diff --git a/.devops/full-cuda.Dockerfile b/.devops/full-cuda.Dockerfile index 2a7da586a..b8a354246 100644 --- a/.devops/full-cuda.Dockerfile +++ b/.devops/full-cuda.Dockerfile @@ -1,18 +1,16 @@ ARG UBUNTU_VERSION=22.04 - # This needs to generally match the container host's environment. -ARG CUDA_VERSION=11.7.1 - +ARG CUDA_VERSION=12.6.0 # Target the CUDA build image ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION} -FROM ${BASE_CUDA_DEV_CONTAINER} as build +FROM ${BASE_CUDA_DEV_CONTAINER} AS build -# Unless otherwise specified, we make a fat build. -ARG CUDA_DOCKER_ARCH=all +# CUDA architecture to build for (defaults to all supported archs) +ARG CUDA_DOCKER_ARCH=default RUN apt-get update && \ - apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1 + apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp1 COPY requirements.txt requirements.txt COPY requirements requirements @@ -24,13 +22,12 @@ WORKDIR /app COPY . . -# Set nvcc architecture -ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH} -# Enable CUDA -ENV GGML_CUDA=1 -# Enable cURL -ENV LLAMA_CURL=1 - -RUN make -j$(nproc) +# Use the default CUDA archs if not specified +RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \ + export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \ + fi && \ + cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \ + cmake --build build --config Release --target llama-cli -j$(nproc) && \ + cp build/bin/* . ENTRYPOINT ["/app/.devops/tools.sh"] diff --git a/.devops/full-rocm.Dockerfile b/.devops/full-rocm.Dockerfile index 5cbd2e7a1..680d1cb92 100644 --- a/.devops/full-rocm.Dockerfile +++ b/.devops/full-rocm.Dockerfile @@ -6,7 +6,7 @@ ARG ROCM_VERSION=5.6 # Target the CUDA build image ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete -FROM ${BASE_ROCM_DEV_CONTAINER} as build +FROM ${BASE_ROCM_DEV_CONTAINER} AS build # Unless otherwise specified, we make a fat build. # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878 diff --git a/.devops/full.Dockerfile b/.devops/full.Dockerfile index 6f19afa9c..2a06f82b7 100644 --- a/.devops/full.Dockerfile +++ b/.devops/full.Dockerfile @@ -1,6 +1,6 @@ ARG UBUNTU_VERSION=22.04 -FROM ubuntu:$UBUNTU_VERSION as build +FROM ubuntu:$UBUNTU_VERSION AS build RUN apt-get update && \ apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1 diff --git a/.devops/llama-cli-cann.Dockerfile b/.devops/llama-cli-cann.Dockerfile new file mode 100644 index 000000000..db5ba2f25 --- /dev/null +++ b/.devops/llama-cli-cann.Dockerfile @@ -0,0 +1,44 @@ +ARG ASCEND_VERSION=8.0.rc2.alpha003-910b-openeuler22.03-py3.8 + +FROM cosdt/cann:$ASCEND_VERSION AS build + +WORKDIR /app + +COPY . . + +RUN yum install -y gcc g++ cmake make +ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest +ENV LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:$LIBRARY_PATH +ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling:${LD_LIBRARY_PATH} +ENV PYTHONPATH=${ASCEND_TOOLKIT_HOME}/python/site-packages:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe:${PYTHONPATH} +ENV PATH=${ASCEND_TOOLKIT_HOME}/bin:${ASCEND_TOOLKIT_HOME}/compiler/ccec_compiler/bin:${PATH} +ENV ASCEND_AICPU_PATH=${ASCEND_TOOLKIT_HOME} +ENV ASCEND_OPP_PATH=${ASCEND_TOOLKIT_HOME}/opp +ENV TOOLCHAIN_HOME=${ASCEND_TOOLKIT_HOME}/toolkit +ENV ASCEND_HOME_PATH=${ASCEND_TOOLKIT_HOME} + +# find libascend_hal.so, because the drive hasn`t been mounted. +ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH + +RUN echo "Building with static libs" && \ + source /usr/local/Ascend/ascend-toolkit/set_env.sh --force && \ + cmake -B build -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF && \ + cmake --build build --config Release --target llama-cli + +# TODO: use image with NNRT +FROM cosdt/cann:$ASCEND_VERSION AS runtime +COPY --from=build /app/build/bin/llama-cli /llama-cli + +ENV LC_ALL=C.utf8 + +ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest +ENV LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:$LIBRARY_PATH +ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling:${LD_LIBRARY_PATH} +ENV PYTHONPATH=${ASCEND_TOOLKIT_HOME}/python/site-packages:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe:${PYTHONPATH} +ENV PATH=${ASCEND_TOOLKIT_HOME}/bin:${ASCEND_TOOLKIT_HOME}/compiler/ccec_compiler/bin:${PATH} +ENV ASCEND_AICPU_PATH=${ASCEND_TOOLKIT_HOME} +ENV ASCEND_OPP_PATH=${ASCEND_TOOLKIT_HOME}/opp +ENV TOOLCHAIN_HOME=${ASCEND_TOOLKIT_HOME}/toolkit +ENV ASCEND_HOME_PATH=${ASCEND_TOOLKIT_HOME} + +ENTRYPOINT ["/llama-cli" ] diff --git a/.devops/llama-cli-cuda.Dockerfile b/.devops/llama-cli-cuda.Dockerfile index bff946cbc..b75163b94 100644 --- a/.devops/llama-cli-cuda.Dockerfile +++ b/.devops/llama-cli-cuda.Dockerfile @@ -1,35 +1,37 @@ ARG UBUNTU_VERSION=22.04 # This needs to generally match the container host's environment. -ARG CUDA_VERSION=11.7.1 +ARG CUDA_VERSION=12.6.0 # Target the CUDA build image ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION} # Target the CUDA runtime image ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION} -FROM ${BASE_CUDA_DEV_CONTAINER} as build +FROM ${BASE_CUDA_DEV_CONTAINER} AS build -# Unless otherwise specified, we make a fat build. -ARG CUDA_DOCKER_ARCH=all +# CUDA architecture to build for (defaults to all supported archs) +ARG CUDA_DOCKER_ARCH=default RUN apt-get update && \ - apt-get install -y build-essential git + apt-get install -y build-essential git cmake WORKDIR /app COPY . . -# Set nvcc architecture -ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH} -# Enable CUDA -ENV GGML_CUDA=1 +# Use the default CUDA archs if not specified +RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \ + export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \ + fi && \ + cmake -B build -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \ + cmake --build build --config Release --target llama-cli -j$(nproc) -RUN make -j$(nproc) llama-cli - -FROM ${BASE_CUDA_RUN_CONTAINER} as runtime +FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime RUN apt-get update && \ apt-get install -y libgomp1 -COPY --from=build /app/llama-cli /llama-cli +COPY --from=build /app/build/ggml/src/libggml.so /libggml.so +COPY --from=build /app/build/src/libllama.so /libllama.so +COPY --from=build /app/build/bin/llama-cli /llama-cli ENTRYPOINT [ "/llama-cli" ] diff --git a/.devops/llama-cli-intel.Dockerfile b/.devops/llama-cli-intel.Dockerfile index bd816f9f5..79dba06a7 100644 --- a/.devops/llama-cli-intel.Dockerfile +++ b/.devops/llama-cli-intel.Dockerfile @@ -1,6 +1,6 @@ ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04 -FROM intel/oneapi-basekit:$ONEAPI_VERSION as build +FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build ARG GGML_SYCL_F16=OFF RUN apt-get update && \ @@ -14,10 +14,12 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \ echo "GGML_SYCL_F16 is set" && \ export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \ fi && \ - cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \ + echo "Building with static libs" && \ + cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx \ + ${OPT_SYCL_F16} -DBUILD_SHARED_LIBS=OFF && \ cmake --build build --config Release --target llama-cli -FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime +FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime COPY --from=build /app/build/bin/llama-cli /llama-cli diff --git a/.devops/llama-cli-rocm.Dockerfile b/.devops/llama-cli-rocm.Dockerfile index caa507b08..c3d1ab067 100644 --- a/.devops/llama-cli-rocm.Dockerfile +++ b/.devops/llama-cli-rocm.Dockerfile @@ -6,7 +6,7 @@ ARG ROCM_VERSION=5.6 # Target the CUDA build image ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete -FROM ${BASE_ROCM_DEV_CONTAINER} as build +FROM ${BASE_ROCM_DEV_CONTAINER} AS build # Unless otherwise specified, we make a fat build. # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878 diff --git a/.devops/llama-cli-vulkan.Dockerfile b/.devops/llama-cli-vulkan.Dockerfile index 6155d5881..9b0dad8bf 100644 --- a/.devops/llama-cli-vulkan.Dockerfile +++ b/.devops/llama-cli-vulkan.Dockerfile @@ -1,6 +1,6 @@ ARG UBUNTU_VERSION=jammy -FROM ubuntu:$UBUNTU_VERSION as build +FROM ubuntu:$UBUNTU_VERSION AS build # Install build tools RUN apt update && apt install -y git build-essential cmake wget libgomp1 diff --git a/.devops/llama-cli.Dockerfile b/.devops/llama-cli.Dockerfile index 38382bfc9..7f741aa46 100644 --- a/.devops/llama-cli.Dockerfile +++ b/.devops/llama-cli.Dockerfile @@ -1,6 +1,6 @@ ARG UBUNTU_VERSION=22.04 -FROM ubuntu:$UBUNTU_VERSION as build +FROM ubuntu:$UBUNTU_VERSION AS build RUN apt-get update && \ apt-get install -y build-essential git @@ -11,7 +11,7 @@ COPY . . RUN make -j$(nproc) llama-cli -FROM ubuntu:$UBUNTU_VERSION as runtime +FROM ubuntu:$UBUNTU_VERSION AS runtime RUN apt-get update && \ apt-get install -y libgomp1 diff --git a/.devops/llama-server-cuda.Dockerfile b/.devops/llama-server-cuda.Dockerfile index d7eaa0925..a40e24205 100644 --- a/.devops/llama-server-cuda.Dockerfile +++ b/.devops/llama-server-cuda.Dockerfile @@ -1,38 +1,41 @@ ARG UBUNTU_VERSION=22.04 # This needs to generally match the container host's environment. -ARG CUDA_VERSION=11.7.1 +ARG CUDA_VERSION=12.6.0 # Target the CUDA build image ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION} # Target the CUDA runtime image ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION} -FROM ${BASE_CUDA_DEV_CONTAINER} as build +FROM ${BASE_CUDA_DEV_CONTAINER} AS build -# Unless otherwise specified, we make a fat build. -ARG CUDA_DOCKER_ARCH=all +# CUDA architecture to build for (defaults to all supported archs) +ARG CUDA_DOCKER_ARCH=default RUN apt-get update && \ - apt-get install -y build-essential git libcurl4-openssl-dev + apt-get install -y build-essential git cmake libcurl4-openssl-dev WORKDIR /app COPY . . -# Set nvcc architecture -ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH} -# Enable CUDA -ENV GGML_CUDA=1 -# Enable cURL -ENV LLAMA_CURL=1 +# Use the default CUDA archs if not specified +RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \ + export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \ + fi && \ + cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \ + cmake --build build --config Release --target llama-server -j$(nproc) -RUN make -j$(nproc) llama-server - -FROM ${BASE_CUDA_RUN_CONTAINER} as runtime +FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime RUN apt-get update && \ apt-get install -y libcurl4-openssl-dev libgomp1 curl -COPY --from=build /app/llama-server /llama-server +COPY --from=build /app/build/ggml/src/libggml.so /libggml.so +COPY --from=build /app/build/src/libllama.so /libllama.so +COPY --from=build /app/build/bin/llama-server /llama-server + +# Must be set to 0.0.0.0 so it can listen to requests from host machine +ENV LLAMA_ARG_HOST=0.0.0.0 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ] diff --git a/.devops/llama-server-intel.Dockerfile b/.devops/llama-server-intel.Dockerfile index 8f8fef8c0..9c355b664 100644 --- a/.devops/llama-server-intel.Dockerfile +++ b/.devops/llama-server-intel.Dockerfile @@ -1,6 +1,6 @@ ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04 -FROM intel/oneapi-basekit:$ONEAPI_VERSION as build +FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build ARG GGML_SYCL_F16=OFF RUN apt-get update && \ @@ -14,10 +14,11 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \ echo "GGML_SYCL_F16 is set" && \ export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \ fi && \ + echo "Building with dynamic libs" && \ cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \ cmake --build build --config Release --target llama-server -FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime +FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime RUN apt-get update && \ apt-get install -y libcurl4-openssl-dev curl @@ -25,6 +26,8 @@ RUN apt-get update && \ COPY --from=build /app/build/bin/llama-server /llama-server ENV LC_ALL=C.utf8 +# Must be set to 0.0.0.0 so it can listen to requests from host machine +ENV LLAMA_ARG_HOST=0.0.0.0 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ] diff --git a/.devops/llama-server-rocm.Dockerfile b/.devops/llama-server-rocm.Dockerfile index af96c3325..fd0e19ad6 100644 --- a/.devops/llama-server-rocm.Dockerfile +++ b/.devops/llama-server-rocm.Dockerfile @@ -6,7 +6,7 @@ ARG ROCM_VERSION=5.6 # Target the CUDA build image ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete -FROM ${BASE_ROCM_DEV_CONTAINER} as build +FROM ${BASE_ROCM_DEV_CONTAINER} AS build # Unless otherwise specified, we make a fat build. # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878 @@ -39,6 +39,8 @@ ENV GPU_TARGETS=${ROCM_DOCKER_ARCH} ENV GGML_HIPBLAS=1 ENV CC=/opt/rocm/llvm/bin/clang ENV CXX=/opt/rocm/llvm/bin/clang++ +# Must be set to 0.0.0.0 so it can listen to requests from host machine +ENV LLAMA_ARG_HOST=0.0.0.0 # Enable cURL ENV LLAMA_CURL=1 diff --git a/.devops/llama-server-vulkan.Dockerfile b/.devops/llama-server-vulkan.Dockerfile index 49062f84b..93c5e0c26 100644 --- a/.devops/llama-server-vulkan.Dockerfile +++ b/.devops/llama-server-vulkan.Dockerfile @@ -1,6 +1,6 @@ ARG UBUNTU_VERSION=jammy -FROM ubuntu:$UBUNTU_VERSION as build +FROM ubuntu:$UBUNTU_VERSION AS build # Install build tools RUN apt update && apt install -y git build-essential cmake wget @@ -23,6 +23,8 @@ RUN cp /app/build/bin/llama-server /llama-server && \ rm -rf /app ENV LC_ALL=C.utf8 +# Must be set to 0.0.0.0 so it can listen to requests from host machine +ENV LLAMA_ARG_HOST=0.0.0.0 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ] diff --git a/.devops/llama-server.Dockerfile b/.devops/llama-server.Dockerfile index a53a5c999..02accc85e 100644 --- a/.devops/llama-server.Dockerfile +++ b/.devops/llama-server.Dockerfile @@ -1,9 +1,9 @@ ARG UBUNTU_VERSION=22.04 -FROM ubuntu:$UBUNTU_VERSION as build +FROM ubuntu:$UBUNTU_VERSION AS build RUN apt-get update && \ - apt-get install -y build-essential git libcurl4-openssl-dev curl + apt-get install -y build-essential git libcurl4-openssl-dev WORKDIR /app @@ -13,14 +13,16 @@ ENV LLAMA_CURL=1 RUN make -j$(nproc) llama-server -FROM ubuntu:$UBUNTU_VERSION as runtime +FROM ubuntu:$UBUNTU_VERSION AS runtime RUN apt-get update && \ - apt-get install -y libcurl4-openssl-dev libgomp1 + apt-get install -y libcurl4-openssl-dev libgomp1 curl COPY --from=build /app/llama-server /llama-server ENV LC_ALL=C.utf8 +# Must be set to 0.0.0.0 so it can listen to requests from host machine +ENV LLAMA_ARG_HOST=0.0.0.0 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ] diff --git a/.devops/nix/apps.nix b/.devops/nix/apps.nix index 897fce4d3..0ecf19fc5 100644 --- a/.devops/nix/apps.nix +++ b/.devops/nix/apps.nix @@ -10,7 +10,6 @@ "llama-embedding" "llama-server" "llama-quantize" - "llama-train-text-from-scratch" ]; mkApp = name: { type = "app"; diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index 49e9b7528..cfffac257 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -13,11 +13,13 @@ mpi, blas, cudaPackages, + autoAddDriverRunpath, darwin, rocmPackages, vulkan-headers, vulkan-loader, curl, + shaderc, useBlas ? builtins.all (x: !x) [ useCuda useMetalKit @@ -89,6 +91,22 @@ let ps.tiktoken ps.torchWithoutCuda ps.transformers + + # server bench + ps.matplotlib + + # server tests + ps.openai + ps.behave + ps.prometheus-client + + # for examples/pydantic-models-to-grammar-examples.py + ps.docstring-parser + ps.pydantic + + # for scripts/compare-llama-bench.py + ps.gitpython + ps.tabulate ] ); @@ -109,16 +127,9 @@ let ++ optionals useMetalKit [ MetalKit ]; cudaBuildInputs = with cudaPackages; [ - cuda_cccl.dev # - - # A temporary hack for reducing the closure size, remove once cudaPackages - # have stopped using lndir: https://github.com/NixOS/nixpkgs/issues/271792 - cuda_cudart.dev - cuda_cudart.lib - cuda_cudart.static - libcublas.dev - libcublas.lib - libcublas.static + cuda_cudart + cuda_cccl # + libcublas ]; rocmBuildInputs = with rocmPackages; [ @@ -130,6 +141,7 @@ let vulkanBuildInputs = [ vulkan-headers vulkan-loader + shaderc ]; in @@ -181,10 +193,7 @@ effectiveStdenv.mkDerivation ( ] ++ optionals useCuda [ cudaPackages.cuda_nvcc - - # TODO: Replace with autoAddDriverRunpath - # once https://github.com/NixOS/nixpkgs/pull/275241 has been merged - cudaPackages.autoAddOpenGLRunpathHook + autoAddDriverRunpath ] ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [ glibc.static diff --git a/.devops/tools.sh b/.devops/tools.sh index 335382f69..24dcfd350 100755 --- a/.devops/tools.sh +++ b/.devops/tools.sh @@ -8,13 +8,11 @@ arg1="$1" shift if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then - python3 ./convert-hf-to-gguf.py "$@" + python3 ./convert_hf_to_gguf.py "$@" elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then ./llama-quantize "$@" elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then ./llama-cli "$@" -elif [[ "$arg1" == '--finetune' || "$arg1" == '-f' ]]; then - ./llama-finetune "$@" elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then echo "Converting PTH to GGML..." for i in `ls $1/$2/ggml-model-f16.bin*`; do @@ -36,8 +34,6 @@ else echo " ex: --outtype f16 \"/models/7B/\" " echo " --quantize (-q): Optimize with quantization process ggml" echo " ex: \"/models/7B/ggml-model-f16.bin\" \"/models/7B/ggml-model-q4_0.bin\" 2" - echo " --finetune (-f): Run finetune command to create a lora finetune of the model" - echo " See documentation for finetune for command-line parameters" echo " --all-in-one (-a): Execute --convert & --quantize" echo " ex: \"/models/\" 7B" echo " --server (-s): Run a model on the server" diff --git a/.ecrc b/.ecrc index a3351f4e6..c68877ec2 100644 --- a/.ecrc +++ b/.ecrc @@ -1,5 +1,5 @@ { - "Exclude": ["^\\.gitmodules$"], + "Exclude": ["^\\.gitmodules$", "stb_image\\.h"], "Disable": { "IndentSize": true } diff --git a/.github/labeler.yml b/.github/labeler.yml index 9c0397d16..89436740d 100644 --- a/.github/labeler.yml +++ b/.github/labeler.yml @@ -16,7 +16,9 @@ SYCL: - any-glob-to-any-file: - ggml/include/ggml-sycl.h - ggml/src/ggml-sycl.cpp - - README-sycl.md + - ggml/src/ggml-sycl/** + - docs/backend/SYCL.md + - examples/sycl/** Nvidia GPU: - changed-files: - any-glob-to-any-file: diff --git a/.github/workflows/bench.yml b/.github/workflows/bench.yml.disabled similarity index 97% rename from .github/workflows/bench.yml rename to .github/workflows/bench.yml.disabled index eb69b82c4..bfdbb4ef5 100644 --- a/.github/workflows/bench.yml +++ b/.github/workflows/bench.yml.disabled @@ -1,3 +1,6 @@ +# TODO: there have been some issues with the workflow, so disabling for now +# https://github.com/ggerganov/llama.cpp/issues/7893 +# # Benchmark name: Benchmark @@ -129,6 +132,8 @@ jobs: - name: Server bench id: server_bench + env: + HEAD_REF: ${{ github.head_ref || github.ref_name }} run: | set -eux @@ -137,7 +142,7 @@ jobs: python bench.py \ --runner-label ${{ env.RUNNER_LABEL }} \ --name ${{ github.job }} \ - --branch ${{ github.head_ref || github.ref_name }} \ + --branch $HEAD_REF \ --commit ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha }} \ --scenario script.js \ --duration ${{ github.event.inputs.duration || env.DURATION }} \ diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 1e344db6b..74b5d4f69 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -47,7 +47,7 @@ jobs: sysctl -a mkdir build cd build - cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF .. + cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF .. cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) - name: Test @@ -105,7 +105,7 @@ jobs: sysctl -a # Metal is disabled due to intermittent failures with Github runners not having a GPU: # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313 - cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF + cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) - name: Test @@ -222,7 +222,7 @@ jobs: run: | mkdir build cd build - cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF + cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF cmake --build . --config Release -j $(nproc) - name: Test @@ -355,8 +355,10 @@ jobs: - name: Dependencies id: depends run: | - sudo apt-get update - sudo apt-get install build-essential libvulkan-dev + wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add - + sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list + sudo apt-get update -y + sudo apt-get install -y build-essential vulkan-sdk - name: Build id: cmake_build @@ -694,22 +696,20 @@ jobs: strategy: matrix: include: - - build: 'rpc-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON' - build: 'noavx-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON' + defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON' - build: 'avx2-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON' + defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON' - build: 'avx-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON' + defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON' - build: 'avx512-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON' + defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON' - build: 'openblas-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"' + defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"' - build: 'kompute-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON' + defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON' - build: 'vulkan-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON' + defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON' - build: 'llvm-arm64' defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON' - build: 'msvc-arm64' @@ -858,6 +858,7 @@ jobs: mkdir build cd build cmake .. -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON + cmake --build . --config Release -j $((${env:NUMBER_OF_PROCESSORS} - 1)) -t ggml cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS} - name: Determine tag name diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml index bf94b2024..56fefd93d 100644 --- a/.github/workflows/docker.yml +++ b/.github/workflows/docker.yml @@ -96,21 +96,12 @@ jobs: env: GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}' - - name: Build and push Docker image (versioned) + - name: Build and push Docker image (tagged + versioned) if: github.event_name == 'push' - uses: docker/build-push-action@v4 + uses: docker/build-push-action@v6 with: context: . push: true platforms: ${{ matrix.config.platforms }} - tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}" - file: ${{ matrix.config.dockerfile }} - - - name: Build and push Docker image (tagged) - uses: docker/build-push-action@v4 - with: - context: . - push: ${{ github.event_name == 'push' }} - platforms: ${{ matrix.config.platforms }} - tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}" + tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}" file: ${{ matrix.config.dockerfile }} diff --git a/.github/workflows/python-check-requirements.yml b/.github/workflows/python-check-requirements.yml index 4e0374fc6..46e80aecd 100644 --- a/.github/workflows/python-check-requirements.yml +++ b/.github/workflows/python-check-requirements.yml @@ -6,15 +6,13 @@ on: - '.github/workflows/python-check-requirements.yml' - 'scripts/check-requirements.sh' - 'convert*.py' - - 'requirements.txt' - - 'requirements/*.txt' + - '**/requirements*.txt' pull_request: paths: - '.github/workflows/python-check-requirements.yml' - 'scripts/check-requirements.sh' - 'convert*.py' - - 'requirements.txt' - - 'requirements/*.txt' + - '**/requirements*.txt' concurrency: group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }} diff --git a/.github/workflows/python-type-check.yml b/.github/workflows/python-type-check.yml new file mode 100644 index 000000000..e5ff5e6d7 --- /dev/null +++ b/.github/workflows/python-type-check.yml @@ -0,0 +1,38 @@ +name: Python Type-Check + +on: + push: + paths: + - '.github/workflows/python-type-check.yml' + - '**.py' + - '**/requirements*.txt' + pull_request: + paths: + - '.github/workflows/python-type-check.yml' + - '**.py' + - '**/requirements*.txt' + +concurrency: + group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }} + cancel-in-progress: true + +jobs: + python-type-check: + runs-on: ubuntu-latest + name: pyright type-check + steps: + - name: Check out source repository + uses: actions/checkout@v4 + - name: Set up Python environment + uses: actions/setup-python@v5 + with: + python-version: "3.11" + - name: Install Python dependencies + # TODO: use a venv + run: pip install -r requirements/requirements-all.txt + - name: Type-check with Pyright + uses: jakebailey/pyright-action@v2 + with: + version: 1.1.370 + level: warning + warnings: true diff --git a/.gitignore b/.gitignore index 6125e350c..9986ac6b1 100644 --- a/.gitignore +++ b/.gitignore @@ -47,8 +47,10 @@ build* !build-info.cpp.in !build-info.sh !build.zig +!docs/build.md /libllama.so /llama-* +/vulkan-shaders-gen android-ndk-* arm_neon.h cmake-build-* @@ -60,6 +62,11 @@ llama-batched-swift out/ tmp/ +# Deprecated + +/main +/server + # CI !.github/workflows/*.yml @@ -72,7 +79,6 @@ models-mnt !models/ggml-vocab-*.gguf* # Zig - zig-out/ zig-cache/ @@ -123,3 +129,6 @@ poetry.toml # Scripts !/scripts/install-oneapi.bat + +# Test models for lora adapters +/lora-tests diff --git a/CMakeLists.txt b/CMakeLists.txt index 4f6cd6872..a31320635 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -50,9 +50,6 @@ endif() # option list # -# general -option(LLAMA_CCACHE "llama: use ccache if available" ON) - # debug option(LLAMA_ALL_WARNINGS "llama: enable all compiler warnings" ON) option(LLAMA_ALL_WARNINGS_3RD_PARTY "llama: enable all compiler warnings in 3rd party libs" OFF) @@ -77,7 +74,6 @@ option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF) include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake) # override ggml options -set(GGML_CCACHE ${LLAMA_CCACHE}) set(GGML_SANITIZE_THREAD ${LLAMA_SANITIZE_THREAD}) set(GGML_SANITIZE_ADDRESS ${LLAMA_SANITIZE_ADDRESS}) set(GGML_SANITIZE_UNDEFINED ${LLAMA_SANITIZE_UNDEFINED}) @@ -110,12 +106,16 @@ llama_option_depr(WARNING LLAMA_NATIVE GGML_NATIVE) llama_option_depr(WARNING LLAMA_RPC GGML_RPC) llama_option_depr(WARNING LLAMA_SYCL GGML_SYCL) llama_option_depr(WARNING LLAMA_SYCL_F16 GGML_SYCL_F16) +llama_option_depr(WARNING LLAMA_CANN GGML_CANN) # # build the library # -add_subdirectory(ggml) +if (NOT TARGET ggml) + add_subdirectory(ggml) + # ... otherwise assume ggml is added by a parent CMakeLists.txt +endif() add_subdirectory(src) # @@ -133,7 +133,17 @@ set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location o set(LLAMA_LIB_INSTALL_DIR ${CMAKE_INSTALL_LIBDIR} CACHE PATH "Location of library files") set(LLAMA_BIN_INSTALL_DIR ${CMAKE_INSTALL_BINDIR} CACHE PATH "Location of binary files") -get_directory_property(LLAMA_TRANSIENT_DEFINES COMPILE_DEFINITIONS) + +# At the moment some compile definitions are placed within the ggml/src +# directory but not exported on the `ggml` target. This could be improved by +# determining _precisely_ which defines are necessary for the llama-config +# package. +# +get_target_property(GGML_DIRECTORY ggml SOURCE_DIR) +get_directory_property(GGML_DIR_DEFINES DIRECTORY ${GGML_DIRECTORY} COMPILE_DEFINITIONS) +get_target_property(GGML_TARGET_DEFINES ggml COMPILE_DEFINITIONS) +set(GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES} ${GGML_DIR_DEFINES}) +get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES) set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h) install(TARGETS llama LIBRARY PUBLIC_HEADER) diff --git a/CMakePresets.json b/CMakePresets.json index bdad38952..ce627b4d3 100644 --- a/CMakePresets.json +++ b/CMakePresets.json @@ -28,6 +28,7 @@ { "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } }, { "name": "reldbg", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } }, { "name": "static", "hidden": true, "cacheVariables": { "GGML_STATIC": "ON" } }, + { "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16": "ON" } }, { "name": "arm64-windows-msvc", "hidden": true, @@ -60,6 +61,8 @@ { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] }, { "name": "x64-windows-sycl-debug" , "inherits": [ "sycl-base", "debug" ] }, - { "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] } + { "name": "x64-windows-sycl-debug-f16", "inherits": [ "sycl-base", "debug", "sycl_f16" ] }, + { "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] }, + { "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] } ] } diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md index 991d85e49..a9e000e52 100644 --- a/CONTRIBUTING.md +++ b/CONTRIBUTING.md @@ -1,14 +1,29 @@ -# Contributing Guidelines +# Pull requests (for contributors) -## Checklist +- Test your changes: + - Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the GGML library + - Execute [the full CI locally on your machine](ci/README.md) before publishing +- Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs. + - The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your convenience +- Consider allowing write access to your branch for faster review +- If your PR becomes stale, don't hesitate to ping the maintainers in the comments -* Make sure your PR follows the [coding guidelines](https://github.com/ggerganov/llama.cpp/blob/master/README.md#coding-guidelines) -* Test your changes using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the GGML library -* Execute [the full CI locally on your machine](ci/README.md) before publishing +# Pull requests (for collaborators) -## PR formatting +- Squash-merge PRs +- Use the following format for the squashed commit title: ` : (#)`. For example: `utils : fix typo in utils.py (#1234)` +- Optionally, pick a `` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules + +# Coding guidelines + +- Avoid adding third-party dependencies, extra files, extra headers, etc. +- Always consider cross-compatibility with other operating systems and architectures +- Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple +- There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit +- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a` +- Naming usually optimizes for common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963) +- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices +- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$ + +![matmul](media/matmul.png) -* Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs. - - The PR template has a series of review complexity checkboxes `[ ]` that you can mark as `[X]` for your conveience. Refer to [About task lists](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) for more information. -* If the pull request only contains documentation changes (e.g., updating READMEs, adding new wiki pages), please add `[no ci]` to the commit title. This will skip unnecessary CI checks and help reduce build times. -* When squashing multiple commits on merge, use the following format for your commit title: ` : (#)`. For example: `utils : Fix typo in utils.py (#1234)` diff --git a/Makefile b/Makefile index 2730d8b60..332496cfc 100644 --- a/Makefile +++ b/Makefile @@ -11,14 +11,15 @@ BUILD_TARGETS = \ llama-embedding \ llama-eval-callback \ llama-export-lora \ - llama-finetune \ llama-gbnf-validator \ llama-gguf \ + llama-gguf-hash \ llama-gguf-split \ llama-gritlm \ llama-imatrix \ llama-infill \ llama-llava-cli \ + llama-minicpmv-cli\ llama-lookahead \ llama-lookup \ llama-lookup-create \ @@ -36,7 +37,6 @@ BUILD_TARGETS = \ llama-simple \ llama-speculative \ llama-tokenize \ - llama-train-text-from-scratch \ llama-vdot \ llama-cvector-generator \ tests/test-c.o @@ -63,9 +63,13 @@ TEST_TARGETS = \ tests/test-tokenizer-1-spm # Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned -LEGACY_TARGETS = main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \ +LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot convert-llama2c-to-ggml \ simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama \ - retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm + retrieval speculative infill tokenize benchmark-matmult parallel export-lora lookahead lookup passkey gritlm + +# Legacy build targets that were renamed in #7809, but we want to build binaries that for them that output a deprecation warning if people try to use them. +# We don't want to clutter things too much, so we only build replacements for the most commonly used binaries. +LEGACY_TARGETS_BUILD = main quantize perplexity embedding server # Deprecation aliases ifdef LLAMA_CUBLAS @@ -192,7 +196,11 @@ ifdef GGML_RPC BUILD_TARGETS += rpc-server endif -default: $(BUILD_TARGETS) +ifdef GGML_VULKAN + BUILD_TARGETS += vulkan-shaders-gen +endif + +default: $(BUILD_TARGETS) $(LEGACY_TARGETS_BUILD) test: $(TEST_TARGETS) @failures=0; \ @@ -227,7 +235,7 @@ test: $(TEST_TARGETS) fi @echo 'All tests passed.' -all: $(BUILD_TARGETS) $(TEST_TARGETS) +all: $(BUILD_TARGETS) $(TEST_TARGETS) $(LEGACY_TARGETS_BUILD) ifdef RISCV_CROSS_COMPILE CC := riscv64-unknown-linux-gnu-gcc @@ -244,17 +252,22 @@ MK_CFLAGS = -std=c11 -fPIC MK_CXXFLAGS = -std=c++11 -fPIC MK_NVCCFLAGS = -std=c++11 -ifndef LLAMA_NO_CCACHE +ifdef LLAMA_NO_CCACHE +GGML_NO_CCACHE := 1 +DEPRECATE_WARNING := 1 +endif + +ifndef GGML_NO_CCACHE CCACHE := $(shell which ccache) ifdef CCACHE export CCACHE_SLOPPINESS = time_macros -$(info I ccache found, compilation results will be cached. Disable with LLAMA_NO_CCACHE.) +$(info I ccache found, compilation results will be cached. Disable with GGML_NO_CCACHE.) CC := $(CCACHE) $(CC) CXX := $(CCACHE) $(CXX) else $(info I ccache not found. Consider installing it for faster compilation.) endif # CCACHE -endif # LLAMA_NO_CCACHE +endif # GGML_NO_CCACHE # clock_gettime came in POSIX.1b (1993) # CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional @@ -313,9 +326,9 @@ ifdef LLAMA_DEBUG endif else MK_CPPFLAGS += -DNDEBUG - MK_CFLAGS += -O3 - MK_CXXFLAGS += -O3 - MK_NVCCFLAGS += -O3 + MK_CFLAGS += -O3 -g + MK_CXXFLAGS += -O3 -g + MK_NVCCFLAGS += -O3 -g endif ifdef LLAMA_SANITIZE_THREAD @@ -516,10 +529,21 @@ ifndef GGML_NO_ACCELERATE endif endif # GGML_NO_ACCELERATE +ifdef GGML_MUSA + CC := clang + CXX := clang++ + GGML_CUDA := 1 + MK_CPPFLAGS += -DGGML_USE_MUSA +endif + ifndef GGML_NO_OPENMP MK_CPPFLAGS += -DGGML_USE_OPENMP MK_CFLAGS += -fopenmp MK_CXXFLAGS += -fopenmp + ifdef GGML_MUSA + MK_CPPFLAGS += -I/usr/lib/llvm-10/include/openmp + MK_LDFLAGS += -L/usr/lib/llvm-10/lib + endif # GGML_MUSA endif # GGML_NO_OPENMP ifdef GGML_OPENBLAS @@ -537,14 +561,20 @@ ifdef GGML_OPENBLAS64 endif # GGML_OPENBLAS64 ifdef GGML_BLIS - MK_CPPFLAGS += -DGGML_USE_BLAS -I/usr/local/include/blis -I/usr/include/blis + MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_BLIS -I/usr/local/include/blis -I/usr/include/blis MK_LDFLAGS += -lblis -L/usr/local/lib OBJ_GGML += ggml/src/ggml-blas.o endif # GGML_BLIS +ifdef GGML_NVPL + MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_NVPL -DNVPL_ILP64 -I/usr/local/include/nvpl_blas -I/usr/include/nvpl_blas + MK_LDFLAGS += -L/usr/local/lib -lnvpl_blas_core -lnvpl_blas_ilp64_gomp + OBJ_GGML += ggml/src/ggml-blas.o +endif # GGML_NVPL + ifndef GGML_NO_LLAMAFILE MK_CPPFLAGS += -DGGML_USE_LLAMAFILE - OBJ_GGML += ggml/src/sgemm.o + OBJ_GGML += ggml/src/llamafile/sgemm.o endif ifdef GGML_RPC @@ -564,15 +594,27 @@ else endif # GGML_CUDA_FA_ALL_QUANTS ifdef GGML_CUDA - ifneq ('', '$(wildcard /opt/cuda)') - CUDA_PATH ?= /opt/cuda - else - CUDA_PATH ?= /usr/local/cuda - endif + ifdef GGML_MUSA + ifneq ('', '$(wildcard /opt/musa)') + CUDA_PATH ?= /opt/musa + else + CUDA_PATH ?= /usr/local/musa + endif - MK_CPPFLAGS += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS - MK_LDFLAGS += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib - MK_NVCCFLAGS += -use_fast_math + MK_CPPFLAGS += -DGGML_USE_CUDA -I$(CUDA_PATH)/include + MK_LDFLAGS += -lmusa -lmublas -lmusart -lpthread -ldl -lrt -L$(CUDA_PATH)/lib -L/usr/lib64 + MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_22 + else + ifneq ('', '$(wildcard /opt/cuda)') + CUDA_PATH ?= /opt/cuda + else + CUDA_PATH ?= /usr/local/cuda + endif + + MK_CPPFLAGS += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS + MK_LDFLAGS += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib + MK_NVCCFLAGS += -use_fast_math + endif # GGML_MUSA OBJ_GGML += ggml/src/ggml-cuda.o OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu)) @@ -582,9 +624,11 @@ ifdef LLAMA_FATAL_WARNINGS MK_NVCCFLAGS += -Werror all-warnings endif # LLAMA_FATAL_WARNINGS +ifndef GGML_MUSA ifndef JETSON_EOL_MODULE_DETECT MK_NVCCFLAGS += --forward-unknown-to-host-compiler endif # JETSON_EOL_MODULE_DETECT +endif # GGML_MUSA ifdef LLAMA_DEBUG MK_NVCCFLAGS += -lineinfo @@ -597,8 +641,12 @@ endif # GGML_CUDA_DEBUG ifdef GGML_CUDA_NVCC NVCC = $(CCACHE) $(GGML_CUDA_NVCC) else - NVCC = $(CCACHE) nvcc -endif #GGML_CUDA_NVCC + ifdef GGML_MUSA + NVCC = $(CCACHE) mcc + else + NVCC = $(CCACHE) nvcc + endif # GGML_MUSA +endif # GGML_CUDA_NVCC ifdef CUDA_DOCKER_ARCH MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH) @@ -669,9 +717,15 @@ define NVCC_COMPILE $(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -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 + ifdef GGML_MUSA +define NVCC_COMPILE + $(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -c $< -o $@ +endef # NVCC_COMPILE + else define NVCC_COMPILE $(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@ endef # NVCC_COMPILE + endif # GGML_MUSA endif # JETSON_EOL_MODULE_DETECT ggml/src/ggml-cuda/%.o: \ @@ -694,8 +748,8 @@ endif # GGML_CUDA ifdef GGML_VULKAN MK_CPPFLAGS += -DGGML_USE_VULKAN - MK_LDFLAGS += -lvulkan - OBJ_GGML += ggml/src/ggml-vulkan.o + MK_LDFLAGS += $(shell pkg-config --libs vulkan) + OBJ_GGML += ggml/src/ggml-vulkan.o ggml/src/ggml-vulkan-shaders.o ifdef GGML_VULKAN_CHECK_RESULTS MK_CPPFLAGS += -DGGML_VULKAN_CHECK_RESULTS @@ -709,6 +763,10 @@ ifdef GGML_VULKAN_MEMORY_DEBUG MK_CPPFLAGS += -DGGML_VULKAN_MEMORY_DEBUG endif +ifdef GGML_VULKAN_PERF + MK_CPPFLAGS += -DGGML_VULKAN_PERF +endif + ifdef GGML_VULKAN_VALIDATE MK_CPPFLAGS += -DGGML_VULKAN_VALIDATE endif @@ -717,10 +775,28 @@ ifdef GGML_VULKAN_RUN_TESTS MK_CPPFLAGS += -DGGML_VULKAN_RUN_TESTS endif -ggml/src/ggml-vulkan.o: \ - ggml/src/ggml-vulkan.cpp \ - ggml/include/ggml-vulkan.h - $(CXX) $(CXXFLAGS) -c $< -o $@ +GLSLC_CMD = glslc +_ggml_vk_genshaders_cmd = $(shell pwd)/vulkan-shaders-gen +_ggml_vk_header = ggml/src/ggml-vulkan-shaders.hpp +_ggml_vk_source = ggml/src/ggml-vulkan-shaders.cpp +_ggml_vk_input_dir = ggml/src/vulkan-shaders +_ggml_vk_shader_deps = $(echo $(_ggml_vk_input_dir)/*.comp) + +ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source) + $(CXX) $(CXXFLAGS) $(shell pkg-config --cflags vulkan) -c $< -o $@ + +$(_ggml_vk_header): $(_ggml_vk_source) + +$(_ggml_vk_source): $(_ggml_vk_shader_deps) vulkan-shaders-gen + $(_ggml_vk_genshaders_cmd) \ + --glslc $(GLSLC_CMD) \ + --input-dir $(_ggml_vk_input_dir) \ + --target-hpp $(_ggml_vk_header) \ + --target-cpp $(_ggml_vk_source) + +vulkan-shaders-gen: ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp + $(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp + endif # GGML_VULKAN ifdef GGML_HIPBLAS @@ -757,6 +833,14 @@ ifdef GGML_CUDA_FORCE_DMMV HIPFLAGS += -DGGML_CUDA_FORCE_DMMV endif # GGML_CUDA_FORCE_DMMV +ifdef GGML_CUDA_FORCE_MMQ + HIPFLAGS += -DGGML_CUDA_FORCE_MMQ +endif # GGML_CUDA_FORCE_MMQ + +ifdef GGML_CUDA_FORCE_CUBLAS + HIPFLAGS += -DGGML_CUDA_FORCE_CUBLAS +endif # GGML_CUDA_FORCE_CUBLAS + ifdef GGML_CUDA_NO_PEER_COPY HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY endif # GGML_CUDA_NO_PEER_COPY @@ -809,15 +893,16 @@ ggml/src/ggml-metal-embed.o: \ ggml/src/ggml-common.h @echo "Embedding Metal library" @sed -e '/#include "ggml-common.h"/r ggml/src/ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml/src/ggml-metal.metal > ggml/src/ggml-metal-embed.metal - $(eval TEMP_ASSEMBLY=$(shell mktemp)) - @echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY) - @echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY) - @echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY) - @echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY) - @echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY) - @echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY) - @$(AS) $(TEMP_ASSEMBLY) -o $@ - @rm -f ${TEMP_ASSEMBLY} + $(eval TEMP_ASSEMBLY=$(shell mktemp -d)) + @echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)/ggml-metal-embed.s + @echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s + @echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s + @echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s + @echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s + @echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s + $(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@ + @rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s + @rmdir ${TEMP_ASSEMBLY} endif endif # GGML_METAL @@ -825,10 +910,14 @@ OBJ_GGML += \ ggml/src/ggml.o \ ggml/src/ggml-alloc.o \ ggml/src/ggml-backend.o \ - ggml/src/ggml-quants.o + ggml/src/ggml-quants.o \ + ggml/src/ggml-aarch64.o OBJ_LLAMA = \ src/llama.o \ + src/llama-vocab.o \ + src/llama-grammar.o \ + src/llama-sampling.o \ src/unicode.o \ src/unicode-data.o @@ -896,6 +985,7 @@ $(info I CXX: $(shell $(CXX) --version | head -n 1)) ifdef GGML_CUDA $(info I NVCC: $(shell $(NVCC) --version | tail -n 1)) CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])') +ifndef GGML_MUSA ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1) ifndef CUDA_DOCKER_ARCH @@ -905,6 +995,7 @@ endif # CUDA_POWER_ARCH endif # CUDA_DOCKER_ARCH endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1) +endif # GGML_MUSA endif # GGML_CUDA $(info ) @@ -925,6 +1016,7 @@ $(info - LLAMA_NO_LLAMAFILE) $(info - LLAMA_NO_ACCELERATE) $(info - LLAMA_NO_OPENMP) $(info - LLAMA_NO_METAL) +$(info - LLAMA_NO_CCACHE) $(info ) endif @@ -958,15 +1050,22 @@ ggml/src/ggml-quants.o: \ ggml/src/ggml-common.h $(CC) $(CFLAGS) -c $< -o $@ +ggml/src/ggml-aarch64.o: \ + ggml/src/ggml-aarch64.c \ + ggml/include/ggml.h \ + ggml/src/ggml-aarch64.h \ + ggml/src/ggml-common.h + $(CC) $(CFLAGS) -c $< -o $@ + ggml/src/ggml-blas.o: \ ggml/src/ggml-blas.cpp \ ggml/include/ggml-blas.h $(CXX) $(CXXFLAGS) -c $< -o $@ ifndef GGML_NO_LLAMAFILE -ggml/src/sgemm.o: \ - ggml/src/sgemm.cpp \ - ggml/src/sgemm.h \ +ggml/src/llamafile/sgemm.o: \ + ggml/src/llamafile/sgemm.cpp \ + ggml/src/llamafile/sgemm.h \ ggml/include/ggml.h $(CXX) $(CXXFLAGS) -c $< -o $@ endif # GGML_NO_LLAMAFILE @@ -1000,6 +1099,10 @@ src/unicode-data.o: \ src/llama.o: \ src/llama.cpp \ + src/llama-impl.h \ + src/llama-vocab.h \ + src/llama-grammar.h \ + src/llama-sampling.h \ src/unicode.h \ include/llama.h \ ggml/include/ggml-cuda.h \ @@ -1009,6 +1112,29 @@ src/llama.o: \ ggml/include/ggml-backend.h $(CXX) $(CXXFLAGS) -c $< -o $@ +src/llama-vocab.o: \ + src/llama-vocab.cpp \ + src/llama-vocab.h \ + src/llama-impl.h \ + include/llama.h + $(CXX) $(CXXFLAGS) -c $< -o $@ + +src/llama-grammar.o: \ + src/llama-grammar.cpp \ + src/llama-grammar.h \ + src/llama-impl.h \ + src/llama-vocab.h \ + src/llama-sampling.h \ + include/llama.h + $(CXX) $(CXXFLAGS) -c $< -o $@ + +src/llama-sampling.o: \ + src/llama-sampling.cpp \ + src/llama-sampling.h \ + src/llama-impl.h \ + include/llama.h + $(CXX) $(CXXFLAGS) -c $< -o $@ + $(LIB_LLAMA): \ $(OBJ_LLAMA) \ $(LIB_GGML) @@ -1085,13 +1211,15 @@ clean: rm -rvf ggml/*.dll rm -rvf ggml/*.so rm -vrf ggml/src/*.o + rm -rvf ggml/src/llamafile/*.o rm -rvf common/build-info.cpp rm -vrf ggml/src/ggml-metal-embed.metal rm -vrf ggml/src/ggml-cuda/*.o rm -vrf ggml/src/ggml-cuda/template-instances/*.o rm -rvf $(BUILD_TARGETS) rm -rvf $(TEST_TARGETS) - rm -rvf $(LEGACY_TARGETS) + rm -f vulkan-shaders-gen ggml/src/ggml-vulkan-shaders.hpp ggml/src/ggml-vulkan-shaders.cpp + rm -rvf $(LEGACY_TARGETS_CLEAN) find examples pocs -type f -name "*.o" -delete # @@ -1178,6 +1306,23 @@ llama-gguf: examples/gguf/gguf.cpp \ $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) +examples/gguf-hash/deps/sha1/sha1.o: \ + examples/gguf-hash/deps/sha1/sha1.c + $(CC) $(CFLAGS) -Iexamples/gguf-hash/deps -c $< -o $@ + +examples/gguf-hash/deps/xxhash/xxhash.o: \ + examples/gguf-hash/deps/xxhash/xxhash.c + $(CC) $(CFLAGS) -Iexamples/gguf-hash/deps -c $< -o $@ + +examples/gguf-hash/deps/sha256/sha256.o: \ + examples/gguf-hash/deps/sha256/sha256.c + $(CC) $(CFLAGS) -Iexamples/gguf-hash/deps -c $< -o $@ + +llama-gguf-hash: examples/gguf-hash/gguf-hash.cpp examples/gguf-hash/deps/sha1/sha1.o examples/gguf-hash/deps/xxhash/xxhash.o examples/gguf-hash/deps/sha256/sha256.o\ + $(OBJ_ALL) + $(CXX) $(CXXFLAGS) -Iexamples/gguf-hash/deps -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + llama-gguf-split: examples/gguf-split/gguf-split.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) @@ -1193,11 +1338,6 @@ llama-cvector-generator: examples/cvector-generator/cvector-generator.cpp \ $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -llama-train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp \ - $(OBJ_ALL) - $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) - llama-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp \ $(OBJ_GGML) $(OBJ_LLAMA) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) @@ -1213,13 +1353,8 @@ llama-baby-llama: examples/baby-llama/baby-llama.cpp \ $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -llama-finetune: examples/finetune/finetune.cpp \ - $(OBJ_ALL) - $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) - llama-export-lora: examples/export-lora/export-lora.cpp \ - $(OBJ_GGML) common/log.h + $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) @@ -1323,15 +1458,20 @@ libllava.a: examples/llava/llava.cpp \ $(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual llama-llava-cli: examples/llava/llava-cli.cpp \ - examples/llava/clip.h \ - examples/llava/clip.cpp \ - examples/llava/llava.h \ examples/llava/llava.cpp \ + examples/llava/llava.h \ + examples/llava/clip.cpp \ + examples/llava/clip.h \ $(OBJ_ALL) - $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual - $(CXX) $(CXXFLAGS) -c examples/llava/llava.cpp -o $(call GET_OBJ_FILE, examples/llava/llava.cpp) - $(CXX) $(CXXFLAGS) $(filter-out %.h $< examples/llava/clip.cpp examples/llava/llava.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) $(call GET_OBJ_FILE, examples/llava/llava.cpp) -o $@ $(LDFLAGS) + $(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual + +llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \ + examples/llava/llava.cpp \ + examples/llava/llava.h \ + examples/llava/clip.cpp \ + examples/llava/clip.h \ + $(OBJ_ALL) + $(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual ifeq ($(UNAME_S),Darwin) swift: examples/batched.swift @@ -1366,7 +1506,7 @@ run-benchmark-matmult: llama-benchmark-matmult .PHONY: run-benchmark-matmult swift tests/test-llama-grammar: tests/test-llama-grammar.cpp \ - $(OBJ_GGML) $(OBJ_COMMON) src/unicode.o src/unicode-data.o + $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) @@ -1470,3 +1610,50 @@ llama-q8dot: pocs/vdot/q8dot.cpp ggml/src/ggml.o \ $(OBJ_GGML) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + +# +# Deprecated binaries that we want to keep around long enough for people to migrate to the new filenames, then these can be removed. +# +# Mark legacy binary targets as .PHONY so that they are always checked. +.PHONY: main quantize perplexity embedding server + +# Define the object file target +examples/deprecation-warning/deprecation-warning.o: examples/deprecation-warning/deprecation-warning.cpp + $(CXX) $(CXXFLAGS) -c $< -o $@ + +# NOTE: We currently will always build the deprecation-warning `main` and `server` binaries to help users migrate. +# Eventually we will want to remove these target from building all the time. +main: examples/deprecation-warning/deprecation-warning.o + $(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS) + @echo "NOTICE: The 'main' binary is deprecated. Please use 'llama-cli' instead." + +server: examples/deprecation-warning/deprecation-warning.o + $(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS) + @echo "NOTICE: The 'server' binary is deprecated. Please use 'llama-server' instead." + +quantize: examples/deprecation-warning/deprecation-warning.o +ifneq (,$(wildcard quantize)) + $(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS) + @echo "#########" + @echo "WARNING: The 'quantize' binary is deprecated. Please use 'llama-quantize' instead." + @echo " Remove the 'quantize' binary to remove this warning." + @echo "#########" +endif + +perplexity: examples/deprecation-warning/deprecation-warning.o +ifneq (,$(wildcard perplexity)) + $(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS) + @echo "#########" + @echo "WARNING: The 'perplexity' binary is deprecated. Please use 'llama-perplexity' instead." + @echo " Remove the 'perplexity' binary to remove this warning." + @echo "#########" +endif + +embedding: examples/deprecation-warning/deprecation-warning.o +ifneq (,$(wildcard embedding)) + $(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS) + @echo "#########" + @echo "WARNING: The 'embedding' binary is deprecated. Please use 'llama-embedding' instead." + @echo " Remove the 'embedding' binary to remove this warning." + @echo "#########" +endif diff --git a/Package.swift b/Package.swift index 77fed86df..1d90b47bf 100644 --- a/Package.swift +++ b/Package.swift @@ -4,12 +4,16 @@ import PackageDescription var sources = [ "src/llama.cpp", + "src/llama-vocab.cpp", + "src/llama-grammar.cpp", + "src/llama-sampling.cpp", "src/unicode.cpp", "src/unicode-data.cpp", "ggml/src/ggml.c", "ggml/src/ggml-alloc.c", "ggml/src/ggml-backend.c", "ggml/src/ggml-quants.c", + "ggml/src/ggml-aarch64.c", ] var resources: [Resource] = [] diff --git a/README.md b/README.md index 3569b2bbb..bb2b93a35 100644 --- a/README.md +++ b/README.md @@ -3,7 +3,7 @@ ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png) [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT) -[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml) +[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml) [![Conan Center](https://shields.io/conan/v/llama-cpp)](https://conan.io/center/llama-cpp) [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml) @@ -13,7 +13,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) > [!IMPORTANT] [2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809) -### Recent API changes +## Recent API changes - [2024 Jun 26] The source code and CMake build scripts have been restructured https://github.com/ggerganov/llama.cpp/pull/8006 - [2024 Apr 21] `llama_token_to_piece` can now optionally render special tokens https://github.com/ggerganov/llama.cpp/pull/6807 @@ -24,9 +24,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) - [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796 - [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849 -### Hot topics +## Hot topics -- **`convert.py` has been deprecated and moved to `examples/convert-legacy-llama.py`, please use `convert-hf-to-gguf.py`** https://github.com/ggerganov/llama.cpp/pull/7430 +- **`convert.py` has been deprecated and moved to `examples/convert_legacy_llama.py`, please use `convert_hf_to_gguf.py`** https://github.com/ggerganov/llama.cpp/pull/7430 - Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021 - BPE pre-tokenization support has been added: https://github.com/ggerganov/llama.cpp/pull/6920 - MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387 @@ -39,37 +39,6 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) ---- -
- Table of Contents -
    -
  1. - Description -
    2. -
  2. - Usage - -
    3. -
  3. Contributing
    4. -
  4. Coding guidelines
    5. -
  5. Docs
    6. -
-
- ## Description The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide @@ -87,14 +56,6 @@ Since its [inception](https://github.com/ggerganov/llama.cpp/issues/33#issuecomm improved significantly thanks to many contributions. It is the main playground for developing new features for the [ggml](https://github.com/ggerganov/ggml) library. -**Supported platforms:** - -- [X] Mac OS -- [X] Linux -- [X] Windows (via CMake) -- [X] Docker -- [X] FreeBSD - **Supported models:** Typically finetunes of the base models below are supported as well. @@ -134,9 +95,20 @@ Typically finetunes of the base models below are supported as well. - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion) - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B) - [x] [OLMo](https://allenai.org/olmo) +- [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330) - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia) +- [x] [Snowflake-Arctic MoE](https://huggingface.co/collections/Snowflake/arctic-66290090abe542894a5ac520) +- [x] [Smaug](https://huggingface.co/models?search=Smaug) +- [x] [Poro 34B](https://huggingface.co/LumiOpen/Poro-34B) +- [x] [Bitnet b1.58 models](https://huggingface.co/1bitLLM) +- [x] [Flan T5](https://huggingface.co/models?search=flan-t5) +- [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca) +- [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b) +- [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966) +- [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct) +- [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a) -(instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md)) +(instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md)) **Multimodal models:** @@ -150,12 +122,6 @@ Typically finetunes of the base models below are supported as well. - [x] [Moondream](https://huggingface.co/vikhyatk/moondream2) - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny) -**HTTP server** - -[llama.cpp web server](./examples/server) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients. - -[simplechat](./examples/server/public_simplechat) is a simple chat client, which can be used to chat with the model exposed using above web server (use --path to point to simplechat), from a local web browser. - **Bindings:** - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python) @@ -176,17 +142,20 @@ Typically finetunes of the base models below are supported as well. - Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig) - Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart) - PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggerganov/llama.cpp/pull/6326) +- Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp) **UI:** Unless otherwise noted these projects are open-source with permissive licensing: +- [MindWorkAI/AI-Studio](https://github.com/MindWorkAI/AI-Studio) (FSL-1.1-MIT) - [iohub/collama](https://github.com/iohub/coLLaMA) - [janhq/jan](https://github.com/janhq/jan) (AGPL) - [nat/openplayground](https://github.com/nat/openplayground) - [Faraday](https://faraday.dev/) (proprietary) - [LMStudio](https://lmstudio.ai/) (proprietary) - [Layla](https://play.google.com/store/apps/details?id=com.laylalite) (proprietary) +- [ramalama](https://github.com/containers/ramalama) (MIT) - [LocalAI](https://github.com/mudler/LocalAI) (MIT) - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL) - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile) @@ -219,14 +188,20 @@ Unless otherwise noted these projects are open-source with permissive licensing: - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption +- [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage **Infrastructure:** - [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp +- [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs ---- +**Games:** +- [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you. -Here is a typical run using LLaMA v2 13B on M2 Ultra: +## Demo + +
+Typical run using LLaMA v2 13B on M2 Ultra ``` $ make -j && ./llama-cli -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e @@ -306,454 +281,85 @@ llama_print_timings: eval time = 24513.59 ms / 399 runs ( 61.44 ms llama_print_timings: total time = 25431.49 ms ``` +
+ +
+Demo of running both LLaMA-7B and whisper.cpp on a single M1 Pro MacBook + And here is another demo of running both LLaMA-7B and [whisper.cpp](https://github.com/ggerganov/whisper.cpp) on a single M1 Pro MacBook: https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8b4f-add84093ffff.mp4 +
+ ## Usage Here are the end-to-end binary build and model conversion steps for most supported models. -### Get the Code +### Basic usage + +Firstly, you need to get the binary. There are different methods that you can follow: +- Method 1: Clone this repository and build locally, see [how to build](./docs/build.md) +- Method 2: If you are using MacOS or Linux, you can install llama.cpp via [brew, flox or nix](./docs/install.md) +- Method 3: Use a Docker image, see [documentation for Docker](./docs/docker.md) +- Method 4: Download pre-built binary from [releases](https://github.com/ggerganov/llama.cpp/releases) + +You can run a basic completion using this command: ```bash -git clone https://github.com/ggerganov/llama.cpp -cd llama.cpp +llama-cli -m your_model.gguf -p "I believe the meaning of life is" -n 128 + +# Output: +# I believe the meaning of life is to find your own truth and to live in accordance with it. For me, this means being true to myself and following my passions, even if they don't align with societal expectations. I think that's what I love about yoga – it's not just a physical practice, but a spiritual one too. It's about connecting with yourself, listening to your inner voice, and honoring your own unique journey. ``` -### Build +See [this page](./examples/main/README.md) for a full list of parameters. -In order to build llama.cpp you have four different options. +### Conversation mode -- Using `make`: - - On Linux or MacOS: - - ```bash - make - ``` - - - On Windows: - - 1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases). - 2. Extract `w64devkit` on your pc. - 3. Run `w64devkit.exe`. - 4. Use the `cd` command to reach the `llama.cpp` folder. - 5. From here you can run: - ```bash - make - ``` - - - Notes: - - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel. - - For faster repeated compilation, install [ccache](https://ccache.dev/). - - For debug builds, run `make LLAMA_DEBUG=1` - -- Using `CMake`: - - ```bash - cmake -B build - cmake --build build --config Release - ``` - - **Notes**: - - - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel. - - For faster repeated compilation, install [ccache](https://ccache.dev/). - - For debug builds, there are two cases: - - 1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag): - - ```bash - cmake -B build -DCMAKE_BUILD_TYPE=Debug - cmake --build build - ``` - - 2. Multi-config generators (`-G` param set to Visual Studio, XCode...): - - ```bash - cmake -B build -G "Xcode" - cmake --build build --config Debug - ``` - -- Using `gmake` (FreeBSD): - - 1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics) - 2. Add your user to **video** group - 3. Install compilation dependencies. - - ```bash - sudo pkg install gmake automake autoconf pkgconf llvm15 openblas - - gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4 - ``` - -### Homebrew - -On Mac and Linux, the homebrew package manager can be used via -``` -brew install llama.cpp -``` -The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668 - -### Nix - -On Mac and Linux, the Nix package manager can be used via -``` -nix profile install nixpkgs#llama-cpp -``` -For flake enabled installs. - -Or -``` -nix-env --file '' --install --attr llama-cpp -``` -For non-flake enabled installs. - -This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164). - -#### Flox - -On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via -``` -flox install llama-cpp -``` -Flox follows the nixpkgs build of llama.cpp. - -### Metal Build - -On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU. -To disable the Metal build at compile time use the `GGML_NO_METAL=1` flag or the `GGML_METAL=OFF` cmake option. - -When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line -argument. - -### BLAS Build - -Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS. There are currently several different BLAS implementations available for build and use: - -- #### Accelerate Framework: - - This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions. - -- #### OpenBLAS: - - This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS installed on your machine. - - - Using `make`: - - On Linux: - ```bash - make GGML_OPENBLAS=1 - ``` - - - On Windows: - - 1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases). - 2. Download the latest version of [OpenBLAS for Windows](https://github.com/xianyi/OpenBLAS/releases). - 3. Extract `w64devkit` on your pc. - 4. From the OpenBLAS zip that you just downloaded copy `libopenblas.a`, located inside the `lib` folder, inside `w64devkit\x86_64-w64-mingw32\lib`. - 5. From the same OpenBLAS zip copy the content of the `include` folder inside `w64devkit\x86_64-w64-mingw32\include`. - 6. Run `w64devkit.exe`. - 7. Use the `cd` command to reach the `llama.cpp` folder. - 8. From here you can run: - - ```bash - make GGML_OPENBLAS=1 - ``` - - - Using `CMake` on Linux: - - ```bash - cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS - cmake --build build --config Release - ``` - -- #### BLIS - - Check [BLIS.md](docs/BLIS.md) for more information. - -- #### SYCL - SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators. - - llama.cpp based on SYCL is used to **support Intel GPU** (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU). - - For detailed info, please refer to [llama.cpp for SYCL](README-sycl.md). - -- #### Intel oneMKL - Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni. Please note that this build config **does not support Intel GPU**. For Intel GPU support, please refer to [llama.cpp for SYCL](./README-sycl.md). - - - Using manual oneAPI installation: - By default, `GGML_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DGGML_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. Otherwise please install oneAPI and follow the below steps: - ```bash - source /opt/intel/oneapi/setvars.sh # You can skip this step if in oneapi-basekit docker image, only required for manual installation - cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_NATIVE=ON - cmake --build build --config Release - ``` - - - Using oneAPI docker image: - If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-basekit](https://hub.docker.com/r/intel/oneapi-basekit). Then, you can use the commands given above. - - Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information. - -- #### CUDA - - This provides GPU acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads). - - For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling. - - - Using `make`: - ```bash - make GGML_CUDA=1 - ``` - - Using `CMake`: - - ```bash - cmake -B build -DGGML_CUDA=ON - cmake --build build --config Release - ``` - - The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. The following compilation options are also available to tweak performance: - - | Option | Legal values | Default | Description | - |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| - | GGML_CUDA_FORCE_DMMV | Boolean | false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. | - | GGML_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. | - | GGML_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. | - | GGML_CUDA_FORCE_MMQ | Boolean | false | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower. | - | GGML_CUDA_FORCE_CUBLAS | Boolean | false | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models | - | GGML_CUDA_F16 | Boolean | false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. | - | GGML_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. | - | GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer | 128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. | - | GGML_CUDA_FA_ALL_QUANTS | Boolean | false | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer. | - -- #### hipBLAS - - This provides BLAS acceleration on HIP-supported AMD GPUs. - Make sure to have ROCm installed. - You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html#rocm-install-quick). - - - Using `make`: - ```bash - make GGML_HIPBLAS=1 - ``` - - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU): - ```bash - HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \ - cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \ - && cmake --build build --config Release -- -j 16 - ``` - On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`. - However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs). - - Note that if you get the following error: - ``` - clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library - ``` - Try searching for a directory under `HIP_PATH` that contains the file - `oclc_abi_version_400.bc`. Then, add the following to the start of the - command: `HIP_DEVICE_LIB_PATH=`, so something - like: - ```bash - HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \ - HIP_DEVICE_LIB_PATH= \ - cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \ - && cmake --build build -- -j 16 - ``` - - - Using `make` (example for target gfx1030, build with 16 CPU threads): - ```bash - make -j16 GGML_HIPBLAS=1 GGML_HIP_UMA=1 AMDGPU_TARGETS=gfx1030 - ``` - - - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU): - ```bash - set PATH=%HIP_PATH%\bin;%PATH% - cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release - cmake --build build - ``` - Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors) - Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`. - - - The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used. - If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3. - The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above): - - | Option | Legal values | Default | Description | - |------------------------|------------------------|---------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| - | GGML_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. | - | GGML_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. | - | GGML_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. | - -- #### Vulkan - - **With docker**: - - You don't need to install Vulkan SDK. It will be installed inside the container. - - ```sh - # Build the image - docker build -t llama-cpp-vulkan -f .devops/llama-cli-vulkan.Dockerfile . - - # Then, use it: - docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 - ``` - - **Without docker**: - - Firstly, you need to make sure you have installed [Vulkan SDK](https://vulkan.lunarg.com/doc/view/latest/linux/getting_started_ubuntu.html) - - For example, on Ubuntu 22.04 (jammy), use the command below: - - ```bash - wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - - wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list - apt update -y - apt-get install -y vulkan-sdk - # To verify the installation, use the command below: - vulkaninfo - ``` - - Alternatively your package manager might be able to provide the appropriate libraries. - For example for Ubuntu 22.04 you can install `libvulkan-dev` instead. - For Fedora 40, you can install `vulkan-devel`, `glslc` and `glslang` packages. - - Then, build llama.cpp using the cmake command below: - - ```bash - cmake -B build -DGGML_VULKAN=1 - cmake --build build --config Release - # Test the output binary (with "-ngl 33" to offload all layers to GPU) - ./bin/llama-cli -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4 - - # You should see in the output, ggml_vulkan detected your GPU. For example: - # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32 - ``` - -### Prepare and Quantize - -> [!NOTE] -> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours. - -To obtain the official LLaMA 2 weights please see the Obtaining and using the Facebook LLaMA 2 model section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face. - -Note: `convert.py` has been moved to `examples/convert-legacy-llama.py` and shouldn't be used for anything other than `Llama/Llama2/Mistral` models and their derivatives. -It does not support LLaMA 3, you can use `convert-hf-to-gguf.py` with LLaMA 3 downloaded from Hugging Face. +If you want a more ChatGPT-like experience, you can run in conversation mode by passing `-cnv` as a parameter: ```bash -# obtain the official LLaMA model weights and place them in ./models -ls ./models -llama-2-7b tokenizer_checklist.chk tokenizer.model -# [Optional] for models using BPE tokenizers -ls ./models - vocab.json -# [Optional] for PyTorch .bin models like Mistral-7B -ls ./models - +llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv -# install Python dependencies -python3 -m pip install -r requirements.txt - -# convert the model to ggml FP16 format -python3 convert-hf-to-gguf.py models/mymodel/ - -# quantize the model to 4-bits (using Q4_K_M method) -./llama-quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M - -# update the gguf filetype to current version if older version is now unsupported -./llama-quantize ./models/mymodel/ggml-model-Q4_K_M.gguf ./models/mymodel/ggml-model-Q4_K_M-v2.gguf COPY +# Output: +# > hi, who are you? +# Hi there! I'm your helpful assistant! I'm an AI-powered chatbot designed to assist and provide information to users like you. I'm here to help answer your questions, provide guidance, and offer support on a wide range of topics. I'm a friendly and knowledgeable AI, and I'm always happy to help with anything you need. What's on your mind, and how can I assist you today? +# +# > what is 1+1? +# Easy peasy! The answer to 1+1 is... 2! ``` -### Run the quantized model +By default, the chat template will be taken from the input model. If you want to use another chat template, pass `--chat-template NAME` as a parameter. See the list of [supported templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) ```bash -# start inference on a gguf model -./llama-cli -m ./models/mymodel/ggml-model-Q4_K_M.gguf -n 128 +./llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv --chat-template chatml ``` -When running the larger models, make sure you have enough disk space to store all the intermediate files. +You can also use your own template via in-prefix, in-suffix and reverse-prompt parameters: -### Running on Windows with prebuilt binaries - -You will find prebuilt Windows binaries on the release page. - -Simply download and extract the latest zip package of choice: (e.g. `llama-b1380-bin-win-avx2-x64.zip`) - -From the unzipped folder, open a terminal/cmd window here and place a pre-converted `.gguf` model file. Test out the main example like so: - -``` -.\main -m llama-2-7b.Q4_0.gguf -n 128 +```bash +./llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv --in-prefix 'User: ' --reverse-prompt 'User:' ``` -### Memory/Disk Requirements +### Web server -As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same. +[llama.cpp web server](./examples/server/README.md) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients. -| Model | Original size | Quantized size (Q4_0) | -|------:|--------------:|----------------------:| -| 7B | 13 GB | 3.9 GB | -| 13B | 24 GB | 7.8 GB | -| 30B | 60 GB | 19.5 GB | -| 65B | 120 GB | 38.5 GB | +Example usage: -### Quantization +```bash +./llama-server -m your_model.gguf --port 8080 -Several quantization methods are supported. They differ in the resulting model disk size and inference speed. - -*(outdated)* - -| Model | Measure | F16 | Q4_0 | Q4_1 | Q5_0 | Q5_1 | Q8_0 | -|------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:| -| 7B | perplexity | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 | -| 7B | file size | 13.0G | 3.5G | 3.9G | 4.3G | 4.7G | 6.7G | -| 7B | ms/tok @ 4th | 127 | 55 | 54 | 76 | 83 | 72 | -| 7B | ms/tok @ 8th | 122 | 43 | 45 | 52 | 56 | 67 | -| 7B | bits/weight | 16.0 | 4.5 | 5.0 | 5.5 | 6.0 | 8.5 | -| 13B | perplexity | 5.2543 | 5.3860 | 5.3608 | 5.2856 | 5.2706 | 5.2548 | -| 13B | file size | 25.0G | 6.8G | 7.6G | 8.3G | 9.1G | 13G | -| 13B | ms/tok @ 4th | - | 103 | 105 | 148 | 160 | 131 | -| 13B | ms/tok @ 8th | - | 73 | 82 | 98 | 105 | 128 | -| 13B | bits/weight | 16.0 | 4.5 | 5.0 | 5.5 | 6.0 | 8.5 | - -- [k-quants](https://github.com/ggerganov/llama.cpp/pull/1684) -- recent k-quants improvements and new i-quants - - [#2707](https://github.com/ggerganov/llama.cpp/pull/2707) - - [#2807](https://github.com/ggerganov/llama.cpp/pull/2807) - - [#4773 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4773) - - [#4856 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4856) - - [#4861 - importance matrix](https://github.com/ggerganov/llama.cpp/pull/4861) - - [#4872 - MoE models](https://github.com/ggerganov/llama.cpp/pull/4872) - - [#4897 - 2-bit quantization](https://github.com/ggerganov/llama.cpp/pull/4897) - - [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930) - - [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957) - - [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969) - - [#4996 - k-qunats tuning](https://github.com/ggerganov/llama.cpp/pull/4996) - - [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060) - - [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196) - - [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361) - -### Perplexity (measuring model quality) - -You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better). -For more information, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity). - -The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512. -The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 threads. - -#### How to run - -1. Download/extract: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip -2. Run `./llama-perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw` -3. Output: +# Basic web UI can be accessed via browser: http://localhost:8080 +# Chat completion endpoint: http://localhost:8080/v1/chat/completions ``` -perplexity : calculating perplexity over 655 chunks -24.43 seconds per pass - ETA 4.45 hours -[1]4.5970,[2]5.1807,[3]6.0382,... -``` -And after 4.45 hours, you will have the final perplexity. ### Interactive mode -If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter. +> [!NOTE] +> If you prefer basic usage, please consider using conversation mode instead of interactive mode + In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMA emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`. Here is an example of a few-shot interaction, invoked with the command @@ -804,18 +410,72 @@ The `grammars/` folder contains a handful of sample grammars. To write your own, For authoring more complex JSON grammars, you can also check out https://grammar.intrinsiclabs.ai/, a browser app that lets you write TypeScript interfaces which it compiles to GBNF grammars that you can save for local use. Note that the app is built and maintained by members of the community, please file any issues or FRs on [its repo](http://github.com/intrinsiclabsai/gbnfgen) and not this one. -### Obtaining and using the Facebook LLaMA 2 model +## Build -- Refer to [Facebook's LLaMA download page](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) if you want to access the model data. -- Alternatively, if you want to save time and space, you can download already converted and quantized models from [TheBloke](https://huggingface.co/TheBloke), including: - - [LLaMA 2 7B base](https://huggingface.co/TheBloke/Llama-2-7B-GGUF) - - [LLaMA 2 13B base](https://huggingface.co/TheBloke/Llama-2-13B-GGUF) - - [LLaMA 2 70B base](https://huggingface.co/TheBloke/Llama-2-70B-GGUF) - - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGUF) - - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGUF) - - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGUF) +Please refer to [Build llama.cpp locally](./docs/build.md) -### Seminal papers and background on the models +## Supported backends + +| Backend | Target devices | +| --- | --- | +| [Metal](./docs/build.md#metal-build) | Apple Silicon | +| [BLAS](./docs/build.md#blas-build) | All | +| [BLIS](./docs/backend/BLIS.md) | All | +| [SYCL](./docs/backend/SYCL.md) | Intel and Nvidia GPU | +| [MUSA](./docs/build.md#musa) | Moore Threads GPU | +| [CUDA](./docs/build.md#cuda) | Nvidia GPU | +| [hipBLAS](./docs/build.md#hipblas) | AMD GPU | +| [Vulkan](./docs/build.md#vulkan) | GPU | +| [CANN](./docs/build.md#cann) | Ascend NPU | + +## Tools + +### Prepare and Quantize + +> [!NOTE] +> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours. + +To obtain the official LLaMA 2 weights please see the Obtaining and using the Facebook LLaMA 2 model section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face. + +Note: `convert.py` has been moved to `examples/convert_legacy_llama.py` and shouldn't be used for anything other than `Llama/Llama2/Mistral` models and their derivatives. +It does not support LLaMA 3, you can use `convert_hf_to_gguf.py` with LLaMA 3 downloaded from Hugging Face. + +To learn more about quantizing model, [read this documentation](./examples/quantize/README.md) + +### Perplexity (measuring model quality) + +You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better). +For more information, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity). + +To learn more how to measure perplexity using llama.cpp, [read this documentation](./examples/perplexity/README.md) + +## Contributing + +- Contributors can open PRs +- Collaborators can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch +- Collaborators will be invited based on contributions +- Any help with managing issues and PRs is very appreciated! +- See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions +- Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information +- Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205) +- A bit of backstory for those who are interested: [Changelog podcast](https://changelog.com/podcast/532) + +## Other documentations + +- [main (cli)](./examples/main/README.md) +- [server](./examples/server/README.md) +- [jeopardy](./examples/jeopardy/README.md) +- [GBNF grammars](./grammars/README.md) + +**Development documentations** + +- [How to build](./docs/build.md) +- [Running on Docker](./docs/docker.md) +- [Build on Android](./docs/android.md) +- [Performance troubleshooting](./docs/development/token_generation_performance_tips.md) +- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks) + +**Seminal papers and background on the models** If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT: - LLaMA: @@ -826,178 +486,3 @@ If your issue is with model generation quality, then please at least scan the fo - GPT-3.5 / InstructGPT / ChatGPT: - [Aligning language models to follow instructions](https://openai.com/research/instruction-following) - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155) - -### Android - -#### Build on Android using Termux -[Termux](https://github.com/termux/termux-app#installation) is a method to execute `llama.cpp` on an Android device (no root required). -``` -apt update && apt upgrade -y -apt install git make cmake -``` - -It's recommended to move your model inside the `~/` directory for best performance: -``` -cd storage/downloads -mv model.gguf ~/ -``` - -[Get the code](https://github.com/ggerganov/llama.cpp#get-the-code) & [follow the Linux build instructions](https://github.com/ggerganov/llama.cpp#build) to build `llama.cpp`. - -#### Building the Project using Android NDK -Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake. - -Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux: -``` -$ mkdir build-android -$ cd build-android -$ export NDK= -$ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod .. -$ make -``` - -Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice). - -Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission: - -(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`) -``` -$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/ -$cd /data/data/com.termux/files/home/bin -$chmod +x ./* -``` - -Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/` - -``` -$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/ -``` - -Now, you can start chatting: -``` -$cd /data/data/com.termux/files/home/bin -$./llama-cli -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml -``` - -Here's a demo of an interactive session running on Pixel 5 phone: - -https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4 - -### Docker - -#### Prerequisites -* Docker must be installed and running on your system. -* Create a folder to store big models & intermediate files (ex. /llama/models) - -#### Images -We have three Docker images available for this project: - -1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`) -2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`) -3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`) - -Additionally, there the following images, similar to the above: - -- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`) -- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`) -- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`) -- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) -- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) -- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) - -The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now). - -#### Usage - -The easiest way to download the models, convert them to ggml and optimize them is with the --all-in-one command which includes the full docker image. - -Replace `/path/to/models` below with the actual path where you downloaded the models. - -```bash -docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B -``` - -On completion, you are ready to play! - -```bash -docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 -``` - -or with a light image: - -```bash -docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 -``` - -or with a server image: - -```bash -docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 -``` - -### Docker With CUDA - -Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container. - -#### Building Locally - -```bash -docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile . -docker build -t local/llama.cpp:light-cuda -f .devops/llama-cli-cuda.Dockerfile . -docker build -t local/llama.cpp:server-cuda -f .devops/llama-server-cuda.Dockerfile . -``` - -You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture. - -The defaults are: - -- `CUDA_VERSION` set to `11.7.1` -- `CUDA_DOCKER_ARCH` set to `all` - -The resulting images, are essentially the same as the non-CUDA images: - -1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. -2. `local/llama.cpp:light-cuda`: This image only includes the main executable file. -3. `local/llama.cpp:server-cuda`: This image only includes the server executable file. - -#### Usage - -After building locally, Usage is similar to the non-CUDA examples, but you'll need to add the `--gpus` flag. You will also want to use the `--n-gpu-layers` flag. - -```bash -docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1 -docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1 -docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1 -``` - -### Contributing - -- Contributors can open PRs -- Collaborators can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch -- Collaborators will be invited based on contributions -- Any help with managing issues and PRs is very appreciated! -- Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205) -- A bit of backstory for those who are interested: [Changelog podcast](https://changelog.com/podcast/532) - -### Coding guidelines - -- Avoid adding third-party dependencies, extra files, extra headers, etc. -- Always consider cross-compatibility with other operating systems and architectures -- Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple -- There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit -- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a` -- See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions -- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices -- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$ - -![matmul](media/matmul.png) - -### Docs - -- [main (cli)](./examples/main/README.md) -- [server](./examples/server/README.md) -- [jeopardy](./examples/jeopardy/README.md) -- [BLIS](./docs/BLIS.md) -- [Performance troubleshooting](./docs/token_generation_performance_tips.md) -- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks) -- [GBNF grammars](./grammars/README.md) diff --git a/ci/run.sh b/ci/run.sh index 067ac405b..751bb0a02 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -13,6 +13,9 @@ # # with SYCL support # GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt # +# # with VULKAN support +# GG_BUILD_VULKAN=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt +# if [ -z "$2" ]; then echo "usage: $0 " @@ -40,7 +43,7 @@ if [ ! -z ${GG_BUILD_METAL} ]; then fi if [ ! -z ${GG_BUILD_CUDA} ]; then - CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=1" + CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native" fi if [ ! -z ${GG_BUILD_SYCL} ]; then @@ -52,6 +55,10 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON" fi + +if [ ! -z ${GG_BUILD_VULKAN} ]; then + CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1" +fi ## helpers # download a file if it does not exist or if it is outdated @@ -103,8 +110,11 @@ function gg_run_ctest_debug { set -e + # Check cmake, make and ctest are installed + gg_check_build_requirements + (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log - (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log + (time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log @@ -131,8 +141,11 @@ function gg_run_ctest_release { set -e + # Check cmake, make and ctest are installed + gg_check_build_requirements + (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log - (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log + (time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log if [ -z ${GG_BUILD_LOW_PERF} ]; then (time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log @@ -260,7 +273,6 @@ function gg_sum_ctest_with_model_release { } # open_llama_7b_v2 -# requires: GG_BUILD_CUDA function gg_run_open_llama_7b_v2 { cd ${SRC} @@ -284,10 +296,10 @@ function gg_run_open_llama_7b_v2 { set -e - (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DGGML_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log - (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log + (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log + (time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log - python3 ../examples/convert-legacy-llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf + python3 ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf model_f16="${path_models}/ggml-model-f16.gguf" model_q8_0="${path_models}/ggml-model-q8_0.gguf" @@ -419,9 +431,9 @@ function gg_run_pythia_1_4b { set -e (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log - (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log + (time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log - python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf + python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf model_f16="${path_models}/ggml-model-f16.gguf" model_q8_0="${path_models}/ggml-model-q8_0.gguf" @@ -529,7 +541,6 @@ function gg_sum_pythia_1_4b { } # pythia_2_8b -# requires: GG_BUILD_CUDA function gg_run_pythia_2_8b { cd ${SRC} @@ -550,10 +561,10 @@ function gg_run_pythia_2_8b { set -e - (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DGGML_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log - (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log + (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log + (time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log - python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf + python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf model_f16="${path_models}/ggml-model-f16.gguf" model_q8_0="${path_models}/ggml-model-q8_0.gguf" @@ -686,7 +697,7 @@ function gg_run_embd_bge_small { set -e (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log - (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log + (time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf @@ -701,6 +712,20 @@ function gg_run_embd_bge_small { set +e } +function gg_check_build_requirements { + if ! command -v cmake &> /dev/null; then + gg_printf 'cmake not found, please install' + fi + + if ! command -v make &> /dev/null; then + gg_printf 'make not found, please install' + fi + + if ! command -v ctest &> /dev/null; then + gg_printf 'ctest not found, please install' + fi +} + function gg_sum_embd_bge_small { gg_printf '### %s\n\n' "${ci}" @@ -741,7 +766,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then fi if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then - if [ -z ${GG_BUILD_CUDA} ]; then + if [ -z ${GG_BUILD_CUDA} ] && [ -z ${GG_BUILD_VULKAN} ]; then test $ret -eq 0 && gg_run pythia_1_4b else test $ret -eq 0 && gg_run pythia_2_8b diff --git a/cmake/llama-config.cmake.in b/cmake/llama-config.cmake.in index 2e7da2f8e..f072b76a3 100644 --- a/cmake/llama-config.cmake.in +++ b/cmake/llama-config.cmake.in @@ -8,6 +8,13 @@ set(GGML_CUDA @GGML_CUDA@) set(GGML_METAL @GGML_METAL@) set(GGML_HIPBLAS @GGML_HIPBLAS@) set(GGML_ACCELERATE @GGML_ACCELERATE@) +set(GGML_VULKAN @GGML_VULKAN@) +set(GGML_VULKAN_CHECK_RESULTS @GGML_VULKAN_CHECK_RESULTS@) +set(GGML_VULKAN_DEBUG @GGML_VULKAN_DEBUG@) +set(GGML_VULKAN_MEMORY_DEBUG @GGML_VULKAN_MEMORY_DEBUG@) +set(GGML_VULKAN_VALIDATE @GGML_VULKAN_VALIDATE@) +set(GGML_SYCL @GGML_SYCL@) +set(GGML_OPENMP @GGML_OPENMP@) @PACKAGE_INIT@ @@ -37,18 +44,36 @@ if (GGML_METAL) find_library(METALKIT_FRAMEWORK MetalKit REQUIRED) endif() +if (GGML_VULKAN) + find_package(Vulkan REQUIRED) +endif() + if (GGML_HIPBLAS) find_package(hip REQUIRED) find_package(hipblas REQUIRED) find_package(rocblas REQUIRED) endif() +if (GGML_SYCL) + find_package(IntelSYCL REQUIRED) + find_package(MKL REQUIRED) +endif() + +if (GGML_OPENMP) + find_package(OpenMP REQUIRED) +endif() + + +find_library(ggml_LIBRARY ggml + REQUIRED + HINTS ${LLAMA_LIB_DIR}) + find_library(llama_LIBRARY llama REQUIRED HINTS ${LLAMA_LIB_DIR}) -set(_llama_link_deps "Threads::Threads" "@LLAMA_EXTRA_LIBS@") -set(_llama_transient_defines "@LLAMA_TRANSIENT_DEFINES@") +set(_llama_link_deps "${ggml_LIBRARY}" "@GGML_LINK_LIBRARIES@") +set(_llama_transient_defines "@GGML_TRANSIENT_DEFINES@") add_library(llama UNKNOWN IMPORTED) diff --git a/common/common.cpp b/common/common.cpp index e6ed77b68..ba003e44b 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1,3 +1,7 @@ +#if defined(_MSC_VER) +#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING +#endif + #include "common.h" // Change JSON_ASSERT from assert() to GGML_ASSERT: #define JSON_ASSERT GGML_ASSERT @@ -73,6 +77,41 @@ using json = nlohmann::ordered_json; +// +// Environment variable utils +// + +template +static typename std::enable_if::value, void>::type +get_env(std::string name, T & target) { + char * value = std::getenv(name.c_str()); + target = value ? std::string(value) : target; +} + +template +static typename std::enable_if::value && std::is_integral::value, void>::type +get_env(std::string name, T & target) { + char * value = std::getenv(name.c_str()); + target = value ? std::stoi(value) : target; +} + +template +static typename std::enable_if::value, void>::type +get_env(std::string name, T & target) { + char * value = std::getenv(name.c_str()); + target = value ? std::stof(value) : target; +} + +template +static typename std::enable_if::value, void>::type +get_env(std::string name, T & target) { + char * value = std::getenv(name.c_str()); + if (value) { + std::string val(value); + target = val == "1" || val == "true"; + } +} + // // CPU utils // @@ -106,8 +145,34 @@ int32_t cpu_get_num_physical_cores() { if (result == 0) { return num_physical_cores; } -#elif defined(_WIN32) - //TODO: Implement +#elif defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later + // TODO: windows + arm64 + mingw64 + unsigned int n_threads_win = std::thread::hardware_concurrency(); + unsigned int default_threads = n_threads_win > 0 ? (n_threads_win <= 4 ? n_threads_win : n_threads_win / 2) : 4; + + DWORD buffer_size = 0; + if (!GetLogicalProcessorInformationEx(RelationProcessorCore, nullptr, &buffer_size)) { + if (GetLastError() != ERROR_INSUFFICIENT_BUFFER) { + return default_threads; + } + } + + std::vector buffer(buffer_size); + if (!GetLogicalProcessorInformationEx(RelationProcessorCore, reinterpret_cast(buffer.data()), &buffer_size)) { + return default_threads; + } + + int32_t num_physical_cores = 0; + PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX info = reinterpret_cast(buffer.data()); + while (buffer_size > 0) { + if (info->Relationship == RelationProcessorCore) { + num_physical_cores += info->Processor.GroupCount; + } + buffer_size -= info->Size; + info = reinterpret_cast(reinterpret_cast(info) + info->Size); + } + + return num_physical_cores > 0 ? num_physical_cores : default_threads; #endif unsigned int n_threads = std::thread::hardware_concurrency(); return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4; @@ -186,6 +251,57 @@ int32_t cpu_get_num_math() { return cpu_get_num_physical_cores(); } +// Helper for setting process priority + +#if defined(_WIN32) + +bool set_process_priority(enum ggml_sched_priority prio) { + if (prio == GGML_SCHED_PRIO_NORMAL) { + return true; + } + + DWORD p = NORMAL_PRIORITY_CLASS; + switch (prio) { + case GGML_SCHED_PRIO_NORMAL: p = NORMAL_PRIORITY_CLASS; break; + case GGML_SCHED_PRIO_MEDIUM: p = ABOVE_NORMAL_PRIORITY_CLASS; break; + case GGML_SCHED_PRIO_HIGH: p = HIGH_PRIORITY_CLASS; break; + case GGML_SCHED_PRIO_REALTIME: p = REALTIME_PRIORITY_CLASS; break; + } + + if (!SetPriorityClass(GetCurrentProcess(), p)) { + fprintf(stderr, "warn: failed to set process priority class %d : (%d)\n", prio, (int) GetLastError()); + return false; + } + + return true; +} + +#else // MacOS and POSIX +#include +#include + +bool set_process_priority(enum ggml_sched_priority prio) { + if (prio == GGML_SCHED_PRIO_NORMAL) { + return true; + } + + int p = 0; + switch (prio) { + case GGML_SCHED_PRIO_NORMAL: p = 0; break; + case GGML_SCHED_PRIO_MEDIUM: p = -5; break; + case GGML_SCHED_PRIO_HIGH: p = -10; break; + case GGML_SCHED_PRIO_REALTIME: p = -20; break; + } + + if (!setpriority(PRIO_PROCESS, 0, p)) { + fprintf(stderr, "warn: failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno); + return false; + } + return true; +} + +#endif + // // CLI argument parsing // @@ -212,6 +328,30 @@ void gpt_params_handle_model_default(gpt_params & params) { } } +void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model) { + int32_t n_set = 0; + + if (cpuparams.n_threads < 0) { + // Assuming everything about cpuparams is invalid + if (role_model != nullptr) { + cpuparams = *role_model; + } else { + cpuparams.n_threads = cpu_get_num_math(); + } + } + + for (int32_t i = 0; i < GGML_MAX_N_THREADS; i++) { + if (cpuparams.cpumask[i]) { + n_set++; + } + } + + if (n_set && n_set < cpuparams.n_threads) { + // Not enough set bits, may experience performance issues. + fprintf(stderr, "warn: Not enough set bits in CPU mask (%d) to satisfy requested thread count: %d\n", n_set, cpuparams.n_threads); + } +} + bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { bool invalid_param = false; std::string arg; @@ -231,12 +371,21 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { } } + postprocess_cpu_params(params.cpuparams, nullptr); + postprocess_cpu_params(params.cpuparams_batch, ¶ms.cpuparams); + postprocess_cpu_params(params.draft_cpuparams, ¶ms.cpuparams); + postprocess_cpu_params(params.draft_cpuparams_batch, ¶ms.cpuparams_batch); + if (params.prompt_cache_all && (params.interactive || params.interactive_first)) { throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n"); } gpt_params_handle_model_default(params); + if (params.hf_token.empty()) { + get_env("HF_TOKEN", params.hf_token); + } + if (params.escape) { string_process_escapes(params.prompt); string_process_escapes(params.input_prefix); @@ -255,6 +404,32 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { return true; } +void gpt_params_parse_from_env(gpt_params & params) { + // we only care about server-related params for now + get_env("LLAMA_ARG_MODEL", params.model); + get_env("LLAMA_ARG_MODEL_URL", params.model_url); + get_env("LLAMA_ARG_MODEL_ALIAS", params.model_alias); + get_env("LLAMA_ARG_HF_REPO", params.hf_repo); + get_env("LLAMA_ARG_HF_FILE", params.hf_file); + get_env("LLAMA_ARG_THREADS", params.cpuparams.n_threads); + get_env("LLAMA_ARG_CTX_SIZE", params.n_ctx); + get_env("LLAMA_ARG_N_PARALLEL", params.n_parallel); + get_env("LLAMA_ARG_BATCH", params.n_batch); + get_env("LLAMA_ARG_UBATCH", params.n_ubatch); + get_env("LLAMA_ARG_N_GPU_LAYERS", params.n_gpu_layers); + get_env("LLAMA_ARG_THREADS_HTTP", params.n_threads_http); + get_env("LLAMA_ARG_CHAT_TEMPLATE", params.chat_template); + get_env("LLAMA_ARG_N_PREDICT", params.n_predict); + get_env("LLAMA_ARG_ENDPOINT_METRICS", params.endpoint_metrics); + get_env("LLAMA_ARG_ENDPOINT_SLOTS", params.endpoint_slots); + get_env("LLAMA_ARG_EMBEDDINGS", params.embedding); + get_env("LLAMA_ARG_FLASH_ATTN", params.flash_attn); + get_env("LLAMA_ARG_DEFRAG_THOLD", params.defrag_thold); + get_env("LLAMA_ARG_CONT_BATCHING", params.cont_batching); + get_env("LLAMA_ARG_HOST", params.hostname); + get_env("LLAMA_ARG_PORT", params.port); +} + bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { const auto params_org = params; // the example can modify the default params @@ -273,6 +448,79 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { return true; } +bool parse_cpu_range(const std::string & range, bool (&boolmask)[GGML_MAX_N_THREADS]) { + size_t dash_loc = range.find('-'); + if (dash_loc == std::string::npos) { + fprintf(stderr, "Format of CPU range is invalid! Expected []-[].\n"); + return false; + } + + size_t start_i; + size_t end_i; + + if (dash_loc == 0) { + start_i = 0; + } else { + start_i = std::stoull(range.substr(0, dash_loc)); + if (start_i >= GGML_MAX_N_THREADS) { + fprintf(stderr, "Start index out of bounds!\n"); + return false; + } + } + + if (dash_loc == range.length() - 1) { + end_i = GGML_MAX_N_THREADS - 1; + } else { + end_i = std::stoull(range.substr(dash_loc + 1)); + if (end_i >= GGML_MAX_N_THREADS) { + fprintf(stderr, "End index out of bounds!\n"); + return false; + } + } + + for (size_t i = start_i; i <= end_i; i++) { + boolmask[i] = true; + } + + return true; +} + +bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREADS]) { + // Discard potential 0x prefix + size_t start_i = 0; + if (mask.length() >= 2 && mask.substr(0, 2) == "0x") { + start_i = 2; + } + + size_t num_digits = mask.length() - start_i; + if (num_digits > 128) num_digits = 128; + + size_t end_i = num_digits + start_i; + + for (size_t i = start_i, n = (num_digits*4 - 1); i < end_i; i++, n-=4) { + char c = mask.at(i); + int8_t id = c; + + if ((c >= '0' && c <= '9')) { + id -= '0'; + } else if (c >= 'a' && c <= 'f') { + id -= 'a' - 10; + } else if (c >= 'A' && c <= 'F') { + id -= 'A' - 10; + } else { + fprintf(stderr, "Invalid hex character '%c' at position %d\n", c, int32_t(i)); + return false; + } + + boolmask[ n ] = boolmask[ n ] || ((id & 8) != 0); + boolmask[n - 1] = boolmask[n - 1] || ((id & 4) != 0); + boolmask[n - 2] = boolmask[n - 2] || ((id & 2) != 0); + boolmask[n - 3] = boolmask[n - 3] || ((id & 1) != 0); + } + + return true; +} + #define CHECK_ARG if (++i >= argc) { invalid_param = true; return true; } bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param) { @@ -289,36 +537,142 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa } if (arg == "-t" || arg == "--threads") { CHECK_ARG - params.n_threads = std::stoi(argv[i]); - if (params.n_threads <= 0) { - params.n_threads = std::thread::hardware_concurrency(); + params.cpuparams.n_threads = std::stoi(argv[i]); + if (params.cpuparams.n_threads <= 0) { + params.cpuparams.n_threads = std::thread::hardware_concurrency(); } return true; } + if (arg == "-C" || arg == "--cpu-mask") { + CHECK_ARG + std::string mask = argv[i]; + params.cpuparams.mask_valid = true; + invalid_param = !parse_cpu_mask(mask, params.cpuparams.cpumask); + return true; + } + if (arg == "-Cr" || arg == "--cpu-range") { + CHECK_ARG + std::string range = argv[i]; + params.cpuparams.mask_valid = true; + invalid_param = !parse_cpu_range(range, params.cpuparams.cpumask); + return true; + } + if (arg == "--prio") { + CHECK_ARG + params.cpuparams.priority = (enum ggml_sched_priority) std::stoul(argv[i]); + return true; + } + if (arg == "--cpu-strict") { + CHECK_ARG + params.cpuparams.strict_cpu = std::stoul(argv[i]); + return true; + } + if (arg == "--poll") { + CHECK_ARG + params.cpuparams.poll = std::stoul(argv[i]); + return true; + } if (arg == "-tb" || arg == "--threads-batch") { CHECK_ARG - params.n_threads_batch = std::stoi(argv[i]); - if (params.n_threads_batch <= 0) { - params.n_threads_batch = std::thread::hardware_concurrency(); + params.cpuparams_batch.n_threads = std::stoi(argv[i]); + if (params.cpuparams_batch.n_threads <= 0) { + params.cpuparams_batch.n_threads = std::thread::hardware_concurrency(); } return true; } + if (arg == "-Cb" || arg == "--cpu-mask-batch") { + CHECK_ARG + std::string mask = argv[i]; + params.cpuparams_batch.mask_valid = true; + invalid_param = !parse_cpu_mask(mask, params.cpuparams_batch.cpumask); + return true; + } + if (arg == "-Crb" || arg == "--cpu-range_batch") { + CHECK_ARG + std::string range = argv[i]; + params.cpuparams_batch.mask_valid = true; + invalid_param = !parse_cpu_range(range, params.cpuparams_batch.cpumask); + return true; + } + if (arg == "--prio-batch") { + CHECK_ARG + params.cpuparams_batch.priority = (enum ggml_sched_priority) std::stoul(argv[i]); + return true; + } + if (arg == "--cpu-strict-batch") { + params.cpuparams_batch.strict_cpu = true; + return true; + } + if (arg == "--poll-batch") { + CHECK_ARG + params.cpuparams_batch.poll = std::stoul(argv[i]); + return true; + } if (arg == "-td" || arg == "--threads-draft") { CHECK_ARG - params.n_threads_draft = std::stoi(argv[i]); - if (params.n_threads_draft <= 0) { - params.n_threads_draft = std::thread::hardware_concurrency(); + params.draft_cpuparams.n_threads = std::stoi(argv[i]); + if (params.draft_cpuparams.n_threads <= 0) { + params.draft_cpuparams.n_threads = std::thread::hardware_concurrency(); } return true; + } + if (arg == "-Cd" || arg == "--cpu-mask-draft") { + CHECK_ARG + std::string mask = argv[i]; + params.draft_cpuparams.mask_valid = true; + invalid_param = !parse_cpu_mask(mask, params.draft_cpuparams.cpumask); + return true; + } + if (arg == "-Crd" || arg == "--cpu-range-draft") { + CHECK_ARG + std::string range = argv[i]; + params.draft_cpuparams.mask_valid = true; + invalid_param = !parse_cpu_range(range, params.draft_cpuparams.cpumask); + return true; + } + if (arg == "--prio-draft") { + CHECK_ARG + params.draft_cpuparams.priority = (enum ggml_sched_priority) std::stoul(argv[i]); + return true; + } + if (arg == "--cpu-strict-draft") { + params.draft_cpuparams.strict_cpu = true; + return true; + } + if (arg == "--poll-draft") { + CHECK_ARG + params.draft_cpuparams.poll = std::stoul(argv[i]); + return true; } if (arg == "-tbd" || arg == "--threads-batch-draft") { CHECK_ARG - params.n_threads_batch_draft = std::stoi(argv[i]); - if (params.n_threads_batch_draft <= 0) { - params.n_threads_batch_draft = std::thread::hardware_concurrency(); + params.draft_cpuparams_batch.n_threads = std::stoi(argv[i]); + if (params.draft_cpuparams_batch.n_threads <= 0) { + params.draft_cpuparams_batch.n_threads = std::thread::hardware_concurrency(); } return true; } + if (arg == "-Crbd" || arg == "--cpu-range-batch-draft") { + CHECK_ARG + std::string range = argv[i]; + params.draft_cpuparams_batch.mask_valid = true; + invalid_param = !parse_cpu_range(range, params.draft_cpuparams_batch.cpumask); + return true; + } + if (arg == "--prio-batch-draft") { + CHECK_ARG + params.draft_cpuparams_batch.priority = (enum ggml_sched_priority) std::stoul(argv[i]); + return true; + } + if (arg == "--cpu-strict-batch-draft") { + params.draft_cpuparams_batch.strict_cpu = true; + return true; + } + if (arg == "--poll-batch-draft") { + CHECK_ARG + params.draft_cpuparams_batch.poll = std::stoul(argv[i]); + return true; + } if (arg == "-p" || arg == "--prompt") { CHECK_ARG params.prompt = argv[i]; @@ -472,6 +826,14 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa else { invalid_param = true; } return true; } + if (arg == "--attention") { + CHECK_ARG + std::string value(argv[i]); + /**/ if (value == "causal") { params.attention_type = LLAMA_ATTENTION_TYPE_CAUSAL; } + else if (value == "non-causal") { params.attention_type = LLAMA_ATTENTION_TYPE_NON_CAUSAL; } + else { invalid_param = true; } + return true; + } if (arg == "--defrag-thold" || arg == "-dt") { CHECK_ARG params.defrag_thold = std::stof(argv[i]); @@ -644,6 +1006,14 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa params.model_url = argv[i]; return true; } + if (arg == "-hft" || arg == "--hf-token") { + if (++i >= argc) { + invalid_param = true; + return true; + } + params.hf_token = argv[i]; + return true; + } if (arg == "-hfr" || arg == "--hf-repo") { CHECK_ARG params.hf_repo = argv[i]; @@ -656,21 +1026,24 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa } if (arg == "--lora") { CHECK_ARG - params.lora_adapter.emplace_back(argv[i], 1.0f); - params.use_mmap = false; + params.lora_adapters.push_back({ + std::string(argv[i]), + 1.0, + }); return true; } if (arg == "--lora-scaled") { CHECK_ARG - const char* lora_adapter = argv[i]; + std::string lora_adapter = argv[i]; CHECK_ARG - params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i])); - params.use_mmap = false; + params.lora_adapters.push_back({ + lora_adapter, + std::stof(argv[i]), + }); return true; } - if (arg == "--lora-base") { - CHECK_ARG - params.lora_base = argv[i]; + if (arg == "--lora-init-without-apply") { + params.lora_init_without_apply = true; return true; } if (arg == "--control-vector") { @@ -769,6 +1142,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa params.cont_batching = true; return true; } + if (arg == "-nocb" || arg == "--no-cont-batching") { + params.cont_batching = false; + return true; + } if (arg == "-fa" || arg == "--flash-attn") { params.flash_attn = true; return true; @@ -790,7 +1167,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa } return true; } - if (arg == "-ngld" || arg == "--gpu-layers-draft" || arg == "--gpu-layers-draft") { + if (arg == "-ngld" || arg == "--gpu-layers-draft" || arg == "--n-gpu-layers-draft") { CHECK_ARG params.n_gpu_layers_draft = std::stoi(argv[i]); if (!llama_supports_gpu_offload()) { @@ -1244,6 +1621,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa CHECK_ARG params.out_file = argv[i]; params.cvector_outfile = argv[i]; + params.lora_outfile = argv[i]; return true; } if (arg == "-ofreq" || arg == "--output-frequency") { @@ -1298,6 +1676,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa else { invalid_param = true; } return true; } + if (arg == "--no-warmup") { + params.warmup = false; + return true; + } #ifndef LOG_DISABLE_LOGS // Parse args for logging parameters if (log_param_single_parse(argv[i])) { @@ -1375,11 +1757,40 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "*", " --no-display-prompt", "don't print prompt at generation (default: %s)", !params.display_prompt ? "true" : "false" }); options.push_back({ "*", "-co, --color", "colorise output to distinguish prompt and user input from generations (default: %s)", params.use_color ? "true" : "false" }); options.push_back({ "*", "-s, --seed SEED", "RNG seed (default: %d, use random seed for < 0)", params.seed }); - options.push_back({ "*", "-t, --threads N", "number of threads to use during generation (default: %d)", params.n_threads }); + options.push_back({ "*", "-t, --threads N", "number of threads to use during generation (default: %d)", params.cpuparams.n_threads }); options.push_back({ "*", "-tb, --threads-batch N", "number of threads to use during batch and prompt processing (default: same as --threads)" }); options.push_back({ "speculative", "-td, --threads-draft N", "number of threads to use during generation (default: same as --threads)" }); - options.push_back({ "speculative", "-tbd, --threads-batch-draft N", - "number of threads to use during batch and prompt processing (default: same as --threads-draft)" }); + options.push_back({ "speculative", "-tbd, --threads-batch-draft N","number of threads to use during batch and prompt processing (default: same as --threads-draft)" }); + +#ifndef GGML_USE_OPENMP + // these options are available only with the internal threadpool + options.push_back({ "*", "-C, --cpu-mask M", "CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: \"\")"}); + options.push_back({ "*", "-Cr, --cpu-range lo-hi", "range of CPUs for affinity. Complements --cpu-mask"}); + options.push_back({ "*", " --cpu-strict <0|1>", "use strict CPU placement (default: %u)\n", (unsigned) params.cpuparams.strict_cpu}); + options.push_back({ "*", " --priority N", "set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.cpuparams.priority}); + options.push_back({ "*", " --poll <0...100>", "use polling level to wait for work (0 - no polling, default: %u)\n", (unsigned) params.cpuparams.poll}); + + options.push_back({ "*", "-Cb, --cpu-mask-batch M", "CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask)"}); + options.push_back({ "*", "-Crb, --cpu-range-batch lo-hi", "ranges of CPUs for affinity. Complements --cpu-mask-batch"}); + options.push_back({ "*", " --cpu-strict-batch <0|1>","use strict CPU placement (default: same as --cpu-strict)"}); + options.push_back({ "*", " --priority-batch N", "set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: --priority)"}); + options.push_back({ "*", " --poll-batch <0|1>", "use polling to wait for work (default: same as --poll"}); + + options.push_back({ "speculative", "-Cd, --cpu-mask-draft M", "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)"}); + options.push_back({ "speculative", "-Crd, --cpu-range-draft lo-hi", "Ranges of CPUs for affinity. Complements --cpu-mask-draft"}); + options.push_back({ "speculative", " --cpu-strict-draft <0|1>","Use strict CPU placement for draft model (default: same as --cpu-strict)"}); + options.push_back({ "speculative", " --priority-draft N", "Set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: same as --priority)"}); + options.push_back({ "speculative", " --poll-draft <0|1>", "Use polling to wait for draft model work (default: same as --poll])"}); + + options.push_back({ "speculative", "-Cbd, --cpu-mask-batch-draft M","Draft model CPU affinity mask. Complements cpu-range-draft-batch (default: same as --cpu-mask-draft)"}); + options.push_back({ "speculative", "-Crbd, --cpu-range-batch-draft lo-hi", + "Ranges of CPUs for affinity. Complements --cpu-mask-draft-batch)"}); + options.push_back({ "speculative", " --cpu-strict-batch-draft <0|1>", + "Use strict CPU placement for draft model (default: --cpu-strict-draft)"}); + options.push_back({ "speculative", " --priority-batch-draft N","Set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: --priority-draft)"}); + options.push_back({ "speculative", " --poll-batch-draft <0|1>","Use polling to wait for draft model work (default: --poll-draft)"}); +#endif // GGML_USE_OPENMP + options.push_back({ "speculative", " --draft N", "number of tokens to draft for speculative decoding (default: %d)", params.n_draft }); options.push_back({ "speculative", "-ps, --p-split N", "speculative decoding split probability (default: %.1f)", (double)params.p_split }); options.push_back({ "*", "-lcs, --lookup-cache-static FNAME", @@ -1394,7 +1805,9 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "*", " --keep N", "number of tokens to keep from the initial prompt (default: %d, -1 = all)", params.n_keep }); options.push_back({ "*", " --chunks N", "max number of chunks to process (default: %d, -1 = all)", params.n_chunks }); options.push_back({ "*", "-fa, --flash-attn", "enable Flash Attention (default: %s)", params.flash_attn ? "enabled" : "disabled" }); - options.push_back({ "*", "-p, --prompt PROMPT", "prompt to start generation with (default: '%s')", params.prompt.c_str() }); + options.push_back({ "*", "-p, --prompt PROMPT", "prompt to start generation with\n" + "in conversation mode, this will be used as system prompt\n" + "(default: '%s')", params.prompt.c_str() }); options.push_back({ "*", "-f, --file FNAME", "a file containing the prompt (default: none)" }); options.push_back({ "*", " --in-file FNAME", "an input file (repeat to specify multiple files)" }); options.push_back({ "*", "-bf, --binary-file FNAME", "binary file containing the prompt (default: none)" }); @@ -1409,13 +1822,16 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param "halt generation at PROMPT, return control in interactive mode\n" "can be specified more than once for multiple prompts" }); options.push_back({ "main", "-sp, --special", "special tokens output enabled (default: %s)", params.special ? "true" : "false" }); - options.push_back({ "main", "-cnv, --conversation", "run in conversation mode (does not print special tokens and suffix/prefix, use default chat template) (default: %s)", params.conversation ? "true" : "false" }); + options.push_back({ "main", "-cnv, --conversation", "run in conversation mode, does not print special tokens and suffix/prefix\n" + "if suffix/prefix are not specified, default chat template will be used\n" + "(default: %s)", params.conversation ? "true" : "false" }); options.push_back({ "main infill", "-i, --interactive", "run in interactive mode (default: %s)", params.interactive ? "true" : "false" }); options.push_back({ "main infill", "-if, --interactive-first", "run in interactive mode and wait for input right away (default: %s)", params.interactive_first ? "true" : "false" }); options.push_back({ "main infill", "-mli, --multiline-input", "allows you to write or paste multiple lines without ending each in '\\'" }); options.push_back({ "main infill", " --in-prefix-bos", "prefix BOS to user inputs, preceding the `--in-prefix` string" }); options.push_back({ "main infill", " --in-prefix STRING", "string to prefix user inputs with (default: empty)" }); options.push_back({ "main infill", " --in-suffix STRING", "string to suffix after user inputs with (default: empty)" }); + options.push_back({ "main", " --no-warmup", "skip warming up the model with an empty run" }); options.push_back({ "server infill", " --spm-infill", "use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: %s)", params.spm_infill ? "enabled" : "disabled" }); @@ -1453,6 +1869,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "main", " --cfg-scale N", "strength of guidance (default: %.1f, 1.0 = disable)", (double)sparams.cfg_scale }); options.push_back({ "main", " --chat-template JINJA_TEMPLATE", "set custom jinja chat template (default: template taken from model's metadata)\n" + "if suffix/prefix are specified, template will be disabled\n" "only commonly used templates are accepted:\n" "https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template" }); options.push_back({ "grammar" }); @@ -1463,8 +1880,10 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param "For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead" }); options.push_back({ "embedding" }); - options.push_back({ "embedding", " --pooling {none,mean,cls}", + options.push_back({ "embedding", " --pooling {none,mean,cls,last}", "pooling type for embeddings, use model default if unspecified" }); + options.push_back({ "embedding", " --attention {causal,non-causal}", + "attention type for embeddings, use model default if unspecified" }); options.push_back({ "context hacking" }); options.push_back({ "*", " --rope-scaling {none,linear,yarn}", @@ -1503,6 +1922,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "*", "-np, --parallel N", "number of parallel sequences to decode (default: %d)", params.n_parallel }); options.push_back({ "*", "-ns, --sequences N", "number of sequences to decode (default: %d)", params.n_sequences }); options.push_back({ "*", "-cb, --cont-batching", "enable continuous batching (a.k.a dynamic batching) (default: %s)", params.cont_batching ? "enabled" : "disabled" }); + options.push_back({ "*", "-nocb, --no-cont-batching", "disable continuous batching" }); options.push_back({ "multi-modality" }); options.push_back({ "*", " --mmproj FILE", "path to a multimodal projector file for LLaVA. see examples/llava/README.md" }); @@ -1545,9 +1965,8 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "*", " --override-kv KEY=TYPE:VALUE", "advanced option to override model metadata by key. may be specified multiple times.\n" "types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false" }); - options.push_back({ "*", " --lora FNAME", "apply LoRA adapter (implies --no-mmap)" }); - options.push_back({ "*", " --lora-scaled FNAME S", "apply LoRA adapter with user defined scaling S (implies --no-mmap)" }); - options.push_back({ "*", " --lora-base FNAME", "optional model to use as a base for the layers modified by the LoRA adapter" }); + options.push_back({ "*", " --lora FNAME", "apply LoRA adapter (can be repeated to use multiple adapters)" }); + options.push_back({ "*", " --lora-scaled FNAME S", "apply LoRA adapter with user defined scaling S (can be repeated to use multiple adapters)" }); options.push_back({ "*", " --control-vector FNAME", "add a control vector\n" "note: this argument can be repeated to add multiple control vectors" }); options.push_back({ "*", " --control-vector-scaled FNAME SCALE", @@ -1561,6 +1980,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "*", "-mu, --model-url MODEL_URL", "model download url (default: unused)" }); options.push_back({ "*", "-hfr, --hf-repo REPO", "Hugging Face model repository (default: unused)" }); options.push_back({ "*", "-hff, --hf-file FILE", "Hugging Face model file (default: unused)" }); + options.push_back({ "*", "-hft, --hf-token TOKEN", "Hugging Face access token (default: value from HF_TOKEN environment variable)" }); options.push_back({ "retrieval" }); options.push_back({ "retrieval", " --context-file FNAME", "file to load context from (repeat to specify multiple files)" }); @@ -1595,7 +2015,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "server", " --host HOST", "ip address to listen (default: %s)", params.hostname.c_str() }); options.push_back({ "server", " --port PORT", "port to listen (default: %d)", params.port }); options.push_back({ "server", " --path PATH", "path to serve static files from (default: %s)", params.public_path.c_str() }); - options.push_back({ "server", " --embedding(s)", "enable embedding endpoint (default: %s)", params.embedding ? "enabled" : "disabled" }); + options.push_back({ "server", " --embedding(s)", "restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled" }); options.push_back({ "server", " --api-key KEY", "API key to use for authentication (default: none)" }); options.push_back({ "server", " --api-key-file FNAME", "path to file containing API keys (default: none)" }); options.push_back({ "server", " --ssl-key-file FNAME", "path to file a PEM-encoded SSL private key" }); @@ -1615,6 +2035,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param "https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template" }); options.push_back({ "server", "-sps, --slot-prompt-similarity SIMILARITY", "how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity }); + options.push_back({ "server", " --lora-init-without-apply", "load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: %s)", params.lora_init_without_apply ? "enabled" : "disabled"}); #ifndef LOG_DISABLE_LOGS options.push_back({ "logging" }); @@ -1637,6 +2058,12 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "cvector", " --pca-iter N", "number of iterations used for PCA (default: %d)", params.n_pca_iterations }); options.push_back({ "cvector", " --method {pca,mean}", "dimensionality reduction method to be used (default: pca)" }); + options.push_back({ "export-lora" }); + options.push_back({ "export-lora", "-m, --model", "model path from which to load base model (default '%s')", params.model.c_str() }); + options.push_back({ "export-lora", " --lora FNAME", "path to LoRA adapter (can be repeated to use multiple adapters)" }); + options.push_back({ "export-lora", " --lora-scaled FNAME S", "path to LoRA adapter with user defined scaling S (can be repeated to use multiple adapters)" }); + options.push_back({ "export-lora", "-o, --output FNAME", "output file (default: '%s')", params.lora_outfile.c_str() }); + printf("usage: %s [options]\n", argv[0]); for (const auto & o : options) { @@ -1666,11 +2093,17 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param std::string gpt_params_get_system_info(const gpt_params & params) { std::ostringstream os; - os << "system_info: n_threads = " << params.n_threads; - if (params.n_threads_batch != -1) { - os << " (n_threads_batch = " << params.n_threads_batch << ")"; + os << "system_info: n_threads = " << params.cpuparams.n_threads; + if (params.cpuparams_batch.n_threads != -1) { + os << " (n_threads_batch = " << params.cpuparams_batch.n_threads << ")"; } +#if defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later + // TODO: windows + arm64 + mingw64 + DWORD logicalProcessorCount = GetActiveProcessorCount(ALL_PROCESSOR_GROUPS); + os << " / " << logicalProcessorCount << " | " << llama_print_system_info(); +#else os << " / " << std::thread::hardware_concurrency() << " | " << llama_print_system_info(); +#endif return os.str(); } @@ -1720,6 +2153,23 @@ std::string string_get_sortable_timestamp() { return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns); } +void string_replace_all(std::string & s, const std::string & search, const std::string & replace) { + if (search.empty()) { + return; + } + std::string builder; + builder.reserve(s.length()); + size_t pos = 0; + size_t last_pos = 0; + while ((pos = s.find(search, last_pos)) != std::string::npos) { + builder.append(s, last_pos, pos - last_pos); + builder.append(replace); + last_pos = pos + search.length(); + } + builder.append(s, last_pos, std::string::npos); + s = std::move(builder); +} + void string_process_escapes(std::string & input) { std::size_t input_len = input.length(); std::size_t output_idx = 0; @@ -1993,23 +2443,23 @@ std::string fs_get_cache_file(const std::string & filename) { // // Model utils // - -std::tuple llama_init_from_gpt_params(gpt_params & params) { +struct llama_init_result llama_init_from_gpt_params(gpt_params & params) { + llama_init_result iparams; auto mparams = llama_model_params_from_gpt_params(params); llama_model * model = nullptr; 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); + model = llama_load_model_from_hf(params.hf_repo.c_str(), params.hf_file.c_str(), params.model.c_str(), params.hf_token.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); + model = llama_load_model_from_url(params.model_url.c_str(), params.model.c_str(), params.hf_token.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); + return iparams; } auto cparams = llama_context_params_from_gpt_params(params); @@ -2018,7 +2468,7 @@ std::tuple llama_init_from_gpt_par if (lctx == NULL) { fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str()); llama_free_model(model); - return std::make_tuple(nullptr, nullptr); + return iparams; } if (!params.control_vectors.empty()) { @@ -2029,7 +2479,7 @@ std::tuple llama_init_from_gpt_par if (cvec.n_embd == -1) { llama_free(lctx); llama_free_model(model); - return std::make_tuple(nullptr, nullptr); + return iparams; } int err = llama_control_vector_apply(lctx, @@ -2041,26 +2491,26 @@ std::tuple llama_init_from_gpt_par if (err) { llama_free(lctx); llama_free_model(model); - return std::make_tuple(nullptr, nullptr); + return iparams; } } - for (unsigned int i = 0; i < params.lora_adapter.size(); ++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(), - lora_scale, - ((i > 0) || params.lora_base.empty()) - ? NULL - : params.lora_base.c_str(), - params.n_threads); - if (err != 0) { - fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__); + // load and optionally apply lora adapters + for (auto & la : params.lora_adapters) { + llama_lora_adapter_container loaded_la; + loaded_la.path = la.path; + loaded_la.scale = la.scale; + loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str()); + if (loaded_la.adapter == nullptr) { + fprintf(stderr, "%s: error: failed to apply lora adapter '%s'\n", __func__, la.path.c_str()); llama_free(lctx); llama_free_model(model); - return std::make_tuple(nullptr, nullptr); + return iparams; } + iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters + } + if (!params.lora_init_without_apply) { + llama_lora_adapters_apply(lctx, iparams.lora_adapters); } if (params.ignore_eos) { @@ -2088,13 +2538,26 @@ std::tuple llama_init_from_gpt_par tmp.clear(); tmp.push_back(decoder_start_token_id); } - llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0)); + if (llama_model_has_decoder(model)) { + llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0)); + } llama_past_clear(lctx); llama_synchronize(lctx); llama_reset_timings(lctx); } - return std::make_tuple(model, lctx); + iparams.model = model; + iparams.context = lctx; + return iparams; +} + +void llama_lora_adapters_apply(struct llama_context * ctx, std::vector & lora_adapters) { + llama_lora_adapter_clear(ctx); + for (auto & la : lora_adapters) { + if (la.scale != 0.0f) { + llama_lora_adapter_set(ctx, la.adapter, la.scale); + } + } } struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) { @@ -2156,8 +2619,9 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param cparams.n_seq_max = params.n_parallel; cparams.n_batch = params.n_batch; cparams.n_ubatch = params.n_ubatch; - cparams.n_threads = params.n_threads; - cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch; + cparams.n_threads = params.cpuparams.n_threads; + cparams.n_threads_batch = params.cpuparams_batch.n_threads == -1 ? + params.cpuparams.n_threads : params.cpuparams_batch.n_threads; cparams.seed = params.seed; cparams.logits_all = params.logits_all; cparams.embeddings = params.embedding; @@ -2170,6 +2634,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param cparams.yarn_beta_slow = params.yarn_beta_slow; cparams.yarn_orig_ctx = params.yarn_orig_ctx; cparams.pooling_type = params.pooling_type; + cparams.attention_type = params.attention_type; cparams.defrag_thold = params.defrag_thold; cparams.cb_eval = params.cb_eval; cparams.cb_eval_user_data = params.cb_eval_user_data; @@ -2182,6 +2647,22 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param return cparams; } +struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params) { + struct ggml_threadpool_params tpp; + + ggml_threadpool_params_init(&tpp, params.n_threads); // setup the defaults + + if (params.mask_valid) { + std::memcpy(&tpp.cpumask, ¶ms.cpumask, GGML_MAX_N_THREADS); + } + + tpp.prio = params.priority; + tpp.poll = params.poll; + tpp.strict_cpu = params.strict_cpu; + + return tpp; +} + #ifdef LLAMA_USE_CURL static bool starts_with(const std::string & str, const std::string & prefix) { @@ -2189,7 +2670,7 @@ static bool starts_with(const std::string & str, const std::string & prefix) { return str.rfind(prefix, 0) == 0; } -static bool llama_download_file(const std::string & url, const std::string & path) { +static bool llama_download_file(const std::string & url, const std::string & path, const std::string & hf_token) { // Initialize libcurl std::unique_ptr curl(curl_easy_init(), &curl_easy_cleanup); @@ -2204,6 +2685,15 @@ static bool llama_download_file(const std::string & url, const std::string & pat curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str()); curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L); + // Check if hf-token or bearer-token was specified + if (!hf_token.empty()) { + std::string auth_header = "Authorization: Bearer "; + auth_header += hf_token.c_str(); + struct curl_slist *http_headers = NULL; + http_headers = curl_slist_append(http_headers, auth_header.c_str()); + curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers); + } + #if defined(_WIN32) // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of // operating system. Currently implemented under MS-Windows. @@ -2399,6 +2889,7 @@ static bool llama_download_file(const std::string & url, const std::string & pat struct llama_model * llama_load_model_from_url( const char * model_url, const char * path_model, + const char * hf_token, const struct llama_model_params & params) { // Basic validation of the model_url if (!model_url || strlen(model_url) == 0) { @@ -2406,7 +2897,7 @@ struct llama_model * llama_load_model_from_url( return NULL; } - if (!llama_download_file(model_url, path_model)) { + if (!llama_download_file(model_url, path_model, hf_token)) { return NULL; } @@ -2454,14 +2945,14 @@ struct llama_model * llama_load_model_from_url( // 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 { + futures_download.push_back(std::async(std::launch::async, [&split_prefix, &split_url_prefix, &n_split, hf_token](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); - return llama_download_file(split_url, split_path); + return llama_download_file(split_url, split_path, hf_token); }, idx)); } @@ -2480,6 +2971,7 @@ struct llama_model * llama_load_model_from_hf( const char * repo, const char * model, const char * path_model, + const char * hf_token, const struct llama_model_params & params) { // construct hugging face model url: // @@ -2495,7 +2987,7 @@ struct llama_model * llama_load_model_from_hf( model_url += "/resolve/main/"; model_url += model; - return llama_load_model_from_url(model_url.c_str(), path_model, params); + return llama_load_model_from_url(model_url.c_str(), path_model, hf_token, params); } #else @@ -2503,6 +2995,7 @@ struct llama_model * llama_load_model_from_hf( struct llama_model * llama_load_model_from_url( const char * /*model_url*/, const char * /*path_model*/, + const char * /*hf_token*/, const struct llama_model_params & /*params*/) { fprintf(stderr, "%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__); return nullptr; @@ -2512,6 +3005,7 @@ struct llama_model * llama_load_model_from_hf( const char * /*repo*/, const char * /*model*/, const char * /*path_model*/, + const char * /*hf_token*/, const struct llama_model_params & /*params*/) { fprintf(stderr, "%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__); return nullptr; @@ -2576,57 +3070,35 @@ std::vector llama_tokenize( } std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) { - std::vector result(8, 0); - const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special); - if (n_tokens < 0) { - result.resize(-n_tokens); - int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special); - GGML_ASSERT(check == -n_tokens); - } else { - result.resize(n_tokens); + std::string piece; + piece.resize(piece.capacity()); // using string internal cache, 15 bytes + '\n' + const int n_chars = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special); + if (n_chars < 0) { + piece.resize(-n_chars); + int check = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special); + GGML_ASSERT(check == -n_chars); + } + else { + piece.resize(n_chars); } - return std::string(result.data(), result.size()); + return piece; } -std::string llama_detokenize_spm(llama_context * ctx, const std::vector & tokens) { - const llama_token bos_id = llama_token_bos(llama_get_model(ctx)); - - std::string piece; - std::string result; - - for (size_t i = 0; i < tokens.size(); ++i) { - piece = llama_token_to_piece(ctx, tokens[i]); - - // remove the leading space of the first non-BOS token - if (((tokens[0] == bos_id && i == 1) || (tokens[0] != bos_id && i == 0)) && piece[0] == ' ') { - piece = piece.substr(1); - } - - result += piece; +std::string llama_detokenize(llama_context * ctx, const std::vector & tokens, bool special) { + std::string text; + text.resize(std::max(text.capacity(), tokens.size())); + int32_t n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special); + if (n_chars < 0) { + text.resize(-n_chars); + n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special); + GGML_ASSERT(n_chars <= (int32_t)text.size()); // whitespace trimming is performed after per-token detokenization } - return result; -} - -std::string llama_detokenize_bpe(llama_context * ctx, const std::vector & tokens) { - std::string piece; - std::string result; - - for (size_t i = 0; i < tokens.size(); ++i) { - piece = llama_token_to_piece(ctx, tokens[i]); - - result += piece; - } + text.resize(n_chars); // NOTE: the original tokenizer decodes bytes after collecting the pieces. - return result; -} - -bool llama_should_add_bos_token(const llama_model * model) { - const int add_bos = llama_add_bos_token(model); - - return add_bos != -1 ? bool(add_bos) : (llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM); + return text; } // @@ -2689,7 +3161,7 @@ std::string llama_chat_format_single(const struct llama_model * model, const llama_chat_msg & new_msg, bool add_ass) { std::ostringstream ss; - auto fmt_past_msg = llama_chat_apply_template(model, tmpl, past_msg, false); + auto fmt_past_msg = past_msg.empty() ? "" : llama_chat_apply_template(model, tmpl, past_msg, false); std::vector chat_new(past_msg); // if the past_msg ends with a newline, we must preserve it in the formatted version if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') { @@ -3121,20 +3593,18 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l } fprintf(stream, "lora:\n"); - for (std::tuple la : params.lora_adapter) { - if (std::get<1>(la) != 1.0f) { - continue; + for (auto & la : params.lora_adapters) { + if (la.scale == 1.0f) { + fprintf(stream, " - %s\n", la.path.c_str()); } - fprintf(stream, " - %s\n", std::get<0>(la).c_str()); } fprintf(stream, "lora_scaled:\n"); - for (std::tuple la : params.lora_adapter) { - if (std::get<1>(la) == 1.0f) { - continue; + for (auto & la : params.lora_adapters) { + if (la.scale != 1.0f) { + fprintf(stream, " - %s: %f\n", la.path.c_str(), la.scale); } - fprintf(stream, " - %s: %f\n", std::get<0>(la).c_str(), std::get<1>(la)); } - fprintf(stream, "lora_base: %s\n", params.lora_base.c_str()); + fprintf(stream, "lora_init_without_apply: %s # default: false\n", params.lora_init_without_apply ? "true" : "false"); fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu); fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep); fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat); @@ -3182,7 +3652,7 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l yaml_dump_vector_float(stream, "tensor_split", tensor_split_vector); fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z); - fprintf(stream, "threads: %d # default: %u\n", params.n_threads, std::thread::hardware_concurrency()); + fprintf(stream, "threads: %d # default: %u\n", params.cpuparams.n_threads, std::thread::hardware_concurrency()); fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k); fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p); fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p); diff --git a/common/common.h b/common/common.h index 65c0ef81a..cb5e7f6df 100644 --- a/common/common.h +++ b/common/common.h @@ -33,6 +33,15 @@ #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf" +struct llama_lora_adapter_info { + std::string path; + float scale; +}; + +struct llama_lora_adapter_container : llama_lora_adapter_info { + struct llama_lora_adapter * adapter; +}; + // build info extern int LLAMA_BUILD_NUMBER; extern char const * LLAMA_COMMIT; @@ -58,13 +67,18 @@ enum dimre_method { DIMRE_METHOD_MEAN, }; +struct cpu_params { + int n_threads = -1; + bool cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask. + bool mask_valid = false; // Default: any CPU + enum ggml_sched_priority priority = GGML_SCHED_PRIO_NORMAL; // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime) + bool strict_cpu = false; // Use strict CPU placement + uint32_t poll = 50; // Polling (busywait) level (0 - no polling, 100 - mostly polling) +}; + struct gpt_params { uint32_t seed = LLAMA_DEFAULT_SEED; // RNG seed - int32_t n_threads = cpu_get_num_math(); - int32_t n_threads_draft = -1; - int32_t n_threads_batch = -1; // number of threads to use for batch processing (-1 = use n_threads) - int32_t n_threads_batch_draft = -1; int32_t n_predict = -1; // new tokens to predict int32_t n_ctx = 0; // context size int32_t n_batch = 2048; // logical batch size for prompt processing (must be >=32 to use BLAS) @@ -91,6 +105,11 @@ struct gpt_params { int32_t yarn_orig_ctx = 0; // YaRN original context length float defrag_thold = -1.0f; // KV cache defragmentation threshold + struct cpu_params cpuparams; + struct cpu_params cpuparams_batch; + struct cpu_params draft_cpuparams; + struct cpu_params draft_cpuparams_batch; + ggml_backend_sched_eval_callback cb_eval = nullptr; void * cb_eval_user_data = nullptr; @@ -99,6 +118,7 @@ struct gpt_params { enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED; enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings + enum llama_attention_type attention_type = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings // // sampling parameters struct llama_sampling_params sparams; @@ -107,6 +127,7 @@ struct gpt_params { 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_token = ""; // HF token std::string hf_repo = ""; // HF repo std::string hf_file = ""; // HF file std::string prompt = ""; @@ -124,9 +145,8 @@ struct gpt_params { std::vector antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts) std::vector kv_overrides; - // TODO: avoid tuple, use struct - std::vector> lora_adapter; // lora adapter path with user defined scale - std::string lora_base = ""; // base model path for the lora adapter + bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply) + std::vector lora_adapters; // lora adapter path with user defined scale std::vector control_vectors; // control vector with user defined scale @@ -194,7 +214,7 @@ struct gpt_params { int32_t port = 8080; // server listens on this network port int32_t timeout_read = 600; // http read timeout in seconds int32_t timeout_write = timeout_read; // http write timeout in seconds - int32_t n_threads_http = -1; // number of threads to process HTTP requests + int n_threads_http = -1; // number of threads to process HTTP requests (TODO: support threadpool) std::string hostname = "127.0.0.1"; std::string public_path = ""; @@ -253,8 +273,11 @@ struct gpt_params { std::string cvector_negative_file = "examples/cvector-generator/negative.txt"; bool spm_infill = false; // suffix/prefix/middle pattern for infill + + std::string lora_outfile = "ggml-lora-merged-f16.gguf"; }; +void gpt_params_parse_from_env(gpt_params & params); void gpt_params_handle_model_default(gpt_params & params); bool gpt_params_parse_ex (int argc, char ** argv, gpt_params & params); @@ -264,6 +287,11 @@ void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params); std::string gpt_params_get_system_info(const gpt_params & params); +bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]); +bool parse_cpu_mask(const std::string& mask, bool(&boolmask)[GGML_MAX_N_THREADS]); +void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model = nullptr); +bool set_process_priority(enum ggml_sched_priority prio); + // // String utils // @@ -273,6 +301,8 @@ std::vector string_split(std::string input, char separator); std::string string_strip(const std::string & str); std::string string_get_sortable_timestamp(); +void string_replace_all(std::string & s, const std::string & search, const std::string & replace); + template static std::vector string_split(const std::string & str, char delim) { std::vector values; @@ -304,14 +334,23 @@ std::string fs_get_cache_file(const std::string & filename); // Model utils // -// TODO: avoid tuplue, use struct -std::tuple llama_init_from_gpt_params(gpt_params & params); +struct llama_init_result { + struct llama_model * model = nullptr; + struct llama_context * context = nullptr; + std::vector lora_adapters; +}; -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_init_result llama_init_from_gpt_params(gpt_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); +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 ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params); + +struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, 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 char * hf_token, const struct llama_model_params & params); + +// clear LoRA adapters from context, then apply new list of adapters +void llama_lora_adapters_apply(struct llama_context * ctx, std::vector & lora_adapters); // Batch utils @@ -349,25 +388,13 @@ std::string llama_token_to_piece( llama_token token, bool special = true); -// TODO: these should be moved in llama.h C-style API under single `llama_detokenize` function -// that takes into account the tokenizer type and decides how to handle the leading space -// // detokenizes a vector of tokens into a string // should work similar to Python's `tokenizer.decode` -// removes the leading space from the first non-BOS token -std::string llama_detokenize_spm( +// optionally renders special/control tokens +std::string llama_detokenize( llama_context * ctx, - const std::vector & tokens); - -// detokenizes a vector of tokens into a string -// should work similar to Python's `tokenizer.decode` -std::string llama_detokenize_bpe( - llama_context * ctx, - const std::vector & tokens); - -// Uses the value from the model metadata if possible, otherwise -// defaults to true when model type is SPM, otherwise false. -bool llama_should_add_bos_token(const llama_model * model); + const std::vector & tokens, + bool special = true); // // Chat template utils diff --git a/common/grammar-parser.cpp b/common/grammar-parser.cpp index a518b766d..438452eab 100644 --- a/common/grammar-parser.cpp +++ b/common/grammar-parser.cpp @@ -369,6 +369,9 @@ namespace grammar_parser { } // Validate the state to ensure that all rules are defined for (const auto & rule : state.rules) { + if (rule.empty()) { + throw std::runtime_error("Undefined rule"); + } for (const auto & elem : rule) { if (elem.type == LLAMA_GRETYPE_RULE_REF) { // Ensure that the rule at that location exists diff --git a/common/log.h b/common/log.h index 09fa63c26..1bc5328ce 100644 --- a/common/log.h +++ b/common/log.h @@ -630,7 +630,7 @@ inline std::string LOG_TOKENS_TOSTR_PRETTY(const C & ctx, const T & tokens) buf << "[ "; bool first = true; - for (const auto &token : tokens) + for (const auto & token : tokens) { if (!first) { buf << ", "; diff --git a/common/ngram-cache.h b/common/ngram-cache.h index e4fa4cbd1..ab4c9b376 100644 --- a/common/ngram-cache.h +++ b/common/ngram-cache.h @@ -37,11 +37,18 @@ struct llama_ngram { } }; +struct llama_token_hash_function { + size_t operator()(const llama_token token) const { + // see https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/ + return token * 11400714819323198485llu; + } +}; + 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]); + size_t hash = llama_token_hash_function{}(ngram.tokens[0]); + for (int i = 1; i < LLAMA_NGRAM_MAX; ++i) { + hash ^= llama_token_hash_function{}(ngram.tokens[i]); } return hash; } diff --git a/common/sampling.cpp b/common/sampling.cpp index 9f332fe57..079e40516 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -282,8 +282,6 @@ static llama_token llama_sampling_sample_impl( GGML_ASSERT(!original_logits.empty()); } llama_token id = 0; - // Get a pointer to the logits - float * logits = llama_get_logits_ith(ctx_main, idx); if (temp < 0.0) { // greedy sampling, with probs @@ -324,12 +322,15 @@ static llama_token llama_sampling_sample_impl( } if (ctx_sampling->grammar != NULL && !is_resampling) { + // Get a pointer to the logits + float * logits = llama_get_logits_ith(ctx_main, idx); + // Create an array with a single token data element for the sampled id llama_token_data single_token_data = {id, logits[id], 0.0f}; llama_token_data_array single_token_data_array = { &single_token_data, 1, false }; // Apply grammar constraints to the single token - llama_sample_grammar(ctx_main, &single_token_data_array, ctx_sampling->grammar); + llama_grammar_sample(ctx_sampling->grammar, ctx_main, &single_token_data_array); // Check if the token is valid according to the grammar by seeing if its logit has been set to -INFINITY bool is_valid = single_token_data_array.data[0].logit != -INFINITY; @@ -377,7 +378,7 @@ static llama_token_data_array llama_sampling_prepare_impl( if (ctx_sampling->grammar != NULL && !apply_grammar) { GGML_ASSERT(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))}; + *original_logits = {logits, logits + n_vocab}; } // apply params.logit_bias map @@ -390,10 +391,10 @@ static llama_token_data_array llama_sampling_prepare_impl( llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale); } - cur.clear(); + cur.resize(n_vocab); 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}); + cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f}; } llama_token_data_array cur_p = { cur.data(), cur.size(), false }; @@ -420,7 +421,7 @@ static llama_token_data_array llama_sampling_prepare_impl( // apply grammar checks before sampling logic if (apply_grammar && ctx_sampling->grammar != NULL) { - llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar); + llama_grammar_sample(ctx_sampling->grammar, ctx_main, &cur_p); } return cur_p; @@ -454,6 +455,6 @@ void llama_sampling_accept( ctx_sampling->prev.push_back(id); if (ctx_sampling->grammar != NULL && apply_grammar) { - llama_grammar_accept_token(ctx_main, ctx_sampling->grammar, id); + llama_grammar_accept_token(ctx_sampling->grammar, ctx_main, id); } } diff --git a/common/stb_image.h b/common/stb_image.h index 4766d7e67..9eedabedc 100644 --- a/common/stb_image.h +++ b/common/stb_image.h @@ -1,4 +1,4 @@ -/* stb_image - v2.28 - public domain image loader - http://nothings.org/stb +/* stb_image - v2.30 - public domain image loader - http://nothings.org/stb no warranty implied; use at your own risk Do this: @@ -48,6 +48,8 @@ LICENSE RECENT REVISION HISTORY: + 2.30 (2024-05-31) avoid erroneous gcc warning + 2.29 (2023-05-xx) optimizations 2.28 (2023-01-29) many error fixes, security errors, just tons of stuff 2.27 (2021-07-11) document stbi_info better, 16-bit PNM support, bug fixes 2.26 (2020-07-13) many minor fixes @@ -371,13 +373,14 @@ RECENT REVISION HISTORY: #define STBI_VERSION 1 -enum { - STBI_default = 0, // only used for desired_channels +enum +{ + STBI_default = 0, // only used for desired_channels - STBI_grey = 1, - STBI_grey_alpha = 2, - STBI_rgb = 3, - STBI_rgb_alpha = 4 + STBI_grey = 1, + STBI_grey_alpha = 2, + STBI_rgb = 3, + STBI_rgb_alpha = 4 }; #include @@ -405,11 +408,11 @@ extern "C" { // load image by filename, open file, or memory buffer // -typedef struct { - int (*read)(void * user, char * data, - int size); // fill 'data' with 'size' bytes. return number of bytes actually read - void (*skip)(void * user, int n); // skip the next 'n' bytes, or 'unget' the last -n bytes if negative - int (*eof)(void * user); // returns nonzero if we are at end of file/data +typedef struct +{ + int (*read) (void *user,char *data,int size); // fill 'data' with 'size' bytes. return number of bytes actually read + void (*skip) (void *user,int n); // skip the next 'n' bytes, or 'unget' the last -n bytes if negative + int (*eof) (void *user); // returns nonzero if we are at end of file/data } stbi_io_callbacks; //////////////////////////////////// @@ -417,24 +420,21 @@ typedef struct { // 8-bits-per-channel interface // -STBIDEF stbi_uc * stbi_load_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * channels_in_file, - int desired_channels); -STBIDEF stbi_uc * stbi_load_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, - int * channels_in_file, int desired_channels); +STBIDEF stbi_uc *stbi_load_from_memory (stbi_uc const *buffer, int len , int *x, int *y, int *channels_in_file, int desired_channels); +STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk , void *user, int *x, int *y, int *channels_in_file, int desired_channels); #ifndef STBI_NO_STDIO -STBIDEF stbi_uc * stbi_load(char const * filename, int * x, int * y, int * channels_in_file, int desired_channels); -STBIDEF stbi_uc * stbi_load_from_file(FILE * f, int * x, int * y, int * channels_in_file, int desired_channels); +STBIDEF stbi_uc *stbi_load (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels); +STBIDEF stbi_uc *stbi_load_from_file (FILE *f, int *x, int *y, int *channels_in_file, int desired_channels); // for stbi_load_from_file, file pointer is left pointing immediately after image #endif #ifndef STBI_NO_GIF -STBIDEF stbi_uc * stbi_load_gif_from_memory(stbi_uc const * buffer, int len, int ** delays, int * x, int * y, int * z, - int * comp, int req_comp); +STBIDEF stbi_uc *stbi_load_gif_from_memory(stbi_uc const *buffer, int len, int **delays, int *x, int *y, int *z, int *comp, int req_comp); #endif #ifdef STBI_WINDOWS_UTF8 -STBIDEF int stbi_convert_wchar_to_utf8(char * buffer, size_t bufferlen, const wchar_t * input); +STBIDEF int stbi_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input); #endif //////////////////////////////////// @@ -442,14 +442,12 @@ STBIDEF int stbi_convert_wchar_to_utf8(char * buffer, size_t bufferlen, const wc // 16-bits-per-channel interface // -STBIDEF stbi_us * stbi_load_16_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * channels_in_file, - int desired_channels); -STBIDEF stbi_us * stbi_load_16_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, - int * channels_in_file, int desired_channels); +STBIDEF stbi_us *stbi_load_16_from_memory (stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels); +STBIDEF stbi_us *stbi_load_16_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels); #ifndef STBI_NO_STDIO -STBIDEF stbi_us * stbi_load_16(char const * filename, int * x, int * y, int * channels_in_file, int desired_channels); -STBIDEF stbi_us * stbi_load_from_file_16(FILE * f, int * x, int * y, int * channels_in_file, int desired_channels); +STBIDEF stbi_us *stbi_load_16 (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels); +STBIDEF stbi_us *stbi_load_from_file_16(FILE *f, int *x, int *y, int *channels_in_file, int desired_channels); #endif //////////////////////////////////// @@ -457,55 +455,56 @@ STBIDEF stbi_us * stbi_load_from_file_16(FILE * f, int * x, int * y, int * chann // float-per-channel interface // #ifndef STBI_NO_LINEAR -STBIDEF float * stbi_loadf_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * channels_in_file, - int desired_channels); -STBIDEF float * stbi_loadf_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, int * channels_in_file, - int desired_channels); + STBIDEF float *stbi_loadf_from_memory (stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels); + STBIDEF float *stbi_loadf_from_callbacks (stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels); -#ifndef STBI_NO_STDIO -STBIDEF float * stbi_loadf(char const * filename, int * x, int * y, int * channels_in_file, int desired_channels); -STBIDEF float * stbi_loadf_from_file(FILE * f, int * x, int * y, int * channels_in_file, int desired_channels); -#endif + #ifndef STBI_NO_STDIO + STBIDEF float *stbi_loadf (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels); + STBIDEF float *stbi_loadf_from_file (FILE *f, int *x, int *y, int *channels_in_file, int desired_channels); + #endif #endif #ifndef STBI_NO_HDR -STBIDEF void stbi_hdr_to_ldr_gamma(float gamma); -STBIDEF void stbi_hdr_to_ldr_scale(float scale); + STBIDEF void stbi_hdr_to_ldr_gamma(float gamma); + STBIDEF void stbi_hdr_to_ldr_scale(float scale); #endif // STBI_NO_HDR #ifndef STBI_NO_LINEAR -STBIDEF void stbi_ldr_to_hdr_gamma(float gamma); -STBIDEF void stbi_ldr_to_hdr_scale(float scale); + STBIDEF void stbi_ldr_to_hdr_gamma(float gamma); + STBIDEF void stbi_ldr_to_hdr_scale(float scale); #endif // STBI_NO_LINEAR // stbi_is_hdr is always defined, but always returns false if STBI_NO_HDR -STBIDEF int stbi_is_hdr_from_callbacks(stbi_io_callbacks const * clbk, void * user); -STBIDEF int stbi_is_hdr_from_memory(stbi_uc const * buffer, int len); +STBIDEF int stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user); +STBIDEF int stbi_is_hdr_from_memory(stbi_uc const *buffer, int len); #ifndef STBI_NO_STDIO -STBIDEF int stbi_is_hdr(char const * filename); -STBIDEF int stbi_is_hdr_from_file(FILE * f); +STBIDEF int stbi_is_hdr (char const *filename); +STBIDEF int stbi_is_hdr_from_file(FILE *f); #endif // STBI_NO_STDIO + // get a VERY brief reason for failure // on most compilers (and ALL modern mainstream compilers) this is threadsafe -STBIDEF const char * stbi_failure_reason(void); +STBIDEF const char *stbi_failure_reason (void); // free the loaded image -- this is just free() -STBIDEF void stbi_image_free(void * retval_from_stbi_load); +STBIDEF void stbi_image_free (void *retval_from_stbi_load); // get image dimensions & components without fully decoding -STBIDEF int stbi_info_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * comp); -STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, int * comp); -STBIDEF int stbi_is_16_bit_from_memory(stbi_uc const * buffer, int len); -STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const * clbk, void * user); +STBIDEF int stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp); +STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp); +STBIDEF int stbi_is_16_bit_from_memory(stbi_uc const *buffer, int len); +STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const *clbk, void *user); #ifndef STBI_NO_STDIO -STBIDEF int stbi_info(char const * filename, int * x, int * y, int * comp); -STBIDEF int stbi_info_from_file(FILE * f, int * x, int * y, int * comp); -STBIDEF int stbi_is_16_bit(char const * filename); -STBIDEF int stbi_is_16_bit_from_file(FILE * f); +STBIDEF int stbi_info (char const *filename, int *x, int *y, int *comp); +STBIDEF int stbi_info_from_file (FILE *f, int *x, int *y, int *comp); +STBIDEF int stbi_is_16_bit (char const *filename); +STBIDEF int stbi_is_16_bit_from_file(FILE *f); #endif + + // for image formats that explicitly notate that they have premultiplied alpha, // we just return the colors as stored in the file. set this flag to force // unpremultiplication. results are undefined if the unpremultiply overflow. @@ -527,14 +526,14 @@ STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_fli // ZLIB client - used by PNG, available for other purposes -STBIDEF char * stbi_zlib_decode_malloc_guesssize(const char * buffer, int len, int initial_size, int * outlen); -STBIDEF char * stbi_zlib_decode_malloc_guesssize_headerflag(const char * buffer, int len, int initial_size, int * outlen, - int parse_header); -STBIDEF char * stbi_zlib_decode_malloc(const char * buffer, int len, int * outlen); -STBIDEF int stbi_zlib_decode_buffer(char * obuffer, int olen, const char * ibuffer, int ilen); +STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen); +STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header); +STBIDEF char *stbi_zlib_decode_malloc(const char *buffer, int len, int *outlen); +STBIDEF int stbi_zlib_decode_buffer(char *obuffer, int olen, const char *ibuffer, int ilen); + +STBIDEF char *stbi_zlib_decode_noheader_malloc(const char *buffer, int len, int *outlen); +STBIDEF int stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen); -STBIDEF char * stbi_zlib_decode_noheader_malloc(const char * buffer, int len, int * outlen); -STBIDEF int stbi_zlib_decode_noheader_buffer(char * obuffer, int olen, const char * ibuffer, int ilen); #ifdef __cplusplus } @@ -547,50 +546,52 @@ STBIDEF int stbi_zlib_decode_noheader_buffer(char * obuffer, int olen, const cha #ifdef STB_IMAGE_IMPLEMENTATION -#if defined(STBI_ONLY_JPEG) || defined(STBI_ONLY_PNG) || defined(STBI_ONLY_BMP) || defined(STBI_ONLY_TGA) || \ - defined(STBI_ONLY_GIF) || defined(STBI_ONLY_PSD) || defined(STBI_ONLY_HDR) || defined(STBI_ONLY_PIC) || \ - defined(STBI_ONLY_PNM) || defined(STBI_ONLY_ZLIB) -#ifndef STBI_ONLY_JPEG -#define STBI_NO_JPEG -#endif -#ifndef STBI_ONLY_PNG -#define STBI_NO_PNG -#endif -#ifndef STBI_ONLY_BMP -#define STBI_NO_BMP -#endif -#ifndef STBI_ONLY_PSD -#define STBI_NO_PSD -#endif -#ifndef STBI_ONLY_TGA -#define STBI_NO_TGA -#endif -#ifndef STBI_ONLY_GIF -#define STBI_NO_GIF -#endif -#ifndef STBI_ONLY_HDR -#define STBI_NO_HDR -#endif -#ifndef STBI_ONLY_PIC -#define STBI_NO_PIC -#endif -#ifndef STBI_ONLY_PNM -#define STBI_NO_PNM -#endif +#if defined(STBI_ONLY_JPEG) || defined(STBI_ONLY_PNG) || defined(STBI_ONLY_BMP) \ + || defined(STBI_ONLY_TGA) || defined(STBI_ONLY_GIF) || defined(STBI_ONLY_PSD) \ + || defined(STBI_ONLY_HDR) || defined(STBI_ONLY_PIC) || defined(STBI_ONLY_PNM) \ + || defined(STBI_ONLY_ZLIB) + #ifndef STBI_ONLY_JPEG + #define STBI_NO_JPEG + #endif + #ifndef STBI_ONLY_PNG + #define STBI_NO_PNG + #endif + #ifndef STBI_ONLY_BMP + #define STBI_NO_BMP + #endif + #ifndef STBI_ONLY_PSD + #define STBI_NO_PSD + #endif + #ifndef STBI_ONLY_TGA + #define STBI_NO_TGA + #endif + #ifndef STBI_ONLY_GIF + #define STBI_NO_GIF + #endif + #ifndef STBI_ONLY_HDR + #define STBI_NO_HDR + #endif + #ifndef STBI_ONLY_PIC + #define STBI_NO_PIC + #endif + #ifndef STBI_ONLY_PNM + #define STBI_NO_PNM + #endif #endif #if defined(STBI_NO_PNG) && !defined(STBI_SUPPORT_ZLIB) && !defined(STBI_NO_ZLIB) #define STBI_NO_ZLIB #endif -#include + #include #include // ptrdiff_t on osx #include #include +#include #if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) -#include // ldexp, pow +#include // ldexp, pow #endif #ifndef STBI_NO_STDIO @@ -608,54 +609,55 @@ STBIDEF int stbi_zlib_decode_noheader_buffer(char * obuffer, int olen, const cha #define STBI_EXTERN extern #endif + #ifndef _MSC_VER -#ifdef __cplusplus -#define stbi_inline inline + #ifdef __cplusplus + #define stbi_inline inline + #else + #define stbi_inline + #endif #else -#define stbi_inline -#endif -#else -#define stbi_inline __forceinline + #define stbi_inline __forceinline #endif #ifndef STBI_NO_THREAD_LOCALS -#if defined(__cplusplus) && __cplusplus >= 201103L -#define STBI_THREAD_LOCAL thread_local -#elif defined(__GNUC__) && __GNUC__ < 5 -#define STBI_THREAD_LOCAL __thread -#elif defined(_MSC_VER) -#define STBI_THREAD_LOCAL __declspec(thread) -#elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L && !defined(__STDC_NO_THREADS__) -#define STBI_THREAD_LOCAL _Thread_local -#endif + #if defined(__cplusplus) && __cplusplus >= 201103L + #define STBI_THREAD_LOCAL thread_local + #elif defined(__GNUC__) && __GNUC__ < 5 + #define STBI_THREAD_LOCAL __thread + #elif defined(_MSC_VER) + #define STBI_THREAD_LOCAL __declspec(thread) + #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 201112L && !defined(__STDC_NO_THREADS__) + #define STBI_THREAD_LOCAL _Thread_local + #endif -#ifndef STBI_THREAD_LOCAL -#if defined(__GNUC__) -#define STBI_THREAD_LOCAL __thread -#endif -#endif + #ifndef STBI_THREAD_LOCAL + #if defined(__GNUC__) + #define STBI_THREAD_LOCAL __thread + #endif + #endif #endif #if defined(_MSC_VER) || defined(__SYMBIAN32__) typedef unsigned short stbi__uint16; -typedef signed short stbi__int16; -typedef unsigned int stbi__uint32; -typedef signed int stbi__int32; +typedef signed short stbi__int16; +typedef unsigned int stbi__uint32; +typedef signed int stbi__int32; #else #include typedef uint16_t stbi__uint16; -typedef int16_t stbi__int16; +typedef int16_t stbi__int16; typedef uint32_t stbi__uint32; -typedef int32_t stbi__int32; +typedef int32_t stbi__int32; #endif // should produce compiler error if size is wrong -typedef unsigned char validate_uint32[sizeof(stbi__uint32) == 4 ? 1 : -1]; +typedef unsigned char validate_uint32[sizeof(stbi__uint32)==4 ? 1 : -1]; #ifdef _MSC_VER -#define STBI_NOTUSED(v) (void)(v) +#define STBI_NOTUSED(v) (void)(v) #else -#define STBI_NOTUSED(v) (void)sizeof(v) +#define STBI_NOTUSED(v) (void)sizeof(v) #endif #ifdef _MSC_VER @@ -663,9 +665,9 @@ typedef unsigned char validate_uint32[sizeof(stbi__uint32) == 4 ? 1 : -1]; #endif #ifdef STBI_HAS_LROTL -#define stbi_lrot(x, y) _lrotl(x, y) + #define stbi_lrot(x,y) _lrotl(x,y) #else -#define stbi_lrot(x, y) (((x) << (y)) | ((x) >> (-(y)&31))) + #define stbi_lrot(x,y) (((x) << (y)) | ((x) >> (-(y) & 31))) #endif #if defined(STBI_MALLOC) && defined(STBI_FREE) && (defined(STBI_REALLOC) || defined(STBI_REALLOC_SIZED)) @@ -677,13 +679,13 @@ typedef unsigned char validate_uint32[sizeof(stbi__uint32) == 4 ? 1 : -1]; #endif #ifndef STBI_MALLOC -#define STBI_MALLOC(sz) malloc(sz) -#define STBI_REALLOC(p, newsz) realloc(p, newsz) -#define STBI_FREE(p) free(p) +#define STBI_MALLOC(sz) malloc(sz) +#define STBI_REALLOC(p,newsz) realloc(p,newsz) +#define STBI_FREE(p) free(p) #endif #ifndef STBI_REALLOC_SIZED -#define STBI_REALLOC_SIZED(p, oldsz, newsz) STBI_REALLOC(p, newsz) +#define STBI_REALLOC_SIZED(p,oldsz,newsz) STBI_REALLOC(p,newsz) #endif // x86/x64 detection @@ -725,31 +727,34 @@ typedef unsigned char validate_uint32[sizeof(stbi__uint32) == 4 ? 