civ/hot-step-cpp-ROCm
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Initial release

Browse Source
This commit is contained in:
civ
2026-08-16 18:24:52 +07:00
commit 876886a39a
13244 changed files with 2353959 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
engine/ggml/src/CMakeLists.txt
+517
View File
@@ -0,0 +1,517 @@
include(CheckCXXCompilerFlag)
include("../cmake/common.cmake")
add_compile_definitions(GGML_SCHED_MAX_COPIES=${GGML_SCHED_MAX_COPIES})
# enable libstdc++ assertions for debug builds
if (CMAKE_SYSTEM_NAME MATCHES "Linux")
add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
endif()
if (NOT MSVC)
if (GGML_SANITIZE_THREAD)
add_compile_options(-fsanitize=thread)
link_libraries (-fsanitize=thread)
endif()
if (GGML_SANITIZE_ADDRESS)
add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
link_libraries (-fsanitize=address)
endif()
if (GGML_SANITIZE_UNDEFINED)
add_compile_options(-fsanitize=undefined)
link_libraries (-fsanitize=undefined)
endif()
endif()
if (GGML_FATAL_WARNINGS)
if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
list(APPEND C_FLAGS -Werror)
list(APPEND CXX_FLAGS -Werror)
elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
add_compile_options(/WX)
endif()
endif()
if (GGML_ALL_WARNINGS)
if (NOT MSVC)
list(APPEND WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
list(APPEND C_FLAGS -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
-Werror=implicit-int -Werror=implicit-function-declaration)
list(APPEND CXX_FLAGS -Wmissing-declarations -Wmissing-noreturn)
list(APPEND C_FLAGS ${WARNING_FLAGS})
list(APPEND CXX_FLAGS ${WARNING_FLAGS})
ggml_get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
add_compile_options("$<$<COMPILE_LANGUAGE:C>:${C_FLAGS};${GF_C_FLAGS}>"
"$<$<COMPILE_LANGUAGE:CXX>:${CXX_FLAGS};${GF_CXX_FLAGS}>")
else()
# todo : msvc
set(C_FLAGS "")
set(CXX_FLAGS "")
endif()
endif()
if (GGML_LTO)
include(CheckIPOSupported)
check_ipo_supported(RESULT result OUTPUT output)
if (result)
set(CMAKE_INTERPROCEDURAL_OPTIMIZATION TRUE)
else()
message(WARNING "IPO is not supported: ${output}")
endif()
endif()
if (GGML_CCACHE AND NOT CMAKE_C_COMPILER_LAUNCHER AND NOT CMAKE_CXX_COMPILER_LAUNCHER)
find_program(GGML_CCACHE_FOUND ccache)
find_program(GGML_SCCACHE_FOUND sccache)
if (GGML_CCACHE_FOUND OR GGML_SCCACHE_FOUND)
if(GGML_CCACHE_FOUND)
set(GGML_CCACHE_VARIANT ccache)
else()
set(GGML_CCACHE_VARIANT sccache)
endif()
# TODO: should not be set globally
if (GGML_SYCL AND GGML_CCACHE_FOUND AND WIN32)
set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "ccache compiler_type=icl")
else ()
set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "${GGML_CCACHE_VARIANT}")
endif ()
set(ENV{CCACHE_SLOPPINESS} time_macros)
message(STATUS "${GGML_CCACHE_VARIANT} found, compilation results will be cached. Disable with GGML_CCACHE=OFF.")
else()
message(STATUS "Warning: ccache not found - consider installing it for faster compilation or disable this warning with GGML_CCACHE=OFF")
endif ()
endif()
# this version of Apple ld64 is buggy
execute_process(
COMMAND ${CMAKE_C_COMPILER} ${CMAKE_EXE_LINKER_FLAGS} -Wl,-v
ERROR_VARIABLE output
OUTPUT_QUIET
)
if (output MATCHES "dyld-1015\.7")
add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
endif()
# architecture specific
# TODO: probably these flags need to be tweaked on some architectures
# feel free to update the Makefile for your architecture and send a pull request or issue
message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
if (MSVC)
string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
else ()
set(CMAKE_GENERATOR_PLATFORM_LWR "")
endif ()
ggml_get_system_arch()
message(STATUS "GGML_SYSTEM_ARCH: ${GGML_SYSTEM_ARCH}")
if (NOT MSVC)
if (GGML_STATIC)
if (UNIX AND NOT APPLE)
set(CMAKE_FIND_LIBRARY_SUFFIXES ".a;.so")
endif()
add_link_options(-static)
if (MINGW)
add_link_options(-static-libgcc -static-libstdc++)
endif()
endif()
if (GGML_GPROF)
add_compile_options(-pg)
endif()
endif()
#
# POSIX conformance
#
# clock_gettime came in POSIX.1b (1993)
# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
# posix_memalign came in POSIX.1-2001 / SUSv3
# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
# Somehow in OpenBSD whenever POSIX conformance is specified
# some string functions rely on locale_t availability,
# which was introduced in POSIX.1-2008, forcing us to go higher
if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
add_compile_definitions(_XOPEN_SOURCE=700)
elseif (CMAKE_SYSTEM_NAME MATCHES "AIX")
# Don't define _XOPEN_SOURCE. We need _ALL_SOURCE, which is the default,
# in order to define _SC_PHYS_PAGES.
else()
add_compile_definitions(_XOPEN_SOURCE=600)
endif()
# Data types, macros and functions related to controlling CPU affinity and
# some memory allocation are available on Linux through GNU extensions in libc
if (CMAKE_SYSTEM_NAME MATCHES "Linux" OR CMAKE_SYSTEM_NAME MATCHES "Android")
add_compile_definitions(_GNU_SOURCE)
endif()
# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
# and on macOS its availability depends on enabling Darwin extensions
# similarly on DragonFly, enabling BSD extensions is necessary
if (
CMAKE_SYSTEM_NAME MATCHES "Darwin" OR
CMAKE_SYSTEM_NAME MATCHES "iOS" OR
CMAKE_SYSTEM_NAME MATCHES "tvOS" OR
CMAKE_SYSTEM_NAME MATCHES "DragonFly"
)
add_compile_definitions(_DARWIN_C_SOURCE)
endif()
# alloca is a non-standard interface that is not visible on BSDs when
# POSIX conformance is specified, but not all of them provide a clean way
# to enable it in such cases
if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
add_compile_definitions(__BSD_VISIBLE)
endif()
if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
add_compile_definitions(_NETBSD_SOURCE)
endif()
if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
add_compile_definitions(_BSD_SOURCE)
endif()
if (WIN32)
add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
endif()
# ggml
if (GGML_BACKEND_DL AND NOT BUILD_SHARED_LIBS)
message(FATAL_ERROR "GGML_BACKEND_DL requires BUILD_SHARED_LIBS")
endif()
add_library(ggml-base
../include/ggml.h
../include/ggml-alloc.h
../include/ggml-backend.h
../include/ggml-cpp.h
../include/ggml-opt.h
../include/gguf.h
ggml.c
ggml.cpp
ggml-alloc.c
ggml-backend.cpp
ggml-backend-meta.cpp
ggml-opt.cpp
ggml-threading.cpp
ggml-threading.h
ggml-quants.c
ggml-quants.h
gguf.cpp)
set_target_properties(ggml-base PROPERTIES
VERSION ${GGML_VERSION}
SOVERSION ${GGML_VERSION_MAJOR}
)
target_include_directories(ggml-base PRIVATE .)
if (GGML_BACKEND_DL)
target_compile_definitions(ggml-base PUBLIC GGML_BACKEND_DL)
endif()
if (GGML_SCHED_NO_REALLOC)
target_compile_definitions(ggml-base PUBLIC GGML_SCHED_NO_REALLOC)
endif()
if (GGML_OPENMP)
find_package(OpenMP)
if (OpenMP_FOUND)
set(GGML_OPENMP_ENABLED "ON" CACHE INTERNAL "")
else()
set(GGML_OPENMP_ENABLED "OFF" CACHE INTERNAL "")
message(WARNING "OpenMP not found")
endif()
else()
set(GGML_OPENMP_ENABLED "OFF" CACHE INTERNAL "")
endif()
if (GGML_OPENMP_ENABLED)
target_compile_definitions(ggml-base PRIVATE GGML_USE_OPENMP)
target_link_libraries(ggml-base PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
endif()
add_library(ggml
ggml-backend-dl.cpp
ggml-backend-reg.cpp)
add_library(ggml::ggml ALIAS ggml)
set_target_properties(ggml PROPERTIES
VERSION ${GGML_VERSION}
SOVERSION ${GGML_VERSION_MAJOR}
)
if (GGML_BACKEND_DIR)
if (NOT GGML_BACKEND_DL)
message(FATAL_ERROR "GGML_BACKEND_DIR requires GGML_BACKEND_DL")
endif()
target_compile_definitions(ggml PUBLIC GGML_BACKEND_DIR="${GGML_BACKEND_DIR}")
endif()
target_link_libraries(ggml PUBLIC ggml-base)
if (CMAKE_SYSTEM_NAME MATCHES "Linux")
target_link_libraries(ggml PRIVATE dl)
endif()
function(ggml_add_backend_library backend)
if (GGML_BACKEND_DL)
add_library(${backend} MODULE ${ARGN})
# write the shared library to the output directory
set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
add_dependencies(ggml ${backend})
if (GGML_BACKEND_DIR)
install(TARGETS ${backend} LIBRARY DESTINATION ${GGML_BACKEND_DIR})
else()
install(TARGETS ${backend} LIBRARY DESTINATION ${CMAKE_INSTALL_BINDIR})
endif()
else()
add_library(${backend} ${ARGN})
target_link_libraries(ggml PUBLIC ${backend})
install(TARGETS ${backend} LIBRARY)
endif()
target_link_libraries(${backend} PRIVATE ggml-base)
target_include_directories(${backend} PRIVATE ..)
if (${BUILD_SHARED_LIBS})
target_compile_definitions(${backend} PRIVATE GGML_BACKEND_BUILD)
target_compile_definitions(${backend} PUBLIC GGML_BACKEND_SHARED)
endif()
# Set versioning properties for all backend libraries
# Building a MODULE library with a version is not supported on macOS (https://gitlab.kitware.com/cmake/cmake/-/issues/20782)
if (NOT (APPLE AND GGML_BACKEND_DL))
set_target_properties(${backend} PROPERTIES
VERSION ${GGML_VERSION}
SOVERSION ${GGML_VERSION_MAJOR}
)
endif()
if(NOT GGML_AVAILABLE_BACKENDS)
set(GGML_AVAILABLE_BACKENDS "${backend}"
CACHE INTERNAL "List of backends for cmake package")
else()
list(FIND GGML_AVAILABLE_BACKENDS "${backend}" has_backend)
if(has_backend EQUAL -1)
set(GGML_AVAILABLE_BACKENDS "${GGML_AVAILABLE_BACKENDS};${backend}"
CACHE INTERNAL "List of backends for cmake package")
endif()
endif()
endfunction()
function(ggml_add_backend backend)
string(TOUPPER "GGML_${backend}" backend_id)
if (${backend_id})
string(TOLOWER "ggml-${backend}" backend_target)
add_subdirectory(${backend_target})
message(STATUS "Including ${backend} backend")
if (NOT GGML_BACKEND_DL)
string(TOUPPER "GGML_USE_${backend}" backend_use)
target_compile_definitions(ggml PUBLIC ${backend_use})
endif()
endif()
endfunction()
function(ggml_add_cpu_backend_variant tag_name)
set(GGML_CPU_TAG_NAME ${tag_name})
# other: OPENMP LLAMAFILE CPU_HBM
if (GGML_SYSTEM_ARCH STREQUAL "x86")
foreach (feat NATIVE
SSE42
AVX AVX2 BMI2 AVX_VNNI FMA F16C
AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
AMX_TILE AMX_INT8 AMX_BF16)
set(GGML_${feat} OFF)
endforeach()
foreach (feat ${ARGN})
set(GGML_${feat} ON)
endforeach()
elseif (GGML_SYSTEM_ARCH STREQUAL "ARM")
foreach (feat ${ARGN})
set(GGML_INTERNAL_${feat} ON)
endforeach()
elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
foreach (feat ${ARGN})
set(GGML_INTERNAL_${feat} ON)
endforeach()
elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
foreach (feat VXE2 NNPA)
set(GGML_INTERNAL_${feat} OFF)
endforeach()
foreach (feat ${ARGN})
set(GGML_INTERNAL_${feat} ON)
endforeach()
elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
foreach (feat RVV)
set(GGML_INTERNAL_${feat} OFF)
endforeach()
foreach (feat ${ARGN})
set(GGML_INTERNAL_${feat} ON)
endforeach()
endif()
ggml_add_cpu_backend_variant_impl(${tag_name})
endfunction()
ggml_add_backend(CPU)
if (GGML_CPU_ALL_VARIANTS)
if (NOT GGML_BACKEND_DL)
message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
elseif (GGML_CPU_ARM_ARCH)
message(FATAL_ERROR "Cannot use both GGML_CPU_ARM_ARCH and GGML_CPU_ALL_VARIANTS")
endif()
if (GGML_SYSTEM_ARCH STREQUAL "x86")
ggml_add_cpu_backend_variant(x64)
ggml_add_cpu_backend_variant(sse42 SSE42)
ggml_add_cpu_backend_variant(sandybridge SSE42 AVX)
if (NOT MSVC)
# __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
ggml_add_cpu_backend_variant(ivybridge SSE42 AVX F16C)
ggml_add_cpu_backend_variant(piledriver SSE42 AVX F16C FMA)
endif()
ggml_add_cpu_backend_variant(haswell SSE42 AVX F16C FMA AVX2 BMI2)
ggml_add_cpu_backend_variant(skylakex SSE42 AVX F16C FMA AVX2 BMI2 AVX512)
ggml_add_cpu_backend_variant(cannonlake SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VBMI)
ggml_add_cpu_backend_variant(cascadelake SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VNNI)
ggml_add_cpu_backend_variant(icelake SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VBMI AVX512_VNNI)
if (NOT MSVC)
# MSVC 2022 doesn't support BF16 intrinsics without `/arch:AVX10.1` ?!
# https://learn.microsoft.com/en-us/cpp/intrinsics/x64-amd64-intrinsics-list?view=msvc-170
# https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170
ggml_add_cpu_backend_variant(cooperlake SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VNNI AVX512_BF16)
ggml_add_cpu_backend_variant(zen4 SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16)
endif()
ggml_add_cpu_backend_variant(alderlake SSE42 AVX F16C FMA AVX2 BMI2 AVX_VNNI)
if (NOT MSVC)
# MSVC doesn't support AMX
ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C FMA AVX2 BMI2 AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
endif()
elseif(GGML_SYSTEM_ARCH STREQUAL "ARM")
if (CMAKE_SYSTEM_NAME MATCHES "Linux")
# Many of these features are optional so we build versions with popular
# combinations and name the backends based on the version they were
# first released with
ggml_add_cpu_backend_variant(armv8.0_1)
ggml_add_cpu_backend_variant(armv8.2_1 DOTPROD)
ggml_add_cpu_backend_variant(armv8.2_2 DOTPROD FP16_VECTOR_ARITHMETIC)
ggml_add_cpu_backend_variant(armv8.2_3 DOTPROD FP16_VECTOR_ARITHMETIC SVE)
ggml_add_cpu_backend_variant(armv8.6_1 DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8)
ggml_add_cpu_backend_variant(armv8.6_2 DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2)
ggml_add_cpu_backend_variant(armv9.2_1 DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SME)
ggml_add_cpu_backend_variant(armv9.2_2 DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2 SME)
elseif (CMAKE_SYSTEM_NAME MATCHES "Android")
# Android-specific backends with SoC-compatible feature sets
ggml_add_cpu_backend_variant(android_armv8.0_1)
ggml_add_cpu_backend_variant(android_armv8.2_1 DOTPROD)
ggml_add_cpu_backend_variant(android_armv8.2_2 DOTPROD FP16_VECTOR_ARITHMETIC)
ggml_add_cpu_backend_variant(android_armv8.6_1 DOTPROD FP16_VECTOR_ARITHMETIC MATMUL_INT8)
ggml_add_cpu_backend_variant(android_armv9.0_1 DOTPROD MATMUL_INT8 FP16_VECTOR_ARITHMETIC SVE2)
ggml_add_cpu_backend_variant(android_armv9.2_1 DOTPROD MATMUL_INT8 FP16_VECTOR_ARITHMETIC SVE SME)
ggml_add_cpu_backend_variant(android_armv9.2_2 DOTPROD MATMUL_INT8 FP16_VECTOR_ARITHMETIC SVE SVE2 SME)
elseif (APPLE)
ggml_add_cpu_backend_variant(apple_m1 DOTPROD)
ggml_add_cpu_backend_variant(apple_m2_m3 DOTPROD MATMUL_INT8)
ggml_add_cpu_backend_variant(apple_m4 DOTPROD MATMUL_INT8 NOSVE SME)
else()
message(FATAL_ERROR "Unsupported ARM target OS: ${CMAKE_SYSTEM_NAME}")
endif()
elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
if (CMAKE_SYSTEM_NAME MATCHES "Linux")
ggml_add_cpu_backend_variant(power0)
ggml_add_cpu_backend_variant(power7_1 POWER7)
ggml_add_cpu_backend_variant(power7_2 POWER7 VSX)
ggml_add_cpu_backend_variant(power8_1 POWER8)
ggml_add_cpu_backend_variant(power8_2 POWER8 VSX)
ggml_add_cpu_backend_variant(power9 POWER9 VSX)
ggml_add_cpu_backend_variant(power10 POWER10 VSX)
# POWER11 backend: only if compiler supports -mcpu=power11
check_cxx_compiler_flag("-mcpu=power11" GGML_CXX_SUPPORTS_POWER11)
if (GGML_CXX_SUPPORTS_POWER11)
message(STATUS "Compiler supports -mcpu=power11, enabling POWER11 backend")
ggml_add_cpu_backend_variant(power11 POWER11 VSX)
else()
message(STATUS "Skipping POWER11 backend: compiler does not support -mcpu=power11")
endif()
else()
message(FATAL_ERROR "Unsupported PowerPC target OS: ${CMAKE_SYSTEM_NAME}")
endif()
elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
if (CMAKE_SYSTEM_NAME MATCHES "Linux")
ggml_add_cpu_backend_variant(z15 Z15 VXE2)
ggml_add_cpu_backend_variant(z16 Z16 VXE2 NNPA)
else()
message(FATAL_ERROR "Unsupported s390x target OS: ${CMAKE_SYSTEM_NAME}")
endif()
elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
if (CMAKE_SYSTEM_NAME MATCHES "Linux")
ggml_add_cpu_backend_variant(riscv64_0)
ggml_add_cpu_backend_variant(riscv64_v RVV)
else()
message(FATAL_ERROR "Unsupported RISC-V target OS: ${CMAKE_SYSTEM_NAME}")
endif()
else()
message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}")
endif()
elseif (GGML_CPU)
ggml_add_cpu_backend_variant_impl("")
endif()
ggml_add_backend(BLAS)
ggml_add_backend(CANN)
ggml_add_backend(CUDA)
ggml_add_backend(HIP)
ggml_add_backend(METAL)
ggml_add_backend(MUSA)
ggml_add_backend(RPC)
ggml_add_backend(VirtGPU)
ggml_add_backend(SYCL)
ggml_add_backend(Vulkan)
ggml_add_backend(WebGPU)
ggml_add_backend(zDNN)
ggml_add_backend(OpenCL)
ggml_add_backend(Hexagon)
ggml_add_backend(ZenDNN)
ggml_add_backend(OPENVINO)
foreach (target ggml-base ggml)
target_include_directories(${target} PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)
target_compile_features (${target} PRIVATE c_std_11 cxx_std_17) # don't bump
endforeach()
target_link_libraries(ggml-base PRIVATE Threads::Threads)
if (DEFINED MATH_LIBRARY)
target_link_libraries(ggml-base PRIVATE ${MATH_LIBRARY})
elseif (NOT WIN32 AND NOT DEFINED ENV{ONEAPI_ROOT})
target_link_libraries(ggml-base PRIVATE m)
endif()
if (CMAKE_SYSTEM_NAME MATCHES "Android")
target_link_libraries(ggml-base PRIVATE dl)
endif()
if(CMAKE_SYSTEM_NAME MATCHES "visionOS")
target_compile_definitions(ggml-base PUBLIC _DARWIN_C_SOURCE)
endif()
if (BUILD_SHARED_LIBS)
foreach (target ggml-base ggml)
set_target_properties(${target} PROPERTIES POSITION_INDEPENDENT_CODE ON)
target_compile_definitions(${target} PRIVATE GGML_BUILD)
target_compile_definitions(${target} PUBLIC GGML_SHARED)
endforeach()
endif()
engine/ggml/src/ggml-alloc.c
+1248
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-backend-dl.cpp
+48
View File
@@ -0,0 +1,48 @@
#include "ggml-backend-dl.h"
#ifdef _WIN32
dl_handle * dl_load_library(const fs::path & path) {
// suppress error dialogs for missing DLLs
DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
HMODULE handle = LoadLibraryW(path.wstring().c_str());
SetErrorMode(old_mode);
return handle;
}
void * dl_get_sym(dl_handle * handle, const char * name) {
DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
void * p = (void *) GetProcAddress(handle, name);
SetErrorMode(old_mode);
return p;
}
const char * dl_error() {
return "";
}
#else
dl_handle * dl_load_library(const fs::path & path) {
dl_handle * handle = dlopen(path.string().c_str(), RTLD_NOW | RTLD_LOCAL);
return handle;
}
void * dl_get_sym(dl_handle * handle, const char * name) {
return dlsym(handle, name);
}
const char * dl_error() {
const char *rslt = dlerror();
return rslt != nullptr ? rslt : "";
}
#endif
engine/ggml/src/ggml-backend-dl.h
+45
View File
@@ -0,0 +1,45 @@
#pragma once
#ifdef _WIN32
# define WIN32_LEAN_AND_MEAN
# ifndef NOMINMAX
# define NOMINMAX
# endif
# include <windows.h>
# include <winevt.h>
#else
# include <dlfcn.h>
# include <unistd.h>
#endif
#include <filesystem>
namespace fs = std::filesystem;
#ifdef _WIN32
using dl_handle = std::remove_pointer_t<HMODULE>;
struct dl_handle_deleter {
void operator()(HMODULE handle) {
FreeLibrary(handle);
}
};
#else
using dl_handle = void;
struct dl_handle_deleter {
void operator()(void * handle) {
dlclose(handle);
}
};
#endif
using dl_handle_ptr = std::unique_ptr<dl_handle, dl_handle_deleter>;
dl_handle * dl_load_library(const fs::path & path);
void * dl_get_sym(dl_handle * handle, const char * name);
const char * dl_error();
engine/ggml/src/ggml-backend-impl.h
+275
View File
@@ -0,0 +1,275 @@
#pragma once
// ggml-backend internal header
#include "ggml-backend.h"
#ifdef __cplusplus
extern "C" {
#endif
#define GGML_BACKEND_API_VERSION 2
//
// Backend buffer type
//
struct ggml_backend_buffer_type_i {
const char * (*get_name) (ggml_backend_buffer_type_t buft);
// allocate a buffer of this type
ggml_backend_buffer_t (*alloc_buffer) (ggml_backend_buffer_type_t buft, size_t size);
// tensor alignment
size_t (*get_alignment) (ggml_backend_buffer_type_t buft);
// (optional) max buffer size that can be allocated (defaults to SIZE_MAX)
size_t (*get_max_size) (ggml_backend_buffer_type_t buft);
// (optional) data size needed to allocate the tensor, including padding (defaults to ggml_nbytes)
size_t (*get_alloc_size)(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
// (optional) check if tensor data is in host memory and uses standard ggml tensor layout (defaults to false)
bool (*is_host) (ggml_backend_buffer_type_t buft);
};
struct ggml_backend_buffer_type {
struct ggml_backend_buffer_type_i iface;
ggml_backend_dev_t device;
void * context;
};
//
// Backend buffer
//
struct ggml_backend_buffer_i {
// (optional) free the buffer
void (*free_buffer) (ggml_backend_buffer_t buffer);
// base address of the buffer
void * (*get_base) (ggml_backend_buffer_t buffer);
// (optional) initialize a tensor in the buffer (eg. add tensor extras)
enum ggml_status (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
// tensor data access
void (*memset_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size);
void (*set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
// (optional) 2d data copies
void (*set_tensor_2d)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
void (*get_tensor_2d)(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
// (optional) tensor copy: dst is in the buffer, src may be in any buffer, including buffers from a different backend (return false if not supported)
bool (*cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst);
// clear the entire buffer
void (*clear) (ggml_backend_buffer_t buffer, uint8_t value);
// (optional) reset any internal state due to tensor initialization, such as tensor extras
void (*reset) (ggml_backend_buffer_t buffer);
};
struct ggml_backend_buffer {
struct ggml_backend_buffer_i iface;
ggml_backend_buffer_type_t buft;
void * context;
size_t size;
enum ggml_backend_buffer_usage usage;
};
GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
ggml_backend_buffer_type_t buft,
struct ggml_backend_buffer_i iface,
void * context,
size_t size);
// do not use directly, use ggml_backend_tensor_copy instead
GGML_API bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
// multi-buffer
// buffer that contains a collection of buffers
GGML_API ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
GGML_API bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
GGML_API void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
//
// Backend (meta)
//
GGML_API bool ggml_backend_is_meta (ggml_backend_t backend);
GGML_API bool ggml_backend_buffer_is_meta(ggml_backend_buffer_t buf);
GGML_API bool ggml_backend_buft_is_meta (ggml_backend_buffer_type_t buft);
GGML_API size_t ggml_backend_meta_n_backends (ggml_backend_t meta_backend);
GGML_API ggml_backend_t ggml_backend_meta_simple_backend(ggml_backend_t meta_backend, size_t index);
// temporary workaround to statically allocate tensors from a context in a deduplicated way:
GGML_API struct ggml_backend_buffer * ggml_backend_meta_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
//
// Backend (stream)
//
struct ggml_backend_i {
const char * (*get_name)(ggml_backend_t backend);
void (*free)(ggml_backend_t backend);
// (optional) asynchronous tensor data access
void (*set_tensor_async) (ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void (*get_tensor_async) (ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
void (*set_tensor_2d_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
void (*get_tensor_2d_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
bool (*cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
// (optional) complete all pending operations (required if the backend supports async operations)
void (*synchronize)(ggml_backend_t backend);
// (optional) graph plans (not used currently)
// compute graph with a plan
ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
void (*graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
// update the plan with a new graph - this should be faster than creating a new plan when the graph has the same topology
void (*graph_plan_update) (ggml_backend_t backend, ggml_backend_graph_plan_t plan, const struct ggml_cgraph * cgraph);
// compute the graph with the plan
enum ggml_status (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
// compute graph (always async if supported by the backend)
enum ggml_status (*graph_compute) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
// (optional) event synchronization
// record an event on this stream
void (*event_record)(ggml_backend_t backend, ggml_backend_event_t event);
// wait for an event on on a different stream
void (*event_wait) (ggml_backend_t backend, ggml_backend_event_t event);
// (optional) sort/optimize the nodes in the graph
void (*graph_optimize) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
};
struct ggml_backend {
ggml_guid_t guid;
struct ggml_backend_i iface;
ggml_backend_dev_t device;
void * context;
};
struct ggml_backend_event {
struct ggml_backend_device * device;
void * context;
};
//
// Backend device
//
// Note: if additional properties are needed, we should add a struct with all of them
// the current functions to obtain the properties can remain, since they are more convenient for often used properties
struct ggml_backend_device_i {
// device name: short identifier for this device, such as "CPU" or "CUDA0"
const char * (*get_name)(ggml_backend_dev_t dev);
// device description: short informative description of the device, could be the model name
const char * (*get_description)(ggml_backend_dev_t dev);
// device memory in bytes: 0 bytes to indicate no memory to report
void (*get_memory)(ggml_backend_dev_t dev, size_t * free, size_t * total);
// device type
enum ggml_backend_dev_type (*get_type)(ggml_backend_dev_t dev);
// device properties
void (*get_props)(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props);
// backend (stream) initialization
ggml_backend_t (*init_backend)(ggml_backend_dev_t dev, const char * params);
// preferred buffer type
ggml_backend_buffer_type_t (*get_buffer_type)(ggml_backend_dev_t dev);
// (optional) host buffer type (in system memory, typically this is a pinned memory buffer for faster transfers between host and device)
ggml_backend_buffer_type_t (*get_host_buffer_type)(ggml_backend_dev_t dev);
// (optional) buffer from pointer: create a buffer from a host pointer (useful for memory mapped models and importing data from other libraries)
ggml_backend_buffer_t (*buffer_from_host_ptr)(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size);
// check if the backend can compute an operation
bool (*supports_op)(ggml_backend_dev_t dev, const struct ggml_tensor * op);
// check if the backend can use tensors allocated in a buffer type
bool (*supports_buft)(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft);
// (optional) check if the backend wants to run an operation, even if the weights are allocated in an incompatible buffer
// these should be expensive operations that may benefit from running on this backend instead of the CPU backend
bool (*offload_op)(ggml_backend_dev_t dev, const struct ggml_tensor * op);
// (optional) event synchronization
ggml_backend_event_t (*event_new) (ggml_backend_dev_t dev);
void (*event_free) (ggml_backend_dev_t dev, ggml_backend_event_t event);
void (*event_synchronize) (ggml_backend_dev_t dev, ggml_backend_event_t event);
};
struct ggml_backend_device {
struct ggml_backend_device_i iface;
ggml_backend_reg_t reg;
void * context;
};
//
// Backend (reg)
//
struct ggml_backend_reg_i {
const char * (*get_name)(ggml_backend_reg_t reg);
// enumerate available devices
size_t (*get_device_count)(ggml_backend_reg_t reg);
ggml_backend_dev_t (*get_device)(ggml_backend_reg_t reg, size_t index);
// (optional) get a pointer to a function in the backend
// backends can add custom functions that are not part of the standard ggml-backend interface
void * (*get_proc_address)(ggml_backend_reg_t reg, const char * name);
};
struct ggml_backend_reg {
int api_version; // initialize to GGML_BACKEND_API_VERSION
struct ggml_backend_reg_i iface;
void * context;
};
// Add backend dynamic loading support to the backend
// Initialize the backend
typedef ggml_backend_reg_t (*ggml_backend_init_t)(void);
// Optional: obtain a score for the backend based on the system configuration
// Higher scores are preferred, 0 means the backend is not supported in the current system
typedef int (*ggml_backend_score_t)(void);
#ifdef GGML_BACKEND_DL
# ifdef __cplusplus
# define GGML_BACKEND_DL_IMPL(reg_fn) \
extern "C" { \
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
} \
ggml_backend_reg_t ggml_backend_init(void) { \
return reg_fn(); \
}
# define GGML_BACKEND_DL_SCORE_IMPL(score_fn) \
extern "C" { \
GGML_BACKEND_API int ggml_backend_score(void); \
} \
int ggml_backend_score(void) { \
return score_fn(); \
}
# else
# define GGML_BACKEND_DL_IMPL(reg_fn) \
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
ggml_backend_reg_t ggml_backend_init(void) { \
return reg_fn(); \
}
# define GGML_BACKEND_DL_SCORE_IMPL(score_fn) \
GGML_BACKEND_API int ggml_backend_score(void); \
int ggml_backend_score(void) { \
return score_fn(); \
}
# endif
#else
# define GGML_BACKEND_DL_IMPL(reg_fn)
# define GGML_BACKEND_DL_SCORE_IMPL(score_fn)
#endif
#ifdef __cplusplus
}
#endif
engine/ggml/src/ggml-backend-meta.cpp
+2263
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-backend-reg.cpp
+586
View File
@@ -0,0 +1,586 @@
#include "ggml-backend-impl.h"
#include "ggml-backend.h"
#include "ggml-backend-dl.h"
#include "ggml-impl.h"
#include <algorithm>
#include <cstring>
#include <filesystem>
#include <memory>
#include <string>
#include <type_traits>
#include <vector>
#include <cctype>
#ifdef _WIN32
# define WIN32_LEAN_AND_MEAN
# ifndef NOMINMAX
# define NOMINMAX
# endif
# include <windows.h>
#elif defined(__APPLE__)
# include <mach-o/dyld.h>
# include <dlfcn.h>
#else
# include <dlfcn.h>
# include <unistd.h>
#endif
// Backend registry
#ifdef GGML_USE_CPU
#include "ggml-cpu.h"
#endif
#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif
#ifdef GGML_USE_METAL
#include "ggml-metal.h"
#endif
#ifdef GGML_USE_SYCL
#include "ggml-sycl.h"
#endif
#ifdef GGML_USE_VULKAN
#include "ggml-vulkan.h"
#endif
#ifdef GGML_USE_WEBGPU
#include "ggml-webgpu.h"
#endif
#ifdef GGML_USE_ZDNN
#include "ggml-zdnn.h"
#endif
#ifdef GGML_USE_OPENCL
#include "ggml-opencl.h"
#endif
#ifdef GGML_USE_HEXAGON
#include "ggml-hexagon.h"
#endif
#ifdef GGML_USE_BLAS
#include "ggml-blas.h"
#endif
#ifdef GGML_USE_RPC
#include "ggml-rpc.h"
#endif
#ifdef GGML_USE_VIRTGPU_FRONTEND
#include "ggml-virtgpu.h"
#endif
#ifdef GGML_USE_CANN
#include "ggml-cann.h"
#endif
#ifdef GGML_USE_ZENDNN
#include "ggml-zendnn.h"
#endif
#ifdef GGML_USE_OPENVINO
#include "ggml-openvino.h"
#endif
namespace fs = std::filesystem;
static std::string path_str(const fs::path & path) {
try {
#if defined(__cpp_lib_char8_t)
// C++20 and later: u8string() returns std::u8string
const std::u8string u8str = path.u8string();
return std::string(reinterpret_cast<const char *>(u8str.data()), u8str.size());
#else
// C++17: u8string() returns std::string
return path.u8string();
#endif
} catch (...) {
return std::string();
}
}
struct ggml_backend_reg_entry {
ggml_backend_reg_t reg;
dl_handle_ptr handle;
};
struct ggml_backend_registry {
std::vector<ggml_backend_reg_entry> backends;
std::vector<ggml_backend_dev_t> devices;
ggml_backend_registry() {
#ifdef GGML_USE_CUDA
register_backend(ggml_backend_cuda_reg());
#endif
#ifdef GGML_USE_METAL
register_backend(ggml_backend_metal_reg());
#endif
#ifdef GGML_USE_SYCL
register_backend(ggml_backend_sycl_reg());
#endif
#ifdef GGML_USE_VULKAN
// Add runtime disable check
if (getenv("GGML_DISABLE_VULKAN") == nullptr) {
register_backend(ggml_backend_vk_reg());
} else {
GGML_LOG_DEBUG("Vulkan backend disabled by GGML_DISABLE_VULKAN environment variable\n");
}
#endif
#ifdef GGML_USE_WEBGPU
register_backend(ggml_backend_webgpu_reg());
#endif
#ifdef GGML_USE_ZDNN
register_backend(ggml_backend_zdnn_reg());
#endif
#ifdef GGML_USE_VIRTGPU_FRONTEND
register_backend(ggml_backend_virtgpu_reg());
#endif
#ifdef GGML_USE_OPENCL
register_backend(ggml_backend_opencl_reg());
#endif
#ifdef GGML_USE_ZENDNN
register_backend(ggml_backend_zendnn_reg());
#endif
#ifdef GGML_USE_HEXAGON
register_backend(ggml_backend_hexagon_reg());
#endif
#ifdef GGML_USE_CANN
register_backend(ggml_backend_cann_reg());
#endif
#ifdef GGML_USE_BLAS
register_backend(ggml_backend_blas_reg());
#endif
#ifdef GGML_USE_RPC
register_backend(ggml_backend_rpc_reg());
#endif
#ifdef GGML_USE_OPENVINO
register_backend(ggml_backend_openvino_reg());
#endif
#ifdef GGML_USE_CPU
register_backend(ggml_backend_cpu_reg());
#endif
}
~ggml_backend_registry() {
// FIXME: backends cannot be safely unloaded without a function to destroy all the backend resources,
// since backend threads may still be running and accessing resources from the dynamic library
for (auto & entry : backends) {
if (entry.handle) {
entry.handle.release(); // NOLINT
}
}
}
void register_backend(ggml_backend_reg_t reg, dl_handle_ptr handle = nullptr) {
if (!reg) {
return;
}
for (auto & entry : backends) {
if (entry.reg == reg) {
return;
}
}
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
__func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
#endif
backends.push_back({ reg, std::move(handle) });
for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
register_device(ggml_backend_reg_dev_get(reg, i));
}
}
void register_device(ggml_backend_dev_t device) {
for (auto & dev : devices) {
if (dev == device) {
return;
}
}
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
#endif
devices.push_back(device);
}
ggml_backend_reg_t load_backend(const fs::path & path, bool silent) {
dl_handle_ptr handle { dl_load_library(path) };
if (!handle) {
if (!silent) {
GGML_LOG_ERROR("%s: failed to load %s: %s\n", __func__, path_str(path).c_str(), dl_error());
}
return nullptr;
}
auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
if (score_fn && score_fn() == 0) {
if (!silent) {
GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, path_str(path).c_str());
}
return nullptr;
}
auto backend_init_fn = (ggml_backend_init_t) dl_get_sym(handle.get(), "ggml_backend_init");
if (!backend_init_fn) {
if (!silent) {
GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, path_str(path).c_str());
}
return nullptr;
}
ggml_backend_reg_t reg = backend_init_fn();
if (!reg || reg->api_version != GGML_BACKEND_API_VERSION) {
if (!silent) {
if (!reg) {
GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n",
__func__, path_str(path).c_str());
} else {
GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n",
__func__, path_str(path).c_str(), reg->api_version, GGML_BACKEND_API_VERSION);
}
}
return nullptr;
}
GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path_str(path).c_str());
register_backend(reg, std::move(handle));
return reg;
}
void unload_backend(ggml_backend_reg_t reg, bool silent) {
auto it = std::find_if(backends.begin(), backends.end(),
[reg](const ggml_backend_reg_entry & entry) { return entry.reg == reg; });
if (it == backends.end()) {
if (!silent) {
GGML_LOG_ERROR("%s: backend not found\n", __func__);
}
return;
}
if (!silent) {
GGML_LOG_DEBUG("%s: unloading %s backend\n", __func__, ggml_backend_reg_name(reg));
}
// remove devices
devices.erase(
std::remove_if(devices.begin(), devices.end(),
[reg](ggml_backend_dev_t dev) { return ggml_backend_dev_backend_reg(dev) == reg; }),
devices.end());
// remove backend
backends.erase(it);
}
};
static ggml_backend_registry & get_reg() {
static ggml_backend_registry reg;
return reg;
}
// Internal API
void ggml_backend_register(ggml_backend_reg_t reg) {
get_reg().register_backend(reg);
}
void ggml_backend_device_register(ggml_backend_dev_t device) {
get_reg().register_device(device);
}
// Backend (reg) enumeration
static bool striequals(const char * a, const char * b) {
for (; *a && *b; a++, b++) {
if (std::tolower(*a) != std::tolower(*b)) {
return false;
}
}
return *a == *b;
}
size_t ggml_backend_reg_count() {
return get_reg().backends.size();
}
ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
GGML_ASSERT(index < ggml_backend_reg_count());
return get_reg().backends[index].reg;
}
ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
ggml_backend_reg_t reg = ggml_backend_reg_get(i);
if (striequals(ggml_backend_reg_name(reg), name)) {
return reg;
}
}
return nullptr;
}
// Device enumeration
size_t ggml_backend_dev_count() {
return get_reg().devices.size();
}
ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
GGML_ASSERT(index < ggml_backend_dev_count());
return get_reg().devices[index];
}
ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
if (striequals(ggml_backend_dev_name(dev), name)) {
return dev;
}
}
return nullptr;
}
ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
if (ggml_backend_dev_type(dev) == type) {
return dev;
}
}
return nullptr;
}
// Convenience functions
ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
if (!dev) {
return nullptr;
}
return ggml_backend_dev_init(dev, params);
}
ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
if (!dev) {
return nullptr;
}
return ggml_backend_dev_init(dev, params);
}
ggml_backend_t ggml_backend_init_best(void) {
ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
dev = dev ? dev : ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU);
dev = dev ? dev : ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
if (!dev) {
return nullptr;
}
return ggml_backend_dev_init(dev, nullptr);
}
// Dynamic loading
ggml_backend_reg_t ggml_backend_load(const char * path) {
return get_reg().load_backend(path, false);
}
void ggml_backend_unload(ggml_backend_reg_t reg) {
get_reg().unload_backend(reg, true);
}
static fs::path get_executable_path() {
#if defined(__APPLE__)
// get executable path
std::vector<char> path;
uint32_t size;
while (true) {
size = path.size();
if (_NSGetExecutablePath(path.data(), &size) == 0) {
break;
}
path.resize(size);
}
std::string base_path(path.data(), size);
// remove executable name
auto last_slash = base_path.find_last_of('/');
if (last_slash != std::string::npos) {
base_path = base_path.substr(0, last_slash);
}
return base_path + "/";
#elif defined(__linux__) || defined(__FreeBSD__)
std::string base_path = ".";
std::vector<char> path(1024);
while (true) {
// get executable path
# if defined(__linux__)
ssize_t len = readlink("/proc/self/exe", path.data(), path.size());
# elif defined(__FreeBSD__)
ssize_t len = readlink("/proc/curproc/file", path.data(), path.size());
# endif
if (len == -1) {
break;
}
if (len < (ssize_t) path.size()) {
base_path = std::string(path.data(), len);
// remove executable name
auto last_slash = base_path.find_last_of('/');
if (last_slash != std::string::npos) {
base_path = base_path.substr(0, last_slash);
}
break;
}
path.resize(path.size() * 2);
}
return base_path + "/";
#elif defined(_WIN32)
std::vector<wchar_t> path(MAX_PATH);
DWORD len = GetModuleFileNameW(NULL, path.data(), path.size());
if (len == 0) {
return {};
}
std::wstring base_path(path.data(), len);
// remove executable name
auto last_slash = base_path.find_last_of('\\');
if (last_slash != std::string::npos) {
base_path = base_path.substr(0, last_slash);
}
return base_path + L"\\";
#else
return {};
#endif
}
static fs::path backend_filename_prefix() {
#ifdef _WIN32
return fs::u8path("ggml-");
#else
return fs::u8path("libggml-");
#endif
}
static fs::path backend_filename_extension() {
#ifdef _WIN32
return fs::u8path(".dll");
#else
return fs::u8path(".so");
#endif
}
static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, const char * user_search_path) {
// enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths
const fs::path name_path = fs::u8path(name);
const fs::path file_prefix = backend_filename_prefix().native() + name_path.native() + fs::u8path("-").native();
const fs::path file_extension = backend_filename_extension();
std::vector<fs::path> search_paths;
if (user_search_path == nullptr) {
#ifdef GGML_BACKEND_DIR
search_paths.push_back(fs::u8path(GGML_BACKEND_DIR));
#endif
// default search paths: executable directory, current directory
search_paths.push_back(get_executable_path());
search_paths.push_back(fs::current_path());
} else {
search_paths.push_back(fs::u8path(user_search_path));
}
int best_score = 0;
fs::path best_path;
std::error_code ec;
for (const auto & search_path : search_paths) {
if (!fs::exists(search_path, ec)) {
if (ec) {
GGML_LOG_DEBUG("%s: posix_stat(%s) failure, error-message: %s\n", __func__, path_str(search_path).c_str(), ec.message().c_str());
} else {
GGML_LOG_DEBUG("%s: search path %s does not exist\n", __func__, path_str(search_path).c_str());
}
continue;
}
fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
for (const auto & entry : dir_it) {
if (entry.is_regular_file(ec)) {
auto filename = entry.path().filename();
auto ext = entry.path().extension();
if (filename.native().find(file_prefix) == 0 && ext == file_extension) {
dl_handle_ptr handle { dl_load_library(entry) };
if (!handle && !silent) {
GGML_LOG_ERROR("%s: failed to load %s: %s\n", __func__, path_str(entry.path()).c_str(), dl_error());
}
if (handle) {
auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
if (score_fn) {
int s = score_fn();
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, path_str(entry.path()).c_str(), s);
#endif
if (s > best_score) {
best_score = s;
best_path = entry.path();
}
} else {
if (!silent) {
GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, path_str(entry.path()).c_str());
}
}
}
}
}
}
}
if (best_score == 0) {
// try to load the base backend
for (const auto & search_path : search_paths) {
fs::path filename = backend_filename_prefix().native() + name_path.native() + backend_filename_extension().native();
fs::path path = search_path / filename;
if (std::error_code ec; fs::exists(path, ec)) {
return get_reg().load_backend(path, silent);
} else {
if (ec) {
GGML_LOG_DEBUG("%s: posix_stat(%s) failure, error-message: %s\n", __func__, path_str(path).c_str(), ec.message().c_str());
}
}
}
return nullptr;
}
return get_reg().load_backend(best_path, silent);
}
void ggml_backend_load_all() {
ggml_backend_load_all_from_path(nullptr);
}
void ggml_backend_load_all_from_path(const char * dir_path) {
#ifdef NDEBUG
bool silent = true;
#else
bool silent = false;
#endif
ggml_backend_load_best("blas", silent, dir_path);
ggml_backend_load_best("zendnn", silent, dir_path);
ggml_backend_load_best("cann", silent, dir_path);
ggml_backend_load_best("cuda", silent, dir_path);
ggml_backend_load_best("hip", silent, dir_path);
ggml_backend_load_best("metal", silent, dir_path);
ggml_backend_load_best("rpc", silent, dir_path);
ggml_backend_load_best("sycl", silent, dir_path);
ggml_backend_load_best("vulkan", silent, dir_path);
ggml_backend_load_best("virtgpu", silent, dir_path);
ggml_backend_load_best("opencl", silent, dir_path);
ggml_backend_load_best("hexagon", silent, dir_path);
ggml_backend_load_best("musa", silent, dir_path);
ggml_backend_load_best("openvino", silent, dir_path);
ggml_backend_load_best("cpu", silent, dir_path);
// check the environment variable GGML_BACKEND_PATH to load an out-of-tree backend
const char * backend_path = std::getenv("GGML_BACKEND_PATH");
if (backend_path) {
ggml_backend_load(backend_path);
}
}
engine/ggml/src/ggml-backend.cpp
+2371
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-blas/CMakeLists.txt
+101
View File
@@ -0,0 +1,101 @@
if (GGML_STATIC)
set(BLA_STATIC ON)
endif()
#if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
# set(BLA_SIZEOF_INTEGER 8)
#endif()
set(BLA_VENDOR ${GGML_BLAS_VENDOR})
find_package(BLAS)
if (BLAS_FOUND)
message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
ggml_add_backend_library(ggml-blas
ggml-blas.cpp
)
if (${GGML_BLAS_VENDOR} MATCHES "Apple")
add_compile_definitions(ACCELERATE_NEW_LAPACK)
add_compile_definitions(ACCELERATE_LAPACK_ILP64)
add_compile_definitions(GGML_BLAS_USE_ACCELERATE)
elseif ("${BLAS_INCLUDE_DIRS}" STREQUAL "")
# BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
# see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
find_package(PkgConfig REQUIRED)
if (${GGML_BLAS_VENDOR} MATCHES "Generic")
pkg_check_modules(DepBLAS blas)
elseif (${GGML_BLAS_VENDOR} MATCHES "OpenBLAS")
# As of openblas v0.3.22, the 64-bit is named openblas64.pc
pkg_check_modules(DepBLAS openblas64)
if (NOT DepBLAS_FOUND)
pkg_check_modules(DepBLAS openblas)
endif()
elseif (${GGML_BLAS_VENDOR} MATCHES "FLAME")
pkg_check_modules(DepBLAS blis)
elseif (${GGML_BLAS_VENDOR} MATCHES "ATLAS")
pkg_check_modules(DepBLAS blas-atlas)
elseif (${GGML_BLAS_VENDOR} MATCHES "FlexiBLAS")
pkg_check_modules(DepBLAS flexiblas_api)
elseif (${GGML_BLAS_VENDOR} MATCHES "Intel")
# all Intel* libraries share the same include path
pkg_check_modules(DepBLAS mkl-sdl)
elseif (${GGML_BLAS_VENDOR} MATCHES "NVHPC")
# this doesn't provide pkg-config
# suggest to assign BLAS_INCLUDE_DIRS on your own
if ("${NVHPC_VERSION}" STREQUAL "")
message(WARNING "Better to set NVHPC_VERSION")
else()
set(DepBLAS_FOUND ON)
set(DepBLAS_INCLUDE_DIRS "/opt/nvidia/hpc_sdk/${CMAKE_SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}/${NVHPC_VERSION}/math_libs/include")
endif()
endif()
if (DepBLAS_FOUND)
set(BLAS_INCLUDE_DIRS ${DepBLAS_INCLUDE_DIRS})
else()
message(WARNING "BLAS_INCLUDE_DIRS neither been provided nor been automatically"
" detected by pkgconfig, trying to find cblas.h from possible paths...")
find_path(BLAS_INCLUDE_DIRS
NAMES cblas.h
HINTS
/usr/include
/usr/local/include
/usr/include/openblas
/opt/homebrew/opt/openblas/include
/usr/local/opt/openblas/include
/usr/include/x86_64-linux-gnu/openblas/include
)
endif()
endif()
message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
target_compile_options(ggml-blas PRIVATE ${BLAS_LINKER_FLAGS})
if ("${GGML_BLAS_VENDOR}" STREQUAL "")
message(WARNING "GGML_BLAS_VENDOR is not set; some methods may not link properly.")
endif()
if ("${GGML_BLAS_VENDOR}" MATCHES "Intel" OR ("${BLAS_INCLUDE_DIRS}" MATCHES "mkl" AND "${GGML_BLAS_VENDOR}" MATCHES "Generic"))
add_compile_definitions(GGML_BLAS_USE_MKL)
endif()
if ("${GGML_BLAS_VENDOR}" MATCHES "OpenBLAS")
add_compile_definitions(GGML_BLAS_USE_OPENBLAS)
endif()
if ("${GGML_BLAS_VENDOR}" MATCHES "FLAME" OR "${GGML_BLAS_VENDOR}" MATCHES "AOCL" OR "${GGML_BLAS_VENDOR}" MATCHES "AOCL_mt")
add_compile_definitions(GGML_BLAS_USE_BLIS)
endif()
if ("${GGML_BLAS_VENDOR}" MATCHES "NVPL")
add_compile_definitions(GGML_BLAS_USE_NVPL)
endif()
target_link_libraries (ggml-blas PRIVATE ${BLAS_LIBRARIES})
target_include_directories(ggml-blas SYSTEM PRIVATE ${BLAS_INCLUDE_DIRS})
else()
message(FATAL_ERROR "BLAS not found, please refer to "
"https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
" to set correct GGML_BLAS_VENDOR")
endif()
engine/ggml/src/ggml-blas/ggml-blas.cpp
+522
View File
@@ -0,0 +1,522 @@
#include "ggml-impl.h"
#include "ggml-blas.h"
#include "ggml-backend-impl.h"
#include <future>
#include <vector>
#include <cstring>
#if defined(GGML_BLAS_USE_ACCELERATE)
# include <Accelerate/Accelerate.h>
#elif defined(GGML_BLAS_USE_MKL)
# include <mkl.h>
#elif defined(GGML_BLAS_USE_BLIS)
# include <blis.h>
#elif defined(GGML_BLAS_USE_NVPL)
# include <nvpl_blas.h>
#else
# include <cblas.h>
#endif
struct ggml_backend_blas_context {
int n_threads = GGML_DEFAULT_N_THREADS;
std::unique_ptr<char[]> work_data;
size_t work_size = 0;
#ifndef GGML_USE_OPENMP
std::vector<std::future<void>> tasks;
#endif
};
static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
const struct ggml_tensor * src0 = dst->src[0];
const struct ggml_tensor * src1 = dst->src[1];
GGML_TENSOR_BINARY_OP_LOCALS
const enum ggml_type type = src0->type;
GGML_ASSERT(ne0 == ne01);
GGML_ASSERT(ne1 == ne11);
GGML_ASSERT(ne2 == ne12);
GGML_ASSERT(ne3 == ne13);
// we don't support permuted src0 or src1
GGML_ASSERT(nb00 == ggml_type_size(type));
GGML_ASSERT(nb10 == ggml_type_size(src1->type));
// dst cannot be transposed or permuted
GGML_ASSERT(nb0 == sizeof(float));
GGML_ASSERT(nb0 <= nb1);
GGML_ASSERT(nb1 <= nb2);
GGML_ASSERT(nb2 <= nb3);
// broadcast factors
const int64_t r2 = ne12/ne02;
const int64_t r3 = ne13/ne03;
const int64_t ne_plane = ne01*ne00;
const size_t desired_wsize = type == GGML_TYPE_F32 ? 0 : ne03*ne02*ne_plane*sizeof(float);
if (ctx->work_size < desired_wsize) {
ctx->work_data.reset(new char[desired_wsize]);
ctx->work_size = desired_wsize;
}
void * wdata = ctx->work_data.get();
// convert src0 to float
if (type != GGML_TYPE_F32) {
const auto * type_traits = ggml_get_type_traits(type);
ggml_to_float_t const to_float = type_traits->to_float;
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
const void * x = (char *) src0->data + i02*nb02 + i03*nb03;
float * const wplane = (float *) wdata + i02*ne_plane + i03*ne02*ne_plane;
const int min_cols_per_thread = 4096;
const int min_rows_per_thread = std::max((int)(min_cols_per_thread/ne00), 1);
const int n_threads = std::max(std::min(ctx->n_threads, (int)(ne01/min_rows_per_thread)), 1);
#ifdef GGML_USE_OPENMP
#pragma omp parallel for num_threads(n_threads)
for (int64_t i01 = 0; i01 < ne01; i01++) {
to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
}
#else
for (int i = 1; i < n_threads; i++) {
const int64_t start = i*ne01/n_threads;
const int64_t end = (i + 1)*ne01/n_threads;
if (start < end) {
ctx->tasks.push_back(std::async(std::launch::async, [=]() {
for (int64_t i01 = start; i01 < end; i01++) {
to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
}
}));
}
}
{
// reuse the current thread for the first task
const int64_t start = 0;
const int64_t end = ne01/n_threads;
for (int64_t i01 = start; i01 < end; i01++) {
to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
}
}
#endif
}
}
#ifndef GGML_USE_OPENMP
// wait for all tasks to finish
for (auto & task : ctx->tasks) {
task.get();
}
ctx->tasks.clear();
#endif
}
#if defined(GGML_BLAS_USE_OPENBLAS)
openblas_set_num_threads(ctx->n_threads);
#elif defined(GGML_BLAS_USE_BLIS)
bli_thread_set_num_threads(ctx->n_threads);
#elif defined(GGML_BLAS_USE_NVPL)
nvpl_blas_set_num_threads(ctx->n_threads);
#elif defined(GGML_BLAS_USE_MKL)
mkl_set_num_threads(ctx->n_threads);
#endif
for (int64_t i13 = 0; i13 < ne13; i13++) {
for (int64_t i12 = 0; i12 < ne12; i12++) {
const int64_t i03 = i13/r3;
const int64_t i02 = i12/r2;
const float * x = (float *) ((char *) src0->data + i02*nb02 + i03*nb03);
const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
if (type != GGML_TYPE_F32) {
x = (float *) wdata + i02*ne_plane + i03*ne02*ne_plane;
}
cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
ne1, ne01, ne10,
1.0f, y, ne10,
x, ne00,
0.0f, d, ne01);
}
}
}
static void ggml_backend_blas_out_prod(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
const struct ggml_tensor * src0 = dst->src[0];
const struct ggml_tensor * src1 = dst->src[1];
GGML_TENSOR_BINARY_OP_LOCALS
GGML_ASSERT(ne0 == ne00);
GGML_ASSERT(ne1 == ne10);
GGML_ASSERT(ne2 == ne02);
GGML_ASSERT(ne02 == ne12);
GGML_ASSERT(ne3 == ne13);
GGML_ASSERT(ne03 == ne13);
// we don't support permuted src0 or src1
GGML_ASSERT(nb00 == sizeof(float));
// dst cannot be transposed or permuted
GGML_ASSERT(nb0 == sizeof(float));
// GGML_ASSERT(nb0 <= nb1);
// GGML_ASSERT(nb1 <= nb2);
// GGML_ASSERT(nb2 <= nb3);
// Arguments to ggml_compute_forward_out_prod (expressed as major,minor)
// src0: (k,n)
// src1: (k,m)
// dst: (m,n)
//
// Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f)
// Also expressed as (major,minor)
// a: (m,k): so src1 transposed
// b: (k,n): so src0
// c: (m,n)
//
// However, if ggml_is_transposed(src1) is true, then
// src1->data already contains a transposed version, so sgemm mustn't
// transpose it further.
