0; j < QK4_1/2; ++j) { - const float x0 = (xi[0 + j] - vmin)*id; - const float x1 = (xi[QK4_1/2 + j] - vmin)*id; - - const uint8_t xi0 = min(15, (int8_t)(x0 + 0.5f)); - const uint8_t xi1 = min(15, (int8_t)(x1 + 0.5f)); - - dsti->qs[j] = xi0; - dsti->qs[j] |= xi1 << 4; - } -} - -template -static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, - const int nb12, const int nb13) { - const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk; - - if (i >= ne) { - return; - } - - const int i03 = i/(ne00 * ne01 * ne02); - const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); - const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; - const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; - const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; - - const int i13 = i/(ne10 * ne11 * ne12); - const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); - const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; - const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; - const int dst_offset = (i10/qk)*nb10 + i11*nb11 + i12*nb12 + i13*nb13; - - cpy_blck(cx + x_offset, cdst + dst_offset); -} - -static __device__ float rope_yarn_ramp(const float low, const float high, const int i0) { - const float y = (i0 / 2 - low) / max(0.001f, high - low); - return 1.0f - min(1.0f, max(0.0f, y)); -} - -struct rope_corr_dims { - float v[4]; -}; - -// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn -// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. -static __device__ void rope_yarn( - float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale, - float * cos_theta, float * sin_theta -) { - // Get n-d rotational scaling corrected for extrapolation - float theta_interp = freq_scale * theta_extrap; - float theta = theta_interp; - if (ext_factor != 0.0f) { - float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor; - theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; - - // Get n-d magnitude scaling corrected for interpolation - mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale); - } - *cos_theta = cosf(theta) * mscale; - *sin_theta = sinf(theta) * mscale; -} - -// rope == RoPE == rotary positional embedding -template -static __global__ void rope( - const T * x, T * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base, - float ext_factor, float attn_factor, rope_corr_dims corr_dims -) { - const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y); - - if (col >= ncols) { - return; - } - - const int row = blockDim.x*blockIdx.x + threadIdx.x; - const int i = row*ncols + col; - const int i2 = row/p_delta_rows; - - const int p = has_pos ? pos[i2] : 0; - const float theta_base = p*powf(freq_base, -float(col)/ncols); - - float cos_theta, sin_theta; - rope_yarn(theta_base, freq_scale, corr_dims, col, ext_factor, attn_factor, &cos_theta, &sin_theta); - - const float x0 = x[i + 0]; - const float x1 = x[i + 1]; - - dst[i + 0] = x0*cos_theta - x1*sin_theta; - dst[i + 1] = x0*sin_theta + x1*cos_theta; -} - -template -static __global__ void rope_neox( - const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, - float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims -) { - const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y); - - if (col >= ncols) { - return; - } - - const int row = blockDim.x*blockIdx.x + threadIdx.x; - const int ib = col / n_dims; - const int ic = col % n_dims; - - if (ib > 0) { - const int i = row*ncols + ib*n_dims + ic; - - dst[i + 0] = x[i + 0]; - dst[i + 1] = x[i + 1]; - - return; - } - - const int i = row*ncols + ib*n_dims + ic/2; - const int i2 = row/p_delta_rows; - - float cur_rot = inv_ndims * ic - ib; - - const int p = has_pos ? pos[i2] : 0; - const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f); - - float cos_theta, sin_theta; - rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta); - - const float x0 = x[i + 0]; - const float x1 = x[i + n_dims/2]; - - dst[i + 0] = x0*cos_theta - x1*sin_theta; - dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta; -} - -static __global__ void rope_glm_f32( - const float * x, float * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base, - int n_ctx -) { - const int col = blockDim.x*blockIdx.x + threadIdx.x; - const int half_n_dims = ncols/4; - - if (col >= half_n_dims) { - return; - } - - const int row = blockDim.y*blockIdx.y + threadIdx.y; - const int i = row*ncols + col; - const int i2 = row/p_delta_rows; - - const float col_theta_scale = powf(freq_base, -2.0f*col/ncols); - // FIXME: this is likely wrong - const int p = pos != nullptr ? pos[i2] : 0; - - const float theta = min(p, n_ctx - 2)*freq_scale*col_theta_scale; - const float sin_theta = sinf(theta); - const float cos_theta = cosf(theta); - - const float x0 = x[i + 0]; - const float x1 = x[i + half_n_dims]; - - dst[i + 0] = x0*cos_theta - x1*sin_theta; - dst[i + half_n_dims] = x0*sin_theta + x1*cos_theta; - - const float block_theta = ((float)max(p - n_ctx - 2, 0))*col_theta_scale; - const float sin_block_theta = sinf(block_theta); - const float cos_block_theta = cosf(block_theta); - - const float x2 = x[i + half_n_dims * 2]; - const float x3 = x[i + half_n_dims * 3]; - - dst[i + half_n_dims * 2] = x2*cos_block_theta - x3*sin_block_theta; - dst[i + half_n_dims * 3] = x2*sin_block_theta + x3*cos_block_theta; -} - -static __global__ void alibi_f32(const float * x, float * dst, const int ncols, const int k_rows, - const int n_heads_log2_floor, const float m0, const float m1) { - const int col = blockDim.x*blockIdx.x + threadIdx.x; - - if (col >= ncols) { - return; - } - - const int row = blockDim.y*blockIdx.y + threadIdx.y; - const int i = row*ncols + col; - - const int k = row/k_rows; - - float m_k; - if (k < n_heads_log2_floor) { - m_k = powf(m0, k + 1); - } else { - m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1); - } - - dst[i] = col * m_k + x[i]; -} - -static __global__ void k_sum_rows_f32(const float * x, float * dst, const int ncols) { - const int row = blockIdx.x; - const int col = threadIdx.x; - - float sum = 0.0f; - for (int i = col; i < ncols; i += blockDim.x) { - sum += x[row * ncols + i]; - } - - sum = warp_reduce_sum(sum); - - if (col == 0) { - dst[row] = sum; - } -} - -template -static inline __device__ void swap(T & a, T & b) { - T tmp = a; - a = b; - b = tmp; -} - -template -static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols) { - // bitonic sort - int col = threadIdx.x; - int row = blockIdx.y; - - if (col >= ncols) return; - - const float * x_row = x + row * ncols; - int * dst_row = dst + row * ncols; - - // initialize indices - if (col < ncols) { - dst_row[col] = col; - } - __syncthreads(); - - for (int k = 2; k <= ncols; k *= 2) { - for (int j = k / 2; j > 0; j /= 2) { - int ixj = col ^ j; - if (ixj > col) { - if ((col & k) == 0) { - if (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]]) { - swap(dst_row[col], dst_row[ixj]); - } - } else { - if (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]]) { - swap(dst_row[col], dst_row[ixj]); - } - } - } - __syncthreads(); - } - } -} - -static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past) { - const int col = blockDim.y*blockIdx.y + threadIdx.y; - const int row = blockDim.x*blockIdx.x + threadIdx.x; - - if (col >= ncols) { - return; - } - - const int i = row*ncols + col; - //dst[i] = col > (n_past + row % rows_per_channel) ? -INFINITY : x[i]; - //dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU - dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX; -} - -template -static __global__ void soft_max_f32(const float * x, const float * mask, const float * pos, float * dst, const int ncols_par, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) { - const int ncols = ncols_template == 0 ? ncols_par : ncols_template; - - const int tid = threadIdx.x; - const int rowx = blockIdx.x; - const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension - - const int block_size = block_size_template == 0 ? blockDim.x : block_size_template; - - const int warp_id = threadIdx.x / WARP_SIZE; - const int lane_id = threadIdx.x % WARP_SIZE; - - float slope = 0.0f; - - // ALiBi - if (max_bias > 0.0f) { - const int h = rowx/nrows_y; // head index - - const float base = h < n_head_log2 ? m0 : m1; - const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; - - slope = powf(base, exp); - } - - extern __shared__ float data_soft_max_f32[]; - float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication - // shared memory buffer to cache values between iterations: - float * vals = vals_smem ? buf_iw + WARP_SIZE : dst + rowx*ncols; - - float max_val = -INFINITY; - -#pragma unroll - for (int col0 = 0; col0 < ncols; col0 += block_size) { - const int col = col0 + tid; - - if (ncols_template == 0 && col >= ncols) { - break; - } - - const int ix = rowx*ncols + col; - const int iy = rowy*ncols + col; - - const float val = x[ix]*scale + (mask ? mask[iy] : 0.0f) + (pos ? slope*pos[col] : 0.0f); - - vals[col] = val; - max_val = max(max_val, val); - } - - // find the max value in the block - max_val = warp_reduce_max(max_val); - if (block_size > WARP_SIZE) { - if (warp_id == 0) { - buf_iw[lane_id] = -INFINITY; - } - __syncthreads(); - - if (lane_id == 0) { - buf_iw[warp_id] = max_val; - } - __syncthreads(); - - max_val = buf_iw[lane_id]; - max_val = warp_reduce_max(max_val); - } - - float tmp = 0.0f; // partial sum - -#pragma unroll - for (int col0 = 0; col0 < ncols; col0 += block_size) { - const int col = col0 + tid; - - if (ncols_template == 0 && col >= ncols) { - break; - } - - const float val = expf(vals[col] - max_val); - tmp += val; - vals[col] = val; - } - - // find the sum of exps in the block - tmp = warp_reduce_sum(tmp); - if (block_size > WARP_SIZE) { - __syncthreads(); - if (warp_id == 0) { - buf_iw[lane_id] = 0.0f; - } - __syncthreads(); - - if (lane_id == 0) { - buf_iw[warp_id] = tmp; - } - __syncthreads(); - - tmp = buf_iw[lane_id]; - tmp = warp_reduce_sum(tmp); - } - - const float inv_sum = 1.0f / tmp; - -#pragma unroll - for (int col0 = 0; col0 < ncols; col0 += block_size) { - const int col = col0 + tid; - - if (ncols_template == 0 && col >= ncols) { - return; - } - - const int idst = rowx*ncols + col; - dst[idst] = vals[col] * inv_sum; - } -} - -static __global__ void scale_f32(const float * x, float * dst, const float scale, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - - dst[i] = scale * x[i]; -} - -static __global__ void clamp_f32(const float * x, float * dst, const float min, const float max, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - - dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]); -} - -template -static __global__ void im2col_kernel( - const float * x, T * dst, int64_t batch_offset, - int64_t offset_delta, int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH, int64_t pelements, int64_t CHW, - int s0, int s1, int p0, int p1, int d0, int d1) { - const int64_t i = threadIdx.x + blockIdx.x * blockDim.x; - if (i >= pelements) { - return; - } - - const int64_t ksize = OW * (KH > 1 ? KW : 1); - const int64_t kx = i / ksize; - const int64_t kd = kx * ksize; - const int64_t ky = (i - kd) / OW; - const int64_t ix = i % OW; - - const int64_t oh = blockIdx.y; - const int64_t batch = blockIdx.z / IC; - const int64_t ic = blockIdx.z % IC; - - const int64_t iiw = ix * s0 + kx * d0 - p0; - const int64_t iih = oh * s1 + ky * d1 - p1; - - const int64_t offset_dst = - ((batch * OH + oh) * OW + ix) * CHW + - (ic * (KW * KH) + ky * KW + kx); - - if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { - dst[offset_dst] = 0.0f; - } else { - const int64_t offset_src = ic * offset_delta + batch * batch_offset; - dst[offset_dst] = x[offset_src + iih * IW + iiw]; - } -} - -template -static __global__ void pool2d_nchw_kernel( - const int ih, const int iw, const int oh, const int ow, - const int kh, const int kw, const int sh, const int sw, - const int ph, const int pw, const int parallel_elements, - const Ti* src, To* dst, const enum ggml_op_pool op) { - int idx = threadIdx.x + blockIdx.x * blockDim.x; - if (idx >= parallel_elements) { - return; - } - - const int I_HW = ih * iw; - const int O_HW = oh * ow; - const int nc = idx / O_HW; - const int cur_oh = idx % O_HW / ow; - const int cur_ow = idx % O_HW % ow; - const Ti* i_ptr = src + nc * I_HW; - To* o_ptr = dst + nc * O_HW; - const int start_h = cur_oh * sh - ph; - const int bh = max(0, start_h); - const int eh = min(ih, start_h + kh); - const int start_w = cur_ow * sw - pw; - const int bw = max(0, start_w); - const int ew = min(iw, start_w + kw); - const To scale = 1. / (kh * kw); - To res = 0; - - switch (op) { - case GGML_OP_POOL_AVG: res = 0; break; - case GGML_OP_POOL_MAX: res = -FLT_MAX; break; - } - - for (int i = bh; i < eh; i += 1) { - for (int j = bw; j < ew; j += 1) { - #if __CUDA_ARCH__ >= 350 - Ti cur = __ldg(i_ptr + i * iw + j); - #else - Ti cur = i_ptr[i * iw + j]; - #endif - switch (op) { - case GGML_OP_POOL_AVG: res += cur * scale; break; - case GGML_OP_POOL_MAX: res = max(res, (To)cur); break; - } - } - } - o_ptr[cur_oh * ow + cur_ow] = res; -} - -template -static void get_rows_cuda(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) { - - GGML_TENSOR_BINARY_OP_LOCALS - - const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1); - const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE); - const dim3 block_nums(block_num_x, ne10, ne11*ne12); - - // strides in elements - //const size_t s0 = nb0 / ggml_element_size(dst); - const size_t s1 = nb1 / ggml_element_size(dst); - const size_t s2 = nb2 / ggml_element_size(dst); - const size_t s3 = nb3 / ggml_element_size(dst); - - const size_t s10 = nb10 / ggml_element_size(src1); - const size_t s11 = nb11 / ggml_element_size(src1); - const size_t s12 = nb12 / ggml_element_size(src1); - //const size_t s13 = nb13 / ggml_element_size(src1); - - GGML_ASSERT(ne00 % 2 == 0); - - k_get_rows<<>>( - src0_dd, src1_dd, dst_dd, - ne00, /*ne01, ne02, ne03,*/ - /*ne10, ne11,*/ ne12, /*ne13,*/ - /* s0,*/ s1, s2, s3, - /* nb00,*/ nb01, nb02, nb03, - s10, s11, s12/*, s13*/); - - (void) dst; -} - -template -static void get_rows_cuda_float(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) { - - GGML_TENSOR_BINARY_OP_LOCALS - - const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1); - const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE; - const dim3 block_nums(block_num_x, ne10, ne11*ne12); - - // strides in elements - //const size_t s0 = nb0 / ggml_element_size(dst); - const size_t s1 = nb1 / ggml_element_size(dst); - const size_t s2 = nb2 / ggml_element_size(dst); - const size_t s3 = nb3 / ggml_element_size(dst); - - const size_t s10 = nb10 / ggml_element_size(src1); - const size_t s11 = nb11 / ggml_element_size(src1); - const size_t s12 = nb12 / ggml_element_size(src1); - //const size_t s13 = nb13 / ggml_element_size(src1); - - k_get_rows_float<<>>( - src0_dd, src1_dd, dst_dd, - ne00, /*ne01, ne02, ne03,*/ - /*ne10, ne11,*/ ne12, /*ne13,*/ - /* s0,*/ s1, s2, s3, - /* nb00,*/ nb01, nb02, nb03, - s10, s11, s12/*, s13*/); - - (void) dst; -} - -template -struct bin_bcast_cuda { - template - void operator()(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, - const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd, - cudaStream_t stream) { - - GGML_TENSOR_BINARY_OP_LOCALS - - int nr0 = ne10/ne0; - int nr1 = ne11/ne1; - int nr2 = ne12/ne2; - int nr3 = ne13/ne3; - - int nr[4] = { nr0, nr1, nr2, nr3 }; - - // collapse dimensions until first broadcast dimension - int64_t cne0[] = {ne0, ne1, ne2, ne3}; - int64_t cne1[] = {ne10, ne11, ne12, ne13}; - size_t cnb0[] = {nb0, nb1, nb2, nb3}; - size_t cnb1[] = {nb10, nb11, nb12, nb13}; - auto collapse = [](int64_t cne[]) { - cne[0] *= cne[1]; - cne[1] = cne[2]; - cne[2] = cne[3]; - cne[3] = 1; - }; - - auto collapse_nb = [](size_t cnb[], const int64_t cne[]) { - cnb[1] *= cne[1]; - cnb[2] *= cne[2]; - cnb[3] *= cne[3]; - }; - - for (int i = 0; i < 4; i++) { - if (nr[i] != 1) { - break; - } - if (i > 0) { - collapse_nb(cnb0, cne0); - collapse_nb(cnb1, cne1); - collapse(cne0); - collapse(cne1); - } - } - { - int64_t ne0 = cne0[0]; - int64_t ne1 = cne0[1]; - int64_t ne2 = cne0[2]; - int64_t ne3 = cne0[3]; - - int64_t ne10 = cne1[0]; - int64_t ne11 = cne1[1]; - int64_t ne12 = cne1[2]; - int64_t ne13 = cne1[3]; - - size_t nb0 = cnb0[0]; - size_t nb1 = cnb0[1]; - size_t nb2 = cnb0[2]; - size_t nb3 = cnb0[3]; - - size_t nb10 = cnb1[0]; - size_t nb11 = cnb1[1]; - size_t nb12 = cnb1[2]; - size_t nb13 = cnb1[3]; - - size_t s0 = nb0 / sizeof(dst_t); - size_t s1 = nb1 / sizeof(dst_t); - size_t s2 = nb2 / sizeof(dst_t); - size_t s3 = nb3 / sizeof(dst_t); - - size_t s10 = nb10 / sizeof(src1_t); - size_t s11 = nb11 / sizeof(src1_t); - size_t s12 = nb12 / sizeof(src1_t); - size_t s13 = nb13 / sizeof(src1_t); - - GGML_ASSERT(s0 == 1); - GGML_ASSERT(s10 == 1); - - const int block_size = 128; - - int64_t hne0 = std::max(ne0/2LL, 1LL); - - dim3 block_dims; - block_dims.x = std::min(hne0, block_size); - block_dims.y = std::min(ne1, block_size / block_dims.x); - block_dims.z = std::min(std::min(ne2*ne3, block_size / block_dims.x / block_dims.y), 64U); - - dim3 block_nums( - (hne0 + block_dims.x - 1) / block_dims.x, - (ne1 + block_dims.y - 1) / block_dims.y, - (ne2*ne3 + block_dims.z - 1) / block_dims.z - ); - - if (block_nums.z > 65535) { - // this is the maximum number of blocks in z direction, fallback to 1D grid kernel - int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size; - k_bin_bcast_unravel<<>>( - src0_dd, src1_dd, dst_dd, - ne0, ne1, ne2, ne3, - ne10, ne11, ne12, ne13, - /* s0, */ s1, s2, s3, - /* s10, */ s11, s12, s13); - } else { - k_bin_bcast<<>>( - src0_dd, src1_dd, dst_dd, - ne0, ne1, ne2, ne3, - ne10, ne11, ne12, ne13, - /* s0, */ s1, s2, s3, - /* s10, */ s11, s12, s13); - } - } - } -}; - -static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements, - const int ne10, const int ne11, const int ne12, - const int nb1, const int nb2, const int offset, cudaStream_t stream) { - int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE; - acc_f32<<>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset); -} - -static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE; - gelu_f32<<>>(x, dst, k); -} - -static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_SILU_BLOCK_SIZE - 1) / CUDA_SILU_BLOCK_SIZE; - silu_f32<<>>(x, dst, k); -} - -static void gelu_quick_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE; - gelu_quick_f32<<>>(x, dst, k); -} - -static void tanh_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_TANH_BLOCK_SIZE - 1) / CUDA_TANH_BLOCK_SIZE; - tanh_f32<<>>(x, dst, k); -} - -static void relu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE; - relu_f32<<>>(x, dst, k); -} - -static void hardsigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_HARDSIGMOID_BLOCK_SIZE - 1) / CUDA_HARDSIGMOID_BLOCK_SIZE; - hardsigmoid_f32<<>>(x, dst, k); -} - -static void hardswish_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_HARDSWISH_BLOCK_SIZE - 1) / CUDA_HARDSWISH_BLOCK_SIZE; - hardswish_f32<<>>(x, dst, k); -} - -static void leaky_relu_f32_cuda(const float * x, float * dst, const int k, const float negative_slope, cudaStream_t stream) { - const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE; - leaky_relu_f32<<>>(x, dst, k, negative_slope); -} - -static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_SQR_BLOCK_SIZE - 1) / CUDA_SQR_BLOCK_SIZE; - sqr_f32<<>>(x, dst, k); -} - -static void norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) { - GGML_ASSERT(ncols % WARP_SIZE == 0); - if (ncols < 1024) { - const dim3 block_dims(WARP_SIZE, 1, 1); - norm_f32<<>>(x, dst, ncols, eps); - } else { - const dim3 block_dims(1024, 1, 1); - norm_f32<1024><<>>(x, dst, ncols, eps); - } -} - -static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const int group_size, const int ne_elements, cudaStream_t stream) { - static const float eps = 1e-6f; - if (group_size < 1024) { - const dim3 block_dims(WARP_SIZE, 1, 1); - group_norm_f32<<>>(x, dst, group_size, ne_elements, eps); - } else { - const dim3 block_dims(1024, 1, 1); - group_norm_f32<1024><<>>(x, dst, group_size, ne_elements, eps); - } -} - -static void concat_f32_cuda(const float * x, const float * y, float * dst, const int ne0, int ne1, int ne2, int ne02, cudaStream_t stream) { - int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE; - dim3 gridDim(num_blocks, ne1, ne2); - concat_f32<<>>(x, y, dst, ne0, ne02); -} - -static void upscale_f32_cuda(const float * x, float * dst, const int ne00, const int ne01, const int ne02, const int ne03, - const int scale_factor, cudaStream_t stream) { - int ne0 = (ne00 * scale_factor); - int num_blocks = (ne0 + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE; - dim3 gridDim(num_blocks, (ne01 * scale_factor), ne02*ne03); - upscale_f32<<>>(x, dst, ne00, ne00 * ne01, scale_factor); -} - -static void pad_f32_cuda(const float * x, float * dst, - const int ne00, const int ne01, const int ne02, const int ne03, - const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) { - int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE; - dim3 gridDim(num_blocks, ne1, ne2*ne3); - pad_f32<<>>(x, dst, ne0, ne00, ne01, ne02, ne03); -} - -static void arange_f32_cuda(float * dst, const int ne0, const float start, const float step, cudaStream_t stream) { - int num_blocks = (ne0 + CUDA_ARANGE_BLOCK_SIZE - 1) / CUDA_ARANGE_BLOCK_SIZE; - arange_f32<<>>(dst, ne0, start, step); -} - -static void timestep_embedding_f32_cuda(const float * x, float * dst, const int ne00, const int nb1, - const int dim, const int max_period, cudaStream_t stream) { - int half_ceil = (dim + 1) / 2; - int num_blocks = (half_ceil + CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE; - dim3 gridDim(num_blocks, ne00, 1); - timestep_embedding_f32<<>>(x, dst, nb1, dim, max_period); -} - -static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) { - GGML_ASSERT(ncols % WARP_SIZE == 0); - if (ncols < 1024) { - const dim3 block_dims(WARP_SIZE, 1, 1); - rms_norm_f32<<>>(x, dst, ncols, eps); - } else { - const dim3 block_dims(1024, 1, 1); - rms_norm_f32<1024><<>>(x, dst, ncols, eps); - } -} - -static void quantize_row_q8_1_cuda(const float * x, void * vy, const int kx, const int ky, const int kx_padded, cudaStream_t stream) { - const int block_num_x = (kx_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE; - const dim3 num_blocks(block_num_x, ky, 1); - const dim3 block_size(CUDA_DEQUANTIZE_BLOCK_SIZE, 1, 1); - quantize_q8_1<<>>(x, vy, kx, kx_padded); -} - -template -static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) { - const int num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE); - dequantize_block<<>>(vx, y, k); -} - -static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * __restrict__ y, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_Q8_0_NE_ALIGN - 1) / CUDA_Q8_0_NE_ALIGN; - if (k % CUDA_Q8_0_NE_ALIGN == 0) { - const bool need_check = false; - dequantize_block_q8_0_f16<<>>(vx, y, k); - } else { - const bool need_check = true; - dequantize_block_q8_0_f16<<>>(vx, y, k); - } -} - -template -static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb = k / QK_K; -#if QK_K == 256 - dequantize_block_q2_K<<>>(vx, y); -#else - dequantize_block_q2_K<<>>(vx, y); -#endif -} - -template -static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb = k / QK_K; -#if QK_K == 256 - dequantize_block_q3_K<<>>(vx, y); -#else - dequantize_block_q3_K<<>>(vx, y); -#endif -} - -template -static void dequantize_row_q4_0_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb32 = k / 32; - const int nb = (k + 255) / 256; - dequantize_block_q4_0<<>>(vx, y, nb32); -} - -template -static void dequantize_row_q4_1_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb32 = k / 32; - const int nb = (k + 255) / 256; - dequantize_block_q4_1<<>>(vx, y, nb32); -} - -template -static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb = k / QK_K; - dequantize_block_q4_K<<>>(vx, y); -} - -template -static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb = k / QK_K; -#if QK_K == 256 - dequantize_block_q5_K<<>>(vx, y); -#else - dequantize_block_q5_K<<>>(vx, y); -#endif -} - -template -static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb = k / QK_K; -#if QK_K == 256 - dequantize_block_q6_K<<>>(vx, y); -#else - dequantize_block_q6_K<<>>(vx, y); -#endif -} - -template -static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb = k / QK_K; - dequantize_block_iq2_xxs<<>>(vx, y); -} - -template -static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb = k / QK_K; - dequantize_block_iq2_xs<<>>(vx, y); -} - -template -static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb = k / QK_K; - dequantize_block_iq2_s<<>>(vx, y); -} - -template -static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb = k / QK_K; - dequantize_block_iq3_xxs<<>>(vx, y); -} - -template -static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb = k / QK_K; - dequantize_block_iq3_s<<>>(vx, y); -} - -template -static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb = k / QK_K; - dequantize_block_iq1_s<<>>(vx, y); -} - -template -static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb = (k + QK_K - 1) / QK_K; - dequantize_block_iq4_nl<<>>(vx, y); -} - -template -static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { - const int nb = (k + QK_K - 1) / QK_K; -#if QK_K == 64 - dequantize_block_iq4_nl<<>>(vx, y); -#else - dequantize_block_iq4_xs<<>>(vx, y); -#endif -} - -template -static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE; - convert_unary<<>>(vx, y, k); -} - -static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) { - int id; - switch (type) { - case GGML_TYPE_Q4_0: - return dequantize_row_q4_0_cuda; - case GGML_TYPE_Q4_1: - return dequantize_row_q4_1_cuda; - case GGML_TYPE_Q5_0: - return dequantize_block_cuda; - case GGML_TYPE_Q5_1: - return dequantize_block_cuda; - case GGML_TYPE_Q8_0: - CUDA_CHECK(cudaGetDevice(&id)); - if (g_device_caps[id].cc >= CC_PASCAL) { - return dequantize_block_q8_0_f16_cuda; - } - return dequantize_block_cuda; - case GGML_TYPE_Q2_K: - return dequantize_row_q2_K_cuda; - case GGML_TYPE_Q3_K: - return dequantize_row_q3_K_cuda; - case GGML_TYPE_Q4_K: - return dequantize_row_q4_K_cuda; - case GGML_TYPE_Q5_K: - return dequantize_row_q5_K_cuda; - case GGML_TYPE_Q6_K: - return dequantize_row_q6_K_cuda; - case GGML_TYPE_IQ2_XXS: - return dequantize_row_iq2_xxs_cuda; - case GGML_TYPE_IQ2_XS: - return dequantize_row_iq2_xs_cuda; - case GGML_TYPE_IQ2_S: - return dequantize_row_iq2_s_cuda; - case GGML_TYPE_IQ3_XXS: - return dequantize_row_iq3_xxs_cuda; - case GGML_TYPE_IQ1_S: - return dequantize_row_iq1_s_cuda; - case GGML_TYPE_IQ4_NL: - return dequantize_row_iq4_nl_cuda; - case GGML_TYPE_IQ4_XS: - return dequantize_row_iq4_xs_cuda; - case GGML_TYPE_IQ3_S: - return dequantize_row_iq3_s_cuda; - case GGML_TYPE_F32: - return convert_unary_cuda; - default: - return nullptr; - } -} - -static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) { - switch (type) { - case GGML_TYPE_Q4_0: - return dequantize_row_q4_0_cuda; - case GGML_TYPE_Q4_1: - return dequantize_row_q4_1_cuda; - case GGML_TYPE_Q5_0: - return dequantize_block_cuda; - case GGML_TYPE_Q5_1: - return dequantize_block_cuda; - case GGML_TYPE_Q8_0: - return dequantize_block_cuda; - case GGML_TYPE_Q2_K: - return dequantize_row_q2_K_cuda; - case GGML_TYPE_Q3_K: - return dequantize_row_q3_K_cuda; - case GGML_TYPE_Q4_K: - return dequantize_row_q4_K_cuda; - case GGML_TYPE_Q5_K: - return dequantize_row_q5_K_cuda; - case GGML_TYPE_Q6_K: - return dequantize_row_q6_K_cuda; - case GGML_TYPE_IQ2_XXS: - return dequantize_row_iq2_xxs_cuda; - case GGML_TYPE_IQ2_XS: - return dequantize_row_iq2_xs_cuda; - case GGML_TYPE_IQ2_S: - return dequantize_row_iq2_s_cuda; - case GGML_TYPE_IQ3_XXS: - return dequantize_row_iq3_xxs_cuda; - case GGML_TYPE_IQ1_S: - return dequantize_row_iq1_s_cuda; - case GGML_TYPE_IQ4_NL: - return dequantize_row_iq4_nl_cuda; - case GGML_TYPE_IQ4_XS: - return dequantize_row_iq4_xs_cuda; - case GGML_TYPE_IQ3_S: - return dequantize_row_iq3_s_cuda; - case GGML_TYPE_F16: - return convert_unary_cuda; - default: - return nullptr; - } -} - -static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { - GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); - const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead - const dim3 block_nums(block_num_y, 1, 1); - const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); - dequantize_mul_mat_vec - <<>>(vx, y, dst, ncols, nrows); -} - -static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { - GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); - const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(block_num_y, 1, 1); - const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); - dequantize_mul_mat_vec - <<>>(vx, y, dst, ncols, nrows); -} - -static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { - GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); - const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(block_num_y, 1, 1); - const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); - dequantize_mul_mat_vec - <<>>(vx, y, dst, ncols, nrows); -} - -static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { - GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); - const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(block_num_y, 1, 1); - const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); - dequantize_mul_mat_vec - <<>>(vx, y, dst, ncols, nrows); -} - -static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { - GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); - const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(block_num_y, 1, 1); - const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); - dequantize_mul_mat_vec - <<>>(vx, y, dst, ncols, nrows); -} - -static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2 - const int block_num_y = (nrows + ny - 1) / ny; - const dim3 block_nums(block_num_y, 1, 1); - const dim3 block_dims(32, ny, 1); - dequantize_mul_mat_vec_q2_k<<>>(vx, y, dst, ncols, nrows); -} - -static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int ny = 2 / K_QUANTS_PER_ITERATION; - const int block_num_y = (nrows + ny - 1) / ny; - const dim3 block_nums(block_num_y, 1, 1); - const dim3 block_dims(32, ny, 1); - dequantize_mul_mat_vec_q3_k<<>>(vx, y, dst, ncols, nrows); -} - -static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int ny = 2 / K_QUANTS_PER_ITERATION; - const int block_num_y = (nrows + ny - 1) / ny; - const dim3 block_nums(block_num_y, 1, 1); - const dim3 block_dims(32, ny, 1); - dequantize_mul_mat_vec_q4_k<<>>(vx, y, dst, ncols, nrows); -} - -static void dequantize_mul_mat_vec_q5_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { - GGML_ASSERT(ncols % QK_K == 0); - const dim3 block_dims(32, 1, 1); - dequantize_mul_mat_vec_q5_k<<>>(vx, y, dst, ncols); -} - -static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int ny = 2 / K_QUANTS_PER_ITERATION; - const int block_num_y = (nrows + ny - 1) / ny; - const dim3 block_nums(block_num_y, 1, 1); - const dim3 block_dims(32, ny, 1); - dequantize_mul_mat_vec_q6_k<<>>(vx, y, dst, ncols, nrows); -} - -static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { - GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); - const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(block_num_y, 1, 1); - const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); - dequantize_mul_mat_vec<1, 1, convert_f16> - <<>>(vx, y, dst, ncols, nrows); -} - -template -static void mul_mat_vec_q_cuda( - const void * vx, const void * vy, float * dst, - const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { - - GGML_ASSERT(ncols_x % qk == 0); - GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE); - - int id; - CUDA_CHECK(cudaGetDevice(&id)); - - int64_t nwarps = 1; - int64_t rows_per_cuda_block = 1; - - if (g_device_caps[id].cc < CC_RDNA2) { // NVIDIA and AMD older than RDNA2 - switch(ncols_y) { - case 1: - nwarps = 4; - rows_per_cuda_block = 1; - break; - case 2: - case 3: - case 4: - nwarps = 4; - rows_per_cuda_block = 2; - break; - case 5: - case 6: - case 7: - case 8: - nwarps = 2; - rows_per_cuda_block = 2; - break; - default: - GGML_ASSERT(false); - break; - } - } - const int64_t nblocks = (nrows_x + rows_per_cuda_block - 1) / rows_per_cuda_block; - const dim3 block_nums(nblocks, 1, 1); - const dim3 block_dims(WARP_SIZE, nwarps, 1); - - switch (ncols_y) { - case 1: - mul_mat_vec_q<1, qk, qi, block_q_t, vdr, vec_dot> - <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); - break; - case 2: - mul_mat_vec_q<2, qk, qi, block_q_t, vdr, vec_dot> - <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); - break; - case 3: - mul_mat_vec_q<3, qk, qi, block_q_t, vdr, vec_dot> - <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); - break; - case 4: - mul_mat_vec_q<4, qk, qi, block_q_t, vdr, vec_dot> - <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); - break; - case 5: - mul_mat_vec_q<5, qk, qi, block_q_t, vdr, vec_dot> - <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); - break; - case 6: - mul_mat_vec_q<6, qk, qi, block_q_t, vdr, vec_dot> - <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); - break; - case 7: - mul_mat_vec_q<7, qk, qi, block_q_t, vdr, vec_dot> - <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); - break; - case 8: - mul_mat_vec_q<8, qk, qi, block_q_t, vdr, vec_dot> - <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); - break; - default: - GGML_ASSERT(false); - break; - } -} - -static void ggml_mul_mat_q4_0_q8_1_cuda( - const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, - const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { - - int id; - CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_device_caps[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_RDNA2) { - mmq_x = MMQ_X_Q4_0_RDNA2; - mmq_y = MMQ_Y_Q4_0_RDNA2; - nwarps = NWARPS_Q4_0_RDNA2; - } else if (compute_capability >= CC_OFFSET_AMD) { - mmq_x = MMQ_X_Q4_0_RDNA1; - mmq_y = MMQ_Y_Q4_0_RDNA1; - nwarps = NWARPS_Q4_0_RDNA1; - } else if (compute_capability >= CC_VOLTA) { - mmq_x = MMQ_X_Q4_0_AMPERE; - mmq_y = MMQ_Y_Q4_0_AMPERE; - nwarps = NWARPS_Q4_0_AMPERE; - } else if (compute_capability >= MIN_CC_DP4A) { - mmq_x = MMQ_X_Q4_0_PASCAL; - mmq_y = MMQ_Y_Q4_0_PASCAL; - nwarps = NWARPS_Q4_0_PASCAL; - } else { - GGML_ASSERT(false); - } - - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const dim3 block_nums(block_num_x, block_num_y, 1); - const dim3 block_dims(WARP_SIZE, nwarps, 1); - - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - mul_mat_q4_0<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } else { - const bool need_check = true; - mul_mat_q4_0<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } -} - -static void ggml_mul_mat_q4_1_q8_1_cuda( - const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, - const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { - - int id; - CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_device_caps[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_RDNA2) { - mmq_x = MMQ_X_Q4_1_RDNA2; - mmq_y = MMQ_Y_Q4_1_RDNA2; - nwarps = NWARPS_Q4_1_RDNA2; - } else if (compute_capability >= CC_OFFSET_AMD) { - mmq_x = MMQ_X_Q4_1_RDNA1; - mmq_y = MMQ_Y_Q4_1_RDNA1; - nwarps = NWARPS_Q4_1_RDNA1; - } else if (compute_capability >= CC_VOLTA) { - mmq_x = MMQ_X_Q4_1_AMPERE; - mmq_y = MMQ_Y_Q4_1_AMPERE; - nwarps = NWARPS_Q4_1_AMPERE; - } else if (compute_capability >= MIN_CC_DP4A) { - mmq_x = MMQ_X_Q4_1_PASCAL; - mmq_y = MMQ_Y_Q4_1_PASCAL; - nwarps = NWARPS_Q4_1_PASCAL; - } else { - GGML_ASSERT(false); - } - - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const dim3 block_nums(block_num_x, block_num_y, 1); - const dim3 block_dims(WARP_SIZE, nwarps, 1); - - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - mul_mat_q4_1<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } else { - const bool need_check = true; - mul_mat_q4_1<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } -} - -static void ggml_mul_mat_q5_0_q8_1_cuda( - const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, - const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { - - int id; - CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_device_caps[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_RDNA2) { - mmq_x = MMQ_X_Q5_0_RDNA2; - mmq_y = MMQ_Y_Q5_0_RDNA2; - nwarps = NWARPS_Q5_0_RDNA2; - } else if (compute_capability >= CC_OFFSET_AMD) { - mmq_x = MMQ_X_Q5_0_RDNA1; - mmq_y = MMQ_Y_Q5_0_RDNA1; - nwarps = NWARPS_Q5_0_RDNA1; - } else if (compute_capability >= CC_VOLTA) { - mmq_x = MMQ_X_Q5_0_AMPERE; - mmq_y = MMQ_Y_Q5_0_AMPERE; - nwarps = NWARPS_Q5_0_AMPERE; - } else if (compute_capability >= MIN_CC_DP4A) { - mmq_x = MMQ_X_Q5_0_PASCAL; - mmq_y = MMQ_Y_Q5_0_PASCAL; - nwarps = NWARPS_Q5_0_PASCAL; - } else { - GGML_ASSERT(false); - } - - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const dim3 block_nums(block_num_x, block_num_y, 1); - const dim3 block_dims(WARP_SIZE, nwarps, 1); - - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - mul_mat_q5_0<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } else { - const bool need_check = true; - mul_mat_q5_0<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } -} - -static void ggml_mul_mat_q5_1_q8_1_cuda( - const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, - const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { - - int id; - CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_device_caps[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_RDNA2) { - mmq_x = MMQ_X_Q5_1_RDNA2; - mmq_y = MMQ_Y_Q5_1_RDNA2; - nwarps = NWARPS_Q5_1_RDNA2; - } else if (compute_capability >= CC_OFFSET_AMD) { - mmq_x = MMQ_X_Q5_1_RDNA1; - mmq_y = MMQ_Y_Q5_1_RDNA1; - nwarps = NWARPS_Q5_1_RDNA1; - } else if (compute_capability >= CC_VOLTA) { - mmq_x = MMQ_X_Q5_1_AMPERE; - mmq_y = MMQ_Y_Q5_1_AMPERE; - nwarps = NWARPS_Q5_1_AMPERE; - } else if (compute_capability >= MIN_CC_DP4A) { - mmq_x = MMQ_X_Q5_1_PASCAL; - mmq_y = MMQ_Y_Q5_1_PASCAL; - nwarps = NWARPS_Q5_1_PASCAL; - } else { - GGML_ASSERT(false); - } - - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const dim3 block_nums(block_num_x, block_num_y, 1); - const dim3 block_dims(WARP_SIZE, nwarps, 1); - - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - mul_mat_q5_1<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } else { - const bool need_check = true; - mul_mat_q5_1<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } -} - -static void ggml_mul_mat_q8_0_q8_1_cuda( - const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, - const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { - - int id; - CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_device_caps[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_RDNA2) { - mmq_x = MMQ_X_Q8_0_RDNA2; - mmq_y = MMQ_Y_Q8_0_RDNA2; - nwarps = NWARPS_Q8_0_RDNA2; - } else if (compute_capability >= CC_OFFSET_AMD) { - mmq_x = MMQ_X_Q8_0_RDNA1; - mmq_y = MMQ_Y_Q8_0_RDNA1; - nwarps = NWARPS_Q8_0_RDNA1; - } else if (compute_capability >= CC_VOLTA) { - mmq_x = MMQ_X_Q8_0_AMPERE; - mmq_y = MMQ_Y_Q8_0_AMPERE; - nwarps = NWARPS_Q8_0_AMPERE; - } else if (compute_capability >= MIN_CC_DP4A) { - mmq_x = MMQ_X_Q8_0_PASCAL; - mmq_y = MMQ_Y_Q8_0_PASCAL; - nwarps = NWARPS_Q8_0_PASCAL; - } else { - GGML_ASSERT(false); - } - - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const dim3 block_nums(block_num_x, block_num_y, 1); - const dim3 block_dims(WARP_SIZE, nwarps, 1); - - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - mul_mat_q8_0<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } else { - const bool need_check = true; - mul_mat_q8_0<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } -} - -static void ggml_mul_mat_q2_K_q8_1_cuda( - const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, - const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { - - int id; - CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_device_caps[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_RDNA2) { - mmq_x = MMQ_X_Q2_K_RDNA2; - mmq_y = MMQ_Y_Q2_K_RDNA2; - nwarps = NWARPS_Q2_K_RDNA2; - } else if (compute_capability >= CC_OFFSET_AMD) { - mmq_x = MMQ_X_Q2_K_RDNA1; - mmq_y = MMQ_Y_Q2_K_RDNA1; - nwarps = NWARPS_Q2_K_RDNA1; - } else if (compute_capability >= CC_VOLTA) { - mmq_x = MMQ_X_Q2_K_AMPERE; - mmq_y = MMQ_Y_Q2_K_AMPERE; - nwarps = NWARPS_Q2_K_AMPERE; - } else if (compute_capability >= MIN_CC_DP4A) { - mmq_x = MMQ_X_Q2_K_PASCAL; - mmq_y = MMQ_Y_Q2_K_PASCAL; - nwarps = NWARPS_Q2_K_PASCAL; - } else { - GGML_ASSERT(false); - } - - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const dim3 block_nums(block_num_x, block_num_y, 1); - const dim3 block_dims(WARP_SIZE, nwarps, 1); - - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - mul_mat_q2_K<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } else { - const bool need_check = true; - mul_mat_q2_K<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } -} - -static void ggml_mul_mat_q3_K_q8_1_cuda( - const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, - const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { - -#if QK_K == 256 - - int id; - CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_device_caps[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_RDNA2) { - mmq_x = MMQ_X_Q3_K_RDNA2; - mmq_y = MMQ_Y_Q3_K_RDNA2; - nwarps = NWARPS_Q3_K_RDNA2; - } else if (compute_capability >= CC_OFFSET_AMD) { - mmq_x = MMQ_X_Q3_K_RDNA1; - mmq_y = MMQ_Y_Q3_K_RDNA1; - nwarps = NWARPS_Q3_K_RDNA1; - } else if (compute_capability >= CC_VOLTA) { - mmq_x = MMQ_X_Q3_K_AMPERE; - mmq_y = MMQ_Y_Q3_K_AMPERE; - nwarps = NWARPS_Q3_K_AMPERE; - } else if (compute_capability >= MIN_CC_DP4A) { - mmq_x = MMQ_X_Q3_K_PASCAL; - mmq_y = MMQ_Y_Q3_K_PASCAL; - nwarps = NWARPS_Q3_K_PASCAL; - } else { - GGML_ASSERT(false); - } - - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const dim3 block_nums(block_num_x, block_num_y, 1); - const dim3 block_dims(WARP_SIZE, nwarps, 1); - - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - mul_mat_q3_K<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } else { - const bool need_check = true; - mul_mat_q3_K<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } -#endif -} - -static void ggml_mul_mat_q4_K_q8_1_cuda( - const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, - const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { - - int id; - CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_device_caps[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_RDNA2) { - mmq_x = MMQ_X_Q4_K_RDNA2; - mmq_y = MMQ_Y_Q4_K_RDNA2; - nwarps = NWARPS_Q4_K_RDNA2; - } else if (compute_capability >= CC_OFFSET_AMD) { - mmq_x = MMQ_X_Q4_K_RDNA1; - mmq_y = MMQ_Y_Q4_K_RDNA1; - nwarps = NWARPS_Q4_K_RDNA1; - } else if (compute_capability >= CC_VOLTA) { - mmq_x = MMQ_X_Q4_K_AMPERE; - mmq_y = MMQ_Y_Q4_K_AMPERE; - nwarps = NWARPS_Q4_K_AMPERE; - } else if (compute_capability >= MIN_CC_DP4A) { - mmq_x = MMQ_X_Q4_K_PASCAL; - mmq_y = MMQ_Y_Q4_K_PASCAL; - nwarps = NWARPS_Q4_K_PASCAL; - } else { - GGML_ASSERT(false); - } - - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const dim3 block_nums(block_num_x, block_num_y, 1); - const dim3 block_dims(WARP_SIZE, nwarps, 1); - - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - mul_mat_q4_K<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } else { - const bool need_check = true; - mul_mat_q4_K<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } -} - -static void ggml_mul_mat_q5_K_q8_1_cuda( - const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, - const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { - - int id; - CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_device_caps[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_RDNA2) { - mmq_x = MMQ_X_Q5_K_RDNA2; - mmq_y = MMQ_Y_Q5_K_RDNA2; - nwarps = NWARPS_Q5_K_RDNA2; - } else if (compute_capability >= CC_OFFSET_AMD) { - mmq_x = MMQ_X_Q5_K_RDNA1; - mmq_y = MMQ_Y_Q5_K_RDNA1; - nwarps = NWARPS_Q5_K_RDNA1; - } else if (compute_capability >= CC_VOLTA) { - mmq_x = MMQ_X_Q5_K_AMPERE; - mmq_y = MMQ_Y_Q5_K_AMPERE; - nwarps = NWARPS_Q5_K_AMPERE; - } else if (compute_capability >= MIN_CC_DP4A) { - mmq_x = MMQ_X_Q5_K_PASCAL; - mmq_y = MMQ_Y_Q5_K_PASCAL; - nwarps = NWARPS_Q5_K_PASCAL; - } else { - GGML_ASSERT(false); - } - - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const dim3 block_nums(block_num_x, block_num_y, 1); - const dim3 block_dims(WARP_SIZE, nwarps, 1); - - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - mul_mat_q5_K<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } else { - const bool need_check = true; - mul_mat_q5_K<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } -} - -static void ggml_mul_mat_q6_K_q8_1_cuda( - const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, - const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { - - int id; - CUDA_CHECK(cudaGetDevice(&id)); - const int compute_capability = g_device_caps[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= CC_RDNA2) { - mmq_x = MMQ_X_Q6_K_RDNA2; - mmq_y = MMQ_Y_Q6_K_RDNA2; - nwarps = NWARPS_Q6_K_RDNA2; - } else if (compute_capability >= CC_OFFSET_AMD) { - mmq_x = MMQ_X_Q6_K_RDNA1; - mmq_y = MMQ_Y_Q6_K_RDNA1; - nwarps = NWARPS_Q6_K_RDNA1; - } else if (compute_capability >= CC_VOLTA) { - mmq_x = MMQ_X_Q6_K_AMPERE; - mmq_y = MMQ_Y_Q6_K_AMPERE; - nwarps = NWARPS_Q6_K_AMPERE; - } else if (compute_capability >= MIN_CC_DP4A) { - mmq_x = MMQ_X_Q6_K_PASCAL; - mmq_y = MMQ_Y_Q6_K_PASCAL; - nwarps = NWARPS_Q6_K_PASCAL; - } else { - GGML_ASSERT(false); - } - - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const dim3 block_nums(block_num_x, block_num_y, 1); - const dim3 block_dims(WARP_SIZE, nwarps, 1); - - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - mul_mat_q6_K<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } else { - const bool need_check = true; - mul_mat_q6_K<<>> - (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); - } -} - static void ggml_mul_mat_p021_f16_f32_cuda( const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y, cudaStream_t stream) { @@ -7312,614 +1168,11 @@ static void ggml_mul_mat_vec_nc_f16_f32_cuda( (vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y/nchannels_x); } - -static void ggml_cpy_f16_f32_cuda( - const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { - - const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; - cpy_f32_f16<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); -} - -static void ggml_cpy_f32_f32_cuda( - const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { - - const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; - cpy_f32_f16<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); -} - -static void ggml_cpy_f32_f16_cuda( - const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { - - const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; - cpy_f32_f16<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); -} - -static void ggml_cpy_f32_q8_0_cuda( - const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { - - GGML_ASSERT(ne % QK8_0 == 0); - const int num_blocks = ne / QK8_0; - cpy_f32_q<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); -} - -static void ggml_cpy_f32_q4_0_cuda( - const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { - - GGML_ASSERT(ne % QK4_0 == 0); - const int num_blocks = ne / QK4_0; - cpy_f32_q<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); -} - -static void ggml_cpy_f32_q4_1_cuda( - const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { - - GGML_ASSERT(ne % QK4_1 == 0); - const int num_blocks = ne / QK4_1; - cpy_f32_q<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); -} - -static void ggml_cpy_f16_f16_cuda( - const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { - - const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; - cpy_f32_f16<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); -} - - - -static void scale_f32_cuda(const float * x, float * dst, const float scale, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE; - scale_f32<<>>(x, dst, scale, k); -} - -static void clamp_f32_cuda(const float * x, float * dst, const float min, const float max, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_CLAMP_BLOCK_SIZE - 1) / CUDA_CLAMP_BLOCK_SIZE; - clamp_f32<<>>(x, dst, min, max, k); -} - -template -static void rope_cuda( - const T * x, T * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, - float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream -) { - GGML_ASSERT(ncols % 2 == 0); - const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1); - const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE); - const dim3 block_nums(nrows, num_blocks_x, 1); - if (pos == nullptr) { - rope<<>>( - x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims - ); - } else { - rope<<>>( - x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims - ); - } -} - -template -static void rope_neox_cuda( - const T * x, T * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, - float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream -) { - GGML_ASSERT(ncols % 2 == 0); - const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1); - const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE); - const dim3 block_nums(nrows, num_blocks_x, 1); - - const float theta_scale = powf(freq_base, -2.0f/n_dims); - const float inv_ndims = -1.0f / n_dims; - - if (pos == nullptr) { - rope_neox<<>>( - x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims, - theta_scale, inv_ndims - ); - } else { - rope_neox<<>>( - x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims, - theta_scale, inv_ndims - ); - } -} - -static void rope_glm_f32_cuda( - const float * x, float * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, - float freq_base, int n_ctx, cudaStream_t stream -) { - GGML_ASSERT(ncols % 4 == 0); - const dim3 block_dims(CUDA_ROPE_BLOCK_SIZE/4, 1, 1); - const int num_blocks_x = (ncols + CUDA_ROPE_BLOCK_SIZE - 1) / CUDA_ROPE_BLOCK_SIZE; - const dim3 block_nums(num_blocks_x, nrows, 1); - rope_glm_f32<<>>(x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, n_ctx); -} - -static void alibi_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, - const int k_rows, const int n_heads_log2_floor, const float m0, - const float m1, cudaStream_t stream) { - const dim3 block_dims(CUDA_ALIBI_BLOCK_SIZE, 1, 1); - const int num_blocks_x = (ncols + CUDA_ALIBI_BLOCK_SIZE - 1) / (CUDA_ALIBI_BLOCK_SIZE); - const dim3 block_nums(num_blocks_x, nrows, 1); - alibi_f32<<>>(x, dst, ncols, k_rows, n_heads_log2_floor, m0, m1); -} - -static void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) { - const dim3 block_dims(WARP_SIZE, 1, 1); - const dim3 block_nums(nrows, 1, 1); - k_sum_rows_f32<<>>(x, dst, ncols); -} - -static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) { - // bitonic sort requires ncols to be power of 2 - GGML_ASSERT((ncols & (ncols - 1)) == 0); - - const dim3 block_dims(ncols, 1, 1); - const dim3 block_nums(1, nrows, 1); - if (order == GGML_SORT_ORDER_ASC) { - k_argsort_f32_i32<<>>(x, dst, ncols); - } else if (order == GGML_SORT_ORDER_DESC) { - k_argsort_f32_i32<<>>(x, dst, ncols); - } else { - GGML_ASSERT(false); - } -} - -static void diag_mask_inf_f32_cuda(const float * x, float * dst, const int ncols_x, const int nrows_x, const int rows_per_channel, const int n_past, cudaStream_t stream) { - const dim3 block_dims(1, CUDA_DIAG_MASK_INF_BLOCK_SIZE, 1); - const int block_num_x = (ncols_x + CUDA_DIAG_MASK_INF_BLOCK_SIZE - 1) / CUDA_DIAG_MASK_INF_BLOCK_SIZE; - const dim3 block_nums(nrows_x, block_num_x, 1); - diag_mask_inf_f32<<>>(x, dst, ncols_x, rows_per_channel, n_past); -} - -static void soft_max_f32_cuda(const float * x, const float * mask, const float * pos, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) { - int nth = WARP_SIZE; - while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2; - const dim3 block_dims(nth, 1, 1); - const dim3 block_nums(nrows_x, 1, 1); - const size_t shmem = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float); - static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted."); - - const uint32_t n_head_kv = nrows_x/nrows_y; - const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv)); - - const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); - const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); - - if (shmem < g_device_caps[g_main_device].smpb) { - switch (ncols_x) { - case 32: - soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); - break; - case 64: - soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); - break; - case 128: - soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); - break; - case 256: - soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); - break; - case 512: - soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); - break; - case 1024: - soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); - break; - case 2048: - soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); - break; - case 4096: - soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); - break; - default: - soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); - break; - } - } else { - const size_t shmem_low = WARP_SIZE*sizeof(float); - soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); - } -} - -template -static void im2col_cuda(const float* x, T* dst, - int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC, - int64_t batch, int64_t batch_offset, int64_t offset_delta, - int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) { - const int parallel_elements = OW * KW * KH; - const int num_blocks = (parallel_elements + CUDA_IM2COL_BLOCK_SIZE - 1) / CUDA_IM2COL_BLOCK_SIZE; - dim3 block_nums(num_blocks, OH, batch * IC); - im2col_kernel<<>>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, (IC * KH * KW), s0, s1, p0, p1, d0, d1); -} - -// buffer pool for cuda -#define MAX_CUDA_BUFFERS 256 - -struct scoped_spin_lock { - std::atomic_flag& lock; - scoped_spin_lock(std::atomic_flag& lock) : lock(lock) { - while (lock.test_and_set(std::memory_order_acquire)) { - ; // spin - } - } - ~scoped_spin_lock() { - lock.clear(std::memory_order_release); - } - scoped_spin_lock(const scoped_spin_lock&) = delete; - scoped_spin_lock& operator=(const scoped_spin_lock&) = delete; -}; - -static std::atomic_flag g_cuda_pool_lock = ATOMIC_FLAG_INIT; - -// #define DEBUG_CUDA_MALLOC -struct ggml_cuda_buffer { - void * ptr = nullptr; - size_t size = 0; -}; - -static ggml_cuda_buffer g_cuda_buffer_pool[GGML_CUDA_MAX_DEVICES][MAX_CUDA_BUFFERS]; -static size_t g_cuda_pool_size[GGML_CUDA_MAX_DEVICES] = {0}; - -static void * ggml_cuda_pool_malloc_leg(int device, size_t size, size_t * actual_size) { - scoped_spin_lock lock(g_cuda_pool_lock); -#ifdef DEBUG_CUDA_MALLOC - int nnz = 0; - size_t max_size = 0; -#endif - size_t best_diff = 1ull << 36; - int ibest = -1; - for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) { - ggml_cuda_buffer& b = g_cuda_buffer_pool[device][i]; - if (b.ptr != nullptr) { -#ifdef DEBUG_CUDA_MALLOC - ++nnz; - if (b.size > max_size) max_size = b.size; -#endif - if (b.size >= size) { - size_t diff = b.size - size; - if (diff < best_diff) { - best_diff = diff; - ibest = i; - if (!best_diff) { - void * ptr = b.ptr; - *actual_size = b.size; - b.ptr = nullptr; - b.size = 0; - return ptr; - } - } - } - } - } - if (ibest >= 0) { - ggml_cuda_buffer& b = g_cuda_buffer_pool[device][ibest]; - void * ptr = b.ptr; - *actual_size = b.size; - b.ptr = nullptr; - b.size = 0; - return ptr; - } - void * ptr; - size_t look_ahead_size = (size_t) (1.05 * size); - look_ahead_size = 256 * ((look_ahead_size + 255)/256); - ggml_cuda_set_device(device); - CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size)); - *actual_size = look_ahead_size; - g_cuda_pool_size[device] += look_ahead_size; -#ifdef DEBUG_CUDA_MALLOC - fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz, - (uint32_t)(max_size/1024/1024), (uint32_t)(g_cuda_pool_size[device]/1024/1024), (uint32_t)(size/1024/1024)); -#endif - return ptr; -} - -static void ggml_cuda_pool_free_leg(int device, void * ptr, size_t size) { - scoped_spin_lock lock(g_cuda_pool_lock); - - for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) { - ggml_cuda_buffer& b = g_cuda_buffer_pool[device][i]; - if (b.ptr == nullptr) { - b.ptr = ptr; - b.size = size; - return; - } - } - fprintf(stderr, "WARNING: cuda buffer pool full, increase MAX_CUDA_BUFFERS\n"); - ggml_cuda_set_device(device); - CUDA_CHECK(cudaFree(ptr)); - g_cuda_pool_size[device] -= size; -} - -#if !defined(GGML_USE_HIPBLAS) -// pool with virtual memory -static CUdeviceptr g_cuda_pool_addr[GGML_CUDA_MAX_DEVICES] = {0}; -static size_t g_cuda_pool_used[GGML_CUDA_MAX_DEVICES] = {0}; -static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB - -static void * ggml_cuda_pool_malloc_vmm(int device, size_t size, size_t * actual_size) { - scoped_spin_lock lock(g_cuda_pool_lock); - - // round up the allocation size to the alignment to ensure that all allocations are aligned for all data types - const size_t alignment = 128; - size = alignment * ((size + alignment - 1) / alignment); - - size_t avail = g_cuda_pool_size[device] - g_cuda_pool_used[device]; - - if (size > avail) { - // round up to the next multiple of the granularity - size_t reserve_size = size - avail; - const size_t granularity = g_device_caps[device].vmm_granularity; - reserve_size = granularity * ((reserve_size + granularity - 1) / granularity); - - GGML_ASSERT(g_cuda_pool_size[device] + reserve_size <= CUDA_POOL_VMM_MAX_SIZE); - - // allocate more physical memory - CUmemAllocationProp prop = {}; - prop.type = CU_MEM_ALLOCATION_TYPE_PINNED; - prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE; - prop.location.id = device; - CUmemGenericAllocationHandle handle; - CU_CHECK(cuMemCreate(&handle, reserve_size, &prop, 0)); - - // reserve virtual address space (if not already reserved) - if (g_cuda_pool_addr[device] == 0) { - CU_CHECK(cuMemAddressReserve(&g_cuda_pool_addr[device], CUDA_POOL_VMM_MAX_SIZE, 0, 0, 0)); - } - - // map at the end of the pool - CU_CHECK(cuMemMap(g_cuda_pool_addr[device] + g_cuda_pool_size[device], reserve_size, 0, handle, 0)); - - // the memory allocation handle is no longer needed after mapping - CU_CHECK(cuMemRelease(handle)); - - // set access - CUmemAccessDesc access = {}; - access.location.type = CU_MEM_LOCATION_TYPE_DEVICE; - access.location.id = device; - access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE; - CU_CHECK(cuMemSetAccess(g_cuda_pool_addr[device] + g_cuda_pool_size[device], reserve_size, &access, 1)); - - // add to the pool - g_cuda_pool_size[device] += reserve_size; - - //printf("cuda pool[%d]: size increased to %llu MB (reserved %llu MB)\n", - // id, (unsigned long long) (g_cuda_pool_size[id]/1024/1024), - // (unsigned long long) (reserve_size/1024/1024)); - } - - GGML_ASSERT(g_cuda_pool_addr[device] != 0); - - void * ptr = (void *) (g_cuda_pool_addr[device] + g_cuda_pool_used[device]); - *actual_size = size; - g_cuda_pool_used[device] += size; - -#ifdef DEBUG_CUDA_MALLOC - printf("cuda pool[%d]: allocated %llu bytes at %llx\n", device, (unsigned long long) size, ptr); -#endif - - return ptr; -} - -static void ggml_cuda_pool_free_vmm(int device, void * ptr, size_t size) { - scoped_spin_lock lock(g_cuda_pool_lock); - -#ifdef DEBUG_CUDA_MALLOC - printf("cuda pool[%d]: freed %llu bytes at %llx\n", device, (unsigned long long) size, ptr); -#endif - - g_cuda_pool_used[device] -= size; - - // all deallocations must be in reverse order of the allocations - GGML_ASSERT(ptr == (void *) (g_cuda_pool_addr[device] + g_cuda_pool_used[device])); -} - -static void * ggml_cuda_pool_malloc(int device, size_t size, size_t * actual_size) { - if (g_device_caps[device].vmm) { - return ggml_cuda_pool_malloc_vmm(device, size, actual_size); - } else { - return ggml_cuda_pool_malloc_leg(device, size, actual_size); - } -} - -static void ggml_cuda_pool_free(int device, void * ptr, size_t size) { - if (g_device_caps[device].vmm) { - ggml_cuda_pool_free_vmm(device, ptr, size); - } else { - ggml_cuda_pool_free_leg(device, ptr, size); - } -} -#else -#define ggml_cuda_pool_malloc ggml_cuda_pool_malloc_leg -#define ggml_cuda_pool_free ggml_cuda_pool_free_leg -#endif // !defined(GGML_USE_HIPBLAS) - -template -struct cuda_pool_alloc { - int device = -1; - T * ptr = nullptr; - size_t actual_size = 0; - - // size is in number of elements - T * alloc(size_t size) { - GGML_ASSERT(ptr == nullptr); - CUDA_CHECK(cudaGetDevice(&device)); - ptr = (T *) ggml_cuda_pool_malloc(device, size * sizeof(T), &this->actual_size); - return ptr; - } - - cuda_pool_alloc(size_t size) { - alloc(size); - } - - ~cuda_pool_alloc() { - if (ptr != nullptr) { - ggml_cuda_pool_free(device, ptr, actual_size); - } - } - - T * get() { - return ptr; - } - - cuda_pool_alloc() = default; - cuda_pool_alloc(const cuda_pool_alloc &) = delete; - cuda_pool_alloc(cuda_pool_alloc &&) = delete; - cuda_pool_alloc& operator=(const cuda_pool_alloc &) = delete; - cuda_pool_alloc& operator=(cuda_pool_alloc &&) = delete; -}; - -static bool g_cublas_loaded = false; - -static void ggml_init_cublas() { - static bool initialized = false; - - if (!initialized) { - -#ifdef __HIP_PLATFORM_AMD__ - // Workaround for a rocBLAS bug when using multiple graphics cards: - // https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346 - rocblas_initialize(); - CUDA_CHECK(cudaDeviceSynchronize()); -#endif - - if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) { - initialized = true; - g_cublas_loaded = false; - fprintf(stderr, "%s: no " GGML_CUDA_NAME " devices found, " GGML_CUDA_NAME " will be disabled\n", __func__); - return; - } - - GGML_ASSERT(g_device_count <= GGML_CUDA_MAX_DEVICES); - int64_t total_vram = 0; -#if defined(GGML_CUDA_FORCE_MMQ) - fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ: yes\n", __func__); -#else - fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ: no\n", __func__); -#endif -#if defined(CUDA_USE_TENSOR_CORES) - fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: yes\n", __func__); -#else - fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: no\n", __func__); -#endif - fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, g_device_count); - for (int id = 0; id < g_device_count; ++id) { - int device_vmm = 0; - -#if !defined(GGML_USE_HIPBLAS) - CUdevice device; - CU_CHECK(cuDeviceGet(&device, id)); - CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device)); - - if (device_vmm) { - CUmemAllocationProp alloc_prop = {}; - alloc_prop.type = CU_MEM_ALLOCATION_TYPE_PINNED; - alloc_prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE; - alloc_prop.location.id = id; - CU_CHECK(cuMemGetAllocationGranularity(&g_device_caps[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED)); - } -#endif // !defined(GGML_USE_HIPBLAS) - g_device_caps[id].vmm = !!device_vmm; - - cudaDeviceProp prop; - CUDA_CHECK(cudaGetDeviceProperties(&prop, id)); - fprintf(stderr, " Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no"); - - g_default_tensor_split[id] = total_vram; - total_vram += prop.totalGlobalMem; - -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) - g_device_caps[id].cc = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD; -#else - g_device_caps[id].cc = 100*prop.major + 10*prop.minor; -#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) - g_device_caps[id].smpb = prop.sharedMemPerBlock; - } - for (int id = 0; id < g_device_count; ++id) { - g_default_tensor_split[id] /= total_vram; - } - - for (int id = 0; id < g_device_count; ++id) { - ggml_cuda_set_device(id); - - // create cuda streams - for (int is = 0; is < MAX_STREAMS; ++is) { - CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams[id][is], cudaStreamNonBlocking)); - } - - // create cublas handle - CUBLAS_CHECK(cublasCreate(&g_cublas_handles[id])); - CUBLAS_CHECK(cublasSetMathMode(g_cublas_handles[id], CUBLAS_TF32_TENSOR_OP_MATH)); - } - - // configure logging to stdout - // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr)); - - initialized = true; - g_cublas_loaded = true; - } -} - -static void * ggml_cuda_host_malloc(size_t size) { - if (getenv("GGML_CUDA_NO_PINNED") != nullptr) { - return nullptr; - } - - void * ptr = nullptr; - cudaError_t err = cudaMallocHost((void **) &ptr, size); - if (err != cudaSuccess) { - // clear the error - cudaGetLastError(); - fprintf(stderr, "%s: warning: failed to allocate %.2f MiB of pinned memory: %s\n", __func__, - size/1024.0/1024.0, cudaGetErrorString(err)); - return nullptr; - } - - return ptr; -} - -static void ggml_cuda_host_free(void * ptr) { - CUDA_CHECK(cudaFreeHost(ptr)); -} - static cudaError_t ggml_cuda_cpy_tensor_2d( void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) { - cudaMemcpyKind kind; - char * src_ptr; - if (src->backend == GGML_BACKEND_TYPE_CPU) { - kind = cudaMemcpyHostToDevice; - src_ptr = (char *) src->data; - } else if (src->backend == GGML_BACKEND_TYPE_GPU || src->backend == GGML_BACKEND_TYPE_GPU_SPLIT) { - GGML_ASSERT(src->backend != GGML_BACKEND_TYPE_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1])); - kind = cudaMemcpyDeviceToDevice; - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra; - int id; - CUDA_CHECK(cudaGetDevice(&id)); - src_ptr = (char *) extra->data_device[id]; - } else { - GGML_ASSERT(false); - } + GGML_ASSERT(ggml_backend_buffer_is_cuda(src->buffer)); + char * src_ptr = (char *) src->data; char * dst_ptr = (char *) dst; const int64_t ne0 = src->ne[0]; @@ -7934,746 +1187,25 @@ static cudaError_t ggml_cuda_cpy_tensor_2d( const char * x = src_ptr + i1_low*nb1 + i2*nb2 + i3*nb3; if (nb0 == ts && nb1 == ts*ne0/bs) { - return cudaMemcpyAsync(dst_ptr, x, i1_diff*nb1, kind, stream); + return cudaMemcpyAsync(dst_ptr, x, i1_diff*nb1, cudaMemcpyDeviceToDevice, stream); } else if (nb0 == ts) { - return