From 60ce97c9d809f4b040e90b597468b839df5728d0 Mon Sep 17 00:00:00 2001 From: Ma Mingfei Date: Fri, 18 Oct 2024 13:34:36 +0800 Subject: [PATCH] add amx kernel for gemm (#8998) add intel amx isa detection add vnni kernel for gemv cases add vnni and amx kernel support for block_q8_0 code cleanup fix packing B issue enable openmp fine tune amx kernel switch to aten parallel pattern add error message for nested parallelism code cleanup add f16 support in ggml-amx add amx kernels for QK_K quant formats: Q4_K, Q5_K, Q6_K and IQ4_XS update CMakeList update README fix some compilation warning fix compiler warning when amx is not enabled minor change ggml-ci move ggml_amx_init from ggml.c to ggml-amx/mmq.cpp ggml-ci update CMakeLists with -mamx-tile, -mamx-int8 and -mamx-bf16 ggml-ci add amx as an ggml-backend update header file, the old path for immintrin.h has changed to ggml-cpu-impl.h minor change update CMakeLists.txt minor change apply weight prepacking in set_tensor method in ggml-backend fix compile error ggml-ci minor change ggml-ci update CMakeLists.txt ggml-ci add march dependency minor change ggml-ci change ggml_backend_buffer_is_host to return false for amx backend ggml-ci fix supports_op use device reg for AMX backend ggml-ci minor change ggml-ci minor change fix rebase set .buffer_from_host_ptr to be false for AMX backend --- CMakeLists.txt | 4 + Makefile | 24 +- README.md | 2 +- ggml/CMakeLists.txt | 4 + ggml/include/ggml-amx.h | 25 + ggml/include/ggml.h | 1 + ggml/src/CMakeLists.txt | 42 + ggml/src/ggml-amx.cpp | 453 +++++++ ggml/src/ggml-amx/common.h | 93 ++ ggml/src/ggml-amx/mmq.cpp | 2509 ++++++++++++++++++++++++++++++++++++ ggml/src/ggml-amx/mmq.h | 17 + ggml/src/ggml-backend.cpp | 12 +- ggml/src/ggml.c | 8 + src/llama.cpp | 17 + 14 files changed, 3204 insertions(+), 7 deletions(-) create mode 100644 ggml/include/ggml-amx.h create mode 100644 ggml/src/ggml-amx.cpp create mode 100644 ggml/src/ggml-amx/common.h create mode 100644 ggml/src/ggml-amx/mmq.cpp create mode 100644 ggml/src/ggml-amx/mmq.h diff --git a/CMakeLists.txt b/CMakeLists.txt index 64a335378e596..ef0932a7b9277 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -88,6 +88,10 @@ if (NOT DEFINED GGML_LLAMAFILE) set(GGML_LLAMAFILE_DEFAULT ON) endif() +if (NOT DEFINED GGML_AMX) + set(GGML_AMX ON) +endif() + if (NOT DEFINED GGML_CUDA_GRAPHS) set(GGML_CUDA_GRAPHS_DEFAULT ON) endif() diff --git a/Makefile b/Makefile index 2793978c3e339..719f45d167463 100644 --- a/Makefile +++ b/Makefile @@ -93,11 +93,6 @@ GGML_METAL := 1 DEPRECATE_WARNING := 1 endif -ifdef LLAMA_OPENMP -GGML_OPENMP := 1 -DEPRECATE_WARNING := 1 -endif - ifdef LLAMA_RPC GGML_RPC := 1 DEPRECATE_WARNING := 1 @@ -584,6 +579,11 @@ ifndef GGML_NO_LLAMAFILE OBJ_GGML += ggml/src/llamafile/sgemm.o endif +ifndef GGML_NO_AMX + MK_CPPFLAGS += -DGGML_USE_AMX + OBJ_GGML += ggml/src/ggml-amx.o ggml/src/ggml-amx/mmq.o +endif + ifdef GGML_RPC MK_CPPFLAGS += -DGGML_USE_RPC OBJ_GGML += ggml/src/ggml-rpc.o @@ -1087,6 +1087,19 @@ ggml/src/llamafile/sgemm.o: \ $(CXX) $(CXXFLAGS) -c $< -o $@ endif # GGML_NO_LLAMAFILE +ifndef GGML_NO_AMX +ggml/src/ggml-amx.o: \ + ggml/src/ggml-amx.cpp \ + ggml/include/ggml-amx.h + $(CXX) $(CXXFLAGS) -c $< -o $@ + +ggml/src/ggml-amx/mmq.o: \ + ggml/src/ggml-amx/mmq.cpp \ + ggml/src/ggml-amx/mmq.h \ + ggml/include/ggml.h + $(CXX) $(CXXFLAGS) -c $< -o $@ +endif + ifdef GGML_RPC ggml/src/ggml-rpc.o: \ ggml/src/ggml-rpc.cpp \ @@ -1238,6 +1251,7 @@ clean: 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 -vrf ggml/src/ggml-amx/*.o rm -rvf $(BUILD_TARGETS) rm -rvf $(TEST_TARGETS) rm -f vulkan-shaders-gen ggml/src/ggml-vulkan-shaders.hpp ggml/src/ggml-vulkan-shaders.cpp diff --git a/README.md b/README.md index 707904ddcfcc6..1088b3338ff17 100644 --- a/README.md +++ b/README.md @@ -29,7 +29,7 @@ variety of hardware - locally and in the cloud. - Plain C/C++ implementation without any dependencies - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks -- AVX, AVX2 and AVX512 support for x86 architectures +- AVX, AVX2, AVX512 and AMX support for x86 architectures - 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use - Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads MTT GPUs via MUSA) - Vulkan and SYCL backend support diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt index 89fdf9d1c11ed..cfa6e3f70e4a3 100644 --- a/ggml/CMakeLists.txt +++ b/ggml/CMakeLists.txt @@ -99,6 +99,9 @@ option(GGML_AVX512 "ggml: enable AVX512" OFF) option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI" OFF) option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI" OFF) option(GGML_AVX512_BF16 "ggml: enable AVX512-BF16" OFF) +option(GGML_AMX_TILE "ggml: enable AMX-TILE" OFF) +option(GGML_AMX_INT8 "ggml: enable AMX-INT8" OFF) +option(GGML_AMX_BF16 "ggml: enable AMX-BF16" OFF) option(GGML_FMA "ggml: enable FMA" ${INS_ENB}) if (NOT MSVC) option(GGML_F16C "ggml: enable F16C" ${INS_ENB}) # in MSVC F16C is implied with AVX2/AVX512 @@ -158,6 +161,7 @@ set (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING set (GGML_METAL_STD "" CACHE STRING "ggml: metal standard version (-std flag)") option(GGML_OPENMP "ggml: use OpenMP" ON) option(GGML_RPC "ggml: use RPC" OFF) +option(GGML_AMX "ggml: use AMX" OFF) option(GGML_SYCL "ggml: use SYCL" OFF) option(GGML_SYCL_F16 "ggml: use 16 bit floats for sycl calculations" OFF) set (GGML_SYCL_TARGET "INTEL" CACHE STRING diff --git a/ggml/include/ggml-amx.h b/ggml/include/ggml-amx.h new file mode 100644 index 0000000000000..22b3f70f43a67 --- /dev/null +++ b/ggml/include/ggml-amx.h @@ -0,0 +1,25 @@ +#pragma once + +#include "ggml.h" +#include "ggml-backend.h" + + +#ifdef __cplusplus +extern "C" { +#endif + +// buffer_type API +GGML_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void); + +GGML_API bool ggml_backend_is_amx(ggml_backend_t backend); + +// backend API +GGML_API ggml_backend_t ggml_backend_amx_init(void); + +GGML_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads); + +GGML_API ggml_backend_reg_t ggml_backend_amx_reg(void); + +#ifdef __cplusplus +} +#endif diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 4508da4fb3a41..de3c706fc251c 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -2488,6 +2488,7 @@ extern "C" { GGML_API int ggml_cpu_has_avx512_vbmi(void); GGML_API int ggml_cpu_has_avx512_vnni(void); GGML_API int ggml_cpu_has_avx512_bf16(void); + GGML_API int ggml_cpu_has_amx_int8 (void); GGML_API int ggml_cpu_has_fma (void); GGML_API int ggml_cpu_has_neon (void); GGML_API int ggml_cpu_has_sve (void); diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index 676f85a369bc6..aa405e4d0fb1a 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -267,6 +267,26 @@ if (GGML_LLAMAFILE) set(GGML_SOURCES_LLAMAFILE llamafile/sgemm.cpp) endif() +if (GGML_AMX) + if (CMAKE_COMPILER_IS_GNUCC AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 11.0) + else() + set(GGML_AMX OFF) + message(WARNING "AMX requires gcc version > 11.0. Turning off GGML_AMX.") + endif() + + if (GGML_AMX) + message(STATUS "Using AMX") + + list(APPEND GGML_CDEF_PUBLIC GGML_USE_AMX) + + file(GLOB GGML_HEADERS_AMX "ggml-amx/*.h") + list(APPEND GGML_HEADERS_AMX "../include/ggml-amx.h") + + file(GLOB GGML_SOURCES_AMX "ggml-amx/*.cpp") + list(APPEND GGML_SOURCES_AMX "ggml-amx.cpp") + endif() +endif() + if (GGML_CUDA) cmake_minimum_required(VERSION 3.18) # for CMAKE_CUDA_ARCHITECTURES @@ -1180,6 +1200,18 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW add_compile_definitions($<$:__AVX512BF16__>) add_compile_definitions($<$:__AVX512BF16__>) endif() + if (GGML_AMX_TILE) + add_compile_definitions($<$:__AMX_TILE__>) + add_compile_definitions($<$:__AMX_TILE__>) + endif() + if (GGML_AMX_INT8) + add_compile_definitions($<$:__AMX_INT8__>) + add_compile_definitions($<$:__AMX_INT8__>) + endif() + if (GGML_AMX_BF16) + add_compile_definitions($<$:__AMX_BF16__>) + add_compile_definitions($<$:__AMX_BF16__>) + endif() elseif (GGML_AVX2) list(APPEND ARCH_FLAGS /arch:AVX2) elseif (GGML_AVX) @@ -1215,6 +1247,15 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW if (GGML_AVX512_BF16) list(APPEND ARCH_FLAGS -mavx512bf16) endif() + if (GGML_AMX_TILE) + list(APPEND ARCH_FLAGS -mamx-tile) + endif() + if (GGML_AMX_INT8) + list(APPEND ARCH_FLAGS -mamx-int8) + endif() + if (GGML_AMX_BF16) + list(APPEND ARCH_FLAGS -mamx-bf16) + endif() endif() elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64") message(STATUS "PowerPC detected") @@ -1340,6 +1381,7 @@ add_library(ggml ${GGML_SOURCES_ROCM} ${GGML_HEADERS_ROCM} ${GGML_SOURCES_BLAS} ${GGML_HEADERS_BLAS} ${GGML_SOURCES_LLAMAFILE} ${GGML_HEADERS_LLAMAFILE} + ${GGML_SOURCES_AMX} ${GGML_HEADERS_AMX} ${GGML_SOURCES_CANN} ${GGML_HEADERS_CANN} ggml-aarch64.c ggml-aarch64.h ) diff --git a/ggml/src/ggml-amx.cpp b/ggml/src/ggml-amx.cpp new file mode 100644 index 0000000000000..ac6ec23426e2c --- /dev/null +++ b/ggml/src/ggml-amx.cpp @@ -0,0 +1,453 @@ +#include "ggml-amx.h" +#include "ggml-amx/common.h" +#include "ggml-amx/mmq.h" +#include "ggml-backend-impl.h" +#include "ggml-impl.h" + +#if defined(__gnu_linux__) +#include +#include +#endif + +#include +#include +#include + +#if defined(__AMX_INT8__) + +// AMX buffer interface +static const char * ggml_backend_amx_buffer_get_name(ggml_backend_buffer_t buffer) { + return "AMX"; + + GGML_UNUSED(buffer); +} + +static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) { + free(buffer->context); +} + +static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) { + return (void *)(buffer->context); +} + +static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { + memset((char *)tensor->data + offset, value, size); + + GGML_UNUSED(buffer); +} + +static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { + if (qtype_has_amx_kernels(tensor->type)) { + ggml_backend_amx_convert_weight(tensor, data, offset, size); + } else { + memcpy((char *)tensor->data + offset, data, size); + } + + GGML_UNUSED(buffer); +} + +static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { + GGML_ASSERT(!qtype_has_amx_kernels(tensor->type)); + memcpy(data, (const char *)tensor->data + offset, size); + + GGML_UNUSED(buffer); +} + +static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) { + if (ggml_backend_buffer_is_host(src->buffer)) { + if (qtype_has_amx_kernels(src->type)) { + ggml_backend_amx_convert_weight(dst, src->data, 0, ggml_backend_amx_get_alloc_size(dst)); + } else { + memcpy(dst->data, src->data, ggml_nbytes(src)); + } + return true; + } + return false; + + GGML_UNUSED(buffer); +} + +static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { + memset(buffer->context, value, buffer->size); +} + +static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = { + /* .get_name = */ ggml_backend_amx_buffer_get_name, + /* .free_buffer = */ ggml_backend_amx_buffer_free_buffer, + /* .get_base = */ ggml_backend_amx_buffer_get_base, + /* .init_tensor = */ NULL, // no initialization required + /* .memset_tensor = */ ggml_backend_amx_buffer_memset_tensor, + /* .set_tensor = */ ggml_backend_amx_buffer_set_tensor, + /* .get_tensor = */ ggml_backend_amx_buffer_get_tensor, + /* .cpy_tensor = */ ggml_backend_amx_buffer_cpy_tensor, + /* .clear = */ ggml_backend_amx_buffer_clear, + /* .reset = */ NULL, +}; + +static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) { + return "AMX"; + + GGML_UNUSED(buft); +} + +static ggml_backend_buffer_t ggml_backend_amx_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { + void * data = aligned_alloc(TENSOR_ALIGNMENT, size); + if (data == NULL) { + fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size); + return NULL; + } + + return ggml_backend_buffer_init(buft, ggml_backend_amx_buffer_interface, data, size); +} + +static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { + return TENSOR_ALIGNMENT; + + GGML_UNUSED(buft); +} + +static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) { + return ggml_backend_amx_get_alloc_size(tensor); + + GGML_UNUSED(buft); +} + +static bool ggml_backend_amx_buffer_type_is_host(ggml_backend_buffer_type_t buft) { + return false; + + GGML_UNUSED(buft); +} + +ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() { + static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = { + /* .iface = */ { + /* .get_name = */ ggml_backend_amx_buffer_type_get_name, + /* .alloc_buffer = */ ggml_backend_amx_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_amx_buffer_type_get_alignment, + /* .get_max_size = */ NULL, // defaults to SIZE_MAX + /* .get_alloc_size = */ ggml_backend_amx_buffer_type_get_alloc_size, + /* .is_host = */ ggml_backend_amx_buffer_type_is_host, + }, + /* .device = */ NULL, + /* .context = */ NULL, + }; + + return &ggml_backend_buffer_type_amx; +} + +// backend interface + +static const char * ggml_backend_amx_name(ggml_backend_t backend) { + return "AMX"; + + GGML_UNUSED(backend); +} + +static void ggml_backend_amx_free(ggml_backend_t backend) { + ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context; + delete ctx; + delete backend; +} + +static ggml_backend_buffer_type_t ggml_backend_amx_get_default_buffer_type(ggml_backend_t backend) { + return ggml_backend_amx_buffer_type(); + + GGML_UNUSED(backend); +} + +static enum ggml_status ggml_backend_amx_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { + ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context; + + for (int i = 0; i < cgraph->n_nodes; i++) { + struct ggml_tensor * node = cgraph->nodes[i]; + + switch (node->op) { + case GGML_OP_MUL_MAT: + ggml_backend_amx_mul_mat(ctx, node); + break; + + case GGML_OP_NONE: + case GGML_OP_RESHAPE: + case GGML_OP_VIEW: + case GGML_OP_PERMUTE: + case GGML_OP_TRANSPOSE: + break; + + default: + fprintf(stderr, "%s: unsupported op %s\n", __func__, ggml_op_desc(node)); + GGML_ASSERT(false); + } + } + + return GGML_STATUS_SUCCESS; + + GGML_UNUSED(backend); +} + +static struct ggml_backend_i ggml_backend_amx_i = { + /* .get_name = */ ggml_backend_amx_name, + /* .free = */ ggml_backend_amx_free, + /* .get_default_buffer_type = */ ggml_backend_amx_get_default_buffer_type, + /* .set_tensor_async = */ NULL, + /* .get_tensor_async = */ NULL, + /* .cpy_tensor_async = */ NULL, + /* .synchronize = */ NULL, + /* .graph_plan_create = */ NULL, + /* .graph_plan_free = */ NULL, + /* .graph_plan_update = */ NULL, + /* .graph_plan_compute = */ NULL, + /* .graph_compute = */ ggml_backend_amx_graph_compute, + /* .supports_op = */ NULL, + /* .supports_buft = */ NULL, + /* .offload_op = */ NULL, + /* .event_record = */ NULL, + /* .event_wait = */ NULL, +}; + +static ggml_guid_t ggml_backend_amx_guid() { + static ggml_guid guid = { 0x13, 0xb8, 0xa4, 0xc4, 0xba, 0xfe, 0x51, 0x67, 0x87, 0x44, 0x55, 0x15, 0xb2, 0x35, 0x62, 0x3e }; + return &guid; +} + +#define ARCH_GET_XCOMP_PERM 0x1022 +#define ARCH_REQ_XCOMP_PERM 0x1023 +#define XFEATURE_XTILECFG 17 +#define XFEATURE_XTILEDATA 18 + +static bool ggml_amx_init() { +#if defined(__gnu_linux__) + if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) { + fprintf(stderr, "AMX is not ready to be used!