int n = src0->ne[0];
int k = src0->ne[1];
int m = src1->ne[0];
CBLAS_TRANSPOSE transposeA;
int lda;
if (!ggml_is_transposed(src1)) {
transposeA = CblasTrans;
lda = m;
} else {
transposeA = CblasNoTrans;
lda = k;
}
float * a = (float *) ((char *) src1->data);
float * b = (float *) ((char *) src0->data);
float * c = (float *) ((char *) dst->data);
cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n);
GGML_UNUSED(ctx);
}
// backend interface
static const char * ggml_backend_blas_get_name(ggml_backend_t backend) {
return "BLAS";
GGML_UNUSED(backend);
}
static void ggml_backend_blas_free(ggml_backend_t backend) {
ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
delete ctx;
delete backend;
}
static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
for (int i = 0; i < cgraph->n_nodes; i++) {
struct ggml_tensor * node = cgraph->nodes[i];
if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
continue;
}
switch (node->op) {
case GGML_OP_MUL_MAT:
ggml_backend_blas_mul_mat(ctx, node);
break;
case GGML_OP_OUT_PROD:
ggml_backend_blas_out_prod(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:
GGML_ABORT("%s: unsupported op %s\n", __func__, ggml_op_desc(node));
}
}
return GGML_STATUS_SUCCESS;
GGML_UNUSED(backend);
}
static struct ggml_backend_i blas_backend_i = {
/* .get_name = */ ggml_backend_blas_get_name,
/* .free = */ ggml_backend_blas_free,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .set_tensor_2d_async = */ NULL,
/* .get_tensor_2d_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_blas_graph_compute,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
/* .graph_optimize = */ NULL,
};
static ggml_guid_t ggml_backend_blas_guid(void) {
static ggml_guid guid = { 0x12, 0xa8, 0xae, 0xf4, 0xc0, 0x1e, 0x61, 0x97, 0x8f, 0xeb, 0x33, 0x04, 0xa1, 0x33, 0x51, 0x2d };
return &guid;
}
ggml_backend_t ggml_backend_blas_init(void) {
ggml_backend_blas_context * ctx = new ggml_backend_blas_context;
ggml_backend_t backend = new ggml_backend {
/* .guid = */ ggml_backend_blas_guid(),
/* .iface = */ blas_backend_i,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_blas_reg(), 0),
/* .context = */ ctx,
};
#if defined(GGML_BLAS_USE_OPENBLAS) && defined(GGML_USE_OPENMP)
if (openblas_get_parallel() != OPENBLAS_OPENMP) {
GGML_LOG_DEBUG("%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
}
#endif
#if defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
GGML_LOG_DEBUG("%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
#endif
return backend;
}
bool ggml_backend_is_blas(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_blas_guid());
}
void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads) {
GGML_ASSERT(ggml_backend_is_blas(backend_blas));
ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend_blas->context;
ctx->n_threads = n_threads;
}
// device interface
static const char * ggml_backend_blas_device_get_name(ggml_backend_dev_t dev) {
return "BLAS";
GGML_UNUSED(dev);
}
static const char * ggml_backend_blas_device_get_description(ggml_backend_dev_t dev) {
#if defined(GGML_BLAS_USE_ACCELERATE)
return "Accelerate";
#elif defined(GGML_BLAS_USE_MKL)
return "MKL";
#elif defined(GGML_BLAS_USE_BLIS)
return "BLIS";
#elif defined(GGML_BLAS_USE_NVPL)
return "NVPL";
#elif defined(GGML_BLAS_USE_OPENBLAS)
return "OpenBLAS";
#else
return "BLAS";
#endif
GGML_UNUSED(dev);
}
static void ggml_backend_blas_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
// no memory to report
*free = 0;
*total = 0;
GGML_UNUSED(dev);
}
static enum ggml_backend_dev_type ggml_backend_blas_device_get_type(ggml_backend_dev_t dev) {
return GGML_BACKEND_DEVICE_TYPE_ACCEL;
GGML_UNUSED(dev);
}
static void ggml_backend_blas_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
props->name = ggml_backend_blas_device_get_name(dev);
props->description = ggml_backend_blas_device_get_description(dev);
props->type = ggml_backend_blas_device_get_type(dev);
ggml_backend_blas_device_get_memory(dev, &props->memory_free, &props->memory_total);
props->caps = {
/* .async = */ false,
/* .host_buffer = */ false,
/* .buffer_from_host_ptr = */ true,
/* .events = */ false,
};
}
static ggml_backend_t ggml_backend_blas_device_init_backend(ggml_backend_dev_t dev, const char * params) {
return ggml_backend_blas_init();
GGML_UNUSED(dev);
GGML_UNUSED(params);
}
static ggml_backend_buffer_type_t ggml_backend_blas_device_get_buffer_type(ggml_backend_dev_t dev) {
return ggml_backend_cpu_buffer_type();
GGML_UNUSED(dev);
}
static ggml_backend_buffer_t ggml_backend_blas_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
return ggml_backend_cpu_buffer_from_ptr(ptr, size);
GGML_UNUSED(dev);
GGML_UNUSED(max_tensor_size);
}
static bool ggml_backend_blas_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * src1 = op->src[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:
{
// BLAS usually is only faster for large matrices
const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * src1 = op->src[1];
const int64_t ne10 = src1->ne[0];
const int64_t ne0 = op->ne[0];
const int64_t ne1 = op->ne[1];
// TODO: find the optimal value
const int64_t min_batch = 32;
return ggml_is_contiguous(src0) &&
ggml_is_contiguous(src1) &&
src1->type == GGML_TYPE_F32 &&
(ne0 >= min_batch && ne1 >= min_batch && ne10 >= min_batch) &&
(src0->type == GGML_TYPE_F32 || ggml_get_type_traits(src0->type)->to_float != NULL);
}
case GGML_OP_OUT_PROD:
return op->src[0]->type == GGML_TYPE_F32 &&
op->src[1]->type == GGML_TYPE_F32 &&
ggml_is_matrix(src0) &&
ggml_is_matrix(src1) &&
ggml_is_contiguous(src0) &&
(ggml_is_contiguous(src1) || ggml_is_transposed(src1)) &&
(src0->type == GGML_TYPE_F32 || ggml_get_type_traits(src0->type)->to_float != NULL);
default:
return false;
}
GGML_UNUSED(dev);
}
static bool ggml_backend_blas_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
return ggml_backend_buft_is_host(buft);
GGML_UNUSED(dev);
}
static const struct ggml_backend_device_i ggml_backend_blas_device_i = {
/* .get_name = */ ggml_backend_blas_device_get_name,
/* .get_description = */ ggml_backend_blas_device_get_description,
/* .get_memory = */ ggml_backend_blas_device_get_memory,
/* .get_type = */ ggml_backend_blas_device_get_type,
/* .get_props = */ ggml_backend_blas_device_get_props,
/* .init_backend = */ ggml_backend_blas_device_init_backend,
/* .get_buffer_type = */ ggml_backend_blas_device_get_buffer_type,
/* .get_host_buffer_type = */ NULL,
/* .buffer_from_host_ptr = */ ggml_backend_blas_device_buffer_from_host_ptr,
/* .supports_op = */ ggml_backend_blas_device_supports_op,
/* .supports_buft = */ ggml_backend_blas_device_supports_buft,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_synchronize = */ NULL,
};
// backend reg interface
static const char * ggml_backend_blas_reg_get_name(ggml_backend_reg_t reg) {
return "BLAS";
GGML_UNUSED(reg);
}
static size_t ggml_backend_blas_reg_get_device_count(ggml_backend_reg_t reg) {
return 1;
GGML_UNUSED(reg);
}
static ggml_backend_dev_t ggml_backend_blas_reg_get_device(ggml_backend_reg_t reg, size_t index) {
GGML_ASSERT(index == 0);
static ggml_backend_device ggml_backend_blas_device = {
/* .iface = */ ggml_backend_blas_device_i,
/* .reg = */ reg,
/* .context = */ nullptr,
};
return &ggml_backend_blas_device;
GGML_UNUSED(reg);
GGML_UNUSED(index);
}
static void * ggml_backend_blas_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_blas_set_n_threads;
}
return NULL;
GGML_UNUSED(reg);
GGML_UNUSED(name);
}
static const struct ggml_backend_reg_i ggml_backend_blas_reg_i = {
/* .get_name = */ ggml_backend_blas_reg_get_name,
/* .get_device_count = */ ggml_backend_blas_reg_get_device_count,
/* .get_device = */ ggml_backend_blas_reg_get_device,
/* .get_proc_address = */ ggml_backend_blas_get_proc_address,
};
ggml_backend_reg_t ggml_backend_blas_reg(void) {
static struct ggml_backend_reg ggml_backend_blas_reg = {
/* .api_version = */ GGML_BACKEND_API_VERSION,
/* .iface = */ ggml_backend_blas_reg_i,
/* .context = */ NULL,
};
return &ggml_backend_blas_reg;
}
GGML_BACKEND_DL_IMPL(ggml_backend_blas_reg)
engine/ggml/src/ggml-cann/CMakeLists.txt
Executable
+89
View File
@@ -0,0 +1,89 @@
if ("cann${CANN_INSTALL_DIR}" STREQUAL "cann" AND DEFINED ENV{ASCEND_TOOLKIT_HOME})
set(CANN_INSTALL_DIR $ENV{ASCEND_TOOLKIT_HOME})
message(STATUS "CANN: updated CANN_INSTALL_DIR from ASCEND_TOOLKIT_HOME=$ENV{ASCEND_TOOLKIT_HOME}")
endif()
# Auto-detech Soc type and Soc version, if detect failed, will abort build
set(SOC_VERSION "")
function(detect_ascend_soc_type SOC_VERSION)
execute_process(
COMMAND bash -c "npu-smi info|awk -F' ' 'NF > 0 && NR==7 {print $3}'"
OUTPUT_VARIABLE npu_info
RESULT_VARIABLE npu_result
OUTPUT_STRIP_TRAILING_WHITESPACE
)
if("${npu_info}" STREQUAL "" OR ${npu_result})
message(FATAL_ERROR "Auto-detech ascend soc type failed, please specify manually or check ascend device working normally.")
endif()
set(${SOC_VERSION} "Ascend${npu_info}" PARENT_SCOPE)
endfunction()
if(NOT SOC_TYPE)
detect_ascend_soc_type(SOC_VERSION)
set(SOC_TYPE "${SOC_VERSION}")
message(STATUS "CANN: SOC_VERSION auto-detected is:${SOC_VERSION}")
endif()
string(TOLOWER ${SOC_TYPE} SOC_VERSION) # SOC_VERSION need lower
# Construct Soc specify compile option: ASCEND_#Soc_Major_SN. Such as ASCEND_910B, ASCEND_310P.
string(REGEX MATCH "[0-9]+[a-zA-Z]" SOC_TYPE_MAJOR_SN "${SOC_VERSION}")
set(SOC_TYPE_COMPILE_OPTION "ASCEND_${SOC_TYPE_MAJOR_SN}")
string(TOUPPER ${SOC_TYPE_COMPILE_OPTION} SOC_TYPE_COMPILE_OPTION)
message(STATUS "CANN: SOC_VERSION = ${SOC_VERSION}")
option(USE_ACL_GRAPH "Enable CANN graph execution (ACL graph mode)" OFF)
if(USE_ACL_GRAPH AND (SOC_TYPE_MAJOR_SN STREQUAL "310P" OR SOC_TYPE_COMPILE_OPTION STREQUAL "ASCEND_310P"))
message(FATAL_ERROR
"CANN Graph (ACL graph mode) is not supported on 310P devices. "
"Please build with -DUSE_ACL_GRAPH=OFF or use a supported SOC.")
endif()
if (CANN_INSTALL_DIR)
# Only Support Linux.
if (NOT UNIX)
message(FATAL_ERROR "CANN: CANN toolkit supports unix but not ${CMAKE_SYSTEM_NAME}")
endif()
# Supported platforms: x86-64, arm64
if (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "amd64")
else()
message(FATAL_ERROR "CANN: CANN toolkit supports x86-64 and arm64 but not ${CMAKE_SYSTEM_PROCESSOR}")
endif()
# Set header and libs
set(CANN_INCLUDE_DIRS
${CANN_INSTALL_DIR}/include
${CANN_INSTALL_DIR}/include/aclnn
${CANN_INSTALL_DIR}/acllib/include
)
list(APPEND CANN_LIBRARIES
ascendcl
nnopbase
opapi
acl_op_compiler
)
file(GLOB GGML_SOURCES_CANN "*.cpp")
ggml_add_backend_library(ggml-cann ${GGML_SOURCES_CANN})
target_link_libraries(ggml-cann PRIVATE ${CANN_LIBRARIES})
target_include_directories(ggml-cann PRIVATE ${CANN_INCLUDE_DIRS})
target_link_directories(ggml-cann PRIVATE ${CANN_INSTALL_DIR}/lib64)
target_compile_definitions(ggml-cann PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
if (USE_ACL_GRAPH)
target_compile_definitions(ggml-cann PRIVATE USE_ACL_GRAPH)
message(STATUS "CANN: USE_ACL_GRAPH is enabled.")
else()
message(STATUS "CANN: USE_ACL_GRAPH is disabled.")
endif()
message(STATUS "CANN: CANN_INCLUDE_DIRS = ${CANN_INCLUDE_DIRS}")
message(STATUS "CANN: CANN_LIBRARIES = ${CANN_LIBRARIES}")
else()
message(FATAL_ERROR "CANN: Can't find CANN_INSTALL_DIR, did you forget to source set_var.sh?")
endif()
engine/ggml/src/ggml-cann/acl_tensor.cpp
+195
View File
@@ -0,0 +1,195 @@
/*
* Copyright (c) 2023-2026 The ggml authors
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
#include "acl_tensor.h"
#include <algorithm>
#include <cstring>
aclDataType ggml_cann_type_mapping(ggml_type type) {
switch (type) {
case GGML_TYPE_F32:
return ACL_FLOAT;
case GGML_TYPE_F16:
return ACL_FLOAT16;
case GGML_TYPE_BF16:
return ACL_BF16;
case GGML_TYPE_I8:
return ACL_INT8;
case GGML_TYPE_I16:
return ACL_INT16;
case GGML_TYPE_I32:
return ACL_INT32;
case GGML_TYPE_Q4_0:
return ACL_INT4;
case GGML_TYPE_Q8_0:
return ACL_INT8;
case GGML_TYPE_I64:
return ACL_INT64;
default:
return ACL_DT_UNDEFINED;
}
}
acl_tensor_ptr ggml_cann_create_tensor(const ggml_tensor * tensor,
int64_t * ne,
size_t * nb,
int64_t dims,
aclFormat format,
size_t offset) {
// If tensor is bcasted, Up to GGML_MAX_DIMS additional dimensions will be
// added.
int64_t acl_ne[GGML_MAX_DIMS * 2], acl_stride[GGML_MAX_DIMS * 2];
if (ne == nullptr) {
for (int i = 0; i < GGML_MAX_DIMS; i++) {
acl_ne[i] = tensor->ne[i];
// The step size of acl is in elements.
acl_stride[i] = tensor->nb[i] / ggml_element_size(tensor);
}
} else {
// With bcast
for (int i = 0; i < dims; i++) {
acl_ne[i] = ne[i];
acl_stride[i] = nb[i] / ggml_element_size(tensor);
}
}
int64_t final_dims = (dims == 0 ? GGML_MAX_DIMS : dims);
int64_t acl_storage_len = 1;
for (int i = 0; i < final_dims; i++) {
acl_storage_len += (acl_ne[i] - 1) * acl_stride[i];
}
size_t elem_offset = offset / ggml_element_size(tensor);
acl_storage_len += elem_offset;
// Reverse ne and stride.
std::reverse(acl_ne, acl_ne + final_dims);
std::reverse(acl_stride, acl_stride + final_dims);
aclTensor * raw = aclCreateTensor(acl_ne, final_dims, ggml_cann_type_mapping(tensor->type), acl_stride, elem_offset,
format, &acl_storage_len, 1, tensor->data);
return acl_tensor_ptr(raw);
}
acl_int_array_ptr ggml_cann_create_int_array(const int64_t * value, uint64_t size) {
aclIntArray * raw = aclCreateIntArray(value, size);
return acl_int_array_ptr(raw);
}
acl_scalar_ptr ggml_cann_create_scalar(void * value, aclDataType dataType) {
aclScalar * raw = aclCreateScalar(value, dataType);
return acl_scalar_ptr(raw);
}
bool ggml_cann_need_bcast(const ggml_tensor * t0, const ggml_tensor * t1) {
for (int i = 0; i < GGML_MAX_DIMS; i++) {
if (t1->ne[i] != t0->ne[i] && t1->ne[i] != 1) {
return true;
}
}
return false;
}
int64_t ggml_cann_get_bcast_shape(const ggml_tensor * src0,
const ggml_tensor * src1,
int64_t * bcast_src0_ne,
int64_t * bcast_src1_ne,
size_t * bcast_src0_nb,
size_t * bcast_src1_nb) {
GGML_ASSERT(ggml_can_repeat(src1, src0));
int bcast_dim_cnt = 0;
for (int i = 0; i < GGML_MAX_DIMS; i++) {
int64_t nr = src0->ne[i] / src1->ne[i];
bcast_src0_ne[bcast_dim_cnt] = src0->ne[i] / nr;
bcast_src1_ne[bcast_dim_cnt] = src1->ne[i];
bcast_src0_nb[bcast_dim_cnt] = src0->nb[i];
bcast_src1_nb[bcast_dim_cnt] = src1->nb[i];
bcast_dim_cnt++;
if (nr != 1) {
// Need to add an extra dim.
bcast_src0_ne[bcast_dim_cnt] = nr;
bcast_src1_ne[bcast_dim_cnt] = 1;
bcast_src0_nb[bcast_dim_cnt] = bcast_src0_nb[bcast_dim_cnt - 1] * bcast_src0_ne[bcast_dim_cnt - 1];
bcast_src1_nb[bcast_dim_cnt] = bcast_src1_nb[bcast_dim_cnt - 1] * bcast_src1_ne[bcast_dim_cnt - 1];
bcast_dim_cnt++;
}
}
return bcast_dim_cnt;
}
int64_t ggml_cann_get_mulmat_bcast_shape(const int64_t * input_ne,
const int64_t * weight_ne,
const int64_t * dst_ne,
const size_t * input_nb,
const size_t * weight_nb,
const size_t * dst_nb,
int64_t * bcast_input_ne,
int64_t * bcast_weight_ne,
int64_t * bcast_dst_ne,
size_t * bcast_input_nb,
size_t * bcast_weight_nb,
size_t * bcast_dst_nb) {
// input and dst shoule in same shape, except first two dims.
GGML_ASSERT(input_ne[2] == dst_ne[2]);
GGML_ASSERT(input_ne[3] == dst_ne[3]);
int bcast_dim_cnt = 0;
// For mul_mat, a dimension needs to be added before the dimension that
// weight needs to be expanded to satisfy the bcast rule of matrix
// multiplication.
for (int i = 0; i < GGML_MAX_DIMS; i++) {
int64_t nr = input_ne[i] / weight_ne[i];
// Do not use bcast in the first two dimensions because we only support
// the bcast batch dimension. Just copy them.
if (i < 2 || nr == 1) {
bcast_input_ne[bcast_dim_cnt] = input_ne[i];
bcast_weight_ne[bcast_dim_cnt] = weight_ne[i];
bcast_dst_ne[bcast_dim_cnt] = dst_ne[i];
bcast_input_nb[bcast_dim_cnt] = input_nb[i];
bcast_weight_nb[bcast_dim_cnt] = weight_nb[i];
bcast_dst_nb[bcast_dim_cnt] = dst_nb[i];
bcast_dim_cnt++;
} else {
// Need to add an extra dim.
bcast_input_ne[bcast_dim_cnt] = nr;
bcast_dst_ne[bcast_dim_cnt] = nr;
bcast_weight_ne[bcast_dim_cnt] = 1;
bcast_input_nb[bcast_dim_cnt] = input_nb[i];
bcast_dst_nb[bcast_dim_cnt] = dst_nb[i];
bcast_weight_nb[bcast_dim_cnt] = weight_nb[i];
bcast_dim_cnt++;
bcast_input_ne[bcast_dim_cnt] = input_ne[i] / nr;
bcast_dst_ne[bcast_dim_cnt] = dst_ne[i] / nr;
bcast_weight_ne[bcast_dim_cnt] = weight_ne[i];
bcast_input_nb[bcast_dim_cnt] = bcast_input_nb[bcast_dim_cnt - 1] * bcast_input_ne[bcast_dim_cnt - 1];
bcast_dst_nb[bcast_dim_cnt] = bcast_dst_nb[bcast_dim_cnt - 1] * bcast_dst_ne[bcast_dim_cnt - 1];
bcast_weight_nb[bcast_dim_cnt] = bcast_weight_nb[bcast_dim_cnt - 1] * bcast_weight_ne[bcast_dim_cnt - 1];
bcast_dim_cnt++;
}
}
return bcast_dim_cnt;
}
engine/ggml/src/ggml-cann/acl_tensor.h
+349
View File
@@ -0,0 +1,349 @@
/*
* Copyright (c) 2023-2026 The ggml authors
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
#ifndef CANN_ACL_TENSOR_H
#define CANN_ACL_TENSOR_H
#include "common.h"
#include <aclnn/aclnn_base.h>
#include <algorithm>
#include <cstring>
/**
* @brief Maps a ggml_type to its corresponding aclDataType.
*
* @details This function takes a ggml_type as input and returns the corresponding
* aclDataType. It supports mapping for various ggml_types. If the input type
* does not match any of the predefined ggml_types, the function returns
* ACL_DT_UNDEFINED.
*
* @param type The ggml_type to be mapped.
* @return The corresponding aclDataType. If the input type is not recognized,
* ACL_DT_UNDEFINED is returned.
*/
aclDataType ggml_cann_type_mapping(ggml_type type);
// Deleter for acl objects.
template <typename T, aclError (*DestroyFunc)(const T *)> struct acl_deleter {
void operator()(T * ptr) const noexcept {
if (ptr) {
ACL_CHECK(DestroyFunc(ptr));
}
}
};
using acl_tensor_ptr = std::unique_ptr<aclTensor, acl_deleter<aclTensor, aclDestroyTensor>>;
using acl_int_array_ptr = std::unique_ptr<aclIntArray, acl_deleter<aclIntArray, aclDestroyIntArray>>;
using acl_scalar_ptr = std::unique_ptr<aclScalar, acl_deleter<aclScalar, aclDestroyScalar>>;
using acl_tensor_list_ptr = std::unique_ptr<aclTensorList, acl_deleter<aclTensorList, aclDestroyTensorList>>;
/**
* @brief Creates an ACL tensor from a ggml_tensor with optional shape.
*
* @details This function creates an ACL tensor based on the properties of the
* provided ggml_tensor. It supports customer shape by adjusting dimensions
* and strides accordingly. If customer shape is applied, additional
* dimensions and strides are calculated based on the provided parameters.
*
* @param tensor Pointer to the ggml_tensor to be converted to ACL tensor.
* @param ne Pointer to an array containing dimensions. Defaults to nullptr
* if no customer shape is applied.
* @param nb Pointer to an array containing strides. Defaults to nullptr
* if no customer shape is applied.
* @param dims Number of dimensions in the tensor. Defaults to 0 if no customer
* shape is applied.
* @param format ACL tensor format. Defaults to ACL_FORMAT_ND.
* @param offset Offset in bytes for the ACL tensor data. Defaults to 0.
* @return Pointer to the created ACL tensor.
*/
acl_tensor_ptr ggml_cann_create_tensor(const ggml_tensor * tensor,
int64_t * ne = nullptr,
size_t * nb = nullptr,
int64_t dims = 0,
aclFormat format = ACL_FORMAT_ND,
size_t offset = 0);
/**
* @brief Template for creating an ACL tensor from provided parameters. typename TYPE
* should be size_t or float.
*
* @details This function creates an ACL tensor using the provided data pointer,
* data type, dimensions, strides, format, offset, and additional parameters.
* It calculates necessary dimensions and strides based on the provided ne and nb
* arrays, adjusting them for the ACL tensor creation. The ACL storage length
* is also calculated based on the provided dimensions and strides.
*
* @param data_ptr Pointer to the data buffer for the ACL tensor.
* @param dtype ACL data type of the tensor.
* @param type_size Size of each element in the tensor data buffer.
* @param ne Pointer to an array containing tensor dimensions.
* @param nb Pointer to an array containing tensor strides.
* @param dims Number of dimensions of the tensor.
* @param format ACL tensor format. Defaults to ACL_FORMAT_ND.
* @param offset Offset in bytes for the ACL tensor data. Defaults to 0.
* @return Pointer to the created ACL tensor.
*/
template <typename TYPE>
acl_tensor_ptr ggml_cann_create_tensor(void * data_ptr,
aclDataType dtype,
TYPE type_size,
int64_t * ne,
TYPE * nb,
int64_t dims,
aclFormat format = ACL_FORMAT_ND,
size_t offset = 0) {
int64_t tmp_ne[GGML_MAX_DIMS * 2];
int64_t tmp_stride[GGML_MAX_DIMS * 2];
memcpy(tmp_ne, ne, dims * sizeof(int64_t));
for (int i = 0; i < dims; i++) {
tmp_stride[i] = nb[i] / type_size;
}
int64_t acl_storage_len = 1;
for (int i = 0; i < dims; i++) {
acl_storage_len += (tmp_ne[i] - 1) * tmp_stride[i];
}
std::reverse(tmp_ne, tmp_ne + dims);
std::reverse(tmp_stride, tmp_stride + dims);
aclTensor * raw =
aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size, format, &acl_storage_len, 1, data_ptr);
return acl_tensor_ptr(raw);
}
/**
* @brief Create an ACL int array resource wrapped in a smart pointer.
*
* This function constructs an aclIntArray from the provided int64_t values
* and returns it as an acl_int_array_ptr (a std::unique_ptr with a custom
* deleter). The returned pointer owns the ACL resource and will automatically
* destroy it via aclDestroyIntArray().
*
* @param value Pointer to the int64_t elements.
* @param size Number of elements in value.
*
* @return A smart pointer managing the created ACL int array.
*/
acl_int_array_ptr ggml_cann_create_int_array(const int64_t * value, uint64_t size);
/**
* @brief Create an ACL scalar resource wrapped in a smart pointer.
*
* This function constructs an aclScalar from the raw value pointer and ACL
* data type, then returns it as an acl_scalar_ptr (a std::unique_ptr with
* a custom deleter). The returned pointer owns the ACL scalar and will
* automatically destroy it via aclDestroyScalar().
*
* @param value Pointer to the raw scalar memory.
* @param dataType ACL data type of the scalar.
*
* @return A smart pointer managing the created ACL scalar.
*/
acl_scalar_ptr ggml_cann_create_scalar(void * value, aclDataType dataType);
/**
* @brief Create an ACL tensor list from multiple tensor smart pointers.
*
* This function accepts a variadic list of acl_tensor_ptr (a unique_ptr with
* custom deleter) and produces an aclTensorList using aclCreateTensorList().
*
* The lifecycle management of the tensor objects changes as follows:
* - aclCreateTensorList() takes ownership of the tensors
* - Each input smart pointer releases ownership using release()
* - As a result, the tensors will NOT be destroyed by unique_ptr
* - Instead, they will be destroyed when aclDestroyTensorList() is called
*
* This ensures correct ownership transfer and prevents double-free situations.
*
* @param acl_tensor_ptr Variadic template parameter; each argument must be
* a unique_ptr-like type supporting get() and release().
*
* @param tensors Variadic list of acl_tensor_ptr objects. Ownership of
* each tensor is transferred away from these smart pointers.
*
* @return A smart pointer (acl_tensor_list_ptr) owning the created ACL tensor list.
*
* @note This implementation is C++11 compatible. The ownership-release process is
* executed using a pack expansion inside an initializer list.
*/
template <typename... acl_tensor_ptr> acl_tensor_list_ptr ggml_cann_create_tensor_list(acl_tensor_ptr &&... tensors) {
aclTensor * raw_tensors[] = { tensors.get()... };
aclTensorList * raw = aclCreateTensorList(raw_tensors, sizeof...(tensors));
// aclTensor will release by aclTensorList, so release ownership without
// destroying the tensor
int dummy[] = { (tensors.release(), 0)... };
GGML_UNUSED(dummy);
return acl_tensor_list_ptr(raw);
}
/**
* @brief Checks if tensors require broadcasting based on their shapes.
*
* @details This function determines if two ggml_tensors need to be broadcasted for
* element-wise operations. Broadcasting is necessary if the shapes of the
* tensors are not identical and no dimension in either tensor equals 1.
*
* @param t0 Pointer to the first ggml_tensor.
* @param t1 Pointer to the second ggml_tensor.
* @return True if broadcasting is needed, False otherwise.
*
* @remarks This function iterates over the dimensions of t0 and t1. It checks if each
* dimension in t1 differs from t0's corresponding dimension and is not equal
* to 1. If such a dimension is found, broadcasting is required to align t1
* with t0 for element-wise operations.
*/
bool ggml_cann_need_bcast(const ggml_tensor * t0, const ggml_tensor * t1);
/**
* @brief Computes broadcast shapes and strides for two ggml_tensors.
*
* @details This function calculates the broadcast shapes and strides for two ggml_tensors,
* following the broadcasting rules similar to numpy. It adjusts dimensions and
* strides to ensure compatibility for element-wise operations where one tensor
* can be broadcasted to match the shape of another tensor.
*
* @param src0 Pointer to the first ggml_tensor.
* @param src1 Pointer to the second ggml_tensor.
* @param bcast_ne_src0 Output array to store broadcasted dimensions for src0.
* @param bcast_ne_src1 Output array to store broadcasted dimensions for src1.
* @param bcast_nb_src0 Output array to store broadcasted strides for src0.
* @param bcast_nb_src1 Output array to store broadcasted strides for src1.
* @return Number of dimensions in the broadcasted shape.
*
* @pre ggml_can_repeat(src1, src0) must return true, indicating src1 can be broadcasted
* to match src0.
*
* @remarks This function iterates over the dimensions of src0 and src1, calculating the
* necessary broadcast dimensions and strides. If a dimension requires broadcasting
* (i.e., its size in src1 is smaller than in src0), an additional dimension is
* added with size calculated to match src0's dimension. This adjustment ensures
* that src1 can be element-wise broadcasted to src0's shape.
*
* How it works:
*
* if dim0 has padding.
* a -> (2, 2) padding = 2
* a: [[1, 2, *, *]
* [2, 3, *, *]]
* nb = (8, 4, 2)
*
* if a should bcast with b -> (2, 4)
* b' -> (2, 2, 2)
* b : [[1, 2, 3, 4, *, *]
* [5, 6, 7, 8, *, *]]
* nb = (12, 6, 1)
*
* after bcast:
* a' -> (2, 1, 2)
* a': [[[1, 2], *, *]
* [[2, 3], *, *]]
* nb = (8, 4, 2, 1)
*
* b' : [[[1, 2], [3, 4], *, *]
* [[5, 6], [7, 8], *, *]]
* nb = (12, 6, 2, 1)
* \endcode
*
* dim1 in a inserted dim, should add nb for dim1,
* and all other nb moves to next in order.
*/
int64_t ggml_cann_get_bcast_shape(const ggml_tensor * src0,
const ggml_tensor * src1,
int64_t * bcast_ne_src0,
int64_t * bcast_ne_src1,
size_t * bcast_nb_src0,
size_t * bcast_nb_src1);
// Bcast macro to avoid duplicate code.
#define BCAST_SHAPE(src0, src1) \
int64_t bcast_##src0##_ne[GGML_MAX_DIMS * 2]; \
int64_t bcast_##src1##_ne[GGML_MAX_DIMS * 2]; \
size_t bcast_##src0##_nb[GGML_MAX_DIMS * 2]; \
size_t bcast_##src1##_nb[GGML_MAX_DIMS * 2]; \
int64_t bcast_dims = ggml_cann_get_bcast_shape(src0, src1, bcast_##src0##_ne, bcast_##src1##_ne, \
bcast_##src0##_nb, bcast_##src1##_nb);
#define BCAST_PARAM(tensor) bcast_##tensor##_ne, bcast_##tensor##_nb, bcast_dims
/**
* @brief Calculates broadcast shapes for matrix multiplication.
*
* @details This function computes the broadcast shapes required for matrix multiplication
* based on the input, weight, and destination tensor shapes. It ensures that the
* dimensions of weight tensors are expanded appropriately to satisfy matrix
* multiplication broadcast rules.
*
* @param input_ne Array containing the dimensions of the input tensor.
* @param weight_ne Array containing the dimensions of the weight tensor.
* @param dst_ne Array containing the dimensions of the destination tensor.
* @param input_nb Array containing the strides of the input tensor.
* @param weight_nb Array containing the strides of the weight tensor.
* @param dst_nb Array containing the strides of the destination tensor.
* @param bcast_input_ne Output array for broadcasted input tensor dimensions.
* @param bcast_weight_ne Output array for broadcasted weight tensor dimensions.
* @param bcast_dst_ne Output array for broadcasted destination tensor dimensions.
* @param bcast_input_nb Output array for broadcasted input tensor strides.
* @param bcast_weight_nb Output array for broadcasted weight tensor strides.
* @param bcast_dst_nb Output array for broadcasted destination tensor strides.
* @return The number of dimensions in the broadcasted tensors.
*
* @remarks This function iterates over the tensor dimensions and calculates the broadcast
* shapes needed for matrix multiplication. It ensures that dimensions where
* weight tensor requires expansion are appropriately handled to conform with
* broadcasting rules.
* @note compare with ggml_cann_get_bcast_shape, mul_mat broadcast need add this new dim
* before cast dim.
* @sa ggml_cann_get_bcast_shape
*/
int64_t ggml_cann_get_mulmat_bcast_shape(const int64_t * input_ne,
const int64_t * weight_ne,
const int64_t * dst_ne,
const size_t * input_nb,
const size_t * weight_nb,
const size_t * dst_nb,
int64_t * bcast_input_ne,
int64_t * bcast_weight_ne,
int64_t * bcast_dst_ne,
size_t * bcast_input_nb,
size_t * bcast_weight_nb,
size_t * bcast_dst_nb);
// Bcast macro to avoid duplicate code.
#define BCAST_MUL_MAT_SHAPE(input, weight, dst) \
int64_t bcast_##input##_ne[GGML_MAX_DIMS * 2]; \
int64_t bcast_##weight##_ne[GGML_MAX_DIMS * 2]; \
int64_t bcast_##dst##_ne[GGML_MAX_DIMS * 2]; \
size_t bcast_##input##_nb[GGML_MAX_DIMS * 2]; \
size_t bcast_##weight##_nb[GGML_MAX_DIMS * 2]; \
size_t bcast_##dst##_nb[GGML_MAX_DIMS * 2]; \
int64_t bcast_dims = ggml_cann_get_mulmat_bcast_shape( \
input->ne, weight->ne, dst->ne, input->nb, weight->nb, dst->nb, bcast_##input##_ne, bcast_##weight##_ne, \
bcast_##dst##_ne, bcast_##input##_nb, bcast_##weight##_nb, bcast_##dst##_nb);
#define BCAST_MUL_MAT_PARAM(tensor) bcast_##tensor##_ne, bcast_##tensor##_nb, bcast_dims
#endif // CANN_ACL_TENSOR_H
engine/ggml/src/ggml-cann/aclnn_ops.cpp
+4436
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cann/aclnn_ops.h
+1190
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cann/common.h
+651
View File
@@ -0,0 +1,651 @@
/*
* Copyright (c) 2023-2026 The ggml authors
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
#ifndef CANN_COMMON_H
#define CANN_COMMON_H
#include "../ggml-impl.h"
#include "../include/ggml-cann.h"
#include "../include/ggml.h"
#include <acl/acl.h>
#include <unistd.h>
#include <atomic>
#include <condition_variable>
#include <cstdio>
#include <functional>
#include <iostream>
#include <list>
#include <map>
#include <memory>
#include <mutex>
#include <optional>
#include <string>
#include <thread>
#include <vector>
#define MATRIX_ROW_PADDING 512
#define GGML_CANN_MAX_STREAMS 8
/**
* @brief Handles CANN-related errors by printing an error message and
* terminating the program.
* @param stmt The statement that caused the error.
* @param func The function in which the error occurred.
* @param file The file in which the error occurred.
* @param line The line number at which the error occurred.
* @param msg The error message.
*/
[[noreturn]] void ggml_cann_error(const char * stmt, const char * func, const char * file, int line, const char * msg);
/**
* @brief Checks the result of a CANN function call and invokes the error
* handler if the call fails.
* @param stmt The CANN function call to check.
* @param success The success code that indicates the call was successful.
* @param error_fn The function to call to retrieve the error message.
*/
#define ACL_CHECK_GEN(stmt, success, error_fn) \
do { \
int err_code = (stmt); \
if (err_code != (success)) { \
ggml_cann_error(#stmt, __func__, __FILE__, __LINE__, error_fn()); \
} \
} while (0);
#define ACL_CHECK(stmt) ACL_CHECK_GEN(stmt, 0, aclGetRecentErrMsg)
/**
* @brief Contains information about CANN devices.
*/
struct ggml_cann_device_info {
/**
* @brief Number of CANN devices available.
*/
int32_t device_count;
/**
* @brief Information about a single CANN device.
*/
struct cann_device_info {
int cc; /**< Compute capability. */
size_t smpb; /**< Maximum shared memory per block. */
bool vmm; /**< Virtual memory support. */
size_t vmm_granularity; /**< Granularity of virtual memory. */
size_t total_vram; /**< Total video RAM available on the device. */
};
cann_device_info devices[GGML_CANN_MAX_DEVICES] = {}; /**< Array of CANN device information. */
};
const ggml_cann_device_info & ggml_cann_info();
void ggml_cann_set_device(int32_t device);
std::optional<std::string> get_env_as_lowercase(const std::string & name);
bool parse_bool(const std::string & value);
int parse_integer(const std::string & value);
/**
* @brief Abstract base class for memory pools used by CANN.
*/
struct ggml_cann_pool {
/**
* @brief Virtual destructor for the memory pool.
*/
virtual ~ggml_cann_pool() = default;
/**
* @brief Allocates memory from the pool.
*
* @param size The size of the memory block to allocate.
* @param actual_size Pointer to a variable where the actual allocated size
* will be stored.
* @return Pointer to the allocated memory block.
*/
virtual void * alloc(size_t size, size_t * actual_size) = 0;
/**
* @brief Frees a previously allocated memory block.
*
* @param ptr Pointer to the memory block to free.
* @param size Size of the memory block to free.
* @note Note that all CANN opertors are running async. Make sure memory is
* still avaiable before this operator finished.
*/
virtual void free(void * ptr, size_t size) = 0;
};
/**
* @brief RAII wrapper for managing memory allocations from a CANN memory pool.
*/
struct ggml_cann_pool_alloc {
ggml_cann_pool * pool = nullptr; /**< Pointer to the memory pool. */
void * ptr = nullptr; /**< Pointer to the allocated memory block. */
size_t actual_size = 0; /**< Actual size of the allocated memory block. */
/**
* @brief Default constructor.
*/
ggml_cann_pool_alloc() = default;
/**
* @brief Constructor that initializes the memory pool.
* @param pool Reference to the memory pool.
*/
explicit ggml_cann_pool_alloc(ggml_cann_pool & pool) : pool(&pool) {}
/**
* @brief Constructor that initializes the memory pool and allocates memory.
* @param pool Reference to the memory pool.
* @param size Size of the memory block to allocate.
*/
ggml_cann_pool_alloc(ggml_cann_pool & pool, size_t size) : pool(&pool) { alloc(size); }
/**
* @brief Destructor that frees the allocated memory block.
*/
~ggml_cann_pool_alloc() {
if (ptr != nullptr) {
pool->free(ptr, actual_size);
}
}
/**
* @brief Allocates memory from the pool.
* @param size Size of the memory block to allocate.
* @return Pointer to the allocated memory block.
*/
void * alloc(size_t size) {
GGML_ASSERT(pool != nullptr);
GGML_ASSERT(ptr == nullptr);
ptr = pool->alloc(size, &this->actual_size);
return ptr;
}
/**
* @brief Allocates memory from a specific memory pool.
* @param pool Reference to the memory pool.
* @param size Size of the memory block to allocate.
* @return Pointer to the allocated memory block.
*/
void * alloc(ggml_cann_pool & pool, size_t size) {
this->pool = &pool;
return alloc(size);
}
/**
* @brief Gets the pointer to the allocated memory block.
* @return Pointer to the allocated memory block.
*/
void * get() { return ptr; }
// Deleted copy constructor
ggml_cann_pool_alloc(const ggml_cann_pool_alloc &) = delete;
// Deleted move constructor
ggml_cann_pool_alloc(ggml_cann_pool_alloc &&) = delete;
// Deleted copy assignment operator
ggml_cann_pool_alloc & operator=(const ggml_cann_pool_alloc &) = delete;
// Deleted move assignment operator
ggml_cann_pool_alloc & operator=(ggml_cann_pool_alloc &&) = delete;
};
#ifdef USE_ACL_GRAPH
struct ggml_graph_node_properties {
// dst tensor
void * node_address;
ggml_type node_type;
int64_t ne[GGML_MAX_DIMS];
size_t nb[GGML_MAX_DIMS];
// src tensor
void * src_address[GGML_MAX_SRC];
ggml_type src_type[GGML_MAX_SRC];
int64_t src_ne[GGML_MAX_SRC][GGML_MAX_DIMS];
size_t src_nb[GGML_MAX_SRC][GGML_MAX_DIMS];
// op
ggml_op node_op;
int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
/**
* @brief Check if a ggml tensor node matches this property set.
*
* This function compares all relevant fields (address, op type, shape, source inputs, op params)
* to determine whether the current node matches these previously recorded properties.
*
* @param node The current ggml tensor node.
* @return true if all fields match (excluding GGML_OP_VIEW); false otherwise.
*/
bool has_matching_properties(ggml_tensor * node) {
if (node->data != this->node_address && node->op != GGML_OP_VIEW) {
return false;
}
if (node->op != this->node_op) {
return false;
}
if (node->type != this->node_type) {
return false;
}
for (int i = 0; i < GGML_MAX_DIMS; i++) {
if (node->ne[i] != this->ne[i]) {
return false;
}
if (node->nb[i] != this->nb[i]) {
return false;
}
}
for (int i = 0; i < GGML_MAX_SRC; i++) {
if (node->src[i]) {
if (node->src[i]->data != this->src_address[i] && node->op != GGML_OP_VIEW) {
return false;
}
if (node->src[i]->type != this->src_type[i]) {
return false;
}
for (int d = 0; d < GGML_MAX_DIMS; d++) {
if (node->src[i]->ne[d] != this->src_ne[i][d]) {
return false;
}
if (node->src[i]->nb[d] != this->src_nb[i][d]) {
return false;
}
}
} else {
if (this->src_address[i] != nullptr) {
return false;
}
}
}
return memcmp(this->op_params, node->op_params, GGML_MAX_OP_PARAMS) == 0;
}
};
struct ggml_cann_graph {
~ggml_cann_graph() {
if (graph != nullptr) {
ACL_CHECK(aclmdlRIDestroy(graph));
}
}
aclmdlRI graph = nullptr;
std::vector<ggml_graph_node_properties> ggml_graph_properties;
/**
* @brief Create a new CANN graph from a ggml computation graph.
*
* This function creates a new ggml_cann_graph object and fills its node properties
* (operation type, dimensions, strides, input sources, and operation parameters)
* based on the current ggml computation graph.
*
* Each node in the ggml graph is mapped to a property entry in the new CANN graph:
* - node address
* - operation type
* - shape (ne) and strides (nb)
* - source tensor addresses
* - operation parameters
*
* @param cgraph The current ggml computation graph.
* @return Pointer to the newly created ggml_cann_graph object.
*/
static ggml_cann_graph * create_from_cgraph(ggml_cgraph * cgraph) {
ggml_cann_graph * new_graph = new ggml_cann_graph();
new_graph->ggml_graph_properties.resize(cgraph->n_nodes);
for (int node_idx = 0; node_idx < cgraph->n_nodes; ++node_idx) {
ggml_tensor * node = cgraph->nodes[node_idx];
auto & prop = new_graph->ggml_graph_properties[node_idx];
prop.node_address = node->data;
prop.node_op = node->op;
prop.node_type = node->type;
std::copy_n(node->ne, GGML_MAX_DIMS, prop.ne);
std::copy_n(node->nb, GGML_MAX_DIMS, prop.nb);
for (int src = 0; src < GGML_MAX_SRC; ++src) {
if (node->src[src]) {
prop.src_address[src] = node->src[src]->data;
prop.src_type[src] = node->src[src]->type;
std::copy_n(node->src[src]->ne, GGML_MAX_DIMS, prop.src_ne[src]);
std::copy_n(node->src[src]->nb, GGML_MAX_DIMS, prop.src_nb[src]);
} else {
prop.src_address[src] = nullptr;
prop.src_type[src] = GGML_TYPE_COUNT;
std::fill_n(prop.src_ne[src], GGML_MAX_DIMS, 0);
std::fill_n(prop.src_nb[src], GGML_MAX_DIMS, 0);
}
}
memcpy(prop.op_params, node->op_params, GGML_MAX_OP_PARAMS);
}
return new_graph;
}
/**
* @brief Check whether this CANN graph matches the given ggml computation graph.
*
* This function compares the number of nodes and each node's properties
* (operation type, dimensions, strides, inputs, and operation parameters)
* to determine whether this CANN graph matches the given ggml graph.
*
* @param cgraph The current ggml computation graph.
* @return true if this CANN graph matches the ggml graph; false otherwise.
*/
bool matches_cgraph(ggml_cgraph * cgraph) {
if (this->ggml_graph_properties.size() != static_cast<size_t>(cgraph->n_nodes)) {
return false;
}
for (int i = 0; i < cgraph->n_nodes; ++i) {
if (!this->ggml_graph_properties[i].has_matching_properties(cgraph->nodes[i])) {
return false;
}
}
return true;
}
};
/**
* @brief LRU cache for managing ggml_cann_graph objects.
*
* This class maintains a list of shared_ptr to ggml_cann_graph objects
* and enforces a maximum capacity. It provides methods to push new graphs,
* move existing graphs to the front (most recently used), and clear the cache.
*/
struct ggml_cann_graph_lru_cache {
size_t capacity; /**< Maximum number of graphs in the cache. */
std::list<ggml_cann_graph *> cache_list; /**< List storing cached graphs as raw pointers. */
ggml_cann_graph_lru_cache() { capacity = parse_integer(get_env_as_lowercase("GGML_CANN_GRAPH_CACHE_CAPACITY").value_or("12")); }
/**
* @brief Push a new graph to the front of the cache.
* If the cache exceeds capacity, the least recently used graph is deleted.
* @param new_node Pointer to the new ggml_cann_graph to cache.
* Ownership is transferred to the cache (cache will delete it).
*/
void push(ggml_cann_graph * new_node) {
if (cache_list.size() >= capacity) {
ggml_cann_graph * old = cache_list.back();
cache_list.pop_back();
delete old; // free the old graph
}
cache_list.push_front(new_node);
}
/**
* @brief Clear all graphs from the cache (also frees memory).
*/
void clear() {
for (auto ptr : cache_list) {
delete ptr;
}
cache_list.clear();
}
/**
* @brief Destructor that clears the cache and frees all cached graphs.
*/
~ggml_cann_graph_lru_cache() { clear(); }
/**
* @brief Find a cached CANN graph that matches the given ggml graph and move it to front.
*
* This function iterates through the cached CANN graphs stored in the LRU cache and
* compares them against the given ggml computation graph. If a matching graph is found,
* it is promoted to the front of the LRU cache and returned. Otherwise, the function
* returns nullptr.
*
* @param cgraph The current ggml computation graph.
* @return true if found; false otherwise.
*/
bool find_and_move_to_front(ggml_cgraph * cgraph) {
for (auto & graph_ptr : this->cache_list) {
if (graph_ptr->matches_cgraph(cgraph)) {
cache_list.remove(graph_ptr);
cache_list.push_front(graph_ptr);
return true;
}
}
return false;
}
};
#endif // USE_ACL_GRAPH
struct ggml_cann_rope_cache {
~ggml_cann_rope_cache() {
if (theta_scale_cache) {
ACL_CHECK(aclrtFree(theta_scale_cache));
}
if (sin_cache) {
ACL_CHECK(aclrtFree(sin_cache));
}
if (cos_cache) {
ACL_CHECK(aclrtFree(cos_cache));
}
if (position_select_index) {
ACL_CHECK(aclrtFree(position_select_index));
}
if (theta_scale_exp_host) {
free(theta_scale_exp_host);
}
if (position_select_index_host) {
free(position_select_index_host);
}
if (yarn_ramp_cache) {
ACL_CHECK(aclrtFree(yarn_ramp_cache));
}
}
bool equal(int64_t theta_scale_length,
int64_t position_length,
float ext_factor,
float theta_scale,
float freq_scale,
float attn_factor,
bool is_neox,
bool indep_sects,
bool mrope_used,
bool is_imrope,
int sections[4]) {
return this->theta_scale_length == theta_scale_length && this->position_length == position_length &&
this->ext_factor == ext_factor && this->theta_scale == theta_scale && this->freq_scale == freq_scale &&
this->attn_factor == attn_factor && this->is_neox == is_neox && this->indep_sects == indep_sects &&
this->mrope_used == mrope_used && this->is_imrope == is_imrope && this->sections[0] == sections[0] &&
this->sections[1] == sections[1] && this->sections[2] == sections[2] && this->sections[3] == sections[3];
}
void set(int64_t theta_scale_length,
int64_t position_length,
float ext_factor,
float theta_scale,
float freq_scale,
float attn_factor,
bool is_neox,
bool indep_sects,
bool mrope_used,
bool is_imrope,
int sections[4]) {
this->theta_scale_length = theta_scale_length;
this->position_length = position_length;
this->ext_factor = ext_factor;
this->theta_scale = theta_scale;
this->freq_scale = freq_scale;
this->attn_factor = attn_factor;
this->is_neox = is_neox;
this->indep_sects = indep_sects;
this->mrope_used = mrope_used;
this->is_imrope = is_imrope;
this->sections[0] = sections[0];
this->sections[1] = sections[1];
this->sections[2] = sections[2];
this->sections[3] = sections[3];
}
// memory cache, prepare before inferencing.
void * theta_scale_cache = nullptr;
float * theta_scale_exp_host = nullptr;
int * position_select_index_host = nullptr;
void * position_select_index = nullptr;
void * yarn_ramp_cache = nullptr;
// sin/cos cache, used only to accelerate first layer on each device
void * sin_cache = nullptr;
void * cos_cache = nullptr;
// Properties to check before reusing the sincos cache
int64_t theta_scale_length = 0;
int64_t position_length = 0;
bool cached = false;
float ext_factor = 0.0f;
float theta_scale = 0.0f;
float freq_scale = 0.0f;
float attn_factor = 0.0f;
bool is_neox = false;
bool indep_sects = false;
bool mrope_used = false;
int sections[4] = { 0, 0, 0, 0 };
bool is_imrope = false;
};
struct ggml_cann_tensor_cache {
~ggml_cann_tensor_cache() {
if (cache != nullptr) {
ACL_CHECK(aclrtFree(cache));
}
}
void * cache = nullptr;
int64_t size = 0;
};
/**
* @brief Context for managing CANN backend operations.
*/
struct ggml_backend_cann_context {
int32_t device; /**< Device ID. */
std::string name; /**< Name of the device. */
std::string description; /**< Description of the device. */
aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
#ifdef USE_ACL_GRAPH
/// Cached CANN ACL graph used for executing the current ggml computation graph.
ggml_cann_graph_lru_cache graph_lru_cache;
bool acl_graph_mode = true;
#endif
bool async_mode;
// Rope Cache
ggml_cann_rope_cache rope_cache;
// Constant Pool
ggml_cann_tensor_cache rms_norm_one_tensor_cache;
ggml_cann_tensor_cache rms_norm_zero_tensor_cache;
aclrtStream streams[GGML_CANN_MAX_STREAMS] = { nullptr }; /**< Array of streams for the device. */
/**
* @brief Constructor for initializing the context with a given device.
* @param device Device ID.
*/
explicit ggml_backend_cann_context(int device) : device(device), name("CANN" + std::to_string(device)) {
ggml_cann_set_device(device);
description = aclrtGetSocName();
#ifdef USE_ACL_GRAPH
acl_graph_mode = parse_bool(get_env_as_lowercase("GGML_CANN_ACL_GRAPH").value_or("on"));
GGML_LOG_INFO("%s: device %d execution mode is %s (%s)\n", __func__, device, acl_graph_mode ? "GRAPH" : "EAGER",
acl_graph_mode ? "acl graph enabled" : "acl graph disabled");
#endif
}
/**
* @brief Destructor for cleaning up resources.
*/
~ggml_backend_cann_context() {
ggml_cann_set_device(device);
if (copy_event != nullptr) {
ACL_CHECK(aclrtDestroyEvent(copy_event));
}
for (int i = 0; i < GGML_CANN_MAX_STREAMS; ++i) {
if (streams[i] != nullptr) {
ACL_CHECK(aclrtDestroyStream(streams[i]));
}
}
}
/**
* @brief Get or create a stream for a given index.
* @param stream Index of the stream.
* @return The stream corresponding to the given index.
*/
aclrtStream stream(int stream) {
if (streams[stream] == nullptr) {
// If the device is not set here, destroying the stream later may cause a mismatch
// between the thread contexts where the stream was created and destroyed.
// However, I printed the device_id, thread_id, and stream, and they are all consistent.
ACL_CHECK(aclrtSetDevice(device));
ACL_CHECK(aclrtCreateStream(&streams[stream]));
}
return streams[stream];
}
/**
* @brief Get or create the default stream (index 0).
* @return The default stream.
*/
aclrtStream stream() { return stream(0); }
// TODO: each stream should have a memory pool.
std::unique_ptr<ggml_cann_pool> mem_pool; /**< Memory pool for the device. */
/**
* @brief Create a new memory pool for a given device.
* @param device Device ID.
* @return A unique pointer to the new memory pool.
*/
static std::unique_ptr<ggml_cann_pool> new_pool_for_device(int device);
/**
* @brief Get or create the memory pool for the context.
* @return Reference to the memory pool.