cudaMemcpy2DAsync(dst_ptr, ts*ne0/bs, x, nb1, ts*ne0/bs, i1_diff, kind, stream); + return cudaMemcpy2DAsync(dst_ptr, ts*ne0/bs, x, nb1, ts*ne0/bs, i1_diff, cudaMemcpyDeviceToDevice, stream); } else { for (int64_t i1 = 0; i1 < i1_diff; i1++) { const void * rx = (const void *) ((const char *) x + i1*nb1); void * rd = (void *) (dst_ptr + i1*ts*ne0/bs); // pretend the row is a matrix with cols=1 - cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, kind, stream); - if (r != cudaSuccess) return r; + cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, cudaMemcpyDeviceToDevice, stream); + if (r != cudaSuccess) { + return r; + } } return cudaSuccess; } } -static void ggml_cuda_op_get_rows( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_d, const float * src1_d, float * dst_d, cudaStream_t stream) { - - GGML_ASSERT(src1->type == GGML_TYPE_I32); - GGML_ASSERT(dst->type == GGML_TYPE_F32); - - GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type)); - GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type)); - GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type)); - - const int32_t * src1_i32 = (const int32_t *) src1_d; - - switch (src0->type) { - case GGML_TYPE_F16: - get_rows_cuda_float(src0, src1, dst, (const half *)src0_d, src1_i32, dst_d, stream); - break; - case GGML_TYPE_F32: - get_rows_cuda_float(src0, src1, dst, src0_d, src1_i32, dst_d, stream); - break; - case GGML_TYPE_Q4_0: - get_rows_cuda(src0, src1, dst, src0_d, src1_i32, dst_d, stream); - break; - case GGML_TYPE_Q4_1: - get_rows_cuda(src0, src1, dst, src0_d, src1_i32, dst_d, stream); - break; - case GGML_TYPE_Q5_0: - get_rows_cuda(src0, src1, dst, src0_d, src1_i32, dst_d, stream); - break; - case GGML_TYPE_Q5_1: - get_rows_cuda(src0, src1, dst, src0_d, src1_i32, dst_d, stream); - break; - case GGML_TYPE_Q8_0: - get_rows_cuda(src0, src1, dst, src0_d, src1_i32, dst_d, stream); - break; - default: - // TODO: k-quants - fprintf(stderr, "%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type)); - GGML_ASSERT(false); - break; - } -} - -template -static void ggml_cuda_op_bin_bcast( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src1->type == GGML_TYPE_F32); - - if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - op()(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); - } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { - op()(src0, src1, dst, (const half *) src0_dd, src1_dd, (half *) dst_dd, main_stream); - } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) { - op()(src0, src1, dst, (const half *) src0_dd, src1_dd, dst_dd, main_stream); - } else { - fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__, - ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type)); - GGML_ASSERT(false); - } -} - -static void ggml_cuda_op_repeat( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_d, const float * src1_d, float * dst_d, cudaStream_t main_stream) { - - ggml_cuda_op_bin_bcast>(dst, src0, dst, nullptr, src0_d, dst_d, main_stream); - - (void) src1; - (void) src1_d; -} - -static void ggml_cuda_op_add( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - ggml_cuda_op_bin_bcast>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); -} - -static void ggml_cuda_op_acc( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported - - int nb1 = dst->op_params[0] / 4; // 4 bytes of float32 - int nb2 = dst->op_params[1] / 4; // 4 bytes of float32 - // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused - int offset = dst->op_params[3] / 4; // offset in bytes - - acc_f32_cuda(src0_dd, src1_dd, dst_dd, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, main_stream); - - (void) dst; -} - -static void ggml_cuda_op_mul( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - ggml_cuda_op_bin_bcast>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); -} - -static void ggml_cuda_op_div( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - ggml_cuda_op_bin_bcast>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); -} - -static void ggml_cuda_op_gelu( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - gelu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_silu( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - silu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_gelu_quick( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - gelu_quick_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_tanh( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - tanh_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_relu( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - relu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_hardsigmoid( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - hardsigmoid_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_hardswish( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - hardswish_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_leaky_relu( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - float negative_slope; - memcpy(&negative_slope, dst->op_params, sizeof(float)); - - leaky_relu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), negative_slope, main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_sqr( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - sqr_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_norm( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - const int64_t ne00 = src0->ne[0]; - const int64_t nrows = ggml_nrows(src0); - - float eps; - memcpy(&eps, dst->op_params, sizeof(float)); - - norm_f32_cuda(src0_dd, dst_dd, ne00, nrows, eps, main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_group_norm( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int num_groups = dst->op_params[0]; - int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups); - group_norm_f32_cuda(src0_dd, dst_dd, num_groups * src0->ne[3], group_size, ggml_nelements(src0), main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_concat( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT(dst->type == GGML_TYPE_F32); - - for (int i3 = 0; i3 < dst->ne[3]; i3++) { - concat_f32_cuda(src0_dd + i3 * (src0->nb[3] / 4), src1_dd + i3 * (src1->nb[3] / 4), dst_dd + i3 * (dst->nb[3] / 4), dst->ne[0], dst->ne[1], dst->ne[2], src0->ne[2], main_stream); - } - - (void) src1; - (void) dst; -} - -static void ggml_cuda_op_upscale( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT(dst->type == GGML_TYPE_F32); - GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors - - const int scale_factor = dst->op_params[0]; - - upscale_f32_cuda(src0_dd, dst_dd, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], scale_factor, main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_pad( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT(dst->type == GGML_TYPE_F32); - GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors - - pad_f32_cuda(src0_dd, dst_dd, - src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], - dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_arange( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(dst->type == GGML_TYPE_F32); - - float start; - float stop; - float step; - memcpy(&start, (float *)dst->op_params + 0, sizeof(float)); - memcpy(&stop, (float *)dst->op_params + 1, sizeof(float)); - memcpy(&step, (float *)dst->op_params + 2, sizeof(float)); - - int64_t steps = (int64_t)ceil((stop - start) / step); - GGML_ASSERT(ggml_nelements(dst) == steps); - - arange_f32_cuda(dst_dd, dst->ne[0], start, step, main_stream); - - (void) src0; - (void) src1; - (void) src0_dd; - (void) src1_dd; -} - -static void ggml_cuda_op_timestep_embedding( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT(dst->type == GGML_TYPE_F32); - - const int dim = dst->op_params[0]; - const int max_period = dst->op_params[1]; - - timestep_embedding_f32_cuda(src0_dd, dst_dd, src0->ne[0], dst->nb[1], dim, max_period, main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_rms_norm( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - const int64_t ne00 = src0->ne[0]; - const int64_t nrows = ggml_nrows(src0); - - float eps; - memcpy(&eps, dst->op_params, sizeof(float)); - - rms_norm_f32_cuda(src0_dd, dst_dd, ne00, nrows, eps, main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_mul_mat_q( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, - const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, - const int64_t src1_padded_row_size, cudaStream_t stream) { - - const int64_t ne00 = src0->ne[0]; - - const int64_t ne10 = src1->ne[0]; - GGML_ASSERT(ne10 % QK8_1 == 0); - - const int64_t ne0 = dst->ne[0]; - - const int64_t row_diff = row_high - row_low; - - int id; - CUDA_CHECK(cudaGetDevice(&id)); - - // the main device has a larger memory buffer to hold the results from all GPUs - // nrows_dst == nrows of the matrix that the kernel writes into - const int64_t nrows_dst = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device ? ne0 : row_diff; - - switch (src0->type) { - case GGML_TYPE_Q4_0: - ggml_mul_mat_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q4_1: - ggml_mul_mat_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q5_0: - ggml_mul_mat_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q5_1: - ggml_mul_mat_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q8_0: - ggml_mul_mat_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q2_K: - ggml_mul_mat_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q3_K: - ggml_mul_mat_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q4_K: - ggml_mul_mat_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q5_K: - ggml_mul_mat_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q6_K: - ggml_mul_mat_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - default: - GGML_ASSERT(false); - break; - } - - (void) src1; - (void) dst; - (void) src1_ddf_i; -} - -static int64_t get_row_rounding(ggml_type type, const std::array & tensor_split) { - int64_t min_compute_capability = INT_MAX; - int64_t max_compute_capability = INT_MIN; - for (int id = 0; id < g_device_count; ++id) { - if (tensor_split[id] < (id + 1 < g_device_count ? tensor_split[id + 1] : 1.0f)) { - if (min_compute_capability > g_device_caps[id].cc) { - min_compute_capability = g_device_caps[id].cc; - } - if (max_compute_capability < g_device_caps[id].cc) { - max_compute_capability = g_device_caps[id].cc; - } - } - } - -#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) - switch(type) { - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_Q5_0: - case GGML_TYPE_Q5_1: - case GGML_TYPE_Q8_0: - return max_compute_capability >= CC_RDNA2 ? 128 : 64; - case GGML_TYPE_F16: - case GGML_TYPE_F32: - return 1; - case GGML_TYPE_Q2_K: - return max_compute_capability >= CC_RDNA2 ? 128 : 32; - case GGML_TYPE_Q3_K: - return min_compute_capability < CC_RDNA2 ? 128 : 64; - case GGML_TYPE_Q4_K: - case GGML_TYPE_Q5_K: - case GGML_TYPE_Q6_K: - case GGML_TYPE_IQ2_XXS: - case GGML_TYPE_IQ2_XS: - case GGML_TYPE_IQ2_S: - case GGML_TYPE_IQ3_XXS: - case GGML_TYPE_IQ1_S: - case GGML_TYPE_IQ4_NL: - case GGML_TYPE_IQ4_XS: - case GGML_TYPE_IQ3_S: - return max_compute_capability >= CC_RDNA2 ? 128 : 64; - default: - GGML_ASSERT(false); - } -#else - switch(type) { - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - return max_compute_capability >= CC_VOLTA ? 128 : 64; - case GGML_TYPE_Q5_0: - case GGML_TYPE_Q5_1: - case GGML_TYPE_Q8_0: - return 64; - case GGML_TYPE_F16: - case GGML_TYPE_F32: - return 1; - case GGML_TYPE_Q2_K: - case GGML_TYPE_Q3_K: - case GGML_TYPE_Q4_K: - case GGML_TYPE_Q5_K: - case GGML_TYPE_IQ2_XXS: - case GGML_TYPE_IQ2_XS: - case GGML_TYPE_IQ2_S: - case GGML_TYPE_IQ3_XXS: - case GGML_TYPE_IQ1_S: - case GGML_TYPE_IQ4_NL: - case GGML_TYPE_IQ4_XS: - case GGML_TYPE_IQ3_S: - return max_compute_capability >= CC_VOLTA ? 128 : 64; - case GGML_TYPE_Q6_K: - return 64; - default: - GGML_ASSERT(false); - } -#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) -} - -static void get_row_split(int64_t * row_low, int64_t * row_high, const ggml_tensor * tensor, const std::array & tensor_split, int id) { - const int64_t nrows = ggml_nrows(tensor); - const int64_t rounding = get_row_rounding(tensor->type, tensor_split); - - *row_low = id == 0 ? 0 : nrows*tensor_split[id]; - *row_low -= *row_low % rounding; - - if (id == g_device_count - 1) { - *row_high = nrows; - } else { - *row_high = nrows*tensor_split[id + 1]; - *row_high -= *row_high % rounding; - } -} - -static void ggml_cuda_op_mul_mat_vec_q( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, - const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, - const int64_t src1_padded_row_size, cudaStream_t stream) { - - const int64_t ne00 = src0->ne[0]; - const int64_t row_diff = row_high - row_low; - - const int64_t ne10 = src1->ne[0]; - GGML_ASSERT(ne10 % QK8_1 == 0); - - const int64_t ne0 = dst->ne[0]; - - int id; - CUDA_CHECK(cudaGetDevice(&id)); - - // the main device has a larger memory buffer to hold the results from all GPUs - // nrows_dst == nrows of the matrix that the kernel writes into - const int64_t nrows_dst = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device ? ne0 : row_diff; - - switch (src0->type) { - case GGML_TYPE_Q4_0: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_Q4_1: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_Q5_0: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_Q5_1: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_Q8_0: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_Q2_K: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_Q3_K: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_Q4_K: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_Q5_K: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_Q6_K: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_IQ2_XXS: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_IQ2_XS: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_IQ2_S: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_IQ3_XXS: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_IQ1_S: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_IQ4_NL: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_IQ4_XS: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - case GGML_TYPE_IQ3_S: - mul_mat_vec_q_cuda - (src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); - break; - default: - GGML_ASSERT(false); - break; - } - - (void) src1; - (void) dst; - (void) src1_ddf_i; - (void) src1_ncols; - (void) src1_padded_row_size; -} - -static void ggml_cuda_op_dequantize_mul_mat_vec( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, - const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, - const int64_t src1_padded_row_size, cudaStream_t stream) { - - const int64_t ne00 = src0->ne[0]; - const int64_t row_diff = row_high - row_low; - - GGML_ASSERT(src1->type == GGML_TYPE_F32); - - // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics -#ifdef GGML_CUDA_F16 - cuda_pool_alloc src1_dfloat_a; - half * src1_dfloat = nullptr; // dfloat == half - - bool src1_convert_f16 = - src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 || - src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 || - src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16; - - if (src1_convert_f16) { - src1_dfloat = src1_dfloat_a.alloc(ne00); - const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type); - GGML_ASSERT(to_fp16_cuda != nullptr); - to_fp16_cuda(src1_ddf_i, src1_dfloat, ne00, stream); - } -#else - const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion -#endif // GGML_CUDA_F16 - - switch (src0->type) { - case GGML_TYPE_Q4_0: - dequantize_mul_mat_vec_q4_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q4_1: - dequantize_mul_mat_vec_q4_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q5_0: - dequantize_mul_mat_vec_q5_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q5_1: - dequantize_mul_mat_vec_q5_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q8_0: - dequantize_mul_mat_vec_q8_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q2_K: - dequantize_mul_mat_vec_q2_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q3_K: - dequantize_mul_mat_vec_q3_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q4_K: - dequantize_mul_mat_vec_q4_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q5_K: - dequantize_mul_mat_vec_q5_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q6_K: - dequantize_mul_mat_vec_q6_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_F16: - convert_mul_mat_vec_f16_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); - break; - default: - GGML_ASSERT(false); - break; - } - - (void) src1; - (void) dst; - (void) src1_ddq_i; - (void) src1_ncols; - (void) src1_padded_row_size; -} - static void ggml_cuda_op_mul_mat_cublas( + ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, const int64_t src1_padded_row_size, cudaStream_t stream) { @@ -8689,19 +1221,17 @@ static void ggml_cuda_op_mul_mat_cublas( const int64_t row_diff = row_high - row_low; - int id; - CUDA_CHECK(cudaGetDevice(&id)); + int id = ggml_cuda_get_device(); // the main device has a larger memory buffer to hold the results from all GPUs // ldc == nrows of the matrix that cuBLAS writes into - int ldc = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device ? ne0 : row_diff; + int ldc = id == ctx.device ? ne0 : row_diff; - const int compute_capability = g_device_caps[id].cc; + const int compute_capability = ggml_cuda_info().devices[id].cc; if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) { - //printf("this branch\n"); // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32 - cuda_pool_alloc src0_as_f16; + ggml_cuda_pool_alloc src0_as_f16(ctx.pool()); if (src0->type != GGML_TYPE_F16) { const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src0->type); GGML_ASSERT(to_fp16_cuda != nullptr); @@ -8711,7 +1241,7 @@ static void ggml_cuda_op_mul_mat_cublas( } const half * src0_ptr = src0->type == GGML_TYPE_F16 ? (const half *) src0_dd_i : src0_as_f16.get(); - cuda_pool_alloc src1_as_f16; + ggml_cuda_pool_alloc src1_as_f16(ctx.pool()); if (src1->type != GGML_TYPE_F16) { const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type); GGML_ASSERT(to_fp16_cuda != nullptr); @@ -8720,14 +1250,14 @@ static void ggml_cuda_op_mul_mat_cublas( to_fp16_cuda(src1_ddf_i, src1_as_f16.get(), ne, stream); } const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddf_i : src1_as_f16.get(); - cuda_pool_alloc dst_f16(row_diff*src1_ncols); + ggml_cuda_pool_alloc dst_f16(ctx.pool(), row_diff*src1_ncols); const half alpha_f16 = 1.0f; const half beta_f16 = 0.0f; - CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], stream)); + CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream)); CUBLAS_CHECK( - cublasGemmEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N, + cublasGemmEx(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N, row_diff, src1_ncols, ne10, &alpha_f16, src0_ptr, CUDA_R_16F, ne00, src1_ptr, CUDA_R_16F, ne10, @@ -8738,8 +1268,8 @@ static void ggml_cuda_op_mul_mat_cublas( const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16); to_fp32_cuda(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream); } else { - cuda_pool_alloc src0_ddq_as_f32; - cuda_pool_alloc src1_ddq_as_f32; + ggml_cuda_pool_alloc src0_ddq_as_f32(ctx.pool(id)); + ggml_cuda_pool_alloc src1_ddq_as_f32(ctx.pool(id)); if (src0->type != GGML_TYPE_F32) { const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type); @@ -8760,366 +1290,21 @@ static void ggml_cuda_op_mul_mat_cublas( const float alpha = 1.0f; const float beta = 0.0f; - CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], stream)); + CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream)); CUBLAS_CHECK( - cublasSgemm(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N, + cublasSgemm(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N, row_diff, src1_ncols, ne10, &alpha, src0_ddf_i, ne00, src1_ddf1_i, ne10, &beta, dst_dd_i, ldc)); } - (void) dst; - (void) src1_ddq_i; - (void) src1_padded_row_size; + GGML_UNUSED(dst); + GGML_UNUSED(src1_ddq_i); + GGML_UNUSED(src1_padded_row_size); } -static void ggml_cuda_op_rope( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); - GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); - GGML_ASSERT(src0->type == dst->type); - - const int64_t ne00 = src0->ne[0]; - const int64_t ne01 = src0->ne[1]; - const int64_t ne2 = dst->ne[2]; - const int64_t nrows = ggml_nrows(src0); - - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - const int n_ctx = ((int32_t *) dst->op_params)[3]; - const int n_orig_ctx = ((int32_t *) dst->op_params)[4]; - - // RoPE alteration for extended context - float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; - memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); - memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); - memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); - memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); - memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); - memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); - - const int32_t * pos = nullptr; - if ((mode & 1) == 0) { - GGML_ASSERT(src1->type == GGML_TYPE_I32); - GGML_ASSERT(src1->ne[0] == ne2); - pos = (const int32_t *) src1_dd; - } - - const bool is_neox = mode & 2; - const bool is_glm = mode & 4; - - rope_corr_dims corr_dims; - ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v); - - // compute - if (is_glm) { - GGML_ASSERT(false); - rope_glm_f32_cuda(src0_dd, dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, n_ctx, main_stream); - } else if (is_neox) { - if (src0->type == GGML_TYPE_F32) { - rope_neox_cuda( - (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor, - attn_factor, corr_dims, main_stream - ); - } else if (src0->type == GGML_TYPE_F16) { - rope_neox_cuda( - (const half *)src0_dd, (half *)dst_dd, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor, - attn_factor, corr_dims, main_stream - ); - } else { - GGML_ASSERT(false); - } - } else { - if (src0->type == GGML_TYPE_F32) { - rope_cuda( - (const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor, - attn_factor, corr_dims, main_stream - ); - } else if (src0->type == GGML_TYPE_F16) { - rope_cuda( - (const half *)src0_dd, (half *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor, - attn_factor, corr_dims, main_stream - ); - } else { - GGML_ASSERT(false); - } - } - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_alibi( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - const int64_t ne00 = src0->ne[0]; - const int64_t ne01 = src0->ne[1]; - const int64_t ne02 = src0->ne[2]; - const int64_t nrows = ggml_nrows(src0); - - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_head = ((int32_t *) dst->op_params)[1]; - float max_bias; - memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float)); - - //GGML_ASSERT(ne01 + n_past == ne00); - GGML_ASSERT(n_head == ne02); - - const int n_heads_log2_floor = 1 << (int) floor(log2(n_head)); - - const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor); - const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor); - - alibi_f32_cuda(src0_dd, dst_dd, ne00, nrows, ne01, n_heads_log2_floor, m0, m1, main_stream); - - (void) src1; - (void) src1_dd; -} - -static void ggml_cuda_op_pool2d( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - const int32_t * opts = (const int32_t *)dst->op_params; - enum ggml_op_pool op = static_cast(opts[0]); - const int k0 = opts[1]; - const int k1 = opts[2]; - const int s0 = opts[3]; - const int s1 = opts[4]; - const int p0 = opts[5]; - const int p1 = opts[6]; - - const int64_t IH = src0->ne[1]; - const int64_t IW = src0->ne[0]; - - const int64_t N = dst->ne[3]; - const int64_t OC = dst->ne[2]; - const int64_t OH = dst->ne[1]; - const int64_t OW = dst->ne[0]; - - const int parallel_elements = N * OC * OH * OW; - const int num_blocks = (parallel_elements + CUDA_POOL2D_BLOCK_SIZE - 1) / CUDA_POOL2D_BLOCK_SIZE; - dim3 block_nums(num_blocks); - pool2d_nchw_kernel<<>>(IH, IW, OH, OW, k1, k0, s1, s0, p1, p0, parallel_elements, src0_dd, dst_dd, op); - - (void) src1; - (void) src1_dd; -} - -static void ggml_cuda_op_im2col( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32); - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[2]; - const int32_t p1 = ((const int32_t*)(dst->op_params))[3]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[4]; - const int32_t d1 = ((const int32_t*)(dst->op_params))[5]; - - const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1; - - const int64_t IC = src1->ne[is_2D ? 2 : 1]; - const int64_t IH = is_2D ? src1->ne[1] : 1; - const int64_t IW = src1->ne[0]; - - const int64_t KH = is_2D ? src0->ne[1] : 1; - const int64_t KW = src0->ne[0]; - - const int64_t OH = is_2D ? dst->ne[2] : 1; - const int64_t OW = dst->ne[1]; - - const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32 - const int64_t batch = src1->ne[3]; - const size_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32 - - if(dst->type == GGML_TYPE_F16) { - im2col_cuda(src1_dd, (half*) dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream); - } else { - im2col_cuda(src1_dd, (float*) dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream); - } - - (void) src0; - (void) src0_dd; -} - -static void ggml_cuda_op_sum_rows( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - const int64_t ncols = src0->ne[0]; - const int64_t nrows = ggml_nrows(src0); - - sum_rows_f32_cuda(src0_dd, dst_dd, ncols, nrows, main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_argsort( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_I32); - - const int64_t ncols = src0->ne[0]; - const int64_t nrows = ggml_nrows(src0); - - enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0]; - - argsort_f32_i32_cuda(src0_dd, (int *)dst_dd, ncols, nrows, order, main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_diag_mask_inf( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - const int64_t ne00 = src0->ne[0]; - const int64_t ne01 = src0->ne[1]; - const int nrows0 = ggml_nrows(src0); - - const int n_past = ((int32_t *) dst->op_params)[0]; - - diag_mask_inf_f32_cuda(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_soft_max( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional - - const int64_t ne00 = src0->ne[0]; - const int64_t nrows_x = ggml_nrows(src0); - const int64_t nrows_y = src0->ne[1]; - - float scale = 1.0f; - float max_bias = 0.0f; - - memcpy(&scale, (float *) dst->op_params + 0, sizeof(float)); - memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float)); - - // positions tensor - float * src2_dd = nullptr; - - ggml_tensor * src2 = dst->src[2]; - const bool use_src2 = src2 != nullptr; - - if (use_src2) { - ggml_tensor_extra_gpu * src2_extra = (ggml_tensor_extra_gpu *) src2->extra; - src2_dd = (float *) src2_extra->data_device[g_main_device]; - } - - soft_max_f32_cuda(src0_dd, src1 ? src1_dd : nullptr, src2_dd, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream); -} - -static void ggml_cuda_op_scale( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - float scale; - memcpy(&scale, dst->op_params, sizeof(float)); - - scale_f32_cuda(src0_dd, dst_dd, scale, ggml_nelements(src0), main_stream); - CUDA_CHECK(cudaGetLastError()); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_clamp( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - float min; - float max; - memcpy(&min, dst->op_params, sizeof(float)); - memcpy(&max, (float *) dst->op_params + 1, sizeof(float)); - - clamp_f32_cuda(src0_dd, dst_dd, min, max, ggml_nelements(src0), main_stream); - CUDA_CHECK(cudaGetLastError()); - - (void) src1; - (void) dst; - (void) src1_dd; -} - -static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const ggml_cuda_op_flatten_t op) { - const int64_t nrows0 = ggml_nrows(src0); - - const bool use_src1 = src1 != nullptr; - const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1; - - GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT); - GGML_ASSERT( dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT); - - ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr; - ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; - - // dd = data device - float * src0_ddf = nullptr; - float * src1_ddf = nullptr; - float * dst_ddf = nullptr; - - ggml_cuda_set_device(g_main_device); - cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; - - src0_ddf = (float *) src0_extra->data_device[g_main_device]; - - if (use_src1) { - src1_ddf = (float *) src1_extra->data_device[g_main_device]; - } - dst_ddf = (float *) dst_extra->data_device[g_main_device]; - - // do the computation - op(src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream); - CUDA_CHECK(cudaGetLastError()); -} - -static void ggml_cuda_set_peer_access(const int n_tokens) { +static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) { static bool peer_access_enabled = false; const bool enable_peer_access = n_tokens <= GGML_CUDA_PEER_MAX_BATCH_SIZE; @@ -9129,19 +1314,19 @@ static void ggml_cuda_set_peer_access(const int n_tokens) { } #ifdef NDEBUG - for (int id = 0; id < g_device_count; ++id) { + for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) { ggml_cuda_set_device(id); CUDA_CHECK(cudaDeviceSynchronize()); } - for (int id = 0; id < g_device_count; ++id) { + for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) { ggml_cuda_set_device(id); - for (int id_other = 0; id_other < g_device_count; ++id_other) { + for (int id_other = 0; id_other < ggml_backend_cuda_get_device_count(); ++id_other) { if (id == id_other) { continue; } - if (id != g_main_device && id_other != g_main_device) { + if (id != main_device && id_other != main_device) { continue; } @@ -9162,17 +1347,17 @@ static void ggml_cuda_set_peer_access(const int n_tokens) { } } } + + ggml_cuda_set_device(main_device); #endif // NDEBUG peer_access_enabled = enable_peer_access; + + GGML_UNUSED(main_device); } -// FIXME: move this somewhere else -struct ggml_backend_cuda_split_buffer_type_context { - std::array tensor_split; -}; - static void ggml_cuda_op_mul_mat( + ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_cuda_op_mul_mat_t op, const bool convert_src1_to_q8_1) { @@ -9195,8 +1380,11 @@ static void ggml_cuda_op_mul_mat( const int nb2 = dst->nb[2]; const int nb3 = dst->nb[3]; - GGML_ASSERT(dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT); - GGML_ASSERT(src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT); + GGML_ASSERT(ggml_backend_buffer_is_cuda(dst->buffer)); + GGML_ASSERT(ggml_backend_buffer_is_cuda(src1->buffer)); + ggml_backend_cuda_buffer_context * src1_ctx = (ggml_backend_cuda_buffer_context *) src1->buffer->context; + ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *) dst->buffer->context; + GGML_ASSERT(src1->type == GGML_TYPE_F32 || (src1->ne[2] == 1 && src1->ne[3] == 1)); GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0); @@ -9208,33 +1396,30 @@ static void ggml_cuda_op_mul_mat( const size_t q8_1_ts = sizeof(block_q8_1); const size_t q8_1_bs = QK8_1; - ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; - ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; - ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; - const bool src0_is_contiguous = ggml_is_contiguous(src0); const bool src1_is_contiguous = ggml_is_contiguous(src1); const int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING); - const bool split = src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT; + const bool split = ggml_backend_buffer_is_cuda_split(src0->buffer); GGML_ASSERT(!(split && ne02 > 1)); GGML_ASSERT(!(split && ne03 > 1)); GGML_ASSERT(!(split && ne02 < ne12)); + ggml_tensor_extra_gpu * src0_extra = split ? (ggml_tensor_extra_gpu *) src0->extra : nullptr; + + std::array tensor_split; if (split) { - // TODO: check that src0->buffer->buft is a split buffer type, replace GGML_BACKEND_TYPE_GPU_SPLIT check - // GGML_ASSERT(src0->buffer != nullptr && src0->buffer->buft == ...); ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context; tensor_split = buft_ctx->tensor_split; } struct dev_data { - cuda_pool_alloc src0_dd_alloc; - cuda_pool_alloc src1_ddf_alloc; - cuda_pool_alloc src1_ddq_alloc; - cuda_pool_alloc dst_dd_alloc; + ggml_cuda_pool_alloc src0_dd_alloc; + ggml_cuda_pool_alloc src1_ddf_alloc; + ggml_cuda_pool_alloc src1_ddq_alloc; + ggml_cuda_pool_alloc dst_dd_alloc; char * src0_dd = nullptr; float * src1_ddf = nullptr; // float @@ -9249,7 +1434,7 @@ static void ggml_cuda_op_mul_mat( int used_devices = 0; - for (int id = 0; id < g_device_count; ++id) { + for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) { // by default, use all rows dev[id].row_low = 0; dev[id].row_high = ne01; @@ -9266,7 +1451,7 @@ static void ggml_cuda_op_mul_mat( } } - if (id != g_device_count - 1) { + if (id != ggml_backend_cuda_get_device_count() - 1) { dev[id].row_high = ne01*tensor_split[id + 1]; if (dev[id].row_high < ne01) { dev[id].row_high -= dev[id].row_high % rounding; @@ -9275,33 +1460,33 @@ static void ggml_cuda_op_mul_mat( } } - for (int id = 0; id < g_device_count; ++id) { - if ((!split && id != g_main_device) || dev[id].row_low == dev[id].row_high) { + for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) { + if ((!split && id != ctx.device) || dev[id].row_low == dev[id].row_high) { continue; } used_devices++; - const bool src1_on_device = id == g_main_device; // TODO: check from buffer - const bool dst_on_device = id == g_main_device; + const bool src1_on_device = id == src1_ctx->device; + const bool dst_on_device = id == dst_ctx->device; ggml_cuda_set_device(id); - cudaStream_t stream = g_cudaStreams[id][0]; + cudaStream_t stream = ctx.stream(id, 0); if (src0_is_contiguous) { - dev[id].src0_dd = (char *) src0_extra->data_device[id]; + dev[id].src0_dd = split ? (char *) src0_extra->data_device[id] : (char *) src0->data; } else { - dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ggml_nbytes(src0)); + dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), ggml_nbytes(src0)); } if (src1_on_device && src1_is_contiguous) { - dev[id].src1_ddf = (float *) src1_extra->data_device[id]; + dev[id].src1_ddf = (float *) src1->data; } else { - dev[id].src1_ddf = dev[id].src1_ddf_alloc.alloc(ggml_nelements(src1)); + dev[id].src1_ddf = dev[id].src1_ddf_alloc.alloc(ctx.pool(id), ggml_nelements(src1)); } if (convert_src1_to_q8_1) { - dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs); + dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(ctx.pool(id), nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs); if (src1_on_device && src1_is_contiguous) { quantize_row_q8_1_cuda(dev[id].src1_ddf, dev[id].src1_ddq, ne10, nrows1, src1_padded_col_size, stream); @@ -9310,40 +1495,40 @@ static void ggml_cuda_op_mul_mat( } if (dst_on_device) { - dev[id].dst_dd = (float *) dst_extra->data_device[id]; + dev[id].dst_dd = (float *) dst->data; } else { const size_t size_dst_ddf = split ? (dev[id].row_high - dev[id].row_low)*ne1 : ggml_nelements(dst); - dev[id].dst_dd = dev[id].dst_dd_alloc.alloc(size_dst_ddf); + dev[id].dst_dd = dev[id].dst_dd_alloc.alloc(ctx.pool(id), size_dst_ddf); } } // if multiple devices are used they need to wait for the main device // here an event is recorded that signals that the main device has finished calculating the input data if (split && used_devices > 1) { - ggml_cuda_set_device(g_main_device); - CUDA_CHECK(cudaEventRecord(src0_extra->events[g_main_device][0], g_cudaStreams[g_main_device][0])); + ggml_cuda_set_device(ctx.device); + CUDA_CHECK(cudaEventRecord(src0_extra->events[ctx.device][0], ctx.stream())); } const int64_t src1_col_stride = split && used_devices > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11; for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) { - const int64_t is = split ? (src1_col_0/src1_col_stride) % MAX_STREAMS : 0; + const int64_t is = split ? (src1_col_0/src1_col_stride) % GGML_CUDA_MAX_STREAMS : 0; const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride; - for (int id = 0; id < g_device_count; ++id) { - if ((!split && id != g_main_device) || dev[id].row_low == dev[id].row_high) { + for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) { + if ((!split && id != ctx.device) || dev[id].row_low == dev[id].row_high) { continue; } - const bool src1_on_device = id == g_main_device; // TODO: check from buffer - const bool dst_on_device = id == g_main_device; + const bool src1_on_device = id == src1_ctx->device; + const bool dst_on_device = id == dst_ctx->device; const int64_t row_diff = dev[id].row_high - dev[id].row_low; ggml_cuda_set_device(id); - cudaStream_t stream = g_cudaStreams[id][is]; + cudaStream_t stream = ctx.stream(id, is); // wait for main GPU data if necessary - if (split && (id != g_main_device || is != 0)) { - CUDA_CHECK(cudaStreamWaitEvent(stream, src0_extra->events[g_main_device][0], 0)); + if (split && (id != ctx.device || is != 0)) { + CUDA_CHECK(cudaStreamWaitEvent(stream, src0_extra->events[ctx.device][0], 0)); } for (int64_t i0 = 0; i0 < ne13*ne12; ++i0) { @@ -9360,21 +1545,21 @@ static void ggml_cuda_op_mul_mat( // the main device memory buffer can be on VRAM scratch, with space for all partial results // in that case an offset on dst_ddf_i is needed - if (id == g_main_device) { + if (id == ctx.device) { dst_dd_i += dev[id].row_low; // offset is 0 if no tensor split } // copy src0, src1 to device if necessary if (src1_is_contiguous) { - if (id != g_main_device) { + if (id != ctx.device) { if (convert_src1_to_q8_1) { - char * src1_ddq_i_source = dev[g_main_device].src1_ddq + src1_ddq_i_offset; - CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddq_i, id, src1_ddq_i_source, g_main_device, + char * src1_ddq_i_source = dev[ctx.device].src1_ddq + src1_ddq_i_offset; + CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddq_i, id, src1_ddq_i_source, ctx.device, src1_ncols*src1_padded_col_size*q8_1_ts/q8_1_bs, stream)); } else { - float * src1_ddf_i_source = (float *) src1_extra->data_device[g_main_device]; + float * src1_ddf_i_source = (float *) src1->data; src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10; - CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddf_i, id, src1_ddf_i_source, g_main_device, + CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddf_i, id, src1_ddf_i_source, ctx.device, src1_ncols*ne10*sizeof(float), stream)); } } @@ -9395,13 +1580,13 @@ static void ggml_cuda_op_mul_mat( } // do the computation - op(src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i, + op(ctx, src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i, dev[id].row_low, dev[id].row_high, src1_ncols, src1_padded_col_size, stream); CUDA_CHECK(cudaGetLastError()); // copy dst to host or other device if necessary if (!dst_on_device) { - void * dst_off_device = dst_extra->data_device[g_main_device]; + void * dst_off_device = dst->data; if (split) { // src0 = weight matrix is saved as a transposed matrix for better memory layout. // dst is NOT transposed. @@ -9414,7 +1599,7 @@ static void ggml_cuda_op_mul_mat( #if !defined(GGML_USE_HIPBLAS) // cudaMemcpy2DAsync may fail with copies between vmm pools of different devices cudaMemcpy3DPeerParms p = {}; - p.dstDevice = g_main_device; + p.dstDevice = ctx.device; p.dstPtr = make_cudaPitchedPtr(dhf_dst_i, ne0*sizeof(float), row_diff, src1_ncols); p.srcDevice = id; p.srcPtr = make_cudaPitchedPtr(dst_dd_i, row_diff*sizeof(float), row_diff, src1_ncols); @@ -9436,7 +1621,7 @@ static void ggml_cuda_op_mul_mat( } // add event for the main device to wait on until other device is done - if (split && (id != g_main_device || is != 0)) { + if (split && (id != ctx.device || is != 0)) { CUDA_CHECK(cudaEventRecord(src0_extra->events[id][is], stream)); } } @@ -9444,130 +1629,25 @@ static void ggml_cuda_op_mul_mat( } // main device waits for all other devices to be finished - if (split && g_device_count > 1) { + if (split && ggml_backend_cuda_get_device_count() > 1) { int64_t is_max = (ne11 + MUL_MAT_SRC1_COL_STRIDE - 1) / MUL_MAT_SRC1_COL_STRIDE; - is_max = is_max <= MAX_STREAMS ? is_max : MAX_STREAMS; + is_max = is_max <= GGML_CUDA_MAX_STREAMS ? is_max : GGML_CUDA_MAX_STREAMS; - ggml_cuda_set_device(g_main_device); - for (int id = 0; id < g_device_count; ++id) { + ggml_cuda_set_device(ctx.device); + for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) { if (dev[id].row_low == dev[id].row_high) { continue; } for (int64_t is = 0; is < is_max; ++is) { - CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[g_main_device][0], src0_extra->events[id][is], 0)); + CUDA_CHECK(cudaStreamWaitEvent(ctx.stream(), src0_extra->events[id][is], 0)); } } } } -static void ggml_cuda_repeat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_repeat); -} - -static void ggml_cuda_get_rows(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_get_rows); -} - -static void ggml_cuda_add(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_add); -} - -static void ggml_cuda_acc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_acc); -} - -static void ggml_cuda_mul(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_mul); -} - -static void ggml_cuda_div(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_div); -} - -static void ggml_cuda_gelu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_gelu); -} - -static void ggml_cuda_silu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_silu); -} - -static void ggml_cuda_gelu_quick(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_gelu_quick); -} - -static void ggml_cuda_tanh(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_tanh); -} - -static void ggml_cuda_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_relu); -} - -static void ggml_cuda_hardsigmoid(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_hardsigmoid); -} - -static void ggml_cuda_hardswish(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_hardswish); -} -static void ggml_cuda_leaky_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_leaky_relu); -} - -static void ggml_cuda_sqr(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_sqr); -} - -static void ggml_cuda_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_norm); -} - -static void ggml_cuda_group_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_group_norm); -} - -static void ggml_cuda_concat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_concat); -} - -static void ggml_cuda_upscale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_upscale); -} - -static void ggml_cuda_pad(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_pad); -} - -static void ggml_cuda_arange(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; - - // dd = data device - float * src0_ddf = nullptr; - float * src1_ddf = nullptr; - float * dst_ddf = nullptr; - - ggml_cuda_set_device(g_main_device); - cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; - - dst_ddf = (float *) dst_extra->data_device[g_main_device]; - - // do the computation - ggml_cuda_op_arange(src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream); - CUDA_CHECK(cudaGetLastError()); -} - -static void ggml_cuda_timestep_embedding(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_timestep_embedding); -} - -static void ggml_cuda_rms_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_rms_norm); -} - -static void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){ +static void ggml_cuda_mul_mat_vec_p021(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){ GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1)); - GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT); + GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer)); GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation GGML_ASSERT(src0->type == GGML_TYPE_F16); @@ -9579,26 +1659,20 @@ static void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tens const int64_t ne12 = src1->ne[2]; - ggml_cuda_set_device(g_main_device); - cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; + cudaStream_t main_stream = ctx.stream(); - ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; - void * src0_ddq = src0_extra->data_device[g_main_device]; - - ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; - float * src1_ddf = (float *) src1_extra->data_device[g_main_device]; - - ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; - float * dst_ddf = (float *) dst_extra->data_device[g_main_device]; + void * src0_ddq = src0->data; + float * src1_ddf = (float *) src1->data; + float * dst_ddf = (float *) dst->data; ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream); } -static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){ +static void ggml_cuda_mul_mat_vec_nc(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){ GGML_ASSERT(!ggml_is_transposed(src0)); GGML_ASSERT(!ggml_is_transposed(src1)); GGML_ASSERT(!ggml_is_permuted(src0)); - GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT); + GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer)); GGML_ASSERT(src0->type == GGML_TYPE_F16); GGML_ASSERT(src1->type == GGML_TYPE_F32); @@ -9611,17 +1685,11 @@ static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor const int64_t ne12 = src1->ne[2]; - ggml_cuda_set_device(g_main_device); - cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; + cudaStream_t main_stream = ctx.stream(); - ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; - void * src0_ddq = src0_extra->data_device[g_main_device]; - - ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; - float * src1_ddf = (float *) src1_extra->data_device[g_main_device]; - - ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; - float * dst_ddf = (float *) dst_extra->data_device[g_main_device]; + void * src0_ddq = src0->data; + float * src1_ddf = (float *) src1->data; + float * dst_ddf = (float *) dst->data; const int64_t row_stride_x = nb01 / sizeof(half); const int64_t channel_stride_x = nb02 / sizeof(half); @@ -9653,34 +1721,28 @@ static __global__ void k_compute_batched_ptrs( ptrs_dst[0*ne23 + i12 + i13*ne12] = ( char *) dst + i12*nbd2 + i13*nbd3; } -static void ggml_cuda_mul_mat_batched_cublas(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { +static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_ASSERT(!ggml_is_transposed(src0)); GGML_ASSERT(!ggml_is_transposed(src1)); - GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT); + GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer)); GGML_ASSERT(src0->type == GGML_TYPE_F16); GGML_TENSOR_BINARY_OP_LOCALS const int64_t ne_dst = ggml_nelements(dst); - ggml_cuda_set_device(g_main_device); - cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; + cudaStream_t main_stream = ctx.stream(); - CUBLAS_CHECK(cublasSetStream(g_cublas_handles[g_main_device], main_stream)); + CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream)); - ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; - void * src0_ddq = src0_extra->data_device[g_main_device]; + void * src0_ddq = src0->data; half * src0_f16 = (half *) src0_ddq; - - ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; - float * src1_ddf = (float *) src1_extra->data_device[g_main_device]; - - ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; - float * dst_ddf = (float *) dst_extra->data_device[g_main_device]; + float * src1_ddf = (float *) src1->data; + float * dst_ddf = (float *) dst->data; // convert src1 to fp16 - cuda_pool_alloc src1_f16_alloc; + ggml_cuda_pool_alloc src1_f16_alloc(ctx.pool()); if (src1->type != GGML_TYPE_F16) { const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type); const int64_t ne_src1 = ggml_nelements(src1); @@ -9690,7 +1752,7 @@ static void ggml_cuda_mul_mat_batched_cublas(const ggml_tensor * src0, const ggm } half * src1_f16 = src1->type == GGML_TYPE_F16 ? (half *) src1_ddf : src1_f16_alloc.get(); - cuda_pool_alloc dst_f16; + ggml_cuda_pool_alloc dst_f16(ctx.pool()); char * dst_t; cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F; @@ -9755,7 +1817,7 @@ static void ggml_cuda_mul_mat_batched_cublas(const ggml_tensor * src0, const ggm // there is no broadcast and src0, src1 are contiguous across dims 2, 3 // use cublasGemmStridedBatchedEx CUBLAS_CHECK( - cublasGemmStridedBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N, + cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N, ne01, ne11, ne10, alpha, (const char *) src0_f16, CUDA_R_16F, nb01/nb00, nb02/nb00, // strideA (const char *) src1_f16, CUDA_R_16F, nb11/nb10, nb12/nb10, // strideB @@ -9767,8 +1829,8 @@ static void ggml_cuda_mul_mat_batched_cublas(const ggml_tensor * src0, const ggm // use cublasGemmBatchedEx const int ne23 = ne12*ne13; - cuda_pool_alloc ptrs_src(2*ne23); - cuda_pool_alloc< void *> ptrs_dst(1*ne23); + ggml_cuda_pool_alloc ptrs_src(ctx.pool(), 2*ne23); + ggml_cuda_pool_alloc< void *> ptrs_dst(ctx.pool(), 1*ne23); dim3 block_dims(ne13, ne12); k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>( @@ -9784,7 +1846,7 @@ static void ggml_cuda_mul_mat_batched_cublas(const ggml_tensor * src0, const ggm CUDA_CHECK(cudaGetLastError()); CUBLAS_CHECK( - cublasGemmBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N, + cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N, ne01, ne11, ne10, alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F, nb01/nb00, (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F, nb11/nb10, @@ -9801,8 +1863,8 @@ static void ggml_cuda_mul_mat_batched_cublas(const ggml_tensor * src0, const ggm } } -static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - const bool split = src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT; +static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + const bool split = ggml_backend_buffer_is_cuda_split(src0->buffer); int64_t min_compute_capability = INT_MAX; @@ -9810,22 +1872,22 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1 if (split) { ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context; auto & tensor_split = buft_ctx->tensor_split; - for (int id = 0; id < g_device_count; ++id) { + for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) { // skip devices that are not going to do any work: - if (tensor_split[id] >= (id + 1 < g_device_count ? tensor_split[id + 1] : 1.0f)) { + if (tensor_split[id] >= (id + 1 < ggml_backend_cuda_get_device_count() ? tensor_split[id + 1] : 1.0f)) { continue; } - if (min_compute_capability > g_device_caps[id].cc) { - min_compute_capability = g_device_caps[id].cc; + if (min_compute_capability > ggml_cuda_info().devices[id].cc) { + min_compute_capability = ggml_cuda_info().devices[id].cc; } - if (g_device_caps[id].cc == 610) { + if (ggml_cuda_info().devices[id].cc == 610) { any_pascal_with_slow_fp16 = true; } } } else { - min_compute_capability = g_device_caps[g_main_device].cc; - any_pascal_with_slow_fp16 = g_device_caps[g_main_device].cc == 610; + min_compute_capability = ggml_cuda_info().devices[ctx.device].cc; + any_pascal_with_slow_fp16 = ggml_cuda_info().devices[ctx.device].cc == 610; } // check data types and tensor shapes for custom matrix multiplication kernels: @@ -9880,21 +1942,21 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1 if (!split && !fp16_performance_good && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) { // KQ single-batch - ggml_cuda_mul_mat_vec_p021(src0, src1, dst); + ggml_cuda_mul_mat_vec_p021(ctx, src0, src1, dst); } else if (!split && !fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) { // KQV single-batch - ggml_cuda_mul_mat_vec_nc(src0, src1, dst); + ggml_cuda_mul_mat_vec_nc(ctx, src0, src1, dst); } else if (!split && fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) { // KQ + KQV multi-batch - ggml_cuda_mul_mat_batched_cublas(src0, src1, dst); + ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst); } else if (use_dequantize_mul_mat_vec) { - ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false); + ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false); } else if (use_mul_mat_vec_q) { - ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true); + ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true); } else if (use_mul_mat_q) { - ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_q, true); + ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_q, true); } else { - ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false); + ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false); } } @@ -10079,12 +2141,15 @@ static void ggml_cuda_mul_mat_id_cublas(ggml_tensor * dst) { } #endif -static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { +static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { #if 0 ggml_cuda_mul_mat_id_cublas(dst); // TODO: mmq/mmv support #endif - cudaStream_t stream = g_cudaStreams[g_main_device][0]; + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + cudaStream_t stream = ctx.stream(); const size_t nb11 = src1->nb[1]; const size_t nb1 = dst->nb[1]; @@ -10094,27 +2159,15 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s const int32_t n_as = ((int32_t *) dst->op_params)[1]; std::vector ids_host(ggml_nbytes(ids)); - const char * ids_dev = (const char *)((const ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device]; + const char * ids_dev = (const char *) ids->data; CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream)); CUDA_CHECK(cudaStreamSynchronize(stream)); - const ggml_tensor_extra_gpu * src1_extra = (const ggml_tensor_extra_gpu *) src1->extra; - const ggml_tensor_extra_gpu * dst_extra = (const ggml_tensor_extra_gpu *) dst->extra; - - ggml_tensor_extra_gpu src1_row_extra; - ggml_tensor_extra_gpu dst_row_extra; - ggml_tensor src1_row = *src1; ggml_tensor dst_row = *dst; - src1_row.backend = GGML_BACKEND_TYPE_GPU; - dst_row.backend = GGML_BACKEND_TYPE_GPU; - - src1_row.extra = &src1_row_extra; - dst_row.extra = &dst_row_extra; - - char * src1_original = (char *) src1_extra->data_device[g_main_device]; - char * dst_original = (char *) dst_extra->data_device[g_main_device]; + char * src1_original = (char *) src1->data; + char * dst_original = (char *) dst->data; if (src1->ne[1] == 1) { for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) { @@ -10124,20 +2177,17 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s const struct ggml_tensor * src0_row = dst->src[row_id + 2]; - src1_row_extra.data_device[g_main_device] = src1_original + i01*src1->nb[1]; - src1_row.data = (char *) src1->data + i01*src1->nb[1]; // TODO why is this set? + src1_row.data = src1_original + i01*src1->nb[1]; + dst_row.data = dst_original + i01*dst->nb[1]; - dst_row_extra.data_device[g_main_device] = dst_original + i01*dst->nb[1]; - dst_row.data = (char *) dst->data + i01*dst->nb[1]; // TODO why is this set? - - ggml_cuda_mul_mat(src0_row, &src1_row, &dst_row); + ggml_cuda_mul_mat(ctx, src0_row, &src1_row, &dst_row); } } else { - cuda_pool_alloc src1_contiguous(sizeof(float)*ggml_nelements(src1)); - cuda_pool_alloc dst_contiguous(sizeof(float)*ggml_nelements(dst)); + ggml_cuda_pool_alloc src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1)); + ggml_cuda_pool_alloc dst_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(dst)); - src1_row_extra.data_device[g_main_device] = src1_contiguous.get(); - dst_row_extra.data_device[g_main_device] = dst_contiguous.get(); + src1_row.data = src1_contiguous.get(); + dst_row.data = dst_contiguous.get(); for (int32_t row_id = 0; row_id < n_as; ++row_id) { const struct ggml_tensor * src0_row = dst->src[row_id + 2]; @@ -10172,7 +2222,7 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s dst_row.nb[2] = num_src1_rows*nb1; dst_row.nb[3] = num_src1_rows*nb1; - ggml_cuda_mul_mat(src0_row, &src1_row, &dst_row); + ggml_cuda_mul_mat(ctx, src0_row, &src1_row, &dst_row); num_src1_rows = 0; for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) { @@ -10192,927 +2242,159 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s } } -static void ggml_cuda_scale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_scale); -} - -static void ggml_cuda_clamp(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_clamp); -} - -static void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - const int64_t ne = ggml_nelements(src0); - GGML_ASSERT(ne == ggml_nelements(src1)); - - GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU); - GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU); - - GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX); - GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX); - - const int64_t ne00 = src0->ne[0]; - const int64_t ne01 = src0->ne[1]; - const int64_t ne02 = src0->ne[2]; - - //GGML_ASSERT(src0->ne[3] == 1); - - const int64_t nb00 = src0->nb[0]; - const int64_t nb01 = src0->nb[1]; - const int64_t nb02 = src0->nb[2]; - const int64_t nb03 = src0->nb[3]; - - const int64_t ne10 = src1->ne[0]; - const int64_t ne11 = src1->ne[1]; - const int64_t ne12 = src1->ne[2]; - - //GGML_ASSERT(src1->ne[3] == 1); - - const int64_t nb10 = src1->nb[0]; - const int64_t nb11 = src1->nb[1]; - const int64_t nb12 = src1->nb[2]; - const int64_t nb13 = src1->nb[3]; - - ggml_cuda_set_device(g_main_device); - cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; - - const ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; - const ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; - - char * src0_ddc = (char *) src0_extra->data_device[g_main_device]; - char * src1_ddc = (char *) src1_extra->data_device[g_main_device]; - - if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) { - ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) { - ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) { - ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) { - ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) { - ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) { - ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) { - ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else { - fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__, - ggml_type_name(src0->type), ggml_type_name(src1->type)); - GGML_ASSERT(false); +static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct ggml_tensor * dst) { + // why is this here instead of mul_mat? + if (dst->src[0] != nullptr && ggml_backend_buffer_is_cuda_split(dst->src[0]->buffer)) { + ggml_cuda_set_peer_access(dst->src[1]->ne[1], ctx.device); } - (void) dst; -} - -static void ggml_cuda_dup(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - // TODO: why do we pass dst as src1 here? - ggml_cuda_cpy(src0, dst, nullptr); - (void) src1; -} - -static void ggml_cuda_diag_mask_inf(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_diag_mask_inf); -} - -static void ggml_cuda_soft_max(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_soft_max); -} - -static void ggml_cuda_rope(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - GGML_ASSERT(ggml_is_contiguous(src0)); // TODO: this restriction is temporary until non-cont support is implemented - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_rope); -} - -static void ggml_cuda_alibi(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_alibi); -} - -static void ggml_cuda_pool2d(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_pool2d); -} - -static void ggml_cuda_im2col(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_im2col); -} - -static void ggml_cuda_sum_rows(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - GGML_ASSERT(ggml_is_contiguous(src0)); - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_sum_rows); -} - -static void ggml_cuda_argsort(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - GGML_ASSERT(ggml_is_contiguous(src0)); - ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_argsort); -} - -static void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - (void) src0; - (void) src1; - (void) dst; -} - -static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) { - static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); - - return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]); -} - -static void ggml_cuda_set_main_device(const int main_device) { - if (main_device >= g_device_count) { - fprintf(stderr, "warning: cannot set main_device=%d because there are only %d devices. Using device %d instead.