From bec2183f2c8d37cf1278c11d1adb9311e9eaa242 Mon Sep 17 00:00:00 2001 From: ag2s20150909 <19373730+ag2s20150909@users.noreply.github.com> Date: Wed, 8 Jan 2025 16:17:29 +0800 Subject: [PATCH 1/6] fix: Vulkan shader gen binary path when Cross-compiling (#11096) * fix: Vulkan shader gen binary path when cross compiling --- ggml/src/ggml-vulkan/CMakeLists.txt | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/ggml/src/ggml-vulkan/CMakeLists.txt b/ggml/src/ggml-vulkan/CMakeLists.txt index 9501de736..61de21d6a 100644 --- a/ggml/src/ggml-vulkan/CMakeLists.txt +++ b/ggml/src/ggml-vulkan/CMakeLists.txt @@ -69,11 +69,15 @@ if (Vulkan_FOUND) file(GLOB _ggml_vk_shader_deps "${_ggml_vk_input_dir}/*.comp") + if (NOT CMAKE_CROSSCOMPILING) + set(_ggml_vk_genshaders_cmd "$/${_ggml_vk_genshaders_cmd}") + endif () + add_custom_command( OUTPUT ${_ggml_vk_header} ${_ggml_vk_source} - COMMAND "$/${_ggml_vk_genshaders_cmd}" + COMMAND ${_ggml_vk_genshaders_cmd} --glslc ${Vulkan_GLSLC_EXECUTABLE} --input-dir ${_ggml_vk_input_dir} --output-dir ${_ggml_vk_output_dir} From 02f04301417e7fb44fa1025bc1b0aef866e2ca89 Mon Sep 17 00:00:00 2001 From: Mathieu Baudier Date: Wed, 8 Jan 2025 09:18:13 +0100 Subject: [PATCH 2/6] Disable GL_KHR_cooperative_matrix Vulkan extension if not available. (#11117) * Disable GL_KHR_cooperative_matrix Vulkan extension if not available. * Perform Vulkan extensions checks in a more sensible order * Remove unnecessary #ifdef directive --- ggml/src/ggml-vulkan/CMakeLists.txt | 14 ++++++++++++++ ggml/src/ggml-vulkan/ggml-vulkan.cpp | 18 +++++++++++++++--- .../vulkan-shaders/test_coopmat_support.comp | 7 +++++++ .../vulkan-shaders/vulkan-shaders-gen.cpp | 2 ++ 4 files changed, 38 insertions(+), 3 deletions(-) create mode 100644 ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat_support.comp diff --git a/ggml/src/ggml-vulkan/CMakeLists.txt b/ggml/src/ggml-vulkan/CMakeLists.txt index 61de21d6a..c0ddaac82 100644 --- a/ggml/src/ggml-vulkan/CMakeLists.txt +++ b/ggml/src/ggml-vulkan/CMakeLists.txt @@ -8,6 +8,20 @@ if (Vulkan_FOUND) ../../include/ggml-vulkan.h ) + # Compile a test shader to determine whether GL_KHR_cooperative_matrix is supported. + # If it's not, there will be an error to stderr. + # If it's supported, set a define to indicate that we should compile those shaders + execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat_support.comp" + OUTPUT_VARIABLE glslc_output + ERROR_VARIABLE glslc_error) + + if (${glslc_error} MATCHES ".*extension not supported: GL_KHR_cooperative_matrix.*") + message(STATUS "GL_KHR_cooperative_matrix not supported by glslc") + else() + message(STATUS "GL_KHR_cooperative_matrix supported by glslc") + add_compile_definitions(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + endif() + # Compile a test shader to determine whether GL_NV_cooperative_matrix2 is supported. # If it's not, there will be an error to stderr. # If it's supported, set a define to indicate that we should compile those shaders diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index d75cd6d61..077452424 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -1645,6 +1645,7 @@ static void ggml_vk_load_shaders(vk_device& device) { #undef CREATE_MM2 } else #endif // defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) +#if defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) if (device->coopmat_support) { // Create 6 variants, {s,m,l}x{unaligned,aligned} #define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ @@ -1739,7 +1740,9 @@ static void ggml_vk_load_shaders(vk_device& device) { } #undef CREATE_MM2 #undef CREATE_MM - } else if (device->fp16) { + } else +#endif // defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + if (device->fp16) { // Create 6 variants, {s,m,l}x{unaligned,aligned} #define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ if (device->mul_mat ## ID ## _l) \ @@ -2242,6 +2245,7 @@ static vk_device ggml_vk_get_device(size_t idx) { last_struct = (VkBaseOutStructure *)&subgroup_size_control_features; } +#if defined(VK_KHR_cooperative_matrix) VkPhysicalDeviceCooperativeMatrixFeaturesKHR coopmat_features; coopmat_features.pNext = nullptr; coopmat_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR; @@ -2251,6 +2255,7 @@ static vk_device ggml_vk_get_device(size_t idx) { last_struct->pNext = (VkBaseOutStructure *)&coopmat_features; last_struct = (VkBaseOutStructure *)&coopmat_features; } +#endif #if defined(VK_NV_cooperative_matrix2) VkPhysicalDeviceCooperativeMatrix2FeaturesNV coopmat2_features {}; @@ -2283,7 +2288,9 @@ static vk_device ggml_vk_get_device(size_t idx) { device_extensions.push_back("VK_EXT_subgroup_size_control"); } +#if defined(VK_KHR_cooperative_matrix) device->coopmat_support = device->coopmat_support && coopmat_features.cooperativeMatrix; +#endif if (coopmat2_support) { #if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) @@ -2376,6 +2383,7 @@ static vk_device ggml_vk_get_device(size_t idx) { device_extensions.push_back("VK_KHR_shader_float16_int8"); } +#if defined(VK_KHR_cooperative_matrix) if (device->coopmat_support) { // Query supported shapes std::vector cm_props; @@ -2442,7 +2450,7 @@ static vk_device ggml_vk_get_device(size_t idx) { if (device->coopmat_support) { device_extensions.push_back("VK_KHR_cooperative_matrix"); } - +#endif