*/
ggml_cann_pool & pool() {
if (mem_pool == nullptr) {
mem_pool = new_pool_for_device(device);
}
return *mem_pool;
}
};
#endif // CANN_COMMON_H
engine/ggml/src/ggml-cann/ggml-cann.cpp
+3062
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-common.h
+1900
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/CMakeLists.txt
+723
View File
@@ -0,0 +1,723 @@
function(ggml_add_cpu_backend_features cpu_name arch)
# The feature detection code is compiled as a separate target so that
# it can be built without the architecture flags
# Since multiple variants of the CPU backend may be included in the same
# build, using set_source_files_properties() to set the arch flags is not possible
set(GGML_CPU_FEATS_NAME ${cpu_name}-feats)
add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/${arch}/cpu-feats.cpp)
target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . ../include)
target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARGN})
target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
# Disable LTO for the feature detection code to prevent cross-module optimization
# from inlining architecture-specific instructions into the score function.
# Without this, LTO can cause SIGILL when loading backends on older CPUs
# (e.g., loading power10 backend on power9 crashes before feature check runs).
target_compile_options(${GGML_CPU_FEATS_NAME} PRIVATE -fno-lto)
target_link_libraries(${cpu_name} PRIVATE ${GGML_CPU_FEATS_NAME})
endfunction()
function(ggml_add_cpu_backend_variant_impl tag_name)
if (tag_name)
set(GGML_CPU_NAME ggml-cpu-${tag_name})
else()
set(GGML_CPU_NAME ggml-cpu)
endif()
ggml_add_backend_library(${GGML_CPU_NAME})
list (APPEND GGML_CPU_SOURCES
ggml-cpu/ggml-cpu.c
ggml-cpu/ggml-cpu.cpp
ggml-cpu/repack.cpp
ggml-cpu/repack.h
ggml-cpu/hbm.cpp
ggml-cpu/hbm.h
ggml-cpu/quants.c
ggml-cpu/quants.h
ggml-cpu/traits.cpp
ggml-cpu/traits.h
ggml-cpu/amx/amx.cpp
ggml-cpu/amx/amx.h
ggml-cpu/amx/mmq.cpp
ggml-cpu/amx/mmq.h
ggml-cpu/ggml-cpu-impl.h
ggml-cpu/common.h
ggml-cpu/binary-ops.h
ggml-cpu/binary-ops.cpp
ggml-cpu/unary-ops.h
ggml-cpu/unary-ops.cpp
ggml-cpu/simd-mappings.h
ggml-cpu/vec.h
ggml-cpu/vec.cpp
ggml-cpu/ops.h
ggml-cpu/ops.cpp
)
target_compile_features(${GGML_CPU_NAME} PRIVATE c_std_11 cxx_std_17)
target_include_directories(${GGML_CPU_NAME} PRIVATE . ggml-cpu)
if (APPLE AND GGML_ACCELERATE)
find_library(ACCELERATE_FRAMEWORK Accelerate)
if (ACCELERATE_FRAMEWORK)
message(STATUS "Accelerate framework found")
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_ACCELERATE)
target_compile_definitions(${GGML_CPU_NAME} PRIVATE ACCELERATE_NEW_LAPACK)
target_compile_definitions(${GGML_CPU_NAME} PRIVATE ACCELERATE_LAPACK_ILP64)
target_link_libraries(${GGML_CPU_NAME} PRIVATE ${ACCELERATE_FRAMEWORK})
else()
message(WARNING "Accelerate framework not found")
endif()
endif()
if (GGML_OPENMP_ENABLED)
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_OPENMP)
target_link_libraries(${GGML_CPU_NAME} PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
endif()
if (GGML_LLAMAFILE)
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_LLAMAFILE)
list(APPEND GGML_CPU_SOURCES
ggml-cpu/llamafile/sgemm.cpp
ggml-cpu/llamafile/sgemm.h)
endif()
if (GGML_CPU_HBM)
find_library(memkind memkind REQUIRED)
message(STATUS "Using memkind for CPU HBM")
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_HBM)
target_link_libraries(${GGML_CPU_NAME} PUBLIC memkind)
endif()
if (GGML_SYSTEM_ARCH STREQUAL "ARM")
message(STATUS "ARM detected")
list(APPEND GGML_CPU_SOURCES
ggml-cpu/arch/arm/quants.c
ggml-cpu/arch/arm/repack.cpp
)
if (MSVC AND NOT CMAKE_C_COMPILER_ID STREQUAL "Clang")
message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
else()
check_cxx_compiler_flag(-mfp16-format=ieee GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E)
if (NOT "${GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
list(APPEND ARCH_FLAGS -mfp16-format=ieee)
endif()
if (GGML_NATIVE)
# -mcpu=native does not always enable all the features in some compilers,
# so we check for them manually and enable them if available
execute_process(
COMMAND ${CMAKE_C_COMPILER} -mcpu=native -E -v -
INPUT_FILE "/dev/null"
OUTPUT_QUIET
ERROR_VARIABLE ARM_MCPU
RESULT_VARIABLE ARM_MCPU_RESULT
)
if (NOT ARM_MCPU_RESULT)
string(REGEX MATCH "-mcpu=[^ ']+" ARM_MCPU_FLAG "${ARM_MCPU}")
string(REGEX MATCH "-march=[^ ']+" ARM_MARCH_FLAG "${ARM_MCPU}")
# on some old GCC we need to read -march=
if (ARM_MARCH_FLAG AND NOT "${ARM_MARCH_FLAG}" STREQUAL "-march=native")
set(ARM_NATIVE_FLAG "${ARM_MARCH_FLAG}")
elseif(ARM_MCPU_FLAG AND NOT "${ARM_MCPU_FLAG}" STREQUAL "-mcpu=native")
set(ARM_NATIVE_FLAG "${ARM_MCPU_FLAG}")
endif()
endif()
if ("${ARM_NATIVE_FLAG}" STREQUAL "")
set(ARM_NATIVE_FLAG -mcpu=native)
message(WARNING "ARM -march/-mcpu not found, -mcpu=native will be used")
else()
message(STATUS "ARM detected flags: ${ARM_NATIVE_FLAG}")
endif()
include(CheckCXXSourceRuns)
macro(check_arm_feature tag feature code)
set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+${tag}")
check_cxx_source_runs("${code}" GGML_MACHINE_SUPPORTS_${tag})
if (GGML_MACHINE_SUPPORTS_${tag})
set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+${tag}")
else()
set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+no${tag}")
check_cxx_source_compiles("int main() { return 0; }" GGML_MACHINE_SUPPORTS_no${tag})
if (GGML_MACHINE_SUPPORTS_no${tag})
set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+no${tag}")
list(APPEND ARCH_FLAGS -U__ARM_FEATURE_${feature})
endif()
endif()
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
endmacro()
check_arm_feature(dotprod DOTPROD "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }")
check_arm_feature(i8mm MATMUL_INT8 "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }")
check_arm_feature(sve SVE "#include <arm_sve.h>\nint main() { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }")
check_arm_feature(sme SME "#include <arm_sme.h>\n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }")
list(APPEND ARCH_FLAGS "${ARM_NATIVE_FLAG}${ARM_NATIVE_FLAG_FIX}")
else()
if (GGML_CPU_ARM_ARCH)
list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
elseif(GGML_CPU_ALL_VARIANTS)
# Begin with the lowest baseline
set(ARM_MCPU "armv8-a")
set(ARCH_TAGS "")
set(ARCH_DEFINITIONS "")
# When a feature is selected, bump the MCPU to the first
# version that supported it
if (GGML_INTERNAL_DOTPROD)
set(ARM_MCPU "armv8.2-a")
set(ARCH_TAGS "${ARCH_TAGS}+dotprod")
list(APPEND ARCH_DEFINITIONS GGML_USE_DOTPROD)
endif()
if (GGML_INTERNAL_FP16_VECTOR_ARITHMETIC)
set(ARM_MCPU "armv8.2-a")
set(ARCH_TAGS "${ARCH_TAGS}+fp16")
list(APPEND ARCH_DEFINITIONS GGML_USE_FP16_VECTOR_ARITHMETIC)
endif()
if (GGML_INTERNAL_SVE)
set(ARM_MCPU "armv8.2-a")
set(ARCH_TAGS "${ARCH_TAGS}+sve")
list(APPEND ARCH_DEFINITIONS GGML_USE_SVE)
endif()
if (GGML_INTERNAL_MATMUL_INT8)
set(ARM_MCPU "armv8.6-a")
set(ARCH_TAGS "${ARCH_TAGS}+i8mm")
list(APPEND ARCH_DEFINITIONS GGML_USE_MATMUL_INT8)
endif()
if (GGML_INTERNAL_SVE2)
set(ARM_MCPU "armv8.6-a")
set(ARCH_TAGS "${ARCH_TAGS}+sve2")
list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2)
endif()
if (GGML_INTERNAL_NOSVE)
set(ARCH_TAGS "${ARCH_TAGS}+nosve")
endif()
if (GGML_INTERNAL_SME)
set(ARM_MCPU "armv9.2-a")
set(ARCH_TAGS "${ARCH_TAGS}+sme")
list(APPEND ARCH_DEFINITIONS GGML_USE_SME)
endif()
list(APPEND ARCH_FLAGS "-march=${ARM_MCPU}${ARCH_TAGS}")
ggml_add_cpu_backend_features(${GGML_CPU_NAME} arm ${ARCH_DEFINITIONS})
endif()
endif()
message(STATUS "Checking for ARM features using flags:")
foreach(flag IN LISTS ARCH_FLAGS)
message(STATUS " ${flag}")
endforeach()
include(CheckCXXSourceCompiles)
set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
string(REPLACE ";" " " ARCH_FLAGS_STR "${ARCH_FLAGS}")
set(CMAKE_REQUIRED_FLAGS "${ARCH_FLAGS_STR}")
foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
set(ARM_FEATURE "HAVE_${feature}")
check_cxx_source_compiles(
"
#if !defined(__ARM_FEATURE_${feature})
# error \"Feature ${feature} is not defined\"
#endif
int main() { return 0; }
"
${ARM_FEATURE}
)
endforeach()
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
endif()
elseif (GGML_SYSTEM_ARCH STREQUAL "x86")
message(STATUS "x86 detected")
list(APPEND GGML_CPU_SOURCES
ggml-cpu/arch/x86/quants.c
ggml-cpu/arch/x86/repack.cpp
)
if (MSVC)
# instruction set detection for MSVC only
if (GGML_NATIVE)
include(ggml-cpu/cmake/FindSIMD.cmake)
endif ()
if (GGML_AVX512)
list(APPEND ARCH_FLAGS /arch:AVX512)
# /arch:AVX512 includes: __AVX512F__, __AVX512CD__, __AVX512BW__, __AVX512DQ__, and __AVX512VL__
# MSVC has no compile-time flags enabling specific
# AVX512 extensions, neither it defines the
# macros corresponding to the extensions.
# Do it manually.
list(APPEND ARCH_DEFINITIONS GGML_AVX512)
if (GGML_AVX512_VBMI)
list(APPEND ARCH_DEFINITIONS __AVX512VBMI__)
if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
list(APPEND ARCH_FLAGS -mavx512vbmi)
endif()
endif()
if (GGML_AVX512_VNNI)
list(APPEND ARCH_DEFINITIONS __AVX512VNNI__ GGML_AVX512_VNNI)
if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
list(APPEND ARCH_FLAGS -mavx512vnni)
endif()
endif()
if (GGML_AVX512_BF16)
list(APPEND ARCH_DEFINITIONS __AVX512BF16__ GGML_AVX512_BF16)
if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
list(APPEND ARCH_FLAGS -mavx512bf16)
endif()
endif()
if (GGML_AMX_TILE)
list(APPEND ARCH_DEFINITIONS __AMX_TILE__ GGML_AMX_TILE)
endif()
if (GGML_AMX_INT8)
list(APPEND ARCH_DEFINITIONS __AMX_INT8__ GGML_AMX_INT8)
endif()
if (GGML_AMX_BF16)
list(APPEND ARCH_DEFINITIONS __AMX_BF16__ GGML_AMX_BF16)
endif()
elseif (GGML_AVX2)
list(APPEND ARCH_FLAGS /arch:AVX2)
list(APPEND ARCH_DEFINITIONS GGML_AVX2 GGML_FMA GGML_F16C)
elseif (GGML_AVX)
list(APPEND ARCH_FLAGS /arch:AVX)
list(APPEND ARCH_DEFINITIONS GGML_AVX)
elseif (GGML_SSE42)
list(APPEND ARCH_FLAGS /arch:SSE4.2)
list(APPEND ARCH_DEFINITIONS GGML_SSE42)
endif()
if (GGML_AVX_VNNI)
list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI)
endif()
if (GGML_BMI2)
# MSVC does not define macro __BMI2__
list(APPEND ARCH_DEFINITIONS __BMI2__ GGML_BMI2)
endif()
else ()
if (GGML_NATIVE)
list(APPEND ARCH_FLAGS -march=native)
else ()
if (GGML_SSE42)
list(APPEND ARCH_FLAGS -msse4.2)
list(APPEND ARCH_DEFINITIONS GGML_SSE42)
endif()
if (GGML_F16C)
list(APPEND ARCH_FLAGS -mf16c)
list(APPEND ARCH_DEFINITIONS GGML_F16C)
endif()
if (GGML_FMA)
list(APPEND ARCH_FLAGS -mfma)
list(APPEND ARCH_DEFINITIONS GGML_FMA)
endif()
if (GGML_BMI2)
list(APPEND ARCH_FLAGS -mbmi2)
list(APPEND ARCH_DEFINITIONS GGML_BMI2)
endif()
if (GGML_AVX)
list(APPEND ARCH_FLAGS -mavx)
list(APPEND ARCH_DEFINITIONS GGML_AVX)
endif()
if (GGML_AVX2)
list(APPEND ARCH_FLAGS -mavx2)
list(APPEND ARCH_DEFINITIONS GGML_AVX2)
endif()
if (GGML_AVX_VNNI)
list(APPEND ARCH_FLAGS -mavxvnni)
list(APPEND ARCH_DEFINITIONS GGML_AVX_VNNI)
endif()
if (GGML_AVX512)
list(APPEND ARCH_FLAGS -mavx512f)
list(APPEND ARCH_FLAGS -mavx512cd)
list(APPEND ARCH_FLAGS -mavx512vl)
list(APPEND ARCH_FLAGS -mavx512dq)
list(APPEND ARCH_FLAGS -mavx512bw)
list(APPEND ARCH_DEFINITIONS GGML_AVX512)
endif()
if (GGML_AVX512_VBMI)
list(APPEND ARCH_FLAGS -mavx512vbmi)
list(APPEND ARCH_DEFINITIONS GGML_AVX512_VBMI)
endif()
if (GGML_AVX512_VNNI)
list(APPEND ARCH_FLAGS -mavx512vnni)
list(APPEND ARCH_DEFINITIONS GGML_AVX512_VNNI)
endif()
if (GGML_AVX512_BF16)
list(APPEND ARCH_FLAGS -mavx512bf16)
list(APPEND ARCH_DEFINITIONS GGML_AVX512_BF16)
endif()
if (GGML_AMX_TILE)
list(APPEND ARCH_FLAGS -mamx-tile)
list(APPEND ARCH_DEFINITIONS GGML_AMX_TILE)
endif()
if (GGML_AMX_INT8)
list(APPEND ARCH_FLAGS -mamx-int8)
list(APPEND ARCH_DEFINITIONS GGML_AMX_INT8)
endif()
if (GGML_AMX_BF16)
list(APPEND ARCH_FLAGS -mamx-bf16)
list(APPEND ARCH_DEFINITIONS GGML_AMX_BF16)
endif()
endif()
endif()
if (GGML_BACKEND_DL)
if (GGML_NATIVE)
# the feature check relies on ARCH_DEFINITIONS, but it is not set with GGML_NATIVE
message(FATAL_ERROR "GGML_NATIVE is not compatible with GGML_BACKEND_DL, consider using GGML_CPU_ALL_VARIANTS")
endif()
ggml_add_cpu_backend_features(${GGML_CPU_NAME} x86 ${ARCH_DEFINITIONS})
endif()
elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
message(STATUS "PowerPC detected")
list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/powerpc/quants.c)
if (GGML_NATIVE)
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
file(READ "/proc/cpuinfo" POWER10_M)
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "powerpc")
execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M)
endif()
string(TOUPPER "${POWER10_M}" POWER10_M_UPPER)
string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M_UPPER}")
string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
if (EXTRACTED_NUMBER EQUAL 10 OR EXTRACTED_NUMBER EQUAL 11)
list(APPEND ARCH_FLAGS -mcpu=power10)
elseif (EXTRACTED_NUMBER EQUAL 9)
list(APPEND ARCH_FLAGS -mcpu=power9)
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
else()
list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
endif()
elseif(GGML_CPU_ALL_VARIANTS)
# Begin with the lowest baseline
set(ARCH_DEFINITIONS "")
# When a feature is selected, bump the MCPU to the first
# version that supported it
foreach(PVER RANGE 7 11)
if(DEFINED GGML_INTERNAL_POWER${PVER})
set(POWERPC_MCPU "power${PVER}")
list(APPEND ARCH_DEFINITIONS GGML_USE_POWER${PVER})
endif()
endforeach()
if (GGML_INTERNAL_VSX)
list(APPEND ARCH_DEFINITIONS GGML_USE_VSX)
list(APPEND ARCH_FLAGS -mvsx)
endif()
if (DEFINED POWERPC_MCPU)
list(APPEND ARCH_FLAGS -mcpu=${POWERPC_MCPU})
endif()
ggml_add_cpu_backend_features(${GGML_CPU_NAME} powerpc ${ARCH_DEFINITIONS})
else()
if (GGML_CPU_POWERPC_CPUTYPE)
list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE})
endif()
endif()
elseif (GGML_SYSTEM_ARCH STREQUAL "loongarch64")
message(STATUS "loongarch64 detected")
list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/loongarch/quants.c)
list(APPEND ARCH_FLAGS -march=loongarch64)
if (GGML_LASX)
list(APPEND ARCH_FLAGS -mlasx)
endif()
if (GGML_LSX)
list(APPEND ARCH_FLAGS -mlsx)
endif()
elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
message(STATUS "riscv64 detected")
list(APPEND GGML_CPU_SOURCES
ggml-cpu/arch/riscv/quants.c
ggml-cpu/arch/riscv/repack.cpp
)
if (GGML_CPU_RISCV64_SPACEMIT)
include(ggml-cpu/cmake/FindSMTIME.cmake)
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_RISCV64_SPACEMIT ${RISCV64_SPACEMIT_IME_SPEC})
list(APPEND GGML_CPU_SOURCES
ggml-cpu/spacemit/ime.cpp
ggml-cpu/spacemit/ime.h
ggml-cpu/spacemit/spine_mem_pool.cpp
ggml-cpu/spacemit/spine_mem_pool.h
ggml-cpu/spacemit/repack.cpp
ggml-cpu/spacemit/repack.h
ggml-cpu/spacemit/ime_env.cpp
ggml-cpu/spacemit/ime_env.h
ggml-cpu/spacemit/ime1_kernels.cpp
ggml-cpu/spacemit/ime2_kernels.cpp
ggml-cpu/spacemit/ime_kernels.h
ggml-cpu/spacemit/rvv_kernels.cpp
ggml-cpu/spacemit/rvv_kernels.h
)
endif()
if(NOT GGML_CPU_ALL_VARIANTS)
set(MARCH_STR "rv64gc")
if (GGML_RVV)
string(APPEND MARCH_STR "v")
endif()
if (GGML_RV_ZFH)
string(APPEND MARCH_STR "_zfh")
endif()
if (GGML_XTHEADVECTOR)
string(APPEND MARCH_STR "_xtheadvector")
elseif (GGML_RVV)
if (GGML_RV_ZVFH)
string(APPEND MARCH_STR "_zvfh")
endif()
if (GGML_RV_ZVFBFWMA)
string(APPEND MARCH_STR "_zvfbfwma")
endif()
endif()
if (GGML_RV_ZICBOP)
string(APPEND MARCH_STR "_zicbop")
endif()
if (GGML_RV_ZIHINTPAUSE)
string(APPEND MARCH_STR "_zihintpause")
endif()
if (GGML_RV_ZBA)
string(APPEND MARCH_STR "_zba")
endif()
if (GGML_CPU_RISCV64_SPACEMIT)
# `xsmtvdotii' is only required for GCC >= 15.
if (CMAKE_C_COMPILER_ID STREQUAL "GNU" AND
CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 15)
string(APPEND MARCH_STR "_xsmtvdotii")
endif()
endif()
list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
else()
# Begin with the lowest baseline
set(ARCH_DEFINITIONS "")
if (GGML_INTERNAL_RVV)
message(STATUS "RVV enabled")
list(APPEND ARCH_DEFINITIONS GGML_USE_RVV)
list(APPEND ARCH_FLAGS -march=rv64gc_v -mabi=lp64d)
endif()
ggml_add_cpu_backend_features(${GGML_CPU_NAME} riscv ${ARCH_DEFINITIONS})
endif()
elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
message(STATUS "s390x detected")
list(APPEND GGML_CPU_SOURCES
ggml-cpu/arch/s390/quants.c)
# for native compilation
if (GGML_NATIVE)
# check machine level to determine target
file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
# TODO: Separation to determine activation of VX/VXE/VXE2
if (${S390X_M} MATCHES "8561|8562")
message(STATUS "z15 target")
list(APPEND ARCH_FLAGS -march=z15)
elseif (${S390X_M} MATCHES "3931")
message(STATUS "z16 target")
list(APPEND ARCH_FLAGS -march=z16)
elseif (${S390X_M} MATCHES "9175|9176")
# NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
# binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
message(STATUS "z17 target")
list(APPEND ARCH_FLAGS -march=arch15)
else()
message(STATUS "Unknown target")
message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
list(APPEND ARCH_FLAGS -march=native -mtune=native)
endif()
# for cross-compilation
elseif(GGML_CPU_ALL_VARIANTS)
# range through IBM z15 to z17
# NOTE: update when a new hardware level is released
foreach (ZHW RANGE 15 17)
if(DEFINED GGML_INTERNAL_Z${ZHW})
message(STATUS "z${ZHW} cross-compile target")
list(APPEND ARCH_FLAGS -march=z${ZHW})
endif()
endforeach()
endif()
if (GGML_VXE OR GGML_INTERNAL_VXE2)
message(STATUS "VXE2 enabled")
list(APPEND ARCH_FLAGS -mvx -mzvector)
list(APPEND ARCH_DEFINITIONS GGML_USE_VXE2)
endif()
if (GGML_INTERNAL_NNPA)
message(STATUS "NNPA enabled")
list(APPEND ARCH_DEFINITIONS GGML_USE_NNPA)
endif()
ggml_add_cpu_backend_features(${GGML_CPU_NAME} s390 ${ARCH_DEFINITIONS})
elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
message(STATUS "Wasm detected")
list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)
else()
message(WARNING "Unknown CPU architecture. Falling back to generic implementations.")
list(APPEND ARCH_FLAGS -DGGML_CPU_GENERIC)
endif()
if (GGML_CPU_REPACK)
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_REPACK)
endif()
if (GGML_CPU_KLEIDIAI)
message(STATUS "Using KleidiAI optimized kernels if applicable")
# Disable the KleidiAI tests
set(KLEIDIAI_BUILD_TESTS OFF)
# Fetch KleidiAI sources:
include(FetchContent)
set(KLEIDIAI_COMMIT_TAG "v1.24.0")
set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/releases/download/${KLEIDIAI_COMMIT_TAG}/kleidiai-${KLEIDIAI_COMMIT_TAG}-src.tar.gz")
set(KLEIDIAI_RELEASE_ARCHIVE_MD5 "2f02ebe29573d45813e671eb304f2a00")
set(KLEIDIAI_FETCH_ARGS
URL ${KLEIDIAI_DOWNLOAD_URL}
URL_HASH MD5=${KLEIDIAI_RELEASE_ARCHIVE_MD5}
)
if (CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
list(APPEND KLEIDIAI_FETCH_ARGS DOWNLOAD_EXTRACT_TIMESTAMP NEW)
endif()
if (CMAKE_VERSION VERSION_GREATER_EQUAL "3.28")
FetchContent_Declare(KleidiAI_Download
${KLEIDIAI_FETCH_ARGS}
EXCLUDE_FROM_ALL
)
FetchContent_MakeAvailable(KleidiAI_Download)
FetchContent_GetProperties(KleidiAI_Download SOURCE_DIR KLEIDIAI_SRC)
else()
FetchContent_Declare(KleidiAI_Download
${KLEIDIAI_FETCH_ARGS}
)
FetchContent_GetProperties(KleidiAI_Download
SOURCE_DIR KLEIDIAI_SRC
POPULATED KLEIDIAI_POPULATED
)
if (NOT KLEIDIAI_POPULATED)
FetchContent_Populate(KleidiAI_Download)
FetchContent_GetProperties(KleidiAI_Download SOURCE_DIR KLEIDIAI_SRC)
endif()
endif()
add_compile_definitions(GGML_USE_CPU_KLEIDIAI)
list(APPEND GGML_CPU_SOURCES
ggml-cpu/kleidiai/kleidiai.cpp
ggml-cpu/kleidiai/kernels.cpp
ggml-cpu/kleidiai/kleidiai.h
ggml-cpu/kleidiai/kernels.h
)
# KleidiAI
include_directories(
${KLEIDIAI_SRC}/
${KLEIDIAI_SRC}/kai/
${KLEIDIAI_SRC}/kai/ukernels/
${KLEIDIAI_SRC}/kai/ukernels/matmul/
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)
set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}")
if (NOT ARCH_FLAGS_TEMP)
string(REGEX MATCH "-march=[^ ]+" ARCH_FLAGS_TEMP "${CMAKE_C_FLAGS}")
endif()
string(FIND "${ARCH_FLAGS_TEMP}" "+dotprod" DOTPROD_ENABLED)
string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED)
string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED)
string(FIND "${ARCH_FLAGS_TEMP}" "+sve" SVE_ENABLED)
set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS_TEMP})
list(APPEND GGML_KLEIDIAI_SOURCES
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c)
if (NOT DOTPROD_ENABLED MATCHES -1)
list(APPEND GGML_KLEIDIAI_SOURCES
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c)
endif()
if (NOT I8MM_ENABLED MATCHES -1)
list(APPEND GGML_KLEIDIAI_SOURCES
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c)
endif()
if (NOT SME_ENABLED MATCHES -1)
list(APPEND GGML_KLEIDIAI_SOURCES
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f16pmrx2_f32_neon.c
${KLEIDIAI_SRC}/kai/kai_common_sme_asm.S)
set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2+sme2+fp16")
endif()
if (NOT SVE_ENABLED MATCHES -1)
list(APPEND GGML_KLEIDIAI_SOURCES
${KLEIDIAI_SRC}/kai/kai_common_sve_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.c
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm_asm.S
${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm.c)
endif()
set_source_files_properties(${GGML_KLEIDIAI_SOURCES} PROPERTIES COMPILE_OPTIONS "${PRIVATE_ARCH_FLAGS}")
list(APPEND GGML_CPU_SOURCES ${GGML_KLEIDIAI_SOURCES})
endif()
message(STATUS "Adding CPU backend variant ${GGML_CPU_NAME}: ${ARCH_FLAGS} ${ARCH_DEFINITIONS}")
target_sources(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_SOURCES})
target_compile_options(${GGML_CPU_NAME} PRIVATE ${ARCH_FLAGS})
target_compile_definitions(${GGML_CPU_NAME} PRIVATE ${ARCH_DEFINITIONS})
if (EMSCRIPTEN)
set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
endif()
if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
# The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
target_compile_options(${GGML_CPU_NAME} PRIVATE "-fno-associative-math")
endif()
endfunction()
engine/ggml/src/ggml-cpu/amx/amx.cpp
+249
View File
@@ -0,0 +1,249 @@
#include "amx.h"
#include "common.h"
#include "mmq.h"
#include "ggml-backend-impl.h"
#include "ggml-backend.h"
#include "ggml-impl.h"
#include "ggml-cpu.h"
#include "traits.h"
#if defined(__linux__)
#include <sys/syscall.h>
#include <unistd.h>
#endif
#include <cstdlib>
#include <cstring>
#include <memory>
#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
// AMX type_trais
namespace ggml::cpu::amx {
class tensor_traits : public ggml::cpu::tensor_traits {
bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
size = ggml_backend_amx_desired_wsize(op);
return true;
}
bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) override {
if (op->op == GGML_OP_MUL_MAT) {
ggml_backend_amx_mul_mat(params, op);
return true;
}
return false;
}
};
static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struct ggml_tensor *) {
static tensor_traits traits;
return &traits;
}
} // namespace ggml::cpu::amx
// AMX buffer interface
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 enum ggml_status ggml_backend_amx_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
tensor->extra = (void *) ggml::cpu::amx::get_tensor_traits(buffer, tensor);
GGML_UNUSED(buffer);
return GGML_STATUS_SUCCESS;
}
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_LOG_DEBUG("%s: amx repack tensor %s of type %s\n", __func__, tensor->name, ggml_type_name(tensor->type));
ggml_backend_amx_convert_weight(tensor, data, offset, size);
} else {
memcpy((char *) tensor->data + offset, data, size);
}
GGML_UNUSED(buffer);
}
/*
// need to figure what we need to do with buffer->extra.
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_nbytes(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 = {
/* .free_buffer = */ ggml_backend_amx_buffer_free_buffer,
/* .get_base = */ ggml_backend_amx_buffer_get_base,
/* .init_tensor = */ ggml_backend_amx_buffer_init_tensor,
/* .memset_tensor = */ ggml_backend_amx_buffer_memset_tensor,
/* .set_tensor = */ ggml_backend_amx_buffer_set_tensor,
/* .get_tensor = */ nullptr,
/* .set_tensor_2d = */ nullptr,
/* .get_tensor_2d = */ nullptr,
/* .cpy_tensor = */ nullptr,
/* .clear = */ ggml_backend_amx_buffer_clear,
/* .reset = */ nullptr,
};
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 = ggml_aligned_malloc(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);
}
namespace ggml::cpu::amx {
class extra_buffer_type : ggml::cpu::extra_buffer_type {
bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
if (op->op != GGML_OP_MUL_MAT) {
return false;
}
auto * src0 = op->src[0];
auto * src1 = op->src[1];
if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
return false;
}
if (!src0->buffer || src0->buffer->buft != ggml_backend_amx_buffer_type()) {
return false;
}
if (src1->buffer && !ggml_backend_buft_is_host(src1->buffer->buft)) {
return false;
}
if (op->ne[0] % (TILE_N * 2)) {
return false;
}
int alignment;
switch (src0->type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q8_0:
alignment = TILE_K;
break;
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
case GGML_TYPE_IQ4_XS:
alignment = 256; // QK_K
break;
case GGML_TYPE_F16:
alignment = 16;
break;
default:
return false;
}
if (src0->ne[0] % alignment) {
return false;
}
if (src1->type != GGML_TYPE_F32) {
return false;
}
return true;
}
ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
if (op->op == GGML_OP_MUL_MAT && op->src[0]->buffer &&
op->src[0]->buffer->buft == ggml_backend_amx_buffer_type()) {
return (ggml::cpu::tensor_traits *) op->src[0]->extra;
}
return nullptr;
}
};
} // namespace ggml::cpu::amx
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);
}
#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(__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;
#else
return false;
#endif
}
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 = */ nullptr, // defaults to SIZE_MAX
/* .get_alloc_size = */ ggml_backend_amx_buffer_type_get_alloc_size,
/* .is_host = */ nullptr,
},
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ new ggml::cpu::amx::extra_buffer_type(),
};
if (!ggml_amx_init()) {
return nullptr;
}
return &ggml_backend_buffer_type_amx;
}
#endif // defined(__AMX_INT8__) && defined(__AVX512VNNI__)
engine/ggml/src/ggml-cpu/amx/amx.h
+8
View File
@@ -0,0 +1,8 @@
#include "ggml-backend.h"
#include "ggml-cpu-impl.h"
// GGML internal header
#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
#endif
engine/ggml/src/ggml-cpu/amx/common.h
+115
View File
@@ -0,0 +1,115 @@
#pragma once
#include "ggml.h"
#include "ggml-cpu-impl.h"
#include <algorithm>
#include <memory>
#include <type_traits>
#if defined(GGML_USE_OPENMP)
#include <omp.h>
#else
#include <thread>
#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 n, const func_t & f) {
if (n <= 0) {
return;
}
#if defined(GGML_USE_OPENMP)
#pragma omp parallel
{
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
int nth = std::thread::hardware_concurrency();
if (nth <= 1) {
f(0, n);
return;
}
if (nth > n) {
nth = n;
}
std::vector<std::thread> threads;
threads.reserve(nth);
for (int ith = 0; ith < nth; ++ith) {
threads.emplace_back([&f, n, ith, nth] {
int tbegin, tend;
balance211(n, nth, ith, tbegin, tend);
f(tbegin, tend);
});
}
for (auto & t : threads) {
t.join();
}
#endif
}
template <typename func_t>
inline void parallel_for_ggml(const ggml_compute_params * params, int n, const func_t & f) {
int tbegin, tend;
balance211(n, params->nth, params->ith, tbegin, tend);
f(tbegin, tend);
}
// 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);
}
engine/ggml/src/ggml-cpu/amx/mmq.cpp
+2512
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/amx/mmq.h
+10
View File
@@ -0,0 +1,10 @@
#pragma once
#include "common.h"
size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst);
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(const struct ggml_compute_params * params, struct ggml_tensor * dst);
engine/ggml/src/ggml-cpu/arch-fallback.h
+347
View File
@@ -0,0 +1,347 @@
#pragma once
// Rename `_generic` functions if no native implementation is available.
// This effectively selects the generic implementation.
#if defined(GGML_CPU_GENERIC)
// quants.c
#define quantize_row_q8_0_generic quantize_row_q8_0
#define quantize_row_q8_1_generic quantize_row_q8_1
#define quantize_row_q8_K_generic quantize_row_q8_K
#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
#elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
// repack.cpp
#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
// quants.c
#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
#elif defined(__POWERPC__) || defined(__powerpc__)
// ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679
// quants.c
#define quantize_row_q8_K_generic quantize_row_q8_K
#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
#elif defined(__loongarch64)
// quants.c
#define quantize_row_q8_K_generic quantize_row_q8_K
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
#elif defined(__riscv)
// quants.c
#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x1_generic ggml_quantize_mat_q8_0_4x1
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_K_4x1_generic ggml_quantize_mat_q8_K_4x1
#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
#elif defined(__s390x__)
// quants.c
#define quantize_row_q8_K_generic quantize_row_q8_K
#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
#elif defined(__wasm__)
// quants.c
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
#endif
engine/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp
+98
View File
@@ -0,0 +1,98 @@
#include "ggml-backend-impl.h"
#if defined(__aarch64__)
#if defined(__linux__)
#include <sys/auxv.h>
#elif defined(__APPLE__)
#include <sys/sysctl.h>
#endif
#if !defined(HWCAP2_SVE2)
#define HWCAP2_SVE2 (1 << 1)
#endif
#if !defined(HWCAP2_I8MM)
#define HWCAP2_I8MM (1 << 13)
#endif
#if !defined(HWCAP2_SME)
#define HWCAP2_SME (1 << 23)
#endif
struct aarch64_features {
// has_neon not needed, aarch64 has NEON guaranteed
bool has_dotprod = false;
bool has_fp16_va = false;
bool has_sve = false;
bool has_sve2 = false;
bool has_i8mm = false;
bool has_sme = false;
aarch64_features() {
#if defined(__linux__)
uint32_t hwcap = getauxval(AT_HWCAP);
uint32_t hwcap2 = getauxval(AT_HWCAP2);
has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
has_fp16_va = !!(hwcap & HWCAP_FPHP);
has_sve = !!(hwcap & HWCAP_SVE);
has_sve2 = !!(hwcap2 & HWCAP2_SVE2);
has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
has_sme = !!(hwcap2 & HWCAP2_SME);
#elif defined(__APPLE__)
int oldp = 0;
size_t size = sizeof(oldp);
if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) == 0) {
has_dotprod = static_cast<bool>(oldp);
}
if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) == 0) {
has_i8mm = static_cast<bool>(oldp);
}
if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) == 0) {
has_sme = static_cast<bool>(oldp);
}
// Apple apparently does not implement SVE yet
#endif
}
};
static int ggml_backend_cpu_aarch64_score() {
int score = 1;
aarch64_features af;
#ifdef GGML_USE_DOTPROD
if (!af.has_dotprod) { return 0; }
score += 1<<1;
#endif
#ifdef GGML_USE_FP16_VECTOR_ARITHMETIC
if (!af.has_fp16_va) { return 0; }
score += 1<<2;
#endif
#ifdef GGML_USE_SVE
if (!af.has_sve) { return 0; }
score += 1<<3;
#endif
#ifdef GGML_USE_MATMUL_INT8
if (!af.has_i8mm) { return 0; }
score += 1<<4;
#endif
#ifdef GGML_USE_SVE2
if (!af.has_sve2) { return 0; }
score += 1<<5;
#endif
#ifdef GGML_USE_SME
if (!af.has_sme) { return 0; }
score += 1<<6;
#endif
return score;
}
GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_aarch64_score)
# endif // defined(__aarch64__)
engine/ggml/src/ggml-cpu/arch/arm/quants.c
+4245
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/arch/arm/repack.cpp
+5156
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/arch/loongarch/quants.c
+2309
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp
+82
View File
@@ -0,0 +1,82 @@
# include "ggml-backend-impl.h"
#if defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)
#if defined(__linux__)
#include <sys/auxv.h>
#endif
#include <string>
struct powerpc_features {
std::string platform = "";
int power_version = -1;
bool has_vsx = false;
powerpc_features() {
#if defined(__linux__)
unsigned long auxval = getauxval(AT_PLATFORM);
if (auxval) {
platform = std::string(reinterpret_cast<const char*>(auxval));
// TBD: Do systems exist that return this in uppercase?
if (platform.substr(0, 5) == "power") {
// Extractt a numeric suffix, if one exists
int vpos = -1;
for (int i = platform.length() - 1; i >= 0; i--) {
if (std::isdigit(platform[i])) {
vpos = i;
} else {
break;
}
}
if (vpos > -1) {
power_version = std::stoi(platform.substr(vpos));
}
}
}
#endif
if (power_version >= 9) {
has_vsx = true;
}
}
};
static int ggml_backend_cpu_powerpc_score() {
int score = 1;
powerpc_features pf;
// Platform scores
#if defined(GGML_USE_POWER7)
if (pf.power_version < 7) { return 0; }
score += 1<<1;
#endif
#if defined(GGML_USE_POWER8)
if (pf.power_version < 8) { return 0; }
score += 1<<2;
#endif
#if defined(GGML_USE_POWER9)
if (pf.power_version < 9) { return 0; }
score += 1<<3;
#endif
#if defined(GGML_USE_POWER10)
if (pf.power_version < 10) { return 0; }
score += 1<<4;
#endif
#if defined(GGML_USE_POWER11)
if (pf.power_version < 11) { return 0; }
score += 1<<5;
#endif
// Feature scores
#if defined(GGML_USE_VSX)
if (!pf.has_vsx) { return 0; }
score += 1<<6;
#endif
return score;
}
GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_powerpc_score)
#endif // defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)
engine/ggml/src/ggml-cpu/arch/powerpc/quants.c
+2304
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/arch/riscv/cpu-feats.cpp
+38
View File
@@ -0,0 +1,38 @@
#include "ggml-backend-impl.h"
#if defined(__riscv) && __riscv_xlen == 64
#include <asm/hwprobe.h>
#include <asm/unistd.h>
#include <unistd.h>
struct riscv64_features {
bool has_rvv = false;
riscv64_features() {
struct riscv_hwprobe probe;
probe.key = RISCV_HWPROBE_KEY_IMA_EXT_0;
probe.value = 0;
int ret = syscall(__NR_riscv_hwprobe, &probe, 1, 0, NULL, 0);
if (0 == ret) {
has_rvv = !!(probe.value & RISCV_HWPROBE_IMA_V);
}
}
};
static int ggml_backend_cpu_riscv64_score() {
int score = 1;
riscv64_features rf;
#ifdef GGML_USE_RVV
if (!rf.has_rvv) { return 0; }
score += 1 << 1;
#endif
return score;
}
GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_riscv64_score)
#endif // __riscv && __riscv_xlen == 64
engine/ggml/src/ggml-cpu/arch/riscv/quants.c
+6596
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/arch/riscv/repack.cpp
+1703
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/arch/s390/cpu-feats.cpp
+50
View File
@@ -0,0 +1,50 @@
#include "ggml-backend-impl.h"
#if defined(__s390x__)
#include <sys/auxv.h>
// find hwcap bits in asm/elf.h
#ifndef HWCAP_VXRS_EXT2
#define HWCAP_VXRS_EXT2 (1 << 15)
#endif
#ifndef HWCAP_NNPA
#define HWCAP_NNPA (1 << 20)
#endif
struct s390x_features {
bool has_vxe2 = false;
bool has_nnpa = false;
s390x_features() {
uint32_t hwcap = getauxval(AT_HWCAP);
// NOTE: use hwcap2 with DFLT for z17 and later
// uint32_t hwcap2 = getauxval(AT_HWCAP2);
has_vxe2 = !!(hwcap & HWCAP_VXRS_EXT2);
has_nnpa = !!(hwcap & HWCAP_NNPA);
}
};
static int ggml_backend_cpu_s390x_score() {
int score = 1;
s390x_features sf;
// IBM z15 / LinuxONE 3
#ifdef GGML_USE_VXE2
if (!sf.has_vxe2) { return 0; }
score += 1 << 1;
#endif
// IBM z16 / LinuxONE 4 and z17 / LinuxONE 5
#ifdef GGML_USE_NNPA
if (!sf.has_nnpa) { return 0; }
score += 1 << 2;
#endif
return score;
}
GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_s390x_score)
#endif // __s390x__
engine/ggml/src/ggml-cpu/arch/s390/quants.c
+1465
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/arch/wasm/quants.c
+1292
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/arch/x86/cpu-feats.cpp
+327
View File
@@ -0,0 +1,327 @@
#include "ggml-backend-impl.h"
#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))
#ifdef _MSC_VER
#include <intrin.h>
#endif
#include <cstring>
#include <vector>
#include <bitset>
#include <array>
#include <string>
// ref: https://cdrdv2-public.intel.com/782156/325383-sdm-vol-2abcd.pdf
struct cpuid_x86 {
bool SSE3(void) { return f_1_ecx[0]; }
bool PCLMULQDQ(void) { return f_1_ecx[1]; }
bool MONITOR(void) { return f_1_ecx[3]; }
bool SSSE3(void) { return f_1_ecx[9]; }
bool FMA(void) { return f_1_ecx[12]; }
bool CMPXCHG16B(void) { return f_1_ecx[13]; }
bool SSE41(void) { return f_1_ecx[19]; }
bool SSE42(void) { return f_1_ecx[20]; }
bool MOVBE(void) { return f_1_ecx[22]; }
bool POPCNT(void) { return f_1_ecx[23]; }
bool AES(void) { return f_1_ecx[25]; }
bool XSAVE(void) { return f_1_ecx[26]; }
bool OSXSAVE(void) { return f_1_ecx[27]; }
bool AVX(void) { return f_1_ecx[28]; }
bool F16C(void) { return f_1_ecx[29]; }
bool RDRAND(void) { return f_1_ecx[30]; }
bool MSR(void) { return f_1_edx[5]; }
bool CX8(void) { return f_1_edx[8]; }
bool SEP(void) { return f_1_edx[11]; }
bool CMOV(void) { return f_1_edx[15]; }
bool CLFSH(void) { return f_1_edx[19]; }
bool MMX(void) { return f_1_edx[23]; }
bool FXSR(void) { return f_1_edx[24]; }
bool SSE(void) { return f_1_edx[25]; }
bool SSE2(void) { return f_1_edx[26]; }
bool FSGSBASE(void) { return f_7_ebx[0]; }
bool BMI1(void) { return f_7_ebx[3]; }
bool HLE(void) { return is_intel && f_7_ebx[4]; }
bool AVX2(void) { return f_7_ebx[5]; }
bool BMI2(void) { return f_7_ebx[8]; }
bool ERMS(void) { return f_7_ebx[9]; }
bool INVPCID(void) { return f_7_ebx[10]; }
bool RTM(void) { return is_intel && f_7_ebx[11]; }
bool AVX512F(void) { return f_7_ebx[16]; }
bool AVX512DQ(void) { return f_7_ebx[17]; }
bool RDSEED(void) { return f_7_ebx[18]; }
bool ADX(void) { return f_7_ebx[19]; }
bool AVX512PF(void) { return f_7_ebx[26]; }
bool AVX512ER(void) { return f_7_ebx[27]; }
bool AVX512CD(void) { return f_7_ebx[28]; }
bool AVX512BW(void) { return f_7_ebx[30]; }
bool AVX512VL(void) { return f_7_ebx[31]; }
bool SHA(void) { return f_7_ebx[29]; }
bool PREFETCHWT1(void) { return f_7_ecx[0]; }
bool LAHF(void) { return f_81_ecx[0]; }
bool LZCNT(void) { return is_intel && f_81_ecx[5]; }
bool ABM(void) { return is_amd && f_81_ecx[5]; }
bool SSE4a(void) { return is_amd && f_81_ecx[6]; }
bool XOP(void) { return is_amd && f_81_ecx[11]; }
bool TBM(void) { return is_amd && f_81_ecx[21]; }
bool SYSCALL(void) { return is_intel && f_81_edx[11]; }
bool MMXEXT(void) { return is_amd && f_81_edx[22]; }
bool RDTSCP(void) { return is_intel && f_81_edx[27]; }
bool _3DNOWEXT(void) { return is_amd && f_81_edx[30]; }
bool _3DNOW(void) { return is_amd && f_81_edx[31]; }
bool AVX512_VBMI(void) { return f_7_ecx[1]; }
bool AVX512_VNNI(void) { return f_7_ecx[11]; }
bool AVX512_FP16(void) { return f_7_edx[23]; }
bool AVX512_BF16(void) { return f_7_1_eax[5]; }
bool AVX_VNNI(void) { return f_7_1_eax[4]; }
bool AMX_TILE(void) { return f_7_edx[24]; }
bool AMX_INT8(void) { return f_7_edx[25]; }
bool AMX_FP16(void) { return f_7_1_eax[21]; }
bool AMX_BF16(void) { return f_7_edx[22]; }
#ifdef _MSC_VER
static void cpuid(int cpu_info[4], int eax) {
__cpuid(cpu_info, eax);
}
static void cpuidex(int cpu_info[4], int eax, int ecx) {
__cpuidex(cpu_info, eax, ecx);
}
#else
static void cpuid(int cpu_info[4], int eax) {
__asm__ __volatile__(
"cpuid"
: "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
: "a"(eax), "c"(0));
}
static void cpuidex(int cpu_info[4], int eax, int ecx) {
__asm__ __volatile__(
"cpuid"
: "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
: "a"(eax), "c"(ecx));
}
#endif
cpuid_x86() {
std::array<int, 4> cpui;
std::vector<std::array<int, 4>> data;
// calling __cpuid with 0x0 as the function_id argument
// gets the number of the highest valid function ID.
cpuid(cpui.data(), 0);
int n_ids = cpui[0];
for (int i = 0; i <= n_ids; ++i) {
cpuidex(cpui.data(), i, 0);
data.push_back(cpui);
}
// capture vendor string
char vendor[0x20] = {};
*reinterpret_cast<int *>(vendor) = data[0][1];
*reinterpret_cast<int *>(vendor + 4) = data[0][3];
*reinterpret_cast<int *>(vendor + 8) = data[0][2];
this->vendor = vendor;
if (this->vendor == "GenuineIntel") {
is_intel = true;
} else if (this->vendor == "AuthenticAMD") {
is_amd = true;
}
// load bitset with flags for function 0x00000001
if (n_ids >= 1) {
f_1_ecx = data[1][2];
f_1_edx = data[1][3];
}
// load bitset with flags for function 0x00000007
if (n_ids >= 7) {
f_7_ebx = data[7][1];
f_7_ecx = data[7][2];
f_7_edx = data[7][3];
cpuidex(cpui.data(), 7, 1);
f_7_1_eax = cpui[0];
}
// calling __cpuid with 0x80000000 as the function_id argument
// gets the number of the highest valid extended ID.
cpuid(cpui.data(), 0x80000000);
unsigned int n_ex_ids = cpui[0];
std::vector<std::array<int, 4>> ext_data;
for (unsigned int i = 0x80000000; i <= n_ex_ids; ++i) {
cpuidex(cpui.data(), i, 0);
ext_data.push_back(cpui);
}
// load bitset with flags for function 0x80000001
if (n_ex_ids >= 0x80000001) {
f_81_ecx = ext_data[1][2];
f_81_edx = ext_data[1][3];
}
// interpret CPU brand string if reported
char brand[0x40] = {};
if (n_ex_ids >= 0x80000004) {
std::memcpy(brand, ext_data[2].data(), sizeof(cpui));
std::memcpy(brand + 16, ext_data[3].data(), sizeof(cpui));
std::memcpy(brand + 32, ext_data[4].data(), sizeof(cpui));
this->brand = brand;
}
}
bool is_intel = false;
bool is_amd = false;
std::string vendor;
std::string brand;
std::bitset<32> f_1_ecx;
std::bitset<32> f_1_edx;
std::bitset<32> f_7_ebx;
std::bitset<32> f_7_ecx;
std::bitset<32> f_7_edx;
std::bitset<32> f_7_1_eax;
std::bitset<32> f_81_ecx;
std::bitset<32> f_81_edx;
};
#if 0
void test_x86_is() {
cpuid_x86 is;
printf("CPU Vendor: %s\n", is.vendor.c_str());
printf("Brand: %s\n", is.brand.c_str());
printf("is_intel: %d\n", is.is_intel);
printf("is_amd: %d\n", is.is_amd);
printf("sse3: %d\n", is.SSE3());
printf("pclmulqdq: %d\n", is.PCLMULQDQ());
printf("ssse3: %d\n", is.SSSE3());
printf("fma: %d\n", is.FMA());
printf("cmpxchg16b: %d\n", is.CMPXCHG16B());
printf("sse41: %d\n", is.SSE41());
printf("sse42: %d\n", is.SSE42());
printf("movbe: %d\n", is.MOVBE());
printf("popcnt: %d\n", is.POPCNT());
printf("aes: %d\n", is.AES());
printf("xsave: %d\n", is.XSAVE());
printf("osxsave: %d\n", is.OSXSAVE());
printf("avx: %d\n", is.AVX());
printf("f16c: %d\n", is.F16C());
printf("rdrand: %d\n", is.RDRAND());
printf("msr: %d\n", is.MSR());
printf("cx8: %d\n", is.CX8());
printf("sep: %d\n", is.SEP());
printf("cmov: %d\n", is.CMOV());
printf("clflush: %d\n", is.CLFSH());
printf("mmx: %d\n", is.MMX());
printf("fxsr: %d\n", is.FXSR());
printf("sse: %d\n", is.SSE());
printf("sse2: %d\n", is.SSE2());
printf("fsgsbase: %d\n", is.FSGSBASE());
printf("bmi1: %d\n", is.BMI1());
printf("hle: %d\n", is.HLE());
printf("avx2: %d\n", is.AVX2());
printf("bmi2: %d\n", is.BMI2());
printf("erms: %d\n", is.ERMS());
printf("invpcid: %d\n", is.INVPCID());
printf("rtm: %d\n", is.RTM());
printf("avx512f: %d\n", is.AVX512F());
printf("rdseed: %d\n", is.RDSEED());
printf("adx: %d\n", is.ADX());
printf("avx512pf: %d\n", is.AVX512PF());
printf("avx512er: %d\n", is.AVX512ER());
printf("avx512cd: %d\n", is.AVX512CD());
printf("sha: %d\n", is.SHA());
printf("prefetchwt1: %d\n", is.PREFETCHWT1());
printf("lahf: %d\n", is.LAHF());
printf("lzcnt: %d\n", is.LZCNT());
printf("abm: %d\n", is.ABM());
printf("sse4a: %d\n", is.SSE4a());
printf("xop: %d\n", is.XOP());
printf("tbm: %d\n", is.TBM());
printf("syscall: %d\n", is.SYSCALL());
printf("mmxext: %d\n", is.MMXEXT());
printf("rdtscp: %d\n", is.RDTSCP());
printf("3dnowext: %d\n", is._3DNOWEXT());
printf("3dnow: %d\n", is._3DNOW());
printf("avx512_vbmi: %d\n", is.AVX512_VBMI());
printf("avx512_vnni: %d\n", is.AVX512_VNNI());
printf("avx512_fp16: %d\n", is.AVX512_FP16());
printf("avx512_bf16: %d\n", is.AVX512_BF16());
printf("amx_tile: %d\n", is.AMX_TILE());
printf("amx_int8: %d\n", is.AMX_INT8());
printf("amx_fp16: %d\n", is.AMX_FP16());
printf("amx_bf16: %d\n", is.AMX_BF16());
}
#endif
static int ggml_backend_cpu_x86_score() {
// FIXME: this does not check for OS support
int score = 1;
cpuid_x86 is;
#ifdef GGML_FMA
if (!is.FMA()) { return 0; }
score += 1;
#endif
#ifdef GGML_F16C
if (!is.F16C()) { return 0; }
score += 1<<1;
#endif
#ifdef GGML_SSE42
if (!is.SSE42()) { return 0; }
score += 1<<2;
#endif
#ifdef GGML_BMI2
if (!is.BMI2()) { return 0; }
score += 1<<3;
#endif
#ifdef GGML_AVX
if (!is.AVX()) { return 0; }
score += 1<<4;
#endif
#ifdef GGML_AVX2
if (!is.AVX2()) { return 0; }
score += 1<<5;
#endif
#ifdef GGML_AVX_VNNI
if (!is.AVX_VNNI()) { return 0; }
score += 1<<6;
#endif
#ifdef GGML_AVX512
if (!is.AVX512F()) { return 0; }
if (!is.AVX512CD()) { return 0; }
if (!is.AVX512VL()) { return 0; }
if (!is.AVX512DQ()) { return 0; }
if (!is.AVX512BW()) { return 0; }
score += 1<<7;
#endif
#ifdef GGML_AVX512_VBMI
if (!is.AVX512_VBMI()) { return 0; }
score += 1<<8;
#endif
#ifdef GGML_AVX512_BF16
if (!is.AVX512_BF16()) { return 0; }
score += 1<<9;
#endif
#ifdef GGML_AVX512_VNNI
if (!is.AVX512_VNNI()) { return 0; }
score += 1<<10;
#endif
#ifdef GGML_AMX_INT8
if (!is.AMX_INT8()) { return 0; }
score += 1<<11;
#endif
return score;
}
GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_x86_score)
#endif // defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))
engine/ggml/src/ggml-cpu/arch/x86/quants.c
+3970
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/arch/x86/repack.cpp
+6407
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/binary-ops.cpp
+154
View File
@@ -0,0 +1,154 @@
#include "binary-ops.h"
#if defined(GGML_USE_ACCELERATE)
#include <Accelerate/Accelerate.h>
using vDSP_fn_t = void (*)(const float *, vDSP_Stride, const float *, vDSP_Stride, float *, vDSP_Stride, vDSP_Length);
#endif
static inline float op_add(float a, float b) {
return a + b;
}
static inline float op_sub(float a, float b) {
return a - b;
}
static inline float op_mul(float a, float b) {
return a * b;
}
static inline float op_div(float a, float b) {
return a / b;
}
template <float (*op)(float, float), typename src0_t, typename src1_t, typename dst_t>
static inline void vec_binary_op_contiguous(const int64_t n, dst_t * z, const src0_t * x, const src1_t * y) {
constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
constexpr auto src1_to_f32 = type_conversion_table<src1_t>::to_f32;
constexpr auto f32_to_dst = type_conversion_table<dst_t >::from_f32;
for (int i = 0; i < n; i++) {
z[i] = f32_to_dst(op(src0_to_f32(x[i]), src1_to_f32(y[i])));
}
}
template <float (*op)(float, float), typename src0_t, typename src1_t, typename dst_t>
static inline void vec_binary_op_non_contiguous(const int64_t n, const int64_t ne10, const int64_t nb10, dst_t * z, const src0_t * x, const src1_t * y) {
constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
constexpr auto src1_to_f32 = type_conversion_table<src1_t>::to_f32;
constexpr auto f32_to_dst = type_conversion_table<dst_t >::from_f32;
for (int i = 0; i < n; i++) {
int i10 = i % ne10;
const src1_t * y_ptr = (const src1_t *)((const char *)y + i10*nb10);
z[i] = f32_to_dst(op(src0_to_f32(x[i]), src1_to_f32(*y_ptr)));
}
}
template <float (*op)(float, float), typename src0_t, typename src1_t, typename dst_t>
static void apply_binary_op(const ggml_compute_params * params, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
GGML_TENSOR_BINARY_OP_LOCALS
GGML_ASSERT( nb0 == sizeof(dst_t));
GGML_ASSERT(nb00 == sizeof(src0_t));
const auto [ir0, ir1] = get_thread_range(params, src0);
const bool is_src1_contiguous_rows = ggml_is_contiguous_rows(src1);
#ifdef GGML_USE_ACCELERATE
vDSP_fn_t vDSP_op = nullptr;
// TODO - avoid the f32-only check using type 'trait' lookup tables and row-based src-to-float conversion functions
if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
if (op == op_add) {
vDSP_op = vDSP_vadd;
} else if (op == op_sub) {
vDSP_op = vDSP_vsub;
} else if (op == op_mul) {
vDSP_op = vDSP_vmul;
} else if (op == op_div) {
vDSP_op = vDSP_vdiv;
}
}
#endif
for (int64_t ir = ir0; ir < ir1; ++ir) {
const int64_t i03 = ir/(ne02*ne01);
const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
const int64_t i13 = i03 % ne13;
const int64_t i12 = i02 % ne12;
const int64_t i11 = i01 % ne11;
dst_t * dst_ptr = (dst_t *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 );
const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
const src1_t * src1_ptr = (const src1_t *) ((const char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
if (is_src1_contiguous_rows) {
// src1 is broadcastable across src0 and dst in i1, i2, i3
const int64_t nr0 = ne00 / ne10;
for (int64_t r = 0; r < nr0; ++r) {
#ifdef GGML_USE_ACCELERATE
if constexpr (std::is_same_v<src0_t, float> && std::is_same_v<src1_t, float> && std::is_same_v<dst_t, float>) {
if (vDSP_op != nullptr) {
vDSP_op(src1_ptr, 1, src0_ptr + r*ne10, 1, dst_ptr + r*ne10, 1, ne10);
continue;
}
}
#endif
vec_binary_op_contiguous<op>(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
}
} else {
vec_binary_op_non_contiguous<op>(ne0, ne10, nb10, dst_ptr, src0_ptr, src1_ptr);
}
}
}
// TODO: Use the 'traits' lookup table (for type conversion fns), instead of a mass of 'if' conditions with long templates
template <float (*op)(float, float)>
static void binary_op(const ggml_compute_params * params, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
/* */ if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { // all f32
apply_binary_op<op, float, float, float>(params, dst);
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { // all f16
apply_binary_op<op, ggml_fp16_t, ggml_fp16_t, ggml_fp16_t>(params, dst);
} else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16
apply_binary_op<op, ggml_bf16_t, ggml_bf16_t, ggml_bf16_t>(params, dst);
} else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_BF16) {
apply_binary_op<op, ggml_bf16_t, float, ggml_bf16_t>(params, dst);
} else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
apply_binary_op<op, ggml_bf16_t, float, float>(params, dst);
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
apply_binary_op<op, ggml_fp16_t, float, ggml_fp16_t>(params, dst);
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
apply_binary_op<op, ggml_fp16_t, float, float>(params, dst);
} else {
GGML_ABORT("%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
}
}
void ggml_compute_forward_add_non_quantized(const ggml_compute_params * params, ggml_tensor * dst) {
binary_op<op_add>(params, dst);
}
void ggml_compute_forward_sub(const ggml_compute_params * params, ggml_tensor * dst) {
binary_op<op_sub>(params, dst);
}
void ggml_compute_forward_mul(const ggml_compute_params * params, ggml_tensor * dst) {
binary_op<op_mul>(params, dst);
}
void ggml_compute_forward_div(const ggml_compute_params * params, ggml_tensor * dst) {
binary_op<op_div>(params, dst);
}
engine/ggml/src/ggml-cpu/binary-ops.h
+16
View File
@@ -0,0 +1,16 @@
#pragma once
#include "common.h"
#ifdef __cplusplus
extern "C" {
#endif
void ggml_compute_forward_add_non_quantized(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sub(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_mul(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_div(const struct ggml_compute_params * params, struct ggml_tensor * dst);
#ifdef __cplusplus
}
#endif
engine/ggml/src/ggml-cpu/cmake/FindSIMD.cmake
+100
View File
@@ -0,0 +1,100 @@
include(CheckCSourceRuns)
set(AVX_CODE "
#include <immintrin.h>
int main()
{
__m256 a;
a = _mm256_set1_ps(0);
return 0;
}
")
set(AVX512_CODE "
#include <immintrin.h>
int main()
{
__m512i a = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0);
__m512i b = a;
__mmask64 equality_mask = _mm512_cmp_epi8_mask(a, b, _MM_CMPINT_EQ);
return 0;
}
")
set(AVX2_CODE "
#include <immintrin.h>
int main()
{
__m256i a = {0};
a = _mm256_abs_epi16(a);
__m256i x;
_mm256_extract_epi64(x, 0); // we rely on this in our AVX2 code
return 0;
}
")
set(FMA_CODE "
#include <immintrin.h>
int main()
{
__m256 acc = _mm256_setzero_ps();
const __m256 d = _mm256_setzero_ps();
const __m256 p = _mm256_setzero_ps();
acc = _mm256_fmadd_ps( d, p, acc );
return 0;
}
")
macro(check_sse type flags)
set(__FLAG_I 1)
set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
foreach (__FLAG ${flags})
if (NOT ${type}_FOUND)
set(CMAKE_REQUIRED_FLAGS ${__FLAG})
check_c_source_runs("${${type}_CODE}" HAS_${type}_${__FLAG_I})
if (HAS_${type}_${__FLAG_I})
set(${type}_FOUND TRUE CACHE BOOL "${type} support")
set(${type}_FLAGS "${__FLAG}" CACHE STRING "${type} flags")
endif()
math(EXPR __FLAG_I "${__FLAG_I}+1")
endif()
endforeach()
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
if (NOT ${type}_FOUND)
set(${type}_FOUND FALSE CACHE BOOL "${type} support")
set(${type}_FLAGS "" CACHE STRING "${type} flags")
endif()
mark_as_advanced(${type}_FOUND ${type}_FLAGS)
endmacro()
# flags are for MSVC only!
check_sse("AVX" " ;/arch:AVX")
if (NOT ${AVX_FOUND})
set(GGML_AVX OFF)
else()
set(GGML_AVX ON)
endif()
check_sse("AVX2" " ;/arch:AVX2")
check_sse("FMA" " ;/arch:AVX2")
if ((NOT ${AVX2_FOUND}) OR (NOT ${FMA_FOUND}))
set(GGML_AVX2 OFF)
else()
set(GGML_AVX2 ON)
endif()
check_sse("AVX512" " ;/arch:AVX512")
if (NOT ${AVX512_FOUND})
set(GGML_AVX512 OFF)
else()
set(GGML_AVX512 ON)
endif()
engine/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake
+32
View File
@@ -0,0 +1,32 @@
include(CheckCSourceRuns)
if (CMAKE_SYSTEM_PROCESSOR MATCHES "^(riscv)" AND GGML_CPU_RISCV64_SPACEMIT)
set(SMT_MARCH_STR "-march=rv64gcv_zfh_zvfh_zba_zicbop")
if (CMAKE_C_COMPILER_ID STREQUAL "GNU" AND
CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 15)
string(APPEND SMT_MARCH_STR "_xsmtvdotii")
endif()
set(CMAKE_REQUIRED_FLAGS "${SMT_MARCH_STR}")
check_c_source_compiles("int main() {__asm__ volatile(\"vmadot v2, v0, v1\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_IME1)
check_c_source_compiles("int main() {__asm__ volatile(\"vmadot v2, v0, v1, i4\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VMADOT_S4)
check_c_source_compiles("int main() {__asm__ volatile(\"vmadot v2, v0, v1, i8\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VMADOT_S8)
check_c_source_compiles("int main() {__asm__ volatile(\"vfwmadot v2, v0, v1, fp16\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VFWMADOT_FP16)
check_c_source_compiles("int main() {__asm__ volatile(\"vmadot.hp v2, v0, v1, v0, 0, i4\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VFMADOT_S4)
check_c_source_compiles("int main() {__asm__ volatile(\"vmadot.hp v2, v0, v1, v0, 0, i8\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VFMADOT_S8)
check_c_source_compiles("int main() {__asm__ volatile(\"vmadot1 v2, v0, v1\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VMADOTN)
check_c_source_compiles("int main() {__asm__ volatile(\"vpack.vv v2, v0, v1, 2\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VPACK)
check_c_source_compiles("int main() {__asm__ volatile(\"vnspack.vv v2, v0, v1, 2\");}" SPACEMIT_RISCV_COMPILER_SUPPORT_VNPACK)
unset(CMAKE_REQUIRED_FLAGS)
list(APPEND RISCV64_SPACEMIT_IME_SPEC "")
if (SPACEMIT_RISCV_COMPILER_SUPPORT_IME1)
set(RISCV64_SPACEMIT_IME_SPEC "RISCV64_SPACEMIT_IME1")
endif()
if (SPACEMIT_RISCV_COMPILER_SUPPORT_VMADOT_S4 AND SPACEMIT_RISCV_COMPILER_SUPPORT_VPACK AND SPACEMIT_RISCV_COMPILER_SUPPORT_VNPACK)
list(APPEND RISCV64_SPACEMIT_IME_SPEC "RISCV64_SPACEMIT_IME2")
endif()
message("RISCV64_SPACEMIT_IME_SPEC: ${RISCV64_SPACEMIT_IME_SPEC}")
endif()
engine/ggml/src/ggml-cpu/common.h
+95
View File
@@ -0,0 +1,95 @@
#pragma once
#include "ggml.h"
#include "traits.h"
#include "ggml-cpu-impl.h"
#include "ggml-impl.h"
#include "simd-mappings.h"
#define GGML_FA_TILE_Q 64
#define GGML_FA_TILE_KV 64
#ifdef __cplusplus
#include <utility>
// convenience functions/macros for use in template calls
// note: these won't be required after the 'traits' lookup table is used.