1 : -1]; #ifdef _MSC_VER -#if _MSC_VER >= 1400 // not VC6 -#include // __cpuid -static int stbi__cpuid3(void) { - int info[4]; - __cpuid(info, 1); - return info[3]; +#if _MSC_VER >= 1400 // not VC6 +#include // __cpuid +static int stbi__cpuid3(void) +{ + int info[4]; + __cpuid(info,1); + return info[3]; } #else -static int stbi__cpuid3(void) { - int res; - __asm { +static int stbi__cpuid3(void) +{ + int res; + __asm { mov eax,1 cpuid mov res,edx - } - return res; + } + return res; } #endif #define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name #if !defined(STBI_NO_JPEG) && defined(STBI_SSE2) -static int stbi__sse2_available(void) { - int info3 = stbi__cpuid3(); - return ((info3 >> 26) & 1) != 0; +static int stbi__sse2_available(void) +{ + int info3 = stbi__cpuid3(); + return ((info3 >> 26) & 1) != 0; } #endif @@ -757,11 +762,12 @@ static int stbi__sse2_available(void) { #define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16))) #if !defined(STBI_NO_JPEG) && defined(STBI_SSE2) -static int stbi__sse2_available(void) { - // If we're even attempting to compile this on GCC/Clang, that means - // -msse2 is on, which means the compiler is allowed to use SSE2 - // instructions at will, and so are we. - return 1; +static int stbi__sse2_available(void) +{ + // If we're even attempting to compile this on GCC/Clang, that means + // -msse2 is on, which means the compiler is allowed to use SSE2 + // instructions at will, and so are we. + return 1; } #endif @@ -796,162 +802,190 @@ static int stbi__sse2_available(void) { // stbi__context structure is our basic context used by all images, so it // contains all the IO context, plus some basic image information -typedef struct { - stbi__uint32 img_x, img_y; - int img_n, img_out_n; +typedef struct +{ + stbi__uint32 img_x, img_y; + int img_n, img_out_n; - stbi_io_callbacks io; - void * io_user_data; + stbi_io_callbacks io; + void *io_user_data; - int read_from_callbacks; - int buflen; - stbi_uc buffer_start[128]; - int callback_already_read; + int read_from_callbacks; + int buflen; + stbi_uc buffer_start[128]; + int callback_already_read; - stbi_uc *img_buffer, *img_buffer_end; - stbi_uc *img_buffer_original, *img_buffer_original_end; + stbi_uc *img_buffer, *img_buffer_end; + stbi_uc *img_buffer_original, *img_buffer_original_end; } stbi__context; -static void stbi__refill_buffer(stbi__context * s); + +static void stbi__refill_buffer(stbi__context *s); // initialize a memory-decode context -static void stbi__start_mem(stbi__context * s, stbi_uc const * buffer, int len) { - s->io.read = NULL; - s->read_from_callbacks = 0; - s->callback_already_read = 0; - s->img_buffer = s->img_buffer_original = (stbi_uc *)buffer; - s->img_buffer_end = s->img_buffer_original_end = (stbi_uc *)buffer + len; +static void stbi__start_mem(stbi__context *s, stbi_uc const *buffer, int len) +{ + s->io.read = NULL; + s->read_from_callbacks = 0; + s->callback_already_read = 0; + s->img_buffer = s->img_buffer_original = (stbi_uc *) buffer; + s->img_buffer_end = s->img_buffer_original_end = (stbi_uc *) buffer+len; } // initialize a callback-based context -static void stbi__start_callbacks(stbi__context * s, stbi_io_callbacks * c, void * user) { - s->io = *c; - s->io_user_data = user; - s->buflen = sizeof(s->buffer_start); - s->read_from_callbacks = 1; - s->callback_already_read = 0; - s->img_buffer = s->img_buffer_original = s->buffer_start; - stbi__refill_buffer(s); - s->img_buffer_original_end = s->img_buffer_end; +static void stbi__start_callbacks(stbi__context *s, stbi_io_callbacks *c, void *user) +{ + s->io = *c; + s->io_user_data = user; + s->buflen = sizeof(s->buffer_start); + s->read_from_callbacks = 1; + s->callback_already_read = 0; + s->img_buffer = s->img_buffer_original = s->buffer_start; + stbi__refill_buffer(s); + s->img_buffer_original_end = s->img_buffer_end; } #ifndef STBI_NO_STDIO -static int stbi__stdio_read(void * user, char * data, int size) { return (int)fread(data, 1, size, (FILE *)user); } - -static void stbi__stdio_skip(void * user, int n) { - int ch; - fseek((FILE *)user, n, SEEK_CUR); - ch = fgetc((FILE *)user); /* have to read a byte to reset feof()'s flag */ - if (ch != EOF) { - ungetc(ch, (FILE *)user); /* push byte back onto stream if valid. */ - } +static int stbi__stdio_read(void *user, char *data, int size) +{ + return (int) fread(data,1,size,(FILE*) user); } -static int stbi__stdio_eof(void * user) { return feof((FILE *)user) || ferror((FILE *)user); } +static void stbi__stdio_skip(void *user, int n) +{ + int ch; + fseek((FILE*) user, n, SEEK_CUR); + ch = fgetc((FILE*) user); /* have to read a byte to reset feof()'s flag */ + if (ch != EOF) { + ungetc(ch, (FILE *) user); /* push byte back onto stream if valid. */ + } +} -static stbi_io_callbacks stbi__stdio_callbacks = { - stbi__stdio_read, - stbi__stdio_skip, - stbi__stdio_eof, +static int stbi__stdio_eof(void *user) +{ + return feof((FILE*) user) || ferror((FILE *) user); +} + +static stbi_io_callbacks stbi__stdio_callbacks = +{ + stbi__stdio_read, + stbi__stdio_skip, + stbi__stdio_eof, }; -static void stbi__start_file(stbi__context * s, FILE * f) { stbi__start_callbacks(s, &stbi__stdio_callbacks, (void *)f); } +static void stbi__start_file(stbi__context *s, FILE *f) +{ + stbi__start_callbacks(s, &stbi__stdio_callbacks, (void *) f); +} -// static void stop_file(stbi__context *s) { } +//static void stop_file(stbi__context *s) { } #endif // !STBI_NO_STDIO -static void stbi__rewind(stbi__context * s) { - // conceptually rewind SHOULD rewind to the beginning of the stream, - // but we just rewind to the beginning of the initial buffer, because - // we only use it after doing 'test', which only ever looks at at most 92 bytes - s->img_buffer = s->img_buffer_original; - s->img_buffer_end = s->img_buffer_original_end; +static void stbi__rewind(stbi__context *s) +{ + // conceptually rewind SHOULD rewind to the beginning of the stream, + // but we just rewind to the beginning of the initial buffer, because + // we only use it after doing 'test', which only ever looks at at most 92 bytes + s->img_buffer = s->img_buffer_original; + s->img_buffer_end = s->img_buffer_original_end; } -enum { STBI_ORDER_RGB, STBI_ORDER_BGR }; +enum +{ + STBI_ORDER_RGB, + STBI_ORDER_BGR +}; -typedef struct { - int bits_per_channel; - int num_channels; - int channel_order; +typedef struct +{ + int bits_per_channel; + int num_channels; + int channel_order; } stbi__result_info; #ifndef STBI_NO_JPEG -static int stbi__jpeg_test(stbi__context * s); -static void * stbi__jpeg_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static int stbi__jpeg_info(stbi__context * s, int * x, int * y, int * comp); +static int stbi__jpeg_test(stbi__context *s); +static void *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp); #endif #ifndef STBI_NO_PNG -static int stbi__png_test(stbi__context * s); -static void * stbi__png_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static int stbi__png_info(stbi__context * s, int * x, int * y, int * comp); -static int stbi__png_is16(stbi__context * s); +static int stbi__png_test(stbi__context *s); +static void *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__png_info(stbi__context *s, int *x, int *y, int *comp); +static int stbi__png_is16(stbi__context *s); #endif #ifndef STBI_NO_BMP -static int stbi__bmp_test(stbi__context * s); -static void * stbi__bmp_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static int stbi__bmp_info(stbi__context * s, int * x, int * y, int * comp); +static int stbi__bmp_test(stbi__context *s); +static void *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp); #endif #ifndef STBI_NO_TGA -static int stbi__tga_test(stbi__context * s); -static void * stbi__tga_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static int stbi__tga_info(stbi__context * s, int * x, int * y, int * comp); +static int stbi__tga_test(stbi__context *s); +static void *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__tga_info(stbi__context *s, int *x, int *y, int *comp); #endif #ifndef STBI_NO_PSD -static int stbi__psd_test(stbi__context * s); -static void * stbi__psd_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri, int bpc); -static int stbi__psd_info(stbi__context * s, int * x, int * y, int * comp); -static int stbi__psd_is16(stbi__context * s); +static int stbi__psd_test(stbi__context *s); +static void *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc); +static int stbi__psd_info(stbi__context *s, int *x, int *y, int *comp); +static int stbi__psd_is16(stbi__context *s); #endif #ifndef STBI_NO_HDR -static int stbi__hdr_test(stbi__context * s); -static float * stbi__hdr_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static int stbi__hdr_info(stbi__context * s, int * x, int * y, int * comp); +static int stbi__hdr_test(stbi__context *s); +static float *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp); #endif #ifndef STBI_NO_PIC -static int stbi__pic_test(stbi__context * s); -static void * stbi__pic_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static int stbi__pic_info(stbi__context * s, int * x, int * y, int * comp); +static int stbi__pic_test(stbi__context *s); +static void *stbi__pic_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__pic_info(stbi__context *s, int *x, int *y, int *comp); #endif #ifndef STBI_NO_GIF -static int stbi__gif_test(stbi__context * s); -static void * stbi__gif_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static void * stbi__load_gif_main(stbi__context * s, int ** delays, int * x, int * y, int * z, int * comp, int req_comp); -static int stbi__gif_info(stbi__context * s, int * x, int * y, int * comp); +static int stbi__gif_test(stbi__context *s); +static void *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static void *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y, int *z, int *comp, int req_comp); +static int stbi__gif_info(stbi__context *s, int *x, int *y, int *comp); #endif #ifndef STBI_NO_PNM -static int stbi__pnm_test(stbi__context * s); -static void * stbi__pnm_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static int stbi__pnm_info(stbi__context * s, int * x, int * y, int * comp); -static int stbi__pnm_is16(stbi__context * s); +static int stbi__pnm_test(stbi__context *s); +static void *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp); +static int stbi__pnm_is16(stbi__context *s); #endif static #ifdef STBI_THREAD_LOCAL - STBI_THREAD_LOCAL +STBI_THREAD_LOCAL #endif - const char * stbi__g_failure_reason; +const char *stbi__g_failure_reason; -STBIDEF const char * stbi_failure_reason(void) { return stbi__g_failure_reason; } +STBIDEF const char *stbi_failure_reason(void) +{ + return stbi__g_failure_reason; +} #ifndef STBI_NO_FAILURE_STRINGS -static int stbi__err(const char * str) { - stbi__g_failure_reason = str; - return 0; +static int stbi__err(const char *str) +{ + stbi__g_failure_reason = str; + return 0; } #endif -static void * stbi__malloc(size_t size) { return STBI_MALLOC(size); } +static void *stbi__malloc(size_t size) +{ + return STBI_MALLOC(size); +} // stb_image uses ints pervasively, including for offset calculations. // therefore the largest decoded image size we can support with the @@ -965,88 +999,88 @@ static void * stbi__malloc(size_t size) { return STBI_MALLOC(size); } // return 1 if the sum is valid, 0 on overflow. // negative terms are considered invalid. -static int stbi__addsizes_valid(int a, int b) { - if (b < 0) - return 0; - // now 0 <= b <= INT_MAX, hence also - // 0 <= INT_MAX - b <= INTMAX. - // And "a + b <= INT_MAX" (which might overflow) is the - // same as a <= INT_MAX - b (no overflow) - return a <= INT_MAX - b; +static int stbi__addsizes_valid(int a, int b) +{ + if (b < 0) return 0; + // now 0 <= b <= INT_MAX, hence also + // 0 <= INT_MAX - b <= INTMAX. + // And "a + b <= INT_MAX" (which might overflow) is the + // same as a <= INT_MAX - b (no overflow) + return a <= INT_MAX - b; } // returns 1 if the product is valid, 0 on overflow. // negative factors are considered invalid. -static int stbi__mul2sizes_valid(int a, int b) { - if (a < 0 || b < 0) - return 0; - if (b == 0) - return 1; // mul-by-0 is always safe - // portable way to check for no overflows in a*b - return a <= INT_MAX / b; +static int stbi__mul2sizes_valid(int a, int b) +{ + if (a < 0 || b < 0) return 0; + if (b == 0) return 1; // mul-by-0 is always safe + // portable way to check for no overflows in a*b + return a <= INT_MAX/b; } #if !defined(STBI_NO_JPEG) || !defined(STBI_NO_PNG) || !defined(STBI_NO_TGA) || !defined(STBI_NO_HDR) // returns 1 if "a*b + add" has no negative terms/factors and doesn't overflow -static int stbi__mad2sizes_valid(int a, int b, int add) { - return stbi__mul2sizes_valid(a, b) && stbi__addsizes_valid(a * b, add); +static int stbi__mad2sizes_valid(int a, int b, int add) +{ + return stbi__mul2sizes_valid(a, b) && stbi__addsizes_valid(a*b, add); } #endif // returns 1 if "a*b*c + add" has no negative terms/factors and doesn't overflow -static int stbi__mad3sizes_valid(int a, int b, int c, int add) { - return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a * b, c) && stbi__addsizes_valid(a * b * c, add); +static int stbi__mad3sizes_valid(int a, int b, int c, int add) +{ + return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) && + stbi__addsizes_valid(a*b*c, add); } // returns 1 if "a*b*c*d + add" has no negative terms/factors and doesn't overflow #if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM) -static int stbi__mad4sizes_valid(int a, int b, int c, int d, int add) { - return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a * b, c) && stbi__mul2sizes_valid(a * b * c, d) && - stbi__addsizes_valid(a * b * c * d, add); +static int stbi__mad4sizes_valid(int a, int b, int c, int d, int add) +{ + return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) && + stbi__mul2sizes_valid(a*b*c, d) && stbi__addsizes_valid(a*b*c*d, add); } #endif #if !defined(STBI_NO_JPEG) || !defined(STBI_NO_PNG) || !defined(STBI_NO_TGA) || !defined(STBI_NO_HDR) // mallocs with size overflow checking -static void * stbi__malloc_mad2(int a, int b, int add) { - if (!stbi__mad2sizes_valid(a, b, add)) - return NULL; - return stbi__malloc(a * b + add); +static void *stbi__malloc_mad2(int a, int b, int add) +{ + if (!stbi__mad2sizes_valid(a, b, add)) return NULL; + return stbi__malloc(a*b + add); } #endif -static void * stbi__malloc_mad3(int a, int b, int c, int add) { - if (!stbi__mad3sizes_valid(a, b, c, add)) - return NULL; - return stbi__malloc(a * b * c + add); +static void *stbi__malloc_mad3(int a, int b, int c, int add) +{ + if (!stbi__mad3sizes_valid(a, b, c, add)) return NULL; + return stbi__malloc(a*b*c + add); } #if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM) -static void * stbi__malloc_mad4(int a, int b, int c, int d, int add) { - if (!stbi__mad4sizes_valid(a, b, c, d, add)) - return NULL; - return stbi__malloc(a * b * c * d + add); +static void *stbi__malloc_mad4(int a, int b, int c, int d, int add) +{ + if (!stbi__mad4sizes_valid(a, b, c, d, add)) return NULL; + return stbi__malloc(a*b*c*d + add); } #endif // returns 1 if the sum of two signed ints is valid (between -2^31 and 2^31-1 inclusive), 0 on overflow. -static int stbi__addints_valid(int a, int b) { - if ((a >= 0) != (b >= 0)) - return 1; // a and b have different signs, so no overflow - if (a < 0 && b < 0) - return a >= INT_MIN - b; // same as a + b >= INT_MIN; INT_MIN - b cannot overflow since b < 0. - return a <= INT_MAX - b; +static int stbi__addints_valid(int a, int b) +{ + if ((a >= 0) != (b >= 0)) return 1; // a and b have different signs, so no overflow + if (a < 0 && b < 0) return a >= INT_MIN - b; // same as a + b >= INT_MIN; INT_MIN - b cannot overflow since b < 0. + return a <= INT_MAX - b; } -// returns 1 if the product of two signed shorts is valid, 0 on overflow. -static int stbi__mul2shorts_valid(short a, short b) { - if (b == 0 || b == -1) - return 1; // multiplication by 0 is always 0; check for -1 so SHRT_MIN/b doesn't overflow - if ((a >= 0) == (b >= 0)) - return a <= SHRT_MAX / b; // product is positive, so similar to mul2sizes_valid - if (b < 0) - return a <= SHRT_MIN / b; // same as a * b >= SHRT_MIN - return a >= SHRT_MIN / b; +// returns 1 if the product of two ints fits in a signed short, 0 on overflow. +static int stbi__mul2shorts_valid(int a, int b) +{ + if (b == 0 || b == -1) return 1; // multiplication by 0 is always 0; check for -1 so SHRT_MIN/b doesn't overflow + if ((a >= 0) == (b >= 0)) return a <= SHRT_MAX/b; // product is positive, so similar to mul2sizes_valid + if (b < 0) return a <= SHRT_MIN / b; // same as a * b >= SHRT_MIN + return a >= SHRT_MIN / b; } // stbi__err - error @@ -1054,411 +1088,423 @@ static int stbi__mul2shorts_valid(short a, short b) { // stbi__errpuc - error returning pointer to unsigned char #ifdef STBI_NO_FAILURE_STRINGS -#define stbi__err(x, y) 0 + #define stbi__err(x,y) 0 #elif defined(STBI_FAILURE_USERMSG) -#define stbi__err(x, y) stbi__err(y) + #define stbi__err(x,y) stbi__err(y) #else -#define stbi__err(x, y) stbi__err(x) + #define stbi__err(x,y) stbi__err(x) #endif -#define stbi__errpf(x, y) ((float *)(size_t)(stbi__err(x, y) ? NULL : NULL)) -#define stbi__errpuc(x, y) ((unsigned char *)(size_t)(stbi__err(x, y) ? NULL : NULL)) +#define stbi__errpf(x,y) ((float *)(size_t) (stbi__err(x,y)?NULL:NULL)) +#define stbi__errpuc(x,y) ((unsigned char *)(size_t) (stbi__err(x,y)?NULL:NULL)) -STBIDEF void stbi_image_free(void * retval_from_stbi_load) { STBI_FREE(retval_from_stbi_load); } +STBIDEF void stbi_image_free(void *retval_from_stbi_load) +{ + STBI_FREE(retval_from_stbi_load); +} #ifndef STBI_NO_LINEAR -static float * stbi__ldr_to_hdr(stbi_uc * data, int x, int y, int comp); +static float *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp); #endif #ifndef STBI_NO_HDR -static stbi_uc * stbi__hdr_to_ldr(float * data, int x, int y, int comp); +static stbi_uc *stbi__hdr_to_ldr(float *data, int x, int y, int comp); #endif static int stbi__vertically_flip_on_load_global = 0; -STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip) { - stbi__vertically_flip_on_load_global = flag_true_if_should_flip; +STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip) +{ + stbi__vertically_flip_on_load_global = flag_true_if_should_flip; } #ifndef STBI_THREAD_LOCAL -#define stbi__vertically_flip_on_load stbi__vertically_flip_on_load_global +#define stbi__vertically_flip_on_load stbi__vertically_flip_on_load_global #else static STBI_THREAD_LOCAL int stbi__vertically_flip_on_load_local, stbi__vertically_flip_on_load_set; -STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip) { - stbi__vertically_flip_on_load_local = flag_true_if_should_flip; - stbi__vertically_flip_on_load_set = 1; +STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip) +{ + stbi__vertically_flip_on_load_local = flag_true_if_should_flip; + stbi__vertically_flip_on_load_set = 1; } -#define stbi__vertically_flip_on_load \ - (stbi__vertically_flip_on_load_set ? stbi__vertically_flip_on_load_local : stbi__vertically_flip_on_load_global) +#define stbi__vertically_flip_on_load (stbi__vertically_flip_on_load_set \ + ? stbi__vertically_flip_on_load_local \ + : stbi__vertically_flip_on_load_global) #endif // STBI_THREAD_LOCAL -static void * stbi__load_main(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri, int bpc) { - memset(ri, 0, sizeof(*ri)); // make sure it's initialized if we add new fields - ri->bits_per_channel = 8; // default is 8 so most paths don't have to be changed - ri->channel_order = STBI_ORDER_RGB; // all current input & output are this, but this is here so we can add BGR order - ri->num_channels = 0; +static void *stbi__load_main(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc) +{ + memset(ri, 0, sizeof(*ri)); // make sure it's initialized if we add new fields + ri->bits_per_channel = 8; // default is 8 so most paths don't have to be changed + ri->channel_order = STBI_ORDER_RGB; // all current input & output are this, but this is here so we can add BGR order + ri->num_channels = 0; -// test the formats with a very explicit header first (at least a FOURCC -// or distinctive magic number first) -#ifndef STBI_NO_PNG - if (stbi__png_test(s)) - return stbi__png_load(s, x, y, comp, req_comp, ri); -#endif -#ifndef STBI_NO_BMP - if (stbi__bmp_test(s)) - return stbi__bmp_load(s, x, y, comp, req_comp, ri); -#endif -#ifndef STBI_NO_GIF - if (stbi__gif_test(s)) - return stbi__gif_load(s, x, y, comp, req_comp, ri); -#endif -#ifndef STBI_NO_PSD - if (stbi__psd_test(s)) - return stbi__psd_load(s, x, y, comp, req_comp, ri, bpc); -#else - STBI_NOTUSED(bpc); -#endif -#ifndef STBI_NO_PIC - if (stbi__pic_test(s)) - return stbi__pic_load(s, x, y, comp, req_comp, ri); -#endif + // test the formats with a very explicit header first (at least a FOURCC + // or distinctive magic number first) + #ifndef STBI_NO_PNG + if (stbi__png_test(s)) return stbi__png_load(s,x,y,comp,req_comp, ri); + #endif + #ifndef STBI_NO_BMP + if (stbi__bmp_test(s)) return stbi__bmp_load(s,x,y,comp,req_comp, ri); + #endif + #ifndef STBI_NO_GIF + if (stbi__gif_test(s)) return stbi__gif_load(s,x,y,comp,req_comp, ri); + #endif + #ifndef STBI_NO_PSD + if (stbi__psd_test(s)) return stbi__psd_load(s,x,y,comp,req_comp, ri, bpc); + #else + STBI_NOTUSED(bpc); + #endif + #ifndef STBI_NO_PIC + if (stbi__pic_test(s)) return stbi__pic_load(s,x,y,comp,req_comp, ri); + #endif -// then the formats that can end up attempting to load with just 1 or 2 -// bytes matching expectations; these are prone to false positives, so -// try them later -#ifndef STBI_NO_JPEG - if (stbi__jpeg_test(s)) - return stbi__jpeg_load(s, x, y, comp, req_comp, ri); -#endif -#ifndef STBI_NO_PNM - if (stbi__pnm_test(s)) - return stbi__pnm_load(s, x, y, comp, req_comp, ri); -#endif + // then the formats that can end up attempting to load with just 1 or 2 + // bytes matching expectations; these are prone to false positives, so + // try them later + #ifndef STBI_NO_JPEG + if (stbi__jpeg_test(s)) return stbi__jpeg_load(s,x,y,comp,req_comp, ri); + #endif + #ifndef STBI_NO_PNM + if (stbi__pnm_test(s)) return stbi__pnm_load(s,x,y,comp,req_comp, ri); + #endif -#ifndef STBI_NO_HDR - if (stbi__hdr_test(s)) { - float * hdr = stbi__hdr_load(s, x, y, comp, req_comp, ri); - return stbi__hdr_to_ldr(hdr, *x, *y, req_comp ? req_comp : *comp); - } -#endif + #ifndef STBI_NO_HDR + if (stbi__hdr_test(s)) { + float *hdr = stbi__hdr_load(s, x,y,comp,req_comp, ri); + return stbi__hdr_to_ldr(hdr, *x, *y, req_comp ? req_comp : *comp); + } + #endif -#ifndef STBI_NO_TGA - // test tga last because it's a crappy test! - if (stbi__tga_test(s)) - return stbi__tga_load(s, x, y, comp, req_comp, ri); -#endif + #ifndef STBI_NO_TGA + // test tga last because it's a crappy test! + if (stbi__tga_test(s)) + return stbi__tga_load(s,x,y,comp,req_comp, ri); + #endif - return stbi__errpuc("unknown image type", "Image not of any known type, or corrupt"); + return stbi__errpuc("unknown image type", "Image not of any known type, or corrupt"); } -static stbi_uc * stbi__convert_16_to_8(stbi__uint16 * orig, int w, int h, int channels) { - int i; - int img_len = w * h * channels; - stbi_uc * reduced; +static stbi_uc *stbi__convert_16_to_8(stbi__uint16 *orig, int w, int h, int channels) +{ + int i; + int img_len = w * h * channels; + stbi_uc *reduced; - reduced = (stbi_uc *)stbi__malloc(img_len); - if (reduced == NULL) - return stbi__errpuc("outofmem", "Out of memory"); + reduced = (stbi_uc *) stbi__malloc(img_len); + if (reduced == NULL) return stbi__errpuc("outofmem", "Out of memory"); - for (i = 0; i < img_len; ++i) - reduced[i] = (stbi_uc)((orig[i] >> 8) & 0xFF); // top half of each byte is sufficient approx of 16->8 bit scaling + for (i = 0; i < img_len; ++i) + reduced[i] = (stbi_uc)((orig[i] >> 8) & 0xFF); // top half of each byte is sufficient approx of 16->8 bit scaling - STBI_FREE(orig); - return reduced; + STBI_FREE(orig); + return reduced; } -static stbi__uint16 * stbi__convert_8_to_16(stbi_uc * orig, int w, int h, int channels) { - int i; - int img_len = w * h * channels; - stbi__uint16 * enlarged; +static stbi__uint16 *stbi__convert_8_to_16(stbi_uc *orig, int w, int h, int channels) +{ + int i; + int img_len = w * h * channels; + stbi__uint16 *enlarged; - enlarged = (stbi__uint16 *)stbi__malloc(img_len * 2); - if (enlarged == NULL) - return (stbi__uint16 *)stbi__errpuc("outofmem", "Out of memory"); + enlarged = (stbi__uint16 *) stbi__malloc(img_len*2); + if (enlarged == NULL) return (stbi__uint16 *) stbi__errpuc("outofmem", "Out of memory"); - for (i = 0; i < img_len; ++i) - enlarged[i] = (stbi__uint16)((orig[i] << 8) + orig[i]); // replicate to high and low byte, maps 0->0, 255->0xffff + for (i = 0; i < img_len; ++i) + enlarged[i] = (stbi__uint16)((orig[i] << 8) + orig[i]); // replicate to high and low byte, maps 0->0, 255->0xffff - STBI_FREE(orig); - return enlarged; + STBI_FREE(orig); + return enlarged; } -static void stbi__vertical_flip(void * image, int w, int h, int bytes_per_pixel) { - int row; - size_t bytes_per_row = (size_t)w * bytes_per_pixel; - stbi_uc temp[2048]; - stbi_uc * bytes = (stbi_uc *)image; +static void stbi__vertical_flip(void *image, int w, int h, int bytes_per_pixel) +{ + int row; + size_t bytes_per_row = (size_t)w * bytes_per_pixel; + stbi_uc temp[2048]; + stbi_uc *bytes = (stbi_uc *)image; - for (row = 0; row < (h >> 1); row++) { - stbi_uc * row0 = bytes + row * bytes_per_row; - stbi_uc * row1 = bytes + (h - row - 1) * bytes_per_row; - // swap row0 with row1 - size_t bytes_left = bytes_per_row; - while (bytes_left) { - size_t bytes_copy = (bytes_left < sizeof(temp)) ? bytes_left : sizeof(temp); - memcpy(temp, row0, bytes_copy); - memcpy(row0, row1, bytes_copy); - memcpy(row1, temp, bytes_copy); - row0 += bytes_copy; - row1 += bytes_copy; - bytes_left -= bytes_copy; - } - } + for (row = 0; row < (h>>1); row++) { + stbi_uc *row0 = bytes + row*bytes_per_row; + stbi_uc *row1 = bytes + (h - row - 1)*bytes_per_row; + // swap row0 with row1 + size_t bytes_left = bytes_per_row; + while (bytes_left) { + size_t bytes_copy = (bytes_left < sizeof(temp)) ? bytes_left : sizeof(temp); + memcpy(temp, row0, bytes_copy); + memcpy(row0, row1, bytes_copy); + memcpy(row1, temp, bytes_copy); + row0 += bytes_copy; + row1 += bytes_copy; + bytes_left -= bytes_copy; + } + } } #ifndef STBI_NO_GIF -static void stbi__vertical_flip_slices(void * image, int w, int h, int z, int bytes_per_pixel) { - int slice; - int slice_size = w * h * bytes_per_pixel; +static void stbi__vertical_flip_slices(void *image, int w, int h, int z, int bytes_per_pixel) +{ + int slice; + int slice_size = w * h * bytes_per_pixel; - stbi_uc * bytes = (stbi_uc *)image; - for (slice = 0; slice < z; ++slice) { - stbi__vertical_flip(bytes, w, h, bytes_per_pixel); - bytes += slice_size; - } + stbi_uc *bytes = (stbi_uc *)image; + for (slice = 0; slice < z; ++slice) { + stbi__vertical_flip(bytes, w, h, bytes_per_pixel); + bytes += slice_size; + } } #endif -static unsigned char * stbi__load_and_postprocess_8bit(stbi__context * s, int * x, int * y, int * comp, int req_comp) { - stbi__result_info ri; - void * result = stbi__load_main(s, x, y, comp, req_comp, &ri, 8); +static unsigned char *stbi__load_and_postprocess_8bit(stbi__context *s, int *x, int *y, int *comp, int req_comp) +{ + stbi__result_info ri; + void *result = stbi__load_main(s, x, y, comp, req_comp, &ri, 8); - if (result == NULL) - return NULL; + if (result == NULL) + return NULL; - // it is the responsibility of the loaders to make sure we get either 8 or 16 bit. - STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16); + // it is the responsibility of the loaders to make sure we get either 8 or 16 bit. + STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16); - if (ri.bits_per_channel != 8) { - result = stbi__convert_16_to_8((stbi__uint16 *)result, *x, *y, req_comp == 0 ? *comp : req_comp); - ri.bits_per_channel = 8; - } + if (ri.bits_per_channel != 8) { + result = stbi__convert_16_to_8((stbi__uint16 *) result, *x, *y, req_comp == 0 ? *comp : req_comp); + ri.bits_per_channel = 8; + } - // @TODO: move stbi__convert_format to here + // @TODO: move stbi__convert_format to here - if (stbi__vertically_flip_on_load) { - int channels = req_comp ? req_comp : *comp; - stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi_uc)); - } + if (stbi__vertically_flip_on_load) { + int channels = req_comp ? req_comp : *comp; + stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi_uc)); + } - return (unsigned char *)result; + return (unsigned char *) result; } -static stbi__uint16 * stbi__load_and_postprocess_16bit(stbi__context * s, int * x, int * y, int * comp, int req_comp) { - stbi__result_info ri; - void * result = stbi__load_main(s, x, y, comp, req_comp, &ri, 16); +static stbi__uint16 *stbi__load_and_postprocess_16bit(stbi__context *s, int *x, int *y, int *comp, int req_comp) +{ + stbi__result_info ri; + void *result = stbi__load_main(s, x, y, comp, req_comp, &ri, 16); - if (result == NULL) - return NULL; + if (result == NULL) + return NULL; - // it is the responsibility of the loaders to make sure we get either 8 or 16 bit. - STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16); + // it is the responsibility of the loaders to make sure we get either 8 or 16 bit. + STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16); - if (ri.bits_per_channel != 16) { - result = stbi__convert_8_to_16((stbi_uc *)result, *x, *y, req_comp == 0 ? *comp : req_comp); - ri.bits_per_channel = 16; - } + if (ri.bits_per_channel != 16) { + result = stbi__convert_8_to_16((stbi_uc *) result, *x, *y, req_comp == 0 ? *comp : req_comp); + ri.bits_per_channel = 16; + } - // @TODO: move stbi__convert_format16 to here - // @TODO: special case RGB-to-Y (and RGBA-to-YA) for 8-bit-to-16-bit case to keep more precision + // @TODO: move stbi__convert_format16 to here + // @TODO: special case RGB-to-Y (and RGBA-to-YA) for 8-bit-to-16-bit case to keep more precision - if (stbi__vertically_flip_on_load) { - int channels = req_comp ? req_comp : *comp; - stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi__uint16)); - } + if (stbi__vertically_flip_on_load) { + int channels = req_comp ? req_comp : *comp; + stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi__uint16)); + } - return (stbi__uint16 *)result; + return (stbi__uint16 *) result; } #if !defined(STBI_NO_HDR) && !defined(STBI_NO_LINEAR) -static void stbi__float_postprocess(float * result, int * x, int * y, int * comp, int req_comp) { - if (stbi__vertically_flip_on_load && result != NULL) { - int channels = req_comp ? req_comp : *comp; - stbi__vertical_flip(result, *x, *y, channels * sizeof(float)); - } +static void stbi__float_postprocess(float *result, int *x, int *y, int *comp, int req_comp) +{ + if (stbi__vertically_flip_on_load && result != NULL) { + int channels = req_comp ? req_comp : *comp; + stbi__vertical_flip(result, *x, *y, channels * sizeof(float)); + } } #endif #ifndef STBI_NO_STDIO #if defined(_WIN32) && defined(STBI_WINDOWS_UTF8) -STBI_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char * str, - int cbmb, wchar_t * widestr, int cchwide); -STBI_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags, - const wchar_t * widestr, int cchwide, char * str, int cbmb, - const char * defchar, int * used_default); +STBI_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char *str, int cbmb, wchar_t *widestr, int cchwide); +STBI_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags, const wchar_t *widestr, int cchwide, char *str, int cbmb, const char *defchar, int *used_default); #endif #if defined(_WIN32) && defined(STBI_WINDOWS_UTF8) -STBIDEF int stbi_convert_wchar_to_utf8(char * buffer, size_t bufferlen, const wchar_t * input) { - return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int)bufferlen, NULL, NULL); +STBIDEF int stbi_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input) +{ + return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int) bufferlen, NULL, NULL); } #endif -static FILE * stbi__fopen(char const * filename, char const * mode) { - FILE * f; +static FILE *stbi__fopen(char const *filename, char const *mode) +{ + FILE *f; #if defined(_WIN32) && defined(STBI_WINDOWS_UTF8) - wchar_t wMode[64]; - wchar_t wFilename[1024]; - if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename) / sizeof(*wFilename))) - return 0; + wchar_t wMode[64]; + wchar_t wFilename[1024]; + if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename)/sizeof(*wFilename))) + return 0; - if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode) / sizeof(*wMode))) - return 0; + if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode)/sizeof(*wMode))) + return 0; #if defined(_MSC_VER) && _MSC_VER >= 1400 - if (0 != _wfopen_s(&f, wFilename, wMode)) - f = 0; + if (0 != _wfopen_s(&f, wFilename, wMode)) + f = 0; #else - f = _wfopen(wFilename, wMode); + f = _wfopen(wFilename, wMode); #endif #elif defined(_MSC_VER) && _MSC_VER >= 1400 - if (0 != fopen_s(&f, filename, mode)) - f = 0; + if (0 != fopen_s(&f, filename, mode)) + f=0; #else - f = fopen(filename, mode); + f = fopen(filename, mode); #endif - return f; + return f; } -STBIDEF stbi_uc * stbi_load(char const * filename, int * x, int * y, int * comp, int req_comp) { - FILE * f = stbi__fopen(filename, "rb"); - unsigned char * result; - if (!f) - return stbi__errpuc("can't fopen", "Unable to open file"); - result = stbi_load_from_file(f, x, y, comp, req_comp); - fclose(f); - return result; + +STBIDEF stbi_uc *stbi_load(char const *filename, int *x, int *y, int *comp, int req_comp) +{ + FILE *f = stbi__fopen(filename, "rb"); + unsigned char *result; + if (!f) return stbi__errpuc("can't fopen", "Unable to open file"); + result = stbi_load_from_file(f,x,y,comp,req_comp); + fclose(f); + return result; } -STBIDEF stbi_uc * stbi_load_from_file(FILE * f, int * x, int * y, int * comp, int req_comp) { - unsigned char * result; - stbi__context s; - stbi__start_file(&s, f); - result = stbi__load_and_postprocess_8bit(&s, x, y, comp, req_comp); - if (result) { - // need to 'unget' all the characters in the IO buffer - fseek(f, -(int)(s.img_buffer_end - s.img_buffer), SEEK_CUR); - } - return result; +STBIDEF stbi_uc *stbi_load_from_file(FILE *f, int *x, int *y, int *comp, int req_comp) +{ + unsigned char *result; + stbi__context s; + stbi__start_file(&s,f); + result = stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp); + if (result) { + // need to 'unget' all the characters in the IO buffer + fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR); + } + return result; } -STBIDEF stbi__uint16 * stbi_load_from_file_16(FILE * f, int * x, int * y, int * comp, int req_comp) { - stbi__uint16 * result; - stbi__context s; - stbi__start_file(&s, f); - result = stbi__load_and_postprocess_16bit(&s, x, y, comp, req_comp); - if (result) { - // need to 'unget' all the characters in the IO buffer - fseek(f, -(int)(s.img_buffer_end - s.img_buffer), SEEK_CUR); - } - return result; +STBIDEF stbi__uint16 *stbi_load_from_file_16(FILE *f, int *x, int *y, int *comp, int req_comp) +{ + stbi__uint16 *result; + stbi__context s; + stbi__start_file(&s,f); + result = stbi__load_and_postprocess_16bit(&s,x,y,comp,req_comp); + if (result) { + // need to 'unget' all the characters in the IO buffer + fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR); + } + return result; } -STBIDEF stbi_us * stbi_load_16(char const * filename, int * x, int * y, int * comp, int req_comp) { - FILE * f = stbi__fopen(filename, "rb"); - stbi__uint16 * result; - if (!f) - return (stbi_us *)stbi__errpuc("can't fopen", "Unable to open file"); - result = stbi_load_from_file_16(f, x, y, comp, req_comp); - fclose(f); - return result; +STBIDEF stbi_us *stbi_load_16(char const *filename, int *x, int *y, int *comp, int req_comp) +{ + FILE *f = stbi__fopen(filename, "rb"); + stbi__uint16 *result; + if (!f) return (stbi_us *) stbi__errpuc("can't fopen", "Unable to open file"); + result = stbi_load_from_file_16(f,x,y,comp,req_comp); + fclose(f); + return result; } -#endif //! STBI_NO_STDIO -STBIDEF stbi_us * stbi_load_16_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * channels_in_file, - int desired_channels) { - stbi__context s; - stbi__start_mem(&s, buffer, len); - return stbi__load_and_postprocess_16bit(&s, x, y, channels_in_file, desired_channels); +#endif //!STBI_NO_STDIO + +STBIDEF stbi_us *stbi_load_16_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels) +{ + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__load_and_postprocess_16bit(&s,x,y,channels_in_file,desired_channels); } -STBIDEF stbi_us * stbi_load_16_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, - int * channels_in_file, int desired_channels) { - stbi__context s; - stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user); - return stbi__load_and_postprocess_16bit(&s, x, y, channels_in_file, desired_channels); +STBIDEF stbi_us *stbi_load_16_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels) +{ + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user); + return stbi__load_and_postprocess_16bit(&s,x,y,channels_in_file,desired_channels); } -STBIDEF stbi_uc * stbi_load_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * comp, int req_comp) { - stbi__context s; - stbi__start_mem(&s, buffer, len); - return stbi__load_and_postprocess_8bit(&s, x, y, comp, req_comp); +STBIDEF stbi_uc *stbi_load_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp); } -STBIDEF stbi_uc * stbi_load_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, int * comp, - int req_comp) { - stbi__context s; - stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user); - return stbi__load_and_postprocess_8bit(&s, x, y, comp, req_comp); +STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user); + return stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp); } #ifndef STBI_NO_GIF -STBIDEF stbi_uc * stbi_load_gif_from_memory(stbi_uc const * buffer, int len, int ** delays, int * x, int * y, int * z, - int * comp, int req_comp) { - unsigned char * result; - stbi__context s; - stbi__start_mem(&s, buffer, len); +STBIDEF stbi_uc *stbi_load_gif_from_memory(stbi_uc const *buffer, int len, int **delays, int *x, int *y, int *z, int *comp, int req_comp) +{ + unsigned char *result; + stbi__context s; + stbi__start_mem(&s,buffer,len); - result = (unsigned char *)stbi__load_gif_main(&s, delays, x, y, z, comp, req_comp); - if (stbi__vertically_flip_on_load) { - stbi__vertical_flip_slices(result, *x, *y, *z, *comp); - } + result = (unsigned char*) stbi__load_gif_main(&s, delays, x, y, z, comp, req_comp); + if (stbi__vertically_flip_on_load) { + stbi__vertical_flip_slices( result, *x, *y, *z, *comp ); + } - return result; + return result; } #endif #ifndef STBI_NO_LINEAR -static float * stbi__loadf_main(stbi__context * s, int * x, int * y, int * comp, int req_comp) { - unsigned char * data; -#ifndef STBI_NO_HDR - if (stbi__hdr_test(s)) { - stbi__result_info ri; - float * hdr_data = stbi__hdr_load(s, x, y, comp, req_comp, &ri); - if (hdr_data) - stbi__float_postprocess(hdr_data, x, y, comp, req_comp); - return hdr_data; - } -#endif - data = stbi__load_and_postprocess_8bit(s, x, y, comp, req_comp); - if (data) - return stbi__ldr_to_hdr(data, *x, *y, req_comp ? req_comp : *comp); - return stbi__errpf("unknown image type", "Image not of any known type, or corrupt"); +static float *stbi__loadf_main(stbi__context *s, int *x, int *y, int *comp, int req_comp) +{ + unsigned char *data; + #ifndef STBI_NO_HDR + if (stbi__hdr_test(s)) { + stbi__result_info ri; + float *hdr_data = stbi__hdr_load(s,x,y,comp,req_comp, &ri); + if (hdr_data) + stbi__float_postprocess(hdr_data,x,y,comp,req_comp); + return hdr_data; + } + #endif + data = stbi__load_and_postprocess_8bit(s, x, y, comp, req_comp); + if (data) + return stbi__ldr_to_hdr(data, *x, *y, req_comp ? req_comp : *comp); + return stbi__errpf("unknown image type", "Image not of any known type, or corrupt"); } -STBIDEF float * stbi_loadf_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * comp, int req_comp) { - stbi__context s; - stbi__start_mem(&s, buffer, len); - return stbi__loadf_main(&s, x, y, comp, req_comp); +STBIDEF float *stbi_loadf_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__loadf_main(&s,x,y,comp,req_comp); } -STBIDEF float * stbi_loadf_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, int * comp, - int req_comp) { - stbi__context s; - stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user); - return stbi__loadf_main(&s, x, y, comp, req_comp); +STBIDEF float *stbi_loadf_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user); + return stbi__loadf_main(&s,x,y,comp,req_comp); } #ifndef STBI_NO_STDIO -STBIDEF float * stbi_loadf(char const * filename, int * x, int * y, int * comp, int req_comp) { - float * result; - FILE * f = stbi__fopen(filename, "rb"); - if (!f) - return stbi__errpf("can't fopen", "Unable to open file"); - result = stbi_loadf_from_file(f, x, y, comp, req_comp); - fclose(f); - return result; +STBIDEF float *stbi_loadf(char const *filename, int *x, int *y, int *comp, int req_comp) +{ + float *result; + FILE *f = stbi__fopen(filename, "rb"); + if (!f) return stbi__errpf("can't fopen", "Unable to open file"); + result = stbi_loadf_from_file(f,x,y,comp,req_comp); + fclose(f); + return result; } -STBIDEF float * stbi_loadf_from_file(FILE * f, int * x, int * y, int * comp, int req_comp) { - stbi__context s; - stbi__start_file(&s, f); - return stbi__loadf_main(&s, x, y, comp, req_comp); +STBIDEF float *stbi_loadf_from_file(FILE *f, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_file(&s,f); + return stbi__loadf_main(&s,x,y,comp,req_comp); } #endif // !STBI_NO_STDIO @@ -1468,208 +1514,222 @@ STBIDEF float * stbi_loadf_from_file(FILE * f, int * x, int * y, int * comp, int // defined, for API simplicity; if STBI_NO_LINEAR is defined, it always // reports false! -STBIDEF int stbi_is_hdr_from_memory(stbi_uc const * buffer, int len) { -#ifndef STBI_NO_HDR - stbi__context s; - stbi__start_mem(&s, buffer, len); - return stbi__hdr_test(&s); -#else - STBI_NOTUSED(buffer); - STBI_NOTUSED(len); - return 0; -#endif +STBIDEF int stbi_is_hdr_from_memory(stbi_uc const *buffer, int len) +{ + #ifndef STBI_NO_HDR + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__hdr_test(&s); + #else + STBI_NOTUSED(buffer); + STBI_NOTUSED(len); + return 0; + #endif } #ifndef STBI_NO_STDIO -STBIDEF int stbi_is_hdr(char const * filename) { - FILE * f = stbi__fopen(filename, "rb"); - int result = 0; - if (f) { - result = stbi_is_hdr_from_file(f); - fclose(f); - } - return result; +STBIDEF int stbi_is_hdr (char const *filename) +{ + FILE *f = stbi__fopen(filename, "rb"); + int result=0; + if (f) { + result = stbi_is_hdr_from_file(f); + fclose(f); + } + return result; } -STBIDEF int stbi_is_hdr_from_file(FILE * f) { -#ifndef STBI_NO_HDR - long pos = ftell(f); - int res; - stbi__context s; - stbi__start_file(&s, f); - res = stbi__hdr_test(&s); - fseek(f, pos, SEEK_SET); - return res; -#else - STBI_NOTUSED(f); - return 0; -#endif +STBIDEF int stbi_is_hdr_from_file(FILE *f) +{ + #ifndef STBI_NO_HDR + long pos = ftell(f); + int res; + stbi__context s; + stbi__start_file(&s,f); + res = stbi__hdr_test(&s); + fseek(f, pos, SEEK_SET); + return res; + #else + STBI_NOTUSED(f); + return 0; + #endif } #endif // !STBI_NO_STDIO -STBIDEF int stbi_is_hdr_from_callbacks(stbi_io_callbacks const * clbk, void * user) { -#ifndef STBI_NO_HDR - stbi__context s; - stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user); - return stbi__hdr_test(&s); -#else - STBI_NOTUSED(clbk); - STBI_NOTUSED(user); - return 0; -#endif +STBIDEF int stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user) +{ + #ifndef STBI_NO_HDR + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user); + return stbi__hdr_test(&s); + #else + STBI_NOTUSED(clbk); + STBI_NOTUSED(user); + return 0; + #endif } #ifndef STBI_NO_LINEAR -static float stbi__l2h_gamma = 2.2f, stbi__l2h_scale = 1.0f; +static float stbi__l2h_gamma=2.2f, stbi__l2h_scale=1.0f; -STBIDEF void stbi_ldr_to_hdr_gamma(float gamma) { stbi__l2h_gamma = gamma; } -STBIDEF void stbi_ldr_to_hdr_scale(float scale) { stbi__l2h_scale = scale; } +STBIDEF void stbi_ldr_to_hdr_gamma(float gamma) { stbi__l2h_gamma = gamma; } +STBIDEF void stbi_ldr_to_hdr_scale(float scale) { stbi__l2h_scale = scale; } #endif -static float stbi__h2l_gamma_i = 1.0f / 2.2f, stbi__h2l_scale_i = 1.0f; +static float stbi__h2l_gamma_i=1.0f/2.2f, stbi__h2l_scale_i=1.0f; + +STBIDEF void stbi_hdr_to_ldr_gamma(float gamma) { stbi__h2l_gamma_i = 1/gamma; } +STBIDEF void stbi_hdr_to_ldr_scale(float scale) { stbi__h2l_scale_i = 1/scale; } -STBIDEF void stbi_hdr_to_ldr_gamma(float gamma) { stbi__h2l_gamma_i = 1 / gamma; } -STBIDEF void stbi_hdr_to_ldr_scale(float scale) { stbi__h2l_scale_i = 1 / scale; } ////////////////////////////////////////////////////////////////////////////// // // Common code used by all image loaders // -enum { STBI__SCAN_load = 0, STBI__SCAN_type, STBI__SCAN_header }; +enum +{ + STBI__SCAN_load=0, + STBI__SCAN_type, + STBI__SCAN_header +}; -static void stbi__refill_buffer(stbi__context * s) { - int n = (s->io.read)(s->io_user_data, (char *)s->buffer_start, s->buflen); - s->callback_already_read += (int)(s->img_buffer - s->img_buffer_original); - if (n == 0) { - // at end of file, treat same as if from memory, but need to handle case - // where s->img_buffer isn't pointing to safe memory, e.g. 0-byte file - s->read_from_callbacks = 0; - s->img_buffer = s->buffer_start; - s->img_buffer_end = s->buffer_start + 1; - *s->img_buffer = 0; - } else { - s->img_buffer = s->buffer_start; - s->img_buffer_end = s->buffer_start + n; - } +static void stbi__refill_buffer(stbi__context *s) +{ + int n = (s->io.read)(s->io_user_data,(char*)s->buffer_start,s->buflen); + s->callback_already_read += (int) (s->img_buffer - s->img_buffer_original); + if (n == 0) { + // at end of file, treat same as if from memory, but need to handle case + // where s->img_buffer isn't pointing to safe memory, e.g. 0-byte file + s->read_from_callbacks = 0; + s->img_buffer = s->buffer_start; + s->img_buffer_end = s->buffer_start+1; + *s->img_buffer = 0; + } else { + s->img_buffer = s->buffer_start; + s->img_buffer_end = s->buffer_start + n; + } } -stbi_inline static stbi_uc stbi__get8(stbi__context * s) { - if (s->img_buffer < s->img_buffer_end) - return *s->img_buffer++; - if (s->read_from_callbacks) { - stbi__refill_buffer(s); - return *s->img_buffer++; - } - return 0; +stbi_inline static stbi_uc stbi__get8(stbi__context *s) +{ + if (s->img_buffer < s->img_buffer_end) + return *s->img_buffer++; + if (s->read_from_callbacks) { + stbi__refill_buffer(s); + return *s->img_buffer++; + } + return 0; } #if defined(STBI_NO_JPEG) && defined(STBI_NO_HDR) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM) // nothing #else -stbi_inline static int stbi__at_eof(stbi__context * s) { - if (s->io.read) { - if (!(s->io.eof)(s->io_user_data)) - return 0; - // if feof() is true, check if buffer = end - // special case: we've only got the special 0 character at the end - if (s->read_from_callbacks == 0) - return 1; - } +stbi_inline static int stbi__at_eof(stbi__context *s) +{ + if (s->io.read) { + if (!(s->io.eof)(s->io_user_data)) return 0; + // if feof() is true, check if buffer = end + // special case: we've only got the special 0 character at the end + if (s->read_from_callbacks == 0) return 1; + } - return s->img_buffer >= s->img_buffer_end; + return s->img_buffer >= s->img_buffer_end; } #endif -#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && \ - defined(STBI_NO_GIF) && defined(STBI_NO_PIC) +#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC) // nothing #else -static void stbi__skip(stbi__context * s, int n) { - if (n == 0) - return; // already there! - if (n < 0) { - s->img_buffer = s->img_buffer_end; - return; - } - if (s->io.read) { - int blen = (int)(s->img_buffer_end - s->img_buffer); - if (blen < n) { - s->img_buffer = s->img_buffer_end; - (s->io.skip)(s->io_user_data, n - blen); - return; - } - } - s->img_buffer += n; +static void stbi__skip(stbi__context *s, int n) +{ + if (n == 0) return; // already there! + if (n < 0) { + s->img_buffer = s->img_buffer_end; + return; + } + if (s->io.read) { + int blen = (int) (s->img_buffer_end - s->img_buffer); + if (blen < n) { + s->img_buffer = s->img_buffer_end; + (s->io.skip)(s->io_user_data, n - blen); + return; + } + } + s->img_buffer += n; } #endif #if defined(STBI_NO_PNG) && defined(STBI_NO_TGA) && defined(STBI_NO_HDR) && defined(STBI_NO_PNM) // nothing #else -static int stbi__getn(stbi__context * s, stbi_uc * buffer, int n) { - if (s->io.read) { - int blen = (int)(s->img_buffer_end - s->img_buffer); - if (blen < n) { - int res, count; +static int stbi__getn(stbi__context *s, stbi_uc *buffer, int n) +{ + if (s->io.read) { + int blen = (int) (s->img_buffer_end - s->img_buffer); + if (blen < n) { + int res, count; - memcpy(buffer, s->img_buffer, blen); + memcpy(buffer, s->img_buffer, blen); - count = (s->io.read)(s->io_user_data, (char *)buffer + blen, n - blen); - res = (count == (n - blen)); - s->img_buffer = s->img_buffer_end; - return res; - } - } + count = (s->io.read)(s->io_user_data, (char*) buffer + blen, n - blen); + res = (count == (n-blen)); + s->img_buffer = s->img_buffer_end; + return res; + } + } - if (s->img_buffer + n <= s->img_buffer_end) { - memcpy(buffer, s->img_buffer, n); - s->img_buffer += n; - return 1; - } else - return 0; + if (s->img_buffer+n <= s->img_buffer_end) { + memcpy(buffer, s->img_buffer, n); + s->img_buffer += n; + return 1; + } else + return 0; } #endif #if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_PSD) && defined(STBI_NO_PIC) // nothing #else -static int stbi__get16be(stbi__context * s) { - int z = stbi__get8(s); - return (z << 8) + stbi__get8(s); +static int stbi__get16be(stbi__context *s) +{ + int z = stbi__get8(s); + return (z << 8) + stbi__get8(s); } #endif #if defined(STBI_NO_PNG) && defined(STBI_NO_PSD) && defined(STBI_NO_PIC) // nothing #else -static stbi__uint32 stbi__get32be(stbi__context * s) { - stbi__uint32 z = stbi__get16be(s); - return (z << 16) + stbi__get16be(s); +static stbi__uint32 stbi__get32be(stbi__context *s) +{ + stbi__uint32 z = stbi__get16be(s); + return (z << 16) + stbi__get16be(s); } #endif #if defined(STBI_NO_BMP) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) // nothing #else -static int stbi__get16le(stbi__context * s) { - int z = stbi__get8(s); - return z + (stbi__get8(s) << 8); +static int stbi__get16le(stbi__context *s) +{ + int z = stbi__get8(s); + return z + (stbi__get8(s) << 8); } #endif #ifndef STBI_NO_BMP -static stbi__uint32 stbi__get32le(stbi__context * s) { - stbi__uint32 z = stbi__get16le(s); - z += (stbi__uint32)stbi__get16le(s) << 16; - return z; +static stbi__uint32 stbi__get32le(stbi__context *s) +{ + stbi__uint32 z = stbi__get16le(s); + z += (stbi__uint32)stbi__get16le(s) << 16; + return z; } #endif -#define STBI__BYTECAST(x) ((stbi_uc)((x)&255)) // truncate int to byte without warnings +#define STBI__BYTECAST(x) ((stbi_uc) ((x) & 255)) // truncate int to byte without warnings -#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && \ - defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM) +#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM) // nothing #else ////////////////////////////////////////////////////////////////////////////// @@ -1683,264 +1743,169 @@ static stbi__uint32 stbi__get32le(stbi__context * s) { // assume data buffer is malloced, so malloc a new one and free that one // only failure mode is malloc failing -static stbi_uc stbi__compute_y(int r, int g, int b) { return (stbi_uc)(((r * 77) + (g * 150) + (29 * b)) >> 8); } +static stbi_uc stbi__compute_y(int r, int g, int b) +{ + return (stbi_uc) (((r*77) + (g*150) + (29*b)) >> 8); +} #endif -#if defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && \ - defined(STBI_NO_PIC) && defined(STBI_NO_PNM) +#if defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM) // nothing #else -static unsigned char * stbi__convert_format(unsigned char * data, int img_n, int req_comp, unsigned int x, unsigned int y) { - int i, j; - unsigned char * good; +static unsigned char *stbi__convert_format(unsigned char *data, int img_n, int req_comp, unsigned int x, unsigned int y) +{ + int i,j; + unsigned char *good; - if (req_comp == img_n) - return data; - STBI_ASSERT(req_comp >= 1 && req_comp <= 4); + if (req_comp == img_n) return data; + STBI_ASSERT(req_comp >= 1 && req_comp <= 4); - good = (unsigned char *)stbi__malloc_mad3(req_comp, x, y, 0); - if (good == NULL) { - STBI_FREE(data); - return stbi__errpuc("outofmem", "Out of memory"); - } + good = (unsigned char *) stbi__malloc_mad3(req_comp, x, y, 0); + if (good == NULL) { + STBI_FREE(data); + return stbi__errpuc("outofmem", "Out of memory"); + } - for (j = 0; j < (int)y; ++j) { - unsigned char * src = data + j * x * img_n; - unsigned char * dest = good + j * x * req_comp; + for (j=0; j < (int) y; ++j) { + unsigned char *src = data + j * x * img_n ; + unsigned char *dest = good + j * x * req_comp; -#define STBI__COMBO(a, b) ((a)*8 + (b)) -#define STBI__CASE(a, b) \ - case STBI__COMBO(a, b): \ - for (i = x - 1; i >= 0; --i, src += a, dest += b) - // convert source image with img_n components to one with req_comp components; - // avoid switch per pixel, so use switch per scanline and massive macros - switch (STBI__COMBO(img_n, req_comp)) { - STBI__CASE(1, 2) { - dest[0] = src[0]; - dest[1] = 255; - } - break; - STBI__CASE(1, 3) { dest[0] = dest[1] = dest[2] = src[0]; } - break; - STBI__CASE(1, 4) { - dest[0] = dest[1] = dest[2] = src[0]; - dest[3] = 255; - } - break; - STBI__CASE(2, 1) { dest[0] = src[0]; } - break; - STBI__CASE(2, 3) { dest[0] = dest[1] = dest[2] = src[0]; } - break; - STBI__CASE(2, 4) { - dest[0] = dest[1] = dest[2] = src[0]; - dest[3] = src[1]; - } - break; - STBI__CASE(3, 4) { - dest[0] = src[0]; - dest[1] = src[1]; - dest[2] = src[2]; - dest[3] = 255; - } - break; - STBI__CASE(3, 1) { dest[0] = stbi__compute_y(src[0], src[1], src[2]); } - break; - STBI__CASE(3, 2) { - dest[0] = stbi__compute_y(src[0], src[1], src[2]); - dest[1] = 255; - } - break; - STBI__CASE(4, 1) { dest[0] = stbi__compute_y(src[0], src[1], src[2]); } - break; - STBI__CASE(4, 2) { - dest[0] = stbi__compute_y(src[0], src[1], src[2]); - dest[1] = src[3]; - } - break; - STBI__CASE(4, 3) { - dest[0] = src[0]; - dest[1] = src[1]; - dest[2] = src[2]; - } - break; - default: - STBI_ASSERT(0); - STBI_FREE(data); - STBI_FREE(good); - return stbi__errpuc("unsupported", "Unsupported format conversion"); - } -#undef STBI__CASE - } + #define STBI__COMBO(a,b) ((a)*8+(b)) + #define STBI__CASE(a,b) case STBI__COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b) + // convert source image with img_n components to one with req_comp components; + // avoid switch per pixel, so use switch per scanline and massive macros + switch (STBI__COMBO(img_n, req_comp)) { + STBI__CASE(1,2) { dest[0]=src[0]; dest[1]=255; } break; + STBI__CASE(1,3) { dest[0]=dest[1]=dest[2]=src[0]; } break; + STBI__CASE(1,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=255; } break; + STBI__CASE(2,1) { dest[0]=src[0]; } break; + STBI__CASE(2,3) { dest[0]=dest[1]=dest[2]=src[0]; } break; + STBI__CASE(2,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=src[1]; } break; + STBI__CASE(3,4) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];dest[3]=255; } break; + STBI__CASE(3,1) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); } break; + STBI__CASE(3,2) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); dest[1] = 255; } break; + STBI__CASE(4,1) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); } break; + STBI__CASE(4,2) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); dest[1] = src[3]; } break; + STBI__CASE(4,3) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2]; } break; + default: STBI_ASSERT(0); STBI_FREE(data); STBI_FREE(good); return stbi__errpuc("unsupported", "Unsupported format conversion"); + } + #undef STBI__CASE + } - STBI_FREE(data); - return good; + STBI_FREE(data); + return good; } #endif #if defined(STBI_NO_PNG) && defined(STBI_NO_PSD) // nothing #else -static stbi__uint16 stbi__compute_y_16(int r, int g, int b) { return (stbi__uint16)(((r * 77) + (g * 150) + (29 * b)) >> 8); } +static stbi__uint16 stbi__compute_y_16(int r, int g, int b) +{ + return (stbi__uint16) (((r*77) + (g*150) + (29*b)) >> 8); +} #endif #if defined(STBI_NO_PNG) && defined(STBI_NO_PSD) // nothing #else -static stbi__uint16 * stbi__convert_format16(stbi__uint16 * data, int img_n, int req_comp, unsigned int x, unsigned int y) { - int i, j; - stbi__uint16 * good; +static stbi__uint16 *stbi__convert_format16(stbi__uint16 *data, int img_n, int req_comp, unsigned int x, unsigned int y) +{ + int i,j; + stbi__uint16 *good; - if (req_comp == img_n) - return data; - STBI_ASSERT(req_comp >= 1 && req_comp <= 4); + if (req_comp == img_n) return data; + STBI_ASSERT(req_comp >= 1 && req_comp <= 4); - good = (stbi__uint16 *)stbi__malloc(req_comp * x * y * 2); - if (good == NULL) { - STBI_FREE(data); - return (stbi__uint16 *)stbi__errpuc("outofmem", "Out of memory"); - } + good = (stbi__uint16 *) stbi__malloc(req_comp * x * y * 2); + if (good == NULL) { + STBI_FREE(data); + return (stbi__uint16 *) stbi__errpuc("outofmem", "Out of memory"); + } - for (j = 0; j < (int)y; ++j) { - stbi__uint16 * src = data + j * x * img_n; - stbi__uint16 * dest = good + j * x * req_comp; + for (j=0; j < (int) y; ++j) { + stbi__uint16 *src = data + j * x * img_n ; + stbi__uint16 *dest = good + j * x * req_comp; -#define STBI__COMBO(a, b) ((a)*8 + (b)) -#define STBI__CASE(a, b) \ - case STBI__COMBO(a, b): \ - for (i = x - 1; i >= 0; --i, src += a, dest += b) - // convert source image with img_n components to one with req_comp components; - // avoid switch per pixel, so use switch per scanline and massive macros - switch (STBI__COMBO(img_n, req_comp)) { - STBI__CASE(1, 2) { - dest[0] = src[0]; - dest[1] = 0xffff; - } - break; - STBI__CASE(1, 3) { dest[0] = dest[1] = dest[2] = src[0]; } - break; - STBI__CASE(1, 4) { - dest[0] = dest[1] = dest[2] = src[0]; - dest[3] = 0xffff; - } - break; - STBI__CASE(2, 1) { dest[0] = src[0]; } - break; - STBI__CASE(2, 3) { dest[0] = dest[1] = dest[2] = src[0]; } - break; - STBI__CASE(2, 4) { - dest[0] = dest[1] = dest[2] = src[0]; - dest[3] = src[1]; - } - break; - STBI__CASE(3, 4) { - dest[0] = src[0]; - dest[1] = src[1]; - dest[2] = src[2]; - dest[3] = 0xffff; - } - break; - STBI__CASE(3, 1) { dest[0] = stbi__compute_y_16(src[0], src[1], src[2]); } - break; - STBI__CASE(3, 2) { - dest[0] = stbi__compute_y_16(src[0], src[1], src[2]); - dest[1] = 0xffff; - } - break; - STBI__CASE(4, 1) { dest[0] = stbi__compute_y_16(src[0], src[1], src[2]); } - break; - STBI__CASE(4, 2) { - dest[0] = stbi__compute_y_16(src[0], src[1], src[2]); - dest[1] = src[3]; - } - break; - STBI__CASE(4, 3) { - dest[0] = src[0]; - dest[1] = src[1]; - dest[2] = src[2]; - } - break; - default: - STBI_ASSERT(0); - STBI_FREE(data); - STBI_FREE(good); - return (stbi__uint16 *)stbi__errpuc("unsupported", "Unsupported format conversion"); - } -#undef STBI__CASE - } + #define STBI__COMBO(a,b) ((a)*8+(b)) + #define STBI__CASE(a,b) case STBI__COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b) + // convert source image with img_n components to one with req_comp components; + // avoid switch per pixel, so use switch per scanline and massive macros + switch (STBI__COMBO(img_n, req_comp)) { + STBI__CASE(1,2) { dest[0]=src[0]; dest[1]=0xffff; } break; + STBI__CASE(1,3) { dest[0]=dest[1]=dest[2]=src[0]; } break; + STBI__CASE(1,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=0xffff; } break; + STBI__CASE(2,1) { dest[0]=src[0]; } break; + STBI__CASE(2,3) { dest[0]=dest[1]=dest[2]=src[0]; } break; + STBI__CASE(2,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=src[1]; } break; + STBI__CASE(3,4) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];dest[3]=0xffff; } break; + STBI__CASE(3,1) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); } break; + STBI__CASE(3,2) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); dest[1] = 0xffff; } break; + STBI__CASE(4,1) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); } break; + STBI__CASE(4,2) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); dest[1] = src[3]; } break; + STBI__CASE(4,3) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2]; } break; + default: STBI_ASSERT(0); STBI_FREE(data); STBI_FREE(good); return (stbi__uint16*) stbi__errpuc("unsupported", "Unsupported format conversion"); + } + #undef STBI__CASE + } - STBI_FREE(data); - return good; + STBI_FREE(data); + return good; } #endif #ifndef STBI_NO_LINEAR -static float * stbi__ldr_to_hdr(stbi_uc * data, int x, int y, int comp) { - int i, k, n; - float * output; - if (!data) - return NULL; - output = (float *)stbi__malloc_mad4(x, y, comp, sizeof(float), 0); - if (output == NULL) { - STBI_FREE(data); - return stbi__errpf("outofmem", "Out of memory"); - } - // compute number of non-alpha components - if (comp & 1) - n = comp; - else - n = comp - 1; - for (i = 0; i < x * y; ++i) { - for (k = 0; k < n; ++k) { - output[i * comp + k] = (float)(pow(data[i * comp + k] / 255.0f, stbi__l2h_gamma) * stbi__l2h_scale); - } - } - if (n < comp) { - for (i = 0; i < x * y; ++i) { - output[i * comp + n] = data[i * comp + n] / 255.0f; - } - } - STBI_FREE(data); - return output; +static float *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp) +{ + int i,k,n; + float *output; + if (!data) return NULL; + output = (float *) stbi__malloc_mad4(x, y, comp, sizeof(float), 0); + if (output == NULL) { STBI_FREE(data); return stbi__errpf("outofmem", "Out of memory"); } + // compute number of non-alpha components + if (comp & 1) n = comp; else n = comp-1; + for (i=0; i < x*y; ++i) { + for (k=0; k < n; ++k) { + output[i*comp + k] = (float) (pow(data[i*comp+k]/255.0f, stbi__l2h_gamma) * stbi__l2h_scale); + } + } + if (n < comp) { + for (i=0; i < x*y; ++i) { + output[i*comp + n] = data[i*comp + n]/255.0f; + } + } + STBI_FREE(data); + return output; } #endif #ifndef STBI_NO_HDR -#define stbi__float2int(x) ((int)(x)) -static stbi_uc * stbi__hdr_to_ldr(float * data, int x, int y, int comp) { - int i, k, n; - stbi_uc * output; - if (!data) - return NULL; - output = (stbi_uc *)stbi__malloc_mad3(x, y, comp, 0); - if (output == NULL) { - STBI_FREE(data); - return stbi__errpuc("outofmem", "Out of memory"); - } - // compute number of non-alpha components - if (comp & 1) - n = comp; - else - n = comp - 1; - for (i = 0; i < x * y; ++i) { - for (k = 0; k < n; ++k) { - float z = (float)pow(data[i * comp + k] * stbi__h2l_scale_i, stbi__h2l_gamma_i) * 255 + 0.5f; - if (z < 0) - z = 0; - if (z > 255) - z = 255; - output[i * comp + k] = (stbi_uc)stbi__float2int(z); - } - if (k < comp) { - float z = data[i * comp + k] * 255 + 0.5f; - if (z < 0) - z = 0; - if (z > 255) - z = 255; - output[i * comp + k] = (stbi_uc)stbi__float2int(z); - } - } - STBI_FREE(data); - return output; +#define stbi__float2int(x) ((int) (x)) +static stbi_uc *stbi__hdr_to_ldr(float *data, int x, int y, int comp) +{ + int i,k,n; + stbi_uc *output; + if (!data) return NULL; + output = (stbi_uc *) stbi__malloc_mad3(x, y, comp, 0); + if (output == NULL) { STBI_FREE(data); return stbi__errpuc("outofmem", "Out of memory"); } + // compute number of non-alpha components + if (comp & 1) n = comp; else n = comp-1; + for (i=0; i < x*y; ++i) { + for (k=0; k < n; ++k) { + float z = (float) pow(data[i*comp+k]*stbi__h2l_scale_i, stbi__h2l_gamma_i) * 255 + 0.5f; + if (z < 0) z = 0; + if (z > 255) z = 255; + output[i*comp + k] = (stbi_uc) stbi__float2int(z); + } + if (k < comp) { + float z = data[i*comp+k] * 255 + 0.5f; + if (z < 0) z = 0; + if (z > 255) z = 255; + output[i*comp + k] = (stbi_uc) stbi__float2int(z); + } + } + STBI_FREE(data); + return output; } #endif @@ -1968,783 +1933,763 @@ static stbi_uc * stbi__hdr_to_ldr(float * data, int x, int y, int comp) { #ifndef STBI_NO_JPEG // huffman decoding acceleration -#define FAST_BITS 9 // larger handles more cases; smaller stomps less cache +#define FAST_BITS 9 // larger handles more cases; smaller stomps less cache -typedef struct { - stbi_uc fast[1 << FAST_BITS]; - // weirdly, repacking this into AoS is a 10% speed loss, instead of a win - stbi__uint16 code[256]; - stbi_uc values[256]; - stbi_uc size[257]; - unsigned int maxcode[18]; - int delta[17]; // old 'firstsymbol' - old 'firstcode' +typedef struct +{ + stbi_uc fast[1 << FAST_BITS]; + // weirdly, repacking this into AoS is a 10% speed loss, instead of a win + stbi__uint16 code[256]; + stbi_uc values[256]; + stbi_uc size[257]; + unsigned int maxcode[18]; + int delta[17]; // old 'firstsymbol' - old 'firstcode' } stbi__huffman; -typedef struct { - stbi__context * s; - stbi__huffman huff_dc[4]; - stbi__huffman huff_ac[4]; - stbi__uint16 dequant[4][64]; - stbi__int16 fast_ac[4][1 << FAST_BITS]; +typedef struct +{ + stbi__context *s; + stbi__huffman huff_dc[4]; + stbi__huffman huff_ac[4]; + stbi__uint16 dequant[4][64]; + stbi__int16 fast_ac[4][1 << FAST_BITS]; - // sizes for components, interleaved MCUs - int img_h_max, img_v_max; - int img_mcu_x, img_mcu_y; - int img_mcu_w, img_mcu_h; +// sizes for components, interleaved MCUs + int img_h_max, img_v_max; + int img_mcu_x, img_mcu_y; + int img_mcu_w, img_mcu_h; - // definition of jpeg image component - struct { - int id; - int h, v; - int tq; - int hd, ha; - int dc_pred; +// definition of jpeg image component + struct + { + int id; + int h,v; + int tq; + int hd,ha; + int dc_pred; - int x, y, w2, h2; - stbi_uc * data; - void *raw_data, *raw_coeff; - stbi_uc * linebuf; - short * coeff; // progressive only - int coeff_w, coeff_h; // number of 8x8 coefficient blocks - } img_comp[4]; + int x,y,w2,h2; + stbi_uc *data; + void *raw_data, *raw_coeff; + stbi_uc *linebuf; + short *coeff; // progressive only + int coeff_w, coeff_h; // number of 8x8 coefficient blocks + } img_comp[4]; - stbi__uint32 code_buffer; // jpeg entropy-coded buffer - int code_bits; // number of valid bits - unsigned char marker; // marker seen while filling entropy buffer - int nomore; // flag if we saw a marker so must stop + stbi__uint32 code_buffer; // jpeg entropy-coded buffer + int code_bits; // number of valid bits + unsigned char marker; // marker seen while filling entropy buffer + int nomore; // flag if we saw a marker so must stop - int progressive; - int spec_start; - int spec_end; - int succ_high; - int succ_low; - int eob_run; - int jfif; - int app14_color_transform; // Adobe APP14 tag - int rgb; + int progressive; + int spec_start; + int spec_end; + int succ_high; + int succ_low; + int eob_run; + int jfif; + int app14_color_transform; // Adobe APP14 tag + int rgb; - int scan_n, order[4]; - int restart_interval, todo; + int scan_n, order[4]; + int restart_interval, todo; - // kernels - void (*idct_block_kernel)(stbi_uc * out, int out_stride, short data[64]); - void (*YCbCr_to_RGB_kernel)(stbi_uc * out, const stbi_uc * y, const stbi_uc * pcb, const stbi_uc * pcr, int count, - int step); - stbi_uc * (*resample_row_hv_2_kernel)(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs); +// kernels + void (*idct_block_kernel)(stbi_uc *out, int out_stride, short data[64]); + void (*YCbCr_to_RGB_kernel)(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step); + stbi_uc *(*resample_row_hv_2_kernel)(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs); } stbi__jpeg; -static int stbi__build_huffman(stbi__huffman * h, int * count) { - int i, j, k = 0; - unsigned int code; - // build size list for each symbol (from JPEG spec) - for (i = 0; i < 16; ++i) { - for (j = 0; j < count[i]; ++j) { - h->size[k++] = (stbi_uc)(i + 1); - if (k >= 257) - return stbi__err("bad size list", "Corrupt JPEG"); - } - } - h->size[k] = 0; +static int stbi__build_huffman(stbi__huffman *h, int *count) +{ + int i,j,k=0; + unsigned int code; + // build size list for each symbol (from JPEG spec) + for (i=0; i < 16; ++i) { + for (j=0; j < count[i]; ++j) { + h->size[k++] = (stbi_uc) (i+1); + if(k >= 257) return stbi__err("bad size list","Corrupt JPEG"); + } + } + h->size[k] = 0; - // compute actual symbols (from jpeg spec) - code = 0; - k = 0; - for (j = 1; j <= 16; ++j) { - // compute delta to add to code to compute symbol id - h->delta[j] = k - code; - if (h->size[k] == j) { - while (h->size[k] == j) - h->code[k++] = (stbi__uint16)(code++); - if (code - 1 >= (1u << j)) - return stbi__err("bad code lengths", "Corrupt JPEG"); - } - // compute largest code + 1 for this size, preshifted as needed later - h->maxcode[j] = code << (16 - j); - code <<= 1; - } - h->maxcode[j] = 0xffffffff; + // compute actual symbols (from jpeg spec) + code = 0; + k = 0; + for(j=1; j <= 16; ++j) { + // compute delta to add to code to compute symbol id + h->delta[j] = k - code; + if (h->size[k] == j) { + while (h->size[k] == j) + h->code[k++] = (stbi__uint16) (code++); + if (code-1 >= (1u << j)) return stbi__err("bad code lengths","Corrupt JPEG"); + } + // compute largest code + 1 for this size, preshifted as needed later + h->maxcode[j] = code << (16-j); + code <<= 1; + } + h->maxcode[j] = 0xffffffff; - // build non-spec acceleration table; 255 is flag for not-accelerated - memset(h->fast, 255, 1 << FAST_BITS); - for (i = 0; i < k; ++i) { - int s = h->size[i]; - if (s <= FAST_BITS) { - int c = h->code[i] << (FAST_BITS - s); - int m = 1 << (FAST_BITS - s); - for (j = 0; j < m; ++j) { - h->fast[c + j] = (stbi_uc)i; - } - } - } - return 1; + // build non-spec acceleration table; 255 is flag for not-accelerated + memset(h->fast, 255, 1 << FAST_BITS); + for (i=0; i < k; ++i) { + int s = h->size[i]; + if (s <= FAST_BITS) { + int c = h->code[i] << (FAST_BITS-s); + int m = 1 << (FAST_BITS-s); + for (j=0; j < m; ++j) { + h->fast[c+j] = (stbi_uc) i; + } + } + } + return 1; } // build a table that decodes both magnitude and value of small ACs in // one go. -static void stbi__build_fast_ac(stbi__int16 * fast_ac, stbi__huffman * h) { - int i; - for (i = 0; i < (1 << FAST_BITS); ++i) { - stbi_uc fast = h->fast[i]; - fast_ac[i] = 0; - if (fast < 255) { - int rs = h->values[fast]; - int run = (rs >> 4) & 15; - int magbits = rs & 15; - int len = h->size[fast]; +static void stbi__build_fast_ac(stbi__int16 *fast_ac, stbi__huffman *h) +{ + int i; + for (i=0; i < (1 << FAST_BITS); ++i) { + stbi_uc fast = h->fast[i]; + fast_ac[i] = 0; + if (fast < 255) { + int rs = h->values[fast]; + int run = (rs >> 4) & 15; + int magbits = rs & 15; + int len = h->size[fast]; - if (magbits && len + magbits <= FAST_BITS) { - // magnitude code followed by receive_extend code - int k = ((i << len) & ((1 << FAST_BITS) - 1)) >> (FAST_BITS - magbits); - int m = 1 << (magbits - 1); - if (k < m) - k += (~0U << magbits) + 1; - // if the result is small enough, we can fit it in fast_ac table - if (k >= -128 && k <= 127) - fast_ac[i] = (stbi__int16)((k * 256) + (run * 16) + (len + magbits)); - } - } - } + if (magbits && len + magbits <= FAST_BITS) { + // magnitude code followed by receive_extend code + int k = ((i << len) & ((1 << FAST_BITS) - 1)) >> (FAST_BITS - magbits); + int m = 1 << (magbits - 1); + if (k < m) k += (~0U << magbits) + 1; + // if the result is small enough, we can fit it in fast_ac table + if (k >= -128 && k <= 127) + fast_ac[i] = (stbi__int16) ((k * 256) + (run * 16) + (len + magbits)); + } + } + } } -static void stbi__grow_buffer_unsafe(stbi__jpeg * j) { - do { - unsigned int b = j->nomore ? 0 : stbi__get8(j->s); - if (b == 0xff) { - int c = stbi__get8(j->s); - while (c == 0xff) - c = stbi__get8(j->s); // consume fill bytes - if (c != 0) { - j->marker = (unsigned char)c; - j->nomore = 1; - return; - } - } - j->code_buffer |= b << (24 - j->code_bits); - j->code_bits += 8; - } while (j->code_bits <= 24); +static void stbi__grow_buffer_unsafe(stbi__jpeg *j) +{ + do { + unsigned int b = j->nomore ? 0 : stbi__get8(j->s); + if (b == 0xff) { + int c = stbi__get8(j->s); + while (c == 0xff) c = stbi__get8(j->s); // consume fill bytes + if (c != 0) { + j->marker = (unsigned char) c; + j->nomore = 1; + return; + } + } + j->code_buffer |= b << (24 - j->code_bits); + j->code_bits += 8; + } while (j->code_bits <= 24); } // (1 << n) - 1 -static const stbi__uint32 stbi__bmask[17] = {0, 1, 3, 7, 15, 31, 63, 127, 255, - 511, 1023, 2047, 4095, 8191, 16383, 32767, 65535}; +static const stbi__uint32 stbi__bmask[17]={0,1,3,7,15,31,63,127,255,511,1023,2047,4095,8191,16383,32767,65535}; // decode a jpeg huffman value from the bitstream -stbi_inline static int stbi__jpeg_huff_decode(stbi__jpeg * j, stbi__huffman * h) { - unsigned int temp; - int c, k; +stbi_inline static int stbi__jpeg_huff_decode(stbi__jpeg *j, stbi__huffman *h) +{ + unsigned int temp; + int c,k; - if (j->code_bits < 16) - stbi__grow_buffer_unsafe(j); + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); - // look at the top FAST_BITS and determine what symbol ID it is, - // if the code is <= FAST_BITS - c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS) - 1); - k = h->fast[c]; - if (k < 255) { - int s = h->size[k]; - if (s > j->code_bits) - return -1; - j->code_buffer <<= s; - j->code_bits -= s; - return h->values[k]; - } + // look at the top FAST_BITS and determine what symbol ID it is, + // if the code is <= FAST_BITS + c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1); + k = h->fast[c]; + if (k < 255) { + int s = h->size[k]; + if (s > j->code_bits) + return -1; + j->code_buffer <<= s; + j->code_bits -= s; + return h->values[k]; + } - // naive test is to shift the code_buffer down so k bits are - // valid, then test against maxcode. To speed this up, we've - // preshifted maxcode left so that it has (16-k) 0s at the - // end; in other words, regardless of the number of bits, it - // wants to be compared against something shifted to have 16; - // that way we don't need to shift inside the loop. - temp = j->code_buffer >> 16; - for (k = FAST_BITS + 1;; ++k) - if (temp < h->maxcode[k]) - break; - if (k == 17) { - // error! code not found - j->code_bits -= 16; - return -1; - } + // naive test is to shift the code_buffer down so k bits are + // valid, then test against maxcode. To speed this up, we've + // preshifted maxcode left so that it has (16-k) 0s at the + // end; in other words, regardless of the number of bits, it + // wants to be compared against something shifted to have 16; + // that way we don't need to shift inside the loop. + temp = j->code_buffer >> 16; + for (k=FAST_BITS+1 ; ; ++k) + if (temp < h->maxcode[k]) + break; + if (k == 17) { + // error! code not found + j->code_bits -= 16; + return -1; + } - if (k > j->code_bits) - return -1; + if (k > j->code_bits) + return -1; - // convert the huffman code to the symbol id - c = ((j->code_buffer >> (32 - k)) & stbi__bmask[k]) + h->delta[k]; - if (c < 0 || c >= 256) // symbol id out of bounds! - return -1; - STBI_ASSERT((((j->code_buffer) >> (32 - h->size[c])) & stbi__bmask[h->size[c]]) == h->code[c]); + // convert the huffman code to the symbol id + c = ((j->code_buffer >> (32 - k)) & stbi__bmask[k]) + h->delta[k]; + if(c < 0 || c >= 256) // symbol id out of bounds! + return -1; + STBI_ASSERT((((j->code_buffer) >> (32 - h->size[c])) & stbi__bmask[h->size[c]]) == h->code[c]); - // convert the id to a symbol - j->code_bits -= k; - j->code_buffer <<= k; - return h->values[c]; + // convert the id to a symbol + j->code_bits -= k; + j->code_buffer <<= k; + return h->values[c]; } // bias[n] = (-1<code_bits < n) - stbi__grow_buffer_unsafe(j); - if (j->code_bits < n) - return 0; // ran out of bits from stream, return 0s intead of continuing +stbi_inline static int stbi__extend_receive(stbi__jpeg *j, int n) +{ + unsigned int k; + int sgn; + if (j->code_bits < n) stbi__grow_buffer_unsafe(j); + if (j->code_bits < n) return 0; // ran out of bits from stream, return 0s intead of continuing - sgn = j->code_buffer >> 31; // sign bit always in MSB; 0 if MSB clear (positive), 1 if MSB set (negative) - k = stbi_lrot(j->code_buffer, n); - j->code_buffer = k & ~stbi__bmask[n]; - k &= stbi__bmask[n]; - j->code_bits -= n; - return k + (stbi__jbias[n] & (sgn - 1)); + sgn = j->code_buffer >> 31; // sign bit always in MSB; 0 if MSB clear (positive), 1 if MSB set (negative) + k = stbi_lrot(j->code_buffer, n); + j->code_buffer = k & ~stbi__bmask[n]; + k &= stbi__bmask[n]; + j->code_bits -= n; + return k + (stbi__jbias[n] & (sgn - 1)); } // get some unsigned bits -stbi_inline static int stbi__jpeg_get_bits(stbi__jpeg * j, int n) { - unsigned int k; - if (j->code_bits < n) - stbi__grow_buffer_unsafe(j); - if (j->code_bits < n) - return 0; // ran out of bits from stream, return 0s intead of continuing - k = stbi_lrot(j->code_buffer, n); - j->code_buffer = k & ~stbi__bmask[n]; - k &= stbi__bmask[n]; - j->code_bits -= n; - return k; +stbi_inline static int stbi__jpeg_get_bits(stbi__jpeg *j, int n) +{ + unsigned int k; + if (j->code_bits < n) stbi__grow_buffer_unsafe(j); + if (j->code_bits < n) return 0; // ran out of bits from stream, return 0s intead of continuing + k = stbi_lrot(j->code_buffer, n); + j->code_buffer = k & ~stbi__bmask[n]; + k &= stbi__bmask[n]; + j->code_bits -= n; + return k; } -stbi_inline static int stbi__jpeg_get_bit(stbi__jpeg * j) { - unsigned int k; - if (j->code_bits < 1) - stbi__grow_buffer_unsafe(j); - if (j->code_bits < 1) - return 0; // ran out of bits from stream, return 0s intead of continuing - k = j->code_buffer; - j->code_buffer <<= 1; - --j->code_bits; - return k & 0x80000000; +stbi_inline static int stbi__jpeg_get_bit(stbi__jpeg *j) +{ + unsigned int k; + if (j->code_bits < 1) stbi__grow_buffer_unsafe(j); + if (j->code_bits < 1) return 0; // ran out of bits from stream, return 0s intead of continuing + k = j->code_buffer; + j->code_buffer <<= 1; + --j->code_bits; + return k & 0x80000000; } // given a value that's at position X in the zigzag stream, // where does it appear in the 8x8 matrix coded as row-major? -static const stbi_uc stbi__jpeg_dezigzag[64 + 15] = { - 0, 1, 8, 16, 9, 2, 3, 10, 17, 24, 32, 25, 18, 11, 4, 5, 12, 19, 26, 33, 40, 48, 41, 34, 27, 20, 13, 6, 7, 14, 21, 28, 35, - 42, 49, 56, 57, 50, 43, 36, 29, 22, 15, 23, 30, 37, 44, 51, 58, 59, 52, 45, 38, 31, 39, 46, 53, 60, 61, 54, 47, 55, 62, 63, - // let corrupt input sample past end - 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63}; +static const stbi_uc stbi__jpeg_dezigzag[64+15] = +{ + 0, 1, 8, 16, 9, 2, 3, 10, + 17, 24, 32, 25, 18, 11, 4, 5, + 12, 19, 26, 33, 40, 48, 41, 34, + 27, 20, 13, 6, 7, 14, 21, 28, + 35, 42, 49, 56, 57, 50, 43, 36, + 29, 22, 15, 23, 30, 37, 44, 51, + 58, 59, 52, 45, 38, 31, 39, 46, + 53, 60, 61, 54, 47, 55, 62, 63, + // let corrupt input sample past end + 63, 63, 63, 63, 63, 63, 63, 63, + 63, 63, 63, 63, 63, 63, 63 +}; // decode one 64-entry block-- -static int stbi__jpeg_decode_block(stbi__jpeg * j, short data[64], stbi__huffman * hdc, stbi__huffman * hac, stbi__int16 * fac, - int b, stbi__uint16 * dequant) { - int diff, dc, k; - int t; +static int stbi__jpeg_decode_block(stbi__jpeg *j, short data[64], stbi__huffman *hdc, stbi__huffman *hac, stbi__int16 *fac, int b, stbi__uint16 *dequant) +{ + int diff,dc,k; + int t; - if (j->code_bits < 16) - stbi__grow_buffer_unsafe(j); - t = stbi__jpeg_huff_decode(j, hdc); - if (t < 0 || t > 15) - return stbi__err("bad huffman code", "Corrupt JPEG"); + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); + t = stbi__jpeg_huff_decode(j, hdc); + if (t < 0 || t > 15) return stbi__err("bad huffman code","Corrupt JPEG"); - // 0 all the ac values now so we can do it 32-bits at a time - memset(data, 0, 64 * sizeof(data[0])); + // 0 all the ac values now so we can do it 32-bits at a time + memset(data,0,64*sizeof(data[0])); - diff = t ? stbi__extend_receive(j, t) : 0; - if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) - return stbi__err("bad delta", "Corrupt JPEG"); - dc = j->img_comp[b].dc_pred + diff; - j->img_comp[b].dc_pred = dc; - if (!stbi__mul2shorts_valid(dc, dequant[0])) - return stbi__err("can't merge dc and ac", "Corrupt JPEG"); - data[0] = (short)(dc * dequant[0]); + diff = t ? stbi__extend_receive(j, t) : 0; + if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) return stbi__err("bad delta","Corrupt JPEG"); + dc = j->img_comp[b].dc_pred + diff; + j->img_comp[b].dc_pred = dc; + if (!stbi__mul2shorts_valid(dc, dequant[0])) return stbi__err("can't merge dc and ac", "Corrupt JPEG"); + data[0] = (short) (dc * dequant[0]); - // decode AC components, see JPEG spec - k = 1; - do { - unsigned int zig; - int c, r, s; - if (j->code_bits < 16) - stbi__grow_buffer_unsafe(j); - c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS) - 1); - r = fac[c]; - if (r) { // fast-AC path - k += (r >> 4) & 15; // run - s = r & 15; // combined length - if (s > j->code_bits) - return stbi__err("bad huffman code", "Combined length longer than code bits available"); - j->code_buffer <<= s; - j->code_bits -= s; + // decode AC components, see JPEG spec + k = 1; + do { + unsigned int zig; + int c,r,s; + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); + c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1); + r = fac[c]; + if (r) { // fast-AC path + k += (r >> 4) & 15; // run + s = r & 15; // combined length + if (s > j->code_bits) return stbi__err("bad huffman code", "Combined length longer than code bits available"); + j->code_buffer <<= s; + j->code_bits -= s; + // decode into unzigzag'd location + zig = stbi__jpeg_dezigzag[k++]; + data[zig] = (short) ((r >> 8) * dequant[zig]); + } else { + int rs = stbi__jpeg_huff_decode(j, hac); + if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG"); + s = rs & 15; + r = rs >> 4; + if (s == 0) { + if (rs != 0xf0) break; // end block + k += 16; + } else { + k += r; // decode into unzigzag'd location zig = stbi__jpeg_dezigzag[k++]; - data[zig] = (short)((r >> 8) * dequant[zig]); - } else { - int rs = stbi__jpeg_huff_decode(j, hac); - if (rs < 0) - return stbi__err("bad huffman code", "Corrupt JPEG"); - s = rs & 15; - r = rs >> 4; - if (s == 0) { - if (rs != 0xf0) - break; // end block - k += 16; - } else { - k += r; - // decode into unzigzag'd location - zig = stbi__jpeg_dezigzag[k++]; - data[zig] = (short)(stbi__extend_receive(j, s) * dequant[zig]); - } - } - } while (k < 64); - return 1; + data[zig] = (short) (stbi__extend_receive(j,s) * dequant[zig]); + } + } + } while (k < 64); + return 1; } -static int stbi__jpeg_decode_block_prog_dc(stbi__jpeg * j, short data[64], stbi__huffman * hdc, int b) { - int diff, dc; - int t; - if (j->spec_end != 0) - return stbi__err("can't merge dc and ac", "Corrupt JPEG"); +static int stbi__jpeg_decode_block_prog_dc(stbi__jpeg *j, short data[64], stbi__huffman *hdc, int b) +{ + int diff,dc; + int t; + if (j->spec_end != 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG"); - if (j->code_bits < 16) - stbi__grow_buffer_unsafe(j); + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); - if (j->succ_high == 0) { - // first scan for DC coefficient, must be first - memset(data, 0, 64 * sizeof(data[0])); // 0 all the ac values now - t = stbi__jpeg_huff_decode(j, hdc); - if (t < 0 || t > 15) - return stbi__err("can't merge dc and ac", "Corrupt JPEG"); - diff = t ? stbi__extend_receive(j, t) : 0; + if (j->succ_high == 0) { + // first scan for DC coefficient, must be first + memset(data,0,64*sizeof(data[0])); // 0 all the ac values now + t = stbi__jpeg_huff_decode(j, hdc); + if (t < 0 || t > 15) return stbi__err("can't merge dc and ac", "Corrupt JPEG"); + diff = t ? stbi__extend_receive(j, t) : 0; - if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) - return stbi__err("bad delta", "Corrupt JPEG"); - dc = j->img_comp[b].dc_pred + diff; - j->img_comp[b].dc_pred = dc; - if (!stbi__mul2shorts_valid(dc, 1 << j->succ_low)) - return stbi__err("can't merge dc and ac", "Corrupt JPEG"); - data[0] = (short)(dc * (1 << j->succ_low)); - } else { - // refinement scan for DC coefficient - if (stbi__jpeg_get_bit(j)) - data[0] += (short)(1 << j->succ_low); - } - return 1; + if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) return stbi__err("bad delta", "Corrupt JPEG"); + dc = j->img_comp[b].dc_pred + diff; + j->img_comp[b].dc_pred = dc; + if (!stbi__mul2shorts_valid(dc, 1 << j->succ_low)) return stbi__err("can't merge dc and ac", "Corrupt JPEG"); + data[0] = (short) (dc * (1 << j->succ_low)); + } else { + // refinement scan for DC coefficient + if (stbi__jpeg_get_bit(j)) + data[0] += (short) (1 << j->succ_low); + } + return 1; } // @OPTIMIZE: store non-zigzagged during the decode passes, // and only de-zigzag when dequantizing -static int stbi__jpeg_decode_block_prog_ac(stbi__jpeg * j, short data[64], stbi__huffman * hac, stbi__int16 * fac) { - int k; - if (j->spec_start == 0) - return stbi__err("can't merge dc and ac", "Corrupt JPEG"); +static int stbi__jpeg_decode_block_prog_ac(stbi__jpeg *j, short data[64], stbi__huffman *hac, stbi__int16 *fac) +{ + int k; + if (j->spec_start == 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG"); - if (j->succ_high == 0) { - int shift = j->succ_low; + if (j->succ_high == 0) { + int shift = j->succ_low; - if (j->eob_run) { - --j->eob_run; - return 1; - } + if (j->eob_run) { + --j->eob_run; + return 1; + } - k = j->spec_start; - do { - unsigned int zig; - int c, r, s; - if (j->code_bits < 16) - stbi__grow_buffer_unsafe(j); - c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS) - 1); - r = fac[c]; - if (r) { // fast-AC path - k += (r >> 4) & 15; // run - s = r & 15; // combined length - if (s > j->code_bits) - return stbi__err("bad huffman code", "Combined length longer than code bits available"); - j->code_buffer <<= s; - j->code_bits -= s; - zig = stbi__jpeg_dezigzag[k++]; - data[zig] = (short)((r >> 8) * (1 << shift)); + k = j->spec_start; + do { + unsigned int zig; + int c,r,s; + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); + c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1); + r = fac[c]; + if (r) { // fast-AC path + k += (r >> 4) & 15; // run + s = r & 15; // combined length + if (s > j->code_bits) return stbi__err("bad huffman code", "Combined length longer than code bits available"); + j->code_buffer <<= s; + j->code_bits -= s; + zig = stbi__jpeg_dezigzag[k++]; + data[zig] = (short) ((r >> 8) * (1 << shift)); + } else { + int rs = stbi__jpeg_huff_decode(j, hac); + if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG"); + s = rs & 15; + r = rs >> 4; + if (s == 0) { + if (r < 15) { + j->eob_run = (1 << r); + if (r) + j->eob_run += stbi__jpeg_get_bits(j, r); + --j->eob_run; + break; + } + k += 16; } else { - int rs = stbi__jpeg_huff_decode(j, hac); - if (rs < 0) - return stbi__err("bad huffman code", "Corrupt JPEG"); - s = rs & 15; - r = rs >> 4; - if (s == 0) { - if (r < 15) { - j->eob_run = (1 << r); - if (r) - j->eob_run += stbi__jpeg_get_bits(j, r); - --j->eob_run; - break; - } - k += 16; - } else { - k += r; - zig = stbi__jpeg_dezigzag[k++]; - data[zig] = (short)(stbi__extend_receive(j, s) * (1 << shift)); - } + k += r; + zig = stbi__jpeg_dezigzag[k++]; + data[zig] = (short) (stbi__extend_receive(j,s) * (1 << shift)); } - } while (k <= j->spec_end); - } else { - // refinement scan for these AC coefficients + } + } while (k <= j->spec_end); + } else { + // refinement scan for these AC coefficients - short bit = (short)(1 << j->succ_low); + short bit = (short) (1 << j->succ_low); - if (j->eob_run) { - --j->eob_run; - for (k = j->spec_start; k <= j->spec_end; ++k) { - short * p = &data[stbi__jpeg_dezigzag[k]]; - if (*p != 0) - if (stbi__jpeg_get_bit(j)) - if ((*p & bit) == 0) { - if (*p > 0) - *p += bit; - else - *p -= bit; - } + if (j->eob_run) { + --j->eob_run; + for (k = j->spec_start; k <= j->spec_end; ++k) { + short *p = &data[stbi__jpeg_dezigzag[k]]; + if (*p != 0) + if (stbi__jpeg_get_bit(j)) + if ((*p & bit)==0) { + if (*p > 0) + *p += bit; + else + *p -= bit; + } + } + } else { + k = j->spec_start; + do { + int r,s; + int rs = stbi__jpeg_huff_decode(j, hac); // @OPTIMIZE see if we can use the fast path here, advance-by-r is so slow, eh + if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG"); + s = rs & 15; + r = rs >> 4; + if (s == 0) { + if (r < 15) { + j->eob_run = (1 << r) - 1; + if (r) + j->eob_run += stbi__jpeg_get_bits(j, r); + r = 64; // force end of block + } else { + // r=15 s=0 should write 16 0s, so we just do + // a run of 15 0s and then write s (which is 0), + // so we don't have to do anything special here + } + } else { + if (s != 1) return stbi__err("bad huffman code", "Corrupt JPEG"); + // sign bit + if (stbi__jpeg_get_bit(j)) + s = bit; + else + s = -bit; } - } else { - k = j->spec_start; - do { - int r, s; - int rs = stbi__jpeg_huff_decode( - j, hac); // @OPTIMIZE see if we can use the fast path here, advance-by-r is so slow, eh - if (rs < 0) - return stbi__err("bad huffman code", "Corrupt JPEG"); - s = rs & 15; - r = rs >> 4; - if (s == 0) { - if (r < 15) { - j->eob_run = (1 << r) - 1; - if (r) - j->eob_run += stbi__jpeg_get_bits(j, r); - r = 64; // force end of block - } else { - // r=15 s=0 should write 16 0s, so we just do - // a run of 15 0s and then write s (which is 0), - // so we don't have to do anything special here - } - } else { - if (s != 1) - return stbi__err("bad huffman code", "Corrupt JPEG"); - // sign bit - if (stbi__jpeg_get_bit(j)) - s = bit; - else - s = -bit; - } - // advance by r - while (k <= j->spec_end) { - short * p = &data[stbi__jpeg_dezigzag[k++]]; - if (*p != 0) { - if (stbi__jpeg_get_bit(j)) - if ((*p & bit) == 0) { - if (*p > 0) - *p += bit; - else - *p -= bit; - } - } else { - if (r == 0) { - *p = (short)s; - break; - } - --r; - } - } - } while (k <= j->spec_end); - } - } - return 1; + // advance by r + while (k <= j->spec_end) { + short *p = &data[stbi__jpeg_dezigzag[k++]]; + if (*p != 0) { + if (stbi__jpeg_get_bit(j)) + if ((*p & bit)==0) { + if (*p > 0) + *p += bit; + else + *p -= bit; + } + } else { + if (r == 0) { + *p = (short) s; + break; + } + --r; + } + } + } while (k <= j->spec_end); + } + } + return 1; } // take a -128..127 value and stbi__clamp it and convert to 0..255 -stbi_inline static stbi_uc stbi__clamp(int x) { - // trick to use a single test to catch both cases - if ((unsigned int)x > 255) { - if (x < 0) - return 0; - if (x > 255) - return 255; - } - return (stbi_uc)x; +stbi_inline static stbi_uc stbi__clamp(int x) +{ + // trick to use a single test to catch both cases + if ((unsigned int) x > 255) { + if (x < 0) return 0; + if (x > 255) return 255; + } + return (stbi_uc) x; } -#define stbi__f2f(x) ((int)(((x)*4096 + 0.5))) -#define stbi__fsh(x) ((x)*4096) +#define stbi__f2f(x) ((int) (((x) * 4096 + 0.5))) +#define stbi__fsh(x) ((x) * 4096) // derived from jidctint -- DCT_ISLOW -#define STBI__IDCT_1D(s0, s1, s2, s3, s4, s5, s6, s7) \ - int t0, t1, t2, t3, p1, p2, p3, p4, p5, x0, x1, x2, x3; \ - p2 = s2; \ - p3 = s6; \ - p1 = (p2 + p3) * stbi__f2f(0.5411961f); \ - t2 = p1 + p3 * stbi__f2f(-1.847759065f); \ - t3 = p1 + p2 * stbi__f2f(0.765366865f); \ - p2 = s0; \ - p3 = s4; \ - t0 = stbi__fsh(p2 + p3); \ - t1 = stbi__fsh(p2 - p3); \ - x0 = t0 + t3; \ - x3 = t0 - t3; \ - x1 = t1 + t2; \ - x2 = t1 - t2; \ - t0 = s7; \ - t1 = s5; \ - t2 = s3; \ - t3 = s1; \ - p3 = t0 + t2; \ - p4 = t1 + t3; \ - p1 = t0 + t3; \ - p2 = t1 + t2; \ - p5 = (p3 + p4) * stbi__f2f(1.175875602f); \ - t0 = t0 * stbi__f2f(0.298631336f); \ - t1 = t1 * stbi__f2f(2.053119869f); \ - t2 = t2 * stbi__f2f(3.072711026f); \ - t3 = t3 * stbi__f2f(1.501321110f); \ - p1 = p5 + p1 * stbi__f2f(-0.899976223f); \ - p2 = p5 + p2 * stbi__f2f(-2.562915447f); \ - p3 = p3 * stbi__f2f(-1.961570560f); \ - p4 = p4 * stbi__f2f(-0.390180644f); \ - t3 += p1 + p4; \ - t2 += p2 + p3; \ - t1 += p2 + p4; \ - t0 += p1 + p3; +#define STBI__IDCT_1D(s0,s1,s2,s3,s4,s5,s6,s7) \ + int t0,t1,t2,t3,p1,p2,p3,p4,p5,x0,x1,x2,x3; \ + p2 = s2; \ + p3 = s6; \ + p1 = (p2+p3) * stbi__f2f(0.5411961f); \ + t2 = p1 + p3*stbi__f2f(-1.847759065f); \ + t3 = p1 + p2*stbi__f2f( 0.765366865f); \ + p2 = s0; \ + p3 = s4; \ + t0 = stbi__fsh(p2+p3); \ + t1 = stbi__fsh(p2-p3); \ + x0 = t0+t3; \ + x3 = t0-t3; \ + x1 = t1+t2; \ + x2 = t1-t2; \ + t0 = s7; \ + t1 = s5; \ + t2 = s3; \ + t3 = s1; \ + p3 = t0+t2; \ + p4 = t1+t3; \ + p1 = t0+t3; \ + p2 = t1+t2; \ + p5 = (p3+p4)*stbi__f2f( 1.175875602f); \ + t0 = t0*stbi__f2f( 0.298631336f); \ + t1 = t1*stbi__f2f( 2.053119869f); \ + t2 = t2*stbi__f2f( 3.072711026f); \ + t3 = t3*stbi__f2f( 1.501321110f); \ + p1 = p5 + p1*stbi__f2f(-0.899976223f); \ + p2 = p5 + p2*stbi__f2f(-2.562915447f); \ + p3 = p3*stbi__f2f(-1.961570560f); \ + p4 = p4*stbi__f2f(-0.390180644f); \ + t3 += p1+p4; \ + t2 += p2+p3; \ + t1 += p2+p4; \ + t0 += p1+p3; -static void stbi__idct_block(stbi_uc * out, int out_stride, short data[64]) { - int i, val[64], *v = val; - stbi_uc * o; - short * d = data; +static void stbi__idct_block(stbi_uc *out, int out_stride, short data[64]) +{ + int i,val[64],*v=val; + stbi_uc *o; + short *d = data; - // columns - for (i = 0; i < 8; ++i, ++d, ++v) { - // if all zeroes, shortcut -- this avoids dequantizing 0s and IDCTing - if (d[8] == 0 && d[16] == 0 && d[24] == 0 && d[32] == 0 && d[40] == 0 && d[48] == 0 && d[56] == 0) { - // no shortcut 0 seconds - // (1|2|3|4|5|6|7)==0 0 seconds - // all separate -0.047 seconds - // 1 && 2|3 && 4|5 && 6|7: -0.047 seconds - int dcterm = d[0] * 4; - v[0] = v[8] = v[16] = v[24] = v[32] = v[40] = v[48] = v[56] = dcterm; - } else { - STBI__IDCT_1D(d[0], d[8], d[16], d[24], d[32], d[40], d[48], d[56]) - // constants scaled things up by 1<<12; let's bring them back - // down, but keep 2 extra bits of precision - x0 += 512; - x1 += 512; - x2 += 512; - x3 += 512; - v[0] = (x0 + t3) >> 10; - v[56] = (x0 - t3) >> 10; - v[8] = (x1 + t2) >> 10; - v[48] = (x1 - t2) >> 10; - v[16] = (x2 + t1) >> 10; - v[40] = (x2 - t1) >> 10; - v[24] = (x3 + t0) >> 10; - v[32] = (x3 - t0) >> 10; - } - } + // columns + for (i=0; i < 8; ++i,++d, ++v) { + // if all zeroes, shortcut -- this avoids dequantizing 0s and IDCTing + if (d[ 8]==0 && d[16]==0 && d[24]==0 && d[32]==0 + && d[40]==0 && d[48]==0 && d[56]==0) { + // no shortcut 0 seconds + // (1|2|3|4|5|6|7)==0 0 seconds + // all separate -0.047 seconds + // 1 && 2|3 && 4|5 && 6|7: -0.047 seconds + int dcterm = d[0]*4; + v[0] = v[8] = v[16] = v[24] = v[32] = v[40] = v[48] = v[56] = dcterm; + } else { + STBI__IDCT_1D(d[ 0],d[ 8],d[16],d[24],d[32],d[40],d[48],d[56]) + // constants scaled things up by 1<<12; let's bring them back + // down, but keep 2 extra bits of precision + x0 += 512; x1 += 512; x2 += 512; x3 += 512; + v[ 0] = (x0+t3) >> 10; + v[56] = (x0-t3) >> 10; + v[ 8] = (x1+t2) >> 10; + v[48] = (x1-t2) >> 10; + v[16] = (x2+t1) >> 10; + v[40] = (x2-t1) >> 10; + v[24] = (x3+t0) >> 10; + v[32] = (x3-t0) >> 10; + } + } - for (i = 0, v = val, o = out; i < 8; ++i, v += 8, o += out_stride) { - // no fast case since the first 1D IDCT spread components out - STBI__IDCT_1D(v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]) - // constants scaled things up by 1<<12, plus we had 1<<2 from first - // loop, plus horizontal and vertical each scale by sqrt(8) so together - // we've got an extra 1<<3, so 1<<17 total we need to remove. - // so we want to round that, which means adding 0.5 * 1<<17, - // aka 65536. Also, we'll end up with -128 to 127 that we want - // to encode as 0..255 by adding 128, so we'll add that before the shift - x0 += 65536 + (128 << 17); - x1 += 65536 + (128 << 17); - x2 += 65536 + (128 << 17); - x3 += 65536 + (128 << 17); - // tried computing the shifts into temps, or'ing the temps to see - // if any were out of range, but that was slower - o[0] = stbi__clamp((x0 + t3) >> 17); - o[7] = stbi__clamp((x0 - t3) >> 17); - o[1] = stbi__clamp((x1 + t2) >> 17); - o[6] = stbi__clamp((x1 - t2) >> 17); - o[2] = stbi__clamp((x2 + t1) >> 17); - o[5] = stbi__clamp((x2 - t1) >> 17); - o[3] = stbi__clamp((x3 + t0) >> 17); - o[4] = stbi__clamp((x3 - t0) >> 17); - } + for (i=0, v=val, o=out; i < 8; ++i,v+=8,o+=out_stride) { + // no fast case since the first 1D IDCT spread components out + STBI__IDCT_1D(v[0],v[1],v[2],v[3],v[4],v[5],v[6],v[7]) + // constants scaled things up by 1<<12, plus we had 1<<2 from first + // loop, plus horizontal and vertical each scale by sqrt(8) so together + // we've got an extra 1<<3, so 1<<17 total we need to remove. + // so we want to round that, which means adding 0.5 * 1<<17, + // aka 65536. Also, we'll end up with -128 to 127 that we want + // to encode as 0..255 by adding 128, so we'll add that before the shift + x0 += 65536 + (128<<17); + x1 += 65536 + (128<<17); + x2 += 65536 + (128<<17); + x3 += 65536 + (128<<17); + // tried computing the shifts into temps, or'ing the temps to see + // if any were out of range, but that was slower + o[0] = stbi__clamp((x0+t3) >> 17); + o[7] = stbi__clamp((x0-t3) >> 17); + o[1] = stbi__clamp((x1+t2) >> 17); + o[6] = stbi__clamp((x1-t2) >> 17); + o[2] = stbi__clamp((x2+t1) >> 17); + o[5] = stbi__clamp((x2-t1) >> 17); + o[3] = stbi__clamp((x3+t0) >> 17); + o[4] = stbi__clamp((x3-t0) >> 17); + } } #ifdef STBI_SSE2 // sse2 integer IDCT. not the fastest possible implementation but it // produces bit-identical results to the generic C version so it's // fully "transparent". -static void stbi__idct_simd(stbi_uc * out, int out_stride, short data[64]) { - // This is constructed to match our regular (generic) integer IDCT exactly. - __m128i row0, row1, row2, row3, row4, row5, row6, row7; - __m128i tmp; +static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64]) +{ + // This is constructed to match our regular (generic) integer IDCT exactly. + __m128i row0, row1, row2, row3, row4, row5, row6, row7; + __m128i tmp; -// dot product constant: even elems=x, odd elems=y -#define dct_const(x, y) _mm_setr_epi16((x), (y), (x), (y), (x), (y), (x), (y)) + // dot product constant: even elems=x, odd elems=y + #define dct_const(x,y) _mm_setr_epi16((x),(y),(x),(y),(x),(y),(x),(y)) -// out(0) = c0[even]*x + c0[odd]*y (c0, x, y 16-bit, out 32-bit) -// out(1) = c1[even]*x + c1[odd]*y -#define dct_rot(out0, out1, x, y, c0, c1) \ - __m128i c0##lo = _mm_unpacklo_epi16((x), (y)); \ - __m128i c0##hi = _mm_unpackhi_epi16((x), (y)); \ - __m128i out0##_l = _mm_madd_epi16(c0##lo, c0); \ - __m128i out0##_h = _mm_madd_epi16(c0##hi, c0); \ - __m128i out1##_l = _mm_madd_epi16(c0##lo, c1); \ - __m128i out1##_h = _mm_madd_epi16(c0##hi, c1) + // out(0) = c0[even]*x + c0[odd]*y (c0, x, y 16-bit, out 32-bit) + // out(1) = c1[even]*x + c1[odd]*y + #define dct_rot(out0,out1, x,y,c0,c1) \ + __m128i c0##lo = _mm_unpacklo_epi16((x),(y)); \ + __m128i c0##hi = _mm_unpackhi_epi16((x),(y)); \ + __m128i out0##_l = _mm_madd_epi16(c0##lo, c0); \ + __m128i out0##_h = _mm_madd_epi16(c0##hi, c0); \ + __m128i out1##_l = _mm_madd_epi16(c0##lo, c1); \ + __m128i out1##_h = _mm_madd_epi16(c0##hi, c1) -// out = in << 12 (in 16-bit, out 32-bit) -#define dct_widen(out, in) \ - __m128i out##_l = _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), (in)), 4); \ - __m128i out##_h = _mm_srai_epi32(_mm_unpackhi_epi16(_mm_setzero_si128(), (in)), 4) + // out = in << 12 (in 16-bit, out 32-bit) + #define dct_widen(out, in) \ + __m128i out##_l = _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), (in)), 4); \ + __m128i out##_h = _mm_srai_epi32(_mm_unpackhi_epi16(_mm_setzero_si128(), (in)), 4) -// wide add -#define dct_wadd(out, a, b) \ - __m128i out##_l = _mm_add_epi32(a##_l, b##_l); \ - __m128i out##_h = _mm_add_epi32(a##_h, b##_h) + // wide add + #define dct_wadd(out, a, b) \ + __m128i out##_l = _mm_add_epi32(a##_l, b##_l); \ + __m128i out##_h = _mm_add_epi32(a##_h, b##_h) -// wide sub -#define dct_wsub(out, a, b) \ - __m128i out##_l = _mm_sub_epi32(a##_l, b##_l); \ - __m128i out##_h = _mm_sub_epi32(a##_h, b##_h) + // wide sub + #define dct_wsub(out, a, b) \ + __m128i out##_l = _mm_sub_epi32(a##_l, b##_l); \ + __m128i out##_h = _mm_sub_epi32(a##_h, b##_h) -// butterfly a/b, add bias, then shift by "s" and pack -#define dct_bfly32o(out0, out1, a, b, bias, s) \ - { \ - __m128i abiased_l = _mm_add_epi32(a##_l, bias); \ - __m128i abiased_h = _mm_add_epi32(a##_h, bias); \ - dct_wadd(sum, abiased, b); \ - dct_wsub(dif, abiased, b); \ - out0 = _mm_packs_epi32(_mm_srai_epi32(sum_l, s), _mm_srai_epi32(sum_h, s)); \ - out1 = _mm_packs_epi32(_mm_srai_epi32(dif_l, s), _mm_srai_epi32(dif_h, s)); \ - } + // butterfly a/b, add bias, then shift by "s" and pack + #define dct_bfly32o(out0, out1, a,b,bias,s) \ + { \ + __m128i abiased_l = _mm_add_epi32(a##_l, bias); \ + __m128i abiased_h = _mm_add_epi32(a##_h, bias); \ + dct_wadd(sum, abiased, b); \ + dct_wsub(dif, abiased, b); \ + out0 = _mm_packs_epi32(_mm_srai_epi32(sum_l, s), _mm_srai_epi32(sum_h, s)); \ + out1 = _mm_packs_epi32(_mm_srai_epi32(dif_l, s), _mm_srai_epi32(dif_h, s)); \ + } -// 8-bit interleave step (for transposes) -#define dct_interleave8(a, b) \ - tmp = a; \ - a = _mm_unpacklo_epi8(a, b); \ - b = _mm_unpackhi_epi8(tmp, b) + // 8-bit interleave step (for transposes) + #define dct_interleave8(a, b) \ + tmp = a; \ + a = _mm_unpacklo_epi8(a, b); \ + b = _mm_unpackhi_epi8(tmp, b) -// 16-bit interleave step (for transposes) -#define dct_interleave16(a, b) \ - tmp = a; \ - a = _mm_unpacklo_epi16(a, b); \ - b = _mm_unpackhi_epi16(tmp, b) + // 16-bit interleave step (for transposes) + #define dct_interleave16(a, b) \ + tmp = a; \ + a = _mm_unpacklo_epi16(a, b); \ + b = _mm_unpackhi_epi16(tmp, b) -#define dct_pass(bias, shift) \ - { \ - /* even part */ \ - dct_rot(t2e, t3e, row2, row6, rot0_0, rot0_1); \ - __m128i sum04 = _mm_add_epi16(row0, row4); \ - __m128i dif04 = _mm_sub_epi16(row0, row4); \ - dct_widen(t0e, sum04); \ - dct_widen(t1e, dif04); \ - dct_wadd(x0, t0e, t3e); \ - dct_wsub(x3, t0e, t3e); \ - dct_wadd(x1, t1e, t2e); \ - dct_wsub(x2, t1e, t2e); \ - /* odd part */ \ - dct_rot(y0o, y2o, row7, row3, rot2_0, rot2_1); \ - dct_rot(y1o, y3o, row5, row1, rot3_0, rot3_1); \ - __m128i sum17 = _mm_add_epi16(row1, row7); \ - __m128i sum35 = _mm_add_epi16(row3, row5); \ - dct_rot(y4o, y5o, sum17, sum35, rot1_0, rot1_1); \ - dct_wadd(x4, y0o, y4o); \ - dct_wadd(x5, y1o, y5o); \ - dct_wadd(x6, y2o, y5o); \ - dct_wadd(x7, y3o, y4o); \ - dct_bfly32o(row0, row7, x0, x7, bias, shift); \ - dct_bfly32o(row1, row6, x1, x6, bias, shift); \ - dct_bfly32o(row2, row5, x2, x5, bias, shift); \ - dct_bfly32o(row3, row4, x3, x4, bias, shift); \ - } + #define dct_pass(bias,shift) \ + { \ + /* even part */ \ + dct_rot(t2e,t3e, row2,row6, rot0_0,rot0_1); \ + __m128i sum04 = _mm_add_epi16(row0, row4); \ + __m128i dif04 = _mm_sub_epi16(row0, row4); \ + dct_widen(t0e, sum04); \ + dct_widen(t1e, dif04); \ + dct_wadd(x0, t0e, t3e); \ + dct_wsub(x3, t0e, t3e); \ + dct_wadd(x1, t1e, t2e); \ + dct_wsub(x2, t1e, t2e); \ + /* odd part */ \ + dct_rot(y0o,y2o, row7,row3, rot2_0,rot2_1); \ + dct_rot(y1o,y3o, row5,row1, rot3_0,rot3_1); \ + __m128i sum17 = _mm_add_epi16(row1, row7); \ + __m128i sum35 = _mm_add_epi16(row3, row5); \ + dct_rot(y4o,y5o, sum17,sum35, rot1_0,rot1_1); \ + dct_wadd(x4, y0o, y4o); \ + dct_wadd(x5, y1o, y5o); \ + dct_wadd(x6, y2o, y5o); \ + dct_wadd(x7, y3o, y4o); \ + dct_bfly32o(row0,row7, x0,x7,bias,shift); \ + dct_bfly32o(row1,row6, x1,x6,bias,shift); \ + dct_bfly32o(row2,row5, x2,x5,bias,shift); \ + dct_bfly32o(row3,row4, x3,x4,bias,shift); \ + } - __m128i rot0_0 = dct_const(stbi__f2f(0.5411961f), stbi__f2f(0.5411961f) + stbi__f2f(-1.847759065f)); - __m128i rot0_1 = dct_const(stbi__f2f(0.5411961f) + stbi__f2f(0.765366865f), stbi__f2f(0.5411961f)); - __m128i rot1_0 = dct_const(stbi__f2f(1.175875602f) + stbi__f2f(-0.899976223f), stbi__f2f(1.175875602f)); - __m128i rot1_1 = dct_const(stbi__f2f(1.175875602f), stbi__f2f(1.175875602f) + stbi__f2f(-2.562915447f)); - __m128i rot2_0 = dct_const(stbi__f2f(-1.961570560f) + stbi__f2f(0.298631336f), stbi__f2f(-1.961570560f)); - __m128i rot2_1 = dct_const(stbi__f2f(-1.961570560f), stbi__f2f(-1.961570560f) + stbi__f2f(3.072711026f)); - __m128i rot3_0 = dct_const(stbi__f2f(-0.390180644f) + stbi__f2f(2.053119869f), stbi__f2f(-0.390180644f)); - __m128i rot3_1 = dct_const(stbi__f2f(-0.390180644f), stbi__f2f(-0.390180644f) + stbi__f2f(1.501321110f)); + __m128i rot0_0 = dct_const(stbi__f2f(0.5411961f), stbi__f2f(0.5411961f) + stbi__f2f(-1.847759065f)); + __m128i rot0_1 = dct_const(stbi__f2f(0.5411961f) + stbi__f2f( 0.765366865f), stbi__f2f(0.5411961f)); + __m128i rot1_0 = dct_const(stbi__f2f(1.175875602f) + stbi__f2f(-0.899976223f), stbi__f2f(1.175875602f)); + __m128i rot1_1 = dct_const(stbi__f2f(1.175875602f), stbi__f2f(1.175875602f) + stbi__f2f(-2.562915447f)); + __m128i rot2_0 = dct_const(stbi__f2f(-1.961570560f) + stbi__f2f( 0.298631336f), stbi__f2f(-1.961570560f)); + __m128i rot2_1 = dct_const(stbi__f2f(-1.961570560f), stbi__f2f(-1.961570560f) + stbi__f2f( 3.072711026f)); + __m128i rot3_0 = dct_const(stbi__f2f(-0.390180644f) + stbi__f2f( 2.053119869f), stbi__f2f(-0.390180644f)); + __m128i rot3_1 = dct_const(stbi__f2f(-0.390180644f), stbi__f2f(-0.390180644f) + stbi__f2f( 1.501321110f)); - // rounding biases in column/row passes, see stbi__idct_block for explanation. - __m128i bias_0 = _mm_set1_epi32(512); - __m128i bias_1 = _mm_set1_epi32(65536 + (128 << 17)); + // rounding biases in column/row passes, see stbi__idct_block for explanation. + __m128i bias_0 = _mm_set1_epi32(512); + __m128i bias_1 = _mm_set1_epi32(65536 + (128<<17)); - // load - row0 = _mm_load_si128((const __m128i *)(data + 0 * 8)); - row1 = _mm_load_si128((const __m128i *)(data + 1 * 8)); - row2 = _mm_load_si128((const __m128i *)(data + 2 * 8)); - row3 = _mm_load_si128((const __m128i *)(data + 3 * 8)); - row4 = _mm_load_si128((const __m128i *)(data + 4 * 8)); - row5 = _mm_load_si128((const __m128i *)(data + 5 * 8)); - row6 = _mm_load_si128((const __m128i *)(data + 6 * 8)); - row7 = _mm_load_si128((const __m128i *)(data + 7 * 8)); + // load + row0 = _mm_load_si128((const __m128i *) (data + 0*8)); + row1 = _mm_load_si128((const __m128i *) (data + 1*8)); + row2 = _mm_load_si128((const __m128i *) (data + 2*8)); + row3 = _mm_load_si128((const __m128i *) (data + 3*8)); + row4 = _mm_load_si128((const __m128i *) (data + 4*8)); + row5 = _mm_load_si128((const __m128i *) (data + 5*8)); + row6 = _mm_load_si128((const __m128i *) (data + 6*8)); + row7 = _mm_load_si128((const __m128i *) (data + 7*8)); - // column pass - dct_pass(bias_0, 10); + // column pass + dct_pass(bias_0, 10); - { - // 16bit 8x8 transpose pass 1 - dct_interleave16(row0, row4); - dct_interleave16(row1, row5); - dct_interleave16(row2, row6); - dct_interleave16(row3, row7); + { + // 16bit 8x8 transpose pass 1 + dct_interleave16(row0, row4); + dct_interleave16(row1, row5); + dct_interleave16(row2, row6); + dct_interleave16(row3, row7); - // transpose pass 2 - dct_interleave16(row0, row2); - dct_interleave16(row1, row3); - dct_interleave16(row4, row6); - dct_interleave16(row5, row7); + // transpose pass 2 + dct_interleave16(row0, row2); + dct_interleave16(row1, row3); + dct_interleave16(row4, row6); + dct_interleave16(row5, row7); - // transpose pass 3 - dct_interleave16(row0, row1); - dct_interleave16(row2, row3); - dct_interleave16(row4, row5); - dct_interleave16(row6, row7); - } + // transpose pass 3 + dct_interleave16(row0, row1); + dct_interleave16(row2, row3); + dct_interleave16(row4, row5); + dct_interleave16(row6, row7); + } - // row pass - dct_pass(bias_1, 17); + // row pass + dct_pass(bias_1, 17); - { - // pack - __m128i p0 = _mm_packus_epi16(row0, row1); // a0a1a2a3...a7b0b1b2b3...b7 - __m128i p1 = _mm_packus_epi16(row2, row3); - __m128i p2 = _mm_packus_epi16(row4, row5); - __m128i p3 = _mm_packus_epi16(row6, row7); + { + // pack + __m128i p0 = _mm_packus_epi16(row0, row1); // a0a1a2a3...a7b0b1b2b3...b7 + __m128i p1 = _mm_packus_epi16(row2, row3); + __m128i p2 = _mm_packus_epi16(row4, row5); + __m128i p3 = _mm_packus_epi16(row6, row7); - // 8bit 8x8 transpose pass 1 - dct_interleave8(p0, p2); // a0e0a1e1... - dct_interleave8(p1, p3); // c0g0c1g1... + // 8bit 8x8 transpose pass 1 + dct_interleave8(p0, p2); // a0e0a1e1... + dct_interleave8(p1, p3); // c0g0c1g1... - // transpose pass 2 - dct_interleave8(p0, p1); // a0c0e0g0... - dct_interleave8(p2, p3); // b0d0f0h0... + // transpose pass 2 + dct_interleave8(p0, p1); // a0c0e0g0... + dct_interleave8(p2, p3); // b0d0f0h0... - // transpose pass 3 - dct_interleave8(p0, p2); // a0b0c0d0... - dct_interleave8(p1, p3); // a4b4c4d4... + // transpose pass 3 + dct_interleave8(p0, p2); // a0b0c0d0... + dct_interleave8(p1, p3); // a4b4c4d4... - // store - _mm_storel_epi64((__m128i *)out, p0); - out += out_stride; - _mm_storel_epi64((__m128i *)out, _mm_shuffle_epi32(p0, 0x4e)); - out += out_stride; - _mm_storel_epi64((__m128i *)out, p2); - out += out_stride; - _mm_storel_epi64((__m128i *)out, _mm_shuffle_epi32(p2, 0x4e)); - out += out_stride; - _mm_storel_epi64((__m128i *)out, p1); - out += out_stride; - _mm_storel_epi64((__m128i *)out, _mm_shuffle_epi32(p1, 0x4e)); - out += out_stride; - _mm_storel_epi64((__m128i *)out, p3); - out += out_stride; - _mm_storel_epi64((__m128i *)out, _mm_shuffle_epi32(p3, 0x4e)); - } + // store + _mm_storel_epi64((__m128i *) out, p0); out += out_stride; + _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p0, 0x4e)); out += out_stride; + _mm_storel_epi64((__m128i *) out, p2); out += out_stride; + _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p2, 0x4e)); out += out_stride; + _mm_storel_epi64((__m128i *) out, p1); out += out_stride; + _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p1, 0x4e)); out += out_stride; + _mm_storel_epi64((__m128i *) out, p3); out += out_stride; + _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p3, 0x4e)); + } #undef dct_const #undef dct_rot @@ -2763,235 +2708,198 @@ static void stbi__idct_simd(stbi_uc * out, int out_stride, short data[64]) { // NEON integer IDCT. should produce bit-identical // results to the generic C version. -static void stbi__idct_simd(stbi_uc * out, int out_stride, short data[64]) { - int16x8_t row0, row1, row2, row3, row4, row5, row6, row7; +static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64]) +{ + int16x8_t row0, row1, row2, row3, row4, row5, row6, row7; - int16x4_t rot0_0 = vdup_n_s16(stbi__f2f(0.5411961f)); - int16x4_t rot0_1 = vdup_n_s16(stbi__f2f(-1.847759065f)); - int16x4_t rot0_2 = vdup_n_s16(stbi__f2f(0.765366865f)); - int16x4_t rot1_0 = vdup_n_s16(stbi__f2f(1.175875602f)); - int16x4_t rot1_1 = vdup_n_s16(stbi__f2f(-0.899976223f)); - int16x4_t rot1_2 = vdup_n_s16(stbi__f2f(-2.562915447f)); - int16x4_t rot2_0 = vdup_n_s16(stbi__f2f(-1.961570560f)); - int16x4_t rot2_1 = vdup_n_s16(stbi__f2f(-0.390180644f)); - int16x4_t rot3_0 = vdup_n_s16(stbi__f2f(0.298631336f)); - int16x4_t rot3_1 = vdup_n_s16(stbi__f2f(2.053119869f)); - int16x4_t rot3_2 = vdup_n_s16(stbi__f2f(3.072711026f)); - int16x4_t rot3_3 = vdup_n_s16(stbi__f2f(1.501321110f)); + int16x4_t rot0_0 = vdup_n_s16(stbi__f2f(0.5411961f)); + int16x4_t rot0_1 = vdup_n_s16(stbi__f2f(-1.847759065f)); + int16x4_t rot0_2 = vdup_n_s16(stbi__f2f( 0.765366865f)); + int16x4_t rot1_0 = vdup_n_s16(stbi__f2f( 1.175875602f)); + int16x4_t rot1_1 = vdup_n_s16(stbi__f2f(-0.899976223f)); + int16x4_t rot1_2 = vdup_n_s16(stbi__f2f(-2.562915447f)); + int16x4_t rot2_0 = vdup_n_s16(stbi__f2f(-1.961570560f)); + int16x4_t rot2_1 = vdup_n_s16(stbi__f2f(-0.390180644f)); + int16x4_t rot3_0 = vdup_n_s16(stbi__f2f( 0.298631336f)); + int16x4_t rot3_1 = vdup_n_s16(stbi__f2f( 2.053119869f)); + int16x4_t rot3_2 = vdup_n_s16(stbi__f2f( 3.072711026f)); + int16x4_t rot3_3 = vdup_n_s16(stbi__f2f( 1.501321110f)); -#define dct_long_mul(out, inq, coeff) \ - int32x4_t out##_l = vmull_s16(vget_low_s16(inq), coeff); \ - int32x4_t out##_h = vmull_s16(vget_high_s16(inq), coeff) +#define dct_long_mul(out, inq, coeff) \ + int32x4_t out##_l = vmull_s16(vget_low_s16(inq), coeff); \ + int32x4_t out##_h = vmull_s16(vget_high_s16(inq), coeff) -#define dct_long_mac(out, acc, inq, coeff) \ - int32x4_t out##_l = vmlal_s16(acc##_l, vget_low_s16(inq), coeff); \ - int32x4_t out##_h = vmlal_s16(acc##_h, vget_high_s16(inq), coeff) +#define dct_long_mac(out, acc, inq, coeff) \ + int32x4_t out##_l = vmlal_s16(acc##_l, vget_low_s16(inq), coeff); \ + int32x4_t out##_h = vmlal_s16(acc##_h, vget_high_s16(inq), coeff) -#define dct_widen(out, inq) \ - int32x4_t out##_l = vshll_n_s16(vget_low_s16(inq), 12); \ - int32x4_t out##_h = vshll_n_s16(vget_high_s16(inq), 12) +#define dct_widen(out, inq) \ + int32x4_t out##_l = vshll_n_s16(vget_low_s16(inq), 12); \ + int32x4_t out##_h = vshll_n_s16(vget_high_s16(inq), 12) // wide add -#define dct_wadd(out, a, b) \ - int32x4_t out##_l = vaddq_s32(a##_l, b##_l); \ - int32x4_t out##_h = vaddq_s32(a##_h, b##_h) +#define dct_wadd(out, a, b) \ + int32x4_t out##_l = vaddq_s32(a##_l, b##_l); \ + int32x4_t out##_h = vaddq_s32(a##_h, b##_h) // wide sub -#define dct_wsub(out, a, b) \ - int32x4_t out##_l = vsubq_s32(a##_l, b##_l); \ - int32x4_t out##_h = vsubq_s32(a##_h, b##_h) +#define dct_wsub(out, a, b) \ + int32x4_t out##_l = vsubq_s32(a##_l, b##_l); \ + int32x4_t out##_h = vsubq_s32(a##_h, b##_h) // butterfly a/b, then shift using "shiftop" by "s" and pack -#define dct_bfly32o(out0, out1, a, b, shiftop, s) \ - { \ - dct_wadd(sum, a, b); \ - dct_wsub(dif, a, b); \ - out0 = vcombine_s16(shiftop(sum_l, s), shiftop(sum_h, s)); \ - out1 = vcombine_s16(shiftop(dif_l, s), shiftop(dif_h, s)); \ - } +#define dct_bfly32o(out0,out1, a,b,shiftop,s) \ + { \ + dct_wadd(sum, a, b); \ + dct_wsub(dif, a, b); \ + out0 = vcombine_s16(shiftop(sum_l, s), shiftop(sum_h, s)); \ + out1 = vcombine_s16(shiftop(dif_l, s), shiftop(dif_h, s)); \ + } -#define dct_pass(shiftop, shift) \ - { \ - /* even part */ \ - int16x8_t sum26 = vaddq_s16(row2, row6); \ - dct_long_mul(p1e, sum26, rot0_0); \ - dct_long_mac(t2e, p1e, row6, rot0_1); \ - dct_long_mac(t3e, p1e, row2, rot0_2); \ - int16x8_t sum04 = vaddq_s16(row0, row4); \ - int16x8_t dif04 = vsubq_s16(row0, row4); \ - dct_widen(t0e, sum04); \ - dct_widen(t1e, dif04); \ - dct_wadd(x0, t0e, t3e); \ - dct_wsub(x3, t0e, t3e); \ - dct_wadd(x1, t1e, t2e); \ - dct_wsub(x2, t1e, t2e); \ - /* odd part */ \ - int16x8_t sum15 = vaddq_s16(row1, row5); \ - int16x8_t sum17 = vaddq_s16(row1, row7); \ - int16x8_t sum35 = vaddq_s16(row3, row5); \ - int16x8_t sum37 = vaddq_s16(row3, row7); \ - int16x8_t sumodd = vaddq_s16(sum17, sum35); \ - dct_long_mul(p5o, sumodd, rot1_0); \ - dct_long_mac(p1o, p5o, sum17, rot1_1); \ - dct_long_mac(p2o, p5o, sum35, rot1_2); \ - dct_long_mul(p3o, sum37, rot2_0); \ - dct_long_mul(p4o, sum15, rot2_1); \ - dct_wadd(sump13o, p1o, p3o); \ - dct_wadd(sump24o, p2o, p4o); \ - dct_wadd(sump23o, p2o, p3o); \ - dct_wadd(sump14o, p1o, p4o); \ - dct_long_mac(x4, sump13o, row7, rot3_0); \ - dct_long_mac(x5, sump24o, row5, rot3_1); \ - dct_long_mac(x6, sump23o, row3, rot3_2); \ - dct_long_mac(x7, sump14o, row1, rot3_3); \ - dct_bfly32o(row0, row7, x0, x7, shiftop, shift); \ - dct_bfly32o(row1, row6, x1, x6, shiftop, shift); \ - dct_bfly32o(row2, row5, x2, x5, shiftop, shift); \ - dct_bfly32o(row3, row4, x3, x4, shiftop, shift); \ - } +#define dct_pass(shiftop, shift) \ + { \ + /* even part */ \ + int16x8_t sum26 = vaddq_s16(row2, row6); \ + dct_long_mul(p1e, sum26, rot0_0); \ + dct_long_mac(t2e, p1e, row6, rot0_1); \ + dct_long_mac(t3e, p1e, row2, rot0_2); \ + int16x8_t sum04 = vaddq_s16(row0, row4); \ + int16x8_t dif04 = vsubq_s16(row0, row4); \ + dct_widen(t0e, sum04); \ + dct_widen(t1e, dif04); \ + dct_wadd(x0, t0e, t3e); \ + dct_wsub(x3, t0e, t3e); \ + dct_wadd(x1, t1e, t2e); \ + dct_wsub(x2, t1e, t2e); \ + /* odd part */ \ + int16x8_t sum15 = vaddq_s16(row1, row5); \ + int16x8_t sum17 = vaddq_s16(row1, row7); \ + int16x8_t sum35 = vaddq_s16(row3, row5); \ + int16x8_t sum37 = vaddq_s16(row3, row7); \ + int16x8_t sumodd = vaddq_s16(sum17, sum35); \ + dct_long_mul(p5o, sumodd, rot1_0); \ + dct_long_mac(p1o, p5o, sum17, rot1_1); \ + dct_long_mac(p2o, p5o, sum35, rot1_2); \ + dct_long_mul(p3o, sum37, rot2_0); \ + dct_long_mul(p4o, sum15, rot2_1); \ + dct_wadd(sump13o, p1o, p3o); \ + dct_wadd(sump24o, p2o, p4o); \ + dct_wadd(sump23o, p2o, p3o); \ + dct_wadd(sump14o, p1o, p4o); \ + dct_long_mac(x4, sump13o, row7, rot3_0); \ + dct_long_mac(x5, sump24o, row5, rot3_1); \ + dct_long_mac(x6, sump23o, row3, rot3_2); \ + dct_long_mac(x7, sump14o, row1, rot3_3); \ + dct_bfly32o(row0,row7, x0,x7,shiftop,shift); \ + dct_bfly32o(row1,row6, x1,x6,shiftop,shift); \ + dct_bfly32o(row2,row5, x2,x5,shiftop,shift); \ + dct_bfly32o(row3,row4, x3,x4,shiftop,shift); \ + } - // load - row0 = vld1q_s16(data + 0 * 8); - row1 = vld1q_s16(data + 1 * 8); - row2 = vld1q_s16(data + 2 * 8); - row3 = vld1q_s16(data + 3 * 8); - row4 = vld1q_s16(data + 4 * 8); - row5 = vld1q_s16(data + 5 * 8); - row6 = vld1q_s16(data + 6 * 8); - row7 = vld1q_s16(data + 7 * 8); + // load + row0 = vld1q_s16(data + 0*8); + row1 = vld1q_s16(data + 1*8); + row2 = vld1q_s16(data + 2*8); + row3 = vld1q_s16(data + 3*8); + row4 = vld1q_s16(data + 4*8); + row5 = vld1q_s16(data + 5*8); + row6 = vld1q_s16(data + 6*8); + row7 = vld1q_s16(data + 7*8); - // add DC bias - row0 = vaddq_s16(row0, vsetq_lane_s16(1024, vdupq_n_s16(0), 0)); + // add DC bias + row0 = vaddq_s16(row0, vsetq_lane_s16(1024, vdupq_n_s16(0), 0)); - // column pass - dct_pass(vrshrn_n_s32, 10); + // column pass + dct_pass(vrshrn_n_s32, 10); - // 16bit 8x8 transpose - { + // 16bit 8x8 transpose + { // these three map to a single VTRN.16, VTRN.32, and VSWP, respectively. // whether compilers actually get this is another story, sadly. -#define dct_trn16(x, y) \ - { \ - int16x8x2_t t = vtrnq_s16(x, y); \ - x = t.val[0]; \ - y = t.val[1]; \ - } -#define dct_trn32(x, y) \ - { \ - int32x4x2_t t = vtrnq_s32(vreinterpretq_s32_s16(x), vreinterpretq_s32_s16(y)); \ - x = vreinterpretq_s16_s32(t.val[0]); \ - y = vreinterpretq_s16_s32(t.val[1]); \ - } -#define dct_trn64(x, y) \ - { \ - int16x8_t x0 = x; \ - int16x8_t y0 = y; \ - x = vcombine_s16(vget_low_s16(x0), vget_low_s16(y0)); \ - y = vcombine_s16(vget_high_s16(x0), vget_high_s16(y0)); \ - } +#define dct_trn16(x, y) { int16x8x2_t t = vtrnq_s16(x, y); x = t.val[0]; y = t.val[1]; } +#define dct_trn32(x, y) { int32x4x2_t t = vtrnq_s32(vreinterpretq_s32_s16(x), vreinterpretq_s32_s16(y)); x = vreinterpretq_s16_s32(t.val[0]); y = vreinterpretq_s16_s32(t.val[1]); } +#define dct_trn64(x, y) { int16x8_t x0 = x; int16x8_t y0 = y; x = vcombine_s16(vget_low_s16(x0), vget_low_s16(y0)); y = vcombine_s16(vget_high_s16(x0), vget_high_s16(y0)); } - // pass 1 - dct_trn16(row0, row1); // a0b0a2b2a4b4a6b6 - dct_trn16(row2, row3); - dct_trn16(row4, row5); - dct_trn16(row6, row7); + // pass 1 + dct_trn16(row0, row1); // a0b0a2b2a4b4a6b6 + dct_trn16(row2, row3); + dct_trn16(row4, row5); + dct_trn16(row6, row7); - // pass 2 - dct_trn32(row0, row2); // a0b0c0d0a4b4c4d4 - dct_trn32(row1, row3); - dct_trn32(row4, row6); - dct_trn32(row5, row7); + // pass 2 + dct_trn32(row0, row2); // a0b0c0d0a4b4c4d4 + dct_trn32(row1, row3); + dct_trn32(row4, row6); + dct_trn32(row5, row7); - // pass 3 - dct_trn64(row0, row4); // a0b0c0d0e0f0g0h0 - dct_trn64(row1, row5); - dct_trn64(row2, row6); - dct_trn64(row3, row7); + // pass 3 + dct_trn64(row0, row4); // a0b0c0d0e0f0g0h0 + dct_trn64(row1, row5); + dct_trn64(row2, row6); + dct_trn64(row3, row7); #undef dct_trn16 #undef dct_trn32 #undef dct_trn64 - } + } - // row pass - // vrshrn_n_s32 only supports shifts up to 16, we need - // 17. so do a non-rounding shift of 16 first then follow - // up with a rounding shift by 1. - dct_pass(vshrn_n_s32, 16); + // row pass + // vrshrn_n_s32 only supports shifts up to 16, we need + // 17. so do a non-rounding shift of 16 first then follow + // up with a rounding shift by 1. + dct_pass(vshrn_n_s32, 16); - { - // pack and round - uint8x8_t p0 = vqrshrun_n_s16(row0, 1); - uint8x8_t p1 = vqrshrun_n_s16(row1, 1); - uint8x8_t p2 = vqrshrun_n_s16(row2, 1); - uint8x8_t p3 = vqrshrun_n_s16(row3, 1); - uint8x8_t p4 = vqrshrun_n_s16(row4, 1); - uint8x8_t p5 = vqrshrun_n_s16(row5, 1); - uint8x8_t p6 = vqrshrun_n_s16(row6, 1); - uint8x8_t p7 = vqrshrun_n_s16(row7, 1); + { + // pack and round + uint8x8_t p0 = vqrshrun_n_s16(row0, 1); + uint8x8_t p1 = vqrshrun_n_s16(row1, 1); + uint8x8_t p2 = vqrshrun_n_s16(row2, 1); + uint8x8_t p3 = vqrshrun_n_s16(row3, 1); + uint8x8_t p4 = vqrshrun_n_s16(row4, 1); + uint8x8_t p5 = vqrshrun_n_s16(row5, 1); + uint8x8_t p6 = vqrshrun_n_s16(row6, 1); + uint8x8_t p7 = vqrshrun_n_s16(row7, 1); - // again, these can translate into one instruction, but often don't. -#define dct_trn8_8(x, y) \ - { \ - uint8x8x2_t t = vtrn_u8(x, y); \ - x = t.val[0]; \ - y = t.val[1]; \ - } -#define dct_trn8_16(x, y) \ - { \ - uint16x4x2_t t = vtrn_u16(vreinterpret_u16_u8(x), vreinterpret_u16_u8(y)); \ - x = vreinterpret_u8_u16(t.val[0]); \ - y = vreinterpret_u8_u16(t.val[1]); \ - } -#define dct_trn8_32(x, y) \ - { \ - uint32x2x2_t t = vtrn_u32(vreinterpret_u32_u8(x), vreinterpret_u32_u8(y)); \ - x = vreinterpret_u8_u32(t.val[0]); \ - y = vreinterpret_u8_u32(t.val[1]); \ - } + // again, these can translate into one instruction, but often don't. +#define dct_trn8_8(x, y) { uint8x8x2_t t = vtrn_u8(x, y); x = t.val[0]; y = t.val[1]; } +#define dct_trn8_16(x, y) { uint16x4x2_t t = vtrn_u16(vreinterpret_u16_u8(x), vreinterpret_u16_u8(y)); x = vreinterpret_u8_u16(t.val[0]); y = vreinterpret_u8_u16(t.val[1]); } +#define dct_trn8_32(x, y) { uint32x2x2_t t = vtrn_u32(vreinterpret_u32_u8(x), vreinterpret_u32_u8(y)); x = vreinterpret_u8_u32(t.val[0]); y = vreinterpret_u8_u32(t.val[1]); } - // sadly can't use interleaved stores here since we only write - // 8 bytes to each scan line! + // sadly can't use interleaved stores here since we only write + // 8 bytes to each scan line! - // 8x8 8-bit transpose pass 1 - dct_trn8_8(p0, p1); - dct_trn8_8(p2, p3); - dct_trn8_8(p4, p5); - dct_trn8_8(p6, p7); + // 8x8 8-bit transpose pass 1 + dct_trn8_8(p0, p1); + dct_trn8_8(p2, p3); + dct_trn8_8(p4, p5); + dct_trn8_8(p6, p7); - // pass 2 - dct_trn8_16(p0, p2); - dct_trn8_16(p1, p3); - dct_trn8_16(p4, p6); - dct_trn8_16(p5, p7); + // pass 2 + dct_trn8_16(p0, p2); + dct_trn8_16(p1, p3); + dct_trn8_16(p4, p6); + dct_trn8_16(p5, p7); - // pass 3 - dct_trn8_32(p0, p4); - dct_trn8_32(p1, p5); - dct_trn8_32(p2, p6); - dct_trn8_32(p3, p7); + // pass 3 + dct_trn8_32(p0, p4); + dct_trn8_32(p1, p5); + dct_trn8_32(p2, p6); + dct_trn8_32(p3, p7); - // store - vst1_u8(out, p0); - out += out_stride; - vst1_u8(out, p1); - out += out_stride; - vst1_u8(out, p2); - out += out_stride; - vst1_u8(out, p3); - out += out_stride; - vst1_u8(out, p4); - out += out_stride; - vst1_u8(out, p5); - out += out_stride; - vst1_u8(out, p6); - out += out_stride; - vst1_u8(out, p7); + // store + vst1_u8(out, p0); out += out_stride; + vst1_u8(out, p1); out += out_stride; + vst1_u8(out, p2); out += out_stride; + vst1_u8(out, p3); out += out_stride; + vst1_u8(out, p4); out += out_stride; + vst1_u8(out, p5); out += out_stride; + vst1_u8(out, p6); out += out_stride; + vst1_u8(out, p7); #undef dct_trn8_8 #undef dct_trn8_16 #undef dct_trn8_32 - } + } #undef dct_long_mul #undef dct_long_mac @@ -3004,1267 +2912,1169 @@ static void stbi__idct_simd(stbi_uc * out, int out_stride, short data[64]) { #endif // STBI_NEON -#define STBI__MARKER_none 0xff +#define STBI__MARKER_none 0xff // if there's a pending marker from the entropy stream, return that // otherwise, fetch from the stream and get a marker. if there's no // marker, return 0xff, which is never a valid marker value -static stbi_uc stbi__get_marker(stbi__jpeg * j) { - stbi_uc x; - if (j->marker != STBI__MARKER_none) { - x = j->marker; - j->marker = STBI__MARKER_none; - return x; - } - x = stbi__get8(j->s); - if (x != 0xff) - return STBI__MARKER_none; - while (x == 0xff) - x = stbi__get8(j->s); // consume repeated 0xff fill bytes - return x; +static stbi_uc stbi__get_marker(stbi__jpeg *j) +{ + stbi_uc x; + if (j->marker != STBI__MARKER_none) { x = j->marker; j->marker = STBI__MARKER_none; return x; } + x = stbi__get8(j->s); + if (x != 0xff) return STBI__MARKER_none; + while (x == 0xff) + x = stbi__get8(j->s); // consume repeated 0xff fill bytes + return x; } // in each scan, we'll have scan_n components, and the order // of the components is specified by order[] -#define STBI__RESTART(x) ((x) >= 0xd0 && (x) <= 0xd7) +#define STBI__RESTART(x) ((x) >= 0xd0 && (x) <= 0xd7) // after a restart interval, stbi__jpeg_reset the entropy decoder and // the dc prediction -static void stbi__jpeg_reset(stbi__jpeg * j) { - j->code_bits = 0; - j->code_buffer = 0; - j->nomore = 0; - j->img_comp[0].dc_pred = j->img_comp[1].dc_pred = j->img_comp[2].dc_pred = j->img_comp[3].dc_pred = 0; - j->marker = STBI__MARKER_none; - j->todo = j->restart_interval ? j->restart_interval : 0x7fffffff; - j->eob_run = 0; - // no more than 1<<31 MCUs if no restart_interal? that's plenty safe, - // since we don't even allow 1<<30 pixels +static void stbi__jpeg_reset(stbi__jpeg *j) +{ + j->code_bits = 0; + j->code_buffer = 0; + j->nomore = 0; + j->img_comp[0].dc_pred = j->img_comp[1].dc_pred = j->img_comp[2].dc_pred = j->img_comp[3].dc_pred = 0; + j->marker = STBI__MARKER_none; + j->todo = j->restart_interval ? j->restart_interval : 0x7fffffff; + j->eob_run = 0; + // no more than 1<<31 MCUs if no restart_interal? that's plenty safe, + // since we don't even allow 1<<30 pixels } -static int stbi__parse_entropy_coded_data(stbi__jpeg * z) { - stbi__jpeg_reset(z); - if (!z->progressive) { - if (z->scan_n == 1) { - int i, j; - STBI_SIMD_ALIGN(short, data[64]); - int n = z->order[0]; - // non-interleaved data, we just need to process one block at a time, - // in trivial scanline order - // number of blocks to do just depends on how many actual "pixels" this - // component has, independent of interleaved MCU blocking and such - int w = (z->img_comp[n].x + 7) >> 3; - int h = (z->img_comp[n].y + 7) >> 3; - for (j = 0; j < h; ++j) { - for (i = 0; i < w; ++i) { - int ha = z->img_comp[n].ha; - if (!stbi__jpeg_decode_block(z, data, z->huff_dc + z->img_comp[n].hd, z->huff_ac + ha, z->fast_ac[ha], n, - z->dequant[z->img_comp[n].tq])) - return 0; - z->idct_block_kernel(z->img_comp[n].data + z->img_comp[n].w2 * j * 8 + i * 8, z->img_comp[n].w2, data); - // every data block is an MCU, so countdown the restart interval - if (--z->todo <= 0) { - if (z->code_bits < 24) - stbi__grow_buffer_unsafe(z); - // if it's NOT a restart, then just bail, so we get corrupt data - // rather than no data - if (!STBI__RESTART(z->marker)) - return 1; - stbi__jpeg_reset(z); - } - } +static int stbi__parse_entropy_coded_data(stbi__jpeg *z) +{ + stbi__jpeg_reset(z); + if (!z->progressive) { + if (z->scan_n == 1) { + int i,j; + STBI_SIMD_ALIGN(short, data[64]); + int n = z->order[0]; + // non-interleaved data, we just need to process one block at a time, + // in trivial scanline order + // number of blocks to do just depends on how many actual "pixels" this + // component has, independent of interleaved MCU blocking and such + int w = (z->img_comp[n].x+7) >> 3; + int h = (z->img_comp[n].y+7) >> 3; + for (j=0; j < h; ++j) { + for (i=0; i < w; ++i) { + int ha = z->img_comp[n].ha; + if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0; + z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data); + // every data block is an MCU, so countdown the restart interval + if (--z->todo <= 0) { + if (z->code_bits < 24) stbi__grow_buffer_unsafe(z); + // if it's NOT a restart, then just bail, so we get corrupt data + // rather than no data + if (!STBI__RESTART(z->marker)) return 1; + stbi__jpeg_reset(z); + } } - return 1; - } else { // interleaved - int i, j, k, x, y; - STBI_SIMD_ALIGN(short, data[64]); - for (j = 0; j < z->img_mcu_y; ++j) { - for (i = 0; i < z->img_mcu_x; ++i) { - // scan an interleaved mcu... process scan_n components in order - for (k = 0; k < z->scan_n; ++k) { - int n = z->order[k]; - // scan out an mcu's worth of this component; that's just determined - // by the basic H and V specified for the component - for (y = 0; y < z->img_comp[n].v; ++y) { - for (x = 0; x < z->img_comp[n].h; ++x) { - int x2 = (i * z->img_comp[n].h + x) * 8; - int y2 = (j * z->img_comp[n].v + y) * 8; - int ha = z->img_comp[n].ha; - if (!stbi__jpeg_decode_block(z, data, z->huff_dc + z->img_comp[n].hd, z->huff_ac + ha, - z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) - return 0; - z->idct_block_kernel(z->img_comp[n].data + z->img_comp[n].w2 * y2 + x2, z->img_comp[n].w2, - data); - } - } - } - // after all interleaved components, that's an interleaved MCU, - // so now count down the restart interval - if (--z->todo <= 0) { - if (z->code_bits < 24) - stbi__grow_buffer_unsafe(z); - if (!STBI__RESTART(z->marker)) - return 1; - stbi__jpeg_reset(z); - } - } - } - return 1; - } - } else { - if (z->scan_n == 1) { - int i, j; - int n = z->order[0]; - // non-interleaved data, we just need to process one block at a time, - // in trivial scanline order - // number of blocks to do just depends on how many actual "pixels" this - // component has, independent of interleaved MCU blocking and such - int w = (z->img_comp[n].x + 7) >> 3; - int h = (z->img_comp[n].y + 7) >> 3; - for (j = 0; j < h; ++j) { - for (i = 0; i < w; ++i) { - short * data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w); - if (z->spec_start == 0) { - if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n)) - return 0; - } else { + } + return 1; + } else { // interleaved + int i,j,k,x,y; + STBI_SIMD_ALIGN(short, data[64]); + for (j=0; j < z->img_mcu_y; ++j) { + for (i=0; i < z->img_mcu_x; ++i) { + // scan an interleaved mcu... process scan_n components in order + for (k=0; k < z->scan_n; ++k) { + int n = z->order[k]; + // scan out an mcu's worth of this component; that's just determined + // by the basic H and V specified for the component + for (y=0; y < z->img_comp[n].v; ++y) { + for (x=0; x < z->img_comp[n].h; ++x) { + int x2 = (i*z->img_comp[n].h + x)*8; + int y2 = (j*z->img_comp[n].v + y)*8; int ha = z->img_comp[n].ha; - if (!stbi__jpeg_decode_block_prog_ac(z, data, &z->huff_ac[ha], z->fast_ac[ha])) - return 0; - } - // every data block is an MCU, so countdown the restart interval - if (--z->todo <= 0) { - if (z->code_bits < 24) - stbi__grow_buffer_unsafe(z); - if (!STBI__RESTART(z->marker)) - return 1; - stbi__jpeg_reset(z); - } - } + if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0; + z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*y2+x2, z->img_comp[n].w2, data); + } + } + } + // after all interleaved components, that's an interleaved MCU, + // so now count down the restart interval + if (--z->todo <= 0) { + if (z->code_bits < 24) stbi__grow_buffer_unsafe(z); + if (!STBI__RESTART(z->marker)) return 1; + stbi__jpeg_reset(z); + } } - return 1; - } else { // interleaved - int i, j, k, x, y; - for (j = 0; j < z->img_mcu_y; ++j) { - for (i = 0; i < z->img_mcu_x; ++i) { - // scan an interleaved mcu... process scan_n components in order - for (k = 0; k < z->scan_n; ++k) { - int n = z->order[k]; - // scan out an mcu's worth of this component; that's just determined - // by the basic H and V specified for the component - for (y = 0; y < z->img_comp[n].v; ++y) { - for (x = 0; x < z->img_comp[n].h; ++x) { - int x2 = (i * z->img_comp[n].h + x); - int y2 = (j * z->img_comp[n].v + y); - short * data = z->img_comp[n].coeff + 64 * (x2 + y2 * z->img_comp[n].coeff_w); - if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n)) - return 0; - } - } - } - // after all interleaved components, that's an interleaved MCU, - // so now count down the restart interval - if (--z->todo <= 0) { - if (z->code_bits < 24) - stbi__grow_buffer_unsafe(z); - if (!STBI__RESTART(z->marker)) - return 1; - stbi__jpeg_reset(z); - } - } + } + return 1; + } + } else { + if (z->scan_n == 1) { + int i,j; + int n = z->order[0]; + // non-interleaved data, we just need to process one block at a time, + // in trivial scanline order + // number of blocks to do just depends on how many actual "pixels" this + // component has, independent of interleaved MCU blocking and such + int w = (z->img_comp[n].x+7) >> 3; + int h = (z->img_comp[n].y+7) >> 3; + for (j=0; j < h; ++j) { + for (i=0; i < w; ++i) { + short *data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w); + if (z->spec_start == 0) { + if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n)) + return 0; + } else { + int ha = z->img_comp[n].ha; + if (!stbi__jpeg_decode_block_prog_ac(z, data, &z->huff_ac[ha], z->fast_ac[ha])) + return 0; + } + // every data block is an MCU, so countdown the restart interval + if (--z->todo <= 0) { + if (z->code_bits < 24) stbi__grow_buffer_unsafe(z); + if (!STBI__RESTART(z->marker)) return 1; + stbi__jpeg_reset(z); + } } - return 1; - } - } + } + return 1; + } else { // interleaved + int i,j,k,x,y; + for (j=0; j < z->img_mcu_y; ++j) { + for (i=0; i < z->img_mcu_x; ++i) { + // scan an interleaved mcu... process scan_n components in order + for (k=0; k < z->scan_n; ++k) { + int n = z->order[k]; + // scan out an mcu's worth of this component; that's just determined + // by the basic H and V specified for the component + for (y=0; y < z->img_comp[n].v; ++y) { + for (x=0; x < z->img_comp[n].h; ++x) { + int x2 = (i*z->img_comp[n].h + x); + int y2 = (j*z->img_comp[n].v + y); + short *data = z->img_comp[n].coeff + 64 * (x2 + y2 * z->img_comp[n].coeff_w); + if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n)) + return 0; + } + } + } + // after all interleaved components, that's an interleaved MCU, + // so now count down the restart interval + if (--z->todo <= 0) { + if (z->code_bits < 24) stbi__grow_buffer_unsafe(z); + if (!STBI__RESTART(z->marker)) return 1; + stbi__jpeg_reset(z); + } + } + } + return 1; + } + } } -static void stbi__jpeg_dequantize(short * data, stbi__uint16 * dequant) { - int i; - for (i = 0; i < 64; ++i) - data[i] *= dequant[i]; +static void stbi__jpeg_dequantize(short *data, stbi__uint16 *dequant) +{ + int i; + for (i=0; i < 64; ++i) + data[i] *= dequant[i]; } -static void stbi__jpeg_finish(stbi__jpeg * z) { - if (z->progressive) { - // dequantize and idct the data - int i, j, n; - for (n = 0; n < z->s->img_n; ++n) { - int w = (z->img_comp[n].x + 7) >> 3; - int h = (z->img_comp[n].y + 7) >> 3; - for (j = 0; j < h; ++j) { - for (i = 0; i < w; ++i) { - short * data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w); - stbi__jpeg_dequantize(data, z->dequant[z->img_comp[n].tq]); - z->idct_block_kernel(z->img_comp[n].data + z->img_comp[n].w2 * j * 8 + i * 8, z->img_comp[n].w2, data); - } +static void stbi__jpeg_finish(stbi__jpeg *z) +{ + if (z->progressive) { + // dequantize and idct the data + int i,j,n; + for (n=0; n < z->s->img_n; ++n) { + int w = (z->img_comp[n].x+7) >> 3; + int h = (z->img_comp[n].y+7) >> 3; + for (j=0; j < h; ++j) { + for (i=0; i < w; ++i) { + short *data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w); + stbi__jpeg_dequantize(data, z->dequant[z->img_comp[n].tq]); + z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data); } - } - } + } + } + } } -static int stbi__process_marker(stbi__jpeg * z, int m) { - int L; - switch (m) { - case STBI__MARKER_none: // no marker found - return stbi__err("expected marker", "Corrupt JPEG"); +static int stbi__process_marker(stbi__jpeg *z, int m) +{ + int L; + switch (m) { + case STBI__MARKER_none: // no marker found + return stbi__err("expected marker","Corrupt JPEG"); - case 0xDD: // DRI - specify restart interval - if (stbi__get16be(z->s) != 4) - return stbi__err("bad DRI len", "Corrupt JPEG"); - z->restart_interval = stbi__get16be(z->s); - return 1; + case 0xDD: // DRI - specify restart interval + if (stbi__get16be(z->s) != 4) return stbi__err("bad DRI len","Corrupt JPEG"); + z->restart_interval = stbi__get16be(z->s); + return 1; - case 0xDB: // DQT - define quantization table - L = stbi__get16be(z->s) - 2; - while (L > 0) { + case 0xDB: // DQT - define quantization table + L = stbi__get16be(z->s)-2; + while (L > 0) { int q = stbi__get8(z->s); int p = q >> 4, sixteen = (p != 0); - int t = q & 15, i; - if (p != 0 && p != 1) - return stbi__err("bad DQT type", "Corrupt JPEG"); - if (t > 3) - return stbi__err("bad DQT table", "Corrupt JPEG"); + int t = q & 15,i; + if (p != 0 && p != 1) return stbi__err("bad DQT type","Corrupt JPEG"); + if (t > 3) return stbi__err("bad DQT table","Corrupt JPEG"); - for (i = 0; i < 64; ++i) - z->dequant[t][stbi__jpeg_dezigzag[i]] = (stbi__uint16)(sixteen ? stbi__get16be(z->s) : stbi__get8(z->s)); + for (i=0; i < 64; ++i) + z->dequant[t][stbi__jpeg_dezigzag[i]] = (stbi__uint16)(sixteen ? stbi__get16be(z->s) : stbi__get8(z->s)); L -= (sixteen ? 129 : 65); - } - return L == 0; + } + return L==0; - case 0xC4: // DHT - define huffman table - L = stbi__get16be(z->s) - 2; - while (L > 0) { - stbi_uc * v; - int sizes[16], i, n = 0; + case 0xC4: // DHT - define huffman table + L = stbi__get16be(z->s)-2; + while (L > 0) { + stbi_uc *v; + int sizes[16],i,n=0; int q = stbi__get8(z->s); int tc = q >> 4; int th = q & 15; - if (tc > 1 || th > 3) - return stbi__err("bad DHT header", "Corrupt JPEG"); - for (i = 0; i < 16; ++i) { - sizes[i] = stbi__get8(z->s); - n += sizes[i]; + if (tc > 1 || th > 3) return stbi__err("bad DHT header","Corrupt JPEG"); + for (i=0; i < 16; ++i) { + sizes[i] = stbi__get8(z->s); + n += sizes[i]; } - if (n > 256) - return stbi__err("bad DHT header", "Corrupt JPEG"); // Loop over i < n would write past end of values! + if(n > 256) return stbi__err("bad DHT header","Corrupt JPEG"); // Loop over i < n would write past end of values! L -= 17; if (tc == 0) { - if (!stbi__build_huffman(z->huff_dc + th, sizes)) - return 0; - v = z->huff_dc[th].values; + if (!stbi__build_huffman(z->huff_dc+th, sizes)) return 0; + v = z->huff_dc[th].values; } else { - if (!stbi__build_huffman(z->huff_ac + th, sizes)) - return 0; - v = z->huff_ac[th].values; + if (!stbi__build_huffman(z->huff_ac+th, sizes)) return 0; + v = z->huff_ac[th].values; } - for (i = 0; i < n; ++i) - v[i] = stbi__get8(z->s); + for (i=0; i < n; ++i) + v[i] = stbi__get8(z->s); if (tc != 0) - stbi__build_fast_ac(z->fast_ac[th], z->huff_ac + th); + stbi__build_fast_ac(z->fast_ac[th], z->huff_ac + th); L -= n; - } - return L == 0; - } + } + return L==0; + } - // check for comment block or APP blocks - if ((m >= 0xE0 && m <= 0xEF) || m == 0xFE) { - L = stbi__get16be(z->s); - if (L < 2) { - if (m == 0xFE) - return stbi__err("bad COM len", "Corrupt JPEG"); - else - return stbi__err("bad APP len", "Corrupt JPEG"); - } - L -= 2; + // check for comment block or APP blocks + if ((m >= 0xE0 && m <= 0xEF) || m == 0xFE) { + L = stbi__get16be(z->s); + if (L < 2) { + if (m == 0xFE) + return stbi__err("bad COM len","Corrupt JPEG"); + else + return stbi__err("bad APP len","Corrupt JPEG"); + } + L -= 2; - if (m == 0xE0 && L >= 5) { // JFIF APP0 segment - static const unsigned char tag[5] = {'J', 'F', 'I', 'F', '\0'}; - int ok = 1; - int i; - for (i = 0; i < 5; ++i) - if (stbi__get8(z->s) != tag[i]) - ok = 0; - L -= 5; - if (ok) - z->jfif = 1; - } else if (m == 0xEE && L >= 12) { // Adobe APP14 segment - static const unsigned char tag[6] = {'A', 'd', 'o', 'b', 'e', '\0'}; - int ok = 1; - int i; - for (i = 0; i < 6; ++i) - if (stbi__get8(z->s) != tag[i]) - ok = 0; + if (m == 0xE0 && L >= 5) { // JFIF APP0 segment + static const unsigned char tag[5] = {'J','F','I','F','\0'}; + int ok = 1; + int i; + for (i=0; i < 5; ++i) + if (stbi__get8(z->s) != tag[i]) + ok = 0; + L -= 5; + if (ok) + z->jfif = 1; + } else if (m == 0xEE && L >= 12) { // Adobe APP14 segment + static const unsigned char tag[6] = {'A','d','o','b','e','\0'}; + int ok = 1; + int i; + for (i=0; i < 6; ++i) + if (stbi__get8(z->s) != tag[i]) + ok = 0; + L -= 6; + if (ok) { + stbi__get8(z->s); // version + stbi__get16be(z->s); // flags0 + stbi__get16be(z->s); // flags1 + z->app14_color_transform = stbi__get8(z->s); // color transform L -= 6; - if (ok) { - stbi__get8(z->s); // version - stbi__get16be(z->s); // flags0 - stbi__get16be(z->s); // flags1 - z->app14_color_transform = stbi__get8(z->s); // color transform - L -= 6; - } - } + } + } - stbi__skip(z->s, L); - return 1; - } + stbi__skip(z->s, L); + return 1; + } - return stbi__err("unknown marker", "Corrupt JPEG"); + return stbi__err("unknown marker","Corrupt JPEG"); } // after we see SOS -static int stbi__process_scan_header(stbi__jpeg * z) { - int i; - int Ls = stbi__get16be(z->s); - z->scan_n = stbi__get8(z->s); - if (z->scan_n < 1 || z->scan_n > 4 || z->scan_n > (int)z->s->img_n) - return stbi__err("bad SOS component count", "Corrupt JPEG"); - if (Ls != 6 + 2 * z->scan_n) - return stbi__err("bad SOS len", "Corrupt JPEG"); - for (i = 0; i < z->scan_n; ++i) { - int id = stbi__get8(z->s), which; - int q = stbi__get8(z->s); - for (which = 0; which < z->s->img_n; ++which) - if (z->img_comp[which].id == id) - break; - if (which == z->s->img_n) - return 0; // no match - z->img_comp[which].hd = q >> 4; - if (z->img_comp[which].hd > 3) - return stbi__err("bad DC huff", "Corrupt JPEG"); - z->img_comp[which].ha = q & 15; - if (z->img_comp[which].ha > 3) - return stbi__err("bad AC huff", "Corrupt JPEG"); - z->order[i] = which; - } +static int stbi__process_scan_header(stbi__jpeg *z) +{ + int i; + int Ls = stbi__get16be(z->s); + z->scan_n = stbi__get8(z->s); + if (z->scan_n < 1 || z->scan_n > 4 || z->scan_n > (int) z->s->img_n) return stbi__err("bad SOS component count","Corrupt JPEG"); + if (Ls != 6+2*z->scan_n) return stbi__err("bad SOS len","Corrupt JPEG"); + for (i=0; i < z->scan_n; ++i) { + int id = stbi__get8(z->s), which; + int q = stbi__get8(z->s); + for (which = 0; which < z->s->img_n; ++which) + if (z->img_comp[which].id == id) + break; + if (which == z->s->img_n) return 0; // no match + z->img_comp[which].hd = q >> 4; if (z->img_comp[which].hd > 3) return stbi__err("bad DC huff","Corrupt JPEG"); + z->img_comp[which].ha = q & 15; if (z->img_comp[which].ha > 3) return stbi__err("bad AC huff","Corrupt JPEG"); + z->order[i] = which; + } - { - int aa; - z->spec_start = stbi__get8(z->s); - z->spec_end = stbi__get8(z->s); // should be 63, but might be 0 - aa = stbi__get8(z->s); - z->succ_high = (aa >> 4); - z->succ_low = (aa & 15); - if (z->progressive) { - if (z->spec_start > 63 || z->spec_end > 63 || z->spec_start > z->spec_end || z->succ_high > 13 || z->succ_low > 13) - return stbi__err("bad SOS", "Corrupt JPEG"); - } else { - if (z->spec_start != 0) - return stbi__err("bad SOS", "Corrupt JPEG"); - if (z->succ_high != 0 || z->succ_low != 0) - return stbi__err("bad SOS", "Corrupt JPEG"); - z->spec_end = 63; - } - } + { + int aa; + z->spec_start = stbi__get8(z->s); + z->spec_end = stbi__get8(z->s); // should be 63, but might be 0 + aa = stbi__get8(z->s); + z->succ_high = (aa >> 4); + z->succ_low = (aa & 15); + if (z->progressive) { + if (z->spec_start > 63 || z->spec_end > 63 || z->spec_start > z->spec_end || z->succ_high > 13 || z->succ_low > 13) + return stbi__err("bad SOS", "Corrupt JPEG"); + } else { + if (z->spec_start != 0) return stbi__err("bad SOS","Corrupt JPEG"); + if (z->succ_high != 0 || z->succ_low != 0) return stbi__err("bad SOS","Corrupt JPEG"); + z->spec_end = 63; + } + } - return 1; + return 1; } -static int stbi__free_jpeg_components(stbi__jpeg * z, int ncomp, int why) { - int i; - for (i = 0; i < ncomp; ++i) { - if (z->img_comp[i].raw_data) { - STBI_FREE(z->img_comp[i].raw_data); - z->img_comp[i].raw_data = NULL; - z->img_comp[i].data = NULL; - } - if (z->img_comp[i].raw_coeff) { - STBI_FREE(z->img_comp[i].raw_coeff); - z->img_comp[i].raw_coeff = 0; - z->img_comp[i].coeff = 0; - } - if (z->img_comp[i].linebuf) { - STBI_FREE(z->img_comp[i].linebuf); - z->img_comp[i].linebuf = NULL; - } - } - return why; +static int stbi__free_jpeg_components(stbi__jpeg *z, int ncomp, int why) +{ + int i; + for (i=0; i < ncomp; ++i) { + if (z->img_comp[i].raw_data) { + STBI_FREE(z->img_comp[i].raw_data); + z->img_comp[i].raw_data = NULL; + z->img_comp[i].data = NULL; + } + if (z->img_comp[i].raw_coeff) { + STBI_FREE(z->img_comp[i].raw_coeff); + z->img_comp[i].raw_coeff = 0; + z->img_comp[i].coeff = 0; + } + if (z->img_comp[i].linebuf) { + STBI_FREE(z->img_comp[i].linebuf); + z->img_comp[i].linebuf = NULL; + } + } + return why; } -static int stbi__process_frame_header(stbi__jpeg * z, int scan) { - stbi__context * s = z->s; - int Lf, p, i, q, h_max = 1, v_max = 1, c; - Lf = stbi__get16be(s); - if (Lf < 11) - return stbi__err("bad SOF len", "Corrupt JPEG"); // JPEG - p = stbi__get8(s); - if (p != 8) - return stbi__err("only 8-bit", "JPEG format not supported: 8-bit only"); // JPEG baseline - s->img_y = stbi__get16be(s); - if (s->img_y == 0) - return stbi__err("no header height", - "JPEG format not supported: delayed height"); // Legal, but we don't handle it--but neither does IJG - s->img_x = stbi__get16be(s); - if (s->img_x == 0) - return stbi__err("0 width", "Corrupt JPEG"); // JPEG requires - if (s->img_y > STBI_MAX_DIMENSIONS) - return stbi__err("too large", "Very large image (corrupt?)"); - if (s->img_x > STBI_MAX_DIMENSIONS) - return stbi__err("too large", "Very large image (corrupt?)"); - c = stbi__get8(s); - if (c != 3 && c != 1 && c != 4) - return stbi__err("bad component count", "Corrupt JPEG"); - s->img_n = c; - for (i = 0; i < c; ++i) { - z->img_comp[i].data = NULL; - z->img_comp[i].linebuf = NULL; - } +static int stbi__process_frame_header(stbi__jpeg *z, int scan) +{ + stbi__context *s = z->s; + int Lf,p,i,q, h_max=1,v_max=1,c; + Lf = stbi__get16be(s); if (Lf < 11) return stbi__err("bad SOF len","Corrupt JPEG"); // JPEG + p = stbi__get8(s); if (p != 8) return stbi__err("only 8-bit","JPEG format not supported: 8-bit only"); // JPEG baseline + s->img_y = stbi__get16be(s); if (s->img_y == 0) return stbi__err("no header height", "JPEG format not supported: delayed height"); // Legal, but we don't handle it--but neither does IJG + s->img_x = stbi__get16be(s); if (s->img_x == 0) return stbi__err("0 width","Corrupt JPEG"); // JPEG requires + if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)"); + if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)"); + c = stbi__get8(s); + if (c != 3 && c != 1 && c != 4) return stbi__err("bad component count","Corrupt JPEG"); + s->img_n = c; + for (i=0; i < c; ++i) { + z->img_comp[i].data = NULL; + z->img_comp[i].linebuf = NULL; + } - if (Lf != 8 + 3 * s->img_n) - return stbi__err("bad SOF len", "Corrupt JPEG"); + if (Lf != 8+3*s->img_n) return stbi__err("bad SOF len","Corrupt JPEG"); - z->rgb = 0; - for (i = 0; i < s->img_n; ++i) { - static const unsigned char rgb[3] = {'R', 'G', 'B'}; - z->img_comp[i].id = stbi__get8(s); - if (s->img_n == 3 && z->img_comp[i].id == rgb[i]) - ++z->rgb; - q = stbi__get8(s); - z->img_comp[i].h = (q >> 4); - if (!z->img_comp[i].h || z->img_comp[i].h > 4) - return stbi__err("bad H", "Corrupt JPEG"); - z->img_comp[i].v = q & 15; - if (!z->img_comp[i].v || z->img_comp[i].v > 4) - return stbi__err("bad V", "Corrupt JPEG"); - z->img_comp[i].tq = stbi__get8(s); - if (z->img_comp[i].tq > 3) - return stbi__err("bad TQ", "Corrupt JPEG"); - } + z->rgb = 0; + for (i=0; i < s->img_n; ++i) { + static const unsigned char rgb[3] = { 'R', 'G', 'B' }; + z->img_comp[i].id = stbi__get8(s); + if (s->img_n == 3 && z->img_comp[i].id == rgb[i]) + ++z->rgb; + q = stbi__get8(s); + z->img_comp[i].h = (q >> 4); if (!z->img_comp[i].h || z->img_comp[i].h > 4) return stbi__err("bad H","Corrupt JPEG"); + z->img_comp[i].v = q & 15; if (!z->img_comp[i].v || z->img_comp[i].v > 4) return stbi__err("bad V","Corrupt JPEG"); + z->img_comp[i].tq = stbi__get8(s); if (z->img_comp[i].tq > 3) return stbi__err("bad TQ","Corrupt JPEG"); + } - if (scan != STBI__SCAN_load) - return 1; + if (scan != STBI__SCAN_load) return 1; - if (!stbi__mad3sizes_valid(s->img_x, s->img_y, s->img_n, 0)) - return stbi__err("too large", "Image too large to decode"); + if (!stbi__mad3sizes_valid(s->img_x, s->img_y, s->img_n, 0)) return stbi__err("too large", "Image too large to decode"); - for (i = 0; i < s->img_n; ++i) { - if (z->img_comp[i].h > h_max) - h_max = z->img_comp[i].h; - if (z->img_comp[i].v > v_max) - v_max = z->img_comp[i].v; - } + for (i=0; i < s->img_n; ++i) { + if (z->img_comp[i].h > h_max) h_max = z->img_comp[i].h; + if (z->img_comp[i].v > v_max) v_max = z->img_comp[i].v; + } - // check that plane subsampling factors are integer ratios; our resamplers can't deal with fractional ratios - // and I've never seen a non-corrupted JPEG file actually use them - for (i = 0; i < s->img_n; ++i) { - if (h_max % z->img_comp[i].h != 0) - return stbi__err("bad H", "Corrupt JPEG"); - if (v_max % z->img_comp[i].v != 0) - return stbi__err("bad V", "Corrupt JPEG"); - } + // check that plane subsampling factors are integer ratios; our resamplers can't deal with fractional ratios + // and I've never seen a non-corrupted JPEG file actually use them + for (i=0; i < s->img_n; ++i) { + if (h_max % z->img_comp[i].h != 0) return stbi__err("bad H","Corrupt JPEG"); + if (v_max % z->img_comp[i].v != 0) return stbi__err("bad V","Corrupt JPEG"); + } - // compute interleaved mcu info - z->img_h_max = h_max; - z->img_v_max = v_max; - z->img_mcu_w = h_max * 8; - z->img_mcu_h = v_max * 8; - // these sizes can't be more than 17 bits - z->img_mcu_x = (s->img_x + z->img_mcu_w - 1) / z->img_mcu_w; - z->img_mcu_y = (s->img_y + z->img_mcu_h - 1) / z->img_mcu_h; + // compute interleaved mcu info + z->img_h_max = h_max; + z->img_v_max = v_max; + z->img_mcu_w = h_max * 8; + z->img_mcu_h = v_max * 8; + // these sizes can't be more than 17 bits + z->img_mcu_x = (s->img_x + z->img_mcu_w-1) / z->img_mcu_w; + z->img_mcu_y = (s->img_y + z->img_mcu_h-1) / z->img_mcu_h; - for (i = 0; i < s->img_n; ++i) { - // number of effective pixels (e.g. for non-interleaved MCU) - z->img_comp[i].x = (s->img_x * z->img_comp[i].h + h_max - 1) / h_max; - z->img_comp[i].y = (s->img_y * z->img_comp[i].v + v_max - 1) / v_max; - // to simplify generation, we'll allocate enough memory to decode - // the bogus oversized data from using interleaved MCUs and their - // big blocks (e.g. a 16x16 iMCU on an image of width 33); we won't - // discard the extra data until colorspace conversion - // - // img_mcu_x, img_mcu_y: <=17 bits; comp[i].h and .v are <=4 (checked earlier) - // so these muls can't overflow with 32-bit ints (which we require) - z->img_comp[i].w2 = z->img_mcu_x * z->img_comp[i].h * 8; - z->img_comp[i].h2 = z->img_mcu_y * z->img_comp[i].v * 8; - z->img_comp[i].coeff = 0; - z->img_comp[i].raw_coeff = 0; - z->img_comp[i].linebuf = NULL; - z->img_comp[i].raw_data = stbi__malloc_mad2(z->img_comp[i].w2, z->img_comp[i].h2, 15); - if (z->img_comp[i].raw_data == NULL) - return stbi__free_jpeg_components(z, i + 1, stbi__err("outofmem", "Out of memory")); - // align blocks for idct using mmx/sse - z->img_comp[i].data = (stbi_uc *)(((size_t)z->img_comp[i].raw_data + 15) & ~15); - if (z->progressive) { - // w2, h2 are multiples of 8 (see above) - z->img_comp[i].coeff_w = z->img_comp[i].w2 / 8; - z->img_comp[i].coeff_h = z->img_comp[i].h2 / 8; - z->img_comp[i].raw_coeff = stbi__malloc_mad3(z->img_comp[i].w2, z->img_comp[i].h2, sizeof(short), 15); - if (z->img_comp[i].raw_coeff == NULL) - return stbi__free_jpeg_components(z, i + 1, stbi__err("outofmem", "Out of memory")); - z->img_comp[i].coeff = (short *)(((size_t)z->img_comp[i].raw_coeff + 15) & ~15); - } - } + for (i=0; i < s->img_n; ++i) { + // number of effective pixels (e.g. for non-interleaved MCU) + z->img_comp[i].x = (s->img_x * z->img_comp[i].h + h_max-1) / h_max; + z->img_comp[i].y = (s->img_y * z->img_comp[i].v + v_max-1) / v_max; + // to simplify generation, we'll allocate enough memory to decode + // the bogus oversized data from using interleaved MCUs and their + // big blocks (e.g. a 16x16 iMCU on an image of width 33); we won't + // discard the extra data until colorspace conversion + // + // img_mcu_x, img_mcu_y: <=17 bits; comp[i].h and .v are <=4 (checked earlier) + // so these muls can't overflow with 32-bit ints (which we require) + z->img_comp[i].w2 = z->img_mcu_x * z->img_comp[i].h * 8; + z->img_comp[i].h2 = z->img_mcu_y * z->img_comp[i].v * 8; + z->img_comp[i].coeff = 0; + z->img_comp[i].raw_coeff = 0; + z->img_comp[i].linebuf = NULL; + z->img_comp[i].raw_data = stbi__malloc_mad2(z->img_comp[i].w2, z->img_comp[i].h2, 15); + if (z->img_comp[i].raw_data == NULL) + return stbi__free_jpeg_components(z, i+1, stbi__err("outofmem", "Out of memory")); + // align blocks for idct using mmx/sse + z->img_comp[i].data = (stbi_uc*) (((size_t) z->img_comp[i].raw_data + 15) & ~15); + if (z->progressive) { + // w2, h2 are multiples of 8 (see above) + z->img_comp[i].coeff_w = z->img_comp[i].w2 / 8; + z->img_comp[i].coeff_h = z->img_comp[i].h2 / 8; + z->img_comp[i].raw_coeff = stbi__malloc_mad3(z->img_comp[i].w2, z->img_comp[i].h2, sizeof(short), 15); + if (z->img_comp[i].raw_coeff == NULL) + return stbi__free_jpeg_components(z, i+1, stbi__err("outofmem", "Out of memory")); + z->img_comp[i].coeff = (short*) (((size_t) z->img_comp[i].raw_coeff + 15) & ~15); + } + } - return 1; + return 1; } // use comparisons since in some cases we handle more than one case (e.g. SOF) -#define stbi__DNL(x) ((x) == 0xdc) -#define stbi__SOI(x) ((x) == 0xd8) -#define stbi__EOI(x) ((x) == 0xd9) -#define stbi__SOF(x) ((x) == 0xc0 || (x) == 0xc1 || (x) == 0xc2) -#define stbi__SOS(x) ((x) == 0xda) +#define stbi__DNL(x) ((x) == 0xdc) +#define stbi__SOI(x) ((x) == 0xd8) +#define stbi__EOI(x) ((x) == 0xd9) +#define stbi__SOF(x) ((x) == 0xc0 || (x) == 0xc1 || (x) == 0xc2) +#define stbi__SOS(x) ((x) == 0xda) -#define stbi__SOF_progressive(x) ((x) == 0xc2) +#define stbi__SOF_progressive(x) ((x) == 0xc2) -static int stbi__decode_jpeg_header(stbi__jpeg * z, int scan) { - int m; - z->jfif = 0; - z->app14_color_transform = -1; // valid values are 0,1,2 - z->marker = STBI__MARKER_none; // initialize cached marker to empty - m = stbi__get_marker(z); - if (!stbi__SOI(m)) - return stbi__err("no SOI", "Corrupt JPEG"); - if (scan == STBI__SCAN_type) - return 1; - m = stbi__get_marker(z); - while (!stbi__SOF(m)) { - if (!stbi__process_marker(z, m)) - return 0; - m = stbi__get_marker(z); - while (m == STBI__MARKER_none) { - // some files have extra padding after their blocks, so ok, we'll scan - if (stbi__at_eof(z->s)) - return stbi__err("no SOF", "Corrupt JPEG"); - m = stbi__get_marker(z); - } - } - z->progressive = stbi__SOF_progressive(m); - if (!stbi__process_frame_header(z, scan)) - return 0; - return 1; +static int stbi__decode_jpeg_header(stbi__jpeg *z, int scan) +{ + int m; + z->jfif = 0; + z->app14_color_transform = -1; // valid values are 0,1,2 + z->marker = STBI__MARKER_none; // initialize cached marker to empty + m = stbi__get_marker(z); + if (!stbi__SOI(m)) return stbi__err("no SOI","Corrupt JPEG"); + if (scan == STBI__SCAN_type) return 1; + m = stbi__get_marker(z); + while (!stbi__SOF(m)) { + if (!stbi__process_marker(z,m)) return 0; + m = stbi__get_marker(z); + while (m == STBI__MARKER_none) { + // some files have extra padding after their blocks, so ok, we'll scan + if (stbi__at_eof(z->s)) return stbi__err("no SOF", "Corrupt JPEG"); + m = stbi__get_marker(z); + } + } + z->progressive = stbi__SOF_progressive(m); + if (!stbi__process_frame_header(z, scan)) return 0; + return 1; } -static int stbi__skip_jpeg_junk_at_end(stbi__jpeg * j) { - // some JPEGs have junk at end, skip over it but if we find what looks - // like a valid marker, resume there - while (!stbi__at_eof(j->s)) { - int x = stbi__get8(j->s); - while (x == 255) { // might be a marker - if (stbi__at_eof(j->s)) - return STBI__MARKER_none; - x = stbi__get8(j->s); - if (x != 0x00 && x != 0xff) { - // not a stuffed zero or lead-in to another marker, looks - // like an actual marker, return it - return x; - } - // stuffed zero has x=0 now which ends the loop, meaning we go - // back to regular scan loop. - // repeated 0xff keeps trying to read the next byte of the marker. - } - } - return STBI__MARKER_none; +static stbi_uc stbi__skip_jpeg_junk_at_end(stbi__jpeg *j) +{ + // some JPEGs have junk at end, skip over it but if we find what looks + // like a valid marker, resume there + while (!stbi__at_eof(j->s)) { + stbi_uc x = stbi__get8(j->s); + while (x == 0xff) { // might be a marker + if (stbi__at_eof(j->s)) return STBI__MARKER_none; + x = stbi__get8(j->s); + if (x != 0x00 && x != 0xff) { + // not a stuffed zero or lead-in to another marker, looks + // like an actual marker, return it + return x; + } + // stuffed zero has x=0 now which ends the loop, meaning we go + // back to regular scan loop. + // repeated 0xff keeps trying to read the next byte of the marker. + } + } + return STBI__MARKER_none; } // decode image to YCbCr format -static int stbi__decode_jpeg_image(stbi__jpeg * j) { - int m; - for (m = 0; m < 4; m++) { - j->img_comp[m].raw_data = NULL; - j->img_comp[m].raw_coeff = NULL; - } - j->restart_interval = 0; - if (!stbi__decode_jpeg_header(j, STBI__SCAN_load)) - return 0; - m = stbi__get_marker(j); - while (!stbi__EOI(m)) { - if (stbi__SOS(m)) { - if (!stbi__process_scan_header(j)) - return 0; - if (!stbi__parse_entropy_coded_data(j)) - return 0; - if (j->marker == STBI__MARKER_none) { - j->marker = stbi__skip_jpeg_junk_at_end(j); - // if we reach eof without hitting a marker, stbi__get_marker() below will fail and we'll eventually return 0 - } +static int stbi__decode_jpeg_image(stbi__jpeg *j) +{ + int m; + for (m = 0; m < 4; m++) { + j->img_comp[m].raw_data = NULL; + j->img_comp[m].raw_coeff = NULL; + } + j->restart_interval = 0; + if (!stbi__decode_jpeg_header(j, STBI__SCAN_load)) return 0; + m = stbi__get_marker(j); + while (!stbi__EOI(m)) { + if (stbi__SOS(m)) { + if (!stbi__process_scan_header(j)) return 0; + if (!stbi__parse_entropy_coded_data(j)) return 0; + if (j->marker == STBI__MARKER_none ) { + j->marker = stbi__skip_jpeg_junk_at_end(j); + // if we reach eof without hitting a marker, stbi__get_marker() below will fail and we'll eventually return 0 + } + m = stbi__get_marker(j); + if (STBI__RESTART(m)) m = stbi__get_marker(j); - if (STBI__RESTART(m)) - m = stbi__get_marker(j); - } else if (stbi__DNL(m)) { - int Ld = stbi__get16be(j->s); - stbi__uint32 NL = stbi__get16be(j->s); - if (Ld != 4) - return stbi__err("bad DNL len", "Corrupt JPEG"); - if (NL != j->s->img_y) - return stbi__err("bad DNL height", "Corrupt JPEG"); - m = stbi__get_marker(j); - } else { - if (!stbi__process_marker(j, m)) - return 1; - m = stbi__get_marker(j); - } - } - if (j->progressive) - stbi__jpeg_finish(j); - return 1; + } else if (stbi__DNL(m)) { + int Ld = stbi__get16be(j->s); + stbi__uint32 NL = stbi__get16be(j->s); + if (Ld != 4) return stbi__err("bad DNL len", "Corrupt JPEG"); + if (NL != j->s->img_y) return stbi__err("bad DNL height", "Corrupt JPEG"); + m = stbi__get_marker(j); + } else { + if (!stbi__process_marker(j, m)) return 1; + m = stbi__get_marker(j); + } + } + if (j->progressive) + stbi__jpeg_finish(j); + return 1; } // static jfif-centered resampling (across block boundaries) -typedef stbi_uc * (*resample_row_func)(stbi_uc * out, stbi_uc * in0, stbi_uc * in1, int w, int hs); +typedef stbi_uc *(*resample_row_func)(stbi_uc *out, stbi_uc *in0, stbi_uc *in1, + int w, int hs); -#define stbi__div4(x) ((stbi_uc)((x) >> 2)) +#define stbi__div4(x) ((stbi_uc) ((x) >> 2)) -static stbi_uc * resample_row_1(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) { - STBI_NOTUSED(out); - STBI_NOTUSED(in_far); - STBI_NOTUSED(w); - STBI_NOTUSED(hs); - return in_near; +static stbi_uc *resample_row_1(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + STBI_NOTUSED(out); + STBI_NOTUSED(in_far); + STBI_NOTUSED(w); + STBI_NOTUSED(hs); + return in_near; } -static stbi_uc * stbi__resample_row_v_2(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) { - // need to generate two samples vertically for every one in input - int i; - STBI_NOTUSED(hs); - for (i = 0; i < w; ++i) - out[i] = stbi__div4(3 * in_near[i] + in_far[i] + 2); - return out; +static stbi_uc* stbi__resample_row_v_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // need to generate two samples vertically for every one in input + int i; + STBI_NOTUSED(hs); + for (i=0; i < w; ++i) + out[i] = stbi__div4(3*in_near[i] + in_far[i] + 2); + return out; } -static stbi_uc * stbi__resample_row_h_2(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) { - // need to generate two samples horizontally for every one in input - int i; - stbi_uc * input = in_near; +static stbi_uc* stbi__resample_row_h_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // need to generate two samples horizontally for every one in input + int i; + stbi_uc *input = in_near; - if (w == 1) { - // if only one sample, can't do any interpolation - out[0] = out[1] = input[0]; - return out; - } + if (w == 1) { + // if only one sample, can't do any interpolation + out[0] = out[1] = input[0]; + return out; + } - out[0] = input[0]; - out[1] = stbi__div4(input[0] * 3 + input[1] + 2); - for (i = 1; i < w - 1; ++i) { - int n = 3 * input[i] + 2; - out[i * 2 + 0] = stbi__div4(n + input[i - 1]); - out[i * 2 + 1] = stbi__div4(n + input[i + 1]); - } - out[i * 2 + 0] = stbi__div4(input[w - 2] * 3 + input[w - 1] + 2); - out[i * 2 + 1] = input[w - 1]; + out[0] = input[0]; + out[1] = stbi__div4(input[0]*3 + input[1] + 2); + for (i=1; i < w-1; ++i) { + int n = 3*input[i]+2; + out[i*2+0] = stbi__div4(n+input[i-1]); + out[i*2+1] = stbi__div4(n+input[i+1]); + } + out[i*2+0] = stbi__div4(input[w-2]*3 + input[w-1] + 2); + out[i*2+1] = input[w-1]; - STBI_NOTUSED(in_far); - STBI_NOTUSED(hs); + STBI_NOTUSED(in_far); + STBI_NOTUSED(hs); - return out; + return out; } -#define stbi__div16(x) ((stbi_uc)((x) >> 4)) +#define stbi__div16(x) ((stbi_uc) ((x) >> 4)) -static stbi_uc * stbi__resample_row_hv_2(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) { - // need to generate 2x2 samples for every one in input - int i, t0, t1; - if (w == 1) { - out[0] = out[1] = stbi__div4(3 * in_near[0] + in_far[0] + 2); - return out; - } +static stbi_uc *stbi__resample_row_hv_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // need to generate 2x2 samples for every one in input + int i,t0,t1; + if (w == 1) { + out[0] = out[1] = stbi__div4(3*in_near[0] + in_far[0] + 2); + return out; + } - t1 = 3 * in_near[0] + in_far[0]; - out[0] = stbi__div4(t1 + 2); - for (i = 1; i < w; ++i) { - t0 = t1; - t1 = 3 * in_near[i] + in_far[i]; - out[i * 2 - 1] = stbi__div16(3 * t0 + t1 + 8); - out[i * 2] = stbi__div16(3 * t1 + t0 + 8); - } - out[w * 2 - 1] = stbi__div4(t1 + 2); + t1 = 3*in_near[0] + in_far[0]; + out[0] = stbi__div4(t1+2); + for (i=1; i < w; ++i) { + t0 = t1; + t1 = 3*in_near[i]+in_far[i]; + out[i*2-1] = stbi__div16(3*t0 + t1 + 8); + out[i*2 ] = stbi__div16(3*t1 + t0 + 8); + } + out[w*2-1] = stbi__div4(t1+2); - STBI_NOTUSED(hs); + STBI_NOTUSED(hs); - return out; + return out; } #if defined(STBI_SSE2) || defined(STBI_NEON) -static stbi_uc * stbi__resample_row_hv_2_simd(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) { - // need to generate 2x2 samples for every one in input - int i = 0, t0, t1; +static stbi_uc *stbi__resample_row_hv_2_simd(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // need to generate 2x2 samples for every one in input + int i=0,t0,t1; - if (w == 1) { - out[0] = out[1] = stbi__div4(3 * in_near[0] + in_far[0] + 2); - return out; - } + if (w == 1) { + out[0] = out[1] = stbi__div4(3*in_near[0] + in_far[0] + 2); + return out; + } - t1 = 3 * in_near[0] + in_far[0]; - // process groups of 8 pixels for as long as we can. - // note we can't handle the last pixel in a row in this loop - // because we need to handle the filter boundary conditions. - for (; i < ((w - 1) & ~7); i += 8) { + t1 = 3*in_near[0] + in_far[0]; + // process groups of 8 pixels for as long as we can. + // note we can't handle the last pixel in a row in this loop + // because we need to handle the filter boundary conditions. + for (; i < ((w-1) & ~7); i += 8) { #if defined(STBI_SSE2) - // load and perform the vertical filtering pass - // this uses 3*x + y = 4*x + (y - x) - __m128i zero = _mm_setzero_si128(); - __m128i farb = _mm_loadl_epi64((__m128i *)(in_far + i)); - __m128i nearb = _mm_loadl_epi64((__m128i *)(in_near + i)); - __m128i farw = _mm_unpacklo_epi8(farb, zero); - __m128i nearw = _mm_unpacklo_epi8(nearb, zero); - __m128i diff = _mm_sub_epi16(farw, nearw); - __m128i nears = _mm_slli_epi16(nearw, 2); - __m128i curr = _mm_add_epi16(nears, diff); // current row + // load and perform the vertical filtering pass + // this uses 3*x + y = 4*x + (y - x) + __m128i zero = _mm_setzero_si128(); + __m128i farb = _mm_loadl_epi64((__m128i *) (in_far + i)); + __m128i nearb = _mm_loadl_epi64((__m128i *) (in_near + i)); + __m128i farw = _mm_unpacklo_epi8(farb, zero); + __m128i nearw = _mm_unpacklo_epi8(nearb, zero); + __m128i diff = _mm_sub_epi16(farw, nearw); + __m128i nears = _mm_slli_epi16(nearw, 2); + __m128i curr = _mm_add_epi16(nears, diff); // current row - // horizontal filter works the same based on shifted vers of current - // row. "prev" is current row shifted right by 1 pixel; we need to - // insert the previous pixel value (from t1). - // "next" is current row shifted left by 1 pixel, with first pixel - // of next block of 8 pixels added in. - __m128i prv0 = _mm_slli_si128(curr, 2); - __m128i nxt0 = _mm_srli_si128(curr, 2); - __m128i prev = _mm_insert_epi16(prv0, t1, 0); - __m128i next = _mm_insert_epi16(nxt0, 3 * in_near[i + 8] + in_far[i + 8], 7); + // horizontal filter works the same based on shifted vers of current + // row. "prev" is current row shifted right by 1 pixel; we need to + // insert the previous pixel value (from t1). + // "next" is current row shifted left by 1 pixel, with first pixel + // of next block of 8 pixels added in. + __m128i prv0 = _mm_slli_si128(curr, 2); + __m128i nxt0 = _mm_srli_si128(curr, 2); + __m128i prev = _mm_insert_epi16(prv0, t1, 0); + __m128i next = _mm_insert_epi16(nxt0, 3*in_near[i+8] + in_far[i+8], 7); - // horizontal filter, polyphase implementation since it's convenient: - // even pixels = 3*cur + prev = cur*4 + (prev - cur) - // odd pixels = 3*cur + next = cur*4 + (next - cur) - // note the shared term. - __m128i bias = _mm_set1_epi16(8); - __m128i curs = _mm_slli_epi16(curr, 2); - __m128i prvd = _mm_sub_epi16(prev, curr); - __m128i nxtd = _mm_sub_epi16(next, curr); - __m128i curb = _mm_add_epi16(curs, bias); - __m128i even = _mm_add_epi16(prvd, curb); - __m128i odd = _mm_add_epi16(nxtd, curb); + // horizontal filter, polyphase implementation since it's convenient: + // even pixels = 3*cur + prev = cur*4 + (prev - cur) + // odd pixels = 3*cur + next = cur*4 + (next - cur) + // note the shared term. + __m128i bias = _mm_set1_epi16(8); + __m128i curs = _mm_slli_epi16(curr, 2); + __m128i prvd = _mm_sub_epi16(prev, curr); + __m128i nxtd = _mm_sub_epi16(next, curr); + __m128i curb = _mm_add_epi16(curs, bias); + __m128i even = _mm_add_epi16(prvd, curb); + __m128i odd = _mm_add_epi16(nxtd, curb); - // interleave even and odd pixels, then undo scaling. - __m128i int0 = _mm_unpacklo_epi16(even, odd); - __m128i int1 = _mm_unpackhi_epi16(even, odd); - __m128i de0 = _mm_srli_epi16(int0, 4); - __m128i de1 = _mm_srli_epi16(int1, 4); + // interleave even and odd pixels, then undo scaling. + __m128i int0 = _mm_unpacklo_epi16(even, odd); + __m128i int1 = _mm_unpackhi_epi16(even, odd); + __m128i de0 = _mm_srli_epi16(int0, 4); + __m128i de1 = _mm_srli_epi16(int1, 4); - // pack and write output - __m128i outv = _mm_packus_epi16(de0, de1); - _mm_storeu_si128((__m128i *)(out + i * 2), outv); + // pack and write output + __m128i outv = _mm_packus_epi16(de0, de1); + _mm_storeu_si128((__m128i *) (out + i*2), outv); #elif defined(STBI_NEON) - // load and perform the vertical filtering pass - // this uses 3*x + y = 4*x + (y - x) - uint8x8_t farb = vld1_u8(in_far + i); - uint8x8_t nearb = vld1_u8(in_near + i); - int16x8_t diff = vreinterpretq_s16_u16(vsubl_u8(farb, nearb)); - int16x8_t nears = vreinterpretq_s16_u16(vshll_n_u8(nearb, 2)); - int16x8_t curr = vaddq_s16(nears, diff); // current row + // load and perform the vertical filtering pass + // this uses 3*x + y = 4*x + (y - x) + uint8x8_t farb = vld1_u8(in_far + i); + uint8x8_t nearb = vld1_u8(in_near + i); + int16x8_t diff = vreinterpretq_s16_u16(vsubl_u8(farb, nearb)); + int16x8_t nears = vreinterpretq_s16_u16(vshll_n_u8(nearb, 2)); + int16x8_t curr = vaddq_s16(nears, diff); // current row - // horizontal filter works the same based on shifted vers of current - // row. "prev" is current row shifted right by 1 pixel; we need to - // insert the previous pixel value (from t1). - // "next" is current row shifted left by 1 pixel, with first pixel - // of next block of 8 pixels added in. - int16x8_t prv0 = vextq_s16(curr, curr, 7); - int16x8_t nxt0 = vextq_s16(curr, curr, 1); - int16x8_t prev = vsetq_lane_s16(t1, prv0, 0); - int16x8_t next = vsetq_lane_s16(3 * in_near[i + 8] + in_far[i + 8], nxt0, 7); + // horizontal filter works the same based on shifted vers of current + // row. "prev" is current row shifted right by 1 pixel; we need to + // insert the previous pixel value (from t1). + // "next" is current row shifted left by 1 pixel, with first pixel + // of next block of 8 pixels added in. + int16x8_t prv0 = vextq_s16(curr, curr, 7); + int16x8_t nxt0 = vextq_s16(curr, curr, 1); + int16x8_t prev = vsetq_lane_s16(t1, prv0, 0); + int16x8_t next = vsetq_lane_s16(3*in_near[i+8] + in_far[i+8], nxt0, 7); - // horizontal filter, polyphase implementation since it's convenient: - // even pixels = 3*cur + prev = cur*4 + (prev - cur) - // odd pixels = 3*cur + next = cur*4 + (next - cur) - // note the shared term. - int16x8_t curs = vshlq_n_s16(curr, 2); - int16x8_t prvd = vsubq_s16(prev, curr); - int16x8_t nxtd = vsubq_s16(next, curr); - int16x8_t even = vaddq_s16(curs, prvd); - int16x8_t odd = vaddq_s16(curs, nxtd); + // horizontal filter, polyphase implementation since it's convenient: + // even pixels = 3*cur + prev = cur*4 + (prev - cur) + // odd pixels = 3*cur + next = cur*4 + (next - cur) + // note the shared term. + int16x8_t curs = vshlq_n_s16(curr, 2); + int16x8_t prvd = vsubq_s16(prev, curr); + int16x8_t nxtd = vsubq_s16(next, curr); + int16x8_t even = vaddq_s16(curs, prvd); + int16x8_t odd = vaddq_s16(curs, nxtd); - // undo scaling and round, then store with even/odd phases interleaved - uint8x8x2_t o; - o.val[0] = vqrshrun_n_s16(even, 4); - o.val[1] = vqrshrun_n_s16(odd, 4); - vst2_u8(out + i * 2, o); + // undo scaling and round, then store with even/odd phases interleaved + uint8x8x2_t o; + o.val[0] = vqrshrun_n_s16(even, 4); + o.val[1] = vqrshrun_n_s16(odd, 4); + vst2_u8(out + i*2, o); #endif - // "previous" value for next iter - t1 = 3 * in_near[i + 7] + in_far[i + 7]; - } + // "previous" value for next iter + t1 = 3*in_near[i+7] + in_far[i+7]; + } - t0 = t1; - t1 = 3 * in_near[i] + in_far[i]; - out[i * 2] = stbi__div16(3 * t1 + t0 + 8); + t0 = t1; + t1 = 3*in_near[i] + in_far[i]; + out[i*2] = stbi__div16(3*t1 + t0 + 8); - for (++i; i < w; ++i) { - t0 = t1; - t1 = 3 * in_near[i] + in_far[i]; - out[i * 2 - 1] = stbi__div16(3 * t0 + t1 + 8); - out[i * 2] = stbi__div16(3 * t1 + t0 + 8); - } - out[w * 2 - 1] = stbi__div4(t1 + 2); + for (++i; i < w; ++i) { + t0 = t1; + t1 = 3*in_near[i]+in_far[i]; + out[i*2-1] = stbi__div16(3*t0 + t1 + 8); + out[i*2 ] = stbi__div16(3*t1 + t0 + 8); + } + out[w*2-1] = stbi__div4(t1+2); - STBI_NOTUSED(hs); + STBI_NOTUSED(hs); - return out; + return out; } #endif -static stbi_uc * stbi__resample_row_generic(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) { - // resample with nearest-neighbor - int i, j; - STBI_NOTUSED(in_far); - for (i = 0; i < w; ++i) - for (j = 0; j < hs; ++j) - out[i * hs + j] = in_near[i]; - return out; +static stbi_uc *stbi__resample_row_generic(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // resample with nearest-neighbor + int i,j; + STBI_NOTUSED(in_far); + for (i=0; i < w; ++i) + for (j=0; j < hs; ++j) + out[i*hs+j] = in_near[i]; + return out; } // this is a reduced-precision calculation of YCbCr-to-RGB introduced // to make sure the code produces the same results in both SIMD and scalar -#define stbi__float2fixed(x) (((int)((x)*4096.0f + 0.5f)) << 8) -static void stbi__YCbCr_to_RGB_row(stbi_uc * out, const stbi_uc * y, const stbi_uc * pcb, const stbi_uc * pcr, int count, - int step) { - int i; - for (i = 0; i < count; ++i) { - int y_fixed = (y[i] << 20) + (1 << 19); // rounding - int r, g, b; - int cr = pcr[i] - 128; - int cb = pcb[i] - 128; - r = y_fixed + cr * stbi__float2fixed(1.40200f); - g = y_fixed + (cr * -stbi__float2fixed(0.71414f)) + ((cb * -stbi__float2fixed(0.34414f)) & 0xffff0000); - b = y_fixed + cb * stbi__float2fixed(1.77200f); - r >>= 20; - g >>= 20; - b >>= 20; - if ((unsigned)r > 255) { - if (r < 0) - r = 0; - else - r = 255; - } - if ((unsigned)g > 255) { - if (g < 0) - g = 0; - else - g = 255; - } - if ((unsigned)b > 255) { - if (b < 0) - b = 0; - else - b = 255; - } - out[0] = (stbi_uc)r; - out[1] = (stbi_uc)g; - out[2] = (stbi_uc)b; - out[3] = 255; - out += step; - } +#define stbi__float2fixed(x) (((int) ((x) * 4096.0f + 0.5f)) << 8) +static void stbi__YCbCr_to_RGB_row(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step) +{ + int i; + for (i=0; i < count; ++i) { + int y_fixed = (y[i] << 20) + (1<<19); // rounding + int r,g,b; + int cr = pcr[i] - 128; + int cb = pcb[i] - 128; + r = y_fixed + cr* stbi__float2fixed(1.40200f); + g = y_fixed + (cr*-stbi__float2fixed(0.71414f)) + ((cb*-stbi__float2fixed(0.34414f)) & 0xffff0000); + b = y_fixed + cb* stbi__float2fixed(1.77200f); + r >>= 20; + g >>= 20; + b >>= 20; + if ((unsigned) r > 255) { if (r < 0) r = 0; else r = 255; } + if ((unsigned) g > 255) { if (g < 0) g = 0; else g = 255; } + if ((unsigned) b > 255) { if (b < 0) b = 0; else b = 255; } + out[0] = (stbi_uc)r; + out[1] = (stbi_uc)g; + out[2] = (stbi_uc)b; + out[3] = 255; + out += step; + } } #if defined(STBI_SSE2) || defined(STBI_NEON) -static void stbi__YCbCr_to_RGB_simd(stbi_uc * out, stbi_uc const * y, stbi_uc const * pcb, stbi_uc const * pcr, int count, - int step) { - int i = 0; +static void stbi__YCbCr_to_RGB_simd(stbi_uc *out, stbi_uc const *y, stbi_uc const *pcb, stbi_uc const *pcr, int count, int step) +{ + int i = 0; #ifdef STBI_SSE2 - // step == 3 is pretty ugly on the final interleave, and i'm not convinced - // it's useful in practice (you wouldn't use it for textures, for example). - // so just accelerate step == 4 case. - if (step == 4) { - // this is a fairly straightforward implementation and not super-optimized. - __m128i signflip = _mm_set1_epi8(-0x80); - __m128i cr_const0 = _mm_set1_epi16((short)(1.40200f * 4096.0f + 0.5f)); - __m128i cr_const1 = _mm_set1_epi16(-(short)(0.71414f * 4096.0f + 0.5f)); - __m128i cb_const0 = _mm_set1_epi16(-(short)(0.34414f * 4096.0f + 0.5f)); - __m128i cb_const1 = _mm_set1_epi16((short)(1.77200f * 4096.0f + 0.5f)); - __m128i y_bias = _mm_set1_epi8((char)(unsigned char)128); - __m128i xw = _mm_set1_epi16(255); // alpha channel + // step == 3 is pretty ugly on the final interleave, and i'm not convinced + // it's useful in practice (you wouldn't use it for textures, for example). + // so just accelerate step == 4 case. + if (step == 4) { + // this is a fairly straightforward implementation and not super-optimized. + __m128i signflip = _mm_set1_epi8(-0x80); + __m128i cr_const0 = _mm_set1_epi16( (short) ( 1.40200f*4096.0f+0.5f)); + __m128i cr_const1 = _mm_set1_epi16( - (short) ( 0.71414f*4096.0f+0.5f)); + __m128i cb_const0 = _mm_set1_epi16( - (short) ( 0.34414f*4096.0f+0.5f)); + __m128i cb_const1 = _mm_set1_epi16( (short) ( 1.77200f*4096.0f+0.5f)); + __m128i y_bias = _mm_set1_epi8((char) (unsigned char) 128); + __m128i xw = _mm_set1_epi16(255); // alpha channel - for (; i + 7 < count; i += 8) { - // load - __m128i y_bytes = _mm_loadl_epi64((__m128i *)(y + i)); - __m128i cr_bytes = _mm_loadl_epi64((__m128i *)(pcr + i)); - __m128i cb_bytes = _mm_loadl_epi64((__m128i *)(pcb + i)); - __m128i cr_biased = _mm_xor_si128(cr_bytes, signflip); // -128 - __m128i cb_biased = _mm_xor_si128(cb_bytes, signflip); // -128 + for (; i+7 < count; i += 8) { + // load + __m128i y_bytes = _mm_loadl_epi64((__m128i *) (y+i)); + __m128i cr_bytes = _mm_loadl_epi64((__m128i *) (pcr+i)); + __m128i cb_bytes = _mm_loadl_epi64((__m128i *) (pcb+i)); + __m128i cr_biased = _mm_xor_si128(cr_bytes, signflip); // -128 + __m128i cb_biased = _mm_xor_si128(cb_bytes, signflip); // -128 - // unpack to short (and left-shift cr, cb by 8) - __m128i yw = _mm_unpacklo_epi8(y_bias, y_bytes); - __m128i crw = _mm_unpacklo_epi8(_mm_setzero_si128(), cr_biased); - __m128i cbw = _mm_unpacklo_epi8(_mm_setzero_si128(), cb_biased); + // unpack to short (and left-shift cr, cb by 8) + __m128i yw = _mm_unpacklo_epi8(y_bias, y_bytes); + __m128i crw = _mm_unpacklo_epi8(_mm_setzero_si128(), cr_biased); + __m128i cbw = _mm_unpacklo_epi8(_mm_setzero_si128(), cb_biased); - // color transform - __m128i yws = _mm_srli_epi16(yw, 4); - __m128i cr0 = _mm_mulhi_epi16(cr_const0, crw); - __m128i cb0 = _mm_mulhi_epi16(cb_const0, cbw); - __m128i cb1 = _mm_mulhi_epi16(cbw, cb_const1); - __m128i cr1 = _mm_mulhi_epi16(crw, cr_const1); - __m128i rws = _mm_add_epi16(cr0, yws); - __m128i gwt = _mm_add_epi16(cb0, yws); - __m128i bws = _mm_add_epi16(yws, cb1); - __m128i gws = _mm_add_epi16(gwt, cr1); + // color transform + __m128i yws = _mm_srli_epi16(yw, 4); + __m128i cr0 = _mm_mulhi_epi16(cr_const0, crw); + __m128i cb0 = _mm_mulhi_epi16(cb_const0, cbw); + __m128i cb1 = _mm_mulhi_epi16(cbw, cb_const1); + __m128i cr1 = _mm_mulhi_epi16(crw, cr_const1); + __m128i rws = _mm_add_epi16(cr0, yws); + __m128i gwt = _mm_add_epi16(cb0, yws); + __m128i bws = _mm_add_epi16(yws, cb1); + __m128i gws = _mm_add_epi16(gwt, cr1); - // descale - __m128i rw = _mm_srai_epi16(rws, 4); - __m128i bw = _mm_srai_epi16(bws, 4); - __m128i gw = _mm_srai_epi16(gws, 4); + // descale + __m128i rw = _mm_srai_epi16(rws, 4); + __m128i bw = _mm_srai_epi16(bws, 4); + __m128i gw = _mm_srai_epi16(gws, 4); - // back to byte, set up for transpose - __m128i brb = _mm_packus_epi16(rw, bw); - __m128i gxb = _mm_packus_epi16(gw, xw); + // back to byte, set up for transpose + __m128i brb = _mm_packus_epi16(rw, bw); + __m128i gxb = _mm_packus_epi16(gw, xw); - // transpose to interleave channels - __m128i t0 = _mm_unpacklo_epi8(brb, gxb); - __m128i t1 = _mm_unpackhi_epi8(brb, gxb); - __m128i o0 = _mm_unpacklo_epi16(t0, t1); - __m128i o1 = _mm_unpackhi_epi16(t0, t1); + // transpose to interleave channels + __m128i t0 = _mm_unpacklo_epi8(brb, gxb); + __m128i t1 = _mm_unpackhi_epi8(brb, gxb); + __m128i o0 = _mm_unpacklo_epi16(t0, t1); + __m128i o1 = _mm_unpackhi_epi16(t0, t1); - // store - _mm_storeu_si128((__m128i *)(out + 0), o0); - _mm_storeu_si128((__m128i *)(out + 16), o1); - out += 32; - } - } + // store + _mm_storeu_si128((__m128i *) (out + 0), o0); + _mm_storeu_si128((__m128i *) (out + 16), o1); + out += 32; + } + } #endif #ifdef STBI_NEON - // in this version, step=3 support would be easy to add. but is there demand? - if (step == 4) { - // this is a fairly straightforward implementation and not super-optimized. - uint8x8_t signflip = vdup_n_u8(0x80); - int16x8_t cr_const0 = vdupq_n_s16((short)(1.40200f * 4096.0f + 0.5f)); - int16x8_t cr_const1 = vdupq_n_s16(-(short)(0.71414f * 4096.0f + 0.5f)); - int16x8_t cb_const0 = vdupq_n_s16(-(short)(0.34414f * 4096.0f + 0.5f)); - int16x8_t cb_const1 = vdupq_n_s16((short)(1.77200f * 4096.0f + 0.5f)); + // in this version, step=3 support would be easy to add. but is there demand? + if (step == 4) { + // this is a fairly straightforward implementation and not super-optimized. + uint8x8_t signflip = vdup_n_u8(0x80); + int16x8_t cr_const0 = vdupq_n_s16( (short) ( 1.40200f*4096.0f+0.5f)); + int16x8_t cr_const1 = vdupq_n_s16( - (short) ( 0.71414f*4096.0f+0.5f)); + int16x8_t cb_const0 = vdupq_n_s16( - (short) ( 0.34414f*4096.0f+0.5f)); + int16x8_t cb_const1 = vdupq_n_s16( (short) ( 1.77200f*4096.0f+0.5f)); - for (; i + 7 < count; i += 8) { - // load - uint8x8_t y_bytes = vld1_u8(y + i); - uint8x8_t cr_bytes = vld1_u8(pcr + i); - uint8x8_t cb_bytes = vld1_u8(pcb + i); - int8x8_t cr_biased = vreinterpret_s8_u8(vsub_u8(cr_bytes, signflip)); - int8x8_t cb_biased = vreinterpret_s8_u8(vsub_u8(cb_bytes, signflip)); + for (; i+7 < count; i += 8) { + // load + uint8x8_t y_bytes = vld1_u8(y + i); + uint8x8_t cr_bytes = vld1_u8(pcr + i); + uint8x8_t cb_bytes = vld1_u8(pcb + i); + int8x8_t cr_biased = vreinterpret_s8_u8(vsub_u8(cr_bytes, signflip)); + int8x8_t cb_biased = vreinterpret_s8_u8(vsub_u8(cb_bytes, signflip)); - // expand to s16 - int16x8_t yws = vreinterpretq_s16_u16(vshll_n_u8(y_bytes, 4)); - int16x8_t crw = vshll_n_s8(cr_biased, 7); - int16x8_t cbw = vshll_n_s8(cb_biased, 7); + // expand to s16 + int16x8_t yws = vreinterpretq_s16_u16(vshll_n_u8(y_bytes, 4)); + int16x8_t crw = vshll_n_s8(cr_biased, 7); + int16x8_t cbw = vshll_n_s8(cb_biased, 7); - // color transform - int16x8_t cr0 = vqdmulhq_s16(crw, cr_const0); - int16x8_t cb0 = vqdmulhq_s16(cbw, cb_const0); - int16x8_t cr1 = vqdmulhq_s16(crw, cr_const1); - int16x8_t cb1 = vqdmulhq_s16(cbw, cb_const1); - int16x8_t rws = vaddq_s16(yws, cr0); - int16x8_t gws = vaddq_s16(vaddq_s16(yws, cb0), cr1); - int16x8_t bws = vaddq_s16(yws, cb1); + // color transform + int16x8_t cr0 = vqdmulhq_s16(crw, cr_const0); + int16x8_t cb0 = vqdmulhq_s16(cbw, cb_const0); + int16x8_t cr1 = vqdmulhq_s16(crw, cr_const1); + int16x8_t cb1 = vqdmulhq_s16(cbw, cb_const1); + int16x8_t rws = vaddq_s16(yws, cr0); + int16x8_t gws = vaddq_s16(vaddq_s16(yws, cb0), cr1); + int16x8_t bws = vaddq_s16(yws, cb1); - // undo scaling, round, convert to byte - uint8x8x4_t o; - o.val[0] = vqrshrun_n_s16(rws, 4); - o.val[1] = vqrshrun_n_s16(gws, 4); - o.val[2] = vqrshrun_n_s16(bws, 4); - o.val[3] = vdup_n_u8(255); + // undo scaling, round, convert to byte + uint8x8x4_t o; + o.val[0] = vqrshrun_n_s16(rws, 4); + o.val[1] = vqrshrun_n_s16(gws, 4); + o.val[2] = vqrshrun_n_s16(bws, 4); + o.val[3] = vdup_n_u8(255); - // store, interleaving r/g/b/a - vst4_u8(out, o); - out += 8 * 4; - } - } + // store, interleaving r/g/b/a + vst4_u8(out, o); + out += 8*4; + } + } #endif - for (; i < count; ++i) { - int y_fixed = (y[i] << 20) + (1 << 19); // rounding - int r, g, b; - int cr = pcr[i] - 128; - int cb = pcb[i] - 128; - r = y_fixed + cr * stbi__float2fixed(1.40200f); - g = y_fixed + cr * -stbi__float2fixed(0.71414f) + ((cb * -stbi__float2fixed(0.34414f)) & 0xffff0000); - b = y_fixed + cb * stbi__float2fixed(1.77200f); - r >>= 20; - g >>= 20; - b >>= 20; - if ((unsigned)r > 255) { - if (r < 0) - r = 0; - else - r = 255; - } - if ((unsigned)g > 255) { - if (g < 0) - g = 0; - else - g = 255; - } - if ((unsigned)b > 255) { - if (b < 0) - b = 0; - else - b = 255; - } - out[0] = (stbi_uc)r; - out[1] = (stbi_uc)g; - out[2] = (stbi_uc)b; - out[3] = 255; - out += step; - } + for (; i < count; ++i) { + int y_fixed = (y[i] << 20) + (1<<19); // rounding + int r,g,b; + int cr = pcr[i] - 128; + int cb = pcb[i] - 128; + r = y_fixed + cr* stbi__float2fixed(1.40200f); + g = y_fixed + cr*-stbi__float2fixed(0.71414f) + ((cb*-stbi__float2fixed(0.34414f)) & 0xffff0000); + b = y_fixed + cb* stbi__float2fixed(1.77200f); + r >>= 20; + g >>= 20; + b >>= 20; + if ((unsigned) r > 255) { if (r < 0) r = 0; else r = 255; } + if ((unsigned) g > 255) { if (g < 0) g = 0; else g = 255; } + if ((unsigned) b > 255) { if (b < 0) b = 0; else b = 255; } + out[0] = (stbi_uc)r; + out[1] = (stbi_uc)g; + out[2] = (stbi_uc)b; + out[3] = 255; + out += step; + } } #endif // set up the kernels -static void stbi__setup_jpeg(stbi__jpeg * j) { - j->idct_block_kernel = stbi__idct_block; - j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_row; - j->resample_row_hv_2_kernel = stbi__resample_row_hv_2; +static void stbi__setup_jpeg(stbi__jpeg *j) +{ + j->idct_block_kernel = stbi__idct_block; + j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_row; + j->resample_row_hv_2_kernel = stbi__resample_row_hv_2; #ifdef STBI_SSE2 - if (stbi__sse2_available()) { - j->idct_block_kernel = stbi__idct_simd; - j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd; - j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd; - } + if (stbi__sse2_available()) { + j->idct_block_kernel = stbi__idct_simd; + j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd; + j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd; + } #endif #ifdef STBI_NEON - j->idct_block_kernel = stbi__idct_simd; - j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd; - j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd; + j->idct_block_kernel = stbi__idct_simd; + j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd; + j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd; #endif } // clean up the temporary component buffers -static void stbi__cleanup_jpeg(stbi__jpeg * j) { stbi__free_jpeg_components(j, j->s->img_n, 0); } +static void stbi__cleanup_jpeg(stbi__jpeg *j) +{ + stbi__free_jpeg_components(j, j->s->img_n, 0); +} -typedef struct { - resample_row_func resample; - stbi_uc *line0, *line1; - int hs, vs; // expansion factor in each axis - int w_lores; // horizontal pixels pre-expansion - int ystep; // how far through vertical expansion we are - int ypos; // which pre-expansion row we're on +typedef struct +{ + resample_row_func resample; + stbi_uc *line0,*line1; + int hs,vs; // expansion factor in each axis + int w_lores; // horizontal pixels pre-expansion + int ystep; // how far through vertical expansion we are + int ypos; // which pre-expansion row we're on } stbi__resample; // fast 0..255 * 0..255 => 0..255 rounded multiplication -static stbi_uc stbi__blinn_8x8(stbi_uc x, stbi_uc y) { - unsigned int t = x * y + 128; - return (stbi_uc)((t + (t >> 8)) >> 8); +static stbi_uc stbi__blinn_8x8(stbi_uc x, stbi_uc y) +{ + unsigned int t = x*y + 128; + return (stbi_uc) ((t + (t >>8)) >> 8); } -static stbi_uc * load_jpeg_image(stbi__jpeg * z, int * out_x, int * out_y, int * comp, int req_comp) { - int n, decode_n, is_rgb; - z->s->img_n = 0; // make stbi__cleanup_jpeg safe +static stbi_uc *load_jpeg_image(stbi__jpeg *z, int *out_x, int *out_y, int *comp, int req_comp) +{ + int n, decode_n, is_rgb; + z->s->img_n = 0; // make stbi__cleanup_jpeg safe - // validate req_comp - if (req_comp < 0 || req_comp > 4) - return stbi__errpuc("bad req_comp", "Internal error"); + // validate req_comp + if (req_comp < 0 || req_comp > 4) return stbi__errpuc("bad req_comp", "Internal error"); - // load a jpeg image from whichever source, but leave in YCbCr format - if (!stbi__decode_jpeg_image(z)) { - stbi__cleanup_jpeg(z); - return NULL; - } + // load a jpeg image from whichever source, but leave in YCbCr format + if (!stbi__decode_jpeg_image(z)) { stbi__cleanup_jpeg(z); return NULL; } - // determine actual number of components to generate - n = req_comp ? req_comp : z->s->img_n >= 3 ? 