diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 8f741b97091..241de9be1b0 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -28,9 +28,9 @@ jobs: -w /workspace ${{ env.ubuntu_image }} /bin/sh -c ' set -e apt update - apt install -y build-essential libsdl2-dev - make - make stream' + apt install -y build-essential libsdl2-dev cmake + cmake -B build + cmake --build build --config Release -j $(nproc)' macOS-latest: runs-on: macOS-latest @@ -42,30 +42,30 @@ jobs: - name: Dependencies run: | brew update - brew install sdl2 + brew install sdl2 cmake - name: Build run: | - make - make stream + cmake -B build + cmake --build build --config Release - freeBSD-latest: - runs-on: macos-12 - - steps: - - name: Clone - uses: actions/checkout@v4 - - - name: Build - uses: cross-platform-actions/action@v0.24.0 - with: - operating_system: freebsd - version: '13.3' - run: | - sudo pkg update - sudo pkg install -y gmake sdl2 - gmake - gmake stream +# freeBSD-latest: +# runs-on: macos-12 +# +# steps: +# - name: Clone +# uses: actions/checkout@v4 +# +# - name: Build +# uses: cross-platform-actions/action@v0.24.0 +# with: +# operating_system: freebsd +# version: '13.3' +# run: | +# sudo pkg update +# sudo pkg install -y gmake sdl2 cmake +# cmake -B build +# cmake --build build --config Release ubuntu-latest-gcc: runs-on: ubuntu-latest @@ -280,21 +280,6 @@ jobs: mingw-w64-${{matrix.env}}-SDL2 mingw-w64-${{matrix.env}}-openblas - - name: Build using make - shell: msys2 {0} - run: | - make -j $(nproc) - - - name: Clean after building using make - shell: msys2 {0} - run: | - make clean - - - name: Build using make w/ OpenBLAS - shell: msys2 {0} - run: | - make GGML_OPENBLAS=1 -j $(nproc) - - name: Build using CMake shell: msys2 {0} run: | @@ -445,71 +430,72 @@ jobs: name: whisper-blas-bin-${{ matrix.arch }} path: build/bin/${{ matrix.build }} - windows-cublas: - runs-on: windows-2019 - - strategy: - matrix: - build: [Release] - arch: [x64] - cublas: [ON] - sdl2: [ON] - cuda-toolkit: [12.2.0, 11.8.0] - include: - - arch: x64 - s2arc: x64 - - sdl2: ON - s2ver: 2.28.5 - - steps: - - name: Clone - uses: actions/checkout@v4 - - - name: Add msbuild to PATH - uses: microsoft/setup-msbuild@v2 - - - name: Install CUDA Toolkit - id: cuda-toolkit - uses: Jimver/cuda-toolkit@v0.2.15 - with: - cuda: '${{ matrix.cuda-toolkit }}' - - - name: Fetch SDL2 and set SDL2_DIR - if: matrix.sdl2 == 'ON' - run: | - C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ matrix.s2ver }}/SDL2-devel-${{ matrix.s2ver }}-VC.zip - 7z x sdl2.zip - echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV - - - name: Configure - run: > - cmake -S . -B ./build -A ${{ matrix.arch }} - -DCMAKE_BUILD_TYPE=${{ matrix.build }} - -DGGML_CUDA=${{ matrix.cublas }} - -DWHISPER_SDL2=${{ matrix.sdl2 }} - - - name: Build ${{ matrix.cuda-toolkit }} - run: | - cd ./build - cmake --build . --config ${{ matrix.build }} - - - name: Copy CUDA DLLs - run: > - Copy-Item -PassThru - -Path "${{ steps.cuda-toolkit.outputs.CUDA_PATH }}/bin/*.dll" - -Include cudart64_*,cublas64_*,cublasLt64_* - -Destination build/bin/${{ matrix.build }} - - - name: Copy SDL2.dll - if: matrix.sdl2 == 'ON' - run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }} - - - name: Upload binaries - if: matrix.sdl2 == 'ON' - uses: actions/upload-artifact@v4 - with: - name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }} - path: build/bin/${{ matrix.build }} +# TODO: fix and re-enable +# windows-cublas: +# runs-on: windows-2019 +# +# strategy: +# matrix: +# build: [Release] +# arch: [x64] +# cublas: [ON] +# sdl2: [ON] +# cuda-toolkit: [12.2.0, 11.8.0] +# include: +# - arch: x64 +# s2arc: x64 +# - sdl2: ON +# s2ver: 2.28.5 +# +# steps: +# - name: Clone +# uses: actions/checkout@v4 +# +# - name: Add msbuild to PATH +# uses: microsoft/setup-msbuild@v2 +# +# - name: Install CUDA Toolkit +# id: cuda-toolkit +# uses: Jimver/cuda-toolkit@v0.2.15 +# with: +# cuda: '${{ matrix.cuda-toolkit }}' +# +# - name: Fetch SDL2 and set SDL2_DIR +# if: matrix.sdl2 == 'ON' +# run: | +# C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ matrix.s2ver }}/SDL2-devel-${{ matrix.s2ver }}-VC.zip +# 7z x sdl2.zip +# echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV +# +# - name: Configure +# run: > +# cmake -S . -B ./build -A ${{ matrix.arch }} +# -DCMAKE_BUILD_TYPE=${{ matrix.build }} +# -DGGML_CUDA=${{ matrix.cublas }} +# -DWHISPER_SDL2=${{ matrix.sdl2 }} +# +# - name: Build ${{ matrix.cuda-toolkit }} +# run: | +# cd ./build +# cmake --build . --config ${{ matrix.build }} +# +# - name: Copy CUDA DLLs +# run: > +# Copy-Item -PassThru +# -Path "${{ steps.cuda-toolkit.outputs.CUDA_PATH }}/bin/*.dll" +# -Include cudart64_*,cublas64_*,cublasLt64_* +# -Destination build/bin/${{ matrix.build }} +# +# - name: Copy SDL2.dll +# if: matrix.sdl2 == 'ON' +# run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }} +# +# - name: Upload binaries +# if: matrix.sdl2 == 'ON' +# uses: actions/upload-artifact@v4 +# with: +# name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }} +# path: build/bin/${{ matrix.build }} emscripten: runs-on: ubuntu-latest @@ -533,7 +519,7 @@ jobs: emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }} make - ios: + ios-xcode-build: runs-on: macos-latest strategy: @@ -541,7 +527,7 @@ jobs: build: [Release] steps: - - name: Clone + - name: Checkout code uses: actions/checkout@v4 - name: Configure @@ -549,40 +535,64 @@ jobs: cp models/for-tests-ggml-base.en.bin models/ggml-base.en.bin mkdir models/ggml-base.en-encoder.mlmodelc - - name: Build objc example - run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphonesimulator build - - - name: Build swiftui example - run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphonesimulator build - - android: - runs-on: ubuntu-latest - - steps: - - name: Clone - uses: actions/checkout@v4 - with: - path: whisper - - - name: Install Java - uses: actions/setup-java@v4 - with: - distribution: zulu - java-version: 21 - - - name: Setup Android SDK - uses: android-actions/setup-android@v3 - - name: Build + id: cmake_build run: | - cd whisper/examples/whisper.android - ./gradlew assembleRelease --no-daemon + sysctl -a + mkdir build + cd build + cmake -G Xcode .. \ + -DGGML_METAL_USE_BF16=ON \ + -DGGML_METAL_EMBED_LIBRARY=ON \ + -DWHISPER_BUILD_EXAMPLES=OFF \ + -DWHISPER_BUILD_TESTS=OFF \ + -DWHISPER_BUILD_SERVER=OFF \ + -DCMAKE_SYSTEM_NAME=iOS \ + -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \ + -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml + cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO + sudo cmake --install . --config Release + + - name: xcodebuild for swift package + id: xcodebuild + run: | + xcodebuild -scheme whisper-Package -destination 'generic/platform=iOS' + +#- name: Build objc example +# run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphoneos build - - name: Build with external ggml - run: | - export PATH_TO_GGML=$PWD/ggml - cd whisper/examples/whisper.android - ./gradlew assembleRelease --no-daemon + - name: Build swiftui example + run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build + +# TODO: update android build and re-enable when it works +# android: +# runs-on: ubuntu-latest +# +# steps: +# - name: Clone +# uses: actions/checkout@v4 +# with: +# path: whisper +# +# - name: Install Java +# uses: actions/setup-java@v4 +# with: +# distribution: zulu +# java-version: 21 +# +# - name: Setup Android SDK +# uses: android-actions/setup-android@v3 +# +# - name: Build +# run: | +# cd whisper/examples/whisper.android +# ./gradlew assembleRelease --no-daemon +# +# - name: Build with external ggml +# run: | +# export PATH_TO_GGML=$PWD/ggml +# cd whisper/examples/whisper.android +# ./gradlew assembleRelease --no-daemon # TODO: disable because of following fail: https://github.com/ggerganov/whisper.cpp/actions/runs/11019444420/job/30627193602 # android_java: @@ -664,5 +674,6 @@ jobs: - name: Test quantize run: | ./models/download-ggml-model.sh tiny.en - make quantize - ./quantize models/ggml-tiny.en.bin models/ggml-tiny.en-q4_0.bin q4_0 + cmake -B build + cmake --build build --config Release + ./build/bin/quantize models/ggml-tiny.en.bin models/ggml-tiny.en-q4_0.bin q4_0 diff --git a/Makefile b/Makefile index 8a6c0282b9a..2743eecd9ab 100644 --- a/Makefile +++ b/Makefile @@ -1,1075 +1,3 @@ -# Define the default target now so that it is always the first target -BUILD_TARGETS = \ - main \ - bench \ - quantize \ - server - -# Binaries only useful for tests -TEST_TARGETS = \ - tests/test-c.o - -# Deprecation aliases -ifdef WHISPER_CUBLAS -$(error WHISPER_CUBLAS is removed. Use GGML_CUDA instead.) -endif - -ifdef WHISPER_CUDA -GGML_CUDA := 1 -DEPRECATE_WARNING := 1 -endif - -ifdef WHISPER_KOMPUTE -GGML_KOMPUTE := 1 -DEPRECATE_WARNING := 1 -endif - -ifdef WHISPER_METAL -GGML_METAL := 1 -DEPRECATE_WARNING := 1 -endif - -ifdef WHISPER_OPENMP -GGML_OPENMP := 1 -DEPRECATE_WARNING := 1 -endif - -ifdef WHISPER_RPC -GGML_RPC := 1 -DEPRECATE_WARNING := 1 -endif - -ifdef WHISPER_SYCL -GGML_SYCL := 1 -DEPRECATE_WARNING := 1 -endif - -ifdef WHISPER_SYCL_F16 -GGML_SYCL_F16 := 1 -DEPRECATE_WARNING := 1 -endif - -ifdef WHISPER_OPENBLAS -GGML_OPENBLAS := 1 -DEPRECATE_WARNING := 1 -endif - -ifdef WHISPER_OPENBLAS64 -GGML_OPENBLAS64 := 1 -DEPRECATE_WARNING := 1 -endif - -ifdef WHISPER_BLIS -GGML_BLIS := 1 -DEPRECATE_WARNING := 1 -endif - -ifdef WHISPER_NO_WHISPERFILE -GGML_NO_WHISPERFILE := 1 -DEPRECATE_WARNING := 1 -endif - -ifdef WHISPER_NO_ACCELERATE -GGML_NO_ACCELERATE := 1 -DEPRECATE_WARNING := 1 -endif - -ifdef WHISPER_NO_OPENMP -GGML_NO_OPENMP := 1 -DEPRECATE_WARNING := 1 -endif - -ifdef WHISPER_NO_METAL -GGML_NO_METAL := 1 -DEPRECATE_WARNING := 1 -endif - -ifndef UNAME_S -UNAME_S := $(shell uname -s) -endif - -ifndef UNAME_P -UNAME_P := $(shell uname -p) -endif - -ifndef UNAME_M -UNAME_M := $(shell uname -m) -endif - -# In GNU make default CXX is g++ instead of c++. Let's fix that so that users -# of non-gcc compilers don't have to provide g++ alias or wrapper. -DEFCC := cc -DEFCXX := c++ -ifeq ($(origin CC),default) -CC := $(DEFCC) -endif -ifeq ($(origin CXX),default) -CXX := $(DEFCXX) -endif - -# Mac OS + Arm can report x86_64 -# ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789 -ifeq ($(UNAME_S),Darwin) - ifndef GGML_NO_METAL - GGML_METAL := 1 - endif - - GGML_NO_OPENMP := 1 - - ifneq ($(UNAME_P),arm) - SYSCTL_M := $(shell sysctl -n hw.optional.arm64 2>/dev/null) - ifeq ($(SYSCTL_M),1) - # UNAME_P := arm - # UNAME_M := arm64 - warn := $(warning Your arch is announced as x86_64, but it seems to actually be ARM64. Not fixing that can lead to bad performance. For more info see: https://github.com/ggerganov/whisper.cpp/issues/66\#issuecomment-1282546789) - endif - endif -endif - -ifdef GGML_METAL - GGML_METAL_EMBED_LIBRARY := 1 -endif - -ifdef GGML_RPC - BUILD_TARGETS += rpc-server -endif - -ifdef GGML_VULKAN - BUILD_TARGETS += vulkan-shaders-gen -endif - -ifeq ($(shell sdl2-config --cflags --libs 2>/dev/null),) -else - BUILD_TARGETS += \ - command \ - stream \ - lsp \ - talk-llama - # talk (TODO: disalbed) -endif - -default: $(BUILD_TARGETS) - -test: $(TEST_TARGETS) - @failures=0; \ - for test_target in $(TEST_TARGETS); do \ - echo "Running test $$test_target..."; \ - ./$$test_target; \ - if [ $$? -ne 0 ]; then \ - printf 'Test %s FAILED!\n\n' $$test_target; \ - failures=$$(( failures + 1 )); \ - else \ - printf 'Test %s passed.\n\n' $$test_target; \ - fi; \ - done; \ - failures=$$(( failures + $$? )); \ - if [ $$failures -gt 0 ]; then \ - printf '\n%s tests failed.\n' $$failures; \ - exit 1; \ - fi - @echo 'All tests passed.' - -all: $(BUILD_TARGETS) $(TEST_TARGETS) - -ifdef RISCV_CROSS_COMPILE - CC := riscv64-unknown-linux-gnu-gcc - CXX := riscv64-unknown-linux-gnu-g++ -endif - -# -# Compile flags -# - -# keep standard at C11 and C++11 -MK_CPPFLAGS = -Iggml/include -Iggml/src -Iinclude -Isrc -Iexamples -MK_CFLAGS = -std=c11 -fPIC -MK_CXXFLAGS = -std=c++11 -fPIC -MK_NVCCFLAGS = -std=c++11 - -ifndef WHISPER_NO_CCACHE -CCACHE := $(shell which ccache) -ifdef CCACHE -export CCACHE_SLOPPINESS = time_macros -$(info I ccache found, compilation results will be cached. Disable with WHISPER_NO_CCACHE.) -CC := $(CCACHE) $(CC) -CXX := $(CCACHE) $(CXX) -else -$(info I ccache not found. Consider installing it for faster compilation.) -endif # CCACHE -endif # WHISPER_NO_CCACHE - -# 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) -MK_CPPFLAGS += -D_XOPEN_SOURCE=600 - -# 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 -ifeq ($(UNAME_S),OpenBSD) - MK_CPPFLAGS += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700 -endif - -# Data types, macros and functions related to controlling CPU affinity and -# some memory allocation are available on Linux through GNU extensions in libc -ifeq ($(UNAME_S),Linux) - MK_CPPFLAGS += -D_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 -ifeq ($(UNAME_S),Darwin) - MK_CPPFLAGS += -D_DARWIN_C_SOURCE -endif -ifeq ($(UNAME_S),DragonFly) - MK_CPPFLAGS += -D__BSD_VISIBLE -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 -ifeq ($(UNAME_S),FreeBSD) - MK_CPPFLAGS += -D__BSD_VISIBLE -endif -ifeq ($(UNAME_S),NetBSD) - MK_CPPFLAGS += -D_NETBSD_SOURCE -endif -ifeq ($(UNAME_S),OpenBSD) - MK_CPPFLAGS += -D_BSD_SOURCE -endif - -ifdef GGML_SCHED_MAX_COPIES - MK_CPPFLAGS += -DGGML_SCHED_MAX_COPIES=$(GGML_SCHED_MAX_COPIES) -endif - -ifdef WHISPER_DEBUG - MK_CFLAGS += -O0 -g - MK_CXXFLAGS += -O0 -g - MK_LDFLAGS += -g - MK_NVCCFLAGS += -O0 -g - - ifeq ($(UNAME_S),Linux) - MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS - endif -else - MK_CPPFLAGS += -DNDEBUG - MK_CFLAGS += -O3 - MK_CXXFLAGS += -O3 - MK_NVCCFLAGS += -O3 -endif - -ifdef WHISPER_SANITIZE_THREAD - MK_CFLAGS += -fsanitize=thread -g - MK_CXXFLAGS += -fsanitize=thread -g - MK_LDFLAGS += -fsanitize=thread -g -endif - -ifdef WHISPER_SANITIZE_ADDRESS - MK_CFLAGS += -fsanitize=address -fno-omit-frame-pointer -g - MK_CXXFLAGS += -fsanitize=address -fno-omit-frame-pointer -g - MK_LDFLAGS += -fsanitize=address -fno-omit-frame-pointer -g -endif - -ifdef WHISPER_SANITIZE_UNDEFINED - MK_CFLAGS += -fsanitize=undefined -g - MK_CXXFLAGS += -fsanitize=undefined -g - MK_LDFLAGS += -fsanitize=undefined -g -endif - -ifdef WHISPER_SERVER_VERBOSE - MK_CPPFLAGS += -DSERVER_VERBOSE=$(WHISPER_SERVER_VERBOSE) -endif - -ifdef WHISPER_SERVER_SSL - MK_CPPFLAGS += -DCPPHTTPLIB_OPENSSL_SUPPORT - MK_LDFLAGS += -lssl -lcrypto -endif - -ifdef WHISPER_DISABLE_LOGS - MK_CPPFLAGS += -DLOG_DISABLE_LOGS -endif # WHISPER_DISABLE_LOGS - -# warnings -WARN_FLAGS = \ - -Wall \ - -Wextra \ - -Wpedantic \ - -Wcast-qual \ - -Wno-unused-function - -MK_CFLAGS += \ - $(WARN_FLAGS) \ - -Wshadow \ - -Wstrict-prototypes \ - -Wpointer-arith \ - -Wmissing-prototypes \ - -Werror=implicit-int \ - -Werror=implicit-function-declaration - -MK_CXXFLAGS += \ - $(WARN_FLAGS) \ - -Wmissing-declarations \ - -Wmissing-noreturn - -ifeq ($(WHISPER_FATAL_WARNINGS),1) - MK_CFLAGS += -Werror - MK_CXXFLAGS += -Werror -endif - -# this version of Apple ld64 is buggy -ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))' - MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER -endif - -# OS specific -# TODO: support Windows -ifneq '' '$(filter $(UNAME_S),Linux Darwin FreeBSD NetBSD OpenBSD Haiku)' - MK_CFLAGS += -pthread - MK_CXXFLAGS += -pthread -endif - -# detect Windows -ifneq ($(findstring _NT,$(UNAME_S)),) - _WIN32 := 1 -endif - -# library name prefix -ifneq ($(_WIN32),1) - LIB_PRE := lib -endif - -# Dynamic Shared Object extension -ifneq ($(_WIN32),1) - DSO_EXT := .so -else - DSO_EXT := .dll -endif - -# Windows Sockets 2 (Winsock) for network-capable apps -ifeq ($(_WIN32),1) - LWINSOCK2 := -lws2_32 -endif - -ifdef WHISPER_GPROF - MK_CFLAGS += -pg - MK_CXXFLAGS += -pg -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 - -ifndef RISCV - -ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64)) - # Use all CPU extensions that are available: - MK_CFLAGS += -march=native -mtune=native - HOST_CXXFLAGS += -march=native -mtune=native - - # Usage AVX-only - #MK_CFLAGS += -mfma -mf16c -mavx - #MK_CXXFLAGS += -mfma -mf16c -mavx - - # Usage SSSE3-only (Not is SSE3!) - #MK_CFLAGS += -mssse3 - #MK_CXXFLAGS += -mssse3 -endif - -ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))' - # The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves. - # https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412 - # https://github.com/ggerganov/llama.cpp/issues/2922 - MK_CFLAGS += -Xassembler -muse-unaligned-vector-move - MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move - - # Target Windows 8 for PrefetchVirtualMemory - MK_CPPFLAGS += -D_WIN32_WINNT=0x602 -endif - -ifneq ($(filter aarch64%,$(UNAME_M)),) - # Apple M1, M2, etc. - # Raspberry Pi 3, 4, Zero 2 (64-bit) - # Nvidia Jetson - MK_CFLAGS += -mcpu=native - MK_CXXFLAGS += -mcpu=native - JETSON_RELEASE_INFO = $(shell jetson_release) - ifdef JETSON_RELEASE_INFO - ifneq ($(filter TX2%,$(JETSON_RELEASE_INFO)),) - JETSON_EOL_MODULE_DETECT = 1 - CC = aarch64-unknown-linux-gnu-gcc - cxx = aarch64-unknown-linux-gnu-g++ - endif - endif -endif - -ifneq ($(filter armv6%,$(UNAME_M)),) - # Raspberry Pi 1, Zero - MK_CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access - MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -endif - -ifneq ($(filter armv7%,$(UNAME_M)),) - # Raspberry Pi 2 - MK_CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations - MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations -endif - -ifneq ($(filter armv8%,$(UNAME_M)),) - # Raspberry Pi 3, 4, Zero 2 (32-bit) - MK_CFLAGS += -mfp16-format=ieee -mno-unaligned-access - MK_CXXFLAGS += -mfp16-format=ieee -mno-unaligned-access -endif - -ifneq ($(filter ppc64%,$(UNAME_M)),) - POWER9_M := $(shell grep "POWER9" /proc/cpuinfo) - ifneq (,$(findstring POWER9,$(POWER9_M))) - MK_CFLAGS += -mcpu=power9 - MK_CXXFLAGS += -mcpu=power9 - endif -endif - -ifneq ($(filter ppc64le%,$(UNAME_M)),) - MK_CFLAGS += -mcpu=powerpc64le - MK_CXXFLAGS += -mcpu=powerpc64le - CUDA_POWER_ARCH = 1 -endif - -ifneq ($(filter loongarch64%,$(UNAME_M)),) - MK_CFLAGS += -mlasx - MK_CXXFLAGS += -mlasx -endif - -else - MK_CFLAGS += -march=rv64gcv -mabi=lp64d - MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d -endif # RISCV - -ifndef GGML_NO_ACCELERATE - # Mac OS - include Accelerate framework. - # `-framework Accelerate` works both with Apple Silicon and Mac Intel - ifeq ($(UNAME_S),Darwin) - MK_CPPFLAGS += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS -DGGML_BLAS_USE_ACCELERATE - MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK - MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64 - MK_LDFLAGS += -framework Accelerate - OBJ_GGML += ggml/src/ggml-blas/ggml-blas.o - endif -endif # GGML_NO_ACCELERATE - -ifndef GGML_NO_OPENMP - MK_CPPFLAGS += -DGGML_USE_OPENMP - MK_CFLAGS += -fopenmp - MK_CXXFLAGS += -fopenmp -endif # GGML_NO_OPENMP - -ifdef WHISPER_COREML - MK_CXXFLAGS += -DWHISPER_USE_COREML - LDFLAGS += -framework Foundation -framework CoreML - -ifdef WHISPER_COREML_ALLOW_FALLBACK - MK_CXXFLAGS += -DWHISPER_COREML_ALLOW_FALLBACK -endif -endif # WHISPER_COREML - -ifdef GGML_OPENBLAS - MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas) - MK_CFLAGS += $(shell pkg-config --cflags-only-other openblas) - MK_LDFLAGS += $(shell pkg-config --libs openblas) - OBJ_GGML += ggml/src/ggml-blas/ggml-blas.o -endif # GGML_OPENBLAS - -ifdef GGML_OPENBLAS64 - MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas64) - MK_CFLAGS += $(shell pkg-config --cflags-only-other openblas64) - MK_LDFLAGS += $(shell pkg-config --libs openblas64) - OBJ_GGML += ggml/src/ggml-blas/ggml-blas.o -endif # GGML_OPENBLAS64 - -ifdef GGML_BLIS - MK_CPPFLAGS += -DGGML_USE_BLAS -I/usr/local/include/blis -I/usr/include/blis - MK_LDFLAGS += -lblis -L/usr/local/lib - OBJ_GGML += ggml/src/ggml-blas/ggml-blas.o -endif # GGML_BLIS - -ifdef GGML_RPC - MK_CPPFLAGS += -DGGML_USE_RPC - OBJ_GGML += ggml/src/ggml-rpc/ggml-rpc.o -endif # GGML_RPC - -OBJ_CUDA_TMPL = $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-wmma*.cu)) -OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/mmq*.cu)) - -ifdef GGML_CUDA_FA_ALL_QUANTS - OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*.cu)) -else - OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu)) - OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu)) - OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*f16-f16.cu)) -endif # GGML_CUDA_FA_ALL_QUANTS - -ifdef GGML_CUDA - ifneq ('', '$(wildcard /opt/cuda)') - CUDA_PATH ?= /opt/cuda - else - CUDA_PATH ?= /usr/local/cuda - endif - - #MK_CPPFLAGS += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS - #MK_LDFLAGS += -lcuda -lcublas -lculibos -lcudart -lcufft -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib - MK_CPPFLAGS += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include - MK_LDFLAGS += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib - MK_NVCCFLAGS += -use_fast_math - - OBJ_GGML += ggml/src/ggml-cuda/ggml-cuda.o - OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu)) - OBJ_GGML += $(OBJ_CUDA_TMPL) -ifdef WHISPER_FATAL_WARNINGS - MK_NVCCFLAGS += -Werror all-warnings -endif # WHISPER_FATAL_WARNINGS - -ifndef JETSON_EOL_MODULE_DETECT - MK_NVCCFLAGS += --forward-unknown-to-host-compiler -endif # JETSON_EOL_MODULE_DETECT - -ifdef WHISPER_DEBUG - MK_NVCCFLAGS += -lineinfo -endif # WHISPER_DEBUG - -ifdef GGML_CUDA_DEBUG - MK_NVCCFLAGS += --device-debug -endif # GGML_CUDA_DEBUG - -ifdef GGML_CUDA_NVCC - NVCC = $(CCACHE) $(GGML_CUDA_NVCC) -else - NVCC = $(CCACHE) nvcc -endif #GGML_CUDA_NVCC - -ifdef CUDA_DOCKER_ARCH - MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH) -else ifndef CUDA_POWER_ARCH - MK_NVCCFLAGS += -arch=native -endif # CUDA_DOCKER_ARCH - -ifdef GGML_CUDA_FORCE_DMMV - MK_NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV -endif # GGML_CUDA_FORCE_DMMV - -ifdef GGML_CUDA_FORCE_MMQ - MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ -endif # GGML_CUDA_FORCE_MMQ - -ifdef GGML_CUDA_DMMV_X - MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X) -else - MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=32 -endif # GGML_CUDA_DMMV_X - -ifdef GGML_CUDA_MMV_Y - MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y) -else ifdef GGML_CUDA_DMMV_Y - MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_DMMV_Y) # for backwards compatibility -else - MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=1 -endif # GGML_CUDA_MMV_Y - -ifdef GGML_CUDA_F16 - MK_NVCCFLAGS += -DGGML_CUDA_F16 -endif # GGML_CUDA_F16 - -ifdef GGML_CUDA_DMMV_F16 - MK_NVCCFLAGS += -DGGML_CUDA_F16 -endif # GGML_CUDA_DMMV_F16 - -ifdef GGML_CUDA_KQUANTS_ITER - MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER) -else - MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2 -endif - -ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE - MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE) -else - MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -endif # GGML_CUDA_PEER_MAX_BATCH_SIZE - -ifdef GGML_CUDA_NO_PEER_COPY - MK_NVCCFLAGS += -DGGML_CUDA_NO_PEER_COPY -endif # GGML_CUDA_NO_PEER_COPY - -ifdef GGML_CUDA_CCBIN - MK_NVCCFLAGS += -ccbin $(GGML_CUDA_CCBIN) -endif # GGML_CUDA_CCBIN - -ifdef GGML_CUDA_FA_ALL_QUANTS - MK_NVCCFLAGS += -DGGML_CUDA_FA_ALL_QUANTS -endif # GGML_CUDA_FA_ALL_QUANTS - -ifdef JETSON_EOL_MODULE_DETECT -define NVCC_COMPILE - $(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@ -endef # NVCC_COMPILE -else -define NVCC_COMPILE - $(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@ -endef # NVCC_COMPILE -endif # JETSON_EOL_MODULE_DETECT - -ggml/src/ggml-cuda/%.o: \ - ggml/src/ggml-cuda/%.cu \ - ggml/include/ggml.h \ - ggml/src/ggml-common.h \ - ggml/src/ggml-cuda/common.cuh - $(NVCC_COMPILE) - -ggml/src/ggml-cuda/ggml-cuda.o: \ - ggml/src/ggml-cuda/ggml-cuda.cu \ - ggml/include/ggml-cuda.h \ - ggml/include/ggml.h \ - ggml/include/ggml-backend.h \ - ggml/src/ggml-backend-impl.h \ - ggml/src/ggml-common.h \ - $(wildcard ggml/src/ggml-cuda/*.cuh) - $(NVCC_COMPILE) -endif # GGML_CUDA - -ifdef GGML_VULKAN - MK_CPPFLAGS += -DGGML_USE_VULKAN - MK_LDFLAGS += $(shell pkg-config --libs vulkan) - OBJ_GGML += ggml/src/ggml-vulkan.o ggml/src/ggml-vulkan-shaders.o - -ifdef GGML_VULKAN_CHECK_RESULTS - MK_CPPFLAGS += -DGGML_VULKAN_CHECK_RESULTS -endif - -ifdef GGML_VULKAN_DEBUG - MK_CPPFLAGS += -DGGML_VULKAN_DEBUG -endif - -ifdef GGML_VULKAN_MEMORY_DEBUG - MK_CPPFLAGS += -DGGML_VULKAN_MEMORY_DEBUG -endif - -ifdef GGML_VULKAN_PERF - MK_CPPFLAGS += -DGGML_VULKAN_PERF -endif - -ifdef GGML_VULKAN_VALIDATE - MK_CPPFLAGS += -DGGML_VULKAN_VALIDATE -endif - -ifdef GGML_VULKAN_RUN_TESTS - MK_CPPFLAGS += -DGGML_VULKAN_RUN_TESTS -endif - -GLSLC_CMD = glslc -_ggml_vk_genshaders_cmd = $(shell pwd)/vulkan-shaders-gen -_ggml_vk_header = ggml/src/ggml-vulkan-shaders.hpp -_ggml_vk_source = ggml/src/ggml-vulkan-shaders.cpp -_ggml_vk_input_dir = ggml/src/vulkan-shaders -_ggml_vk_shader_deps = $(echo $(_ggml_vk_input_dir)/*.comp) - -ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source) - $(CXX) $(CXXFLAGS) $(shell pkg-config --cflags vulkan) -c $< -o $@ - -$(_ggml_vk_header): $(_ggml_vk_source) - -$(_ggml_vk_source): $(_ggml_vk_shader_deps) vulkan-shaders-gen - $(_ggml_vk_genshaders_cmd) \ - --glslc $(GLSLC_CMD) \ - --input-dir $(_ggml_vk_input_dir) \ - --target-hpp $(_ggml_vk_header) \ - --target-cpp $(_ggml_vk_source) - -vulkan-shaders-gen: ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp - $(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp - -endif # GGML_VULKAN - -ifdef GGML_HIPBLAS - ifeq ($(wildcard /opt/rocm),) - ROCM_PATH ?= /usr - AMDGPU_TARGETS ?= $(shell $(shell which amdgpu-arch)) - else - ROCM_PATH ?= /opt/rocm - AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch) - endif - - GGML_CUDA_DMMV_X ?= 32 - GGML_CUDA_MMV_Y ?= 1 - GGML_CUDA_KQUANTS_ITER ?= 2 - - MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA - -ifdef GGML_HIP_UMA - MK_CPPFLAGS += -DGGML_HIP_UMA -endif # GGML_HIP_UMA - - MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib - MK_LDFLAGS += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64 - MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas - - HIPCC ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc - - HIPFLAGS += $(addprefix --offload-arch=,$(AMDGPU_TARGETS)) - HIPFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X) - HIPFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y) - HIPFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER) - -ifdef GGML_CUDA_FORCE_DMMV - HIPFLAGS += -DGGML_CUDA_FORCE_DMMV -endif # GGML_CUDA_FORCE_DMMV - -ifdef GGML_CUDA_NO_PEER_COPY - HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY -endif # GGML_CUDA_NO_PEER_COPY - - OBJ_GGML += ggml/src/ggml-cuda.o - OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu)) - OBJ_GGML += $(OBJ_CUDA_TMPL) - -ggml/src/ggml-cuda.o: \ - ggml/src/ggml-cuda.cu \ - ggml/include/ggml.h \ - ggml/include/ggml-backend.h \ - ggml/include/ggml-cuda.h \ - ggml/src/ggml-backend-impl.h \ - ggml/src/ggml-common.h \ - $(wildcard ggml/src/ggml-cuda/*.cuh) - $(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $< - -ggml/src/ggml-cuda/%.o: \ - ggml/src/ggml-cuda/%.cu \ - ggml/include/ggml.h \ - ggml/src/ggml-common.h \ - ggml/src/ggml-cuda/common.cuh - $(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $< -endif # GGML_HIPBLAS - -ifdef GGML_METAL - MK_CPPFLAGS += -DGGML_USE_METAL - MK_LDFLAGS += -framework Foundation -framework Metal -framework MetalKit - OBJ_GGML += ggml/src/ggml-metal/ggml-metal.o -ifdef GGML_METAL_NDEBUG - MK_CPPFLAGS += -DGGML_METAL_NDEBUG -endif - -ifdef GGML_METAL_EMBED_LIBRARY - MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY - OBJ_GGML += ggml/src/ggml-metal/ggml-metal-embed.o -endif -endif # GGML_METAL - -ifdef GGML_METAL -ggml/src/ggml-metal/ggml-metal.o: \ - ggml/src/ggml-metal/ggml-metal.m \ - ggml/src/ggml-metal/ggml-metal-impl.h \ - ggml/include/ggml-metal.h \ - ggml/include/ggml.h - $(CC) $(CFLAGS) -c $< -o $@ - -ifdef GGML_METAL_EMBED_LIBRARY -ggml/src/ggml-metal/ggml-metal-embed.o: \ - ggml/src/ggml-metal/ggml-metal.metal \ - ggml/src/ggml-metal/ggml-metal-impl.h \ - ggml/src/ggml-common.h - @echo "Embedding Metal library" - @sed -e '/__embed_ggml-common.h__/r ggml/src/ggml-common.h' -e '/__embed_ggml-common.h__/d' < ggml/src/ggml-metal/ggml-metal.metal > ggml/src/ggml-metal/ggml-metal-embed.metal.tmp - @sed -e '/#include "ggml-metal-impl.h"/r ggml/src/ggml-metal/ggml-metal-impl.h' -e '/#include "ggml-metal-impl.h"/d' < ggml/src/ggml-metal/ggml-metal-embed.metal.tmp > ggml/src/ggml-metal/ggml-metal-embed.metal - $(eval TEMP_ASSEMBLY=$(shell mktemp -d)) - @echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)/ggml-metal-embed.s - @echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s - @echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s - @echo ".incbin \"ggml/src/ggml-metal/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s - @echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s - @echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s - $(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@ - @rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s - @rmdir ${TEMP_ASSEMBLY} -endif -endif # GGML_METAL - -ifdef WHISPER_COREML -src/coreml/whisper-encoder.o: src/coreml/whisper-encoder.mm src/coreml/whisper-encoder.h - $(CXX) -O3 -I . -fobjc-arc -c src/coreml/whisper-encoder.mm -o src/coreml/whisper-encoder.o - -src/coreml/whisper-encoder-impl.o: src/coreml/whisper-encoder-impl.m src/coreml/whisper-encoder-impl.h - $(CXX) -O3 -I . -fobjc-arc -c src/coreml/whisper-encoder-impl.m -o src/coreml/whisper-encoder-impl.o - -OBJ_WHISPER += src/coreml/whisper-encoder.o src/coreml/whisper-encoder-impl.o -endif - -OBJ_GGML += \ - ggml/src/ggml.o \ - ggml/src/ggml-aarch64.o \ - ggml/src/ggml-alloc.o \ - ggml/src/ggml-backend.o \ - ggml/src/ggml-backend-reg.o \ - ggml/src/ggml-opt.o \ - ggml/src/ggml-quants.o \ - ggml/src/ggml-threading.o \ - ggml/src/ggml-cpu/ggml-cpu.o \ - ggml/src/ggml-cpu/ggml-cpu-cpp.o \ - ggml/src/ggml-cpu/ggml-cpu-aarch64.o \ - ggml/src/ggml-cpu/ggml-cpu-quants.o - -OBJ_WHISPER += \ - src/whisper.o - -OBJ_COMMON += \ - examples/common.o \ - examples/common-ggml.o \ - examples/grammar-parser.o - -OBJ_SDL += \ - examples/common-sdl.o - -OBJ_ALL = $(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL) - -LIB_GGML = $(LIB_PRE)ggml$(DSO_EXT) -LIB_GGML_S = $(LIB_PRE)ggml.a - -LIB_WHISPER = $(LIB_PRE)whisper$(DSO_EXT) -LIB_WHISPER_S = $(LIB_PRE)whisper.a - -LIB_COMMON = $(LIB_PRE)common$(DSO_EXT) -LIB_COMMON_S = $(LIB_PRE)common.a - -LIB_COMMON_SDL = $(LIB_PRE)common-sdl$(DSO_EXT) -LIB_COMMON_SDL_S = $(LIB_PRE)common-sdl.a - -LIB_ALL = $(LIB_GGML) $(LIB_WHISPER) $(LIB_COMMON) $(LIB_COMMON_SDL) -LIB_ALL_S = $(LIB_GGML_S) $(LIB_WHISPER_S) $(LIB_COMMON_S) $(LIB_COMMON_SDL_S) - -GF_CC := $(CC) -include scripts/get-flags.mk - -# combine build flags with cmdline overrides -override CPPFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) -override CFLAGS := $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS) -BASE_CXXFLAGS := $(MK_CXXFLAGS) $(CXXFLAGS) -override CXXFLAGS := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS) $(CPPFLAGS) -override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS) -override LDFLAGS := $(MK_LDFLAGS) $(LDFLAGS) - -# identify CUDA host compiler -ifdef GGML_CUDA -GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler -include scripts/get-flags.mk -CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic -endif - -ifdef WHISPER_CURL -override CXXFLAGS := $(CXXFLAGS) -DWHISPER_USE_CURL -override LDFLAGS := $(LDFLAGS) -lcurl -endif - -# -# Print build information -# - -$(info I whisper.cpp build info: ) -$(info I UNAME_S: $(UNAME_S)) -$(info I UNAME_P: $(UNAME_P)) -$(info I UNAME_M: $(UNAME_M)) -$(info I CFLAGS: $(CFLAGS)) -$(info I CXXFLAGS: $(CXXFLAGS)) -$(info I NVCCFLAGS: $(NVCCFLAGS)) -$(info I LDFLAGS: $(LDFLAGS)) -$(info I CC: $(shell $(CC) --version | head -n 1)) -$(info I CXX: $(shell $(CXX) --version | head -n 1)) -ifdef GGML_CUDA -$(info I NVCC: $(shell $(NVCC) --version | tail -n 1)) -CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])') -ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1) - -ifndef CUDA_DOCKER_ARCH -ifndef CUDA_POWER_ARCH -$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via environment variable CUDA_DOCKER_ARCH, e.g. by running "export CUDA_DOCKER_ARCH=compute_XX" on Unix-like systems, where XX is the minimum compute capability that the code needs to run on. A list with compute capabilities can be found here: https://developer.nvidia.com/cuda-gpus ) -endif # CUDA_POWER_ARCH -endif # CUDA_DOCKER_ARCH - -endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1) -endif # GGML_CUDA -$(info ) - -ifdef DEPRECATE_WARNING -$(info !!! DEPRECATION WARNING !!!) -$(info The following WHISPER_ options are deprecated and will be removed in the future. Use the GGML_ prefix instead) -$(info - WHISPER_CUDA) -$(info - WHISPER_METAL) -$(info - WHISPER_OPENMP) -$(info - WHISPER_RPC) -$(info - WHISPER_SYCL) -$(info - WHISPER_SYCL_F16) -$(info - WHISPER_OPENBLAS) -$(info - WHISPER_OPENBLAS64) -$(info - WHISPER_BLIS) -$(info - WHISPER_NO_LLAMAFILE) -$(info - WHISPER_NO_ACCELERATE) -$(info - WHISPER_NO_OPENMP) -$(info - WHISPER_NO_METAL) -$(info ) -endif - -# -# Build libraries -# - -LIB_GGML = libggml.so -LIB_GGML_S = libggml.a - -LIB_LLAMA = libllama.so -LIB_LLAMA_S = libllama.a - -LIB_COMMON = libcommon.so -LIB_COMMON_S = libcommon.a - -LIB_COMMON_SDL = libcommon-sdl.so -LIB_COMMON_SDL_S = libcommon-sdl.a - -# Targets -BUILD_TARGETS += $(LIB_GGML) $(LIB_GGML_S) $(LIB_LLAMA) $(LIB_LLAMA_S) $(LIB_COMMON) $(LIB_COMMON_S) - -# Dependency files -DEP_FILES = $(OBJ_GGML:.o=.d) $(OBJ_LLAMA:.o=.d) $(OBJ_COMMON:.o=.d) - -# Default target -all: $(BUILD_TARGETS) - -# Note: need this exception because `ggml-cpu.c` and `ggml-cpu.cpp` both produce the same obj/dep files -# g++ -M -I ./ggml/include/ -I ./ggml/src ggml/src/ggml-cpu/ggml-cpu.cpp | grep ggml -ggml/src/ggml-cpu/ggml-cpu-cpp.o: \ - ggml/src/ggml-cpu/ggml-cpu.cpp \ - ggml/include/ggml-backend.h \ - ggml/include/ggml.h \ - ggml/include/ggml-alloc.h \ - ggml/src/ggml-backend-impl.h \ - ggml/include/ggml-cpu.h \ - ggml/src/ggml-impl.h - $(CXX) $(CXXFLAGS) -c $< -o $@ - -# Rules for building object files -ggml/%.o: ggml/%.c - $(CC) $(CFLAGS) -MMD -c $< -o $@ - -ggml/%.o: ggml/%.cpp - $(CXX) $(CXXFLAGS) -MMD -c $< -o $@ - -src/%.o: src/%.cpp - $(CXX) $(CXXFLAGS) -MMD -c $< -o $@ - -examples/%.o: examples/%.cpp - $(CXX) $(CXXFLAGS) -MMD -c $< -o $@ - -# Rules for building libraries -$(LIB_GGML): $(OBJ_GGML) - $(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) - -$(LIB_GGML_S): $(OBJ_GGML) - ar rcs $(LIB_GGML_S) $^ - -$(LIB_LLAMA): $(OBJ_LLAMA) $(LIB_GGML) - $(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) - -$(LIB_LLAMA_S): $(OBJ_LLAMA) - ar rcs $(LIB_LLAMA_S) $^ - -# common-sdl - -CFLAGS_SDL=$(shell sdl2-config --cflags) -LDFLAGS_SDL=$(shell sdl2-config --libs) - -examples/common-sdl.o: \ - examples/common-sdl.cpp \ - examples/common-sdl.h - $(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $@ - -$(LIB_COMMON): $(OBJ_COMMON) $(LIB_LLAMA) $(LIB_GGML) - $(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) - -$(LIB_COMMON_S): $(OBJ_COMMON) - ar rcs $(LIB_COMMON_S) $^ - -# Include dependency files --include $(DEP_FILES) - -# Clean rule -clean: - rm -vrf $(BUILD_TARGETS) $(TEST_TARGETS) - rm -rvf *.a *.dll *.so *.dot - find ggml src tests examples -type f -name "*.o" -delete - find ggml src tests examples -type f -name "*.d" -delete - -# -# Examples -# - -# $< is the first prerequisite, i.e. the source file. -# Explicitly compile this to an object file so that it can be cached with ccache. -# The source file is then filtered out from $^ (the list of all prerequisites) and the object file is added instead. - -# Helper function that replaces .c, .cpp, and .cu file endings with .o: -GET_OBJ_FILE = $(patsubst %.c,%.o,$(patsubst %.cpp,%.o,$(patsubst %.cu,%.o,$(1)))) - -main: examples/main/main.cpp \ - $(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) - $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) - -bench: examples/bench/bench.cpp \ - $(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) - $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) - -quantize: examples/quantize/quantize.cpp \ - $(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) - $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) - -server: examples/server/server.cpp \ - $(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) - $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2) - -command: examples/command/command.cpp \ - $(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL) - $(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL) - -stream: examples/stream/stream.cpp \ - $(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL) - $(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL) - -lsp: examples/lsp/lsp.cpp \ - $(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL) - $(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL) - -# TODO: disabled until update -# https://github.com/ggerganov/whisper.cpp/issues/1818 -#talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp \ -# $(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL) -# $(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<) -# $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL) - -talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp examples/talk-llama/llama-vocab.cpp examples/talk-llama/llama-grammar.cpp examples/talk-llama/llama-sampling.cpp examples/talk-llama/unicode.cpp examples/talk-llama/unicode-data.cpp \ - $(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL) - $(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL) - -# -# Tests -# - -tests: $(TEST_TARGETS) - -tests/test-c.o: tests/test-c.c include/whisper.h - $(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@ - # # Audio samples # @@ -1117,8 +45,10 @@ samples: .PHONY: large-v3 .PHONY: large-v3-turbo -tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large-v3 large-v3-turbo: main +tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large-v3 large-v3-turbo: bash ./models/download-ggml-model.sh $@ + cmake -B build + cmake --build build --config Release @echo "" @echo "===============================================" @echo "Running $@ on all samples in ./samples ..." @@ -1129,6 +59,6 @@ tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 larg echo "[+] Running $@ on $$f ... (run 'ffplay $$f' to listen)" ; \ echo "----------------------------------------------" ; \ echo "" ; \ - ./main -m models/ggml-$@.bin -f $$f ; \ + ./build/bin/main -m models/ggml-$@.bin -f $$f ; \ echo "" ; \ done diff --git a/Package.swift b/Package.swift index 367ce03b5a3..303ed0d78fb 100644 --- a/Package.swift +++ b/Package.swift @@ -14,55 +14,6 @@ let package = Package( .library(name: "whisper", targets: ["whisper"]), ], targets: [ - .target( - name: "whisper", - path: ".", - exclude: [ - "build", - "bindings", - "cmake", - "examples", - "scripts", - "models", - "samples", - "tests", - "CMakeLists.txt", - "Makefile", - "ggml/src/ggml-metal/ggml-metal-embed.metal" - ], - sources: [ - "ggml/src/ggml.c", - "src/whisper.cpp", - "ggml/src/ggml-aarch64.c", - "ggml/src/ggml-alloc.c", - "ggml/src/ggml-backend.cpp", - "ggml/src/ggml-backend-reg.cpp", - "ggml/src/ggml-cpu/ggml-cpu.c", - "ggml/src/ggml-cpu/ggml-cpu.cpp", - "ggml/src/ggml-cpu/ggml-cpu-aarch64.c", - "ggml/src/ggml-cpu/ggml-cpu-quants.c", - "ggml/src/ggml-quants.c", - "ggml/src/ggml-threading.cpp", - "ggml/src/ggml-metal/ggml-metal.m" - ], - resources: [.process("ggml/src/ggml-metal/ggml-metal.metal")], - publicHeadersPath: "spm-headers", - cSettings: [ - .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]), - .unsafeFlags(["-fno-objc-arc"]), - .headerSearchPath("ggml/src"), - .define("GGML_USE_ACCELERATE"), - .define("GGML_USE_METAL") - // NOTE: NEW_LAPACK will required iOS version 16.4+ - // We should consider add this in the future when we drop support for iOS 14 - // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc) - // .define("ACCELERATE_NEW_LAPACK"), - // .define("ACCELERATE_LAPACK_ILP64") - ], - linkerSettings: [ - .linkedFramework("Accelerate") - ] - ) - ], - cxxLanguageStandard: .cxx11 + .systemLibrary(name: "whisper", pkgConfig: "whisper"), + ] ) diff --git a/README.md b/README.md index c02f66b8163..a70e58df5dc 100644 --- a/README.md +++ b/README.md @@ -89,10 +89,11 @@ Now build the [main](examples/main) example and transcribe an audio file like th ```bash # build the main example -make -j +cmake -B build +cmake --build build --config Release # transcribe an audio file -./main -f samples/jfk.wav +./build/bin/main -f samples/jfk.wav ``` --- @@ -265,11 +266,12 @@ Here are the steps for creating and using a quantized model: ```bash # quantize a model with Q5_0 method -make -j quantize -./quantize models/ggml-base.en.bin models/ggml-base.en-q5_0.bin q5_0 +cmake -B build +cmake --build build --config Release +./build/bin/quantize models/ggml-base.en.bin models/ggml-base.en-q5_0.bin q5_0 # run the examples as usual, specifying the quantized model file -./main -m models/ggml-base.en-q5_0.bin ./samples/gb0.wav +./build/bin/main -m models/ggml-base.en-q5_0.bin ./samples/gb0.wav ``` ## Core ML support @@ -303,10 +305,6 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in - Build `whisper.cpp` with Core ML support: ```bash - # using Makefile - make clean - WHISPER_COREML=1 make -j - # using CMake cmake -B build -DWHISPER_COREML=1 cmake --build build -j --config Release @@ -426,8 +424,8 @@ First, make sure you have installed `cuda`: https://developer.nvidia.com/cuda-do Now build `whisper.cpp` with CUDA support: ``` -make clean -GGML_CUDA=1 make -j +cmake -B build -DGGML_CUDA=1 +cmake --build build -j --config Release ``` ## Vulkan GPU support @@ -436,8 +434,8 @@ First, make sure your graphics card driver provides support for Vulkan API. Now build `whisper.cpp` with Vulkan support: ``` -make clean -make GGML_VULKAN=1 -j +cmake -B build -DGGML_VULKAN=1 +cmake --build build -j --config Release ``` ## BLAS CPU support via OpenBLAS @@ -448,28 +446,13 @@ First, make sure you have installed `openblas`: https://www.openblas.net/ Now build `whisper.cpp` with OpenBLAS support: ``` -make clean -GGML_OPENBLAS=1 make -j -``` - -## BLAS CPU support via Intel MKL - -Encoder processing can be accelerated on the CPU via the BLAS compatible interface of Intel's Math Kernel Library. -First, make sure you have installed Intel's MKL runtime and development packages: https://www.intel.com/content/www/us/en/developer/tools/oneapi/onemkl-download.html - -Now build `whisper.cpp` with Intel MKL BLAS support: - -``` -source /opt/intel/oneapi/setvars.sh -mkdir build -cd build -cmake -DWHISPER_MKL=ON .. -WHISPER_MKL=1 make -j +cmake -B build -DGGML_BLAS=1 +cmake --build build -j --config Release ``` ## Ascend NPU support -Ascend NPU provides inference acceleration via [`CANN`](https://www.hiascend.com/en/software/cann) and AI cores. +Ascend NPU provides inference acceleration via [`CANN`](https://www.hiascend.com/en/software/cann) and AI cores. First, check if your Ascend NPU device is supported: @@ -483,10 +466,8 @@ Then, make sure you have installed [`CANN toolkit`](https://www.hiascend.com/en/ Now build `whisper.cpp` with CANN support: ``` -mkdir build -cd build -cmake .. -D GGML_CANN=on -make -j +cmake -B build -DGGML_CANN=1 +cmake --build build -j --config Release ``` Run the inference examples as usual, for example: @@ -636,8 +617,9 @@ The [stream](examples/stream) tool samples the audio every half a second and run More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10). ```bash -make stream -j -./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000 +cmake -B build +cmake --build build --config Release +./build/bin/stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000 ``` https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a80f-28ba83be7d09.mp4 diff --git a/Sources/whisper/module.modulemap b/Sources/whisper/module.modulemap new file mode 100644 index 00000000000..a266c0f2476 --- /dev/null +++ b/Sources/whisper/module.modulemap @@ -0,0 +1,5 @@ +module whisper [system] { + header "whisper.h" + link "whisper" + export * +} diff --git a/Sources/whisper/whisper.h b/Sources/whisper/whisper.h new file mode 100644 index 00000000000..e503c8a38ad --- /dev/null +++ b/Sources/whisper/whisper.h @@ -0,0 +1,4 @@ +#pragma once + +#include + diff --git a/cmake/whisper.pc.in b/cmake/whisper.pc.in index 50ab4f8285a..00ec7912014 100644 --- a/cmake/whisper.pc.in +++ b/cmake/whisper.pc.in @@ -1,10 +1,10 @@ prefix=@CMAKE_INSTALL_PREFIX@ exec_prefix=${prefix} -libdir=@CMAKE_INSTALL_FULL_LIBDIR@ +libdir=${exec_prefix}/lib includedir=${prefix}/include Name: whisper Description: Port of OpenAI's Whisper model in C/C++ Version: @PROJECT_VERSION@ -Libs: -L${libdir} -lwhisper +Libs: -L${libdir} -lggml -lggml-base -lwhisper Cflags: -I${includedir} diff --git a/examples/common-ggml.cpp b/examples/common-ggml.cpp index 760cd1f44be..bfc7609afa1 100644 --- a/examples/common-ggml.cpp +++ b/examples/common-ggml.cpp @@ -217,6 +217,7 @@ bool ggml_common_quantize_0( case GGML_TYPE_Q4_0_8_8: case GGML_TYPE_TQ1_0: case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ4_NL_4_4: case GGML_TYPE_COUNT: { fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype)); diff --git a/examples/talk-llama/llama.cpp b/examples/talk-llama/llama.cpp index c51b36e6604..00f78639e6c 100644 --- a/examples/talk-llama/llama.cpp +++ b/examples/talk-llama/llama.cpp @@ -179,7 +179,7 @@ enum llm_arch { LLM_ARCH_COMMAND_R, LLM_ARCH_DBRX, LLM_ARCH_OLMO, - LLM_ARCH_OLMO_1124, + LLM_ARCH_OLMO2, LLM_ARCH_OLMOE, LLM_ARCH_OPENELM, LLM_ARCH_ARCTIC, @@ -233,7 +233,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_COMMAND_R, "command-r" }, { LLM_ARCH_DBRX, "dbrx" }, { LLM_ARCH_OLMO, "olmo" }, - { LLM_ARCH_OLMO_1124, "olmo_1124" }, + { LLM_ARCH_OLMO2, "olmo2" }, { LLM_ARCH_OLMOE, "olmoe" }, { LLM_ARCH_OPENELM, "openelm" }, { LLM_ARCH_ARCTIC, "arctic" }, @@ -1036,6 +1036,8 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, { LLM_TENSOR_OUTPUT, "output" }, { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ROPE_FACTORS_LONG, "rope_factors_long" }, + { LLM_TENSOR_ROPE_FACTORS_SHORT, "rope_factors_short" }, { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, @@ -1210,7 +1212,7 @@ static const std::map> LLM_TENSOR_N }, }, { - LLM_ARCH_OLMO_1124, + LLM_ARCH_OLMO2, { { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, @@ -1549,6 +1551,67 @@ static const std::map> LLM_TENSOR_N }, }; +enum llm_chat_template { + LLM_CHAT_TEMPLATE_CHATML, + LLM_CHAT_TEMPLATE_LLAMA_2, + LLM_CHAT_TEMPLATE_LLAMA_2_SYS, + LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS, + LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP, + LLM_CHAT_TEMPLATE_MISTRAL_V1, + LLM_CHAT_TEMPLATE_MISTRAL_V3, + LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN, + LLM_CHAT_TEMPLATE_MISTRAL_V7, + LLM_CHAT_TEMPLATE_PHI_3, + LLM_CHAT_TEMPLATE_ZEPHYR, + LLM_CHAT_TEMPLATE_MONARCH, + LLM_CHAT_TEMPLATE_GEMMA, + LLM_CHAT_TEMPLATE_ORION, + LLM_CHAT_TEMPLATE_OPENCHAT, + LLM_CHAT_TEMPLATE_VICUNA, + LLM_CHAT_TEMPLATE_VICUNA_ORCA, + LLM_CHAT_TEMPLATE_DEEPSEEK, + LLM_CHAT_TEMPLATE_DEEPSEEK_2, + LLM_CHAT_TEMPLATE_COMMAND_R, + LLM_CHAT_TEMPLATE_LLAMA_3, + LLM_CHAT_TEMPLATE_CHATGML_3, + LLM_CHAT_TEMPLATE_CHATGML_4, + LLM_CHAT_TEMPLATE_MINICPM, + LLM_CHAT_TEMPLATE_EXAONE_3, + LLM_CHAT_TEMPLATE_RWKV_WORLD, + LLM_CHAT_TEMPLATE_GRANITE, + LLM_CHAT_TEMPLATE_UNKNOWN, +}; + +static const std::map LLM_CHAT_TEMPLATES = { + { "chatml", LLM_CHAT_TEMPLATE_CHATML }, + { "llama2", LLM_CHAT_TEMPLATE_LLAMA_2 }, + { "llama2-sys", LLM_CHAT_TEMPLATE_LLAMA_2_SYS }, + { "llama2-sys-bos", LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS }, + { "llama2-sys-strip", LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP }, + { "mistral-v1", LLM_CHAT_TEMPLATE_MISTRAL_V1 }, + { "mistral-v3", LLM_CHAT_TEMPLATE_MISTRAL_V3 }, + { "mistral-v3-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN }, + { "mistral-v7", LLM_CHAT_TEMPLATE_MISTRAL_V7 }, + { "phi3", LLM_CHAT_TEMPLATE_PHI_3 }, + { "zephyr", LLM_CHAT_TEMPLATE_ZEPHYR }, + { "monarch", LLM_CHAT_TEMPLATE_MONARCH }, + { "gemma", LLM_CHAT_TEMPLATE_GEMMA }, + { "orion", LLM_CHAT_TEMPLATE_ORION }, + { "openchat", LLM_CHAT_TEMPLATE_OPENCHAT }, + { "vicuna", LLM_CHAT_TEMPLATE_VICUNA }, + { "vicuna-orca", LLM_CHAT_TEMPLATE_VICUNA_ORCA }, + { "deepseek", LLM_CHAT_TEMPLATE_DEEPSEEK }, + { "deepseek2", LLM_CHAT_TEMPLATE_DEEPSEEK_2 }, + { "command-r", LLM_CHAT_TEMPLATE_COMMAND_R }, + { "llama3", LLM_CHAT_TEMPLATE_LLAMA_3 }, + { "chatglm3", LLM_CHAT_TEMPLATE_CHATGML_3 }, + { "chatglm4", LLM_CHAT_TEMPLATE_CHATGML_4 }, + { "minicpm", LLM_CHAT_TEMPLATE_MINICPM }, + { "exaone3", LLM_CHAT_TEMPLATE_EXAONE_3 }, + { "rwkv-world", LLM_CHAT_TEMPLATE_RWKV_WORLD }, + { "granite", LLM_CHAT_TEMPLATE_GRANITE }, +}; + static llm_arch llm_arch_from_string(const std::string & name) { for (const auto & kv : LLM_ARCH_NAMES) { // NOLINT if (kv.second == name) { @@ -1622,9 +1685,10 @@ struct LLM_TN { // static const std::map LLAMA_ROPE_SCALING_TYPES = { - { LLAMA_ROPE_SCALING_TYPE_NONE, "none" }, - { LLAMA_ROPE_SCALING_TYPE_LINEAR, "linear" }, - { LLAMA_ROPE_SCALING_TYPE_YARN, "yarn" }, + { LLAMA_ROPE_SCALING_TYPE_NONE, "none" }, + { LLAMA_ROPE_SCALING_TYPE_LINEAR, "linear" }, + { LLAMA_ROPE_SCALING_TYPE_YARN, "yarn" }, + { LLAMA_ROPE_SCALING_TYPE_LONGROPE, "longrope" }, }; static llama_rope_scaling_type llama_rope_scaling_type_from_string(const std::string & name) { @@ -2341,6 +2405,7 @@ enum e_model { MODEL_16B, MODEL_20B, MODEL_30B, + MODEL_32B, MODEL_34B, MODEL_35B, MODEL_40B, @@ -4866,7 +4931,9 @@ struct llama_model_loader { mappings.reserve(files.size()); mmaps_used.reserve(files.size()); for (const auto & file : files) { - std::unique_ptr mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, ggml_is_numa())); + auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU)); + auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa"); + std::unique_ptr mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, is_numa_fn())); mmaps_used.emplace_back(mapping->size, 0); if (mlock_mmaps) { std::unique_ptr mlock_mmap(new llama_mlock()); @@ -5328,6 +5395,7 @@ static const char * llama_model_type_name(e_model type) { case MODEL_16B: return "16B"; case MODEL_20B: return "20B"; case MODEL_30B: return "30B"; + case MODEL_32B: return "32B"; case MODEL_34B: return "34B"; case MODEL_35B: return "35B"; case MODEL_40B: return "40B"; @@ -5515,8 +5583,12 @@ static void llm_load_hparams( case LLM_ARCH_MINICPM: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale); + ml.get_key(LLM_KV_RESIDUAL_SCALE, hparams.f_residual_scale); + ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale); switch (hparams.n_layer) { + case 52: model.type = e_model::MODEL_1B; break; case 40: model.type = e_model::MODEL_2B; break; default: model.type = e_model::MODEL_UNKNOWN; } @@ -5688,7 +5760,10 @@ static void llm_load_hparams( case 24: model.type = hparams.n_embd == 1024 ? e_model::MODEL_0_5B : e_model::MODEL_1B; break; case 28: model.type = hparams.n_embd == 1536 ? e_model::MODEL_1_5B : e_model::MODEL_7B; break; case 32: model.type = e_model::MODEL_7B; break; + case 36: model.type = e_model::MODEL_3B; break; case 40: model.type = hparams.n_head() == 20 ? e_model::MODEL_4B : e_model::MODEL_13B; break; + case 48: model.type = e_model::MODEL_14B; break; + case 64: model.type = e_model::MODEL_32B; break; case 80: model.type = e_model::MODEL_70B; break; default: model.type = e_model::MODEL_UNKNOWN; } @@ -5898,7 +5973,7 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; - case LLM_ARCH_OLMO_1124: + case LLM_ARCH_OLMO2: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); @@ -6997,7 +7072,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp); } - if (model.arch == LLM_ARCH_GRANITE || model.arch == LLM_ARCH_GRANITE_MOE) { + if (model.arch == LLM_ARCH_MINICPM || model.arch == LLM_ARCH_GRANITE || model.arch == LLM_ARCH_GRANITE_MOE) { LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale); LLAMA_LOG_INFO("%s: f_residual_scale = %f\n", __func__, hparams.f_residual_scale); LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale); @@ -7181,12 +7256,12 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w } break; case GGML_OP_ADD: { - ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, w->ne[0], 512); + ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], w->ne[1], w->ne[2], w->ne[3]); op_tensor = ggml_add(ctx, a, w); } break; case GGML_OP_MUL: { - ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, w->ne[0], 512); + ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], w->ne[1], w->ne[2], w->ne[3]); op_tensor = ggml_mul(ctx, a, w); } break; case GGML_OP_DIV: @@ -7622,7 +7697,13 @@ static bool llm_load_tensors( layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); - layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0)); + if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) { + layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0)); + layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0)); + } + else { + layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0)); + } if (n_expert == 0) { layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); @@ -8591,7 +8672,7 @@ static bool llm_load_tensors( layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); } } break; - case LLM_ARCH_OLMO_1124: + case LLM_ARCH_OLMO2: { model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -9190,7 +9271,7 @@ static bool llm_load_tensors( ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft); if (!dev) { // FIXME: workaround for CPU backend buft having a NULL device - dev = ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0); + dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); } ggml_backend_dev_props props; ggml_backend_dev_get_props(dev, &props); @@ -13429,153 +13510,6 @@ struct llm_build_context { return gf; } - // ref: https://arxiv.org/abs/2203.03466 - // https://github.com/ggerganov/llama.cpp/issues/5276#issuecomment-1925774738 - // based on the original build_llama() function - struct ggml_cgraph * build_minicpm() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - const int64_t n_embd = hparams.n_embd; - //TODO: if the model varies, these parameters need to be read from the model - const int64_t n_embd_base = 256; - const float scale_embd = 12.0f; - const float scale_depth = 1.4f; - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // scale the input embeddings - inpL = ggml_scale(ctx0, inpL, scale_embd); - cb(inpL, "inp_scaled", -1); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - // scale_res - scale the hidden states for residual connection - const float scale_res = scale_depth/sqrtf(float(n_layer)); - cur = ggml_scale(ctx0, cur, scale_res); - cb(cur, "hidden_scaled", -1); - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - // scale the hidden states for residual connection - cur = ggml_scale(ctx0, cur, scale_res); - cb(cur, "hidden_scaled_ffn", -1); - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head scaling - const float scale_lmhead = float(n_embd_base)/float(n_embd); - cur = ggml_scale(ctx0, cur, scale_lmhead); - cb(cur, "lmhead_scaling", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - struct ggml_cgraph * build_minicpm3() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); @@ -14481,7 +14415,7 @@ struct llm_build_context { return gf; } - struct ggml_cgraph * build_olmo_1124() { + struct ggml_cgraph * build_olmo2() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); // mutable variable, needed during the last layer of the computation to skip unused tokens @@ -16674,6 +16608,7 @@ static struct ggml_cgraph * llama_build_graph( switch (model.arch) { case LLM_ARCH_LLAMA: + case LLM_ARCH_MINICPM: case LLM_ARCH_GRANITE: case LLM_ARCH_GRANITE_MOE: { @@ -16757,10 +16692,6 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_internlm2(); } break; - case LLM_ARCH_MINICPM: - { - result = llm.build_minicpm(); - } break; case LLM_ARCH_MINICPM3: { result = llm.build_minicpm3(); @@ -16797,9 +16728,9 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_olmo(); } break; - case LLM_ARCH_OLMO_1124: + case LLM_ARCH_OLMO2: { - result = llm.build_olmo_1124(); + result = llm.build_olmo2(); } break; case LLM_ARCH_OLMOE: { @@ -17443,8 +17374,9 @@ static enum ggml_status llama_graph_compute( int n_threads, ggml_threadpool * threadpool) { if (lctx.backend_cpu != nullptr) { - ggml_backend_cpu_set_threadpool(lctx.backend_cpu, threadpool); - ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data); + auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(lctx.backend_cpu)); + auto * set_threadpool_fn = (decltype(ggml_backend_cpu_set_threadpool) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool"); + set_threadpool_fn(lctx.backend_cpu, threadpool); } // set the number of threads for all the backends @@ -18211,13 +18143,13 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) { static void llama_kv_cache_update_internal(struct llama_context & lctx) { bool need_reserve = false; - // apply K-shift if needed - if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE && lctx.kv_self.has_shift) { + if (lctx.kv_self.has_shift) { if (!llama_kv_cache_can_shift(&lctx)) { - GGML_ABORT("Deepseek2 does not support K-shift"); + GGML_ABORT("The current context does not support K-shift"); } - { + // apply K-shift if needed + if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE) { ggml_backend_sched_reset(lctx.sched.get()); ggml_cgraph * gf = llama_build_graph_k_shift(lctx); @@ -19361,6 +19293,7 @@ void llama_lora_adapter_free(struct llama_lora_adapter * adapter) { // struct llama_model_params llama_model_default_params() { struct llama_model_params result = { + /*.devices =*/ nullptr, /*.n_gpu_layers =*/ 0, /*.split_mode =*/ LLAMA_SPLIT_MODE_LAYER, /*.main_gpu =*/ 0, @@ -19478,7 +19411,11 @@ void llama_backend_init(void) { void llama_numa_init(enum ggml_numa_strategy numa) { if (numa != GGML_NUMA_STRATEGY_DISABLED) { - ggml_numa_init(numa); + auto * dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); + GGML_ASSERT(dev && "CPU backend is not loaded"); + auto * reg = ggml_backend_dev_backend_reg(dev); + auto * numa_init_fn = (decltype(ggml_numa_init) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_numa_init"); + numa_init_fn(numa); } } @@ -19569,19 +19506,24 @@ struct llama_model * llama_load_model_from_file( } // create list of devices to use with this model - // currently, we use all available devices - // TODO: rework API to give user more control over device selection - for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { - ggml_backend_dev_t dev = ggml_backend_dev_get(i); - switch (ggml_backend_dev_type(dev)) { - case GGML_BACKEND_DEVICE_TYPE_CPU: - case GGML_BACKEND_DEVICE_TYPE_ACCEL: - // skip CPU backends since they are handled separately - break; + if (params.devices) { + for (ggml_backend_dev_t * dev = params.devices; *dev; ++dev) { + model->devices.push_back(*dev); + } + } else { + // use all available devices + for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { + ggml_backend_dev_t dev = ggml_backend_dev_get(i); + switch (ggml_backend_dev_type(dev)) { + case GGML_BACKEND_DEVICE_TYPE_CPU: + case GGML_BACKEND_DEVICE_TYPE_ACCEL: + // skip CPU backends since they are handled separately + break; - case GGML_BACKEND_DEVICE_TYPE_GPU: - model->devices.push_back(dev); - break; + case GGML_BACKEND_DEVICE_TYPE_GPU: + model->devices.push_back(dev); + break; + } } } @@ -19752,9 +19694,6 @@ struct llama_context * llama_new_context_with_model( __func__, n_ctx_per_seq, hparams.n_ctx_train); } - ctx->abort_callback = params.abort_callback; - ctx->abort_callback_data = params.abort_callback_data; - ctx->logits_all = params.logits_all; // build worst-case graph for encoder if a model contains encoder @@ -19803,7 +19742,7 @@ struct llama_context * llama_new_context_with_model( } // add CPU backend - ctx->backend_cpu = ggml_backend_cpu_init(); + ctx->backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr); if (ctx->backend_cpu == nullptr) { LLAMA_LOG_ERROR("%s: failed to initialize CPU backend\n", __func__); llama_free(ctx); @@ -19823,6 +19762,8 @@ struct llama_context * llama_new_context_with_model( } } + llama_set_abort_callback(ctx, params.abort_callback, params.abort_callback_data); + if (!llama_kv_cache_init(ctx->kv_self, ctx, type_k, type_v, kv_size, cparams.offload_kqv)) { LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__); llama_free(ctx); @@ -19868,7 +19809,8 @@ struct llama_context * llama_new_context_with_model( std::vector backend_ptrs; for (auto & backend : ctx->backends) { auto * buft = ggml_backend_get_default_buffer_type(backend.get()); - if (ggml_backend_is_cpu(backend.get()) && !model->devices.empty()) { + auto backend_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get())); + if (backend_type == GGML_BACKEND_DEVICE_TYPE_CPU && !model->devices.empty()) { // use the host buffer of the first device CPU for faster transfer of the intermediate state auto * dev = model->devices[0]; auto * host_buft = ggml_backend_dev_host_buffer_type(dev); @@ -19896,7 +19838,8 @@ struct llama_context * llama_new_context_with_model( // pipeline parallelism requires support for async compute and events in all devices if (pipeline_parallel) { for (auto & backend : ctx->backends) { - if (ggml_backend_is_cpu(backend.get())) { + auto dev_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get())); + if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU) { // ignore CPU backend continue; } @@ -20070,7 +20013,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_QWEN: case LLM_ARCH_QWEN2: case LLM_ARCH_QWEN2MOE: - case LLM_ARCH_OLMO_1124: + case LLM_ARCH_OLMO2: case LLM_ARCH_OLMOE: case LLM_ARCH_PHI2: case LLM_ARCH_PHI3: @@ -20463,7 +20406,7 @@ void llama_kv_cache_update(struct llama_context * ctx) { } bool llama_kv_cache_can_shift(struct llama_context * ctx) { - return ctx->model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA + return !ctx->kv_self.recurrent && ctx->model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA } // deprecated @@ -21450,6 +21393,14 @@ int32_t llama_n_threads_batch(struct llama_context * ctx) { void llama_set_abort_callback(struct llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) { ctx->abort_callback = abort_callback; ctx->abort_callback_data = abort_callback_data; + + for (auto & backend : ctx->backends) { + auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend.get())); + auto * set_abort_callback_fn = (ggml_backend_set_abort_callback_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback"); + if (set_abort_callback_fn) { + set_abort_callback_fn(backend.get(), ctx->abort_callback, ctx->abort_callback_data); + } + } } void llama_set_embeddings(struct llama_context * ctx, bool embeddings) { @@ -21816,18 +21767,109 @@ int32_t llama_detokenize( // chat templates // +static llm_chat_template llama_chat_detect_template(const std::string & tmpl) { + if (LLM_CHAT_TEMPLATES.find(tmpl) != LLM_CHAT_TEMPLATES.end()) { + return LLM_CHAT_TEMPLATES.at(tmpl); + } + auto tmpl_contains = [&tmpl](const char * haystack) -> bool { + return tmpl.find(haystack) != std::string::npos; + }; + if (tmpl_contains("<|im_start|>")) { + return LLM_CHAT_TEMPLATE_CHATML; + } else if (tmpl.find("mistral") == 0 || tmpl_contains("[INST]")) { + if (tmpl_contains("[SYSTEM_PROMPT]")) { + return LLM_CHAT_TEMPLATE_MISTRAL_V7; + } else if ( + // catches official 'v1' template + tmpl_contains("' [INST] ' + system_message") + // catches official 'v3' and 'v3-tekken' templates + || tmpl_contains("[AVAILABLE_TOOLS]") + ) { + // Official mistral 'v1', 'v3' and 'v3-tekken' templates + // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/chat_templates.md + // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/templates.md + if (tmpl_contains(" [INST]")) { + return LLM_CHAT_TEMPLATE_MISTRAL_V1; + } else if (tmpl_contains("\"[INST]\"")) { + return LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN; + } + return LLM_CHAT_TEMPLATE_MISTRAL_V3; + } else { + // llama2 template and its variants + // [variant] support system message + // See: https://huggingface.co/blog/llama2#how-to-prompt-llama-2 + bool support_system_message = tmpl_contains("<>"); + bool add_bos_inside_history = tmpl_contains("bos_token + '[INST]"); + bool strip_message = tmpl_contains("content.strip()"); + if (strip_message) { + return LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP; + } else if (add_bos_inside_history) { + return LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS; + } else if (support_system_message) { + return LLM_CHAT_TEMPLATE_LLAMA_2_SYS; + } else { + return LLM_CHAT_TEMPLATE_LLAMA_2; + } + } + } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) { + return LLM_CHAT_TEMPLATE_PHI_3; + } else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) { + return LLM_CHAT_TEMPLATE_ZEPHYR; + } else if (tmpl_contains("bos_token + message['role']")) { + return LLM_CHAT_TEMPLATE_MONARCH; + } else if (tmpl_contains("")) { + return LLM_CHAT_TEMPLATE_GEMMA; + } else if (tmpl_contains("'\\n\\nAssistant: ' + eos_token")) { + // OrionStarAI/Orion-14B-Chat + return LLM_CHAT_TEMPLATE_ORION; + } else if (tmpl_contains("GPT4 Correct ")) { + // openchat/openchat-3.5-0106 + return LLM_CHAT_TEMPLATE_OPENCHAT; + } else if (tmpl_contains("USER: ") && tmpl_contains("ASSISTANT: ")) { + // eachadea/vicuna-13b-1.1 (and Orca variant) + if (tmpl_contains("SYSTEM: ")) { + return LLM_CHAT_TEMPLATE_VICUNA_ORCA; + } + return LLM_CHAT_TEMPLATE_VICUNA; + } else if (tmpl_contains("### Instruction:") && tmpl_contains("<|EOT|>")) { + // deepseek-ai/deepseek-coder-33b-instruct + return LLM_CHAT_TEMPLATE_DEEPSEEK; + } else if (tmpl_contains("<|START_OF_TURN_TOKEN|>") && tmpl_contains("<|USER_TOKEN|>")) { + // CohereForAI/c4ai-command-r-plus + return LLM_CHAT_TEMPLATE_COMMAND_R; + } else if (tmpl_contains("<|start_header_id|>") && tmpl_contains("<|end_header_id|>")) { + return LLM_CHAT_TEMPLATE_LLAMA_3; + } else if (tmpl_contains("[gMASK]sop")) { + // chatglm3-6b + return LLM_CHAT_TEMPLATE_CHATGML_3; + } else if (tmpl_contains("[gMASK]")) { + return LLM_CHAT_TEMPLATE_CHATGML_4; + } else if (tmpl_contains(LU8("<用户>"))) { + // MiniCPM-3B-OpenHermes-2.5-v2-GGUF + return LLM_CHAT_TEMPLATE_MINICPM; + } else if (tmpl_contains("'Assistant: ' + message['content'] + eos_token")) { + return LLM_CHAT_TEMPLATE_DEEPSEEK_2; + } else if (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]")) { + // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb + // EXAONE-3.0-7.8B-Instruct + return LLM_CHAT_TEMPLATE_EXAONE_3; + } else if (tmpl_contains("rwkv-world")) { + return LLM_CHAT_TEMPLATE_RWKV_WORLD; + } else if (tmpl_contains("<|start_of_role|>")) { + return LLM_CHAT_TEMPLATE_GRANITE; + } + return LLM_CHAT_TEMPLATE_UNKNOWN; +} + // Simple version of "llama_apply_chat_template" that only works with strings // This function uses heuristic checks to determine commonly used template. It is not a jinja parser. static int32_t llama_chat_apply_template_internal( - const std::string & tmpl, + const llm_chat_template tmpl, const std::vector & chat, std::string & dest, bool add_ass) { // Taken from the research: https://github.com/ggerganov/llama.cpp/issues/5527 std::stringstream ss; - auto tmpl_contains = [&tmpl](std::string haystack) -> bool { - return tmpl.find(haystack) != std::string::npos; - }; - if (tmpl == "chatml" || tmpl_contains("<|im_start|>")) { + if (tmpl == LLM_CHAT_TEMPLATE_CHATML) { // chatml template for (auto message : chat) { ss << "<|im_start|>" << message->role << "\n" << message->content << "<|im_end|>\n"; @@ -21835,16 +21877,59 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|im_start|>assistant\n"; } - } else if (tmpl == "llama2" || tmpl == "mistral" || tmpl_contains("[INST]")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7) { + // Official mistral 'v7' template + // See: https://huggingface.co/mistralai/Mistral-Large-Instruct-2411#basic-instruct-template-v7 + for (auto message : chat) { + std::string role(message->role); + std::string content(message->content); + if (role == "system") { + ss << "[SYSTEM_PROMPT] " << content << "[/SYSTEM_PROMPT]"; + } else if (role == "user") { + ss << "[INST] " << content << "[/INST]"; + } + else { + ss << " " << content << ""; + } + } + } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1 + || tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3 + || tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN) { + // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/chat_templates.md + // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/templates.md + std::string leading_space = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1 ? " " : ""; + std::string trailing_space = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN ? "" : " "; + bool trim_assistant_message = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3; + bool is_inside_turn = false; + for (auto message : chat) { + if (!is_inside_turn) { + ss << leading_space << "[INST]" << trailing_space; + is_inside_turn = true; + } + std::string role(message->role); + std::string content(message->content); + if (role == "system") { + ss << content << "\n\n"; + } else if (role == "user") { + ss << content << leading_space << "[/INST]"; + } else { + ss << trailing_space << (trim_assistant_message ? trim(content) : content) << ""; + is_inside_turn = false; + } + } + } else if ( + tmpl == LLM_CHAT_TEMPLATE_LLAMA_2 + || tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS + || tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS + || tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP) { // llama2 template and its variants // [variant] support system message - bool support_system_message = tmpl_contains("<>") || tmpl == "mistral"; - // [variant] space before + after response - bool space_around_response = tmpl_contains("' ' + eos_token"); + // See: https://huggingface.co/blog/llama2#how-to-prompt-llama-2 + bool support_system_message = tmpl != LLM_CHAT_TEMPLATE_LLAMA_2; // [variant] add BOS inside history - bool add_bos_inside_history = tmpl_contains("bos_token + '[INST]"); + bool add_bos_inside_history = tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS; // [variant] trim spaces from the input message - bool strip_message = tmpl_contains("content.strip()"); + bool strip_message = tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP; // construct the prompt bool is_inside_turn = true; // skip BOS at the beginning ss << "[INST] "; @@ -21865,12 +21950,11 @@ static int32_t llama_chat_apply_template_internal( } else if (role == "user") { ss << content << " [/INST]"; } else { - ss << (space_around_response ? " " : "") << content << (space_around_response ? " " : "") << ""; + ss << content << ""; is_inside_turn = false; } } - // llama2 templates seem to not care about "add_generation_prompt" - } else if (tmpl == "phi3" || (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>"))) { + } else if (tmpl == LLM_CHAT_TEMPLATE_PHI_3) { // Phi 3 for (auto message : chat) { std::string role(message->role); @@ -21879,7 +21963,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|assistant|>\n"; } - } else if (tmpl == "zephyr" || tmpl_contains("<|user|>")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_ZEPHYR) { // zephyr template for (auto message : chat) { ss << "<|" << message->role << "|>" << "\n" << message->content << "<|endoftext|>\n"; @@ -21887,7 +21971,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|assistant|>\n"; } - } else if (tmpl == "monarch" || tmpl_contains("bos_token + message['role']")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_MONARCH) { // mlabonne/AlphaMonarch-7B template (the is included inside history) for (auto message : chat) { std::string bos = (message == chat.front()) ? "" : ""; // skip BOS for first message @@ -21896,7 +21980,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "assistant\n"; } - } else if (tmpl == "gemma" || tmpl == "gemma2" || tmpl_contains("")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_GEMMA) { // google/gemma-7b-it std::string system_prompt = ""; for (auto message : chat) { @@ -21918,7 +22002,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "model\n"; } - } else if (tmpl == "orion" || tmpl_contains("'\\n\\nAssistant: ' + eos_token")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_ORION) { // OrionStarAI/Orion-14B-Chat std::string system_prompt = ""; for (auto message : chat) { @@ -21938,7 +22022,7 @@ static int32_t llama_chat_apply_template_internal( ss << message->content << ""; } } - } else if (tmpl == "openchat" || tmpl_contains("GPT4 Correct ")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_OPENCHAT) { // openchat/openchat-3.5-0106, for (auto message : chat) { std::string role(message->role); @@ -21952,13 +22036,13 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "GPT4 Correct Assistant:"; } - } else if (tmpl == "vicuna" || tmpl == "vicuna-orca" || (tmpl_contains("USER: ") && tmpl_contains("ASSISTANT: "))) { + } else if (tmpl == LLM_CHAT_TEMPLATE_VICUNA || tmpl == LLM_CHAT_TEMPLATE_VICUNA_ORCA) { // eachadea/vicuna-13b-1.1 (and Orca variant) for (auto message : chat) { std::string role(message->role); if (role == "system") { // Orca-Vicuna variant uses a system prefix - if (tmpl == "vicuna-orca" || tmpl_contains("SYSTEM: ")) { + if (tmpl == LLM_CHAT_TEMPLATE_VICUNA_ORCA) { ss << "SYSTEM: " << message->content << "\n"; } else { ss << message->content << "\n\n"; @@ -21972,7 +22056,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "ASSISTANT:"; } - } else if (tmpl == "deepseek" || (tmpl_contains("### Instruction:") && tmpl_contains("<|EOT|>"))) { + } else if (tmpl == LLM_CHAT_TEMPLATE_DEEPSEEK) { // deepseek-ai/deepseek-coder-33b-instruct for (auto message : chat) { std::string role(message->role); @@ -21987,7 +22071,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "### Response:\n"; } - } else if (tmpl == "command-r" || (tmpl_contains("<|START_OF_TURN_TOKEN|>") && tmpl_contains("<|USER_TOKEN|>"))) { + } else if (tmpl == LLM_CHAT_TEMPLATE_COMMAND_R) { // CohereForAI/c4ai-command-r-plus for (auto message : chat) { std::string role(message->role); @@ -22002,7 +22086,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"; } - } else if (tmpl == "llama3" || (tmpl_contains("<|start_header_id|>") && tmpl_contains("<|end_header_id|>"))) { + } else if (tmpl == LLM_CHAT_TEMPLATE_LLAMA_3) { // Llama 3 for (auto message : chat) { std::string role(message->role); @@ -22011,7 +22095,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|start_header_id|>assistant<|end_header_id|>\n\n"; } - } else if (tmpl == "chatglm3" || tmpl_contains("[gMASK]sop")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_3) { // chatglm3-6b ss << "[gMASK]" << "sop"; for (auto message : chat) { @@ -22021,7 +22105,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|assistant|>"; } - } else if (tmpl == "chatglm4" || tmpl_contains("[gMASK]")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_4) { ss << "[gMASK]" << ""; for (auto message : chat) { std::string role(message->role); @@ -22030,7 +22114,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|assistant|>"; } - } else if (tmpl == "minicpm" || tmpl_contains(LU8("<用户>"))) { + } else if (tmpl == LLM_CHAT_TEMPLATE_MINICPM) { // MiniCPM-3B-OpenHermes-2.5-v2-GGUF for (auto message : chat) { std::string role(message->role); @@ -22042,7 +22126,7 @@ static int32_t llama_chat_apply_template_internal( ss << trim(message->content); } } - } else if (tmpl == "deepseek2" || tmpl_contains("'Assistant: ' + message['content'] + eos_token")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_DEEPSEEK_2) { // DeepSeek-V2 for (auto message : chat) { std::string role(message->role); @@ -22057,7 +22141,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "Assistant:"; } - } else if (tmpl == "exaone3" || (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]"))) { + } else if (tmpl == LLM_CHAT_TEMPLATE_EXAONE_3) { // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb // EXAONE-3.0-7.8B-Instruct for (auto message : chat) { @@ -22073,7 +22157,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "[|assistant|]"; } - } else if (tmpl == "rwkv-world" || tmpl_contains("rwkv-world")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) { // this template requires the model to have "\n\n" as EOT token for (auto message : chat) { std::string role(message->role); @@ -22083,7 +22167,7 @@ static int32_t llama_chat_apply_template_internal( ss << message->content << "\n\n"; } } - } else if (tmpl == "granite" || tmpl_contains("<|start_of_role|>")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_GRANITE) { // IBM Granite template for (const auto & message : chat) { std::string role(message->role); @@ -22135,7 +22219,11 @@ int32_t llama_chat_apply_template( } std::string formatted_chat; - int32_t res = llama_chat_apply_template_internal(curr_tmpl, chat_vec, formatted_chat, add_ass); + llm_chat_template detected_tmpl = llama_chat_detect_template(curr_tmpl); + if (detected_tmpl == LLM_CHAT_TEMPLATE_UNKNOWN) { + return -1; + } + int32_t res = llama_chat_apply_template_internal(detected_tmpl, chat_vec, formatted_chat, add_ass); if (res < 0) { return res; } @@ -22145,6 +22233,15 @@ int32_t llama_chat_apply_template( return res; } +int32_t llama_chat_builtin_templates(const char ** output, size_t len) { + auto it = LLM_CHAT_TEMPLATES.begin(); + for (size_t i = 0; i < std::min(len, LLM_CHAT_TEMPLATES.size()); i++) { + output[i] = it->first.c_str(); + std::advance(it, 1); + } + return (int32_t) LLM_CHAT_TEMPLATES.size(); +} + // // sampling // @@ -22191,32 +22288,23 @@ int llama_split_prefix(char * dest, size_t maxlen, const char * split_path, int } const char * llama_print_system_info(void) { - ggml_cpu_init(); // some ARM features are detected at runtime - static std::string s; - s = ""; - s += "AVX = " + std::to_string(ggml_cpu_has_avx()) + " | "; - s += "AVX_VNNI = " + std::to_string(ggml_cpu_has_avx_vnni()) + " | "; - s += "AVX2 = " + std::to_string(ggml_cpu_has_avx2()) + " | "; - s += "AVX512 = " + std::to_string(ggml_cpu_has_avx512()) + " | "; - s += "AVX512_VBMI = " + std::to_string(ggml_cpu_has_avx512_vbmi()) + " | "; - s += "AVX512_VNNI = " + std::to_string(ggml_cpu_has_avx512_vnni()) + " | "; - s += "AVX512_BF16 = " + std::to_string(ggml_cpu_has_avx512_bf16()) + " | "; - s += "AMX_INT8 = " + std::to_string(ggml_cpu_has_amx_int8()) + " | "; - s += "FMA = " + std::to_string(ggml_cpu_has_fma()) + " | "; - s += "NEON = " + std::to_string(ggml_cpu_has_neon()) + " | "; - s += "SVE = " + std::to_string(ggml_cpu_has_sve()) + " | "; - s += "ARM_FMA = " + std::to_string(ggml_cpu_has_arm_fma()) + " | "; - s += "F16C = " + std::to_string(ggml_cpu_has_f16c()) + " | "; - s += "FP16_VA = " + std::to_string(ggml_cpu_has_fp16_va()) + " | "; - s += "RISCV_VECT = " + std::to_string(ggml_cpu_has_riscv_v()) + " | "; - s += "WASM_SIMD = " + std::to_string(ggml_cpu_has_wasm_simd()) + " | "; - s += "SSE3 = " + std::to_string(ggml_cpu_has_sse3()) + " | "; - s += "SSSE3 = " + std::to_string(ggml_cpu_has_ssse3()) + " | "; - s += "VSX = " + std::to_string(ggml_cpu_has_vsx()) + " | "; - s += "MATMUL_INT8 = " + std::to_string(ggml_cpu_has_matmul_int8()) + " | "; - s += "LLAMAFILE = " + std::to_string(ggml_cpu_has_llamafile()) + " | "; + for (size_t i = 0; i < ggml_backend_reg_count(); i++) { + auto * reg = ggml_backend_reg_get(i); + auto * get_features_fn = (ggml_backend_get_features_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_get_features"); + if (get_features_fn) { + ggml_backend_feature * features = get_features_fn(reg); + s += ggml_backend_reg_name(reg); + s += " : "; + for (; features->name; features++) { + s += features->name; + s += " = "; + s += features->value; + s += " | "; + } + } + } return s.c_str(); } diff --git a/examples/talk-llama/llama.h b/examples/talk-llama/llama.h index 90791d5f5ea..168c3fa1f6e 100644 --- a/examples/talk-llama/llama.h +++ b/examples/talk-llama/llama.h @@ -185,7 +185,8 @@ extern "C" { LLAMA_ROPE_SCALING_TYPE_NONE = 0, LLAMA_ROPE_SCALING_TYPE_LINEAR = 1, LLAMA_ROPE_SCALING_TYPE_YARN = 2, - LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN, + LLAMA_ROPE_SCALING_TYPE_LONGROPE = 3, + LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_LONGROPE, }; enum llama_pooling_type { @@ -272,6 +273,9 @@ extern "C" { }; struct llama_model_params { + // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used) + ggml_backend_dev_t * devices; + int32_t n_gpu_layers; // number of layers to store in VRAM enum llama_split_mode split_mode; // how to split the model across multiple GPUs @@ -987,6 +991,9 @@ extern "C" { char * buf, int32_t length); + // Get list of built-in chat templates + LLAMA_API int32_t llama_chat_builtin_templates(const char ** output, size_t len); + // // Sampling API // diff --git a/examples/talk-llama/unicode.cpp b/examples/talk-llama/unicode.cpp index 50b35bbbc91..3d459263525 100644 --- a/examples/talk-llama/unicode.cpp +++ b/examples/talk-llama/unicode.cpp @@ -201,7 +201,18 @@ static std::unordered_map unicode_utf8_to_byte_map() { } static inline std::wstring unicode_wstring_from_utf8(const std::string & s) { +#if defined(__clang__) + // disable C++17 deprecation warning for std::codecvt_utf8 +# pragma clang diagnostic push +# pragma clang diagnostic ignored "-Wdeprecated-declarations" +#endif + std::wstring_convert> conv; + +#if defined(__clang__) +# pragma clang diagnostic pop +#endif + return conv.from_bytes(s); } diff --git a/examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt b/examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt index 1243d732c34..f5e555c1e4a 100644 --- a/examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt +++ b/examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt @@ -2,11 +2,11 @@ cmake_minimum_required(VERSION 3.10) project(whisper.cpp) -set(CMAKE_CXX_STANDARD 11) +set(CMAKE_CXX_STANDARD 17) set(WHISPER_LIB_DIR ${CMAKE_SOURCE_DIR}/../../../../../../..) # Path to external GGML, otherwise uses the copy in whisper.cpp. -option(GGML_HOME "whisper: Path to external GGML source" OFF) +option(GGML_HOME "whisper: Path to external GGML source" OFF) set( SOURCE_FILES @@ -14,6 +14,8 @@ set( ${CMAKE_SOURCE_DIR}/jni.c ) +# TODO: this needs to be updated to work with the new ggml CMakeLists + if (NOT GGML_HOME) set( SOURCE_FILES diff --git a/examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj b/examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj index 11447909695..fcf0614179e 100644 --- a/examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj +++ b/examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj @@ -363,7 +363,7 @@ GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE; GCC_WARN_UNUSED_FUNCTION = YES; GCC_WARN_UNUSED_VARIABLE = YES; - HEADER_SEARCH_PATHS = ""; + HEADER_SEARCH_PATHS = ../../../ggml/src/; IPHONEOS_DEPLOYMENT_TARGET = 16.0; MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE; MTL_FAST_MATH = YES; @@ -417,7 +417,7 @@ GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE; GCC_WARN_UNUSED_FUNCTION = YES; GCC_WARN_UNUSED_VARIABLE = YES; - HEADER_SEARCH_PATHS = ""; + HEADER_SEARCH_PATHS = ../../../ggml/src/; IPHONEOS_DEPLOYMENT_TARGET = 16.0; MTL_ENABLE_DEBUG_INFO = NO; MTL_FAST_MATH = YES; diff --git a/examples/whisper.swiftui/whisper.cpp.swift/LibWhisper.swift b/examples/whisper.swiftui/whisper.cpp.swift/LibWhisper.swift index 125be4f881c..3c769e7af8e 100644 --- a/examples/whisper.swiftui/whisper.cpp.swift/LibWhisper.swift +++ b/examples/whisper.swiftui/whisper.cpp.swift/LibWhisper.swift @@ -66,7 +66,7 @@ actor WhisperContext { private func systemInfo() -> String { var info = "" - if (ggml_cpu_has_neon() != 0) { info += "NEON " } + //if (ggml_cpu_has_neon() != 0) { info += "NEON " } return String(info.dropLast()) } @@ -75,45 +75,45 @@ actor WhisperContext { if (whisper_set_mel(context, nil, 0, nMels) != 0) { return "error: failed to set mel" } - + // heat encoder if (whisper_encode(context, 0, nThreads) != 0) { return "error: failed to encode" } - + var tokens = [whisper_token](repeating: 0, count: 512) - + // prompt heat if (whisper_decode(context, &tokens, 256, 0, nThreads) != 0) { return "error: failed to decode" } - + // text-generation heat if (whisper_decode(context, &tokens, 1, 256, nThreads) != 0) { return "error: failed to decode" } - + whisper_reset_timings(context) - + // actual run if (whisper_encode(context, 0, nThreads) != 0) { return "error: failed to encode" } - + // text-generation for i in 0..<256 { if (whisper_decode(context, &tokens, 1, Int32(i), nThreads) != 0) { return "error: failed to decode" } } - + // batched decoding for _ in 0..<64 { if (whisper_decode(context, &tokens, 5, 0, nThreads) != 0) { return "error: failed to decode" } } - + // prompt processing for _ in 0..<16 { if (whisper_decode(context, &tokens, 256, 0, nThreads) != 0) { diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt index 9ab91421a7d..1b3d9896781 100644 --- a/ggml/CMakeLists.txt +++ b/ggml/CMakeLists.txt @@ -33,6 +33,7 @@ else() endif() option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT}) +option(GGML_BACKEND_DL "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF) # # option list @@ -91,28 +92,33 @@ else() set(INS_ENB ON) endif() -option(GGML_CPU_HBM "ggml: use memkind for CPU HBM" OFF) -option(GGML_CPU_AARCH64 "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON) - -option(GGML_AVX "ggml: enable AVX" ${INS_ENB}) -option(GGML_AVX2 "ggml: enable AVX2" ${INS_ENB}) -option(GGML_AVX512 "ggml: enable AVX512" OFF) -option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI" OFF) -option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI" OFF) -option(GGML_AVX512_BF16 "ggml: enable AVX512-BF16" OFF) -option(GGML_AMX_TILE "ggml: enable AMX-TILE" OFF) -option(GGML_AMX_INT8 "ggml: enable AMX-INT8" OFF) -option(GGML_AMX_BF16 "ggml: enable AMX-BF16" OFF) -option(GGML_FMA "ggml: enable FMA" ${INS_ENB}) +option(GGML_CPU_HBM "ggml: use memkind for CPU HBM" OFF) +option(GGML_CPU_AARCH64 "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON) +option(GGML_AVX "ggml: enable AVX" ${INS_ENB}) +option(GGML_AVX_VNNI "ggml: enable AVX-VNNI" OFF) +option(GGML_AVX2 "ggml: enable AVX2" ${INS_ENB}) +option(GGML_AVX512 "ggml: enable AVX512F" OFF) +option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI" OFF) +option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI" OFF) +option(GGML_AVX512_BF16 "ggml: enable AVX512-BF16" OFF) if (NOT MSVC) - option(GGML_F16C "ggml: enable F16C" ${INS_ENB}) # in MSVC F16C is implied with AVX2/AVX512 + # in MSVC F16C and FMA is implied with AVX2/AVX512 + option(GGML_FMA "ggml: enable FMA" ${INS_ENB}) + option(GGML_F16C "ggml: enable F16C" ${INS_ENB}) + # MSVC does not seem to support AMX + option(GGML_AMX_TILE "ggml: enable AMX-TILE" OFF) + option(GGML_AMX_INT8 "ggml: enable AMX-INT8" OFF) + option(GGML_AMX_BF16 "ggml: enable AMX-BF16" OFF) endif() -option(GGML_LASX "ggml: enable lasx" ON) -option(GGML_LSX "ggml: enable lsx" ON) -option(GGML_SVE "ggml: enable SVE" OFF) +option(GGML_LASX "ggml: enable lasx" ON) +option(GGML_LSX "ggml: enable lsx" ON) +option(GGML_RVV "ggml: enable rvv" ON) +option(GGML_SVE "ggml: enable SVE" OFF) +option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF) + if (WIN32) - set(GGML_WIN_VER "0x602" CACHE STRING "ggml: Windows Version") + set(GGML_WIN_VER "0x602" CACHE STRING "ggml: Windows version") endif() # ggml core @@ -159,7 +165,6 @@ set (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING set (GGML_METAL_STD "" CACHE STRING "ggml: metal standard version (-std flag)") option(GGML_OPENMP "ggml: use OpenMP" ON) option(GGML_RPC "ggml: use RPC" OFF) -option(GGML_AMX "ggml: use AMX" OFF) option(GGML_SYCL "ggml: use SYCL" OFF) option(GGML_SYCL_F16 "ggml: use 16 bit floats for sycl calculations" OFF) set (GGML_SYCL_TARGET "INTEL" CACHE STRING @@ -178,11 +183,7 @@ option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE}) set(CMAKE_C_STANDARD 11) set(CMAKE_C_STANDARD_REQUIRED true) -if (GGML_SYCL) - set(CMAKE_CXX_STANDARD 17) -else() - set(CMAKE_CXX_STANDARD 11) -endif() +set(CMAKE_CXX_STANDARD 17) set(CMAKE_CXX_STANDARD_REQUIRED true) set(THREADS_PREFER_PTHREAD_FLAG ON) diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h index cef164764bb..19881a5059f 100644 --- a/ggml/include/ggml-backend.h +++ b/ggml/include/ggml-backend.h @@ -190,6 +190,14 @@ extern "C" { typedef void (*ggml_backend_set_n_threads_t)(ggml_backend_t backend, int n_threads); // Get additional buffer types provided by the device (returns a NULL-terminated array) typedef ggml_backend_buffer_type_t * (*ggml_backend_dev_get_extra_bufts_t)(ggml_backend_dev_t device); + // Set the abort callback for the backend + typedef void (*ggml_backend_set_abort_callback_t)(ggml_backend_t backend, ggml_abort_callback abort_callback, void * abort_callback_data); + // Get a list of feature flags supported by the backend (returns a NULL-terminated array) + struct ggml_backend_feature { + const char * name; + const char * value; + }; + typedef struct ggml_backend_feature * (*ggml_backend_get_features_t)(ggml_backend_reg_t reg); // // Backend registry @@ -214,6 +222,13 @@ extern "C" { // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU) OR ggml_backend_dev_by_type(CPU), NULL) GGML_API ggml_backend_t ggml_backend_init_best(void); + // Load a backend from a dynamic library and register it + GGML_API ggml_backend_reg_t ggml_backend_load(const char * path); + // Unload a backend if loaded dynamically and unregister it + GGML_API void ggml_backend_unload(ggml_backend_reg_t reg); + // Load all known backends from dynamic libraries + GGML_API void ggml_backend_load_all(void); + // // Backend scheduler // diff --git a/ggml/include/ggml-cpu.h b/ggml/include/ggml-cpu.h index 7571ef97983..e14ea9ea530 100644 --- a/ggml/include/ggml-cpu.h +++ b/ggml/include/ggml-cpu.h @@ -7,29 +7,6 @@ extern "C" { #endif - // Scheduling priorities - enum ggml_sched_priority { - GGML_SCHED_PRIO_NORMAL, - GGML_SCHED_PRIO_MEDIUM, - GGML_SCHED_PRIO_HIGH, - GGML_SCHED_PRIO_REALTIME - }; - - // Threadpool params - // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults - struct ggml_threadpool_params { - bool cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings) - int n_threads; // number of threads - enum ggml_sched_priority prio; // thread priority - uint32_t poll; // polling level (0 - no polling, 100 - aggressive polling) - bool strict_cpu; // strict cpu placement - bool paused; // start in paused state - }; - - struct ggml_threadpool; // forward declaration, see ggml.c - - typedef struct ggml_threadpool * ggml_threadpool_t; - // the compute plan that needs to be prepared for ggml_graph_compute() // since https://github.com/ggerganov/ggml/issues/287 struct ggml_cplan { @@ -75,14 +52,11 @@ extern "C" { GGML_BACKEND_API float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3); GGML_BACKEND_API void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value); - GGML_BACKEND_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads); - GGML_BACKEND_API void ggml_threadpool_params_init (struct ggml_threadpool_params * p, int n_threads); - GGML_BACKEND_API bool ggml_threadpool_params_match (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1); - GGML_BACKEND_API struct ggml_threadpool * ggml_threadpool_new (struct ggml_threadpool_params * params); - GGML_BACKEND_API void ggml_threadpool_free (struct ggml_threadpool * threadpool); - GGML_BACKEND_API int ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool); - GGML_BACKEND_API void ggml_threadpool_pause (struct ggml_threadpool * threadpool); - GGML_BACKEND_API void ggml_threadpool_resume (struct ggml_threadpool * threadpool); + GGML_BACKEND_API struct ggml_threadpool * ggml_threadpool_new (struct ggml_threadpool_params * params); + GGML_BACKEND_API void ggml_threadpool_free (struct ggml_threadpool * threadpool); + GGML_BACKEND_API int ggml_threadpool_get_n_threads (struct ggml_threadpool * threadpool); + GGML_BACKEND_API void ggml_threadpool_pause (struct ggml_threadpool * threadpool); + GGML_BACKEND_API void ggml_threadpool_resume (struct ggml_threadpool * threadpool); // ggml_graph_plan() has to be called before ggml_graph_compute() // when plan.work_size > 0, caller must allocate memory for plan.work_data @@ -104,10 +78,10 @@ extern "C" { GGML_BACKEND_API int ggml_cpu_has_sse3 (void); GGML_BACKEND_API int ggml_cpu_has_ssse3 (void); GGML_BACKEND_API int ggml_cpu_has_avx (void); + GGML_BACKEND_API int ggml_cpu_has_avx_vnni (void); GGML_BACKEND_API int ggml_cpu_has_avx2 (void); GGML_BACKEND_API int ggml_cpu_has_f16c (void); GGML_BACKEND_API int ggml_cpu_has_fma (void); - GGML_BACKEND_API int ggml_cpu_has_avx_vnni (void); GGML_BACKEND_API int ggml_cpu_has_avx512 (void); GGML_BACKEND_API int ggml_cpu_has_avx512_vbmi(void); GGML_BACKEND_API int ggml_cpu_has_avx512_vnni(void); @@ -117,6 +91,7 @@ extern "C" { GGML_BACKEND_API int ggml_cpu_has_neon (void); GGML_BACKEND_API int ggml_cpu_has_arm_fma (void); GGML_BACKEND_API int ggml_cpu_has_fp16_va (void); + GGML_BACKEND_API int ggml_cpu_has_dotprod (void); GGML_BACKEND_API int ggml_cpu_has_matmul_int8(void); GGML_BACKEND_API int ggml_cpu_has_sve (void); GGML_BACKEND_API int ggml_cpu_get_sve_cnt (void); // sve vector length in bytes diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 69e6a24344b..1c8cc11b6b2 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -389,6 +389,9 @@ extern "C" { GGML_TYPE_Q4_0_8_8 = 33, GGML_TYPE_TQ1_0 = 34, GGML_TYPE_TQ2_0 = 35, + GGML_TYPE_IQ4_NL_4_4 = 36, + // GGML_TYPE_IQ4_NL_4_8 = 37, + // GGML_TYPE_IQ4_NL_8_8 = 38, GGML_TYPE_COUNT, }; @@ -496,6 +499,7 @@ extern "C" { GGML_OP_POOL_2D_BACK, GGML_OP_UPSCALE, // nearest interpolate GGML_OP_PAD, + GGML_OP_PAD_REFLECT_1D, GGML_OP_ARANGE, GGML_OP_TIMESTEP_EMBEDDING, GGML_OP_ARGSORT, @@ -1692,6 +1696,13 @@ extern "C" { int p2, int p3); + // pad each dimension with reflection: [a, b, c, d] -> [b, a, b, c, d, c] + GGML_API struct ggml_tensor * ggml_pad_reflect_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int p0, + int p1); + // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151 // timesteps: [N,] // return: [N, dim] @@ -2215,6 +2226,37 @@ extern "C" { GGML_API const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type); + // ggml threadpool + // TODO: currently, only a few functions are in the base ggml API, while the rest are in the CPU backend + // the goal should be to create an API that other backends can use move everything to the ggml base + + // scheduling priorities + enum ggml_sched_priority { + GGML_SCHED_PRIO_NORMAL, + GGML_SCHED_PRIO_MEDIUM, + GGML_SCHED_PRIO_HIGH, + GGML_SCHED_PRIO_REALTIME + }; + + // threadpool params + // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults + struct ggml_threadpool_params { + bool cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings) + int n_threads; // number of threads + enum ggml_sched_priority prio; // thread priority + uint32_t poll; // polling level (0 - no polling, 100 - aggressive polling) + bool strict_cpu; // strict cpu placement + bool paused; // start in paused state + }; + + struct ggml_threadpool; // forward declaration, see ggml.c + + typedef struct ggml_threadpool * ggml_threadpool_t; + + GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads); + GGML_API void ggml_threadpool_params_init (struct ggml_threadpool_params * p, int n_threads); + GGML_API bool ggml_threadpool_params_match (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1); + #ifdef __cplusplus } #endif diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index 8df0e85c0d0..f07533fdb31 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -24,7 +24,7 @@ if (NOT MSVC) endif() endif() -function(get_flags CCID CCVER) +function(ggml_get_flags CCID CCVER) set(C_FLAGS "") set(CXX_FLAGS "") @@ -41,6 +41,7 @@ function(get_flags CCID CCVER) elseif (CCID STREQUAL "GNU") set(C_FLAGS -Wdouble-promotion) set(CXX_FLAGS -Wno-array-bounds) + if (CCVER VERSION_GREATER_EQUAL 8.1.0) list(APPEND CXX_FLAGS -Wextra-semi) endif() @@ -69,7 +70,7 @@ if (GGML_ALL_WARNINGS) list(APPEND C_FLAGS ${WARNING_FLAGS}) list(APPEND CXX_FLAGS ${WARNING_FLAGS}) - get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION}) + ggml_get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION}) add_compile_options("$<$:${C_FLAGS};${GF_C_FLAGS}>" "$<$:${CXX_FLAGS};${GF_CXX_FLAGS}>") @@ -202,6 +203,10 @@ endif() # ggml +if (GGML_BACKEND_DL AND NOT BUILD_SHARED_LIBS) + message(FATAL_ERROR "GGML_BACKEND_DL requires BUILD_SHARED_LIBS") +endif() + add_library(ggml-base ../include/ggml.h ../include/ggml-alloc.h @@ -226,44 +231,94 @@ add_library(ggml target_link_libraries(ggml PUBLIC ggml-base) +if (CMAKE_SYSTEM_NAME MATCHES "Linux") + target_link_libraries(ggml PRIVATE dl) +endif() + +function(ggml_add_backend_library backend) + if (GGML_BACKEND_DL) + add_library(${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) + 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() +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}) - # check again in case the backend disabled itself - # note that this should NOT be the normal behavior, in case of errors the backend should fail the build - # however, currently it is necessary for AMX, since it is enabled by default on llama.cpp - if (${backend_id}) - message(STATUS "Including ${backend} backend") - if (${BUILD_SHARED_LIBS}) - target_compile_definitions(${backend_target} PRIVATE GGML_BACKEND_BUILD) - target_compile_definitions(${backend_target} PUBLIC GGML_BACKEND_SHARED) - endif() - install(TARGETS ${backend_target} LIBRARY) - target_link_libraries(ggml PUBLIC ${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 ${tag_name}) + # other: OPENMP LLAMAFILE CPU_HBM + foreach (feat NATIVE + AVX AVX2 AVX_VNNI FMA F16C + AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 + AMX_TILE AMX_INT8 AMX_BF16) + set(GGML_${feat} OFF) + endforeach() + + foreach (feat ${ARGN}) + set(GGML_${feat} ON) + endforeach() + + ggml_add_cpu_backend_variant_impl(${tag_name}) +endfunction() + ggml_add_backend(CPU) -ggml_add_backend(AMX) + +if (GGML_CPU_ALL_VARIANTS) + if (NOT GGML_BACKEND_DL) + message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL") + endif() + ggml_add_cpu_backend_variant(sandybridge AVX) + ggml_add_cpu_backend_variant(haswell AVX F16C AVX2 FMA) + ggml_add_cpu_backend_variant(skylakex AVX F16C AVX2 FMA AVX512) + ggml_add_cpu_backend_variant(icelake AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI) + if (NOT MSVC) + # MSVC doesn't support AVX-VNNI or AMX + ggml_add_cpu_backend_variant(alderlake AVX F16C AVX2 FMA AVX_VNNI) + ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8) + endif() +else () + ggml_add_cpu_backend_variant_impl("") +endif() + ggml_add_backend(BLAS) ggml_add_backend(CANN) ggml_add_backend(CUDA) ggml_add_backend(HIP) ggml_add_backend(Kompute) ggml_add_backend(METAL) +ggml_add_backend(MUSA) ggml_add_backend(RPC) ggml_add_backend(SYCL) ggml_add_backend(Vulkan) -ggml_add_backend(MUSA) foreach (target ggml-base ggml) target_include_directories(${target} PUBLIC $ $) - target_compile_features (${target} PRIVATE c_std_11) # don't bump + target_compile_features (${target} PRIVATE c_std_11 cxx_std_17) # don't bump endforeach() target_link_libraries(ggml-base PRIVATE Threads::Threads) diff --git a/ggml/src/ggml-backend-impl.h b/ggml/src/ggml-backend-impl.h index fa8d5b7fb68..36d72e95f02 100644 --- a/ggml/src/ggml-backend-impl.h +++ b/ggml/src/ggml-backend-impl.h @@ -8,6 +8,8 @@ extern "C" { #endif + #define GGML_BACKEND_API_VERSION 1 + // // Backend buffer type // @@ -63,20 +65,20 @@ extern "C" { enum ggml_backend_buffer_usage usage; }; - ggml_backend_buffer_t ggml_backend_buffer_init( + 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 - bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst); + 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_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers); - bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer); - void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage); + 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 (stream) @@ -199,17 +201,55 @@ extern "C" { }; struct ggml_backend_reg { - // int api_version; // TODO: for dynamic loading + int api_version; // initialize to GGML_BACKEND_API_VERSION struct ggml_backend_reg_i iface; void * context; }; - // Internal backend registry API - void ggml_backend_register(ggml_backend_reg_t reg); - void ggml_backend_device_register(ggml_backend_dev_t device); - // TODO: backends can be loaded as a dynamic library, in which case it needs to export this function - // typedef ggml_backend_register_t * (*ggml_backend_init)(void); + GGML_API void ggml_backend_register(ggml_backend_reg_t reg); + GGML_API void ggml_backend_device_register(ggml_backend_dev_t device); + + // 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 } diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp index 63e9d820174..5cb0fb9d159 100644 --- a/ggml/src/ggml-backend-reg.cpp +++ b/ggml/src/ggml-backend-reg.cpp @@ -1,11 +1,34 @@ #include "ggml-backend-impl.h" #include "ggml-backend.h" -#include "ggml-cpu.h" #include "ggml-impl.h" +#include +#include #include +#include +#include +#include +#include +#include #include +#ifdef _WIN32 +# define WIN32_LEAN_AND_MEAN +# ifndef NOMINMAX +# define NOMINMAX +# endif +# include +#elif defined(__APPLE__) +# include +# include +#else +# include +# include +#endif + // Backend registry +#ifdef GGML_USE_CPU +#include "ggml-cpu.h" +#endif #ifdef GGML_USE_CUDA #include "ggml-cuda.h" @@ -31,10 +54,6 @@ #include "ggml-rpc.h" #endif -#ifdef GGML_USE_AMX -# include "ggml-amx.h" -#endif - #ifdef GGML_USE_CANN #include "ggml-cann.h" #endif @@ -43,8 +62,75 @@ #include "ggml-kompute.h" #endif +#ifdef _WIN32 + +using dl_handle = std::remove_pointer_t; + +struct dl_handle_deleter { + void operator()(HMODULE handle) { + FreeLibrary(handle); + } +}; + +static dl_handle * dl_load_library(const std::wstring & path) { + // suppress error dialogs for missing DLLs + DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS); + SetErrorMode(old_mode | SEM_FAILCRITICALERRORS); + + HMODULE handle = LoadLibraryW(path.c_str()); + + SetErrorMode(old_mode); + + return handle; +} + +static dl_handle * dl_load_library(const std::string & path) { + std::wstring_convert> converter; + return dl_load_library(converter.from_bytes(path)); +} + +static 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; +} + +#else + +using dl_handle = void; + +struct dl_handle_deleter { + void operator()(void * handle) { + dlclose(handle); + } +}; + +static void * dl_load_library(const std::string & path) { + dl_handle * handle = dlopen(path.c_str(), RTLD_NOW | RTLD_LOCAL); + + return handle; +} + +static void * dl_get_sym(dl_handle * handle, const char * name) { + return dlsym(handle, name); +} + +#endif + +using dl_handle_ptr = std::unique_ptr; + +struct ggml_backend_reg_entry { + ggml_backend_reg_t reg; + dl_handle_ptr handle; +}; + struct ggml_backend_registry { - std::vector backends; + std::vector backends; std::vector devices; ggml_backend_registry() { @@ -69,17 +155,25 @@ struct ggml_backend_registry { #ifdef GGML_USE_RPC register_backend(ggml_backend_rpc_reg()); #endif -#ifdef GGML_USE_AMX - register_backend(ggml_backend_amx_reg()); -#endif #ifdef GGML_USE_KOMPUTE register_backend(ggml_backend_kompute_reg()); #endif - +#ifdef GGML_USE_CPU register_backend(ggml_backend_cpu_reg()); +#endif } - void register_backend(ggml_backend_reg_t reg) { + ~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; } @@ -88,7 +182,7 @@ struct ggml_backend_registry { 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); + 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)); } @@ -100,6 +194,76 @@ struct ggml_backend_registry { #endif devices.push_back(device); } + + ggml_backend_reg_t load_backend(const char * path, bool silent) { + dl_handle_ptr handle { dl_load_library(path) }; + if (!handle) { + if (!silent) { + GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path); + } + 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); + } + 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); + } + 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); + } else { + GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n", + __func__, path, 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); + + 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() { @@ -117,23 +281,32 @@ void ggml_backend_device_register(ggml_backend_dev_t 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]; + 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 (std::strcmp(ggml_backend_reg_name(reg), name) == 0) { + if (striequals(ggml_backend_reg_name(reg), name)) { return reg; } } - return NULL; + return nullptr; } // Device enumeration @@ -149,11 +322,11 @@ ggml_backend_dev_t ggml_backend_dev_get(size_t 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 (strcmp(ggml_backend_dev_name(dev), name) == 0) { + if (striequals(ggml_backend_dev_name(dev), name)) { return dev; } } - return NULL; + return nullptr; } ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) { @@ -163,14 +336,14 @@ ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) { return dev; } } - return NULL; + 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 NULL; + return nullptr; } return ggml_backend_dev_init(dev, params); } @@ -178,7 +351,7 @@ ggml_backend_t ggml_backend_init_by_name(const char * name, const char * 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 NULL; + return nullptr; } return ggml_backend_dev_init(dev, params); } @@ -189,7 +362,168 @@ ggml_backend_t ggml_backend_init_best(void) { dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); } if (!dev) { - return NULL; + 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 std::string get_executable_path() { +#if defined(__APPLE__) + // get executable path + std::vector 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__) + std::string base_path = "."; + std::vector path(1024); + while (true) { + // get executable path + ssize_t len = readlink("/proc/self/exe", path.data(), path.size()); + 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 ggml_backend_dev_init(dev, NULL); + + return base_path + "/"; +#elif defined(_WIN32) + std::vector path(MAX_PATH); + DWORD len = GetModuleFileNameA(NULL, path.data(), path.size()); + if (len == 0) { + return ""; + } + std::string 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 + "\\"; +#endif +} + +static std::string backend_filename_prefix() { +#ifdef _WIN32 + return "ggml-"; +#else + return "libggml-"; +#endif +} + +static std::string backend_filename_suffix() { +#ifdef _WIN32 + return ".dll"; +#else + return ".so"; +#endif +} + +static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent) { + // enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths + // TODO: search system paths + std::vector search_paths = { "./", get_executable_path() }; + std::string file_prefix = backend_filename_prefix() + name + "-"; + + int best_score = 0; + std::string best_path; + + namespace fs = std::filesystem; + for (const auto & search_path : search_paths) { + if (!fs::exists(search_path)) { + continue; + } + for (const auto & entry : fs::directory_iterator(search_path)) { + if (entry.is_regular_file()) { + std::string filename = entry.path().filename().string(); + std::string ext = entry.path().extension().string(); + if (filename.find(file_prefix) == 0 && ext == backend_filename_suffix()) { + dl_handle_ptr handle { dl_load_library(entry.path().c_str()) }; + if (!handle && !silent) { + GGML_LOG_ERROR("%s: failed to load %s\n", __func__, entry.path().string().c_str()); + } + 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__, entry.path().string().c_str(), s); +#endif + if (s > best_score) { + best_score = s; + best_path = entry.path().string(); + } + } else { + if (!silent) { + GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, entry.path().string().c_str()); + } + } + } + } + } + } + } + + if (best_score == 0) { + // try to load the base backend + for (const auto & search_path : search_paths) { + std::string path = search_path + backend_filename_prefix() + name + backend_filename_suffix(); + if (fs::exists(path)) { + return get_reg().load_backend(path.c_str(), silent); + } + } + return nullptr; + } + + return get_reg().load_backend(best_path.c_str(), silent); +} + +void ggml_backend_load_all() { +#ifdef NDEBUG + bool silent = true; +#else + bool silent = false; +#endif + + ggml_backend_load_best("blas", silent); + ggml_backend_load_best("cann", silent); + ggml_backend_load_best("cuda", silent); + ggml_backend_load_best("hip", silent); + ggml_backend_load_best("kompute", silent); + ggml_backend_load_best("metal", silent); + ggml_backend_load_best("rpc", silent); + ggml_backend_load_best("sycl", silent); + ggml_backend_load_best("vulkan", silent); + ggml_backend_load_best("musa", silent); + ggml_backend_load_best("cpu", silent); } diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp index fd8d0dc0dc9..fdb4b986f61 100644 --- a/ggml/src/ggml-backend.cpp +++ b/ggml/src/ggml-backend.cpp @@ -252,6 +252,7 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten } void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { + GGML_ASSERT(tensor); ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; if (size == 0) { @@ -266,6 +267,7 @@ void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, siz } void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { + GGML_ASSERT(tensor); ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; if (size == 0) { @@ -740,7 +742,8 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st if (tensor->buffer || (tensor->view_src && tensor->view_src->buffer)) { // since the tensor is pre-allocated, it cannot be moved to another backend - GGML_ABORT("pre-allocated tensor (%s) in a backend that cannot run the operation", tensor->name); + ggml_backend_buffer_t buffer = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; + GGML_ABORT("pre-allocated tensor (%s) in a buffer (%s) that cannot run the operation (%s)", tensor->name, ggml_backend_buffer_name(buffer), ggml_op_name(tensor->op)); } // graph input diff --git a/ggml/src/ggml-blas/CMakeLists.txt b/ggml/src/ggml-blas/CMakeLists.txt index e2cbabf0dae..0bf3c05d93a 100644 --- a/ggml/src/ggml-blas/CMakeLists.txt +++ b/ggml/src/ggml-blas/CMakeLists.txt @@ -11,12 +11,9 @@ find_package(BLAS) if (BLAS_FOUND) message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}") - add_library(ggml-blas - ggml-blas.cpp - ) - - target_link_libraries(ggml-blas PRIVATE ggml-base) - target_include_directories(ggml-blas PRIVATE . ..) + ggml_add_backend_library(ggml-blas + ggml-blas.cpp + ) if (${GGML_BLAS_VENDOR} MATCHES "Apple") add_compile_definitions(ACCELERATE_NEW_LAPACK) diff --git a/ggml/src/ggml-blas/ggml-blas.cpp b/ggml/src/ggml-blas/ggml-blas.cpp index 648c9d875e3..ec158dfac6e 100644 --- a/ggml/src/ggml-blas/ggml-blas.cpp +++ b/ggml/src/ggml-blas/ggml-blas.cpp @@ -506,9 +506,12 @@ static const struct ggml_backend_reg_i ggml_backend_blas_reg_i = { ggml_backend_reg_t ggml_backend_blas_reg(void) { static struct ggml_backend_reg ggml_backend_blas_reg = { - /* .iface = */ ggml_backend_blas_reg_i, - /* .context = */ NULL, + /* .