static inline ggml_fp16_t f32_to_f16(float x) {
return GGML_CPU_FP32_TO_FP16(x);
}
static inline float f16_to_f32(ggml_fp16_t x) {
return GGML_CPU_FP16_TO_FP32(x);
}
static inline ggml_bf16_t f32_to_bf16(float x) {
return GGML_FP32_TO_BF16(x);
}
static inline float bf16_to_f32(ggml_bf16_t x) {
return GGML_BF16_TO_FP32(x);
}
static inline float i32_to_f32(int32_t x) {
return x;
}
static inline int32_t f32_to_i32(float x) {
return x;
}
static inline float f32_to_f32(float x) {
return x;
}
// TODO - merge this into the traits table, after using row-based conversions
template <class T>
struct type_conversion_table;
template <>
struct type_conversion_table<ggml_fp16_t> {
static constexpr float (*to_f32)(ggml_fp16_t) = f16_to_f32;
static constexpr ggml_fp16_t (*from_f32)(float) = f32_to_f16;
};
template <>
struct type_conversion_table<float> {
static constexpr float (*to_f32)(float) = f32_to_f32;
static constexpr float (*from_f32)(float) = f32_to_f32;
};
template <>
struct type_conversion_table<ggml_bf16_t> {
static constexpr float (*to_f32)(ggml_bf16_t) = bf16_to_f32;
static constexpr ggml_bf16_t (*from_f32)(float) = f32_to_bf16;
};
template <>
struct type_conversion_table<int32_t> {
static constexpr float (*to_f32)(int32_t) = i32_to_f32;
static constexpr int32_t (*from_f32)(float) = f32_to_i32;
};
static std::pair<int64_t, int64_t> get_thread_range(const struct ggml_compute_params * params, const struct ggml_tensor * src0) {
const int64_t ith = params->ith;
const int64_t nth = params->nth;
const int64_t nr = ggml_nrows(src0);
// rows per thread
const int64_t dr = (nr + nth - 1)/nth;
// row range for this thread
const int64_t ir0 = dr*ith;
const int64_t ir1 = MIN(ir0 + dr, nr);
return {ir0, ir1};
}
struct ggml_fa_tile_config {
static constexpr size_t Q = GGML_FA_TILE_Q;
static constexpr size_t KV = GGML_FA_TILE_KV;
};
#endif
engine/ggml/src/ggml-cpu/ggml-cpu-impl.h
+539
View File
@@ -0,0 +1,539 @@
#pragma once
// GGML CPU internal header
#include "ggml.h"
#include "ggml-impl.h"
#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
//#include <stddef.h>
#include <stdbool.h>
#include <string.h> // memcpy
#include <math.h> // fabsf
#ifdef __cplusplus
extern "C" {
#endif
struct ggml_compute_params {
// ith = thread index, nth = number of threads
int ith, nth;
// work buffer for all threads
size_t wsize;
void * wdata;
struct ggml_threadpool * threadpool;
// use reference implementation
bool use_ref;
};
#if defined(_MSC_VER)
#define m512bh(p) p
#define m512i(p) p
#else
#define m512bh(p) (__m512bh)(p)
#define m512i(p) (__m512i)(p)
#endif
// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
#ifndef __FMA__
#define __FMA__
#endif
#ifndef __F16C__
#define __F16C__
#endif
#endif
// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
#ifndef __SSE3__
#define __SSE3__
#endif
#ifndef __SSSE3__
#define __SSSE3__
#endif
#endif
#if defined(__s390x__) && defined(__VEC__)
#ifndef __VXE__
#define __VXE__
#endif // __VXE__
#ifndef __VXE2__
#define __VXE2__
#endif // __VXE2__
#endif // __s390x__ && __VEC__
#if defined(__ARM_FEATURE_SVE) && defined(__linux__)
#include <sys/prctl.h>
#endif
#if defined(__ARM_NEON)
// ref: https://github.com/ggml-org/llama.cpp/pull/5404
#ifdef _MSC_VER
#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
#else
#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
#endif // _MSC_VER
#if !defined(__aarch64__)
// 32-bit ARM compatibility
// vaddlvq_s16
// vpaddq_s16
// vpaddq_s32
// vaddvq_s32
// vaddvq_f32
// vmaxvq_f32
// vcvtnq_s32_f32
// vzip1_u8
// vzip2_u8
inline static int32_t vaddlvq_s16(int16x8_t v) {
int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
}
inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
return vcombine_s16(a0, b0);
}
inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
return vcombine_s32(a0, b0);
}
inline static int32_t vaddvq_s32(int32x4_t v) {
return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
}
inline static float vaddvq_f32(float32x4_t v) {
return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
}
inline static float vmaxvq_f32(float32x4_t v) {
return
MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
}
inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
int32x4_t res;
res[0] = roundf(vgetq_lane_f32(v, 0));
res[1] = roundf(vgetq_lane_f32(v, 1));
res[2] = roundf(vgetq_lane_f32(v, 2));
res[3] = roundf(vgetq_lane_f32(v, 3));
return res;
}
inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
uint8x8_t res;
res[0] = a[0]; res[1] = b[0];
res[2] = a[1]; res[3] = b[1];
res[4] = a[2]; res[5] = b[2];
res[6] = a[3]; res[7] = b[3];
return res;
}
inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
uint8x8_t res;
res[0] = a[4]; res[1] = b[4];
res[2] = a[5]; res[3] = b[5];
res[4] = a[6]; res[5] = b[6];
res[6] = a[7]; res[7] = b[7];
return res;
}
// vld1q_s16_x2
// vld1q_u8_x2
// vld1q_u8_x4
// vld1q_s8_x2
// vld1q_s8_x4
// TODO: double-check these work correctly
typedef struct ggml_int16x8x2_t {
int16x8_t val[2];
} ggml_int16x8x2_t;
inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
ggml_int16x8x2_t res;
res.val[0] = vld1q_s16(ptr + 0);
res.val[1] = vld1q_s16(ptr + 8);
return res;
}
typedef struct ggml_uint8x16x2_t {
uint8x16_t val[2];
} ggml_uint8x16x2_t;
inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
ggml_uint8x16x2_t res;
res.val[0] = vld1q_u8(ptr + 0);
res.val[1] = vld1q_u8(ptr + 16);
return res;
}
typedef struct ggml_uint8x16x4_t {
uint8x16_t val[4];
} ggml_uint8x16x4_t;
inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
ggml_uint8x16x4_t res;
res.val[0] = vld1q_u8(ptr + 0);
res.val[1] = vld1q_u8(ptr + 16);
res.val[2] = vld1q_u8(ptr + 32);
res.val[3] = vld1q_u8(ptr + 48);
return res;
}
typedef struct ggml_int8x16x2_t {
int8x16_t val[2];
} ggml_int8x16x2_t;
inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
ggml_int8x16x2_t res;
res.val[0] = vld1q_s8(ptr + 0);
res.val[1] = vld1q_s8(ptr + 16);
return res;
}
typedef struct ggml_int8x16x4_t {
int8x16_t val[4];
} ggml_int8x16x4_t;
inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
ggml_int8x16x4_t res;
res.val[0] = vld1q_s8(ptr + 0);
res.val[1] = vld1q_s8(ptr + 16);
res.val[2] = vld1q_s8(ptr + 32);
res.val[3] = vld1q_s8(ptr + 48);
return res;
}
// NOTE: not tested
inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
int8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
// NOTE: not tested
inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
uint8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
#else
#define ggml_int16x8x2_t int16x8x2_t
#define ggml_uint8x16x2_t uint8x16x2_t
#define ggml_uint8x16x4_t uint8x16x4_t
#define ggml_int8x16x2_t int8x16x2_t
#define ggml_int8x16x4_t int8x16x4_t
#define ggml_vld1q_s16_x2 vld1q_s16_x2
#define ggml_vld1q_u8_x2 vld1q_u8_x2
#define ggml_vld1q_u8_x4 vld1q_u8_x4
#define ggml_vld1q_s8_x2 vld1q_s8_x2
#define ggml_vld1q_s8_x4 vld1q_s8_x4
#define ggml_vqtbl1q_s8 vqtbl1q_s8
#define ggml_vqtbl1q_u8 vqtbl1q_u8
#endif // !defined(__aarch64__)
#if !defined(__ARM_FEATURE_DOTPROD)
// NOTE: this fallback produces the same total sum as native vdotq_s32 but with different per-lane grouping — do not use when individual lane values matter.
inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
}
#else
#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
#endif // !defined(__ARM_FEATURE_DOTPROD)
static inline int32x4_t ggml_nvfp4_dot8(const int8x8_t q4_lo, const int8x8_t q8_lo,
const int8x8_t q4_hi, const int8x8_t q8_hi) {
const int16x8_t p_lo = vmull_s8(q4_lo, q8_lo);
const int16x8_t p_hi = vmull_s8(q4_hi, q8_hi);
const int32x4_t sum_lo = vpaddlq_s16(p_lo);
const int32x4_t sum_hi = vpaddlq_s16(p_hi);
return vaddq_s32(sum_lo, sum_hi);
}
#endif // defined(__ARM_NEON)
#ifdef __wasm_simd128__
#include <wasm_simd128.h>
#endif
#ifdef __POWER9_VECTOR__
#include <altivec.h>
#endif
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <intrin.h>
#elif defined(__SSE__) || defined(__SSE3__) || defined(__SSSE3__) || defined(__AVX__) || defined(__F16C__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX512BF16__)
#include <immintrin.h>
#endif
#ifdef __riscv_v_intrinsic
#include <riscv_vector.h>
#endif
#if defined(__loongarch64)
#if defined(__loongarch_asx)
#include <lasxintrin.h>
#endif
#if defined(__loongarch_sx)
#include <lsxintrin.h>
#endif
#endif
#if defined(__VXE__) || defined(__VXE2__)
#include <vecintrin.h>
#define vec_neg(a) (-(a)) // Vector Negate
#define vec_add(a, b) ((a) + (b)) // Vector Add
#define vec_sub(a, b) ((a) - (b)) // Vector Subtract
#define vec_mul(a, b) ((a) * (b)) // Vector Multiply
#define vec_div(a, b) ((a) / (b)) // Vector Divide
#define vec_sl(a, b) ((a) << (b)) // Vector Shift Left
#define vec_sra(a, b) ((a) >> (b)) // Vector Shift Right
#define vec_sr(a, b) ((a) >> (b)) // Vector Shift Right Algebraic
#define vec_slo(a, b) vec_slb(a, (b) << 64) // Vector Shift Left by Octet
#define vec_sro(a, b) vec_srb(a, (b) << 64) // Vector Shift Right by Octet
#ifndef vec_and
#define vec_and(a, b) ((a) & (b)) // Vector AND
#endif
#ifndef vec_or
#define vec_or(a, b) ((a) | (b)) // Vector OR
#endif
#ifndef vec_xor
#define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
#endif
typedef signed char char8x16_t __attribute__((vector_size(16)));
typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
typedef int8_t int8x16_t __attribute__((vector_size(16)));
typedef int16_t int16x8_t __attribute__((vector_size(16)));
typedef int32_t int32x4_t __attribute__((vector_size(16)));
typedef uint8_t uint8x16_t __attribute__((vector_size(16)));
typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
typedef float float32x4_t __attribute__((vector_size(16)));
typedef double double64x2_t __attribute__((vector_size(16)));
typedef signed long long long64x2_t __attribute__((vector_size(16)));
typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
typedef struct ggml_uint8x16x2_t {
uint8x16_t val[2];
} ggml_uint8x16x2_t;
inline static ggml_uint8x16x2_t ggml_vec_xl_u8x2(const uint8_t * ptr) {
ggml_uint8x16x2_t res;
res.val[0] = vec_xl( 0, ptr);
res.val[1] = vec_xl(16, ptr);
return res;
}
typedef struct ggml_uint8x16x4_t {
uint8x16_t val[4];
} ggml_uint8x16x4_t;
inline static ggml_uint8x16x4_t ggml_vec_xl_u8x4(const uint8_t * ptr) {
ggml_uint8x16x4_t res;
res.val[0] = vec_xl( 0, ptr);
res.val[1] = vec_xl(16, ptr);
res.val[2] = vec_xl(32, ptr);
res.val[3] = vec_xl(48, ptr);
return res;
}
typedef struct ggml_int8x16x4_t {
int8x16_t val[4];
} ggml_int8x16x4_t;
inline static ggml_int8x16x4_t ggml_vec_xl_s8x4(const int8_t * ptr) {
ggml_int8x16x4_t res;
res.val[0] = vec_xl( 0, ptr);
res.val[1] = vec_xl(16, ptr);
res.val[2] = vec_xl(32, ptr);
res.val[3] = vec_xl(48, ptr);
return res;
}
typedef struct ggml_int16x8x2_t {
int16x8_t val[2];
} ggml_int16x8x2_t;
inline static ggml_int16x8x2_t ggml_vec_xl_s16x2(const int16_t * ptr) {
ggml_int16x8x2_t res;
res.val[0] = vec_xl( 0, ptr);
res.val[1] = vec_xl(16, ptr);
return res;
}
/*
! WARNING: Very slow. Use vec_perm if possible. Refer to iq4_xs
! or iq4_nl for example implementation.
*/
inline static int8x16_t ggml_vec_tbl(int8x16_t a, uint8x16_t b) {
int8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
inline static int16x8_t vec_padd_s16(int16x8_t a, int16x8_t b) {
const uchar8x16_t v_maske = { 0, 1, 4, 5, 8, 9, 12, 13,
16, 17, 20, 21, 24, 25, 28, 29 };
const int16x8_t v_abo = vec_pack((int32x4_t)a, (int32x4_t)b);
const int16x8_t v_abe = vec_perm(a, b, v_maske);
return v_abo + v_abe;
}
/**
* @see https://github.com/ggml-org/llama.cpp/pull/14037
*/
inline static float vec_hsum_f32x4(float32x4_t v) {
float32x4_t v_temp = v + vec_reve(v);
return v_temp[0] + v_temp[1];
}
inline static int32_t vec_hsum_i32x4(int32x4_t v) {
int32x4_t v_temp = v + vec_reve(v);
return v_temp[0] + v_temp[1];
}
inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
const int16x8_t p = vec_mule(a, b) + vec_mulo(a, b);
return acc + (vec_unpackh(p) + vec_unpackl(p));
}
#endif
#if defined(__loongarch_sx)
/* float type data load instructions */
static __m128 __lsx_vreplfr2vr_s(const float val) {
v4f32 res = {val, val, val, val};
return (__m128)res;
}
#endif
#if defined(__loongarch_asx)
static __m256 __lasx_xvreplfr2vr_s(const float val) {
v8f32 res = {val, val, val, val, val, val, val, val};
return (__m256)res;
}
#endif
// TODO: move to ggml-threading
void ggml_barrier(struct ggml_threadpool * tp);
void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value);
int ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value);
#ifdef __cplusplus
}
#endif
engine/ggml/src/ggml-cpu/ggml-cpu.c
+3887
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/ggml-cpu.cpp
+703
View File
@@ -0,0 +1,703 @@
#include "ggml-backend.h"
#include "ggml-backend-impl.h"
#include "ggml-cpu.h"
#include "repack.h"
#include "traits.h"
#include "ggml-impl.h"
#include "amx/amx.h"
#include <cctype>
#include <string>
#include <vector>
#ifdef GGML_USE_CPU_HBM
# include "hbm.h"
#endif
#ifdef GGML_USE_CPU_KLEIDIAI
# include "kleidiai/kleidiai.h"
#endif
#ifdef GGML_USE_CPU_RISCV64_SPACEMIT
# include "spacemit/ime.h"
#endif
#if defined(_WIN32)
# define WIN32_LEAN_AND_MEAN
# ifndef NOMINMAX
# define NOMINMAX
# endif
# include <windows.h>
#else
# include <unistd.h>
#endif
#if defined(__APPLE__)
# include <sys/sysctl.h>
# include <sys/types.h>
#endif
// ggml-backend interface
std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffer_types() {
static std::vector<ggml_backend_buffer_type_t> bufts = []() {
std::vector<ggml_backend_buffer_type_t> bufts;
#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
if (ggml_backend_amx_buffer_type()) {
bufts.push_back(ggml_backend_amx_buffer_type());
}
#endif
#ifdef GGML_USE_CPU_RISCV64_SPACEMIT
if (ggml_backend_cpu_riscv64_spacemit_buffer_type()) {
bufts.push_back(ggml_backend_cpu_riscv64_spacemit_buffer_type());
}
#endif
#ifdef GGML_USE_CPU_KLEIDIAI
if (ggml_backend_cpu_kleidiai_buffer_type()) {
bufts.push_back(ggml_backend_cpu_kleidiai_buffer_type());
}
#endif
#ifdef GGML_USE_CPU_REPACK
if (ggml_backend_cpu_repack_buffer_type()) {
bufts.push_back(ggml_backend_cpu_repack_buffer_type());
}
#endif
return bufts;
}();
return bufts;
}
static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_type(ggml_backend_dev_t device) {
static std::vector<ggml_backend_buffer_type_t> extra_bufts = [] {
std::vector<ggml_backend_buffer_type_t> bufts = ggml_backend_cpu_get_extra_buffer_types();
bufts.push_back(nullptr);
return bufts;
}();
return extra_bufts.data();
GGML_UNUSED(device);
}
static bool ggml_backend_cpu_is_extra_buffer_type(ggml_backend_buffer_type_t buft) {
for (auto * extra : ggml_backend_cpu_get_extra_buffer_types()) {
if (extra == buft) {
return true;
}
}
return false;
}
// CPU backend - backend (stream)
struct ggml_backend_cpu_context {
int n_threads;
ggml_threadpool_t threadpool;
uint8_t * work_data;
size_t work_size;
ggml_abort_callback abort_callback;
void * abort_callback_data;
bool use_ref; // use reference implementation
};
static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
return "CPU";
GGML_UNUSED(backend);
}
static void ggml_backend_cpu_free(ggml_backend_t backend) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
delete[] cpu_ctx->work_data;
delete cpu_ctx;
delete backend;
}
struct ggml_backend_plan_cpu {
struct ggml_cplan cplan;
struct ggml_cgraph cgraph;
};
static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;
cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
cpu_plan->cgraph = *cgraph; // FIXME: deep copy
if (cpu_plan->cplan.work_size > 0) {
cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
if (cpu_plan->cplan.work_data == NULL) {
delete cpu_plan;
return NULL;
}
}
cpu_plan->cplan.abort_callback = cpu_ctx->abort_callback;
cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
cpu_plan->cplan.use_ref = cpu_ctx->use_ref;
return cpu_plan;
}
static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
delete[] cpu_plan->cplan.work_data;
delete cpu_plan;
GGML_UNUSED(backend);
}
static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
GGML_UNUSED(backend);
}
static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
if (cpu_ctx->work_size < cplan.work_size) {
delete[] cpu_ctx->work_data;
cpu_ctx->work_data = new uint8_t[cplan.work_size];
if (cpu_ctx->work_data == NULL) {
cpu_ctx->work_size = 0;
return GGML_STATUS_ALLOC_FAILED;
}
cpu_ctx->work_size = cplan.work_size;
}
cplan.work_data = (uint8_t *)cpu_ctx->work_data;
cplan.abort_callback = cpu_ctx->abort_callback;
cplan.abort_callback_data = cpu_ctx->abort_callback_data;
cplan.use_ref = cpu_ctx->use_ref;
return ggml_graph_compute(cgraph, &cplan);
}
static const struct ggml_backend_i ggml_backend_cpu_i = {
/* .get_name = */ ggml_backend_cpu_get_name,
/* .free = */ ggml_backend_cpu_free,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .set_tensor_2d_async = */ NULL,
/* .get_tensor_2d_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create,
/* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free,
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ ggml_backend_cpu_graph_plan_compute,
/* .graph_compute = */ ggml_backend_cpu_graph_compute,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
/* .graph_optimize = */ NULL,
};
static ggml_guid_t ggml_backend_cpu_guid(void) {
static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
return &guid;
}
ggml_backend_t ggml_backend_cpu_init(void) {
// initialize CPU backend now to avoid slowing the first graph computation
ggml_cpu_init();
struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
if (ctx == NULL) {
return NULL;
}
ctx->n_threads = GGML_DEFAULT_N_THREADS;
ctx->threadpool = NULL;
ctx->work_data = NULL;
ctx->work_size = 0;
ctx->abort_callback = NULL;
ctx->abort_callback_data = NULL;
ctx->use_ref = false;
ggml_backend_t cpu_backend = new ggml_backend {
/* .guid = */ ggml_backend_cpu_guid(),
/* .iface = */ ggml_backend_cpu_i,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ ctx,
};
if (cpu_backend == NULL) {
delete ctx;
return NULL;
}
return cpu_backend;
}
bool ggml_backend_is_cpu(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
}
void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->n_threads = n_threads;
}
void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
if (ctx->threadpool && ctx->threadpool != threadpool) {
// already had a different threadpool, pause/suspend it before switching
ggml_threadpool_pause(ctx->threadpool);
}
ctx->threadpool = threadpool;
}
void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->abort_callback = abort_callback;
ctx->abort_callback_data = abort_callback_data;
}
void ggml_backend_cpu_set_use_ref(ggml_backend_t backend_cpu, bool use_ref) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->use_ref = use_ref;
}
// CPU backend - device
struct ggml_backend_cpu_device_context {
std::string description = "CPU";
ggml_backend_cpu_device_context() {
#ifdef __APPLE__
size_t len = 0;
if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) {
description.resize(len);
sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT
}
#elif defined(__linux__)
FILE * f = fopen("/proc/cpuinfo", "r");
if (f) {
char buf[1024];
while (fgets(buf, sizeof(buf), f)) {
if (strncmp(buf, "model name", 10) == 0) {
char * p = strchr(buf, ':');
if (p) {
p++;
while (std::isspace(*p)) {
p++;
}
while (std::isspace(p[strlen(p) - 1])) {
p[strlen(p) - 1] = '\0';
}
description = p;
break;
}
}
}
fclose(f);
}
#elif defined(_WIN32)
HKEY hKey;
if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
0,
KEY_READ,
&hKey) == ERROR_SUCCESS) {
DWORD cpu_brand_size = 0;
if (RegQueryValueExA(hKey,
"ProcessorNameString",
NULL,
NULL,
NULL,
&cpu_brand_size) == ERROR_SUCCESS) {
description.resize(cpu_brand_size);
if (RegQueryValueExA(hKey,
"ProcessorNameString",
NULL,
NULL,
(LPBYTE)&description[0], // NOLINT
&cpu_brand_size) == ERROR_SUCCESS) {
if (description.find('\0') != std::string::npos) {
description.resize(description.find('\0'));
}
}
}
RegCloseKey(hKey);
}
#endif
}
};
static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
return "CPU";
GGML_UNUSED(dev);
}
static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context;
return ctx->description.c_str();
}
static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
#ifdef _WIN32
MEMORYSTATUSEX status;
status.dwLength = sizeof(status);
GlobalMemoryStatusEx(&status);
*total = status.ullTotalPhys;
*free = status.ullAvailPhys;
#else
long pages = sysconf(_SC_PHYS_PAGES);
long page_size = sysconf(_SC_PAGE_SIZE);
*total = pages * page_size;
// "free" system memory is ill-defined, for practical purposes assume that all of it is free:
*free = *total;
#endif // _WIN32
GGML_UNUSED(dev);
}
static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) {
return GGML_BACKEND_DEVICE_TYPE_CPU;
GGML_UNUSED(dev);
}
static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
props->name = ggml_backend_cpu_device_get_name(dev);
props->description = ggml_backend_cpu_device_get_description(dev);
props->type = ggml_backend_cpu_device_get_type(dev);
ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
props->caps = {
/* .async = */ false,
/* .host_buffer = */ false,
/* .buffer_from_host_ptr = */ true,
/* .events = */ false,
};
}
static ggml_backend_t ggml_backend_cpu_device_init_backend(ggml_backend_dev_t dev, const char * params) {
return ggml_backend_cpu_init();
GGML_UNUSED(dev);
GGML_UNUSED(params);
}
static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_backend_dev_t dev) {
return ggml_backend_cpu_buffer_type();
GGML_UNUSED(dev);
}
static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
return ggml_backend_cpu_buffer_from_ptr(ptr, size);
GGML_UNUSED(dev);
GGML_UNUSED(max_tensor_size);
}
static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * src1 = op->src[1];
if (op->op == GGML_OP_NONE || op->op == GGML_OP_RESHAPE || op->op == GGML_OP_VIEW || op->op == GGML_OP_PERMUTE || op->op == GGML_OP_TRANSPOSE) {
return true;
}
// check extra buffer types
// note: only the first sources are checked for extra buffer types to reduce overhead, increase if necessary
for (int i = 0; i < 4; i++) {
if (op->src[i] && op->src[i]->buffer &&
ggml_backend_cpu_is_extra_buffer_type(op->src[i]->buffer->buft)) {
auto * buf_extra = (ggml::cpu::extra_buffer_type *) op->src[i]->buffer->buft->context;
return buf_extra->supports_op(dev, op);
}
}
switch (op->op) {
case GGML_OP_CPY:
case GGML_OP_SET_ROWS:
return
op->type != GGML_TYPE_IQ3_XXS &&
op->type != GGML_TYPE_IQ3_S &&
op->type != GGML_TYPE_IQ2_XXS &&
op->type != GGML_TYPE_IQ2_XS &&
op->type != GGML_TYPE_IQ2_S &&
op->type != GGML_TYPE_IQ1_S &&
op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
case GGML_OP_MUL_MAT:
return src1->type == GGML_TYPE_F32 || src1->type == ggml_get_type_traits_cpu(src0->type)->vec_dot_type;
case GGML_OP_SOFT_MAX_BACK: {
if (op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type != GGML_TYPE_F32) {
return false;
}
float max_bias = 0.0f;
memcpy(&max_bias, (const float *) op->op_params + 1, sizeof(float));
return max_bias == 0.0f;
}
case GGML_OP_IM2COL_BACK:
return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
case GGML_OP_GET_ROWS_BACK:
return src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16;
case GGML_OP_OUT_PROD:
return (src0->type == GGML_TYPE_F32 || (ggml_is_quantized(src0->type) && src0->ne[2] == src1->ne[2] && src0->ne[3] == src1->ne[3])) &&
src1->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
default:
return true;
}
}
static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
return ggml_backend_buft_is_host(buft) || ggml_backend_cpu_is_extra_buffer_type(buft);
GGML_UNUSED(dev);
}
static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
/* .get_name = */ ggml_backend_cpu_device_get_name,
/* .get_description = */ ggml_backend_cpu_device_get_description,
/* .get_memory = */ ggml_backend_cpu_device_get_memory,
/* .get_type = */ ggml_backend_cpu_device_get_type,
/* .get_props = */ ggml_backend_cpu_device_get_props,
/* .init_backend = */ ggml_backend_cpu_device_init_backend,
/* .get_buffer_type = */ ggml_backend_cpu_device_get_buffer_type,
/* .get_host_buffer_type = */ NULL,
/* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,
/* .supports_op = */ ggml_backend_cpu_device_supports_op,
/* .supports_buft = */ ggml_backend_cpu_device_supports_buft,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_synchronize = */ NULL,
};
// CPU backend - backend (reg)
static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
return "CPU";
GGML_UNUSED(reg);
}
static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
return 1;
GGML_UNUSED(reg);
}
static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
GGML_ASSERT(index == 0);
static ggml_backend_cpu_device_context ctx;
static ggml_backend_device ggml_backend_cpu_device = {
/* .iface = */ ggml_backend_cpu_device_i,
/* .reg = */ reg,
/* .context = */ &ctx,
};
return &ggml_backend_cpu_device;
}
// This is intended to replace the the ggml_cpu_has_* functions when loading the CPU backend dynamically,
// and additionally to allow other backends to expose their own list of features that applications can query using the same API
static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t reg) {
static std::vector<ggml_backend_feature> features = []() {
ggml_cpu_init();
std::vector<ggml_backend_feature> features;
if (ggml_cpu_has_sse3()) {
features.push_back({ "SSE3", "1" });
}
if (ggml_cpu_has_ssse3()) {
features.push_back({ "SSSE3", "1" });
}
if (ggml_cpu_has_avx()) {
features.push_back({ "AVX", "1" });
}
if (ggml_cpu_has_avx_vnni()) {
features.push_back({ "AVX_VNNI", "1" });
}
if (ggml_cpu_has_avx2()) {
features.push_back({ "AVX2", "1" });
}
if (ggml_cpu_has_f16c()) {
features.push_back({ "F16C", "1" });
}
if (ggml_cpu_has_fma()) {
features.push_back({ "FMA", "1" });
}
if (ggml_cpu_has_bmi2()) {
features.push_back({ "BMI2", "1" });
}
if (ggml_cpu_has_avx512()) {
features.push_back({ "AVX512", "1" });
}
if (ggml_cpu_has_avx512_vbmi()) {
features.push_back({ "AVX512_VBMI", "1" });
}
if (ggml_cpu_has_avx512_vnni()) {
features.push_back({ "AVX512_VNNI", "1" });
}
if (ggml_cpu_has_avx512_bf16()) {
features.push_back({ "AVX512_BF16", "1" });
}
if (ggml_cpu_has_amx_int8()) {
features.push_back({ "AMX_INT8", "1" });
}
if (ggml_cpu_has_neon()) {
features.push_back({ "NEON", "1" });
}
if (ggml_cpu_has_arm_fma()) {
features.push_back({ "ARM_FMA", "1" });
}
if (ggml_cpu_has_fp16_va()) {
features.push_back({ "FP16_VA", "1" });
}
if (ggml_cpu_has_matmul_int8()) {
features.push_back({ "MATMUL_INT8", "1" });
}
if (ggml_cpu_has_sve()) {
features.push_back({ "SVE", "1" });
}
if (ggml_cpu_has_dotprod()) {
features.push_back({ "DOTPROD", "1" });
}
if (ggml_cpu_get_sve_cnt() > 0) {
static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt());
features.push_back({ "SVE_CNT", sve_cnt.c_str() });
}
if (ggml_cpu_has_sme()) {
features.push_back({ "SME", "1" });
}
if (ggml_cpu_has_riscv_v()) {
features.push_back({ "RISCV_V", "1" });
}
if (ggml_cpu_get_rvv_vlen() > 0) {
static std::string rvv_vlen = std::to_string(ggml_cpu_get_rvv_vlen());
features.push_back({ "RVV_VLEN", rvv_vlen.c_str() });
}
if (ggml_cpu_has_vsx()) {
features.push_back({ "VSX", "1" });
}
if (ggml_cpu_has_vxe()) {
features.push_back({ "VXE", "1" });
}
if (ggml_cpu_has_wasm_simd()) {
features.push_back({ "WASM_SIMD", "1" });
}
if (ggml_cpu_has_llamafile()) {
features.push_back({ "LLAMAFILE", "1" });
}
#ifdef GGML_USE_ACCELERATE
features.push_back({ "ACCELERATE", "1" });
#endif
#ifdef GGML_USE_CPU_HBM
features.push_back({ "CPU_HBM", "1" });
#endif
#ifdef GGML_USE_OPENMP
features.push_back({ "OPENMP", "1" });
#endif
#ifdef GGML_USE_CPU_KLEIDIAI
features.push_back({ "KLEIDIAI", "1" });
#endif
#ifdef GGML_USE_CPU_REPACK
features.push_back({ "REPACK", "1" });
#endif
features.push_back({ nullptr, nullptr });
return features;
}();
return features.data();
GGML_UNUSED(reg);
}
static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) {
if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
ggml_backend_set_n_threads_t fct = ggml_backend_cpu_set_n_threads;
return (void *)fct;
}
if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
ggml_backend_dev_get_extra_bufts_t fct = ggml_backend_cpu_device_get_extra_buffers_type;
return (void *)fct;
}
if (strcmp(name, "ggml_backend_get_features") == 0) {
return (void *)ggml_backend_cpu_get_features;
}
if (strcmp(name, "ggml_backend_set_abort_callback") == 0) {
return (void *)ggml_backend_cpu_set_abort_callback;
}
if (strcmp(name, "ggml_backend_cpu_numa_init") == 0) {
return (void *)ggml_numa_init;
}
if (strcmp(name, "ggml_backend_cpu_is_numa") == 0) {
return (void *)ggml_is_numa;
}
if (strcmp(name, "ggml_backend_cpu_set_use_ref") == 0) {
return (void *)ggml_backend_cpu_set_use_ref;
}
// threadpool - TODO: move to ggml-base
if (strcmp(name, "ggml_threadpool_new") == 0) {
return (void *)ggml_threadpool_new;
}
if (strcmp(name, "ggml_threadpool_free") == 0) {
return (void *)ggml_threadpool_free;
}
if (strcmp(name, "ggml_backend_cpu_set_threadpool") == 0) {
return (void *)ggml_backend_cpu_set_threadpool;
}
return NULL;
GGML_UNUSED(reg);
}
static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
/* .get_name = */ ggml_backend_cpu_reg_get_name,
/* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
/* .get_device = */ ggml_backend_cpu_reg_get_device,
/* .get_proc_address = */ ggml_backend_cpu_get_proc_address,
};
ggml_backend_reg_t ggml_backend_cpu_reg(void) {
// init CPU feature detection
ggml_cpu_init();
static struct ggml_backend_reg ggml_backend_cpu_reg = {
/* .api_version = */ GGML_BACKEND_API_VERSION,
/* .iface = */ ggml_backend_cpu_reg_i,
/* .context = */ NULL,
};
return &ggml_backend_cpu_reg;
}
GGML_BACKEND_DL_IMPL(ggml_backend_cpu_reg)
engine/ggml/src/ggml-cpu/hbm.cpp
+55
View File
@@ -0,0 +1,55 @@
#ifdef GGML_USE_CPU_HBM
#include "ggml-backend.h"
#include "ggml-backend-impl.h"
#include "ggml-cpu.h"
#include "ggml-impl.h"
#include "hbm.h"
// buffer type HBM
#include <hbwmalloc.h>
static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_HBM";
GGML_UNUSED(buft);
}
static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
hbw_free(buffer->context);
}
static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
size_t size) {
void * ptr;
int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
if (result != 0) {
GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
return NULL;
}
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
buffer->buft = buft;
buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_hbm_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment,
/* .get_max_size = */ nullptr, // defaults to SIZE_MAX
/* .get_alloc_size = */ nullptr, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .context = */ nullptr,
};
return &ggml_backend_cpu_buffer_type_hbm;
}
#endif
engine/ggml/src/ggml-cpu/hbm.h
+8
View File
@@ -0,0 +1,8 @@
#pragma once
#include "ggml-backend.h"
#include "ggml.h"
// GGML CPU internal header
ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
engine/ggml/src/ggml-cpu/kleidiai/kernels.cpp
+939
View File
@@ -0,0 +1,939 @@
// SPDX-FileCopyrightText: Copyright 2025-2026 Arm Limited and/or its affiliates <open-source-office@arm.com>
// SPDX-License-Identifier: MIT
//
// KleidiAI micro-kernels
#include "kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_interface.h"
#include "kai_matmul_clamp_f32_qai8dxp_qsi8cxp_interface.h"
#include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h"
#include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h"
#include "kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.h"
#include "kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.h"
#include "kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.h"
#include "kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.h"
#include "kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.h"
#include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm.h"
#include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.h"
#include "kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.h"
#include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
#include "kai_lhs_quant_pack_qsi8d32p_f32.h"
#include "kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.h"
#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
#include "kai_lhs_quant_pack_qai8dxp_f32.h"
#include "kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.h"
#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
#include "kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.h"
#include "kai_lhs_pack_f16pmrx2_f32_neon.h"
#include "kai_common.h"
#include "simd-mappings.h"
#define GGML_COMMON_DECL_CPP
#include "ggml-common.h"
#include "kernels.h"
#define NELEMS(x) (sizeof(x) / sizeof(*x))
template<size_t(*Fn)(size_t,size_t,size_t)>
static inline size_t kernel_offs_fn3(size_t a, size_t b, size_t c) {
return Fn(a, b, c);
}
template<size_t(*Fn)(size_t,size_t)>
static inline size_t kernel_offs_fn2(size_t a, size_t b, size_t) {
return Fn(a, b);
}
template<void(*Fn)(size_t,size_t,size_t,size_t,const void*,const void*,float*,size_t,size_t,float,float)>
static inline void kernel_run_fn11(size_t m, size_t n, size_t k, size_t bl,
const void* lhs, const void* rhs, void* dst,
size_t dst_stride_row, size_t dst_stride_col,
float clamp_min, float clamp_max) {
Fn(m, n, k, bl, lhs, rhs, static_cast<float*>(dst), dst_stride_row, dst_stride_col, clamp_min, clamp_max);
}
template<void(*Fn)(size_t,size_t,size_t,const void*,const void*,void*,size_t,size_t,float,float)>
static inline void kernel_run_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
const void* lhs, const void* rhs, void* dst,
size_t dst_stride_row, size_t dst_stride_col,
float clamp_min, float clamp_max) {
Fn(m, n, k, lhs, rhs, dst, dst_stride_row, dst_stride_col, clamp_min, clamp_max);
}
template<void(*Fn)(size_t,size_t,size_t,const void*,const void*,float*,size_t,size_t,float,float)>
static inline void kernel_run_float_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
const void* lhs, const void* rhs, void* dst,
size_t dst_stride_row, size_t dst_stride_col,
float clamp_min, float clamp_max) {
Fn(m, n, k, lhs, rhs, static_cast<float*>(dst), dst_stride_row, dst_stride_col, clamp_min, clamp_max);
}
template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t)>
static inline size_t lhs_ps_fn6(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr) {
return Fn(m, k, bl, mr, kr, sr);
}
template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t)>
static inline size_t lhs_ps_fn5(size_t m, size_t k, size_t /*bl*/, size_t mr, size_t kr, size_t sr) {
return Fn(m, k, mr, kr, sr);
}
template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t)>
static inline size_t lhs_offs_fn6(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr) {
return Fn(m_idx, k, bl, mr, kr, sr);
}
template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t)>
static inline size_t lhs_offs_fn5(size_t m_idx, size_t k, size_t /*bl*/, size_t mr, size_t kr, size_t sr) {
return Fn(m_idx, k, mr, kr, sr);
}
template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,size_t,const float*,size_t,void*)>
static inline void lhs_pack_float_fn10(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr,
size_t m_idx_start, const void* lhs, size_t lhs_stride, void* lhs_packed) {
Fn(m, k, bl, mr, kr, sr, m_idx_start, static_cast<const float*>(lhs), lhs_stride, lhs_packed);
}
template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,size_t,const void*,size_t,void*)>
static inline void lhs_pack_void_fn10(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr,
size_t m_idx_start, const void* lhs, size_t lhs_stride, void* lhs_packed) {
Fn(m, k, bl, mr, kr, sr, m_idx_start, lhs, lhs_stride, lhs_packed);
}
template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,const void*,size_t,void*)>
static inline void lhs_pack_void_fn9(size_t m, size_t k, size_t /*bl*/, size_t mr, size_t kr, size_t sr,
size_t m_idx_start, const void* lhs, size_t lhs_stride, void* lhs_packed) {
Fn(m, k, mr, kr, sr, m_idx_start, lhs, lhs_stride, lhs_packed);
}
template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,const float*,size_t,void*)>
static inline void lhs_pack_float_fn9_no_bl(size_t m, size_t k, size_t /*bl*/, size_t mr, size_t kr, size_t sr,
size_t m_idx_start, const void * lhs, size_t lhs_stride, void * lhs_packed) {
Fn(m, k, mr, kr, sr, m_idx_start, static_cast<const float*>(lhs), lhs_stride, lhs_packed);
}
template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t)>
static inline size_t rhs_ps_fn5(size_t n, size_t k, size_t nr, size_t kr, size_t bl) {
return Fn(n, k, nr, kr, bl);
}
template<size_t(*Fn)(size_t,size_t)>
static inline size_t rhs_ps_fn2(size_t n, size_t k, size_t /*nr*/, size_t /*kr*/, size_t /*bl*/) {
return Fn(n, k);
}
template<size_t(*Fn)(size_t,size_t,size_t,size_t)>
static inline size_t rhs_stride_fn4(size_t k, size_t nr, size_t kr, size_t bl) {
return Fn(k, nr, kr, bl);
}
template<size_t(*Fn)(size_t)>
static inline size_t rhs_stride_fn1(size_t k, size_t /*nr*/, size_t /*kr*/, size_t /*bl*/) {
return Fn(k);
}
template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,size_t,const uint8_t*,const float*,void*,size_t,const struct kai_rhs_pack_qs4cxs1s0_param*)>
static inline void rhs_pack_fn12(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl,
size_t /*rhs_stride*/, const void* rhs, const void* bias, const void* /*scale*/,
void* rhs_packed, size_t extra_bytes, const void* params) {
Fn(num_groups, n, k, nr, kr, sr, bl,
static_cast<const uint8_t*>(rhs),
static_cast<const float*>(bias),
rhs_packed, extra_bytes,
static_cast<const kai_rhs_pack_qs4cxs1s0_param*>(params));
}
template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,const int8_t*,const float*,const float*,void*,size_t,const struct kai_rhs_pack_qsi8cx_params*)>
static inline void rhs_pack_scale_fn12(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t /*bl*/,
size_t /*rhs_stride*/, const void* rhs, const void* bias, const void* scale,
void* rhs_packed, size_t extra_bytes, const void* params) {
Fn(num_groups, n, k, nr, kr, sr,
static_cast<const int8_t*>(rhs),
static_cast<const float*>(bias),
static_cast<const float*>(scale),
rhs_packed, extra_bytes,
static_cast<const kai_rhs_pack_qsi8cx_params*>(params));
}
template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,size_t,const void*,const void*,const void*,void*,size_t,const void*)>
static inline void rhs_pack_fn13(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t /*bl*/,
size_t rhs_stride, const void* rhs, const void* bias, const void* scale,
void* rhs_packed, size_t extra_bytes, const void* params) {
Fn(num_groups, n, k, nr, kr, sr, rhs_stride, rhs, bias, scale, rhs_packed, extra_bytes, params);
}
static const size_t INT4_PER_BYTE = 2;
static const size_t INT4_BITS = 4;
static const int Q4_0_ZERO_POINT = 8;
const size_t INT4_PER_UINT16 = 4;
static void dequantize_row_qsi4c32pscalef16(
const void *packed_data,
int32_t row_idx,
int64_t nc,
float *out,
size_t nr_pack,
size_t packed_row_stride,
size_t kr,
size_t bl,
size_t num_bytes_multiplier
) {
size_t group_idx = row_idx / nr_pack;
size_t row_in_group = row_idx % nr_pack;
const uint8_t *packed_group = (const uint8_t *)packed_data + group_idx * packed_row_stride;
size_t num_blocks = nc / bl;
const uint8_t *block_ptr = packed_group;
for (size_t b = 0; b < num_blocks; ++b) {
uint16_t scale_f16 = *((const uint16_t *)(block_ptr + row_in_group * num_bytes_multiplier));
float scale = GGML_CPU_FP16_TO_FP32(scale_f16);
const uint8_t *segment_ptr = block_ptr + nr_pack * num_bytes_multiplier;
size_t num_segments = bl / kr;
size_t num_bytes_per_segment = kr / INT4_PER_BYTE;
for (size_t s = 0; s < num_segments; ++s) {
const uint8_t *seg_base = segment_ptr + s * nr_pack * num_bytes_per_segment;
const uint8_t *qbytes = seg_base + row_in_group * num_bytes_per_segment;
for (size_t k = 0; k < num_bytes_per_segment; ++k) {
uint8_t byte = qbytes[k] ^ 0x88;
int x0 = (byte & 0x0F) - Q4_0_ZERO_POINT;
int x1 = (byte >> INT4_BITS) - Q4_0_ZERO_POINT;
out[b * bl + s * num_bytes_per_segment + k] = x0 * scale;
out[b * bl + s * num_bytes_per_segment + k + bl/2] = x1 * scale;
}
}
block_ptr += nr_pack * num_bytes_multiplier + num_segments * nr_pack * num_bytes_per_segment;
}
}
static void dequantize_row_qsi4c32ps1s0scalef16(
const void *packed_data,
int32_t row_idx,
int64_t k,
float *out,
size_t nr,
size_t packed_row_stride,
size_t kr,
size_t bl,
size_t num_bytes_multiplier
) {
const size_t num_blocks = k / bl;
const size_t bl4 = bl / INT4_PER_UINT16;
size_t group_idx = row_idx / nr;
size_t row_in_group = row_idx % nr;
const uint8_t *packed_group = (const uint8_t *)packed_data + group_idx * packed_row_stride;
const uint16_t *qdata = (const uint16_t *)packed_group;
const uint16_t *scales = (const uint16_t *)(packed_group + packed_row_stride - (nr * num_blocks * num_bytes_multiplier));
for (size_t block_idx = 0; block_idx < num_blocks; ++block_idx) {
uint16_t scale_f16 = scales[row_in_group + block_idx * nr];
float scale = GGML_CPU_FP16_TO_FP32(scale_f16);
for (size_t bl4_idx = 0; bl4_idx < bl4; ++bl4_idx) {
uint16_t q = qdata[(block_idx * bl4 + bl4_idx) * nr + row_in_group];
for (size_t qidx = 0; qidx < INT4_PER_UINT16; ++qidx) {
int v = ((q >> (qidx * 4)) & 0xF) - Q4_0_ZERO_POINT;
out[block_idx * bl + bl4_idx * INT4_BITS + qidx] = v * scale;
}
}
}
GGML_UNUSED(kr);
}
static void dequantize_row_qsi8cxp(
const void *packed_data,
int32_t row_idx,
int64_t k,
float *out,
size_t nr,
size_t packed_row_stride,
size_t kr,
size_t bl,
size_t num_bytes_multiplier
) {
GGML_UNUSED(bl);
GGML_UNUSED(num_bytes_multiplier);
const size_t k_internal = ((size_t) k + QK8_0 - 1) / QK8_0 * QK8_0;
const size_t group_idx = row_idx / nr;
const size_t row_in_group = row_idx % nr;
const uint8_t * group_ptr = static_cast<const uint8_t *>(packed_data) + group_idx * packed_row_stride;
const int8_t * data_base = reinterpret_cast<const int8_t *>(group_ptr);
const size_t num_blocks = k_internal / kr;
for (size_t block = 0; block < num_blocks; ++block) {
const int8_t * block_ptr = data_base + (block * nr + row_in_group) * kr;
for (size_t i = 0; i < kr; ++i) {
const size_t k_idx = block * kr + i;
if (k_idx < (size_t) k) {
out[k_idx] = static_cast<float>(block_ptr[i]);
}
}
}
const uint8_t * sums_ptr = group_ptr + nr * k_internal;
GGML_UNUSED(sums_ptr);
const float * scale_ptr = reinterpret_cast<const float *>(sums_ptr + nr * sizeof(int32_t));
const float scale = scale_ptr[row_in_group];
if (scale == 0.0f) {
for (size_t i = 0; i < (size_t) k; ++i) {
out[i] = 0.0f;
}
return;
}
for (size_t i = 0; i < (size_t) k; ++i) {
out[i] *= scale;
}
}
static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
#if defined(__ARM_FEATURE_SME)
{
/* SME GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
/* .get_lhs_offset_ex = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa>,
/* .run_kernel_ex = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa>,
},
/* .gemm_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_pack_f16pmrx2_f32_neon,
/* .get_packed_offset_ex = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_pack_f16pmrx2_f32_neon>,
/* .packed_size_ex = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_pack_f16pmrx2_f32_neon>,
/* .pack_func_ex = */ &lhs_pack_void_fn10<kai_run_lhs_pack_f16pmrx2_f32_neon>,
},
/* SME GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
/* .get_lhs_offset_ex = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot>,
/* .run_kernel_ex = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot>,
},
/* .gemv_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32_neon,
/* .get_packed_offset_ex = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon>,
/* .packed_size_ex = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon>,
/* .pack_func_ex = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32_neon>,
},
/* .rhs_info = */ {
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
/* .to_float = */ dequantize_row_qsi4c32ps1s0scalef16,
/* .packed_size_ex = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon>,
/* .packed_stride_ex = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon>,
/* .pack_func_ex = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon>,
},
/* .required_cpu = */ CPU_FEATURE_SME,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
{
/* SME GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_lhs_offset_ex = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa>,
/* .run_kernel_ex = */ &kernel_run_fn10<kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa>,
},
/* .gemm_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_pack_bf16p2vlx2_f32_sme,
/* .get_packed_offset_ex = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme>,
/* .packed_size_ex = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme>,
/* .pack_func_ex = */ &lhs_pack_void_fn9<kai_run_lhs_pack_bf16p2vlx2_f32_sme>,
},
/* SME GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
/* .get_lhs_offset_ex = */ nullptr,
/* .get_rhs_packed_offset_ex = */ nullptr,
/* .run_kernel_ex = */ nullptr,
},
/* .gemv_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_pack_bf16p2vlx2_f32_sme,
/* .get_packed_offset_ex = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme>,
/* .packed_size_ex = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme>,
/* .pack_func_ex = */ &lhs_pack_void_fn9<kai_run_lhs_pack_bf16p2vlx2_f32_sme>,
},
/* .rhs_info = */ {
/* .packed_stride = */ nullptr,
/* .to_float = */ nullptr,
/* .packed_size_ex = */ &rhs_ps_fn2<kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme>,
/* .packed_stride_ex = */ &rhs_stride_fn1<kai_get_rhs_packed_stride_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme>,
/* .pack_func_ex = */ &rhs_pack_fn13<kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme>,
},
/* .required_cpu = */ CPU_FEATURE_SME,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_F16,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#if defined(__APPLE__)
#if defined(__ARM_FEATURE_DOTPROD)
{
/* DOTPROD GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_lhs_offset_ex = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
/* .run_kernel_ex = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
},
/* .gemm_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset_ex = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
/* .packed_size_ex = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
/* .pack_func_ex = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
},
/* DOTPROD GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_lhs_offset_ex = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
/* .run_kernel_ex = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
},
/* .gemv_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset_ex = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
/* .packed_size_ex = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
/* .pack_func_ex = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
},
/* .rhs_info = */ {
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .to_float = */ dequantize_row_qsi4c32pscalef16,
/* .packed_size_ex = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
/* .packed_stride_ex = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
/* .pack_func_ex = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#if defined(__ARM_FEATURE_MATMUL_INT8)
{
/* i8mm GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_lhs_offset_ex = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
/* .run_kernel_ex = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
},
/* .gemm_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
/* .get_packed_offset_ex = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
/* .packed_size_ex = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
/* .pack_func_ex = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
},
/* i8mm GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_lhs_offset_ex = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
/* .run_kernel_ex = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
},
/* .gemv_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset_ex = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
/* .packed_size_ex = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
/* .pack_func_ex = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
},
/* .rhs_info = */ {
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .to_float = */ dequantize_row_qsi4c32pscalef16,
/* .packed_size_ex = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
/* .packed_stride_ex = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
/* .pack_func_ex = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
},
/* .required_cpu = */ CPU_FEATURE_I8MM,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#else
#if defined(__ARM_FEATURE_SVE)
{
/* SVE i8mm GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
/* .get_lhs_offset_ex = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm>,
/* .run_kernel_ex = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm>,
},
/* .gemm_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
/* .get_packed_offset_ex = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
/* .packed_size_ex = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
/* .pack_func_ex = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
},
/* SVE dotprod GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
/* .get_lhs_offset_ex = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod>,
/* .run_kernel_ex = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod>,
},
/* .gemv_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset_ex = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
/* .packed_size_ex = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
/* .pack_func_ex = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
},
/* .rhs_info = */ {
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .to_float = */ dequantize_row_qsi4c32pscalef16,
/* .packed_size_ex = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
/* .packed_stride_ex = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
/* .pack_func_ex = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
},
/* .required_cpu = */ CPU_FEATURE_SVE | CPU_FEATURE_I8MM | CPU_FEATURE_DOTPROD,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#if defined(__ARM_FEATURE_MATMUL_INT8)
{
/* i8mm GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
/* .get_lhs_offset_ex = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