3 : 1; + // determine actual number of components to generate + n = req_comp ? req_comp : z->s->img_n >= 3 ? 3 : 1; - is_rgb = z->s->img_n == 3 && (z->rgb == 3 || (z->app14_color_transform == 0 && !z->jfif)); + is_rgb = z->s->img_n == 3 && (z->rgb == 3 || (z->app14_color_transform == 0 && !z->jfif)); - if (z->s->img_n == 3 && n < 3 && !is_rgb) - decode_n = 1; - else - decode_n = z->s->img_n; + if (z->s->img_n == 3 && n < 3 && !is_rgb) + decode_n = 1; + else + decode_n = z->s->img_n; - // nothing to do if no components requested; check this now to avoid - // accessing uninitialized coutput[0] later - if (decode_n <= 0) { - stbi__cleanup_jpeg(z); - return NULL; - } + // nothing to do if no components requested; check this now to avoid + // accessing uninitialized coutput[0] later + if (decode_n <= 0) { stbi__cleanup_jpeg(z); return NULL; } - // resample and color-convert - { - int k; - unsigned int i, j; - stbi_uc * output; - stbi_uc * coutput[4] = {NULL, NULL, NULL, NULL}; + // resample and color-convert + { + int k; + unsigned int i,j; + stbi_uc *output; + stbi_uc *coutput[4] = { NULL, NULL, NULL, NULL }; - stbi__resample res_comp[4]; + stbi__resample res_comp[4]; - for (k = 0; k < decode_n; ++k) { - stbi__resample * r = &res_comp[k]; + for (k=0; k < decode_n; ++k) { + stbi__resample *r = &res_comp[k]; - // allocate line buffer big enough for upsampling off the edges - // with upsample factor of 4 - z->img_comp[k].linebuf = (stbi_uc *)stbi__malloc(z->s->img_x + 3); - if (!z->img_comp[k].linebuf) { - stbi__cleanup_jpeg(z); - return stbi__errpuc("outofmem", "Out of memory"); + // allocate line buffer big enough for upsampling off the edges + // with upsample factor of 4 + z->img_comp[k].linebuf = (stbi_uc *) stbi__malloc(z->s->img_x + 3); + if (!z->img_comp[k].linebuf) { stbi__cleanup_jpeg(z); return stbi__errpuc("outofmem", "Out of memory"); } + + r->hs = z->img_h_max / z->img_comp[k].h; + r->vs = z->img_v_max / z->img_comp[k].v; + r->ystep = r->vs >> 1; + r->w_lores = (z->s->img_x + r->hs-1) / r->hs; + r->ypos = 0; + r->line0 = r->line1 = z->img_comp[k].data; + + if (r->hs == 1 && r->vs == 1) r->resample = resample_row_1; + else if (r->hs == 1 && r->vs == 2) r->resample = stbi__resample_row_v_2; + else if (r->hs == 2 && r->vs == 1) r->resample = stbi__resample_row_h_2; + else if (r->hs == 2 && r->vs == 2) r->resample = z->resample_row_hv_2_kernel; + else r->resample = stbi__resample_row_generic; + } + + // can't error after this so, this is safe + output = (stbi_uc *) stbi__malloc_mad3(n, z->s->img_x, z->s->img_y, 1); + if (!output) { stbi__cleanup_jpeg(z); return stbi__errpuc("outofmem", "Out of memory"); } + + // now go ahead and resample + for (j=0; j < z->s->img_y; ++j) { + stbi_uc *out = output + n * z->s->img_x * j; + for (k=0; k < decode_n; ++k) { + stbi__resample *r = &res_comp[k]; + int y_bot = r->ystep >= (r->vs >> 1); + coutput[k] = r->resample(z->img_comp[k].linebuf, + y_bot ? r->line1 : r->line0, + y_bot ? r->line0 : r->line1, + r->w_lores, r->hs); + if (++r->ystep >= r->vs) { + r->ystep = 0; + r->line0 = r->line1; + if (++r->ypos < z->img_comp[k].y) + r->line1 += z->img_comp[k].w2; } - - r->hs = z->img_h_max / z->img_comp[k].h; - r->vs = z->img_v_max / z->img_comp[k].v; - r->ystep = r->vs >> 1; - r->w_lores = (z->s->img_x + r->hs - 1) / r->hs; - r->ypos = 0; - r->line0 = r->line1 = z->img_comp[k].data; - - if (r->hs == 1 && r->vs == 1) - r->resample = resample_row_1; - else if (r->hs == 1 && r->vs == 2) - r->resample = stbi__resample_row_v_2; - else if (r->hs == 2 && r->vs == 1) - r->resample = stbi__resample_row_h_2; - else if (r->hs == 2 && r->vs == 2) - r->resample = z->resample_row_hv_2_kernel; - else - r->resample = stbi__resample_row_generic; - } - - // can't error after this so, this is safe - output = (stbi_uc *)stbi__malloc_mad3(n, z->s->img_x, z->s->img_y, 1); - if (!output) { - stbi__cleanup_jpeg(z); - return stbi__errpuc("outofmem", "Out of memory"); - } - - // now go ahead and resample - for (j = 0; j < z->s->img_y; ++j) { - stbi_uc * out = output + n * z->s->img_x * j; - for (k = 0; k < decode_n; ++k) { - stbi__resample * r = &res_comp[k]; - int y_bot = r->ystep >= (r->vs >> 1); - coutput[k] = r->resample(z->img_comp[k].linebuf, y_bot ? r->line1 : r->line0, y_bot ? r->line0 : r->line1, - r->w_lores, r->hs); - if (++r->ystep >= r->vs) { - r->ystep = 0; - r->line0 = r->line1; - if (++r->ypos < z->img_comp[k].y) - r->line1 += z->img_comp[k].w2; - } - } - if (n >= 3) { - stbi_uc * y = coutput[0]; - if (z->s->img_n == 3) { - if (is_rgb) { - for (i = 0; i < z->s->img_x; ++i) { - out[0] = y[i]; - out[1] = coutput[1][i]; - out[2] = coutput[2][i]; - out[3] = 255; - out += n; - } - } else { - z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); - } - } else if (z->s->img_n == 4) { - if (z->app14_color_transform == 0) { // CMYK - for (i = 0; i < z->s->img_x; ++i) { - stbi_uc m = coutput[3][i]; - out[0] = stbi__blinn_8x8(coutput[0][i], m); - out[1] = stbi__blinn_8x8(coutput[1][i], m); - out[2] = stbi__blinn_8x8(coutput[2][i], m); - out[3] = 255; - out += n; - } - } else if (z->app14_color_transform == 2) { // YCCK - z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); - for (i = 0; i < z->s->img_x; ++i) { - stbi_uc m = coutput[3][i]; - out[0] = stbi__blinn_8x8(255 - out[0], m); - out[1] = stbi__blinn_8x8(255 - out[1], m); - out[2] = stbi__blinn_8x8(255 - out[2], m); - out += n; - } - } else { // YCbCr + alpha? Ignore the fourth channel for now - z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); - } - } else - for (i = 0; i < z->s->img_x; ++i) { - out[0] = out[1] = out[2] = y[i]; - out[3] = 255; // not used if n==3 - out += n; - } + } + if (n >= 3) { + stbi_uc *y = coutput[0]; + if (z->s->img_n == 3) { + if (is_rgb) { + for (i=0; i < z->s->img_x; ++i) { + out[0] = y[i]; + out[1] = coutput[1][i]; + out[2] = coutput[2][i]; + out[3] = 255; + out += n; + } + } else { + z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); + } + } else if (z->s->img_n == 4) { + if (z->app14_color_transform == 0) { // CMYK + for (i=0; i < z->s->img_x; ++i) { + stbi_uc m = coutput[3][i]; + out[0] = stbi__blinn_8x8(coutput[0][i], m); + out[1] = stbi__blinn_8x8(coutput[1][i], m); + out[2] = stbi__blinn_8x8(coutput[2][i], m); + out[3] = 255; + out += n; + } + } else if (z->app14_color_transform == 2) { // YCCK + z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); + for (i=0; i < z->s->img_x; ++i) { + stbi_uc m = coutput[3][i]; + out[0] = stbi__blinn_8x8(255 - out[0], m); + out[1] = stbi__blinn_8x8(255 - out[1], m); + out[2] = stbi__blinn_8x8(255 - out[2], m); + out += n; + } + } else { // YCbCr + alpha? Ignore the fourth channel for now + z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); + } + } else + for (i=0; i < z->s->img_x; ++i) { + out[0] = out[1] = out[2] = y[i]; + out[3] = 255; // not used if n==3 + out += n; + } + } else { + if (is_rgb) { + if (n == 1) + for (i=0; i < z->s->img_x; ++i) + *out++ = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]); + else { + for (i=0; i < z->s->img_x; ++i, out += 2) { + out[0] = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]); + out[1] = 255; + } + } + } else if (z->s->img_n == 4 && z->app14_color_transform == 0) { + for (i=0; i < z->s->img_x; ++i) { + stbi_uc m = coutput[3][i]; + stbi_uc r = stbi__blinn_8x8(coutput[0][i], m); + stbi_uc g = stbi__blinn_8x8(coutput[1][i], m); + stbi_uc b = stbi__blinn_8x8(coutput[2][i], m); + out[0] = stbi__compute_y(r, g, b); + out[1] = 255; + out += n; + } + } else if (z->s->img_n == 4 && z->app14_color_transform == 2) { + for (i=0; i < z->s->img_x; ++i) { + out[0] = stbi__blinn_8x8(255 - coutput[0][i], coutput[3][i]); + out[1] = 255; + out += n; + } } else { - if (is_rgb) { - if (n == 1) - for (i = 0; i < z->s->img_x; ++i) - *out++ = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]); - else { - for (i = 0; i < z->s->img_x; ++i, out += 2) { - out[0] = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]); - out[1] = 255; - } - } - } else if (z->s->img_n == 4 && z->app14_color_transform == 0) { - for (i = 0; i < z->s->img_x; ++i) { - stbi_uc m = coutput[3][i]; - stbi_uc r = stbi__blinn_8x8(coutput[0][i], m); - stbi_uc g = stbi__blinn_8x8(coutput[1][i], m); - stbi_uc b = stbi__blinn_8x8(coutput[2][i], m); - out[0] = stbi__compute_y(r, g, b); - out[1] = 255; - out += n; - } - } else if (z->s->img_n == 4 && z->app14_color_transform == 2) { - for (i = 0; i < z->s->img_x; ++i) { - out[0] = stbi__blinn_8x8(255 - coutput[0][i], coutput[3][i]); - out[1] = 255; - out += n; - } - } else { - stbi_uc * y = coutput[0]; - if (n == 1) - for (i = 0; i < z->s->img_x; ++i) - out[i] = y[i]; - else - for (i = 0; i < z->s->img_x; ++i) { - *out++ = y[i]; - *out++ = 255; - } - } + stbi_uc *y = coutput[0]; + if (n == 1) + for (i=0; i < z->s->img_x; ++i) out[i] = y[i]; + else + for (i=0; i < z->s->img_x; ++i) { *out++ = y[i]; *out++ = 255; } } - } - stbi__cleanup_jpeg(z); - *out_x = z->s->img_x; - *out_y = z->s->img_y; - if (comp) - *comp = z->s->img_n >= 3 ? 3 : 1; // report original components, not output - return output; - } + } + } + stbi__cleanup_jpeg(z); + *out_x = z->s->img_x; + *out_y = z->s->img_y; + if (comp) *comp = z->s->img_n >= 3 ? 3 : 1; // report original components, not output + return output; + } } -static void * stbi__jpeg_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) { - unsigned char * result; - stbi__jpeg * j = (stbi__jpeg *)stbi__malloc(sizeof(stbi__jpeg)); - if (!j) - return stbi__errpuc("outofmem", "Out of memory"); - memset(j, 0, sizeof(stbi__jpeg)); - STBI_NOTUSED(ri); - j->s = s; - stbi__setup_jpeg(j); - result = load_jpeg_image(j, x, y, comp, req_comp); - STBI_FREE(j); - return result; +static void *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + unsigned char* result; + stbi__jpeg* j = (stbi__jpeg*) stbi__malloc(sizeof(stbi__jpeg)); + if (!j) return stbi__errpuc("outofmem", "Out of memory"); + memset(j, 0, sizeof(stbi__jpeg)); + STBI_NOTUSED(ri); + j->s = s; + stbi__setup_jpeg(j); + result = load_jpeg_image(j, x,y,comp,req_comp); + STBI_FREE(j); + return result; } -static int stbi__jpeg_test(stbi__context * s) { - int r; - stbi__jpeg * j = (stbi__jpeg *)stbi__malloc(sizeof(stbi__jpeg)); - if (!j) - return stbi__err("outofmem", "Out of memory"); - memset(j, 0, sizeof(stbi__jpeg)); - j->s = s; - stbi__setup_jpeg(j); - r = stbi__decode_jpeg_header(j, STBI__SCAN_type); - stbi__rewind(s); - STBI_FREE(j); - return r; +static int stbi__jpeg_test(stbi__context *s) +{ + int r; + stbi__jpeg* j = (stbi__jpeg*)stbi__malloc(sizeof(stbi__jpeg)); + if (!j) return stbi__err("outofmem", "Out of memory"); + memset(j, 0, sizeof(stbi__jpeg)); + j->s = s; + stbi__setup_jpeg(j); + r = stbi__decode_jpeg_header(j, STBI__SCAN_type); + stbi__rewind(s); + STBI_FREE(j); + return r; } -static int stbi__jpeg_info_raw(stbi__jpeg * j, int * x, int * y, int * comp) { - if (!stbi__decode_jpeg_header(j, STBI__SCAN_header)) { - stbi__rewind(j->s); - return 0; - } - if (x) - *x = j->s->img_x; - if (y) - *y = j->s->img_y; - if (comp) - *comp = j->s->img_n >= 3 ? 3 : 1; - return 1; +static int stbi__jpeg_info_raw(stbi__jpeg *j, int *x, int *y, int *comp) +{ + if (!stbi__decode_jpeg_header(j, STBI__SCAN_header)) { + stbi__rewind( j->s ); + return 0; + } + if (x) *x = j->s->img_x; + if (y) *y = j->s->img_y; + if (comp) *comp = j->s->img_n >= 3 ? 3 : 1; + return 1; } -static int stbi__jpeg_info(stbi__context * s, int * x, int * y, int * comp) { - int result; - stbi__jpeg * j = (stbi__jpeg *)(stbi__malloc(sizeof(stbi__jpeg))); - if (!j) - return stbi__err("outofmem", "Out of memory"); - memset(j, 0, sizeof(stbi__jpeg)); - j->s = s; - result = stbi__jpeg_info_raw(j, x, y, comp); - STBI_FREE(j); - return result; +static int stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp) +{ + int result; + stbi__jpeg* j = (stbi__jpeg*) (stbi__malloc(sizeof(stbi__jpeg))); + if (!j) return stbi__err("outofmem", "Out of memory"); + memset(j, 0, sizeof(stbi__jpeg)); + j->s = s; + result = stbi__jpeg_info_raw(j, x, y, comp); + STBI_FREE(j); + return result; } #endif @@ -4278,81 +4088,84 @@ static int stbi__jpeg_info(stbi__context * s, int * x, int * y, int * comp) { #ifndef STBI_NO_ZLIB // fast-way is faster to check than jpeg huffman, but slow way is slower -#define STBI__ZFAST_BITS 9 // accelerate all cases in default tables -#define STBI__ZFAST_MASK ((1 << STBI__ZFAST_BITS) - 1) +#define STBI__ZFAST_BITS 9 // accelerate all cases in default tables +#define STBI__ZFAST_MASK ((1 << STBI__ZFAST_BITS) - 1) #define STBI__ZNSYMS 288 // number of symbols in literal/length alphabet // zlib-style huffman encoding // (jpegs packs from left, zlib from right, so can't share code) -typedef struct { - stbi__uint16 fast[1 << STBI__ZFAST_BITS]; - stbi__uint16 firstcode[16]; - int maxcode[17]; - stbi__uint16 firstsymbol[16]; - stbi_uc size[STBI__ZNSYMS]; - stbi__uint16 value[STBI__ZNSYMS]; +typedef struct +{ + stbi__uint16 fast[1 << STBI__ZFAST_BITS]; + stbi__uint16 firstcode[16]; + int maxcode[17]; + stbi__uint16 firstsymbol[16]; + stbi_uc size[STBI__ZNSYMS]; + stbi__uint16 value[STBI__ZNSYMS]; } stbi__zhuffman; -stbi_inline static int stbi__bitreverse16(int n) { - n = ((n & 0xAAAA) >> 1) | ((n & 0x5555) << 1); - n = ((n & 0xCCCC) >> 2) | ((n & 0x3333) << 2); - n = ((n & 0xF0F0) >> 4) | ((n & 0x0F0F) << 4); - n = ((n & 0xFF00) >> 8) | ((n & 0x00FF) << 8); - return n; +stbi_inline static int stbi__bitreverse16(int n) +{ + n = ((n & 0xAAAA) >> 1) | ((n & 0x5555) << 1); + n = ((n & 0xCCCC) >> 2) | ((n & 0x3333) << 2); + n = ((n & 0xF0F0) >> 4) | ((n & 0x0F0F) << 4); + n = ((n & 0xFF00) >> 8) | ((n & 0x00FF) << 8); + return n; } -stbi_inline static int stbi__bit_reverse(int v, int bits) { - STBI_ASSERT(bits <= 16); - // to bit reverse n bits, reverse 16 and shift - // e.g. 11 bits, bit reverse and shift away 5 - return stbi__bitreverse16(v) >> (16 - bits); +stbi_inline static int stbi__bit_reverse(int v, int bits) +{ + STBI_ASSERT(bits <= 16); + // to bit reverse n bits, reverse 16 and shift + // e.g. 11 bits, bit reverse and shift away 5 + return stbi__bitreverse16(v) >> (16-bits); } -static int stbi__zbuild_huffman(stbi__zhuffman * z, const stbi_uc * sizelist, int num) { - int i, k = 0; - int code, next_code[16], sizes[17]; +static int stbi__zbuild_huffman(stbi__zhuffman *z, const stbi_uc *sizelist, int num) +{ + int i,k=0; + int code, next_code[16], sizes[17]; - // DEFLATE spec for generating codes - memset(sizes, 0, sizeof(sizes)); - memset(z->fast, 0, sizeof(z->fast)); - for (i = 0; i < num; ++i) - ++sizes[sizelist[i]]; - sizes[0] = 0; - for (i = 1; i < 16; ++i) - if (sizes[i] > (1 << i)) - return stbi__err("bad sizes", "Corrupt PNG"); - code = 0; - for (i = 1; i < 16; ++i) { - next_code[i] = code; - z->firstcode[i] = (stbi__uint16)code; - z->firstsymbol[i] = (stbi__uint16)k; - code = (code + sizes[i]); - if (sizes[i]) - if (code - 1 >= (1 << i)) - return stbi__err("bad codelengths", "Corrupt PNG"); - z->maxcode[i] = code << (16 - i); // preshift for inner loop - code <<= 1; - k += sizes[i]; - } - z->maxcode[16] = 0x10000; // sentinel - for (i = 0; i < num; ++i) { - int s = sizelist[i]; - if (s) { - int c = next_code[s] - z->firstcode[s] + z->firstsymbol[s]; - stbi__uint16 fastv = (stbi__uint16)((s << 9) | i); - z->size[c] = (stbi_uc)s; - z->value[c] = (stbi__uint16)i; - if (s <= STBI__ZFAST_BITS) { - int j = stbi__bit_reverse(next_code[s], s); - while (j < (1 << STBI__ZFAST_BITS)) { - z->fast[j] = fastv; - j += (1 << s); - } + // DEFLATE spec for generating codes + memset(sizes, 0, sizeof(sizes)); + memset(z->fast, 0, sizeof(z->fast)); + for (i=0; i < num; ++i) + ++sizes[sizelist[i]]; + sizes[0] = 0; + for (i=1; i < 16; ++i) + if (sizes[i] > (1 << i)) + return stbi__err("bad sizes", "Corrupt PNG"); + code = 0; + for (i=1; i < 16; ++i) { + next_code[i] = code; + z->firstcode[i] = (stbi__uint16) code; + z->firstsymbol[i] = (stbi__uint16) k; + code = (code + sizes[i]); + if (sizes[i]) + if (code-1 >= (1 << i)) return stbi__err("bad codelengths","Corrupt PNG"); + z->maxcode[i] = code << (16-i); // preshift for inner loop + code <<= 1; + k += sizes[i]; + } + z->maxcode[16] = 0x10000; // sentinel + for (i=0; i < num; ++i) { + int s = sizelist[i]; + if (s) { + int c = next_code[s] - z->firstcode[s] + z->firstsymbol[s]; + stbi__uint16 fastv = (stbi__uint16) ((s << 9) | i); + z->size [c] = (stbi_uc ) s; + z->value[c] = (stbi__uint16) i; + if (s <= STBI__ZFAST_BITS) { + int j = stbi__bit_reverse(next_code[s],s); + while (j < (1 << STBI__ZFAST_BITS)) { + z->fast[j] = fastv; + j += (1 << s); } - ++next_code[s]; - } - } - return 1; + } + ++next_code[s]; + } + } + return 1; } // zlib-from-memory implementation for PNG reading @@ -4361,298 +4174,297 @@ static int stbi__zbuild_huffman(stbi__zhuffman * z, const stbi_uc * sizelist, in // we require PNG read all the IDATs and combine them into a single // memory buffer -typedef struct { - stbi_uc *zbuffer, *zbuffer_end; - int num_bits; - stbi__uint32 code_buffer; +typedef struct +{ + stbi_uc *zbuffer, *zbuffer_end; + int num_bits; + int hit_zeof_once; + stbi__uint32 code_buffer; - char * zout; - char * zout_start; - char * zout_end; - int z_expandable; + char *zout; + char *zout_start; + char *zout_end; + int z_expandable; - stbi__zhuffman z_length, z_distance; + stbi__zhuffman z_length, z_distance; } stbi__zbuf; -stbi_inline static int stbi__zeof(stbi__zbuf * z) { return (z->zbuffer >= z->zbuffer_end); } - -stbi_inline static stbi_uc stbi__zget8(stbi__zbuf * z) { return stbi__zeof(z) ? 0 : *z->zbuffer++; } - -static void stbi__fill_bits(stbi__zbuf * z) { - do { - if (z->code_buffer >= (1U << z->num_bits)) { - z->zbuffer = z->zbuffer_end; /* treat this as EOF so we fail. */ - return; - } - z->code_buffer |= (unsigned int)stbi__zget8(z) << z->num_bits; - z->num_bits += 8; - } while (z->num_bits <= 24); -} - -stbi_inline static unsigned int stbi__zreceive(stbi__zbuf * z, int n) { - unsigned int k; - if (z->num_bits < n) - stbi__fill_bits(z); - k = z->code_buffer & ((1 << n) - 1); - z->code_buffer >>= n; - z->num_bits -= n; - return k; -} - -static int stbi__zhuffman_decode_slowpath(stbi__zbuf * a, stbi__zhuffman * z) { - int b, s, k; - // not resolved by fast table, so compute it the slow way - // use jpeg approach, which requires MSbits at top - k = stbi__bit_reverse(a->code_buffer, 16); - for (s = STBI__ZFAST_BITS + 1;; ++s) - if (k < z->maxcode[s]) - break; - if (s >= 16) - return -1; // invalid code! - // code size is s, so: - b = (k >> (16 - s)) - z->firstcode[s] + z->firstsymbol[s]; - if (b >= STBI__ZNSYMS) - return -1; // some data was corrupt somewhere! - if (z->size[b] != s) - return -1; // was originally an assert, but report failure instead. - a->code_buffer >>= s; - a->num_bits -= s; - return z->value[b]; -} - -stbi_inline static int stbi__zhuffman_decode(stbi__zbuf * a, stbi__zhuffman * z) { - int b, s; - if (a->num_bits < 16) { - if (stbi__zeof(a)) { - return -1; /* report error for unexpected end of data. */ - } - stbi__fill_bits(a); - } - b = z->fast[a->code_buffer & STBI__ZFAST_MASK]; - if (b) { - s = b >> 9; - a->code_buffer >>= s; - a->num_bits -= s; - return b & 511; - } - return stbi__zhuffman_decode_slowpath(a, z); -} - -static int stbi__zexpand(stbi__zbuf * z, char * zout, int n) // need to make room for n bytes +stbi_inline static int stbi__zeof(stbi__zbuf *z) { - char * q; - unsigned int cur, limit, old_limit; - z->zout = zout; - if (!z->z_expandable) - return stbi__err("output buffer limit", "Corrupt PNG"); - cur = (unsigned int)(z->zout - z->zout_start); - limit = old_limit = (unsigned)(z->zout_end - z->zout_start); - if (UINT_MAX - cur < (unsigned)n) - return stbi__err("outofmem", "Out of memory"); - while (cur + n > limit) { - if (limit > UINT_MAX / 2) - return stbi__err("outofmem", "Out of memory"); - limit *= 2; - } - q = (char *)STBI_REALLOC_SIZED(z->zout_start, old_limit, limit); - STBI_NOTUSED(old_limit); - if (q == NULL) - return stbi__err("outofmem", "Out of memory"); - z->zout_start = q; - z->zout = q + cur; - z->zout_end = q + limit; - return 1; + return (z->zbuffer >= z->zbuffer_end); } -static const int stbi__zlength_base[31] = {3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 15, 17, 19, 23, 27, 31, - 35, 43, 51, 59, 67, 83, 99, 115, 131, 163, 195, 227, 258, 0, 0}; - -static const int stbi__zlength_extra[31] = {0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, - 3, 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5, 0, 0, 0}; - -static const int stbi__zdist_base[32] = {1, 2, 3, 4, 5, 7, 9, 13, 17, 25, 33, - 49, 65, 97, 129, 193, 257, 385, 513, 769, 1025, 1537, - 2049, 3073, 4097, 6145, 8193, 12289, 16385, 24577, 0, 0}; - -static const int stbi__zdist_extra[32] = {0, 0, 0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, - 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13}; - -static int stbi__parse_huffman_block(stbi__zbuf * a) { - char * zout = a->zout; - for (;;) { - int z = stbi__zhuffman_decode(a, &a->z_length); - if (z < 256) { - if (z < 0) - return stbi__err("bad huffman code", "Corrupt PNG"); // error in huffman codes - if (zout >= a->zout_end) { - if (!stbi__zexpand(a, zout, 1)) - return 0; - zout = a->zout; - } - *zout++ = (char)z; - } else { - stbi_uc * p; - int len, dist; - if (z == 256) { - a->zout = zout; - return 1; - } - if (z >= 286) - return stbi__err("bad huffman code", - "Corrupt PNG"); // per DEFLATE, length codes 286 and 287 must not appear in compressed data - z -= 257; - len = stbi__zlength_base[z]; - if (stbi__zlength_extra[z]) - len += stbi__zreceive(a, stbi__zlength_extra[z]); - z = stbi__zhuffman_decode(a, &a->z_distance); - if (z < 0 || z >= 30) - return stbi__err("bad huffman code", - "Corrupt PNG"); // per DEFLATE, distance codes 30 and 31 must not appear in compressed data - dist = stbi__zdist_base[z]; - if (stbi__zdist_extra[z]) - dist += stbi__zreceive(a, stbi__zdist_extra[z]); - if (zout - a->zout_start < dist) - return stbi__err("bad dist", "Corrupt PNG"); - if (zout + len > a->zout_end) { - if (!stbi__zexpand(a, zout, len)) - return 0; - zout = a->zout; - } - p = (stbi_uc *)(zout - dist); - if (dist == 1) { // run of one byte; common in images. - stbi_uc v = *p; - if (len) { - do - *zout++ = v; - while (--len); - } - } else { - if (len) { - do - *zout++ = *p++; - while (--len); - } - } - } - } +stbi_inline static stbi_uc stbi__zget8(stbi__zbuf *z) +{ + return stbi__zeof(z) ? 0 : *z->zbuffer++; } -static int stbi__compute_huffman_codes(stbi__zbuf * a) { - static const stbi_uc length_dezigzag[19] = {16, 17, 18, 0, 8, 7, 9, 6, 10, 5, 11, 4, 12, 3, 13, 2, 14, 1, 15}; - stbi__zhuffman z_codelength; - stbi_uc lencodes[286 + 32 + 137]; // padding for maximum single op - stbi_uc codelength_sizes[19]; - int i, n; +static void stbi__fill_bits(stbi__zbuf *z) +{ + do { + if (z->code_buffer >= (1U << z->num_bits)) { + z->zbuffer = z->zbuffer_end; /* treat this as EOF so we fail. */ + return; + } + z->code_buffer |= (unsigned int) stbi__zget8(z) << z->num_bits; + z->num_bits += 8; + } while (z->num_bits <= 24); +} - int hlit = stbi__zreceive(a, 5) + 257; - int hdist = stbi__zreceive(a, 5) + 1; - int hclen = stbi__zreceive(a, 4) + 4; - int ntot = hlit + hdist; +stbi_inline static unsigned int stbi__zreceive(stbi__zbuf *z, int n) +{ + unsigned int k; + if (z->num_bits < n) stbi__fill_bits(z); + k = z->code_buffer & ((1 << n) - 1); + z->code_buffer >>= n; + z->num_bits -= n; + return k; +} - memset(codelength_sizes, 0, sizeof(codelength_sizes)); - for (i = 0; i < hclen; ++i) { - int s = stbi__zreceive(a, 3); - codelength_sizes[length_dezigzag[i]] = (stbi_uc)s; - } - if (!stbi__zbuild_huffman(&z_codelength, codelength_sizes, 19)) - return 0; +static int stbi__zhuffman_decode_slowpath(stbi__zbuf *a, stbi__zhuffman *z) +{ + int b,s,k; + // not resolved by fast table, so compute it the slow way + // use jpeg approach, which requires MSbits at top + k = stbi__bit_reverse(a->code_buffer, 16); + for (s=STBI__ZFAST_BITS+1; ; ++s) + if (k < z->maxcode[s]) + break; + if (s >= 16) return -1; // invalid code! + // code size is s, so: + b = (k >> (16-s)) - z->firstcode[s] + z->firstsymbol[s]; + if (b >= STBI__ZNSYMS) return -1; // some data was corrupt somewhere! + if (z->size[b] != s) return -1; // was originally an assert, but report failure instead. + a->code_buffer >>= s; + a->num_bits -= s; + return z->value[b]; +} - n = 0; - while (n < ntot) { - int c = stbi__zhuffman_decode(a, &z_codelength); - if (c < 0 || c >= 19) +stbi_inline static int stbi__zhuffman_decode(stbi__zbuf *a, stbi__zhuffman *z) +{ + int b,s; + if (a->num_bits < 16) { + if (stbi__zeof(a)) { + if (!a->hit_zeof_once) { + // This is the first time we hit eof, insert 16 extra padding btis + // to allow us to keep going; if we actually consume any of them + // though, that is invalid data. This is caught later. + a->hit_zeof_once = 1; + a->num_bits += 16; // add 16 implicit zero bits + } else { + // We already inserted our extra 16 padding bits and are again + // out, this stream is actually prematurely terminated. + return -1; + } + } else { + stbi__fill_bits(a); + } + } + b = z->fast[a->code_buffer & STBI__ZFAST_MASK]; + if (b) { + s = b >> 9; + a->code_buffer >>= s; + a->num_bits -= s; + return b & 511; + } + return stbi__zhuffman_decode_slowpath(a, z); +} + +static int stbi__zexpand(stbi__zbuf *z, char *zout, int n) // need to make room for n bytes +{ + char *q; + unsigned int cur, limit, old_limit; + z->zout = zout; + if (!z->z_expandable) return stbi__err("output buffer limit","Corrupt PNG"); + cur = (unsigned int) (z->zout - z->zout_start); + limit = old_limit = (unsigned) (z->zout_end - z->zout_start); + if (UINT_MAX - cur < (unsigned) n) return stbi__err("outofmem", "Out of memory"); + while (cur + n > limit) { + if(limit > UINT_MAX / 2) return stbi__err("outofmem", "Out of memory"); + limit *= 2; + } + q = (char *) STBI_REALLOC_SIZED(z->zout_start, old_limit, limit); + STBI_NOTUSED(old_limit); + if (q == NULL) return stbi__err("outofmem", "Out of memory"); + z->zout_start = q; + z->zout = q + cur; + z->zout_end = q + limit; + return 1; +} + +static const int stbi__zlength_base[31] = { + 3,4,5,6,7,8,9,10,11,13, + 15,17,19,23,27,31,35,43,51,59, + 67,83,99,115,131,163,195,227,258,0,0 }; + +static const int stbi__zlength_extra[31]= +{ 0,0,0,0,0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3,4,4,4,4,5,5,5,5,0,0,0 }; + +static const int stbi__zdist_base[32] = { 1,2,3,4,5,7,9,13,17,25,33,49,65,97,129,193, +257,385,513,769,1025,1537,2049,3073,4097,6145,8193,12289,16385,24577,0,0}; + +static const int stbi__zdist_extra[32] = +{ 0,0,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8,8,9,9,10,10,11,11,12,12,13,13}; + +static int stbi__parse_huffman_block(stbi__zbuf *a) +{ + char *zout = a->zout; + for(;;) { + int z = stbi__zhuffman_decode(a, &a->z_length); + if (z < 256) { + if (z < 0) return stbi__err("bad huffman code","Corrupt PNG"); // error in huffman codes + if (zout >= a->zout_end) { + if (!stbi__zexpand(a, zout, 1)) return 0; + zout = a->zout; + } + *zout++ = (char) z; + } else { + stbi_uc *p; + int len,dist; + if (z == 256) { + a->zout = zout; + if (a->hit_zeof_once && a->num_bits < 16) { + // The first time we hit zeof, we inserted 16 extra zero bits into our bit + // buffer so the decoder can just do its speculative decoding. But if we + // actually consumed any of those bits (which is the case when num_bits < 16), + // the stream actually read past the end so it is malformed. + return stbi__err("unexpected end","Corrupt PNG"); + } + return 1; + } + if (z >= 286) return stbi__err("bad huffman code","Corrupt PNG"); // per DEFLATE, length codes 286 and 287 must not appear in compressed data + z -= 257; + len = stbi__zlength_base[z]; + if (stbi__zlength_extra[z]) len += stbi__zreceive(a, stbi__zlength_extra[z]); + z = stbi__zhuffman_decode(a, &a->z_distance); + if (z < 0 || z >= 30) return stbi__err("bad huffman code","Corrupt PNG"); // per DEFLATE, distance codes 30 and 31 must not appear in compressed data + dist = stbi__zdist_base[z]; + if (stbi__zdist_extra[z]) dist += stbi__zreceive(a, stbi__zdist_extra[z]); + if (zout - a->zout_start < dist) return stbi__err("bad dist","Corrupt PNG"); + if (len > a->zout_end - zout) { + if (!stbi__zexpand(a, zout, len)) return 0; + zout = a->zout; + } + p = (stbi_uc *) (zout - dist); + if (dist == 1) { // run of one byte; common in images. + stbi_uc v = *p; + if (len) { do *zout++ = v; while (--len); } + } else { + if (len) { do *zout++ = *p++; while (--len); } + } + } + } +} + +static int stbi__compute_huffman_codes(stbi__zbuf *a) +{ + static const stbi_uc length_dezigzag[19] = { 16,17,18,0,8,7,9,6,10,5,11,4,12,3,13,2,14,1,15 }; + stbi__zhuffman z_codelength; + stbi_uc lencodes[286+32+137];//padding for maximum single op + stbi_uc codelength_sizes[19]; + int i,n; + + int hlit = stbi__zreceive(a,5) + 257; + int hdist = stbi__zreceive(a,5) + 1; + int hclen = stbi__zreceive(a,4) + 4; + int ntot = hlit + hdist; + + memset(codelength_sizes, 0, sizeof(codelength_sizes)); + for (i=0; i < hclen; ++i) { + int s = stbi__zreceive(a,3); + codelength_sizes[length_dezigzag[i]] = (stbi_uc) s; + } + if (!stbi__zbuild_huffman(&z_codelength, codelength_sizes, 19)) return 0; + + n = 0; + while (n < ntot) { + int c = stbi__zhuffman_decode(a, &z_codelength); + if (c < 0 || c >= 19) return stbi__err("bad codelengths", "Corrupt PNG"); + if (c < 16) + lencodes[n++] = (stbi_uc) c; + else { + stbi_uc fill = 0; + if (c == 16) { + c = stbi__zreceive(a,2)+3; + if (n == 0) return stbi__err("bad codelengths", "Corrupt PNG"); + fill = lencodes[n-1]; + } else if (c == 17) { + c = stbi__zreceive(a,3)+3; + } else if (c == 18) { + c = stbi__zreceive(a,7)+11; + } else { return stbi__err("bad codelengths", "Corrupt PNG"); - if (c < 16) - lencodes[n++] = (stbi_uc)c; - else { - stbi_uc fill = 0; - if (c == 16) { - c = stbi__zreceive(a, 2) + 3; - if (n == 0) - return stbi__err("bad codelengths", "Corrupt PNG"); - fill = lencodes[n - 1]; - } else if (c == 17) { - c = stbi__zreceive(a, 3) + 3; - } else if (c == 18) { - c = stbi__zreceive(a, 7) + 11; - } else { - return stbi__err("bad codelengths", "Corrupt PNG"); - } - if (ntot - n < c) - return stbi__err("bad codelengths", "Corrupt PNG"); - memset(lencodes + n, fill, c); - n += c; - } - } - if (n != ntot) - return stbi__err("bad codelengths", "Corrupt PNG"); - if (!stbi__zbuild_huffman(&a->z_length, lencodes, hlit)) - return 0; - if (!stbi__zbuild_huffman(&a->z_distance, lencodes + hlit, hdist)) - return 0; - return 1; + } + if (ntot - n < c) return stbi__err("bad codelengths", "Corrupt PNG"); + memset(lencodes+n, fill, c); + n += c; + } + } + if (n != ntot) return stbi__err("bad codelengths","Corrupt PNG"); + if (!stbi__zbuild_huffman(&a->z_length, lencodes, hlit)) return 0; + if (!stbi__zbuild_huffman(&a->z_distance, lencodes+hlit, hdist)) return 0; + return 1; } -static int stbi__parse_uncompressed_block(stbi__zbuf * a) { - stbi_uc header[4]; - int len, nlen, k; - if (a->num_bits & 7) - stbi__zreceive(a, a->num_bits & 7); // discard - // drain the bit-packed data into header - k = 0; - while (a->num_bits > 0) { - header[k++] = (stbi_uc)(a->code_buffer & 255); // suppress MSVC run-time check - a->code_buffer >>= 8; - a->num_bits -= 8; - } - if (a->num_bits < 0) - return stbi__err("zlib corrupt", "Corrupt PNG"); - // now fill header the normal way - while (k < 4) - header[k++] = stbi__zget8(a); - len = header[1] * 256 + header[0]; - nlen = header[3] * 256 + header[2]; - if (nlen != (len ^ 0xffff)) - return stbi__err("zlib corrupt", "Corrupt PNG"); - if (a->zbuffer + len > a->zbuffer_end) - return stbi__err("read past buffer", "Corrupt PNG"); - if (a->zout + len > a->zout_end) - if (!stbi__zexpand(a, a->zout, len)) - return 0; - memcpy(a->zout, a->zbuffer, len); - a->zbuffer += len; - a->zout += len; - return 1; +static int stbi__parse_uncompressed_block(stbi__zbuf *a) +{ + stbi_uc header[4]; + int len,nlen,k; + if (a->num_bits & 7) + stbi__zreceive(a, a->num_bits & 7); // discard + // drain the bit-packed data into header + k = 0; + while (a->num_bits > 0) { + header[k++] = (stbi_uc) (a->code_buffer & 255); // suppress MSVC run-time check + a->code_buffer >>= 8; + a->num_bits -= 8; + } + if (a->num_bits < 0) return stbi__err("zlib corrupt","Corrupt PNG"); + // now fill header the normal way + while (k < 4) + header[k++] = stbi__zget8(a); + len = header[1] * 256 + header[0]; + nlen = header[3] * 256 + header[2]; + if (nlen != (len ^ 0xffff)) return stbi__err("zlib corrupt","Corrupt PNG"); + if (a->zbuffer + len > a->zbuffer_end) return stbi__err("read past buffer","Corrupt PNG"); + if (a->zout + len > a->zout_end) + if (!stbi__zexpand(a, a->zout, len)) return 0; + memcpy(a->zout, a->zbuffer, len); + a->zbuffer += len; + a->zout += len; + return 1; } -static int stbi__parse_zlib_header(stbi__zbuf * a) { - int cmf = stbi__zget8(a); - int cm = cmf & 15; - /* int cinfo = cmf >> 4; */ - int flg = stbi__zget8(a); - if (stbi__zeof(a)) - return stbi__err("bad zlib header", "Corrupt PNG"); // zlib spec - if ((cmf * 256 + flg) % 31 != 0) - return stbi__err("bad zlib header", "Corrupt PNG"); // zlib spec - if (flg & 32) - return stbi__err("no preset dict", "Corrupt PNG"); // preset dictionary not allowed in png - if (cm != 8) - return stbi__err("bad compression", "Corrupt PNG"); // DEFLATE required for png - // window = 1 << (8 + cinfo)... but who cares, we fully buffer output - return 1; +static int stbi__parse_zlib_header(stbi__zbuf *a) +{ + int cmf = stbi__zget8(a); + int cm = cmf & 15; + /* int cinfo = cmf >> 4; */ + int flg = stbi__zget8(a); + if (stbi__zeof(a)) return stbi__err("bad zlib header","Corrupt PNG"); // zlib spec + if ((cmf*256+flg) % 31 != 0) return stbi__err("bad zlib header","Corrupt PNG"); // zlib spec + if (flg & 32) return stbi__err("no preset dict","Corrupt PNG"); // preset dictionary not allowed in png + if (cm != 8) return stbi__err("bad compression","Corrupt PNG"); // DEFLATE required for png + // window = 1 << (8 + cinfo)... but who cares, we fully buffer output + return 1; } -static const stbi_uc stbi__zdefault_length[STBI__ZNSYMS] = { - 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, - 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, - 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, - 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, - 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, - 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, - 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, - 9, 9, 9, 9, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 8, 8}; -static const stbi_uc stbi__zdefault_distance[32] = {5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, - 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5}; +static const stbi_uc stbi__zdefault_length[STBI__ZNSYMS] = +{ + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, + 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, + 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, + 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, + 7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7, 7,7,7,7,7,7,7,7,8,8,8,8,8,8,8,8 +}; +static const stbi_uc stbi__zdefault_distance[32] = +{ + 5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5 +}; /* Init algorithm: { @@ -4666,122 +4478,118 @@ Init algorithm: } */ -static int stbi__parse_zlib(stbi__zbuf * a, int parse_header) { - int final, type; - if (parse_header) - if (!stbi__parse_zlib_header(a)) - return 0; - a->num_bits = 0; - a->code_buffer = 0; - do { - final = stbi__zreceive(a, 1); - type = stbi__zreceive(a, 2); - if (type == 0) { - if (!stbi__parse_uncompressed_block(a)) - return 0; - } else if (type == 3) { - return 0; - } else { - if (type == 1) { - // use fixed code lengths - if (!stbi__zbuild_huffman(&a->z_length, stbi__zdefault_length, STBI__ZNSYMS)) - return 0; - if (!stbi__zbuild_huffman(&a->z_distance, stbi__zdefault_distance, 32)) - return 0; - } else { - if (!stbi__compute_huffman_codes(a)) - return 0; - } - if (!stbi__parse_huffman_block(a)) - return 0; - } - } while (!final); - return 1; +static int stbi__parse_zlib(stbi__zbuf *a, int parse_header) +{ + int final, type; + if (parse_header) + if (!stbi__parse_zlib_header(a)) return 0; + a->num_bits = 0; + a->code_buffer = 0; + a->hit_zeof_once = 0; + do { + final = stbi__zreceive(a,1); + type = stbi__zreceive(a,2); + if (type == 0) { + if (!stbi__parse_uncompressed_block(a)) return 0; + } else if (type == 3) { + return 0; + } else { + if (type == 1) { + // use fixed code lengths + if (!stbi__zbuild_huffman(&a->z_length , stbi__zdefault_length , STBI__ZNSYMS)) return 0; + if (!stbi__zbuild_huffman(&a->z_distance, stbi__zdefault_distance, 32)) return 0; + } else { + if (!stbi__compute_huffman_codes(a)) return 0; + } + if (!stbi__parse_huffman_block(a)) return 0; + } + } while (!final); + return 1; } -static int stbi__do_zlib(stbi__zbuf * a, char * obuf, int olen, int exp, int parse_header) { - a->zout_start = obuf; - a->zout = obuf; - a->zout_end = obuf + olen; - a->z_expandable = exp; +static int stbi__do_zlib(stbi__zbuf *a, char *obuf, int olen, int exp, int parse_header) +{ + a->zout_start = obuf; + a->zout = obuf; + a->zout_end = obuf + olen; + a->z_expandable = exp; - return stbi__parse_zlib(a, parse_header); + return stbi__parse_zlib(a, parse_header); } -STBIDEF char * stbi_zlib_decode_malloc_guesssize(const char * buffer, int len, int initial_size, int * outlen) { - stbi__zbuf a; - char * p = (char *)stbi__malloc(initial_size); - if (p == NULL) - return NULL; - a.zbuffer = (stbi_uc *)buffer; - a.zbuffer_end = (stbi_uc *)buffer + len; - if (stbi__do_zlib(&a, p, initial_size, 1, 1)) { - if (outlen) - *outlen = (int)(a.zout - a.zout_start); - return a.zout_start; - } else { - STBI_FREE(a.zout_start); - return NULL; - } +STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen) +{ + stbi__zbuf a; + char *p = (char *) stbi__malloc(initial_size); + if (p == NULL) return NULL; + a.zbuffer = (stbi_uc *) buffer; + a.zbuffer_end = (stbi_uc *) buffer + len; + if (stbi__do_zlib(&a, p, initial_size, 1, 1)) { + if (outlen) *outlen = (int) (a.zout - a.zout_start); + return a.zout_start; + } else { + STBI_FREE(a.zout_start); + return NULL; + } } -STBIDEF char * stbi_zlib_decode_malloc(char const * buffer, int len, int * outlen) { - return stbi_zlib_decode_malloc_guesssize(buffer, len, 16384, outlen); +STBIDEF char *stbi_zlib_decode_malloc(char const *buffer, int len, int *outlen) +{ + return stbi_zlib_decode_malloc_guesssize(buffer, len, 16384, outlen); } -STBIDEF char * stbi_zlib_decode_malloc_guesssize_headerflag(const char * buffer, int len, int initial_size, int * outlen, - int parse_header) { - stbi__zbuf a; - char * p = (char *)stbi__malloc(initial_size); - if (p == NULL) - return NULL; - a.zbuffer = (stbi_uc *)buffer; - a.zbuffer_end = (stbi_uc *)buffer + len; - if (stbi__do_zlib(&a, p, initial_size, 1, parse_header)) { - if (outlen) - *outlen = (int)(a.zout - a.zout_start); - return a.zout_start; - } else { - STBI_FREE(a.zout_start); - return NULL; - } +STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header) +{ + stbi__zbuf a; + char *p = (char *) stbi__malloc(initial_size); + if (p == NULL) return NULL; + a.zbuffer = (stbi_uc *) buffer; + a.zbuffer_end = (stbi_uc *) buffer + len; + if (stbi__do_zlib(&a, p, initial_size, 1, parse_header)) { + if (outlen) *outlen = (int) (a.zout - a.zout_start); + return a.zout_start; + } else { + STBI_FREE(a.zout_start); + return NULL; + } } -STBIDEF int stbi_zlib_decode_buffer(char * obuffer, int olen, char const * ibuffer, int ilen) { - stbi__zbuf a; - a.zbuffer = (stbi_uc *)ibuffer; - a.zbuffer_end = (stbi_uc *)ibuffer + ilen; - if (stbi__do_zlib(&a, obuffer, olen, 0, 1)) - return (int)(a.zout - a.zout_start); - else - return -1; +STBIDEF int stbi_zlib_decode_buffer(char *obuffer, int olen, char const *ibuffer, int ilen) +{ + stbi__zbuf a; + a.zbuffer = (stbi_uc *) ibuffer; + a.zbuffer_end = (stbi_uc *) ibuffer + ilen; + if (stbi__do_zlib(&a, obuffer, olen, 0, 1)) + return (int) (a.zout - a.zout_start); + else + return -1; } -STBIDEF char * stbi_zlib_decode_noheader_malloc(char const * buffer, int len, int * outlen) { - stbi__zbuf a; - char * p = (char *)stbi__malloc(16384); - if (p == NULL) - return NULL; - a.zbuffer = (stbi_uc *)buffer; - a.zbuffer_end = (stbi_uc *)buffer + len; - if (stbi__do_zlib(&a, p, 16384, 1, 0)) { - if (outlen) - *outlen = (int)(a.zout - a.zout_start); - return a.zout_start; - } else { - STBI_FREE(a.zout_start); - return NULL; - } +STBIDEF char *stbi_zlib_decode_noheader_malloc(char const *buffer, int len, int *outlen) +{ + stbi__zbuf a; + char *p = (char *) stbi__malloc(16384); + if (p == NULL) return NULL; + a.zbuffer = (stbi_uc *) buffer; + a.zbuffer_end = (stbi_uc *) buffer+len; + if (stbi__do_zlib(&a, p, 16384, 1, 0)) { + if (outlen) *outlen = (int) (a.zout - a.zout_start); + return a.zout_start; + } else { + STBI_FREE(a.zout_start); + return NULL; + } } -STBIDEF int stbi_zlib_decode_noheader_buffer(char * obuffer, int olen, const char * ibuffer, int ilen) { - stbi__zbuf a; - a.zbuffer = (stbi_uc *)ibuffer; - a.zbuffer_end = (stbi_uc *)ibuffer + ilen; - if (stbi__do_zlib(&a, obuffer, olen, 0, 0)) - return (int)(a.zout - a.zout_start); - else - return -1; +STBIDEF int stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen) +{ + stbi__zbuf a; + a.zbuffer = (stbi_uc *) ibuffer; + a.zbuffer_end = (stbi_uc *) ibuffer + ilen; + if (stbi__do_zlib(&a, obuffer, olen, 0, 0)) + return (int) (a.zout - a.zout_start); + else + return -1; } #endif @@ -4796,1303 +4604,1131 @@ STBIDEF int stbi_zlib_decode_noheader_buffer(char * obuffer, int olen, const cha // - uses stb_zlib, a PD zlib implementation with fast huffman decoding #ifndef STBI_NO_PNG -typedef struct { - stbi__uint32 length; - stbi__uint32 type; +typedef struct +{ + stbi__uint32 length; + stbi__uint32 type; } stbi__pngchunk; -static stbi__pngchunk stbi__get_chunk_header(stbi__context * s) { - stbi__pngchunk c; - c.length = stbi__get32be(s); - c.type = stbi__get32be(s); - return c; +static stbi__pngchunk stbi__get_chunk_header(stbi__context *s) +{ + stbi__pngchunk c; + c.length = stbi__get32be(s); + c.type = stbi__get32be(s); + return c; } -static int stbi__check_png_header(stbi__context * s) { - static const stbi_uc png_sig[8] = {137, 80, 78, 71, 13, 10, 26, 10}; - int i; - for (i = 0; i < 8; ++i) - if (stbi__get8(s) != png_sig[i]) - return stbi__err("bad png sig", "Not a PNG"); - return 1; +static int stbi__check_png_header(stbi__context *s) +{ + static const stbi_uc png_sig[8] = { 137,80,78,71,13,10,26,10 }; + int i; + for (i=0; i < 8; ++i) + if (stbi__get8(s) != png_sig[i]) return stbi__err("bad png sig","Not a PNG"); + return 1; } -typedef struct { - stbi__context * s; - stbi_uc *idata, *expanded, *out; - int depth; +typedef struct +{ + stbi__context *s; + stbi_uc *idata, *expanded, *out; + int depth; } stbi__png; + enum { - STBI__F_none = 0, - STBI__F_sub = 1, - STBI__F_up = 2, - STBI__F_avg = 3, - STBI__F_paeth = 4, - // synthetic filters used for first scanline to avoid needing a dummy row of 0s - STBI__F_avg_first, - STBI__F_paeth_first + STBI__F_none=0, + STBI__F_sub=1, + STBI__F_up=2, + STBI__F_avg=3, + STBI__F_paeth=4, + // synthetic filter used for first scanline to avoid needing a dummy row of 0s + STBI__F_avg_first }; -static stbi_uc first_row_filter[5] = {STBI__F_none, STBI__F_sub, STBI__F_none, STBI__F_avg_first, STBI__F_paeth_first}; +static stbi_uc first_row_filter[5] = +{ + STBI__F_none, + STBI__F_sub, + STBI__F_none, + STBI__F_avg_first, + STBI__F_sub // Paeth with b=c=0 turns out to be equivalent to sub +}; -static int stbi__paeth(int a, int b, int c) { - int p = a + b - c; - int pa = abs(p - a); - int pb = abs(p - b); - int pc = abs(p - c); - if (pa <= pb && pa <= pc) - return a; - if (pb <= pc) - return b; - return c; +static int stbi__paeth(int a, int b, int c) +{ + // This formulation looks very different from the reference in the PNG spec, but is + // actually equivalent and has favorable data dependencies and admits straightforward + // generation of branch-free code, which helps performance significantly. + int thresh = c*3 - (a + b); + int lo = a < b ? a : b; + int hi = a < b ? b : a; + int t0 = (hi <= thresh) ? lo : c; + int t1 = (thresh <= lo) ? hi : t0; + return t1; } -static const stbi_uc stbi__depth_scale_table[9] = {0, 0xff, 0x55, 0, 0x11, 0, 0, 0, 0x01}; +static const stbi_uc stbi__depth_scale_table[9] = { 0, 0xff, 0x55, 0, 0x11, 0,0,0, 0x01 }; + +// adds an extra all-255 alpha channel +// dest == src is legal +// img_n must be 1 or 3 +static void stbi__create_png_alpha_expand8(stbi_uc *dest, stbi_uc *src, stbi__uint32 x, int img_n) +{ + int i; + // must process data backwards since we allow dest==src + if (img_n == 1) { + for (i=x-1; i >= 0; --i) { + dest[i*2+1] = 255; + dest[i*2+0] = src[i]; + } + } else { + STBI_ASSERT(img_n == 3); + for (i=x-1; i >= 0; --i) { + dest[i*4+3] = 255; + dest[i*4+2] = src[i*3+2]; + dest[i*4+1] = src[i*3+1]; + dest[i*4+0] = src[i*3+0]; + } + } +} // create the png data from post-deflated data -static int stbi__create_png_image_raw(stbi__png * a, stbi_uc * raw, stbi__uint32 raw_len, int out_n, stbi__uint32 x, - stbi__uint32 y, int depth, int color) { - int bytes = (depth == 16 ? 2 : 1); - stbi__context * s = a->s; - stbi__uint32 i, j, stride = x * out_n * bytes; - stbi__uint32 img_len, img_width_bytes; - int k; - int img_n = s->img_n; // copy it into a local for later +static int stbi__create_png_image_raw(stbi__png *a, stbi_uc *raw, stbi__uint32 raw_len, int out_n, stbi__uint32 x, stbi__uint32 y, int depth, int color) +{ + int bytes = (depth == 16 ? 2 : 1); + stbi__context *s = a->s; + stbi__uint32 i,j,stride = x*out_n*bytes; + stbi__uint32 img_len, img_width_bytes; + stbi_uc *filter_buf; + int all_ok = 1; + int k; + int img_n = s->img_n; // copy it into a local for later - int output_bytes = out_n * bytes; - int filter_bytes = img_n * bytes; - int width = x; + int output_bytes = out_n*bytes; + int filter_bytes = img_n*bytes; + int width = x; - STBI_ASSERT(out_n == s->img_n || out_n == s->img_n + 1); - a->out = (stbi_uc *)stbi__malloc_mad3(x, y, output_bytes, 0); // extra bytes to write off the end into - if (!a->out) - return stbi__err("outofmem", "Out of memory"); + STBI_ASSERT(out_n == s->img_n || out_n == s->img_n+1); + a->out = (stbi_uc *) stbi__malloc_mad3(x, y, output_bytes, 0); // extra bytes to write off the end into + if (!a->out) return stbi__err("outofmem", "Out of memory"); - if (!stbi__mad3sizes_valid(img_n, x, depth, 7)) - return stbi__err("too large", "Corrupt PNG"); - img_width_bytes = (((img_n * x * depth) + 7) >> 3); - img_len = (img_width_bytes + 1) * y; + // note: error exits here don't need to clean up a->out individually, + // stbi__do_png always does on error. + if (!stbi__mad3sizes_valid(img_n, x, depth, 7)) return stbi__err("too large", "Corrupt PNG"); + img_width_bytes = (((img_n * x * depth) + 7) >> 3); + if (!stbi__mad2sizes_valid(img_width_bytes, y, img_width_bytes)) return stbi__err("too large", "Corrupt PNG"); + img_len = (img_width_bytes + 1) * y; - // we used to check for exact match between raw_len and img_len on non-interlaced PNGs, - // but issue #276 reported a PNG in the wild that had extra data at the end (all zeros), - // so just check for raw_len < img_len always. - if (raw_len < img_len) - return stbi__err("not enough pixels", "Corrupt PNG"); + // we used to check for exact match between raw_len and img_len on non-interlaced PNGs, + // but issue #276 reported a PNG in the wild that had extra data at the end (all zeros), + // so just check for raw_len < img_len always. + if (raw_len < img_len) return stbi__err("not enough pixels","Corrupt PNG"); - for (j = 0; j < y; ++j) { - stbi_uc * cur = a->out + stride * j; - stbi_uc * prior; - int filter = *raw++; + // Allocate two scan lines worth of filter workspace buffer. + filter_buf = (stbi_uc *) stbi__malloc_mad2(img_width_bytes, 2, 0); + if (!filter_buf) return stbi__err("outofmem", "Out of memory"); - if (filter > 4) - return stbi__err("invalid filter", "Corrupt PNG"); + // Filtering for low-bit-depth images + if (depth < 8) { + filter_bytes = 1; + width = img_width_bytes; + } - if (depth < 8) { - if (img_width_bytes > x) - return stbi__err("invalid width", "Corrupt PNG"); - cur += x * out_n - img_width_bytes; // store output to the rightmost img_len bytes, so we can decode in place - filter_bytes = 1; - width = img_width_bytes; - } - prior = cur - stride; // bugfix: need to compute this after 'cur +=' computation above + for (j=0; j < y; ++j) { + // cur/prior filter buffers alternate + stbi_uc *cur = filter_buf + (j & 1)*img_width_bytes; + stbi_uc *prior = filter_buf + (~j & 1)*img_width_bytes; + stbi_uc *dest = a->out + stride*j; + int nk = width * filter_bytes; + int filter = *raw++; - // if first row, use special filter that doesn't sample previous row - if (j == 0) - filter = first_row_filter[filter]; + // check filter type + if (filter > 4) { + all_ok = stbi__err("invalid filter","Corrupt PNG"); + break; + } - // handle first byte explicitly - for (k = 0; k < filter_bytes; ++k) { - switch (filter) { - case STBI__F_none: - cur[k] = raw[k]; - break; - case STBI__F_sub: - cur[k] = raw[k]; - break; - case STBI__F_up: - cur[k] = STBI__BYTECAST(raw[k] + prior[k]); - break; - case STBI__F_avg: - cur[k] = STBI__BYTECAST(raw[k] + (prior[k] >> 1)); - break; - case STBI__F_paeth: - cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(0, prior[k], 0)); - break; - case STBI__F_avg_first: - cur[k] = raw[k]; - break; - case STBI__F_paeth_first: - cur[k] = raw[k]; - break; + // if first row, use special filter that doesn't sample previous row + if (j == 0) filter = first_row_filter[filter]; + + // perform actual filtering + switch (filter) { + case STBI__F_none: + memcpy(cur, raw, nk); + break; + case STBI__F_sub: + memcpy(cur, raw, filter_bytes); + for (k = filter_bytes; k < nk; ++k) + cur[k] = STBI__BYTECAST(raw[k] + cur[k-filter_bytes]); + break; + case STBI__F_up: + for (k = 0; k < nk; ++k) + cur[k] = STBI__BYTECAST(raw[k] + prior[k]); + break; + case STBI__F_avg: + for (k = 0; k < filter_bytes; ++k) + cur[k] = STBI__BYTECAST(raw[k] + (prior[k]>>1)); + for (k = filter_bytes; k < nk; ++k) + cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k-filter_bytes])>>1)); + break; + case STBI__F_paeth: + for (k = 0; k < filter_bytes; ++k) + cur[k] = STBI__BYTECAST(raw[k] + prior[k]); // prior[k] == stbi__paeth(0,prior[k],0) + for (k = filter_bytes; k < nk; ++k) + cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k-filter_bytes], prior[k], prior[k-filter_bytes])); + break; + case STBI__F_avg_first: + memcpy(cur, raw, filter_bytes); + for (k = filter_bytes; k < nk; ++k) + cur[k] = STBI__BYTECAST(raw[k] + (cur[k-filter_bytes] >> 1)); + break; + } + + raw += nk; + + // expand decoded bits in cur to dest, also adding an extra alpha channel if desired + if (depth < 8) { + stbi_uc scale = (color == 0) ? stbi__depth_scale_table[depth] : 1; // scale grayscale values to 0..255 range + stbi_uc *in = cur; + stbi_uc *out = dest; + stbi_uc inb = 0; + stbi__uint32 nsmp = x*img_n; + + // expand bits to bytes first + if (depth == 4) { + for (i=0; i < nsmp; ++i) { + if ((i & 1) == 0) inb = *in++; + *out++ = scale * (inb >> 4); + inb <<= 4; } - } - - if (depth == 8) { - if (img_n != out_n) - cur[img_n] = 255; // first pixel - raw += img_n; - cur += out_n; - prior += out_n; - } else if (depth == 16) { - if (img_n != out_n) { - cur[filter_bytes] = 255; // first pixel top byte - cur[filter_bytes + 1] = 255; // first pixel bottom byte + } else if (depth == 2) { + for (i=0; i < nsmp; ++i) { + if ((i & 3) == 0) inb = *in++; + *out++ = scale * (inb >> 6); + inb <<= 2; } - raw += filter_bytes; - cur += output_bytes; - prior += output_bytes; - } else { - raw += 1; - cur += 1; - prior += 1; - } - - // this is a little gross, so that we don't switch per-pixel or per-component - if (depth < 8 || img_n == out_n) { - int nk = (width - 1) * filter_bytes; -#define STBI__CASE(f) \ - case f: \ - for (k = 0; k < nk; ++k) - switch (filter) { - // "none" filter turns into a memcpy here; make that explicit. - case STBI__F_none: - memcpy(cur, raw, nk); - break; - STBI__CASE(STBI__F_sub) { cur[k] = STBI__BYTECAST(raw[k] + cur[k - filter_bytes]); } - break; - STBI__CASE(STBI__F_up) { cur[k] = STBI__BYTECAST(raw[k] + prior[k]); } - break; - STBI__CASE(STBI__F_avg) { cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k - filter_bytes]) >> 1)); } - break; - STBI__CASE(STBI__F_paeth) { - cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k - filter_bytes], prior[k], prior[k - filter_bytes])); - } - break; - STBI__CASE(STBI__F_avg_first) { cur[k] = STBI__BYTECAST(raw[k] + (cur[k - filter_bytes] >> 1)); } - break; - STBI__CASE(STBI__F_paeth_first) { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k - filter_bytes], 0, 0)); } - break; + } else { + STBI_ASSERT(depth == 1); + for (i=0; i < nsmp; ++i) { + if ((i & 7) == 0) inb = *in++; + *out++ = scale * (inb >> 7); + inb <<= 1; } -#undef STBI__CASE - raw += nk; - } else { - STBI_ASSERT(img_n + 1 == out_n); -#define STBI__CASE(f) \ - case f: \ - for (i = x - 1; i >= 1; --i, cur[filter_bytes] = 255, raw += filter_bytes, cur += output_bytes, prior += output_bytes) \ - for (k = 0; k < filter_bytes; ++k) - switch (filter) { - STBI__CASE(STBI__F_none) { cur[k] = raw[k]; } - break; - STBI__CASE(STBI__F_sub) { cur[k] = STBI__BYTECAST(raw[k] + cur[k - output_bytes]); } - break; - STBI__CASE(STBI__F_up) { cur[k] = STBI__BYTECAST(raw[k] + prior[k]); } - break; - STBI__CASE(STBI__F_avg) { cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k - output_bytes]) >> 1)); } - break; - STBI__CASE(STBI__F_paeth) { - cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k - output_bytes], prior[k], prior[k - output_bytes])); - } - break; - STBI__CASE(STBI__F_avg_first) { cur[k] = STBI__BYTECAST(raw[k] + (cur[k - output_bytes] >> 1)); } - break; - STBI__CASE(STBI__F_paeth_first) { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k - output_bytes], 0, 0)); } - break; + } + + // insert alpha=255 values if desired + if (img_n != out_n) + stbi__create_png_alpha_expand8(dest, dest, x, img_n); + } else if (depth == 8) { + if (img_n == out_n) + memcpy(dest, cur, x*img_n); + else + stbi__create_png_alpha_expand8(dest, cur, x, img_n); + } else if (depth == 16) { + // convert the image data from big-endian to platform-native + stbi__uint16 *dest16 = (stbi__uint16*)dest; + stbi__uint32 nsmp = x*img_n; + + if (img_n == out_n) { + for (i = 0; i < nsmp; ++i, ++dest16, cur += 2) + *dest16 = (cur[0] << 8) | cur[1]; + } else { + STBI_ASSERT(img_n+1 == out_n); + if (img_n == 1) { + for (i = 0; i < x; ++i, dest16 += 2, cur += 2) { + dest16[0] = (cur[0] << 8) | cur[1]; + dest16[1] = 0xffff; + } + } else { + STBI_ASSERT(img_n == 3); + for (i = 0; i < x; ++i, dest16 += 4, cur += 6) { + dest16[0] = (cur[0] << 8) | cur[1]; + dest16[1] = (cur[2] << 8) | cur[3]; + dest16[2] = (cur[4] << 8) | cur[5]; + dest16[3] = 0xffff; + } } -#undef STBI__CASE + } + } + } - // the loop above sets the high byte of the pixels' alpha, but for - // 16 bit png files we also need the low byte set. we'll do that here. - if (depth == 16) { - cur = a->out + stride * j; // start at the beginning of the row again - for (i = 0; i < x; ++i, cur += output_bytes) { - cur[filter_bytes + 1] = 255; - } - } - } - } + STBI_FREE(filter_buf); + if (!all_ok) return 0; - // we make a separate pass to expand bits to pixels; for performance, - // this could run two scanlines behind the above code, so it won't - // intefere with filtering but will still be in the cache. - if (depth < 8) { - for (j = 0; j < y; ++j) { - stbi_uc * cur = a->out + stride * j; - stbi_uc * in = a->out + stride * j + x * out_n - img_width_bytes; - // unpack 1/2/4-bit into a 8-bit buffer. allows us to keep the common 8-bit path optimal at minimal cost for - // 1/2/4-bit png guarante byte alignment, if width is not multiple of 8/4/2 we'll decode dummy trailing data that - // will be skipped in the later loop - stbi_uc scale = (color == 0) ? stbi__depth_scale_table[depth] : 1; // scale grayscale values to 0..255 range - - // note that the final byte might overshoot and write more data than desired. - // we can allocate enough data that this never writes out of memory, but it - // could also overwrite the next scanline. can it overwrite non-empty data - // on the next scanline? yes, consider 1-pixel-wide scanlines with 1-bit-per-pixel. - // so we need to explicitly clamp the final ones - - if (depth == 4) { - for (k = x * img_n; k >= 2; k -= 2, ++in) { - *cur++ = scale * ((*in >> 4)); - *cur++ = scale * ((*in) & 0x0f); - } - if (k > 0) - *cur++ = scale * ((*in >> 4)); - } else if (depth == 2) { - for (k = x * img_n; k >= 4; k -= 4, ++in) { - *cur++ = scale * ((*in >> 6)); - *cur++ = scale * ((*in >> 4) & 0x03); - *cur++ = scale * ((*in >> 2) & 0x03); - *cur++ = scale * ((*in) & 0x03); - } - if (k > 0) - *cur++ = scale * ((*in >> 6)); - if (k > 1) - *cur++ = scale * ((*in >> 4) & 0x03); - if (k > 2) - *cur++ = scale * ((*in >> 2) & 0x03); - } else if (depth == 1) { - for (k = x * img_n; k >= 8; k -= 8, ++in) { - *cur++ = scale * ((*in >> 7)); - *cur++ = scale * ((*in >> 6) & 0x01); - *cur++ = scale * ((*in >> 5) & 0x01); - *cur++ = scale * ((*in >> 4) & 0x01); - *cur++ = scale * ((*in >> 3) & 0x01); - *cur++ = scale * ((*in >> 2) & 0x01); - *cur++ = scale * ((*in >> 1) & 0x01); - *cur++ = scale * ((*in) & 0x01); - } - if (k > 0) - *cur++ = scale * ((*in >> 7)); - if (k > 1) - *cur++ = scale * ((*in >> 6) & 0x01); - if (k > 2) - *cur++ = scale * ((*in >> 5) & 0x01); - if (k > 3) - *cur++ = scale * ((*in >> 4) & 0x01); - if (k > 4) - *cur++ = scale * ((*in >> 3) & 0x01); - if (k > 5) - *cur++ = scale * ((*in >> 2) & 0x01); - if (k > 6) - *cur++ = scale * ((*in >> 1) & 0x01); - } - if (img_n != out_n) { - int q; - // insert alpha = 255 - cur = a->out + stride * j; - if (img_n == 1) { - for (q = x - 1; q >= 0; --q) { - cur[q * 2 + 1] = 255; - cur[q * 2 + 0] = cur[q]; - } - } else { - STBI_ASSERT(img_n == 3); - for (q = x - 1; q >= 0; --q) { - cur[q * 4 + 3] = 255; - cur[q * 4 + 2] = cur[q * 3 + 2]; - cur[q * 4 + 1] = cur[q * 3 + 1]; - cur[q * 4 + 0] = cur[q * 3 + 0]; - } - } - } - } - } else if (depth == 16) { - // force the image data from big-endian to platform-native. - // this is done in a separate pass due to the decoding relying - // on the data being untouched, but could probably be done - // per-line during decode if care is taken. - stbi_uc * cur = a->out; - stbi__uint16 * cur16 = (stbi__uint16 *)cur; - - for (i = 0; i < x * y * out_n; ++i, cur16++, cur += 2) { - *cur16 = (cur[0] << 8) | cur[1]; - } - } - - return 1; + return 1; } -static int stbi__create_png_image(stbi__png * a, stbi_uc * image_data, stbi__uint32 image_data_len, int out_n, int depth, - int color, int interlaced) { - int bytes = (depth == 16 ? 2 : 1); - int out_bytes = out_n * bytes; - stbi_uc * final; - int p; - if (!interlaced) - return stbi__create_png_image_raw(a, image_data, image_data_len, out_n, a->s->img_x, a->s->img_y, depth, color); +static int stbi__create_png_image(stbi__png *a, stbi_uc *image_data, stbi__uint32 image_data_len, int out_n, int depth, int color, int interlaced) +{ + int bytes = (depth == 16 ? 2 : 1); + int out_bytes = out_n * bytes; + stbi_uc *final; + int p; + if (!interlaced) + return stbi__create_png_image_raw(a, image_data, image_data_len, out_n, a->s->img_x, a->s->img_y, depth, color); - // de-interlacing - final = (stbi_uc *)stbi__malloc_mad3(a->s->img_x, a->s->img_y, out_bytes, 0); - if (!final) - return stbi__err("outofmem", "Out of memory"); - for (p = 0; p < 7; ++p) { - int xorig[] = {0, 4, 0, 2, 0, 1, 0}; - int yorig[] = {0, 0, 4, 0, 2, 0, 1}; - int xspc[] = {8, 8, 4, 4, 2, 2, 1}; - int yspc[] = {8, 8, 8, 4, 4, 2, 2}; - int i, j, x, y; - // pass1_x[4] = 0, pass1_x[5] = 1, pass1_x[12] = 1 - x = (a->s->img_x - xorig[p] + xspc[p] - 1) / xspc[p]; - y = (a->s->img_y - yorig[p] + yspc[p] - 1) / yspc[p]; - if (x && y) { - stbi__uint32 img_len = ((((a->s->img_n * x * depth) + 7) >> 3) + 1) * y; - if (!stbi__create_png_image_raw(a, image_data, image_data_len, out_n, x, y, depth, color)) { - STBI_FREE(final); - return 0; + // de-interlacing + final = (stbi_uc *) stbi__malloc_mad3(a->s->img_x, a->s->img_y, out_bytes, 0); + if (!final) return stbi__err("outofmem", "Out of memory"); + for (p=0; p < 7; ++p) { + int xorig[] = { 0,4,0,2,0,1,0 }; + int yorig[] = { 0,0,4,0,2,0,1 }; + int xspc[] = { 8,8,4,4,2,2,1 }; + int yspc[] = { 8,8,8,4,4,2,2 }; + int i,j,x,y; + // pass1_x[4] = 0, pass1_x[5] = 1, pass1_x[12] = 1 + x = (a->s->img_x - xorig[p] + xspc[p]-1) / xspc[p]; + y = (a->s->img_y - yorig[p] + yspc[p]-1) / yspc[p]; + if (x && y) { + stbi__uint32 img_len = ((((a->s->img_n * x * depth) + 7) >> 3) + 1) * y; + if (!stbi__create_png_image_raw(a, image_data, image_data_len, out_n, x, y, depth, color)) { + STBI_FREE(final); + return 0; + } + for (j=0; j < y; ++j) { + for (i=0; i < x; ++i) { + int out_y = j*yspc[p]+yorig[p]; + int out_x = i*xspc[p]+xorig[p]; + memcpy(final + out_y*a->s->img_x*out_bytes + out_x*out_bytes, + a->out + (j*x+i)*out_bytes, out_bytes); } - for (j = 0; j < y; ++j) { - for (i = 0; i < x; ++i) { - int out_y = j * yspc[p] + yorig[p]; - int out_x = i * xspc[p] + xorig[p]; - memcpy(final + out_y * a->s->img_x * out_bytes + out_x * out_bytes, a->out + (j * x + i) * out_bytes, - out_bytes); - } - } - STBI_FREE(a->out); - image_data += img_len; - image_data_len -= img_len; - } - } - a->out = final; + } + STBI_FREE(a->out); + image_data += img_len; + image_data_len -= img_len; + } + } + a->out = final; - return 1; + return 1; } -static int stbi__compute_transparency(stbi__png * z, stbi_uc tc[3], int out_n) { - stbi__context * s = z->s; - stbi__uint32 i, pixel_count = s->img_x * s->img_y; - stbi_uc * p = z->out; +static int stbi__compute_transparency(stbi__png *z, stbi_uc tc[3], int out_n) +{ + stbi__context *s = z->s; + stbi__uint32 i, pixel_count = s->img_x * s->img_y; + stbi_uc *p = z->out; - // compute color-based transparency, assuming we've - // already got 255 as the alpha value in the output - STBI_ASSERT(out_n == 2 || out_n == 4); + // compute color-based transparency, assuming we've + // already got 255 as the alpha value in the output + STBI_ASSERT(out_n == 2 || out_n == 4); - if (out_n == 2) { - for (i = 0; i < pixel_count; ++i) { - p[1] = (p[0] == tc[0] ? 0 : 255); - p += 2; - } - } else { - for (i = 0; i < pixel_count; ++i) { - if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2]) - p[3] = 0; - p += 4; - } - } - return 1; + if (out_n == 2) { + for (i=0; i < pixel_count; ++i) { + p[1] = (p[0] == tc[0] ? 0 : 255); + p += 2; + } + } else { + for (i=0; i < pixel_count; ++i) { + if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2]) + p[3] = 0; + p += 4; + } + } + return 1; } -static int stbi__compute_transparency16(stbi__png * z, stbi__uint16 tc[3], int out_n) { - stbi__context * s = z->s; - stbi__uint32 i, pixel_count = s->img_x * s->img_y; - stbi__uint16 * p = (stbi__uint16 *)z->out; +static int stbi__compute_transparency16(stbi__png *z, stbi__uint16 tc[3], int out_n) +{ + stbi__context *s = z->s; + stbi__uint32 i, pixel_count = s->img_x * s->img_y; + stbi__uint16 *p = (stbi__uint16*) z->out; - // compute color-based transparency, assuming we've - // already got 65535 as the alpha value in the output - STBI_ASSERT(out_n == 2 || out_n == 4); + // compute color-based transparency, assuming we've + // already got 65535 as the alpha value in the output + STBI_ASSERT(out_n == 2 || out_n == 4); - if (out_n == 2) { - for (i = 0; i < pixel_count; ++i) { - p[1] = (p[0] == tc[0] ? 0 : 65535); - p += 2; - } - } else { - for (i = 0; i < pixel_count; ++i) { - if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2]) - p[3] = 0; - p += 4; - } - } - return 1; + if (out_n == 2) { + for (i = 0; i < pixel_count; ++i) { + p[1] = (p[0] == tc[0] ? 0 : 65535); + p += 2; + } + } else { + for (i = 0; i < pixel_count; ++i) { + if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2]) + p[3] = 0; + p += 4; + } + } + return 1; } -static int stbi__expand_png_palette(stbi__png * a, stbi_uc * palette, int len, int pal_img_n) { - stbi__uint32 i, pixel_count = a->s->img_x * a->s->img_y; - stbi_uc *p, *temp_out, *orig = a->out; +static int stbi__expand_png_palette(stbi__png *a, stbi_uc *palette, int len, int pal_img_n) +{ + stbi__uint32 i, pixel_count = a->s->img_x * a->s->img_y; + stbi_uc *p, *temp_out, *orig = a->out; - p = (stbi_uc *)stbi__malloc_mad2(pixel_count, pal_img_n, 0); - if (p == NULL) - return stbi__err("outofmem", "Out of memory"); + p = (stbi_uc *) stbi__malloc_mad2(pixel_count, pal_img_n, 0); + if (p == NULL) return stbi__err("outofmem", "Out of memory"); - // between here and free(out) below, exitting would leak - temp_out = p; + // between here and free(out) below, exitting would leak + temp_out = p; - if (pal_img_n == 3) { - for (i = 0; i < pixel_count; ++i) { - int n = orig[i] * 4; - p[0] = palette[n]; - p[1] = palette[n + 1]; - p[2] = palette[n + 2]; - p += 3; - } - } else { - for (i = 0; i < pixel_count; ++i) { - int n = orig[i] * 4; - p[0] = palette[n]; - p[1] = palette[n + 1]; - p[2] = palette[n + 2]; - p[3] = palette[n + 3]; - p += 4; - } - } - STBI_FREE(a->out); - a->out = temp_out; + if (pal_img_n == 3) { + for (i=0; i < pixel_count; ++i) { + int n = orig[i]*4; + p[0] = palette[n ]; + p[1] = palette[n+1]; + p[2] = palette[n+2]; + p += 3; + } + } else { + for (i=0; i < pixel_count; ++i) { + int n = orig[i]*4; + p[0] = palette[n ]; + p[1] = palette[n+1]; + p[2] = palette[n+2]; + p[3] = palette[n+3]; + p += 4; + } + } + STBI_FREE(a->out); + a->out = temp_out; - STBI_NOTUSED(len); + STBI_NOTUSED(len); - return 1; + return 1; } static int stbi__unpremultiply_on_load_global = 0; static int stbi__de_iphone_flag_global = 0; -STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply) { - stbi__unpremultiply_on_load_global = flag_true_if_should_unpremultiply; +STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply) +{ + stbi__unpremultiply_on_load_global = flag_true_if_should_unpremultiply; } -STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert) { - stbi__de_iphone_flag_global = flag_true_if_should_convert; +STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert) +{ + stbi__de_iphone_flag_global = flag_true_if_should_convert; } #ifndef STBI_THREAD_LOCAL -#define stbi__unpremultiply_on_load stbi__unpremultiply_on_load_global -#define stbi__de_iphone_flag stbi__de_iphone_flag_global +#define stbi__unpremultiply_on_load stbi__unpremultiply_on_load_global +#define stbi__de_iphone_flag stbi__de_iphone_flag_global #else static STBI_THREAD_LOCAL int stbi__unpremultiply_on_load_local, stbi__unpremultiply_on_load_set; static STBI_THREAD_LOCAL int stbi__de_iphone_flag_local, stbi__de_iphone_flag_set; -STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply) { - stbi__unpremultiply_on_load_local = flag_true_if_should_unpremultiply; - stbi__unpremultiply_on_load_set = 1; +STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply) +{ + stbi__unpremultiply_on_load_local = flag_true_if_should_unpremultiply; + stbi__unpremultiply_on_load_set = 1; } -STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert) { - stbi__de_iphone_flag_local = flag_true_if_should_convert; - stbi__de_iphone_flag_set = 1; +STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert) +{ + stbi__de_iphone_flag_local = flag_true_if_should_convert; + stbi__de_iphone_flag_set = 1; } -#define stbi__unpremultiply_on_load \ - (stbi__unpremultiply_on_load_set ? stbi__unpremultiply_on_load_local : stbi__unpremultiply_on_load_global) -#define stbi__de_iphone_flag (stbi__de_iphone_flag_set ? stbi__de_iphone_flag_local : stbi__de_iphone_flag_global) +#define stbi__unpremultiply_on_load (stbi__unpremultiply_on_load_set \ + ? stbi__unpremultiply_on_load_local \ + : stbi__unpremultiply_on_load_global) +#define stbi__de_iphone_flag (stbi__de_iphone_flag_set \ + ? stbi__de_iphone_flag_local \ + : stbi__de_iphone_flag_global) #endif // STBI_THREAD_LOCAL -static void stbi__de_iphone(stbi__png * z) { - stbi__context * s = z->s; - stbi__uint32 i, pixel_count = s->img_x * s->img_y; - stbi_uc * p = z->out; +static void stbi__de_iphone(stbi__png *z) +{ + stbi__context *s = z->s; + stbi__uint32 i, pixel_count = s->img_x * s->img_y; + stbi_uc *p = z->out; - if (s->img_out_n == 3) { // convert bgr to rgb - for (i = 0; i < pixel_count; ++i) { + if (s->img_out_n == 3) { // convert bgr to rgb + for (i=0; i < pixel_count; ++i) { + stbi_uc t = p[0]; + p[0] = p[2]; + p[2] = t; + p += 3; + } + } else { + STBI_ASSERT(s->img_out_n == 4); + if (stbi__unpremultiply_on_load) { + // convert bgr to rgb and unpremultiply + for (i=0; i < pixel_count; ++i) { + stbi_uc a = p[3]; + stbi_uc t = p[0]; + if (a) { + stbi_uc half = a / 2; + p[0] = (p[2] * 255 + half) / a; + p[1] = (p[1] * 255 + half) / a; + p[2] = ( t * 255 + half) / a; + } else { + p[0] = p[2]; + p[2] = t; + } + p += 4; + } + } else { + // convert bgr to rgb + for (i=0; i < pixel_count; ++i) { stbi_uc t = p[0]; p[0] = p[2]; p[2] = t; - p += 3; - } - } else { - STBI_ASSERT(s->img_out_n == 4); - if (stbi__unpremultiply_on_load) { - // convert bgr to rgb and unpremultiply - for (i = 0; i < pixel_count; ++i) { - stbi_uc a = p[3]; - stbi_uc t = p[0]; - if (a) { - stbi_uc half = a / 2; - p[0] = (p[2] * 255 + half) / a; - p[1] = (p[1] * 255 + half) / a; - p[2] = (t * 255 + half) / a; - } else { - p[0] = p[2]; - p[2] = t; - } - p += 4; - } - } else { - // convert bgr to rgb - for (i = 0; i < pixel_count; ++i) { - stbi_uc t = p[0]; - p[0] = p[2]; - p[2] = t; - p += 4; - } - } - } + p += 4; + } + } + } } -#define STBI__PNG_TYPE(a, b, c, d) (((unsigned)(a) << 24) + ((unsigned)(b) << 16) + ((unsigned)(c) << 8) + (unsigned)(d)) +#define STBI__PNG_TYPE(a,b,c,d) (((unsigned) (a) << 24) + ((unsigned) (b) << 16) + ((unsigned) (c) << 8) + (unsigned) (d)) -static int stbi__parse_png_file(stbi__png * z, int scan, int req_comp) { - stbi_uc palette[1024], pal_img_n = 0; - stbi_uc has_trans = 0, tc[3] = {0}; - stbi__uint16 tc16[3]; - stbi__uint32 ioff = 0, idata_limit = 0, i, pal_len = 0; - int first = 1, k, interlace = 0, color = 0, is_iphone = 0; - stbi__context * s = z->s; +static int stbi__parse_png_file(stbi__png *z, int scan, int req_comp) +{ + stbi_uc palette[1024], pal_img_n=0; + stbi_uc has_trans=0, tc[3]={0}; + stbi__uint16 tc16[3]; + stbi__uint32 ioff=0, idata_limit=0, i, pal_len=0; + int first=1,k,interlace=0, color=0, is_iphone=0; + stbi__context *s = z->s; - z->expanded = NULL; - z->idata = NULL; - z->out = NULL; + z->expanded = NULL; + z->idata = NULL; + z->out = NULL; - if (!stbi__check_png_header(s)) - return 0; + if (!stbi__check_png_header(s)) return 0; - if (scan == STBI__SCAN_type) - return 1; + if (scan == STBI__SCAN_type) return 1; - for (;;) { - stbi__pngchunk c = stbi__get_chunk_header(s); - switch (c.type) { - case STBI__PNG_TYPE('C', 'g', 'B', 'I'): + for (;;) { + stbi__pngchunk c = stbi__get_chunk_header(s); + switch (c.type) { + case STBI__PNG_TYPE('C','g','B','I'): is_iphone = 1; stbi__skip(s, c.length); break; - case STBI__PNG_TYPE('I', 'H', 'D', 'R'): { - int comp, filter; - if (!first) - return stbi__err("multiple IHDR", "Corrupt PNG"); + case STBI__PNG_TYPE('I','H','D','R'): { + int comp,filter; + if (!first) return stbi__err("multiple IHDR","Corrupt PNG"); first = 0; - if (c.length != 13) - return stbi__err("bad IHDR len", "Corrupt PNG"); + if (c.length != 13) return stbi__err("bad IHDR len","Corrupt PNG"); s->img_x = stbi__get32be(s); s->img_y = stbi__get32be(s); - if (s->img_y > STBI_MAX_DIMENSIONS) - return stbi__err("too large", "Very large image (corrupt?)"); - if (s->img_x > STBI_MAX_DIMENSIONS) - return stbi__err("too large", "Very large image (corrupt?)"); - z->depth = stbi__get8(s); - if (z->depth != 1 && z->depth != 2 && z->depth != 4 && z->depth != 8 && z->depth != 16) - return stbi__err("1/2/4/8/16-bit only", "PNG not supported: 1/2/4/8/16-bit only"); - color = stbi__get8(s); - if (color > 6) - return stbi__err("bad ctype", "Corrupt PNG"); - if (color == 3 && z->depth == 16) - return stbi__err("bad ctype", "Corrupt PNG"); - if (color == 3) - pal_img_n = 3; - else if (color & 1) - return stbi__err("bad ctype", "Corrupt PNG"); - comp = stbi__get8(s); - if (comp) - return stbi__err("bad comp method", "Corrupt PNG"); - filter = stbi__get8(s); - if (filter) - return stbi__err("bad filter method", "Corrupt PNG"); - interlace = stbi__get8(s); - if (interlace > 1) - return stbi__err("bad interlace method", "Corrupt PNG"); - if (!s->img_x || !s->img_y) - return stbi__err("0-pixel image", "Corrupt PNG"); + if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)"); + if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)"); + z->depth = stbi__get8(s); if (z->depth != 1 && z->depth != 2 && z->depth != 4 && z->depth != 8 && z->depth != 16) return stbi__err("1/2/4/8/16-bit only","PNG not supported: 1/2/4/8/16-bit only"); + color = stbi__get8(s); if (color > 6) return stbi__err("bad ctype","Corrupt PNG"); + if (color == 3 && z->depth == 16) return stbi__err("bad ctype","Corrupt PNG"); + if (color == 3) pal_img_n = 3; else if (color & 1) return stbi__err("bad ctype","Corrupt PNG"); + comp = stbi__get8(s); if (comp) return stbi__err("bad comp method","Corrupt PNG"); + filter= stbi__get8(s); if (filter) return stbi__err("bad filter method","Corrupt PNG"); + interlace = stbi__get8(s); if (interlace>1) return stbi__err("bad interlace method","Corrupt PNG"); + if (!s->img_x || !s->img_y) return stbi__err("0-pixel image","Corrupt PNG"); if (!pal_img_n) { - s->img_n = (color & 2 ? 3 : 1) + (color & 4 ? 1 : 0); - if ((1 << 30) / s->img_x / s->img_n < s->img_y) - return stbi__err("too large", "Image too large to decode"); + s->img_n = (color & 2 ? 3 : 1) + (color & 4 ? 1 : 0); + if ((1 << 30) / s->img_x / s->img_n < s->img_y) return stbi__err("too large", "Image too large to decode"); } else { - // if paletted, then pal_n is our final components, and - // img_n is # components to decompress/filter. - s->img_n = 1; - if ((1 << 30) / s->img_x / 4 < s->img_y) - return stbi__err("too large", "Corrupt PNG"); + // if paletted, then pal_n is our final components, and + // img_n is # components to decompress/filter. + s->img_n = 1; + if ((1 << 30) / s->img_x / 4 < s->img_y) return stbi__err("too large","Corrupt PNG"); } // even with SCAN_header, have to scan to see if we have a tRNS break; - } + } - case STBI__PNG_TYPE('P', 'L', 'T', 'E'): { - if (first) - return stbi__err("first not IHDR", "Corrupt PNG"); - if (c.length > 256 * 3) - return stbi__err("invalid PLTE", "Corrupt PNG"); + case STBI__PNG_TYPE('P','L','T','E'): { + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); + if (c.length > 256*3) return stbi__err("invalid PLTE","Corrupt PNG"); pal_len = c.length / 3; - if (pal_len * 3 != c.length) - return stbi__err("invalid PLTE", "Corrupt PNG"); - for (i = 0; i < pal_len; ++i) { - palette[i * 4 + 0] = stbi__get8(s); - palette[i * 4 + 1] = stbi__get8(s); - palette[i * 4 + 2] = stbi__get8(s); - palette[i * 4 + 3] = 255; + if (pal_len * 3 != c.length) return stbi__err("invalid PLTE","Corrupt PNG"); + for (i=0; i < pal_len; ++i) { + palette[i*4+0] = stbi__get8(s); + palette[i*4+1] = stbi__get8(s); + palette[i*4+2] = stbi__get8(s); + palette[i*4+3] = 255; } break; - } + } - case STBI__PNG_TYPE('t', 'R', 'N', 'S'): { - if (first) - return stbi__err("first not IHDR", "Corrupt PNG"); - if (z->idata) - return stbi__err("tRNS after IDAT", "Corrupt PNG"); + case STBI__PNG_TYPE('t','R','N','S'): { + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); + if (z->idata) return stbi__err("tRNS after IDAT","Corrupt PNG"); if (pal_img_n) { - if (scan == STBI__SCAN_header) { - s->img_n = 4; - return 1; - } - if (pal_len == 0) - return stbi__err("tRNS before PLTE", "Corrupt PNG"); - if (c.length > pal_len) - return stbi__err("bad tRNS len", "Corrupt PNG"); - pal_img_n = 4; - for (i = 0; i < c.length; ++i) - palette[i * 4 + 3] = stbi__get8(s); + if (scan == STBI__SCAN_header) { s->img_n = 4; return 1; } + if (pal_len == 0) return stbi__err("tRNS before PLTE","Corrupt PNG"); + if (c.length > pal_len) return stbi__err("bad tRNS len","Corrupt PNG"); + pal_img_n = 4; + for (i=0; i < c.length; ++i) + palette[i*4+3] = stbi__get8(s); } else { - if (!(s->img_n & 1)) - return stbi__err("tRNS with alpha", "Corrupt PNG"); - if (c.length != (stbi__uint32)s->img_n * 2) - return stbi__err("bad tRNS len", "Corrupt PNG"); - has_trans = 1; - // non-paletted with tRNS = constant alpha. if header-scanning, we can stop now. - if (scan == STBI__SCAN_header) { - ++s->img_n; - return 1; - } - if (z->depth == 16) { - for (k = 0; k < s->img_n; ++k) - tc16[k] = (stbi__uint16)stbi__get16be(s); // copy the values as-is - } else { - for (k = 0; k < s->img_n; ++k) - tc[k] = (stbi_uc)(stbi__get16be(s) & 255) * - stbi__depth_scale_table[z->depth]; // non 8-bit images will be larger - } + if (!(s->img_n & 1)) return stbi__err("tRNS with alpha","Corrupt PNG"); + if (c.length != (stbi__uint32) s->img_n*2) return stbi__err("bad tRNS len","Corrupt PNG"); + has_trans = 1; + // non-paletted with tRNS = constant alpha. if header-scanning, we can stop now. + if (scan == STBI__SCAN_header) { ++s->img_n; return 1; } + if (z->depth == 16) { + for (k = 0; k < s->img_n && k < 3; ++k) // extra loop test to suppress false GCC warning + tc16[k] = (stbi__uint16)stbi__get16be(s); // copy the values as-is + } else { + for (k = 0; k < s->img_n && k < 3; ++k) + tc[k] = (stbi_uc)(stbi__get16be(s) & 255) * stbi__depth_scale_table[z->depth]; // non 8-bit images will be larger + } } break; - } + } - case STBI__PNG_TYPE('I', 'D', 'A', 'T'): { - if (first) - return stbi__err("first not IHDR", "Corrupt PNG"); - if (pal_img_n && !pal_len) - return stbi__err("no PLTE", "Corrupt PNG"); + case STBI__PNG_TYPE('I','D','A','T'): { + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); + if (pal_img_n && !pal_len) return stbi__err("no PLTE","Corrupt PNG"); if (scan == STBI__SCAN_header) { - // header scan definitely stops at first IDAT - if (pal_img_n) - s->img_n = pal_img_n; - return 1; + // header scan definitely stops at first IDAT + if (pal_img_n) + s->img_n = pal_img_n; + return 1; } - if (c.length > (1u << 30)) - return stbi__err("IDAT size limit", "IDAT section larger than 2^30 bytes"); - if ((int)(ioff + c.length) < (int)ioff) - return 0; + if (c.length > (1u << 30)) return stbi__err("IDAT size limit", "IDAT section larger than 2^30 bytes"); + if ((int)(ioff + c.length) < (int)ioff) return 0; if (ioff + c.length > idata_limit) { - stbi__uint32 idata_limit_old = idata_limit; - stbi_uc * p; - if (idata_limit == 0) - idata_limit = c.length > 4096 ? c.length : 4096; - while (ioff + c.length > idata_limit) - idata_limit *= 2; - STBI_NOTUSED(idata_limit_old); - p = (stbi_uc *)STBI_REALLOC_SIZED(z->idata, idata_limit_old, idata_limit); - if (p == NULL) - return stbi__err("outofmem", "Out of memory"); - z->idata = p; + stbi__uint32 idata_limit_old = idata_limit; + stbi_uc *p; + if (idata_limit == 0) idata_limit = c.length > 4096 ? c.length : 4096; + while (ioff + c.length > idata_limit) + idata_limit *= 2; + STBI_NOTUSED(idata_limit_old); + p = (stbi_uc *) STBI_REALLOC_SIZED(z->idata, idata_limit_old, idata_limit); if (p == NULL) return stbi__err("outofmem", "Out of memory"); + z->idata = p; } - if (!stbi__getn(s, z->idata + ioff, c.length)) - return stbi__err("outofdata", "Corrupt PNG"); + if (!stbi__getn(s, z->idata+ioff,c.length)) return stbi__err("outofdata","Corrupt PNG"); ioff += c.length; break; - } + } - case STBI__PNG_TYPE('I', 'E', 'N', 'D'): { + case STBI__PNG_TYPE('I','E','N','D'): { stbi__uint32 raw_len, bpl; - if (first) - return stbi__err("first not IHDR", "Corrupt PNG"); - if (scan != STBI__SCAN_load) - return 1; - if (z->idata == NULL) - return stbi__err("no IDAT", "Corrupt PNG"); + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); + if (scan != STBI__SCAN_load) return 1; + if (z->idata == NULL) return stbi__err("no IDAT","Corrupt PNG"); // initial guess for decoded data size to avoid unnecessary reallocs bpl = (s->img_x * z->depth + 7) / 8; // bytes per line, per component raw_len = bpl * s->img_y * s->img_n /* pixels */ + s->img_y /* filter mode per row */; - z->expanded = (stbi_uc *)stbi_zlib_decode_malloc_guesssize_headerflag((char *)z->idata, ioff, raw_len, - (int *)&raw_len, !is_iphone); - if (z->expanded == NULL) - return 0; // zlib should set error - STBI_FREE(z->idata); - z->idata = NULL; - if ((req_comp == s->img_n + 1 && req_comp != 3 && !pal_img_n) || has_trans) - s->img_out_n = s->img_n + 1; + z->expanded = (stbi_uc *) stbi_zlib_decode_malloc_guesssize_headerflag((char *) z->idata, ioff, raw_len, (int *) &raw_len, !is_iphone); + if (z->expanded == NULL) return 0; // zlib should set error + STBI_FREE(z->idata); z->idata = NULL; + if ((req_comp == s->img_n+1 && req_comp != 3 && !pal_img_n) || has_trans) + s->img_out_n = s->img_n+1; else - s->img_out_n = s->img_n; - if (!stbi__create_png_image(z, z->expanded, raw_len, s->img_out_n, z->depth, color, interlace)) - return 0; + s->img_out_n = s->img_n; + if (!stbi__create_png_image(z, z->expanded, raw_len, s->img_out_n, z->depth, color, interlace)) return 0; if (has_trans) { - if (z->depth == 16) { - if (!stbi__compute_transparency16(z, tc16, s->img_out_n)) - return 0; - } else { - if (!stbi__compute_transparency(z, tc, s->img_out_n)) - return 0; - } + if (z->depth == 16) { + if (!stbi__compute_transparency16(z, tc16, s->img_out_n)) return 0; + } else { + if (!stbi__compute_transparency(z, tc, s->img_out_n)) return 0; + } } if (is_iphone && stbi__de_iphone_flag && s->img_out_n > 2) - stbi__de_iphone(z); + stbi__de_iphone(z); if (pal_img_n) { - // pal_img_n == 3 or 4 - s->img_n = pal_img_n; // record the actual colors we had - s->img_out_n = pal_img_n; - if (req_comp >= 3) - s->img_out_n = req_comp; - if (!stbi__expand_png_palette(z, palette, pal_len, s->img_out_n)) - return 0; + // pal_img_n == 3 or 4 + s->img_n = pal_img_n; // record the actual colors we had + s->img_out_n = pal_img_n; + if (req_comp >= 3) s->img_out_n = req_comp; + if (!stbi__expand_png_palette(z, palette, pal_len, s->img_out_n)) + return 0; } else if (has_trans) { - // non-paletted image with tRNS -> source image has (constant) alpha - ++s->img_n; + // non-paletted image with tRNS -> source image has (constant) alpha + ++s->img_n; } - STBI_FREE(z->expanded); - z->expanded = NULL; + STBI_FREE(z->expanded); z->expanded = NULL; // end of PNG chunk, read and skip CRC stbi__get32be(s); return 1; - } + } - default: + default: // if critical, fail - if (first) - return stbi__err("first not IHDR", "Corrupt PNG"); + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); if ((c.type & (1 << 29)) == 0) { -#ifndef STBI_NO_FAILURE_STRINGS - // not threadsafe - static char invalid_chunk[] = "XXXX PNG chunk not known"; - invalid_chunk[0] = STBI__BYTECAST(c.type >> 24); - invalid_chunk[1] = STBI__BYTECAST(c.type >> 16); - invalid_chunk[2] = STBI__BYTECAST(c.type >> 8); - invalid_chunk[3] = STBI__BYTECAST(c.type >> 0); -#endif - return stbi__err(invalid_chunk, "PNG not supported: unknown PNG chunk type"); + #ifndef STBI_NO_FAILURE_STRINGS + // not threadsafe + static char invalid_chunk[] = "XXXX PNG chunk not known"; + invalid_chunk[0] = STBI__BYTECAST(c.type >> 24); + invalid_chunk[1] = STBI__BYTECAST(c.type >> 16); + invalid_chunk[2] = STBI__BYTECAST(c.type >> 8); + invalid_chunk[3] = STBI__BYTECAST(c.type >> 0); + #endif + return stbi__err(invalid_chunk, "PNG not supported: unknown PNG chunk type"); } stbi__skip(s, c.length); break; - } - // end of PNG chunk, read and skip CRC - stbi__get32be(s); - } + } + // end of PNG chunk, read and skip CRC + stbi__get32be(s); + } } -static void * stbi__do_png(stbi__png * p, int * x, int * y, int * n, int req_comp, stbi__result_info * ri) { - void * result = NULL; - if (req_comp < 0 || req_comp > 4) - return stbi__errpuc("bad req_comp", "Internal error"); - if (stbi__parse_png_file(p, STBI__SCAN_load, req_comp)) { - if (p->depth <= 8) - ri->bits_per_channel = 8; - else if (p->depth == 16) - ri->bits_per_channel = 16; - else - return stbi__errpuc("bad bits_per_channel", "PNG not supported: unsupported color depth"); - result = p->out; - p->out = NULL; - if (req_comp && req_comp != p->s->img_out_n) { - if (ri->bits_per_channel == 8) - result = stbi__convert_format((unsigned char *)result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y); - else - result = stbi__convert_format16((stbi__uint16 *)result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y); - p->s->img_out_n = req_comp; - if (result == NULL) - return result; - } - *x = p->s->img_x; - *y = p->s->img_y; - if (n) - *n = p->s->img_n; - } - STBI_FREE(p->out); - p->out = NULL; - STBI_FREE(p->expanded); - p->expanded = NULL; - STBI_FREE(p->idata); - p->idata = NULL; +static void *stbi__do_png(stbi__png *p, int *x, int *y, int *n, int req_comp, stbi__result_info *ri) +{ + void *result=NULL; + if (req_comp < 0 || req_comp > 4) return stbi__errpuc("bad req_comp", "Internal error"); + if (stbi__parse_png_file(p, STBI__SCAN_load, req_comp)) { + if (p->depth <= 8) + ri->bits_per_channel = 8; + else if (p->depth == 16) + ri->bits_per_channel = 16; + else + return stbi__errpuc("bad bits_per_channel", "PNG not supported: unsupported color depth"); + result = p->out; + p->out = NULL; + if (req_comp && req_comp != p->s->img_out_n) { + if (ri->bits_per_channel == 8) + result = stbi__convert_format((unsigned char *) result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y); + else + result = stbi__convert_format16((stbi__uint16 *) result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y); + p->s->img_out_n = req_comp; + if (result == NULL) return result; + } + *x = p->s->img_x; + *y = p->s->img_y; + if (n) *n = p->s->img_n; + } + STBI_FREE(p->out); p->out = NULL; + STBI_FREE(p->expanded); p->expanded = NULL; + STBI_FREE(p->idata); p->idata = NULL; - return result; + return result; } -static void * stbi__png_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) { - stbi__png p; - p.s = s; - return stbi__do_png(&p, x, y, comp, req_comp, ri); +static void *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + stbi__png p; + p.s = s; + return stbi__do_png(&p, x,y,comp,req_comp, ri); } -static int stbi__png_test(stbi__context * s) { - int r; - r = stbi__check_png_header(s); - stbi__rewind(s); - return r; +static int stbi__png_test(stbi__context *s) +{ + int r; + r = stbi__check_png_header(s); + stbi__rewind(s); + return r; } -static int stbi__png_info_raw(stbi__png * p, int * x, int * y, int * comp) { - if (!stbi__parse_png_file(p, STBI__SCAN_header, 0)) { - stbi__rewind(p->s); - return 0; - } - if (x) - *x = p->s->img_x; - if (y) - *y = p->s->img_y; - if (comp) - *comp = p->s->img_n; - return 1; +static int stbi__png_info_raw(stbi__png *p, int *x, int *y, int *comp) +{ + if (!stbi__parse_png_file(p, STBI__SCAN_header, 0)) { + stbi__rewind( p->s ); + return 0; + } + if (x) *x = p->s->img_x; + if (y) *y = p->s->img_y; + if (comp) *comp = p->s->img_n; + return 1; } -static int stbi__png_info(stbi__context * s, int * x, int * y, int * comp) { - stbi__png p; - p.s = s; - return stbi__png_info_raw(&p, x, y, comp); +static int stbi__png_info(stbi__context *s, int *x, int *y, int *comp) +{ + stbi__png p; + p.s = s; + return stbi__png_info_raw(&p, x, y, comp); } -static int stbi__png_is16(stbi__context * s) { - stbi__png p; - p.s = s; - if (!stbi__png_info_raw(&p, NULL, NULL, NULL)) - return 0; - if (p.depth != 16) { - stbi__rewind(p.s); - return 0; - } - return 1; +static int stbi__png_is16(stbi__context *s) +{ + stbi__png p; + p.s = s; + if (!stbi__png_info_raw(&p, NULL, NULL, NULL)) + return 0; + if (p.depth != 16) { + stbi__rewind(p.s); + return 0; + } + return 1; } #endif // Microsoft/Windows BMP image #ifndef STBI_NO_BMP -static int stbi__bmp_test_raw(stbi__context * s) { - int r; - int sz; - if (stbi__get8(s) != 'B') - return 0; - if (stbi__get8(s) != 'M') - return 0; - stbi__get32le(s); // discard filesize - stbi__get16le(s); // discard reserved - stbi__get16le(s); // discard reserved - stbi__get32le(s); // discard data offset - sz = stbi__get32le(s); - r = (sz == 12 || sz == 40 || sz == 56 || sz == 108 || sz == 124); - return r; +static int stbi__bmp_test_raw(stbi__context *s) +{ + int r; + int sz; + if (stbi__get8(s) != 'B') return 0; + if (stbi__get8(s) != 'M') return 0; + stbi__get32le(s); // discard filesize + stbi__get16le(s); // discard reserved + stbi__get16le(s); // discard reserved + stbi__get32le(s); // discard data offset + sz = stbi__get32le(s); + r = (sz == 12 || sz == 40 || sz == 56 || sz == 108 || sz == 124); + return r; } -static int stbi__bmp_test(stbi__context * s) { - int r = stbi__bmp_test_raw(s); - stbi__rewind(s); - return r; +static int stbi__bmp_test(stbi__context *s) +{ + int r = stbi__bmp_test_raw(s); + stbi__rewind(s); + return r; } + // returns 0..31 for the highest set bit -static int stbi__high_bit(unsigned int z) { - int n = 0; - if (z == 0) - return -1; - if (z >= 0x10000) { - n += 16; - z >>= 16; - } - if (z >= 0x00100) { - n += 8; - z >>= 8; - } - if (z >= 0x00010) { - n += 4; - z >>= 4; - } - if (z >= 0x00004) { - n += 2; - z >>= 2; - } - if (z >= 0x00002) { - n += 1; /* >>= 1;*/ - } - return n; +static int stbi__high_bit(unsigned int z) +{ + int n=0; + if (z == 0) return -1; + if (z >= 0x10000) { n += 16; z >>= 16; } + if (z >= 0x00100) { n += 8; z >>= 8; } + if (z >= 0x00010) { n += 4; z >>= 4; } + if (z >= 0x00004) { n += 2; z >>= 2; } + if (z >= 0x00002) { n += 1;/* >>= 1;*/ } + return n; } -static int stbi__bitcount(unsigned int a) { - a = (a & 0x55555555) + ((a >> 1) & 0x55555555); // max 2 - a = (a & 0x33333333) + ((a >> 2) & 0x33333333); // max 4 - a = (a + (a >> 4)) & 0x0f0f0f0f; // max 8 per 4, now 8 bits - a = (a + (a >> 8)); // max 16 per 8 bits - a = (a + (a >> 16)); // max 32 per 8 bits - return a & 0xff; +static int stbi__bitcount(unsigned int a) +{ + a = (a & 0x55555555) + ((a >> 1) & 0x55555555); // max 2 + a = (a & 0x33333333) + ((a >> 2) & 0x33333333); // max 4 + a = (a + (a >> 4)) & 0x0f0f0f0f; // max 8 per 4, now 8 bits + a = (a + (a >> 8)); // max 16 per 8 bits + a = (a + (a >> 16)); // max 32 per 8 bits + return a & 0xff; } // extract an arbitrarily-aligned N-bit value (N=bits) // from v, and then make it 8-bits long and fractionally // extend it to full full range. -static int stbi__shiftsigned(unsigned int v, int shift, int bits) { - static unsigned int mul_table[9] = { - 0, - 0xff /*0b11111111*/, - 0x55 /*0b01010101*/, - 0x49 /*0b01001001*/, - 0x11 /*0b00010001*/, - 0x21 /*0b00100001*/, - 0x41 /*0b01000001*/, - 0x81 /*0b10000001*/, - 0x01 /*0b00000001*/, - }; - static unsigned int shift_table[9] = { - 0, 0, 0, 1, 0, 2, 4, 6, 0, - }; - if (shift < 0) - v <<= -shift; - else - v >>= shift; - STBI_ASSERT(v < 256); - v >>= (8 - bits); - STBI_ASSERT(bits >= 0 && bits <= 8); - return (int)((unsigned)v * mul_table[bits]) >> shift_table[bits]; +static int stbi__shiftsigned(unsigned int v, int shift, int bits) +{ + static unsigned int mul_table[9] = { + 0, + 0xff/*0b11111111*/, 0x55/*0b01010101*/, 0x49/*0b01001001*/, 0x11/*0b00010001*/, + 0x21/*0b00100001*/, 0x41/*0b01000001*/, 0x81/*0b10000001*/, 0x01/*0b00000001*/, + }; + static unsigned int shift_table[9] = { + 0, 0,0,1,0,2,4,6,0, + }; + if (shift < 0) + v <<= -shift; + else + v >>= shift; + STBI_ASSERT(v < 256); + v >>= (8-bits); + STBI_ASSERT(bits >= 0 && bits <= 8); + return (int) ((unsigned) v * mul_table[bits]) >> shift_table[bits]; } -typedef struct { - int bpp, offset, hsz; - unsigned int mr, mg, mb, ma, all_a; - int extra_read; +typedef struct +{ + int bpp, offset, hsz; + unsigned int mr,mg,mb,ma, all_a; + int extra_read; } stbi__bmp_data; -static int stbi__bmp_set_mask_defaults(stbi__bmp_data * info, int compress) { - // BI_BITFIELDS specifies masks explicitly, don't override - if (compress == 3) - return 1; +static int stbi__bmp_set_mask_defaults(stbi__bmp_data *info, int compress) +{ + // BI_BITFIELDS specifies masks explicitly, don't override + if (compress == 3) + return 1; - if (compress == 0) { - if (info->bpp == 16) { - info->mr = 31u << 10; - info->mg = 31u << 5; - info->mb = 31u << 0; - } else if (info->bpp == 32) { - info->mr = 0xffu << 16; - info->mg = 0xffu << 8; - info->mb = 0xffu << 0; - info->ma = 0xffu << 24; - info->all_a = 0; // if all_a is 0 at end, then we loaded alpha channel but it was all 0 - } else { - // otherwise, use defaults, which is all-0 - info->mr = info->mg = info->mb = info->ma = 0; - } - return 1; - } - return 0; // error + if (compress == 0) { + if (info->bpp == 16) { + info->mr = 31u << 10; + info->mg = 31u << 5; + info->mb = 31u << 0; + } else if (info->bpp == 32) { + info->mr = 0xffu << 16; + info->mg = 0xffu << 8; + info->mb = 0xffu << 0; + info->ma = 0xffu << 24; + info->all_a = 0; // if all_a is 0 at end, then we loaded alpha channel but it was all 0 + } else { + // otherwise, use defaults, which is all-0 + info->mr = info->mg = info->mb = info->ma = 0; + } + return 1; + } + return 0; // error } -static void * stbi__bmp_parse_header(stbi__context * s, stbi__bmp_data * info) { - int hsz; - if (stbi__get8(s) != 'B' || stbi__get8(s) != 'M') - return stbi__errpuc("not BMP", "Corrupt BMP"); - stbi__get32le(s); // discard filesize - stbi__get16le(s); // discard reserved - stbi__get16le(s); // discard reserved - info->offset = stbi__get32le(s); - info->hsz = hsz = stbi__get32le(s); - info->mr = info->mg = info->mb = info->ma = 0; - info->extra_read = 14; +static void *stbi__bmp_parse_header(stbi__context *s, stbi__bmp_data *info) +{ + int hsz; + if (stbi__get8(s) != 'B' || stbi__get8(s) != 'M') return stbi__errpuc("not BMP", "Corrupt BMP"); + stbi__get32le(s); // discard filesize + stbi__get16le(s); // discard reserved + stbi__get16le(s); // discard reserved + info->offset = stbi__get32le(s); + info->hsz = hsz = stbi__get32le(s); + info->mr = info->mg = info->mb = info->ma = 0; + info->extra_read = 14; - if (info->offset < 0) - return stbi__errpuc("bad BMP", "bad BMP"); + if (info->offset < 0) return stbi__errpuc("bad BMP", "bad BMP"); - if (hsz != 12 && hsz != 40 && hsz != 56 && hsz != 108 && hsz != 124) - return stbi__errpuc("unknown BMP", "BMP type not supported: unknown"); - if (hsz == 12) { - s->img_x = stbi__get16le(s); - s->img_y = stbi__get16le(s); - } else { - s->img_x = stbi__get32le(s); - s->img_y = stbi__get32le(s); - } - if (stbi__get16le(s) != 1) - return stbi__errpuc("bad BMP", "bad BMP"); - info->bpp = stbi__get16le(s); - if (hsz != 12) { - int compress = stbi__get32le(s); - if (compress == 1 || compress == 2) - return stbi__errpuc("BMP RLE", "BMP type not supported: RLE"); - if (compress >= 4) - return stbi__errpuc("BMP JPEG/PNG", - "BMP type not supported: unsupported compression"); // this includes PNG/JPEG modes - if (compress == 3 && info->bpp != 16 && info->bpp != 32) - return stbi__errpuc("bad BMP", "bad BMP"); // bitfields requires 16 or 32 bits/pixel - stbi__get32le(s); // discard sizeof - stbi__get32le(s); // discard hres - stbi__get32le(s); // discard vres - stbi__get32le(s); // discard colorsused - stbi__get32le(s); // discard max important - if (hsz == 40 || hsz == 56) { - if (hsz == 56) { - stbi__get32le(s); - stbi__get32le(s); - stbi__get32le(s); - stbi__get32le(s); - } - if (info->bpp == 16 || info->bpp == 32) { - if (compress == 0) { - stbi__bmp_set_mask_defaults(info, compress); - } else if (compress == 3) { - info->mr = stbi__get32le(s); - info->mg = stbi__get32le(s); - info->mb = stbi__get32le(s); - info->extra_read += 12; - // not documented, but generated by photoshop and handled by mspaint - if (info->mr == info->mg && info->mg == info->mb) { - // ?!?!? - return stbi__errpuc("bad BMP", "bad BMP"); - } - } else - return stbi__errpuc("bad BMP", "bad BMP"); - } - } else { - // V4/V5 header - int i; - if (hsz != 108 && hsz != 124) - return stbi__errpuc("bad BMP", "bad BMP"); - info->mr = stbi__get32le(s); - info->mg = stbi__get32le(s); - info->mb = stbi__get32le(s); - info->ma = stbi__get32le(s); - if (compress != 3) // override mr/mg/mb unless in BI_BITFIELDS mode, as per docs - stbi__bmp_set_mask_defaults(info, compress); - stbi__get32le(s); // discard color space - for (i = 0; i < 12; ++i) - stbi__get32le(s); // discard color space parameters - if (hsz == 124) { - stbi__get32le(s); // discard rendering intent - stbi__get32le(s); // discard offset of profile data - stbi__get32le(s); // discard size of profile data - stbi__get32le(s); // discard reserved - } - } - } - return (void *)1; + if (hsz != 12 && hsz != 40 && hsz != 56 && hsz != 108 && hsz != 124) return stbi__errpuc("unknown BMP", "BMP type not supported: unknown"); + if (hsz == 12) { + s->img_x = stbi__get16le(s); + s->img_y = stbi__get16le(s); + } else { + s->img_x = stbi__get32le(s); + s->img_y = stbi__get32le(s); + } + if (stbi__get16le(s) != 1) return stbi__errpuc("bad BMP", "bad BMP"); + info->bpp = stbi__get16le(s); + if (hsz != 12) { + int compress = stbi__get32le(s); + if (compress == 1 || compress == 2) return stbi__errpuc("BMP RLE", "BMP type not supported: RLE"); + if (compress >= 4) return stbi__errpuc("BMP JPEG/PNG", "BMP type not supported: unsupported compression"); // this includes PNG/JPEG modes + if (compress == 3 && info->bpp != 16 && info->bpp != 32) return stbi__errpuc("bad BMP", "bad BMP"); // bitfields requires 16 or 32 bits/pixel + stbi__get32le(s); // discard sizeof + stbi__get32le(s); // discard hres + stbi__get32le(s); // discard vres + stbi__get32le(s); // discard colorsused + stbi__get32le(s); // discard max important + if (hsz == 40 || hsz == 56) { + if (hsz == 56) { + stbi__get32le(s); + stbi__get32le(s); + stbi__get32le(s); + stbi__get32le(s); + } + if (info->bpp == 16 || info->bpp == 32) { + if (compress == 0) { + stbi__bmp_set_mask_defaults(info, compress); + } else if (compress == 3) { + info->mr = stbi__get32le(s); + info->mg = stbi__get32le(s); + info->mb = stbi__get32le(s); + info->extra_read += 12; + // not documented, but generated by photoshop and handled by mspaint + if (info->mr == info->mg && info->mg == info->mb) { + // ?!?!? + return stbi__errpuc("bad BMP", "bad BMP"); + } + } else + return stbi__errpuc("bad BMP", "bad BMP"); + } + } else { + // V4/V5 header + int i; + if (hsz != 108 && hsz != 124) + return stbi__errpuc("bad BMP", "bad BMP"); + info->mr = stbi__get32le(s); + info->mg = stbi__get32le(s); + info->mb = stbi__get32le(s); + info->ma = stbi__get32le(s); + if (compress != 3) // override mr/mg/mb unless in BI_BITFIELDS mode, as per docs + stbi__bmp_set_mask_defaults(info, compress); + stbi__get32le(s); // discard color space + for (i=0; i < 12; ++i) + stbi__get32le(s); // discard color space parameters + if (hsz == 124) { + stbi__get32le(s); // discard rendering intent + stbi__get32le(s); // discard offset of profile data + stbi__get32le(s); // discard size of profile data + stbi__get32le(s); // discard reserved + } + } + } + return (void *) 1; } -static void * stbi__bmp_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) { - stbi_uc * out; - unsigned int mr = 0, mg = 0, mb = 0, ma = 0, all_a; - stbi_uc pal[256][4]; - int psize = 0, i, j, width; - int flip_vertically, pad, target; - stbi__bmp_data info; - STBI_NOTUSED(ri); - info.all_a = 255; - if (stbi__bmp_parse_header(s, &info) == NULL) - return NULL; // error code already set +static void *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + stbi_uc *out; + unsigned int mr=0,mg=0,mb=0,ma=0, all_a; + stbi_uc pal[256][4]; + int psize=0,i,j,width; + int flip_vertically, pad, target; + stbi__bmp_data info; + STBI_NOTUSED(ri); - flip_vertically = ((int)s->img_y) > 0; - s->img_y = abs((int)s->img_y); + info.all_a = 255; + if (stbi__bmp_parse_header(s, &info) == NULL) + return NULL; // error code already set - if (s->img_y > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); - if (s->img_x > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); + flip_vertically = ((int) s->img_y) > 0; + s->img_y = abs((int) s->img_y); - mr = info.mr; - mg = info.mg; - mb = info.mb; - ma = info.ma; - all_a = info.all_a; + if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); + if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); - if (info.hsz == 12) { - if (info.bpp < 24) - psize = (info.offset - info.extra_read - 24) / 3; - } else { - if (info.bpp < 16) - psize = (info.offset - info.extra_read - info.hsz) >> 2; - } - if (psize == 0) { - // accept some number of extra bytes after the header, but if the offset points either to before - // the header ends or implies a large amount of extra data, reject the file as malformed - int bytes_read_so_far = s->callback_already_read + (int)(s->img_buffer - s->img_buffer_original); - int header_limit = 1024; // max we actually read is below 256 bytes currently. - int extra_data_limit = 256 * 4; // what ordinarily goes here is a palette; 256 entries*4 bytes is its max size. - if (bytes_read_so_far <= 0 || bytes_read_so_far > header_limit) { - return stbi__errpuc("bad header", "Corrupt BMP"); - } - // we established that bytes_read_so_far is positive and sensible. - // the first half of this test rejects offsets that are either too small positives, or - // negative, and guarantees that info.offset >= bytes_read_so_far > 0. this in turn - // ensures the number computed in the second half of the test can't overflow. - if (info.offset < bytes_read_so_far || info.offset - bytes_read_so_far > extra_data_limit) { - return stbi__errpuc("bad offset", "Corrupt BMP"); - } else { - stbi__skip(s, info.offset - bytes_read_so_far); - } - } + mr = info.mr; + mg = info.mg; + mb = info.mb; + ma = info.ma; + all_a = info.all_a; - if (info.bpp == 24 && ma == 0xff000000) - s->img_n = 3; - else - s->img_n = ma ? 4 : 3; - if (req_comp && req_comp >= 3) // we can directly decode 3 or 4 - target = req_comp; - else - target = s->img_n; // if they want monochrome, we'll post-convert + if (info.hsz == 12) { + if (info.bpp < 24) + psize = (info.offset - info.extra_read - 24) / 3; + } else { + if (info.bpp < 16) + psize = (info.offset - info.extra_read - info.hsz) >> 2; + } + if (psize == 0) { + // accept some number of extra bytes after the header, but if the offset points either to before + // the header ends or implies a large amount of extra data, reject the file as malformed + int bytes_read_so_far = s->callback_already_read + (int)(s->img_buffer - s->img_buffer_original); + int header_limit = 1024; // max we actually read is below 256 bytes currently. + int extra_data_limit = 256*4; // what ordinarily goes here is a palette; 256 entries*4 bytes is its max size. + if (bytes_read_so_far <= 0 || bytes_read_so_far > header_limit) { + return stbi__errpuc("bad header", "Corrupt BMP"); + } + // we established that bytes_read_so_far is positive and sensible. + // the first half of this test rejects offsets that are either too small positives, or + // negative, and guarantees that info.offset >= bytes_read_so_far > 0. this in turn + // ensures the number computed in the second half of the test can't overflow. + if (info.offset < bytes_read_so_far || info.offset - bytes_read_so_far > extra_data_limit) { + return stbi__errpuc("bad offset", "Corrupt BMP"); + } else { + stbi__skip(s, info.offset - bytes_read_so_far); + } + } - // sanity-check size - if (!stbi__mad3sizes_valid(target, s->img_x, s->img_y, 0)) - return stbi__errpuc("too large", "Corrupt BMP"); + if (info.bpp == 24 && ma == 0xff000000) + s->img_n = 3; + else + s->img_n = ma ? 4 : 3; + if (req_comp && req_comp >= 3) // we can directly decode 3 or 4 + target = req_comp; + else + target = s->img_n; // if they want monochrome, we'll post-convert - out = (stbi_uc *)stbi__malloc_mad3(target, s->img_x, s->img_y, 0); - if (!out) - return stbi__errpuc("outofmem", "Out of memory"); - if (info.bpp < 16) { - int z = 0; - if (psize == 0 || psize > 256) { - STBI_FREE(out); - return stbi__errpuc("invalid", "Corrupt BMP"); - } - for (i = 0; i < psize; ++i) { - pal[i][2] = stbi__get8(s); - pal[i][1] = stbi__get8(s); - pal[i][0] = stbi__get8(s); - if (info.hsz != 12) - stbi__get8(s); - pal[i][3] = 255; - } - stbi__skip(s, info.offset - info.extra_read - info.hsz - psize * (info.hsz == 12 ? 3 : 4)); - if (info.bpp == 1) - width = (s->img_x + 7) >> 3; - else if (info.bpp == 4) - width = (s->img_x + 1) >> 1; - else if (info.bpp == 8) - width = s->img_x; - else { - STBI_FREE(out); - return stbi__errpuc("bad bpp", "Corrupt BMP"); - } - pad = (-width) & 3; - if (info.bpp == 1) { - for (j = 0; j < (int)s->img_y; ++j) { - int bit_offset = 7, v = stbi__get8(s); - for (i = 0; i < (int)s->img_x; ++i) { - int color = (v >> bit_offset) & 0x1; - out[z++] = pal[color][0]; - out[z++] = pal[color][1]; - out[z++] = pal[color][2]; - if (target == 4) - out[z++] = 255; - if (i + 1 == (int)s->img_x) - break; - if ((--bit_offset) < 0) { - bit_offset = 7; - v = stbi__get8(s); - } - } - stbi__skip(s, pad); - } - } else { - for (j = 0; j < (int)s->img_y; ++j) { - for (i = 0; i < (int)s->img_x; i += 2) { - int v = stbi__get8(s), v2 = 0; - if (info.bpp == 4) { - v2 = v & 15; - v >>= 4; - } - out[z++] = pal[v][0]; - out[z++] = pal[v][1]; - out[z++] = pal[v][2]; - if (target == 4) - out[z++] = 255; - if (i + 1 == (int)s->img_x) - break; - v = (info.bpp == 8) ? stbi__get8(s) : v2; - out[z++] = pal[v][0]; - out[z++] = pal[v][1]; - out[z++] = pal[v][2]; - if (target == 4) - out[z++] = 255; - } - stbi__skip(s, pad); - } - } - } else { - int rshift = 0, gshift = 0, bshift = 0, ashift = 0, rcount = 0, gcount = 0, bcount = 0, acount = 0; - int z = 0; - int easy = 0; - stbi__skip(s, info.offset - info.extra_read - info.hsz); - if (info.bpp == 24) - width = 3 * s->img_x; - else if (info.bpp == 16) - width = 2 * s->img_x; - else /* bpp = 32 and pad = 0 */ - width = 0; - pad = (-width) & 3; - if (info.bpp == 24) { - easy = 1; - } else if (info.bpp == 32) { - if (mb == 0xff && mg == 0xff00 && mr == 0x00ff0000 && ma == 0xff000000) - easy = 2; - } - if (!easy) { - if (!mr || !mg || !mb) { - STBI_FREE(out); - return stbi__errpuc("bad masks", "Corrupt BMP"); - } - // right shift amt to put high bit in position #7 - rshift = stbi__high_bit(mr) - 7; - rcount = stbi__bitcount(mr); - gshift = stbi__high_bit(mg) - 7; - gcount = stbi__bitcount(mg); - bshift = stbi__high_bit(mb) - 7; - bcount = stbi__bitcount(mb); - ashift = stbi__high_bit(ma) - 7; - acount = stbi__bitcount(ma); - if (rcount > 8 || gcount > 8 || bcount > 8 || acount > 8) { - STBI_FREE(out); - return stbi__errpuc("bad masks", "Corrupt BMP"); - } - } - for (j = 0; j < (int)s->img_y; ++j) { - if (easy) { - for (i = 0; i < (int)s->img_x; ++i) { - unsigned char a; - out[z + 2] = stbi__get8(s); - out[z + 1] = stbi__get8(s); - out[z + 0] = stbi__get8(s); - z += 3; - a = (easy == 2 ? stbi__get8(s) : 255); - all_a |= a; - if (target == 4) - out[z++] = a; - } - } else { - int bpp = info.bpp; - for (i = 0; i < (int)s->img_x; ++i) { - stbi__uint32 v = (bpp == 16 ? (stbi__uint32)stbi__get16le(s) : stbi__get32le(s)); - unsigned int a; - out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mr, rshift, rcount)); - out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mg, gshift, gcount)); - out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mb, bshift, bcount)); - a = (ma ? stbi__shiftsigned(v & ma, ashift, acount) : 255); - all_a |= a; - if (target == 4) - out[z++] = STBI__BYTECAST(a); - } + // sanity-check size + if (!stbi__mad3sizes_valid(target, s->img_x, s->img_y, 0)) + return stbi__errpuc("too large", "Corrupt BMP"); + + out = (stbi_uc *) stbi__malloc_mad3(target, s->img_x, s->img_y, 0); + if (!out) return stbi__errpuc("outofmem", "Out of memory"); + if (info.bpp < 16) { + int z=0; + if (psize == 0 || psize > 256) { STBI_FREE(out); return stbi__errpuc("invalid", "Corrupt BMP"); } + for (i=0; i < psize; ++i) { + pal[i][2] = stbi__get8(s); + pal[i][1] = stbi__get8(s); + pal[i][0] = stbi__get8(s); + if (info.hsz != 12) stbi__get8(s); + pal[i][3] = 255; + } + stbi__skip(s, info.offset - info.extra_read - info.hsz - psize * (info.hsz == 12 ? 3 : 4)); + if (info.bpp == 1) width = (s->img_x + 7) >> 3; + else if (info.bpp == 4) width = (s->img_x + 1) >> 1; + else if (info.bpp == 8) width = s->img_x; + else { STBI_FREE(out); return stbi__errpuc("bad bpp", "Corrupt BMP"); } + pad = (-width)&3; + if (info.bpp == 1) { + for (j=0; j < (int) s->img_y; ++j) { + int bit_offset = 7, v = stbi__get8(s); + for (i=0; i < (int) s->img_x; ++i) { + int color = (v>>bit_offset)&0x1; + out[z++] = pal[color][0]; + out[z++] = pal[color][1]; + out[z++] = pal[color][2]; + if (target == 4) out[z++] = 255; + if (i+1 == (int) s->img_x) break; + if((--bit_offset) < 0) { + bit_offset = 7; + v = stbi__get8(s); + } } stbi__skip(s, pad); - } - } - - // if alpha channel is all 0s, replace with all 255s - if (target == 4 && all_a == 0) - for (i = 4 * s->img_x * s->img_y - 1; i >= 0; i -= 4) - out[i] = 255; - - if (flip_vertically) { - stbi_uc t; - for (j = 0; j < (int)s->img_y >> 1; ++j) { - stbi_uc * p1 = out + j * s->img_x * target; - stbi_uc * p2 = out + (s->img_y - 1 - j) * s->img_x * target; - for (i = 0; i < (int)s->img_x * target; ++i) { - t = p1[i]; - p1[i] = p2[i]; - p2[i] = t; + } + } else { + for (j=0; j < (int) s->img_y; ++j) { + for (i=0; i < (int) s->img_x; i += 2) { + int v=stbi__get8(s),v2=0; + if (info.bpp == 4) { + v2 = v & 15; + v >>= 4; + } + out[z++] = pal[v][0]; + out[z++] = pal[v][1]; + out[z++] = pal[v][2]; + if (target == 4) out[z++] = 255; + if (i+1 == (int) s->img_x) break; + v = (info.bpp == 8) ? stbi__get8(s) : v2; + out[z++] = pal[v][0]; + out[z++] = pal[v][1]; + out[z++] = pal[v][2]; + if (target == 4) out[z++] = 255; } - } - } + stbi__skip(s, pad); + } + } + } else { + int rshift=0,gshift=0,bshift=0,ashift=0,rcount=0,gcount=0,bcount=0,acount=0; + int z = 0; + int easy=0; + stbi__skip(s, info.offset - info.extra_read - info.hsz); + if (info.bpp == 24) width = 3 * s->img_x; + else if (info.bpp == 16) width = 2*s->img_x; + else /* bpp = 32 and pad = 0 */ width=0; + pad = (-width) & 3; + if (info.bpp == 24) { + easy = 1; + } else if (info.bpp == 32) { + if (mb == 0xff && mg == 0xff00 && mr == 0x00ff0000 && ma == 0xff000000) + easy = 2; + } + if (!easy) { + if (!mr || !mg || !mb) { STBI_FREE(out); return stbi__errpuc("bad masks", "Corrupt BMP"); } + // right shift amt to put high bit in position #7 + rshift = stbi__high_bit(mr)-7; rcount = stbi__bitcount(mr); + gshift = stbi__high_bit(mg)-7; gcount = stbi__bitcount(mg); + bshift = stbi__high_bit(mb)-7; bcount = stbi__bitcount(mb); + ashift = stbi__high_bit(ma)-7; acount = stbi__bitcount(ma); + if (rcount > 8 || gcount > 8 || bcount > 8 || acount > 8) { STBI_FREE(out); return stbi__errpuc("bad masks", "Corrupt BMP"); } + } + for (j=0; j < (int) s->img_y; ++j) { + if (easy) { + for (i=0; i < (int) s->img_x; ++i) { + unsigned char a; + out[z+2] = stbi__get8(s); + out[z+1] = stbi__get8(s); + out[z+0] = stbi__get8(s); + z += 3; + a = (easy == 2 ? stbi__get8(s) : 255); + all_a |= a; + if (target == 4) out[z++] = a; + } + } else { + int bpp = info.bpp; + for (i=0; i < (int) s->img_x; ++i) { + stbi__uint32 v = (bpp == 16 ? (stbi__uint32) stbi__get16le(s) : stbi__get32le(s)); + unsigned int a; + out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mr, rshift, rcount)); + out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mg, gshift, gcount)); + out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mb, bshift, bcount)); + a = (ma ? stbi__shiftsigned(v & ma, ashift, acount) : 255); + all_a |= a; + if (target == 4) out[z++] = STBI__BYTECAST(a); + } + } + stbi__skip(s, pad); + } + } - if (req_comp && req_comp != target) { - out = stbi__convert_format(out, target, req_comp, s->img_x, s->img_y); - if (out == NULL) - return out; // stbi__convert_format frees input on failure - } + // if alpha channel is all 0s, replace with all 255s + if (target == 4 && all_a == 0) + for (i=4*s->img_x*s->img_y-1; i >= 0; i -= 4) + out[i] = 255; - *x = s->img_x; - *y = s->img_y; - if (comp) - *comp = s->img_n; - return out; + if (flip_vertically) { + stbi_uc t; + for (j=0; j < (int) s->img_y>>1; ++j) { + stbi_uc *p1 = out + j *s->img_x*target; + stbi_uc *p2 = out + (s->img_y-1-j)*s->img_x*target; + for (i=0; i < (int) s->img_x*target; ++i) { + t = p1[i]; p1[i] = p2[i]; p2[i] = t; + } + } + } + + if (req_comp && req_comp != target) { + out = stbi__convert_format(out, target, req_comp, s->img_x, s->img_y); + if (out == NULL) return out; // stbi__convert_format frees input on failure + } + + *x = s->img_x; + *y = s->img_y; + if (comp) *comp = s->img_n; + return out; } #endif @@ -6100,74 +5736,68 @@ static void * stbi__bmp_load(stbi__context * s, int * x, int * y, int * comp, in // by Jonathan Dummer #ifndef STBI_NO_TGA // returns STBI_rgb or whatever, 0 on error -static int stbi__tga_get_comp(int bits_per_pixel, int is_grey, int * is_rgb16) { - // only RGB or RGBA (incl. 16bit) or grey allowed - if (is_rgb16) - *is_rgb16 = 0; - switch (bits_per_pixel) { - case 8: - return STBI_grey; - case 16: - if (is_grey) - return STBI_grey_alpha; - // fallthrough - case 15: - if (is_rgb16) - *is_rgb16 = 1; - return STBI_rgb; - case 24: // fallthrough - case 32: - return bits_per_pixel / 8; - default: - return 0; - } +static int stbi__tga_get_comp(int bits_per_pixel, int is_grey, int* is_rgb16) +{ + // only RGB or RGBA (incl. 16bit) or grey allowed + if (is_rgb16) *is_rgb16 = 0; + switch(bits_per_pixel) { + case 8: return STBI_grey; + case 16: if(is_grey) return STBI_grey_alpha; + // fallthrough + case 15: if(is_rgb16) *is_rgb16 = 1; + return STBI_rgb; + case 24: // fallthrough + case 32: return bits_per_pixel/8; + default: return 0; + } } -static int stbi__tga_info(stbi__context * s, int * x, int * y, int * comp) { +static int stbi__tga_info(stbi__context *s, int *x, int *y, int *comp) +{ int tga_w, tga_h, tga_comp, tga_image_type, tga_bits_per_pixel, tga_colormap_bpp; int sz, tga_colormap_type; - stbi__get8(s); // discard Offset + stbi__get8(s); // discard Offset tga_colormap_type = stbi__get8(s); // colormap type - if (tga_colormap_type > 1) { + if( tga_colormap_type > 1 ) { stbi__rewind(s); - return 0; // only RGB or indexed allowed + return 0; // only RGB or indexed allowed } tga_image_type = stbi__get8(s); // image type - if (tga_colormap_type == 1) { // colormapped (paletted) image + if ( tga_colormap_type == 1 ) { // colormapped (paletted) image if (tga_image_type != 1 && tga_image_type != 9) { stbi__rewind(s); return 0; } - stbi__skip(s, 4); // skip index of first colormap entry and number of entries - sz = stbi__get8(s); // check bits per palette color entry - if ((sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32)) { + stbi__skip(s,4); // skip index of first colormap entry and number of entries + sz = stbi__get8(s); // check bits per palette color entry + if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) { stbi__rewind(s); return 0; } - stbi__skip(s, 4); // skip image x and y origin + stbi__skip(s,4); // skip image x and y origin tga_colormap_bpp = sz; } else { // "normal" image w/o colormap - only RGB or grey allowed, +/- RLE - if ((tga_image_type != 2) && (tga_image_type != 3) && (tga_image_type != 10) && (tga_image_type != 11)) { + if ( (tga_image_type != 2) && (tga_image_type != 3) && (tga_image_type != 10) && (tga_image_type != 11) ) { stbi__rewind(s); return 0; // only RGB or grey allowed, +/- RLE } - stbi__skip(s, 9); // skip colormap specification and image x/y origin + stbi__skip(s,9); // skip colormap specification and image x/y origin tga_colormap_bpp = 0; } tga_w = stbi__get16le(s); - if (tga_w < 1) { + if( tga_w < 1 ) { stbi__rewind(s); - return 0; // test width + return 0; // test width } tga_h = stbi__get16le(s); - if (tga_h < 1) { + if( tga_h < 1 ) { stbi__rewind(s); - return 0; // test height + return 0; // test height } tga_bits_per_pixel = stbi__get8(s); // bits per pixel - stbi__get8(s); // ignore alpha bits + stbi__get8(s); // ignore alpha bits if (tga_colormap_bpp != 0) { - if ((tga_bits_per_pixel != 8) && (tga_bits_per_pixel != 16)) { + if((tga_bits_per_pixel != 8) && (tga_bits_per_pixel != 16)) { // when using a colormap, tga_bits_per_pixel is the size of the indexes // I don't think anything but 8 or 16bit indexes makes sense stbi__rewind(s); @@ -6177,268 +5807,270 @@ static int stbi__tga_info(stbi__context * s, int * x, int * y, int * comp) { } else { tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3) || (tga_image_type == 11), NULL); } - if (!tga_comp) { - stbi__rewind(s); - return 0; + if(!tga_comp) { + stbi__rewind(s); + return 0; } - if (x) - *x = tga_w; - if (y) - *y = tga_h; - if (comp) - *comp = tga_comp; - return 1; // seems to have passed everything + if (x) *x = tga_w; + if (y) *y = tga_h; + if (comp) *comp = tga_comp; + return 1; // seems to have passed everything } -static int stbi__tga_test(stbi__context * s) { - int res = 0; - int sz, tga_color_type; - stbi__get8(s); // discard Offset - tga_color_type = stbi__get8(s); // color type - if (tga_color_type > 1) - goto errorEnd; // only RGB or indexed allowed - sz = stbi__get8(s); // image type - if (tga_color_type == 1) { // colormapped (paletted) image - if (sz != 1 && sz != 9) - goto errorEnd; // colortype 1 demands image type 1 or 9 - stbi__skip(s, 4); // skip index of first colormap entry and number of entries - sz = stbi__get8(s); // check bits per palette color entry - if ((sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32)) - goto errorEnd; - stbi__skip(s, 4); // skip image x and y origin - } else { // "normal" image w/o colormap - if ((sz != 2) && (sz != 3) && (sz != 10) && (sz != 11)) - goto errorEnd; // only RGB or grey allowed, +/- RLE - stbi__skip(s, 9); // skip colormap specification and image x/y origin - } - if (stbi__get16le(s) < 1) - goto errorEnd; // test width - if (stbi__get16le(s) < 1) - goto errorEnd; // test height - sz = stbi__get8(s); // bits per pixel - if ((tga_color_type == 1) && (sz != 8) && (sz != 16)) - goto errorEnd; // for colormapped images, bpp is size of an index - if ((sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32)) - goto errorEnd; +static int stbi__tga_test(stbi__context *s) +{ + int res = 0; + int sz, tga_color_type; + stbi__get8(s); // discard Offset + tga_color_type = stbi__get8(s); // color type + if ( tga_color_type > 1 ) goto errorEnd; // only RGB or indexed allowed + sz = stbi__get8(s); // image type + if ( tga_color_type == 1 ) { // colormapped (paletted) image + if (sz != 1 && sz != 9) goto errorEnd; // colortype 1 demands image type 1 or 9 + stbi__skip(s,4); // skip index of first colormap entry and number of entries + sz = stbi__get8(s); // check bits per palette color entry + if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) goto errorEnd; + stbi__skip(s,4); // skip image x and y origin + } else { // "normal" image w/o colormap + if ( (sz != 2) && (sz != 3) && (sz != 10) && (sz != 11) ) goto errorEnd; // only RGB or grey allowed, +/- RLE + stbi__skip(s,9); // skip colormap specification and image x/y origin + } + if ( stbi__get16le(s) < 1 ) goto errorEnd; // test width + if ( stbi__get16le(s) < 1 ) goto errorEnd; // test height + sz = stbi__get8(s); // bits per pixel + if ( (tga_color_type == 1) && (sz != 8) && (sz != 16) ) goto errorEnd; // for colormapped images, bpp is size of an index + if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) goto errorEnd; - res = 1; // if we got this far, everything's good and we can return 1 instead of 0 + res = 1; // if we got this far, everything's good and we can return 1 instead of 0 errorEnd: - stbi__rewind(s); - return res; + stbi__rewind(s); + return res; } // read 16bit value and convert to 24bit RGB -static void stbi__tga_read_rgb16(stbi__context * s, stbi_uc * out) { - stbi__uint16 px = (stbi__uint16)stbi__get16le(s); - stbi__uint16 fiveBitMask = 31; - // we have 3 channels with 5bits each - int r = (px >> 10) & fiveBitMask; - int g = (px >> 5) & fiveBitMask; - int b = px & fiveBitMask; - // Note that this saves the data in RGB(A) order, so it doesn't need to be swapped later - out[0] = (stbi_uc)((r * 255) / 31); - out[1] = (stbi_uc)((g * 255) / 31); - out[2] = (stbi_uc)((b * 255) / 31); +static void stbi__tga_read_rgb16(stbi__context *s, stbi_uc* out) +{ + stbi__uint16 px = (stbi__uint16)stbi__get16le(s); + stbi__uint16 fiveBitMask = 31; + // we have 3 channels with 5bits each + int r = (px >> 10) & fiveBitMask; + int g = (px >> 5) & fiveBitMask; + int b = px & fiveBitMask; + // Note that this saves the data in RGB(A) order, so it doesn't need to be swapped later + out[0] = (stbi_uc)((r * 255)/31); + out[1] = (stbi_uc)((g * 255)/31); + out[2] = (stbi_uc)((b * 255)/31); - // some people claim that the most significant bit might be used for alpha - // (possibly if an alpha-bit is set in the "image descriptor byte") - // but that only made 16bit test images completely translucent.. - // so let's treat all 15 and 16bit TGAs as RGB with no alpha. + // some people claim that the most significant bit might be used for alpha + // (possibly if an alpha-bit is set in the "image descriptor byte") + // but that only made 16bit test images completely translucent.. + // so let's treat all 15 and 16bit TGAs as RGB with no alpha. } -static void * stbi__tga_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) { - // read in the TGA header stuff - int tga_offset = stbi__get8(s); - int tga_indexed = stbi__get8(s); - int tga_image_type = stbi__get8(s); - int tga_is_RLE = 0; - int tga_palette_start = stbi__get16le(s); - int tga_palette_len = stbi__get16le(s); - int tga_palette_bits = stbi__get8(s); - int tga_x_origin = stbi__get16le(s); - int tga_y_origin = stbi__get16le(s); - int tga_width = stbi__get16le(s); - int tga_height = stbi__get16le(s); - int tga_bits_per_pixel = stbi__get8(s); - int tga_comp, tga_rgb16 = 0; - int tga_inverted = stbi__get8(s); - // int tga_alpha_bits = tga_inverted & 15; // the 4 lowest bits - unused (useless?) - // image data - unsigned char * tga_data; - unsigned char * tga_palette = NULL; - int i, j; - unsigned char raw_data[4] = {0}; - int RLE_count = 0; - int RLE_repeating = 0; - int read_next_pixel = 1; - STBI_NOTUSED(ri); - STBI_NOTUSED(tga_x_origin); // @TODO - STBI_NOTUSED(tga_y_origin); // @TODO +static void *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + // read in the TGA header stuff + int tga_offset = stbi__get8(s); + int tga_indexed = stbi__get8(s); + int tga_image_type = stbi__get8(s); + int tga_is_RLE = 0; + int tga_palette_start = stbi__get16le(s); + int tga_palette_len = stbi__get16le(s); + int tga_palette_bits = stbi__get8(s); + int tga_x_origin = stbi__get16le(s); + int tga_y_origin = stbi__get16le(s); + int tga_width = stbi__get16le(s); + int tga_height = stbi__get16le(s); + int tga_bits_per_pixel = stbi__get8(s); + int tga_comp, tga_rgb16=0; + int tga_inverted = stbi__get8(s); + // int tga_alpha_bits = tga_inverted & 15; // the 4 lowest bits - unused (useless?) + // image data + unsigned char *tga_data; + unsigned char *tga_palette = NULL; + int i, j; + unsigned char raw_data[4] = {0}; + int RLE_count = 0; + int RLE_repeating = 0; + int read_next_pixel = 1; + STBI_NOTUSED(ri); + STBI_NOTUSED(tga_x_origin); // @TODO + STBI_NOTUSED(tga_y_origin); // @TODO - if (tga_height > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); - if (tga_width > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); + if (tga_height > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); + if (tga_width > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); - // do a tiny bit of precessing - if (tga_image_type >= 8) { - tga_image_type -= 8; - tga_is_RLE = 1; - } - tga_inverted = 1 - ((tga_inverted >> 5) & 1); + // do a tiny bit of precessing + if ( tga_image_type >= 8 ) + { + tga_image_type -= 8; + tga_is_RLE = 1; + } + tga_inverted = 1 - ((tga_inverted >> 5) & 1); - // If I'm paletted, then I'll use the number of bits from the palette - if (tga_indexed) - tga_comp = stbi__tga_get_comp(tga_palette_bits, 0, &tga_rgb16); - else - tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3), &tga_rgb16); + // If I'm paletted, then I'll use the number of bits from the palette + if ( tga_indexed ) tga_comp = stbi__tga_get_comp(tga_palette_bits, 0, &tga_rgb16); + else tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3), &tga_rgb16); - if (!tga_comp) // shouldn't really happen, stbi__tga_test() should have ensured basic consistency - return stbi__errpuc("bad format", "Can't find out TGA pixelformat"); + if(!tga_comp) // shouldn't really happen, stbi__tga_test() should have ensured basic consistency + return stbi__errpuc("bad format", "Can't find out TGA pixelformat"); - // tga info - *x = tga_width; - *y = tga_height; - if (comp) - *comp = tga_comp; + // tga info + *x = tga_width; + *y = tga_height; + if (comp) *comp = tga_comp; - if (!stbi__mad3sizes_valid(tga_width, tga_height, tga_comp, 0)) - return stbi__errpuc("too large", "Corrupt TGA"); + if (!stbi__mad3sizes_valid(tga_width, tga_height, tga_comp, 0)) + return stbi__errpuc("too large", "Corrupt TGA"); - tga_data = (unsigned char *)stbi__malloc_mad3(tga_width, tga_height, tga_comp, 0); - if (!tga_data) - return stbi__errpuc("outofmem", "Out of memory"); + tga_data = (unsigned char*)stbi__malloc_mad3(tga_width, tga_height, tga_comp, 0); + if (!tga_data) return stbi__errpuc("outofmem", "Out of memory"); - // skip to the data's starting position (offset usually = 0) - stbi__skip(s, tga_offset); + // skip to the data's starting position (offset usually = 0) + stbi__skip(s, tga_offset ); - if (!tga_indexed && !tga_is_RLE && !tga_rgb16) { - for (i = 0; i < tga_height; ++i) { - int row = tga_inverted ? tga_height - i - 1 : i; - stbi_uc * tga_row = tga_data + row * tga_width * tga_comp; - stbi__getn(s, tga_row, tga_width * tga_comp); - } - } else { - // do I need to load a palette? - if (tga_indexed) { - if (tga_palette_len == 0) { /* you have to have at least one entry! */ - STBI_FREE(tga_data); - return stbi__errpuc("bad palette", "Corrupt TGA"); + if ( !tga_indexed && !tga_is_RLE && !tga_rgb16 ) { + for (i=0; i < tga_height; ++i) { + int row = tga_inverted ? tga_height -i - 1 : i; + stbi_uc *tga_row = tga_data + row*tga_width*tga_comp; + stbi__getn(s, tga_row, tga_width * tga_comp); + } + } else { + // do I need to load a palette? + if ( tga_indexed) + { + if (tga_palette_len == 0) { /* you have to have at least one entry! */ + STBI_FREE(tga_data); + return stbi__errpuc("bad palette", "Corrupt TGA"); + } + + // any data to skip? (offset usually = 0) + stbi__skip(s, tga_palette_start ); + // load the palette + tga_palette = (unsigned char*)stbi__malloc_mad2(tga_palette_len, tga_comp, 0); + if (!tga_palette) { + STBI_FREE(tga_data); + return stbi__errpuc("outofmem", "Out of memory"); + } + if (tga_rgb16) { + stbi_uc *pal_entry = tga_palette; + STBI_ASSERT(tga_comp == STBI_rgb); + for (i=0; i < tga_palette_len; ++i) { + stbi__tga_read_rgb16(s, pal_entry); + pal_entry += tga_comp; } - - // any data to skip? (offset usually = 0) - stbi__skip(s, tga_palette_start); - // load the palette - tga_palette = (unsigned char *)stbi__malloc_mad2(tga_palette_len, tga_comp, 0); - if (!tga_palette) { - STBI_FREE(tga_data); - return stbi__errpuc("outofmem", "Out of memory"); + } else if (!stbi__getn(s, tga_palette, tga_palette_len * tga_comp)) { + STBI_FREE(tga_data); + STBI_FREE(tga_palette); + return stbi__errpuc("bad palette", "Corrupt TGA"); + } + } + // load the data + for (i=0; i < tga_width * tga_height; ++i) + { + // if I'm in RLE mode, do I need to get a RLE stbi__pngchunk? + if ( tga_is_RLE ) + { + if ( RLE_count == 0 ) + { + // yep, get the next byte as a RLE command + int RLE_cmd = stbi__get8(s); + RLE_count = 1 + (RLE_cmd & 127); + RLE_repeating = RLE_cmd >> 7; + read_next_pixel = 1; + } else if ( !RLE_repeating ) + { + read_next_pixel = 1; } - if (tga_rgb16) { - stbi_uc * pal_entry = tga_palette; - STBI_ASSERT(tga_comp == STBI_rgb); - for (i = 0; i < tga_palette_len; ++i) { - stbi__tga_read_rgb16(s, pal_entry); - pal_entry += tga_comp; - } - } else if (!stbi__getn(s, tga_palette, tga_palette_len * tga_comp)) { - STBI_FREE(tga_data); - STBI_FREE(tga_palette); - return stbi__errpuc("bad palette", "Corrupt TGA"); - } - } - // load the data - for (i = 0; i < tga_width * tga_height; ++i) { - // if I'm in RLE mode, do I need to get a RLE stbi__pngchunk? - if (tga_is_RLE) { - if (RLE_count == 0) { - // yep, get the next byte as a RLE command - int RLE_cmd = stbi__get8(s); - RLE_count = 1 + (RLE_cmd & 127); - RLE_repeating = RLE_cmd >> 7; - read_next_pixel = 1; - } else if (!RLE_repeating) { - read_next_pixel = 1; - } + } else + { + read_next_pixel = 1; + } + // OK, if I need to read a pixel, do it now + if ( read_next_pixel ) + { + // load however much data we did have + if ( tga_indexed ) + { + // read in index, then perform the lookup + int pal_idx = (tga_bits_per_pixel == 8) ? stbi__get8(s) : stbi__get16le(s); + if ( pal_idx >= tga_palette_len ) { + // invalid index + pal_idx = 0; + } + pal_idx *= tga_comp; + for (j = 0; j < tga_comp; ++j) { + raw_data[j] = tga_palette[pal_idx+j]; + } + } else if(tga_rgb16) { + STBI_ASSERT(tga_comp == STBI_rgb); + stbi__tga_read_rgb16(s, raw_data); } else { - read_next_pixel = 1; + // read in the data raw + for (j = 0; j < tga_comp; ++j) { + raw_data[j] = stbi__get8(s); + } } - // OK, if I need to read a pixel, do it now - if (read_next_pixel) { - // load however much data we did have - if (tga_indexed) { - // read in index, then perform the lookup - int pal_idx = (tga_bits_per_pixel == 8) ? stbi__get8(s) : stbi__get16le(s); - if (pal_idx >= tga_palette_len) { - // invalid index - pal_idx = 0; - } - pal_idx *= tga_comp; - for (j = 0; j < tga_comp; ++j) { - raw_data[j] = tga_palette[pal_idx + j]; - } - } else if (tga_rgb16) { - STBI_ASSERT(tga_comp == STBI_rgb); - stbi__tga_read_rgb16(s, raw_data); - } else { - // read in the data raw - for (j = 0; j < tga_comp; ++j) { - raw_data[j] = stbi__get8(s); - } - } - // clear the reading flag for the next pixel - read_next_pixel = 0; - } // end of reading a pixel + // clear the reading flag for the next pixel + read_next_pixel = 0; + } // end of reading a pixel - // copy data - for (j = 0; j < tga_comp; ++j) - tga_data[i * tga_comp + j] = raw_data[j]; + // copy data + for (j = 0; j < tga_comp; ++j) + tga_data[i*tga_comp+j] = raw_data[j]; - // in case we're in RLE mode, keep counting down - --RLE_count; - } - // do I need to invert the image? - if (tga_inverted) { - for (j = 0; j * 2 < tga_height; ++j) { - int index1 = j * tga_width * tga_comp; - int index2 = (tga_height - 1 - j) * tga_width * tga_comp; - for (i = tga_width * tga_comp; i > 0; --i) { - unsigned char temp = tga_data[index1]; - tga_data[index1] = tga_data[index2]; - tga_data[index2] = temp; - ++index1; - ++index2; - } + // in case we're in RLE mode, keep counting down + --RLE_count; + } + // do I need to invert the image? + if ( tga_inverted ) + { + for (j = 0; j*2 < tga_height; ++j) + { + int index1 = j * tga_width * tga_comp; + int index2 = (tga_height - 1 - j) * tga_width * tga_comp; + for (i = tga_width * tga_comp; i > 0; --i) + { + unsigned char temp = tga_data[index1]; + tga_data[index1] = tga_data[index2]; + tga_data[index2] = temp; + ++index1; + ++index2; } - } - // clear my palette, if I had one - if (tga_palette != NULL) { - STBI_FREE(tga_palette); - } - } + } + } + // clear my palette, if I had one + if ( tga_palette != NULL ) + { + STBI_FREE( tga_palette ); + } + } - // swap RGB - if the source data was RGB16, it already is in the right order - if (tga_comp >= 3 && !tga_rgb16) { - unsigned char * tga_pixel = tga_data; - for (i = 0; i < tga_width * tga_height; ++i) { - unsigned char temp = tga_pixel[0]; - tga_pixel[0] = tga_pixel[2]; - tga_pixel[2] = temp; - tga_pixel += tga_comp; - } - } + // swap RGB - if the source data was RGB16, it already is in the right order + if (tga_comp >= 3 && !tga_rgb16) + { + unsigned char* tga_pixel = tga_data; + for (i=0; i < tga_width * tga_height; ++i) + { + unsigned char temp = tga_pixel[0]; + tga_pixel[0] = tga_pixel[2]; + tga_pixel[2] = temp; + tga_pixel += tga_comp; + } + } - // convert to target component count - if (req_comp && req_comp != tga_comp) - tga_data = stbi__convert_format(tga_data, tga_comp, req_comp, tga_width, tga_height); + // convert to target component count + if (req_comp && req_comp != tga_comp) + tga_data = stbi__convert_format(tga_data, tga_comp, req_comp, tga_width, tga_height); - // the things I do to get rid of an error message, and yet keep - // Microsoft's C compilers happy... [8^( - tga_palette_start = tga_palette_len = tga_palette_bits = tga_x_origin = tga_y_origin = 0; - STBI_NOTUSED(tga_palette_start); - // OK, done - return tga_data; + // the things I do to get rid of an error message, and yet keep + // Microsoft's C compilers happy... [8^( + tga_palette_start = tga_palette_len = tga_palette_bits = + tga_x_origin = tga_y_origin = 0; + STBI_NOTUSED(tga_palette_start); + // OK, done + return tga_data; } #endif @@ -6446,253 +6078,250 @@ static void * stbi__tga_load(stbi__context * s, int * x, int * y, int * comp, in // Photoshop PSD loader -- PD by Thatcher Ulrich, integration by Nicolas Schulz, tweaked by STB #ifndef STBI_NO_PSD -static int stbi__psd_test(stbi__context * s) { - int r = (stbi__get32be(s) == 0x38425053); - stbi__rewind(s); - return r; +static int stbi__psd_test(stbi__context *s) +{ + int r = (stbi__get32be(s) == 0x38425053); + stbi__rewind(s); + return r; } -static int stbi__psd_decode_rle(stbi__context * s, stbi_uc * p, int pixelCount) { - int count, nleft, len; +static int stbi__psd_decode_rle(stbi__context *s, stbi_uc *p, int pixelCount) +{ + int count, nleft, len; - count = 0; - while ((nleft = pixelCount - count) > 0) { - len = stbi__get8(s); - if (len == 128) { - // No-op. - } else if (len < 128) { - // Copy next len+1 bytes literally. - len++; - if (len > nleft) - return 0; // corrupt data - count += len; - while (len) { - *p = stbi__get8(s); - p += 4; - len--; - } - } else if (len > 128) { - stbi_uc val; - // Next -len+1 bytes in the dest are replicated from next source byte. - // (Interpret len as a negative 8-bit int.) - len = 257 - len; - if (len > nleft) - return 0; // corrupt data - val = stbi__get8(s); - count += len; - while (len) { - *p = val; - p += 4; - len--; - } - } - } + count = 0; + while ((nleft = pixelCount - count) > 0) { + len = stbi__get8(s); + if (len == 128) { + // No-op. + } else if (len < 128) { + // Copy next len+1 bytes literally. + len++; + if (len > nleft) return 0; // corrupt data + count += len; + while (len) { + *p = stbi__get8(s); + p += 4; + len--; + } + } else if (len > 128) { + stbi_uc val; + // Next -len+1 bytes in the dest are replicated from next source byte. + // (Interpret len as a negative 8-bit int.) + len = 257 - len; + if (len > nleft) return 0; // corrupt data + val = stbi__get8(s); + count += len; + while (len) { + *p = val; + p += 4; + len--; + } + } + } - return 1; + return 1; } -static void * stbi__psd_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri, int bpc) { - int pixelCount; - int channelCount, compression; - int channel, i; - int bitdepth; - int w, h; - stbi_uc * out; - STBI_NOTUSED(ri); +static void *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc) +{ + int pixelCount; + int channelCount, compression; + int channel, i; + int bitdepth; + int w,h; + stbi_uc *out; + STBI_NOTUSED(ri); - // Check identifier - if (stbi__get32be(s) != 0x38425053) // "8BPS" - return stbi__errpuc("not PSD", "Corrupt PSD image"); + // Check identifier + if (stbi__get32be(s) != 0x38425053) // "8BPS" + return stbi__errpuc("not PSD", "Corrupt PSD image"); - // Check file type version. - if (stbi__get16be(s) != 1) - return stbi__errpuc("wrong version", "Unsupported version of PSD image"); + // Check file type version. + if (stbi__get16be(s) != 1) + return stbi__errpuc("wrong version", "Unsupported version of PSD image"); - // Skip 6 reserved bytes. - stbi__skip(s, 6); + // Skip 6 reserved bytes. + stbi__skip(s, 6 ); - // Read the number of channels (R, G, B, A, etc). - channelCount = stbi__get16be(s); - if (channelCount < 0 || channelCount > 16) - return stbi__errpuc("wrong channel count", "Unsupported number of channels in PSD image"); + // Read the number of channels (R, G, B, A, etc). + channelCount = stbi__get16be(s); + if (channelCount < 0 || channelCount > 16) + return stbi__errpuc("wrong channel count", "Unsupported number of channels in PSD image"); - // Read the rows and columns of the image. - h = stbi__get32be(s); - w = stbi__get32be(s); + // Read the rows and columns of the image. + h = stbi__get32be(s); + w = stbi__get32be(s); - if (h > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); - if (w > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); + if (h > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); + if (w > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); - // Make sure the depth is 8 bits. - bitdepth = stbi__get16be(s); - if (bitdepth != 8 && bitdepth != 16) - return stbi__errpuc("unsupported bit depth", "PSD bit depth is not 8 or 16 bit"); + // Make sure the depth is 8 bits. + bitdepth = stbi__get16be(s); + if (bitdepth != 8 && bitdepth != 16) + return stbi__errpuc("unsupported bit depth", "PSD bit depth is not 8 or 16 bit"); - // Make sure the color mode is RGB. - // Valid options are: - // 0: Bitmap - // 1: Grayscale - // 2: Indexed color - // 3: RGB color - // 4: CMYK color - // 7: Multichannel - // 8: Duotone - // 9: Lab color - if (stbi__get16be(s) != 3) - return stbi__errpuc("wrong color format", "PSD is not in RGB color format"); + // Make sure the color mode is RGB. + // Valid options are: + // 0: Bitmap + // 1: Grayscale + // 2: Indexed color + // 3: RGB color + // 4: CMYK color + // 7: Multichannel + // 8: Duotone + // 9: Lab color + if (stbi__get16be(s) != 3) + return stbi__errpuc("wrong color format", "PSD is not in RGB color format"); - // Skip the Mode Data. (It's the palette for indexed color; other info for other modes.) - stbi__skip(s, stbi__get32be(s)); + // Skip the Mode Data. (It's the palette for indexed color; other info for other modes.) + stbi__skip(s,stbi__get32be(s) ); - // Skip the image resources. (resolution, pen tool paths, etc) - stbi__skip(s, stbi__get32be(s)); + // Skip the image resources. (resolution, pen tool paths, etc) + stbi__skip(s, stbi__get32be(s) ); - // Skip the reserved data. - stbi__skip(s, stbi__get32be(s)); + // Skip the reserved data. + stbi__skip(s, stbi__get32be(s) ); - // Find out if the data is compressed. - // Known values: - // 0: no compression - // 1: RLE compressed - compression = stbi__get16be(s); - if (compression > 1) - return stbi__errpuc("bad compression", "PSD has an unknown compression format"); + // Find out if the data is compressed. + // Known values: + // 0: no compression + // 1: RLE compressed + compression = stbi__get16be(s); + if (compression > 1) + return stbi__errpuc("bad compression", "PSD has an unknown compression format"); - // Check size - if (!stbi__mad3sizes_valid(4, w, h, 0)) - return stbi__errpuc("too large", "Corrupt PSD"); + // Check size + if (!stbi__mad3sizes_valid(4, w, h, 0)) + return stbi__errpuc("too large", "Corrupt PSD"); - // Create the destination image. + // Create the destination image. - if (!compression && bitdepth == 16 && bpc == 16) { - out = (stbi_uc *)stbi__malloc_mad3(8, w, h, 0); - ri->bits_per_channel = 16; - } else - out = (stbi_uc *)stbi__malloc(4 * w * h); + if (!compression && bitdepth == 16 && bpc == 16) { + out = (stbi_uc *) stbi__malloc_mad3(8, w, h, 0); + ri->bits_per_channel = 16; + } else + out = (stbi_uc *) stbi__malloc(4 * w*h); - if (!out) - return stbi__errpuc("outofmem", "Out of memory"); - pixelCount = w * h; + if (!out) return stbi__errpuc("outofmem", "Out of memory"); + pixelCount = w*h; - // Initialize the data to zero. - // memset( out, 0, pixelCount * 4 ); + // Initialize the data to zero. + //memset( out, 0, pixelCount * 4 ); - // Finally, the image data. - if (compression) { - // RLE as used by .PSD and .TIFF - // Loop until you get the number of unpacked bytes you are expecting: - // Read the next source byte into n. - // If n is between 0 and 127 inclusive, copy the next n+1 bytes literally. - // Else if n is between -127 and -1 inclusive, copy the next byte -n+1 times. - // Else if n is 128, noop. - // Endloop + // Finally, the image data. + if (compression) { + // RLE as used by .PSD and .TIFF + // Loop until you get the number of unpacked bytes you are expecting: + // Read the next source byte into n. + // If n is between 0 and 127 inclusive, copy the next n+1 bytes literally. + // Else if n is between -127 and -1 inclusive, copy the next byte -n+1 times. + // Else if n is 128, noop. + // Endloop - // The RLE-compressed data is preceded by a 2-byte data count for each row in the data, - // which we're going to just skip. - stbi__skip(s, h * channelCount * 2); + // The RLE-compressed data is preceded by a 2-byte data count for each row in the data, + // which we're going to just skip. + stbi__skip(s, h * channelCount * 2 ); - // Read the RLE data by channel. - for (channel = 0; channel < 4; channel++) { - stbi_uc * p; + // Read the RLE data by channel. + for (channel = 0; channel < 4; channel++) { + stbi_uc *p; - p = out + channel; - if (channel >= channelCount) { - // Fill this channel with default data. - for (i = 0; i < pixelCount; i++, p += 4) - *p = (channel == 3 ? 255 : 0); + p = out+channel; + if (channel >= channelCount) { + // Fill this channel with default data. + for (i = 0; i < pixelCount; i++, p += 4) + *p = (channel == 3 ? 255 : 0); + } else { + // Read the RLE data. + if (!stbi__psd_decode_rle(s, p, pixelCount)) { + STBI_FREE(out); + return stbi__errpuc("corrupt", "bad RLE data"); + } + } + } + + } else { + // We're at the raw image data. It's each channel in order (Red, Green, Blue, Alpha, ...) + // where each channel consists of an 8-bit (or 16-bit) value for each pixel in the image. + + // Read the data by channel. + for (channel = 0; channel < 4; channel++) { + if (channel >= channelCount) { + // Fill this channel with default data. + if (bitdepth == 16 && bpc == 16) { + stbi__uint16 *q = ((stbi__uint16 *) out) + channel; + stbi__uint16 val = channel == 3 ? 65535 : 0; + for (i = 0; i < pixelCount; i++, q += 4) + *q = val; } else { - // Read the RLE data. - if (!stbi__psd_decode_rle(s, p, pixelCount)) { - STBI_FREE(out); - return stbi__errpuc("corrupt", "bad RLE data"); - } + stbi_uc *p = out+channel; + stbi_uc val = channel == 3 ? 255 : 0; + for (i = 0; i < pixelCount; i++, p += 4) + *p = val; } - } - } else { - // We're at the raw image data. It's each channel in order (Red, Green, Blue, Alpha, ...) - // where each channel consists of an 8-bit (or 16-bit) value for each pixel in the image. - - // Read the data by channel. - for (channel = 0; channel < 4; channel++) { - if (channel >= channelCount) { - // Fill this channel with default data. - if (bitdepth == 16 && bpc == 16) { - stbi__uint16 * q = ((stbi__uint16 *)out) + channel; - stbi__uint16 val = channel == 3 ? 65535 : 0; - for (i = 0; i < pixelCount; i++, q += 4) - *q = val; - } else { - stbi_uc * p = out + channel; - stbi_uc val = channel == 3 ? 255 : 0; - for (i = 0; i < pixelCount; i++, p += 4) - *p = val; - } + } else { + if (ri->bits_per_channel == 16) { // output bpc + stbi__uint16 *q = ((stbi__uint16 *) out) + channel; + for (i = 0; i < pixelCount; i++, q += 4) + *q = (stbi__uint16) stbi__get16be(s); } else { - if (ri->bits_per_channel == 16) { // output bpc - stbi__uint16 * q = ((stbi__uint16 *)out) + channel; - for (i = 0; i < pixelCount; i++, q += 4) - *q = (stbi__uint16)stbi__get16be(s); - } else { - stbi_uc * p = out + channel; - if (bitdepth == 16) { // input bpc - for (i = 0; i < pixelCount; i++, p += 4) - *p = (stbi_uc)(stbi__get16be(s) >> 8); - } else { - for (i = 0; i < pixelCount; i++, p += 4) - *p = stbi__get8(s); - } - } + stbi_uc *p = out+channel; + if (bitdepth == 16) { // input bpc + for (i = 0; i < pixelCount; i++, p += 4) + *p = (stbi_uc) (stbi__get16be(s) >> 8); + } else { + for (i = 0; i < pixelCount; i++, p += 4) + *p = stbi__get8(s); + } } - } - } + } + } + } - // remove weird white matte from PSD - if (channelCount >= 4) { - if (ri->bits_per_channel == 16) { - for (i = 0; i < w * h; ++i) { - stbi__uint16 * pixel = (stbi__uint16 *)out + 4 * i; - if (pixel[3] != 0 && pixel[3] != 65535) { - float a = pixel[3] / 65535.0f; - float ra = 1.0f / a; - float inv_a = 65535.0f * (1 - ra); - pixel[0] = (stbi__uint16)(pixel[0] * ra + inv_a); - pixel[1] = (stbi__uint16)(pixel[1] * ra + inv_a); - pixel[2] = (stbi__uint16)(pixel[2] * ra + inv_a); - } + // remove weird white matte from PSD + if (channelCount >= 4) { + if (ri->bits_per_channel == 16) { + for (i=0; i < w*h; ++i) { + stbi__uint16 *pixel = (stbi__uint16 *) out + 4*i; + if (pixel[3] != 0 && pixel[3] != 65535) { + float a = pixel[3] / 65535.0f; + float ra = 1.0f / a; + float inv_a = 65535.0f * (1 - ra); + pixel[0] = (stbi__uint16) (pixel[0]*ra + inv_a); + pixel[1] = (stbi__uint16) (pixel[1]*ra + inv_a); + pixel[2] = (stbi__uint16) (pixel[2]*ra + inv_a); } - } else { - for (i = 0; i < w * h; ++i) { - unsigned char * pixel = out + 4 * i; - if (pixel[3] != 0 && pixel[3] != 255) { - float a = pixel[3] / 255.0f; - float ra = 1.0f / a; - float inv_a = 255.0f * (1 - ra); - pixel[0] = (unsigned char)(pixel[0] * ra + inv_a); - pixel[1] = (unsigned char)(pixel[1] * ra + inv_a); - pixel[2] = (unsigned char)(pixel[2] * ra + inv_a); - } + } + } else { + for (i=0; i < w*h; ++i) { + unsigned char *pixel = out + 4*i; + if (pixel[3] != 0 && pixel[3] != 255) { + float a = pixel[3] / 255.0f; + float ra = 1.0f / a; + float inv_a = 255.0f * (1 - ra); + pixel[0] = (unsigned char) (pixel[0]*ra + inv_a); + pixel[1] = (unsigned char) (pixel[1]*ra + inv_a); + pixel[2] = (unsigned char) (pixel[2]*ra + inv_a); } - } - } + } + } + } - // convert to desired output format - if (req_comp && req_comp != 4) { - if (ri->bits_per_channel == 16) - out = (stbi_uc *)stbi__convert_format16((stbi__uint16 *)out, 4, req_comp, w, h); - else - out = stbi__convert_format(out, 4, req_comp, w, h); - if (out == NULL) - return out; // stbi__convert_format frees input on failure - } + // convert to desired output format + if (req_comp && req_comp != 4) { + if (ri->bits_per_channel == 16) + out = (stbi_uc *) stbi__convert_format16((stbi__uint16 *) out, 4, req_comp, w, h); + else + out = stbi__convert_format(out, 4, req_comp, w, h); + if (out == NULL) return out; // stbi__convert_format frees input on failure + } - if (comp) - *comp = 4; - *y = h; - *x = w; + if (comp) *comp = 4; + *y = h; + *x = w; - return out; + return out; } #endif @@ -6704,221 +6333,216 @@ static void * stbi__psd_load(stbi__context * s, int * x, int * y, int * comp, in // See http://ozviz.wasp.uwa.edu.au/~pbourke/dataformats/softimagepic/ #ifndef STBI_NO_PIC -static int stbi__pic_is4(stbi__context * s, const char * str) { - int i; - for (i = 0; i < 4; ++i) - if (stbi__get8(s) != (stbi_uc)str[i]) - return 0; +static int stbi__pic_is4(stbi__context *s,const char *str) +{ + int i; + for (i=0; i<4; ++i) + if (stbi__get8(s) != (stbi_uc)str[i]) + return 0; - return 1; + return 1; } -static int stbi__pic_test_core(stbi__context * s) { - int i; +static int stbi__pic_test_core(stbi__context *s) +{ + int i; - if (!stbi__pic_is4(s, "\x53\x80\xF6\x34")) - return 0; + if (!stbi__pic_is4(s,"\x53\x80\xF6\x34")) + return 0; - for (i = 0; i < 84; ++i) - stbi__get8(s); + for(i=0;i<84;++i) + stbi__get8(s); - if (!stbi__pic_is4(s, "PICT")) - return 0; + if (!stbi__pic_is4(s,"PICT")) + return 0; - return 1; + return 1; } -typedef struct { - stbi_uc size, type, channel; +typedef struct +{ + stbi_uc size,type,channel; } stbi__pic_packet; -static stbi_uc * stbi__readval(stbi__context * s, int channel, stbi_uc * dest) { - int mask = 0x80, i; +static stbi_uc *stbi__readval(stbi__context *s, int channel, stbi_uc *dest) +{ + int mask=0x80, i; - for (i = 0; i < 4; ++i, mask >>= 1) { - if (channel & mask) { - if (stbi__at_eof(s)) - return stbi__errpuc("bad file", "PIC file too short"); - dest[i] = stbi__get8(s); - } - } + for (i=0; i<4; ++i, mask>>=1) { + if (channel & mask) { + if (stbi__at_eof(s)) return stbi__errpuc("bad file","PIC file too short"); + dest[i]=stbi__get8(s); + } + } - return dest; + return dest; } -static void stbi__copyval(int channel, stbi_uc * dest, const stbi_uc * src) { - int mask = 0x80, i; +static void stbi__copyval(int channel,stbi_uc *dest,const stbi_uc *src) +{ + int mask=0x80,i; - for (i = 0; i < 4; ++i, mask >>= 1) - if (channel & mask) - dest[i] = src[i]; + for (i=0;i<4; ++i, mask>>=1) + if (channel&mask) + dest[i]=src[i]; } -static stbi_uc * stbi__pic_load_core(stbi__context * s, int width, int height, int * comp, stbi_uc * result) { - int act_comp = 0, num_packets = 0, y, chained; - stbi__pic_packet packets[10]; +static stbi_uc *stbi__pic_load_core(stbi__context *s,int width,int height,int *comp, stbi_uc *result) +{ + int act_comp=0,num_packets=0,y,chained; + stbi__pic_packet packets[10]; - // this will (should...) cater for even some bizarre stuff like having data + // this will (should...) cater for even some bizarre stuff like having data // for the same channel in multiple packets. - do { - stbi__pic_packet * packet; + do { + stbi__pic_packet *packet; - if (num_packets == sizeof(packets) / sizeof(packets[0])) - return stbi__errpuc("bad format", "too many packets"); + if (num_packets==sizeof(packets)/sizeof(packets[0])) + return stbi__errpuc("bad format","too many packets"); - packet = &packets[num_packets++]; + packet = &packets[num_packets++]; - chained = stbi__get8(s); - packet->size = stbi__get8(s); - packet->type = stbi__get8(s); - packet->channel = stbi__get8(s); + chained = stbi__get8(s); + packet->size = stbi__get8(s); + packet->type = stbi__get8(s); + packet->channel = stbi__get8(s); - act_comp |= packet->channel; + act_comp |= packet->channel; - if (stbi__at_eof(s)) - return stbi__errpuc("bad file", "file too short (reading packets)"); - if (packet->size != 8) - return stbi__errpuc("bad format", "packet isn't 8bpp"); - } while (chained); + if (stbi__at_eof(s)) return stbi__errpuc("bad file","file too short (reading packets)"); + if (packet->size != 8) return stbi__errpuc("bad format","packet isn't 8bpp"); + } while (chained); - *comp = (act_comp & 0x10 ? 4 : 3); // has alpha channel? + *comp = (act_comp & 0x10 ? 4 : 3); // has alpha channel? - for (y = 0; y < height; ++y) { - int packet_idx; + for(y=0; ytype) { + switch (packet->type) { default: - return stbi__errpuc("bad format", "packet has bad compression type"); + return stbi__errpuc("bad format","packet has bad compression type"); - case 0: { // uncompressed - int x; + case 0: {//uncompressed + int x; - for (x = 0; x < width; ++x, dest += 4) - if (!stbi__readval(s, packet->channel, dest)) - return 0; - break; + for(x=0;xchannel,dest)) + return 0; + break; } - case 1: // Pure RLE - { - int left = width, i; + case 1://Pure RLE + { + int left=width, i; - while (left > 0) { - stbi_uc count, value[4]; + while (left>0) { + stbi_uc count,value[4]; - count = stbi__get8(s); - if (stbi__at_eof(s)) - return stbi__errpuc("bad file", "file too short (pure read count)"); + count=stbi__get8(s); + if (stbi__at_eof(s)) return stbi__errpuc("bad file","file too short (pure read count)"); - if (count > left) - count = (stbi_uc)left; + if (count > left) + count = (stbi_uc) left; - if (!stbi__readval(s, packet->channel, value)) + if (!stbi__readval(s,packet->channel,value)) return 0; + + for(i=0; ichannel,dest,value); + left -= count; + } + } + break; + + case 2: {//Mixed RLE + int left=width; + while (left>0) { + int count = stbi__get8(s), i; + if (stbi__at_eof(s)) return stbi__errpuc("bad file","file too short (mixed read count)"); + + if (count >= 128) { // Repeated + stbi_uc value[4]; + + if (count==128) + count = stbi__get16be(s); + else + count -= 127; + if (count > left) + return stbi__errpuc("bad file","scanline overrun"); + + if (!stbi__readval(s,packet->channel,value)) return 0; - for (i = 0; i < count; ++i, dest += 4) - stbi__copyval(packet->channel, dest, value); - left -= count; - } - } break; + for(i=0;ichannel,dest,value); + } else { // Raw + ++count; + if (count>left) return stbi__errpuc("bad file","scanline overrun"); - case 2: { // Mixed RLE - int left = width; - while (left > 0) { - int count = stbi__get8(s), i; - if (stbi__at_eof(s)) - return stbi__errpuc("bad file", "file too short (mixed read count)"); - - if (count >= 128) { // Repeated - stbi_uc value[4]; - - if (count == 128) - count = stbi__get16be(s); - else - count -= 127; - if (count > left) - return stbi__errpuc("bad file", "scanline overrun"); - - if (!stbi__readval(s, packet->channel, value)) - return 0; - - for (i = 0; i < count; ++i, dest += 4) - stbi__copyval(packet->channel, dest, value); - } else { // Raw - ++count; - if (count > left) - return stbi__errpuc("bad file", "scanline overrun"); - - for (i = 0; i < count; ++i, dest += 4) - if (!stbi__readval(s, packet->channel, dest)) - return 0; - } - left -= count; - } - break; + for(i=0;ichannel,dest)) + return 0; + } + left-=count; + } + break; } - } - } - } + } + } + } - return result; + return result; } -static void * stbi__pic_load(stbi__context * s, int * px, int * py, int * comp, int req_comp, stbi__result_info * ri) { - stbi_uc * result; - int i, x, y, internal_comp; - STBI_NOTUSED(ri); +static void *stbi__pic_load(stbi__context *s,int *px,int *py,int *comp,int req_comp, stbi__result_info *ri) +{ + stbi_uc *result; + int i, x,y, internal_comp; + STBI_NOTUSED(ri); - if (!comp) - comp = &internal_comp; + if (!comp) comp = &internal_comp; - for (i = 0; i < 92; ++i) - stbi__get8(s); + for (i=0; i<92; ++i) + stbi__get8(s); - x = stbi__get16be(s); - y = stbi__get16be(s); + x = stbi__get16be(s); + y = stbi__get16be(s); - if (y > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); - if (x > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); + if (y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); + if (x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); - if (stbi__at_eof(s)) - return stbi__errpuc("bad file", "file too short (pic header)"); - if (!stbi__mad3sizes_valid(x, y, 4, 0)) - return stbi__errpuc("too large", "PIC image too large to decode"); + if (stbi__at_eof(s)) return stbi__errpuc("bad file","file too short (pic header)"); + if (!stbi__mad3sizes_valid(x, y, 4, 0)) return stbi__errpuc("too large", "PIC image too large to decode"); - stbi__get32be(s); // skip `ratio' - stbi__get16be(s); // skip `fields' - stbi__get16be(s); // skip `pad' + stbi__get32be(s); //skip `ratio' + stbi__get16be(s); //skip `fields' + stbi__get16be(s); //skip `pad' - // intermediate buffer is RGBA - result = (stbi_uc *)stbi__malloc_mad3(x, y, 4, 0); - if (!result) - return stbi__errpuc("outofmem", "Out of memory"); - memset(result, 0xff, x * y * 4); + // intermediate buffer is RGBA + result = (stbi_uc *) stbi__malloc_mad3(x, y, 4, 0); + if (!result) return stbi__errpuc("outofmem", "Out of memory"); + memset(result, 0xff, x*y*4); - if (!stbi__pic_load_core(s, x, y, comp, result)) { - STBI_FREE(result); - result = 0; - } - *px = x; - *py = y; - if (req_comp == 0) - req_comp = *comp; - result = stbi__convert_format(result, 4, req_comp, x, y); + if (!stbi__pic_load_core(s,x,y,comp, result)) { + STBI_FREE(result); + result=0; + } + *px = x; + *py = y; + if (req_comp == 0) req_comp = *comp; + result=stbi__convert_format(result,4,req_comp,x,y); - return result; + return result; } -static int stbi__pic_test(stbi__context * s) { - int r = stbi__pic_test_core(s); - stbi__rewind(s); - return r; +static int stbi__pic_test(stbi__context *s) +{ + int r = stbi__pic_test_core(s); + stbi__rewind(s); + return r; } #endif @@ -6926,968 +6550,931 @@ static int stbi__pic_test(stbi__context * s) { // GIF loader -- public domain by Jean-Marc Lienher -- simplified/shrunk by stb #ifndef STBI_NO_GIF -typedef struct { - stbi__int16 prefix; - stbi_uc first; - stbi_uc suffix; +typedef struct +{ + stbi__int16 prefix; + stbi_uc first; + stbi_uc suffix; } stbi__gif_lzw; -typedef struct { - int w, h; - stbi_uc * out; // output buffer (always 4 components) - stbi_uc * background; // The current "background" as far as a gif is concerned - stbi_uc * history; - int flags, bgindex, ratio, transparent, eflags; - stbi_uc pal[256][4]; - stbi_uc lpal[256][4]; - stbi__gif_lzw codes[8192]; - stbi_uc * color_table; - int parse, step; - int lflags; - int start_x, start_y; - int max_x, max_y; - int cur_x, cur_y; - int line_size; - int delay; +typedef struct +{ + int w,h; + stbi_uc *out; // output buffer (always 4 components) + stbi_uc *background; // The current "background" as far as a gif is concerned + stbi_uc *history; + int flags, bgindex, ratio, transparent, eflags; + stbi_uc pal[256][4]; + stbi_uc lpal[256][4]; + stbi__gif_lzw codes[8192]; + stbi_uc *color_table; + int parse, step; + int lflags; + int start_x, start_y; + int max_x, max_y; + int cur_x, cur_y; + int line_size; + int delay; } stbi__gif; -static int stbi__gif_test_raw(stbi__context * s) { - int sz; - if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8') - return 0; - sz = stbi__get8(s); - if (sz != '9' && sz != '7') - return 0; - if (stbi__get8(s) != 'a') - return 0; - return 1; +static int stbi__gif_test_raw(stbi__context *s) +{ + int sz; + if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8') return 0; + sz = stbi__get8(s); + if (sz != '9' && sz != '7') return 0; + if (stbi__get8(s) != 'a') return 0; + return 1; } -static int stbi__gif_test(stbi__context * s) { - int r = stbi__gif_test_raw(s); - stbi__rewind(s); - return r; +static int stbi__gif_test(stbi__context *s) +{ + int r = stbi__gif_test_raw(s); + stbi__rewind(s); + return r; } -static void stbi__gif_parse_colortable(stbi__context * s, stbi_uc pal[256][4], int num_entries, int transp) { - int i; - for (i = 0; i < num_entries; ++i) { - pal[i][2] = stbi__get8(s); - pal[i][1] = stbi__get8(s); - pal[i][0] = stbi__get8(s); - pal[i][3] = transp == i ? 0 : 255; - } +static void stbi__gif_parse_colortable(stbi__context *s, stbi_uc pal[256][4], int num_entries, int transp) +{ + int i; + for (i=0; i < num_entries; ++i) { + pal[i][2] = stbi__get8(s); + pal[i][1] = stbi__get8(s); + pal[i][0] = stbi__get8(s); + pal[i][3] = transp == i ? 