device->name = GGML_VK_NAME + std::to_string(idx); device_create_info = { @@ -2553,9 +2561,11 @@ static void ggml_vk_print_gpu_info(size_t idx) { fp16_storage = true; } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) { fp16_compute = true; - } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 && +#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 && !getenv("GGML_VK_DISABLE_COOPMAT")) { coopmat_support = true; +#endif #if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) } else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 && !getenv("GGML_VK_DISABLE_COOPMAT2")) { @@ -2593,6 +2603,7 @@ static void ggml_vk_print_gpu_info(size_t idx) { // Pointer to the last chain element VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&vk12_features; +#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) VkPhysicalDeviceCooperativeMatrixFeaturesKHR coopmat_features; coopmat_features.pNext = nullptr; coopmat_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR; @@ -2608,6 +2619,7 @@ static void ggml_vk_print_gpu_info(size_t idx) { fp16 = fp16 && vk12_features.shaderFloat16; coopmat_support = coopmat_support && coopmat_features.cooperativeMatrix; +#endif std::string matrix_cores = coopmat2_support ? "NV_coopmat2" : coopmat_support ? "KHR_coopmat" : "none"; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat_support.comp b/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat_support.comp new file mode 100644 index 000000000..8c5dd1bd1 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat_support.comp @@ -0,0 +1,7 @@ +#version 460 + +#extension GL_KHR_cooperative_matrix : require + +void main() +{ +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp index 8111c0638..7b5044798 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -342,9 +342,11 @@ void process_shaders() { matmul_shaders(true, matmul_id, false, false, false); matmul_shaders(true, matmul_id, false, false, true); +#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) // Coopmat, fp32acc and fp16acc matmul_shaders(true, matmul_id, true, false, false); matmul_shaders(true, matmul_id, true, false, true); +#endif #if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) // Coopmat2, fp32acc and fp16acc From 0d52a69e4bf0d6181beec7853307bdcdeec9905b Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 8 Jan 2025 11:29:34 +0200 Subject: [PATCH 3/6] ci : fix cmake option (#11125) --- .github/workflows/build.yml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 602cf5220..02a193b86 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -665,7 +665,7 @@ jobs: - build: 'llvm-arm64' defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON' - build: 'msvc-arm64' - defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=O' + defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON' - build: 'llvm-arm64-opencl-adreno' defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON' From 8cef75c743ba13ebbd6d380c531200c768a8b8aa Mon Sep 17 00:00:00 2001 From: amritahs-ibm Date: Wed, 8 Jan 2025 16:24:19 +0530 Subject: [PATCH 4/6] llamafile : ppc64le MMA INT8 implementation (#10912) This change upstreams llamafile's cpu matrix multiplication kernels for ppc64le using MMA builtins for quantised int8 datatype. This change results in 10% - 70% improvement in total speed(ie all tokens/total time), across various batch sizes. The patch is tested with Meta-Lllama-3-8B, Mistral-7B, Llama-2-7B-chat-hf models on a IBM POWER10 machine. Signed-off-by: Amrita H S --- ggml/src/ggml-cpu/llamafile/sgemm.cpp | 836 ++++++++++++++++++++++++-- 1 file changed, 770 insertions(+), 66 deletions(-) diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.cpp b/ggml/src/ggml-cpu/llamafile/sgemm.cpp index 8fce576c3..c22a66287 100644 --- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp +++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp @@ -54,6 +54,7 @@ #include "ggml-quants.h" #include +#include #ifdef _MSC_VER #define NOINLINE __declspec(noinline) @@ -1051,6 +1052,704 @@ class tinyBLAS_Q0_AVX { } \ } \ +template +class tinyBLAS_Q0_PPC { + public: + tinyBLAS_Q0_PPC(int64_t k, + const TA *A, int64_t lda, + const TB *B, int64_t ldb, + TC *C, int64_t ldc, + int ith, int nth) + : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) { + } + + void matmul(int64_t m, int64_t n) { + mnpack(0, m, 0, n); + } + + private: + + template + inline void save_res(int ii, int jj, int idx, vector float* fin_res) { + for (int I = 0; I < RM; I++) { + for (int J = 0; J < RN; J++) { + *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&fin_res[idx+I]+J); + } + } + } + + template + inline void compute(acc_t* ACC, int c_idx, int s_idx, std::array& comparray, vector float* vs, vector float* fin_res) { + vector signed int vec_C[4]; + vector float CA[4] = {0}; + vector float res[4] = {0}; + __builtin_mma_disassemble_acc(vec_C, ACC); + for (int i = 0; i < 4; i++) { + CA[i] = vec_splats((float)(((double)comparray[c_idx+i]) * -128.0)); + res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]); + fin_res[s_idx+i] = vec_madd(res[i], vs[s_idx+i], fin_res[s_idx+i]); + } + } + + template + void packNormal(const TA* a, int64_t lda, int rows, int cols, VA* vec, bool flip) { + int64_t i, j; + TA *aoffset = NULL; + VA *vecOffset = NULL; + TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL; + TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL; + __vector_pair C1, C2, C3, C4, C5, C6, C7, C8; + VB c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2]={0}; + VB c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2]={0}; + VB t1, t2, t3, t4, t5, t6, t7, t8; + vector unsigned char xor_vector; + uint8_t flip_vec = 0x80; + xor_vector = vec_splats(flip_vec); + vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23}; + vector unsigned char swiz2 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31}; + vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27}; + vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31}; + + aoffset = const_cast(a); + vecOffset = vec; + j = (rows >> 3); + if (j > 0) { + do { + aoffset1 = aoffset; + aoffset2 = aoffset1 + lda; + aoffset3 = aoffset2 + lda; + aoffset4 = aoffset3 + lda; + aoffset5 = aoffset4 + lda; + aoffset6 = aoffset5 + lda; + aoffset7 = aoffset6 + lda; + aoffset8 = aoffset7 + lda; + aoffset += 8 * lda; + + i = (cols >> 3); + if (i > 0) { + do { + C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1->qs); + C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2->qs); + C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs); + C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4->qs); + C5 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset5->qs); + C6 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset6->qs); + C7 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset7->qs); + C8 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset8->qs); + + __builtin_vsx_disassemble_pair(c1, &C1); + __builtin_vsx_disassemble_pair(c2, &C2); + __builtin_vsx_disassemble_pair(c3, &C3); + __builtin_vsx_disassemble_pair(c4, &C4); + __builtin_vsx_disassemble_pair(c5, &C5); + __builtin_vsx_disassemble_pair(c6, &C6); + __builtin_vsx_disassemble_pair(c7, &C7); + __builtin_vsx_disassemble_pair(c8, &C8); + + t1 = vec_perm(c1[0], c2[0], swiz1); + t2 = vec_perm(c1[0], c2[0], swiz2); + t3 = vec_perm(c3[0], c4[0], swiz1); + t4 = vec_perm(c3[0], c4[0], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + if (flip == true) { + t5 = vec_xor(t5, xor_vector); + t6 = vec_xor(t6, xor_vector); + t7 = vec_xor(t7, xor_vector); + t8 = vec_xor(t8, xor_vector); + } + vec_xst(t5, 0, vecOffset); + vec_xst(t6, 0, vecOffset+16); + vec_xst(t7, 0, vecOffset+32); + vec_xst(t8, 0, vecOffset+48); + + t1 = vec_perm(c1[1], c2[1], swiz1); + t2 = vec_perm(c1[1], c2[1], swiz2); + t3 = vec_perm(c3[1], c4[1], swiz1); + t4 = vec_perm(c3[1], c4[1], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + if (flip == true) { + t5 = vec_xor(t5, xor_vector); + t6 = vec_xor(t6, xor_vector); + t7 = vec_xor(t7, xor_vector); + t8 = vec_xor(t8, xor_vector); + } + vec_xst(t5, 0, vecOffset+64); + vec_xst(t6, 0, vecOffset+80); + vec_xst(t7, 0, vecOffset+96); + vec_xst(t8, 0, vecOffset+112); + + t1 = vec_perm(c5[0], c6[0], swiz1); + t2 = vec_perm(c5[0], c6[0], swiz2); + t3 = vec_perm(c7[0], c8[0], swiz1); + t4 = vec_perm(c7[0], c8[0], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + if (flip == true) { + t5 = vec_xor(t5, xor_vector); + t6 = vec_xor(t6, xor_vector); + t7 = vec_xor(t7, xor_vector); + t8 = vec_xor(t8, xor_vector); + } + vec_xst(t5, 0, vecOffset+128); + vec_xst(t6, 0, vecOffset+144); + vec_xst(t7, 0, vecOffset+160); + vec_xst(t8, 0, vecOffset+176); + + t1 = vec_perm(c5[1], c6[1], swiz1); + t2 = vec_perm(c5[1], c6[1], swiz2); + t3 = vec_perm(c7[1], c8[1], swiz1); + t4 = vec_perm(c7[1], c8[1], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + if (flip == true) { + t5 = vec_xor(t5, xor_vector); + t6 = vec_xor(t6, xor_vector); + t7 = vec_xor(t7, xor_vector); + t8 = vec_xor(t8, xor_vector); + } + vec_xst(t5, 0, vecOffset+192); + vec_xst(t6, 0, vecOffset+208); + vec_xst(t7, 0, vecOffset+224); + vec_xst(t8, 0, vecOffset+240); + + aoffset1 += lda; + aoffset2 += lda; + aoffset3 += lda; + aoffset4 += lda; + aoffset5 += lda; + aoffset6 += lda; + aoffset7 += lda; + aoffset8 += lda; + vecOffset += 256; + i--; + } while(i > 0); + } + j--; + } while(j > 0); + } + + if (rows & 4) { + aoffset1 = aoffset; + aoffset2 = aoffset1 + lda; + aoffset3 = aoffset2 + lda; + aoffset4 = aoffset3 + lda; + aoffset += 4 * lda; + + i = (cols >> 3); + if (i > 0) { + do { + C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1->qs); + C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2->qs); + C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs); + C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4->qs); + + __builtin_vsx_disassemble_pair(c1, &C1); + __builtin_vsx_disassemble_pair(c2, &C2); + __builtin_vsx_disassemble_pair(c3, &C3); + __builtin_vsx_disassemble_pair(c4, &C4); + + t1 = vec_perm(c1[0], c2[0], swiz1); + t2 = vec_perm(c1[0], c2[0], swiz2); + t3 = vec_perm(c3[0], c4[0], swiz1); + t4 = vec_perm(c3[0], c4[0], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + if (flip == true) { + t5 = vec_xor(t5, xor_vector); + t6 = vec_xor(t6, xor_vector); + t7 = vec_xor(t7, xor_vector); + t8 = vec_xor(t8, xor_vector); + } + vec_xst(t5, 0, vecOffset); + vec_xst(t6, 0, vecOffset+16); + vec_xst(t7, 0, vecOffset+32); + vec_xst(t8, 0, vecOffset+48); + + t1 = vec_perm(c1[1], c2[1], swiz1); + t2 = vec_perm(c1[1], c2[1], swiz2); + t3 = vec_perm(c3[1], c4[1], swiz1); + t4 = vec_perm(c3[1], c4[1], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + if (flip == true) { + t5 = vec_xor(t5, xor_vector); + t6 = vec_xor(t6, xor_vector); + t7 = vec_xor(t7, xor_vector); + t8 = vec_xor(t8, xor_vector); + } + vec_xst(t5, 0, vecOffset+64); + vec_xst(t6, 0, vecOffset+80); + vec_xst(t7, 0, vecOffset+96); + vec_xst(t8, 0, vecOffset+112); + + aoffset1 += lda; + aoffset2 += lda; + aoffset3 += lda; + aoffset4 += lda; + vecOffset += 128; + i--; + } while(i > 0); + } + } + if (rows & 3) { + aoffset1 = aoffset; + aoffset2 = aoffset1 + lda; + aoffset3 = aoffset2 + lda; + i = (cols >> 3); + if (i > 0) { + do { + switch(rows) { + case 3: C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs); + __builtin_vsx_disassemble_pair(c3, &C3); + case 2: C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2->qs); + __builtin_vsx_disassemble_pair(c2, &C2); + case 1: C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1->qs); + __builtin_vsx_disassemble_pair(c1, &C1); + break; + } + t1 = vec_perm(c1[0], c2[0], swiz1); + t2 = vec_perm(c1[0], c2[0], swiz2); + t3 = vec_perm(c3[0], c4[0], swiz1); + t4 = vec_perm(c3[0], c4[0], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + if (flip == true) { + t5 = vec_xor(t5, xor_vector); + t6 = vec_xor(t6, xor_vector); + t7 = vec_xor(t7, xor_vector); + t8 = vec_xor(t8, xor_vector); + } + vec_xst(t5, 0, vecOffset); + vec_xst(t6, 0, vecOffset+16); + vec_xst(t7, 0, vecOffset+32); + vec_xst(t8, 0, vecOffset+48); + + t1 = vec_perm(c1[1], c2[1], swiz1); + t2 = vec_perm(c1[1], c2[1], swiz2); + t3 = vec_perm(c3[1], c4[1], swiz1); + t4 = vec_perm(c3[1], c4[1], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + if (flip == true) { + t5 = vec_xor(t5, xor_vector); + t6 = vec_xor(t6, xor_vector); + t7 = vec_xor(t7, xor_vector); + t8 = vec_xor(t8, xor_vector); + } + vec_xst(t5, 0, vecOffset+64); + vec_xst(t6, 0, vecOffset+80); + vec_xst(t7, 0, vecOffset+96); + vec_xst(t8, 0, vecOffset+112); + + aoffset1 += lda; + aoffset2 += lda; + aoffset3 += lda; + vecOffset += 128; + i--; + } while(i > 0); + } + } + } + + void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) { + int64_t mc, nc, mp, np; + int m_rem = MIN(m - m0, 8); + int n_rem = MIN(n - n0, 8); + // TO-DO: KERNEL_16x8 and KERNEL_8x16 are having some performance + // issues. After resolving them, below code will be enabled. + /*if (m_rem >= 16 && n_rem >= 8) { + mc = 16; + nc = 8; + gemm<16,8>(m0, m, n0, n); + } else if(m_rem >= 8 && n_rem >= 16) { + mc = 8; + nc = 16; + gemm<8,16>(m0, m, n0, n); + }*/ + if (m_rem >= 8 && n_rem >= 8) { + mc = 8; + nc = 8; + gemm<8,8>(m0, m, n0, n); + } else if (m_rem >= 4 && n_rem >= 8) { + mc = 4; + nc = 8; + gemm<4,8>(m0, m, n0, n); + } else if (m_rem >= 8 && n_rem >= 4) { + mc = 8; + nc = 4; + gemm<8,4>(m0, m, n0, n); + } else if (m_rem >= 4 && n_rem >= 4) { + mc = 4; + nc = 4; + gemm_small<4, 4>(m0, m, n0, n); + } else if ((m_rem < 4) && (n_rem > 4)) { + nc = 4; + switch(m_rem) { + case 1: + mc = 1; + gemm_small<1, 4>(m0, m, n0, n); + break; + case 2: + mc = 2; + gemm_small<2, 4>(m0, m, n0, n); + break; + case 3: + mc = 3; + gemm_small<3, 4>(m0, m, n0, n); + break; + default: + return; + } + } else if ((m_rem > 4) && (n_rem < 4)) { + mc = 4; + switch(n_rem) { + case 1: + nc = 1; + gemm_small<4, 1>(m0, m, n0, n); + break; + case 2: + nc = 2; + gemm_small<4, 2>(m0, m, n0, n); + break; + case 3: + nc = 3; + gemm_small<4, 3>(m0, m, n0, n); + break; + default: + return; + } + } else { + switch((m_rem << 4) | n_rem) { + case 0x43: + mc = 4; + nc = 3; + gemm_small<4, 3>(m0, m, n0, n); + break; + case 0x42: + mc = 4; + nc = 2; + gemm_small<4, 2>(m0, m, n0, n); + break; + case 0x41: + mc = 4; + nc = 1; + gemm_small<4, 1>(m0, m, n0, n); + break; + case 0x34: + mc = 3; + nc = 4; + gemm_small<3, 4>(m0, m, n0, n); + break; + case 0x33: + mc = 3; + nc = 3; + gemm_small<3, 3>(m0, m, n0, n); + break; + case 0x32: + mc = 3; + nc = 2; + gemm_small<3, 2>(m0, m, n0, n); + break; + case 0x31: + mc = 3; + nc = 1; + gemm_small<3, 1>(m0, m, n0, n); + break; + case 0x24: + mc = 2; + nc = 4; + gemm_small<2, 4>(m0, m, n0, n); + break; + case 0x23: + mc = 2; + nc = 3; + gemm_small<2, 3>(m0, m, n0, n); + break; + case 0x22: + mc = 2; + nc = 2; + gemm_small<2, 2>(m0, m, n0, n); + break; + case 0x21: + mc = 2; + nc = 1; + gemm_small<2, 1>(m0, m, n0, n); + break; + case 0x14: + mc = 1; + nc = 4; + gemm_small<1, 4>(m0, m, n0, n); + break; + case 0x13: + mc = 1; + nc = 3; + gemm_small<1, 3>(m0, m, n0, n); + break; + case 0x12: + mc = 1; + nc = 2; + gemm_small<1, 2>(m0, m, n0, n); + break; + case 0x11: + mc = 1; + nc = 1; + gemm_small<1, 1>(m0, m, n0, n); + break; + default: + return; + } + } + mp = m0 + (m - m0) / mc * mc; + np = n0 + (n - n0) / nc * nc; + mnpack(mp, m, n0, np); + mnpack(m0, m, np, n); + } + + void KERNEL_4x8(int64_t ii, int64_t jj) { + vec_t vec_A[8], vec_B[16] = {0}; + acc_t acc_0, acc_1; + std::array comparray; + vector float fin_res[8] = {0}; + vector float vs[8] = {0}; + for (int l = 0; l < k; l++) { + __builtin_mma_xxsetaccz(&acc_0); + __builtin_mma_xxsetaccz(&acc_1); + packNormal((A+(ii*lda)+l), lda, 4, 8, (int8_t*)vec_A, false); + packNormal((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true); + for(int x = 