\n", - main_device, g_device_count, g_main_device); - return; - } - - if (g_main_device != main_device && g_device_count > 1) { - g_main_device = main_device; - //cudaDeviceProp prop; - //CUDA_CHECK(cudaGetDeviceProperties(&prop, g_main_device)); - //fprintf(stderr, "%s: using device %d (%s) as main device\n", __func__, g_main_device, prop.name); - } -} - -static bool ggml_cuda_compute_forward(struct ggml_tensor * tensor) { - if (!g_cublas_loaded) return false; - - if (tensor->op == GGML_OP_MUL_MAT) { - if (tensor->src[0]->ne[3] != tensor->src[1]->ne[3]) { -#ifndef NDEBUG - fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, tensor->name, tensor->src[0]->ne[3], tensor->src[1]->ne[3]); -#endif - return false; - } - } - - ggml_cuda_func_t func; - - switch (tensor->op) { + switch (dst->op) { case GGML_OP_REPEAT: - func = ggml_cuda_repeat; + ggml_cuda_op_repeat(ctx, dst); break; case GGML_OP_GET_ROWS: - func = ggml_cuda_get_rows; + ggml_cuda_op_get_rows(ctx, dst); break; case GGML_OP_DUP: - func = ggml_cuda_dup; + ggml_cuda_dup(ctx, dst); + break; + case GGML_OP_CPY: + ggml_cuda_cpy(ctx, dst->src[0], dst->src[1]); + break; + case GGML_OP_CONT: + ggml_cuda_dup(ctx, dst); break; case GGML_OP_ADD: - func = ggml_cuda_add; + ggml_cuda_op_add(ctx, dst); break; case GGML_OP_ACC: - func = ggml_cuda_acc; + ggml_cuda_op_acc(ctx, dst); break; case GGML_OP_MUL: - func = ggml_cuda_mul; + ggml_cuda_op_mul(ctx, dst); break; case GGML_OP_DIV: - func = ggml_cuda_div; + ggml_cuda_op_div(ctx, dst); break; case GGML_OP_UNARY: - switch (ggml_get_unary_op(tensor)) { + switch (ggml_get_unary_op(dst)) { case GGML_UNARY_OP_GELU: - func = ggml_cuda_gelu; + ggml_cuda_op_gelu(ctx, dst); break; case GGML_UNARY_OP_SILU: - func = ggml_cuda_silu; + ggml_cuda_op_silu(ctx, dst); break; case GGML_UNARY_OP_GELU_QUICK: - func = ggml_cuda_gelu_quick; + ggml_cuda_op_gelu_quick(ctx, dst); break; case GGML_UNARY_OP_TANH: - func = ggml_cuda_tanh; + ggml_cuda_op_tanh(ctx, dst); break; case GGML_UNARY_OP_RELU: - func = ggml_cuda_relu; + ggml_cuda_op_relu(ctx, dst); break; case GGML_UNARY_OP_HARDSIGMOID: - func = ggml_cuda_hardsigmoid; + ggml_cuda_op_hardsigmoid(ctx, dst); break; case GGML_UNARY_OP_HARDSWISH: - func = ggml_cuda_hardswish; + ggml_cuda_op_hardswish(ctx, dst); break; default: return false; } break; case GGML_OP_NORM: - func = ggml_cuda_norm; + ggml_cuda_op_norm(ctx, dst); break; case GGML_OP_GROUP_NORM: - func = ggml_cuda_group_norm; + ggml_cuda_op_group_norm(ctx, dst); break; case GGML_OP_CONCAT: - func = ggml_cuda_concat; + ggml_cuda_op_concat(ctx, dst); break; case GGML_OP_UPSCALE: - func = ggml_cuda_upscale; + ggml_cuda_op_upscale(ctx, dst); break; case GGML_OP_PAD: - func = ggml_cuda_pad; + ggml_cuda_op_pad(ctx, dst); break; case GGML_OP_ARANGE: - func = ggml_cuda_arange; + ggml_cuda_op_arange(ctx, dst); break; case GGML_OP_TIMESTEP_EMBEDDING: - func = ggml_cuda_timestep_embedding; + ggml_cuda_op_timestep_embedding(ctx, dst); break; case GGML_OP_LEAKY_RELU: - func = ggml_cuda_leaky_relu; + ggml_cuda_op_leaky_relu(ctx, dst); break; case GGML_OP_RMS_NORM: - func = ggml_cuda_rms_norm; + ggml_cuda_op_rms_norm(ctx, dst); break; case GGML_OP_MUL_MAT: - func = ggml_cuda_mul_mat; + if (dst->src[0]->ne[3] != dst->src[1]->ne[3]) { + fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, dst->name, dst->src[0]->ne[3], dst->src[1]->ne[3]); + return false; + } else { + ggml_cuda_mul_mat(ctx, dst->src[0], dst->src[1], dst); + } break; case GGML_OP_MUL_MAT_ID: - func = ggml_cuda_mul_mat_id; + ggml_cuda_mul_mat_id(ctx, dst); break; case GGML_OP_SCALE: - func = ggml_cuda_scale; + ggml_cuda_op_scale(ctx, dst); break; case GGML_OP_SQR: - func = ggml_cuda_sqr; + ggml_cuda_op_sqr(ctx, dst); break; case GGML_OP_CLAMP: - func = ggml_cuda_clamp; - break; - case GGML_OP_CPY: - func = ggml_cuda_cpy; - break; - case GGML_OP_CONT: - func = ggml_cuda_dup; + ggml_cuda_op_clamp(ctx, dst); break; case GGML_OP_NONE: case GGML_OP_RESHAPE: case GGML_OP_VIEW: case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: - func = ggml_cuda_nop; - break; + break; case GGML_OP_DIAG_MASK_INF: - func = ggml_cuda_diag_mask_inf; + ggml_cuda_op_diag_mask_inf(ctx, dst); break; case GGML_OP_SOFT_MAX: - func = ggml_cuda_soft_max; + ggml_cuda_op_soft_max(ctx, dst); break; case GGML_OP_ROPE: - func = ggml_cuda_rope; + ggml_cuda_op_rope(ctx, dst); break; case GGML_OP_ALIBI: - func = ggml_cuda_alibi; + ggml_cuda_op_alibi(ctx, dst); break; case GGML_OP_IM2COL: - func = ggml_cuda_im2col; + ggml_cuda_op_im2col(ctx, dst); break; case GGML_OP_POOL_2D: - func = ggml_cuda_pool2d; + ggml_cuda_op_pool2d(ctx, dst); break; case GGML_OP_SUM_ROWS: - func = ggml_cuda_sum_rows; + ggml_cuda_op_sum_rows(ctx, dst); break; case GGML_OP_ARGSORT: - func = ggml_cuda_argsort; + ggml_cuda_op_argsort(ctx, dst); break; default: return false; } - if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT) { - ggml_cuda_set_peer_access(tensor->src[1]->ne[1]); + cudaError_t err = cudaGetLastError(); + if (err != cudaSuccess) { + fprintf(stderr, "%s: %s failed\n", __func__, ggml_op_desc(dst)); + GGML_ASSERT(false); } - func(tensor->src[0], tensor->src[1], tensor); return true; } -static int ggml_cuda_get_device_count() { - int device_count; - if (cudaGetDeviceCount(&device_count) != cudaSuccess) { - return 0; - } - return device_count; -} - -static void ggml_cuda_get_device_description(int device, char * description, size_t description_size) { - cudaDeviceProp prop; - CUDA_CHECK(cudaGetDeviceProperties(&prop, device)); - snprintf(description, description_size, "%s", prop.name); -} - //////////////////////////////////////////////////////////////////////////////// -// backend interface - -#define UNUSED GGML_UNUSED - -struct ggml_backend_cuda_context { - explicit ggml_backend_cuda_context(int device) : - device(device), - name(GGML_CUDA_NAME + std::to_string(device)) { - } - - ~ggml_backend_cuda_context() { - if (copy_event != nullptr) { - CUDA_CHECK(cudaEventDestroy(copy_event)); - } - } - - int device; - std::string name; - cudaEvent_t copy_event = nullptr; -}; - -// cuda buffer - -struct ggml_backend_cuda_buffer_context { - int device; - void * dev_ptr = nullptr; - ggml_tensor_extra_gpu * temp_tensor_extras = nullptr; - size_t temp_tensor_extra_index = 0; - std::string name; - - ggml_backend_cuda_buffer_context(int device, void * dev_ptr) : - device(device), dev_ptr(dev_ptr), - name(GGML_CUDA_NAME + std::to_string(device)) { - } - - ~ggml_backend_cuda_buffer_context() { - delete[] temp_tensor_extras; - } - - ggml_tensor_extra_gpu * ggml_cuda_alloc_temp_tensor_extra() { - // TODO: remove GGML_CUDA_MAX_NODES, allocate dynamically and reuse in backend_buffer_reset - if (temp_tensor_extras == nullptr) { - temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_CUDA_MAX_NODES]; - } - - size_t alloc_index = temp_tensor_extra_index; - temp_tensor_extra_index = (temp_tensor_extra_index + 1) % GGML_CUDA_MAX_NODES; - ggml_tensor_extra_gpu * extra = &temp_tensor_extras[alloc_index]; - memset(extra, 0, sizeof(*extra)); - - return extra; - } -}; - -GGML_CALL static const char * ggml_backend_cuda_buffer_get_name(ggml_backend_buffer_t buffer) { - ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; - return ctx->name.c_str(); -} - -GGML_CALL static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) { - return buffer->iface.get_name == ggml_backend_cuda_buffer_get_name; -} - -GGML_CALL static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) { - ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; - CUDA_CHECK(cudaFree(ctx->dev_ptr)); - delete ctx; -} - -GGML_CALL static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) { - ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; - return ctx->dev_ptr; -} - -GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { - ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; - - if (tensor->view_src != NULL && tensor->view_offs == 0) { - assert(tensor->view_src->buffer->buft == buffer->buft); - tensor->backend = tensor->view_src->backend; - tensor->extra = tensor->view_src->extra; - return; - } - - ggml_tensor_extra_gpu * extra = ctx->ggml_cuda_alloc_temp_tensor_extra(); - - extra->data_device[ctx->device] = tensor->data; - - tensor->backend = GGML_BACKEND_TYPE_GPU; - tensor->extra = extra; - - if (ggml_is_quantized(tensor->type)) { - // initialize padding to 0 to avoid possible NaN values - size_t original_size = ggml_nbytes(tensor); - size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor); - - if (padded_size > original_size && tensor->view_src == nullptr) { - ggml_cuda_set_device(ctx->device); - CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size)); - } - } -} - -GGML_CALL static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { - GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU); - - ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; - - ggml_cuda_set_device(ctx->device); - CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cudaStreamPerThread)); - CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); -} - -GGML_CALL static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { - GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU); - - ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; - - ggml_cuda_set_device(ctx->device); - CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cudaStreamPerThread)); - CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); -} - -GGML_CALL static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) { - if (ggml_backend_buffer_is_cuda(src->buffer)) { - ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src->buffer->context; - ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)dst->buffer->context; - if (src_ctx->device == dst_ctx->device) { - CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(src), cudaMemcpyDeviceToDevice, cudaStreamPerThread)); - } else { - CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, dst_ctx->device, src->data, src_ctx->device, ggml_nbytes(src), cudaStreamPerThread)); - } - CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); - return true; - } - return false; - - UNUSED(buffer); -} - -GGML_CALL static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { - ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; - - ggml_cuda_set_device(ctx->device); - CUDA_CHECK(cudaDeviceSynchronize()); - CUDA_CHECK(cudaMemset(ctx->dev_ptr, value, buffer->size)); - CUDA_CHECK(cudaDeviceSynchronize()); -} - -static ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = { - /* .get_name = */ ggml_backend_cuda_buffer_get_name, - /* .free_buffer = */ ggml_backend_cuda_buffer_free_buffer, - /* .get_base = */ ggml_backend_cuda_buffer_get_base, - /* .init_tensor = */ ggml_backend_cuda_buffer_init_tensor, - /* .set_tensor = */ ggml_backend_cuda_buffer_set_tensor, - /* .get_tensor = */ ggml_backend_cuda_buffer_get_tensor, - /* .cpy_tensor = */ ggml_backend_cuda_buffer_cpy_tensor, - /* .clear = */ ggml_backend_cuda_buffer_clear, - /* .reset = */ NULL, -}; - -// cuda buffer type -struct ggml_backend_cuda_buffer_type_context { - int device; - std::string name; -}; - -GGML_CALL static const char * ggml_backend_cuda_buffer_type_name(ggml_backend_buffer_type_t buft) { - ggml_backend_cuda_buffer_type_context * ctx = (ggml_backend_cuda_buffer_type_context *)buft->context; - - return ctx->name.c_str(); -} - -GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { - ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context; - - ggml_cuda_set_device(buft_ctx->device); - - size = std::max(size, (size_t)1); // cudaMalloc returns null for size 0 - - void * dev_ptr; - cudaError_t err = cudaMalloc(&dev_ptr, size); - if (err != cudaSuccess) { - fprintf(stderr, "%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size/1024.0/1024.0, buft_ctx->device, cudaGetErrorString(err)); - return nullptr; - } - - ggml_backend_cuda_buffer_context * ctx = new ggml_backend_cuda_buffer_context(buft_ctx->device, dev_ptr); - - return ggml_backend_buffer_init(buft, ggml_backend_cuda_buffer_interface, ctx, size); -} - -GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { - return 128; - - UNUSED(buft); -} - -GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { - size_t size = ggml_nbytes(tensor); - int64_t ne0 = tensor->ne[0]; - - if (ggml_is_quantized(tensor->type)) { - if (ne0 % MATRIX_ROW_PADDING != 0) { - size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); - } - } - - return size; - - UNUSED(buft); -} - -GGML_CALL static bool ggml_backend_cuda_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) { - if (!ggml_backend_is_cuda(backend)) { - return false; - } - - ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context; - ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; - - return buft_ctx->device == cuda_ctx->device; -} - -static ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = { - /* .get_name = */ ggml_backend_cuda_buffer_type_name, - /* .alloc_buffer = */ ggml_backend_cuda_buffer_type_alloc_buffer, - /* .get_alignment = */ ggml_backend_cuda_buffer_type_get_alignment, - /* .get_max_size = */ NULL, // defaults to SIZE_MAX - /* .get_alloc_size = */ ggml_backend_cuda_buffer_type_get_alloc_size, - /* .supports_backend = */ ggml_backend_cuda_buffer_type_supports_backend, - /* .is_host = */ NULL, -}; - -GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) { - ggml_init_cublas(); - - // FIXME: this is not thread safe - if (device >= ggml_backend_cuda_get_device_count()) { - return nullptr; - } - - static ggml_backend_buffer_type ggml_backend_cuda_buffer_types[GGML_CUDA_MAX_DEVICES]; - - static bool ggml_backend_cuda_buffer_type_initialized = false; - - if (!ggml_backend_cuda_buffer_type_initialized) { - for (int i = 0; i < GGML_CUDA_MAX_DEVICES; i++) { - ggml_backend_cuda_buffer_types[i] = { - /* .iface = */ ggml_backend_cuda_buffer_type_interface, - /* .context = */ new ggml_backend_cuda_buffer_type_context{i, GGML_CUDA_NAME + std::to_string(i)}, - }; - } - ggml_backend_cuda_buffer_type_initialized = true; - } - - return &ggml_backend_cuda_buffer_types[device]; -} - -// cuda split buffer - -struct ggml_backend_cuda_split_buffer_context { - ~ggml_backend_cuda_split_buffer_context() { - for (ggml_tensor_extra_gpu * extra : tensor_extras) { - for (int id = 0; id < g_device_count; ++id) { - for (int64_t is = 0; is < MAX_STREAMS; ++is) { - if (extra->events[id][is] != nullptr) { - CUDA_CHECK(cudaEventDestroy(extra->events[id][is])); - } - } - if (extra->data_device[id] != nullptr) { - CUDA_CHECK(cudaFree(extra->data_device[id])); - } - } - delete extra; - } - } - - std::vector tensor_extras; -}; - -GGML_CALL static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backend_buffer_t buffer) { - return GGML_CUDA_NAME "_Split"; - - UNUSED(buffer); -} - -static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) { - return buffer->iface.get_name == ggml_backend_cuda_split_buffer_get_name; - UNUSED(ggml_backend_buffer_is_cuda_split); // only used in debug builds currently, avoid unused function warning in release builds -} - -GGML_CALL static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) { - ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context; - delete ctx; -} - -GGML_CALL static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) { - // the pointers are stored in the tensor extras, this is just a dummy address and never dereferenced - return (void *)0x1000; - - UNUSED(buffer); -} - -GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { - GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported - - ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context; - ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context; - - const int64_t ne0 = tensor->ne[0]; - - ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{}; - - ctx->tensor_extras.push_back(extra); - - for (int id = 0; id < g_device_count; ++id) { - int64_t row_low, row_high; - get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id); - - int64_t nrows_split = row_high - row_low; - if (nrows_split == 0) { - continue; - } - - size_t size = ggml_nbytes_split(tensor, nrows_split); - const size_t original_size = size; - - // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses - if (ne0 % MATRIX_ROW_PADDING != 0) { - size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); - } - - // FIXME: do not crash if cudaMalloc fails - // currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first - ggml_cuda_set_device(id); - char * buf; - CUDA_CHECK(cudaMalloc(&buf, size)); - - // set padding to 0 to avoid possible NaN values - if (size > original_size) { - CUDA_CHECK(cudaMemset(buf + original_size, 0, size - original_size)); - } - - extra->data_device[id] = buf; - - for (int64_t is = 0; is < MAX_STREAMS; ++is) { - CUDA_CHECK(cudaEventCreateWithFlags(&extra->events[id][is], cudaEventDisableTiming)); - } - } - tensor->backend = GGML_BACKEND_TYPE_GPU_SPLIT; - tensor->extra = extra; -} - -GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { - // split tensors must always be set in their entirety at once - GGML_ASSERT(offset == 0); - GGML_ASSERT(size == ggml_nbytes(tensor)); - - ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context; - - const int64_t ne0 = tensor->ne[0]; - const size_t nb1 = tensor->nb[1]; - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra; - - for (int id = 0; id < g_device_count; ++id) { - int64_t row_low, row_high; - get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id); - - int64_t nrows_split = row_high - row_low; - if (nrows_split == 0) { - continue; - } - - const size_t offset_split = row_low*nb1; - size_t size = ggml_nbytes_split(tensor, nrows_split); - const size_t original_size = size; - - // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses - if (ne0 % MATRIX_ROW_PADDING != 0) { - size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); - } - - const char * buf_host = (const char *)data + offset_split; - CUDA_CHECK(cudaMemcpyAsync(extra->data_device[id], buf_host, original_size, cudaMemcpyHostToDevice, cudaStreamPerThread)); - } - - for (int id = 0; id < g_device_count; ++id) { - CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); - } -} - -GGML_CALL static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { - // split tensors must always be set in their entirety at once - GGML_ASSERT(offset == 0); - GGML_ASSERT(size == ggml_nbytes(tensor)); - - ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context; - - const int64_t ne0 = tensor->ne[0]; - const size_t nb1 = tensor->nb[1]; - ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra; - - for (int id = 0; id < g_device_count; ++id) { - int64_t row_low, row_high; - get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id); - - int64_t nrows_split = row_high - row_low; - if (nrows_split == 0) { - continue; - } - - const size_t offset_split = row_low*nb1; - size_t size = ggml_nbytes_split(tensor, nrows_split); - const size_t original_size = size; - - // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses - if (ne0 % MATRIX_ROW_PADDING != 0) { - size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); - } - - char * buf_host = (char *)data + offset_split; - CUDA_CHECK(cudaMemcpyAsync(buf_host, extra->data_device[id], original_size, cudaMemcpyDeviceToHost, cudaStreamPerThread)); - } - - for (int id = 0; id < g_device_count; ++id) { - CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); - } -} - -GGML_CALL static void ggml_backend_cuda_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { - UNUSED(buffer); - UNUSED(value); -} - -static struct ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = { - /* .get_name = */ ggml_backend_cuda_split_buffer_get_name, - /* .free_buffer = */ ggml_backend_cuda_split_buffer_free_buffer, - /* .get_base = */ ggml_backend_cuda_split_buffer_get_base, - /* .init_tensor = */ ggml_backend_cuda_split_buffer_init_tensor, - /* .set_tensor = */ ggml_backend_cuda_split_buffer_set_tensor, - /* .get_tensor = */ ggml_backend_cuda_split_buffer_get_tensor, - /* .cpy_tensor = */ NULL, - /* .clear = */ ggml_backend_cuda_split_buffer_clear, - /* .reset = */ NULL, -}; - -// cuda split buffer type - -GGML_CALL static const char * ggml_backend_cuda_split_buffer_type_name(ggml_backend_buffer_type_t buft) { - return GGML_CUDA_NAME "_Split"; - - UNUSED(buft); -} - -GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { - // since we don't know the exact split after rounding, we cannot allocate the device buffers at this point - // instead, we allocate them for each tensor separately in init_tensor - // however, the size still represents the maximum cumulative size of all the device buffers after the tensors are allocated, - // as returned by get_alloc_size. this limit is enforced during tensor allocation by ggml-alloc, so it must be correct. - ggml_backend_cuda_split_buffer_context * ctx = new ggml_backend_cuda_split_buffer_context(); - - return ggml_backend_buffer_init(buft, ggml_backend_cuda_split_buffer_interface, ctx, size); -} - -GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { - return 128; - - UNUSED(buft); -} - -GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { - ggml_backend_cuda_split_buffer_type_context * ctx = (ggml_backend_cuda_split_buffer_type_context *)buft->context; - - size_t total_size = 0; - - const int64_t ne0 = tensor->ne[0]; - - for (int id = 0; id < g_device_count; ++id) { - int64_t row_low, row_high; - get_row_split(&row_low, &row_high, tensor, ctx->tensor_split, id); - - int64_t nrows_split = row_high - row_low; - if (nrows_split == 0) { - continue; - } - - total_size += ggml_nbytes_split(tensor, nrows_split); - - // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses - if (ne0 % MATRIX_ROW_PADDING != 0) { - total_size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); - } - } - - return total_size; -} - -GGML_CALL static bool ggml_backend_cuda_split_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) { - return ggml_backend_is_cuda(backend); - - UNUSED(buft); -} - -GGML_CALL static bool ggml_backend_cuda_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) { - return false; - - UNUSED(buft); -} - -static ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface = { - /* .get_name = */ ggml_backend_cuda_split_buffer_type_name, - /* .alloc_buffer = */ ggml_backend_cuda_split_buffer_type_alloc_buffer, - /* .get_alignment = */ ggml_backend_cuda_split_buffer_type_get_alignment, - /* .get_max_size = */ NULL, // defaults to SIZE_MAX - /* .get_alloc_size = */ ggml_backend_cuda_split_buffer_type_get_alloc_size, - /* .supports_backend = */ ggml_backend_cuda_split_buffer_type_supports_backend, - /* .is_host = */ ggml_backend_cuda_split_buffer_type_is_host, -}; - -GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split) { - ggml_init_cublas(); - - // FIXME: this is not thread safe - static std::map, struct ggml_backend_buffer_type> buft_map; - - std::array tensor_split_arr = {}; - - bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + GGML_CUDA_MAX_DEVICES, [](float x) { return x == 0.0f; }); - if (all_zero) { - tensor_split_arr = g_default_tensor_split; - } else { - float split_sum = 0.0f; - for (int i = 0; i < g_device_count; ++i) { - tensor_split_arr[i] = split_sum; - split_sum += tensor_split[i]; - } - for (int i = 0; i < g_device_count; ++i) { - tensor_split_arr[i] /= split_sum; - } - } - - auto it = buft_map.find(tensor_split_arr); - if (it != buft_map.end()) { - return &it->second; - } - - struct ggml_backend_buffer_type buft { - /* .iface = */ ggml_backend_cuda_split_buffer_type_interface, - /* .context = */ new ggml_backend_cuda_split_buffer_type_context{tensor_split_arr}, - }; - - auto result = buft_map.emplace(tensor_split_arr, buft); - return &result.first->second; -} - -// host buffer type - -GGML_CALL static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_type_t buft) { - return GGML_CUDA_NAME "_Host"; - - UNUSED(buft); -} - -GGML_CALL static const char * ggml_backend_cuda_host_buffer_name(ggml_backend_buffer_t buffer) { - return GGML_CUDA_NAME "_Host"; - - UNUSED(buffer); -} - -GGML_CALL static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) { - ggml_cuda_host_free(buffer->context); -} - -GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { - void * ptr = ggml_cuda_host_malloc(size); - - if (ptr == nullptr) { - // fallback to cpu buffer - return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size); - } - - ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size); - buffer->buft = buft; - buffer->iface.get_name = ggml_backend_cuda_host_buffer_name; - buffer->iface.free_buffer = ggml_backend_cuda_host_buffer_free_buffer; - - return buffer; -} - -GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() { - static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_type_host = { - /* .iface = */ { - /* .get_name = */ ggml_backend_cuda_host_buffer_type_name, - /* .alloc_buffer = */ ggml_backend_cuda_host_buffer_type_alloc_buffer, - /* .get_alignment = */ ggml_backend_cpu_buffer_type()->iface.get_alignment, - /* .get_max_size = */ NULL, // defaults to SIZE_MAX - /* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size, - /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend, - /* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host, - }, - /* .context = */ nullptr, - }; - - return &ggml_backend_cuda_buffer_type_host; -} - -//static bool ggml_backend_buffer_is_cuda_host(ggml_backend_buffer_t buffer) { -// return buffer->buft->iface.get_name == ggml_backend_cuda_host_buffer_type_name; -//} - // backend GGML_CALL static const char * ggml_backend_cuda_name(ggml_backend_t backend) { @@ -11139,9 +2421,8 @@ GGML_CALL static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type"); - GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU); - CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, g_cudaStreams[cuda_ctx->device][0])); + CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cuda_ctx->stream())); } GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) { @@ -11149,9 +2430,8 @@ GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type"); - GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU); - CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, g_cudaStreams[cuda_ctx->device][0])); + CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cuda_ctx->stream())); } GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) { @@ -11179,26 +2459,30 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_ GGML_ASSERT(cuda_ctx_src->device == buf_ctx_src->device); GGML_ASSERT(cuda_ctx_dst->device == buf_ctx_dst->device); + // copy on src stream + if (cuda_ctx_src->device == cuda_ctx_dst->device) { + CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream())); + } else { +#ifdef GGML_CUDA_NO_PEER_COPY + return false; +#else + CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, cuda_ctx_dst->device, src->data, cuda_ctx_src->device, ggml_nbytes(dst), cuda_ctx_src->stream())); +#endif + } + + // record event on src stream if (!cuda_ctx_src->copy_event) { ggml_cuda_set_device(cuda_ctx_src->device); CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_src->copy_event, cudaEventDisableTiming)); } - // copy on src stream - if (cuda_ctx_src->device == cuda_ctx_dst->device) { - CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, g_cudaStreams[cuda_ctx_dst->device][0])); - } else { - CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, cuda_ctx_dst->device, src->data, cuda_ctx_src->device, ggml_nbytes(dst), g_cudaStreams[cuda_ctx_src->device][0])); - } - - // record event on src stream - CUDA_CHECK(cudaEventRecord(cuda_ctx_src->copy_event, g_cudaStreams[cuda_ctx_src->device][0])); + CUDA_CHECK(cudaEventRecord(cuda_ctx_src->copy_event, cuda_ctx_src->stream())); // wait on dst stream for the copy to complete - CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[cuda_ctx_dst->device][0], cuda_ctx_src->copy_event, 0)); + CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx_dst->stream(), cuda_ctx_src->copy_event, 0)); } else { // src and dst are on the same backend - CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, g_cudaStreams[cuda_ctx_dst->device][0])); + CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream())); } return true; } @@ -11206,15 +2490,15 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_ GGML_CALL static void ggml_backend_cuda_synchronize(ggml_backend_t backend) { ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; - CUDA_CHECK(cudaStreamSynchronize(g_cudaStreams[cuda_ctx->device][0])); + CUDA_CHECK(cudaStreamSynchronize(cuda_ctx->stream())); - UNUSED(backend); + GGML_UNUSED(backend); } GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; - ggml_cuda_set_main_device(cuda_ctx->device); + ggml_cuda_set_device(cuda_ctx->device); for (int i = 0; i < cgraph->n_nodes; i++) { ggml_tensor * node = cgraph->nodes[i]; @@ -11224,20 +2508,15 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t } #ifndef NDEBUG - assert(node->backend == GGML_BACKEND_TYPE_GPU || node->backend == GGML_BACKEND_TYPE_GPU_SPLIT); assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device)); - assert(node->extra != nullptr); - for (int j = 0; j < GGML_MAX_SRC; j++) { if (node->src[j] != nullptr) { - assert(node->src[j]->backend == GGML_BACKEND_TYPE_GPU || node->src[j]->backend == GGML_BACKEND_TYPE_GPU_SPLIT); assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) || ggml_backend_buffer_is_cuda_split(node->src[j]->buffer)); - assert(node->src[j]->extra != nullptr); } } #endif - bool ok = ggml_cuda_compute_forward(node); + bool ok = ggml_cuda_compute_forward(*cuda_ctx, node); if (!ok) { fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op)); } @@ -11322,6 +2601,15 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_1) { return true; } + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_0) { + return true; + } + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_1) { + return true; + } + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) { + return true; + } if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) { return true; } @@ -11371,7 +2659,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons return false; } - UNUSED(backend); + GGML_UNUSED(backend); } GGML_CALL static bool ggml_backend_cuda_offload_op(ggml_backend_t backend, const ggml_tensor * op) { @@ -11379,10 +2667,13 @@ GGML_CALL static bool ggml_backend_cuda_offload_op(ggml_backend_t backend, const return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS; - UNUSED(backend); + GGML_UNUSED(backend); } static ggml_backend_event_t ggml_backend_cuda_event_new(ggml_backend_t backend) { +#ifdef GGML_CUDA_NO_PEER_COPY + return nullptr; +#else ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; ggml_cuda_set_device(cuda_ctx->device); @@ -11394,6 +2685,7 @@ static ggml_backend_event_t ggml_backend_cuda_event_new(ggml_backend_t backend) /* .backend = */ backend, /* .context = */ event, }; +#endif } static void ggml_backend_cuda_event_free(ggml_backend_event_t event) { @@ -11405,14 +2697,14 @@ static void ggml_backend_cuda_event_free(ggml_backend_event_t event) { static void ggml_backend_cuda_event_record(ggml_backend_event_t event) { ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)event->backend->context; - CUDA_CHECK(cudaEventRecord((cudaEvent_t)event->context, g_cudaStreams[cuda_ctx->device][0])); + CUDA_CHECK(cudaEventRecord((cudaEvent_t)event->context, cuda_ctx->stream())); } static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_event_t event) { ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; if (ggml_backend_is_cuda(event->backend)) { - CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[cuda_ctx->device][0], (cudaEvent_t)event->context, 0)); + CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx->stream(), (cudaEvent_t)event->context, 0)); } else { #if 0 // untested @@ -11421,7 +2713,7 @@ static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_ev ggml_backend_event_synchronize(event); }; - CUDA_CHECK(cudaLaunchHostFunc(g_cudaStreams[cuda_ctx->device][0], wait_fn, event)); + CUDA_CHECK(cudaLaunchHostFunc(cuda_ctx->stream(), wait_fn, event)); #endif GGML_ASSERT(false); } @@ -11458,16 +2750,11 @@ static ggml_guid_t ggml_backend_cuda_guid() { } GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) { - ggml_init_cublas(); - - if (device < 0 || device >= ggml_cuda_get_device_count()) { + if (device < 0 || device >= ggml_backend_cuda_get_device_count()) { fprintf(stderr, "%s: error: invalid device %d\n", __func__, device); return nullptr; } - // not strictly necessary, but it may reduce the overhead of the first graph_compute - ggml_cuda_set_main_device(device); - ggml_backend_cuda_context * ctx = new ggml_backend_cuda_context(device); if (ctx == nullptr) { fprintf(stderr, "%s: error: failed to allocate context\n", __func__); @@ -11488,11 +2775,13 @@ GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend) { } GGML_CALL int ggml_backend_cuda_get_device_count() { - return ggml_cuda_get_device_count(); + return ggml_cuda_info().device_count; } GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size) { - ggml_cuda_get_device_description(device, description, description_size); + cudaDeviceProp prop; + CUDA_CHECK(cudaGetDeviceProperties(&prop, device)); + snprintf(description, description_size, "%s", prop.name); } GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total) { @@ -11502,7 +2791,7 @@ GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, si } GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size) { - if (getenv("GGML_CUDA_NO_PINNED") != nullptr) { + if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) { return false; } @@ -11519,6 +2808,10 @@ GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size } GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer) { + if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) { + return; + } + cudaError_t err = cudaHostUnregister(buffer); if (err != cudaSuccess) { // clear the error @@ -11531,13 +2824,13 @@ GGML_CALL static ggml_backend_t ggml_backend_reg_cuda_init(const char * params, ggml_backend_t cuda_backend = ggml_backend_cuda_init((int) (intptr_t) user_data); return cuda_backend; - UNUSED(params); + GGML_UNUSED(params); } extern "C" GGML_CALL int ggml_backend_cuda_reg_devices(); GGML_CALL int ggml_backend_cuda_reg_devices() { - int device_count = ggml_cuda_get_device_count(); + int device_count = ggml_backend_cuda_get_device_count(); //int device_count = 1; // DEBUG: some tools require delaying CUDA initialization for (int i = 0; i < device_count; i++) { char name[128]; diff --git a/ggml-cuda/acc.cu b/ggml-cuda/acc.cu new file mode 100644 index 000000000..96bfe1c9d --- /dev/null +++ b/ggml-cuda/acc.cu @@ -0,0 +1,47 @@ +#include "acc.cuh" + +static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne, + const int ne10, const int ne11, const int ne12, + const int nb1, const int nb2, int offset) { + const int i = blockDim.x * blockIdx.x + threadIdx.x; + if (i >= ne) { + return; + } + int src1_idx = i - offset; + int oz = src1_idx / nb2; + int oy = (src1_idx - (oz * nb2)) / nb1; + int ox = src1_idx % nb1; + if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) { + dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11]; + } else { + dst[i] = x[i]; + } +} + +static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements, + const int ne10, const int ne11, const int ne12, + const int nb1, const int nb2, const int offset, cudaStream_t stream) { + int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE; + acc_f32<<>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset); +} + +void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + const float * src0_d = (const float *)src0->data; + const float * src1_d = (const float *)src1->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported + + int nb1 = dst->op_params[0] / 4; // 4 bytes of float32 + int nb2 = dst->op_params[1] / 4; // 4 bytes of float32 + // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused + int offset = dst->op_params[3] / 4; // offset in bytes + + acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, stream); +} diff --git a/ggml-cuda/acc.cuh b/ggml-cuda/acc.cuh new file mode 100644 index 000000000..1168ea1b2 --- /dev/null +++ b/ggml-cuda/acc.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_ACC_BLOCK_SIZE 256 + +void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/alibi.cu b/ggml-cuda/alibi.cu new file mode 100644 index 000000000..6c7f1fd95 --- /dev/null +++ b/ggml-cuda/alibi.cu @@ -0,0 +1,63 @@ +#include "alibi.cuh" + +static __global__ void alibi_f32(const float * x, float * dst, const int ncols, const int k_rows, + const int n_heads_log2_floor, const float m0, const float m1) { + const int col = blockDim.x*blockIdx.x + threadIdx.x; + + if (col >= ncols) { + return; + } + + const int row = blockDim.y*blockIdx.y + threadIdx.y; + const int i = row*ncols + col; + + const int k = row/k_rows; + + float m_k; + if (k < n_heads_log2_floor) { + m_k = powf(m0, k + 1); + } else { + m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1); + } + + dst[i] = col * m_k + x[i]; +} + +static void alibi_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, + const int k_rows, const int n_heads_log2_floor, const float m0, + const float m1, cudaStream_t stream) { + const dim3 block_dims(CUDA_ALIBI_BLOCK_SIZE, 1, 1); + const int num_blocks_x = (ncols + CUDA_ALIBI_BLOCK_SIZE - 1) / (CUDA_ALIBI_BLOCK_SIZE); + const dim3 block_nums(num_blocks_x, nrows, 1); + alibi_f32<<>>(x, dst, ncols, k_rows, n_heads_log2_floor, m0, m1); +} + +void ggml_cuda_op_alibi(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t ne02 = src0->ne[2]; + const int64_t nrows = ggml_nrows(src0); + + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_head = ((int32_t *) dst->op_params)[1]; + float max_bias; + memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float)); + + //GGML_ASSERT(ne01 + n_past == ne00); + GGML_ASSERT(n_head == ne02); + + const int n_heads_log2_floor = 1 << (int) floor(log2(n_head)); + + const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor); + const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor); + + alibi_f32_cuda(src0_d, dst_d, ne00, nrows, ne01, n_heads_log2_floor, m0, m1, stream); +} diff --git a/ggml-cuda/alibi.cuh b/ggml-cuda/alibi.cuh new file mode 100644 index 000000000..630adfc7f --- /dev/null +++ b/ggml-cuda/alibi.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_ALIBI_BLOCK_SIZE 32 + +void ggml_cuda_op_alibi(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/arange.cu b/ggml-cuda/arange.cu new file mode 100644 index 000000000..b5e495a24 --- /dev/null +++ b/ggml-cuda/arange.cu @@ -0,0 +1,34 @@ +#include "arange.cuh" + +static __global__ void arange_f32(float * dst, const int ne0, const float start, const float step) { + // blockIDx.x: idx of ne0 / BLOCK_SIZE + int nidx = threadIdx.x + blockIdx.x * blockDim.x; + if (nidx >= ne0) { + return; + } + dst[nidx] = start + step * nidx; +} + +static void arange_f32_cuda(float * dst, const int ne0, const float start, const float step, cudaStream_t stream) { + int num_blocks = (ne0 + CUDA_ARANGE_BLOCK_SIZE - 1) / CUDA_ARANGE_BLOCK_SIZE; + arange_f32<<>>(dst, ne0, start, step); +} + +void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + float start; + float stop; + float step; + memcpy(&start, (float *)dst->op_params + 0, sizeof(float)); + memcpy(&stop, (float *)dst->op_params + 1, sizeof(float)); + memcpy(&step, (float *)dst->op_params + 2, sizeof(float)); + + int64_t steps = (int64_t)ceil((stop - start) / step); + GGML_ASSERT(ggml_nelements(dst) == steps); + + arange_f32_cuda(dst_d, dst->ne[0], start, step, stream); +} diff --git a/ggml-cuda/arange.cuh b/ggml-cuda/arange.cuh new file mode 100644 index 000000000..41e74fdfc --- /dev/null +++ b/ggml-cuda/arange.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_ARANGE_BLOCK_SIZE 256 + +void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/argsort.cu b/ggml-cuda/argsort.cu new file mode 100644 index 000000000..1333287e4 --- /dev/null +++ b/ggml-cuda/argsort.cu @@ -0,0 +1,77 @@ +#include "argsort.cuh" + +template +static inline __device__ void ggml_cuda_swap(T & a, T & b) { + T tmp = a; + a = b; + b = tmp; +} + +template +static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols) { + // bitonic sort + int col = threadIdx.x; + int row = blockIdx.y; + + if (col >= ncols) return; + + const float * x_row = x + row * ncols; + int * dst_row = dst + row * ncols; + + // initialize indices + if (col < ncols) { + dst_row[col] = col; + } + __syncthreads(); + + for (int k = 2; k <= ncols; k *= 2) { + for (int j = k / 2; j > 0; j /= 2) { + int ixj = col ^ j; + if (ixj > col) { + if ((col & k) == 0) { + if (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]]) { + ggml_cuda_swap(dst_row[col], dst_row[ixj]); + } + } else { + if (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]]) { + ggml_cuda_swap(dst_row[col], dst_row[ixj]); + } + } + } + __syncthreads(); + } + } +} + +static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) { + // bitonic sort requires ncols to be power of 2 + GGML_ASSERT((ncols & (ncols - 1)) == 0); + + const dim3 block_dims(ncols, 1, 1); + const dim3 block_nums(1, nrows, 1); + if (order == GGML_SORT_ORDER_ASC) { + k_argsort_f32_i32<<>>(x, dst, ncols); + } else if (order == GGML_SORT_ORDER_DESC) { + k_argsort_f32_i32<<>>(x, dst, ncols); + } else { + GGML_ASSERT(false); + } +} + +void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_is_contiguous(src0)); + + const int64_t ncols = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0]; + + argsort_f32_i32_cuda(src0_d, (int *)dst_d, ncols, nrows, order, stream); +} diff --git a/ggml-cuda/argsort.cuh b/ggml-cuda/argsort.cuh new file mode 100644 index 000000000..68a001547 --- /dev/null +++ b/ggml-cuda/argsort.cuh @@ -0,0 +1,3 @@ +#include "common.cuh" + +void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/binbcast.cu b/ggml-cuda/binbcast.cu new file mode 100644 index 000000000..959eaed95 --- /dev/null +++ b/ggml-cuda/binbcast.cu @@ -0,0 +1,236 @@ +#include "binbcast.cuh" + +static __device__ __forceinline__ float op_repeat(const float a, const float b) { + return b; + GGML_UNUSED(a); +} + +static __device__ __forceinline__ float op_add(const float a, const float b) { + return a + b; +} + +static __device__ __forceinline__ float op_mul(const float a, const float b) { + return a * b; +} + +static __device__ __forceinline__ float op_div(const float a, const float b) { + return a / b; +} + +template +static __global__ void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst, + int ne0, int ne1, int ne2, int ne3, + int ne10, int ne11, int ne12, int ne13, + /*int s0, */ int s1, int s2, int s3, + /*int s10,*/ int s11, int s12, int s13) { + const int i0s = blockDim.x*blockIdx.x + threadIdx.x; + const int i1 = (blockDim.y*blockIdx.y + threadIdx.y); + const int i2 = (blockDim.z*blockIdx.z + threadIdx.z) / ne3; + const int i3 = (blockDim.z*blockIdx.z + threadIdx.z) % ne3; + + if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) { + return; + } + + const int i11 = i1 % ne11; + const int i12 = i2 % ne12; + const int i13 = i3 % ne13; + + const size_t i_src0 = i3*s3 + i2*s2 + i1*s1; + const size_t i_src1 = i13*s13 + i12*s12 + i11*s11; + const size_t i_dst = i_src0; + + const src0_t * src0_row = src0 + i_src0; + const src1_t * src1_row = src1 + i_src1; + dst_t * dst_row = dst + i_dst; + + for (int i0 = i0s; i0 < ne0; i0 += blockDim.x*gridDim.x) { + const int i10 = i0 % ne10; + dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]); + } +} + +template +static __global__ void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst, + int ne0, int ne1, int ne2, int ne3, + int ne10, int ne11, int ne12, int ne13, + /*int s0, */ int s1, int s2, int s3, + /*int s10,*/ int s11, int s12, int s13) { + + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + const int i3 = i/(ne2*ne1*ne0); + const int i2 = (i/(ne1*ne0)) % ne2; + const int i1 = (i/ne0) % ne1; + const int i0 = i % ne0; + + if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) { + return; + } + + const int i11 = i1 % ne11; + const int i12 = i2 % ne12; + const int i13 = i3 % ne13; + + const size_t i_src0 = i3*s3 + i2*s2 + i1*s1; + const size_t i_src1 = i13*s13 + i12*s12 + i11*s11; + const size_t i_dst = i_src0; + + const src0_t * src0_row = src0 + i_src0; + const src1_t * src1_row = src1 + i_src1; + dst_t * dst_row = dst + i_dst; + + const int i10 = i0 % ne10; + dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]); +} + +template +struct bin_bcast_cuda { + template + void operator()(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, + const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd, + cudaStream_t stream) { + + GGML_TENSOR_BINARY_OP_LOCALS + + int nr0 = ne10/ne0; + int nr1 = ne11/ne1; + int nr2 = ne12/ne2; + int nr3 = ne13/ne3; + + int nr[4] = { nr0, nr1, nr2, nr3 }; + + // collapse dimensions until first broadcast dimension + int64_t cne0[] = {ne0, ne1, ne2, ne3}; + int64_t cne1[] = {ne10, ne11, ne12, ne13}; + size_t cnb0[] = {nb0, nb1, nb2, nb3}; + size_t cnb1[] = {nb10, nb11, nb12, nb13}; + auto collapse = [](int64_t cne[]) { + cne[0] *= cne[1]; + cne[1] = cne[2]; + cne[2] = cne[3]; + cne[3] = 1; + }; + + auto collapse_nb = [](size_t cnb[], const int64_t cne[]) { + cnb[1] *= cne[1]; + cnb[2] *= cne[2]; + cnb[3] *= cne[3]; + }; + + for (int i = 0; i < 4; i++) { + if (nr[i] != 1) { + break; + } + if (i > 0) { + collapse_nb(cnb0, cne0); + collapse_nb(cnb1, cne1); + collapse(cne0); + collapse(cne1); + } + } + { + int64_t ne0 = cne0[0]; + int64_t ne1 = cne0[1]; + int64_t ne2 = cne0[2]; + int64_t ne3 = cne0[3]; + + int64_t ne10 = cne1[0]; + int64_t ne11 = cne1[1]; + int64_t ne12 = cne1[2]; + int64_t ne13 = cne1[3]; + + size_t nb0 = cnb0[0]; + size_t nb1 = cnb0[1]; + size_t nb2 = cnb0[2]; + size_t nb3 = cnb0[3]; + + size_t nb10 = cnb1[0]; + size_t nb11 = cnb1[1]; + size_t nb12 = cnb1[2]; + size_t nb13 = cnb1[3]; + + size_t s0 = nb0 / sizeof(dst_t); + size_t s1 = nb1 / sizeof(dst_t); + size_t s2 = nb2 / sizeof(dst_t); + size_t s3 = nb3 / sizeof(dst_t); + + size_t s10 = nb10 / sizeof(src1_t); + size_t s11 = nb11 / sizeof(src1_t); + size_t s12 = nb12 / sizeof(src1_t); + size_t s13 = nb13 / sizeof(src1_t); + + GGML_ASSERT(s0 == 1); + GGML_ASSERT(s10 == 1); + + const int block_size = 128; + + int64_t hne0 = std::max(ne0/2LL, 1LL); + + dim3 block_dims; + block_dims.x = std::min(hne0, block_size); + block_dims.y = std::min(ne1, block_size / block_dims.x); + block_dims.z = std::min(std::min(ne2*ne3, block_size / block_dims.x / block_dims.y), 64U); + + dim3 block_nums( + (hne0 + block_dims.x - 1) / block_dims.x, + (ne1 + block_dims.y - 1) / block_dims.y, + (ne2*ne3 + block_dims.z - 1) / block_dims.z + ); + + if (block_nums.z > 65535) { + // this is the maximum number of blocks in z direction, fallback to 1D grid kernel + int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size; + k_bin_bcast_unravel<<>>( + src0_dd, src1_dd, dst_dd, + ne0, ne1, ne2, ne3, + ne10, ne11, ne12, ne13, + /* s0, */ s1, s2, s3, + /* s10, */ s11, s12, s13); + } else { + k_bin_bcast<<>>( + src0_dd, src1_dd, dst_dd, + ne0, ne1, ne2, ne3, + ne10, ne11, ne12, ne13, + /* s0, */ s1, s2, s3, + /* s10, */ s11, s12, s13); + } + } + } +}; + +template +static void ggml_cuda_op_bin_bcast( + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const void * src0_dd, const void * src1_dd, void * dst_dd, cudaStream_t stream) { + + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + op()(src0, src1, dst, (const float *)src0_dd, (const float *)src1_dd, (float *)dst_dd, stream); + } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { + op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (half *) dst_dd, stream); + } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) { + op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (float *)dst_dd, stream); + } else { + fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__, + ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type)); + GGML_ASSERT(false); + } +} + +void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_bin_bcast>(dst, dst->src[0], dst, nullptr, dst->src[0]->data, dst->data, ctx.stream()); +} + +void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_bin_bcast>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream()); +} + +void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_bin_bcast>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream()); +} + +void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_bin_bcast>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream()); +} diff --git a/ggml-cuda/binbcast.cuh b/ggml-cuda/binbcast.cuh new file mode 100644 index 000000000..4f63d6372 --- /dev/null +++ b/ggml-cuda/binbcast.cuh @@ -0,0 +1,6 @@ +#include "common.cuh" + +void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst); +void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst); +void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst); +void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/clamp.cu b/ggml-cuda/clamp.cu new file mode 100644 index 000000000..379ded042 --- /dev/null +++ b/ggml-cuda/clamp.cu @@ -0,0 +1,35 @@ +#include "clamp.cuh" + +static __global__ void clamp_f32(const float * x, float * dst, const float min, const float max, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + + dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]); +} + +static void clamp_f32_cuda(const float * x, float * dst, const float min, const float max, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_CLAMP_BLOCK_SIZE - 1) / CUDA_CLAMP_BLOCK_SIZE; + clamp_f32<<>>(x, dst, min, max, k); +} + + +void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + float min; + float max; + memcpy(&min, dst->op_params, sizeof(float)); + memcpy(&max, (float *) dst->op_params + 1, sizeof(float)); + + clamp_f32_cuda(src0_d, dst_d, min, max, ggml_nelements(src0), stream); + CUDA_CHECK(cudaGetLastError()); +} diff --git a/ggml-cuda/clamp.cuh b/ggml-cuda/clamp.cuh new file mode 100644 index 000000000..7f9559dd1 --- /dev/null +++ b/ggml-cuda/clamp.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_CLAMP_BLOCK_SIZE 256 + +void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/common.cuh b/ggml-cuda/common.cuh new file mode 100644 index 000000000..33c8ed1da --- /dev/null +++ b/ggml-cuda/common.cuh @@ -0,0 +1,550 @@ +#pragma once + +#include "../ggml.h" +#include "../ggml-cuda.h" +#include + +#if defined(GGML_USE_HIPBLAS) +#define GGML_COMMON_DECL_HIP +#define GGML_COMMON_IMPL_HIP +#else +#define GGML_COMMON_DECL_CUDA +#define GGML_COMMON_IMPL_CUDA +#endif +#include "../ggml-common.h" + +#include +#include +#include +#include +#include + +#if defined(GGML_USE_HIPBLAS) +#include +#include +#include +#ifdef __HIP_PLATFORM_AMD__ +// for rocblas_initialize() +#include "rocblas/rocblas.h" +#endif // __HIP_PLATFORM_AMD__ +#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F +#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F +#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F +#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT +#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT +#define CUBLAS_OP_N HIPBLAS_OP_N +#define CUBLAS_OP_T HIPBLAS_OP_T +#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS +#define CUBLAS_TF32_TENSOR_OP_MATH 0 +#define CUDA_R_16F HIPBLAS_R_16F +#define CUDA_R_32F HIPBLAS_R_32F +#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width) +#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6 +#define cublasCreate hipblasCreate +#define cublasDestroy hipblasDestroy +#define cublasGemmEx hipblasGemmEx +#define cublasGemmBatchedEx hipblasGemmBatchedEx +#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx +#define cublasHandle_t hipblasHandle_t +#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS +#define cublasSetStream hipblasSetStream +#define cublasSgemm hipblasSgemm +#define cublasStatus_t hipblasStatus_t +#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6 +#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer +#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess +#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess +#define cudaDeviceProp hipDeviceProp_t +#define cudaDeviceSynchronize hipDeviceSynchronize +#define cudaError_t hipError_t +#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled +#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled +#define cudaEventCreateWithFlags hipEventCreateWithFlags +#define cudaEventDisableTiming hipEventDisableTiming +#define cudaEventRecord hipEventRecord +#define cudaEventSynchronize hipEventSynchronize +#define cudaEvent_t hipEvent_t +#define cudaEventDestroy hipEventDestroy +#define cudaFree hipFree +#define cudaFreeHost hipHostFree +#define cudaGetDevice hipGetDevice +#define cudaGetDeviceCount hipGetDeviceCount +#define cudaGetDeviceProperties hipGetDeviceProperties +#define cudaGetErrorString hipGetErrorString +#define cudaGetLastError hipGetLastError +#define cudaHostRegister hipHostRegister +#define cudaHostRegisterPortable hipHostRegisterPortable +#define cudaHostRegisterReadOnly hipHostRegisterReadOnly +#define cudaHostUnregister hipHostUnregister +#define cudaLaunchHostFunc hipLaunchHostFunc +#ifdef GGML_HIP_UMA +#define cudaMalloc hipMallocManaged +#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size) +#else +#define cudaMalloc hipMalloc +#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault) +#endif +#define cudaMemcpy hipMemcpy +#define cudaMemcpyAsync hipMemcpyAsync +#define cudaMemcpyPeerAsync hipMemcpyPeerAsync +#define cudaMemcpy2DAsync hipMemcpy2DAsync +#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice +#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost +#define cudaMemcpyHostToDevice hipMemcpyHostToDevice +#define cudaMemcpyKind hipMemcpyKind +#define cudaMemset hipMemset +#define cudaMemsetAsync hipMemsetAsync +#define cudaMemGetInfo hipMemGetInfo +#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize +#define cudaSetDevice hipSetDevice +#define cudaStreamCreateWithFlags hipStreamCreateWithFlags +#define cudaStreamDestroy hipStreamDestroy +#define cudaStreamFireAndForget hipStreamFireAndForget +#define cudaStreamNonBlocking hipStreamNonBlocking +#define cudaStreamPerThread hipStreamPerThread +#define cudaStreamSynchronize hipStreamSynchronize +#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags) +#define cudaStream_t hipStream_t +#define cudaSuccess hipSuccess +#define __trap abort +#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS +#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED +#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED +#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE +#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH +#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR +#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED +#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR +#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED +#else +#include +#include +#include +#include + +#if CUDART_VERSION < 11020 +#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED +#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH +#define CUBLAS_COMPUTE_16F CUDA_R_16F +#define CUBLAS_COMPUTE_32F CUDA_R_32F +#define cublasComputeType_t cudaDataType_t +#endif // CUDART_VERSION < 11020 + +#endif // defined(GGML_USE_HIPBLAS) + +#define STRINGIZE_IMPL(...) #__VA_ARGS__ +#define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__) + +#define WARP_SIZE 32 +#define CUDART_HMAX 11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed) + +#define CC_PASCAL 600 +#define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products +#define CC_VOLTA 700 +#define CC_OFFSET_AMD 1000000 +#define CC_RDNA1 (CC_OFFSET_AMD + 1010) +#define CC_RDNA2 (CC_OFFSET_AMD + 1030) +#define CC_RDNA3 (CC_OFFSET_AMD + 1100) + +// define this if you want to always fallback to MMQ kernels and not use cuBLAS for matrix multiplication +// on modern hardware, using cuBLAS is recommended as it utilizes F16 tensor cores which are very performant +// for large computational tasks. the drawback is that this requires some extra amount of VRAM: +// - 7B quantum model: +100-200 MB +// - 13B quantum model: +200-400 MB +// +//#define GGML_CUDA_FORCE_MMQ + +// TODO: improve this to be correct for more hardware +// for example, currently fails for GeForce GTX 1660 which is TURING arch (> VOLTA) but does not have tensor cores +#if !defined(GGML_CUDA_FORCE_MMQ) +#define CUDA_USE_TENSOR_CORES +#endif + +#define MMVQ_MAX_BATCH_SIZE 8 // max batch size to use MMVQ kernels +#define MMQ_MAX_BATCH_SIZE 32 // max batch size to use MMQ kernels when tensor cores are available + +#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses + +#if defined(_MSC_VER) +#pragma warning(disable: 4244 4267) // possible loss of data +#endif + +#define GGML_CUDA_MAX_STREAMS 8 + +[[noreturn]] +void ggml_cuda_error(const char * stmt, const char * func, const char * file, int line, const char * msg); + +#define CUDA_CHECK_GEN(err, success, error_fn) \ + do { \ + auto err_ = (err); \ + if (err_ != (success)) { \ + ggml_cuda_error(#err, __func__, __FILE__, __LINE__, error_fn(err_)); \ + } \ + } while (0) + +#define CUDA_CHECK(err) CUDA_CHECK_GEN(err, cudaSuccess, cudaGetErrorString) + +#if CUDART_VERSION >= 12000 + static const char * cublas_get_error_str(const cublasStatus_t err) { + return cublasGetStatusString(err); + } +#else + static const char * cublas_get_error_str(const cublasStatus_t err) { + switch (err) { + case CUBLAS_STATUS_SUCCESS: return "CUBLAS_STATUS_SUCCESS"; + case CUBLAS_STATUS_NOT_INITIALIZED: return "CUBLAS_STATUS_NOT_INITIALIZED"; + case CUBLAS_STATUS_ALLOC_FAILED: return "CUBLAS_STATUS_ALLOC_FAILED"; + case CUBLAS_STATUS_INVALID_VALUE: return "CUBLAS_STATUS_INVALID_VALUE"; + case CUBLAS_STATUS_ARCH_MISMATCH: return "CUBLAS_STATUS_ARCH_MISMATCH"; + case CUBLAS_STATUS_MAPPING_ERROR: return "CUBLAS_STATUS_MAPPING_ERROR"; + case CUBLAS_STATUS_EXECUTION_FAILED: return "CUBLAS_STATUS_EXECUTION_FAILED"; + case CUBLAS_STATUS_INTERNAL_ERROR: return "CUBLAS_STATUS_INTERNAL_ERROR"; + case CUBLAS_STATUS_NOT_SUPPORTED: return "CUBLAS_STATUS_NOT_SUPPORTED"; + default: return "unknown error"; + } + } +#endif // CUDART_VERSION >= 12000 + +#define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str) + +#if !defined(GGML_USE_HIPBLAS) +static const char * cu_get_error_str(CUresult err) { + const char * err_str; + cuGetErrorString(err, &err_str); + return err_str; +} +#define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str) +#endif + +#if CUDART_VERSION >= 11100 +#define GGML_CUDA_ASSUME(x) __builtin_assume(x) +#else +#define GGML_CUDA_ASSUME(x) +#endif // CUDART_VERSION >= 11100 + +#ifdef GGML_CUDA_F16 +typedef half dfloat; // dequantize float +typedef half2 dfloat2; +#else +typedef float dfloat; // dequantize float +typedef float2 dfloat2; +#endif //GGML_CUDA_F16 + +[[noreturn]] +static __device__ void no_device_code( + const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) { + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n", + file_name, line, function_name, arch); + GGML_UNUSED(arch_list); +#else + printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n", + file_name, line, function_name, arch, arch_list); +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + __trap(); + + GGML_UNUSED(no_device_code); // suppress unused function warning +} + +#ifdef __CUDA_ARCH__ +#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__)) +#else +#define NO_DEVICE_CODE //GGML_ASSERT(false && "NO_DEVICE_CODE not valid in host code.") +#endif // __CUDA_ARCH__ + +static __device__ __forceinline__ float warp_reduce_sum(float x) { +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + x += __shfl_xor_sync(0xffffffff, x, mask, 32); + } + return x; +} + +static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) { +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32); + a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32); + } + return a; +} + +#ifdef GGML_CUDA_F16 +static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) { +#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32)); + } + return a; +#else + GGML_UNUSED(a); + NO_DEVICE_CODE; +#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL +} +#endif // GGML_CUDA_F16 + +static __device__ __forceinline__ float warp_reduce_max(float x) { +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32)); + } + return x; +} + +//static __device__ __forceinline__ half2 warp_reduce_max(half2 x) { +//#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX +//#pragma unroll +// for (int mask = 16; mask > 0; mask >>= 1) { +// x = __hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32)); +// } +// return x; +//#else +// GGML_UNUSED(x); +// NO_DEVICE_CODE; +//#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX +//} + + +#if defined(GGML_USE_HIPBLAS) +#define __CUDA_ARCH__ 1300 + +#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \ + defined(__gfx1150__) || defined(__gfx1151__) +#define RDNA3 +#endif + +#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \ + defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__) +#define RDNA2 +#endif + +#ifndef __has_builtin + #define __has_builtin(x) 0 +#endif + +typedef int8_t int8x4_t __attribute__((ext_vector_type(4))); +typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4))); +static __device__ __forceinline__ int __vsubss4(const int a, const int b) { + const int8x4_t va = reinterpret_cast(a); + const int8x4_t vb = reinterpret_cast(b); +#if __has_builtin(__builtin_elementwise_sub_sat) + const int8x4_t c = __builtin_elementwise_sub_sat(va, vb); + return reinterpret_cast(c); +#else + int8x4_t c; + int16_t tmp; +#pragma unroll + for (int i = 0; i < 4; i++) { + tmp = va[i] - vb[i]; + if(tmp > std::numeric_limits::max()) tmp = std::numeric_limits::max(); + if(tmp < std::numeric_limits::min()) tmp = std::numeric_limits::min(); + c[i] = tmp; + } + return reinterpret_cast(c); +#endif // __has_builtin(__builtin_elementwise_sub_sat) +} + +static __device__ __forceinline__ int __vsub4(const int a, const int b) { + return __vsubss4(a, b); +} + +static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigned int b) { + const uint8x4_t& va = reinterpret_cast(a); + const uint8x4_t& vb = reinterpret_cast(b); + unsigned int c; + uint8x4_t& vc = reinterpret_cast(c); +#pragma unroll + for (int i = 0; i < 4; ++i) { + vc[i] = va[i] == vb[i] ? 0xff : 0x00; + } + return c; +} + +static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) { +#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) + c = __builtin_amdgcn_sdot4(a, b, c, false); +#elif defined(RDNA3) + c = __builtin_amdgcn_sudot4( true, a, true, b, c, false); +#elif defined(__gfx1010__) || defined(__gfx900__) + int tmp1; + int tmp2; + asm("\n \ + v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \ + v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \ + v_add3_u32 %0, %1, %2, %0 \n \ + v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \ + v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \ + v_add3_u32 %0, %1, %2, %0 \n \ + " + : "+v"(c), "=&v"(tmp1), "=&v"(tmp2) + : "v"(a), "v"(b) + ); +#else + const int8x4_t va = reinterpret_cast(a); + const int8x4_t vb = reinterpret_cast(b); + c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3]; +#endif + return c; +} +#endif // defined(GGML_USE_HIPBLAS) + +// TODO: move to ggml-common.h +static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113}; + +typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v); + + +////////////////////// + +struct ggml_cuda_device_info { + int device_count; + + struct cuda_device_info { + int cc; // compute capability + size_t smpb; // max. shared memory per block + bool vmm; // virtual memory support + size_t vmm_granularity; // granularity of virtual memory + size_t total_vram; + }; + + cuda_device_info devices[GGML_CUDA_MAX_DEVICES] = {}; + + std::array default_tensor_split = {}; +}; + +const ggml_cuda_device_info & ggml_cuda_info(); + +void ggml_cuda_set_device(int device); +int ggml_cuda_get_device(); + +struct ggml_cuda_pool { + virtual ~ggml_cuda_pool() = default; + + virtual void * alloc(size_t size, size_t * actual_size) = 0; + virtual void free(void * ptr, size_t size) = 0; +}; + +template +struct ggml_cuda_pool_alloc { + ggml_cuda_pool * pool = nullptr; + T * ptr = nullptr; + size_t actual_size = 0; + + ggml_cuda_pool_alloc() = default; + + explicit ggml_cuda_pool_alloc(ggml_cuda_pool & pool) : pool(&pool) { + } + + ggml_cuda_pool_alloc(ggml_cuda_pool & pool, size_t size) : pool(&pool) { + alloc(size); + } + + ~ggml_cuda_pool_alloc() { + if (ptr != nullptr) { + pool->free(ptr, actual_size); + } + } + + // size is in number of elements + T * alloc(size_t size) { + GGML_ASSERT(pool != nullptr); + GGML_ASSERT(ptr == nullptr); + ptr = (T *) pool->alloc(size * sizeof(T), &this->actual_size); + return ptr; + } + + T * alloc(ggml_cuda_pool & pool, size_t size) { + this->pool = &pool; + return alloc(size); + } + + T * get() { + return ptr; + } + + ggml_cuda_pool_alloc(const ggml_cuda_pool_alloc &) = delete; + ggml_cuda_pool_alloc(ggml_cuda_pool_alloc &&) = delete; + ggml_cuda_pool_alloc& operator=(const ggml_cuda_pool_alloc &) = delete; + ggml_cuda_pool_alloc& operator=(ggml_cuda_pool_alloc &&) = delete; +}; + + +// backend interface + +struct ggml_tensor_extra_gpu { + void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors + cudaEvent_t events[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS]; // events for synchronizing multiple GPUs +}; + +struct ggml_backend_cuda_context { + int device; + std::string name; + cudaEvent_t copy_event = nullptr; + + cudaStream_t streams[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = { { nullptr } }; + cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr}; + + explicit ggml_backend_cuda_context(int device) : + device(device), + name(GGML_CUDA_NAME + std::to_string(device)) { + } + + ~ggml_backend_cuda_context() { + if (copy_event != nullptr) { + CUDA_CHECK(cudaEventDestroy(copy_event)); + } + for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) { + for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) { + if (streams[i][j] != nullptr) { + CUDA_CHECK(cudaStreamDestroy(streams[i][j])); + } + } + if (cublas_handles[i] != nullptr) { + CUBLAS_CHECK(cublasDestroy(cublas_handles[i])); + } + } + } + + cudaStream_t stream(int device, int stream) { + if (streams[device][stream] == nullptr) { + ggml_cuda_set_device(device); + CUDA_CHECK(cudaStreamCreateWithFlags(&streams[device][stream], cudaStreamNonBlocking)); + } + return streams[device][stream]; + } + + cudaStream_t stream() { + return stream(device, 0); + } + + cublasHandle_t cublas_handle(int device) { + if (cublas_handles[device] == nullptr) { + ggml_cuda_set_device(device); + CUBLAS_CHECK(cublasCreate(&cublas_handles[device])); + CUBLAS_CHECK(cublasSetMathMode(cublas_handles[device], CUBLAS_TF32_TENSOR_OP_MATH)); + } + return cublas_handles[device]; + } + + cublasHandle_t cublas_handle() { + return cublas_handle(device); + } + + // pool + std::unique_ptr pools[GGML_CUDA_MAX_DEVICES]; + + static std::unique_ptr new_pool_for_device(int device); + + ggml_cuda_pool & pool(int device) { + if (pools[device] == nullptr) { + pools[device] = new_pool_for_device(device); + } + return *pools[device]; + } + + ggml_cuda_pool & pool() { + return pool(device); + } +}; diff --git a/ggml-cuda/concat.cu b/ggml-cuda/concat.cu new file mode 100644 index 000000000..2941d2f17 --- /dev/null +++ b/ggml-cuda/concat.cu @@ -0,0 +1,49 @@ +#include "concat.cuh" + +static __global__ void concat_f32(const float * x,const float * y, float * dst, const int ne0, const int ne02) { + int nidx = threadIdx.x + blockIdx.x * blockDim.x; + if (nidx >= ne0) { + return; + } + // operation + int offset_dst = + nidx + + blockIdx.y * ne0 + + blockIdx.z * ne0 * gridDim.y; + if (blockIdx.z < ne02) { // src0 + int offset_src = + nidx + + blockIdx.y * ne0 + + blockIdx.z * ne0 * gridDim.y; + dst[offset_dst] = x[offset_src]; + } else { + int offset_src = + nidx + + blockIdx.y * ne0 + + (blockIdx.z - ne02) * ne0 * gridDim.y; + dst[offset_dst] = y[offset_src]; + } +} + +static void concat_f32_cuda(const float * x, const float * y, float * dst, const int ne0, int ne1, int ne2, int ne02, cudaStream_t stream) { + int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE; + dim3 gridDim(num_blocks, ne1, ne2); + concat_f32<<>>(x, y, dst, ne0, ne02); +} + +void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + const float * src0_d = (const float *)src0->data; + const float * src1_d = (const float *)src1->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + for (int i3 = 0; i3 < dst->ne[3]; i3++) { + concat_f32_cuda(src0_d + i3 * (src0->nb[3] / 4), src1_d + i3 * (src1->nb[3] / 4), dst_d + i3 * (dst->nb[3] / 4), dst->ne[0], dst->ne[1], dst->ne[2], src0->ne[2], stream); + } +} diff --git a/ggml-cuda/concat.cuh b/ggml-cuda/concat.cuh new file mode 100644 index 000000000..aa506a05f --- /dev/null +++ b/ggml-cuda/concat.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_CONCAT_BLOCK_SIZE 256 + +void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/convert.cu b/ggml-cuda/convert.cu new file mode 100644 index 000000000..2516ecddd --- /dev/null +++ b/ggml-cuda/convert.cu @@ -0,0 +1,783 @@ +#include "convert.cuh" +#include "dequantize.cuh" + +#define CUDA_Q8_0_NE_ALIGN 2048 + +template +static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k) { + const int i = 2*(blockDim.x*blockIdx.x + threadIdx.x); + + if (i >= k) { + return; + } + + const int ib = i/qk; // block index + const int iqs = (i%qk)/qr; // quant index + const int iybs = i - i%qk; // y block start index + const int y_offset = qr == 1 ? 