0 : 255; + } } -static int stbi__gif_header(stbi__context * s, stbi__gif * g, int * comp, int is_info) { - stbi_uc version; - if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8') - return stbi__err("not GIF", "Corrupt GIF"); +static int stbi__gif_header(stbi__context *s, stbi__gif *g, int *comp, int is_info) +{ + stbi_uc version; + if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8') + return stbi__err("not GIF", "Corrupt GIF"); - version = stbi__get8(s); - if (version != '7' && version != '9') - return stbi__err("not GIF", "Corrupt GIF"); - if (stbi__get8(s) != 'a') - return stbi__err("not GIF", "Corrupt GIF"); + version = stbi__get8(s); + if (version != '7' && version != '9') return stbi__err("not GIF", "Corrupt GIF"); + if (stbi__get8(s) != 'a') return stbi__err("not GIF", "Corrupt GIF"); - stbi__g_failure_reason = ""; - g->w = stbi__get16le(s); - g->h = stbi__get16le(s); - g->flags = stbi__get8(s); - g->bgindex = stbi__get8(s); - g->ratio = stbi__get8(s); - g->transparent = -1; + stbi__g_failure_reason = ""; + g->w = stbi__get16le(s); + g->h = stbi__get16le(s); + g->flags = stbi__get8(s); + g->bgindex = stbi__get8(s); + g->ratio = stbi__get8(s); + g->transparent = -1; - if (g->w > STBI_MAX_DIMENSIONS) - return stbi__err("too large", "Very large image (corrupt?)"); - if (g->h > STBI_MAX_DIMENSIONS) - return stbi__err("too large", "Very large image (corrupt?)"); + if (g->w > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)"); + if (g->h > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)"); - if (comp != 0) - *comp = 4; // can't actually tell whether it's 3 or 4 until we parse the comments + if (comp != 0) *comp = 4; // can't actually tell whether it's 3 or 4 until we parse the comments - if (is_info) - return 1; + if (is_info) return 1; - if (g->flags & 0x80) - stbi__gif_parse_colortable(s, g->pal, 2 << (g->flags & 7), -1); + if (g->flags & 0x80) + stbi__gif_parse_colortable(s,g->pal, 2 << (g->flags & 7), -1); - return 1; + return 1; } -static int stbi__gif_info_raw(stbi__context * s, int * x, int * y, int * comp) { - stbi__gif * g = (stbi__gif *)stbi__malloc(sizeof(stbi__gif)); - if (!g) - return stbi__err("outofmem", "Out of memory"); - if (!stbi__gif_header(s, g, comp, 1)) { - STBI_FREE(g); - stbi__rewind(s); - return 0; - } - if (x) - *x = g->w; - if (y) - *y = g->h; - STBI_FREE(g); - return 1; +static int stbi__gif_info_raw(stbi__context *s, int *x, int *y, int *comp) +{ + stbi__gif* g = (stbi__gif*) stbi__malloc(sizeof(stbi__gif)); + if (!g) return stbi__err("outofmem", "Out of memory"); + if (!stbi__gif_header(s, g, comp, 1)) { + STBI_FREE(g); + stbi__rewind( s ); + return 0; + } + if (x) *x = g->w; + if (y) *y = g->h; + STBI_FREE(g); + return 1; } -static void stbi__out_gif_code(stbi__gif * g, stbi__uint16 code) { - stbi_uc *p, *c; - int idx; +static void stbi__out_gif_code(stbi__gif *g, stbi__uint16 code) +{ + stbi_uc *p, *c; + int idx; - // recurse to decode the prefixes, since the linked-list is backwards, - // and working backwards through an interleaved image would be nasty - if (g->codes[code].prefix >= 0) - stbi__out_gif_code(g, g->codes[code].prefix); + // recurse to decode the prefixes, since the linked-list is backwards, + // and working backwards through an interleaved image would be nasty + if (g->codes[code].prefix >= 0) + stbi__out_gif_code(g, g->codes[code].prefix); - if (g->cur_y >= g->max_y) - return; + if (g->cur_y >= g->max_y) return; - idx = g->cur_x + g->cur_y; - p = &g->out[idx]; - g->history[idx / 4] = 1; + idx = g->cur_x + g->cur_y; + p = &g->out[idx]; + g->history[idx / 4] = 1; - c = &g->color_table[g->codes[code].suffix * 4]; - if (c[3] > 128) { // don't render transparent pixels; - p[0] = c[2]; - p[1] = c[1]; - p[2] = c[0]; - p[3] = c[3]; - } - g->cur_x += 4; + c = &g->color_table[g->codes[code].suffix * 4]; + if (c[3] > 128) { // don't render transparent pixels; + p[0] = c[2]; + p[1] = c[1]; + p[2] = c[0]; + p[3] = c[3]; + } + g->cur_x += 4; - if (g->cur_x >= g->max_x) { - g->cur_x = g->start_x; - g->cur_y += g->step; + if (g->cur_x >= g->max_x) { + g->cur_x = g->start_x; + g->cur_y += g->step; - while (g->cur_y >= g->max_y && g->parse > 0) { - g->step = (1 << g->parse) * g->line_size; - g->cur_y = g->start_y + (g->step >> 1); - --g->parse; - } - } + while (g->cur_y >= g->max_y && g->parse > 0) { + g->step = (1 << g->parse) * g->line_size; + g->cur_y = g->start_y + (g->step >> 1); + --g->parse; + } + } } -static stbi_uc * stbi__process_gif_raster(stbi__context * s, stbi__gif * g) { - stbi_uc lzw_cs; - stbi__int32 len, init_code; - stbi__uint32 first; - stbi__int32 codesize, codemask, avail, oldcode, bits, valid_bits, clear; - stbi__gif_lzw * p; +static stbi_uc *stbi__process_gif_raster(stbi__context *s, stbi__gif *g) +{ + stbi_uc lzw_cs; + stbi__int32 len, init_code; + stbi__uint32 first; + stbi__int32 codesize, codemask, avail, oldcode, bits, valid_bits, clear; + stbi__gif_lzw *p; - lzw_cs = stbi__get8(s); - if (lzw_cs > 12) - return NULL; - clear = 1 << lzw_cs; - first = 1; - codesize = lzw_cs + 1; - codemask = (1 << codesize) - 1; - bits = 0; - valid_bits = 0; - for (init_code = 0; init_code < clear; init_code++) { - g->codes[init_code].prefix = -1; - g->codes[init_code].first = (stbi_uc)init_code; - g->codes[init_code].suffix = (stbi_uc)init_code; - } + lzw_cs = stbi__get8(s); + if (lzw_cs > 12) return NULL; + clear = 1 << lzw_cs; + first = 1; + codesize = lzw_cs + 1; + codemask = (1 << codesize) - 1; + bits = 0; + valid_bits = 0; + for (init_code = 0; init_code < clear; init_code++) { + g->codes[init_code].prefix = -1; + g->codes[init_code].first = (stbi_uc) init_code; + g->codes[init_code].suffix = (stbi_uc) init_code; + } - // support no starting clear code - avail = clear + 2; - oldcode = -1; + // support no starting clear code + avail = clear+2; + oldcode = -1; - len = 0; - for (;;) { - if (valid_bits < codesize) { - if (len == 0) { - len = stbi__get8(s); // start new block - if (len == 0) - return g->out; + len = 0; + for(;;) { + if (valid_bits < codesize) { + if (len == 0) { + len = stbi__get8(s); // start new block + if (len == 0) + return g->out; + } + --len; + bits |= (stbi__int32) stbi__get8(s) << valid_bits; + valid_bits += 8; + } else { + stbi__int32 code = bits & codemask; + bits >>= codesize; + valid_bits -= codesize; + // @OPTIMIZE: is there some way we can accelerate the non-clear path? + if (code == clear) { // clear code + codesize = lzw_cs + 1; + codemask = (1 << codesize) - 1; + avail = clear + 2; + oldcode = -1; + first = 0; + } else if (code == clear + 1) { // end of stream code + stbi__skip(s, len); + while ((len = stbi__get8(s)) > 0) + stbi__skip(s,len); + return g->out; + } else if (code <= avail) { + if (first) { + return stbi__errpuc("no clear code", "Corrupt GIF"); } - --len; - bits |= (stbi__int32)stbi__get8(s) << valid_bits; - valid_bits += 8; - } else { - stbi__int32 code = bits & codemask; - bits >>= codesize; - valid_bits -= codesize; - // @OPTIMIZE: is there some way we can accelerate the non-clear path? - if (code == clear) { // clear code - codesize = lzw_cs + 1; - codemask = (1 << codesize) - 1; - avail = clear + 2; - oldcode = -1; - first = 0; - } else if (code == clear + 1) { // end of stream code - stbi__skip(s, len); - while ((len = stbi__get8(s)) > 0) - stbi__skip(s, len); - return g->out; - } else if (code <= avail) { - if (first) { - return stbi__errpuc("no clear code", "Corrupt GIF"); - } - if (oldcode >= 0) { - p = &g->codes[avail++]; - if (avail > 8192) { - return stbi__errpuc("too many codes", "Corrupt GIF"); - } + if (oldcode >= 0) { + p = &g->codes[avail++]; + if (avail > 8192) { + return stbi__errpuc("too many codes", "Corrupt GIF"); + } - p->prefix = (stbi__int16)oldcode; - p->first = g->codes[oldcode].first; - p->suffix = (code == avail) ? p->first : g->codes[code].first; - } else if (code == avail) - return stbi__errpuc("illegal code in raster", "Corrupt GIF"); + p->prefix = (stbi__int16) oldcode; + p->first = g->codes[oldcode].first; + p->suffix = (code == avail) ? p->first : g->codes[code].first; + } else if (code == avail) + return stbi__errpuc("illegal code in raster", "Corrupt GIF"); - stbi__out_gif_code(g, (stbi__uint16)code); + stbi__out_gif_code(g, (stbi__uint16) code); - if ((avail & codemask) == 0 && avail <= 0x0FFF) { - codesize++; - codemask = (1 << codesize) - 1; - } - - oldcode = code; - } else { - return stbi__errpuc("illegal code in raster", "Corrupt GIF"); + if ((avail & codemask) == 0 && avail <= 0x0FFF) { + codesize++; + codemask = (1 << codesize) - 1; } - } - } + + oldcode = code; + } else { + return stbi__errpuc("illegal code in raster", "Corrupt GIF"); + } + } + } } // this function is designed to support animated gifs, although stb_image doesn't support it // two back is the image from two frames ago, used for a very specific disposal format -static stbi_uc * stbi__gif_load_next(stbi__context * s, stbi__gif * g, int * comp, int req_comp, stbi_uc * two_back) { - int dispose; - int first_frame; - int pi; - int pcount; - STBI_NOTUSED(req_comp); +static stbi_uc *stbi__gif_load_next(stbi__context *s, stbi__gif *g, int *comp, int req_comp, stbi_uc *two_back) +{ + int dispose; + int first_frame; + int pi; + int pcount; + STBI_NOTUSED(req_comp); - // on first frame, any non-written pixels get the background colour (non-transparent) - first_frame = 0; - if (g->out == 0) { - if (!stbi__gif_header(s, g, comp, 0)) - return 0; // stbi__g_failure_reason set by stbi__gif_header - if (!stbi__mad3sizes_valid(4, g->w, g->h, 0)) - return stbi__errpuc("too large", "GIF image is too large"); - pcount = g->w * g->h; - g->out = (stbi_uc *)stbi__malloc(4 * pcount); - g->background = (stbi_uc *)stbi__malloc(4 * pcount); - g->history = (stbi_uc *)stbi__malloc(pcount); - if (!g->out || !g->background || !g->history) - return stbi__errpuc("outofmem", "Out of memory"); + // on first frame, any non-written pixels get the background colour (non-transparent) + first_frame = 0; + if (g->out == 0) { + if (!stbi__gif_header(s, g, comp,0)) return 0; // stbi__g_failure_reason set by stbi__gif_header + if (!stbi__mad3sizes_valid(4, g->w, g->h, 0)) + return stbi__errpuc("too large", "GIF image is too large"); + pcount = g->w * g->h; + g->out = (stbi_uc *) stbi__malloc(4 * pcount); + g->background = (stbi_uc *) stbi__malloc(4 * pcount); + g->history = (stbi_uc *) stbi__malloc(pcount); + if (!g->out || !g->background || !g->history) + return stbi__errpuc("outofmem", "Out of memory"); - // image is treated as "transparent" at the start - ie, nothing overwrites the current background; - // background colour is only used for pixels that are not rendered first frame, after that "background" - // color refers to the color that was there the previous frame. - memset(g->out, 0x00, 4 * pcount); - memset(g->background, 0x00, 4 * pcount); // state of the background (starts transparent) - memset(g->history, 0x00, pcount); // pixels that were affected previous frame - first_frame = 1; - } else { - // second frame - how do we dispose of the previous one? - dispose = (g->eflags & 0x1C) >> 2; - pcount = g->w * g->h; + // image is treated as "transparent" at the start - ie, nothing overwrites the current background; + // background colour is only used for pixels that are not rendered first frame, after that "background" + // color refers to the color that was there the previous frame. + memset(g->out, 0x00, 4 * pcount); + memset(g->background, 0x00, 4 * pcount); // state of the background (starts transparent) + memset(g->history, 0x00, pcount); // pixels that were affected previous frame + first_frame = 1; + } else { + // second frame - how do we dispose of the previous one? + dispose = (g->eflags & 0x1C) >> 2; + pcount = g->w * g->h; - if ((dispose == 3) && (two_back == 0)) { - dispose = 2; // if I don't have an image to revert back to, default to the old background - } + if ((dispose == 3) && (two_back == 0)) { + dispose = 2; // if I don't have an image to revert back to, default to the old background + } - if (dispose == 3) { // use previous graphic - for (pi = 0; pi < pcount; ++pi) { - if (g->history[pi]) { - memcpy(&g->out[pi * 4], &two_back[pi * 4], 4); - } + if (dispose == 3) { // use previous graphic + for (pi = 0; pi < pcount; ++pi) { + if (g->history[pi]) { + memcpy( &g->out[pi * 4], &two_back[pi * 4], 4 ); } - } else if (dispose == 2) { - // restore what was changed last frame to background before that frame; - for (pi = 0; pi < pcount; ++pi) { - if (g->history[pi]) { - memcpy(&g->out[pi * 4], &g->background[pi * 4], 4); - } + } + } else if (dispose == 2) { + // restore what was changed last frame to background before that frame; + for (pi = 0; pi < pcount; ++pi) { + if (g->history[pi]) { + memcpy( &g->out[pi * 4], &g->background[pi * 4], 4 ); } - } else { - // This is a non-disposal case eithe way, so just - // leave the pixels as is, and they will become the new background - // 1: do not dispose - // 0: not specified. - } + } + } else { + // This is a non-disposal case eithe way, so just + // leave the pixels as is, and they will become the new background + // 1: do not dispose + // 0: not specified. + } - // background is what out is after the undoing of the previou frame; - memcpy(g->background, g->out, 4 * g->w * g->h); - } + // background is what out is after the undoing of the previou frame; + memcpy( g->background, g->out, 4 * g->w * g->h ); + } - // clear my history; - memset(g->history, 0x00, g->w * g->h); // pixels that were affected previous frame + // clear my history; + memset( g->history, 0x00, g->w * g->h ); // pixels that were affected previous frame - for (;;) { - int tag = stbi__get8(s); - switch (tag) { - case 0x2C: /* Image Descriptor */ - { + for (;;) { + int tag = stbi__get8(s); + switch (tag) { + case 0x2C: /* Image Descriptor */ + { stbi__int32 x, y, w, h; - stbi_uc * o; + stbi_uc *o; x = stbi__get16le(s); y = stbi__get16le(s); w = stbi__get16le(s); h = stbi__get16le(s); if (((x + w) > (g->w)) || ((y + h) > (g->h))) - return stbi__errpuc("bad Image Descriptor", "Corrupt GIF"); + return stbi__errpuc("bad Image Descriptor", "Corrupt GIF"); g->line_size = g->w * 4; g->start_x = x * 4; g->start_y = y * g->line_size; - g->max_x = g->start_x + w * 4; - g->max_y = g->start_y + h * g->line_size; - g->cur_x = g->start_x; - g->cur_y = g->start_y; + g->max_x = g->start_x + w * 4; + g->max_y = g->start_y + h * g->line_size; + g->cur_x = g->start_x; + g->cur_y = g->start_y; // if the width of the specified rectangle is 0, that means // we may not see *any* pixels or the image is malformed; // to make sure this is caught, move the current y down to // max_y (which is what out_gif_code checks). if (w == 0) - g->cur_y = g->max_y; + g->cur_y = g->max_y; g->lflags = stbi__get8(s); if (g->lflags & 0x40) { - g->step = 8 * g->line_size; // first interlaced spacing - g->parse = 3; + g->step = 8 * g->line_size; // first interlaced spacing + g->parse = 3; } else { - g->step = g->line_size; - g->parse = 0; + g->step = g->line_size; + g->parse = 0; } if (g->lflags & 0x80) { - stbi__gif_parse_colortable(s, g->lpal, 2 << (g->lflags & 7), g->eflags & 0x01 ? g->transparent : -1); - g->color_table = (stbi_uc *)g->lpal; + stbi__gif_parse_colortable(s,g->lpal, 2 << (g->lflags & 7), g->eflags & 0x01 ? g->transparent : -1); + g->color_table = (stbi_uc *) g->lpal; } else if (g->flags & 0x80) { - g->color_table = (stbi_uc *)g->pal; + g->color_table = (stbi_uc *) g->pal; } else - return stbi__errpuc("missing color table", "Corrupt GIF"); + return stbi__errpuc("missing color table", "Corrupt GIF"); o = stbi__process_gif_raster(s, g); - if (!o) - return NULL; + if (!o) return NULL; // if this was the first frame, pcount = g->w * g->h; if (first_frame && (g->bgindex > 0)) { - // if first frame, any pixel not drawn to gets the background color - for (pi = 0; pi < pcount; ++pi) { - if (g->history[pi] == 0) { - g->pal[g->bgindex][3] = - 255; // just in case it was made transparent, undo that; It will be reset next frame if need be; - memcpy(&g->out[pi * 4], &g->pal[g->bgindex], 4); - } - } + // if first frame, any pixel not drawn to gets the background color + for (pi = 0; pi < pcount; ++pi) { + if (g->history[pi] == 0) { + g->pal[g->bgindex][3] = 255; // just in case it was made transparent, undo that; It will be reset next frame if need be; + memcpy( &g->out[pi * 4], &g->pal[g->bgindex], 4 ); + } + } } return o; - } + } - case 0x21: // Comment Extension. - { + case 0x21: // Comment Extension. + { int len; int ext = stbi__get8(s); if (ext == 0xF9) { // Graphic Control Extension. - len = stbi__get8(s); - if (len == 4) { - g->eflags = stbi__get8(s); - g->delay = 10 * stbi__get16le(s); // delay - 1/100th of a second, saving as 1/1000ths. + len = stbi__get8(s); + if (len == 4) { + g->eflags = stbi__get8(s); + g->delay = 10 * stbi__get16le(s); // delay - 1/100th of a second, saving as 1/1000ths. - // unset old transparent - if (g->transparent >= 0) { - g->pal[g->transparent][3] = 255; - } - if (g->eflags & 0x01) { - g->transparent = stbi__get8(s); - if (g->transparent >= 0) { - g->pal[g->transparent][3] = 0; - } - } else { - // don't need transparent - stbi__skip(s, 1); - g->transparent = -1; - } - } else { - stbi__skip(s, len); - break; - } + // unset old transparent + if (g->transparent >= 0) { + g->pal[g->transparent][3] = 255; + } + if (g->eflags & 0x01) { + g->transparent = stbi__get8(s); + if (g->transparent >= 0) { + g->pal[g->transparent][3] = 0; + } + } else { + // don't need transparent + stbi__skip(s, 1); + g->transparent = -1; + } + } else { + stbi__skip(s, len); + break; + } } while ((len = stbi__get8(s)) != 0) { - stbi__skip(s, len); + stbi__skip(s, len); } break; - } + } - case 0x3B: // gif stream termination code - return (stbi_uc *)s; // using '1' causes warning on some compilers + case 0x3B: // gif stream termination code + return (stbi_uc *) s; // using '1' causes warning on some compilers - default: + default: return stbi__errpuc("unknown code", "Corrupt GIF"); - } - } + } + } } -static void * stbi__load_gif_main_outofmem(stbi__gif * g, stbi_uc * out, int ** delays) { - STBI_FREE(g->out); - STBI_FREE(g->history); - STBI_FREE(g->background); +static void *stbi__load_gif_main_outofmem(stbi__gif *g, stbi_uc *out, int **delays) +{ + STBI_FREE(g->out); + STBI_FREE(g->history); + STBI_FREE(g->background); - if (out) - STBI_FREE(out); - if (delays && *delays) - STBI_FREE(*delays); - return stbi__errpuc("outofmem", "Out of memory"); + if (out) STBI_FREE(out); + if (delays && *delays) STBI_FREE(*delays); + return stbi__errpuc("outofmem", "Out of memory"); } -static void * stbi__load_gif_main(stbi__context * s, int ** delays, int * x, int * y, int * z, int * comp, int req_comp) { - if (stbi__gif_test(s)) { - int layers = 0; - stbi_uc * u = 0; - stbi_uc * out = 0; - stbi_uc * two_back = 0; - stbi__gif g; - int stride; - int out_size = 0; - int delays_size = 0; +static void *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y, int *z, int *comp, int req_comp) +{ + if (stbi__gif_test(s)) { + int layers = 0; + stbi_uc *u = 0; + stbi_uc *out = 0; + stbi_uc *two_back = 0; + stbi__gif g; + int stride; + int out_size = 0; + int delays_size = 0; - STBI_NOTUSED(out_size); - STBI_NOTUSED(delays_size); + STBI_NOTUSED(out_size); + STBI_NOTUSED(delays_size); - memset(&g, 0, sizeof(g)); - if (delays) { - *delays = 0; - } + memset(&g, 0, sizeof(g)); + if (delays) { + *delays = 0; + } - do { - u = stbi__gif_load_next(s, &g, comp, req_comp, two_back); - if (u == (stbi_uc *)s) - u = 0; // end of animated gif marker + do { + u = stbi__gif_load_next(s, &g, comp, req_comp, two_back); + if (u == (stbi_uc *) s) u = 0; // end of animated gif marker - if (u) { - *x = g.w; - *y = g.h; - ++layers; - stride = g.w * g.h * 4; + if (u) { + *x = g.w; + *y = g.h; + ++layers; + stride = g.w * g.h * 4; - if (out) { - void * tmp = (stbi_uc *)STBI_REALLOC_SIZED(out, out_size, layers * stride); - if (!tmp) - return stbi__load_gif_main_outofmem(&g, out, delays); - else { - out = (stbi_uc *)tmp; - out_size = layers * stride; - } + if (out) { + void *tmp = (stbi_uc*) STBI_REALLOC_SIZED( out, out_size, layers * stride ); + if (!tmp) + return stbi__load_gif_main_outofmem(&g, out, delays); + else { + out = (stbi_uc*) tmp; + out_size = layers * stride; + } - if (delays) { - int * new_delays = (int *)STBI_REALLOC_SIZED(*delays, delays_size, sizeof(int) * layers); - if (!new_delays) - return stbi__load_gif_main_outofmem(&g, out, delays); - *delays = new_delays; - delays_size = layers * sizeof(int); - } - } else { - out = (stbi_uc *)stbi__malloc(layers * stride); - if (!out) - return stbi__load_gif_main_outofmem(&g, out, delays); - out_size = layers * stride; - if (delays) { - *delays = (int *)stbi__malloc(layers * sizeof(int)); - if (!*delays) - return stbi__load_gif_main_outofmem(&g, out, delays); - delays_size = layers * sizeof(int); - } - } - memcpy(out + ((layers - 1) * stride), u, stride); - if (layers >= 2) { - two_back = out - 2 * stride; - } - - if (delays) { - (*delays)[layers - 1U] = g.delay; - } + if (delays) { + int *new_delays = (int*) STBI_REALLOC_SIZED( *delays, delays_size, sizeof(int) * layers ); + if (!new_delays) + return stbi__load_gif_main_outofmem(&g, out, delays); + *delays = new_delays; + delays_size = layers * sizeof(int); + } + } else { + out = (stbi_uc*)stbi__malloc( layers * stride ); + if (!out) + return stbi__load_gif_main_outofmem(&g, out, delays); + out_size = layers * stride; + if (delays) { + *delays = (int*) stbi__malloc( layers * sizeof(int) ); + if (!*delays) + return stbi__load_gif_main_outofmem(&g, out, delays); + delays_size = layers * sizeof(int); + } + } + memcpy( out + ((layers - 1) * stride), u, stride ); + if (layers >= 2) { + two_back = out - 2 * stride; } - } while (u != 0); - // free temp buffer; - STBI_FREE(g.out); - STBI_FREE(g.history); - STBI_FREE(g.background); + if (delays) { + (*delays)[layers - 1U] = g.delay; + } + } + } while (u != 0); - // do the final conversion after loading everything; - if (req_comp && req_comp != 4) - out = stbi__convert_format(out, 4, req_comp, layers * g.w, g.h); + // free temp buffer; + STBI_FREE(g.out); + STBI_FREE(g.history); + STBI_FREE(g.background); - *z = layers; - return out; - } else { - return stbi__errpuc("not GIF", "Image was not as a gif type."); - } + // do the final conversion after loading everything; + if (req_comp && req_comp != 4) + out = stbi__convert_format(out, 4, req_comp, layers * g.w, g.h); + + *z = layers; + return out; + } else { + return stbi__errpuc("not GIF", "Image was not as a gif type."); + } } -static void * stbi__gif_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) { - stbi_uc * u = 0; - stbi__gif g; - memset(&g, 0, sizeof(g)); - STBI_NOTUSED(ri); +static void *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + stbi_uc *u = 0; + stbi__gif g; + memset(&g, 0, sizeof(g)); + STBI_NOTUSED(ri); - u = stbi__gif_load_next(s, &g, comp, req_comp, 0); - if (u == (stbi_uc *)s) - u = 0; // end of animated gif marker - if (u) { - *x = g.w; - *y = g.h; + u = stbi__gif_load_next(s, &g, comp, req_comp, 0); + if (u == (stbi_uc *) s) u = 0; // end of animated gif marker + if (u) { + *x = g.w; + *y = g.h; - // moved conversion to after successful load so that the same - // can be done for multiple frames. - if (req_comp && req_comp != 4) - u = stbi__convert_format(u, 4, req_comp, g.w, g.h); - } else if (g.out) { - // if there was an error and we allocated an image buffer, free it! - STBI_FREE(g.out); - } + // moved conversion to after successful load so that the same + // can be done for multiple frames. + if (req_comp && req_comp != 4) + u = stbi__convert_format(u, 4, req_comp, g.w, g.h); + } else if (g.out) { + // if there was an error and we allocated an image buffer, free it! + STBI_FREE(g.out); + } - // free buffers needed for multiple frame loading; - STBI_FREE(g.history); - STBI_FREE(g.background); + // free buffers needed for multiple frame loading; + STBI_FREE(g.history); + STBI_FREE(g.background); - return u; + return u; } -static int stbi__gif_info(stbi__context * s, int * x, int * y, int * comp) { return stbi__gif_info_raw(s, x, y, comp); } +static int stbi__gif_info(stbi__context *s, int *x, int *y, int *comp) +{ + return stbi__gif_info_raw(s,x,y,comp); +} #endif // ************************************************************************************************* // Radiance RGBE HDR loader // originally by Nicolas Schulz #ifndef STBI_NO_HDR -static int stbi__hdr_test_core(stbi__context * s, const char * signature) { - int i; - for (i = 0; signature[i]; ++i) - if (stbi__get8(s) != signature[i]) - return 0; - stbi__rewind(s); - return 1; +static int stbi__hdr_test_core(stbi__context *s, const char *signature) +{ + int i; + for (i=0; signature[i]; ++i) + if (stbi__get8(s) != signature[i]) + return 0; + stbi__rewind(s); + return 1; } -static int stbi__hdr_test(stbi__context * s) { - int r = stbi__hdr_test_core(s, "#?RADIANCE\n"); - stbi__rewind(s); - if (!r) { - r = stbi__hdr_test_core(s, "#?RGBE\n"); - stbi__rewind(s); - } - return r; +static int stbi__hdr_test(stbi__context* s) +{ + int r = stbi__hdr_test_core(s, "#?RADIANCE\n"); + stbi__rewind(s); + if(!r) { + r = stbi__hdr_test_core(s, "#?RGBE\n"); + stbi__rewind(s); + } + return r; } -#define STBI__HDR_BUFLEN 1024 -static char * stbi__hdr_gettoken(stbi__context * z, char * buffer) { - int len = 0; - char c = '\0'; +#define STBI__HDR_BUFLEN 1024 +static char *stbi__hdr_gettoken(stbi__context *z, char *buffer) +{ + int len=0; + char c = '\0'; - c = (char)stbi__get8(z); + c = (char) stbi__get8(z); - while (!stbi__at_eof(z) && c != '\n') { - buffer[len++] = c; - if (len == STBI__HDR_BUFLEN - 1) { - // flush to end of line - while (!stbi__at_eof(z) && stbi__get8(z) != '\n') - ; - break; - } - c = (char)stbi__get8(z); - } + while (!stbi__at_eof(z) && c != '\n') { + buffer[len++] = c; + if (len == STBI__HDR_BUFLEN-1) { + // flush to end of line + while (!stbi__at_eof(z) && stbi__get8(z) != '\n') + ; + break; + } + c = (char) stbi__get8(z); + } - buffer[len] = 0; - return buffer; + buffer[len] = 0; + return buffer; } -static void stbi__hdr_convert(float * output, stbi_uc * input, int req_comp) { - if (input[3] != 0) { - float f1; - // Exponent - f1 = (float)ldexp(1.0f, input[3] - (int)(128 + 8)); - if (req_comp <= 2) - output[0] = (input[0] + input[1] + input[2]) * f1 / 3; - else { - output[0] = input[0] * f1; - output[1] = input[1] * f1; - output[2] = input[2] * f1; - } - if (req_comp == 2) - output[1] = 1; - if (req_comp == 4) - output[3] = 1; - } else { - switch (req_comp) { - case 4: - output[3] = 1; /* fallthrough */ - case 3: - output[0] = output[1] = output[2] = 0; - break; - case 2: - output[1] = 1; /* fallthrough */ - case 1: - output[0] = 0; - break; - } - } +static void stbi__hdr_convert(float *output, stbi_uc *input, int req_comp) +{ + if ( input[3] != 0 ) { + float f1; + // Exponent + f1 = (float) ldexp(1.0f, input[3] - (int)(128 + 8)); + if (req_comp <= 2) + output[0] = (input[0] + input[1] + input[2]) * f1 / 3; + else { + output[0] = input[0] * f1; + output[1] = input[1] * f1; + output[2] = input[2] * f1; + } + if (req_comp == 2) output[1] = 1; + if (req_comp == 4) output[3] = 1; + } else { + switch (req_comp) { + case 4: output[3] = 1; /* fallthrough */ + case 3: output[0] = output[1] = output[2] = 0; + break; + case 2: output[1] = 1; /* fallthrough */ + case 1: output[0] = 0; + break; + } + } } -static float * stbi__hdr_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) { - char buffer[STBI__HDR_BUFLEN]; - char * token; - int valid = 0; - int width, height; - stbi_uc * scanline; - float * hdr_data; - int len; - unsigned char count, value; - int i, j, k, c1, c2, z; - const char * headerToken; - STBI_NOTUSED(ri); +static float *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + char buffer[STBI__HDR_BUFLEN]; + char *token; + int valid = 0; + int width, height; + stbi_uc *scanline; + float *hdr_data; + int len; + unsigned char count, value; + int i, j, k, c1,c2, z; + const char *headerToken; + STBI_NOTUSED(ri); - // Check identifier - headerToken = stbi__hdr_gettoken(s, buffer); - if (strcmp(headerToken, "#?RADIANCE") != 0 && strcmp(headerToken, "#?RGBE") != 0) - return stbi__errpf("not HDR", "Corrupt HDR image"); + // Check identifier + headerToken = stbi__hdr_gettoken(s,buffer); + if (strcmp(headerToken, "#?RADIANCE") != 0 && strcmp(headerToken, "#?RGBE") != 0) + return stbi__errpf("not HDR", "Corrupt HDR image"); - // Parse header - for (;;) { - token = stbi__hdr_gettoken(s, buffer); - if (token[0] == 0) - break; - if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) - valid = 1; - } + // Parse header + for(;;) { + token = stbi__hdr_gettoken(s,buffer); + if (token[0] == 0) break; + if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) valid = 1; + } - if (!valid) - return stbi__errpf("unsupported format", "Unsupported HDR format"); + if (!valid) return stbi__errpf("unsupported format", "Unsupported HDR format"); - // Parse width and height - // can't use sscanf() if we're not using stdio! - token = stbi__hdr_gettoken(s, buffer); - if (strncmp(token, "-Y ", 3)) - return stbi__errpf("unsupported data layout", "Unsupported HDR format"); - token += 3; - height = (int)strtol(token, &token, 10); - while (*token == ' ') - ++token; - if (strncmp(token, "+X ", 3)) - return stbi__errpf("unsupported data layout", "Unsupported HDR format"); - token += 3; - width = (int)strtol(token, NULL, 10); + // Parse width and height + // can't use sscanf() if we're not using stdio! + token = stbi__hdr_gettoken(s,buffer); + if (strncmp(token, "-Y ", 3)) return stbi__errpf("unsupported data layout", "Unsupported HDR format"); + token += 3; + height = (int) strtol(token, &token, 10); + while (*token == ' ') ++token; + if (strncmp(token, "+X ", 3)) return stbi__errpf("unsupported data layout", "Unsupported HDR format"); + token += 3; + width = (int) strtol(token, NULL, 10); - if (height > STBI_MAX_DIMENSIONS) - return stbi__errpf("too large", "Very large image (corrupt?)"); - if (width > STBI_MAX_DIMENSIONS) - return stbi__errpf("too large", "Very large image (corrupt?)"); + if (height > STBI_MAX_DIMENSIONS) return stbi__errpf("too large","Very large image (corrupt?)"); + if (width > STBI_MAX_DIMENSIONS) return stbi__errpf("too large","Very large image (corrupt?)"); - *x = width; - *y = height; + *x = width; + *y = height; - if (comp) - *comp = 3; - if (req_comp == 0) - req_comp = 3; + if (comp) *comp = 3; + if (req_comp == 0) req_comp = 3; - if (!stbi__mad4sizes_valid(width, height, req_comp, sizeof(float), 0)) - return stbi__errpf("too large", "HDR image is too large"); + if (!stbi__mad4sizes_valid(width, height, req_comp, sizeof(float), 0)) + return stbi__errpf("too large", "HDR image is too large"); - // Read data - hdr_data = (float *)stbi__malloc_mad4(width, height, req_comp, sizeof(float), 0); - if (!hdr_data) - return stbi__errpf("outofmem", "Out of memory"); + // Read data + hdr_data = (float *) stbi__malloc_mad4(width, height, req_comp, sizeof(float), 0); + if (!hdr_data) + return stbi__errpf("outofmem", "Out of memory"); - // Load image data - // image data is stored as some number of sca - if (width < 8 || width >= 32768) { - // Read flat data - for (j = 0; j < height; ++j) { - for (i = 0; i < width; ++i) { - stbi_uc rgbe[4]; - main_decode_loop: - stbi__getn(s, rgbe, 4); - stbi__hdr_convert(hdr_data + j * width * req_comp + i * req_comp, rgbe, req_comp); - } - } - } else { - // Read RLE-encoded data - scanline = NULL; + // Load image data + // image data is stored as some number of sca + if ( width < 8 || width >= 32768) { + // Read flat data + for (j=0; j < height; ++j) { + for (i=0; i < width; ++i) { + stbi_uc rgbe[4]; + main_decode_loop: + stbi__getn(s, rgbe, 4); + stbi__hdr_convert(hdr_data + j * width * req_comp + i * req_comp, rgbe, req_comp); + } + } + } else { + // Read RLE-encoded data + scanline = NULL; - for (j = 0; j < height; ++j) { - c1 = stbi__get8(s); - c2 = stbi__get8(s); - len = stbi__get8(s); - if (c1 != 2 || c2 != 2 || (len & 0x80)) { - // not run-length encoded, so we have to actually use THIS data as a decoded - // pixel (note this can't be a valid pixel--one of RGB must be >= 128) - stbi_uc rgbe[4]; - rgbe[0] = (stbi_uc)c1; - rgbe[1] = (stbi_uc)c2; - rgbe[2] = (stbi_uc)len; - rgbe[3] = (stbi_uc)stbi__get8(s); - stbi__hdr_convert(hdr_data, rgbe, req_comp); - i = 1; - j = 0; - STBI_FREE(scanline); - goto main_decode_loop; // yes, this makes no sense - } - len <<= 8; - len |= stbi__get8(s); - if (len != width) { - STBI_FREE(hdr_data); - STBI_FREE(scanline); - return stbi__errpf("invalid decoded scanline length", "corrupt HDR"); - } - if (scanline == NULL) { - scanline = (stbi_uc *)stbi__malloc_mad2(width, 4, 0); - if (!scanline) { - STBI_FREE(hdr_data); - return stbi__errpf("outofmem", "Out of memory"); - } - } - - for (k = 0; k < 4; ++k) { - int nleft; - i = 0; - while ((nleft = width - i) > 0) { - count = stbi__get8(s); - if (count > 128) { - // Run - value = stbi__get8(s); - count -= 128; - if ((count == 0) || (count > nleft)) { - STBI_FREE(hdr_data); - STBI_FREE(scanline); - return stbi__errpf("corrupt", "bad RLE data in HDR"); - } - for (z = 0; z < count; ++z) - scanline[i++ * 4 + k] = value; - } else { - // Dump - if ((count == 0) || (count > nleft)) { - STBI_FREE(hdr_data); - STBI_FREE(scanline); - return stbi__errpf("corrupt", "bad RLE data in HDR"); - } - for (z = 0; z < count; ++z) - scanline[i++ * 4 + k] = stbi__get8(s); - } - } - } - for (i = 0; i < width; ++i) - stbi__hdr_convert(hdr_data + (j * width + i) * req_comp, scanline + i * 4, req_comp); - } - if (scanline) + for (j = 0; j < height; ++j) { + c1 = stbi__get8(s); + c2 = stbi__get8(s); + len = stbi__get8(s); + if (c1 != 2 || c2 != 2 || (len & 0x80)) { + // not run-length encoded, so we have to actually use THIS data as a decoded + // pixel (note this can't be a valid pixel--one of RGB must be >= 128) + stbi_uc rgbe[4]; + rgbe[0] = (stbi_uc) c1; + rgbe[1] = (stbi_uc) c2; + rgbe[2] = (stbi_uc) len; + rgbe[3] = (stbi_uc) stbi__get8(s); + stbi__hdr_convert(hdr_data, rgbe, req_comp); + i = 1; + j = 0; STBI_FREE(scanline); - } + goto main_decode_loop; // yes, this makes no sense + } + len <<= 8; + len |= stbi__get8(s); + if (len != width) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("invalid decoded scanline length", "corrupt HDR"); } + if (scanline == NULL) { + scanline = (stbi_uc *) stbi__malloc_mad2(width, 4, 0); + if (!scanline) { + STBI_FREE(hdr_data); + return stbi__errpf("outofmem", "Out of memory"); + } + } - return hdr_data; + for (k = 0; k < 4; ++k) { + int nleft; + i = 0; + while ((nleft = width - i) > 0) { + count = stbi__get8(s); + if (count > 128) { + // Run + value = stbi__get8(s); + count -= 128; + if ((count == 0) || (count > nleft)) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("corrupt", "bad RLE data in HDR"); } + for (z = 0; z < count; ++z) + scanline[i++ * 4 + k] = value; + } else { + // Dump + if ((count == 0) || (count > nleft)) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("corrupt", "bad RLE data in HDR"); } + for (z = 0; z < count; ++z) + scanline[i++ * 4 + k] = stbi__get8(s); + } + } + } + for (i=0; i < width; ++i) + stbi__hdr_convert(hdr_data+(j*width + i)*req_comp, scanline + i*4, req_comp); + } + if (scanline) + STBI_FREE(scanline); + } + + return hdr_data; } -static int stbi__hdr_info(stbi__context * s, int * x, int * y, int * comp) { - char buffer[STBI__HDR_BUFLEN]; - char * token; - int valid = 0; - int dummy; +static int stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp) +{ + char buffer[STBI__HDR_BUFLEN]; + char *token; + int valid = 0; + int dummy; - if (!x) - x = &dummy; - if (!y) - y = &dummy; - if (!comp) - comp = &dummy; + if (!x) x = &dummy; + if (!y) y = &dummy; + if (!comp) comp = &dummy; - if (stbi__hdr_test(s) == 0) { - stbi__rewind(s); - return 0; - } + if (stbi__hdr_test(s) == 0) { + stbi__rewind( s ); + return 0; + } - for (;;) { - token = stbi__hdr_gettoken(s, buffer); - if (token[0] == 0) - break; - if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) - valid = 1; - } + for(;;) { + token = stbi__hdr_gettoken(s,buffer); + if (token[0] == 0) break; + if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) valid = 1; + } - if (!valid) { - stbi__rewind(s); - return 0; - } - token = stbi__hdr_gettoken(s, buffer); - if (strncmp(token, "-Y ", 3)) { - stbi__rewind(s); - return 0; - } - token += 3; - *y = (int)strtol(token, &token, 10); - while (*token == ' ') - ++token; - if (strncmp(token, "+X ", 3)) { - stbi__rewind(s); - return 0; - } - token += 3; - *x = (int)strtol(token, NULL, 10); - *comp = 3; - return 1; + if (!valid) { + stbi__rewind( s ); + return 0; + } + token = stbi__hdr_gettoken(s,buffer); + if (strncmp(token, "-Y ", 3)) { + stbi__rewind( s ); + return 0; + } + token += 3; + *y = (int) strtol(token, &token, 10); + while (*token == ' ') ++token; + if (strncmp(token, "+X ", 3)) { + stbi__rewind( s ); + return 0; + } + token += 3; + *x = (int) strtol(token, NULL, 10); + *comp = 3; + return 1; } #endif // STBI_NO_HDR #ifndef STBI_NO_BMP -static int stbi__bmp_info(stbi__context * s, int * x, int * y, int * comp) { - void * p; - stbi__bmp_data info; +static int stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp) +{ + void *p; + stbi__bmp_data info; - info.all_a = 255; - p = stbi__bmp_parse_header(s, &info); - if (p == NULL) { - stbi__rewind(s); - return 0; - } - if (x) - *x = s->img_x; - if (y) - *y = s->img_y; - if (comp) { - if (info.bpp == 24 && info.ma == 0xff000000) - *comp = 3; - else - *comp = info.ma ? 4 : 3; - } - return 1; + info.all_a = 255; + p = stbi__bmp_parse_header(s, &info); + if (p == NULL) { + stbi__rewind( s ); + return 0; + } + if (x) *x = s->img_x; + if (y) *y = s->img_y; + if (comp) { + if (info.bpp == 24 && info.ma == 0xff000000) + *comp = 3; + else + *comp = info.ma ? 4 : 3; + } + return 1; } #endif #ifndef STBI_NO_PSD -static int stbi__psd_info(stbi__context * s, int * x, int * y, int * comp) { - int channelCount, dummy, depth; - if (!x) - x = &dummy; - if (!y) - y = &dummy; - if (!comp) - comp = &dummy; - if (stbi__get32be(s) != 0x38425053) { - stbi__rewind(s); - return 0; - } - if (stbi__get16be(s) != 1) { - stbi__rewind(s); - return 0; - } - stbi__skip(s, 6); - channelCount = stbi__get16be(s); - if (channelCount < 0 || channelCount > 16) { - stbi__rewind(s); - return 0; - } - *y = stbi__get32be(s); - *x = stbi__get32be(s); - depth = stbi__get16be(s); - if (depth != 8 && depth != 16) { - stbi__rewind(s); - return 0; - } - if (stbi__get16be(s) != 3) { - stbi__rewind(s); - return 0; - } - *comp = 4; - return 1; +static int stbi__psd_info(stbi__context *s, int *x, int *y, int *comp) +{ + int channelCount, dummy, depth; + if (!x) x = &dummy; + if (!y) y = &dummy; + if (!comp) comp = &dummy; + if (stbi__get32be(s) != 0x38425053) { + stbi__rewind( s ); + return 0; + } + if (stbi__get16be(s) != 1) { + stbi__rewind( s ); + return 0; + } + stbi__skip(s, 6); + channelCount = stbi__get16be(s); + if (channelCount < 0 || channelCount > 16) { + stbi__rewind( s ); + return 0; + } + *y = stbi__get32be(s); + *x = stbi__get32be(s); + depth = stbi__get16be(s); + if (depth != 8 && depth != 16) { + stbi__rewind( s ); + return 0; + } + if (stbi__get16be(s) != 3) { + stbi__rewind( s ); + return 0; + } + *comp = 4; + return 1; } -static int stbi__psd_is16(stbi__context * s) { - int channelCount, depth; - if (stbi__get32be(s) != 0x38425053) { - stbi__rewind(s); - return 0; - } - if (stbi__get16be(s) != 1) { - stbi__rewind(s); - return 0; - } - stbi__skip(s, 6); - channelCount = stbi__get16be(s); - if (channelCount < 0 || channelCount > 16) { - stbi__rewind(s); - return 0; - } - STBI_NOTUSED(stbi__get32be(s)); - STBI_NOTUSED(stbi__get32be(s)); - depth = stbi__get16be(s); - if (depth != 16) { - stbi__rewind(s); - return 0; - } - return 1; +static int stbi__psd_is16(stbi__context *s) +{ + int channelCount, depth; + if (stbi__get32be(s) != 0x38425053) { + stbi__rewind( s ); + return 0; + } + if (stbi__get16be(s) != 1) { + stbi__rewind( s ); + return 0; + } + stbi__skip(s, 6); + channelCount = stbi__get16be(s); + if (channelCount < 0 || channelCount > 16) { + stbi__rewind( s ); + return 0; + } + STBI_NOTUSED(stbi__get32be(s)); + STBI_NOTUSED(stbi__get32be(s)); + depth = stbi__get16be(s); + if (depth != 16) { + stbi__rewind( s ); + return 0; + } + return 1; } #endif #ifndef STBI_NO_PIC -static int stbi__pic_info(stbi__context * s, int * x, int * y, int * comp) { - int act_comp = 0, num_packets = 0, chained, dummy; - stbi__pic_packet packets[10]; +static int stbi__pic_info(stbi__context *s, int *x, int *y, int *comp) +{ + int act_comp=0,num_packets=0,chained,dummy; + stbi__pic_packet packets[10]; - if (!x) - x = &dummy; - if (!y) - y = &dummy; - if (!comp) - comp = &dummy; + if (!x) x = &dummy; + if (!y) y = &dummy; + if (!comp) comp = &dummy; - if (!stbi__pic_is4(s, "\x53\x80\xF6\x34")) { - stbi__rewind(s); - return 0; - } + if (!stbi__pic_is4(s,"\x53\x80\xF6\x34")) { + stbi__rewind(s); + return 0; + } - stbi__skip(s, 88); + stbi__skip(s, 88); - *x = stbi__get16be(s); - *y = stbi__get16be(s); - if (stbi__at_eof(s)) { - stbi__rewind(s); - return 0; - } - if ((*x) != 0 && (1 << 28) / (*x) < (*y)) { - stbi__rewind(s); - return 0; - } + *x = stbi__get16be(s); + *y = stbi__get16be(s); + if (stbi__at_eof(s)) { + stbi__rewind( s); + return 0; + } + if ( (*x) != 0 && (1 << 28) / (*x) < (*y)) { + stbi__rewind( s ); + return 0; + } - stbi__skip(s, 8); + stbi__skip(s, 8); - do { - stbi__pic_packet * packet; + do { + stbi__pic_packet *packet; - if (num_packets == sizeof(packets) / sizeof(packets[0])) - return 0; + if (num_packets==sizeof(packets)/sizeof(packets[0])) + return 0; - packet = &packets[num_packets++]; - chained = stbi__get8(s); - packet->size = stbi__get8(s); - packet->type = stbi__get8(s); - packet->channel = stbi__get8(s); - act_comp |= packet->channel; + packet = &packets[num_packets++]; + chained = stbi__get8(s); + packet->size = stbi__get8(s); + packet->type = stbi__get8(s); + packet->channel = stbi__get8(s); + act_comp |= packet->channel; - if (stbi__at_eof(s)) { - stbi__rewind(s); - return 0; - } - if (packet->size != 8) { - stbi__rewind(s); - return 0; - } - } while (chained); + if (stbi__at_eof(s)) { + stbi__rewind( s ); + return 0; + } + if (packet->size != 8) { + stbi__rewind( s ); + return 0; + } + } while (chained); - *comp = (act_comp & 0x10 ? 4 : 3); + *comp = (act_comp & 0x10 ? 4 : 3); - return 1; + return 1; } #endif @@ -7904,271 +7491,272 @@ static int stbi__pic_info(stbi__context * s, int * x, int * y, int * comp) { #ifndef STBI_NO_PNM -static int stbi__pnm_test(stbi__context * s) { - char p, t; - p = (char)stbi__get8(s); - t = (char)stbi__get8(s); - if (p != 'P' || (t != '5' && t != '6')) { - stbi__rewind(s); - return 0; - } - return 1; +static int stbi__pnm_test(stbi__context *s) +{ + char p, t; + p = (char) stbi__get8(s); + t = (char) stbi__get8(s); + if (p != 'P' || (t != '5' && t != '6')) { + stbi__rewind( s ); + return 0; + } + return 1; } -static void * stbi__pnm_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) { - stbi_uc * out; - STBI_NOTUSED(ri); +static void *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + stbi_uc *out; + STBI_NOTUSED(ri); - ri->bits_per_channel = stbi__pnm_info(s, (int *)&s->img_x, (int *)&s->img_y, (int *)&s->img_n); - if (ri->bits_per_channel == 0) - return 0; + ri->bits_per_channel = stbi__pnm_info(s, (int *)&s->img_x, (int *)&s->img_y, (int *)&s->img_n); + if (ri->bits_per_channel == 0) + return 0; - if (s->img_y > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); - if (s->img_x > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); + if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); + if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); - *x = s->img_x; - *y = s->img_y; - if (comp) - *comp = s->img_n; + *x = s->img_x; + *y = s->img_y; + if (comp) *comp = s->img_n; - if (!stbi__mad4sizes_valid(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0)) - return stbi__errpuc("too large", "PNM too large"); + if (!stbi__mad4sizes_valid(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0)) + return stbi__errpuc("too large", "PNM too large"); - out = (stbi_uc *)stbi__malloc_mad4(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0); - if (!out) - return stbi__errpuc("outofmem", "Out of memory"); - if (!stbi__getn(s, out, s->img_n * s->img_x * s->img_y * (ri->bits_per_channel / 8))) { - STBI_FREE(out); - return stbi__errpuc("bad PNM", "PNM file truncated"); - } + out = (stbi_uc *) stbi__malloc_mad4(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0); + if (!out) return stbi__errpuc("outofmem", "Out of memory"); + if (!stbi__getn(s, out, s->img_n * s->img_x * s->img_y * (ri->bits_per_channel / 8))) { + STBI_FREE(out); + return stbi__errpuc("bad PNM", "PNM file truncated"); + } - if (req_comp && req_comp != s->img_n) { - if (ri->bits_per_channel == 16) { - out = (stbi_uc *)stbi__convert_format16((stbi__uint16 *)out, s->img_n, req_comp, s->img_x, s->img_y); - } else { - out = stbi__convert_format(out, s->img_n, req_comp, s->img_x, s->img_y); - } - if (out == NULL) - return out; // stbi__convert_format frees input on failure - } - return out; + if (req_comp && req_comp != s->img_n) { + if (ri->bits_per_channel == 16) { + out = (stbi_uc *) stbi__convert_format16((stbi__uint16 *) out, s->img_n, req_comp, s->img_x, s->img_y); + } else { + out = stbi__convert_format(out, s->img_n, req_comp, s->img_x, s->img_y); + } + if (out == NULL) return out; // stbi__convert_format frees input on failure + } + return out; } -static int stbi__pnm_isspace(char c) { return c == ' ' || c == '\t' || c == '\n' || c == '\v' || c == '\f' || c == '\r'; } - -static void stbi__pnm_skip_whitespace(stbi__context * s, char * c) { - for (;;) { - while (!stbi__at_eof(s) && stbi__pnm_isspace(*c)) - *c = (char)stbi__get8(s); - - if (stbi__at_eof(s) || *c != '#') - break; - - while (!stbi__at_eof(s) && *c != '\n' && *c != '\r') - *c = (char)stbi__get8(s); - } +static int stbi__pnm_isspace(char c) +{ + return c == ' ' || c == '\t' || c == '\n' || c == '\v' || c == '\f' || c == '\r'; } -static int stbi__pnm_isdigit(char c) { return c >= '0' && c <= '9'; } +static void stbi__pnm_skip_whitespace(stbi__context *s, char *c) +{ + for (;;) { + while (!stbi__at_eof(s) && stbi__pnm_isspace(*c)) + *c = (char) stbi__get8(s); -static int stbi__pnm_getinteger(stbi__context * s, char * c) { - int value = 0; + if (stbi__at_eof(s) || *c != '#') + break; - while (!stbi__at_eof(s) && stbi__pnm_isdigit(*c)) { - value = value * 10 + (*c - '0'); - *c = (char)stbi__get8(s); - if ((value > 214748364) || (value == 214748364 && *c > '7')) - return stbi__err("integer parse overflow", "Parsing an integer in the PPM header overflowed a 32-bit int"); - } - - return value; + while (!stbi__at_eof(s) && *c != '\n' && *c != '\r' ) + *c = (char) stbi__get8(s); + } } -static int stbi__pnm_info(stbi__context * s, int * x, int * y, int * comp) { - int maxv, dummy; - char c, p, t; - - if (!x) - x = &dummy; - if (!y) - y = &dummy; - if (!comp) - comp = &dummy; - - stbi__rewind(s); - - // Get identifier - p = (char)stbi__get8(s); - t = (char)stbi__get8(s); - if (p != 'P' || (t != '5' && t != '6')) { - stbi__rewind(s); - return 0; - } - - *comp = (t == '6') ? 3 : 1; // '5' is 1-component .pgm; '6' is 3-component .ppm - - c = (char)stbi__get8(s); - stbi__pnm_skip_whitespace(s, &c); - - *x = stbi__pnm_getinteger(s, &c); // read width - if (*x == 0) - return stbi__err("invalid width", "PPM image header had zero or overflowing width"); - stbi__pnm_skip_whitespace(s, &c); - - *y = stbi__pnm_getinteger(s, &c); // read height - if (*y == 0) - return stbi__err("invalid width", "PPM image header had zero or overflowing width"); - stbi__pnm_skip_whitespace(s, &c); - - maxv = stbi__pnm_getinteger(s, &c); // read max value - if (maxv > 65535) - return stbi__err("max value > 65535", "PPM image supports only 8-bit and 16-bit images"); - else if (maxv > 255) - return 16; - else - return 8; +static int stbi__pnm_isdigit(char c) +{ + return c >= '0' && c <= '9'; } -static int stbi__pnm_is16(stbi__context * s) { - if (stbi__pnm_info(s, NULL, NULL, NULL) == 16) - return 1; - return 0; +static int stbi__pnm_getinteger(stbi__context *s, char *c) +{ + int value = 0; + + while (!stbi__at_eof(s) && stbi__pnm_isdigit(*c)) { + value = value*10 + (*c - '0'); + *c = (char) stbi__get8(s); + if((value > 214748364) || (value == 214748364 && *c > '7')) + return stbi__err("integer parse overflow", "Parsing an integer in the PPM header overflowed a 32-bit int"); + } + + return value; +} + +static int stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp) +{ + int maxv, dummy; + char c, p, t; + + if (!x) x = &dummy; + if (!y) y = &dummy; + if (!comp) comp = &dummy; + + stbi__rewind(s); + + // Get identifier + p = (char) stbi__get8(s); + t = (char) stbi__get8(s); + if (p != 'P' || (t != '5' && t != '6')) { + stbi__rewind(s); + return 0; + } + + *comp = (t == '6') ? 3 : 1; // '5' is 1-component .pgm; '6' is 3-component .ppm + + c = (char) stbi__get8(s); + stbi__pnm_skip_whitespace(s, &c); + + *x = stbi__pnm_getinteger(s, &c); // read width + if(*x == 0) + return stbi__err("invalid width", "PPM image header had zero or overflowing width"); + stbi__pnm_skip_whitespace(s, &c); + + *y = stbi__pnm_getinteger(s, &c); // read height + if (*y == 0) + return stbi__err("invalid width", "PPM image header had zero or overflowing width"); + stbi__pnm_skip_whitespace(s, &c); + + maxv = stbi__pnm_getinteger(s, &c); // read max value + if (maxv > 65535) + return stbi__err("max value > 65535", "PPM image supports only 8-bit and 16-bit images"); + else if (maxv > 255) + return 16; + else + return 8; +} + +static int stbi__pnm_is16(stbi__context *s) +{ + if (stbi__pnm_info(s, NULL, NULL, NULL) == 16) + return 1; + return 0; } #endif -static int stbi__info_main(stbi__context * s, int * x, int * y, int * comp) { -#ifndef STBI_NO_JPEG - if (stbi__jpeg_info(s, x, y, comp)) - return 1; -#endif +static int stbi__info_main(stbi__context *s, int *x, int *y, int *comp) +{ + #ifndef STBI_NO_JPEG + if (stbi__jpeg_info(s, x, y, comp)) return 1; + #endif -#ifndef STBI_NO_PNG - if (stbi__png_info(s, x, y, comp)) - return 1; -#endif + #ifndef STBI_NO_PNG + if (stbi__png_info(s, x, y, comp)) return 1; + #endif -#ifndef STBI_NO_GIF - if (stbi__gif_info(s, x, y, comp)) - return 1; -#endif + #ifndef STBI_NO_GIF + if (stbi__gif_info(s, x, y, comp)) return 1; + #endif -#ifndef STBI_NO_BMP - if (stbi__bmp_info(s, x, y, comp)) - return 1; -#endif + #ifndef STBI_NO_BMP + if (stbi__bmp_info(s, x, y, comp)) return 1; + #endif -#ifndef STBI_NO_PSD - if (stbi__psd_info(s, x, y, comp)) - return 1; -#endif + #ifndef STBI_NO_PSD + if (stbi__psd_info(s, x, y, comp)) return 1; + #endif -#ifndef STBI_NO_PIC - if (stbi__pic_info(s, x, y, comp)) - return 1; -#endif + #ifndef STBI_NO_PIC + if (stbi__pic_info(s, x, y, comp)) return 1; + #endif -#ifndef STBI_NO_PNM - if (stbi__pnm_info(s, x, y, comp)) - return 1; -#endif + #ifndef STBI_NO_PNM + if (stbi__pnm_info(s, x, y, comp)) return 1; + #endif -#ifndef STBI_NO_HDR - if (stbi__hdr_info(s, x, y, comp)) - return 1; -#endif + #ifndef STBI_NO_HDR + if (stbi__hdr_info(s, x, y, comp)) return 1; + #endif -// test tga last because it's a crappy test! -#ifndef STBI_NO_TGA - if (stbi__tga_info(s, x, y, comp)) - return 1; -#endif - return stbi__err("unknown image type", "Image not of any known type, or corrupt"); + // test tga last because it's a crappy test! + #ifndef STBI_NO_TGA + if (stbi__tga_info(s, x, y, comp)) + return 1; + #endif + return stbi__err("unknown image type", "Image not of any known type, or corrupt"); } -static int stbi__is_16_main(stbi__context * s) { -#ifndef STBI_NO_PNG - if (stbi__png_is16(s)) - return 1; -#endif +static int stbi__is_16_main(stbi__context *s) +{ + #ifndef STBI_NO_PNG + if (stbi__png_is16(s)) return 1; + #endif -#ifndef STBI_NO_PSD - if (stbi__psd_is16(s)) - return 1; -#endif + #ifndef STBI_NO_PSD + if (stbi__psd_is16(s)) return 1; + #endif -#ifndef STBI_NO_PNM - if (stbi__pnm_is16(s)) - return 1; -#endif - return 0; + #ifndef STBI_NO_PNM + if (stbi__pnm_is16(s)) return 1; + #endif + return 0; } #ifndef STBI_NO_STDIO -STBIDEF int stbi_info(char const * filename, int * x, int * y, int * comp) { - FILE * f = stbi__fopen(filename, "rb"); +STBIDEF int stbi_info(char const *filename, int *x, int *y, int *comp) +{ + FILE *f = stbi__fopen(filename, "rb"); int result; - if (!f) - return stbi__err("can't fopen", "Unable to open file"); + if (!f) return stbi__err("can't fopen", "Unable to open file"); result = stbi_info_from_file(f, x, y, comp); fclose(f); return result; } -STBIDEF int stbi_info_from_file(FILE * f, int * x, int * y, int * comp) { - int r; - stbi__context s; - long pos = ftell(f); - stbi__start_file(&s, f); - r = stbi__info_main(&s, x, y, comp); - fseek(f, pos, SEEK_SET); - return r; +STBIDEF int stbi_info_from_file(FILE *f, int *x, int *y, int *comp) +{ + int r; + stbi__context s; + long pos = ftell(f); + stbi__start_file(&s, f); + r = stbi__info_main(&s,x,y,comp); + fseek(f,pos,SEEK_SET); + return r; } -STBIDEF int stbi_is_16_bit(char const * filename) { - FILE * f = stbi__fopen(filename, "rb"); +STBIDEF int stbi_is_16_bit(char const *filename) +{ + FILE *f = stbi__fopen(filename, "rb"); int result; - if (!f) - return stbi__err("can't fopen", "Unable to open file"); + if (!f) return stbi__err("can't fopen", "Unable to open file"); result = stbi_is_16_bit_from_file(f); fclose(f); return result; } -STBIDEF int stbi_is_16_bit_from_file(FILE * f) { - int r; - stbi__context s; - long pos = ftell(f); - stbi__start_file(&s, f); - r = stbi__is_16_main(&s); - fseek(f, pos, SEEK_SET); - return r; +STBIDEF int stbi_is_16_bit_from_file(FILE *f) +{ + int r; + stbi__context s; + long pos = ftell(f); + stbi__start_file(&s, f); + r = stbi__is_16_main(&s); + fseek(f,pos,SEEK_SET); + return r; } #endif // !STBI_NO_STDIO -STBIDEF int stbi_info_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * comp) { - stbi__context s; - stbi__start_mem(&s, buffer, len); - return stbi__info_main(&s, x, y, comp); +STBIDEF int stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp) +{ + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__info_main(&s,x,y,comp); } -STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const * c, void * user, int * x, int * y, int * comp) { - stbi__context s; - stbi__start_callbacks(&s, (stbi_io_callbacks *)c, user); - return stbi__info_main(&s, x, y, comp); +STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const *c, void *user, int *x, int *y, int *comp) +{ + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *) c, user); + return stbi__info_main(&s,x,y,comp); } -STBIDEF int stbi_is_16_bit_from_memory(stbi_uc const * buffer, int len) { - stbi__context s; - stbi__start_mem(&s, buffer, len); - return stbi__is_16_main(&s); +STBIDEF int stbi_is_16_bit_from_memory(stbi_uc const *buffer, int len) +{ + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__is_16_main(&s); } -STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const * c, void * user) { - stbi__context s; - stbi__start_callbacks(&s, (stbi_io_callbacks *)c, user); - return stbi__is_16_main(&s); +STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const *c, void *user) +{ + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *) c, user); + return stbi__is_16_main(&s); } #endif // STB_IMAGE_IMPLEMENTATION @@ -8279,9 +7867,12 @@ STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const * c, void * us 1.30 (2011-06-11) added ability to load files via callbacks to accomidate custom input streams (Ben Wenger) removed deprecated format-specific test/load functions - removed support for installable file formats (stbi_loader) -- would have been broken for IO callbacks - anyway error cases in bmp and tga give messages and don't leak (Raymond Barbiero, grisha) fix inefficiency in - decoding 32-bit BMP (David Woo) 1.29 (2010-08-16) various warning fixes from Aurelien Pocheville 1.28 (2010-08-01) + removed support for installable file formats (stbi_loader) -- would have been broken for IO callbacks anyway + error cases in bmp and tga give messages and don't leak (Raymond Barbiero, grisha) + fix inefficiency in decoding 32-bit BMP (David Woo) + 1.29 (2010-08-16) + various warning fixes from Aurelien Pocheville + 1.28 (2010-08-01) fix bug in GIF palette transparency (SpartanJ) 1.27 (2010-08-01) cast-to-stbi_uc to fix warnings @@ -8353,6 +7944,7 @@ STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const * c, void * us first released version */ + /* ------------------------------------------------------------------------------ This software is available under 2 licenses -- choose whichever you prefer. diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index dad6f707d..2e433741f 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -48,34 +48,40 @@ class Model: dir_model: Path ftype: gguf.LlamaFileType + fname_out: Path is_big_endian: bool endianess: gguf.GGUFEndian use_temp_file: bool lazy: bool - model_name: str | None part_names: list[str] is_safetensors: bool hparams: dict[str, Any] block_count: int tensor_map: gguf.TensorNameMap tensor_names: set[str] | None - fname_out: Path gguf_writer: gguf.GGUFWriter + model_name: str | None + metadata_override: Path | None + dir_model_card: Path + is_lora: bool # subclasses should define this! model_arch: gguf.MODEL_ARCH - def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool, use_temp_file: bool, eager: bool, - model_name: str | None, split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False): + def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False, + use_temp_file: bool = False, eager: bool = False, + metadata_override: Path | None = None, model_name: str | None = None, + split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False, is_lora: bool = False): if type(self) is Model: raise TypeError(f"{type(self).__name__!r} should not be directly instantiated") + self.dir_model = dir_model self.ftype = ftype + self.fname_out = fname_out self.is_big_endian = is_big_endian self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE self.use_temp_file = use_temp_file self.lazy = not eager - self.model_name = model_name self.part_names = Model.get_model_part_names(self.dir_model, "model", ".safetensors") self.is_safetensors = len(self.part_names) > 0 if not self.is_safetensors: @@ -84,6 +90,12 @@ class Model: self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"]) self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count) self.tensor_names = None + self.metadata_override = metadata_override + self.model_name = model_name + self.dir_model_card = dir_model # overridden in convert_lora_to_gguf.py + self.is_lora = is_lora # true if model is used inside convert_lora_to_gguf.py + + # Apply heuristics to figure out typical tensor encoding based on first layer tensor encoding type if self.ftype == gguf.LlamaFileType.GUESSED: # NOTE: can't use field "torch_dtype" in config.json, because some finetunes lie. _, first_tensor = next(self.get_tensors()) @@ -93,10 +105,8 @@ class Model: else: logger.info(f"choosing --outtype bf16 from first tensor type ({first_tensor.dtype})") self.ftype = gguf.LlamaFileType.MOSTLY_BF16 - ftype_up: str = self.ftype.name.partition("_")[2].upper() - ftype_lw: str = ftype_up.lower() - # allow templating the file name with the output ftype, useful with the "auto" ftype - self.fname_out = fname_out.parent / fname_out.name.format(ftype_lw, outtype=ftype_lw, ftype=ftype_lw, OUTTYPE=ftype_up, FTYPE=ftype_up) + + # Configure GGUF Writer self.gguf_writer = gguf.GGUFWriter(path=None, arch=gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file, split_max_tensors=split_max_tensors, split_max_size=split_max_size, dry_run=dry_run, small_first_shard=small_first_shard) @@ -148,9 +158,16 @@ class Model: tensor_names_from_parts.update(model_part.keys()) for name in model_part.keys(): - data = model_part.get_tensor(name) if self.is_safetensors else model_part[name] - if self.lazy: - data = LazyTorchTensor.from_eager(data) + if self.is_safetensors: + if self.lazy: + data = model_part.get_slice(name) + data = LazyTorchTensor.from_safetensors_slice(data) + else: + data = model_part.get_tensor(name) + else: + data = model_part[name] + if self.lazy: + data = LazyTorchTensor.from_eager(data) yield name, data # only verify tensor name presence; it doesn't matter if they are not in the right files @@ -186,7 +203,6 @@ class Model: return new_name def set_gguf_parameters(self): - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_block_count(self.block_count) if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None: @@ -225,6 +241,10 @@ class Model: self.gguf_writer.add_expert_used_count(n_experts_used) logger.info(f"gguf: experts used count = {n_experts_used}") + if (head_dim := self.hparams.get("head_dim")) is not None: + self.gguf_writer.add_key_length(head_dim) + self.gguf_writer.add_value_length(head_dim) + self.gguf_writer.add_file_type(self.ftype) logger.info(f"gguf: file type = {self.ftype}") @@ -233,17 +253,12 @@ class Model: return [(self.map_tensor_name(name), data_torch)] - def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool: + def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool: del name, new_name, bid, n_dims # unused return False - def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool: - del name, new_name, bid, n_dims # unused - - return False - - def write_tensors(self): + def prepare_tensors(self): max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,") for name, data_torch in self.get_tensors(): @@ -265,57 +280,50 @@ class Model: break for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)): - data: np.ndarray = data # type hint + data: np.ndarray # type hint n_dims = len(data.shape) - data_dtype = data.dtype - data_qtype: gguf.GGMLQuantizationType | None = None - - # when both are True, f32 should win - extra_f32 = self.extra_f32_tensors(name, new_name, bid, n_dims) - extra_f16 = self.extra_f16_tensors(name, new_name, bid, n_dims) + data_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims) # Most of the codebase that takes in 1D tensors or norms only handles F32 tensors - # Conditions should closely match those in llama_model_quantize_internal in llama.cpp - extra_f32 = any(cond for cond in ( - extra_f32, - n_dims == 1, - new_name.endswith("_norm.weight"), - )) - - # Some tensor types are always in float32 - extra_f32 = extra_f32 or any(self.match_model_tensor_name(new_name, key, bid) for key in ( - gguf.MODEL_TENSOR.FFN_GATE_INP, - gguf.MODEL_TENSOR.POS_EMBD, - gguf.MODEL_TENSOR.TOKEN_TYPES, - )) - - # if f16 desired, convert any float32 2-dim weight tensors to float16 - extra_f16 = any(cond for cond in ( - extra_f16, - (name.endswith(".weight") and n_dims >= 2), - )) - - if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32: - if self.ftype == gguf.LlamaFileType.MOSTLY_BF16: - data = gguf.quantize_bf16(data) - assert data.dtype == np.int16 - data_qtype = gguf.GGMLQuantizationType.BF16 - - elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0 and gguf.can_quantize_to_q8_0(data): - data = gguf.quantize_q8_0(data) - assert data.dtype == np.uint8 - data_qtype = gguf.GGMLQuantizationType.Q8_0 - - else: # default to float16 for quantized tensors - if data_dtype != np.float16: - data = data.astype(np.float16) - data_qtype = gguf.GGMLQuantizationType.F16 - - if data_qtype is None: # by default, convert to float32 - if data_dtype != np.float32: - data = data.astype(np.float32) + if n_dims <= 1 or new_name.endswith("_norm.weight"): data_qtype = gguf.GGMLQuantizationType.F32 + # Conditions should closely match those in llama_model_quantize_internal in llama.cpp + # Some tensor types are always in float32 + if data_qtype is False and ( + any( + self.match_model_tensor_name(new_name, key, bid) + for key in ( + gguf.MODEL_TENSOR.FFN_GATE_INP, + gguf.MODEL_TENSOR.POS_EMBD, + gguf.MODEL_TENSOR.TOKEN_TYPES, + gguf.MODEL_TENSOR.SSM_CONV1D, + ) + ) + or not name.endswith(".weight") + ): + data_qtype = gguf.GGMLQuantizationType.F32 + + # No override (data_qtype is False), or wants to be quantized (data_qtype is True) + if isinstance(data_qtype, bool): + if self.ftype == gguf.LlamaFileType.ALL_F32: + data_qtype = gguf.GGMLQuantizationType.F32 + elif self.ftype == gguf.LlamaFileType.MOSTLY_F16: + data_qtype = gguf.GGMLQuantizationType.F16 + elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16: + data_qtype = gguf.GGMLQuantizationType.BF16 + elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0: + data_qtype = gguf.GGMLQuantizationType.Q8_0 + else: + raise ValueError(f"Unknown file type: {self.ftype.name}") + + try: + data = gguf.quants.quantize(data, data_qtype) + except gguf.QuantError as e: + logger.warning("%s, %s", e, "falling back to F16") + data_qtype = gguf.GGMLQuantizationType.F16 + data = gguf.quants.quantize(data, data_qtype) + shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape # reverse shape to make it similar to the internal ggml dimension order @@ -326,9 +334,62 @@ class Model: self.gguf_writer.add_tensor(new_name, data, raw_dtype=data_qtype) + def set_type(self): + self.gguf_writer.add_type(gguf.GGUFType.MODEL) + + def prepare_metadata(self, vocab_only: bool): + + total_params, shared_params, expert_params, expert_count = self.gguf_writer.get_total_parameter_count() + + self.metadata = gguf.Metadata.load(self.metadata_override, self.dir_model_card, self.model_name, total_params) + + # Fallback to model directory name if metadata name is still missing + if self.metadata.name is None: + self.metadata.name = self.dir_model.name + + # Generate parameter weight class (useful for leader boards) if not yet determined + if self.metadata.size_label is None and total_params > 0: + self.metadata.size_label = gguf.size_label(total_params, shared_params, expert_params, expert_count) + + # Extract the encoding scheme from the file type name. e.g. 'gguf.LlamaFileType.MOSTLY_Q8_0' --> 'Q8_0' + output_type: str = self.ftype.name.partition("_")[2] + + # Filename Output + if self.fname_out.is_dir(): + # Generate default filename based on model specification and available metadata + if not vocab_only: + fname_default: str = gguf.naming_convention(self.metadata.name, self.metadata.basename, self.metadata.finetune, self.metadata.version, self.metadata.size_label, output_type, model_type="LoRA" if total_params < 0 else None) + else: + fname_default: str = gguf.naming_convention(self.metadata.name, self.metadata.basename, self.metadata.finetune, self.metadata.version, size_label=None, output_type=None, model_type="vocab") + + # Use the default filename + self.fname_out = self.fname_out / f"{fname_default}.gguf" + else: + # Output path is a custom defined templated filename + # Note: `not is_dir()` is used because `.is_file()` will not detect + # file template strings as it doesn't actually exist as a file + + # Process templated file name with the output ftype, useful with the "auto" ftype + self.fname_out = self.fname_out.parent / gguf.fill_templated_filename(self.fname_out.name, output_type) + + self.set_type() + + logger.info("Set meta model") + self.metadata.set_gguf_meta_model(self.gguf_writer) + + logger.info("Set model parameters") + self.set_gguf_parameters() + + logger.info("Set model tokenizer") + self.set_vocab() + + logger.info("Set model quantization version") + self.gguf_writer.add_quantization_version(gguf.GGML_QUANT_VERSION) + def write(self): - self.write_tensors() - self.gguf_writer.write_header_to_file(self.fname_out) + self.prepare_tensors() + self.prepare_metadata(vocab_only=False) + self.gguf_writer.write_header_to_file(path=self.fname_out) self.gguf_writer.write_kv_data_to_file() self.gguf_writer.write_tensors_to_file(progress=True) self.gguf_writer.close() @@ -336,7 +397,9 @@ class Model: def write_vocab(self): if len(self.gguf_writer.tensors) != 1: raise ValueError('Splitting the vocabulary is not supported') - self.gguf_writer.write_header_to_file(self.fname_out) + + self.prepare_metadata(vocab_only=True) + self.gguf_writer.write_header_to_file(path=self.fname_out) self.gguf_writer.write_kv_data_to_file() self.gguf_writer.close() @@ -373,6 +436,29 @@ class Model: except KeyError: raise NotImplementedError(f'Architecture {arch!r} not supported!') from None + def does_token_look_special(self, token: str | bytes) -> bool: + if isinstance(token, (bytes, bytearray)): + token_text = token.decode(encoding="utf-8") + elif isinstance(token, memoryview): + token_text = token.tobytes().decode(encoding="utf-8") + else: + token_text = token + + # Some models mark some added tokens which ought to be control tokens as not special. + # (e.g. command-r, command-r-plus, deepseek-coder, gemma{,-2}) + seems_special = token_text in ( + "", # deepseek-coder + "", "<2mass>", "[@BOS@]", # gemma{,-2} + ) + + seems_special = seems_special or (token_text.startswith("<|") and token_text.endswith("|>")) + seems_special = seems_special or (token_text.startswith("<｜") and token_text.endswith("｜>")) # deepseek-coder + + # TODO: should these be marked as UNUSED instead? (maybe not) + seems_special = seems_special or (token_text.startswith("")) # gemma{,-2} + + return seems_special + # used for GPT-2 BPE and WordPiece vocabs def get_vocab_base(self) -> tuple[list[str], list[int], str]: tokens: list[str] = [] @@ -391,20 +477,22 @@ class Model: for i in range(vocab_size): if i not in reverse_vocab: tokens.append(f"[PAD{i}]") - toktypes.append(gguf.TokenType.USER_DEFINED) - elif reverse_vocab[i] in added_vocab: - tokens.append(reverse_vocab[i]) - if tokenizer.added_tokens_decoder[i].special: - toktypes.append(gguf.TokenType.CONTROL) - else: - toktypes.append(gguf.TokenType.USER_DEFINED) + toktypes.append(gguf.TokenType.UNUSED) else: - tokens.append(reverse_vocab[i]) - toktypes.append(gguf.TokenType.NORMAL) + token: str = reverse_vocab[i] + if token in added_vocab: + if tokenizer.added_tokens_decoder[i].special or self.does_token_look_special(token): + toktypes.append(gguf.TokenType.CONTROL) + else: + token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ") # pre-normalize user-defined spaces + toktypes.append(gguf.TokenType.USER_DEFINED) + else: + toktypes.append(gguf.TokenType.NORMAL) + tokens.append(token) return tokens, toktypes, tokpre - # NOTE: this function is generated by convert-hf-to-gguf-update.py + # NOTE: this function is generated by convert_hf_to_gguf_update.py # do not modify it manually! # ref: https://github.com/ggerganov/llama.cpp/pull/6920 # Marker: Start get_vocab_base_pre @@ -424,7 +512,7 @@ class Model: res = None - # NOTE: if you get an error here, you need to update the convert-hf-to-gguf-update.py script + # NOTE: if you get an error here, you need to update the convert_hf_to_gguf_update.py script # or pull the latest version of the model from Huggingface # don't edit the hashes manually! if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5": @@ -487,21 +575,42 @@ class Model: if chkhsh == "7967bfa498ade6b757b064f31e964dddbb80f8f9a4d68d4ba7998fcf281c531a": # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-code res = "jina-v2-code" + if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b": + # ref: https://huggingface.co/THUDM/glm-4-9b-chat + res = "chatglm-bpe" if chkhsh == "7fc505bd3104ca1083b150b17d088b59534ede9bde81f0dd2090967d7fe52cee": # ref: https://huggingface.co/LumiOpen/Viking-7B res = "viking" if chkhsh == "b53802fb28e26d645c3a310b34bfe07da813026ec7c7716883404d5e0f8b1901": # ref: https://huggingface.co/core42/jais-13b res = "jais" + if chkhsh == "7b3e7548e4308f52a76e8229e4e6cc831195d0d1df43aed21ac6c93da05fec5f": + # ref: https://huggingface.co/WisdomShell/CodeShell-7B + res = "codeshell" + if chkhsh == "63b97e4253352e6f357cc59ea5b583e3a680eaeaf2632188c2b952de2588485e": + # ref: https://huggingface.co/mistralai/Mistral-Nemo-Base-2407 + res = "tekken" + if chkhsh == "855059429035d75a914d1eda9f10a876752e281a054a7a3d421ef0533e5b6249": + # ref: https://huggingface.co/HuggingFaceTB/SmolLM-135M + res = "smollm" + if chkhsh == "3c30d3ad1d6b64202cd222813e7736c2db6e1bd6d67197090fc1211fbc612ae7": + # ref: https://huggingface.co/bigscience/bloom + res = "bloom" + if chkhsh == "bc01ce58980e1db43859146dc51b1758b3b88729b217a74792e9f8d43e479d21": + # ref: https://huggingface.co/TurkuNLP/gpt3-finnish-small + res = "gpt3-finnish" + if chkhsh == "4e2b24cc4770243d65a2c9ec19770a72f08cffc161adbb73fcbb6b7dd45a0aae": + # ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct + res = "exaone" if res is None: logger.warning("\n") logger.warning("**************************************************************************************") logger.warning("** WARNING: The BPE pre-tokenizer was not recognized!") logger.warning("** There are 2 possible reasons for this:") - logger.warning("** - the model has not been added to convert-hf-to-gguf-update.py yet") + logger.warning("** - the model has not been added to convert_hf_to_gguf_update.py yet") logger.warning("** - the pre-tokenization config has changed upstream") - logger.warning("** Check your model files and convert-hf-to-gguf-update.py and update them accordingly.") + logger.warning("** Check your model files and convert_hf_to_gguf_update.py and update them accordingly.") logger.warning("** ref: https://github.com/ggerganov/llama.cpp/pull/6920") logger.warning("**") logger.warning(f"** chkhsh: {chkhsh}") @@ -556,7 +665,7 @@ class Model: for i in range(vocab_size): if i not in reverse_vocab: tokens.append(f"[PAD{i}]") - toktypes.append(gguf.TokenType.USER_DEFINED) + toktypes.append(gguf.TokenType.UNUSED) elif reverse_vocab[i] in added_vocab: tokens.append(reverse_vocab[i]) toktypes.append(gguf.TokenType.CONTROL) @@ -596,10 +705,6 @@ class Model: tokenizer_path = self.dir_model / 'tokenizer.model' - tokens: list[bytes] = [] - scores: list[float] = [] - toktypes: list[int] = [] - if not tokenizer_path.is_file(): raise FileNotFoundError(f"File not found: {tokenizer_path}") @@ -610,7 +715,7 @@ class Model: tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)] scores: list[float] = [-10000.0] * vocab_size - toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size + toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size for token_id in range(tokenizer.vocab_size()): piece = tokenizer.IdToPiece(token_id) @@ -637,7 +742,7 @@ class Model: added_tokens_json = json.load(f) for key in added_tokens_json: token_id = added_tokens_json[key] - if (token_id >= vocab_size): + if token_id >= vocab_size: logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') continue @@ -645,6 +750,26 @@ class Model: scores[token_id] = -1000.0 toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED + tokenizer_config_file = self.dir_model / 'tokenizer_config.json' + if tokenizer_config_file.is_file(): + with open(tokenizer_config_file, "r", encoding="utf-8") as f: + tokenizer_config_json = json.load(f) + added_tokens_decoder = tokenizer_config_json.get("added_tokens_decoder", {}) + for token_id, token_data in added_tokens_decoder.items(): + token_id = int(token_id) + token: str = token_data["content"] + if toktypes[token_id] != SentencePieceTokenTypes.UNUSED: + if tokens[token_id] != token.encode("utf-8"): + logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token!r}') + if token_data.get("special") or self.does_token_look_special(token): + toktypes[token_id] = SentencePieceTokenTypes.CONTROL + else: + token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ") # pre-normalize user-defined spaces + toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED + + scores[token_id] = -1000.0 + tokens[token_id] = token.encode("utf-8") + if vocab_size > len(tokens): pad_count = vocab_size - len(tokens) logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]") @@ -730,7 +855,6 @@ class GPTNeoXModel(Model): def set_gguf_parameters(self): block_count = self.hparams["num_hidden_layers"] - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"]) self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) self.gguf_writer.add_block_count(block_count) @@ -781,12 +905,11 @@ class GPTNeoXModel(Model): return tensors -@Model.register("BloomForCausalLM") +@Model.register("BloomForCausalLM", "BloomModel") class BloomModel(Model): model_arch = gguf.MODEL_ARCH.BLOOM def set_gguf_parameters(self): - self.gguf_writer.add_name("Bloom") n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed")) n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads")) self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed)) @@ -863,7 +986,6 @@ class MPTModel(Model): def set_gguf_parameters(self): block_count = self.hparams["n_layers"] - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_context_length(self.hparams["max_seq_len"]) self.gguf_writer.add_embedding_length(self.hparams["d_model"]) self.gguf_writer.add_block_count(block_count) @@ -902,7 +1024,6 @@ class OrionModel(Model): block_count = self.hparams["num_hidden_layers"] head_count = self.hparams["num_attention_heads"] head_count_kv = self.hparams.get("num_key_value_heads", head_count) - hf_repo = self.hparams.get("_name_or_path", "") ctx_length = 0 if "max_sequence_length" in self.hparams: @@ -915,8 +1036,6 @@ class OrionModel(Model): raise ValueError("gguf: can not find ctx length parameter.") self.gguf_writer.add_file_type(self.ftype) - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) - self.gguf_writer.add_source_hf_repo(hf_repo) self.gguf_writer.add_tensor_data_layout("Meta AI original pth") self.gguf_writer.add_context_length(ctx_length) self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) @@ -940,7 +1059,6 @@ class BaichuanModel(Model): block_count = self.hparams["num_hidden_layers"] head_count = self.hparams["num_attention_heads"] head_count_kv = self.hparams.get("num_key_value_heads", head_count) - hf_repo = self.hparams.get("_name_or_path", "") ctx_length = 0 if "max_sequence_length" in self.hparams: @@ -952,8 +1070,6 @@ class BaichuanModel(Model): else: raise ValueError("gguf: can not find ctx length parameter.") - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) - self.gguf_writer.add_source_hf_repo(hf_repo) self.gguf_writer.add_tensor_data_layout("Meta AI original pth") self.gguf_writer.add_context_length(ctx_length) self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) @@ -1067,7 +1183,6 @@ class XverseModel(Model): block_count = self.hparams["num_hidden_layers"] head_count = self.hparams["num_attention_heads"] head_count_kv = self.hparams.get("num_key_value_heads", head_count) - hf_repo = self.hparams.get("_name_or_path", "") ctx_length = 0 if "max_sequence_length" in self.hparams: @@ -1079,8 +1194,6 @@ class XverseModel(Model): else: raise ValueError("gguf: can not find ctx length parameter.") - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) - self.gguf_writer.add_source_hf_repo(hf_repo) self.gguf_writer.add_tensor_data_layout("Meta AI original pth") self.gguf_writer.add_context_length(ctx_length) self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) @@ -1139,7 +1252,6 @@ class FalconModel(Model): if n_head_kv is None: n_head_kv = self.hparams.get("n_head_kv", 1) # old name - self.gguf_writer.add_name("Falcon") self.gguf_writer.add_context_length(2048) # not in config.json self.gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) @@ -1184,7 +1296,6 @@ class StarCoderModel(Model): def set_gguf_parameters(self): block_count = self.hparams["n_layer"] - self.gguf_writer.add_name("StarCoder") self.gguf_writer.add_context_length(self.hparams["n_positions"]) self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"]) @@ -1204,11 +1315,11 @@ class RefactModel(Model): # TODO: how to determine special FIM tokens automatically? special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False, - special_token_types = ['prefix', 'suffix', 'middle', 'fsep', 'eot']) + special_token_types = ['prefix', 'suffix', 'middle', 'eot']) special_vocab._set_special_token("prefix", 1) special_vocab._set_special_token("suffix", 3) special_vocab._set_special_token("middle", 2) - special_vocab._set_special_token("fsep", 4) # is this correct? + special_vocab.chat_template = None # do not add it twice special_vocab.add_to_gguf(self.gguf_writer) def set_gguf_parameters(self): @@ -1220,7 +1331,6 @@ class RefactModel(Model): block_count = self.hparams["n_layer"] - self.gguf_writer.add_name("Refact") # refact uses Alibi. So this is from config.json which might be used by training. self.gguf_writer.add_context_length(self.hparams["n_positions"]) self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) @@ -1268,14 +1378,13 @@ class StableLMModel(Model): if (self.dir_model / "tokenizer.json").is_file(): self._set_vocab_gpt2() else: - # StableLM 2 1.6B uses a vocab in a similar format to Qwen's vocab + # StableLM 2 1.6B used to have a vocab in a similar format to Qwen's vocab self._set_vocab_qwen() def set_gguf_parameters(self): hparams = self.hparams block_count = hparams["num_hidden_layers"] - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_context_length(hparams["max_position_embeddings"]) self.gguf_writer.add_embedding_length(hparams["hidden_size"]) self.gguf_writer.add_block_count(block_count) @@ -1337,8 +1446,8 @@ class StableLMModel(Model): return [(new_name, data_torch)] - def write_tensors(self): - super().write_tensors() + def prepare_tensors(self): + super().prepare_tensors() if self._q_norms is not None or self._k_norms is not None: # flatten two `list[dict[str, Tensor]]` into a single `list[str]` @@ -1357,7 +1466,7 @@ class LlamaModel(Model): def set_vocab(self): try: - self. _set_vocab_sentencepiece() + self._set_vocab_sentencepiece() except FileNotFoundError: try: self._set_vocab_llama_hf() @@ -1381,7 +1490,12 @@ class LlamaModel(Model): super().set_gguf_parameters() hparams = self.hparams self.gguf_writer.add_vocab_size(hparams["vocab_size"]) - self.gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"]) + + if "head_dim" in hparams: + rope_dim = hparams["head_dim"] + else: + rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"] + self.gguf_writer.add_rope_dimension_count(rope_dim) if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: if self.hparams["rope_scaling"].get("type") == "linear": @@ -1454,8 +1568,37 @@ class LlamaModel(Model): return [(self.map_tensor_name(name), data_torch)] - def write_tensors(self): - super().write_tensors() + def prepare_tensors(self): + if rope_scaling := self.find_hparam(["rope_scaling"], optional=True): + if rope_scaling.get("rope_type", '').lower() == "llama3": + base = self.hparams.get("rope_theta", 10000.0) + dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"]) + freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + + factor = rope_scaling.get("factor", 8.0) + low_freq_factor = rope_scaling.get("low_freq_factor", 1.0) + high_freq_factor = rope_scaling.get("high_freq_factor", 4.0) + old_context_len = self.hparams.get("original_max_position_embeddings", 8192) + + low_freq_wavelen = old_context_len / low_freq_factor + high_freq_wavelen = old_context_len / high_freq_factor + assert low_freq_wavelen != high_freq_wavelen + + rope_factors = [] + for freq in freqs: + wavelen = 2 * math.pi / freq + if wavelen < high_freq_wavelen: + rope_factors.append(1) + elif wavelen > low_freq_wavelen: + rope_factors.append(factor) + else: + smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor) + rope_factors.append(1 / ((1 - smooth) / factor + smooth)) + + if not self.is_lora: + self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32)) + + super().prepare_tensors() if self._experts is not None: # flatten `list[dict[str, Tensor]]` into `list[str]` @@ -1518,7 +1661,6 @@ class GrokModel(Model): def set_gguf_parameters(self): super().set_gguf_parameters() - self.gguf_writer.add_name("Grok") _experts: list[dict[str, Tensor]] | None = None @@ -1567,7 +1709,6 @@ class DbrxModel(Model): def set_gguf_parameters(self): ffn_config = self.hparams["ffn_config"] attn_config = self.hparams["attn_config"] - self.gguf_writer.add_name(self.hparams["model_type"]) self.gguf_writer.add_block_count(self.hparams["n_layers"]) self.gguf_writer.add_context_length(self.hparams["max_seq_len"]) @@ -1580,7 +1721,6 @@ class DbrxModel(Model): self.gguf_writer.add_rope_freq_base(attn_config["rope_theta"]) self.gguf_writer.add_clamp_kqv(attn_config["clip_qkv"]) - self.gguf_writer.add_file_type(self.ftype) self.gguf_writer.add_expert_count(ffn_config["moe_num_experts"]) self.gguf_writer.add_expert_used_count(ffn_config["moe_top_k"]) @@ -1625,7 +1765,7 @@ class DbrxModel(Model): return [(new_name, data_torch)] - def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool: + def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool: del name, new_name, bid # unused return n_dims > 1 @@ -1637,7 +1777,6 @@ class MiniCPMModel(Model): def set_gguf_parameters(self): block_count = self.hparams["num_hidden_layers"] - self.gguf_writer.add_name("MiniCPM") self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"]) self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) self.gguf_writer.add_block_count(block_count) @@ -1707,7 +1846,6 @@ class QwenModel(Model): self._set_vocab_qwen() def set_gguf_parameters(self): - self.gguf_writer.add_name("Qwen") self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"]) self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"]) self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) @@ -1783,8 +1921,8 @@ class Qwen2MoeModel(Model): return [(self.map_tensor_name(name), data_torch)] - def write_tensors(self): - super().write_tensors() + def prepare_tensors(self): + super().prepare_tensors() if self._experts is not None: # flatten `list[dict[str, Tensor]]` into `list[str]` @@ -1798,7 +1936,6 @@ class GPT2Model(Model): model_arch = gguf.MODEL_ARCH.GPT2 def set_gguf_parameters(self): - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_block_count(self.hparams["n_layer"]) self.gguf_writer.add_context_length(self.hparams["n_ctx"]) self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) @@ -1841,7 +1978,6 @@ class Phi2Model(Model): n_embd = self.find_hparam(["hidden_size", "n_embd"]) n_head = self.find_hparam(["num_attention_heads", "n_head"]) - self.gguf_writer.add_name("Phi2") self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"])) self.gguf_writer.add_embedding_length(n_embd) @@ -1874,7 +2010,7 @@ class Phi3MiniModel(Model): tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)] scores: list[float] = [-10000.0] * vocab_size - toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size + toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size for token_id in range(tokenizer.vocab_size()): @@ -1903,7 +2039,7 @@ class Phi3MiniModel(Model): for key in added_tokens_json: token_id = added_tokens_json[key] - if (token_id >= vocab_size): + if token_id >= vocab_size: logger.debug(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') continue @@ -1919,8 +2055,9 @@ class Phi3MiniModel(Model): for token_id, foken_data in added_tokens_decoder.items(): token_id = int(token_id) token = foken_data["content"].encode("utf-8") - if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN: - assert tokens[token_id] == token + if toktypes[token_id] != SentencePieceTokenTypes.UNUSED: + if tokens[token_id] != token: + logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token.decode("utf-8")!r}') tokens[token_id] = token scores[token_id] = -1000.0 toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED @@ -1935,8 +2072,9 @@ class Phi3MiniModel(Model): for foken_data in added_tokens: token_id = int(foken_data["id"]) token = foken_data["content"].encode("utf-8") - if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN: - assert tokens[token_id] == token + if toktypes[token_id] != SentencePieceTokenTypes.UNUSED: + if tokens[token_id] != token: + logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token.decode("utf-8")!r}') tokens[token_id] = token scores[token_id] = -1000.0 toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED @@ -1963,7 +2101,6 @@ class Phi3MiniModel(Model): orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"]) rope_dims = n_embd // n_head - self.gguf_writer.add_name("Phi3") self.gguf_writer.add_context_length(max_pos_embds) self.gguf_writer.add_rope_scaling_orig_ctx_len(orig_max_pos_embds) self.gguf_writer.add_embedding_length(n_embd) @@ -1975,10 +2112,11 @@ class Phi3MiniModel(Model): self.gguf_writer.add_rope_dimension_count(rope_dims) self.gguf_writer.add_rope_freq_base(self.find_hparam(["rope_theta"])) self.gguf_writer.add_file_type(self.ftype) + self.gguf_writer.add_sliding_window(self.find_hparam(["sliding_window"])) # write rope scaling for long context (128k) model rope_scaling = self.find_hparam(['rope_scaling'], True) - if (rope_scaling is None): + if rope_scaling is None: return scale = max_pos_embds / orig_max_pos_embds @@ -2005,8 +2143,9 @@ class Phi3MiniModel(Model): if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2: raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}') - self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG] + ".weight", np.array(long_factors, dtype=np.float32)) - self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32)) + if not self.is_lora: + self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG] + ".weight", np.array(long_factors, dtype=np.float32)) + self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32)) @Model.register("PlamoForCausalLM") @@ -2020,7 +2159,6 @@ class PlamoModel(Model): hparams = self.hparams block_count = hparams["num_hidden_layers"] - self.gguf_writer.add_name("PLaMo") self.gguf_writer.add_context_length(4096) # not in config.json self.gguf_writer.add_embedding_length(hparams["hidden_size"]) self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"]) @@ -2065,7 +2203,6 @@ class CodeShellModel(Model): def set_gguf_parameters(self): block_count = self.hparams["n_layer"] - self.gguf_writer.add_name("CodeShell") self.gguf_writer.add_context_length(self.hparams["n_positions"]) self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"]) @@ -2117,7 +2254,7 @@ class InternLM2Model(Model): logger.error(f'Error: Missing {tokenizer_path}') sys.exit(1) - sentencepiece_model = model.ModelProto() + sentencepiece_model = model.ModelProto() # pyright: ignore[reportAttributeAccessIssue] sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read()) add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix @@ -2145,6 +2282,9 @@ class InternLM2Model(Model): toktype = SentencePieceTokenTypes.UNUSED elif tokenizer.IsByte(token_id): toktype = SentencePieceTokenTypes.BYTE + # take care of ununsed raw token + if piece.startswith('[UNUSED'): + toktype = SentencePieceTokenTypes.UNUSED tokens.append(text) scores.append(score) @@ -2160,6 +2300,49 @@ class InternLM2Model(Model): scores.append(-1000.0) toktypes.append(SentencePieceTokenTypes.USER_DEFINED) + chat_eos_token = '<|im_end|>' + chat_eos_token_id = None + + tokenizer_config_file = self.dir_model / 'tokenizer_config.json' + if tokenizer_config_file.is_file(): + with open(tokenizer_config_file, "r", encoding="utf-8") as f: + tokenizer_config_json = json.load(f) + added_tokens_decoder = tokenizer_config_json.get("added_tokens_decoder", {}) + for token_id, foken_data in added_tokens_decoder.items(): + token_id = int(token_id) + token = foken_data["content"] + if token == chat_eos_token: + chat_eos_token_id = token_id + token = token.encode("utf-8") + if toktypes[token_id] != SentencePieceTokenTypes.UNUSED: + if tokens[token_id] != token: + logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token.decode("utf-8")!r}') + tokens[token_id] = token + scores[token_id] = -1000.0 + toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED + if foken_data.get("special"): + toktypes[token_id] = SentencePieceTokenTypes.CONTROL + + tokenizer_file = self.dir_model / 'tokenizer.json' + if tokenizer_file.is_file(): + with open(tokenizer_file, "r", encoding="utf-8") as f: + tokenizer_json = json.load(f) + added_tokens = tokenizer_json.get("added_tokens", []) + for foken_data in added_tokens: + token_id = int(foken_data["id"]) + token = foken_data["content"] + if token == chat_eos_token: + chat_eos_token_id = token_id + token = token.encode("utf-8") + if toktypes[token_id] != SentencePieceTokenTypes.UNUSED: + if tokens[token_id] != token: + logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token.decode("utf-8")!r}') + tokens[token_id] = token + scores[token_id] = -1000.0 + toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED + if foken_data.get("special"): + toktypes[token_id] = SentencePieceTokenTypes.CONTROL + self.gguf_writer.add_tokenizer_model("llama") self.gguf_writer.add_tokenizer_pre("default") self.gguf_writer.add_token_list(tokens) @@ -2169,37 +2352,17 @@ class InternLM2Model(Model): special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) old_eos = special_vocab.special_token_ids["eos"] - if "chat" in os.path.basename(self.dir_model.absolute()): + if chat_eos_token_id is not None: # For the chat model, we replace the eos with '<|im_end|>'. # TODO: this is a hack, should be fixed # https://github.com/ggerganov/llama.cpp/pull/6745#issuecomment-2067687048 - special_vocab.special_token_ids["eos"] = self._try_get_sft_eos(tokenizer) - logger.warning(f"Replace eos:{old_eos} with a special token:{special_vocab.special_token_ids['eos']} \ -in chat mode so that the conversation can end normally.") + special_vocab.special_token_ids["eos"] = chat_eos_token_id + logger.warning(f"Replace eos:{old_eos} with a special token:{chat_eos_token_id}" + " in chat mode so that the conversation can end normally.") special_vocab.add_to_gguf(self.gguf_writer) - def _try_get_sft_eos(self, tokenizer): - unused_145_list = tokenizer.Encode('[UNUSED_TOKEN_145]') - im_end_list = tokenizer.Encode('<|im_end|>') - eos_token = None - assert (len(unused_145_list) == 1) ^ (len(im_end_list) == 1) - if len(unused_145_list) == 1: - eos_token = unused_145_list[0] - if len(im_end_list) == 1: - eos_token = im_end_list[0] - assert eos_token - return eos_token - - def _hf_permute_qk(self, weights, n_head: int, n_head_kv: int): - if n_head_kv is not None and n_head != n_head_kv: - n_head = n_head_kv - return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) - .swapaxes(1, 2) - .reshape(weights.shape)) - def set_gguf_parameters(self): - self.gguf_writer.add_name("InternLM2") self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"]) self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"]) self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) @@ -2209,30 +2372,30 @@ in chat mode so that the conversation can end normally.") self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"]) self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"]) self.gguf_writer.add_file_type(self.ftype) + if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: + if self.hparams["rope_scaling"].get("type") == "linear": + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: num_heads = self.hparams["num_attention_heads"] num_kv_heads = self.hparams["num_key_value_heads"] - hidden_size = self.hparams["hidden_size"] + n_embd = self.hparams["hidden_size"] q_per_kv = num_heads // num_kv_heads - head_dim = hidden_size // num_heads + head_dim = n_embd // num_heads num_groups = num_heads // q_per_kv - qkv_pattern = r"model\.layers\.(\d+)\.attention\.wqkv" - - if re.match(qkv_pattern, name): - bid = re.findall(qkv_pattern, name)[0] + if bid is not None and f"model.layers.{bid}.attention.wqkv" in name: qkv = data_torch - # qkv = rearrange(qkv.T, " o (g n i) ->o g n i", g=num_groups, n=q_per_kv + 2, i=head_dim) - qkv = qkv.T.reshape((-1, num_groups, q_per_kv + 2, head_dim)) - q, k, v = qkv[..., : q_per_kv, :], qkv[..., q_per_kv: q_per_kv + 1, :], qkv[..., q_per_kv + 1: q_per_kv + 2, :] + + qkv = qkv.reshape((num_groups, q_per_kv + 2, head_dim, n_embd)) + q, k, v = qkv[:, : q_per_kv], qkv[:, -2], qkv[:, -1] + # The model weights of q and k equire additional reshape. - # q = self._hf_permute_qk(rearrange(q, " o g n i -> o (g n i)").T, num_heads, num_heads) - q = self._hf_permute_qk(q.reshape((q.shape[0], -1)).T, num_heads, num_heads) - # k = self._hf_permute_qk(rearrange(k, " o g n i -> o (g n i)").T, num_heads, num_kv_heads) - k = self._hf_permute_qk(k.reshape((k.shape[0], -1)).T, num_heads, num_kv_heads) - # v = rearrange(v, " o g n i -> o (g n i)").T - v = v.reshape((v.shape[0], -1)).T + q = LlamaModel.permute(q.reshape((-1, q.shape[-1])), num_heads, num_heads) + k = LlamaModel.permute(k.reshape((-1, k.shape[-1])), num_heads, num_kv_heads) + v = v.reshape((-1, v.shape[-1])) + return [ (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), q), (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), k), @@ -2344,6 +2507,112 @@ class NomicBertModel(BertModel): self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"]) +@Model.register("XLMRobertaModel") +class XLMRobertaModel(BertModel): + model_arch = gguf.MODEL_ARCH.BERT + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + # we need the pad_token_id to know how to chop down position_embd matrix + if (pad_token_id := self.hparams.get("pad_token_id")) is not None: + self._position_offset = 1 + pad_token_id + if "max_position_embeddings" in self.hparams: + self.hparams["max_position_embeddings"] -= self._position_offset + else: + self._position_offset = None + + def set_vocab(self): + # to avoid TypeError: Descriptors cannot be created directly + # exception when importing sentencepiece_model_pb2 + os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python" + from sentencepiece import SentencePieceProcessor + from sentencepiece import sentencepiece_model_pb2 as model + + tokenizer_path = self.dir_model / 'sentencepiece.bpe.model' + if not tokenizer_path.is_file(): + raise FileNotFoundError(f"File not found: {tokenizer_path}") + + sentencepiece_model = model.ModelProto() # pyright: ignore[reportAttributeAccessIssue] + sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read()) + assert sentencepiece_model.trainer_spec.model_type == 1 # UNIGRAM + + add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix + remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces + precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap + + tokenizer = SentencePieceProcessor() + tokenizer.LoadFromFile(str(tokenizer_path)) + + vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size()) + + tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)] + scores: list[float] = [-10000.0] * vocab_size + toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size + + for token_id in range(tokenizer.vocab_size()): + piece = tokenizer.IdToPiece(token_id) + text = piece.encode("utf-8") + score = tokenizer.GetScore(token_id) + + toktype = SentencePieceTokenTypes.NORMAL + if tokenizer.IsUnknown(token_id): + toktype = SentencePieceTokenTypes.UNKNOWN + elif tokenizer.IsControl(token_id): + toktype = SentencePieceTokenTypes.CONTROL + elif tokenizer.IsUnused(token_id): + toktype = SentencePieceTokenTypes.UNUSED + elif tokenizer.IsByte(token_id): + toktype = SentencePieceTokenTypes.BYTE + + tokens[token_id] = text + scores[token_id] = score + toktypes[token_id] = toktype + + if vocab_size > len(tokens): + pad_count = vocab_size - len(tokens) + logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]") + for i in range(1, pad_count + 1): + tokens.append(bytes(f"[PAD{i}]", encoding="utf-8")) + scores.append(-1000.0) + toktypes.append(SentencePieceTokenTypes.UNUSED) + + # realign tokens (see HF tokenizer code) + tokens = [b'', b'', b'', b''] + tokens[3:-1] + scores = [0.0, 0.0, 0.0, 0.0] + scores[3:-1] + toktypes = [ + SentencePieceTokenTypes.CONTROL, + SentencePieceTokenTypes.CONTROL, + SentencePieceTokenTypes.CONTROL, + SentencePieceTokenTypes.UNKNOWN, + ] + toktypes[3:-1] + + self.gguf_writer.add_tokenizer_model("t5") + self.gguf_writer.add_tokenizer_pre("default") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_scores(scores) + self.gguf_writer.add_token_types(toktypes) + self.gguf_writer.add_add_space_prefix(add_prefix) + self.gguf_writer.add_token_type_count(1) + self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces) + if precompiled_charsmap: + self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap) + + special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) + special_vocab.add_to_gguf(self.gguf_writer) + + self.gguf_writer.add_add_bos_token(True) + self.gguf_writer.add_add_eos_token(True) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + # position embeddings start at pad_token_id + 1, so just chop down the weight tensor + if name == "embeddings.position_embeddings.weight": + if self._position_offset is not None: + data_torch = data_torch[self._position_offset:,:] + + return super().modify_tensors(data_torch, name, bid) + + @Model.register("GemmaForCausalLM") class GemmaModel(Model): model_arch = gguf.MODEL_ARCH.GEMMA @@ -2359,6 +2628,7 @@ class GemmaModel(Model): special_vocab._set_special_token("middle", 68) special_vocab._set_special_token("fsep", 70) special_vocab._set_special_token("eot", 107) + special_vocab.chat_template = None # do not add it twice special_vocab.add_to_gguf(self.gguf_writer) self.gguf_writer.add_add_space_prefix(False) @@ -2367,7 +2637,6 @@ class GemmaModel(Model): hparams = self.hparams block_count = hparams["num_hidden_layers"] - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_context_length(hparams["max_position_embeddings"]) self.gguf_writer.add_embedding_length(hparams["hidden_size"]) self.gguf_writer.add_block_count(block_count) @@ -2400,19 +2669,7 @@ class Gemma2Model(Model): model_arch = gguf.MODEL_ARCH.GEMMA2 def set_vocab(self): - tokens, scores, toktypes = self._create_vocab_sentencepiece() - # hack: This is required so that we can properly use start/end-of-turn for chat template - for i in range(108): - # including , , - toktypes[i] = SentencePieceTokenTypes.CONTROL - self.gguf_writer.add_tokenizer_model("llama") - self.gguf_writer.add_tokenizer_pre("default") - self.gguf_writer.add_token_list(tokens) - self.gguf_writer.add_token_scores(scores) - self.gguf_writer.add_token_types(toktypes) - - special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) - special_vocab.add_to_gguf(self.gguf_writer) + self._set_vocab_sentencepiece() self.gguf_writer.add_add_space_prefix(False) @@ -2420,7 +2677,6 @@ class Gemma2Model(Model): hparams = self.hparams block_count = hparams["num_hidden_layers"] - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_context_length(hparams["max_position_embeddings"]) self.gguf_writer.add_embedding_length(hparams["hidden_size"]) self.gguf_writer.add_block_count(block_count) @@ -2439,13 +2695,8 @@ class Gemma2Model(Model): ) self.gguf_writer.add_sliding_window(self.hparams["sliding_window"]) - # sanity check - attn_scalar = self.hparams["query_pre_attn_scalar"] - if attn_scalar != hparams["hidden_size"] / hparams["num_attention_heads"]: - raise ValueError("query_pre_attn_scalar must be equal to n_embd / n_head") - def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: - del bid # unusem + del bid # unused # lm_head is not used in llama.cpp, while autoawq will include this tensor in model # To prevent errors, skip loading lm_head.weight. @@ -2465,7 +2716,7 @@ class StarCoder2Model(Model): model_arch = gguf.MODEL_ARCH.STARCODER2 -@Model.register("MambaForCausalLM", "MambaLMHeadModel") +@Model.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM") class MambaModel(Model): model_arch = gguf.MODEL_ARCH.MAMBA @@ -2496,11 +2747,13 @@ class MambaModel(Model): # ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58 dt_rank = self.find_hparam(["time_step_rank", "dt_rank"], optional=True) or -(d_model // -16) rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5 - + use_dt_b_c_norm = False + # For falconmamba we do apply RMS norm on B / DT and C layers + if self.find_hparam(["model_type"], optional=True) in ("falcon_mamba",): + use_dt_b_c_norm = True # Fail early for models which don't have a block expansion factor of 2 assert d_inner == 2 * d_model - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default self.gguf_writer.add_embedding_length(d_model) self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading @@ -2511,6 +2764,7 @@ class MambaModel(Model): self.gguf_writer.add_ssm_state_size(d_state) self.gguf_writer.add_ssm_time_step_rank(dt_rank) self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps) + self.gguf_writer.add_ssm_dt_b_c_rms(use_dt_b_c_norm) # For classic Mamba we don't apply rms norm on B / DT layers self.gguf_writer.add_file_type(self.ftype) _tok_embd = None @@ -2537,19 +2791,6 @@ class MambaModel(Model): return [(new_name, data_torch)] - def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool: - del n_dims # unused - - return bid is not None and new_name in ( - self.format_tensor_name(n, bid, ".weight" if name.endswith(".weight") else "") for n in [ - gguf.MODEL_TENSOR.SSM_CONV1D, - gguf.MODEL_TENSOR.SSM_X, - gguf.MODEL_TENSOR.SSM_DT, - gguf.MODEL_TENSOR.SSM_A, - gguf.MODEL_TENSOR.SSM_D, - ] - ) - @Model.register("JambaForCausalLM") class JambaModel(Model): @@ -2749,7 +2990,7 @@ class JinaBertV2Model(BertModel): yield name, data - def set_vocab(self, *args, **kwargs): + def set_vocab(self): tokenizer_class = 'BertTokenizer' with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f: tokenizer_class = json.load(f)['tokenizer_class'] @@ -2808,7 +3049,6 @@ class OpenELMModel(Model): assert self.block_count == len(self._num_query_heads) assert self.block_count == len(self._ffn_dims) - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_block_count(self.block_count) self.gguf_writer.add_context_length(self.hparams["max_context_length"]) self.gguf_writer.add_embedding_length(n_embd) @@ -2865,7 +3105,7 @@ class ArcticModel(Model): tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)] scores: list[float] = [-10000.0] * vocab_size - toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size + toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size for token_id in range(tokenizer.vocab_size()): @@ -2898,7 +3138,7 @@ class ArcticModel(Model): added_tokens_decoder = tokenizer_config_json["added_tokens_decoder"] for token_id, token_json in added_tokens_decoder.items(): token_id = int(token_id) - if (token_id >= vocab_size): + if token_id >= vocab_size: logger.debug(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') continue @@ -2982,8 +3222,8 @@ class ArcticModel(Model): return [(self.map_tensor_name(name), data_torch)] - def write_tensors(self): - super().write_tensors() + def prepare_tensors(self): + super().prepare_tensors() if self._experts is not None: # flatten `list[dict[str, Tensor]]` into `list[str]` @@ -3061,8 +3301,8 @@ class DeepseekV2Model(Model): return [(self.map_tensor_name(name), data_torch)] - def write_tensors(self): - super().write_tensors() + def prepare_tensors(self): + super().prepare_tensors() if self._experts is not None: # flatten `list[dict[str, Tensor]]` into `list[str]` @@ -3098,16 +3338,16 @@ class T5Model(Model): if not tokenizer_path.is_file(): raise FileNotFoundError(f"File not found: {tokenizer_path}") - sentencepiece_model = model.ModelProto() + sentencepiece_model = model.ModelProto() # pyright: ignore[reportAttributeAccessIssue] sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read()) # some models like Pile-T5 family use BPE tokenizer instead of Unigram - if sentencepiece_model.trainer_spec.model_type == 2: # BPE + if sentencepiece_model.trainer_spec.model_type == 2: # BPE # assure the tokenizer model file name is correct assert tokenizer_path.name == 'tokenizer.model' return self._set_vocab_sentencepiece() else: - assert sentencepiece_model.trainer_spec.model_type == 1 # UNIGRAM + assert sentencepiece_model.trainer_spec.model_type == 1 # UNIGRAM add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces @@ -3120,7 +3360,7 @@ class T5Model(Model): tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)] scores: list[float] = [-10000.0] * vocab_size - toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size + toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size for token_id in range(tokenizer.vocab_size()): piece = tokenizer.IdToPiece(token_id) @@ -3147,7 +3387,7 @@ class T5Model(Model): added_tokens_json = json.load(f) for key in added_tokens_json: token_id = added_tokens_json[key] - if (token_id >= vocab_size): + if token_id >= vocab_size: logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') continue @@ -3180,7 +3420,6 @@ class T5Model(Model): self.gguf_writer.add_add_eos_token(True) def set_gguf_parameters(self): - self.gguf_writer.add_name("T5") if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None: logger.warning("Couldn't find context length in config.json, assuming default value of 512") n_ctx = 512 @@ -3215,6 +3454,145 @@ class T5Model(Model): return [(self.map_tensor_name(name), data_torch)] +@Model.register("T5EncoderModel") +class T5EncoderModel(Model): + model_arch = gguf.MODEL_ARCH.T5ENCODER + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.shared_token_embeddings_found = False + + def set_vocab(self): + # to avoid TypeError: Descriptors cannot be created directly + # exception when importing sentencepiece_model_pb2 + os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python" + from sentencepiece import SentencePieceProcessor + from sentencepiece import sentencepiece_model_pb2 as model + + tokenizer_path = self.dir_model / 'tokenizer.model' + + # many older models use spiece.model tokenizer model filename + if not tokenizer_path.is_file(): + tokenizer_path = self.dir_model / 'spiece.model' + + if not tokenizer_path.is_file(): + raise FileNotFoundError(f"File not found: {tokenizer_path}") + + sentencepiece_model = model.ModelProto() # pyright: ignore[reportAttributeAccessIssue] + sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read()) + + # some models like Pile-T5 family use BPE tokenizer instead of Unigram + if sentencepiece_model.trainer_spec.model_type == 2: # BPE + # assure the tokenizer model file name is correct + assert tokenizer_path.name == 'tokenizer.model' + return self._set_vocab_sentencepiece() + else: + assert sentencepiece_model.trainer_spec.model_type == 1 # UNIGRAM + + add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix + remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces + precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap + + tokenizer = SentencePieceProcessor() + tokenizer.LoadFromFile(str(tokenizer_path)) + + vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size()) + + tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)] + scores: list[float] = [-10000.0] * vocab_size + toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size + + for token_id in range(tokenizer.vocab_size()): + piece = tokenizer.IdToPiece(token_id) + text = piece.encode("utf-8") + score = tokenizer.GetScore(token_id) + + toktype = SentencePieceTokenTypes.NORMAL + if tokenizer.IsUnknown(token_id): + toktype = SentencePieceTokenTypes.UNKNOWN + elif tokenizer.IsControl(token_id): + toktype = SentencePieceTokenTypes.CONTROL + elif tokenizer.IsUnused(token_id): + toktype = SentencePieceTokenTypes.UNUSED + elif tokenizer.IsByte(token_id): + toktype = SentencePieceTokenTypes.BYTE + + tokens[token_id] = text + scores[token_id] = score + toktypes[token_id] = toktype + + added_tokens_file = self.dir_model / 'added_tokens.json' + if added_tokens_file.is_file(): + with open(added_tokens_file, "r", encoding="utf-8") as f: + added_tokens_json = json.load(f) + for key in added_tokens_json: + token_id = added_tokens_json[key] + if token_id >= vocab_size: + logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') + continue + + tokens[token_id] = key.encode("utf-8") + scores[token_id] = -1000.0 + toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED + + if vocab_size > len(tokens): + pad_count = vocab_size - len(tokens) + logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]") + for i in range(1, pad_count + 1): + tokens.append(bytes(f"[PAD{i}]", encoding="utf-8")) + scores.append(-1000.0) + toktypes.append(SentencePieceTokenTypes.UNUSED) + + self.gguf_writer.add_tokenizer_model("t5") + self.gguf_writer.add_tokenizer_pre("default") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_scores(scores) + self.gguf_writer.add_token_types(toktypes) + self.gguf_writer.add_add_space_prefix(add_prefix) + self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces) + if precompiled_charsmap: + self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap) + + special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) + special_vocab.add_to_gguf(self.gguf_writer) + + self.gguf_writer.add_add_bos_token(False) + self.gguf_writer.add_add_eos_token(True) + + def set_gguf_parameters(self): + if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None: + logger.warning("Couldn't find context length in config.json, assuming default value of 512") + n_ctx = 512 + self.gguf_writer.add_context_length(n_ctx) + self.gguf_writer.add_embedding_length(self.hparams["d_model"]) + self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"]) + self.gguf_writer.add_block_count(self.hparams["num_layers"]) + self.gguf_writer.add_head_count(self.hparams["num_heads"]) + self.gguf_writer.add_key_length(self.hparams["d_kv"]) + self.gguf_writer.add_value_length(self.hparams["d_kv"]) + self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"]) + self.gguf_writer.add_relative_attn_buckets_count(self.hparams["relative_attention_num_buckets"]) + self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"]) + self.gguf_writer.add_file_type(self.ftype) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + del bid # unused + + # T5 based models contain shared token embeddings tensors saved randomly as either "encoder.embed_tokens.weight", + # "decoder.embed_tokens.weight" or "shared.weight" tensor. In some models there are even multiple of them stored + # in the safetensors files. We use the first tensor from these three as the token embeddings for both encoder + # and decoder and ignore the remaining ones. + if name in ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight", "shared.weight"]: + if not self.shared_token_embeddings_found: + name = "shared.weight" + self.shared_token_embeddings_found = True + else: + logger.debug(f"Skipping shared tensor {name!r} in safetensors so that convert can end normally.") + return [] + + return [(self.map_tensor_name(name), data_torch)] + + @Model.register("JAISLMHeadModel") class JaisModel(Model): model_arch = gguf.MODEL_ARCH.JAIS @@ -3254,7 +3632,6 @@ class JaisModel(Model): self._set_vocab_gpt2() def set_gguf_parameters(self): - self.gguf_writer.add_name(self.dir_model.name) self.gguf_writer.add_block_count(self.hparams["n_layer"]) self.gguf_writer.add_context_length(self.hparams["n_positions"]) self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) @@ -3278,7 +3655,7 @@ class JaisModel(Model): # but Jais's PyTorch model simply precalculates the slope values and places them # in relative_pes.slopes n_head_closest_log2 = 2 ** math.floor(math.log2(self.hparams["n_head"])) - first_val = float(data_torch._data[0]) + first_val = float(data_torch[0].item()) self.max_alibi_bias = -round(math.log2(first_val) * n_head_closest_log2) return tensors @@ -3300,13 +3677,308 @@ class JaisModel(Model): return tensors - def write_tensors(self): - super().write_tensors() + def prepare_tensors(self): + super().prepare_tensors() self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias) -###### CONVERSION LOGIC ###### +@Model.register("ChatGLMModel", "ChatGLMForConditionalGeneration") +class ChatGLMModel(Model): + model_arch = gguf.MODEL_ARCH.CHATGLM + def set_vocab_chatglm3(self): + dir_model = self.dir_model + hparams = self.hparams + tokens: list[bytes] = [] + toktypes: list[int] = [] + scores: list[float] = [] + + from transformers import AutoTokenizer + tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True) + vocab_size = hparams.get("padded_vocab_size", len(tokenizer.get_vocab())) + assert max(tokenizer.get_vocab().values()) < vocab_size + role_special_tokens = ["<|system|>", "<|user|>", "<|assistant|>", "<|observation|>"] + special_tokens = ["[MASK]", "[gMASK]", "[sMASK]", "sop", "eop"] + role_special_tokens + for token_id in range(vocab_size): + piece = tokenizer._convert_id_to_token(token_id) + if token_id == 0: + piece = "" + elif token_id == 1: + piece = "" + elif token_id == 2: + piece = "" + + text = piece.encode("utf-8") + score = 0.0 + # Referencing the tokenizer Python implementation(https://huggingface.co/THUDM/chatglm3-6b/blob/main/tokenization_chatglm.py), + # it is only valid if it is less than tokenizer.tokenizer.sp_model.vocab_size() + if len(piece) != 0 and token_id < tokenizer.tokenizer.sp_model.vocab_size(): + score = tokenizer.tokenizer.sp_model.get_score(token_id) + + if token_id >= tokenizer.tokenizer.sp_model.vocab_size(): + if piece in special_tokens: + toktype = SentencePieceTokenTypes.CONTROL + elif len(piece) == 0: + text = f"[PAD{token_id}]".encode("utf-8") + toktype = SentencePieceTokenTypes.UNUSED + else: + toktype = SentencePieceTokenTypes.USER_DEFINED + tokens.append(text) + scores.append(score) + toktypes.append(toktype) + continue + + toktype = SentencePieceTokenTypes.NORMAL + if tokenizer.tokenizer.sp_model.is_unknown(token_id): + toktype = SentencePieceTokenTypes.UNKNOWN + elif tokenizer.tokenizer.sp_model.is_control(token_id): + toktype = SentencePieceTokenTypes.CONTROL + elif tokenizer.tokenizer.sp_model.is_unused(token_id): + toktype = SentencePieceTokenTypes.UNUSED + elif tokenizer.tokenizer.sp_model.is_byte(token_id): + toktype = SentencePieceTokenTypes.BYTE + + tokens.append(text) + scores.append(score) + toktypes.append(toktype) + + self.gguf_writer.add_tokenizer_model("llama") + # glm3 needs prefix and suffix formatted as: + # prompt = "[gMASK]sop<|user|>\n" + prompt + "<|assistant|>" + self.gguf_writer.add_tokenizer_pre("chatglm-spm") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_scores(scores) + self.gguf_writer.add_token_types(toktypes) + + special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) + special_vocab.add_to_gguf(self.gguf_writer) + + @staticmethod + def token_bytes_to_string(b): + from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode + byte_encoder = bytes_to_unicode() + return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')]) + + @staticmethod + def bpe(mergeable_ranks: dict[bytes, int], token: bytes, max_rank: int | None = None) -> list[bytes]: + parts = [bytes([b]) for b in token] + while True: + min_idx = None + min_rank = None + for i, pair in enumerate(zip(parts[:-1], parts[1:])): + rank = mergeable_ranks.get(pair[0] + pair[1]) + if rank is not None and (min_rank is None or rank < min_rank): + min_idx = i + min_rank = rank + if min_rank is None or (max_rank is not None and min_rank >= max_rank): + break + assert min_idx is not None + parts = parts[:min_idx] + [parts[min_idx] + parts[min_idx + 1]] + parts[min_idx + 2:] + return parts + + def set_vocab(self): + if "THUDM/chatglm3-6b" in self.hparams.get("_name_or_path", ""): + self.set_vocab_chatglm3() + return + + dir_model = self.dir_model + hparams = self.hparams + tokens: list[str] = [] + toktypes: list[int] = [] + + from transformers import AutoTokenizer + tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True) + vocab_size = hparams["padded_vocab_size"] + assert max(tokenizer.get_vocab().values()) < vocab_size + + tokpre = self.get_vocab_base_pre(tokenizer) + + merges = [] + vocab = {} + mergeable_ranks = tokenizer.mergeable_ranks + for token, rank in mergeable_ranks.items(): + vocab[ChatGLMModel.token_bytes_to_string(token)] = rank + if len(token) == 1: + continue + merged = ChatGLMModel.bpe(mergeable_ranks, token, max_rank=rank) + assert len(merged) >= 2 and len(merged) <= 7 + merges.append(' '.join(map(ChatGLMModel.token_bytes_to_string, merged))) + + # for this kind of tokenizer, added_vocab is not a subset of vocab, so they need to be combined + added_vocab = tokenizer.get_added_vocab() + reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **added_vocab}.items()} + + for i in range(vocab_size): + if i not in reverse_vocab: + tokens.append(f"[PAD{i}]") + toktypes.append(gguf.TokenType.UNUSED) + elif reverse_vocab[i] in added_vocab: + tokens.append(reverse_vocab[i]) + if tokenizer.added_tokens_decoder[i].special: + toktypes.append(gguf.TokenType.CONTROL) + else: + toktypes.append(gguf.TokenType.USER_DEFINED) + else: + tokens.append(reverse_vocab[i]) + toktypes.append(gguf.TokenType.NORMAL) + + self.gguf_writer.add_tokenizer_model("gpt2") + self.gguf_writer.add_tokenizer_pre(tokpre) + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_types(toktypes) + + special_vocab = gguf.SpecialVocab(dir_model, load_merges=False) + special_vocab.merges = merges + # only add special tokens when they were not already loaded from config.json + special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"]) + special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"]) + # this one is usually not in config.json anyway + special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"]) + special_vocab.add_to_gguf(self.gguf_writer) + + def set_gguf_parameters(self): + n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed")) + n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads")) + n_head_kv = self.hparams.get("multi_query_group_num", n_head) + self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed)) + self.gguf_writer.add_embedding_length(n_embed) + self.gguf_writer.add_feed_forward_length(self.hparams.get("ffn_hidden_size", 4 * n_embed)) + self.gguf_writer.add_block_count(self.hparams["num_layers"]) + self.gguf_writer.add_head_count(n_head) + self.gguf_writer.add_head_count_kv(n_head_kv) + self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layernorm_epsilon"]) + self.gguf_writer.add_file_type(self.ftype) + self.gguf_writer.add_rope_dimension_count(64) + self.gguf_writer.add_add_bos_token(False) + rope_freq = 10000 + if "rope_ratio" in self.hparams: + rope_freq = rope_freq * self.hparams["rope_ratio"] + self.gguf_writer.add_rope_freq_base(rope_freq) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + del bid # unused + + if name.endswith(".rotary_pos_emb.inv_freq"): + return [] + + name = name.removeprefix("transformer.") + return [(self.map_tensor_name(name), data_torch)] + + +@Model.register("NemotronForCausalLM") +class NemotronModel(Model): + model_arch = gguf.MODEL_ARCH.NEMOTRON + + def set_vocab(self): + self._set_vocab_sentencepiece() + self.gguf_writer.add_pad_token_id(0) + self.gguf_writer.add_unk_token_id(1) + + def set_gguf_parameters(self): + super().set_gguf_parameters() + hparams = self.hparams + self.gguf_writer.add_vocab_size(hparams["vocab_size"]) + + f_norm_eps = self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon", "norm_eps"]) + self.gguf_writer.add_layer_norm_eps(f_norm_eps) + + # * Partial RoPE + rot_pct = self.find_hparam(["partial_rotary_factor", "rope_pct", "rope_percent"]) + n_embd = self.find_hparam(["hidden_size", "n_embd"]) + n_head = self.find_hparam(["num_attention_heads", "n_head"]) + self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head) + + # * RopeScaling for Nemotron + if "rope_scaling" not in self.hparams or self.hparams["rope_scaling"] is None: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE) + else: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(self.hparams["factor"]) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + # * Adding +1 to LayerNorm's weights here to implement layernorm1p w/o changing anything on the GGML engine side + # model.layers.{l}.input_layernorm.weight + # model.layers.{l}.post_attention_layernorm.weight + # model.norm.weight + if name.endswith("norm.weight"): + data_torch = data_torch + 1 + + return [(self.map_tensor_name(name), data_torch)] + + +@Model.register("ExaoneForCausalLM") +class ExaoneModel(Model): + model_arch = gguf.MODEL_ARCH.EXAONE + + def set_gguf_parameters(self): + hparams = self.hparams + + assert (hparams["activation_function"] == "silu") + + max_position_embeddings = hparams["max_position_embeddings"] + embed_dim = hparams["hidden_size"] + num_heads = hparams["num_attention_heads"] + num_kv_heads = hparams.get("num_key_value_heads", num_heads) + layer_norm_eps = hparams["layer_norm_epsilon"] + intermediate_size = hparams["intermediate_size"] if "intermediate_size" in hparams else 4 * embed_dim + num_layers = hparams["num_layers"] + # ignore for now as EXAONE-3.0-7.8B-Instruct attentino_dropout is 0.0 + # attention_dropout_rate = hparams["attention_dropout"] + # ignore for now as EXAONE-3.0-7.8B-Instruct embed_dropout is 0.0 + # embed_dropout_rate = hparams["embed_dropout"] + self.gguf_writer.add_embedding_length(embed_dim) + self.gguf_writer.add_head_count(num_heads) + self.gguf_writer.add_head_count_kv(num_kv_heads) + self.gguf_writer.add_context_length(max_position_embeddings) + self.gguf_writer.add_layer_norm_rms_eps(layer_norm_eps) + self.gguf_writer.add_feed_forward_length(intermediate_size) + self.gguf_writer.add_block_count(num_layers) + self.gguf_writer.add_file_type(self.ftype) + + if (rope_theta := self.hparams.get("rope_theta")) is not None: + self.gguf_writer.add_rope_freq_base(rope_theta) + rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"], optional=True) + rotary_factor = rotary_factor if rotary_factor is not None else 1.0 + self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"]))) + if hparams.get("rope_scaling") is not None and "factor" in hparams["rope_scaling"]: + if hparams["rope_scaling"].get("type") == "linear": + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"]) + + def prepare_tensors(self): + if rope_scaling := self.find_hparam(["rope_scaling"], optional=True): + if rope_scaling.get("rope_type", '').lower() == "llama3": + base = self.hparams.get("rope_theta", 10000.0) + dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"]) + freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + + factor = rope_scaling.get("factor", 8.0) + low_freq_factor = rope_scaling.get("low_freq_factor", 1.0) + high_freq_factor = rope_scaling.get("high_freq_factor", 4.0) + old_context_len = self.hparams.get("original_max_position_embeddings", 8192) + + low_freq_wavelen = old_context_len / low_freq_factor + high_freq_wavelen = old_context_len / high_freq_factor + assert low_freq_wavelen != high_freq_wavelen + + rope_factors = [] + for freq in freqs: + wavelen = 2 * math.pi / freq + if wavelen < high_freq_wavelen: + rope_factors.append(1) + elif wavelen > low_freq_wavelen: + rope_factors.append(factor) + else: + smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor) + rope_factors.append(1 / ((1 - smooth) / factor + smooth)) + + if not self.is_lora: + self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32)) + + super().prepare_tensors() + + +###### CONVERSION LOGIC ###### # tree of lazy tensors class LazyTorchTensor(gguf.LazyBase): @@ -3321,19 +3993,46 @@ class LazyTorchTensor(gguf.LazyBase): torch.float32: np.float32, } + # used for safetensors slices + # ref: https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/src/lib.rs#L1046 + # TODO: uncomment U64, U32, and U16, ref: https://github.com/pytorch/pytorch/issues/58734 + _dtype_str_map: dict[str, torch.dtype] = { + "F64": torch.float64, + "F32": torch.float32, + "BF16": torch.bfloat16, + "F16": torch.float16, + # "U64": torch.uint64, + "I64": torch.int64, + # "U32": torch.uint32, + "I32": torch.int32, + # "U16": torch.uint16, + "I16": torch.int16, + "U8": torch.uint8, + "I8": torch.int8, + "BOOL": torch.bool, + "F8_E4M3": torch.float8_e4m3fn, + "F8_E5M2": torch.float8_e5m2, + } + def numpy(self) -> gguf.LazyNumpyTensor: dtype = self._dtype_map[self.dtype] return gguf.LazyNumpyTensor( meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(dtype, self.shape), - lazy=self._lazy, args=(self,), - func=(lambda s: s[0].numpy()) + func=(lambda s: s.numpy()) ) @classmethod - def meta_with_dtype_and_shape(cls, dtype: torch.dtype, shape: torch.Size) -> Tensor: + def meta_with_dtype_and_shape(cls, dtype: torch.dtype, shape: tuple[int, ...]) -> Tensor: return torch.empty(size=shape, dtype=dtype, device="meta") + @classmethod + def from_safetensors_slice(cls, st_slice: Any) -> Tensor: + dtype = cls._dtype_str_map[st_slice.get_dtype()] + shape: tuple[int, ...] = tuple(st_slice.get_shape()) + lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[:]) + return cast(torch.Tensor, lazy) + @classmethod def __torch_function__(cls, func, types, args=(), kwargs=None): del types # unused @@ -3344,7 +4043,7 @@ class LazyTorchTensor(gguf.LazyBase): if func is torch.Tensor.numpy: return args[0].numpy() - return LazyTorchTensor._wrap_fn(func)(*args, **kwargs) + return cls._wrap_fn(func)(*args, **kwargs) def parse_args() -> argparse.Namespace: @@ -3354,10 +4053,6 @@ def parse_args() -> argparse.Namespace: "--vocab-only", action="store_true", help="extract only the vocab", ) - parser.add_argument( - "--awq-path", type=Path, default=None, - help="Path to scale awq cache file", - ) parser.add_argument( "--outfile", type=Path, help="path to write to; default: based on input. {ftype} will be replaced by the outtype.", @@ -3406,6 +4101,10 @@ def parse_args() -> argparse.Namespace: "--no-tensor-first-split", action="store_true", help="do not add tensors to the first split (disabled by default)" ) + parser.add_argument( + "--metadata", type=Path, + help="Specify the path for an authorship metadata override file" + ) return parser.parse_args() @@ -3431,23 +4130,13 @@ def split_str_to_n_bytes(split_str: str) -> int: def main() -> None: args = parse_args() - logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + if args.verbose: + logging.basicConfig(level=logging.DEBUG) + else: + logging.basicConfig(level=logging.INFO) dir_model = args.model - if args.awq_path: - sys.path.insert(1, str(Path(__file__).parent / 'awq-py')) - from awq.apply_awq import add_scale_weights # type: ignore[import-not-found] - tmp_model_path = args.model / "weighted_model" - dir_model = tmp_model_path - if tmp_model_path.is_dir(): - logger.info(f"{tmp_model_path} exists as a weighted model.") - else: - tmp_model_path.mkdir(parents=True, exist_ok=True) - logger.info("Saving new weighted model ...") - add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path)) - logger.info(f"Saved weighted model at {tmp_model_path}.") - if not dir_model.is_dir(): logger.error(f'Error: {args.model} is not a directory') sys.exit(1) @@ -3468,33 +4157,30 @@ def main() -> None: if args.outfile is not None: fname_out = args.outfile else: - # output in the same directory as the model by default - fname_out = dir_model / 'ggml-model-{ftype}.gguf' + fname_out = dir_model logger.info(f"Loading model: {dir_model.name}") hparams = Model.load_hparams(dir_model) with torch.inference_mode(): + output_type = ftype_map[args.outtype] + model_architecture = hparams["architectures"][0] + try: - model_class = Model.from_model_architecture(hparams["architectures"][0]) + model_class = Model.from_model_architecture(model_architecture) except NotImplementedError: - logger.error(f"Model {hparams['architectures'][0]} is not supported") + logger.error(f"Model {model_architecture} is not supported") sys.exit(1) - model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file, - args.no_lazy, args.model_name, split_max_tensors=args.split_max_tensors, + model_instance = model_class(dir_model=dir_model, ftype=output_type, fname_out=fname_out, + is_big_endian=args.bigendian, use_temp_file=args.use_temp_file, + eager=args.no_lazy, + metadata_override=args.metadata, model_name=args.model_name, + split_max_tensors=args.split_max_tensors, split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run, small_first_shard=args.no_tensor_first_split) - logger.info("Set model parameters") - model_instance.set_gguf_parameters() - - logger.info("Set model tokenizer") - model_instance.set_vocab() - - model_instance.gguf_writer.add_quantization_version(gguf.GGML_QUANT_VERSION) - if args.vocab_only: logger.info("Exporting model vocab...") model_instance.write_vocab() diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py index 21a306255..ff4955f9c 100755 --- a/convert_hf_to_gguf_update.py +++ b/convert_hf_to_gguf_update.py @@ -2,7 +2,7 @@ # -*- coding: utf-8 -*- # This script downloads the tokenizer models of the specified models from Huggingface and -# generates the get_vocab_base_pre() function for convert-hf-to-gguf.py +# generates the get_vocab_base_pre() function for convert_hf_to_gguf.py # # This is necessary in order to analyze the type of pre-tokenizer used by the model and # provide the necessary information to llama.cpp via the GGUF header in order to implement @@ -15,9 +15,9 @@ # - Add a new model to the "models" list # - Run the script with your huggingface token: # -# python3 convert-hf-to-gguf-update.py +# python3 convert_hf_to_gguf_update.py # -# - Copy-paste the generated get_vocab_base_pre() function into convert-hf-to-gguf.py +# - Copy-paste the generated get_vocab_base_pre() function into convert_hf_to_gguf.py # - Update llama.cpp with the new pre-tokenizer if necessary # # TODO: generate tokenizer tests for llama.cpp @@ -37,7 +37,7 @@ from enum import IntEnum, auto from transformers import AutoTokenizer logging.basicConfig(level=logging.DEBUG) -logger = logging.getLogger("convert-hf-to-gguf-update") +logger = logging.getLogger("convert_hf_to_gguf_update") sess = requests.Session() @@ -50,16 +50,16 @@ class TOKENIZER_TYPE(IntEnum): # TODO: this string has to exercise as much pre-tokenizer functionality as possible # will be updated with time - contributions welcome -chktxt = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL' +CHK_TXT = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL' if len(sys.argv) == 2: token = sys.argv[1] if not token.startswith("hf_"): logger.info("Huggingface token seems invalid") - logger.info("Usage: python convert-hf-to-gguf-update.py ") + logger.info("Usage: python convert_hf_to_gguf_update.py ") sys.exit(1) else: - logger.info("Usage: python convert-hf-to-gguf-update.py ") + logger.info("Usage: python convert_hf_to_gguf_update.py ") sys.exit(1) # TODO: add models here, base models preferred @@ -91,6 +91,12 @@ models = [ {"name": "gemma-2", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", }, {"name": "jais", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", }, {"name": "t5", "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", }, + {"name": "codeshell", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", }, + {"name": "tekken", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", }, + {"name": "smollm", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", }, + {'name': "bloom", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigscience/bloom", }, + {'name': "gpt3-finnish", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", }, + {"name": "exaone", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", }, ] @@ -99,8 +105,8 @@ def download_file_with_auth(url, token, save_path): response = sess.get(url, headers=headers) response.raise_for_status() os.makedirs(os.path.dirname(save_path), exist_ok=True) - with open(save_path, 'wb') as f: - f.write(response.content) + with open(save_path, 'wb') as downloaded_file: + downloaded_file.write(response.content) logger.info(f"File {save_path} downloaded successfully") @@ -134,7 +140,7 @@ for model in models: logger.error(f"Failed to download model {model['name']}. Error: {e}") -# generate the source code for the convert-hf-to-gguf.py:get_vocab_base_pre() function: +# generate the source code for the convert_hf_to_gguf.py:get_vocab_base_pre() function: src_ifs = "" for model in models: @@ -159,7 +165,7 @@ for model in models: logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}") continue # Skip to the next model if the tokenizer can't be loaded - chktok = tokenizer.encode(chktxt) + chktok = tokenizer.encode(CHK_TXT) chkhsh = sha256(str(chktok).encode()).hexdigest() logger.info(f"model: {name}") @@ -191,7 +197,7 @@ src_func = f""" # we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can # use in llama.cpp to implement the same pre-tokenizer - chktxt = {repr(chktxt)} + chktxt = {repr(CHK_TXT)} chktok = tokenizer.encode(chktxt) chkhsh = sha256(str(chktok).encode()).hexdigest() @@ -201,7 +207,7 @@ src_func = f""" res = None - # NOTE: if you get an error here, you need to update the convert-hf-to-gguf-update.py script + # NOTE: if you get an error here, you need to update the convert_hf_to_gguf_update.py script # or pull the latest version of the model from Huggingface # don't edit the hashes manually! {src_ifs} @@ -210,9 +216,9 @@ src_func = f""" logger.warning("**************************************************************************************") logger.warning("** WARNING: The BPE pre-tokenizer was not recognized!") logger.warning("** There are 2 possible reasons for this:") - logger.warning("** - the model has not been added to convert-hf-to-gguf-update.py yet") + logger.warning("** - the model has not been added to convert_hf_to_gguf_update.py yet") logger.warning("** - the pre-tokenization config has changed upstream") - logger.warning("** Check your model files and convert-hf-to-gguf-update.py and update them accordingly.") + logger.warning("** Check your model files and convert_hf_to_gguf_update.py and update them accordingly.") logger.warning("** ref: https://github.com/ggerganov/llama.cpp/pull/6920") logger.warning("**") logger.warning(f"** chkhsh: {{chkhsh}}") @@ -226,7 +232,7 @@ src_func = f""" return res """ -convert_py_pth = pathlib.Path("convert-hf-to-gguf.py") +convert_py_pth = pathlib.Path("convert_hf_to_gguf.py") convert_py = convert_py_pth.read_text(encoding="utf-8") convert_py = re.sub( r"(# Marker: Start get_vocab_base_pre)(.+?)( +# Marker: End get_vocab_base_pre)", @@ -237,7 +243,7 @@ convert_py = re.sub( convert_py_pth.write_text(convert_py, encoding="utf-8") -logger.info("+++ convert-hf-to-gguf.py was updated") +logger.info("+++ convert_hf_to_gguf.py was updated") # generate tests for each tokenizer model @@ -287,7 +293,7 @@ tests = [ "333333333", "Cửa Việt", # llama-bpe fails on this " discards", - chktxt, + CHK_TXT, ] # write the tests to ./models/ggml-vocab-{name}.gguf.inp @@ -343,6 +349,6 @@ logger.info("\nRun the following commands to generate the vocab files for testin for model in models: name = model["name"] - print(f"python3 convert-hf-to-gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only") # noqa: NP100 + print(f"python3 convert_hf_to_gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only") # noqa: NP100 logger.info("\n") diff --git a/convert_llama_ggml_to_gguf.py b/convert_llama_ggml_to_gguf.py index 9349de3b3..29b14e98d 100755 --- a/convert_llama_ggml_to_gguf.py +++ b/convert_llama_ggml_to_gguf.py @@ -116,7 +116,7 @@ class Tensor: assert quant is not None, 'Unknown tensor type' (blksize, tysize) = quant offset += 12 - self.dtype= dtype + self.dtype= gguf.GGMLQuantizationType(dtype) self.dims = struct.unpack(f'<{n_dims}I', data[offset:offset + (4 * n_dims)]) offset += 4 * n_dims self.name = bytes(data[offset:offset + name_len]) @@ -132,6 +132,10 @@ class Tensor: class GGMLModel: + + file_format: GGMLFormat + format_version: int + def __init__(self): self.hyperparameters = None self.vocab = None @@ -290,7 +294,7 @@ class GGMLToGGUF: if self.vocab_override is not None: vo = self.vocab_override logger.info('* Adding vocab item(s)') - for (idx, (vbytes, score, ttype)) in enumerate(vo.all_tokens()): + for (_, (vbytes, score, ttype)) in enumerate(vo.all_tokens()): tokens.append(vbytes) scores.append(score) toktypes.append(ttype) @@ -354,7 +358,8 @@ class GGMLToGGUF: def handle_metadata(cfg, hp): - import convert + import examples.convert_legacy_llama as convert + assert cfg.model_metadata_dir.is_dir(), 'Metadata dir is not a directory' hf_config_path = cfg.model_metadata_dir / "config.json" orig_config_path = cfg.model_metadata_dir / "params.json" diff --git a/convert_lora_to_gguf.py b/convert_lora_to_gguf.py new file mode 100755 index 000000000..ddd347a2a --- /dev/null +++ b/convert_lora_to_gguf.py @@ -0,0 +1,394 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- + +from __future__ import annotations + +from dataclasses import dataclass +import logging +import argparse +import os +import sys +import json +from math import prod +from pathlib import Path +from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast + +import torch + +if TYPE_CHECKING: + from torch import Tensor + +if 'NO_LOCAL_GGUF' not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) +import gguf + +# reuse model definitions from convert_hf_to_gguf.py +from convert_hf_to_gguf import LazyTorchTensor, Model + +logger = logging.getLogger("lora-to-gguf") + + +@dataclass +class PartialLoraTensor: + A: Tensor | None = None + B: Tensor | None = None + + +# magic to support tensor shape modifications and splitting +class LoraTorchTensor: + _lora_A: Tensor # (n_rank, row_size) + _lora_B: Tensor # (col_size, n_rank) + _rank: int + + def __init__(self, A: Tensor, B: Tensor): + assert len(A.shape) == len(B.shape) + assert A.shape[-2] == B.shape[-1] + if A.dtype != B.dtype: + A = A.to(torch.float32) + B = B.to(torch.float32) + self._lora_A = A + self._lora_B = B + self._rank = B.shape[-1] + + def get_lora_A_B(self) -> tuple[Tensor, Tensor]: + return (self._lora_A, self._lora_B) + + def __getitem__( + self, + indices: ( + SupportsIndex + | slice + | tuple[SupportsIndex | slice | Tensor, ...] # TODO: add ellipsis in the type signature + ), + ) -> LoraTorchTensor: + shape = self.shape + if isinstance(indices, SupportsIndex): + if len(shape) > 2: + return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices]) + else: + raise NotImplementedError # can't return a vector + elif isinstance(indices, slice): + if len(shape) > 2: + return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices]) + else: + return LoraTorchTensor(self._lora_A, self._lora_B[indices]) + elif isinstance(indices, tuple): + assert len(indices) > 0 + if indices[-1] is Ellipsis: + return self[indices[:-1]] + # expand ellipsis + indices = tuple( + u + for v in ( + ( + (slice(None, None) for _ in range(len(indices) - 1)) + if i is Ellipsis + else (i,) + ) + for i in indices + ) + for u in v + ) + + if len(indices) < len(shape): + indices = (*indices, *(slice(None, None) for _ in range(len(indices), len(shape)))) + + # TODO: make sure this is correct + indices_A = ( + *( + ( + j.__index__() % self._lora_A.shape[i] + if isinstance(j, SupportsIndex) + else slice(None, None) + ) + for i, j in enumerate(indices[:-2]) + ), + slice(None, None), + indices[-1], + ) + indices_B = indices[:-1] + return LoraTorchTensor(self._lora_A[indices_A], self._lora_B[indices_B]) + else: + raise NotImplementedError # unknown indice type + + @property + def dtype(self) -> torch.dtype: + assert self._lora_A.dtype == self._lora_B.dtype + return self._lora_A.dtype + + @property + def shape(self) -> tuple[int, ...]: + assert len(self._lora_A.shape) == len(self._lora_B.shape) + return (*self._lora_B.shape[:-1], self._lora_A.shape[-1]) + + def size(self, dim=None): + assert dim is None + return self.shape + + def reshape(self, *shape: int | tuple[int, ...]) -> LoraTorchTensor: + if isinstance(shape[0], tuple): + new_shape: tuple[int, ...] = shape[0] + else: + new_shape = cast(tuple[int, ...], shape) + orig_shape = self.shape + if len(new_shape) < 2: + raise NotImplementedError # can't become a vector + + # expand -1 in the shape + if any(dim == -1 for dim in new_shape): + n_elems = prod(orig_shape) + n_new_elems = prod(dim if dim != -1 else 1 for dim in new_shape) + assert n_elems % n_new_elems == 0 + new_shape = (*(dim if dim != -1 else n_elems // n_new_elems for dim in new_shape),) + + if new_shape[-1] != orig_shape[-1]: + raise NotImplementedError # can't reshape the row size trivially + + shape_A = (*(1 for _ in new_shape[:-2]), self._rank, orig_shape[-1]) + shape_B = (*new_shape[:-1], self._rank) + return LoraTorchTensor( + self._lora_A.reshape(shape_A), + self._lora_B.reshape(shape_B), + ) + + def reshape_as(self, other: Tensor) -> LoraTorchTensor: + return self.reshape(*other.shape) + + def view(self, *size: int) -> LoraTorchTensor: + return self.reshape(*size) + + def permute(self, *dims: int) -> LoraTorchTensor: + shape = self.shape + dims = tuple(dim - len(shape) if dim >= 0 else dim for dim in dims) + if dims[-1] == -1: + # TODO: support higher dimensional A shapes bigger than 1 + assert all(dim == 1 for dim in self._lora_A.shape[:-2]) + return LoraTorchTensor(self._lora_A, self._lora_B.permute(*dims)) + if len(shape) == 2 and dims[-1] == -2 and dims[-2] == -1: + return LoraTorchTensor(self._lora_B.permute(*dims), self._lora_A.permute(*dims)) + else: + # TODO: compose the above two + raise NotImplementedError + + def transpose(self, dim0: int, dim1: int) -> LoraTorchTensor: + shape = self.shape + dims = [i for i in range(len(shape))] + dims[dim0], dims[dim1] = dims[dim1], dims[dim0] + return self.permute(*dims) + + def swapaxes(self, axis0: int, axis1: int) -> LoraTorchTensor: + return self.transpose(axis0, axis1) + + def to(self, *args, **kwargs): + return LoraTorchTensor(self._lora_A.to(*args, **kwargs), self._lora_B.to(*args, **kwargs)) + + @classmethod + def __torch_function__(cls, func: Callable, types, args=(), kwargs=None): + del types # unused + + if kwargs is None: + kwargs = {} + + if func is torch.permute: + return type(args[0]).permute(*args, **kwargs) + elif func is torch.reshape: + return type(args[0]).reshape(*args, **kwargs) + elif func is torch.stack: + assert isinstance(args[0], Sequence) + dim = kwargs.get("dim", 0) + assert dim == 0 + return LoraTorchTensor( + torch.stack([a._lora_A for a in args[0]], dim), + torch.stack([b._lora_B for b in args[0]], dim), + ) + elif func is torch.cat: + assert isinstance(args[0], Sequence) + dim = kwargs.get("dim", 0) + assert dim == 0 + if len(args[0][0].shape) > 2: + return LoraTorchTensor( + torch.cat([a._lora_A for a in args[0]], dim), + torch.cat([b._lora_B for b in args[0]], dim), + ) + elif all(torch.equal(args[0][0]._lora_A, t._lora_A) for t in args[0][1:]): + return LoraTorchTensor( + args[0][0]._lora_A, + torch.cat([b._lora_B for b in args[0]], dim), + ) + else: + raise NotImplementedError + else: + raise NotImplementedError + + +def get_base_tensor_name(lora_tensor_name: str) -> str: + base_name = lora_tensor_name.replace("base_model.model.", "") + base_name = base_name.replace(".lora_A.weight", ".weight") + base_name = base_name.replace(".lora_B.weight", ".weight") + return base_name + + +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser( + description="Convert a huggingface PEFT LoRA adapter to a GGML compatible file") + parser.add_argument( + "--outfile", type=Path, + help="path to write to; default: based on input. {ftype} will be replaced by the outtype.", + ) + parser.add_argument( + "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16", + help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type", + ) + parser.add_argument( + "--bigendian", action="store_true", + help="model is executed on big endian machine", + ) + parser.add_argument( + "--no-lazy", action="store_true", + help="use more RAM by computing all outputs before writing (use in case lazy evaluation is broken)", + ) + parser.add_argument( + "--verbose", action="store_true", + help="increase output verbosity", + ) + parser.add_argument( + "--dry-run", action="store_true", + help="only print out what will be done, without writing any new files", + ) + parser.add_argument( + "--base", type=Path, required=True, + help="directory containing base model file", + ) + parser.add_argument( + "lora_path", type=Path, + help="directory containing LoRA adapter file", + ) + + return parser.parse_args() + + +if __name__ == '__main__': + args = parse_args() + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + + ftype_map: dict[str, gguf.LlamaFileType] = { + "f32": gguf.LlamaFileType.ALL_F32, + "f16": gguf.LlamaFileType.MOSTLY_F16, + "bf16": gguf.LlamaFileType.MOSTLY_BF16, + "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0, + "auto": gguf.LlamaFileType.GUESSED, + } + + ftype = ftype_map[args.outtype] + + dir_base_model: Path = args.base + dir_lora: Path = args.lora_path + lora_config = dir_lora / "adapter_config.json" + input_model = dir_lora / "adapter_model.safetensors" + + if args.outfile is not None: + fname_out = args.outfile + else: + # output in the same directory as the model by default + fname_out = dir_lora + + if os.path.exists(input_model): + # lazy import load_file only if lora is in safetensors format. + from safetensors.torch import load_file + + lora_model = load_file(input_model, device="cpu") + else: + input_model = os.path.join(dir_lora, "adapter_model.bin") + lora_model = torch.load(input_model, map_location="cpu", weights_only=True) + + # load base model + logger.info(f"Loading base model: {dir_base_model.name}") + hparams = Model.load_hparams(dir_base_model) + with torch.inference_mode(): + try: + model_class = Model.from_model_architecture(hparams["architectures"][0]) + except NotImplementedError: + logger.error(f"Model {hparams['architectures'][0]} is not supported") + sys.exit(1) + + class LoraModel(model_class): + model_arch = model_class.model_arch + + lora_alpha: float + + def __init__(self, *args, dir_lora_model: Path, lora_alpha: float, **kwargs): + + super().__init__(*args, **kwargs) + + self.dir_model_card = dir_lora_model + self.lora_alpha = float(lora_alpha) + + def set_type(self): + self.gguf_writer.add_type(gguf.GGUFType.ADAPTER) + self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora") + + def set_gguf_parameters(self): + self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha) + super().set_gguf_parameters() + + def get_tensors(self) -> Iterator[tuple[str, Tensor]]: + tensor_map: dict[str, PartialLoraTensor] = {} + + for name, tensor in lora_model.items(): + if self.lazy: + tensor = LazyTorchTensor.from_eager(tensor) + base_name = get_base_tensor_name(name) + is_lora_a = ".lora_A.weight" in name + is_lora_b = ".lora_B.weight" in name + if not is_lora_a and not is_lora_b: + if ".base_layer.weight" in name: + continue + logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor") + sys.exit(1) + + if base_name in tensor_map: + if is_lora_a: + tensor_map[base_name].A = tensor + else: + tensor_map[base_name].B = tensor + else: + if is_lora_a: + tensor_map[base_name] = PartialLoraTensor(A=tensor) + else: + tensor_map[base_name] = PartialLoraTensor(B=tensor) + + for name, tensor in tensor_map.items(): + assert tensor.A is not None + assert tensor.B is not None + yield (name, cast(torch.Tensor, LoraTorchTensor(tensor.A, tensor.B))) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + dest = super().modify_tensors(data_torch, name, bid) + for dest_name, dest_data in dest: + assert isinstance(dest_data, LoraTorchTensor) + lora_a, lora_b = dest_data.get_lora_A_B() + + yield (dest_name + ".lora_a", lora_a) + yield (dest_name + ".lora_b", lora_b) + + with open(lora_config, "r") as f: + lparams: dict[str, Any] = json.load(f) + + alpha: float = lparams["lora_alpha"] + + model_instance = LoraModel( + dir_base_model, + ftype, + fname_out, + is_big_endian=args.bigendian, + use_temp_file=False, + eager=args.no_lazy, + dry_run=args.dry_run, + dir_lora_model=dir_lora, + lora_alpha=alpha, + is_lora=True, + ) + + logger.info("Exporting model...") + model_instance.write() + logger.info(f"Model successfully exported to {model_instance.fname_out}") diff --git a/docs/android.md b/docs/android.md new file mode 100644 index 000000000..cec4358d9 --- /dev/null +++ b/docs/android.md @@ -0,0 +1,56 @@ + +# Android + +## Build on Android using Termux +[Termux](https://github.com/termux/termux-app#installation) is a method to execute `llama.cpp` on an Android device (no root required). +``` +apt update && apt upgrade -y +apt install git make cmake +``` + +It's recommended to move your model inside the `~/` directory for best performance: +``` +cd storage/downloads +mv model.gguf ~/ +``` + +[Get the code](https://github.com/ggerganov/llama.cpp#get-the-code) & [follow the Linux build instructions](https://github.com/ggerganov/llama.cpp#build) to build `llama.cpp`. + +## Building the Project using Android NDK +Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake. + +Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux: +``` +$ mkdir build-android +$ cd build-android +$ export NDK= +$ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod .. +$ make +``` + +Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice). + +Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission: + +(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`) +``` +$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/ +$cd /data/data/com.termux/files/home/bin +$chmod +x ./* +``` + +Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/` + +``` +$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/ +``` + +Now, you can start chatting: +``` +$cd /data/data/com.termux/files/home/bin +$./llama-cli -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml +``` + +Here's a demo of an interactive session running on Pixel 5 phone: + +https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4 diff --git a/docs/BLIS.md b/docs/backend/BLIS.md similarity index 100% rename from docs/BLIS.md rename to docs/backend/BLIS.md diff --git a/docs/backend/CANN.md b/docs/backend/CANN.md new file mode 100644 index 000000000..6bdd9d2da --- /dev/null +++ b/docs/backend/CANN.md @@ -0,0 +1,259 @@ +# llama.cpp for CANN + + - [Background](#background) + - [News](#news) + - [OS](#os) + - [Hardware](#hardware) + - [Model Supports](#model-supports) + - [DataType Supports](#datatype-supports) + - [Docker](#docker) + - [Linux](#linux) + - [TODO](#todo) + + +## Background + +**Ascend NPU** is a range of AI processors using Neural Processing Unit. It will efficiently handle matrix-matrix multiplication, dot-product and scalars. + +**CANN** (Compute Architecture for Neural Networks) is a heterogeneous computing architecture for AI scenarios, providing support for multiple AI frameworks on the top and serving AI processors and programming at the bottom. It plays a crucial role in bridging the gap between upper and lower layers, and is a key platform for improving the computing efficiency of Ascend AI processors. Meanwhile, it offers a highly efficient and easy-to-use programming interface for diverse application scenarios, allowing users to rapidly build AI applications and services based on the Ascend platform. + +**Llama.cpp + CANN** + +The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the ability of AscendC and ACLNN which are intergrated to CANN Toolkit and kernels to using Ascend NPU directly. + +## News + +- 2024.8 + - Support `Q4_0` and `Q8_0` data type for Ascend NPU. +- 2024.7 + - Create CANN backend for Ascend NPU. + +## OS + +| OS | Status | Verified | +|:-------:|:-------:|:----------------------------------------------:| +| Linux | Support | Ubuntu 22.04, OpenEuler22.03 | + + +## Hardware + +### Ascend NPU + +**Verified devices** +| Ascend NPU | Status | +|:-----------------------------:|:-------:| +| Atlas 300T A2 | Support | + +*Notes:* + +- If you have trouble with Ascend NPU device, please create a issue with **[CANN]** prefix/tag. +- If you run successfully with your Ascend NPU device, please help update the upper table. + + +## Model Supports + +| Model Name | FP16 | Q8_0 | Q4_0 | +|:----------------------------|:-----:|:----:|:----:| +| AquilaChat2-7B | √ | √ | √ | +| Baichuan-7b | √ | √ | √ | +| Baichuan2-7B-Chat | √ | √ | √ | +| bitnet_b1_58-large | √ | √ | √ | +| bloom-560m | √ | x | √ | +| bloomz-alpaca-560m | √ | x | √ | +| c4ai-command-r-35B-v01 | x | x | x | +| chatglm3-6B | x | x | x | +| chinese-alpaca-2-1.3b | √ | √ | √ | +| CodeShell-7B | √ | √ | √ | +| deepseek-ai_deepseek-coder-1.3B-base | x | x | x | +| deepseek-ai_DeepSeek-V2-Lite | x | x | x | +| deepseek-coder-6.7B-instruct | x | x | x | +| DeepSeek-V2-Lite-64x1.5B | x | x | x | +| falcon-7b-instruct | √ | √ | √ | +| flan-t5-large | √ | √ | √ | +| gemma-2-9b-it | √ | √ | √ | +| glm-4-9B | x | x | x | +| gpt2 | √ | √ | √ | +| Gpt2-163M | √ | √ | √ | +| granite-3B-code-instruct | √ | √ | √ | +| GritLM-7B | √ | √ | √ | +| internlm2_5-7b-chat | √ | √ | √ | +| koala-7B-HF | √ | √ | √ | +| Llama-2-7b-chat-hf | √ | √ | √ | +| Llama-3-Smaug-8B | √ | √ | √ | +| Llama2-Chinese-7b-Chat | √ | √ | √ | +| Llama3-8B | √ | √ | √ | +| Llama3-8b-chinese | √ | √ | √ | +| mamba-130m-hf | √ | √ | √ | +| Mistral-7B-Instruct-v0.2 | √ | √ | √ | +| Mixtral-8x7B-Instruct-v0.1 | x | √ | √ | +| mpt-7B | √ | √ | √ | +| OLMo-1B-hf | √ | √ | √ | +| OpenELM-3B-Instruct | √ | √ | √ | +| Orion-14b-base | √ | √ | √ | +| phi1 | x | x | x | +| phi2 | x | x | x | +| Phi-3-mini-4k-instruct | √ | √ | √ | +| plamo-13b | √ | √ | √ | +| pythia-70M | x | x | x | +| Qwen-7B | √ | √ | √ | +| Qwen2-1.5B-Instruct | √ | x | √ | +| Refact-1_6B-fim | √ | √ | √ | +| SmolLM-135M | √ | √ | √ | +| stablelm-zephyr | x | x | x | +| stablelm-2-zephyr-1_6b | x | x | x | +| starcoderbase-1b | √ | √ | √ | +| starcoder2-3b | √ | √ | √ | +| vigogne-7b-chat | √ | √ | √ | +| xverse-7b-chat | √ | √ | √ | +| Yi-6b-Chat | √ | √ | √ | + + + +## DataType Supports + +| DataType | Status | +|:----------------------:|:-------:| +| FP16 | Support | +| Q8_0 | Support | +| Q4_0 | Support | + +## Docker + +### Build Images +You can get a image with llama.cpp in one command. +```sh +docker build -t llama-cpp-cann -f .devops/llama-cli-cann.Dockerfile . +``` + +### Run container + +```sh +# Find all cards. +npu-smi info + +# Select the cards that you want to use, make sure these cards are not used by someone. +# Following using cards of device0. +docker run --name llamacpp --device /dev/davinci0 --device /dev/davinci_manager --device /dev/devmm_svm --device /dev/hisi_hdc -v /usr/local/dcmi:/usr/local/dcmi -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info -v /PATH_TO_YOUR_MODELS/:/app/models -it llama-cpp-cann -m /app/models/MODEL_PATH -ngl 32 -p "Building a website can be done in 10 simple steps:" +``` + +*Notes:* + +- You may need to install Ascend Driver and firmware on the **host** machine *(Please refer to the [Linux configuration](#linux) for details)*. + +## Linux + +### I. Setup Environment + +1. **Install Ascend Driver and firmware** + + ```sh + # create driver running user. + sudo groupadd -g HwHiAiUser + sudo useradd -g HwHiAiUser -d /home/HwHiAiUser -m HwHiAiUser -s /bin/bash + sudo usermod -aG HwHiAiUser $USER + + # download driver from https://www.hiascend.com/hardware/firmware-drivers/community according to your system + # and install driver. + sudo sh Ascend-hdk-910b-npu-driver_x.x.x_linux-{arch}.run --full --install-for-all + ``` + + Once installed, run `npu-smi info` to check whether driver is installed successfully. + ```sh + +-------------------------------------------------------------------------------------------+ + | npu-smi 24.1.rc2 Version: 24.1.rc2 | + +----------------------+---------------+----------------------------------------------------+ + | NPU Name | Health | Power(W) Temp(C) Hugepages-Usage(page)| + | Chip | Bus-Id | AICore(%) Memory-Usage(MB) HBM-Usage(MB) | + +======================+===============+====================================================+ + | 2 xxx | OK | 64.4 51 15 / 15 | + | 0 | 0000:01:00.0 | 0 1873 / 15077 0 / 32768 | + +======================+===============+====================================================+ + | 5 xxx | OK | 64.0 52 15 / 15 | + | 0 | 0000:81:00.0 | 0 1874 / 15077 0 / 32768 | + +======================+===============+====================================================+ + | No running processes found in NPU 2 | + +======================+===============+====================================================+ + | No running processes found in NPU 5 | + +======================+===============+====================================================+ + ``` + +2. **Install Ascend Firmware** + ```sh + # download driver from https://www.hiascend.com/hardware/firmware-drivers/community according to your system + # and install driver. + sudo sh Ascend-hdk-910b-npu-firmware_x.x.x.x.X.run --full + ``` + If the following messaage appers, firmware is installed successfully. + ```sh + Firmware package installed successfully! + ``` + + +3. **Install CANN toolkit and kernels** + + CANN toolkit and kernels can be obtained from the official [CANN Toolkit](https://www.hiascend.com/zh/developer/download/community/result?module=cann) page. + + Please download the corresponding version that satified your system. The minimum version required is 8.0.RC2.alpha002 and here is the install command. + ```sh + pip3 install attrs numpy decorator sympy cffi pyyaml pathlib2 psutil protobuf scipy requests absl-py wheel typing_extensions + sh Ascend-cann-toolkit_8.0.RC2.alpha002_linux-aarch64.run --install + sh Ascend-cann-kernels-910b_8.0.RC2.alpha002_linux.run --install + ``` + + Set Ascend Variables: + ```sh + echo "source ~/Ascend/ascend-toolkit/set_env.sh" >> ~/.bashrc + source ~/.bashrc + ``` + +Upon a successful installation, CANN is enabled for the available ascend devices. + +### II. Build llama.cpp + +```sh +cmake -B build -DGGML_CANN=on -DCMAKE_BUILD_TYPE=release +cmake --build build --config release +``` + +### III. Run the inference + +1. **Retrieve and prepare model** + + You can refer to the general [*Prepare and Quantize*](../../README.md#prepare-and-quantize) guide for model prepration. + + **Notes**: + + - CANN backend only supports FP16/Q4_0/Q8_0 models currently. + +2. **Launch inference** + + There are two device selection modes: + + - Single device: Use one device target specified by the user. + - Multiple devices: Automatically choose the devices with the same backend. + + | Device selection | Parameter | + |:----------------:|:--------------------------------------:| + | Single device | --split-mode none --main-gpu DEVICE_ID | + | Multiple devices | --split-mode layer (default) | + + Examples: + + - Use device 0: + + ```sh + ./build/bin/llama-cli -m path_to_model -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0 + ``` + + - Use multiple devices: + + ```sh + ./build/bin/llama-cli -m path_to_model -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer + ``` + +### **GitHub contribution**: +Please add the **[CANN]** prefix/tag in issues/PRs titles to help the CANN-team check/address them without delay. + + +## TODO +- Support more models and data types. diff --git a/README-sycl.md b/docs/backend/SYCL.md similarity index 83% rename from README-sycl.md rename to docs/backend/SYCL.md index 885983e92..e3b9572cc 100644 --- a/README-sycl.md +++ b/docs/backend/SYCL.md @@ -20,7 +20,7 @@ **oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include: - **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers. -- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL - Math Kernel Library)*. +- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL and oneDNN)*. - **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs. - **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets. @@ -28,10 +28,6 @@ The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it could support other vendor GPUs: Nvidia GPU (*AMD GPU coming*). -When targeting **Intel CPU**, it is recommended to use llama.cpp for [Intel oneMKL](README.md#intel-onemkl) backend. - -It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS, cuBLAS, etc..*. In beginning work, the oneAPI's [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) open-source migration tool (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) was used for this purpose. - ## Recommended Release The SYCL backend would be broken by some PRs due to no online CI. @@ -45,6 +41,10 @@ The following release is verified with good quality: ## News + +- 2024.8 + - Use oneDNN as the default GEMM library, improve the compatibility for new Intel GPUs. + - 2024.5 - Performance is increased: 34 -> 37 tokens/s of llama-2-7b.Q4_0 on Arc770. - Arch Linux is verified successfully. @@ -80,7 +80,14 @@ The following release is verified with good quality: ### Intel GPU -**Verified devices** +SYCL backend supports Intel GPU Family: + +- Intel Data Center Max Series +- Intel Flex Series, Arc Series +- Intel Built-in Arc GPU +- Intel iGPU in Core CPU (11th Generation Core CPU and newer, refer to [oneAPI supported GPU](https://www.intel.com/content/www/us/en/developer/articles/system-requirements/intel-oneapi-base-toolkit-system-requirements.html#inpage-nav-1-1)). + +#### Verified devices | Intel GPU | Status | Verified Model | |-------------------------------|---------|---------------------------------------| @@ -88,7 +95,7 @@ The following release is verified with good quality: | Intel Data Center Flex Series | Support | Flex 170 | | Intel Arc Series | Support | Arc 770, 730M, Arc A750 | | Intel built-in Arc GPU | Support | built-in Arc GPU in Meteor Lake | -| Intel iGPU | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 | +| Intel iGPU | Support | iGPU in 13700k, i5-1250P, i7-1260P, i7-1165G7 | *Notes:* @@ -189,7 +196,7 @@ Please follow the instructions for downloading and installing the Toolkit for Li Following guidelines/code snippets assume the default installation values. Otherwise, please make sure the necessary changes are reflected where applicable. -Upon a successful installation, SYCL is enabled for the available intel devices, along with relevant libraries such as oneAPI MKL for intel GPUs. +Upon a successful installation, SYCL is enabled for the available intel devices, along with relevant libraries such as oneAPI oneDNN for Intel GPUs. - **Adding support to Nvidia GPUs** @@ -237,12 +244,17 @@ Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA devic ### II. Build llama.cpp #### Intel GPU + +``` +./examples/sycl/build.sh +``` + +or + ```sh # Export relevant ENV variables source /opt/intel/oneapi/setvars.sh -# Build LLAMA with MKL BLAS acceleration for intel GPU - # Option 1: Use FP32 (recommended for better performance in most cases) cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx @@ -276,48 +288,71 @@ cmake --build build --config Release -j -v ### III. Run the inference -1. Retrieve and prepare model +#### Retrieve and prepare model You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example. -2. Enable oneAPI running environment +##### Check device + +1. Enable oneAPI running environment ```sh source /opt/intel/oneapi/setvars.sh ``` -3. List devices information +2. List devices information Similar to the native `sycl-ls`, available SYCL devices can be queried as follow: ```sh ./build/bin/llama-ls-sycl-device ``` -A example of such log in a system with 1 *intel CPU* and 1 *intel GPU* can look like the following: + +This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following: ``` -found 6 SYCL devices: +found 2 SYCL devices: + | | | |Compute |Max compute|Max work|Max sub| | |ID| Device Type| Name|capability|units |group |group |Global mem size| |--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------| | 0|[level_zero:gpu:0]| Intel(R) Arc(TM) A770 Graphics| 1.3| 512| 1024| 32| 16225243136| | 1|[level_zero:gpu:1]| Intel(R) UHD Graphics 770| 1.3| 32| 512| 32| 53651849216| -| 2| [opencl:gpu:0]| Intel(R) Arc(TM) A770 Graphics| 3.0| 512| 1024| 32| 16225243136| -| 3| [opencl:gpu:1]| Intel(R) UHD Graphics 770| 3.0| 32| 512| 32| 53651849216| -| 4| [opencl:cpu:0]| 13th Gen Intel(R) Core(TM) i7-13700K| 3.0| 24| 8192| 64| 67064815616| -| 5| [opencl:acc:0]| Intel(R) FPGA Emulation Device| 1.2| 24|67108864| 64| 67064815616| ``` -| Attribute | Note | -|------------------------|-------------------------------------------------------------| -| compute capability 1.3 | Level-zero driver/runtime, recommended | -| compute capability 3.0 | OpenCL driver/runtime, slower than level-zero in most cases | +#### Choose level-zero devices -4. Launch inference +|Chosen Device ID|Setting| +|-|-| +|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action| +|1|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"`| +|0 & 1|`export ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`| + +#### Execute + +Choose one of following methods to run. + +1. Script + +- Use device 0: + +```sh +./examples/sycl/run-llama2.sh 0 +``` +- Use multiple devices: + +```sh +./examples/sycl/run-llama2.sh +``` + +2. Command line +Launch inference There are two device selection modes: -- Single device: Use one device target specified by the user. -- Multiple devices: Automatically select the devices with the same largest Max compute-units. +- Single device: Use one device assigned by user. Default device id is 0. +- Multiple devices: Automatically choose the devices with the same backend. + +In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR. | Device selection | Parameter | |------------------|----------------------------------------| @@ -331,11 +366,6 @@ Examples: ```sh ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0 ``` -or run by script: - -```sh -./examples/sycl/run_llama2.sh 0 -``` - Use multiple devices: @@ -343,12 +373,6 @@ or run by script: ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer ``` -Otherwise, you can run the script: - -```sh -./examples/sycl/run_llama2.sh -``` - *Notes:* - Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow: @@ -395,7 +419,7 @@ c. Verify installation In the oneAPI command line, run the following to print the available SYCL devices: ``` -sycl-ls +sycl-ls.exe ``` There should be one or more *level-zero* GPU devices displayed as **[ext_oneapi_level_zero:gpu]**. Below is example of such output detecting an *intel Iris Xe* GPU as a Level-zero SYCL device: @@ -416,6 +440,18 @@ b. The new Visual Studio will install Ninja as default. (If not, please install ### II. Build llama.cpp +You could download the release package for Windows directly, which including binary files and depended oneAPI dll files. + +Choose one of following methods to build from source code. + +1. Script + +```sh +.\examples\sycl\win-build-sycl.bat +``` + +2. CMake + On the oneAPI command line window, step into the llama.cpp main directory and run the following: ``` @@ -430,12 +466,8 @@ cmake -B build -G "Ninja" -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPI cmake --build build --config Release -j ``` -Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions: -```sh -.\examples\sycl\win-build-sycl.bat -``` - Or, use CMake presets to build: + ```sh cmake --preset x64-windows-sycl-release cmake --build build-x64-windows-sycl-release -j --target llama-cli @@ -447,7 +479,9 @@ cmake --preset x64-windows-sycl-debug cmake --build build-x64-windows-sycl-debug -j --target llama-cli ``` -Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project. +3. Visual Studio + +You can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project. *Notes:* @@ -455,52 +489,65 @@ Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choos ### III. Run the inference -1. Retrieve and prepare model +#### Retrieve and prepare model -You can refer to the general [*Prepare and Quantize*](README#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example. +You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example. -2. Enable oneAPI running environment +##### Check device + +1. Enable oneAPI running environment On the oneAPI command line window, run the following and step into the llama.cpp directory: ``` "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 ``` -3. List devices information +2. List devices information Similar to the native `sycl-ls`, available SYCL devices can be queried as follow: ``` -build\bin\ls-sycl-device.exe +build\bin\llama-ls-sycl-device.exe ``` -The output of this command in a system with 1 *intel CPU* and 1 *intel GPU* would look like the following: +This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following: ``` -found 6 SYCL devices: +found 2 SYCL devices: | | | |Compute |Max compute|Max work|Max sub| | |ID| Device Type| Name|capability|units |group |group |Global mem size| |--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------| | 0|[level_zero:gpu:0]| Intel(R) Arc(TM) A770 Graphics| 1.3| 512| 1024| 32| 16225243136| | 1|[level_zero:gpu:1]| Intel(R) UHD Graphics 770| 1.3| 32| 512| 32| 53651849216| -| 2| [opencl:gpu:0]| Intel(R) Arc(TM) A770 Graphics| 3.0| 512| 1024| 32| 16225243136| -| 3| [opencl:gpu:1]| Intel(R) UHD Graphics 770| 3.0| 32| 512| 32| 53651849216| -| 4| [opencl:cpu:0]| 13th Gen Intel(R) Core(TM) i7-13700K| 3.0| 24| 8192| 64| 67064815616| -| 5| [opencl:acc:0]| Intel(R) FPGA Emulation Device| 1.2| 24|67108864| 64| 67064815616| ``` +#### Choose level-zero devices -| Attribute | Note | -|------------------------|-----------------------------------------------------------| -| compute capability 1.3 | Level-zero running time, recommended | -| compute capability 3.0 | OpenCL running time, slower than level-zero in most cases | +|Chosen Device ID|Setting| +|-|-| +|0|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action| +|1|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"`| +|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`| +#### Execute -4. Launch inference +Choose one of following methods to run. + +1. Script + +``` +examples\sycl\win-run-llama2.bat +``` + +2. Command line + +Launch inference There are two device selection modes: -- Single device: Use one device assigned by user. -- Multiple devices: Automatically choose the devices with the same biggest Max compute units. +- Single device: Use one device assigned by user. Default device id is 0. +- Multiple devices: Automatically choose the devices with the same backend. + +In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR. | Device selection | Parameter | |------------------|----------------------------------------| @@ -520,11 +567,7 @@ build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website ca ``` build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer ``` -Otherwise, run the following wrapper script: -``` -.\examples\sycl\win-run-llama2.bat -``` Note: @@ -538,17 +581,18 @@ Or use 1 SYCL GPUs: [0] with Max compute units:512 ``` + ## Environment Variable #### Build | Name | Value | Function | |--------------------|-----------------------------------|---------------------------------------------| -| GGML_SYCL | ON (mandatory) | Enable build with SYCL code path. | +| GGML_SYCL | ON (mandatory) | Enable build with SYCL code path.