0; x < 8; x++) { + __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]); + __builtin_mma_xvi8ger4pp(&acc_1, vec_A[x], vec_B[x+8]); + } + for (int I = 0; I<4; I++) { + for (int J = 0; J<4; J++) { + *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d)); + *((float*)&vs[I+4]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d)); + } + } + auto aoffset = A+(ii*lda)+l; + for (int i = 0; i < 4; i++) { + comparray[i] = 0; + int ca = 0; + const int8_t *at = aoffset->qs; + for (int j = 0; j < 32; j++) + ca += (int)*at++; + comparray[i] = ca; + aoffset += lda; + } + compute<4>(&acc_0, 0, 0, comparray, vs, fin_res); + compute<4>(&acc_1, 0, 4, comparray, vs, fin_res); + } + save_res<4, 4>(ii, jj, 0, fin_res); + save_res<4, 4>(ii, jj+4, 4, fin_res); + } + + void KERNEL_8x4(int64_t ii, int64_t jj) { + vec_t vec_A[16], vec_B[8] = {0}; + acc_t acc_0, acc_1; + std::array comparray; + vector float fin_res[8] = {0}; + vector float vs[8] = {0}; + for (int l = 0; l < k; l++) { + __builtin_mma_xxsetaccz(&acc_0); + __builtin_mma_xxsetaccz(&acc_1); + packNormal((A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false); + packNormal((B+(jj*ldb)+l), ldb, 4, 8, (uint8_t*)vec_B, true); + for(int x = 0; x < 8; x++) { + __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]); + __builtin_mma_xvi8ger4pp(&acc_1, vec_A[x+8], vec_B[x]); + } + for (int I = 0; I<8; I++) { + for (int J = 0; J<4; J++) { + *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d)); + } + } + auto aoffset = A+(ii*lda)+l; + for (int i = 0; i < 8; i++) { + comparray[i] = 0; + int ca = 0; + const int8_t *at = aoffset->qs; + for (int j = 0; j < 32; j++) + ca += (int)*at++; + comparray[i] = ca; + aoffset += lda; + } + compute<8>(&acc_0, 0, 0, comparray, vs, fin_res); + compute<8>(&acc_1, 4, 4, comparray, vs, fin_res); + } + save_res<4, 4>(ii, jj, 0, fin_res); + save_res<4, 4>(ii+4, jj, 4, fin_res); + } + + void KERNEL_8x8(int64_t ii, int64_t jj) { + vec_t vec_A[16], vec_B[16] = {0}; + acc_t acc_0, acc_1, acc_2, acc_3; + std::array comparray; + vector float fin_res[16] = {0}; + vector float vs[16] = {0}; + for (int l = 0; l < k; l++) { + __builtin_mma_xxsetaccz(&acc_0); + __builtin_mma_xxsetaccz(&acc_1); + __builtin_mma_xxsetaccz(&acc_2); + __builtin_mma_xxsetaccz(&acc_3); + packNormal((A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false); + packNormal((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true); + for(int x = 0; x < 8; x++) { + __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]); + __builtin_mma_xvi8ger4pp(&acc_1, vec_A[x+8], vec_B[x]); + __builtin_mma_xvi8ger4pp(&acc_2, vec_A[x], vec_B[x+8]); + __builtin_mma_xvi8ger4pp(&acc_3, vec_A[x+8], vec_B[x+8]); + } + for (int I = 0; I<8; I++) { + for (int J = 0; J<4; J++) { + *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d)); + *((float*)&vs[I+8]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d)); + } + } + auto aoffset = A+(ii*lda)+l; + for (int i = 0; i < 8; i++) { + comparray[i] = 0; + int ca = 0; + const int8_t *at = aoffset->qs; + for (int j = 0; j < 32; j++) + ca += (int)*at++; + comparray[i] = ca; + aoffset += lda; + } + compute<8>(&acc_0, 0, 0, comparray, vs, fin_res); + compute<8>(&acc_1, 4, 4, comparray, vs, fin_res); + compute<8>(&acc_2, 0, 8, comparray, vs, fin_res); + compute<8>(&acc_3, 4, 12, comparray, vs, fin_res); + } + save_res<4, 4>(ii, jj, 0, fin_res); + save_res<4, 4>(ii+4, jj, 4, fin_res); + save_res<4, 4>(ii, jj+4, 8, fin_res); + save_res<4, 4>(ii+4, jj+4, 12, fin_res); + } + + template + void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n) { + int64_t ytiles = (m - m0) / RM; + int64_t xtiles = (n - n0) / RN; + int64_t tiles = xtiles * ytiles; + int64_t duty = (tiles + nth - 1) / nth; + int64_t start = duty * ith; + int64_t end = start + duty; + vec_t vec_A[8], vec_B[8] = {0}; + vector signed int vec_C[4]; + acc_t acc_0; + + if (end > tiles) + end = tiles; + for (int64_t job = start; job < end; ++job) { + int64_t ii = m0 + job / xtiles * RM; + int64_t jj = n0 + job % xtiles * RN; + std::array comparray; + vector float res[4] = {0}; + vector float fin_res[4] = {0}; + vector float vs[4] = {0}; + vector float CA[4] = {0}; + __builtin_prefetch((A+(ii*lda)+0)->qs, 0, 1); // prefetch first value + __builtin_prefetch((B+(jj*ldb)+0)->qs, 0, 1); // prefetch first value + for (int l = 0; l < k; l++) { + __builtin_prefetch((A+(ii*lda)+(l+1))->qs, 0, 1); // prefetch one loop ahead + __builtin_prefetch((B+(jj*ldb)+(l+1))->qs, 0, 1); // prefetch one loop ahead + __builtin_mma_xxsetaccz(&acc_0); + packNormal((A+(ii*lda)+l), lda, RM, 8, (int8_t*)vec_A, false); + packNormal((B+(jj*ldb)+l), ldb, RN, 8, (uint8_t*)vec_B, true); + for(int x = 0; x < 8; x+=4) { + __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]); + __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+1], vec_B[x+1]); + __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+2], vec_B[x+2]); + __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+3], vec_B[x+3]); + } + for (int I = 0; Id) * unhalf((B+((jj+J)*ldb)+l)->d)); + } + } + __builtin_mma_disassemble_acc(vec_C, &acc_0); + auto aoffset = A+(ii*lda)+l; + for (int i = 0; i < RM; i++) { + comparray[i] = 0; + int ca = 0; + const int8_t *at = aoffset->qs; + for (int j = 0; j < 32; j++) + ca += (int)*at++; + comparray[i] = ca; + aoffset += lda; + } + + for (int i = 0; i < RM; i++) { + CA[i] = vec_splats((float)(((double)comparray[i]) * -128.0)); + res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]); + fin_res[i] = vec_madd(res[i], vs[i], fin_res[i]); + } + } + save_res(ii, jj, 0, fin_res); + } + } + + template + inline void kernel(int64_t ii, int64_t jj) { + if constexpr(RM == 4 && RN == 8) { + KERNEL_4x8(ii,jj); + } else if constexpr(RM == 8 && RN == 4) { + KERNEL_8x4(ii,jj); + } else if constexpr(RM == 8 && RN == 8) { + KERNEL_8x8(ii,jj); + } else { + static_assert(false, "RN/RM values not supported"); + } + } + + template + NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) { + int64_t ytiles = (m - m0) / RM; + int64_t xtiles = (n - n0) / RN; + int64_t tiles = xtiles * ytiles; + int64_t duty = (tiles + nth - 1) / nth; + int64_t start = duty * ith; + int64_t end = start + duty; + if (end > tiles) + end = tiles; + for (int64_t job = start; job < end; ++job) { + int64_t ii = m0 + job / xtiles * RM; + int64_t jj = n0 + job % xtiles * RN; + kernel(ii, jj); + } + } + + const TA *const A; + const TB *const B; + TC *C; + TA *At; + TB *Bt; + const int64_t k; + const int64_t lda; + const int64_t ldb; + const int64_t ldc; + const int ith; + const int nth; +}; + template class tinyBLAS_PPC { public: @@ -1070,13 +1769,17 @@ class tinyBLAS_PPC { void (tinyBLAS_PPC::*kernel)(int64_t, int64_t); - void READ_BLOCK(const float* a, int64_t lda, int rows, int cols, float* vec) { + template + void packTranspose(const TA* a, int64_t lda, int rows, int cols, TA* vec) { int64_t i, j; - float *aoffset = NULL, *boffset = NULL; - float *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL; - float *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL; - - aoffset = const_cast(a); + TA *aoffset = NULL, *boffset = NULL; + TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL; + TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL; + __vector_pair C1, C2, C3, C4, C5, C6, C7, C8; + VA c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2] = {0}; + VA c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2] = {0}; + VA t1, t2, t3, t4, t5, t6, t7, t8; + aoffset = const_cast(a); boffset = vec; j = (rows >> 3); if (j > 0) { @@ -1092,9 +1795,6 @@ class tinyBLAS_PPC { aoffset += 8 * lda; i = (cols >> 3); if (i > 0) { - __vector_pair C1, C2, C3, C4, C5, C6, C7, C8; - vector float c1[2], c2[2], c3[2], c4[2], c5[2], c6[2], c7[2], c8[2]; - vector float t1, t2, t3, t4, t5, t6, t7, t8; do { C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1); C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2); @@ -1174,21 +1874,19 @@ class tinyBLAS_PPC { } while(i > 0); } if (cols & 4) { - vector float c1, c2, c3, c4, c5, c6, c7, c8; - vector float t1, t2, t3, t4, t5, t6, t7, t8; - c1 = vec_xl(0, aoffset1); - c2 = vec_xl(0, aoffset2); - c3 = vec_xl(0, aoffset3); - c4 = vec_xl(0, aoffset4); - c5 = vec_xl(0, aoffset5); - c6 = vec_xl(0, aoffset6); - c7 = vec_xl(0, aoffset7); - c8 = vec_xl(0, aoffset8); + c1[0] = vec_xl(0, aoffset1); + c2[0] = vec_xl(0, aoffset2); + c3[0] = vec_xl(0, aoffset3); + c4[0] = vec_xl(0, aoffset4); + c5[0] = vec_xl(0, aoffset5); + c6[0] = vec_xl(0, aoffset6); + c7[0] = vec_xl(0, aoffset7); + c8[0] = vec_xl(0, aoffset8); - t1 = vec_mergeh(c1, c2); - t2 = vec_mergeh(c3, c4); - t3 = vec_mergeh(c5, c6); - t4 = vec_mergeh(c7, c8); + t1 = vec_mergeh(c1[0], c2[0]); + t2 = vec_mergeh(c3[0], c4[0]); + t3 = vec_mergeh(c5[0], c6[0]); + t4 = vec_mergeh(c7[0], c8[0]); t5 = vec_xxpermdi(t1, t2, 0); t6 = vec_xxpermdi(t3, t4, 0); t7 = vec_xxpermdi(t1, t2, 3); @@ -1198,10 +1896,10 @@ class tinyBLAS_PPC { vec_xst(t7, 0, boffset+8); vec_xst(t8, 0, boffset+12); - t1 = vec_mergel(c1, c2); - t2 = vec_mergel(c3, c4); - t3 = vec_mergel(c5, c6); - t4 = vec_mergel(c7, c8); + t1 = vec_mergel(c1[0], c2[0]); + t2 = vec_mergel(c3[0], c4[0]); + t3 = vec_mergel(c5[0], c6[0]); + t4 = vec_mergel(c7[0], c8[0]); t5 = vec_xxpermdi(t1, t2, 0); t6 = vec_xxpermdi(t3, t4, 0); t7 = vec_xxpermdi(t1, t2, 3); @@ -1223,9 +1921,6 @@ class tinyBLAS_PPC { aoffset += 4 * lda; i = (cols >> 3); if (i > 0) { - __vector_pair C1, C2, C3, C4; - vector float c1[2], c2[2], c3[2], c4[2]; - vector float t1, t2, t3, t4, t5, t6, t7, t8; do { C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1); C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2); @@ -1272,22 +1967,20 @@ class tinyBLAS_PPC { } if (cols & 4) { - vector float c1, c2, c3, c4; - vector float t1, t2, t3, t4; - c1 = vec_xl(0, aoffset1); - c2 = vec_xl(0, aoffset2); - c3 = vec_xl(0, aoffset3); - c4 = vec_xl(0, aoffset4); + c1[0] = vec_xl(0, aoffset1); + c2[0] = vec_xl(0, aoffset2); + c3[0] = vec_xl(0, aoffset3); + c4[0] = vec_xl(0, aoffset4); - t1 = vec_mergeh(c1, c2); - t2 = vec_mergeh(c3, c4); + t1 = vec_mergeh(c1[0], c2[0]); + t2 = vec_mergeh(c3[0], c4[0]); t3 = vec_xxpermdi(t1, t2, 0); t4 = vec_xxpermdi(t1, t2, 3); vec_xst(t3, 0, boffset); vec_xst(t4, 0, boffset+4); - t1 = vec_mergel(c1, c2); - t2 = vec_mergel(c3, c4); + t1 = vec_mergel(c1[0], c2[0]); + t2 = vec_mergel(c3[0], c4[0]); t3 = vec_xxpermdi(t1, t2, 0); t4 = vec_xxpermdi(t1, t2, 3); vec_xst(t3, 0, boffset+8); @@ -1299,21 +1992,19 @@ class tinyBLAS_PPC { aoffset2 = aoffset1 + lda; aoffset3 = aoffset2 + lda; if (cols & 4) { - vector