\n"); + return false; + } + return true; +#elif defined(_WIN32) + return true; +#endif +} + +ggml_backend_t ggml_backend_amx_init() { + + // invoke a Linux system call to request access to AMX features + ggml_amx_init(); + + // backend context + ggml_backend_amx_context * ctx = new ggml_backend_amx_context; + + // ggml amx backend + ggml_backend_t backend = new ggml_backend { + /* .guid = */ ggml_backend_amx_guid(), + /* .interface = */ ggml_backend_amx_i, + /* .device = */ ggml_backend_reg_dev_get(ggml_backend_amx_reg(), 0), + /* .context = */ ctx, + }; + + return backend; +} + +bool ggml_backend_is_amx(ggml_backend_t backend) { + return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_amx_guid()); +} + +void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) { + GGML_ASSERT(ggml_backend_is_amx(backend_amx)); + + ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend_amx->context; + ctx->n_threads = n_threads; +} + +// device interface + +static const char * ggml_backend_amx_device_get_name(ggml_backend_dev_t dev) { + return "AMX"; + + GGML_UNUSED(dev); +} + +static const char * ggml_backend_amx_device_get_description(ggml_backend_dev_t dev) { + return "Intel Advanced Matrix Extensions"; + + GGML_UNUSED(dev); +} + +static void ggml_backend_amx_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) { + // TODO + *free = 0; + *total = 0; + + GGML_UNUSED(dev); +} + +static enum ggml_backend_dev_type ggml_backend_amx_device_get_type(ggml_backend_dev_t dev) { + return GGML_BACKEND_DEVICE_TYPE_CPU; + + GGML_UNUSED(dev); +} + +static void ggml_backend_amx_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) { + props->name = ggml_backend_amx_device_get_name(dev); + props->description = ggml_backend_amx_device_get_description(dev); + props->type = ggml_backend_amx_device_get_type(dev); + ggml_backend_amx_device_get_memory(dev, &props->memory_free, &props->memory_total); + + // `buffer_from_host_ptr` is intended to be used in mmap, when memory layout unchanged + props->caps = { + /* .async = */ false, + /* .host_buffer = */ false, + /* .buffer_from_host_ptr = */ false, + /* .events = */ false, + }; +} + +static ggml_backend_t ggml_backend_amx_device_init(ggml_backend_dev_t dev, const char * params) { + return ggml_backend_amx_init(); + + GGML_UNUSED(dev); + GGML_UNUSED(params); +} + +static ggml_backend_buffer_type_t ggml_backend_amx_device_get_buffer_type(ggml_backend_dev_t dev) { + return ggml_backend_amx_buffer_type(); + + GGML_UNUSED(dev); +} + +static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) { + + // handle only 2d gemm for now + auto is_contiguous_2d = [](const struct ggml_tensor * t) { + return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1; + }; + + switch (op->op) { + case GGML_OP_NONE: + case GGML_OP_RESHAPE: + case GGML_OP_VIEW: + case GGML_OP_PERMUTE: + case GGML_OP_TRANSPOSE: + return true; + + case GGML_OP_MUL_MAT: { + const struct ggml_tensor * src0 = op->src[0]; + const struct ggml_tensor * src1 = op->src[1]; + + const enum ggml_type type = src0->type; + const int64_t ne0 = op->ne[0]; + + bool is_training = src0->grad || src1->grad; + + // amx kernels enables for Q4_0, Q4_1, Q8_0, F16 + // Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256 + bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16); + + bool can_use_amx = + is_contiguous_2d(src0) && // src0 must be contiguous + is_contiguous_2d(src1) && // src1 must be contiguous + !is_training && // inference only + src1->type == GGML_TYPE_F32 && // src1 must be float32 + has_amx_kernels && // with amx kernel impls + ne0 % (TILE_N * 2) == 0; // out_features is 32x + + return can_use_amx; + } + default: + return false; + } + + GGML_UNUSED(dev); +} + +static bool ggml_backend_amx_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) { + return buft->iface.get_name == ggml_backend_amx_buffer_type_get_name; + + GGML_UNUSED(dev); +} + +static const struct ggml_backend_device_i ggml_backend_amx_device_i = { + /* .get_name = */ ggml_backend_amx_device_get_name, + /* .get_description = */ ggml_backend_amx_device_get_description, + /* .get_memory = */ ggml_backend_amx_device_get_memory, + /* .get_type = */ ggml_backend_amx_device_get_type, + /* .get_props = */ ggml_backend_amx_device_get_props, + /* .init_backend = */ ggml_backend_amx_device_init, + /* .get_buffer_type = */ ggml_backend_amx_device_get_buffer_type, + /* .get_host_buffer_type = */ NULL, + /* .buffer_from_host_ptr = */ NULL, + /* .supports_op = */ ggml_backend_amx_device_supports_op, + /* .supports_buft = */ ggml_backend_amx_device_supports_buft, + /* .offload_op = */ NULL, + /* .event_new = */ NULL, + /* .event_free = */ NULL, + /* .event_synchronize = */ NULL, +}; + +// backend reg interface + +static const char * ggml_backend_amx_reg_get_name(ggml_backend_reg_t reg) { + return "AMX"; + + GGML_UNUSED(reg); +} + +static size_t ggml_backend_amx_reg_get_device_count(ggml_backend_reg_t reg) { + return 1; + + GGML_UNUSED(reg); +} + +static ggml_backend_dev_t ggml_backend_amx_reg_get_device(ggml_backend_reg_t reg, size_t index) { + GGML_ASSERT(index == 0); + + static ggml_backend_device ggml_backend_amx_device = { + /* .iface = */ ggml_backend_amx_device_i, + /* .reg = */ reg, + /* .context = */ nullptr, + }; + + return &ggml_backend_amx_device; + + GGML_UNUSED(reg); + GGML_UNUSED(index); +} + +static void * ggml_backend_amx_get_proc_address(ggml_backend_reg_t reg, const char * name) { + if (std::strcmp(name, "ggml_backend_set_n_threads") == 0) { + return (void *)ggml_backend_amx_set_n_threads; + } + return NULL; + + GGML_UNUSED(reg); + GGML_UNUSED(name); +} + +static const struct ggml_backend_reg_i ggml_backend_amx_reg_i = { + /* .get_name = */ ggml_backend_amx_reg_get_name, + /* .get_device_count = */ ggml_backend_amx_reg_get_device_count, + /* .get_device = */ ggml_backend_amx_reg_get_device, + /* .get_proc_address = */ ggml_backend_amx_get_proc_address, +}; + +ggml_backend_reg_t ggml_backend_amx_reg(void) { + static struct ggml_backend_reg ggml_backend_amx_reg = { + /* .iface = */ ggml_backend_amx_reg_i, + /* .context = */ NULL, + }; + + return &ggml_backend_amx_reg; +} + +#else // if defined(__AMX_INT8__) + +ggml_backend_t ggml_backend_amx_init(void) { + fprintf(stderr, "GGML is not compiled with AMX support!\n"); + return ggml_backend_t{}; +} + +void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) { + fprintf(stderr, "GGML is not compiled with AMX support!\n"); + + GGML_UNUSED(backend_amx); + GGML_UNUSED(n_threads); +} + +#endif diff --git a/ggml/src/ggml-amx/common.h b/ggml/src/ggml-amx/common.h new file mode 100644 index 0000000000000..2b6c635270451 --- /dev/null +++ b/ggml/src/ggml-amx/common.h @@ -0,0 +1,93 @@ +#pragma once + +#include "ggml.h" +#include "ggml-cpu-impl.h" // + +#include +#include +#include + +#if defined(_OPENMP) +#include +#endif + +#define TILE_M 16 +#define TILE_N 16 +#define TILE_K 32 +#define VNNI_BLK 4 + +#define AMX_BLK_SIZE 32 + +#define TMM0 0 +#define TMM1 1 +#define TMM2 2 +#define TMM3 3 +#define TMM4 4 +#define TMM5 5 +#define TMM6 6 +#define TMM7 7 + +// parallel routines +template ::value, int>::type = 0> +inline T div_up(T x, T y) { return (x + y - 1) / y; } + +template +inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) { +#if 0 + // onednn partition pattern + T& n_my = n_end; + if (nth <= 1 || n == 0) { + n_start = 0; + n_my = n; + } else { + T n1 = div_up(n, nth); + T n2 = n1 - 1; + T T1 = n - n2 * nth; + n_my = ith < T1 ? n1 : n2; + n_start = ith <= T1 ? ith*n1 : T1 * n1 + (ith - T1) * n2; + } + n_end += n_start; +#else + // pytorch aten partition pattern + T n_my = div_up(n, nth); + n_start = ith * n_my; + n_end = std::min(n_start + n_my, n); +#endif +} + +template +inline void parallel_for(int nth, int n, const func_t& f) { +#if defined(_OPENMP) +#pragma omp parallel num_threads(nth) +{ + //int nth = omp_get_num_threads(); + int ith = omp_get_thread_num(); + int tbegin, tend; + balance211(n, nth, ith, tbegin, tend); + f(tbegin, tend); +} +#else + f(0, n); + + GGML_UNUSED(nth); +#endif +} + +// quantized types that have AMX support +inline bool qtype_has_amx_kernels(const enum ggml_type type) { + // TODO: fix padding for vnni format + return (type == GGML_TYPE_Q4_0) || + (type == GGML_TYPE_Q4_1); + //(type == GGML_TYPE_Q8_0) || + //(type == GGML_TYPE_Q4_K) || + //(type == GGML_TYPE_Q5_K) || + //(type == GGML_TYPE_Q6_K) || + //(type == GGML_TYPE_IQ4_XS); +} + +// ggml backend context +struct ggml_backend_amx_context { + int n_threads = GGML_DEFAULT_N_THREADS; + std::unique_ptr work_data; + size_t work_size = 0; +}; diff --git a/ggml/src/ggml-amx/mmq.cpp b/ggml/src/ggml-amx/mmq.cpp new file mode 100644 index 0000000000000..239d15121a602 --- /dev/null +++ b/ggml/src/ggml-amx/mmq.cpp @@ -0,0 +1,2509 @@ + +#if defined(__GNUC__) +#pragma GCC diagnostic ignored "-Wpedantic" +#pragma GCC diagnostic ignored "-Wunused-local-typedefs" +#endif + +#include "mmq.h" +#include "ggml-impl.h" +#include "ggml-quants.h" +#include +#include + +#if defined(__gnu_linux__) +#include +#include +#endif + +#if defined(_OPENMP) +#include +#endif + +#if (defined(_WIN32) || defined(_WIN64)) +#define RESTRICT __restrict +#else +#define RESTRICT __restrict__ +#endif + +#if (defined(_WIN32) || defined(_WIN64)) +#define ALWAYS_INLINE __forceinline +#elif __has_attribute(always_inline) || defined(__GNUC__) +#define ALWAYS_INLINE __attribute__((__always_inline__)) inline +#else +#define ALWAYS_INLINE inline +#endif + +#if defined(__AMX_INT8__) + +namespace { + +// Forced unrolling +template +struct Unroll { + template + ALWAYS_INLINE void operator()(const Func& f, Args... args) const { + Unroll{}(f, args...); + f(std::integral_constant{}, args...); + } +}; + +template <> +struct Unroll<1> { + template + ALWAYS_INLINE void operator()(const Func& f, Args... args) const { + f(std::integral_constant{}, args...); + } +}; + +// type traits +template struct PackedTypes {}; +template <> struct PackedTypes { using type = int8_t; }; +template <> struct PackedTypes { using type = uint8_t; }; +template <> struct PackedTypes { using type = int8_t; }; +template using packed_B_type = typename PackedTypes::type; + +template +struct do_compensate : std::integral_constant::value> {}; + +template +struct do_unpack : std::integral_constant::value || + std::is_same::value> {}; + +template +struct is_type_qkk : std::integral_constant::value || + std::is_same::value || + std::is_same::value || + std::is_same::value> {}; + +#define GGML_DISPATCH_FLOATING_TYPES(TYPE, ...) \ + [&] { \ + switch (TYPE) { \ + case GGML_TYPE_F16: { \ + using type = ggml_fp16_t; \ + constexpr int blck_size = 16; \ + return __VA_ARGS__(); \ + } \ + case GGML_TYPE_BF16: { \ + using type = ggml_bf16_t; \ + constexpr int blck_size = 32; \ + return __VA_ARGS__(); \ + } \ + default: \ + fprintf(stderr, "Unsupported floating data type\n"); \ + } \ + }() + +#define GGML_DISPATCH_QTYPES(QT, ...) \ + [&] { \ + switch (QT) { \ + case GGML_TYPE_Q4_0: { \ + using type = block_q4_0; \ + using vec_dot_type = block_q8_0; \ + constexpr int blck_size = QK4_0; \ + return __VA_ARGS__(); \ + } \ + case GGML_TYPE_Q4_1: { \ + using type = block_q4_1; \ + using vec_dot_type = block_q8_1; \ + constexpr int blck_size = QK4_1; \ + return __VA_ARGS__(); \ + } \ + case GGML_TYPE_Q8_0: { \ + using type = block_q8_0; \ + using vec_dot_type = block_q8_0; \ + constexpr int blck_size = QK8_0; \ + return __VA_ARGS__(); \ + } \ + case GGML_TYPE_Q4_K: { \ + using type = block_q4_K; \ + using vec_dot_type = block_q8_K; \ + constexpr int blck_size = QK_K; \ + return __VA_ARGS__(); \ + } \ + case GGML_TYPE_Q5_K: { \ + using type = block_q5_K; \ + using vec_dot_type = block_q8_K; \ + constexpr int blck_size = QK_K; \ + return __VA_ARGS__(); \ + } \ + case GGML_TYPE_Q6_K: { \ + using type = block_q6_K; \ + using vec_dot_type = block_q8_K; \ + constexpr int blck_size = QK_K; \ + return __VA_ARGS__(); \ + } \ + case GGML_TYPE_IQ4_XS: { \ + using type = block_iq4_xs; \ + using vec_dot_type = block_q8_K; \ + constexpr int blck_size = QK_K; \ + return __VA_ARGS__(); \ + } \ + default: \ + fprintf(stderr, "Unsupported quantized data type: %d\n", int(TYPE)); \ + } \ + }() + +#define GGML_DISPATCH_BOOL(BOOL_V, BOOL_NAME, ...) \ + [&] { \ + if (BOOL_V) { \ + constexpr bool BOOL_NAME = true; \ + return __VA_ARGS__(); \ + } else { \ + constexpr bool BOOL_NAME = false; \ + return __VA_ARGS__(); \ + } \ + }() + +// define amx tile config data structure +struct tile_config_t{ + uint8_t palette_id = 0; + uint8_t start_row = 0; + uint8_t reserved_0[14] = {0}; + uint16_t colsb[16] = {0}; + uint8_t rows[16] = {0}; +}; + +// Notes: amx tile config +// +// Typically, TMUL calculates A and B of size 16 x 64 containing INT8 values, +// and accumulate the result to a 16 x 16 matrix C containing INT32 values, +// +// As many GGUF quantized types as `block_size` of 32, so a 16-16-32 config is used +// instead of the normally used 16-16-64 config. +// +// Block A: {16, 32}, dtype = int8_t +// Block B: {16, 32}, dtype = uint8_t/int8_t +// Block C: {16, 16}, dtype = int32_t +// +// Block B needs to be prepacked to vnni format before feeding into TMUL: +// packed_B: from {n, k} to {k/vnni_blk, n, vnni_blck}, viewed in 2d, we get {8, 64} +// +// Therefore, we get tileconfig: +// A B C +// rows 16 8 16 +// colsb 32 64 16 +// +// For tile distribution, follow a 2-2-4 pattern, e.g. A used TMM2-TMM3, B used TMM0-TMM1, +// C used TMM4-TMM7: +// B TMM0 B TMM1 +// A TMM2 C TMM4 C TMM6 +// A TMM3 C TMM5 C TMM7 +// +// Each `amx` kernel handles 4 blocks at a time: 2MB * 2NB, when m < 2 * BLOCK_M, unpack A +// will be needed. +// +// Here another commonly used pattern 1-3-3 is skipped, as it is mostly used when m <=16; +// and the sinlge batch gemm (m=1) has a special fast path with `avx512-vnni`. +// +// ref: https://www.intel.com/content/www/us/en/developer/articles/code-sample/ +// advanced-matrix-extensions-intrinsics-functions.html +// + +#define TC_CONFIG_TILE(i, r, cb) tc.rows[i] = r; tc.colsb[i] = cb +void ggml_tile_config_init(void) { + static thread_local bool is_first_time = true; + + if (!is_first_time) { + return; + } + + static thread_local tile_config_t tc; + tile_config_t current_tc; + _tile_storeconfig(¤t_tc); + + // load only when config changes + if (tc.palette_id == 0 || (memcmp(¤t_tc.colsb, &tc.colsb, sizeof(uint16_t) * 8) != 0 && + memcmp(¤t_tc.rows, &tc.rows, sizeof(uint8_t) * 8) != 0)) { + tc.palette_id = 1; + tc.start_row = 0; + TC_CONFIG_TILE(TMM0, 8, 64); + TC_CONFIG_TILE(TMM1, 8, 64); + TC_CONFIG_TILE(TMM2, 16, 32); + TC_CONFIG_TILE(TMM3, 16, 32); + TC_CONFIG_TILE(TMM4, 16, 64); + TC_CONFIG_TILE(TMM5, 16, 64); + TC_CONFIG_TILE(TMM6, 16, 64); + TC_CONFIG_TILE(TMM7, 16, 64); + _tile_loadconfig(&tc); + } + + is_first_time = false; +} + +// we need an extra 16 * 4B (TILE_N * int32_t) for each NB/KB block for compensation. +// See the notes `s8s8 igemm compensation in avx512-vnni` for detail. +template +int get_tile_size() { + int tile_size = TILE_N * sizeof(TB); + if (do_compensate::value) { + tile_size += TILE_N * sizeof(int32_t); + } + if (std::is_same::value || + std::is_same::value) { + tile_size += TILE_N * 4; + } + if (std::is_same::value) { + tile_size += TILE_N * 2; + } + return tile_size; +} + +template +int get_row_size(int K) { + int KB = K / BLOCK_K; + int row_size = KB * sizeof(TB); + if (do_compensate::value) { + row_size += KB * sizeof(int32_t); + } + if (std::is_same::value || + std::is_same::value) { + row_size += KB * 4; + } + if (std::is_same::value) { + row_size += KB * 2; + } + return row_size; +} + +// vectorized dtype conversion +inline float FP16_TO_FP32(ggml_half val) { + __m256i v = _mm256_setr_epi16( + val, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0); + __m512 o = _mm512_cvtph_ps(v); + return _mm512_cvtss_f32(o); +} + +inline __m512 FP16_TO_FP32_VEC(ggml_half val) { + __m256i v = _mm256_set1_epi16(val); + return _mm512_cvtph_ps(v); +} + +// horizontal reduce +inline float _mm512_reduce_max_ps(const __m512 x) { + __m512 v = x; + __m512 v1 = _mm512_shuffle_f32x4(v, v, 0x4E); + v = _mm512_max_ps(v, v1); + v1 = _mm512_shuffle_f32x4(v, v, 0xB1); + v = _mm512_max_ps(v, v1); + v1 = _mm512_shuffle_ps(v, v, 0x4E); + v = _mm512_max_ps(v, v1); + v1 = _mm512_shuffle_ps(v, v, 0xB1); + v = _mm512_max_ps(v, v1); + return _mm512_cvtss_f32(v); +} + +// transpose utils +#define SHUFFLE_EPI32(a, b, mask) \ + _mm256_castps_si256(_mm256_shuffle_ps(_mm256_castsi256_ps(a), _mm256_castsi256_ps(b), mask)) +inline void transpose_8x8_32bit(__m256i * v, __m256i * v1) { + // unpacking and 32-bit elements + v1[0] = _mm256_unpacklo_epi32(v[0], v[1]); + v1[1] = _mm256_unpackhi_epi32(v[0], v[1]); + v1[2] = _mm256_unpacklo_epi32(v[2], v[3]); + v1[3] = _mm256_unpackhi_epi32(v[2], v[3]); + v1[4] = _mm256_unpacklo_epi32(v[4], v[5]); + v1[5] = _mm256_unpackhi_epi32(v[4], v[5]); + v1[6] = _mm256_unpacklo_epi32(v[6], v[7]); + v1[7] = _mm256_unpackhi_epi32(v[6], v[7]); + + // shuffling the 32-bit elements + v[0] = SHUFFLE_EPI32(v1[0], v1[2], 0x44); + v[1] = SHUFFLE_EPI32(v1[0], v1[2], 0xee); + v[2] = SHUFFLE_EPI32(v1[4], v1[6], 0x44); + v[3] = SHUFFLE_EPI32(v1[4], v1[6], 0xee); + v[4] = SHUFFLE_EPI32(v1[1], v1[3], 0x44); + v[5] = SHUFFLE_EPI32(v1[1], v1[3], 0xee); + v[6] = SHUFFLE_EPI32(v1[5], v1[7], 0x44); + v[7] = SHUFFLE_EPI32(v1[5], v1[7], 0xee); + + // shuffling 128-bit elements + v1[0] = _mm256_permute2f128_si256(v[2], v[0], 0x02); + v1[1] = _mm256_permute2f128_si256(v[3], v[1], 0x02); + v1[2] = _mm256_permute2f128_si256(v[6], v[4], 0x02); + v1[3] = _mm256_permute2f128_si256(v[7], v[5], 0x02); + v1[4] = _mm256_permute2f128_si256(v[2], v[0], 0x13); + v1[5] = _mm256_permute2f128_si256(v[3], v[1], 0x13); + v1[6] = _mm256_permute2f128_si256(v[6], v[4], 0x13); + v1[7] = _mm256_permute2f128_si256(v[7], v[5], 0x13); +} + +inline void transpose_16x4_32bit(__m512i * r, __m512i * d) { + + static const __m512i index1 = _mm512_set_epi32( + 0x0f, 0x0b, 0x07, 0x03, + 0x0e, 0x0a, 0x06, 0x02, + 0x0d, 0x09, 0x05, 0x01, + 0x0c, 0x08, 0x04, 0x00); + + d[0] = _mm512_permutexvar_epi32(index1, r[0]); + d[1] = _mm512_permutexvar_epi32(index1, r[1]); + d[2] = _mm512_permutexvar_epi32(index1, r[2]); + d[3] = _mm512_permutexvar_epi32(index1, r[3]); + + r[0] = _mm512_shuffle_i32x4(d[0], d[1], 0x44); + r[1] = _mm512_shuffle_i32x4(d[0], d[1], 0xee); + r[2] = _mm512_shuffle_i32x4(d[2], d[3], 0x44); + r[3] = _mm512_shuffle_i32x4(d[2], d[3], 0xee); + + d[0] = _mm512_shuffle_i32x4(r[0], r[2], 0x88); + d[1] = _mm512_shuffle_i32x4(r[0], r[2], 0xdd); + d[2] = _mm512_shuffle_i32x4(r[1], r[3], 0x88); + d[3] = _mm512_shuffle_i32x4(r[1], r[3], 0xdd); +} + +inline void transpose_16x16_32bit(__m512i * v) { + __m512i v1[16]; + v1[0] = _mm512_unpacklo_epi32(v[0], v[1]); + v1[1] = _mm512_unpackhi_epi32(v[0], v[1]); + v1[2] = _mm512_unpacklo_epi32(v[2], v[3]); + v1[3] = _mm512_unpackhi_epi32(v[2], v[3]); + v1[4] = _mm512_unpacklo_epi32(v[4], v[5]); + v1[5] = _mm512_unpackhi_epi32(v[4], v[5]); + v1[6] = _mm512_unpacklo_epi32(v[6], v[7]); + v1[7] = _mm512_unpackhi_epi32(v[6], v[7]); + v1[8] = _mm512_unpacklo_epi32(v[8], v[9]); + v1[9] = _mm512_unpackhi_epi32(v[8], v[9]); + v1[10] = _mm512_unpacklo_epi32(v[10], v[11]); + v1[11] = _mm512_unpackhi_epi32(v[10], v[11]); + v1[12] = _mm512_unpacklo_epi32(v[12], v[13]); + v1[13] = _mm512_unpackhi_epi32(v[12], v[13]); + v1[14] = _mm512_unpacklo_epi32(v[14], v[15]); + v1[15] = _mm512_unpackhi_epi32(v[14], v[15]); + + v[0] = _mm512_unpacklo_epi64(v1[0], v1[2]); + v[1] = _mm512_unpackhi_epi64(v1[0], v1[2]); + v[2] = _mm512_unpacklo_epi64(v1[1], v1[3]); + v[3] = _mm512_unpackhi_epi64(v1[1], v1[3]); + v[4] = _mm512_unpacklo_epi64(v1[4], v1[6]); + v[5] = _mm512_unpackhi_epi64(v1[4], v1[6]); + v[6] = _mm512_unpacklo_epi64(v1[5], v1[7]); + v[7] = _mm512_unpackhi_epi64(v1[5], v1[7]); + v[8] = _mm512_unpacklo_epi64(v1[8], v1[10]); + v[9] = _mm512_unpackhi_epi64(v1[8], v1[10]); + v[10] = _mm512_unpacklo_epi64(v1[9], v1[11]); + v[11] = _mm512_unpackhi_epi64(v1[9], v1[11]); + v[12] = _mm512_unpacklo_epi64(v1[12], v1[14]); + v[13] = _mm512_unpackhi_epi64(v1[12], v1[14]); + v[14] = _mm512_unpacklo_epi64(v1[13], v1[15]); + v[15] = _mm512_unpackhi_epi64(v1[13], v1[15]); + + v1[0] = _mm512_shuffle_i32x4(v[0], v[4], 0x88); + v1[1] = _mm512_shuffle_i32x4(v[1], v[5], 0x88); + v1[2] = _mm512_shuffle_i32x4(v[2], v[6], 0x88); + v1[3] = _mm512_shuffle_i32x4(v[3], v[7], 0x88); + v1[4] = _mm512_shuffle_i32x4(v[0], v[4], 0xdd); + v1[5] = _mm512_shuffle_i32x4(v[1], v[5], 0xdd); + v1[6] = _mm512_shuffle_i32x4(v[2], v[6], 0xdd); + v1[7] = _mm512_shuffle_i32x4(v[3], v[7], 0xdd); + v1[8] = _mm512_shuffle_i32x4(v[8], v[12], 0x88); + v1[9] = _mm512_shuffle_i32x4(v[9], v[13], 0x88); + v1[10] = _mm512_shuffle_i32x4(v[10], v[14], 0x88); + v1[11] = _mm512_shuffle_i32x4(v[11], v[15], 0x88); + v1[12] = _mm512_shuffle_i32x4(v[8], v[12], 0xdd); + v1[13] = _mm512_shuffle_i32x4(v[9], v[13], 0xdd); + v1[14] = _mm512_shuffle_i32x4(v[10], v[14], 0xdd); + v1[15] = _mm512_shuffle_i32x4(v[11], v[15], 0xdd); + + v[0] = _mm512_shuffle_i32x4(v1[0], v1[8], 0x88); + v[1] = _mm512_shuffle_i32x4(v1[1], v1[9], 0x88); + v[2] = _mm512_shuffle_i32x4(v1[2], v1[10], 0x88); + v[3] = _mm512_shuffle_i32x4(v1[3], v1[11], 0x88); + v[4] = _mm512_shuffle_i32x4(v1[4], v1[12], 0x88); + v[5] = _mm512_shuffle_i32x4(v1[5], v1[13], 0x88); + v[6] = _mm512_shuffle_i32x4(v1[6], v1[14], 0x88); + v[7] = _mm512_shuffle_i32x4(v1[7], v1[15], 0x88); + v[8] = _mm512_shuffle_i32x4(v1[0], v1[8], 0xdd); + v[9] = _mm512_shuffle_i32x4(v1[1], v1[9], 0xdd); + v[10] = _mm512_shuffle_i32x4(v1[2], v1[10], 0xdd); + v[11] = _mm512_shuffle_i32x4(v1[3], v1[11], 0xdd); + v[12] = _mm512_shuffle_i32x4(v1[4], v1[12], 0xdd); + v[13] = _mm512_shuffle_i32x4(v1[5], v1[13], 0xdd); + v[14] = _mm512_shuffle_i32x4(v1[6], v1[14], 0xdd); + v[15] = _mm512_shuffle_i32x4(v1[7], v1[15], 0xdd); +} + +void quantize_row_q8_K_vnni(const float * RESTRICT x, void * RESTRICT vy, int64_t k) { + assert(k % QK_K == 0); + const int KB = k / QK_K; + constexpr int kVecs = QK_K / 16; + + block_q8_K * y = reinterpret_cast(vy); + + // hold 16 float vecs from x + __m512 v[kVecs]; + + // hold the quants vecs + __m512i vq[kVecs / 4]; + + // hold the packed quants vecs + __m512i vq_packed[kVecs / 4]; + + const __m512 signBit = _mm512_set1_ps(-0.f); + + for (int i = 0; i < KB; ++i) { + // Compute max(abs(e)) for the block + __m512 vamax = _mm512_set1_ps(0.f); + for (int j = 0; j < kVecs; ++j) { + v[j] = _mm512_loadu_ps(x); x += 16; + vamax = _mm512_max_ps(vamax, _mm512_andnot_ps(signBit, v[j])); + } + const float amax = _mm512_reduce_max_ps(vamax); + + // Quantize these floats + const float iscale = 127.f / amax; + y[i].d = GGML_FP32_TO_FP16(1 / iscale); + const float id = ( amax != 0.0f ) ? iscale : 0.f; + const __m512 vscale = _mm512_set1_ps(id); + + // Apply multiplier and round to nearest integer + for (int j = 0; j < kVecs; ++j) { + v[j] = _mm512_mul_ps(v[j], vscale); + v[j] = _mm512_roundscale_ps(v[j], (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + } + + // Pack to epi8 vecs + for (int j = 0; j < kVecs / 4; ++j) { + __m128i q8_0 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 0])); + __m128i q8_1 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 1])); + __m128i q8_2 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 2])); + __m128i q8_3 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 3])); + + __m256i q8_01 = _mm256_insertf128_si256(_mm256_castsi128_si256(q8_0), (q8_1), 1); + __m256i q8_23 = _mm256_insertf128_si256(_mm256_castsi128_si256(q8_2), (q8_3), 1); + + vq[j] = _mm512_inserti32x8(_mm512_castsi256_si512(q8_01), q8_23, 1); + _mm512_storeu_si512((__m512i *)(y[i].qs + j * 64), vq[j]); + } + + // Compute the bsums with vnni + transpose_16x4_32bit(vq, vq_packed); + + const __m512i one = _mm512_set1_epi8(1); + __m512i sum = _mm512_setzero_si512(); + for (int k = 0; k < 4; ++k) { + sum = _mm512_dpbusd_epi32(sum, one, vq_packed[k]); + } + _mm256_storeu_si256((__m256i *)(y[i].bsums), _mm512_cvtepi32_epi16(sum)); + } +} + +// quantize A from float to `vec_dot_type` +template +inline void from_float(const float * x, char * vy, int64_t k); + +template <> +inline void from_float(const float * x, char * vy, int64_t k) { + quantize_row_q8_0(x, vy, k); +} + +template <> +inline void from_float(const float * x, char * vy, int64_t k) { + quantize_row_q8_1(x, vy, k); +} + +template <> +inline void from_float(const float * x, char * vy, int64_t k) { +#if 1 + // TODO: this is reference impl! + quantize_row_q8_K(x, vy, k); +#else + quantize_row_q8_K_vnni(x, vy, k); +#endif +} + +// load A from memory to array when nrows can not fill in whole tile +void unpack_A(int8_t * RESTRICT tile, const block_q8_0 * RESTRICT A, int lda, int nr) { + assert(nr != TILE_M); + for (int m = 0; m < nr; ++m) { + const __m256i v = _mm256_loadu_si256((const __m256i *)(A[m * lda].qs)); + _mm256_storeu_si256((__m256i *)(tile + m * TILE_K), v); + } +} + +void unpack_A(int8_t * RESTRICT tile, const block_q8_1 * RESTRICT A, int lda, int nr) { + assert(nr != TILE_M); + for (int m = 0; m < nr; ++m) { + const __m256i v = _mm256_loadu_si256((const __m256i *)(A[m * lda].qs)); + _mm256_storeu_si256((__m256i *)(tile + m * TILE_K), v); + } +} + +template +void unpack_A(int8_t * RESTRICT tile, const block_q8_K * RESTRICT A, int lda, int k, int nr) { + assert(nr <= TILE_M); + for (int m = 0; m < nr; ++m) { + const __m256i v = _mm256_loadu_si256((const __m256i *)(A[m * lda].qs + k * 32)); + _mm256_storeu_si256((__m256i *)(tile + m * TILE_K), v); + } +} + +template <> +void unpack_A(int8_t * RESTRICT tile, const block_q8_K * RESTRICT A, int lda, int k, int nr) { + assert(nr <= TILE_M); + // zero padding k from 16 to 32, so that we don't have to re-config amx + const __m128i zero = _mm_setzero_si128(); + for (int m = 0; m < nr; ++m) { + const __m128i v = _mm_loadu_si128((const __m128i *)(A[m * lda].qs + k * 16)); + const __m256i r = _mm256_insertf128_si256(_mm256_castsi128_si256(v), zero, 1); + _mm256_storeu_si256((__m256i *)(tile + m * TILE_K), r); + } +} + +#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1) +inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) { + const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi); + const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp); + const __m256i lowMask = _mm256_set1_epi8(0xF); + return _mm256_and_si256(lowMask, bytes); +} + +// used for block_q4_K +inline __m512i bytes_from_nibbles_64(const uint8_t * rsi) { + const __m256i tmp = _mm256_loadu_si256((const __m256i *)rsi); + const __m256i lowMask = _mm256_set1_epi8(0xF); + const __m256i q4l = _mm256_and_si256(tmp, lowMask); + const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(tmp, 4), lowMask); + return _mm512_inserti32x8(_mm512_castsi256_si512(q4l), q4h, 1); +} + +// used for block_q5_K +inline __m512i bytes_from_nibbles_64(const uint8_t * qs, const uint8_t * qh, int k) { + const __m256i lowMask = _mm256_set1_epi8(0xF); + __m256i hmask = _mm256_set1_epi8(1); + hmask = _mm256_slli_epi16(hmask, k); + + const __m256i q5bits = _mm256_loadu_si256((const __m256i *)qs); + const __m256i hbits = _mm256_loadu_si256((const __m256i *)qh); + + const __m256i q5l_0 = _mm256_and_si256(q5bits, lowMask); + const __m256i q5h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), k + 0), 4); + const __m256i q5_0 = _mm256_add_epi8(q5l_0, q5h_0); + hmask = _mm256_slli_epi16(hmask, 1); + + const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), lowMask); + const __m256i q5h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), k + 1), 4); + const __m256i q5_1 = _mm256_add_epi8(q5l_1, q5h_1); + + return _mm512_inserti32x8(_mm512_castsi256_si512(q5_0), q5_1, 1); +} + +// used for block_q6_K +inline void bytes_from_nibbles_128(__m512i& r0, __m512i& r1, const uint8_t * qs, const uint8_t * qh) { + const __m256i m4 = _mm256_set1_epi8(0xF); + const __m256i m2 = _mm256_set1_epi8(0x3); + + const __m256i q6bits1 = _mm256_loadu_si256((const __m256i *)qs); + const __m256i q6bits2 = _mm256_loadu_si256((const __m256i *)(qs + 32)); + const __m256i q6bitsH = _mm256_loadu_si256((const __m256i *)qh); + + const __m256i q6h_0 = _mm256_slli_epi16(_mm256_and_si256( q6bitsH, m2), 4); + const __m256i q6h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q6bitsH, 2), m2), 4); + const __m256i q6h_2 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q6bitsH, 4), m2), 4); + const __m256i q6h_3 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q6bitsH, 6), m2), 4); + + const __m256i q6_0 = _mm256_or_si256(_mm256_and_si256(q6bits1, m4), q6h_0); + const __m256i q6_1 = _mm256_or_si256(_mm256_and_si256(q6bits2, m4), q6h_1); + const __m256i q6_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q6bits1, 4), m4), q6h_2); + const __m256i q6_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q6bits2, 4), m4), q6h_3); + + r0 = _mm512_inserti32x8(_mm512_castsi256_si512(q6_0), q6_1, 1); + r1 = _mm512_inserti32x8(_mm512_castsi256_si512(q6_2), q6_3, 1); +} + +inline __m512i packNibbles(__m512i r0, __m512i r1) { + return _mm512_or_si512(r0, _mm512_slli_epi16(r1, 4)); +} + +template +inline void pack_qs(void * RESTRICT packed_B, const TB * RESTRICT B, int KB) { + int8_t tmp[8 * 64]; + __m256i v[8], v2[8]; + for (int n = 0; n < 8; ++n) { + v[n] = bytes_from_nibbles_32(B[n * KB].qs); + } + transpose_8x8_32bit(v, v2); + for (int n = 0; n < 8; ++n) { + _mm256_storeu_si256((__m256i *)(tmp + n * 64), v2[n]); + } + for (int n = 0; n < 8; ++n) { + v[n] = bytes_from_nibbles_32(B[(n + 8) * KB].qs); + } + transpose_8x8_32bit(v, v2); + for (int n = 0; n < 8; ++n) { + _mm256_storeu_si256((__m256i *)(tmp + n * 64 + 32), v2[n]); + } + + // pack again with 128 to fully utilize vector length + for (int n = 0; n < 8; n += 2) { + __m512i r0 = _mm512_loadu_si512((const __m512i *)(tmp + n * 64)); + __m512i r1 = _mm512_loadu_si512((const __m512i *)(tmp + n * 64 + 64)); + __m512i r1r0 = packNibbles(r0, r1); + _mm512_storeu_si512((__m512i *)((char *)packed_B + n * 32), r1r0); + } +} + +template <> +inline void pack_qs(void * RESTRICT packed_B, const block_q8_0 * RESTRICT B, int KB) { + __m256i v[8], v2[8]; + for (int n = 0; n < 8; ++n) { + v[n] = _mm256_loadu_si256((const __m256i *)(B[n * KB].qs)); + } + transpose_8x8_32bit(v, v2); + for (int n = 0; n < 8; ++n) { + _mm256_storeu_si256((__m256i *)((char *)packed_B + n * 64), v2[n]); + } + for (int n = 0; n < 8; ++n) { + v[n] = _mm256_loadu_si256((const __m256i *)(B[(n + 8) * KB].qs)); + } + transpose_8x8_32bit(v, v2); + for (int n = 0; n < 8; ++n) { + _mm256_storeu_si256((__m256i *)((char *)packed_B + n * 64 + 32), v2[n]); + } +} + +template <> +inline void pack_qs(void * RESTRICT packed_B, const block_q4_K * RESTRICT B, int KB) { + __m512i v[16]; + // QK_K 256 with 8 groups, handle 2 groups at a time + char * pb = (char *)packed_B; + for (int k = 0; k < QK_K / 64; ++k) { + // pack 2 groups { n, g, k} to {g, k/4, 4n} + // e.g. {16, 2, 32} to {2, 8, 64} + for (int n = 0; n < TILE_N; ++n) { + v[n] = bytes_from_nibbles_64(B[n * KB].qs + k * 32); + } + + transpose_16x16_32bit(v); + + // pack again with 128 to fully utilize vector length + for (int n = 0; n < TILE_N; n += 2) { + _mm512_storeu_si512((__m512i *)pb, packNibbles(v[n], v[n + 1])); + pb += 64; + } + } +} + +template <> +inline void pack_qs(void * RESTRICT packed_B, const block_q5_K * RESTRICT B, int KB) { + __m512i v[16]; + const __m512i lowMask = _mm512_set1_epi8(0xF); + // QK_K 256 with 8 groups, handle 2 groups at a time + char * pb = (char *)packed_B; + char * ph = (char *)packed_B + (QK_K / 2) * TILE_N; + for (int k = 0; k < QK_K / 64; ++k) { + // pack 2 groups { n, g, k} to {g, k/4, 4n} + // e.g. {16, 2, 32} to {2, 8, 64} + for (int n = 0; n < TILE_N; ++n) { + v[n] = bytes_from_nibbles_64(B[n * KB].qs + k * 32, B[n * KB].qh, /* group */2 * k); + } + + transpose_16x16_32bit(v); + + // 1. pack lower 4bits with 2 groups + for (int n = 0; n < TILE_N; n += 2) { + // get lower 4 bits + const __m512i r0 = _mm512_and_si512(v[n], lowMask); + const __m512i r1 = _mm512_and_si512(v[n + 1], lowMask); + _mm512_storeu_si512((__m512i *)pb, packNibbles(r0, r1)); pb += 64; + } + + // 2. pack higher 1bit with 2 groups + const __m512i hmask = _mm512_set1_epi8(0x10); + for (int g = 0; g < 2; ++g) { + __m512i hbits = _mm512_setzero_si512(); + hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 0], hmask), 4)); + hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 1], hmask), 3)); + hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 2], hmask), 2)); + hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 3], hmask), 1)); + hbits = _mm512_add_epi8(hbits, _mm512_and_si512(v[g * 8 + 4], hmask) ); + hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 8 + 5], hmask), 1)); + hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 8 + 6], hmask), 2)); + hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 8 + 7], hmask), 3)); + _mm512_storeu_si512((__m512i *)ph, hbits); ph += 64; + } + } +} + +template <> +inline void pack_qs(void * RESTRICT packed_B, const block_q6_K * RESTRICT B, int KB) { + __m512i v[32]; + const __m512i lowMask = _mm512_set1_epi8(0xF); + // QK_K 256 with 8 groups, handle 4 groups at a time + char * pb = (char *)packed_B; + char * ph = (char *)packed_B + (QK_K / 2) * TILE_N; + for (int k = 0; k < QK_K / 128; ++k) { + for (int n = 0; n < TILE_N; ++n) { + bytes_from_nibbles_128(v[n], v[n + 16], B[n * KB].ql + k * 64, B[n * KB].qh + k * 32); + } + + // top half: group 0,1 or 4,5; bottom half: group 2,3 or 6,7 + transpose_16x16_32bit(v); + transpose_16x16_32bit(v + 16); + + // 1. pack lower 4bits with 4 groups + for (int n = 0; n < 32; n += 2) { + const __m512i r0 = _mm512_and_si512(v[n], lowMask); + const __m512i r1 = _mm512_and_si512(v[n + 1], lowMask); + _mm512_storeu_si512((__m512i *)pb, packNibbles(r0, r1)); pb += 64; + } + + // 2. pack higher 2bit with 4 groups + const __m512i hmask = _mm512_set1_epi8(0x30); + for (int g = 0; g < 8; ++g) { + __m512i hbits = _mm512_setzero_si512(); + hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 4 + 0], hmask), 4)); + hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 4 + 1], hmask), 2)); + hbits = _mm512_add_epi8(hbits, _mm512_and_si512(v[g * 4 + 2], hmask) ); + hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 4 + 3], hmask), 2)); + _mm512_storeu_si512((__m512i *)ph, hbits); ph += 64; + } + } +} + +template <> +inline void pack_qs(void * RESTRICT packed_B, const block_iq4_xs * RESTRICT B, int KB) { + __m512i v[16]; + char * pb = (char *)packed_B; + for (int k = 0; k < QK_K / 64; ++k) { + for (int n = 0; n < TILE_N; ++n) { + __m256i r0 = bytes_from_nibbles_32(B[n * KB].qs + k * 32 + 0); + __m256i r1 = bytes_from_nibbles_32(B[n * KB].qs + k * 32 + 16); + v[n] = _mm512_inserti32x8(_mm512_castsi256_si512(r0), r1, 1); + } + + transpose_16x16_32bit(v); + + // pack again with 128 to fully utilize vector length + for (int n = 0; n < TILE_N; n += 2) { + _mm512_storeu_si512((__m512i *)pb, packNibbles(v[n], v[n + 1])); + pb += 64; + } + } +} + +// pack B to vnni formats in 4bits or 8 bits +void pack_B(void * RESTRICT packed_B, const block_q4_0 * RESTRICT B, int KB) { + pack_qs(packed_B, B, KB); + ggml_half * d0 = reinterpret_cast((char *)packed_B + TILE_N * TILE_K / 2); + for (int n = 0; n < TILE_N; ++n) { + d0[n] = B[n * KB].d; + } +} + +void pack_B(void * RESTRICT packed_B, const block_q4_1 * RESTRICT B, int KB) { + pack_qs(packed_B, B, KB); + ggml_half * d0 = reinterpret_cast((char *)packed_B + TILE_N * TILE_K / 2); + ggml_half * m0 = d0 + TILE_N; + for (int n = 0; n < TILE_N; ++n) { + d0[n] = B[n * KB].d; + m0[n] = B[n * KB].m; + } +} + +inline void s8s8_compensation(void * RESTRICT packed_B) { + // packed_B layout: + // quants {TILE_N, TILEK} int8_t + // d0 {TILE_N} ggml_half + // comp {TILE_N} int32_t + const int offset = TILE_N * TILE_K + TILE_N * sizeof(ggml_half); + __m512i vcomp = _mm512_setzero_si512(); + const __m512i off = _mm512_set1_epi8(static_cast(0x80)); + for (int k = 0; k < 8; ++k) { + __m512i vb = _mm512_loadu_si512((const __m512i *)((const char *)packed_B + k * 64)); + vcomp = _mm512_dpbusd_epi32(vcomp, off, vb); + } + _mm512_storeu_si512((__m512i *)((char *)(packed_B) + offset), vcomp); +} + +void pack_B(void * RESTRICT packed_B, const block_q8_0 * RESTRICT B, int KB) { + pack_qs(packed_B, B, KB); + ggml_half * d0 = reinterpret_cast((char *)packed_B + TILE_N * TILE_K); + for (int n = 0; n < TILE_N; ++n) { + d0[n] = B[n * KB].d; + } + s8s8_compensation(packed_B); +} + +// convert 8 * {min, scale} from int6 to int8 +inline void unpack_mins_and_scales(const uint8_t * scales, uint32_t * utmp) { + const uint32_t kmask1 = 0x3f3f3f3f; + const uint32_t kmask2 = 0x0f0f0f0f; + const uint32_t kmask3 = 0x03030303; + + memcpy(utmp, scales, 12); + utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); + const uint32_t uaux = utmp[1] & kmask1; + utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); + utmp[2] = uaux; + utmp[0] &= kmask1; +} + +// packed_B layout: +// quants {8, TILE_N, 16} uint8 +// scales {8, TILE_N} uint8 +// mins {8, TILE_N} uint8 +// d {TILE_N} ggml_half +// dmin {TILE_N} ggml_half +void pack_B(void * RESTRICT packed_B, const block_q4_K * RESTRICT B, int KB) { + pack_qs(packed_B, B, KB); + + uint8_t * scales = reinterpret_cast((char *)packed_B + (QK_K / 2) * TILE_N); + uint8_t * mins = scales + 8 * TILE_N; + ggml_half * d = reinterpret_cast(mins + 8 * TILE_N); + ggml_half * dmin = d + TILE_N; + + union { + uint32_t u32[4]; + uint8_t u8[16]; + } s; + + for (int n = 0; n < TILE_N; ++n) { + unpack_mins_and_scales(B[n * KB].scales, s.u32); + for (int k = 0; k < 8; ++k) { + scales[k * TILE_N + n] = s.u8[k]; + mins[(k >> 1) * TILE_N * 2 + n * 2 + (k & 0x1)] = s.u8[k + 8]; + } + d[n] = B[n * KB].d; + dmin[n] = B[n * KB].dmin; + } +} + +// packed_B layout: +// quants {8, TILE_N, 16} uint8 +// qh {8, TILE_N, 4} uint8 +// scales {8, TILE_N} uint8 +// mins {8, TILE_N} uint8 +// d {TILE_N} ggml_half +// dmin {TILE_N} ggml_half +void pack_B(void * RESTRICT packed_B, const block_q5_K * RESTRICT B, int KB) { + pack_qs(packed_B, B, KB); + + uint8_t * scales = reinterpret_cast((char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N); + uint8_t * mins = scales + 8 * TILE_N; + ggml_half * d = reinterpret_cast(mins + 8 * TILE_N); + ggml_half * dmin = d + TILE_N; + + union { + uint32_t u32[4]; + uint8_t u8[16]; + } s; + + for (int n = 0; n < TILE_N; ++n) { + unpack_mins_and_scales(B[n * KB].scales, s.u32); + for (int k = 0; k < 8; ++k) { + scales[k * TILE_N + n] = s.u8[k]; + mins[(k >> 1) * TILE_N * 2 + n * 2 + (k & 0x1)] = s.u8[k + 8]; + } + d[n] = B[n * KB].d; + dmin[n] = B[n * KB].dmin; + } +} + +// packed_B layout: +// quants {16, TILE_N, 8} uint8 +// qh {16, TILE_N, 4} uint8 +// scales {16, TILE_N} uint8 +// d {TILE_N} ggml_half +void pack_B(void * RESTRICT packed_B, const block_q6_K * RESTRICT B, int KB) { + pack_qs(packed_B, B, KB); + + uint8_t * scales = reinterpret_cast((char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N); + ggml_half * d = reinterpret_cast(scales + 16 * TILE_N); + for (int n = 0; n < TILE_N; ++n) { + const int8_t * ps = B[n * KB].scales; + for (int k = 0; k < 16; ++k) { + scales[k * TILE_N + n] = ps[k]; + } + d[n] = B[n * KB].d; + } +} + +// packed_B layout: +// quants {8, TILE_N, 16} uint8 +// scales {8, TILE_N} int8 +// d {TILE_N} ggml_half +void pack_B(void * RESTRICT packed_B, const block_iq4_xs * RESTRICT B, int KB) { + pack_qs(packed_B, B, KB); + + int8_t * scales = reinterpret_cast((char *)packed_B + (QK_K / 2) * TILE_N); + ggml_half * d = reinterpret_cast(scales + 8 * TILE_N); + + // pack