FP32 path - recommended for better perforemance than FP16 on quantized model| | GGML_SYCL_TARGET | INTEL *(default)* \| NVIDIA | Set the SYCL target device type. | | GGML_SYCL_F16 | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path. | -| CMAKE_C_COMPILER | icx | Set *icx* compiler for SYCL code path. | -| CMAKE_CXX_COMPILER | icpx *(Linux)*, icx *(Windows)* | Set `icpx/icx` compiler for SYCL code path. | +| CMAKE_C_COMPILER | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path. | +| CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)* | Set `icpx/icx` compiler for SYCL code path. | #### Runtime @@ -584,9 +628,18 @@ use 1 SYCL GPUs: [0] with Max compute units:512 ``` Otherwise, please double-check the GPU driver installation steps. +- Can I report Ollama issue on Intel GPU to llama.cpp SYCL backend? + + No. We can't support Ollama issue directly, because we aren't familiar with Ollama. + + Sugguest reproducing on llama.cpp and report similar issue to llama.cpp. We will surpport it. + + It's same for other projects including llama.cpp SYCL backend. + + ### **GitHub contribution**: Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay. ## TODO -- Support row layer split for multiple card runs. +- NA diff --git a/docs/build.md b/docs/build.md new file mode 100644 index 000000000..152d46d6f --- /dev/null +++ b/docs/build.md @@ -0,0 +1,382 @@ +# Build llama.cpp locally + +**To get the Code:** + +```bash +git clone https://github.com/ggerganov/llama.cpp +cd llama.cpp +``` + +In order to build llama.cpp you have four different options. + +- Using `make`: + - On Linux or MacOS: + + ```bash + make + ``` + + - On Windows (x86/x64 only, arm64 requires cmake): + + 1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases). + 2. Extract `w64devkit` on your pc. + 3. Run `w64devkit.exe`. + 4. Use the `cd` command to reach the `llama.cpp` folder. + 5. From here you can run: + ```bash + make + ``` + + - Notes: + - For `Q4_0_4_4` quantization type build, add the `GGML_NO_LLAMAFILE=1` flag. For example, use `make GGML_NO_LLAMAFILE=1`. + - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel. + - For faster repeated compilation, install [ccache](https://ccache.dev/). + - For debug builds, run `make LLAMA_DEBUG=1` + +- Using `CMake`: + + ```bash + cmake -B build + cmake --build build --config Release + ``` + + **Notes**: + + - For `Q4_0_4_4` quantization type build, add the `-DGGML_LLAMAFILE=OFF` cmake option. For example, use `cmake -B build -DGGML_LLAMAFILE=OFF`. + - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel. + - For faster repeated compilation, install [ccache](https://ccache.dev/). + - For debug builds, there are two cases: + + 1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag): + + ```bash + cmake -B build -DCMAKE_BUILD_TYPE=Debug + cmake --build build + ``` + + 2. Multi-config generators (`-G` param set to Visual Studio, XCode...): + + ```bash + cmake -B build -G "Xcode" + cmake --build build --config Debug + ``` + - Building for Windows (x86, x64 and arm64) with MSVC or clang as compilers: + - Install Visual Studio 2022, e.g. via the [Community Edition](https://visualstudio.microsoft.com/de/vs/community/). In the installer, select at least the following options (this also automatically installs the required additional tools like CMake,...): + - Tab Workload: Desktop-development with C++ + - Tab Components (select quickly via search): C++-_CMake_ Tools for Windows, _Git_ for Windows, C++-_Clang_ Compiler for Windows, MS-Build Support for LLVM-Toolset (clang) + - Please remember to always use a Developer Command Prompt / PowerShell for VS2022 for git, build, test + - For Windows on ARM (arm64, WoA) build with: + ```bash + cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF + cmake --build build-arm64-windows-llvm-release + ``` + Note: Building for arm64 could also be done just with MSVC (with the build-arm64-windows-MSVC preset, or the standard CMake build instructions). But MSVC does not support inline ARM assembly-code, used e.g. for the accelerated Q4_0_4_8 CPU kernels. + +- Using `gmake` (FreeBSD): + + 1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics) + 2. Add your user to **video** group + 3. Install compilation dependencies. + + ```bash + sudo pkg install gmake automake autoconf pkgconf llvm15 openblas + + gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4 + ``` + +## Metal Build + +On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU. +To disable the Metal build at compile time use the `GGML_NO_METAL=1` flag or the `GGML_METAL=OFF` cmake option. + +When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line +argument. + +## BLAS Build + +Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS. There are currently several different BLAS implementations available for build and use: + +### Accelerate Framework: + +This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions. + +### OpenBLAS: + +This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS installed on your machine. + +- Using `make`: + - On Linux: + ```bash + make GGML_OPENBLAS=1 + ``` + + - On Windows: + + 1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases). + 2. Download the latest version of [OpenBLAS for Windows](https://github.com/xianyi/OpenBLAS/releases). + 3. Extract `w64devkit` on your pc. + 4. From the OpenBLAS zip that you just downloaded copy `libopenblas.a`, located inside the `lib` folder, inside `w64devkit\x86_64-w64-mingw32\lib`. + 5. From the same OpenBLAS zip copy the content of the `include` folder inside `w64devkit\x86_64-w64-mingw32\include`. + 6. Run `w64devkit.exe`. + 7. Use the `cd` command to reach the `llama.cpp` folder. + 8. From here you can run: + + ```bash + make GGML_OPENBLAS=1 + ``` + +- Using `CMake` on Linux: + + ```bash + cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS + cmake --build build --config Release + ``` + +### BLIS + +Check [BLIS.md](./backend/BLIS.md) for more information. + +### SYCL + +SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators. + +llama.cpp based on SYCL is used to **support Intel GPU** (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU). + +For detailed info, please refer to [llama.cpp for SYCL](./backend/SYCL.md). + +### Intel oneMKL + +Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni. Please note that this build config **does not support Intel GPU**. For Intel GPU support, please refer to [llama.cpp for SYCL](./backend/SYCL.md). + +- Using manual oneAPI installation: + By default, `GGML_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DGGML_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. Otherwise please install oneAPI and follow the below steps: + ```bash + source /opt/intel/oneapi/setvars.sh # You can skip this step if in oneapi-basekit docker image, only required for manual installation + cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_NATIVE=ON + cmake --build build --config Release + ``` + +- Using oneAPI docker image: + If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-basekit](https://hub.docker.com/r/intel/oneapi-basekit). Then, you can use the commands given above. + +Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information. + +### CUDA + +This provides GPU acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads). + +For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling. + +- Using `make`: + ```bash + make GGML_CUDA=1 + ``` +- Using `CMake`: + + ```bash + cmake -B build -DGGML_CUDA=ON + cmake --build build --config Release + ``` + +The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. + +The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`. + +The following compilation options are also available to tweak performance: + +| Option | Legal values | Default | Description | +|-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| +| GGML_CUDA_FORCE_DMMV | Boolean | false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. | +| GGML_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. | +| GGML_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. | +| GGML_CUDA_FORCE_MMQ | Boolean | false | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower. | +| GGML_CUDA_FORCE_CUBLAS | Boolean | false | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models | +| GGML_CUDA_F16 | Boolean | false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. | +| GGML_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. | +| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer | 128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. | +| GGML_CUDA_FA_ALL_QUANTS | Boolean | false | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer. | + +### MUSA + +- Using `make`: + ```bash + make GGML_MUSA=1 + ``` +- Using `CMake`: + + ```bash + cmake -B build -DGGML_MUSA=ON + cmake --build build --config Release + ``` + +### hipBLAS + +This provides BLAS acceleration on HIP-supported AMD GPUs. +Make sure to have ROCm installed. +You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html#rocm-install-quick). + +- Using `make`: + ```bash + make GGML_HIPBLAS=1 + ``` +- Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU): + ```bash + HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \ + cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \ + && cmake --build build --config Release -- -j 16 + ``` + On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`. + However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs). + + Note that if you get the following error: + ``` + clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library + ``` + Try searching for a directory under `HIP_PATH` that contains the file + `oclc_abi_version_400.bc`. Then, add the following to the start of the + command: `HIP_DEVICE_LIB_PATH=`, so something + like: + ```bash + HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \ + HIP_DEVICE_LIB_PATH= \ + cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \ + && cmake --build build -- -j 16 + ``` + +- Using `make` (example for target gfx1030, build with 16 CPU threads): + ```bash + make -j16 GGML_HIPBLAS=1 GGML_HIP_UMA=1 AMDGPU_TARGETS=gfx1030 + ``` + +- Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU): + ```bash + set PATH=%HIP_PATH%\bin;%PATH% + cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release + cmake --build build + ``` + Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors) + Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`. + + +The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used. +If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3. +The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above): + +| Option | Legal values | Default | Description | +|------------------------|------------------------|---------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| +| GGML_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. | +| GGML_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. | +| GGML_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. | + +### Vulkan + +**Windows** + +#### w64devkit + +Download and extract [w64devkit](https://github.com/skeeto/w64devkit/releases). + +Download and install the [Vulkan SDK](https://vulkan.lunarg.com/sdk/home#windows). When selecting components, only the Vulkan SDK Core is required. + +Launch `w64devkit.exe` and run the following commands to copy Vulkan dependencies: +```sh +SDK_VERSION=1.3.283.0 +cp /VulkanSDK/$SDK_VERSION/Bin/glslc.exe $W64DEVKIT_HOME/bin/ +cp /VulkanSDK/$SDK_VERSION/Lib/vulkan-1.lib $W64DEVKIT_HOME/x86_64-w64-mingw32/lib/ +cp -r /VulkanSDK/$SDK_VERSION/Include/* $W64DEVKIT_HOME/x86_64-w64-mingw32/include/ +cat > $W64DEVKIT_HOME/x86_64-w64-mingw32/lib/pkgconfig/vulkan.pc <' --install --attr llama-cpp +``` + +For non-flake enabled installs. + +This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164). + +## Flox + +On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via + +```sh +flox install llama-cpp +``` + +Flox follows the nixpkgs build of llama.cpp. diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index 7d9ab3457..67b3d2774 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -21,8 +21,8 @@ else() add_subdirectory(embedding) add_subdirectory(eval-callback) add_subdirectory(export-lora) - add_subdirectory(finetune) add_subdirectory(gbnf-validator) + add_subdirectory(gguf-hash) add_subdirectory(gguf-split) add_subdirectory(gguf) add_subdirectory(gritlm) @@ -52,5 +52,4 @@ else() add_subdirectory(simple) add_subdirectory(speculative) add_subdirectory(tokenize) - add_subdirectory(train-text-from-scratch) endif() diff --git a/examples/baby-llama/baby-llama.cpp b/examples/baby-llama/baby-llama.cpp index 4f6c3746a..3ce91070b 100644 --- a/examples/baby-llama/baby-llama.cpp +++ b/examples/baby-llama/baby-llama.cpp @@ -1,7 +1,6 @@ #include "ggml.h" #include "train.h" -#include #include #include #include @@ -19,7 +18,7 @@ constexpr float rms_norm_eps = 5e-6f; #endif static void ggml_graph_compute_helper(std::vector & buf, ggml_cgraph * graph, int n_threads) { - struct ggml_cplan plan = ggml_graph_plan(graph, n_threads); + struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr); if (plan.work_size > 0) { buf.resize(plan.work_size); diff --git a/examples/batched-bench/batched-bench.cpp b/examples/batched-bench/batched-bench.cpp index 114dd811e..f3ce4964f 100644 --- a/examples/batched-bench/batched-bench.cpp +++ b/examples/batched-bench/batched-bench.cpp @@ -69,7 +69,7 @@ int main(int argc, char ** argv) { llama_context_params ctx_params = llama_context_params_from_gpt_params(params); // ensure enough sequences are available - ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end()); + ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end()); llama_context * ctx = llama_new_context_with_model(model, ctx_params); diff --git a/examples/batched.swift/Sources/main.swift b/examples/batched.swift/Sources/main.swift index 443a03d57..2a7324e6b 100644 --- a/examples/batched.swift/Sources/main.swift +++ b/examples/batched.swift/Sources/main.swift @@ -229,7 +229,7 @@ private func tokenize(text: String, add_bos: Bool) -> [llama_token] { private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String? { var result = [CChar](repeating: 0, count: 8) - let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count), false) + let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count), 0, false) if nTokens < 0 { let actualTokensCount = -Int(nTokens) result = .init(repeating: 0, count: actualTokensCount) @@ -238,6 +238,7 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String token, &result, Int32(result.count), + 0, false ) assert(check == actualTokensCount) diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp index 3e4ea9e78..00cd744d4 100644 --- a/examples/batched/batched.cpp +++ b/examples/batched/batched.cpp @@ -31,7 +31,7 @@ int main(int argc, char ** argv) { int n_parallel = params.n_parallel; // total length of the sequences including the prompt - int n_predict = 32; + int n_predict = params.n_predict; // init LLM diff --git a/examples/benchmark/benchmark-matmult.cpp b/examples/benchmark/benchmark-matmult.cpp index 47cb16c69..97622f4f4 100644 --- a/examples/benchmark/benchmark-matmult.cpp +++ b/examples/benchmark/benchmark-matmult.cpp @@ -21,7 +21,7 @@ #endif static void ggml_graph_compute_helper(std::vector & buf, ggml_cgraph * graph, int n_threads) { - struct ggml_cplan plan = ggml_graph_plan(graph, n_threads); + struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr); if (plan.work_size > 0) { buf.resize(plan.work_size); @@ -54,7 +54,7 @@ static void tensor_dump(const ggml_tensor * tensor, const char * name) { #define TENSOR_DUMP(tensor) tensor_dump(tensor, #tensor) struct benchmark_params_struct { - int32_t n_threads = 1; + int n_threads = 1; int32_t n_iterations = 10; }; diff --git a/examples/convert-legacy-llama.py b/examples/convert_legacy_llama.py similarity index 88% rename from examples/convert-legacy-llama.py rename to examples/convert_legacy_llama.py index 721a57c00..9ab9ab06e 100755 --- a/examples/convert-legacy-llama.py +++ b/examples/convert_legacy_llama.py @@ -24,7 +24,7 @@ from abc import ABC, abstractmethod from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor from dataclasses import dataclass from pathlib import Path -from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar, Optional +from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar import numpy as np @@ -346,42 +346,6 @@ class Params: return params -@dataclass -class Metadata: - name: Optional[str] = None - author: Optional[str] = None - version: Optional[str] = None - url: Optional[str] = None - description: Optional[str] = None - licence: Optional[str] = None - source_url: Optional[str] = None - source_hf_repo: Optional[str] = None - - @staticmethod - def load(metadata_path: Path) -> Metadata: - if metadata_path is None or not metadata_path.exists(): - return Metadata() - - with open(metadata_path, 'r') as file: - data = json.load(file) - - # Create a new Metadata instance - metadata = Metadata() - - # Assigning values to Metadata attributes if they exist in the JSON file - # This is based on LLM_KV_NAMES mapping in llama.cpp - metadata.name = data.get("general.name") - metadata.author = data.get("general.author") - metadata.version = data.get("general.version") - metadata.url = data.get("general.url") - metadata.description = data.get("general.description") - metadata.license = data.get("general.license") - metadata.source_url = data.get("general.source.url") - metadata.source_hf_repo = data.get("general.source.huggingface.repository") - - return metadata - - # # data loading # TODO: reuse (probably move to gguf.py?) @@ -492,12 +456,13 @@ class LazyTensor: LazyModel: TypeAlias = 'dict[str, LazyTensor]' +ModelFormat: TypeAlias = Literal['ggml', 'torch', 'safetensors', 'none'] @dataclass class ModelPlus: model: LazyModel paths: list[Path] # Where this was read from. - format: Literal['ggml', 'torch', 'safetensors', 'none'] + format: ModelFormat vocab: BaseVocab | None # For GGML models (which have vocab built in), the vocab. @@ -536,7 +501,7 @@ def merge_sharded(models: list[LazyModel]) -> LazyModel: def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus: - formats = set(mp.format for mp in models_plus) + formats: set[ModelFormat] = set(mp.format for mp in models_plus) assert len(formats) == 1, "different formats?" format = formats.pop() paths = [path for mp in models_plus for path in mp.paths] @@ -555,7 +520,7 @@ def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus: else: model = merge_sharded([mp.model for mp in models_plus]) - return ModelPlus(model, paths, format, vocab) # pytype: disable=wrong-arg-types + return ModelPlus(model, paths, format, vocab) def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTensor: @@ -805,7 +770,7 @@ class OutputFile: def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE): self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess) - def add_meta_model(self, params: Params, metadata: Metadata) -> None: + def add_meta_model(self, params: Params, metadata: gguf.Metadata | None) -> None: # Metadata About The Model And Its Provenence name = "LLaMA" if metadata is not None and metadata.name is not None: @@ -823,16 +788,73 @@ class OutputFile: self.gguf.add_author(metadata.author) if metadata.version is not None: self.gguf.add_version(metadata.version) - if metadata.url is not None: - self.gguf.add_url(metadata.url) + if metadata.organization is not None: + self.gguf.add_organization(metadata.organization) + + if metadata.finetune is not None: + self.gguf.add_finetune(metadata.finetune) + if metadata.basename is not None: + self.gguf.add_basename(metadata.basename) + if metadata.description is not None: self.gguf.add_description(metadata.description) - if metadata.licence is not None: - self.gguf.add_licence(metadata.licence) + if metadata.quantized_by is not None: + self.gguf.add_quantized_by(metadata.quantized_by) + + if metadata.size_label is not None: + self.gguf.add_size_label(metadata.size_label) + + if metadata.license is not None: + self.gguf.add_license(metadata.license) + if metadata.license_name is not None: + self.gguf.add_license_name(metadata.license_name) + if metadata.license_link is not None: + self.gguf.add_license_link(metadata.license_link) + + if metadata.url is not None: + self.gguf.add_url(metadata.url) + if metadata.doi is not None: + self.gguf.add_doi(metadata.doi) + if metadata.uuid is not None: + self.gguf.add_uuid(metadata.uuid) + if metadata.repo_url is not None: + self.gguf.add_repo_url(metadata.repo_url) + if metadata.source_url is not None: self.gguf.add_source_url(metadata.source_url) - if metadata.source_hf_repo is not None: - self.gguf.add_source_hf_repo(metadata.source_hf_repo) + if metadata.source_doi is not None: + self.gguf.add_source_doi(metadata.source_doi) + if metadata.source_uuid is not None: + self.gguf.add_source_uuid(metadata.source_uuid) + if metadata.source_repo_url is not None: + self.gguf.add_source_repo_url(metadata.source_repo_url) + + if metadata.base_models is not None: + self.gguf.add_base_model_count(len(metadata.base_models)) + for key, base_model_entry in enumerate(metadata.base_models): + if "name" in base_model_entry: + self.gguf.add_base_model_name(key, base_model_entry["name"]) + if "author" in base_model_entry: + self.gguf.add_base_model_author(key, base_model_entry["author"]) + if "version" in base_model_entry: + self.gguf.add_base_model_version(key, base_model_entry["version"]) + if "organization" in base_model_entry: + self.gguf.add_base_model_organization(key, base_model_entry["organization"]) + if "url" in base_model_entry: + self.gguf.add_base_model_url(key, base_model_entry["url"]) + if "doi" in base_model_entry: + self.gguf.add_base_model_doi(key, base_model_entry["doi"]) + if "uuid" in base_model_entry: + self.gguf.add_base_model_uuid(key, base_model_entry["uuid"]) + if "repo_url" in base_model_entry: + self.gguf.add_base_model_repo_url(key, base_model_entry["repo_url"]) + + if metadata.tags is not None: + self.gguf.add_tags(metadata.tags) + if metadata.languages is not None: + self.gguf.add_languages(metadata.languages) + if metadata.datasets is not None: + self.gguf.add_datasets(metadata.datasets) def add_meta_arch(self, params: Params) -> None: # Metadata About The Neural Architecture Itself @@ -943,7 +965,7 @@ class OutputFile: @staticmethod def write_vocab_only( fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab, - endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: Metadata = None, + endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: gguf.Metadata | None = None, ) -> None: check_vocab_size(params, vocab, pad_vocab=pad_vocab) @@ -977,7 +999,7 @@ class OutputFile: fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, - metadata: Metadata = None, + metadata: gguf.Metadata | None = None, ) -> None: check_vocab_size(params, vocab, pad_vocab=pad_vocab) @@ -1020,35 +1042,32 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT raise ValueError(f"Unexpected combination of types: {name_to_type}") -def model_parameter_count(model: LazyModel) -> int: - total_model_parameters = 0 - for i, (name, lazy_tensor) in enumerate(model.items()): - sum_weights_in_tensor = 1 +def per_model_weight_count_estimation(tensors: Iterable[tuple[str, LazyTensor]]) -> tuple[int, int, int]: + total_params = 0 + shared_params = 0 + expert_params = 0 + + for name, lazy_tensor in tensors: + # We don't need these + if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")): + continue + + # Got A Tensor + sum_weights_in_tensor: int = 1 + + # Tensor Volume for dim in lazy_tensor.shape: sum_weights_in_tensor *= dim - total_model_parameters += sum_weights_in_tensor - return total_model_parameters + if ".experts." in name: + if ".experts.0." in name: + expert_params += sum_weights_in_tensor + else: + shared_params += sum_weights_in_tensor -def model_parameter_count_rounded_notation(model_params_count: int) -> str: - if model_params_count > 1e12 : - # Trillions Of Parameters - scaled_model_params = model_params_count * 1e-12 - scale_suffix = "T" - elif model_params_count > 1e9 : - # Billions Of Parameters - scaled_model_params = model_params_count * 1e-9 - scale_suffix = "B" - elif model_params_count > 1e6 : - # Millions Of Parameters - scaled_model_params = model_params_count * 1e-6 - scale_suffix = "M" - else: - # Thousands Of Parameters - scaled_model_params = model_params_count * 1e-3 - scale_suffix = "K" + total_params += sum_weights_in_tensor - return f"{round(scaled_model_params)}{scale_suffix}" + return total_params, shared_params, expert_params def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel: @@ -1230,34 +1249,24 @@ class VocabFactory: return vocab, special_vocab -def default_convention_outfile(file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> str: - quantization = { +def default_convention_outfile(file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> str: + name = metadata.name if metadata.name is not None else None + basename = metadata.basename if metadata.basename is not None else None + finetune = metadata.finetune if metadata.finetune is not None else None + version = metadata.version if metadata.version is not None else None + size_label = metadata.size_label if metadata.size_label is not None else gguf.size_label(*model_params_count, expert_count=expert_count or 0) + + output_type = { GGMLFileType.AllF32: "F32", GGMLFileType.MostlyF16: "F16", GGMLFileType.MostlyQ8_0: "Q8_0", }[file_type] - parameters = model_parameter_count_rounded_notation(model_params_count) - - expert_count = "" - if params.n_experts is not None: - expert_count = f"{params.n_experts}x" - - version = "" - if metadata is not None and metadata.version is not None: - version = f"-{metadata.version}" - - name = "ggml-model" - if metadata is not None and metadata.name is not None: - name = metadata.name - elif params.path_model is not None: - name = params.path_model.name - - return f"{name}{version}-{expert_count}{parameters}-{quantization}" + return gguf.naming_convention(name, basename, finetune, version, size_label, output_type) -def default_outfile(model_paths: list[Path], file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> Path: - default_filename = default_convention_outfile(file_type, params, model_params_count, metadata) +def default_outfile(model_paths: list[Path], file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> Path: + default_filename = default_convention_outfile(file_type, expert_count, model_params_count, metadata) ret = model_paths[0].parent / f"{default_filename}.gguf" if ret in model_paths: logger.error( @@ -1296,8 +1305,9 @@ def main(args_in: list[str] | None = None) -> None: parser.add_argument("--pad-vocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides") parser.add_argument("--skip-unknown", action="store_true", help="skip unknown tensor names instead of failing") parser.add_argument("--verbose", action="store_true", help="increase output verbosity") - parser.add_argument("--metadata", type=Path, help="Specify the path for a metadata file") + parser.add_argument("--metadata", type=Path, help="Specify the path for an authorship metadata override file") parser.add_argument("--get-outfile", action="store_true", help="get calculated default outfile name") + parser.add_argument("--model-name", type=str, default=None, help="name of the model") args = parser.parse_args(args_in) @@ -1309,32 +1319,36 @@ def main(args_in: list[str] | None = None) -> None: else: logging.basicConfig(level=logging.INFO) - metadata = Metadata.load(args.metadata) + model_name = args.model_name + dir_model = args.model + + metadata = gguf.Metadata.load(args.metadata, dir_model, model_name) if args.get_outfile: - model_plus = load_some_model(args.model) + model_plus = load_some_model(dir_model) params = Params.load(model_plus) - model = convert_model_names(model_plus.model, params, args.skip_unknown) - model_params_count = model_parameter_count(model_plus.model) - ftype = pick_output_type(model, args.outtype) - print(f"{default_convention_outfile(ftype, params, model_params_count, metadata)}") # noqa: NP100 + model = convert_model_names(model_plus.model, params, args.skip_unknown) + model_params_count = per_model_weight_count_estimation(model_plus.model.items()) + ftype = pick_output_type(model, args.outtype) + + if (metadata is None or metadata.name is None) and params.path_model is not None: + metadata.name = params.path_model.name + + print(f"{default_convention_outfile(ftype, params.n_experts, model_params_count, metadata)}") # noqa: NP100 return if args.no_vocab and args.vocab_only: raise ValueError("--vocab-only does not make sense with --no-vocab") if args.dump_single: - model_plus = lazy_load_file(args.model) + model_plus = lazy_load_file(dir_model) do_dump_model(model_plus) return if not args.vocab_only: - model_plus = load_some_model(args.model) + model_plus = load_some_model(dir_model) else: - model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None) - - model_params_count = model_parameter_count(model_plus.model) - logger.info(f"model parameters count : {model_params_count} ({model_parameter_count_rounded_notation(model_params_count)})") + model_plus = ModelPlus(model = {}, paths = [dir_model / 'dummy'], format = 'none', vocab = None) if args.dump: do_dump_model(model_plus) @@ -1367,7 +1381,7 @@ def main(args_in: list[str] | None = None) -> None: logger.info(f"params = {params}") model_parent_path = model_plus.paths[0].parent - vocab_path = Path(args.vocab_dir or args.model or model_parent_path) + vocab_path = Path(args.vocab_dir or dir_model or model_parent_path) vocab_factory = VocabFactory(vocab_path) vocab_types = None if args.no_vocab else args.vocab_type.split(",") vocab, special_vocab = vocab_factory.load_vocab(vocab_types, model_parent_path) @@ -1396,13 +1410,23 @@ def main(args_in: list[str] | None = None) -> None: if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab: vocab = model_plus.vocab + assert params is not None + + if metadata.name is None and params.path_model is not None: + metadata.name = params.path_model.name + + model_params_count = per_model_weight_count_estimation(model_plus.model.items()) + logger.info(f"model parameters count : {model_params_count} ({gguf.model_weight_count_rounded_notation(model_params_count[0])})") + logger.info(f"Vocab info: {vocab}") logger.info(f"Special vocab info: {special_vocab}") model = model_plus.model model = convert_model_names(model, params, args.skip_unknown) ftype = pick_output_type(model, args.outtype) model = convert_to_output_type(model, ftype) - outfile = args.outfile or default_outfile(model_plus.paths, ftype, params, model_params_count, metadata) + outfile = args.outfile or default_outfile(model_plus.paths, ftype, params.n_experts, model_params_count, metadata=metadata) + + metadata.size_label = gguf.size_label(*model_params_count, expert_count=params.n_experts or 0) params.ftype = ftype logger.info(f"Writing {outfile}, format {ftype}") diff --git a/examples/cvector-generator/cvector-generator.cpp b/examples/cvector-generator/cvector-generator.cpp index 7d27d86ec..0acdfdf38 100644 --- a/examples/cvector-generator/cvector-generator.cpp +++ b/examples/cvector-generator/cvector-generator.cpp @@ -271,7 +271,7 @@ struct tokenized_prompt { size_t max_seq_len; tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) { - const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); + const bool add_bos = llama_add_bos_token(llama_get_model(ctx)); tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true); tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true); max_seq_len = std::max(tokens_pos.size(), tokens_neg.size()); @@ -414,9 +414,10 @@ int main(int argc, char ** argv) { llama_numa_init(params.numa); // load the model to get hparams - llama_model * model; - llama_context * ctx; - std::tie(model, ctx) = llama_init_from_gpt_params(params); + llama_init_result llama_init = llama_init_from_gpt_params(params); + + llama_model * model = llama_init.model; + llama_context * ctx = llama_init.context; // int n_ctx = llama_n_ctx(ctx); int n_layers = llama_n_layer(model); @@ -485,8 +486,8 @@ int main(int argc, char ** argv) { if (use_pca) { // run PCA PCA::pca_params pca_params; - pca_params.n_threads = params.n_threads; - pca_params.n_batch = params.n_pca_batch; + pca_params.n_threads = params.cpuparams.n_threads; + pca_params.n_batch = params.n_pca_batch; pca_params.n_iterations = params.n_pca_iterations; PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final); } else { diff --git a/examples/deprecation-warning/README.md b/examples/deprecation-warning/README.md new file mode 100644 index 000000000..59918ec2b --- /dev/null +++ b/examples/deprecation-warning/README.md @@ -0,0 +1,49 @@ +# Migration notice for binary filenames + +> [!IMPORTANT] +[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809) + +This migration was important, but it is a breaking change that may not always be immediately obvious to users. + +Please update all scripts and workflows to use the new binary names. + +| Old Filename | New Filename | +| ---- | ---- | +| main | llama-cli | +| server | llama-server | +| llama-bench | llama-bench | +| embedding | llama-embedding | +| quantize | llama-quantize | +| tokenize | llama-tokenize | +| export-lora | llama-export-lora | +| libllava.a | libllava.a | +| baby-llama | llama-baby-llama | +| batched | llama-batched | +| batched-bench | llama-batched-bench | +| benchmark-matmult | llama-benchmark-matmult | +| convert-llama2c-to-ggml | llama-convert-llama2c-to-ggml | +| eval-callback | llama-eval-callback | +| gbnf-validator | llama-gbnf-validator | +| gguf | llama-gguf | +| gguf-split | llama-gguf-split | +| gritlm | llama-gritlm | +| imatrix | llama-imatrix | +| infill | llama-infill | +| llava-cli | llama-llava-cli | +| lookahead | llama-lookahead | +| lookup | llama-lookup | +| lookup-create | llama-lookup-create | +| lookup-merge | llama-lookup-merge | +| lookup-stats | llama-lookup-stats | +| parallel | llama-parallel | +| passkey | llama-passkey | +| perplexity | llama-perplexity | +| q8dot | llama-q8dot | +| quantize-stats | llama-quantize-stats | +| retrieval | llama-retrieval | +| save-load-state | llama-save-load-state | +| simple | llama-simple | +| speculative | llama-speculative | +| vdot | llama-vdot | +| tests/test-c.o | tests/test-c.o | + diff --git a/examples/deprecation-warning/deprecation-warning.cpp b/examples/deprecation-warning/deprecation-warning.cpp new file mode 100644 index 000000000..11b35d2c2 --- /dev/null +++ b/examples/deprecation-warning/deprecation-warning.cpp @@ -0,0 +1,35 @@ +// Warns users that this filename was deprecated, and provides a link for more information. + +#include +#include +#include + +// Main +int main(int argc, char** argv) { + std::string filename = "main"; + if (argc >= 1) { + filename = argv[0]; + } + + // Get only the program name from the full path + auto pos = filename.find_last_of('/'); + if (pos != std::string::npos) { + filename = filename.substr(pos+1); + } + + // Append "llama-" to the beginning of filename to get the replacemnt filename + auto replacement_filename = "llama-" + filename; + + // The exception is if the filename is "main", then our replacement filename is "llama-cli" + if (filename == "main") { + replacement_filename = "llama-cli"; + } + + fprintf(stdout, "\n"); + fprintf(stdout, "WARNING: The binary '%s' is deprecated.\n", filename.c_str()); + fprintf(stdout, " Please use '%s' instead.\n", replacement_filename.c_str()); + fprintf(stdout, " See https://github.com/ggerganov/llama.cpp/tree/master/examples/deprecation-warning/README.md for more information.\n"); + fprintf(stdout, "\n"); + + return EXIT_FAILURE; +} diff --git a/examples/embedding/README.md b/examples/embedding/README.md index e3705b454..12b372bf1 100644 --- a/examples/embedding/README.md +++ b/examples/embedding/README.md @@ -9,13 +9,13 @@ To get started right away, run the following command, making sure to use the cor ### Unix-based systems (Linux, macOS, etc.): ```bash -./llama-embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null +./llama-embedding -m ./path/to/model --pooling mean --log-disable -p "Hello World!" 2>/dev/null ``` ### Windows: ```powershell -llama-embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null +llama-embedding.exe -m ./path/to/model --pooling mean --log-disable -p "Hello World!" 2>$null ``` The above command will output space-separated float values. @@ -50,11 +50,11 @@ The above command will output space-separated float values. ### Unix-based systems (Linux, macOS, etc.): ```bash -./embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2 --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null +./llama-embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --pooling mean --embd-separator '<#sep#>' --embd-normalize 2 --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null ``` ### Windows: ```powershell -embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2 --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null +llama-embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --pooling mean --embd-separator '<#sep#>' --embd-normalize 2 --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null ``` diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index e11246fd0..a98b0811d 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -31,13 +31,24 @@ static void batch_add_seq(llama_batch & batch, const std::vector & toke } static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd, int embd_norm) { + const enum llama_pooling_type pooling_type = llama_pooling_type(ctx); + const struct llama_model * model = llama_get_model(ctx); + // clear previous kv_cache values (irrelevant for embeddings) llama_past_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__); + if (llama_model_has_encoder(model) && !llama_model_has_decoder(model)) { + // encoder-only model + if (llama_encode(ctx, batch) < 0) { + fprintf(stderr, "%s : failed to encode\n", __func__); + } + } else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) { + // decoder-only model + if (llama_decode(ctx, batch) < 0) { + fprintf(stderr, "%s : failed to decode\n", __func__); + } } for (int i = 0; i < batch.n_tokens; i++) { @@ -45,11 +56,22 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu 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]); - GGML_ASSERT(embd != NULL && "failed to get sequence embeddings"); + const float * embd = nullptr; + int embd_pos = 0; - float * out = output + batch.seq_id[i][0] * n_embd; + if (pooling_type == LLAMA_POOLING_TYPE_NONE) { + // try to get token embeddings + embd = llama_get_embeddings_ith(ctx, i); + embd_pos = i; + GGML_ASSERT(embd != NULL && "failed to get token embeddings"); + } else { + // try to get sequence embeddings - supported only when pooling_type is not NONE + embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]); + embd_pos = batch.seq_id[i][0]; + GGML_ASSERT(embd != NULL && "failed to get sequence embeddings"); + } + + float * out = output + embd_pos * n_embd; llama_embd_normalize(embd, out, n_embd, embd_norm); } } @@ -79,11 +101,11 @@ int main(int argc, char ** argv) { 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); + llama_init_result llama_init = llama_init_from_gpt_params(params); + + llama_model * model = llama_init.model; + llama_context * ctx = llama_init.context; if (model == NULL) { fprintf(stderr, "%s: error: unable to load model\n", __func__); return 1; @@ -93,8 +115,9 @@ int main(int argc, char ** argv) { const int n_ctx = llama_n_ctx(ctx); const enum llama_pooling_type pooling_type = llama_pooling_type(ctx); - if (pooling_type == LLAMA_POOLING_TYPE_NONE) { - fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__); + + if (llama_model_has_encoder(model) && llama_model_has_decoder(model)) { + fprintf(stderr, "%s: error: computing embeddings in encoder-decoder models is not supported\n", __func__); return 1; } @@ -153,13 +176,23 @@ int main(int argc, char ** argv) { const int n_prompts = prompts.size(); struct llama_batch batch = llama_batch_init(n_batch, 0, 1); + // count number of embeddings + int n_embd_count = 0; + if (pooling_type == LLAMA_POOLING_TYPE_NONE) { + for (int k = 0; k < n_prompts; k++) { + n_embd_count += inputs[k].size(); + } + } else { + n_embd_count = n_prompts; + } + // allocate output const int n_embd = llama_n_embd(model); - std::vector embeddings(n_prompts * n_embd, 0); + std::vector embeddings(n_embd_count * n_embd, 0); float * emb = embeddings.data(); // break into batches - int p = 0; // number of prompts processed already + int e = 0; // number of embeddings already stored int s = 0; // number of prompts in current batch for (int k = 0; k < n_prompts; k++) { // clamp to n_batch tokens @@ -169,11 +202,11 @@ int main(int argc, char ** argv) { // encode if at capacity if (batch.n_tokens + n_toks > n_batch) { - float * out = emb + p * n_embd; + float * out = emb + e * n_embd; batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize); - llama_batch_clear(batch); - p += s; + e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s; s = 0; + llama_batch_clear(batch); } // add to batch @@ -182,40 +215,63 @@ int main(int argc, char ** argv) { } // final batch - float * out = emb + p * n_embd; + float * out = emb + e * n_embd; batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize); if (params.embd_out.empty()) { - // print the first part of the embeddings or for a single prompt, the full embedding fprintf(stdout, "\n"); - for (int j = 0; j < n_prompts; j++) { - fprintf(stdout, "embedding %d: ", j); - for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) { - if (params.embd_normalize == 0) { - fprintf(stdout, "%6.0f ", emb[j * n_embd + i]); - } else { - fprintf(stdout, "%9.6f ", emb[j * n_embd + i]); - } - } - fprintf(stdout, "\n"); - } - // print cosine similarity matrix - if (n_prompts > 1) { - fprintf(stdout, "\n"); - printf("cosine similarity matrix:\n\n"); - for (int i = 0; i < n_prompts; i++) { - fprintf(stdout, "%6.6s ", prompts[i].c_str()); - } - fprintf(stdout, "\n"); - for (int i = 0; i < n_prompts; i++) { - for (int j = 0; j < n_prompts; j++) { - float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd); - fprintf(stdout, "%6.2f ", sim); + if (pooling_type == LLAMA_POOLING_TYPE_NONE) { + for (int j = 0; j < n_embd_count; j++) { + fprintf(stdout, "embedding %d: ", j); + for (int i = 0; i < std::min(3, n_embd); i++) { + if (params.embd_normalize == 0) { + fprintf(stdout, "%6.0f ", emb[j * n_embd + i]); + } else { + fprintf(stdout, "%9.6f ", emb[j * n_embd + i]); + } + } + fprintf(stdout, " ... "); + for (int i = n_embd - 3; i < n_embd; i++) { + if (params.embd_normalize == 0) { + fprintf(stdout, "%6.0f ", emb[j * n_embd + i]); + } else { + fprintf(stdout, "%9.6f ", emb[j * n_embd + i]); + } } - fprintf(stdout, "%1.10s", prompts[i].c_str()); fprintf(stdout, "\n"); } + } else { + // print the first part of the embeddings or for a single prompt, the full embedding + for (int j = 0; j < n_prompts; j++) { + fprintf(stdout, "embedding %d: ", j); + for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) { + if (params.embd_normalize == 0) { + fprintf(stdout, "%6.0f ", emb[j * n_embd + i]); + } else { + fprintf(stdout, "%9.6f ", emb[j * n_embd + i]); + } + } + fprintf(stdout, "\n"); + } + + // print cosine similarity matrix + if (n_prompts > 1) { + fprintf(stdout, "\n"); + printf("cosine similarity matrix:\n\n"); + for (int i = 0; i < n_prompts; i++) { + fprintf(stdout, "%6.6s ", prompts[i].c_str()); + } + fprintf(stdout, "\n"); + for (int i = 0; i < n_prompts; i++) { + for (int j = 0; j < n_prompts; j++) { + float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd); + fprintf(stdout, "%6.2f ", sim); + } + fprintf(stdout, "%1.10s", prompts[i].c_str()); + fprintf(stdout, "\n"); + } + } } } @@ -233,23 +289,23 @@ int main(int argc, char ** argv) { } fprintf(stdout, notArray ? "]\n }" : "]"); j++; - if (j < n_prompts) fprintf(stdout, notArray ? ",\n" : ","); else break; + if (j < n_embd_count) fprintf(stdout, notArray ? ",\n" : ","); else break; } fprintf(stdout, notArray ? "\n ]" : "]\n"); if (params.embd_out == "json+" && n_prompts > 1) { fprintf(stdout, ",\n \"cosineSimilarity\": [\n"); - for (int i = 0;;) { // at least two iteration (n_prompts > 1) + for (int i = 0;;) { // at least two iteration (n_embd_count > 1) fprintf(stdout, " ["); - for (int j = 0;;) { // at least two iteration (n_prompts > 1) + for (int j = 0;;) { // at least two iteration (n_embd_count > 1) float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd); fprintf(stdout, "%6.2f", sim); j++; - if (j < n_prompts) fprintf(stdout, ", "); else break; + if (j < n_embd_count) fprintf(stdout, ", "); else break; } fprintf(stdout, " ]"); i++; - if (i < n_prompts) fprintf(stdout, ",\n"); else break; + if (i < n_embd_count) fprintf(stdout, ",\n"); else break; } fprintf(stdout, "\n ]"); } diff --git a/examples/eval-callback/eval-callback.cpp b/examples/eval-callback/eval-callback.cpp index 64cd338c2..5e89988e2 100644 --- a/examples/eval-callback/eval-callback.cpp +++ b/examples/eval-callback/eval-callback.cpp @@ -62,7 +62,7 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne } else if (type == GGML_TYPE_I8) { v = (float) *(int8_t *) &data[i]; } else { - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } printf("%12.4f", v); sum += v; @@ -99,7 +99,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) { char src1_str[128] = {0}; if (src1) { - sprintf(src1_str, "%s{%s}", src1->name, ggml_ne_string(src1).c_str()); + snprintf(src1_str, sizeof(src1_str), "%s{%s}", src1->name, ggml_ne_string(src1).c_str()); } printf("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, @@ -127,7 +127,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) { } static bool run(llama_context * ctx, const gpt_params & params) { - const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); + const bool add_bos = llama_add_bos_token(llama_get_model(ctx)); std::vector tokens = ::llama_tokenize(ctx, params.prompt, add_bos); @@ -163,9 +163,10 @@ int main(int argc, char ** argv) { params.warmup = false; // init - llama_model * model; - llama_context * ctx; - std::tie(model, ctx) = llama_init_from_gpt_params(params); + llama_init_result llama_init = llama_init_from_gpt_params(params); + + llama_model * model = llama_init.model; + llama_context * ctx = llama_init.context; if (model == nullptr || ctx == nullptr) { fprintf(stderr, "%s : failed to init\n", __func__); return 1; diff --git a/examples/export-lora/README.md b/examples/export-lora/README.md index 1fb17feec..7dce99c9a 100644 --- a/examples/export-lora/README.md +++ b/examples/export-lora/README.md @@ -6,21 +6,28 @@ Apply LORA adapters to base model and export the resulting model. usage: llama-export-lora [options] options: - -h, --help show this help message and exit - -m FNAME, --model-base FNAME model path from which to load base model (default '') - -o FNAME, --model-out FNAME path to save exported model (default '') - -l FNAME, --lora FNAME apply LoRA adapter - -s FNAME S, --lora-scaled FNAME S apply LoRA adapter with user defined scaling S - -t N, --threads N number of threads to use during computation (default: 4) + -m, --model model path from which to load base model (default '') + --lora FNAME path to LoRA adapter (can be repeated to use multiple adapters) + --lora-scaled FNAME S path to LoRA adapter with user defined scaling S (can be repeated to use multiple adapters) + -t, --threads N number of threads to use during computation (default: 4) + -o, --output FNAME output file (default: 'ggml-lora-merged-f16.gguf') ``` For example: ```bash ./bin/llama-export-lora \ - -m open-llama-3b-v2-q8_0.gguf \ - -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \ - -l lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.bin + -m open-llama-3b-v2.gguf \ + -o open-llama-3b-v2-english2tokipona-chat.gguf \ + --lora lora-open-llama-3b-v2-english2tokipona-chat-LATEST.gguf ``` -Multiple LORA adapters can be applied by passing multiple `-l FN` or `-s FN S` command line parameters. +Multiple LORA adapters can be applied by passing multiple `--lora FNAME` or `--lora-scaled FNAME S` command line parameters: + +```bash +./bin/llama-export-lora \ + -m your_base_model.gguf \ + -o your_merged_model.gguf \ + --lora-scaled lora_task_A.gguf 0.5 \ + --lora-scaled lora_task_B.gguf 0.5 +``` diff --git a/examples/export-lora/export-lora.cpp b/examples/export-lora/export-lora.cpp index 08413f57e..8df457e21 100644 --- a/examples/export-lora/export-lora.cpp +++ b/examples/export-lora/export-lora.cpp @@ -1,462 +1,423 @@ - #include "common.h" #include "ggml.h" #include "ggml-alloc.h" +#include #include #include #include +#include -struct lora_info { - std::string filename; +static bool g_verbose = false; + +struct tensor_transformation { + struct ggml_tensor * in; + struct ggml_tensor * out; + bool is_copy; +}; + +static std::string get_kv_str(struct gguf_context * ctx_gguf, const std::string & key){ + int id = gguf_find_key(ctx_gguf, key.c_str()); + return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id)); +} + +static float get_kv_f32(struct gguf_context * ctx_gguf, const std::string & key) { + int id = gguf_find_key(ctx_gguf, key.c_str()); + return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf, id); +} + +static void zeros(std::ofstream & file, size_t n) { + char zero = 0; + for (size_t i = 0; i < n; ++i) { + file.write(&zero, 1); + } +} + +static std::string ggml_ne_string(const ggml_tensor * t) { + std::string str; + for (int i = 0; i < GGML_MAX_DIMS; ++i) { + str += std::to_string(t->ne[i]); + if (i + 1 < GGML_MAX_DIMS) { + str += ", "; + } + } + return str; +} + +static struct gguf_context * load_gguf(std::string & fname, struct ggml_context ** ctx_ggml) { + struct gguf_init_params params = { + /*.no_alloc = */ true, + /*.ctx = */ ctx_ggml, + }; + struct gguf_context * ctx_gguf = gguf_init_from_file(fname.c_str(), params); + if (!ctx_gguf) { + throw std::runtime_error("failed to load input GGUF from " + fname); + } + return ctx_gguf; +} + +struct file_input { + struct ggml_context * ctx_meta = nullptr; + struct gguf_context * ctx_gguf = nullptr; + std::ifstream f_in; + std::map tensors; + float alpha; float scale; + + file_input(std::string & fname, float scale): f_in(fname, std::ios::binary), scale(scale) { + if (!f_in.is_open()) { + throw std::runtime_error("failed to open input gguf from " + fname); + } + + ctx_gguf = load_gguf(fname, &ctx_meta); + alpha = get_kv_f32(ctx_gguf, "adapter.lora.alpha"); + printf("%s: loaded gguf from %s\n", __func__, fname.c_str()); + + for (ggml_tensor * cur = ggml_get_first_tensor(ctx_meta); cur; cur = ggml_get_next_tensor(ctx_meta, cur)) { + std::string name(cur->name); + tensors[name] = cur; + if (g_verbose) { + printf("%s: %s\n", __func__, cur->name); + } + } + } + + ggml_tensor * get_tensor(std::string name) { + if (tensors.find(name) == tensors.end()) { + return nullptr; + } + return tensors[name]; + } + + void read_tensor_data(std::string name, std::vector & buf) { + if (tensors.find(name) == tensors.end()) { + throw std::runtime_error("cannot find tensor with name: " + name); + } + auto len = ggml_nbytes(tensors[name]); + if (buf.size() < len) { + buf.resize(len); + } + auto i_tensor_in = gguf_find_tensor(ctx_gguf, name.c_str()); // idx of tensor in the input file + auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor_in); + f_in.seekg(offset); + f_in.read((char* )buf.data(), len); + } + + ~file_input() { + gguf_free(ctx_gguf); + ggml_free(ctx_meta); + } }; -struct export_lora_params { - std::string fn_model_base; - std::string fn_model_out; - std::vector lora; +struct lora_merge_ctx { + // input base model + adapters + file_input base_model; + std::vector> adapters; + + // for computing merged tensor int n_threads; -}; + ggml_backend_t backend = nullptr; + ggml_gallocr_t allocr = nullptr; + std::vector read_buf; -struct lora_data { - struct lora_info info; - std::vector data; - struct ggml_context * ctx; + // output file + struct gguf_context * ctx_out; + struct ggml_context * ctx_out_ggml; + std::ofstream fout; - uint32_t lora_r; - uint32_t lora_alpha; -}; + lora_merge_ctx( + std::string & base_fname, + std::vector & lora_files, + std::string & outfile, + int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) { + fout.exceptions(std::ofstream::failbit); // fail fast on write errors -struct llama_file { - // use FILE * so we don't have to re-open the file to mmap - FILE * fp; - size_t size; + if (gguf_find_key(base_model.ctx_gguf, LLM_KV_SPLIT_COUNT) >= 0) { + throw std::runtime_error("split model is not yet supported"); + } - llama_file(const char * fname, const char * mode) { - fp = std::fopen(fname, mode); - if (fp == NULL) { - size = 0; + for (auto & lora_inp : lora_files) { + auto fname = lora_inp.path; + auto scale = lora_inp.scale; + std::unique_ptr adapter(new file_input(fname, scale)); + check_metadata_lora(adapter.get()); + adapters.push_back(std::move(adapter)); + } + + ctx_out = gguf_init_empty(); + struct ggml_init_params params = { + /*.mem_size =*/ gguf_get_n_tensors(base_model.ctx_gguf)*ggml_tensor_overhead(), + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, + }; + ctx_out_ggml = ggml_init(params); + backend = ggml_backend_cpu_init(); + allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend)); + } + + void check_metadata_lora(file_input * adapter) { + auto general_type = get_kv_str(adapter->ctx_gguf, "general.type"); + if (general_type != "adapter") { + throw std::runtime_error("expect general.type to be 'adapter', but got: " + general_type); + } + + auto adapter_type = get_kv_str(adapter->ctx_gguf, "adapter.type"); + if (adapter_type != "lora") { + throw std::runtime_error("expect adapter.type to be 'lora', but got: " + adapter_type); + } + + auto general_arch_base = get_kv_str(base_model.ctx_gguf, "general.architecture"); + auto general_arch_lora = get_kv_str(adapter->ctx_gguf, "general.architecture"); + if (general_arch_base != general_arch_lora) { + throw std::runtime_error("model arch and LoRA arch mismatch"); + } + } + + ggml_type get_out_tensor_type(struct ggml_tensor * t) { + if (t->type == GGML_TYPE_F32) { + return GGML_TYPE_F32; } else { - seek(0, SEEK_END); - size = tell(); - seek(0, SEEK_SET); + return GGML_TYPE_F16; } } - size_t tell() const { -#ifdef _WIN32 - __int64 ret = _ftelli64(fp); -#else - long ret = std::ftell(fp); -#endif - GGML_ASSERT(ret != -1); // this really shouldn't fail - return (size_t) ret; - } + void run_merge() { + // prepare metadata + gguf_set_kv(ctx_out, base_model.ctx_gguf); + // output is forced to f16 for now + gguf_set_val_u32(ctx_out, "general.file_type", LLAMA_FTYPE_MOSTLY_F16); - void seek(size_t offset, int whence) { -#ifdef _WIN32 - int ret = _fseeki64(fp, (__int64) offset, whence); -#else - int ret = std::fseek(fp, (long) offset, whence); -#endif - GGML_ASSERT(ret == 0); // same - } - - void read_raw(void * ptr, size_t size) { - if (size == 0) { - return; + // check if all lora adapters have the same tensors + // TODO: remove this when we can support merging subset of adapters. Ref: https://github.com/ggerganov/llama.cpp/pull/8607#discussion_r1686027777 + static const char * err_no_subset_adapter = "Input adapters do not have the same list of tensors. This is not yet supported. Please merge the adapter one-by-one instead of merging all at once."; + if (adapters.size() > 1) { + for (size_t i = 1; i < adapters.size(); ++i) { + if (adapters[0]->tensors.size() != adapters[i]->tensors.size()) { + throw std::runtime_error(err_no_subset_adapter); + } + for (auto & it : adapters[i]->tensors) { + if (adapters[0]->get_tensor(it.first) == nullptr) { + throw std::runtime_error(err_no_subset_adapter); + } + } + } } - errno = 0; - std::size_t ret = std::fread(ptr, size, 1, fp); - if (ferror(fp)) { - die_fmt("read error: %s", strerror(errno)); + + // mapping base tensor to out tensor (same shape with base, but different type) + std::vector trans; + for (auto & it : base_model.tensors) { + bool t_a = true; + bool t_b = true; + for (auto & adapter : adapters) { + t_a &= nullptr != adapter->get_tensor(it.first + ".lora_a"); + t_b &= nullptr != adapter->get_tensor(it.first + ".lora_b"); + } + auto base_tensor = it.second; + if (!t_a && !t_b) { + // only copy + struct ggml_tensor * cpy_tensor = ggml_dup_tensor(ctx_out_ggml, base_tensor); + ggml_set_name(cpy_tensor, base_tensor->name); + trans.push_back({ + cpy_tensor, + cpy_tensor, + true, + }); + gguf_add_tensor(ctx_out, cpy_tensor); + } else if (t_a && t_b) { + // need merging + struct ggml_tensor * out_tensor = ggml_new_tensor( + ctx_out_ggml, get_out_tensor_type(base_tensor), GGML_MAX_DIMS, base_tensor->ne); + ggml_set_name(out_tensor, base_tensor->name); + trans.push_back({ + base_tensor, + out_tensor, + false, + }); + gguf_add_tensor(ctx_out, out_tensor); + } else { + throw std::runtime_error("tensor " + it.first + " missing either lora_a or lora_b"); + } } - if (ret != 1) { - die("unexpectedly reached end of file"); + + // placeholder for the meta data + { + size_t meta_size = gguf_get_meta_size(ctx_out); + zeros(fout, meta_size); } - } - std::uint32_t read_u32() { - std::uint32_t ret; - read_raw(&ret, sizeof(ret)); - return ret; - } - - std::string read_string(std::uint32_t len) { - std::vector chars(len); - read_raw(chars.data(), len); - return std::string(chars.data(), len); - } - - void write_raw(const void * ptr, size_t size) { - if (size == 0) { - return; + // process base model tensors + size_t n_merged = 0; + for (auto & it : trans) { + if (!it.is_copy) { + merge_tensor(it.in, it.out); + n_merged++; + } else { + copy_tensor(it.in); + } } - errno = 0; - size_t ret = std::fwrite(ptr, size, 1, fp); - if (ret != 1) { - die_fmt("write error: %s", strerror(errno)); + + // write output metadata + { + std::vector data(gguf_get_meta_size(ctx_out)); + gguf_get_meta_data(ctx_out, data.data()); + fout.seekp(0); + fout.write((const char *)data.data(), data.size()); } + + printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged); + printf("%s : wrote %ld tensors to output file\n", __func__, trans.size()); } - void write_u32(std::uint32_t val) { - write_raw(&val, sizeof(val)); + void copy_tensor(struct ggml_tensor * base) { + printf("%s : %s [%s]\n", __func__, base->name, ggml_ne_string(base).c_str()); + size_t len = ggml_nbytes(base); + base_model.read_tensor_data(base->name, read_buf); + fout.write((char* )read_buf.data(), len); + zeros(fout, GGML_PAD(len, GGUF_DEFAULT_ALIGNMENT) - len); } - bool eof() { - return tell() >= size; - } + void merge_tensor(struct ggml_tensor * base, struct ggml_tensor * out) { + std::string name_base(base->name); + std::string name_lora_a = name_base + ".lora_a"; + std::string name_lora_b = name_base + ".lora_b"; - ~llama_file() { - if (fp) { - std::fclose(fp); + printf("%s : %s [%s]\n", __func__, base->name, ggml_ne_string(base).c_str()); + + // context for input tensor + std::vector inp_a(adapters.size()); + std::vector inp_b(adapters.size()); + struct ggml_init_params params { + /*.mem_size =*/ ggml_tensor_overhead()*(2+adapters.size()*2), + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, + }; + struct ggml_context * ctx = ggml_init(params); + + // alloc tensors + struct ggml_tensor * inp_base = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, base->ne); + for (size_t i = 0; i < adapters.size(); ++i) { + auto t_a = adapters[i]->get_tensor(name_lora_a); + auto t_b = adapters[i]->get_tensor(name_lora_b); + // TODO: add support for quantized lora + if (ggml_is_quantized(t_a->type) || ggml_is_quantized(t_b->type)) { + throw std::runtime_error("quantized LoRA adapters is not supported, please retry with f16 or f32"); + } + inp_a[i] = ggml_dup_tensor(ctx, t_a); + inp_b[i] = ggml_dup_tensor(ctx, t_b); } + ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx, backend); + + // load base tensor to backend buffer + base_model.read_tensor_data(name_base, read_buf); + if (base->type != GGML_TYPE_F32) { + // optionally dequantize it + printf("%s : + dequantize base tensor from %s to F32\n", __func__, ggml_type_name(base->type)); + auto nels = ggml_nelements(inp_base); + ggml_type_traits_t qtype = ggml_internal_get_type_traits(base->type); + std::vector dequant_buf(nels * sizeof(float)); + qtype.to_float(read_buf.data(), (float *)dequant_buf.data(), nels); + ggml_backend_tensor_set(inp_base, dequant_buf.data(), 0, dequant_buf.size()); + } else { + ggml_backend_tensor_set(inp_base, read_buf.data(), 0, ggml_nbytes(inp_base)); + } + + // load lora tensors to backend buffer + for (size_t i = 0; i < adapters.size(); ++i) { + adapters[i]->read_tensor_data(name_lora_a, read_buf); + ggml_backend_tensor_set(inp_a[i], read_buf.data(), 0, ggml_nbytes(inp_a[i])); + adapters[i]->read_tensor_data(name_lora_b, read_buf); + ggml_backend_tensor_set(inp_b[i], read_buf.data(), 0, ggml_nbytes(inp_b[i])); + } + + // build graph + struct ggml_cgraph * gf; + { + static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead(); + static std::vector buf(buf_size); + struct ggml_init_params params0 = { + /*.mem_size =*/ buf_size, + /*.mem_buffer =*/ buf.data(), + /*.no_alloc =*/ true, + }; + struct ggml_context * ctx0 = ggml_init(params0); + gf = ggml_new_graph(ctx0); + struct ggml_tensor * cur = inp_base; + for (size_t i = 0; i < adapters.size(); ++i) { + struct ggml_tensor * a_T = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_cast(ctx0, inp_a[i], GGML_TYPE_F32))); + struct ggml_tensor * delta = ggml_mul_mat(ctx0, a_T, ggml_cast(ctx0, inp_b[i], GGML_TYPE_F32)); + // scale + const float alpha = adapters[i]->alpha; + const float rank = (float) inp_b[i]->ne[0]; + const float scale = alpha ? adapters[i]->scale * alpha / rank : adapters[i]->scale; + delta = ggml_scale(ctx0, delta, scale); + cur = ggml_add(ctx0, delta, cur); + printf("%s : + merging from adapter[%ld] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type)); + printf("%s : input_scale=%f calculated_scale=%f rank=%d\n", __func__, adapters[i]->scale, scale, (int) inp_b[i]->ne[0]); + } + cur = ggml_cast(ctx0, cur, out->type); + printf("%s : + output type is %s\n", __func__, ggml_type_name(out->type)); + ggml_build_forward_expand(gf, cur); + ggml_free(ctx0); + } + + // compute + { + ggml_gallocr_alloc_graph(allocr, gf); + ggml_backend_cpu_set_n_threads(backend, n_threads); + ggml_backend_graph_compute(backend, gf); + } + + // write data to output file + { + auto result = gf->nodes[gf->n_nodes - 1]; + size_t len = ggml_nbytes(result); + if (read_buf.size() < len) { + read_buf.resize(len); + } + ggml_backend_tensor_get(result, read_buf.data(), 0, len); + fout.write((char* )read_buf.data(), len); + zeros(fout, GGML_PAD(len, GGUF_DEFAULT_ALIGNMENT) - len); + } + + ggml_free(ctx); + ggml_backend_buffer_free(buffer); + } + + ~lora_merge_ctx() { + ggml_gallocr_free(allocr); + ggml_backend_free(backend); + gguf_free(ctx_out); + ggml_free(ctx_out_ggml); } }; -static struct export_lora_params get_default_export_lora_params() { - struct export_lora_params result; - result.fn_model_base = ""; - result.fn_model_out = ""; - result.n_threads = GGML_DEFAULT_N_THREADS; - return result; -} +static void print_usage(int argc, char ** argv, const gpt_params & params) { + gpt_params_print_usage(argc, argv, params); -static void export_lora_print_usage(int /*argc*/, char ** argv, const struct export_lora_params * params) { - fprintf(stderr, "usage: %s [options]\n", argv[0]); - fprintf(stderr, "\n"); - fprintf(stderr, "options:\n"); - fprintf(stderr, " -h, --help show this help message and exit\n"); - fprintf(stderr, " -m FNAME, --model-base FNAME model path from which to load base model (default '%s')\n", params->fn_model_base.c_str()); - fprintf(stderr, " -o FNAME, --model-out FNAME path to save exported model (default '%s')\n", params->fn_model_out.c_str()); - fprintf(stderr, " -l FNAME, --lora FNAME apply LoRA adapter\n"); - fprintf(stderr, " -s FNAME S, --lora-scaled FNAME S apply LoRA adapter with user defined scaling S\n"); - fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params->n_threads); -} - -static bool export_lora_params_parse(int argc, char ** argv, struct export_lora_params * params) { - bool invalid_param = false; - std::string arg; - struct export_lora_params default_params = get_default_export_lora_params(); - const std::string arg_prefix = "--"; - - for (int i = 1; i < argc; i++) { - arg = argv[i]; - if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) { - std::replace(arg.begin(), arg.end(), '_', '-'); - } - - if (arg == "-m" || arg == "--model-base") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->fn_model_base = argv[i]; - } else if (arg == "-o" || arg == "--model-out") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->fn_model_out = argv[i]; - } else if (arg == "-l" || arg == "--lora") { - if (++i >= argc) { - invalid_param = true; - break; - } - struct lora_info lora; - lora.filename = argv[i]; - lora.scale = 1.0f; - params->lora.push_back(lora); - } else if (arg == "-s" || arg == "--lora-scaled") { - if (++i >= argc) { - invalid_param = true; - break; - } - struct lora_info lora; - lora.filename = argv[i]; - if (++i >= argc) { - invalid_param = true; - break; - } - lora.scale = std::stof(argv[i]); - params->lora.push_back(lora); - } else if (arg == "-t" || arg == "--threads") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_threads = std::stoi(argv[i]); - if (params->n_threads <= 0) { - params->n_threads = std::thread::hardware_concurrency(); - } - } else { - fprintf(stderr, "error: unknown argument: '%s'\n", arg.c_str()); - export_lora_print_usage(argc, argv, &default_params); - exit(1); - } - } - - if (params->fn_model_base == default_params.fn_model_base) { - fprintf(stderr, "error: please specify a filename for model-base.\n"); - export_lora_print_usage(argc, argv, &default_params); - exit(1); - } - if (params->fn_model_out == default_params.fn_model_out) { - fprintf(stderr, "error: please specify a filename for model-out.\n"); - export_lora_print_usage(argc, argv, &default_params); - exit(1); - } - if (invalid_param) { - fprintf(stderr, "error: invalid parameter for argument: '%s'\n", arg.c_str()); - export_lora_print_usage(argc, argv, &default_params); - exit(1); - } - return true; -} - -static void free_lora(struct lora_data * lora) { - if (lora->ctx != NULL) { - ggml_free(lora->ctx); - } - delete lora; -} - -static struct lora_data * load_lora(struct lora_info * info) { - struct lora_data * result = new struct lora_data; - result->info = *info; - result->ctx = NULL; - result->lora_r = 1; - result->lora_alpha = 1; - - struct llama_file file(info->filename.c_str(), "rb"); - if (file.fp == NULL) { - fprintf(stderr, "warning: Could not open lora adapter '%s'. Ignoring this adapter.\n", - info->filename.c_str()); - free_lora(result); - return NULL; - } - - struct ggml_init_params params_ggml; - params_ggml.mem_size = ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE; - params_ggml.mem_buffer = NULL; - params_ggml.no_alloc = true; - result->ctx = ggml_init(params_ggml); - - uint32_t magic = file.read_u32(); - if (magic != LLAMA_FILE_MAGIC_GGLA) { - die_fmt("unexpected lora header file magic in '%s'", info->filename.c_str()); - } - uint32_t version = file.read_u32(); - if (version != 1) { - die_fmt("unexpected lora file version '%u' in '%s'", (unsigned) version, info->filename.c_str()); - } - result->lora_r = file.read_u32(); - result->lora_alpha = file.read_u32(); - // read tensor infos from file - std::vector name_buf; - std::vector tensors; - std::vector tensors_offset; - size_t total_nbytes_pad = 0; - while(!file.eof()) { - int64_t ne[4] = {1,1,1,1}; - uint32_t n_dims = file.read_u32(); - uint32_t namelen = file.read_u32(); - uint32_t type = file.read_u32(); - for (uint32_t k = 0; k < n_dims; ++k) { - ne[k] = (int64_t)file.read_u32(); - } - name_buf.clear(); - name_buf.resize(namelen + 1, '\0'); - file.read_raw(name_buf.data(), namelen); - file.seek((0-file.tell()) & 31, SEEK_CUR); - size_t offset = file.tell(); - struct ggml_tensor * tensor = ggml_new_tensor(result->ctx, (enum ggml_type) type, n_dims, ne); - ggml_set_name(tensor, name_buf.data()); - size_t nbytes = ggml_nbytes(tensor); - size_t nbytes_pad = ggml_nbytes_pad(tensor); - total_nbytes_pad += nbytes_pad; - tensors.push_back(tensor); - tensors_offset.push_back(offset); - file.seek(nbytes, SEEK_CUR); - } - // read tensor data - result->data.resize(total_nbytes_pad); - size_t data_offset = 0; - for (size_t i = 0; i < tensors.size(); ++i) { - struct ggml_tensor * tensor = tensors[i]; - size_t offset = tensors_offset[i]; - size_t nbytes = ggml_nbytes(tensor); - size_t nbytes_pad = ggml_nbytes_pad(tensor); - file.seek(offset, SEEK_SET); - tensor->data = result->data.data() + data_offset; - file.read_raw(tensor->data, nbytes); - data_offset += nbytes_pad; - } - return result; -} - - -static struct ggml_cgraph * build_graph_lora( - struct ggml_context * ctx, - struct ggml_tensor * tensor, - struct ggml_tensor * lora_a, - struct ggml_tensor * lora_b, - float scaling -) { - struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b); - if (scaling != 1.0f) { - ab = ggml_scale(ctx, ab, scaling); - } - struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab); - - struct ggml_cgraph * gf = ggml_new_graph(ctx); - ggml_build_forward_expand (gf, res); - return gf; -} - -static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int n_threads) { - if (lora->ctx == NULL) { - return false; - } - std::string name = ggml_get_name(tensor); - std::string name_a = name + std::string(".loraA"); - std::string name_b = name + std::string(".loraB"); - struct ggml_tensor * lora_a = ggml_get_tensor(lora->ctx, name_a.c_str()); - struct ggml_tensor * lora_b = ggml_get_tensor(lora->ctx, name_b.c_str()); - if (lora_a == NULL || lora_b == NULL) { - return false; - } - - float scaling = lora->info.scale * (float)lora->lora_alpha / (float)lora->lora_r; - - struct ggml_init_params params; - params.mem_size = GGML_OBJECT_SIZE + ggml_graph_overhead() + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5; - params.mem_buffer = NULL; - params.no_alloc = true; - struct ggml_context * ctx = NULL; - struct ggml_gallocr * alloc = NULL; - struct ggml_cgraph * gf = NULL; - - ctx = ggml_init(params); - alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type()); - gf = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling); - - ggml_gallocr_alloc_graph(alloc, gf); - - struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads); - static std::vector data_work; - data_work.resize(cplan.work_size); - cplan.work_data = data_work.data(); - - ggml_graph_compute(gf, &cplan); - - ggml_gallocr_free(alloc); - ggml_free(ctx); - return true; -} - -static void export_lora(struct export_lora_params * params) { - // load all loras - std::vector loras; - for (size_t i = 0; i < params->lora.size(); ++i) { - struct lora_data * lora = load_lora(¶ms->lora[i]); - if (lora != NULL) { - loras.push_back(lora); - } - } - if (loras.size() == 0) { - fprintf(stderr, "warning: no lora adapters will be applied.\n"); - } - - // open input file - struct llama_file fin(params->fn_model_base.c_str(), "rb"); - if (!fin.fp) { - die_fmt("Could not open file '%s'\n", params->fn_model_base.c_str()); - } - - // open base model gguf, read tensors without their data - struct ggml_context * ctx_in; - struct gguf_init_params params_gguf; - params_gguf.no_alloc = true; - params_gguf.ctx = &ctx_in; - struct gguf_context * gguf_in = gguf_init_from_file(params->fn_model_base.c_str(), params_gguf); - - // create new gguf - struct gguf_context * gguf_out = gguf_init_empty(); - - // copy meta data from base model: kv and tensors - gguf_set_kv(gguf_out, gguf_in); - int n_tensors = gguf_get_n_tensors(gguf_in); - for (int i=0; i < n_tensors; ++i) { - const char * name = gguf_get_tensor_name(gguf_in, i); - struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name); - gguf_add_tensor(gguf_out, tensor); - } - - // create output file - struct llama_file fout(params->fn_model_out.c_str(), "wb"); - if (!fout.fp) { - die_fmt("Could not create file '%s'\n", params->fn_model_out.c_str()); - } - - // write gguf meta data - std::vector meta; - meta.resize(gguf_get_meta_size(gguf_out)); - gguf_get_meta_data(gguf_out, meta.data()); - fout.write_raw(meta.data(), meta.size()); - - std::vector data; - std::vector padding; - for (int i=0; i < n_tensors; ++i) { - const char * name = gguf_get_tensor_name(gguf_in, i); - struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name); - - // read tensor data - data.resize(ggml_nbytes(tensor)); - tensor->data = data.data(); - size_t offset = gguf_get_tensor_offset(gguf_in, i); - fin.seek(offset + meta.size(), SEEK_SET); - fin.read_raw(data.data(), data.size()); - - // apply all loras - for (size_t k = 0; k < loras.size(); ++k) { - apply_lora(tensor, loras[k], params->n_threads); - } - - // write tensor data + padding - padding.clear(); - padding.resize(GGML_PAD(data.size(), gguf_get_alignment(gguf_out)) - data.size(), 0); - - GGML_ASSERT(fout.tell() == offset + meta.size()); - // fout.seek(offset + meta.size(), SEEK_SET); - fout.write_raw(data.data(), data.size()); - fout.write_raw(padding.data(), padding.size()); - - if (i % 2 == 0) { - printf("."); - } - } + printf("\nexample usage:\n"); + printf("\n %s -m base-model.gguf --lora lora-file.gguf -o merged-model-f16.gguf\n", argv[0]); + printf("\nNOTE: output model is F16\n"); printf("\n"); - - // close gguf - gguf_free(gguf_out); - gguf_free(gguf_in); - - // free loras - for (size_t i = 0; i < loras.size(); ++i) { - free_lora(loras[i]); - } } int main(int argc, char ** argv) { - struct export_lora_params params = get_default_export_lora_params(); + gpt_params params; - if (!export_lora_params_parse(argc, argv, ¶ms)) { + if (!gpt_params_parse(argc, argv, params)) { + print_usage(argc, argv, params); return 1; } - export_lora(¶ms); + g_verbose = (params.verbosity == 1); + try { + lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.cpuparams.n_threads); + ctx.run_merge(); + } catch (const std::exception & err) { + fprintf(stderr, "%s\n", err.what()); + exit(EXIT_FAILURE); + } + + printf("done, output file is %s\n", params.lora_outfile.c_str()); return 0; } diff --git a/examples/finetune/CMakeLists.txt b/examples/finetune/CMakeLists.txt deleted file mode 100644 index 64afe6ddc..000000000 --- a/examples/finetune/CMakeLists.txt +++ /dev/null @@ -1,5 +0,0 @@ -set(TARGET llama-finetune) -add_executable(${TARGET} finetune.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/finetune/README.md b/examples/finetune/README.md deleted file mode 100644 index a6ae64983..000000000 --- a/examples/finetune/README.md +++ /dev/null @@ -1,90 +0,0 @@ -# finetune - -Basic usage instructions: - -```bash -# get training data -wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/shakespeare.txt - -# finetune LORA adapter -./bin/llama-finetune \ - --model-base open-llama-3b-v2-q8_0.gguf \ - --checkpoint-in chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf \ - --checkpoint-out chk-lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.gguf \ - --lora-out lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.bin \ - --train-data "shakespeare.txt" \ - --save-every 10 \ - --threads 6 --adam-iter 30 --batch 4 --ctx 64 \ - --use-checkpointing - -# predict -./bin/llama-cli -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin -``` - -**Only llama based models are supported!** The output files will be saved every N iterations (config with `--save-every N`). -The pattern 'ITERATION' in the output filenames will be replaced with the iteration number and with 'LATEST' for the latest output. -So in above example after 10 iterations these files will be written: -- chk-lora-open-llama-3b-v2-q8_0-shakespeare-10.gguf -- chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf -- lora-open-llama-3b-v2-q8_0-shakespeare-10.bin -- lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin - -After 10 more iterations: -- chk-lora-open-llama-3b-v2-q8_0-shakespeare-20.gguf -- chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf -- lora-open-llama-3b-v2-q8_0-shakespeare-20.bin -- lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin - -Checkpoint files (`--checkpoint-in FN`, `--checkpoint-out FN`) store the training process. When the input checkpoint file does not exist, it will begin finetuning a new randomly initialized adapter. - -llama.cpp compatible LORA adapters will be saved with filename specified by `--lora-out FN`. -These LORA adapters can then be used by `llama-cli` together with the base model, like in the 'predict' example command above. - -In `llama-cli` you can also load multiple LORA adapters, which will then be mixed together. - -For example if you have two LORA adapters `lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin` and `lora-open-llama-3b-v2-q8_0-bible-LATEST.bin`, you can mix them together like this: - -```bash -./bin/llama-cli -m open-llama-3b-v2-q8_0.gguf \ - --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin \ - --lora lora-open-llama-3b-v2-q8_0-bible-LATEST.bin -``` - -You can change how strong each LORA adapter is applied to the base model by using `--lora-scaled FN SCALE` instead of `--lora FN`. - -For example to apply 40% of the 'shakespeare' LORA adapter, 80% of the 'bible' LORA adapter and 100% of yet another one: - -```bash -./bin/llama-cli -m open-llama-3b-v2-q8_0.gguf \ - --lora-scaled lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin 0.4 \ - --lora-scaled lora-open-llama-3b-v2-q8_0-bible-LATEST.bin 0.8 \ - --lora lora-open-llama-3b-v2-q8_0-yet-another-one-LATEST.bin -``` - -The scale numbers don't need to add up to one, and you can also use numbers greater than 1 to further increase the influence of an adapter. But making the values too big will sometimes result in worse output. Play around to find good values. - -Gradient checkpointing reduces the memory requirements by ~50% but increases the runtime. -If you have enough RAM, you can make finetuning a bit faster by disabling checkpointing with `--no-checkpointing`. - -The default LORA rank can be specified with `--lora-r N`. -The LORA rank can be configured for each model tensor type separately with these command line options: - -```bash - --lora-r N LORA r: default rank. Also specifies resulting scaling together with lora-alpha. (default 4) - --rank-att-norm N LORA rank for attention norm tensor (default 1) - --rank-ffn-norm N LORA rank for feed-forward norm tensor (default 1) - --rank-out-norm N LORA rank for output norm tensor (default 1) - --rank-tok-embd N LORA rank for token embeddings tensor (default 4) - --rank-out N LORA rank for output tensor (default 4) - --rank-wq N LORA rank for wq tensor (default 4) - --rank-wk N LORA rank for wk tensor (default 4) - --rank-wv N LORA rank for wv tensor (default 4) - --rank-wo N LORA rank for wo tensor (default 4) - --rank-ffn_gate N LORA rank for ffn_gate tensor (default 4) - --rank-ffn_down N LORA rank for ffn_down tensor (default 4) - --rank-ffn_up N LORA rank for ffn_up tensor (default 4) -``` - -The LORA rank of 'norm' tensors should always be 1. - -To see all available options use `finetune --help`. diff --git a/examples/finetune/convert-finetune-checkpoint-to-gguf.py b/examples/finetune/convert-finetune-checkpoint-to-gguf.py deleted file mode 100644 index c89090918..000000000 --- a/examples/finetune/convert-finetune-checkpoint-to-gguf.py +++ /dev/null @@ -1,487 +0,0 @@ -#!/usr/bin/env python3 -# finetune checkpoint --> gguf conversion - -import argparse -import gguf -import struct -import numpy as np -from pathlib import Path - -# gguf constants -LLM_KV_OPTIMIZER_TYPE = "optimizer.type" -LLM_KV_OPTIMIZER_TYPE_ADAM = "adam" -LLM_KV_OPTIMIZER_TYPE_LBFGS = "lbfgs" -LLM_KV_OPTIMIZER_FILE_VERSION = "optimizer.file_version" -LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT = "optimizer.convergence_past_count" -LLM_KV_OPTIMIZER_PARAMETER_COUNT = "optimizer.parameter_count" -LLM_KV_OPTIMIZER_ITERATION_COUNT = "optimizer.iteration_count" -LLM_KV_OPTIMIZER_JUST_INITIALIZED = "optimizer.just_initialized" -LLM_KV_OPTIMIZER_ADAM_BEST_LOSS = "optimizer.adam.best_loss" -LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS = "optimizer.adam.previous_loss" -LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT = "optimizer.adam.no_improvement_count" -LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT = "optimizer.lbfgs.approx_hessian_count" -LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS = "optimizer.lbfgs.best_loss" -LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP = "optimizer.lbfgs.line_search_step" -LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J = "optimizer.lbfgs.line_search_j" -LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K = "optimizer.lbfgs.line_search_k" -LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END = "optimizer.lbfgs.line_search_end" -LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT = "optimizer.lbfgs.no_improvement_count" - -LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS = "optimizer.adam.first_moments" -LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS = "optimizer.adam.second_moments" -LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES = "optimizer.adam.past_loss_values" - -LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS = "optimizer.lbfgs.current_parameters" -LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS = "optimizer.lbfgs.previous_parameters" -LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS = "optimizer.lbfgs.current_gradients" -LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS = "optimizer.lbfgs.previous_gradients" -LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION = "optimizer.lbfgs.search_direction" -LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES = "optimizer.lbfgs.past_loss_values" -LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA = "optimizer.lbfgs.memory_alpha" -LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS = "optimizer.lbfgs.memory_ys" -LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S = "optimizer.lbfgs.memory_s" -LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y = "optimizer.lbfgs.memory_y" - -LLM_KV_TRAINING_TYPE_TRAIN_MODEL = "train_model" -LLM_KV_TRAINING_TYPE_FINETUNE_LORA = "finetune_lora" -LLM_KV_TRAINING_TYPE = "training.type" -LLM_KV_TRAINING_FILE_VERSION = "training.file_version" -LLM_KV_TRAINING_ITERATION_COUNT = "training.iteration_count" -LLM_KV_TRAINING_SAMPLE_COUNT = "training.sample_count" -LLM_KV_TRAINING_TOKEN_COUNT = "training.token_count" - -LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD = "training.lora.rank.token_embd" -LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM = "training.lora.rank.output_norm" -LLM_KV_TRAINING_LORA_RANK_OUTPUT = "training.lora.rank.output" -LLM_KV_TRAINING_LORA_RANK_ATTN_NORM = "training.lora.rank.attn_norm" -LLM_KV_TRAINING_LORA_RANK_ATTN_Q = "training.lora.rank.attn_q" -LLM_KV_TRAINING_LORA_RANK_ATTN_K = "training.lora.rank.attn_k" -LLM_KV_TRAINING_LORA_RANK_ATTN_V = "training.lora.rank.attn_v" -LLM_KV_TRAINING_LORA_RANK_ATTN_OUT = "training.lora.rank.attn_output" -LLM_KV_TRAINING_LORA_RANK_FFN_NORM = "training.lora.rank.ffn_norm" -LLM_KV_TRAINING_LORA_RANK_FFN_GATE = "training.lora.rank.ffn_gate" -LLM_KV_TRAINING_LORA_RANK_FFN_DOWN = "training.lora.rank.ffn_down" -LLM_KV_TRAINING_LORA_RANK_FFN_UP = "training.lora.rank.ffn_up" - -class Tensor: - def __init__(self, dtype='f', ne=None): - if ne is None: - ne = [] - self.dtype = dtype - self.ne = ne - self.nbytes = 0 - if self.dtype == 'f': - if len(self.ne) == 0: - self.nbytes = 0 - else: - self.nbytes = int(np.product(self.ne)) * 4 - else: - raise ValueError(f"Unhandled data type '{self.dtype}'") - - def load(self, data, offset): - nd = struct.unpack(' 0 else []) - - self.lbfgs_x = Tensor('f', [self.nx]) - self.lbfgs_xp = Tensor('f', [self.nx]) - self.lbfgs_g = Tensor('f', [self.nx]) - self.lbfgs_gp = Tensor('f', [self.nx]) - self.lbfgs_d = Tensor('f', [self.nx]) - self.lbfgs_pf = Tensor('f', [self.past] if self.past > 0 else []) - self.lbfgs_lmal = Tensor('f', [self.lbfgs_m]) - self.lbfgs_lmys = Tensor('f', [self.lbfgs_m]) - self.lbfgs_lms = Tensor('f', [self.nx, self.lbfgs_m]) - self.lbfgs_lmy = Tensor('f', [self.nx, self.lbfgs_m]) - - # forgot to save type in version 1: - # guess self.type from number of remaining bytes - size_type_0 = 12 + sum([t.max_storage_size() for t in - [self.adam_m, self.adam_v] - +([self.adam_pf] if (self.past > 0) else [])]) - size_type_1 = 24 + sum([t.max_storage_size() for t in - [self.lbfgs_x, self.lbfgs_xp, self.lbfgs_g, - self.lbfgs_gp, self.lbfgs_d, self.lbfgs_pf, - self.lbfgs_lmal, self.lbfgs_lmys, - self.lbfgs_lms, self.lbfgs_lmy] - +([self.lbfgs_pf] if (self.past > 0) else [])]) - # due to alignment padding the size might not by exact - # but the difference in size for both types is significant, - # so we can just use whichever is closest - remaining = len(data) - offset - if abs(remaining - size_type_0) < abs(remaining - size_type_1): - self.type = 0 - else: - self.type = 1 - - if self.type == 0: - offset = self.adam_m.load(data, offset) - offset = self.adam_v.load(data, offset) - offset = self.adam_pf.load(data,offset) - - self.adam_fx_best = struct.unpack(' 0: - self.adam_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES) - - elif self.type == 1: - gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS) - gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, self.lbfgs_m) - gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS, self.lbfgs_fx_best) - gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP, self.lbfgs_step) - gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J, self.lbfgs_j) - gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K, self.lbfgs_k) - gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END, self.lbfgs_end) - gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT, self.lbfgs_n_no_improvement) - - self.lbfgs_x.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS) - self.lbfgs_xp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS) - self.lbfgs_g.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS) - self.lbfgs_gp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS) - self.lbfgs_d.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION) - if self.past > 0: - self.lbfgs_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES) - self.lbfgs_lmal.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA) - self.lbfgs_lmys.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS) - self.lbfgs_lms.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S) - self.lbfgs_lmy.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y) - else: - raise ValueError('Unknown optimizer type') - -class LoraParams: - def __init__(self): - pass - - def load(self, data, offset): - self.n_rank_attention_norm = struct.unpack(' -#include -#include -#include -#include - -#if defined(_MSC_VER) -#pragma warning(disable: 4244 4267) // possible loss of data -#endif - -struct my_llama_hparams { - uint32_t n_vocab = 32000; - uint32_t n_ctx = 512; - uint32_t n_embd = 4096; - uint32_t n_ff = 11008; - uint32_t n_head = 32; - uint32_t n_head_kv = 32; - uint32_t n_layer = 32; - - // float f_norm_eps = 1e-5f; // falcon - float f_norm_rms_eps = 1e-5f; // llama - - float rope_freq_base = 10000.0f; - float rope_freq_scale = 1.0f; - - uint32_t n_gqa() const { - return n_head/n_head_kv; - } - - uint32_t n_embd_head() const { - return n_embd/n_head; - } - - uint32_t n_embd_gqa() const { - return n_embd/n_gqa(); - } - - bool operator!=(const my_llama_hparams& other) const { - return memcmp(this, &other, sizeof(other)); - } -}; - -struct my_llama_layer { - // normalization - struct ggml_tensor * attention_norm; - - // attention - struct ggml_tensor * wq; - struct ggml_tensor * wk; - struct ggml_tensor * wv; - struct ggml_tensor * wo; - - // normalization - struct ggml_tensor * ffn_norm; - - // ff - struct ggml_tensor * ffn_gate; // w1 - struct ggml_tensor * ffn_down; // w2 - struct ggml_tensor * ffn_up; // w3 -}; - -struct my_llama_model { - struct my_llama_hparams hparams; - - struct ggml_tensor * tok_embeddings; - - struct ggml_tensor * norm; - struct ggml_tensor * output; - - std::vector layers; -}; - -struct my_llama_lora_hparams { - uint32_t lora_r = 1; - uint32_t lora_alpha = 1; - uint32_t n_rank_attention_norm = 1; - uint32_t n_rank_wq = 4; - uint32_t n_rank_wk = 4; - uint32_t n_rank_wv = 4; - uint32_t n_rank_wo = 4; - uint32_t n_rank_ffn_norm = 1; - uint32_t n_rank_ffn_gate = 4; - uint32_t n_rank_ffn_down = 4; - uint32_t n_rank_ffn_up = 4; - uint32_t n_rank_tok_embeddings = 4; - uint32_t n_rank_norm = 1; - uint32_t n_rank_output = 4; - - bool operator!=(const my_llama_lora_hparams& other) const { - return memcmp(this, &other, sizeof(other)); - } -}; - -struct my_llama_lora_layer { - // normalization - struct ggml_tensor * attention_norm_a; - struct ggml_tensor * attention_norm_b; - - // attention - struct ggml_tensor * wq_a; - struct ggml_tensor * wq_b; - struct ggml_tensor * wk_a; - struct ggml_tensor * wk_b; - struct ggml_tensor * wv_a; - struct ggml_tensor * wv_b; - struct ggml_tensor * wo_a; - struct ggml_tensor * wo_b; - - // normalization - struct ggml_tensor * ffn_norm_a; - struct ggml_tensor * ffn_norm_b; - - // ff - struct ggml_tensor * ffn_gate_a; - struct ggml_tensor * ffn_gate_b; - struct ggml_tensor * ffn_down_a; - struct ggml_tensor * ffn_down_b; - struct ggml_tensor * ffn_up_a; - struct ggml_tensor * ffn_up_b; -}; - -struct my_llama_lora { - struct ggml_context * ctx = NULL; - ggml_backend_buffer_t data; - - my_llama_lora_hparams hparams; - - struct ggml_tensor * tok_embeddings_a; - struct ggml_tensor * tok_embeddings_b; - - struct ggml_tensor * norm_a; - struct ggml_tensor * norm_b; - struct ggml_tensor * output_a; - struct ggml_tensor * output_b; - - std::vector layers; -}; - -// gguf constants -static const char * LLM_KV_TRAINING_TYPE_FINETUNE_LORA = "finetune_lora"; -static const char * LLM_KV_TRAINING_TYPE = "training.type"; - -static const char * LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD = "training.lora.rank.token_embd"; -static const char * LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM = "training.lora.rank.output_norm"; -static const char * LLM_KV_TRAINING_LORA_RANK_OUTPUT = "training.lora.rank.output"; -static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_NORM = "training.lora.rank.attn_norm"; -static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_Q = "training.lora.rank.attn_q"; -static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_K = "training.lora.rank.attn_k"; -static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_V = "training.lora.rank.attn_v"; -static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_OUT = "training.lora.rank.attn_output"; -static const char * LLM_KV_TRAINING_LORA_RANK_FFN_NORM = "training.lora.rank.ffn_norm"; -static const char * LLM_KV_TRAINING_LORA_RANK_FFN_GATE = "training.lora.rank.ffn_gate"; -static const char * LLM_KV_TRAINING_LORA_RANK_FFN_DOWN = "training.lora.rank.ffn_down"; -static const char * LLM_KV_TRAINING_LORA_RANK_FFN_UP = "training.lora.rank.ffn_up"; - -// gguf constants (sync with gguf.py) - -static const char * LLM_KV_GENERAL_ARCHITECTURE = "general.architecture"; -static const char * LLM_KV_GENERAL_FILE_TYPE = "general.file_type"; - -static const char * LLM_KV_CONTEXT_LENGTH = "%s.context_length"; -static const char * LLM_KV_EMBEDDING_LENGTH = "%s.embedding_length"; -static const char * LLM_KV_BLOCK_COUNT = "%s.block_count"; -static const char * LLM_KV_FEED_FORWARD_LENGTH = "%s.feed_forward_length"; -static const char * LLM_KV_ATTENTION_HEAD_COUNT = "%s.attention.head_count"; -static const char * LLM_KV_ATTENTION_HEAD_COUNT_KV = "%s.attention.head_count_kv"; -static const char * LLM_KV_ATTENTION_LAYERNORM_RMS_EPS = "%s.attention.layer_norm_rms_epsilon"; -static const char * LLM_KV_ROPE_DIMENSION_COUNT = "%s.rope.dimension_count"; -static const char * LLM_KV_ROPE_FREQ_BASE = "%s.rope.freq_base"; // TODO load in llama.cpp -static const char * LLM_KV_ROPE_SCALE_LINEAR = "%s.rope.scale_linear"; - -static const char * LLM_TENSOR_TOKEN_EMBD = "token_embd"; -static const char * LLM_TENSOR_OUTPUT_NORM = "output_norm"; -static const char * LLM_TENSOR_OUTPUT = "output"; -static const char * LLM_TENSOR_ATTN_NORM = "blk.%d.attn_norm"; -static const char * LLM_TENSOR_ATTN_Q = "blk.%d.attn_q"; -static const char * LLM_TENSOR_ATTN_K = "blk.%d.attn_k"; -static const char * LLM_TENSOR_ATTN_V = "blk.%d.attn_v"; -static const char * LLM_TENSOR_ATTN_OUT = "blk.%d.attn_output"; -static const char * LLM_TENSOR_FFN_NORM = "blk.%d.ffn_norm"; -static const char * LLM_TENSOR_FFN_GATE = "blk.%d.ffn_gate"; -static const char * LLM_TENSOR_FFN_DOWN = "blk.%d.ffn_down"; -static const char * LLM_TENSOR_FFN_UP = "blk.%d.ffn_up"; - -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_ff : %u\n", __func__, params->n_ff); - printf("%s: n_head : %u\n", __func__, params->n_head); - printf("%s: n_head_kv : %u\n", __func__, params->n_head_kv); - printf("%s: n_layer : %u\n", __func__, params->n_layer); - printf("%s: norm_rms_eps : %f\n", __func__, params->f_norm_rms_eps); - printf("%s: rope_freq_base : %f\n", __func__, params->rope_freq_base); - printf("%s: rope_freq_scale : %f\n", __func__, params->rope_freq_scale); -} - -static void print_lora_params(struct my_llama_lora_hparams * params) { - printf("%s: n_rank_attention_norm : %u\n", __func__, params->n_rank_attention_norm); - printf("%s: n_rank_wq : %u\n", __func__, params->n_rank_wq); - printf("%s: n_rank_wk : %u\n", __func__, params->n_rank_wk); - printf("%s: n_rank_wv : %u\n", __func__, params->n_rank_wv); - printf("%s: n_rank_wo : %u\n", __func__, params->n_rank_wo); - printf("%s: n_rank_ffn_norm : %u\n", __func__, params->n_rank_ffn_norm); - printf("%s: n_rank_ffn_gate : %u\n", __func__, params->n_rank_ffn_gate); - printf("%s: n_rank_ffn_down : %u\n", __func__, params->n_rank_ffn_down); - printf("%s: n_rank_ffn_up : %u\n", __func__, params->n_rank_ffn_up); - printf("%s: n_rank_tok_embeddings : %u\n", __func__, params->n_rank_tok_embeddings); - printf("%s: n_rank_norm : %u\n", __func__, params->n_rank_norm); - printf("%s: n_rank_output : %u\n", __func__, params->n_rank_output); -} - -#define GGUF_GET_KEY(ctx, dst, func, type, req, key) \ -{ \ - const std::string skey(key); \ - const int kid = gguf_find_key(ctx, skey.c_str()); \ - if (kid >= 0) { \ - enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \ - if (ktype != (type)) { \ - die_fmt("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype)); \ - } \ - (dst) = func(ctx, kid); \ - } else if (req) { \ - die_fmt("key not found in model: %s", skey.c_str()); \ - } \ -} - -static void load_model_hparams_gguf(struct gguf_context * ctx, struct my_llama_hparams * hparams, const char * expected_arch) { - std::string arch; - - GGUF_GET_KEY(ctx, arch, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_GENERAL_ARCHITECTURE); - if (expected_arch != NULL) { - if (arch != expected_arch) { - printf("%s: arch=%s expected_arch=%s\n", __func__, arch.c_str(), expected_arch); - } - GGML_ASSERT(arch == expected_arch); - } - - std::vector keybuf; - keybuf.resize(512); - auto kv = [&arch, &keybuf](const char * key) -> const char * { - snprintf(keybuf.data(), keybuf.size(), key, arch.c_str()); - return keybuf.data(); - }; - - GGUF_GET_KEY(ctx, hparams->n_embd, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_EMBEDDING_LENGTH)); - GGUF_GET_KEY(ctx, hparams->n_ctx, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_CONTEXT_LENGTH)); - GGUF_GET_KEY(ctx, hparams->n_ff, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_FEED_FORWARD_LENGTH)); - GGUF_GET_KEY(ctx, hparams->n_head, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_ATTENTION_HEAD_COUNT)); - GGUF_GET_KEY(ctx, hparams->n_layer, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_BLOCK_COUNT)); - - // n_head_kv is optional, default to n_head - hparams->n_head_kv = hparams->n_head; - GGUF_GET_KEY(ctx, hparams->n_head_kv, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV)); - - float rope_freq_scale = 1.0f; - GGUF_GET_KEY(ctx, hparams->f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS)); - GGUF_GET_KEY(ctx, hparams->rope_freq_base, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE)); - GGUF_GET_KEY(ctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR)); - if (rope_freq_scale != 1.0f) { - hparams->rope_freq_scale = 1.0f / rope_freq_scale; - } -} - -static void init_model(struct llama_model * input, struct my_llama_model * model, const char * fn_model, uint32_t n_ctx) { - auto & hparams = model->hparams; - - std::vector tn_buf; - tn_buf.resize(GGML_MAX_NAME); - auto tn = [&tn_buf](const char * key) -> const char * { - snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", key); - return tn_buf.data(); - }; - auto tni = [&tn_buf](const char * key, int bid) -> const char * { - snprintf(tn_buf.data(), tn_buf.size(), key, bid); - std::string s = tn_buf.data(); - snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", s.c_str()); - return tn_buf.data(); - }; - - - // get parameters directly from gguf file - { - struct gguf_init_params params = { - /*.no_alloc = */ false, - /*.ctx = */ NULL, - }; - struct gguf_context * mctx = gguf_init_from_file(fn_model, params); - - load_model_hparams_gguf(mctx, &hparams, "llama"); - - gguf_free(mctx); - } - hparams.n_vocab = llama_n_vocab(input); - hparams.n_ctx = n_ctx; - - // get tensors from llama_model (possibly mmapped) - model->tok_embeddings = llama_get_model_tensor(input, tn(LLM_TENSOR_TOKEN_EMBD)); - model->norm = llama_get_model_tensor(input, tn(LLM_TENSOR_OUTPUT_NORM)); - model->output = llama_get_model_tensor(input, tn(LLM_TENSOR_OUTPUT)); - - assert_shape_2d(model->tok_embeddings, hparams.n_embd, hparams.n_vocab); - assert_shape_1d(model->norm, hparams.n_embd); - assert_shape_2d(model->output, hparams.n_embd, hparams.n_vocab); - - model->layers.resize(hparams.n_layer); - for (uint32_t i = 0; i < hparams.n_layer; ++i) { - auto & layer = model->layers[i]; - - layer.attention_norm = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_NORM, i)); - layer.wq = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_Q, i)); - layer.wk = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_K, i)); - layer.wv = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_V, i)); - layer.wo = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_OUT, i)); - layer.ffn_norm = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_NORM, i)); - layer.ffn_gate = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i)); - layer.ffn_down = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i)); - layer.ffn_up = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i)); - - assert_shape_1d(layer.attention_norm, hparams.n_embd); - assert_shape_2d(layer.wq, hparams.n_embd, hparams.n_embd); - assert_shape_2d(layer.wk, hparams.n_embd, hparams.n_embd_gqa()); - assert_shape_2d(layer.wv, hparams.n_embd, hparams.n_embd_gqa()); - assert_shape_2d(layer.wo, hparams.n_embd, hparams.n_embd); - assert_shape_1d(layer.ffn_norm, hparams.n_embd); - assert_shape_2d(layer.ffn_gate, hparams.n_embd, hparams.n_ff); - assert_shape_2d(layer.ffn_down, hparams.n_ff, hparams.n_embd); - assert_shape_2d(layer.ffn_up, hparams.n_embd, hparams.n_ff); - } -} - -static void set_param_lora(struct my_llama_lora * lora) { - const uint32_t n_layer = lora->layers.size(); - - struct ggml_context* ctx = lora->ctx; - - ggml_set_param(ctx, lora->tok_embeddings_a); - ggml_set_param(ctx, lora->tok_embeddings_b); - ggml_set_param(ctx, lora->norm_a); - ggml_set_param(ctx, lora->norm_b); - ggml_set_param(ctx, lora->output_a); - ggml_set_param(ctx, lora->output_b); - - for (uint32_t i = 0; i < n_layer; ++i) { - auto & layer = lora->layers[i]; - - ggml_set_param(ctx, layer.attention_norm_a); - ggml_set_param(ctx, layer.attention_norm_b); - ggml_set_param(ctx, layer.wq_a); - ggml_set_param(ctx, layer.wq_b); - ggml_set_param(ctx, layer.wk_a); - ggml_set_param(ctx, layer.wk_b); - ggml_set_param(ctx, layer.wv_a); - ggml_set_param(ctx, layer.wv_b); - ggml_set_param(ctx, layer.wo_a); - ggml_set_param(ctx, layer.wo_b); - ggml_set_param(ctx, layer.ffn_norm_a); - ggml_set_param(ctx, layer.ffn_norm_b); - ggml_set_param(ctx, layer.ffn_gate_a); - ggml_set_param(ctx, layer.ffn_gate_b); - ggml_set_param(ctx, layer.ffn_down_a); - ggml_set_param(ctx, layer.ffn_down_b); - ggml_set_param(ctx, layer.ffn_up_a); - ggml_set_param(ctx, layer.ffn_up_b); - } -} - -static void init_lora(const struct my_llama_model * model, struct my_llama_lora * lora) { - const auto & lparams = lora->hparams; - - const uint32_t n_embd = model->hparams.n_embd; - const uint32_t n_embd_gqa = model->hparams.n_embd_gqa(); - const uint32_t n_layer = model->hparams.n_layer; - const uint32_t n_vocab = model->hparams.n_vocab; - const uint32_t n_ff = model->hparams.n_ff; - - std::vector tn_buf; - tn_buf.resize(GGML_MAX_NAME); - auto tn = [&tn_buf](const char * key, const char * suffix) -> const char * { - snprintf(tn_buf.data(), tn_buf.size(), "%s%s", key, suffix); - return tn_buf.data(); - }; - auto tni = [&tn_buf](const char * key, const char * suffix, int bid) -> const char * { - snprintf(tn_buf.data(), tn_buf.size(), key, bid); - std::string s = tn_buf.data(); - snprintf(tn_buf.data(), tn_buf.size(), "%s%s", s.c_str(), suffix); - return tn_buf.data(); - }; - - // context for lora tensors without their data - struct ggml_init_params ctx_lora_params; - ctx_lora_params.mem_size = ggml_tensor_overhead()*2*(6 + n_layer*18); - ctx_lora_params.mem_buffer = NULL; - ctx_lora_params.no_alloc = true; - - struct ggml_context * ctx = ggml_init(ctx_lora_params); - lora->ctx = ctx; - - lora->tok_embeddings_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_tok_embeddings, n_embd); - lora->tok_embeddings_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_tok_embeddings, n_vocab); - lora->norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_norm, n_embd); - lora->norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_norm, 1); - lora->output_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_output, n_embd); - lora->output_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_output, n_vocab); - - ggml_set_name(lora->tok_embeddings_a, tn(LLM_TENSOR_TOKEN_EMBD, ".weight.lora_a")); - ggml_set_name(lora->tok_embeddings_b, tn(LLM_TENSOR_TOKEN_EMBD, ".weight.lora_b")); - ggml_set_name(lora->norm_a, tn(LLM_TENSOR_OUTPUT_NORM, ".weight.lora_a")); - ggml_set_name(lora->norm_b, tn(LLM_TENSOR_OUTPUT_NORM, ".weight.lora_b")); - ggml_set_name(lora->output_a, tn(LLM_TENSOR_OUTPUT, ".weight.lora_a")); - ggml_set_name(lora->output_b, tn(LLM_TENSOR_OUTPUT, ".weight.lora_b")); - - lora->layers.resize(n_layer); - for (uint32_t i = 0; i < n_layer; ++i) { - auto & layer = lora->layers[i]; - - layer.attention_norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_attention_norm, n_embd); - layer.attention_norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_attention_norm, 1); - - layer.wq_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wq, n_embd); - layer.wq_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wq, n_embd); - layer.wk_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wk, n_embd); - layer.wk_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wk, n_embd_gqa); - layer.wv_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wv, n_embd); - layer.wv_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wv, n_embd_gqa); - layer.wo_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wo, n_embd); - layer.wo_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wo, n_embd); - - layer.ffn_norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, n_embd); - layer.ffn_norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, 1); - - layer.ffn_gate_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_embd); - layer.ffn_gate_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_ff); - layer.ffn_down_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_ff); - layer.ffn_down_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_embd); - layer.ffn_up_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up, n_embd); - layer.ffn_up_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up, n_ff); - - ggml_set_name(layer.attention_norm_a, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_a", i)); - ggml_set_name(layer.attention_norm_b, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_b", i)); - ggml_set_name(layer.wq_a, tni(LLM_TENSOR_ATTN_Q, ".weight.lora_a", i)); - ggml_set_name(layer.wq_b, tni(LLM_TENSOR_ATTN_Q, ".weight.lora_b", i)); - ggml_set_name(layer.wk_a, tni(LLM_TENSOR_ATTN_K, ".weight.lora_a", i)); - ggml_set_name(layer.wk_b, tni(LLM_TENSOR_ATTN_K, ".weight.lora_b", i)); - ggml_set_name(layer.wv_a, tni(LLM_TENSOR_ATTN_V, ".weight.lora_a", i)); - ggml_set_name(layer.wv_b, tni(LLM_TENSOR_ATTN_V, ".weight.lora_b", i)); - ggml_set_name(layer.wo_a, tni(LLM_TENSOR_ATTN_OUT, ".weight.lora_a", i)); - ggml_set_name(layer.wo_b, tni(LLM_TENSOR_ATTN_OUT, ".weight.lora_b", i)); - ggml_set_name(layer.ffn_norm_a, tni(LLM_TENSOR_FFN_NORM, ".weight.lora_a", i)); - ggml_set_name(layer.ffn_norm_b, tni(LLM_TENSOR_FFN_NORM, ".weight.lora_b", i)); - ggml_set_name(layer.ffn_gate_a, tni(LLM_TENSOR_FFN_GATE, ".weight.lora_a", i)); - ggml_set_name(layer.ffn_gate_b, tni(LLM_TENSOR_FFN_GATE, ".weight.lora_b", i)); - ggml_set_name(layer.ffn_down_a, tni(LLM_TENSOR_FFN_DOWN, ".weight.lora_a", i)); - ggml_set_name(layer.ffn_down_b, tni(LLM_TENSOR_FFN_DOWN, ".weight.lora_b", i)); - ggml_set_name(layer.ffn_up_a, tni(LLM_TENSOR_FFN_UP, ".weight.lora_a", i)); - ggml_set_name(layer.ffn_up_b, tni(LLM_TENSOR_FFN_UP, ".weight.lora_b", i)); - } - - set_param_lora(lora); - - // allocate data for lora tensors - lora->data = ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_cpu_buffer_type()); -} - -static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, float std, float min, float max) { - const uint32_t n_layer = lora->layers.size(); - - struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max); - - randomize_tensor_normal(lora->tok_embeddings_a, rnd); - ggml_set_zero(lora->tok_embeddings_b); - randomize_tensor_normal(lora->norm_a, rnd); - ggml_set_zero(lora->norm_b); - randomize_tensor_normal(lora->output_a, rnd); - ggml_set_zero(lora->output_b); - - for (uint32_t i = 0; i < n_layer; ++i) { - auto & layer = lora->layers[i]; - randomize_tensor_normal(layer.attention_norm_a, rnd); - ggml_set_zero(layer.attention_norm_b); - - randomize_tensor_normal(layer.wq_a, rnd); - ggml_set_zero(layer.wq_b); - randomize_tensor_normal(layer.wk_a, rnd); - ggml_set_zero(layer.wk_b); - randomize_tensor_normal(layer.wv_a, rnd); - ggml_set_zero(layer.wv_b); - randomize_tensor_normal(layer.wo_a, rnd); - ggml_set_zero(layer.wo_b); - - randomize_tensor_normal(layer.ffn_norm_a, rnd); - ggml_set_zero(layer.ffn_norm_b); - - randomize_tensor_normal(layer.ffn_gate_a, rnd); - ggml_set_zero(layer.ffn_gate_b); - randomize_tensor_normal(layer.ffn_down_a, rnd); - ggml_set_zero(layer.ffn_down_b); - randomize_tensor_normal(layer.ffn_up_a, rnd); - ggml_set_zero(layer.ffn_up_b); - } - - free_random_normal_distribution(rnd); -} - -static struct ggml_tensor * llama_build_lora_finetune_graphs( - struct my_llama_model * model, - struct my_llama_lora * lora, - ggml_gallocr_t alloc, - struct ggml_context * ctx, - struct ggml_cgraph * gf, - struct ggml_cgraph * gb, - struct ggml_cgraph * gb_tmp, - struct ggml_tensor * * logits, - struct ggml_tensor * tokens_input, - struct ggml_tensor * targets, - const int n_tokens, - const int n_batch, - const bool enable_flash_attn, - const bool enable_checkpointing, - const bool measure_only) { - - ggml_set_scratch(ctx, { 0, 0, nullptr, }); - const int n_past = 0; - const int N = n_tokens; - const auto & hparams = model->hparams; - const int n_ctx = hparams.n_ctx; - const int n_vocab = hparams.n_vocab; - const int n_embd = hparams.n_embd; - const int n_layer = hparams.n_layer; - const int n_head = hparams.n_head; - const int n_head_kv = hparams.n_head_kv; - const int n_ff = hparams.n_ff; - const int n_rot = hparams.n_embd_head(); - const int n_embd_head = hparams.n_embd_head(); - const int n_embd_gqa = hparams.n_embd_gqa(); - - const float rms_norm_eps = hparams.f_norm_rms_eps; - const float rope_freq_base = hparams.rope_freq_base; - const float rope_freq_scale = hparams.rope_freq_scale; - - GGML_ASSERT((size_t) n_layer == lora->layers.size()); - - auto set_name = [](struct ggml_tensor * t, const char * n) { - ggml_set_name(t, n); - if (t->grad) { - ggml_format_name(t->grad, "%s->grad", n); - } - }; - - // KQ_pos - contains the positions - struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N); - ggml_set_input(KQ_pos); - - // rope has so much parameters that we make a custom function for it - auto rope = [ctx, KQ_pos, n_rot, n_ctx, rope_freq_base, rope_freq_scale] - (struct ggml_tensor * t) -> struct ggml_tensor * { - // not capturing these, to silcence warnings - const int rope_mode = 0; - - return ggml_rope_ext(ctx, - t, KQ_pos, nullptr, n_rot, rope_mode, n_ctx, - rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f - ); - }; - - set_name(tokens_input, "tokens_input"); - set_name(targets, "targets"); - - GGML_ASSERT(tokens_input->type == GGML_TYPE_I32); - - auto add_to_f32 = [] (struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) { - if (ggml_is_quantized(a->type) || a->type == GGML_TYPE_F16 || a->type == GGML_TYPE_BF16) { - return ggml_add_cast(ctx, a, b, GGML_TYPE_F32); - } else if (a->type == GGML_TYPE_F32) { - return ggml_add(ctx, a, b); - } else { - die_fmt("%s: Finetuning on tensors with type '%s' is not yet supported.\n", - __func__, ggml_type_name(a->type)); - } - }; - - struct ggml_tensor * tok_embeddings = add_to_f32(ctx, model->tok_embeddings, ggml_mul_mat(ctx, lora->tok_embeddings_a, lora->tok_embeddings_b)); - struct ggml_tensor * norm = add_to_f32(ctx, model->norm, ggml_mul_mat(ctx, lora->norm_a, lora->norm_b)); - struct ggml_tensor * output = add_to_f32(ctx, model->output, ggml_mul_mat(ctx, lora->output_a, lora->output_b)); - - struct ggml_tensor * t00 = ggml_reshape_1d(ctx, tokens_input, N*n_batch); set_name(t00, "t00"); assert_shape_1d(t00, N*n_batch); - struct ggml_tensor * t01 = ggml_get_rows(ctx, tok_embeddings, t00); set_name(t01, "t01"); assert_shape_2d(t01, n_embd, N*n_batch); - - struct ggml_tensor * cur = t01; - - std::vector checkpoints; - if (enable_checkpointing) { - checkpoints.push_back(tokens_input); - checkpoints.push_back(targets); - checkpoints.push_back(t00); - checkpoints.push_back(t01); - } - - const float kv_scale = 1.0f/sqrtf(float(n_embd)/n_head); - - for (int il = 0; il < n_layer; ++il) { - struct my_llama_layer & layer = model->layers[il]; - struct my_llama_lora_layer & llayer = lora->layers[il]; - - struct ggml_tensor * attention_norm = add_to_f32(ctx, layer.attention_norm, ggml_mul_mat(ctx, llayer.attention_norm_a, llayer.attention_norm_b)); - struct ggml_tensor * ffn_norm = add_to_f32(ctx, layer.ffn_norm, ggml_mul_mat(ctx, llayer.ffn_norm_a, llayer.ffn_norm_b)); - struct ggml_tensor * wq = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b)); - struct ggml_tensor * wk = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b)); - struct ggml_tensor * wv = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b)); - struct ggml_tensor * wo = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b)); - struct ggml_tensor * ffn_gate = add_to_f32(ctx, layer.ffn_gate, ggml_mul_mat(ctx, llayer.ffn_gate_a, llayer.ffn_gate_b)); - struct ggml_tensor * ffn_down = add_to_f32(ctx, layer.ffn_down, ggml_mul_mat(ctx, llayer.ffn_down_a, llayer.ffn_down_b)); - struct ggml_tensor * ffn_up = add_to_f32(ctx, layer.ffn_up, ggml_mul_mat(ctx, llayer.ffn_up_a, llayer.ffn_up_b)); - - struct ggml_tensor * t02 = ggml_rms_norm (ctx, cur, rms_norm_eps); set_name(t02, "t02"); assert_shape_2d(t02, n_embd, N*n_batch); - struct ggml_tensor * t03 = ggml_repeat (ctx, attention_norm, t02); set_name(t03, "t03"); assert_shape_2d(t03, n_embd, N*n_batch); - struct ggml_tensor * t04 = ggml_mul (ctx, t03, t02); set_name(t04, "t04"); assert_shape_2d(t04, n_embd, N*n_batch); - struct ggml_tensor * t05 = ggml_mul_mat (ctx, wq, t04); set_name(t05, "t05"); assert_shape_2d(t05, n_embd, N*n_batch); - struct ggml_tensor * t06 = ggml_reshape_4d (ctx, t05, n_embd_head, n_head, N, n_batch); set_name(t06, "t06"); assert_shape_4d(t06, n_embd_head, n_head, N, n_batch); - struct ggml_tensor * t07 = rope (t06); set_name(t07, "t07"); assert_shape_4d(t07, n_embd_head, n_head, N, n_batch); - struct ggml_tensor * t08 = ggml_mul_mat (ctx, wk, t04); set_name(t08, "t08"); assert_shape_2d(t08, n_embd_gqa, N*n_batch); - struct ggml_tensor * t09 = ggml_reshape_4d (ctx, t08, n_embd_head, n_head_kv, N, n_batch); set_name(t09, "t09"); assert_shape_4d(t09, n_embd_head, n_head_kv, N, n_batch); - struct ggml_tensor * t10 = rope (t09); set_name(t10, "t10"); assert_shape_4d(t10, n_embd_head, n_head_kv, N, n_batch); - - struct ggml_tensor * t11; - if (ggml_is_quantized(wv->type)) { - struct ggml_tensor * t11_1 = ggml_mul_mat (ctx, wv, t04); set_name(t11_1, "t11_1"); assert_shape_2d(t11_1, n_embd_gqa, N*n_batch); - struct ggml_tensor * t11_2 = ggml_transpose(ctx, t11_1); set_name(t11_2, "t11_2"); assert_shape_2d(t11_2, N*n_batch, n_embd_gqa); - t11 = ggml_cont (ctx, t11_2); set_name(t11, "t11"); assert_shape_2d(t11, N*n_batch, n_embd_gqa); - } else { - t11 = ggml_mul_mat (ctx, t04, wv); set_name(t11, "t11"); assert_shape_2d(t11, N*n_batch, n_embd_gqa); - } - - struct ggml_tensor * t12 = ggml_reshape_4d (ctx, t11, N, n_batch, n_embd_head, n_head_kv); set_name(t12, "t12"); assert_shape_4d(t12, N, n_batch, n_embd_head, n_head_kv); - struct ggml_tensor * t13 = ggml_permute (ctx, t07, 0, 2, 1, 3); set_name(t13, "t13"); assert_shape_4d(t13, n_embd_head, N, n_head, n_batch); - struct ggml_tensor * t14 = ggml_permute (ctx, t10, 0, 2, 1, 3); set_name(t14, "t14"); assert_shape_4d(t14, n_embd_head, N, n_head_kv, n_batch); - struct ggml_tensor * t15 = ggml_permute (ctx, t12, 0, 3, 1, 2); set_name(t15, "t15"); assert_shape_4d(t15, N, n_embd_head, n_head_kv, n_batch); - struct ggml_tensor * t16; - if (enable_flash_attn) { - GGML_ASSERT(false && "TODO: ggml_flash_attn_ext() not yet supported"); - //t16 = ggml_flash_attn(ctx, t13, t14, t15, true); set_name(t16, "t16"); assert_shape_4d(t16, n_embd_head, N, n_head, n_batch); - } else { - struct ggml_tensor * t16_0 = ggml_mul_mat (ctx, t14, t13); set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch); - struct ggml_tensor * t16_1 = ggml_scale_inplace (ctx, t16_0, kv_scale); set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch); - struct ggml_tensor * t16_2 = ggml_diag_mask_inf_inplace(ctx, t16_1, n_past); set_name(t16_2, "t16_2"); assert_shape_4d(t16_2, N, N, n_head, n_batch); - struct ggml_tensor * t16_3 = ggml_soft_max_inplace (ctx, t16_2); set_name(t16_3, "t16_3"); assert_shape_4d(t16_3, N, N, n_head, n_batch); - t16 = ggml_mul_mat(ctx, t15, t16_3); set_name(t16, "t16"); assert_shape_4d(t16, n_embd_head, N, n_head, n_batch); - } - struct ggml_tensor * t17 = ggml_permute (ctx, t16, 0, 2, 1, 3); set_name(t17, "t17"); assert_shape_4d(t17, n_embd_head, n_head, N, n_batch); - struct ggml_tensor * t18 = ggml_cont (ctx, t17); set_name(t18, "t18"); assert_shape_4d(t18, n_embd_head, n_head, N, n_batch); - struct ggml_tensor * t19 = ggml_reshape_2d (ctx, t18, n_embd, N*n_batch); set_name(t19, "t19"); assert_shape_2d(t19, n_embd, N*n_batch); - struct ggml_tensor * t20 = ggml_mul_mat (ctx, wo, t19); set_name(t20, "t20"); assert_shape_2d(t20, n_embd, N*n_batch); - struct ggml_tensor * t21 = ggml_add (ctx, t20, cur); set_name(t21, "t21"); assert_shape_2d(t21, n_embd, N*n_batch); - struct ggml_tensor * t22 = ggml_rms_norm (ctx, t21, rms_norm_eps); set_name(t22, "t22"); assert_shape_2d(t22, n_embd, N*n_batch); - struct ggml_tensor * t23 = ggml_repeat (ctx, ffn_norm, t22); set_name(t23, "t23"); assert_shape_2d(t23, n_embd, N*n_batch); - struct ggml_tensor * t24 = ggml_mul (ctx, t23, t22); set_name(t24, "t24"); assert_shape_2d(t24, n_embd, N*n_batch); - struct ggml_tensor * t25 = ggml_mul_mat (ctx, ffn_up, t24); set_name(t25, "t25"); assert_shape_2d(t25, n_ff, N*n_batch); - struct ggml_tensor * t26 = ggml_mul_mat (ctx, ffn_gate, t24); set_name(t26, "t26"); assert_shape_2d(t26, n_ff, N*n_batch); - struct ggml_tensor * t27 = ggml_silu (ctx, t26); set_name(t27, "t27"); assert_shape_2d(t27, n_ff, N*n_batch); - struct ggml_tensor * t28 = ggml_mul (ctx, t27, t25); set_name(t28, "t28"); assert_shape_2d(t28, n_ff, N*n_batch); - struct ggml_tensor * t29 = ggml_mul_mat (ctx, ffn_down, t28); set_name(t29, "t29"); assert_shape_2d(t29, n_embd, N*n_batch); - struct ggml_tensor * t30 = ggml_add (ctx, t29, t21); set_name(t30, "t30"); assert_shape_2d(t30, n_embd, N*n_batch); - cur = t30; - if (enable_checkpointing) { - checkpoints.push_back(cur); - } - } - struct ggml_tensor * t31 = ggml_rms_norm (ctx, cur, rms_norm_eps); set_name(t31, "t31"); assert_shape_2d(t31, n_embd, N*n_batch); - struct ggml_tensor * t32 = ggml_repeat (ctx, norm, t31); set_name(t32, "t32"); assert_shape_2d(t32, n_embd, N*n_batch); - struct ggml_tensor * t33 = ggml_mul (ctx, t32, t31); set_name(t33, "t33"); assert_shape_2d(t33, n_embd, N*n_batch); - struct ggml_tensor * t34 = ggml_mul_mat (ctx, output, t33); set_name(t34, "t34"); assert_shape_2d(t34, n_vocab, N*n_batch); - struct ggml_tensor * t35 = ggml_reshape_3d (ctx, t34, n_vocab, N, n_batch); set_name(t35, "t35"); assert_shape_3d(t35, n_vocab, N, n_batch); - struct ggml_tensor * t36 = ggml_cross_entropy_loss(ctx, t35, targets); set_name(t36, "t36"); assert_shape_1d(t36, 1); - - if (enable_checkpointing) { - checkpoints.push_back(t31); - checkpoints.push_back(t32); - checkpoints.push_back(t33); - checkpoints.push_back(t34); - checkpoints.push_back(t35); - checkpoints.push_back(t36); - } - - ggml_build_forward_expand(gf, t36); - - if (enable_checkpointing) { - ggml_build_backward_gradient_checkpointing(ctx, gf, gb, gb_tmp, checkpoints.data(), (int) checkpoints.size()); - } else { - ggml_graph_cpy(gf, gb); - ggml_build_backward_expand(ctx, gf, gb, true); - } - - GGML_ASSERT(alloc != NULL); - - // make sure some tensors are not reallocated by inserting new temporary nodes depending on them - int n_leafs_before = gb->n_leafs; - int n_nodes_before = gb->n_nodes; - - // output tensors - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, 1.0f)); - // input gradient - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f)); - GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL); - ggml_set_input(t36->grad); - // KQ_pos - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f)); - - // make sure base model tensors data cannot be used in viewable operations - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, 1.0f)); - for (int il = 0; il < n_layer; ++il) { - struct my_llama_layer & layer = model->layers[il]; - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_gate, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_down, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_up, 1.0f)); - } - - // allocating checkpoints in one block to reduce memory fragmentation - // note: they will be freed in reverse order - for (unsigned int i = 0; i < checkpoints.size(); ++i) { - if (checkpoints[i]->data == NULL && checkpoints[i]->view_src == NULL) { - ggml_set_input(checkpoints[i]); - } - } - - if (measure_only) { - ggml_gallocr_reserve(alloc, gb); - } else { - ggml_gallocr_alloc_graph(alloc, gb); - - // set KQ_pos - { - int * data = (int *) KQ_pos->data; - for (int i = 0; i < N; ++i) { - data[i] = n_past + i; - } - } - } - - // remove the additional nodes and leafs - for (int i = n_leafs_before; i < gb->n_leafs; ++i) { - gb->leafs[i] = NULL; - } - for (int i = n_nodes_before; i < gb->n_nodes; ++i) { - gb->nodes[i] = NULL; - } - gb->n_leafs = n_leafs_before; - gb->n_nodes = n_nodes_before; - - *logits = t35; - return t36; -} - -static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model, struct my_llama_lora * lora) { - // NOTE: gguf_context must be initialized with f_ggml_ctx and no_alloc=false, otherwise tensor data can not be read - - std::string arch; - - std::vector keybuf; - keybuf.resize(512); - - GGUF_GET_KEY(fctx, arch, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_GENERAL_ARCHITECTURE); - GGML_ASSERT(arch == "llama"); - - uint32_t ftype_u; - GGUF_GET_KEY(fctx, ftype_u, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_GENERAL_FILE_TYPE); - GGML_ASSERT((enum llama_ftype) ftype_u == LLAMA_FTYPE_ALL_F32); - - struct my_llama_hparams hparams; - load_model_hparams_gguf(fctx, &hparams, arch.c_str()); - - // parameters that define tensor shapes must match - GGML_ASSERT(hparams.n_embd == model->hparams.n_embd); - GGML_ASSERT(hparams.n_ff == model->hparams.n_ff); - GGML_ASSERT(hparams.n_head == model->hparams.n_head); - GGML_ASSERT(hparams.n_head_kv == model->hparams.n_head_kv); - GGML_ASSERT(hparams.n_layer == model->hparams.n_layer); - - GGUF_GET_KEY(fctx, lora->hparams.n_rank_tok_embeddings, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_norm, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_output, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_OUTPUT); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_attention_norm, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_NORM); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_wq, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_Q); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_wk, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_K); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_wv, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_V); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_wo, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_norm, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_NORM); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_gate, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_down, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_up, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP); - - init_lora(model, lora); - - copy_tensor_by_name(lora->tok_embeddings_a, f_ggml_ctx, ggml_get_name(lora->tok_embeddings_a)); - copy_tensor_by_name(lora->tok_embeddings_b, f_ggml_ctx, ggml_get_name(lora->tok_embeddings_b)); - copy_tensor_by_name(lora->norm_a, f_ggml_ctx, ggml_get_name(lora->norm_a)); - copy_tensor_by_name(lora->norm_b, f_ggml_ctx, ggml_get_name(lora->norm_b)); - copy_tensor_by_name(lora->output_a, f_ggml_ctx, ggml_get_name(lora->output_a)); - copy_tensor_by_name(lora->output_b, f_ggml_ctx, ggml_get_name(lora->output_b)); - - for (uint32_t i = 0; i < lora->layers.size(); ++i) { - auto & layer = lora->layers[i]; - copy_tensor_by_name(layer.attention_norm_a, f_ggml_ctx, ggml_get_name(layer.attention_norm_a)); - copy_tensor_by_name(layer.attention_norm_b, f_ggml_ctx, ggml_get_name(layer.attention_norm_b)); - copy_tensor_by_name(layer.wq_a, f_ggml_ctx, ggml_get_name(layer.wq_a)); - copy_tensor_by_name(layer.wq_b, f_ggml_ctx, ggml_get_name(layer.wq_b)); - copy_tensor_by_name(layer.wk_a, f_ggml_ctx, ggml_get_name(layer.wk_a)); - copy_tensor_by_name(layer.wk_b, f_ggml_ctx, ggml_get_name(layer.wk_b)); - copy_tensor_by_name(layer.wv_a, f_ggml_ctx, ggml_get_name(layer.wv_a)); - copy_tensor_by_name(layer.wv_b, f_ggml_ctx, ggml_get_name(layer.wv_b)); - copy_tensor_by_name(layer.wo_a, f_ggml_ctx, ggml_get_name(layer.wo_a)); - copy_tensor_by_name(layer.wo_b, f_ggml_ctx, ggml_get_name(layer.wo_b)); - copy_tensor_by_name(layer.ffn_norm_a, f_ggml_ctx, ggml_get_name(layer.ffn_norm_a)); - copy_tensor_by_name(layer.ffn_norm_b, f_ggml_ctx, ggml_get_name(layer.ffn_norm_b)); - copy_tensor_by_name(layer.ffn_gate_a, f_ggml_ctx, ggml_get_name(layer.ffn_gate_a)); - copy_tensor_by_name(layer.ffn_gate_b, f_ggml_ctx, ggml_get_name(layer.ffn_gate_b)); - copy_tensor_by_name(layer.ffn_down_a, f_ggml_ctx, ggml_get_name(layer.ffn_down_a)); - copy_tensor_by_name(layer.ffn_down_b, f_ggml_ctx, ggml_get_name(layer.ffn_down_b)); - copy_tensor_by_name(layer.ffn_up_a, f_ggml_ctx, ggml_get_name(layer.ffn_up_a)); - copy_tensor_by_name(layer.ffn_up_b, f_ggml_ctx, ggml_get_name(layer.ffn_up_b)); - } -} - -static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_model * model, struct my_llama_lora * lora) { - const char * arch = "llama"; - enum llama_ftype ftype = LLAMA_FTYPE_ALL_F32; - - std::vector keybuf; - keybuf.resize(512); - auto kv = [arch, &keybuf](const char * key) -> const char * { - snprintf(keybuf.data(), keybuf.size(), key, arch); - return keybuf.data(); - }; - - gguf_set_val_str(fctx, LLM_KV_GENERAL_ARCHITECTURE, arch); - gguf_set_val_u32(fctx, LLM_KV_GENERAL_FILE_TYPE, ftype); - - gguf_set_val_u32(fctx, kv(LLM_KV_CONTEXT_LENGTH), model->hparams.n_ctx); - gguf_set_val_u32(fctx, kv(LLM_KV_EMBEDDING_LENGTH), model->hparams.n_embd); - gguf_set_val_u32(fctx, kv(LLM_KV_FEED_FORWARD_LENGTH), model->hparams.n_ff); - gguf_set_val_u32(fctx, kv(LLM_KV_ATTENTION_HEAD_COUNT), model->hparams.n_head); - gguf_set_val_u32(fctx, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV), model->hparams.n_head_kv); - gguf_set_val_u32(fctx, kv(LLM_KV_BLOCK_COUNT), model->hparams.n_layer); - gguf_set_val_u32(fctx, kv(LLM_KV_ROPE_DIMENSION_COUNT), model->hparams.n_embd_head()); - gguf_set_val_f32(fctx, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS), model->hparams.f_norm_rms_eps); - gguf_set_val_f32(fctx, kv(LLM_KV_ROPE_FREQ_BASE), model->hparams.rope_freq_base); - gguf_set_val_f32(fctx, kv(LLM_KV_ROPE_SCALE_LINEAR), model->hparams.rope_freq_scale); - - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD, lora->hparams.n_rank_tok_embeddings); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM, lora->hparams.n_rank_norm); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_OUTPUT, lora->hparams.n_rank_output); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_NORM, lora->hparams.n_rank_attention_norm); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_Q, lora->hparams.n_rank_wq); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_K, lora->hparams.n_rank_wk); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_V, lora->hparams.n_rank_wv); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT, lora->hparams.n_rank_wo); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_NORM, lora->hparams.n_rank_ffn_norm); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE, lora->hparams.n_rank_ffn_gate); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN, lora->hparams.n_rank_ffn_down); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP, lora->hparams.n_rank_ffn_up); - - gguf_add_tensor(fctx, lora->tok_embeddings_a); - gguf_add_tensor(fctx, lora->tok_embeddings_b); - gguf_add_tensor(fctx, lora->norm_a); - gguf_add_tensor(fctx, lora->norm_b); - gguf_add_tensor(fctx, lora->output_a); - gguf_add_tensor(fctx, lora->output_b); - - for (uint32_t i = 0; i < lora->layers.size(); ++i) { - auto & layer = lora->layers[i]; - - gguf_add_tensor(fctx, layer.attention_norm_a); - gguf_add_tensor(fctx, layer.attention_norm_b); - gguf_add_tensor(fctx, layer.wq_a); - gguf_add_tensor(fctx, layer.wq_b); - gguf_add_tensor(fctx, layer.wk_a); - gguf_add_tensor(fctx, layer.wk_b); - gguf_add_tensor(fctx, layer.wv_a); - gguf_add_tensor(fctx, layer.wv_b); - gguf_add_tensor(fctx, layer.wo_a); - gguf_add_tensor(fctx, layer.wo_b); - gguf_add_tensor(fctx, layer.ffn_norm_a); - gguf_add_tensor(fctx, layer.ffn_norm_b); - gguf_add_tensor(fctx, layer.ffn_gate_a); - gguf_add_tensor(fctx, layer.ffn_gate_b); - gguf_add_tensor(fctx, layer.ffn_down_a); - gguf_add_tensor(fctx, layer.ffn_down_b); - gguf_add_tensor(fctx, layer.ffn_up_a); - gguf_add_tensor(fctx, layer.ffn_up_b); - } -} - -static void load_checkpoint_lora_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) { - std::string train_type = LLM_KV_TRAINING_TYPE_FINETUNE_LORA; - GGUF_GET_KEY(fctx, train_type, gguf_get_val_str, GGUF_TYPE_STRING, false, LLM_KV_TRAINING_TYPE); - GGML_ASSERT(train_type == LLM_KV_TRAINING_TYPE_FINETUNE_LORA); - - load_train_state_gguf(fctx, f_ggml_ctx, train); - load_llama_lora_gguf(fctx, f_ggml_ctx, model, lora); -} - -static void save_checkpoint_lora_gguf(struct gguf_context * fctx, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) { - gguf_set_val_str(fctx, LLM_KV_TRAINING_TYPE, LLM_KV_TRAINING_TYPE_FINETUNE_LORA); - save_llama_lora_gguf(fctx, model, lora); - save_train_state_gguf(fctx, train); -} - -static bool load_checkpoint_lora_file(const char * filename, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) { - struct ggml_context * f_ggml_ctx; - struct gguf_init_params params; - params.no_alloc = false; - params.ctx = &f_ggml_ctx; - struct gguf_context * fctx = gguf_init_from_file(filename, params); - if (fctx == NULL) { - return false; - } - - load_checkpoint_lora_gguf(fctx, f_ggml_ctx, model, lora, train); - - gguf_free(fctx); - return true; -} - -static void save_checkpoint_lora_file(const char * filename, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) { - printf("%s: saving to %s\n", __func__, filename); - struct gguf_context * fctx = gguf_init_empty(); - - save_checkpoint_lora_gguf(fctx, model, lora, train); - - // write file - const bool only_meta = false; - gguf_write_to_file(fctx, filename, only_meta); - gguf_free(fctx); -} - -struct llama_file { - // use FILE * so we don't have to re-open the file to mmap - FILE * fp; - size_t size; - - llama_file(const char * fname, const char * mode) { - fp = std::fopen(fname, mode); - if (fp == NULL) { - size = 0; - } else { - seek(0, SEEK_END); - size = tell(); - seek(0, SEEK_SET); - } - } - - size_t tell() const { -#ifdef _WIN32 - __int64 ret = _ftelli64(fp); -#else - long ret = std::ftell(fp); -#endif - GGML_ASSERT(ret != -1); // this really shouldn't fail - return (size_t) ret; - } - - void seek(size_t offset, int whence) { -#ifdef _WIN32 - int ret = _fseeki64(fp, (__int64) offset, whence); -#else - int ret = std::fseek(fp, (long) offset, whence); -#endif - GGML_ASSERT(ret == 0); // same - } - - void read_raw(void * ptr, size_t size) { - if (size == 0) { - return; - } - errno = 0; - std::size_t ret = std::fread(ptr, size, 1, fp); - if (ferror(fp)) { - die_fmt("read error: %s", strerror(errno)); - } - if (ret != 1) { - die("unexpectedly reached end of file"); - } - } - - std::uint32_t read_u32() { - std::uint32_t ret; - read_raw(&ret, sizeof(ret)); - return ret; - } - - std::string read_string(std::uint32_t len) { - std::vector chars(len); - read_raw(chars.data(), len); - return std::string(chars.data(), len); - } - - void write_raw(const void * ptr, size_t size) { - if (size == 0) { - return; - } - errno = 0; - size_t ret = std::fwrite(ptr, size, 1, fp); - if (ret != 1) { - die_fmt("write error: %s", strerror(errno)); - } - } - - void write_u32(std::uint32_t val) { - write_raw(&val, sizeof(val)); - } - - ~llama_file() { - if (fp) { - std::fclose(fp); - } - } -}; - -static void write_tensor(struct llama_file * file, struct ggml_tensor * tensor, const char * name) { - if (tensor == NULL) { - file->write_u32(0); - file->write_u32(0); - file->write_u32(GGML_TYPE_F32); - file->seek((0-file->tell()) & 31, SEEK_CUR); - return; - } - if (name == NULL) { - name = ggml_get_name(tensor); - } - uint32_t name_len = strlen(name); - uint32_t nd = ggml_n_dims(tensor); - uint32_t ne[4] = { (uint32_t)tensor->ne[0], - (uint32_t)tensor->ne[1], - (uint32_t)tensor->ne[2], - (uint32_t)tensor->ne[3] }; - file->write_u32(nd); - file->write_u32(name_len); - file->write_u32(tensor->type); - file->write_raw(ne, sizeof(ne[0]) * nd); - file->write_raw(name, name_len); - file->seek((0-file->tell()) & 31, SEEK_CUR); - file->write_raw(tensor->data, ggml_nbytes(tensor)); -} - -static void save_as_llama_lora(const char * filename, struct my_llama_lora * lora) { - printf("%s: saving to %s\n", __func__, filename); - struct llama_file file(filename, "wb"); - if (file.fp == NULL) { - return; - } - - std::vector tn_buf; - tn_buf.resize(GGML_MAX_NAME); - - auto tn = [&tn_buf](const char * key, const char * suffix) -> const char * { - snprintf(tn_buf.data(), tn_buf.size(), "%s%s", key, suffix); - return tn_buf.data(); - }; - - auto tni = [&tn_buf](const char * key, int bid, const char * suffix) -> const char * { - snprintf(tn_buf.data(), tn_buf.size(), key, bid); - std::string s = tn_buf.data(); - snprintf(tn_buf.data(), tn_buf.size(), "%s%s", s.c_str(), suffix); - return tn_buf.data(); - }; - - // write_magic - file.write_u32(LLAMA_FILE_MAGIC_GGLA); // magic - file.write_u32(1); // version - // write_hparams - file.write_u32(lora->hparams.lora_r); - file.write_u32(lora->hparams.lora_alpha); - // write tensors - write_tensor(&file, lora->tok_embeddings_a, tn(LLM_TENSOR_TOKEN_EMBD, ".weight.loraA")); - write_tensor(&file, lora->tok_embeddings_b, tn(LLM_TENSOR_TOKEN_EMBD, ".weight.loraB")); - write_tensor(&file, lora->norm_a, tn(LLM_TENSOR_OUTPUT_NORM, ".weight.loraA")); - write_tensor(&file, lora->norm_b, tn(LLM_TENSOR_OUTPUT_NORM, ".weight.loraB")); - write_tensor(&file, lora->output_a, tn(LLM_TENSOR_OUTPUT, ".weight.loraA")); - write_tensor(&file, lora->output_b, tn(LLM_TENSOR_OUTPUT, ".weight.loraB")); - for (uint32_t i = 0; i < lora->layers.size(); ++i) { - auto & layer = lora->layers[i]; - write_tensor(&file, layer.attention_norm_a, tni(LLM_TENSOR_ATTN_NORM, i, ".weight.loraA")); - write_tensor(&file, layer.attention_norm_b, tni(LLM_TENSOR_ATTN_NORM, i, ".weight.loraB")); - write_tensor(&file, layer.wq_a, tni(LLM_TENSOR_ATTN_Q, i, ".weight.loraA")); - write_tensor(&file, layer.wq_b, tni(LLM_TENSOR_ATTN_Q, i, ".weight.loraB")); - write_tensor(&file, layer.wk_a, tni(LLM_TENSOR_ATTN_K, i, ".weight.loraA")); - write_tensor(&file, layer.wk_b, tni(LLM_TENSOR_ATTN_K, i, ".weight.loraB")); - write_tensor(&file, layer.wv_a, tni(LLM_TENSOR_ATTN_V, i, ".weight.loraA")); - write_tensor(&file, layer.wv_b, tni(LLM_TENSOR_ATTN_V, i, ".weight.loraB")); - write_tensor(&file, layer.wo_a, tni(LLM_TENSOR_ATTN_OUT, i, ".weight.loraA")); - write_tensor(&file, layer.wo_b, tni(LLM_TENSOR_ATTN_OUT, i, ".weight.loraB")); - write_tensor(&file, layer.ffn_norm_a, tni(LLM_TENSOR_FFN_NORM, i, ".weight.loraA")); - write_tensor(&file, layer.ffn_norm_b, tni(LLM_TENSOR_FFN_NORM, i, ".weight.loraB")); - write_tensor(&file, layer.ffn_gate_a, tni(LLM_TENSOR_FFN_GATE, i, ".weight.loraA")); - write_tensor(&file, layer.ffn_gate_b, tni(LLM_TENSOR_FFN_GATE, i, ".weight.loraB")); - write_tensor(&file, layer.ffn_down_a, tni(LLM_TENSOR_FFN_DOWN, i, ".weight.loraA")); - write_tensor(&file, layer.ffn_down_b, tni(LLM_TENSOR_FFN_DOWN, i, ".weight.loraB")); - write_tensor(&file, layer.ffn_up_a, tni(LLM_TENSOR_FFN_UP, i, ".weight.loraA")); - write_tensor(&file, layer.ffn_up_b, tni(LLM_TENSOR_FFN_UP, i, ".weight.loraB")); - } -} - -struct train_params { - struct train_params_common common; - - const char * fn_model_base; - const char * fn_lora_out; - - bool only_write_lora; - - float f_norm_rms_eps; - float rope_freq_base; - float rope_freq_scale; - - bool custom_f_norm_rms_eps; - bool custom_rope_freq_base; - bool custom_rope_freq_scale; - - int32_t lora_r; - int32_t lora_alpha; - bool custom_lora_alpha; - - uint32_t n_rank_attention_norm; - uint32_t n_rank_wq; - uint32_t n_rank_wk; - uint32_t n_rank_wv; - uint32_t n_rank_wo; - uint32_t n_rank_ffn_norm; - uint32_t n_rank_ffn_gate; - uint32_t n_rank_ffn_down; - uint32_t n_rank_ffn_up; - uint32_t n_rank_tok_embeddings; - uint32_t n_rank_norm; - uint32_t n_rank_output; - - bool custom_n_rank_attention_norm; - bool custom_n_rank_wq; - bool custom_n_rank_wk; - bool custom_n_rank_wv; - bool custom_n_rank_wo; - bool custom_n_rank_ffn_norm; - bool custom_n_rank_ffn_gate; - bool custom_n_rank_ffn_down; - bool custom_n_rank_ffn_up; - bool custom_n_rank_tok_embeddings; - bool custom_n_rank_norm; - bool custom_n_rank_output; -}; - -static struct train_params get_default_train_params() { - struct train_params params; - params.common = get_default_train_params_common(); - params.fn_model_base = ""; - params.fn_lora_out = "ggml-lora-ITERATION-f32.gguf"; - - params.only_write_lora = false; - - params.f_norm_rms_eps = 1e-5f; - params.rope_freq_base = 10000.0f; - params.rope_freq_scale = 1.0f; - - params.custom_f_norm_rms_eps = false; - params.custom_rope_freq_base = false; - params.custom_rope_freq_scale = false; - - params.lora_r = 4; - params.lora_alpha = 4; - params.custom_lora_alpha = false; - - params.n_rank_attention_norm = 1; - params.n_rank_wq = 4; - params.n_rank_wk = 4; - params.n_rank_wv = 4; - params.n_rank_wo = 4; - params.n_rank_ffn_norm = 1; - params.n_rank_ffn_gate = 4; - params.n_rank_ffn_down = 4; - params.n_rank_ffn_up = 4; - params.n_rank_tok_embeddings = 4; - params.n_rank_norm = 1; - params.n_rank_output = 4; - - params.custom_n_rank_attention_norm = false; - params.custom_n_rank_wq = false; - params.custom_n_rank_wk = false; - params.custom_n_rank_wv = false; - params.custom_n_rank_wo = false; - params.custom_n_rank_ffn_norm = false; - params.custom_n_rank_ffn_gate = false; - params.custom_n_rank_ffn_down = false; - params.custom_n_rank_ffn_up = false; - params.custom_n_rank_tok_embeddings = false; - params.custom_n_rank_norm = false; - params.custom_n_rank_output = false; - - return params; -} - -static void train_print_usage(int argc, char ** argv, const struct train_params * params) { - fprintf(stderr, "usage: %s [options]\n", argv[0]); - fprintf(stderr, "\n"); - fprintf(stderr, "options:\n"); - fprintf(stderr, " -h, --help show this help message and exit\n"); - - fprintf(stderr, " --model-base FNAME model path from which to load base model (default '%s')\n", params->fn_model_base); - fprintf(stderr, " --lora-out FNAME path to save llama lora (default '%s')\n", params->fn_lora_out); - fprintf(stderr, " --only-write-lora only save llama lora, don't do any training. use this if you only want to convert a checkpoint to a lora adapter.\n"); - fprintf(stderr, " --norm-rms-eps F RMS-Norm epsilon value (default %f)\n", params->f_norm_rms_eps); - fprintf(stderr, " --rope-freq-base F Frequency base for ROPE (default %f)\n", params->rope_freq_base); - fprintf(stderr, " --rope-freq-scale F Frequency scale for ROPE (default %f)\n", params->rope_freq_scale); - fprintf(stderr, " --lora-alpha N LORA alpha : resulting LORA scaling is alpha/r. (default %d)\n", params->lora_alpha); - fprintf(stderr, " --lora-r N LORA r: default rank. Also specifies resulting scaling together with lora-alpha. (default %d)\n", params->lora_r); - fprintf(stderr, " --rank-att-norm N LORA rank for attention norm tensor, overrides default rank. Norm tensors should generally have rank 1.\n"); - fprintf(stderr, " --rank-ffn-norm N LORA rank for feed-forward norm tensor, overrides default rank. Norm tensors should generally have rank 1.\n"); - fprintf(stderr, " --rank-out-norm N LORA rank for output norm tensor, overrides default rank. Norm tensors should generally have rank 1.\n"); - fprintf(stderr, " --rank-tok-embd N LORA rank for token embeddings tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-out N LORA rank for output tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-wq N LORA rank for wq tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-wk N LORA rank for wk tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-wv N LORA rank for wv tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-wo N LORA rank for wo tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-ffn_gate N LORA rank for ffn_gate tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-ffn_down N LORA rank for ffn_down tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-ffn_up N LORA rank for ffn_up tensor, overrides default rank.\n"); - - print_common_train_usage(argc, argv, ¶ms->common); -} - -static bool train_params_parse(int argc, char ** argv, struct train_params * params) { - bool invalid_param = false; - std::string arg; - struct train_params default_params = get_default_train_params(); - const std::string arg_prefix = "--"; - - for (int i = 1; i < argc; i++) { - arg = argv[i]; - if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) { - std::replace(arg.begin(), arg.end(), '_', '-'); - } - - if (consume_common_train_arg(argc, argv, &i, ¶ms->common, &invalid_param)) { - if (invalid_param) { - break; - } else if (params->common.print_usage) { - train_print_usage(argc, argv, &default_params); - exit(0); - } - } else if (arg == "--model-base") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->fn_model_base = argv[i]; - } else if (arg == "--lora-out") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->fn_lora_out = argv[i]; - } else if (arg == "--only-write-lora") { - params->only_write_lora = true; - } else if (arg == "--norm-rms-eps") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->f_norm_rms_eps = std::stof(argv[i]); - params->custom_f_norm_rms_eps = true; - } else if (arg == "--rope-freq-base") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->rope_freq_base = std::stof(argv[i]); - params->custom_rope_freq_base = true; - } else if (arg == "--rope-freq-scale") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->rope_freq_scale = std::stof(argv[i]); - params->custom_rope_freq_scale = true; - } else if (arg == "--lora-alpha") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->lora_alpha = std::stoi(argv[i]); - params->custom_lora_alpha = true; - } else if (arg == "--lora-r") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->lora_r = std::stoi(argv[i]); - } else if (arg == "--rank-att-norm") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_attention_norm = std::stoi(argv[i]); - params->custom_n_rank_attention_norm = true; - } else if (arg == "--rank-ffn-norm") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_ffn_norm = std::stoi(argv[i]); - params->custom_n_rank_ffn_norm = true; - } else if (arg == "--rank-out-norm") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_norm = std::stoi(argv[i]); - params->custom_n_rank_norm = true; - } else if (arg == "--rank-tok-embd") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_tok_embeddings = std::stoi(argv[i]); - params->custom_n_rank_tok_embeddings = true; - } else if (arg == "--rank-out") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_output = std::stoi(argv[i]); - params->custom_n_rank_output = true; - } else if (arg == "--rank-wq") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_wq = std::stoi(argv[i]); - params->custom_n_rank_wq = true; - } else if (arg == "--rank-wk") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_wk = std::stoi(argv[i]); - params->custom_n_rank_wk = true; - } else if (arg == "--rank-wv") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_wv = std::stoi(argv[i]); - params->custom_n_rank_wv = true; - } else if (arg == "--rank-wo") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_wo = std::stoi(argv[i]); - params->custom_n_rank_wo = true; - } else if (arg == "--rank-ffn_gate") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_ffn_gate = std::stoi(argv[i]); - params->custom_n_rank_ffn_gate = true; - } else if (arg == "--rank-ffn_down") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_ffn_down = std::stoi(argv[i]); - params->custom_n_rank_ffn_down = true; - } else if (arg == "--rank-ffn_up") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_ffn_up = std::stoi(argv[i]); - params->custom_n_rank_ffn_up = true; - } else { - fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); - train_print_usage(argc, argv, &default_params); - exit(1); - } - } - if (invalid_param) { - fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str()); - train_print_usage(argc, argv, &default_params); - exit(1); - } - finish_processing_train_args(¶ms->common); - return true; -} - -struct save_train_files_data { - const char * fn_checkpoint_out; - const char * fn_lora_out; - const char * pattern_fn_it; - const char * fn_latest; - struct my_llama_model * model; - struct my_llama_lora * lora; -}; - -static void save_train_files(void * vdata, struct train_state * train) { - struct save_train_files_data * data = (struct save_train_files_data *) vdata; - - int64_t iter = train->opt->iter; - - if (strlen(data->fn_checkpoint_out) > 0) { - save_checkpoint_lora_file(get_train_filename(data->fn_checkpoint_out, data->pattern_fn_it, data->fn_latest, iter).c_str(), data->model, data->lora, train); - save_checkpoint_lora_file(get_train_filename(data->fn_checkpoint_out, data->pattern_fn_it, data->fn_latest, -1 ).c_str(), data->model, data->lora, train); - } - if (strlen(data->fn_lora_out) > 0) { - save_as_llama_lora(get_train_filename(data->fn_lora_out, data->pattern_fn_it, data->fn_latest, iter).c_str(), data->lora); - save_as_llama_lora(get_train_filename(data->fn_lora_out, data->pattern_fn_it, data->fn_latest, -1 ).c_str(), data->lora); - } -} - -static int64_t get_parameter_count(struct my_llama_lora* lora) { - int64_t nx = 0; - nx += ggml_nelements(lora->tok_embeddings_a); - nx += ggml_nelements(lora->tok_embeddings_b); - nx += ggml_nelements(lora->norm_a); - nx += ggml_nelements(lora->norm_b); - nx += ggml_nelements(lora->output_a); - nx += ggml_nelements(lora->output_b); - - for (uint32_t i = 0; i < lora->layers.size(); ++i) { - auto & layer = lora->layers[i]; - nx += ggml_nelements(layer.attention_norm_a); - nx += ggml_nelements(layer.attention_norm_b); - nx += ggml_nelements(layer.wq_a); - nx += ggml_nelements(layer.wq_b); - nx += ggml_nelements(layer.wk_a); - nx += ggml_nelements(layer.wk_b); - nx += ggml_nelements(layer.wv_a); - nx += ggml_nelements(layer.wv_b); - nx += ggml_nelements(layer.wo_a); - nx += ggml_nelements(layer.wo_b); - nx += ggml_nelements(layer.ffn_norm_a); - nx += ggml_nelements(layer.ffn_norm_b); - nx += ggml_nelements(layer.ffn_gate_a); - nx += ggml_nelements(layer.ffn_gate_b); - nx += ggml_nelements(layer.ffn_down_a); - nx += ggml_nelements(layer.ffn_down_b); - nx += ggml_nelements(layer.ffn_up_a); - nx += ggml_nelements(layer.ffn_up_b); - } - return nx; -} - -int main(int argc, char ** argv) { - struct train_params params = get_default_train_params(); - - if (!train_params_parse(argc, argv, ¶ms)) { - return 1; - } - - if (params.common.seed == LLAMA_DEFAULT_SEED) { - params.common.seed = time(NULL); - } - printf("%s: seed: %u\n", __func__, params.common.seed); - srand(params.common.seed); - - struct llama_model_params llama_mparams = llama_model_default_params(); - llama_mparams.n_gpu_layers = params.common.n_gpu_layers; - llama_mparams.vocab_only = false; - - printf("%s: model base = '%s'\n", __func__, params.fn_model_base); - struct llama_model * lmodel = llama_load_model_from_file(params.fn_model_base, llama_mparams); - - struct llama_context_params llama_cparams = llama_context_default_params(); - struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_cparams); - - struct my_llama_model model; - init_model(lmodel, &model, params.fn_model_base, params.common.n_ctx); - - struct my_llama_lora lora; - - struct train_state * train = init_train_state(); - struct ggml_opt_context * opt = train->opt; - - // set params from command line - if (params.custom_f_norm_rms_eps) { - model.hparams.f_norm_rms_eps = params.f_norm_rms_eps; - } - if (params.custom_rope_freq_base) { - model.hparams.rope_freq_base = params.rope_freq_base; - } - if (params.custom_rope_freq_scale) { - model.hparams.rope_freq_scale = params.rope_freq_scale; - } - lora.hparams.lora_r = params.lora_r; - lora.hparams.lora_alpha = params.custom_lora_alpha ? params.lora_alpha : params.lora_r; - uint32_t n_rank_attention_norm = params.custom_n_rank_attention_norm ? params.n_rank_attention_norm : 1; - uint32_t n_rank_wq = params.custom_n_rank_wq ? params.n_rank_wq : params.lora_r; - uint32_t n_rank_wk = params.custom_n_rank_wk ? params.n_rank_wk : params.lora_r; - uint32_t n_rank_wv = params.custom_n_rank_wv ? params.n_rank_wv : params.lora_r; - uint32_t n_rank_wo = params.custom_n_rank_wo ? params.n_rank_wo : params.lora_r; - uint32_t n_rank_ffn_norm = params.custom_n_rank_ffn_norm ? params.n_rank_ffn_norm : 1; - uint32_t n_rank_ffn_gate = params.custom_n_rank_ffn_gate ? params.n_rank_ffn_gate : params.lora_r; - uint32_t n_rank_ffn_down = params.custom_n_rank_ffn_down ? params.n_rank_ffn_down : params.lora_r; - uint32_t n_rank_ffn_up = params.custom_n_rank_ffn_up ? params.n_rank_ffn_up : params.lora_r; - uint32_t n_rank_tok_embeddings = params.custom_n_rank_tok_embeddings ? params.n_rank_tok_embeddings : params.lora_r; - uint32_t n_rank_norm = params.custom_n_rank_norm ? params.n_rank_norm : 1; - uint32_t n_rank_output = params.custom_n_rank_output ? params.n_rank_output : params.lora_r; - lora.hparams.n_rank_attention_norm = n_rank_attention_norm; - lora.hparams.n_rank_wq = n_rank_wq; - lora.hparams.n_rank_wk = n_rank_wk; - lora.hparams.n_rank_wv = n_rank_wv; - lora.hparams.n_rank_wo = n_rank_wo; - lora.hparams.n_rank_ffn_norm = n_rank_ffn_norm; - lora.hparams.n_rank_ffn_gate = n_rank_ffn_gate; - lora.hparams.n_rank_ffn_down = n_rank_ffn_down; - lora.hparams.n_rank_ffn_up = n_rank_ffn_up; - lora.hparams.n_rank_tok_embeddings = n_rank_tok_embeddings; - lora.hparams.n_rank_norm = n_rank_norm; - lora.hparams.n_rank_output = n_rank_output; - - // set opt params from command line - opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM); - opt->params.print_forward_graph = false; - opt->params.print_backward_graph = false; - opt->params.graph_size = LLAMA_TRAIN_MAX_NODES; - opt->params.n_threads = params.common.n_threads; - opt->params.past = params.common.opt_past; - opt->params.delta = params.common.opt_delta; - opt->params.max_no_improvement = params.common.opt_max_no_improvement; - opt->params.n_gradient_accumulation = params.common.n_gradient_accumulation; - opt->params.adam.n_iter = params.common.adam_n_iter; - opt->params.adam.sched = 1.0f; - opt->params.adam.alpha = params.common.adam_alpha; - opt->params.adam.decay = params.common.adam_decay; - opt->params.adam.decay_min_ndim = params.common.adam_decay_min_ndim; - opt->params.adam.beta1 = params.common.adam_beta1; - opt->params.adam.beta2 = params.common.adam_beta2; - opt->params.adam.gclip = params.common.adam_gclip; - opt->params.adam.eps_f = params.common.adam_eps_f; - - printf("%s: init model\n", __func__); - bool existed = load_checkpoint_lora_file(params.common.fn_checkpoint_in, &model, &lora, train); - - if (existed) { - // overwrite last n_ctx with user provided n_ctx - if (params.common.custom_n_ctx) { - model.hparams.n_ctx = params.common.n_ctx; - } - - const bool opt_param_count_changed = ( - (lora.hparams.n_rank_attention_norm != n_rank_attention_norm) - || (lora.hparams.n_rank_wq != n_rank_wq) - || (lora.hparams.n_rank_wk != n_rank_wk) - || (lora.hparams.n_rank_wv != n_rank_wv) - || (lora.hparams.n_rank_wo != n_rank_wo) - || (lora.hparams.n_rank_ffn_norm != n_rank_ffn_norm) - || (lora.hparams.n_rank_ffn_gate != n_rank_ffn_gate) - || (lora.hparams.n_rank_ffn_down != n_rank_ffn_down) - || (lora.hparams.n_rank_ffn_up != n_rank_ffn_up) - || (lora.hparams.n_rank_tok_embeddings != n_rank_tok_embeddings) - || (lora.hparams.n_rank_norm != n_rank_norm) - || (lora.hparams.n_rank_output != n_rank_output) - ); - - const bool opt_past_changed = opt->params.past != params.common.opt_past; - - if (opt_param_count_changed) { - print_lora_params(&lora.hparams); - die("Provided rank differs from checkpoint file. To use different rank start finetune from scratch with empty input checkpoint, e.g --checkpoint-in ''. Aborting."); - // need to discard previous optimizer gradient statistics and opt_init with new shapes - // TODO - } - if (opt_past_changed) { - die("Optimizer parameter '--opt-past N' differs from checkpoint file. To use different value finetune from scratch with empty input checkpoint, e.g --checkpoint-in ''. Aborting"); - // need to discard previous optimizer past function value statistics and opt_init with new shapes - // TODO - } - } else { // existed == false - init_lora(&model, &lora); - randomize_lora(&lora, params.common.seed, 0.0f, 1.0f, -1.0f, +1.0f); - if (!params.only_write_lora) { - ggml_opt_init(opt->ctx, opt, opt->params, get_parameter_count(&lora)); - } - } - opt->iter = train->train_its; - - print_params(&model.hparams); - print_lora_params(&lora.hparams); - printf("%s: total train_iterations %llu\n", __func__, (long long unsigned) train->train_its); - printf("%s: seen train_samples %llu\n", __func__, (long long unsigned) train->train_samples); - printf("%s: seen train_tokens %llu\n", __func__, (long long unsigned) train->train_tokens); - printf("%s: completed train_epochs %llu\n", __func__, (long long unsigned) train->train_epochs); - printf("%s: lora_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(lora.ctx) + ggml_backend_buffer_get_size(lora.data)), (float) (ggml_used_mem(lora.ctx) + ggml_backend_buffer_get_size(lora.data)) / (1024.0f*1024.0f)); - - if (params.only_write_lora) { - save_train_files_data save_data; - save_data.fn_checkpoint_out = ""; - save_data.fn_lora_out = params.fn_lora_out; - save_data.pattern_fn_it = params.common.pattern_fn_it; - save_data.fn_latest = params.common.fn_latest; - save_data.model = &model; - save_data.lora = &lora; - - save_train_files(&save_data, train); - - free_train_state(train); - ggml_free(lora.ctx); - llama_free(lctx); - llama_free_model(lmodel); - return 0; - } - - printf("%s: opt_size = %zu bytes (%.1f MB)\n", __func__, ggml_get_mem_size(opt->ctx), (float) ggml_get_mem_size(opt->ctx) / (1024.0f*1024.0f)); - printf("%s: opt iter %d\n", __func__, opt->iter); - - int n_tokens = model.hparams.n_ctx; - int n_vocab = model.hparams.n_vocab; - int n_batch = params.common.n_batch; - - // context for input tensors without their data - struct ggml_init_params ctx_input_params = { - ggml_tensor_overhead() * 2, // mem_size - NULL, // mem_buffer - true, // no_alloc - }; - struct ggml_context * ctx_input = ggml_init(ctx_input_params); - - // the input tensors - struct ggml_tensor * tokens_input = ggml_new_tensor_2d(ctx_input, GGML_TYPE_I32, n_tokens, n_batch); - struct ggml_tensor * target_probs = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab, n_tokens, n_batch); - - // allocate input tensors - // measure required memory for input tensors - ggml_backend_buffer_t input_data = ggml_backend_alloc_ctx_tensors_from_buft(ctx_input, ggml_backend_cpu_buffer_type()); - size_t max_input_size = ggml_backend_buffer_get_size(input_data); - printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f)); - - // context for compute tensors without their data - const size_t estimated_compute_size_wo_data = ( - 2*LLAMA_TRAIN_MAX_NODES*ggml_tensor_overhead() + - (params.common.use_checkpointing ? 3 : 2)*(GGML_OBJECT_SIZE+ggml_graph_overhead_custom(LLAMA_TRAIN_MAX_NODES, true)) - ); - struct ggml_init_params ctx_compute_params = { - estimated_compute_size_wo_data, // mem_size - NULL, // mem_buffer - true, // no_alloc - }; - struct ggml_context * ctx_compute = NULL; - - struct ggml_tensor * loss = NULL; - struct ggml_tensor * logits = NULL; - - struct ggml_cgraph * gf = NULL; - struct ggml_cgraph * gb = NULL; - struct ggml_cgraph * gb_tmp = NULL; - - // measure required memory for compute tensors - size_t best_compute_size = SIZE_MAX; - enum ggml_cgraph_eval_order best_order = GGML_CGRAPH_EVAL_ORDER_COUNT; - // find best evaluation order - for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) { - ctx_compute = ggml_init(ctx_compute_params); - ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type()); - gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); - gf->order = (enum ggml_cgraph_eval_order) order; - gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); - gb_tmp = params.common.use_checkpointing - ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true) - : NULL; - loss = llama_build_lora_finetune_graphs( - &model, &lora, alloc, ctx_compute, - gf, gb, gb_tmp, - &logits, tokens_input, target_probs, - n_tokens, n_batch, - params.common.use_flash, - params.common.use_checkpointing, - true - ); - size_t max_compute_size = ggml_gallocr_get_buffer_size(alloc, 0); // FIXME: this will still allocate the buffer - if (max_compute_size < best_compute_size) { - best_compute_size = max_compute_size; - best_order = gf->order; - } - ggml_gallocr_free(alloc); - ggml_free(ctx_compute); - } - size_t max_compute_size = best_compute_size; - printf("%s: compute_size = %zu bytes (%.1f MB)\n", __func__, max_compute_size, (float) max_compute_size / (1024.0f*1024.0f)); - printf("%s: evaluation order = %s\n", __func__, - (best_order == GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT) ? "LEFT_TO_RIGHT" : - (best_order == GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT) ? "RIGHT_TO_LEFT" : - "invalid"); - - // allocate compute tensors - ctx_compute = ggml_init(ctx_compute_params); - ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type()); - gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); - gf->order = best_order; - gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); - gb_tmp = params.common.use_checkpointing - ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true) - : NULL; - loss = llama_build_lora_finetune_graphs( - &model, &lora, alloc, ctx_compute, - gf, gb, gb_tmp, - &logits, tokens_input, target_probs, - n_tokens, n_batch, - params.common.use_flash, - params.common.use_checkpointing, - false - ); - - // tokenize data - std::vector train_tokens; - std::vector train_samples_begin; - std::vector train_samples_size; - printf("%s: tokenize training data from %s\n", __func__, params.common.fn_train_data); - printf("%s: sample-start: %s\n", __func__, params.common.sample_start.c_str()); - printf("%s: include-sample-start: %s\n", __func__, params.common.include_sample_start ? "true" : "false"); - tokenize_file(lctx, - params.common.fn_train_data, - params.common.sample_start, - params.common.include_sample_start, - params.common.overlapping_samples, - n_tokens, - train_tokens, - train_samples_begin, - train_samples_size); - GGML_ASSERT(train_samples_begin.size() == train_samples_size.size()); - - printf("%s: number of training tokens: %zu\n", __func__, train_tokens.size()); - - std::vector token_noccurs; - token_noccurs.resize(model.hparams.n_vocab, 0); - for (unsigned int i = 0; i < train_tokens.size(); ++i) { - ++token_noccurs[train_tokens[i]]; - } - int n_unique_tokens = 0; - for (unsigned int i = 0; i < token_noccurs.size(); ++i) { - if (token_noccurs[i] == 0) continue; - ++n_unique_tokens; - } - printf("%s: number of unique tokens: %d\n", __func__, n_unique_tokens); - - size_t shuffle_samples_hash = compute_samples_hash(params.common.fn_train_data, train_samples_begin.data(), train_samples_size.data(), train_samples_size.size()); - const bool changed_train_data = (shuffle_samples_hash != train->shuffle_samples_hash) || (train->shuffle_sample_count != train_samples_size.size()); - if (changed_train_data) { - printf("%s: train data seems to have changed. restarting shuffled epoch.\n", __func__); - } - if (params.common.force_reshuffle) { - printf("%s: forced reshuffling of data. restarting with newly shuffled epoch.\n", __func__); - } - if ((train->shuffle_rng_state_current == "") || changed_train_data || params.common.force_reshuffle) { - train->shuffle_rng_state_current = mt19937_seed_to_state(params.common.seed); - train->shuffle_sample_count = train_samples_size.size(); - train->shuffle_next_sample = 0; - train->shuffle_samples_hash = shuffle_samples_hash; - } - std::vector train_shuffled_samples_offs; - std::vector train_shuffled_samples_begin; - std::vector train_shuffled_samples_size; - train_shuffled_samples_offs.resize(train_samples_begin.size()); - train_shuffled_samples_begin.resize(train_samples_begin.size()); - train_shuffled_samples_size.resize(train_samples_size.size()); - train->shuffle_rng_state_next = shuffle_samples( - train->shuffle_rng_state_current, - train_shuffled_samples_offs.data(), - train_shuffled_samples_begin.data(), - train_shuffled_samples_size.data(), - train_samples_begin.data(), - train_samples_size.data(), - train_samples_size.size()); - - printf("%s: begin training\n", __func__); - - save_train_files_data save_data; - save_data.fn_checkpoint_out = params.common.fn_checkpoint_out; - save_data.fn_lora_out = params.fn_lora_out; - save_data.pattern_fn_it = params.common.pattern_fn_it; - save_data.fn_latest = params.common.fn_latest; - save_data.model = &model; - save_data.lora = &lora; - - struct train_opt_callback_data opt_cb_data; - opt_cb_data.params = ¶ms.common; - opt_cb_data.train = train; - opt_cb_data.save_cb = &save_train_files; - opt_cb_data.save_data = &save_data; - opt_cb_data.lctx = lctx; - opt_cb_data.last_save_iter = opt->iter; - opt_cb_data.tokens_data = train_tokens.data(); - opt_cb_data.tokens_size = train_tokens.size(); - opt_cb_data.samples_begin = train_samples_begin.data(); - opt_cb_data.samples_size = train_samples_size.data(); - opt_cb_data.shuffled_samples_offs = train_shuffled_samples_offs.data(); - opt_cb_data.shuffled_samples_begin = train_shuffled_samples_begin.data(); - opt_cb_data.shuffled_samples_size = train_shuffled_samples_size.data(); - opt_cb_data.samples_count = train_samples_size.size(); - opt_cb_data.tokens_input = tokens_input; - opt_cb_data.target_probs = target_probs; - opt_cb_data.first_iter = opt->iter; - opt_cb_data.first_epoch = train->train_epochs; - opt_cb_data.iter_at_last_epoch = -1; - opt_cb_data.last_time = ggml_time_ms(); - opt_cb_data.millis_per_iter = 0.0; - - // measure required memory for work buffer - size_t max_work_size = ggml_graph_plan(gb, params.common.n_threads).work_size + GGML_OBJECT_SIZE; - printf("%s: work_size = %zu bytes (%.1f MB)\n", __func__, max_work_size, (float) max_work_size / (1024.0f*1024.0f)); - - // context for work buffer - struct ggml_init_params ctx_work_params = { - max_work_size, // mem_size - NULL, // mem_buffer - false, // no_alloc - }; - struct ggml_context * ctx_work = ggml_init(ctx_work_params); - - int64_t t0 = ggml_time_ms(); - - ggml_opt_resume_g(ctx_work, opt, loss, gf, gb, &train_opt_callback, (void *) &opt_cb_data); - - ggml_free(ctx_work); - ggml_free(ctx_compute); - ggml_free(ctx_input); - ggml_gallocr_free(alloc); - - - int64_t t1 = ggml_time_ms(); - printf("%s: total training time: ", __func__); - print_duration((double) (t1 - t0)); - printf("\n"); - - int new_iters = opt->iter - opt_cb_data.last_save_iter; - if (new_iters > 0) { - train->train_its += new_iters; - train->train_tokens += new_iters * opt->params.n_gradient_accumulation * n_batch * n_tokens; - - save_train_files(&save_data, train); - opt_cb_data.last_save_iter = opt->iter; - } - - ggml_free(opt->ctx); - free_train_state(train); - ggml_free(lora.ctx); - llama_free(lctx); - llama_free_model(lmodel); - return 0; -} diff --git a/examples/finetune/finetune.sh b/examples/finetune/finetune.sh deleted file mode 100644 index d7f2165e5..000000000 --- a/examples/finetune/finetune.sh +++ /dev/null @@ -1,34 +0,0 @@ -#!