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) diff --git a/ggml/src/ggml-cann/CMakeLists.txt b/ggml/src/ggml-cann/CMakeLists.txt index c8e15c6d40c..05cf06bfab4 100644 --- a/ggml/src/ggml-cann/CMakeLists.txt +++ b/ggml/src/ggml-cann/CMakeLists.txt @@ -3,6 +3,34 @@ if ("cann${CANN_INSTALL_DIR}" STREQUAL "cann" AND DEFINED ENV{ASCEND_TOOLKIT_HOM message(STATUS "CANN: updated CANN_INSTALL_DIR from ASCEND_TOOLKIT_HOME=$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) + if (CANN_INSTALL_DIR) # Only Support Linux. if (NOT UNIX) @@ -34,11 +62,13 @@ if (CANN_INSTALL_DIR) file(GLOB GGML_SOURCES_CANN "*.cpp") - add_library(ggml-cann ${GGML_SOURCES_CANN}) - target_link_libraries(ggml-cann PRIVATE ggml-base ${CANN_LIBRARIES}) - target_include_directories(ggml-cann PRIVATE . .. ${CANN_INCLUDE_DIRS}) + 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}") + message(STATUS "CANN: CANN_INCLUDE_DIRS = ${CANN_INCLUDE_DIRS}") message(STATUS "CANN: CANN_LIBRARIES = ${CANN_LIBRARIES}") else() diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp index a4ec8418e2a..b2d857e1e54 100644 --- a/ggml/src/ggml-cann/aclnn_ops.cpp +++ b/ggml/src/ggml-cann/aclnn_ops.cpp @@ -22,11 +22,14 @@ #include "aclnn_ops.h" +#include #include +#include #include #include #include #include +#include #include #include #include @@ -34,6 +37,7 @@ #include #include #include +#include #include #include #include @@ -53,6 +57,7 @@ #include #include +#include "ggml-impl.h" #include "kernels/ascendc_kernels.h" #define GGML_COMMON_DECL_C @@ -241,10 +246,14 @@ void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst) { aclTensor* acl_src1 = ggml_cann_create_tensor(src1); aclTensor* acl_dst = ggml_cann_create_tensor(dst); - int64_t concat_dim = 1; + const int32_t dim = ggml_get_op_params_i32(dst, 0); + + GGML_ASSERT(dim >= 0 && dim < 4); + int32_t acl_dim = 3 - dim; + aclTensor* tensors[] = {acl_src0, acl_src1}; aclTensorList* tensorList = aclCreateTensorList(tensors, 2); - aclnn_concat(ctx, tensorList, acl_dst, concat_dim); + aclnn_concat(ctx, tensorList, acl_dst, acl_dim); ACL_CHECK(aclDestroyTensorList(tensorList)); ACL_CHECK(aclDestroyTensor(acl_dst)); @@ -1096,9 +1105,9 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer, } /** - * @brief Creates an ACL tensor initialized with ones using a provided buffer. + * @brief Creates an ACL tensor initialized with value using a provided buffer. * - * This function initializes a tensor with ones using the specified buffer and + * This function initializes a tensor with value using the specified buffer and * tensor parameters. * * @param ctx The context for the CANN backend operations. @@ -1111,12 +1120,12 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer, * @param type_size The size of each element in the tensor data type. * @param value The value to be used for initializing the tensor (default * is 1.0). - * @return An ACL tensor initialized with ones. + * @return An ACL tensor initialized with value. */ -static aclTensor* aclnn_ones(ggml_backend_cann_context& ctx, void* buffer, - size_t n_bytes, int64_t* ne, int64_t dims, - aclDataType type, size_t type_size, - float value = 1.0f) { +static aclTensor* aclnn_values(ggml_backend_cann_context& ctx, void* buffer, + size_t n_bytes, int64_t* ne, int64_t dims, + aclDataType type, size_t type_size, + float value = 1.0f) { aclTensor* acl_tensor = aclnn_zero(ctx, buffer, n_bytes, ne, dims, type, type_size); float alpha_host = 1.0f; @@ -1158,7 +1167,7 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) { size_t one_tensor_n_bytes = src->ne[0] * ggml_element_size(src); ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes); - aclTensor* acl_gamma = aclnn_ones( + aclTensor* acl_gamma = aclnn_values( ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne, 1, ggml_cann_type_mapping(src->type), ggml_element_size(src)); @@ -1202,9 +1211,9 @@ void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst, ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes); aclTensor* mask_tensor = - aclnn_ones(ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne, - GGML_MAX_DIMS, ggml_cann_type_mapping(src->type), - ggml_element_size(src), value); + aclnn_values(ctx, one_tensor_allocator.get(), one_tensor_n_bytes, + src->ne, GGML_MAX_DIMS, ggml_cann_type_mapping(src->type), + ggml_element_size(src), value); uint64_t workspaceSize = 0; aclOpExecutor* executor; @@ -1437,10 +1446,6 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ggml_tensor* src0 = dst->src[0]; // kernel ggml_tensor* src1 = dst->src[1]; // input - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32); - GGML_TENSOR_BINARY_OP_LOCALS; // aclnnIm2col only works on 2D. set s1, p1, d1 to 1 to perform 2D @@ -1462,9 +1467,6 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) { const int64_t OH = is_2D ? ne2 : 1; const int64_t OW = ne1; - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(float)); - // memory allocated increased to 3x when is_2D == false const int64_t n_bytes_factor = is_2D ? 1 : 3; @@ -1768,6 +1770,92 @@ static void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src, ACL_CHECK(aclnnSin(workspaceAddr, workspaceSize, executor, ctx.stream())); } +/** + * @brief Performs element-wise division of tensor1 by tensor2 , multiplies the + result by the scalar value and adds it to self . + * + * Performs element-wise division of tensor1 by tensor2, + * multiplies the result by the scalar value and adds it to self . + * The operation is defined as: + * \f[ + * \text{out}_i = \text{selft}_i + \text{value} \times + \frac{\text{tensor1}_i}{\text{tensor2}_i} + * \f] + + * @param ctx The context for the CANN backend operations. + * @param acl_self The source tensor on which the addcdiv function will be + applied. + * @param tensor1 Numerator tensor. + * @param tensor2 Denominator tensor. + * @param value The value to be used for coefficient. + */ +static void aclnn_inplace_addcdiv(ggml_backend_cann_context& ctx, + aclTensor* acl_self, aclTensor* tensor1, + aclTensor* tensor2, float value) { + uint64_t workspaceSize = 0; + aclOpExecutor* executor; + void* workspaceAddr = nullptr; + aclScalar* acl_value = aclCreateScalar(&value, aclDataType::ACL_FLOAT); + + ACL_CHECK(aclnnInplaceAddcdivGetWorkspaceSize( + acl_self, tensor1, tensor2, acl_value, &workspaceSize, &executor)); + if (workspaceSize > 0) { + ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); + workspaceAddr = workspace_allocator.get(); + } + + ACL_CHECK(aclnnInplaceAddcdiv(workspaceAddr, workspaceSize, executor, + ctx.stream())); +} + +/** + * @brief Matrix division, optionally in-place. + * + * This function division each element of the source tensor `acl_src` by the + * tensor `acl_other` and stores the result in the destination tensor `acl_dst`. + * If `inplace` is true, `acl_dst` will not be used and the operation is + * performed in-place on `acl_src`. The operation is defined as: \f[ + * \text{dst}_i = \frac{\text{acl_src}_i}{\text{acl_other}_i} + * \f] + * + * @param ctx The context for the CANN backend operations. + * @param acl_src Numerator tensor.. + * @param acl_other Denominator tensor. + * @param acl_dst The destination tensor where the result will be stored if + * `inplace` is false. + * @param inplace Flag indicating whether to perform the operation in-place on + * `acl_src`. + */ +static void aclnn_div_tensor(ggml_backend_cann_context& ctx, aclTensor* acl_src, + aclTensor* acl_other, aclTensor* acl_dst, + bool inplace) { + uint64_t workspaceSize = 0; + aclOpExecutor* executor; + void* workspaceAddr = nullptr; + + if (inplace) { + ACL_CHECK(aclnnInplaceDivGetWorkspaceSize(acl_src, acl_other, + &workspaceSize, &executor)); + if (workspaceSize > 0) { + ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); + workspaceAddr = workspace_allocator.get(); + } + + ACL_CHECK(aclnnInplaceDiv(workspaceAddr, workspaceSize, executor, + ctx.stream())); + } else { + ACL_CHECK(aclnnDivGetWorkspaceSize(acl_src, acl_other, acl_dst, + &workspaceSize, &executor)); + if (workspaceSize > 0) { + ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); + workspaceAddr = workspace_allocator.get(); + } + + ACL_CHECK( + aclnnDiv(workspaceAddr, workspaceSize, executor, ctx.stream())); + } +} + void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx, ggml_tensor* dst) { const ggml_tensor* src = dst->src[0]; @@ -2311,7 +2399,16 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ctx.stream())); switch (src0->type) { - case GGML_TYPE_F32: + case GGML_TYPE_F32: { +#ifdef ASCEND_310P + // Special operation for get_row_f32 kernel of 310P: clear the + // content of dest data buffer when row is not aligned to 32 bytes + if ((src0->ne[0] % 8) != 0) { + size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] * + src0->ne[0] * ggml_type_size(GGML_TYPE_F32); + ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len)); + } +#endif aclrtlaunch_ascendc_get_row_f32( 24, ctx.stream(), src0->data, src1->data, dst->data, ((ggml_tensor*)src0->extra)->ne, @@ -2320,7 +2417,19 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne, ((ggml_tensor*)dst->extra)->nb); break; - case GGML_TYPE_F16: + } + case GGML_TYPE_F16: { +#ifdef ASCEND_310P + // Special operation for get_row_f16 kernel of 310P: clear the + // content of dest data buffer when row is not aligned to 32 bytes + if ((src0->ne[0] % 16) != 0) { + size_t dst_len = + src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] * + ggml_type_size( + GGML_TYPE_F32); // out is also f32, even input is f16 + ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len)); + } +#endif aclrtlaunch_ascendc_get_row_f16( 24, ctx.stream(), src0->data, src1->data, dst->data, ((ggml_tensor*)src0->extra)->ne, @@ -2329,6 +2438,7 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne, ((ggml_tensor*)dst->extra)->nb); break; + } case GGML_TYPE_Q4_0: aclrtlaunch_ascendc_get_row_q4_0( 24, ctx.stream(), src0->data, src1->data, dst->data, @@ -2407,7 +2517,6 @@ static void aclnn_mat_mul(ggml_backend_cann_context& ctx, aclTensor* acl_input, aclTensor* acl_weight, aclTensor* acl_dst) { int8_t cube_math_type = 1; // ALLOW_FP32_DOWN_PRECISION, when input is // fp32, atlas a2 will transpose it to HFLOAT32. - uint64_t workspaceSize = 0; aclOpExecutor* executor; void* workspaceAddr = nullptr; @@ -2425,6 +2534,81 @@ static void aclnn_mat_mul(ggml_backend_cann_context& ctx, aclTensor* acl_input, aclnnMatmul(workspaceAddr, workspaceSize, executor, ctx.stream())); } +/** + * @brief Performs matrix multiplication of two 2D tensors. + * + * This function computes the matrix multiplication of the input tensor + * `acl_input` and the weight tensor `acl_weight`, and stores the result in the + * destination tensor `acl_dst`. + * The operation is defined as: + * \f[ + * \text {acl_dst}=\text {acl_input@acl_weight} + * \f] + * + * @param ctx The context for the CANN backend operations. + * @param acl_input The input tensor for the matrix multiplication. + * @param acl_weight The weight tensor for the matrix multiplication. + * @param acl_dst The destination tensor where the result of the matrix + * multiplication will be stored. + */ +static void aclnn_mat_mul_2d(ggml_backend_cann_context& ctx, + aclTensor* acl_input, aclTensor* acl_weight, + aclTensor* acl_dst) { + int8_t cube_math_type = 2; + uint64_t workspaceSize = 0; + aclOpExecutor* executor; + void* workspaceAddr = nullptr; + + ACL_CHECK(aclnnMmGetWorkspaceSize(acl_input, acl_weight, acl_dst, + cube_math_type, &workspaceSize, + &executor)); + + if (workspaceSize > 0) { + ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); + workspaceAddr = workspace_allocator.get(); + } + + ACL_CHECK(aclnnMm(workspaceAddr, workspaceSize, executor, ctx.stream())); +} + +/** + * @brief Performs matrix multiplication of two 3D tensors. + * + * This function computes the matrix multiplication of the input tensor + * `acl_input` and the weight tensor `acl_weight`, and stores the result in the + * destination tensor `acl_dst`. + * The operation is defined as: + * \f[ + * \text {acl_dst}=\text {acl_input@acl_weight} + * \f] + * + * @param ctx The context for the CANN backend operations. + * @param acl_input The input tensor for the matrix multiplication. + * @param acl_weight The weight tensor for the matrix multiplication. + * @param acl_dst The destination tensor where the result of the matrix + * multiplication will be stored. + */ +static void aclnn_mat_mul_3d(ggml_backend_cann_context& ctx, + aclTensor* acl_input, aclTensor* acl_weight, + aclTensor* acl_dst) { + int8_t cube_math_type = 2; + uint64_t workspaceSize = 0; + aclOpExecutor* executor; + void* workspaceAddr = nullptr; + + ACL_CHECK(aclnnBatchMatMulGetWorkspaceSize(acl_input, acl_weight, acl_dst, + cube_math_type, &workspaceSize, + &executor)); + + if (workspaceSize > 0) { + ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); + workspaceAddr = workspace_allocator.get(); + } + + ACL_CHECK( + aclnnBatchMatMul(workspaceAddr, workspaceSize, executor, ctx.stream())); +} + /** * @brief Performs matrix multiplication with floating-point precision on * tensors using the CANN backend. @@ -2446,20 +2630,39 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx, // broadcast, when weight ne2 or ne3 is not 1, weight need repeat. BCAST_MUL_MAT_SHAPE(input, weight, dst); - // transpose weight: [1,2,3,4] -> [1,2,4,3] + int64_t n_dims = bcast_dims; + if (bcast_input_ne[3] == bcast_weight_ne[3] && bcast_input_ne[3] == 1) { + if (bcast_input_ne[2] == 1 && bcast_weight_ne[2] == 1) { + n_dims = 2; + } else if (bcast_input_ne[2] == 1) { + n_dims = 3; + } + } + + aclTensor* acl_input_tensor = + ggml_cann_create_tensor(input, bcast_input_ne, bcast_input_nb, n_dims); int64_t transpose_ne[] = {bcast_weight_ne[1], bcast_weight_ne[0], bcast_weight_ne[2], bcast_weight_ne[3], bcast_weight_ne[4], bcast_weight_ne[5]}; size_t transpose_nb[] = {bcast_weight_nb[1], bcast_weight_nb[0], bcast_weight_nb[2], bcast_weight_nb[3], bcast_weight_nb[4], bcast_weight_nb[5]}; - aclTensor* acl_weight_tensor = - ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, bcast_dims); - aclTensor* acl_input_tensor = - ggml_cann_create_tensor(input, BCAST_MUL_MAT_PARAM(input)); - aclTensor* acl_dst = ggml_cann_create_tensor(dst, BCAST_MUL_MAT_PARAM(dst)); - aclnn_mat_mul(ctx, acl_input_tensor, acl_weight_tensor, acl_dst); + ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims); + aclTensor* acl_dst = + ggml_cann_create_tensor(dst, bcast_dst_ne, bcast_dst_nb, n_dims); + + switch (n_dims) { + case 2: + aclnn_mat_mul_2d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst); + break; + case 3: + aclnn_mat_mul_3d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst); + break; + default: + aclnn_mat_mul(ctx, acl_input_tensor, acl_weight_tensor, acl_dst); + break; + } ACL_CHECK(aclDestroyTensor(acl_weight_tensor)); ACL_CHECK(aclDestroyTensor(acl_input_tensor)); @@ -2480,51 +2683,47 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx, * multiplication will be stored. */ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx, - ggml_tensor* dst, - const enum ggml_type type) { + ggml_tensor* dst, + const enum ggml_type type) { ggml_tensor* src0 = dst->src[0]; // weight ggml_tensor* src1 = dst->src[1]; // input - // The shape of the weight is NCHW. Matrix multiplication uses HW dims. HC - // is regarded as batch. weight need transpose. - int64_t weight_ne[] = {src0->ne[1], src0->ne[0]}; + // The shape of the weight is NCHW. + // Matrix multiplication uses HW dims. + // HC is regarded as batch. + // weight need transpose. float weight_elem_size; if (type == GGML_TYPE_Q4_0) { weight_elem_size = float(sizeof(uint8_t)) / 2; - } - else if (type == GGML_TYPE_Q8_0) { + } else if (type == GGML_TYPE_Q8_0) { weight_elem_size = float(sizeof(uint8_t)); - } - else { + } else { GGML_ABORT("Only support Q4_0 and Q8_0 MUL_MAT"); } - float weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size}; - - // size of one matrix is element_size * height * width. - size_t weight_stride = weight_elem_size * src0->ne[0] * src0->ne[1]; + float weight_nb[] = {src0->ne[0] * weight_elem_size, weight_elem_size}; + size_t weight_stride = src0->ne[1] * src0->ne[0] * weight_elem_size; size_t weight_size = weight_stride * src0->ne[2] * src0->ne[3]; // scale stored at the end of weight. Also need transpose. - GGML_ASSERT(QK4_0 == QK8_0); - int64_t scale_ne[] = {src0->ne[1], src0->ne[0] / QK8_0}; size_t scale_elem_size = sizeof(uint16_t); size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size, scale_elem_size}; - size_t scale_stride = scale_elem_size * src0->ne[0] * src0->ne[1] / QK8_0; + size_t scale_stride = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size; char* scale_offset = (char*)src0->data + weight_size; // input - void* input_buffer; size_t input_elem_size = sizeof(uint16_t); int64_t input_ne[] = {src1->ne[0], src1->ne[1]}; - size_t input_nb[] = {input_elem_size, input_elem_size * src1->ne[0]}; - size_t input_stride = input_elem_size * src1->ne[0] * src1->ne[1]; - + size_t input_nb[] = {input_elem_size, input_ne[0] * input_elem_size}; + size_t input_stride = input_ne[0] * input_ne[1] * input_elem_size; ggml_cann_pool_alloc input_alloctor(ctx.pool()); + void* input_buffer = src1->data; + + // case in if (src1->type != GGML_TYPE_F16) { aclTensor* acl_src1_tensor = ggml_cann_create_tensor(src1); - input_alloctor.alloc(ggml_nelements(src1) * input_elem_size); - input_buffer = input_alloctor.get(); + input_buffer = + input_alloctor.alloc(ggml_nelements(src1) * input_elem_size); int64_t* input_cast_ne = src1->ne; size_t input_cast_nb[GGML_MAX_DIMS]; @@ -2537,85 +2736,136 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx, input_buffer, ACL_FLOAT16, input_elem_size, input_cast_ne, input_cast_nb, GGML_MAX_DIMS); aclnn_cast(ctx, acl_src1_tensor, acl_input_tensor, ACL_FLOAT16); + ACL_CHECK(aclDestroyTensor(acl_input_tensor)); ACL_CHECK(aclDestroyTensor(acl_src1_tensor)); - } else { - input_buffer = src1->data; } // output size_t output_elem_size = sizeof(uint16_t); - int64_t output_ne[] = {dst->ne[0], dst->ne[1]}; - size_t output_nb[] = {output_elem_size, output_elem_size * dst->ne[0]}; - ggml_cann_pool_alloc output_alloctor( - ctx.pool(), ggml_nelements(dst) * output_elem_size); - void* output_buffer = output_alloctor.get(); - size_t output_stride = output_elem_size * dst->ne[0] * dst->ne[1]; + size_t output_nb[] = {output_elem_size, dst->ne[0] * output_elem_size}; + ggml_cann_pool_alloc output_allocator(ctx.pool()); + void* output_buffer = + output_allocator.alloc(ggml_nelements(dst) * output_elem_size); + size_t output_stride = dst->ne[0] * dst->ne[1] * output_elem_size; // aclnn + int64_t max_elem_size = 65535; + int64_t split_size = (src0->ne[1] / max_elem_size) + 1; + ggml_cann_pool_alloc workspace_allocator(ctx.pool()); + aclOpExecutor* executor = nullptr; uint64_t workspaceSize = 0; - aclOpExecutor* executor; void* workspaceAddr = nullptr; - for (int64_t n1 = 0; n1 < src1->ne[3]; n1++) { for (int64_t c1 = 0; c1 < src1->ne[2]; c1++) { int64_t n0 = n1 / (src1->ne[3] / src0->ne[3]); int64_t c0 = c1 / (src1->ne[2] / src0->ne[2]); - int64_t batch1 = n1 * src1->ne[2] + c1; - int64_t batch0 = n0 * src0->ne[2] + c0; + int64_t batch1 = (n1 * src1->ne[2]) + c1; + int64_t batch0 = (n0 * src0->ne[2]) + c0; aclTensor* acl_input_tensor = ggml_cann_create_tensor( (char*)input_buffer + batch1 * input_stride, ACL_FLOAT16, input_elem_size, input_ne, input_nb, 2); + + // first split + int64_t weight_ne_offset = 0; + int64_t weight_ne[2] = { + max_elem_size > src0->ne[1] ? src0->ne[1] : max_elem_size, + src0->ne[0]}; + int64_t scale_ne_offset = 0; + int64_t scale_ne[2] = {weight_ne[0], weight_ne[1] / QK8_0}; + int64_t output_ne_offset = 0; + int64_t output_ne[2] = {weight_ne[0], dst->ne[1]}; + aclTensor* acl_weight_tensor = ggml_cann_create_tensor( (char*)src0->data + batch0 * weight_stride, ggml_cann_type_mapping(type), weight_elem_size, weight_ne, - weight_nb, 2); + weight_nb, 2, ACL_FORMAT_ND, weight_ne_offset); aclTensor* acl_scale_tensor = ggml_cann_create_tensor( scale_offset + batch0 * scale_stride, ACL_FLOAT16, - scale_elem_size, scale_ne, scale_nb, 2); + scale_elem_size, scale_ne, scale_nb, 2, ACL_FORMAT_ND, + scale_ne_offset); aclTensor* acl_output_tensor = ggml_cann_create_tensor( (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16, - output_elem_size, output_ne, output_nb, 2); + output_elem_size, output_ne, output_nb, 2, ACL_FORMAT_ND, + output_ne_offset); ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize( acl_input_tensor, acl_weight_tensor, acl_scale_tensor, nullptr, nullptr, nullptr, nullptr, QK8_0, acl_output_tensor, &workspaceSize, &executor)); - - if (workspaceSize > 0 && workspaceAddr == nullptr) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), - workspaceSize); - workspaceAddr = workspace_allocator.get(); + if (workspaceAddr == nullptr) { + workspaceAddr = workspace_allocator.alloc(workspaceSize); } - ACL_CHECK(aclnnWeightQuantBatchMatmulV2( workspaceAddr, workspaceSize, executor, ctx.stream())); - ACL_CHECK(aclDestroyTensor(acl_input_tensor)); ACL_CHECK(aclDestroyTensor(acl_weight_tensor)); ACL_CHECK(aclDestroyTensor(acl_scale_tensor)); ACL_CHECK(aclDestroyTensor(acl_output_tensor)); + + // other splits + for (int64_t split = 1; split < split_size; split++) { + weight_ne_offset += + weight_elem_size * weight_ne[0] * weight_ne[1]; + weight_ne[0] = max_elem_size * (split + 1) > src0->ne[1] + ? src0->ne[1] - (max_elem_size * split) + : max_elem_size; + scale_ne_offset += scale_elem_size * scale_ne[0] * scale_ne[1]; + scale_ne[0] = weight_ne[0]; + output_ne_offset += + output_elem_size * output_ne[0] * output_ne[1]; + output_ne[0] = weight_ne[0]; + + acl_weight_tensor = ggml_cann_create_tensor( + (char*)src0->data + batch0 * weight_stride, + ggml_cann_type_mapping(type), weight_elem_size, weight_ne, + weight_nb, 2, ACL_FORMAT_ND, weight_ne_offset); + acl_scale_tensor = ggml_cann_create_tensor( + scale_offset + batch0 * scale_stride, ACL_FLOAT16, + scale_elem_size, scale_ne, scale_nb, 2, ACL_FORMAT_ND, + scale_ne_offset); + acl_output_tensor = ggml_cann_create_tensor( + (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16, + output_elem_size, output_ne, output_nb, 2, ACL_FORMAT_ND, + output_ne_offset); + + ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize( + acl_input_tensor, acl_weight_tensor, acl_scale_tensor, + nullptr, nullptr, nullptr, nullptr, QK8_0, + acl_output_tensor, &workspaceSize, &executor)); + ACL_CHECK(aclnnWeightQuantBatchMatmulV2( + workspaceAddr, workspaceSize, executor, ctx.stream())); + + ACL_CHECK(aclDestroyTensor(acl_weight_tensor)); + ACL_CHECK(aclDestroyTensor(acl_scale_tensor)); + ACL_CHECK(aclDestroyTensor(acl_output_tensor)); + } + + ACL_CHECK(aclDestroyTensor(acl_input_tensor)); } } // cast out - int64_t* output_cast_ne = dst->ne; - size_t output_cast_nb[GGML_MAX_DIMS]; - output_cast_nb[0] = sizeof(uint16_t); - for (int i = 1; i < GGML_MAX_DIMS; i++) { - output_cast_nb[i] = output_cast_nb[i - 1] * output_cast_ne[i - 1]; - } + if (dst->type != GGML_TYPE_F16) { + int64_t* output_cast_ne = dst->ne; + size_t output_cast_nb[GGML_MAX_DIMS]; + output_cast_nb[0] = sizeof(uint16_t); + for (int i = 1; i < GGML_MAX_DIMS; i++) { + output_cast_nb[i] = output_cast_nb[i - 1] * output_cast_ne[i - 1]; + } - aclTensor* acl_output_tensor = - ggml_cann_create_tensor(output_buffer, ACL_FLOAT16, output_elem_size, - output_cast_ne, output_cast_nb, GGML_MAX_DIMS); - aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst); - aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor, ACL_FLOAT); + aclTensor* acl_output_tensor = ggml_cann_create_tensor( + output_buffer, ACL_FLOAT16, output_elem_size, output_cast_ne, + output_cast_nb, GGML_MAX_DIMS); + aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst); + aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor, + ggml_cann_type_mapping(dst->type)); - ACL_CHECK(aclDestroyTensor(acl_output_tensor)); - ACL_CHECK(aclDestroyTensor(acl_dst_tensor)); + ACL_CHECK(aclDestroyTensor(acl_output_tensor)); + ACL_CHECK(aclDestroyTensor(acl_dst_tensor)); + } } void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst) { @@ -2714,12 +2964,14 @@ static void aclnn_index_fill_tensor(ggml_backend_cann_context& ctx, static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst, aclTensor* acl_cos_repeat_tensor, aclTensor* acl_sin_repeat_tensor, - float theta_scale, bool is_neox) { + float theta_scale, float freq_scale, + float attn_factor, bool is_neox) { // int sin/cos cache, cache has different repeat method depond on // @param.is_neox ggml_tensor* src0 = dst->src[0]; // input ggml_tensor* src1 = dst->src[1]; // position + ggml_tensor* src2 = dst->src[2]; // freq_factors // arange, [0,1,...,ne0/2] int64_t arange_length = src0->ne[0] / 2; @@ -2748,11 +3000,26 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst, ggml_cann_pool_alloc theta_scale_allocator(ctx.pool(), arange_length * sizeof(float_t)); void* theta_scale_buffer = theta_scale_allocator.get(); - aclTensor* acl_theta_scale_tensor = aclnn_ones( + aclTensor* acl_theta_scale_tensor = aclnn_values( ctx, theta_scale_buffer, arange_length * sizeof(float_t), arange_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), theta_scale); aclnn_pow_tensor_tensor(ctx, acl_theta_scale_tensor, acl_arange_tensor); + // freq_scale + if (freq_scale != 1) { + aclnn_muls(ctx, acl_theta_scale_tensor, freq_scale, nullptr, true); + } + + // freq_factors + if (src2) { + aclTensor* acl_freq_factors_tensor = ggml_cann_create_tensor( + src2->data, ggml_cann_type_mapping(src2->type), + ggml_type_size(src2->type), arange_ne, arange_nb, GGML_MAX_DIMS); + aclnn_div_tensor(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor, + nullptr, true); + ACL_CHECK(aclDestroyTensor(acl_freq_factors_tensor)); + } + // position GGML_ASSERT(src1->type == GGML_TYPE_I32); int64_t position_length = src1->ne[0]; @@ -2816,6 +3083,12 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst, GGML_MAX_DIMS, ACL_FORMAT_ND); aclnn_cos(ctx, acl_permute_tensor, acl_cos_tensor); + // attn_factor + if (attn_factor != 1) { + aclnn_muls(ctx, acl_sin_tensor, attn_factor, nullptr, true); + aclnn_muls(ctx, acl_cos_tensor, attn_factor, nullptr, true); + } + // repeat if (is_neox) { int64_t repeatsArray[] = {1, 1, 1, 2}; @@ -2841,15 +3114,27 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst, ACL_CHECK(aclDestroyTensor(acl_cos_tensor)); } +#ifdef __cplusplus +extern "C" { +#endif +aclnnStatus aclnnRotaryPositionEmbeddingGetWorkspaceSize( + const aclTensor* x, const aclTensor* cos, const aclTensor* sin, + int64_t mode, const aclTensor* yOut, uint64_t* workspaceSize, + aclOpExecutor** executor); +aclnnStatus aclnnRotaryPositionEmbedding(void* workspace, + uint64_t workspaceSize, + aclOpExecutor* executor, + aclrtStream stream); +#ifdef __cplusplus +} +#endif + void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) { // TODO: use ascendc // Only test with LLAMA model. ggml_tensor* src0 = dst->src[0]; // input ggml_tensor* src2 = dst->src[2]; // freq_factors - // TODO: with freq_factors - GGML_ASSERT(src2 == NULL); - // param float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; // const int n_past = ((int32_t *) dst->op_params)[0]; @@ -2867,13 +3152,11 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) { memcpy(&beta_fast, (int32_t*)dst->op_params + 9, sizeof(float)); memcpy(&beta_slow, (int32_t*)dst->op_params + 10, sizeof(float)); - GGML_ASSERT(n_dims <= ne0); + // TODO: n_dims <= ne0 + GGML_ASSERT(n_dims == ne0); GGML_ASSERT(n_dims % 2 == 0); - // TODO: ext_factor != 0 GGML_ASSERT(ext_factor == 0); - // TODO: freq_scale != 1 - GGML_ASSERT(freq_scale == 1); const float theta_scale = powf(freq_base, -2.0f / n_dims); @@ -2904,7 +3187,13 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float_t), sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS); aclnn_cache_init(ctx, dst, acl_cos_reshape_tensor, acl_sin_reshape_tensor, - theta_scale, is_neox); + theta_scale, freq_scale, attn_factor, is_neox); + + aclTensor* acl_src = ggml_cann_create_tensor(src0); + aclTensor* acl_dst = ggml_cann_create_tensor(dst); + +#ifdef ASCEND_310P + // Special ROPE operation for 310P // roll input void* input_roll_buffer; @@ -2947,7 +3236,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) { for (int i = 1; i < GGML_MAX_DIMS; i++) { minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1]; } - acl_minus_one_tensor = aclnn_ones( + acl_minus_one_tensor = aclnn_values( ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0], minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1); int64_t dim = 3; @@ -2974,17 +3263,15 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ACL_CHECK(aclDestroyTensor(acl_input_roll_tensor)); ACL_CHECK(aclDestroyTensor(acl_input_tensor)); - // init [-1, -1, -1, 1, 1，1，...] minus_one_scale_buffer = minus_one_scale_allocator.get(); - int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1}; size_t minus_one_nb[GGML_MAX_DIMS]; minus_one_nb[0] = sizeof(float_t); for (int i = 1; i < GGML_MAX_DIMS; i++) { minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1]; } - acl_minus_one_tensor = aclnn_ones( + acl_minus_one_tensor = aclnn_values( ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0], minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1); // -1 * first half @@ -3026,14 +3313,12 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) { acl_input_roll_mul_scale_tensor); // output - aclTensor* acl_src0 = ggml_cann_create_tensor(src0); - aclTensor* acl_dst = ggml_cann_create_tensor(dst); void* output_fp32_buffer; if (src0->type == GGML_TYPE_F32) { - aclnn_inplace_mul(ctx, acl_src0, acl_cos_reshape_tensor); + aclnn_inplace_mul(ctx, acl_src, acl_cos_reshape_tensor); aclnn_inplace_mul(ctx, acl_input_roll_mul_scale_tensor, acl_sin_reshape_tensor); - aclnn_add(ctx, acl_src0, acl_input_roll_mul_scale_tensor, acl_dst); + aclnn_add(ctx, acl_src, acl_input_roll_mul_scale_tensor, acl_dst); // TODO: ne0 != n_dims in mode2 } else if (src0->type == GGML_TYPE_F16) { size_t input_fp32_nb[GGML_MAX_DIMS]; @@ -3060,7 +3345,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) { aclTensor* output_fp32_tensor = ggml_cann_create_tensor( output_fp32_buffer, ACL_FLOAT, sizeof(float_t), dst->ne, input_fp32_nb, GGML_MAX_DIMS); - aclnn_mul(ctx, acl_src0, acl_cos_reshape_tensor, input_fp32_tensor1); + aclnn_mul(ctx, acl_src, acl_cos_reshape_tensor, input_fp32_tensor1); aclnn_mul(ctx, acl_input_roll_mul_scale_tensor, acl_sin_reshape_tensor, input_fp32_tensor2); aclnn_add(ctx, input_fp32_tensor1, input_fp32_tensor2, @@ -3070,13 +3355,73 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ACL_CHECK(aclDestroyTensor(input_fp32_tensor1)); ACL_CHECK(aclDestroyTensor(input_fp32_tensor2)); ACL_CHECK(aclDestroyTensor(output_fp32_tensor)); + ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor)); + ACL_CHECK(aclDestroyTensor(acl_minus_one_tensor)); + ACL_CHECK(aclDestroyTensor(acl_input_roll_mul_scale_tensor)); + ACL_CHECK(aclDestroyTensor(acl_input_roll_reshape_tensor)); + ACL_CHECK(aclDestroyTensor(acl_src)); } + return; +#endif - ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor)); + // src0 == GGML_TYPE_F16 + // TODO: optimization this `if` code + if (src0->type == GGML_TYPE_F16) { + ggml_cann_pool_alloc sin_final_allocator( + ctx.pool(), src0->ne[0] * src0->ne[2] * ggml_type_size(src0->type)); + ggml_cann_pool_alloc cos_final_allocator( + ctx.pool(), src0->ne[0] * src0->ne[2] * ggml_type_size(src0->type)); + void* sin_final_buffer = sin_final_allocator.get(); + void* cos_final_buffer = cos_final_allocator.get(); + + int64_t sin_final_ne[4] = {src0->ne[0], 1, src0->ne[2], 1}; + size_t sin_final_nb[GGML_MAX_DIMS]; + sin_final_nb[0] = ggml_type_size(src0->type); + for (int i = 1; i < GGML_MAX_DIMS; i++) { + sin_final_nb[i] = sin_final_nb[i - 1] * sin_final_ne[i - 1]; + } + aclTensor* acl_sin_final_tensor = ggml_cann_create_tensor( + sin_final_buffer, ggml_cann_type_mapping(src0->type), + ggml_type_size(src0->type), sin_final_ne, sin_final_nb, + GGML_MAX_DIMS); + aclTensor* acl_cos_final_tensor = ggml_cann_create_tensor( + cos_final_buffer, ggml_cann_type_mapping(src0->type), + ggml_type_size(src0->type), sin_final_ne, sin_final_nb, + GGML_MAX_DIMS); + + aclnn_cast(ctx, acl_sin_reshape_tensor, acl_sin_final_tensor, + ggml_cann_type_mapping(src0->type)); + aclnn_cast(ctx, acl_cos_reshape_tensor, acl_cos_final_tensor, + ggml_cann_type_mapping(src0->type)); + ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor)); + ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor)); + acl_sin_reshape_tensor = acl_sin_final_tensor; + acl_cos_reshape_tensor = acl_cos_final_tensor; + } + + uint64_t workspaceSize = 0; + aclOpExecutor* executor; + + void* workspaceAddr = nullptr; + + int acl_mode = mode; + if (mode == 0) { + acl_mode = 1; + } + + ACL_CHECK(aclnnRotaryPositionEmbeddingGetWorkspaceSize( + acl_src, acl_cos_reshape_tensor, acl_sin_reshape_tensor, acl_mode, + acl_dst, &workspaceSize, &executor)); + if (workspaceSize > 0) { + ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); + workspaceAddr = workspace_allocator.get(); + } + + ACL_CHECK(aclnnRotaryPositionEmbedding(workspaceAddr, workspaceSize, + executor, ctx.stream())); + + ACL_CHECK(aclDestroyTensor(acl_src)); ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor)); - ACL_CHECK(aclDestroyTensor(acl_minus_one_tensor)); - ACL_CHECK(aclDestroyTensor(acl_input_roll_mul_scale_tensor)); - ACL_CHECK(aclDestroyTensor(acl_input_roll_reshape_tensor)); - ACL_CHECK(aclDestroyTensor(acl_src0)); + ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor)); ACL_CHECK(aclDestroyTensor(acl_dst)); } diff --git a/ggml/src/ggml-cann/common.h b/ggml/src/ggml-cann/common.h index edfa496148f..5164cb74ec9 100644 --- a/ggml/src/ggml-cann/common.h +++ b/ggml/src/ggml-cann/common.h @@ -211,17 +211,20 @@ struct ggml_cann_pool_alloc { 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. */ - aclrtStream streams[GGML_CANN_MAX_STREAMS] = { - {nullptr}}; /**< Array of streams for the device. */ + 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)) {} + : device(device), name("CANN" + std::to_string(device)) { + ggml_cann_set_device(device); + description = aclrtGetSocName(); + } /** * @brief Destructor for cleaning up resources. diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp index 77634088143..04e25b8ab1a 100644 --- a/ggml/src/ggml-cann/ggml-cann.cpp +++ b/ggml/src/ggml-cann/ggml-cann.cpp @@ -122,6 +122,10 @@ static ggml_cann_device_info ggml_cann_init() { ACL_CHECK(aclrtMemGetAllocationGranularity( &prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED, &info.devices[id].vmm_granularity)); + + size_t free, total; + ggml_backend_cann_get_device_memory(id, &free, &total); + info.devices[id].total_vram = free; } // TODO: add more device info later. @@ -208,6 +212,11 @@ struct ggml_cann_pool_leg : public ggml_cann_pool { * @return A pointer to the allocated buffer. */ void* alloc(size_t size, size_t* actual_size) override { + const size_t alignment = 128; + size = GGML_PAD(size, alignment); + if (size == 0) { + size = alignment; + } #ifdef DEBUG_CANN_MALLOC int nnz = 0; size_t max_size = 0; @@ -246,13 +255,11 @@ struct ggml_cann_pool_leg : public ggml_cann_pool { return ptr; } void* ptr; - size_t look_ahead_size = (size_t)(1.05 * size); - look_ahead_size = 256 * ((look_ahead_size + 255) / 256); ggml_cann_set_device(device); ACL_CHECK( - aclrtMalloc(&ptr, look_ahead_size, ACL_MEM_MALLOC_HUGE_FIRST)); - *actual_size = look_ahead_size; - pool_size += look_ahead_size; + aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST)); + *actual_size = size; + pool_size += size; #ifdef DEBUG_CANN_MALLOC GGML_LOG_INFO( "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, " @@ -296,7 +303,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool { /** * @brief The maximum size of the virtual memory pool (32 GB). */ - static const size_t CANN_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB + size_t max_size; /** * @brief The device ID associated with this buffer pool. @@ -341,7 +348,11 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool { */ explicit ggml_cann_pool_vmm(int device) : device(device), - granularity(ggml_cann_info().devices[device].vmm_granularity) {} + granularity(ggml_cann_info().devices[device].vmm_granularity) { + auto dev = ggml_cann_info().devices[device]; + granularity = dev.vmm_granularity; + max_size = dev.total_vram; + } /** * @brief Destructor to free all buffers in the virtual memory pool. @@ -370,17 +381,19 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool { // round up the allocation size to the alignment to ensure that all // allocations are aligned for all data types const size_t alignment = 128; - size = alignment * ((size + alignment - 1) / alignment); + size = GGML_PAD(size, alignment); + if (size == 0) { + size = alignment; + } size_t avail = pool_size - pool_used; if (size > avail) { // round up to the next multiple of the granularity size_t reserve_size = size - avail; - reserve_size = - granularity * ((reserve_size + granularity - 1) / granularity); + reserve_size = GGML_PAD(reserve_size, granularity); - GGML_ASSERT(pool_size + reserve_size <= CANN_POOL_VMM_MAX_SIZE); + GGML_ASSERT(pool_size + reserve_size <= max_size); // allocate more physical memory aclrtPhysicalMemProp prop = {}; @@ -396,7 +409,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool { // reserve virtual address space (if not already reserved) if (pool_addr == 0) { ACL_CHECK(aclrtReserveMemAddress( - &pool_addr, CANN_POOL_VMM_MAX_SIZE, 0, NULL, 1)); + &pool_addr, max_size, 0, NULL, 1)); } // map at the end of the pool @@ -409,10 +422,11 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool { // add to the pool pool_size += reserve_size; - // GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB ( - // reserved %llu MB)\n", - // device, (unsigned long long) (pool_size/1024/1024), - // (unsigned long long) (reserve_size/1024/1024)); +#ifdef DEBUG_CANN_MALLOC + GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB (reserved %llu MB)\n", + device, (unsigned long long) (pool_size/1024/1024), + (unsigned long long) (reserve_size/1024/1024)); +#endif } GGML_ASSERT(pool_addr != 0); @@ -457,7 +471,6 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool { */ std::unique_ptr ggml_backend_cann_context::new_pool_for_device( int device) { - // return std::unique_ptr(new ggml_cann_pool_leg(device)); return std::unique_ptr(new ggml_cann_pool_vmm(device)); } @@ -1130,10 +1143,10 @@ ggml_backend_cann_buffer_type(int32_t device) { static bool ggml_backend_cann_buffer_type_initialized = false; if (!ggml_backend_cann_buffer_type_initialized) { - for (int32_t i = 0; i < GGML_CANN_MAX_DEVICES; i++) { + for (int32_t i = 0; i < ggml_cann_info().device_count; i++) { ggml_backend_cann_buffer_types[i] = { /* .iface = */ ggml_backend_cann_buffer_type_interface, - /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device), + /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), i), /* .context = */ new ggml_backend_cann_buffer_type_context{ i, "CANN" + std::to_string(i)}, @@ -1199,10 +1212,15 @@ static void * ggml_cann_host_malloc(size_t size) { return nullptr; } + const size_t alignment = 128; + size = GGML_PAD(size, alignment); + if (size == 0) { + size = alignment; + } + void * hostPtr = nullptr; aclError err = aclrtMallocHost((void **) &hostPtr, size); if (err != ACL_SUCCESS) { - GGML_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__, size / 1024.0 / 1024.0, aclGetRecentErrMsg()); return nullptr; @@ -1669,12 +1687,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, } case GGML_OP_MUL_MAT: { switch (op->src[0]->type) { - case GGML_TYPE_F16: - case GGML_TYPE_F32: case GGML_TYPE_Q8_0: - // TODO: fix me // Current groupsize should not be greater than k-1 in - // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(). + // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize + if (op->src[0]->ne[0] <= QK8_0) { + return false; + } + case GGML_TYPE_F16: + case GGML_TYPE_F32: case GGML_TYPE_Q4_0: return true; default: @@ -1706,9 +1726,41 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, return false; } } + case GGML_OP_CONT: { + // TODO: support GGML_TYPE_BF16 + switch (op->src[0]->type) { + case GGML_TYPE_F32: + case GGML_TYPE_F16: + return true; + default: + return false; + } + } + case GGML_OP_ROPE: { + // TODO: with ops-test v == 1 + float * ext_factor = (float*)((int32_t*)op->op_params + 7); + // TODO: n_dims <= ne0 + if (op->src[0]->ne[0] != op->op_params[1]) { + return false; + } + // TODO: ext_factor != 0 + if (*ext_factor != 0) { + return false; + } + return true; + } + case GGML_OP_UPSCALE: { + // aclnnUpsampleNearest2dGetWorkspaceSize not support + // selfDimN[2]/outDimN[2] or selfDimC[3]/outDimC[3] not equal + if (op->src[0]->ne[2] * op->ne[3] != op->src[0]->ne[3] * op->ne[2]) { + return false; + } + return true; + } + case GGML_OP_IM2COL: + case GGML_OP_CONCAT: case GGML_OP_DUP: case GGML_OP_REPEAT: - case GGML_OP_CONCAT: case GGML_OP_NONE: case GGML_OP_RESHAPE: case GGML_OP_VIEW: @@ -1722,17 +1774,13 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, case GGML_OP_SCALE: case GGML_OP_SQR: case GGML_OP_CLAMP: - case GGML_OP_CONT: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: - case GGML_OP_ROPE: - case GGML_OP_IM2COL: case GGML_OP_POOL_2D: case GGML_OP_SUM_ROWS: case GGML_OP_ARGSORT: case GGML_OP_ACC: case GGML_OP_GROUP_NORM: - case GGML_OP_UPSCALE: case GGML_OP_PAD: case GGML_OP_ARANGE: case GGML_OP_TIMESTEP_EMBEDDING: @@ -2064,16 +2112,17 @@ ggml_backend_reg_t ggml_backend_cann_reg() { dev_ctx->name = GGML_CANN_NAME + std::to_string(i); ggml_cann_set_device(i); ggml_backend_dev_t dev = new ggml_backend_device { - /* .interface = */ ggml_backend_cann_device_interface, - /* .reg = */ ®, - /* .context = */ dev_ctx + /* .iface = */ ggml_backend_cann_device_interface, + /* .reg = */ ®, + /* .context = */ dev_ctx }; ctx->devices.push_back(dev); } reg = ggml_backend_reg { - /* .interface = */ ggml_backend_cann_reg_interface, - /* .context = */ ctx + /* .api_version = */ GGML_BACKEND_API_VERSION, + /* .iface = */ ggml_backend_cann_reg_interface, + /* .context = */ ctx }; } @@ -2126,3 +2175,5 @@ void ggml_backend_cann_get_device_memory(int32_t device, size_t* free, ggml_cann_set_device(device); ACL_CHECK(aclrtGetMemInfo(ACL_HBM_MEM, free, total)); } + +GGML_BACKEND_DL_IMPL(ggml_backend_cann_reg) diff --git a/ggml/src/ggml-cann/kernels/CMakeLists.txt b/ggml/src/ggml-cann/kernels/CMakeLists.txt index 5b4fef91b58..d687220c3c5 100644 --- a/ggml/src/ggml-cann/kernels/CMakeLists.txt +++ b/ggml/src/ggml-cann/kernels/CMakeLists.txt @@ -1,7 +1,3 @@ -if (NOT SOC_TYPE) - set (SOC_TYPE "Ascend910B3") -endif() - file(GLOB SRC_FILES get_row_f32.cpp get_row_f16.cpp @@ -13,7 +9,6 @@ file(GLOB SRC_FILES dup.cpp ) -string(TOLOWER ${SOC_TYPE} SOC_VERSION) set(ASCEND_CANN_PACKAGE_PATH ${CANN_INSTALL_DIR}) set(RUN_MODE "npu" CACHE STRING "run mode: npu/sim") @@ -30,4 +25,6 @@ ascendc_library(ascendc_kernels STATIC ${SRC_FILES} ) +message(STATUS "CANN: compile ascend kernels witch SOC_TYPE:${SOC_TYPE}, SOC_VERSION:${SOC_VERSION}, compile macro:-D${SOC_TYPE_COMPILE_OPTION}.") +ascendc_compile_definitions(ascendc_kernels PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}") # ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP) diff --git a/ggml/src/ggml-cann/kernels/dup.cpp b/ggml/src/ggml-cann/kernels/dup.cpp index e2c651152f4..c7ba38d10a0 100644 --- a/ggml/src/ggml-cann/kernels/dup.cpp +++ b/ggml/src/ggml-cann/kernels/dup.cpp @@ -5,6 +5,7 @@ using namespace AscendC; #define BUFFER_NUM 2 +const int64_t SUPPORTED_MAX_DIM = 65535; // currently the limit of max block dim supportted by dup kernel is 65535template template class DupByRows { @@ -51,24 +52,36 @@ class DupByRows { __aicore__ inline void copy_in() { LocalTensor src_local = src_queue.AllocTensor(); - - DataCopyExtParams dataCopyParams; - dataCopyParams.blockCount = 1; - dataCopyParams.blockLen = num_elem * sizeof(SRC_T); - DataCopyPadExtParams padParams; - DataCopyPad(src_local, src_gm, dataCopyParams, padParams); - + const size_t elem_per_block = 32 / sizeof(SRC_T); + size_t tail = num_elem % elem_per_block; + size_t cpy_elements_len = tail > 0 ? num_elem + 1 : num_elem; + DataCopy(src_local, src_gm, cpy_elements_len); src_queue.EnQue(src_local); } __aicore__ inline void copy_out() { LocalTensor dst_local = dst_queue.DeQue(); - +#ifdef ASCEND_310P + const size_t elem_per_block = 32 / sizeof(DST_T); + size_t tail = num_elem % elem_per_block; + size_t len = num_elem & ~(elem_per_block - 1); + if (len > 0) { + DataCopy(dst_gm, dst_local, len); + } + if(tail != 0) { + for (size_t i = tail; i < elem_per_block; i++) { + dst_local[len + i].SetValue(0, 0); + } + SetAtomicAdd(); + DataCopy(dst_gm[len], dst_local[len], elem_per_block); + SetAtomicNone(); + } +#else DataCopyExtParams dataCopyParams; dataCopyParams.blockCount = 1; dataCopyParams.blockLen = num_elem * sizeof(DST_T); DataCopyPad(dst_gm, dst_local, dataCopyParams); - +#endif dst_queue.FreeTensor(dst_local); } diff --git a/ggml/src/ggml-cann/kernels/get_row_f16.cpp b/ggml/src/ggml-cann/kernels/get_row_f16.cpp index c704b5b2ec0..416b45104de 100644 --- a/ggml/src/ggml-cann/kernels/get_row_f16.cpp +++ b/ggml/src/ggml-cann/kernels/get_row_f16.cpp @@ -14,7 +14,7 @@ class GET_ROW_F16 { int64_t *output_ne_ub, size_t *output_nb_ub) { // TODO, use template for F16/f32 int64_t op_block_num = GetBlockNum(); - int64_t op_block_idx = GetBlockIdx(); + op_block_idx = GetBlockIdx(); for (int i = 0; i < 4; i++) { input_ne[i] = input_ne_ub[i]; @@ -59,32 +59,42 @@ class GET_ROW_F16 { } __aicore__ inline void copy_in(uint32_t offset, size_t len) { + size_t origin_len = len; LocalTensor input_local = input_queue.AllocTensor(); - size_t tail = len % 32; - len = len & ~31; - DataCopy(input_local, input_gm[offset], len); + const size_t elem_per_block = 32 / sizeof(half); + size_t tail = len % elem_per_block; + len = len & ~(elem_per_block - 1); if(tail != 0) { - DataCopyExtParams dataCopyParams; - dataCopyParams.blockCount = 1; - dataCopyParams.blockLen = tail * sizeof(half); - DataCopyPadExtParams padParams; - DataCopyPad(input_local[len], input_gm[offset + len], - dataCopyParams, padParams); + len += elem_per_block; } + DataCopy(input_local, input_gm[offset], len); input_queue.EnQue(input_local); } __aicore__ inline void copy_out(uint32_t offset, size_t len) { LocalTensor output_local = output_queue.DeQue(); - size_t tail = len % 32; - len = len & ~31; - DataCopy(output_gm[offset], output_local, len); + const size_t elem_per_block = 32 / sizeof(float); + size_t tail = len % elem_per_block; + len = len & ~(elem_per_block - 1); + if (len > 0) { + DataCopy(output_gm[offset], output_local, len); + } + if(tail != 0) { +#ifdef ASCEND_310P + for (size_t i = tail; i < elem_per_block; i++) { + output_local[len + i].SetValue(0, 0); + } + SetAtomicAdd(); + DataCopy(output_gm[offset + len], output_local[len], elem_per_block); + SetAtomicNone(); +#else DataCopyExtParams dataCopyParams; dataCopyParams.blockCount = 1; dataCopyParams.blockLen = tail * sizeof(float); DataCopyPad(output_gm[offset + len], output_local[len], dataCopyParams); +#endif } output_queue.FreeTensor(output_local); } @@ -150,6 +160,7 @@ class GET_ROW_F16 { GlobalTensor output_gm; TQue input_queue; TQue output_queue; + int64_t op_block_idx; }; template diff --git a/ggml/src/ggml-cann/kernels/get_row_f32.cpp b/ggml/src/ggml-cann/kernels/get_row_f32.cpp index 9db080af369..02116905b18 100644 --- a/ggml/src/ggml-cann/kernels/get_row_f32.cpp +++ b/ggml/src/ggml-cann/kernels/get_row_f32.cpp @@ -13,7 +13,7 @@ class GET_ROW_F32 { int64_t *indices_ne_ub, size_t *indices_nb_ub, int64_t *output_ne_ub, size_t *output_nb_ub) { int64_t op_block_num = GetBlockNum(); - int64_t op_block_idx = GetBlockIdx(); + op_block_idx = GetBlockIdx(); for (int i = 0; i < 4; i++) { input_ne[i] = input_ne_ub[i]; @@ -55,31 +55,40 @@ class GET_ROW_F32 { __aicore__ inline void copy_in(uint32_t offset, size_t len) { LocalTensor input_local = input_queue.AllocTensor(); - size_t tail = len % 32; - len = len & ~31; - DataCopy(input_local, input_gm[offset], len); + const size_t elem_per_block = 32 / sizeof(float); + size_t tail = len % elem_per_block; + len = len & ~(elem_per_block - 1); if(tail != 0) { - DataCopyExtParams dataCopyParams; - dataCopyParams.blockCount = 1; - dataCopyParams.blockLen = tail * sizeof(float); - DataCopyPadExtParams padParams; - DataCopyPad(input_local[len], input_gm[offset + len], - dataCopyParams, padParams); + len += elem_per_block; } + DataCopy(input_local, input_gm[offset], len); input_queue.EnQue(input_local); } __aicore__ inline void copy_out(uint32_t offset, size_t len) { LocalTensor output_local = output_queue.DeQue(); - size_t tail = len % 32; - len = len & ~31; - DataCopy(output_gm[offset], output_local, len); + const size_t elem_per_block = 32 / sizeof(float); + size_t tail = len % elem_per_block; + len = len & ~(elem_per_block - 1); + if (len > 0) { + DataCopy(output_gm[offset], output_local, len); + } + if(tail != 0) { +#ifdef ASCEND_310P + for (size_t i = tail; i < elem_per_block; i++) { + output_local[len + i].SetValue(0, 0); + } + SetAtomicAdd(); + DataCopy(output_gm[offset + len], output_local[len], elem_per_block); + SetAtomicNone(); +#else DataCopyExtParams dataCopyParams; dataCopyParams.blockCount = 1; dataCopyParams.blockLen = tail * sizeof(float); DataCopyPad(output_gm[offset + len], output_local[len], dataCopyParams); +#endif } output_queue.FreeTensor(output_local); } @@ -144,6 +153,7 @@ class GET_ROW_F32 { GlobalTensor output_gm; TQue input_queue; TQue output_queue; + int64_t op_block_idx; }; template diff --git a/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp b/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp index a80bfeec241..4fbe722086c 100644 --- a/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp +++ b/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp @@ -2,6 +2,15 @@ // optimize me. Use template to avoid copy code. using namespace AscendC; +#ifdef ASCEND_310P // 310P not support 4bit get row + extern "C" __global__ __aicore__ void ascendc_get_row_q4_0( + GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm, + GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm, + GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) { + // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed. + printf("Ascend310P not support 4bit get row.