/* .run_kernel_ex = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
},
/* .gemm_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
/* .get_packed_offset_ex = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
/* .packed_size_ex = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
/* .pack_func_ex = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
},
/* i8mm GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
/* .get_lhs_offset_ex = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
/* .run_kernel_ex = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
},
/* .gemv_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset_ex = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
/* .packed_size_ex = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
/* .pack_func_ex = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
},
/* .rhs_info = */ {
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .to_float = */ dequantize_row_qsi4c32pscalef16,
/* .packed_size_ex = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
/* .packed_stride_ex = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
/* .pack_func_ex = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
},
/* .required_cpu = */ CPU_FEATURE_I8MM,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif // __ARM_FEATURE_MATMUL_INT8
#if defined(__ARM_FEATURE_DOTPROD)
{
/* DOTPROD GEMM */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
/* .get_lhs_offset_ex = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
/* .run_kernel_ex = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
},
/* .gemm_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset_ex = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
/* .packed_size_ex = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
/* .pack_func_ex = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
},
/* DOTPROD GEMV */
/* .kern_info = */ {
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
/* .get_lhs_offset_ex = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
/* .run_kernel_ex = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
},
/* .gemv_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
/* .get_packed_offset_ex = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
/* .packed_size_ex = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
/* .pack_func_ex = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
},
/* .rhs_info = */ {
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
/* .to_float = */ dequantize_row_qsi4c32pscalef16,
/* .packed_size_ex = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
/* .packed_stride_ex = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
/* .pack_func_ex = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q4_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#endif
{ /* Sentinel */ }
};
static ggml_kleidiai_kernels gemm_gemv_kernels_q8[] = {
#if defined(__ARM_FEATURE_SME)
{
/* SME GEMM */
{
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
/* .get_lhs_offset_ex = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa>,
/* .run_kernel_ex = */ &kernel_run_float_fn10<kai_run_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa>,
},
/* .gemm_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32,
/* .get_packed_offset_ex = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32>,
/* .packed_size_ex = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
/* .pack_func_ex = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
},
/* SME GEMV */
{
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
/* .get_lhs_offset_ex = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot>,
/* .run_kernel_ex = */ &kernel_run_float_fn10<kai_run_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot>,
},
/* .gemv_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32,
/* .get_packed_offset_ex = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32>,
/* .packed_size_ex = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
/* .pack_func_ex = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
},
/* .rhs_info = */ {
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon,
/* .to_float = */ dequantize_row_qsi8cxp,
/* .packed_size_ex = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
/* .packed_stride_ex = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
/* .pack_func_ex = */ &rhs_pack_scale_fn12<kai_run_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
},
/* .required_cpu = */ CPU_FEATURE_SME,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q8_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#if defined(__ARM_FEATURE_MATMUL_INT8)
{
/* I8MM GEMM */
{
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm,
/* .get_lhs_offset_ex = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm>,
/* .run_kernel_ex = */ &kernel_run_float_fn10<kai_run_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm>,
},
/* .gemm_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32,
/* .get_packed_offset_ex = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32>,
/* .packed_size_ex = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
/* .pack_func_ex = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
},
/* I8MM GEMV (dotprod fallback) */
{
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod,
/* .get_lhs_offset_ex = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod>,
/* .run_kernel_ex = */ &kernel_run_float_fn10<kai_run_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod>,
},
/* .gemv_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32,
/* .get_packed_offset_ex = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32>,
/* .packed_size_ex = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
/* .pack_func_ex = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
},
/* .rhs_info = */ {
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon,
/* .to_float = */ dequantize_row_qsi8cxp,
/* .packed_size_ex = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
/* .packed_stride_ex = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
/* .pack_func_ex = */ &rhs_pack_scale_fn12<kai_run_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
},
/* .required_cpu = */ CPU_FEATURE_I8MM,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q8_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
#if defined(__ARM_FEATURE_DOTPROD)
{
/* DOTPROD GEMM */
{
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod,
/* .get_lhs_offset_ex = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod>,
/* .run_kernel_ex = */ &kernel_run_float_fn10<kai_run_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod>,
},
/* .gemm_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32,
/* .get_packed_offset_ex = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32>,
/* .packed_size_ex = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
/* .pack_func_ex = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
},
/* DOTPROD GEMV */
{
/* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
/* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
/* .get_mr = */ kai_get_mr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
/* .get_nr = */ kai_get_nr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
/* .get_kr = */ kai_get_kr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
/* .get_sr = */ kai_get_sr_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
/* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
/* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod,
/* .get_lhs_offset_ex = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod>,
/* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod>,
/* .run_kernel_ex = */ &kernel_run_float_fn10<kai_run_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod>,
},
/* .gemv_lhs_info = */ {
/* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32,
/* .get_packed_offset_ex = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32>,
/* .packed_size_ex = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
/* .pack_func_ex = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
},
/* .rhs_info = */ {
/* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon,
/* .to_float = */ dequantize_row_qsi8cxp,
/* .packed_size_ex = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
/* .packed_stride_ex = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
/* .pack_func_ex = */ &rhs_pack_scale_fn12<kai_run_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
},
/* .required_cpu = */ CPU_FEATURE_DOTPROD,
/* .lhs_type = */ GGML_TYPE_F32,
/* .rhs_type = */ GGML_TYPE_Q8_0,
/* .op_type = */ GGML_TYPE_F32,
},
#endif
{ /* Sentinel */ }
};
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor) {
ggml_kleidiai_kernels * kernel = nullptr;
if (tensor->op == GGML_OP_MUL_MAT && tensor->src[0] != nullptr && tensor->src[1] != nullptr) {
#if defined(__ARM_FEATURE_SME) || \
defined(__ARM_FEATURE_DOTPROD) || \
defined(__ARM_FEATURE_MATMUL_INT8) || \
defined(__ARM_FEATURE_SVE)
auto try_table = [&](auto & table) {
for (size_t i = 0; i < NELEMS(table) - 1; ++i) {
if ((cpu_features & table[i].required_cpu) == table[i].required_cpu &&
table[i].lhs_type == tensor->src[1]->type &&
table[i].rhs_type == tensor->src[0]->type &&
table[i].op_type == tensor->type) {
kernel = &table[i];
return true;
}
}
return false;
};
if (tensor->src[0]->type == GGML_TYPE_Q8_0) {
try_table(gemm_gemv_kernels_q8);
} else {
try_table(gemm_gemv_kernels);
}
#else
GGML_UNUSED(gemm_gemv_kernels);
GGML_UNUSED(gemm_gemv_kernels_q8);
GGML_UNUSED(cpu_features);
#endif
}
return kernel;
}
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features) {
ggml_kleidiai_kernels * kernels = nullptr;
#if defined(__ARM_FEATURE_SME) || \
defined(__ARM_FEATURE_DOTPROD) || \
defined(__ARM_FEATURE_MATMUL_INT8) || \
defined(__ARM_FEATURE_SVE)
for (size_t i = 0; i < NELEMS(gemm_gemv_kernels) - 1; ++i) {
if ((features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu) {
kernels = &gemm_gemv_kernels[i];
break;
}
}
#else
GGML_UNUSED(features);
#endif
return kernels;
}
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q8_0(cpu_feature features) {
ggml_kleidiai_kernels * kernels = nullptr;
#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_DOTPROD) || defined(__ARM_FEATURE_MATMUL_INT8)
for (size_t i = 0; i < NELEMS(gemm_gemv_kernels_q8) - 1; ++i) {
if ((features & gemm_gemv_kernels_q8[i].required_cpu) == gemm_gemv_kernels_q8[i].required_cpu) {
kernels = &gemm_gemv_kernels_q8[i];
break;
}
}
#else
GGML_UNUSED(features);
#endif
return kernels;
}
engine/ggml/src/ggml-cpu/kleidiai/kernels.h
+90
View File
@@ -0,0 +1,90 @@
// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
// SPDX-License-Identifier: MIT
//
#pragma once
#include "ggml.h"
enum cpu_feature {
CPU_FEATURE_NONE = 0,
CPU_FEATURE_DOTPROD = 1,
CPU_FEATURE_I8MM = 2,
CPU_FEATURE_SVE = 4,
CPU_FEATURE_SME = 8
};
inline cpu_feature& operator|=(cpu_feature& lhs, cpu_feature rhs) {
lhs = static_cast<cpu_feature>(lhs | rhs);
return lhs;
}
inline cpu_feature operator|(cpu_feature lhs, cpu_feature rhs) {
return static_cast<cpu_feature>(static_cast<int>(lhs) | static_cast<int>(rhs));
}
struct kernel_info {
size_t (*get_m_step)(void);
size_t (*get_n_step)(void);
size_t (*get_mr)(void);
size_t (*get_nr)(void);
size_t (*get_kr)(void);
size_t (*get_sr)(void);
size_t (*get_dst_offset)(size_t m_idx, size_t n_idx, size_t stride);
size_t (*get_dst_size)(size_t m, size_t n);
size_t (*get_lhs_offset_ex)(size_t m_idx, size_t k, size_t bl);
size_t (*get_rhs_packed_offset_ex)(size_t n_idx, size_t k, size_t bl);
void (*run_kernel_ex)(
size_t m, size_t n, size_t k, size_t bl,
const void* lhs_packed, const void* rhs_packed,
void* dst, size_t dst_stride_row, size_t dst_stride_col,
float clamp_min, float clamp_max);
};
struct lhs_packing_info {
size_t (*get_offset)(size_t m_idx, size_t lhs_stride);
size_t (*get_packed_offset_ex)(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
size_t (*packed_size_ex)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
void (*pack_func_ex)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr,
size_t m_idx_start, const void * lhs, size_t lhs_stride, void * lhs_packed);
};
struct rhs_packing_info {
size_t (*packed_stride)(size_t k, size_t nr, size_t kr, size_t bl);
void (*to_float)(const void *packed_data, int32_t row_idx, int64_t nc, float *out,
size_t nr_pack, size_t packed_row_stride, size_t kr, size_t bl,
size_t num_bytes_multiplier);
size_t (*packed_size_ex)(size_t n, size_t k, size_t nr, size_t kr, size_t bl);
size_t (*packed_stride_ex)(size_t k, size_t nr, size_t kr, size_t bl);
void (*pack_func_ex)(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl,
size_t rhs_stride, const void * rhs, const void * bias, const void * scale, void * rhs_packed, size_t extra_bytes, const void * params);
};
struct ggml_kleidiai_kernels {
kernel_info gemm;
lhs_packing_info gemm_lhs_info;
kernel_info gemv;
lhs_packing_info gemv_lhs_info;
rhs_packing_info rhs_info;
cpu_feature required_cpu;
ggml_type lhs_type;
ggml_type rhs_type;
ggml_type op_type;
};
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor);
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features);
ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q8_0(cpu_feature features);
engine/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
+1523
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/kleidiai/kleidiai.h
+17
View File
@@ -0,0 +1,17 @@
// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
// SPDX-License-Identifier: MIT
//
#pragma once
#include "ggml-alloc.h"
#ifdef __cplusplus
extern "C" {
#endif
ggml_backend_buffer_type_t ggml_backend_cpu_kleidiai_buffer_type(void);
#ifdef __cplusplus
}
#endif
engine/ggml/src/ggml-cpu/llamafile/sgemm.cpp
+4052
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/llamafile/sgemm.h
+25
View File
@@ -0,0 +1,25 @@
#pragma once
#include <stdint.h>
#include <stdbool.h>
#if defined(__VXE__) || defined(__VXE2__)
#include <vecintrin.h>
#endif
#ifdef _MSC_VER
#define NOINLINE __declspec(noinline)
#else
#define NOINLINE __attribute__((__noinline__))
#endif
#ifdef __cplusplus
extern "C" {
#endif
bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t, int64_t, int64_t,
const void *, int64_t, const void *, int64_t, void *, int64_t,
int, int, int);
#ifdef __cplusplus
}
#endif
engine/ggml/src/ggml-cpu/ops.cpp
+11540
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/ops.h
+125
View File
@@ -0,0 +1,125 @@
#pragma once
#include "ggml.h"
//
// cache line
//
#if defined(__cpp_lib_hardware_interference_size)
#define CACHE_LINE_SIZE std::hardware_destructive_interference_size
#else
#if defined(__POWER9_VECTOR__)
#define CACHE_LINE_SIZE 128
#elif defined(__VXE__) || defined(__VXE2__)
#define CACHE_LINE_SIZE 256
#else
#define CACHE_LINE_SIZE 64
#endif
#endif
static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
// Work buffer size for im2col operations in CONV2D
#define GGML_IM2COL_WORK_SIZE (16 * 1024 * 1024)
#ifdef __cplusplus
extern "C" {
#endif
void ggml_compute_forward_dup(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add_id(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_acc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sum(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sum_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cumsum(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_mean(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_argmax(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_count_equal(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_repeat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_repeat_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_concat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_silu_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rms_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rms_norm_mul_fused(const struct ggml_compute_params * params, struct ggml_tensor * dst_rms_norm, struct ggml_tensor * dst_mul);
void ggml_compute_forward_rms_norm_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_group_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_l2_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_out_prod(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_scale(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_set(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cpy(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cont(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_get_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_get_rows_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_set_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_diag(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_diag_mask_inf(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_diag_mask_zero(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_soft_max(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_soft_max_ext_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rope(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rope_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_clamp(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_im2col_3d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_col2im_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_snake_fused(const struct ggml_compute_params * params,
const struct ggml_tensor * x,
const struct ggml_tensor * a,
const struct ggml_tensor * inv_b,
struct ggml_tensor * dst);
void ggml_compute_forward_conv_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_3d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_2d_dw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pool_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_upscale(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pad(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_pad_reflect_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_roll(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_arange(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_top_k(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_leaky_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_tri(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_fill(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_ext(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_flash_attn_back(
const struct ggml_compute_params * params,
const bool masked,
struct ggml_tensor * dst);
void ggml_compute_forward_ssm_conv(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_ssm_scan(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_win_part(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_win_unpart(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_unary(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_glu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_get_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_add_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rwkv_wkv6(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rwkv_wkv7(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_solve_tri(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_gla(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_gated_delta_net(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_map_custom1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_map_custom2(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_map_custom3(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_custom(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_fwht(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_opt_step_sgd(const struct ggml_compute_params * params, struct ggml_tensor * dst);
#ifdef __cplusplus
}
#endif
engine/ggml/src/ggml-cpu/quants.c
+1288
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/quants.h
+103
View File
@@ -0,0 +1,103 @@
#pragma once
#define GGML_COMMON_DECL_C
#include "ggml-common.h"
#include "ggml.h"
// GGML CPU internal header
#ifdef __cplusplus
extern "C" {
#endif
// Quantization
void quantize_row_q1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_nvfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
// Dot product
void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_nvfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq1_m_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
// Generic implementation
void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void ggml_vec_dot_q1_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_mxfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_nvfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
#ifdef __cplusplus
}
#endif
engine/ggml/src/ggml-cpu/repack.cpp
+4836
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/repack.h
+245
View File
@@ -0,0 +1,245 @@
#pragma once
#define GGML_COMMON_DECL_CPP
#include "ggml-common.h"
#include "traits.h"
#include "ggml.h"
// GGML internal header
ggml_backend_buffer_type_t ggml_backend_cpu_repack_buffer_type(void);
template <int K> constexpr int QK_0() {
if constexpr (K == 4) {
return QK4_0;
}
if constexpr (K == 8) {
return QK8_0;
}
return -1;
}
template <int K, int N> struct block {
ggml_half d[N]; // deltas for N qK_0 blocks
int8_t qs[(QK_0<K>() * N * K) / 8]; // quants for N qK_0 blocks
};
// control size
static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
static_assert(sizeof(block<4, 16>) == 16 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<4,16> size/padding");
static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
static_assert(sizeof(block<8, 16>) == 16 * sizeof(ggml_half) + QK8_0 * 16, "wrong block<8,16> size/padding");
using block_q4_0x4 = block<4, 4>;
using block_q4_0x8 = block<4, 8>;
using block_q4_0x16 = block<4, 16>;
using block_q8_0x4 = block<8, 4>;
using block_q8_0x8 = block<8, 8>;
using block_q8_0x16 = block<8, 16>;
struct block_q4_Kx8 {
ggml_half d[8]; // super-block scale for quantized scales
ggml_half dmin[8]; // super-block scale for quantized mins
uint8_t scales[96]; // scales and mins, quantized with 6 bits
uint8_t qs[1024]; // 4--bit quants
};
static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
struct block_q4_Kx16 {
ggml_half d[16]; // super-block scale for quantized scales
ggml_half dmin[16]; // super-block scale for quantized mins
uint8_t scales[192]; // scales and mins, quantized with 6 bits
uint8_t qs[2048]; // 4--bit quants
};
static_assert(sizeof(block_q4_Kx16) == sizeof(ggml_half) * 32 + K_SCALE_SIZE * 16 + QK_K * 8, "wrong q4_K block size/padding");
struct block_q2_Kx8 {
ggml_half d[8]; // super-block scale for quantized scales
ggml_half dmin[8]; // super-block scale for quantized mins
uint8_t scales[128]; // scales and mins, quantized with 4 bits
uint8_t qs[512]; // 2--bit quants
};
static_assert(sizeof(block_q2_Kx8) == sizeof(ggml_half) * 16 + QK_K/2 + QK_K * 2, "wrong q2_K block size/padding");
struct block_q2_Kx16 {
ggml_half d[16]; // Super-block scale for quantized scales
ggml_half dmin[16]; // Super-block scale for quantized mins
uint8_t scales[256]; // Sub-block scales (16 cols * 16 sub-blocks)
uint8_t qs[1024]; // Data (16 cols * 64 bytes per block)
};
static_assert(sizeof(block_q2_Kx16) == sizeof(ggml_half) * 32 + QK_K + QK_K * 4, "wrong q2_K block size/padding");
struct block_q5_Kx8 {
ggml_half d[8]; // super-block scale for quantized scales
ggml_half dmin[8]; // super-block scale for quantized mins
uint8_t scales[96]; // scales and mins, quantized with 6 bits
uint8_t qh[QK_K * 8 / 8]; // high bits of 5-bit quants
uint8_t qs[QK_K * 8 / 2]; // low bits of 5-bit quants (in groups of 4)
};
static_assert(sizeof(block_q5_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 5,
"wrong q5_K block size/padding");
struct block_q6_Kx8 {
ggml_half d[8];
int8_t scales[QK_K / 16 * 8];
uint8_t ql[QK_K / 2 * 8]; // low bits of 6-bit quants (groups of 2)
uint8_t qh[QK_K / 4 * 8]; // high bits of 6-bit quants (groups of 4)
};
static_assert(sizeof(block_q6_Kx8) == sizeof(ggml_half) * 8 + QK_K / 16 * 8 + 3 * QK_K / 4 * 8,
"wrong q6_K block size/padding");
struct block_q8_Kx4 {
float d[4]; // delta
int8_t qs[QK_K * 4]; // quants
int16_t bsums[QK_K / 4]; // sum of quants in groups of 16
};
static_assert(sizeof(block_q8_Kx4) == sizeof(float) * 4 + QK_K * 4 + (QK_K / 4) * sizeof(int16_t), "wrong q8_K block size/padding");
struct block_iq4_nlx4 {
ggml_half d[4]; // deltas for 4 iq4_nl blocks
uint8_t qs[QK4_NL * 2]; // nibbles / quants for 4 iq4_nl blocks
};
static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
struct block_iq4_nlx8 {
ggml_half d[8]; // deltas for 8 iq4_nl blocks
uint8_t qs[QK4_NL * 4]; // nibbles / quants for 8 iq4_nl blocks
};
static_assert(sizeof(block_iq4_nlx8) == 8 * sizeof(ggml_half) + QK4_NL * 4, "wrong iq4_nlx8 block size/padding");
struct block_iq4_nlx16 {
ggml_half d[16]; // deltas for 16 iq4_nl blocks
uint8_t qs[QK4_NL * 8]; // nibbles / quants for 16 iq4_nl blocks
};
static_assert(sizeof(block_iq4_nlx16) == 16 * sizeof(ggml_half) + QK4_NL * 8, "wrong iq4_nlx16 block size/padding");
struct block_mxfp4x4 {
uint8_t e[4];
uint8_t qs[QK_MXFP4 * 2];
};
static_assert(sizeof(block_mxfp4x4) == 4 + QK_MXFP4 * 2, "wrong mxfp4x4 block size/padding");
struct block_mxfp4x8 {
uint8_t e[8];
uint8_t qs[QK_MXFP4 * 4];
};
static_assert(sizeof(block_mxfp4x8) == 8 + QK_MXFP4 * 4, "wrong mxfp4x8 block size/padding");
#if defined(__cplusplus)
extern "C" {
#endif
void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_quantize_mat_q8_K_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q5_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_mxfp4_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q5_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_mxfp4_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q8_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
#if defined __riscv_zvfh
void ggml_quantize_mat_q8_0_4x1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_quantize_mat_q8_K_4x1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_gemv_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
#endif
// Native implementations
void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_quantize_mat_q8_K_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q5_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q5_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_mxfp4_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q8_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
#if defined __riscv_zvfh
void ggml_quantize_mat_q8_0_4x1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_quantize_mat_q8_K_4x1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
void ggml_gemv_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q8_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q8_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
#endif
#if defined(__cplusplus)
} // extern "C"
#endif
engine/ggml/src/ggml-cpu/simd-gemm.h
+226
View File
@@ -0,0 +1,226 @@
#pragma once
// Computes C[M x N] += A[M x K] * B[K x N]
#include "simd-mappings.h"
// TODO: add support for sizeless vector types
#if defined(GGML_SIMD) && !defined(__ARM_FEATURE_SVE) && !defined(__riscv_v_intrinsic)
// TODO: untested on avx512
// These are in units of GGML_F32_EPR
#if defined(__AVX512F__) || defined (__ARM_NEON__)
static constexpr int GEMM_RM = 4;
static constexpr int GEMM_RN = 4; // 16+4+1 = 25/32
#elif defined(__AVX2__) || defined(__AVX__)
static constexpr int GEMM_RM = 6;
static constexpr int GEMM_RN = 2; // 12+2+1 = 15/16
#else
static constexpr int GEMM_RM = 2;
static constexpr int GEMM_RN = 2;
#endif
template <int RM, int RN>
static inline void simd_gemm_ukernel(
float * GGML_RESTRICT C,
const float * GGML_RESTRICT A,
const float * GGML_RESTRICT B,
int K, int N)
{
static constexpr int KN = GGML_F32_EPR;
GGML_F32_VEC acc[RM][RN];
for (int64_t i = 0; i < RM; i++) {
for (int r = 0; r < RN; r++) {
acc[i][r] = GGML_F32_VEC_LOAD(C + i * N + r * KN);
}
}
for (int64_t kk = 0; kk < K; kk++) {
GGML_F32_VEC Bv[RN];
for (int r = 0; r < RN; r++) {
Bv[r] = GGML_F32_VEC_LOAD(B + kk * N + r * KN);
}
for (int64_t i = 0; i < RM; i++) {
GGML_F32_VEC p = GGML_F32_VEC_SET1(A[i * K + kk]);
for (int r = 0; r < RN; r++) {
acc[i][r] = GGML_F32_VEC_FMA(acc[i][r], Bv[r], p);
}
}
}
for (int64_t i = 0; i < RM; i++) {
for (int r = 0; r < RN; r++) {
GGML_F32_VEC_STORE(C + i * N + r * KN, acc[i][r]);
}
}
}
// C[M x N] += A[M x K] * B[K x N]
static void simd_gemm(
float * GGML_RESTRICT C,
const float * GGML_RESTRICT A,
const float * GGML_RESTRICT B,
int M, int K, int N)
{
static constexpr int KN = GGML_F32_EPR;
int64_t ii = 0;
for (; ii + GEMM_RM <= M; ii += GEMM_RM) {
int64_t jj = 0;
for (; jj + GEMM_RN * KN <= N; jj += GEMM_RN * KN) {
simd_gemm_ukernel<GEMM_RM, GEMM_RN>(C + jj, A, B + jj, K, N);
}
for (; jj + KN <= N; jj += KN) {
simd_gemm_ukernel<GEMM_RM, 1>(C + jj, A, B + jj, K, N);
}
for (; jj < N; jj++) {
for (int64_t i = 0; i < GEMM_RM; i++) {
float a = C[i * N + jj];
for (int64_t kk = 0; kk < K; kk++) {
a += A[i + kk] * B[kk * N + jj];
}
C[i * N + jj] = a;
}
}
A += GEMM_RM * K;
C += GEMM_RM * N;
}
// Tail rows: one at a time
for (; ii < M; ii++) {
int64_t jj = 0;
for (; jj + GEMM_RN * KN <= N; jj += GEMM_RN * KN) {
simd_gemm_ukernel<1, GEMM_RN>(C + jj, A, B + jj, K, N);
}
for (; jj + KN <= N; jj += KN) {
simd_gemm_ukernel<1, 1>(C + jj, A, B + jj, K, N);
}
for (; jj < N; jj++) {
float a = C[jj];
for (int64_t kk = 0; kk < K; kk++) {
a += A[kk] * B[kk * N + jj];
}
C[jj] = a;
}
A += K;
C += N;
}
}
#elif defined(GGML_SIMD) && defined(__riscv_v_intrinsic)
// RM accumulators + 1 B vector = RM + 1 <= 8 => RM <= 7
// Microkernel: C[RM x vl] += A[RM x K] * B[K x N]
template <int RM>
static inline void rvv_simd_gemm_ukernel(
float * GGML_RESTRICT C,
const float * GGML_RESTRICT A,
const float * GGML_RESTRICT B,
int K, int N, size_t vl)
{
static_assert(RM >= 1 && RM <= 7, "RM must be 1..7 for LMUL=4");
vfloat32m4_t acc_0 = __riscv_vle32_v_f32m4(C + 0 * N, vl);
vfloat32m4_t acc_1, acc_2, acc_3, acc_4, acc_5, acc_6;
if constexpr (RM > 1) acc_1 = __riscv_vle32_v_f32m4(C + 1 * N, vl);
if constexpr (RM > 2) acc_2 = __riscv_vle32_v_f32m4(C + 2 * N, vl);
if constexpr (RM > 3) acc_3 = __riscv_vle32_v_f32m4(C + 3 * N, vl);
if constexpr (RM > 4) acc_4 = __riscv_vle32_v_f32m4(C + 4 * N, vl);
if constexpr (RM > 5) acc_5 = __riscv_vle32_v_f32m4(C + 5 * N, vl);
if constexpr (RM > 6) acc_6 = __riscv_vle32_v_f32m4(C + 6 * N, vl);
for (int kk = 0; kk < K; kk++) {
vfloat32m4_t b_0 = __riscv_vle32_v_f32m4(B + kk * N, vl);
acc_0 = __riscv_vfmacc_vf_f32m4(acc_0, A[0 * K + kk], b_0, vl);
if constexpr (RM > 1) acc_1 = __riscv_vfmacc_vf_f32m4(acc_1, A[1 * K + kk], b_0, vl);
if constexpr (RM > 2) acc_2 = __riscv_vfmacc_vf_f32m4(acc_2, A[2 * K + kk], b_0, vl);
if constexpr (RM > 3) acc_3 = __riscv_vfmacc_vf_f32m4(acc_3, A[3 * K + kk], b_0, vl);
if constexpr (RM > 4) acc_4 = __riscv_vfmacc_vf_f32m4(acc_4, A[4 * K + kk], b_0, vl);
if constexpr (RM > 5) acc_5 = __riscv_vfmacc_vf_f32m4(acc_5, A[5 * K + kk], b_0, vl);
if constexpr (RM > 6) acc_6 = __riscv_vfmacc_vf_f32m4(acc_6, A[6 * K + kk], b_0, vl);
}
__riscv_vse32_v_f32m4(C + 0 * N, acc_0, vl);
if constexpr (RM > 1) __riscv_vse32_v_f32m4(C + 1 * N, acc_1, vl);
if constexpr (RM > 2) __riscv_vse32_v_f32m4(C + 2 * N, acc_2, vl);
if constexpr (RM > 3) __riscv_vse32_v_f32m4(C + 3 * N, acc_3, vl);
if constexpr (RM > 4) __riscv_vse32_v_f32m4(C + 4 * N, acc_4, vl);
if constexpr (RM > 5) __riscv_vse32_v_f32m4(C + 5 * N, acc_5, vl);
if constexpr (RM > 6) __riscv_vse32_v_f32m4(C + 6 * N, acc_6, vl);
}
template <int RM>
static inline void rvv_simd_gemm_dispatch_tail(
float * GGML_RESTRICT C,
const float * GGML_RESTRICT A,
const float * GGML_RESTRICT B,
int K, int N, int KN, int remaining_rows)
{
if constexpr (RM > 0) {
if (remaining_rows == RM) {
int64_t jj = 0;
for (; jj + KN <= N; jj += KN) {
rvv_simd_gemm_ukernel<RM>(C + jj, A, B + jj, K, N, KN);
}
if (jj < N) {
rvv_simd_gemm_ukernel<RM>(C + jj, A, B + jj, K, N, N - jj);
}
} else {
rvv_simd_gemm_dispatch_tail<RM - 1>(C, A, B, K, N, KN, remaining_rows);
}
}
}
static constexpr int GEMM_RM = 7;
// C[M x N] += A[M x K] * B[K x N]
static void simd_gemm(
float * GGML_RESTRICT C,
const float * GGML_RESTRICT A,
const float * GGML_RESTRICT B,
int M, int K, int N)
{
const int KN = (int)__riscv_vlenb();
int64_t ii = 0;
for (; ii + GEMM_RM <= M; ii += GEMM_RM) {
int64_t jj = 0;
for (; jj + KN <= N; jj += KN) {
rvv_simd_gemm_ukernel<GEMM_RM>(C + jj, A, B + jj, K, N, KN);
}
if (jj < N) {
rvv_simd_gemm_ukernel<GEMM_RM>(C + jj, A, B + jj, K, N, N - jj);
}
A += GEMM_RM * K;
C += GEMM_RM * N;
}
int remaining_rows = M - ii;
rvv_simd_gemm_dispatch_tail<GEMM_RM - 1>(C, A, B, K, N, KN, remaining_rows);
}
#if defined(__GNUC__) && !defined(__clang__)
#pragma GCC diagnostic pop
#endif
#else // scalar path
static void simd_gemm(
float * GGML_RESTRICT C,
const float * GGML_RESTRICT A,
const float * GGML_RESTRICT B,
int M, int K, int N)
{
for (int64_t i = 0; i < M; i++) {
for (int64_t j = 0; j < N; j++) {
float sum = C[i * N + j];
for (int64_t kk = 0; kk < K; kk++) {
sum += A[i * K + kk] * B[kk * N + j];
}
C[i * N + j] = sum;
}
}
}
#endif // GGML_SIMD
engine/ggml/src/ggml-cpu/simd-mappings.h
+1306
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/spacemit/ime.cpp
+1740
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/spacemit/ime.h
+21
View File
@@ -0,0 +1,21 @@
#pragma once
#include "ggml-alloc.h"
#ifdef __cplusplus
extern "C" {
#endif
ggml_backend_buffer_type_t ggml_backend_cpu_riscv64_spacemit_buffer_type(void);
void ggml_backend_cpu_riscv64_spacemit_set_numa_thread_affinity(int thread_n);
void ggml_backend_cpu_riscv64_spacemit_clear_numa_thread_affinity_threaded(int thread_n);
void * ggml_backend_cpu_riscv64_spacemit_alloc_shared(size_t size, size_t alignment);
void ggml_backend_cpu_riscv64_spacemit_free_shared(void * ptr);
#ifdef __cplusplus
}
#endif
engine/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp
+1027
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp
+5768
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/spacemit/ime_env.cpp
+320
View File
@@ -0,0 +1,320 @@
#include "ime_env.h"
#include "ggml-impl.h"
#include "spine_mem_pool.h"
#include <fcntl.h>
#include <unistd.h>
#include <algorithm>
#include <array>
#include <cctype>
#include <fstream>
#include <string>
#include <thread>
#include <unordered_map>
namespace ggml::cpu::riscv64_spacemit {
bool spine_core_info::get_spine_core_info(std::vector<spine_core_info> & result) {
static std::unordered_map<uint64_t, spine_core_arch_id> spine_march_mapping_ = {
{0x8000000058000001, spine_core_arch_id::core_arch_x60 },
{ 0x8000000041000001, spine_core_arch_id::core_arch_a60 },
{ 0x8000000058000002, spine_core_arch_id::core_arch_x100},
{ 0x8000000041000002, spine_core_arch_id::core_arch_a100},
};
result.clear();
std::ifstream file("/proc/cpuinfo");
std::string line;
std::vector<std::array<uint64_t, 2>> cpu_info_list;
uint64_t current_processor = spine_invalid_core_id;
uint64_t current_marchid = 0;
bool has_processor = false;
bool has_marchid = false;
if (!file.is_open()) {
return false;
}
while (std::getline(file, line)) {
if (line.substr(0, 9) == "processor") {
if (has_processor && has_marchid) {
cpu_info_list.push_back({ current_processor, current_marchid });
}
size_t colon_pos = line.find(':');
if (colon_pos != std::string::npos) {
current_processor = std::stoi(line.substr(colon_pos + 1));
has_processor = true;
}
has_marchid = false;
} else if (line.substr(0, 7) == "marchid") {
size_t colon_pos = line.find(':');
if (colon_pos != std::string::npos) {
std::string marchid_str = line.substr(colon_pos + 1);
marchid_str.erase(std::remove_if(marchid_str.begin(), marchid_str.end(), isspace), marchid_str.end());
current_marchid = std::stoull(marchid_str, nullptr, 16);
has_marchid = true;
}
}
}
if (has_processor && has_marchid) {
cpu_info_list.push_back({ current_processor, current_marchid });
}
if (has_processor && has_marchid) {
for (auto & cpu_info : cpu_info_list) {
if (cpu_info[0] != spine_invalid_core_id &&
spine_march_mapping_.find(cpu_info[1]) != spine_march_mapping_.end()) {
auto core_info = spine_core_info();
core_info.core_id = cpu_info[0];
core_info.arch_id = spine_core_arch_id(spine_march_mapping_[cpu_info[1]]);
result.push_back(core_info);
}
}
}
return has_processor && has_marchid;
}
namespace {
uint16_t hex_string_to_u16(const std::string & hex_str) {
try {
size_t pos = 0;
if (hex_str.substr(0, 2) == "0x" || hex_str.substr(0, 2) == "0X") {
pos = 2;
}
unsigned long result = std::stoul(hex_str.substr(pos), nullptr, 16);
if (result > std::numeric_limits<uint16_t>::max()) {
throw std::out_of_range("Converted value is out of range for uint16_t");
}
return static_cast<uint16_t>(result);
} catch (const std::invalid_argument & e) {
throw std::invalid_argument("Invalid hexadecimal string");
} catch (const std::out_of_range & e) {
throw;
}
}
const char * spine_mem_pool_backend_to_string(spine_mem_pool_backend backend) {
switch (backend) {
case spine_mem_pool_backend::none:
return "NONE";
case spine_mem_pool_backend::posix_memalign:
return "POSIX";
case spine_mem_pool_backend::transparent_hugepage:
return "HPAGE";
case spine_mem_pool_backend::hugetlb_1g:
return "HPAGE1GB";
}
return "unknown";
}
spine_mem_pool_backend parse_mem_backend(const char * mem_backend_str) {
if (mem_backend_str == nullptr || mem_backend_str[0] == '\0') {
return spine_mem_pool_backend::transparent_hugepage;
}
std::string value(mem_backend_str);
std::transform(value.begin(), value.end(), value.begin(),
[](unsigned char ch) { return static_cast<char>(std::tolower(ch)); });
if (value == "none") {
return spine_mem_pool_backend::none;
}
if (value == "posix") {
return spine_mem_pool_backend::posix_memalign;
}
if (value == "hpage") {
return spine_mem_pool_backend::transparent_hugepage;
}
if (value == "hpage1gb") {
return spine_mem_pool_backend::hugetlb_1g;
}
throw std::runtime_error("invalid SPACEMIT_MEM_BACKEND: " + value + ", expected NONE, POSIX, HPAGE or HPAGE1GB");
}
} // namespace
spine_env_info::spine_env_info() {
num_cores = static_cast<int>(std::thread::hardware_concurrency());
spine_core_info::get_spine_core_info(core_info_list);
// special for x60 K1
if (core_info_list.size() == 8 && core_info_list[0].arch_id == spine_core_arch_id::core_arch_x60) {
for (int i = 0; i < 4; i++) {
core_info_list[i].arch_id = spine_core_arch_id::core_arch_a60;
}
}
// special for qemu
if (core_info_list.size() == 0) {
char * spine_core_arch_str = getenv("SPACEMIT_CORE_ARCH");
if (spine_core_arch_str != nullptr) {
auto arch_id = hex_string_to_u16(spine_core_arch_str);
for (int i = 0; i < num_cores; i++) {
auto core_info = spine_core_info();
core_info.core_id = i;
core_info.arch_id = spine_core_arch_id{ arch_id };
core_info_list.push_back(core_info);
}
}
}
if (core_info_list.size() == 0) {
throw std::runtime_error(
"Failed to get SPACEMIT_CORE_ARCH from environment or failed to parse it from /proc/cpuinfo");
}
char * spine_perfer_core_arch_str = getenv("SPACEMIT_PERFER_CORE_ARCH");
if (spine_perfer_core_arch_str != nullptr && spine_perfer_core_arch_str != "") {
perfer_core_arch_id = spine_core_arch_id{ hex_string_to_u16(spine_perfer_core_arch_str) };
}
char * spine_perfer_core_id_str = getenv("SPACEMIT_PERFER_CORE_ID");
std::vector<int> perfer_core_id_vec;
if (spine_perfer_core_id_str != nullptr && spine_perfer_core_id_str != "") {
std::string perfer_core_id_str(spine_perfer_core_id_str);
size_t start = 0;
size_t end = 0;
while ((end = perfer_core_id_str.find(',', start)) != std::string::npos) {
std::string core_id_substr = perfer_core_id_str.substr(start, end - start);
perfer_core_id_vec.push_back(std::stoi(core_id_substr));
start = end + 1;
}
std::string core_id_substr = perfer_core_id_str.substr(start);
perfer_core_id_vec.push_back(std::stoi(core_id_substr));
}
perfer_core_ids.reserve(num_cores);
if (perfer_core_arch_id == spine_core_arch_id::core_arch_none) {
for (auto & core_info : core_info_list) {
auto core_arch_id = core_info.arch_id;
auto core_arch_head = (uint16_t) (core_arch_id) >> 12;
if (core_arch_head == 0xA) {
num_perfer_cores++;
perfer_core_arch_id = core_arch_id;
cpu_mask |= (1ULL << core_info.core_id);
perfer_core_ids.push_back(core_info.core_id);
}
}
} else {
for (auto & core_info : core_info_list) {
auto core_arch_id = core_info.arch_id;
if (core_arch_id == perfer_core_arch_id) {
num_perfer_cores++;
cpu_mask |= (1ULL << core_info.core_id);
auto core_arch_head = (uint16_t) (core_arch_id) >> 12;
if (core_arch_head == 0xA) {
perfer_core_ids.push_back(core_info.core_id);
}
}
}
if (num_perfer_cores == 0) {
GGML_ABORT("can not find core with arch id %x for SPACEMIT_PERFER_CORE_ARCH in core info list\n",
(uint16_t) perfer_core_arch_id);
}
}
if (perfer_core_id_vec.size() > 0) {
perfer_core_ids.clear();
cpu_mask = 0;
num_perfer_cores = 0;
for (int core_id : perfer_core_id_vec) {
if (core_id < 0 || core_id >= num_cores) {
GGML_ABORT("invalid core id in SPACEMIT_PERFER_CORE_ID: %d, should be between 0 and %d\n", core_id,
num_cores - 1);
}
auto core_info = core_info_list[core_id];
auto core_arch_id = core_info.arch_id;
if (core_arch_id == perfer_core_arch_id) {
cpu_mask |= (1ULL << core_id);
perfer_core_ids.push_back(core_id);
} else {
GGML_ABORT(
"core id %d in SPACEMIT_PERFER_CORE_ID has arch id %x which does not match "
"SPACEMIT_PERFER_CORE_ARCH %x\n",
core_id, (uint16_t) core_arch_id, (uint16_t) perfer_core_arch_id);
}
}
std::string perfer_core_id_vec_str;
for (int core_id : perfer_core_id_vec) {
perfer_core_id_vec_str += std::to_string(core_id) + ",";
}
perfer_core_id_vec_str.pop_back();
GGML_LOG_DEBUG("SPACEMIT_PERFER_CORE_ID is set, perferred core ids: %s\n", perfer_core_id_vec_str.c_str());
num_perfer_cores = static_cast<int>(perfer_core_id_vec.size());
}
use_ime1 = perfer_core_arch_id == spine_core_arch_id::core_arch_a60 ||
perfer_core_arch_id == spine_core_arch_id::core_arch_x100;
use_ime2 = perfer_core_arch_id == spine_core_arch_id::core_arch_a100;
mem_backend = parse_mem_backend(getenv("SPACEMIT_MEM_BACKEND"));
char * spine_disable_tcm_str = getenv("SPACEMIT_DISABLE_TCM");
auto user_disable_tcm = spine_disable_tcm_str != nullptr && strcmp(spine_disable_tcm_str, "0") != 0;
if (!user_disable_tcm) {
spine_mem_pool_tcm_info tcm_info;
if (spine_mem_pool_tcm_init(&tcm_info)) {
use_tcm = tcm_info.available;
tcm_blk_size = tcm_info.blk_size;
GGML_LOG_DEBUG("CPU_RISCV64_SPACEMIT: tcm is available, blk_size: %zu, blk_num: %zu, is_fake_tcm: %d\n",
tcm_info.blk_size, tcm_info.blk_num, tcm_info.is_fake_tcm);
for (auto & core_info : core_info_list) {
auto core_arch_head = (uint16_t) (core_info.arch_id) >> 12;
if (core_arch_head != 0xA) {
aicpu_id_offset++;
} else {
break;
}
}
}
}
GGML_LOG_DEBUG(
"CPU_RISCV64_SPACEMIT: num_cores: %d, num_perfer_cores: %d, perfer_core_arch_id: %x, exclude_main_thread: %d, "
"use_ime1: %d, use_ime2: %d, mem_backend: %s, cpu_mask: %lx, aicpu_id_offset: %d\n",
num_cores, num_perfer_cores, (uint16_t) perfer_core_arch_id, exclude_main_thread, use_ime1, use_ime2,
spine_mem_pool_backend_to_string(mem_backend), cpu_mask, aicpu_id_offset);
const size_t init_barrier_size = sizeof(spine_barrier_t) * spine_init_barrier_count;
init_barrier =
static_cast<spine_barrier_t *>(spine_mem_pool_shared_mem_alloc(init_barrier_size, alignof(spine_barrier_t)));
if (init_barrier != nullptr) {
init_barrier_is_shared_mem = true;
} else {
GGML_LOG_WARN("CPU_RISCV64_SPACEMIT: failed to allocate init_barrier from shared mem, falling back to heap\n",
__func__);
init_barrier = new spine_barrier_t[spine_init_barrier_count];
}
spine_barrier_init(init_barrier, spine_init_barrier_count, 2);
}
spine_env_info::~spine_env_info() {
if (init_barrier_is_shared_mem) {
spine_mem_pool_shared_mem_free(init_barrier);
} else {
delete[] init_barrier;
}
init_barrier = nullptr;
init_barrier_is_shared_mem = false;
}
spine_env_info global_spine_env_info;
} // namespace ggml::cpu::riscv64_spacemit
engine/ggml/src/ggml-cpu/spacemit/ime_env.h
+55
View File
@@ -0,0 +1,55 @@
#pragma once
#include "spine_barrier.h"
#include "spine_mem_pool.h"
#include <cstddef>
#include <cstdint>
#include <vector>
namespace ggml::cpu::riscv64_spacemit {
constexpr uint64_t spine_invalid_core_id = 0xFFFFFFFF;
constexpr size_t spine_init_barrier_count = 16;
enum class spine_core_arch_id : uint16_t {
core_arch_none = 0,
core_arch_x60 = 0x503C,
core_arch_x100 = 0x5064,
core_arch_x200 = 0x50C8,
core_arch_a60 = 0xA03C,
core_arch_a100 = 0xA064,
core_arch_a200 = 0xA0C8,
};
struct spine_core_info {
uint64_t core_id{ spine_invalid_core_id };
spine_core_arch_id arch_id{ spine_core_arch_id::core_arch_none };
static bool get_spine_core_info(std::vector<spine_core_info> & result);
};
struct spine_env_info {
std::vector<spine_core_info> core_info_list;
std::vector<int> perfer_core_ids;
int aicpu_id_offset{ 0 };
int num_cores{ 0 };
int num_perfer_cores{ 0 };
spine_core_arch_id perfer_core_arch_id{ spine_core_arch_id::core_arch_none };
bool exclude_main_thread{ false };
bool use_ime2{ false };
bool use_ime1{ false };
bool use_tcm{ false };
spine_mem_pool_backend mem_backend{ spine_mem_pool_backend::transparent_hugepage };
uint64_t tcm_blk_size{ 0 };
uint64_t cpu_mask{ 0 };
spine_barrier_t * init_barrier{ nullptr };
bool init_barrier_is_shared_mem{ false };
spine_env_info();
~spine_env_info();
};
extern spine_env_info global_spine_env_info;
} // namespace ggml::cpu::riscv64_spacemit
engine/ggml/src/ggml-cpu/spacemit/ime_kernels.h
+189
View File
@@ -0,0 +1,189 @@
#pragma once
#include <cassert>
#include <cstddef>
#include <functional>
namespace spacemit_kernels {
#define BLOCK_QNK_LEN 256
template <int N> struct nrow_block_q2_k {
// [4bit scale + 4bit zp] * N * 16
uint8_t scales[N * BLOCK_QNK_LEN / 16];
// [b0, b16, b32, b48] [b1, b17, b33, b49] ... [b15, b31, b47, b63]
// [b64, b80, b96, b112] ...[b79, b95, b111, b127]
// [b128, b144, b160, b176] ...[b143, b159, b175, b191]
// [b192, b208, b224, b240] ...[b207, b223, b239, b255]
uint8_t qs[N * BLOCK_QNK_LEN / 4];
uint16_t scales16[N];
uint16_t zeros16[N];
};
template <int N> struct nrow_block_q3_k {
// [8bit scale] * N * 16
int8_t scales[N * 16];
// [b0, b1, b2, b3, b4, b5, b6, b7] ... [b248, b249, b250, b251, b252, b253, b254, b255]
uint8_t hmask[N * BLOCK_QNK_LEN / 8];
// [b0, b16, b32, b48] [b1, b17, b33, b49] ... [b15, b31, b47, b63]
// [b64, b80, b96, b112] ...[b79, b95, b111, b127]
// [b128, b144, b160, b176] ...[b143, b159, b175, b191]
// [b192, b208, b224, b240] ...[b207, b223, b239, b255]
uint8_t qs[N * BLOCK_QNK_LEN / 4];
uint16_t scales16[N];
};
template <int N> struct nrow_block_mxfp4 {
uint8_t e[N];
uint8_t qh[4 * N];
uint8_t qs[16 * N];
};
template <int N> struct __attribute__((packed)) nrow_block_q5_1 {
uint16_t scales16[N];
uint8_t zp[N];
// n0 [bh0, bh1, bh2, bh3, bh4, bh5, bh6, bh7] ....
uint8_t qh[4 * N];
// n0 [b0, b1], [b2, b3] .... [b30, b31]
// n1 [b0, b1], [b2, b3] .... [b30, b31]
uint8_t qs[16 * N];
};
static_assert(sizeof(nrow_block_q5_1<1>) == sizeof(uint8_t) + 22, "wrong nrow_block_q5_1 block size/padding");
template <int N> struct __attribute__((packed)) nrow_block_q5_0 {
uint16_t scales16[N];
// n0 [bh0, bh1, bh2, bh3, bh4, bh5, bh6, bh7] ....
uint8_t qh[4 * N];
// n0 [b0, b1], [b2, b3] .... [b30, b31]
// n1 [b0, b1], [b2, b3] .... [b30, b31]
uint8_t qs[16 * N];
};
static_assert(sizeof(nrow_block_q5_0<1>) == 22, "wrong nrow_block_q5_0 block size/padding");
using gemm_kernel_quantize_def = std::function<
size_t(size_t, const uint8_t *, const uint8_t *, const uint8_t *, float *, size_t, size_t, size_t, size_t)>;
using moe_gemm_kernel_quantize_def = std::function<
size_t(size_t, const uint8_t **, const uint8_t *, const uint8_t *, float **, size_t, size_t, size_t, size_t)>;
namespace ime1 {
size_t gemm_kernel_i8i4(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
void quantize_a_row_i8(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
void quantize_a_4row_i8(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
} // namespace ime1
namespace ime2 {
size_t gemm_kernel_i8i2k(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t gemm_kernel_i8i3k(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t gemm_kernel_i8i4(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t gemm_kernel_i8i4_hp(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t moe_m2_gemm_kernel_i8i4(size_t blk_len,
const uint8_t ** quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float ** c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t gemm_kernel_i8i8(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t gemm_kernel_i8mxfp4(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t moe_m2_gemm_kernel_i8mxfp4(size_t blk_len,
const uint8_t ** quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float ** c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t gemm_kernel_i8i5(size_t blk_len,
const uint8_t * quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float * c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
size_t moe_m2_gemm_kernel_i8i5(size_t blk_len,
const uint8_t ** quant_a_ptr,
const uint8_t * quant_b_data,
const uint8_t * quant_b_zp,
float ** c_ptr,
size_t count_m,
size_t count_n,
size_t k_blks,
size_t ldc);
} // namespace ime2
} // namespace spacemit_kernels
engine/ggml/src/ggml-cpu/spacemit/repack.cpp
+1795
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/spacemit/repack.h
+14
View File
@@ -0,0 +1,14 @@
#pragma once
#include "ggml-common.h"
#include "ggml.h"
#include <cstddef>
#include <cstdint>
namespace ggml::cpu::riscv64_spacemit {
template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
int repack(ggml_tensor * t, const void * data, size_t data_size);
} // namespace ggml::cpu::riscv64_spacemit
engine/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp
+3178
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cpu/spacemit/rvv_kernels.h
+95
View File
@@ -0,0 +1,95 @@
#pragma once
#include "ggml-cpu-impl.h"
#include <cassert>
#include <cstddef>
#include <cstdint>
#include <functional>
namespace spacemit_kernels {
constexpr auto div_round_up(auto up, auto down) {
return (up + down - 1) / down;
}
// Q8 Blk [f32] [s16] [int8 * blk_len]
// Q8 Blk N [f32 * N] [s16 * N] [int8 * blk_len * N]
constexpr size_t q8_blk_size(size_t blk_len, bool with_blk_sum = false) {
const size_t blk_size = sizeof(float) + blk_len * sizeof(int8_t) + (with_blk_sum ? sizeof(int16_t) : 0);
return blk_size;
}
// Q8 HP row block: K is split into K32 subblocks.