/bin/bash -cd `dirname $0` -cd ../.. - -EXE="./llama-finetune" - -if [[ ! $LLAMA_MODEL_DIR ]]; then LLAMA_MODEL_DIR="./models"; fi -if [[ ! $LLAMA_TRAINING_DIR ]]; then LLAMA_TRAINING_DIR="."; fi - -# MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2-q8_0.gguf" # This is the model the readme uses. -MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2.gguf" # An f16 model. Note in this case with "-g", you get an f32-format .BIN file that isn't yet supported if you use it with "main --lora" with GPU inferencing. - -while getopts "dg" opt; do - case $opt in - d) - DEBUGGER="gdb --args" - ;; - g) - EXE="./build/bin/Release/finetune" - GPUARG="--gpu-layers 25" - ;; - esac -done - -$DEBUGGER $EXE \ - --model-base $MODEL \ - $GPUARG \ - --checkpoint-in chk-ol3b-shakespeare-LATEST.gguf \ - --checkpoint-out chk-ol3b-shakespeare-ITERATION.gguf \ - --lora-out lora-ol3b-shakespeare-ITERATION.bin \ - --train-data "$LLAMA_TRAINING_DIR\shakespeare.txt" \ - --save-every 10 \ - --threads 10 --adam-iter 30 --batch 4 --ctx 64 \ - --use-checkpointing diff --git a/examples/gbnf-validator/gbnf-validator.cpp b/examples/gbnf-validator/gbnf-validator.cpp index dd53ba9b1..48a705e15 100644 --- a/examples/gbnf-validator/gbnf-validator.cpp +++ b/examples/gbnf-validator/gbnf-validator.cpp @@ -16,20 +16,25 @@ static bool llama_sample_grammar_string(struct llama_grammar * grammar, const st auto decoded = decode_utf8(input_str, {}); const auto & code_points = decoded.first; + const llama_grammar_rules & rules = llama_grammar_get_rules (grammar); + llama_grammar_stacks & cur_stacks = llama_grammar_get_stacks(grammar); + size_t pos = 0; for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) { - auto prev_stacks = grammar->stacks; - llama_grammar_accept(grammar->rules, prev_stacks, *it, grammar->stacks); - if (grammar->stacks.empty()) { + const llama_grammar_stacks prev_stacks = llama_grammar_get_stacks(grammar); // copy + + llama_grammar_accept(rules, prev_stacks, *it, cur_stacks); + + if (cur_stacks.empty()) { error_pos = pos; error_msg = "Unexpected character '" + unicode_cpt_to_utf8(*it) + "'"; - grammar->stacks = prev_stacks; + cur_stacks = prev_stacks; return false; } ++pos; } - for (const auto & stack : grammar->stacks) { + for (const auto & stack : cur_stacks) { if (stack.empty()) { return true; } diff --git a/examples/gguf-hash/CMakeLists.txt b/examples/gguf-hash/CMakeLists.txt new file mode 100644 index 000000000..633f45535 --- /dev/null +++ b/examples/gguf-hash/CMakeLists.txt @@ -0,0 +1,15 @@ +set(TARGET llama-gguf-hash) +add_executable(${TARGET} gguf-hash.cpp) +install(TARGETS ${TARGET} RUNTIME) + +# clibs dependencies +include_directories(deps/) +add_library(xxhash OBJECT deps/xxhash/xxhash.c deps/xxhash/xxhash.h) +target_link_libraries(${TARGET} PRIVATE xxhash) +add_library(sha1 OBJECT deps/sha1/sha1.c deps/sha1/sha1.h) +target_link_libraries(${TARGET} PRIVATE sha1) +add_library(sha256 OBJECT deps/sha256/sha256.c deps/sha256/sha256.h) +target_link_libraries(${TARGET} PRIVATE sha256) + +target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/gguf-hash/README.md b/examples/gguf-hash/README.md new file mode 100644 index 000000000..9871651e3 --- /dev/null +++ b/examples/gguf-hash/README.md @@ -0,0 +1,206 @@ + +# llama-gguf-hash + +CLI to hash GGUF files to detect difference on a per model and per tensor level. + +**Command line options:** + +- `--help`: display help message +- `--xxh64`: use xhash 64bit hash mode (default) +- `--sha1`: use sha1 +- `--uuid`: use uuid +- `--sha256`: use sha256 +- `--all`: use all hash +- `--no-layer`: exclude per layer hash +- `--uuid`: generate UUIDv5 ID +- `-c`, `--check `: verify against a manifest + +## About + +While most POSIX systems already have hash checking programs like sha256sum, it +is designed to check entire files. This is not ideal for our purpose if we want +to check for consistency of the tensor data even if the metadata content of the +gguf KV store has been updated. + +This program is designed to hash a gguf tensor payload on a 'per tensor layer' +in addition to a 'entire tensor model' hash. The intent is that the entire +tensor layer can be checked first but if there is any detected inconsistencies, +then the per tensor hash can be used to narrow down the specific tensor layer +that has inconsistencies. + +For Maintainers: +- Detection of tensor inconsistency during development and automated tests + - This is served by xxh64 which is fast + - This is also served by having per tensor layer to assist in narrowing down + the location of the faulty tensor layer + - This is also served by sha1 which is much slower but more widely supported + +For Model Creators: +- Optional consistent UUID generation based on model tensor content + - This is served by UUIDv5 which is useful for databases keys + - llama.cpp UUIDv5 Namespace: `ef001206-dadc-5f6d-a15f-3359e577d4e5` + - Made via UUIDv5 URL namespace of `en.wikipedia.org/wiki/Llama.cpp` + +For Model Users: +- Assurance of tensor layer integrity even if metadata was updated + - This is served by sha256 which is still considered very secure as of 2024 + +### Design Note + +- The default behavior of this program if no arguments is provided is to hash + using xxhash's xxh32 mode because it is very fast and is primarily targeted + towards maintainers who may want to use this in automated tests. +- xxhash support xxh32 and xxh128 for 32bit hash and 128bit hash respectively + however we picked 64bit xxhash as most computers are 64bit as of 2024 and thus + would have a better affinity to calculating hash that is 64bit in size. + +## Compile Example + +```bash +cmake -B build -DCMAKE_BUILD_TYPE=Debug -DLLAMA_FATAL_WARNINGS=ON +make -C build clean +make -C build llama-gguf-hash VERBOSE=1 +./build/bin/llama-gguf-hash test.gguf +./build/bin/llama-gguf-hash --xxh64 test.gguf +./build/bin/llama-gguf-hash --sha1 test.gguf +./build/bin/llama-gguf-hash --uuid test.gguf +./build/bin/llama-gguf-hash --sha256 test.gguf +``` + +## Generation and Verification Example + +To generate we may use this command + +```bash +./llama-gguf-hash --all test.gguf > test.gguf.manifest +``` + +Which would generate a manifest that looks like below, which contains multiple hash type and per tensor layer hashes as well +(This excludes UUID as that is an ID not a hash) + +```bash +xxh64 f66e9cd66a4396a0 test.gguf:tensor_0 +sha1 59f79ecefd8125a996fdf419239051a7e99e5f20 test.gguf:tensor_0 +sha256 c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd test.gguf:tensor_0 +xxh64 7d3a1f9ac04d0537 test.gguf:tensor_1 +sha1 4765f592eacf096df4628ba59476af94d767080a test.gguf:tensor_1 +sha256 8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7 test.gguf:tensor_1 +xxh64 a0af5d700049693b test.gguf:tensor_2 +sha1 25cbfbad4513cc348e2c95ebdee69d6ff2fd8753 test.gguf:tensor_2 +sha256 947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180 test.gguf:tensor_2 +xxh64 e83fddf559d7b6a6 test.gguf:tensor_3 +sha1 a9cba73e2d90f2ee3dae2548caa42bef3fe6a96c test.gguf:tensor_3 +sha256 423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1 test.gguf:tensor_3 +xxh64 1257733306b7992d test.gguf:tensor_4 +sha1 d7bc61db93bb685ce9d598da89717c66729b7543 test.gguf:tensor_4 +sha256 79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf test.gguf:tensor_4 +xxh64 d238d16ba4711e58 test.gguf:tensor_5 +sha1 0706566c198fe1072f37e0a5135b4b5f23654c52 test.gguf:tensor_5 +sha256 60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b test.gguf:tensor_5 +xxh64 3fbc3b65ab8c7f39 test.gguf:tensor_6 +sha1 73922a0727226a409049f6fc3172a52219ca6f00 test.gguf:tensor_6 +sha256 574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0 test.gguf:tensor_6 +xxh64 c22021c29854f093 test.gguf:tensor_7 +sha1 efc39cece6a951188fc41e354c73bbfe6813d447 test.gguf:tensor_7 +sha256 4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75 test.gguf:tensor_7 +xxh64 936df61f5d64261f test.gguf:tensor_8 +sha1 c2490296d789a4f34398a337fed8377d943d9f06 test.gguf:tensor_8 +sha256 c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01 test.gguf:tensor_8 +xxh64 93fd20c64421c081 test.gguf:tensor_9 +sha1 7047ce1e78437a6884337a3751c7ee0421918a65 test.gguf:tensor_9 +sha256 23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514 test.gguf:tensor_9 +xxh64 5a54d3aad816f302 test.gguf +sha1 d15be52c4ff213e823cb6dd13af7ee2f978e7042 test.gguf +sha256 7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4 test.gguf +``` + +We can then use the normal check command which will by default check for the highest security strength hash and verify against that: + +```bash +$ ./llama-gguf-hash --check test.gguf.manifest test.gguf +manifest test.gguf.manifest sha256 sha1 xxh64 +sha256 c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd test.gguf:tensor_0 - Ok +sha256 8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7 test.gguf:tensor_1 - Ok +sha256 947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180 test.gguf:tensor_2 - Ok +sha256 423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1 test.gguf:tensor_3 - Ok +sha256 79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf test.gguf:tensor_4 - Ok +sha256 60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b test.gguf:tensor_5 - Ok +sha256 574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0 test.gguf:tensor_6 - Ok +sha256 4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75 test.gguf:tensor_7 - Ok +sha256 c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01 test.gguf:tensor_8 - Ok +sha256 23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514 test.gguf:tensor_9 - Ok +sha256 7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4 test.gguf - Ok + +Verification results for test.gguf.manifest - Success +``` + +Or we may explicitly ask for a faster hash like: + +```bash +$ ./llama-gguf-hash --check test.gguf.manifest --xxh64 test.gguf +manifest test.gguf.manifest sha256 sha1 xxh64 +xxh64 f66e9cd66a4396a0 test.gguf:tensor_0 - Ok +xxh64 7d3a1f9ac04d0537 test.gguf:tensor_1 - Ok +xxh64 a0af5d700049693b test.gguf:tensor_2 - Ok +xxh64 e83fddf559d7b6a6 test.gguf:tensor_3 - Ok +xxh64 1257733306b7992d test.gguf:tensor_4 - Ok +xxh64 d238d16ba4711e58 test.gguf:tensor_5 - Ok +xxh64 3fbc3b65ab8c7f39 test.gguf:tensor_6 - Ok +xxh64 c22021c29854f093 test.gguf:tensor_7 - Ok +xxh64 936df61f5d64261f test.gguf:tensor_8 - Ok +xxh64 93fd20c64421c081 test.gguf:tensor_9 - Ok +xxh64 5a54d3aad816f302 test.gguf - Ok + +Verification results for test.gguf.manifest - Success +``` + +Or maybe we want to just check that all the hash is valid: + +```bash +$./llama-gguf-hash --check test.gguf.manifest --all test.gguf.manifest +manifest test.gguf.manifest sha256 sha1 xxh64 +xxh64 f66e9cd66a4396a0 test.gguf:tensor_0 - Ok +sha1 59f79ecefd8125a996fdf419239051a7e99e5f20 test.gguf:tensor_0 - Ok +sha256 c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd test.gguf:tensor_0 - Ok +xxh64 7d3a1f9ac04d0537 test.gguf:tensor_1 - Ok +sha1 4765f592eacf096df4628ba59476af94d767080a test.gguf:tensor_1 - Ok +sha256 8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7 test.gguf:tensor_1 - Ok +xxh64 a0af5d700049693b test.gguf:tensor_2 - Ok +sha1 25cbfbad4513cc348e2c95ebdee69d6ff2fd8753 test.gguf:tensor_2 - Ok +sha256 947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180 test.gguf:tensor_2 - Ok +xxh64 e83fddf559d7b6a6 test.gguf:tensor_3 - Ok +sha1 a9cba73e2d90f2ee3dae2548caa42bef3fe6a96c test.gguf:tensor_3 - Ok +sha256 423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1 test.gguf:tensor_3 - Ok +xxh64 1257733306b7992d test.gguf:tensor_4 - Ok +sha1 d7bc61db93bb685ce9d598da89717c66729b7543 test.gguf:tensor_4 - Ok +sha256 79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf test.gguf:tensor_4 - Ok +xxh64 d238d16ba4711e58 test.gguf:tensor_5 - Ok +sha1 0706566c198fe1072f37e0a5135b4b5f23654c52 test.gguf:tensor_5 - Ok +sha256 60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b test.gguf:tensor_5 - Ok +xxh64 3fbc3b65ab8c7f39 test.gguf:tensor_6 - Ok +sha1 73922a0727226a409049f6fc3172a52219ca6f00 test.gguf:tensor_6 - Ok +sha256 574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0 test.gguf:tensor_6 - Ok +xxh64 c22021c29854f093 test.gguf:tensor_7 - Ok +sha1 efc39cece6a951188fc41e354c73bbfe6813d447 test.gguf:tensor_7 - Ok +sha256 4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75 test.gguf:tensor_7 - Ok +xxh64 936df61f5d64261f test.gguf:tensor_8 - Ok +sha1 c2490296d789a4f34398a337fed8377d943d9f06 test.gguf:tensor_8 - Ok +sha256 c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01 test.gguf:tensor_8 - Ok +xxh64 93fd20c64421c081 test.gguf:tensor_9 - Ok +sha1 7047ce1e78437a6884337a3751c7ee0421918a65 test.gguf:tensor_9 - Ok +sha256 23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514 test.gguf:tensor_9 - Ok +xxh64 5a54d3aad816f302 test.gguf - Ok +sha1 d15be52c4ff213e823cb6dd13af7ee2f978e7042 test.gguf - Ok +sha256 7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4 test.gguf - Ok + +Verification results for test.gguf.manifest - Success +``` + + +## Crypto/Hash Libraries Used + +These micro c libraries dependencies was installed via the [clib c package manager](https://github.com/clibs) + +- https://github.com/Cyan4973/xxHash +- https://github.com/clibs/sha1/ +- https://github.com/jb55/sha256.c diff --git a/examples/gguf-hash/deps/rotate-bits/package.json b/examples/gguf-hash/deps/rotate-bits/package.json new file mode 100644 index 000000000..74c0bef68 --- /dev/null +++ b/examples/gguf-hash/deps/rotate-bits/package.json @@ -0,0 +1,13 @@ +{ + "name": "rotate-bits", + "version": "0.1.1", + "repo": "jb55/rotate-bits.h", + "description": "rotate bits", + "keywords": ["rotl", "rotr"], + "src": ["rotate-bits.h"], + "license": "Public Domain", + "development": { + "thlorenz/tap.c": "*" + } +} + diff --git a/examples/gguf-hash/deps/rotate-bits/rotate-bits.h b/examples/gguf-hash/deps/rotate-bits/rotate-bits.h new file mode 100644 index 000000000..75c4881fc --- /dev/null +++ b/examples/gguf-hash/deps/rotate-bits/rotate-bits.h @@ -0,0 +1,46 @@ + + +#ifndef __ROTATE_DEFS_H +#define __ROTATE_DEFS_H + +#ifdef _MSC_VER + +#include + +#define ROTL32(v, n) _rotl((v), (n)) +#define ROTL64(v, n) _rotl64((v), (n)) + +#define ROTR32(v, n) _rotr((v), (n)) +#define ROTR64(v, n) _rotr64((v), (n)) + +#else + +#include + +#define U8V(v) ((uint8_t)(v) & 0xFFU) +#define U16V(v) ((uint16_t)(v) & 0xFFFFU) +#define U32V(v) ((uint32_t)(v) & 0xFFFFFFFFU) +#define U64V(v) ((uint64_t)(v) & 0xFFFFFFFFFFFFFFFFU) + +#define ROTL32(v, n) \ + (U32V((uint32_t)(v) << (n)) | ((uint32_t)(v) >> (32 - (n)))) + +// tests fail if we don't have this cast... +#define ROTL64(v, n) \ + (U64V((uint64_t)(v) << (n)) | ((uint64_t)(v) >> (64 - (n)))) + +#define ROTR32(v, n) ROTL32(v, 32 - (n)) +#define ROTR64(v, n) ROTL64(v, 64 - (n)) + +#endif + +#define ROTL8(v, n) \ + (U8V((uint8_t)(v) << (n)) | ((uint8_t)(v) >> (8 - (n)))) + +#define ROTL16(v, n) \ + (U16V((uint16_t)(v) << (n)) | ((uint16_t)(v) >> (16 - (n)))) + +#define ROTR8(v, n) ROTL8(v, 8 - (n)) +#define ROTR16(v, n) ROTL16(v, 16 - (n)) + +#endif diff --git a/examples/gguf-hash/deps/sha1/package.json b/examples/gguf-hash/deps/sha1/package.json new file mode 100644 index 000000000..6a5843dd1 --- /dev/null +++ b/examples/gguf-hash/deps/sha1/package.json @@ -0,0 +1,9 @@ +{ + "name": "sha1", + "version": "0.0.1", + "repo": "clibs/sha1", + "description": "sha1 hash algorithm", + "keywords": ["sha1", "hash"], + "license": "public domain", + "src": ["sha1.c", "sha1.h"] +} diff --git a/examples/gguf-hash/deps/sha1/sha1.c b/examples/gguf-hash/deps/sha1/sha1.c new file mode 100644 index 000000000..76cd6ca33 --- /dev/null +++ b/examples/gguf-hash/deps/sha1/sha1.c @@ -0,0 +1,295 @@ +/* +SHA-1 in C +By Steve Reid +100% Public Domain + +Test Vectors (from FIPS PUB 180-1) +"abc" + A9993E36 4706816A BA3E2571 7850C26C 9CD0D89D +"abcdbcdecdefdefgefghfghighijhijkijkljklmklmnlmnomnopnopq" + 84983E44 1C3BD26E BAAE4AA1 F95129E5 E54670F1 +A million repetitions of "a" + 34AA973C D4C4DAA4 F61EEB2B DBAD2731 6534016F +*/ + +/* #define LITTLE_ENDIAN * This should be #define'd already, if true. */ +/* #define SHA1HANDSOFF * Copies data before messing with it. */ + +#define SHA1HANDSOFF + +#include +#include + +/* for uint32_t */ +#include + +#include "sha1.h" + + +#define rol(value, bits) (((value) << (bits)) | ((value) >> (32 - (bits)))) + +/* blk0() and blk() perform the initial expand. */ +/* I got the idea of expanding during the round function from SSLeay */ +#if BYTE_ORDER == LITTLE_ENDIAN +#define blk0(i) (block->l[i] = (rol(block->l[i],24)&0xFF00FF00) \ + |(rol(block->l[i],8)&0x00FF00FF)) +#elif BYTE_ORDER == BIG_ENDIAN +#define blk0(i) block->l[i] +#else +#error "Endianness not defined!" +#endif +#define blk(i) (block->l[i&15] = rol(block->l[(i+13)&15]^block->l[(i+8)&15] \ + ^block->l[(i+2)&15]^block->l[i&15],1)) + +/* (R0+R1), R2, R3, R4 are the different operations used in SHA1 */ +#define R0(v,w,x,y,z,i) z+=((w&(x^y))^y)+blk0(i)+0x5A827999+rol(v,5);w=rol(w,30); +#define R1(v,w,x,y,z,i) z+=((w&(x^y))^y)+blk(i)+0x5A827999+rol(v,5);w=rol(w,30); +#define R2(v,w,x,y,z,i) z+=(w^x^y)+blk(i)+0x6ED9EBA1+rol(v,5);w=rol(w,30); +#define R3(v,w,x,y,z,i) z+=(((w|x)&y)|(w&x))+blk(i)+0x8F1BBCDC+rol(v,5);w=rol(w,30); +#define R4(v,w,x,y,z,i) z+=(w^x^y)+blk(i)+0xCA62C1D6+rol(v,5);w=rol(w,30); + + +/* Hash a single 512-bit block. This is the core of the algorithm. */ + +void SHA1Transform( + uint32_t state[5], + const unsigned char buffer[64] +) +{ + uint32_t a, b, c, d, e; + + typedef union + { + unsigned char c[64]; + uint32_t l[16]; + } CHAR64LONG16; + +#ifdef SHA1HANDSOFF + CHAR64LONG16 block[1]; /* use array to appear as a pointer */ + + memcpy(block, buffer, 64); +#else + /* The following had better never be used because it causes the + * pointer-to-const buffer to be cast into a pointer to non-const. + * And the result is written through. I threw a "const" in, hoping + * this will cause a diagnostic. + */ + CHAR64LONG16 *block = (const CHAR64LONG16 *) buffer; +#endif + /* Copy context->state[] to working vars */ + a = state[0]; + b = state[1]; + c = state[2]; + d = state[3]; + e = state[4]; + /* 4 rounds of 20 operations each. Loop unrolled. */ + R0(a, b, c, d, e, 0); + R0(e, a, b, c, d, 1); + R0(d, e, a, b, c, 2); + R0(c, d, e, a, b, 3); + R0(b, c, d, e, a, 4); + R0(a, b, c, d, e, 5); + R0(e, a, b, c, d, 6); + R0(d, e, a, b, c, 7); + R0(c, d, e, a, b, 8); + R0(b, c, d, e, a, 9); + R0(a, b, c, d, e, 10); + R0(e, a, b, c, d, 11); + R0(d, e, a, b, c, 12); + R0(c, d, e, a, b, 13); + R0(b, c, d, e, a, 14); + R0(a, b, c, d, e, 15); + R1(e, a, b, c, d, 16); + R1(d, e, a, b, c, 17); + R1(c, d, e, a, b, 18); + R1(b, c, d, e, a, 19); + R2(a, b, c, d, e, 20); + R2(e, a, b, c, d, 21); + R2(d, e, a, b, c, 22); + R2(c, d, e, a, b, 23); + R2(b, c, d, e, a, 24); + R2(a, b, c, d, e, 25); + R2(e, a, b, c, d, 26); + R2(d, e, a, b, c, 27); + R2(c, d, e, a, b, 28); + R2(b, c, d, e, a, 29); + R2(a, b, c, d, e, 30); + R2(e, a, b, c, d, 31); + R2(d, e, a, b, c, 32); + R2(c, d, e, a, b, 33); + R2(b, c, d, e, a, 34); + R2(a, b, c, d, e, 35); + R2(e, a, b, c, d, 36); + R2(d, e, a, b, c, 37); + R2(c, d, e, a, b, 38); + R2(b, c, d, e, a, 39); + R3(a, b, c, d, e, 40); + R3(e, a, b, c, d, 41); + R3(d, e, a, b, c, 42); + R3(c, d, e, a, b, 43); + R3(b, c, d, e, a, 44); + R3(a, b, c, d, e, 45); + R3(e, a, b, c, d, 46); + R3(d, e, a, b, c, 47); + R3(c, d, e, a, b, 48); + R3(b, c, d, e, a, 49); + R3(a, b, c, d, e, 50); + R3(e, a, b, c, d, 51); + R3(d, e, a, b, c, 52); + R3(c, d, e, a, b, 53); + R3(b, c, d, e, a, 54); + R3(a, b, c, d, e, 55); + R3(e, a, b, c, d, 56); + R3(d, e, a, b, c, 57); + R3(c, d, e, a, b, 58); + R3(b, c, d, e, a, 59); + R4(a, b, c, d, e, 60); + R4(e, a, b, c, d, 61); + R4(d, e, a, b, c, 62); + R4(c, d, e, a, b, 63); + R4(b, c, d, e, a, 64); + R4(a, b, c, d, e, 65); + R4(e, a, b, c, d, 66); + R4(d, e, a, b, c, 67); + R4(c, d, e, a, b, 68); + R4(b, c, d, e, a, 69); + R4(a, b, c, d, e, 70); + R4(e, a, b, c, d, 71); + R4(d, e, a, b, c, 72); + R4(c, d, e, a, b, 73); + R4(b, c, d, e, a, 74); + R4(a, b, c, d, e, 75); + R4(e, a, b, c, d, 76); + R4(d, e, a, b, c, 77); + R4(c, d, e, a, b, 78); + R4(b, c, d, e, a, 79); + /* Add the working vars back into context.state[] */ + state[0] += a; + state[1] += b; + state[2] += c; + state[3] += d; + state[4] += e; + /* Wipe variables */ + a = b = c = d = e = 0; +#ifdef SHA1HANDSOFF + memset(block, '\0', sizeof(block)); +#endif +} + + +/* SHA1Init - Initialize new context */ + +void SHA1Init( + SHA1_CTX * context +) +{ + /* SHA1 initialization constants */ + context->state[0] = 0x67452301; + context->state[1] = 0xEFCDAB89; + context->state[2] = 0x98BADCFE; + context->state[3] = 0x10325476; + context->state[4] = 0xC3D2E1F0; + context->count[0] = context->count[1] = 0; +} + + +/* Run your data through this. */ + +void SHA1Update( + SHA1_CTX * context, + const unsigned char *data, + uint32_t len +) +{ + uint32_t i; + + uint32_t j; + + j = context->count[0]; + if ((context->count[0] += len << 3) < j) + context->count[1]++; + context->count[1] += (len >> 29); + j = (j >> 3) & 63; + if ((j + len) > 63) + { + memcpy(&context->buffer[j], data, (i = 64 - j)); + SHA1Transform(context->state, context->buffer); + for (; i + 63 < len; i += 64) + { + SHA1Transform(context->state, &data[i]); + } + j = 0; + } + else + i = 0; + memcpy(&context->buffer[j], &data[i], len - i); +} + + +/* Add padding and return the message digest. */ + +void SHA1Final( + unsigned char digest[20], + SHA1_CTX * context +) +{ + unsigned i; + + unsigned char finalcount[8]; + + unsigned char c; + +#if 0 /* untested "improvement" by DHR */ + /* Convert context->count to a sequence of bytes + * in finalcount. Second element first, but + * big-endian order within element. + * But we do it all backwards. + */ + unsigned char *fcp = &finalcount[8]; + + for (i = 0; i < 2; i++) + { + uint32_t t = context->count[i]; + + int j; + + for (j = 0; j < 4; t >>= 8, j++) + *--fcp = (unsigned char) t} +#else + for (i = 0; i < 8; i++) + { + finalcount[i] = (unsigned char) ((context->count[(i >= 4 ? 0 : 1)] >> ((3 - (i & 3)) * 8)) & 255); /* Endian independent */ + } +#endif + c = 0200; + SHA1Update(context, &c, 1); + while ((context->count[0] & 504) != 448) + { + c = 0000; + SHA1Update(context, &c, 1); + } + SHA1Update(context, finalcount, 8); /* Should cause a SHA1Transform() */ + for (i = 0; i < 20; i++) + { + digest[i] = (unsigned char) + ((context->state[i >> 2] >> ((3 - (i & 3)) * 8)) & 255); + } + /* Wipe variables */ + memset(context, '\0', sizeof(*context)); + memset(&finalcount, '\0', sizeof(finalcount)); +} + +void SHA1( + char *hash_out, + const char *str, + uint32_t len) +{ + SHA1_CTX ctx; + unsigned int ii; + + SHA1Init(&ctx); + for (ii=0; ii + 100% Public Domain + */ + +#include "stdint.h" + +#if defined(__cplusplus) +extern "C" { +#endif + +typedef struct +{ + uint32_t state[5]; + uint32_t count[2]; + unsigned char buffer[64]; +} SHA1_CTX; + +void SHA1Transform( + uint32_t state[5], + const unsigned char buffer[64] + ); + +void SHA1Init( + SHA1_CTX * context + ); + +void SHA1Update( + SHA1_CTX * context, + const unsigned char *data, + uint32_t len + ); + +void SHA1Final( + unsigned char digest[20], + SHA1_CTX * context + ); + +void SHA1( + char *hash_out, + const char *str, + uint32_t len); + +#if defined(__cplusplus) +} +#endif + +#endif /* SHA1_H */ diff --git a/examples/gguf-hash/deps/sha256/package.json b/examples/gguf-hash/deps/sha256/package.json new file mode 100644 index 000000000..b92a04127 --- /dev/null +++ b/examples/gguf-hash/deps/sha256/package.json @@ -0,0 +1,15 @@ +{ + "name": "sha256", + "version": "0.0.2", + "repo": "jb55/sha256.c", + "description": "sha256 in c", + "keywords": ["sha256", "sha2"], + "src": ["sha256.c", "sha256.h"], + "dependencies": { + "jb55/rotate-bits.h": "0.1.1" + }, + "development": { + "thlorenz/tap.c": "*" + } +} + diff --git a/examples/gguf-hash/deps/sha256/sha256.c b/examples/gguf-hash/deps/sha256/sha256.c new file mode 100644 index 000000000..a7a87aeb2 --- /dev/null +++ b/examples/gguf-hash/deps/sha256/sha256.c @@ -0,0 +1,221 @@ +/* Crypto/Sha256.c -- SHA-256 Hash +2010-06-11 : Igor Pavlov : Public domain +This code is based on public domain code from Wei Dai's Crypto++ library. */ + +#include "rotate-bits/rotate-bits.h" +#include "sha256.h" + +/* define it for speed optimization */ +#define _SHA256_UNROLL +#define _SHA256_UNROLL2 + +void +sha256_init(sha256_t *p) +{ + p->state[0] = 0x6a09e667; + p->state[1] = 0xbb67ae85; + p->state[2] = 0x3c6ef372; + p->state[3] = 0xa54ff53a; + p->state[4] = 0x510e527f; + p->state[5] = 0x9b05688c; + p->state[6] = 0x1f83d9ab; + p->state[7] = 0x5be0cd19; + p->count = 0; +} + +#define S0(x) (ROTR32(x, 2) ^ ROTR32(x,13) ^ ROTR32(x, 22)) +#define S1(x) (ROTR32(x, 6) ^ ROTR32(x,11) ^ ROTR32(x, 25)) +#define s0(x) (ROTR32(x, 7) ^ ROTR32(x,18) ^ (x >> 3)) +#define s1(x) (ROTR32(x,17) ^ ROTR32(x,19) ^ (x >> 10)) + +#define blk0(i) (W[i] = data[i]) +#define blk2(i) (W[i&15] += s1(W[(i-2)&15]) + W[(i-7)&15] + s0(W[(i-15)&15])) + +#define Ch(x,y,z) (z^(x&(y^z))) +#define Maj(x,y,z) ((x&y)|(z&(x|y))) + +#define a(i) T[(0-(i))&7] +#define b(i) T[(1-(i))&7] +#define c(i) T[(2-(i))&7] +#define d(i) T[(3-(i))&7] +#define e(i) T[(4-(i))&7] +#define f(i) T[(5-(i))&7] +#define g(i) T[(6-(i))&7] +#define h(i) T[(7-(i))&7] + + +#ifdef _SHA256_UNROLL2 + +#define R(a,b,c,d,e,f,g,h, i) h += S1(e) + Ch(e,f,g) + K[i+j] + (j?blk2(i):blk0(i));\ + d += h; h += S0(a) + Maj(a, b, c) + +#define RX_8(i) \ + R(a,b,c,d,e,f,g,h, i); \ + R(h,a,b,c,d,e,f,g, (i+1)); \ + R(g,h,a,b,c,d,e,f, (i+2)); \ + R(f,g,h,a,b,c,d,e, (i+3)); \ + R(e,f,g,h,a,b,c,d, (i+4)); \ + R(d,e,f,g,h,a,b,c, (i+5)); \ + R(c,d,e,f,g,h,a,b, (i+6)); \ + R(b,c,d,e,f,g,h,a, (i+7)) + +#else + +#define R(i) h(i) += S1(e(i)) + Ch(e(i),f(i),g(i)) + K[i+j] + (j?blk2(i):blk0(i));\ + d(i) += h(i); h(i) += S0(a(i)) + Maj(a(i), b(i), c(i)) + +#ifdef _SHA256_UNROLL + +#define RX_8(i) R(i+0); R(i+1); R(i+2); R(i+3); R(i+4); R(i+5); R(i+6); R(i+7); + +#endif + +#endif + +static const uint32_t K[64] = { + 0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5, + 0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5, + 0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3, + 0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174, + 0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc, + 0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da, + 0x983e5152, 0xa831c66d, 0xb00327c8, 0xbf597fc7, + 0xc6e00bf3, 0xd5a79147, 0x06ca6351, 0x14292967, + 0x27b70a85, 0x2e1b2138, 0x4d2c6dfc, 0x53380d13, + 0x650a7354, 0x766a0abb, 0x81c2c92e, 0x92722c85, + 0xa2bfe8a1, 0xa81a664b, 0xc24b8b70, 0xc76c51a3, + 0xd192e819, 0xd6990624, 0xf40e3585, 0x106aa070, + 0x19a4c116, 0x1e376c08, 0x2748774c, 0x34b0bcb5, + 0x391c0cb3, 0x4ed8aa4a, 0x5b9cca4f, 0x682e6ff3, + 0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208, + 0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2 +}; + +static void +sha256_transform(uint32_t *state, const uint32_t *data) +{ + uint32_t W[16] = {0}; + unsigned j; + #ifdef _SHA256_UNROLL2 + uint32_t a,b,c,d,e,f,g,h; + a = state[0]; + b = state[1]; + c = state[2]; + d = state[3]; + e = state[4]; + f = state[5]; + g = state[6]; + h = state[7]; + #else + uint32_t T[8]; + for (j = 0; j < 8; j++) + T[j] = state[j]; + #endif + + for (j = 0; j < 64; j += 16) + { + #if defined(_SHA256_UNROLL) || defined(_SHA256_UNROLL2) + RX_8(0); RX_8(8); + #else + unsigned i; + for (i = 0; i < 16; i++) { R(i); } + #endif + } + + #ifdef _SHA256_UNROLL2 + state[0] += a; + state[1] += b; + state[2] += c; + state[3] += d; + state[4] += e; + state[5] += f; + state[6] += g; + state[7] += h; + #else + for (j = 0; j < 8; j++) + state[j] += T[j]; + #endif + + /* Wipe variables */ + /* memset(W, 0, sizeof(W)); */ + /* memset(T, 0, sizeof(T)); */ +} + +#undef S0 +#undef S1 +#undef s0 +#undef s1 + +static void +sha256_write_byte_block(sha256_t *p) +{ + uint32_t data32[16]; + unsigned i; + for (i = 0; i < 16; i++) + data32[i] = + ((uint32_t)(p->buffer[i * 4 ]) << 24) + + ((uint32_t)(p->buffer[i * 4 + 1]) << 16) + + ((uint32_t)(p->buffer[i * 4 + 2]) << 8) + + ((uint32_t)(p->buffer[i * 4 + 3])); + sha256_transform(p->state, data32); +} + + +void +sha256_hash(unsigned char *buf, const unsigned char *data, size_t size) +{ + sha256_t hash; + sha256_init(&hash); + sha256_update(&hash, data, size); + sha256_final(&hash, buf); +} + + +void +sha256_update(sha256_t *p, const unsigned char *data, size_t size) +{ + uint32_t curBufferPos = (uint32_t)p->count & 0x3F; + while (size > 0) + { + p->buffer[curBufferPos++] = *data++; + p->count++; + size--; + if (curBufferPos == 64) + { + curBufferPos = 0; + sha256_write_byte_block(p); + } + } +} + + +void +sha256_final(sha256_t *p, unsigned char *digest) +{ + uint64_t lenInBits = (p->count << 3); + uint32_t curBufferPos = (uint32_t)p->count & 0x3F; + unsigned i; + p->buffer[curBufferPos++] = 0x80; + while (curBufferPos != (64 - 8)) + { + curBufferPos &= 0x3F; + if (curBufferPos == 0) + sha256_write_byte_block(p); + p->buffer[curBufferPos++] = 0; + } + for (i = 0; i < 8; i++) + { + p->buffer[curBufferPos++] = (unsigned char)(lenInBits >> 56); + lenInBits <<= 8; + } + sha256_write_byte_block(p); + + for (i = 0; i < 8; i++) + { + *digest++ = (unsigned char)(p->state[i] >> 24); + *digest++ = (unsigned char)(p->state[i] >> 16); + *digest++ = (unsigned char)(p->state[i] >> 8); + *digest++ = (unsigned char)(p->state[i]); + } + sha256_init(p); +} diff --git a/examples/gguf-hash/deps/sha256/sha256.h b/examples/gguf-hash/deps/sha256/sha256.h new file mode 100644 index 000000000..21657e66b --- /dev/null +++ b/examples/gguf-hash/deps/sha256/sha256.h @@ -0,0 +1,24 @@ +/* Sha256.h -- SHA-256 Hash +2010-06-11 : Igor Pavlov : Public domain */ + +#ifndef __CRYPTO_SHA256_H +#define __CRYPTO_SHA256_H + +#include +#include + +#define SHA256_DIGEST_SIZE 32 + +typedef struct sha256_t +{ + uint32_t state[8]; + uint64_t count; + unsigned char buffer[64]; +} sha256_t; + +void sha256_init(sha256_t *p); +void sha256_update(sha256_t *p, const unsigned char *data, size_t size); +void sha256_final(sha256_t *p, unsigned char *digest); +void sha256_hash(unsigned char *buf, const unsigned char *data, size_t size); + +#endif diff --git a/examples/gguf-hash/deps/xxhash/clib.json b/examples/gguf-hash/deps/xxhash/clib.json new file mode 100644 index 000000000..242343c5d --- /dev/null +++ b/examples/gguf-hash/deps/xxhash/clib.json @@ -0,0 +1,12 @@ +{ + "name": "xxhash", + "version": "0.8.2", + "repo": "Cyan4973/xxhash", + "description": "Extremely fast non-cryptographic hash algorithm", + "keywords": ["xxhash", "hashing"], + "license": "BSD-2-Clause", + "src": [ + "xxhash.c", + "xxhash.h" + ] +} diff --git a/examples/gguf-hash/deps/xxhash/xxhash.c b/examples/gguf-hash/deps/xxhash/xxhash.c new file mode 100644 index 000000000..e60cc37f1 --- /dev/null +++ b/examples/gguf-hash/deps/xxhash/xxhash.c @@ -0,0 +1,42 @@ +/* + * xxHash - Extremely Fast Hash algorithm + * Copyright (C) 2012-2023 Yann Collet + * + * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php) + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are + * met: + * + * * Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * * Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following disclaimer + * in the documentation and/or other materials provided with the + * distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + * You can contact the author at: + * - xxHash homepage: https://www.xxhash.com + * - xxHash source repository: https://github.com/Cyan4973/xxHash + */ + +/* + * xxhash.c instantiates functions defined in xxhash.h + */ + +#define XXH_STATIC_LINKING_ONLY /* access advanced declarations */ +#define XXH_IMPLEMENTATION /* access definitions */ + +#include "xxhash.h" diff --git a/examples/gguf-hash/deps/xxhash/xxhash.h b/examples/gguf-hash/deps/xxhash/xxhash.h new file mode 100644 index 000000000..c0fafe20d --- /dev/null +++ b/examples/gguf-hash/deps/xxhash/xxhash.h @@ -0,0 +1,7093 @@ +/* + * xxHash - Extremely Fast Hash algorithm + * Header File + * Copyright (C) 2012-2023 Yann Collet + * + * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php) + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are + * met: + * + * * Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * * Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following disclaimer + * in the documentation and/or other materials provided with the + * distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + * You can contact the author at: + * - xxHash homepage: https://www.xxhash.com + * - xxHash source repository: https://github.com/Cyan4973/xxHash + */ + +/*! + * @mainpage xxHash + * + * xxHash is an extremely fast non-cryptographic hash algorithm, working at RAM speed + * limits. + * + * It is proposed in four flavors, in three families: + * 1. @ref XXH32_family + * - Classic 32-bit hash function. Simple, compact, and runs on almost all + * 32-bit and 64-bit systems. + * 2. @ref XXH64_family + * - Classic 64-bit adaptation of XXH32. Just as simple, and runs well on most + * 64-bit systems (but _not_ 32-bit systems). + * 3. @ref XXH3_family + * - Modern 64-bit and 128-bit hash function family which features improved + * strength and performance across the board, especially on smaller data. + * It benefits greatly from SIMD and 64-bit without requiring it. + * + * Benchmarks + * --- + * The reference system uses an Intel i7-9700K CPU, and runs Ubuntu x64 20.04. + * The open source benchmark program is compiled with clang v10.0 using -O3 flag. + * + * | Hash Name | ISA ext | Width | Large Data Speed | Small Data Velocity | + * | -------------------- | ------- | ----: | ---------------: | ------------------: | + * | XXH3_64bits() | @b AVX2 | 64 | 59.4 GB/s | 133.1 | + * | MeowHash | AES-NI | 128 | 58.2 GB/s | 52.5 | + * | XXH3_128bits() | @b AVX2 | 128 | 57.9 GB/s | 118.1 | + * | CLHash | PCLMUL | 64 | 37.1 GB/s | 58.1 | + * | XXH3_64bits() | @b SSE2 | 64 | 31.5 GB/s | 133.1 | + * | XXH3_128bits() | @b SSE2 | 128 | 29.6 GB/s | 118.1 | + * | RAM sequential read | | N/A | 28.0 GB/s | N/A | + * | ahash | AES-NI | 64 | 22.5 GB/s | 107.2 | + * | City64 | | 64 | 22.0 GB/s | 76.6 | + * | T1ha2 | | 64 | 22.0 GB/s | 99.0 | + * | City128 | | 128 | 21.7 GB/s | 57.7 | + * | FarmHash | AES-NI | 64 | 21.3 GB/s | 71.9 | + * | XXH64() | | 64 | 19.4 GB/s | 71.0 | + * | SpookyHash | | 64 | 19.3 GB/s | 53.2 | + * | Mum | | 64 | 18.0 GB/s | 67.0 | + * | CRC32C | SSE4.2 | 32 | 13.0 GB/s | 57.9 | + * | XXH32() | | 32 | 9.7 GB/s | 71.9 | + * | City32 | | 32 | 9.1 GB/s | 66.0 | + * | Blake3* | @b AVX2 | 256 | 4.4 GB/s | 8.1 | + * | Murmur3 | | 32 | 3.9 GB/s | 56.1 | + * | SipHash* | | 64 | 3.0 GB/s | 43.2 | + * | Blake3* | @b SSE2 | 256 | 2.4 GB/s | 8.1 | + * | HighwayHash | | 64 | 1.4 GB/s | 6.0 | + * | FNV64 | | 64 | 1.2 GB/s | 62.7 | + * | Blake2* | | 256 | 1.1 GB/s | 5.1 | + * | SHA1* | | 160 | 0.8 GB/s | 5.6 | + * | MD5* | | 128 | 0.6 GB/s | 7.8 | + * @note + * - Hashes which require a specific ISA extension are noted. SSE2 is also noted, + * even though it is mandatory on x64. + * - Hashes with an asterisk are cryptographic. Note that MD5 is non-cryptographic + * by modern standards. + * - Small data velocity is a rough average of algorithm's efficiency for small + * data. For more accurate information, see the wiki. + * - More benchmarks and strength tests are found on the wiki: + * https://github.com/Cyan4973/xxHash/wiki + * + * Usage + * ------ + * All xxHash variants use a similar API. Changing the algorithm is a trivial + * substitution. + * + * @pre + * For functions which take an input and length parameter, the following + * requirements are assumed: + * - The range from [`input`, `input + length`) is valid, readable memory. + * - The only exception is if the `length` is `0`, `input` may be `NULL`. + * - For C++, the objects must have the *TriviallyCopyable* property, as the + * functions access bytes directly as if it was an array of `unsigned char`. + * + * @anchor single_shot_example + * **Single Shot** + * + * These functions are stateless functions which hash a contiguous block of memory, + * immediately returning the result. They are the easiest and usually the fastest + * option. + * + * XXH32(), XXH64(), XXH3_64bits(), XXH3_128bits() + * + * @code{.c} + * #include + * #include "xxhash.h" + * + * // Example for a function which hashes a null terminated string with XXH32(). + * XXH32_hash_t hash_string(const char* string, XXH32_hash_t seed) + * { + * // NULL pointers are only valid if the length is zero + * size_t length = (string == NULL) ? 0 : strlen(string); + * return XXH32(string, length, seed); + * } + * @endcode + * + * + * @anchor streaming_example + * **Streaming** + * + * These groups of functions allow incremental hashing of unknown size, even + * more than what would fit in a size_t. + * + * XXH32_reset(), XXH64_reset(), XXH3_64bits_reset(), XXH3_128bits_reset() + * + * @code{.c} + * #include + * #include + * #include "xxhash.h" + * // Example for a function which hashes a FILE incrementally with XXH3_64bits(). + * XXH64_hash_t hashFile(FILE* f) + * { + * // Allocate a state struct. Do not just use malloc() or new. + * XXH3_state_t* state = XXH3_createState(); + * assert(state != NULL && "Out of memory!"); + * // Reset the state to start a new hashing session. + * XXH3_64bits_reset(state); + * char buffer[4096]; + * size_t count; + * // Read the file in chunks + * while ((count = fread(buffer, 1, sizeof(buffer), f)) != 0) { + * // Run update() as many times as necessary to process the data + * XXH3_64bits_update(state, buffer, count); + * } + * // Retrieve the finalized hash. This will not change the state. + * XXH64_hash_t result = XXH3_64bits_digest(state); + * // Free the state. Do not use free(). + * XXH3_freeState(state); + * return result; + * } + * @endcode + * + * Streaming functions generate the xxHash value from an incremental input. + * This method is slower than single-call functions, due to state management. + * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized. + * + * An XXH state must first be allocated using `XXH*_createState()`. + * + * Start a new hash by initializing the state with a seed using `XXH*_reset()`. + * + * Then, feed the hash state by calling `XXH*_update()` as many times as necessary. + * + * The function returns an error code, with 0 meaning OK, and any other value + * meaning there is an error. + * + * Finally, a hash value can be produced anytime, by using `XXH*_digest()`. + * This function returns the nn-bits hash as an int or long long. + * + * It's still possible to continue inserting input into the hash state after a + * digest, and generate new hash values later on by invoking `XXH*_digest()`. + * + * When done, release the state using `XXH*_freeState()`. + * + * + * @anchor canonical_representation_example + * **Canonical Representation** + * + * The default return values from XXH functions are unsigned 32, 64 and 128 bit + * integers. + * This the simplest and fastest format for further post-processing. + * + * However, this leaves open the question of what is the order on the byte level, + * since little and big endian conventions will store the same number differently. + * + * The canonical representation settles this issue by mandating big-endian + * convention, the same convention as human-readable numbers (large digits first). + * + * When writing hash values to storage, sending them over a network, or printing + * them, it's highly recommended to use the canonical representation to ensure + * portability across a wider range of systems, present and future. + * + * The following functions allow transformation of hash values to and from + * canonical format. + * + * XXH32_canonicalFromHash(), XXH32_hashFromCanonical(), + * XXH64_canonicalFromHash(), XXH64_hashFromCanonical(), + * XXH128_canonicalFromHash(), XXH128_hashFromCanonical(), + * + * @code{.c} + * #include + * #include "xxhash.h" + * + * // Example for a function which prints XXH32_hash_t in human readable format + * void printXxh32(XXH32_hash_t hash) + * { + * XXH32_canonical_t cano; + * XXH32_canonicalFromHash(&cano, hash); + * size_t i; + * for(i = 0; i < sizeof(cano.digest); ++i) { + * printf("%02x", cano.digest[i]); + * } + * printf("\n"); + * } + * + * // Example for a function which converts XXH32_canonical_t to XXH32_hash_t + * XXH32_hash_t convertCanonicalToXxh32(XXH32_canonical_t cano) + * { + * XXH32_hash_t hash = XXH32_hashFromCanonical(&cano); + * return hash; + * } + * @endcode + * + * + * @file xxhash.h + * xxHash prototypes and implementation + */ + +#if defined (__cplusplus) +extern "C" { +#endif + +/* **************************** + * INLINE mode + ******************************/ +/*! + * @defgroup public Public API + * Contains details on the public xxHash functions. + * @{ + */ +#ifdef XXH_DOXYGEN +/*! + * @brief Gives access to internal state declaration, required for static allocation. + * + * Incompatible with dynamic linking, due to risks of ABI changes. + * + * Usage: + * @code{.c} + * #define XXH_STATIC_LINKING_ONLY + * #include "xxhash.h" + * @endcode + */ +# define XXH_STATIC_LINKING_ONLY +/* Do not undef XXH_STATIC_LINKING_ONLY for Doxygen */ + +/*! + * @brief Gives access to internal definitions. + * + * Usage: + * @code{.c} + * #define XXH_STATIC_LINKING_ONLY + * #define XXH_IMPLEMENTATION + * #include "xxhash.h" + * @endcode + */ +# define XXH_IMPLEMENTATION +/* Do not undef XXH_IMPLEMENTATION for Doxygen */ + +/*! + * @brief Exposes the implementation and marks all functions as `inline`. + * + * Use these build macros to inline xxhash into the target unit. + * Inlining improves performance on small inputs, especially when the length is + * expressed as a compile-time constant: + * + * https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html + * + * It also keeps xxHash symbols private to the unit, so they are not exported. + * + * Usage: + * @code{.c} + * #define XXH_INLINE_ALL + * #include "xxhash.h" + * @endcode + * Do not compile and link xxhash.o as a separate object, as it is not useful. + */ +# define XXH_INLINE_ALL +# undef XXH_INLINE_ALL +/*! + * @brief Exposes the implementation without marking functions as inline. + */ +# define XXH_PRIVATE_API +# undef XXH_PRIVATE_API +/*! + * @brief Emulate a namespace by transparently prefixing all symbols. + * + * If you want to include _and expose_ xxHash functions from within your own + * library, but also want to avoid symbol collisions with other libraries which + * may also include xxHash, you can use @ref XXH_NAMESPACE to automatically prefix + * any public symbol from xxhash library with the value of @ref XXH_NAMESPACE + * (therefore, avoid empty or numeric values). + * + * Note that no change is required within the calling program as long as it + * includes `xxhash.h`: Regular symbol names will be automatically translated + * by this header. + */ +# define XXH_NAMESPACE /* YOUR NAME HERE */ +# undef XXH_NAMESPACE +#endif + +#if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \ + && !defined(XXH_INLINE_ALL_31684351384) + /* this section should be traversed only once */ +# define XXH_INLINE_ALL_31684351384 + /* give access to the advanced API, required to compile implementations */ +# undef XXH_STATIC_LINKING_ONLY /* avoid macro redef */ +# define XXH_STATIC_LINKING_ONLY + /* make all functions private */ +# undef XXH_PUBLIC_API +# if defined(__GNUC__) +# define XXH_PUBLIC_API static __inline __attribute__((__unused__)) +# elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) +# define XXH_PUBLIC_API static inline +# elif defined(_MSC_VER) +# define XXH_PUBLIC_API static __inline +# else + /* note: this version may generate warnings for unused static functions */ +# define XXH_PUBLIC_API static +# endif + + /* + * This part deals with the special case where a unit wants to inline xxHash, + * but "xxhash.h" has previously been included without XXH_INLINE_ALL, + * such as part of some previously included *.h header file. + * Without further action, the new include would just be ignored, + * and functions would effectively _not_ be inlined (silent failure). + * The following macros solve this situation by prefixing all inlined names, + * avoiding naming collision with previous inclusions. + */ + /* Before that, we unconditionally #undef all symbols, + * in case they were already defined with XXH_NAMESPACE. + * They will then be redefined for XXH_INLINE_ALL + */ +# undef XXH_versionNumber + /* XXH32 */ +# undef XXH32 +# undef XXH32_createState +# undef XXH32_freeState +# undef XXH32_reset +# undef XXH32_update +# undef XXH32_digest +# undef XXH32_copyState +# undef XXH32_canonicalFromHash +# undef XXH32_hashFromCanonical + /* XXH64 */ +# undef XXH64 +# undef XXH64_createState +# undef XXH64_freeState +# undef XXH64_reset +# undef XXH64_update +# undef XXH64_digest +# undef XXH64_copyState +# undef XXH64_canonicalFromHash +# undef XXH64_hashFromCanonical + /* XXH3_64bits */ +# undef XXH3_64bits +# undef XXH3_64bits_withSecret +# undef XXH3_64bits_withSeed +# undef XXH3_64bits_withSecretandSeed +# undef XXH3_createState +# undef XXH3_freeState +# undef XXH3_copyState +# undef XXH3_64bits_reset +# undef XXH3_64bits_reset_withSeed +# undef XXH3_64bits_reset_withSecret +# undef XXH3_64bits_update +# undef XXH3_64bits_digest +# undef XXH3_generateSecret + /* XXH3_128bits */ +# undef XXH128 +# undef XXH3_128bits +# undef XXH3_128bits_withSeed +# undef XXH3_128bits_withSecret +# undef XXH3_128bits_reset +# undef XXH3_128bits_reset_withSeed +# undef XXH3_128bits_reset_withSecret +# undef XXH3_128bits_reset_withSecretandSeed +# undef XXH3_128bits_update +# undef XXH3_128bits_digest +# undef XXH128_isEqual +# undef XXH128_cmp +# undef XXH128_canonicalFromHash +# undef XXH128_hashFromCanonical + /* Finally, free the namespace itself */ +# undef XXH_NAMESPACE + + /* employ the namespace for XXH_INLINE_ALL */ +# define XXH_NAMESPACE XXH_INLINE_ + /* + * Some identifiers (enums, type names) are not symbols, + * but they must nonetheless be renamed to avoid redeclaration. + * Alternative solution: do not redeclare them. + * However, this requires some #ifdefs, and has a more dispersed impact. + * Meanwhile, renaming can be achieved in a single place. + */ +# define XXH_IPREF(Id) XXH_NAMESPACE ## Id +# define XXH_OK XXH_IPREF(XXH_OK) +# define XXH_ERROR XXH_IPREF(XXH_ERROR) +# define XXH_errorcode XXH_IPREF(XXH_errorcode) +# define XXH32_canonical_t XXH_IPREF(XXH32_canonical_t) +# define XXH64_canonical_t XXH_IPREF(XXH64_canonical_t) +# define XXH128_canonical_t XXH_IPREF(XXH128_canonical_t) +# define XXH32_state_s XXH_IPREF(XXH32_state_s) +# define XXH32_state_t XXH_IPREF(XXH32_state_t) +# define XXH64_state_s XXH_IPREF(XXH64_state_s) +# define XXH64_state_t XXH_IPREF(XXH64_state_t) +# define XXH3_state_s XXH_IPREF(XXH3_state_s) +# define XXH3_state_t XXH_IPREF(XXH3_state_t) +# define XXH128_hash_t XXH_IPREF(XXH128_hash_t) + /* Ensure the header is parsed again, even if it was previously included */ +# undef XXHASH_H_5627135585666179 +# undef XXHASH_H_STATIC_13879238742 +#endif /* XXH_INLINE_ALL || XXH_PRIVATE_API */ + +/* **************************************************************** + * Stable API + *****************************************************************/ +#ifndef XXHASH_H_5627135585666179 +#define XXHASH_H_5627135585666179 1 + +/*! @brief Marks a global symbol. */ +#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API) +# if defined(_WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT)) +# ifdef XXH_EXPORT +# define XXH_PUBLIC_API __declspec(dllexport) +# elif XXH_IMPORT +# define XXH_PUBLIC_API __declspec(dllimport) +# endif +# else +# define XXH_PUBLIC_API /* do nothing */ +# endif +#endif + +#ifdef XXH_NAMESPACE +# define XXH_CAT(A,B) A##B +# define XXH_NAME2(A,B) XXH_CAT(A,B) +# define XXH_versionNumber XXH_NAME2(XXH_NAMESPACE, XXH_versionNumber) +/* XXH32 */ +# define XXH32 XXH_NAME2(XXH_NAMESPACE, XXH32) +# define XXH32_createState XXH_NAME2(XXH_NAMESPACE, XXH32_createState) +# define XXH32_freeState XXH_NAME2(XXH_NAMESPACE, XXH32_freeState) +# define XXH32_reset XXH_NAME2(XXH_NAMESPACE, XXH32_reset) +# define XXH32_update XXH_NAME2(XXH_NAMESPACE, XXH32_update) +# define XXH32_digest XXH_NAME2(XXH_NAMESPACE, XXH32_digest) +# define XXH32_copyState XXH_NAME2(XXH_NAMESPACE, XXH32_copyState) +# define XXH32_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH32_canonicalFromHash) +# define XXH32_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH32_hashFromCanonical) +/* XXH64 */ +# define XXH64 XXH_NAME2(XXH_NAMESPACE, XXH64) +# define XXH64_createState XXH_NAME2(XXH_NAMESPACE, XXH64_createState) +# define XXH64_freeState XXH_NAME2(XXH_NAMESPACE, XXH64_freeState) +# define XXH64_reset XXH_NAME2(XXH_NAMESPACE, XXH64_reset) +# define XXH64_update XXH_NAME2(XXH_NAMESPACE, XXH64_update) +# define XXH64_digest XXH_NAME2(XXH_NAMESPACE, XXH64_digest) +# define XXH64_copyState XXH_NAME2(XXH_NAMESPACE, XXH64_copyState) +# define XXH64_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH64_canonicalFromHash) +# define XXH64_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH64_hashFromCanonical) +/* XXH3_64bits */ +# define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits) +# define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret) +# define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed) +# define XXH3_64bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecretandSeed) +# define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState) +# define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState) +# define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState) +# define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset) +# define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed) +# define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret) +# define XXH3_64bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecretandSeed) +# define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update) +# define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest) +# define XXH3_generateSecret XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret) +# define XXH3_generateSecret_fromSeed XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret_fromSeed) +/* XXH3_128bits */ +# define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128) +# define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits) +# define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed) +# define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret) +# define XXH3_128bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecretandSeed) +# define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset) +# define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed) +# define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret) +# define XXH3_128bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecretandSeed) +# define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update) +# define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest) +# define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual) +# define XXH128_cmp XXH_NAME2(XXH_NAMESPACE, XXH128_cmp) +# define XXH128_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash) +# define XXH128_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical) +#endif + + +/* ************************************* +* Compiler specifics +***************************************/ + +/* specific declaration modes for Windows */ +#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API) +# if defined(_WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT)) +# ifdef XXH_EXPORT +# define XXH_PUBLIC_API __declspec(dllexport) +# elif XXH_IMPORT +# define XXH_PUBLIC_API __declspec(dllimport) +# endif +# else +# define XXH_PUBLIC_API /* do nothing */ +# endif +#endif + +#if defined (__GNUC__) +# define XXH_CONSTF __attribute__((__const__)) +# define XXH_PUREF __attribute__((__pure__)) +# define XXH_MALLOCF __attribute__((__malloc__)) +#else +# define XXH_CONSTF /* disable */ +# define XXH_PUREF +# define XXH_MALLOCF +#endif + +/* ************************************* +* Version +***************************************/ +#define XXH_VERSION_MAJOR 0 +#define XXH_VERSION_MINOR 8 +#define XXH_VERSION_RELEASE 3 +/*! @brief Version number, encoded as two digits each */ +#define XXH_VERSION_NUMBER (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE) + +/*! + * @brief Obtains the xxHash version. + * + * This is mostly useful when xxHash is compiled as a shared library, + * since the returned value comes from the library, as opposed to header file. + * + * @return @ref XXH_VERSION_NUMBER of the invoked library. + */ +XXH_PUBLIC_API XXH_CONSTF unsigned XXH_versionNumber (void); + + +/* **************************** +* Common basic types +******************************/ +#include /* size_t */ +/*! + * @brief Exit code for the streaming API. + */ +typedef enum { + XXH_OK = 0, /*!< OK */ + XXH_ERROR /*!< Error */ +} XXH_errorcode; + + +/*-********************************************************************** +* 32-bit hash +************************************************************************/ +#if defined(XXH_DOXYGEN) /* Don't show include */ +/*! + * @brief An unsigned 32-bit integer. + * + * Not necessarily defined to `uint32_t` but functionally equivalent. + */ +typedef uint32_t XXH32_hash_t; + +#elif !defined (__VMS) \ + && (defined (__cplusplus) \ + || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) +# ifdef _AIX +# include +# else +# include +# endif + typedef uint32_t XXH32_hash_t; + +#else +# include +# if UINT_MAX == 0xFFFFFFFFUL + typedef unsigned int XXH32_hash_t; +# elif ULONG_MAX == 0xFFFFFFFFUL + typedef unsigned long XXH32_hash_t; +# else +# error "unsupported platform: need a 32-bit type" +# endif +#endif + +/*! + * @} + * + * @defgroup XXH32_family XXH32 family + * @ingroup public + * Contains functions used in the classic 32-bit xxHash algorithm. + * + * @note + * XXH32 is useful for older platforms, with no or poor 64-bit performance. + * Note that the @ref XXH3_family provides competitive speed for both 32-bit + * and 64-bit systems, and offers true 64/128 bit hash results. + * + * @see @ref XXH64_family, @ref XXH3_family : Other xxHash families + * @see @ref XXH32_impl for implementation details + * @{ + */ + +/*! + * @brief Calculates the 32-bit hash of @p input using xxHash32. + * + * @param input The block of data to be hashed, at least @p length bytes in size. + * @param length The length of @p input, in bytes. + * @param seed The 32-bit seed to alter the hash's output predictably. + * + * @pre + * The memory between @p input and @p input + @p length must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p input may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * @return The calculated 32-bit xxHash32 value. + * + * @see @ref single_shot_example "Single Shot Example" for an example. + */ +XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed); + +#ifndef XXH_NO_STREAM +/*! + * @typedef struct XXH32_state_s XXH32_state_t + * @brief The opaque state struct for the XXH32 streaming API. + * + * @see XXH32_state_s for details. + * @see @ref streaming_example "Streaming Example" + */ +typedef struct XXH32_state_s XXH32_state_t; + +/*! + * @brief Allocates an @ref XXH32_state_t. + * + * @return An allocated pointer of @ref XXH32_state_t on success. + * @return `NULL` on failure. + * + * @note Must be freed with XXH32_freeState(). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_MALLOCF XXH32_state_t* XXH32_createState(void); +/*! + * @brief Frees an @ref XXH32_state_t. + * + * @param statePtr A pointer to an @ref XXH32_state_t allocated with @ref XXH32_createState(). + * + * @return @ref XXH_OK. + * + * @note @p statePtr must be allocated with XXH32_createState(). + * + * @see @ref streaming_example "Streaming Example" + * + */ +XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr); +/*! + * @brief Copies one @ref XXH32_state_t to another. + * + * @param dst_state The state to copy to. + * @param src_state The state to copy from. + * @pre + * @p dst_state and @p src_state must not be `NULL` and must not overlap. + */ +XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state); + +/*! + * @brief Resets an @ref XXH32_state_t to begin a new hash. + * + * @param statePtr The state struct to reset. + * @param seed The 32-bit seed to alter the hash result predictably. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note This function resets and seeds a state. Call it before @ref XXH32_update(). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH32_reset (XXH32_state_t* statePtr, XXH32_hash_t seed); + +/*! + * @brief Consumes a block of @p input to an @ref XXH32_state_t. + * + * @param statePtr The state struct to update. + * @param input The block of data to be hashed, at least @p length bytes in size. + * @param length The length of @p input, in bytes. + * + * @pre + * @p statePtr must not be `NULL`. + * @pre + * The memory between @p input and @p input + @p length must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p input may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note Call this to incrementally consume blocks of data. + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length); + +/*! + * @brief Returns the calculated hash value from an @ref XXH32_state_t. + * + * @param statePtr The state struct to calculate the hash from. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return The calculated 32-bit xxHash32 value from that state. + * + * @note + * Calling XXH32_digest() will not affect @p statePtr, so you can update, + * digest, and update again. + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr); +#endif /* !XXH_NO_STREAM */ + +/******* Canonical representation *******/ + +/*! + * @brief Canonical (big endian) representation of @ref XXH32_hash_t. + */ +typedef struct { + unsigned char digest[4]; /*!< Hash bytes, big endian */ +} XXH32_canonical_t; + +/*! + * @brief Converts an @ref XXH32_hash_t to a big endian @ref XXH32_canonical_t. + * + * @param dst The @ref XXH32_canonical_t pointer to be stored to. + * @param hash The @ref XXH32_hash_t to be converted. + * + * @pre + * @p dst must not be `NULL`. + * + * @see @ref canonical_representation_example "Canonical Representation Example" + */ +XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash); + +/*! + * @brief Converts an @ref XXH32_canonical_t to a native @ref XXH32_hash_t. + * + * @param src The @ref XXH32_canonical_t to convert. + * + * @pre + * @p src must not be `NULL`. + * + * @return The converted hash. + * + * @see @ref canonical_representation_example "Canonical Representation Example" + */ +XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src); + + +/*! @cond Doxygen ignores this part */ +#ifdef __has_attribute +# define XXH_HAS_ATTRIBUTE(x) __has_attribute(x) +#else +# define XXH_HAS_ATTRIBUTE(x) 0 +#endif +/*! @endcond */ + +/*! @cond Doxygen ignores this part */ +/* + * C23 __STDC_VERSION__ number hasn't been specified yet. For now + * leave as `201711L` (C17 + 1). + * TODO: Update to correct value when its been specified. + */ +#define XXH_C23_VN 201711L +/*! @endcond */ + +/*! @cond Doxygen ignores this part */ +/* C-language Attributes are added in C23. */ +#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) && defined(__has_c_attribute) +# define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x) +#else +# define XXH_HAS_C_ATTRIBUTE(x) 0 +#endif +/*! @endcond */ + +/*! @cond Doxygen ignores this part */ +#if defined(__cplusplus) && defined(__has_cpp_attribute) +# define XXH_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x) +#else +# define XXH_HAS_CPP_ATTRIBUTE(x) 0 +#endif +/*! @endcond */ + +/*! @cond Doxygen ignores this part */ +/* + * Define XXH_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute + * introduced in CPP17 and C23. + * CPP17 : https://en.cppreference.com/w/cpp/language/attributes/fallthrough + * C23 : https://en.cppreference.com/w/c/language/attributes/fallthrough + */ +#if XXH_HAS_C_ATTRIBUTE(fallthrough) || XXH_HAS_CPP_ATTRIBUTE(fallthrough) +# define XXH_FALLTHROUGH [[fallthrough]] +#elif XXH_HAS_ATTRIBUTE(__fallthrough__) +# define XXH_FALLTHROUGH __attribute__ ((__fallthrough__)) +#else +# define XXH_FALLTHROUGH /* fallthrough */ +#endif +/*! @endcond */ + +/*! @cond Doxygen ignores this part */ +/* + * Define XXH_NOESCAPE for annotated pointers in public API. + * https://clang.llvm.org/docs/AttributeReference.html#noescape + * As of writing this, only supported by clang. + */ +#if XXH_HAS_ATTRIBUTE(noescape) +# define XXH_NOESCAPE __attribute__((__noescape__)) +#else +# define XXH_NOESCAPE +#endif +/*! @endcond */ + + +/*! + * @} + * @ingroup public + * @{ + */ + +#ifndef XXH_NO_LONG_LONG +/*-********************************************************************** +* 64-bit hash +************************************************************************/ +#if defined(XXH_DOXYGEN) /* don't include */ +/*! + * @brief An unsigned 64-bit integer. + * + * Not necessarily defined to `uint64_t` but functionally equivalent. + */ +typedef uint64_t XXH64_hash_t; +#elif !defined (__VMS) \ + && (defined (__cplusplus) \ + || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) +# ifdef _AIX +# include +# else +# include +# endif + typedef uint64_t XXH64_hash_t; +#else +# include +# if defined(__LP64__) && ULONG_MAX == 0xFFFFFFFFFFFFFFFFULL + /* LP64 ABI says uint64_t is unsigned long */ + typedef unsigned long XXH64_hash_t; +# else + /* the following type must have a width of 64-bit */ + typedef unsigned long long XXH64_hash_t; +# endif +#endif + +/*! + * @} + * + * @defgroup XXH64_family XXH64 family + * @ingroup public + * @{ + * Contains functions used in the classic 64-bit xxHash algorithm. + * + * @note + * XXH3 provides competitive speed for both 32-bit and 64-bit systems, + * and offers true 64/128 bit hash results. + * It provides better speed for systems with vector processing capabilities. + */ + +/*! + * @brief Calculates the 64-bit hash of @p input using xxHash64. + * + * @param input The block of data to be hashed, at least @p length bytes in size. + * @param length The length of @p input, in bytes. + * @param seed The 64-bit seed to alter the hash's output predictably. + * + * @pre + * The memory between @p input and @p input + @p length must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p input may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * @return The calculated 64-bit xxHash64 value. + * + * @see @ref single_shot_example "Single Shot Example" for an example. + */ +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed); + +/******* Streaming *******/ +#ifndef XXH_NO_STREAM +/*! + * @brief The opaque state struct for the XXH64 streaming API. + * + * @see XXH64_state_s for details. + * @see @ref streaming_example "Streaming Example" + */ +typedef struct XXH64_state_s XXH64_state_t; /* incomplete type */ + +/*! + * @brief Allocates an @ref XXH64_state_t. + * + * @return An allocated pointer of @ref XXH64_state_t on success. + * @return `NULL` on failure. + * + * @note Must be freed with XXH64_freeState(). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_MALLOCF XXH64_state_t* XXH64_createState(void); + +/*! + * @brief Frees an @ref XXH64_state_t. + * + * @param statePtr A pointer to an @ref XXH64_state_t allocated with @ref XXH64_createState(). + * + * @return @ref XXH_OK. + * + * @note @p statePtr must be allocated with XXH64_createState(). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr); + +/*! + * @brief Copies one @ref XXH64_state_t to another. + * + * @param dst_state The state to copy to. + * @param src_state The state to copy from. + * @pre + * @p dst_state and @p src_state must not be `NULL` and must not overlap. + */ +XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dst_state, const XXH64_state_t* src_state); + +/*! + * @brief Resets an @ref XXH64_state_t to begin a new hash. + * + * @param statePtr The state struct to reset. + * @param seed The 64-bit seed to alter the hash result predictably. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note This function resets and seeds a state. Call it before @ref XXH64_update(). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH64_reset (XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed); + +/*! + * @brief Consumes a block of @p input to an @ref XXH64_state_t. + * + * @param statePtr The state struct to update. + * @param input The block of data to be hashed, at least @p length bytes in size. + * @param length The length of @p input, in bytes. + * + * @pre + * @p statePtr must not be `NULL`. + * @pre + * The memory between @p input and @p input + @p length must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p input may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note Call this to incrementally consume blocks of data. + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH_NOESCAPE XXH64_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length); + +/*! + * @brief Returns the calculated hash value from an @ref XXH64_state_t. + * + * @param statePtr The state struct to calculate the hash from. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return The calculated 64-bit xxHash64 value from that state. + * + * @note + * Calling XXH64_digest() will not affect @p statePtr, so you can update, + * digest, and update again. + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_digest (XXH_NOESCAPE const XXH64_state_t* statePtr); +#endif /* !XXH_NO_STREAM */ +/******* Canonical representation *******/ + +/*! + * @brief Canonical (big endian) representation of @ref XXH64_hash_t. + */ +typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t; + +/*! + * @brief Converts an @ref XXH64_hash_t to a big endian @ref XXH64_canonical_t. + * + * @param dst The @ref XXH64_canonical_t pointer to be stored to. + * @param hash The @ref XXH64_hash_t to be converted. + * + * @pre + * @p dst must not be `NULL`. + * + * @see @ref canonical_representation_example "Canonical Representation Example" + */ +XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash); + +/*! + * @brief Converts an @ref XXH64_canonical_t to a native @ref XXH64_hash_t. + * + * @param src The @ref XXH64_canonical_t to convert. + * + * @pre + * @p src must not be `NULL`. + * + * @return The converted hash. + * + * @see @ref canonical_representation_example "Canonical Representation Example" + */ +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src); + +#ifndef XXH_NO_XXH3 + +/*! + * @} + * ************************************************************************ + * @defgroup XXH3_family XXH3 family + * @ingroup public + * @{ + * + * XXH3 is a more recent hash algorithm featuring: + * - Improved speed for both small and large inputs + * - True 64-bit and 128-bit outputs + * - SIMD acceleration + * - Improved 32-bit viability + * + * Speed analysis methodology is explained here: + * + * https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html + * + * Compared to XXH64, expect XXH3 to run approximately + * ~2x faster on large inputs and >3x faster on small ones, + * exact differences vary depending on platform. + * + * XXH3's speed benefits greatly from SIMD and 64-bit arithmetic, + * but does not require it. + * Most 32-bit and 64-bit targets that can run XXH32 smoothly can run XXH3 + * at competitive speeds, even without vector support. Further details are + * explained in the implementation. + * + * XXH3 has a fast scalar implementation, but it also includes accelerated SIMD + * implementations for many common platforms: + * - AVX512 + * - AVX2 + * - SSE2 + * - ARM NEON + * - WebAssembly SIMD128 + * - POWER8 VSX + * - s390x ZVector + * This can be controlled via the @ref XXH_VECTOR macro, but it automatically + * selects the best version according to predefined macros. For the x86 family, an + * automatic runtime dispatcher is included separately in @ref xxh_x86dispatch.c. + * + * XXH3 implementation is portable: + * it has a generic C90 formulation that can be compiled on any platform, + * all implementations generate exactly the same hash value on all platforms. + * Starting from v0.8.0, it's also labelled "stable", meaning that + * any future version will also generate the same hash value. + * + * XXH3 offers 2 variants, _64bits and _128bits. + * + * When only 64 bits are needed, prefer invoking the _64bits variant, as it + * reduces the amount of mixing, resulting in faster speed on small inputs. + * It's also generally simpler to manipulate a scalar return type than a struct. + * + * The API supports one-shot hashing, streaming mode, and custom secrets. + */ +/*-********************************************************************** +* XXH3 64-bit variant +************************************************************************/ + +/*! + * @brief Calculates 64-bit unseeded variant of XXH3 hash of @p input. + * + * @param input The block of data to be hashed, at least @p length bytes in size. + * @param length The length of @p input, in bytes. + * + * @pre + * The memory between @p input and @p input + @p length must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p input may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * @return The calculated 64-bit XXH3 hash value. + * + * @note + * This is equivalent to @ref XXH3_64bits_withSeed() with a seed of `0`, however + * it may have slightly better performance due to constant propagation of the + * defaults. + * + * @see + * XXH3_64bits_withSeed(), XXH3_64bits_withSecret(): other seeding variants + * @see @ref single_shot_example "Single Shot Example" for an example. + */ +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length); + +/*! + * @brief Calculates 64-bit seeded variant of XXH3 hash of @p input. + * + * @param input The block of data to be hashed, at least @p length bytes in size. + * @param length The length of @p input, in bytes. + * @param seed The 64-bit seed to alter the hash result predictably. + * + * @pre + * The memory between @p input and @p input + @p length must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p input may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * @return The calculated 64-bit XXH3 hash value. + * + * @note + * seed == 0 produces the same results as @ref XXH3_64bits(). + * + * This variant generates a custom secret on the fly based on default secret + * altered using the @p seed value. + * + * While this operation is decently fast, note that it's not completely free. + * + * @see @ref single_shot_example "Single Shot Example" for an example. + */ +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed); + +/*! + * The bare minimum size for a custom secret. + * + * @see + * XXH3_64bits_withSecret(), XXH3_64bits_reset_withSecret(), + * XXH3_128bits_withSecret(), XXH3_128bits_reset_withSecret(). + */ +#define XXH3_SECRET_SIZE_MIN 136 + +/*! + * @brief Calculates 64-bit variant of XXH3 with a custom "secret". + * + * @param data The block of data to be hashed, at least @p len bytes in size. + * @param len The length of @p data, in bytes. + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * + * @return The calculated 64-bit XXH3 hash value. + * + * @pre + * The memory between @p data and @p data + @p len must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p data may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * It's possible to provide any blob of bytes as a "secret" to generate the hash. + * This makes it more difficult for an external actor to prepare an intentional collision. + * The main condition is that @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN). + * However, the quality of the secret impacts the dispersion of the hash algorithm. + * Therefore, the secret _must_ look like a bunch of random bytes. + * Avoid "trivial" or structured data such as repeated sequences or a text document. + * Whenever in doubt about the "randomness" of the blob of bytes, + * consider employing @ref XXH3_generateSecret() instead (see below). + * It will generate a proper high entropy secret derived from the blob of bytes. + * Another advantage of using XXH3_generateSecret() is that + * it guarantees that all bits within the initial blob of bytes + * will impact every bit of the output. + * This is not necessarily the case when using the blob of bytes directly + * because, when hashing _small_ inputs, only a portion of the secret is employed. + * + * @see @ref single_shot_example "Single Shot Example" for an example. + */ +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize); + + +/******* Streaming *******/ +#ifndef XXH_NO_STREAM +/* + * Streaming requires state maintenance. + * This operation costs memory and CPU. + * As a consequence, streaming is slower than one-shot hashing. + * For better performance, prefer one-shot functions whenever applicable. + */ + +/*! + * @brief The opaque state struct for the XXH3 streaming API. + * + * @see XXH3_state_s for details. + * @see @ref streaming_example "Streaming Example" + */ +typedef struct XXH3_state_s XXH3_state_t; +XXH_PUBLIC_API XXH_MALLOCF XXH3_state_t* XXH3_createState(void); +XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr); + +/*! + * @brief Copies one @ref XXH3_state_t to another. + * + * @param dst_state The state to copy to. + * @param src_state The state to copy from. + * @pre + * @p dst_state and @p src_state must not be `NULL` and must not overlap. + */ +XXH_PUBLIC_API void XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state); + +/*! + * @brief Resets an @ref XXH3_state_t to begin a new hash. + * + * @param statePtr The state struct to reset. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note + * - This function resets `statePtr` and generate a secret with default parameters. + * - Call this function before @ref XXH3_64bits_update(). + * - Digest will be equivalent to `XXH3_64bits()`. + * + * @see @ref streaming_example "Streaming Example" + * + */ +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr); + +/*! + * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash. + * + * @param statePtr The state struct to reset. + * @param seed The 64-bit seed to alter the hash result predictably. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note + * - This function resets `statePtr` and generate a secret from `seed`. + * - Call this function before @ref XXH3_64bits_update(). + * - Digest will be equivalent to `XXH3_64bits_withSeed()`. + * + * @see @ref streaming_example "Streaming Example" + * + */ +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed); + +/*! + * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash. + * + * @param statePtr The state struct to reset. + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note + * `secret` is referenced, it _must outlive_ the hash streaming session. + * + * Similar to one-shot API, `secretSize` must be >= @ref XXH3_SECRET_SIZE_MIN, + * and the quality of produced hash values depends on secret's entropy + * (secret's content should look like a bunch of random bytes). + * When in doubt about the randomness of a candidate `secret`, + * consider employing `XXH3_generateSecret()` instead (see below). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize); + +/*! + * @brief Consumes a block of @p input to an @ref XXH3_state_t. + * + * @param statePtr The state struct to update. + * @param input The block of data to be hashed, at least @p length bytes in size. + * @param length The length of @p input, in bytes. + * + * @pre + * @p statePtr must not be `NULL`. + * @pre + * The memory between @p input and @p input + @p length must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p input may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note Call this to incrementally consume blocks of data. + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length); + +/*! + * @brief Returns the calculated XXH3 64-bit hash value from an @ref XXH3_state_t. + * + * @param statePtr The state struct to calculate the hash from. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return The calculated XXH3 64-bit hash value from that state. + * + * @note + * Calling XXH3_64bits_digest() will not affect @p statePtr, so you can update, + * digest, and update again. + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr); +#endif /* !XXH_NO_STREAM */ + +/* note : canonical representation of XXH3 is the same as XXH64 + * since they both produce XXH64_hash_t values */ + + +/*-********************************************************************** +* XXH3 128-bit variant +************************************************************************/ + +/*! + * @brief The return value from 128-bit hashes. + * + * Stored in little endian order, although the fields themselves are in native + * endianness. + */ +typedef struct { + XXH64_hash_t low64; /*!< `value & 0xFFFFFFFFFFFFFFFF` */ + XXH64_hash_t high64; /*!< `value >> 64` */ +} XXH128_hash_t; + +/*! + * @brief Calculates 128-bit unseeded variant of XXH3 of @p data. + * + * @param data The block of data to be hashed, at least @p length bytes in size. + * @param len The length of @p data, in bytes. + * + * @return The calculated 128-bit variant of XXH3 value. + * + * The 128-bit variant of XXH3 has more strength, but it has a bit of overhead + * for shorter inputs. + * + * This is equivalent to @ref XXH3_128bits_withSeed() with a seed of `0`, however + * it may have slightly better performance due to constant propagation of the + * defaults. + * + * @see XXH3_128bits_withSeed(), XXH3_128bits_withSecret(): other seeding variants + * @see @ref single_shot_example "Single Shot Example" for an example. + */ +XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* data, size_t len); +/*! @brief Calculates 128-bit seeded variant of XXH3 hash of @p data. + * + * @param data The block of data to be hashed, at least @p length bytes in size. + * @param len The length of @p data, in bytes. + * @param seed The 64-bit seed to alter the hash result predictably. + * + * @return The calculated 128-bit variant of XXH3 value. + * + * @note + * seed == 0 produces the same results as @ref XXH3_64bits(). + * + * This variant generates a custom secret on the fly based on default secret + * altered using the @p seed value. + * + * While this operation is decently fast, note that it's not completely free. + * + * @see XXH3_128bits(), XXH3_128bits_withSecret(): other seeding variants + * @see @ref single_shot_example "Single Shot Example" for an example. + */ +XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSeed(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed); +/*! + * @brief Calculates 128-bit variant of XXH3 with a custom "secret". + * + * @param data The block of data to be hashed, at least @p len bytes in size. + * @param len The length of @p data, in bytes. + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * + * @return The calculated 128-bit variant of XXH3 value. + * + * It's possible to provide any blob of bytes as a "secret" to generate the hash. + * This makes it more difficult for an external actor to prepare an intentional collision. + * The main condition is that @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN). + * However, the quality of the secret impacts the dispersion of the hash algorithm. + * Therefore, the secret _must_ look like a bunch of random bytes. + * Avoid "trivial" or structured data such as repeated sequences or a text document. + * Whenever in doubt about the "randomness" of the blob of bytes, + * consider employing @ref XXH3_generateSecret() instead (see below). + * It will generate a proper high entropy secret derived from the blob of bytes. + * Another advantage of using XXH3_generateSecret() is that + * it guarantees that all bits within the initial blob of bytes + * will impact every bit of the output. + * This is not necessarily the case when using the blob of bytes directly + * because, when hashing _small_ inputs, only a portion of the secret is employed. + * + * @see @ref single_shot_example "Single Shot Example" for an example. + */ +XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize); + +/******* Streaming *******/ +#ifndef XXH_NO_STREAM +/* + * Streaming requires state maintenance. + * This operation costs memory and CPU. + * As a consequence, streaming is slower than one-shot hashing. + * For better performance, prefer one-shot functions whenever applicable. + * + * XXH3_128bits uses the same XXH3_state_t as XXH3_64bits(). + * Use already declared XXH3_createState() and XXH3_freeState(). + * + * All reset and streaming functions have same meaning as their 64-bit counterpart. + */ + +/*! + * @brief Resets an @ref XXH3_state_t to begin a new hash. + * + * @param statePtr The state struct to reset. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note + * - This function resets `statePtr` and generate a secret with default parameters. + * - Call it before @ref XXH3_128bits_update(). + * - Digest will be equivalent to `XXH3_128bits()`. + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr); + +/*! + * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash. + * + * @param statePtr The state struct to reset. + * @param seed The 64-bit seed to alter the hash result predictably. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note + * - This function resets `statePtr` and generate a secret from `seed`. + * - Call it before @ref XXH3_128bits_update(). + * - Digest will be equivalent to `XXH3_128bits_withSeed()`. + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed); +/*! + * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash. + * + * @param statePtr The state struct to reset. + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * `secret` is referenced, it _must outlive_ the hash streaming session. + * Similar to one-shot API, `secretSize` must be >= @ref XXH3_SECRET_SIZE_MIN, + * and the quality of produced hash values depends on secret's entropy + * (secret's content should look like a bunch of random bytes). + * When in doubt about the randomness of a candidate `secret`, + * consider employing `XXH3_generateSecret()` instead (see below). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize); + +/*! + * @brief Consumes a block of @p input to an @ref XXH3_state_t. + * + * Call this to incrementally consume blocks of data. + * + * @param statePtr The state struct to update. + * @param input The block of data to be hashed, at least @p length bytes in size. + * @param length The length of @p input, in bytes. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note + * The memory between @p input and @p input + @p length must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p input may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + */ +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length); + +/*! + * @brief Returns the calculated XXH3 128-bit hash value from an @ref XXH3_state_t. + * + * @param statePtr The state struct to calculate the hash from. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return The calculated XXH3 128-bit hash value from that state. + * + * @note + * Calling XXH3_128bits_digest() will not affect @p statePtr, so you can update, + * digest, and update again. + * + */ +XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr); +#endif /* !XXH_NO_STREAM */ + +/* Following helper functions make it possible to compare XXH128_hast_t values. + * Since XXH128_hash_t is a structure, this capability is not offered by the language. + * Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */ + +/*! + * @brief Check equality of two XXH128_hash_t values + * + * @param h1 The 128-bit hash value. + * @param h2 Another 128-bit hash value. + * + * @return `1` if `h1` and `h2` are equal. + * @return `0` if they are not. + */ +XXH_PUBLIC_API XXH_PUREF int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2); + +/*! + * @brief Compares two @ref XXH128_hash_t + * + * This comparator is compatible with stdlib's `qsort()`/`bsearch()`. + * + * @param h128_1 Left-hand side value + * @param h128_2 Right-hand side value + * + * @return >0 if @p h128_1 > @p h128_2 + * @return =0 if @p h128_1 == @p h128_2 + * @return <0 if @p h128_1 < @p h128_2 + */ +XXH_PUBLIC_API XXH_PUREF int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2); + + +/******* Canonical representation *******/ +typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t; + + +/*! + * @brief Converts an @ref XXH128_hash_t to a big endian @ref XXH128_canonical_t. + * + * @param dst The @ref XXH128_canonical_t pointer to be stored to. + * @param hash The @ref XXH128_hash_t to be converted. + * + * @pre + * @p dst must not be `NULL`. + * @see @ref canonical_representation_example "Canonical Representation Example" + */ +XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash); + +/*! + * @brief Converts an @ref XXH128_canonical_t to a native @ref XXH128_hash_t. + * + * @param src The @ref XXH128_canonical_t to convert. + * + * @pre + * @p src must not be `NULL`. + * + * @return The converted hash. + * @see @ref canonical_representation_example "Canonical Representation Example" + */ +XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src); + + +#endif /* !XXH_NO_XXH3 */ +#endif /* XXH_NO_LONG_LONG */ + +/*! + * @} + */ +#endif /* XXHASH_H_5627135585666179 */ + + + +#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) +#define XXHASH_H_STATIC_13879238742 +/* **************************************************************************** + * This section contains declarations which are not guaranteed to remain stable. + * They may change in future versions, becoming incompatible with a different + * version of the library. + * These declarations should only be used with static linking. + * Never use them in association with dynamic linking! + ***************************************************************************** */ + +/* + * These definitions are only present to allow static allocation + * of XXH states, on stack or in a struct, for example. + * Never **ever** access their members directly. + */ + +/*! + * @internal + * @brief Structure for XXH32 streaming API. + * + * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY, + * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is + * an opaque type. This allows fields to safely be changed. + * + * Typedef'd to @ref XXH32_state_t. + * Do not access the members of this struct directly. + * @see XXH64_state_s, XXH3_state_s + */ +struct XXH32_state_s { + XXH32_hash_t total_len_32; /*!< Total length hashed, modulo 2^32 */ + XXH32_hash_t large_len; /*!< Whether the hash is >= 16 (handles @ref total_len_32 overflow) */ + XXH32_hash_t v[4]; /*!< Accumulator lanes */ + XXH32_hash_t mem32[4]; /*!< Internal buffer for partial reads. Treated as unsigned char[16]. */ + XXH32_hash_t memsize; /*!< Amount of data in @ref mem32 */ + XXH32_hash_t reserved; /*!< Reserved field. Do not read nor write to it. */ +}; /* typedef'd to XXH32_state_t */ + + +#ifndef XXH_NO_LONG_LONG /* defined when there is no 64-bit support */ + +/*! + * @internal + * @brief Structure for XXH64 streaming API. + * + * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY, + * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is + * an opaque type. This allows fields to safely be changed. + * + * Typedef'd to @ref XXH64_state_t. + * Do not access the members of this struct directly. + * @see XXH32_state_s, XXH3_state_s + */ +struct XXH64_state_s { + XXH64_hash_t total_len; /*!< Total length hashed. This is always 64-bit. */ + XXH64_hash_t v[4]; /*!< Accumulator lanes */ + XXH64_hash_t mem64[4]; /*!< Internal buffer for partial reads. Treated as unsigned char[32]. */ + XXH32_hash_t memsize; /*!< Amount of data in @ref mem64 */ + XXH32_hash_t reserved32; /*!< Reserved field, needed for padding anyways*/ + XXH64_hash_t reserved64; /*!< Reserved field. Do not read or write to it. */ +}; /* typedef'd to XXH64_state_t */ + +#ifndef XXH_NO_XXH3 + +/* Windows SDK under 10.0.22000 is missing stdalign.h so we add a check + before allowing the windows compiler to use the C11 form. + Reference: https://github.com/Cyan4973/xxHash/issues/955 */ +#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) \ + && (defined(_MSC_VER) && (_MSC_VER >= 1000) || !defined(_MSC_VER)) /* >= C11 */ +# include +# define XXH_ALIGN(n) alignas(n) +#elif defined(__cplusplus) && (__cplusplus >= 201103L) /* >= C++11 */ +/* In C++ alignas() is a keyword */ +# define XXH_ALIGN(n) alignas(n) +#elif defined(__GNUC__) +# define XXH_ALIGN(n) __attribute__ ((aligned(n))) +#elif defined(_MSC_VER) +# define XXH_ALIGN(n) __declspec(align(n)) +#else +# define XXH_ALIGN(n) /* disabled */ +#endif + +/* Old GCC versions only accept the attribute after the type in structures. */ +#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)) /* C11+ */ \ + && ! (defined(__cplusplus) && (__cplusplus >= 201103L)) /* >= C++11 */ \ + && defined(__GNUC__) +# define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align) +#else +# define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type +#endif + +/*! + * @brief The size of the internal XXH3 buffer. + * + * This is the optimal update size for incremental hashing. + * + * @see XXH3_64b_update(), XXH3_128b_update(). + */ +#define XXH3_INTERNALBUFFER_SIZE 256 + +/*! + * @internal + * @brief Default size of the secret buffer (and @ref XXH3_kSecret). + * + * This is the size used in @ref XXH3_kSecret and the seeded functions. + * + * Not to be confused with @ref XXH3_SECRET_SIZE_MIN. + */ +#define XXH3_SECRET_DEFAULT_SIZE 192 + +/*! + * @internal + * @brief Structure for XXH3 streaming API. + * + * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY, + * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. + * Otherwise it is an opaque type. + * Never use this definition in combination with dynamic library. + * This allows fields to safely be changed in the future. + * + * @note ** This structure has a strict alignment requirement of 64 bytes!! ** + * Do not allocate this with `malloc()` or `new`, + * it will not be sufficiently aligned. + * Use @ref XXH3_createState() and @ref XXH3_freeState(), or stack allocation. + * + * Typedef'd to @ref XXH3_state_t. + * Do never access the members of this struct directly. + * + * @see XXH3_INITSTATE() for stack initialization. + * @see XXH3_createState(), XXH3_freeState(). + * @see XXH32_state_s, XXH64_state_s + */ +struct XXH3_state_s { + XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]); + /*!< The 8 accumulators. See @ref XXH32_state_s::v and @ref XXH64_state_s::v */ + XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]); + /*!< Used to store a custom secret generated from a seed. */ + XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]); + /*!< The internal buffer. @see XXH32_state_s::mem32 */ + XXH32_hash_t bufferedSize; + /*!< The amount of memory in @ref buffer, @see XXH32_state_s::memsize */ + XXH32_hash_t useSeed; + /*!< Reserved field. Needed for padding on 64-bit. */ + size_t nbStripesSoFar; + /*!< Number or stripes processed. */ + XXH64_hash_t totalLen; + /*!< Total length hashed. 64-bit even on 32-bit targets. */ + size_t nbStripesPerBlock; + /*!< Number of stripes per block. */ + size_t secretLimit; + /*!< Size of @ref customSecret or @ref extSecret */ + XXH64_hash_t seed; + /*!< Seed for _withSeed variants. Must be zero otherwise, @see XXH3_INITSTATE() */ + XXH64_hash_t reserved64; + /*!< Reserved field. */ + const unsigned char* extSecret; + /*!< Reference to an external secret for the _withSecret variants, NULL + * for other variants. */ + /* note: there may be some padding at the end due to alignment on 64 bytes */ +}; /* typedef'd to XXH3_state_t */ + +#undef XXH_ALIGN_MEMBER + +/*! + * @brief Initializes a stack-allocated `XXH3_state_s`. + * + * When the @ref XXH3_state_t structure is merely emplaced on stack, + * it should be initialized with XXH3_INITSTATE() or a memset() + * in case its first reset uses XXH3_NNbits_reset_withSeed(). + * This init can be omitted if the first reset uses default or _withSecret mode. + * This operation isn't necessary when the state is created with XXH3_createState(). + * Note that this doesn't prepare the state for a streaming operation, + * it's still necessary to use XXH3_NNbits_reset*() afterwards. + */ +#define XXH3_INITSTATE(XXH3_state_ptr) \ + do { \ + XXH3_state_t* tmp_xxh3_state_ptr = (XXH3_state_ptr); \ + tmp_xxh3_state_ptr->seed = 0; \ + tmp_xxh3_state_ptr->extSecret = NULL; \ + } while(0) + + +/*! + * @brief Calculates the 128-bit hash of @p data using XXH3. + * + * @param data The block of data to be hashed, at least @p len bytes in size. + * @param len The length of @p data, in bytes. + * @param seed The 64-bit seed to alter the hash's output predictably. + * + * @pre + * The memory between @p data and @p data + @p len must be valid, + * readable, contiguous memory. However, if @p len is `0`, @p data may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * @return The calculated 128-bit XXH3 value. + * + * @see @ref single_shot_example "Single Shot Example" for an example. + */ +XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed); + + +/* === Experimental API === */ +/* Symbols defined below must be considered tied to a specific library version. */ + +/*! + * @brief Derive a high-entropy secret from any user-defined content, named customSeed. + * + * @param secretBuffer A writable buffer for derived high-entropy secret data. + * @param secretSize Size of secretBuffer, in bytes. Must be >= XXH3_SECRET_SIZE_MIN. + * @param customSeed A user-defined content. + * @param customSeedSize Size of customSeed, in bytes. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * The generated secret can be used in combination with `*_withSecret()` functions. + * The `_withSecret()` variants are useful to provide a higher level of protection + * than 64-bit seed, as it becomes much more difficult for an external actor to + * guess how to impact the calculation logic. + * + * The function accepts as input a custom seed of any length and any content, + * and derives from it a high-entropy secret of length @p secretSize into an + * already allocated buffer @p secretBuffer. + * + * The generated secret can then be used with any `*_withSecret()` variant. + * The functions @ref XXH3_128bits_withSecret(), @ref XXH3_64bits_withSecret(), + * @ref XXH3_128bits_reset_withSecret() and @ref XXH3_64bits_reset_withSecret() + * are part of this list. They all accept a `secret` parameter + * which must be large enough for implementation reasons (>= @ref XXH3_SECRET_SIZE_MIN) + * _and_ feature very high entropy (consist of random-looking bytes). + * These conditions can be a high bar to meet, so @ref XXH3_generateSecret() can + * be employed to ensure proper quality. + * + * @p customSeed can be anything. It can have any size, even small ones, + * and its content can be anything, even "poor entropy" sources such as a bunch + * of zeroes. The resulting `secret` will nonetheless provide all required qualities. + * + * @pre + * - @p secretSize must be >= @ref XXH3_SECRET_SIZE_MIN + * - When @p customSeedSize > 0, supplying NULL as customSeed is undefined behavior. + * + * Example code: + * @code{.c} + * #include + * #include + * #include + * #define XXH_STATIC_LINKING_ONLY // expose unstable API + * #include "xxhash.h" + * // Hashes argv[2] using the entropy from argv[1]. + * int main(int argc, char* argv[]) + * { + * char secret[XXH3_SECRET_SIZE_MIN]; + * if (argv != 3) { return 1; } + * XXH3_generateSecret(secret, sizeof(secret), argv[1], strlen(argv[1])); + * XXH64_hash_t h = XXH3_64bits_withSecret( + * argv[2], strlen(argv[2]), + * secret, sizeof(secret) + * ); + * printf("%016llx\n", (unsigned long long) h); + * } + * @endcode + */ +XXH_PUBLIC_API XXH_errorcode XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize); + +/*! + * @brief Generate the same secret as the _withSeed() variants. + * + * @param secretBuffer A writable buffer of @ref XXH3_SECRET_DEFAULT_SIZE bytes + * @param seed The 64-bit seed to alter the hash result predictably. + * + * The generated secret can be used in combination with + *`*_withSecret()` and `_withSecretandSeed()` variants. + * + * Example C++ `std::string` hash class: + * @code{.cpp} + * #include + * #define XXH_STATIC_LINKING_ONLY // expose unstable API + * #include "xxhash.h" + * // Slow, seeds each time + * class HashSlow { + * XXH64_hash_t seed; + * public: + * HashSlow(XXH64_hash_t s) : seed{s} {} + * size_t operator()(const std::string& x) const { + * return size_t{XXH3_64bits_withSeed(x.c_str(), x.length(), seed)}; + * } + * }; + * // Fast, caches the seeded secret for future uses. + * class HashFast { + * unsigned char secret[XXH3_SECRET_DEFAULT_SIZE]; + * public: + * HashFast(XXH64_hash_t s) { + * XXH3_generateSecret_fromSeed(secret, seed); + * } + * size_t operator()(const std::string& x) const { + * return size_t{ + * XXH3_64bits_withSecret(x.c_str(), x.length(), secret, sizeof(secret)) + * }; + * } + * }; + * @endcode + */ +XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed); + +/*! + * @brief Maximum size of "short" key in bytes. + */ +#define XXH3_MIDSIZE_MAX 240 + +/*! + * @brief Calculates 64/128-bit seeded variant of XXH3 hash of @p data. + * + * @param data The block of data to be hashed, at least @p len bytes in size. + * @param len The length of @p data, in bytes. + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * @param seed The 64-bit seed to alter the hash result predictably. + * + * These variants generate hash values using either: + * - @p seed for "short" keys (< @ref XXH3_MIDSIZE_MAX = 240 bytes) + * - @p secret for "large" keys (>= @ref XXH3_MIDSIZE_MAX). + * + * This generally benefits speed, compared to `_withSeed()` or `_withSecret()`. + * `_withSeed()` has to generate the secret on the fly for "large" keys. + * It's fast, but can be perceptible for "not so large" keys (< 1 KB). + * `_withSecret()` has to generate the masks on the fly for "small" keys, + * which requires more instructions than _withSeed() variants. + * Therefore, _withSecretandSeed variant combines the best of both worlds. + * + * When @p secret has been generated by XXH3_generateSecret_fromSeed(), + * this variant produces *exactly* the same results as `_withSeed()` variant, + * hence offering only a pure speed benefit on "large" input, + * by skipping the need to regenerate the secret for every large input. + * + * Another usage scenario is to hash the secret to a 64-bit hash value, + * for example with XXH3_64bits(), which then becomes the seed, + * and then employ both the seed and the secret in _withSecretandSeed(). + * On top of speed, an added benefit is that each bit in the secret + * has a 50% chance to swap each bit in the output, via its impact to the seed. + * + * This is not guaranteed when using the secret directly in "small data" scenarios, + * because only portions of the secret are employed for small data. + */ +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t +XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* data, size_t len, + XXH_NOESCAPE const void* secret, size_t secretSize, + XXH64_hash_t seed); + +/*! + * @brief Calculates 128-bit seeded variant of XXH3 hash of @p data. + * + * @param data The memory segment to be hashed, at least @p len bytes in size. + * @param length The length of @p data, in bytes. + * @param secret The secret used to alter hash result predictably. + * @param secretSize The length of @p secret, in bytes (must be >= XXH3_SECRET_SIZE_MIN) + * @param seed64 The 64-bit seed to alter the hash result predictably. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @see XXH3_64bits_withSecretandSeed(): contract is the same. + */ +XXH_PUBLIC_API XXH_PUREF XXH128_hash_t +XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length, + XXH_NOESCAPE const void* secret, size_t secretSize, + XXH64_hash_t seed64); + +#ifndef XXH_NO_STREAM +/*! + * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash. + * + * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState(). + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * @param seed64 The 64-bit seed to alter the hash result predictably. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @see XXH3_64bits_withSecretandSeed(). Contract is identical. + */ +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, + XXH_NOESCAPE const void* secret, size_t secretSize, + XXH64_hash_t seed64); + +/*! + * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash. + * + * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState(). + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * @param seed64 The 64-bit seed to alter the hash result predictably. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @see XXH3_64bits_withSecretandSeed(). Contract is identical. + * + * Note: there was a bug in an earlier version of this function (<= v0.8.2) + * that would make it generate an incorrect hash value + * when @p seed == 0 and @p length < XXH3_MIDSIZE_MAX + * and @p secret is different from XXH3_generateSecret_fromSeed(). + * As stated in the contract, the correct hash result must be + * the same as XXH3_128bits_withSeed() when @p length <= XXH3_MIDSIZE_MAX. + * Results generated by this older version are wrong, hence not comparable. + */ +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, + XXH_NOESCAPE const void* secret, size_t secretSize, + XXH64_hash_t seed64); + +#endif /* !XXH_NO_STREAM */ + +#endif /* !XXH_NO_XXH3 */ +#endif /* XXH_NO_LONG_LONG */ +#if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) +# define XXH_IMPLEMENTATION +#endif + +#endif /* defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) */ + + +/* ======================================================================== */ +/* ======================================================================== */ +/* ======================================================================== */ + + +/*-********************************************************************** + * xxHash implementation + *-********************************************************************** + * xxHash's implementation used to be hosted inside xxhash.c. + * + * However, inlining requires implementation to be visible to the compiler, + * hence be included alongside the header. + * Previously, implementation was hosted inside xxhash.c, + * which was then #included when inlining was activated. + * This construction created issues with a few build and install systems, + * as it required xxhash.c to be stored in /include directory. + * + * xxHash implementation is now directly integrated within xxhash.h. + * As a consequence, xxhash.c is no longer needed in /include. + * + * xxhash.c is still available and is still useful. + * In a "normal" setup, when xxhash is not inlined, + * xxhash.h only exposes the prototypes and public symbols, + * while xxhash.c can be built into an object file xxhash.o + * which can then be linked into the final binary. + ************************************************************************/ + +#if ( defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) \ + || defined(XXH_IMPLEMENTATION) ) && !defined(XXH_IMPLEM_13a8737387) +# define XXH_IMPLEM_13a8737387 + +/* ************************************* +* Tuning parameters +***************************************/ + +/*! + * @defgroup tuning Tuning parameters + * @{ + * + * Various macros to control xxHash's behavior. + */ +#ifdef XXH_DOXYGEN +/*! + * @brief Define this to disable 64-bit code. + * + * Useful if only using the @ref XXH32_family and you have a strict C90 compiler. + */ +# define XXH_NO_LONG_LONG +# undef XXH_NO_LONG_LONG /* don't actually */ +/*! + * @brief Controls how unaligned memory is accessed. + * + * By default, access to unaligned memory is controlled by `memcpy()`, which is + * safe and portable. + * + * Unfortunately, on some target/compiler combinations, the generated assembly + * is sub-optimal. + * + * The below switch allow selection of a different access method + * in the search for improved performance. + * + * @par Possible options: + * + * - `XXH_FORCE_MEMORY_ACCESS=0` (default): `memcpy` + * @par + * Use `memcpy()`. Safe and portable. Note that most modern compilers will + * eliminate the function call and treat it as an unaligned access. + * + * - `XXH_FORCE_MEMORY_ACCESS=1`: `__attribute__((aligned(1)))` + * @par + * Depends on compiler extensions and is therefore not portable. + * This method is safe _if_ your compiler supports it, + * and *generally* as fast or faster than `memcpy`. + * + * - `XXH_FORCE_MEMORY_ACCESS=2`: Direct cast + * @par + * Casts directly and dereferences. This method doesn't depend on the + * compiler, but it violates the C standard as it directly dereferences an + * unaligned pointer. It can generate buggy code on targets which do not + * support unaligned memory accesses, but in some circumstances, it's the + * only known way to get the most performance. + * + * - `XXH_FORCE_MEMORY_ACCESS=3`: Byteshift + * @par + * Also portable. This can generate the best code on old compilers which don't + * inline small `memcpy()` calls, and it might also be faster on big-endian + * systems which lack a native byteswap instruction. However, some compilers + * will emit literal byteshifts even if the target supports unaligned access. + * + * + * @warning + * Methods 1 and 2 rely on implementation-defined behavior. Use these with + * care, as what works on one compiler/platform/optimization level may cause + * another to read garbage data or even crash. + * + * See https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html for details. + * + * Prefer these methods in priority order (0 > 3 > 1 > 2) + */ +# define XXH_FORCE_MEMORY_ACCESS 0 + +/*! + * @def XXH_SIZE_OPT + * @brief Controls how much xxHash optimizes for size. + * + * xxHash, when compiled, tends to result in a rather large binary size. This + * is mostly due to heavy usage to forced inlining and constant folding of the + * @ref XXH3_family to increase performance. + * + * However, some developers prefer size over speed. This option can + * significantly reduce the size of the generated code. When using the `-Os` + * or `-Oz` options on GCC or Clang, this is defined to 1 by default, + * otherwise it is defined to 0. + * + * Most of these size optimizations can be controlled manually. + * + * This is a number from 0-2. + * - `XXH_SIZE_OPT` == 0: Default. xxHash makes no size optimizations. Speed + * comes first. + * - `XXH_SIZE_OPT` == 1: Default for `-Os` and `-Oz`. xxHash is more + * conservative and disables hacks that increase code size. It implies the + * options @ref XXH_NO_INLINE_HINTS == 1, @ref XXH_FORCE_ALIGN_CHECK == 0, + * and @ref XXH3_NEON_LANES == 8 if they are not already defined. + * - `XXH_SIZE_OPT` == 2: xxHash tries to make itself as small as possible. + * Performance may cry. For example, the single shot functions just use the + * streaming API. + */ +# define XXH_SIZE_OPT 0 + +/*! + * @def XXH_FORCE_ALIGN_CHECK + * @brief If defined to non-zero, adds a special path for aligned inputs (XXH32() + * and XXH64() only). + * + * This is an important performance trick for architectures without decent + * unaligned memory access performance. + * + * It checks for input alignment, and when conditions are met, uses a "fast + * path" employing direct 32-bit/64-bit reads, resulting in _dramatically + * faster_ read speed. + * + * The check costs one initial branch per hash, which is generally negligible, + * but not zero. + * + * Moreover, it's not useful to generate an additional code path if memory + * access uses the same instruction for both aligned and unaligned + * addresses (e.g. x86 and aarch64). + * + * In these cases, the alignment check can be removed by setting this macro to 0. + * Then the code will always use unaligned memory access. + * Align check is automatically disabled on x86, x64, ARM64, and some ARM chips + * which are platforms known to offer good unaligned memory accesses performance. + * + * It is also disabled by default when @ref XXH_SIZE_OPT >= 1. + * + * This option does not affect XXH3 (only XXH32 and XXH64). + */ +# define XXH_FORCE_ALIGN_CHECK 0 + +/*! + * @def XXH_NO_INLINE_HINTS + * @brief When non-zero, sets all functions to `static`. + * + * By default, xxHash tries to force the compiler to inline almost all internal + * functions. + * + * This can usually improve performance due to reduced jumping and improved + * constant folding, but significantly increases the size of the binary which + * might not be favorable. + * + * Additionally, sometimes the forced inlining can be detrimental to performance, + * depending on the architecture. + * + * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the + * compiler full control on whether to inline or not. + * + * When not optimizing (-O0), using `-fno-inline` with GCC or Clang, or if + * @ref XXH_SIZE_OPT >= 1, this will automatically be defined. + */ +# define XXH_NO_INLINE_HINTS 0 + +/*! + * @def XXH3_INLINE_SECRET + * @brief Determines whether to inline the XXH3 withSecret code. + * + * When the secret size is known, the compiler can improve the performance + * of XXH3_64bits_withSecret() and XXH3_128bits_withSecret(). + * + * However, if the secret size is not known, it doesn't have any benefit. This + * happens when xxHash is compiled into a global symbol. Therefore, if + * @ref XXH_INLINE_ALL is *not* defined, this will be defined to 0. + * + * Additionally, this defaults to 0 on GCC 12+, which has an issue with function pointers + * that are *sometimes* force inline on -Og, and it is impossible to automatically + * detect this optimization level. + */ +# define XXH3_INLINE_SECRET 0 + +/*! + * @def XXH32_ENDJMP + * @brief Whether to use a jump for `XXH32_finalize`. + * + * For performance, `XXH32_finalize` uses multiple branches in the finalizer. + * This is generally preferable for performance, + * but depending on exact architecture, a jmp may be preferable. + * + * This setting is only possibly making a difference for very small inputs. + */ +# define XXH32_ENDJMP 0 + +/*! + * @internal + * @brief Redefines old internal names. + * + * For compatibility with code that uses xxHash's internals before the names + * were changed to improve namespacing. There is no other reason to use this. + */ +# define XXH_OLD_NAMES +# undef XXH_OLD_NAMES /* don't actually use, it is ugly. */ + +/*! + * @def XXH_NO_STREAM + * @brief Disables the streaming API. + * + * When xxHash is not inlined and the streaming functions are not used, disabling + * the streaming functions can improve code size significantly, especially with + * the @ref XXH3_family which tends to make constant folded copies of itself. + */ +# define XXH_NO_STREAM +# undef XXH_NO_STREAM /* don't actually */ +#endif /* XXH_DOXYGEN */ +/*! + * @} + */ + +#ifndef XXH_FORCE_MEMORY_ACCESS /* can be defined externally, on command line for example */ + /* prefer __packed__ structures (method 1) for GCC + * < ARMv7 with unaligned access (e.g. Raspbian armhf) still uses byte shifting, so we use memcpy + * which for some reason does unaligned loads. */ +# if defined(__GNUC__) && !(defined(__ARM_ARCH) && __ARM_ARCH < 7 && defined(__ARM_FEATURE_UNALIGNED)) +# define XXH_FORCE_MEMORY_ACCESS 1 +# endif +#endif + +#ifndef XXH_SIZE_OPT + /* default to 1 for -Os or -Oz */ +# if (defined(__GNUC__) || defined(__clang__)) && defined(__OPTIMIZE_SIZE__) +# define XXH_SIZE_OPT 1 +# else +# define XXH_SIZE_OPT 0 +# endif +#endif + +#ifndef XXH_FORCE_ALIGN_CHECK /* can be defined externally */ + /* don't check on sizeopt, x86, aarch64, or arm when unaligned access is available */ +# if XXH_SIZE_OPT >= 1 || \ + defined(__i386) || defined(__x86_64__) || defined(__aarch64__) || defined(__ARM_FEATURE_UNALIGNED) \ + || defined(_M_IX86) || defined(_M_X64) || defined(_M_ARM64) || defined(_M_ARM) /* visual */ +# define XXH_FORCE_ALIGN_CHECK 0 +# else +# define XXH_FORCE_ALIGN_CHECK 1 +# endif +#endif + +#ifndef XXH_NO_INLINE_HINTS +# if XXH_SIZE_OPT >= 1 || defined(__NO_INLINE__) /* -O0, -fno-inline */ +# define XXH_NO_INLINE_HINTS 1 +# else +# define XXH_NO_INLINE_HINTS 0 +# endif +#endif + +#ifndef XXH3_INLINE_SECRET +# if (defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 12) \ + || !defined(XXH_INLINE_ALL) +# define XXH3_INLINE_SECRET 0 +# else +# define XXH3_INLINE_SECRET 1 +# endif +#endif + +#ifndef XXH32_ENDJMP +/* generally preferable for performance */ +# define XXH32_ENDJMP 0 +#endif + +/*! + * @defgroup impl Implementation + * @{ + */ + + +/* ************************************* +* Includes & Memory related functions +***************************************/ +#if defined(XXH_NO_STREAM) +/* nothing */ +#elif defined(XXH_NO_STDLIB) + +/* When requesting to disable any mention of stdlib, + * the library loses the ability to invoked malloc / free. + * In practice, it means that functions like `XXH*_createState()` + * will always fail, and return NULL. + * This flag is useful in situations where + * xxhash.h is integrated into some kernel, embedded or limited environment + * without access to dynamic allocation. + */ + +static XXH_CONSTF void* XXH_malloc(size_t s) { (void)s; return NULL; } +static void XXH_free(void* p) { (void)p; } + +#else + +/* + * Modify the local functions below should you wish to use + * different memory routines for malloc() and free() + */ +#include + +/*! + * @internal + * @brief Modify this function to use a different routine than malloc(). + */ +static XXH_MALLOCF void* XXH_malloc(size_t s) { return malloc(s); } + +/*! + * @internal + * @brief Modify this function to use a different routine than free(). + */ +static void XXH_free(void* p) { free(p); } + +#endif /* XXH_NO_STDLIB */ + +#include + +/*! + * @internal + * @brief Modify this function to use a different routine than memcpy(). + */ +static void* XXH_memcpy(void* dest, const void* src, size_t size) +{ + return memcpy(dest,src,size); +} + +#include /* ULLONG_MAX */ + + +/* ************************************* +* Compiler Specific Options +***************************************/ +#ifdef _MSC_VER /* Visual Studio warning fix */ +# pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */ +#endif + +#if XXH_NO_INLINE_HINTS /* disable inlining hints */ +# if defined(__GNUC__) || defined(__clang__) +# define XXH_FORCE_INLINE static __attribute__((__unused__)) +# else +# define XXH_FORCE_INLINE static +# endif +# define XXH_NO_INLINE static +/* enable inlining hints */ +#elif defined(__GNUC__) || defined(__clang__) +# define XXH_FORCE_INLINE static __inline__ __attribute__((__always_inline__, __unused__)) +# define XXH_NO_INLINE static __attribute__((__noinline__)) +#elif defined(_MSC_VER) /* Visual Studio */ +# define XXH_FORCE_INLINE static __forceinline +# define XXH_NO_INLINE static __declspec(noinline) +#elif defined (__cplusplus) \ + || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) /* C99 */ +# define XXH_FORCE_INLINE static inline +# define XXH_NO_INLINE static +#else +# define XXH_FORCE_INLINE static +# define XXH_NO_INLINE static +#endif + +#if XXH3_INLINE_SECRET +# define XXH3_WITH_SECRET_INLINE XXH_FORCE_INLINE +#else +# define XXH3_WITH_SECRET_INLINE XXH_NO_INLINE +#endif + + +/* ************************************* +* Debug +***************************************/ +/*! + * @ingroup tuning + * @def XXH_DEBUGLEVEL + * @brief Sets the debugging level. + * + * XXH_DEBUGLEVEL is expected to be defined externally, typically via the + * compiler's command line options. The value must be a number. + */ +#ifndef XXH_DEBUGLEVEL +# ifdef DEBUGLEVEL /* backwards compat */ +# define XXH_DEBUGLEVEL DEBUGLEVEL +# else +# define XXH_DEBUGLEVEL 0 +# endif +#endif + +#if (XXH_DEBUGLEVEL>=1) +# include /* note: can still be disabled with NDEBUG */ +# define XXH_ASSERT(c) assert(c) +#else +# if defined(__INTEL_COMPILER) +# define XXH_ASSERT(c) XXH_ASSUME((unsigned char) (c)) +# else +# define XXH_ASSERT(c) XXH_ASSUME(c) +# endif +#endif + +/* note: use after variable declarations */ +#ifndef XXH_STATIC_ASSERT +# if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* C11 */ +# define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { _Static_assert((c),m); } while(0) +# elif defined(__cplusplus) && (__cplusplus >= 201103L) /* C++11 */ +# define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { static_assert((c),m); } while(0) +# else +# define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { struct xxh_sa { char x[(c) ? 1 : -1]; }; } while(0) +# endif +# define XXH_STATIC_ASSERT(c) XXH_STATIC_ASSERT_WITH_MESSAGE((c),#c) +#endif + +/*! + * @internal + * @def XXH_COMPILER_GUARD(var) + * @brief Used to prevent unwanted optimizations for @p var. + * + * It uses an empty GCC inline assembly statement with a register constraint + * which forces @p var into a general purpose register (eg eax, ebx, ecx + * on x86) and marks it as modified. + * + * This is used in a few places to avoid unwanted autovectorization (e.g. + * XXH32_round()). All vectorization we want is explicit via intrinsics, + * and _usually_ isn't wanted elsewhere. + * + * We also use it to prevent unwanted constant folding for AArch64 in + * XXH3_initCustomSecret_scalar(). + */ +#if defined(__GNUC__) || defined(__clang__) +# define XXH_COMPILER_GUARD(var) __asm__("" : "+r" (var)) +#else +# define XXH_COMPILER_GUARD(var) ((void)0) +#endif + +/* Specifically for NEON vectors which use the "w" constraint, on + * Clang. */ +#if defined(__clang__) && defined(__ARM_ARCH) && !defined(__wasm__) +# define XXH_COMPILER_GUARD_CLANG_NEON(var) __asm__("" : "+w" (var)) +#else +# define XXH_COMPILER_GUARD_CLANG_NEON(var) ((void)0) +#endif + +/* ************************************* +* Basic Types +***************************************/ +#if !defined (__VMS) \ + && (defined (__cplusplus) \ + || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) +# ifdef _AIX +# include +# else +# include +# endif + typedef uint8_t xxh_u8; +#else + typedef unsigned char xxh_u8; +#endif +typedef XXH32_hash_t xxh_u32; + +#ifdef XXH_OLD_NAMES +# warning "XXH_OLD_NAMES is planned to be removed starting v0.9. If the program depends on it, consider moving away from it by employing newer type names directly" +# define BYTE xxh_u8 +# define U8 xxh_u8 +# define U32 xxh_u32 +#endif + +/* *** Memory access *** */ + +/*! + * @internal + * @fn xxh_u32 XXH_read32(const void* ptr) + * @brief Reads an unaligned 32-bit integer from @p ptr in native endianness. + * + * Affected by @ref XXH_FORCE_MEMORY_ACCESS. + * + * @param ptr The pointer to read from. + * @return The 32-bit native endian integer from the bytes at @p ptr. + */ + +/*! + * @internal + * @fn xxh_u32 XXH_readLE32(const void* ptr) + * @brief Reads an unaligned 32-bit little endian integer from @p ptr. + * + * Affected by @ref XXH_FORCE_MEMORY_ACCESS. + * + * @param ptr The pointer to read from. + * @return The 32-bit little endian integer from the bytes at @p ptr. + */ + +/*! + * @internal + * @fn xxh_u32 XXH_readBE32(const void* ptr) + * @brief Reads an unaligned 32-bit big endian integer from @p ptr. + * + * Affected by @ref XXH_FORCE_MEMORY_ACCESS. + * + * @param ptr The pointer to read from. + * @return The 32-bit big endian integer from the bytes at @p ptr. + */ + +/*! + * @internal + * @fn xxh_u32 XXH_readLE32_align(const void* ptr, XXH_alignment align) + * @brief Like @ref XXH_readLE32(), but has an option for aligned reads. + * + * Affected by @ref XXH_FORCE_MEMORY_ACCESS. + * Note that when @ref XXH_FORCE_ALIGN_CHECK == 0, the @p align parameter is + * always @ref XXH_alignment::XXH_unaligned. + * + * @param ptr The pointer to read from. + * @param align Whether @p ptr is aligned. + * @pre + * If @p align == @ref XXH_alignment::XXH_aligned, @p ptr must be 4 byte + * aligned. + * @return The 32-bit little endian integer from the bytes at @p ptr. + */ + +#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3)) +/* + * Manual byteshift. Best for old compilers which don't inline memcpy. + * We actually directly use XXH_readLE32 and XXH_readBE32. + */ +#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2)) + +/* + * Force direct memory access. Only works on CPU which support unaligned memory + * access in hardware. + */ +static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; } + +#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1)) + +/* + * __attribute__((aligned(1))) is supported by gcc and clang. Originally the + * documentation claimed that it only increased the alignment, but actually it + * can decrease it on gcc, clang, and icc: + * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502, + * https://gcc.godbolt.org/z/xYez1j67Y. + */ +#ifdef XXH_OLD_NAMES +typedef union { xxh_u32 u32; } __attribute__((__packed__)) unalign; +#endif +static xxh_u32 XXH_read32(const void* ptr) +{ + typedef __attribute__((__aligned__(1))) xxh_u32 xxh_unalign32; + return *((const xxh_unalign32*)ptr); +} + +#else + +/* + * Portable and safe solution. Generally efficient. + * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html + */ +static xxh_u32 XXH_read32(const void* memPtr) +{ + xxh_u32 val; + XXH_memcpy(&val, memPtr, sizeof(val)); + return val; +} + +#endif /* XXH_FORCE_DIRECT_MEMORY_ACCESS */ + + +/* *** Endianness *** */ + +/*! + * @ingroup tuning + * @def XXH_CPU_LITTLE_ENDIAN + * @brief Whether the target is little endian. + * + * Defined to 1 if the target is little endian, or 0 if it is big endian. + * It can be defined externally, for example on the compiler command line. + * + * If it is not defined, + * a runtime check (which is usually constant folded) is used instead. + * + * @note + * This is not necessarily defined to an integer constant. + * + * @see XXH_isLittleEndian() for the runtime check. + */ +#ifndef XXH_CPU_LITTLE_ENDIAN +/* + * Try to detect endianness automatically, to avoid the nonstandard behavior + * in `XXH_isLittleEndian()` + */ +# if defined(_WIN32) /* Windows is always little endian */ \ + || defined(__LITTLE_ENDIAN__) \ + || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) +# define XXH_CPU_LITTLE_ENDIAN 1 +# elif defined(__BIG_ENDIAN__) \ + || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) +# define XXH_CPU_LITTLE_ENDIAN 0 +# else +/*! + * @internal + * @brief Runtime check for @ref XXH_CPU_LITTLE_ENDIAN. + * + * Most compilers will constant fold this. + */ +static int XXH_isLittleEndian(void) +{ + /* + * Portable and well-defined behavior. + * Don't use static: it is detrimental to performance. + */ + const union { xxh_u32 u; xxh_u8 c[4]; } one = { 1 }; + return one.c[0]; +} +# define XXH_CPU_LITTLE_ENDIAN XXH_isLittleEndian() +# endif +#endif + + + + +/* **************************************** +* Compiler-specific Functions and Macros +******************************************/ +#define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__) + +#ifdef __has_builtin +# define XXH_HAS_BUILTIN(x) __has_builtin(x) +#else +# define XXH_HAS_BUILTIN(x) 0 +#endif + + + +/* + * C23 and future versions have standard "unreachable()". + * Once it has been implemented reliably we can add it as an + * additional case: + * + * ``` + * #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) + * # include + * # ifdef unreachable + * # define XXH_UNREACHABLE() unreachable() + * # endif + * #endif + * ``` + * + * Note C++23 also has std::unreachable() which can be detected + * as follows: + * ``` + * #if defined(__cpp_lib_unreachable) && (__cpp_lib_unreachable >= 202202L) + * # include + * # define XXH_UNREACHABLE() std::unreachable() + * #endif + * ``` + * NB: `__cpp_lib_unreachable` is defined in the `` header. + * We don't use that as including `` in `extern "C"` blocks + * doesn't work on GCC12 + */ + +#if XXH_HAS_BUILTIN(__builtin_unreachable) +# define XXH_UNREACHABLE() __builtin_unreachable() + +#elif defined(_MSC_VER) +# define XXH_UNREACHABLE() __assume(0) + +#else +# define XXH_UNREACHABLE() +#endif + +#if XXH_HAS_BUILTIN(__builtin_assume) +# define XXH_ASSUME(c) __builtin_assume(c) +#else +# define XXH_ASSUME(c) if (!(c)) { XXH_UNREACHABLE(); } +#endif + +/*! + * @internal + * @def XXH_rotl32(x,r) + * @brief 32-bit rotate left. + * + * @param x The 32-bit integer to be rotated. + * @param r The number of bits to rotate. + * @pre + * @p r > 0 && @p r < 32 + * @note + * @p x and @p r may be evaluated multiple times. + * @return The rotated result. + */ +#if !defined(NO_CLANG_BUILTIN) && XXH_HAS_BUILTIN(__builtin_rotateleft32) \ + && XXH_HAS_BUILTIN(__builtin_rotateleft64) +# define XXH_rotl32 __builtin_rotateleft32 +# define XXH_rotl64 __builtin_rotateleft64 +/* Note: although _rotl exists for minGW (GCC under windows), performance seems poor */ +#elif defined(_MSC_VER) +# define XXH_rotl32(x,r) _rotl(x,r) +# define XXH_rotl64(x,r) _rotl64(x,r) +#else +# define XXH_rotl32(x,r) (((x) << (r)) | ((x) >> (32 - (r)))) +# define XXH_rotl64(x,r) (((x) << (r)) | ((x) >> (64 - (r)))) +#endif + +/*! + * @internal + * @fn xxh_u32 XXH_swap32(xxh_u32 x) + * @brief A 32-bit byteswap. + * + * @param x The 32-bit integer to byteswap. + * @return @p x, byteswapped. + */ +#if defined(_MSC_VER) /* Visual Studio */ +# define XXH_swap32 _byteswap_ulong +#elif XXH_GCC_VERSION >= 403 +# define XXH_swap32 __builtin_bswap32 +#else +static xxh_u32 XXH_swap32 (xxh_u32 x) +{ + return ((x << 24) & 0xff000000 ) | + ((x << 8) & 0x00ff0000 ) | + ((x >> 8) & 0x0000ff00 ) | + ((x >> 24) & 0x000000ff ); +} +#endif + + +/* *************************** +* Memory reads +*****************************/ + +/*! + * @internal + * @brief Enum to indicate whether a pointer is aligned. + */ +typedef enum { + XXH_aligned, /*!< Aligned */ + XXH_unaligned /*!< Possibly unaligned */ +} XXH_alignment; + +/* + * XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. + * + * This is ideal for older compilers which don't inline memcpy. + */ +#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3)) + +XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* memPtr) +{ + const xxh_u8* bytePtr = (const xxh_u8 *)memPtr; + return bytePtr[0] + | ((xxh_u32)bytePtr[1] << 8) + | ((xxh_u32)bytePtr[2] << 16) + | ((xxh_u32)bytePtr[3] << 24); +} + +XXH_FORCE_INLINE xxh_u32 XXH_readBE32(const void* memPtr) +{ + const xxh_u8* bytePtr = (const xxh_u8 *)memPtr; + return bytePtr[3] + | ((xxh_u32)bytePtr[2] << 8) + | ((xxh_u32)bytePtr[1] << 16) + | ((xxh_u32)bytePtr[0] << 24); +} + +#else +XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* ptr) +{ + return XXH_CPU_LITTLE_ENDIAN ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr)); +} + +static xxh_u32 XXH_readBE32(const void* ptr) +{ + return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr); +} +#endif + +XXH_FORCE_INLINE xxh_u32 +XXH_readLE32_align(const void* ptr, XXH_alignment align) +{ + if (align==XXH_unaligned) { + return XXH_readLE32(ptr); + } else { + return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u32*)ptr : XXH_swap32(*(const xxh_u32*)ptr); + } +} + + +/* ************************************* +* Misc +***************************************/ +/*! @ingroup public */ +XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; } + + +/* ******************************************************************* +* 32-bit hash functions +*********************************************************************/ +/*! + * @} + * @defgroup XXH32_impl XXH32 implementation + * @ingroup impl + * + * Details on the XXH32 implementation. + * @{ + */ + /* #define instead of static const, to be used as initializers */ +#define XXH_PRIME32_1 0x9E3779B1U /*!< 0b10011110001101110111100110110001 */ +#define XXH_PRIME32_2 0x85EBCA77U /*!< 0b10000101111010111100101001110111 */ +#define XXH_PRIME32_3 0xC2B2AE3DU /*!< 0b11000010101100101010111000111101 */ +#define XXH_PRIME32_4 0x27D4EB2FU /*!< 0b00100111110101001110101100101111 */ +#define XXH_PRIME32_5 0x165667B1U /*!< 0b00010110010101100110011110110001 */ + +#ifdef XXH_OLD_NAMES +# define PRIME32_1 XXH_PRIME32_1 +# define PRIME32_2 XXH_PRIME32_2 +# define PRIME32_3 XXH_PRIME32_3 +# define PRIME32_4 XXH_PRIME32_4 +# define PRIME32_5 XXH_PRIME32_5 +#endif + +/*! + * @internal + * @brief Normal stripe processing routine. + * + * This shuffles the bits so that any bit from @p input impacts several bits in + * @p acc. + * + * @param acc The accumulator lane. + * @param input The stripe of input to mix. + * @return The mixed accumulator lane. + */ +static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input) +{ + acc += input * XXH_PRIME32_2; + acc = XXH_rotl32(acc, 13); + acc *= XXH_PRIME32_1; +#if (defined(__SSE4_1__) || defined(__aarch64__) || defined(__wasm_simd128__)) && !defined(XXH_ENABLE_AUTOVECTORIZE) + /* + * UGLY HACK: + * A compiler fence is used to prevent GCC and Clang from + * autovectorizing the XXH32 loop (pragmas and attributes don't work for some + * reason) without globally disabling SSE4.1. + * + * The reason we want to avoid vectorization is because despite working on + * 4 integers at a time, there are multiple factors slowing XXH32 down on + * SSE4: + * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on + * newer chips!) making it slightly slower to multiply four integers at + * once compared to four integers independently. Even when pmulld was + * fastest, Sandy/Ivy Bridge, it is still not worth it to go into SSE + * just to multiply unless doing a long operation. + * + * - Four instructions are required to rotate, + * movqda tmp, v // not required with VEX encoding + * pslld tmp, 13 // tmp <<= 13 + * psrld v, 19 // x >>= 19 + * por v, tmp // x |= tmp + * compared to one for scalar: + * roll v, 13 // reliably fast across the board + * shldl v, v, 13 // Sandy Bridge and later prefer this for some reason + * + * - Instruction level parallelism is actually more beneficial here because + * the SIMD actually serializes this operation: While v1 is rotating, v2 + * can load data, while v3 can multiply. SSE forces them to operate + * together. + * + * This is also enabled on AArch64, as Clang is *very aggressive* in vectorizing + * the loop. NEON is only faster on the A53, and with the newer cores, it is less + * than half the speed. + * + * Additionally, this is used on WASM SIMD128 because it JITs to the same + * SIMD instructions and has the same issue. + */ + XXH_COMPILER_GUARD(acc); +#endif + return acc; +} + +/*! + * @internal + * @brief Mixes all bits to finalize the hash. + * + * The final mix ensures that all input bits have a chance to impact any bit in + * the output digest, resulting in an unbiased distribution. + * + * @param hash The hash to avalanche. + * @return The avalanched hash. + */ +static xxh_u32 XXH32_avalanche(xxh_u32 hash) +{ + hash ^= hash >> 15; + hash *= XXH_PRIME32_2; + hash ^= hash >> 13; + hash *= XXH_PRIME32_3; + hash ^= hash >> 16; + return hash; +} + +#define XXH_get32bits(p) XXH_readLE32_align(p, align) + +/*! + * @internal + * @brief Processes the last 0-15 bytes of @p ptr. + * + * There may be up to 15 bytes remaining to consume from the input. + * This final stage will digest them to ensure that all input bytes are present + * in the final mix. + * + * @param hash The hash to finalize. + * @param ptr The pointer to the remaining input. + * @param len The remaining length, modulo 16. + * @param align Whether @p ptr is aligned. + * @return The finalized hash. + * @see XXH64_finalize(). + */ +static XXH_PUREF xxh_u32 +XXH32_finalize(xxh_u32 hash, const xxh_u8* ptr, size_t len, XXH_alignment align) +{ +#define XXH_PROCESS1 do { \ + hash += (*ptr++) * XXH_PRIME32_5; \ + hash = XXH_rotl32(hash, 11) * XXH_PRIME32_1; \ +} while (0) + +#define XXH_PROCESS4 do { \ + hash += XXH_get32bits(ptr) * XXH_PRIME32_3; \ + ptr += 4; \ + hash = XXH_rotl32(hash, 17) * XXH_PRIME32_4; \ +} while (0) + + if (ptr==NULL) XXH_ASSERT(len == 0); + + /* Compact rerolled version; generally faster */ + if (!XXH32_ENDJMP) { + len &= 15; + while (len >= 4) { + XXH_PROCESS4; + len -= 4; + } + while (len > 0) { + XXH_PROCESS1; + --len; + } + return XXH32_avalanche(hash); + } else { + switch(len&15) /* or switch(bEnd - p) */ { + case 12: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 8: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 4: XXH_PROCESS4; + return XXH32_avalanche(hash); + + case 13: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 9: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 5: XXH_PROCESS4; + XXH_PROCESS1; + return XXH32_avalanche(hash); + + case 14: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 10: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 6: XXH_PROCESS4; + XXH_PROCESS1; + XXH_PROCESS1; + return XXH32_avalanche(hash); + + case 15: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 11: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 7: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 3: XXH_PROCESS1; + XXH_FALLTHROUGH; /* fallthrough */ + case 2: XXH_PROCESS1; + XXH_FALLTHROUGH; /* fallthrough */ + case 1: XXH_PROCESS1; + XXH_FALLTHROUGH; /* fallthrough */ + case 0: return XXH32_avalanche(hash); + } + XXH_ASSERT(0); + return hash; /* reaching this point is deemed impossible */ + } +} + +#ifdef XXH_OLD_NAMES +# define PROCESS1 XXH_PROCESS1 +# define PROCESS4 XXH_PROCESS4 +#else +# undef XXH_PROCESS1 +# undef XXH_PROCESS4 +#endif + +/*! + * @internal + * @brief The implementation for @ref XXH32(). + * + * @param input , len , seed Directly passed from @ref XXH32(). + * @param align Whether @p input is aligned. + * @return The calculated hash. + */ +XXH_FORCE_INLINE XXH_PUREF xxh_u32 +XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align) +{ + xxh_u32 h32; + + if (input==NULL) XXH_ASSERT(len == 0); + + if (len>=16) { + const xxh_u8* const bEnd = input + len; + const xxh_u8* const limit = bEnd - 15; + xxh_u32 v1 = seed + XXH_PRIME32_1 + XXH_PRIME32_2; + xxh_u32 v2 = seed + XXH_PRIME32_2; + xxh_u32 v3 = seed + 0; + xxh_u32 v4 = seed - XXH_PRIME32_1; + + do { + v1 = XXH32_round(v1, XXH_get32bits(input)); input += 4; + v2 = XXH32_round(v2, XXH_get32bits(input)); input += 4; + v3 = XXH32_round(v3, XXH_get32bits(input)); input += 4; + v4 = XXH32_round(v4, XXH_get32bits(input)); input += 4; + } while (input < limit); + + h32 = XXH_rotl32(v1, 1) + XXH_rotl32(v2, 7) + + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18); + } else { + h32 = seed + XXH_PRIME32_5; + } + + h32 += (xxh_u32)len; + + return XXH32_finalize(h32, input, len&15, align); +} + +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed) +{ +#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2 + /* Simple version, good for code maintenance, but unfortunately slow for small inputs */ + XXH32_state_t state; + XXH32_reset(&state, seed); + XXH32_update(&state, (const xxh_u8*)input, len); + return XXH32_digest(&state); +#else + if (XXH_FORCE_ALIGN_CHECK) { + if ((((size_t)input) & 3) == 0) { /* Input is 4-bytes aligned, leverage the speed benefit */ + return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_aligned); + } } + + return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned); +#endif +} + + + +/******* Hash streaming *******/ +#ifndef XXH_NO_STREAM +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void) +{ + return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t)); +} +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr) +{ + XXH_free(statePtr); + return XXH_OK; +} + +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState) +{ + XXH_memcpy(dstState, srcState, sizeof(*dstState)); +} + +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed) +{ + XXH_ASSERT(statePtr != NULL); + memset(statePtr, 0, sizeof(*statePtr)); + statePtr->v[0] = seed + XXH_PRIME32_1 + XXH_PRIME32_2; + statePtr->v[1] = seed + XXH_PRIME32_2; + statePtr->v[2] = seed + 0; + statePtr->v[3] = seed - XXH_PRIME32_1; + return XXH_OK; +} + + +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API XXH_errorcode +XXH32_update(XXH32_state_t* state, const void* input, size_t len) +{ + if (input==NULL) { + XXH_ASSERT(len == 0); + return XXH_OK; + } + + { const xxh_u8* p = (const xxh_u8*)input; + const xxh_u8* const bEnd = p + len; + + state->total_len_32 += (XXH32_hash_t)len; + state->large_len |= (XXH32_hash_t)((len>=16) | (state->total_len_32>=16)); + + if (state->memsize + len < 16) { /* fill in tmp buffer */ + XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, len); + state->memsize += (XXH32_hash_t)len; + return XXH_OK; + } + + if (state->memsize) { /* some data left from previous update */ + XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, 16-state->memsize); + { const xxh_u32* p32 = state->mem32; + state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p32)); p32++; + state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p32)); p32++; + state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p32)); p32++; + state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p32)); + } + p += 16-state->memsize; + state->memsize = 0; + } + + if (p <= bEnd-16) { + const xxh_u8* const limit = bEnd - 16; + + do { + state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p)); p+=4; + state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p)); p+=4; + state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p)); p+=4; + state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p)); p+=4; + } while (p<=limit); + + } + + if (p < bEnd) { + XXH_memcpy(state->mem32, p, (size_t)(bEnd-p)); + state->memsize = (unsigned)(bEnd-p); + } + } + + return XXH_OK; +} + + +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t* state) +{ + xxh_u32 h32; + + if (state->large_len) { + h32 = XXH_rotl32(state->v[0], 1) + + XXH_rotl32(state->v[1], 7) + + XXH_rotl32(state->v[2], 12) + + XXH_rotl32(state->v[3], 18); + } else { + h32 = state->v[2] /* == seed */ + XXH_PRIME32_5; + } + + h32 += state->total_len_32; + + return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned); +} +#endif /* !XXH_NO_STREAM */ + +/******* Canonical representation *******/ + +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash) +{ + XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t)); + if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash); + XXH_memcpy(dst, &hash, sizeof(*dst)); +} +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src) +{ + return XXH_readBE32(src); +} + + +#ifndef XXH_NO_LONG_LONG + +/* ******************************************************************* +* 64-bit hash functions +*********************************************************************/ +/*! + * @} + * @ingroup impl + * @{ + */ +/******* Memory access *******/ + +typedef XXH64_hash_t xxh_u64; + +#ifdef XXH_OLD_NAMES +# define U64 xxh_u64 +#endif + +#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3)) +/* + * Manual byteshift. Best for old compilers which don't inline memcpy. + * We actually directly use XXH_readLE64 and XXH_readBE64. + */ +#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2)) + +/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */ +static xxh_u64 XXH_read64(const void* memPtr) +{ + return *(const xxh_u64*) memPtr; +} + +#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1)) + +/* + * __attribute__((aligned(1))) is supported by gcc and clang. Originally the + * documentation claimed that it only increased the alignment, but actually it + * can decrease it on gcc, clang, and icc: + * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502, + * https://gcc.godbolt.org/z/xYez1j67Y. + */ +#ifdef XXH_OLD_NAMES +typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((__packed__)) unalign64; +#endif +static xxh_u64 XXH_read64(const void* ptr) +{ + typedef __attribute__((__aligned__(1))) xxh_u64 xxh_unalign64; + return *((const xxh_unalign64*)ptr); +} + +#else + +/* + * Portable and safe solution. Generally efficient. + * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html + */ +static xxh_u64 XXH_read64(const void* memPtr) +{ + xxh_u64 val; + XXH_memcpy(&val, memPtr, sizeof(val)); + return val; +} + +#endif /* XXH_FORCE_DIRECT_MEMORY_ACCESS */ + +#if defined(_MSC_VER) /* Visual Studio */ +# define XXH_swap64 _byteswap_uint64 +#elif XXH_GCC_VERSION >= 403 +# define XXH_swap64 __builtin_bswap64 +#else +static xxh_u64 XXH_swap64(xxh_u64 x) +{ + return ((x << 56) & 0xff00000000000000ULL) | + ((x << 40) & 0x00ff000000000000ULL) | + ((x << 24) & 0x0000ff0000000000ULL) | + ((x << 8) & 0x000000ff00000000ULL) | + ((x >> 8) & 0x00000000ff000000ULL) | + ((x >> 24) & 0x0000000000ff0000ULL) | + ((x >> 40) & 0x000000000000ff00ULL) | + ((x >> 56) & 0x00000000000000ffULL); +} +#endif + + +/* XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. */ +#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3)) + +XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* memPtr) +{ + const xxh_u8* bytePtr = (const xxh_u8 *)memPtr; + return bytePtr[0] + | ((xxh_u64)bytePtr[1] << 8) + | ((xxh_u64)bytePtr[2] << 16) + | ((xxh_u64)bytePtr[3] << 24) + | ((xxh_u64)bytePtr[4] << 32) + | ((xxh_u64)bytePtr[5] << 40) + | ((xxh_u64)bytePtr[6] << 48) + | ((xxh_u64)bytePtr[7] << 56); +} + +XXH_FORCE_INLINE xxh_u64 XXH_readBE64(const void* memPtr) +{ + const xxh_u8* bytePtr = (const xxh_u8 *)memPtr; + return bytePtr[7] + | ((xxh_u64)bytePtr[6] << 8) + | ((xxh_u64)bytePtr[5] << 16) + | ((xxh_u64)bytePtr[4] << 24) + | ((xxh_u64)bytePtr[3] << 32) + | ((xxh_u64)bytePtr[2] << 40) + | ((xxh_u64)bytePtr[1] << 48) + | ((xxh_u64)bytePtr[0] << 56); +} + +#else +XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* ptr) +{ + return XXH_CPU_LITTLE_ENDIAN ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr)); +} + +static xxh_u64 XXH_readBE64(const void* ptr) +{ + return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr); +} +#endif + +XXH_FORCE_INLINE xxh_u64 +XXH_readLE64_align(const void* ptr, XXH_alignment align) +{ + if (align==XXH_unaligned) + return XXH_readLE64(ptr); + else + return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u64*)ptr : XXH_swap64(*(const xxh_u64*)ptr); +} + + +/******* xxh64 *******/ +/*! + * @} + * @defgroup XXH64_impl XXH64 implementation + * @ingroup impl + * + * Details on the XXH64 implementation. + * @{ + */ +/* #define rather that static const, to be used as initializers */ +#define XXH_PRIME64_1 0x9E3779B185EBCA87ULL /*!< 0b1001111000110111011110011011000110000101111010111100101010000111 */ +#define XXH_PRIME64_2 0xC2B2AE3D27D4EB4FULL /*!< 0b1100001010110010101011100011110100100111110101001110101101001111 */ +#define XXH_PRIME64_3 0x165667B19E3779F9ULL /*!< 0b0001011001010110011001111011000110011110001101110111100111111001 */ +#define XXH_PRIME64_4 0x85EBCA77C2B2AE63ULL /*!< 0b1000010111101011110010100111011111000010101100101010111001100011 */ +#define XXH_PRIME64_5 0x27D4EB2F165667C5ULL /*!< 0b0010011111010100111010110010111100010110010101100110011111000101 */ + +#ifdef XXH_OLD_NAMES +# define PRIME64_1 XXH_PRIME64_1 +# define PRIME64_2 XXH_PRIME64_2 +# define PRIME64_3 XXH_PRIME64_3 +# define PRIME64_4 XXH_PRIME64_4 +# define PRIME64_5 XXH_PRIME64_5 +#endif + +/*! @copydoc XXH32_round */ +static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input) +{ + acc += input * XXH_PRIME64_2; + acc = XXH_rotl64(acc, 31); + acc *= XXH_PRIME64_1; +#if (defined(__AVX512F__)) && !defined(XXH_ENABLE_AUTOVECTORIZE) + /* + * DISABLE AUTOVECTORIZATION: + * A compiler fence is used to prevent GCC and Clang from + * autovectorizing the XXH64 loop (pragmas and attributes don't work for some + * reason) without globally disabling AVX512. + * + * Autovectorization of XXH64 tends to be detrimental, + * though the exact outcome may change depending on exact cpu and compiler version. + * For information, it has been reported as detrimental for Skylake-X, + * but possibly beneficial for Zen4. + * + * The default is to disable auto-vectorization, + * but you can select to enable it instead using `XXH_ENABLE_AUTOVECTORIZE` build variable. + */ + XXH_COMPILER_GUARD(acc); +#endif + return acc; +} + +static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val) +{ + val = XXH64_round(0, val); + acc ^= val; + acc = acc * XXH_PRIME64_1 + XXH_PRIME64_4; + return acc; +} + +/*! @copydoc XXH32_avalanche */ +static xxh_u64 XXH64_avalanche(xxh_u64 hash) +{ + hash ^= hash >> 33; + hash *= XXH_PRIME64_2; + hash ^= hash >> 29; + hash *= XXH_PRIME64_3; + hash ^= hash >> 32; + return hash; +} + + +#define XXH_get64bits(p) XXH_readLE64_align(p, align) + +/*! + * @internal + * @brief Processes the last 0-31 bytes of @p ptr. + * + * There may be up to 31 bytes remaining to consume from the input. + * This final stage will digest them to ensure that all input bytes are present + * in the final mix. + * + * @param hash The hash to finalize. + * @param ptr The pointer to the remaining input. + * @param len The remaining length, modulo 32. + * @param align Whether @p ptr is aligned. + * @return The finalized hash + * @see XXH32_finalize(). + */ +static XXH_PUREF xxh_u64 +XXH64_finalize(xxh_u64 hash, const xxh_u8* ptr, size_t len, XXH_alignment align) +{ + if (ptr==NULL) XXH_ASSERT(len == 0); + len &= 31; + while (len >= 8) { + xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr)); + ptr += 8; + hash ^= k1; + hash = XXH_rotl64(hash,27) * XXH_PRIME64_1 + XXH_PRIME64_4; + len -= 8; + } + if (len >= 4) { + hash ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1; + ptr += 4; + hash = XXH_rotl64(hash, 23) * XXH_PRIME64_2 + XXH_PRIME64_3; + len -= 4; + } + while (len > 0) { + hash ^= (*ptr++) * XXH_PRIME64_5; + hash = XXH_rotl64(hash, 11) * XXH_PRIME64_1; + --len; + } + return XXH64_avalanche(hash); +} + +#ifdef XXH_OLD_NAMES +# define PROCESS1_64 XXH_PROCESS1_64 +# define PROCESS4_64 XXH_PROCESS4_64 +# define PROCESS8_64 XXH_PROCESS8_64 +#else +# undef XXH_PROCESS1_64 +# undef XXH_PROCESS4_64 +# undef XXH_PROCESS8_64 +#endif + +/*! + * @internal + * @brief The implementation for @ref XXH64(). + * + * @param input , len , seed Directly passed from @ref XXH64(). + * @param align Whether @p input is aligned. + * @return The calculated hash. + */ +XXH_FORCE_INLINE XXH_PUREF xxh_u64 +XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align) +{ + xxh_u64 h64; + if (input==NULL) XXH_ASSERT(len == 0); + + if (len>=32) { + const xxh_u8* const bEnd = input + len; + const xxh_u8* const limit = bEnd - 31; + xxh_u64 v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2; + xxh_u64 v2 = seed + XXH_PRIME64_2; + xxh_u64 v3 = seed + 0; + xxh_u64 v4 = seed - XXH_PRIME64_1; + + do { + v1 = XXH64_round(v1, XXH_get64bits(input)); input+=8; + v2 = XXH64_round(v2, XXH_get64bits(input)); input+=8; + v3 = XXH64_round(v3, XXH_get64bits(input)); input+=8; + v4 = XXH64_round(v4, XXH_get64bits(input)); input+=8; + } while (input= 2 + /* Simple version, good for code maintenance, but unfortunately slow for small inputs */ + XXH64_state_t state; + XXH64_reset(&state, seed); + XXH64_update(&state, (const xxh_u8*)input, len); + return XXH64_digest(&state); +#else + if (XXH_FORCE_ALIGN_CHECK) { + if ((((size_t)input) & 7)==0) { /* Input is aligned, let's leverage the speed advantage */ + return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_aligned); + } } + + return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned); + +#endif +} + +/******* Hash Streaming *******/ +#ifndef XXH_NO_STREAM +/*! @ingroup XXH64_family*/ +XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void) +{ + return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t)); +} +/*! @ingroup XXH64_family */ +XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr) +{ + XXH_free(statePtr); + return XXH_OK; +} + +/*! @ingroup XXH64_family */ +XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dstState, const XXH64_state_t* srcState) +{ + XXH_memcpy(dstState, srcState, sizeof(*dstState)); +} + +/*! @ingroup XXH64_family */ +XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed) +{ + XXH_ASSERT(statePtr != NULL); + memset(statePtr, 0, sizeof(*statePtr)); + statePtr->v[0] = seed + XXH_PRIME64_1 + XXH_PRIME64_2; + statePtr->v[1] = seed + XXH_PRIME64_2; + statePtr->v[2] = seed + 0; + statePtr->v[3] = seed - XXH_PRIME64_1; + return XXH_OK; +} + +/*! @ingroup XXH64_family */ +XXH_PUBLIC_API XXH_errorcode +XXH64_update (XXH_NOESCAPE XXH64_state_t* state, XXH_NOESCAPE const void* input, size_t len) +{ + if (input==NULL) { + XXH_ASSERT(len == 0); + return XXH_OK; + } + + { const xxh_u8* p = (const xxh_u8*)input; + const xxh_u8* const bEnd = p + len; + + state->total_len += len; + + if (state->memsize + len < 32) { /* fill in tmp buffer */ + XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, len); + state->memsize += (xxh_u32)len; + return XXH_OK; + } + + if (state->memsize) { /* tmp buffer is full */ + XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, 32-state->memsize); + state->v[0] = XXH64_round(state->v[0], XXH_readLE64(state->mem64+0)); + state->v[1] = XXH64_round(state->v[1], XXH_readLE64(state->mem64+1)); + state->v[2] = XXH64_round(state->v[2], XXH_readLE64(state->mem64+2)); + state->v[3] = XXH64_round(state->v[3], XXH_readLE64(state->mem64+3)); + p += 32 - state->memsize; + state->memsize = 0; + } + + if (p+32 <= bEnd) { + const xxh_u8* const limit = bEnd - 32; + + do { + state->v[0] = XXH64_round(state->v[0], XXH_readLE64(p)); p+=8; + state->v[1] = XXH64_round(state->v[1], XXH_readLE64(p)); p+=8; + state->v[2] = XXH64_round(state->v[2], XXH_readLE64(p)); p+=8; + state->v[3] = XXH64_round(state->v[3], XXH_readLE64(p)); p+=8; + } while (p<=limit); + + } + + if (p < bEnd) { + XXH_memcpy(state->mem64, p, (size_t)(bEnd-p)); + state->memsize = (unsigned)(bEnd-p); + } + } + + return XXH_OK; +} + + +/*! @ingroup XXH64_family */ +XXH_PUBLIC_API XXH64_hash_t XXH64_digest(XXH_NOESCAPE const XXH64_state_t* state) +{ + xxh_u64 h64; + + if (state->total_len >= 32) { + h64 = XXH_rotl64(state->v[0], 1) + XXH_rotl64(state->v[1], 7) + XXH_rotl64(state->v[2], 12) + XXH_rotl64(state->v[3], 18); + h64 = XXH64_mergeRound(h64, state->v[0]); + h64 = XXH64_mergeRound(h64, state->v[1]); + h64 = XXH64_mergeRound(h64, state->v[2]); + h64 = XXH64_mergeRound(h64, state->v[3]); + } else { + h64 = state->v[2] /*seed*/ + XXH_PRIME64_5; + } + + h64 += (xxh_u64) state->total_len; + + return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned); +} +#endif /* !XXH_NO_STREAM */ + +/******* Canonical representation *******/ + +/*! @ingroup XXH64_family */ +XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash) +{ + XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t)); + if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash); + XXH_memcpy(dst, &hash, sizeof(*dst)); +} + +/*! @ingroup XXH64_family */ +XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src) +{ + return XXH_readBE64(src); +} + +#ifndef XXH_NO_XXH3 + +/* ********************************************************************* +* XXH3 +* New generation hash designed for speed on small keys and vectorization +************************************************************************ */ +/*! + * @} + * @defgroup XXH3_impl XXH3 implementation + * @ingroup impl + * @{ + */ + +/* === Compiler specifics === */ + +#if ((defined(sun) || defined(__sun)) && __cplusplus) /* Solaris includes __STDC_VERSION__ with C++. Tested with GCC 5.5 */ +# define XXH_RESTRICT /* disable */ +#elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L /* >= C99 */ +# define XXH_RESTRICT restrict +#elif (defined (__GNUC__) && ((__GNUC__ > 3) || (__GNUC__ == 3 && __GNUC_MINOR__ >= 1))) \ + || (defined (__clang__)) \ + || (defined (_MSC_VER) && (_MSC_VER >= 1400)) \ + || (defined (__INTEL_COMPILER) && (__INTEL_COMPILER >= 1300)) +/* + * There are a LOT more compilers that recognize __restrict but this + * covers the major ones. + */ +# define XXH_RESTRICT __restrict +#else +# define XXH_RESTRICT /* disable */ +#endif + +#if (defined(__GNUC__) && (__GNUC__ >= 3)) \ + || (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) \ + || defined(__clang__) +# define XXH_likely(x) __builtin_expect(x, 1) +# define XXH_unlikely(x) __builtin_expect(x, 0) +#else +# define XXH_likely(x) (x) +# define XXH_unlikely(x) (x) +#endif + +#ifndef XXH_HAS_INCLUDE +# ifdef __has_include +/* + * Not defined as XXH_HAS_INCLUDE(x) (function-like) because + * this causes segfaults in Apple Clang 4.2 (on Mac OS X 10.7 Lion) + */ +# define XXH_HAS_INCLUDE __has_include +# else +# define XXH_HAS_INCLUDE(x) 0 +# endif +#endif + +#if defined(__GNUC__) || defined(__clang__) +# if defined(__ARM_FEATURE_SVE) +# include +# endif +# if defined(__ARM_NEON__) || defined(__ARM_NEON) \ + || (defined(_M_ARM) && _M_ARM >= 7) \ + || defined(_M_ARM64) || defined(_M_ARM64EC) \ + || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE()) /* WASM SIMD128 via SIMDe */ +# define inline __inline__ /* circumvent a clang bug */ +# include +# undef inline +# elif defined(__AVX2__) +# include +# elif defined(__SSE2__) +# include +# endif +#endif + +#if defined(_MSC_VER) +# include +#endif + +/* + * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while + * remaining a true 64-bit/128-bit hash function. + * + * This is done by prioritizing a subset of 64-bit operations that can be + * emulated without too many steps on the average 32-bit machine. + * + * For example, these two lines seem similar, and run equally fast on 64-bit: + * + * xxh_u64 x; + * x ^= (x >> 47); // good + * x ^= (x >> 13); // bad + * + * However, to a 32-bit machine, there is a major difference. + * + * x ^= (x >> 47) looks like this: + * + * x.lo ^= (x.hi >> (47 - 32)); + * + * while x ^= (x >> 13) looks like this: + * + * // note: funnel shifts are not usually cheap. + * x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13)); + * x.hi ^= (x.hi >> 13); + * + * The first one is significantly faster than the second, simply because the + * shift is larger than 32. This means: + * - All the bits we need are in the upper 32 bits, so we can ignore the lower + * 32 bits in the shift. + * - The shift result will always fit in the lower 32 bits, and therefore, + * we can ignore the upper 32 bits in the xor. + * + * Thanks to this optimization, XXH3 only requires these features to be efficient: + * + * - Usable unaligned access + * - A 32-bit or 64-bit ALU + * - If 32-bit, a decent ADC instruction + * - A 32 or 64-bit multiply with a 64-bit result + * - For the 128-bit variant, a decent byteswap helps short inputs. + * + * The first two are already required by XXH32, and almost all 32-bit and 64-bit + * platforms which can run XXH32 can run XXH3 efficiently. + * + * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one + * notable exception. + * + * First of all, Thumb-1 lacks support for the UMULL instruction which + * performs the important long multiply. This means numerous __aeabi_lmul + * calls. + * + * Second of all, the 8 functional registers are just not enough. + * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need + * Lo registers, and this shuffling results in thousands more MOVs than A32. + * + * A32 and T32 don't have this limitation. They can access all 14 registers, + * do a 32->64 multiply with UMULL, and the flexible operand allowing free + * shifts is helpful, too. + * + * Therefore, we do a quick sanity check. + * + * If compiling Thumb-1 for a target which supports ARM instructions, we will + * emit a warning, as it is not a "sane" platform to compile for. + * + * Usually, if this happens, it is because of an accident and you probably need + * to specify -march, as you likely meant to compile for a newer architecture. + * + * Credit: large sections of the vectorial and asm source code paths + * have been contributed by @easyaspi314 + */ +#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM) +# warning "XXH3 is highly inefficient without ARM or Thumb-2." +#endif + +/* ========================================== + * Vectorization detection + * ========================================== */ + +#ifdef XXH_DOXYGEN +/*! + * @ingroup tuning + * @brief Overrides the vectorization implementation chosen for XXH3. + * + * Can be defined to 0 to disable SIMD or any of the values mentioned in + * @ref XXH_VECTOR_TYPE. + * + * If this is not defined, it uses predefined macros to determine the best + * implementation. + */ +# define XXH_VECTOR XXH_SCALAR +/*! + * @ingroup tuning + * @brief Possible values for @ref XXH_VECTOR. + * + * Note that these are actually implemented as macros. + * + * If this is not defined, it is detected automatically. + * internal macro XXH_X86DISPATCH overrides this. + */ +enum XXH_VECTOR_TYPE /* fake enum */ { + XXH_SCALAR = 0, /*!< Portable scalar version */ + XXH_SSE2 = 1, /*!< + * SSE2 for Pentium 4, Opteron, all x86_64. + * + * @note SSE2 is also guaranteed on Windows 10, macOS, and + * Android x86. + */ + XXH_AVX2 = 2, /*!< AVX2 for Haswell and Bulldozer */ + XXH_AVX512 = 3, /*!< AVX512 for Skylake and Icelake */ + XXH_NEON = 4, /*!< + * NEON for most ARMv7-A, all AArch64, and WASM SIMD128 + * via the SIMDeverywhere polyfill provided with the + * Emscripten SDK. + */ + XXH_VSX = 5, /*!< VSX and ZVector for POWER8/z13 (64-bit) */ + XXH_SVE = 6, /*!< SVE for some ARMv8-A and ARMv9-A */ +}; +/*! + * @ingroup tuning + * @brief Selects the minimum alignment for XXH3's accumulators. + * + * When using SIMD, this should match the alignment required for said vector + * type, so, for example, 32 for AVX2. + * + * Default: Auto detected. + */ +# define XXH_ACC_ALIGN 8 +#endif + +/* Actual definition */ +#ifndef XXH_DOXYGEN +# define XXH_SCALAR 0 +# define XXH_SSE2 1 +# define XXH_AVX2 2 +# define XXH_AVX512 3 +# define XXH_NEON 4 +# define XXH_VSX 5 +# define XXH_SVE 6 +#endif + +#ifndef XXH_VECTOR /* can be defined on command line */ +# if defined(__ARM_FEATURE_SVE) +# define XXH_VECTOR XXH_SVE +# elif ( \ + defined(__ARM_NEON__) || defined(__ARM_NEON) /* gcc */ \ + || defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM64EC) /* msvc */ \ + || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE()) /* wasm simd128 via SIMDe */ \ + ) && ( \ + defined(_WIN32) || defined(__LITTLE_ENDIAN__) /* little endian only */ \ + || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \ + ) +# define XXH_VECTOR XXH_NEON +# elif defined(__AVX512F__) +# define XXH_VECTOR XXH_AVX512 +# elif defined(__AVX2__) +# define XXH_VECTOR XXH_AVX2 +# elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2)) +# define XXH_VECTOR XXH_SSE2 +# elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \ + || (defined(__s390x__) && defined(__VEC__)) \ + && defined(__GNUC__) /* TODO: IBM XL */ +# define XXH_VECTOR XXH_VSX +# else +# define XXH_VECTOR XXH_SCALAR +# endif +#endif + +/* __ARM_FEATURE_SVE is only supported by GCC & Clang. */ +#if (XXH_VECTOR == XXH_SVE) && !defined(__ARM_FEATURE_SVE) +# ifdef _MSC_VER +# pragma warning(once : 4606) +# else +# warning "__ARM_FEATURE_SVE isn't supported. Use SCALAR instead." +# endif +# undef XXH_VECTOR +# define XXH_VECTOR XXH_SCALAR +#endif + +/* + * Controls the alignment of the accumulator, + * for compatibility with aligned vector loads, which are usually faster. + */ +#ifndef XXH_ACC_ALIGN +# if defined(XXH_X86DISPATCH) +# define XXH_ACC_ALIGN 64 /* for compatibility with avx512 */ +# elif XXH_VECTOR == XXH_SCALAR /* scalar */ +# define XXH_ACC_ALIGN 8 +# elif XXH_VECTOR == XXH_SSE2 /* sse2 */ +# define XXH_ACC_ALIGN 16 +# elif XXH_VECTOR == XXH_AVX2 /* avx2 */ +# define XXH_ACC_ALIGN 32 +# elif XXH_VECTOR == XXH_NEON /* neon */ +# define XXH_ACC_ALIGN 16 +# elif XXH_VECTOR == XXH_VSX /* vsx */ +# define XXH_ACC_ALIGN 16 +# elif XXH_VECTOR == XXH_AVX512 /* avx512 */ +# define XXH_ACC_ALIGN 64 +# elif XXH_VECTOR == XXH_SVE /* sve */ +# define XXH_ACC_ALIGN 64 +# endif +#endif + +#if defined(XXH_X86DISPATCH) || XXH_VECTOR == XXH_SSE2 \ + || XXH_VECTOR == XXH_AVX2 || XXH_VECTOR == XXH_AVX512 +# define XXH_SEC_ALIGN XXH_ACC_ALIGN +#elif XXH_VECTOR == XXH_SVE +# define XXH_SEC_ALIGN XXH_ACC_ALIGN +#else +# define XXH_SEC_ALIGN 8 +#endif + +#if defined(__GNUC__) || defined(__clang__) +# define XXH_ALIASING __attribute__((__may_alias__)) +#else +# define XXH_ALIASING /* nothing */ +#endif + +/* + * UGLY HACK: + * GCC usually generates the best code with -O3 for xxHash. + * + * However, when targeting AVX2, it is overzealous in its unrolling resulting + * in code roughly 3/4 the speed of Clang. + * + * There are other issues, such as GCC splitting _mm256_loadu_si256 into + * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which + * only applies to Sandy and Ivy Bridge... which don't even support AVX2. + * + * That is why when compiling the AVX2 version, it is recommended to use either + * -O2 -mavx2 -march=haswell + * or + * -O2 -mavx2 -mno-avx256-split-unaligned-load + * for decent performance, or to use Clang instead. + * + * Fortunately, we can control the first one with a pragma that forces GCC into + * -O2, but the other one we can't control without "failed to inline always + * inline function due to target mismatch" warnings. + */ +#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \ + && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ + && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */ +# pragma GCC push_options +# pragma GCC optimize("-O2") +#endif + +#if XXH_VECTOR == XXH_NEON + +/* + * UGLY HACK: While AArch64 GCC on Linux does not seem to care, on macOS, GCC -O3 + * optimizes out the entire hashLong loop because of the aliasing violation. + * + * However, GCC is also inefficient at load-store optimization with vld1q/vst1q, + * so the only option is to mark it as aliasing. + */ +typedef uint64x2_t xxh_aliasing_uint64x2_t XXH_ALIASING; + +/*! + * @internal + * @brief `vld1q_u64` but faster and alignment-safe. + * + * On AArch64, unaligned access is always safe, but on ARMv7-a, it is only + * *conditionally* safe (`vld1` has an alignment bit like `movdq[ua]` in x86). + * + * GCC for AArch64 sees `vld1q_u8` as an intrinsic instead of a load, so it + * prohibits load-store optimizations. Therefore, a direct dereference is used. + * + * Otherwise, `vld1q_u8` is used with `vreinterpretq_u8_u64` to do a safe + * unaligned load. + */ +#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__) +XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr) /* silence -Wcast-align */ +{ + return *(xxh_aliasing_uint64x2_t const *)ptr; +} +#else +XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr) +{ + return vreinterpretq_u64_u8(vld1q_u8((uint8_t const*)ptr)); +} +#endif + +/*! + * @internal + * @brief `vmlal_u32` on low and high halves of a vector. + * + * This is a workaround for AArch64 GCC < 11 which implemented arm_neon.h with + * inline assembly and were therefore incapable of merging the `vget_{low, high}_u32` + * with `vmlal_u32`. + */ +#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 11 +XXH_FORCE_INLINE uint64x2_t +XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs) +{ + /* Inline assembly is the only way */ + __asm__("umlal %0.2d, %1.2s, %2.2s" : "+w" (acc) : "w" (lhs), "w" (rhs)); + return acc; +} +XXH_FORCE_INLINE uint64x2_t +XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs) +{ + /* This intrinsic works as expected */ + return vmlal_high_u32(acc, lhs, rhs); +} +#else +/* Portable intrinsic versions */ +XXH_FORCE_INLINE uint64x2_t +XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs) +{ + return vmlal_u32(acc, vget_low_u32(lhs), vget_low_u32(rhs)); +} +/*! @copydoc XXH_vmlal_low_u32 + * Assume the compiler converts this to vmlal_high_u32 on aarch64 */ +XXH_FORCE_INLINE uint64x2_t +XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs) +{ + return vmlal_u32(acc, vget_high_u32(lhs), vget_high_u32(rhs)); +} +#endif + +/*! + * @ingroup tuning + * @brief Controls the NEON to scalar ratio for XXH3 + * + * This can be set to 2, 4, 6, or 8. + * + * ARM Cortex CPUs are _very_ sensitive to how their pipelines are used. + * + * For example, the Cortex-A73 can dispatch 3 micro-ops per cycle, but only 2 of those + * can be NEON. If you are only using NEON instructions, you are only using 2/3 of the CPU + * bandwidth. + * + * This is even more noticeable on the more advanced cores like the Cortex-A76 which + * can dispatch 8 micro-ops per cycle, but still only 2 NEON micro-ops at once. + * + * Therefore, to make the most out of the pipeline, it is beneficial to run 6 NEON lanes + * and 2 scalar lanes, which is chosen by default. + * + * This does not apply to Apple processors or 32-bit processors, which run better with + * full NEON. These will default to 8. Additionally, size-optimized builds run 8 lanes. + * + * This change benefits CPUs with large micro-op buffers without negatively affecting + * most other CPUs: + * + * | Chipset | Dispatch type | NEON only | 6:2 hybrid | Diff. | + * |:----------------------|:--------------------|----------:|-----------:|------:| + * | Snapdragon 730 (A76) | 2 NEON/8 micro-ops | 8.8 GB/s | 10.1 GB/s | ~16% | + * | Snapdragon 835 (A73) | 2 NEON/3 micro-ops | 5.1 GB/s | 5.3 GB/s | ~5% | + * | Marvell PXA1928 (A53) | In-order dual-issue | 1.9 GB/s | 1.9 GB/s | 0% | + * | Apple M1 | 4 NEON/8 micro-ops | 37.3 GB/s | 36.1 GB/s | ~-3% | + * + * It also seems to fix some bad codegen on GCC, making it almost as fast as clang. + * + * When using WASM SIMD128, if this is 2 or 6, SIMDe will scalarize 2 of the lanes meaning + * it effectively becomes worse 4. + * + * @see XXH3_accumulate_512_neon() + */ +# ifndef XXH3_NEON_LANES +# if (defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || defined(_M_ARM64EC)) \ + && !defined(__APPLE__) && XXH_SIZE_OPT <= 0 +# define XXH3_NEON_LANES 6 +# else +# define XXH3_NEON_LANES XXH_ACC_NB +# endif +# endif +#endif /* XXH_VECTOR == XXH_NEON */ + +/* + * VSX and Z Vector helpers. + * + * This is very messy, and any pull requests to clean this up are welcome. + * + * There are a lot of problems with supporting VSX and s390x, due to + * inconsistent intrinsics, spotty coverage, and multiple endiannesses. + */ +#if XXH_VECTOR == XXH_VSX +/* Annoyingly, these headers _may_ define three macros: `bool`, `vector`, + * and `pixel`. This is a problem for obvious reasons. + * + * These keywords are unnecessary; the spec literally says they are + * equivalent to `__bool`, `__vector`, and `__pixel` and may be undef'd + * after including the header. + * + * We use pragma push_macro/pop_macro to keep the namespace clean. */ +# pragma push_macro("bool") +# pragma push_macro("vector") +# pragma push_macro("pixel") +/* silence potential macro redefined warnings */ +# undef bool +# undef vector +# undef pixel + +# if defined(__s390x__) +# include +# else +# include +# endif + +/* Restore the original macro values, if applicable. */ +# pragma pop_macro("pixel") +# pragma pop_macro("vector") +# pragma pop_macro("bool") + +typedef __vector unsigned long long xxh_u64x2; +typedef __vector unsigned char xxh_u8x16; +typedef __vector unsigned xxh_u32x4; + +/* + * UGLY HACK: Similar to aarch64 macOS GCC, s390x GCC has the same aliasing issue. + */ +typedef xxh_u64x2 xxh_aliasing_u64x2 XXH_ALIASING; + +# ifndef XXH_VSX_BE +# if defined(__BIG_ENDIAN__) \ + || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) +# define XXH_VSX_BE 1 +# elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__ +# warning "-maltivec=be is not recommended. Please use native endianness." +# define XXH_VSX_BE 1 +# else +# define XXH_VSX_BE 0 +# endif +# endif /* !defined(XXH_VSX_BE) */ + +# if XXH_VSX_BE +# if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__)) +# define XXH_vec_revb vec_revb +# else +/*! + * A polyfill for POWER9's vec_revb(). + */ +XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val) +{ + xxh_u8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00, + 0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 }; + return vec_perm(val, val, vByteSwap); +} +# endif +# endif /* XXH_VSX_BE */ + +/*! + * Performs an unaligned vector load and byte swaps it on big endian. + */ +XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr) +{ + xxh_u64x2 ret; + XXH_memcpy(&ret, ptr, sizeof(xxh_u64x2)); +# if XXH_VSX_BE + ret = XXH_vec_revb(ret); +# endif + return ret; +} + +/* + * vec_mulo and vec_mule are very problematic intrinsics on PowerPC + * + * These intrinsics weren't added until GCC 8, despite existing for a while, + * and they are endian dependent. Also, their meaning swap depending on version. + * */ +# if defined(__s390x__) + /* s390x is always big endian, no issue on this platform */ +# define XXH_vec_mulo vec_mulo +# define XXH_vec_mule vec_mule +# elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw) && !defined(__ibmxl__) +/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */ + /* The IBM XL Compiler (which defined __clang__) only implements the vec_* operations */ +# define XXH_vec_mulo __builtin_altivec_vmulouw +# define XXH_vec_mule __builtin_altivec_vmuleuw +# else +/* gcc needs inline assembly */ +/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */ +XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b) +{ + xxh_u64x2 result; + __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b)); + return result; +} +XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b) +{ + xxh_u64x2 result; + __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b)); + return result; +} +# endif /* XXH_vec_mulo, XXH_vec_mule */ +#endif /* XXH_VECTOR == XXH_VSX */ + +#if XXH_VECTOR == XXH_SVE +#define ACCRND(acc, offset) \ +do { \ + svuint64_t input_vec = svld1_u64(mask, xinput + offset); \ + svuint64_t secret_vec = svld1_u64(mask, xsecret + offset); \ + svuint64_t mixed = sveor_u64_x(mask, secret_vec, input_vec); \ + svuint64_t swapped = svtbl_u64(input_vec, kSwap); \ + svuint64_t mixed_lo = svextw_u64_x(mask, mixed); \ + svuint64_t mixed_hi = svlsr_n_u64_x(mask, mixed, 32); \ + svuint64_t mul = svmad_u64_x(mask, mixed_lo, mixed_hi, swapped); \ + acc = svadd_u64_x(mask, acc, mul); \ +} while (0) +#endif /* XXH_VECTOR == XXH_SVE */ + +/* prefetch + * can be disabled, by declaring XXH_NO_PREFETCH build macro */ +#if defined(XXH_NO_PREFETCH) +# define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */ +#else +# if XXH_SIZE_OPT >= 1 +# define XXH_PREFETCH(ptr) (void)(ptr) +# elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86)) /* _mm_prefetch() not defined outside of x86/x64 */ +# include /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */ +# define XXH_PREFETCH(ptr) _mm_prefetch((const char*)(ptr), _MM_HINT_T0) +# elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) ) +# define XXH_PREFETCH(ptr) __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */) +# else +# define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */ +# endif +#endif /* XXH_NO_PREFETCH */ + + +/* ========================================== + * XXH3 default settings + * ========================================== */ + +#define XXH_SECRET_DEFAULT_SIZE 192 /* minimum XXH3_SECRET_SIZE_MIN */ + +#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN) +# error "default keyset is not large enough" +#endif + +/*! Pseudorandom secret taken directly from FARSH. */ +XXH_ALIGN(64) static const xxh_u8 XXH3_kSecret[XXH_SECRET_DEFAULT_SIZE] = { + 0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c, + 0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f, + 0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21, + 0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c, + 0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3, + 0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8, + 0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d, + 0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64, + 0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb, + 0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e, + 0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce, + 0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e, +}; + +static const xxh_u64 PRIME_MX1 = 0x165667919E3779F9ULL; /*!< 0b0001011001010110011001111001000110011110001101110111100111111001 */ +static const xxh_u64 PRIME_MX2 = 0x9FB21C651E98DF25ULL; /*!