1 : qk/2; + + // dequantize + dfloat2 v; + dequantize_kernel(vx, ib, iqs, v); + + y[iybs + iqs + 0] = v.x; + y[iybs + iqs + y_offset] = v.y; +} + +template +static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, half * __restrict__ y, const int k) { +#if __CUDA_ARCH__ >= CC_PASCAL + constexpr int nint = CUDA_Q8_0_NE_ALIGN/sizeof(int) + WARP_SIZE; + + const int i0 = CUDA_Q8_0_NE_ALIGN*blockIdx.x; + const int * x0 = ((int *) vx) + blockIdx.x * nint; + half2 * y2 = (half2 *) (y + i0); + + __shared__ int vals[nint]; + +#pragma unroll + for (int ix0 = 0; ix0 < nint; ix0 += WARP_SIZE) { + if (need_check && i0*sizeof(block_q8_0)/QK8_0 + sizeof(int)*(ix0 + threadIdx.x) >= k*sizeof(block_q8_0)/QK8_0) { + break; + } + + const int ix = ix0 + threadIdx.x; + vals[ix] = x0[ix]; + } + +#pragma unroll + for (int iy = 0; iy < CUDA_Q8_0_NE_ALIGN; iy += 2*WARP_SIZE) { + if (need_check && i0 + iy + 2*threadIdx.x >= k) { + return; + } + + const half * b0 = ((const half *) vals) + (sizeof(block_q8_0)/sizeof(half)) * ((iy + 2*threadIdx.x)/QK8_0); + const half d = *b0; + const char2 qs = ((const char2 *) (b0 + 1))[threadIdx.x % (QK8_0/2)]; + + y2[iy/2 + threadIdx.x] = __hmul2(make_half2(qs.x, qs.y), __half2half2(d)); + } +#else + GGML_UNUSED(vx); + GGML_UNUSED(y); + GGML_UNUSED(k); + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= CC_PASCAL +} + +template +static __global__ void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) { + + const int i = blockIdx.x; + + // assume 32 threads + const int tid = threadIdx.x; + const int il = tid/8; + const int ir = tid%8; + const int ib = 8*i + ir; + if (ib >= nb32) { + return; + } + + dst_t * y = yy + 256*i + 32*ir + 4*il; + + const block_q4_0 * x = (const block_q4_0 *)vx + ib; + const float d = __half2float(x->d); + const float dm = -8*d; + + const uint8_t * q = x->qs + 4*il; + + for (int l = 0; l < 4; ++l) { + y[l+ 0] = d * (q[l] & 0xF) + dm; + y[l+16] = d * (q[l] >> 4) + dm; + } +} + +template +static __global__ void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) { + + const int i = blockIdx.x; + + // assume 32 threads + const int tid = threadIdx.x; + const int il = tid/8; + const int ir = tid%8; + const int ib = 8*i + ir; + if (ib >= nb32) { + return; + } + + dst_t * y = yy + 256*i + 32*ir + 4*il; + + const block_q4_1 * x = (const block_q4_1 *)vx + ib; + const float2 d = __half22float2(x->dm); + + const uint8_t * q = x->qs + 4*il; + + for (int l = 0; l < 4; ++l) { + y[l+ 0] = d.x * (q[l] & 0xF) + d.y; + y[l+16] = d.x * (q[l] >> 4) + d.y; + } +} + +//================================== k-quants + +template +static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { + + const int i = blockIdx.x; + const block_q2_K * x = (const block_q2_K *) vx; + + const int tid = threadIdx.x; +#if QK_K == 256 + const int n = tid/32; + const int l = tid - 32*n; + const int is = 8*n + l/16; + + const uint8_t q = x[i].qs[32*n + l]; + dst_t * y = yy + i*QK_K + 128*n; + + float dall = __low2half(x[i].dm); + float dmin = __high2half(x[i].dm); + y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4); + y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4); + y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4); + y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4); +#else + const int is = tid/16; // 0 or 1 + const int il = tid%16; // 0...15 + const uint8_t q = x[i].qs[il] >> (2*is); + dst_t * y = yy + i*QK_K + 16*is + il; + float dall = __low2half(x[i].dm); + float dmin = __high2half(x[i].dm); + y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4); + y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4); +#endif + +} + +template +static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { + + const int i = blockIdx.x; + const block_q3_K * x = (const block_q3_K *) vx; + +#if QK_K == 256 + const int r = threadIdx.x/4; + const int tid = r/2; + const int is0 = r%2; + const int l0 = 16*is0 + 4*(threadIdx.x%4); + const int n = tid / 4; + const int j = tid - 4*n; + + uint8_t m = 1 << (4*n + j); + int is = 8*n + 2*j + is0; + int shift = 2*j; + + int8_t us = is < 4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) : + is < 8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) : + is < 12 ? (x[i].scales[is-8] >> 4) | (((x[i].scales[is+0] >> 4) & 3) << 4) : + (x[i].scales[is-8] >> 4) | (((x[i].scales[is-4] >> 6) & 3) << 4); + float d_all = x[i].d; + float dl = d_all * (us - 32); + + dst_t * y = yy + i*QK_K + 128*n + 32*j; + const uint8_t * q = x[i].qs + 32*n; + const uint8_t * hm = x[i].hmask; + + for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)); +#else + const int tid = threadIdx.x; + const int is = tid/16; // 0 or 1 + const int il = tid%16; // 0...15 + const int im = il/8; // 0...1 + const int in = il%8; // 0...7 + + dst_t * y = yy + i*QK_K + 16*is + il; + + const uint8_t q = x[i].qs[il] >> (2*is); + const uint8_t h = x[i].hmask[in] >> (2*is + im); + const float d = (float)x[i].d; + + if (is == 0) { + y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4)); + y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4)); + } else { + y[ 0] = d * ((x[i].scales[0] >> 4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4)); + y[32] = d * ((x[i].scales[1] >> 4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4)); + } +#endif + +} + +#if QK_K == 256 +static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) { + if (j < 4) { + d = q[j] & 63; m = q[j + 4] & 63; + } else { + d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4); + m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4); + } +} +#endif + +template +static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { + const block_q4_K * x = (const block_q4_K *) vx; + + const int i = blockIdx.x; + +#if QK_K == 256 + // assume 32 threads + const int tid = threadIdx.x; + const int il = tid/8; + const int ir = tid%8; + const int is = 2*il; + const int n = 4; + + dst_t * y = yy + i*QK_K + 64*il + n*ir; + + const float dall = __low2half(x[i].dm); + const float dmin = __high2half(x[i].dm); + + const uint8_t * q = x[i].qs + 32*il + n*ir; + + uint8_t sc, m; + get_scale_min_k4(is + 0, x[i].scales, sc, m); + const float d1 = dall * sc; const float m1 = dmin * m; + get_scale_min_k4(is + 1, x[i].scales, sc, m); + const float d2 = dall * sc; const float m2 = dmin * m; + for (int l = 0; l < n; ++l) { + y[l + 0] = d1 * (q[l] & 0xF) - m1; + y[l +32] = d2 * (q[l] >> 4) - m2; + } +#else + const int tid = threadIdx.x; + const uint8_t * q = x[i].qs; + dst_t * y = yy + i*QK_K; + const float d = (float)x[i].dm[0]; + const float m = (float)x[i].dm[1]; + y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4); + y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >> 4) - m * (x[i].scales[1] >> 4); +#endif +} + +template +static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { + const block_q5_K * x = (const block_q5_K *) vx; + + const int i = blockIdx.x; + +#if QK_K == 256 + // assume 64 threads - this is very slightly better than the one below + const int tid = threadIdx.x; + const int il = tid/16; // il is in 0...3 + const int ir = tid%16; // ir is in 0...15 + const int is = 2*il; // is is in 0...6 + + dst_t * y = yy + i*QK_K + 64*il + 2*ir; + + const float dall = __low2half(x[i].dm); + const float dmin = __high2half(x[i].dm); + + const uint8_t * ql = x[i].qs + 32*il + 2*ir; + const uint8_t * qh = x[i].qh + 2*ir; + + uint8_t sc, m; + get_scale_min_k4(is + 0, x[i].scales, sc, m); + const float d1 = dall * sc; const float m1 = dmin * m; + get_scale_min_k4(is + 1, x[i].scales, sc, m); + const float d2 = dall * sc; const float m2 = dmin * m; + + uint8_t hm = 1 << (2*il); + y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1; + y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1; + hm <<= 1; + y[32] = d2 * ((ql[ 0] >> 4) + (qh[ 0] & hm ? 16 : 0)) - m2; + y[33] = d2 * ((ql[ 1] >> 4) + (qh[ 1] & hm ? 16 : 0)) - m2; +#else + const int tid = threadIdx.x; + const uint8_t q = x[i].qs[tid]; + const int im = tid/8; // 0...3 + const int in = tid%8; // 0...7 + const int is = tid/16; // 0 or 1 + const uint8_t h = x[i].qh[in] >> im; + const float d = x[i].d; + dst_t * y = yy + i*QK_K + tid; + y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16)); + y[32] = d * x[i].scales[is+2] * ((q >> 4) - ((h >> 4) & 1 ? 0 : 16)); +#endif +} + +template +static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { + const block_q6_K * x = (const block_q6_K *) vx; + + const int i = blockIdx.x; +#if QK_K == 256 + + // assume 64 threads - this is very slightly better than the one below + const int tid = threadIdx.x; + const int ip = tid/32; // ip is 0 or 1 + const int il = tid - 32*ip; // 0...32 + const int is = 8*ip + il/16; + + dst_t * y = yy + i*QK_K + 128*ip + il; + + const float d = x[i].d; + + const uint8_t * ql = x[i].ql + 64*ip + il; + const uint8_t qh = x[i].qh[32*ip + il]; + const int8_t * sc = x[i].scales + is; + + y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32); + y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32); + y[64] = d * sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32); + y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32); +#else + + // assume 32 threads + const int tid = threadIdx.x; + const int ip = tid/16; // 0 or 1 + const int il = tid - 16*ip; // 0...15 + + dst_t * y = yy + i*QK_K + 16*ip + il; + + const float d = x[i].d; + + const uint8_t ql = x[i].ql[16*ip + il]; + const uint8_t qh = x[i].qh[il] >> (2*ip); + const int8_t * sc = x[i].scales; + + y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32); + y[32] = d * sc[ip+2] * ((int8_t)((ql >> 4) | (((qh >> 4) & 3) << 4)) - 32); +#endif +} + +template +static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) { + + const int i = blockIdx.x; + const block_iq2_xxs * x = (const block_iq2_xxs *) vx; + + const int tid = threadIdx.x; +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint16_t * q2 = x[i].qs + 4*ib; + const uint8_t * aux8 = (const uint8_t *)q2; + const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[il]); + const uint32_t aux32 = q2[2] | (q2[3] << 16); + const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f; + const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127]; + for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); +#else + assert(false); +#endif + +} + +template +static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) { + + const int i = blockIdx.x; + const block_iq2_xs * x = (const block_iq2_xs *) vx; + + const int tid = threadIdx.x; +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint16_t * q2 = x[i].qs + 4*ib; + const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511)); + const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; + const uint8_t signs = ksigns_iq2xs[q2[il] >> 9]; + for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); +#else + assert(false); +#endif + +} + +template +static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_t * __restrict__ yy) { + + const int i = blockIdx.x; + const block_iq2_s * x = (const block_iq2_s *) vx; + + const int tid = threadIdx.x; +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300))); + const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; + const uint8_t signs = x[i].qs[QK_K/8+4*ib+il]; + for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); +#else + assert(false); +#endif + +} + +template +static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) { + + const int i = blockIdx.x; + const block_iq3_xxs * x = (const block_iq3_xxs *) vx; + + const int tid = threadIdx.x; +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint8_t * q3 = x[i].qs + 8*ib; + const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib; + const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]); + const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]); + const uint32_t aux32 = gas[0] | (gas[1] << 16); + const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f; + const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127]; + for (int j = 0; j < 4; ++j) { + y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); + y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); + } +#else + assert(false); +#endif + +} + +template +static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) { + + const int i = blockIdx.x; + const block_iq3_s * x = (const block_iq3_s *) vx; + + const int tid = threadIdx.x; +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint8_t * qs = x[i].qs + 8*ib; + const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256))); + const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256))); + const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf)); + const uint8_t signs = x[i].signs[4*ib + il]; + for (int j = 0; j < 4; ++j) { + y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); + y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); + } +#else + assert(false); +#endif + +} + +template +static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) { + + const int i = blockIdx.x; + const block_iq1_s * x = (const block_iq1_s *) vx; + + const int tid = threadIdx.x; +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA; + const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 7) + 1); + uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32; + grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)]; + grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f; + grid32[0] &= 0x0f0f0f0f; + for (int j = 0; j < 8; ++j) { + y[j] = d * (q[j] + delta); + } +#else + assert(false); +#endif + +} + +template +static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) { + + const int i = blockIdx.x; + const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL); + + const int tid = threadIdx.x; + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 4*il; + const uint8_t * q4 = x[ib].qs + 4*il; + const float d = (float)x[ib].d; + for (int j = 0; j < 4; ++j) { + y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf]; + y[j+16] = d * kvalues_iq4nl[q4[j] >> 4]; + } + +} + +#if QK_K != 64 +template +static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) { + const int i = blockIdx.x; + const block_iq4_xs * x = (const block_iq4_xs *)vx; + + const int tid = threadIdx.x; + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 4*il; + const uint8_t * q4 = x[i].qs + 16*ib + 4*il; + const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32); + for (int j = 0; j < 4; ++j) { + y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf]; + y[j+16] = d * kvalues_iq4nl[q4[j] >> 4]; + } +} +#endif + +template +static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) { + const int num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE); + dequantize_block<<>>(vx, y, k); +} + +static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * __restrict__ y, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_Q8_0_NE_ALIGN - 1) / CUDA_Q8_0_NE_ALIGN; + if (k % CUDA_Q8_0_NE_ALIGN == 0) { + const bool need_check = false; + dequantize_block_q8_0_f16<<>>(vx, y, k); + } else { + const bool need_check = true; + dequantize_block_q8_0_f16<<>>(vx, y, k); + } +} + +template +static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb = k / QK_K; +#if QK_K == 256 + dequantize_block_q2_K<<>>(vx, y); +#else + dequantize_block_q2_K<<>>(vx, y); +#endif +} + +template +static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb = k / QK_K; +#if QK_K == 256 + dequantize_block_q3_K<<>>(vx, y); +#else + dequantize_block_q3_K<<>>(vx, y); +#endif +} + +template +static void dequantize_row_q4_0_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb32 = k / 32; + const int nb = (k + 255) / 256; + dequantize_block_q4_0<<>>(vx, y, nb32); +} + +template +static void dequantize_row_q4_1_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb32 = k / 32; + const int nb = (k + 255) / 256; + dequantize_block_q4_1<<>>(vx, y, nb32); +} + +template +static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb = k / QK_K; + dequantize_block_q4_K<<>>(vx, y); +} + +template +static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb = k / QK_K; +#if QK_K == 256 + dequantize_block_q5_K<<>>(vx, y); +#else + dequantize_block_q5_K<<>>(vx, y); +#endif +} + +template +static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb = k / QK_K; +#if QK_K == 256 + dequantize_block_q6_K<<>>(vx, y); +#else + dequantize_block_q6_K<<>>(vx, y); +#endif +} + +template +static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb = k / QK_K; + dequantize_block_iq2_xxs<<>>(vx, y); +} + +template +static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb = k / QK_K; + dequantize_block_iq2_xs<<>>(vx, y); +} + +template +static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb = k / QK_K; + dequantize_block_iq2_s<<>>(vx, y); +} + +template +static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb = k / QK_K; + dequantize_block_iq3_xxs<<>>(vx, y); +} + +template +static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb = k / QK_K; + dequantize_block_iq3_s<<>>(vx, y); +} + +template +static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb = k / QK_K; + dequantize_block_iq1_s<<>>(vx, y); +} + +template +static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb = (k + QK_K - 1) / QK_K; + dequantize_block_iq4_nl<<>>(vx, y); +} + +template +static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) { + const int nb = (k + QK_K - 1) / QK_K; +#if QK_K == 64 + dequantize_block_iq4_nl<<>>(vx, y); +#else + dequantize_block_iq4_xs<<>>(vx, y); +#endif +} + +template +static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + + const src_t * x = (src_t *) vx; + + y[i] = x[i]; +} + +template +static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE; + convert_unary<<>>(vx, y, k); +} + +to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) { + int id; + switch (type) { + case GGML_TYPE_Q4_0: + return dequantize_row_q4_0_cuda; + case GGML_TYPE_Q4_1: + return dequantize_row_q4_1_cuda; + case GGML_TYPE_Q5_0: + return dequantize_block_cuda; + case GGML_TYPE_Q5_1: + return dequantize_block_cuda; + case GGML_TYPE_Q8_0: + CUDA_CHECK(cudaGetDevice(&id)); + if (ggml_cuda_info().devices[id].cc >= CC_PASCAL) { + return dequantize_block_q8_0_f16_cuda; + } + return dequantize_block_cuda; + case GGML_TYPE_Q2_K: + return dequantize_row_q2_K_cuda; + case GGML_TYPE_Q3_K: + return dequantize_row_q3_K_cuda; + case GGML_TYPE_Q4_K: + return dequantize_row_q4_K_cuda; + case GGML_TYPE_Q5_K: + return dequantize_row_q5_K_cuda; + case GGML_TYPE_Q6_K: + return dequantize_row_q6_K_cuda; + case GGML_TYPE_IQ2_XXS: + return dequantize_row_iq2_xxs_cuda; + case GGML_TYPE_IQ2_XS: + return dequantize_row_iq2_xs_cuda; + case GGML_TYPE_IQ2_S: + return dequantize_row_iq2_s_cuda; + case GGML_TYPE_IQ3_XXS: + return dequantize_row_iq3_xxs_cuda; + case GGML_TYPE_IQ1_S: + return dequantize_row_iq1_s_cuda; + case GGML_TYPE_IQ4_NL: + return dequantize_row_iq4_nl_cuda; + case GGML_TYPE_IQ4_XS: + return dequantize_row_iq4_xs_cuda; + case GGML_TYPE_IQ3_S: + return dequantize_row_iq3_s_cuda; + case GGML_TYPE_F32: + return convert_unary_cuda; + default: + return nullptr; + } +} + +to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) { + switch (type) { + case GGML_TYPE_Q4_0: + return dequantize_row_q4_0_cuda; + case GGML_TYPE_Q4_1: + return dequantize_row_q4_1_cuda; + case GGML_TYPE_Q5_0: + return dequantize_block_cuda; + case GGML_TYPE_Q5_1: + return dequantize_block_cuda; + case GGML_TYPE_Q8_0: + return dequantize_block_cuda; + case GGML_TYPE_Q2_K: + return dequantize_row_q2_K_cuda; + case GGML_TYPE_Q3_K: + return dequantize_row_q3_K_cuda; + case GGML_TYPE_Q4_K: + return dequantize_row_q4_K_cuda; + case GGML_TYPE_Q5_K: + return dequantize_row_q5_K_cuda; + case GGML_TYPE_Q6_K: + return dequantize_row_q6_K_cuda; + case GGML_TYPE_IQ2_XXS: + return dequantize_row_iq2_xxs_cuda; + case GGML_TYPE_IQ2_XS: + return dequantize_row_iq2_xs_cuda; + case GGML_TYPE_IQ2_S: + return dequantize_row_iq2_s_cuda; + case GGML_TYPE_IQ3_XXS: + return dequantize_row_iq3_xxs_cuda; + case GGML_TYPE_IQ1_S: + return dequantize_row_iq1_s_cuda; + case GGML_TYPE_IQ4_NL: + return dequantize_row_iq4_nl_cuda; + case GGML_TYPE_IQ4_XS: + return dequantize_row_iq4_xs_cuda; + case GGML_TYPE_IQ3_S: + return dequantize_row_iq3_s_cuda; + case GGML_TYPE_F16: + return convert_unary_cuda; + default: + return nullptr; + } +} diff --git a/ggml-cuda/convert.cuh b/ggml-cuda/convert.cuh new file mode 100644 index 000000000..db34c0be9 --- /dev/null +++ b/ggml-cuda/convert.cuh @@ -0,0 +1,13 @@ +#include "common.cuh" + +#define CUDA_DEQUANTIZE_BLOCK_SIZE 256 + +template +using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int k, cudaStream_t stream); + +typedef to_t_cuda_t to_fp32_cuda_t; +typedef to_t_cuda_t to_fp16_cuda_t; + +to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type); + +to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type); diff --git a/ggml-cuda/cpy.cu b/ggml-cuda/cpy.cu new file mode 100644 index 000000000..16d9c8fff --- /dev/null +++ b/ggml-cuda/cpy.cu @@ -0,0 +1,461 @@ +#include "cpy.cuh" + +typedef void (*cpy_kernel_t)(const char * cx, char * cdst); + +static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + float * dsti = (float *) cdsti; + + *dsti = *xi; +} + +static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + half * dsti = (half *) cdsti; + + *dsti = __float2half(*xi); +} + +static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) { + const half * xi = (const half *) cxi; + half * dsti = (half *) cdsti; + + *dsti = *xi; +} + +static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) { + const half * xi = (const half *) cxi; + float * dsti = (float *) cdsti; + + *dsti = *xi; +} + +template +static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13) { + const int64_t i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= ne) { + return; + } + + // determine indices i03/i13, i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor + // then combine those indices with the corresponding byte offsets to get the total offsets + const int64_t i03 = i/(ne00 * ne01 * ne02); + const int64_t i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); + const int64_t i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; + const int64_t i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; + const int64_t x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; + + const int64_t i13 = i/(ne10 * ne11 * ne12); + const int64_t i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); + const int64_t i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; + const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; + const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13; + + cpy_1(cx + x_offset, cdst + dst_offset); +} + +static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q8_0 * dsti = (block_q8_0 *) cdsti; + + float amax = 0.0f; // absolute max + + for (int j = 0; j < QK8_0; j++) { + const float v = xi[j]; + amax = fmaxf(amax, fabsf(v)); + } + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + dsti->d = d; + + for (int j = 0; j < QK8_0; ++j) { + const float x0 = xi[j]*id; + + dsti->qs[j] = roundf(x0); + } +} + +static __device__ void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q4_0 * dsti = (block_q4_0 *) cdsti; + + float amax = 0.0f; + float vmax = 0.0f; + + for (int j = 0; j < QK4_0; ++j) { + const float v = xi[j]; + if (amax < fabsf(v)) { + amax = fabsf(v); + vmax = v; + } + } + + const float d = vmax / -8; + const float id = d ? 1.0f/d : 0.0f; + + dsti->d = d; + + for (int j = 0; j < QK4_0/2; ++j) { + const float x0 = xi[0 + j]*id; + const float x1 = xi[QK4_0/2 + j]*id; + + const uint8_t xi0 = min(15, (int8_t)(x0 + 8.5f)); + const uint8_t xi1 = min(15, (int8_t)(x1 + 8.5f)); + + dsti->qs[j] = xi0; + dsti->qs[j] |= xi1 << 4; + } +} + +static __device__ void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q4_1 * dsti = (block_q4_1 *) cdsti; + + float vmin = FLT_MAX; + float vmax = -FLT_MAX; + + for (int j = 0; j < QK4_1; ++j) { + const float v = xi[j]; + + if (v < vmin) vmin = v; + if (v > vmax) vmax = v; + } + + const float d = (vmax - vmin) / ((1 << 4) - 1); + const float id = d ? 1.0f/d : 0.0f; + + dsti->dm.x = d; + dsti->dm.y = vmin; + + for (int j = 0; j < QK4_1/2; ++j) { + const float x0 = (xi[0 + j] - vmin)*id; + const float x1 = (xi[QK4_1/2 + j] - vmin)*id; + + const uint8_t xi0 = min(15, (int8_t)(x0 + 0.5f)); + const uint8_t xi1 = min(15, (int8_t)(x1 + 0.5f)); + + dsti->qs[j] = xi0; + dsti->qs[j] |= xi1 << 4; + } +} + +static __device__ void cpy_blck_f32_q5_0(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q5_0 * dsti = (block_q5_0 *) cdsti; + + float amax = 0.0f; + float vmax = 0.0f; + + for (int j = 0; j < QK5_0; ++j) { + const float v = xi[j]; + if (amax < fabsf(v)) { + amax = fabsf(v); + vmax = v; + } + } + + const float d = vmax / -16; + const float id = d ? 1.0f/d : 0.0f; + + dsti->d = d; + + uint32_t qh = 0; + for (int j = 0; j < QK5_0/2; ++j) { + const float x0 = xi[0 + j]*id; + const float x1 = xi[QK5_0/2 + j]*id; + + const uint8_t xi0 = min(31, (int8_t)(x0 + 16.5f)); + const uint8_t xi1 = min(31, (int8_t)(x1 + 16.5f)); + + dsti->qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4); + qh |= ((xi0 & 0x10u) >> 4) << (j + 0); + qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2); + } + memcpy(dsti->qh, &qh, sizeof(qh)); +} + +static __device__ void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q5_1 * dsti = (block_q5_1 *) cdsti; + + float min = xi[0]; + float max = xi[0]; + + for (int j = 1; j < QK5_1; ++j) { + const float v = xi[j]; + min = v < min ? v : min; + max = v > max ? v : max; + } + + const float d = (max - min) / 31; + const float id = d ? 1.0f/d : 0.0f; + + dsti->dm.x = d; + dsti->dm.y = min; + + uint32_t qh = 0; + for (int j = 0; j < QK5_1/2; ++j) { + const float x0 = (xi[0 + j] - min)*id; + const float x1 = (xi[QK5_1/2 + j] - min)*id; + + const uint8_t xi0 = (uint8_t)(x0 + 0.5f); + const uint8_t xi1 = (uint8_t)(x1 + 0.5f); + + dsti->qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4); + qh |= ((xi0 & 0x10u) >> 4) << (j + 0); + qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2); + } + memcpy(dsti->qh, &qh, sizeof(qh)); +} + + +static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) { + if (x <= val[0]) return 0; + if (x >= val[n-1]) return n-1; + int ml = 0, mu = n-1; + while (mu-ml > 1) { + int mav = (ml+mu)/2; + if (x < val[mav]) mu = mav; else ml = mav; + } + return x - val[mu-1] < val[mu] - x ? mu-1 : mu; +} + +static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_iq4_nl * dsti = (block_iq4_nl *) cdsti; + + float amax = 0.0f; + float vmax = 0.0f; + + for (int j = 0; j < QK4_NL; ++j) { + const float v = xi[j]; + if (amax < fabsf(v)) { + amax = fabsf(v); + vmax = v; + } + } + + float d = vmax / kvalues_iq4nl[0]; + const float id = d ? 1.0f/d : 0.0f; + + float sumqx = 0, sumq2 = 0; + for (int j = 0; j < QK4_NL/2; ++j) { + const float x0 = xi[0 + j]*id; + const float x1 = xi[QK4_NL/2 + j]*id; + const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl, x0); + const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl, x1); + dsti->qs[j] = xi0 | (xi1 << 4); + const float v0 = kvalues_iq4nl[xi0]; + const float v1 = kvalues_iq4nl[xi1]; + const float w0 = xi[0 + j]*xi[0 + j]; + const float w1 = xi[QK4_NL/2 + j]*xi[QK4_NL/2 + j]; + sumqx += w0*v0*xi[j] + w1*v1*xi[QK4_NL/2 + j]; + sumq2 += w0*v0*v0 + w1*v1*v1; + } + + dsti->d = sumq2 > 0 ? sumqx/sumq2 : d; +} + +template +static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13) { + const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk; + + if (i >= ne) { + return; + } + + const int i03 = i/(ne00 * ne01 * ne02); + const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); + const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; + const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; + const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; + + const int i13 = i/(ne10 * ne11 * ne12); + const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); + const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; + const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; + const int dst_offset = (i10/qk)*nb10 + i11*nb11 + i12*nb12 + i13*nb13; + + cpy_blck(cx + x_offset, cdst + dst_offset); +} + +static void ggml_cpy_f16_f32_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + + const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; + cpy_f32_f16<<>> + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); +} + +static void ggml_cpy_f32_f32_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + + const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; + cpy_f32_f16<<>> + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); +} + +static void ggml_cpy_f32_f16_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + + const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; + cpy_f32_f16<<>> + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); +} + +static void ggml_cpy_f32_q8_0_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + + GGML_ASSERT(ne % QK8_0 == 0); + const int num_blocks = ne / QK8_0; + cpy_f32_q<<>> + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); +} + +static void ggml_cpy_f32_q4_0_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + + GGML_ASSERT(ne % QK4_0 == 0); + const int num_blocks = ne / QK4_0; + cpy_f32_q<<>> + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); +} + +static void ggml_cpy_f32_q4_1_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + + GGML_ASSERT(ne % QK4_1 == 0); + const int num_blocks = ne / QK4_1; + cpy_f32_q<<>> + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); +} + +static void ggml_cpy_f32_q5_0_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + + GGML_ASSERT(ne % QK5_0 == 0); + const int num_blocks = ne / QK5_0; + cpy_f32_q<<>> + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); +} + +static void ggml_cpy_f32_q5_1_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + + GGML_ASSERT(ne % QK5_1 == 0); + const int num_blocks = ne / QK5_1; + cpy_f32_q<<>> + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); +} + +static void ggml_cpy_f32_iq4_nl_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + + GGML_ASSERT(ne % QK4_NL == 0); + const int num_blocks = ne / QK4_NL; + cpy_f32_q<<>> + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); +} + +static void ggml_cpy_f16_f16_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + + const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; + cpy_f32_f16<<>> + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); +} + +void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) { + const int64_t ne = ggml_nelements(src0); + GGML_ASSERT(ne == ggml_nelements(src1)); + + GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX); + GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t ne02 = src0->ne[2]; + + //GGML_ASSERT(src0->ne[3] == 1); + + const int64_t nb00 = src0->nb[0]; + const int64_t nb01 = src0->nb[1]; + const int64_t nb02 = src0->nb[2]; + const int64_t nb03 = src0->nb[3]; + + const int64_t ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + const int64_t ne12 = src1->ne[2]; + + //GGML_ASSERT(src1->ne[3] == 1); + + const int64_t nb10 = src1->nb[0]; + const int64_t nb11 = src1->nb[1]; + const int64_t nb12 = src1->nb[2]; + const int64_t nb13 = src1->nb[3]; + + cudaStream_t main_stream = ctx.stream(); + + char * src0_ddc = (char *) src0->data; + char * src1_ddc = (char *) src1->data; + + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) { + ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) { + ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) { + ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) { + ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) { + ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) { + ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) { + ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) { + ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) { + ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) { + ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + } else { + fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__, + ggml_type_name(src0->type), ggml_type_name(src1->type)); + GGML_ASSERT(false); + } +} + +void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + ggml_cuda_cpy(ctx, src0, dst); +} diff --git a/ggml-cuda/cpy.cuh b/ggml-cuda/cpy.cuh new file mode 100644 index 000000000..f0b2c453b --- /dev/null +++ b/ggml-cuda/cpy.cuh @@ -0,0 +1,7 @@ +#include "common.cuh" + +#define CUDA_CPY_BLOCK_SIZE 32 + +void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1); + +void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/dequantize.cuh b/ggml-cuda/dequantize.cuh new file mode 100644 index 000000000..b54400632 --- /dev/null +++ b/ggml-cuda/dequantize.cuh @@ -0,0 +1,103 @@ +#include "common.cuh" + +static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, dfloat2 & v){ + const block_q4_0 * x = (const block_q4_0 *) vx; + + const dfloat d = x[ib].d; + + const int vui = x[ib].qs[iqs]; + + v.x = vui & 0xF; + v.y = vui >> 4; + +#ifdef GGML_CUDA_F16 + v = __hsub2(v, {8.0f, 8.0f}); + v = __hmul2(v, {d, d}); +#else + v.x = (v.x - 8.0f) * d; + v.y = (v.y - 8.0f) * d; +#endif // GGML_CUDA_F16 +} + +static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, dfloat2 & v){ + const block_q4_1 * x = (const block_q4_1 *) vx; + + const dfloat d = __low2half(x[ib].dm); + const dfloat m = __high2half(x[ib].dm); + + const int vui = x[ib].qs[iqs]; + + v.x = vui & 0xF; + v.y = vui >> 4; + +#ifdef GGML_CUDA_F16 + v = __hmul2(v, {d, d}); + v = __hadd2(v, {m, m}); +#else + v.x = (v.x * d) + m; + v.y = (v.y * d) + m; +#endif // GGML_CUDA_F16 +} + +static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, dfloat2 & v){ + const block_q5_0 * x = (const block_q5_0 *) vx; + + const dfloat d = x[ib].d; + + uint32_t qh; + memcpy(&qh, x[ib].qh, sizeof(qh)); + + const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; + const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10; + + v.x = ((x[ib].qs[iqs] & 0xf) | xh_0); + v.y = ((x[ib].qs[iqs] >> 4) | xh_1); + +#ifdef GGML_CUDA_F16 + v = __hsub2(v, {16.0f, 16.0f}); + v = __hmul2(v, {d, d}); +#else + v.x = (v.x - 16.0f) * d; + v.y = (v.y - 16.0f) * d; +#endif // GGML_CUDA_F16 +} + +static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, dfloat2 & v){ + const block_q5_1 * x = (const block_q5_1 *) vx; + + const dfloat d = __low2half(x[ib].dm); + const dfloat m = __high2half(x[ib].dm); + + uint32_t qh; + memcpy(&qh, x[ib].qh, sizeof(qh)); + + const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; + const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10; + + v.x = ((x[ib].qs[iqs] & 0xf) | xh_0); + v.y = ((x[ib].qs[iqs] >> 4) | xh_1); + +#ifdef GGML_CUDA_F16 + v = __hmul2(v, {d, d}); + v = __hadd2(v, {m, m}); +#else + v.x = (v.x * d) + m; + v.y = (v.y * d) + m; +#endif // GGML_CUDA_F16 +} + +static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, dfloat2 & v){ + const block_q8_0 * x = (const block_q8_0 *) vx; + + const dfloat d = x[ib].d; + + v.x = x[ib].qs[iqs + 0]; + v.y = x[ib].qs[iqs + 1]; + +#ifdef GGML_CUDA_F16 + v = __hmul2(v, {d, d}); +#else + v.x *= d; + v.y *= d; +#endif // GGML_CUDA_F16 +} diff --git a/ggml-cuda/diagmask.cu b/ggml-cuda/diagmask.cu new file mode 100644 index 000000000..4b713ba22 --- /dev/null +++ b/ggml-cuda/diagmask.cu @@ -0,0 +1,40 @@ +#include "diagmask.cuh" + +static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past) { + const int col = blockDim.y*blockIdx.y + threadIdx.y; + const int row = blockDim.x*blockIdx.x + threadIdx.x; + + if (col >= ncols) { + return; + } + + const int i = row*ncols + col; + //dst[i] = col > (n_past + row % rows_per_channel) ? -INFINITY : x[i]; + //dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU + dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX; +} + +static void diag_mask_inf_f32_cuda(const float * x, float * dst, const int ncols_x, const int nrows_x, const int rows_per_channel, const int n_past, cudaStream_t stream) { + const dim3 block_dims(1, CUDA_DIAG_MASK_INF_BLOCK_SIZE, 1); + const int block_num_x = (ncols_x + CUDA_DIAG_MASK_INF_BLOCK_SIZE - 1) / CUDA_DIAG_MASK_INF_BLOCK_SIZE; + const dim3 block_nums(nrows_x, block_num_x, 1); + diag_mask_inf_f32<<>>(x, dst, ncols_x, rows_per_channel, n_past); +} + +void ggml_cuda_op_diag_mask_inf(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int nrows0 = ggml_nrows(src0); + + const int n_past = ((int32_t *) dst->op_params)[0]; + + diag_mask_inf_f32_cuda(src0_d, dst_d, ne00, nrows0, ne01, n_past, stream); +} diff --git a/ggml-cuda/diagmask.cuh b/ggml-cuda/diagmask.cuh new file mode 100644 index 000000000..6cdbef17e --- /dev/null +++ b/ggml-cuda/diagmask.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_DIAG_MASK_INF_BLOCK_SIZE 32 + +void ggml_cuda_op_diag_mask_inf(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/dmmv.cu b/ggml-cuda/dmmv.cu new file mode 100644 index 000000000..7156c9219 --- /dev/null +++ b/ggml-cuda/dmmv.cu @@ -0,0 +1,821 @@ +#include "dmmv.cuh" +#include "dequantize.cuh" +#include "convert.cuh" + +// dmmv = dequantize_mul_mat_vec +#ifndef GGML_CUDA_DMMV_X +#define GGML_CUDA_DMMV_X 32 +#endif +#ifndef GGML_CUDA_MMV_Y +#define GGML_CUDA_MMV_Y 1 +#endif + +#ifndef K_QUANTS_PER_ITERATION +#define K_QUANTS_PER_ITERATION 2 +#else +static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2"); +#endif + +static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) { + + static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION"); + + const int row = blockIdx.x*blockDim.y + threadIdx.y; + if (row > nrows) return; + + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q2_K * x = (const block_q2_K *)vx + ib0; + + float tmp = 0; // partial sum for thread in warp + +#if QK_K == 256 + const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...15 + const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0,1 + + const int step = 16/K_QUANTS_PER_ITERATION; + + const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... + const int in = tid - step*im; // 0...15 or 0...7 + + const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 or 0...14 in steps of 2 + const int q_offset = 32*im + l0; + const int s_offset = 8*im; + const int y_offset = 128*im + l0; + + uint32_t aux[4]; + const uint8_t * d = (const uint8_t *)aux; + const uint8_t * m = (const uint8_t *)(aux + 2); + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + y_offset; + const uint8_t * q = x[i].qs + q_offset; + + const float dall = __low2half(x[i].dm); + const float dmin = __high2half(x[i].dm); + + const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset); + aux[0] = a[0] & 0x0f0f0f0f; + aux[1] = a[1] & 0x0f0f0f0f; + aux[2] = (a[0] >> 4) & 0x0f0f0f0f; + aux[3] = (a[1] >> 4) & 0x0f0f0f0f; + + float sum1 = 0, sum2 = 0; + for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { + sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3) + + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3) + + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3) + + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3) + + y[l+16] * d[1] * ((q[l+16] >> 0) & 3) + + y[l+48] * d[3] * ((q[l+16] >> 2) & 3) + + y[l+80] * d[5] * ((q[l+16] >> 4) & 3) + +y[l+112] * d[7] * ((q[l+16] >> 6) & 3); + sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6] + + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7]; + + } + tmp += dall * sum1 - dmin * sum2; + + } +#else + const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7 + const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3 + const int offset = tid * K_QUANTS_PER_ITERATION; + + uint32_t uaux[2]; + const uint8_t * d = (const uint8_t *)uaux; + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + offset; + const uint8_t * q = x[i].qs + offset; + const uint32_t * s = (const uint32_t *)x[i].scales; + + uaux[0] = s[0] & 0x0f0f0f0f; + uaux[1] = (s[0] >> 4) & 0x0f0f0f0f; + + const float2 dall = __half22float2(x[i].dm); + + float sum1 = 0, sum2 = 0; + for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { + const uint8_t ql = q[l]; + sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3) + + y[l+16] * d[1] * ((ql >> 2) & 3) + + y[l+32] * d[2] * ((ql >> 4) & 3) + + y[l+48] * d[3] * ((ql >> 6) & 3); + sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7]; + } + tmp += dall.x * sum1 - dall.y * sum2; + } +#endif + + // sum up partial sums and write back result + tmp = warp_reduce_sum(tmp); + + if (threadIdx.x == 0) { + dst[row] = tmp; + } +} + +static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) { + + const int row = blockIdx.x*blockDim.y + threadIdx.y; + if (row > nrows) return; + + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q3_K * x = (const block_q3_K *)vx + ib0; + + float tmp = 0; // partial sum for thread in warp + +#if QK_K == 256 + + const uint16_t kmask1 = 0x0303; + const uint16_t kmask2 = 0x0f0f; + + const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16 + const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0,1 + + const int n = K_QUANTS_PER_ITERATION; // iterations in the inner loop + const int step = 16/K_QUANTS_PER_ITERATION; + const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... + const int in = tid - step*im; // 0....15 or 0...7 + + const uint8_t m = 1 << (4*im); + + const int l0 = n*in; // 0...15 or 0...14 in steps of 2 + const int q_offset = 32*im + l0; + const int y_offset = 128*im + l0; + + uint16_t utmp[4]; + const int8_t * s = (const int8_t *)utmp; + + const uint16_t s_shift = 4*im; + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + y_offset; + const uint8_t * q = x[i].qs + q_offset; + const uint8_t * h = x[i].hmask + l0; + + const uint16_t * a = (const uint16_t *)x[i].scales; + utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4); + utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4); + utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4); + utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4); + + const float d = x[i].d; + + float sum = 0; + for (int l = 0; l < n; ++l) { + sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4)) + + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4)) + + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4)) + + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4)); + sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4)) + + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4)) + + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4)) + + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4)); + } + tmp += d * sum; + + } +#else + + const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7 + const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3 + const int offset = tid * K_QUANTS_PER_ITERATION; // 0...15 or 0...14 + const int in = offset/8; // 0 or 1 + const int im = offset%8; // 0...7 + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + offset; + const uint8_t * q = x[i].qs + offset; + const uint8_t * s = x[i].scales; + + const float dall = (float)x[i].d; + + float sum = 0; + for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { + const uint8_t hl = x[i].hmask[im+l] >> in; + const uint8_t ql = q[l]; + sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4)) + + y[l+16] * dall * ((s[0] >> 4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4)) + + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4)) + + y[l+48] * dall * ((s[1] >> 4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4)); + } + tmp += sum; + } +#endif + + // sum up partial sums and write back result + tmp = warp_reduce_sum(tmp); + + if (threadIdx.x == 0) { + dst[row] = tmp; + } +} + +static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) { + + const int row = blockIdx.x*blockDim.y + threadIdx.y; + if (row > nrows) return; + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q4_K * x = (const block_q4_K *)vx + ib0; + +#if QK_K == 256 + const uint16_t kmask1 = 0x3f3f; + const uint16_t kmask2 = 0x0f0f; + const uint16_t kmask3 = 0xc0c0; + + const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16 + const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0,1 + + const int step = 8/K_QUANTS_PER_ITERATION; // 8 or 4 + + const int il = tid/step; // 0...3 + const int ir = tid - step*il; // 0...7 or 0...3 + const int n = 2 * K_QUANTS_PER_ITERATION; // 2 or 4 + + const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 + const int in = il%2; + + const int l0 = n*(2*ir + in); + const int q_offset = 32*im + l0; + const int y_offset = 64*im + l0; + + uint16_t aux[4]; + const uint8_t * sc = (const uint8_t *)aux; + +#if K_QUANTS_PER_ITERATION == 2 + uint32_t q32[4]; + const uint8_t * q4 = (const uint8_t *)q32; +#else + uint16_t q16[4]; + const uint8_t * q4 = (const uint8_t *)q16; +#endif + + float tmp = 0; // partial sum for thread in warp + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + const float * y1 = yy + i*QK_K + y_offset; + const float * y2 = y1 + 128; + + const float dall = __low2half(x[i].dm); + const float dmin = __high2half(x[i].dm); + + const uint16_t * a = (const uint16_t *)x[i].scales; + aux[0] = a[im+0] & kmask1; + aux[1] = a[im+2] & kmask1; + aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2); + aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2); + +#if K_QUANTS_PER_ITERATION == 2 + const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset); + const uint32_t * q2 = q1 + 16; + + q32[0] = q1[0] & 0x0f0f0f0f; + q32[1] = q1[0] & 0xf0f0f0f0; + q32[2] = q2[0] & 0x0f0f0f0f; + q32[3] = q2[0] & 0xf0f0f0f0; + + float4 s = {0.f, 0.f, 0.f, 0.f}; + float smin = 0; + for (int l = 0; l < 4; ++l) { + s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+ 4]; + s.z += y2[l] * q4[l+8]; s.w += y2[l+32] * q4[l+12]; + smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7]; + } + tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin; +#else + const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset); + const uint16_t * q2 = q1 + 32; + + q16[0] = q1[0] & 0x0f0f; + q16[1] = q1[0] & 0xf0f0; + q16[2] = q2[0] & 0x0f0f; + q16[3] = q2[0] & 0xf0f0; + + float4 s = {0.f, 0.f, 0.f, 0.f}; + float smin = 0; + for (int l = 0; l < 2; ++l) { + s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2]; + s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6]; + smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7]; + } + tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin; +#endif + + } +#else + const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 + const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); + + const int step = tid * K_QUANTS_PER_ITERATION; + + uint16_t aux16[2]; + const uint8_t * s = (const uint8_t *)aux16; + + float tmp = 0; + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + const uint8_t * q = x[i].qs + step; + const float * y = yy + i*QK_K + step; + const uint16_t * a = (const uint16_t *)x[i].scales; + aux16[0] = a[0] & 0x0f0f; + aux16[1] = (a[0] >> 4) & 0x0f0f; + const float d = (float)x[i].dm[0]; + const float m = (float)x[i].dm[1]; + float sum = 0.f; + for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { + sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2]) + + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2]) + + y[j+32] * (d * s[1] * (q[j+ 0] >> 4) - m * s[3]) + + y[j+48] * (d * s[1] * (q[j+16] >> 4) - m * s[3]); + } + tmp += sum; + } + +#endif + + // sum up partial sums and write back result + tmp = warp_reduce_sum(tmp); + + if (tid == 0) { + dst[row] = tmp; + } +} + +static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols) { + + const int row = blockIdx.x; + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q5_K * x = (const block_q5_K *)vx + ib0; + + float tmp = 0; // partial sum for thread in warp + +#if QK_K == 256 + const uint16_t kmask1 = 0x3f3f; + const uint16_t kmask2 = 0x0f0f; + const uint16_t kmask3 = 0xc0c0; + + const int tid = threadIdx.x/2; // 0...15 + const int ix = threadIdx.x%2; + + const int il = tid/4; // 0...3 + const int ir = tid - 4*il;// 0...3 + const int n = 2; + + const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 + const int in = il%2; + + const int l0 = n*(2*ir + in); + const int q_offset = 32*im + l0; + const int y_offset = 64*im + l0; + + const uint8_t hm1 = 1 << (2*im); + const uint8_t hm2 = hm1 << 4; + + uint16_t aux[4]; + const uint8_t * sc = (const uint8_t *)aux; + + uint16_t q16[8]; + const uint8_t * q4 = (const uint8_t *)q16; + + for (int i = ix; i < num_blocks_per_row; i += 2) { + + const uint8_t * ql1 = x[i].qs + q_offset; + const uint8_t * qh = x[i].qh + l0; + const float * y1 = yy + i*QK_K + y_offset; + const float * y2 = y1 + 128; + + const float dall = __low2half(x[i].dm); + const float dmin = __high2half(x[i].dm); + + const uint16_t * a = (const uint16_t *)x[i].scales; + aux[0] = a[im+0] & kmask1; + aux[1] = a[im+2] & kmask1; + aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2); + aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2); + + float4 sum = {0.f, 0.f, 0.f, 0.f}; + float smin = 0; + const uint16_t * q1 = (const uint16_t *)ql1; + const uint16_t * q2 = q1 + 32; + q16[0] = q1[0] & 0x0f0f; + q16[1] = q1[8] & 0x0f0f; + q16[2] = (q1[0] >> 4) & 0x0f0f; + q16[3] = (q1[8] >> 4) & 0x0f0f; + q16[4] = q2[0] & 0x0f0f; + q16[5] = q2[8] & 0x0f0f; + q16[6] = (q2[0] >> 4) & 0x0f0f; + q16[7] = (q2[8] >> 4) & 0x0f0f; + for (int l = 0; l < n; ++l) { + sum.x += y1[l+ 0] * (q4[l +0] + (qh[l+ 0] & (hm1 << 0) ? 16 : 0)) + + y1[l+16] * (q4[l +2] + (qh[l+16] & (hm1 << 0) ? 16 : 0)); + sum.y += y1[l+32] * (q4[l +4] + (qh[l+ 0] & (hm1 << 1) ? 16 : 0)) + + y1[l+48] * (q4[l +6] + (qh[l+16] & (hm1 << 1) ? 16 : 0)); + sum.z += y2[l+ 0] * (q4[l +8] + (qh[l+ 0] & (hm2 << 0) ? 16 : 0)) + + y2[l+16] * (q4[l+10] + (qh[l+16] & (hm2 << 0) ? 16 : 0)); + sum.w += y2[l+32] * (q4[l+12] + (qh[l+ 0] & (hm2 << 1) ? 16 : 0)) + + y2[l+48] * (q4[l+14] + (qh[l+16] & (hm2 << 1) ? 16 : 0)); + smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3] + + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7]; + } + tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin; + } + +#else + const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 + const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); + const int step = tid * K_QUANTS_PER_ITERATION; + const int im = step/8; + const int in = step%8; + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + const uint8_t * q = x[i].qs + step; + const int8_t * s = x[i].scales; + const float * y = yy + i*QK_K + step; + const float d = x[i].d; + float sum = 0.f; + for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { + const uint8_t h = x[i].qh[in+j] >> im; + sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16)) + + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16)) + + y[j+32] * d * s[2] * ((q[j+ 0] >> 4) - ((h >> 4) & 1 ? 0 : 16)) + + y[j+48] * d * s[3] * ((q[j+16] >> 4) - ((h >> 6) & 1 ? 0 : 16)); + } + tmp += sum; + } +#endif + + // sum up partial sums and write back result + tmp = warp_reduce_sum(tmp); + + if (threadIdx.x == 0) { + dst[row] = tmp; + } +} + +static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) { + + static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION"); + + const int row = blockIdx.x*blockDim.y + threadIdx.y; + if (row > nrows) return; + + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q6_K * x = (const block_q6_K *)vx + ib0; + +#if QK_K == 256 + + const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16 + const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1 + + const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8 + + const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... + const int in = tid - step*im; // 0...15 or 0...7 + +#if K_QUANTS_PER_ITERATION == 1 + const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 + const int is = 0; +#else + const int l0 = 4 * in; // 0, 4, 8, ..., 28 + const int is = in / 4; +#endif + const int ql_offset = 64*im + l0; + const int qh_offset = 32*im + l0; + const int s_offset = 8*im + is; + const int y_offset = 128*im + l0; + + float tmp = 0; // partial sum for thread in warp + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + y_offset; + const uint8_t * ql = x[i].ql + ql_offset; + const uint8_t * qh = x[i].qh + qh_offset; + const int8_t * s = x[i].scales + s_offset; + + const float d = x[i].d; + +#if K_QUANTS_PER_ITERATION == 1 + float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32) + + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32) + + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32) + + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32) + + y[64] * s[4] * d * ((int8_t)((ql[ 0] >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32) + + y[80] * s[5] * d * ((int8_t)((ql[16] >> 4) | ((qh[16] & 0x30) >> 0)) - 32) + + y[96] * s[6] * d * ((int8_t)((ql[32] >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32) + +y[112] * s[7] * d * ((int8_t)((ql[48] >> 4) | ((qh[16] & 0xc0) >> 2)) - 32); + tmp += sum; +#else + float sum = 0; + for (int l = 0; l < 4; ++l) { + sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32) + + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32) + + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32) + + y[l+96] * s[6] * d * ((int8_t)((ql[l+32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32); + } + tmp += sum; +#endif + + } + +#else + + const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...7 + const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0...3 + + const int step = tid * K_QUANTS_PER_ITERATION; + + float tmp = 0; // partial sum for thread in warp + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + step; + const uint8_t * ql = x[i].ql + step; + const uint8_t * qh = x[i].qh + step; + const int8_t * s = x[i].scales; + + const float d = x[i+0].d; + + float sum = 0; + for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { + sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32) + + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32) + + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >> 4) | ((qh[j] & 0x30) >> 0)) - 32) + + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >> 4) | ((qh[j] & 0xc0) >> 2)) - 32); + } + tmp += sum; + + } + +#endif + + // sum up partial sums and write back result + tmp = warp_reduce_sum(tmp); + + if (tid == 0) { + dst[row] = tmp; + } +} + +static __device__ void convert_f16(const void * vx, const int ib, const int iqs, dfloat2 & v){ + const half * x = (const half *) vx; + + // automatic half -> float type cast if dfloat == float + v.x = x[ib + iqs + 0]; + v.y = x[ib + iqs + 1]; +} + +template +static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) { + // qk = quantized weights per x block + // qr = number of quantized weights per data value in x block + const int row = blockIdx.x*blockDim.y + threadIdx.y; + + if (row >= nrows) { + return; + } + + const int tid = threadIdx.x; + + const int iter_stride = 2*GGML_CUDA_DMMV_X; + const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter + const int y_offset = qr == 1 ? 1 : qk/2; + +// partial sum for each thread +#ifdef GGML_CUDA_F16 + half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics +#else + float tmp = 0.0f; +#endif // GGML_CUDA_F16 + + for (int i = 0; i < ncols; i += iter_stride) { + const int col = i + vals_per_iter*tid; + const int ib = (row*ncols + col)/qk; // x block index + const int iqs = (col%qk)/qr; // x quant index + const int iybs = col - col%qk; // y block start index + +// processing >2 values per i iter is faster for fast GPUs +#pragma unroll + for (int j = 0; j < vals_per_iter; j += 2) { + // process 2 vals per j iter + + // dequantize + // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val + dfloat2 v; + dequantize_kernel(vx, ib, iqs + j/qr, v); + + // matrix multiplication + // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2 +#ifdef GGML_CUDA_F16 + tmp += __hmul2(v, { + y[iybs + iqs + j/qr + 0], + y[iybs + iqs + j/qr + y_offset] + }); +#else + tmp += v.x * y[iybs + iqs + j/qr + 0]; + tmp += v.y * y[iybs + iqs + j/qr + y_offset]; +#endif // GGML_CUDA_F16 + } + } + + // sum up partial sums and write back result + tmp = warp_reduce_sum(tmp); + + if (tid == 0) { +#ifdef GGML_CUDA_F16 + dst[row] = tmp.x + tmp.y; +#else + dst[row] = tmp; +#endif // GGML_CUDA_F16 + } +} + +static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); + const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; + // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead + const dim3 block_nums(block_num_y, 1, 1); + const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols, nrows); +} + +static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); + const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; + const dim3 block_nums(block_num_y, 1, 1); + const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols, nrows); +} + +static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); + const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; + const dim3 block_nums(block_num_y, 1, 1); + const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols, nrows); +} + +static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); + const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; + const dim3 block_nums(block_num_y, 1, 1); + const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols, nrows); +} + +static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); + const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; + const dim3 block_nums(block_num_y, 1, 1); + const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols, nrows); +} + +static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2 + const int block_num_y = (nrows + ny - 1) / ny; + const dim3 block_nums(block_num_y, 1, 1); + const dim3 block_dims(32, ny, 1); + dequantize_mul_mat_vec_q2_k<<>>(vx, y, dst, ncols, nrows); +} + +static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int ny = 2 / K_QUANTS_PER_ITERATION; + const int block_num_y = (nrows + ny - 1) / ny; + const dim3 block_nums(block_num_y, 1, 1); + const dim3 block_dims(32, ny, 1); + dequantize_mul_mat_vec_q3_k<<>>(vx, y, dst, ncols, nrows); +} + +static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int ny = 2 / K_QUANTS_PER_ITERATION; + const int block_num_y = (nrows + ny - 1) / ny; + const dim3 block_nums(block_num_y, 1, 1); + const dim3 block_dims(32, ny, 1); + dequantize_mul_mat_vec_q4_k<<>>(vx, y, dst, ncols, nrows); +} + +static void dequantize_mul_mat_vec_q5_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % QK_K == 0); + const dim3 block_dims(32, 1, 1); + dequantize_mul_mat_vec_q5_k<<>>(vx, y, dst, ncols); +} + +static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int ny = 2 / K_QUANTS_PER_ITERATION; + const int block_num_y = (nrows + ny - 1) / ny; + const dim3 block_nums(block_num_y, 1, 1); + const dim3 block_dims(32, ny, 1); + dequantize_mul_mat_vec_q6_k<<>>(vx, y, dst, ncols, nrows); +} + +static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); + const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; + const dim3 block_nums(block_num_y, 1, 1); + const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); + dequantize_mul_mat_vec<1, 1, convert_f16> + <<>>(vx, y, dst, ncols, nrows); +} + +void ggml_cuda_op_dequantize_mul_mat_vec( + ggml_backend_cuda_context & ctx, + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, + const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, + const int64_t src1_padded_row_size, cudaStream_t stream) { + GGML_UNUSED(ctx); + const int64_t ne00 = src0->ne[0]; + const int64_t row_diff = row_high - row_low; + + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics +#ifdef GGML_CUDA_F16 + ggml_cuda_pool_alloc src1_dfloat_a(ctx.pool()); + half * src1_dfloat = nullptr; // dfloat == half + + bool src1_convert_f16 = + src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 || + src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 || + src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16; + + if (src1_convert_f16) { + src1_dfloat = src1_dfloat_a.alloc(ne00); + const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type); + GGML_ASSERT(to_fp16_cuda != nullptr); + to_fp16_cuda(src1_ddf_i, src1_dfloat, ne00, stream); + } +#else + const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion +#endif // GGML_CUDA_F16 + + switch (src0->type) { + case GGML_TYPE_Q4_0: + dequantize_mul_mat_vec_q4_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_1: + dequantize_mul_mat_vec_q4_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_0: + dequantize_mul_mat_vec_q5_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_1: + dequantize_mul_mat_vec_q5_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q8_0: + dequantize_mul_mat_vec_q8_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q2_K: + dequantize_mul_mat_vec_q2_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q3_K: + dequantize_mul_mat_vec_q3_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_K: + dequantize_mul_mat_vec_q4_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_K: + dequantize_mul_mat_vec_q5_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q6_K: + dequantize_mul_mat_vec_q6_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_F16: + convert_mul_mat_vec_f16_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + default: + GGML_ASSERT(false); + break; + } + + GGML_UNUSED(src1); + GGML_UNUSED(dst); + GGML_UNUSED(src1_ddq_i); + GGML_UNUSED(src1_ncols); + GGML_UNUSED(src1_padded_row_size); +} diff --git a/ggml-cuda/dmmv.cuh b/ggml-cuda/dmmv.cuh new file mode 100644 index 000000000..3802678ff --- /dev/null +++ b/ggml-cuda/dmmv.cuh @@ -0,0 +1,7 @@ +#include "common.cuh" + +void ggml_cuda_op_dequantize_mul_mat_vec( + ggml_backend_cuda_context & ctx, + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, + const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, + const int64_t src1_padded_row_size, cudaStream_t stream); diff --git a/ggml-cuda/getrows.cu b/ggml-cuda/getrows.cu new file mode 100644 index 000000000..55af195fd --- /dev/null +++ b/ggml-cuda/getrows.cu @@ -0,0 +1,178 @@ +#include "getrows.cuh" +#include "dequantize.cuh" + +template +static __global__ void k_get_rows( + const void * src0, const int32_t * src1, dst_t * dst, + int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ + /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ + /*size_t s0,*/ size_t s1, size_t s2, size_t s3, + /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, + size_t s10, size_t s11, size_t s12/*, size_t s13*/) { + + const int i00 = (blockIdx.x*blockDim.x + threadIdx.x)*2; + const int i10 = blockDim.y*blockIdx.y + threadIdx.y; + const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12; + const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12; + + if (i00 >= ne00) { + return; + } + + const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; + + dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; + const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03; + + const int ib = i00/qk; // block index + const int iqs = (i00%qk)/qr; // quant index + const int iybs = i00 - i00%qk; // dst block start index + const int y_offset = qr == 1 ? 1 : qk/2; + + // dequantize + dfloat2 v; + dequantize_kernel(src0_row, ib, iqs, v); + + dst_row[iybs + iqs + 0] = v.x; + dst_row[iybs + iqs + y_offset] = v.y; +} + +template +static __global__ void k_get_rows_float( + const src0_t * src0, const int32_t * src1, dst_t * dst, + int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ + /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ + /*size_t s0,*/ size_t s1, size_t s2, size_t s3, + /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, + size_t s10, size_t s11, size_t s12/*, size_t s13*/) { + + const int i00 = blockIdx.x*blockDim.x + threadIdx.x; + const int i10 = blockDim.y*blockIdx.y + threadIdx.y; + const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12; + const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12; + + if (i00 >= ne00) { + return; + } + + const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; + + dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; + const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03); + + dst_row[i00] = src0_row[i00]; +} + +template +static void get_rows_cuda(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) { + + GGML_TENSOR_BINARY_OP_LOCALS + + const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1); + const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE); + const dim3 block_nums(block_num_x, ne10, ne11*ne12); + + // strides in elements + //const size_t s0 = nb0 / ggml_element_size(dst); + const size_t s1 = nb1 / ggml_element_size(dst); + const size_t s2 = nb2 / ggml_element_size(dst); + const size_t s3 = nb3 / ggml_element_size(dst); + + const size_t s10 = nb10 / ggml_element_size(src1); + const size_t s11 = nb11 / ggml_element_size(src1); + const size_t s12 = nb12 / ggml_element_size(src1); + //const size_t s13 = nb13 / ggml_element_size(src1); + + GGML_ASSERT(ne00 % 2 == 0); + + k_get_rows<<>>( + src0_dd, src1_dd, dst_dd, + ne00, /*ne01, ne02, ne03,*/ + /*ne10, ne11,*/ ne12, /*ne13,*/ + /* s0,*/ s1, s2, s3, + /* nb00,*/ nb01, nb02, nb03, + s10, s11, s12/*, s13*/); + + GGML_UNUSED(dst); +} + +template +static void get_rows_cuda_float(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) { + + GGML_TENSOR_BINARY_OP_LOCALS + + const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1); + const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE; + const dim3 block_nums(block_num_x, ne10, ne11*ne12); + + // strides in elements + //const size_t s0 = nb0 / ggml_element_size(dst); + const size_t s1 = nb1 / ggml_element_size(dst); + const size_t s2 = nb2 / ggml_element_size(dst); + const size_t s3 = nb3 / ggml_element_size(dst); + + const size_t s10 = nb10 / ggml_element_size(src1); + const size_t s11 = nb11 / ggml_element_size(src1); + const size_t s12 = nb12 / ggml_element_size(src1); + //const size_t s13 = nb13 / ggml_element_size(src1); + + k_get_rows_float<<>>( + src0_dd, src1_dd, dst_dd, + ne00, /*ne01, ne02, ne03,*/ + /*ne10, ne11,*/ ne12, /*ne13,*/ + /* s0,*/ s1, s2, s3, + /* nb00,*/ nb01, nb02, nb03, + s10, s11, s12/*, s13*/); + + GGML_UNUSED(dst); +} + +void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + const float * src0_d = (const float *)src0->data; + const float * src1_d = (const float *)src1->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + + GGML_ASSERT(src1->type == GGML_TYPE_I32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type)); + GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type)); + GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type)); + + const int32_t * src1_i32 = (const int32_t *) src1_d; + + switch (src0->type) { + case GGML_TYPE_F16: + get_rows_cuda_float(src0, src1, dst, (const half *)src0_d, src1_i32, dst_d, stream); + break; + case GGML_TYPE_F32: + get_rows_cuda_float(src0, src1, dst, src0_d, src1_i32, dst_d, stream); + break; + case GGML_TYPE_Q4_0: + get_rows_cuda(src0, src1, dst, src0_d, src1_i32, dst_d, stream); + break; + case GGML_TYPE_Q4_1: + get_rows_cuda(src0, src1, dst, src0_d, src1_i32, dst_d, stream); + break; + case GGML_TYPE_Q5_0: + get_rows_cuda(src0, src1, dst, src0_d, src1_i32, dst_d, stream); + break; + case GGML_TYPE_Q5_1: + get_rows_cuda(src0, src1, dst, src0_d, src1_i32, dst_d, stream); + break; + case GGML_TYPE_Q8_0: + get_rows_cuda(src0, src1, dst, src0_d, src1_i32, dst_d, stream); + break; + default: + // TODO: k-quants + fprintf(stderr, "%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type)); + GGML_ASSERT(false); + break; + } +} diff --git a/ggml-cuda/getrows.cuh b/ggml-cuda/getrows.cuh new file mode 100644 index 000000000..bbf130232 --- /dev/null +++ b/ggml-cuda/getrows.