\n"); + } +#else #define BUFFER_NUM 2 @@ -191,3 +200,5 @@ extern "C" __global__ __aicore__ void ascendc_get_row_q4_0( indices_nb_ub, output_ne_ub, output_nb_ub); op.calculate(); } + +#endif // #ifdef ASCEND_310P diff --git a/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp b/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp index 8423b3f02a8..504b43afaa1 100644 --- a/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp +++ b/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp @@ -1,6 +1,14 @@ #include "kernel_operator.h" using namespace AscendC; +#ifdef ASCEND_310P + extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0( + GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm, + GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) { + // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed. + printf("Ascend310P not support f16->8bit quantization.\n"); + } +#else #define BUFFER_NUM 2 #define QK8_0 32 @@ -206,3 +214,5 @@ extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0( op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub); op.calculate(); } + +#endif // #ifdef ASCEND_310P diff --git a/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp b/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp index b7c575093e9..05b0bc1df59 100644 --- a/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp +++ b/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp @@ -1,6 +1,14 @@ #include "kernel_operator.h" using namespace AscendC; +#ifdef ASCEND_310P // 310P not support f32->8bit quantization + extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0( + GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm, + GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) { + // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed. + printf("Ascend310P not support f32->8bit quantization.\n"); + } +#else #define BUFFER_NUM 2 #define QK8_0 32 @@ -204,3 +212,5 @@ extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0( op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub); op.calculate(); } + +#endif // #ifdef ASCEND_310P diff --git a/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp b/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp index 9c8c86b66ad..1188937b744 100644 --- a/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp +++ b/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp @@ -1,6 +1,21 @@ #include "kernel_operator.h" using namespace AscendC; +#ifdef ASCEND_310P // 310P not support float->4bit quantization + extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0( + GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm, + GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) { + // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed. + printf("Ascend310P not support f32->4bit quantization.\n"); + } + + extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0( + GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm, + GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) { + // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed. + printf("Ascend310P not support f16->4bit quantization.\n"); + } +#else #define BUFFER_NUM 2 #define Group_Size 32 @@ -276,3 +291,5 @@ extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0( op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub); op.calculate(); } + +#endif // #ifdef ASCEND_310P diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index 05016139345..27253a6c2b3 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -418,6 +418,12 @@ typedef struct { } block_iq4_xs; static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding"); +typedef struct { + ggml_half d[4]; // deltas for 4 iq4_nl blocks + uint8_t qs[QK4_NL * 2];// nibbles / quants for 4 iq4_nl blocks +} block_iq4_nlx4; +static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding"); + #endif // GGML_COMMON_DECL #endif // GGML_COMMON_DECL diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt index cef41a0743c..bc326c05930 100644 --- a/ggml/src/ggml-cpu/CMakeLists.txt +++ b/ggml/src/ggml-cpu/CMakeLists.txt @@ -1,261 +1,354 @@ -add_library(ggml-cpu - ggml-cpu.c - ggml-cpu.cpp - ggml-cpu-aarch64.c - ggml-cpu-aarch64.h - ggml-cpu-quants.c - ggml-cpu-quants.h - ) - -target_link_libraries(ggml-cpu PRIVATE ggml-base) -target_include_directories(ggml-cpu PRIVATE . ..) - -if (APPLE AND GGML_ACCELERATE) - find_library(ACCELERATE_FRAMEWORK Accelerate) - if (ACCELERATE_FRAMEWORK) - message(STATUS "Accelerate framework found") - - add_compile_definitions(GGML_USE_ACCELERATE) - add_compile_definitions(ACCELERATE_NEW_LAPACK) - add_compile_definitions(ACCELERATE_LAPACK_ILP64) - - target_link_libraries(ggml-cpu PRIVATE ${ACCELERATE_FRAMEWORK}) +function(ggml_add_cpu_backend_variant_impl tag_name) + if (tag_name) + set(GGML_CPU_NAME ggml-cpu-${tag_name}) else() - message(WARNING "Accelerate framework not found") + set(GGML_CPU_NAME ggml-cpu) endif() -endif() -if (GGML_OPENMP) - find_package(OpenMP) - if (OpenMP_FOUND) - message(STATUS "OpenMP found") + ggml_add_backend_library(${GGML_CPU_NAME}) - add_compile_definitions(GGML_USE_OPENMP) + list (APPEND GGML_CPU_SOURCES + ggml-cpu/ggml-cpu.c + ggml-cpu/ggml-cpu.cpp + ggml-cpu/ggml-cpu-aarch64.c + ggml-cpu/ggml-cpu-aarch64.h + ggml-cpu/ggml-cpu-quants.c + ggml-cpu/ggml-cpu-quants.h + ggml-cpu/amx/amx.cpp + ggml-cpu/amx/amx.h + ggml-cpu/amx/mmq.cpp + ggml-cpu/amx/mmq.h + ggml-cpu/ggml-cpu-impl.h + ) - target_link_libraries(ggml-cpu PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX) + target_compile_features(${GGML_CPU_NAME} PRIVATE c_std_11 cxx_std_17) + target_include_directories(${GGML_CPU_NAME} PRIVATE . ggml-cpu) - # FIXME: should be replaced with a compiler id check - #if (GGML_MUSA) - # list(APPEND GGML_CPU_EXTRA_INCLUDES "/usr/lib/llvm-14/lib/clang/14.0.0/include") - # list(APPEND GGML_CPU_EXTRA_LIBS_PRIVATE "/usr/lib/llvm-14/lib/libomp.so") - #endif() - else() - message(WARNING "OpenMP not found") + if (APPLE AND GGML_ACCELERATE) + find_library(ACCELERATE_FRAMEWORK Accelerate) + if (ACCELERATE_FRAMEWORK) + message(STATUS "Accelerate framework found") + + target_compile_definitions(${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 ${ACCELERATE_FRAMEWORK}) + else() + message(WARNING "Accelerate framework not found") + endif() endif() -endif() -if (GGML_LLAMAFILE) - message(STATUS "Using llamafile") + if (GGML_OPENMP) + find_package(OpenMP) + if (OpenMP_FOUND) + target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_OPENMP) - add_compile_definitions(GGML_USE_LLAMAFILE) + target_link_libraries(${GGML_CPU_NAME} PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX) + else() + message(WARNING "OpenMP not found") + endif() + endif() - target_sources(ggml-cpu PRIVATE - llamafile/sgemm.cpp - llamafile/sgemm.h) -endif() + if (GGML_LLAMAFILE) + target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_LLAMAFILE) -if (GGML_CPU_HBM) - find_library(memkind memkind REQUIRED) + list(APPEND GGML_CPU_SOURCES + ggml-cpu/llamafile/sgemm.cpp + ggml-cpu/llamafile/sgemm.h) + endif() - message(STATUS "Using memkind for CPU HBM") + if (GGML_CPU_HBM) + find_library(memkind memkind REQUIRED) - add_compile_definitions(GGML_USE_CPU_HBM) + message(STATUS "Using memkind for CPU HBM") - target_link_libraries(ggml-cpu PUBLIC memkind) -endif() + target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_HBM) -if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR - CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR - (NOT CMAKE_OSX_ARCHITECTURES AND - NOT CMAKE_GENERATOR_PLATFORM_LWR AND - CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$")) + target_link_libraries(${GGML_CPU_NAME} PUBLIC memkind) + endif() - message(STATUS "ARM detected") + if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR + CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR + (NOT CMAKE_OSX_ARCHITECTURES AND + NOT CMAKE_GENERATOR_PLATFORM_LWR AND + CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$")) - if (MSVC) - add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead - add_compile_definitions(__ARM_NEON) - add_compile_definitions(__ARM_FEATURE_FMA) + message(STATUS "ARM detected") - set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS}) - string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2") + if (MSVC) + list(APPEND ARCH_DEFINITIONS __aarch64__) # MSVC defines _M_ARM64 instead + list(APPEND ARCH_DEFINITIONS __ARM_NEON) + list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_FMA) - check_cxx_source_compiles("#include \nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD) - if (GGML_COMPILER_SUPPORT_DOTPROD) - add_compile_definitions(__ARM_FEATURE_DOTPROD) - endif () + set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS}) + string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2") - check_cxx_source_compiles("#include \nint main() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8) + check_cxx_source_compiles("#include \nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD) + if (GGML_COMPILER_SUPPORT_DOTPROD) + list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_DOTPROD) - if (GGML_COMPILER_SUPPORT_MATMUL_INT8) - add_compile_definitions(__ARM_FEATURE_MATMUL_INT8) - endif () + message(STATUS "ARM feature DOTPROD enabled") + endif () - check_cxx_source_compiles("#include \nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC) - if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC) - add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) - endif () + check_cxx_source_compiles("#include \nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8) - set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV}) - else() - check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E) - if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "") - list(APPEND ARCH_FLAGS -mfp16-format=ieee) - endif() - if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6") - # Raspberry Pi 1, Zero - list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access) - endif() - if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7") - if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android") - # Android armeabi-v7a - list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations) - else() - # Raspberry Pi 2 - list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations) + if (GGML_COMPILER_SUPPORT_MATMUL_INT8) + list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_MATMUL_INT8) + + message(STATUS "ARM feature MATMUL_INT8 enabled") + endif () + + check_cxx_source_compiles("#include \nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC) + if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC) + list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_FP16_VECTOR_ARITHMETIC) + + message(STATUS "ARM feature FP16_VECTOR_ARITHMETIC enabled") + endif () + + set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV}) + elseif (APPLE) + if (GGML_NATIVE) + set(USER_PROVIDED_MARCH FALSE) + foreach(flag_var IN ITEMS CMAKE_C_FLAGS CMAKE_CXX_FLAGS CMAKE_REQUIRED_FLAGS) + if ("${${flag_var}}" MATCHES "-march=[a-zA-Z0-9+._-]+") + set(USER_PROVIDED_MARCH TRUE) + break() + endif() + endforeach() + + if (NOT USER_PROVIDED_MARCH) + set(MARCH_FLAGS "-march=armv8.2a") + + check_cxx_source_compiles("#include \nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD) + if (GGML_COMPILER_SUPPORT_DOTPROD) + set(MARCH_FLAGS "${MARCH_FLAGS}+dotprod") + list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_DOTPROD) + + message(STATUS "ARM feature DOTPROD enabled") + endif () + + set(TEST_I8MM_FLAGS "-march=armv8.2a+i8mm") + + set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS}) + set(CMAKE_REQUIRED_FLAGS "${CMAKE_REQUIRED_FLAGS} ${TEST_I8MM_FLAGS}") + + check_cxx_source_compiles("#include \nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8) + if (GGML_COMPILER_SUPPORT_MATMUL_INT8) + set(MARCH_FLAGS "${MARCH_FLAGS}+i8mm") + list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_MATMUL_INT8) + + message(STATUS "ARM feature MATMUL_INT8 enabled") + endif () + + set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE}) + + list(APPEND ARCH_FLAGS "${MARCH_FLAGS}") + endif () + endif () + else() + check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E) + if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "") + list(APPEND ARCH_FLAGS -mfp16-format=ieee) + endif() + if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6") + # Raspberry Pi 1, Zero + list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access) + endif() + if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7") + if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android") + # Android armeabi-v7a + list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations) + else() + # Raspberry Pi 2 + list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations) + endif() + endif() + if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8") + # Android arm64-v8a + # Raspberry Pi 3, 4, Zero 2 (32-bit) + list(APPEND ARCH_FLAGS -mno-unaligned-access) + endif() + if (GGML_SVE) + list(APPEND ARCH_FLAGS -march=armv8.6-a+sve) endif() endif() - if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8") - # Android arm64-v8a - # Raspberry Pi 3, 4, Zero 2 (32-bit) - list(APPEND ARCH_FLAGS -mno-unaligned-access) - endif() - if (GGML_SVE) - list(APPEND ARCH_FLAGS -march=armv8.6-a+sve) - endif() - endif() -elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR - (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND - CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$")) - message(STATUS "x86 detected") - if (MSVC) - # instruction set detection for MSVC only - if (GGML_NATIVE) - # TODO: improve, should not reference files from the parent folder - include(cmake/FindSIMD.cmake) - endif () - if (GGML_AVX512) - list(APPEND ARCH_FLAGS /arch:AVX512) - # MSVC has no compile-time flags enabling specific - # AVX512 extensions, neither it defines the - # macros corresponding to the extensions. - # Do it manually. - if (GGML_AVX512_VBMI) - add_compile_definitions($<$:__AVX512VBMI__>) - add_compile_definitions($<$:__AVX512VBMI__>) - if (CMAKE_C_COMPILER_ID STREQUAL "Clang") - list(APPEND ARCH_FLAGS -mavx512vbmi) + elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR + (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND + CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$")) + 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) + else () + list(APPEND ARCH_FLAGS /arch:SSE4.2) + list(APPEND ARCH_DEFINITIONS GGML_SSE42) endif() - if (GGML_AVX512_VNNI) - add_compile_definitions($<$:__AVX512VNNI__>) - add_compile_definitions($<$:__AVX512VNNI__>) - if (CMAKE_C_COMPILER_ID STREQUAL "Clang") + if (GGML_AVX_VNNI) + # MSVC generates AVX512 with AVX-VNNI intrinsics even with /arch:AVX2 + #list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI) + endif() + else () + if (GGML_NATIVE) + list(APPEND ARCH_FLAGS -march=native) + else () + list(APPEND ARCH_FLAGS -msse4.2) + list(APPEND ARCH_DEFINITIONS GGML_SSE42) + 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_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() - endif() - if (GGML_AVX512_BF16) - add_compile_definitions($<$:__AVX512BF16__>) - add_compile_definitions($<$:__AVX512BF16__>) - if (CMAKE_C_COMPILER_ID STREQUAL "Clang") + 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() - if (GGML_AMX_TILE) - add_compile_definitions($<$:__AMX_TILE__>) - add_compile_definitions($<$:__AMX_TILE__>) - endif() - if (GGML_AMX_INT8) - add_compile_definitions($<$:__AMX_INT8__>) - add_compile_definitions($<$:__AMX_INT8__>) - endif() - if (GGML_AMX_BF16) - add_compile_definitions($<$:__AMX_BF16__>) - add_compile_definitions($<$:__AMX_BF16__>) - endif() - elseif (GGML_AVX2) - list(APPEND ARCH_FLAGS /arch:AVX2) - elseif (GGML_AVX) - list(APPEND ARCH_FLAGS /arch:AVX) - endif() - else() - if (GGML_NATIVE) - list(APPEND ARCH_FLAGS -march=native) - endif() - if (GGML_F16C) - list(APPEND ARCH_FLAGS -mf16c) - endif() - if (GGML_FMA) - list(APPEND ARCH_FLAGS -mfma) - endif() - if (GGML_AVX) - list(APPEND ARCH_FLAGS -mavx) - endif() - if (GGML_AVX2) - list(APPEND ARCH_FLAGS -mavx2) - endif() - if (GGML_AVX512) - list(APPEND ARCH_FLAGS -mavx512f) - list(APPEND ARCH_FLAGS -mavx512dq) - list(APPEND ARCH_FLAGS -mavx512bw) - endif() - if (GGML_AVX512_VBMI) - list(APPEND ARCH_FLAGS -mavx512vbmi) endif() - if (GGML_AVX512_VNNI) - list(APPEND ARCH_FLAGS -mavx512vnni) + elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64") + message(STATUS "PowerPC detected") + execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1" OUTPUT_VARIABLE POWER10_M) + string(FIND "${POWER10_M}" "POWER10" substring_index) + if (NOT DEFINED substring_index OR "${substring_index}" STREQUAL "") + set(substring_index -1) endif() - if (GGML_AVX512_BF16) - list(APPEND ARCH_FLAGS -mavx512bf16) + + if (${substring_index} GREATER_EQUAL 0) + list(APPEND ARCH_FLAGS -mcpu=power10) + elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le") + list(APPEND ARCH_FLAGS -mcpu=powerpc64le) + else() + list(APPEND ARCH_FLAGS -mcpu=native -mtune=native) + # TODO: Add targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be) endif() - if (GGML_AMX_TILE) - list(APPEND ARCH_FLAGS -mamx-tile) + elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64") + message(STATUS "loongarch64 detected") + + list(APPEND ARCH_FLAGS -march=loongarch64) + if (GGML_LASX) + list(APPEND ARCH_FLAGS -mlasx) endif() - if (GGML_AMX_INT8) - list(APPEND ARCH_FLAGS -mamx-int8) + if (GGML_LSX) + list(APPEND ARCH_FLAGS -mlsx) endif() - if (GGML_AMX_BF16) - list(APPEND ARCH_FLAGS -mamx-bf16) + elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64") + message(STATUS "RISC-V detected") + if (GGML_RVV) + list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d) endif() - endif() -elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64") - message(STATUS "PowerPC detected") - execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1" OUTPUT_VARIABLE POWER10_M) - string(FIND "${POWER10_M}" "POWER10" substring_index) - if (NOT DEFINED substring_index OR "${substring_index}" STREQUAL "") - set(substring_index -1) - endif() - - if (${substring_index} GREATER_EQUAL 0) - list(APPEND ARCH_FLAGS -mcpu=power10) - elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le") - list(APPEND ARCH_FLAGS -mcpu=powerpc64le) else() - list(APPEND ARCH_FLAGS -mcpu=native -mtune=native) - #TODO: Add targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be) + message(STATUS "Unknown architecture") endif() -elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64") - message(STATUS "loongarch64 detected") - list(APPEND ARCH_FLAGS -march=loongarch64) - if (GGML_LASX) - list(APPEND ARCH_FLAGS -mlasx) - endif() - if (GGML_LSX) - list(APPEND ARCH_FLAGS -mlsx) + if (GGML_CPU_AARCH64) + target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_AARCH64) endif() -else() - message(STATUS "Unknown architecture") -endif() -if (GGML_CPU_AARCH64) - message(STATUS "Using runtime weight conversion of Q4_0 to Q4_0_x_x to enable optimized GEMM/GEMV kernels") - add_compile_definitions(GGML_USE_CPU_AARCH64) -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 ${ARCH_DEFINITIONS}) -target_compile_options(ggml-cpu PRIVATE "$<$:${ARCH_FLAGS}>") -target_compile_options(ggml-cpu PRIVATE "$<$:${ARCH_FLAGS}>") + if (GGML_BACKEND_DL) + # 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 ${GGML_CPU_NAME}-feats) + add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/cpu-feats-x86.cpp) + target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include) + target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARCH_DEFINITIONS}) + target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED) + set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON) + target_link_libraries(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_FEATS_NAME}) + endif() -if (EMSCRIPTEN) - set_target_properties(ggml-cpu PROPERTIES COMPILE_FLAGS "-msimd128") -endif() + if (EMSCRIPTEN) + set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128") + endif() +endfunction() diff --git a/ggml/src/ggml-cpu/amx/amx.cpp b/ggml/src/ggml-cpu/amx/amx.cpp new file mode 100644 index 00000000000..09c0df0f5c2 --- /dev/null +++ b/ggml/src/ggml-cpu/amx/amx.cpp @@ -0,0 +1,196 @@ +#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" + +#if defined(__gnu_linux__) +#include +#include +#endif + +#include +#include +#include + +#if defined(__AMX_INT8__) && defined(__AVX512VNNI__) + +// 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 void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { + memset((char *)tensor->data + offset, value, size); + + GGML_UNUSED(buffer); +} + +static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { + if (qtype_has_amx_kernels(tensor->type)) { + ggml_backend_amx_convert_weight(tensor, data, offset, size); + } else { + memcpy((char *)tensor->data + offset, data, size); + } + + GGML_UNUSED(buffer); +} + +static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { + GGML_ASSERT(!qtype_has_amx_kernels(tensor->type)); + memcpy(data, (const char *)tensor->data + offset, size); + + GGML_UNUSED(buffer); +} + +static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) { + if (ggml_backend_buffer_is_host(src->buffer)) { + if (qtype_has_amx_kernels(src->type)) { + ggml_backend_amx_convert_weight(dst, src->data, 0, ggml_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 = */ NULL, // no initialization required + /* .memset_tensor = */ ggml_backend_amx_buffer_memset_tensor, + /* .set_tensor = */ ggml_backend_amx_buffer_set_tensor, + /* .get_tensor = */ ggml_backend_amx_buffer_get_tensor, + /* .cpy_tensor = */ ggml_backend_amx_buffer_cpy_tensor, + /* .clear = */ ggml_backend_amx_buffer_clear, + /* .reset = */ NULL, +}; + +static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) { + return "AMX"; + + GGML_UNUSED(buft); +} + +static ggml_backend_buffer_t ggml_backend_amx_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { + void * data = aligned_alloc(TENSOR_ALIGNMENT, size); + if (data == NULL) { + fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size); + return NULL; + } + + return ggml_backend_buffer_init(buft, ggml_backend_amx_buffer_interface, data, size); +} + +static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { + return TENSOR_ALIGNMENT; + + GGML_UNUSED(buft); +} + +static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) { + return ggml_backend_amx_get_alloc_size(tensor); + + GGML_UNUSED(buft); +} + +static bool ggml_backend_amx_buffer_type_is_host(ggml_backend_buffer_type_t buft) { + return false; + + GGML_UNUSED(buft); +} + +#define ARCH_GET_XCOMP_PERM 0x1022 +#define ARCH_REQ_XCOMP_PERM 0x1023 +#define XFEATURE_XTILECFG 17 +#define XFEATURE_XTILEDATA 18 + +static bool ggml_amx_init() { +#if defined(__gnu_linux__) + if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) { + fprintf(stderr, "AMX is not ready to be used!\n"); + return false; + } + return true; +#elif defined(_WIN32) + return true; +#endif +} +ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() { + static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = { + /* .iface = */ { + /* .get_name = */ ggml_backend_amx_buffer_type_get_name, + /* .alloc_buffer = */ ggml_backend_amx_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_amx_buffer_type_get_alignment, + /* .get_max_size = */ NULL, // defaults to SIZE_MAX + /* .get_alloc_size = */ ggml_backend_amx_buffer_type_get_alloc_size, + /* .is_host = */ ggml_backend_amx_buffer_type_is_host, + }, + /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0), + /* .context = */ NULL, + }; + + if (!ggml_amx_init()) { + return NULL; + } + + return &ggml_backend_buffer_type_amx; +} + +bool ggml_backend_amx_buft_is_amx(ggml_backend_buffer_type_t buft) { + return buft->iface.get_name == ggml_backend_amx_buffer_type_get_name; +} + +bool ggml_backend_amx_device_supports_op(const struct ggml_tensor * op) { + // handle only 2d gemm for now + auto is_contiguous_2d = [](const struct ggml_tensor * t) { + return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1; + }; + + switch (op->op) { + case GGML_OP_NONE: + case GGML_OP_RESHAPE: + case GGML_OP_VIEW: + case GGML_OP_PERMUTE: + case GGML_OP_TRANSPOSE: + return true; + + case GGML_OP_MUL_MAT: { + const struct ggml_tensor * src0 = op->src[0]; + const struct ggml_tensor * src1 = op->src[1]; + + const enum ggml_type type = src0->type; + const int64_t ne0 = op->ne[0]; + + // amx kernels enables for Q4_0, Q4_1, Q8_0, F16 + // Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256 + bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16); + + bool can_use_amx = + is_contiguous_2d(src0) && // src0 must be contiguous + is_contiguous_2d(src1) && // src1 must be contiguous + src1->type == GGML_TYPE_F32 && // src1 must be float32 + has_amx_kernels && // with amx kernel impls + ne0 % (TILE_N * 2) == 0; // out_features is 32x + + return can_use_amx; + } + default: + return false; + } +} + +#endif // defined(__AMX_INT8__) && defined(__AVX512VNNI__) diff --git a/ggml/src/ggml-cpu/amx/amx.h b/ggml/src/ggml-cpu/amx/amx.h new file mode 100644 index 00000000000..c4354627361 --- /dev/null +++ b/ggml/src/ggml-cpu/amx/amx.h @@ -0,0 +1,20 @@ +#include "ggml-backend.h" +#include "ggml-cpu-impl.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#if defined(__AMX_INT8__) && defined(__AVX512VNNI__) + +ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void); +bool ggml_backend_amx_buft_is_amx(ggml_backend_buffer_type_t buft); +bool ggml_backend_amx_device_supports_op(const struct ggml_tensor * op); +void ggml_backend_amx_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst); +size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst); + +#endif + +#ifdef __cplusplus +} +#endif diff --git a/ggml/src/ggml-cpu/amx/common.h b/ggml/src/ggml-cpu/amx/common.h new file mode 100644 index 00000000000..40074c3fc92 --- /dev/null +++ b/ggml/src/ggml-cpu/amx/common.h @@ -0,0 +1,100 @@ +#pragma once + +#include "ggml.h" +#include "ggml-cpu-impl.h" + +#include +#include +#include + +#if defined(_OPENMP) +#include +#endif + +#define TILE_M 16 +#define TILE_N 16 +#define TILE_K 32 +#define VNNI_BLK 4 + +#define AMX_BLK_SIZE 32 + +#define TMM0 0 +#define TMM1 1 +#define TMM2 2 +#define TMM3 3 +#define TMM4 4 +#define TMM5 5 +#define TMM6 6 +#define TMM7 7 + +// parallel routines +template ::value, int>::type = 0> +inline T div_up(T x, T y) { return (x + y - 1) / y; } + +template +inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) { +#if 0 + // onednn partition pattern + T& n_my = n_end; + if (nth <= 1 || n == 0) { + n_start = 0; + n_my = n; + } else { + T n1 = div_up(n, nth); + T n2 = n1 - 1; + T T1 = n - n2 * nth; + n_my = ith < T1 ? n1 : n2; + n_start = ith <= T1 ? ith*n1 : T1 * n1 + (ith - T1) * n2; + } + n_end += n_start; +#else + // pytorch aten partition pattern + T n_my = div_up(n, nth); + n_start = ith * n_my; + n_end = std::min(n_start + n_my, n); +#endif +} + +template +inline void parallel_for(int nth, int n, const func_t& f) { +#if defined(_OPENMP) +#pragma omp parallel num_threads(nth) +{ + //int nth = omp_get_num_threads(); + int ith = omp_get_thread_num(); + int tbegin, tend; + balance211(n, nth, ith, tbegin, tend); + f(tbegin, tend); +} +#else + f(0, n); + + GGML_UNUSED(nth); +#endif +} + +template +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); +} + +// ggml backend context +struct ggml_backend_amx_context { + int n_threads = GGML_DEFAULT_N_THREADS; + std::unique_ptr work_data; + size_t work_size = 0; +}; diff --git a/ggml/src/ggml-cpu/amx/mmq.cpp b/ggml/src/ggml-cpu/amx/mmq.cpp new file mode 100644 index 00000000000..0ec3aa86df8 --- /dev/null +++ b/ggml/src/ggml-cpu/amx/mmq.cpp @@ -0,0 +1,2522 @@ + +#if defined(__GNUC__) +#pragma GCC diagnostic ignored "-Wpedantic" +#pragma GCC diagnostic ignored "-Wunused-local-typedefs" +#endif + +#include "amx.h" +#include "mmq.h" +#include "ggml-impl.h" +#include "ggml-cpu-impl.h" +#include "ggml-cpu-quants.h" +#include "ggml-quants.h" +#include +#include + +#if defined(__gnu_linux__) +#include +#include +#endif + +#if defined(_OPENMP) +#include +#endif + +#if (defined(_WIN32) || defined(_WIN64)) +#define RESTRICT __restrict +#else +#define RESTRICT __restrict__ +#endif + +#if (defined(_WIN32) || defined(_WIN64)) +#define ALWAYS_INLINE __forceinline +#elif __has_attribute(always_inline) || defined(__GNUC__) +#define ALWAYS_INLINE __attribute__((__always_inline__)) inline +#else +#define ALWAYS_INLINE inline +#endif + +#if defined(__AMX_INT8__) && defined(__AVX512VNNI__) + +namespace { + +// Forced unrolling +template +struct Unroll { + template + ALWAYS_INLINE void operator()(const Func& f, Args... args) const { + Unroll{}(f, args...); + f(std::integral_constant{}, args...); + } +}; + +template <> +struct Unroll<1> { + template + ALWAYS_INLINE void operator()(const Func& f, Args... args) const { + f(std::integral_constant{}, args...); + } +}; + +// type traits +template struct PackedTypes {}; +template <> struct PackedTypes { using type = int8_t; }; +template <> struct PackedTypes { using type = uint8_t; }; +template <> struct PackedTypes { using type = int8_t; }; +template using packed_B_type = typename PackedTypes::type; + +template +struct do_compensate : std::integral_constant::value> {}; + +template +struct do_unpack : std::integral_constant::value || + std::is_same::value> {}; + +template +struct is_type_qkk : std::integral_constant::value || + std::is_same::value || + std::is_same::value || + std::is_same::value> {}; + +#define GGML_DISPATCH_FLOATING_TYPES(TYPE, ...) \ + [&] { \ + switch (TYPE) { \ + case GGML_TYPE_F16: { \ + using type = ggml_fp16_t; \ + constexpr int blck_size = 16; \ + return __VA_ARGS__(); \ + } \ + case GGML_TYPE_BF16: { \ + using type = ggml_bf16_t; \ + constexpr int blck_size = 32; \ + return __VA_ARGS__(); \ + } \ + default: \ + fprintf(stderr, "Unsupported floating data type\n"); \ + } \ + }() + +#define GGML_DISPATCH_QTYPES(QT, ...) \ + [&] { \ + switch (QT) { \ + case GGML_TYPE_Q4_0: { \ + using type = block_q4_0; \ + using vec_dot_type = block_q8_0; \ + constexpr int blck_size = QK4_0; \ + return __VA_ARGS__(); \ + } \ + case GGML_TYPE_Q4_1: { \ + using type = block_q4_1; \ + using vec_dot_type = block_q8_1; \ + constexpr int blck_size = QK4_1; \ + return __VA_ARGS__(); \ + } \ + case GGML_TYPE_Q8_0: { \ + using type = block_q8_0; \ + using vec_dot_type = block_q8_0; \ + constexpr int blck_size = QK8_0; \ + return __VA_ARGS__(); \ + } \ + case GGML_TYPE_Q4_K: { \ + using type = block_q4_K; \ + using vec_dot_type = block_q8_K; \ + constexpr int blck_size = QK_K; \ + return __VA_ARGS__(); \ + } \ + case GGML_TYPE_Q5_K: { \ + using type = block_q5_K; \ + using vec_dot_type = block_q8_K; \ + constexpr int blck_size = QK_K; \ + return __VA_ARGS__(); \ + } \ + case GGML_TYPE_Q6_K: { \ + using type = block_q6_K; \ + using vec_dot_type = block_q8_K; \ + constexpr int blck_size = QK_K; \ + return __VA_ARGS__(); \ + } \ + case GGML_TYPE_IQ4_XS: { \ + using type = block_iq4_xs; \ + using vec_dot_type = block_q8_K; \ + constexpr int blck_size = QK_K; \ + return __VA_ARGS__(); \ + } \ + default: \ + fprintf(stderr, "Unsupported quantized data type: %d\n", int(TYPE)); \ + } \ + }() + +#define GGML_DISPATCH_BOOL(BOOL_V, BOOL_NAME, ...) \ + [&] { \ + if (BOOL_V) { \ + constexpr bool BOOL_NAME = true; \ + return __VA_ARGS__(); \ + } else { \ + constexpr bool BOOL_NAME = false; \ + return __VA_ARGS__(); \ + } \ + }() + +// define amx tile config data structure +struct tile_config_t{ + uint8_t palette_id = 0; + uint8_t start_row = 0; + uint8_t reserved_0[14] = {0}; + uint16_t colsb[16] = {0}; + uint8_t rows[16] = {0}; +}; + +// Notes: amx tile config +// +// Typically, TMUL calculates A and B of size 16 x 64 containing INT8 values, +// and accumulate the result to a 16 x 16 matrix C containing INT32 values, +// +// As many GGUF quantized types as `block_size` of 32, so a 16-16-32 config is used +// instead of the normally used 16-16-64 config. +// +// Block A: {16, 32}, dtype = int8_t +// Block B: {16, 32}, dtype = uint8_t/int8_t +// Block C: {16, 16}, dtype = int32_t +// +// Block B needs to be prepacked to vnni format before feeding into TMUL: +// packed_B: from {n, k} to {k/vnni_blk, n, vnni_blck}, viewed in 2d, we get {8, 64} +// +// Therefore, we get tileconfig: +// A B C +// rows 16 8 16 +// colsb 32 64 16 +// +// For tile distribution, follow a 2-2-4 pattern, e.g. A used TMM2-TMM3, B used TMM0-TMM1, +// C used TMM4-TMM7: +// B TMM0 B TMM1 +// A TMM2 C TMM4 C TMM6 +// A TMM3 C TMM5 C TMM7 +// +// Each `amx` kernel handles 4 blocks at a time: 2MB * 2NB, when m < 2 * BLOCK_M, unpack A +// will be needed. +// +// Here another commonly used pattern 1-3-3 is skipped, as it is mostly used when m <=16; +// and the sinlge batch gemm (m=1) has a special fast path with `avx512-vnni`. +// +// ref: https://www.intel.com/content/www/us/en/developer/articles/code-sample/ +// advanced-matrix-extensions-intrinsics-functions.html +// + +#define TC_CONFIG_TILE(i, r, cb) tc.rows[i] = r; tc.colsb[i] = cb +void ggml_tile_config_init(void) { + static thread_local bool is_first_time = true; + + if (!is_first_time) { + return; + } + + static thread_local tile_config_t tc; + tile_config_t current_tc; + _tile_storeconfig(¤t_tc); + + // load only when config changes + if (tc.palette_id == 0 || (memcmp(¤t_tc.colsb, &tc.colsb, sizeof(uint16_t) * 8) != 0 && + memcmp(¤t_tc.rows, &tc.rows, sizeof(uint8_t) * 8) != 0)) { + tc.palette_id = 1; + tc.start_row = 0; + TC_CONFIG_TILE(TMM0, 8, 64); + TC_CONFIG_TILE(TMM1, 8, 64); + TC_CONFIG_TILE(TMM2, 16, 32); + TC_CONFIG_TILE(TMM3, 16, 32); + TC_CONFIG_TILE(TMM4, 16, 64); + TC_CONFIG_TILE(TMM5, 16, 64); + TC_CONFIG_TILE(TMM6, 16, 64); + TC_CONFIG_TILE(TMM7, 16, 64); + _tile_loadconfig(&tc); + } + + is_first_time = false; +} + +// we need an extra 16 * 4B (TILE_N * int32_t) for each NB/KB block for compensation. +// See the notes `s8s8 igemm compensation in avx512-vnni` for detail. +template +int get_tile_size() { + int tile_size = TILE_N * sizeof(TB); + if (do_compensate::value) { + tile_size += TILE_N * sizeof(int32_t); + } + if (std::is_same::value || + std::is_same::value) { + tile_size += TILE_N * 4; + } + if (std::is_same::value) { + tile_size += TILE_N * 2; + } + return tile_size; +} + +template +int get_row_size(int K) { + int KB = K / BLOCK_K; + int row_size = KB * sizeof(TB); + if (do_compensate::value) { + row_size += KB * sizeof(int32_t); + } + if (std::is_same::value || + std::is_same::value) { + row_size += KB * 4; + } + if (std::is_same::value) { + row_size += KB * 2; + } + return row_size; +} + +// vectorized dtype conversion +inline float FP16_TO_FP32(ggml_half val) { + __m256i v = _mm256_setr_epi16( + val, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0); + __m512 o = _mm512_cvtph_ps(v); + return _mm512_cvtss_f32(o); +} + +inline __m512 FP16_TO_FP32_VEC(ggml_half val) { + __m256i v = _mm256_set1_epi16(val); + return _mm512_cvtph_ps(v); +} + +// horizontal reduce +inline float _mm512_reduce_max_ps(const __m512 x) { + __m512 v = x; + __m512 v1 = _mm512_shuffle_f32x4(v, v, 0x4E); + v = _mm512_max_ps(v, v1); + v1 = _mm512_shuffle_f32x4(v, v, 0xB1); + v = _mm512_max_ps(v, v1); + v1 = _mm512_shuffle_ps(v, v, 0x4E); + v = _mm512_max_ps(v, v1); + v1 = _mm512_shuffle_ps(v, v, 0xB1); + v = _mm512_max_ps(v, v1); + return _mm512_cvtss_f32(v); +} + +// transpose utils +#define SHUFFLE_EPI32(a, b, mask) \ + _mm256_castps_si256(_mm256_shuffle_ps(_mm256_castsi256_ps(a), _mm256_castsi256_ps(b), mask)) +inline void transpose_8x8_32bit(__m256i * v, __m256i * v1) { + // unpacking and 32-bit elements + v1[0] = _mm256_unpacklo_epi32(v[0], v[1]); + v1[1] = _mm256_unpackhi_epi32(v[0], v[1]); + v1[2] = _mm256_unpacklo_epi32(v[2], v[3]); + v1[3] = _mm256_unpackhi_epi32(v[2], v[3]); + v1[4] = _mm256_unpacklo_epi32(v[4], v[5]); + v1[5] = _mm256_unpackhi_epi32(v[4], v[5]); + v1[6] = _mm256_unpacklo_epi32(v[6], v[7]); + v1[7] = _mm256_unpackhi_epi32(v[6], v[7]); + + // shuffling the 32-bit elements + v[0] = SHUFFLE_EPI32(v1[0], v1[2], 0x44); + v[1] = SHUFFLE_EPI32(v1[0], v1[2], 0xee); + v[2] = SHUFFLE_EPI32(v1[4], v1[6], 0x44); + v[3] = SHUFFLE_EPI32(v1[4], v1[6], 0xee); + v[4] = SHUFFLE_EPI32(v1[1], v1[3], 0x44); + v[5] = SHUFFLE_EPI32(v1[1], v1[3], 0xee); + v[6] = SHUFFLE_EPI32(v1[5], v1[7], 0x44); + v[7] = SHUFFLE_EPI32(v1[5], v1[7], 0xee); + + // shuffling 128-bit elements + v1[0] = _mm256_permute2f128_si256(v[2], v[0], 0x02); + v1[1] = _mm256_permute2f128_si256(v[3], v[1], 0x02); + v1[2] = _mm256_permute2f128_si256(v[6], v[4], 0x02); + v1[3] = _mm256_permute2f128_si256(v[7], v[5], 0x02); + v1[4] = _mm256_permute2f128_si256(v[2], v[0], 0x13); + v1[5] = _mm256_permute2f128_si256(v[3], v[1], 0x13); + v1[6] = _mm256_permute2f128_si256(v[6], v[4], 0x13); + v1[7] = _mm256_permute2f128_si256(v[7], v[5], 0x13); +} + +inline void transpose_16x4_32bit(__m512i * r, __m512i * d) { + + static const __m512i index1 = _mm512_set_epi32( + 0x0f, 0x0b, 0x07, 0x03, + 0x0e, 0x0a, 0x06, 0x02, + 0x0d, 0x09, 0x05, 0x01, + 0x0c, 0x08, 0x04, 0x00); + + d[0] = _mm512_permutexvar_epi32(index1, r[0]); + d[1] = _mm512_permutexvar_epi32(index1, r[1]); + d[2] = _mm512_permutexvar_epi32(index1, r[2]); + d[3] = _mm512_permutexvar_epi32(index1, r[3]); + + r[0] = _mm512_shuffle_i32x4(d[0], d[1], 0x44); + r[1] = _mm512_shuffle_i32x4(d[0], d[1], 0xee); + r[2] = _mm512_shuffle_i32x4(d[2], d[3], 0x44); + r[3] = _mm512_shuffle_i32x4(d[2], d[3], 0xee); + + d[0] = _mm512_shuffle_i32x4(r[0], r[2], 0x88); + d[1] = _mm512_shuffle_i32x4(r[0], r[2], 0xdd); + d[2] = _mm512_shuffle_i32x4(r[1], r[3], 0x88); + d[3] = _mm512_shuffle_i32x4(r[1], r[3], 0xdd); +} + +inline void transpose_16x16_32bit(__m512i * v) { + __m512i v1[16]; + v1[0] = _mm512_unpacklo_epi32(v[0], v[1]); + v1[1] = _mm512_unpackhi_epi32(v[0], v[1]); + v1[2] = _mm512_unpacklo_epi32(v[2], v[3]); + v1[3] = _mm512_unpackhi_epi32(v[2], v[3]); + v1[4] = _mm512_unpacklo_epi32(v[4], v[5]); + v1[5] = _mm512_unpackhi_epi32(v[4], v[5]); + v1[6] = _mm512_unpacklo_epi32(v[6], v[7]); + v1[7] = _mm512_unpackhi_epi32(v[6], v[7]); + v1[8] = _mm512_unpacklo_epi32(v[8], v[9]); + v1[9] = _mm512_unpackhi_epi32(v[8], v[9]); + v1[10] = _mm512_unpacklo_epi32(v[10], v[11]); + v1[11] = _mm512_unpackhi_epi32(v[10], v[11]); + v1[12] = _mm512_unpacklo_epi32(v[12], v[13]); + v1[13] = _mm512_unpackhi_epi32(v[12], v[13]); + v1[14] = _mm512_unpacklo_epi32(v[14], v[15]); + v1[15] = _mm512_unpackhi_epi32(v[14], v[15]); + + v[0] = _mm512_unpacklo_epi64(v1[0], v1[2]); + v[1] = _mm512_unpackhi_epi64(v1[0], v1[2]); + v[2] = _mm512_unpacklo_epi64(v1[1], v1[3]); + v[3] = _mm512_unpackhi_epi64(v1[1], v1[3]); + v[4] = _mm512_unpacklo_epi64(v1[4], v1[6]); + v[5] = _mm512_unpackhi_epi64(v1[4], v1[6]); + v[6] = _mm512_unpacklo_epi64(v1[5], v1[7]); + v[7] = _mm512_unpackhi_epi64(v1[5], v1[7]); + v[8] = _mm512_unpacklo_epi64(v1[8], v1[10]); + v[9] = _mm512_unpackhi_epi64(v1[8], v1[10]); + v[10] = _mm512_unpacklo_epi64(v1[9], v1[11]); + v[11] = _mm512_unpackhi_epi64(v1[9], v1[11]); + v[12] = _mm512_unpacklo_epi64(v1[12], v1[14]); + v[13] = _mm512_unpackhi_epi64(v1[12], v1[14]); + v[14] = _mm512_unpacklo_epi64(v1[13], v1[15]); + v[15] = _mm512_unpackhi_epi64(v1[13], v1[15]); + + v1[0] = _mm512_shuffle_i32x4(v[0], v[4], 0x88); + v1[1] = _mm512_shuffle_i32x4(v[1], v[5], 0x88); + v1[2] = _mm512_shuffle_i32x4(v[2], v[6], 0x88); + v1[3] = _mm512_shuffle_i32x4(v[3], v[7], 0x88); + v1[4] = _mm512_shuffle_i32x4(v[0], v[4], 0xdd); + v1[5] = _mm512_shuffle_i32x4(v[1], v[5], 0xdd); + v1[6] = _mm512_shuffle_i32x4(v[2], v[6], 0xdd); + v1[7] = _mm512_shuffle_i32x4(v[3], v[7], 0xdd); + v1[8] = _mm512_shuffle_i32x4(v[8], v[12], 0x88); + v1[9] = _mm512_shuffle_i32x4(v[9], v[13], 0x88); + v1[10] = _mm512_shuffle_i32x4(v[10], v[14], 0x88); + v1[11] = _mm512_shuffle_i32x4(v[11], v[15], 0x88); + v1[12] = _mm512_shuffle_i32x4(v[8], v[12], 0xdd); + v1[13] = _mm512_shuffle_i32x4(v[9], v[13], 0xdd); + v1[14] = _mm512_shuffle_i32x4(v[10], v[14], 0xdd); + v1[15] = _mm512_shuffle_i32x4(v[11], v[15], 0xdd); + + v[0] = _mm512_shuffle_i32x4(v1[0], v1[8], 0x88); + v[1] = _mm512_shuffle_i32x4(v1[1], v1[9], 0x88); + v[2] = _mm512_shuffle_i32x4(v1[2], v1[10], 0x88); + v[3] = _mm512_shuffle_i32x4(v1[3], v1[11], 0x88); + v[4] = _mm512_shuffle_i32x4(v1[4], v1[12], 0x88); + v[5] = _mm512_shuffle_i32x4(v1[5], v1[13], 0x88); + v[6] = _mm512_shuffle_i32x4(v1[6], v1[14], 0x88); + v[7] = _mm512_shuffle_i32x4(v1[7], v1[15], 0x88); + v[8] = _mm512_shuffle_i32x4(v1[0], v1[8], 0xdd); + v[9] = _mm512_shuffle_i32x4(v1[1], v1[9], 0xdd); + v[10] = _mm512_shuffle_i32x4(v1[2], v1[10], 0xdd); + v[11] = _mm512_shuffle_i32x4(v1[3], v1[11], 0xdd); + v[12] = _mm512_shuffle_i32x4(v1[4], v1[12], 0xdd); + v[13] = _mm512_shuffle_i32x4(v1[5], v1[13], 0xdd); + v[14] = _mm512_shuffle_i32x4(v1[6], v1[14], 0xdd); + v[15] = _mm512_shuffle_i32x4(v1[7], v1[15], 0xdd); +} + +void quantize_row_q8_K_vnni(const float * RESTRICT x, void * RESTRICT vy, int64_t k) { + assert(k % QK_K == 0); + const int KB = k / QK_K; + constexpr int kVecs = QK_K / 16; + + block_q8_K * y = reinterpret_cast(vy); + + // hold 16 float vecs from x + __m512 v[kVecs]; + + // hold the quants vecs + __m512i vq[kVecs / 4]; + + // hold the packed quants vecs + __m512i vq_packed[kVecs / 4]; + + const __m512 signBit = _mm512_set1_ps(-0.f); + + for (int i = 0; i < KB; ++i) { + // Compute max(abs(e)) for the block + __m512 vamax = _mm512_set1_ps(0.f); + for (int j = 0; j < kVecs; ++j) { + v[j] = _mm512_loadu_ps(x); x += 16; + vamax = _mm512_max_ps(vamax, _mm512_andnot_ps(signBit, v[j])); + } + const float amax = _mm512_reduce_max_ps(vamax); + + // Quantize these floats + const float iscale = 127.f / amax; + y[i].d = GGML_FP32_TO_FP16(1 / iscale); + const float id = ( amax != 0.0f ) ? iscale : 0.f; + const __m512 vscale = _mm512_set1_ps(id); + + // Apply multiplier and round to nearest integer + for (int j = 0; j < kVecs; ++j) { + v[j] = _mm512_mul_ps(v[j], vscale); + v[j] = _mm512_roundscale_ps(v[j], (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + } + + // Pack to epi8 vecs + for (int j = 0; j < kVecs / 4; ++j) { + __m128i q8_0 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 0])); + __m128i q8_1 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 1])); + __m128i q8_2 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 2])); + __m128i q8_3 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 3])); + + __m256i q8_01 = _mm256_insertf128_si256(_mm256_castsi128_si256(q8_0), (q8_1), 1); + __m256i q8_23 = _mm256_insertf128_si256(_mm256_castsi128_si256(q8_2), (q8_3), 1); + + vq[j] = _mm512_inserti32x8(_mm512_castsi256_si512(q8_01), q8_23, 1); + _mm512_storeu_si512((__m512i *)(y[i].qs + j * 64), vq[j]); + } + + // Compute the bsums with vnni + transpose_16x4_32bit(vq, vq_packed); + + const __m512i one = _mm512_set1_epi8(1); + __m512i sum = _mm512_setzero_si512(); + for (int k = 0; k < 4; ++k) { + sum = _mm512_dpbusd_epi32(sum, one, vq_packed[k]); + } + _mm256_storeu_si256((__m256i *)(y[i].bsums), _mm512_cvtepi32_epi16(sum)); + } +} + +// quantize A from float to `vec_dot_type` +template +inline void from_float(const float * x, char * vy, int64_t k); + +template <> +inline void from_float(const float * x, char * vy, int64_t k) { + quantize_row_q8_0(x, (block_q8_0 *)vy, k); +} + +template <> +inline void from_float(const float * x, char * vy, int64_t k) { + quantize_row_q8_1(x, (block_q8_1 *)vy, k); +} + +template <> +inline void from_float(const float * x, char * vy, int64_t k) { +#if 1 + // TODO: this is reference impl! + quantize_row_q8_K_ref(x, (block_q8_K *)vy, k); +#else + quantize_row_q8_K_vnni(x, vy, k); +#endif +} + +// load A from memory to array when nrows can not fill in whole tile +void unpack_A(int8_t * RESTRICT tile, const block_q8_0 * RESTRICT A, int lda, int nr) { + assert(nr != TILE_M); + for (int m = 0; m < nr; ++m) { + const __m256i v = _mm256_loadu_si256((const __m256i *)(A[m * lda].qs)); + _mm256_storeu_si256((__m256i *)(tile + m * TILE_K), v); + } +} + +void unpack_A(int8_t * RESTRICT tile, const block_q8_1 * RESTRICT A, int lda, int nr) { + assert(nr != TILE_M); + for (int m = 0; m < nr; ++m) { + const __m256i v = _mm256_loadu_si256((const __m256i *)(A[m * lda].qs)); + _mm256_storeu_si256((__m256i *)(tile + m * TILE_K), v); + } +} + +template +void unpack_A(int8_t * RESTRICT tile, const block_q8_K * RESTRICT A, int lda, int k, int nr) { + assert(nr <= TILE_M); + for (int m = 0; m < nr; ++m) { + const __m256i v = _mm256_loadu_si256((const __m256i *)(A[m * lda].qs + k * 32)); + _mm256_storeu_si256((__m256i *)(tile + m * TILE_K), v); + } +} + +template <> +void unpack_A(int8_t * RESTRICT tile, const block_q8_K * RESTRICT A, int lda, int k, int nr) { + assert(nr <= TILE_M); + // zero padding k from 16 to 32, so that we don't have to re-config amx + const __m128i zero = _mm_setzero_si128(); + for (int m = 0; m < nr; ++m) { + const __m128i v = _mm_loadu_si128((const __m128i *)(A[m * lda].qs + k * 16)); + const __m256i r = _mm256_insertf128_si256(_mm256_castsi128_si256(v), zero, 1); + _mm256_storeu_si256((__m256i *)(tile + m * TILE_K), r); + } +} + +#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1) +inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) { + const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi); + const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp); + const __m256i lowMask = _mm256_set1_epi8(0xF); + return _mm256_and_si256(lowMask, bytes); +} + +// used for block_q4_K +inline __m512i bytes_from_nibbles_64(const uint8_t * rsi) { + const __m256i tmp = _mm256_loadu_si256((const __m256i *)rsi); + const __m256i lowMask = _mm256_set1_epi8(0xF); + const __m256i q4l = _mm256_and_si256(tmp, lowMask); + const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(tmp, 4), lowMask); + return _mm512_inserti32x8(_mm512_castsi256_si512(q4l), q4h, 1); +} + +// used for block_q5_K +inline __m512i bytes_from_nibbles_64(const uint8_t * qs, const uint8_t * qh, int k) { + const __m256i lowMask = _mm256_set1_epi8(0xF); + __m256i hmask = _mm256_set1_epi8(1); + hmask = _mm256_slli_epi16(hmask, k); + + const __m256i q5bits = _mm256_loadu_si256((const __m256i *)qs); + const __m256i hbits = _mm256_loadu_si256((const __m256i *)qh); + + const __m256i q5l_0 = _mm256_and_si256(q5bits, lowMask); + const __m256i q5h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), k + 0), 4); + const __m256i q5_0 = _mm256_add_epi8(q5l_0, q5h_0); + hmask = _mm256_slli_epi16(hmask, 1); + + const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), lowMask); + const __m256i q5h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), k + 1), 4); + const __m256i q5_1 = _mm256_add_epi8(q5l_1, q5h_1); + + return _mm512_inserti32x8(_mm512_castsi256_si512(q5_0), q5_1, 1); +} + +// used for block_q6_K +inline void bytes_from_nibbles_128(__m512i& r0, __m512i& r1, const uint8_t * qs, const uint8_t * qh) { + const __m256i m4 = _mm256_set1_epi8(0xF); + const __m256i m2 = _mm256_set1_epi8(0x3); + + const __m256i q6bits1 = _mm256_loadu_si256((const __m256i *)qs); + const __m256i q6bits2 = _mm256_loadu_si256((const __m256i *)(qs + 32)); + const __m256i q6bitsH = _mm256_loadu_si256((const __m256i *)qh); + + const __m256i q6h_0 = _mm256_slli_epi16(_mm256_and_si256( q6bitsH, m2), 4); + const __m256i q6h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q6bitsH, 2), m2), 4); + const __m256i q6h_2 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q6bitsH, 4), m2), 4); + const __m256i q6h_3 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q6bitsH, 6), m2), 4); + + const __m256i q6_0 = _mm256_or_si256(_mm256_and_si256(q6bits1, m4), q6h_0); + const __m256i q6_1 = _mm256_or_si256(_mm256_and_si256(q6bits2, m4), q6h_1); + const __m256i q6_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q6bits1, 4), m4), q6h_2); + const __m256i q6_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q6bits2, 4), m4), q6h_3); + + r0 = _mm512_inserti32x8(_mm512_castsi256_si512(q6_0), q6_1, 1); + r1 = _mm512_inserti32x8(_mm512_castsi256_si512(q6_2), q6_3, 1); +} + +inline __m512i packNibbles(__m512i r0, __m512i r1) { + return _mm512_or_si512(r0, _mm512_slli_epi16(r1, 4)); +} + +template +inline void pack_qs(void * RESTRICT packed_B, const TB * RESTRICT B, int KB) { + int8_t tmp[8 * 64]; + __m256i v[8], v2[8]; + for (int n = 0; n < 8; ++n) { + v[n] = bytes_from_nibbles_32(B[n * KB].qs); + } + transpose_8x8_32bit(v, v2); + for (int n = 0; n < 8; ++n) { + _mm256_storeu_si256((__m256i *)(tmp + n * 64), v2[n]); + } + for (int n = 0; n < 8; ++n) { + v[n] = bytes_from_nibbles_32(B[(n + 8) * KB].qs); + } + transpose_8x8_32bit(v, v2); + for (int n = 0; n < 8; ++n) { + _mm256_storeu_si256((__m256i *)(tmp + n * 64 + 32), v2[n]); + } + + // pack again with 128 to fully utilize vector length + for (int n = 0; n < 8; n += 2) { + __m512i r0 = _mm512_loadu_si512((const __m512i *)(tmp + n * 64)); + __m512i r1 = _mm512_loadu_si512((const __m512i *)(tmp + n * 64 + 64)); + __m512i r1r0 = packNibbles(r0, r1); + _mm512_storeu_si512((__m512i *)((char *)packed_B + n * 32), r1r0); + } +} + +template <> +inline void pack_qs(void * RESTRICT packed_B, const block_q8_0 * RESTRICT B, int KB) { + __m256i v[8], v2[8]; + for (int n = 0; n < 8; ++n) { + v[n] = _mm256_loadu_si256((const __m256i *)(B[n * KB].qs)); + } + transpose_8x8_32bit(v, v2); + for (int n = 0; n < 8; ++n) { + _mm256_storeu_si256((__m256i *)((char *)packed_B + n * 64), v2[n]); + } + for (int n = 0; n < 8; ++n) { + v[n] = _mm256_loadu_si256((const __m256i *)(B[(n + 8) * KB].qs)); + } + transpose_8x8_32bit(v, v2); + for (int n = 0; n < 8; ++n) { + _mm256_storeu_si256((__m256i *)((char *)packed_B + n * 64 + 32), v2[n]); + } +} + +template <> +inline void pack_qs(void * RESTRICT packed_B, const block_q4_K * RESTRICT B, int KB) { + __m512i v[16]; + // QK_K 256 with 8 groups, handle 2 groups at a time + char * pb = (char *)packed_B; + for (int k = 0; k < QK_K / 64; ++k) { + // pack 2 groups { n, g, k} to {g, k/4, 4n} + // e.g. {16, 2, 32} to {2, 8, 64} + for (int n = 0; n < TILE_N; ++n) { + v[n] = bytes_from_nibbles_64(B[n * KB].qs + k * 32); + } + + transpose_16x16_32bit(v); + + // pack again with 128 to fully utilize vector length + for (int n = 0; n < TILE_N; n += 2) { + _mm512_storeu_si512((__m512i *)pb, packNibbles(v[n], v[n + 1])); + pb += 64; + } + } +} + +template <> +inline void pack_qs(void * RESTRICT packed_B, const block_q5_K * RESTRICT B, int KB) { + __m512i v[16]; + const __m512i lowMask = _mm512_set1_epi8(0xF); + // QK_K 256 with 8 groups, handle 2 groups at a time + char * pb = (char *)packed_B; + char * ph = (char *)packed_B + (QK_K / 2) * TILE_N; + for (int k = 0; k < QK_K / 