< 0b1001111110110010000111000110010100011110100110001101111100100101 */ + +#ifdef XXH_OLD_NAMES +# define kSecret XXH3_kSecret +#endif + +#ifdef XXH_DOXYGEN +/*! + * @brief Calculates a 32-bit to 64-bit long multiply. + * + * Implemented as a macro. + * + * Wraps `__emulu` on MSVC x86 because it tends to call `__allmul` when it doesn't + * need to (but it shouldn't need to anyways, it is about 7 instructions to do + * a 64x64 multiply...). Since we know that this will _always_ emit `MULL`, we + * use that instead of the normal method. + * + * If you are compiling for platforms like Thumb-1 and don't have a better option, + * you may also want to write your own long multiply routine here. + * + * @param x, y Numbers to be multiplied + * @return 64-bit product of the low 32 bits of @p x and @p y. + */ +XXH_FORCE_INLINE xxh_u64 +XXH_mult32to64(xxh_u64 x, xxh_u64 y) +{ + return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF); +} +#elif defined(_MSC_VER) && defined(_M_IX86) +# define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y)) +#else +/* + * Downcast + upcast is usually better than masking on older compilers like + * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers. + * + * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands + * and perform a full 64x64 multiply -- entirely redundant on 32-bit. + */ +# define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y)) +#endif + +/*! + * @brief Calculates a 64->128-bit long multiply. + * + * Uses `__uint128_t` and `_umul128` if available, otherwise uses a scalar + * version. + * + * @param lhs , rhs The 64-bit integers to be multiplied + * @return The 128-bit result represented in an @ref XXH128_hash_t. + */ +static XXH128_hash_t +XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs) +{ + /* + * GCC/Clang __uint128_t method. + * + * On most 64-bit targets, GCC and Clang define a __uint128_t type. + * This is usually the best way as it usually uses a native long 64-bit + * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64. + * + * Usually. + * + * Despite being a 32-bit platform, Clang (and emscripten) define this type + * despite not having the arithmetic for it. This results in a laggy + * compiler builtin call which calculates a full 128-bit multiply. + * In that case it is best to use the portable one. + * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677 + */ +#if (defined(__GNUC__) || defined(__clang__)) && !defined(__wasm__) \ + && defined(__SIZEOF_INT128__) \ + || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) + + __uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs; + XXH128_hash_t r128; + r128.low64 = (xxh_u64)(product); + r128.high64 = (xxh_u64)(product >> 64); + return r128; + + /* + * MSVC for x64's _umul128 method. + * + * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct); + * + * This compiles to single operand MUL on x64. + */ +#elif (defined(_M_X64) || defined(_M_IA64)) && !defined(_M_ARM64EC) + +#ifndef _MSC_VER +# pragma intrinsic(_umul128) +#endif + xxh_u64 product_high; + xxh_u64 const product_low = _umul128(lhs, rhs, &product_high); + XXH128_hash_t r128; + r128.low64 = product_low; + r128.high64 = product_high; + return r128; + + /* + * MSVC for ARM64's __umulh method. + * + * This compiles to the same MUL + UMULH as GCC/Clang's __uint128_t method. + */ +#elif defined(_M_ARM64) || defined(_M_ARM64EC) + +#ifndef _MSC_VER +# pragma intrinsic(__umulh) +#endif + XXH128_hash_t r128; + r128.low64 = lhs * rhs; + r128.high64 = __umulh(lhs, rhs); + return r128; + +#else + /* + * Portable scalar method. Optimized for 32-bit and 64-bit ALUs. + * + * This is a fast and simple grade school multiply, which is shown below + * with base 10 arithmetic instead of base 0x100000000. + * + * 9 3 // D2 lhs = 93 + * x 7 5 // D2 rhs = 75 + * ---------- + * 1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15 + * 4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45 + * 2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21 + * + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63 + * --------- + * 2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27 + * + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67 + * --------- + * 6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975 + * + * The reasons for adding the products like this are: + * 1. It avoids manual carry tracking. Just like how + * (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX. + * This avoids a lot of complexity. + * + * 2. It hints for, and on Clang, compiles to, the powerful UMAAL + * instruction available in ARM's Digital Signal Processing extension + * in 32-bit ARMv6 and later, which is shown below: + * + * void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm) + * { + * xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm; + * *RdLo = (xxh_u32)(product & 0xFFFFFFFF); + * *RdHi = (xxh_u32)(product >> 32); + * } + * + * This instruction was designed for efficient long multiplication, and + * allows this to be calculated in only 4 instructions at speeds + * comparable to some 64-bit ALUs. + * + * 3. It isn't terrible on other platforms. Usually this will be a couple + * of 32-bit ADD/ADCs. + */ + + /* First calculate all of the cross products. */ + xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF); + xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32, rhs & 0xFFFFFFFF); + xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32); + xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32, rhs >> 32); + + /* Now add the products together. These will never overflow. */ + xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi; + xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32) + hi_hi; + xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF); + + XXH128_hash_t r128; + r128.low64 = lower; + r128.high64 = upper; + return r128; +#endif +} + +/*! + * @brief Calculates a 64-bit to 128-bit multiply, then XOR folds it. + * + * The reason for the separate function is to prevent passing too many structs + * around by value. This will hopefully inline the multiply, but we don't force it. + * + * @param lhs , rhs The 64-bit integers to multiply + * @return The low 64 bits of the product XOR'd by the high 64 bits. + * @see XXH_mult64to128() + */ +static xxh_u64 +XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs) +{ + XXH128_hash_t product = XXH_mult64to128(lhs, rhs); + return product.low64 ^ product.high64; +} + +/*! Seems to produce slightly better code on GCC for some reason. */ +XXH_FORCE_INLINE XXH_CONSTF xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift) +{ + XXH_ASSERT(0 <= shift && shift < 64); + return v64 ^ (v64 >> shift); +} + +/* + * This is a fast avalanche stage, + * suitable when input bits are already partially mixed + */ +static XXH64_hash_t XXH3_avalanche(xxh_u64 h64) +{ + h64 = XXH_xorshift64(h64, 37); + h64 *= PRIME_MX1; + h64 = XXH_xorshift64(h64, 32); + return h64; +} + +/* + * This is a stronger avalanche, + * inspired by Pelle Evensen's rrmxmx + * preferable when input has not been previously mixed + */ +static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len) +{ + /* this mix is inspired by Pelle Evensen's rrmxmx */ + h64 ^= XXH_rotl64(h64, 49) ^ XXH_rotl64(h64, 24); + h64 *= PRIME_MX2; + h64 ^= (h64 >> 35) + len ; + h64 *= PRIME_MX2; + return XXH_xorshift64(h64, 28); +} + + +/* ========================================== + * Short keys + * ========================================== + * One of the shortcomings of XXH32 and XXH64 was that their performance was + * sub-optimal on short lengths. It used an iterative algorithm which strongly + * favored lengths that were a multiple of 4 or 8. + * + * Instead of iterating over individual inputs, we use a set of single shot + * functions which piece together a range of lengths and operate in constant time. + * + * Additionally, the number of multiplies has been significantly reduced. This + * reduces latency, especially when emulating 64-bit multiplies on 32-bit. + * + * Depending on the platform, this may or may not be faster than XXH32, but it + * is almost guaranteed to be faster than XXH64. + */ + +/* + * At very short lengths, there isn't enough input to fully hide secrets, or use + * the entire secret. + * + * There is also only a limited amount of mixing we can do before significantly + * impacting performance. + * + * Therefore, we use different sections of the secret and always mix two secret + * samples with an XOR. This should have no effect on performance on the + * seedless or withSeed variants because everything _should_ be constant folded + * by modern compilers. + * + * The XOR mixing hides individual parts of the secret and increases entropy. + * + * This adds an extra layer of strength for custom secrets. + */ +XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t +XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(1 <= len && len <= 3); + XXH_ASSERT(secret != NULL); + /* + * len = 1: combined = { input[0], 0x01, input[0], input[0] } + * len = 2: combined = { input[1], 0x02, input[0], input[1] } + * len = 3: combined = { input[2], 0x03, input[0], input[1] } + */ + { xxh_u8 const c1 = input[0]; + xxh_u8 const c2 = input[len >> 1]; + xxh_u8 const c3 = input[len - 1]; + xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2 << 24) + | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8); + xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed; + xxh_u64 const keyed = (xxh_u64)combined ^ bitflip; + return XXH64_avalanche(keyed); + } +} + +XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t +XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(4 <= len && len <= 8); + seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32; + { xxh_u32 const input1 = XXH_readLE32(input); + xxh_u32 const input2 = XXH_readLE32(input + len - 4); + xxh_u64 const bitflip = (XXH_readLE64(secret+8) ^ XXH_readLE64(secret+16)) - seed; + xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32); + xxh_u64 const keyed = input64 ^ bitflip; + return XXH3_rrmxmx(keyed, len); + } +} + +XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t +XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(9 <= len && len <= 16); + { xxh_u64 const bitflip1 = (XXH_readLE64(secret+24) ^ XXH_readLE64(secret+32)) + seed; + xxh_u64 const bitflip2 = (XXH_readLE64(secret+40) ^ XXH_readLE64(secret+48)) - seed; + xxh_u64 const input_lo = XXH_readLE64(input) ^ bitflip1; + xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2; + xxh_u64 const acc = len + + XXH_swap64(input_lo) + input_hi + + XXH3_mul128_fold64(input_lo, input_hi); + return XXH3_avalanche(acc); + } +} + +XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t +XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(len <= 16); + { if (XXH_likely(len > 8)) return XXH3_len_9to16_64b(input, len, secret, seed); + if (XXH_likely(len >= 4)) return XXH3_len_4to8_64b(input, len, secret, seed); + if (len) return XXH3_len_1to3_64b(input, len, secret, seed); + return XXH64_avalanche(seed ^ (XXH_readLE64(secret+56) ^ XXH_readLE64(secret+64))); + } +} + +/* + * DISCLAIMER: There are known *seed-dependent* multicollisions here due to + * multiplication by zero, affecting hashes of lengths 17 to 240. + * + * However, they are very unlikely. + * + * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all + * unseeded non-cryptographic hashes, it does not attempt to defend itself + * against specially crafted inputs, only random inputs. + * + * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes + * cancelling out the secret is taken an arbitrary number of times (addressed + * in XXH3_accumulate_512), this collision is very unlikely with random inputs + * and/or proper seeding: + * + * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a + * function that is only called up to 16 times per hash with up to 240 bytes of + * input. + * + * This is not too bad for a non-cryptographic hash function, especially with + * only 64 bit outputs. + * + * The 128-bit variant (which trades some speed for strength) is NOT affected + * by this, although it is always a good idea to use a proper seed if you care + * about strength. + */ +XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input, + const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64) +{ +#if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ + && defined(__i386__) && defined(__SSE2__) /* x86 + SSE2 */ \ + && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable like XXH32 hack */ + /* + * UGLY HACK: + * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in + * slower code. + * + * By forcing seed64 into a register, we disrupt the cost model and + * cause it to scalarize. See `XXH32_round()` + * + * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600, + * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on + * GCC 9.2, despite both emitting scalar code. + * + * GCC generates much better scalar code than Clang for the rest of XXH3, + * which is why finding a more optimal codepath is an interest. + */ + XXH_COMPILER_GUARD(seed64); +#endif + { xxh_u64 const input_lo = XXH_readLE64(input); + xxh_u64 const input_hi = XXH_readLE64(input+8); + return XXH3_mul128_fold64( + input_lo ^ (XXH_readLE64(secret) + seed64), + input_hi ^ (XXH_readLE64(secret+8) - seed64) + ); + } +} + +/* For mid range keys, XXH3 uses a Mum-hash variant. */ +XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t +XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH64_hash_t seed) +{ + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; + XXH_ASSERT(16 < len && len <= 128); + + { xxh_u64 acc = len * XXH_PRIME64_1; +#if XXH_SIZE_OPT >= 1 + /* Smaller and cleaner, but slightly slower. */ + unsigned int i = (unsigned int)(len - 1) / 32; + do { + acc += XXH3_mix16B(input+16 * i, secret+32*i, seed); + acc += XXH3_mix16B(input+len-16*(i+1), secret+32*i+16, seed); + } while (i-- != 0); +#else + if (len > 32) { + if (len > 64) { + if (len > 96) { + acc += XXH3_mix16B(input+48, secret+96, seed); + acc += XXH3_mix16B(input+len-64, secret+112, seed); + } + acc += XXH3_mix16B(input+32, secret+64, seed); + acc += XXH3_mix16B(input+len-48, secret+80, seed); + } + acc += XXH3_mix16B(input+16, secret+32, seed); + acc += XXH3_mix16B(input+len-32, secret+48, seed); + } + acc += XXH3_mix16B(input+0, secret+0, seed); + acc += XXH3_mix16B(input+len-16, secret+16, seed); +#endif + return XXH3_avalanche(acc); + } +} + +XXH_NO_INLINE XXH_PUREF XXH64_hash_t +XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH64_hash_t seed) +{ + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; + XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); + + #define XXH3_MIDSIZE_STARTOFFSET 3 + #define XXH3_MIDSIZE_LASTOFFSET 17 + + { xxh_u64 acc = len * XXH_PRIME64_1; + xxh_u64 acc_end; + unsigned int const nbRounds = (unsigned int)len / 16; + unsigned int i; + XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); + for (i=0; i<8; i++) { + acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed); + } + /* last bytes */ + acc_end = XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed); + XXH_ASSERT(nbRounds >= 8); + acc = XXH3_avalanche(acc); +#if defined(__clang__) /* Clang */ \ + && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \ + && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable */ + /* + * UGLY HACK: + * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86. + * In everywhere else, it uses scalar code. + * + * For 64->128-bit multiplies, even if the NEON was 100% optimal, it + * would still be slower than UMAAL (see XXH_mult64to128). + * + * Unfortunately, Clang doesn't handle the long multiplies properly and + * converts them to the nonexistent "vmulq_u64" intrinsic, which is then + * scalarized into an ugly mess of VMOV.32 instructions. + * + * This mess is difficult to avoid without turning autovectorization + * off completely, but they are usually relatively minor and/or not + * worth it to fix. + * + * This loop is the easiest to fix, as unlike XXH32, this pragma + * _actually works_ because it is a loop vectorization instead of an + * SLP vectorization. + */ + #pragma clang loop vectorize(disable) +#endif + for (i=8 ; i < nbRounds; i++) { + /* + * Prevents clang for unrolling the acc loop and interleaving with this one. + */ + XXH_COMPILER_GUARD(acc); + acc_end += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed); + } + return XXH3_avalanche(acc + acc_end); + } +} + + +/* ======= Long Keys ======= */ + +#define XXH_STRIPE_LEN 64 +#define XXH_SECRET_CONSUME_RATE 8 /* nb of secret bytes consumed at each accumulation */ +#define XXH_ACC_NB (XXH_STRIPE_LEN / sizeof(xxh_u64)) + +#ifdef XXH_OLD_NAMES +# define STRIPE_LEN XXH_STRIPE_LEN +# define ACC_NB XXH_ACC_NB +#endif + +#ifndef XXH_PREFETCH_DIST +# ifdef __clang__ +# define XXH_PREFETCH_DIST 320 +# else +# if (XXH_VECTOR == XXH_AVX512) +# define XXH_PREFETCH_DIST 512 +# else +# define XXH_PREFETCH_DIST 384 +# endif +# endif /* __clang__ */ +#endif /* XXH_PREFETCH_DIST */ + +/* + * These macros are to generate an XXH3_accumulate() function. + * The two arguments select the name suffix and target attribute. + * + * The name of this symbol is XXH3_accumulate_() and it calls + * XXH3_accumulate_512_(). + * + * It may be useful to hand implement this function if the compiler fails to + * optimize the inline function. + */ +#define XXH3_ACCUMULATE_TEMPLATE(name) \ +void \ +XXH3_accumulate_##name(xxh_u64* XXH_RESTRICT acc, \ + const xxh_u8* XXH_RESTRICT input, \ + const xxh_u8* XXH_RESTRICT secret, \ + size_t nbStripes) \ +{ \ + size_t n; \ + for (n = 0; n < nbStripes; n++ ) { \ + const xxh_u8* const in = input + n*XXH_STRIPE_LEN; \ + XXH_PREFETCH(in + XXH_PREFETCH_DIST); \ + XXH3_accumulate_512_##name( \ + acc, \ + in, \ + secret + n*XXH_SECRET_CONSUME_RATE); \ + } \ +} + + +XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64) +{ + if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64); + XXH_memcpy(dst, &v64, sizeof(v64)); +} + +/* Several intrinsic functions below are supposed to accept __int64 as argument, + * as documented in https://software.intel.com/sites/landingpage/IntrinsicsGuide/ . + * However, several environments do not define __int64 type, + * requiring a workaround. + */ +#if !defined (__VMS) \ + && (defined (__cplusplus) \ + || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) + typedef int64_t xxh_i64; +#else + /* the following type must have a width of 64-bit */ + typedef long long xxh_i64; +#endif + + +/* + * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized. + * + * It is a hardened version of UMAC, based off of FARSH's implementation. + * + * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD + * implementations, and it is ridiculously fast. + * + * We harden it by mixing the original input to the accumulators as well as the product. + * + * This means that in the (relatively likely) case of a multiply by zero, the + * original input is preserved. + * + * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve + * cross-pollination, as otherwise the upper and lower halves would be + * essentially independent. + * + * This doesn't matter on 64-bit hashes since they all get merged together in + * the end, so we skip the extra step. + * + * Both XXH3_64bits and XXH3_128bits use this subroutine. + */ + +#if (XXH_VECTOR == XXH_AVX512) \ + || (defined(XXH_DISPATCH_AVX512) && XXH_DISPATCH_AVX512 != 0) + +#ifndef XXH_TARGET_AVX512 +# define XXH_TARGET_AVX512 /* disable attribute target */ +#endif + +XXH_FORCE_INLINE XXH_TARGET_AVX512 void +XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + __m512i* const xacc = (__m512i *) acc; + XXH_ASSERT((((size_t)acc) & 63) == 0); + XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i)); + + { + /* data_vec = input[0]; */ + __m512i const data_vec = _mm512_loadu_si512 (input); + /* key_vec = secret[0]; */ + __m512i const key_vec = _mm512_loadu_si512 (secret); + /* data_key = data_vec ^ key_vec; */ + __m512i const data_key = _mm512_xor_si512 (data_vec, key_vec); + /* data_key_lo = data_key >> 32; */ + __m512i const data_key_lo = _mm512_srli_epi64 (data_key, 32); + /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */ + __m512i const product = _mm512_mul_epu32 (data_key, data_key_lo); + /* xacc[0] += swap(data_vec); */ + __m512i const data_swap = _mm512_shuffle_epi32(data_vec, (_MM_PERM_ENUM)_MM_SHUFFLE(1, 0, 3, 2)); + __m512i const sum = _mm512_add_epi64(*xacc, data_swap); + /* xacc[0] += product; */ + *xacc = _mm512_add_epi64(product, sum); + } +} +XXH_FORCE_INLINE XXH_TARGET_AVX512 XXH3_ACCUMULATE_TEMPLATE(avx512) + +/* + * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing. + * + * Multiplication isn't perfect, as explained by Google in HighwayHash: + * + * // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to + * // varying degrees. In descending order of goodness, bytes + * // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32. + * // As expected, the upper and lower bytes are much worse. + * + * Source: https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291 + * + * Since our algorithm uses a pseudorandom secret to add some variance into the + * mix, we don't need to (or want to) mix as often or as much as HighwayHash does. + * + * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid + * extraction. + * + * Both XXH3_64bits and XXH3_128bits use this subroutine. + */ + +XXH_FORCE_INLINE XXH_TARGET_AVX512 void +XXH3_scrambleAcc_avx512(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 63) == 0); + XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i)); + { __m512i* const xacc = (__m512i*) acc; + const __m512i prime32 = _mm512_set1_epi32((int)XXH_PRIME32_1); + + /* xacc[0] ^= (xacc[0] >> 47) */ + __m512i const acc_vec = *xacc; + __m512i const shifted = _mm512_srli_epi64 (acc_vec, 47); + /* xacc[0] ^= secret; */ + __m512i const key_vec = _mm512_loadu_si512 (secret); + __m512i const data_key = _mm512_ternarylogic_epi32(key_vec, acc_vec, shifted, 0x96 /* key_vec ^ acc_vec ^ shifted */); + + /* xacc[0] *= XXH_PRIME32_1; */ + __m512i const data_key_hi = _mm512_srli_epi64 (data_key, 32); + __m512i const prod_lo = _mm512_mul_epu32 (data_key, prime32); + __m512i const prod_hi = _mm512_mul_epu32 (data_key_hi, prime32); + *xacc = _mm512_add_epi64(prod_lo, _mm512_slli_epi64(prod_hi, 32)); + } +} + +XXH_FORCE_INLINE XXH_TARGET_AVX512 void +XXH3_initCustomSecret_avx512(void* XXH_RESTRICT customSecret, xxh_u64 seed64) +{ + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 63) == 0); + XXH_STATIC_ASSERT(XXH_SEC_ALIGN == 64); + XXH_ASSERT(((size_t)customSecret & 63) == 0); + (void)(&XXH_writeLE64); + { int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m512i); + __m512i const seed_pos = _mm512_set1_epi64((xxh_i64)seed64); + __m512i const seed = _mm512_mask_sub_epi64(seed_pos, 0xAA, _mm512_set1_epi8(0), seed_pos); + + const __m512i* const src = (const __m512i*) ((const void*) XXH3_kSecret); + __m512i* const dest = ( __m512i*) customSecret; + int i; + XXH_ASSERT(((size_t)src & 63) == 0); /* control alignment */ + XXH_ASSERT(((size_t)dest & 63) == 0); + for (i=0; i < nbRounds; ++i) { + dest[i] = _mm512_add_epi64(_mm512_load_si512(src + i), seed); + } } +} + +#endif + +#if (XXH_VECTOR == XXH_AVX2) \ + || (defined(XXH_DISPATCH_AVX2) && XXH_DISPATCH_AVX2 != 0) + +#ifndef XXH_TARGET_AVX2 +# define XXH_TARGET_AVX2 /* disable attribute target */ +#endif + +XXH_FORCE_INLINE XXH_TARGET_AVX2 void +XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 31) == 0); + { __m256i* const xacc = (__m256i *) acc; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */ + const __m256i* const xinput = (const __m256i *) input; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */ + const __m256i* const xsecret = (const __m256i *) secret; + + size_t i; + for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) { + /* data_vec = xinput[i]; */ + __m256i const data_vec = _mm256_loadu_si256 (xinput+i); + /* key_vec = xsecret[i]; */ + __m256i const key_vec = _mm256_loadu_si256 (xsecret+i); + /* data_key = data_vec ^ key_vec; */ + __m256i const data_key = _mm256_xor_si256 (data_vec, key_vec); + /* data_key_lo = data_key >> 32; */ + __m256i const data_key_lo = _mm256_srli_epi64 (data_key, 32); + /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */ + __m256i const product = _mm256_mul_epu32 (data_key, data_key_lo); + /* xacc[i] += swap(data_vec); */ + __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2)); + __m256i const sum = _mm256_add_epi64(xacc[i], data_swap); + /* xacc[i] += product; */ + xacc[i] = _mm256_add_epi64(product, sum); + } } +} +XXH_FORCE_INLINE XXH_TARGET_AVX2 XXH3_ACCUMULATE_TEMPLATE(avx2) + +XXH_FORCE_INLINE XXH_TARGET_AVX2 void +XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 31) == 0); + { __m256i* const xacc = (__m256i*) acc; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */ + const __m256i* const xsecret = (const __m256i *) secret; + const __m256i prime32 = _mm256_set1_epi32((int)XXH_PRIME32_1); + + size_t i; + for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) { + /* xacc[i] ^= (xacc[i] >> 47) */ + __m256i const acc_vec = xacc[i]; + __m256i const shifted = _mm256_srli_epi64 (acc_vec, 47); + __m256i const data_vec = _mm256_xor_si256 (acc_vec, shifted); + /* xacc[i] ^= xsecret; */ + __m256i const key_vec = _mm256_loadu_si256 (xsecret+i); + __m256i const data_key = _mm256_xor_si256 (data_vec, key_vec); + + /* xacc[i] *= XXH_PRIME32_1; */ + __m256i const data_key_hi = _mm256_srli_epi64 (data_key, 32); + __m256i const prod_lo = _mm256_mul_epu32 (data_key, prime32); + __m256i const prod_hi = _mm256_mul_epu32 (data_key_hi, prime32); + xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32)); + } + } +} + +XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2(void* XXH_RESTRICT customSecret, xxh_u64 seed64) +{ + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 31) == 0); + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE / sizeof(__m256i)) == 6); + XXH_STATIC_ASSERT(XXH_SEC_ALIGN <= 64); + (void)(&XXH_writeLE64); + XXH_PREFETCH(customSecret); + { __m256i const seed = _mm256_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64, (xxh_i64)(0U - seed64), (xxh_i64)seed64); + + const __m256i* const src = (const __m256i*) ((const void*) XXH3_kSecret); + __m256i* dest = ( __m256i*) customSecret; + +# if defined(__GNUC__) || defined(__clang__) + /* + * On GCC & Clang, marking 'dest' as modified will cause the compiler: + * - do not extract the secret from sse registers in the internal loop + * - use less common registers, and avoid pushing these reg into stack + */ + XXH_COMPILER_GUARD(dest); +# endif + XXH_ASSERT(((size_t)src & 31) == 0); /* control alignment */ + XXH_ASSERT(((size_t)dest & 31) == 0); + + /* GCC -O2 need unroll loop manually */ + dest[0] = _mm256_add_epi64(_mm256_load_si256(src+0), seed); + dest[1] = _mm256_add_epi64(_mm256_load_si256(src+1), seed); + dest[2] = _mm256_add_epi64(_mm256_load_si256(src+2), seed); + dest[3] = _mm256_add_epi64(_mm256_load_si256(src+3), seed); + dest[4] = _mm256_add_epi64(_mm256_load_si256(src+4), seed); + dest[5] = _mm256_add_epi64(_mm256_load_si256(src+5), seed); + } +} + +#endif + +/* x86dispatch always generates SSE2 */ +#if (XXH_VECTOR == XXH_SSE2) || defined(XXH_X86DISPATCH) + +#ifndef XXH_TARGET_SSE2 +# define XXH_TARGET_SSE2 /* disable attribute target */ +#endif + +XXH_FORCE_INLINE XXH_TARGET_SSE2 void +XXH3_accumulate_512_sse2( void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + /* SSE2 is just a half-scale version of the AVX2 version. */ + XXH_ASSERT((((size_t)acc) & 15) == 0); + { __m128i* const xacc = (__m128i *) acc; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ + const __m128i* const xinput = (const __m128i *) input; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ + const __m128i* const xsecret = (const __m128i *) secret; + + size_t i; + for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) { + /* data_vec = xinput[i]; */ + __m128i const data_vec = _mm_loadu_si128 (xinput+i); + /* key_vec = xsecret[i]; */ + __m128i const key_vec = _mm_loadu_si128 (xsecret+i); + /* data_key = data_vec ^ key_vec; */ + __m128i const data_key = _mm_xor_si128 (data_vec, key_vec); + /* data_key_lo = data_key >> 32; */ + __m128i const data_key_lo = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1)); + /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */ + __m128i const product = _mm_mul_epu32 (data_key, data_key_lo); + /* xacc[i] += swap(data_vec); */ + __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2)); + __m128i const sum = _mm_add_epi64(xacc[i], data_swap); + /* xacc[i] += product; */ + xacc[i] = _mm_add_epi64(product, sum); + } } +} +XXH_FORCE_INLINE XXH_TARGET_SSE2 XXH3_ACCUMULATE_TEMPLATE(sse2) + +XXH_FORCE_INLINE XXH_TARGET_SSE2 void +XXH3_scrambleAcc_sse2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 15) == 0); + { __m128i* const xacc = (__m128i*) acc; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ + const __m128i* const xsecret = (const __m128i *) secret; + const __m128i prime32 = _mm_set1_epi32((int)XXH_PRIME32_1); + + size_t i; + for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) { + /* xacc[i] ^= (xacc[i] >> 47) */ + __m128i const acc_vec = xacc[i]; + __m128i const shifted = _mm_srli_epi64 (acc_vec, 47); + __m128i const data_vec = _mm_xor_si128 (acc_vec, shifted); + /* xacc[i] ^= xsecret[i]; */ + __m128i const key_vec = _mm_loadu_si128 (xsecret+i); + __m128i const data_key = _mm_xor_si128 (data_vec, key_vec); + + /* xacc[i] *= XXH_PRIME32_1; */ + __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1)); + __m128i const prod_lo = _mm_mul_epu32 (data_key, prime32); + __m128i const prod_hi = _mm_mul_epu32 (data_key_hi, prime32); + xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32)); + } + } +} + +XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2(void* XXH_RESTRICT customSecret, xxh_u64 seed64) +{ + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0); + (void)(&XXH_writeLE64); + { int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m128i); + +# if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900 + /* MSVC 32bit mode does not support _mm_set_epi64x before 2015 */ + XXH_ALIGN(16) const xxh_i64 seed64x2[2] = { (xxh_i64)seed64, (xxh_i64)(0U - seed64) }; + __m128i const seed = _mm_load_si128((__m128i const*)seed64x2); +# else + __m128i const seed = _mm_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64); +# endif + int i; + + const void* const src16 = XXH3_kSecret; + __m128i* dst16 = (__m128i*) customSecret; +# if defined(__GNUC__) || defined(__clang__) + /* + * On GCC & Clang, marking 'dest' as modified will cause the compiler: + * - do not extract the secret from sse registers in the internal loop + * - use less common registers, and avoid pushing these reg into stack + */ + XXH_COMPILER_GUARD(dst16); +# endif + XXH_ASSERT(((size_t)src16 & 15) == 0); /* control alignment */ + XXH_ASSERT(((size_t)dst16 & 15) == 0); + + for (i=0; i < nbRounds; ++i) { + dst16[i] = _mm_add_epi64(_mm_load_si128((const __m128i *)src16+i), seed); + } } +} + +#endif + +#if (XXH_VECTOR == XXH_NEON) + +/* forward declarations for the scalar routines */ +XXH_FORCE_INLINE void +XXH3_scalarRound(void* XXH_RESTRICT acc, void const* XXH_RESTRICT input, + void const* XXH_RESTRICT secret, size_t lane); + +XXH_FORCE_INLINE void +XXH3_scalarScrambleRound(void* XXH_RESTRICT acc, + void const* XXH_RESTRICT secret, size_t lane); + +/*! + * @internal + * @brief The bulk processing loop for NEON and WASM SIMD128. + * + * The NEON code path is actually partially scalar when running on AArch64. This + * is to optimize the pipelining and can have up to 15% speedup depending on the + * CPU, and it also mitigates some GCC codegen issues. + * + * @see XXH3_NEON_LANES for configuring this and details about this optimization. + * + * NEON's 32-bit to 64-bit long multiply takes a half vector of 32-bit + * integers instead of the other platforms which mask full 64-bit vectors, + * so the setup is more complicated than just shifting right. + * + * Additionally, there is an optimization for 4 lanes at once noted below. + * + * Since, as stated, the most optimal amount of lanes for Cortexes is 6, + * there needs to be *three* versions of the accumulate operation used + * for the remaining 2 lanes. + * + * WASM's SIMD128 uses SIMDe's arm_neon.h polyfill because the intrinsics overlap + * nearly perfectly. + */ + +XXH_FORCE_INLINE void +XXH3_accumulate_512_neon( void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 15) == 0); + XXH_STATIC_ASSERT(XXH3_NEON_LANES > 0 && XXH3_NEON_LANES <= XXH_ACC_NB && XXH3_NEON_LANES % 2 == 0); + { /* GCC for darwin arm64 does not like aliasing here */ + xxh_aliasing_uint64x2_t* const xacc = (xxh_aliasing_uint64x2_t*) acc; + /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */ + uint8_t const* xinput = (const uint8_t *) input; + uint8_t const* xsecret = (const uint8_t *) secret; + + size_t i; +#ifdef __wasm_simd128__ + /* + * On WASM SIMD128, Clang emits direct address loads when XXH3_kSecret + * is constant propagated, which results in it converting it to this + * inside the loop: + * + * a = v128.load(XXH3_kSecret + 0 + $secret_offset, offset = 0) + * b = v128.load(XXH3_kSecret + 16 + $secret_offset, offset = 0) + * ... + * + * This requires a full 32-bit address immediate (and therefore a 6 byte + * instruction) as well as an add for each offset. + * + * Putting an asm guard prevents it from folding (at the cost of losing + * the alignment hint), and uses the free offset in `v128.load` instead + * of adding secret_offset each time which overall reduces code size by + * about a kilobyte and improves performance. + */ + XXH_COMPILER_GUARD(xsecret); +#endif + /* Scalar lanes use the normal scalarRound routine */ + for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) { + XXH3_scalarRound(acc, input, secret, i); + } + i = 0; + /* 4 NEON lanes at a time. */ + for (; i+1 < XXH3_NEON_LANES / 2; i+=2) { + /* data_vec = xinput[i]; */ + uint64x2_t data_vec_1 = XXH_vld1q_u64(xinput + (i * 16)); + uint64x2_t data_vec_2 = XXH_vld1q_u64(xinput + ((i+1) * 16)); + /* key_vec = xsecret[i]; */ + uint64x2_t key_vec_1 = XXH_vld1q_u64(xsecret + (i * 16)); + uint64x2_t key_vec_2 = XXH_vld1q_u64(xsecret + ((i+1) * 16)); + /* data_swap = swap(data_vec) */ + uint64x2_t data_swap_1 = vextq_u64(data_vec_1, data_vec_1, 1); + uint64x2_t data_swap_2 = vextq_u64(data_vec_2, data_vec_2, 1); + /* data_key = data_vec ^ key_vec; */ + uint64x2_t data_key_1 = veorq_u64(data_vec_1, key_vec_1); + uint64x2_t data_key_2 = veorq_u64(data_vec_2, key_vec_2); + + /* + * If we reinterpret the 64x2 vectors as 32x4 vectors, we can use a + * de-interleave operation for 4 lanes in 1 step with `vuzpq_u32` to + * get one vector with the low 32 bits of each lane, and one vector + * with the high 32 bits of each lane. + * + * The intrinsic returns a double vector because the original ARMv7-a + * instruction modified both arguments in place. AArch64 and SIMD128 emit + * two instructions from this intrinsic. + * + * [ dk11L | dk11H | dk12L | dk12H ] -> [ dk11L | dk12L | dk21L | dk22L ] + * [ dk21L | dk21H | dk22L | dk22H ] -> [ dk11H | dk12H | dk21H | dk22H ] + */ + uint32x4x2_t unzipped = vuzpq_u32( + vreinterpretq_u32_u64(data_key_1), + vreinterpretq_u32_u64(data_key_2) + ); + /* data_key_lo = data_key & 0xFFFFFFFF */ + uint32x4_t data_key_lo = unzipped.val[0]; + /* data_key_hi = data_key >> 32 */ + uint32x4_t data_key_hi = unzipped.val[1]; + /* + * Then, we can split the vectors horizontally and multiply which, as for most + * widening intrinsics, have a variant that works on both high half vectors + * for free on AArch64. A similar instruction is available on SIMD128. + * + * sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi + */ + uint64x2_t sum_1 = XXH_vmlal_low_u32(data_swap_1, data_key_lo, data_key_hi); + uint64x2_t sum_2 = XXH_vmlal_high_u32(data_swap_2, data_key_lo, data_key_hi); + /* + * Clang reorders + * a += b * c; // umlal swap.2d, dkl.2s, dkh.2s + * c += a; // add acc.2d, acc.2d, swap.2d + * to + * c += a; // add acc.2d, acc.2d, swap.2d + * c += b * c; // umlal acc.2d, dkl.2s, dkh.2s + * + * While it would make sense in theory since the addition is faster, + * for reasons likely related to umlal being limited to certain NEON + * pipelines, this is worse. A compiler guard fixes this. + */ + XXH_COMPILER_GUARD_CLANG_NEON(sum_1); + XXH_COMPILER_GUARD_CLANG_NEON(sum_2); + /* xacc[i] = acc_vec + sum; */ + xacc[i] = vaddq_u64(xacc[i], sum_1); + xacc[i+1] = vaddq_u64(xacc[i+1], sum_2); + } + /* Operate on the remaining NEON lanes 2 at a time. */ + for (; i < XXH3_NEON_LANES / 2; i++) { + /* data_vec = xinput[i]; */ + uint64x2_t data_vec = XXH_vld1q_u64(xinput + (i * 16)); + /* key_vec = xsecret[i]; */ + uint64x2_t key_vec = XXH_vld1q_u64(xsecret + (i * 16)); + /* acc_vec_2 = swap(data_vec) */ + uint64x2_t data_swap = vextq_u64(data_vec, data_vec, 1); + /* data_key = data_vec ^ key_vec; */ + uint64x2_t data_key = veorq_u64(data_vec, key_vec); + /* For two lanes, just use VMOVN and VSHRN. */ + /* data_key_lo = data_key & 0xFFFFFFFF; */ + uint32x2_t data_key_lo = vmovn_u64(data_key); + /* data_key_hi = data_key >> 32; */ + uint32x2_t data_key_hi = vshrn_n_u64(data_key, 32); + /* sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi; */ + uint64x2_t sum = vmlal_u32(data_swap, data_key_lo, data_key_hi); + /* Same Clang workaround as before */ + XXH_COMPILER_GUARD_CLANG_NEON(sum); + /* xacc[i] = acc_vec + sum; */ + xacc[i] = vaddq_u64 (xacc[i], sum); + } + } +} +XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(neon) + +XXH_FORCE_INLINE void +XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 15) == 0); + + { xxh_aliasing_uint64x2_t* xacc = (xxh_aliasing_uint64x2_t*) acc; + uint8_t const* xsecret = (uint8_t const*) secret; + + size_t i; + /* WASM uses operator overloads and doesn't need these. */ +#ifndef __wasm_simd128__ + /* { prime32_1, prime32_1 } */ + uint32x2_t const kPrimeLo = vdup_n_u32(XXH_PRIME32_1); + /* { 0, prime32_1, 0, prime32_1 } */ + uint32x4_t const kPrimeHi = vreinterpretq_u32_u64(vdupq_n_u64((xxh_u64)XXH_PRIME32_1 << 32)); +#endif + + /* AArch64 uses both scalar and neon at the same time */ + for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) { + XXH3_scalarScrambleRound(acc, secret, i); + } + for (i=0; i < XXH3_NEON_LANES / 2; i++) { + /* xacc[i] ^= (xacc[i] >> 47); */ + uint64x2_t acc_vec = xacc[i]; + uint64x2_t shifted = vshrq_n_u64(acc_vec, 47); + uint64x2_t data_vec = veorq_u64(acc_vec, shifted); + + /* xacc[i] ^= xsecret[i]; */ + uint64x2_t key_vec = XXH_vld1q_u64(xsecret + (i * 16)); + uint64x2_t data_key = veorq_u64(data_vec, key_vec); + /* xacc[i] *= XXH_PRIME32_1 */ +#ifdef __wasm_simd128__ + /* SIMD128 has multiply by u64x2, use it instead of expanding and scalarizing */ + xacc[i] = data_key * XXH_PRIME32_1; +#else + /* + * Expanded version with portable NEON intrinsics + * + * lo(x) * lo(y) + (hi(x) * lo(y) << 32) + * + * prod_hi = hi(data_key) * lo(prime) << 32 + * + * Since we only need 32 bits of this multiply a trick can be used, reinterpreting the vector + * as a uint32x4_t and multiplying by { 0, prime, 0, prime } to cancel out the unwanted bits + * and avoid the shift. + */ + uint32x4_t prod_hi = vmulq_u32 (vreinterpretq_u32_u64(data_key), kPrimeHi); + /* Extract low bits for vmlal_u32 */ + uint32x2_t data_key_lo = vmovn_u64(data_key); + /* xacc[i] = prod_hi + lo(data_key) * XXH_PRIME32_1; */ + xacc[i] = vmlal_u32(vreinterpretq_u64_u32(prod_hi), data_key_lo, kPrimeLo); +#endif + } + } +} +#endif + +#if (XXH_VECTOR == XXH_VSX) + +XXH_FORCE_INLINE void +XXH3_accumulate_512_vsx( void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + /* presumed aligned */ + xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc; + xxh_u8 const* const xinput = (xxh_u8 const*) input; /* no alignment restriction */ + xxh_u8 const* const xsecret = (xxh_u8 const*) secret; /* no alignment restriction */ + xxh_u64x2 const v32 = { 32, 32 }; + size_t i; + for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) { + /* data_vec = xinput[i]; */ + xxh_u64x2 const data_vec = XXH_vec_loadu(xinput + 16*i); + /* key_vec = xsecret[i]; */ + xxh_u64x2 const key_vec = XXH_vec_loadu(xsecret + 16*i); + xxh_u64x2 const data_key = data_vec ^ key_vec; + /* shuffled = (data_key << 32) | (data_key >> 32); */ + xxh_u32x4 const shuffled = (xxh_u32x4)vec_rl(data_key, v32); + /* product = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)shuffled & 0xFFFFFFFF); */ + xxh_u64x2 const product = XXH_vec_mulo((xxh_u32x4)data_key, shuffled); + /* acc_vec = xacc[i]; */ + xxh_u64x2 acc_vec = xacc[i]; + acc_vec += product; + + /* swap high and low halves */ +#ifdef __s390x__ + acc_vec += vec_permi(data_vec, data_vec, 2); +#else + acc_vec += vec_xxpermdi(data_vec, data_vec, 2); +#endif + xacc[i] = acc_vec; + } +} +XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(vsx) + +XXH_FORCE_INLINE void +XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 15) == 0); + + { xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc; + const xxh_u8* const xsecret = (const xxh_u8*) secret; + /* constants */ + xxh_u64x2 const v32 = { 32, 32 }; + xxh_u64x2 const v47 = { 47, 47 }; + xxh_u32x4 const prime = { XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1 }; + size_t i; + for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) { + /* xacc[i] ^= (xacc[i] >> 47); */ + xxh_u64x2 const acc_vec = xacc[i]; + xxh_u64x2 const data_vec = acc_vec ^ (acc_vec >> v47); + + /* xacc[i] ^= xsecret[i]; */ + xxh_u64x2 const key_vec = XXH_vec_loadu(xsecret + 16*i); + xxh_u64x2 const data_key = data_vec ^ key_vec; + + /* xacc[i] *= XXH_PRIME32_1 */ + /* prod_lo = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)prime & 0xFFFFFFFF); */ + xxh_u64x2 const prod_even = XXH_vec_mule((xxh_u32x4)data_key, prime); + /* prod_hi = ((xxh_u64x2)data_key >> 32) * ((xxh_u64x2)prime >> 32); */ + xxh_u64x2 const prod_odd = XXH_vec_mulo((xxh_u32x4)data_key, prime); + xacc[i] = prod_odd + (prod_even << v32); + } } +} + +#endif + +#if (XXH_VECTOR == XXH_SVE) + +XXH_FORCE_INLINE void +XXH3_accumulate_512_sve( void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + uint64_t *xacc = (uint64_t *)acc; + const uint64_t *xinput = (const uint64_t *)(const void *)input; + const uint64_t *xsecret = (const uint64_t *)(const void *)secret; + svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1); + uint64_t element_count = svcntd(); + if (element_count >= 8) { + svbool_t mask = svptrue_pat_b64(SV_VL8); + svuint64_t vacc = svld1_u64(mask, xacc); + ACCRND(vacc, 0); + svst1_u64(mask, xacc, vacc); + } else if (element_count == 2) { /* sve128 */ + svbool_t mask = svptrue_pat_b64(SV_VL2); + svuint64_t acc0 = svld1_u64(mask, xacc + 0); + svuint64_t acc1 = svld1_u64(mask, xacc + 2); + svuint64_t acc2 = svld1_u64(mask, xacc + 4); + svuint64_t acc3 = svld1_u64(mask, xacc + 6); + ACCRND(acc0, 0); + ACCRND(acc1, 2); + ACCRND(acc2, 4); + ACCRND(acc3, 6); + svst1_u64(mask, xacc + 0, acc0); + svst1_u64(mask, xacc + 2, acc1); + svst1_u64(mask, xacc + 4, acc2); + svst1_u64(mask, xacc + 6, acc3); + } else { + svbool_t mask = svptrue_pat_b64(SV_VL4); + svuint64_t acc0 = svld1_u64(mask, xacc + 0); + svuint64_t acc1 = svld1_u64(mask, xacc + 4); + ACCRND(acc0, 0); + ACCRND(acc1, 4); + svst1_u64(mask, xacc + 0, acc0); + svst1_u64(mask, xacc + 4, acc1); + } +} + +XXH_FORCE_INLINE void +XXH3_accumulate_sve(xxh_u64* XXH_RESTRICT acc, + const xxh_u8* XXH_RESTRICT input, + const xxh_u8* XXH_RESTRICT secret, + size_t nbStripes) +{ + if (nbStripes != 0) { + uint64_t *xacc = (uint64_t *)acc; + const uint64_t *xinput = (const uint64_t *)(const void *)input; + const uint64_t *xsecret = (const uint64_t *)(const void *)secret; + svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1); + uint64_t element_count = svcntd(); + if (element_count >= 8) { + svbool_t mask = svptrue_pat_b64(SV_VL8); + svuint64_t vacc = svld1_u64(mask, xacc + 0); + do { + /* svprfd(svbool_t, void *, enum svfprop); */ + svprfd(mask, xinput + 128, SV_PLDL1STRM); + ACCRND(vacc, 0); + xinput += 8; + xsecret += 1; + nbStripes--; + } while (nbStripes != 0); + + svst1_u64(mask, xacc + 0, vacc); + } else if (element_count == 2) { /* sve128 */ + svbool_t mask = svptrue_pat_b64(SV_VL2); + svuint64_t acc0 = svld1_u64(mask, xacc + 0); + svuint64_t acc1 = svld1_u64(mask, xacc + 2); + svuint64_t acc2 = svld1_u64(mask, xacc + 4); + svuint64_t acc3 = svld1_u64(mask, xacc + 6); + do { + svprfd(mask, xinput + 128, SV_PLDL1STRM); + ACCRND(acc0, 0); + ACCRND(acc1, 2); + ACCRND(acc2, 4); + ACCRND(acc3, 6); + xinput += 8; + xsecret += 1; + nbStripes--; + } while (nbStripes != 0); + + svst1_u64(mask, xacc + 0, acc0); + svst1_u64(mask, xacc + 2, acc1); + svst1_u64(mask, xacc + 4, acc2); + svst1_u64(mask, xacc + 6, acc3); + } else { + svbool_t mask = svptrue_pat_b64(SV_VL4); + svuint64_t acc0 = svld1_u64(mask, xacc + 0); + svuint64_t acc1 = svld1_u64(mask, xacc + 4); + do { + svprfd(mask, xinput + 128, SV_PLDL1STRM); + ACCRND(acc0, 0); + ACCRND(acc1, 4); + xinput += 8; + xsecret += 1; + nbStripes--; + } while (nbStripes != 0); + + svst1_u64(mask, xacc + 0, acc0); + svst1_u64(mask, xacc + 4, acc1); + } + } +} + +#endif + +/* scalar variants - universal */ + +#if defined(__aarch64__) && (defined(__GNUC__) || defined(__clang__)) +/* + * In XXH3_scalarRound(), GCC and Clang have a similar codegen issue, where they + * emit an excess mask and a full 64-bit multiply-add (MADD X-form). + * + * While this might not seem like much, as AArch64 is a 64-bit architecture, only + * big Cortex designs have a full 64-bit multiplier. + * + * On the little cores, the smaller 32-bit multiplier is used, and full 64-bit + * multiplies expand to 2-3 multiplies in microcode. This has a major penalty + * of up to 4 latency cycles and 2 stall cycles in the multiply pipeline. + * + * Thankfully, AArch64 still provides the 32-bit long multiply-add (UMADDL) which does + * not have this penalty and does the mask automatically. + */ +XXH_FORCE_INLINE xxh_u64 +XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc) +{ + xxh_u64 ret; + /* note: %x = 64-bit register, %w = 32-bit register */ + __asm__("umaddl %x0, %w1, %w2, %x3" : "=r" (ret) : "r" (lhs), "r" (rhs), "r" (acc)); + return ret; +} +#else +XXH_FORCE_INLINE xxh_u64 +XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc) +{ + return XXH_mult32to64((xxh_u32)lhs, (xxh_u32)rhs) + acc; +} +#endif + +/*! + * @internal + * @brief Scalar round for @ref XXH3_accumulate_512_scalar(). + * + * This is extracted to its own function because the NEON path uses a combination + * of NEON and scalar. + */ +XXH_FORCE_INLINE void +XXH3_scalarRound(void* XXH_RESTRICT acc, + void const* XXH_RESTRICT input, + void const* XXH_RESTRICT secret, + size_t lane) +{ + xxh_u64* xacc = (xxh_u64*) acc; + xxh_u8 const* xinput = (xxh_u8 const*) input; + xxh_u8 const* xsecret = (xxh_u8 const*) secret; + XXH_ASSERT(lane < XXH_ACC_NB); + XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0); + { + xxh_u64 const data_val = XXH_readLE64(xinput + lane * 8); + xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + lane * 8); + xacc[lane ^ 1] += data_val; /* swap adjacent lanes */ + xacc[lane] = XXH_mult32to64_add64(data_key /* & 0xFFFFFFFF */, data_key >> 32, xacc[lane]); + } +} + +/*! + * @internal + * @brief Processes a 64 byte block of data using the scalar path. + */ +XXH_FORCE_INLINE void +XXH3_accumulate_512_scalar(void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + size_t i; + /* ARM GCC refuses to unroll this loop, resulting in a 24% slowdown on ARMv6. */ +#if defined(__GNUC__) && !defined(__clang__) \ + && (defined(__arm__) || defined(__thumb2__)) \ + && defined(__ARM_FEATURE_UNALIGNED) /* no unaligned access just wastes bytes */ \ + && XXH_SIZE_OPT <= 0 +# pragma GCC unroll 8 +#endif + for (i=0; i < XXH_ACC_NB; i++) { + XXH3_scalarRound(acc, input, secret, i); + } +} +XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(scalar) + +/*! + * @internal + * @brief Scalar scramble step for @ref XXH3_scrambleAcc_scalar(). + * + * This is extracted to its own function because the NEON path uses a combination + * of NEON and scalar. + */ +XXH_FORCE_INLINE void +XXH3_scalarScrambleRound(void* XXH_RESTRICT acc, + void const* XXH_RESTRICT secret, + size_t lane) +{ + xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned */ + const xxh_u8* const xsecret = (const xxh_u8*) secret; /* no alignment restriction */ + XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0); + XXH_ASSERT(lane < XXH_ACC_NB); + { + xxh_u64 const key64 = XXH_readLE64(xsecret + lane * 8); + xxh_u64 acc64 = xacc[lane]; + acc64 = XXH_xorshift64(acc64, 47); + acc64 ^= key64; + acc64 *= XXH_PRIME32_1; + xacc[lane] = acc64; + } +} + +/*! + * @internal + * @brief Scrambles the accumulators after a large chunk has been read + */ +XXH_FORCE_INLINE void +XXH3_scrambleAcc_scalar(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + size_t i; + for (i=0; i < XXH_ACC_NB; i++) { + XXH3_scalarScrambleRound(acc, secret, i); + } +} + +XXH_FORCE_INLINE void +XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64) +{ + /* + * We need a separate pointer for the hack below, + * which requires a non-const pointer. + * Any decent compiler will optimize this out otherwise. + */ + const xxh_u8* kSecretPtr = XXH3_kSecret; + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0); + +#if defined(__GNUC__) && defined(__aarch64__) + /* + * UGLY HACK: + * GCC and Clang generate a bunch of MOV/MOVK pairs for aarch64, and they are + * placed sequentially, in order, at the top of the unrolled loop. + * + * While MOVK is great for generating constants (2 cycles for a 64-bit + * constant compared to 4 cycles for LDR), it fights for bandwidth with + * the arithmetic instructions. + * + * I L S + * MOVK + * MOVK + * MOVK + * MOVK + * ADD + * SUB STR + * STR + * By forcing loads from memory (as the asm line causes the compiler to assume + * that XXH3_kSecretPtr has been changed), the pipelines are used more + * efficiently: + * I L S + * LDR + * ADD LDR + * SUB STR + * STR + * + * See XXH3_NEON_LANES for details on the pipsline. + * + * XXH3_64bits_withSeed, len == 256, Snapdragon 835 + * without hack: 2654.4 MB/s + * with hack: 3202.9 MB/s + */ + XXH_COMPILER_GUARD(kSecretPtr); +#endif + { int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16; + int i; + for (i=0; i < nbRounds; i++) { + /* + * The asm hack causes the compiler to assume that kSecretPtr aliases with + * customSecret, and on aarch64, this prevented LDP from merging two + * loads together for free. Putting the loads together before the stores + * properly generates LDP. + */ + xxh_u64 lo = XXH_readLE64(kSecretPtr + 16*i) + seed64; + xxh_u64 hi = XXH_readLE64(kSecretPtr + 16*i + 8) - seed64; + XXH_writeLE64((xxh_u8*)customSecret + 16*i, lo); + XXH_writeLE64((xxh_u8*)customSecret + 16*i + 8, hi); + } } +} + + +typedef void (*XXH3_f_accumulate)(xxh_u64* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, size_t); +typedef void (*XXH3_f_scrambleAcc)(void* XXH_RESTRICT, const void*); +typedef void (*XXH3_f_initCustomSecret)(void* XXH_RESTRICT, xxh_u64); + + +#if (XXH_VECTOR == XXH_AVX512) + +#define XXH3_accumulate_512 XXH3_accumulate_512_avx512 +#define XXH3_accumulate XXH3_accumulate_avx512 +#define XXH3_scrambleAcc XXH3_scrambleAcc_avx512 +#define XXH3_initCustomSecret XXH3_initCustomSecret_avx512 + +#elif (XXH_VECTOR == XXH_AVX2) + +#define XXH3_accumulate_512 XXH3_accumulate_512_avx2 +#define XXH3_accumulate XXH3_accumulate_avx2 +#define XXH3_scrambleAcc XXH3_scrambleAcc_avx2 +#define XXH3_initCustomSecret XXH3_initCustomSecret_avx2 + +#elif (XXH_VECTOR == XXH_SSE2) + +#define XXH3_accumulate_512 XXH3_accumulate_512_sse2 +#define XXH3_accumulate XXH3_accumulate_sse2 +#define XXH3_scrambleAcc XXH3_scrambleAcc_sse2 +#define XXH3_initCustomSecret XXH3_initCustomSecret_sse2 + +#elif (XXH_VECTOR == XXH_NEON) + +#define XXH3_accumulate_512 XXH3_accumulate_512_neon +#define XXH3_accumulate XXH3_accumulate_neon +#define XXH3_scrambleAcc XXH3_scrambleAcc_neon +#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar + +#elif (XXH_VECTOR == XXH_VSX) + +#define XXH3_accumulate_512 XXH3_accumulate_512_vsx +#define XXH3_accumulate XXH3_accumulate_vsx +#define XXH3_scrambleAcc XXH3_scrambleAcc_vsx +#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar + +#elif (XXH_VECTOR == XXH_SVE) +#define XXH3_accumulate_512 XXH3_accumulate_512_sve +#define XXH3_accumulate XXH3_accumulate_sve +#define XXH3_scrambleAcc XXH3_scrambleAcc_scalar +#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar + +#else /* scalar */ + +#define XXH3_accumulate_512 XXH3_accumulate_512_scalar +#define XXH3_accumulate XXH3_accumulate_scalar +#define XXH3_scrambleAcc XXH3_scrambleAcc_scalar +#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar + +#endif + +#if XXH_SIZE_OPT >= 1 /* don't do SIMD for initialization */ +# undef XXH3_initCustomSecret +# define XXH3_initCustomSecret XXH3_initCustomSecret_scalar +#endif + +XXH_FORCE_INLINE void +XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc, + const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH3_f_accumulate f_acc, + XXH3_f_scrambleAcc f_scramble) +{ + size_t const nbStripesPerBlock = (secretSize - XXH_STRIPE_LEN) / XXH_SECRET_CONSUME_RATE; + size_t const block_len = XXH_STRIPE_LEN * nbStripesPerBlock; + size_t const nb_blocks = (len - 1) / block_len; + + size_t n; + + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + + for (n = 0; n < nb_blocks; n++) { + f_acc(acc, input + n*block_len, secret, nbStripesPerBlock); + f_scramble(acc, secret + secretSize - XXH_STRIPE_LEN); + } + + /* last partial block */ + XXH_ASSERT(len > XXH_STRIPE_LEN); + { size_t const nbStripes = ((len - 1) - (block_len * nb_blocks)) / XXH_STRIPE_LEN; + XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE)); + f_acc(acc, input + nb_blocks*block_len, secret, nbStripes); + + /* last stripe */ + { const xxh_u8* const p = input + len - XXH_STRIPE_LEN; +#define XXH_SECRET_LASTACC_START 7 /* not aligned on 8, last secret is different from acc & scrambler */ + XXH3_accumulate_512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START); + } } +} + +XXH_FORCE_INLINE xxh_u64 +XXH3_mix2Accs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret) +{ + return XXH3_mul128_fold64( + acc[0] ^ XXH_readLE64(secret), + acc[1] ^ XXH_readLE64(secret+8) ); +} + +static XXH64_hash_t +XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, xxh_u64 start) +{ + xxh_u64 result64 = start; + size_t i = 0; + + for (i = 0; i < 4; i++) { + result64 += XXH3_mix2Accs(acc+2*i, secret + 16*i); +#if defined(__clang__) /* Clang */ \ + && (defined(__arm__) || defined(__thumb__)) /* ARMv7 */ \ + && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \ + && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable */ + /* + * UGLY HACK: + * Prevent autovectorization on Clang ARMv7-a. Exact same problem as + * the one in XXH3_len_129to240_64b. Speeds up shorter keys > 240b. + * XXH3_64bits, len == 256, Snapdragon 835: + * without hack: 2063.7 MB/s + * with hack: 2560.7 MB/s + */ + XXH_COMPILER_GUARD(result64); +#endif + } + + return XXH3_avalanche(result64); +} + +#define XXH3_INIT_ACC { XXH_PRIME32_3, XXH_PRIME64_1, XXH_PRIME64_2, XXH_PRIME64_3, \ + XXH_PRIME64_4, XXH_PRIME32_2, XXH_PRIME64_5, XXH_PRIME32_1 } + +XXH_FORCE_INLINE XXH64_hash_t +XXH3_hashLong_64b_internal(const void* XXH_RESTRICT input, size_t len, + const void* XXH_RESTRICT secret, size_t secretSize, + XXH3_f_accumulate f_acc, + XXH3_f_scrambleAcc f_scramble) +{ + XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC; + + XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, f_acc, f_scramble); + + /* converge into final hash */ + XXH_STATIC_ASSERT(sizeof(acc) == 64); + /* do not align on 8, so that the secret is different from the accumulator */ +#define XXH_SECRET_MERGEACCS_START 11 + XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); + return XXH3_mergeAccs(acc, (const xxh_u8*)secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * XXH_PRIME64_1); +} + +/* + * It's important for performance to transmit secret's size (when it's static) + * so that the compiler can properly optimize the vectorized loop. + * This makes a big performance difference for "medium" keys (<1 KB) when using AVX instruction set. + * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE + * breaks -Og, this is XXH_NO_INLINE. + */ +XXH3_WITH_SECRET_INLINE XXH64_hash_t +XXH3_hashLong_64b_withSecret(const void* XXH_RESTRICT input, size_t len, + XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen) +{ + (void)seed64; + return XXH3_hashLong_64b_internal(input, len, secret, secretLen, XXH3_accumulate, XXH3_scrambleAcc); +} + +/* + * It's preferable for performance that XXH3_hashLong is not inlined, + * as it results in a smaller function for small data, easier to the instruction cache. + * Note that inside this no_inline function, we do inline the internal loop, + * and provide a statically defined secret size to allow optimization of vector loop. + */ +XXH_NO_INLINE XXH_PUREF XXH64_hash_t +XXH3_hashLong_64b_default(const void* XXH_RESTRICT input, size_t len, + XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen) +{ + (void)seed64; (void)secret; (void)secretLen; + return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_accumulate, XXH3_scrambleAcc); +} + +/* + * XXH3_hashLong_64b_withSeed(): + * Generate a custom key based on alteration of default XXH3_kSecret with the seed, + * and then use this key for long mode hashing. + * + * This operation is decently fast but nonetheless costs a little bit of time. + * Try to avoid it whenever possible (typically when seed==0). + * + * It's important for performance that XXH3_hashLong is not inlined. Not sure + * why (uop cache maybe?), but the difference is large and easily measurable. + */ +XXH_FORCE_INLINE XXH64_hash_t +XXH3_hashLong_64b_withSeed_internal(const void* input, size_t len, + XXH64_hash_t seed, + XXH3_f_accumulate f_acc, + XXH3_f_scrambleAcc f_scramble, + XXH3_f_initCustomSecret f_initSec) +{ +#if XXH_SIZE_OPT <= 0 + if (seed == 0) + return XXH3_hashLong_64b_internal(input, len, + XXH3_kSecret, sizeof(XXH3_kSecret), + f_acc, f_scramble); +#endif + { XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE]; + f_initSec(secret, seed); + return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret), + f_acc, f_scramble); + } +} + +/* + * It's important for performance that XXH3_hashLong is not inlined. + */ +XXH_NO_INLINE XXH64_hash_t +XXH3_hashLong_64b_withSeed(const void* XXH_RESTRICT input, size_t len, + XXH64_hash_t seed, const xxh_u8* XXH_RESTRICT secret, size_t secretLen) +{ + (void)secret; (void)secretLen; + return XXH3_hashLong_64b_withSeed_internal(input, len, seed, + XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret); +} + + +typedef XXH64_hash_t (*XXH3_hashLong64_f)(const void* XXH_RESTRICT, size_t, + XXH64_hash_t, const xxh_u8* XXH_RESTRICT, size_t); + +XXH_FORCE_INLINE XXH64_hash_t +XXH3_64bits_internal(const void* XXH_RESTRICT input, size_t len, + XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen, + XXH3_hashLong64_f f_hashLong) +{ + XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN); + /* + * If an action is to be taken if `secretLen` condition is not respected, + * it should be done here. + * For now, it's a contract pre-condition. + * Adding a check and a branch here would cost performance at every hash. + * Also, note that function signature doesn't offer room to return an error. + */ + if (len <= 16) + return XXH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64); + if (len <= 128) + return XXH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64); + if (len <= XXH3_MIDSIZE_MAX) + return XXH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64); + return f_hashLong(input, len, seed64, (const xxh_u8*)secret, secretLen); +} + + +/* === Public entry point === */ + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length) +{ + return XXH3_64bits_internal(input, length, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH64_hash_t +XXH3_64bits_withSecret(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize) +{ + return XXH3_64bits_internal(input, length, 0, secret, secretSize, XXH3_hashLong_64b_withSecret); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH64_hash_t +XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed) +{ + return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed); +} + +XXH_PUBLIC_API XXH64_hash_t +XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed) +{ + if (length <= XXH3_MIDSIZE_MAX) + return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL); + return XXH3_hashLong_64b_withSecret(input, length, seed, (const xxh_u8*)secret, secretSize); +} + + +/* === XXH3 streaming === */ +#ifndef XXH_NO_STREAM +/* + * Malloc's a pointer that is always aligned to align. + * + * This must be freed with `XXH_alignedFree()`. + * + * malloc typically guarantees 16 byte alignment on 64-bit systems and 8 byte + * alignment on 32-bit. This isn't enough for the 32 byte aligned loads in AVX2 + * or on 32-bit, the 16 byte aligned loads in SSE2 and NEON. + * + * This underalignment previously caused a rather obvious crash which went + * completely unnoticed due to XXH3_createState() not actually being tested. + * Credit to RedSpah for noticing this bug. + * + * The alignment is done manually: Functions like posix_memalign or _mm_malloc + * are avoided: To maintain portability, we would have to write a fallback + * like this anyways, and besides, testing for the existence of library + * functions without relying on external build tools is impossible. + * + * The method is simple: Overallocate, manually align, and store the offset + * to the original behind the returned pointer. + * + * Align must be a power of 2 and 8 <= align <= 128. + */ +static XXH_MALLOCF void* XXH_alignedMalloc(size_t s, size_t align) +{ + XXH_ASSERT(align <= 128 && align >= 8); /* range check */ + XXH_ASSERT((align & (align-1)) == 0); /* power of 2 */ + XXH_ASSERT(s != 0 && s < (s + align)); /* empty/overflow */ + { /* Overallocate to make room for manual realignment and an offset byte */ + xxh_u8* base = (xxh_u8*)XXH_malloc(s + align); + if (base != NULL) { + /* + * Get the offset needed to align this pointer. + * + * Even if the returned pointer is aligned, there will always be + * at least one byte to store the offset to the original pointer. + */ + size_t offset = align - ((size_t)base & (align - 1)); /* base % align */ + /* Add the offset for the now-aligned pointer */ + xxh_u8* ptr = base + offset; + + XXH_ASSERT((size_t)ptr % align == 0); + + /* Store the offset immediately before the returned pointer. */ + ptr[-1] = (xxh_u8)offset; + return ptr; + } + return NULL; + } +} +/* + * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass + * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout. + */ +static void XXH_alignedFree(void* p) +{ + if (p != NULL) { + xxh_u8* ptr = (xxh_u8*)p; + /* Get the offset byte we added in XXH_malloc. */ + xxh_u8 offset = ptr[-1]; + /* Free the original malloc'd pointer */ + xxh_u8* base = ptr - offset; + XXH_free(base); + } +} +/*! @ingroup XXH3_family */ +/*! + * @brief Allocate an @ref XXH3_state_t. + * + * @return An allocated pointer of @ref XXH3_state_t on success. + * @return `NULL` on failure. + * + * @note Must be freed with XXH3_freeState(). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void) +{ + XXH3_state_t* const state = (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64); + if (state==NULL) return NULL; + XXH3_INITSTATE(state); + return state; +} + +/*! @ingroup XXH3_family */ +/*! + * @brief Frees an @ref XXH3_state_t. + * + * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState(). + * + * @return @ref XXH_OK. + * + * @note Must be allocated with XXH3_createState(). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr) +{ + XXH_alignedFree(statePtr); + return XXH_OK; +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API void +XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state) +{ + XXH_memcpy(dst_state, src_state, sizeof(*dst_state)); +} + +static void +XXH3_reset_internal(XXH3_state_t* statePtr, + XXH64_hash_t seed, + const void* secret, size_t secretSize) +{ + size_t const initStart = offsetof(XXH3_state_t, bufferedSize); + size_t const initLength = offsetof(XXH3_state_t, nbStripesPerBlock) - initStart; + XXH_ASSERT(offsetof(XXH3_state_t, nbStripesPerBlock) > initStart); + XXH_ASSERT(statePtr != NULL); + /* set members from bufferedSize to nbStripesPerBlock (excluded) to 0 */ + memset((char*)statePtr + initStart, 0, initLength); + statePtr->acc[0] = XXH_PRIME32_3; + statePtr->acc[1] = XXH_PRIME64_1; + statePtr->acc[2] = XXH_PRIME64_2; + statePtr->acc[3] = XXH_PRIME64_3; + statePtr->acc[4] = XXH_PRIME64_4; + statePtr->acc[5] = XXH_PRIME32_2; + statePtr->acc[6] = XXH_PRIME64_5; + statePtr->acc[7] = XXH_PRIME32_1; + statePtr->seed = seed; + statePtr->useSeed = (seed != 0); + statePtr->extSecret = (const unsigned char*)secret; + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + statePtr->secretLimit = secretSize - XXH_STRIPE_LEN; + statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE; +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr) +{ + if (statePtr == NULL) return XXH_ERROR; + XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE); + return XXH_OK; +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize) +{ + if (statePtr == NULL) return XXH_ERROR; + XXH3_reset_internal(statePtr, 0, secret, secretSize); + if (secret == NULL) return XXH_ERROR; + if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; + return XXH_OK; +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed) +{ + if (statePtr == NULL) return XXH_ERROR; + if (seed==0) return XXH3_64bits_reset(statePtr); + if ((seed != statePtr->seed) || (statePtr->extSecret != NULL)) + XXH3_initCustomSecret(statePtr->customSecret, seed); + XXH3_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE); + return XXH_OK; +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed64) +{ + if (statePtr == NULL) return XXH_ERROR; + if (secret == NULL) return XXH_ERROR; + if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; + XXH3_reset_internal(statePtr, seed64, secret, secretSize); + statePtr->useSeed = 1; /* always, even if seed64==0 */ + return XXH_OK; +} + +/*! + * @internal + * @brief Processes a large input for XXH3_update() and XXH3_digest_long(). + * + * Unlike XXH3_hashLong_internal_loop(), this can process data that overlaps a block. + * + * @param acc Pointer to the 8 accumulator lanes + * @param nbStripesSoFarPtr In/out pointer to the number of leftover stripes in the block* + * @param nbStripesPerBlock Number of stripes in a block + * @param input Input pointer + * @param nbStripes Number of stripes to process + * @param secret Secret pointer + * @param secretLimit Offset of the last block in @p secret + * @param f_acc Pointer to an XXH3_accumulate implementation + * @param f_scramble Pointer to an XXH3_scrambleAcc implementation + * @return Pointer past the end of @p input after processing + */ +XXH_FORCE_INLINE const xxh_u8 * +XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc, + size_t* XXH_RESTRICT nbStripesSoFarPtr, size_t nbStripesPerBlock, + const xxh_u8* XXH_RESTRICT input, size_t nbStripes, + const xxh_u8* XXH_RESTRICT secret, size_t secretLimit, + XXH3_f_accumulate f_acc, + XXH3_f_scrambleAcc f_scramble) +{ + const xxh_u8* initialSecret = secret + *nbStripesSoFarPtr * XXH_SECRET_CONSUME_RATE; + /* Process full blocks */ + if (nbStripes >= (nbStripesPerBlock - *nbStripesSoFarPtr)) { + /* Process the initial partial block... */ + size_t nbStripesThisIter = nbStripesPerBlock - *nbStripesSoFarPtr; + + do { + /* Accumulate and scramble */ + f_acc(acc, input, initialSecret, nbStripesThisIter); + f_scramble(acc, secret + secretLimit); + input += nbStripesThisIter * XXH_STRIPE_LEN; + nbStripes -= nbStripesThisIter; + /* Then continue the loop with the full block size */ + nbStripesThisIter = nbStripesPerBlock; + initialSecret = secret; + } while (nbStripes >= nbStripesPerBlock); + *nbStripesSoFarPtr = 0; + } + /* Process a partial block */ + if (nbStripes > 0) { + f_acc(acc, input, initialSecret, nbStripes); + input += nbStripes * XXH_STRIPE_LEN; + *nbStripesSoFarPtr += nbStripes; + } + /* Return end pointer */ + return input; +} + +#ifndef XXH3_STREAM_USE_STACK +# if XXH_SIZE_OPT <= 0 && !defined(__clang__) /* clang doesn't need additional stack space */ +# define XXH3_STREAM_USE_STACK 1 +# endif +#endif +/* + * Both XXH3_64bits_update and XXH3_128bits_update use this routine. + */ +XXH_FORCE_INLINE XXH_errorcode +XXH3_update(XXH3_state_t* XXH_RESTRICT const state, + const xxh_u8* XXH_RESTRICT input, size_t len, + XXH3_f_accumulate f_acc, + XXH3_f_scrambleAcc f_scramble) +{ + if (input==NULL) { + XXH_ASSERT(len == 0); + return XXH_OK; + } + + XXH_ASSERT(state != NULL); + { const xxh_u8* const bEnd = input + len; + const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret; +#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1 + /* For some reason, gcc and MSVC seem to suffer greatly + * when operating accumulators directly into state. + * Operating into stack space seems to enable proper optimization. + * clang, on the other hand, doesn't seem to need this trick */ + XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[8]; + XXH_memcpy(acc, state->acc, sizeof(acc)); +#else + xxh_u64* XXH_RESTRICT const acc = state->acc; +#endif + state->totalLen += len; + XXH_ASSERT(state->bufferedSize <= XXH3_INTERNALBUFFER_SIZE); + + /* small input : just fill in tmp buffer */ + if (len <= XXH3_INTERNALBUFFER_SIZE - state->bufferedSize) { + XXH_memcpy(state->buffer + state->bufferedSize, input, len); + state->bufferedSize += (XXH32_hash_t)len; + return XXH_OK; + } + + /* total input is now > XXH3_INTERNALBUFFER_SIZE */ + #define XXH3_INTERNALBUFFER_STRIPES (XXH3_INTERNALBUFFER_SIZE / XXH_STRIPE_LEN) + XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % XXH_STRIPE_LEN == 0); /* clean multiple */ + + /* + * Internal buffer is partially filled (always, except at beginning) + * Complete it, then consume it. + */ + if (state->bufferedSize) { + size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize; + XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize); + input += loadSize; + XXH3_consumeStripes(acc, + &state->nbStripesSoFar, state->nbStripesPerBlock, + state->buffer, XXH3_INTERNALBUFFER_STRIPES, + secret, state->secretLimit, + f_acc, f_scramble); + state->bufferedSize = 0; + } + XXH_ASSERT(input < bEnd); + if (bEnd - input > XXH3_INTERNALBUFFER_SIZE) { + size_t nbStripes = (size_t)(bEnd - 1 - input) / XXH_STRIPE_LEN; + input = XXH3_consumeStripes(acc, + &state->nbStripesSoFar, state->nbStripesPerBlock, + input, nbStripes, + secret, state->secretLimit, + f_acc, f_scramble); + XXH_memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN); + + } + /* Some remaining input (always) : buffer it */ + XXH_ASSERT(input < bEnd); + XXH_ASSERT(bEnd - input <= XXH3_INTERNALBUFFER_SIZE); + XXH_ASSERT(state->bufferedSize == 0); + XXH_memcpy(state->buffer, input, (size_t)(bEnd-input)); + state->bufferedSize = (XXH32_hash_t)(bEnd-input); +#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1 + /* save stack accumulators into state */ + XXH_memcpy(state->acc, acc, sizeof(acc)); +#endif + } + + return XXH_OK; +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len) +{ + return XXH3_update(state, (const xxh_u8*)input, len, + XXH3_accumulate, XXH3_scrambleAcc); +} + + +XXH_FORCE_INLINE void +XXH3_digest_long (XXH64_hash_t* acc, + const XXH3_state_t* state, + const unsigned char* secret) +{ + xxh_u8 lastStripe[XXH_STRIPE_LEN]; + const xxh_u8* lastStripePtr; + + /* + * Digest on a local copy. This way, the state remains unaltered, and it can + * continue ingesting more input afterwards. + */ + XXH_memcpy(acc, state->acc, sizeof(state->acc)); + if (state->bufferedSize >= XXH_STRIPE_LEN) { + /* Consume remaining stripes then point to remaining data in buffer */ + size_t const nbStripes = (state->bufferedSize - 1) / XXH_STRIPE_LEN; + size_t nbStripesSoFar = state->nbStripesSoFar; + XXH3_consumeStripes(acc, + &nbStripesSoFar, state->nbStripesPerBlock, + state->buffer, nbStripes, + secret, state->secretLimit, + XXH3_accumulate, XXH3_scrambleAcc); + lastStripePtr = state->buffer + state->bufferedSize - XXH_STRIPE_LEN; + } else { /* bufferedSize < XXH_STRIPE_LEN */ + /* Copy to temp buffer */ + size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize; + XXH_ASSERT(state->bufferedSize > 0); /* there is always some input buffered */ + XXH_memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize); + XXH_memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize); + lastStripePtr = lastStripe; + } + /* Last stripe */ + XXH3_accumulate_512(acc, + lastStripePtr, + secret + state->secretLimit - XXH_SECRET_LASTACC_START); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* state) +{ + const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret; + if (state->totalLen > XXH3_MIDSIZE_MAX) { + XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB]; + XXH3_digest_long(acc, state, secret); + return XXH3_mergeAccs(acc, + secret + XXH_SECRET_MERGEACCS_START, + (xxh_u64)state->totalLen * XXH_PRIME64_1); + } + /* totalLen <= XXH3_MIDSIZE_MAX: digesting a short input */ + if (state->useSeed) + return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed); + return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen), + secret, state->secretLimit + XXH_STRIPE_LEN); +} +#endif /* !XXH_NO_STREAM */ + + +/* ========================================== + * XXH3 128 bits (a.k.a XXH128) + * ========================================== + * XXH3's 128-bit variant has better mixing and strength than the 64-bit variant, + * even without counting the significantly larger output size. + * + * For example, extra steps are taken to avoid the seed-dependent collisions + * in 17-240 byte inputs (See XXH3_mix16B and XXH128_mix32B). + * + * This strength naturally comes at the cost of some speed, especially on short + * lengths. Note that longer hashes are about as fast as the 64-bit version + * due to it using only a slight modification of the 64-bit loop. + * + * XXH128 is also more oriented towards 64-bit machines. It is still extremely + * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64). + */ + +XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t +XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + /* A doubled version of 1to3_64b with different constants. */ + XXH_ASSERT(input != NULL); + XXH_ASSERT(1 <= len && len <= 3); + XXH_ASSERT(secret != NULL); + /* + * len = 1: combinedl = { input[0], 0x01, input[0], input[0] } + * len = 2: combinedl = { input[1], 0x02, input[0], input[1] } + * len = 3: combinedl = { input[2], 0x03, input[0], input[1] } + */ + { xxh_u8 const c1 = input[0]; + xxh_u8 const c2 = input[len >> 1]; + xxh_u8 const c3 = input[len - 1]; + xxh_u32 const combinedl = ((xxh_u32)c1 <<16) | ((xxh_u32)c2 << 24) + | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8); + xxh_u32 const combinedh = XXH_rotl32(XXH_swap32(combinedl), 13); + xxh_u64 const bitflipl = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed; + xxh_u64 const bitfliph = (XXH_readLE32(secret+8) ^ XXH_readLE32(secret+12)) - seed; + xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ bitflipl; + xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ bitfliph; + XXH128_hash_t h128; + h128.low64 = XXH64_avalanche(keyed_lo); + h128.high64 = XXH64_avalanche(keyed_hi); + return h128; + } +} + +XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t +XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(4 <= len && len <= 8); + seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32; + { xxh_u32 const input_lo = XXH_readLE32(input); + xxh_u32 const input_hi = XXH_readLE32(input + len - 4); + xxh_u64 const input_64 = input_lo + ((xxh_u64)input_hi << 32); + xxh_u64 const bitflip = (XXH_readLE64(secret+16) ^ XXH_readLE64(secret+24)) + seed; + xxh_u64 const keyed = input_64 ^ bitflip; + + /* Shift len to the left to ensure it is even, this avoids even multiplies. */ + XXH128_hash_t m128 = XXH_mult64to128(keyed, XXH_PRIME64_1 + (len << 2)); + + m128.high64 += (m128.low64 << 1); + m128.low64 ^= (m128.high64 >> 3); + + m128.low64 = XXH_xorshift64(m128.low64, 35); + m128.low64 *= PRIME_MX2; + m128.low64 = XXH_xorshift64(m128.low64, 28); + m128.high64 = XXH3_avalanche(m128.high64); + return m128; + } +} + +XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t +XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(9 <= len && len <= 16); + { xxh_u64 const bitflipl = (XXH_readLE64(secret+32) ^ XXH_readLE64(secret+40)) - seed; + xxh_u64 const bitfliph = (XXH_readLE64(secret+48) ^ XXH_readLE64(secret+56)) + seed; + xxh_u64 const input_lo = XXH_readLE64(input); + xxh_u64 input_hi = XXH_readLE64(input + len - 8); + XXH128_hash_t m128 = XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, XXH_PRIME64_1); + /* + * Put len in the middle of m128 to ensure that the length gets mixed to + * both the low and high bits in the 128x64 multiply below. + */ + m128.low64 += (xxh_u64)(len - 1) << 54; + input_hi ^= bitfliph; + /* + * Add the high 32 bits of input_hi to the high 32 bits of m128, then + * add the long product of the low 32 bits of input_hi and XXH_PRIME32_2 to + * the high 64 bits of m128. + * + * The best approach to this operation is different on 32-bit and 64-bit. + */ + if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */ + /* + * 32-bit optimized version, which is more readable. + * + * On 32-bit, it removes an ADC and delays a dependency between the two + * halves of m128.high64, but it generates an extra mask on 64-bit. + */ + m128.high64 += (input_hi & 0xFFFFFFFF00000000ULL) + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2); + } else { + /* + * 64-bit optimized (albeit more confusing) version. + * + * Uses some properties of addition and multiplication to remove the mask: + * + * Let: + * a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF) + * b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000) + * c = XXH_PRIME32_2 + * + * a + (b * c) + * Inverse Property: x + y - x == y + * a + (b * (1 + c - 1)) + * Distributive Property: x * (y + z) == (x * y) + (x * z) + * a + (b * 1) + (b * (c - 1)) + * Identity Property: x * 1 == x + * a + b + (b * (c - 1)) + * + * Substitute a, b, and c: + * input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1)) + * + * Since input_hi.hi + input_hi.lo == input_hi, we get this: + * input_hi + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1)) + */ + m128.high64 += input_hi + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2 - 1); + } + /* m128 ^= XXH_swap64(m128 >> 64); */ + m128.low64 ^= XXH_swap64(m128.high64); + + { /* 128x64 multiply: h128 = m128 * XXH_PRIME64_2; */ + XXH128_hash_t h128 = XXH_mult64to128(m128.low64, XXH_PRIME64_2); + h128.high64 += m128.high64 * XXH_PRIME64_2; + + h128.low64 = XXH3_avalanche(h128.low64); + h128.high64 = XXH3_avalanche(h128.high64); + return h128; + } } +} + +/* + * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN + */ +XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t +XXH3_len_0to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(len <= 16); + { if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed); + if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed); + if (len) return XXH3_len_1to3_128b(input, len, secret, seed); + { XXH128_hash_t h128; + xxh_u64 const bitflipl = XXH_readLE64(secret+64) ^ XXH_readLE64(secret+72); + xxh_u64 const bitfliph = XXH_readLE64(secret+80) ^ XXH_readLE64(secret+88); + h128.low64 = XXH64_avalanche(seed ^ bitflipl); + h128.high64 = XXH64_avalanche( seed ^ bitfliph); + return h128; + } } +} + +/* + * A bit slower than XXH3_mix16B, but handles multiply by zero better. + */ +XXH_FORCE_INLINE XXH128_hash_t +XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2, + const xxh_u8* secret, XXH64_hash_t seed) +{ + acc.low64 += XXH3_mix16B (input_1, secret+0, seed); + acc.low64 ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8); + acc.high64 += XXH3_mix16B (input_2, secret+16, seed); + acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8); + return acc; +} + + +XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t +XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH64_hash_t seed) +{ + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; + XXH_ASSERT(16 < len && len <= 128); + + { XXH128_hash_t acc; + acc.low64 = len * XXH_PRIME64_1; + acc.high64 = 0; + +#if XXH_SIZE_OPT >= 1 + { + /* Smaller, but slightly slower. */ + unsigned int i = (unsigned int)(len - 1) / 32; + do { + acc = XXH128_mix32B(acc, input+16*i, input+len-16*(i+1), secret+32*i, seed); + } while (i-- != 0); + } +#else + if (len > 32) { + if (len > 64) { + if (len > 96) { + acc = XXH128_mix32B(acc, input+48, input+len-64, secret+96, seed); + } + acc = XXH128_mix32B(acc, input+32, input+len-48, secret+64, seed); + } + acc = XXH128_mix32B(acc, input+16, input+len-32, secret+32, seed); + } + acc = XXH128_mix32B(acc, input, input+len-16, secret, seed); +#endif + { XXH128_hash_t h128; + h128.low64 = acc.low64 + acc.high64; + h128.high64 = (acc.low64 * XXH_PRIME64_1) + + (acc.high64 * XXH_PRIME64_4) + + ((len - seed) * XXH_PRIME64_2); + h128.low64 = XXH3_avalanche(h128.low64); + h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64); + return h128; + } + } +} + +XXH_NO_INLINE XXH_PUREF XXH128_hash_t +XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH64_hash_t seed) +{ + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; + XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); + + { XXH128_hash_t acc; + unsigned i; + acc.low64 = len * XXH_PRIME64_1; + acc.high64 = 0; + /* + * We set as `i` as offset + 32. We do this so that unchanged + * `len` can be used as upper bound. This reaches a sweet spot + * where both x86 and aarch64 get simple agen and good codegen + * for the loop. + */ + for (i = 32; i < 160; i += 32) { + acc = XXH128_mix32B(acc, + input + i - 32, + input + i - 16, + secret + i - 32, + seed); + } + acc.low64 = XXH3_avalanche(acc.low64); + acc.high64 = XXH3_avalanche(acc.high64); + /* + * NB: `i <= len` will duplicate the last 32-bytes if + * len % 32 was zero. This is an unfortunate necessity to keep + * the hash result stable. + */ + for (i=160; i <= len; i += 32) { + acc = XXH128_mix32B(acc, + input + i - 32, + input + i - 16, + secret + XXH3_MIDSIZE_STARTOFFSET + i - 160, + seed); + } + /* last bytes */ + acc = XXH128_mix32B(acc, + input + len - 16, + input + len - 32, + secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16, + (XXH64_hash_t)0 - seed); + + { XXH128_hash_t h128; + h128.low64 = acc.low64 + acc.high64; + h128.high64 = (acc.low64 * XXH_PRIME64_1) + + (acc.high64 * XXH_PRIME64_4) + + ((len - seed) * XXH_PRIME64_2); + h128.low64 = XXH3_avalanche(h128.low64); + h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64); + return h128; + } + } +} + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_hashLong_128b_internal(const void* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH3_f_accumulate f_acc, + XXH3_f_scrambleAcc f_scramble) +{ + XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC; + + XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, secret, secretSize, f_acc, f_scramble); + + /* converge into final hash */ + XXH_STATIC_ASSERT(sizeof(acc) == 64); + XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); + { XXH128_hash_t h128; + h128.low64 = XXH3_mergeAccs(acc, + secret + XXH_SECRET_MERGEACCS_START, + (xxh_u64)len * XXH_PRIME64_1); + h128.high64 = XXH3_mergeAccs(acc, + secret + secretSize + - sizeof(acc) - XXH_SECRET_MERGEACCS_START, + ~((xxh_u64)len * XXH_PRIME64_2)); + return h128; + } +} + +/* + * It's important for performance that XXH3_hashLong() is not inlined. + */ +XXH_NO_INLINE XXH_PUREF XXH128_hash_t +XXH3_hashLong_128b_default(const void* XXH_RESTRICT input, size_t len, + XXH64_hash_t seed64, + const void* XXH_RESTRICT secret, size_t secretLen) +{ + (void)seed64; (void)secret; (void)secretLen; + return XXH3_hashLong_128b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), + XXH3_accumulate, XXH3_scrambleAcc); +} + +/* + * It's important for performance to pass @p secretLen (when it's static) + * to the compiler, so that it can properly optimize the vectorized loop. + * + * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE + * breaks -Og, this is XXH_NO_INLINE. + */ +XXH3_WITH_SECRET_INLINE XXH128_hash_t +XXH3_hashLong_128b_withSecret(const void* XXH_RESTRICT input, size_t len, + XXH64_hash_t seed64, + const void* XXH_RESTRICT secret, size_t secretLen) +{ + (void)seed64; + return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, secretLen, + XXH3_accumulate, XXH3_scrambleAcc); +} + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_hashLong_128b_withSeed_internal(const void* XXH_RESTRICT input, size_t len, + XXH64_hash_t seed64, + XXH3_f_accumulate f_acc, + XXH3_f_scrambleAcc f_scramble, + XXH3_f_initCustomSecret f_initSec) +{ + if (seed64 == 0) + return XXH3_hashLong_128b_internal(input, len, + XXH3_kSecret, sizeof(XXH3_kSecret), + f_acc, f_scramble); + { XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE]; + f_initSec(secret, seed64); + return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, sizeof(secret), + f_acc, f_scramble); + } +} + +/* + * It's important for performance that XXH3_hashLong is not inlined. + */ +XXH_NO_INLINE XXH128_hash_t +XXH3_hashLong_128b_withSeed(const void* input, size_t len, + XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen) +{ + (void)secret; (void)secretLen; + return XXH3_hashLong_128b_withSeed_internal(input, len, seed64, + XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret); +} + +typedef XXH128_hash_t (*XXH3_hashLong128_f)(const void* XXH_RESTRICT, size_t, + XXH64_hash_t, const void* XXH_RESTRICT, size_t); + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_128bits_internal(const void* input, size_t len, + XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen, + XXH3_hashLong128_f f_hl128) +{ + XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN); + /* + * If an action is to be taken if `secret` conditions are not respected, + * it should be done here. + * For now, it's a contract pre-condition. + * Adding a check and a branch here would cost performance at every hash. + */ + if (len <= 16) + return XXH3_len_0to16_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64); + if (len <= 128) + return XXH3_len_17to128_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64); + if (len <= XXH3_MIDSIZE_MAX) + return XXH3_len_129to240_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64); + return f_hl128(input, len, seed64, secret, secretLen); +} + + +/* === Public XXH128 API === */ + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* input, size_t len) +{ + return XXH3_128bits_internal(input, len, 0, + XXH3_kSecret, sizeof(XXH3_kSecret), + XXH3_hashLong_128b_default); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH128_hash_t +XXH3_128bits_withSecret(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize) +{ + return XXH3_128bits_internal(input, len, 0, + (const xxh_u8*)secret, secretSize, + XXH3_hashLong_128b_withSecret); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH128_hash_t +XXH3_128bits_withSeed(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed) +{ + return XXH3_128bits_internal(input, len, seed, + XXH3_kSecret, sizeof(XXH3_kSecret), + XXH3_hashLong_128b_withSeed); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH128_hash_t +XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed) +{ + if (len <= XXH3_MIDSIZE_MAX) + return XXH3_128bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL); + return XXH3_hashLong_128b_withSecret(input, len, seed, secret, secretSize); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH128_hash_t +XXH128(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed) +{ + return XXH3_128bits_withSeed(input, len, seed); +} + + +/* === XXH3 128-bit streaming === */ +#ifndef XXH_NO_STREAM +/* + * All initialization and update functions are identical to 64-bit streaming variant. + * The only difference is the finalization routine. + */ + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr) +{ + return XXH3_64bits_reset(statePtr); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize) +{ + return XXH3_64bits_reset_withSecret(statePtr, secret, secretSize); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed) +{ + return XXH3_64bits_reset_withSeed(statePtr, seed); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed) +{ + return XXH3_64bits_reset_withSecretandSeed(statePtr, secret, secretSize, seed); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len) +{ + return XXH3_64bits_update(state, input, len); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* state) +{ + const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret; + if (state->totalLen > XXH3_MIDSIZE_MAX) { + XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB]; + XXH3_digest_long(acc, state, secret); + XXH_ASSERT(state->secretLimit + XXH_STRIPE_LEN >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); + { XXH128_hash_t h128; + h128.low64 = XXH3_mergeAccs(acc, + secret + XXH_SECRET_MERGEACCS_START, + (xxh_u64)state->totalLen * XXH_PRIME64_1); + h128.high64 = XXH3_mergeAccs(acc, + secret + state->secretLimit + XXH_STRIPE_LEN + - sizeof(acc) - XXH_SECRET_MERGEACCS_START, + ~((xxh_u64)state->totalLen * XXH_PRIME64_2)); + return h128; + } + } + /* len <= XXH3_MIDSIZE_MAX : short code */ + if (state->useSeed) + return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed); + return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen), + secret, state->secretLimit + XXH_STRIPE_LEN); +} +#endif /* !XXH_NO_STREAM */ +/* 128-bit utility functions */ + +#include /* memcmp, memcpy */ + +/* return : 1 is equal, 0 if different */ +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2) +{ + /* note : XXH128_hash_t is compact, it has no padding byte */ + return !(memcmp(&h1, &h2, sizeof(h1))); +} + +/* This prototype is compatible with stdlib's qsort(). + * @return : >0 if *h128_1 > *h128_2 + * <0 if *h128_1 < *h128_2 + * =0 if *h128_1 == *h128_2 */ +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2) +{ + XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1; + XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2; + int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64); + /* note : bets that, in most cases, hash values are different */ + if (hcmp) return hcmp; + return (h1.low64 > h2.low64) - (h2.low64 > h1.low64); +} + + +/*====== Canonical representation ======*/ +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API void +XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash) +{ + XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t)); + if (XXH_CPU_LITTLE_ENDIAN) { + hash.high64 = XXH_swap64(hash.high64); + hash.low64 = XXH_swap64(hash.low64); + } + XXH_memcpy(dst, &hash.high64, sizeof(hash.high64)); + XXH_memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64)); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH128_hash_t +XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src) +{ + XXH128_hash_t h; + h.high64 = XXH_readBE64(src); + h.low64 = XXH_readBE64(src->digest + 8); + return h; +} + + + +/* ========================================== + * Secret generators + * ========================================== + */ +#define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x)) + +XXH_FORCE_INLINE void XXH3_combine16(void* dst, XXH128_hash_t h128) +{ + XXH_writeLE64( dst, XXH_readLE64(dst) ^ h128.low64 ); + XXH_writeLE64( (char*)dst+8, XXH_readLE64((char*)dst+8) ^ h128.high64 ); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize) +{ +#if (XXH_DEBUGLEVEL >= 1) + XXH_ASSERT(secretBuffer != NULL); + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); +#else + /* production mode, assert() are disabled */ + if (secretBuffer == NULL) return XXH_ERROR; + if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; +#endif + + if (customSeedSize == 0) { + customSeed = XXH3_kSecret; + customSeedSize = XXH_SECRET_DEFAULT_SIZE; + } +#if (XXH_DEBUGLEVEL >= 1) + XXH_ASSERT(customSeed != NULL); +#else + if (customSeed == NULL) return XXH_ERROR; +#endif + + /* Fill secretBuffer with a copy of customSeed - repeat as needed */ + { size_t pos = 0; + while (pos < secretSize) { + size_t const toCopy = XXH_MIN((secretSize - pos), customSeedSize); + memcpy((char*)secretBuffer + pos, customSeed, toCopy); + pos += toCopy; + } } + + { size_t const nbSeg16 = secretSize / 16; + size_t n; + XXH128_canonical_t scrambler; + XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0)); + for (n=0; n /* abort() */ +#include +#include +#include +#include +#include +#include + +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +#include "xxhash/xxhash.h" +#include "sha1/sha1.h" +#include "sha256/sha256.h" + +#ifdef __cplusplus +} +#endif + + +// uuid.uuid5(uuid.NAMESPACE_URL, 'en.wikipedia.org/wiki/Llama.cpp') +#define UUID_NAMESPACE_LLAMA_CPP "ef001206-dadc-5f6d-a15f-3359e577d4e5" +#define UUID_NAMESPACE_LLAMA_CPP_HEX 0xef, 0x00, 0x12, 0x06, 0xda, 0xdc, 0x5f, 0x6d, 0xa1, 0x5f, 0x33, 0x59, 0xe5, 0x77, 0xd4, 0xe5 + + +#define HASH_TYPE_SHA256_STR "sha256" +#define HASH_TYPE_SHA1_STR "sha1" +#define HASH_TYPE_XXH64_STR "xxh64" +#define HASH_TYPE_UUID_STR "uuid" + + +typedef enum { + HASH_EXIT_SUCCESS = 0, // All hash has been generated or validated + HASH_EXIT_FAILURE = 1, // Generic Failure + HASH_EXIT_MISMATCH = 2, // Hash mismatched during validation + HASH_EXIT_MANIFEST_MISSING_ENTRY = 3, // Hash attempted validation but missing entry in manifest + HASH_EXIT_MANIFEST_UNKNOWN_HASH = 4, // Manifest is present, but we do not know any hash format within it + HASH_EXIT_MANIFEST_FILE_ERROR = 5 // Manifest is either missing or not a known format +} hash_exit_code_t; + + +typedef enum { + HASH_MANIFEST_NOT_FOUND, + HASH_MANIFEST_MISMATCH, + HASH_MANIFEST_OK, +} hash_manifest_result_t; + + +struct hash_params { + std::string input; + bool xxh64 = false; + bool sha1 = false; + bool sha256 = false; + bool uuid = false; + + bool no_layer = false; + + bool manifest_is_usable = false; + std::string manifest_file; +}; + +struct manifest_check_params { + bool xxh64 = false; + bool sha1 = false; + bool sha256 = false; + bool uuid = false; +}; + +static char const * hash_manifest_result_to_str(hash_manifest_result_t value) { + switch (value) { + case HASH_MANIFEST_NOT_FOUND: return "Not Found"; + case HASH_MANIFEST_MISMATCH: return "Mismatch"; + case HASH_MANIFEST_OK: return "Ok"; + } + return "?"; +} + +static char const * hash_exit_code_to_str(hash_exit_code_t value) { + switch (value) { + case HASH_EXIT_SUCCESS: return "Success"; + case HASH_EXIT_FAILURE: return "Failure"; + case HASH_EXIT_MISMATCH: return "Mismatch"; + case HASH_EXIT_MANIFEST_MISSING_ENTRY: return "Manifest Missing Entry"; + case HASH_EXIT_MANIFEST_UNKNOWN_HASH: return "Manifest Unknown Hash"; + case HASH_EXIT_MANIFEST_FILE_ERROR: return "Manifest File Error"; + } + return "?"; +} + +static void hash_print_usage(const char * executable) { + const hash_params default_params; + printf("\n"); + printf("usage: %s [options] GGUF_IN\n", executable); + printf("\n"); + printf("Hash a GGUF file"); + printf("\n"); + printf("options:\n"); + printf(" -h, --help show this help message and exit\n"); + printf(" --xxh64 use xxh64 hash\n"); + printf(" --sha1 use sha1 hash\n"); + printf(" --sha256 use sha256 hash\n"); + printf(" --all use all hash\n"); + printf(" --no-layer exclude per layer hash\n"); + printf(" --uuid generate UUIDv5 ID\n"); + printf(" -c, --check verify against a manifest\n"); + printf("\n"); +} + +static void hash_params_parse_ex(int argc, const char ** argv, hash_params & params) { + std::string arg; + bool invalid_param = false; + const std::string arg_prefix = "--"; + + int arg_idx = 1; + for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) { + arg = argv[arg_idx]; + if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) { + std::replace(arg.begin(), arg.end(), '_', '-'); + } + + bool arg_found = false; + if (arg == "-h" || arg == "--help") { + hash_print_usage(argv[0]); + exit(0); + } + + if (arg == "--xxh64") { + arg_found = true; + params.xxh64 = true; + } + + if (arg == "--sha1") { + arg_found = true; + params.sha1 = true; + } + + if (arg == "--uuid") { + arg_found = true; + params.uuid = true; + } + + if (arg == "--sha256") { + arg_found = true; + params.sha256 = true; + } + + if (arg == "--all") { + arg_found = true; + params.sha256 = true; + params.sha1 = true; + params.xxh64 = true; + } + + if (arg == "--no-layer") { + arg_found = true; + params.no_layer = true; + } + + if (arg == "-c" || arg == "--check") { + if (++arg_idx >= argc) { + invalid_param = true; + break; + } + arg_found = true; + params.manifest_file = argv[arg_idx]; + } + + if (!arg_found) { + throw std::invalid_argument("error: unknown argument: " + arg); + } + } + + if (invalid_param) { + throw std::invalid_argument("error: invalid parameter for argument:" + arg); + } + + if (argc - arg_idx < 1) { + throw std::invalid_argument("error: bad arguments"); + } + + params.input = argv[arg_idx++]; +} + +static bool hash_params_parse(int argc, const char ** argv, hash_params & params) { + bool result = true; + try { + hash_params_parse_ex(argc, argv, params); + } + catch (const std::invalid_argument & ex) { + fprintf(stderr, "%s\n", ex.what()); + hash_print_usage(argv[0]); + exit(EXIT_FAILURE); + } + return result; +} + +static bool manifest_type(const std::string & manifest_file, manifest_check_params & manifest_check) { + if (manifest_file.empty()) { + return false; + } + + std::ifstream file(manifest_file); + if (!file.is_open()) { + return false; + } + + std::string manifest_entry_line; + while (getline(file, manifest_entry_line)) { + // hash_type_str hash_str tensor_name + // e.g. 'xxh64 f66e9cd66a4396a0 test.gguf:tensor_0' + std::istringstream line_stream(manifest_entry_line); + std::string file_hash_type; + if (line_stream >> file_hash_type) { + if (file_hash_type == HASH_TYPE_SHA256_STR) { + manifest_check.sha256 = true; + } else if (file_hash_type == HASH_TYPE_SHA1_STR) { + manifest_check.sha1 = true; + } else if (file_hash_type == HASH_TYPE_XXH64_STR) { + manifest_check.xxh64 = true; + } else if (file_hash_type == HASH_TYPE_UUID_STR) { + manifest_check.uuid = true; + } + } + } + + return true; +} + +static hash_manifest_result_t manifest_verify(const std::string& manifest_file, const std::string& hash_type_str, const std::string& hash_str, const std::string& tensor_name) { + if (manifest_file.empty()) { + return HASH_MANIFEST_NOT_FOUND; + } + + std::ifstream file(manifest_file); + if (!file.is_open()) { + return HASH_MANIFEST_NOT_FOUND; + } + + std::string manifest_entry_line; + while (getline(file, manifest_entry_line)) { + std::istringstream line_stream(manifest_entry_line); + std::string file_hash_type; + std::string file_hash; + std::string file_tensor_name; + if (line_stream >> file_hash_type >> file_hash >> file_tensor_name) { + // Line parsed. Check hash validity + + if (file_hash_type != hash_type_str) { + continue; + } + + if (file_tensor_name != tensor_name) { + continue; + } + + return (file_hash == hash_str) ? HASH_MANIFEST_OK : HASH_MANIFEST_MISMATCH; + } + } + + return HASH_MANIFEST_NOT_FOUND; +} + +static void generate_uuidv5(const unsigned char sha1_digest[20], unsigned char uuid[16]) { + // Ref: https://www.rfc-editor.org/rfc/rfc9562.html#section-5.5 + // Assumes that digest was processed correctly with the expected namespace + for (int i = 0; i < 16; i++) { + uuid[i] = sha1_digest[i]; + } + + // Set bits corresponding to UUID ver 5 + uuid[ 6] &= ~(0xF << 4); + uuid[ 6] |= (5 << 4); + + // Set bits corresponding to UUID variant 0b10XX + uuid[ 8] &= ~(0xc << 4); + uuid[ 8] |= (0x8 << 4); +} + +static hash_exit_code_t gguf_hash(const hash_params & hash_params) { + const std::string & fname = hash_params.input; + struct ggml_context * ctx_data = NULL; + + struct gguf_init_params params = { + /*.no_alloc = */ false, + /*.ctx = */ &ctx_data, + }; + + // xxh64 init + XXH64_state_t* xxh64_model_hash_state = NULL; + if (hash_params.xxh64) { + xxh64_model_hash_state = XXH64_createState(); + if (xxh64_model_hash_state==NULL) { + abort(); + } + + XXH64_hash_t const seed = 0; + if (XXH64_reset(xxh64_model_hash_state, seed) == XXH_ERROR) { + abort(); + } + } + + // sha1 init + SHA1_CTX sha1_model_hash_ctx; + if (hash_params.sha1) { + SHA1Init(&sha1_model_hash_ctx); + } + + // sha256 init + sha256_t sha256_model_hash_ctx; + if (hash_params.sha256) { + sha256_init(&sha256_model_hash_ctx); + } + + // sha1 for uuid init + SHA1_CTX sha1_for_uuid_ctx; + if (hash_params.uuid) { + unsigned char const uuidv5_namespace[] = {UUID_NAMESPACE_LLAMA_CPP_HEX}; + SHA1Init(&sha1_for_uuid_ctx); + SHA1Update( &sha1_for_uuid_ctx, (unsigned char const *)uuidv5_namespace, sizeof(uuidv5_namespace)); + } + + struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params); + const int n_tensors = gguf_get_n_tensors(ctx); + bool tensor_layer_in_manifest = false; + bool model_in_manifest = false; + bool tensor_layer_has_mismatch = false; + bool model_has_mismatch = false; + for (int i = 0; i < n_tensors; ++i) { + const char * name = gguf_get_tensor_name(ctx, i); + struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name); + auto n_bytes = ggml_nbytes(cur); + auto *raw_data = cur->data; + const std::string tensor_layer_name = fname + ":" + name; + + if (hash_params.xxh64) { + + if (!hash_params.no_layer) { + // Per Layer Hash + XXH64_hash_t hash = XXH64(raw_data, n_bytes, 0); + + char hex_result[17]; + for (int offset = 0; offset < 8; offset++) { + unsigned int shift_bits_by = (8 * (8 - offset - 1)); + snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", (unsigned char) (hash >> shift_bits_by)&0xff); + } + + if (hash_params.manifest_is_usable) { + hash_manifest_result_t verify_result = manifest_verify(hash_params.manifest_file, HASH_TYPE_XXH64_STR, hex_result, tensor_layer_name); + + switch (verify_result) { + case HASH_MANIFEST_NOT_FOUND: + break; + case HASH_MANIFEST_MISMATCH: + tensor_layer_in_manifest = true; + tensor_layer_has_mismatch = true; + break; + case HASH_MANIFEST_OK: + tensor_layer_in_manifest = true; + break; + } + + printf("%-8s %-s %s - %s\n", HASH_TYPE_XXH64_STR, hex_result, tensor_layer_name.c_str(), hash_manifest_result_to_str(verify_result)); + } else { + printf("%-8s %-s %s\n", HASH_TYPE_XXH64_STR, hex_result, tensor_layer_name.c_str()); + } + } + + // Overall Model Hash + if (XXH64_update(xxh64_model_hash_state, raw_data, n_bytes) == XXH_ERROR) abort(); + } + + if (hash_params.sha1) { + + if (!hash_params.no_layer) { + // Per Layer Hash + char result[21]; // sha1 outputs 20 bytes + SHA1( result, (const char *)raw_data, n_bytes); + + char hex_result[41] = {0}; + for (int offset = 0; offset < 20; offset++) { + snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff); + } + + if (hash_params.manifest_is_usable) { + hash_manifest_result_t verify_result = manifest_verify(hash_params.manifest_file, HASH_TYPE_SHA1_STR, hex_result, tensor_layer_name); + + switch (verify_result) { + case HASH_MANIFEST_NOT_FOUND: + break; + case HASH_MANIFEST_MISMATCH: + tensor_layer_in_manifest = true; + tensor_layer_has_mismatch = true; + break; + case HASH_MANIFEST_OK: + tensor_layer_in_manifest = true; + break; + } + + printf("%-8s %-s %s - %s\n", HASH_TYPE_SHA1_STR, hex_result, tensor_layer_name.c_str(), hash_manifest_result_to_str(verify_result)); + } else { + printf("%-8s %-s %s\n", HASH_TYPE_SHA1_STR, hex_result, tensor_layer_name.c_str()); + } + } + + // Overall Model Hash + SHA1Update( &sha1_model_hash_ctx, (unsigned char const *)raw_data, n_bytes); + } + + if (hash_params.sha256) { + + if (!hash_params.no_layer) { + // Per Layer Hash + unsigned char result[SHA256_DIGEST_SIZE]; // sha256 outputs 32 bytes + sha256_hash((unsigned char*) result, (const unsigned char *)raw_data, n_bytes); + + char hex_result[SHA256_DIGEST_SIZE * 2 + 1] = {0}; + for (int offset = 0; offset < SHA256_DIGEST_SIZE; offset++) { + snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff); + } + + if (hash_params.manifest_is_usable) { + hash_manifest_result_t verify_result = manifest_verify(hash_params.manifest_file, HASH_TYPE_SHA256_STR, hex_result, tensor_layer_name); + + switch (verify_result) { + case HASH_MANIFEST_NOT_FOUND: + break; + case HASH_MANIFEST_MISMATCH: + tensor_layer_in_manifest = true; + tensor_layer_has_mismatch = true; + break; + case HASH_MANIFEST_OK: + tensor_layer_in_manifest = true; + break; + } + + printf("%-8s %-s %s - %s\n", HASH_TYPE_SHA256_STR, hex_result, tensor_layer_name.c_str(), hash_manifest_result_to_str(verify_result)); + } else { + printf("%-8s %-s %s\n", HASH_TYPE_SHA256_STR, hex_result, tensor_layer_name.c_str()); + } + } + + // Overall Model Hash + sha256_update( &sha256_model_hash_ctx, (unsigned char const *)raw_data, n_bytes); + } + + if (hash_params.uuid) { + SHA1Update( &sha1_for_uuid_ctx, (unsigned char const *)raw_data, n_bytes); + } + } + + if (hash_params.xxh64) { + XXH64_hash_t const hash = XXH64_digest(xxh64_model_hash_state); + + char hex_result[17]; + for (int offset = 0; offset < 8; offset++) { + unsigned int shift_bits_by = (8 * (8 - offset - 1)); + snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", (unsigned char) (hash >> shift_bits_by)&0xff); + } + + if (hash_params.manifest_is_usable) { + hash_manifest_result_t verify_result = manifest_verify(hash_params.manifest_file, HASH_TYPE_XXH64_STR, hex_result, fname); + + switch (verify_result) { + case HASH_MANIFEST_NOT_FOUND: + break; + case HASH_MANIFEST_MISMATCH: + model_in_manifest = true; + model_has_mismatch = true; + break; + case HASH_MANIFEST_OK: + model_in_manifest = true; + break; + } + + printf("%-8s %-s %s - %s\n", HASH_TYPE_XXH64_STR, hex_result, fname.c_str(), hash_manifest_result_to_str(verify_result)); + } else { + printf("%-8s %-s %s\n", HASH_TYPE_XXH64_STR, hex_result, fname.c_str()); + } + } + + if (hash_params.sha1) { + unsigned char result[21]; + SHA1Final(result, &sha1_model_hash_ctx); + + char hex_result[41]; + for (int offset = 0; offset < 20; offset++) { + snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff); + } + + if (hash_params.manifest_is_usable) { + hash_manifest_result_t verify_result = manifest_verify(hash_params.manifest_file, HASH_TYPE_SHA1_STR, hex_result, fname); + + switch (verify_result) { + case HASH_MANIFEST_NOT_FOUND: + break; + case HASH_MANIFEST_MISMATCH: + model_in_manifest = true; + model_has_mismatch = true; + break; + case HASH_MANIFEST_OK: + model_in_manifest = true; + break; + } + + printf("%-8s %-s %s - %s\n", HASH_TYPE_SHA1_STR, hex_result, fname.c_str(), hash_manifest_result_to_str(verify_result)); + } else { + printf("%-8s %-s %s\n", HASH_TYPE_SHA1_STR, hex_result, fname.c_str()); + } + } + + if (hash_params.sha256) { + unsigned char result[SHA256_DIGEST_SIZE]; // sha256 outputs 32 bytes + sha256_final( &sha256_model_hash_ctx, result); + + char hex_result[SHA256_DIGEST_SIZE * 2 + 1] = {0}; + for (int offset = 0; offset < SHA256_DIGEST_SIZE; offset++) { + snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff); + } + + if (hash_params.manifest_is_usable) { + hash_manifest_result_t verify_result = manifest_verify(hash_params.manifest_file, HASH_TYPE_SHA256_STR, hex_result, fname); + + switch (verify_result) { + case HASH_MANIFEST_NOT_FOUND: + break; + case HASH_MANIFEST_MISMATCH: + model_in_manifest = true; + model_has_mismatch = true; + break; + case HASH_MANIFEST_OK: + model_in_manifest = true; + break; + } + + printf("%-8s %-s %s - %s\n", HASH_TYPE_SHA256_STR, hex_result, fname.c_str(), hash_manifest_result_to_str(verify_result)); + } else { + printf("%-8s %-s %s\n", HASH_TYPE_SHA256_STR, hex_result, fname.c_str()); + } + } + + if (hash_params.uuid) { + unsigned char result[21]; + SHA1Final(result, &sha1_for_uuid_ctx); + + unsigned char uuid[16]; + generate_uuidv5(result, uuid); + + char string_buffer[37] = {0}; + snprintf(string_buffer, sizeof(string_buffer), "%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x", + uuid[0], uuid[1], uuid[2], uuid[3], + uuid[4], uuid[5], uuid[6], uuid[7], + uuid[8], uuid[9], uuid[10], uuid[11], + uuid[12], uuid[13], uuid[14], uuid[15]); + + if (hash_params.manifest_is_usable) { + hash_manifest_result_t verify_result = manifest_verify(hash_params.manifest_file, HASH_TYPE_SHA256_STR, string_buffer, fname); + + switch (verify_result) { + case HASH_MANIFEST_NOT_FOUND: + break; + case HASH_MANIFEST_MISMATCH: + model_in_manifest = true; + model_has_mismatch = true; + break; + case HASH_MANIFEST_OK: + model_in_manifest = true; + break; + } + + printf("%-8s %-s %s - %s\n", HASH_TYPE_UUID_STR, string_buffer, fname.c_str(), hash_manifest_result_to_str(verify_result)); + } else { + printf("%-8s %-s %s\n", HASH_TYPE_UUID_STR, string_buffer, fname.c_str()); + } + } + + + ggml_free(ctx_data); + gguf_free(ctx); + + + if (hash_params.manifest_is_usable) { + // In hash verification mode + + if (!model_in_manifest) { + // model missing in manifest? + + // Check tensor layer... + if (!tensor_layer_in_manifest) { + // Still missing? Maybe we are reading the wrong manifest. + return HASH_EXIT_MANIFEST_MISSING_ENTRY; + } + + if (tensor_layer_has_mismatch) { + // Per tensor check found error + return HASH_EXIT_FAILURE; + } + + // All per tensor layer checks passed? Sounds good enough. + return HASH_EXIT_SUCCESS; + } + + // Overall model check passed, but let's check per layer just in case + // If missing, we don't care too much as the overall model checked + if (tensor_layer_in_manifest && tensor_layer_has_mismatch) { + return HASH_EXIT_FAILURE; + } + + if (model_has_mismatch) { + // model has failed hash somewhere in the model + return HASH_EXIT_FAILURE; + } + + // All checks appears to be fine + return HASH_EXIT_SUCCESS; + } + + // In hash generation mode + return HASH_EXIT_SUCCESS; +} + +int main(int argc, const char ** argv) { + hash_params params; + manifest_check_params manifest_check; + hash_params_parse(argc, argv, params); + + if (!params.manifest_file.empty()) { + if (!manifest_type(params.manifest_file, manifest_check)) { + printf("ERROR cannot open manifest %s", params.manifest_file.c_str()); + return HASH_EXIT_MANIFEST_FILE_ERROR; + } + + if (!manifest_check.sha256 && !manifest_check.sha1 && !manifest_check.xxh64 && !manifest_check.uuid) { + printf("ERROR manifest does not have any known hash format in %s", params.manifest_file.c_str()); + return HASH_EXIT_MANIFEST_UNKNOWN_HASH; + } + + printf("manifest %s", params.manifest_file.c_str()); + + if (manifest_check.sha256) { + printf(" sha256"); + } + + if (manifest_check.sha1) { + printf(" sha1"); + } + + if (manifest_check.xxh64) { + printf(" xxh64"); + } + + if (manifest_check.uuid) { + printf(" uuid"); + } + + printf("\n"); + + // Autoselect the highest security hash if manifest is provided but + // the user has not specifically defined the hash they care about + if (!params.xxh64 && !params.sha1 && !params.uuid && !params.sha256) { + // User has not selected a specific value, pick most secure hash + if (manifest_check.sha256) { + params.sha256 = true; + } else if (manifest_check.sha1) { + params.sha1 = true; + } else if (manifest_check.xxh64) { + params.xxh64 = true; + } else if (manifest_check.uuid) { + params.uuid = true; + } + } + + params.manifest_is_usable = true; + } + + // By default if no swich argument provided, assume xxh64 + if (!params.xxh64 && !params.sha1 && !params.uuid && !params.sha256) { + params.xxh64 = true; + } + + hash_exit_code_t exit_code = gguf_hash(params); + + if (params.manifest_is_usable) { + printf("\nVerification results for %s - %s\n", params.manifest_file.c_str(), hash_exit_code_to_str(exit_code)); + } + + return exit_code; +} diff --git a/examples/gguf/gguf.cpp b/examples/gguf/gguf.cpp index 575143771..7498f85ef 100644 --- a/examples/gguf/gguf.cpp +++ b/examples/gguf/gguf.cpp @@ -92,6 +92,11 @@ static bool gguf_ex_read_0(const std::string & fname) { struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params); + if (!ctx) { + fprintf(stderr, "%s: failed to load '%s'\n", __func__, fname.c_str()); + return false; + } + printf("%s: version: %d\n", __func__, gguf_get_version(ctx)); printf("%s: alignment: %zu\n", __func__, gguf_get_alignment(ctx)); printf("%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx)); diff --git a/examples/imatrix/README.md b/examples/imatrix/README.md index 29602881a..bb5faec94 100644 --- a/examples/imatrix/README.md +++ b/examples/imatrix/README.md @@ -1,6 +1,6 @@ # llama.cpp/examples/imatrix -Compute an importance matrix for a model and given text dataset. Can be used during quantization to enchance the quality of the quantum models. +Compute an importance matrix for a model and given text dataset. Can be used during quantization to enchance the quality of the quantized models. More information is available here: https://github.com/ggerganov/llama.cpp/pull/4861 ## Usage diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp index f2f205418..48a7c366e 100644 --- a/examples/imatrix/imatrix.cpp +++ b/examples/imatrix/imatrix.cpp @@ -127,7 +127,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * } else if (e.values.size() != (size_t)src1->ne[0]*n_as) { fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as); - exit(1); //GGML_ASSERT(false); + exit(1); //GGML_ABORT("fatal error"); } if (m_params.verbosity > 1) { printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type); @@ -176,7 +176,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * } else if (e.values.size() != (size_t)src1->ne[0]) { fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]); - exit(1); //GGML_ASSERT(false); + exit(1); //GGML_ABORT("fatal error"); } ++e.ncall; if (m_params.verbosity > 1) { @@ -433,8 +433,8 @@ static void process_logits( } static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { - const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); - GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1); + const bool add_bos = llama_add_bos_token(llama_get_model(ctx)); + GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx))); const int n_ctx = llama_n_ctx(ctx); auto tim1 = std::chrono::high_resolution_clock::now(); @@ -611,10 +611,10 @@ int main(int argc, char ** argv) { params.warmup = false; // init - llama_model * model; - llama_context * ctx; + llama_init_result llama_init = llama_init_from_gpt_params(params); - std::tie(model, ctx) = llama_init_from_gpt_params(params); + llama_model * model = llama_init.model; + llama_context * ctx = llama_init.context; if (model == nullptr || ctx == nullptr) { fprintf(stderr, "%s : failed to init\n", __func__); return 1; diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index d638473e4..778461b1f 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -179,7 +179,10 @@ int main(int argc, char ** argv) { // load the model and apply lora adapter, if any LOG("%s: load the model and apply lora adapter, if any\n", __func__); - std::tie(model, ctx) = llama_init_from_gpt_params(params); + llama_init_result llama_init = llama_init_from_gpt_params(params); + + model = llama_init.model; + ctx = llama_init.context; if (model == NULL) { LOG_TEE("%s: error: unable to load model\n", __func__); @@ -200,25 +203,21 @@ int main(int argc, char ** argv) { LOG_TEE("\n"); LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str()); } - const bool add_bos = llama_should_add_bos_token(model); - GGML_ASSERT(llama_add_eos_token(model) != 1); + const bool add_bos = llama_add_bos_token(model); + GGML_ASSERT(!llama_add_eos_token(model)); LOG("add_bos: %d\n", add_bos); - bool suff_rm_leading_spc = params.escape; - if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) { - params.input_suffix.erase(0, 1); - suff_rm_leading_spc = false; - } std::vector embd_inp; std::vector embd_end; std::vector inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false); std::vector inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false); - const int space_token = 29871; - if (suff_rm_leading_spc && inp_sfx[0] == space_token) { - inp_sfx.erase(inp_sfx.begin()); - } + + GGML_ASSERT(llama_token_prefix(model) >= 0); + GGML_ASSERT(llama_token_suffix(model) >= 0); + inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model)); inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model)); + embd_inp = params.spm_infill ? inp_sfx : inp_pfx; embd_end = params.spm_infill ? inp_pfx : inp_sfx; if (add_bos) { @@ -516,19 +515,14 @@ int main(int argc, char ** argv) { string_process_escapes(params.input_prefix); string_process_escapes(params.input_suffix); } - suff_rm_leading_spc = params.escape; - if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) { - params.input_suffix.erase(0, 1); - suff_rm_leading_spc = false; - } + // tokenize new prefix and suffix std::vector inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false); std::vector inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false); - if (suff_rm_leading_spc && inp_sfx[0] == space_token) { - inp_sfx.erase(inp_sfx.begin()); - } + inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model)); inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model)); + embd_inp = params.spm_infill ? inp_sfx : inp_pfx; embd_end = params.spm_infill ? inp_pfx : inp_sfx; if (add_bos) { diff --git a/examples/json-schema-pydantic-example.py b/examples/json_schema_pydantic_example.py similarity index 95% rename from examples/json-schema-pydantic-example.py rename to examples/json_schema_pydantic_example.py index 2a24f8118..19c0bdb5b 100644 --- a/examples/json-schema-pydantic-example.py +++ b/examples/json_schema_pydantic_example.py @@ -1,9 +1,9 @@ # Usage: #! ./llama-server -m some-model.gguf & #! pip install pydantic -#! python json-schema-pydantic-example.py +#! python json_schema_pydantic_example.py -from pydantic import BaseModel, Extra, TypeAdapter +from pydantic import BaseModel, Field, TypeAdapter from annotated_types import MinLen from typing import Annotated, List, Optional import json, requests @@ -17,6 +17,9 @@ if True: The response_model param takes a type (+ supports Pydantic) and behaves just as w/ Instructor (see below) ''' + response_format = None + type_adapter = None + if response_model: type_adapter = TypeAdapter(response_model) schema = type_adapter.json_schema() diff --git a/examples/json_schema_to_grammar.py b/examples/json_schema_to_grammar.py index 072a230f7..a8779bf3b 100755 --- a/examples/json_schema_to_grammar.py +++ b/examples/json_schema_to_grammar.py @@ -1,4 +1,6 @@ #!/usr/bin/env python3 +from __future__ import annotations + import argparse import itertools import json @@ -188,7 +190,7 @@ def _generate_min_max_int(min_value: Optional[int], max_value: Optional[int], ou raise RuntimeError("At least one of min_value or max_value must be set") class BuiltinRule: - def __init__(self, content: str, deps: list = None): + def __init__(self, content: str, deps: list | None = None): self.content = content self.deps = deps or [] @@ -248,7 +250,7 @@ class SchemaConverter: def _format_literal(self, literal): escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub( - lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), literal + lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)) or m.group(0), literal ) return f'"{escaped}"' @@ -403,11 +405,11 @@ class SchemaConverter: i = 0 length = len(pattern) - def to_rule(s: Tuple[str, bool]) -> str: + def to_rule(s: tuple[str, bool]) -> str: (txt, is_literal) = s return "\"" + txt + "\"" if is_literal else txt - def transform() -> Tuple[str, bool]: + def transform() -> tuple[str, bool]: ''' Parse a unit at index i (advancing it), and return its string representation + whether it's a literal. ''' @@ -420,7 +422,7 @@ class SchemaConverter: # 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]] = [] + seq: list[tuple[str, bool]] = [] def get_dot(): if self._dotall: diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index 924c411fd..d86bae1f3 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -16,6 +16,7 @@ #include #include #include +#include #include "ggml.h" #include "llama.h" @@ -23,6 +24,18 @@ #include "ggml-cuda.h" #include "ggml-sycl.h" +#ifdef GGML_USE_CANN +#include "ggml-cann.h" +#endif + +#ifdef _WIN32 +#define WIN32_LEAN_AND_MEAN +#ifndef NOMINMAX +# define NOMINMAX +#endif +#include +#endif + // utils static uint64_t get_time_ns() { using clock = std::chrono::high_resolution_clock; @@ -92,6 +105,27 @@ static std::string get_cpu_info() { } fclose(f); } +#elif defined(_WIN32) + HKEY hKey; + if (RegOpenKeyEx(HKEY_LOCAL_MACHINE, + TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"), + 0, + KEY_READ, + &hKey) != ERROR_SUCCESS) { + // fail to open registry key + return ""; + } + char cpu_brand[256]; + DWORD cpu_brand_size = sizeof(cpu_brand); + if (RegQueryValueExA(hKey, + TEXT("ProcessorNameString"), + NULL, + NULL, + (LPBYTE)cpu_brand, + &cpu_brand_size) == ERROR_SUCCESS) { + id.assign(cpu_brand, cpu_brand_size); + } + RegCloseKey(hKey); #endif // TODO: other platforms return id; @@ -120,6 +154,17 @@ static std::string get_gpu_info() { id += "/"; } } +#endif +#ifdef GGML_USE_CANN + uint32_t count = ggml_backend_cann_get_device_count(); + for (uint32_t i = 0; i < count; i++) { + char buf[128]; + ggml_backend_cann_get_device_description(i, buf, sizeof(buf)); + id += buf; + if (i < count - 1) { + id += "/"; + } + } #endif // TODO: other backends return id; @@ -135,7 +180,7 @@ static const char * output_format_str(output_formats format) { case JSON: return "json"; case MARKDOWN: return "md"; case SQL: return "sql"; - default: GGML_ASSERT(!"invalid output format"); + default: GGML_ABORT("invalid output format"); } } @@ -161,7 +206,7 @@ static const char * split_mode_str(llama_split_mode mode) { case LLAMA_SPLIT_MODE_NONE: return "none"; case LLAMA_SPLIT_MODE_LAYER: return "layer"; case LLAMA_SPLIT_MODE_ROW: return "row"; - default: GGML_ASSERT(!"invalid split mode"); + default: GGML_ABORT("invalid split mode"); } } @@ -181,6 +226,9 @@ struct cmd_params { std::vector type_k; std::vector type_v; std::vector n_threads; + std::vector cpu_mask; + std::vector cpu_strict; + std::vector poll; std::vector n_gpu_layers; std::vector rpc_servers; std::vector split_mode; @@ -192,6 +240,8 @@ struct cmd_params { std::vector embeddings; ggml_numa_strategy numa; int reps; + ggml_sched_priority prio; + int delay; bool verbose; output_formats output_format; output_formats output_format_stderr; @@ -207,6 +257,9 @@ static const cmd_params cmd_params_defaults = { /* type_k */ {GGML_TYPE_F16}, /* type_v */ {GGML_TYPE_F16}, /* n_threads */ {cpu_get_num_math()}, + /* cpu_mask */ {"0x0"}, + /* cpu_strict */ {false}, + /* poll */ {50}, /* n_gpu_layers */ {99}, /* rpc_servers */ {""}, /* split_mode */ {LLAMA_SPLIT_MODE_LAYER}, @@ -218,6 +271,8 @@ static const cmd_params cmd_params_defaults = { /* embeddings */ {false}, /* numa */ GGML_NUMA_STRATEGY_DISABLED, /* reps */ 5, + /* prio */ GGML_SCHED_PRIO_NORMAL, + /* delay */ 0, /* verbose */ false, /* output_format */ MARKDOWN, /* output_format_stderr */ NONE, @@ -237,6 +292,9 @@ static void print_usage(int /* argc */, char ** argv) { printf(" -ctk, --cache-type-k (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str()); printf(" -ctv, --cache-type-v (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str()); printf(" -t, --threads (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str()); + printf(" -C, --cpu-mask (default: %s)\n", join(cmd_params_defaults.cpu_mask, ",").c_str()); + printf(" --cpu-strict <0|1> (default: %s)\n", join(cmd_params_defaults.cpu_strict, ",").c_str()); + printf(" --poll <0...100> (default: %s)\n", join(cmd_params_defaults.poll, ",").c_str()); printf(" -ngl, --n-gpu-layers (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str()); printf(" -rpc, --rpc (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str()); printf(" -sm, --split-mode (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str()); @@ -248,6 +306,8 @@ static void print_usage(int /* argc */, char ** argv) { printf(" -embd, --embeddings <0|1> (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str()); printf(" -ts, --tensor-split (default: 0)\n"); printf(" -r, --repetitions (default: %d)\n", cmd_params_defaults.reps); + printf(" --prio <0|1|2|3> (default: %d)\n", cmd_params_defaults.prio); + printf(" --delay <0...N> (seconds) (default: %d)\n", cmd_params_defaults.delay); printf(" -o, --output (default: %s)\n", output_format_str(cmd_params_defaults.output_format)); printf(" -oe, --output-err (default: %s)\n", output_format_str(cmd_params_defaults.output_format_stderr)); printf(" -v, --verbose (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0"); @@ -294,6 +354,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { params.output_format_stderr = cmd_params_defaults.output_format_stderr; params.reps = cmd_params_defaults.reps; params.numa = cmd_params_defaults.numa; + params.prio = cmd_params_defaults.prio; + params.delay = cmd_params_defaults.delay; for (int i = 1; i < argc; i++) { arg = argv[i]; @@ -389,6 +451,27 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { } auto p = string_split(argv[i], split_delim); params.n_threads.insert(params.n_threads.end(), p.begin(), p.end()); + } else if (arg == "-C" || arg == "--cpu-mask") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + params.cpu_mask.insert(params.cpu_mask.end(), p.begin(), p.end()); + } else if (arg == "--cpu-strict") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + params.cpu_strict.insert(params.cpu_strict.end(), p.begin(), p.end()); + } else if (arg == "--poll") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + params.poll.insert(params.poll.end(), p.begin(), p.end()); } else if (arg == "-ngl" || arg == "--n-gpu-layers") { if (++i >= argc) { invalid_param = true; @@ -497,6 +580,18 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { break; } params.reps = std::stoi(argv[i]); + } else if (arg == "--prio") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.prio = (enum ggml_sched_priority) std::stoi(argv[i]); + } else if (arg == "--delay") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.delay = std::stoi(argv[i]); } else if (arg == "-o" || arg == "--output") { if (++i >= argc) { invalid_param = true; @@ -541,6 +636,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { if (params.use_mmap.empty()) { params.use_mmap = cmd_params_defaults.use_mmap; } if (params.embeddings.empty()) { params.embeddings = cmd_params_defaults.embeddings; } if (params.n_threads.empty()) { params.n_threads = cmd_params_defaults.n_threads; } + if (params.cpu_mask.empty()) { params.cpu_mask = cmd_params_defaults.cpu_mask; } + if (params.cpu_strict.empty()) { params.cpu_strict = cmd_params_defaults.cpu_strict; } + if (params.poll.empty()) { params.poll = cmd_params_defaults.poll; } return params; } @@ -554,6 +652,9 @@ struct cmd_params_instance { ggml_type type_k; ggml_type type_v; int n_threads; + std::string cpu_mask; + bool cpu_strict; + int poll; int n_gpu_layers; std::string rpc_servers; llama_split_mode split_mode; @@ -623,7 +724,10 @@ static std::vector get_cmd_params_instances(const cmd_param for (const auto & tv : params.type_v) for (const auto & nkvo : params.no_kv_offload) for (const auto & fa : params.flash_attn) - for (const auto & nt : params.n_threads) { + for (const auto & nt : params.n_threads) + for (const auto & cm : params.cpu_mask) + for (const auto & cs : params.cpu_strict) + for (const auto & pl : params.poll) { for (const auto & n_prompt : params.n_prompt) { if (n_prompt == 0) { continue; @@ -637,6 +741,9 @@ static std::vector get_cmd_params_instances(const cmd_param /* .type_k = */ tk, /* .type_v = */ tv, /* .n_threads = */ nt, + /* .cpu_mask = */ cm, + /* .cpu_strict = */ cs, + /* .poll = */ pl, /* .n_gpu_layers = */ nl, /* .rpc_servers = */ rpc, /* .split_mode = */ sm, @@ -663,6 +770,9 @@ static std::vector get_cmd_params_instances(const cmd_param /* .type_k = */ tk, /* .type_v = */ tv, /* .n_threads = */ nt, + /* .cpu_mask = */ cm, + /* .cpu_strict = */ cs, + /* .poll = */ pl, /* .n_gpu_layers = */ nl, /* .rpc_servers = */ rpc, /* .split_mode = */ sm, @@ -689,6 +799,9 @@ static std::vector get_cmd_params_instances(const cmd_param /* .type_k = */ tk, /* .type_v = */ tv, /* .n_threads = */ nt, + /* .cpu_mask = */ cm, + /* .cpu_strict = */ cs, + /* .poll = */ pl, /* .n_gpu_layers = */ nl, /* .rpc_servers = */ rpc, /* .split_mode = */ sm, @@ -725,6 +838,9 @@ struct test { int n_batch; int n_ubatch; int n_threads; + std::string cpu_mask; + bool cpu_strict; + int poll; bool has_rpc; ggml_type type_k; ggml_type type_v; @@ -751,6 +867,9 @@ struct test { n_batch = inst.n_batch; n_ubatch = inst.n_ubatch; n_threads = inst.n_threads; + cpu_mask = inst.cpu_mask; + cpu_strict = inst.cpu_strict; + poll = inst.poll; has_rpc = !inst.rpc_servers.empty(); type_k = inst.type_k; type_v = inst.type_v; @@ -828,13 +947,14 @@ struct test { "cpu_info", "gpu_info", "model_filename", "model_type", "model_size", "model_n_params", "n_batch", "n_ubatch", - "n_threads", "type_k", "type_v", + "n_threads", "cpu_mask", "cpu_strict", "poll", + "type_k", "type_v", "n_gpu_layers", "split_mode", "main_gpu", "no_kv_offload", "flash_attn", "tensor_split", "use_mmap", "embeddings", "n_prompt", "n_gen", "test_time", "avg_ns", "stddev_ns", - "avg_ts", "stddev_ts" + "avg_ts", "stddev_ts", }; return fields; } @@ -843,7 +963,7 @@ struct test { static field_type get_field_type(const std::string & field) { if (field == "build_number" || field == "n_batch" || field == "n_ubatch" || - field == "n_threads" || + field == "n_threads" || field == "poll" || field == "model_size" || field == "model_n_params" || field == "n_gpu_layers" || field == "main_gpu" || field == "n_prompt" || field == "n_gen" || @@ -852,6 +972,7 @@ struct test { } if (field == "cuda" || field == "vulkan" || field == "kompute" || field == "metal" || field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" || + field == "cpu_strict" || field == "flash_attn" || field == "use_mmap" || field == "embeddings") { return BOOL; } @@ -884,7 +1005,8 @@ struct test { cpu_info, gpu_info, model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params), std::to_string(n_batch), std::to_string(n_ubatch), - std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v), + std::to_string(n_threads), cpu_mask, std::to_string(cpu_strict), std::to_string(poll), + ggml_type_name(type_k), ggml_type_name(type_v), std::to_string(n_gpu_layers), split_mode_str(split_mode), std::to_string(main_gpu), std::to_string(no_kv_offload), std::to_string(flash_attn), tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings), @@ -1023,7 +1145,7 @@ struct markdown_printer : public printer { return -30; } if (field == "t/s") { - return 16; + return 20; } if (field == "size" || field == "params") { return 10; @@ -1105,6 +1227,15 @@ struct markdown_printer : public printer { if (params.n_threads.size() > 1 || params.n_threads != cmd_params_defaults.n_threads || is_cpu_backend) { fields.emplace_back("n_threads"); } + if (params.cpu_mask.size() > 1 || params.cpu_mask != cmd_params_defaults.cpu_mask) { + fields.emplace_back("cpu_mask"); + } + if (params.cpu_strict.size() > 1 || params.cpu_strict != cmd_params_defaults.cpu_strict) { + fields.emplace_back("cpu_strict"); + } + if (params.poll.size() > 1 || params.poll != cmd_params_defaults.poll) { + fields.emplace_back("poll"); + } if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) { fields.emplace_back("n_batch"); } @@ -1311,7 +1442,7 @@ static std::unique_ptr create_printer(output_formats format) { case SQL: return std::unique_ptr(new sql_printer()); } - GGML_ASSERT(false); + GGML_ABORT("fatal error"); } int main(int argc, char ** argv) { @@ -1339,6 +1470,8 @@ int main(int argc, char ** argv) { llama_backend_init(); llama_numa_init(params.numa); + set_process_priority(params.prio); + // initialize printer std::unique_ptr p = create_printer(params.output_format); std::unique_ptr p_err = create_printer(params.output_format_stderr); @@ -1384,6 +1517,28 @@ int main(int argc, char ** argv) { llama_past_clear(ctx); + // cool off before the test + if (params.delay) { + std::this_thread::sleep_for(std::chrono::seconds(params.delay)); + } + + struct ggml_threadpool_params tpp = ggml_threadpool_params_default(t.n_threads); + if (!parse_cpu_mask(t.cpu_mask, tpp.cpumask)) { + LOG_TEE("%s: failed to parse cpu-mask: %s\n", __func__, t.cpu_mask.c_str()); + exit(1); + } + tpp.strict_cpu = t.cpu_strict; + tpp.poll = t.poll; + tpp.prio = params.prio; + + struct ggml_threadpool* threadpool = ggml_threadpool_new(&tpp); + if (!threadpool) { + LOG_TEE("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads); + exit(1); + } + + llama_attach_threadpool(ctx, threadpool, NULL); + // warmup run if (t.n_prompt > 0) { //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads); @@ -1422,6 +1577,8 @@ int main(int argc, char ** argv) { llama_print_timings(ctx); llama_free(ctx); + + ggml_threadpool_free(threadpool); } llama_free_model(lmodel); diff --git a/examples/llama.android/llama/src/main/cpp/llama-android.cpp b/examples/llama.android/llama/src/main/cpp/llama-android.cpp index 09bf6cc30..c5366d97f 100644 --- a/examples/llama.android/llama/src/main/cpp/llama-android.cpp +++ b/examples/llama.android/llama/src/main/cpp/llama-android.cpp @@ -409,7 +409,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop( const auto n_cur = env->CallIntMethod(intvar_ncur, la_int_var_value); if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) { - return env->NewStringUTF(""); + return nullptr; } auto new_token_chars = llama_token_to_piece(context, new_token_id); diff --git a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift index 50fcaa12d..f893de7a5 100644 --- a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift +++ b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift @@ -26,11 +26,12 @@ actor LlamaContext { private var context: OpaquePointer private var batch: llama_batch private var tokens_list: [llama_token] + var is_done: Bool = false /// This variable is used to store temporarily invalid cchars private var temporary_invalid_cchars: [CChar] - var n_len: Int32 = 64 + var n_len: Int32 = 1024 var n_cur: Int32 = 0 var n_decode: Int32 = 0 @@ -70,8 +71,8 @@ actor LlamaContext { var ctx_params = llama_context_default_params() ctx_params.seed = 1234 ctx_params.n_ctx = 2048 - ctx_params.n_threads = UInt32(n_threads) - ctx_params.n_threads_batch = UInt32(n_threads) + ctx_params.n_threads = Int32(n_threads) + ctx_params.n_threads_batch = Int32(n_threads) let context = llama_new_context_with_model(model, ctx_params) guard let context else { @@ -160,6 +161,7 @@ actor LlamaContext { if llama_token_is_eog(model, new_token_id) || n_cur == n_len { print("\n") + is_done = true let new_token_str = String(cString: temporary_invalid_cchars + [0]) temporary_invalid_cchars.removeAll() return new_token_str @@ -322,7 +324,7 @@ actor LlamaContext { defer { result.deallocate() } - let nTokens = llama_token_to_piece(model, token, result, 8, false) + let nTokens = llama_token_to_piece(model, token, result, 8, 0, false) if nTokens < 0 { let newResult = UnsafeMutablePointer.allocate(capacity: Int(-nTokens)) @@ -330,7 +332,7 @@ actor LlamaContext { defer { newResult.deallocate() } - let nNewTokens = llama_token_to_piece(model, token, newResult, -nTokens, false) + let nNewTokens = llama_token_to_piece(model, token, newResult, -nTokens, 0, false) let bufferPointer = UnsafeBufferPointer(start: newResult, count: Int(nNewTokens)) return Array(bufferPointer) } else { diff --git a/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift b/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift index 2c1e3f61b..b8f6a31d5 100644 --- a/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift +++ b/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift @@ -132,7 +132,7 @@ class LlamaState: ObservableObject { messageLog += "\(text)" Task.detached { - while await llamaContext.n_cur < llamaContext.n_len { + while await !llamaContext.is_done { let result = await llamaContext.completion_loop() await MainActor.run { self.messageLog += "\(result)" diff --git a/examples/llava/CMakeLists.txt b/examples/llava/CMakeLists.txt index e9fa73acb..bbf5fec58 100644 --- a/examples/llava/CMakeLists.txt +++ b/examples/llava/CMakeLists.txt @@ -36,3 +36,10 @@ set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-cli) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) + +set(TARGET llama-minicpmv-cli) +add_executable(${TARGET} minicpmv-cli.cpp) +set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-minicpmv-cli) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/llava/MobileVLM-README.md b/examples/llava/MobileVLM-README.md index f6c619c87..06a65fba4 100644 --- a/examples/llava/MobileVLM-README.md +++ b/examples/llava/MobileVLM-README.md @@ -30,16 +30,16 @@ git clone https://huggingface.co/mtgv/MobileVLM-1.7B git clone https://huggingface.co/openai/clip-vit-large-patch14-336 ``` -2. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents: +2. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents: ```sh -python ./examples/llava/llava-surgery.py -m path/to/MobileVLM-1.7B +python ./examples/llava/llava_surgery.py -m path/to/MobileVLM-1.7B ``` -3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF: +3. Use `convert_image_encoder_to_gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF: ```sh -python ./examples/llava/convert-image-encoder-to-gguf \ +python ./examples/llava/convert_image_encoder_to_gguf \ -m path/to/clip-vit-large-patch14-336 \ --llava-projector path/to/MobileVLM-1.7B/llava.projector \ --output-dir path/to/MobileVLM-1.7B \ @@ -47,17 +47,17 @@ python ./examples/llava/convert-image-encoder-to-gguf \ ``` ```sh -python ./examples/llava/convert-image-encoder-to-gguf \ +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 `examples/convert-legacy-llama.py` to convert the LLaMA part of LLaVA to GGUF: +4. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF: ```sh -python ./examples/convert-legacy-llama.py path/to/MobileVLM-1.7B +python ./examples/convert_legacy_llama.py path/to/MobileVLM-1.7B ``` 5. Use `quantize` to convert LLaMA part's DataType from `fp16` to `q4_k` diff --git a/examples/llava/README-minicpmv2.5.md b/examples/llava/README-minicpmv2.5.md new file mode 100644 index 000000000..1c8498ff9 --- /dev/null +++ b/examples/llava/README-minicpmv2.5.md @@ -0,0 +1,99 @@ +## MiniCPM-Llama3-V 2.5 + +### Prepare models and code + +Download [MiniCPM-Llama3-V-2_5](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5) PyTorch model from huggingface to "MiniCPM-Llama3-V-2_5" folder. + +Clone llama.cpp: +```bash +git clone https://github.com/ggerganov/llama.cpp +cd llama.cpp +``` + +### Usage + +Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us) + +```bash +python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5 +python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2 +python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model + +# quantize int4 version +./llama-quantize ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf Q4_K_M +``` + +Build for Linux or Mac + +```bash +make +make llama-minicpmv-cli +``` + +Inference on Linux or Mac +``` +# run f16 version +./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" + +# run quantized int4 version +./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" + +# or run in interactive mode +./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i +``` + +### Android + +#### Build on Android device using Termux +We found that build on Android device would bring better runtime performance, so we recommend to build on device. + +[Termux](https://github.com/termux/termux-app#installation) is a terminal app on Android device (no root required). + +Install tools in Termux: +``` +apt update && apt upgrade -y +apt install git make cmake +``` + +It's recommended to move your model inside the `~/` directory for best performance: +``` +cd storage/downloads +mv model.gguf ~/ +``` + +#### Building the Project using Android NDK +Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake. + +Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux: + +```bash +mkdir build-android +cd build-android +export NDK=/your_ndk_path +cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod .. +make +``` + +Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice). + +Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission: + +(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`) +``` +$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/ +$cd /data/data/com.termux/files/home/bin +$chmod +x ./* +``` + +Download models and push them to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/` + +``` +$mv /sdcard/llama.cpp/ggml-model-Q4_K_M.gguf /data/data/com.termux/files/home/model/ +$mv /sdcard/llama.cpp/mmproj-model-f16.gguf /data/data/com.termux/files/home/model/ +``` + +Now, you can start chatting: +``` +$cd /data/data/com.termux/files/home/bin +$./llama-minicpmv-cli -m ../model/ggml-model-Q4_K_M.gguf --mmproj ../model/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" +``` diff --git a/examples/llava/README-minicpmv2.6.md b/examples/llava/README-minicpmv2.6.md new file mode 100644 index 000000000..c4be5e5dd --- /dev/null +++ b/examples/llava/README-minicpmv2.6.md @@ -0,0 +1,107 @@ +## MiniCPM-V 2.6 + +### Prepare models and code + +Download [MiniCPM-V-2_6](https://huggingface.co/openbmb/MiniCPM-V-2_6) PyTorch model from huggingface to "MiniCPM-V-2_6" folder. + +Clone llama.cpp: +```bash +git clone git@github.com:OpenBMB/llama.cpp.git +cd llama.cpp +git checkout minicpmv-main +``` + +### Usage of MiniCPM-V 2.6 + +Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) by us) + +```bash +python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-V-2_6 +python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3 +python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model + +# quantize int4 version +./llama-quantize ../MiniCPM-V-2_6/model/ggml-model-f16.gguf ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M +``` + +Build for Linux or Mac + +```bash +make +make llama-minicpmv-cli +``` + +Inference on Linux or Mac +``` +# run f16 version +./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" + +# run quantized int4 version +./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" + +# or run in interactive mode +./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i +``` + +### Video +Install FFmpeg +``` +brew install ffmpeg +brew install pkg-config +``` + +### Android + +#### Build on Android device using Termux +We found that build on Android device would bring better runtime performance, so we recommend to build on device. + +[Termux](https://github.com/termux/termux-app#installation) is a terminal app on Android device (no root required). + +Install tools in Termux: +``` +apt update && apt upgrade -y +apt install git make cmake +``` + +It's recommended to move your model inside the `~/` directory for best performance: +``` +cd storage/downloads +mv model.gguf ~/ +``` + +#### Building the Project using Android NDK +Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake. + +Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux: + +```bash +mkdir build-android +cd build-android +export NDK=/your_ndk_path +cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod .. +make +``` + +Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice). + +Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission: + +(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`) +``` +$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/ +$cd /data/data/com.termux/files/home/bin +$chmod +x ./* +``` + +Download models and push them to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/` + +``` +$mv /sdcard/llama.cpp/ggml-model-Q4_K_M.gguf /data/data/com.termux/files/home/model/ +$mv /sdcard/llama.cpp/mmproj-model-f16.gguf /data/data/com.termux/files/home/model/ +``` + +Now, you can start chatting: +``` +$cd /data/data/com.termux/files/home/bin +$./llama-minicpmv-cli -m ../model/ggml-model-Q4_K_M.gguf --mmproj ../model/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" +``` diff --git a/examples/llava/README.md b/examples/llava/README.md index f4554de67..012451361 100644 --- a/examples/llava/README.md +++ b/examples/llava/README.md @@ -38,22 +38,22 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336 pip install -r examples/llava/requirements.txt ``` -3. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents: +3. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents: ```sh -python ./examples/llava/llava-surgery.py -m ../llava-v1.5-7b +python ./examples/llava/llava_surgery.py -m ../llava-v1.5-7b ``` -4. Use `convert-image-encoder-to-gguf.py` to convert the LLaVA image encoder to GGUF: +4. Use `convert_image_encoder_to_gguf.py` to convert the LLaVA image encoder to GGUF: ```sh -python ./examples/llava/convert-image-encoder-to-gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b +python ./examples/llava/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b ``` -5. Use `examples/convert-legacy-llama.py` to convert the LLaMA part of LLaVA to GGUF: +5. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF: ```sh -python ./examples/convert-legacy-llama.py ../llava-v1.5-7b --skip-unknown +python ./examples/convert_legacy_llama.py ../llava-v1.5-7b --skip-unknown ``` Now both the LLaMA part and the image encoder are in the `llava-v1.5-7b` directory. @@ -70,9 +70,9 @@ git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b pip install -r examples/llava/requirements.txt ``` -3) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models: +3) Use `llava_surgery_v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models: ```console -python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/ +python examples/llava/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/ ``` - you will find a llava.projector and a llava.clip file in your model directory @@ -86,13 +86,13 @@ curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.jso 5) Create the visual gguf model: ```console -python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision +python ./examples/llava/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision ``` - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP 6) Then convert the model to gguf format: ```console -python ./examples/convert-legacy-llama.py ../llava-v1.6-vicuna-7b/ --skip-unknown +python ./examples/convert_legacy_llama.py ../llava-v1.6-vicuna-7b/ --skip-unknown ``` 7) And finally we can run the llava cli using the 1.6 model version: diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index d6882eec3..9b890571e 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -16,6 +16,14 @@ #include "ggml-metal.h" #endif +#ifdef GGML_USE_CANN +#include "ggml-cann.h" +#endif + +#ifdef GGML_USE_VULKAN +#include "ggml-vulkan.h" +#endif + #define STB_IMAGE_IMPLEMENTATION #include "stb_image.h" @@ -70,26 +78,28 @@ static std::string format(const char * fmt, ...) { // key constants // -#define KEY_FTYPE "general.file_type" -#define KEY_NAME "general.name" -#define KEY_DESCRIPTION "general.description" -#define KEY_HAS_TEXT_ENC "clip.has_text_encoder" -#define KEY_HAS_VIS_ENC "clip.has_vision_encoder" -#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector" -#define KEY_USE_GELU "clip.use_gelu" -#define KEY_N_EMBD "clip.%s.embedding_length" -#define KEY_N_FF "clip.%s.feed_forward_length" -#define KEY_N_BLOCK "clip.%s.block_count" -#define KEY_N_HEAD "clip.%s.attention.head_count" -#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon" -#define KEY_PROJ_DIM "clip.%s.projection_dim" -#define KEY_TOKENS "tokenizer.ggml.tokens" -#define KEY_N_POSITIONS "clip.text.context_length" -#define KEY_IMAGE_SIZE "clip.vision.image_size" -#define KEY_PATCH_SIZE "clip.vision.patch_size" -#define KEY_IMAGE_MEAN "clip.vision.image_mean" -#define KEY_IMAGE_STD "clip.vision.image_std" -#define KEY_PROJ_TYPE "clip.projector_type" +#define KEY_FTYPE "general.file_type" +#define KEY_NAME "general.name" +#define KEY_DESCRIPTION "general.description" +#define KEY_HAS_TEXT_ENC "clip.has_text_encoder" +#define KEY_HAS_VIS_ENC "clip.has_vision_encoder" +#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector" +#define KEY_HAS_MINICPMV_PROJ "clip.has_minicpmv_projector" +#define KEY_MINICPMV_VERSION "clip.minicpmv_version" +#define KEY_USE_GELU "clip.use_gelu" +#define KEY_N_EMBD "clip.%s.embedding_length" +#define KEY_N_FF "clip.%s.feed_forward_length" +#define KEY_N_BLOCK "clip.%s.block_count" +#define KEY_N_HEAD "clip.%s.attention.head_count" +#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon" +#define KEY_PROJ_DIM "clip.%s.projection_dim" +#define KEY_TOKENS "tokenizer.ggml.tokens" +#define KEY_N_POSITIONS "clip.text.context_length" +#define KEY_IMAGE_SIZE "clip.vision.image_size" +#define KEY_PATCH_SIZE "clip.vision.patch_size" +#define KEY_IMAGE_MEAN "clip.vision.image_mean" +#define KEY_IMAGE_STD "clip.vision.image_std" +#define KEY_PROJ_TYPE "clip.projector_type" #define KEY_MM_PATCH_MERGE_TYPE "clip.vision.mm_patch_merge_type" #define KEY_IMAGE_GRID_PINPOINTS "clip.vision.image_grid_pinpoints" @@ -123,12 +133,20 @@ static std::string format(const char * fmt, ...) { #define TN_MVLM_PROJ_PEG "mm.model.peg.%d.%s" #define TN_IMAGE_NEWLINE "model.image_newline" +#define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k" +#define TN_MINICPMV_QUERY "resampler.query" +#define TN_MINICPMV_PROJ "resampler.proj.weight" +#define TN_MINICPMV_KV_PROJ "resampler.kv.weight" +#define TN_MINICPMV_ATTN "resampler.attn.%s.%s" +#define TN_MINICPMV_LN "resampler.ln_%s.%s" + enum projector_type { PROJECTOR_TYPE_MLP, PROJECTOR_TYPE_MLP_NORM, PROJECTOR_TYPE_LDP, PROJECTOR_TYPE_LDPV2, + PROJECTOR_TYPE_RESAMPLER, PROJECTOR_TYPE_UNKNOWN, }; @@ -136,6 +154,7 @@ static std::map PROJECTOR_TYPE_NAMES = { { PROJECTOR_TYPE_MLP, "mlp" }, { PROJECTOR_TYPE_LDP, "ldp" }, { PROJECTOR_TYPE_LDPV2, "ldpv2"}, + { PROJECTOR_TYPE_RESAMPLER, "resampler"}, }; @@ -196,17 +215,20 @@ static std::string gguf_data_to_str(enum gguf_type type, const void * data, int } static void replace_all(std::string & s, const std::string & search, const std::string & replace) { - std::string result; - for (size_t pos = 0; ; pos += search.length()) { - auto new_pos = s.find(search, pos); - if (new_pos == std::string::npos) { - result += s.substr(pos, s.size() - pos); - break; - } - result += s.substr(pos, new_pos - pos) + replace; - pos = new_pos; + if (search.empty()) { + return; } - s = std::move(result); + std::string builder; + builder.reserve(s.length()); + size_t pos = 0; + size_t last_pos = 0; + while ((pos = s.find(search, last_pos)) != std::string::npos) { + builder.append(s, last_pos, pos - last_pos); + builder.append(replace); + last_pos = pos + search.length(); + } + builder.append(s, last_pos, std::string::npos); + s = std::move(builder); } static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) { @@ -488,12 +510,34 @@ struct clip_vision_model { struct ggml_tensor * mm_model_mlp_2_b; struct ggml_tensor * mm_model_peg_0_w; struct ggml_tensor * mm_model_peg_0_b; + + // MINICPMV projection + struct ggml_tensor * mm_model_pos_embed_k; + struct ggml_tensor * mm_model_query; + struct ggml_tensor * mm_model_proj; + struct ggml_tensor * mm_model_kv_proj; + struct ggml_tensor * mm_model_attn_q_w; + struct ggml_tensor * mm_model_attn_q_b; + struct ggml_tensor * mm_model_attn_k_w; + struct ggml_tensor * mm_model_attn_k_b; + struct ggml_tensor * mm_model_attn_v_w; + struct ggml_tensor * mm_model_attn_v_b; + struct ggml_tensor * mm_model_attn_o_w; + struct ggml_tensor * mm_model_attn_o_b; + struct ggml_tensor * mm_model_ln_q_w; + struct ggml_tensor * mm_model_ln_q_b; + struct ggml_tensor * mm_model_ln_kv_w; + struct ggml_tensor * mm_model_ln_kv_b; + struct ggml_tensor * mm_model_ln_post_w; + struct ggml_tensor * mm_model_ln_post_b; }; struct clip_ctx { bool has_text_encoder = false; bool has_vision_encoder = false; bool has_llava_projector = false; + bool has_minicpmv_projector = false; + int minicpmv_version = 2; struct clip_vision_model vision_model; projector_type proj_type = PROJECTOR_TYPE_MLP; @@ -518,9 +562,11 @@ struct clip_ctx { ggml_backend_t backend = NULL; ggml_gallocr_t compute_alloc = NULL; + + struct clip_image_size * load_image_size; }; -static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs) { +static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) { if (!ctx->has_vision_encoder) { LOG_TEE("This gguf file seems to have no vision encoder\n"); return nullptr; @@ -529,20 +575,33 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 const auto & model = ctx->vision_model; const auto & hparams = model.hparams; - const int image_size = hparams.image_size; + const int image_size = hparams.image_size; + int image_size_width = image_size; + int image_size_height = image_size; + if (ctx->has_minicpmv_projector) { + if (load_image_size == nullptr) { + load_image_size = clip_image_size_init(); + } + LOG_TEE("%s: %d %d\n", __func__, load_image_size->width, load_image_size->height); + image_size_width = load_image_size->width; + image_size_height = load_image_size->height; + if (is_inf) { + image_size_width = imgs->data->nx; + image_size_height = imgs->data->ny; + } + } const int patch_size = hparams.patch_size; - const int num_patches = ((image_size / patch_size) * (image_size / patch_size)); - const int num_patches_per_side = image_size / patch_size; GGML_UNUSED(num_patches_per_side); + const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size)); const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0); const int hidden_size = hparams.hidden_size; const int n_head = hparams.n_head; const int d_head = hidden_size / n_head; - const int n_layer = hparams.n_layer; + int n_layer = hparams.n_layer; const float eps = hparams.eps; const int batch_size = imgs->size; - if (ctx->has_llava_projector) { + if (ctx->has_llava_projector || ctx->has_minicpmv_projector) { GGML_ASSERT(batch_size == 1); } @@ -555,7 +614,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 struct ggml_context * ctx0 = ggml_init(params); struct ggml_cgraph * gf = ggml_new_graph(ctx0); - struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size, image_size, 3, batch_size); + struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3, batch_size); ggml_set_name(inp_raw, "inp_raw"); ggml_set_input(inp_raw); @@ -568,19 +627,21 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp)); inp = ggml_add(ctx0, inp, model.patch_bias); } - - // concat class_embeddings and patch_embeddings struct ggml_tensor * embeddings = inp; - if (ctx->has_class_embedding) { - embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size); - ggml_set_name(embeddings, "embeddings"); - ggml_set_input(embeddings); - embeddings = ggml_acc(ctx0, embeddings, model.class_embedding, - embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0); - embeddings = ggml_acc(ctx0, embeddings, inp, - embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]); - } + struct ggml_tensor * pos_embed = nullptr; + if (ctx->has_llava_projector) { + // concat class_embeddings and patch_embeddings + if (ctx->has_class_embedding) { + embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size); + ggml_set_name(embeddings, "embeddings"); + ggml_set_input(embeddings); + embeddings = ggml_acc(ctx0, embeddings, model.class_embedding, + embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0); + embeddings = ggml_acc(ctx0, embeddings, inp, + embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]); + } + } struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions); ggml_set_name(positions, "positions"); @@ -589,6 +650,19 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 embeddings = ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions)); + if (ctx->has_minicpmv_projector) { + int pos_w = image_size_width/patch_size; + int pos_h = image_size_height/patch_size; + if (ctx->minicpmv_version == 2) { + pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 4096, pos_w * pos_h, 1); + } + else if (ctx->minicpmv_version == 3) { + pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 3584, pos_w * pos_h, 1); + } + ggml_set_name(pos_embed, "pos_embed"); + ggml_set_input(pos_embed); + } + // pre-layernorm if (ctx->has_pre_norm) { embeddings = ggml_norm(ctx0, embeddings, eps); @@ -598,6 +672,9 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 } // loop over layers + if (ctx->has_minicpmv_projector) { + n_layer += 1; + } for (int il = 0; il < n_layer - 1; il++) { struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states @@ -687,7 +764,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 } // llava projector - { + if (ctx->has_llava_projector) { embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]); struct ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_patches); @@ -708,8 +785,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 embeddings = ggml_gelu(ctx0, embeddings); embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings); embeddings = ggml_add(ctx0, embeddings, model.mm_2_b); - - } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { + } + else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false); @@ -864,6 +941,75 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 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_ABORT("fatal error"); + } + } + // minicpmv projector + else if (ctx->has_minicpmv_projector) + { + if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) { + struct ggml_tensor * q = model.mm_model_query; + { // layernorm + q = ggml_norm(ctx0, q, eps); + q = ggml_add(ctx0, ggml_mul(ctx0, q, model.mm_model_ln_q_w), model.mm_model_ln_q_b); + } + struct ggml_tensor * v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings); + { // layernorm + v = ggml_norm(ctx0, v, eps); + v = ggml_add(ctx0, ggml_mul(ctx0, v, model.mm_model_ln_kv_w), model.mm_model_ln_kv_b); + } + struct ggml_tensor * k; + { // position + // q = ggml_add(ctx0, q, model.mm_model_pos_embed); + k = ggml_add(ctx0, v, pos_embed); + } + + { // attention + int hidden_size = 4096; + const int d_head = 128; + int n_head = hidden_size/d_head; + int num_query = 96; + if (ctx->minicpmv_version == 2) { + hidden_size = 4096; + n_head = hidden_size/d_head; + num_query = 96; + } + else if (ctx->minicpmv_version == 3) { + hidden_size = 3584; + n_head = hidden_size/d_head; + num_query = 64; + } + + struct ggml_tensor * Q = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q), model.mm_model_attn_q_b); + Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head)); + struct ggml_tensor * K = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_k_w, k), model.mm_model_attn_k_b); + struct ggml_tensor * V = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_v_w, v), model.mm_model_attn_v_b); + // permute + Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_query, batch_size); + Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3)); + Q = ggml_reshape_3d(ctx0, Q, d_head, num_query, n_head * batch_size); + K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size); + K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3)); + K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size); + V = ggml_reshape_4d(ctx0, V, d_head, n_head, num_positions, batch_size); + V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3)); + V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size); + struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + KQ = ggml_soft_max_inplace(ctx0, KQ); + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ); + KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_query, n_head, batch_size); + KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3); + KQV = ggml_cont_3d(ctx0, KQV, hidden_size, num_query, batch_size); + + embeddings = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_o_w, KQV), model.mm_model_attn_o_b); + } + { // layernorm + embeddings = ggml_norm(ctx0, embeddings, eps); + embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_post_w), model.mm_model_ln_post_b); + } + embeddings = ggml_mul_mat(ctx0, model.mm_model_proj, embeddings); + } else { GGML_ASSERT(false); } @@ -972,7 +1118,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { } } - clip_ctx * new_clip = new clip_ctx; + clip_ctx * new_clip = new clip_ctx{}; // update projector type { @@ -1001,6 +1147,15 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { LOG_TEE("%s: CLIP using Metal backend\n", __func__); #endif +#ifdef GGML_USE_CANN + new_clip->backend = ggml_backend_cann_init(0); + LOG_TEE("%s: CLIP using CANN backend\n", __func__); +#endif + +#ifdef GGML_USE_VULKAN + new_clip->backend = ggml_backend_vk_init(0); + LOG_TEE("%s: CLIP using Vulkan backend\n", __func__); +#endif if (!new_clip->backend) { new_clip->backend = ggml_backend_cpu_init(); @@ -1020,7 +1175,18 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { new_clip->has_llava_projector = gguf_get_val_bool(ctx, idx); } - GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search + idx = gguf_find_key(ctx, KEY_HAS_MINICPMV_PROJ); + if (idx != -1) { + new_clip->has_minicpmv_projector = gguf_get_val_bool(ctx, idx); + } + + idx = gguf_find_key(ctx, KEY_MINICPMV_VERSION); + if (idx != -1) { + new_clip->minicpmv_version = gguf_get_val_i32(ctx, idx); + } + + // GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search + GGML_ASSERT(new_clip->has_vision_encoder); GGML_ASSERT(!new_clip->has_text_encoder); @@ -1031,6 +1197,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { LOG_TEE("%s: text_encoder: %d\n", __func__, new_clip->has_text_encoder); LOG_TEE("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder); LOG_TEE("%s: llava_projector: %d\n", __func__, new_clip->has_llava_projector); + LOG_TEE("%s: minicpmv_projector: %d\n", __func__, new_clip->has_minicpmv_projector); LOG_TEE("%s: model size: %.2f MB\n", __func__, model_size / 1024.0 / 1024.0); LOG_TEE("%s: metadata size: %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0); } @@ -1272,6 +1439,27 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { 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 if (new_clip->proj_type == PROJECTOR_TYPE_RESAMPLER) { + // vision_model.mm_model_pos_embed = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD); + vision_model.mm_model_pos_embed_k = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD_K); + vision_model.mm_model_query = get_tensor(new_clip->ctx_data, TN_MINICPMV_QUERY); + vision_model.mm_model_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_PROJ); + vision_model.mm_model_kv_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_KV_PROJ); + vision_model.mm_model_attn_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "weight")); + vision_model.mm_model_attn_k_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "weight")); + vision_model.mm_model_attn_v_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "weight")); + vision_model.mm_model_attn_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "bias")); + vision_model.mm_model_attn_k_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "bias")); + vision_model.mm_model_attn_v_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "bias")); + vision_model.mm_model_attn_o_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "weight")); + vision_model.mm_model_attn_o_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "bias")); + vision_model.mm_model_ln_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "weight")); + vision_model.mm_model_ln_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "bias")); + vision_model.mm_model_ln_kv_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "weight")); + vision_model.mm_model_ln_kv_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "bias")); + vision_model.mm_model_ln_post_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "weight")); + vision_model.mm_model_ln_post_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "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())); @@ -1310,7 +1498,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend)); clip_image_f32_batch batch; batch.size = 1; - ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch); + ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false); ggml_gallocr_reserve(new_clip->compute_alloc, gf); size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0); LOG_TEE("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0); @@ -1319,6 +1507,17 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { return new_clip; } +void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size) { + ctx_clip->load_image_size = load_image_size; +} + +struct clip_image_size * clip_image_size_init() { + struct clip_image_size * load_image_size = new struct clip_image_size(); + load_image_size->width = 448; + load_image_size->height = 448; + return load_image_size; +} + struct clip_image_u8 * clip_image_u8_init() { return new clip_image_u8(); } @@ -1424,7 +1623,7 @@ static void normalize_image_u8_to_f32(const clip_image_u8* src, clip_image_f32* } } -inline float clip(float x, float lower, float upper) { +inline int clip(int x, int lower, int upper) { return std::max(lower, std::min(x, upper)); } @@ -1589,9 +1788,182 @@ static std::vector divide_to_patches_u8(const clip_image_u8 & im return patches; } +static int ensure_divide(int length, int patch_size) { + return std::max(static_cast(std::round(static_cast(length) / patch_size) * patch_size), patch_size); +} + +static std::pair uhd_find_best_resize(std::pair original_size, int scale_resolution, int patch_size, bool allow_upscale = false) { + int width = original_size.first; + int height = original_size.second; + if ((width * height > scale_resolution * scale_resolution) || allow_upscale) { + float r = static_cast(width) / height; + height = static_cast(scale_resolution / std::sqrt(r)); + width = static_cast(height * r); + } + int best_width = ensure_divide(width, patch_size); + int best_height = ensure_divide(height, patch_size); + return std::make_pair(best_width, best_height); +} + +static std::pair uhd_get_refine_size(std::pair original_size, std::pair grid, int scale_resolution, int patch_size, bool allow_upscale = false) { + int width, height; + std::tie(width, height) = original_size; + int grid_x, grid_y; + std::tie(grid_x, grid_y) = grid; + + int refine_width = ensure_divide(width, grid_x); + int refine_height = ensure_divide(height, grid_y); + + int grid_width = refine_width / grid_x; + int grid_height = refine_height / grid_y; + + // auto best_grid_size = find_best_resize(std::make_tuple(grid_width, grid_height), scale_resolution, patch_size, allow_upscale); (old line) + auto best_grid_size = uhd_find_best_resize(std::make_pair(grid_width, grid_height), scale_resolution, patch_size, allow_upscale); // (new line) => fixes conversion for make_tuple to make_pair + int best_grid_width, best_grid_height; + std::tie(best_grid_width, best_grid_height) = best_grid_size; + + // std::pair refine_size = std::make_tuple(best_grid_width * grid_x, best_grid_height * grid_y); (old line) + std::pair refine_size = std::make_pair(best_grid_width * grid_x, best_grid_height * grid_y); // (new line) + return refine_size; +} + +static std::pair uhd_best_grid(const int max_slice_nums, const int multiple, const float log_ratio) { + std::vector candidate_split_grids_nums; + for (int i : {multiple - 1, multiple, multiple + 1}) { + if (i == 1 || i > max_slice_nums) { + continue; + } + candidate_split_grids_nums.push_back(i); + } + + std::vector> candidate_grids; + for (int split_grids_nums : candidate_split_grids_nums) { + int m = 1; + while (m <= split_grids_nums) { + if (split_grids_nums % m == 0) { + candidate_grids.emplace_back(m, split_grids_nums / m); + } + ++m; + } + } + + std::pair best_grid{1, 1}; + float min_error = std::numeric_limits::infinity(); + for (const auto& grid : candidate_grids) { + float error = std::abs(log_ratio - std::log(1.0 * grid.first / grid.second)); + if (error < min_error) { + best_grid = grid; + min_error = error; + } + } + return best_grid; +} + +// inspired from LLaVA-UHD: +// -> https://arxiv.org/pdf/2403.11703 +// -> https://github.com/thunlp/LLaVA-UHD +// -> https://github.com/thunlp/LLaVA-UHD/blob/302301bc2175f7e717fb8548516188e89f649753/llava_uhd/train/llava-uhd/slice_logic.py#L118 +static std::vector> uhd_slice_image(const clip_image_u8 * img, const int max_slice_nums=9, const int scale_resolution=448, const int patch_size=14) { + const std::pair original_size={img->nx,img->ny}; + const int original_width = img->nx; + const int original_height = img->ny; + const float log_ratio = log(1.0*original_width/original_height); + const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution); + const int multiple = fmin(ceil(ratio), max_slice_nums); + + std::vector> images; + LOG_TEE("%s: multiple %d\n", __func__, multiple); + images.push_back(std::vector()); + + if (multiple <= 1) { + auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size, true); + clip_image_u8 * source_image = clip_image_u8_init(); + bicubic_resize(*img, *source_image, best_size.first, best_size.second); + // source_image = image.resize(best_size, Image.Resampling.BICUBIC) + images[images.size()-1].push_back(source_image); + } + else if (multiple > 1) { + auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size); + clip_image_u8 * source_image = clip_image_u8_init(); + bicubic_resize(*img, *source_image, best_size.first, best_size.second); + // source_image = image.copy().resize(best_resize, Image.Resampling.BICUBIC) + LOG_TEE("%s: image_size: %d %d; source_image size: %d %d\n", __func__, img->nx, img->ny, best_size.first, best_size.second); + images[images.size()-1].push_back(source_image); + + std::pair best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio); + LOG_TEE("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second); + + auto refine_size = uhd_get_refine_size(original_size, best_grid, scale_resolution, patch_size, true); + clip_image_u8 * refine_image = clip_image_u8_init(); + bicubic_resize(*img, *refine_image, refine_size.first, refine_size.second); + + LOG_TEE("%s: refine_image_size: %d %d; refine_size: %d %d\n", __func__, refine_image->nx, refine_image->ny, refine_size.first, refine_size.second); + + // split_to_patches + int width = refine_image->nx; + int height = refine_image->ny; + int grid_x = int(width / best_grid.first); + int grid_y = int(height / best_grid.second); + for (int patches_i = 0, ic = 0; patches_i < height && ic < best_grid.second; patches_i += grid_y, ic += 1){ + images.push_back(std::vector()); + for(int patches_j = 0, jc = 0; patches_j < width && jc < best_grid.first; patches_j += grid_x, jc += 1){ + clip_image_u8 * patch = clip_image_u8_init(); + patch->nx = grid_x; + patch->ny = grid_y; + patch->buf.resize(3 * patch->nx * patch->ny); + for (int y = patches_i; y < patches_i + grid_y; ++y) { + for (int x = patches_j; x < patches_j + grid_x; ++x) { + const int i = 3 * (y * refine_image->nx + x); + const int j = 3 * ((y-patches_i) * patch->nx + (x-patches_j)); + patch->buf[j] = refine_image->buf[i]; + patch->buf[j+1] = refine_image->buf[i+1]; + patch->buf[j+2] = refine_image->buf[i+2]; + } + } + images[images.size()-1].push_back(patch); + } + } + } + return images; +} + +int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) { + const int max_slice_nums=9; + const int scale_resolution=448; + const int original_width = ctx_clip->load_image_size->width; + const int original_height = ctx_clip->load_image_size->height; + const float log_ratio = log(1.0*original_width/original_height); + const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution); + const int multiple = fmin(ceil(ratio), max_slice_nums); + std::pair best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio); + return best_grid.first; +} + // returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector // res_imgs memory is being allocated here, previous allocations will be freed if found bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) { + + if(clip_is_minicpmv(ctx)){ + int max_slice_nums = 9; + std::vector> imgs = uhd_slice_image(img, max_slice_nums); + res_imgs->size = 0; + for (size_t i = 0; i < imgs.size(); ++i){ + res_imgs->size += imgs[i].size(); + } + res_imgs->data = new clip_image_f32[res_imgs->size]; + int idx = 0; + for (size_t i = 0; i < imgs.size(); ++i) { + for (size_t j = 0; j < imgs[i].size(); ++j) { + LOG_TEE("%s: %d %d\n", __func__,imgs[i][j]->nx,imgs[i][j]->ny); + clip_image_f32 * res = clip_image_f32_init(); + normalize_image_u8_to_f32(imgs[i][j], res, ctx->image_mean, ctx->image_std); + res_imgs->data[idx++] = *res; + clip_image_f32_free(res); + } + } + return true; + } + bool pad_to_square = true; if (!ctx->has_vision_encoder) { LOG_TEE("This gguf file seems to have no vision encoder\n"); @@ -1807,11 +2179,104 @@ int clip_n_patches(const struct clip_ctx * ctx) { if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) { n_patches /= 4; + } else if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) { + if (ctx->minicpmv_version == 2) { + n_patches = 96; + } + else if (ctx->minicpmv_version == 3) { + n_patches = 64; + } } return n_patches; } +static std::vector>> get_1d_sincos_pos_embed_from_grid_new(int embed_dim, const std::vector> & pos) { + assert(embed_dim % 2 == 0); + int H = pos.size(); + int W = pos[0].size(); + + std::vector omega(embed_dim / 2); + for (int i = 0; i < embed_dim / 2; ++i) { + omega[i] = 1.0 / pow(10000.0, static_cast(i) / (embed_dim / 2)); + } + + std::vector>> emb(H, std::vector>(W, std::vector(embed_dim))); + for (int h = 0; h < H; ++h) { + for (int w = 0; w < W; ++w) { + for (int d = 0; d < embed_dim / 2; ++d) { + float out_value = pos[h][w] * omega[d]; + emb[h][w][d] = sin(out_value); + emb[h][w][d + embed_dim / 2] = cos(out_value); + } + } + } + + return emb; +} + +static std::vector>> get_2d_sincos_pos_embed_from_grid(int embed_dim, const std::vector>> & grid) { + assert(embed_dim % 2 == 0); + std::vector>> emb_h = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[0]); // (H, W, D/2) + std::vector>> emb_w = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[1]); // (H, W, D/2) + + int H = emb_h.size(); + int W = emb_h[0].size(); + std::vector>> emb(H, std::vector>(W, std::vector(embed_dim))); + + for (int h = 0; h < H; ++h) { + for (int w = 0; w < W; ++w) { + for (int d = 0; d < embed_dim / 2; ++d) { + emb[h][w][d] = emb_h[h][w][d]; + emb[h][w][d + embed_dim / 2] = emb_w[h][w][d]; + } + } + } + return emb; +} + +static std::vector> get_2d_sincos_pos_embed(int embed_dim, const std::pair image_size) { + int grid_h_size = image_size.first; + int grid_w_size = image_size.second; + + std::vector grid_h(grid_h_size); + std::vector grid_w(grid_w_size); + + for (int i = 0; i < grid_h_size; ++i) { + grid_h[i] = static_cast(i); + } + for (int i = 0; i < grid_w_size; ++i) { + grid_w[i] = static_cast(i); + } + + std::vector> grid(grid_h_size, std::vector(grid_w_size)); + for (int h = 0; h < grid_h_size; ++h) { + for (int w = 0; w < grid_w_size; ++w) { + grid[h][w] = grid_w[w]; + } + } + std::vector>> grid_2d = {grid, grid}; + for (int h = 0; h < grid_h_size; ++h) { + for (int w = 0; w < grid_w_size; ++w) { + grid_2d[0][h][w] = grid_h[h]; + grid_2d[1][h][w] = grid_w[w]; + } + } + + std::vector>> pos_embed_3d = get_2d_sincos_pos_embed_from_grid(embed_dim, grid_2d); + + int H = image_size.first; + int W = image_size.second; + std::vector> pos_embed_2d(H * W, std::vector(embed_dim)); + for (int h = 0; h < H; ++h) { + for (int w = 0; w < W; ++w) { + pos_embed_2d[w * H + h] = pos_embed_3d[h][w]; + } + } + + return pos_embed_2d; +} + bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) { if (!ctx->has_vision_encoder) { LOG_TEE("This gguf file seems to have no vision encoder\n"); @@ -1834,19 +2299,33 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima if (ctx->has_llava_projector) { GGML_ASSERT(batch_size == 1); // TODO: support multiple images } + if (ctx->has_minicpmv_projector) { + GGML_ASSERT(batch_size == 1); + } // build the inference graph - ggml_cgraph * gf = clip_image_build_graph(ctx, imgs); + ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true); ggml_gallocr_alloc_graph(ctx->compute_alloc, gf); // set inputs const auto & model = ctx->vision_model; const auto & hparams = model.hparams; - const int image_size = hparams.image_size; + const int image_size = hparams.image_size; + int image_size_width = image_size; + int image_size_height = image_size; + if (ctx->has_minicpmv_projector) { + image_size_width = imgs->data[0].nx; + image_size_height = imgs->data[0].ny; + } const int patch_size = hparams.patch_size; - const int num_patches = ((image_size / patch_size) * (image_size / patch_size)); + const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size)); const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0); + if(ctx->load_image_size==nullptr){ + ctx->load_image_size= clip_image_size_init(); + } + const int pos_w = ctx->load_image_size->width/patch_size; + const int pos_h = ctx->load_image_size->height/patch_size; { struct ggml_tensor * inp_raw = ggml_graph_get_tensor(gf, "inp_raw"); @@ -1855,7 +2334,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima for (size_t i = 0; i < imgs->size; i++) { const int nx = imgs->data[i].nx; const int ny = imgs->data[i].ny; - GGML_ASSERT(nx == image_size && ny == image_size); + if (!ctx->has_minicpmv_projector) { + GGML_ASSERT(nx == image_size && ny == image_size); + } const int n = nx * ny; @@ -1872,37 +2353,87 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima ggml_backend_tensor_set(inp_raw, data, 0, ggml_nbytes(inp_raw)); free(data); } + if (ctx->has_minicpmv_projector) { + { + // inspired from siglip: + // -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit + // -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316 + struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); + int* positions_data = (int*)malloc(ggml_nbytes(positions)); + int bucket_coords_h[70]; + int bucket_coords_w[70]; + for (int i = 0; i < pos_h; i++){ + bucket_coords_h[i] = std::floor(70.0*i/pos_h); + } + for (int i = 0; i < pos_w; i++){ + bucket_coords_w[i] = std::floor(70.0*i/pos_w); + } + for (int i = 0, id = 0; i < pos_h; i++){ + for (int j = 0; j < pos_w; j++){ + positions_data[id++] = bucket_coords_h[i]*70 + bucket_coords_w[j]; + } + } + ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); + free(positions_data); + } - { - if (ctx->has_class_embedding) { - struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings"); + { + // inspired from resampler of Qwen-VL: + // -> https://huggingface.co/Qwen/Qwen-VL/tree/main + // -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23 + struct ggml_tensor * pos_embed = ggml_graph_get_tensor(gf, "pos_embed"); + int embed_dim = 4096; + if (ctx->minicpmv_version == 2) { + embed_dim = 4096; + } + else if (ctx->minicpmv_version == 3) { + embed_dim = 3584; + } + auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h)); - void* zero_mem = malloc(ggml_nbytes(embeddings)); - memset(zero_mem, 0, ggml_nbytes(embeddings)); - ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings)); - free(zero_mem); + float * pos_embed_data = (float *)malloc(ggml_nbytes(pos_embed)); + for(int i=0;ihas_class_embedding) { + struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings"); - { - struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); - - int* positions_data = (int*)malloc(ggml_nbytes(positions)); - for (int i = 0; i < num_positions; i++) { - positions_data[i] = i; + void* zero_mem = malloc(ggml_nbytes(embeddings)); + memset(zero_mem, 0, ggml_nbytes(embeddings)); + ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings)); + free(zero_mem); + } } - ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); - free(positions_data); - } - { - struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches"); - int* patches_data = (int*)malloc(ggml_nbytes(patches)); - for (int i = 0; i < num_patches; i++) { - patches_data[i] = i + 1; + { + struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); + + int* positions_data = (int*)malloc(ggml_nbytes(positions)); + for (int i = 0; i < num_positions; i++) { + positions_data[i] = i; + } + ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); + free(positions_data); + } + + { + struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches"); + int* patches_data = (int*)malloc(ggml_nbytes(patches)); + for (int i = 0; i < num_patches; i++) { + patches_data[i] = i + 1; + } + ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches)); + free(patches_data); } - ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches)); - free(patches_data); } if (ggml_backend_is_cpu(ctx->backend)) { @@ -2072,7 +2603,22 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) { if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { return ctx->vision_model.mm_3_b->ne[0]; } + if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) { + if (ctx->minicpmv_version == 2) { + return 4096; + } + else if (ctx->minicpmv_version == 3) { + return 3584; + } + } std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type]; throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); } + +int clip_is_minicpmv(const struct clip_ctx * ctx) { + if (ctx->has_minicpmv_projector) { + return ctx->minicpmv_version; + } + return 0; +} diff --git a/examples/llava/clip.h b/examples/llava/clip.h index ca3631384..78588bdf1 100644 --- a/examples/llava/clip.h +++ b/examples/llava/clip.h @@ -18,14 +18,17 @@ # define CLIP_API #endif -struct clip_ctx; - #ifdef __cplusplus extern "C" { #endif struct clip_ctx; +struct clip_image_size { + int width; + int height; +}; + struct clip_image_u8_batch { struct clip_image_u8 * data; size_t size; @@ -55,6 +58,10 @@ CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx); CLIP_API int clip_n_patches (const struct clip_ctx * ctx); CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx); +CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip); +CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size); + +CLIP_API struct clip_image_size * clip_image_size_init(); CLIP_API struct clip_image_u8 * clip_image_u8_init (); CLIP_API struct clip_image_f32 * clip_image_f32_init(); @@ -78,6 +85,8 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype); +CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx); + #ifdef __cplusplus } #endif diff --git a/examples/llava/convert-image-encoder-to-gguf.py b/examples/llava/convert_image_encoder_to_gguf.py similarity index 97% rename from examples/llava/convert-image-encoder-to-gguf.py rename to examples/llava/convert_image_encoder_to_gguf.py index b00bf7c6d..36f6b92fb 100644 --- a/examples/llava/convert-image-encoder-to-gguf.py +++ b/examples/llava/convert_image_encoder_to_gguf.py @@ -185,6 +185,8 @@ else: fout.add_description("two-tower CLIP model") if has_text_encoder: + assert t_hparams is not None + assert tokens is not None # text_model hparams fout.add_uint32(k(KEY_CONTEXT_LENGTH, TEXT), t_hparams["max_position_embeddings"]) fout.add_uint32(k(KEY_EMBEDDING_LENGTH, TEXT), t_hparams["hidden_size"]) @@ -259,8 +261,8 @@ if has_vision_encoder: if processor is not None: - image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean - image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std + image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean # pyright: ignore[reportAttributeAccessIssue] + image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std # pyright: ignore[reportAttributeAccessIssue] else: image_mean = args.image_mean if args.image_mean is not None else default_image_mean image_std = args.image_std if args.image_std is not None else default_image_std @@ -272,7 +274,7 @@ fout.add_bool("clip.use_gelu", use_gelu) if has_llava_projector: - model.vision_model.encoder.layers.pop(-1) + model.vision_model.encoder.layers.pop(-1) # pyright: ignore[reportAttributeAccessIssue] projector = torch.load(args.llava_projector) for name, data in projector.items(): name = get_tensor_name(name) @@ -286,7 +288,7 @@ if has_llava_projector: print("Projector tensors added\n") -state_dict = model.state_dict() +state_dict = model.state_dict() # pyright: ignore[reportAttributeAccessIssue] for name, data in state_dict.items(): if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_llava_projector): # we don't need this diff --git a/examples/llava/llava-cli.cpp b/examples/llava/llava-cli.cpp index 8c7dd2ae3..86b39f20e 100644 --- a/examples/llava/llava-cli.cpp +++ b/examples/llava/llava-cli.cpp @@ -129,14 +129,14 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para if (!params->image.empty()) { LOG_TEE("using base64 encoded image instead of command line image path\n"); } - embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->n_threads, prompt); + embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt); if (!embed) { LOG_TEE("%s: can't load image from prompt\n", __func__); return NULL; } params->prompt = remove_image_from_prompt(prompt); } else { - embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->n_threads, fname.c_str()); + embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->cpuparams.n_threads, fname.c_str()); if (!embed) { fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str()); return NULL; diff --git a/examples/llava/llava.cpp b/examples/llava/llava.cpp index 63878d176..851af0f00 100644 --- a/examples/llava/llava.cpp +++ b/examples/llava/llava.cpp @@ -202,6 +202,33 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector return true; } +static clip_image_f32 * only_v2_5_reshape_by_patch(clip_image_f32 * image, int patch_size) { + int width = image->nx; + int height = image->ny; + int num_patches = (height / patch_size) * (width / patch_size); + clip_image_f32 * patch = clip_image_f32_init(); + patch->nx = patch_size * num_patches; + patch->ny = patch_size; + patch->buf.resize(3 * patch->nx * patch->ny); + + int patch_index = 0; + + for (int i = 0; i < height; i += patch_size) { + for (int j = 0; j < width; j += patch_size) { + for (int pi = 0; pi < patch_size; ++pi) { + for (int pj = 0; pj < patch_size; ++pj) { + int input_index = ((i + pi) * width + (j + pj)) * 3; + int output_index = (pi * patch_size * num_patches + patch_index * patch_size + pj) * 3; + patch->buf[output_index] = image->buf[input_index]; + patch->buf[output_index+1] = image->buf[input_index+1]; + patch->buf[output_index+2] = image->buf[input_index+2]; + } + } + patch_index++; + } + } + return patch; +} static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) { // std::vector img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336 @@ -218,7 +245,51 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip); - if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) { + if (clip_is_minicpmv(ctx_clip)) { + std::vector image_embd_v; + image_embd_v.resize(img_res_v.size); + struct clip_image_size * load_image_size = clip_image_size_init(); + for (size_t i = 0; i < img_res_v.size; i++) { + const int64_t t_img_enc_step_start_us = ggml_time_us(); + image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); + int patch_size=14; + load_image_size->width = img_res_v.data[i].nx; + load_image_size->height = img_res_v.data[i].ny; + clip_add_load_image_size(ctx_clip, load_image_size); + bool encoded = false; + int has_minicpmv_projector = clip_is_minicpmv(ctx_clip); + if (has_minicpmv_projector == 2) { + encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]); + } + else if (has_minicpmv_projector == 3) { + encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); + } + if (!encoded) { + LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size); + return false; + } + const int64_t t_img_enc_steop_batch_us = ggml_time_us(); + LOG_TEE("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0); + } + const int64_t t_img_enc_batch_us = ggml_time_us(); + LOG_TEE("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0); + + int n_img_pos_out = 0; + for (size_t i = 0; i < image_embd_v.size(); i++) { + std::memcpy(image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], clip_embd_nbytes(ctx_clip)); + n_img_pos_out += clip_n_patches(ctx_clip); + } + *n_img_pos = n_img_pos_out; + for (size_t i = 0; i < image_embd_v.size(); i++) { + free(image_embd_v[i]); + } + image_embd_v.clear(); + load_image_size->width = img->nx; + load_image_size->height = img->ny; + clip_add_load_image_size(ctx_clip, load_image_size); + LOG_TEE("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height); + } + else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) { // flat / default llava-1.5 type embedding *n_img_pos = clip_n_patches(ctx_clip); bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096 @@ -228,7 +299,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli return false; } - } else { + } + else { // spatial_unpad llava-1.6 type embedding // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working std::vector image_embd_v; @@ -297,7 +369,11 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * } bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) { - float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model + int num_max_patches = 6; + if (clip_is_minicpmv(ctx_clip)) { + num_max_patches = 10; + } + float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model if (!image_embd) { LOG_TEE("Unable to allocate memory for image embeddings\n"); return false; diff --git a/examples/llava/llava.h b/examples/llava/llava.h index 19212f6e9..b6feb3027 100644 --- a/examples/llava/llava.h +++ b/examples/llava/llava.h @@ -17,12 +17,11 @@ # define LLAVA_API #endif -struct clip_ctx; - #ifdef __cplusplus extern "C" { #endif +struct clip_ctx; struct llava_image_embed { float * embed; int n_image_pos; @@ -37,8 +36,8 @@ LLAVA_API bool llava_image_embed_make_with_clip_img(struct clip_ctx * ctx_clip, LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length); /** build an image embed from a path to an image filename */ LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path); -LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed); /** free an embedding made with llava_image_embed_make_* */ +LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed); /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */ LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past); diff --git a/examples/llava/llava-surgery.py b/examples/llava/llava_surgery.py similarity index 100% rename from examples/llava/llava-surgery.py rename to examples/llava/llava_surgery.py diff --git a/examples/llava/llava-surgery-v2.py b/examples/llava/llava_surgery_v2.py similarity index 94% rename from examples/llava/llava-surgery-v2.py rename to examples/llava/llava_surgery_v2.py index eb56d6988..2d5b32fe6 100644 --- a/examples/llava/llava-surgery-v2.py +++ b/examples/llava/llava_surgery_v2.py @@ -2,7 +2,9 @@ import argparse import glob import os import torch -from safetensors.torch import load as safe_load, save as safe_save, safe_open, save_file +from safetensors import safe_open +from safetensors.torch import save_file +from typing import Any, ContextManager, cast # Function to determine if file is a SafeTensor file def is_safetensor_file(file_path): @@ -13,7 +15,7 @@ def is_safetensor_file(file_path): def load_model(file_path): if is_safetensor_file(file_path): tensors = {} - with safe_open(file_path, framework="pt", device="cpu") as f: + with cast(ContextManager[Any], safe_open(file_path, framework="pt", device="cpu")) as f: for key in f.keys(): tensors[key] = f.get_tensor(key).clone() # output shape @@ -134,7 +136,7 @@ if len(mm_tensors) == 0: if last_checkpoint is not None: for k, v in last_checkpoint.items(): print(k) - print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint)} tensors.") + print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint) if last_checkpoint is not None else 0} tensors.") print("No tensors found. Is this a LLaVA model?") exit() @@ -143,8 +145,10 @@ print(f"Found additional {len(first_mm_tensors)} tensors to extract.") # projector = {name: checkpoint.