the scales + for (int n = 0; n < TILE_N; ++n) { + uint16_t sh = B[n * KB].scales_h; + for (int k = 0; k < 8; k += 2) { + const int16_t ls1 = ((B[n * KB].scales_l[k / 2] & 0xf) | ((sh << 4) & 0x30)) - 32; + const int16_t ls2 = ((B[n * KB].scales_l[k / 2] >> 4) | ((sh << 2) & 0x30)) - 32; + scales[(k + 0) * TILE_N + n] = ls1; + scales[(k + 1) * TILE_N + n] = ls2; + sh >>= 4; + } + d[n] = B[n * KB].d; + } +} + +template> +void unpack_B(packed_B_t * RESTRICT tile, const void * RESTRICT packed_B) { + GGML_UNUSED(tile); + GGML_UNUSED(packed_B); +}; + +template <> +void unpack_B(int8_t * RESTRICT tile, const void * RESTRICT packed_B) { + const __m512i off = _mm512_set1_epi8(8); + const __m512i lowMask = _mm512_set1_epi8(0xF); + for (int n = 0; n < 8; n += 2) { + __m512i bytes = _mm512_loadu_si512((const __m512i *)((const char *)packed_B + n * 32)); + const __m512i r0 = _mm512_sub_epi8(_mm512_and_si512(bytes, lowMask), off); + const __m512i r1 = _mm512_sub_epi8(_mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask), off); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1); + } +} + +template <> +void unpack_B(uint8_t * RESTRICT tile, const void * RESTRICT packed_B) { + const __m512i lowMask = _mm512_set1_epi8(0xF); + for (int n = 0; n < 8; n += 2) { + __m512i bytes = _mm512_loadu_si512((const __m512i *)((const char *)packed_B + n * 32)); + const __m512i r0 = _mm512_and_si512(bytes, lowMask); + const __m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1); + } +} + +// packed_B_t for QKK is int8_t +template +void unpack_B(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) { + const int packed_B_group_size = QK_K / 2 * TILE_N / 8; + const char * packed_B_group = (const char *)packed_B + k * packed_B_group_size; + const __m512i lowMask = _mm512_set1_epi8(0xF); + for (int n = 0; n < 8; n += 2) { + __m512i bytes = _mm512_loadu_si512(packed_B_group + n * 32); + const __m512i r0 = _mm512_and_si512(bytes, lowMask); + const __m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1); + } +} + +template <> +void unpack_B(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) { + // lower 4bits, stride 256 bytes + const int packed_l4_group_size = QK_K / 2 * TILE_N / 8; + const char * pb = (const char *)packed_B + k * packed_l4_group_size; + + // higher 1bit, stride 64 bytes + const int packed_h1_group_size = QK_K / 8 * TILE_N / 8; + const char * ph = (const char *)packed_B + (QK_K / 2) * TILE_N + k * packed_h1_group_size; + const __m512i hbits = _mm512_loadu_si512(ph); + + const __m512i lowMask = _mm512_set1_epi8(0xF); + __m512i hmask0 = _mm512_set1_epi8(0x1); + __m512i hmask1 = _mm512_set1_epi8(0x2); + + for (int n = 0; n < 8; n += 2) { + __m512i bytes = _mm512_loadu_si512(pb + n * 32); + __m512i r0 = _mm512_and_si512(bytes, lowMask); + __m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + __m512i h0 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask0), n), 4); + __m512i h1 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask1), n + 1), 4); + + hmask0 = _mm512_slli_epi16(hmask0, 2); + hmask1 = _mm512_slli_epi16(hmask1, 2); + r0 = _mm512_add_epi8(r0, h0); + r1 = _mm512_add_epi8(r1, h1); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1); + } +} + +template <> +void unpack_B(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) { + // lower 4bits, stride 128 bytes + const int packed_l4_group_size = QK_K / 2 * TILE_N / 16; + const char * pb = (const char *)packed_B + k * packed_l4_group_size; + + // higher 2bits, stride 64 bytes + const int packed_h2_group_size = QK_K / 4 * TILE_N / 16; + const char * ph = (const char *)packed_B + (QK_K / 2) * TILE_N + k * packed_h2_group_size; + const __m512i hbits = _mm512_loadu_si512(ph); + + const __m512i off = _mm512_set1_epi8(32); + const __m512i lowMask = _mm512_set1_epi8(0xF); + __m512i hmask0 = _mm512_set1_epi8(0x3); // 0011 + __m512i hmask1 = _mm512_set1_epi8(0xC); // 1100 + + // notes: skip zero padding from row4 to row7 as we have done so in `unpack_A` + __m512i bytes = _mm512_loadu_si512(pb); + __m512i r0 = _mm512_and_si512(bytes, lowMask); + __m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + __m512i h0 = _mm512_slli_epi16(_mm512_and_si512(hbits, hmask0), 4); + __m512i h1 = _mm512_slli_epi16(_mm512_and_si512(hbits, hmask1), 2); + _mm512_storeu_si512((__m512i *)(tile + 0), _mm512_sub_epi8(_mm512_add_epi8(r0, h0), off)); + _mm512_storeu_si512((__m512i *)(tile + 64), _mm512_sub_epi8(_mm512_add_epi8(r1, h1), off)); + + hmask0 = _mm512_slli_epi16(hmask0, 4); + hmask1 = _mm512_slli_epi16(hmask1, 4); + + bytes = _mm512_loadu_si512(pb + 64); + r0 = _mm512_and_si512(bytes, lowMask); + r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + h0 = _mm512_and_si512(hbits, hmask0); + h1 = _mm512_srli_epi16(_mm512_and_si512(hbits, hmask1), 2); + _mm512_storeu_si512((__m512i *)(tile + 128), _mm512_sub_epi8(_mm512_add_epi8(r0, h0), off)); + _mm512_storeu_si512((__m512i *)(tile + 192), _mm512_sub_epi8(_mm512_add_epi8(r1, h1), off)); +} + +template <> +void unpack_B(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) { + static const __m512i values128 = _mm512_set_epi8( + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127, + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127, + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127, + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127 + ); + + const int packed_B_group_size = QK_K / 2 * TILE_N / 8; + const char * pb = (const char *)packed_B + k * packed_B_group_size; + const __m512i lowMask = _mm512_set1_epi8(0xF); + + for (int n = 0; n < 8; n += 2) { + __m512i bytes = _mm512_loadu_si512(pb + n * 32); + const __m512i r0 = _mm512_shuffle_epi8(values128, _mm512_and_si512(bytes, lowMask)); + const __m512i r1 = _mm512_shuffle_epi8(values128, _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask)); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1); + } +} + +template +struct acc_C {}; + +template +struct acc_C { + static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_0 * A, int lda, const void * packed_B, int nr) { + const int offset = TILE_N * TILE_K / 2; + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset))); + + for (int m = 0; m < nr; ++m) { + const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d)); + const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); + + __m512 vsum; + if (is_acc) { + vsum = _mm512_loadu_ps(C + m * ldc); + } else { + vsum = _mm512_set1_ps(0.f); + } + vsum = _mm512_fmadd_ps(vtile, _mm512_mul_ps(vd0, vd1), vsum); + _mm512_storeu_ps(C + m * ldc, vsum); + } + } +}; + +template +struct acc_C { + static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_1 * A, int lda, const void * packed_B, int nr) { + const int offset = TILE_N * TILE_K / 2; + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset))); + const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset + TILE_N * sizeof(ggml_half)))); + + for (int m = 0; m < nr; ++m) { + const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d)); + const __m512 vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].s)); + const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); + + __m512 vsum; + if (is_acc) { + vsum = _mm512_loadu_ps(C + m * ldc); + } else { + vsum = _mm512_set1_ps(0.f); + } + vsum = _mm512_fmadd_ps(vtile, _mm512_mul_ps(vd0, vd1), vsum); + vsum = _mm512_fmadd_ps(vm0, vs1, vsum); + _mm512_storeu_ps(C + m * ldc, vsum); + } + } +}; + +template +struct acc_C { + static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_0 * A, int lda, const void * packed_B, int nr) { + const int offset = TILE_N * TILE_K; + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset))); + + for (int m = 0; m < nr; ++m) { + const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d)); + const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); + + __m512 vsum; + if (is_acc) { + vsum = _mm512_loadu_ps(C + m * ldc); + } else { + vsum = _mm512_set1_ps(0.f); + } + vsum = _mm512_fmadd_ps(vtile, _mm512_mul_ps(vd0, vd1), vsum); + _mm512_storeu_ps(C + m * ldc, vsum); + } + } +}; + +template +struct acc_C { + static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) { + const uint8_t * scales = reinterpret_cast((const char *)packed_B + (QK_K / 2) * TILE_N); + const uint8_t * mins = scales + 8 * TILE_N; + const ggml_half * d0 = reinterpret_cast(mins + 8 * TILE_N); + const ggml_half * dmin = d0 + TILE_N; + + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0)); + const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)dmin)); + + for (int m = 0; m < nr; ++m) { + const float d1 = A[m * lda].d; + const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0); + const __m512 vdm = _mm512_mul_ps(_mm512_set1_ps(-d1), vdmin); + const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); + + __m512 vsum; + if (is_acc) { + vsum = _mm512_loadu_ps(C + m * ldc); + } else { + vsum = _mm512_set1_ps(0.f); + } + + const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[m * lda].bsums); + const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1)); + + __m512i acc_m = _mm512_setzero_si512(); + for (int k = 0; k < 4; ++k) { + __m512i vmask = _mm512_set1_epi32(k); + __m512i va = _mm512_permutexvar_epi32(vmask, _mm512_castsi128_si512(q8s)); + __m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(mins + k * 32))); + acc_m = _mm512_dpwssds_epi32(acc_m, va, vb); + } + + vsum = _mm512_fmadd_ps(vtile, vd, vsum); + vsum = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc_m), vdm, vsum); + _mm512_storeu_ps(C + m * ldc, vsum); + } + } +}; + +template +struct acc_C { + static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) { + const uint8_t * scales = reinterpret_cast((const char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N); + const uint8_t * mins = scales + 8 * TILE_N; + const ggml_half * d0 = reinterpret_cast(mins + 8 * TILE_N); + const ggml_half * dmin = d0 + TILE_N; + + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0)); + const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)dmin)); + + for (int m = 0; m < nr; ++m) { + const float d1 = A[m * lda].d; + const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0); + const __m512 vdm = _mm512_mul_ps(_mm512_set1_ps(-d1), vdmin); + const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); + + __m512 vsum; + if (is_acc) { + vsum = _mm512_loadu_ps(C + m * ldc); + } else { + vsum = _mm512_set1_ps(0.f); + } + + const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[m * lda].bsums); + const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1)); + + __m512i acc_m = _mm512_setzero_si512(); + for (int k = 0; k < 4; ++k) { + __m512i vmask = _mm512_set1_epi32(k); + __m512i va = _mm512_permutexvar_epi32(vmask, _mm512_castsi128_si512(q8s)); + __m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(mins + k * 32))); + acc_m = _mm512_dpwssds_epi32(acc_m, va, vb); + } + + vsum = _mm512_fmadd_ps(vtile, vd, vsum); + vsum = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc_m), vdm, vsum); + _mm512_storeu_ps(C + m * ldc, vsum); + } + } +}; + +template +struct acc_C { + static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) { + const uint8_t * scales = reinterpret_cast((const char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N); + const ggml_half * d0 = reinterpret_cast(scales + 16 * TILE_N); + + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0)); + + for (int m = 0; m < nr; ++m) { + const float d1 = A[m * lda].d; + const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0); + const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); + + __m512 vsum; + if (is_acc) { + vsum = _mm512_loadu_ps(C + m * ldc); + } else { + vsum = _mm512_set1_ps(0.f); + } + + vsum = _mm512_fmadd_ps(vtile, vd, vsum); + _mm512_storeu_ps(C + m * ldc, vsum); + } + } +}; + +template +struct acc_C { + static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) { + const int8_t * scales = reinterpret_cast((const char *)packed_B + (QK_K / 2) * TILE_N); + const ggml_half * d0 = reinterpret_cast(scales + 8 * TILE_N); + + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0)); + + for (int m = 0; m < nr; ++m) { + const float d1 = A[m * lda].d; + const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0); + const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); + + __m512 vsum; + if (is_acc) { + vsum = _mm512_loadu_ps(C + m * ldc); + } else { + vsum = _mm512_set1_ps(0.f); + } + + vsum = _mm512_fmadd_ps(vtile, vd, vsum); + _mm512_storeu_ps(C + m * ldc, vsum); + } + } +}; + +template constexpr int get_quants_size(); +template <> constexpr int get_quants_size() { return (QK_K / 2) * TILE_N; } +template <> constexpr int get_quants_size() { return (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N; } +template <> constexpr int get_quants_size() { return (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N; } +template <> constexpr int get_quants_size() { return (QK_K / 2) * TILE_N; } + +// used for QKK format +template ::value, int>::type = 0> +inline void scale_C(const int32_t * RESTRICT tile, int32_t * RESTRICT sumi, const void * packed_B, int k, int nr) { + const uint8_t * scales = reinterpret_cast((const char *)packed_B + get_quants_size()); + const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(scales + k * TILE_N))); + + for (int m = 0; m < nr; ++m) { + __m512i vsumi; + if (is_acc) { + vsumi = _mm512_loadu_si512(sumi + m * TILE_N); + } else { + vsumi = _mm512_setzero_si512(); + } + __m512i vtile = _mm512_loadu_si512(tile + m * TILE_N); + vsumi = _mm512_add_epi32(vsumi, _mm512_mullo_epi32(vtile, vscale)); + _mm512_storeu_si512((__m512i *)(sumi + m * TILE_N), vsumi); + } +} + +template +struct tinygemm_kernel_avx { + static