// Each subblock stores [f32 scale] [int8 * 32], with an optional fp16 sum trailer per subblock.
constexpr size_t q8_hp_blk_size(size_t blk_len, bool with_blk_sum = false, bool with_blk_scale = false) {
const size_t subblk_count = div_round_up(blk_len, size_t(32));
const size_t blk_size = blk_len * sizeof(int8_t) + subblk_count * sizeof(_Float16) +
(with_blk_sum ? subblk_count * sizeof(_Float16) : 0) +
(with_blk_scale ? sizeof(_Float16) : 0);
return blk_size;
}
// Q8K Blk [f32] [s16 * (blk_len / 16)] [int8 * blk_len]
// Q8K Blk N [f32 * N] [s16 * (blk_len / 16) * N] [int8 * blk_len * N]
constexpr size_t q8k_blk_size(size_t blk_len) {
const size_t blk_size = sizeof(float) + blk_len * sizeof(int8_t) + sizeof(int16_t) * blk_len / 16;
return blk_size;
}
using quantize_a_row_def = std::function<void(size_t, const float *, size_t, uint8_t *)>;
namespace rvv {
void memcpy1d(void * dst, const void * src, int64_t size);
void memcpy2d(void * dst, int64_t dst_stride, const void * src, int64_t src_stride, int64_t tile_rows, int64_t size);
void forward_flash_attn_ext_f16_one_chunk_vlen1024_vf16(const ggml_compute_params * params,
ggml_tensor * dst,
int ir0,
int ir1,
void * tcm_buffer,
size_t tcm_buffer_size);
void forward_flash_attn_ext_f16_tiled_vlen1024_vf16(const ggml_compute_params * params,
ggml_tensor * dst,
int ir0,
int ir1,
void * tcm_buffer,
size_t tcm_buffer_size);
void forward_rms_norm_f32(ggml_compute_params * params, ggml_tensor * op);
void forward_norm_f32(ggml_compute_params * params, ggml_tensor * op);
void forward_cont_with_permute(ggml_compute_params * params, ggml_tensor * op);
void forward_cpy_with_permute(ggml_compute_params * params, ggml_tensor * op);
template <typename T> void forward_get_rows(ggml_compute_params * params, ggml_tensor * op);
template <typename T> void forward_concat(ggml_compute_params * params, ggml_tensor * op);
template <ggml_op op_type, typename T> void forward_binary(ggml_compute_params * params, ggml_tensor * op);
template <typename T> void forward_sum_rows(const ggml_compute_params * params, ggml_tensor * op);
template <typename T> void forward_repeat_nrows(ggml_compute_params * params, ggml_tensor * op);
template <typename T> void forward_repeat_dim1(ggml_compute_params * params, ggml_tensor * op);
void quantize_a_row_i8(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
void quantize_a_4row_i8(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
void quantize_a_row_i8_hp(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
void quantize_a_4row_i8_hp(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
void quantize_a_row_i8k(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
void quantize_a_4row_i8k(size_t blk_len, const float * a_ptr, size_t count_k, uint8_t * quant_a_ptr);
} // namespace rvv
} // namespace spacemit_kernels
engine/ggml/src/ggml-cpu/spacemit/spine_barrier.h
+34
View File
@@ -0,0 +1,34 @@
#pragma once
#include <atomic>
#include <cstdint>
#define SPINE_CACHE_LINE 64
#define SPINE_CACHE_ALIGN __attribute__((aligned(SPINE_CACHE_LINE)))
struct spine_barrier_t {
SPINE_CACHE_ALIGN std::atomic<int64_t> pending_;
SPINE_CACHE_ALIGN std::atomic<int64_t> rounds_;
SPINE_CACHE_ALIGN int64_t total_;
};
inline void spine_barrier_wait(spine_barrier_t * b) {
auto cur_round = b->rounds_.load(std::memory_order_acquire);
auto cnt = --b->pending_;
if (cnt == 0) {
b->pending_.store(b->total_);
b->rounds_.store(cur_round + 1);
} else {
while (cur_round == b->rounds_.load(std::memory_order_relaxed)) {
__asm__ volatile("pause " ::: "memory");
}
}
}
inline void spine_barrier_init(spine_barrier_t * b, int num_barriers, uint64_t thread_count) {
for (int i = 0; i < num_barriers; i++) {
b[i].total_ = thread_count;
b[i].pending_.store(thread_count);
b[i].rounds_.store(0);
}
}
engine/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp
+760
View File
@@ -0,0 +1,760 @@
#include "spine_mem_pool.h"
#include "common.h"
#include "ime_env.h"
#include "spine_tcm.h"
#include <fcntl.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <unistd.h>
#include <algorithm>
#include <cerrno>
#include <cstdint>
#include <cstdlib>
#include <limits>
#include <memory>
#include <mutex>
#include <unordered_map>
#include <vector>
namespace ggml::cpu::riscv64_spacemit {
namespace {
constexpr size_t SPINE_MEM_POOL_CHUNK_SIZE = 512ull * 1024ull * 1024ull;
constexpr size_t SPINE_SHARE_MEM_POOL_CHUNK_SIZE = 512ull * 1024ull;
constexpr size_t SPINE_MEM_POOL_1G_REGION_SIZE = 1ull << 30;
constexpr uint64_t HUGETLB_1G_FLAG_REQUIRE_PUD = 1ull << 0;
constexpr char SPINE_MEM_POOL_HUGETLB_1G_DEV[] = "/dev/hugetlb_1g";
constexpr char SPINE_MEM_POOL_TCM_SYNC_MEM_DEV[] = "/dev/tcm_sync_mem";
struct hugetlb_1g_region {
uint64_t size{ 0 };
uint64_t dma_addr{ 0 };
uint64_t flags{ 0 };
uint64_t reserved{ 0 };
};
#define HUGETLB_1G_IOC_MAGIC 'M'
#define HUGETLB_1G_IOC_ALLOC _IOWR(HUGETLB_1G_IOC_MAGIC, 0x00, struct hugetlb_1g_region)
#define HUGETLB_1G_IOC_FREE _IO(HUGETLB_1G_IOC_MAGIC, 0x01)
struct free_block {
size_t offset{ 0 };
size_t size{ 0 };
};
struct pool_chunk {
uint8_t * base{ nullptr };
size_t size{ 0 };
int fd{ -1 };
std::vector<free_block> free_blocks;
};
struct pool_allocation {
void * chunk_base{ nullptr };
size_t chunk_size{ 0 };
void * base{ nullptr };
size_t size{ 0 };
};
bool is_power_of_two(size_t value) {
return value != 0 && (value & (value - 1)) == 0;
}
bool align_up(size_t value, size_t alignment, size_t * aligned_value) {
if (aligned_value == nullptr || alignment == 0) {
return false;
}
const size_t remainder = value % alignment;
if (remainder == 0) {
*aligned_value = value;
return true;
}
const size_t padding = alignment - remainder;
if (value > std::numeric_limits<size_t>::max() - padding) {
return false;
}
*aligned_value = value + padding;
return true;
}
bool align_up_uintptr(uintptr_t value, size_t alignment, uintptr_t * aligned_value) {
if (aligned_value == nullptr || alignment == 0) {
return false;
}
const uintptr_t remainder = value % alignment;
if (remainder == 0) {
*aligned_value = value;
return true;
}
const uintptr_t padding = alignment - remainder;
if (value > std::numeric_limits<uintptr_t>::max() - padding) {
return false;
}
*aligned_value = value + padding;
return true;
}
class spine_mem_pool_manager {
public:
explicit spine_mem_pool_manager(size_t default_chunk_size) : default_chunk_size_(default_chunk_size) {}
virtual ~spine_mem_pool_manager() = default;
void * alloc(size_t size, size_t alignment) {
if (size == 0 || !is_power_of_two(alignment)) {
return nullptr;
}
size_t aligned_size = 0;
if (!align_up(size, alignment, &aligned_size)) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: align_up failed for size %zu alignment %zu\n", __func__, size,
alignment);
return nullptr;
}
pool_allocation allocation;
std::lock_guard<std::mutex> lock(mutex_);
if (!try_alloc_locked(aligned_size, alignment, &allocation)) {
if (!add_chunk_locked(aligned_size, alignment)) {
return nullptr;
}
if (!try_alloc_locked(aligned_size, alignment, &allocation)) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: allocation retry failed for size %zu alignment %zu\n",
__func__, aligned_size, alignment);
return nullptr;
}
}
try {
const auto [allocation_it, inserted] = allocations_.emplace(allocation.base, allocation);
if (!inserted) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: duplicate allocation key %p\n", __func__, allocation.base);
rollback_allocation_locked(allocation);
return nullptr;
}
} catch (const std::bad_alloc &) {
rollback_allocation_locked(allocation);
throw;
}
return allocation.base;
}
void free(void * base) {
if (base == nullptr) {
return;
}
std::lock_guard<std::mutex> lock(mutex_);
auto allocation_it = allocations_.find(base);
if (allocation_it == allocations_.end()) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: unknown allocation %p\n", __func__, base);
return;
}
pool_allocation allocation = allocation_it->second;
allocations_.erase(allocation_it);
auto chunk_it = find_chunk_locked(allocation);
if (chunk_it == chunks_.end()) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: unknown chunk for allocation %p size %zu\n", __func__,
allocation.base, allocation.size);
return;
}
auto * chunk_base = chunk_it->base;
auto * alloc_base = static_cast<uint8_t *>(allocation.base);
if (alloc_base < chunk_base || alloc_base >= chunk_base + chunk_it->size) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: allocation %p out of chunk range %p..%p\n", __func__,
allocation.base, chunk_base, chunk_base + chunk_it->size);
return;
}
const size_t offset = static_cast<size_t>(alloc_base - chunk_base);
if (offset > chunk_it->size || allocation.size > chunk_it->size - offset) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: allocation %p size %zu exceeds chunk size %zu\n", __func__,
allocation.base, allocation.size, chunk_it->size);
return;
}
insert_free_block_locked(*chunk_it, { offset, allocation.size });
maybe_release_empty_chunk_locked(chunk_it);
}
protected:
void release_chunks() {
std::lock_guard<std::mutex> lock(mutex_);
allocations_.clear();
for (auto & chunk : chunks_) {
dealloc_chunk(&chunk);
}
chunks_.clear();
}
size_t default_chunk_size() const { return default_chunk_size_; }
static void clear_chunk(pool_chunk * chunk) {
chunk->base = nullptr;
chunk->size = 0;
chunk->fd = -1;
chunk->free_blocks.clear();
}
virtual bool alloc_chunk(size_t min_size, size_t alignment, void * hint_addr, pool_chunk * chunk) = 0;
virtual void dealloc_chunk(pool_chunk * chunk) = 0;
private:
struct alloc_candidate {
size_t chunk_index{ 0 };
size_t block_index{ 0 };
size_t aligned_offset{ 0 };
uintptr_t address{ std::numeric_limits<uintptr_t>::max() };
bool valid{ false };
};
std::vector<pool_chunk>::iterator find_chunk_locked(const pool_allocation & allocation) {
return std::find_if(chunks_.begin(), chunks_.end(), [&](const pool_chunk & chunk) {
return chunk.base == allocation.chunk_base && chunk.size == allocation.chunk_size;
});
}
bool add_chunk_locked(size_t min_size, size_t alignment) {
pool_chunk chunk;
const size_t chunk_request = default_chunk_size_ == 0 ? min_size : std::max(min_size, default_chunk_size_);
void * hint_addr = nullptr;
for (const auto & existing_chunk : chunks_) {
auto * chunk_end = existing_chunk.base + existing_chunk.size;
if (hint_addr == nullptr || chunk_end > hint_addr) {
hint_addr = chunk_end;
}
}
if (!alloc_chunk(chunk_request, alignment, hint_addr, &chunk)) {
return false;
}
if (chunk.base == nullptr || chunk.size < min_size) {
GGML_LOG_ERROR(
"CPU_RISCV64_SPACEMIT: %s: invalid chunk returned for request size %zu, chunk_base=%p chunk_size=%zu\n",
__func__, min_size, chunk.base, chunk.size);
dealloc_chunk(&chunk);
return false;
}
try {
chunk.free_blocks.push_back({ 0, chunk.size });
chunks_.push_back(std::move(chunk));
} catch (const std::bad_alloc &) {
dealloc_chunk(&chunk);
throw;
}
return true;
}
void rollback_allocation_locked(const pool_allocation & allocation) {
auto chunk_it = find_chunk_locked(allocation);
if (chunk_it == chunks_.end()) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: failed to rollback allocation %p, owning chunk not found\n",
__func__, allocation.base);
return;
}
auto * chunk_base = chunk_it->base;
auto * alloc_base = static_cast<uint8_t *>(allocation.base);
if (alloc_base < chunk_base || alloc_base >= chunk_base + chunk_it->size) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: failed to rollback allocation %p, chunk range is invalid\n",
__func__, allocation.base);
return;
}
const size_t offset = static_cast<size_t>(alloc_base - chunk_base);
if (offset > chunk_it->size || allocation.size > chunk_it->size - offset) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: failed to rollback allocation %p size %zu\n", __func__,
allocation.base, allocation.size);
return;
}
insert_free_block_locked(*chunk_it, { offset, allocation.size });
maybe_release_empty_chunk_locked(chunk_it);
}
bool try_alloc_locked(size_t size, size_t alignment, pool_allocation * allocation) {
alloc_candidate best;
for (size_t chunk_index = 0; chunk_index < chunks_.size(); ++chunk_index) {
const auto & chunk = chunks_[chunk_index];
for (size_t block_index = 0; block_index < chunk.free_blocks.size(); ++block_index) {
const auto & block = chunk.free_blocks[block_index];
uintptr_t aligned_addr = 0;
const auto block_addr = reinterpret_cast<uintptr_t>(chunk.base + block.offset);
if (!align_up_uintptr(block_addr, alignment, &aligned_addr)) {
continue;
}
if (aligned_addr < block_addr) {
continue;
}
const size_t aligned_offset = block.offset + static_cast<size_t>(aligned_addr - block_addr);
const size_t padding = aligned_offset - block.offset;
if (padding > block.size || size > block.size - padding) {
continue;
}
if (!best.valid || aligned_addr < best.address) {
best.chunk_index = chunk_index;
best.block_index = block_index;
best.aligned_offset = aligned_offset;
best.address = aligned_addr;
best.valid = true;
}
}
}
if (!best.valid) {
return false;
}
auto & chunk = chunks_[best.chunk_index];
const free_block block = chunk.free_blocks[best.block_index];
const size_t padding = best.aligned_offset - block.offset;
const size_t alloc_end = best.aligned_offset + size;
const size_t block_end = block.offset + block.size;
chunk.free_blocks.erase(chunk.free_blocks.begin() + best.block_index);
auto insert_it = chunk.free_blocks.begin() + best.block_index;
if (padding != 0) {
insert_it = chunk.free_blocks.insert(insert_it, { block.offset, padding });
++insert_it;
}
if (alloc_end < block_end) {
chunk.free_blocks.insert(insert_it, { alloc_end, block_end - alloc_end });
}
allocation->chunk_base = chunk.base;
allocation->chunk_size = chunk.size;
allocation->base = chunk.base + best.aligned_offset;
allocation->size = size;
return true;
}
void maybe_release_empty_chunk_locked(std::vector<pool_chunk>::iterator chunk_it) {
if (chunk_it->free_blocks.size() != 1) {
return;
}
const auto & block = chunk_it->free_blocks.front();
if (block.offset != 0 || block.size != chunk_it->size) {
return;
}
dealloc_chunk(&*chunk_it);
chunks_.erase(chunk_it);
}
void insert_free_block_locked(pool_chunk & chunk, free_block block) {
auto it = chunk.free_blocks.begin();
while (it != chunk.free_blocks.end() && it->offset < block.offset) {
++it;
}
if (it != chunk.free_blocks.begin()) {
const auto & prev = *(it - 1);
if (prev.offset + prev.size > block.offset) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: overlapping free block at offset %zu size %zu\n", __func__,
block.offset, block.size);
return;
}
}
if (it != chunk.free_blocks.end() && block.offset + block.size > it->offset) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: overlapping next free block at offset %zu size %zu\n", __func__,
block.offset, block.size);
return;
}
it = chunk.free_blocks.insert(it, block);
if (it != chunk.free_blocks.begin()) {
auto prev = it - 1;
if (prev->offset + prev->size == it->offset) {
it->offset = prev->offset;
it->size += prev->size;
it = chunk.free_blocks.erase(prev);
}
}
if (it + 1 != chunk.free_blocks.end() && it->offset + it->size == (it + 1)->offset) {
it->size += (it + 1)->size;
chunk.free_blocks.erase(it + 1);
}
}
std::mutex mutex_;
std::vector<pool_chunk> chunks_;
std::unordered_map<void *, pool_allocation> allocations_;
size_t default_chunk_size_{ 0 };
};
class spine_mem_pool_posix final : public spine_mem_pool_manager {
public:
spine_mem_pool_posix() : spine_mem_pool_manager(0) {}
~spine_mem_pool_posix() override { release_chunks(); }
private:
bool alloc_chunk(size_t min_size, size_t alignment, void * hint_addr, pool_chunk * chunk) override {
(void) hint_addr;
const size_t alloc_alignment = std::max(alignment, sizeof(void *));
void * base = nullptr;
const int rc = posix_memalign(&base, alloc_alignment, min_size);
if (rc != 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: posix_memalign failed for size %zu alignment %zu, rc=%d\n",
__func__, min_size, alloc_alignment, rc);
return false;
}
chunk->base = static_cast<uint8_t *>(base);
chunk->size = min_size;
chunk->fd = -1;
return true;
}
void dealloc_chunk(pool_chunk * chunk) override {
std::free(chunk->base);
clear_chunk(chunk);
}
};
class spine_mem_pool_transparent_hugepage final : public spine_mem_pool_manager {
public:
spine_mem_pool_transparent_hugepage() : spine_mem_pool_manager(SPINE_MEM_POOL_CHUNK_SIZE) {}
~spine_mem_pool_transparent_hugepage() override { release_chunks(); }
private:
bool alloc_chunk(size_t min_size, size_t alignment, void * hint_addr, pool_chunk * chunk) override {
(void) alignment;
size_t chunk_size = 0;
if (!align_up(min_size, default_chunk_size(), &chunk_size)) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: failed to round chunk size for %zu\n", __func__, min_size);
return false;
}
void * map_addr = mmap(hint_addr, chunk_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
if (map_addr == MAP_FAILED) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: mmap failed for chunk size %zu, errno=%d\n", __func__, chunk_size,
errno);
return false;
}
if (madvise(map_addr, chunk_size, MADV_HUGEPAGE) != 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: madvise(MADV_HUGEPAGE) failed for chunk size %zu, errno=%d\n",
__func__, chunk_size, errno);
munmap(map_addr, chunk_size);
return false;
}
chunk->base = static_cast<uint8_t *>(map_addr);
chunk->size = chunk_size;
chunk->fd = -1;
return true;
}
void dealloc_chunk(pool_chunk * chunk) override {
if (chunk->base != nullptr && chunk->size != 0 && munmap(chunk->base, chunk->size) != 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: munmap failed for chunk %p size %zu, errno=%d\n", __func__,
chunk->base, chunk->size, errno);
}
clear_chunk(chunk);
}
};
class spine_mem_pool_hugetlb_1g final : public spine_mem_pool_manager {
public:
spine_mem_pool_hugetlb_1g() : spine_mem_pool_manager(SPINE_MEM_POOL_1G_REGION_SIZE) {}
~spine_mem_pool_hugetlb_1g() override { release_chunks(); }
private:
bool alloc_chunk(size_t min_size, size_t alignment, void * hint_addr, pool_chunk * chunk) override {
(void) alignment;
(void) hint_addr;
size_t region_size = 0;
if (!align_up(min_size, SPINE_MEM_POOL_1G_REGION_SIZE, &region_size)) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: failed to round hugetlb_1g size for %zu\n", __func__, min_size);
return false;
}
const int fd = open(SPINE_MEM_POOL_HUGETLB_1G_DEV, O_RDWR);
if (fd < 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: open(%s) failed, errno=%d\n", __func__,
SPINE_MEM_POOL_HUGETLB_1G_DEV, errno);
return false;
}
hugetlb_1g_region region;
region.size = region_size;
region.flags = HUGETLB_1G_FLAG_REQUIRE_PUD;
if (ioctl(fd, HUGETLB_1G_IOC_ALLOC, &region) < 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: HUGETLB_1G_IOC_ALLOC failed for size %zu, errno=%d\n", __func__,
region_size, errno);
close(fd);
return false;
}
void * map_addr = mmap(nullptr, region.size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
if (map_addr == MAP_FAILED) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: mmap failed for hugetlb_1g size %llu, errno=%d\n", __func__,
static_cast<unsigned long long>(region.size), errno);
ioctl(fd, HUGETLB_1G_IOC_FREE);
close(fd);
return false;
}
chunk->base = static_cast<uint8_t *>(map_addr);
chunk->size = region.size;
chunk->fd = fd;
return true;
}
void dealloc_chunk(pool_chunk * chunk) override {
if (chunk->base != nullptr && chunk->size != 0 && munmap(chunk->base, chunk->size) != 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: munmap failed for hugetlb_1g chunk %p size %zu, errno=%d\n",
__func__, chunk->base, chunk->size, errno);
}
if (chunk->fd >= 0) {
if (ioctl(chunk->fd, HUGETLB_1G_IOC_FREE) < 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: HUGETLB_1G_IOC_FREE failed for chunk %p, errno=%d\n",
__func__, chunk->base, errno);
}
close(chunk->fd);
}
clear_chunk(chunk);
}
};
class spine_mem_pool_shared_mem final : public spine_mem_pool_manager {
public:
spine_mem_pool_shared_mem() : spine_mem_pool_manager(SPINE_SHARE_MEM_POOL_CHUNK_SIZE) {}
~spine_mem_pool_shared_mem() override { release_chunks(); }
private:
bool alloc_chunk(size_t min_size, size_t alignment, void * hint_addr, pool_chunk * chunk) override {
(void) alignment;
if (hint_addr != nullptr) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: shared_mem does not support multiple active chunks\n", __func__);
return false;
}
if (min_size > default_chunk_size()) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: shared_mem request %zu exceeds chunk size %zu\n", __func__,
min_size, default_chunk_size());
return false;
}
const int fd = open(SPINE_MEM_POOL_TCM_SYNC_MEM_DEV, O_RDWR | O_SYNC);
if (fd < 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: open(%s) failed, errno=%d\n", __func__,
SPINE_MEM_POOL_TCM_SYNC_MEM_DEV, errno);
return false;
}
void * map_addr = mmap(nullptr, default_chunk_size(), PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
if (map_addr == MAP_FAILED) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: mmap failed for %s size %zu, errno=%d\n", __func__,
SPINE_MEM_POOL_TCM_SYNC_MEM_DEV, default_chunk_size(), errno);
close(fd);
return false;
}
chunk->base = static_cast<uint8_t *>(map_addr);
chunk->size = default_chunk_size();
chunk->fd = fd;
return true;
}
void dealloc_chunk(pool_chunk * chunk) override {
if (chunk->base != nullptr && chunk->size != 0 && munmap(chunk->base, chunk->size) != 0) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: munmap failed for shared_mem chunk %p size %zu, errno=%d\n",
__func__, chunk->base, chunk->size, errno);
}
if (chunk->fd >= 0) {
close(chunk->fd);
}
clear_chunk(chunk);
}
};
spine_mem_pool_manager & get_spine_mem_pool_manager() {
static std::once_flag pool_once;
static std::unique_ptr<spine_mem_pool_manager> selected_pool;
static spine_mem_pool_backend selected_backend = spine_mem_pool_backend::none;
spine_mem_pool_backend backend = global_spine_env_info.mem_backend;
if (backend == spine_mem_pool_backend::none) {
backend = spine_mem_pool_backend::transparent_hugepage;
}
std::call_once(pool_once, [&]() {
selected_backend = backend;
switch (selected_backend) {
case spine_mem_pool_backend::posix_memalign:
selected_pool = std::make_unique<spine_mem_pool_posix>();
break;
case spine_mem_pool_backend::transparent_hugepage:
selected_pool = std::make_unique<spine_mem_pool_transparent_hugepage>();
break;
case spine_mem_pool_backend::hugetlb_1g:
selected_pool = std::make_unique<spine_mem_pool_hugetlb_1g>();
break;
case spine_mem_pool_backend::none:
selected_backend = spine_mem_pool_backend::transparent_hugepage;
selected_pool = std::make_unique<spine_mem_pool_transparent_hugepage>();
break;
}
});
if (backend != selected_backend) {
GGML_LOG_ERROR(
"CPU_RISCV64_SPACEMIT: %s: mem pool backend is process-global and mutually exclusive, requested=%d but "
"selected=%d\n",
__func__, static_cast<int>(backend), static_cast<int>(selected_backend));
}
if (selected_pool) {
return *selected_pool;
}
throw std::bad_alloc();
}
spine_mem_pool_manager & get_spine_mem_pool_shared_mem_manager() {
static std::once_flag shared_mem_pool_once;
static std::unique_ptr<spine_mem_pool_shared_mem> shared_mem_pool;
std::call_once(shared_mem_pool_once, [&]() { shared_mem_pool = std::make_unique<spine_mem_pool_shared_mem>(); });
if (shared_mem_pool) {
return *shared_mem_pool;
}
throw std::bad_alloc();
}
} // namespace
bool spine_mem_pool_tcm_init(spine_mem_pool_tcm_info * info) noexcept {
if (info == nullptr) {
return false;
}
*info = {};
if (spine_tcm_open_handle(NULL) != 0 || !spine_tcm_is_available()) {
return false;
}
spine_tcm_mem_info_t mem_info;
if (spine_tcm_mem_info(&mem_info) != 0) {
return false;
}
info->available = true;
info->blk_size = mem_info.blk_size;
info->blk_num = mem_info.blk_num;
info->is_fake_tcm = mem_info.is_fake_tcm != 0;
return true;
}
void * spine_mem_pool_tcm_mem_get(int cpu_id) noexcept {
return spine_tcm_mem_get(cpu_id);
}
void * spine_mem_pool_tcm_mem_wait(int cpu_id) noexcept {
return spine_tcm_mem_try_wait(cpu_id, 1000 * 1000);
}
int spine_mem_pool_tcm_mem_release(int cpu_id) noexcept {
return spine_tcm_mem_release(cpu_id);
}
void * spine_mem_pool_alloc(size_t size, size_t alignment) noexcept {
try {
return get_spine_mem_pool_manager().alloc(size, alignment);
} catch (const std::bad_alloc &) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: bad_alloc while allocating size %zu\n", __func__, size);
return nullptr;
}
}
void * spine_mem_pool_shared_mem_alloc(size_t size, size_t alignment) noexcept {
try {
return get_spine_mem_pool_shared_mem_manager().alloc(size, alignment);
} catch (const std::bad_alloc &) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: bad_alloc while allocating shared memory size %zu\n", __func__, size);
return nullptr;
}
}
void spine_mem_pool_free(void * base) noexcept {
try {
get_spine_mem_pool_manager().free(base);
} catch (const std::bad_alloc &) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: bad_alloc while freeing allocation %p\n", __func__, base);
}
}
void spine_mem_pool_shared_mem_free(void * base) noexcept {
try {
get_spine_mem_pool_shared_mem_manager().free(base);
} catch (const std::bad_alloc &) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: bad_alloc while freeing shared allocation %p\n", __func__, base);
}
}
} // namespace ggml::cpu::riscv64_spacemit
extern "C" {
void * ggml_backend_cpu_riscv64_spacemit_alloc_shared(size_t size, size_t alignment) {
void * result = ggml::cpu::riscv64_spacemit::spine_mem_pool_shared_mem_alloc(size, alignment);
if (result == nullptr) {
GGML_LOG_ERROR("CPU_RISCV64_SPACEMIT: %s: failed to allocate shared memory size %zu alignment %zu\n", __func__,
size, alignment);
}
return result;
}
void ggml_backend_cpu_riscv64_spacemit_free_shared(void * ptr) {
ggml::cpu::riscv64_spacemit::spine_mem_pool_shared_mem_free(ptr);
}
}
engine/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h
+32
View File
@@ -0,0 +1,32 @@
#pragma once
#include <cstddef>
#include <cstdint>
namespace ggml::cpu::riscv64_spacemit {
enum class spine_mem_pool_backend : uint8_t {
none,
posix_memalign,
transparent_hugepage,
hugetlb_1g,
};
struct spine_mem_pool_tcm_info {
bool available{ false };
size_t blk_size{ 0 };
size_t blk_num{ 0 };
bool is_fake_tcm{ false };
};
bool spine_mem_pool_tcm_init(spine_mem_pool_tcm_info * info) noexcept;
void * spine_mem_pool_tcm_mem_get(int cpu_id) noexcept;
void * spine_mem_pool_tcm_mem_wait(int cpu_id) noexcept;
int spine_mem_pool_tcm_mem_release(int cpu_id) noexcept;
void * spine_mem_pool_alloc(size_t size, size_t alignment) noexcept;
void * spine_mem_pool_shared_mem_alloc(size_t size, size_t alignment) noexcept;
void spine_mem_pool_free(void * base) noexcept;
void spine_mem_pool_shared_mem_free(void * base) noexcept;
} // namespace ggml::cpu::riscv64_spacemit
engine/ggml/src/ggml-cpu/spacemit/spine_tcm.h
+409
View File
@@ -0,0 +1,409 @@
#ifndef SPINE_TCM_PUBLIC_H_
#define SPINE_TCM_PUBLIC_H_
/*
* spine_tcm public API
*
* Usage:
* 1. Direct link mode
* Define SPINE_TCM_DIRECT_LINK and link against libspine_tcm.so.
*
* if (spine_tcm_is_available()) {
* void *buffer = spine_tcm_mem_get(0);
* spine_tcm_mem_free(0);
* }
*
* 2. Header-only loader mode
* Include this header without linking libspine_tcm.so. The loader first
* tries to reuse a process-global spine_tcm instance and falls back to
* dlopen("libspine_tcm.so") when needed.
*
* spine_tcm_open_handle(NULL); // optional pre-bind
* if (spine_tcm_is_available()) {
* void *buffer = spine_tcm_mem_get(0);
* spine_tcm_mem_free(0);
* }
*/
#include <stddef.h>
#include <stdint.h>
#include <string.h>
#if !defined(SPINE_TCM_BUILD_SHARED) && !defined(SPINE_TCM_DIRECT_LINK)
# include <dlfcn.h>
#endif
#ifdef __cplusplus
extern "C" {
#endif
#if defined(_WIN32)
# if defined(SPINE_TCM_BUILD_SHARED)
# define SPINE_TCM_API __declspec(dllexport)
# else
# define SPINE_TCM_API __declspec(dllimport)
# endif
#else
# define SPINE_TCM_API __attribute__((visibility("default")))
#endif
typedef struct spine_tcm_mem_info {
size_t blk_size;
size_t blk_num;
int is_fake_tcm;
} spine_tcm_mem_info_t;
typedef struct spine_tcm_block_info {
int id;
void * va;
size_t size;
uint64_t phys_addr;
uint64_t cpu_affinity_mask;
int owner_tid;
int is_acquired;
} spine_tcm_block_info_t;
/* Shared-library runtime ABI exported by libspine_tcm.so. */
SPINE_TCM_API const char * spine_tcm_runtime_version(void);
SPINE_TCM_API int spine_tcm_runtime_is_available(void);
SPINE_TCM_API int spine_tcm_runtime_layout_info(spine_tcm_mem_info_t * info);
SPINE_TCM_API int spine_tcm_runtime_mem_info(int id, spine_tcm_block_info_t * info);
SPINE_TCM_API void * spine_tcm_runtime_mem_get(int id);
SPINE_TCM_API int spine_tcm_runtime_mem_free(int id);
SPINE_TCM_API void * spine_tcm_runtime_mem_try_wait(int id, size_t timeout_us);
SPINE_TCM_API int spine_tcm_runtime_mem_release(int id);
SPINE_TCM_API int spine_tcm_runtime_mem_force_release(int id);
SPINE_TCM_API int spine_tcm_runtime_mem_query(int id);
#if defined(SPINE_TCM_DIRECT_LINK)
/* Optional no-op in direct-link mode. */
static inline int spine_tcm_open_handle(const char * so_path) {
(void) so_path;
return 0;
}
static inline const char * spine_tcm_version(void) {
return spine_tcm_runtime_version();
}
/* Returns 1 when the runtime driver is available, otherwise 0. */
static inline int spine_tcm_is_available(void) {
return spine_tcm_runtime_is_available();
}
/* Returns runtime memory geometry and whether the current backend is fake TCM. */
static inline int spine_tcm_mem_info(spine_tcm_mem_info_t * info) {
return spine_tcm_runtime_layout_info(info);
}
/* Returns per-block runtime metadata for the given TCM id. */
static inline int spine_tcm_block_info(int id, spine_tcm_block_info_t * info) {
return spine_tcm_runtime_mem_info(id, info);
}
/* Returns a cached buffer for the given TCM id, or NULL on failure. */
static inline void * spine_tcm_mem_get(int id) {
return spine_tcm_runtime_mem_get(id);
}
/* Releases one reference acquired by spine_tcm_mem_get(id). */
static inline int spine_tcm_mem_free(int id) {
return spine_tcm_runtime_mem_free(id);
}
/* Waits for a TCM block handoff and returns the driver-owned buffer when available. */
static inline void * spine_tcm_mem_try_wait(int id, size_t over_time) {
return spine_tcm_runtime_mem_try_wait(id, over_time);
}
/* Releases a buffer acquired by spine_tcm_mem_try_wait(id, over_time). */
static inline int spine_tcm_mem_release(int id) {
return spine_tcm_runtime_mem_release(id);
}
/* Forces a release for the given TCM id when the backend supports it. */
static inline int spine_tcm_mem_force_release(int id) {
return spine_tcm_runtime_mem_force_release(id);
}
/* Returns whether the given TCM id is currently acquired. */
static inline int spine_tcm_mem_query(int id) {
return spine_tcm_runtime_mem_query(id);
}
#elif !defined(SPINE_TCM_BUILD_SHARED)
typedef struct spine_tcm_handle {
void * module_handle;
int use_global_scope;
int owns_module_handle;
const char * (*runtime_version)(void);
int (*runtime_is_available)(void);
int (*runtime_layout_info)(spine_tcm_mem_info_t * info);
int (*runtime_mem_info)(int id, spine_tcm_block_info_t * info);
void * (*runtime_mem_get)(int id);
int (*runtime_mem_free)(int id);
void * (*runtime_mem_try_wait)(int id, size_t over_time);
int (*runtime_mem_release)(int id);
int (*runtime_mem_force_release)(int id);
int (*runtime_mem_query)(int id);
} spine_tcm_handle_t;
static inline spine_tcm_handle_t * spine_tcm_default_handle(void) {
static spine_tcm_handle_t handle = { 0 };
return &handle;
}
static inline void spine_tcm_handle_reset(spine_tcm_handle_t * handle) {
if (handle != NULL) {
memset(handle, 0, sizeof(*handle));
}
}
static inline int spine_tcm_handle_bind(spine_tcm_handle_t * handle) {
void * symbol_scope = handle->use_global_scope ? RTLD_DEFAULT : handle->module_handle;
handle->runtime_version = (const char * (*) (void) ) dlsym(symbol_scope, "spine_tcm_runtime_version");
handle->runtime_is_available = (int (*)(void)) dlsym(symbol_scope, "spine_tcm_runtime_is_available");
handle->runtime_layout_info =
(int (*)(spine_tcm_mem_info_t *)) dlsym(symbol_scope, "spine_tcm_runtime_layout_info");
handle->runtime_mem_info =
(int (*)(int, spine_tcm_block_info_t *)) dlsym(symbol_scope, "spine_tcm_runtime_mem_info");
handle->runtime_mem_get = (void * (*) (int) ) dlsym(symbol_scope, "spine_tcm_runtime_mem_get");
handle->runtime_mem_free = (int (*)(int)) dlsym(symbol_scope, "spine_tcm_runtime_mem_free");
handle->runtime_mem_try_wait = (void * (*) (int, size_t)) dlsym(symbol_scope, "spine_tcm_runtime_mem_try_wait");
handle->runtime_mem_release = (int (*)(int)) dlsym(symbol_scope, "spine_tcm_runtime_mem_release");
handle->runtime_mem_force_release = (int (*)(int)) dlsym(symbol_scope, "spine_tcm_runtime_mem_force_release");
handle->runtime_mem_query = (int (*)(int)) dlsym(symbol_scope, "spine_tcm_runtime_mem_query");
return handle->runtime_version != NULL && handle->runtime_is_available != NULL &&
handle->runtime_layout_info != NULL && handle->runtime_mem_info != NULL &&
handle->runtime_mem_get != NULL && handle->runtime_mem_free != NULL &&
handle->runtime_mem_try_wait != NULL && handle->runtime_mem_release != NULL &&
handle->runtime_mem_force_release != NULL && handle->runtime_mem_query != NULL ?
0 :
-1;
}
/*
* Try to bind against an already-loaded process-global spine_tcm instance.
* The shared library exports spine_tcm_runtime_marker only for this probe.
*/
static inline int spine_tcm_try_bind_global(spine_tcm_handle_t * handle) {
if (dlsym(RTLD_DEFAULT, "spine_tcm_runtime_marker") == NULL) {
return -1;
}
handle->use_global_scope = 1;
return spine_tcm_handle_bind(handle);
}
/*
* Optional pre-bind entry point.
*
* Behavior:
* - Reuses an already-loaded global spine_tcm instance when available.
* - Otherwise loads the shared library from so_path or the default soname.
* - Repeated calls are safe and return 0 after the first successful bind.
*/
static inline int spine_tcm_open_handle(const char * so_path) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
const char * library = (so_path != NULL && so_path[0] != '\0') ? so_path : "libspine_tcm.so";
if (resolved->module_handle != NULL || resolved->use_global_scope) {
return 0;
}
if (spine_tcm_try_bind_global(resolved) == 0) {
return 0;
}
spine_tcm_handle_reset(resolved);
resolved->module_handle = dlopen(library, RTLD_LAZY | RTLD_GLOBAL);
resolved->owns_module_handle = resolved->module_handle != NULL ? 1 : 0;
if (resolved->module_handle == NULL) {
spine_tcm_handle_reset(resolved);
return -1;
}
if (spine_tcm_handle_bind(resolved) != 0) {
if (resolved->owns_module_handle) {
dlclose(resolved->module_handle);
}
spine_tcm_handle_reset(resolved);
return -1;
}
return 0;
}
/* Returns 1 when the runtime driver is available, otherwise 0. */
static inline int spine_tcm_is_available(void) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) || resolved->runtime_is_available == NULL) {
return 0;
}
return resolved->runtime_is_available();
}
/* Returns runtime memory geometry and whether the current backend is fake TCM. */
static inline int spine_tcm_mem_info(spine_tcm_mem_info_t * info) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) || resolved->runtime_layout_info == NULL) {
return -1;
}
return resolved->runtime_layout_info(info);
}
static inline const char * spine_tcm_version(void) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) || resolved->runtime_version == NULL) {
return "unknown";
}
return resolved->runtime_version();
}
/* Returns per-block runtime metadata for the given TCM id. */
static inline int spine_tcm_block_info(int id, spine_tcm_block_info_t * info) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) || resolved->runtime_mem_info == NULL) {
return -1;
}
return resolved->runtime_mem_info(id, info);
}
/* Returns a cached buffer for the given TCM id, or NULL on failure. */
static inline void * spine_tcm_mem_get(int id) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
return NULL;
}
if (resolved->runtime_mem_get == NULL) {
return NULL;
}
return resolved->runtime_mem_get(id);
}
/* Releases one reference acquired by spine_tcm_mem_get(id). */
static inline int spine_tcm_mem_free(int id) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) || resolved->runtime_mem_free == NULL) {
return -1;
}
return resolved->runtime_mem_free(id);
}
/* Waits for a TCM block handoff and returns the driver-owned buffer when available. */
static inline void * spine_tcm_mem_try_wait(int id, size_t over_time) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
return NULL;
}
if (resolved->runtime_mem_try_wait == NULL) {
return NULL;
}
return resolved->runtime_mem_try_wait(id, over_time);
}
/* Releases a buffer acquired by spine_tcm_mem_try_wait(id, over_time). */
static inline int spine_tcm_mem_release(int id) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) || resolved->runtime_mem_release == NULL) {
return -1;
}
return resolved->runtime_mem_release(id);
}
/* Forces a release for the given TCM id when the backend supports it. */
static inline int spine_tcm_mem_force_release(int id) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) ||
resolved->runtime_mem_force_release == NULL) {
return -1;
}
return resolved->runtime_mem_force_release(id);
}
/* Returns whether the given TCM id is currently acquired. */
static inline int spine_tcm_mem_query(int id) {
spine_tcm_handle_t * resolved = spine_tcm_default_handle();
if (resolved->module_handle == NULL && !resolved->use_global_scope) {
(void) spine_tcm_open_handle(NULL);
}
if ((resolved->module_handle == NULL && !resolved->use_global_scope) || resolved->runtime_mem_query == NULL) {
return -1;
}
return resolved->runtime_mem_query(id);
}
#else
static inline const char * spine_tcm_version(void) {
return spine_tcm_runtime_version();
}
#endif
#define SPINE_TCM_VERSION (spine_tcm_version())
#ifdef __cplusplus
}
#endif
#endif
engine/ggml/src/ggml-cpu/traits.cpp
+36
View File
@@ -0,0 +1,36 @@
#include "traits.h"
#include "ggml-backend-impl.h"
#include "ggml-backend.h"
namespace ggml::cpu {
tensor_traits::~tensor_traits() {}
extra_buffer_type::~extra_buffer_type() {}
} // namespace ggml::cpu
bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) {
for (auto extra : ggml_backend_cpu_get_extra_buffer_types()) {
if (extra && extra->context) {
auto buf_extra = (ggml::cpu::extra_buffer_type *) extra->context;
auto tensor_traits = buf_extra->get_tensor_traits(op);
if (tensor_traits && tensor_traits->compute_forward(params, op)) {
return true;
}
}
}
return false;
}
bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size) {
for (auto extra : ggml_backend_cpu_get_extra_buffer_types()) {
if (extra && extra->context) {
auto buf_extra = (ggml::cpu::extra_buffer_type *) extra->context;
auto tensor_traits = buf_extra->get_tensor_traits(op);
if (tensor_traits && tensor_traits->work_size(n_threads, op, *size)) {
return true;
}
}
}
return false;
}
engine/ggml/src/ggml-cpu/traits.h
+38
View File
@@ -0,0 +1,38 @@
#pragma once
#include "ggml-backend-impl.h"
#include "ggml-cpu-impl.h"
#include "ggml.h"
#ifdef __cplusplus
# include <vector>
extern "C" {
#endif
// return true if op part of extra "accelerator"
bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op);
bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size);
#ifdef __cplusplus
}
namespace ggml::cpu {
// register in tensor->extra
class tensor_traits {
public:
virtual ~tensor_traits();
virtual bool work_size(int n_threads, const struct ggml_tensor * op, size_t & size) = 0;
virtual bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) = 0;
};
class extra_buffer_type {
public:
virtual ~extra_buffer_type();
virtual bool supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) = 0;
virtual tensor_traits * get_tensor_traits(const struct ggml_tensor * op) = 0;
};
} // namespace ggml::cpu
// implemented in ggml-cpu.cpp.
std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffer_types();
#endif
engine/ggml/src/ggml-cpu/unary-ops.cpp
+337
View File
@@ -0,0 +1,337 @@
#include "unary-ops.h"
static inline float op_abs(float x) {
return fabsf(x);
}
static inline float op_sgn(float x) {
return (x > 0.f) ? 1.f : ((x < 0.f) ? -1.f : 0.f);
}
static inline float op_neg(float x) {
return -x;
}
static inline float op_step(float x) {
return (x > 0.f) ? 1.f : 0.f;
}
static inline float op_tanh(float x) {
return tanhf(x);
}
static inline float op_elu(float x) {
return (x > 0.f) ? x : expm1f(x);
}
static inline float op_relu(float x) {
return (x > 0.f) ? x : 0.f;
}
static inline float op_sigmoid(float x) {
return 1.f / (1.f + expf(-x));
}
static inline float op_hardsigmoid(float x) {
return fminf(1.0f, fmaxf(0.0f, (x + 3.0f) / 6.0f));
}
static inline float op_exp(float x) {
return expf(x);
}
static inline float op_hardswish(float x) {
return x * fminf(1.0f, fmaxf(0.0f, (x + 3.0f) / 6.0f));
}
static inline float op_sqr(float x) {
return x * x;
}
static inline float op_sqrt(float x) {
return sqrtf(x);
}
static inline float op_xielu(float x, float alpha_n, float alpha_p, float beta, float eps) {
if (x > 0.0f) {
return alpha_p * x * x + beta * x;
} else {
const float min_x_eps = fminf(x, eps);
return (expm1f(min_x_eps) - x) * alpha_n + beta * x;
}
}
static inline float op_sin(float x) {
return sinf(x);
}
static inline float op_cos(float x) {
return cosf(x);
}
static inline float op_log(float x) {
return logf(x);
}
static inline float op_expm1(float x) {
return expf(x) - 1.0f;
}
static inline float op_softplus(float x) {
return (x > 20.0f) ? x : logf(1.0f + expf(x));
}
static inline float op_floor(float x) {
return floorf(x);
}
static inline float op_ceil(float x) {
return ceilf(x);
}
static inline float op_round(float x) {
return roundf(x);
}
static inline float op_trunc(float x) {
return truncf(x);
}
template <float (*op)(float), typename src0_t, typename dst_t>
static inline void vec_unary_op(int64_t n, dst_t * y, const src0_t * x) {
constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
constexpr auto f32_to_dst = type_conversion_table<dst_t >::from_f32;
for (int i = 0; i < n; i++) {
y[i] = f32_to_dst(op(src0_to_f32(x[i])));
}
}
template <float (*op)(float), typename src0_t, typename dst_t>
static void apply_unary_op(const ggml_compute_params * params, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
GGML_ASSERT(ggml_is_contiguous_rows(src0) && ggml_is_contiguous_rows(dst) && ggml_are_same_shape(src0, dst));
GGML_TENSOR_UNARY_OP_LOCALS
GGML_ASSERT( nb0 == sizeof(dst_t));
GGML_ASSERT(nb00 == sizeof(src0_t));
const auto [ir0, ir1] = get_thread_range(params, src0);
for (int64_t ir = ir0; ir < ir1; ++ir) {
const int64_t i03 = ir/(ne02*ne01);
const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
dst_t * dst_ptr = (dst_t *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 );
const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
vec_unary_op<op>(ne0, dst_ptr, src0_ptr);
}
}
// TODO: Use the 'traits' lookup table (for type conversion fns), instead of a mass of 'if' conditions with long templates
template <float (*op)(float)>
static void unary_op(const ggml_compute_params * params, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
/* */ if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { // all f32
apply_unary_op<op, float, float>(params, dst);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { // all f16
apply_unary_op<op, ggml_fp16_t, ggml_fp16_t>(params, dst);
} else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16
apply_unary_op<op, ggml_bf16_t, ggml_bf16_t>(params, dst);
} else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_F32) {
apply_unary_op<op, ggml_bf16_t, float>(params, dst);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
apply_unary_op<op, ggml_fp16_t, float>(params, dst);
} else {
fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s\n", __func__,
ggml_type_name(dst->type), ggml_type_name(src0->type));
GGML_ABORT("fatal error");
}
}
template <float (*op)(float, ggml_tensor *)>
static void unary_op_params(const ggml_compute_params * params, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
/* */ if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { // all f32
apply_unary_op<op, float, float>(params, dst);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { // all f16
apply_unary_op<op, ggml_fp16_t, ggml_fp16_t>(params, dst);
} else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16
apply_unary_op<op, ggml_bf16_t, ggml_bf16_t>(params, dst);
} else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_F32) {
apply_unary_op<op, ggml_bf16_t, float>(params, dst);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
apply_unary_op<op, ggml_fp16_t, float>(params, dst);
} else {
fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s\n", __func__,
ggml_type_name(dst->type), ggml_type_name(src0->type));
GGML_ABORT("fatal error");
}
}
// Extend vec_unary_op to support functors
template <typename Op, typename src0_t, typename dst_t>
static inline void vec_unary_op_functor(int64_t n, dst_t * y, const src0_t * x, Op op) {
constexpr auto src0_to_f32 = type_conversion_table<src0_t>::to_f32;
constexpr auto f32_to_dst = type_conversion_table<dst_t >::from_f32;
for (int i = 0; i < n; i++) {
y[i] = f32_to_dst(op(src0_to_f32(x[i])));
}
}
// Extend apply_unary_op to support functors
template <typename Op, typename src0_t, typename dst_t>
static void apply_unary_op_functor(const ggml_compute_params * params, ggml_tensor * dst, Op op) {
const ggml_tensor * src0 = dst->src[0];
GGML_ASSERT(ggml_is_contiguous_1(src0) && ggml_is_contiguous_1(dst) && ggml_are_same_shape(src0, dst));
GGML_TENSOR_UNARY_OP_LOCALS
GGML_ASSERT( nb0 == sizeof(dst_t));
GGML_ASSERT(nb00 == sizeof(src0_t));
const auto [ir0, ir1] = get_thread_range(params, src0);
for (int64_t ir = ir0; ir < ir1; ++ir) {
const int64_t i03 = ir/(ne02*ne01);
const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
dst_t * dst_ptr = (dst_t *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 );
const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
vec_unary_op_functor(ne0, dst_ptr, src0_ptr, op);
}
}
// Generic dispatcher for functors
template <typename Op>
static void unary_op_functor(const ggml_compute_params * params, ggml_tensor * dst, Op op) {
const ggml_tensor * src0 = dst->src[0];
/* */ if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { // all f32
apply_unary_op_functor<Op, float, float>(params, dst, op);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { // all f16
apply_unary_op_functor<Op, ggml_fp16_t, ggml_fp16_t>(params, dst, op);
} else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16
apply_unary_op_functor<Op, ggml_bf16_t, ggml_bf16_t>(params, dst, op);
} else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_F32) {
apply_unary_op_functor<Op, ggml_bf16_t, float>(params, dst, op);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
apply_unary_op_functor<Op, ggml_fp16_t, float>(params, dst, op);
} else {
fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s\n", __func__,
ggml_type_name(dst->type), ggml_type_name(src0->type));
GGML_ABORT("fatal error");
}
}
void ggml_compute_forward_abs(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_abs>(params, dst);
}
void ggml_compute_forward_sgn(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_sgn>(params, dst);
}
void ggml_compute_forward_neg(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_neg>(params, dst);
}
void ggml_compute_forward_step(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_step>(params, dst);
}
void ggml_compute_forward_tanh(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_tanh>(params, dst);
}
void ggml_compute_forward_elu(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_elu>(params, dst);
}
void ggml_compute_forward_relu(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_relu>(params, dst);
}
void ggml_compute_forward_sigmoid(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_sigmoid>(params, dst);
}
void ggml_compute_forward_hardsigmoid(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_hardsigmoid>(params, dst);
}
void ggml_compute_forward_exp(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_exp>(params, dst);
}
void ggml_compute_forward_hardswish(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_hardswish>(params, dst);
}
void ggml_compute_forward_sqr(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_sqr>(params, dst);
}
void ggml_compute_forward_sqrt(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_sqrt>(params, dst);
}
void ggml_compute_forward_sin(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_sin>(params, dst);
}
void ggml_compute_forward_cos(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_cos>(params, dst);
}
void ggml_compute_forward_log(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_log>(params, dst);
}
void ggml_compute_forward_expm1(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_expm1>(params, dst);
}
void ggml_compute_forward_softplus(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_softplus>(params, dst);
}
void ggml_compute_forward_floor(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_floor>(params, dst);
}
void ggml_compute_forward_ceil(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_ceil>(params, dst);
}
void ggml_compute_forward_round(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_round>(params, dst);
}
void ggml_compute_forward_trunc(const ggml_compute_params * params, ggml_tensor * dst) {
unary_op<op_trunc>(params, dst);
}
void ggml_compute_forward_xielu(const ggml_compute_params * params, ggml_tensor * dst) {
const float alpha_n = ggml_get_op_params_f32(dst, 1);
const float alpha_p = ggml_get_op_params_f32(dst, 2);
const float beta = ggml_get_op_params_f32(dst, 3);
const float eps = ggml_get_op_params_f32(dst, 4);
const auto xielu_op_params = [alpha_n, alpha_p, beta, eps](float f) {
return op_xielu(f, alpha_n, alpha_p, beta, eps);
};
unary_op_functor(params, dst, xielu_op_params);
}
engine/ggml/src/ggml-cpu/unary-ops.h
+35
View File
@@ -0,0 +1,35 @@
#pragma once
#include "common.h"
#ifdef __cplusplus
extern "C" {
#endif
void ggml_compute_forward_abs(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sgn(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_neg(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_step(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_tanh(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_elu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sigmoid(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_hardsigmoid(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_exp(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_hardswish(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sqr(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sqrt(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_sin(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_log(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_expm1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_softplus(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_floor(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_ceil(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_round(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_trunc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_xielu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
#ifdef __cplusplus
}
#endif
engine/ggml/src/ggml-cpu/vec.cpp
+613
View File
@@ -0,0 +1,613 @@
#include "vec.h"
#include <cassert>
// precomputed gelu table for f16 (128 KB)
ggml_fp16_t ggml_table_gelu_f16[1 << 16];
// precomputed quick gelu table for f16 (128 KB)
ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc) {
assert(nrc == 1);
GGML_UNUSED(nrc);
GGML_UNUSED(bx);
GGML_UNUSED(by);
GGML_UNUSED(bs);
#if defined(GGML_SIMD)
float sumf = 0.0f;
#if defined(__ARM_FEATURE_SVE)
const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
const int ggml_f32_epr = sve_register_length / 32;//8;//svcntw(); // SVE128:4, SVE256:8, SVE512:16
const int ggml_f32_step = 8 * ggml_f32_epr; // choose 8 SVE registers
const int np = (n & ~(ggml_f32_step - 1));
svfloat32_t sum1 = svdup_n_f32(0.0f);
svfloat32_t sum2 = svdup_n_f32(0.0f);
svfloat32_t sum3 = svdup_n_f32(0.0f);
svfloat32_t sum4 = svdup_n_f32(0.0f);
svfloat32_t sum5 = svdup_n_f32(0.0f);
svfloat32_t sum6 = svdup_n_f32(0.0f);
svfloat32_t sum7 = svdup_n_f32(0.0f);
svfloat32_t sum8 = svdup_n_f32(0.0f);
svfloat32_t ax1,ax2,ax3,ax4,ax5,ax6,ax7,ax8;
svfloat32_t ay1,ay2,ay3,ay4,ay5,ay6,ay7,ay8;
for (int i = 0; i < np; i += ggml_f32_step) {
ax1 = GGML_F32_VEC_LOAD(x + i);
ay1 = GGML_F32_VEC_LOAD(y + i);
sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
sum2 = GGML_F32_VEC_FMA(sum2, ax2, ay2);
ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
sum3 = GGML_F32_VEC_FMA(sum3, ax3, ay3);
ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
sum4 = GGML_F32_VEC_FMA(sum4, ax4, ay4);
ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