64; ++k) { + // pack 2 groups { n, g, k} to {g, k/4, 4n} + // e.g. {16, 2, 32} to {2, 8, 64} + for (int n = 0; n < TILE_N; ++n) { + v[n] = bytes_from_nibbles_64(B[n * KB].qs + k * 32, B[n * KB].qh, /* group */2 * k); + } + + transpose_16x16_32bit(v); + + // 1. pack lower 4bits with 2 groups + for (int n = 0; n < TILE_N; n += 2) { + // get lower 4 bits + const __m512i r0 = _mm512_and_si512(v[n], lowMask); + const __m512i r1 = _mm512_and_si512(v[n + 1], lowMask); + _mm512_storeu_si512((__m512i *)pb, packNibbles(r0, r1)); pb += 64; + } + + // 2. pack higher 1bit with 2 groups + const __m512i hmask = _mm512_set1_epi8(0x10); + for (int g = 0; g < 2; ++g) { + __m512i hbits = _mm512_setzero_si512(); + hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 0], hmask), 4)); + hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 1], hmask), 3)); + hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 2], hmask), 2)); + hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 3], hmask), 1)); + hbits = _mm512_add_epi8(hbits, _mm512_and_si512(v[g * 8 + 4], hmask) ); + hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 8 + 5], hmask), 1)); + hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 8 + 6], hmask), 2)); + hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 8 + 7], hmask), 3)); + _mm512_storeu_si512((__m512i *)ph, hbits); ph += 64; + } + } +} + +template <> +inline void pack_qs(void * RESTRICT packed_B, const block_q6_K * RESTRICT B, int KB) { + __m512i v[32]; + const __m512i lowMask = _mm512_set1_epi8(0xF); + // QK_K 256 with 8 groups, handle 4 groups at a time + char * pb = (char *)packed_B; + char * ph = (char *)packed_B + (QK_K / 2) * TILE_N; + for (int k = 0; k < QK_K / 128; ++k) { + for (int n = 0; n < TILE_N; ++n) { + bytes_from_nibbles_128(v[n], v[n + 16], B[n * KB].ql + k * 64, B[n * KB].qh + k * 32); + } + + // top half: group 0,1 or 4,5; bottom half: group 2,3 or 6,7 + transpose_16x16_32bit(v); + transpose_16x16_32bit(v + 16); + + // 1. pack lower 4bits with 4 groups + for (int n = 0; n < 32; n += 2) { + const __m512i r0 = _mm512_and_si512(v[n], lowMask); + const __m512i r1 = _mm512_and_si512(v[n + 1], lowMask); + _mm512_storeu_si512((__m512i *)pb, packNibbles(r0, r1)); pb += 64; + } + + // 2. pack higher 2bit with 4 groups + const __m512i hmask = _mm512_set1_epi8(0x30); + for (int g = 0; g < 8; ++g) { + __m512i hbits = _mm512_setzero_si512(); + hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 4 + 0], hmask), 4)); + hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 4 + 1], hmask), 2)); + hbits = _mm512_add_epi8(hbits, _mm512_and_si512(v[g * 4 + 2], hmask) ); + hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 4 + 3], hmask), 2)); + _mm512_storeu_si512((__m512i *)ph, hbits); ph += 64; + } + } +} + +template <> +inline void pack_qs(void * RESTRICT packed_B, const block_iq4_xs * RESTRICT B, int KB) { + __m512i v[16]; + char * pb = (char *)packed_B; + for (int k = 0; k < QK_K / 64; ++k) { + for (int n = 0; n < TILE_N; ++n) { + __m256i r0 = bytes_from_nibbles_32(B[n * KB].qs + k * 32 + 0); + __m256i r1 = bytes_from_nibbles_32(B[n * KB].qs + k * 32 + 16); + v[n] = _mm512_inserti32x8(_mm512_castsi256_si512(r0), r1, 1); + } + + transpose_16x16_32bit(v); + + // pack again with 128 to fully utilize vector length + for (int n = 0; n < TILE_N; n += 2) { + _mm512_storeu_si512((__m512i *)pb, packNibbles(v[n], v[n + 1])); + pb += 64; + } + } +} + +// pack B to vnni formats in 4bits or 8 bits +void pack_B(void * RESTRICT packed_B, const block_q4_0 * RESTRICT B, int KB) { + pack_qs(packed_B, B, KB); + ggml_half * d0 = reinterpret_cast((char *)packed_B + TILE_N * TILE_K / 2); + for (int n = 0; n < TILE_N; ++n) { + d0[n] = B[n * KB].d; + } +} + +void pack_B(void * RESTRICT packed_B, const block_q4_1 * RESTRICT B, int KB) { + pack_qs(packed_B, B, KB); + ggml_half * d0 = reinterpret_cast((char *)packed_B + TILE_N * TILE_K / 2); + ggml_half * m0 = d0 + TILE_N; + for (int n = 0; n < TILE_N; ++n) { + d0[n] = B[n * KB].d; + m0[n] = B[n * KB].m; + } +} + +inline void s8s8_compensation(void * RESTRICT packed_B) { + // packed_B layout: + // quants {TILE_N, TILEK} int8_t + // d0 {TILE_N} ggml_half + // comp {TILE_N} int32_t + const int offset = TILE_N * TILE_K + TILE_N * sizeof(ggml_half); + __m512i vcomp = _mm512_setzero_si512(); + const __m512i off = _mm512_set1_epi8(static_cast(0x80)); + for (int k = 0; k < 8; ++k) { + __m512i vb = _mm512_loadu_si512((const __m512i *)((const char *)packed_B + k * 64)); + vcomp = _mm512_dpbusd_epi32(vcomp, off, vb); + } + _mm512_storeu_si512((__m512i *)((char *)(packed_B) + offset), vcomp); +} + +void pack_B(void * RESTRICT packed_B, const block_q8_0 * RESTRICT B, int KB) { + pack_qs(packed_B, B, KB); + ggml_half * d0 = reinterpret_cast((char *)packed_B + TILE_N * TILE_K); + for (int n = 0; n < TILE_N; ++n) { + d0[n] = B[n * KB].d; + } + s8s8_compensation(packed_B); +} + +// convert 8 * {min, scale} from int6 to int8 +inline void unpack_mins_and_scales(const uint8_t * scales, uint32_t * utmp) { + const uint32_t kmask1 = 0x3f3f3f3f; + const uint32_t kmask2 = 0x0f0f0f0f; + const uint32_t kmask3 = 0x03030303; + + memcpy(utmp, scales, 12); + utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); + const uint32_t uaux = utmp[1] & kmask1; + utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); + utmp[2] = uaux; + utmp[0] &= kmask1; +} + +// packed_B layout: +// quants {8, TILE_N, 16} uint8 +// scales {8, TILE_N} uint8 +// mins {8, TILE_N} uint8 +// d {TILE_N} ggml_half +// dmin {TILE_N} ggml_half +void pack_B(void * RESTRICT packed_B, const block_q4_K * RESTRICT B, int KB) { + pack_qs(packed_B, B, KB); + + uint8_t * scales = reinterpret_cast((char *)packed_B + (QK_K / 2) * TILE_N); + uint8_t * mins = scales + 8 * TILE_N; + ggml_half * d = reinterpret_cast(mins + 8 * TILE_N); + ggml_half * dmin = d + TILE_N; + + union { + uint32_t u32[4]; + uint8_t u8[16]; + } s; + + for (int n = 0; n < TILE_N; ++n) { + unpack_mins_and_scales(B[n * KB].scales, s.u32); + for (int k = 0; k < 8; ++k) { + scales[k * TILE_N + n] = s.u8[k]; + mins[(k >> 1) * TILE_N * 2 + n * 2 + (k & 0x1)] = s.u8[k + 8]; + } + d[n] = B[n * KB].d; + dmin[n] = B[n * KB].dmin; + } +} + +// packed_B layout: +// quants {8, TILE_N, 16} uint8 +// qh {8, TILE_N, 4} uint8 +// scales {8, TILE_N} uint8 +// mins {8, TILE_N} uint8 +// d {TILE_N} ggml_half +// dmin {TILE_N} ggml_half +void pack_B(void * RESTRICT packed_B, const block_q5_K * RESTRICT B, int KB) { + pack_qs(packed_B, B, KB); + + uint8_t * scales = reinterpret_cast((char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N); + uint8_t * mins = scales + 8 * TILE_N; + ggml_half * d = reinterpret_cast(mins + 8 * TILE_N); + ggml_half * dmin = d + TILE_N; + + union { + uint32_t u32[4]; + uint8_t u8[16]; + } s; + + for (int n = 0; n < TILE_N; ++n) { + unpack_mins_and_scales(B[n * KB].scales, s.u32); + for (int k = 0; k < 8; ++k) { + scales[k * TILE_N + n] = s.u8[k]; + mins[(k >> 1) * TILE_N * 2 + n * 2 + (k & 0x1)] = s.u8[k + 8]; + } + d[n] = B[n * KB].d; + dmin[n] = B[n * KB].dmin; + } +} + +// packed_B layout: +// quants {16, TILE_N, 8} uint8 +// qh {16, TILE_N, 4} uint8 +// scales {16, TILE_N} uint8 +// d {TILE_N} ggml_half +void pack_B(void * RESTRICT packed_B, const block_q6_K * RESTRICT B, int KB) { + pack_qs(packed_B, B, KB); + + uint8_t * scales = reinterpret_cast((char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N); + ggml_half * d = reinterpret_cast(scales + 16 * TILE_N); + for (int n = 0; n < TILE_N; ++n) { + const int8_t * ps = B[n * KB].scales; + for (int k = 0; k < 16; ++k) { + scales[k * TILE_N + n] = ps[k]; + } + d[n] = B[n * KB].d; + } +} + +// packed_B layout: +// quants {8, TILE_N, 16} uint8 +// scales {8, TILE_N} int8 +// d {TILE_N} ggml_half +void pack_B(void * RESTRICT packed_B, const block_iq4_xs * RESTRICT B, int KB) { + pack_qs(packed_B, B, KB); + + int8_t * scales = reinterpret_cast((char *)packed_B + (QK_K / 2) * TILE_N); + ggml_half * d = reinterpret_cast(scales + 8 * TILE_N); + + // pack the scales + for (int n = 0; n < TILE_N; ++n) { + uint16_t sh = B[n * KB].scales_h; + for (int k = 0; k < 8; k += 2) { + const int16_t ls1 = ((B[n * KB].scales_l[k / 2] & 0xf) | ((sh << 4) & 0x30)) - 32; + const int16_t ls2 = ((B[n * KB].scales_l[k / 2] >> 4) | ((sh << 2) & 0x30)) - 32; + scales[(k + 0) * TILE_N + n] = ls1; + scales[(k + 1) * TILE_N + n] = ls2; + sh >>= 4; + } + d[n] = B[n * KB].d; + } +} + +template> +void unpack_B(packed_B_t * RESTRICT tile, const void * RESTRICT packed_B) { + GGML_UNUSED(tile); + GGML_UNUSED(packed_B); +} + +template <> +void unpack_B(int8_t * RESTRICT tile, const void * RESTRICT packed_B) { + const __m512i off = _mm512_set1_epi8(8); + const __m512i lowMask = _mm512_set1_epi8(0xF); + for (int n = 0; n < 8; n += 2) { + __m512i bytes = _mm512_loadu_si512((const __m512i *)((const char *)packed_B + n * 32)); + const __m512i r0 = _mm512_sub_epi8(_mm512_and_si512(bytes, lowMask), off); + const __m512i r1 = _mm512_sub_epi8(_mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask), off); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1); + } +} + +template <> +void unpack_B(uint8_t * RESTRICT tile, const void * RESTRICT packed_B) { + const __m512i lowMask = _mm512_set1_epi8(0xF); + for (int n = 0; n < 8; n += 2) { + __m512i bytes = _mm512_loadu_si512((const __m512i *)((const char *)packed_B + n * 32)); + const __m512i r0 = _mm512_and_si512(bytes, lowMask); + const __m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1); + } +} + +// packed_B_t for QKK is int8_t +template +void unpack_B(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) { + const int packed_B_group_size = QK_K / 2 * TILE_N / 8; + const char * packed_B_group = (const char *)packed_B + k * packed_B_group_size; + const __m512i lowMask = _mm512_set1_epi8(0xF); + for (int n = 0; n < 8; n += 2) { + __m512i bytes = _mm512_loadu_si512(packed_B_group + n * 32); + const __m512i r0 = _mm512_and_si512(bytes, lowMask); + const __m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1); + } +} + +template <> +void unpack_B(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) { + // lower 4bits, stride 256 bytes + const int packed_l4_group_size = QK_K / 2 * TILE_N / 8; + const char * pb = (const char *)packed_B + k * packed_l4_group_size; + + // higher 1bit, stride 64 bytes + const int packed_h1_group_size = QK_K / 8 * TILE_N / 8; + const char * ph = (const char *)packed_B + (QK_K / 2) * TILE_N + k * packed_h1_group_size; + const __m512i hbits = _mm512_loadu_si512(ph); + + const __m512i lowMask = _mm512_set1_epi8(0xF); + __m512i hmask0 = _mm512_set1_epi8(0x1); + __m512i hmask1 = _mm512_set1_epi8(0x2); + + for (int n = 0; n < 8; n += 2) { + __m512i bytes = _mm512_loadu_si512(pb + n * 32); + __m512i r0 = _mm512_and_si512(bytes, lowMask); + __m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + __m512i h0 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask0), n), 4); + __m512i h1 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask1), n + 1), 4); + + hmask0 = _mm512_slli_epi16(hmask0, 2); + hmask1 = _mm512_slli_epi16(hmask1, 2); + r0 = _mm512_add_epi8(r0, h0); + r1 = _mm512_add_epi8(r1, h1); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1); + } +} + +template <> +void unpack_B(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) { + // lower 4bits, stride 128 bytes + const int packed_l4_group_size = QK_K / 2 * TILE_N / 16; + const char * pb = (const char *)packed_B + k * packed_l4_group_size; + + // higher 2bits, stride 64 bytes + const int packed_h2_group_size = QK_K / 4 * TILE_N / 16; + const char * ph = (const char *)packed_B + (QK_K / 2) * TILE_N + k * packed_h2_group_size; + const __m512i hbits = _mm512_loadu_si512(ph); + + const __m512i off = _mm512_set1_epi8(32); + const __m512i lowMask = _mm512_set1_epi8(0xF); + __m512i hmask0 = _mm512_set1_epi8(0x3); // 0011 + __m512i hmask1 = _mm512_set1_epi8(0xC); // 1100 + + // notes: skip zero padding from row4 to row7 as we have done so in `unpack_A` + __m512i bytes = _mm512_loadu_si512(pb); + __m512i r0 = _mm512_and_si512(bytes, lowMask); + __m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + __m512i h0 = _mm512_slli_epi16(_mm512_and_si512(hbits, hmask0), 4); + __m512i h1 = _mm512_slli_epi16(_mm512_and_si512(hbits, hmask1), 2); + _mm512_storeu_si512((__m512i *)(tile + 0), _mm512_sub_epi8(_mm512_add_epi8(r0, h0), off)); + _mm512_storeu_si512((__m512i *)(tile + 64), _mm512_sub_epi8(_mm512_add_epi8(r1, h1), off)); + + hmask0 = _mm512_slli_epi16(hmask0, 4); + hmask1 = _mm512_slli_epi16(hmask1, 4); + + bytes = _mm512_loadu_si512(pb + 64); + r0 = _mm512_and_si512(bytes, lowMask); + r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + h0 = _mm512_and_si512(hbits, hmask0); + h1 = _mm512_srli_epi16(_mm512_and_si512(hbits, hmask1), 2); + _mm512_storeu_si512((__m512i *)(tile + 128), _mm512_sub_epi8(_mm512_add_epi8(r0, h0), off)); + _mm512_storeu_si512((__m512i *)(tile + 192), _mm512_sub_epi8(_mm512_add_epi8(r1, h1), off)); +} + +template <> +void unpack_B(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) { + static const __m512i values128 = _mm512_set_epi8( + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127, + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127, + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127, + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127 + ); + + const int packed_B_group_size = QK_K / 2 * TILE_N / 8; + const char * pb = (const char *)packed_B + k * packed_B_group_size; + const __m512i lowMask = _mm512_set1_epi8(0xF); + + for (int n = 0; n < 8; n += 2) { + __m512i bytes = _mm512_loadu_si512(pb + n * 32); + const __m512i r0 = _mm512_shuffle_epi8(values128, _mm512_and_si512(bytes, lowMask)); + const __m512i r1 = _mm512_shuffle_epi8(values128, _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask)); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0); + _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1); + } +} + +template +struct acc_C {}; + +template +struct acc_C { + static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_0 * A, int lda, const void * packed_B, int nr) { + const int offset = TILE_N * TILE_K / 2; + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset))); + + for (int m = 0; m < nr; ++m) { + const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d)); + const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); + + __m512 vsum; + if (is_acc) { + vsum = _mm512_loadu_ps(C + m * ldc); + } else { + vsum = _mm512_set1_ps(0.f); + } + vsum = _mm512_fmadd_ps(vtile, _mm512_mul_ps(vd0, vd1), vsum); + _mm512_storeu_ps(C + m * ldc, vsum); + } + } +}; + +template +struct acc_C { + static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_1 * A, int lda, const void * packed_B, int nr) { + const int offset = TILE_N * TILE_K / 2; + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset))); + const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset + TILE_N * sizeof(ggml_half)))); + + for (int m = 0; m < nr; ++m) { + const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d)); + const __m512 vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].s)); + const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); + + __m512 vsum; + if (is_acc) { + vsum = _mm512_loadu_ps(C + m * ldc); + } else { + vsum = _mm512_set1_ps(0.f); + } + vsum = _mm512_fmadd_ps(vtile, _mm512_mul_ps(vd0, vd1), vsum); + vsum = _mm512_fmadd_ps(vm0, vs1, vsum); + _mm512_storeu_ps(C + m * ldc, vsum); + } + } +}; + +template +struct acc_C { + static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_0 * A, int lda, const void * packed_B, int nr) { + const int offset = TILE_N * TILE_K; + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset))); + + for (int m = 0; m < nr; ++m) { + const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d)); + const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); + + __m512 vsum; + if (is_acc) { + vsum = _mm512_loadu_ps(C + m * ldc); + } else { + vsum = _mm512_set1_ps(0.f); + } + vsum = _mm512_fmadd_ps(vtile, _mm512_mul_ps(vd0, vd1), vsum); + _mm512_storeu_ps(C + m * ldc, vsum); + } + } +}; + +template +struct acc_C { + static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) { + const uint8_t * scales = reinterpret_cast((const char *)packed_B + (QK_K / 2) * TILE_N); + const uint8_t * mins = scales + 8 * TILE_N; + const ggml_half * d0 = reinterpret_cast(mins + 8 * TILE_N); + const ggml_half * dmin = d0 + TILE_N; + + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0)); + const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)dmin)); + + for (int m = 0; m < nr; ++m) { + const float d1 = A[m * lda].d; + const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0); + const __m512 vdm = _mm512_mul_ps(_mm512_set1_ps(-d1), vdmin); + const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); + + __m512 vsum; + if (is_acc) { + vsum = _mm512_loadu_ps(C + m * ldc); + } else { + vsum = _mm512_set1_ps(0.f); + } + + const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[m * lda].bsums); + const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1)); + + __m512i acc_m = _mm512_setzero_si512(); + for (int k = 0; k < 4; ++k) { + __m512i vmask = _mm512_set1_epi32(k); + __m512i va = _mm512_permutexvar_epi32(vmask, _mm512_castsi128_si512(q8s)); + __m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(mins + k * 32))); + acc_m = _mm512_dpwssds_epi32(acc_m, va, vb); + } + + vsum = _mm512_fmadd_ps(vtile, vd, vsum); + vsum = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc_m), vdm, vsum); + _mm512_storeu_ps(C + m * ldc, vsum); + } + } +}; + +template +struct acc_C { + static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) { + const uint8_t * scales = reinterpret_cast((const char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N); + const uint8_t * mins = scales + 8 * TILE_N; + const ggml_half * d0 = reinterpret_cast(mins + 8 * TILE_N); + const ggml_half * dmin = d0 + TILE_N; + + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0)); + const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)dmin)); + + for (int m = 0; m < nr; ++m) { + const float d1 = A[m * lda].d; + const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0); + const __m512 vdm = _mm512_mul_ps(_mm512_set1_ps(-d1), vdmin); + const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); + + __m512 vsum; + if (is_acc) { + vsum = _mm512_loadu_ps(C + m * ldc); + } else { + vsum = _mm512_set1_ps(0.f); + } + + const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[m * lda].bsums); + const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1)); + + __m512i acc_m = _mm512_setzero_si512(); + for (int k = 0; k < 4; ++k) { + __m512i vmask = _mm512_set1_epi32(k); + __m512i va = _mm512_permutexvar_epi32(vmask, _mm512_castsi128_si512(q8s)); + __m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(mins + k * 32))); + acc_m = _mm512_dpwssds_epi32(acc_m, va, vb); + } + + vsum = _mm512_fmadd_ps(vtile, vd, vsum); + vsum = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc_m), vdm, vsum); + _mm512_storeu_ps(C + m * ldc, vsum); + } + } +}; + +template +struct acc_C { + static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) { + const uint8_t * scales = reinterpret_cast((const char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N); + const ggml_half * d0 = reinterpret_cast(scales + 16 * TILE_N); + + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0)); + + for (int m = 0; m < nr; ++m) { + const float d1 = A[m * lda].d; + const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0); + const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); + + __m512 vsum; + if (is_acc) { + vsum = _mm512_loadu_ps(C + m * ldc); + } else { + vsum = _mm512_set1_ps(0.f); + } + + vsum = _mm512_fmadd_ps(vtile, vd, vsum); + _mm512_storeu_ps(C + m * ldc, vsum); + } + } +}; + +template +struct acc_C { + static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) { + const int8_t * scales = reinterpret_cast((const char *)packed_B + (QK_K / 2) * TILE_N); + const ggml_half * d0 = reinterpret_cast(scales + 8 * TILE_N); + + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0)); + + for (int m = 0; m < nr; ++m) { + const float d1 = A[m * lda].d; + const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0); + const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); + + __m512 vsum; + if (is_acc) { + vsum = _mm512_loadu_ps(C + m * ldc); + } else { + vsum = _mm512_set1_ps(0.f); + } + + vsum = _mm512_fmadd_ps(vtile, vd, vsum); + _mm512_storeu_ps(C + m * ldc, vsum); + } + } +}; + +template constexpr int get_quants_size(); +template <> constexpr int get_quants_size() { return (QK_K / 2) * TILE_N; } +template <> constexpr int get_quants_size() { return (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N; } +template <> constexpr int get_quants_size() { return (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N; } +template <> constexpr int get_quants_size() { return (QK_K / 2) * TILE_N; } + +// used for QKK format +template ::value, int>::type = 0> +inline void scale_C(const int32_t * RESTRICT tile, int32_t * RESTRICT sumi, const void * packed_B, int k, int nr) { + const uint8_t * scales = reinterpret_cast((const char *)packed_B + get_quants_size()); + const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(scales + k * TILE_N))); + + for (int m = 0; m < nr; ++m) { + __m512i vsumi; + if (is_acc) { + vsumi = _mm512_loadu_si512(sumi + m * TILE_N); + } else { + vsumi = _mm512_setzero_si512(); + } + __m512i vtile = _mm512_loadu_si512(tile + m * TILE_N); + vsumi = _mm512_add_epi32(vsumi, _mm512_mullo_epi32(vtile, vscale)); + _mm512_storeu_si512((__m512i *)(sumi + m * TILE_N), vsumi); + } +} + +template +struct tinygemm_kernel_avx { + static void apply(int K, const TA * RESTRICT A, const TB * RESTRICT B, TC * RESTRICT C, int ldc) { + GGML_UNUSED(K); + GGML_UNUSED(A); + GGML_UNUSED(B); + GGML_UNUSED(C); + GGML_UNUSED(ldc); + } +}; + +template +struct tinygemm_kernel_avx { + static void apply(int K, const float * RESTRICT A, const ggml_fp16_t * RESTRICT B, float * RESTRICT C, int ldc) { + constexpr int ROWS = BLOCK_M; + constexpr int COLS = BLOCK_N; + assert(BLOCK_K == 16); + + __m512 va; + __m512 vb[COLS]; + __m512 vc[ROWS * COLS]; + + auto loadc = [&](auto idx) { + vc[idx] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + auto compute = [&](auto idx, auto k) { + constexpr int row = idx / COLS; + constexpr int col = idx % COLS; + + if constexpr (col == 0) { + va = _mm512_loadu_ps(A + row * K + k); + } + if constexpr (row == 0) { + vb[col] = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(B + col * K + k))); + } + vc[idx] = _mm512_fmadd_ps(va, vb[col], vc[idx]); + }; + + for (int k = 0; k < K; k += 16) { + Unroll{}(compute, k); + } + + auto storec = [&](auto idx) { + constexpr int row = idx / COLS; + constexpr int col = idx % COLS; + C[row * ldc + col] = _mm512_reduce_add_ps(vc[idx]); + }; + Unroll{}(storec); + } +}; + +#define LAUNCH_TINYGEMM_KERNEL_AVX(MB_SIZE, NB_SIZE) \ + tinygemm_kernel_avx::apply( \ + K, (const float *)src1->data + mb_start * K, \ + (const type *)src0->data + nb_start * K, \ + (float *)dst->data + mb_start * ldc + nb_start, ldc); + + +// re-organize in the format {NB, KB, TILE_SIZE}: +#define PACKED_INDEX(n, k, KB, tile_size) (n * KB + k) * tile_size + +template +void convert_B_packed_format(void * RESTRICT packed_B, const TB * RESTRICT B, int N, int K, int n_threads) { + const int NB = N / TILE_N; + const int KB = K / BLOCK_K; + const int TILE_SIZE = get_tile_size(); + + // parallel on NB should be enough + parallel_for(n_threads, NB, [&](int begin, int end) { + for (int n = begin; n < end; ++n) { + for (int k = 0; k < KB; ++k) { + int n0 = n * TILE_N; + pack_B((char *)packed_B + PACKED_INDEX(n, k, KB, TILE_SIZE), &B[n0 * KB + k], KB); + } + } + }); +} + +template +struct tinygemm_kernel_vnni {}; + +template +struct tinygemm_kernel_vnni { + static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + + constexpr int COLS = BLOCK_N / 16; + const int TILE_SIZE = TILE_N * sizeof(block_q4_0); + + const block_q8_0 * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + __m512i va[8]; + __m512 vc[COLS]; + __m512 vd1; + + // sum of offsets, shared across COLS + // + // avx512-vnni does not have `_mm512_dpbssd_epi32`, + // need to transfrom ss to us: + // a * (b - 8) is equavilent to b * a - 8 * a + // s u u u s u s + // + __m512i vcomp; + + const __m512i off = _mm512_set1_epi8(8); + const __m512i lowMask = _mm512_set1_epi8(0xF); + + auto loadc = [&](auto col) { + vc[col] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + auto compute = [&](auto col, auto i) { + // load a and compute compensation + if constexpr (col == 0) { + const int32_t * a_ptr = reinterpret_cast(A[0 * KB + i].qs); + vcomp = _mm512_setzero_si512(); + for (int k = 0; k < 8; ++k) { + va[k] = _mm512_set1_epi32(a_ptr[k]); + vcomp = _mm512_dpbusd_epi32(vcomp, off, va[k]); + } + vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d)); + } + + // load b + __m512i vsum = _mm512_setzero_si512(); + const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE); + for (int k = 0; k < 8; k += 2) { + __m512i bytes = _mm512_loadu_si512((const __m512i *)(b_ptr + k * 32)); + __m512i vb0 = _mm512_and_si512(bytes, lowMask); + vsum = _mm512_dpbusd_epi32(vsum, vb0, va[k + 0]); + __m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + vsum = _mm512_dpbusd_epi32(vsum, vb1, va[k + 1]); + } + const int offset = TILE_N * TILE_K / 2; + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset))); + vsum = _mm512_sub_epi32(vsum, vcomp); + + vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(vsum), _mm512_mul_ps(vd0, vd1), vc[col]); + }; + + for (int i = 0; i < KB; ++i) { + Unroll{}(compute, i); + } + + //store to C + auto storec = [&](auto col) { + _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]); + }; + Unroll{}(storec); + } +}; + +template +struct tinygemm_kernel_vnni { + static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + + constexpr int COLS = BLOCK_N / 16; + const int TILE_SIZE = TILE_N * sizeof(block_q4_1); + + const block_q8_1 * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + __m512i va[8]; + __m512i vb[8]; + __m512 vc[COLS]; + __m512 vd1, vs1; + + const __m512i lowMask = _mm512_set1_epi8(0xF); + + auto loadc = [&](auto col) { + vc[col] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + auto compute = [&](auto col, auto i) { + // load a + if constexpr (col == 0) { + const int32_t * a_ptr = reinterpret_cast(A[0 * KB + i].qs); + for (int k = 0; k < 8; ++k) { + va[k] = _mm512_set1_epi32(a_ptr[k]); + } + vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d)); + vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].s)); + } + + // load b + const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE); + for (int k = 0; k < 8; k += 2) { + __m512i bytes = _mm512_loadu_si512((const __m512i *)(b_ptr + k * 32)); + vb[k + 0] = _mm512_and_si512(bytes, lowMask); + vb[k + 1] = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + } + const int offset = TILE_N * TILE_K / 2; + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset))); + const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset + TILE_N * sizeof(ggml_half)))); + + __m512i vsum = _mm512_setzero_si512(); + for (int k = 0; k < 8; ++k) { + vsum = _mm512_dpbusd_epi32(vsum, vb[k], va[k]); + } + + vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(vsum), _mm512_mul_ps(vd0, vd1), vc[col]); + vc[col] = _mm512_fmadd_ps(vm0, vs1, vc[col]); + }; + + for (int i = 0; i < KB; ++i) { + Unroll{}(compute, i); + } + + //store to C + auto storec = [&](auto col) { + _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]); + }; + Unroll{}(storec); + } +}; + +template +struct tinygemm_kernel_vnni { + static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + + constexpr int COLS = BLOCK_N / 16; + const int TILE_SIZE = TILE_N * sizeof(block_q8_0) + TILE_N * sizeof(int32_t); + + const block_q8_0 * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + __m512i va[8]; + __m512i vb[8]; + __m512 vc[COLS]; + __m512 vd1; + + // Notes: s8s8 igemm compensation in avx512-vnni + // change s8s8 to u8s8 with compensate + // a * b = (a + 128) * b - 128 * b + // s s u s u s + // + // (128 * b is pre-computed when packing B to vnni formats) + // + const __m512i off = _mm512_set1_epi8(static_cast(0x80)); + + auto loadc = [&](auto col) { + vc[col] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + auto compute = [&](auto col, auto i) { + // load a and add offset 128 + if constexpr (col == 0) { + const int32_t * a_ptr = reinterpret_cast(A[0 * KB + i].qs); + for (int k = 0; k < 8; ++k) { + va[k] = _mm512_set1_epi32(a_ptr[k]); + va[k] = _mm512_add_epi8(va[k], off); + } + vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d)); + } + + // load b + const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE); + for (int k = 0; k < 8; ++k) { + vb[k] = _mm512_loadu_si512((const __m512i *)(b_ptr + k * 64)); + } + const int offset = TILE_N * TILE_K; + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset))); + const int offset2 = TILE_N * TILE_K + TILE_N * sizeof(ggml_half); + const __m512i vcomp = _mm512_loadu_si512((const __m512i *)(b_ptr + offset2)); + + __m512i vsum = _mm512_setzero_si512(); + for (int k = 0; k < 8; ++k) { + vsum = _mm512_dpbusd_epi32(vsum, va[k], vb[k]); + } + vsum = _mm512_sub_epi32(vsum, vcomp); + + vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(vsum), _mm512_mul_ps(vd0, vd1), vc[col]); + }; + + for (int i = 0; i < KB; ++i) { + Unroll{}(compute, i); + } + + //store to C + auto storec = [&](auto col) { + _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]); + }; + Unroll{}(storec); + } +}; + +template +struct tinygemm_kernel_vnni { + static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + + constexpr int COLS = BLOCK_N / 16; + const int TILE_SIZE = TILE_N * sizeof(block_q4_K) + TILE_N * 4; + + const block_q8_K * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + // a.qs: 8 groups, 32 bytes each group (m256i) + __m512i va[8]; + // a.bsum: 8 groups, 2 bytes each group (m128i) + __m512i va_bsum; + __m512 vc[COLS]; + __m512 vd1; + + // packed_B: + const int offset_scales = (QK_K / 2) * TILE_N; + const int offset_mins = (QK_K / 2) * TILE_N + 8 * TILE_N; + const int offset_d0 = (QK_K / 2) * TILE_N + 16 * TILE_N; + const int offset_dmin = (QK_K / 2) * TILE_N + 16 * TILE_N + TILE_N * sizeof(ggml_half); + + const __m512i lowMask = _mm512_set1_epi8(0xF); + + auto loadc = [&](auto col) { + vc[col] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + // Notes: vnni formats in QK_K + // a) quants vnni format + // int8 {k/4, n, 4}, viewed as 2d {k/4, 4n}, k = 32 + // from {16, 32} to {8, 64} + // + // b) min vnni format + // int16 {k/2, n, 2}, viewed as 2d {k/2, 2n}, k = 8 + // from {16, 8} to {4, 32} + // + auto compute = [&](auto col, auto i) { + // load a + if constexpr (col == 0) { + for (int k_group = 0; k_group < QK_K / 32; ++k_group) { + va[k_group] = _mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)(A[0 * KB + i].qs + k_group * 32))); + } + const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums); + const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1)); + va_bsum = _mm512_castsi128_si512(q8s); + vd1 = _mm512_set1_ps(A[0 * KB + i].d); + } + + // step 1: accumultate the quants + __m512i acc = _mm512_setzero_si512(); + const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE); + const char * b_qs = b_ptr; + for (int k_group = 0; k_group < QK_K / 32; ++k_group) { + __m512i vsum = _mm512_setzero_si512(); + for (int k = 0; k < 8; k += 2) { + __m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 0), va[k_group]); + __m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 1), va[k_group]); + + __m512i bytes = _mm512_loadu_si512((const __m512i *)b_qs); + __m512i vb0 = _mm512_and_si512(bytes, lowMask); + vsum = _mm512_dpbusd_epi32(vsum, vb0, va0); + __m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + vsum = _mm512_dpbusd_epi32(vsum, vb1, va1); + + b_qs += 64; + } + // vacc += scale * (q8 @ q4) + const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N))); + acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale)); + } + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0))); + vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]); + + // step 2: accumulate the mins + __m512i acc_m = _mm512_setzero_si512(); + for (int k = 0; k < 4; ++k) { + __m512i vmask = _mm512_set1_epi32(k); + __m512i va = _mm512_permutexvar_epi32(vmask, va_bsum); + __m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_mins + k * 32))); + acc_m = _mm512_dpwssds_epi32(acc_m, va, vb); + } + const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_dmin))); + vc[col] = _mm512_fnmadd_ps(_mm512_cvtepi32_ps(acc_m), _mm512_mul_ps(vdmin, vd1), vc[col]); + }; + + for (int i = 0; i < KB; ++i) { + Unroll{}(compute, i); + } + + //store to C + auto storec = [&](auto col) { + _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]); + }; + Unroll{}(storec); + } +}; + +template +struct tinygemm_kernel_vnni { + static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + + constexpr int COLS = BLOCK_N / 16; + const int TILE_SIZE = TILE_N * sizeof(block_q5_K) + TILE_N * 4; + + const block_q8_K * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + // a.qs: 8 groups, 32 bytes each group (m256i) + __m512i va[8]; + // a.bsum: 8 groups, 2 bytes each group (m128i) + __m512i va_bsum; + __m512 vc[COLS]; + __m512 vd1; + + // packed_B: + const int offset_qh = (QK_K / 2) * TILE_N; + const int offset_scales = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N; + const int offset_mins = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N + 8 * TILE_N; + const int offset_d0 = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N + 16 * TILE_N; + const int offset_dmin = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N + 16 * TILE_N + TILE_N * sizeof(ggml_half); + + const __m512i lowMask = _mm512_set1_epi8(0xF); + + auto loadc = [&](auto col) { + vc[col] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + // Q5_K and Q4_K shares the same vnni formats, refer to notes above. + auto compute = [&](auto col, auto i) { + // load a + if constexpr (col == 0) { + for (int k_group = 0; k_group < QK_K / 32; ++k_group) { + va[k_group] = _mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)(A[0 * KB + i].qs + k_group * 32))); + } + const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums); + const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1)); + va_bsum = _mm512_castsi128_si512(q8s); + vd1 = _mm512_set1_ps(A[0 * KB + i].d); + } + + // step 1: accumultate the quants + __m512i acc = _mm512_setzero_si512(); + const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE); + const char * b_qs = b_ptr; + const char * b_qh = b_ptr + offset_qh; + for (int k_group = 0; k_group < QK_K / 32; ++k_group) { + __m512i vsum = _mm512_setzero_si512(); + __m512i hmask0 = _mm512_set1_epi8(0x1); + __m512i hmask1 = _mm512_set1_epi8(0x2); + __m512i hbits = _mm512_loadu_si512((const __m512i *)(b_qh + k_group * 64)); + for (int k = 0; k < 8; k += 2) { + __m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 0), va[k_group]); + __m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 1), va[k_group]); + + __m512i bytes = _mm512_loadu_si512((const __m512i *)b_qs); + __m512i vb0 = _mm512_and_si512(bytes, lowMask); + __m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + + __m512i vh0 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask0), k), 4); + __m512i vh1 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask1), k + 1), 4); + + hmask0 = _mm512_slli_epi16(hmask0, 2); + hmask1 = _mm512_slli_epi16(hmask1, 2); + vb0 = _mm512_add_epi8(vb0, vh0); + vb1 = _mm512_add_epi8(vb1, vh1); + + vsum = _mm512_dpbusd_epi32(vsum, vb0, va0); + vsum = _mm512_dpbusd_epi32(vsum, vb1, va1); + + b_qs += 64; + } + // vacc += scale * (q8 @ q5) + const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N))); + acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale)); + } + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0))); + vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]); + + // step 2: accumulate the mins + __m512i acc_m = _mm512_setzero_si512(); + for (int k = 0; k < 4; ++k) { + __m512i vmask = _mm512_set1_epi32(k); + __m512i va = _mm512_permutexvar_epi32(vmask, va_bsum); + __m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_mins + k * 32))); + acc_m = _mm512_dpwssds_epi32(acc_m, va, vb); + } + const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_dmin))); + vc[col] = _mm512_fnmadd_ps(_mm512_cvtepi32_ps(acc_m), _mm512_mul_ps(vdmin, vd1), vc[col]); + }; + + for (int i = 0; i < KB; ++i) { + Unroll{}(compute, i); + } + + //store to C + auto storec = [&](auto col) { + _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]); + }; + Unroll{}(storec); + } +}; + +template +struct tinygemm_kernel_vnni { + static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + + constexpr int COLS = BLOCK_N / 16; + const int TILE_SIZE = TILE_N * sizeof(block_q6_K); + + const block_q8_K * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + // load the 256 bytes from A to 4 avx512 vectors + __m512i va[4]; + __m512 vc[COLS]; + __m512 vd1; + + // packed_B: + const int offset_qh = (QK_K / 2) * TILE_N; + const int offset_scales = (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N; + const int offset_d0 = (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N + 16 * TILE_N; + + // compensation + __m512i vcomp; + + const __m512i m32s = _mm512_set1_epi32(32); + const __m512i lowMask = _mm512_set1_epi8(0xF); + + auto loadc = [&](auto col) { + vc[col] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + auto compute = [&](auto col, auto i) { + if constexpr (col == 0) { + // load a + va[0] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 0)); + va[1] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 64)); + va[2] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 128)); + va[3] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 192)); + + const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums); + vcomp = _mm512_mullo_epi32(_mm512_cvtepi16_epi32(q8sums), m32s); + vd1 = _mm512_set1_ps(A[0 * KB + i].d); + } + + // accmulate the quants + __m512i acc = _mm512_setzero_si512(); + const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE); + const char * b_qs = b_ptr; + const char * b_qh = b_ptr + offset_qh; + int mask = 0; + for (int k_group = 0; k_group < QK_K / 16; ++k_group) { + int r = k_group >> 2; + __m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]); + __m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]); + + __m512i vsum = _mm512_setzero_si512(); + __m512i hmask = _mm512_set1_epi8(0x3); + + __m512i bytes = _mm512_loadu_si512(b_qs); + __m512i hbits = _mm512_loadu_si512(b_qh); + __m512i vb0 = _mm512_and_si512(bytes, lowMask); + __m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + __m512i vh0 = _mm512_slli_epi16(_mm512_and_si512(hbits, hmask), 4); + __m512i vh1 = _mm512_slli_epi16(_mm512_and_si512(hbits, _mm512_slli_epi16(hmask, 2)), 2); + + vb0 = _mm512_add_epi8(vb0, vh0); + vb1 = _mm512_add_epi8(vb1, vh1); + vsum = _mm512_dpbusd_epi32(vsum, vb0, va0); + vsum = _mm512_dpbusd_epi32(vsum, vb1, va1); + b_qs += 64; + + va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]); + va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]); + + bytes = _mm512_loadu_si512(b_qs); + vb0 = _mm512_and_si512(bytes, lowMask); + vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask); + vh0 = _mm512_and_si512(hbits, _mm512_slli_epi16(hmask, 4)); + vh1 = _mm512_srli_epi16(_mm512_and_si512(hbits, _mm512_slli_epi16(hmask, 6)), 2); + vb0 = _mm512_add_epi8(vb0, vh0); + vb1 = _mm512_add_epi8(vb1, vh1); + vsum = _mm512_dpbusd_epi32(vsum, vb0, va0); + vsum = _mm512_dpbusd_epi32(vsum, vb1, va1); + b_qs += 64; + b_qh += 64; + + // B * A - 32 * A + __m512i vmask = _mm512_set1_epi32(k_group); + vsum = _mm512_sub_epi32(vsum, _mm512_permutexvar_epi32(vmask, vcomp)); + + // vacc += scale * (q8 @ q6) + const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N))); + acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale)); + } + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0))); + vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]); + }; + + for (int i = 0; i < KB; ++i) { + Unroll{}(compute, i); + } + + //store to C + auto storec = [&](int col) { + _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]); + }; + Unroll{}(storec); + } +}; + +template +struct tinygemm_kernel_vnni { + static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + + constexpr int COLS = BLOCK_N / 16; + const int TILE_SIZE = TILE_N * sizeof(block_iq4_xs) + TILE_N * 2; + + const block_q8_K * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + // load the 256 bytes from A to 4 avx512 vectors + __m512i va[4]; + __m512 vc[COLS]; + __m512 vd1; + + // packed_B: + const int offset_scales = (QK_K / 2) * TILE_N ; + const int offset_d0 = (QK_K / 2) * TILE_N + 8 * TILE_N; + + // compensation + __m512i vcomp; + + const __m256i m128s = _mm256_set1_epi16(128); + const __m512i lowMask = _mm512_set1_epi8(0xF); + + const __m512i values128 = _mm512_set_epi8( + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127, + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127, + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127, + 113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127 + ); + const __m512i off = _mm512_set1_epi8(static_cast(0x80)); + const __m512i values256 = _mm512_add_epi8(values128, off); + + auto loadc = [&](auto col) { + vc[col] = _mm512_setzero_ps(); + }; + Unroll{}(loadc); + + auto compute = [&](auto col, auto i) { + if constexpr (col == 0) { + // load a + va[0] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 0)); + va[1] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 64)); + va[2] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 128)); + va[3] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 192)); + + // compensation: 128 * A + const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums); + vcomp = _mm512_castsi256_si512(_mm256_madd_epi16(q8sums, m128s)); + vd1 = _mm512_set1_ps(A[0 * KB + i].d); + } + + // accmulate the quants + __m512i acc = _mm512_setzero_si512(); + const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE); + const char * b_qs = b_ptr; + int mask = 0; + for (int k_group = 0; k_group < QK_K / 32; ++k_group) { + int r = k_group >> 1; + __m512i vmask = _mm512_set1_epi32(k_group); + __m512i vsum = _mm512_setzero_si512(); + for (int k = 0; k < 8; k += 2) { + __m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]); + __m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]); + + __m512i bytes = _mm512_loadu_si512(b_qs); + __m512i vb0 = _mm512_shuffle_epi8(values256, _mm512_and_si512(bytes, lowMask)); + __m512i vb1 = _mm512_shuffle_epi8(values256, _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask)); + + vsum = _mm512_dpbusd_epi32(vsum, vb0, va0); + vsum = _mm512_dpbusd_epi32(vsum, vb1, va1); + b_qs += 64; + } + // (B + 128) * A - 128 * A + vsum = _mm512_sub_epi32(vsum, _mm512_permutexvar_epi32(vmask, vcomp)); + + // vacc += scale * (q8 @ q4) + const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N))); + acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale)); + } + const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0))); + vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]); + }; + + for (int i = 0; i < KB; ++i) { + Unroll{}(compute, i); + } + + //store to C + auto storec = [&](auto col) { + _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]); + }; + Unroll{}(storec); + } +}; + +#define LAUNCH_TINYGEMM_KERNEL_VNNI(NB_SIZE) \ + tinygemm_kernel_vnni::apply( \ + KB, (const char *)wdata + 0 * row_size_A, \ + (const char *)src0->data + PACKED_INDEX(nb * kTilesN, 0, KB, TILE_SIZE), \ + (float *) dst->data + 0 * N + nb_start, ldc) + +template ::value, int>::type = 0> +void tinygemm_kernel_amx(int M, int N, int KB, const void * RESTRICT _A, const void * RESTRICT _B, TC * RESTRICT C, int ldc) { + using packed_B_t = packed_B_type; + const int TILE_SIZE = get_tile_size(); + const bool need_unpack = do_unpack::value; + + GGML_ASSERT(M <= 2 * TILE_M && N == 2 * TILE_N); + const TA * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + const int m0 = std::min(M, TILE_M); + const int m1 = std::max(M - TILE_M, 0); + const int lda = KB * sizeof(TA); + //const int ldb = KB * sizeof(TB); + + static thread_local packed_B_t Tile0[TILE_N * TILE_K]; + static thread_local packed_B_t Tile1[TILE_N * TILE_K]; + static thread_local int8_t Tile23[TILE_M * TILE_K]; + + static thread_local int32_t TileC0[TILE_M * TILE_N * 4]; + static thread_local int32_t TileC1[TILE_M * TILE_N * 4]; + + // double buffering C to interleave avx512 and amx + int32_t * C_cur = TileC0; + int32_t * C_pre = TileC1; + + auto Tile4 = [&](int32_t * base) { return base; }; + auto Tile5 = [&](int32_t * base) { return base + TILE_M * TILE_N; }; + auto Tile6 = [&](int32_t * base) { return base + 2 * TILE_M * TILE_N; }; + auto Tile7 = [&](int32_t * base) { return base + 3 * TILE_M * TILE_N; }; + + if (M == 2 * TILE_M) { + // i = 0 + const char * B_blk0 = B + PACKED_INDEX(0, 0, KB, TILE_SIZE); + const char * B_blk1 = B + PACKED_INDEX(1, 0, KB, TILE_SIZE); + if (need_unpack) { + unpack_B(Tile0, B_blk0); + _tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK); + } else { + _tile_loadd(TMM0, B_blk0, TILE_N * VNNI_BLK); + } + + _tile_zero(TMM4); + _tile_loadd(TMM2, A[0].qs, lda); + _tile_dpbssd(TMM4, TMM2, TMM0); + _tile_stored(TMM4, Tile4(C_pre), TILE_N * sizeof(int32_t)); + + _tile_zero(TMM5); + _tile_loadd(TMM3, A[TILE_M * KB + 0].qs, lda); + _tile_dpbssd(TMM5, TMM3, TMM0); + _tile_stored(TMM5, Tile5(C_pre), TILE_N * sizeof(int32_t)); + + if (need_unpack) { + unpack_B(Tile1, B_blk0); + _tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK); + } else { + _tile_loadd(TMM1, B_blk1, TILE_N * VNNI_BLK); + } + + _tile_zero(TMM6); + _tile_dpbssd(TMM6, TMM2, TMM1); + _tile_stored(TMM6, Tile6(C_pre), TILE_N * sizeof(int32_t)); + + _tile_zero(TMM7); + _tile_dpbssd(TMM7, TMM3, TMM1); + _tile_stored(TMM7, Tile7(C_pre), TILE_N * sizeof(int32_t)); + + for (int i = 1; i < KB; ++i) { + // index of previous iter + const int ii = i - 1; + const char * B_blk0 = B + PACKED_INDEX(0, i, KB, TILE_SIZE); + const char * B_blk1 = B + PACKED_INDEX(1, i, KB, TILE_SIZE); + GGML_DISPATCH_BOOL(ii > 0, is_acc, [&] { + if (need_unpack) { + unpack_B(Tile0, B_blk0); + _tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK); + } else { + _tile_loadd(TMM0, B_blk0, TILE_N * VNNI_BLK); + } + _tile_zero(TMM4); + _tile_loadd(TMM2, A[i].qs, lda); + acc_C::apply(C, ldc, Tile4(C_pre), &A[ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M); + + _tile_dpbssd(TMM4, TMM2, TMM0); + _tile_stored(TMM4, Tile4(C_cur), TILE_N * sizeof(int32_t)); + + _tile_zero(TMM5); + _tile_loadd(TMM3, A[TILE_M * KB + i].qs, lda); + acc_C::apply(C + TILE_M * ldc, ldc, Tile5(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M); + + _tile_dpbssd(TMM5, TMM3, TMM0); + _tile_stored(TMM5, Tile5(C_cur), TILE_N * sizeof(int32_t)); + + if (need_unpack) { + unpack_B(Tile1, B_blk1); + _tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK); + } else { + _tile_loadd(TMM1, B_blk1, TILE_N * VNNI_BLK); + } + _tile_zero(TMM6); + acc_C::apply(C + TILE_N, ldc, Tile6(C_pre), &A[ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M); + + _tile_dpbssd(TMM6, TMM2, TMM1); + _tile_stored(TMM6, Tile6(C_cur), TILE_N * sizeof(int32_t)); + + _tile_zero(TMM7); + acc_C::apply(C + TILE_M * ldc + TILE_N, ldc, Tile7(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M); + + _tile_dpbssd(TMM7, TMM3, TMM1); + _tile_stored(TMM7, Tile7(C_cur), TILE_N * sizeof(int32_t)); + + std::swap(C_cur, C_pre); + }); + } + // final accumulation + { + int ii = KB - 1; + acc_C::apply(C, ldc, Tile4(C_pre), &A[ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M); + acc_C::apply(C + TILE_M * ldc, ldc, Tile5(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M); + acc_C::apply(C + TILE_N, ldc, Tile6(C_pre), &A[ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M); + acc_C::apply(C + TILE_M * ldc + TILE_N, ldc, Tile7(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M); + } + } else { + for (int i = 0; i < KB; ++i) { + _tile_zero(TMM4); + _tile_zero(TMM6); + if (m1 != 0) { + _tile_zero(TMM5); + _tile_zero(TMM7); + } + + const char * B_blk0 = B + PACKED_INDEX(0, i, KB, TILE_SIZE); + const char * B_blk1 = B + PACKED_INDEX(1, i, KB, TILE_SIZE); + if (need_unpack) { + unpack_B(Tile0, B_blk0); + _tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK); + } else { + _tile_loadd(TMM0, B_blk0, TILE_N * VNNI_BLK); + } + + if (need_unpack) { + unpack_B(Tile1, B_blk1); + _tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK); + } else { + _tile_loadd(TMM1, B_blk1, TILE_N * VNNI_BLK); + } + + if (m0 == TILE_M) { + _tile_loadd(TMM2, A[i].qs, lda); + } else { + unpack_A(Tile23, &A[i], KB, m0); + _tile_loadd(TMM2, Tile23, TILE_K); + } + + _tile_dpbssd(TMM4, TMM2, TMM0); + _tile_dpbssd(TMM6, TMM2, TMM1); + + _tile_stored(TMM4, Tile4(C_cur), TILE_N * sizeof(int32_t)); + _tile_stored(TMM6, Tile6(C_cur), TILE_N * sizeof(int32_t)); + + GGML_DISPATCH_BOOL(i > 0, is_acc, [&] { + acc_C::apply(C, ldc, Tile4(C_cur), &A[i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m0); + acc_C::apply(C + TILE_N, ldc, Tile6(C_cur), &A[i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m0); + }); + + if (m1 != 0) { + unpack_A(Tile23, &A[TILE_M * KB + i], KB, m1); + _tile_loadd(TMM3, Tile23, TILE_K); + + _tile_dpbssd(TMM5, TMM3, TMM0); + _tile_dpbssd(TMM7, TMM3, TMM1); + _tile_stored(TMM5, Tile5(C_cur), TILE_N * sizeof(int32_t)); + _tile_stored(TMM7, Tile7(C_cur), TILE_N * sizeof(int32_t)); + GGML_DISPATCH_BOOL(i > 0, is_acc, [&] { + acc_C::apply(C + TILE_M * ldc, ldc, Tile5(C_cur), &A[TILE_M * KB + i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m1); + acc_C::apply(C + TILE_M * ldc + TILE_N, ldc, Tile7(C_cur), &A[TILE_M * KB + i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m1); + }); + } + } + } + return; +} + +template ::value, int>::type = 0> +void tinygemm_kernel_amx(int M, int N, int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) { + static_assert(std::is_same::value); + const int TILE_SIZE = get_tile_size(); + + GGML_ASSERT(M <= 2 * TILE_M && N == 2 * TILE_N); + const TA * RESTRICT A = static_cast(_A); + const char * RESTRICT B = static_cast(_B); + + const int m0 = std::min(M, TILE_M); + const int m1 = std::max(M - TILE_M, 0); + //const int lda = KB * sizeof(TA); + + static thread_local int8_t Tile0[TILE_N * TILE_K]; + static thread_local int8_t Tile1[TILE_N * TILE_K]; + static thread_local int8_t Tile23[TILE_M * TILE_K]; + + // mat mul result for each group + static thread_local int32_t Tile4[TILE_M * TILE_N]; + static thread_local int32_t Tile5[TILE_M * TILE_N]; + static thread_local int32_t Tile6[TILE_M * TILE_N]; + static thread_local int32_t Tile7[TILE_M * TILE_N]; + + // sum of each QK_K block, contains 8 groups, int32 + static thread_local int32_t Sumi4[TILE_M * TILE_N]; + static thread_local int32_t Sumi5[TILE_M * TILE_N]; + static thread_local int32_t Sumi6[TILE_M * TILE_N]; + static thread_local int32_t Sumi7[TILE_M * TILE_N]; + + const int k_group_size = std::is_same::value ? 