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_GET_ROWS_BLOCK_SIZE 256 + +void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/im2col.cu b/ggml-cuda/im2col.cu new file mode 100644 index 000000000..3d0d8d4e6 --- /dev/null +++ b/ggml-cuda/im2col.cu @@ -0,0 +1,104 @@ +#include "im2col.cuh" + +template +static __global__ void im2col_kernel( + const float * x, T * dst, int64_t batch_offset, + int64_t offset_delta, int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH, int64_t pelements, int64_t CHW, + int s0, int s1, int p0, int p1, int d0, int d1) { + const int64_t i = threadIdx.x + blockIdx.x * blockDim.x; + if (i >= pelements) { + return; + } + + const int64_t ksize = OW * (KH > 1 ? KW : 1); + const int64_t kx = i / ksize; + const int64_t kd = kx * ksize; + const int64_t ky = (i - kd) / OW; + const int64_t ix = i % OW; + + const int64_t oh = blockIdx.y; + const int64_t batch = blockIdx.z / IC; + const int64_t ic = blockIdx.z % IC; + + const int64_t iiw = ix * s0 + kx * d0 - p0; + const int64_t iih = oh * s1 + ky * d1 - p1; + + const int64_t offset_dst = + ((batch * OH + oh) * OW + ix) * CHW + + (ic * (KW * KH) + ky * KW + kx); + + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst[offset_dst] = 0.0f; + } else { + const int64_t offset_src = ic * offset_delta + batch * batch_offset; + dst[offset_dst] = x[offset_src + iih * IW + iiw]; + } +} + +template +static void im2col_cuda(const float * x, T* dst, + int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC, + int64_t batch, int64_t batch_offset, int64_t offset_delta, + int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) { + const int parallel_elements = OW * KW * KH; + const int num_blocks = (parallel_elements + CUDA_IM2COL_BLOCK_SIZE - 1) / CUDA_IM2COL_BLOCK_SIZE; + dim3 block_nums(num_blocks, OH, batch * IC); + im2col_kernel<<>>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, (IC * KH * KW), s0, s1, p0, p1, d0, d1); +} + +static void im2col_cuda_f16(const float * x, half * dst, + int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC, + int64_t batch, int64_t batch_offset, int64_t offset_delta, + int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) { + + im2col_cuda(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1, d0, d1, stream); +} + +static void im2col_cuda_f32(const float * x, float * dst, + int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC, + int64_t batch, int64_t batch_offset, int64_t offset_delta, + int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) { + + im2col_cuda(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1, d0, d1, stream); +} + +void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + const float * src1_d = (const float *)src1->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32); + + const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t*)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t*)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t*)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t*)(dst->op_params))[5]; + + const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1; + + const int64_t IC = src1->ne[is_2D ? 2 : 1]; + const int64_t IH = is_2D ? src1->ne[1] : 1; + const int64_t IW = src1->ne[0]; + + const int64_t KH = is_2D ? src0->ne[1] : 1; + const int64_t KW = src0->ne[0]; + + const int64_t OH = is_2D ? dst->ne[2] : 1; + const int64_t OW = dst->ne[1]; + + const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32 + const int64_t batch = src1->ne[3]; + const size_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32 + + if(dst->type == GGML_TYPE_F16) { + im2col_cuda_f16(src1_d, (half *) dst_d, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream); + } else { + im2col_cuda_f32(src1_d, (float *) dst_d, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream); + } +} diff --git a/ggml-cuda/im2col.cuh b/ggml-cuda/im2col.cuh new file mode 100644 index 000000000..1ce8fae4d --- /dev/null +++ b/ggml-cuda/im2col.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_IM2COL_BLOCK_SIZE 256 + +void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/mmq.cu b/ggml-cuda/mmq.cu new file mode 100644 index 000000000..60d6616a8 --- /dev/null +++ b/ggml-cuda/mmq.cu @@ -0,0 +1,2265 @@ +#include "mmq.cuh" +#include "vecdotq.cuh" + +typedef void (*allocate_tiles_cuda_t)(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc); +typedef void (*load_tiles_cuda_t)( + const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, + int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row); +typedef float (*vec_dot_q_mul_mat_cuda_t)( + const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, + const int * __restrict__ y_qs, const half2 * __restrict__ y_ms, const int & i, const int & j, const int & k); +typedef void (*dot_kernel_k_t)(const void * __restrict__ vx, const int ib, const int iqs, const float * __restrict__ y, float & v); + +template static __device__ __forceinline__ void allocate_tiles_q4_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + GGML_UNUSED(x_qh); + GGML_UNUSED(x_sc); + + __shared__ int tile_x_qs[mmq_y * (WARP_SIZE) + mmq_y]; + __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI4_0) + mmq_y/QI4_0]; + + *x_ql = tile_x_qs; + *x_dm = (half2 *) tile_x_d; +} + +template static __device__ __forceinline__ void load_tiles_q4_0( + const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, + int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + GGML_UNUSED(x_qh); GGML_UNUSED(x_sc); + GGML_CUDA_ASSUME(i_offset >= 0); + GGML_CUDA_ASSUME(i_offset < nwarps); + GGML_CUDA_ASSUME(k >= 0); + GGML_CUDA_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI4_0; + const int kqsx = k % QI4_0; + + const block_q4_0 * bx0 = (const block_q4_0 *) vx; + + float * x_dmf = (float *) x_dm; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = min(i, i_max); + } + + const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx); + // x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbx] = bxi->d; + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI4_0; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_0) { + int i = i0 + i_offset * QI4_0 + k / blocks_per_tile_x_row; + + if (need_check) { + i = min(i, i_max); + } + + const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbxd] = bxi->d; + } +} + +static __device__ __forceinline__ float vec_dot_q4_0_q8_1_mul_mat( + const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, + const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + GGML_UNUSED(x_qh); GGML_UNUSED(x_sc); + + const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); + const float * x_dmf = (const float *) x_dm; + + int u[2*VDR_Q4_0_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < VDR_Q4_0_Q8_1_MMQ; ++l) { + u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; + u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_0) % WARP_SIZE]; + } + + return vec_dot_q4_0_q8_1_impl + (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dmf[i * (WARP_SIZE/QI4_0) + i/QI4_0 + k/QI4_0], + y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); +} + +template static __device__ __forceinline__ void allocate_tiles_q4_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + GGML_UNUSED(x_qh); GGML_UNUSED(x_sc); + + __shared__ int tile_x_qs[mmq_y * (WARP_SIZE) + + mmq_y]; + __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI4_1) + mmq_y/QI4_1]; + + *x_ql = tile_x_qs; + *x_dm = tile_x_dm; +} + +template static __device__ __forceinline__ void load_tiles_q4_1( + const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, + int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + GGML_UNUSED(x_qh); GGML_UNUSED(x_sc); + + GGML_CUDA_ASSUME(i_offset >= 0); + GGML_CUDA_ASSUME(i_offset < nwarps); + GGML_CUDA_ASSUME(k >= 0); + GGML_CUDA_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI4_1; + const int kqsx = k % QI4_1; + + const block_q4_1 * bx0 = (const block_q4_1 *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = min(i, i_max); + } + + const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI4_1; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_1) { + int i = i0 + i_offset * QI4_1 + k / blocks_per_tile_x_row; + + if (need_check) { + i = min(i, i_max); + } + + const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dm[i * (WARP_SIZE/QI4_1) + i / QI4_1 + kbxd] = bxi->dm; + } +} + +static __device__ __forceinline__ float vec_dot_q4_1_q8_1_mul_mat( + const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, + const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + GGML_UNUSED(x_qh); GGML_UNUSED(x_sc); + + const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); + + int u[2*VDR_Q4_1_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) { + u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; + u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_1) % WARP_SIZE]; + } + + return vec_dot_q4_1_q8_1_impl + (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dm[i * (WARP_SIZE/QI4_1) + i/QI4_1 + k/QI4_1], + y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); +} + +template static __device__ __forceinline__ void allocate_tiles_q5_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + GGML_UNUSED(x_qh); GGML_UNUSED(x_sc); + + __shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y]; + __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI5_0) + mmq_y/QI5_0]; + + *x_ql = tile_x_ql; + *x_dm = (half2 *) tile_x_d; +} + +template static __device__ __forceinline__ void load_tiles_q5_0( + const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, + int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + GGML_UNUSED(x_qh); GGML_UNUSED(x_sc); + + GGML_CUDA_ASSUME(i_offset >= 0); + GGML_CUDA_ASSUME(i_offset < nwarps); + GGML_CUDA_ASSUME(k >= 0); + GGML_CUDA_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI5_0; + const int kqsx = k % QI5_0; + + const block_q5_0 * bx0 = (const block_q5_0 *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = min(i, i_max); + } + + const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbx; + + const int ql = get_int_from_uint8(bxi->qs, kqsx); + const int qh = get_int_from_uint8(bxi->qh, 0) >> (4 * (k % QI5_0)); + + int qs0 = (ql >> 0) & 0x0F0F0F0F; + qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 + qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 + qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 + qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 + qs0 = __vsubss4(qs0, 0x10101010); // subtract 16 + + x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0; + + int qs1 = (ql >> 4) & 0x0F0F0F0F; + qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 + qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 + qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 + qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 + qs1 = __vsubss4(qs1, 0x10101010); // subtract 16 + + x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1; + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI5_0; + const int kbxd = k % blocks_per_tile_x_row; + float * x_dmf = (float *) x_dm; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_0) { + int i = i0 + i_offset * QI5_0 + k / blocks_per_tile_x_row; + + if (need_check) { + i = min(i, i_max); + } + + const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI5_0) + i / QI5_0 + kbxd] = bxi->d; + } +} + +static __device__ __forceinline__ float vec_dot_q5_0_q8_1_mul_mat( + const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, + const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + GGML_UNUSED(x_qh); GGML_UNUSED(x_sc); + + const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); + const int index_bx = i * (WARP_SIZE/QI5_0) + i/QI5_0 + k/QI5_0; + const float * x_dmf = (const float *) x_dm; + const float * y_df = (const float *) y_ds; + + int u[2*VDR_Q5_0_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < VDR_Q5_0_Q8_1_MMQ; ++l) { + u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; + u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_0) % WARP_SIZE]; + } + + return vec_dot_q8_0_q8_1_impl + (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dmf[index_bx], y_df[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); +} + + +template static __device__ __forceinline__ void allocate_tiles_q5_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + GGML_UNUSED(x_qh); GGML_UNUSED(x_sc); + + __shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y]; + __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI5_1) + mmq_y/QI5_1]; + + *x_ql = tile_x_ql; + *x_dm = tile_x_dm; +} + +template static __device__ __forceinline__ void load_tiles_q5_1( + const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, + int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + GGML_UNUSED(x_qh); GGML_UNUSED(x_sc); + + GGML_CUDA_ASSUME(i_offset >= 0); + GGML_CUDA_ASSUME(i_offset < nwarps); + GGML_CUDA_ASSUME(k >= 0); + GGML_CUDA_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI5_1; + const int kqsx = k % QI5_1; + + const block_q5_1 * bx0 = (const block_q5_1 *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = min(i, i_max); + } + + const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbx; + + const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx); + const int qh = get_int_from_uint8_aligned(bxi->qh, 0) >> (4 * (k % QI5_1)); + + int qs0 = (ql >> 0) & 0x0F0F0F0F; + qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 + qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 + qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 + qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 + + x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0; + + int qs1 = (ql >> 4) & 0x0F0F0F0F; + qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 + qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 + qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 + qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 + + x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1; + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI5_1; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_1) { + int i = i0 + i_offset * QI5_1 + k / blocks_per_tile_x_row; + + if (need_check) { + i = min(i, i_max); + } + + const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dm[i * (WARP_SIZE/QI5_1) + i / QI5_1 + kbxd] = bxi->dm; + } +} + +static __device__ __forceinline__ float vec_dot_q5_1_q8_1_mul_mat( + const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, + const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + GGML_UNUSED(x_qh); GGML_UNUSED(x_sc); + + const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); + const int index_bx = i * (WARP_SIZE/QI5_1) + + i/QI5_1 + k/QI5_1; + + int u[2*VDR_Q5_1_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < VDR_Q5_1_Q8_1_MMQ; ++l) { + u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; + u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_1) % WARP_SIZE]; + } + + return vec_dot_q8_1_q8_1_impl + (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dm[index_bx], y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); +} + +template static __device__ __forceinline__ void allocate_tiles_q8_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + GGML_UNUSED(x_qh); GGML_UNUSED(x_sc); + + __shared__ int tile_x_qs[mmq_y * (WARP_SIZE) + mmq_y]; + __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI8_0) + mmq_y/QI8_0]; + + *x_ql = tile_x_qs; + *x_dm = (half2 *) tile_x_d; +} + +template static __device__ __forceinline__ void load_tiles_q8_0( + const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, + int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + GGML_UNUSED(x_qh); GGML_UNUSED(x_sc); + + GGML_CUDA_ASSUME(i_offset >= 0); + GGML_CUDA_ASSUME(i_offset < nwarps); + GGML_CUDA_ASSUME(k >= 0); + GGML_CUDA_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI8_0; + const int kqsx = k % QI8_0; + float * x_dmf = (float *) x_dm; + + const block_q8_0 * bx0 = (const block_q8_0 *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = min(i, i_max); + } + + const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_int8(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI8_0; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI8_0) { + int i = i0 + i_offset * QI8_0 + k / blocks_per_tile_x_row; + + if (need_check) { + i = min(i, i_max); + } + + const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI8_0) + i / QI8_0 + kbxd] = bxi->d; + } +} + +static __device__ __forceinline__ float vec_dot_q8_0_q8_1_mul_mat( + const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, + const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + GGML_UNUSED(x_qh); GGML_UNUSED(x_sc); + + const float * x_dmf = (const float *) x_dm; + const float * y_df = (const float *) y_ds; + + return vec_dot_q8_0_q8_1_impl + (&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[j * WARP_SIZE + k], x_dmf[i * (WARP_SIZE/QI8_0) + i/QI8_0 + k/QI8_0], + y_df[j * (WARP_SIZE/QI8_1) + k/QI8_1]); +} + +template static __device__ __forceinline__ void allocate_tiles_q2_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + GGML_UNUSED(x_qh); + + __shared__ int tile_x_ql[mmq_y * (WARP_SIZE) + mmq_y]; + __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI2_K) + mmq_y/QI2_K]; + __shared__ int tile_x_sc[mmq_y * (WARP_SIZE/4) + mmq_y/4]; + + *x_ql = tile_x_ql; + *x_dm = tile_x_dm; + *x_sc = tile_x_sc; +} + +template static __device__ __forceinline__ void load_tiles_q2_K( + const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, + int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + GGML_UNUSED(x_qh); + + GGML_CUDA_ASSUME(i_offset >= 0); + GGML_CUDA_ASSUME(i_offset < nwarps); + GGML_CUDA_ASSUME(k >= 0); + GGML_CUDA_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI2_K; + const int kqsx = k % QI2_K; + + const block_q2_K * bx0 = (const block_q2_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = min(i, i_max); + } + + const block_q2_K * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI2_K; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI2_K) { + int i = (i0 + i_offset * QI2_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = min(i, i_max); + } + + const block_q2_K * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dm[i * (WARP_SIZE/QI2_K) + i / QI2_K + kbxd] = bxi->dm; + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) { + int i = i0 + i_offset * 4 + k / (WARP_SIZE/4); + + if (need_check) { + i = min(i, i_max); + } + + const block_q2_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI2_K/4); + + x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = get_int_from_uint8_aligned(bxi->scales, k % (QI2_K/4)); + } +} + +static __device__ __forceinline__ float vec_dot_q2_K_q8_1_mul_mat( + const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, + const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + GGML_UNUSED(x_qh); + + const int kbx = k / QI2_K; + const int ky = (k % QI2_K) * QR2_K; + const float * y_df = (const float *) y_ds; + + int v[QR2_K*VDR_Q2_K_Q8_1_MMQ]; + + const int kqsx = i * (WARP_SIZE + 1) + kbx*QI2_K + (QI2_K/2) * (ky/(2*QI2_K)) + ky % (QI2_K/2); + const int shift = 2 * ((ky % (2*QI2_K)) / (QI2_K/2)); + +#pragma unroll + for (int l = 0; l < QR2_K*VDR_Q2_K_Q8_1_MMQ; ++l) { + v[l] = (x_ql[kqsx + l] >> shift) & 0x03030303; + } + + const uint8_t * scales = ((const uint8_t *) &x_sc[i * (WARP_SIZE/4) + i/4 + kbx*4]) + ky/4; + + const int index_y = j * WARP_SIZE + (QR2_K*k) % WARP_SIZE; + return vec_dot_q2_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dm[i * (WARP_SIZE/QI2_K) + i/QI2_K + kbx], y_df[index_y/QI8_1]); +} + +template static __device__ __forceinline__ void allocate_tiles_q3_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + + __shared__ int tile_x_ql[mmq_y * (WARP_SIZE) + mmq_y]; + __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI3_K) + mmq_y/QI3_K]; + __shared__ int tile_x_qh[mmq_y * (WARP_SIZE/2) + mmq_y/2]; + __shared__ int tile_x_sc[mmq_y * (WARP_SIZE/4) + mmq_y/4]; + + *x_ql = tile_x_ql; + *x_dm = tile_x_dm; + *x_qh = tile_x_qh; + *x_sc = tile_x_sc; +} + +template static __device__ __forceinline__ void load_tiles_q3_K( + const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, + int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + + GGML_CUDA_ASSUME(i_offset >= 0); + GGML_CUDA_ASSUME(i_offset < nwarps); + GGML_CUDA_ASSUME(k >= 0); + GGML_CUDA_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI3_K; + const int kqsx = k % QI3_K; + + const block_q3_K * bx0 = (const block_q3_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = min(i, i_max); + } + + const block_q3_K * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI3_K; + const int kbxd = k % blocks_per_tile_x_row; + float * x_dmf = (float *) x_dm; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI3_K) { + int i = (i0 + i_offset * QI3_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = min(i, i_max); + } + + const block_q3_K * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI3_K) + i / QI3_K + kbxd] = bxi->d; + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 2) { + int i = i0 + i_offset * 2 + k / (WARP_SIZE/2); + + if (need_check) { + i = min(i, i_max); + } + + const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/2)) / (QI3_K/2); + + // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted + x_qh[i * (WARP_SIZE/2) + i / 2 + k % (WARP_SIZE/2)] = ~get_int_from_uint8(bxi->hmask, k % (QI3_K/2)); + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) { + int i = i0 + i_offset * 4 + k / (WARP_SIZE/4); + + if (need_check) { + i = min(i, i_max); + } + + const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI3_K/4); + + const int ksc = k % (QI3_K/4); + + const int ksc_low = ksc % (QI3_K/8); + const int shift_low = 4 * (ksc / (QI3_K/8)); + const int sc_low = (get_int_from_uint8(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F; + + const int ksc_high = QI3_K/8; + const int shift_high = 2 * ksc; + const int sc_high = ((get_int_from_uint8(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030; + + const int sc = __vsubss4(sc_low | sc_high, 0x20202020); + + x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = sc; + } +} + +static __device__ __forceinline__ float vec_dot_q3_K_q8_1_mul_mat( + const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, + const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + + const int kbx = k / QI3_K; + const int ky = (k % QI3_K) * QR3_K; + const float * x_dmf = (const float *) x_dm; + const float * y_df = (const float *) y_ds; + + const int8_t * scales = ((const int8_t *) (x_sc + i * (WARP_SIZE/4) + i/4 + kbx*4)) + ky/4; + + int v[QR3_K*VDR_Q3_K_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < QR3_K*VDR_Q3_K_Q8_1_MMQ; ++l) { + const int kqsx = i * (WARP_SIZE + 1) + kbx*QI3_K + (QI3_K/2) * (ky/(2*QI3_K)) + ky % (QI3_K/2); + const int shift = 2 * ((ky % 32) / 8); + const int vll = (x_ql[kqsx + l] >> shift) & 0x03030303; + + const int vh = x_qh[i * (WARP_SIZE/2) + i/2 + kbx * (QI3_K/2) + (ky+l)%8] >> ((ky+l) / 8); + const int vlh = (vh << 2) & 0x04040404; + + v[l] = __vsubss4(vll, vlh); + } + + const int index_y = j * WARP_SIZE + (k*QR3_K) % WARP_SIZE; + return vec_dot_q3_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dmf[i * (WARP_SIZE/QI3_K) + i/QI3_K + kbx], y_df[index_y/QI8_1]); +} + +template static __device__ __forceinline__ void allocate_tiles_q4_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + GGML_UNUSED(x_qh); + + __shared__ int tile_x_ql[mmq_y * (WARP_SIZE) + mmq_y]; + __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI4_K) + mmq_y/QI4_K]; + __shared__ int tile_x_sc[mmq_y * (WARP_SIZE/8) + mmq_y/8]; + + *x_ql = tile_x_ql; + *x_dm = tile_x_dm; + *x_sc = tile_x_sc; +} + +template static __device__ __forceinline__ void load_tiles_q4_K( + const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, + int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + GGML_UNUSED(x_qh); + + GGML_CUDA_ASSUME(i_offset >= 0); + GGML_CUDA_ASSUME(i_offset < nwarps); + GGML_CUDA_ASSUME(k >= 0); + GGML_CUDA_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI4_K; // == 0 if QK_K == 256 + const int kqsx = k % QI4_K; // == k if QK_K == 256 + + const block_q4_K * bx0 = (const block_q4_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = min(i, i_max); + } + + const block_q4_K * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI4_K; // == 1 if QK_K == 256 + const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_K) { + int i = (i0 + i_offset * QI4_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = min(i, i_max); + } + + const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd; + +#if QK_K == 256 + x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm; +#else + x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]}; +#endif + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { + int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; + + if (need_check) { + i = min(i, i_max); + } + + const block_q4_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI4_K/8); + + const int * scales = (const int *) bxi->scales; + + const int ksc = k % (WARP_SIZE/8); + + // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8 + int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits + scales8 |= (scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030; // upper 2 bits + + x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8; + } +} + +static __device__ __forceinline__ float vec_dot_q4_K_q8_1_mul_mat( + const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, + const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + GGML_UNUSED(x_qh); + + const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2*((k % 16) / 8); + + const int index_y = j * WARP_SIZE + (QR4_K*k) % WARP_SIZE; + return vec_dot_q4_K_q8_1_impl_mmq(&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[index_y], sc, sc+8, + x_dm[i * (WARP_SIZE/QI4_K) + i/QI4_K], &y_ds[index_y/QI8_1]); +} + +template static __device__ __forceinline__ void allocate_tiles_q5_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + GGML_UNUSED(x_qh); + + __shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y]; + __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI5_K) + mmq_y/QI5_K]; + __shared__ int tile_x_sc[mmq_y * (WARP_SIZE/8) + mmq_y/8]; + + *x_ql = tile_x_ql; + *x_dm = tile_x_dm; + *x_sc = tile_x_sc; +} + +template static __device__ __forceinline__ void load_tiles_q5_K( + const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, + int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + GGML_UNUSED(x_qh); + + GGML_CUDA_ASSUME(i_offset >= 0); + GGML_CUDA_ASSUME(i_offset < nwarps); + GGML_CUDA_ASSUME(k >= 0); + GGML_CUDA_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI5_K; // == 0 if QK_K == 256 + const int kqsx = k % QI5_K; // == k if QK_K == 256 + + const block_q5_K * bx0 = (const block_q5_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = min(i, i_max); + } + + const block_q5_K * bxi = bx0 + i*blocks_per_row + kbx; + const int ky = QR5_K*kqsx; + + const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx); + const int ql0 = (ql >> 0) & 0x0F0F0F0F; + const int ql1 = (ql >> 4) & 0x0F0F0F0F; + + const int qh = get_int_from_uint8_aligned(bxi->qh, kqsx % (QI5_K/4)); + const int qh0 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 0)) << 4) & 0x10101010; + const int qh1 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 1)) << 4) & 0x10101010; + + const int kq0 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + 0; + const int kq1 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + (QI5_K/4); + + x_ql[i * (2*WARP_SIZE + 1) + kq0] = ql0 | qh0; + x_ql[i * (2*WARP_SIZE + 1) + kq1] = ql1 | qh1; + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI5_K; // == 1 if QK_K == 256 + const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_K) { + int i = (i0 + i_offset * QI5_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = min(i, i_max); + } + + const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd; + +#if QK_K == 256 + x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm; +#endif + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { + int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; + + if (need_check) { + i = min(i, i_max); + } + + const block_q5_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI5_K/8); + + const int * scales = (const int *) bxi->scales; + + const int ksc = k % (WARP_SIZE/8); + + // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8 + int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits + scales8 |= (scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030; // upper 2 bits + + x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8; + } +} + +static __device__ __forceinline__ float vec_dot_q5_K_q8_1_mul_mat( + const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, + const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + GGML_UNUSED(x_qh); + + const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2 * ((k % 16) / 8); + + const int index_x = i * (QR5_K*WARP_SIZE + 1) + QR5_K*k; + const int index_y = j * WARP_SIZE + (QR5_K*k) % WARP_SIZE; + return vec_dot_q5_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, sc+8, + x_dm[i * (WARP_SIZE/QI5_K) + i/QI5_K], &y_ds[index_y/QI8_1]); +} + +template static __device__ __forceinline__ void allocate_tiles_q6_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + GGML_UNUSED(x_qh); + + __shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y]; + __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI6_K) + mmq_y/QI6_K]; + __shared__ int tile_x_sc[mmq_y * (WARP_SIZE/8) + mmq_y/8]; + + *x_ql = tile_x_ql; + *x_dm = tile_x_dm; + *x_sc = tile_x_sc; +} + +template static __device__ __forceinline__ void load_tiles_q6_K( + const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, + int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + GGML_UNUSED(x_qh); + + GGML_CUDA_ASSUME(i_offset >= 0); + GGML_CUDA_ASSUME(i_offset < nwarps); + GGML_CUDA_ASSUME(k >= 0); + GGML_CUDA_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI6_K; // == 0 if QK_K == 256 + const int kqsx = k % QI6_K; // == k if QK_K == 256 + + const block_q6_K * bx0 = (const block_q6_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = min(i, i_max); + } + + const block_q6_K * bxi = bx0 + i*blocks_per_row + kbx; + const int ky = QR6_K*kqsx; + + const int ql = get_int_from_uint8(bxi->ql, kqsx); + const int ql0 = (ql >> 0) & 0x0F0F0F0F; + const int ql1 = (ql >> 4) & 0x0F0F0F0F; + + const int qh = get_int_from_uint8(bxi->qh, (QI6_K/4) * (kqsx / (QI6_K/2)) + kqsx % (QI6_K/4)); + const int qh0 = ((qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) << 4) & 0x30303030; + const int qh1 = (qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) & 0x30303030; + + const int kq0 = ky - ky % QI6_K + k % (QI6_K/2) + 0; + const int kq1 = ky - ky % QI6_K + k % (QI6_K/2) + (QI6_K/2); + + x_ql[i * (2*WARP_SIZE + 1) + kq0] = __vsubss4(ql0 | qh0, 0x20202020); + x_ql[i * (2*WARP_SIZE + 1) + kq1] = __vsubss4(ql1 | qh1, 0x20202020); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI6_K; // == 1 if QK_K == 256 + const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 + float * x_dmf = (float *) x_dm; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI6_K) { + int i = (i0 + i_offset * QI6_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = min(i, i_max); + } + + const block_q6_K * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI6_K) + i / QI6_K + kbxd] = bxi->d; + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { + int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; + + if (need_check) { + i = min(i, i_max); + } + + const block_q6_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / 4; + + x_sc[i * (WARP_SIZE/8) + i / 8 + k % (WARP_SIZE/8)] = get_int_from_int8(bxi->scales, k % (QI6_K/8)); + } +} + +static __device__ __forceinline__ float vec_dot_q6_K_q8_1_mul_mat( + const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, + const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + GGML_UNUSED(x_qh); + + const float * x_dmf = (const float *) x_dm; + const float * y_df = (const float *) y_ds; + + const int8_t * sc = ((const int8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/8]); + + const int index_x = i * (QR6_K*WARP_SIZE + 1) + QR6_K*k; + const int index_y = j * WARP_SIZE + (QR6_K*k) % WARP_SIZE; + return vec_dot_q6_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, x_dmf[i * (WARP_SIZE/QI6_K) + i/QI6_K], &y_df[index_y/QI8_1]); +} + +#define MMQ_X_Q4_0_RDNA2 64 +#define MMQ_Y_Q4_0_RDNA2 128 +#define NWARPS_Q4_0_RDNA2 8 +#define MMQ_X_Q4_0_RDNA1 64 +#define MMQ_Y_Q4_0_RDNA1 64 +#define NWARPS_Q4_0_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q4_0_AMPERE 4 +#define MMQ_Y_Q4_0_AMPERE 32 +#define NWARPS_Q4_0_AMPERE 4 +#else +#define MMQ_X_Q4_0_AMPERE 64 +#define MMQ_Y_Q4_0_AMPERE 128 +#define NWARPS_Q4_0_AMPERE 4 +#endif +#define MMQ_X_Q4_0_PASCAL 64 +#define MMQ_Y_Q4_0_PASCAL 64 +#define NWARPS_Q4_0_PASCAL 8 + +template +static __device__ __forceinline__ void mul_mat_q( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + const int blocks_per_row_x = ncols_x / qk; + const int blocks_per_col_y = nrows_y / QK8_1; + const int blocks_per_warp = WARP_SIZE / qi; + + const int & ncols_dst = ncols_y; + + const int row_dst_0 = blockIdx.x*mmq_y; + const int & row_x_0 = row_dst_0; + + const int col_dst_0 = blockIdx.y*mmq_x; + const int & col_y_0 = col_dst_0; + + int * tile_x_ql = nullptr; + half2 * tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + + allocate_tiles(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc); + + __shared__ int tile_y_qs[mmq_x * WARP_SIZE]; + __shared__ half2 tile_y_ds[mmq_x * WARP_SIZE/QI8_1]; + + float sum[mmq_y/WARP_SIZE][mmq_x/nwarps] = {{0.0f}}; + + for (int ib0 = 0; ib0 < blocks_per_row_x; ib0 += blocks_per_warp) { + + load_tiles(x + row_x_0*blocks_per_row_x + ib0, tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc, + threadIdx.y, nrows_x-row_x_0-1, threadIdx.x, blocks_per_row_x); + +#pragma unroll + for (int ir = 0; ir < qr; ++ir) { + const int kqs = ir*WARP_SIZE + threadIdx.x; + const int kbxd = kqs / QI8_1; + +#pragma unroll + for (int i = 0; i < mmq_x; i += nwarps) { + const int col_y_eff = min(col_y_0 + threadIdx.y + i, ncols_y-1); // to prevent out-of-bounds memory accesses + + const block_q8_1 * by0 = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + kbxd]; + + const int index_y = (threadIdx.y + i) * WARP_SIZE + kqs % WARP_SIZE; + tile_y_qs[index_y] = get_int_from_int8_aligned(by0->qs, threadIdx.x % QI8_1); + } + +#pragma unroll + for (int ids0 = 0; ids0 < mmq_x; ids0 += nwarps * QI8_1) { + const int ids = (ids0 + threadIdx.y * QI8_1 + threadIdx.x / (WARP_SIZE/QI8_1)) % mmq_x; + const int kby = threadIdx.x % (WARP_SIZE/QI8_1); + const int col_y_eff = min(col_y_0 + ids, ncols_y-1); + + // if the sum is not needed it's faster to transform the scale to f32 ahead of time + const half2 * dsi_src = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + ir*(WARP_SIZE/QI8_1) + kby].ds; + half2 * dsi_dst = &tile_y_ds[ids * (WARP_SIZE/QI8_1) + kby]; + if (need_sum) { + *dsi_dst = *dsi_src; + } else { + float * dfi_dst = (float *) dsi_dst; + *dfi_dst = __low2float(*dsi_src); + } + } + + __syncthreads(); + +// #pragma unroll // unrolling this loop causes too much register pressure + for (int k = ir*WARP_SIZE/qr; k < (ir+1)*WARP_SIZE/qr; k += vdr) { +#pragma unroll + for (int j = 0; j < mmq_x; j += nwarps) { +#pragma unroll + for (int i = 0; i < mmq_y; i += WARP_SIZE) { + sum[i/WARP_SIZE][j/nwarps] += vec_dot( + tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc, tile_y_qs, tile_y_ds, + threadIdx.x + i, threadIdx.y + j, k); + } + } + } + + __syncthreads(); + } + } + +#pragma unroll + for (int j = 0; j < mmq_x; j += nwarps) { + const int col_dst = col_dst_0 + j + threadIdx.y; + + if (col_dst >= ncols_dst) { + return; + } + +#pragma unroll + for (int i = 0; i < mmq_y; i += WARP_SIZE) { + const int row_dst = row_dst_0 + threadIdx.x + i; + + if (row_dst >= nrows_dst) { + continue; + } + + dst[col_dst*nrows_dst + row_dst] = sum[i/WARP_SIZE][j/nwarps]; + } + } +} + +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q4_0_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + mul_mat_q4_0( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q4_0_RDNA2; + const int mmq_y = MMQ_Y_Q4_0_RDNA2; + const int nwarps = NWARPS_Q4_0_RDNA2; +#else + const int mmq_x = MMQ_X_Q4_0_RDNA1; + const int mmq_y = MMQ_Y_Q4_0_RDNA1; + const int nwarps = NWARPS_Q4_0_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q4_0, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_VOLTA + const int mmq_x = MMQ_X_Q4_0_AMPERE; + const int mmq_y = MMQ_Y_Q4_0_AMPERE; + const int nwarps = NWARPS_Q4_0_AMPERE; + + mul_mat_q, + load_tiles_q4_0, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= MIN_CC_DP4A + const int mmq_x = MMQ_X_Q4_0_PASCAL; + const int mmq_y = MMQ_Y_Q4_0_PASCAL; + const int nwarps = NWARPS_Q4_0_PASCAL; + + mul_mat_q, + load_tiles_q4_0, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); +#else + GGML_UNUSED(vec_dot_q4_0_q8_1_mul_mat); + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= CC_VOLTA +} + +#define MMQ_X_Q4_1_RDNA2 64 +#define MMQ_Y_Q4_1_RDNA2 128 +#define NWARPS_Q4_1_RDNA2 8 +#define MMQ_X_Q4_1_RDNA1 64 +#define MMQ_Y_Q4_1_RDNA1 64 +#define NWARPS_Q4_1_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q4_1_AMPERE 4 +#define MMQ_Y_Q4_1_AMPERE 32 +#define NWARPS_Q4_1_AMPERE 4 +#else +#define MMQ_X_Q4_1_AMPERE 64 +#define MMQ_Y_Q4_1_AMPERE 128 +#define NWARPS_Q4_1_AMPERE 4 +#endif +#define MMQ_X_Q4_1_PASCAL 64 +#define MMQ_Y_Q4_1_PASCAL 64 +#define NWARPS_Q4_1_PASCAL 8 + +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q4_1_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#elif __CUDA_ARCH__ < CC_VOLTA + __launch_bounds__(WARP_SIZE*NWARPS_Q4_1_PASCAL, 2) +#endif // __CUDA_ARCH__ < CC_VOLTA + mul_mat_q4_1( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q4_1_RDNA2; + const int mmq_y = MMQ_Y_Q4_1_RDNA2; + const int nwarps = NWARPS_Q4_1_RDNA2; +#else + const int mmq_x = MMQ_X_Q4_1_RDNA1; + const int mmq_y = MMQ_Y_Q4_1_RDNA1; + const int nwarps = NWARPS_Q4_1_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q4_1, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_VOLTA + const int mmq_x = MMQ_X_Q4_1_AMPERE; + const int mmq_y = MMQ_Y_Q4_1_AMPERE; + const int nwarps = NWARPS_Q4_1_AMPERE; + + mul_mat_q, + load_tiles_q4_1, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= MIN_CC_DP4A + const int mmq_x = MMQ_X_Q4_1_PASCAL; + const int mmq_y = MMQ_Y_Q4_1_PASCAL; + const int nwarps = NWARPS_Q4_1_PASCAL; + + mul_mat_q, + load_tiles_q4_1, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); +#else + GGML_UNUSED(vec_dot_q4_1_q8_1_mul_mat); + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= CC_VOLTA +} + +#define MMQ_X_Q5_0_RDNA2 64 +#define MMQ_Y_Q5_0_RDNA2 128 +#define NWARPS_Q5_0_RDNA2 8 +#define MMQ_X_Q5_0_RDNA1 64 +#define MMQ_Y_Q5_0_RDNA1 64 +#define NWARPS_Q5_0_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q5_0_AMPERE 4 +#define MMQ_Y_Q5_0_AMPERE 32 +#define NWARPS_Q5_0_AMPERE 4 +#else +#define MMQ_X_Q5_0_AMPERE 128 +#define MMQ_Y_Q5_0_AMPERE 64 +#define NWARPS_Q5_0_AMPERE 4 +#endif +#define MMQ_X_Q5_0_PASCAL 64 +#define MMQ_Y_Q5_0_PASCAL 64 +#define NWARPS_Q5_0_PASCAL 8 + +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q5_0_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + mul_mat_q5_0( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q5_0_RDNA2; + const int mmq_y = MMQ_Y_Q5_0_RDNA2; + const int nwarps = NWARPS_Q5_0_RDNA2; +#else + const int mmq_x = MMQ_X_Q5_0_RDNA1; + const int mmq_y = MMQ_Y_Q5_0_RDNA1; + const int nwarps = NWARPS_Q5_0_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q5_0, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_VOLTA + const int mmq_x = MMQ_X_Q5_0_AMPERE; + const int mmq_y = MMQ_Y_Q5_0_AMPERE; + const int nwarps = NWARPS_Q5_0_AMPERE; + + mul_mat_q, + load_tiles_q5_0, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= MIN_CC_DP4A + const int mmq_x = MMQ_X_Q5_0_PASCAL; + const int mmq_y = MMQ_Y_Q5_0_PASCAL; + const int nwarps = NWARPS_Q5_0_PASCAL; + + mul_mat_q, + load_tiles_q5_0, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); +#else + GGML_UNUSED(vec_dot_q5_0_q8_1_mul_mat); + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= CC_VOLTA +} + +#define MMQ_X_Q5_1_RDNA2 64 +#define MMQ_Y_Q5_1_RDNA2 128 +#define NWARPS_Q5_1_RDNA2 8 +#define MMQ_X_Q5_1_RDNA1 64 +#define MMQ_Y_Q5_1_RDNA1 64 +#define NWARPS_Q5_1_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q5_1_AMPERE 4 +#define MMQ_Y_Q5_1_AMPERE 32 +#define NWARPS_Q5_1_AMPERE 4 +#else +#define MMQ_X_Q5_1_AMPERE 128 +#define MMQ_Y_Q5_1_AMPERE 64 +#define NWARPS_Q5_1_AMPERE 4 +#endif +#define MMQ_X_Q5_1_PASCAL 64 +#define MMQ_Y_Q5_1_PASCAL 64 +#define NWARPS_Q5_1_PASCAL 8 + +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q5_1_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +mul_mat_q5_1( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q5_1_RDNA2; + const int mmq_y = MMQ_Y_Q5_1_RDNA2; + const int nwarps = NWARPS_Q5_1_RDNA2; +#else + const int mmq_x = MMQ_X_Q5_1_RDNA1; + const int mmq_y = MMQ_Y_Q5_1_RDNA1; + const int nwarps = NWARPS_Q5_1_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q5_1, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_VOLTA + const int mmq_x = MMQ_X_Q5_1_AMPERE; + const int mmq_y = MMQ_Y_Q5_1_AMPERE; + const int nwarps = NWARPS_Q5_1_AMPERE; + + mul_mat_q, + load_tiles_q5_1, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= MIN_CC_DP4A + const int mmq_x = MMQ_X_Q5_1_PASCAL; + const int mmq_y = MMQ_Y_Q5_1_PASCAL; + const int nwarps = NWARPS_Q5_1_PASCAL; + + mul_mat_q, + load_tiles_q5_1, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); +#else + GGML_UNUSED(vec_dot_q5_1_q8_1_mul_mat); + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= CC_VOLTA +} + +#define MMQ_X_Q8_0_RDNA2 64 +#define MMQ_Y_Q8_0_RDNA2 128 +#define NWARPS_Q8_0_RDNA2 8 +#define MMQ_X_Q8_0_RDNA1 64 +#define MMQ_Y_Q8_0_RDNA1 64 +#define NWARPS_Q8_0_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q8_0_AMPERE 4 +#define MMQ_Y_Q8_0_AMPERE 32 +#define NWARPS_Q8_0_AMPERE 4 +#else +#define MMQ_X_Q8_0_AMPERE 128 +#define MMQ_Y_Q8_0_AMPERE 64 +#define NWARPS_Q8_0_AMPERE 4 +#endif +#define MMQ_X_Q8_0_PASCAL 64 +#define MMQ_Y_Q8_0_PASCAL 64 +#define NWARPS_Q8_0_PASCAL 8 + +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q8_0_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + mul_mat_q8_0( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q8_0_RDNA2; + const int mmq_y = MMQ_Y_Q8_0_RDNA2; + const int nwarps = NWARPS_Q8_0_RDNA2; +#else + const int mmq_x = MMQ_X_Q8_0_RDNA1; + const int mmq_y = MMQ_Y_Q8_0_RDNA1; + const int nwarps = NWARPS_Q8_0_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q8_0, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_VOLTA + const int mmq_x = MMQ_X_Q8_0_AMPERE; + const int mmq_y = MMQ_Y_Q8_0_AMPERE; + const int nwarps = NWARPS_Q8_0_AMPERE; + + mul_mat_q, + load_tiles_q8_0, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= MIN_CC_DP4A + const int mmq_x = MMQ_X_Q8_0_PASCAL; + const int mmq_y = MMQ_Y_Q8_0_PASCAL; + const int nwarps = NWARPS_Q8_0_PASCAL; + + mul_mat_q, + load_tiles_q8_0, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); +#else + GGML_UNUSED(vec_dot_q8_0_q8_1_mul_mat); + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= CC_VOLTA +} + +#define MMQ_X_Q2_K_RDNA2 64 +#define MMQ_Y_Q2_K_RDNA2 128 +#define NWARPS_Q2_K_RDNA2 8 +#define MMQ_X_Q2_K_RDNA1 128 +#define MMQ_Y_Q2_K_RDNA1 32 +#define NWARPS_Q2_K_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q2_K_AMPERE 4 +#define MMQ_Y_Q2_K_AMPERE 32 +#define NWARPS_Q2_K_AMPERE 4 +#else +#define MMQ_X_Q2_K_AMPERE 64 +#define MMQ_Y_Q2_K_AMPERE 128 +#define NWARPS_Q2_K_AMPERE 4 +#endif +#define MMQ_X_Q2_K_PASCAL 64 +#define MMQ_Y_Q2_K_PASCAL 64 +#define NWARPS_Q2_K_PASCAL 8 + +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q2_K_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +mul_mat_q2_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q2_K_RDNA2; + const int mmq_y = MMQ_Y_Q2_K_RDNA2; + const int nwarps = NWARPS_Q2_K_RDNA2; +#else + const int mmq_x = MMQ_X_Q2_K_RDNA1; + const int mmq_y = MMQ_Y_Q2_K_RDNA1; + const int nwarps = NWARPS_Q2_K_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q2_K, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_VOLTA + const int mmq_x = MMQ_X_Q2_K_AMPERE; + const int mmq_y = MMQ_Y_Q2_K_AMPERE; + const int nwarps = NWARPS_Q2_K_AMPERE; + + mul_mat_q, + load_tiles_q2_K, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= MIN_CC_DP4A + const int mmq_x = MMQ_X_Q2_K_PASCAL; + const int mmq_y = MMQ_Y_Q2_K_PASCAL; + const int nwarps = NWARPS_Q2_K_PASCAL; + + mul_mat_q, + load_tiles_q2_K, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); +#else + GGML_UNUSED(vec_dot_q2_K_q8_1_mul_mat); + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= CC_VOLTA +} + +#define MMQ_X_Q3_K_RDNA2 128 +#define MMQ_Y_Q3_K_RDNA2 64 +#define NWARPS_Q3_K_RDNA2 8 +#define MMQ_X_Q3_K_RDNA1 32 +#define MMQ_Y_Q3_K_RDNA1 128 +#define NWARPS_Q3_K_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q3_K_AMPERE 4 +#define MMQ_Y_Q3_K_AMPERE 32 +#define NWARPS_Q3_K_AMPERE 4 +#else +#define MMQ_X_Q3_K_AMPERE 128 +#define MMQ_Y_Q3_K_AMPERE 128 +#define NWARPS_Q3_K_AMPERE 4 +#endif +#define MMQ_X_Q3_K_PASCAL 64 +#define MMQ_Y_Q3_K_PASCAL 64 +#define NWARPS_Q3_K_PASCAL 8 + +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q3_K_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#elif __CUDA_ARCH__ < CC_VOLTA + __launch_bounds__(WARP_SIZE*NWARPS_Q3_K_PASCAL, 2) +#endif // __CUDA_ARCH__ < CC_VOLTA + mul_mat_q3_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q3_K_RDNA2; + const int mmq_y = MMQ_Y_Q3_K_RDNA2; + const int nwarps = NWARPS_Q3_K_RDNA2; +#else + const int mmq_x = MMQ_X_Q3_K_RDNA1; + const int mmq_y = MMQ_Y_Q3_K_RDNA1; + const int nwarps = NWARPS_Q3_K_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q3_K, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_VOLTA + const int mmq_x = MMQ_X_Q3_K_AMPERE; + const int mmq_y = MMQ_Y_Q3_K_AMPERE; + const int nwarps = NWARPS_Q3_K_AMPERE; + + mul_mat_q, + load_tiles_q3_K, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= MIN_CC_DP4A + const int mmq_x = MMQ_X_Q3_K_PASCAL; + const int mmq_y = MMQ_Y_Q3_K_PASCAL; + const int nwarps = NWARPS_Q3_K_PASCAL; + + mul_mat_q, + load_tiles_q3_K, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); +#else + GGML_UNUSED(vec_dot_q3_K_q8_1_mul_mat); + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= CC_VOLTA +} + +#define MMQ_X_Q4_K_RDNA2 64 +#define MMQ_Y_Q4_K_RDNA2 128 +#define NWARPS_Q4_K_RDNA2 8 +#define MMQ_X_Q4_K_RDNA1 32 +#define MMQ_Y_Q4_K_RDNA1 64 +#define NWARPS_Q4_K_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q4_K_AMPERE 4 +#define MMQ_Y_Q4_K_AMPERE 32 +#define NWARPS_Q4_K_AMPERE 4 +#else +#define MMQ_X_Q4_K_AMPERE 64 +#define MMQ_Y_Q4_K_AMPERE 128 +#define NWARPS_Q4_K_AMPERE 4 +#endif +#define MMQ_X_Q4_K_PASCAL 64 +#define MMQ_Y_Q4_K_PASCAL 64 +#define NWARPS_Q4_K_PASCAL 8 + +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q4_K_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#elif __CUDA_ARCH__ < CC_VOLTA + __launch_bounds__(WARP_SIZE*NWARPS_Q4_K_PASCAL, 2) +#endif // __CUDA_ARCH__ < CC_VOLTA + mul_mat_q4_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q4_K_RDNA2; + const int mmq_y = MMQ_Y_Q4_K_RDNA2; + const int nwarps = NWARPS_Q4_K_RDNA2; +#else + const int mmq_x = MMQ_X_Q4_K_RDNA1; + const int mmq_y = MMQ_Y_Q4_K_RDNA1; + const int nwarps = NWARPS_Q4_K_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q4_K, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_VOLTA + const int mmq_x = MMQ_X_Q4_K_AMPERE; + const int mmq_y = MMQ_Y_Q4_K_AMPERE; + const int nwarps = NWARPS_Q4_K_AMPERE; + + mul_mat_q, + load_tiles_q4_K, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= MIN_CC_DP4A + const int mmq_x = MMQ_X_Q4_K_PASCAL; + const int mmq_y = MMQ_Y_Q4_K_PASCAL; + const int nwarps = NWARPS_Q4_K_PASCAL; + + mul_mat_q, + load_tiles_q4_K, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); +#else + GGML_UNUSED(vec_dot_q4_K_q8_1_mul_mat); + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= CC_VOLTA +} + +#define MMQ_X_Q5_K_RDNA2 64 +#define MMQ_Y_Q5_K_RDNA2 128 +#define NWARPS_Q5_K_RDNA2 8 +#define MMQ_X_Q5_K_RDNA1 32 +#define MMQ_Y_Q5_K_RDNA1 64 +#define NWARPS_Q5_K_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q5_K_AMPERE 4 +#define MMQ_Y_Q5_K_AMPERE 32 +#define NWARPS_Q5_K_AMPERE 4 +#else +#define MMQ_X_Q5_K_AMPERE 64 +#define MMQ_Y_Q5_K_AMPERE 128 +#define NWARPS_Q5_K_AMPERE 4 +#endif +#define MMQ_X_Q5_K_PASCAL 64 +#define MMQ_Y_Q5_K_PASCAL 64 +#define NWARPS_Q5_K_PASCAL 8 + +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q5_K_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +mul_mat_q5_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q5_K_RDNA2; + const int mmq_y = MMQ_Y_Q5_K_RDNA2; + const int nwarps = NWARPS_Q5_K_RDNA2; +#else + const int mmq_x = MMQ_X_Q5_K_RDNA1; + const int mmq_y = MMQ_Y_Q5_K_RDNA1; + const int nwarps = NWARPS_Q5_K_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q5_K, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_VOLTA + const int mmq_x = MMQ_X_Q5_K_AMPERE; + const int mmq_y = MMQ_Y_Q5_K_AMPERE; + const int nwarps = NWARPS_Q5_K_AMPERE; + + mul_mat_q, + load_tiles_q5_K, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= MIN_CC_DP4A + const int mmq_x = MMQ_X_Q5_K_PASCAL; + const int mmq_y = MMQ_Y_Q5_K_PASCAL; + const int nwarps = NWARPS_Q5_K_PASCAL; + + mul_mat_q, + load_tiles_q5_K, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); +#else + GGML_UNUSED(vec_dot_q5_K_q8_1_mul_mat); + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= CC_VOLTA +} + +#define MMQ_X_Q6_K_RDNA2 64 +#define MMQ_Y_Q6_K_RDNA2 128 +#define NWARPS_Q6_K_RDNA2 8 +#define MMQ_X_Q6_K_RDNA1 32 +#define MMQ_Y_Q6_K_RDNA1 64 +#define NWARPS_Q6_K_RDNA1 8 +#if defined(CUDA_USE_TENSOR_CORES) +#define MMQ_X_Q6_K_AMPERE 4 +#define MMQ_Y_Q6_K_AMPERE 32 +#define NWARPS_Q6_K_AMPERE 4 +#else +#define MMQ_X_Q6_K_AMPERE 64 +#define MMQ_Y_Q6_K_AMPERE 64 +#define NWARPS_Q6_K_AMPERE 4 +#endif +#define MMQ_X_Q6_K_PASCAL 64 +#define MMQ_Y_Q6_K_PASCAL 64 +#define NWARPS_Q6_K_PASCAL 8 + +template static __global__ void +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + __launch_bounds__(WARP_SIZE*NWARPS_Q6_K_RDNA2, 2) +#endif // defined(RDNA3) || defined(RDNA2) +#elif __CUDA_ARCH__ < CC_VOLTA + __launch_bounds__(WARP_SIZE*NWARPS_Q6_K_PASCAL, 2) +#endif // __CUDA_ARCH__ < CC_VOLTA + mul_mat_q6_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) { + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) +#if defined(RDNA3) || defined(RDNA2) + const int mmq_x = MMQ_X_Q6_K_RDNA2; + const int mmq_y = MMQ_Y_Q6_K_RDNA2; + const int nwarps = NWARPS_Q6_K_RDNA2; +#else + const int mmq_x = MMQ_X_Q6_K_RDNA1; + const int mmq_y = MMQ_Y_Q6_K_RDNA1; + const int nwarps = NWARPS_Q6_K_RDNA1; +#endif // defined(RDNA3) || defined(RDNA2) + + mul_mat_q, + load_tiles_q6_K, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= CC_VOLTA + const int mmq_x = MMQ_X_Q6_K_AMPERE; + const int mmq_y = MMQ_Y_Q6_K_AMPERE; + const int nwarps = NWARPS_Q6_K_AMPERE; + + mul_mat_q, + load_tiles_q6_K, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + +#elif __CUDA_ARCH__ >= MIN_CC_DP4A + const int mmq_x = MMQ_X_Q6_K_PASCAL; + const int mmq_y = MMQ_Y_Q6_K_PASCAL; + const int nwarps = NWARPS_Q6_K_PASCAL; + + mul_mat_q, + load_tiles_q6_K, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); +#else + GGML_UNUSED(vec_dot_q6_K_q8_1_mul_mat); + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= CC_VOLTA +} + +static void ggml_mul_mat_q4_0_q8_1_cuda( + const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, + const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { + + int id; + CUDA_CHECK(cudaGetDevice(&id)); + const int compute_capability = ggml_cuda_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q4_0_RDNA2; + mmq_y = MMQ_Y_Q4_0_RDNA2; + nwarps = NWARPS_Q4_0_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q4_0_RDNA1; + mmq_y = MMQ_Y_Q4_0_RDNA1; + nwarps = NWARPS_Q4_0_RDNA1; + } else if (compute_capability >= CC_VOLTA) { + mmq_x = MMQ_X_Q4_0_AMPERE; + mmq_y = MMQ_Y_Q4_0_AMPERE; + nwarps = NWARPS_Q4_0_AMPERE; + } else if (compute_capability >= MIN_CC_DP4A) { + mmq_x = MMQ_X_Q4_0_PASCAL; + mmq_y = MMQ_Y_Q4_0_PASCAL; + nwarps = NWARPS_Q4_0_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const dim3 block_nums(block_num_x, block_num_y, 1); + const dim3 block_dims(WARP_SIZE, nwarps, 1); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + mul_mat_q4_0<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } else { + const bool need_check = true; + mul_mat_q4_0<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } +} + +static void ggml_mul_mat_q4_1_q8_1_cuda( + const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, + const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { + + int id; + CUDA_CHECK(cudaGetDevice(&id)); + const int compute_capability = ggml_cuda_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q4_1_RDNA2; + mmq_y = MMQ_Y_Q4_1_RDNA2; + nwarps = NWARPS_Q4_1_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q4_1_RDNA1; + mmq_y = MMQ_Y_Q4_1_RDNA1; + nwarps = NWARPS_Q4_1_RDNA1; + } else if (compute_capability >= CC_VOLTA) { + mmq_x = MMQ_X_Q4_1_AMPERE; + mmq_y = MMQ_Y_Q4_1_AMPERE; + nwarps = NWARPS_Q4_1_AMPERE; + } else if (compute_capability >= MIN_CC_DP4A) { + mmq_x = MMQ_X_Q4_1_PASCAL; + mmq_y = MMQ_Y_Q4_1_PASCAL; + nwarps = NWARPS_Q4_1_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const dim3 block_nums(block_num_x, block_num_y, 1); + const dim3 block_dims(WARP_SIZE, nwarps, 1); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + mul_mat_q4_1<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } else { + const bool need_check = true; + mul_mat_q4_1<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } +} + +static void ggml_mul_mat_q5_0_q8_1_cuda( + const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, + const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { + + int id; + CUDA_CHECK(cudaGetDevice(&id)); + const int compute_capability = ggml_cuda_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q5_0_RDNA2; + mmq_y = MMQ_Y_Q5_0_RDNA2; + nwarps = NWARPS_Q5_0_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q5_0_RDNA1; + mmq_y = MMQ_Y_Q5_0_RDNA1; + nwarps = NWARPS_Q5_0_RDNA1; + } else if (compute_capability >= CC_VOLTA) { + mmq_x = MMQ_X_Q5_0_AMPERE; + mmq_y = MMQ_Y_Q5_0_AMPERE; + nwarps = NWARPS_Q5_0_AMPERE; + } else if (compute_capability >= MIN_CC_DP4A) { + mmq_x = MMQ_X_Q5_0_PASCAL; + mmq_y = MMQ_Y_Q5_0_PASCAL; + nwarps = NWARPS_Q5_0_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const dim3 block_nums(block_num_x, block_num_y, 1); + const dim3 block_dims(WARP_SIZE, nwarps, 1); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + mul_mat_q5_0<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } else { + const bool need_check = true; + mul_mat_q5_0<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } +} + +static void ggml_mul_mat_q5_1_q8_1_cuda( + const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, + const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { + + int id; + CUDA_CHECK(cudaGetDevice(&id)); + const int compute_capability = ggml_cuda_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q5_1_RDNA2; + mmq_y = MMQ_Y_Q5_1_RDNA2; + nwarps = NWARPS_Q5_1_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q5_1_RDNA1; + mmq_y = MMQ_Y_Q5_1_RDNA1; + nwarps = NWARPS_Q5_1_RDNA1; + } else if (compute_capability >= CC_VOLTA) { + mmq_x = MMQ_X_Q5_1_AMPERE; + mmq_y = MMQ_Y_Q5_1_AMPERE; + nwarps = NWARPS_Q5_1_AMPERE; + } else if (compute_capability >= MIN_CC_DP4A) { + mmq_x = MMQ_X_Q5_1_PASCAL; + mmq_y = MMQ_Y_Q5_1_PASCAL; + nwarps = NWARPS_Q5_1_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const dim3 block_nums(block_num_x, block_num_y, 1); + const dim3 block_dims(WARP_SIZE, nwarps, 1); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + mul_mat_q5_1<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } else { + const bool need_check = true; + mul_mat_q5_1<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } +} + +static void ggml_mul_mat_q8_0_q8_1_cuda( + const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, + const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { + + int id; + CUDA_CHECK(cudaGetDevice(&id)); + const int compute_capability = ggml_cuda_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q8_0_RDNA2; + mmq_y = MMQ_Y_Q8_0_RDNA2; + nwarps = NWARPS_Q8_0_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q8_0_RDNA1; + mmq_y = MMQ_Y_Q8_0_RDNA1; + nwarps = NWARPS_Q8_0_RDNA1; + } else if (compute_capability >= CC_VOLTA) { + mmq_x = MMQ_X_Q8_0_AMPERE; + mmq_y = MMQ_Y_Q8_0_AMPERE; + nwarps = NWARPS_Q8_0_AMPERE; + } else if (compute_capability >= MIN_CC_DP4A) { + mmq_x = MMQ_X_Q8_0_PASCAL; + mmq_y = MMQ_Y_Q8_0_PASCAL; + nwarps = NWARPS_Q8_0_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const dim3 block_nums(block_num_x, block_num_y, 1); + const dim3 block_dims(WARP_SIZE, nwarps, 1); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + mul_mat_q8_0<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } else { + const bool need_check = true; + mul_mat_q8_0<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } +} + +static void ggml_mul_mat_q2_K_q8_1_cuda( + const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, + const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { + + int id; + CUDA_CHECK(cudaGetDevice(&id)); + const int compute_capability = ggml_cuda_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q2_K_RDNA2; + mmq_y = MMQ_Y_Q2_K_RDNA2; + nwarps = NWARPS_Q2_K_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q2_K_RDNA1; + mmq_y = MMQ_Y_Q2_K_RDNA1; + nwarps = NWARPS_Q2_K_RDNA1; + } else if (compute_capability >= CC_VOLTA) { + mmq_x = MMQ_X_Q2_K_AMPERE; + mmq_y = MMQ_Y_Q2_K_AMPERE; + nwarps = NWARPS_Q2_K_AMPERE; + } else if (compute_capability >= MIN_CC_DP4A) { + mmq_x = MMQ_X_Q2_K_PASCAL; + mmq_y = MMQ_Y_Q2_K_PASCAL; + nwarps = NWARPS_Q2_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const dim3 block_nums(block_num_x, block_num_y, 1); + const dim3 block_dims(WARP_SIZE, nwarps, 1); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + mul_mat_q2_K<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } else { + const bool need_check = true; + mul_mat_q2_K<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } +} + +static void ggml_mul_mat_q3_K_q8_1_cuda( + const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, + const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { + +#if QK_K == 256 + + int id; + CUDA_CHECK(cudaGetDevice(&id)); + const int compute_capability = ggml_cuda_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q3_K_RDNA2; + mmq_y = MMQ_Y_Q3_K_RDNA2; + nwarps = NWARPS_Q3_K_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q3_K_RDNA1; + mmq_y = MMQ_Y_Q3_K_RDNA1; + nwarps = NWARPS_Q3_K_RDNA1; + } else if (compute_capability >= CC_VOLTA) { + mmq_x = MMQ_X_Q3_K_AMPERE; + mmq_y = MMQ_Y_Q3_K_AMPERE; + nwarps = NWARPS_Q3_K_AMPERE; + } else if (compute_capability >= MIN_CC_DP4A) { + mmq_x = MMQ_X_Q3_K_PASCAL; + mmq_y = MMQ_Y_Q3_K_PASCAL; + nwarps = NWARPS_Q3_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const dim3 block_nums(block_num_x, block_num_y, 1); + const dim3 block_dims(WARP_SIZE, nwarps, 1); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + mul_mat_q3_K<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } else { + const bool need_check = true; + mul_mat_q3_K<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } +#endif +} + +static void ggml_mul_mat_q4_K_q8_1_cuda( + const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, + const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { + + int id; + CUDA_CHECK(cudaGetDevice(&id)); + const int compute_capability = ggml_cuda_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q4_K_RDNA2; + mmq_y = MMQ_Y_Q4_K_RDNA2; + nwarps = NWARPS_Q4_K_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q4_K_RDNA1; + mmq_y = MMQ_Y_Q4_K_RDNA1; + nwarps = NWARPS_Q4_K_RDNA1; + } else if (compute_capability >= CC_VOLTA) { + mmq_x = MMQ_X_Q4_K_AMPERE; + mmq_y = MMQ_Y_Q4_K_AMPERE; + nwarps = NWARPS_Q4_K_AMPERE; + } else if (compute_capability >= MIN_CC_DP4A) { + mmq_x = MMQ_X_Q4_K_PASCAL; + mmq_y = MMQ_Y_Q4_K_PASCAL; + nwarps = NWARPS_Q4_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const dim3 block_nums(block_num_x, block_num_y, 1); + const dim3 block_dims(WARP_SIZE, nwarps, 1); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + mul_mat_q4_K<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } else { + const bool need_check = true; + mul_mat_q4_K<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } +} + +static void ggml_mul_mat_q5_K_q8_1_cuda( + const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, + const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { + + int id; + CUDA_CHECK(cudaGetDevice(&id)); + const int compute_capability = ggml_cuda_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q5_K_RDNA2; + mmq_y = MMQ_Y_Q5_K_RDNA2; + nwarps = NWARPS_Q5_K_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q5_K_RDNA1; + mmq_y = MMQ_Y_Q5_K_RDNA1; + nwarps = NWARPS_Q5_K_RDNA1; + } else if (compute_capability >= CC_VOLTA) { + mmq_x = MMQ_X_Q5_K_AMPERE; + mmq_y = MMQ_Y_Q5_K_AMPERE; + nwarps = NWARPS_Q5_K_AMPERE; + } else if (compute_capability >= MIN_CC_DP4A) { + mmq_x = MMQ_X_Q5_K_PASCAL; + mmq_y = MMQ_Y_Q5_K_PASCAL; + nwarps = NWARPS_Q5_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const dim3 block_nums(block_num_x, block_num_y, 1); + const dim3 block_dims(WARP_SIZE, nwarps, 1); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + mul_mat_q5_K<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } else { + const bool need_check = true; + mul_mat_q5_K<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } +} + +static void ggml_mul_mat_q6_K_q8_1_cuda( + const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, + const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { + + int id; + CUDA_CHECK(cudaGetDevice(&id)); + const int compute_capability = ggml_cuda_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= CC_RDNA2) { + mmq_x = MMQ_X_Q6_K_RDNA2; + mmq_y = MMQ_Y_Q6_K_RDNA2; + nwarps = NWARPS_Q6_K_RDNA2; + } else if (compute_capability >= CC_OFFSET_AMD) { + mmq_x = MMQ_X_Q6_K_RDNA1; + mmq_y = MMQ_Y_Q6_K_RDNA1; + nwarps = NWARPS_Q6_K_RDNA1; + } else if (compute_capability >= CC_VOLTA) { + mmq_x = MMQ_X_Q6_K_AMPERE; + mmq_y = MMQ_Y_Q6_K_AMPERE; + nwarps = NWARPS_Q6_K_AMPERE; + } else if (compute_capability >= MIN_CC_DP4A) { + mmq_x = MMQ_X_Q6_K_PASCAL; + mmq_y = MMQ_Y_Q6_K_PASCAL; + nwarps = NWARPS_Q6_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const dim3 block_nums(block_num_x, block_num_y, 1); + const dim3 block_dims(WARP_SIZE, nwarps, 1); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + mul_mat_q6_K<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } else { + const bool need_check = true; + mul_mat_q6_K<<>> + (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); + } +} + +void ggml_cuda_op_mul_mat_q( + ggml_backend_cuda_context & ctx, + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, + const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, + const int64_t src1_padded_row_size, cudaStream_t stream) { + + const int64_t ne00 = src0->ne[0]; + + const int64_t ne10 = src1->ne[0]; + GGML_ASSERT(ne10 % QK8_1 == 0); + + const int64_t ne0 = dst->ne[0]; + + const int64_t row_diff = row_high - row_low; + + int id = ggml_cuda_get_device(); + + // the main device has a larger memory buffer to hold the results from all GPUs + // nrows_dst == nrows of the matrix that the kernel writes into + const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff; + + switch (src0->type) { + case GGML_TYPE_Q4_0: + ggml_mul_mat_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q4_1: + ggml_mul_mat_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q5_0: + ggml_mul_mat_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q5_1: + ggml_mul_mat_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q8_0: + ggml_mul_mat_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q2_K: + ggml_mul_mat_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q3_K: + ggml_mul_mat_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q4_K: + ggml_mul_mat_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q5_K: + ggml_mul_mat_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q6_K: + ggml_mul_mat_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + default: + GGML_ASSERT(false); + break; + } + + GGML_UNUSED(src1); + GGML_UNUSED(dst); + GGML_UNUSED(src1_ddf_i); +} + +bool ggml_cuda_supports_mmq(enum ggml_type type) { + switch (type) { + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + return true; + default: + return false; + } +} diff --git a/ggml-cuda/mmq.cuh b/ggml-cuda/mmq.cuh new file mode 100644 index 000000000..807817c4a --- /dev/null +++ b/ggml-cuda/mmq.cuh @@ -0,0 +1,9 @@ +#include "common.cuh" + +void ggml_cuda_op_mul_mat_q( + ggml_backend_cuda_context & ctx, + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, + const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, + const int64_t src1_padded_row_size, cudaStream_t stream); + +bool ggml_cuda_supports_mmq(enum ggml_type type); diff --git a/ggml-cuda/mmvq.cu b/ggml-cuda/mmvq.cu new file mode 100644 index 000000000..8b2d7a7ff --- /dev/null +++ b/ggml-cuda/mmvq.cu @@ -0,0 +1,395 @@ +#include "mmvq.cuh" +#include "vecdotq.cuh" + +typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs); + +template +#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) +// tell the compiler to use as many registers as it wants, see nwarps definition below +__launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1) +#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) +static __global__ void mul_mat_vec_q( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) { + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3)) + constexpr int nwarps = 1; + constexpr int rows_per_cuda_block = 1; +#else + constexpr int nwarps = ncols_y <= 4 ? 