sum5 = GGML_F32_VEC_FMA(sum5, ax5, ay5);
ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
sum6 = GGML_F32_VEC_FMA(sum6, ax6, ay6);
ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
sum7 = GGML_F32_VEC_FMA(sum7, ax7, ay7);
ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
sum8 = GGML_F32_VEC_FMA(sum8, ax8, ay8);
}
// leftovers
// Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop
const int np2 = (n & ~(ggml_f32_epr - 1));
for (int i = np; i < np2; i += ggml_f32_epr) {
ax1 = GGML_F32_VEC_LOAD(x + i);
ay1 = GGML_F32_VEC_LOAD(y + i);
sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
}
// maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
if (np2 < n) {
svbool_t pg = svwhilelt_b32(np2, n);
ax1 = svld1_f32(pg, x + np2);
ay1 = svld1_f32(pg, y + np2);
sum1 = svmad_f32_m(pg, ax1, ay1, sum1);
}
// reduce sum1,sum2 to sum1
GGML_F32_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8);
#elif defined(__riscv_v_intrinsic)
int vl = __riscv_vsetvlmax_e32m8();
vfloat32m1_t vs = __riscv_vfmv_v_f_f32m1(0.0f, 1);
vfloat32m8_t vsum;
vfloat32m8_t ax;
vfloat32m8_t ay;
vsum = __riscv_vfmv_v_f_f32m8_tu(vsum, 0.0f, vl);
for (int i = 0; i < n; i += vl) {
vl = __riscv_vsetvl_e32m8(n - i);
ax = __riscv_vle32_v_f32m8_tu(ax, &x[i], vl);
ay = __riscv_vle32_v_f32m8_tu(ay, &y[i], vl);
vsum = __riscv_vfmacc_vv_f32m8_tu(vsum, ax, ay, vl);
}
vl = __riscv_vsetvlmax_e32m8();
vs = __riscv_vfredusum_vs_f32m8_f32m1(vsum, vs, vl);
sumf += __riscv_vfmv_f_s_f32m1_f32(vs);
#else
const int np = (n & ~(GGML_F32_STEP - 1));
GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
GGML_F32_VEC ax[GGML_F32_ARR];
GGML_F32_VEC ay[GGML_F32_ARR];
for (int i = 0; i < np; i += GGML_F32_STEP) {
for (int j = 0; j < GGML_F32_ARR; j++) {
ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], ay[j]);
}
}
// reduce sum0..sum3 to sum0
GGML_F32_VEC_REDUCE(sumf, sum);
// leftovers
for (int i = np; i < n; ++i) {
sumf += x[i]*y[i];
}
#endif
#else
// scalar
ggml_float sumf = 0.0;
for (int i = 0; i < n; ++i) {
sumf += (ggml_float)(x[i]*y[i]);
}
#endif
*s = sumf;
}
void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc) {
assert(nrc == 1);
GGML_UNUSED(nrc);
GGML_UNUSED(bx);
GGML_UNUSED(by);
GGML_UNUSED(bs);
int i = 0;
ggml_float sumf = 0;
#if defined(__AVX512BF16__)
__m512 c1 = _mm512_setzero_ps();
__m512 c2 = _mm512_setzero_ps();
for (; i + 64 <= n; i += 64) {
c1 = _mm512_dpbf16_ps(c1, m512bh(_mm512_loadu_si512((x + i))),
m512bh(_mm512_loadu_si512((y + i))));
c2 = _mm512_dpbf16_ps(c2, m512bh(_mm512_loadu_si512((x + i + 32))),
m512bh(_mm512_loadu_si512((y + i + 32))));
}
sumf += (ggml_float)_mm512_reduce_add_ps(c1);
sumf += (ggml_float)_mm512_reduce_add_ps(c2);
#elif defined(__AVX512F__)
#define LOAD(p) _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm256_loadu_si256((const __m256i *)(p))), 16))
__m512 c1 = _mm512_setzero_ps();
__m512 c2 = _mm512_setzero_ps();
for (; i + 32 <= n; i += 32) {
c1 = _mm512_add_ps(_mm512_mul_ps(LOAD(x + i), LOAD(y + i)), c1);
c2 = _mm512_add_ps(_mm512_mul_ps(LOAD(x + i + 16), LOAD(y + i + 16)), c2);
}
sumf += (ggml_float)_mm512_reduce_add_ps(c1);
sumf += (ggml_float)_mm512_reduce_add_ps(c2);
#undef LOAD
#elif defined(__AVX2__) || defined(__AVX__)
#if defined(__AVX2__)
#define LOAD(p) _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16))
#else
#define LOAD(p) _mm256_castsi256_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16)), (_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_bsrli_si128(_mm_loadu_si128((const __m128i *)(p)), 8)), 16)), 1))
#endif
__m256 c1 = _mm256_setzero_ps();
__m256 c2 = _mm256_setzero_ps();
__m256 c3 = _mm256_setzero_ps();
__m256 c4 = _mm256_setzero_ps();
for (; i + 32 <= n; i += 32) {
c1 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i), LOAD(y + i)), c1);
c2 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 8), LOAD(y + i + 8)), c2);
c3 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 16), LOAD(y + i + 16)), c3);
c4 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 24), LOAD(y + i + 24)), c4);
}
__m128 g;
c1 = _mm256_add_ps(_mm256_add_ps(c1, c3),
_mm256_add_ps(c2, c4));
g = _mm_add_ps(_mm256_extractf128_ps(c1, 1),
_mm256_castps256_ps128(c1));
g = _mm_add_ps(g, _mm_movehl_ps(g, g));
g = _mm_add_ss(g, _mm_movehdup_ps(g));
sumf += (ggml_float)_mm_cvtss_f32(g);
#undef LOAD
#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfbfwma)
size_t vl = __riscv_vsetvlmax_e32m4();
// initialize accumulators to all zeroes
vfloat32m4_t vsum0 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
vfloat32m4_t vsum1 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
// calculate step size
const size_t epr = __riscv_vsetvlmax_e16m2();
const size_t step = epr * 2;
const int np = (n & ~(step - 1));
// unroll by 2
for (; i < np; i += step) {
vbfloat16m2_t ax0 = __riscv_vle16_v_bf16m2((const __bf16 *)&x[i], epr);
vbfloat16m2_t ay0 = __riscv_vle16_v_bf16m2((const __bf16 *)&y[i], epr);
vsum0 = __riscv_vfwmaccbf16_vv_f32m4(vsum0, ax0, ay0, epr);
__asm__ __volatile__ ("" ::: "memory");
vbfloat16m2_t ax1 = __riscv_vle16_v_bf16m2((const __bf16 *)&x[i + epr], epr);
vbfloat16m2_t ay1 = __riscv_vle16_v_bf16m2((const __bf16 *)&y[i + epr], epr);
vsum1 = __riscv_vfwmaccbf16_vv_f32m4(vsum1, ax1, ay1, epr);
__asm__ __volatile__ ("" ::: "memory");
}
// accumulate in 1 register
vsum0 = __riscv_vfadd_vv_f32m4(vsum0, vsum1, vl);
// leftovers
for (i = np; i < n; i += vl) {
vl = __riscv_vsetvl_e16m2(n - i);
vbfloat16m2_t ax0 = __riscv_vle16_v_bf16m2((const __bf16 *)&x[i], vl);
vbfloat16m2_t ay0 = __riscv_vle16_v_bf16m2((const __bf16 *)&y[i], vl);
vsum0 = __riscv_vfwmaccbf16_vv_f32m4(vsum0, ax0, ay0, vl);
}
// reduce
vl = __riscv_vsetvlmax_e32m4();
vfloat32m1_t redsum = __riscv_vfredusum_vs_f32m4_f32m1(vsum0, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
sumf += __riscv_vfmv_f_s_f32m1_f32(redsum);
#elif defined(__POWER9_VECTOR__) || defined(__VXE__) || defined(__VXE2__)
const int np = (n & ~(GGML_BF16_STEP - 1));
if (np > 0) {
GGML_F32_VEC sum[4] = {GGML_F32_VEC_ZERO};
for (; i < np; i += GGML_BF16_STEP) {
GGML_BF16_VEC vx0 = GGML_BF16_VEC_LOAD(x + i);
GGML_BF16_VEC vx1 = GGML_BF16_VEC_LOAD(x + i + 8);
GGML_BF16_VEC vy0 = GGML_BF16_VEC_LOAD(y + i);
GGML_BF16_VEC vy1 = GGML_BF16_VEC_LOAD(y + i + 8);
GGML_BF16_FMA_LO(sum[0], vx0, vy0);
GGML_BF16_FMA_HI(sum[1], vx0, vy0);
GGML_BF16_FMA_LO(sum[2], vx1, vy1);
GGML_BF16_FMA_HI(sum[3], vx1, vy1);
}
GGML_F32x4_REDUCE_4(sumf, sum[0], sum[1], sum[2], sum[3]);
}
#endif
for (; i < n; ++i) {
sumf += (ggml_float)(GGML_BF16_TO_FP32(x[i]) *
GGML_BF16_TO_FP32(y[i]));
}
*s = sumf;
}
void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc) {
assert(nrc == 1);
GGML_UNUSED(nrc);
GGML_UNUSED(bx);
GGML_UNUSED(by);
GGML_UNUSED(bs);
ggml_float sumf = 0.0;
#if defined(GGML_SIMD)
#if defined(__ARM_FEATURE_SVE)
const int ggml_f16_epr = svcnth();
const int ggml_f16_step = 8 * ggml_f16_epr;
const int np = n - (n % ggml_f16_step);
const int np2 = n - (n % ggml_f16_epr);
svfloat32_t sum1_lo = svdup_n_f32(0.0f);
svfloat32_t sum1_hi = svdup_n_f32(0.0f);
svfloat32_t sum2_lo = svdup_n_f32(0.0f);
svfloat32_t sum2_hi = svdup_n_f32(0.0f);
svfloat32_t sum3_lo = svdup_n_f32(0.0f);
svfloat32_t sum3_hi = svdup_n_f32(0.0f);
svfloat32_t sum4_lo = svdup_n_f32(0.0f);
svfloat32_t sum4_hi = svdup_n_f32(0.0f);
for (int i = 0; i < np; i += ggml_f16_step) {
ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0), GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0));
ggml_sve_f16_fma_widened(&sum2_lo, &sum2_hi, GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1), GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1));
ggml_sve_f16_fma_widened(&sum3_lo, &sum3_hi, GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2), GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2));
ggml_sve_f16_fma_widened(&sum4_lo, &sum4_hi, GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3), GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3));
ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4), GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4));
ggml_sve_f16_fma_widened(&sum2_lo, &sum2_hi, GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5), GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5));
ggml_sve_f16_fma_widened(&sum3_lo, &sum3_hi, GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6), GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6));
ggml_sve_f16_fma_widened(&sum4_lo, &sum4_hi, GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7), GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7));
}
for (int i = np; i < np2; i += ggml_f16_epr) {
ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, GGML_F16x_VEC_LOAD(x + i, 0), GGML_F16x_VEC_LOAD(y + i, 0));
}
if (np2 < n) {
const svbool_t pg = svwhilelt_b16(np2, n);
const svfloat16_t rx = svld1_f16(pg, (const __fp16 *)(x + np2));
const svfloat16_t ry = svld1_f16(pg, (const __fp16 *)(y + np2));
ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, rx, ry);
}
sum1_lo = svadd_f32_m(DEFAULT_PG32, sum1_lo, sum2_lo);
sum1_hi = svadd_f32_m(DEFAULT_PG32, sum1_hi, sum2_hi);
sum3_lo = svadd_f32_m(DEFAULT_PG32, sum3_lo, sum4_lo);
sum3_hi = svadd_f32_m(DEFAULT_PG32, sum3_hi, sum4_hi);
sum1_lo = svadd_f32_m(DEFAULT_PG32, sum1_lo, sum3_lo);
sum1_hi = svadd_f32_m(DEFAULT_PG32, sum1_hi, sum3_hi);
sumf = ggml_sve_sum_f32x2(sum1_lo, sum1_hi);
#elif defined(__riscv_v_intrinsic)
#if defined(__riscv_zvfh)
int vl = __riscv_vsetvlmax_e32m2();
vfloat32m1_t vs = __riscv_vfmv_v_f_f32m1(0.0f, 1);
vfloat32m2_t vsum;
vfloat16m1_t ax;
vfloat16m1_t ay;
vsum = __riscv_vreinterpret_v_u32m2_f32m2(__riscv_vmv_v_x_u32m2(0, vl));
for (int i = 0; i < n; i += vl) {
vl = __riscv_vsetvl_e16m1(n - i);
ax = __riscv_vle16_v_f16m1_tu(ax, (const _Float16 *)&x[i], vl);
ay = __riscv_vle16_v_f16m1_tu(ay, (const _Float16 *)&y[i], vl);
vsum = __riscv_vfwmacc_vv_f32m2_tu(vsum, ax, ay, vl);
}
vl = __riscv_vsetvlmax_e32m1();
vfloat32m1_t ac0 = __riscv_vfadd_vv_f32m1(__riscv_vget_v_f32m2_f32m1(vsum, 0), __riscv_vget_v_f32m2_f32m1(vsum, 1), vl);
vs = __riscv_vfredusum_vs_f32m1_f32m1(ac0, vs, vl);
sumf += __riscv_vfmv_f_s_f32m1_f32(vs);
#else
for (int i = 0; i < n; ++i) {
sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
}
#endif // __riscv_zvfh
#else
const int np = (n & ~(GGML_F16_STEP - 1));
GGML_F16_VEC sum[GGML_F16_ARR] = { GGML_F16_VEC_ZERO };
GGML_F16_VEC ax[GGML_F16_ARR];
GGML_F16_VEC ay[GGML_F16_ARR];
for (int i = 0; i < np; i += GGML_F16_STEP) {
for (int j = 0; j < GGML_F16_ARR; j++) {
ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
sum[j] = GGML_F16_VEC_FMA(sum[j], ax[j], ay[j]);
}
}
// reduce sum0..sum3 to sum0
GGML_F16_VEC_REDUCE(sumf, sum);
// leftovers
for (int i = np; i < n; ++i) {
sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
}
// if you hit this, you are likely running outside the FP range
assert(!isnan(sumf) && !isinf(sumf));
#endif
#else
for (int i = 0; i < n; ++i) {
sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
}
#endif // GGML_SIMD
*s = sumf;
}
void ggml_vec_silu_f32(const int n, float * y, const float * x) {
int i = 0;
#if defined(__AVX512F__) && defined(__AVX512DQ__)
for (; i + 15 < n; i += 16) {
_mm512_storeu_ps(y + i, ggml_v_silu(_mm512_loadu_ps(x + i)));
}
#elif defined(__AVX2__) && defined(__FMA__)
for (; i + 7 < n; i += 8) {
_mm256_storeu_ps(y + i, ggml_v_silu(_mm256_loadu_ps(x + i)));
}
#elif defined(__SSE2__)
for (; i + 3 < n; i += 4) {
_mm_storeu_ps(y + i, ggml_v_silu(_mm_loadu_ps(x + i)));
}
#elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
const int vlen = svcntw();
for (; i < n; i += vlen) {
const svbool_t pg = svwhilelt_b32_s32(i, n);
svst1_f32(pg, y + i, ggml_v_silu(pg, svld1_f32(pg, x + i)));
}
#elif defined(__ARM_NEON) && defined(__aarch64__)
for (; i + 3 < n; i += 4) {
vst1q_f32(y + i, ggml_v_silu(vld1q_f32(x + i)));
}
#elif defined(__riscv_v_intrinsic)
for (int vl; i < n; i += vl) {
vl = __riscv_vsetvl_e32m2(n - i);
vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
vfloat32m2_t vy = ggml_v_silu_m2(vx, vl);
__riscv_vse32_v_f32m2(&y[i], vy, vl);
}
#endif
for (; i < n; ++i) {
y[i] = ggml_silu_f32(x[i]);
}
}
void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float * g) {
int i = 0;
#if defined(__AVX512F__) && defined(__AVX512DQ__)
for (; i + 15 < n; i += 16) {
_mm512_storeu_ps(y + i, _mm512_mul_ps(ggml_v_silu(_mm512_loadu_ps(x + i)), _mm512_loadu_ps(g + i)));
}
#elif defined(__AVX2__) && defined(__FMA__)
for (; i + 7 < n; i += 8) {
_mm256_storeu_ps(y + i, _mm256_mul_ps(ggml_v_silu(_mm256_loadu_ps(x + i)), _mm256_loadu_ps(g + i)));
}
#elif defined(__SSE2__)
for (; i + 3 < n; i += 4) {
_mm_storeu_ps(y + i, _mm_mul_ps(ggml_v_silu(_mm_loadu_ps(x + i)), _mm_loadu_ps(g + i)));
}
#elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
const int vlen = svcntw();
for (; i < n; i += vlen) {
const svbool_t pg = svwhilelt_b32_s32(i, n);
svst1_f32(pg, y + i, svmul_f32_x(pg, ggml_v_silu(pg, svld1_f32(pg, x + i)), svld1_f32(pg, g + i)));
}
#elif defined(__ARM_NEON) && defined(__aarch64__)
for (; i + 3 < n; i += 4) {
vst1q_f32(y + i, vmulq_f32(ggml_v_silu(vld1q_f32(x + i)), vld1q_f32(g + i)));
}
#elif defined(__riscv_v_intrinsic)
for (int vl; i < n; i += vl) {
vl = __riscv_vsetvl_e32m2(n - i);
vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
vfloat32m2_t vg = __riscv_vle32_v_f32m2(&g[i], vl);
vfloat32m2_t vy = __riscv_vfmul_vv_f32m2(ggml_v_silu_m2(vx, vl), vg, vl);
__riscv_vse32_v_f32m2(&y[i], vy, vl);
}
#endif
for (; i < n; ++i) {
y[i] = ggml_silu_f32(x[i]) * g[i];
}
}
ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const float mean) {
int i = 0;
ggml_float sum = 0;
// TODO: optimize to process the remaining elements in groups using the smaller vector sizes from AVX2 and SSE
// ref: https://github.com/ggml-org/llama.cpp/pull/15953#pullrequestreview-3310928344
#if defined(__AVX512F__) && defined(__AVX512DQ__)
for (; i + 15 < n; i += 16) {
__m512 val = _mm512_sub_ps(_mm512_loadu_ps(x + i),
_mm512_set1_ps(mean));
_mm512_storeu_ps(y + i, val);
sum += (ggml_float)_mm512_reduce_add_ps(_mm512_mul_ps(val, val));
}
#elif defined(__AVX2__) && defined(__FMA__)
for (; i + 7 < n; i += 8) {
__m256 val = _mm256_sub_ps(_mm256_loadu_ps(x + i),
_mm256_set1_ps(mean));
_mm256_storeu_ps(y + i, val);
val = _mm256_mul_ps(val,val);
__m128 val2 = _mm_add_ps(_mm256_extractf128_ps(val, 1),
_mm256_castps256_ps128(val));
val2 = _mm_add_ps(val2, _mm_movehl_ps(val2, val2));
val2 = _mm_add_ss(val2, _mm_movehdup_ps(val2));
sum += (ggml_float)_mm_cvtss_f32(val2);
}
#elif defined(__SSE2__)
for (; i + 3 < n; i += 4) {
__m128 val = _mm_sub_ps(_mm_loadu_ps(x + i),
_mm_set1_ps(mean));
_mm_storeu_ps(y + i, val);
val = _mm_mul_ps(val, val);
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
val = _mm_add_ps(val, _mm_movehl_ps(val, val));
val = _mm_add_ss(val, _mm_movehdup_ps(val));
#else
__m128 tmp = _mm_shuffle_ps(val, val, _MM_SHUFFLE(2, 3, 0, 1));
val = _mm_add_ps(val, tmp);
tmp = _mm_movehl_ps(tmp, val);
val = _mm_add_ss(val, tmp);
#endif // __AVX__ || __AVX2__ || __AVX512F__
sum += (ggml_float)_mm_cvtss_f32(val);
}
#elif defined(__ARM_NEON) && defined(__aarch64__)
for (; i + 3 < n; i += 4) {
float32x4_t val = vsubq_f32(vld1q_f32(x + i),
vdupq_n_f32(mean));
vst1q_f32(y + i, val);
val = vmulq_f32(val, val);
sum += (ggml_float)vaddvq_f32(val);
}
#elif defined(__VXE__) || defined(__VXE2__)
for (; i + 3 < n; i += 4) {
float32x4_t val = vec_sub(vec_xl(0, x + i), vec_splats(mean));
vec_xst(val, 0, y + i);
val = vec_mul(val, val);
sum += (ggml_float)vec_hsum_f32x4(val);
}
#elif defined(__riscv_v_intrinsic)
vfloat64m1_t vsum = __riscv_vfmv_v_f_f64m1(0, 1);
for (int vl; i < n; i += vl) {
vl = __riscv_vsetvl_e32m2(n - i);
vfloat32m2_t val = __riscv_vfsub_vf_f32m2(__riscv_vle32_v_f32m2(&x[i], vl), mean, vl);
__riscv_vse32_v_f32m2(&y[i], val, vl);
val = __riscv_vfmul_vv_f32m2(val, val, vl);
vsum = __riscv_vfwredusum_vs_f32m2_f64m1(val, vsum, vl);
}
sum = (ggml_float)__riscv_vfmv_f_s_f64m1_f64(vsum);
#endif
for (; i < n; ++i) {
float val = x[i] - mean;
y[i] = val;
val *= val;
sum += (ggml_float)val;
}
return sum/n;
}
ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max) {
int i = 0;
ggml_float sum = 0;
#if defined(__AVX512F__) && defined(__AVX512DQ__)
for (; i + 15 < n; i += 16) {
__m512 val = ggml_v_expf(_mm512_sub_ps(_mm512_loadu_ps(x + i),
_mm512_set1_ps(max)));
_mm512_storeu_ps(y + i, val);
sum += (ggml_float)_mm512_reduce_add_ps(val);
}
#elif defined(__AVX2__) && defined(__FMA__)
for (; i + 7 < n; i += 8) {
__m256 val = ggml_v_expf(_mm256_sub_ps(_mm256_loadu_ps(x + i),
_mm256_set1_ps(max)));
_mm256_storeu_ps(y + i, val);
__m128 val2 = _mm_add_ps(_mm256_extractf128_ps(val, 1),
_mm256_castps256_ps128(val));
val2 = _mm_add_ps(val2, _mm_movehl_ps(val2, val2));
val2 = _mm_add_ss(val2, _mm_movehdup_ps(val2));
sum += (ggml_float)_mm_cvtss_f32(val2);
}
#elif defined(__SSE2__)
for (; i + 3 < n; i += 4) {
__m128 val = ggml_v_expf(_mm_sub_ps(_mm_loadu_ps(x + i),
_mm_set1_ps(max)));
_mm_storeu_ps(y + i, val);
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
val = _mm_add_ps(val, _mm_movehl_ps(val, val));
val = _mm_add_ss(val, _mm_movehdup_ps(val));
#else
__m128 tmp = _mm_shuffle_ps(val, val, _MM_SHUFFLE(2, 3, 0, 1));
val = _mm_add_ps(val, tmp);
tmp = _mm_movehl_ps(tmp, val);
val = _mm_add_ss(val, tmp);
#endif
sum += (ggml_float)_mm_cvtss_f32(val);
}
#elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
const int vlen = svcntw();
for (; i < n; i += vlen) {
const svbool_t pg = svwhilelt_b32_s32(i, n);
svfloat32_t val = ggml_v_expf(pg, svsub_f32_x(pg, svld1_f32(pg, x + i),
svdup_n_f32_x(pg, max)));
svst1_f32(pg, y + i, val);
sum += (ggml_float)svaddv_f32(pg, val);
}
#elif defined(__ARM_NEON) && defined(__aarch64__)
for (; i + 3 < n; i += 4) {
float32x4_t val = ggml_v_expf(vsubq_f32(vld1q_f32(x + i),
vdupq_n_f32(max)));
vst1q_f32(y + i, val);
sum += (ggml_float)vaddvq_f32(val);
}
#elif defined(__riscv_v_intrinsic)
vfloat64m1_t vsum = __riscv_vfmv_v_f_f64m1(0, 1);
for (int avl; i < n; i += avl) {
avl = __riscv_vsetvl_e32m2(n - i);
vfloat32m2_t val = ggml_v_expf_m2(__riscv_vfsub_vf_f32m2(__riscv_vle32_v_f32m2(&x[i], avl), max, avl), avl);
__riscv_vse32_v_f32m2(&y[i], val, avl);
vsum = __riscv_vfwredusum_vs_f32m2_f64m1(val, vsum, avl);
}
return (ggml_float)__riscv_vfmv_f_s_f64m1_f64(vsum);
#endif
for (; i < n; ++i) {
float val = expf(x[i] - max);
sum += (ggml_float)val;
y[i] = val;
}
return sum;
}
ggml_float ggml_vec_log_soft_max_f32(const int n, float * y, const float * x, float max) {
// log(soft_max) = log(soft_max_i / soft_max_sum) = log(soft_max_i) - log(soft_max_sum) = (logit_i - max) - log(soft_max_i)
int i = 0;
ggml_float sum = 0;
for (; i < n; ++i) {
float val = x[i] - max;
y[i] = val;
sum += (ggml_float)expf(val);
}
return sum = (ggml_float)logf(sum);
}
engine/ggml/src/ggml-cpu/vec.h
+1570
View File
File diff suppressed because it is too large Load Diff
engine/ggml/src/ggml-cuda/CMakeLists.txt
+269
View File
@@ -0,0 +1,269 @@
cmake_minimum_required(VERSION 3.18) # for CMAKE_CUDA_ARCHITECTURES
find_package(CUDAToolkit)
if (CUDAToolkit_FOUND)
message(STATUS "CUDA Toolkit found")
if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
# native == GPUs available at build time
# 50 == Maxwell, lowest CUDA 12 standard
# 60 == P100, FP16 CUDA intrinsics
# 61 == Pascal, __dp4a instruction (per-byte integer dot product)
# 70 == V100, FP16 tensor cores
# 75 == Turing, int8 tensor cores
# 80 == Ampere, asynchronous data loading, faster tensor core instructions
# 86 == RTX 3000, needs CUDA v11.1
# 89 == RTX 4000, needs CUDA v11.8
# 90 == Hopper H100/200, needs CUDA v11.8
# 120 == Blackwell, needs CUDA v12.8, FP4 tensor cores
#
# XX-virtual == compile CUDA code as PTX, do JIT compilation to binary code on first run
# XX-real == compile CUDA code as device code for this specific architecture
# no suffix == compile as both PTX and device code
#
# The default behavior for a non-native is to build virtual architectures as needed to cover all features needed
# for best performance and to also build real architectures for the most commonly used GPUs.
if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
set(CMAKE_CUDA_ARCHITECTURES "native")
else()
if (CUDAToolkit_VERSION VERSION_LESS "13")
list(APPEND CMAKE_CUDA_ARCHITECTURES 50-virtual 61-virtual 70-virtual)
endif ()
list(APPEND CMAKE_CUDA_ARCHITECTURES 75-virtual 80-virtual 86-real)
if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
list(APPEND CMAKE_CUDA_ARCHITECTURES 89-real 90-virtual)
endif()
if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")
# The CUDA architecture 120f-virtual would in principle work for Blackwell support
# but the newly added "f" suffix conflicted with a preexising regex for validating CUDA architectures in CMake.
# So either a recent CMake version or one with the backported fix is needed.
# The following versions should work:
# - CMake >= v3.31.8 && CMake < v4.0.0
# - CMake >= v4.0.2
# This is NOT documented in the CMake release notes,
# check Modules/Internal/CMakeCUDAArchitecturesValidate.cmake in the CMake git repository instead.
# However, the architectures 120a-real and 121a-real should work with basically any CMake version and
# until the release of e.g. Rubin there is no benefit to shipping virtual architectures for Blackwell.
list(APPEND CMAKE_CUDA_ARCHITECTURES 120a-real)
endif()
if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.9")
list(APPEND CMAKE_CUDA_ARCHITECTURES 121a-real)
endif()
endif()
endif()
enable_language(CUDA)
# TODO: Remove once CCCL 3.2 has been released and bundled with CUDA Toolkit
if (GGML_CUDA_CUB_3DOT2)
include(FetchContent)
FetchContent_Declare(
CCCL
GIT_REPOSITORY https://github.com/nvidia/cccl.git
GIT_TAG v3.2.0
GIT_SHALLOW TRUE
)
FetchContent_MakeAvailable(CCCL)
endif()
# Replace any plain 12X CUDA architectures with their "architecture-specific" equivalents 12Xa.
# 12X is forwards-compatible, 12Xa is not.
# Notably the Blackwell FP4 tensor core instructions are not forwards compatible and therefore need 12Xa.
# But while 12X vs. 12Xa can be checked in device code there is (to my knowledge) no easy way to do the same check in host code.
# So for now just replace all instances of 12X with 12Xa, this should be fine until Rubin is released.
foreach(ARCHS IN ITEMS CMAKE_CUDA_ARCHITECTURES CMAKE_CUDA_ARCHITECTURES_NATIVE)
set(FIXED_ARCHS "")
foreach(ARCH IN LISTS ${ARCHS})
if (ARCH MATCHES "^12[0-9](-real|-virtual)?$")
string(REGEX REPLACE "^(12[0-9])((-real|-virtual)?)$" "\\1a\\2" FIXED_ARCH ${ARCH})
message(STATUS "Replacing ${ARCH} in ${ARCHS} with ${FIXED_ARCH}")
list(APPEND FIXED_ARCHS "${FIXED_ARCH}")
else()
list(APPEND FIXED_ARCHS "${ARCH}")
endif()
endforeach()
set(${ARCHS} ${FIXED_ARCHS})
endforeach()
# If we try to compile a "native" build it will use the 12X architectures and fail.
# So we should instead use the native architectures as determined by CMake after replacing 12X with 12Xa.
# But if at the time of the build no GPUs are connected at all CMAKE_CUDA_ARCHITECTURES will contain garbage that we should not use.
if (CMAKE_CUDA_ARCHITECTURES STREQUAL "native" AND CMAKE_CUDA_ARCHITECTURES_NATIVE MATCHES "^[0-9]+(a|f)?(-real|-virtual)?(;[0-9]+(a|f)?(-real|-virtual)?|;)*$")
set(CMAKE_CUDA_ARCHITECTURES ${CMAKE_CUDA_ARCHITECTURES_NATIVE})
endif()
message(STATUS "Using CMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} CMAKE_CUDA_ARCHITECTURES_NATIVE=${CMAKE_CUDA_ARCHITECTURES_NATIVE}")
file(GLOB GGML_HEADERS_CUDA "*.cuh")
list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
file(GLOB GGML_SOURCES_CUDA "*.cu")
file(GLOB SRCS "template-instances/fattn-tile*.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
file(GLOB SRCS "template-instances/fattn-mma*.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
file(GLOB SRCS "template-instances/mmq*.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
file(GLOB SRCS "template-instances/mmf*.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
if (GGML_CUDA_FA_ALL_QUANTS)
file(GLOB SRCS "template-instances/fattn-vec*.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
else()
list(APPEND GGML_SOURCES_CUDA
template-instances/fattn-vec-instance-f16-f16.cu
template-instances/fattn-vec-instance-q4_0-q4_0.cu
template-instances/fattn-vec-instance-q8_0-q8_0.cu
template-instances/fattn-vec-instance-bf16-bf16.cu)
endif()
ggml_add_backend_library(ggml-cuda
${GGML_HEADERS_CUDA}
${GGML_SOURCES_CUDA}
)
add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
if (GGML_CUDA_GRAPHS)
add_compile_definitions(GGML_CUDA_USE_GRAPHS)
endif()
if (GGML_CUDA_FORCE_MMQ)
add_compile_definitions(GGML_CUDA_FORCE_MMQ)
endif()
if (GGML_CUDA_FORCE_CUBLAS)
add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
endif()
if (GGML_CUDA_NO_VMM)
add_compile_definitions(GGML_CUDA_NO_VMM)
endif()
if (NOT GGML_CUDA_FA)
add_compile_definitions(GGML_CUDA_NO_FA)
endif()
if (GGML_CUDA_NO_PEER_COPY)
add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
endif()
if (GGML_STATIC)
if (WIN32)
# As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas)
else ()
if (GGML_CUDA_CUB_3DOT2)
target_link_libraries(ggml-cuda PRIVATE CCCL::CCCL)
endif()
if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "10.1")
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
else()
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas_static)
endif()
endif()
else()
if (GGML_CUDA_CUB_3DOT2)
target_link_libraries(ggml-cuda PRIVATE CCCL::CCCL)
endif()
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas)
endif()
if (GGML_CUDA_NO_VMM)
# No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
else()
target_link_libraries(ggml-cuda PRIVATE CUDA::cuda_driver)
endif()
if (GGML_CUDA_NCCL)
find_package(NCCL)
if (NCCL_FOUND)
add_compile_definitions(GGML_USE_NCCL)
target_link_libraries(ggml-cuda PRIVATE NCCL::NCCL)
else()
message(STATUS "Warning: NCCL not found, performance for multiple CUDA GPUs will be suboptimal")
endif()
endif()
set(CUDA_CXX_FLAGS "")
set(CUDA_FLAGS -use_fast_math -extended-lambda)
if (GGML_CUDA_DEBUG)
list(APPEND CUDA_FLAGS -lineinfo)
add_compile_definitions(GGML_CUDA_DEBUG)
endif()
if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")
# Options are:
# - none (not recommended)
# - speed (nvcc's default)
# - balance
# - size
list(APPEND CUDA_FLAGS -compress-mode=${GGML_CUDA_COMPRESSION_MODE})
endif()
if (GGML_FATAL_WARNINGS)
list(APPEND CUDA_FLAGS -Werror all-warnings)
endif()
if (GGML_ALL_WARNINGS AND NOT MSVC)
set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
endif()
execute_process(
COMMAND ${NVCC_CMD} -Xcompiler --version
OUTPUT_VARIABLE CUDA_CCFULLVER
ERROR_QUIET
)
if (NOT CUDA_CCFULLVER MATCHES clang)
set(CUDA_CCID "GNU")
execute_process(
COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
OUTPUT_VARIABLE CUDA_CCVER
ERROR_QUIET
OUTPUT_STRIP_TRAILING_WHITESPACE
)
else()
if (CUDA_CCFULLVER MATCHES Apple)
set(CUDA_CCID "AppleClang")
else()
set(CUDA_CCID "Clang")
endif()
string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
endif()
message(STATUS "CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
ggml_get_flags(${CUDA_CCID} ${CUDA_CCVER})
list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS}) # This is passed to -Xcompiler later
endif()
if (NOT MSVC)
list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
else()
# CCCL 3.2 onwards will require a cpp-standard-compliant preprocessor for MSVC
# https://github.com/NVIDIA/cccl/pull/6827
list(APPEND CUDA_CXX_FLAGS /Zc:preprocessor)
endif()
list(JOIN CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED) # pass host compiler flags as a single argument
if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
endif()
target_compile_options(ggml-cuda PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
else()
message(FATAL_ERROR "CUDA Toolkit not found")
endif()
engine/ggml/src/ggml-cuda/acc.cu
+61
View File
@@ -0,0 +1,61 @@
#include "acc.cuh"
static __global__ void acc_f32(const float * x, const float * y, float * dst, const int64_t ne,
const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
const int64_t s11, const int64_t s12, const int64_t s13, const int64_t offset) {
const int64_t i = blockDim.x * blockIdx.x + threadIdx.x;
if (i >= ne) {
return;
}
int64_t src1_idx = i - offset;
int64_t tmp = src1_idx;
const int64_t i13 = tmp / s13;
tmp -= i13 * s13;
const int64_t i12 = tmp / s12;
tmp -= i12 * s12;
const int64_t i11 = tmp / s11;
tmp -= i11 * s11;
const int64_t i10 = tmp;
float val = x[i];
if (src1_idx >= 0 && i10 < ne10 && i11 < ne11 && i12 < ne12 && i13 < ne13) {
val += y[((i13*ne12 + i12) * ne11 + i11) * ne10 + i10];
}
dst[i] = val;
}
static void acc_f32_cuda(const float * x, const float * y, float * dst, const int64_t n_elements,
const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
const int64_t s1, const int64_t s2, const int64_t s3, const int64_t offset, cudaStream_t stream) {
const int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, ne13, s1, s2, s3, offset);
}
void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const float * src0_d = (const float *) src0->data;
const float * src1_d = (const float *) src1->data;
float * dst_d = (float *) dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_is_contiguous(src1));
GGML_ASSERT(dst->nb[0] == ggml_element_size(dst));
GGML_ASSERT(ggml_is_contiguously_allocated(dst));
const int64_t s1 = dst->op_params[0] / sizeof(float);
const int64_t s2 = dst->op_params[1] / sizeof(float);
const int64_t s3 = dst->op_params[2] / sizeof(float);
const int64_t offset = dst->op_params[3] / sizeof(float);
acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], s1, s2, s3, offset, stream);
}
engine/ggml/src/ggml-cuda/acc.cuh
+5
View File
@@ -0,0 +1,5 @@
#include "common.cuh"
#define CUDA_ACC_BLOCK_SIZE 256
void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
engine/ggml/src/ggml-cuda/add-id.cu
+58
View File
@@ -0,0 +1,58 @@
#include "add-id.cuh"
static __global__ void add_id_kernel(
const float * src0, const float * src1, const int32_t * src2, float * dst,
int64_t ne0, int64_t ne1,
size_t nb01, size_t nb02,
size_t nb11,
size_t nb21
) {
const int64_t i1 = blockIdx.x;
const int64_t i2 = blockIdx.y;
const int i11 = *(const int32_t *) ((const char *) src2 + i1*sizeof(int32_t) + i2*nb21);
const size_t nb1 = ne0 * sizeof(float);
const size_t nb2 = ne1 * nb1;
float * dst_row = (float *)((char *)dst + i1*nb1 + i2*nb2);
const float * src0_row = (const float *)((const char *)src0 + i1*nb01 + i2*nb02);
const float * src1_row = (const float *)((const char *)src1 + i11*nb11);
for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
dst_row[i0] = src0_row[i0] + src1_row[i0];
}
}
void ggml_cuda_op_add_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
const ggml_tensor * src2 = dst->src[2];
GGML_TENSOR_TERNARY_OP_LOCALS
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(src2->type == GGML_TYPE_I32);
GGML_ASSERT(nb00 == sizeof(float));
GGML_ASSERT(nb10 == sizeof(float));
GGML_ASSERT(nb20 == sizeof(int32_t));
const float * src0_d = (const float *)src0->data;
const float * src1_d = (const float *)src1->data;
const int32_t * src2_d = (const int32_t *)src2->data;
float * dst_d = (float *)dst->data;
int threads = std::min((int)ne00, 768); // cols
dim3 blocks(ne01, ne02); // n_experts_used, n_tokens
add_id_kernel<<<blocks, threads, 0, ctx.stream()>>>(
src0_d, src1_d, src2_d, dst_d,
ne0, ne1,
nb01, nb02,
nb11,
nb21
);
}
engine/ggml/src/ggml-cuda/add-id.cuh
+3
View File
@@ -0,0 +1,3 @@
#include "common.cuh"
void ggml_cuda_op_add_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
engine/ggml/src/ggml-cuda/allreduce.cu
+971
View File
@@ -0,0 +1,971 @@
#include "allreduce.cuh"
#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
#include "convert.cuh"
#include "ggml-impl.h"
#include <algorithm>
#include <cstdlib>
#include <cstring>
#include <limits>
// ---------------------------------------------------------------------------
// CUDA AllReduce for tensor-parallel inference across two GPUs.
//
// Provides an in-place sum reduction over matching tensors on two CUDA
// devices in the same process. Used by the tensor-split path alongside
// NCCL; targets setups without NVLink, where data is exchanged between the
// GPUs by staging it through pinned host memory over PCIe.
//
// Two reduction strategies are selected per call by tensor size:
//
// * Chunked kernel path (small reductions): a single CUDA kernel both
// stages data through pinned host memory and performs the local sum.
// Cross-GPU synchronization happens *inside the kernel* (busy-wait on
// a host-memory flag), which keeps launch overhead low for the
// latency-sensitive token-generation case.
//
// * Copy-engine path (large reductions): the transfer is split into
// D2H + H2D cudaMemcpyAsync chunks driven by the GPU's copy engine,
// followed by a small device-side add kernel. Cross-GPU
// synchronization happens *outside the kernel*, via CUDA events
// between streams. This keeps the compute engine free while large
// transfers are in flight, which matters for prefill-sized tensors.
// Reductions larger than the per-call inner cap are processed by an
// outer chunker that issues sequential inner calls.
// ---------------------------------------------------------------------------
// ---------------------------------------------------------------------------
// Cross-GPU signal mechanism
//
// One int per (slot, rank) pair in pinned host memory. Each AR call writes a
// strictly increasing token (= the AR call number) into its own arrival int.
// The peer spins until its read of the other's arrival int equals the token
// it expects for this call -- a mismatch means the peer hasn't arrived yet.
// Tokens never repeat over realistic call rates (32-bit int wraps in tens of
// days at thousands of ARs/sec), so arrival ints don't need to be reset
// between calls; we initialize once at pipeline init and let the values
// accumulate.
//
// There is exactly one writer (the owning GPU) and one reader (the peer), so
// we don't need atomics. A volatile store paired with __threadfence_system()
// provides the release ordering that makes the D2H writes visible system-wide
// before the arrival token is observed.
//
// atomicAdd_system() requires hostNativeAtomicSupported, which is unavailable
// on PCIe-attached consumer GPUs without NVLink, so the volatile path is the
// portable choice.
// ---------------------------------------------------------------------------
static __device__ __forceinline__ void ggml_cuda_ar_signal_set(int * p, int token) {
*(volatile int *)p = token;
}
static __device__ __forceinline__ int ggml_cuda_ar_signal_get(const int * p) {
return *(const volatile int *)p;
}
// Byte spacing between adjacent arrival ints. 64 bytes (one cache line)
// ensures each GPU/block's arrival slot lives on its own line, preventing
// false-sharing stalls on the polling GPU.
static constexpr size_t GGML_CUDA_AR_ARRIVAL_STRIDE = 64;
// Number of blocks the chunked kernel launches with. Each block stripes a
// disjoint slice of the data and synchronizes through its own arrival-token
// slot so multiple SMs can pump PCIe stores in parallel.
static constexpr int GGML_CUDA_AR_KERNEL_BLOCKS = 8;
// ---------------------------------------------------------------------------
// Chunked kernel AllReduce -- 2 GPUs, supports float, half, and bfloat16.
//
// Both GPUs run this kernel simultaneously on independent streams. sendbuf
// and recvbuf live in T_dst (the caller's tensor type); host_mine / host_other
// carry data in T_wire (the on-wire type, possibly narrower than T_dst -- e.g.
// T_dst=F32 with T_wire=BF16 halves the bytes pushed across PCIe). When
// T_dst == T_wire the casts below are no-ops.
//
// Each GPU runs three phases:
//
// Phase 1 (all threads): cast sendbuf (T_dst) -> T_wire and store as
// single-instruction-width vectors into host_mine.
// __threadfence_system() commits these writes to host
// memory.
// Phase 2 (thread 0): write token to arrival_mine; spin until
// arrival_other == token.
// Phase 3 (all threads): read T_wire vectors from host_other, cast
// each element to T_dst, and sum with the local
// sendbuf value (also rounded through T_wire so that
// both GPUs truncate identically -- this guarantees
// bit-equivalent results across the two devices).
//
// Multi-block: blocks stripe vectors across (gridDim.x * blockDim.x) global
// threads to keep multiple SMs issuing PCIe stores in parallel. Each block
// has its own arrival-token slot (offset by blockIdx.x * ARRIVAL_STRIDE);
// thread 0 of each block signals/spins on that slot independently of other
// blocks. Tail elements (the leftover < ELEMS_PER_VEC at the end) are
// handled only by block 0 to avoid cross-block writes to the same slots.
// ---------------------------------------------------------------------------
template <typename T_dst, typename T_wire>
static __global__ void ggml_cuda_ar_kernel(
const T_dst * sendbuf,
T_dst * recvbuf,
T_wire * __restrict__ host_mine,
const T_wire * __restrict__ host_other,
int count,
int * arrival_mine,
int * arrival_other,
int token) {
// Vector unit for the wire type, sized to the arch's widest single-instruction
// copy (16 B on Volta+). Each phase-1 iter writes one vector to host memory;
// each phase-3 iter reads one and produces ELEMS_PER_VEC sums.
constexpr int ELEMS_PER_VEC = ggml_cuda_get_max_cpy_bytes() / sizeof(T_wire);
constexpr int ARRIVAL_INTS = (int)(GGML_CUDA_AR_ARRIVAL_STRIDE / sizeof(int));
const int tid = threadIdx.x;
const int nt = blockDim.x;
const int bid = blockIdx.x;
const int gtid = bid * nt + tid;
const int gnt = gridDim.x * nt;
const int count_vec = count / ELEMS_PER_VEC;
const int tail = count_vec * ELEMS_PER_VEC;
// Phase 1: cast sendbuf (T_dst) -> host_mine (T_wire) and store as vectors.
{
for (int i = gtid; i < count_vec; i += gnt) {
const int off = i * ELEMS_PER_VEC;
T_wire wire[ELEMS_PER_VEC];
#pragma unroll
for (int k = 0; k < ELEMS_PER_VEC; ++k) {
wire[k] = ggml_cuda_cast<T_wire>(sendbuf[off + k]);
}
ggml_cuda_memcpy_1<sizeof(wire)>(&host_mine[off], wire);
}
if (bid == 0 && tid < count - tail) {
host_mine[tail + tid] = ggml_cuda_cast<T_wire>(sendbuf[tail + tid]);
}
}
// Commit this block's host writes before signalling.
__threadfence_system();
__syncthreads();
// Phase 2: thread 0 of each block signals on its own arrival slot, then
// spins for the matching slot from peer. Per-block tokens mean blocks
// proceed independently -- no inter-block barrier needed.
if (tid == 0) {
int * my_slot = arrival_mine + bid * ARRIVAL_INTS;
const int * other_slot = arrival_other + bid * ARRIVAL_INTS;
ggml_cuda_ar_signal_set(my_slot, token);
__threadfence_system(); // make our signal visible system-wide
while (ggml_cuda_ar_signal_get(other_slot) != token) {
#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
__nanosleep(100);
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
}
}
__syncthreads();
// Acquire peer's host_other writes (this block's stripe of them).
__threadfence_system();
// Phase 3: read peer's T_wire vector, cast both sides through T_wire for
// bit-equivalence, sum in T_dst precision, and write back to recvbuf.
{
for (int i = gtid; i < count_vec; i += gnt) {
const int off = i * ELEMS_PER_VEC;
T_wire wire[ELEMS_PER_VEC];
ggml_cuda_memcpy_1<sizeof(wire)>(wire, &host_other[off]);
#pragma unroll
for (int k = 0; k < ELEMS_PER_VEC; ++k) {
const T_wire d_low = ggml_cuda_cast<T_wire>(sendbuf[off + k]);
recvbuf[off + k] = ggml_cuda_cast<T_dst>(
ggml_cuda_cast<float>(d_low) + ggml_cuda_cast<float>(wire[k]));
}
}
if (bid == 0 && tid < count - tail) {
const T_wire d_low = ggml_cuda_cast<T_wire>(sendbuf[tail + tid]);
recvbuf[tail + tid] = ggml_cuda_cast<T_dst>(
ggml_cuda_cast<float>(d_low) +
ggml_cuda_cast<float>(host_other[tail + tid]));
}
}
}
// Combined load-convert-add kernel. The peer's contribution arrives as T_src
// (which may be a lower-precision type than T_dst when the BF16 round-trip is
// active). For bit-equivalence between the two GPUs, dst is first rounded
// through T_src's precision via ggml_cuda_cast -- peer already truncated its
// own value the same way before sending -- so both sides perform identical
// arithmetic. When T_dst == T_src the round-trip cast is a no-op.
template <typename T_dst, typename T_src>
static __global__ void ggml_cuda_ar_add_kernel(
T_dst * __restrict__ dst,
const T_src * __restrict__ src,
int count) {
const int tid = blockIdx.x * blockDim.x + threadIdx.x;
const int nt = gridDim.x * blockDim.x;
for (int i = tid; i < count; i += nt) {
const T_src d_low = ggml_cuda_cast<T_src>(dst[i]);
dst[i] = ggml_cuda_cast<T_dst>(
ggml_cuda_cast<float>(d_low) + ggml_cuda_cast<float>(src[i]));
}
}
// ---------------------------------------------------------------------------
// Pipeline structure
// ---------------------------------------------------------------------------
// Number of slots in the event / arrival ring. Two slots is sufficient:
// lockstep guarantees the two GPUs are at most one AR (or chunk) apart, so
// slot[N%2] is always safe to reuse -- peer has already consumed slot[N%2]
// from AR N-2 by the time we get to AR N. acquire_slot's
// cudaEventSynchronize on ev.ker for both devices makes that consumption
// explicit before we overwrite host_buf[slot] for the new AR.
static constexpr int GGML_CUDA_AR_POOL_SIZE = 2;
// Maximum chunk size (bytes per GPU) handled by one chunked kernel launch.
// Larger tensors are reduced by issuing multiple chunked launches.
static constexpr size_t GGML_CUDA_AR_MAX_BYTES = 1024 * 1024; // 1 MB
// Copy-engine path: largest tensor accepted on this path; sets host_large /
// dev_tmp allocation size.
static constexpr size_t GGML_CUDA_AR_COPY_MAX_BYTES = 32 * 1024 * 1024; // 32 MB
// AR wire size at which the copy-engine path takes over from the chunked-
// kernel path. Override via GGML_CUDA_AR_COPY_THRESHOLD.
static constexpr size_t GGML_CUDA_AR_COPY_THRESHOLD_DEFAULT = 1024 * 1024; // 1 MB
// Per-call CE chunk-size heuristic: chunk_bytes = clamp(nbytes / 4, MIN, MAX).
// The /4 keeps ~4 chunks in flight at any moment (good D2H/H2D overlap with
// the peer); the clamps cover the cases where nbytes/4 is too small (per-
// memcpy fixed cost dominates) or too large (chunk-level pipelining stalls).
// Env var GGML_CUDA_AR_COPY_CHUNK_BYTES can override with a fixed value.
static constexpr size_t GGML_CUDA_AR_COPY_CHUNK_BYTES_HEURISTIC_MIN = 512 * 1024; // 512 KB
static constexpr size_t GGML_CUDA_AR_COPY_CHUNK_BYTES_HEURISTIC_MAX = 2 * 1024 * 1024; // 2 MB
// Absolute floor that an env-var override is allowed to set; this caps the
// per-slot copy-event array. 256 KB -> up to 128 chunks per 32 MB tensor.
static constexpr size_t GGML_CUDA_AR_COPY_CHUNK_BYTES_MIN = 256 * 1024;
static constexpr int GGML_CUDA_AR_COPY_MAX_CHUNKS =
static_cast<int>((GGML_CUDA_AR_COPY_MAX_BYTES + GGML_CUDA_AR_COPY_CHUNK_BYTES_MIN - 1) /
GGML_CUDA_AR_COPY_CHUNK_BYTES_MIN);
struct ggml_cuda_ar_event_slot {
cudaEvent_t app = nullptr; // upstream computation complete
cudaEvent_t cpy[GGML_CUDA_AR_COPY_MAX_CHUNKS] = {}; // copy-engine D2H chunks complete
cudaEvent_t h2d = nullptr; // copy-engine H2Ds complete (handoff AR stream -> compute stream)
cudaEvent_t ker = nullptr; // AllReduce kernel complete
};
// Mapped pinned host allocation: cudaHostAlloc + cudaHostGetDevicePointer
// in one place, with the host handle preserved for cudaFreeHost. Used where
// the CPU never touches the buffer -- only the device reads/writes via the
// mapped device pointer. Required on systems where cudaDevAttrCanUseHost-
// PointerForRegisteredMem is 0 and the host pointer can't be used as a
// device pointer.
struct ggml_cuda_ar_host_mapping {
uint8_t * host = nullptr; // cudaFreeHost handle; also the H-side ptr for cudaMemcpyAsync
uint8_t * dev = nullptr; // device-side pointer for kernels / cudaMemset
cudaError_t alloc(size_t bytes) {
cudaError_t rc = cudaHostAlloc(reinterpret_cast<void **>(&host), bytes,
cudaHostAllocPortable | cudaHostAllocMapped);
if (rc != cudaSuccess) {
host = nullptr;
return rc;
}
rc = cudaHostGetDevicePointer(reinterpret_cast<void **>(&dev), host, 0);
if (rc != cudaSuccess) {
cudaFreeHost(host);
host = nullptr;
dev = nullptr;
}
return rc;
}
void free() {
if (host) {
cudaFreeHost(host);
host = nullptr;
dev = nullptr;
}
}
};
struct ggml_cuda_ar_pipeline {
int n_devices;
int devices[GGML_CUDA_MAX_DEVICES];
size_t buf_bytes; // bytes per device in host_buf[]
size_t copy_bytes; // bytes per device in host_large[] / dev_tmp[]
size_t copy_threshold;
size_t copy_chunk_bytes;
size_t bf16_threshold; // tensors >= this size (bytes) are reduced via FP32->BF16 round-trip; 0 disables
uint64_t call_count;
// Per-device resources.
ggml_cuda_ar_host_mapping host_buf[GGML_CUDA_MAX_DEVICES]; // pinned staging (chunked kernel)
ggml_cuda_ar_host_mapping host_large[GGML_CUDA_MAX_DEVICES]; // pinned staging (copy-engine)
char * dev_tmp[GGML_CUDA_MAX_DEVICES]; // device scratch for copy-engine path
cudaStream_t streams[GGML_CUDA_MAX_DEVICES]; // non-blocking
ggml_cuda_ar_event_slot ev_pool[GGML_CUDA_MAX_DEVICES][GGML_CUDA_AR_POOL_SIZE];
// Copy-engine: per-device "I finished reading my peer's host_large"
// event. Indexed by RECORDER device. Recorded same-device on streams[i]
// after stage 2's last H2D from host_large[peer]. Waited cross-device
// by peer's stage-1 stream before the next AR overwrites host_large[peer].
cudaEvent_t host_large_read_done[GGML_CUDA_MAX_DEVICES];
bool host_large_read_done_valid;
// Copy-engine: per-device "my add_kernel is done with dev_tmp" event.
// Recorded on the compute stream after each add_kernel; the AR stream
// waits on it before the next copy_impl's H2D overwrites dev_tmp. Lets us
// single-buffer dev_tmp despite add_kernel running on a separate stream.
cudaEvent_t dev_tmp_kernel_done[GGML_CUDA_MAX_DEVICES];
bool dev_tmp_kernel_done_valid;
// Arrival ring: ARRIVAL_STRIDE bytes between adjacent ints. Mapped pinned
// memory; CPU never reads/writes -- only the kernel and cudaMemset.
// Use ggml_cuda_ar_arrival_ptr() to index.
ggml_cuda_ar_host_mapping arrival;
};
// Base pointer for the (slot, rank) per-block token block. The kernel adds
// blockIdx.x * (ARRIVAL_STRIDE/sizeof(int)) internally to land on its own slot.
static int * ggml_cuda_ar_arrival_ptr(const ggml_cuda_ar_pipeline * p, int slot, int rank) {
const size_t offset = ((size_t)slot * p->n_devices + rank) *
GGML_CUDA_AR_KERNEL_BLOCKS * GGML_CUDA_AR_ARRIVAL_STRIDE;
return reinterpret_cast<int *>(p->arrival.dev + offset);
}
static uint64_t ggml_cuda_ar_env_u64(const char * name, uint64_t default_value) {
const char * value = getenv(name);
if (value == nullptr || value[0] == '\0') {
return default_value;
}
char * end = nullptr;
const unsigned long long parsed = strtoull(value, &end, 10);
return end != value ? (uint64_t) parsed : default_value;
}
struct ggml_cuda_ar_slot_info {
int slot;
int token;
};
static ggml_cuda_ar_slot_info ggml_cuda_ar_acquire_slot(ggml_cuda_ar_pipeline * p) {
const int slot = static_cast<int>(p->call_count % GGML_CUDA_AR_POOL_SIZE);
const bool pool_lapped = p->call_count >= GGML_CUDA_AR_POOL_SIZE;
p->call_count++;
if (pool_lapped) {
for (int i = 0; i < p->n_devices; ++i) {
ggml_cuda_set_device(p->devices[i]);
CUDA_CHECK(cudaEventSynchronize(p->ev_pool[i][slot].ker));
}
}
return { slot, (int) p->call_count };
}
// Per-AR copy-engine chunk size: env-var override if set, else heuristic
// (clamp(nbytes/4, HEURISTIC_MIN, HEURISTIC_MAX)).
static size_t ggml_cuda_ar_chunk_bytes(const ggml_cuda_ar_pipeline * p, size_t nbytes) {
if (p->copy_chunk_bytes > 0) {
return p->copy_chunk_bytes;
}
return std::min(GGML_CUDA_AR_COPY_CHUNK_BYTES_HEURISTIC_MAX,
std::max(GGML_CUDA_AR_COPY_CHUNK_BYTES_HEURISTIC_MIN, nbytes / 4));
}
static void ggml_cuda_ar_wait_for_compute(
ggml_cuda_ar_pipeline * p, ggml_backend_cuda_context * cuda_ctx, int rank, int slot) {
ggml_cuda_ar_event_slot & ev = p->ev_pool[rank][slot];
CUDA_CHECK(cudaEventRecord(ev.app, cuda_ctx->stream()));
CUDA_CHECK(cudaStreamWaitEvent(p->streams[rank], ev.app));
}
// ---------------------------------------------------------------------------
// Init / free
// ---------------------------------------------------------------------------
ggml_cuda_ar_pipeline * ggml_cuda_ar_pipeline_init(const int * devices, size_t n_devices) {
if (n_devices != 2) {
GGML_LOG_DEBUG("%s: internal AllReduce only supports n_devices=2 (got %zu); "
"falling back\n", __func__, n_devices);
return nullptr;
}
// The chunked kernel uses __nanosleep, which is sm70+ (Volta+).
for (size_t i = 0; i < n_devices; ++i) {
const int cc = ggml_cuda_info().devices[devices[i]].cc;
if (cc < GGML_CUDA_CC_VOLTA) {
GGML_LOG_DEBUG("%s: internal AllReduce requires compute capability >= %d "
"(device %d has cc=%d); falling back\n",
__func__, GGML_CUDA_CC_VOLTA, devices[i], cc);
return nullptr;
}
}
auto * p = new ggml_cuda_ar_pipeline{};
p->n_devices = n_devices;
p->copy_bytes = GGML_CUDA_AR_COPY_MAX_BYTES;
p->copy_threshold = ggml_cuda_ar_env_u64("GGML_CUDA_AR_COPY_THRESHOLD", GGML_CUDA_AR_COPY_THRESHOLD_DEFAULT);
// 0 = use the per-call heuristic (default). Non-zero env value forces a
// fixed chunk size for diagnostics, with a floor at COPY_CHUNK_BYTES_MIN.
p->copy_chunk_bytes = ggml_cuda_ar_env_u64("GGML_CUDA_AR_COPY_CHUNK_BYTES", 0);
if (p->copy_chunk_bytes > 0 && p->copy_chunk_bytes < GGML_CUDA_AR_COPY_CHUNK_BYTES_MIN) {
GGML_LOG_WARN("%s: GGML_CUDA_AR_COPY_CHUNK_BYTES=%zu below minimum %zu; clamping\n",
__func__, p->copy_chunk_bytes, GGML_CUDA_AR_COPY_CHUNK_BYTES_MIN);
p->copy_chunk_bytes = GGML_CUDA_AR_COPY_CHUNK_BYTES_MIN;
}
// Default 1: BF16 round-trip is always on for F32 inputs (any non-zero
// ne). Set GGML_CUDA_AR_BF16_THRESHOLD=0 to disable, or to a larger
// byte threshold to opt out for small tensors.
p->bf16_threshold = ggml_cuda_ar_env_u64("GGML_CUDA_AR_BF16_THRESHOLD", 1);
for (size_t i = 0; i < n_devices; ++i) {
p->devices[i] = devices[i];
}
// Per-device streams and event pools.
for (size_t i = 0; i < n_devices; ++i) {
ggml_cuda_set_device(p->devices[i]);
cudaStream_t stream = nullptr;
if (cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking) != cudaSuccess) {
GGML_LOG_ERROR("%s: cudaStreamCreateWithFlags failed for device %d\n",
__func__, p->devices[i]);
ggml_cuda_ar_pipeline_free(p);
return nullptr;
}
p->streams[i] = stream;
for (int s = 0; s < GGML_CUDA_AR_POOL_SIZE; ++s) {
bool ok =
cudaEventCreateWithFlags(&p->ev_pool[i][s].app, cudaEventDisableTiming) == cudaSuccess &&
cudaEventCreateWithFlags(&p->ev_pool[i][s].h2d, cudaEventDisableTiming) == cudaSuccess &&
cudaEventCreateWithFlags(&p->ev_pool[i][s].ker, cudaEventDisableTiming) == cudaSuccess;
for (int c = 0; ok && c < GGML_CUDA_AR_COPY_MAX_CHUNKS; ++c) {
ok = cudaEventCreateWithFlags(&p->ev_pool[i][s].cpy[c], cudaEventDisableTiming) == cudaSuccess;
}
if (!ok) {
GGML_LOG_ERROR("%s: cudaEventCreate failed for device %d slot %d\n",
__func__, p->devices[i], s);
ggml_cuda_ar_pipeline_free(p);
return nullptr;
}
}
if (cudaEventCreateWithFlags(&p->host_large_read_done[i], cudaEventDisableTiming) != cudaSuccess) {
GGML_LOG_ERROR("%s: cudaEventCreate for host_large_read_done failed for device %d\n",
__func__, p->devices[i]);
ggml_cuda_ar_pipeline_free(p);
return nullptr;
}
if (cudaEventCreateWithFlags(&p->dev_tmp_kernel_done[i], cudaEventDisableTiming) != cudaSuccess) {
GGML_LOG_ERROR("%s: cudaEventCreate for dev_tmp_kernel_done failed for device %d\n",
__func__, p->devices[i]);
ggml_cuda_ar_pipeline_free(p);
return nullptr;
}
}
// Arrival ring: cache-line padded so each GPU's int is on its own line.
const size_t arrival_bytes =
(size_t)GGML_CUDA_AR_POOL_SIZE * n_devices *
GGML_CUDA_AR_KERNEL_BLOCKS * GGML_CUDA_AR_ARRIVAL_STRIDE;
if (p->arrival.alloc(arrival_bytes) != cudaSuccess) {
GGML_LOG_ERROR("%s: alloc for arrival ring failed (%zu bytes)\n",
__func__, arrival_bytes);
ggml_cuda_ar_pipeline_free(p);
return nullptr;
}
ggml_cuda_set_device(p->devices[0]);
if (cudaMemset(p->arrival.dev, 0, arrival_bytes) != cudaSuccess) {
GGML_LOG_ERROR("%s: cudaMemset for arrival ring failed (%zu bytes)\n",
__func__, arrival_bytes);
ggml_cuda_ar_pipeline_free(p);
return nullptr;
}
// Per-device pinned staging buffers -- POOL_SIZE-deep ring so the chunked-
// kernel can write the next slot's data while the peer is still reading
// the previous slot's. Indexed by (slot * buf_bytes) at the call site.
p->buf_bytes = GGML_CUDA_AR_MAX_BYTES;
const size_t host_buf_total = (size_t) GGML_CUDA_AR_POOL_SIZE * p->buf_bytes;
for (size_t i = 0; i < n_devices; ++i) {
if (p->host_buf[i].alloc(host_buf_total) != cudaSuccess) {
GGML_LOG_ERROR("%s: alloc for staging failed (%zu bytes)\n",
__func__, host_buf_total);
ggml_cuda_ar_pipeline_free(p);
return nullptr;
}
}
// Copy-engine path: pinned host staging + device scratch, sized for the
// largest tensor we accept on this path (GGML_CUDA_AR_COPY_MAX_BYTES).