float c1, c2, c3, c4 = {0}; - vector float t1, t2, t3, t4; - c1 = vec_xl(0, aoffset1); - c2 = vec_xl(0, aoffset2); - c3 = vec_xl(0, aoffset3); + c1[0] = vec_xl(0, aoffset1); + c2[0] = vec_xl(0, aoffset2); + c3[0] = vec_xl(0, aoffset3); - t1 = vec_mergeh(c1, c2); - t2 = vec_mergeh(c3, c4); + t1 = vec_mergeh(c1[0], c2[0]); + t2 = vec_mergeh(c3[0], c4[0]); t3 = vec_xxpermdi(t1, t2, 0); t4 = vec_xxpermdi(t1, t2, 3); vec_xst(t3, 0, boffset); vec_xst(t4, 0, boffset+4); - t1 = vec_mergel(c1, c2); - t2 = vec_mergel(c3, c4); + t1 = vec_mergel(c1[0], c2[0]); + t2 = vec_mergel(c3[0], c4[0]); t3 = vec_xxpermdi(t1, t2, 0); t4 = vec_xxpermdi(t1, t2, 3); vec_xst(t3, 0, boffset+8); @@ -1321,14 +2012,13 @@ class tinyBLAS_PPC { } } } - void KERNEL_4x4(int64_t ii, int64_t jj) { vec_t vec_A[4], vec_B[4], vec_C[4]; acc_t acc_0; __builtin_mma_xxsetaccz(&acc_0); for (int l = 0; l < k; l+=4) { - READ_BLOCK(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A); - READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B); + packTranspose(A+(ii*lda)+l, lda, 4, 4, (TA*)vec_A); + packTranspose(B+(jj*ldb)+l, ldb, 4, 4, (TA*)vec_B); __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]); __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]); __builtin_mma_xvf32gerpp(&acc_0, vec_A[2], vec_B[2]); @@ -1343,8 +2033,8 @@ class tinyBLAS_PPC { __builtin_mma_xxsetaccz(&acc_0); __builtin_mma_xxsetaccz(&acc_1); for (int64_t l = 0; l < k; l+=4) { - READ_BLOCK(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A); - READ_BLOCK(B+(jj*ldb)+l, ldb, 8, 4, (float*)vec_B); + packTranspose(A+(ii*lda)+l, lda, 4, 4, (TA*)vec_A); + packTranspose(B+(jj*ldb)+l, ldb, 8, 4, (TA*)vec_B); __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], (vec_t)vec_B[0]); __builtin_mma_xvf32gerpp(&acc_1, vec_A[0], (vec_t)vec_B[1]); __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], (vec_t)vec_B[2]); @@ -1364,8 +2054,8 @@ class tinyBLAS_PPC { __builtin_mma_xxsetaccz(&acc_0); __builtin_mma_xxsetaccz(&acc_1); for (int64_t l = 0; l < k; l+=4) { - READ_BLOCK(A+(ii*lda)+l, lda, 8, 4, (float*)vec_A); - READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B); + packTranspose(A+(ii*lda)+l, lda, 8, 4, (TA*)vec_A); + packTranspose(B+(jj*ldb)+l, ldb, 4, 4, (TA*)vec_B); __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[0], vec_B[0]); __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[1], vec_B[0]); __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[2], vec_B[1]); @@ -1387,8 +2077,8 @@ class tinyBLAS_PPC { __builtin_mma_xxsetaccz(&acc_2); __builtin_mma_xxsetaccz(&acc_3); for (int l = 0; l < k; l+=8) { - READ_BLOCK(A+(ii*lda)+l, lda, 8, 8, (float*)vec_A); - READ_BLOCK(B+(jj*ldb)+l, ldb, 8, 8, (float*)vec_B); + packTranspose(A+(ii*lda)+l, lda, 8, 8, (TA*)vec_A); + packTranspose(B+(jj*ldb)+l, ldb, 8, 8, (TA*)vec_B); for(int x = 0; x < 16; x+=2) { __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[x], vec_B[x]); __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[x], vec_B[x+1]); @@ -1571,15 +2261,15 @@ class tinyBLAS_PPC { vec_t vec_A[4], vec_B[4]; for (int l=0; l= 4 && RM == 1) { - float* a = const_cast(A+(ii)*lda+l); - READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B); + TA* a = const_cast(A+(ii)*lda+l); + packTranspose(B+(jj*ldb)+l, ldb, 4, 4, (TA*)vec_B); vec_A[0] = (vec_t)vec_xl(0,a); - vec_A[1] = (vec_t)vec_splats(*((float*)&vec_A+1)); - vec_A[2] = (vec_t)vec_splats(*((float*)&vec_A+2)); - vec_A[3] = (vec_t)vec_splats(*((float*)&vec_A+3)); + vec_A[1] = (vec_t)vec_splats(*((TA*)&vec_A+1)); + vec_A[2] = (vec_t)vec_splats(*((TA*)&vec_A+2)); + vec_A[3] = (vec_t)vec_splats(*((TA*)&vec_A+3)); } else { - READ_BLOCK(A+(ii*lda)+l, lda, RM, 4, (float*)vec_A); - READ_BLOCK(B+(jj*ldb)+l, ldb, RN, 4, (float*)vec_B); + packTranspose(A+(ii*lda)+l, lda, RM, 4, (TA*)vec_A); + packTranspose(B+(jj*ldb)+l, ldb, RN, 4, (TA*)vec_B); } __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]); __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]); @@ -1589,7 +2279,7 @@ class tinyBLAS_PPC { __builtin_mma_disassemble_acc(vec_C, &acc_0); for (int I = 0; I < RM; I++) { for (int J = 0; J < RN; J++) { - *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&vec_C[I]+J); + *((TC*)(C+ii+((jj+J)*ldc)+I)) = *((TC*)&vec_C[I]+J); } } } @@ -1812,6 +2502,20 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64 params->ith, params->nth}; tb.matmul(m, n); return true; + +#elif defined(__MMA__) + if (n < 8 && n != 4) + return false; + if (m < 8 && m != 4) + return false; + tinyBLAS_Q0_PPC tb{ + k, (const block_q8_0 *)A, lda, + (const block_q8_0 *)B, ldb, + (float *)C, ldc, + params->ith, params->nth}; + tb.matmul(m, n); + return true; + #else return false; #endif From a3c1232c3f475f0a77b9cc5225516ac31c567a06 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 8 Jan 2025 12:55:36 +0200 Subject: [PATCH 5/6] arg : option to exclude arguments from specific examples (#11136) * arg : option to exclude arguments from specific examples ggml-ci * readme : remove old args [no ci] --- common/arg.cpp | 17 +++++++++++++---- common/arg.h | 3 +++ examples/server/README.md | 3 --- 3 files changed, 16 insertions(+), 7 deletions(-) diff --git a/common/arg.cpp b/common/arg.cpp index c81b15217..27886b84e 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -22,6 +22,11 @@ common_arg & common_arg::set_examples(std::initializer_list return *this; } +common_arg & common_arg::set_excludes(std::initializer_list excludes) { + this->excludes = std::move(excludes); + return *this; +} + common_arg & common_arg::set_env(const char * env) { help = help + "\n(env: " + env + ")"; this->env = env; @@ -37,6 +42,10 @@ bool common_arg::in_example(enum llama_example ex) { return examples.find(ex) != examples.end(); } +bool common_arg::is_exclude(enum llama_example ex) { + return excludes.find(ex) != excludes.end(); +} + bool common_arg::get_value_from_env(std::string & output) { if (env == nullptr) return false; char * value = std::getenv(env); @@ -420,7 +429,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex * - if both {LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_*,} are set, we will prioritize the LLAMA_EXAMPLE_* matching current example */ auto add_opt = [&](common_arg arg) { - if (arg.in_example(ex) || arg.in_example(LLAMA_EXAMPLE_COMMON)) { + if ((arg.in_example(ex) || arg.in_example(LLAMA_EXAMPLE_COMMON)) && !arg.is_exclude(ex)) { ctx_arg.options.push_back(std::move(arg)); } }; @@ -649,7 +658,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params, const std::string & value) { params.prompt = value; } - )); + ).set_excludes({LLAMA_EXAMPLE_SERVER})); add_opt(common_arg( {"--no-perf"}, string_format("disable internal libllama performance timings (default: %s)", params.no_perf ? "true" : "false"), @@ -673,7 +682,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.prompt.pop_back(); } } - )); + ).set_excludes({LLAMA_EXAMPLE_SERVER})); add_opt(common_arg( {"--in-file"}, "FNAME", "an input file (repeat to specify multiple files)", @@ -700,7 +709,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.prompt = ss.str(); fprintf(stderr, "Read %zu bytes from binary file %s\n", params.prompt.size(), value.c_str()); } - )); + ).set_excludes({LLAMA_EXAMPLE_SERVER})); add_opt(common_arg( {"-e", "--escape"}, string_format("process escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\) (default: %s)", params.escape ? "true" : "false"), diff --git a/common/arg.h b/common/arg.h index a6700d323..49ab8667b 100644 --- a/common/arg.h +++ b/common/arg.h @@ -12,6 +12,7 @@ struct common_arg { std::set examples = {LLAMA_EXAMPLE_COMMON}; + std::set excludes = {}; std::vector args; const char * value_hint = nullptr; // help text or example for arg value const char * value_hint_2 = nullptr; // for second arg value @@ -53,9 +54,11 @@ struct common_arg { ) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {} common_arg & set_examples(std::initializer_list examples); + common_arg & set_excludes(std::initializer_list excludes); common_arg & set_env(const char * env); common_arg & set_sparam(); bool in_example(enum llama_example ex); + bool is_exclude(enum llama_example ex); bool get_value_from_env(std::string & output); bool has_value_from_env(); std::string to_string(); diff --git a/examples/server/README.md b/examples/server/README.md index 3ce16945a..1f0a27d96 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -45,10 +45,7 @@ The project is under active development, and we are [looking for feedback and co | `-ub, --ubatch-size N` | physical maximum batch size (default: 512)
(env: LLAMA_ARG_UBATCH) | | `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) | | `-fa, --flash-attn` | enable Flash Attention (default: disabled)
(env: LLAMA_ARG_FLASH_ATTN) | -| `-p, --prompt PROMPT` | prompt to start generation with | | `--no-perf` | disable internal libllama performance timings (default: false)
(env: LLAMA_ARG_NO_PERF) | -| `-f, --file FNAME` | a file containing the prompt (default: none) | -| `-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 | | `--rope-scaling {none,linear,yarn}` | RoPE frequency scaling method, defaults to linear unless specified by the model
(env: LLAMA_ARG_ROPE_SCALING_TYPE) | From 80ccf5d725571035b454659e3c1b4b2b07b65e71 Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Wed, 8 Jan 2025 12:07:20 +0100 Subject: [PATCH 6/6] ci : pin dependency to specific version (#11137) * ci : pin dependency to specific version * will this fix ec? --- .github/workflows/docker.yml | 2 +- .github/workflows/editorconfig.yml | 4 +++- 2 files changed, 4 insertions(+), 2 deletions(-) diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml index 41f1a89ee..f5af72d0b 100644 --- a/.github/workflows/docker.yml +++ b/.github/workflows/docker.yml @@ -100,7 +100,7 @@ jobs: # https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example - name: Free Disk Space (Ubuntu) if: ${{ matrix.config.free_disk_space == true }} - uses: jlumbroso/free-disk-space@main + uses: jlumbroso/free-disk-space@v1.3.1 with: # this might remove tools that are actually needed, # if set to "true" but frees about 6 GB diff --git a/.github/workflows/editorconfig.yml b/.github/workflows/editorconfig.yml index ae86e9927..f02b7c219 100644 --- a/.github/workflows/editorconfig.yml +++ b/.github/workflows/editorconfig.yml @@ -23,5 +23,7 @@ jobs: runs-on: ubuntu-latest steps: - uses: actions/checkout@v4 - - uses: editorconfig-checker/action-editorconfig-checker@main + - uses: editorconfig-checker/action-editorconfig-checker@v2 + with: + version: v3.0.3 - run: editorconfig-checker