4 : 2; + constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2; +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3) + + const int tid = WARP_SIZE*threadIdx.y + threadIdx.x; + const int row0 = rows_per_cuda_block*blockIdx.x; + const int blocks_per_row_x = ncols_x / qk; + const int blocks_per_col_y = nrows_y / QK8_1; + constexpr int blocks_per_iter = vdr * nwarps*WARP_SIZE / qi; + +// partial sum for each thread + float tmp[ncols_y][rows_per_cuda_block] = {0.0f}; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) { + const int kby = kbx * (qk/QK8_1); // y block index that aligns with kbx + + // x block quant index when casting the quants to int + const int kqs = vdr * (tid % (qi/vdr)); + +#pragma unroll + for (int j = 0; j < ncols_y; ++j) { +#pragma unroll + for (int i = 0; i < rows_per_cuda_block; ++i) { + tmp[j][i] += vec_dot_q_cuda( + &x[kbx + (row0 + i)*blocks_per_row_x], &y[j*blocks_per_col_y + kby], kqs); + } + } + } + + __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][WARP_SIZE]; + if (threadIdx.y > 0) { +#pragma unroll + for (int j = 0; j < ncols_y; ++j) { +#pragma unroll + for (int i = 0; i < rows_per_cuda_block; ++i) { + tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i]; + } + } + } + __syncthreads(); + if (threadIdx.y > 0) { + return; + } + + // sum up partial sums and write back result +#pragma unroll + for (int j = 0; j < ncols_y; ++j) { +#pragma unroll + for (int i = 0; i < rows_per_cuda_block; ++i) { +#pragma unroll + for (int l = 0; l < nwarps-1; ++l) { + tmp[j][i] += tmp_shared[l][j][i][threadIdx.x]; + } + tmp[j][i] = warp_reduce_sum(tmp[j][i]); + } + + if (threadIdx.x < rows_per_cuda_block) { + dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x]; + } + } +} + +template +static void mul_mat_vec_q_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + GGML_ASSERT(ncols_x % qk == 0); + GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE); + + int id; + CUDA_CHECK(cudaGetDevice(&id)); + + int64_t nwarps = 1; + int64_t rows_per_cuda_block = 1; + + if (ggml_cuda_info().devices[id].cc < CC_RDNA2) { // NVIDIA and AMD older than RDNA2 + switch(ncols_y) { + case 1: + nwarps = 4; + rows_per_cuda_block = 1; + break; + case 2: + case 3: + case 4: + nwarps = 4; + rows_per_cuda_block = 2; + break; + case 5: + case 6: + case 7: + case 8: + nwarps = 2; + rows_per_cuda_block = 2; + break; + default: + GGML_ASSERT(false); + break; + } + } + const int64_t nblocks = (nrows_x + rows_per_cuda_block - 1) / rows_per_cuda_block; + const dim3 block_nums(nblocks, 1, 1); + const dim3 block_dims(WARP_SIZE, nwarps, 1); + + switch (ncols_y) { + case 1: + mul_mat_vec_q<1, qk, qi, block_q_t, vdr, vec_dot> + <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + break; + case 2: + mul_mat_vec_q<2, qk, qi, block_q_t, vdr, vec_dot> + <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + break; + case 3: + mul_mat_vec_q<3, qk, qi, block_q_t, vdr, vec_dot> + <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + break; + case 4: + mul_mat_vec_q<4, qk, qi, block_q_t, vdr, vec_dot> + <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + break; + case 5: + mul_mat_vec_q<5, qk, qi, block_q_t, vdr, vec_dot> + <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + break; + case 6: + mul_mat_vec_q<6, qk, qi, block_q_t, vdr, vec_dot> + <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + break; + case 7: + mul_mat_vec_q<7, qk, qi, block_q_t, vdr, vec_dot> + <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + break; + case 8: + mul_mat_vec_q<8, qk, qi, block_q_t, vdr, vec_dot> + <<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + break; + default: + GGML_ASSERT(false); + break; + } +} + +static void mul_mat_vec_q4_0_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_q4_1_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_q5_0_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_q5_1_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_q8_0_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_q2_K_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_q3_K_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_q4_K_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_q5_K_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_q6_K_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_iq2_xxs_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_iq2_xs_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_iq2_s_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_iq3_xxs_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_iq1_s_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_iq4_nl_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_iq4_xs_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +static void mul_mat_vec_iq3_s_q8_1_cuda( + const void * vx, const void * vy, float * dst, + const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) { + + mul_mat_vec_q_cuda + (vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream); +} + +void ggml_cuda_op_mul_mat_vec_q( + ggml_backend_cuda_context & ctx, + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, + const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, + const int64_t src1_padded_row_size, cudaStream_t stream) { + + const int64_t ne00 = src0->ne[0]; + const int64_t row_diff = row_high - row_low; + + const int64_t ne10 = src1->ne[0]; + GGML_ASSERT(ne10 % QK8_1 == 0); + + const int64_t ne0 = dst->ne[0]; + + int id; + CUDA_CHECK(cudaGetDevice(&id)); + + // the main device has a larger memory buffer to hold the results from all GPUs + // nrows_dst == nrows of the matrix that the kernel writes into + const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff; + + switch (src0->type) { + case GGML_TYPE_Q4_0: + mul_mat_vec_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_Q4_1: + mul_mat_vec_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_Q5_0: + mul_mat_vec_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_Q5_1: + mul_mat_vec_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_Q8_0: + mul_mat_vec_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_Q2_K: + mul_mat_vec_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_Q3_K: + mul_mat_vec_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_Q4_K: + mul_mat_vec_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_Q5_K: + mul_mat_vec_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_Q6_K: + mul_mat_vec_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_IQ2_XXS: + mul_mat_vec_iq2_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_IQ2_XS: + mul_mat_vec_iq2_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_IQ2_S: + mul_mat_vec_iq2_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_IQ3_XXS: + mul_mat_vec_iq3_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_IQ1_S: + mul_mat_vec_iq1_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_IQ4_NL: + mul_mat_vec_iq4_nl_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_IQ4_XS: + mul_mat_vec_iq4_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + case GGML_TYPE_IQ3_S: + mul_mat_vec_iq3_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream); + break; + default: + GGML_ASSERT(false); + break; + } + + GGML_UNUSED(src1); + GGML_UNUSED(dst); + GGML_UNUSED(src1_ddf_i); + GGML_UNUSED(src1_ncols); + GGML_UNUSED(src1_padded_row_size); +} diff --git a/ggml-cuda/mmvq.cuh b/ggml-cuda/mmvq.cuh new file mode 100644 index 000000000..88c42c4b7 --- /dev/null +++ b/ggml-cuda/mmvq.cuh @@ -0,0 +1,7 @@ +#include "common.cuh" + +void ggml_cuda_op_mul_mat_vec_q( + ggml_backend_cuda_context & ctx, + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, + const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, + const int64_t src1_padded_row_size, cudaStream_t stream); diff --git a/ggml-cuda/norm.cu b/ggml-cuda/norm.cu new file mode 100644 index 000000000..86f774534 --- /dev/null +++ b/ggml-cuda/norm.cu @@ -0,0 +1,215 @@ +#include "norm.cuh" + +template +static __global__ void norm_f32(const float * x, float * dst, const int ncols, const float eps) { + const int row = blockIdx.x*blockDim.y + threadIdx.y; + const int tid = threadIdx.x; + + float2 mean_var = make_float2(0.f, 0.f); + + for (int col = tid; col < ncols; col += block_size) { + const float xi = x[row*ncols + col]; + mean_var.x += xi; + mean_var.y += xi * xi; + } + + // sum up partial sums + mean_var = warp_reduce_sum(mean_var); + if (block_size > WARP_SIZE) { + __shared__ float2 s_sum[32]; + int warp_id = threadIdx.x / WARP_SIZE; + int lane_id = threadIdx.x % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = mean_var; + } + __syncthreads(); + mean_var = s_sum[lane_id]; + mean_var = warp_reduce_sum(mean_var); + } + + const float mean = mean_var.x / ncols; + const float var = mean_var.y / ncols - mean * mean; + const float inv_std = rsqrtf(var + eps); + + for (int col = tid; col < ncols; col += block_size) { + dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_std; + } +} + +template +static __global__ void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps) { + // blockIdx.x: num_groups idx + // threadIdx.x: block_size idx + int start = blockIdx.x * group_size; + int end = start + group_size; + + start += threadIdx.x; + + if (end >= ne_elements) { + end = ne_elements; + } + + float tmp = 0.0f; // partial sum for thread in warp + + for (int j = start; j < end; j += block_size) { + tmp += x[j]; + } + + tmp = warp_reduce_sum(tmp); + if (block_size > WARP_SIZE) { + __shared__ float s_sum[32]; + int warp_id = threadIdx.x / WARP_SIZE; + int lane_id = threadIdx.x % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = tmp; + } + __syncthreads(); + tmp = s_sum[lane_id]; + tmp = warp_reduce_sum(tmp); + } + + float mean = tmp / group_size; + tmp = 0.0f; + + for (int j = start; j < end; j += block_size) { + float xi = x[j] - mean; + dst[j] = xi; + tmp += xi * xi; + } + + tmp = warp_reduce_sum(tmp); + if (block_size > WARP_SIZE) { + __shared__ float s_sum[32]; + int warp_id = threadIdx.x / WARP_SIZE; + int lane_id = threadIdx.x % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = tmp; + } + __syncthreads(); + tmp = s_sum[lane_id]; + tmp = warp_reduce_sum(tmp); + } + + float variance = tmp / group_size; + float scale = rsqrtf(variance + eps); + for (int j = start; j < end; j += block_size) { + dst[j] *= scale; + } +} + +template +static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps) { + const int row = blockIdx.x*blockDim.y + threadIdx.y; + const int tid = threadIdx.x; + + float tmp = 0.0f; // partial sum for thread in warp + + for (int col = tid; col < ncols; col += block_size) { + const float xi = x[row*ncols + col]; + tmp += xi * xi; + } + + // sum up partial sums + tmp = warp_reduce_sum(tmp); + if (block_size > WARP_SIZE) { + __shared__ float s_sum[32]; + int warp_id = threadIdx.x / WARP_SIZE; + int lane_id = threadIdx.x % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = tmp; + } + __syncthreads(); + tmp = s_sum[lane_id]; + tmp = warp_reduce_sum(tmp); + } + + const float mean = tmp / ncols; + const float scale = rsqrtf(mean + eps); + + for (int col = tid; col < ncols; col += block_size) { + dst[row*ncols + col] = scale * x[row*ncols + col]; + } +} + +static void norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) { + GGML_ASSERT(ncols % WARP_SIZE == 0); + if (ncols < 1024) { + const dim3 block_dims(WARP_SIZE, 1, 1); + norm_f32<<>>(x, dst, ncols, eps); + } else { + const dim3 block_dims(1024, 1, 1); + norm_f32<1024><<>>(x, dst, ncols, eps); + } +} + +static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const int group_size, const int ne_elements, cudaStream_t stream) { + static const float eps = 1e-6f; + if (group_size < 1024) { + const dim3 block_dims(WARP_SIZE, 1, 1); + group_norm_f32<<>>(x, dst, group_size, ne_elements, eps); + } else { + const dim3 block_dims(1024, 1, 1); + group_norm_f32<1024><<>>(x, dst, group_size, ne_elements, eps); + } +} + +static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) { + GGML_ASSERT(ncols % WARP_SIZE == 0); + if (ncols < 1024) { + const dim3 block_dims(WARP_SIZE, 1, 1); + rms_norm_f32<<>>(x, dst, ncols, eps); + } else { + const dim3 block_dims(1024, 1, 1); + rms_norm_f32<1024><<>>(x, dst, ncols, eps); + } +} + +void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + norm_f32_cuda(src0_d, dst_d, ne00, nrows, eps, stream); +} + +void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + int num_groups = dst->op_params[0]; + int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups); + group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], group_size, ggml_nelements(src0), stream); +} + +void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + rms_norm_f32_cuda(src0_d, dst_d, ne00, nrows, eps, stream); +} diff --git a/ggml-cuda/norm.cuh b/ggml-cuda/norm.cuh new file mode 100644 index 000000000..431a8f74d --- /dev/null +++ b/ggml-cuda/norm.cuh @@ -0,0 +1,7 @@ +#include "common.cuh" + +void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/pad.cu b/ggml-cuda/pad.cu new file mode 100644 index 000000000..aba539e8d --- /dev/null +++ b/ggml-cuda/pad.cu @@ -0,0 +1,49 @@ +#include "pad.cuh" + +static __global__ void pad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02, const int ne03) { + // blockIdx.z: idx of ne2*ne3, aka ne02*ne03 + // blockIdx.y: idx of ne1 + // blockIDx.x: idx of ne0 / BLOCK_SIZE + int nidx = threadIdx.x + blockIdx.x * blockDim.x; + if (nidx >= ne0) { + return; + } + + // operation + int offset_dst = + nidx + + blockIdx.y * ne0 + + blockIdx.z * ne0 * gridDim.y; + if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) { + int offset_src = + nidx + + blockIdx.y * ne00 + + blockIdx.z * ne00 * ne01; + dst[offset_dst] = x[offset_src]; + } else { + dst[offset_dst] = 0.0f; + } +} + +static void pad_f32_cuda(const float * x, float * dst, + const int ne00, const int ne01, const int ne02, const int ne03, + const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) { + int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE; + dim3 gridDim(num_blocks, ne1, ne2*ne3); + pad_f32<<>>(x, dst, ne0, ne00, ne01, ne02, ne03); +} + +void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors + + pad_f32_cuda(src0_d, dst_d, + src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], + dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream); +} diff --git a/ggml-cuda/pad.cuh b/ggml-cuda/pad.cuh new file mode 100644 index 000000000..8fd386b00 --- /dev/null +++ b/ggml-cuda/pad.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_PAD_BLOCK_SIZE 256 + +void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/pool2d.cu b/ggml-cuda/pool2d.cu new file mode 100644 index 000000000..c6d51e4d6 --- /dev/null +++ b/ggml-cuda/pool2d.cu @@ -0,0 +1,94 @@ +#include "pool2d.cuh" + +template +static __global__ void pool2d_nchw_kernel( + const int ih, const int iw, const int oh, const int ow, + const int kh, const int kw, const int sh, const int sw, + const int ph, const int pw, const int parallel_elements, + const Ti* src, To* dst, const enum ggml_op_pool op) { + int idx = threadIdx.x + blockIdx.x * blockDim.x; + if (idx >= parallel_elements) { + return; + } + + const int I_HW = ih * iw; + const int O_HW = oh * ow; + const int nc = idx / O_HW; + const int cur_oh = idx % O_HW / ow; + const int cur_ow = idx % O_HW % ow; + const Ti* i_ptr = src + nc * I_HW; + To* o_ptr = dst + nc * O_HW; + const int start_h = cur_oh * sh - ph; + const int bh = max(0, start_h); + const int eh = min(ih, start_h + kh); + const int start_w = cur_ow * sw - pw; + const int bw = max(0, start_w); + const int ew = min(iw, start_w + kw); + const To scale = 1. / (kh * kw); + To res = 0; + + switch (op) { + case GGML_OP_POOL_AVG: res = 0; break; + case GGML_OP_POOL_MAX: res = -FLT_MAX; break; + default: assert(false); + } + + for (int i = bh; i < eh; i += 1) { + for (int j = bw; j < ew; j += 1) { +#if __CUDA_ARCH__ >= 350 + Ti cur = __ldg(i_ptr + i * iw + j); +#else + Ti cur = i_ptr[i * iw + j]; +#endif + switch (op) { + case GGML_OP_POOL_AVG: res += cur * scale; break; + case GGML_OP_POOL_MAX: res = max(res, (To)cur); break; + default: assert(false); + } + } + } + o_ptr[cur_oh * ow + cur_ow] = res; +} + +static void pool2d_nchw_kernel_f32_f32_cuda( + const int ih, const int iw, const int oh, const int ow, + const int kh, const int kw, const int sh, const int sw, + const int ph, const int pw, const int parallel_elements, + const float * src, float * dst, const enum ggml_op_pool op, + cudaStream_t stream) { + + const int num_blocks = (parallel_elements + CUDA_POOL2D_BLOCK_SIZE - 1) / CUDA_POOL2D_BLOCK_SIZE; + dim3 block_nums(num_blocks); + pool2d_nchw_kernel<<>>(ih, iw, oh, ow, kh, kw, sh, sw, ph, pw, parallel_elements, src, dst, op); +} + +void ggml_cuda_op_pool2d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + const int32_t * opts = (const int32_t *)dst->op_params; + enum ggml_op_pool op = static_cast(opts[0]); + const int k0 = opts[1]; + const int k1 = opts[2]; + const int s0 = opts[3]; + const int s1 = opts[4]; + const int p0 = opts[5]; + const int p1 = opts[6]; + + const int64_t IH = src0->ne[1]; + const int64_t IW = src0->ne[0]; + + const int64_t N = dst->ne[3]; + const int64_t OC = dst->ne[2]; + const int64_t OH = dst->ne[1]; + const int64_t OW = dst->ne[0]; + + const int parallel_elements = N * OC * OH * OW; + + pool2d_nchw_kernel_f32_f32_cuda(IH, IW, OH, OW, k1, k0, s1, s0, p1, p0, parallel_elements, src0_d, dst_d, op, stream); +} diff --git a/ggml-cuda/pool2d.cuh b/ggml-cuda/pool2d.cuh new file mode 100644 index 000000000..7841292bc --- /dev/null +++ b/ggml-cuda/pool2d.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_POOL2D_BLOCK_SIZE 256 + +void ggml_cuda_op_pool2d(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/quantize.cu b/ggml-cuda/quantize.cu new file mode 100644 index 000000000..a1fbc9932 --- /dev/null +++ b/ggml-cuda/quantize.cu @@ -0,0 +1,45 @@ +#include "quantize.cuh" + +static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded) { + const int ix = blockDim.x*blockIdx.x + threadIdx.x; + + if (ix >= kx_padded) { + return; + } + + const int iy = blockDim.y*blockIdx.y + threadIdx.y; + + const int i_padded = iy*kx_padded + ix; + + block_q8_1 * y = (block_q8_1 *) vy; + + const int ib = i_padded / QK8_1; // block index + const int iqs = i_padded % QK8_1; // quant index + + const float xi = ix < kx ? x[iy*kx + ix] : 0.0f; + float amax = fabsf(xi); + float sum = xi; + + amax = warp_reduce_max(amax); + sum = warp_reduce_sum(sum); + + const float d = amax / 127; + const int8_t q = amax == 0.0f ? 0 : roundf(xi / d); + + y[ib].qs[iqs] = q; + + if (iqs > 0) { + return; + } + + reinterpret_cast(y[ib].ds.x) = d; + reinterpret_cast(y[ib].ds.y) = sum; +} + +void quantize_row_q8_1_cuda(const float * x, void * vy, const int kx, const int ky, const int kx_padded, cudaStream_t stream) { + const int block_num_x = (kx_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE; + const dim3 num_blocks(block_num_x, ky, 1); + const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1); + quantize_q8_1<<>>(x, vy, kx, kx_padded); +} + diff --git a/ggml-cuda/quantize.cuh b/ggml-cuda/quantize.cuh new file mode 100644 index 000000000..adb89c83a --- /dev/null +++ b/ggml-cuda/quantize.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_QUANTIZE_BLOCK_SIZE 256 + +void quantize_row_q8_1_cuda(const float * x, void * vy, const int kx, const int ky, const int kx_padded, cudaStream_t stream); diff --git a/ggml-cuda/rope.cu b/ggml-cuda/rope.cu new file mode 100644 index 000000000..4b0d2e5ad --- /dev/null +++ b/ggml-cuda/rope.cu @@ -0,0 +1,308 @@ +#include "rope.cuh" + +struct rope_corr_dims { + float v[4]; +}; + +static __device__ float rope_yarn_ramp(const float low, const float high, const int i0) { + const float y = (i0 / 2 - low) / max(0.001f, high - low); + return 1.0f - min(1.0f, max(0.0f, y)); +} + +// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn +// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. +static __device__ void rope_yarn( + float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale, + float * cos_theta, float * sin_theta +) { + // Get n-d rotational scaling corrected for extrapolation + float theta_interp = freq_scale * theta_extrap; + float theta = theta_interp; + if (ext_factor != 0.0f) { + float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor; + theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; + + // Get n-d magnitude scaling corrected for interpolation + mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale); + } + *cos_theta = cosf(theta) * mscale; + *sin_theta = sinf(theta) * mscale; +} + +// rope == RoPE == rotary positional embedding +template +static __global__ void rope( + const T * x, T * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base, + float ext_factor, float attn_factor, rope_corr_dims corr_dims +) { + const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y); + + if (col >= ncols) { + return; + } + + const int row = blockDim.x*blockIdx.x + threadIdx.x; + const int i = row*ncols + col; + const int i2 = row/p_delta_rows; + + const int p = has_pos ? pos[i2] : 0; + const float theta_base = p*powf(freq_base, -float(col)/ncols); + + float cos_theta, sin_theta; + rope_yarn(theta_base, freq_scale, corr_dims, col, ext_factor, attn_factor, &cos_theta, &sin_theta); + + const float x0 = x[i + 0]; + const float x1 = x[i + 1]; + + dst[i + 0] = x0*cos_theta - x1*sin_theta; + dst[i + 1] = x0*sin_theta + x1*cos_theta; +} + +template +static __global__ void rope_neox( + const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, + float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims +) { + const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y); + + if (col >= ncols) { + return; + } + + const int row = blockDim.x*blockIdx.x + threadIdx.x; + const int ib = col / n_dims; + const int ic = col % n_dims; + + if (ib > 0) { + const int i = row*ncols + ib*n_dims + ic; + + dst[i + 0] = x[i + 0]; + dst[i + 1] = x[i + 1]; + + return; + } + + const int i = row*ncols + ib*n_dims + ic/2; + const int i2 = row/p_delta_rows; + + float cur_rot = inv_ndims * ic - ib; + + const int p = has_pos ? pos[i2] : 0; + const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f); + + float cos_theta, sin_theta; + rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta); + + const float x0 = x[i + 0]; + const float x1 = x[i + n_dims/2]; + + dst[i + 0] = x0*cos_theta - x1*sin_theta; + dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta; +} + +static __global__ void rope_glm_f32( + const float * x, float * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base, + int n_ctx +) { + const int col = blockDim.x*blockIdx.x + threadIdx.x; + const int half_n_dims = ncols/4; + + if (col >= half_n_dims) { + return; + } + + const int row = blockDim.y*blockIdx.y + threadIdx.y; + const int i = row*ncols + col; + const int i2 = row/p_delta_rows; + + const float col_theta_scale = powf(freq_base, -2.0f*col/ncols); + // FIXME: this is likely wrong + const int p = pos != nullptr ? pos[i2] : 0; + + const float theta = min(p, n_ctx - 2)*freq_scale*col_theta_scale; + const float sin_theta = sinf(theta); + const float cos_theta = cosf(theta); + + const float x0 = x[i + 0]; + const float x1 = x[i + half_n_dims]; + + dst[i + 0] = x0*cos_theta - x1*sin_theta; + dst[i + half_n_dims] = x0*sin_theta + x1*cos_theta; + + const float block_theta = ((float)max(p - n_ctx - 2, 0))*col_theta_scale; + const float sin_block_theta = sinf(block_theta); + const float cos_block_theta = cosf(block_theta); + + const float x2 = x[i + half_n_dims * 2]; + const float x3 = x[i + half_n_dims * 3]; + + dst[i + half_n_dims * 2] = x2*cos_block_theta - x3*sin_block_theta; + dst[i + half_n_dims * 3] = x2*sin_block_theta + x3*cos_block_theta; +} + + +template +static void rope_cuda( + const T * x, T * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, + float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream +) { + GGML_ASSERT(ncols % 2 == 0); + const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1); + const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE); + const dim3 block_nums(nrows, num_blocks_x, 1); + if (pos == nullptr) { + rope<<>>( + x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims + ); + } else { + rope<<>>( + x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims + ); + } +} + +template +static void rope_neox_cuda( + const T * x, T * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, + float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream +) { + GGML_ASSERT(ncols % 2 == 0); + const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1); + const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE); + const dim3 block_nums(nrows, num_blocks_x, 1); + + const float theta_scale = powf(freq_base, -2.0f/n_dims); + const float inv_ndims = -1.0f / n_dims; + + if (pos == nullptr) { + rope_neox<<>>( + x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims, + theta_scale, inv_ndims + ); + } else { + rope_neox<<>>( + x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims, + theta_scale, inv_ndims + ); + } +} + +static void rope_glm_f32_cuda( + const float * x, float * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, + float freq_base, int n_ctx, cudaStream_t stream +) { + GGML_ASSERT(ncols % 4 == 0); + const dim3 block_dims(CUDA_ROPE_BLOCK_SIZE/4, 1, 1); + const int num_blocks_x = (ncols + CUDA_ROPE_BLOCK_SIZE - 1) / CUDA_ROPE_BLOCK_SIZE; + const dim3 block_nums(num_blocks_x, nrows, 1); + rope_glm_f32<<>>(x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, n_ctx); +} + +static void rope_cuda_f16( + const half * x, half * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, + float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream) { + + rope_cuda(x, dst, ncols, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream); +} + +static void rope_cuda_f32( + const float * x, float * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, + float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream) { + + rope_cuda(x, dst, ncols, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream); +} + +static void rope_neox_cuda_f16( + const half * x, half * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, + float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream) { + + rope_neox_cuda(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream); +} + +static void rope_neox_cuda_f32( + const float * x, float * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, + float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream +) { + + rope_neox_cuda(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream); +} + +void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + const float * src0_d = (const float *)src0->data; + const float * src1_d = (const float *)src1->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); + GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); + GGML_ASSERT(src0->type == dst->type); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t ne2 = dst->ne[2]; + const int64_t nrows = ggml_nrows(src0); + + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + const int n_ctx = ((int32_t *) dst->op_params)[3]; + const int n_orig_ctx = ((int32_t *) dst->op_params)[4]; + + // RoPE alteration for extended context + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; + memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); + + const int32_t * pos = nullptr; + if ((mode & 1) == 0) { + GGML_ASSERT(src1->type == GGML_TYPE_I32); + GGML_ASSERT(src1->ne[0] == ne2); + pos = (const int32_t *) src1_d; + } + + const bool is_neox = mode & 2; + const bool is_glm = mode & 4; + + rope_corr_dims corr_dims; + ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v); + + // compute + if (is_glm) { + GGML_ASSERT(false); + rope_glm_f32_cuda(src0_d, dst_d, ne00, nrows, pos, freq_scale, ne01, freq_base, n_ctx, stream); + } else if (is_neox) { + if (src0->type == GGML_TYPE_F32) { + rope_neox_cuda_f32( + (const float *)src0_d, (float *)dst_d, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, stream + ); + } else if (src0->type == GGML_TYPE_F16) { + rope_neox_cuda_f16( + (const half *)src0_d, (half *)dst_d, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, stream + ); + } else { + GGML_ASSERT(false); + } + } else { + if (src0->type == GGML_TYPE_F32) { + rope_cuda_f32( + (const float *)src0_d, (float *)dst_d, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, stream + ); + } else if (src0->type == GGML_TYPE_F16) { + rope_cuda_f16( + (const half *)src0_d, (half *)dst_d, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, stream + ); + } else { + GGML_ASSERT(false); + } + } +} diff --git a/ggml-cuda/rope.cuh b/ggml-cuda/rope.cuh new file mode 100644 index 000000000..0f787a0b2 --- /dev/null +++ b/ggml-cuda/rope.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_ROPE_BLOCK_SIZE 256 + +void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/scale.cu b/ggml-cuda/scale.cu new file mode 100644 index 000000000..6e3617d1c --- /dev/null +++ b/ggml-cuda/scale.cu @@ -0,0 +1,32 @@ +#include "scale.cuh" + +static __global__ void scale_f32(const float * x, float * dst, const float scale, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + + dst[i] = scale * x[i]; +} + +static void scale_f32_cuda(const float * x, float * dst, const float scale, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE; + scale_f32<<>>(x, dst, scale, k); +} + +void ggml_cuda_op_scale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + float scale; + memcpy(&scale, dst->op_params, sizeof(float)); + + scale_f32_cuda(src0_d, dst_d, scale, ggml_nelements(src0), stream); + CUDA_CHECK(cudaGetLastError()); +} diff --git a/ggml-cuda/scale.cuh b/ggml-cuda/scale.cuh new file mode 100644 index 000000000..8ff75c829 --- /dev/null +++ b/ggml-cuda/scale.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_SCALE_BLOCK_SIZE 256 + +void ggml_cuda_op_scale(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/softmax.cu b/ggml-cuda/softmax.cu new file mode 100644 index 000000000..9bda18e58 --- /dev/null +++ b/ggml-cuda/softmax.cu @@ -0,0 +1,201 @@ +#include "softmax.cuh" + +template +static __global__ void soft_max_f32(const float * x, const float * mask, const float * pos, float * dst, const int ncols_par, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) { + const int ncols = ncols_template == 0 ? ncols_par : ncols_template; + + const int tid = threadIdx.x; + const int rowx = blockIdx.x; + const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension + + const int block_size = block_size_template == 0 ? blockDim.x : block_size_template; + + const int warp_id = threadIdx.x / WARP_SIZE; + const int lane_id = threadIdx.x % WARP_SIZE; + + float slope = 0.0f; + + // ALiBi + if (max_bias > 0.0f) { + const int h = rowx/nrows_y; // head index + + const float base = h < n_head_log2 ? m0 : m1; + const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + + slope = powf(base, exp); + } + + extern __shared__ float data_soft_max_f32[]; + float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication + // shared memory buffer to cache values between iterations: + float * vals = vals_smem ? buf_iw + WARP_SIZE : dst + rowx*ncols; + + float max_val = -INFINITY; + +#pragma unroll + for (int col0 = 0; col0 < ncols; col0 += block_size) { + const int col = col0 + tid; + + if (ncols_template == 0 && col >= ncols) { + break; + } + + const int ix = rowx*ncols + col; + const int iy = rowy*ncols + col; + + const float val = x[ix]*scale + (mask ? mask[iy] : 0.0f) + (pos ? slope*pos[col] : 0.0f); + + vals[col] = val; + max_val = max(max_val, val); + } + + // find the max value in the block + max_val = warp_reduce_max(max_val); + if (block_size > WARP_SIZE) { + if (warp_id == 0) { + buf_iw[lane_id] = -INFINITY; + } + __syncthreads(); + + if (lane_id == 0) { + buf_iw[warp_id] = max_val; + } + __syncthreads(); + + max_val = buf_iw[lane_id]; + max_val = warp_reduce_max(max_val); + } + + float tmp = 0.0f; // partial sum + +#pragma unroll + for (int col0 = 0; col0 < ncols; col0 += block_size) { + const int col = col0 + tid; + + if (ncols_template == 0 && col >= ncols) { + break; + } + + const float val = expf(vals[col] - max_val); + tmp += val; + vals[col] = val; + } + + // find the sum of exps in the block + tmp = warp_reduce_sum(tmp); + if (block_size > WARP_SIZE) { + __syncthreads(); + if (warp_id == 0) { + buf_iw[lane_id] = 0.0f; + } + __syncthreads(); + + if (lane_id == 0) { + buf_iw[warp_id] = tmp; + } + __syncthreads(); + + tmp = buf_iw[lane_id]; + tmp = warp_reduce_sum(tmp); + } + + const float inv_sum = 1.0f / tmp; + +#pragma unroll + for (int col0 = 0; col0 < ncols; col0 += block_size) { + const int col = col0 + tid; + + if (ncols_template == 0 && col >= ncols) { + return; + } + + const int idst = rowx*ncols + col; + dst[idst] = vals[col] * inv_sum; + } +} + +static void soft_max_f32_cuda(const float * x, const float * mask, const float * pos, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) { + int nth = WARP_SIZE; + while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2; + const dim3 block_dims(nth, 1, 1); + const dim3 block_nums(nrows_x, 1, 1); + const size_t shmem = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float); + static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted."); + + const uint32_t n_head_kv = nrows_x/nrows_y; + const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv)); + + const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); + const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); + + if (shmem < ggml_cuda_info().devices[ggml_cuda_get_device()].smpb) { + switch (ncols_x) { + case 32: + soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); + break; + case 64: + soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); + break; + case 128: + soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); + break; + case 256: + soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); + break; + case 512: + soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); + break; + case 1024: + soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); + break; + case 2048: + soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); + break; + case 4096: + soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); + break; + default: + soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); + break; + } + } else { + const size_t shmem_low = WARP_SIZE*sizeof(float); + soft_max_f32<<>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2); + } +} + +void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + const float * src0_d = (const float *)src0->data; + const float * src1_d = src1 ? (const float *)src1->data : nullptr; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional + + const int64_t ne00 = src0->ne[0]; + const int64_t nrows_x = ggml_nrows(src0); + const int64_t nrows_y = src0->ne[1]; + + float scale = 1.0f; + float max_bias = 0.0f; + + memcpy(&scale, (float *) dst->op_params + 0, sizeof(float)); + memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float)); + + // positions tensor + float * src2_dd = nullptr; + + ggml_tensor * src2 = dst->src[2]; + const bool use_src2 = src2 != nullptr; + + if (use_src2) { + src2_dd = (float *)src2->data; + } + + soft_max_f32_cuda(src0_d, src1_d, src2_dd, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream); +} diff --git a/ggml-cuda/softmax.cuh b/ggml-cuda/softmax.cuh new file mode 100644 index 000000000..4ef4ff86c --- /dev/null +++ b/ggml-cuda/softmax.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_SOFT_MAX_BLOCK_SIZE 1024 + +void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/sumrows.cu b/ggml-cuda/sumrows.cu new file mode 100644 index 000000000..82e8e875f --- /dev/null +++ b/ggml-cuda/sumrows.cu @@ -0,0 +1,40 @@ +#include "sumrows.cuh" + +static __global__ void k_sum_rows_f32(const float * x, float * dst, const int ncols) { + const int row = blockIdx.x; + const int col = threadIdx.x; + + float sum = 0.0f; + for (int i = col; i < ncols; i += blockDim.x) { + sum += x[row * ncols + i]; + } + + sum = warp_reduce_sum(sum); + + if (col == 0) { + dst[row] = sum; + } +} + +static void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + const dim3 block_dims(WARP_SIZE, 1, 1); + const dim3 block_nums(nrows, 1, 1); + k_sum_rows_f32<<>>(x, dst, ncols); +} + +void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + GGML_ASSERT(ggml_is_contiguous(src0)); + + + const int64_t ncols = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + sum_rows_f32_cuda(src0_d, dst_d, ncols, nrows, stream); +} diff --git a/ggml-cuda/sumrows.cuh b/ggml-cuda/sumrows.cuh new file mode 100644 index 000000000..e7545f83c --- /dev/null +++ b/ggml-cuda/sumrows.cuh @@ -0,0 +1,3 @@ +#include "common.cuh" + +void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/tsembd.cu b/ggml-cuda/tsembd.cu new file mode 100644 index 000000000..153ddbcda --- /dev/null +++ b/ggml-cuda/tsembd.cu @@ -0,0 +1,47 @@ +#include "tsembd.cuh" + +static __global__ void timestep_embedding_f32(const float * timesteps, float * dst, const int nb1, const int dim, const int max_period) { + // blockIDx.y: idx of timesteps->ne[0] + // blockIDx.x: idx of ((dim + 1) / 2) / BLOCK_SIZE + int i = blockIdx.y; + int j = threadIdx.x + blockIdx.x * blockDim.x; + float * embed_data = (float *)((char *)dst + i*nb1); + + if (dim % 2 != 0 && j == ((dim + 1) / 2)) { + embed_data[dim] = 0.f; + } + + int half = dim / 2; + if (j >= half) { + return; + } + + float timestep = timesteps[i]; + float freq = (float)expf(-logf(max_period) * j / half); + float arg = timestep * freq; + embed_data[j] = cosf(arg); + embed_data[j + half] = sinf(arg); +} + +static void timestep_embedding_f32_cuda(const float * x, float * dst, const int ne00, const int nb1, + const int dim, const int max_period, cudaStream_t stream) { + int half_ceil = (dim + 1) / 2; + int num_blocks = (half_ceil + CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE; + dim3 gridDim(num_blocks, ne00, 1); + timestep_embedding_f32<<>>(x, dst, nb1, dim, max_period); +} + +void ggml_cuda_op_timestep_embedding(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + const int dim = dst->op_params[0]; + const int max_period = dst->op_params[1]; + + timestep_embedding_f32_cuda(src0_d, dst_d, src0->ne[0], dst->nb[1], dim, max_period, stream); +} diff --git a/ggml-cuda/tsembd.cuh b/ggml-cuda/tsembd.cuh new file mode 100644 index 000000000..84340e3d7 --- /dev/null +++ b/ggml-cuda/tsembd.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE 256 + +void ggml_cuda_op_timestep_embedding(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/unary.cu b/ggml-cuda/unary.cu new file mode 100644 index 000000000..1a7f09469 --- /dev/null +++ b/ggml-cuda/unary.cu @@ -0,0 +1,240 @@ +#include "unary.cuh" + +static __global__ void gelu_f32(const float * x, float * dst, const int k) { + const float GELU_COEF_A = 0.044715f; + const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f; + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + + float xi = x[i]; + dst[i] = 0.5f*xi*(1.0f + tanhf(SQRT_2_OVER_PI*xi*(1.0f + GELU_COEF_A*xi*xi))); +} + +static __global__ void gelu_quick_f32(const float * x, float * dst, int k) { + const float GELU_QUICK_COEF = -1.702f; + const int i = blockDim.x*blockIdx.x + threadIdx.x; + if (i >= k) { + return; + } + dst[i] = x[i] * (1.0f / (1.0f + expf(GELU_QUICK_COEF * x[i]))); +} + +static __global__ void silu_f32(const float * x, float * dst, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + dst[i] = x[i] / (1.0f + expf(-x[i])); +} + +static __global__ void tanh_f32(const float * x, float * dst, int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + if (i >= k) { + return; + } + dst[i] = tanhf(x[i]); +} + +static __global__ void relu_f32(const float * x, float * dst, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + dst[i] = fmaxf(x[i], 0); +} + +static __global__ void hardsigmoid_f32(const float * x, float * dst, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + dst[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); +} + +static __global__ void hardswish_f32(const float * x, float * dst, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + dst[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); +} + +static __global__ void leaky_relu_f32(const float * x, float * dst, const int k, const float negative_slope) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + if (i >= k) { + return; + } + dst[i] = fmaxf(x[i], 0) + fminf(x[i], 0.0f) * negative_slope; +} + +static __global__ void sqr_f32(const float * x, float * dst, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + dst[i] = x[i] * x[i]; +} + +static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE; + gelu_f32<<>>(x, dst, k); +} + +static void gelu_quick_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE; + gelu_quick_f32<<>>(x, dst, k); +} + +static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_SILU_BLOCK_SIZE - 1) / CUDA_SILU_BLOCK_SIZE; + silu_f32<<>>(x, dst, k); +} + +static void tanh_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_TANH_BLOCK_SIZE - 1) / CUDA_TANH_BLOCK_SIZE; + tanh_f32<<>>(x, dst, k); +} + +static void relu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE; + relu_f32<<>>(x, dst, k); +} + +static void hardsigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_HARDSIGMOID_BLOCK_SIZE - 1) / CUDA_HARDSIGMOID_BLOCK_SIZE; + hardsigmoid_f32<<>>(x, dst, k); +} + +static void hardswish_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_HARDSWISH_BLOCK_SIZE - 1) / CUDA_HARDSWISH_BLOCK_SIZE; + hardswish_f32<<>>(x, dst, k); +} + +static void leaky_relu_f32_cuda(const float * x, float * dst, const int k, const float negative_slope, cudaStream_t stream) { + const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE; + leaky_relu_f32<<>>(x, dst, k, negative_slope); +} + +static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_SQR_BLOCK_SIZE - 1) / CUDA_SQR_BLOCK_SIZE; + sqr_f32<<>>(x, dst, k); +} + +void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + gelu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +} + +void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + silu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +} + +void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + gelu_quick_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +} + +void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + tanh_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +} + +void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + relu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +} + +void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + hardsigmoid_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +} + +void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + hardswish_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +} + +void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + float negative_slope; + memcpy(&negative_slope, dst->op_params, sizeof(float)); + + leaky_relu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), negative_slope, stream); +} + +void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + sqr_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +} diff --git a/ggml-cuda/unary.cuh b/ggml-cuda/unary.cuh new file mode 100644 index 000000000..2002ed989 --- /dev/null +++ b/ggml-cuda/unary.cuh @@ -0,0 +1,27 @@ +#include "common.cuh" + +#define CUDA_GELU_BLOCK_SIZE 256 +#define CUDA_SILU_BLOCK_SIZE 256 +#define CUDA_TANH_BLOCK_SIZE 256 +#define CUDA_RELU_BLOCK_SIZE 256 +#define CUDA_HARDSIGMOID_BLOCK_SIZE 256 +#define CUDA_HARDSWISH_BLOCK_SIZE 256 +#define CUDA_SQR_BLOCK_SIZE 256 + +void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/upscale.cu b/ggml-cuda/upscale.cu new file mode 100644 index 000000000..2f62fed48 --- /dev/null +++ b/ggml-cuda/upscale.cu @@ -0,0 +1,48 @@ +#include "upscale.cuh" + +static __global__ void upscale_f32(const float * x, float * dst, const int ne00, const int ne00xne01, const int scale_factor) { + // blockIdx.z: idx of ne02*ne03 + // blockIdx.y: idx of ne01*scale_factor， aka ne1 + // blockIDx.x: idx of ne00*scale_factor / BLOCK_SIZE + // ne00xne01: ne00 * ne01 + int ne0 = ne00 * scale_factor; + int nidx = threadIdx.x + blockIdx.x * blockDim.x; + if (nidx >= ne0) { + return; + } + // operation + int i00 = nidx / scale_factor; + int i01 = blockIdx.y / scale_factor; + int offset_src = + i00 + + i01 * ne00 + + blockIdx.z * ne00xne01; + int offset_dst = + nidx + + blockIdx.y * ne0 + + blockIdx.z * ne0 * gridDim.y; + dst[offset_dst] = x[offset_src]; +} + +static void upscale_f32_cuda(const float * x, float * dst, const int ne00, const int ne01, const int ne02, const int ne03, + const int scale_factor, cudaStream_t stream) { + int ne0 = (ne00 * scale_factor); + int num_blocks = (ne0 + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE; + dim3 gridDim(num_blocks, (ne01 * scale_factor), ne02*ne03); + upscale_f32<<>>(x, dst, ne00, ne00 * ne01, scale_factor); +} + +void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors + + const int scale_factor = dst->op_params[0]; + + upscale_f32_cuda(src0_d, dst_d, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], scale_factor, stream); +} diff --git a/ggml-cuda/upscale.cuh b/ggml-cuda/upscale.cuh new file mode 100644 index 000000000..d4d765230 --- /dev/null +++ b/ggml-cuda/upscale.cuh @@ -0,0 +1,5 @@ +#include "common.cuh" + +#define CUDA_UPSCALE_BLOCK_SIZE 256 + +void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/vecdotq.cuh b/ggml-cuda/vecdotq.cuh new file mode 100644 index 000000000..d911d851d --- /dev/null +++ b/ggml-cuda/vecdotq.cuh @@ -0,0 +1,1284 @@ +#include "common.cuh" + +static __device__ __forceinline__ int get_int_from_int8(const int8_t * x8, const int & i32) { + const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment + + int x32 = 0; + x32 |= x16[0] << 0; + x32 |= x16[1] << 16; + + return x32; +} + +static __device__ __forceinline__ int get_int_from_uint8(const uint8_t * x8, const int & i32) { + const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment + + int x32 = 0; + x32 |= x16[0] << 0; + x32 |= x16[1] << 16; + + return x32; +} + +static __device__ __forceinline__ int get_int_from_int8_aligned(const int8_t * x8, const int & i32) { + return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment +} + +static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t * x8, const int & i32) { + return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment +} + + +// VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called +// MMVQ = mul_mat_vec_q, MMQ = mul_mat_q + +#define VDR_Q4_0_Q8_1_MMVQ 2 +#define VDR_Q4_0_Q8_1_MMQ 4 + +template static __device__ __forceinline__ float vec_dot_q4_0_q8_1_impl( + const int * v, const int * u, const float & d4, const half2 & ds8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + const int vi0 = (v[i] >> 0) & 0x0F0F0F0F; + const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; + + // SIMD dot product of quantized values + sumi = __dp4a(vi0, u[2*i+0], sumi); + sumi = __dp4a(vi1, u[2*i+1], sumi); + } + + const float2 ds8f = __half22float2(ds8); + + // second part effectively subtracts 8 from each quant value + return d4 * (sumi * ds8f.x - (8*vdr/QI4_0) * ds8f.y); +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +#define VDR_Q4_1_Q8_1_MMVQ 2 +#define VDR_Q4_1_Q8_1_MMQ 4 + +template static __device__ __forceinline__ float vec_dot_q4_1_q8_1_impl( + const int * v, const int * u, const half2 & dm4, const half2 & ds8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + const int vi0 = (v[i] >> 0) & 0x0F0F0F0F; + const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; + + // SIMD dot product of quantized values + sumi = __dp4a(vi0, u[2*i+0], sumi); + sumi = __dp4a(vi1, u[2*i+1], sumi); + } + +#ifdef GGML_CUDA_F16 + const float2 tmp = __half22float2(__hmul2(dm4, ds8)); + const float d4d8 = tmp.x; + const float m4s8 = tmp.y; +#else + const float2 dm4f = __half22float2(dm4); + const float2 ds8f = __half22float2(ds8); + const float d4d8 = dm4f.x * ds8f.x; + const float m4s8 = dm4f.y * ds8f.y; +#endif // GGML_CUDA_F16 + + // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it + return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1)); +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +#define VDR_Q5_0_Q8_1_MMVQ 2 +#define VDR_Q5_0_Q8_1_MMQ 4 + +template static __device__ __forceinline__ float vec_dot_q5_0_q8_1_impl( + const int * vl, const int * vh, const int * u, const float & d5, const half2 & ds8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits + vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4 + vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12 + vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20 + vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28 + sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values + + int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits + vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4 + vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12 + vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20 + vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28 + sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values + } + + const float2 ds8f = __half22float2(ds8); + + // second part effectively subtracts 16 from each quant value + return d5 * (sumi * ds8f.x - (16*vdr/QI5_0) * ds8f.y); +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +#define VDR_Q5_1_Q8_1_MMVQ 2 +#define VDR_Q5_1_Q8_1_MMQ 4 + +template static __device__ __forceinline__ float vec_dot_q5_1_q8_1_impl( + const int * vl, const int * vh, const int * u, const half2 & dm5, const half2 & ds8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits + vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4 + vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12 + vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20 + vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28 + sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values + + int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits + vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4 + vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12 + vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20 + vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28 + sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values + } + +#ifdef GGML_CUDA_F16 + const float2 tmp = __half22float2(__hmul2(dm5, ds8)); + const float d5d8 = tmp.x; + const float m5s8 = tmp.y; +#else + const float2 dm5f = __half22float2(dm5); + const float2 ds8f = __half22float2(ds8); + const float d5d8 = dm5f.x * ds8f.x; + const float m5s8 = dm5f.y * ds8f.y; +#endif // GGML_CUDA_F16 + + // scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it + return sumi*d5d8 + m5s8 / (QI5_1 / vdr); + +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +#define VDR_Q8_0_Q8_1_MMVQ 2 +#define VDR_Q8_0_Q8_1_MMQ 8 + +template static __device__ __forceinline__ float vec_dot_q8_0_q8_1_impl( + const int * v, const int * u, const float & d8_0, const float & d8_1) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + // SIMD dot product of quantized values + sumi = __dp4a(v[i], u[i], sumi); + } + + return d8_0*d8_1 * sumi; +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +template static __device__ __forceinline__ float vec_dot_q8_1_q8_1_impl( + const int * v, const int * u, const half2 & dm8, const half2 & ds8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + // SIMD dot product of quantized values + sumi = __dp4a(v[i], u[i], sumi); + } + +#ifdef GGML_CUDA_F16 + const float2 tmp = __half22float2(__hmul2(dm8, ds8)); + const float d8d8 = tmp.x; + const float m8s8 = tmp.y; +#else + const float2 dm8f = __half22float2(dm8); + const float2 ds8f = __half22float2(ds8); + const float d8d8 = dm8f.x * ds8f.x; + const float m8s8 = dm8f.y * ds8f.y; +#endif // GGML_CUDA_F16 + + // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it + return sumi*d8d8 + m8s8 / (QI8_1 / vdr); +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +#define VDR_Q2_K_Q8_1_MMVQ 1 +#define VDR_Q2_K_Q8_1_MMQ 2 + +// contiguous v/x values +static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq( + const int & v, const int * __restrict__ u, const uint8_t * __restrict__ scales, + const half2 & dm2, const float * __restrict__ d8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR2_K; ++i) { + const int sc = scales[2*i]; + + const int vi = (v >> (2*i)) & 0x03030303; + + sumf_d += d8[i] * (__dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product + + // fill int with 4x m + int m = sc >> 4; + m |= m << 8; + m |= m << 16; + sumf_m += d8[i] * __dp4a(m, u[i], 0); // multiply constant q2_K part with sum of q8_1 values + } + + const float2 dm2f = __half22float2(dm2); + + return dm2f.x*sumf_d - dm2f.y*sumf_m; +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +// contiguous u/y values +static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq( + const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ scales, + const half2 & dm2, const float & d8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + int sumi_d = 0; + int sumi_m = 0; + +#pragma unroll + for (int i0 = 0; i0 < QI8_1; i0 += QI8_1/2) { + int sumi_d_sc = 0; + + const int sc = scales[i0 / (QI8_1/2)]; + + // fill int with 4x m + int m = sc >> 4; + m |= m << 8; + m |= m << 16; + +#pragma unroll + for (int i = i0; i < i0 + QI8_1/2; ++i) { + sumi_d_sc = __dp4a(v[i], u[i], sumi_d_sc); // SIMD dot product + sumi_m = __dp4a(m, u[i], sumi_m); // multiply sum of q8_1 values with m + } + + sumi_d += sumi_d_sc * (sc & 0xF); + } + + const float2 dm2f = __half22float2(dm2); + + return d8 * (dm2f.x*sumi_d - dm2f.y*sumi_m); +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +#define VDR_Q3_K_Q8_1_MMVQ 1 +#define VDR_Q3_K_Q8_1_MMQ 2 + +// contiguous v/x values +static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq( + const int & vl, const int & vh, const int * __restrict__ u, const uint8_t * __restrict__ scales, + const int & scale_offset, const float & d3, const float * __restrict__ d8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + float sumf = 0.0f; + +#pragma unroll + for (int i = 0; i < QR3_K; ++i) { + const int isc = scale_offset + 2*i; + + const int isc_low = isc % (QK_K/32); + const int sc_shift_low = 4 * (isc / (QK_K/32)); + const int sc_low = (scales[isc_low] >> sc_shift_low) & 0xF; + + const int isc_high = isc % (QK_K/64); + const int sc_shift_high = 2 * (isc / (QK_K/64)); + const int sc_high = ((scales[(QK_K/32) + isc_high] >> sc_shift_high) & 3) << 4; + + const int sc = (sc_low | sc_high) - 32; + + const int vil = (vl >> (2*i)) & 0x03030303; + + const int vih = ((vh >> i) << 2) & 0x04040404; + + const int vi = __vsubss4(vil, vih); + + sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product + } + + return d3 * sumf; +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +// contiguous u/y values +static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq( + const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ scales, + const float & d3, const float & d8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + int sumi = 0; + +#pragma unroll + for (int i0 = 0; i0 < QR3_K*VDR_Q3_K_Q8_1_MMQ; i0 += QI8_1/2) { + int sumi_sc = 0; + + for (int i = i0; i < i0 + QI8_1/2; ++i) { + sumi_sc = __dp4a(v[i], u[i], sumi_sc); // SIMD dot product + } + + sumi += sumi_sc * scales[i0 / (QI8_1/2)]; + } + + return d3*d8 * sumi; +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +#define VDR_Q4_K_Q8_1_MMVQ 2 +#define VDR_Q4_K_Q8_1_MMQ 8 + +// contiguous v/x values +static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq( + const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc, + const uint8_t * __restrict__ m, const half2 & dm4, const float * __restrict__ d8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR4_K; ++i) { + const int v0i = (v[0] >> (4*i)) & 0x0F0F0F0F; + const int v1i = (v[1] >> (4*i)) & 0x0F0F0F0F; + + const int dot1 = __dp4a(v1i, u[2*i+1], __dp4a(v0i, u[2*i+0], 0)); // SIMD dot product + const int dot2 = __dp4a(0x01010101, u[2*i+1], __dp4a(0x01010101, u[2*i+0], 0)); // sum of u + + sumf_d += d8[i] * (dot1 * sc[i]); + sumf_m += d8[i] * (dot2 * m[i]); // multiply constant part of q4_K with sum of q8_1 values + } + + const float2 dm4f = __half22float2(dm4); + + return dm4f.x*sumf_d - dm4f.y*sumf_m; + +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +// contiguous u/y values +static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq( + const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc, + const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR4_K*VDR_Q4_K_Q8_1_MMQ/QI8_1; ++i) { + int sumi_d = 0; + +#pragma unroll + for (int j = 0; j < QI8_1; ++j) { + sumi_d = __dp4a((v[j] >> (4*i)) & 0x0F0F0F0F, u[i*QI8_1 + j], sumi_d); // SIMD dot product + } + + const float2 ds8f = __half22float2(ds8[i]); + + sumf_d += ds8f.x * (sc[i] * sumi_d); + sumf_m += ds8f.y * m[i]; // sum of q8_1 block * q4_K min val + } + + const float2 dm4f = __half22float2(dm4); + + return dm4f.x*sumf_d - dm4f.y*sumf_m; + +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +#define VDR_Q5_K_Q8_1_MMVQ 2 +#define VDR_Q5_K_Q8_1_MMQ 8 + +// contiguous v/x values +static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq( + const int * __restrict__ vl, const int * __restrict__ vh, const int * __restrict__ u, const uint8_t * __restrict__ sc, + const uint8_t * __restrict__ m, const half2 & dm5, const float * __restrict__ d8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR5_K; ++i) { + const int vl0i = (vl[0] >> (4*i)) & 0x0F0F0F0F; + const int vl1i = (vl[1] >> (4*i)) & 0x0F0F0F0F; + + const int vh0i = ((vh[0] >> i) << 4) & 0x10101010; + const int vh1i = ((vh[1] >> i) << 4) & 0x10101010; + + const int v0i = vl0i | vh0i; + const int v1i = vl1i | vh1i; + + const int dot1 = __dp4a(v0i, u[2*i+0], __dp4a(v1i, u[2*i+1], 0)); // SIMD dot product + const int dot2 = __dp4a(0x01010101, u[2*i+0], __dp4a(0x01010101, u[2*i+1], 0)); // sum of u + + sumf_d += d8[i] * (dot1 * sc[i]); + sumf_m += d8[i] * (dot2 * m[i]); + + } + + const float2 dm5f = __half22float2(dm5); + + return dm5f.x*sumf_d - dm5f.y*sumf_m; + +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +// contiguous u/y values +static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq( + const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc, + const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR5_K*VDR_Q5_K_Q8_1_MMQ/QI8_1; ++i) { + int sumi_d = 0; + +#pragma unroll + for (int j = 0; j < QI8_1; ++j) { + sumi_d = __dp4a(v[i*QI8_1 + j], u[i*QI8_1 + j], sumi_d); // SIMD dot product + } + + const float2 ds8f = __half22float2(ds8[i]); + + sumf_d += ds8f.x * (sc[i] * sumi_d); + sumf_m += ds8f.y * m[i]; // sum of q8_1 block * q4_K min val + } + + const float2 dm4f = __half22float2(dm4); + + return dm4f.x*sumf_d - dm4f.y*sumf_m; + +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +#define VDR_Q6_K_Q8_1_MMVQ 1 +#define VDR_Q6_K_Q8_1_MMQ 8 + +// contiguous v/x values +static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq( + const int & vl, const int & vh, const int * __restrict__ u, const int8_t * __restrict__ scales, + const float & d, const float * __restrict__ d8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + float sumf = 0.0f; + +#pragma unroll + for (int i = 0; i < QR6_K; ++i) { + const int sc = scales[4*i]; + + const int vil = (vl >> (4*i)) & 0x0F0F0F0F; + + const int vih = ((vh >> (4*i)) << 4) & 0x30303030; + + const int vi = __vsubss4((vil | vih), 0x20202020); // vi = (vil | vih) - 32 + + sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product + } + + return d*sumf; +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +// contiguous u/y values +static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq( + const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ sc, + const float & d6, const float * __restrict__ d8) { + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + float sumf_d = 0.0f; + +#pragma unroll + for (int i0 = 0; i0 < VDR_Q6_K_Q8_1_MMQ; i0 += 4) { + int2 sumi_d = {0, 0}; // 2 q6_K scales per q8_1 scale + +#pragma unroll + for (int i = i0; i < i0 + 2; ++i) { + sumi_d.x = __dp4a(v[2*i+0], u[2*i+0], sumi_d.x); // SIMD dot product + sumi_d.x = __dp4a(v[2*i+1], u[2*i+1], sumi_d.x); // SIMD dot product + + sumi_d.y = __dp4a(v[2*i+4], u[2*i+4], sumi_d.y); // SIMD dot product + sumi_d.y = __dp4a(v[2*i+5], u[2*i+5], sumi_d.y); // SIMD dot product + } + + sumf_d += d8[i0/4] * (sc[i0/2+0]*sumi_d.x + sc[i0/2+1]*sumi_d.y); + } + + return d6 * sumf_d; + +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A +} + +static __device__ __forceinline__ float vec_dot_q4_0_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { + + const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq; + + int v[VDR_Q4_0_Q8_1_MMVQ]; + int u[2*VDR_Q4_0_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) { + v[i] = get_int_from_uint8(bq4_0->qs, iqs + i); + u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0); + } + + return vec_dot_q4_0_q8_1_impl(v, u, bq4_0->d, bq8_1->ds); +} + + +static __device__ __forceinline__ float vec_dot_q4_1_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { + + const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq; + + int v[VDR_Q4_1_Q8_1_MMVQ]; + int u[2*VDR_Q4_1_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q4_1_Q8_1_MMVQ; ++i) { + v[i] = get_int_from_uint8_aligned(bq4_1->qs, iqs + i); + u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_1); + } + + return vec_dot_q4_1_q8_1_impl(v, u, bq4_1->dm, bq8_1->ds); +} + +static __device__ __forceinline__ float vec_dot_q5_0_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { + + const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq; + + int vl[VDR_Q5_0_Q8_1_MMVQ]; + int vh[VDR_Q5_0_Q8_1_MMVQ]; + int u[2*VDR_Q5_0_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q5_0_Q8_1_MMVQ; ++i) { + vl[i] = get_int_from_uint8(bq5_0->qs, iqs + i); + vh[i] = get_int_from_uint8(bq5_0->qh, 0) >> (4 * (iqs + i)); + u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_0); + } + + return vec_dot_q5_0_q8_1_impl(vl, vh, u, bq5_0->d, bq8_1->ds); +} + +static __device__ __forceinline__ float vec_dot_q5_1_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { + + const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq; + + int vl[VDR_Q5_1_Q8_1_MMVQ]; + int vh[VDR_Q5_1_Q8_1_MMVQ]; + int u[2*VDR_Q5_1_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q5_1_Q8_1_MMVQ; ++i) { + vl[i] = get_int_from_uint8_aligned(bq5_1->qs, iqs + i); + vh[i] = get_int_from_uint8_aligned(bq5_1->qh, 0) >> (4 * (iqs + i)); + u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_1); + } + + return vec_dot_q5_1_q8_1_impl(vl, vh, u, bq5_1->dm, bq8_1->ds); +} + +static __device__ __forceinline__ float vec_dot_q8_0_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { + + const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq; + + int v[VDR_Q8_0_Q8_1_MMVQ]; + int u[VDR_Q8_0_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q8_0_Q8_1_MMVQ; ++i) { + v[i] = get_int_from_int8(bq8_0->qs, iqs + i); + u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + } + + return vec_dot_q8_0_q8_1_impl(v, u, bq8_0->d, __low2half(bq8_1->ds)); +} + +static __device__ __forceinline__ float vec_dot_q2_K_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { + + const block_q2_K * bq2_K = (const block_q2_K *) vbq; + + const int bq8_offset = QR2_K * (iqs / QI8_1); + const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2); + + const uint8_t * scales = bq2_K->scales + scale_offset; + + const int v = get_int_from_uint8_aligned(bq2_K->qs, iqs); + int u[QR2_K]; + float d8[QR2_K]; + +#pragma unroll + for (int i = 0; i < QR2_K; ++ i) { + u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); + d8[i] = __low2float(bq8_1[bq8_offset + i].ds); + } + + return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8); +} + +static __device__ __forceinline__ float vec_dot_q3_K_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { + + const block_q3_K * bq3_K = (const block_q3_K *) vbq; + + const int bq8_offset = QR3_K * (iqs / (QI3_K/2)); + const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2); + + const float d = bq3_K->d; + + const int vl = get_int_from_uint8(bq3_K->qs, iqs); + + // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted + const int vh = ~get_int_from_uint8(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset; + + int u[QR3_K]; + float d8[QR3_K]; + +#pragma unroll + for (int i = 0; i < QR3_K; ++i) { + u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); + d8[i] = __low2float(bq8_1[bq8_offset + i].ds); + } + + return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8); +} + +static __device__ __forceinline__ float vec_dot_q4_K_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { + +#ifndef GGML_QKK_64 + const block_q4_K * bq4_K = (const block_q4_K *) vbq; + + int v[2]; + int u[2*QR4_K]; + float d8[QR4_K]; + + // iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6 + const int bq8_offset = QR4_K * ((iqs/2) / (QI8_1/2)); + + // iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12 + // iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44 + // iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76 + // iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108 + + const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4)); + v[0] = q4[0]; + v[1] = q4[4]; + + const uint16_t * scales = (const uint16_t *)bq4_K->scales; + uint16_t aux[2]; + const int j = bq8_offset/2; + if (j < 2) { + aux[0] = scales[j+0] & 0x3f3f; + aux[1] = scales[j+2] & 0x3f3f; + } else { + aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2); + aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2); + } + const uint8_t * sc = (const uint8_t *)aux; + const uint8_t * m = sc + 2; + + for (int i = 0; i < QR4_K; ++i) { + const