root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-11-13 14:29:52 +00:00
Code Issues Actions 11 Packages Projects Releases Wiki Activity

add amx kernel for gemm (#8998)

Browse Source 
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
This commit is contained in:
Ma Mingfei 2024-10-18 13:34:36 +08:00 committed by GitHub
parent 8901755ba3
commit 60ce97c9d8
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
14 changed files with 3204 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
CMakeLists.txt
View File

@ -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()

24
Makefile
View File

@ -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

2
README.md
View File

@ -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

4
ggml/CMakeLists.txt
View File

@ -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

25
ggml/include/ggml-amx.h Normal file
View File

@ -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

1
ggml/include/ggml.h
View File

@ -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);

42
ggml/src/CMakeLists.txt
View File

@ -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($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
endif()
if (GGML_AMX_TILE)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
endif()
if (GGML_AMX_INT8)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
endif()
if (GGML_AMX_BF16)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__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
)

453
ggml/src/ggml-amx.cpp Normal file
View File

@ -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 <sys/syscall.h>
#include <unistd.h>
#endif
#include <cstdlib>
#include <cstring>
#include <memory>
#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

93
ggml/src/ggml-amx/common.h Normal file
View File

@ -0,0 +1,93 @@
#pragma once
#include "ggml.h"
#include "ggml-cpu-impl.h" // <immintrin.h>
#include <algorithm>
#include <memory>
#include <type_traits>
#if defined(_OPENMP)
#include <omp.h>
#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 <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
inline T div_up(T x, T y) { return (x + y - 1) / y; }
template <typename T>
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 <typename func_t>
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<char[]> work_data;
size_t work_size = 0;
};

2509
ggml/src/ggml-amx/mmq.cpp Normal file
View File

File diff suppressed because it is too large Load Diff

17
ggml/src/ggml-amx/mmq.h Normal file
View File

@ -0,0 +1,17 @@
#pragma once
#include "common.h"
#include <stdint.h>
#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

12
ggml/src/ggml-backend.cpp
View File

@ -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<ggml_backend_reg_t> backends;
std::vector<ggml_backend_dev_t> 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

8
ggml/src/ggml.c
View File

@ -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;

17
src/llama.cpp
View File

@ -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()) + " | ";