16 : 32; + for (int i = 0; i < KB; ++i) { + // step 1: accumulate the quants across 8 groups, each group with 32 + for (int k = 0; k < QK_K / k_group_size; ++k) { + GGML_DISPATCH_BOOL(k > 0, is_acc, [&] { + _tile_zero(TMM4); + _tile_zero(TMM6); + + unpack_B(Tile0, B + PACKED_INDEX(0, i, KB, TILE_SIZE), k); + _tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK); + + unpack_B(Tile1, B + PACKED_INDEX(1, i, KB, TILE_SIZE), k); + _tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK); + + unpack_A(Tile23, &A[i], KB, k, m0); + _tile_loadd(TMM2, Tile23, TILE_K); + + _tile_dpbssd(TMM4, TMM2, TMM0); + _tile_dpbssd(TMM6, TMM2, TMM1); + + _tile_stored(TMM4, Tile4, TILE_N * sizeof(int32_t)); + _tile_stored(TMM6, Tile6, TILE_N * sizeof(int32_t)); + + scale_C(Tile4, Sumi4, B + PACKED_INDEX(0, i, KB, TILE_SIZE), k, m0); + scale_C(Tile6, Sumi6, B + PACKED_INDEX(1, i, KB, TILE_SIZE), k, m0); + + if (m1 != 0) { + _tile_zero(TMM5); + _tile_zero(TMM7); + + unpack_A(Tile23, &A[TILE_M * KB + i], KB, k, m1); + _tile_loadd(TMM3, Tile23, TILE_K); + + _tile_dpbssd(TMM5, TMM3, TMM0); + _tile_dpbssd(TMM7, TMM3, TMM1); + + _tile_stored(TMM5, Tile5, TILE_N * sizeof(int32_t)); + _tile_stored(TMM7, Tile7, TILE_N * sizeof(int32_t)); + + scale_C(Tile5, Sumi5, B + PACKED_INDEX(0, i, KB, TILE_SIZE), k, m1); + scale_C(Tile7, Sumi7, B + PACKED_INDEX(1, i, KB, TILE_SIZE), k, m1); + } + }); + } + + // step 2: accmulate the mins + GGML_DISPATCH_BOOL(i > 0, is_acc, [&] { + acc_C::apply(C, ldc, Sumi4, &A[i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m0); + acc_C::apply(C + TILE_N, ldc, Sumi6, &A[i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m0); + if (m1 != 0) { + acc_C::apply(C + TILE_M * ldc, ldc, Sumi5, &A[TILE_M * KB + i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m1); + acc_C::apply(C + TILE_M * ldc + TILE_N, ldc, Sumi7, &A[TILE_M * KB + i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m1); + } + }); + } + return; +} + +} // anonymous namespace + +// get the packed tensor size for quantized weights +size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor) { + const enum ggml_type TYPE = tensor->type; + + const int K = tensor->ne[0]; // ne0: in_features + const int N = tensor->ne[1]; // ne1: out_features + + auto get_tensor_size = [&] { + size_t row_size_B{0}; + GGML_DISPATCH_QTYPES(TYPE, [&] { + row_size_B = get_row_size(K); + }); + return N * row_size_B; + }; + + if (qtype_has_amx_kernels(TYPE)) { + return get_tensor_size(); + } else { + // for f16, bf16 we don't do packing + return ggml_nbytes(tensor); + } +} + +// pack weight to vnni format +void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { + GGML_ASSERT(offset == 0 && size == ggml_nbytes(tensor)); // only full tensor conversion is supported for now + + const enum ggml_type TYPE = tensor->type; + + const int K = tensor->ne[0]; // ne0: in_features + const int N = tensor->ne[1]; // ne1: out_features + +#if defined(_OPENMP) + // the buffer ctx is not initialized when .set_tensor is called + int n_threads = omp_get_num_threads(); +#else + int n_threads = 1; +#endif + + GGML_DISPATCH_QTYPES(TYPE, [&] { + convert_B_packed_format((void *)((char *)tensor->data + offset), (const type *)data, N, K, n_threads); + }); +} + +size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst) { + struct ggml_tensor * src0 = dst->src[0]; + + const enum ggml_type TYPE = src0->type; + + const bool is_floating_type = TYPE == GGML_TYPE_F16; + if (is_floating_type) { + return 0; + } + + const int M = dst->ne[1]; + const int K = src0->ne[0]; + + size_t desired_wsize = 0; + + GGML_DISPATCH_QTYPES(TYPE, [&] { + const size_t row_size_A = K / blck_size * sizeof(vec_dot_type); + desired_wsize = M * row_size_A; + }); + + return desired_wsize; +} + +// NB: mixed dtype gemm with Advanced Matrix Extensions (Intel AMX) +// +// src0: weight in shape of {N, K}, quantized +// src1: input in shape of {M, K}, float32 +// dst: output in shape of {M, N}, float32 +// +// the function performs: dst = src1 @ src0.T +// +void ggml_backend_amx_mul_mat(const ggml_compute_params * params, struct ggml_tensor * dst) { + struct ggml_tensor * src0 = dst->src[0]; + struct ggml_tensor * src1 = dst->src[1]; + + const enum ggml_type TYPE = src0->type; + + // f16 only has avx512 kernels for now, + // amx kernels will be added once 6th gen xeon is released. + const bool is_floating_type = TYPE == GGML_TYPE_F16; + + const int M = dst->ne[1]; + const int N = dst->ne[0]; + const int K = src0->ne[0]; + const int ldc = dst->nb[1] / dst->nb[0]; + + if (is_floating_type) { + constexpr int BLOCK_M = 4; + constexpr int BLOCK_N = 6; + const int MB = div_up(M, BLOCK_M); + const int NB = div_up(N, BLOCK_N); + + parallel_for_ggml(params, MB * NB, [&](int begin, int end) { + GGML_DISPATCH_FLOATING_TYPES(TYPE, [&] { + for (int i = begin; i < end; ++i) { + int mb = i / NB; + int nb = i % NB; + + int mb_start = mb * BLOCK_M; + int mb_size = std::min(BLOCK_M, M - mb_start); + int nb_start = nb * BLOCK_N; + int nb_size = std::min(BLOCK_N, N - nb_start); + + switch (mb_size << 4 | nb_size) { + case 0x12: LAUNCH_TINYGEMM_KERNEL_AVX(1, 2); break; + case 0x14: LAUNCH_TINYGEMM_KERNEL_AVX(1, 4); break; + case 0x16: LAUNCH_TINYGEMM_KERNEL_AVX(1, 6); break; + case 0x22: LAUNCH_TINYGEMM_KERNEL_AVX(2, 2); break; + case 0x24: LAUNCH_TINYGEMM_KERNEL_AVX(2, 4); break; + case 0x26: LAUNCH_TINYGEMM_KERNEL_AVX(2, 6); break; + case 0x32: LAUNCH_TINYGEMM_KERNEL_AVX(3, 2); break; + case 0x34: LAUNCH_TINYGEMM_KERNEL_AVX(3, 4); break; + case 0x36: LAUNCH_TINYGEMM_KERNEL_AVX(3, 6); break; + case 0x42: LAUNCH_TINYGEMM_KERNEL_AVX(4, 2); break; + case 0x44: LAUNCH_TINYGEMM_KERNEL_AVX(4, 4); break; + case 0x46: LAUNCH_TINYGEMM_KERNEL_AVX(4, 6); break; + default: fprintf(stderr, "Unexpected block size!\n"); + } + } + }); + }); + return; + } + + // pointer to work space, used convert A from float to quantized type + void * wdata = params->wdata; + + //TODO: performance improvement: merge quant A + if (params->ith == 0) { + GGML_DISPATCH_QTYPES(TYPE, [&] { + const size_t row_size_A = K / blck_size * sizeof(vec_dot_type); + const size_t desired_wsize = M * row_size_A; + if (params->wsize < desired_wsize) { + GGML_ABORT("insufficient work space size"); + } + + // Q4_0, Q4_1, Q8_0 handles 1 TILE_K per blck_size + // Q4_K, Q5_K, Q6_K, IQ4_XS handles 8 TILE_K per blck_size + GGML_ASSERT(TILE_K == blck_size || TILE_K * 8 == blck_size); + + const float * A_data = static_cast(src1->data); + for (int m = 0; m < M; ++m) { + from_float(A_data + m * K, (char *)wdata + m * row_size_A, K); + } + }); + } + + ggml_barrier(params->threadpool); + + if (M == 1) { + // MB = 1 and handle 8 tiles in each block + constexpr int kTilesN = 4; + constexpr int BLOCK_N = TILE_N * kTilesN; + const int NB = div_up(N, BLOCK_N); + + parallel_for_ggml(params, NB, [&](int begin, int end) { + GGML_DISPATCH_QTYPES(TYPE, [&] { + const int KB = K / blck_size; + const int TILE_SIZE = get_tile_size(); + const int row_size_A = KB * sizeof(vec_dot_type); + for (int i = begin; i < end; ++i) { + int nb = i; + int nb_start = nb * BLOCK_N; + int nb_size = std::min(BLOCK_N, N - nb_start); // 32, 64, 96 + + switch (nb_size) { + //case 160: LAUNCH_TINYGEMM_KERNEL_VNNI(160); break; + case 128: LAUNCH_TINYGEMM_KERNEL_VNNI(128); break; + case 96: LAUNCH_TINYGEMM_KERNEL_VNNI(96); break; + case 64: LAUNCH_TINYGEMM_KERNEL_VNNI(64); break; + case 32: LAUNCH_TINYGEMM_KERNEL_VNNI(32); break; + default: fprintf(stderr, "Unexpected n block size!\n"); + } + } + }); + }); + return; + } + + // handle 4 tiles at a tile + constexpr int BLOCK_M = TILE_M * 2; + constexpr int BLOCK_N = TILE_N * 2; + const int MB = div_up(M, BLOCK_M); + const int NB = div_up(N, BLOCK_N); + + parallel_for_ggml(params, MB * NB, [&](int begin, int end) { + // init tile config for each thread + ggml_tile_config_init(); + + GGML_DISPATCH_QTYPES(TYPE, [&] { + const int KB = K / blck_size; + const int TILE_SIZE = get_tile_size(); + const int row_size_A = KB * sizeof(vec_dot_type); + + for (int i = begin; i < end; ++i) { + int mb = i / NB; + int nb = i % NB; + + int mb_start = mb * BLOCK_M; + int mb_size = std::min(BLOCK_M, M - mb_start); + int nb_start = nb * BLOCK_N; + int nb_size = BLOCK_N; + + tinygemm_kernel_amx( + mb_size, nb_size, KB, + (const char *)wdata + mb_start * row_size_A, + (const char *)src0->data + PACKED_INDEX(nb * 2, 0, KB, TILE_SIZE), + (float *) dst->data + mb_start * N + nb_start, ldc); + } + }); + }); +} + +#endif // if defined(__AMX_INT8__) && defined(__AVX512VNNI__) diff --git a/ggml/src/ggml-cpu/amx/mmq.h b/ggml/src/ggml-cpu/amx/mmq.h new file mode 100644 index 00000000000..f3736609315 --- /dev/null +++ b/ggml/src/ggml-cpu/amx/mmq.h @@ -0,0 +1,16 @@ +#pragma once +#include "common.h" + +#ifdef __cplusplus +extern "C" { +#endif + +size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor); + +void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); + +void ggml_backend_amx_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst); + +#ifdef __cplusplus +} +#endif diff --git a/ggml/src/ggml-cpu/cpu-feats-x86.cpp b/ggml/src/ggml-cpu/cpu-feats-x86.cpp new file mode 100644 index 00000000000..e8133d411fd --- /dev/null +++ b/ggml/src/ggml-cpu/cpu-feats-x86.cpp @@ -0,0 +1,323 @@ +#include "ggml-backend-impl.h" + +#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64)) + +#ifdef _MSC_VER +#include +#endif + +#include +#include +#include +#include +#include + +// 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 cpui; + std::vector> 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(vendor) = data[0][1]; + *reinterpret_cast(vendor + 4) = data[0][3]; + *reinterpret_cast(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> 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 = 0; + 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_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)) diff --git a/ggml/src/ggml-cpu/ggml-cpu-aarch64.c b/ggml/src/ggml-cpu/ggml-cpu-aarch64.c index 96a16dfba1f..11152385e67 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-aarch64.c +++ b/ggml/src/ggml-cpu/ggml-cpu-aarch64.c @@ -1,7 +1,3 @@ -// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates -// SPDX-License-Identifier: MIT -// - #define GGML_COMMON_IMPL_C #include "ggml-common.h" @@ -132,7 +128,7 @@ static inline __m512i sum_i16_pairs_int_32x16(const __m512i x) { } static inline __m512i mul_sum_us8_pairs_int32x16(const __m512i ax, const __m512i sy) { -#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__)) +#if defined(__AVX512VNNI__) const __m512i zero = _mm512_setzero_si512(); return _mm512_dpbusd_epi32(zero, ax, sy); #else @@ -187,6 +183,8 @@ static inline __m256i mul_sum_i8_pairs_int32x8(const __m256i x, const __m256i y) } #endif +static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113}; + static void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k) { assert(QK8_0 == 32); assert(k % QK8_0 == 0); @@ -527,67 +525,47 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * UNUSED(ncols_interleaved); UNUSED(blocklen); -#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) - if (ggml_cpu_has_neon()) { - const void * b_ptr = vx; - const void * a_ptr = vy; - float * res_ptr = s; - - __asm__ __volatile__( - "movi v31.16b, #0x4\n" - "movi v30.16b, #0xf0\n" - "add %x[b_ptr], %x[b_ptr], #0x8\n" - "1:" // Column loop - "add x22, %x[a_ptr], #0x2\n" - "movi v29.16b, #0x0\n" - "mov x21, %x[nb]\n" - "2:" // Block loop - "ldr q28, [%x[b_ptr], #0x0]\n" - "ldr q27, [x22, #0x0]\n" - "movi v26.4s, #0x0\n" - "sub x20, x22, #0x2\n" - "ldr q25, [x22, #0x10]\n" - "ldr q24, [%x[b_ptr], #0x10]\n" - "sub x21, x21, #0x1\n" - "add x22, x22, #0x22\n" - "ldr q23, [%x[b_ptr], #0x20]\n" - "ldr q22, [%x[b_ptr], #0x30]\n" - "ld1r { v21.8h }, [x20]\n" - "ldr q20, [%x[b_ptr], #-0x8]\n" - "sshl v16.16b, v28.16b, v31.16b\n" - "and v28.16b, v28.16b, v30.16b\n" - "sshl v19.16b, v24.16b, v31.16b\n" - "and v24.16b, v24.16b, v30.16b\n" - "add %x[b_ptr], %x[b_ptr], #0x48\n" - "sshl v18.16b, v23.16b, v31.16b\n" - "and v23.16b, v23.16b, v30.16b\n" - ".inst 0x4f9be21a // sdot v26.4s, v16.16b, v27.4b[0]\n" - "sshl v17.16b, v22.16b, v31.16b\n" - "and v22.16b, v22.16b, v30.16b\n" - "fcvtl v21.4s, v21.4h\n" - "fcvtl v16.4s, v20.4h\n" - ".inst 0x4f99e39a // sdot v26.4s, v28.16b, v25.4b[0]\n" - "fmul v16.4s, v16.4s, v21.4s\n" - ".inst 0x4fbbe27a // sdot v26.4s, v19.16b, v27.4b[1]\n" - ".inst 0x4fb9e31a // sdot v26.4s, v24.16b, v25.4b[1]\n" - ".inst 0x4f9bea5a // sdot v26.4s, v18.16b, v27.4b[2]\n" - ".inst 0x4f99eafa // sdot v26.4s, v23.16b, v25.4b[2]\n" - ".inst 0x4fbbea3a // sdot v26.4s, v17.16b, v27.4b[3]\n" - ".inst 0x4fb9eada // sdot v26.4s, v22.16b, v25.4b[3]\n" - "scvtf v26.4s, v26.4s, #0x4\n" - "fmla v29.4s, v26.4s, v16.4s\n" - "cbnz x21, 2b\n" - "sub %x[nc], %x[nc], #0x4\n" - "str q29, [%x[res_ptr], #0x0]\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "cbnz %x[nc], 1b\n" - : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc) - : [a_ptr] "r" (a_ptr), [nb] "r" (nb) - : "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22" - ); +#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) + if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { + const block_q4_0x4 * b_ptr = (const block_q4_0x4 *)vx; + + for (int c = 0; c < nc; c += ncols_interleaved) { + const block_q8_0 * a_ptr = (const block_q8_0 *)vy; + float32x4_t acc = vdupq_n_f32(0); + for (int b = 0; b < nb; b++) { + int8x16_t b0 = vld1q_s8((const int8_t *)b_ptr->qs); + int8x16_t b1 = vld1q_s8((const int8_t *)b_ptr->qs + 16); + int8x16_t b2 = vld1q_s8((const int8_t *)b_ptr->qs + 32); + int8x16_t b3 = vld1q_s8((const int8_t *)b_ptr->qs + 48); + float16x4_t bd = vld1_f16((const __fp16 *)b_ptr->d); + + int8x16_t a0 = vld1q_s8(a_ptr->qs); + int8x16_t a1 = vld1q_s8(a_ptr->qs + qk/2); + float16x4_t ad = vld1_dup_f16((const __fp16 *)&a_ptr->d); + + int32x4_t ret = vdupq_n_s32(0); + + ret = vdotq_laneq_s32(ret, b0 << 4, a0, 0); + ret = vdotq_laneq_s32(ret, b1 << 4, a0, 1); + ret = vdotq_laneq_s32(ret, b2 << 4, a0, 2); + ret = vdotq_laneq_s32(ret, b3 << 4, a0, 3); + + ret = vdotq_laneq_s32(ret, b0 & 0xf0U, a1, 0); + ret = vdotq_laneq_s32(ret, b1 & 0xf0U, a1, 1); + ret = vdotq_laneq_s32(ret, b2 & 0xf0U, a1, 2); + ret = vdotq_laneq_s32(ret, b3 & 0xf0U, a1, 3); + + acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4), + vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd))); + a_ptr++; + b_ptr++; + } + vst1q_f32(s, acc); + s += ncols_interleaved; + } return; } -#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) +#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) float sumf[4]; int sumi; @@ -996,6 +974,102 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * } } +void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 4; + + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) + if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { + const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl); + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + float * res_ptr = s; + + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); + + float32x4_t sumf = vdupq_n_f32(0); + for (int l = 0; l < nb; l++) { + uint8x16_t b_0 = vld1q_u8(b_ptr[l].qs + 0); + uint8x16_t b_1 = vld1q_u8(b_ptr[l].qs + 16); + uint8x16_t b_2 = vld1q_u8(b_ptr[l].qs + 32); + uint8x16_t b_3 = vld1q_u8(b_ptr[l].qs + 48); + + int8x16_t b_0_hi = vqtbl1q_s8(kvalues, b_0 >> 4); + int8x16_t b_0_lo = vqtbl1q_s8(kvalues, b_0 & 0x0F); + int8x16_t b_1_hi = vqtbl1q_s8(kvalues, b_1 >> 4); + int8x16_t b_1_lo = vqtbl1q_s8(kvalues, b_1 & 0x0F); + int8x16_t b_2_hi = vqtbl1q_s8(kvalues, b_2 >> 4); + int8x16_t b_2_lo = vqtbl1q_s8(kvalues, b_2 & 0x0F); + int8x16_t b_3_hi = vqtbl1q_s8(kvalues, b_3 >> 4); + int8x16_t b_3_lo = vqtbl1q_s8(kvalues, b_3 & 0x0F); + + int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 0); + int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16); + + int32x4_t sumi = vdupq_n_s32(0); + sumi = vdotq_laneq_s32(sumi, b_0_lo, a_0, 0); + sumi = vdotq_laneq_s32(sumi, b_0_hi, a_1, 0); + sumi = vdotq_laneq_s32(sumi, b_1_lo, a_0, 1); + sumi = vdotq_laneq_s32(sumi, b_1_hi, a_1, 1); + sumi = vdotq_laneq_s32(sumi, b_2_lo, a_0, 2); + sumi = vdotq_laneq_s32(sumi, b_2_hi, a_1, 2); + sumi = vdotq_laneq_s32(sumi, b_3_lo, a_0, 3); + sumi = vdotq_laneq_s32(sumi, b_3_hi, a_1, 3); + + float32x4_t a_d = vcvt_f32_f16(vld1_dup_f16((const float16_t *)&a_ptr[l].d)); + float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d)); + float32x4_t d = a_d * b_d; + + sumf = vmlaq_f32(sumf, d, vcvtq_f32_s32(sumi)); + } + + vst1q_f32(res_ptr + x * 4, sumf); + } + return; + } +#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) + { + float sumf[4]; + int sumi; + + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; + const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); + } + sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; + } + } +} + void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; @@ -1017,7 +1091,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * UNUSED(blocklen); #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) - if (ggml_cpu_has_neon()) { + if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { const void * b_ptr = vx; const void * a_ptr = vy; float * res_ptr = s; @@ -3386,6 +3460,117 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * } } +void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 4; + const int blocklen = 4; + + assert (n % qk == 0); + assert (nr % 4 == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) + if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { + const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl); + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); + + float32x4_t sumf[4]; + for (int m = 0; m < 4; m++) { + sumf[m] = vdupq_n_f32(0); + } + + for (int l = 0; l < nb; l++) { + float32x4_t a_d = vcvt_f32_f16(vld1_f16((const float16_t *)a_ptr[l].d)); + float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d)); + + int32x4_t sumi_0 = vdupq_n_s32(0); + int32x4_t sumi_1 = vdupq_n_s32(0); + int32x4_t sumi_2 = vdupq_n_s32(0); + int32x4_t sumi_3 = vdupq_n_s32(0); + + for (int k = 0; k < 4; k++) { + int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 16 * k + 0); + int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16 * k + 64); + + uint8x16_t b = vld1q_u8(b_ptr[l].qs + 16 * k); + int8x16_t b_hi = vqtbl1q_s8(kvalues, b >> 4); + int8x16_t b_lo = vqtbl1q_s8(kvalues, b & 0xF); + + sumi_0 = vdotq_laneq_s32(sumi_0, b_lo, a_0, 0); + sumi_1 = vdotq_laneq_s32(sumi_1, b_lo, a_0, 1); + sumi_2 = vdotq_laneq_s32(sumi_2, b_lo, a_0, 2); + sumi_3 = vdotq_laneq_s32(sumi_3, b_lo, a_0, 3); + sumi_0 = vdotq_laneq_s32(sumi_0, b_hi, a_1, 0); + sumi_1 = vdotq_laneq_s32(sumi_1, b_hi, a_1, 1); + sumi_2 = vdotq_laneq_s32(sumi_2, b_hi, a_1, 2); + sumi_3 = vdotq_laneq_s32(sumi_3, b_hi, a_1, 3); + } + + sumf[0] = vmlaq_f32(sumf[0], vmulq_laneq_f32(b_d, a_d, 0), vcvtq_f32_s32(sumi_0)); + sumf[1] = vmlaq_f32(sumf[1], vmulq_laneq_f32(b_d, a_d, 1), vcvtq_f32_s32(sumi_1)); + sumf[2] = vmlaq_f32(sumf[2], vmulq_laneq_f32(b_d, a_d, 2), vcvtq_f32_s32(sumi_2)); + sumf[3] = vmlaq_f32(sumf[3], vmulq_laneq_f32(b_d, a_d, 3), vcvtq_f32_s32(sumi_3)); + } + + for (int m = 0; m < 4; m++) { + vst1q_f32(s + (y * 4 + m) * bs + x * 4, sumf[m]); + } + } + } + return; + } +#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) + { + float sumf[4][4]; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; + const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); + } + sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); + } + } + } + } + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + } + } + } + } +} + // FIXME: this code is duplicated from ggml-aarch64.c static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) { block_q4_0x4 out; @@ -3518,6 +3703,70 @@ static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor *t, int interleave_block, GGML_UNUSED(data_size); } +static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_size_interleave) { + block_iq4_nlx4 out; + + for (int i = 0; i < 4; i++) { + out.d[i] = in[i].d; + } + + const int end = QK4_NL * 2 / blck_size_interleave; + + if (blck_size_interleave == 8) { + for (int i = 0; i < end; ++i) { + int src_id = i % 4; + int src_offset = (i / 4) * blck_size_interleave; + int dst_offset = i * blck_size_interleave; + + // Using memcpy to avoid unaligned memory accesses + memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t)); + } + } else if (blck_size_interleave == 4) { + for (int i = 0; i < end; ++i) { + int src_id = i % 4; + int src_offset = (i / 4) * blck_size_interleave; + int dst_offset = i * blck_size_interleave; + + memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint32_t)); + } + } else { + GGML_ASSERT(false); + } + + return out; +} + +static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * restrict data, size_t data_size) { + GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL); + GGML_ASSERT(interleave_block == 4 || interleave_block == 8); + + block_iq4_nlx4 * dst = (block_iq4_nlx4 *)t->data; + const block_iq4_nl * src = (const block_iq4_nl *)data; + block_iq4_nl dst_tmp[4]; + int nrow = t->ne[1]; // Number of rows + int nrows_interleaved = 4; + int nblocks = t->ne[0] / QK4_0; + + GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl)); + + if (nrow % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { + return -1; + } + + for (int b = 0; b < nrow; b += nrows_interleaved) { + for (int64_t x = 0; x < nblocks; x++) { + for (int i = 0; i < nrows_interleaved; i++) { + dst_tmp[i] = src[x + i * nblocks]; + } + *dst++ = make_block_iq4_nlx4(dst_tmp, interleave_block); + } + src += nrows_interleaved * nblocks; + } + return 0; + + GGML_UNUSED(data_size); +} + // Prepare for optimized kernels if applicable void ggml_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_type, const void * restrict data, size_t data_size) { if (cur->type == repack_type) { @@ -3525,20 +3774,30 @@ void ggml_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_ return; } - GGML_ASSERT(cur->type == GGML_TYPE_Q4_0); - - switch (repack_type) { - case GGML_TYPE_Q4_0_8_8: - repack_q4_0_to_q4_0_8_bl(cur, 8, data, data_size); - break; - case GGML_TYPE_Q4_0_4_8: - repack_q4_0_to_q4_0_4_bl(cur, 8, data, data_size); - break; - case GGML_TYPE_Q4_0_4_4: - repack_q4_0_to_q4_0_4_bl(cur, 4, data, data_size); - break; - default: - GGML_ABORT("Unsupported type"); + if (cur->type == GGML_TYPE_Q4_0) { + switch (repack_type) { + case GGML_TYPE_Q4_0_8_8: + repack_q4_0_to_q4_0_8_bl(cur, 8, data, data_size); + break; + case GGML_TYPE_Q4_0_4_8: + repack_q4_0_to_q4_0_4_bl(cur, 8, data, data_size); + break; + case GGML_TYPE_Q4_0_4_4: + repack_q4_0_to_q4_0_4_bl(cur, 4, data, data_size); + break; + default: + GGML_ABORT("Unsupported type"); + } + } else if (cur->type == GGML_TYPE_IQ4_NL) { + switch (repack_type) { + case GGML_TYPE_IQ4_NL_4_4: + repack_iq4_nl_to_iq4_nl_4_bl(cur, 4, data, data_size); + break; + default: + GGML_ABORT("Unsupported type"); + } + } else { + GGML_ABORT("Unsupported type"); } } @@ -3551,9 +3810,13 @@ enum ggml_type ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * c if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { return GGML_TYPE_Q4_0_4_8; } - if (ggml_cpu_has_neon()) { + if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { return GGML_TYPE_Q4_0_4_4; } + } else if (cur->type == GGML_TYPE_IQ4_NL) { + if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { + return GGML_TYPE_IQ4_NL_4_4; + } } return cur->type; diff --git a/ggml/src/ggml-cpu/ggml-cpu-aarch64.h b/ggml/src/ggml-cpu/ggml-cpu-aarch64.h index 53b30c1dd2d..3d9db6a19eb 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-aarch64.h +++ b/ggml/src/ggml-cpu/ggml-cpu-aarch64.h @@ -15,11 +15,13 @@ void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in 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_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); // GEMM 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_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_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_type, const void * data, size_t data_size); enum ggml_type ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur); diff --git a/ggml/src/ggml-cpu/ggml-cpu-impl.h b/ggml/src/ggml-cpu/ggml-cpu-impl.h index 27a530b2279..d71076ad12b 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-impl.h +++ b/ggml/src/ggml-cpu/ggml-cpu-impl.h @@ -15,6 +15,18 @@ 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; +}; + + #if defined(_MSC_VER) #define m512bh(p) p @@ -366,6 +378,9 @@ static __m256 __lasx_xvreplfr2vr_s(float val) { } #endif +// TODO: move to ggml-threading +void ggml_barrier(struct ggml_threadpool * tp); + #ifdef __cplusplus } #endif diff --git a/ggml/src/ggml-cpu/ggml-cpu-quants.c b/ggml/src/ggml-cpu/ggml-cpu-quants.c index f0e276b6987..634c5fa1162 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-quants.c +++ b/ggml/src/ggml-cpu/ggml-cpu-quants.c @@ -1791,11 +1791,12 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r const int8x16_t y1_l = vld1q_s8(b_y1->qs); const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16); - float32_t _scale[4] = { GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d), - GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d), - GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d), - GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)}; - + float32_t _scale[4] = { + GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d), + GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d), + GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d), + GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d) + }; float32x4_t scale = vld1q_f32(_scale); int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l))); @@ -1811,13 +1812,15 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h))); sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)), - l1, r1)), l2, r2)), l3, r3))), scale); + l1, r1)), l2, r2)), l3, r3))), scale); } - float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2); + + float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2); float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1); - vst1_f32(s, vget_low_f32(sumv2)); + vst1_f32(s, vget_low_f32 (sumv2)); vst1_f32(s + bs, vget_high_f32(sumv2)); + return; } #endif @@ -2345,10 +2348,12 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r const block_q8_1 * restrict b_y0 = &vy0[i]; const block_q8_1 * restrict b_y1 = &vy1[i]; - float32_t summs_t[4] = {GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s), - GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s), - GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s), - GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)}; + float32_t summs_t[4] = { + GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s), + GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s), + GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s), + GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s) + }; summs0 = vaddq_f32(summs0, vld1q_f32(summs_t)); const uint8x16_t m4b = vdupq_n_u8(0x0F); @@ -2369,10 +2374,12 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16); // mmla into int32x4_t - float32_t _scale[4] = {GGML_FP16_TO_FP32(b_x0->d)*b_y0->d, - GGML_FP16_TO_FP32(b_x0->d)*b_y1->d, - GGML_FP16_TO_FP32(b_x1->d)*b_y0->d, - GGML_FP16_TO_FP32(b_x1->d)*b_y1->d}; + float32_t _scale[4] = { + GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d), + GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d), + GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d), + GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d) + }; float32x4_t scale = vld1q_f32(_scale); int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l))); @@ -2387,15 +2394,17 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h))); int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h))); sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)), - l1, r1)), l2, r2)), l3, r3))), scale); + l1, r1)), l2, r2)), l3, r3))), scale); } - float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2); + float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2); float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1); + sumv2 = vaddq_f32(sumv2, summs0); vst1_f32(s, vget_low_f32 (sumv2)); vst1_f32(s + bs, vget_high_f32(sumv2)); + return; } #endif @@ -3372,10 +3381,12 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r const int8x16_t y1_l = vld1q_s8(b_y1->qs); const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16); - float32_t _scale[4] = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d), - GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d), - GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d), - GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)}; + float32_t _scale[4] = { + GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d), + GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d), + GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d), + GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d) + }; float32x4_t scale = vld1q_f32(_scale); int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l))); @@ -3391,13 +3402,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h))); sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)), - l1, r1)), l2, r2)), l3, r3))), scale); + l1, r1)), l2, r2)), l3, r3))), scale); } - float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2); + + float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2); float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1); - vst1_f32(s, vget_low_f32(sumv2)); + vst1_f32(s, vget_low_f32 (sumv2)); vst1_f32(s + bs, vget_high_f32(sumv2)); + return; } #endif diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c index 0d23669c25b..f12b62e3bd8 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.c +++ b/ggml/src/ggml-cpu/ggml-cpu.c @@ -10,6 +10,7 @@ #include "ggml-quants.h" #include "ggml-cpu-quants.h" #include "ggml-threading.h" +#include "amx/amx.h" #include "ggml.h" #if defined(_MSC_VER) || defined(__MINGW32__) @@ -109,10 +110,11 @@ static ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16]; #if defined(__ARM_ARCH) struct ggml_arm_arch_features_type { int has_neon; + int has_dotprod; int has_i8mm; int has_sve; int sve_cnt; -} ggml_arm_arch_features = {-1, -1, -1, 0}; +} ggml_arm_arch_features = {-1, -1, -1, -1, 0}; #endif @@ -446,6 +448,15 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = { .vec_dot_type = GGML_TYPE_Q8_K, .nrows = 1, }, + [GGML_TYPE_IQ4_NL_4_4] = { + .from_float = NULL, + .vec_dot = NULL, + .vec_dot_type = GGML_TYPE_Q8_0, + .nrows = 1, + .ncols = 4, + .gemv = ggml_gemv_iq4_nl_4x4_q8_0, + .gemm = ggml_gemm_iq4_nl_4x4_q8_0, + }, }; const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type) { @@ -614,7 +625,7 @@ do { \ for (int i = 0; i < offset; ++i) { \ x[i] = _mm512_add_ps(x[i], x[offset+i]); \ } \ - res = _mm512_reduce_add_ps(x[0]); \ + res = (ggml_float) _mm512_reduce_add_ps(x[0]); \ } while (0) // TODO: is this optimal ? @@ -664,7 +675,7 @@ do { \ for (int i = 0; i < offset; ++i) { \ x[i] = _mm512_add_ps(x[i], x[offset+i]); \ } \ - res = _mm512_reduce_add_ps(x[0]); \ + res = (ggml_float) _mm512_reduce_add_ps(x[0]); \ } while (0) #define GGML_F16_VEC GGML_F32Cx16 @@ -675,8 +686,8 @@ do { \ #define GGML_F16_VEC_FMA GGML_F32Cx16_FMA #define GGML_F16_VEC_ADD GGML_F32Cx16_ADD #define GGML_F16_VEC_MUL GGML_F32Cx16_MUL -#define GGML_F16_VEC_REDUCE GGML_F32Cx16_REDUCE +#define GGML_F16_VEC_REDUCE GGML_F32Cx16_REDUCE #elif defined(__AVX__) #define GGML_SIMD @@ -745,7 +756,7 @@ do { \ #define GGML_F32Cx8_LOAD(x) _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x))) #define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0)) #else -static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) { +static inline __m256 __avx_f32cx8_load(const ggml_fp16_t * x) { float tmp[8]; for (int i = 0; i < 8; i++) { @@ -1168,28 +1179,28 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) { #define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a) #define GGML_F32x4_ADD __lsx_vfadd_s #define GGML_F32x4_MUL __lsx_vfmul_s -#define GGML_F32x4_REDUCE(res, x) \ -{ \ - int offset = GGML_F32_ARR >> 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = __lsx_vfadd_s(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = __lsx_vfadd_s(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = __lsx_vfadd_s(x[i], x[offset+i]); \ - } \ - __m128i tmp = __lsx_vsrli_d((__m128i)x[0], 32); \ - tmp = (__m128i)__lsx_vfadd_s((__m128)tmp, x[0]); \ - tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \ - const __m128 t0 = __lsx_vshuf4i_w(tmp, 0x88); \ - tmp = __lsx_vsrli_d((__m128i)t0, 32); \ - tmp = (__m128i)__lsx_vfadd_s((__m128)tmp, t0); \ - tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \ - res = (ggml_float) __lsx_vpickve2gr_w(__lsx_vshuf4i_w(tmp, 0x88), 0); \ +#define GGML_F32x4_REDUCE(res, x) \ +{ \ + int offset = GGML_F32_ARR >> 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = __lsx_vfadd_s(x[i], x[offset + i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = __lsx_vfadd_s(x[i], x[offset + i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = __lsx_vfadd_s(x[i], x[offset + i]); \ + } \ + __m128i tmp = __lsx_vsrli_d((__m128i) x[0], 32); \ + tmp = (__m128i) __lsx_vfadd_s((__m128) tmp, x[0]); \ + tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \ + const __m128 t0 = __lsx_vshuf4i_w(tmp, 0x88); \ + tmp = __lsx_vsrli_d((__m128i) t0, 32); \ + tmp = (__m128i) __lsx_vfadd_s((__m128) tmp, t0); \ + tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \ + res = (ggml_float) __lsx_vpickve2gr_w(__lsx_vshuf4i_w(tmp, 0x88), 0); \ } #define GGML_F32_VEC GGML_F32x4 @@ -1357,31 +1368,18 @@ struct ggml_compute_state { int ith; }; -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; -}; - // // fundamental operations // inline static void ggml_vec_set_i8(const int n, int8_t * x, const int8_t v) { for (int i = 0; i < n; ++i) x[i] = v; } - inline static void ggml_vec_set_i16(const int n, int16_t * x, const int16_t v) { for (int i = 0; i < n; ++i) x[i] = v; } -inline static void ggml_vec_set_i32(const int n, int32_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; } +inline static void ggml_vec_set_i32(const int n, int32_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; } +inline static void ggml_vec_cpy_i32(const int n, int32_t * y, const int32_t * x) { for (int i = 0; i < n; ++i) y[i] = x[i]; } inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; } - inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf16_t v) { for (int i = 0; i < n; ++i) x[i] = v; } - inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] + y[i]; } inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float v) { for (int i = 0; i < n; ++i) z[i] = x[i] + v; } inline static void ggml_vec_acc_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] += x[i]; } @@ -2276,7 +2274,7 @@ struct ggml_state { static struct ggml_state g_state = {0}; -static void ggml_barrier(struct ggml_threadpool * tp) { +void ggml_barrier(struct ggml_threadpool * tp) { int n_threads = atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed); if (n_threads == 1) { return; @@ -2430,7 +2428,7 @@ bool ggml_is_numa(void) { #endif #if !defined(HWCAP2_I8MM) -#define HWCAP2_I8MM 0 +#define HWCAP2_I8MM (1 << 13) #endif static void ggml_init_arm_arch_features(void) { @@ -2439,6 +2437,7 @@ static void ggml_init_arm_arch_features(void) { uint32_t hwcap2 = getauxval(AT_HWCAP2); ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD); + ggml_arm_arch_features.has_dotprod = !!(hwcap && HWCAP_ASIMDDP); ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM); ggml_arm_arch_features.has_sve = !!(hwcap & HWCAP_SVE); @@ -2453,6 +2452,11 @@ static void ggml_init_arm_arch_features(void) { } ggml_arm_arch_features.has_neon = oldp; + if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) != 0) { + oldp = 0; + } + ggml_arm_arch_features.has_dotprod = oldp; + if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) { oldp = 0; } @@ -7439,6 +7443,13 @@ static void ggml_compute_forward_mul_mat( type = (enum ggml_type)(intptr_t)src0->extra; } +#if defined(__AMX_INT8__) && defined(__AVX512VNNI__) + if (src0->buffer && ggml_backend_amx_buft_is_amx(src0->buffer->buft)) { + ggml_backend_amx_mul_mat(params, dst); + return; + } +#endif + enum ggml_type const vec_dot_type = type_traits_cpu[type].vec_dot_type; ggml_from_float_t const from_float = type_traits_cpu[vec_dot_type].from_float; ggml_from_float_to_mat_t const from_float_to_mat = type_traits_cpu[vec_dot_type].from_float_to_mat; @@ -7560,14 +7571,6 @@ UseGgmlGemm2:; // This is the size of the rest of the dimensions of the result const int64_t nr1 = ne1 * ne2 * ne3; - // dot kernels can handle 1 row and col at a time, but mmla kernels can process 2 rows and cols - int64_t num_rows_per_vec_dot = vec_dot_num_rows; - // TODO: currently the mmla kernels support only even numbered rows/cols. - // this check can be removed once they are extended to support odd numbered rows/cols too - if ((nr0 % 2 != 0) || (ne11 % 2 != 0)) { - num_rows_per_vec_dot = 1; - } - // Now select a reasonable chunk size. int chunk_size = 16; @@ -7630,6 +7633,15 @@ UseGgmlGemm2:; const int64_t ir1_start = dr1 * ith1; const int64_t ir1_end = MIN(ir1_start + dr1, nr1); + // dot kernels can handle 1 row and col at a time, but mmla kernels can process 2 rows and cols + int64_t num_rows_per_vec_dot = vec_dot_num_rows; + + // these checks are needed to avoid crossing dim1 boundaries + // can be optimized, but the logic would become more complicated, so keeping it like this for simplicity + if ((nr0 % 2 != 0) || (ne11 % 2 != 0) || ((ir0_end - ir0_start) % 2 != 0) || ((ir1_end - ir1_start) % 2 != 0)) { + num_rows_per_vec_dot = 1; + } + ggml_compute_forward_mul_mat_one_chunk(params, dst, type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end); if (nth >= nchunk0 * nchunk1) { @@ -8239,6 +8251,77 @@ static void ggml_compute_forward_set_f32( } } +static void ggml_compute_forward_set_i32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + + // view src0 and dst with these strides and data offset inbytes during set + // nb0 is implicitly element_size because src0 and dst are contiguous + size_t nb1 = ((int32_t *) dst->op_params)[0]; + size_t nb2 = ((int32_t *) dst->op_params)[1]; + size_t nb3 = ((int32_t *) dst->op_params)[2]; + size_t offset = ((int32_t *) dst->op_params)[3]; + bool inplace = (bool) ((int32_t *) dst->op_params)[4]; + + if (!inplace) { + if (params->ith == 0) { + // memcpy needs to be synchronized across threads to avoid race conditions. + // => do it in INIT phase + memcpy( + ((char *) dst->data), + ((char *) src0->data), + ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + } + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src1); + const int nc = src1->ne[0]; + + GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) + GGML_TENSOR_LOCALS(size_t, nb1, src1, nb) + + // src0 and dst as viewed during set + const size_t nb0 = ggml_element_size(src0); + + const int im0 = (ne10 == 0 ? 0 : ne10-1); + const int im1 = (ne11 == 0 ? 0 : ne11-1); + const int im2 = (ne12 == 0 ? 0 : ne12-1); + const int im3 = (ne13 == 0 ? 0 : ne13-1); + + GGML_ASSERT(offset + im0*nb0 + im1*nb1 + im2*nb2 + im3*nb3 <= ggml_nbytes(dst)); + + GGML_ASSERT(nb10 == sizeof(int32_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are viewed with shape of src1 and offset + // => same indices + const int i3 = ir/(ne12*ne11); + const int i2 = (ir - i3*ne12*ne11)/ne11; + const int i1 = (ir - i3*ne12*ne11 - i2*ne11); + + ggml_vec_cpy_i32(nc, + (int32_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), + (int32_t *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); + } +} + static void ggml_compute_forward_set( const struct ggml_compute_params * params, struct ggml_tensor * dst) { @@ -8250,6 +8333,10 @@ static void ggml_compute_forward_set( { ggml_compute_forward_set_f32(params, dst); } break; + case GGML_TYPE_I32: + { + ggml_compute_forward_set_i32(params, dst); + } break; case GGML_TYPE_F16: case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: @@ -9133,6 +9220,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_Q4_0_4_4: case GGML_TYPE_Q4_0_4_8: case GGML_TYPE_Q4_0_8_8: + case GGML_TYPE_IQ4_NL_4_4: case GGML_TYPE_I8: case GGML_TYPE_I16: case GGML_TYPE_I32: @@ -10429,6 +10517,40 @@ static void ggml_compute_forward_pad( } } +// ggml_compute_forward_pad_reflect_1d + +static void ggml_compute_forward_pad_reflect_1d( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + const int ith = params->ith; + const int nth = params->nth; + + const int32_t * opts = (const int32_t *) dst->op_params; + const int p0 = opts[0]; + const int p1 = opts[1]; + + GGML_TENSOR_UNARY_OP_LOCALS + + for (int64_t i3 = 0; i3 < ne3; i3++) { + for (int64_t i2 = 0; i2 < ne2; i2++) { + for (int64_t i1 = ith; i1 < ne1; i1 += nth) { + float * left = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + p0*nb0); + float * right = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (ne0-p1-1)*nb0); + + ggml_vec_cpy_f32(ne00, left, (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01)); + + for (int i0 = 1; i0 <= p0; i0++) { left[-i0] = left[i0]; } + for (int i0 = 1; i0 <= p1; i0++) { right[i0] = right[-i0]; } + } + } + } +} // ggml_compute_forward_arange @@ -12525,6 +12647,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_pad(params, tensor); } break; + case GGML_OP_PAD_REFLECT_1D: + { + ggml_compute_forward_pad_reflect_1d(params, tensor); + } break; case GGML_OP_ARANGE: { ggml_compute_forward_arange(params, tensor); @@ -12867,6 +12993,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { } break; case GGML_OP_UPSCALE: case GGML_OP_PAD: + case GGML_OP_PAD_REFLECT_1D: case GGML_OP_ARANGE: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_ARGSORT: @@ -13276,10 +13403,16 @@ struct ggml_cplan ggml_graph_plan( } break; case GGML_OP_MUL_MAT: { +#if defined(__AMX_INT8__) && defined(__AVX512VNNI__) + if (node->src[0]->buffer && ggml_backend_amx_buft_is_amx(node->src[0]->buffer->buft)) { + cur = ggml_backend_amx_desired_wsize(node); + } +#endif const enum ggml_type vec_dot_type = type_traits_cpu[node->src[0]->type].vec_dot_type; if (node->src[1]->type != vec_dot_type) { - cur = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1])); + size_t cur2 = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1])); + cur = MAX(cur, cur2); } } break; case GGML_OP_MUL_MAT_ID: @@ -13578,29 +13711,6 @@ static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool, int #endif // GGML_USE_OPENMP -void ggml_threadpool_params_init(struct ggml_threadpool_params * p, int n_threads) { - p->n_threads = n_threads; - p->prio = 0; // default priority (usually means normal or inherited) - p->poll = 50; // hybrid-polling enabled - p->strict_cpu = false; // no strict placement (all threads share same cpumask) - p->paused = false; // threads are ready to go - memset(p->cpumask, 0, GGML_MAX_N_THREADS); // all-zero means use the default affinity (usually inherited) -} - -struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads) { - struct ggml_threadpool_params p; - ggml_threadpool_params_init(&p, n_threads); - return p; -} - -bool ggml_threadpool_params_match(const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1) { - if (p0->n_threads != p1->n_threads ) return false; - if (p0->prio != p1->prio ) return false; - if (p0->poll != p1->poll ) return false; - if (p0->strict_cpu != p1->strict_cpu ) return false; - return memcmp(p0->cpumask, p1->cpumask, GGML_MAX_N_THREADS) == 0; -} - static struct ggml_threadpool * ggml_threadpool_new_impl( struct ggml_threadpool_params * tpp, struct ggml_cgraph * cgraph, @@ -13896,15 +14006,23 @@ int ggml_cpu_has_vsx(void) { } int ggml_cpu_has_neon(void) { -#if defined(__ARM_ARCH) +#if defined(__ARM_ARCH) && defined(__ARM_NEON) return ggml_arm_arch_features.has_neon; #else return 0; #endif } +int ggml_cpu_has_dotprod(void) { +#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_DOTPROD) + return ggml_arm_arch_features.has_dotprod; +#else + return 0; +#endif +} + int ggml_cpu_has_sve(void) { -#if defined(__ARM_ARCH) +#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE) return ggml_arm_arch_features.has_sve; #else return 0; @@ -13912,7 +14030,7 @@ int ggml_cpu_has_sve(void) { } int ggml_cpu_has_matmul_int8(void) { -#if defined(__ARM_ARCH) +#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8) return ggml_arm_arch_features.has_i8mm; #else return 0; @@ -13920,7 +14038,7 @@ int ggml_cpu_has_matmul_int8(void) { } int ggml_cpu_get_sve_cnt(void) { -#if defined(__ARM_ARCH) +#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE) return ggml_arm_arch_features.sve_cnt; #else return 0; diff --git a/ggml/src/ggml-cpu/ggml-cpu.cpp b/ggml/src/ggml-cpu/ggml-cpu.cpp index 573b7c5b9b3..d3b4bdb9650 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.cpp +++ b/ggml/src/ggml-cpu/ggml-cpu.cpp @@ -3,6 +3,7 @@ #include "ggml-cpu.h" #include "ggml-cpu-aarch64.h" #include "ggml-impl.h" +#include "amx/amx.h" #include #include #include @@ -134,12 +135,16 @@ static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backen static std::vector bufts = []() { std::vector bufts; -#ifdef GGML_USE_CPU_HBM - bufts.push_back(ggml_backend_cpu_hbm_buffer_type()); +#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_AARCH64 - bufts.push_back(ggml_backend_cpu_aarch64_buffer_type()); + if (ggml_backend_cpu_aarch64_buffer_type()) { + bufts.push_back(ggml_backend_cpu_aarch64_buffer_type()); + } #endif bufts.push_back(NULL); @@ -456,12 +461,27 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st 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; + } + if (src0 && src0->buffer && ggml_backend_cpu_buft_is_aarch64(src0->buffer->buft)) { - if (op->op != GGML_OP_MUL_MAT || src0->type != GGML_TYPE_Q4_0 || ggml_aarch64_get_optimal_repack_type(src0) == GGML_TYPE_Q4_0) { + if (op->op != GGML_OP_MUL_MAT || src0->type == ggml_aarch64_get_optimal_repack_type(src0)) { return false; } } +#if defined(__AMX_INT8__) && defined(__AVX512VNNI__) + if (src0 && src0->buffer && ggml_backend_amx_buft_is_amx(src0->buffer->buft)) { + return ggml_backend_amx_device_supports_op(op); + } + for (int i = 1; i < GGML_MAX_SRC; i++) { + if (op->src[i] && op->src[i]->buffer && ggml_backend_amx_buft_is_amx(op->src[i]->buffer->buft)) { + return false; + } + } +#endif + for (int i = 1; i < GGML_MAX_SRC; i++) { if (op->src[i] && op->src[i]->buffer && ggml_backend_cpu_buft_is_aarch64(op->src[i]->buffer->buft)) { return false; @@ -491,7 +511,13 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st } 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_buft_is_aarch64(buft); + bool supported = ggml_backend_buft_is_host(buft) || ggml_backend_cpu_buft_is_aarch64(buft); + +#if defined(__AMX_INT8__) && defined(__AVX512VNNI__) + supported = supported || ggml_backend_amx_buft_is_amx(buft); +#endif + + return supported; GGML_UNUSED(dev); } @@ -541,16 +567,12 @@ static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg return &ggml_backend_cpu_device; } -struct ggml_backend_feature { - const char * name; - const char * value; -}; - -// Not used yet // 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. +// 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 features = []() { + ggml_cpu_init(); + std::vector features; if (ggml_cpu_has_sse3()) { features.push_back({ "SSE3", "1" }); @@ -561,6 +583,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r 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" }); } @@ -570,9 +595,6 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r if (ggml_cpu_has_fma()) { features.push_back({ "FMA", "1" }); } - if (ggml_cpu_has_avx_vnni()) { - features.push_back({ "AVX_VNNI", "1" }); - } if (ggml_cpu_has_avx512()) { features.push_back({ "AVX512", "1" }); } @@ -619,6 +641,18 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r 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_AARCH64 + features.push_back({ "AARCH64_REPACK", "1" }); + #endif features.push_back({ nullptr, nullptr }); @@ -637,6 +671,29 @@ static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const ch if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) { return (void *)ggml_backend_cpu_get_extra_bufts; } + 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; + } + + // 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; @@ -655,9 +712,12 @@ ggml_backend_reg_t ggml_backend_cpu_reg(void) { ggml_cpu_init(); static struct ggml_backend_reg ggml_backend_cpu_reg = { - /* .iface = */ ggml_backend_cpu_reg_i, - /* .context = */ NULL, + /* .