block_q8_1 * bq8i = bq8_1 + bq8_offset + i; + d8[i] = __low2float(bq8i->ds); + + const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4); + u[2*i+0] = q8[0]; + u[2*i+1] = q8[4]; + } + + return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8); + +#else + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + const block_q4_K * bq4_K = (const block_q4_K *) vbq; + + float sumf_d = 0.0f; + float sumf_m = 0.0f; + + uint16_t aux16[2]; + const uint8_t * s = (const uint8_t *)aux16; + + const uint16_t * a = (const uint16_t *)bq4_K->scales; + aux16[0] = a[0] & 0x0f0f; + aux16[1] = (a[0] >> 4) & 0x0f0f; + + const float dall = bq4_K->dm[0]; + const float dmin = bq4_K->dm[1]; + + const float d8_1 = __low2float(bq8_1[0].ds); + const float d8_2 = __low2float(bq8_1[1].ds); + + const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); + const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); + const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2)); + const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4); + + const int * q4 = (const int *)bq4_K->qs + (iqs/2); + const int v1 = q4[0]; + const int v2 = q4[4]; + + const int dot1 = __dp4a(ui2, v2 & 0x0f0f0f0f, __dp4a(ui1, v1 & 0x0f0f0f0f, 0)); + const int dot2 = __dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, __dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0)); + const int dot3 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0)); + const int dot4 = __dp4a(0x01010101, ui4, __dp4a(0x01010101, ui3, 0)); + + sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]); + sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]); + + return dall * sumf_d - dmin * sumf_m; + +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A + +#endif +} + +static __device__ __forceinline__ float vec_dot_q5_K_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { + +#ifndef GGML_QKK_64 + const block_q5_K * bq5_K = (const block_q5_K *) vbq; + + int vl[2]; + int vh[2]; + int u[2*QR5_K]; + float d8[QR5_K]; + + const int bq8_offset = QR5_K * ((iqs/2) / (QI8_1/2)); + const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4)); + const int * qh = (const int *)(bq5_K->qh + 4 * ((iqs/2)%4)); + + vl[0] = ql[0]; + vl[1] = ql[4]; + + vh[0] = qh[0] >> bq8_offset; + vh[1] = qh[4] >> bq8_offset; + + const uint16_t * scales = (const uint16_t *)bq5_K->scales; + uint16_t aux[2]; + const int j = bq8_offset/2; + if (j < 2) { + aux[0] = scales[j+0] & 0x3f3f; + aux[1] = scales[j+2] & 0x3f3f; + } else { + aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2); + aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2); + } + const uint8_t * sc = (const uint8_t *)aux; + const uint8_t * m = sc + 2; + +#pragma unroll + for (int i = 0; i < QR5_K; ++i) { + const block_q8_1 * bq8i = bq8_1 + bq8_offset + i; + d8[i] = __low2float(bq8i->ds); + + const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4); + u[2*i+0] = q8[0]; + u[2*i+1] = q8[4]; + } + + return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8); + +#else + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + const block_q5_K * bq5_K = (const block_q5_K *) vbq; + + const int8_t * s = bq5_K->scales; + + const float d = bq5_K->d; + + const float d8_1 = __low2half(bq8_1[0].ds); + const float d8_2 = __low2half(bq8_1[1].ds); + + const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); + const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); + const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2)); + const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4); + + const int * ql = (const int *)bq5_K->qs + (iqs/2); + const int vl1 = ql[0]; + const int vl2 = ql[4]; + + const int step = 4 * (iqs/2); // 0, 4, 8, 12 + const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6 + const int in = step%8; // 0, 4, 0, 4 + const int vh = (*((const int *)(bq5_K->qh + in))) >> im; + + const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f); + const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f); + const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f); + const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f); + + const float sumf_d = d8_1 * (__dp4a(ui1, v1, 0) * s[0] + __dp4a(ui2, v2, 0) * s[1]) + + d8_2 * (__dp4a(ui3, v3, 0) * s[2] + __dp4a(ui4, v4, 0) * s[3]); + + return d * sumf_d; + +#else + NO_DEVICE_CODE; +#endif // __CUDA_ARCH__ >= MIN_CC_DP4A + +#endif +} + +static __device__ __forceinline__ float vec_dot_q6_K_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { + + const block_q6_K * bq6_K = (const block_q6_K *) vbq; + + const int bq8_offset = 2 * QR6_K * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/4); + const int scale_offset = (QI6_K/4) * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/8); + const int vh_shift = 2 * ((iqs % (QI6_K/2)) / (QI6_K/4)); + + const int vl = get_int_from_uint8(bq6_K->ql, iqs); + const int vh = get_int_from_uint8(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift; + + const int8_t * scales = bq6_K->scales + scale_offset; + + int u[QR6_K]; + float d8[QR6_K]; + +#pragma unroll + for (int i = 0; i < QR6_K; ++i) { + u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1); + d8[i] = __low2float(bq8_1[bq8_offset + 2*i].ds); + } + + return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8); +} + +static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { +#if QK_K == 256 + const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq; + +#if QR2_XXS == 8 + const int ib32 = iqs; + const uint16_t * q2 = bq2->qs + 4*ib32; + const uint8_t * aux8 = (const uint8_t *)q2; + const int8_t * q8 = bq8_1[ib32].qs; + uint32_t aux32 = q2[2] | (q2[3] << 16); + int sumi = 0; + for (int l = 0; l < 4; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); + const uint8_t signs = ksigns_iq2xs[aux32 & 127]; + for (int j = 0; j < 8; ++j) { + sumi += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1); + } + q8 += 8; + aux32 >>= 7; + } + const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.25f; + return d * sumi; +#else + // iqs is 0...15 + const int ib32 = iqs/2; + const int il = iqs%2; + const uint16_t * q2 = bq2->qs + 4*ib32; + const uint8_t * aux8 = (const uint8_t *)q2; + const uint8_t * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]); + const uint8_t * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]); + const uint32_t aux32 = q2[2] | (q2[3] << 16); + const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * __low2float(bq8_1[ib32].ds) * 0.25f; + const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127]; + const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127]; + const int8_t * q8 = bq8_1[ib32].qs + 16*il; + int sumi1 = 0, sumi2 = 0; + for (int j = 0; j < 8; ++j) { + sumi1 += q8[j+0] * grid1[j] * (signs1 & kmask_iq2xs[j] ? -1 : 1); + sumi2 += q8[j+8] * grid2[j] * (signs2 & kmask_iq2xs[j] ? -1 : 1); + } + return d * (sumi1 + sumi2); +#endif +#else + assert(false); + return 0.f; +#endif +} + +static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics +#if QK_K == 256 + const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq; + + const int ib32 = iqs; + const uint16_t * q2 = bq2->qs + 4*ib32; + const int8_t * q8 = bq8_1[ib32].qs; + const uint8_t ls1 = bq2->scales[ib32] & 0xf; + const uint8_t ls2 = bq2->scales[ib32] >> 4; + int sumi1 = 0; + for (int l = 0; l < 2; ++l) { + const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511)); + const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9)); + const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]); + const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]); + sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1); + sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1); + q8 += 8; + } + int sumi2 = 0; + for (int l = 2; l < 4; ++l) { + const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511)); + const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9)); + const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]); + const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]); + sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2); + sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2); + q8 += 8; + } + const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f; + return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); +#else + GGML_UNUSED(ksigns64); + assert(false); + return 0.f; +#endif +#else + GGML_UNUSED(ksigns64); + assert(false); + return 0.f; +#endif +} + +// TODO +static __device__ __forceinline__ float vec_dot_iq2_s_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics +#if QK_K == 256 + const block_iq2_s * bq2 = (const block_iq2_s *) vbq; + + const int ib32 = iqs; + const int8_t * q8 = bq8_1[ib32].qs; + const uint8_t * signs = bq2->qs + QK_K/8 + 4*ib32; + const uint8_t ls1 = bq2->scales[ib32] & 0xf; + const uint8_t ls2 = bq2->scales[ib32] >> 4; + int sumi1 = 0; + for (int l = 0; l < 2; ++l) { + const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300))); + const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201); + const uint32_t signs1 = __vcmpeq4(((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201); + const int grid_l = __vsub4(grid[0] ^ signs0, signs0); + const int grid_h = __vsub4(grid[1] ^ signs1, signs1); + sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1); + sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1); + q8 += 8; + } + int sumi2 = 0; + for (int l = 2; l < 4; ++l) { + const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300))); + const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201); + const uint32_t signs1 = __vcmpeq4(((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201); + const int grid_l = __vsub4(grid[0] ^ signs0, signs0); + const int grid_h = __vsub4(grid[1] ^ signs1, signs1); + sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2); + sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2); + q8 += 8; + } + const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f; + return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); +#else + GGML_UNUSED(ksigns64); + assert(false); + return 0.f; +#endif +#else + GGML_UNUSED(ksigns64); + assert(false); + return 0.f; +#endif +} + +static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics +#if QK_K == 256 + const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq; + + const int ib32 = iqs; + const uint8_t * q3 = bq2->qs + 8*ib32; + const uint16_t * gas = (const uint16_t *)(bq2->qs + QK_K/4) + 2*ib32; + const int8_t * q8 = bq8_1[ib32].qs; + uint32_t aux32 = gas[0] | (gas[1] << 16); + int sumi = 0; + for (int l = 0; l < 4; ++l) { + const uint32_t * grid1 = iq3xxs_grid + q3[2*l+0]; + const uint32_t * grid2 = iq3xxs_grid + q3[2*l+1]; + const uint32_t * signs = (const uint32_t *)(ksigns64 + (aux32 & 127)); + const int grid_l = __vsub4(grid1[0] ^ signs[0], signs[0]); + const int grid_h = __vsub4(grid2[0] ^ signs[1], signs[1]); + sumi = __dp4a(grid_l, *((int *)q8+0), sumi); + sumi = __dp4a(grid_h, *((int *)q8+1), sumi); + q8 += 8; + aux32 >>= 7; + } + const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.5f; + return d * sumi; +#else + assert(false); + return 0.f; +#endif +#else + assert(false); + return 0.f; +#endif +} + +// TODO: don't use lookup table for signs +static __device__ __forceinline__ float vec_dot_iq3_s_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics +#if QK_K == 256 + const block_iq3_s * bq2 = (const block_iq3_s *) vbq; + + const int ib32 = iqs; + const uint8_t * qs = bq2->qs + 8*ib32; + const int8_t * q8 = bq8_1[ib32].qs; + int sumi = 0; + for (int l = 0; l < 4; ++l) { + const uint32_t * grid1 = iq3s_grid + (qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256)); + const uint32_t * grid2 = iq3s_grid + (qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256)); + uint32_t signs0 = __vcmpeq4(((bq2->signs[4*ib32+l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201); + uint32_t signs1 = __vcmpeq4(((bq2->signs[4*ib32+l] >> 4) * 0x01010101) & 0x08040201, 0x08040201); + const int grid_l = __vsub4(grid1[0] ^ signs0, signs0); + const int grid_h = __vsub4(grid2[0] ^ signs1, signs1); + sumi = __dp4a(grid_l, *((int *)q8+0), sumi); + sumi = __dp4a(grid_h, *((int *)q8+1), sumi); + q8 += 8; + } + const float d = (float)bq2->d * (1 + 2*((bq2->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * __low2float(bq8_1[ib32].ds); + return d * sumi; +#else + assert(false); + return 0.f; +#endif +#else + assert(false); + return 0.f; +#endif +} + +static __device__ __forceinline__ float vec_dot_iq1_s_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { +#if QK_K == 256 + const block_iq1_s * bq1 = (const block_iq1_s *) vbq; + + const int ib32 = iqs; + int sumi = 0; +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + const int * q8 = (const int *)bq8_1[ib32].qs; + for (int l = 0; l < 4; ++l) { + const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8))); + int grid0 = grid[0] & 0x0f0f0f0f; + int grid1 = (grid[0] >> 4) & 0x0f0f0f0f; + sumi = __dp4a(q8[2*l+1], grid1, __dp4a(q8[2*l+0], grid0, sumi)); + } +#else + const int8_t * q8 = bq8_1[ib32].qs; + for (int l = 0; l < 4; ++l) { + const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8))); + for (int j = 0; j < 4; ++j) { + sumi += q8[j] * (grid[j] & 0xf) + q8[j+4] * (grid[j] >> 4); + } + q8 += 8; + } +#endif + const float delta = bq1->qh[ib32] & 0x8000 ? -1-IQ1S_DELTA : -1+IQ1S_DELTA; + const float d1q = (float)bq1->d * (2*((bq1->qh[ib32] >> 12) & 7) + 1); + const float d = d1q * __low2float (bq8_1[ib32].ds); + const float m = d1q * __high2float(bq8_1[ib32].ds); + return d * sumi + m * delta; +#else + assert(false); + return 0.f; +#endif +} + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics +static __device__ __forceinline__ void get_int_from_table_16(const uint32_t & q4, const uint8_t * values, + int & val1, int & val2) { + + uint32_t aux32; const uint8_t * q8 = (const uint8_t *)&aux32; + aux32 = q4 & 0x0f0f0f0f; + uint16_t v1 = values[q8[0]] | (values[q8[1]] << 8); + uint16_t v2 = values[q8[2]] | (values[q8[3]] << 8); + val1 = v1 | (v2 << 16); + aux32 = (q4 >> 4) & 0x0f0f0f0f; + v1 = values[q8[0]] | (values[q8[1]] << 8); + v2 = values[q8[2]] | (values[q8[3]] << 8); + val2 = v1 | (v2 << 16); +} +#endif + +static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { + + const block_iq4_nl * bq = (const block_iq4_nl *) vbq; + +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + const uint16_t * q4 = (const uint16_t *)bq->qs + 2*iqs; + const int32_t * q8 = (const int32_t *)bq8_1->qs + iqs; + + const uint8_t * values = (const uint8_t *)kvalues_iq4nl; + + int v1, v2; + int sumi1 = 0, sumi2 = 0; + for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) { + const uint32_t aux = q4[2*l] | (q4[2*l+1] << 16); + get_int_from_table_16(aux, values, v1, v2); + sumi1 = __dp4a(v1, q8[l+0], sumi1); + sumi2 = __dp4a(v2, q8[l+4], sumi2); + } + +#else + const uint8_t * q4 = bq->qs + 4*iqs; + const int8_t * q8 = bq8_1->qs + 4*iqs; + + int sumi1 = 0, sumi2 = 0; + for (int l = 0; l < 4*VDR_Q4_0_Q8_1_MMVQ; ++l) { + sumi1 += q8[l+ 0] * kvalues_iq4nl[q4[l] & 0xf]; + sumi2 += q8[l+16] * kvalues_iq4nl[q4[l] >> 4]; + } +#endif + const float d = (float)bq->d * __low2float(bq8_1->ds); + return d * (sumi1 + sumi2); +} + +static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1( + const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) { + +#if QK_K == 256 +#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics + + const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq; + const uint8_t * values = (const uint8_t *)kvalues_iq4nl; + + //// iqs is 0...7 + //const int ib64 = iqs/2; + //const int il = iqs%2; + //const int32_t * q8_1 = (const int *)bq8_1[2*ib64+0].qs + 2*il; + //const int32_t * q8_2 = (const int *)bq8_1[2*ib64+1].qs + 2*il; + //const uint32_t * q4_1 = (const uint32_t *)bq4->qs + 8*ib64 + 2*il; + //const uint32_t * q4_2 = q4_1 + 4; + //const int8_t ls1 = (bq4->scales_l[ib64] & 0xf) | (((bq4->scales_h >> (4*ib64+0)) & 3) << 4); + //const int8_t ls2 = (bq4->scales_l[ib64] >> 4) | (((bq4->scales_h >> (4*ib64+2)) & 3) << 4); + //const float d1 = (float)bq4->d * (ls1 - 32) * __low2float(bq8_1[2*ib64+0].ds); + //const float d2 = (float)bq4->d * (ls2 - 32) * __low2float(bq8_1[2*ib64+1].ds); + //int v1, v2; + //int sumi1 = 0, sumi2 = 0; + //for (int j = 0; j < 2; ++j) { + // get_int_from_table_16(q4_1[j], values, v1, v2); + // sumi1 = __dp4a(v2, q8_1[j+4], __dp4a(v1, q8_1[j+0], sumi1)); + // get_int_from_table_16(q4_2[j], values, v1, v2); + // sumi2 = __dp4a(v2, q8_2[j+4], __dp4a(v1, q8_2[j+0], sumi2)); + //} + //return d1 * sumi1 + d2 * sumi2; + + // iqs is 0...7 + const int ib32 = iqs; + const int32_t * q8 = (const int *)bq8_1[ib32].qs; + const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32; + const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4); + const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds); + int v1, v2; + int sumi1 = 0, sumi2 = 0; + for (int j = 0; j < 4; ++j) { + get_int_from_table_16(q4[j], values, v1, v2); + sumi1 = __dp4a(v1, q8[j+0], sumi1); + sumi2 = __dp4a(v2, q8[j+4], sumi2); + } + return d * (sumi1 + sumi2); + + //// iqs is 0...15 + //const int ib32 = iqs/2; + //const int il = iqs%2; + //const int32_t * q8 = (const int *)bq8_1[ib32].qs + 2*il; + //const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32 + 2*il; + //const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4); + //const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds); + //int v1, v2; + //int sumi1 = 0, sumi2 = 0; + //for (int j = 0; j < 2; ++j) { + // get_int_from_table_16(q4[j], values, v1, v2); + // sumi1 = __dp4a(v1, q8[j+0], sumi1); + // sumi2 = __dp4a(v2, q8[j+4], sumi2); + //} + //return d * (sumi1 + sumi2); +#else + assert(false); + return 0.f; +#endif +#else + return vec_dot_iq4_xs_q8_1(vbq, bq8_1, iqs); +#endif +} diff --git a/ggml-metal.m b/ggml-metal.m index 109e5fe6b..416b24532 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -173,8 +173,9 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0, GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1, - //GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0, - //GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1, + GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0, + GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1, + GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL, GGML_METAL_KERNEL_TYPE_CPY_F16_F16, GGML_METAL_KERNEL_TYPE_CPY_F16_F32, GGML_METAL_KERNEL_TYPE_CONCAT, @@ -598,8 +599,9 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0, cpy_f32_q4_0, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1, cpy_f32_q4_1, true); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0, cpy_f32_q5_0, true); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1, cpy_f32_q5_1, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0, cpy_f32_q5_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1, cpy_f32_q5_1, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL, cpy_f32_iq4_nl, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16, cpy_f16_f16, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32, cpy_f16_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT, concat, true); @@ -739,6 +741,9 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const case GGML_TYPE_Q8_0: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_IQ4_NL: return true; default: return false; @@ -1387,6 +1392,14 @@ static enum ggml_status ggml_metal_graph_compute( (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) { //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12); + // some Metal matrix data types require aligned pointers + // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5) + switch (src0->type) { + case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break; + case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8 == 0); break; + default: break; + } + id pipeline = nil; switch (src0->type) { @@ -1701,6 +1714,14 @@ static enum ggml_status ggml_metal_graph_compute( ne20 % 32 == 0 && ne20 >= 64 && ne11 > ne11_mm_min) { + // some Metal matrix data types require aligned pointers + // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5) + switch (src0->type) { + case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break; + case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8 == 0); break; + default: break; + } + id pipeline = nil; switch (src2->type) { @@ -2431,13 +2452,14 @@ static enum ggml_status ggml_metal_graph_compute( GGML_ASSERT(ne0 % ggml_blck_size(dst->type) == 0); switch (dstt) { - case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F16].pipeline; break; - case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline; break; - case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0].pipeline; break; - case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0].pipeline; break; - case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1].pipeline; break; - //case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0].pipeline; break; - //case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F16].pipeline; break; + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline; break; + case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0].pipeline; break; + case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0].pipeline; break; + case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1].pipeline; break; + case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0].pipeline; break; + case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1].pipeline; break; + case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL].pipeline; break; default: GGML_ASSERT(false && "not implemented"); }; } break; diff --git a/ggml-metal.metal b/ggml-metal.metal index 63de56325..748f0acef 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -2388,6 +2388,242 @@ kernel void kernel_cpy_f32_q4_1( } } +kernel void kernel_cpy_f32_q5_0( + device const float * src0, + device void * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant int64_t & ne03, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant uint64_t & nb03, + constant int64_t & ne0, + constant int64_t & ne1, + constant int64_t & ne2, + constant int64_t & ne3, + constant uint64_t & nb0, + constant uint64_t & nb1, + constant uint64_t & nb2, + constant uint64_t & nb3, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + const int64_t i03 = tgpig[2]; + const int64_t i02 = tgpig[1]; + const int64_t i01 = tgpig[0]; + + const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; + + const int64_t i3 = n / (ne2*ne1*ne0); + const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0); + const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0; + const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK5_0; + + device block_q5_0 * dst_data = (device block_q5_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + for (int64_t i00 = tpitg.x*QK5_0; i00 < ne00; i00 += ntg.x*QK5_0) { + device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00); + + float amax = 0.0f; // absolute max + float max = 0.0f; + + for (int j = 0; j < QK5_0; j++) { + const float v = src[j]; + if (amax < fabs(v)) { + amax = fabs(v); + max = v; + } + } + + const float d = max / -16; + const float id = d ? 1.0f/d : 0.0f; + + dst_data[i00/QK5_0].d = d; + + uint32_t qh = 0; + for (int j = 0; j < QK5_0/2; ++j) { + const float x0 = src[0 + j]*id; + const float x1 = src[QK5_0/2 + j]*id; + + const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f)); + const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f)); + + dst_data[i00/QK5_0].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4); + qh |= ((xi0 & 0x10u) >> 4) << (j + 0); + qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2); + } + thread const uint8_t * qh8 = (thread const uint8_t *)&qh; + for (int j = 0; j < 4; ++j) { + dst_data[i00/QK5_0].qh[j] = qh8[j]; + } + } +} + +kernel void kernel_cpy_f32_q5_1( + device const float * src0, + device void * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant int64_t & ne03, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant uint64_t & nb03, + constant int64_t & ne0, + constant int64_t & ne1, + constant int64_t & ne2, + constant int64_t & ne3, + constant uint64_t & nb0, + constant uint64_t & nb1, + constant uint64_t & nb2, + constant uint64_t & nb3, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + const int64_t i03 = tgpig[2]; + const int64_t i02 = tgpig[1]; + const int64_t i01 = tgpig[0]; + + const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; + + const int64_t i3 = n / (ne2*ne1*ne0); + const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0); + const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0; + const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK5_1; + + device block_q5_1 * dst_data = (device block_q5_1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + for (int64_t i00 = tpitg.x*QK5_1; i00 < ne00; i00 += ntg.x*QK5_1) { + device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00); + + float max = src[0]; + float min = src[0]; + + for (int j = 1; j < QK5_1; j++) { + const float v = src[j]; + min = v < min ? v : min; + max = v > max ? v : max; + } + + const float d = (max - min) / 31; + const float id = d ? 1.0f/d : 0.0f; + + dst_data[i00/QK5_1].d = d; + dst_data[i00/QK5_1].m = min; + + uint32_t qh = 0; + for (int j = 0; j < QK5_1/2; ++j) { + const float x0 = (src[0 + j] - min)*id; + const float x1 = (src[QK5_1/2 + j] - min)*id; + + const uint8_t xi0 = (uint8_t)(x0 + 0.5f); + const uint8_t xi1 = (uint8_t)(x1 + 0.5f); + + dst_data[i00/QK5_1].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4); + qh |= ((xi0 & 0x10u) >> 4) << (j + 0); + qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2); + } + thread const uint8_t * qh8 = (thread const uint8_t *)&qh; + for (int j = 0; j < 4; ++j) { + dst_data[i00/QK5_1].qh[j] = qh8[j]; + } + } +} + +static inline int best_index_int8(int n, constant float * val, float x) { + if (x <= val[0]) return 0; + if (x >= val[n-1]) return n-1; + int ml = 0, mu = n-1; + while (mu-ml > 1) { + int mav = (ml+mu)/2; + if (x < val[mav]) mu = mav; else ml = mav; + } + return x - val[mu-1] < val[mu] - x ? mu-1 : mu; +} + +constexpr constant static float kvalues_iq4nl_f[16] = { + -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f +}; + +kernel void kernel_cpy_f32_iq4_nl( + device const float * src0, + device void * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant int64_t & ne03, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant uint64_t & nb03, + constant int64_t & ne0, + constant int64_t & ne1, + constant int64_t & ne2, + constant int64_t & ne3, + constant uint64_t & nb0, + constant uint64_t & nb1, + constant uint64_t & nb2, + constant uint64_t & nb3, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + const int64_t i03 = tgpig[2]; + const int64_t i02 = tgpig[1]; + const int64_t i01 = tgpig[0]; + + const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; + + const int64_t i3 = n / (ne2*ne1*ne0); + const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0); + const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0; + const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_NL; + + device block_iq4_nl * dst_data = (device block_iq4_nl *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + for (int64_t i00 = tpitg.x*QK4_NL; i00 < ne00; i00 += ntg.x*QK4_NL) { + device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00); + + float amax = 0.0f; // absolute max + float max = 0.0f; + + for (int j = 0; j < QK4_0; j++) { + const float v = src[j]; + if (amax < fabs(v)) { + amax = fabs(v); + max = v; + } + } + + const float d = max / kvalues_iq4nl_f[0]; + const float id = d ? 1.0f/d : 0.0f; + + float sumqx = 0, sumq2 = 0; + for (int j = 0; j < QK4_NL/2; ++j) { + const float x0 = src[0 + j]*id; + const float x1 = src[QK4_NL/2 + j]*id; + + const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0); + const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1); + + dst_data[i00/QK4_NL].qs[j] = xi0 | (xi1 << 4); + + const float v0 = kvalues_iq4nl_f[xi0]; + const float v1 = kvalues_iq4nl_f[xi1]; + const float w0 = src[0 + j]*src[0 + j]; + const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j]; + sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j]; + sumq2 += w0*v0*v0 + w1*v1*v1; + + } + + dst_data[i00/QK4_NL].d = sumq2 > 0 ? sumqx/sumq2 : d; + + } +} + kernel void kernel_concat( device const char * src0, device const char * src1, @@ -4220,10 +4456,6 @@ void kernel_mul_mv_iq1_s_f32_impl( } } -constexpr constant static float kvalues_iq4nl_f[16] = { - -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f -}; - void kernel_mul_mv_iq4_nl_f32_impl( device const void * src0, device const float * src1, @@ -5528,18 +5760,18 @@ typedef void (mat_mm_t)( threadgroup uchar *, uint3, uint, uint); -template [[host_name("kernel_mul_mm_f32_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_f32_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_iq2_xs_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm; @@ -5588,18 +5820,18 @@ typedef void (mat_mm_id_t)( threadgroup uchar *, uint3, uint, uint); -template [[host_name("kernel_mul_mm_id_f32_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_f16_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q4_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q4_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q5_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q5_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q8_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q2_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q3_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q4_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_f32_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_f16_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q4_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q4_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q5_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q5_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q8_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q2_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q3_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q4_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; diff --git a/ggml-quants.c b/ggml-quants.c index 109dd6660..f26798acc 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -132,7 +132,7 @@ static inline __m256 sum_i16_pairs_float(const __m256i x) { } static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) { -#if __AVXVNNI__ +#if defined(__AVXVNNI__) || defined(__AVX512VNNI__) const __m256i zero = _mm256_setzero_si256(); const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy); return _mm256_cvtepi32_ps(summed_pairs); @@ -11705,9 +11705,8 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block ggml_fp16_t * dh, uint8_t * q4, uint16_t * scales_h, uint8_t * scales_l, float * scales, float * weight, uint8_t * L, const int8_t * values, - const float * quant_weights) { - - const int ntry = 7; + const float * quant_weights, + const int ntry) { float sigma2 = 0; for (int j = 0; j < super_block_size; ++j) sigma2 += x[j]*x[j]; @@ -11719,6 +11718,7 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block float max_scale = 0, amax_scale = 0; for (int ib = 0; ib < super_block_size/block_size; ++ib) { const float * xb = x + ib*block_size; + uint8_t * Lb = L + ib*block_size; if (quant_weights) { const float * qw = quant_weights + ib*block_size; for (int j = 0; j < block_size; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]); @@ -11736,12 +11736,13 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block scales[ib] = 0; continue; } - float d = -max/values[0]; + float d = ntry > 0 ? -max/values[0] : max/values[0]; float id = 1/d; float sumqx = 0, sumq2 = 0; for (int j = 0; j < block_size; ++j) { float al = id*xb[j]; int l = best_index_int8(16, values, al); + Lb[j] = l; float q = values[l]; float w = weight[j]; sumqx += w*q*xb[j]; @@ -11796,9 +11797,11 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block } } else { dh[0] = GGML_FP32_TO_FP16(scales[0]); - float id = scales[0] ? 1/scales[0] : 0; - for (int j = 0; j < super_block_size; ++j) { - L[j] = best_index_int8(16, values, id*x[j]); + if (ntry > 0) { + float id = scales[0] ? 1/scales[0] : 0; + for (int j = 0; j < super_block_size; ++j) { + L[j] = best_index_int8(16, values, id*x[j]); + } } } @@ -11823,7 +11826,7 @@ size_t quantize_iq4_nl(const float * restrict src, void * restrict dst, int nrow for (int ibl = 0; ibl < nblock; ++ibl) { const float * qw = quant_weights ? quant_weights + QK4_NL*ibl : NULL; quantize_row_iq4_nl_impl(QK4_NL, 32, src + QK4_NL*ibl, &iq4[ibl].d, iq4[ibl].qs, &unused_h, unused_l, - &scale, weight, L, kvalues_iq4nl, qw); + &scale, weight, L, kvalues_iq4nl, qw, 7); } src += n_per_row; qrow += nblock*sizeof(block_iq4_nl); @@ -11832,14 +11835,23 @@ size_t quantize_iq4_nl(const float * restrict src, void * restrict dst, int nrow } void quantize_row_iq4_nl(const float * restrict x, void * restrict vy, int k) { - assert(k % QK4_NL == 0); - block_iq4_nl * restrict y = vy; - quantize_row_iq4_nl_reference(x, y, k); + GGML_ASSERT(k%QK4_NL == 0); + int nblock = k/QK4_NL; + uint8_t L[QK4_NL]; + float weight[QK4_NL]; + uint16_t unused_h; + uint8_t * unused_l = NULL; + float scale; + block_iq4_nl * iq4 = (block_iq4_nl *)vy; + for (int ibl = 0; ibl < nblock; ++ibl) { + quantize_row_iq4_nl_impl(QK4_NL, 32, x + QK4_NL*ibl, &iq4[ibl].d, iq4[ibl].qs, &unused_h, unused_l, + &scale, weight, L, kvalues_iq4nl, NULL, -1); + } } void quantize_row_iq4_nl_reference(const float * restrict x, block_iq4_nl * restrict y, int k) { assert(k % QK4_NL == 0); - quantize_iq4_nl(x, y, 1, k, NULL); + quantize_row_iq4_nl(x, y, k); } size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { @@ -11857,7 +11869,7 @@ size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int nrow for (int ibl = 0; ibl < nblock; ++ibl) { const float * qw = quant_weights ? quant_weights + QK_K*ibl : NULL; quantize_row_iq4_nl_impl(QK_K, 32, src + QK_K*ibl, &iq4[ibl].d, iq4[ibl].qs, &iq4[ibl].scales_h, iq4[ibl].scales_l, - scales, weight, L, kvalues_iq4nl, qw); + scales, weight, L, kvalues_iq4nl, qw, 7); } src += n_per_row; qrow += nblock*sizeof(block_iq4_xs); diff --git a/ggml-sycl.cpp b/ggml-sycl.cpp index d51f23b41..fc4d2964c 100644 --- a/ggml-sycl.cpp +++ b/ggml-sycl.cpp @@ -740,11 +740,7 @@ namespace dpct sycl::queue &default_queue() { -#ifdef DPCT_USM_LEVEL_NONE - return out_of_order_queue(); -#else return in_order_queue(); -#endif // DPCT_USM_LEVEL_NONE } void queues_wait_and_throw() @@ -763,11 +759,7 @@ namespace dpct sycl::queue *create_queue(bool enable_exception_handler = false) { -#ifdef DPCT_USM_LEVEL_NONE - return create_out_of_order_queue(enable_exception_handler); -#else return create_in_order_queue(enable_exception_handler); -#endif // DPCT_USM_LEVEL_NONE } sycl::queue *create_queue(sycl::context context, sycl::device device, @@ -977,8 +969,10 @@ namespace dpct static int convert_backend_index(std::string & backend) { if (backend == "ext_oneapi_level_zero:gpu") return 0; if (backend == "opencl:gpu") return 1; - if (backend == "opencl:cpu") return 2; - if (backend == "opencl:acc") return 3; + if (backend == "ext_oneapi_cuda:gpu") return 2; + if (backend == "ext_oneapi_hip:gpu") return 3; + if (backend == "opencl:cpu") return 4; + if (backend == "opencl:acc") return 5; printf("convert_backend_index: can't handle backend=%s\n", backend.c_str()); GGML_ASSERT(false); } @@ -1073,11 +1067,6 @@ namespace dpct static pointer_access_attribute get_pointer_attribute(sycl::queue &q, const void *ptr) { -#ifdef DPCT_USM_LEVEL_NONE - return mem_mgr::instance().is_device_ptr(ptr) - ? pointer_access_attribute::device_only - : pointer_access_attribute::host_only; -#else switch (sycl::get_pointer_type(ptr, q.get_context())) { case sycl::usm::alloc::unknown: @@ -1088,7 +1077,6 @@ namespace dpct case sycl::usm::alloc::host: return pointer_access_attribute::host_device; } -#endif } template @@ -1271,11 +1259,7 @@ namespace dpct static inline void *dpct_malloc(size_t size, sycl::queue &q) { -#ifdef DPCT_USM_LEVEL_NONE - return mem_mgr::instance().mem_alloc(size * sizeof(byte_t)); -#else return sycl::malloc_device(size, q.get_device(), q.get_context()); -#endif // DPCT_USM_LEVEL_NONE } #define PITCH_DEFAULT_ALIGN(x) (((x) + 31) & ~(0x1F)) @@ -1299,25 +1283,7 @@ namespace dpct static inline sycl::event dpct_memset(sycl::queue &q, void *dev_ptr, valueT value, size_t size) { -#ifdef DPCT_USM_LEVEL_NONE - auto &mm = mem_mgr::instance(); - assert(mm.is_device_ptr(dev_ptr)); - auto alloc = mm.translate_ptr(dev_ptr); - size_t offset = (valueT *)dev_ptr - (valueT *)alloc.alloc_ptr; - - return q.submit([&](sycl::handler &cgh) - { - auto r = sycl::range<1>(size); - auto o = sycl::id<1>(offset); - auto new_buffer = alloc.buffer.reinterpret( - sycl::range<1>(alloc.size / sizeof(valueT))); - sycl::accessor - acc(new_buffer, cgh, r, o); - cgh.fill(acc, value); }); -#else return q.fill(dev_ptr, value, size); -#endif // DPCT_USM_LEVEL_NONE } /** @@ -1411,72 +1377,8 @@ namespace dpct { if (!size) return sycl::event{}; -#ifdef DPCT_USM_LEVEL_NONE - auto &mm = mem_mgr::instance(); - auto real_direction = deduce_memcpy_direction(q, to_ptr, from_ptr, direction); - - switch (real_direction) - { - case host_to_host: - return q.submit([&](sycl::handler &cgh) - { - cgh.depends_on(dep_events); - cgh.host_task([=] { std::memcpy(to_ptr, from_ptr, size); }); }); - case host_to_device: - { - auto alloc = mm.translate_ptr(to_ptr); - size_t offset = (byte_t *)to_ptr - alloc.alloc_ptr; - return q.submit([&](sycl::handler &cgh) - { - cgh.depends_on(dep_events); - auto r = sycl::range<1>(size); - auto o = sycl::id<1>(offset); - sycl::accessor - acc(alloc.buffer, cgh, r, o); - cgh.copy(from_ptr, acc); }); - } - case device_to_host: - { - auto alloc = mm.translate_ptr(from_ptr); - size_t offset = (byte_t *)from_ptr - alloc.alloc_ptr; - return q.submit([&](sycl::handler &cgh) - { - cgh.depends_on(dep_events); - auto r = sycl::range<1>(size); - auto o = sycl::id<1>(offset); - sycl::accessor - acc(alloc.buffer, cgh, r, o); - cgh.copy(acc, to_ptr); }); - } - case device_to_device: - { - auto to_alloc = mm.translate_ptr(to_ptr); - auto from_alloc = mm.translate_ptr(from_ptr); - size_t to_offset = (byte_t *)to_ptr - to_alloc.alloc_ptr; - size_t from_offset = (byte_t *)from_ptr - from_alloc.alloc_ptr; - return q.submit([&](sycl::handler &cgh) - { - cgh.depends_on(dep_events); - auto r = sycl::range<1>(size); - auto to_o = sycl::id<1>(to_offset); - auto from_o = sycl::id<1>(from_offset); - sycl::accessor - to_acc(to_alloc.buffer, cgh, r, to_o); - sycl::accessor - from_acc(from_alloc.buffer, cgh, r, from_o); - cgh.copy(from_acc, to_acc); }); - } - default: - throw std::runtime_error("dpct_memcpy: invalid direction value"); - } -#else return q.memcpy(to_ptr, from_ptr, size, dep_events); GGML_UNUSED(direction); -#endif // DPCT_USM_LEVEL_NONE } // Get actual copy range and make sure it will not exceed range. @@ -1616,45 +1518,15 @@ namespace dpct break; } case device_to_device: -#ifdef DPCT_USM_LEVEL_NONE - { - auto &mm = mem_mgr::instance(); - auto to_alloc = mm.translate_ptr(to_surface); - auto from_alloc = mm.translate_ptr(from_surface); - size_t to_offset = (byte_t *)to_surface - to_alloc.alloc_ptr; - size_t from_offset = (byte_t *)from_surface - from_alloc.alloc_ptr; - event_list.push_back(q.submit([&](sycl::handler &cgh) - { - cgh.depends_on(dep_events); - auto to_o = sycl::id<1>(to_offset); - auto from_o = sycl::id<1>(from_offset); - sycl::accessor - to_acc(to_alloc.buffer, cgh, - get_copy_range(size, to_slice, to_range.get(0)), to_o); - sycl::accessor - from_acc(from_alloc.buffer, cgh, - get_copy_range(size, from_slice, from_range.get(0)), from_o); - cgh.parallel_for( - size, - [=](sycl::id<3> id) { - to_acc[get_offset(id, to_slice, to_range.get(0))] = - from_acc[get_offset(id, from_slice, from_range.get(0))]; - }); })); - } -#else - event_list.push_back(q.submit([&](sycl::handler &cgh) - { - cgh.depends_on(dep_events); - cgh.parallel_for( - size, - [=](sycl::id<3> id) { - to_surface[get_offset(id, to_slice, to_range.get(0))] = - from_surface[get_offset(id, from_slice, from_range.get(0))]; - }); })); -#endif - break; + event_list.push_back(q.submit([&](sycl::handler &cgh){ + cgh.depends_on(dep_events); + cgh.parallel_for( + size, + [=](sycl::id<3> id) { + to_surface[get_offset(id, to_slice, to_range.get(0))] = + from_surface[get_offset(id, from_slice, from_range.get(0))]; + }); })); + break; default: throw std::runtime_error("dpct_memcpy: invalid direction value"); } @@ -1752,11 +1624,7 @@ namespace dpct { if (ptr) { -#ifdef DPCT_USM_LEVEL_NONE - detail::mem_mgr::instance().mem_free(ptr); -#else sycl::free(ptr, q.get_context()); -#endif // DPCT_USM_LEVEL_NONE } } @@ -1764,11 +1632,7 @@ namespace dpct inline auto get_memory(const void *x) { T *new_x = reinterpret_cast(const_cast(x)); -#ifdef DPCT_USM_LEVEL_NONE - return dpct::get_buffer>(new_x); -#else return new_x; -#endif } template @@ -2220,72 +2084,8 @@ namespace dpct { if (!size) return sycl::event{}; -#ifdef DPCT_USM_LEVEL_NONE - auto &mm = mem_mgr::instance(); - auto real_direction = deduce_memcpy_direction(q, to_ptr, from_ptr, direction); - - switch (real_direction) - { - case host_to_host: - return q.submit([&](sycl::handler &cgh) - { - cgh.depends_on(dep_events); - cgh.host_task([=] { std::memcpy(to_ptr, from_ptr, size); }); }); - case host_to_device: - { - auto alloc = mm.translate_ptr(to_ptr); - size_t offset = (byte_t *)to_ptr - alloc.alloc_ptr; - return q.submit([&](sycl::handler &cgh) - { - cgh.depends_on(dep_events); - auto r = sycl::range<1>(size); - auto o = sycl::id<1>(offset); - sycl::accessor - acc(alloc.buffer, cgh, r, o); - cgh.copy(from_ptr, acc); }); - } - case device_to_host: - { - auto alloc = mm.translate_ptr(from_ptr); - size_t offset = (byte_t *)from_ptr - alloc.alloc_ptr; - return q.submit([&](sycl::handler &cgh) - { - cgh.depends_on(dep_events); - auto r = sycl::range<1>(size); - auto o = sycl::id<1>(offset); - sycl::accessor - acc(alloc.buffer, cgh, r, o); - cgh.copy(acc, to_ptr); }); - } - case device_to_device: - { - auto to_alloc = mm.translate_ptr(to_ptr); - auto from_alloc = mm.translate_ptr(from_ptr); - size_t to_offset = (byte_t *)to_ptr - to_alloc.alloc_ptr; - size_t from_offset = (byte_t *)from_ptr - from_alloc.alloc_ptr; - return q.submit([&](sycl::handler &cgh) - { - cgh.depends_on(dep_events); - auto r = sycl::range<1>(size); - auto to_o = sycl::id<1>(to_offset); - auto from_o = sycl::id<1>(from_offset); - sycl::accessor - to_acc(to_alloc.buffer, cgh, r, to_o); - sycl::accessor - from_acc(from_alloc.buffer, cgh, r, from_o); - cgh.copy(from_acc, to_acc); }); - } - default: - throw std::runtime_error("dpct_memcpy: invalid direction value"); - } -#else return q.memcpy(to_ptr, from_ptr, size, dep_events); GGML_UNUSED(direction); -#endif // DPCT_USM_LEVEL_NONE } // Get actual copy range and make sure it will not exceed range. @@ -2425,34 +2225,6 @@ namespace dpct break; } case device_to_device: -#ifdef DPCT_USM_LEVEL_NONE - { - auto &mm = mem_mgr::instance(); - auto to_alloc = mm.translate_ptr(to_surface); - auto from_alloc = mm.translate_ptr(from_surface); - size_t to_offset = (byte_t *)to_surface - to_alloc.alloc_ptr; - size_t from_offset = (byte_t *)from_surface - from_alloc.alloc_ptr; - event_list.push_back(q.submit([&](sycl::handler &cgh) - { - cgh.depends_on(dep_events); - auto to_o = sycl::id<1>(to_offset); - auto from_o = sycl::id<1>(from_offset); - sycl::accessor - to_acc(to_alloc.buffer, cgh, - get_copy_range(size, to_slice, to_range.get(0)), to_o); - sycl::accessor - from_acc(from_alloc.buffer, cgh, - get_copy_range(size, from_slice, from_range.get(0)), from_o); - cgh.parallel_for( - size, - [=](sycl::id<3> id) { - to_acc[get_offset(id, to_slice, to_range.get(0))] = - from_acc[get_offset(id, from_slice, from_range.get(0))]; - }); })); - } -#else event_list.push_back(q.submit([&](sycl::handler &cgh) { cgh.depends_on(dep_events); @@ -2462,7 +2234,6 @@ namespace dpct to_surface[get_offset(id, to_slice, to_range.get(0))] = from_surface[get_offset(id, from_slice, from_range.get(0))]; }); })); -#endif break; default: throw std::runtime_error("dpct_memcpy: invalid direction value"); @@ -2653,9 +2424,6 @@ namespace dpct void *c[], library_data_t c_type, int ldc, int batch_size, library_data_t scaling_type) { -#ifdef DPCT_USM_LEVEL_NONE - throw std::runtime_error("this API is unsupported when USM level is none"); -#else if (scaling_type == library_data_t::real_float && c_type == library_data_t::complex_float) { @@ -2790,7 +2558,6 @@ namespace dpct default: throw std::runtime_error("the combination of data type is unsupported"); } -#endif } /// Computes a batch of matrix-matrix product with general matrices. @@ -3129,24 +2896,9 @@ namespace dpct template typename std::enable_if::type &operator[](size_t index) { init(); - #ifdef DPCT_USM_LEVEL_NONE - return dpct::get_buffer::type>( - _device_ptr) - .template get_access()[index]; - #else return _device_ptr[index]; - #endif // DPCT_USM_LEVEL_NONE } - #ifdef DPCT_USM_LEVEL_NONE - /// Get sycl::accessor for the device memory object when usm is not used. - accessor_t get_access(sycl::handler &cgh) { - return get_buffer(_device_ptr) - .template reinterpret(_range) - .template get_access::mode, - detail::memory_traits::target>(cgh); - } - #else /// Get dpct::accessor with dimension info for the device memory object /// when usm is used and dimension is greater than 1. template @@ -3154,7 +2906,6 @@ namespace dpct get_access(sycl::handler &cgh) { return dpct_accessor_t((T *)_device_ptr, _range); } - #endif // DPCT_USM_LEVEL_NONE private: device_memory(value_t *memory_ptr, size_t size) @@ -3199,15 +2950,6 @@ namespace dpct /// Default constructor device_memory() : base(1) {} - - #ifdef DPCT_USM_LEVEL_NONE - /// Get sycl::accessor for the device memory object when usm is not used. - accessor_t get_access(sycl::handler &cgh) { - auto buf = get_buffer(base::get_ptr()) - .template reinterpret(sycl::range<1>(1)); - return accessor_t(buf, cgh); - } - #endif // DPCT_USM_LEVEL_NONE }; } // namespace detail @@ -13179,7 +12921,7 @@ int get_work_group_size(int user_device_id) { return prop.get_max_work_group_size(); } -void ggml_init_sycl() try { +static void ggml_init_sycl() try { static bool initialized = false; if (!initialized) { @@ -16675,6 +16417,7 @@ static ggml_backend_buffer_type_i ggml_backend_sycl_buffer_type_interface = { }; ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device_index) { + ggml_init_sycl(); if (device_index>=g_device_count or device_index<0) { printf("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n", device_index, g_device_count-1); @@ -17044,6 +16787,7 @@ static ggml_backend_buffer_type_i ggml_backend_sycl_split_buffer_type_interface }; GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split) { + ggml_init_sycl(); // FIXME: this is not thread safe static std::map, struct ggml_backend_buffer_type> buft_map; @@ -17377,6 +17121,13 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons UNUSED(backend); } +GGML_CALL static bool ggml_backend_sycl_offload_op(ggml_backend_t backend, const ggml_tensor * op) { + const int min_batch_size = 32; + return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS; + GGML_UNUSED(backend); +} + + static ggml_backend_i ggml_backend_sycl_interface = { /* .get_name = */ ggml_backend_sycl_name, /* .free = */ ggml_backend_sycl_free, @@ -17390,7 +17141,7 @@ static ggml_backend_i ggml_backend_sycl_interface = { /* .graph_plan_compute = */ NULL, /* .graph_compute = */ ggml_backend_sycl_graph_compute, /* .supports_op = */ ggml_backend_sycl_supports_op, - /* .offload_op = */ NULL, + /* .offload_op = */ ggml_backend_sycl_offload_op, /* .event_new = */ NULL, /* .event_free = */ NULL, /* .event_record = */ NULL, @@ -17404,7 +17155,7 @@ static ggml_guid_t ggml_backend_sycl_guid() { } GGML_CALL ggml_backend_t ggml_backend_sycl_init(int device) { - ggml_init_sycl(); // TODO: remove from ggml.c + ggml_init_sycl(); check_allow_gpu_index(device); diff --git a/ggml-sycl.h b/ggml-sycl.h index 1c9d52115..a9f776fc1 100644 --- a/ggml-sycl.h +++ b/ggml-sycl.h @@ -16,16 +16,22 @@ extern "C" { #define GGML_SYCL_MAX_DEVICES 48 #define GGML_SYCL_NAME "SYCL" -GGML_API void ggml_init_sycl(void); -GGML_API bool ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor); +// backend API GGML_API ggml_backend_t ggml_backend_sycl_init(int device); + +// devide buffer GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device); + +// split tensor buffer that splits matrices by rows across multiple devices +GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split); + +// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void); + GGML_API void ggml_backend_sycl_print_sycl_devices(void); GGML_API GGML_CALL void ggml_sycl_get_gpu_list(int *id_list, int max_len); GGML_API GGML_CALL void ggml_sycl_get_device_description(int device, char *description, size_t description_size); GGML_API GGML_CALL int ggml_backend_sycl_get_device_count(); -GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split); GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total); GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id); @@ -34,6 +40,10 @@ GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id); GGML_API GGML_CALL int ggml_backend_sycl_get_device_id(int device_index); GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id); GGML_API GGML_CALL void ggml_backend_sycl_set_mul_device_mode(); + +// SYCL doesn't support registering host memory, keep here for reference +// GGML_API GGML_CALL bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size); +// GGML_API GGML_CALL void ggml_backend_sycl_unregister_host_buffer(void * buffer); #ifdef __cplusplus } #endif diff --git a/ggml.c b/ggml.c index 1d5854960..203a9e540 100644 --- a/ggml.c +++ b/ggml.c @@ -3,6 +3,7 @@ #include "ggml-impl.h" #include "ggml-quants.h" +#include "ggml.h" #if defined(_MSC_VER) || defined(__MINGW32__) #include // using malloc.h with MSC/MINGW @@ -43,6 +44,10 @@ #if defined(_WIN32) +#define WIN32_LEAN_AND_MEAN +#ifndef NOMINMAX + #define NOMINMAX +#endif #include typedef volatile LONG atomic_int; @@ -286,8 +291,6 @@ inline static void * ggml_calloc(size_t num, size_t size) { #include "ggml-opencl.h" #elif defined(GGML_USE_VULKAN) #include "ggml-vulkan.h" -#elif defined(GGML_USE_SYCL) -#include "ggml-sycl.h" #endif // floating point type used to accumulate sums @@ -430,6 +433,57 @@ int64_t ggml_cycles_per_ms(void) { #define ggml_perf_cycles_per_ms() 0 #endif +// +// cross-platform UTF-8 file paths +// + +#ifdef _WIN32 +static wchar_t * ggml_mbstowcs(const char * mbs) { + int wlen = MultiByteToWideChar(CP_UTF8, 0, mbs, -1, NULL, 0); + if (!wlen) { + errno = EINVAL; + return NULL; + } + + wchar_t * wbuf = GGML_MALLOC(wlen * sizeof(wchar_t)); + wlen = MultiByteToWideChar(CP_UTF8, 0, mbs, -1, wbuf, wlen); + if (!wlen) { + GGML_FREE(wbuf); + errno = EINVAL; + return NULL; + } + + return wbuf; +} +#endif + +FILE * ggml_fopen(const char * fname, const char * mode) { +#ifdef _WIN32 + FILE * file = NULL; + + // convert fname (UTF-8) + wchar_t * wfname = ggml_mbstowcs(fname); + if (wfname) { + // convert mode (ANSI) + wchar_t * wmode = GGML_MALLOC((strlen(mode) + 1) * sizeof(wchar_t)); + wchar_t * wmode_p = wmode; + do { + *wmode_p++ = (wchar_t)*mode; + } while (*mode++); + + // open file + file = _wfopen(wfname, wmode); + + GGML_FREE(wfname); + GGML_FREE(wmode); + } + + return file; +#else + return fopen(fname, mode); +#endif +} + // // cache line // @@ -2642,8 +2696,6 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { ggml_cl_init(); #elif defined(GGML_USE_VULKAN) ggml_vk_init_cpu_assist(); -#elif defined(GGML_USE_SYCL) - ggml_init_sycl(); #endif ggml_setup_op_has_task_pass(); @@ -16059,12 +16111,6 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_TYPE_CPU); #endif // GGML_USE_VULKAN -#ifdef GGML_USE_SYCL - bool skip_cpu = ggml_sycl_compute_forward(params, tensor); - if (skip_cpu) { - return; - } -#endif // GGML_USE_SYCL switch (tensor->op) { case GGML_OP_DUP: { @@ -18739,7 +18785,7 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) { // write binary data { - FILE * fout = fopen(fname, "wb"); + FILE * fout = ggml_fopen(fname, "wb"); if (!fout) { fprintf(stderr, "%s: failed to open %s\n", __func__, fname); @@ -18877,7 +18923,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context * // read file into data { - FILE * fin = fopen(fname, "rb"); + FILE * fin = ggml_fopen(fname, "rb"); if (!fin) { fprintf(stderr, "%s: failed to open %s\n", __func__, fname); return result; @@ -19213,7 +19259,7 @@ static void ggml_graph_dump_dot_leaf_edge(FILE * fp, struct ggml_tensor * node, void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename) { char color[16]; - FILE * fp = fopen(filename, "w"); + FILE * fp = ggml_fopen(filename, "w"); GGML_ASSERT(fp); fprintf(fp, "digraph G {\n"); @@ -20531,7 +20577,7 @@ struct gguf_context * gguf_init_empty(void) { } struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params) { - FILE * file = fopen(fname, "rb"); + FILE * file = ggml_fopen(fname, "rb"); if (!file) { return NULL; } @@ -21486,7 +21532,7 @@ static void gguf_write_to_buf(const struct gguf_context * ctx, struct gguf_buf * } void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta) { - FILE * file = fopen(fname, "wb"); + FILE * file = ggml_fopen(fname, "wb"); if (!file) { GGML_ASSERT(false && "failed to open file for writing"); } diff --git a/ggml.h b/ggml.h index c937d4a53..0a5af7205 100644 --- a/ggml.h +++ b/ggml.h @@ -214,9 +214,10 @@ # define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__))) #endif -#include -#include #include +#include +#include +#include #define GGML_FILE_MAGIC 0x67676d6c // "ggml" #define GGML_FILE_VERSION 1 @@ -708,6 +709,9 @@ extern "C" { GGML_API void ggml_print_backtrace(void); + // accepts a UTF-8 path, even on Windows + GGML_API FILE * ggml_fopen(const char * fname, const char * mode); + GGML_API void ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems GGML_API bool ggml_is_numa(void); // true if init detected that system has >1 NUMA node diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 4a4facb06..e47896e2a 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -100,6 +100,7 @@ class MODEL_ARCH(IntEnum): LLAMA = auto() FALCON = auto() BAICHUAN = auto() + GROK = auto() GPT2 = auto() GPTJ = auto() GPTNEOX = auto() @@ -167,6 +168,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.LLAMA: "llama", MODEL_ARCH.FALCON: "falcon", MODEL_ARCH.BAICHUAN: "baichuan", + MODEL_ARCH.GROK: "grok", MODEL_ARCH.GPT2: "gpt2", MODEL_ARCH.GPTJ: "gptj", MODEL_ARCH.GPTNEOX: "gptneox", @@ -251,6 +253,28 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN_EXP, MODEL_TENSOR.FFN_UP_EXP, ], + MODEL_ARCH.GROK: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.ATTN_OUT_NORM, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.LAYER_OUT_NORM, + ], MODEL_ARCH.GPTNEOX: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index ed89955d8..11fd34b8b 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -23,6 +23,7 @@ class TensorNameMap: "model.embedding", # mamba-qbert "backbone.embedding", # mamba "backbone.embeddings", # mamba-hf + "transformer.in_out_embed", # Grok ), # Token type embeddings @@ -66,6 +67,7 @@ class TensorNameMap: "lm_head.ln", # phi2 "model.norm_f", # mamba-qbert "backbone.norm_f", # mamba + "transformer.rms_norm", # Grok ), # Rope frequencies @@ -93,6 +95,7 @@ class TensorNameMap: "model.layers.{bid}.attention_norm", # internlm2 "model.layers.{bid}.norm", # mamba-qbert "backbone.layers.{bid}.norm", # mamba + "transformer.decoder_layer.{bid}.rms_norm", # Grok ), # Attention norm 2 @@ -116,32 +119,35 @@ class TensorNameMap: # Attention query MODEL_TENSOR.ATTN_Q: ( - "model.layers.{bid}.self_attn.q_proj", # llama-hf - "layers.{bid}.attention.wq", # llama-pth - "encoder.layer.{bid}.attention.self.query", # bert - "transformer.h.{bid}.attn.q_proj", # gpt-j - "model.layers.layers.{bid}.self_attn.q_proj", # plamo - "model.layers.{bid}.attention.wq" # internlm2 + "model.layers.{bid}.self_attn.q_proj", # llama-hf + "layers.{bid}.attention.wq", # llama-pth + "encoder.layer.{bid}.attention.self.query", # bert + "transformer.h.{bid}.attn.q_proj", # gpt-j + "model.layers.layers.{bid}.self_attn.q_proj", # plamo + "model.layers.{bid}.attention.wq", # internlm2 + "transformer.decoder_layer.{bid}.multi_head_attention.query" # Grok ), # Attention key MODEL_TENSOR.ATTN_K: ( - "model.layers.{bid}.self_attn.k_proj", # llama-hf - "layers.{bid}.attention.wk", # llama-pth - "encoder.layer.{bid}.attention.self.key", # bert - "transformer.h.{bid}.attn.k_proj", # gpt-j - "model.layers.layers.{bid}.self_attn.k_proj", # plamo - "model.layers.{bid}.attention.wk" # internlm2 + "model.layers.{bid}.self_attn.k_proj", # llama-hf + "layers.{bid}.attention.wk", # llama-pth + "encoder.layer.{bid}.attention.self.key", # bert + "transformer.h.{bid}.attn.k_proj", # gpt-j + "model.layers.layers.{bid}.self_attn.k_proj", # plamo + "model.layers.{bid}.attention.wk", # internlm2 + "transformer.decoder_layer.{bid}.multi_head_attention.key" # Grok ), # Attention value MODEL_TENSOR.ATTN_V: ( - "model.layers.{bid}.self_attn.v_proj", # llama-hf - "layers.{bid}.attention.wv", # llama-pth - "encoder.layer.{bid}.attention.self.value", # bert - "transformer.h.{bid}.attn.v_proj", # gpt-j - "model.layers.layers.{bid}.self_attn.v_proj", # plamo - "model.layers.{bid}.attention.wv" # internlm2 + "model.layers.{bid}.self_attn.v_proj", # llama-hf + "layers.{bid}.attention.wv", # llama-pth + "encoder.layer.{bid}.attention.self.value", # bert + "transformer.h.{bid}.attn.v_proj", # gpt-j + "model.layers.layers.{bid}.self_attn.v_proj", # plamo + "model.layers.{bid}.attention.wv", # internlm2 + "transformer.decoder_layer.{bid}.multi_head_attention.value" # Grok ), # Attention output @@ -162,12 +168,14 @@ class TensorNameMap: "model.layers.layers.{bid}.self_attn.o_proj", # plamo "model.layers.{bid}.attention.wo", # internlm2 "encoder.layers.{bid}.attn.out_proj", # nomic-bert + "transformer.decoder_layer.{bid}.multi_head_attention.linear"# Grok ), # Attention output norm MODEL_TENSOR.ATTN_OUT_NORM: ( "encoder.layer.{bid}.attention.output.LayerNorm", # bert "encoder.layers.{bid}.norm1", # nomic-bert + "transformer.decoder_layer.{bid}.rms_norm_1", # Grok ), # Rotary embeddings @@ -190,11 +198,13 @@ class TensorNameMap: "model.layers.{bid}.ln2", # yi "h.{bid}.ln_2", # gpt2 "model.layers.{bid}.ffn_norm", # internlm2 + "transformer.decoder_layer.{bid}.rms_norm_2", # Grok ), MODEL_TENSOR.FFN_GATE_INP: ( "layers.{bid}.feed_forward.gate", # mixtral "model.layers.{bid}.block_sparse_moe.gate", # mixtral + "transformer.decoder_layer.{bid}.router" # Grok ), # Feed-forward up @@ -223,6 +233,7 @@ class TensorNameMap: MODEL_TENSOR.FFN_UP_EXP: ( "layers.{bid}.feed_forward.experts.{xid}.w3", # mixtral "model.layers.{bid}.block_sparse_moe.experts.{xid}.w3", # mixtral + "transformer.decoder_layer.{bid}.moe.{xid}.linear_v", # Grok ), # AWQ-activation gate @@ -243,6 +254,7 @@ class TensorNameMap: MODEL_TENSOR.FFN_GATE_EXP: ( "layers.{bid}.feed_forward.experts.{xid}.w1", # mixtral "model.layers.{bid}.block_sparse_moe.experts.{xid}.w1", # mixtral + "transformer.decoder_layer.{bid}.moe.{xid}.linear" # Grok ), # Feed-forward down @@ -270,6 +282,8 @@ class TensorNameMap: MODEL_TENSOR.FFN_DOWN_EXP: ( "layers.{bid}.feed_forward.experts.{xid}.w2", # mixtral "model.layers.{bid}.block_sparse_moe.experts.{xid}.w2", # mixtral + "transformer.decoder_layer.{bid}.moe.{xid}.linear_1", # Grok + ), MODEL_TENSOR.ATTN_Q_NORM: ( @@ -287,8 +301,9 @@ class TensorNameMap: ), MODEL_TENSOR.LAYER_OUT_NORM: ( - "encoder.layer.{bid}.output.LayerNorm", # bert - "encoder.layers.{bid}.norm2", # nomic-bert + "encoder.layer.{bid}.output.LayerNorm", # bert + "encoder.layers.{bid}.norm2", # nomic-bert + "transformer.decoder_layer.{bid}.rms_norm_3", # Grok ), MODEL_TENSOR.SSM_IN: ( diff --git a/llama.cpp b/llama.cpp index 1a9fe0c4d..61587cb7a 100644 --- a/llama.cpp +++ b/llama.cpp @@ -52,6 +52,9 @@ #define NOMINMAX #endif #include + #ifndef PATH_MAX + #define PATH_MAX MAX_PATH + #endif #include #endif @@ -192,6 +195,7 @@ enum llm_arch { LLM_ARCH_LLAMA, LLM_ARCH_FALCON, LLM_ARCH_BAICHUAN, + LLM_ARCH_GROK, LLM_ARCH_GPT2, LLM_ARCH_GPTJ, LLM_ARCH_GPTNEOX, @@ -221,6 +225,7 @@ enum llm_arch { static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_LLAMA, "llama" }, { LLM_ARCH_FALCON, "falcon" }, + { LLM_ARCH_GROK, "grok" }, { LLM_ARCH_GPT2, "gpt2" }, { LLM_ARCH_GPTJ, "gptj" }, { LLM_ARCH_GPTNEOX, "gptneox" }, @@ -290,6 +295,10 @@ enum llm_kv { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, LLM_KV_ROPE_SCALING_FINETUNED, + LLM_KV_SPLIT_NO, + LLM_KV_SPLIT_COUNT, + LLM_KV_SPLIT_TENSORS_COUNT, + LLM_KV_SSM_INNER_SIZE, LLM_KV_SSM_CONV_KERNEL, LLM_KV_SSM_STATE_SIZE, @@ -355,6 +364,10 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" }, { LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" }, + { LLM_KV_SPLIT_NO, "split.no" }, + { LLM_KV_SPLIT_COUNT, "split.count" }, + { LLM_KV_SPLIT_TENSORS_COUNT, "split.tensors.count" }, + { LLM_KV_SSM_CONV_KERNEL, "%s.ssm.conv_kernel" }, { LLM_KV_SSM_INNER_SIZE, "%s.ssm.inner_size" }, { LLM_KV_SSM_STATE_SIZE, "%s.ssm.state_size" }, @@ -483,6 +496,28 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_GROK, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_EXP, "blk.%d.ffn_gate.%d" }, + { LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" }, + { LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" }, + { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" }, + { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" }, + }, + }, { LLM_ARCH_GPT2, { @@ -1030,7 +1065,7 @@ struct llama_file { size_t size; llama_file(const char * fname, const char * mode) { - fp = std::fopen(fname, mode); + fp = ggml_fopen(fname, mode); if (fp == NULL) { throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno))); } @@ -1099,6 +1134,7 @@ struct llama_file { } } }; +using llama_files = std::vector>; struct llama_mmap { void * addr; @@ -1299,6 +1335,7 @@ struct llama_mmap { } #endif }; +using llama_mmaps = std::vector>; // Represents some region of memory being locked using mlock or VirtualLock; // will automatically unlock on destruction. @@ -1448,6 +1485,7 @@ struct llama_mlock { static void raw_unlock(const void * addr, size_t len) {} #endif }; +using llama_mlocks = std::vector>; static std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) { std::vector result(8, 0); @@ -1621,6 +1659,7 @@ enum e_model { MODEL_40B, MODEL_65B, MODEL_70B, + MODEL_314B, MODEL_SMALL, MODEL_MEDIUM, MODEL_LARGE, @@ -2023,12 +2062,12 @@ struct llama_model { // the model memory buffers for the tensor data std::vector bufs; - // model memory mapped file - std::unique_ptr mapping; + // model memory mapped files + llama_mmaps mappings; // objects representing data potentially being locked in memory - std::vector> mlock_bufs; - llama_mlock mlock_mmap; + llama_mlocks mlock_bufs; + llama_mlocks mlock_mmaps; // for quantize-stats only std::vector> tensors_by_name; @@ -2792,6 +2831,8 @@ namespace GGUFMeta { }; } +using llama_buf_map = std::unordered_map; + struct llama_model_loader { int n_kv = 0; int n_tensors = 0; @@ -2802,54 +2843,133 @@ struct llama_model_loader { bool use_mmap = false; - llama_file file; + llama_files files; llama_ftype ftype; llama_fver fver; - std::unique_ptr mapping; + llama_mmaps mappings; + + // Holds information on a model weights + struct llama_tensor_weights { + uint16_t idx; // source file index + size_t offs; // tensor data offset in the original file + + ggml_tensor * tensor; + + llama_tensor_weights(uint16_t idx, const char * name, const struct gguf_context * gguf_ctx, ggml_tensor * tensor) : idx(idx), tensor(tensor) { + const int tensor_idx = gguf_find_tensor(gguf_ctx, name); + offs = gguf_get_data_offset(gguf_ctx) + gguf_get_tensor_offset(gguf_ctx, tensor_idx); + } + }; + std::vector weights; + std::unordered_map kv_overrides; - struct gguf_context * ctx_gguf = NULL; - struct ggml_context * ctx_meta = NULL; + struct gguf_context * meta = NULL; + std::vector contexts; std::string arch_name; LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN); - llama_model_loader(const std::string & fname, bool use_mmap, const struct llama_model_kv_override * param_overrides_p) : file(fname.c_str(), "rb") { + llama_model_loader(const std::string & fname, bool use_mmap, const struct llama_model_kv_override * param_overrides_p) { int trace = 0; if (getenv("LLAMA_TRACE")) { trace = atoi(getenv("LLAMA_TRACE")); } - struct gguf_init_params params = { - /*.no_alloc = */ true, - /*.ctx = */ &ctx_meta, - }; - if (param_overrides_p != nullptr) { for (const struct llama_model_kv_override *p = param_overrides_p; p->key[0] != 0; p++) { kv_overrides.insert({std::string(p->key), *p}); } } - ctx_gguf = gguf_init_from_file(fname.c_str(), params); - if (!ctx_gguf) { + struct ggml_context * ctx = NULL; + struct gguf_init_params params = { + /*.no_alloc = */ true, + /*.ctx = */ &ctx, + }; + + meta = gguf_init_from_file(fname.c_str(), params); + if (!meta) { throw std::runtime_error(format("%s: failed to load model from %s\n", __func__, fname.c_str())); } get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false); llm_kv = LLM_KV(llm_arch_from_string(arch_name)); - n_kv = gguf_get_n_kv(ctx_gguf); - n_tensors = gguf_get_n_tensors(ctx_gguf); + // Save tensors data offset of the main file. + // For subsidiary files, `meta` tensor data offset must not be used, + // so we build a unified tensors index for weights. + for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) { + weights.emplace_back(llama_tensor_weights(0, cur->name, meta, cur)); + } + files.emplace_back(new llama_file(fname.c_str(), "rb")); + contexts.emplace_back(ctx); - fver = (enum llama_fver ) gguf_get_version(ctx_gguf); + uint16_t n_split = 0; + get_key(llm_kv(LLM_KV_SPLIT_COUNT), n_split, false); - for (int i = 0; i < n_tensors; i++) { - const char * name = gguf_get_tensor_name(ctx_gguf, i); - struct ggml_tensor * t = ggml_get_tensor(ctx_meta, name); - n_elements += ggml_nelements(t); - n_bytes += ggml_nbytes(t); + // Load additional GGML contexts + if (n_split > 1) { + uint16_t idx = 0; + get_key(llm_kv(LLM_KV_SPLIT_NO), idx); + if (idx != 0) { + throw std::runtime_error(format("illegal split file: %d, model must be loaded with the first split", idx)); + } + + char split_prefix[PATH_MAX] = {0}; + if (!llama_split_prefix(split_prefix, sizeof(split_prefix), fname.c_str(), idx, n_split)) { + throw std::runtime_error(format("invalid split file: %s", fname.c_str())); + } + + if (trace > 0) { + LLAMA_LOG_INFO("%s: loading additional %d GGUFs\n", __func__, n_split); + } + + char split_path[PATH_MAX] = {0}; + for (idx = 1; idx < n_split; idx++) { + llama_split_path(split_path, sizeof(split_path), split_prefix, idx, n_split); + + struct gguf_init_params split_params = { + /*.no_alloc = */ true, + /*.ctx = */ &ctx, + }; + struct gguf_context * ctx_gguf = gguf_init_from_file(split_path, split_params); + if (!ctx_gguf) { + throw std::runtime_error(format("%s: failed to load GGUF split from %s\n", __func__, split_path)); + } + + // Save tensors data offset info of the shard. + for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) { + weights.emplace_back(llama_tensor_weights(idx, cur->name, ctx_gguf, cur)); + } + files.emplace_back(new llama_file(split_path, "rb")); + contexts.emplace_back(ctx); + + gguf_free(ctx_gguf); + } + + get_key(llm_kv(LLM_KV_SPLIT_TENSORS_COUNT), n_tensors); + + // sanity check + { + const int n_tensors_loaded = (int) weights.size(); + if (n_tensors != n_tensors_loaded) { + throw std::runtime_error(format("corrupted model: %d tensors expected but %d found", n_tensors, n_tensors_loaded)); + } + } + + LLAMA_LOG_INFO("%s: additional %d GGUFs metadata loaded.