void apply(int K, const TA * RESTRICT A, const TB * RESTRICT B, TC * RESTRICT C, int ldc) { + GGML_UNUSED(K); + GGML_UNUSED(A); + GGML_UNUSED(B); + GGML_UNUSED(C); + GGML_UNUSED(ldc); + } +}; + +template +struct tinygemm_kernel_avx { + static void apply(int K, const float * RESTRICT A, const ggml_fp16_t * RESTRICT B, float * RESTRICT C, int ldc) { + constexpr int ROWS = BLOCK_M; + constexpr int COLS = BLOCK_N; + assert(BLOCK_K == 16); + + __m512 va; + __m512 vb[COLS]; + __m512 vc[ROWS * COLS]; + + auto loadc = [&](int idx) { + vc[idx] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + auto compute = [&](int idx, int k) { + // TODO: use `constexpr` here to get rid of interger div + // when upgraded to C++17 + const int row = idx / COLS; + const int col = idx % COLS; + + if (col == 0) { + va = _mm512_loadu_ps(A + row * K + k); + } + if (row == 0) { + vb[col] = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(B + col * K + k))); + } + vc[idx] = _mm512_fmadd_ps(va, vb[col], vc[idx]); + }; + + for (int k = 0; k < K; k += 16) { + Unroll{}(compute, k); + } + + auto storec = [&](int idx) { + const int row = idx / COLS; + const int col = idx % COLS; + C[row * ldc + col] = _mm512_reduce_add_ps(vc[idx]); + }; + Unroll{}(storec); + } +}; + +#define LAUNCH_TINYGEMM_KERNEL_AVX(MB_SIZE, NB_SIZE) \ + tinygemm_kernel_avx::apply( \ + K, (const float *)src1->data + mb_start * K, \ + (const type *)src0->data + nb_start * K, \ + (float *)dst->data + mb_start * ldc + nb_start, ldc); + + +// re-organize in the format {NB, KB, TILE_SIZE}: +#define PACKED_INDEX(n, k, KB, tile_size) (n * KB + k) * tile_size + +template +void convert_B_packed_format(void * RESTRICT packed_B, const TB * RESTRICT B, int N, int K, int n_threads) { + const int NB = N / TILE_N; + const int KB = K / BLOCK_K; + const int TILE_SIZE = get_tile_size(); + + // parallel on NB should be enough + parallel_for(n_threads, NB, [&](int begin, int end) { + for (int n = begin; n < end; ++n) { + for (int k = 0; k < KB; ++k) { + int n0 = n * TILE_N; + pack_B((char *)packed_B + PACKED_INDEX(n, k, KB, TILE_SIZE), &B[n0 * KB + k], KB); + } + } + }); +} + +template +struct tinygemm_kernel_vnni {}; + +template +struct tinygemm_kernel_vnni { + static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + + constexpr int COLS = BLOCK_N / 16; + const int TILE_SIZE = TILE_N * sizeof(block_q4_0); + + const block_q8_0 * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + __m512i va[8]; + __m512 vc[COLS]; + __m512 vd1; + + // sum of offsets, shared across COLS + // + // avx512-vnni does not have `_mm512_dpbssd_epi32`, + // need to transfrom ss to us: + // a * (b - 8) is equavilent to b * a - 8 * a + // s u u u s u s + // + __m512i vcomp; + + const __m512i off = _mm512_set1_epi8(8); + const __m512i lowMask = _mm512_set1_epi8(0xF); + + auto loadc = [&](int col) { + vc[col] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + auto compute = [&](int col, int i) { + // load a and compute compensation + if (col == 0) { + const int32_t * a_ptr = reinterpret_cast(A[0 * KB + i].qs); + vcomp = _mm512_setzero_si512(); + for (int k = 0; k < 8; ++k) { + va[k] = _mm512_set1_epi32(a_ptr[k]); + vcomp = _mm512_dpbusd_epi32(vcomp, off, va[k]); + } + vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d)); + } + + // load b + __m512i vsum = _mm512_setzero_si512(); + const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE); + for (int k = 0; k < 8; k += 2) { + __m512i bytes = _mm512_loadu_si512((const __m512i *)(b_ptr + k * 32)); + __m512i vb0 = _mm512_and_si512(bytes, lowMask); + vsum = _mm512_dpbusd_epi32(vsum, vb0, va[k + 0]); + __m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + vsum = _mm512_dpbusd_epi32(vsum, vb1, va[k + 1]); + } + const int offset = TILE_N * TILE_K / 2; + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset))); + vsum = _mm512_sub_epi32(vsum, vcomp); + + vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(vsum), _mm512_mul_ps(vd0, vd1), vc[col]); + }; + + for (int i = 0; i < KB; ++i) { + Unroll{}(compute, i); + } + + //store to C + auto storec = [&](int col) { + _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]); + }; + Unroll{}(storec); + } +}; + +template +struct tinygemm_kernel_vnni { + static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + + constexpr int COLS = BLOCK_N / 16; + const int TILE_SIZE = TILE_N * sizeof(block_q4_1); + + const block_q8_1 * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + __m512i va[8]; + __m512i vb[8]; + __m512 vc[COLS]; + __m512 vd1, vs1; + + const __m512i lowMask = _mm512_set1_epi8(0xF); + + auto loadc = [&](int col) { + vc[col] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + auto compute = [&](int col, int i) { + // load a + if (col == 0) { + const int32_t * a_ptr = reinterpret_cast(A[0 * KB + i].qs); + for (int k = 0; k < 8; ++k) { + va[k] = _mm512_set1_epi32(a_ptr[k]); + } + vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d)); + vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].s)); + } + + // load b + const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE); + for (int k = 0; k < 8; k += 2) { + __m512i bytes = _mm512_loadu_si512((const __m512i *)(b_ptr + k * 32)); + vb[k + 0] = _mm512_and_si512(bytes, lowMask); + vb[k + 1] = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + } + const int offset = TILE_N * TILE_K / 2; + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset))); + const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset + TILE_N * sizeof(ggml_half)))); + + __m512i vsum = _mm512_setzero_si512(); + for (int k = 0; k < 8; ++k) { + vsum = _mm512_dpbusd_epi32(vsum, vb[k], va[k]); + } + + vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(vsum), _mm512_mul_ps(vd0, vd1), vc[col]); + vc[col] = _mm512_fmadd_ps(vm0, vs1, vc[col]); + }; + + for (int i = 0; i < KB; ++i) { + Unroll{}(compute, i); + } + + //store to C + auto storec = [&](int col) { + _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]); + }; + Unroll{}(storec); + } +}; + +template +struct tinygemm_kernel_vnni { + static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + + constexpr int COLS = BLOCK_N / 16; + const int TILE_SIZE = TILE_N * sizeof(block_q8_0) + TILE_N * sizeof(int32_t); + + const block_q8_0 * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + __m512i va[8]; + __m512i vb[8]; + __m512 vc[COLS]; + __m512 vd1; + + // Notes: s8s8 igemm compensation in avx512-vnni + // change s8s8 to u8s8 with compensate + // a * b = (a + 128) * b - 128 * b + // s s u s u s + // + // (128 * b is pre-computed when packing B to vnni formats) + // + const __m512i off = _mm512_set1_epi8(static_cast(0x80)); + + auto loadc = [&](int col) { + vc[col] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + auto compute = [&](int col, int i) { + // load a and add offset 128 + if (col == 0) { + const int32_t * a_ptr = reinterpret_cast(A[0 * KB + i].qs); + for (int k = 0; k < 8; ++k) { + va[k] = _mm512_set1_epi32(a_ptr[k]); + va[k] = _mm512_add_epi8(va[k], off); + } + vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d)); + } + + // load b + const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE); + for (int k = 0; k < 8; ++k) { + vb[k] = _mm512_loadu_si512((const __m512i *)(b_ptr + k * 64)); + } + const int offset = TILE_N * TILE_K; + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset))); + const int offset2 = TILE_N * TILE_K + TILE_N * sizeof(ggml_half); + const __m512i vcomp = _mm512_loadu_si512((const __m512i *)(b_ptr + offset2)); + + __m512i vsum = _mm512_setzero_si512(); + for (int k = 0; k < 8; ++k) { + vsum = _mm512_dpbusd_epi32(vsum, va[k], vb[k]); + } + vsum = _mm512_sub_epi32(vsum, vcomp); + + vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(vsum), _mm512_mul_ps(vd0, vd1), vc[col]); + }; + + for (int i = 0; i < KB; ++i) { + Unroll{}(compute, i); + } + + //store to C + auto storec = [&](int col) { + _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]); + }; + Unroll{}(storec); + } +}; + +template +struct tinygemm_kernel_vnni { + static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + + constexpr int COLS = BLOCK_N / 16; + const int TILE_SIZE = TILE_N * sizeof(block_q4_K) + TILE_N * 4; + + const block_q8_K * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + // a.qs: 8 groups, 32 bytes each group (m256i) + __m512i va[8]; + // a.bsum: 8 groups, 2 bytes each group (m128i) + __m512i va_bsum; + __m512 vc[COLS]; + __m512 vd1; + + // packed_B: + const int offset_scales = (QK_K / 2) * TILE_N; + const int offset_mins = (QK_K / 2) * TILE_N + 8 * TILE_N; + const int offset_d0 = (QK_K / 2) * TILE_N + 16 * TILE_N; + const int offset_dmin = (QK_K / 2) * TILE_N + 16 * TILE_N + TILE_N * sizeof(ggml_half); + + const __m512i lowMask = _mm512_set1_epi8(0xF); + + auto loadc = [&](int col) { + vc[col] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + // Notes: vnni formats in QK_K + // a) quants vnni format + // int8 {k/4, n, 4}, viewed as 2d {k/4, 4n}, k = 32 + // from {16, 32} to {8, 64} + // + // b) min vnni format + // int16 {k/2, n, 2}, viewed as 2d {k/2, 2n}, k = 8 + // from {16, 8} to {4, 32} + // + auto compute = [&](int col, int i) { + // load a + if (col == 0) { + for (int k_group = 0; k_group < QK_K / 32; ++k_group) { + va[k_group] = _mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)(A[0 * KB + i].qs + k_group * 32))); + } + const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums); + const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1)); + va_bsum = _mm512_castsi128_si512(q8s); + vd1 = _mm512_set1_ps(A[0 * KB + i].d); + } + + // step 1: accumultate the quants + __m512i acc = _mm512_setzero_si512(); + const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE); + const char * b_qs = b_ptr; + for (int k_group = 0; k_group < QK_K / 32; ++k_group) { + __m512i vsum = _mm512_setzero_si512(); + for (int k = 0; k < 8; k += 2) { + __m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 0), va[k_group]); + __m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 1), va[k_group]); + + __m512i bytes = _mm512_loadu_si512((const __m512i *)b_qs); + __m512i vb0 = _mm512_and_si512(bytes, lowMask); + vsum = _mm512_dpbusd_epi32(vsum, vb0, va0); + __m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + vsum = _mm512_dpbusd_epi32(vsum, vb1, va1); + + b_qs += 64; + } + // vacc += scale * (q8 @ q4) + const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N))); + acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale)); + } + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0))); + vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]); + + // step 2: accumulate the mins + __m512i acc_m = _mm512_setzero_si512(); + for (int k = 0; k < 4; ++k) { + __m512i vmask = _mm512_set1_epi32(k); + __m512i va = _mm512_permutexvar_epi32(vmask, va_bsum); + __m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_mins + k * 32))); + acc_m = _mm512_dpwssds_epi32(acc_m, va, vb); + } + const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_dmin))); + vc[col] = _mm512_fnmadd_ps(_mm512_cvtepi32_ps(acc_m), _mm512_mul_ps(vdmin, vd1), vc[col]); + }; + + for (int i = 0; i < KB; ++i) { + Unroll{}(compute, i); + } + + //store to C + auto storec = [&](int col) { + _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]); + }; + Unroll{}(storec); + } +}; + +template +struct tinygemm_kernel_vnni { + static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + + constexpr int COLS = BLOCK_N / 16; + const int TILE_SIZE = TILE_N * sizeof(block_q5_K) + TILE_N * 4; + + const block_q8_K * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + // a.qs: 8 groups, 32 bytes each group (m256i) + __m512i va[8]; + // a.bsum: 8 groups, 2 bytes each group (m128i) + __m512i va_bsum; + __m512 vc[COLS]; + __m512 vd1; + + // packed_B: + const int offset_qh = (QK_K / 2) * TILE_N; + const int offset_scales = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N; + const int offset_mins = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N + 8 * TILE_N; + const int offset_d0 = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N + 16 * TILE_N; + const int offset_dmin = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N + 16 * TILE_N + TILE_N * sizeof(ggml_half); + + const __m512i lowMask = _mm512_set1_epi8(0xF); + + auto loadc = [&](int col) { + vc[col] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + // Q5_K and Q4_K shares the same vnni formats, refer to notes above. + auto compute = [&](int col, int i) { + // load a + if (col == 0) { + for (int k_group = 0; k_group < QK_K / 32; ++k_group) { + va[k_group] = _mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)(A[0 * KB + i].qs + k_group * 32))); + } + const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums); + const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1)); + va_bsum = _mm512_castsi128_si512(q8s); + vd1 = _mm512_set1_ps(A[0 * KB + i].d); + } + + // step 1: accumultate the quants + __m512i acc = _mm512_setzero_si512(); + const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE); + const char * b_qs = b_ptr; + const char * b_qh = b_ptr + offset_qh; + for (int k_group = 0; k_group < QK_K / 32; ++k_group) { + __m512i vsum = _mm512_setzero_si512(); + __m512i hmask0 = _mm512_set1_epi8(0x1); + __m512i hmask1 = _mm512_set1_epi8(0x2); + __m512i hbits = _mm512_loadu_si512((const __m512i *)(b_qh + k_group * 64)); + for (int k = 0; k < 8; k += 2) { + __m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 0), va[k_group]); + __m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 1), va[k_group]); + + __m512i bytes = _mm512_loadu_si512((const __m512i *)b_qs); + __m512i vb0 = _mm512_and_si512(bytes, lowMask); + __m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + + __m512i vh0 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask0), k), 4); + __m512i vh1 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask1), k + 1), 4); + + hmask0 = _mm512_slli_epi16(hmask0, 2); + hmask1 = _mm512_slli_epi16(hmask1, 2); + vb0 = _mm512_add_epi8(vb0, vh0); + vb1 = _mm512_add_epi8(vb1, vh1); + + vsum = _mm512_dpbusd_epi32(vsum, vb0, va0); + vsum = _mm512_dpbusd_epi32(vsum, vb1, va1); + + b_qs += 64; + } + // vacc += scale * (q8 @ q5) + const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N))); + acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale)); + } + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0))); + vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]); + + // step 2: accumulate the mins + __m512i acc_m = _mm512_setzero_si512(); + for (int k = 0; k < 4; ++k) { + __m512i vmask = _mm512_set1_epi32(k); + __m512i va = _mm512_permutexvar_epi32(vmask, va_bsum); + __m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_mins + k * 32))); + acc_m = _mm512_dpwssds_epi32(acc_m, va, vb); + } + const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_dmin))); + vc[col] = _mm512_fnmadd_ps(_mm512_cvtepi32_ps(acc_m), _mm512_mul_ps(vdmin, vd1), vc[col]); + }; + + for (int i = 0; i < KB; ++i) { + Unroll{}(compute, i); + } + + //store to C + auto storec = [&](int col) { + _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]); + }; + Unroll{}(storec); + } +}; + +template +struct tinygemm_kernel_vnni { + static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + + constexpr int COLS = BLOCK_N / 16; + const int TILE_SIZE = TILE_N * sizeof(block_q6_K); + + const block_q8_K * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + // load the 256 bytes from A to 4 avx512 vectors + __m512i va[4]; + __m512 vc[COLS]; + __m512 vd1; + + // packed_B: + const int offset_qh = (QK_K / 2) * TILE_N; + const int offset_scales = (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N; + const int offset_d0 = (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N + 16 * TILE_N; + + // compensation + __m512i vcomp; + + const __m512i m32s = _mm512_set1_epi32(32); + const __m512i lowMask = _mm512_set1_epi8(0xF); + + auto loadc = [&](int col) { + vc[col] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + auto compute = [&](int col, int i) { + if (col == 0) { + // load a + va[0] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 0)); + va[1] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 64)); + va[2] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 128)); + va[3] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 192)); + + const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums); + vcomp = _mm512_mullo_epi32(_mm512_cvtepi16_epi32(q8sums), m32s); + vd1 = _mm512_set1_ps(A[0 * KB + i].d); + } + + // accmulate the quants + __m512i acc = _mm512_setzero_si512(); + const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE); + const char * b_qs = b_ptr; + const char * b_qh = b_ptr + offset_qh; + int mask = 0; + for (int k_group = 0; k_group < QK_K / 16; ++k_group) { + int r = k_group >> 2; + __m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]); + __m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]); + + __m512i vsum = _mm512_setzero_si512(); + __m512i hmask = _mm512_set1_epi8(0x3); + + __m512i bytes = _mm512_loadu_si512(b_qs); + __m512i hbits = _mm512_loadu_si512(b_qh); + __m512i vb0 = _mm512_and_si512(bytes, lowMask); + __m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + __m512i vh0 = _mm512_slli_epi16(_mm512_and_si512(hbits, hmask), 4); + __m512i vh1 = _mm512_slli_epi16(_mm512_and_si512(hbits, _mm512_slli_epi16(hmask, 2)), 2); + + vb0 = _mm512_add_epi8(vb0, vh0); + vb1 = _mm512_add_epi8(vb1, vh1); + vsum = _mm512_dpbusd_epi32(vsum, vb0, va0); + vsum = _mm512_dpbusd_epi32(vsum, vb1, va1); + b_qs += 64; + + va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]); + va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]); + + bytes = _mm512_loadu_si512(b_qs); + vb0 = _mm512_and_si512(bytes, lowMask); + vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + vh0 = _mm512_and_si512(hbits, _mm512_slli_epi16(hmask, 4)); + vh1 = _mm512_srli_epi16(_mm512_and_si512(hbits, _mm512_slli_epi16(hmask, 6)), 2); + vb0 = _mm512_add_epi8(vb0, vh0); + vb1 = _mm512_add_epi8(vb1, vh1); + vsum = _mm512_dpbusd_epi32(vsum, vb0, va0); + vsum = _mm512_dpbusd_epi32(vsum, vb1, va1); + b_qs += 64; + b_qh += 64; + + // B * A - 32 * A + __m512i vmask = _mm512_set1_epi32(k_group); + vsum = _mm512_sub_epi32(vsum, _mm512_permutexvar_epi32(vmask, vcomp)); + + // vacc += scale * (q8 @ q6) + const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N))); + acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale)); + } + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0))); + vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]); + }; + + for (int i = 0; i < KB; ++i) { + Unroll{}(compute, i); + } + + //store to C + auto storec = [&](int col) { + _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]); + }; + Unroll{}(storec); + } +}; + +template +struct tinygemm_kernel_vnni { + static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + + constexpr int COLS = BLOCK_N / 16; + const int TILE_SIZE = TILE_N * sizeof(block_iq4_xs) + TILE_N * 2; + + const block_q8_K * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + // load the 256 bytes from A to 4 avx512 vectors + __m512i va[4]; + __m512 vc[COLS]; + __m512 vd1; + + // packed_B: + const int offset_scales = (QK_K / 2) * TILE_N ; + const int offset_d0 = (QK_K / 2) * TILE_N + 8 * TILE_N; + + // compensation + __m512i vcomp; + + const __m256i m128s = _mm256_set1_epi16(128); + const __m512i lowMask = _mm512_set1_epi8(0xF); + + const __m512i values128 = _mm512_set_epi8( + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127, + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127, + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127, + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127 + ); + const __m512i off = _mm512_set1_epi8(static_cast(0x80)); + const __m512i values256 = _mm512_add_epi8(values128, off); + + auto loadc = [&](int col) { + vc[col] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + auto compute = [&](int col, int i) { + if (col == 0) { + // load a + va[0] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 0)); + va[1] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 64)); + va[2] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 128)); + va[3] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 192)); + + // compensation: 128 * A + const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums); + vcomp = _mm512_castsi256_si512(_mm256_madd_epi16(q8sums, m128s)); + vd1 = _mm512_set1_ps(A[0 * KB + i].d); + } + + // accmulate the quants + __m512i acc = _mm512_setzero_si512(); + const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE); + const char * b_qs = b_ptr; + int mask = 0; + for (int k_group = 0; k_group < QK_K / 32; ++k_group) { + int r = k_group >> 1; + __m512i vmask = _mm512_set1_epi32(k_group); + __m512i vsum = _mm512_setzero_si512(); + for (int k = 0; k < 8; k += 2) { + __m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]); + __m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]); + + __m512i bytes = _mm512_loadu_si512(b_qs); + __m512i vb0 = _mm512_shuffle_epi8(values256, _mm512_and_si512(bytes, lowMask)); + __m512i vb1 = _mm512_shuffle_epi8(values256, _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask)); + + vsum = _mm512_dpbusd_epi32(vsum, vb0, va0); + vsum = _mm512_dpbusd_epi32(vsum, vb1, va1); + b_qs += 64; + } + // (B + 128) * A - 128 * A + vsum = _mm512_sub_epi32(vsum, _mm512_permutexvar_epi32(vmask, vcomp)); + + // vacc += scale * (q8 @ q4) + const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N))); + acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale)); + } + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0))); + vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]); + }; + + for (int i = 0; i < KB; ++i) { + Unroll{}(compute, i); + } + + //store to C + auto storec = [&](int col) { + _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]); + }; + Unroll{}(storec); + } +}; + +#define LAUNCH_TINYGEMM_KERNEL_VNNI(NB_SIZE) \ + tinygemm_kernel_vnni::apply( \ + KB, (const char *)wdata + 0 * row_size_A, \ + (const char *)src0->data + PACKED_INDEX(nb * kTilesN, 0, KB, TILE_SIZE), \ + (float *) dst->data + 0 * N + nb_start, ldc) + +template ::value, int>::type = 0> +void tinygemm_kernel_amx(int M, int N, int KB, const void * RESTRICT _A, const void * RESTRICT _B, TC * RESTRICT C, int ldc) { + using packed_B_t = packed_B_type; + const int TILE_SIZE = get_tile_size(); + const bool need_unpack = do_unpack::value; + + GGML_ASSERT(M <= 2 * TILE_M && N == 2 * TILE_N); + const TA * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + const int m0 = std::min(M, TILE_M); + const int m1 = std::max(M - TILE_M, 0); + const int lda = KB * sizeof(TA); + //const int ldb = KB * sizeof(TB); + + static thread_local packed_B_t Tile0[TILE_N * TILE_K]; + static thread_local packed_B_t Tile1[TILE_N * TILE_K]; + static thread_local int8_t Tile23[TILE_M * TILE_K]; + + static thread_local int32_t TileC0[TILE_M * TILE_N * 4]; + static thread_local int32_t TileC1[TILE_M * TILE_N * 4]; + + // double buffering C to interleave avx512 and amx + int32_t * C_cur = TileC0; + int32_t * C_pre = TileC1; + + auto Tile4 = [&](int32_t * base) { return base; }; + auto Tile5 = [&](int32_t * base) { return base + TILE_M * TILE_N; }; + auto Tile6 = [&](int32_t * base) { return base + 2 * TILE_M * TILE_N; }; + auto Tile7 = [&](int32_t * base) { return base + 3 * TILE_M * TILE_N; }; + + if (M == 2 * TILE_M) { + // i = 0 + const char * B_blk0 = B + PACKED_INDEX(0, 0, KB, TILE_SIZE); + const char * B_blk1 = B + PACKED_INDEX(1, 0, KB, TILE_SIZE); + if (need_unpack) { + unpack_B(Tile0, B_blk0); + _tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK); + } else { + _tile_loadd(TMM0, B_blk0, TILE_N * VNNI_BLK); + } + + _tile_zero(TMM4); + _tile_loadd(TMM2, A[0].qs, lda); + _tile_dpbssd(TMM4, TMM2, TMM0); + _tile_stored(TMM4, Tile4(C_pre), TILE_N * sizeof(int32_t)); + + _tile_zero(TMM5); + _tile_loadd(TMM3, A[TILE_M * KB + 0].qs, lda); + _tile_dpbssd(TMM5, TMM3, TMM0); + _tile_stored(TMM5, Tile5(C_pre), TILE_N * sizeof(int32_t)); + + if (need_unpack) { + unpack_B(Tile1, B_blk0); + _tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK); + } else { + _tile_loadd(TMM1, B_blk1, TILE_N * VNNI_BLK); + } + + _tile_zero(TMM6); + _tile_dpbssd(TMM6, TMM2, TMM1); + _tile_stored(TMM6, Tile6(C_pre), TILE_N * sizeof(int32_t)); + + _tile_zero(TMM7); + _tile_dpbssd(TMM7, TMM3, TMM1); + _tile_stored(TMM7, Tile7(C_pre), TILE_N * sizeof(int32_t)); + + for (int i = 1; i < KB; ++i) { + // index of previous iter + const int ii = i - 1; + const char * B_blk0 = B + PACKED_INDEX(0, i, KB, TILE_SIZE); + const char * B_blk1 = B + PACKED_INDEX(1, i, KB, TILE_SIZE); + GGML_DISPATCH_BOOL(ii > 0, is_acc, [&] { + if (need_unpack) { + unpack_B(Tile0, B_blk0); + _tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK); + } else { + _tile_loadd(TMM0, B_blk0, TILE_N * VNNI_BLK); + } + _tile_zero(TMM4); + _tile_loadd(TMM2, A[i].qs, lda); + acc_C::apply(C, ldc, Tile4(C_pre), &A[ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M); + + _tile_dpbssd(TMM4, TMM2, TMM0); + _tile_stored(TMM4, Tile4(C_cur), TILE_N * sizeof(int32_t)); + + _tile_zero(TMM5); + _tile_loadd(TMM3, A[TILE_M * KB + i].qs, lda); + acc_C::apply(C + TILE_M * ldc, ldc, Tile5(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M); + + _tile_dpbssd(TMM5, TMM3, TMM0); + _tile_stored(TMM5, Tile5(C_cur), TILE_N * sizeof(int32_t)); + + if (need_unpack) { + unpack_B(Tile1, B_blk1); + _tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK); + } else { + _tile_loadd(TMM1, B_blk1, TILE_N * VNNI_BLK); + } + _tile_zero(TMM6); + acc_C::apply(C + TILE_N, ldc, Tile6(C_pre), &A[ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M); + + _tile_dpbssd(TMM6, TMM2, TMM1); + _tile_stored(TMM6, Tile6(C_cur), TILE_N * sizeof(int32_t)); + + _tile_zero(TMM7); + acc_C::apply(C + TILE_M * ldc + TILE_N, ldc, Tile7(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M); + + _tile_dpbssd(TMM7, TMM3, TMM1); + _tile_stored(TMM7, Tile7(C_cur), TILE_N * sizeof(int32_t)); + + std::swap(C_cur, C_pre); + }); + } + // final accumulation + { + int ii = KB - 1; + acc_C::apply(C, ldc, Tile4(C_pre), &A[ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M); + acc_C::apply(C + TILE_M * ldc, ldc, Tile5(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M); + acc_C::apply(C + TILE_N, ldc, Tile6(C_pre), &A[ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M); + acc_C::apply(C + TILE_M * ldc + TILE_N, ldc, Tile7(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M); + } + } else { + for (int i = 0; i < KB; ++i) { + _tile_zero(TMM4); + _tile_zero(TMM6); + if (m1 != 0) { + _tile_zero(TMM5); + _tile_zero(TMM7); + } + + const char * B_blk0 = B + PACKED_INDEX(0, i, KB, TILE_SIZE); + const char * B_blk1 = B + PACKED_INDEX(1, i, KB, TILE_SIZE); + if (need_unpack) { + unpack_B(Tile0, B_blk0); + _tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK); + } else { + _tile_loadd(TMM0, B_blk0, TILE_N * VNNI_BLK); + } + + if (need_unpack) { + unpack_B(Tile1, B_blk1); + _tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK); + } else { + _tile_loadd(TMM1, B_blk1, TILE_N * VNNI_BLK); + } + + if (m0 == TILE_M) { + _tile_loadd(TMM2, A[i].qs, lda); + } else { + unpack_A(Tile23, &A[i], KB, m0); + _tile_loadd(TMM2, Tile23, TILE_K); + } + + _tile_dpbssd(TMM4, TMM2, TMM0); + _tile_dpbssd(TMM6, TMM2, TMM1); + + _tile_stored(TMM4, Tile4(C_cur), TILE_N * sizeof(int32_t)); + _tile_stored(TMM6, Tile6(C_cur), TILE_N * sizeof(int32_t)); + + GGML_DISPATCH_BOOL(i > 0, is_acc, [&] { + acc_C::apply(C, ldc, Tile4(C_cur), &A[i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m0); + acc_C::apply(C + TILE_N, ldc, Tile6(C_cur), &A[i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m0); + }); + + if (m1 != 0) { + unpack_A(Tile23, &A[TILE_M * KB + i], KB, m1); + _tile_loadd(TMM3, Tile23, TILE_K); + + _tile_dpbssd(TMM5, TMM3, TMM0); + _tile_dpbssd(TMM7, TMM3, TMM1); + _tile_stored(TMM5, Tile5(C_cur), TILE_N * sizeof(int32_t)); + _tile_stored(TMM7, Tile7(C_cur), TILE_N * sizeof(int32_t)); + GGML_DISPATCH_BOOL(i > 0, is_acc, [&] { + acc_C::apply(C + TILE_M * ldc, ldc, Tile5(C_cur), &A[TILE_M * KB + i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m1); + acc_C::apply(C + TILE_M * ldc + TILE_N, ldc, Tile7(C_cur), &A[TILE_M * KB + i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m1); + }); + } + } + } + return; +} + +template ::value, int>::type = 0> +void tinygemm_kernel_amx(int M, int N, int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + static_assert(std::is_same::value); + const int TILE_SIZE = get_tile_size(); + + GGML_ASSERT(M <= 2 * TILE_M && N == 2 * TILE_N); + const TA * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + const int m0 = std::min(M, TILE_M); + const int m1 = std::max(M - TILE_M, 0); + //const int lda = KB * sizeof(TA); + + static thread_local int8_t Tile0[TILE_N * TILE_K]; + static thread_local int8_t Tile1[TILE_N * TILE_K]; + static thread_local int8_t Tile23[TILE_M * TILE_K]; + + // mat mul result for each group + static thread_local int32_t Tile4[TILE_M * TILE_N]; + static thread_local int32_t Tile5[TILE_M * TILE_N]; + static thread_local int32_t Tile6[TILE_M * TILE_N]; + static thread_local int32_t Tile7[TILE_M * TILE_N]; + + // sum of each QK_K block, contains 8 groups, int32 + static thread_local int32_t Sumi4[TILE_M * TILE_N]; + static thread_local int32_t Sumi5[TILE_M * TILE_N]; + static thread_local int32_t Sumi6[TILE_M * TILE_N]; + static thread_local int32_t Sumi7[TILE_M * TILE_N]; + + const int k_group_size = std::is_same::value ? 16 : 32; + for (int i = 0; i < KB; ++i) { + // step 1: accumulate the quants across 8 groups, each group with 32 + for (int k = 0; k < QK_K / k_group_size; ++k) { + GGML_DISPATCH_BOOL(k > 0, is_acc, [&] { + _tile_zero(TMM4); + _tile_zero(TMM6); + + unpack_B(Tile0, B + PACKED_INDEX(0, i, KB, TILE_SIZE), k); + _tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK); + + unpack_B(Tile1, B + PACKED_INDEX(1, i, KB, TILE_SIZE), k); + _tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK); + + unpack_A(Tile23, &A[i], KB, k, m0); + _tile_loadd(TMM2, Tile23, TILE_K); + + _tile_dpbssd(TMM4, TMM2, TMM0); + _tile_dpbssd(TMM6, TMM2, TMM1); + + _tile_stored(TMM4, Tile4, TILE_N * sizeof(int32_t)); + _tile_stored(TMM6, Tile6, TILE_N * sizeof(int32_t)); + + scale_C(Tile4, Sumi4, B + PACKED_INDEX(0, i, KB, TILE_SIZE), k, m0); + scale_C(Tile6, Sumi6, B + PACKED_INDEX(1, i, KB, TILE_SIZE), k, m0); + + if (m1 != 0) { + _tile_zero(TMM5); + _tile_zero(TMM7); + + unpack_A(Tile23, &A[TILE_M * KB + i], KB, k, m1); + _tile_loadd(TMM3, Tile23, TILE_K); + + _tile_dpbssd(TMM5, TMM3, TMM0); + _tile_dpbssd(TMM7, TMM3, TMM1); + + _tile_stored(TMM5, Tile5, TILE_N * sizeof(int32_t)); + _tile_stored(TMM7, Tile7, TILE_N * sizeof(int32_t)); + + scale_C(Tile5, Sumi5, B + PACKED_INDEX(0, i, KB, TILE_SIZE), k, m1); + scale_C(Tile7, Sumi7, B + PACKED_INDEX(1, i, KB, TILE_SIZE), k, m1); + } + }); + } + + // step 2: accmulate the mins + GGML_DISPATCH_BOOL(i > 0, is_acc, [&] { + acc_C::apply(C, ldc, Sumi4, &A[i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m0); + acc_C::apply(C + TILE_N, ldc, Sumi6, &A[i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m0); + if (m1 != 0) { + acc_C::apply(C + TILE_M * ldc, ldc, Sumi5, &A[TILE_M * KB + i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m1); + acc_C::apply(C + TILE_M * ldc + TILE_N, ldc, Sumi7, &A[TILE_M * KB + i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m1); + } + }); + } + return; +} + +} // anonymous namespace + +// get the packed tensor size for quantized weights +size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor) { + const enum ggml_type TYPE = tensor->type; + + const int K = tensor->ne[0]; // ne0: in_features + const int N = tensor->ne[1]; // ne1: out_features + + auto get_tensor_size = [&] { + size_t row_size_B{0}; + GGML_DISPATCH_QTYPES(TYPE, [&] { + row_size_B = get_row_size(K); + }); + return N * row_size_B; + }; + + if (qtype_has_amx_kernels(TYPE)) { + return get_tensor_size(); + } else { + // for f16, bf16 we don't do packing + return ggml_nbytes(tensor); + } +} + +// pack weight to vnni format +void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { + + size_t alloc_size = ggml_backend_amx_get_alloc_size(tensor); + GGML_ASSERT(alloc_size == size); + + const enum ggml_type TYPE = tensor->type; + + const int K = tensor->ne[0]; // ne0: in_features + const int N = tensor->ne[1]; // ne1: out_features + +#if defined(_OPENMP) + // the buffer ctx is not initialized when .set_tensor is called + int n_threads = omp_get_num_threads(); +#else + int n_threads = 1; +#endif + + GGML_DISPATCH_QTYPES(TYPE, [&] { + convert_B_packed_format((void *)((char *)tensor->data + offset), (const type *)data, N, K, n_threads); + }); +} + +// NB: mixed dtype gemm with Advanced Matrix Extensions (Intel AMX) +// +// src0: weight in shape of {N, K}, quantized +// src1: input in shape of {M, K}, float32 +// dst: output in shape of {M, N}, float32 +// +// the function performs: dst = src1 @ src0.T +// +void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor * dst) { + struct ggml_tensor * src0 = dst->src[0]; + struct ggml_tensor * src1 = dst->src[1]; + + const enum ggml_type TYPE = src0->type; + + const int n_threads = ctx->n_threads; + + // f16 only has avx512 kernels for now, + // amx kernels will be added once 6th gen xeon is released. + const bool is_floating_type = TYPE == GGML_TYPE_F16; + + const int M = dst->ne[1]; + const int N = dst->ne[0]; + const int K = src0->ne[0]; + const int ldc = dst->nb[1] / dst->nb[0]; + + if (is_floating_type) { + constexpr int BLOCK_M = 4; + constexpr int BLOCK_N = 6; + const int MB = div_up(M, BLOCK_M); + const int NB = div_up(N, BLOCK_N); + + parallel_for(n_threads, MB * NB, [&](int begin, int end) { + GGML_DISPATCH_FLOATING_TYPES(TYPE, [&] { + for (int i = begin; i < end; ++i) { + int mb = i / NB; + int nb = i % NB; + + int mb_start = mb * BLOCK_M; + int mb_size = std::min(BLOCK_M, M - mb_start); + int nb_start = nb * BLOCK_N; + int nb_size = std::min(BLOCK_N, N - nb_start); + + switch (mb_size << 4 | nb_size) { + case 0x12: LAUNCH_TINYGEMM_KERNEL_AVX(1, 2); break; + case 0x14: LAUNCH_TINYGEMM_KERNEL_AVX(1, 4); break; + case 0x16: LAUNCH_TINYGEMM_KERNEL_AVX(1, 6); break; + case 0x22: LAUNCH_TINYGEMM_KERNEL_AVX(2, 2); break; + case 0x24: LAUNCH_TINYGEMM_KERNEL_AVX(2, 4); break; + case 0x26: LAUNCH_TINYGEMM_KERNEL_AVX(2, 6); break; + case 0x32: LAUNCH_TINYGEMM_KERNEL_AVX(3, 2); break; + case 0x34: LAUNCH_TINYGEMM_KERNEL_AVX(3, 4); break; + case 0x36: LAUNCH_TINYGEMM_KERNEL_AVX(3, 6); break; + case 0x42: LAUNCH_TINYGEMM_KERNEL_AVX(4, 2); break; + case 0x44: LAUNCH_TINYGEMM_KERNEL_AVX(4, 4); break; + case 0x46: LAUNCH_TINYGEMM_KERNEL_AVX(4, 6); break; + default: fprintf(stderr, "Unexpected block size!\n"); + } + } + }); + }); + return; + } + + // pointer to work space, used convert A from float to quantized type + void * wdata = nullptr; + + //TODO: performance improvement: merge quant A + GGML_DISPATCH_QTYPES(TYPE, [&] { + const size_t row_size_A = K / blck_size * sizeof(vec_dot_type); + const size_t desired_wsize = M * row_size_A; + if (ctx->work_size < desired_wsize) { + ctx->work_data.reset(new char[desired_wsize]); + ctx->work_size = desired_wsize; + } + wdata = ctx->work_data.get(); + + // Q4_0, Q4_1, Q8_0 handles 1 TILE_K per blck_size + // Q4_K, Q5_K, Q6_K, IQ4_XS handles 8 TILE_K per blck_size + GGML_ASSERT(TILE_K == blck_size || TILE_K * 8 == blck_size); + + const float * A_data = static_cast(src1->data); + for (int m = 0; m < M; ++m) { + from_float(A_data + m * K, (char *)wdata + m * row_size_A, K); + } + }); + + if (M == 1) { + // MB = 1 and handle 8 tiles in each block + constexpr int kTilesN = 4; + constexpr int BLOCK_N = TILE_N * kTilesN; + const int NB = div_up(N, BLOCK_N); + + parallel_for(n_threads, NB, [&](int begin, int end) { + GGML_DISPATCH_QTYPES(TYPE, [&] { + const int KB = K / blck_size; + const int TILE_SIZE = get_tile_size(); + const int row_size_A = KB * sizeof(vec_dot_type); + for (int i = begin; i < end; ++i) { + int nb = i; + int nb_start = nb * BLOCK_N; + int nb_size = std::min(BLOCK_N, N - nb_start); // 32, 64, 96 + + switch (nb_size) { + //case 160: LAUNCH_TINYGEMM_KERNEL_VNNI(160); break; + case 128: LAUNCH_TINYGEMM_KERNEL_VNNI(128); break; + case 96: LAUNCH_TINYGEMM_KERNEL_VNNI(96); break; + case 64: LAUNCH_TINYGEMM_KERNEL_VNNI(64); break; + case 32: LAUNCH_TINYGEMM_KERNEL_VNNI(32); break; + default: fprintf(stderr, "Unexpected n block size!\n"); + } + } + }); + }); + return; + } + + // handle 4 tiles at a tile + constexpr int BLOCK_M = TILE_M * 2; + constexpr int BLOCK_N = TILE_N * 2; + const int MB = div_up(M, BLOCK_M); + const int NB = div_up(N, BLOCK_N); + + parallel_for(n_threads, MB * NB, [&](int begin, int end) { + // init tile config for each thread + ggml_tile_config_init(); + + GGML_DISPATCH_QTYPES(TYPE, [&] { + const int KB = K / blck_size; + const int TILE_SIZE = get_tile_size(); + const int row_size_A = KB * sizeof(vec_dot_type); + + for (int i = begin; i < end; ++i) { + int mb = i / NB; + int nb = i % NB; + + int mb_start = mb * BLOCK_M; + int mb_size = std::min(BLOCK_M, M - mb_start); + int nb_start = nb * BLOCK_N; + int nb_size = BLOCK_N; + + tinygemm_kernel_amx( + mb_size, nb_size, KB, + (const char *)wdata + mb_start * row_size_A, + (const char *)src0->data + PACKED_INDEX(nb * 2, 0, KB, TILE_SIZE), + (float *) dst->data + mb_start * N + nb_start, ldc); + } + }); + }); +} + +#else // if defined(__AMX_INT8__) + +void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor * dst) { + fprintf(stderr, "GGML is not compiled with AMX support!\n"); + + GGML_UNUSED(ctx); + GGML_UNUSED(dst); +} + +#endif // if defined(__AMX_INT8__) diff --git a/ggml/src/ggml-amx/mmq.h b/ggml/src/ggml-amx/mmq.h new file mode 100644 index 0000000000000..cf09206206370 --- /dev/null +++ b/ggml/src/ggml-amx/mmq.h @@ -0,0 +1,17 @@ +#pragma once +#include "common.h" +#include + +#ifdef __cplusplus +extern "C" { +#endif + +size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor); + +void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); + +void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor * dst); + +#ifdef __cplusplus +} +#endif diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp index a3bc79a46b871..1c17dde30ae0d 100644 --- a/ggml/src/ggml-backend.cpp +++ b/ggml/src/ggml-backend.cpp @@ -329,7 +329,6 @@ bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type if (backend->device) { return ggml_backend_dev_supports_buft(backend->device, buft); } - return backend->iface.supports_buft(backend, buft); } @@ -550,6 +549,14 @@ void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * na #include "ggml-rpc.h" #endif +#ifndef __AMX_INT8__ +#undef GGML_USE_AMX +#endif + +#ifdef GGML_USE_AMX +# include "ggml-amx.h" +#endif + struct ggml_backend_registry { std::vector backends; std::vector devices; @@ -570,6 +577,9 @@ struct ggml_backend_registry { #ifdef GGML_USE_RPC register_backend(ggml_backend_rpc_reg()); #endif +#ifdef GGML_USE_AMX + register_backend(ggml_backend_amx_reg()); +#endif // TODO: sycl, kompute, cann diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 779b38d1213f2..7e24313edb269 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -23252,6 +23252,14 @@ int ggml_cpu_has_avx512_bf16(void) { #endif } +int ggml_cpu_has_amx_int8(void) { +#if defined(__AMX_INT8__) + return 1; +#else + return 0; +#endif +} + int ggml_cpu_has_fma(void) { #if defined(__FMA__) return 1; diff --git a/src/llama.cpp b/src/llama.cpp index dcb015d12544b..0025e94b89d20 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -16,6 +16,14 @@ # include "ggml-cann.h" #endif +#ifndef __AMX_INT8__ +#undef GGML_USE_AMX +#endif + +#ifdef GGML_USE_AMX +# include "ggml-amx.h" +#endif + // TODO: replace with ggml API call #define QK_K 256 @@ -3533,6 +3541,7 @@ static size_t llama_get_device_memory(const llama_model & model, int device) { #else return 1; #endif + GGML_UNUSED(model); GGML_UNUSED(device); } @@ -7031,7 +7040,14 @@ static bool llm_load_tensors( // assign cpu layers for (int i = 0; i < i_gpu_start; ++i) { +#ifdef GGML_USE_AMX + model.buft_layer[i] = { + ggml_backend_amx_buffer_type(), + llama_default_buffer_type_cpu(model, true) + }; +#else model.buft_layer[i] = llama_default_buffer_type_cpu(model, true); +#endif } if (split_mode == LLAMA_SPLIT_MODE_LAYER) { @@ -21839,6 +21855,7 @@ const char * llama_print_system_info(void) { s += "AVX512_VBMI = " + std::to_string(ggml_cpu_has_avx512_vbmi()) + " | "; s += "AVX512_VNNI = " + std::to_string(ggml_cpu_has_avx512_vnni()) + " | "; s += "AVX512_BF16 = " + std::to_string(ggml_cpu_has_avx512_bf16()) + " | "; + s += "AMX_INT8 = " + std::to_string(ggml_cpu_has_amx_int8()) + " | "; s += "FMA = " + std::to_string(ggml_cpu_has_fma()) + " | "; s += "NEON = " + std::to_string(ggml_cpu_has_neon()) + " | "; s += "SVE = " + std::to_string(ggml_cpu_has_sve()) + " | ";