[name].float() for name in mm_tensors} projector = {} for name in mm_tensors: + assert last_checkpoint is not None projector[name] = last_checkpoint[name].float() for name in first_mm_tensors: + assert first_checkpoint is not None projector[name] = first_checkpoint[name].float() if len(projector) > 0: diff --git a/examples/llava/minicpmv-cli.cpp b/examples/llava/minicpmv-cli.cpp new file mode 100644 index 000000000..f500ea5b9 --- /dev/null +++ b/examples/llava/minicpmv-cli.cpp @@ -0,0 +1,329 @@ +#include "ggml.h" +#include "log.h" +#include "common.h" +#include "clip.h" +#include "llava.h" +#include "llama.h" + +#include +#include +#include + +struct llava_context { + struct clip_ctx * ctx_clip = NULL; + struct llama_context * ctx_llama = NULL; + struct llama_model * model = NULL; +}; + +static void show_additional_info(int /*argc*/, char ** argv) { + LOG_TEE("\n example usage: %s -m --mmproj --image --image [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]); + LOG_TEE(" note: a lower temperature value like 0.1 is recommended for better quality.\n"); +} + +static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) { + (void) level; + (void) user_data; + LOG_TEE("%s", text); +} + +static struct llama_model * llava_init(gpt_params * params) { + llama_backend_init(); + llama_numa_init(params->numa); + + llama_model_params model_params = llama_model_params_from_gpt_params(*params); + + llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params); + if (model == NULL) { + LOG_TEE("%s: error: unable to load model\n" , __func__); + return NULL; + } + return model; +} + +static struct llava_context * llava_init_context(gpt_params * params, llama_model * model) { + auto prompt = params->prompt; + if (prompt.empty()) { + prompt = "describe the image in detail."; + } + + llama_context_params ctx_params = llama_context_params_from_gpt_params(*params); + if (params->n_ctx < 2048) { + // warn user here, "Image processing requires at least 2048 context, setting context to 2048" + LOG_TEE("%s: warn: Image processing requires at least 2048 context, setting context to 2048\n" , __func__); + ctx_params.n_ctx = 2048; + } else { + ctx_params.n_ctx = params->n_ctx; + } + + llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params); + + if (ctx_llama == NULL) { + LOG_TEE("%s: error: failed to create the llama_context\n" , __func__); + return NULL; + } + + auto ctx_llava = (struct llava_context *)malloc(sizeof(llava_context)); + + ctx_llava->ctx_llama = ctx_llama; + ctx_llava->model = model; + return ctx_llava; +} + +static void llava_free(struct llava_context * ctx_llava) { + if (ctx_llava->ctx_clip) { + clip_free(ctx_llava->ctx_clip); + ctx_llava->ctx_clip = NULL; + } + + llama_free(ctx_llava->ctx_llama); + llama_free_model(ctx_llava->model); + llama_backend_free(); +} + +static struct clip_ctx * clip_init_context(gpt_params * params) { + const char * clip_path = params->mmproj.c_str(); + + auto prompt = params->prompt; + if (prompt.empty()) { + prompt = "describe the image in detail."; + } + auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1); + return ctx_clip; +} + +static bool eval_tokens(struct llama_context * ctx_llama, std::vector tokens, int n_batch, int * n_past) { + int N = (int) tokens.size(); + for (int i = 0; i < N; i += n_batch) { + int n_eval = (int) tokens.size() - i; + if (n_eval > n_batch) { + n_eval = n_batch; + } + if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) { + LOG_TEE("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past); + return false; + } + *n_past += n_eval; + } + return true; +} + +static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) { + std::vector tokens; + tokens.push_back(id); + return eval_tokens(ctx_llama, tokens, 1, n_past); +} + +static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){ + std::string str2 = str; + std::vector embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos, true); + return eval_tokens(ctx_llama, embd_inp, n_batch, n_past); +} + +static void process_eval_image_embed(struct llava_context * ctx_llava, const struct llava_image_embed * embeds, int n_batch, int * n_past, int idx) { + float * image_embed = (float *)malloc(clip_embd_nbytes(ctx_llava->ctx_clip)); + std::memcpy(image_embed, embeds->embed + idx * clip_n_patches(ctx_llava->ctx_clip) * clip_n_mmproj_embd(ctx_llava->ctx_clip), clip_embd_nbytes(ctx_llava->ctx_clip)); + + auto slice_embed = (llava_image_embed*)malloc(sizeof(llava_image_embed)); + slice_embed->embed = image_embed; + slice_embed->n_image_pos = clip_n_patches(ctx_llava->ctx_clip); + llava_eval_image_embed(ctx_llava->ctx_llama, slice_embed, n_batch, n_past); + llava_image_embed_free(slice_embed); +} + +static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, gpt_params * params, int &n_past) { + std::string system_prompt; + int idx = 0; + int num_image_embeds = embeds->n_image_pos / clip_n_patches(ctx_llava->ctx_clip); + int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip); + if (has_minicpmv_projector == 2) { + system_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n"; + } + else if (has_minicpmv_projector == 3) { + system_prompt = "<|im_start|>user\n"; + } + LOG_TEE("%s: image token past: %d\n", __func__, n_past); + eval_string(ctx_llava->ctx_llama, (system_prompt+"").c_str(), params->n_batch, &n_past, false); + process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++); + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + if (num_image_embeds > 1) { + size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip); + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) { + for (size_t j = 0; j < num_image_embeds_col; ++j) { + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++); + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + if (j == num_image_embeds_col - 1) { + eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false); + } + } + } + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + } + LOG_TEE("%s: image token past: %d\n", __func__, n_past); +} + +static const char * sample(struct llama_sampling_context * ctx_sampling, + struct llama_context * ctx_llama, + int * n_past) { + const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL); + llama_sampling_accept(ctx_sampling, ctx_llama, id, true); + static std::string ret; + if (llama_token_is_eog(llama_get_model(ctx_llama), id)) { + ret = ""; + } else { + ret = llama_token_to_piece(ctx_llama, id); + } + eval_id(ctx_llama, id, n_past); + return ret.c_str(); +} + +static struct llava_context * minicpmv_init(gpt_params * params, const std::string & fname, int &n_past){ + auto ctx_clip = clip_init_context(params); + auto embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str()); + if (!embeds) { + std::cerr << "error: failed to load image " << fname << ". Terminating\n\n"; + return NULL; + } + + // process the prompt + if (params->prompt.empty() && params->interactive == false) { + LOG_TEE("prompt should be given or interactive mode should be on"); + return NULL; + } + + auto model = llava_init(params); + if (model == NULL) { + fprintf(stderr, "%s: error: failed to init minicpmv model\n", __func__); + return NULL; + } + const int64_t t_llava_init_start_us = ggml_time_us(); + auto ctx_llava = llava_init_context(params, model); + ctx_llava->ctx_clip = ctx_clip; + const int64_t t_llava_init_end_us = ggml_time_us(); + float t_llava_init_ms = (t_llava_init_end_us - t_llava_init_start_us) / 1000.0; + LOG_TEE("\n%s: llava init in %8.2f ms.\n", __func__, t_llava_init_ms); + + const int64_t t_process_image_start_us = ggml_time_us(); + process_image(ctx_llava, embeds, params, n_past); + const int64_t t_process_image_end_us = ggml_time_us(); + float t_process_image_ms = (t_process_image_end_us - t_process_image_start_us) / 1000.0; + LOG_TEE("\n%s: llama process image in %8.2f ms.\n", __func__, t_process_image_ms); + + llava_image_embed_free(embeds); + return ctx_llava; +} + +static struct llama_sampling_context * llama_init(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false){ + std::string user_prompt = prompt; + int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip); + if (!is_first) { + if (has_minicpmv_projector == 2) { + user_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" + prompt; + } + else if (has_minicpmv_projector == 3) { + user_prompt = "<|im_start|>user\n" + prompt; + } + } + + eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false); + if (has_minicpmv_projector == 2) { + eval_string(ctx_llava->ctx_llama, "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", params->n_batch, &n_past, false); + } + else if (has_minicpmv_projector == 3) { + eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false); + } + + // generate the response + + LOG_TEE("\n"); + + struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams); + return ctx_sampling; +} + +static const char * llama_loop(struct llava_context * ctx_llava,struct llama_sampling_context * ctx_sampling, int &n_past){ + + const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past); + return tmp; +} + +int main(int argc, char ** argv) { + ggml_time_init(); + + gpt_params params; + + if (!gpt_params_parse(argc, argv, params)) { + show_additional_info(argc, argv); + return 1; + } + +#ifndef LOG_DISABLE_LOGS + log_set_target(log_filename_generator("llava", "log")); + LOG_TEE("Log start\n"); + log_dump_cmdline(argc, argv); + llama_log_set(llama_log_callback_logTee, nullptr); +#endif // LOG_DISABLE_LOGS + + if (params.mmproj.empty() || (params.image.empty())) { + gpt_params_print_usage(argc, argv, params); + show_additional_info(argc, argv); + return 1; + } + + for (auto & image : params.image) { + int n_past = 0; + auto ctx_llava = minicpmv_init(¶ms, image, n_past); + + if (!params.prompt.empty()) { + LOG_TEE("%s\n", params.prompt.c_str()); + LOG_TEE(""); + auto ctx_sampling = llama_init(ctx_llava, ¶ms, params.prompt.c_str(), n_past, true); + const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict; + std::string response = ""; + bool have_tmp = false; + for (int i = 0; i < max_tgt_len; i++) { + auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past); + response += tmp; + if (strcmp(tmp, "") == 0){ + if(!have_tmp)continue; + else break; + } + if (strstr(tmp, "###")) break; // Yi-VL behavior + have_tmp = true; + printf("%s", tmp); + if (strstr(response.c_str(), "")) break; // minicpm-v + + fflush(stdout); + } + llama_sampling_free(ctx_sampling); + }else { + while (true) { + LOG_TEE(""); + std::string prompt; + std::getline(std::cin, prompt); + LOG_TEE(""); + auto ctx_sampling = llama_init(ctx_llava, ¶ms, prompt, n_past, true); + const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict; + std::string response = ""; + for (int i = 0; i < max_tgt_len; i++) { + auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past); + response += tmp; + if (strcmp(tmp, "") == 0) break; + if (strstr(tmp, "###")) break; // Yi-VL behavior + printf("%s", tmp);// mistral llava-1.6 + if (strstr(response.c_str(), "")) break; // minicpm-v + fflush(stdout); + } + llama_sampling_free(ctx_sampling); + } + } + printf("\n"); + llama_print_timings(ctx_llava->ctx_llama); + + ctx_llava->model = NULL; + llava_free(ctx_llava); + } + + return 0; +} diff --git a/examples/llava/minicpmv-convert-image-encoder-to-gguf.py b/examples/llava/minicpmv-convert-image-encoder-to-gguf.py new file mode 100644 index 000000000..ea773742a --- /dev/null +++ b/examples/llava/minicpmv-convert-image-encoder-to-gguf.py @@ -0,0 +1,806 @@ +# coding=utf-8 +# Copyright 2024 Google AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" PyTorch Siglip model. """ +# Copied from HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit and add tgt_sizes + + +import os +import math +import warnings + +import numpy as np +import torch +import torch.nn.functional as F +import torch.utils.checkpoint +from torch import nn +from torch.nn.init import _calculate_fan_in_and_fan_out + +from transformers.activations import ACT2FN +from transformers.modeling_utils import PreTrainedModel +from transformers.configuration_utils import PretrainedConfig +from transformers.utils import ( + logging, +) +from transformers.utils import logging + +logger = logging.get_logger(__name__) + +class SiglipVisionConfig(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`SiglipVisionModel`]. It is used to instantiate a + Siglip vision encoder according to the specified arguments, defining the model architecture. Instantiating a + configuration with the defaults will yield a similar configuration to that of the vision encoder of the Siglip + [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture. + Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the + documentation from [`PretrainedConfig`] for more information. + Args: + hidden_size (`int`, *optional*, defaults to 768): + Dimensionality of the encoder layers and the pooler layer. + intermediate_size (`int`, *optional*, defaults to 3072): + Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder. + num_hidden_layers (`int`, *optional*, defaults to 12): + Number of hidden layers in the Transformer encoder. + num_attention_heads (`int`, *optional*, defaults to 12): + Number of attention heads for each attention layer in the Transformer encoder. + num_channels (`int`, *optional*, defaults to 3): + Number of channels in the input images. + image_size (`int`, *optional*, defaults to 224): + The size (resolution) of each image. + patch_size (`int`, *optional*, defaults to 16): + The size (resolution) of each patch. + hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`): + The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`, + `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported. + layer_norm_eps (`float`, *optional*, defaults to 1e-06): + The epsilon used by the layer normalization layers. + attention_dropout (`float`, *optional*, defaults to 0.0): + The dropout ratio for the attention probabilities. + Example: + ```python + >>> from transformers import SiglipVisionConfig, SiglipVisionModel + >>> # Initializing a SiglipVisionConfig with google/siglip-base-patch16-224 style configuration + >>> configuration = SiglipVisionConfig() + >>> # Initializing a SiglipVisionModel (with random weights) from the google/siglip-base-patch16-224 style configuration + >>> model = SiglipVisionModel(configuration) + >>> # Accessing the model configuration + >>> configuration = model.config + ```""" + + model_type = "siglip_vision_model" + + def __init__( + self, + hidden_size=768, + intermediate_size=3072, + num_hidden_layers=12, + num_attention_heads=12, + num_channels=3, + image_size=224, + patch_size=16, + hidden_act="gelu_pytorch_tanh", + layer_norm_eps=1e-6, + attention_dropout=0.0, + **kwargs, + ): + super().__init__(**kwargs) + + self.hidden_size = hidden_size + self.intermediate_size = intermediate_size + self.num_hidden_layers = num_hidden_layers + self.num_attention_heads = num_attention_heads + self.num_channels = num_channels + self.patch_size = patch_size + self.image_size = image_size + self.attention_dropout = attention_dropout + self.layer_norm_eps = layer_norm_eps + self.hidden_act = hidden_act + +_CHECKPOINT_FOR_DOC = "google/siglip-base-patch16-224" + +SIGLIP_PRETRAINED_MODEL_ARCHIVE_LIST = [ + "google/siglip-base-patch16-224", + # See all SigLIP models at https://huggingface.co/models?filter=siglip +] + +# Copied from transformers.models.llama.modeling_llama._get_unpad_data +def _get_unpad_data(attention_mask): + seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32) + indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten() + max_seqlen_in_batch = seqlens_in_batch.max().item() + cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0)) + return ( + indices, + cu_seqlens, + max_seqlen_in_batch, + ) + + +def _trunc_normal_(tensor, mean, std, a, b): + # Cut & paste from PyTorch official master until it's in a few official releases - RW + # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf + def norm_cdf(x): + # Computes standard normal cumulative distribution function + return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0 + + if (mean < a - 2 * std) or (mean > b + 2 * std): + warnings.warn( + "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. " + "The distribution of values may be incorrect.", + stacklevel=2, + ) + + # Values are generated by using a truncated uniform distribution and + # then using the inverse CDF for the normal distribution. + # Get upper and lower cdf values + l = norm_cdf((a - mean) / std) + u = norm_cdf((b - mean) / std) + + # Uniformly fill tensor with values from [l, u], then translate to + # [2l-1, 2u-1]. + tensor.uniform_(2 * l - 1, 2 * u - 1) + + # Use inverse cdf transform for normal distribution to get truncated + # standard normal + if tensor.dtype in [torch.float16, torch.bfloat16]: + # The `erfinv_` op is not (yet?) defined in float16+cpu, bfloat16+gpu + og_dtype = tensor.dtype + tensor = tensor.to(torch.float32) + tensor.erfinv_() + tensor = tensor.to(og_dtype) + else: + tensor.erfinv_() + + # Transform to proper mean, std + tensor.mul_(std * math.sqrt(2.0)) + tensor.add_(mean) + + # Clamp to ensure it's in the proper range + if tensor.dtype == torch.float16: + # The `clamp_` op is not (yet?) defined in float16+cpu + tensor = tensor.to(torch.float32) + tensor.clamp_(min=a, max=b) + tensor = tensor.to(torch.float16) + else: + tensor.clamp_(min=a, max=b) + + +def trunc_normal_tf_( + tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0 +): + """Fills the input Tensor with values drawn from a truncated + normal distribution. The values are effectively drawn from the + normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)` + with values outside :math:`[a, b]` redrawn until they are within + the bounds. The method used for generating the random values works + best when :math:`a \\leq \text{mean} \\leq b`. + NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the + bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0 + and the result is subsquently scaled and shifted by the mean and std args. + Args: + tensor: an n-dimensional `torch.Tensor` + mean: the mean of the normal distribution + std: the standard deviation of the normal distribution + a: the minimum cutoff value + b: the maximum cutoff value + """ + with torch.no_grad(): + _trunc_normal_(tensor, 0, 1.0, a, b) + tensor.mul_(std).add_(mean) + + +def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"): + fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor) + denom = fan_in + if mode == "fan_in": + denom = fan_in + elif mode == "fan_out": + denom = fan_out + elif mode == "fan_avg": + denom = (fan_in + fan_out) / 2 + + variance = scale / denom + + if distribution == "truncated_normal": + # constant is stddev of standard normal truncated to (-2, 2) + trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978) + elif distribution == "normal": + with torch.no_grad(): + tensor.normal_(std=math.sqrt(variance)) + elif distribution == "uniform": + bound = math.sqrt(3 * variance) + with torch.no_grad(): + tensor.uniform_(-bound, bound) + else: + raise ValueError(f"invalid distribution {distribution}") + + +def lecun_normal_(tensor): + variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal") + + +def default_flax_embed_init(tensor): + variance_scaling_(tensor, mode="fan_in", distribution="normal") + +class SiglipVisionEmbeddings(nn.Module): + def __init__(self, config: SiglipVisionConfig): + super().__init__() + self.config = config + self.embed_dim = config.hidden_size + self.image_size = config.image_size + self.patch_size = config.patch_size + + self.patch_embedding = nn.Conv2d( + in_channels=config.num_channels, + out_channels=self.embed_dim, + kernel_size=self.patch_size, + stride=self.patch_size, + padding="valid", + ) + + self.num_patches_per_side = self.image_size // self.patch_size + self.num_patches = self.num_patches_per_side**2 + self.num_positions = self.num_patches + self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim) + +class SiglipAttention(nn.Module): + """Multi-headed attention from 'Attention Is All You Need' paper""" + + # Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__ + def __init__(self, config): + super().__init__() + self.config = config + self.embed_dim = config.hidden_size + self.num_heads = config.num_attention_heads + self.head_dim = self.embed_dim // self.num_heads + if self.head_dim * self.num_heads != self.embed_dim: + raise ValueError( + f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:" + f" {self.num_heads})." + ) + self.scale = self.head_dim**-0.5 + self.dropout = config.attention_dropout + + self.k_proj = nn.Linear(self.embed_dim, self.embed_dim) + self.v_proj = nn.Linear(self.embed_dim, self.embed_dim) + self.q_proj = nn.Linear(self.embed_dim, self.embed_dim) + self.out_proj = nn.Linear(self.embed_dim, self.embed_dim) + +# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Siglip +class SiglipMLP(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + self.activation_fn = ACT2FN[config.hidden_act] + self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size) + self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size) + + +# Copied from transformers.models.clip.modeling_clip.CLIPEncoderLayer with CLIP->Siglip +class SiglipEncoderLayer(nn.Module): + def __init__(self, config: SiglipVisionConfig): + super().__init__() + self.embed_dim = config.hidden_size + self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2" + self.self_attn = ( + SiglipAttention(config) + ) + self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps) + self.mlp = SiglipMLP(config) + self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps) + +class SiglipPreTrainedModel(PreTrainedModel): + """ + An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained + models. + """ + + config_class = SiglipVisionConfig + base_model_prefix = "siglip" + supports_gradient_checkpointing = True + + def _init_weights(self, module): + """Initialize the weights""" + + if isinstance(module, SiglipVisionEmbeddings): + width = self.config.hidden_size + nn.init.normal_(module.position_embedding.weight, std=1 / np.sqrt(width)) + elif isinstance(module, nn.Embedding): + default_flax_embed_init(module.weight) + elif isinstance(module, SiglipAttention): + nn.init.normal_(module.q_proj.weight) + nn.init.normal_(module.k_proj.weight) + nn.init.normal_(module.v_proj.weight) + nn.init.normal_(module.out_proj.weight) + nn.init.zeros_(module.q_proj.bias) + nn.init.zeros_(module.k_proj.bias) + nn.init.zeros_(module.v_proj.bias) + nn.init.zeros_(module.out_proj.bias) + elif isinstance(module, SiglipMLP): + nn.init.normal_(module.fc1.weight) + nn.init.normal_(module.fc2.weight) + nn.init.normal_(module.fc1.bias, std=1e-6) + nn.init.normal_(module.fc2.bias, std=1e-6) + elif isinstance(module, (nn.Linear, nn.Conv2d)): + lecun_normal_(module.weight) + if module.bias is not None: + nn.init.zeros_(module.bias) + elif isinstance(module, nn.LayerNorm): + module.bias.data.zero_() + module.weight.data.fill_(1.0) + + +SIGLIP_START_DOCSTRING = r""" + This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the + library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads + etc.) + This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. + Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage + and behavior. + Parameters: + config ([`SiglipVisionConfig`]): Model configuration class with all the parameters of the model. + Initializing with a config file does not load the weights associated with the model, only the + configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. +""" + + +SIGLIP_VISION_INPUTS_DOCSTRING = r""" + Args: + pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): + Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using + [`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details. + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned + tensors for more detail. + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for + more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. +""" + + +# Copied from transformers.models.clip.modeling_clip.CLIPEncoder with CLIP->Siglip +class SiglipEncoder(nn.Module): + """ + Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a + [`SiglipEncoderLayer`]. + Args: + config: SiglipConfig + """ + + def __init__(self, config: SiglipVisionConfig): + super().__init__() + self.config = config + self.layers = nn.ModuleList([SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)]) + self.gradient_checkpointing = False + +class SiglipVisionTransformer(SiglipPreTrainedModel): + config_class = SiglipVisionConfig + main_input_name = "pixel_values" + _supports_flash_attn_2 = True + + def __init__(self, config: SiglipVisionConfig): + super().__init__(config) + self.config = config + embed_dim = config.hidden_size + + self.embeddings = SiglipVisionEmbeddings(config) + self.encoder = SiglipEncoder(config) + self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps) + self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2" + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self) -> nn.Module: + return self.embeddings.patch_embedding + +import argparse +import json +import re + +import numpy as np +from gguf import * +from transformers.models.idefics2.modeling_idefics2 import Idefics2VisionTransformer, Idefics2VisionConfig + +TEXT = "clip.text" +VISION = "clip.vision" + + +def add_key_str(raw_key: str, arch: str) -> str: + return raw_key.format(arch=arch) + + +def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_minicpmv: bool) -> bool: + if name in ( + "logit_scale", + "text_model.embeddings.position_ids", + "vision_model.embeddings.position_ids", + ): + return True + + if has_minicpmv and name in ["visual_projection.weight"]: + return True + + if name.startswith("v") and not has_vision: + return True + + if name.startswith("t") and not has_text: + return True + + return False + + +def get_tensor_name(name: str) -> str: + if "projection" in name: + return name + if "mm_projector" in name: + 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") + + +def bytes_to_unicode(): + """ + Returns list of utf-8 byte and a corresponding list of unicode strings. + The reversible bpe codes work on unicode strings. + This means you need a large # of unicode characters in your vocab if you want to avoid UNKs. + When you're at something like a 10B token dataset you end up needing around 5K for decent coverage. + This is a significant percentage of your normal, say, 32K bpe vocab. + To avoid that, we want lookup tables between utf-8 bytes and unicode strings. + And avoids mapping to whitespace/control characters the bpe code barfs on. + """ + bs = ( + list(range(ord("!"), ord("~") + 1)) + + list(range(ord("¡"), ord("¬") + 1)) + + list(range(ord("®"), ord("ÿ") + 1)) + ) + cs = bs[:] + n = 0 + for b in range(2**8): + if b not in bs: + bs.append(b) + cs.append(2**8 + n) + n += 1 + cs = [chr(n) for n in cs] + return dict(zip(bs, cs)) + + +ap = argparse.ArgumentParser() +ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True) +ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16") +ap.add_argument("--text-only", action="store_true", required=False, + help="Save a text-only model. It can't be used to encode images") +ap.add_argument("--vision-only", action="store_true", required=False, + help="Save a vision-only model. It can't be used to encode texts") +ap.add_argument("--clip-model-is-vision", action="store_true", required=False, + help="The clip model is a pure vision model (ShareGPT4V vision extract for example)") +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("--minicpmv-projector", help="Path to minicpmv.projector file. If specified, save an image encoder for MiniCPM-V models.") +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 +default_image_mean = [0.48145466, 0.4578275, 0.40821073] +default_image_std = [0.26862954, 0.26130258, 0.27577711] +ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None) +ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None) +ap.add_argument('--minicpmv_version', type=int, help='minicpmv_version: MiniCPM-V-2 use 1; MiniCPM-V-2.5 use 2; MiniCPM-V-2.6 use 3', default=2) + +# with proper +args = ap.parse_args() + + +if args.text_only and args.vision_only: + print("--text-only and --image-only arguments cannot be specified at the same time.") + exit(1) + +if args.use_f32: + print("WARNING: Weights for the convolution op is always saved in f16, as the convolution op in GGML does not support 32-bit kernel weights yet.") + +# output in the same directory as the model if output_dir is None +dir_model = args.model_dir + +if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip: + vocab = None + tokens = None +else: + with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f: + vocab = json.load(f) + tokens = [key for key in vocab] + +# possible data types +# ftype == 0 -> float32 +# ftype == 1 -> float16 +# +# map from ftype to string +ftype_str = ["f32", "f16"] + +ftype = 1 +if args.use_f32: + ftype = 0 + +# if args.clip_model_is_vision or args.clip_model_is_openclip: +# model = CLIPVisionModel.from_pretrained(dir_model) +# processor = None +# else: +# model = CLIPModel.from_pretrained(dir_model) +# processor = CLIPProcessor.from_pretrained(dir_model) + +minicpmv_version = args.minicpmv_version +emb_dim = 4096 +if minicpmv_version == 1: + emb_dim = 2304 +elif minicpmv_version == 2: + emb_dim = 4096 +elif minicpmv_version == 3: + emb_dim = 3584 + +default_vision_config = { + "hidden_size": 1152, + "image_size": 980, + "intermediate_size": 4304, + "model_type": "idefics2", + "num_attention_heads": 16, + "num_hidden_layers": 27, + "patch_size": 14, + } + +vision_config = Idefics2VisionConfig(**default_vision_config) +model = Idefics2VisionTransformer(vision_config) +if minicpmv_version == 3: + vision_config = SiglipVisionConfig(**default_vision_config) + model = SiglipVisionTransformer(vision_config) + +processor = None +# if model.attn_pool is not None: +# model.attn_pool = torch.nn.Identity() + +# model.blocks = model.blocks[:-1] +model.load_state_dict(torch.load(os.path.join(dir_model, "minicpmv.clip"))) + +fname_middle = None +has_text_encoder = True +has_vision_encoder = True +has_minicpmv_projector = False + +if args.text_only: + fname_middle = "text-" + has_vision_encoder = False +elif args.minicpmv_projector is not None: + fname_middle = "mmproj-" + has_text_encoder = False + has_minicpmv_projector = True + minicpmv_version = 3 +elif args.vision_only: + fname_middle = "vision-" + has_text_encoder = False +else: + fname_middle = "" + +output_dir = args.output_dir if args.output_dir is not None else dir_model +os.makedirs(output_dir, exist_ok=True) +output_prefix = os.path.basename(output_dir).replace("ggml_", "") +fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf") +fout = GGUFWriter(path=fname_out, arch="clip") + +fout.add_bool("clip.has_text_encoder", has_text_encoder) +fout.add_bool("clip.has_vision_encoder", has_vision_encoder) +fout.add_bool("clip.has_minicpmv_projector", has_minicpmv_projector) +fout.add_file_type(ftype) +if args.text_only: + fout.add_description("text-only CLIP model") +elif args.vision_only and not has_minicpmv_projector: + fout.add_description("vision-only CLIP model") +elif has_minicpmv_projector: + fout.add_description("image encoder for MiniCPM-V") + # add projector type + fout.add_string("clip.projector_type", "resampler") + fout.add_int32("clip.minicpmv_version", minicpmv_version) +else: + fout.add_description("two-tower CLIP model") + +if has_vision_encoder: + # vision_model hparams + fout.add_uint32("clip.vision.image_size", 448) + fout.add_uint32("clip.vision.patch_size", 14) + fout.add_uint32(add_key_str(KEY_EMBEDDING_LENGTH, VISION), 1152) + fout.add_uint32(add_key_str(KEY_FEED_FORWARD_LENGTH, VISION), 4304) + fout.add_uint32("clip.vision.projection_dim", 0) + fout.add_uint32(add_key_str(KEY_ATTENTION_HEAD_COUNT, VISION), 16) + fout.add_float32(add_key_str(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6) + block_count = 26 + fout.add_uint32(add_key_str(KEY_BLOCK_COUNT, VISION), block_count) + + if processor is not None: + image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean + image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std + else: + image_mean = args.image_mean if args.image_mean is not None else default_image_mean + image_std = args.image_std if args.image_std is not None else default_image_std + fout.add_array("clip.vision.image_mean", image_mean) + fout.add_array("clip.vision.image_std", image_std) + +use_gelu = True +fout.add_bool("clip.use_gelu", use_gelu) + +def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): + """ + embed_dim: output dimension for each position + pos: a list of positions to be encoded: size (M,) + out: (M, D) + """ + assert embed_dim % 2 == 0 + omega = np.arange(embed_dim // 2, dtype=np.float32) + omega /= embed_dim / 2. + omega = 1. / 10000 ** omega # (D/2,) + + pos = pos.reshape(-1) # (M,) + out = np.einsum('m,d->md', pos, omega) # (M, D/2), outer product + + emb_sin = np.sin(out) # (M, D/2) + emb_cos = np.cos(out) # (M, D/2) + + emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D) + return emb + +def get_2d_sincos_pos_embed_from_grid(embed_dim, grid): + assert embed_dim % 2 == 0 + + # use half of dimensions to encode grid_h + emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2) + emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2) + + emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D) + return emb + + +# https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20 +def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False): + """ + grid_size: int of the grid height and width + return: + pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token) + """ + if isinstance(grid_size, int): + grid_h_size, grid_w_size = grid_size, grid_size + else: + grid_h_size, grid_w_size = grid_size[0], grid_size[1] + + grid_h = np.arange(grid_h_size, dtype=np.float32) + grid_w = np.arange(grid_w_size, dtype=np.float32) + grid = np.meshgrid(grid_w, grid_h) # here w goes first + grid = np.stack(grid, axis=0) + + grid = grid.reshape([2, 1, grid_h_size, grid_w_size]) + pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid) + if cls_token: + pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0) + return pos_embed + +def _replace_name_resampler(s, v): + if re.match("resampler.pos_embed", s): + return { + s: v, + re.sub("pos_embed", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(emb_dim, (70, 70))), + } + if re.match("resampler.proj", s): + return { + re.sub("proj", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(emb_dim, (70, 70))), + re.sub("proj", "proj.weight", s): v.transpose(-1, -2).contiguous(), + } + if re.match("resampler.attn.in_proj_.*", s): + return { + re.sub("attn.in_proj_", "attn.q.", s): v.chunk(3, dim=0)[0], + re.sub("attn.in_proj_", "attn.k.", s): v.chunk(3, dim=0)[1], + re.sub("attn.in_proj_", "attn.v.", s): v.chunk(3, dim=0)[2], + } + return {s: v} + +if has_minicpmv_projector: + projector = torch.load(args.minicpmv_projector) + new_state_dict = {} + for k, v in projector.items(): + kvs = _replace_name_resampler(k, v) + for nk, nv in kvs.items(): + new_state_dict[nk] = nv + projector = new_state_dict + ftype_cur = 0 + for name, data in projector.items(): + name = get_tensor_name(name) + data = data.squeeze().numpy() + + n_dims = len(data.shape) + if ftype == 1: + if name[-7:] == ".weight" and n_dims == 2: + print(" Converting to float16") + data = data.astype(np.float16) + ftype_cur = 1 + else: + print(" Converting to float32") + data = data.astype(np.float32) + ftype_cur = 0 + else: + if data.dtype != np.float32: + print(" Converting to float32") + data = data.astype(np.float32) + ftype_cur = 0 + + fout.add_tensor(name, data) + print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}") + + print("Projector tensors added\n") + +def _replace_name(s, v): + s = "vision_model." + s + if re.match("vision_model.embeddings.position_embedding", s): + v = v.unsqueeze(0) + return {s: v} + + return {s: v} + +state_dict = model.state_dict() +new_state_dict = {} +for k, v in state_dict.items(): + kvs = _replace_name(k, v) + for nk, nv in kvs.items(): + new_state_dict[nk] = nv +state_dict = new_state_dict +for name, data in state_dict.items(): + if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_minicpmv_projector): + # we don't need this + print(f"skipping parameter: {name}") + continue + + name = get_tensor_name(name) + data = data.squeeze().numpy() + + n_dims = len(data.shape) + + # ftype == 0 -> float32, ftype == 1 -> float16 + ftype_cur = 0 + if n_dims == 4: + print(f"tensor {name} is always saved in f16") + data = data.astype(np.float16) + ftype_cur = 1 + elif ftype == 1: + if name[-7:] == ".weight" and n_dims == 2: + print(" Converting to float16") + data = data.astype(np.float16) + ftype_cur = 1 + else: + print(" Converting to float32") + data = data.astype(np.float32) + ftype_cur = 0 + else: + if data.dtype != np.float32: + print(" Converting to float32") + data = data.astype(np.float32) + ftype_cur = 0 + + print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}") + fout.add_tensor(name, data) + + +fout.write_header_to_file() +fout.write_kv_data_to_file() +fout.write_tensors_to_file() +fout.close() + +print("Done. Output file: " + fname_out) diff --git a/examples/llava/minicpmv-surgery.py b/examples/llava/minicpmv-surgery.py new file mode 100644 index 000000000..748ff5c57 --- /dev/null +++ b/examples/llava/minicpmv-surgery.py @@ -0,0 +1,45 @@ +import argparse +import os +import torch +from transformers import AutoModel, AutoTokenizer + +ap = argparse.ArgumentParser() +ap.add_argument("-m", "--model", help="Path to MiniCPM-V model") +args = ap.parse_args() + +# find the model part that includes the the multimodal projector weights +model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True) +checkpoint = model.state_dict() + +# get a list of mm tensor names +mm_tensors = [k for k, v in checkpoint.items() if k.startswith("resampler")] + +# store these tensors in a new dictionary and torch.save them +projector = {name: checkpoint[name].float() for name in mm_tensors} +torch.save(projector, f"{args.model}/minicpmv.projector") + +clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vpm")] +if len(clip_tensors) > 0: + clip = {name.replace("vpm.", ""): checkpoint[name].float() for name in clip_tensors} + torch.save(clip, f"{args.model}/minicpmv.clip") + + # added tokens should be removed to be able to convert Mistral models + if os.path.exists(f"{args.model}/added_tokens.json"): + with open(f"{args.model}/added_tokens.json", "w") as f: + f.write("{}\n") + +config = model.llm.config +config.auto_map = { + "AutoConfig": "configuration_minicpm.MiniCPMConfig", + "AutoModel": "modeling_minicpm.MiniCPMModel", + "AutoModelForCausalLM": "modeling_minicpm.MiniCPMForCausalLM", + "AutoModelForSeq2SeqLM": "modeling_minicpm.MiniCPMForCausalLM", + "AutoModelForSequenceClassification": "modeling_minicpm.MiniCPMForSequenceClassification" +} +model.llm.save_pretrained(f"{args.model}/model") +tok = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True) +tok.save_pretrained(f"{args.model}/model") + +print("Done!") +print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.") +print(f"Also, use {args.model}/minicpmv.projector to prepare a minicpmv-encoder.gguf file.") diff --git a/examples/llava/requirements.txt b/examples/llava/requirements.txt index 21149d6fe..cbcbf26c9 100644 --- a/examples/llava/requirements.txt +++ b/examples/llava/requirements.txt @@ -1,3 +1,5 @@ --r ../../requirements/requirements-convert-legacy-llama.txt +-r ../../requirements/requirements-convert_legacy_llama.txt +--extra-index-url https://download.pytorch.org/whl/cpu pillow~=10.2.0 torch~=2.2.1 +torchvision~=0.17.1 diff --git a/examples/lookahead/lookahead.cpp b/examples/lookahead/lookahead.cpp index 7f6e42e8d..49288ab35 100644 --- a/examples/lookahead/lookahead.cpp +++ b/examples/lookahead/lookahead.cpp @@ -58,11 +58,11 @@ int main(int argc, char ** argv) { llama_backend_init(); llama_numa_init(params.numa); - llama_model * model = NULL; - llama_context * ctx = NULL; - // load the target model - std::tie(model, ctx) = llama_init_from_gpt_params(params); + llama_init_result llama_init = llama_init_from_gpt_params(params); + + llama_model * model = llama_init.model; + llama_context * ctx = llama_init.context; // Tokenize the prompt std::vector inp; diff --git a/examples/lookup/lookup-create.cpp b/examples/lookup/lookup-create.cpp index d713f6f21..5f04709f5 100644 --- a/examples/lookup/lookup-create.cpp +++ b/examples/lookup/lookup-create.cpp @@ -22,11 +22,11 @@ int main(int argc, char ** argv){ 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_init_result llama_init = llama_init_from_gpt_params(params); + + llama_model * model = llama_init.model; + llama_context * ctx = llama_init.context; GGML_ASSERT(model != nullptr); // tokenize the prompt diff --git a/examples/lookup/lookup-stats.cpp b/examples/lookup/lookup-stats.cpp index 0b171c872..400f3e0b0 100644 --- a/examples/lookup/lookup-stats.cpp +++ b/examples/lookup/lookup-stats.cpp @@ -26,12 +26,11 @@ int main(int argc, char ** argv){ 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(llama_n_vocab(model) < (1 << 16)); + llama_init_result llama_init = llama_init_from_gpt_params(params); + + llama_model * model = llama_init.model; + llama_context * ctx = llama_init.context; // tokenize the prompt std::vector inp; @@ -65,7 +64,7 @@ int main(int argc, char ** argv){ } const int n_input = inp.size(); - const int n_ctx = params.n_ctx; + const int n_ctx = llama_n_ctx(ctx); int n_drafted = 0; int n_accept = 0; diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp index db861d6ad..4f1bb5140 100644 --- a/examples/lookup/lookup.cpp +++ b/examples/lookup/lookup.cpp @@ -34,12 +34,11 @@ int main(int argc, char ** argv){ 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(llama_n_vocab(model) < (1 << 16)); + llama_init_result llama_init = llama_init_from_gpt_params(params); + + llama_model * model = llama_init.model; + llama_context * ctx = llama_init.context; // tokenize the prompt std::vector inp; diff --git a/examples/main/README.md b/examples/main/README.md index 61e4a42f7..9396a34fa 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -1,6 +1,6 @@ # llama.cpp/examples/main -This example program allows you to use various LLaMA language models in an easy and efficient way. It is specifically designed to work with the [llama.cpp](https://github.com/ggerganov/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts. +This example program allows you to use various LLaMA language models easily and efficiently. It is specifically designed to work with the [llama.cpp](https://github.com/ggerganov/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts. ## Table of Contents @@ -17,60 +17,59 @@ This example program allows you to use various LLaMA language models in an easy To get started right away, run the following command, making sure to use the correct path for the model you have: -#### Unix-based systems (Linux, macOS, etc.): +First, we will need to download a model. In these examples, we will use the Gemma model from the ggml-org repo on Hugging Face. +[https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true](https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true) + +Once downloaded, place your model in the models folder in llama.cpp. + +### Unix-based systems (Linux, macOS, etc.): + +##### Input prompt (One-and-done) ```bash -./llama-cli -m models/7B/ggml-model.bin --prompt "Once upon a time" +./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --prompt "Once upon a time" ``` - -#### Windows: - -```powershell -llama-cli.exe -m models\7B\ggml-model.bin --prompt "Once upon a time" -``` - -For an interactive experience, try this command: - -#### Unix-based systems (Linux, macOS, etc.): +##### Conversation mode (Allow for continuous interaction with the model) ```bash -./llama-cli -m models/7B/ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -i -p \ -'User: Hi -AI: Hello. I am an AI chatbot. Would you like to talk? -User: Sure! -AI: What would you like to talk about? -User:' +./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf -cnv --chat-template gemma ``` -#### Windows: - -```powershell -llama-cli.exe -m models\7B\ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -i -e -p "User: Hi\nAI: Hello. I am an AI chatbot. Would you like to talk?\nUser: Sure!\nAI: What would you like to talk about?\nUser:" -``` - -The following command generates "infinite" text from a starting prompt (you can use `Ctrl-C` to stop it): - -#### Unix-based systems (Linux, macOS, etc.): - +##### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it): ```bash -./llama-cli -m models/7B/ggml-model.bin --ignore-eos -n -1 +./llama-cli -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1 ``` -#### Windows: +### Windows: + +##### Input prompt (One-and-done) +```powershell +./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --prompt "Once upon a time" +``` +##### Conversation mode (Allow for continuous interaction with the model) ```powershell -llama-cli.exe -m models\7B\ggml-model.bin --ignore-eos -n -1 +./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf -cnv --chat-template gemma +``` + +#### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it): + +```powershell +llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1 ``` ## Common Options In this section, we cover the most commonly used options for running the `llama-cli` program with the LLaMA models: -- `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`; inferred from `--model-url` if set). -- `-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). +- `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/gemma-1.1-7b-it.Q4_K_M.gguf`; inferred from `--model-url` if set). +- `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g [https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true](https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true)). - `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses. - `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text. - `-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. +- `-mli, --multiline-input`: Allows you to write or paste multiple lines without ending each in '\' +- `-t N, --threads N`: Set the number of threads to use during generation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has. +- - `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance. ## Input Prompts @@ -90,6 +89,7 @@ In interactive mode, users can participate in text generation by injecting their - `-i, --interactive`: Run the program in interactive mode, allowing users to engage in real-time conversations or provide specific instructions to the model. - `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation. +- `-cnv, --conversation`: Run the program in conversation mode (does not print special tokens and suffix/prefix, use default chat template) (default: false) - `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text. By understanding and utilizing these interaction options, you can create engaging and dynamic experiences with the LLaMA models, tailoring the text generation process to your specific needs. @@ -117,6 +117,13 @@ The `--in-suffix` flag is used to add a suffix after your input. This is useful ```sh ./llama-cli -r "User:" --in-prefix " " --in-suffix "Assistant:" ``` +When --in-prefix or --in-suffix options are enabled the chat template ( --chat-template ) is disabled + +### Chat templates + + `--chat-template JINJA_TEMPLATE`: This option sets a custom jinja chat template. It accepts a string, not a file name. Default: template taken from model's metadata. Llama.cpp only supports [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template). These include llama2, llama3, gemma, monarch, chatml, orion, vicuna, vicuna-orca, deepseek, command-r, zephyr. When --in-prefix or --in-suffix options are enabled the chat template ( --chat-template ) is disabled. + + Example usage: `--chat-template gemma` ## Context Management @@ -124,9 +131,7 @@ During text generation, LLaMA models have a limited context size, which means th ### Context Size -The `--ctx-size` option allows you to set the size of the prompt context used by the LLaMA models during text generation. A larger context size helps the model to better comprehend and generate responses for longer input or conversations. - -- `-c N, --ctx-size N`: Set the size of the prompt context (default: 512). The LLaMA models were built with a context of 2048, which will yield the best results on longer input/inference. However, increasing the context size beyond 2048 may lead to unpredictable results. +- `-c N, --ctx-size N`: Set the size of the prompt context (default: 0, 0 = loaded from model). The LLaMA models were built with a context of 2048-8192, which will yield the best results on longer input/inference. ### Extended Context Size @@ -148,15 +153,15 @@ The following options allow you to control the text generation process and fine- ### Number of Tokens to Predict -- `-n N, --n-predict N`: Set the number of tokens to predict when generating text (default: 128, -1 = infinity, -2 = until context filled) +- `-n N, --predict N`: Set the number of tokens to predict when generating text (default: -1, -1 = infinity, -2 = until context filled) -The `--n-predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text. +The `--predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text. -A value of -1 will enable infinite text generation, even though we have a finite context window. When the context window is full, some of the earlier tokens (half of the tokens after `--n-keep`) will be discarded. The context must then be re-evaluated before generation can resume. On large models and/or large context windows, this will result in significant pause in output. +A value of -1 will enable infinite text generation, even though we have a finite context window. When the context window is full, some of the earlier tokens (half of the tokens after `--keep`) will be discarded. The context must then be re-evaluated before generation can resume. On large models and/or large context windows, this will result in a significant pause in output. If the pause is undesirable, a value of -2 will stop generation immediately when the context is filled. -It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `n-predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the `--ignore-eos` parameter. +It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode, text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `--predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the `--ignore-eos` parameter. ### Temperature @@ -164,15 +169,15 @@ It is important to note that the generated text may be shorter than the specifie Temperature is a hyperparameter that controls the randomness of the generated text. It affects the probability distribution of the model's output tokens. A higher temperature (e.g., 1.5) makes the output more random and creative, while a lower temperature (e.g., 0.5) makes the output more focused, deterministic, and conservative. The default value is 0.8, which provides a balance between randomness and determinism. At the extreme, a temperature of 0 will always pick the most likely next token, leading to identical outputs in each run. -Example usage: `--temp 0.5` +Example usage: `--temp 0` ### Repeat Penalty -- `--repeat-penalty N`: Control the repetition of token sequences in the generated text (default: 1.1). +- `--repeat-penalty N`: Control the repetition of token sequences in the generated text default: 1.0, 1.0 = disabled). - `--repeat-last-n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size). - `--no-penalize-nl`: Disable penalization for newline tokens when applying the repeat penalty. -The `repeat-penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.1. +The `repeat-penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1. The `repeat-last-n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`ctx-size`). @@ -196,19 +201,19 @@ Top-p sampling, also known as nucleus sampling, is another text generation metho Example usage: `--top-p 0.95` -### Min P Sampling +### Min-P Sampling -- `--min-p N`: Sets a minimum base probability threshold for token selection (default: 0.05). +- `--min-p N`: Sets a minimum base probability threshold for token selection (default: 0.1). The Min-P sampling method was designed as an alternative to Top-P, and aims to ensure a balance of quality and variety. The parameter *p* represents the minimum probability for a token to be considered, relative to the probability of the most likely token. For example, with *p*=0.05 and the most likely token having a probability of 0.9, logits with a value less than 0.045 are filtered out. Example usage: `--min-p 0.05` -### Tail Free Sampling (TFS) +### Tail-Free Sampling (TFS) - `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled). -Tail free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. Similar to Top-P it tries to determine the bulk of the most likely tokens dynamically. But TFS filters out logits based on the second derivative of their probabilities. Adding tokens is stopped after the sum of the second derivatives reaches the parameter z. In short: TFS looks how quickly the probabilities of the tokens decrease and cuts off the tail of unlikely tokens using the parameter z. Typical values for z are in the range of 0.9 to 0.95. A value of 1.0 would include all tokens, and thus disables the effect of TFS. +Tail-free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. Similar to Top-P it tries to determine the bulk of the most likely tokens dynamically. But TFS filters out logits based on the second derivative of their probabilities. Adding tokens is stopped after the sum of the second derivatives reaches the parameter z. In short: TFS looks at how quickly the probabilities of the tokens decrease and cuts off the tail of unlikely tokens using the parameter z. Typical values for z are in the range of 0.9 to 0.95. A value of 1.0 would include all tokens and thus disables the effect of TFS. Example usage: `--tfs 0.95` @@ -307,10 +312,8 @@ These options provide extra functionality and customization when running the LLa - `-h, --help`: Display a help message showing all available options and their default values. This is particularly useful for checking the latest options and default values, as they can change frequently, and the information in this document may become outdated. - `--verbose-prompt`: Print the prompt before generating text. -- `-ngl N, --n-gpu-layers N`: When compiled with GPU support, 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. - `-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. - `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains. - `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation. - - `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable or in an OS-specific local cache. diff --git a/examples/main/main.cpp b/examples/main/main.cpp index fb1963fde..6304ec4fc 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -37,7 +37,8 @@ static gpt_params * g_params; static std::vector * g_input_tokens; static std::ostringstream * g_output_ss; static std::vector * g_output_tokens; -static bool is_interacting = false; +static bool is_interacting = false; +static bool need_insert_eot = false; static bool file_exists(const std::string & path) { std::ifstream f(path.c_str()); @@ -99,7 +100,8 @@ static void write_logfile( static void sigint_handler(int signo) { if (signo == SIGINT) { if (!is_interacting && g_params->interactive) { - is_interacting = true; + is_interacting = true; + need_insert_eot = true; } else { console::cleanup(); printf("\n"); @@ -122,6 +124,7 @@ static std::string chat_add_and_format(struct llama_model * model, std::vectorchat_template, chat_msgs, new_msg, role == "user"); chat_msgs.push_back({role, content}); + LOG("formatted: %s\n", formatted.c_str()); return formatted; } @@ -204,7 +207,10 @@ int main(int argc, char ** argv) { // load the model and apply lora adapter, if any LOG("%s: load the model and apply lora adapter, if any\n", __func__); - std::tie(model, ctx) = llama_init_from_gpt_params(params); + llama_init_result llama_init = llama_init_from_gpt_params(params); + + model = llama_init.model; + ctx = llama_init.context; if (sparams.cfg_scale > 1.f) { struct llama_context_params lparams = llama_context_params_from_gpt_params(params); ctx_guidance = llama_new_context_with_model(model, lparams); @@ -215,6 +221,40 @@ int main(int argc, char ** argv) { return 1; } + LOG("%s: llama threadpool init = n_threads = %d\n", + __func__, + (int) params.cpuparams.n_threads + ); + struct ggml_threadpool_params tpp_batch = + ggml_threadpool_params_from_cpu_params(params.cpuparams_batch); + struct ggml_threadpool_params tpp = + ggml_threadpool_params_from_cpu_params(params.cpuparams); + + set_process_priority(params.cpuparams.priority); + + struct ggml_threadpool * threadpool_batch = NULL; + if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) { + threadpool_batch = ggml_threadpool_new(&tpp_batch); + if (!threadpool_batch) { + LOG_TEE("%s: batch threadpool create failed : n_threads %d\n", __func__, tpp_batch.n_threads); + exit(1); + } + + // Start the non-batch threadpool in the paused state + tpp.paused = true; + } + + struct ggml_threadpool * threadpool = ggml_threadpool_new(&tpp); + if (!threadpool) { + LOG_TEE("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads); + exit(1); + } + + llama_attach_threadpool(ctx, threadpool, threadpool_batch); + if (ctx_guidance) { + llama_attach_threadpool(ctx_guidance, threadpool, threadpool_batch); + } + const int n_ctx_train = llama_n_ctx_train(model); const int n_ctx = llama_n_ctx(ctx); LOG("n_ctx: %d\n", n_ctx); @@ -224,7 +264,14 @@ int main(int argc, char ** argv) { __func__, n_ctx_train, n_ctx); } - LOG_TEE("%s: chat template example: %s\n", __func__, llama_chat_format_example(model, params.chat_template).c_str()); + // print chat template example in conversation mode + if (params.conversation) { + if (params.enable_chat_template) { + LOG_TEE("%s: chat template example: %s\n", __func__, llama_chat_format_example(model, params.chat_template).c_str()); + } else { + LOG_TEE("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__); + } + } // print system information { @@ -254,16 +301,16 @@ int main(int argc, char ** argv) { } } - const bool add_bos = llama_should_add_bos_token(model); + const bool add_bos = llama_add_bos_token(model); if (!llama_model_has_encoder(model)) { - GGML_ASSERT(llama_add_eos_token(model) != 1); + GGML_ASSERT(!llama_add_eos_token(model)); } LOG("add_bos: %d\n", add_bos); std::vector embd_inp; { - auto prompt = (params.conversation && params.enable_chat_template) + auto prompt = (params.conversation && params.enable_chat_template && !params.prompt.empty()) ? chat_add_and_format(model, chat_msgs, "system", params.prompt) // format the system prompt in conversation mode : params.prompt; if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) { @@ -280,8 +327,13 @@ int main(int argc, char ** argv) { // Should not run without any tokens if (embd_inp.empty()) { - embd_inp.push_back(llama_token_bos(model)); - LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str()); + if (add_bos) { + embd_inp.push_back(llama_token_bos(model)); + LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str()); + } else { + LOG_TEE("error: input is empty\n"); + return -1; + } } // Tokenize negative prompt @@ -910,6 +962,13 @@ int main(int argc, char ** argv) { LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str()); + // if user stop generation mid-way, we must add EOT to finish model's last response + if (need_insert_eot && format_chat) { + llama_token eot = llama_token_eot(model); + embd_inp.push_back(eot == -1 ? llama_token_eos(model) : eot); + need_insert_eot = false; + } + embd_inp.insert(embd_inp.end(), line_pfx.begin(), line_pfx.end()); embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end()); embd_inp.insert(embd_inp.end(), line_sfx.begin(), line_sfx.end()); @@ -969,6 +1028,9 @@ int main(int argc, char ** argv) { llama_sampling_free(ctx_sampling); llama_backend_free(); + ggml_threadpool_free(threadpool); + ggml_threadpool_free(threadpool_batch); + #ifndef LOG_DISABLE_LOGS LOG_TEE("Log end\n"); #endif // LOG_DISABLE_LOGS diff --git a/examples/parallel/parallel.cpp b/examples/parallel/parallel.cpp index f68478804..32c3ba2b0 100644 --- a/examples/parallel/parallel.cpp +++ b/examples/parallel/parallel.cpp @@ -129,11 +129,11 @@ int main(int argc, char ** argv) { llama_backend_init(); llama_numa_init(params.numa); - llama_model * model = NULL; - llama_context * ctx = NULL; - // load the target model - std::tie(model, ctx) = llama_init_from_gpt_params(params); + llama_init_result llama_init = llama_init_from_gpt_params(params); + + llama_model * model = llama_init.model; + llama_context * ctx = llama_init.context; // load the prompts from an external file if there are any if (params.prompt.empty()) { diff --git a/examples/passkey/README.md b/examples/passkey/README.md index a48a6283a..2b8e910f9 100644 --- a/examples/passkey/README.md +++ b/examples/passkey/README.md @@ -1,5 +1,8 @@ # llama.cpp/example/passkey +A passkey retrieval task is an evaluation method used to measure a language +models ability to recall information from long contexts. + See the following PRs for more info: - https://github.com/ggerganov/llama.cpp/pull/3856 diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 454296894..f2b0b9df8 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -340,8 +340,8 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & // Output: `perplexity: 13.5106 [114/114]` // BOS tokens will be added for each chunk before eval - const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); - GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1); + const bool add_bos = llama_add_bos_token(llama_get_model(ctx)); + GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx))); fprintf(stderr, "%s: tokenizing the input ..\n", __func__); @@ -480,8 +480,8 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par // Output: `perplexity: 13.5106 [114/114]` // BOS tokens will be added for each chunk before eval - const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); - GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1); + const bool add_bos = llama_add_bos_token(llama_get_model(ctx)); + GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx))); std::ofstream logits_stream; if (!params.logits_file.empty()) { @@ -1733,8 +1733,8 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { const int n_batch = params.n_batch; const int num_batches = (n_ctx + n_batch - 1)/n_batch; const int nv = 2*((n_vocab + 1)/2) + 4; - const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); - GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1); + const bool add_bos = llama_add_bos_token(llama_get_model(ctx)); + GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx))); std::vector log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv); std::vector kld_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk); @@ -2018,11 +2018,11 @@ int main(int argc, char ** argv) { llama_backend_init(); llama_numa_init(params.numa); - llama_model * model; - llama_context * ctx; - // load the model and apply lora adapter, if any - std::tie(model, ctx) = llama_init_from_gpt_params(params); + llama_init_result llama_init = llama_init_from_gpt_params(params); + + llama_model * model = llama_init.model; + llama_context * ctx = llama_init.context; if (model == NULL) { fprintf(stderr, "%s: error: unable to load model\n", __func__); return 1; diff --git a/examples/pydantic-models-to-grammar-examples.py b/examples/pydantic-models-to-grammar-examples.py deleted file mode 100644 index 160966649..000000000 --- a/examples/pydantic-models-to-grammar-examples.py +++ /dev/null @@ -1,224 +0,0 @@ -# Function calling example using pydantic models. -import datetime -import importlib -import json -from enum import Enum -from typing import Optional, Union - -import requests -from pydantic import BaseModel, Field -from pydantic_models_to_grammar import (add_run_method_to_dynamic_model, convert_dictionary_to_pydantic_model, - create_dynamic_model_from_function, generate_gbnf_grammar_and_documentation) - - -# Function to get completion on the llama.cpp server with grammar. -def create_completion(prompt, grammar): - headers = {"Content-Type": "application/json"} - data = {"prompt": prompt, "grammar": grammar} - - response = requests.post("http://127.0.0.1:8080/completion", headers=headers, json=data) - data = response.json() - - print(data["content"]) - return data["content"] - - -# A function for the agent to send a message to the user. -class SendMessageToUser(BaseModel): - """ - Send a message to the User. - """ - chain_of_thought: str = Field(..., description="Your chain of thought while sending the message.") - message: str = Field(..., description="Message you want to send to the user.") - - def run(self): - print(self.message) - - -# Enum for the calculator tool. -class MathOperation(Enum): - ADD = "add" - SUBTRACT = "subtract" - MULTIPLY = "multiply" - DIVIDE = "divide" - - -# Simple pydantic calculator tool for the agent that can add, subtract, multiply, and divide. Docstring and description of fields will be used in system prompt. -class Calculator(BaseModel): - """ - Perform a math operation on two numbers. - """ - number_one: Union[int, float] = Field(..., description="First number.") - operation: MathOperation = Field(..., description="Math operation to perform.") - number_two: Union[int, float] = Field(..., description="Second number.") - - def run(self): - if self.operation == MathOperation.ADD: - return self.number_one + self.number_two - elif self.operation == MathOperation.SUBTRACT: - return self.number_one - self.number_two - elif self.operation == MathOperation.MULTIPLY: - return self.number_one * self.number_two - elif self.operation == MathOperation.DIVIDE: - return self.number_one / self.number_two - else: - raise ValueError("Unknown operation.") - - -# Here the grammar gets generated by passing the available function models to generate_gbnf_grammar_and_documentation function. This also generates a documentation usable by the LLM. -# pydantic_model_list is the list of pydanitc models -# outer_object_name is an optional name for an outer object around the actual model object. Like a "function" object with "function_parameters" which contains the actual model object. If None, no outer object will be generated -# outer_object_content is the name of outer object content. -# model_prefix is the optional prefix for models in the documentation. (Default="Output Model") -# fields_prefix is the prefix for the model fields in the documentation. (Default="Output Fields") -gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation( - pydantic_model_list=[SendMessageToUser, Calculator], outer_object_name="function", - outer_object_content="function_parameters", model_prefix="Function", fields_prefix="Parameters") - -print(gbnf_grammar) -print(documentation) - -system_message = "You are an advanced AI, tasked to assist the user by calling functions in JSON format. The following are the available functions and their parameters and types:\n\n" + documentation - -user_message = "What is 42 * 42?" -prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant" - -text = create_completion(prompt=prompt, grammar=gbnf_grammar) -# This should output something like this: -# { -# "function": "calculator", -# "function_parameters": { -# "number_one": 42, -# "operation": "multiply", -# "number_two": 42 -# } -# } -function_dictionary = json.loads(text) -if function_dictionary["function"] == "calculator": - function_parameters = {**function_dictionary["function_parameters"]} - - print(Calculator(**function_parameters).run()) - # This should output: 1764 - - -# A example structured output based on pydantic models. The LLM will create an entry for a Book database out of an unstructured text. -class Category(Enum): - """ - The category of the book. - """ - Fiction = "Fiction" - NonFiction = "Non-Fiction" - - -class Book(BaseModel): - """ - Represents an entry about a book. - """ - title: str = Field(..., description="Title of the book.") - author: str = Field(..., description="Author of the book.") - published_year: Optional[int] = Field(..., description="Publishing year of the book.") - keywords: list[str] = Field(..., description="A list of keywords.") - category: Category = Field(..., description="Category of the book.") - summary: str = Field(..., description="Summary of the book.") - - -# We need no additional parameters other than our list of pydantic models. -gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation([Book]) - -system_message = "You are an advanced AI, tasked to create a dataset entry in JSON for a Book. The following is the expected output model:\n\n" + documentation - -text = """The Feynman Lectures on Physics is a physics textbook based on some lectures by Richard Feynman, a Nobel laureate who has sometimes been called "The Great Explainer". The lectures were presented before undergraduate students at the California Institute of Technology (Caltech), during 1961–1963. The book's co-authors are Feynman, Robert B. Leighton, and Matthew Sands.""" -prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant" - -text = create_completion(prompt=prompt, grammar=gbnf_grammar) - -json_data = json.loads(text) - -print(Book(**json_data)) -# An example for parallel function calling with a Python function, a pydantic function model and an OpenAI like function definition. - -def get_current_datetime(output_format: Optional[str] = None): - """ - Get the current date and time in the given format. - Args: - output_format: formatting string for the date and time, defaults to '%Y-%m-%d %H:%M:%S' - """ - if output_format is None: - output_format = '%Y-%m-%d %H:%M:%S' - return datetime.datetime.now().strftime(output_format) - - -# Example function to get the weather -def get_current_weather(location, unit): - """Get the current weather in a given location""" - if "London" in location: - return json.dumps({"location": "London", "temperature": "42", "unit": unit.value}) - elif "New York" in location: - return json.dumps({"location": "New York", "temperature": "24", "unit": unit.value}) - elif "North Pole" in location: - return json.dumps({"location": "North Pole", "temperature": "-42", "unit": unit.value}) - else: - return json.dumps({"location": location, "temperature": "unknown"}) - - -# Here is a function definition in OpenAI style -current_weather_tool = { - "type": "function", - "function": { - "name": "get_current_weather", - "description": "Get the current weather in a given location", - "parameters": { - "type": "object", - "properties": { - "location": { - "type": "string", - "description": "The city and state, e.g. San Francisco, CA", - }, - "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}, - }, - "required": ["location"], - }, - }, -} - -# Convert OpenAI function definition into pydantic model -current_weather_tool_model = convert_dictionary_to_pydantic_model(current_weather_tool) -# Add the actual function to a pydantic model -current_weather_tool_model = add_run_method_to_dynamic_model(current_weather_tool_model, get_current_weather) - -# Convert normal Python function to a pydantic model -current_datetime_model = create_dynamic_model_from_function(get_current_datetime) - -tool_list = [SendMessageToUser, Calculator, current_datetime_model, current_weather_tool_model] - - -gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation( - pydantic_model_list=tool_list, outer_object_name="function", - outer_object_content="params", model_prefix="Function", fields_prefix="Parameters", list_of_outputs=True) - -system_message = "You are an advanced AI assistant. You are interacting with the user and with your environment by calling functions. You call functions by writing JSON objects, which represent specific function calls.\nBelow is a list of your available function calls:\n\n" + documentation - - -text = """Get the date and time, get the current weather in celsius in London and solve the following calculation: 42 * 42""" -prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant" - -text = create_completion(prompt=prompt, grammar=gbnf_grammar) - -json_data = json.loads(text) - -print(json_data) -# Should output something like this: -# [{'function': 'get_current_datetime', 'params': {'output_format': '%Y-%m-%d %H:%M:%S'}}, {'function': 'get_current_weather', 'params': {'location': 'London', 'unit': 'celsius'}}, {'function': 'Calculator', 'params': {'number_one': 42, 'operation': 'multiply', 'number_two': 42}}] - - -for call in json_data: - if call["function"] == "Calculator": - print(Calculator(**call["params"]).run()) - elif call["function"] == "get_current_datetime": - print(current_datetime_model(**call["params"]).run()) - elif call["function"] == "get_current_weather": - print(current_weather_tool_model(**call["params"]).run()) -# Should output something like this: -# 2024-01-14 13:36:06 -# {"location": "London", "temperature": "42", "unit": "celsius"} -# 1764 diff --git a/examples/pydantic_models_to_grammar.py b/examples/pydantic_models_to_grammar.py index f029c73a2..93e5dcb6c 100644 --- a/examples/pydantic_models_to_grammar.py +++ b/examples/pydantic_models_to_grammar.py @@ -9,7 +9,7 @@ from inspect import getdoc, isclass from typing import TYPE_CHECKING, Any, Callable, List, Optional, Union, get_args, get_origin, get_type_hints from docstring_parser import parse -from pydantic import BaseModel, Field, create_model +from pydantic import BaseModel, create_model if TYPE_CHECKING: from types import GenericAlias @@ -17,6 +17,9 @@ else: # python 3.8 compat from typing import _GenericAlias as GenericAlias +# TODO: fix this +# pyright: reportAttributeAccessIssue=information + class PydanticDataType(Enum): """ @@ -50,35 +53,38 @@ class PydanticDataType(Enum): def map_pydantic_type_to_gbnf(pydantic_type: type[Any]) -> str: - if isclass(pydantic_type) and issubclass(pydantic_type, str): + origin_type = get_origin(pydantic_type) + origin_type = pydantic_type if origin_type is None else origin_type + + if isclass(origin_type) and issubclass(origin_type, str): return PydanticDataType.STRING.value - elif isclass(pydantic_type) and issubclass(pydantic_type, bool): + elif isclass(origin_type) and issubclass(origin_type, bool): return PydanticDataType.BOOLEAN.value - elif isclass(pydantic_type) and issubclass(pydantic_type, int): + elif isclass(origin_type) and issubclass(origin_type, int): return PydanticDataType.INTEGER.value - elif isclass(pydantic_type) and issubclass(pydantic_type, float): + elif isclass(origin_type) and issubclass(origin_type, float): return PydanticDataType.FLOAT.value - elif isclass(pydantic_type) and issubclass(pydantic_type, Enum): + elif isclass(origin_type) and issubclass(origin_type, Enum): return PydanticDataType.ENUM.value - elif isclass(pydantic_type) and issubclass(pydantic_type, BaseModel): - return format_model_and_field_name(pydantic_type.__name__) - elif get_origin(pydantic_type) is list: + elif isclass(origin_type) and issubclass(origin_type, BaseModel): + return format_model_and_field_name(origin_type.__name__) + elif origin_type is list: element_type = get_args(pydantic_type)[0] return f"{map_pydantic_type_to_gbnf(element_type)}-list" - elif get_origin(pydantic_type) is set: + elif origin_type is set: element_type = get_args(pydantic_type)[0] return f"{map_pydantic_type_to_gbnf(element_type)}-set" - elif get_origin(pydantic_type) is Union: + elif origin_type is Union: union_types = get_args(pydantic_type) union_rules = [map_pydantic_type_to_gbnf(ut) for ut in union_types] return f"union-{'-or-'.join(union_rules)}" - elif get_origin(pydantic_type) is Optional: + elif origin_type is Optional: element_type = get_args(pydantic_type)[0] return f"optional-{map_pydantic_type_to_gbnf(element_type)}" - elif isclass(pydantic_type): - return f"{PydanticDataType.CUSTOM_CLASS.value}-{format_model_and_field_name(pydantic_type.__name__)}" - elif get_origin(pydantic_type) is dict: + elif isclass(origin_type): + return f"{PydanticDataType.CUSTOM_CLASS.value}-{format_model_and_field_name(origin_type.__name__)}" + elif origin_type is dict: key_type, value_type = get_args(pydantic_type) return f"custom-dict-key-type-{format_model_and_field_name(map_pydantic_type_to_gbnf(key_type))}-value-type-{format_model_and_field_name(map_pydantic_type_to_gbnf(value_type))}" else: @@ -115,7 +121,7 @@ def get_members_structure(cls, rule_name): # Modify this comprehension members = [ f' "\\"{name}\\"" ":" {map_pydantic_type_to_gbnf(param_type)}' - for name, param_type in cls.__annotations__.items() + for name, param_type in get_type_hints(cls).items() if name != "self" ] @@ -234,8 +240,9 @@ def generate_gbnf_float_rules(max_digit=None, min_digit=None, max_precision=None # Define the integer part rule integer_part_rule = ( - "integer-part" + (f"-max{max_digit}" if max_digit is not None else "") + ( - f"-min{min_digit}" if min_digit is not None else "") + "integer-part" + + (f"-max{max_digit}" if max_digit is not None else "") + + (f"-min{min_digit}" if min_digit is not None else "") ) # Define the fractional part rule based on precision constraints @@ -293,17 +300,20 @@ def generate_gbnf_rule_for_type( field_name = format_model_and_field_name(field_name) gbnf_type = map_pydantic_type_to_gbnf(field_type) - if isclass(field_type) and issubclass(field_type, BaseModel): + origin_type = get_origin(field_type) + origin_type = field_type if origin_type is None else origin_type + + if isclass(origin_type) and issubclass(origin_type, BaseModel): nested_model_name = format_model_and_field_name(field_type.__name__) nested_model_rules, _ = generate_gbnf_grammar(field_type, processed_models, created_rules) rules.extend(nested_model_rules) gbnf_type, rules = nested_model_name, rules - elif isclass(field_type) and issubclass(field_type, Enum): + elif isclass(origin_type) and issubclass(origin_type, Enum): enum_values = [f'"\\"{e.value}\\""' for e in field_type] # Adding escaped quotes enum_rule = f"{model_name}-{field_name} ::= {' | '.join(enum_values)}" rules.append(enum_rule) gbnf_type, rules = model_name + "-" + field_name, rules - elif get_origin(field_type) == list: # Array + elif origin_type is list: # Array element_type = get_args(field_type)[0] element_rule_name, additional_rules = generate_gbnf_rule_for_type( model_name, f"{field_name}-element", element_type, is_optional, processed_models, created_rules @@ -313,7 +323,7 @@ def generate_gbnf_rule_for_type( rules.append(array_rule) gbnf_type, rules = model_name + "-" + field_name, rules - elif get_origin(field_type) == set or field_type == set: # Array + elif origin_type is set: # Array element_type = get_args(field_type)[0] element_rule_name, additional_rules = generate_gbnf_rule_for_type( model_name, f"{field_name}-element", element_type, is_optional, processed_models, created_rules @@ -367,7 +377,7 @@ def generate_gbnf_rule_for_type( gbnf_type = f"{model_name}-{field_name}-optional" else: gbnf_type = f"{model_name}-{field_name}-union" - elif isclass(field_type) and issubclass(field_type, str): + elif isclass(origin_type) and issubclass(origin_type, str): if field_info and hasattr(field_info, "json_schema_extra") and field_info.json_schema_extra is not None: triple_quoted_string = field_info.json_schema_extra.get("triple_quoted_string", False) markdown_string = field_info.json_schema_extra.get("markdown_code_block", False) @@ -383,8 +393,8 @@ def generate_gbnf_rule_for_type( gbnf_type = PydanticDataType.STRING.value elif ( - isclass(field_type) - and issubclass(field_type, float) + isclass(origin_type) + and issubclass(origin_type, float) and field_info and hasattr(field_info, "json_schema_extra") and field_info.json_schema_extra is not None @@ -409,8 +419,8 @@ def generate_gbnf_rule_for_type( ) elif ( - isclass(field_type) - and issubclass(field_type, int) + isclass(origin_type) + and issubclass(origin_type, int) and field_info and hasattr(field_info, "json_schema_extra") and field_info.json_schema_extra is not None @@ -458,7 +468,7 @@ def generate_gbnf_grammar(model: type[BaseModel], processed_models: set[type[Bas if not issubclass(model, BaseModel): # For non-Pydantic classes, generate model_fields from __annotations__ or __init__ if hasattr(model, "__annotations__") and model.__annotations__: - model_fields = {name: (typ, ...) for name, typ in model.__annotations__.items()} + model_fields = {name: (typ, ...) for name, typ in get_type_hints(model).items()} else: init_signature = inspect.signature(model.__init__) parameters = init_signature.parameters @@ -466,7 +476,7 @@ def generate_gbnf_grammar(model: type[BaseModel], processed_models: set[type[Bas name != "self"} else: # For Pydantic models, use model_fields and check for ellipsis (required fields) - model_fields = model.__annotations__ + model_fields = get_type_hints(model) model_rule_parts = [] nested_rules = [] @@ -680,7 +690,7 @@ def generate_markdown_documentation( str: Generated text documentation. """ documentation = "" - pyd_models = [(model, True) for model in pydantic_models] + pyd_models: list[tuple[type[BaseModel], bool]] = [(model, True) for model in pydantic_models] for model, add_prefix in pyd_models: if add_prefix: documentation += f"{model_prefix}: {model.__name__}\n" @@ -700,9 +710,9 @@ def generate_markdown_documentation( # Indenting the fields section documentation += f" {fields_prefix}:\n" else: - documentation += f" Fields:\n" + documentation += f" Fields:\n" # noqa: F541 if isclass(model) and issubclass(model, BaseModel): - for name, field_type in model.__annotations__.items(): + for name, field_type in get_type_hints(model).items(): # if name == "markdown_code_block": # continue if get_origin(field_type) == list: @@ -750,14 +760,17 @@ def generate_field_markdown( field_info = model.model_fields.get(field_name) field_description = field_info.description if field_info and field_info.description else "" - if get_origin(field_type) == list: + origin_type = get_origin(field_type) + origin_type = field_type if origin_type is None else origin_type + + if origin_type == list: element_type = get_args(field_type)[0] field_text = f"{indent}{field_name} ({format_model_and_field_name(field_type.__name__)} of {format_model_and_field_name(element_type.__name__)})" if field_description != "": field_text += ":\n" else: field_text += "\n" - elif get_origin(field_type) == Union: + elif origin_type == Union: element_types = get_args(field_type) types = [] for element_type in element_types: @@ -778,7 +791,7 @@ def generate_field_markdown( return field_text if field_description != "": - field_text += f" Description: " + field_description + "\n" + field_text += f" Description: {field_description}\n" # Check for and include field-specific examples if available if hasattr(model, "Config") and hasattr(model.Config, @@ -788,9 +801,9 @@ def generate_field_markdown( example_text = f"'{field_example}'" if isinstance(field_example, str) else field_example field_text += f"{indent} Example: {example_text}\n" - if isclass(field_type) and issubclass(field_type, BaseModel): + if isclass(origin_type) and issubclass(origin_type, BaseModel): field_text += f"{indent} Details:\n" - for name, type_ in field_type.__annotations__.items(): + for name, type_ in get_type_hints(field_type).items(): field_text += generate_field_markdown(name, type_, field_type, depth + 2) return field_text @@ -833,7 +846,7 @@ def generate_text_documentation( str: Generated text documentation. """ documentation = "" - pyd_models = [(model, True) for model in pydantic_models] + pyd_models: list[tuple[type[BaseModel], bool]] = [(model, True) for model in pydantic_models] for model, add_prefix in pyd_models: if add_prefix: documentation += f"{model_prefix}: {model.__name__}\n" @@ -851,7 +864,7 @@ def generate_text_documentation( if isclass(model) and issubclass(model, BaseModel): documentation_fields = "" - for name, field_type in model.__annotations__.items(): + for name, field_type in get_type_hints(model).items(): # if name == "markdown_code_block": # continue if get_origin(field_type) == list: @@ -944,7 +957,7 @@ def generate_field_text( if isclass(field_type) and issubclass(field_type, BaseModel): field_text += f"{indent} Details:\n" - for name, type_ in field_type.__annotations__.items(): + for name, type_ in get_type_hints(field_type).items(): field_text += generate_field_text(name, type_, field_type, depth + 2) return field_text @@ -1164,7 +1177,7 @@ def create_dynamic_model_from_function(func: Callable[..., Any]): dynamic_fields[param.name] = ( param.annotation if param.annotation != inspect.Parameter.empty else str, default_value) # Creating the dynamic model - dynamic_model = create_model(f"{func.__name__}", **dynamic_fields) # type: ignore[call-overload] + dynamic_model = create_model(f"{func.__name__}", **dynamic_fields) for name, param_doc in param_docs: dynamic_model.model_fields[name].description = param_doc.description @@ -1228,9 +1241,6 @@ def map_grammar_names_to_pydantic_model_class(pydantic_model_list): return output -from enum import Enum - - def json_schema_to_python_types(schema): type_map = { "any": Any, @@ -1275,7 +1285,7 @@ def convert_dictionary_to_pydantic_model(dictionary: dict[str, Any], model_name: if items != {}: array = {"properties": items} array_type = convert_dictionary_to_pydantic_model(array, f"{model_name}_{field_name}_items") - fields[field_name] = (List[array_type], ...) # type: ignore[valid-type] + fields[field_name] = (List[array_type], ...) else: fields[field_name] = (list, ...) elif field_type == "object": @@ -1285,7 +1295,8 @@ def convert_dictionary_to_pydantic_model(dictionary: dict[str, Any], model_name: required = field_data.get("enum", []) for key, field in fields.items(): if key not in required: - fields[key] = (Optional[fields[key][0]], ...) + optional_type = fields[key][0] + fields[key] = (Optional[optional_type], ...) else: field_type = json_schema_to_python_types(field_type) fields[field_name] = (field_type, ...) @@ -1305,6 +1316,7 @@ def convert_dictionary_to_pydantic_model(dictionary: dict[str, Any], model_name: required = dictionary.get("required", []) for key, field in fields.items(): if key not in required: - fields[key] = (Optional[fields[key][0]], ...) + optional_type = fields[key][0] + fields[key] = (Optional[optional_type], ...) custom_model = create_model(model_name, **fields) return custom_model diff --git a/examples/pydantic_models_to_grammar_examples.py b/examples/pydantic_models_to_grammar_examples.py new file mode 100755 index 000000000..eb000d5cc --- /dev/null +++ b/examples/pydantic_models_to_grammar_examples.py @@ -0,0 +1,312 @@ +#!/usr/bin/env python3 + +"""Function calling example using pydantic models.""" + +from __future__ import annotations + +import argparse +import datetime +import json +import logging +import textwrap +import sys +from enum import Enum +from typing import Optional, Union + +import requests +from pydantic import BaseModel, Field +from pydantic_models_to_grammar import (add_run_method_to_dynamic_model, convert_dictionary_to_pydantic_model, + create_dynamic_model_from_function, generate_gbnf_grammar_and_documentation) + + +def create_completion(host, prompt, gbnf_grammar): + """Calls the /completion API on llama-server. + + See + https://github.com/ggerganov/llama.cpp/tree/HEAD/examples/server#api-endpoints + """ + print(f" Request:\n Grammar:\n{textwrap.indent(gbnf_grammar, ' ')}\n Prompt:\n{textwrap.indent(prompt.rstrip(), ' ')}") + headers = {"Content-Type": "application/json"} + data = {"prompt": prompt, "grammar": gbnf_grammar} + result = requests.post(f"http://{host}/completion", headers=headers, json=data).json() + assert data.get("error") is None, data + logging.info("Result: %s", result) + content = result["content"] + print(f" Model: {result['model']}") + print(f" Result:\n{textwrap.indent(json.dumps(json.loads(content), indent=2), ' ')}") + return content + + +# A function for the agent to send a message to the user. +class SendMessageToUser(BaseModel): + """Send a message to the User.""" + chain_of_thought: str = Field(..., description="Your chain of thought while sending the message.") + message: str = Field(..., description="Message you want to send to the user.") + + def run(self): + print(f"SendMessageToUser: {self.message}") + + +def example_rce(host): + """Minimal test case where the LLM call an arbitrary python function.""" + print("- example_rce") + tools = [SendMessageToUser] + gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation( + pydantic_model_list=tools, outer_object_name="function", + outer_object_content="function_parameters", model_prefix="Function", fields_prefix="Parameters") + system_message = "You are an advanced AI, tasked to assist the user by calling functions in JSON format. The following are the available functions and their parameters and types:\n\n" + documentation + user_message = "What is 42 * 42?" + prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant" + text = create_completion(host, prompt, gbnf_grammar) + json_data = json.loads(text) + tools_map = {tool.__name__:tool for tool in tools} + # This finds "SendMessageToUser": + tool = tools_map.get(json_data["function"]) + if not tool: + print(f"Error: unknown tool {json_data['function']}") + return 1 + tool(**json_data["function_parameters"]).run() + return 0 + + +# Enum for the calculator tool. +class MathOperation(Enum): + ADD = "add" + SUBTRACT = "subtract" + MULTIPLY = "multiply" + DIVIDE = "divide" + + +# Simple pydantic calculator tool for the agent that can add, subtract, +# multiply, and divide. Docstring and description of fields will be used in +# system prompt. +class Calculator(BaseModel): + """Perform a math operation on two numbers.""" + number_one: Union[int, float] = Field(..., description="First number.") + operation: MathOperation = Field(..., description="Math operation to perform.") + number_two: Union[int, float] = Field(..., description="Second number.") + + def run(self): + if self.operation == MathOperation.ADD: + return self.number_one + self.number_two + elif self.operation == MathOperation.SUBTRACT: + return self.number_one - self.number_two + elif self.operation == MathOperation.MULTIPLY: + return self.number_one * self.number_two + elif self.operation == MathOperation.DIVIDE: + return self.number_one / self.number_two + else: + raise ValueError("Unknown operation.") + + +def example_calculator(host): + """Have the LLM ask to get a calculation done. + + Here the grammar gets generated by passing the available function models to + generate_gbnf_grammar_and_documentation function. This also generates a + documentation usable by the LLM. + + pydantic_model_list is the list of pydantic models outer_object_name is an + optional name for an outer object around the actual model object. Like a + "function" object with "function_parameters" which contains the actual model + object. If None, no outer object will be generated outer_object_content is + the name of outer object content. + + model_prefix is the optional prefix for models in the documentation. (Default="Output Model") + fields_prefix is the prefix for the model fields in the documentation. (Default="Output Fields") + """ + print("- example_calculator") + tools = [SendMessageToUser, Calculator] + gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation( + pydantic_model_list=tools, outer_object_name="function", + outer_object_content="function_parameters", model_prefix="Function", fields_prefix="Parameters") + system_message = "You are an advanced AI, tasked to assist the user by calling functions in JSON format. The following are the available functions and their parameters and types:\n\n" + documentation + user_message1 = "What is 42 * 42?" + prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message1}<|im_end|>\n<|im_start|>assistant" + text = create_completion(host, prompt, gbnf_grammar) + json_data = json.loads(text) + expected = { + "function": "Calculator", + "function_parameters": { + "number_one": 42, + "operation": "multiply", + "number_two": 42 + } + } + if json_data != expected: + print(" Result is not as expected!") + tools_map = {tool.__name__:tool for tool in tools} + # This finds "Calculator": + tool = tools_map.get(json_data["function"]) + if not tool: + print(f"Error: unknown tool {json_data['function']}") + return 1 + result = tool(**json_data["function_parameters"]).run() + print(f" Call {json_data['function']} gave result {result}") + return 0 + + +class Category(Enum): + """The category of the book.""" + Fiction = "Fiction" + NonFiction = "Non-Fiction" + + +class Book(BaseModel): + """Represents an entry about a book.""" + title: str = Field(..., description="Title of the book.") + author: str = Field(..., description="Author of the book.") + published_year: Optional[int] = Field(..., description="Publishing year of the book.") + keywords: list[str] = Field(..., description="A list of keywords.") + category: Category = Field(..., description="Category of the book.") + summary: str = Field(..., description="Summary of the book.") + + +def example_struct(host): + """A example structured output based on pydantic models. + + The LLM will create an entry for a Book database out of an unstructured + text. We need no additional parameters other than our list of pydantic + models. + """ + print("- example_struct") + tools = [Book] + gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation(pydantic_model_list=tools) + system_message = "You are an advanced AI, tasked to create a dataset entry in JSON for a Book. The following is the expected output model:\n\n" + documentation + text = """The Feynman Lectures on Physics is a physics textbook based on some lectures by Richard Feynman, a Nobel laureate who has sometimes been called "The Great Explainer". The lectures were presented before undergraduate students at the California Institute of Technology (Caltech), during 1961–1963. The book's co-authors are Feynman, Robert B. Leighton, and Matthew Sands.""" + prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant" + text = create_completion(host, prompt, gbnf_grammar) + json_data = json.loads(text) + # In this case, there's no function nor function_parameters. + # Here the result will vary based on the LLM used. + keys = sorted(["title", "author", "published_year", "keywords", "category", "summary"]) + if keys != sorted(json_data.keys()): + print(f"Unexpected result: {sorted(json_data.keys())}") + return 1 + book = Book(**json_data) + print(f" As a Book object: %s" % book) + return 0 + + +def get_current_datetime(output_format: Optional[str] = None): + """Get the current date and time in the given format. + + Args: + output_format: formatting string for the date and time, defaults to '%Y-%m-%d %H:%M:%S' + """ + return datetime.datetime.now().strftime(output_format or "%Y-%m-%d %H:%M:%S") + + +# Example function to get the weather. +def get_current_weather(location, unit): + """Get the current weather in a given location""" + if "London" in location: + return json.dumps({"location": "London", "temperature": "42", "unit": unit.value}) + elif "New York" in location: + return json.dumps({"location": "New York", "temperature": "24", "unit": unit.value}) + elif "North Pole" in location: + return json.dumps({"location": "North Pole", "temperature": "-42", "unit": unit.value}) + return json.dumps({"location": location, "temperature": "unknown"}) + + +def example_concurrent(host): + """An example for parallel function calling with a Python function, a pydantic + function model and an OpenAI like function definition. + """ + print("- example_concurrent") + # Function definition in OpenAI style. + current_weather_tool = { + "type": "function", + "function": { + "name": "get_current_weather", + "description": "Get the current weather in a given location", + "parameters": { + "type": "object", + "properties": { + "location": { + "type": "string", + "description": "The city and state, e.g. San Francisco, CA", + }, + "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}, + }, + "required": ["location"], + }, + }, + } + # Convert OpenAI function definition into pydantic model. + current_weather_tool_model = convert_dictionary_to_pydantic_model(current_weather_tool) + # Add the actual function to a pydantic model. + current_weather_tool_model = add_run_method_to_dynamic_model(current_weather_tool_model, get_current_weather) + + # Convert normal Python function to a pydantic model. + current_datetime_model = create_dynamic_model_from_function(get_current_datetime) + + tools = [SendMessageToUser, Calculator, current_datetime_model, current_weather_tool_model] + gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation( + pydantic_model_list=tools, outer_object_name="function", + outer_object_content="params", model_prefix="Function", fields_prefix="Parameters", list_of_outputs=True) + system_message = "You are an advanced AI assistant. You are interacting with the user and with your environment by calling functions. You call functions by writing JSON objects, which represent specific function calls.\nBelow is a list of your available function calls:\n\n" + documentation + text = """Get the date and time, get the current weather in celsius in London and solve the following calculation: 42 * 42""" + prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant" + text = create_completion(host, prompt, gbnf_grammar) + json_data = json.loads(text) + expected = [ + { + "function": "get_current_datetime", + "params": { + "output_format": "%Y-%m-%d %H:%M:%S" + } + }, + { + "function": "get_current_weather", + "params": { + "location": "London", + "unit": "celsius" + } + }, + { + "function": "Calculator", + "params": { + "number_one": 42, + "operation": "multiply", + "number_two": 42 + } + } + ] + res = 0 + if json_data != expected: + print(" Result is not as expected!") + print(" This can happen on highly quantized models") + res = 1 + tools_map = {tool.__name__:tool for tool in tools} + for call in json_data: + tool = tools_map.get(call["function"]) + if not tool: + print(f"Error: unknown tool {call['function']}") + return 1 + result = tool(**call["params"]).run() + print(f" Call {call['function']} returned {result}") + # Should output something like this: + # Call get_current_datetime returned 2024-07-15 09:50:38 + # Call get_current_weather returned {"location": "London", "temperature": "42", "unit": "celsius"} + # Call Calculator returned 1764 + return res + + +def main(): + parser = argparse.ArgumentParser(description=sys.modules[__name__].__doc__) + parser.add_argument("--host", default="localhost:8080", help="llama.cpp server") + parser.add_argument("-v", "--verbose", action="store_true", help="enables logging") + args = parser.parse_args() + logging.basicConfig(level=logging.INFO if args.verbose else logging.ERROR) + ret = 0 + # Comment out below to only run the example you want. + ret = ret or example_rce(args.host) + ret = ret or example_calculator(args.host) + ret = ret or example_struct(args.host) + ret = ret or example_concurrent(args.host) + return ret + + +if __name__ == "__main__": + sys.exit(main()) diff --git a/examples/quantize-stats/quantize-stats.cpp b/examples/quantize-stats/quantize-stats.cpp index 746df8446..68cf8d359 100644 --- a/examples/quantize-stats/quantize-stats.cpp +++ b/examples/quantize-stats/quantize-stats.cpp @@ -154,7 +154,7 @@ static void test_roundtrip_on_chunk( } if (use_reference) { - qfns.from_float_reference(input_scratch, quantized_scratch, chunk_size); + qfns.from_float_ref(input_scratch, quantized_scratch, chunk_size); } else { qfns.from_float(input_scratch, quantized_scratch, chunk_size); } diff --git a/examples/quantize/README.md b/examples/quantize/README.md index b78ece4e7..5d1e11c67 100644 --- a/examples/quantize/README.md +++ b/examples/quantize/README.md @@ -4,7 +4,89 @@ You can also use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf- Note: It is synced from llama.cpp `main` every 6 hours. -## Llama 2 7B +Example usage: + +```bash +# obtain the official LLaMA model weights and place them in ./models +ls ./models +llama-2-7b tokenizer_checklist.chk tokenizer.model +# [Optional] for models using BPE tokenizers +ls ./models + vocab.json +# [Optional] for PyTorch .bin models like Mistral-7B +ls ./models + + +# install Python dependencies +python3 -m pip install -r requirements.txt + +# convert the model to ggml FP16 format +python3 convert_hf_to_gguf.py models/mymodel/ + +# quantize the model to 4-bits (using Q4_K_M method) +./llama-quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M + +# update the gguf filetype to current version if older version is now unsupported +./llama-quantize ./models/mymodel/ggml-model-Q4_K_M.gguf ./models/mymodel/ggml-model-Q4_K_M-v2.gguf COPY +``` + +Run the quantized model: + +```bash +# start inference on a gguf model +./llama-cli -m ./models/mymodel/ggml-model-Q4_K_M.gguf -cnv -p "You are a helpful assistant" +``` + +When running the larger models, make sure you have enough disk space to store all the intermediate files. + +## Memory/Disk Requirements + +As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same. + +| Model | Original size | Quantized size (Q4_0) | +|------:|--------------:|----------------------:| +| 7B | 13 GB | 3.9 GB | +| 13B | 24 GB | 7.8 GB | +| 30B | 60 GB | 19.5 GB | +| 65B | 120 GB | 38.5 GB | + +## Quantization + +Several quantization methods are supported. They differ in the resulting model disk size and inference speed. + +*(outdated)* + +| Model | Measure | F16 | Q4_0 | Q4_1 | Q5_0 | Q5_1 | Q8_0 | +|------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:| +| 7B | perplexity | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 | +| 7B | file size | 13.0G | 3.5G | 3.9G | 4.3G | 4.7G | 6.7G | +| 7B | ms/tok @ 4th | 127 | 55 | 54 | 76 | 83 | 72 | +| 7B | ms/tok @ 8th | 122 | 43 | 45 | 52 | 56 | 67 | +| 7B | bits/weight | 16.0 | 4.5 | 5.0 | 5.5 | 6.0 | 8.5 | +| 13B | perplexity | 5.2543 | 5.3860 | 5.3608 | 5.2856 | 5.2706 | 5.2548 | +| 13B | file size | 25.0G | 6.8G | 7.6G | 8.3G | 9.1G | 13G | +| 13B | ms/tok @ 4th | - | 103 | 105 | 148 | 160 | 131 | +| 13B | ms/tok @ 8th | - | 73 | 82 | 98 | 105 | 128 | +| 13B | bits/weight | 16.0 | 4.5 | 5.0 | 5.5 | 6.0 | 8.5 | + +- [k-quants](https://github.com/ggerganov/llama.cpp/pull/1684) +- recent k-quants improvements and new i-quants + - [#2707](https://github.com/ggerganov/llama.cpp/pull/2707) + - [#2807](https://github.com/ggerganov/llama.cpp/pull/2807) + - [#4773 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4773) + - [#4856 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4856) + - [#4861 - importance matrix](https://github.com/ggerganov/llama.cpp/pull/4861) + - [#4872 - MoE models](https://github.com/ggerganov/llama.cpp/pull/4872) + - [#4897 - 2-bit quantization](https://github.com/ggerganov/llama.cpp/pull/4897) + - [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930) + - [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957) + - [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969) + - [#4996 - k-qunats tuning](https://github.com/ggerganov/llama.cpp/pull/4996) + - [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060) + - [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196) + - [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361) + +**Llama 2 7B** | Quantization | Bits per Weight (BPW) | |--------------|-----------------------| @@ -18,7 +100,8 @@ Note: It is synced from llama.cpp `main` every 6 hours. | Q5_K_M | 5.68 | | Q6_K | 6.56 | -## Llama 2 13B +**Llama 2 13B** + Quantization | Bits per Weight (BPW) -- | -- Q2_K | 3.34 @@ -31,7 +114,7 @@ Q5_K_S | 5.51 Q5_K_M | 5.67 Q6_K | 6.56 -# Llama 2 70B +**Llama 2 70B** Quantization | Bits per Weight (BPW) -- | -- diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 76e2052d5..202346310 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -16,41 +16,44 @@ struct quant_option { }; static const std::vector QUANT_OPTIONS = { - { "Q4_0", LLAMA_FTYPE_MOSTLY_Q4_0, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, - { "Q4_1", LLAMA_FTYPE_MOSTLY_Q4_1, " 4.78G, +0.4511 ppl @ Llama-3-8B", }, - { "Q5_0", LLAMA_FTYPE_MOSTLY_Q5_0, " 5.21G, +0.1316 ppl @ Llama-3-8B", }, - { "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 5.65G, +0.1062 ppl @ Llama-3-8B", }, - { "IQ2_XXS",LLAMA_FTYPE_MOSTLY_IQ2_XXS," 2.06 bpw quantization", }, - { "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization", }, - { "IQ2_S", LLAMA_FTYPE_MOSTLY_IQ2_S, " 2.5 bpw quantization", }, - { "IQ2_M", LLAMA_FTYPE_MOSTLY_IQ2_M, " 2.7 bpw quantization", }, - { "IQ1_S", LLAMA_FTYPE_MOSTLY_IQ1_S, " 1.56 bpw quantization", }, - { "IQ1_M", LLAMA_FTYPE_MOSTLY_IQ1_M, " 1.75 bpw quantization", }, - { "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.96G, +3.5199 ppl @ Llama-3-8B", }, - { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.96G, +3.1836 ppl @ Llama-3-8B", }, - { "IQ3_XXS",LLAMA_FTYPE_MOSTLY_IQ3_XXS," 3.06 bpw quantization", }, - { "IQ3_S", LLAMA_FTYPE_MOSTLY_IQ3_S, " 3.44 bpw quantization", }, - { "IQ3_M", LLAMA_FTYPE_MOSTLY_IQ3_M, " 3.66 bpw quantization mix", }, - { "Q3_K", LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" }, - { "IQ3_XS", LLAMA_FTYPE_MOSTLY_IQ3_XS, " 3.3 bpw quantization", }, - { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 3.41G, +1.6321 ppl @ Llama-3-8B", }, - { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.74G, +0.6569 ppl @ Llama-3-8B", }, - { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 4.03G, +0.5562 ppl @ Llama-3-8B", }, - { "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.50 bpw non-linear quantization", }, - { "IQ4_XS", LLAMA_FTYPE_MOSTLY_IQ4_XS, " 4.25 bpw non-linear quantization", }, - { "Q4_K", LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", }, - { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 4.37G, +0.2689 ppl @ Llama-3-8B", }, - { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 4.58G, +0.1754 ppl @ Llama-3-8B", }, - { "Q5_K", LLAMA_FTYPE_MOSTLY_Q5_K_M, "alias for Q5_K_M", }, - { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 5.21G, +0.1049 ppl @ Llama-3-8B", }, - { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 5.33G, +0.0569 ppl @ Llama-3-8B", }, - { "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 6.14G, +0.0217 ppl @ Llama-3-8B", }, - { "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 7.96G, +0.0026 ppl @ Llama-3-8B", }, - { "F16", LLAMA_FTYPE_MOSTLY_F16, "14.00G, +0.0020 ppl @ Mistral-7B", }, - { "BF16", LLAMA_FTYPE_MOSTLY_BF16, "14.00G, -0.0050 ppl @ Mistral-7B", }, - { "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", }, + { "Q4_0", LLAMA_FTYPE_MOSTLY_Q4_0, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, + { "Q4_1", LLAMA_FTYPE_MOSTLY_Q4_1, " 4.78G, +0.4511 ppl @ Llama-3-8B", }, + { "Q5_0", LLAMA_FTYPE_MOSTLY_Q5_0, " 5.21G, +0.1316 ppl @ Llama-3-8B", }, + { "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 5.65G, +0.1062 ppl @ Llama-3-8B", }, + { "IQ2_XXS", LLAMA_FTYPE_MOSTLY_IQ2_XXS, " 2.06 bpw quantization", }, + { "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization", }, + { "IQ2_S", LLAMA_FTYPE_MOSTLY_IQ2_S, " 2.5 bpw quantization", }, + { "IQ2_M", LLAMA_FTYPE_MOSTLY_IQ2_M, " 2.7 bpw quantization", }, + { "IQ1_S", LLAMA_FTYPE_MOSTLY_IQ1_S, " 1.56 bpw quantization", }, + { "IQ1_M", LLAMA_FTYPE_MOSTLY_IQ1_M, " 1.75 bpw quantization", }, + { "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.96G, +3.5199 ppl @ Llama-3-8B", }, + { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.96G, +3.1836 ppl @ Llama-3-8B", }, + { "IQ3_XXS", LLAMA_FTYPE_MOSTLY_IQ3_XXS, " 3.06 bpw quantization", }, + { "IQ3_S", LLAMA_FTYPE_MOSTLY_IQ3_S, " 3.44 bpw quantization", }, + { "IQ3_M", LLAMA_FTYPE_MOSTLY_IQ3_M, " 3.66 bpw quantization mix", }, + { "Q3_K", LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" }, + { "IQ3_XS", LLAMA_FTYPE_MOSTLY_IQ3_XS, " 3.3 bpw quantization", }, + { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 3.41G, +1.6321 ppl @ Llama-3-8B", }, + { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.74G, +0.6569 ppl @ Llama-3-8B", }, + { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 4.03G, +0.5562 ppl @ Llama-3-8B", }, + { "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.50 bpw non-linear quantization", }, + { "IQ4_XS", LLAMA_FTYPE_MOSTLY_IQ4_XS, " 4.25 bpw non-linear quantization", }, + { "Q4_K", LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", }, + { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 4.37G, +0.2689 ppl @ Llama-3-8B", }, + { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 4.58G, +0.1754 ppl @ Llama-3-8B", }, + { "Q5_K", LLAMA_FTYPE_MOSTLY_Q5_K_M, "alias for Q5_K_M", }, + { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 5.21G, +0.1049 ppl @ Llama-3-8B", }, + { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 5.33G, +0.0569 ppl @ Llama-3-8B", }, + { "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 6.14G, +0.0217 ppl @ Llama-3-8B", }, + { "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 7.96G, +0.0026 ppl @ Llama-3-8B", }, + { "Q4_0_4_4", LLAMA_FTYPE_MOSTLY_Q4_0_4_4, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, + { "Q4_0_4_8", LLAMA_FTYPE_MOSTLY_Q4_0_4_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, + { "Q4_0_8_8", LLAMA_FTYPE_MOSTLY_Q4_0_8_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, + { "F16", LLAMA_FTYPE_MOSTLY_F16, "14.00G, +0.0020 ppl @ Mistral-7B", }, + { "BF16", LLAMA_FTYPE_MOSTLY_BF16, "14.00G, -0.0050 ppl @ Mistral-7B", }, + { "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", }, // Note: Ensure COPY comes after F32 to avoid ftype 0 from matching. - { "COPY", LLAMA_FTYPE_ALL_F32, "only copy tensors, no quantizing", }, + { "COPY", LLAMA_FTYPE_ALL_F32, "only copy tensors, no quantizing", }, }; static const char * const LLM_KV_QUANTIZE_IMATRIX_FILE = "quantize.imatrix.file"; @@ -88,7 +91,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp } // usage: -// ./quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads] +// ./llama-quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads] // [[noreturn]] static void usage(const char * executable) { @@ -101,7 +104,7 @@ static void usage(const char * executable) { printf(" --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n"); printf(" --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor\n"); printf(" --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n"); - printf(" --keep-split: will generate quatized model in the same shards as input"); + printf(" --keep-split: will generate quantized model in the same shards as input\n"); printf(" --override-kv KEY=TYPE:VALUE\n"); printf(" Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n"); printf("Note: --include-weights and --exclude-weights cannot be used together\n"); diff --git a/examples/regex-to-grammar.py b/examples/regex_to_grammar.py similarity index 100% rename from examples/regex-to-grammar.py rename to examples/regex_to_grammar.py diff --git a/examples/retrieval/retrieval.cpp b/examples/retrieval/retrieval.cpp index 94fc7e54b..54bdf5d1f 100644 --- a/examples/retrieval/retrieval.cpp +++ b/examples/retrieval/retrieval.cpp @@ -148,11 +148,12 @@ int main(int argc, char ** argv) { 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); + llama_init_result llama_init = llama_init_from_gpt_params(params); + + llama_model * model = llama_init.model; + llama_context * ctx = llama_init.context; + if (model == NULL) { fprintf(stderr, "%s: error: unable to load model\n", __func__); return 1; @@ -252,6 +253,8 @@ int main(int argc, char ** argv) { chunks[i].tokens.clear(); } + struct llama_batch query_batch = llama_batch_init(n_batch, 0, 1); + // start loop, receive query and return top k similar chunks based on cosine similarity std::string query; while (true) { @@ -259,7 +262,6 @@ int main(int argc, char ** argv) { 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); @@ -292,6 +294,7 @@ int main(int argc, char ** argv) { } // clean up + llama_batch_free(query_batch); llama_print_timings(ctx); llama_free(ctx); llama_free_model(model); diff --git a/examples/rpc/README.md b/examples/rpc/README.md index e1da801f2..adedc8909 100644 --- a/examples/rpc/README.md +++ b/examples/rpc/README.md @@ -1,5 +1,9 @@ ## Overview +> [!IMPORTANT] +> This example and the RPC backend are currently in a proof-of-concept development stage. As such, the functionality is fragile and +> insecure. **Never run the RPC server on an open network or in a sensitive environment!** + The `rpc-server` allows running `ggml` backend on a remote host. The RPC backend communicates with one or several instances of `rpc-server` and offloads computations to them. This can be used for distributed LLM inference with `llama.cpp` in the following way: diff --git a/examples/rpc/rpc-server.cpp b/examples/rpc/rpc-server.cpp index 7c15d2aa4..6342e6488 100644 --- a/examples/rpc/rpc-server.cpp +++ b/examples/rpc/rpc-server.cpp @@ -16,7 +16,7 @@ #include struct rpc_server_params { - std::string host = "0.0.0.0"; + std::string host = "127.0.0.1"; int port = 50052; size_t backend_mem = 0; }; @@ -114,6 +114,17 @@ int main(int argc, char * argv[]) { fprintf(stderr, "Invalid parameters\n"); return 1; } + + if (params.host != "127.0.0.1") { + fprintf(stderr, "\n"); + fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n"); + fprintf(stderr, "WARNING: Host ('%s') is != '127.0.0.1'\n", params.host.c_str()); + fprintf(stderr, " Never expose the RPC server to an open network!\n"); + fprintf(stderr, " This is an experimental feature and is not secure!\n"); + fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n"); + fprintf(stderr, "\n"); + } + ggml_backend_t backend = create_backend(); if (!backend) { fprintf(stderr, "Failed to create backend\n"); diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp index 974dc3c3e..906a1970b 100644 --- a/examples/save-load-state/save-load-state.cpp +++ b/examples/save-load-state/save-load-state.cpp @@ -28,10 +28,11 @@ int main(int argc, char ** argv) { std::string result2; // init - llama_model * model; - llama_context * ctx; + llama_init_result llama_init = llama_init_from_gpt_params(params); + + llama_model * model = llama_init.model; + llama_context * ctx = llama_init.context; - std::tie(model, ctx) = llama_init_from_gpt_params(params); if (model == nullptr || ctx == nullptr) { fprintf(stderr, "%s : failed to init\n", __func__); return 1; @@ -47,7 +48,7 @@ int main(int argc, char ** argv) { // save state (rng, logits, embedding and kv_cache) to file { std::vector state_mem(llama_state_get_size(ctx)); - const size_t written = llama_state_get_data(ctx, state_mem.data()); + const size_t written = llama_state_get_data(ctx, state_mem.data(), state_mem.size()); FILE *fp_write = fopen("dump_state.bin", "wb"); fwrite(state_mem.data(), 1, written, fp_write); @@ -99,13 +100,16 @@ int main(int argc, char ** argv) { // load state (rng, logits, embedding and kv_cache) from file { - std::vector state_mem(llama_state_get_size(ctx2)); + std::vector state_mem; FILE * fp_read = fopen("dump_state.bin", "rb"); + fseek(fp_read, 0, SEEK_END); + state_mem.resize(ftell(fp_read)); + fseek(fp_read, 0, SEEK_SET); const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read); fclose(fp_read); - if (read != llama_state_set_data(ctx2, state_mem.data())) { + if (read != llama_state_set_data(ctx2, state_mem.data(), state_mem.size())) { fprintf(stderr, "\n%s : failed to read state\n", __func__); llama_free(ctx2); llama_free_model(model); @@ -159,13 +163,16 @@ int main(int argc, char ** argv) { // load state (rng, logits, embedding and kv_cache) from file { - std::vector state_mem(llama_state_get_size(ctx3)); + std::vector state_mem; FILE * fp_read = fopen("dump_state.bin", "rb"); + fseek(fp_read, 0, SEEK_END); + state_mem.resize(ftell(fp_read)); + fseek(fp_read, 0, SEEK_SET); const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read); fclose(fp_read); - if (read != llama_state_set_data(ctx3, state_mem.data())) { + if (read != llama_state_set_data(ctx3, state_mem.data(), state_mem.size())) { fprintf(stderr, "\n%s : failed to read state\n", __func__); llama_free(ctx3); llama_free_model(model); @@ -182,7 +189,7 @@ int main(int argc, char ** argv) { { // save kv of seq 0 std::vector seq_store(llama_state_seq_get_size(ctx3, 0)); - const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), 0); + const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), seq_store.size(), 0); if (ncopy != seq_store.size()) { fprintf(stderr, "\n%s : seq copy data length %zd does not match expected length %zd\n", __func__, ncopy, seq_store.size()); llama_free(ctx3); @@ -196,7 +203,7 @@ int main(int argc, char ** argv) { fprintf(stderr, "%s : kv cache cleared\n", __func__); // restore kv into seq 1 - const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), 1); + const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), seq_store.size(), 1); if (nset != seq_store.size()) { fprintf(stderr, "\n%s : seq set data length %zd does not match expected length %zd\n", __func__, nset, seq_store.size()); llama_free(ctx3); diff --git a/examples/server/README.md b/examples/server/README.md index 4fab006bb..805e05b4a 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -5,7 +5,7 @@ Fast, lightweight, pure C/C++ HTTP server based on [httplib](https://github.com/ Set of LLM REST APIs and a simple web front end to interact with llama.cpp. **Features:** - * LLM inference of F16 and quantum models on GPU and CPU + * LLM inference of F16 and quantized models on GPU and CPU * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes * Parallel decoding with multi-user support * Continuous batching @@ -15,69 +15,283 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp. The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggerganov/llama.cpp/issues/4216). -**Command line options:** +## Usage -- `-v`, `--verbose`: Enable verbose server output. When using the `/completion` endpoint, this includes the tokenized prompt, the full request and the full response. -- `-t N`, `--threads N`: Set the number of threads to use by CPU layers during generation. Not used by model layers that are offloaded to GPU. This option has no effect when using the maximum number of GPU layers. Default: `std::thread::hardware_concurrency()` (number of CPU cores). -- `-tb N, --threads-batch N`: Set the number of threads to use by CPU layers during batch and prompt processing (>= 32 tokens). This option has no effect if a GPU is available. Default: `--threads`. -- `--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. 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 GPU support, 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. -- `-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. -- `-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` -- `--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. -- `--numa STRATEGY`: Attempt one of the below optimization strategies that may help on some NUMA systems -- `--numa distribute`: Spread execution evenly over all nodes -- `--numa isolate`: Only spawn threads on CPUs on the node that execution started on -- `--numa numactl`: Use the CPU map provided by numactl. If run without this previously, it is recommended to drop the system page cache before using this. See https://github.com/ggerganov/llama.cpp/issues/1437 -- `--numa`: Attempt optimizations that may help on some NUMA systems. -- `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains. -- `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation. -- `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600` -- `--host`: Set the hostname or ip address to listen. Default `127.0.0.1` -- `--port`: Set the port to listen. Default: `8080` -- `--path`: Path from which to serve static files. Default: disabled -- `--api-key`: Set an api key for request authorization. By default, the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys. -- `--api-key-file`: Path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`s. -- `--embeddings`: Enable embedding vector output and the OAI compatible endpoint /v1/embeddings. Physical batch size (`--ubatch-size`) must be carefully defined. Default: disabled -- `-np N`, `--parallel N`: Set the number of slots for process requests. Default: `1`. Values > 1 will allow for higher throughput with multiple parallel requests but the results will **not** be deterministic due to differences in rounding error. -- `-cb`, `--cont-batching`: Enable continuous batching (a.k.a dynamic batching). Default: disabled -- `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load a system prompt (initial prompt of all slots). This is useful for chat applications. [See more](#change-system-prompt-on-runtime) -- `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA. -- `--grp-attn-n`: Set the group attention factor to extend context size through self-extend. Used together with group attention width `--grp-attn-w`. Default: `1`, which is disabled. -- `--grp-attn-w`: Set the group attention width to extend context size through self-extend. Used together with group attention factor `--grp-attn-n`. Default: `512` -- `-n N, --n-predict N`: Set the maximum tokens to predict. Default: `-1` -- `--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 -- `--slot-save-path PATH`: Specifies the path where the state of slots (the prompt cache) can be stored. If not provided, the slot management endpoints will be 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, not to `llama.log`. Default: enabled -- `--log-format FORMAT`: Define the log output to FORMAT: json or text Default: `json` -- `--rope-scaling` : RoPE scaling method. Defaults to linear unless otherwise specified by the model. Options are `none`, `linear`, `yarn` -- `--rope-freq-base N` : RoPE frequency base (default: loaded from model) -- `--rope-freq-scale N`: RoPE frequency scaling factor, expands context by a factor of 1/N (e.g. 0.25) -- `--yarn-ext-factor N` : YaRN: extrapolation mix factor (Default: 1.0, 0.0 = full interpolation) -- `--yarn-attn-factor N` : YaRN: scale sqrt(t) or attention magnitude (default: 1.0) -- `--yarn-beta-slow N`: YaRN: High correction dim or alpha (default: 1.0) -- `--yarn-beta-fast N`: YaRN: low correction dim or beta (default: 32.0) -- `--pooling` : Pooling type for embeddings, use model default if unspecified. Options are `none`, `mean`, `cls` -- `-dt N`, `--defrag-thold N`: KV cache defragmentation threshold (default: -1.0, < 0 = disabled) -- `-fa`, `--flash-attn` : enable flash attention (default: disabled). -- `-ctk TYPE`, `--cache-type-k TYPE` : KV cache data type for K (default: `f16`, options `f32`, `f16`, `q8_0`, `q4_0`, `q4_1`, `iq4_nl`, `q5_0`, or `q5_1`) -- `-ctv TYPE`, `--cache-type-v TYPE` : KV cache type for V (default `f16`, see `-ctk` for options) -- `--spm-infill` : Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. +``` +usage: ./llama-server [options] -**If compiled with `LLAMA_SERVER_SSL=ON`** -- `--ssl-key-file FNAME`: path to file a PEM-encoded SSL private key -- `--ssl-cert-file FNAME`: path to file a PEM-encoded SSL certificate +general: + + -h, --help, --usage print usage and exit + --version show version and build info + -v, --verbose print verbose information + --verbosity N set specific verbosity level (default: 0) + --verbose-prompt print a verbose prompt before generation (default: false) + --no-display-prompt don't print prompt at generation (default: false) + -co, --color colorise output to distinguish prompt and user input from generations (default: false) + -s, --seed SEED RNG seed (default: -1, use random seed for < 0) + -t, --threads N number of threads to use during generation (default: 8) + -tb, --threads-batch N number of threads to use during batch and prompt processing (default: same as --threads) + -td, --threads-draft N number of threads to use during generation (default: same as --threads) + -tbd, --threads-batch-draft N number of threads to use during batch and prompt processing (default: same as --threads-draft) + --draft N number of tokens to draft for speculative decoding (default: 5) + -ps, --p-split N speculative decoding split probability (default: 0.1) + -lcs, --lookup-cache-static FNAME + path to static lookup cache to use for lookup decoding (not updated by generation) + -lcd, --lookup-cache-dynamic FNAME + path to dynamic lookup cache to use for lookup decoding (updated by generation) + -c, --ctx-size N size of the prompt context (default: 0, 0 = loaded from model) + -n, --predict N number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled) + -b, --batch-size N logical maximum batch size (default: 2048) + -ub, --ubatch-size N physical maximum batch size (default: 512) + --keep N number of tokens to keep from the initial prompt (default: 0, -1 = all) + --chunks N max number of chunks to process (default: -1, -1 = all) + -fa, --flash-attn enable Flash Attention (default: disabled) + -p, --prompt PROMPT prompt to start generation with + in conversation mode, this will be used as system prompt + (default: '') + -f, --file FNAME a file containing the prompt (default: none) + --in-file FNAME an input file (repeat to specify multiple files) + -bf, --binary-file FNAME binary file containing the prompt (default: none) + -e, --escape process escapes sequences (\n, \r, \t, \', \", \\) (default: true) + --no-escape do not process escape sequences + -ptc, --print-token-count N print token count every N tokens (default: -1) + --prompt-cache FNAME file to cache prompt state for faster startup (default: none) + --prompt-cache-all if specified, saves user input and generations to cache as well + not supported with --interactive or other interactive options + --prompt-cache-ro if specified, uses the prompt cache but does not update it + -r, --reverse-prompt PROMPT halt generation at PROMPT, return control in interactive mode + can be specified more than once for multiple prompts + -sp, --special special tokens output enabled (default: false) + -cnv, --conversation run in conversation mode, does not print special tokens and suffix/prefix + if suffix/prefix are not specified, default chat template will be used + (default: false) + -i, --interactive run in interactive mode (default: false) + -if, --interactive-first run in interactive mode and wait for input right away (default: false) + -mli, --multiline-input allows you to write or paste multiple lines without ending each in '\' + --in-prefix-bos prefix BOS to user inputs, preceding the `--in-prefix` string + --in-prefix STRING string to prefix user inputs with (default: empty) + --in-suffix STRING string to suffix after user inputs with (default: empty) + --spm-infill use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) + +sampling: + + --samplers SAMPLERS samplers that will be used for generation in the order, separated by ';' + (default: top_k;tfs_z;typical_p;top_p;min_p;temperature) + --sampling-seq SEQUENCE simplified sequence for samplers that will be used (default: kfypmt) + --ignore-eos ignore end of stream token and continue generating (implies --logit-bias EOS-inf) + --penalize-nl penalize newline tokens (default: false) + --temp N temperature (default: 0.8) + --top-k N top-k sampling (default: 40, 0 = disabled) + --top-p N top-p sampling (default: 0.9, 1.0 = disabled) + --min-p N min-p sampling (default: 0.1, 0.0 = disabled) + --tfs N tail free sampling, parameter z (default: 1.0, 1.0 = disabled) + --typical N locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) + --repeat-last-n N last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) + --repeat-penalty N penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) + --presence-penalty N repeat alpha presence penalty (default: 0.0, 0.0 = disabled) + --frequency-penalty N repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) + --dynatemp-range N dynamic temperature range (default: 0.0, 0.0 = disabled) + --dynatemp-exp N dynamic temperature exponent (default: 1.0) + --mirostat N use Mirostat sampling. + Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used. + (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) + --mirostat-lr N Mirostat learning rate, parameter eta (default: 0.1) + --mirostat-ent N Mirostat target entropy, parameter tau (default: 5.0) + -l TOKEN_ID(+/-)BIAS modifies the likelihood of token appearing in the completion, + i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello', + or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' + --cfg-negative-prompt PROMPT + negative prompt to use for guidance (default: '') + --cfg-negative-prompt-file FNAME + negative prompt file to use for guidance + --cfg-scale N strength of guidance (default: 1.0, 1.0 = disable) + --chat-template JINJA_TEMPLATE + set custom jinja chat template (default: template taken from model's metadata) + if suffix/prefix are specified, template will be disabled + only commonly used templates are accepted: + https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template + +grammar: + + --grammar GRAMMAR BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') + --grammar-file FNAME file to read grammar from + -j, --json-schema SCHEMA JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object + For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead + +embedding: + + --pooling {none,mean,cls,last} + pooling type for embeddings, use model default if unspecified + --attention {causal,non-causal} + attention type for embeddings, use model default if unspecified + +context hacking: + + --rope-scaling {none,linear,yarn} + RoPE frequency scaling method, defaults to linear unless specified by the model + --rope-scale N RoPE context scaling factor, expands context by a factor of N + --rope-freq-base N RoPE base frequency, used by NTK-aware scaling (default: loaded from model) + --rope-freq-scale N RoPE frequency scaling factor, expands context by a factor of 1/N + --yarn-orig-ctx N YaRN: original context size of model (default: 0 = model training context size) + --yarn-ext-factor N YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation) + --yarn-attn-factor N YaRN: scale sqrt(t) or attention magnitude (default: 1.0) + --yarn-beta-slow N YaRN: high correction dim or alpha (default: 1.0) + --yarn-beta-fast N YaRN: low correction dim or beta (default: 32.0) + -gan, --grp-attn-n N group-attention factor (default: 1) + -gaw, --grp-attn-w N group-attention width (default: 512.0) + -dkvc, --dump-kv-cache verbose print of the KV cache + -nkvo, --no-kv-offload disable KV offload + -ctk, --cache-type-k TYPE KV cache data type for K (default: f16) + -ctv, --cache-type-v TYPE KV cache data type for V (default: f16) + +perplexity: + + --all-logits return logits for all tokens in the batch (default: false) + --hellaswag compute HellaSwag score over random tasks from datafile supplied with -f + --hellaswag-tasks N number of tasks to use when computing the HellaSwag score (default: 400) + --winogrande compute Winogrande score over random tasks from datafile supplied with -f + --winogrande-tasks N number of tasks to use when computing the Winogrande score (default: 0) + --multiple-choice compute multiple choice score over random tasks from datafile supplied with -f + --multiple-choice-tasks N + number of tasks to use when computing the multiple choice score (default: 0) + --kl-divergence computes KL-divergence to logits provided via --kl-divergence-base + --ppl-stride N stride for perplexity calculation (default: 0) + --ppl-output-type {0,1} output type for perplexity calculation (default: 0) + +parallel: + + -dt, --defrag-thold N KV cache defragmentation threshold (default: -1.0, < 0 - disabled) + -np, --parallel N number of parallel sequences to decode (default: 1) + -ns, --sequences N number of sequences to decode (default: 1) + -cb, --cont-batching enable continuous batching (a.k.a dynamic batching) (default: enabled) + +multi-modality: + + --mmproj FILE path to a multimodal projector file for LLaVA. see examples/llava/README.md + --image FILE path to an image file. use with multimodal models. Specify multiple times for batching + +backend: + + --rpc SERVERS comma separated list of RPC servers + --mlock force system to keep model in RAM rather than swapping or compressing + --no-mmap do not memory-map model (slower load but may reduce pageouts if not using mlock) + --numa TYPE attempt optimizations that help on some NUMA systems + - distribute: spread execution evenly over all nodes + - isolate: only spawn threads on CPUs on the node that execution started on + - numactl: use the CPU map provided by numactl + if run without this previously, it is recommended to drop the system page cache before using this + see https://github.com/ggerganov/llama.cpp/issues/1437 + +model: + + --check-tensors check model tensor data for invalid values (default: false) + --override-kv KEY=TYPE:VALUE + advanced option to override model metadata by key. may be specified multiple times. + types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false + --lora FNAME apply LoRA adapter (implies --no-mmap) + --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap) + --lora-base FNAME optional model to use as a base for the layers modified by the LoRA adapter + --control-vector FNAME add a control vector + note: this argument can be repeated to add multiple control vectors + --control-vector-scaled FNAME SCALE + add a control vector with user defined scaling SCALE + note: this argument can be repeated to add multiple scaled control vectors + --control-vector-layer-range START END + layer range to apply the control vector(s) to, start and end inclusive + -m, --model FNAME model path (default: models/$filename with filename from --hf-file + or --model-url if set, otherwise models/7B/ggml-model-f16.gguf) + -md, --model-draft FNAME draft model for speculative decoding (default: unused) + -mu, --model-url MODEL_URL model download url (default: unused) + -hfr, --hf-repo REPO Hugging Face model repository (default: unused) + -hff, --hf-file FILE Hugging Face model file (default: unused) + -hft, --hf-token TOKEN Hugging Face access token (default: value from HF_TOKEN environment variable) + +server: + + --host HOST ip address to listen (default: 127.0.0.1) + --port PORT port to listen (default: 8080) + --path PATH path to serve static files from (default: ) + --embedding(s) restrict to only support embedding use case; use only with dedicated embedding models (default: disabled) + --api-key KEY API key to use for authentication (default: none) + --api-key-file FNAME path to file containing API keys (default: none) + --ssl-key-file FNAME path to file a PEM-encoded SSL private key + --ssl-cert-file FNAME path to file a PEM-encoded SSL certificate + --timeout N server read/write timeout in seconds (default: 600) + --threads-http N number of threads used to process HTTP requests (default: -1) + --system-prompt-file FNAME + set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications + --log-format {text,json} + log output format: json or text (default: json) + --metrics enable prometheus compatible metrics endpoint (default: disabled) + --no-slots disables slots monitoring endpoint (default: enabled) + --slot-save-path PATH path to save slot kv cache (default: disabled) + --chat-template JINJA_TEMPLATE + set custom jinja chat template (default: template taken from model's metadata) + only commonly used templates are accepted: + https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template + -sps, --slot-prompt-similarity SIMILARITY + how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled) + --lora-init-without-apply + load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) + +logging: + + --simple-io use basic IO for better compatibility in subprocesses and limited consoles + -ld, --logdir LOGDIR path under which to save YAML logs (no logging if unset) + --log-test Run simple logging test + --log-disable Disable trace logs + --log-enable Enable trace logs + --log-file FNAME Specify a log filename (without extension) + --log-new Create a separate new log file on start. Each log file will have unique name: "..log" + --log-append Don't truncate the old log file. +``` + +Available environment variables (if specified, these variables will override parameters specified in arguments): + +- `LLAMA_CACHE`: cache directory, used by `--hf-repo` +- `HF_TOKEN`: Hugging Face access token, used when accessing a gated model with `--hf-repo` +- `LLAMA_ARG_MODEL`: equivalent to `-m` +- `LLAMA_ARG_MODEL_URL`: equivalent to `-mu` +- `LLAMA_ARG_MODEL_ALIAS`: equivalent to `-a` +- `LLAMA_ARG_HF_REPO`: equivalent to `--hf-repo` +- `LLAMA_ARG_HF_FILE`: equivalent to `--hf-file` +- `LLAMA_ARG_THREADS`: equivalent to `-t` +- `LLAMA_ARG_CTX_SIZE`: equivalent to `-c` +- `LLAMA_ARG_N_PARALLEL`: equivalent to `-np` +- `LLAMA_ARG_BATCH`: equivalent to `-b` +- `LLAMA_ARG_UBATCH`: equivalent to `-ub` +- `LLAMA_ARG_N_GPU_LAYERS`: equivalent to `-ngl` +- `LLAMA_ARG_THREADS_HTTP`: equivalent to `--threads-http` +- `LLAMA_ARG_CHAT_TEMPLATE`: equivalent to `--chat-template` +- `LLAMA_ARG_N_PREDICT`: equivalent to `-n` +- `LLAMA_ARG_ENDPOINT_METRICS`: if set to `1`, it will enable metrics endpoint (equivalent to `--metrics`) +- `LLAMA_ARG_ENDPOINT_SLOTS`: if set to `0`, it will **disable** slots endpoint (equivalent to `--no-slots`). This feature is enabled by default. +- `LLAMA_ARG_EMBEDDINGS`: if set to `1`, it will enable embeddings endpoint (equivalent to `--embeddings`) +- `LLAMA_ARG_FLASH_ATTN`: if set to `1`, it will enable flash attention (equivalent to `-fa`) +- `LLAMA_ARG_CONT_BATCHING`: if set to `0`, it will **disable** continuous batching (equivalent to `--no-cont-batching`). This feature is enabled by default. +- `LLAMA_ARG_DEFRAG_THOLD`: equivalent to `-dt` +- `LLAMA_ARG_HOST`: equivalent to `--host` +- `LLAMA_ARG_PORT`: equivalent to `--port` + +Example usage of docker compose with environment variables: + +```yml +services: + llamacpp-server: + image: ghcr.io/ggerganov/llama.cpp:server + ports: + - 8080:8080 + volumes: + - ./models:/models + environment: + # alternatively, you can use "LLAMA_ARG_MODEL_URL" to download the model + LLAMA_ARG_MODEL: /models/my_model.gguf + LLAMA_ARG_CTX_SIZE: 4096 + LLAMA_ARG_N_PARALLEL: 2 + LLAMA_ARG_ENDPOINT_METRICS: 1 # to disable, either remove or set to 0 + LLAMA_ARG_PORT: 8080 +``` ## Build @@ -199,16 +413,18 @@ node index.js ## API Endpoints -- **GET** `/health`: Returns the current state of the server: - - 503 -> `{"status": "loading model"}` if the model is still being loaded. - - 500 -> `{"status": "error"}` if the model failed to load. - - 200 -> `{"status": "ok", "slots_idle": 1, "slots_processing": 2 }` if the model is successfully loaded and the server is ready for further requests mentioned below. - - 200 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slots are currently available. - - 503 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if the query parameter `fail_on_no_slot` is provided and no slots are currently available. +### GET `/health`: Returns heath check result - If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set. +**Response format** -- **POST** `/completion`: Given a `prompt`, it returns the predicted completion. +- HTTP status code 503 + - Body: `{"error": {"code": 503, "message": "Loading model", "type": "unavailable_error"}}` + - Explanation: the model is still being loaded. +- HTTP status code 200 + - Body: `{"status": "ok" }` + - Explanation: the model is successfully loaded and the server is ready. + +### POST `/completion`: Given a `prompt`, it returns the predicted completion. *Options:* @@ -232,7 +448,7 @@ node index.js `n_predict`: Set the maximum number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. Default: `-1`, where `-1` is infinity. - `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded. + `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded. The number excludes the BOS token. By default, this value is set to `0`, meaning no tokens are kept. Use `-1` to retain all tokens from the prompt. `stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`. @@ -286,7 +502,7 @@ node index.js `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values. -### Result JSON +**Response format** - Note: When using streaming mode (`stream`), only `content` and `stop` will be returned until end of completion. @@ -325,7 +541,7 @@ Notice that each `probs` is an array of length `n_probs`. - `tokens_evaluated`: Number of tokens evaluated in total from the prompt - `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`) -- **POST** `/tokenize`: Tokenize a given text. +### POST `/tokenize`: Tokenize a given text *Options:* @@ -333,13 +549,15 @@ Notice that each `probs` is an array of length `n_probs`. `add_special`: Boolean indicating if special tokens, i.e. `BOS`, should be inserted. Default: `false` -- **POST** `/detokenize`: Convert tokens to text. +### POST `/detokenize`: Convert tokens to text *Options:* `tokens`: Set the tokens to detokenize. -- **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does. +### POST `/embedding`: Generate embedding of a given text + +The same as [the embedding example](../embedding) does. *Options:* @@ -347,7 +565,9 @@ Notice that each `probs` is an array of length `n_probs`. `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `content`. You can determine the place of the image in the content as in the following: `Image: [img-21].\nCaption: This is a picture of a house`. In this case, `[img-21]` will be replaced by the embeddings of the image with id `21` in the following `image_data` array: `{..., "image_data": [{"data": "", "id": 21}]}`. Use `image_data` only with multimodal models, e.g., LLaVA. -- **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream. +### POST `/infill`: For code infilling. + +Takes a prefix and a suffix and returns the predicted completion as stream. *Options:* @@ -359,14 +579,15 @@ Notice that each `probs` is an array of length `n_probs`. - **GET** `/props`: Return current server settings. -### Result JSON +**Response format** ```json { "assistant_name": "", "user_name": "", "default_generation_settings": { ... }, - "total_slots": 1 + "total_slots": 1, + "chat_template": "" } ``` @@ -374,8 +595,11 @@ Notice that each `probs` is an array of length `n_probs`. - `user_name` - the required anti-prompt to generate the prompt in case you have specified a system prompt for all slots. - `default_generation_settings` - the default generation settings for the `/completion` endpoint, which 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) +- `chat_template` - the model's original Jinja2 prompt template -- **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. +### 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 models with a [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, the ChatML template will be used. *Options:* @@ -427,7 +651,7 @@ Notice that each `probs` is an array of length `n_probs`. }' ``` -- **POST** `/v1/embeddings`: OpenAI-compatible embeddings API. +### POST `/v1/embeddings`: OpenAI-compatible embeddings API *Options:* @@ -461,9 +685,15 @@ Notice that each `probs` is an array of length `n_probs`. }' ``` -- **GET** `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`. +### GET `/slots`: Returns the current slots processing state -### Result JSON +This endpoint can be disabled with `--no-slots` + +If query param `?fail_on_no_slot=1` is set, this endpoint will respond with status code 503 if there is no available slots. + +**Response format** + +Example: ```json [ @@ -524,7 +754,13 @@ Notice that each `probs` is an array of length `n_probs`. ] ``` -- **GET** `/metrics`: [Prometheus](https://prometheus.io/) compatible metrics exporter endpoint if `--metrics` is enabled: +Possible values for `slot[i].state` are: +- `0`: SLOT_STATE_IDLE +- `1`: SLOT_STATE_PROCESSING + +### GET `/metrics`: Prometheus compatible metrics exporter + +This endpoint is only accessible if `--metrics` is set. Available metrics: - `llamacpp:prompt_tokens_total`: Number of prompt tokens processed. @@ -536,13 +772,13 @@ Available metrics: - `llamacpp:requests_processing`: Number of requests processing. - `llamacpp:requests_deferred`: Number of requests deferred. -- **POST** `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file. +### POST `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file. *Options:* `filename`: Name of the file to save the slot's prompt cache. The file will be saved in the directory specified by the `--slot-save-path` server parameter. -### Result JSON +**Response format** ```json { @@ -556,13 +792,13 @@ Available metrics: } ``` -- **POST** `/slots/{id_slot}?action=restore`: Restore the prompt cache of the specified slot from a file. +### POST `/slots/{id_slot}?action=restore`: Restore the prompt cache of the specified slot from a file. *Options:* `filename`: Name of the file to restore the slot's prompt cache from. The file should be located in the directory specified by the `--slot-save-path` server parameter. -### Result JSON +**Response format** ```json { @@ -576,9 +812,9 @@ Available metrics: } ``` -- **POST** `/slots/{id_slot}?action=erase`: Erase the prompt cache of the specified slot. +### POST `/slots/{id_slot}?action=erase`: Erase the prompt cache of the specified slot. -### Result JSON +**Response format** ```json { @@ -587,6 +823,46 @@ Available metrics: } ``` +### GET `/lora-adapters`: Get list of all LoRA adapters + +This endpoint returns the loaded LoRA adapters. You can add adapters using `--lora` when starting the server, for example: `--lora my_adapter_1.gguf --lora my_adapter_2.gguf ...` + +By default, all adapters will be loaded with scale set to 1. To initialize all adapters scale to 0, add `--lora-init-without-apply` + +If an adapter is disabled, the scale will be set to 0. + +**Response format** + +```json +[ + { + "id": 0, + "path": "my_adapter_1.gguf", + "scale": 0.0 + }, + { + "id": 1, + "path": "my_adapter_2.gguf", + "scale": 0.0 + } +] +``` + +### POST `/lora-adapters`: Set list of LoRA adapters + +To disable an adapter, either remove it from the list below, or set scale to 0. + +**Request format** + +To know the `id` of the adapter, use GET `/lora-adapters` + +```json +[ + {"id": 0, "scale": 0.2}, + {"id": 1, "scale": 0.8} +] +``` + ## More examples ### Change system prompt on runtime diff --git a/examples/server/bench/bench.py b/examples/server/bench/bench.py index 4fbbb2032..2daac0884 100644 --- a/examples/server/bench/bench.py +++ b/examples/server/bench/bench.py @@ -1,3 +1,5 @@ +from __future__ import annotations + import argparse import json import os @@ -59,10 +61,11 @@ def main(args_in: list[str] | None = None) -> None: sys.exit(1) # start the benchmark + iterations = 0 + data = {} try: start_benchmark(args) - iterations = 0 with open("results.github.env", 'w') as github_env: # parse output with open('k6-results.json', 'r') as bench_results: @@ -129,7 +132,7 @@ def main(args_in: list[str] | None = None) -> None: timestamps, metric_values = zip(*values) metric_values = [float(value) for value in metric_values] prometheus_metrics[metric] = metric_values - timestamps_dt = [datetime.fromtimestamp(int(ts)) for ts in timestamps] + timestamps_dt = [str(datetime.fromtimestamp(int(ts))) for ts in timestamps] plt.figure(figsize=(16, 10), dpi=80) plt.plot(timestamps_dt, metric_values, label=metric) plt.xticks(rotation=0, fontsize=14, horizontalalignment='center', alpha=.7) @@ -156,7 +159,7 @@ def main(args_in: list[str] | None = None) -> None: plt.close() # Mermaid format in case images upload failed - with (open(f"{metric}.mermaid", 'w') as mermaid_f): + with open(f"{metric}.mermaid", 'w') as mermaid_f: mermaid = ( f"""--- config: @@ -278,7 +281,7 @@ def start_server_background(args): } server_process = subprocess.Popen( args, - **pkwargs) + **pkwargs) # pyright: ignore[reportArgumentType, reportCallIssue] def server_log(in_stream, out_stream): for line in iter(in_stream.readline, b''): diff --git a/examples/server/public/completion.js b/examples/server/public/completion.js index 987b9a3b8..36818f764 100644 --- a/examples/server/public/completion.js +++ b/examples/server/public/completion.js @@ -21,7 +21,7 @@ let generation_settings = null; // export async function* llama(prompt, params = {}, config = {}) { let controller = config.controller; - const api_url = config.api_url || ""; + const api_url = config.api_url?.replace(/\/+$/, '') || ""; if (!controller) { controller = new AbortController(); @@ -196,7 +196,7 @@ export const llamaComplete = async (params, controller, callback) => { // Get the model info from the server. This is useful for getting the context window and so on. export const llamaModelInfo = async (config = {}) => { if (!generation_settings) { - const api_url = config.api_url || ""; + const api_url = config.api_url?.replace(/\/+$/, '') || ""; const props = await fetch(`${api_url}/props`).then(r => r.json()); generation_settings = props.default_generation_settings; } diff --git a/examples/server/public/index-new.html b/examples/server/public/index-new.html index 5513e9121..c87dd8f1e 100644 --- a/examples/server/public/index-new.html +++ b/examples/server/public/index-new.html @@ -14,10 +14,10 @@