\n", __func__, n_split - 1); + } + + n_kv = gguf_get_n_kv(meta); + n_tensors = weights.size(); + + fver = (enum llama_fver) gguf_get_version(meta); + + for (auto & w : weights) { + n_elements += ggml_nelements(w.tensor); + n_bytes += ggml_nbytes(w.tensor); } LLAMA_LOG_INFO("%s: loaded meta data with %d key-value pairs and %d tensors from %s (version %s)\n", @@ -2864,7 +2984,8 @@ struct llama_model_loader { enum ggml_type type_max = GGML_TYPE_F32; for (int i = 0; i < n_tensors; i++) { - enum ggml_type type = gguf_get_tensor_type(ctx_gguf, i); + const ggml_tensor * tensor = weights.at(i).tensor; + enum ggml_type type = tensor->type; n_type[type]++; @@ -2874,8 +2995,8 @@ struct llama_model_loader { } if (trace > 0) { - struct ggml_tensor * meta = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_gguf, i)); - LLAMA_LOG_INFO("%s: - tensor %4d: %32s %-8s [ %s ]\n", __func__, i, ggml_get_name(meta), ggml_type_name(type), llama_format_tensor_shape(meta).c_str()); + const uint16_t sid = weights.at(i).idx; + LLAMA_LOG_INFO("%s: - tensor %4d, split %2d: %32s %-8s [ %s ]\n", __func__, i, sid, ggml_get_name(tensor), ggml_type_name(type), llama_format_tensor_shape(tensor).c_str()); } } @@ -2911,22 +3032,23 @@ struct llama_model_loader { ftype = (llama_ftype) (ftype | LLAMA_FTYPE_GUESSED); { - const int kid = gguf_find_key(ctx_gguf, "general.file_type"); + const int kid = gguf_find_key(meta, "general.file_type"); if (kid >= 0) { - ftype = (llama_ftype) gguf_get_val_u32(ctx_gguf, kid); + ftype = (llama_ftype) gguf_get_val_u32(meta, kid); } } LLAMA_LOG_INFO("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__); + for (int i = 0; i < n_kv; i++) { - const char * name = gguf_get_key(ctx_gguf, i); - const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); + const char * name = gguf_get_key(meta, i); + const enum gguf_type type = gguf_get_kv_type(meta, i); const std::string type_name = type == GGUF_TYPE_ARRAY - ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(ctx_gguf, i)), gguf_get_arr_n(ctx_gguf, i)) + ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(meta, i)), gguf_get_arr_n(meta, i)) : gguf_type_name(type); - std::string value = gguf_kv_to_str(ctx_gguf, i); + std::string value = gguf_kv_to_str(meta, i); const size_t MAX_VALUE_LEN = 40; if (value.size() > MAX_VALUE_LEN) { value = format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str()); @@ -2955,18 +3077,18 @@ struct llama_model_loader { } ~llama_model_loader() { - if (ctx_gguf) { - gguf_free(ctx_gguf); + if (meta) { + gguf_free(meta); } - if (ctx_meta) { - ggml_free(ctx_meta); + for (auto * ctx : contexts) { + ggml_free(ctx); } } template typename std::enable_if::value, bool>::type get_arr_n(const std::string & key, T & result, const bool required = true) { - const int kid = gguf_find_key(ctx_gguf, key.c_str()); + const int kid = gguf_find_key(meta, key.c_str()); if (kid < 0) { if (required) { @@ -2976,7 +3098,7 @@ struct llama_model_loader { } struct GGUFMeta::ArrayInfo arr_info = - GGUFMeta::GKV::get_kv(ctx_gguf, kid); + GGUFMeta::GKV::get_kv(meta, kid); result = arr_info.length; @@ -2996,7 +3118,7 @@ struct llama_model_loader { const struct llama_model_kv_override * override = it != kv_overrides.end() ? &it->second : nullptr; - const bool found = GGUFMeta::GKV::set(ctx_gguf, key, result, override); + const bool found = GGUFMeta::GKV::set(meta, key, result, override); if (required && !found) { throw std::runtime_error(format("key not found in model: %s", key.c_str())); @@ -3019,20 +3141,33 @@ struct llama_model_loader { } const char * get_tensor_name(int i) const { - return gguf_get_tensor_name(ctx_gguf, i); + return weights.at(i).tensor->name; + } + + const llama_tensor_weights & get_weights(const char * name) const { + for (const auto & weight : weights) { + if (strcmp(name, weight.tensor->name) == 0) { + return weight; + } + } + throw std::runtime_error(format("tensor %s not found", name)); } struct ggml_tensor * get_tensor_meta(const char * name) const { - return ggml_get_tensor(ctx_meta, name); + try { + return get_weights(name).tensor; + } catch (const std::runtime_error & e) { + return NULL; + } } struct ggml_tensor * get_tensor_meta(int i) const { return get_tensor_meta(get_tensor_name(i)); } - struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, struct ggml_tensor * meta) { - struct ggml_tensor * tensor = ggml_dup_tensor(ctx, meta); - ggml_set_name(tensor, ggml_get_name(meta)); + struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, const struct ggml_tensor * cur) { + struct ggml_tensor * tensor = ggml_dup_tensor(ctx, cur); + ggml_set_name(tensor, ggml_get_name(cur)); n_created++; @@ -3040,7 +3175,7 @@ struct llama_model_loader { } struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector & ne, bool required = true) { - struct ggml_tensor * cur = ggml_get_tensor(ctx_meta, name.c_str()); + const struct ggml_tensor * cur = get_tensor_meta(name.c_str()); if (cur == NULL) { if (!required) { @@ -3075,76 +3210,79 @@ struct llama_model_loader { } } - size_t file_offset(const char * name) const { - const int idx = gguf_find_tensor(ctx_gguf, name); - - if (idx < 0) { - throw std::runtime_error(format("%s: tensor '%s' not found in the file", __func__, name)); - } - - return gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, idx); - } - - void init_mapping(bool prefetch = true, llama_mlock * lmlock = nullptr) { - // prefetch the whole file - all the data is needed anyway + void init_mappings(bool prefetch = true, llama_mlocks * mlock_mmaps = nullptr) { if (use_mmap) { - mapping.reset(new llama_mmap(&file, prefetch ? -1 : 0, ggml_is_numa())); + mappings.reserve(files.size()); + mmaps_used.reserve(files.size()); + for (const auto & file : files) { + std::unique_ptr mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, ggml_is_numa())); + mmaps_used.emplace_back(std::make_pair(mapping->size, 0)); + if (mlock_mmaps) { + std::unique_ptr mlock_mmap(new llama_mlock()); + mlock_mmap->init(mapping->addr); + mlock_mmaps->emplace_back(std::move(mlock_mmap)); + } + mappings.emplace_back(std::move(mapping)); + } } // compute the total size of all tensors for progress reporting - for (int i = 0; i < gguf_get_n_tensors(ctx_gguf); i++) { - struct ggml_tensor * cur = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_gguf, i)); - size_data += ggml_nbytes(cur); - } - - if (use_mmap && mapping) { - if (lmlock) { - lmlock->init(mapping->addr); - } - mmap_used_first = mapping->size; + for (auto & w : weights) { + size_data += ggml_nbytes(w.tensor); } } - void get_mapping_range(size_t * first, size_t * last, ggml_context * ctx) const { - GGML_ASSERT(mapping); + void get_mapping_range(size_t * first, size_t * last, void ** addr, int idx, ggml_context * ctx) const { + GGML_ASSERT(!mappings.empty()); + const auto & mapping = mappings.at(idx); *first = mapping->size; *last = 0; + *addr = mapping->addr; for (ggml_tensor * tensor = ggml_get_first_tensor(ctx); tensor; tensor = ggml_get_next_tensor(ctx, tensor)) { - const size_t offs = file_offset(ggml_get_name(tensor)); - *first = std::min(*first, offs); - *last = std::max(*last, offs + ggml_nbytes(tensor)); + const auto & w = get_weights(ggml_get_name(tensor)); + if (w.idx != idx) { + continue; + } + *first = std::min(*first, w.offs); + *last = std::max(*last, w.offs + ggml_nbytes(tensor)); } } // for backwards compatibility, does not support ggml-backend void load_data_for(struct ggml_tensor * cur) const { - const size_t offs = file_offset(ggml_get_name(cur)); + const auto & w = get_weights(ggml_get_name(cur)); - if (use_mmap && mapping) { + if (use_mmap) { + const auto & mapping = mappings.at(w.idx); if (cur->data == nullptr) { - cur->data = (uint8_t *)mapping->addr + offs; + cur->data = (uint8_t *)mapping->addr + w.offs; } else { - memcpy(cur->data, (uint8_t *)mapping->addr + offs, ggml_nbytes(cur)); + memcpy(cur->data, (uint8_t *)mapping->addr + w.offs, ggml_nbytes(cur)); } } else { GGML_ASSERT(cur->data != nullptr); - file.seek(offs, SEEK_SET); - file.read_raw(cur->data, ggml_nbytes(cur)); + GGML_ASSERT(w.idx < files.size()); + const auto & file = files.at(w.idx); + file->seek(w.offs, SEEK_SET); + file->read_raw(cur->data, ggml_nbytes(cur)); } } size_t size_done = 0; size_t size_data = 0; - size_t mmap_used_first = -1; - size_t mmap_used_last = 0; + std::vector> mmaps_used; // Returns false if cancelled by progress_callback - bool load_all_data(struct ggml_context * ctx, llama_progress_callback progress_callback, void * progress_callback_user_data, ggml_backend_buffer_t buf_mmap, llama_mlock * lmlock) { - GGML_ASSERT(size_data != 0 && "call init_mapping() first"); + bool load_all_data( + struct ggml_context * ctx, + llama_buf_map & bufs_mmap, + llama_mlocks * lmlocks, + llama_progress_callback progress_callback, + void * progress_callback_user_data) { + GGML_ASSERT(size_data != 0 && "call init_mappings() first"); std::vector> read_buf; - for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur != NULL; cur = ggml_get_next_tensor(ctx, cur)) { if (progress_callback) { if (!progress_callback((float) size_done / size_data, progress_callback_user_data)) { @@ -3152,41 +3290,57 @@ struct llama_model_loader { } } - const size_t offs = file_offset(ggml_get_name(cur)); + const auto & w = get_weights(ggml_get_name(cur)); + size_t n_size = ggml_nbytes(cur); - if (use_mmap && mapping) { + if (use_mmap) { + const auto & mapping = mappings.at(w.idx); + ggml_backend_buffer_t buf_mmap = nullptr; + if (bufs_mmap.count(w.idx)) { + buf_mmap = bufs_mmap.at(w.idx); + } + GGML_ASSERT(buf_mmap || cur->data); // either we have a buffer to allocate the tensor in, or it is already allocated if (buf_mmap && cur->data == nullptr) { - ggml_backend_tensor_alloc(buf_mmap, cur, (uint8_t *) mapping->addr + offs); - if (lmlock) { - lmlock->grow_to(offs + ggml_nbytes(cur)); + ggml_backend_tensor_alloc(buf_mmap, cur, (uint8_t *) mapping->addr + w.offs); + if (lmlocks) { + const auto & lmlock = lmlocks->at(w.idx); + lmlock->grow_to(w.offs + ggml_nbytes(cur)); } - mmap_used_first = std::min(mmap_used_first, offs); - mmap_used_last = std::max(mmap_used_last, offs + ggml_nbytes(cur)); + + auto & mmap_used = mmaps_used[w.idx]; + mmap_used.first = std::min(mmap_used.first, w.offs); + mmap_used.second = std::max(mmap_used.second, w.offs + n_size); } else { - ggml_backend_tensor_set(cur, (uint8_t *) mapping->addr + offs, 0, ggml_nbytes(cur)); + ggml_backend_tensor_set(cur, (uint8_t *) mapping->addr + w.offs, 0, n_size); } } else { + GGML_ASSERT(w.idx < files.size()); + const auto & file = files.at(w.idx); if (ggml_backend_buffer_is_host(cur->buffer)) { - file.seek(offs, SEEK_SET); - file.read_raw(cur->data, ggml_nbytes(cur)); + file->seek(w.offs, SEEK_SET); + file->read_raw(cur->data, ggml_nbytes(cur)); } else { read_buf.resize(ggml_nbytes(cur)); - file.seek(offs, SEEK_SET); - file.read_raw(read_buf.data(), ggml_nbytes(cur)); - ggml_backend_tensor_set(cur, read_buf.data(), 0, ggml_nbytes(cur)); + file->seek(w.offs, SEEK_SET); + file->read_raw(read_buf.data(), ggml_nbytes(cur)); + ggml_backend_tensor_set(cur, read_buf.data(), 0, n_size); } } - size_done += ggml_nbytes(cur); + size_done += n_size; } // check if this is the last call and do final cleanup if (size_done >= size_data) { // unmap offloaded tensors and metadata - if (use_mmap && mapping) { - mapping->unmap_fragment(0, mmap_used_first); - if (mmap_used_last != 0) { - mapping->unmap_fragment(mmap_used_last, mapping->size); + if (use_mmap) { + for (uint32_t idx = 0; idx < mappings.size(); idx++) { + const auto & mmap_used = mmaps_used.at(idx); + auto & mapping = mappings.at(idx); + mapping->unmap_fragment(0, mmap_used.first); + if (mmap_used.second != 0) { + mapping->unmap_fragment(mmap_used.second, mapping->size); + } } } if (progress_callback) { @@ -3290,6 +3444,7 @@ static const char * llama_model_type_name(e_model type) { case MODEL_40B: return "40B"; case MODEL_65B: return "65B"; case MODEL_70B: return "70B"; + case MODEL_314B: return "314B"; case MODEL_SMALL: return "0.1B"; case MODEL_MEDIUM: return "0.4B"; case MODEL_LARGE: return "0.8B"; @@ -3319,7 +3474,7 @@ static void llm_load_hparams( llama_model_loader & ml, llama_model & model) { auto & hparams = model.hparams; - const gguf_context * ctx = ml.ctx_gguf; + const gguf_context * ctx = ml.meta; // get metadata as string for (int i = 0; i < gguf_get_n_kv(ctx); i++) { @@ -3428,6 +3583,15 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_GROK: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 64: model.type = e_model::MODEL_314B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; case LLM_ARCH_FALCON: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); @@ -3709,7 +3873,7 @@ static void llm_load_vocab( llama_model & model) { auto & vocab = model.vocab; - struct gguf_context * ctx = ml.ctx_gguf; + struct gguf_context * ctx = ml.meta; const auto kv = LLM_KV(model.arch); @@ -3842,7 +4006,7 @@ static void llm_load_vocab( } else if (vocab.type == LLAMA_VOCAB_TYPE_WPM) { vocab.linefeed_id = vocab.special_pad_id; } else { - const std::vector ids = llama_tokenize_internal(vocab, "\u010A", false); + const std::vector ids = llama_tokenize_internal(vocab, "\xC4\x8A", false); // U+010A GGML_ASSERT(!ids.empty() && "model vocab missing newline token"); vocab.linefeed_id = ids[0]; } @@ -4265,6 +4429,54 @@ static bool llm_load_tensors( } } } break; + case LLM_ARCH_GROK: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false); + // if output is NULL, init from the input tok embed + if (model.output == NULL) { + model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + ml.n_created--; // artificial tensor + ml.size_data += ggml_nbytes(model.output); + } + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + + layer.attn_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + + layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd}); + + GGML_ASSERT(hparams.n_expert > 0); + GGML_ASSERT(hparams.n_expert_used > 0); + + // MoE branch + for (uint32_t x = 0; x < hparams.n_expert; ++x) { + layer.ffn_gate_exp[x] = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, x), {n_embd, n_ff}); + layer.ffn_down_exp[x] = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, x), { n_ff, n_embd}); + layer.ffn_up_exp[x] = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXP, "weight", i, x), {n_embd, n_ff}); + } + + layer.layer_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}); + } + } break; case LLM_ARCH_BAICHUAN: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); @@ -4319,10 +4531,8 @@ static bool llm_load_tensors( layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}); - if (gguf_find_tensor(ml.ctx_gguf, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i).c_str()) >= 0) { - layer.attn_norm_2 = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}); - layer.attn_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}); - } + layer.attn_norm_2 = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, false); + layer.attn_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, false); layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}); layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); @@ -5024,56 +5234,97 @@ static bool llm_load_tensors( ml.done_getting_tensors(); - ml.init_mapping(true, use_mlock ? &model.mlock_mmap : nullptr); + ml.init_mappings(true, &model.mlock_mmaps); + model.mappings.reserve(ml.mappings.size()); // create the backend buffers - std::vector> ctx_bufs; + std::vector> ctx_bufs; + ctx_bufs.reserve(ctx_map.size()); + + // Ensure we have enough capacity for the maximum backend buffer we will potentially create + size_t n_max_backend_buffer = ctx_map.size() * ml.files.size(); + model.bufs.reserve(n_max_backend_buffer); for (auto & it : ctx_map) { ggml_backend_buffer_type_t buft = it.first; - ggml_context * ctx = it.second; - ggml_backend_buffer_t buf = nullptr; + ggml_context * ctx = it.second; + + llama_buf_map bufs; + bufs.reserve(n_max_backend_buffer); // only the mmap region containing the tensors in the model is mapped to the backend buffer // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size if (ml.use_mmap && buft == llama_default_buffer_type_cpu(true)) { - size_t first, last; - ml.get_mapping_range(&first, &last, ctx); - buf = ggml_backend_cpu_buffer_from_ptr((char *) ml.mapping->addr + first, last - first); + for (uint32_t idx = 0; idx < ml.files.size(); idx++) { + void * addr = nullptr; + size_t first, last; + ml.get_mapping_range(&first, &last, &addr, idx, ctx); + if (first >= last) { + continue; + } + ggml_backend_buffer_t buf = ggml_backend_cpu_buffer_from_ptr((char *) addr + first, last - first); + if (buf == nullptr) { + throw std::runtime_error("unable to allocate backend CPU buffer"); + } + model.bufs.push_back(buf); + bufs.emplace(idx, buf); #ifdef GGML_USE_CUBLAS - if (n_layer >= n_gpu_layers) { - ggml_backend_cuda_register_host_buffer( + if (n_layer >= n_gpu_layers) { + ggml_backend_cuda_register_host_buffer( ggml_backend_buffer_get_base(buf), ggml_backend_buffer_get_size(buf)); - } + } #endif + } } #ifdef GGML_USE_METAL else if (ml.use_mmap && buft == ggml_backend_metal_buffer_type()) { - const size_t max_size = ggml_get_max_tensor_size(ctx); - size_t first, last; - ml.get_mapping_range(&first, &last, ctx); - buf = ggml_backend_metal_buffer_from_ptr((char *) ml.mapping->addr + first, last - first, max_size); + for (uint32_t idx = 0; idx < ml.files.size(); idx++) { + const size_t max_size = ggml_get_max_tensor_size(ctx); + void * addr = nullptr; + size_t first, last; + ml.get_mapping_range(&first, &last, &addr, idx, ctx); + if (first >= last) { + continue; + } + ggml_backend_buffer_t buf = ggml_backend_metal_buffer_from_ptr((char *) addr + first, last - first, max_size); + if (buf == nullptr) { + throw std::runtime_error("unable to allocate backend metal buffer"); + } + model.bufs.push_back(buf); + bufs.emplace(idx, buf); + } } #endif else { - buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); - if (buf != nullptr && use_mlock && ggml_backend_buffer_is_host(buf)) { + ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); + if (buf == nullptr) { + throw std::runtime_error("unable to allocate backend buffer"); + } + model.bufs.push_back(buf); + if (use_mlock && ggml_backend_buffer_is_host(buf)) { model.mlock_bufs.emplace_back(new llama_mlock); auto & mlock_buf = model.mlock_bufs.back(); mlock_buf->init (ggml_backend_buffer_get_base(buf)); mlock_buf->grow_to(ggml_backend_buffer_get_size(buf)); } + for (uint32_t idx = 0; idx < ml.files.size(); idx++) { + bufs.emplace(idx, buf); + } } - if (buf == nullptr) { + + if (bufs.empty()) { throw std::runtime_error("failed to allocate buffer"); } - // indicate that this buffer contains weights - // this is used by ggml_backend_sched to improve op scheduling -> ops that use a weight are preferably scheduled to the backend that contains the weight - ggml_backend_buffer_set_usage(buf, GGML_BACKEND_BUFFER_USAGE_WEIGHTS); - model.bufs.push_back(buf); - ctx_bufs.emplace_back(ctx, buf); + + for (auto & buf : bufs) { + // indicate that this buffer contains weights + // this is used by ggml_backend_sched to improve op scheduling -> ops that use a weight are preferably scheduled to the backend that contains the weight + ggml_backend_buffer_set_usage(buf.second, GGML_BACKEND_BUFFER_USAGE_WEIGHTS); + } + + ctx_bufs.emplace_back(ctx, bufs); } if (llama_supports_gpu_offload()) { @@ -5105,13 +5356,15 @@ static bool llm_load_tensors( // load tensor data for (auto & it : ctx_bufs) { ggml_context * ctx = it.first; - ggml_backend_buffer_t buf = it.second; - if (!ml.load_all_data(ctx, progress_callback, progress_callback_user_data, buf, use_mlock ? &model.mlock_mmap : NULL)) { + auto & bufs = it.second; + if (!ml.load_all_data(ctx, bufs, use_mlock ? &model.mlock_mmaps : NULL, progress_callback, progress_callback_user_data)) { return false; } } - model.mapping = std::move(ml.mapping); + for (auto & mapping : ml.mappings) { + model.mappings.emplace_back(std::move(mapping)); + } // loading time will be recalculate after the first eval, so // we take page faults deferred by mmap() into consideration @@ -5451,6 +5704,20 @@ static struct ggml_tensor * llm_build_kqv( ggml_mul_mat_set_prec(kq, GGML_PREC_F32); } + if (model.arch == LLM_ARCH_GROK) { + // need to do the following: + // multiply by attn_output_multiplyer of 0.08838834764831845 + // and then : + // kq = 30 * tanh(kq / 30) + // before the softmax below + + //try from phi2 + //ggml_mul_mat_set_prec(kq, GGML_PREC_F32); + + kq = ggml_tanh(ctx, ggml_scale(ctx, kq, 0.08838834764831845f/30.0f)); + kq = ggml_scale(ctx, kq, 30); + } + #if defined(GGML_USE_KOMPUTE) #pragma message("TODO: ALiBi support in ggml_soft_max_ext is not implemented for Kompute") #pragma message(" Falling back to ggml_alibi(). Will become an error in Mar 2024") @@ -6225,6 +6492,203 @@ struct llm_build_context { return gf; } + struct ggml_cgraph * build_grok() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // multiply by embedding_multiplier_scale of 78.38367176906169 + inpL = ggml_scale(ctx0, inpL, 78.38367176906169f); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = build_inp_pos(); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + + // self-attention + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il); + } + + // Grok + // if attn_out_norm is present then apply it before adding the input + if (model.layers[il].attn_out_norm) { + cur = llm_build_norm(ctx0, cur, hparams, + model.layers[il].attn_out_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_out_norm", il); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + // MoE branch + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * logits = ggml_mul_mat(ctx0, model.layers[il].ffn_gate_inp, cur); // [n_tokens, num_experts] + cb(logits, "ffn_moe_logits", il); + + ggml_tensor * probs = ggml_soft_max(ctx0, logits); // [n_tokens, num_experts] + cb(probs, "ffn_moe_probs", il); + + // select experts + ggml_tensor * selected_experts = ggml_top_k(ctx0, probs, n_expert_used); // [n_tokens, num_experts_per_tok] + cb(selected_experts->src[0], "ffn_moe_argsort", il); + + ggml_tensor * weights = ggml_get_rows(ctx0, + ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); + cb(weights, "ffn_moe_weights", il); + + weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); // [n_tokens, num_experts_per_tok] + + ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); + cb(weights_sum, "ffn_moe_weights_sum", il); + + weights = ggml_div(ctx0, weights, weights_sum); // [n_tokens, num_experts_per_tok] + cb(weights, "ffn_moe_weights_norm", il); + + // compute expert outputs + ggml_tensor * moe_out = nullptr; + + for (int i = 0; i < n_expert_used; ++i) { + ggml_tensor * cur_expert; + + ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exp, n_expert, selected_experts, i, cur); + cb(cur_up, "ffn_moe_up", il); + + ggml_tensor * cur_gate = ggml_mul_mat_id(ctx0, model.layers[il].ffn_gate_exp, n_expert, selected_experts, i, cur); + cb(cur_gate, "ffn_moe_gate", il); + + //GeLU + cur_gate = ggml_gelu(ctx0, cur_gate); + cb(cur_gate, "ffn_moe_gelu", il); + + cur_expert = ggml_mul(ctx0, cur_up, cur_gate); // [n_tokens, n_embd] + cb(cur_expert, "ffn_moe_gate_par", il); + + cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exp, n_expert, selected_experts, i, cur_expert); // [n_tokens, n_embd] + cb(cur_expert, "ffn_moe_down", il); + + cur_expert = ggml_mul(ctx0, cur_expert, + ggml_view_2d(ctx0, weights, 1, n_tokens, weights->nb[1], i*weights->nb[0])); + cb(cur_expert, "ffn_moe_weighted", il); + + if (i == 0) { + moe_out = cur_expert; + } else { + moe_out = ggml_add(ctx0, moe_out, cur_expert); + cb(moe_out, "ffn_moe_out", il); + } + } + + cur = moe_out; + + // Grok + // if layer_out_norm is present then apply it before adding the input + // Idea: maybe ffn_out_norm is a better name + if (model.layers[il].layer_out_norm) { + cur = llm_build_norm(ctx0, cur, hparams, + model.layers[il].layer_out_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "layer_out_norm", il); + } + + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + ggml_tensor * layer_dir = lctx.cvec.tensor_for(il); + if (layer_dir != nullptr) { + cur = ggml_add(ctx0, cur, layer_dir); + } + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + + // Grok + // multiply logits by output_multiplier_scale of 0.5773502691896257 + + cur = ggml_scale(ctx0, cur, 0.5773502691896257f); + + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + struct ggml_cgraph * build_starcoder() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); @@ -8648,6 +9112,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_falcon(); } break; + case LLM_ARCH_GROK: + { + result = llm.build_grok(); + } break; case LLM_ARCH_STARCODER: { result = llm.build_starcoder(); @@ -11971,27 +12439,34 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n // for arches that share the same tensor between the token embeddings and the output, we quantize the token embeddings // with the quantization of the output tensor if (name == tn(LLM_TENSOR_OUTPUT, "weight") || (!qs.has_output && name == tn(LLM_TENSOR_TOKEN_EMBD, "weight"))) { - int nx = tensor->ne[0]; - if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) { - new_type = GGML_TYPE_Q8_0; - } - else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || - ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) { - new_type = GGML_TYPE_Q5_K; - } - else if (new_type != GGML_TYPE_Q8_0) { - new_type = GGML_TYPE_Q6_K; + if (qs.params->output_tensor_type < GGML_TYPE_COUNT) { + new_type = qs.params->output_tensor_type; + } else { + int nx = tensor->ne[0]; + if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) { + new_type = GGML_TYPE_Q8_0; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || + ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) { + new_type = GGML_TYPE_Q5_K; + } + else if (new_type != GGML_TYPE_Q8_0) { + new_type = GGML_TYPE_Q6_K; + } } } else if (name == "token_embd.weight") { - if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || - ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) { - new_type = GGML_TYPE_Q2_K; - } - else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) { - new_type = GGML_TYPE_IQ3_S; - } - else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) { - new_type = GGML_TYPE_IQ3_S; + if (qs.params->token_embedding_type < GGML_TYPE_COUNT) { + new_type = qs.params->token_embedding_type; + } else { + if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) { + new_type = GGML_TYPE_Q2_K; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) { + new_type = GGML_TYPE_IQ3_S; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) { + new_type = GGML_TYPE_IQ3_S; + } } } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) { @@ -12027,13 +12502,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) { new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : !qs.has_imatrix ? GGML_TYPE_IQ3_S : GGML_TYPE_IQ3_XXS; } - else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_S && qs.model.hparams.n_gqa() >= 4) { - new_type = GGML_TYPE_Q4_K; - } - else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) { - new_type = GGML_TYPE_Q4_K; - } - else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_S && qs.model.hparams.n_gqa() >= 4) { + else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S) && qs.model.hparams.n_gqa() >= 4) { new_type = GGML_TYPE_Q4_K; } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) { @@ -12308,7 +12777,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s #endif llama_model_loader ml(fname_inp, use_mmap, NULL); - ml.init_mapping(false); // no prefetching? + ml.init_mappings(false); // no prefetching? llama_model model; llm_load_arch(ml, model); @@ -12332,12 +12801,12 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s struct gguf_context * ctx_out = gguf_init_empty(); // copy the KV pairs from the input file - gguf_set_kv (ctx_out, ml.ctx_gguf); + gguf_set_kv (ctx_out, ml.meta); gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION); gguf_set_val_u32(ctx_out, "general.file_type", ftype); for (int i = 0; i < ml.n_tensors; ++i) { - struct ggml_tensor * meta = ml.get_tensor_meta(i); + const struct ggml_tensor * meta = ml.get_tensor_meta(i); const std::string name = ggml_get_name(meta); @@ -12377,7 +12846,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s // populate the original tensors so we get an initial meta data for (int i = 0; i < ml.n_tensors; ++i) { - struct ggml_tensor * meta = ml.get_tensor_meta(i); + const struct ggml_tensor * meta = ml.get_tensor_meta(i); gguf_add_tensor(ctx_out, meta); } @@ -12582,7 +13051,7 @@ static int llama_apply_lora_from_file_internal( if (path_base_model) { LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model); ml.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*kv_overrides*/ nullptr)); - ml->init_mapping(/*prefetch*/ false); // no prefetching + ml->init_mappings(/*prefetch*/ false); // no prefetching } struct tensor_meta { @@ -12703,7 +13172,7 @@ static int llama_apply_lora_from_file_internal( ggml_tensor * base_t; if (ml) { - if (gguf_find_tensor(ml->ctx_gguf, base_name.c_str()) < 0) { + if (!ml->get_tensor_meta(base_name.c_str())) { LLAMA_LOG_ERROR("%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str()); return 1; } @@ -12887,6 +13356,8 @@ struct llama_model_quantize_params llama_model_quantize_default_params() { struct llama_model_quantize_params result = { /*.nthread =*/ 0, /*.ftype =*/ LLAMA_FTYPE_MOSTLY_Q5_1, + /*.output_tensor_type =*/ GGML_TYPE_COUNT, + /*.token_embedding_type =*/ GGML_TYPE_COUNT, /*.allow_requantize =*/ false, /*.quantize_output_tensor =*/ true, /*.only_copy =*/ false, @@ -13044,6 +13515,9 @@ struct llama_context * llama_new_context_with_model( cparams.rope_freq_base = params.rope_freq_base == 0.0f ? hparams.rope_freq_base_train : params.rope_freq_base; cparams.rope_freq_scale = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale; + // this is necessary due to kv_self.n being padded later during inference + cparams.n_ctx = GGML_PAD(cparams.n_ctx, 32); + // with causal attention, the batch size is limited by the context size cparams.n_batch = hparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch; cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch); @@ -13158,30 +13632,28 @@ struct llama_context * llama_new_context_with_model( } } #elif defined(GGML_USE_SYCL) - if (model->n_gpu_layers > 0) { - // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used - if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) { - ggml_backend_t backend = ggml_backend_sycl_init(model->main_gpu); + // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used + if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) { + ggml_backend_t backend = ggml_backend_sycl_init(model->main_gpu); + if (backend == nullptr) { + int main_gpu_id = ggml_backend_sycl_get_device_id(model->main_gpu); + LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d) backend\n", __func__, main_gpu_id, model->main_gpu); + llama_free(ctx); + return nullptr; + } + ctx->backends.push_back(backend); + } else { + // LLAMA_SPLIT_LAYER requires a backend for each GPU + for (int i = 0; i < ggml_backend_sycl_get_device_count(); ++i) { + ggml_backend_t backend = ggml_backend_sycl_init(i); if (backend == nullptr) { - int main_gpu_id = ggml_backend_sycl_get_device_id(model->main_gpu); - LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d) backend\n", __func__, main_gpu_id, model->main_gpu); + int id_list[GGML_SYCL_MAX_DEVICES]; + ggml_sycl_get_gpu_list(id_list, GGML_SYCL_MAX_DEVICES); + LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d) backend\n", __func__, id_list[i], i); llama_free(ctx); return nullptr; } ctx->backends.push_back(backend); - } else { - // LLAMA_SPLIT_LAYER requires a backend for each GPU - for (int i = 0; i < ggml_backend_sycl_get_device_count(); ++i) { - ggml_backend_t backend = ggml_backend_sycl_init(i); - if (backend == nullptr) { - int id_list[GGML_SYCL_MAX_DEVICES]; - ggml_sycl_get_gpu_list(id_list, GGML_SYCL_MAX_DEVICES); - LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d) backend\n", __func__, id_list[i], i); - llama_free(ctx); - return nullptr; - } - ctx->backends.push_back(backend); - } } } #elif defined(GGML_USE_KOMPUTE) @@ -13385,6 +13857,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { // the pairs of head values are offset by n_rot/2 case LLM_ARCH_FALCON: + case LLM_ARCH_GROK: case LLM_ARCH_PERSIMMON: case LLM_ARCH_BERT: case LLM_ARCH_NOMIC_BERT: @@ -14648,6 +15121,30 @@ LLAMA_API int32_t llama_chat_apply_template( return res; } +LLAMA_API int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count) { + static const char * const SPLIT_PATH_FORMAT = "%s-%05d-of-%05d.gguf"; + if (snprintf(split_path, maxlen, SPLIT_PATH_FORMAT, path_prefix, split_no + 1, split_count)) { + return strlen(split_path); + } + return 0; +} + +int llama_split_prefix(char * dest, size_t maxlen, const char * split_path, int split_no, int split_count) { + std::string str_split_path(split_path); + char postfix[32]; + snprintf(postfix, 32, "-%05d-of-%05d.gguf", split_no + 1, split_count); + std::string str_postfix(postfix); + + // check if dest ends with postfix + int size_prefix = str_split_path.size() - str_postfix.size(); + if (size_prefix > 0 && str_split_path.find(str_postfix, size_prefix) != std::string::npos) { + snprintf(dest, std::min((size_t) size_prefix + 1, maxlen), "%s", split_path); + return size_prefix; + } + + return 0; +} + struct llama_timings llama_get_timings(struct llama_context * ctx) { struct llama_timings result = { /*.t_start_ms =*/ 1e-3 * ctx->t_start_us, diff --git a/llama.h b/llama.h index 40dcf54e3..74f0e56de 100644 --- a/llama.h +++ b/llama.h @@ -275,13 +275,15 @@ extern "C" { // model quantization parameters typedef struct llama_model_quantize_params { - int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() - enum llama_ftype ftype; // quantize to this llama_ftype - bool allow_requantize; // allow quantizing non-f32/f16 tensors - bool quantize_output_tensor; // quantize output.weight - bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored - bool pure; // quantize all tensors to the default type - void * imatrix; // pointer to importance matrix data + int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() + enum llama_ftype ftype; // quantize to this llama_ftype + enum ggml_type output_tensor_type; // output tensor type + enum ggml_type token_embedding_type; // itoken embeddings tensor type + bool allow_requantize; // allow quantizing non-f32/f16 tensors + bool quantize_output_tensor; // quantize output.weight + bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored + bool pure; // quantize all tensors to the default type + void * imatrix; // pointer to importance matrix data } llama_model_quantize_params; // grammar types @@ -960,6 +962,16 @@ extern "C" { int32_t n_past, int32_t n_predict); + /// @details Build a split GGUF final path for this chunk. + /// llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf" + // Returns the split_path length. + LLAMA_API int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count); + + /// @details Extract the path prefix from the split_path if and only if the split_no and split_count match. + /// llama_split_prefix(split_prefix, 64, "/models/ggml-model-q4_0-00002-of-00004.gguf", 2, 4) => split_prefix = "/models/ggml-model-q4_0" + // Returns the split_prefix length. + LLAMA_API int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count); + // Performance information LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx); diff --git a/scripts/get-wikitext-103.sh b/scripts/get-wikitext-103.sh new file mode 100755 index 000000000..880dd5cbe --- /dev/null +++ b/scripts/get-wikitext-103.sh @@ -0,0 +1,10 @@ +#!/bin/bash + +wget https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-103-raw-v1.zip + +echo "Usage:" +echo "" +echo " ./perplexity -m model.gguf -f wiki.test.raw [other params]" +echo "" + +exit 0 diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index 10326d531..a43439aed 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -1,66 +1,75 @@ -function(llama_build_executable source) - get_filename_component(TEST_TARGET ${source} NAME_WE) +# Builds and runs a test source file. +# Optional args: +# - NAME: name of the executable & test target (defaults to the source file name without extension) +# - LABEL: label for the test (defaults to main) +# - ARGS: arguments to pass to the test executable +# - WORKING_DIRECTORY +function(llama_test source) + include(CMakeParseArguments) + set(options) + set(oneValueArgs NAME LABEL WORKING_DIRECTORY) + set(multiValueArgs ARGS) + cmake_parse_arguments(LLAMA_TEST "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + + if (NOT DEFINED LLAMA_TEST_LABEL) + set(LLAMA_TEST_LABEL "main") + endif() + if (NOT DEFINED LLAMA_TEST_WORKING_DIRECTORY) + set(LLAMA_TEST_WORKING_DIRECTORY .) + endif() + if (DEFINED LLAMA_TEST_NAME) + set(TEST_TARGET ${LLAMA_TEST_NAME}) + else() + get_filename_component(TEST_TARGET ${source} NAME_WE) + endif() + add_executable(${TEST_TARGET} ${source} get-model.cpp) install(TARGETS ${TEST_TARGET} RUNTIME) - target_link_libraries(${TEST_TARGET} PRIVATE common) + target_link_libraries(${TEST_TARGET} PRIVATE common json-schema-to-grammar) + add_test( + NAME ${TEST_TARGET} + WORKING_DIRECTORY ${LLAMA_TEST_WORKING_DIRECTORY} + COMMAND $ + ${LLAMA_TEST_ARGS}) + + set_property(TEST ${TEST_TARGET} PROPERTY LABELS ${LLAMA_TEST_LABEL}) endfunction() -function(llama_test_executable name source) - get_filename_component(TEST_TARGET ${source} NAME_WE) - add_test(NAME ${name} COMMAND $ ${ARGN}) - set_property(TEST ${name} PROPERTY LABELS "main") -endfunction() +# llama_test(test-double-float.cpp) # SLOW +llama_test(test-quantize-fns.cpp) +llama_test(test-quantize-perf.cpp) +llama_test(test-sampling.cpp) +llama_test(test-chat-template.cpp) -function(llama_build_and_test_executable source) - llama_build_and_test_executable_with_label(${source} "main") -endfunction() +llama_test(test-tokenizer-0-llama.cpp NAME test-tokenizer-0-llama ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf) +llama_test(test-tokenizer-0-falcon.cpp NAME test-tokenizer-0-falcon ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf) -function(llama_build_and_test_executable_with_label source label) - get_filename_component(TEST_TARGET ${source} NAME_WE) - add_executable(${TEST_TARGET} ${source} get-model.cpp) - install(TARGETS ${TEST_TARGET} RUNTIME) - target_link_libraries(${TEST_TARGET} PRIVATE common) - add_test(NAME ${TEST_TARGET} COMMAND $ ${ARGN}) - set_property(TEST ${TEST_TARGET} PROPERTY LABELS ${label}) -endfunction() +llama_test(test-tokenizer-1-llama.cpp NAME test-tokenizer-1-llama ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf) +llama_test(test-tokenizer-1-llama.cpp NAME test-tokenizer-1-baichuan ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf) -# llama_build_and_test_executable(test-double-float.cpp) # SLOW -llama_build_and_test_executable(test-quantize-fns.cpp) -llama_build_and_test_executable(test-quantize-perf.cpp) -llama_build_and_test_executable(test-sampling.cpp) -llama_build_and_test_executable(test-chat-template.cpp) +llama_test(test-tokenizer-1-bpe.cpp NAME test-tokenizer-1-falcon ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf) +llama_test(test-tokenizer-1-bpe.cpp NAME test-tokenizer-1-aquila ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf) +llama_test(test-tokenizer-1-bpe.cpp NAME test-tokenizer-1-mpt ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf) +llama_test(test-tokenizer-1-bpe.cpp NAME test-tokenizer-1-stablelm-3b-4e1t ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-stablelm-3b-4e1t.gguf) +llama_test(test-tokenizer-1-bpe.cpp NAME test-tokenizer-1-gpt-neox ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf) +llama_test(test-tokenizer-1-bpe.cpp NAME test-tokenizer-1-refact ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf) +llama_test(test-tokenizer-1-bpe.cpp NAME test-tokenizer-1-starcoder ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf) +llama_test(test-tokenizer-1-bpe.cpp NAME test-tokenizer-1-gpt2 ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt2.gguf) +#llama_test(test-tokenizer-1-bpe.cpp NAME test-tokenizer-1-bloom ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bloom.gguf) # BIG -llama_build_executable(test-tokenizer-0-llama.cpp) -llama_test_executable (test-tokenizer-0-llama test-tokenizer-0-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf) +llama_test(test-grammar-parser.cpp) +llama_test(test-llama-grammar.cpp) +llama_test(test-grad0.cpp) +# llama_test(test-opt.cpp) # SLOW +llama_test(test-backend-ops.cpp) -llama_build_executable(test-tokenizer-0-falcon.cpp) -llama_test_executable (test-tokenizer-0-falcon test-tokenizer-0-falcon.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf) +llama_test(test-rope.cpp) -llama_build_executable(test-tokenizer-1-llama.cpp) -llama_test_executable (test-tokenizer-1-llama test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf) -llama_test_executable (test-tokenizer-1-baichuan test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf) +llama_test(test-model-load-cancel.cpp LABEL "model") +llama_test(test-autorelease.cpp LABEL "model") -llama_build_executable(test-tokenizer-1-bpe.cpp) -llama_test_executable (test-tokenizer-1-falcon test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf) -llama_test_executable (test-tokenizer-1-aquila test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf) -llama_test_executable (test-tokenizer-1-mpt test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf) -llama_test_executable (test-tokenizer-1-stablelm-3b-4e1t test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-stablelm-3b-4e1t.gguf) -llama_test_executable (test-tokenizer-1-gpt-neox test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf) -llama_test_executable (test-tokenizer-1-refact test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf) -llama_test_executable (test-tokenizer-1-starcoder test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf) -llama_test_executable (test-tokenizer-1-gpt2 test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt2.gguf) -# llama_test_executable (test-tokenizer-1-bloom test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bloom.gguf) # BIG - -llama_build_and_test_executable(test-grammar-parser.cpp) -llama_build_and_test_executable(test-llama-grammar.cpp) -llama_build_and_test_executable(test-grad0.cpp) -# llama_build_and_test_executable(test-opt.cpp) # SLOW -llama_build_and_test_executable(test-backend-ops.cpp) - -llama_build_and_test_executable(test-rope.cpp) - -llama_build_and_test_executable_with_label(test-model-load-cancel.cpp "model") -llama_build_and_test_executable_with_label(test-autorelease.cpp "model") +llama_test(test-json-schema-to-grammar.cpp WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..) +target_include_directories(test-json-schema-to-grammar PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../examples/server) # dummy executable - not installed get_filename_component(TEST_TARGET test-c.c NAME_WE) diff --git a/tests/run-json-schema-to-grammar.mjs b/tests/run-json-schema-to-grammar.mjs new file mode 100644 index 000000000..71bf62ed3 --- /dev/null +++ b/tests/run-json-schema-to-grammar.mjs @@ -0,0 +1,10 @@ +import { readFileSync } from "fs" +import { SchemaConverter } from "../examples/server/public/json-schema-to-grammar.mjs" + +const [, , file] = process.argv +const url = `file://${file}` +let schema = JSON.parse(readFileSync(file, "utf8")); +const converter = new SchemaConverter({}) +schema = await converter.resolveRefs(schema, url) +converter.visit(schema, '') +console.log(converter.formatGrammar()) diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index c2916c3e4..1998e1cbc 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -2091,6 +2091,13 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op } } + test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 64, 2, 128, { 8, 1}, {1, 1})); + test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 83, 2, 128, { 8, 1}, {4, 1})); + test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 64, 2, 64, { 8, 1}, {4, 1})); + test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 83, 2, 64, { 8, 1}, {4, 1})); + test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 64, 45, 128, { 8, 1}, {4, 1})); + test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 45, 64, { 8, 1}, {4, 1})); + for (ggml_type type_a : all_types) { for (ggml_type type_b : {GGML_TYPE_F32 /*, GGML_TYPE_F16 */}) { for (int n_mats : {2, 4, 8}) { diff --git a/tests/test-json-schema-to-grammar.cpp b/tests/test-json-schema-to-grammar.cpp new file mode 100755 index 000000000..4d110cf5a --- /dev/null +++ b/tests/test-json-schema-to-grammar.cpp @@ -0,0 +1,842 @@ +#ifdef NDEBUG +#undef NDEBUG +#endif + +#include +#include +#include + +#include "json-schema-to-grammar.h" +#include "grammar-parser.h" + +static std::string trim(const std::string & source) { + std::string s(source); + s.erase(0,s.find_first_not_of(" \n\r\t")); + s.erase(s.find_last_not_of(" \n\r\t")+1); + return std::regex_replace(s, std::regex("(^|\n)[ \t]+"), "$1"); +} + +enum TestCaseStatus { + SUCCESS, + FAILURE +}; + +struct TestCase { + TestCaseStatus expected_status; + std::string name; + std::string schema; + std::string expected_grammar; + + void _print_failure_header() const { + fprintf(stderr, "#\n# Test '%s' failed.\n#\n%s\n", name.c_str(), schema.c_str()); + } + void verify(const std::string & actual_grammar) const { + if (trim(actual_grammar) != trim(expected_grammar)) { + _print_failure_header(); + fprintf(stderr, "# EXPECTED:\n%s\n# ACTUAL:\n%s\n", expected_grammar.c_str(), actual_grammar.c_str()); + assert(false); + } + } + void verify_expectation_parseable() const { + try { + auto state = grammar_parser::parse(expected_grammar.c_str()); + if (state.symbol_ids.find("root") == state.symbol_ids.end()) { + throw std::runtime_error("Grammar failed to parse:\n" + expected_grammar); + } + } catch (const std::runtime_error & ex) { + _print_failure_header(); + fprintf(stderr, "# GRAMMAR ERROR: %s\n", ex.what()); + assert(false); + } + } + void verify_status(TestCaseStatus status) const { + if (status != expected_status) { + _print_failure_header(); + fprintf(stderr, "# EXPECTED STATUS: %s\n", expected_status == SUCCESS ? "SUCCESS" : "FAILURE"); + fprintf(stderr, "# ACTUAL STATUS: %s\n", status == SUCCESS ? "SUCCESS" : "FAILURE"); + assert(false); + } + } +}; + +static void write(const std::string & file, const std::string & content) { + std::ofstream f; + f.open(file.c_str()); + f << content.c_str(); + f.close(); +} + +static std::string read(const std::string & file) { + std::ostringstream actuals; + actuals << std::ifstream(file.c_str()).rdbuf(); + return actuals.str(); +} + +static void test_all(const std::string & lang, std::function runner) { + fprintf(stderr, "#\n# Testing JSON schema conversion (%s)\n#\n", lang.c_str()); + auto test = [&](const TestCase & tc) { + fprintf(stderr, "- %s%s\n", tc.name.c_str(), tc.expected_status == FAILURE ? " (failure expected)" : ""); + runner(tc); + }; + + test({ + FAILURE, + "unknown type", + R"""({ + "type": "kaboom" + })""", + "" + }); + + test({ + FAILURE, + "invalid type", + R"""({ + "type": 123 + })""", + "" + }); + + test({ + SUCCESS, + "empty schema (object)", + "{}", + R"""( + array ::= "[" space ( value ("," space value)* )? "]" space + boolean ::= ("true" | "false") space + null ::= "null" space + number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space + object ::= "{" space ( string ":" space value ("," space string ":" space value)* )? "}" space + root ::= object + space ::= " "? + string ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + value ::= object | array | string | number | boolean + )""" + }); + + test({ + SUCCESS, + "exotic formats", + R"""({ + "items": [ + { "format": "date" }, + { "format": "uuid" }, + { "format": "time" }, + { "format": "date-time" } + ] + })""", + R"""( + date ::= [0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( "0" [1-9] | [1-2] [0-9] | "3" [0-1] ) + date-string ::= "\"" date "\"" space + date-time ::= date "T" time + date-time-string ::= "\"" date-time "\"" space + root ::= "[" space date-string "," space uuid "," space time-string "," space date-time-string "]" space + space ::= " "? + time ::= ([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] ) + time-string ::= "\"" time "\"" space + uuid ::= "\"" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "-" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "-" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "-" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "-" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "\"" space + )""" + }); + + test({ + SUCCESS, + "string", + R"""({ + "type": "string" + })""", + R"""( + root ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "boolean", + R"""({ + "type": "boolean" + })""", + R"""( + root ::= ("true" | "false") space + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "integer", + R"""({ + "type": "integer" + })""", + R"""( + root ::= ("-"? ([0-9] | [1-9] [0-9]*)) space + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "string const", + R"""({ + "const": "foo" + })""", + R"""( + root ::= "\"foo\"" + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "non-string const", + R"""({ + "const": 123 + })""", + R"""( + root ::= "123" + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "non-string enum", + R"""({ + "enum": ["red", "amber", "green", null, 42, ["foo"]] + })""", + R"""( + root ::= "\"red\"" | "\"amber\"" | "\"green\"" | "null" | "42" | "[\"foo\"]" + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "tuple1", + R"""({ + "prefixItems": [{ "type": "string" }] + })""", + R"""( + root ::= "[" space string "]" space + space ::= " "? + string ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + )""" + }); + + test({ + SUCCESS, + "tuple2", + R"""({ + "prefixItems": [{ "type": "string" }, { "type": "number" }] + })""", + R"""( + number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space + root ::= "[" space string "," space number "]" space + space ::= " "? + string ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + )""" + }); + + test({ + SUCCESS, + "number", + R"""({ + "type": "number" + })""", + R"""( + root ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "minItems", + R"""({ + "items": { + "type": "boolean" + }, + "minItems": 2 + })""", + R"""( + boolean ::= ("true" | "false") space + root ::= "[" space boolean ( "," space boolean )( "," space boolean )* "]" space + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "maxItems 1", + R"""({ + "items": { + "type": "boolean" + }, + "maxItems": 1 + })""", + R"""( + boolean ::= ("true" | "false") space + root ::= "[" space ( boolean )? "]" space + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "maxItems 2", + R"""({ + "items": { + "type": "boolean" + }, + "maxItems": 2 + })""", + R"""( + boolean ::= ("true" | "false") space + root ::= "[" space ( boolean ( "," space boolean )? )? "]" space + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "min + maxItems", + R"""({ + "items": { + "type": ["number", "integer"] + }, + "minItems": 3, + "maxItems": 5 + })""", + R"""( + integer ::= ("-"? ([0-9] | [1-9] [0-9]*)) space + item ::= number | integer + number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space + root ::= "[" space item ( "," space item )( "," space item )( "," space item )?( "," space item )? "]" space + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "simple regexp", + R"""({ + "type": "string", + "pattern": "^abc?d*efg+(hij)?kl$" + })""", + R"""( + root ::= "\"" "ab" "c"? "d"* "ef" "g"+ ("hij")? "kl" "\"" space + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "regexp escapes", + R"""({ + "type": "string", + "pattern": "^\\[\\]\\{\\}\\(\\)\\|\\+\\*\\?$" + })""", + R"""( + root ::= "\"" "[]{}()|+*?" "\"" space + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "regexp quote", + R"""({ + "type": "string", + "pattern": "^\"$" + })""", + R"""( + root ::= "\"" "\"" "\"" space + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "regexp", + R"""({ + "type": "string", + "pattern": "^(\\([0-9]{1,3}\\))?[0-9]{3}-[0-9]{4} and...$" + })""", + R"""( + dot ::= [\U00000000-\x09\x0B\x0C\x0E-\U0010FFFF] + root ::= "\"" ("(" root-1 root-1? root-1? ")")? root-1 root-1 root-1 "-" root-1 root-1 root-1 root-1 " and" dot dot dot "\"" space + root-1 ::= [0-9] + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "required props in original order", + R"""({ + "type": "object", + "properties": { + "b": {"type": "string"}, + "c": {"type": "string"}, + "a": {"type": "string"} + }, + "required": [ + "a", + "b", + "c" + ], + "additionalProperties": false, + "definitions": {} + })""", + R"""( + a-kv ::= "\"a\"" space ":" space string + b-kv ::= "\"b\"" space ":" space string + c-kv ::= "\"c\"" space ":" space string + root ::= "{" space b-kv "," space c-kv "," space a-kv "}" space + space ::= " "? + string ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + )""" + }); + + test({ + SUCCESS, + "1 optional prop", + R"""({ + "properties": { + "a": { + "type": "string" + } + }, + "additionalProperties": false + })""", + R"""( + a-kv ::= "\"a\"" space ":" space string + root ::= "{" space (a-kv )? "}" space + space ::= " "? + string ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + )""" + }); + + test({ + SUCCESS, + "N optional props", + R"""({ + "properties": { + "a": {"type": "string"}, + "b": {"type": "string"}, + "c": {"type": "string"} + }, + "additionalProperties": false + })""", + R"""( + a-kv ::= "\"a\"" space ":" space string + a-rest ::= ( "," space b-kv )? b-rest + b-kv ::= "\"b\"" space ":" space string + b-rest ::= ( "," space c-kv )? + c-kv ::= "\"c\"" space ":" space string + root ::= "{" space (a-kv a-rest | b-kv b-rest | c-kv )? "}" space + space ::= " "? + string ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + )""" + }); + + test({ + SUCCESS, + "required + optional props each in original order", + R"""({ + "properties": { + "b": {"type": "string"}, + "a": {"type": "string"}, + "d": {"type": "string"}, + "c": {"type": "string"} + }, + "required": ["a", "b"], + "additionalProperties": false + })""", + R"""( + a-kv ::= "\"a\"" space ":" space string + b-kv ::= "\"b\"" space ":" space string + c-kv ::= "\"c\"" space ":" space string + d-kv ::= "\"d\"" space ":" space string + d-rest ::= ( "," space c-kv )? + root ::= "{" space b-kv "," space a-kv ( "," space ( d-kv d-rest | c-kv ) )? "}" space + space ::= " "? + string ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + )""" + }); + + test({ + SUCCESS, + "additional props", + R"""({ + "type": "object", + "additionalProperties": {"type": "array", "items": {"type": "number"}} + })""", + R"""( + additional-kv ::= string ":" space additional-value + additional-kvs ::= additional-kv ( "," space additional-kv )* + additional-value ::= "[" space ( number ( "," space number )* )? "]" space + number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space + root ::= "{" space (additional-kvs )? "}" space + space ::= " "? + string ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + )""" + }); + + test({ + SUCCESS, + "additional props (true)", + R"""({ + "type": "object", + "additionalProperties": true + })""", + R"""( + array ::= "[" space ( value ("," space value)* )? "]" space + boolean ::= ("true" | "false") space + null ::= "null" space + number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space + object ::= "{" space ( string ":" space value ("," space string ":" space value)* )? "}" space + root ::= object + space ::= " "? + string ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + value ::= object | array | string | number | boolean + )""" + }); + + test({ + SUCCESS, + "additional props (implicit)", + R"""({ + "type": "object" + })""", + R"""( + array ::= "[" space ( value ("," space value)* )? "]" space + boolean ::= ("true" | "false") space + null ::= "null" space + number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space + object ::= "{" space ( string ":" space value ("," space string ":" space value)* )? "}" space + root ::= object + space ::= " "? + string ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + value ::= object | array | string | number | boolean + )""" + }); + + test({ + SUCCESS, + "empty w/o additional props", + R"""({ + "type": "object", + "additionalProperties": false + })""", + R"""( + root ::= "{" space "}" space + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "required + additional props", + R"""({ + "type": "object", + "properties": { + "a": {"type": "number"} + }, + "required": ["a"], + "additionalProperties": {"type": "string"} + })""", + R"""( + a-kv ::= "\"a\"" space ":" space number + additional-kv ::= string ":" space string + additional-kvs ::= additional-kv ( "," space additional-kv )* + number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space + root ::= "{" space a-kv ( "," space ( additional-kvs ) )? "}" space + space ::= " "? + string ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + )""" + }); + + test({ + SUCCESS, + "optional + additional props", + R"""({ + "type": "object", + "properties": { + "a": {"type": "number"} + }, + "additionalProperties": {"type": "number"} + })""", + R"""( + a-kv ::= "\"a\"" space ":" space number + a-rest ::= additional-kvs + additional-kv ::= string ":" space number + additional-kvs ::= additional-kv ( "," space additional-kv )* + number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space + root ::= "{" space (a-kv a-rest | additional-kvs )? "}" space + space ::= " "? + string ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + )""" + }); + + test({ + SUCCESS, + "required + optional + additional props", + R"""({ + "type": "object", + "properties": { + "a": {"type": "number"}, + "b": {"type": "number"} + }, + "required": ["a"], + "additionalProperties": {"type": "number"} + })""", + R"""( + a-kv ::= "\"a\"" space ":" space number + additional-kv ::= string ":" space number + additional-kvs ::= additional-kv ( "," space additional-kv )* + b-kv ::= "\"b\"" space ":" space number + b-rest ::= additional-kvs + number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space + root ::= "{" space a-kv ( "," space ( b-kv b-rest | additional-kvs ) )? "}" space + space ::= " "? + string ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + )""" + }); + + test({ + SUCCESS, + "top-level $ref", + R"""({ + "$ref": "#/definitions/MyType", + "definitions": { + "MyType": { + "type": "object", + "properties": { + "a": { + "type": "string" + } + }, + "required": [ + "a" + ], + "additionalProperties": false + } + } + })""", + R"""( + MyType ::= "{" space MyType-a-kv "}" space + MyType-a-kv ::= "\"a\"" space ":" space string + root ::= MyType + space ::= " "? + string ::= "\"" ( + [^"\\] | + "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) + )* "\"" space + )""" + }); + + test({ + SUCCESS, + "anyOf", + R"""({ + "anyOf": [ + {"$ref": "#/definitions/foo"}, + {"$ref": "#/definitions/bar"} + ], + "definitions": { + "foo": { + "properties": {"a": {"type": "number"}} + }, + "bar": { + "properties": {"b": {"type": "number"}} + } + }, + "type": "object" + })""", + R"""( + alternative-0 ::= foo + alternative-1 ::= bar + bar ::= "{" space (bar-b-kv )? "}" space + bar-b-kv ::= "\"b\"" space ":" space number + foo ::= "{" space (foo-a-kv )? "}" space + foo-a-kv ::= "\"a\"" space ":" space number + number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space + root ::= alternative-0 | alternative-1 + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "mix of allOf, anyOf and $ref (similar to https://json.schemastore.org/tsconfig.json)", + R"""({ + "allOf": [ + {"$ref": "#/definitions/foo"}, + {"$ref": "#/definitions/bar"}, + { + "anyOf": [ + {"$ref": "#/definitions/baz"}, + {"$ref": "#/definitions/bam"} + ] + } + ], + "definitions": { + "foo": { + "properties": {"a": {"type": "number"}} + }, + "bar": { + "properties": {"b": {"type": "number"}} + }, + "bam": { + "properties": {"c": {"type": "number"}} + }, + "baz": { + "properties": {"d": {"type": "number"}} + } + }, + "type": "object" + })""", + R"""( + a-kv ::= "\"a\"" space ":" space number + b-kv ::= "\"b\"" space ":" space number + c-kv ::= "\"c\"" space ":" space number + d-kv ::= "\"d\"" space ":" space number + d-rest ::= ( "," space c-kv )? + number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space + root ::= "{" space a-kv "," space b-kv ( "," space ( d-kv d-rest | c-kv ) )? "}" space + space ::= " "? + )""" + }); + + test({ + SUCCESS, + "conflicting names", + R"""({ + "type": "object", + "properties": { + "number": { + "type": "object", + "properties": { + "number": { + "type": "object", + "properties": { + "root": { + "type": "number" + } + }, + "required": [ + "root" + ], + "additionalProperties": false + } + }, + "required": [ + "number" + ], + "additionalProperties": false + } + }, + "required": [ + "number" + ], + "additionalProperties": false, + "definitions": {} + })""", + R"""( + number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space + number- ::= "{" space number-number-kv "}" space + number-kv ::= "\"number\"" space ":" space number- + number-number ::= "{" space number-number-root-kv "}" space + number-number-kv ::= "\"number\"" space ":" space number-number + number-number-root-kv ::= "\"root\"" space ":" space number + root ::= "{" space number-kv "}" space + space ::= " "? + )""" + }); +} + +int main() { + fprintf(stderr, "LLAMA_NODE_AVAILABLE = %s\n", getenv("LLAMA_NODE_AVAILABLE") ? "true" : "false"); + fprintf(stderr, "LLAMA_PYTHON_AVAILABLE = %s\n", getenv("LLAMA_PYTHON_AVAILABLE") ? "true" : "false"); + + test_all("C++", [](const TestCase & tc) { + try { + tc.verify(json_schema_to_grammar(nlohmann::ordered_json::parse(tc.schema))); + tc.verify_status(SUCCESS); + } catch (const std::runtime_error & ex) { + fprintf(stderr, "Error: %s\n", ex.what()); + tc.verify_status(FAILURE); + } + }); + + if (getenv("LLAMA_PYTHON_AVAILABLE") || (std::system("python --version") == 0)) { + test_all("Python", [](const TestCase & tc) { + write("test-json-schema-input.tmp", tc.schema); + tc.verify_status(std::system( + "python ./examples/json-schema-to-grammar.py test-json-schema-input.tmp > test-grammar-output.tmp") == 0 ? SUCCESS : FAILURE); + tc.verify(read("test-grammar-output.tmp")); + }); + } else { + fprintf(stderr, "\033[33mWARNING: Python not found, skipping Python JSON schema -> grammar tests.\n\033[0m"); + } + + if (getenv("LLAMA_NODE_AVAILABLE") || (std::system("node --version") == 0)) { + test_all("JavaScript", [](const TestCase & tc) { + write("test-json-schema-input.tmp", tc.schema); + tc.verify_status(std::system( + "node ./tests/run-json-schema-to-grammar.mjs test-json-schema-input.tmp > test-grammar-output.tmp") == 0 ? SUCCESS : FAILURE); + tc.verify(read("test-grammar-output.tmp")); + }); + } else { + fprintf(stderr, "\033[33mWARNING: Node not found, skipping JavaScript JSON schema -> grammar tests.\n\033[0m"); + } + + test_all("Check Expectations Validity", [](const TestCase & tc) { + if (tc.expected_status == SUCCESS) { + tc.verify_expectation_parseable(); + } + }); +} diff --git a/tests/test-quantize-fns.cpp b/tests/test-quantize-fns.cpp index f615b612d..e690ac6c8 100644 --- a/tests/test-quantize-fns.cpp +++ b/tests/test-quantize-fns.cpp @@ -138,11 +138,6 @@ int main(int argc, char * argv[]) { const ggml_type ei = (ggml_type)i; - if (ei == GGML_TYPE_IQ2_XXS || ei == GGML_TYPE_IQ2_XS) { - printf("Skip %s due to missing quantization functionality\n", ggml_type_name(ei)); - continue; - } - printf("Testing %s\n", ggml_type_name((ggml_type) i)); ggml_quantize_init(ei);