// dev_tmp is single-buffered; cross-AR safety is enforced by an explicit
// cross-stream wait in copy_impl on the prior AR's add_kernel-done event.
for (size_t i = 0; i < n_devices; ++i) {
ggml_cuda_set_device(p->devices[i]);
if (p->host_large[i].alloc(p->copy_bytes) != cudaSuccess) {
GGML_LOG_ERROR("%s: alloc for large staging failed (%zu bytes)\n",
__func__, p->copy_bytes);
ggml_cuda_ar_pipeline_free(p);
return nullptr;
}
if (cudaMalloc(reinterpret_cast<void **>(&p->dev_tmp[i]), p->copy_bytes) != cudaSuccess) {
GGML_LOG_ERROR("%s: cudaMalloc for copy scratch failed (%zu bytes) on device %d\n",
__func__, p->copy_bytes, p->devices[i]);
ggml_cuda_ar_pipeline_free(p);
return nullptr;
}
}
GGML_LOG_INFO("%s: initialized AllReduce pipeline: %zu GPUs, "
"%zu KB chunked kernel staging + %zu MB copy-engine staging per GPU\n",
__func__, n_devices, p->buf_bytes >> 10, p->copy_bytes >> 20);
return p;
}
void ggml_cuda_ar_pipeline_free(ggml_cuda_ar_pipeline * p) {
if (!p) {
return;
}
// Drain all in-flight kernels before tearing down resources.
for (int i = 0; i < p->n_devices; ++i) {
if (p->streams[i]) {
ggml_cuda_set_device(p->devices[i]);
cudaStreamSynchronize(p->streams[i]);
}
}
for (int i = 0; i < p->n_devices; ++i) {
p->host_buf[i].free();
p->host_large[i].free();
if (p->dev_tmp[i]) {
ggml_cuda_set_device(p->devices[i]);
cudaFree(p->dev_tmp[i]);
}
ggml_cuda_set_device(p->devices[i]);
for (int s = 0; s < GGML_CUDA_AR_POOL_SIZE; ++s) {
if (p->ev_pool[i][s].app) { cudaEventDestroy(p->ev_pool[i][s].app); }
for (int c = 0; c < GGML_CUDA_AR_COPY_MAX_CHUNKS; ++c) {
if (p->ev_pool[i][s].cpy[c]) { cudaEventDestroy(p->ev_pool[i][s].cpy[c]); }
}
if (p->ev_pool[i][s].h2d) { cudaEventDestroy(p->ev_pool[i][s].h2d); }
if (p->ev_pool[i][s].ker) { cudaEventDestroy(p->ev_pool[i][s].ker); }
}
if (p->host_large_read_done[i]) {
ggml_cuda_set_device(p->devices[i]);
cudaEventDestroy(p->host_large_read_done[i]);
}
if (p->dev_tmp_kernel_done[i]) {
ggml_cuda_set_device(p->devices[i]);
cudaEventDestroy(p->dev_tmp_kernel_done[i]);
}
if (p->streams[i]) {
ggml_cuda_set_device(p->devices[i]);
cudaStreamDestroy(p->streams[i]);
}
}
p->arrival.free();
delete p;
}
// ---------------------------------------------------------------------------
// Dispatch
// ---------------------------------------------------------------------------
// Asymmetric copy_impl: data sent over PCIe in T_src precision (one element of
// nbytes per ne element); accumulated locally into a T_dst buffer. When
// T_src == T_dst this is the original homogeneous reduction. When they differ
// (e.g. BF16 wire / F32 accumulator) the add kernel rounds dst through T_src
// for bit-equivalence between GPUs and we skip the otherwise-needed
// post-conversion entirely.
template <typename T_src, typename T_dst>
static bool ggml_cuda_ar_allreduce_copy_impl(
ggml_cuda_ar_pipeline * p,
ggml_backend_t * backends,
T_src * const src_buf[GGML_CUDA_MAX_DEVICES],
T_dst * const dst_buf[GGML_CUDA_MAX_DEVICES],
const bool compute[GGML_CUDA_MAX_DEVICES],
int64_t ne,
size_t nbytes) {
GGML_ASSERT(p->n_devices == 2);
GGML_ASSERT(nbytes <= p->copy_bytes);
GGML_ASSERT(ne <= std::numeric_limits<int>::max());
const size_t chunk_bytes = ggml_cuda_ar_chunk_bytes(p, nbytes);
GGML_ASSERT(chunk_bytes > 0);
const int slot = ggml_cuda_ar_acquire_slot(p).slot;
const size_t copy_chunks = (nbytes + chunk_bytes - 1) / chunk_bytes;
GGML_ASSERT(copy_chunks <= GGML_CUDA_AR_COPY_MAX_CHUNKS);
ggml_backend_cuda_context * cuda_ctx[2] = {};
// Stage 1: both GPUs copy their local contribution to pinned host memory.
for (int i = 0; i < 2; ++i) {
ggml_cuda_set_device(p->devices[i]);
cuda_ctx[i] = static_cast<ggml_backend_cuda_context *>(backends[i]->context);
GGML_ASSERT(cuda_ctx[i]->device == p->devices[i]);
ggml_cuda_ar_wait_for_compute(p, cuda_ctx[i], i, slot);
// Wait for peer's H2D from our host_large[i] (recorded in the
// previous AR's stage 2) to complete before we overwrite host_large[i].
// host_large_read_done[peer] = peer finished reading host_large[i].
// No-op on the first AR -- no prior record exists.
if (p->host_large_read_done_valid) {
const int peer = 1 - i;
CUDA_CHECK(cudaStreamWaitEvent(p->streams[i], p->host_large_read_done[peer]));
}
if (!compute[i]) {
CUDA_CHECK(cudaMemsetAsync(src_buf[i], 0, nbytes, p->streams[i]));
}
for (size_t c = 0; c < copy_chunks; ++c) {
const size_t offset = c * chunk_bytes;
const size_t this_bytes = (nbytes - offset) < chunk_bytes ?
(nbytes - offset) : chunk_bytes;
CUDA_CHECK(cudaMemcpyAsync(
p->host_large[i].host + offset, reinterpret_cast<char *>(src_buf[i]) + offset, this_bytes,
cudaMemcpyDeviceToHost, p->streams[i]));
CUDA_CHECK(cudaEventRecord(p->ev_pool[i][slot].cpy[c], p->streams[i]));
}
}
// Stage 2: each GPU waits for each peer D2H chunk, pulls that chunk back to
// local device scratch (dev_tmp), then performs one device-local add over
// the assembled peer tensor. The H2Ds run on the AR stream (copy engine)
// and the add_kernel runs on the caller's compute stream, so the AR stream
// stays pure-copy and avoids an in-stream copy->compute engine switch every
// AR. dev_tmp is single-buffered: the AR stream waits cross-stream on the
// prior AR's add_kernel-done event before overwriting it.
for (int i = 0; i < 2; ++i) {
const int peer = 1 - i;
ggml_cuda_set_device(p->devices[i]);
// Wait for the previous AR's add_kernel (on the compute stream) to
// finish reading dev_tmp before our H2D overwrites it. No-op on the
// first copy_impl call.
if (p->dev_tmp_kernel_done_valid) {
CUDA_CHECK(cudaStreamWaitEvent(p->streams[i], p->dev_tmp_kernel_done[i]));
}
for (size_t c = 0; c < copy_chunks; ++c) {
const size_t offset = c * chunk_bytes;
const size_t this_bytes = (nbytes - offset) < chunk_bytes ?
(nbytes - offset) : chunk_bytes;
CUDA_CHECK(cudaStreamWaitEvent(p->streams[i], p->ev_pool[peer][slot].cpy[c]));
CUDA_CHECK(cudaMemcpyAsync(
p->dev_tmp[i] + offset, p->host_large[peer].host + offset, this_bytes,
cudaMemcpyHostToDevice, p->streams[i]));
}
// Mark our reads of host_large[peer] complete so peer's next AR can
// safely overwrite it.
CUDA_CHECK(cudaEventRecord(p->host_large_read_done[i], p->streams[i]));
// Hand off from AR stream (copy engine) to compute stream: compute
// stream waits for all H2Ds to finish, then runs the add_kernel.
CUDA_CHECK(cudaEventRecord(p->ev_pool[i][slot].h2d, p->streams[i]));
CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx[i]->stream(), p->ev_pool[i][slot].h2d));
const int block_size = 256;
int n_blocks = (int) ((ne + block_size - 1) / block_size);
if (n_blocks > 1024) {
n_blocks = 1024;
}
ggml_cuda_ar_add_kernel<T_dst, T_src><<<n_blocks, block_size, 0, cuda_ctx[i]->stream()>>>(
dst_buf[i],
reinterpret_cast<const T_src *>(p->dev_tmp[i]),
(int) ne);
CUDA_CHECK(cudaGetLastError());
// Record dev_tmp-released on the compute stream so the next copy_impl
// can wait for the kernel to finish before overwriting dev_tmp. Also
// record AR-done as ev.ker for acquire_slot's pool-wraparound sync.
CUDA_CHECK(cudaEventRecord(p->dev_tmp_kernel_done[i], cuda_ctx[i]->stream()));
CUDA_CHECK(cudaEventRecord(p->ev_pool[i][slot].ker, cuda_ctx[i]->stream()));
}
p->host_large_read_done_valid = true;
p->dev_tmp_kernel_done_valid = true;
return true;
}
// Outer-level chunker: copy_impl handles up to copy_bytes per call (limited by
// the host_large / dev_tmp allocation size). When the full AR exceeds that,
// slice the tensor into copy_bytes-sized pieces and call copy_impl repeatedly.
// Each slice goes through its own stage 1 -> stage 2 cycle and acquires its own
// slot, so cross-AR fences and pool wraparound work the same way as for any
// other sequence of small ARs.
template <typename T_src, typename T_dst>
static bool ggml_cuda_ar_allreduce_copy_outer(
ggml_cuda_ar_pipeline * p,
ggml_backend_t * backends,
T_src * const src_buf[GGML_CUDA_MAX_DEVICES],
T_dst * const dst_buf[GGML_CUDA_MAX_DEVICES],
const bool compute[GGML_CUDA_MAX_DEVICES],
int64_t ne) {
const int64_t outer_max_elems = (int64_t) (p->copy_bytes / sizeof(T_src));
GGML_ASSERT(outer_max_elems > 0);
bool ok = true;
for (int64_t outer_start = 0; outer_start < ne && ok; outer_start += outer_max_elems) {
const int64_t outer_ne = std::min(outer_max_elems, ne - outer_start);
const size_t outer_nbytes = (size_t) outer_ne * sizeof(T_src);
T_src * src[GGML_CUDA_MAX_DEVICES] = {};
T_dst * dst[GGML_CUDA_MAX_DEVICES] = {};
for (int i = 0; i < p->n_devices; ++i) {
src[i] = src_buf[i] + outer_start;
dst[i] = dst_buf[i] + outer_start;
}
ok = ggml_cuda_ar_allreduce_copy_impl<T_src, T_dst>(
p, backends, src, dst, compute, outer_ne, outer_nbytes);
}
return ok;
}
bool ggml_cuda_ar_allreduce(
ggml_cuda_ar_pipeline * p,
ggml_backend_t * backends,
ggml_tensor ** tensors) {
GGML_ASSERT(p != nullptr);
const int n = p->n_devices;
GGML_ASSERT(n == 2);
const ggml_type input_type = tensors[0]->type;
GGML_ASSERT(input_type == GGML_TYPE_F32 || input_type == GGML_TYPE_F16 || input_type == GGML_TYPE_BF16);
const int64_t ne = ggml_nelements(tensors[0]);
GGML_ASSERT(ne > 0);
const size_t input_nbytes = ggml_nbytes(tensors[0]);
// BF16 round-trip: F32 inputs >= bf16_threshold are converted to BF16 for
// the reduction (chunked or copy-engine), halving on-wire bytes. Matches
// NCCL's behaviour. The pre-conversion zeroes inactive shards so the
// inner paths see them as already-prepared compute tensors.
const bool use_bf16 =
input_type == GGML_TYPE_F32 &&
p->bf16_threshold > 0 &&
input_nbytes >= p->bf16_threshold;
const ggml_type kernel_type = use_bf16 ? GGML_TYPE_BF16 : input_type;
const size_t type_size = ggml_type_size(kernel_type);
GGML_ASSERT(p->buf_bytes >= type_size);
const size_t nbytes = (size_t) ne * type_size;
bool compute_flag[GGML_CUDA_MAX_DEVICES] = {};
for (int i = 0; i < n; ++i) {
compute_flag[i] = (tensors[i]->flags & GGML_TENSOR_FLAG_COMPUTE) != 0;
}
// Decide between copy-engine and chunked kernel paths based on the working
// type's actual byte count. No upper bound: copy_outer slices reductions
// larger than copy_bytes into copy_bytes-sized pieces.
const bool use_copy_engine =
p->copy_threshold > 0 &&
nbytes >= p->copy_threshold;
// BF16 inactive-shard zeroing: when use_bf16 is on, the combined kernel
// (chunked kernel path) and the combined add kernel (copy_engine path)
// both accumulate into the F32 tensor data directly, so an inactive
// shard's accumulator must start at zero.
if (use_bf16) {
for (int i = 0; i < n; ++i) {
if (!compute_flag[i]) {
auto * cuda_ctx = static_cast<ggml_backend_cuda_context *>(backends[i]->context);
GGML_ASSERT(cuda_ctx->device == p->devices[i]);
ggml_cuda_set_device(p->devices[i]);
CUDA_CHECK(cudaMemsetAsync(tensors[i]->data, 0, (size_t) ne * sizeof(float), cuda_ctx->stream()));
}
}
}
// Pre-convert F32 -> BF16 into bf16_tmp ONLY for the copy_engine + use_bf16
// path; the chunked kernel path's combined kernel does the conversion
// inline as it writes to host_buf.
ggml_cuda_pool_alloc<nv_bfloat16> bf16_tmp[GGML_CUDA_MAX_DEVICES];
void * copy_src_ptr[GGML_CUDA_MAX_DEVICES] = {};
if (use_copy_engine && use_bf16) {
to_bf16_cuda_t to_bf16 = ggml_get_to_bf16_cuda(GGML_TYPE_F32);
for (int i = 0; i < n; ++i) {
auto * cuda_ctx = static_cast<ggml_backend_cuda_context *>(backends[i]->context);
GGML_ASSERT(cuda_ctx->device == p->devices[i]);
bf16_tmp[i].pool = &cuda_ctx->pool();
bf16_tmp[i].alloc(ne);
ggml_cuda_set_device(p->devices[i]);
if (compute_flag[i]) {
to_bf16(tensors[i]->data, bf16_tmp[i].get(), ne, cuda_ctx->stream());
CUDA_CHECK(cudaGetLastError());
} else {
CUDA_CHECK(cudaMemsetAsync(bf16_tmp[i].get(), 0, nbytes, cuda_ctx->stream()));
}
copy_src_ptr[i] = bf16_tmp[i].get();
}
}
bool ok = true;
if (use_copy_engine) {
// After up-front BF16 conversion, the tmp buffers already hold the
// (possibly zeroed-for-inactive) data, so the inner path can treat
// every shard as compute.
bool inner_compute[GGML_CUDA_MAX_DEVICES];
for (int i = 0; i < n; ++i) {
inner_compute[i] = use_bf16 ? true : compute_flag[i];
}
// Dispatch into copy_impl with explicit src/dst types. When use_bf16
// is on, the wire type is BF16 (src = bf16_tmp) and the accumulator
// is F32 (dst = tensors[i]->data); the combined add kernel rounds dst
// through BF16 for bit-equivalence and writes F32 directly, so no
// post-conversion is needed. Otherwise src == dst (same native type).
if (use_bf16) {
GGML_ASSERT(kernel_type == GGML_TYPE_BF16);
nv_bfloat16 * src[GGML_CUDA_MAX_DEVICES] = {};
float * dst[GGML_CUDA_MAX_DEVICES] = {};
for (int i = 0; i < n; ++i) {
src[i] = static_cast<nv_bfloat16 *>(copy_src_ptr[i]);
dst[i] = static_cast<float *>(tensors[i]->data);
}
ok = ggml_cuda_ar_allreduce_copy_outer<nv_bfloat16, float>(
p, backends, src, dst, inner_compute, ne);
} else {
switch (kernel_type) {
case GGML_TYPE_F32: {
float * buf[GGML_CUDA_MAX_DEVICES] = {};
for (int i = 0; i < n; ++i) {
buf[i] = static_cast<float *>(tensors[i]->data);
}
ok = ggml_cuda_ar_allreduce_copy_outer<float, float>(
p, backends, buf, buf, inner_compute, ne);
break;
}
case GGML_TYPE_BF16: {
nv_bfloat16 * buf[GGML_CUDA_MAX_DEVICES] = {};
for (int i = 0; i < n; ++i) {
buf[i] = static_cast<nv_bfloat16 *>(tensors[i]->data);
}
ok = ggml_cuda_ar_allreduce_copy_outer<nv_bfloat16, nv_bfloat16>(
p, backends, buf, buf, inner_compute, ne);
break;
}
case GGML_TYPE_F16: {
half * buf[GGML_CUDA_MAX_DEVICES] = {};
for (int i = 0; i < n; ++i) {
buf[i] = static_cast<half *>(tensors[i]->data);
}
ok = ggml_cuda_ar_allreduce_copy_outer<half, half>(
p, backends, buf, buf, inner_compute, ne);
break;
}
default:
GGML_ASSERT(false);
}
}
} else {
// host_buf carries T_wire-typed data; max_chunk_elems is the count that
// fits in one host_buf at the wire size.
const size_t max_chunk_elems = p->buf_bytes / type_size;
const size_t input_type_size = ggml_type_size(input_type);
// Chunked kernel path runs entirely on the caller's compute stream:
// since AR is a barrier here, same-stream ordering subsumes any
// cross-stream event handshake that the copy-engine path needs, and
// skips the cross-stream scheduling overhead that was hurting the
// small-tensor (tg) latency on the AR-stream variant. Only ev.ker is
// still recorded at end-of-AR for acquire_slot's pool-wraparound check.
for (int64_t chunk_start = 0; chunk_start < ne; chunk_start += (int64_t) max_chunk_elems) {
const size_t remaining_elems = (size_t) (ne - chunk_start);
const size_t chunk_elems = remaining_elems < max_chunk_elems ? remaining_elems : max_chunk_elems;
const size_t chunk_dst_bytes = chunk_elems * input_type_size;
const auto [slot, token] = ggml_cuda_ar_acquire_slot(p);
const bool last_chunk = chunk_start + (int64_t) chunk_elems == ne;
for (int i = 0; i < n; ++i) {
const int peer = 1 - i; // valid for n == 2 only
ggml_cuda_set_device(p->devices[i]);
auto * cuda_ctx = static_cast<ggml_backend_cuda_context *>(backends[i]->context);
GGML_ASSERT(cuda_ctx->device == p->devices[i]);
cudaStream_t stream = cuda_ctx->stream();
char * data = static_cast<char *>(tensors[i]->data) + chunk_start * (int64_t) input_type_size;
// Match NCCL/meta-backend semantics: inactive shards contribute
// zeros. On the BF16 path the F32 tensor data was already
// zeroed up-front (above), so per-chunk zeroing isn't needed.
if (!compute_flag[i] && !use_bf16) {
CUDA_CHECK(cudaMemsetAsync(data, 0, chunk_dst_bytes, stream));
}
#define LAUNCH_AR_KERNEL(T_dst, T_wire) \
ggml_cuda_ar_kernel<T_dst, T_wire><<<dim3(GGML_CUDA_AR_KERNEL_BLOCKS), dim3(256), 0, stream>>>( \
reinterpret_cast<const T_dst *>(data), \
reinterpret_cast<T_dst *>(data), \
reinterpret_cast<T_wire *>(p->host_buf[i].dev + (size_t) slot * p->buf_bytes), \
reinterpret_cast<const T_wire *>(p->host_buf[peer].dev + (size_t) slot * p->buf_bytes), \
static_cast<int>(chunk_elems), \
ggml_cuda_ar_arrival_ptr(p, slot, i), \
ggml_cuda_ar_arrival_ptr(p, slot, peer), \
token)
if (use_bf16) {
GGML_ASSERT(input_type == GGML_TYPE_F32);
LAUNCH_AR_KERNEL(float, nv_bfloat16);
} else {
switch (input_type) {
case GGML_TYPE_F32: LAUNCH_AR_KERNEL(float, float); break;
case GGML_TYPE_F16: LAUNCH_AR_KERNEL(half, half); break;
case GGML_TYPE_BF16: LAUNCH_AR_KERNEL(nv_bfloat16, nv_bfloat16); break;
default: GGML_ASSERT(false);
}
}
#undef LAUNCH_AR_KERNEL
CUDA_CHECK(cudaGetLastError());
if (last_chunk) {
CUDA_CHECK(cudaEventRecord(p->ev_pool[i][slot].ker, stream));
}
}
}
}
return ok;
}
#else // defined(GGML_USE_HIP) || defined(GGML_USE_MUSA)
// HIP and MUSA lack the host-mapped pinned-memory APIs (cudaHostAllocPortable
// / cudaHostAllocMapped / cudaHostGetDevicePointer) and __nanosleep that this
// implementation relies on, so the internal AllReduce is a CUDA-only feature.
// The dispatcher in ggml-cuda.cu treats a nullptr pipeline as "init failed"
// and silently falls back to the meta backend's generic AllReduce.
ggml_cuda_ar_pipeline * ggml_cuda_ar_pipeline_init(const int *, size_t) {
return nullptr;
}
void ggml_cuda_ar_pipeline_free(ggml_cuda_ar_pipeline *) {
}
bool ggml_cuda_ar_allreduce(ggml_cuda_ar_pipeline *, ggml_backend_t *, ggml_tensor **) {
return false;
}
#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
engine/ggml/src/ggml-cuda/allreduce.cuh
+29
View File
@@ -0,0 +1,29 @@
#pragma once
#include "common.cuh"
#include "ggml-backend-impl.h"
#include <cstddef>
// Opaque pipeline context -- owns all pinned buffers, streams, and events.
struct ggml_cuda_ar_pipeline;
// Allocate a pipeline for n_devices GPUs.
// devices[] holds the CUDA device IDs in rank order.
// Returns nullptr on allocation failure.
ggml_cuda_ar_pipeline * ggml_cuda_ar_pipeline_init(
const int * devices, size_t n_devices);
// Release all resources owned by the pipeline.
void ggml_cuda_ar_pipeline_free(ggml_cuda_ar_pipeline * pipeline);
// Execute an in-place AllReduce (sum) across tensors[0..n_devices-1].
// tensors[i] must live on the device managed by backends[i] and be
// contiguous F32, F16, or BF16.
// Preconditions are checked by the CUDA comm dispatcher before calling this.
// Returns true once the reduction work has been enqueued successfully.
bool ggml_cuda_ar_allreduce(
ggml_cuda_ar_pipeline * pipeline,
ggml_backend_t * backends,
ggml_tensor ** tensors);
engine/ggml/src/ggml-cuda/arange.cu
+34
View File
@@ -0,0 +1,34 @@
#include "arange.cuh"
static __global__ void arange_f32(float * dst, const int ne0, const float start, const float step) {
// blockIDx.x: idx of ne0 / BLOCK_SIZE
int nidx = threadIdx.x + blockIdx.x * blockDim.x;
if (nidx >= ne0) {
return;
}
dst[nidx] = start + step * nidx;
}
static void arange_f32_cuda(float * dst, const int ne0, const float start, const float step, cudaStream_t stream) {
int num_blocks = (ne0 + CUDA_ARANGE_BLOCK_SIZE - 1) / CUDA_ARANGE_BLOCK_SIZE;
arange_f32<<<num_blocks, CUDA_ARANGE_BLOCK_SIZE, 0, stream>>>(dst, ne0, start, step);
}
void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(dst->type == GGML_TYPE_F32);
float start;
float stop;
float step;
memcpy(&start, (float *)dst->op_params + 0, sizeof(float));
memcpy(&stop, (float *)dst->op_params + 1, sizeof(float));
memcpy(&step, (float *)dst->op_params + 2, sizeof(float));
int64_t steps = (int64_t)ceil((stop - start) / step);
GGML_ASSERT(ggml_nelements(dst) == steps);
arange_f32_cuda(dst_d, dst->ne[0], start, step, stream);
}
engine/ggml/src/ggml-cuda/arange.cuh
+5
View File
@@ -0,0 +1,5 @@
#include "common.cuh"
#define CUDA_ARANGE_BLOCK_SIZE 256
void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
engine/ggml/src/ggml-cuda/argmax.cu
+91
View File
@@ -0,0 +1,91 @@
#include <algorithm>
#include <cstdint>
#include "argmax.cuh"
#include "common.cuh"
#include "sum.cuh"
static __global__ void argmax_f32(const float * __restrict__ x, int32_t * __restrict__ dst, const int64_t ncols) {
const int64_t row = blockIdx.x;
float maxval = -FLT_MAX;
int argmax = -1;
const float * rowx = x + row * ncols;
for (int32_t col = threadIdx.x; col < ncols; col += blockDim.x) {
const float val = rowx[col];
if (val > maxval) {
maxval = val;
argmax = col;
}
}
#pragma unroll
for (int offset = WARP_SIZE/2; offset > 0; offset >>= 1) {
const float val = __shfl_xor_sync(0xFFFFFFFF, maxval, offset, WARP_SIZE);
const int col = __shfl_xor_sync(0xFFFFFFFF, argmax, offset, WARP_SIZE);
if (val > maxval) {
maxval = val;
argmax = col;
}
}
const int n_warps = blockDim.x / WARP_SIZE;
const int lane_id = threadIdx.x % WARP_SIZE;
const int warp_id = threadIdx.x / WARP_SIZE;
if (n_warps > 1) {
constexpr int max_warps = 1024 / WARP_SIZE;
__shared__ float shared_maxval[max_warps];
__shared__ int shared_argmax[max_warps];
if (lane_id == 0) {
shared_maxval[warp_id] = maxval;
shared_argmax[warp_id] = argmax;
}
__syncthreads();
if (warp_id == 0) {
if (lane_id < n_warps) {
maxval = shared_maxval[lane_id];
argmax = shared_argmax[lane_id];
}
#pragma unroll
for (int offset = WARP_SIZE/2; offset > 0; offset >>= 1) {
const float val = __shfl_xor_sync(0xFFFFFFFF, maxval, offset, WARP_SIZE);
const int col = __shfl_xor_sync(0xFFFFFFFF, argmax, offset, WARP_SIZE);
if (val > maxval) {
maxval = val;
argmax = col;
}
}
}
}
if (warp_id == 0 && lane_id == 0) {
dst[row] = argmax;
}
}
void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_I32);
GGML_ASSERT(ggml_is_contiguous(src0));
const int64_t ne00 = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
const float * src0_d = (const float *) src0->data;
int32_t * dst_d = (int32_t *) dst->data;
cudaStream_t stream = ctx.stream();
const int64_t num_blocks = nrows;
const int64_t num_threads = std::min<int64_t>(1024, (ne00 + WARP_SIZE - 1) / WARP_SIZE * WARP_SIZE);
const dim3 blocks_dim(num_threads, 1, 1);
const dim3 blocks_num(num_blocks, 1, 1);
argmax_f32<<<blocks_num, blocks_dim, 0, stream>>>(src0_d, dst_d, ne00);
}
engine/ggml/src/ggml-cuda/argmax.cuh
+3
View File
@@ -0,0 +1,3 @@
#include "common.cuh"
void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
engine/ggml/src/ggml-cuda/argsort.cu
+266
View File
@@ -0,0 +1,266 @@
#include "argsort.cuh"
#ifdef GGML_CUDA_USE_CUB
# include <cub/cub.cuh>
# if (CCCL_MAJOR_VERSION >= 3 && CCCL_MINOR_VERSION >= 1)
# define STRIDED_ITERATOR_AVAILABLE
# include <cuda/iterator>
# endif
using namespace cub;
#endif // GGML_CUDA_USE_CUB
static __global__ void init_indices(int * indices, const int ncols, const int nrows) {
const int col = blockIdx.x * blockDim.x + threadIdx.x;
const int row = blockIdx.y;
if (col < ncols && row < nrows) {
indices[row * ncols + col] = col;
}
}
#ifndef STRIDED_ITERATOR_AVAILABLE
static __global__ void init_offsets(int * offsets, const int ncols, const int nrows) {
const int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx <= nrows) {
offsets[idx] = idx * ncols;
}
}
#endif // STRIDED_ITERATOR_AVAILABLE
#ifdef GGML_CUDA_USE_CUB
void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
const float * x,
int * dst,
const int ncols,
const int nrows,
ggml_sort_order order,
cudaStream_t stream) {
ggml_cuda_pool_alloc<int> temp_indices_alloc(pool, ncols * nrows);
ggml_cuda_pool_alloc<float> temp_keys_alloc(pool, ncols * nrows);
int * temp_indices = temp_indices_alloc.get();
float * temp_keys = temp_keys_alloc.get();
static const int block_size = 256;
const dim3 grid_size((ncols + block_size - 1) / block_size, nrows);
init_indices<<<grid_size, block_size, 0, stream>>>(temp_indices, ncols, nrows);
#ifdef STRIDED_ITERATOR_AVAILABLE
auto offset_iterator = cuda::make_strided_iterator(cuda::make_counting_iterator(0), ncols);
#else
// offset_iterator needs to populate nrows + 1 elements, so we also have to ceildiv nrows + 1 by block_size
const int nrows_offset = nrows + 1;
ggml_cuda_pool_alloc<int> offsets_alloc(pool, nrows_offset);
int * offset_iterator = offsets_alloc.get();
const dim3 offset_grid((nrows_offset + block_size - 1) / block_size);
init_offsets<<<offset_grid, block_size, 0, stream>>>(offset_iterator, ncols, nrows);
#endif
CUDA_CHECK(cudaMemcpyAsync(temp_keys, x, ncols * nrows * sizeof(float), cudaMemcpyDeviceToDevice, stream));
size_t temp_storage_bytes = 0;
bool is_capturing = false;
#ifdef USE_CUDA_GRAPH
// Currently (confirmed for CCCL <= 3.2) DeviceSegmentedSort does not support stream capture, while DeviceSegmentedRadixSort does.
// See https://github.com/NVIDIA/cccl/issues/5661#issuecomment-3229037149
// TODO: constrain this to the CCCL versions that have this issue once it's resolved in a future CCCL release.
cudaStreamCaptureStatus capture_status;
CUDA_CHECK(cudaStreamIsCapturing(stream, &capture_status));
is_capturing = (capture_status != cudaStreamCaptureStatusNone);
#endif // USE_CUDA_GRAPH
if (order == GGML_SORT_ORDER_ASC) {
if (nrows == 1) {
CUDA_CHECK(DeviceRadixSort::SortPairs(nullptr, temp_storage_bytes, temp_keys, temp_keys, // keys (in-place)
temp_indices, dst, // values (indices)
ncols, 0, sizeof(float) * 8, stream));
} else if (is_capturing) {
CUDA_CHECK(DeviceSegmentedRadixSort::SortPairs(
nullptr, temp_storage_bytes, temp_keys, temp_keys, // keys (in-place)
temp_indices, dst, // values (indices)
ncols * nrows, nrows, // num items, num segments
offset_iterator, offset_iterator + 1, 0, sizeof(float) * 8, stream));
} else {
CUDA_CHECK(DeviceSegmentedSort::SortPairs(nullptr, temp_storage_bytes, temp_keys,
temp_keys, // keys (in-place)
temp_indices, dst, // values (indices)
ncols * nrows, nrows, // num items, num segments
offset_iterator, offset_iterator + 1, stream));
}
} else {
if (nrows == 1) {
CUDA_CHECK(DeviceRadixSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys,
temp_keys, // keys (in-place)
temp_indices, dst, // values (indices)
ncols, 0, sizeof(float) * 8, stream));
} else if (is_capturing) {
CUDA_CHECK(DeviceSegmentedRadixSort::SortPairsDescending(
nullptr, temp_storage_bytes, temp_keys, temp_keys, temp_indices, dst, ncols * nrows, nrows,
offset_iterator, offset_iterator + 1, 0, sizeof(float) * 8, stream));
} else {
CUDA_CHECK(DeviceSegmentedSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys,
temp_indices, dst, ncols * nrows, nrows,
offset_iterator, offset_iterator + 1, stream));
}
}
ggml_cuda_pool_alloc<uint8_t> temp_storage_alloc(pool, temp_storage_bytes);
void * d_temp_storage = temp_storage_alloc.get();
if (order == GGML_SORT_ORDER_ASC) {
if (nrows == 1) {
CUDA_CHECK(DeviceRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys,
temp_keys, // keys (in-place)
temp_indices, dst, // values (indices)
ncols, 0, sizeof(float) * 8, stream));
} else if (is_capturing) {
CUDA_CHECK(DeviceSegmentedRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,
temp_indices, dst, ncols * nrows, nrows, offset_iterator,
offset_iterator + 1, 0, sizeof(float) * 8, stream));
} else {
CUDA_CHECK(DeviceSegmentedSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,
temp_indices, dst, ncols * nrows, nrows, offset_iterator,
offset_iterator + 1, stream));
}
} else {
if (nrows == 1) {
CUDA_CHECK(DeviceRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys,
temp_keys, // keys (in-place)
temp_indices, dst, // values (indices)
ncols, 0, sizeof(float) * 8, stream));
} else if (is_capturing) {
CUDA_CHECK(DeviceSegmentedRadixSort::SortPairsDescending(
d_temp_storage, temp_storage_bytes, temp_keys, temp_keys, temp_indices, dst, ncols * nrows, nrows,
offset_iterator, offset_iterator + 1, 0, sizeof(float) * 8, stream));
} else {
CUDA_CHECK(DeviceSegmentedSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys,
temp_keys, temp_indices, dst, ncols * nrows, nrows,
offset_iterator, offset_iterator + 1, stream));
}
}
}
#endif // GGML_CUDA_USE_CUB
// Bitonic sort implementation
template<typename T>
static inline __device__ void ggml_cuda_swap(T & a, T & b) {
T tmp = a;
a = b;
b = tmp;
}
template<ggml_sort_order order>
static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols, int ncols_pad) {
// bitonic sort
int col = threadIdx.x;
int row = blockIdx.x;
if (col >= ncols_pad) {
return;
}
const float * x_row = x + row * ncols;
extern __shared__ int dst_row[];
// initialize indices
dst_row[col] = col;
__syncthreads();
for (int k = 2; k <= ncols_pad; k *= 2) {
for (int j = k / 2; j > 0; j /= 2) {
int ixj = col ^ j;
if (ixj > col) {
if ((col & k) == 0) {
if (dst_row[col] >= ncols ||
(dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ?
x_row[dst_row[col]] > x_row[dst_row[ixj]] :
x_row[dst_row[col]] < x_row[dst_row[ixj]]))
) {
ggml_cuda_swap(dst_row[col], dst_row[ixj]);
}
} else {
if (dst_row[ixj] >= ncols ||
(dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ?
x_row[dst_row[col]] < x_row[dst_row[ixj]] :
x_row[dst_row[col]] > x_row[dst_row[ixj]]))
) {
ggml_cuda_swap(dst_row[col], dst_row[ixj]);
}
}
}
__syncthreads();
}
}
// copy the result to dst without the padding
if (col < ncols) {
dst[row * ncols + col] = dst_row[col];
}
}
static int next_power_of_2(int x) {
int n = 1;
while (n < x) {
n *= 2;
}
return n;
}
void argsort_f32_i32_cuda_bitonic(const float * x,
int * dst,
const int ncols,
const int nrows,
ggml_sort_order order,
cudaStream_t stream) {
// bitonic sort requires ncols to be power of 2
const int ncols_pad = next_power_of_2(ncols);
const dim3 block_dims(ncols_pad, 1, 1);
const dim3 block_nums(nrows, 1, 1);
const size_t shared_mem = ncols_pad * sizeof(int);
// FIXME: this limit could be raised by ~2-4x on Ampere or newer
GGML_ASSERT(shared_mem <= ggml_cuda_info().devices[ggml_cuda_get_device()].smpb);
if (order == GGML_SORT_ORDER_ASC) {
k_argsort_f32_i32<GGML_SORT_ORDER_ASC>
<<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
} else if (order == GGML_SORT_ORDER_DESC) {
k_argsort_f32_i32<GGML_SORT_ORDER_DESC>
<<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
} else {
GGML_ABORT("fatal error");
}
}
void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_I32);
GGML_ASSERT(ggml_is_contiguous(src0));
const int64_t ncols = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
#ifdef GGML_CUDA_USE_CUB
const int ncols_pad = next_power_of_2(ncols);
const size_t shared_mem = ncols_pad * sizeof(int);
const size_t max_shared_mem = ggml_cuda_info().devices[ggml_cuda_get_device()].smpb;
if (shared_mem > max_shared_mem || ncols > 1024) {
ggml_cuda_pool & pool = ctx.pool();
argsort_f32_i32_cuda_cub(pool, src0_d, (int *) dst_d, ncols, nrows, order, stream);
} else {
argsort_f32_i32_cuda_bitonic(src0_d, (int *) dst_d, ncols, nrows, order, stream);
}
#else
argsort_f32_i32_cuda_bitonic(src0_d, (int *) dst_d, ncols, nrows, order, stream);
#endif
}
engine/ggml/src/ggml-cuda/argsort.cuh
+19
View File
@@ -0,0 +1,19 @@
#include "common.cuh"
void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
#ifdef GGML_CUDA_USE_CUB
void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
const float * x,
int * dst,
const int ncols,
const int nrows,
ggml_sort_order order,
cudaStream_t stream);
#endif // GGML_CUDA_USE_CUB
void argsort_f32_i32_cuda_bitonic(const float * x,
int * dst,
const int ncols,
const int nrows,
ggml_sort_order order,
cudaStream_t stream);
engine/ggml/src/ggml-cuda/binbcast.cu
+530
View File
@@ -0,0 +1,530 @@
#include "binbcast.cuh"
#include <cstdint>
#include <utility>
template<typename T, size_t>
using type_for_index = T;
static __device__ __forceinline__ float op_repeat(const float a, const float b) {
return b;
GGML_UNUSED(a);
}
static __device__ __forceinline__ float op_add(const float a, const float b) {
return a + b;
}
static __device__ __forceinline__ float op_sub(const float a, const float b) {
return a - b;
}
static __device__ __forceinline__ float op_mul(const float a, const float b) {
return a * b;
}
static __device__ __forceinline__ float op_div(const float a, const float b) {
return a / b;
}
template <float (*bin_op)(const float, const float),
typename src0_t,
typename src1_t,
typename dst_t,
typename... src1_ptrs>
static __global__ void k_bin_bcast(const src0_t * src0,
const src1_t * src1,
dst_t * dst,
const int ne0,
const int ne1,
const int ne2,
const uint3 ne3,
const uint3 ne10,
const uint3 ne11,
const uint3 ne12,
const uint3 ne13,
/*const int s0,*/
const int s1,
const int s2,
const int s3,
const int s00,
const int s01,
const int s02,
const int s03,
const int s10,
const int s11,
const int s12,
const int s13,
src1_ptrs... src1s) {
ggml_cuda_pdl_lc();
const uint32_t i0s = blockDim.x * blockIdx.x + threadIdx.x;
const uint32_t i1 = (blockDim.y * blockIdx.y + threadIdx.y);
const uint32_t i2 = fastdiv((blockDim.z * blockIdx.z + threadIdx.z), ne3);
const uint32_t i3 = (blockDim.z * blockIdx.z + threadIdx.z) - (i2 * ne3.z);
if (i0s >= (uint32_t)ne0 || i1 >= (uint32_t)ne1 || i2 >= (uint32_t)ne2 || i3 >= ne3.z) {
return;
}
const uint32_t i11 = fastmodulo(i1, ne11);
const uint32_t i12 = fastmodulo(i2, ne12);
const uint32_t i13 = fastmodulo(i3, ne13);
const size_t i_src0 = i3*s03 + i2*s02 + i1*s01;
const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
const size_t i_dst = i3*s3 + i2*s2 + i1*s1;
const src0_t * src0_row = src0 ? (src0 + i_src0) : nullptr;
dst_t * dst_row = dst + i_dst;
ggml_cuda_pdl_sync();
for (int i0 = i0s; i0 < ne0; i0 += blockDim.x * gridDim.x) {
const uint32_t i10 = fastmodulo(i0, ne10);
float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
if constexpr (sizeof...(src1_ptrs) > 0) {
result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
} else {
result = bin_op(result, (float)src1[i_src1 + i10*s10]);
}
dst_row[i0] = (dst_t) result;
}
}
template <float (*bin_op)(const float, const float),
typename src0_t,
typename src1_t,
typename dst_t,
typename... src1_ptrs>
static __global__ void k_bin_bcast_unravel(const src0_t * src0,
const src1_t * src1,
dst_t * dst,
const uint3 ne0,
const uint3 ne1,
const uint3 ne2,
const uint32_t ne3,
const uint3 prod_012,
const uint3 prod_01,
const uint3 ne10,
const uint3 ne11,
const uint3 ne12,
const uint3 ne13,
/*const int s0,*/
const int s1,
const int s2,
const int s3,
const int s00,
const int s01,
const int s02,
const int s03,
const int s10,
const int s11,
const int s12,
const int s13,
src1_ptrs... src1s) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
const uint32_t i3 = fastdiv(i, prod_012);
const uint32_t i2 = fastdiv(i - i3 * prod_012.z, prod_01);
const uint32_t i1 = fastdiv(i - i3 * prod_012.z - i2 * prod_01.z, ne0);
const uint32_t i0 = i - i3 * prod_012.z - i2 * prod_01.z - i1 * ne0.z;
if (i0 >= ne0.z || i1 >= ne1.z || i2 >= ne2.z || i3 >= ne3) {
return;
}
const int i11 = fastmodulo(i1, ne11);
const int i12 = fastmodulo(i2, ne12);
const int i13 = fastmodulo(i3, ne13);
const size_t i_src0 = i3*s03 + i2*s02 + i1*s01;
const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
const size_t i_dst = i3*s3 + i2*s2 + i1*s1;
const src0_t * src0_row = src0 ? (src0 + i_src0) : nullptr;
dst_t * dst_row = dst + i_dst;
const int i10 = fastmodulo(i0, ne10);
ggml_cuda_pdl_sync();
float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
if constexpr (sizeof...(src1_ptrs) > 0) {
result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
} else {
result = bin_op(result, (float)src1[i_src1 + i10*s10]);
}
dst_row[i0] = (dst_t) result;
}
template <float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t, size_t... I>
static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd,
cudaStream_t stream, std::index_sequence<I...>) {
GGML_TENSOR_BINARY_OP_LOCALS
int nr0 = ne10 / ne0;
int nr1 = ne11 / ne1;
int nr2 = ne12 / ne2;
int nr3 = ne13 / ne3;
int nr[4] = { nr0, nr1, nr2, nr3 };
int64_t cne[] = { ne0, ne1, ne2, ne3 };
int64_t cne0[] = { ne00, ne01, ne02, ne03 };
int64_t cne1[] = { ne10, ne11, ne12, ne13 };
size_t cnb[] = { nb0, nb1, nb2, nb3 };
size_t cnb0[] = { nb00, nb01, nb02, nb03 };
size_t cnb1[] = { nb10, nb11, nb12, nb13 };
auto collapse = [](int64_t cne[]) {
cne[0] *= cne[1];
cne[1] = cne[2];
cne[2] = cne[3];
cne[3] = 1;
};
auto collapse_nb = [](size_t cnb[], const int64_t cne[]) {
cnb[1] *= cne[1];
cnb[2] *= cne[2];
cnb[3] *= cne[3];
};
if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && !ggml_is_permuted(src0) && !ggml_is_permuted(src1)) {
for (int i = 0; i < 4; i++) {
if (nr[i] != 1) {
break;
}
if (i > 0) {
collapse_nb(cnb, cne);
collapse_nb(cnb0, cne0);
collapse_nb(cnb1, cne1);
collapse(cne);
collapse(cne0);
collapse(cne1);
}
}
}
{
int64_t ne0 = cne[0];
int64_t ne1 = cne[1];
int64_t ne2 = cne[2];
int64_t ne3 = cne[3];
//int64_t ne00 = cne0[0]; GGML_UNUSED(ne00);
//int64_t ne01 = cne0[1]; GGML_UNUSED(ne01);
//int64_t ne02 = cne0[2]; GGML_UNUSED(ne02);
//int64_t ne03 = cne0[3]; GGML_UNUSED(ne03);
size_t nb0 = cnb[0];
size_t nb1 = cnb[1];
size_t nb2 = cnb[2];
size_t nb3 = cnb[3];
size_t nb00 = cnb0[0];
size_t nb01 = cnb0[1];
size_t nb02 = cnb0[2];
size_t nb03 = cnb0[3];
size_t nb10 = cnb1[0];
size_t nb11 = cnb1[1];
size_t nb12 = cnb1[2];
size_t nb13 = cnb1[3];
//size_t s0 = nb0 / sizeof(dst_t);
size_t s1 = nb1 / sizeof(dst_t);
size_t s2 = nb2 / sizeof(dst_t);
size_t s3 = nb3 / sizeof(dst_t);
size_t s10 = nb10 / sizeof(src1_t);
size_t s11 = nb11 / sizeof(src1_t);
size_t s12 = nb12 / sizeof(src1_t);
size_t s13 = nb13 / sizeof(src1_t);
size_t s00 = nb00 / sizeof(src0_t);
size_t s01 = nb01 / sizeof(src0_t);
size_t s02 = nb02 / sizeof(src0_t);
size_t s03 = nb03 / sizeof(src0_t);
GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
GGML_ASSERT(nb3 % sizeof(dst_t) == 0);
GGML_ASSERT(nb00 % sizeof(src0_t) == 0);
GGML_ASSERT(nb01 % sizeof(src0_t) == 0);
GGML_ASSERT(nb02 % sizeof(src0_t) == 0);
GGML_ASSERT(nb03 % sizeof(src0_t) == 0);
GGML_ASSERT(nb10 % sizeof(src1_t) == 0);
GGML_ASSERT(nb11 % sizeof(src1_t) == 0);
GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
const int block_size = 128;
int64_t hne0 = std::max(ne0 / 2LL, 1LL);
dim3 block_dims;
block_dims.x = std::min<unsigned int>(hne0, block_size);
block_dims.y = std::min<unsigned int>(ne1, block_size / block_dims.x);
block_dims.z = std::min(std::min<unsigned int>(ne2 * ne3, block_size / block_dims.x / block_dims.y), 64U);
dim3 block_nums((hne0 + block_dims.x - 1) / block_dims.x, (ne1 + block_dims.y - 1) / block_dims.y,
(ne2 * ne3 + block_dims.z - 1) / block_dims.z);
const uint3 ne10 = init_fastdiv_values((uint32_t) cne1[0]);
const uint3 ne11 = init_fastdiv_values((uint32_t) cne1[1]);
const uint3 ne12 = init_fastdiv_values((uint32_t) cne1[2]);
const uint3 ne13 = init_fastdiv_values((uint32_t) cne1[3]);
if (block_nums.z > 65535 || block_nums.y > 65535) {
int block_num = (ne0 * ne1 * ne2 * ne3 + block_size - 1) / block_size;
const uint3 prod_012 = init_fastdiv_values((uint32_t) (ne0 * ne1 * ne2));
const uint3 prod_01 = init_fastdiv_values((uint32_t) (ne0 * ne1));
const uint3 ne0_fastdiv = init_fastdiv_values((uint32_t) ne0);
const uint3 ne1_fastdiv = init_fastdiv_values((uint32_t) ne1);
const uint3 ne2_fastdiv = init_fastdiv_values((uint32_t) ne2);
{
const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)block_num, block_size, 0, stream);
ggml_cuda_kernel_launch(k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t, type_for_index<const src1_t *, I>...>, launch_params,
src0_dd, src1_dd, dst_dd, ne0_fastdiv, ne1_fastdiv, ne2_fastdiv, ne3, prod_012, prod_01, ne10, ne11,
ne12, ne13,
/*s0,*/ s1, s2, s3,
s00, s01, s02, s03,
s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
}
} else {
const uint3 ne3_fastdiv = init_fastdiv_values((uint32_t) ne3);
{
const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
ggml_cuda_kernel_launch(k_bin_bcast<bin_op, src0_t, src1_t, dst_t, type_for_index<const src1_t *, I>...>, launch_params,
src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3_fastdiv, ne10, ne11, ne12, ne13,
/*s0,*/ s1, s2, s3,
s00, s01, s02, s03,
s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
}
}
}
}
template <typename T>
static __global__ void k_repeat_back(
const T * __restrict__ src, T * __restrict__ dst, const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
const size_t s00, const size_t s01, const size_t s02, const size_t s03,
const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3) {
const int64_t tid0 = int64_t(blockIdx.x)*blockDim.x + threadIdx.x;
const int64_t tid1 = int64_t(blockIdx.y)*blockDim.y + threadIdx.y;
const int64_t tid23 = int64_t(blockIdx.z)*blockDim.z + threadIdx.z;
const int64_t tid2 = tid23 % ne2;
const int64_t tid3 = tid23 / ne2;
if (tid0 >= ne0) {
return;
}
T sum = 0;
ggml_cuda_pdl_sync();
for (int64_t i3 = tid3; i3 < ne03; i3 += ne3) {
for (int64_t i2 = tid2; i2 < ne02; i2 += ne2) {
for (int64_t i1 = tid1; i1 < ne01; i1 += ne1) {
for (int64_t i0 = tid0; i0 < ne00; i0 += ne0) {
sum += src[i3*s03 + i2*s02 + i1*s01 + i0*s00];
}
}
}
}
dst[tid3*ne2*ne1*ne0 + tid2*ne1*ne0 + tid1*ne0 + tid0] = sum;
}
template <float (*bin_op)(const float, const float), int n_fuse = 1>
struct bin_bcast_cuda {
template<typename src0_t, typename src1_t, typename dst_t>
void operator()(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst,
const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd,
cudaStream_t stream) {
launch_bin_bcast_pack<bin_op, src0_t, src1_t, dst_t>(
src0, src1, dst, src0_dd, src1_dd, dst_dd, stream, std::make_index_sequence<n_fuse>{});
}
};
template <typename T>
static void repeat_back_cuda(
const T * src, T * dst, const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
const size_t s00, const size_t s01, const size_t s02, const size_t s03,
const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
const dim3 block_dims(WARP_SIZE, 1, 1);
const dim3 block_nums((ne0 + WARP_SIZE - 1) / WARP_SIZE, ne1, ne2*ne3);
k_repeat_back<T><<<block_nums, block_dims, 0, stream>>>
(src, dst, ne00, ne01, ne02, ne03, s00, s01, s02, s03, ne0, ne1, ne2, ne3);
}
template<class op>
static void ggml_cuda_op_bin_bcast(
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
const void * src0_dd, const void * src1_dd, void * dst_dd, cudaStream_t stream) {
GGML_ASSERT(src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
op()(src0, src1, dst, (const float *)src0_dd, (const float *)src1_dd, (float *)dst_dd, stream);
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
op()(src0, src1, dst, (const half *) src0_dd, (const half *)src1_dd, (half *) dst_dd, stream);
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (half *) dst_dd, stream);
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (float *)dst_dd, stream);
} else {
fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
GGML_ABORT("fatal error");
}
}
void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_repeat, 0>>(dst, dst->src[0], dst, nullptr, dst->src[0]->data, dst->data, ctx.stream());
}
void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_add>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
}
void ggml_cuda_op_sub(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_sub>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
}
void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_mul>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
}
void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_div>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
}
template <float (*op)(const float, const float), int n_fuse>
static void ggml_cuda_op_fused_binbcast_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
cudaStream_t stream = ctx.stream();
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
launch_bin_bcast_pack<op, float, float, float>(src0, src1, dst,
(const float *) src0->data, (const float *) src1->data, (float *) dst->data,
stream, std::make_index_sequence<n_fuse>{});
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
launch_bin_bcast_pack<op, half, half, half>(src0, src1, dst,
(const half *) src0->data, (const half *) src1->data, (half *) dst->data,
stream, std::make_index_sequence<n_fuse>{});
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
launch_bin_bcast_pack<op, half, float, half>(src0, src1, dst,
(const half *) src0->data, (const float *) src1->data, (half *) dst->data,
stream, std::make_index_sequence<n_fuse>{});
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
launch_bin_bcast_pack<op, half, float, float>(src0, src1, dst,
(const half *) src0->data, (const float *) src1->data, (float *) dst->data,
stream, std::make_index_sequence<n_fuse>{});
} else {
fprintf(stderr,
"%s: unsupported types for fusion: dst: %s, src0: %s, src1: %s\n",
__func__, ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
GGML_ABORT("fatal error");
}
}
void ggml_cuda_op_fused_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst, int n_fuse) {
GGML_ASSERT(2 <= n_fuse && n_fuse <= 8);
switch (n_fuse) {
case 2:
ggml_cuda_op_fused_binbcast_impl<op_add, 2>(ctx, dst);
break;
case 3:
ggml_cuda_op_fused_binbcast_impl<op_add, 3>(ctx, dst);
break;
case 4:
ggml_cuda_op_fused_binbcast_impl<op_add, 4>(ctx, dst);
break;
case 5:
ggml_cuda_op_fused_binbcast_impl<op_add, 5>(ctx, dst);
break;
case 6:
ggml_cuda_op_fused_binbcast_impl<op_add, 6>(ctx, dst);
break;
case 7:
ggml_cuda_op_fused_binbcast_impl<op_add, 7>(ctx, dst);
break;
case 8:
ggml_cuda_op_fused_binbcast_impl<op_add, 8>(ctx, dst);
break;
default:
GGML_ASSERT(false && "Unsupported n_fuse value");
}
}
void ggml_cuda_op_fused_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst, int n_fuse) {
GGML_ASSERT(2 <= n_fuse && n_fuse <= 8);
switch (n_fuse) {
case 2:
ggml_cuda_op_fused_binbcast_impl<op_mul, 2>(ctx, dst);
break;
case 3:
ggml_cuda_op_fused_binbcast_impl<op_mul, 3>(ctx, dst);
break;
case 4:
ggml_cuda_op_fused_binbcast_impl<op_mul, 4>(ctx, dst);
break;
case 5:
ggml_cuda_op_fused_binbcast_impl<op_mul, 5>(ctx, dst);
break;
case 6:
ggml_cuda_op_fused_binbcast_impl<op_mul, 6>(ctx, dst);
break;
case 7:
ggml_cuda_op_fused_binbcast_impl<op_mul, 7>(ctx, dst);
break;
case 8:
ggml_cuda_op_fused_binbcast_impl<op_mul, 8>(ctx, dst);
break;
default:
GGML_ASSERT(false && "Unsupported n_fuse value");
}
}
void ggml_cuda_op_repeat_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
GGML_ASSERT(src0->type == dst->type);
GGML_ASSERT(ggml_is_contiguous(dst));
GGML_ASSERT(ggml_can_repeat(dst, src0));
cudaStream_t stream = ctx.stream();
GGML_TENSOR_UNARY_OP_LOCALS;
GGML_ASSERT(ne2*ne3 <= (1 << 15));
const size_t ts = ggml_type_size(src0->type);
const size_t s00 = nb00 / ts;
const size_t s01 = nb01 / ts;
const size_t s02 = nb02 / ts;
const size_t s03 = nb03 / ts;
switch (dst->type) {
case GGML_TYPE_F32: {
const float * src0_d = (const float *) src0->data;
float * dst_d = (float *) dst->data;
repeat_back_cuda(src0_d, dst_d, ne00, ne01, ne02, ne03, s00, s01, s02, s03, ne0, ne1, ne2, ne3, stream);
} break;
default: {
GGML_ASSERT(false);
} break;
}
}
engine/ggml/src/ggml-cuda/binbcast.cuh
+12
View File
@@ -0,0 +1,12 @@
#include "common.cuh"
void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_sub(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_repeat_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_fused_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst, int n_fuse);
void ggml_cuda_op_fused_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst, int n_fuse);

Some files were not shown because too many files have changed in this diff Show More