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) diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.cpp b/ggml/src/ggml-cpu/llamafile/sgemm.cpp index b2ce2e66494..da4146ec4f6 100644 --- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp +++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp @@ -50,8 +50,7 @@ #include "sgemm.h" #include "ggml-impl.h" -// hack until moved into the CPU backend -#include "../ggml-cpu-impl.h" +#include "ggml-cpu-impl.h" #include "ggml-quants.h" #ifdef _MSC_VER diff --git a/ggml/src/ggml-cuda/argmax.cu b/ggml/src/ggml-cuda/argmax.cu index aab04eca7a3..5340eedc089 100644 --- a/ggml/src/ggml-cuda/argmax.cu +++ b/ggml/src/ggml-cuda/argmax.cu @@ -1,57 +1,69 @@ -#include "common.cuh" +#include +#include + #include "argmax.cuh" +#include "common.cuh" #include "sum.cuh" -#include +static __global__ void argmax_f32(const float * __restrict__ x, int32_t * __restrict__ dst, const int64_t ncols) { + const int64_t row = blockIdx.x; -static __global__ void argmax_f32( - const float * x, int32_t * dst, const int64_t ncols, const int64_t nrows) { + float maxval = -FLT_MAX; + int argmax = -1; + const float * rowx = x + row * ncols; - int argmax_thread = 0; - const int64_t row0 = (int64_t)blockIdx.x*WARP_SIZE; + 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 (int64_t row1 = 0; row1 < WARP_SIZE; ++row1) { - const int64_t row = row0 + row1; - - if (row >= nrows) { - break; + for (int offset = 16; 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; } + } - float maxval = -FLT_MAX; - int argmax = -1; - - for (int32_t col = threadIdx.x; col < ncols; col += WARP_SIZE) { - const float val = x[row*ncols + col]; - const int bigger = val > maxval; - const int not_bigger = bigger ^ 0x00000001; - - maxval = maxval*not_bigger + val*bigger; - argmax = argmax*not_bigger + col*bigger; + 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 mask = 16; mask > 0; mask >>= 1) { - const float val = __shfl_xor_sync(0xFFFFFFFF, maxval, mask, WARP_SIZE); - const int col = __shfl_xor_sync(0xFFFFFFFF, argmax, mask, WARP_SIZE); - const int bigger = val > maxval; - const int not_bigger = bigger ^ 0x00000001; - - maxval = maxval*not_bigger + val*bigger; - argmax = argmax*not_bigger + col*bigger; + for (int offset = 16; 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 store = row1 == threadIdx.x; - argmax_thread += store*argmax; } - const int row = row0 + threadIdx.x; - - if (row >= nrows) { - return; + if (warp_id == 0 && lane_id == 0) { + dst[row] = argmax; } - - dst[row] = argmax_thread; } void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { @@ -70,10 +82,10 @@ void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { cudaStream_t stream = ctx.stream(); - const int64_t num_blocks = (nrows + WARP_SIZE - 1) / WARP_SIZE; - - const dim3 blocks_dim(WARP_SIZE, 1, 1); + const int64_t num_blocks = nrows; + const int64_t num_threads = std::min(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<<>>(src0_d, dst_d, ne00, nrows); + argmax_f32<<>>(src0_d, dst_d, ne00); } diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index e146c691c6f..535118d8792 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -47,9 +47,20 @@ #define CC_TURING 750 #define CC_AMPERE 800 #define CC_OFFSET_AMD 1000000 -#define CC_RDNA1 (CC_OFFSET_AMD + 1010) -#define CC_RDNA2 (CC_OFFSET_AMD + 1030) -#define CC_RDNA3 (CC_OFFSET_AMD + 1100) + +// GCN/CNDA, wave size is 64 +#define CC_GCN4 (CC_OFFSET_AMD + 803) // Tonga, Fiji, Polaris, minimum for fast fp16 +#define CC_VEGA (CC_OFFSET_AMD + 900) // Vega56/64, minimum for fp16 dual issue +#define CC_VEGA20 (CC_OFFSET_AMD + 906) // MI50/Radeon VII, minimum for dp4a +#define CC_CDNA (CC_OFFSET_AMD + 908) // MI100, minimum for MFMA, acc registers +#define CC_CDNA2 (CC_OFFSET_AMD + 910) // MI210, minimum acc register renameing +#define CC_CDNA3 (CC_OFFSET_AMD + 942) // MI300 + +// RNDA removes MFMA, dp4a, xnack, acc registers, wave size is 32 +#define CC_RDNA1 (CC_OFFSET_AMD + 1010) // RX 5000 +#define CC_RDNA2 (CC_OFFSET_AMD + 1030) // RX 6000, minimum for dp4a +#define CC_RDNA3 (CC_OFFSET_AMD + 1100) // RX 7000, minimum for WMMA + #define CC_QY1 210 #define CC_QY2 220 @@ -180,8 +191,8 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) { return __reduce_add_sync(0xffffffff, x); #else #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - x += __shfl_xor_sync(0xffffffff, x, mask, 32); + for (int offset = 16; offset > 0; offset >>= 1) { + x += __shfl_xor_sync(0xffffffff, x, offset, 32); } return x; #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE @@ -189,17 +200,17 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) { static __device__ __forceinline__ float warp_reduce_sum(float x) { #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - x += __shfl_xor_sync(0xffffffff, x, mask, 32); + for (int offset = 16; offset > 0; offset >>= 1) { + x += __shfl_xor_sync(0xffffffff, x, offset, 32); } return x; } static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) { #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32); - a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32); + for (int offset = 16; offset > 0; offset >>= 1) { + a.x += __shfl_xor_sync(0xffffffff, a.x, offset, 32); + a.y += __shfl_xor_sync(0xffffffff, a.y, offset, 32); } return a; } @@ -209,16 +220,16 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) { #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - const half2 a_other = __shfl_xor_sync(0xffffffff, a, mask, 32); + for (int offset = 16; offset > 0; offset >>= 1) { + const half2 a_other = __shfl_xor_sync(0xffffffff, a, offset, 32); reinterpret_cast(a.x) += __low2half(a_other); reinterpret_cast(a.y) += __high2half(a_other); } return a; #else #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32)); + for (int offset = 16; offset > 0; offset >>= 1) { + a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, offset, 32)); } return a; #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) @@ -231,8 +242,8 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) { static __device__ __forceinline__ float warp_reduce_max(float x) { #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32)); + for (int offset = 16; offset > 0; offset >>= 1) { + x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, offset, 32)); } return x; } @@ -275,8 +286,8 @@ static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const hal static __device__ __forceinline__ half2 warp_reduce_max(half2 x) { #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32)); + for (int offset = 16; offset > 0; offset >>= 1) { + x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, offset, 32)); } return x; #else diff --git a/ggml/src/ggml-cuda/fattn-vec-f16.cuh b/ggml/src/ggml-cuda/fattn-vec-f16.cuh index 5ec3b91ae2b..34a2992c769 100644 --- a/ggml/src/ggml-cuda/fattn-vec-f16.cuh +++ b/ggml/src/ggml-cuda/fattn-vec-f16.cuh @@ -220,7 +220,6 @@ static __global__ void flash_attn_vec_ext_f16( for (int j = 0; j < ncols; ++j) { half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j]; - kqmax_new_j = warp_reduce_max(kqmax_new_j); if (threadIdx.x == 0) { kqmax_shared[j][threadIdx.y] = kqmax_new_j; } diff --git a/ggml/src/ggml-cuda/fattn-vec-f32.cuh b/ggml/src/ggml-cuda/fattn-vec-f32.cuh index 3d93f4a8acd..a28fc8b7fc8 100644 --- a/ggml/src/ggml-cuda/fattn-vec-f32.cuh +++ b/ggml/src/ggml-cuda/fattn-vec-f32.cuh @@ -206,7 +206,6 @@ static __global__ void flash_attn_vec_ext_f32( for (int j = 0; j < ncols; ++j) { float kqmax_new_j = kqmax_new_arr[j]; - kqmax_new_j = warp_reduce_max(kqmax_new_j); if (threadIdx.x == 0) { kqmax_shared[j][threadIdx.y] = kqmax_new_j; } diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index dd94ab03d5b..d6e4bfdd0d4 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -1107,6 +1107,11 @@ static void ggml_cuda_op_mul_mat_cublas( const half alpha_f16 = 1.0f; const half beta_f16 = 0.0f; + cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F; + if (ggml_cuda_info().devices[ctx.device].cc == CC_CDNA) { + cu_compute_type = CUBLAS_COMPUTE_32F; + } + CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream)); CUBLAS_CHECK( cublasGemmEx(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N, @@ -1114,7 +1119,7 @@ static void ggml_cuda_op_mul_mat_cublas( &alpha_f16, src0_ptr, CUDA_R_16F, ne00, src1_ptr, CUDA_R_16F, ne10, &beta_f16, dst_f16.get(), CUDA_R_16F, ldc, - CUBLAS_COMPUTE_16F, + cu_compute_type, CUBLAS_GEMM_DEFAULT_TENSOR_OP)); const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16); @@ -1607,6 +1612,10 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F; cudaDataType_t cu_data_type = CUDA_R_16F; + if (ggml_cuda_info().devices[ctx.device].cc == CC_CDNA) { + cu_compute_type = CUBLAS_COMPUTE_32F; + } + // dst strides size_t nbd2 = dst->nb[2]; size_t nbd3 = dst->nb[3]; @@ -3126,6 +3135,61 @@ static ggml_backend_dev_t ggml_backend_cuda_reg_get_device(ggml_backend_reg_t re return ctx->devices[index]; } +static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t reg) { + static std::vector features = []() { + std::vector features; + #define _STRINGIFY(...) #__VA_ARGS__ + #define STRINGIFY(...) _STRINGIFY(__VA_ARGS__) + + #ifdef __CUDA_ARCH_LIST__ + features.push_back({ "ARCHS", STRINGIFY(__CUDA_ARCH_LIST__) }); + #endif + + #ifdef GGML_CUDA_FORCE_MMQ + features.push_back({ "FORCE_MMQ", "1" }); + #endif + + #ifdef GGML_CUDA_FORCE_CUBLAS + features.push_back({ "FORCE_CUBLAS", "1" }); + #endif + + #ifdef GGML_CUDA_NO_VMM + features.push_back({ "NO_VMM", "1" }); + #endif + + #ifdef GGML_CUDA_NO_PEER_COPY + features.push_back({ "NO_PEER_COPY", "1" }); + #endif + + #ifdef GGML_CUDA_F16 + features.push_back({ "F16", "1" }); + #endif + + #ifdef GGML_CUDA_USE_GRAPHS + features.push_back({ "USE_GRAPHS", "1" }); + #endif + + #ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE + features.push_back({ "PEER_MAX_BATCH_SIZE", STRINGIFY(GGML_CUDA_PEER_MAX_BATCH_SIZE) }); + #endif + + #ifdef GGML_CUDA_FA_ALL_QUANTS + features.push_back({ "FA_ALL_QUANTS", "1" }); + #endif + + #undef _STRINGIFY + #undef STRINGIFY + + features.push_back({ nullptr, nullptr }); + + return features; + }(); + + return features.data(); + + GGML_UNUSED(reg); +} + static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) { GGML_UNUSED(reg); if (strcmp(name, "ggml_backend_split_buffer_type") == 0) { @@ -3137,6 +3201,9 @@ static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, con if (strcmp(name, "ggml_backend_unregister_host_buffer") == 0) { return (void *)ggml_backend_cuda_unregister_host_buffer; } + if (strcmp(name, "ggml_backend_get_features") == 0) { + return (void *)ggml_backend_cuda_get_features; + } return nullptr; } @@ -3169,16 +3236,17 @@ ggml_backend_reg_t ggml_backend_cuda_reg() { dev_ctx->description = prop.name; ggml_backend_dev_t dev = new ggml_backend_device { - /* .interface = */ ggml_backend_cuda_device_interface, - /* .reg = */ ®, - /* .context = */ dev_ctx + /* .iface = */ ggml_backend_cuda_device_interface, + /* .reg = */ ®, + /* .context = */ dev_ctx }; ctx->devices.push_back(dev); } reg = ggml_backend_reg { - /* .interface = */ ggml_backend_cuda_reg_interface, - /* .context = */ ctx + /* .api_version = */ GGML_BACKEND_API_VERSION, + /* .iface = */ ggml_backend_cuda_reg_interface, + /* .context = */ ctx }; } @@ -3209,3 +3277,5 @@ ggml_backend_t ggml_backend_cuda_init(int device) { return cuda_backend; } + +GGML_BACKEND_DL_IMPL(ggml_backend_cuda_reg) diff --git a/ggml/src/ggml-cuda/ggml/CMakeLists.txt b/ggml/src/ggml-cuda/ggml/CMakeLists.txt index e1482a269d6..14761650f8a 100644 --- a/ggml/src/ggml-cuda/ggml/CMakeLists.txt +++ b/ggml/src/ggml-cuda/ggml/CMakeLists.txt @@ -46,13 +46,10 @@ if (CUDAToolkit_FOUND) list(APPEND GGML_SOURCES_CUDA ${SRCS}) endif() - add_library(ggml-cuda - ${GGML_HEADERS_CUDA} - ${GGML_SOURCES_CUDA} - ) - - target_link_libraries(ggml-cuda PRIVATE ggml-base) - target_include_directories(ggml-cuda PRIVATE . ..) + 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}) @@ -135,7 +132,7 @@ if (CUDAToolkit_FOUND) message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}") - get_flags(${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() diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu index ae5c68ab351..7f7c8c90b6f 100644 --- a/ggml/src/ggml-cuda/mmq.cu +++ b/ggml/src/ggml-cuda/mmq.cu @@ -148,5 +148,5 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) { return cc < CC_VOLTA || ne11 < MMQ_DP4A_MAX_BATCH_SIZE; } - return cc < CC_RDNA3 || ne11 < MMQ_DP4A_MAX_BATCH_SIZE; + return (cc < CC_RDNA3 && cc != CC_CDNA && cc != CC_VEGA20) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE; } diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh index 425acb20da3..8d8867121f3 100644 --- a/ggml/src/ggml-cuda/mmq.cuh +++ b/ggml/src/ggml-cuda/mmq.cuh @@ -2570,9 +2570,9 @@ static __device__ void mul_mat_q_process_tile( template #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) +#if defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN) __launch_bounds__(WARP_SIZE*nwarps, 2) -#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN) #else #if __CUDA_ARCH__ >= CC_VOLTA __launch_bounds__(WARP_SIZE*nwarps, 1) diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu index 735975c160d..02d1509836c 100644 --- a/ggml/src/ggml-cuda/mmvq.cu +++ b/ggml/src/ggml-cuda/mmvq.cu @@ -142,7 +142,7 @@ static void mul_mat_vec_q_cuda( int64_t nwarps = 1; int64_t rows_per_cuda_block = 1; - if (ggml_cuda_info().devices[id].cc < CC_RDNA2) { // NVIDIA and AMD older than RDNA2 + if (ggml_cuda_info().devices[id].cc < CC_CDNA || ggml_cuda_info().devices[id].cc == CC_RDNA1) { // NVIDIA and AMD older than RDNA2 but not CDNA switch(ncols_y) { case 1: nwarps = 4; diff --git a/ggml/src/ggml-cuda/quantize.cu b/ggml/src/ggml-cuda/quantize.cu index 45408ce8684..1702e4ce2fe 100644 --- a/ggml/src/ggml-cuda/quantize.cu +++ b/ggml/src/ggml-cuda/quantize.cu @@ -69,8 +69,8 @@ static __global__ void quantize_mmq_q8_1( // Exchange max. abs. value between vals_per_scale/4 threads. #pragma unroll - for (int mask = vals_per_scale/8; mask > 0; mask >>= 1) { - amax = fmaxf(amax, __shfl_xor_sync(0xFFFFFFFF, amax, mask, WARP_SIZE)); + for (int offset = vals_per_scale/8; offset > 0; offset >>= 1) { + amax = fmaxf(amax, __shfl_xor_sync(0xFFFFFFFF, amax, offset, WARP_SIZE)); } float sum; @@ -79,8 +79,8 @@ static __global__ void quantize_mmq_q8_1( // Exchange calculate sum across vals_per_sum/4 threads. #pragma unroll - for (int mask = vals_per_sum/8; mask > 0; mask >>= 1) { - sum += __shfl_xor_sync(0xFFFFFFFF, sum, mask, WARP_SIZE); + for (int offset = vals_per_sum/8; offset > 0; offset >>= 1) { + sum += __shfl_xor_sync(0xFFFFFFFF, sum, offset, WARP_SIZE); } } diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h index 1f3c70c2e69..3205534d66f 100644 --- a/ggml/src/ggml-cuda/vendors/hip.h +++ b/ggml/src/ggml-cuda/vendors/hip.h @@ -95,6 +95,14 @@ #define __CUDA_ARCH__ 1300 +#if defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) +#define GCN +#endif + +#if defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx942__) +#define CDNA +#endif + #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \ defined(__gfx1150__) || defined(__gfx1151__) #define RDNA3 diff --git a/ggml/src/ggml-hip/CMakeLists.txt b/ggml/src/ggml-hip/CMakeLists.txt index fccf8eb8440..b15fbd24d6b 100644 --- a/ggml/src/ggml-hip/CMakeLists.txt +++ b/ggml/src/ggml-hip/CMakeLists.txt @@ -64,12 +64,10 @@ else() list(APPEND GGML_SOURCES_ROCM ${SRCS}) endif() -add_library(ggml-hip - ${GGML_HEADERS_ROCM} - ${GGML_SOURCES_ROCM}) - -target_link_libraries(ggml-hip PRIVATE ggml-base) -target_include_directories(ggml-hip PRIVATE . ..) +ggml_add_backend_library(ggml-hip + ${GGML_HEADERS_ROCM} + ${GGML_SOURCES_ROCM} + ) # TODO: do not use CUDA definitions for HIP target_compile_definitions(ggml PUBLIC GGML_USE_CUDA) diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h index 92a64fe5a59..00a1546a7d5 100644 --- a/ggml/src/ggml-impl.h +++ b/ggml/src/ggml-impl.h @@ -14,7 +14,7 @@ #include #endif // __ARM_FEATURE_SVE -#if defined(__ARM_NEON) +#if defined(__ARM_NEON) && !defined(__CUDACC__) // if YCM cannot find , make a symbolic link to it, for example: // // $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/ @@ -30,11 +30,13 @@ extern "C" { #endif -#undef MIN -#undef MAX +#ifndef MIN +# define MIN(a, b) ((a) < (b) ? (a) : (b)) +#endif -#define MIN(a, b) ((a) < (b) ? (a) : (b)) -#define MAX(a, b) ((a) > (b) ? (a) : (b)) +#ifndef MAX +# define MAX(a, b) ((a) > (b) ? (a) : (b)) +#endif // required for mmap as gguf only guarantees 32-byte alignment #define TENSOR_ALIGNMENT 32 @@ -295,6 +297,9 @@ struct ggml_cgraph { enum ggml_cgraph_eval_order order; }; +// returns a slice of cgraph with nodes [i0, i1) +// the slice does not have leafs or gradients +// if you need the gradients, get them from the original graph struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1); // Memory allocation @@ -305,14 +310,14 @@ void ggml_aligned_free(void * ptr, size_t size); // FP16 to FP32 conversion #if defined(__ARM_NEON) - #ifdef _MSC_VER + #if defined(_MSC_VER) || (defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) typedef uint16_t ggml_fp16_internal_t; #else typedef __fp16 ggml_fp16_internal_t; #endif #endif -#if defined(__ARM_NEON) && !defined(_MSC_VER) +#if defined(__ARM_NEON) && !defined(_MSC_VER) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) diff --git a/ggml/src/ggml-kompute/CMakeLists.txt b/ggml/src/ggml-kompute/CMakeLists.txt index 0bd027c7f53..c9109d5e8ee 100644 --- a/ggml/src/ggml-kompute/CMakeLists.txt +++ b/ggml/src/ggml-kompute/CMakeLists.txt @@ -6,13 +6,13 @@ if (NOT glslc_executable) message(FATAL_ERROR "glslc not found") endif() -add_library(ggml-kompute - ggml-kompute.cpp - ../../include/ggml-kompute.h - ) +ggml_add_backend_library(ggml-kompute + ggml-kompute.cpp + ../../include/ggml-kompute.h + ) target_link_libraries(ggml-kompute PRIVATE ggml-base kompute) -target_include_directories(ggml-kompute PRIVATE . .. ${CMAKE_CURRENT_BINARY_DIR}) +target_include_directories(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR}) add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1) @@ -105,8 +105,10 @@ if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt") kompute-shaders/op_getrows_q4_0.comp kompute-shaders/op_getrows_q4_1.comp kompute-shaders/op_getrows_q6_k.comp - kompute-shaders/op_rope_f16.comp - kompute-shaders/op_rope_f32.comp + kompute-shaders/op_rope_norm_f16.comp + kompute-shaders/op_rope_norm_f32.comp + kompute-shaders/op_rope_neox_f16.comp + kompute-shaders/op_rope_neox_f32.comp kompute-shaders/op_cpy_f16_f16.comp kompute-shaders/op_cpy_f16_f32.comp kompute-shaders/op_cpy_f32_f16.comp @@ -139,8 +141,10 @@ if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt") shaderop_getrows_q4_0.h shaderop_getrows_q4_1.h shaderop_getrows_q6_k.h - shaderop_rope_f16.h - shaderop_rope_f32.h + shaderop_rope_norm_f16.h + shaderop_rope_norm_f32.h + shaderop_rope_neox_f16.h + shaderop_rope_neox_f32.h shaderop_cpy_f16_f16.h shaderop_cpy_f16_f32.h shaderop_cpy_f32_f16.h diff --git a/ggml/src/ggml-kompute/ggml-kompute.cpp b/ggml/src/ggml-kompute/ggml-kompute.cpp index 2fea9e4cc8d..28ceecfc40d 100644 --- a/ggml/src/ggml-kompute/ggml-kompute.cpp +++ b/ggml/src/ggml-kompute/ggml-kompute.cpp @@ -28,8 +28,10 @@ #include "shaderop_getrows_q4_0.h" #include "shaderop_getrows_q4_1.h" #include "shaderop_getrows_q6_k.h" -#include "shaderop_rope_f16.h" -#include "shaderop_rope_f32.h" +#include "shaderop_rope_norm_f16.h" +#include "shaderop_rope_norm_f32.h" +#include "shaderop_rope_neox_f16.h" +#include "shaderop_rope_neox_f32.h" #include "shaderop_cpy_f16_f16.h" #include "shaderop_cpy_f16_f32.h" #include "shaderop_cpy_f32_f16.h" @@ -345,7 +347,7 @@ void ggml_vk_allocate_descriptor_pool(struct ggml_kompute_context * ctx, size_t std::vector descriptorPoolSizes = { vk::DescriptorPoolSize( vk::DescriptorType::eStorageBuffer, - 3 * size // Descriptor count is number of possible tensors to pass into an algorithm + 4 * size // Descriptor count is number of possible tensors to pass into an algorithm ) }; @@ -788,7 +790,8 @@ static void ggml_vk_soft_max( const std::shared_ptr& out, uint32_t inAOff, uint32_t inBOff, uint32_t outOff, int32_t ne00, int32_t ne01, int32_t ne02, uint32_t ne03, - float scale + float scale, float max_bias, float m0, float m1, + uint32_t n_head_log2 ) { const static auto spirv = getSpirvShader(kp::shader_data::op_softmax_comp_spv, kp::shader_data::op_softmax_comp_spv_len); @@ -796,12 +799,14 @@ static void ggml_vk_soft_max( struct PushConstants { uint32_t inAOff, inBOff, outOff; int32_t ne00, ne01, ne02; - float scale; + float scale, max_bias, m0, m1; + uint32_t n_head_log2; int32_t mask; } pushConsts { safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4), ne00, ne01, ne02, - scale, + scale, max_bias, m0, m1, + n_head_log2, bool(inB) }; @@ -911,9 +916,9 @@ static void ggml_vk_mul_mat_f16( const std::shared_ptr& out, uint32_t inAOff, uint32_t inBOff, uint32_t outOff, int32_t ne00, int32_t ne01, int32_t ne02, - uint32_t nb00, uint32_t nb01, uint32_t nb02, + uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03, int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13, - uint32_t nb10, uint32_t nb11, uint32_t nb12, + uint32_t nb10, uint32_t nb11, uint32_t nb12, uint32_t nb13, int32_t ne0, int32_t ne1, uint32_t r2, uint32_t r3 ) { @@ -923,17 +928,17 @@ static void ggml_vk_mul_mat_f16( struct PushConstants { uint32_t inAOff, inBOff, outOff; int32_t ne00, ne01, ne02; - uint32_t nb00, nb01, nb02; + uint32_t nb00, nb01, nb02, nb03; int32_t ne10, ne11, ne12; - uint32_t nb10, nb11, nb12; + uint32_t nb10, nb11, nb12, nb13; int32_t ne0, ne1; uint32_t r2, r3; } pushConsts { safe_divide(inAOff, 2), safe_divide(inBOff, 4), safe_divide(outOff, 4), ne00, ne01, ne02, - nb00, nb01, nb02, + nb00, nb01, nb02, nb03, ne10, ne11, ne12, - nb10, nb11, nb12, + nb10, nb11, nb12, nb13, ne0, ne1, r2, r3 }; @@ -1013,6 +1018,8 @@ static void ggml_vk_mul_mat_impl( int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13, int32_t ne0, int32_t ne1, + uint32_t nb01, uint32_t nb02, uint32_t nb03, + uint32_t nb11, uint32_t nb12, uint32_t nb13, uint32_t r2, uint32_t r3 ) { struct PushConstants { @@ -1020,19 +1027,23 @@ static void ggml_vk_mul_mat_impl( int32_t ne00, ne01, ne02; int32_t ne10, ne12; int32_t ne0, ne1; + uint32_t nb01, nb02, nb03; + uint32_t nb11, nb12, nb13; uint32_t r2, r3; } pushConsts { safe_divide(inAOff, block_size), safe_divide(inBOff, 4), safe_divide(outOff, 4), ne00, ne01, ne02, ne10, ne12, ne0, ne1, + nb01, nb02, nb03, + nb11, nb12, nb13, r2, r3 }; auto name = std::string(__func__) + "_" + suffix; std::shared_ptr s_algo = nullptr; if (!komputeManager()->hasAlgorithm(name)) { - const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2; + const uint32_t local_x = (ggml_vk_current_device().subgroupSize * 2) / 8; s_algo = komputeManager()->algorithm(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)}, {local_x}, {pushConsts}); } else { s_algo = komputeManager()->getAlgorithm(name); @@ -1074,19 +1085,26 @@ static void ggml_vk_mul_mat_q4_k( const std::shared_ptr& inB, const std::shared_ptr& out, uint32_t inAOff, uint32_t inBOff, uint32_t outOff, - int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne10, - int32_t ne11, int32_t ne12, int32_t ne13, int32_t ne0, - int32_t ne1, int32_t r2, int32_t r3 + int32_t ne00, int32_t ne01, int32_t ne02, + int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13, + int32_t ne0, int32_t ne1, + uint32_t nb01, uint32_t nb02, uint32_t nb03, + uint32_t nb11, uint32_t nb12, uint32_t nb13, + uint32_t r2, uint32_t r3 ) { const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_k_comp_spv, kp::shader_data::op_mul_mat_q4_k_comp_spv_len); struct PushConstants { uint32_t inAOff, inBOff, outOff; - int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12, r2, r3; + int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12; + uint32_t nb01, nb02, nb03, nb11, nb12, nb13; + uint32_t r2, r3; } pushConsts { - 0, 0, 0, - ne00, ne10, ne0, ne1, ne01, ne02, ne12, r2, r3 + inAOff, safe_divide(inBOff, 4), safe_divide(outOff, 4), + ne00, ne10, ne0, ne1, ne01, ne02, ne12, + nb01, nb02, nb03, nb11, nb12, nb13, + r2, r3 }; std::shared_ptr s_algo = nullptr; @@ -1108,28 +1126,37 @@ static void ggml_vk_mul_mat_q6_k( const std::shared_ptr& inB, const std::shared_ptr& out, uint32_t inAOff, uint32_t inBOff, uint32_t outOff, - int32_t ne00, int32_t ne10, int32_t ne0, int32_t ne1, - int32_t ne01, int32_t ne11, int32_t ne12, int32_t ne02 + int32_t ne00, int32_t ne01, int32_t ne02, + int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13, + int32_t ne0, int32_t ne1, + uint32_t nb01, uint32_t nb02, uint32_t nb03, + uint32_t nb11, uint32_t nb12, uint32_t nb13, + uint32_t r2, uint32_t r3 ) { const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q6_k_comp_spv, kp::shader_data::op_mul_mat_q6_k_comp_spv_len); struct PushConstants { uint32_t inAOff, inBOff, outOff; - int32_t ne00, ne10, ne0, ne1, ne01, gqa; + int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12; + uint32_t nb01, nb02, nb03, nb11, nb12, nb13; + uint32_t r2, r3; } pushConsts { inAOff, safe_divide(inBOff, 4), safe_divide(outOff, 4), - ne00, ne10, ne0, ne1, ne01, ne12/ne02 + ne00, ne10, ne0, ne1, ne01, ne02, ne12, + nb01, nb02, nb03, nb11, nb12, nb13, + r2, r3 }; std::shared_ptr s_algo = nullptr; if (!komputeManager()->hasAlgorithm(__func__)) { - const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2; - s_algo = komputeManager()->algorithm(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)}, {local_x}, {pushConsts}); + const uint32_t local_x = 2; + const uint32_t local_y = ggml_vk_current_device().subgroupSize; + s_algo = komputeManager()->algorithm(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)*unsigned(ne13)}, {local_x, local_y}, {pushConsts}); } else { s_algo = komputeManager()->getAlgorithm(__func__); s_algo->setTensors({inA, inB, out}); - s_algo->setWorkgroup({unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)}); + s_algo->setWorkgroup({unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)*unsigned(ne13)}); s_algo->setPushConstants({pushConsts}); s_algo->updateDescriptors(s_kompute_context->pool.get()); } @@ -1217,10 +1244,11 @@ static void ggml_vk_rope( kp::Sequence& seq, const std::shared_ptr& inA, const std::shared_ptr& inB, + const std::shared_ptr& inC, const std::shared_ptr& out, - uint32_t inAOff, uint32_t inBOff, uint32_t outOff, + uint32_t inAOff, uint32_t inBOff, uint32_t inCOff, uint32_t outOff, ggml_type src0t, int32_t n_dims, int32_t mode, int32_t n_ctx_orig, - float freq_base, float freq_scale, float ext_factor, float attn_factor, float beta_fast, float beta_slow, + float freq_base, float freq_scale, bool has_freq_factors, float ext_factor, float attn_factor, float beta_fast, float beta_slow, int32_t ne01, int32_t ne02, int32_t ne03, uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03, int32_t ne0, @@ -1228,11 +1256,17 @@ static void ggml_vk_rope( ) { GGML_ASSERT(src0t == GGML_TYPE_F16 || src0t == GGML_TYPE_F32); - static const auto spirv_f16 = getSpirvShader( - kp::shader_data::op_rope_f16_comp_spv, kp::shader_data::op_rope_f16_comp_spv_len + static const auto spirv_norm_f16 = getSpirvShader( + kp::shader_data::op_rope_norm_f16_comp_spv, kp::shader_data::op_rope_norm_f16_comp_spv_len + ); + static const auto spirv_norm_f32 = getSpirvShader( + kp::shader_data::op_rope_norm_f32_comp_spv, kp::shader_data::op_rope_norm_f32_comp_spv_len + ); + static const auto spirv_neox_f16 = getSpirvShader( + kp::shader_data::op_rope_neox_f16_comp_spv, kp::shader_data::op_rope_neox_f16_comp_spv_len ); - static const auto spirv_f32 = getSpirvShader( - kp::shader_data::op_rope_f32_comp_spv, kp::shader_data::op_rope_f32_comp_spv_len + static const auto spirv_neox_f32 = getSpirvShader( + kp::shader_data::op_rope_neox_f32_comp_spv, kp::shader_data::op_rope_neox_f32_comp_spv_len ); int type_size = src0t == GGML_TYPE_F16 ? 2 : 4; @@ -1247,32 +1281,40 @@ static void ggml_vk_rope( GGML_ASSERT(nb0 % type_size == 0); struct PushConstants { - uint32_t inAOff, inBOff, outOff; + uint32_t inAOff, inBOff, inCOff, outOff; int32_t n_dims, mode, n_ctx_orig; - float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; + float freq_base, freq_scale; + bool has_freq_factors; + float ext_factor, attn_factor, beta_fast, beta_slow; uint32_t nb00, nb01, nb02, nb03; int32_t ne0; uint32_t nb0, nb1, nb2, nb3; } pushConsts { - safe_divide(inAOff, type_size), safe_divide(inBOff, 4), safe_divide(outOff, type_size), + safe_divide(inAOff, type_size), safe_divide(inBOff, 4), safe_divide(inCOff, type_size), safe_divide(outOff, type_size), n_dims, mode, n_ctx_orig, - freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, + freq_base, freq_scale, + has_freq_factors, + ext_factor, attn_factor, beta_fast, beta_slow, nb00, nb01, nb02, nb03, ne0, nb0, nb1, nb2, nb3 }; - auto name = std::string(__func__) + (src0t == GGML_TYPE_F16 ? "_f16" : "_f32"); + auto & inC_ = inC ? inC : inA; + const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; + const bool is_f16 = src0t == GGML_TYPE_F16; + + auto name = std::string(__func__) + (is_neox ? "_neox" : "_norm") + (src0t == GGML_TYPE_F16 ? "_f16" : "_f32"); std::shared_ptr s_algo = nullptr; if (!komputeManager()->hasAlgorithm(name)) { + auto & spirv = is_neox ? is_f16 ? spirv_neox_f16 : spirv_neox_f32 : is_f16 ? spirv_norm_f16 : spirv_norm_f32; s_algo = komputeManager()->algorithm( - name, s_kompute_context->pool.get(), {inA, inB, out}, - src0t == GGML_TYPE_F16 ? spirv_f16 : spirv_f32, + name, s_kompute_context->pool.get(), {inA, inB, inC_, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts} ); } else { s_algo = komputeManager()->getAlgorithm(name); - s_algo->setTensors({inA, inB, out}); + s_algo->setTensors({inA, inB, inC_, out}); s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)}); s_algo->setPushConstants({pushConsts}); s_algo->updateDescriptors(s_kompute_context->pool.get()); @@ -1351,11 +1393,15 @@ static void ggml_vk_cpy_f16_f32(Args&&... args) { } static bool ggml_backend_kompute_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) { + int64_t n = ggml_nelements(op); switch (op->op) { case GGML_OP_UNARY: + if (n % 4 != 0) return false; switch (ggml_get_unary_op(op)) { - case GGML_UNARY_OP_RELU: case GGML_UNARY_OP_GELU: + if (n % 8 != 0) return false; + // fall through + case GGML_UNARY_OP_RELU: case GGML_UNARY_OP_SILU: return ggml_is_contiguous(op->src[0]); default: @@ -1413,8 +1459,8 @@ static bool ggml_backend_kompute_device_supports_op(ggml_backend_dev_t dev, cons switch (op->src[0]->type) { case GGML_TYPE_F32: - case GGML_TYPE_Q6_K: return op->ne[3] == 1; + case GGML_TYPE_Q6_K: case GGML_TYPE_F16: case GGML_TYPE_Q8_0: case GGML_TYPE_Q4_0: @@ -1515,9 +1561,11 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml const static std::shared_ptr nullTensor = nullptr; uint32_t off_src0 = 0; uint32_t off_src1 = 0; + uint32_t off_src2 = 0; uint32_t off_dst = 0; const std::shared_ptr& id_src0 = src0 ? ggml_vk_get_tensor(src0, &off_src0) : nullTensor; const std::shared_ptr& id_src1 = src1 ? ggml_vk_get_tensor(src1, &off_src1) : nullTensor; + const std::shared_ptr& id_src2 = src2 ? ggml_vk_get_tensor(src2, &off_src2) : nullTensor; const std::shared_ptr& id_dst = dst ? ggml_vk_get_tensor(dst, &off_dst) : nullTensor; switch (dst->op) { @@ -1593,11 +1641,16 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml #pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5021") GGML_ASSERT(!src1 || src1t == GGML_TYPE_F32); -#pragma message("TODO: add ALiBi support") -#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/7192") - GGML_ASSERT(max_bias == 0.0f); + const int64_t nrows_x = ggml_nrows(src0); + const int64_t nrows_y = src0->ne[1]; + + const uint32_t n_head = nrows_x/nrows_y; + const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head)); + + const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); + const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); - ggml_vk_soft_max(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale); + ggml_vk_soft_max(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale, max_bias, m0, m1, n_head_log2); } break; case GGML_OP_DIAG_MASK_INF: { @@ -1649,38 +1702,44 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml case GGML_TYPE_F16: ggml_vk_mul_mat_f16( seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, - ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, ne13, nb10, nb11, nb12, + ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, ne13, nb10, nb11, nb12, nb13, ne0, ne1, r2, r3 ); break; case GGML_TYPE_Q8_0: ggml_vk_mul_mat_q8_0( seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, - ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3 + ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, + nb01, nb02, nb03, nb11, nb12, nb13, r2, r3 ); break; case GGML_TYPE_Q4_0: ggml_vk_mul_mat_q4_0( seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, - ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3 + ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, + nb01, nb02, nb03, nb11, nb12, nb13, r2, r3 ); break; case GGML_TYPE_Q4_1: ggml_vk_mul_mat_q4_1( seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, - ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3 + ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, + nb01, nb02, nb03, nb11, nb12, nb13, r2, r3 ); break; case GGML_TYPE_Q4_K: ggml_vk_mul_mat_q4_k( seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, - ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, ne12/ne02, ne13/ne03 + ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, + nb01, nb02, nb03, nb11, nb12, nb13, r2, r3 ); break; case GGML_TYPE_Q6_K: ggml_vk_mul_mat_q6_k( seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, - ne00, ne10, ne0, ne1, ne01, ne11, ne12, ne02 + ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, + nb01, nb02, nb03, nb11, nb12, nb13, r2, r3 ); break; default: { @@ -1709,13 +1768,6 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml } break; case GGML_OP_ROPE: { -#pragma message("TODO: implement phi3 frequency factors support") -#pragma message(" https://github.com/ggerganov/llama.cpp/pull/7225") - GGML_ASSERT(dst->src[2] == nullptr && "phi3 frequency factors not implemented yet"); - -#pragma message("TODO: update rope NORM mode to match NEOX mode") -#pragma message(" https://github.com/ggerganov/llama.cpp/pull/7634") - GGML_ASSERT(ne10 == ne02); GGML_ASSERT(src0t == dstt); // const int n_past = ((int32_t *) dst->op_params)[0]; @@ -1724,6 +1776,8 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml // skip 3, n_ctx used in GLM RoPE, unimplemented in Vulkan const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; + const bool has_freq_factors = dst->src[2] != nullptr; + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); @@ -1732,8 +1786,8 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); ggml_vk_rope( - seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, src0t, n_dims, mode, n_ctx_orig, - freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, + seq, id_src0, id_src1, id_src2, id_dst, off_src0, off_src1, off_src2, off_dst, src0t, n_dims, mode, n_ctx_orig, + freq_base, freq_scale, has_freq_factors, ext_factor, attn_factor, beta_fast, beta_slow, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, nb0, nb1, nb2, nb3 ); } break; @@ -2176,9 +2230,12 @@ static const struct ggml_backend_reg_i ggml_backend_kompute_reg_i = { ggml_backend_reg_t ggml_backend_kompute_reg() { static ggml_backend_reg reg = { - /* .iface = */ ggml_backend_kompute_reg_i, - /* .context = */ nullptr, + /* .api_version = */ GGML_BACKEND_API_VERSION, + /* .iface = */ ggml_backend_kompute_reg_i, + /* .context = */ nullptr, }; return ® } + +GGML_BACKEND_DL_IMPL(ggml_backend_kompute_reg) diff --git a/ggml/src/ggml-kompute/kompute-shaders/common.comp b/ggml/src/ggml-kompute/kompute-shaders/common.comp index 2aaddf704a7..dbe4cf804e6 100644 --- a/ggml/src/ggml-kompute/kompute-shaders/common.comp +++ b/ggml/src/ggml-kompute/kompute-shaders/common.comp @@ -3,6 +3,7 @@ #extension GL_EXT_shader_explicit_arithmetic_types_float16: require #extension GL_EXT_shader_explicit_arithmetic_types_int8: require #extension GL_EXT_shader_explicit_arithmetic_types_int16: require +#extension GL_EXT_shader_explicit_arithmetic_types_int64: require #extension GL_EXT_control_flow_attributes: enable #extension GL_KHR_shader_subgroup_arithmetic : require #extension GL_EXT_debug_printf : enable diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp index 8f0a9031f7a..0ab1b2fc20e 100644 --- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp +++ b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp @@ -20,12 +20,14 @@ layout (push_constant) uniform parameter { uint nb00; uint nb01; uint nb02; + uint nb03; int ne10; int ne11; int ne12; uint nb10; uint nb11; uint nb12; + uint nb13; int ne0; int ne1; uint r2; @@ -42,7 +44,7 @@ void main() { const uint i12 = im%pcs.ne12; const uint i13 = im/pcs.ne12; - const uint offset0 = r0*pcs.nb01 + (i12/pcs.r2)*pcs.nb02 + (i13/pcs.r3)*pcs.nb02*pcs.ne02; + const uint offset0 = r0*pcs.nb01 + (i12/pcs.r2)*pcs.nb02 + (i13/pcs.r3)*pcs.nb03; const uint x = offset0 / 2 + pcs.inAOff; // Based from inA @@ -52,7 +54,7 @@ void main() { break; } - const uint y = (r1*pcs.nb11 + im*pcs.nb12) / 4 + pcs.inBOff; // Based from inB + const uint y = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff; float sumf = 0; for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) { diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp index fc8e45aa977..a5752a3a006 100644 --- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp +++ b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp @@ -24,8 +24,14 @@ layout (push_constant) uniform parameter { int ne01; int ne02; int ne12; - int r2; - int r3; + uint nb01; + uint nb02; + uint nb03; + uint nb11; + uint nb12; + uint nb13; + uint r2; + uint r3; } pcs; void main() { @@ -50,10 +56,11 @@ void main() { const uint i12 = im%pcs.ne12; const uint i13 = im/pcs.ne12; - const uint offset0 = (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02); + const uint offset0 = first_row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK); + const uint offset1 = r1*pcs.nb11 + (i12 )*pcs.nb12 + (i13 )*pcs.nb13; - const uint xblk = ib_row + offset0 + pcs.inAOff; - const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff; + const uint xblk = offset0 + pcs.inAOff; + const uint y = (offset1 / 4) + pcs.inBOff; float yl[16]; float yh[16]; @@ -74,7 +81,7 @@ void main() { } for (int row = 0; row < N_DST; row++) { - uint row_idx = row * nb; + uint row_idx = row * (pcs.nb01 / SIZE_OF_BLOCK); uint16_t sc_0 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 0); uint16_t sc_1 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 2); diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp index c9baebdf4ba..d331d1a7057 100644 --- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp +++ b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp @@ -21,7 +21,16 @@ layout (push_constant) uniform parameter { int ne0; int ne1; int ne01; - int gqa; + int ne02; + int ne12; + uint nb01; + uint nb02; + uint nb03; + uint nb11; + uint nb12; + uint nb13; + uint r2; + uint r3; } pcs; void main() { @@ -34,12 +43,15 @@ void main() { const uint r0 = gl_WorkGroupID.x; const uint r1 = gl_WorkGroupID.y; - const uint r2 = gl_WorkGroupID.z; + const uint im = gl_WorkGroupID.z; const uint row = (r0 * gl_NumSubgroups + gl_SubgroupID); - const uint offset0 = r2/pcs.gqa*(nb*pcs.ne0); - const uint x = row * nb + offset0; // Based from inA without base offset - const uint yy = r1*pcs.ne10 + r2*pcs.ne00*pcs.ne1+pcs.inBOff; // Based from inB + + const uint i12 = im%pcs.ne12; + const uint i13 = im/pcs.ne12; + + const uint x = row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK); + const uint yy = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff; float sumf = 0; @@ -89,6 +101,6 @@ void main() { const float tot = subgroupAdd(sumf); if (subgroupElect()) { - out_[r1*pcs.ne0 + r2*pcs.ne0*pcs.ne1 + row + pcs.outOff] = tot; + out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + row + pcs.outOff] = tot; } } diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp index 440b5ab2c81..a6517cc1f19 100644 --- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp +++ b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp @@ -14,10 +14,15 @@ void main() { const uint i12 = im%pcs.ne12; const uint i13 = im/pcs.ne12; - const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02); + // pointers to src0 rows + uint ax[N_ROWS]; + for (int row = 0; row < N_ROWS; ++row) { + const uint offset0 = (first_row + row)*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK); + + ax[row] = offset0 + pcs.inAOff; + } - const uint x = offset0; // Based from inA without base offset - const uint y = r1*uint(pcs.ne10)+im*pcs.ne00*pcs.ne1+pcs.inBOff; // Based from inB + const uint y = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff; float sumf[N_ROWS] = {0.0f, 0.0f, 0.0f, 0.0f}; @@ -32,8 +37,7 @@ void main() { for (uint ib = ix; ib < nb; ib += 16) { for (int row = 0; row < N_ROWS; row++) { - const uint block_index = x + ib + row * nb; - sumf[row] += block_q_n_dot_y(block_index, yb, il); + sumf[row] += block_q_n_dot_y(ax[row] + ib, yb, il); } yb += BLOCKS_IN_QUANT * 16; diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp index 7912b09ac69..a9a2f22180f 100644 --- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp +++ b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp @@ -1,5 +1,5 @@ layout(local_size_x_id = 0) in; -layout(local_size_y = 1) in; +layout(local_size_y = 8) in; layout(local_size_z = 1) in; layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; }; @@ -17,6 +17,12 @@ layout (push_constant) uniform parameter { int ne12; int ne0; int ne1; + uint nb01; + uint nb02; + uint nb03; + uint nb11; + uint nb12; + uint nb13; uint r2; uint r3; } pcs; diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_f16.comp deleted file mode 100644 index 0ecfb2eab52..00000000000 --- a/ggml/src/ggml-kompute/kompute-shaders/op_rope_f16.comp +++ /dev/null @@ -1,73 +0,0 @@ -#version 450 - -#include "rope_common.comp" - -layout(binding = 0) buffer restrict readonly tensorInA { float16_t inA[]; }; -layout(binding = 1) buffer restrict readonly tensorInB { int inB[]; }; -layout(binding = 2) buffer restrict writeonly tensorOut { float16_t out_[]; }; - -void main() { - const uint i3 = gl_WorkGroupID.z; - const uint i2 = gl_WorkGroupID.y; - const uint i1 = gl_WorkGroupID.x; - - const bool is_neox = (pcs.mode & GGML_ROPE_TYPE_NEOX) != 0; - - float corr_dims[2]; - rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims); - - const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims); - - const int p = inB[pcs.inBOff + i2]; - - float theta = float(p); - - if (!is_neox) { - for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) { - float cos_theta, sin_theta; - rope_yarn(theta, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta); - - theta *= theta_scale; - - const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in - const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 2) + pcs.outOff; // Based from out_ - - const float x0 = float(inA[src]); - const float x1 = float(inA[src+1]); - - out_[dst_data] = float16_t(x0*cos_theta - x1*sin_theta); - out_[dst_data+1] = float16_t(x0*sin_theta + x1*cos_theta); - } - } else { - const float inv_ndims = -1.f/pcs.n_dims; - for (uint ic = 0; ic < pcs.n_dims; ic += 2) { - const uint cur_rot = ic; - - float cos_theta, sin_theta; - rope_yarn(theta, pcs.freq_scale, corr_dims, cur_rot, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta); - - theta *= theta_scale; - - const uint i0 = ic/2; - - const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in - const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 2) + pcs.outOff; // Based from out_ - - const float x0 = float(inA[src]); - const float x1 = float(inA[src+pcs.n_dims/2]); - - out_[dst_data] = float16_t(x0*cos_theta - x1*sin_theta); - out_[dst_data+pcs.n_dims/2] = float16_t(x0*sin_theta + x1*cos_theta); - } - - for (uint ic = pcs.n_dims; ic < pcs.ne0; ic += 2) { - const uint i0 = ic; - - const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in - const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 2) + pcs.outOff; // Based from out_ - - out_[dst_data + 0] = inA[src + 0]; - out_[dst_data + 1] = inA[src + 1]; - } - } -} diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_f32.comp deleted file mode 100644 index cec0fd9a5d1..00000000000 --- a/ggml/src/ggml-kompute/kompute-shaders/op_rope_f32.comp +++ /dev/null @@ -1,73 +0,0 @@ -#version 450 - -#include "rope_common.comp" - -layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; }; -layout(binding = 1) buffer restrict readonly tensorInB { int inB[]; }; -layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; }; - -void main() { - const uint i3 = gl_WorkGroupID.z; - const uint i2 = gl_WorkGroupID.y; - const uint i1 = gl_WorkGroupID.x; - - const bool is_neox = (pcs.mode & GGML_ROPE_TYPE_NEOX) != 0; - - float corr_dims[2]; - rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims); - - const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims); - - const int p = inB[pcs.inBOff + i2]; - - float theta = float(p); - - if (!is_neox) { - for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) { - float cos_theta, sin_theta; - rope_yarn(theta, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta); - - theta *= theta_scale; - - const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in - const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_ - - const float x0 = inA[src]; - const float x1 = inA[src+1]; - - out_[dst_data] = x0*cos_theta - x1*sin_theta; - out_[dst_data+1] = x0*sin_theta + x1*cos_theta; - } - } else { - const float inv_ndims = -1.f/pcs.n_dims; - for (uint ic = 0; ic < pcs.n_dims; ic += 2) { - const uint cur_rot = ic; - - float cos_theta, sin_theta; - rope_yarn(theta, pcs.freq_scale, corr_dims, cur_rot, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta); - - theta *= theta_scale; - - const uint i0 = ic/2; - - const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in - const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_ - - const float x0 = inA[src]; - const float x1 = inA[src+pcs.n_dims/2]; - - out_[dst_data] = x0*cos_theta - x1*sin_theta; - out_[dst_data+pcs.n_dims/2] = x0*sin_theta + x1*cos_theta; - } - - for (uint ic = pcs.n_dims; ic < pcs.ne0; ic += 2) { - const uint i0 = ic; - - const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in - const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_ - - out_[dst_data + 0] = inA[src + 0]; - out_[dst_data + 1] = inA[src + 1]; - } - } -} diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp new file mode 100644 index 00000000000..63659cbfe55 --- /dev/null +++ b/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp @@ -0,0 +1,52 @@ +#version 450 + +#include "rope_common.comp" + +layout(binding = 0) buffer restrict readonly tensorInA { float16_t inA[]; }; +layout(binding = 1) buffer restrict readonly tensorInB { int inB[]; }; +layout(binding = 2) buffer restrict readonly tensorInC { float inC[]; }; +layout(binding = 3) buffer restrict writeonly tensorOut { float16_t out_[]; }; + +void main() { + const uint i3 = gl_WorkGroupID.z; + const uint i2 = gl_WorkGroupID.y; + const uint i1 = gl_WorkGroupID.x; + + float corr_dims[2]; + rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims); + + const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims); + + float theta_base = float(inB[pcs.inBOff + i2]); + float inv_ndims = -1.f/pcs.n_dims; + + float cos_theta; + float sin_theta; + + for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) { + if (i0 < pcs.n_dims) { + uint ic = i0/2; + + float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0); + + const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f; + + rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta); + + const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + ic*pcs.nb00) / 2) + pcs.inAOff; // Based from in + const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + ic*pcs.nb0) / 2) + pcs.outOff; // Based from out_ + + const float x0 = float(inA[src]); + const float x1 = float(inA[src+pcs.n_dims/2]); + + out_[dst_data] = float16_t(x0*cos_theta - x1*sin_theta); + out_[dst_data+pcs.n_dims/2] = float16_t(x0*sin_theta + x1*cos_theta); + } else { + const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in + const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 2) + pcs.outOff; // Based from out_ + + out_[dst_data] = inA[src]; + out_[dst_data+1] = inA[src+1]; + } + } +} diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp new file mode 100644 index 00000000000..4df56204d72 --- /dev/null +++ b/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp @@ -0,0 +1,52 @@ +#version 450 + +#include "rope_common.comp" + +layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; }; +layout(binding = 1) buffer restrict readonly tensorInB { int inB[]; }; +layout(binding = 2) buffer restrict readonly tensorInC { float inC[]; }; +layout(binding = 3) buffer restrict writeonly tensorOut { float out_[]; }; + +void main() { + const uint i3 = gl_WorkGroupID.z; + const uint i2 = gl_WorkGroupID.y; + const uint i1 = gl_WorkGroupID.x; + + float corr_dims[2]; + rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims); + + const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims); + + float theta_base = float(inB[pcs.inBOff + i2]); + float inv_ndims = -1.f/pcs.n_dims; + + float cos_theta; + float sin_theta; + + for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) { + if (i0 < pcs.n_dims) { + uint ic = i0/2; + + float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0); + + const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f; + + rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta); + + const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + ic*pcs.nb00) / 4) + pcs.inAOff; // Based from in + const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + ic*pcs.nb0) / 4) + pcs.outOff; // Based from out_ + + const float x0 = inA[src]; + const float x1 = inA[src+pcs.n_dims/2]; + + out_[dst_data] = x0*cos_theta - x1*sin_theta; + out_[dst_data+pcs.n_dims/2] = x0*sin_theta + x1*cos_theta; + } else { + const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in + const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_ + + out_[dst_data] = inA[src]; + out_[dst_data+1] = inA[src+1]; + } + } +} diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp new file mode 100644 index 00000000000..a3c0eda8bd3 --- /dev/null +++ b/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp @@ -0,0 +1,52 @@ +#version 450 + +#include "rope_common.comp" + +layout(binding = 0) buffer restrict readonly tensorInA { float16_t inA[]; }; +layout(binding = 1) buffer restrict readonly tensorInB { int inB[]; }; +layout(binding = 2) buffer restrict readonly tensorInC { float inC[]; }; +layout(binding = 3) buffer restrict writeonly tensorOut { float16_t out_[]; }; + +void main() { + const uint i3 = gl_WorkGroupID.z; + const uint i2 = gl_WorkGroupID.y; + const uint i1 = gl_WorkGroupID.x; + + float corr_dims[2]; + rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims); + + const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims); + + float theta_base = float(inB[pcs.inBOff + i2]); + float inv_ndims = -1.f/pcs.n_dims; + + float cos_theta; + float sin_theta; + + for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) { + if (i0 < pcs.n_dims) { + uint ic = i0/2; + + float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0); + + const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f; + + rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta); + + const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in + const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 2) + pcs.outOff; // Based from out_ + + const float x0 = float(inA[src]); + const float x1 = float(inA[src+1]); + + out_[dst_data] = float16_t(x0*cos_theta - x1*sin_theta); + out_[dst_data+1] = float16_t(x0*sin_theta + x1*cos_theta); + } else { + const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in + const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 2) + pcs.outOff; // Based from out_ + + out_[dst_data] = inA[src]; + out_[dst_data+1] = inA[src+1]; + } + } +} diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp new file mode 100644 index 00000000000..b7963ae7253 --- /dev/null +++ b/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp @@ -0,0 +1,52 @@ +#version 450 + +#include "rope_common.comp" + +layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; }; +layout(binding = 1) buffer restrict readonly tensorInB { int inB[]; }; +layout(binding = 2) buffer restrict readonly tensorInC { float inC[]; }; +layout(binding = 3) buffer restrict writeonly tensorOut { float out_[]; }; + +void main() { + const uint i3 = gl_WorkGroupID.z; + const uint i2 = gl_WorkGroupID.y; + const uint i1 = gl_WorkGroupID.x; + + float corr_dims[2]; + rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims); + + const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims); + + float theta_base = float(inB[pcs.inBOff + i2]); + float inv_ndims = -1.f/pcs.n_dims; + + float cos_theta; + float sin_theta; + + for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) { + if (i0 < pcs.n_dims) { + uint ic = i0/2; + + float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0); + + const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f; + + rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta); + + const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in + const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_ + + const float x0 = inA[src]; + const float x1 = inA[src+1]; + + out_[dst_data] = x0*cos_theta - x1*sin_theta; + out_[dst_data+1] = x0*sin_theta + x1*cos_theta; + } else { + const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in + const uint dst_data = uint((i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_ + + out_[dst_data] = inA[src]; + out_[dst_data+1] = inA[src+1]; + } + } +} diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp b/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp index 7bc9176caba..4165295bf4b 100644 --- a/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp +++ b/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp @@ -18,6 +18,10 @@ layout(push_constant) uniform PushConstants { int ne01; int ne02; float scale; + float max_bias; + float m0; + float m1; + uint n_head_log2; int mask; } pcs; @@ -34,17 +38,29 @@ void main() { const uint pmask = i01*pcs.ne00 + pcs.inBOff; // Based from inB const uint pdst = extra_off + pcs.outOff; // Based from out_ + float slope = 1.0f; + + // ALiBi + if (pcs.max_bias > 0.0f) { + int64_t h = i02; + + float base = h < pcs.n_head_log2 ? pcs.m0 : pcs.m1; + int64_t exp = h < pcs.n_head_log2 ? h + 1 : 2*(h - pcs.n_head_log2) + 1; + + slope = pow(base, float(exp)); + } + // parallel max float localMax = uintBitsToFloat(0xFF800000); for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) { - localMax = max(localMax, inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? inB[pmask + i00] : 0.0f)); + localMax = max(localMax, inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? slope*inB[pmask + i00] : 0.0f)); } float max_ = subgroupMax(localMax); // parallel sum float localSum = 0.0f; for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) { - const float exp_psrc0 = exp(inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? inB[pmask + i00] : 0.0f) - max_); + const float exp_psrc0 = exp(inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? slope*inB[pmask + i00] : 0.0f) - max_); localSum += exp_psrc0; out_[pdst + i00] = exp_psrc0; } diff --git a/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp b/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp index df4702896d4..0fca640dcc2 100644 --- a/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp +++ b/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp @@ -8,12 +8,14 @@ layout(local_size_x = 1) in; layout (push_constant) uniform parameter { uint inAOff; uint inBOff; + uint inCOff; uint outOff; int n_dims; int mode; int n_ctx_orig; float freq_base; float freq_scale; + bool has_freq_factors; float ext_factor; float attn_factor; float beta_fast; diff --git a/ggml/src/ggml-metal/CMakeLists.txt b/ggml/src/ggml-metal/CMakeLists.txt index b237d79f47d..1bad2720682 100644 --- a/ggml/src/ggml-metal/CMakeLists.txt +++ b/ggml/src/ggml-metal/CMakeLists.txt @@ -4,19 +4,16 @@ find_library(METALKIT_FRAMEWORK MetalKit REQUIRED) message(STATUS "Metal framework found") -add_library(ggml-metal - ggml-metal.m - ) +ggml_add_backend_library(ggml-metal + ggml-metal.m + ) target_link_libraries(ggml-metal PRIVATE - ggml-base ${FOUNDATION_LIBRARY} ${METAL_FRAMEWORK} ${METALKIT_FRAMEWORK} ) -target_include_directories(ggml-metal PRIVATE . ..) - if (GGML_METAL_NDEBUG) add_compile_definitions(GGML_METAL_NDEBUG) endif() diff --git a/ggml/src/ggml-metal/ggml-metal-impl.h b/ggml/src/ggml-metal/ggml-metal-impl.h index 53c13549650..e3dc25f1686 100644 --- a/ggml/src/ggml-metal/ggml-metal-impl.h +++ b/ggml/src/ggml-metal/ggml-metal-impl.h @@ -102,6 +102,21 @@ typedef struct { uint64_t nb3; } ggml_metal_kargs_cpy; +typedef struct { + int64_t ne10; + int64_t ne11; + int64_t ne12; + uint64_t nb10; + uint64_t nb11; + uint64_t nb12; + uint64_t nb13; + uint64_t nb1; + uint64_t nb2; + uint64_t nb3; + uint64_t offs; + bool inplace; +} ggml_metal_kargs_set; + typedef struct { int32_t ne00; int32_t ne01; @@ -192,6 +207,30 @@ typedef struct { int16_t r3; } ggml_metal_kargs_mul_mv; +typedef struct { + int32_t ne00; + int32_t ne01; + int32_t ne02; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + uint64_t nb03; + int32_t ne10; + int32_t ne11; + int32_t ne12; + uint64_t nb10; + uint64_t nb11; + uint64_t nb12; + uint64_t nb13; + int32_t ne0; + int32_t ne1; + int16_t r2; + int16_t r3; + int16_t nsg; + int16_t nxpsg; + int16_t r1ptg; +} ggml_metal_kargs_mul_mv_ext; + typedef struct { int32_t nei0; int32_t nei1; diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m index d1abb3cef0e..f80fda7a4ca 100644 --- a/ggml/src/ggml-metal/ggml-metal.m +++ b/ggml/src/ggml-metal/ggml-metal.m @@ -175,6 +175,46 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_3, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_4, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_5, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_2, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_3, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_4, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_5, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_2, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_3, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_4, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_5, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_2, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_3, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_4, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_5, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_2, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_3, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_4, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_5, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_2, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_3, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_4, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_5, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_2, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_3, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_4, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_5, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_2, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_3, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_4, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_5, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_2, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_3, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_4, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_5, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_2, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_3, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_4, + GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_5, GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, @@ -266,8 +306,11 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte GGML_METAL_KERNEL_TYPE_IM2COL_F32, GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F16, GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F32, + GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F32_F32, + GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F16_F32, GGML_METAL_KERNEL_TYPE_UPSCALE_F32, GGML_METAL_KERNEL_TYPE_PAD_F32, + GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32, GGML_METAL_KERNEL_TYPE_ARANGE_F32, GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32, GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, @@ -329,6 +372,8 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H256, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256, + GGML_METAL_KERNEL_TYPE_SET_I32, + GGML_METAL_KERNEL_TYPE_SET_F32, GGML_METAL_KERNEL_TYPE_CPY_F32_F32, GGML_METAL_KERNEL_TYPE_CPY_F32_F16, GGML_METAL_KERNEL_TYPE_CPY_F32_BF16, @@ -350,6 +395,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte GGML_METAL_KERNEL_TYPE_SUM_ROWS, GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32, GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32, + GGML_METAL_KERNEL_TYPE_ARGMAX, GGML_METAL_KERNEL_TYPE_COUNT }; @@ -699,6 +745,46 @@ @implementation GGMLMetalClass GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, has_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, has_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2, mul_mv_ext_f16_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_3, mul_mv_ext_f16_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_4, mul_mv_ext_f16_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_5, mul_mv_ext_f16_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_2, mul_mv_ext_q4_0_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_3, mul_mv_ext_q4_0_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_4, mul_mv_ext_q4_0_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_5, mul_mv_ext_q4_0_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_2, mul_mv_ext_q4_1_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_3, mul_mv_ext_q4_1_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_4, mul_mv_ext_q4_1_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_5, mul_mv_ext_q4_1_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_2, mul_mv_ext_q5_0_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_3, mul_mv_ext_q5_0_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_4, mul_mv_ext_q5_0_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_5, mul_mv_ext_q5_0_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_2, mul_mv_ext_q5_1_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_3, mul_mv_ext_q5_1_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_4, mul_mv_ext_q5_1_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_5, mul_mv_ext_q5_1_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_2, mul_mv_ext_q8_0_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_3, mul_mv_ext_q8_0_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_4, mul_mv_ext_q8_0_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_5, mul_mv_ext_q8_0_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_2, mul_mv_ext_q4_K_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_3, mul_mv_ext_q4_K_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_4, mul_mv_ext_q4_K_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_5, mul_mv_ext_q4_K_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_2, mul_mv_ext_q5_K_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_3, mul_mv_ext_q5_K_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_4, mul_mv_ext_q5_K_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_5, mul_mv_ext_q5_K_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_2, mul_mv_ext_q6_K_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_3, mul_mv_ext_q6_K_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_4, mul_mv_ext_q6_K_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_5, mul_mv_ext_q6_K_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_2, mul_mv_ext_iq4_nl_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_3, mul_mv_ext_iq4_nl_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_4, mul_mv_ext_iq4_nl_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_5, mul_mv_ext_iq4_nl_f32_r1_5, has_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, has_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, has_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, has_simdgroup_reduction); @@ -790,8 +876,11 @@ @implementation GGMLMetalClass GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32, im2col_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F16, im2col_ext_f16, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F32, im2col_ext_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F32_F32, conv_transpose_1d_f32_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F16_F32, conv_transpose_1d_f16_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32, upscale_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32, pad_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32, pad_reflect_1d_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32, timestep_embedding_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32, arange_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true); @@ -853,6 +942,8 @@ @implementation GGMLMetalClass GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256, flash_attn_ext_vec_q5_0_h256, has_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H256, flash_attn_ext_vec_q5_1_h256, has_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256, flash_attn_ext_vec_q8_0_h256, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_F32, set_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_I32, set_i32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_BF16, cpy_f32_bf16, use_bfloat); @@ -872,6 +963,7 @@ @implementation GGMLMetalClass GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIN, sin, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS, cos, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX, argmax, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32, pool_2d_avg_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32, pool_2d_max_f32, true); } @@ -989,6 +1081,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex case GGML_OP_REPEAT: case GGML_OP_SCALE: case GGML_OP_CLAMP: + case GGML_OP_CONV_TRANSPOSE_1D: return true; case GGML_OP_SQR: case GGML_OP_SQRT: @@ -997,9 +1090,11 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex return ggml_is_contiguous(op->src[0]); case GGML_OP_SUM_ROWS: case GGML_OP_SOFT_MAX: - case GGML_OP_RMS_NORM: case GGML_OP_GROUP_NORM: return has_simdgroup_reduction; + case GGML_OP_RMS_NORM: + return has_simdgroup_reduction && (op->ne[0] % 4 == 0); + case GGML_OP_ARGMAX: case GGML_OP_NORM: case GGML_OP_ROPE: return true; @@ -1010,6 +1105,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex case GGML_OP_POOL_2D: case GGML_OP_UPSCALE: case GGML_OP_PAD: + case GGML_OP_PAD_REFLECT_1D: case GGML_OP_ARANGE: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_ARGSORT: @@ -1067,6 +1163,16 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex return false; }; } + case GGML_OP_SET: + { + switch (op->src[0]->type) { + case GGML_TYPE_F32: + case GGML_TYPE_I32: + return true; + default: + return false; + }; + } case GGML_OP_DIAG_MASK_INF: case GGML_OP_GET_ROWS: { @@ -1927,340 +2033,490 @@ static void ggml_metal_encode_node( // find the break-even point where the matrix-matrix kernel becomes more efficient compared // to the matrix-vector kernel - int ne11_mm_min = 1; + const int ne11_mm_min = 4; + + // first try to use small-batch mat-mv kernels + // these should be efficient for BS [2, ~8] + if (src1t == GGML_TYPE_F32 && (ne00%256 == 0) && + ( + ( + ( + src0t == GGML_TYPE_F16 || // TODO: helper function + src0t == GGML_TYPE_Q4_0 || + src0t == GGML_TYPE_Q4_1 || + src0t == GGML_TYPE_Q5_0 || + src0t == GGML_TYPE_Q5_1 || + src0t == GGML_TYPE_Q8_0 || + src0t == GGML_TYPE_IQ4_NL || + false) && (ne11 >= 2 && ne11 <= 8) + ) || + ( + ( + src0t == GGML_TYPE_Q4_K || + src0t == GGML_TYPE_Q5_K || + src0t == GGML_TYPE_Q6_K || + false) && (ne11 >= 4 && ne11 <= 8) + ) + ) + ) { + // TODO: determine the optimal parameters based on grid utilization + // I still don't know why we should not always use the maximum available threads: + // + // nsg = pipeline.maxTotalThreadsPerThreadgroup / 32 + // + // my current hypothesis is that the work grid is not evenly divisible for different nsg + // values and there can be some tail effects when nsg is high. need to confirm this + // + const int nsg = 2; // num simdgroups per threadgroup + const int nxpsg = ne11 < 3 ? 16 : 8; // num threads along row per simdgroup + const int nypsg = 32/nxpsg; // num threads along col per simdgroup (i.e. a simdgroup processes that many src0 rows at a time) + const int r0ptg = nypsg*nsg; // num src0 rows per threadgroup + int r1ptg = 4; // num src1 rows per threadgroup + + // note: not sure how optimal are those across all different hardware. there might be someting cleverer + switch (ne11) { + case 2: + r1ptg = 2; break; + case 3: + case 6: + r1ptg = 3; break; + case 4: + case 7: + case 8: + r1ptg = 4; break; + case 5: + r1ptg = 5; break; + }; -#if 0 - // the numbers below are measured on M2 Ultra for 7B and 13B models - // these numbers do not translate to other devices or model sizes - // TODO: need to find a better approach - if ([device.name isEqualToString:@"Apple M2 Ultra"]) { - switch (src0t) { - case GGML_TYPE_F16: ne11_mm_min = 2; break; - case GGML_TYPE_Q8_0: ne11_mm_min = 7; break; - case GGML_TYPE_Q2_K: ne11_mm_min = 15; break; - case GGML_TYPE_Q3_K: ne11_mm_min = 7; break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: ne11_mm_min = 15; break; - case GGML_TYPE_Q4_K: ne11_mm_min = 11; break; - case GGML_TYPE_Q5_0: // not tested yet - case GGML_TYPE_Q5_1: ne11_mm_min = 13; break; // not tested yet - case GGML_TYPE_Q5_K: ne11_mm_min = 7; break; - case GGML_TYPE_Q6_K: ne11_mm_min = 7; break; - default: ne11_mm_min = 1; break; - } - } -#endif + id pipeline = nil; - // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs - // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel - if ([device supportsFamily:MTLGPUFamilyApple7] && - !ggml_is_transposed(src0) && - !ggml_is_transposed(src1) && - src1t == GGML_TYPE_F32 && - ne00 % 32 == 0 && ne00 >= 64 && - (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) { - //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12); - - // some Metal matrix data types require aligned pointers - // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5) - switch (src0->type) { - case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break; - case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8 == 0); break; - case GGML_TYPE_BF16: GGML_ASSERT(nb01 % 8 == 0); break; - default: break; - } + switch (src0->type) { + case GGML_TYPE_F16: + switch (r1ptg) { + case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2].pipeline; break; + case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_3].pipeline; break; + case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_4].pipeline; break; + case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_5].pipeline; break; + default: GGML_ABORT("not implemented"); + } break; + case GGML_TYPE_Q4_0: + switch (r1ptg) { + case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_2].pipeline; break; + case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_3].pipeline; break; + case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_4].pipeline; break; + case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_5].pipeline; break; + default: GGML_ABORT("not implemented"); + } break; + case GGML_TYPE_Q4_1: + switch (r1ptg) { + case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_2].pipeline; break; + case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_3].pipeline; break; + case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_4].pipeline; break; + case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_5].pipeline; break; + default: GGML_ABORT("not implemented"); + } break; + case GGML_TYPE_Q5_0: + switch (r1ptg) { + case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_2].pipeline; break; + case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_3].pipeline; break; + case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_4].pipeline; break; + case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_5].pipeline; break; + default: GGML_ABORT("not implemented"); + } break; + case GGML_TYPE_Q5_1: + switch (r1ptg) { + case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_2].pipeline; break; + case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_3].pipeline; break; + case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_4].pipeline; break; + case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_5].pipeline; break; + default: GGML_ABORT("not implemented"); + } break; + case GGML_TYPE_Q8_0: + switch (r1ptg) { + case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_2].pipeline; break; + case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_3].pipeline; break; + case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_4].pipeline; break; + case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_5].pipeline; break; + default: GGML_ABORT("not implemented"); + } break; + case GGML_TYPE_Q4_K: + switch (r1ptg) { + case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_2].pipeline; break; + case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_3].pipeline; break; + case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_4].pipeline; break; + case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_5].pipeline; break; + default: GGML_ABORT("not implemented"); + } break; + case GGML_TYPE_Q5_K: + switch (r1ptg) { + case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_2].pipeline; break; + case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_3].pipeline; break; + case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_4].pipeline; break; + case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_5].pipeline; break; + default: GGML_ABORT("not implemented"); + } break; + case GGML_TYPE_Q6_K: + switch (r1ptg) { + case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_2].pipeline; break; + case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_3].pipeline; break; + case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_4].pipeline; break; + case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_5].pipeline; break; + default: GGML_ABORT("not implemented"); + } break; + case GGML_TYPE_IQ4_NL: + switch (r1ptg) { + case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_2].pipeline; break; + case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_3].pipeline; break; + case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_4].pipeline; break; + case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_5].pipeline; break; + default: GGML_ABORT("not implemented"); + } break; + default: GGML_ABORT("not implemented"); + } - id pipeline = nil; - - switch (src0->type) { - case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32 ].pipeline; break; - case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32 ].pipeline; break; - case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_BF16_F32 ].pipeline; break; - case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32 ].pipeline; break; - case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32 ].pipeline; break; - case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32 ].pipeline; break; - case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32 ].pipeline; break; - case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32 ].pipeline; break; - case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32 ].pipeline; break; - case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32 ].pipeline; break; - case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32 ].pipeline; break; - case GGML_TYPE_Q5_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32 ].pipeline; break; - case GGML_TYPE_Q6_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32 ].pipeline; break; - case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break; - case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break; - case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break; - case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32 ].pipeline; break; - case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32 ].pipeline; break; - case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32 ].pipeline; break; - case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32 ].pipeline; break; - case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break; - case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break; - default: GGML_ABORT("MUL MAT-MAT not implemented"); - } + ggml_metal_kargs_mul_mv_ext args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.nb03 =*/ nb03, + /*.ne10 =*/ ne10, + /*.ne11 =*/ ne11, + /*.ne12 =*/ ne12, + /*.nb10 =*/ nb10, + /*.nb11 =*/ nb11, + /*.nb12 =*/ nb12, + /*.nb13 =*/ nb13, + /*.ne0 =*/ ne0, + /*.ne1 =*/ ne1, + /*.r2 =*/ r2, + /*.r3 =*/ r3, + /*.nsg =*/ nsg, + /*.nxpsg =*/ nxpsg, + /*.r1ptg =*/ r1ptg, + }; - ggml_metal_kargs_mul_mm args = { - /*.ne00 =*/ ne00, - /*.ne02 =*/ ne02, - /*.nb01 =*/ nb01, - /*.nb02 =*/ nb02, - /*.nb03 =*/ nb03, - /*.ne12 =*/ ne12, - /*.nb10 =*/ nb10, - /*.nb11 =*/ nb11, - /*.nb12 =*/ nb12, - /*.nb13 =*/ nb13, - /*.ne0 =*/ ne0, - /*.ne1 =*/ ne1, - /*.r2 =*/ r2, - /*.r3 =*/ r3, - }; + [encoder setComputePipelineState:pipeline]; + [encoder setBytes:&args length:sizeof(args) atIndex:0]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:3]; + + //printf("ne01 = %lld nr0ptg = %d\n", ne01, nr0ptg); + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + r0ptg - 1)/r0ptg, (ne11 + r1ptg - 1)/r1ptg, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)]; + } else + // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs + // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel + if ([device supportsFamily:MTLGPUFamilyApple7] && + !ggml_is_transposed(src0) && + !ggml_is_transposed(src1) && + src1t == GGML_TYPE_F32 && + ne00 % 32 == 0 && ne00 >= 64 && + (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) { + //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12); + + // some Metal matrix data types require aligned pointers + // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5) + switch (src0->type) { + case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break; + case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8 == 0); break; + case GGML_TYPE_BF16: GGML_ASSERT(nb01 % 8 == 0); break; + default: break; + } + + id pipeline = nil; + + switch (src0->type) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32 ].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32 ].pipeline; break; + case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_BF16_F32 ].pipeline; break; + case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32 ].pipeline; break; + case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32 ].pipeline; break; + case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32 ].pipeline; break; + case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32 ].pipeline; break; + case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32 ].pipeline; break; + case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32 ].pipeline; break; + case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32 ].pipeline; break; + case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32 ].pipeline; break; + case GGML_TYPE_Q5_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32 ].pipeline; break; + case GGML_TYPE_Q6_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32 ].pipeline; break; + case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break; + case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break; + case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break; + case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32 ].pipeline; break; + case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32 ].pipeline; break; + case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32 ].pipeline; break; + case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32 ].pipeline; break; + case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break; + case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break; + default: GGML_ABORT("MUL MAT-MAT not implemented"); + } + + ggml_metal_kargs_mul_mm args = { + /*.ne00 =*/ ne00, + /*.ne02 =*/ ne02, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.nb03 =*/ nb03, + /*.ne12 =*/ ne12, + /*.nb10 =*/ nb10, + /*.nb11 =*/ nb11, + /*.nb12 =*/ nb12, + /*.nb13 =*/ nb13, + /*.ne0 =*/ ne0, + /*.ne1 =*/ ne1, + /*.r2 =*/ r2, + /*.r3 =*/ r3, + }; + + [encoder setComputePipelineState:pipeline]; + [encoder setBytes:&args length:sizeof(args) atIndex:0]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:3]; + + [encoder setThreadgroupMemoryLength:8192 atIndex:0]; + [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)]; + } else { + int nth0 = 32; + int nth1 = 1; + int nrows = 1; + //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12); - [encoder setComputePipelineState:pipeline]; - [encoder setBytes:&args length:sizeof(args) atIndex:0]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:3]; - - [encoder setThreadgroupMemoryLength:8192 atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)]; - } else { - int nth0 = 32; - int nth1 = 1; - int nrows = 1; - //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12); - - id pipeline = nil; - - // use custom matrix x vector kernel - switch (src0t) { - case GGML_TYPE_F32: - { - GGML_ASSERT(src1t == GGML_TYPE_F32); - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline; + id pipeline = nil; + + // use custom matrix x vector kernel + switch (src0t) { + case GGML_TYPE_F32: + { + GGML_ASSERT(src1t == GGML_TYPE_F32); + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline; + nrows = 4; + } break; + case GGML_TYPE_F16: + { + nth0 = 32; + nth1 = 1; + if (src1t == GGML_TYPE_F32) { + if (ne11 * ne12 < 4) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline; + } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline; + nrows = ne11; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32].pipeline; nrows = 4; - } break; - case GGML_TYPE_F16: - { - nth0 = 32; - nth1 = 1; - if (src1t == GGML_TYPE_F32) { - if (ne11 * ne12 < 4) { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline; - } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline; - nrows = ne11; - } else { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32].pipeline; - nrows = 4; - } - } else { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16].pipeline; - nrows = 4; - } - } break; - case GGML_TYPE_BF16: - { - nth0 = 32; - nth1 = 1; - if (src1t == GGML_TYPE_F32) { - if (ne11 * ne12 < 4) { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW].pipeline; - } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4].pipeline; - nrows = ne11; - } else { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32].pipeline; - nrows = 4; - } - } else { - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16].pipeline; - nrows = 4; - } - } break; - case GGML_TYPE_Q4_0: - { - nth0 = 8; - nth1 = 8; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32].pipeline; - } break; - case GGML_TYPE_Q4_1: - { - nth0 = 8; - nth1 = 8; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32].pipeline; - } break; - case GGML_TYPE_Q5_0: - { - nth0 = 8; - nth1 = 8; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32].pipeline; - } break; - case GGML_TYPE_Q5_1: - { - nth0 = 8; - nth1 = 8; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32].pipeline; - } break; - case GGML_TYPE_Q8_0: - { - nth0 = 8; - nth1 = 8; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline; - } break; - case GGML_TYPE_Q2_K: - { - nth0 = 2; - nth1 = 32; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32].pipeline; - } break; - case GGML_TYPE_Q3_K: - { - nth0 = 2; - nth1 = 32; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32].pipeline; - } break; - case GGML_TYPE_Q4_K: - { - nth0 = 4; //1; - nth1 = 8; //32; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32].pipeline; - } break; - case GGML_TYPE_Q5_K: - { - nth0 = 2; - nth1 = 32; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32].pipeline; - } break; - case GGML_TYPE_Q6_K: - { - nth0 = 2; - nth1 = 32; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32].pipeline; - } break; - case GGML_TYPE_IQ2_XXS: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32].pipeline; - } break; - case GGML_TYPE_IQ2_XS: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32].pipeline; - } break; - case GGML_TYPE_IQ3_XXS: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline; - } break; - case GGML_TYPE_IQ3_S: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline; - } break; - case GGML_TYPE_IQ2_S: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline; - } break; - case GGML_TYPE_IQ1_S: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline; - } break; - case GGML_TYPE_IQ1_M: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline; - } break; - case GGML_TYPE_IQ4_NL: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline; - } break; - case GGML_TYPE_IQ4_XS: - { - nth0 = 4; - nth1 = 16; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline; - } break; - default: - { - GGML_LOG_ERROR("Asserting on type %d\n", (int)src0t); - GGML_ABORT("not implemented"); } - }; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16].pipeline; + nrows = 4; + } + } break; + case GGML_TYPE_BF16: + { + nth0 = 32; + nth1 = 1; + if (src1t == GGML_TYPE_F32) { + if (ne11 * ne12 < 4) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW].pipeline; + } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4].pipeline; + nrows = ne11; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32].pipeline; + nrows = 4; + } + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16].pipeline; + nrows = 4; + } + } break; + case GGML_TYPE_Q4_0: + { + nth0 = 8; + nth1 = 8; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32].pipeline; + } break; + case GGML_TYPE_Q4_1: + { + nth0 = 8; + nth1 = 8; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32].pipeline; + } break; + case GGML_TYPE_Q5_0: + { + nth0 = 8; + nth1 = 8; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32].pipeline; + } break; + case GGML_TYPE_Q5_1: + { + nth0 = 8; + nth1 = 8; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32].pipeline; + } break; + case GGML_TYPE_Q8_0: + { + nth0 = 8; + nth1 = 8; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline; + } break; + case GGML_TYPE_Q2_K: + { + nth0 = 2; + nth1 = 32; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32].pipeline; + } break; + case GGML_TYPE_Q3_K: + { + nth0 = 2; + nth1 = 32; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32].pipeline; + } break; + case GGML_TYPE_Q4_K: + { + nth0 = 4; //1; + nth1 = 8; //32; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32].pipeline; + } break; + case GGML_TYPE_Q5_K: + { + nth0 = 2; + nth1 = 32; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32].pipeline; + } break; + case GGML_TYPE_Q6_K: + { + nth0 = 2; + nth1 = 32; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32].pipeline; + } break; + case GGML_TYPE_IQ2_XXS: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32].pipeline; + } break; + case GGML_TYPE_IQ2_XS: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32].pipeline; + } break; + case GGML_TYPE_IQ3_XXS: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline; + } break; + case GGML_TYPE_IQ3_S: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline; + } break; + case GGML_TYPE_IQ2_S: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline; + } break; + case GGML_TYPE_IQ1_S: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline; + } break; + case GGML_TYPE_IQ1_M: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline; + } break; + case GGML_TYPE_IQ4_NL: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline; + } break; + case GGML_TYPE_IQ4_XS: + { + nth0 = 4; + nth1 = 16; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline; + } break; + default: + { + GGML_LOG_ERROR("Asserting on type %d\n", (int)src0t); + GGML_ABORT("not implemented"); + } + }; - ggml_metal_kargs_mul_mv args = { - /*.ne00 =*/ ne00, - /*.ne01 =*/ ne01, - /*.ne02 =*/ ne02, - /*.nb00 =*/ nb00, - /*.nb01 =*/ nb01, - /*.nb02 =*/ nb02, - /*.nb03 =*/ nb03, - /*.ne10 =*/ ne10, - /*.ne11 =*/ ne11, - /*.ne12 =*/ ne12, - /*.nb10 =*/ nb10, - /*.nb11 =*/ nb11, - /*.nb12 =*/ nb12, - /*.nb13 =*/ nb13, - /*.ne0 =*/ ne0, - /*.ne1 =*/ ne1, - /*.r2 =*/ r2, - /*.r3 =*/ r3, - }; + ggml_metal_kargs_mul_mv args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.nb03 =*/ nb03, + /*.ne10 =*/ ne10, + /*.ne11 =*/ ne11, + /*.ne12 =*/ ne12, + /*.nb10 =*/ nb10, + /*.nb11 =*/ nb11, + /*.nb12 =*/ nb12, + /*.nb13 =*/ nb13, + /*.ne0 =*/ ne0, + /*.ne1 =*/ ne1, + /*.r2 =*/ r2, + /*.r3 =*/ r3, + }; - [encoder setComputePipelineState:pipeline]; - [encoder setBytes:&args length:sizeof(args) atIndex:0]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:3]; + [encoder setComputePipelineState:pipeline]; + [encoder setBytes:&args length:sizeof(args) atIndex:0]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:3]; - if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q5_0 || - src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K || - src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) { - const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128; - [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) { - const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4; - [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) { - const int mem_size = 32*sizeof(float); - [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q4_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q3_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q5_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q6_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } else { - const int64_t ny = (ne11 + nrows - 1)/nrows; - [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - } + if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q5_0 || + src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K || + src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) { + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) { + const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128; + [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) { + const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4; + [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) { + const int mem_size = 32*sizeof(float); + [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_Q4_K) { + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_Q3_K) { + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_Q5_K) { + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_Q6_K) { + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } else { + const int64_t ny = (ne11 + nrows - 1)/nrows; + [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + } } break; case GGML_OP_MUL_MAT_ID: { @@ -2672,7 +2928,6 @@ static void ggml_metal_encode_node( } break; case GGML_OP_GROUP_NORM: { - GGML_ASSERT(ne00 % 4 == 0); GGML_ASSERT(ggml_is_contiguous(src0)); float eps; @@ -2908,6 +3163,49 @@ static void ggml_metal_encode_node( [encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)]; } } break; + case GGML_OP_CONV_TRANSPOSE_1D: + { + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(src1)); + GGML_ASSERT(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; + + const int32_t IC = src1->ne[1]; + const int32_t IL = src1->ne[0]; + + const int32_t K = src0->ne[0]; + + const int32_t OL = dst->ne[0]; + const int32_t OC = dst->ne[1]; + + id pipeline; + + switch (src0->type) { + case GGML_TYPE_F32: { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F32_F32].pipeline; + } break; + case GGML_TYPE_F16: { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F16_F32].pipeline; + } break; + default: GGML_ABORT("fatal error"); + }; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; + [encoder setBytes:&IC length:sizeof( int32_t) atIndex:3]; + [encoder setBytes:&IL length:sizeof( int32_t) atIndex:4]; + [encoder setBytes:&K length:sizeof( int32_t) atIndex:5]; + [encoder setBytes:&s0 length:sizeof( int32_t) atIndex:6]; + [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:7]; + [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:8]; + + [encoder dispatchThreadgroups:MTLSizeMake(OL, OC, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; case GGML_OP_UPSCALE: { GGML_ASSERT(src0->type == GGML_TYPE_F32); @@ -2977,6 +3275,38 @@ static void ggml_metal_encode_node( const int nth = MIN(1024, ne0); + [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_PAD_REFLECT_1D: + { + GGML_ASSERT(src0->type == GGML_TYPE_F32); + + const int32_t p0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t p1 = ((const int32_t *)(dst->op_params))[1]; + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; + [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; + [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:6]; + [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9]; + [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10]; + [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:11]; + [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:12]; + [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:13]; + [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:14]; + [encoder setBytes:&p0 length:sizeof(p0) atIndex:15]; + [encoder setBytes:&p1 length:sizeof(p1) atIndex:16]; + + const int nth = MIN(1024, ne0); + [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; } break; case GGML_OP_ARANGE: @@ -3508,6 +3838,68 @@ static void ggml_metal_encode_node( [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; } break; + case GGML_OP_SET: + { + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + + // src0 and dst as viewed during set + const size_t dst_nb0 = ggml_element_size(src0); + + const size_t dst_nb1 = ((int32_t *) dst->op_params)[0]; + const size_t dst_nb2 = ((int32_t *) dst->op_params)[1]; + const size_t dst_nb3 = ((int32_t *) dst->op_params)[2]; + const size_t offset = ((int32_t *) dst->op_params)[3]; + const bool inplace = (bool) ((int32_t *) dst->op_params)[4]; + + if (!inplace) { + memcpy(((char *) dst->data), ((char *) src0->data), ggml_nbytes(dst)); + } + + const int im0 = (ne10 == 0 ? 0 : ne10-1); + const int im1 = (ne11 == 0 ? 0 : ne11-1); + const int im2 = (ne12 == 0 ? 0 : ne12-1); + const int im3 = (ne13 == 0 ? 0 : ne13-1); + + GGML_ASSERT(offset + im0*dst_nb0 + im1*dst_nb1 + im2*dst_nb2 + im3*dst_nb3 <= ggml_nbytes(dst)); + + id pipeline = nil; + + switch (src0t) { + case GGML_TYPE_F32: + GGML_ASSERT(nb10 == sizeof(float)); + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_F32].pipeline; break; + case GGML_TYPE_I32: + GGML_ASSERT(nb10 == sizeof(int32_t)); + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_I32].pipeline; break; + default: GGML_ABORT("fatal error"); + } + + ggml_metal_kargs_set args = { + /*.ne10 =*/ ne10, + /*.ne11 =*/ ne11, + /*.ne12 =*/ ne12, + /*.nb10 =*/ nb10, + /*.nb11 =*/ nb11, + /*.nb12 =*/ nb12, + /*.nb13 =*/ nb13, + /*.nb1 =*/ dst_nb1, + /*.nb2 =*/ dst_nb2, + /*.nb3 =*/ dst_nb3, + /*.offs =*/ offset, + /*.inplace =*/ inplace, + }; + + const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne10); + + [encoder setComputePipelineState:pipeline]; + [encoder setBytes:&args length:sizeof(args) atIndex:0]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:3]; + + [encoder dispatchThreadgroups:MTLSizeMake(ne11, ne12, ne13) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; case GGML_OP_POOL_2D: { GGML_ASSERT(ggml_is_contiguous(src0)); @@ -3567,6 +3959,31 @@ static void ggml_metal_encode_node( [encoder dispatchThreadgroups:MTLSizeMake(n_tg, 1, 1) threadsPerThreadgroup:MTLSizeMake(n_threads, 1, 1)]; } break; + case GGML_OP_ARGMAX: + { + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(ggml_is_contiguous_1(src0)); + GGML_ASSERT(nb00 == ggml_type_size(src0->type)); + + const int64_t nrows = ggml_nrows(src0); + + int nth = 32; // SIMD width + while (nth < ne00 && nth*ne01*ne02*ne03 < 256) { + nth *= 2; + } + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGMAX].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3]; + [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0]; + [encoder setThreadgroupMemoryLength:32*sizeof(int32_t) atIndex:1]; + + [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; default: { GGML_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op)); @@ -4372,19 +4789,45 @@ static ggml_backend_dev_t ggml_backend_metal_reg_device_get(ggml_backend_reg_t r GGML_UNUSED(index); } +static struct ggml_backend_feature g_ggml_backend_metal_features[] = { +#if defined(GGML_METAL_EMBED_LIBRARY) + { "EMBED_LIBRARY", "1" }, +#endif +#if defined(GGML_METAL_USE_BF16) + { "BF16", "1" }, +#endif + { nil, nil }, +}; + +static struct ggml_backend_feature * ggml_backend_metal_get_features(ggml_backend_reg_t reg) { + return g_ggml_backend_metal_features; + + GGML_UNUSED(reg); +} + +static void * ggml_backend_metal_get_proc_address(ggml_backend_reg_t reg, const char * name) { + if (strcmp(name, "ggml_backend_get_features") == 0) { + return (void *)ggml_backend_metal_get_features; + } + + return NULL; + + GGML_UNUSED(reg); +} static struct ggml_backend_reg_i ggml_backend_metal_reg_i = { /* .get_name = */ ggml_backend_metal_reg_get_name, /* .device_count = */ ggml_backend_metal_reg_device_count, /* .device_get = */ ggml_backend_metal_reg_device_get, - /* .get_proc_address = */ NULL, + /* .get_proc_address = */ ggml_backend_metal_get_proc_address, }; ggml_backend_reg_t ggml_backend_metal_reg(void) { // TODO: make this thread-safe somehow? { g_ggml_backend_metal_reg = (struct ggml_backend_reg) { - /* .iface = */ ggml_backend_metal_reg_i, - /* .context = */ NULL, + /* .api_version = */ GGML_BACKEND_API_VERSION, + /* .iface = */ ggml_backend_metal_reg_i, + /* .context = */ NULL, }; g_ggml_backend_metal_device = (struct ggml_backend_device) { @@ -4396,3 +4839,5 @@ ggml_backend_reg_t ggml_backend_metal_reg(void) { return &g_ggml_backend_metal_reg; } + +GGML_BACKEND_DL_IMPL(ggml_backend_metal_reg) diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal index 971f5054bce..8ba43904d0c 100644 --- a/ggml/src/ggml-metal/ggml-metal.metal +++ b/ggml/src/ggml-metal/ggml-metal.metal @@ -47,6 +47,11 @@ void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) reg = (type4x4)(*src); } +template +void dequantize_f16_t4(device const half4 * src, short il, thread type4 & reg) { + reg = (type4)(*(src + il)); +} + #if defined(GGML_METAL_USE_BF16) template void dequantize_bf16(device const bfloat4x4 * src, short il, thread type4x4 & reg) { @@ -55,7 +60,7 @@ void dequantize_bf16(device const bfloat4x4 * src, short il, thread type4x4 & re #endif template -void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) { +void dequantize_q4_0(device const block_q4_0 * xb, short il, thread type4x4 & reg) { device const uint16_t * qs = ((device const uint16_t *)xb + 1); const float d1 = il ? (xb->d / 16.h) : xb->d; const float d2 = d1 / 256.f; @@ -73,8 +78,23 @@ void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg reg = (type4x4) reg_f; } +template +void dequantize_q4_0_t4(device const block_q4_0 * xb, short il, thread type4 & reg) { + device const uint16_t * qs = ((device const uint16_t *)xb + 1); + const float d1 = (il/4) ? (xb->d / 16.h) : xb->d; + const float d2 = d1 / 256.f; + const float md = -8.h * xb->d; + const ushort mask0 = (il/4) ? 0x00F0 : 0x000F; + const ushort mask1 = mask0 << 8; + + for (int i = 0; i < 2; i++) { + reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + md; + reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + md; + } +} + template -void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg) { +void dequantize_q4_1(device const block_q4_1 * xb, short il, thread type4x4 & reg) { device const uint16_t * qs = ((device const uint16_t *)xb + 2); const float d1 = il ? (xb->d / 16.h) : xb->d; const float d2 = d1 / 256.f; @@ -92,8 +112,23 @@ void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg reg = (type4x4) reg_f; } +template +void dequantize_q4_1_t4(device const block_q4_1 * xb, short il, thread type4 & reg) { + device const uint16_t * qs = ((device const uint16_t *)xb + 2); + const float d1 = (il/4) ? (xb->d / 16.h) : xb->d; + const float d2 = d1 / 256.f; + const float m = xb->m; + const ushort mask0 = (il/4) ? 0x00F0 : 0x000F; + const ushort mask1 = mask0 << 8; + + for (int i = 0; i < 2; i++) { + reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + m; + reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + m; + } +} + template -void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg) { +void dequantize_q5_0(device const block_q5_0 * xb, short il, thread type4x4 & reg) { device const uint16_t * qs = ((device const uint16_t *)xb + 3); const float d = xb->d; const float md = -16.h * xb->d; @@ -124,8 +159,38 @@ void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg reg = (type4x4) reg_f; } +template +void dequantize_q5_0_t4(device const block_q5_0 * xb, short il, thread type4 & reg) { + device const uint16_t * qs = ((device const uint16_t *)xb + 3); + const float d = xb->d; + const float md = -16.h * xb->d; + const ushort mask = (il/4) ? 0x00F0 : 0x000F; + + const uint32_t qh = *((device const uint32_t *)xb->qh); + + const int x_mv = (il/4) ? 4 : 0; + + const int gh_mv = (il/4) ? 12 : 0; + const int gh_bk = (il/4) ? 0 : 4; + + for (int ii = 0; ii < 2; ii++) { + int i = 2*(il%4) + ii; + + // extract the 5-th bits for x0 and x1 + const uint8_t xh_0 = ((qh >> (gh_mv + 2*i )) << gh_bk) & 0x10; + const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10; + + // combine the 4-bits from qs with the 5th bit + const int32_t x0 = ((((qs[i] ) & mask) >> x_mv) | xh_0); + const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1); + + reg[2*ii + 0] = d * x0 + md; + reg[2*ii + 1] = d * x1 + md; + } +} + template -void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg) { +void dequantize_q5_1(device const block_q5_1 * xb, short il, thread type4x4 & reg) { device const uint16_t * qs = ((device const uint16_t *)xb + 4); const float d = xb->d; const float m = xb->m; @@ -156,10 +221,40 @@ void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg reg = (type4x4) reg_f; } +template +void dequantize_q5_1_t4(device const block_q5_1 * xb, short il, thread type4 & reg) { + device const uint16_t * qs = ((device const uint16_t *)xb + 4); + const float d = xb->d; + const float m = xb->m; + const ushort mask = (il/4) ? 0x00F0 : 0x000F; + + const uint32_t qh = *((device const uint32_t *)xb->qh); + + const int x_mv = (il/4) ? 4 : 0; + + const int gh_mv = (il/4) ? 12 : 0; + const int gh_bk = (il/4) ? 0 : 4; + + for (int ii = 0; ii < 2; ii++) { + int i = 2*(il%4) + ii; + + // extract the 5-th bits for x0 and x1 + const uint8_t xh_0 = ((qh >> (gh_mv + 2*i )) << gh_bk) & 0x10; + const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10; + + // combine the 4-bits from qs with the 5th bit + const int32_t x0 = ((((qs[i] ) & mask) >> x_mv) | xh_0); + const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1); + + reg[2*ii + 0] = d * x0 + m; + reg[2*ii + 1] = d * x1 + m; + } +} + template void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) { device const int8_t * qs = ((device const int8_t *)xb->qs); - const half d = xb->d; + const float d = xb->d; float4x4 reg_f; @@ -170,6 +265,16 @@ void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg reg = (type4x4) reg_f; } +template +void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & reg) { + device const int8_t * qs = ((device const int8_t *)xb->qs); + const float d = xb->d; + + for (int i = 0; i < 4; i++) { + reg[i] = (qs[4*(il%4) + i + 16*(il/4)] * d); + } +} + template void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) { const float d = xb->d; @@ -224,7 +329,7 @@ static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q } template -void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) { +void dequantize_q4_K(device const block_q4_K * xb, short il, thread type4x4 & reg) { device const uchar * q = xb->qs; short is = (il/4) * 2; @@ -236,7 +341,7 @@ void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg const float dl = d * sc[0]; const float ml = min * sc[1]; - const ushort mask = il<2 ? 0x0F : 0xF0; + const ushort mask = il < 2 ? 0x0F : 0xF0; for (int i = 0; i < 16; ++i) { reg[i/4][i%4] = dl * (q[i] & mask) - ml; } @@ -469,6 +574,19 @@ void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 } } +template +void dequantize_iq4_nl_t4(device const block_iq4_nl * xb, short il, thread type4 & reg) { + device const uint16_t * q4 = (device const uint16_t *)xb->qs; + const float d = xb->d; + uint32_t aux32; + thread const uint8_t * q8 = (thread const uint8_t *)&aux32; + aux32 = ((q4[2*(il%4)] | (q4[2*(il%4)+1] << 16)) >> 4*(il/4)) & 0x0f0f0f0f; + reg[0] = d * kvalues_iq4nl_f[q8[0]]; + reg[1] = d * kvalues_iq4nl_f[q8[1]]; + reg[2] = d * kvalues_iq4nl_f[q8[2]]; + reg[3] = d * kvalues_iq4nl_f[q8[3]]; +} + template void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) { // il is 0...15 for QK_K = 256 => index of block of 32 is il/2 @@ -1248,6 +1366,63 @@ kernel void kernel_ssm_scan_f32( } } +kernel void kernel_argmax( + device const void * x, + device int32_t * dst, + constant int64_t & ncols, + constant uint64_t & nb01, + threadgroup float * shared_maxval [[threadgroup(0)]], + threadgroup int32_t * shared_argmax [[threadgroup(1)]], + uint tgpig[[threadgroup_position_in_grid]], + uint tpitg[[thread_position_in_threadgroup]], + uint sgitg[[simdgroup_index_in_threadgroup]], + uint tiisg[[thread_index_in_simdgroup]], + uint ntg[[threads_per_threadgroup]]) { + device const float * x_row = (device const float *) ((device const char *) x + tgpig * nb01); + + float lmax = -INFINITY; + int32_t larg = -1; + + for (int i00 = tpitg; i00 < ncols; i00 += ntg) { + if (x_row[i00] > lmax) { + lmax = x_row[i00]; + larg = i00; + } + } + + // find the argmax value in the block + float max_val = simd_max(lmax); + int32_t arg_val = simd_max(select(-1, larg, lmax == max_val)); + + if (ntg > N_SIMDWIDTH) { + if (sgitg == 0) { + shared_maxval[tiisg] = -INFINITY; + shared_argmax[tiisg] = -1; + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + + if (tiisg == 0) { + shared_maxval[sgitg] = max_val; + shared_argmax[sgitg] = arg_val; + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + + max_val = shared_maxval[tiisg]; + arg_val = shared_argmax[tiisg]; + + float max_val_reduced = simd_max(max_val); + int32_t arg_val_reduced = simd_max(select(-1, arg_val, max_val == max_val_reduced)); + + dst[tgpig] = arg_val_reduced; + + return; + } + + dst[tgpig] = arg_val; +} + kernel void kernel_norm( constant ggml_metal_kargs_norm & args, device const char * src0, @@ -1752,6 +1927,301 @@ kernel void kernel_mul_mv_q8_0_f32( kernel_mul_mv_q8_0_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); } +// mat-vec kernel processing in chunks of float4 +// chpb - chunks per quantization block +template +void kernel_mul_mv_ext_q4_f32_impl( + constant ggml_metal_kargs_mul_mv_ext & args, + device const char * src0, + device const char * src1, + device char * dst, + uint3 tgpig[[threadgroup_position_in_grid]], + ushort tiisg[[thread_index_in_simdgroup]], + ushort sgitg[[simdgroup_index_in_threadgroup]]) { + const short chpt = 4; // chunks per thread + + //const short nxpsg = (32); + const short nypsg = (32/nxpsg); + + const short tx = tiisg%nxpsg; + const short ty = tiisg/nxpsg; + + const int i01 = tgpig.x*(nypsg*args.nsg) + nypsg*sgitg + ty; + const int i11 = tgpig.y*r1ptg; + const int i1m = tgpig.z; + + const int i12 = i1m%args.ne12; + const int i13 = i1m/args.ne12; + + const uint64_t offset0 = i01*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03; + const uint64_t offset1 = i11*args.nb11 + (i12 )*args.nb12 + (i13 )*args.nb13; + + device const q_t * xq = (i01 < args.ne01) ? (device const q_t *) (src0 + offset0) + tx/chpb : (device const q_t *) src0; + + device const float4 * y4[r1ptg]; + + for (int ir1 = 0; ir1 < r1ptg; ++ir1) { + y4[ir1] = (i11 + ir1 < args.ne11) ? (device const float4 *) (src1 + offset1 + ir1*args.nb11) + tx : (device const float4 *) src1; + } + + float sumf[r1ptg] = { [ 0 ... r1ptg - 1 ] = 0.0f }; + + short cch = tx%chpb; // current chunk index + + for (int ich = tx; 4*ich < args.ne00; ich += chpt*nxpsg) { + float4 lx[chpt]; + +#pragma unroll(chpt) + for (short ch = 0; ch < chpt; ++ch) { + deq_t4(xq, cch, lx[ch]); + + cch += nxpsg; + if (cch >= chpb) { + xq += cch/chpb; + cch %= chpb; + } + } + +#pragma unroll(chpt) + for (short ch = 0; ch < chpt; ++ch) { +#pragma unroll(r1ptg) + for (short ir1 = 0; ir1 < r1ptg; ++ir1) { + sumf[ir1] += dot(lx[ch], y4[ir1][ch*nxpsg]); + + } + } + +#pragma unroll(r1ptg) + for (short ir1 = 0; ir1 < r1ptg; ++ir1) { + y4[ir1] += chpt*nxpsg; + } + } + + // reduce only the threads in each row + for (short ir1 = 0; ir1 < r1ptg; ++ir1) { + if (nxpsg >= 32) { + sumf[ir1] += simd_shuffle_down(sumf[ir1], 16); + } + if (nxpsg >= 16) { + sumf[ir1] += simd_shuffle_down(sumf[ir1], 8); + } + if (nxpsg >= 8) { + sumf[ir1] += simd_shuffle_down(sumf[ir1], 4); + } + if (nxpsg >= 4) { + sumf[ir1] += simd_shuffle_down(sumf[ir1], 2); + } + if (nxpsg >= 2) { + sumf[ir1] += simd_shuffle_down(sumf[ir1], 1); + } + + //sumf[ir1] = simd_sum(sumf[ir1]); + } + + if (tx == 0) { + for (short ir1 = 0; ir1 < r1ptg && i11 + ir1 < args.ne11; ++ir1) { + device float * dst_f32 = (device float *) dst + (uint64_t)i1m*args.ne0*args.ne1 + (uint64_t)(i11 + ir1)*args.ne0; + + if (i01 < args.ne01) { + dst_f32[i01] = sumf[ir1]; + } + } + } +} + +// mat-vec kernel processing in chunks of float4x4 +template +void kernel_mul_mv_ext_q4x4_f32_impl( + constant ggml_metal_kargs_mul_mv_ext & args, + device const char * src0, + device const char * src1, + device char * dst, + uint3 tgpig[[threadgroup_position_in_grid]], + ushort tiisg[[thread_index_in_simdgroup]], + ushort sgitg[[simdgroup_index_in_threadgroup]]) { + const short chpt = 1; + + //const short nxpsg = (32); + const short nypsg = (32/nxpsg); + + const short tx = tiisg%nxpsg; + const short ty = tiisg/nxpsg; + + const int i01 = tgpig.x*(nypsg*args.nsg) + nypsg*sgitg + ty; + const int i11 = tgpig.y*r1ptg; + const int i1m = tgpig.z; + + const int i12 = i1m%args.ne12; + const int i13 = i1m/args.ne12; + + const uint64_t offset0 = i01*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03; + const uint64_t offset1 = i11*args.nb11 + (i12 )*args.nb12 + (i13 )*args.nb13; + + device const q_t * xq = (i01 < args.ne01) ? (device const q_t *) (src0 + offset0) + tx/chpb : (device const q_t *) src0; + + device const float4x4 * y4x4[r1ptg]; + + for (int ir1 = 0; ir1 < r1ptg; ++ir1) { + y4x4[ir1] = (i11 + ir1 < args.ne11) ? (device const float4x4 *) (src1 + offset1 + ir1*args.nb11) + tx : (device const float4x4 *) src1; + } + + float sumf[r1ptg] = { [ 0 ... r1ptg - 1 ] = 0.0f }; + + short cch = tx%chpb; + + for (int ich = tx; 16*ich < args.ne00; ich += chpt*nxpsg) { + float4x4 lx[chpt]; + +#pragma unroll(chpt) + for (short ch = 0; ch < chpt; ++ch) { + deq_t4x4(xq, cch, lx[ch]); + + cch += nxpsg; + if (cch >= chpb) { + xq += cch/chpb; + cch %= chpb; + } + } + +#pragma unroll(chpt) + for (short ch = 0; ch < chpt; ++ch) { +#pragma unroll(r1ptg) + for (short ir1 = 0; ir1 < r1ptg; ++ir1) { + sumf[ir1] += + dot(lx[ch][0], y4x4[ir1][ch*nxpsg][0]) + + dot(lx[ch][1], y4x4[ir1][ch*nxpsg][1]) + + dot(lx[ch][2], y4x4[ir1][ch*nxpsg][2]) + + dot(lx[ch][3], y4x4[ir1][ch*nxpsg][3]); + + } + } + +#pragma unroll(r1ptg) + for (short ir1 = 0; ir1 < r1ptg; ++ir1) { + y4x4[ir1] += chpt*nxpsg; + } + } + + for (short ir1 = 0; ir1 < r1ptg; ++ir1) { + if (nxpsg >= 32) { + sumf[ir1] += simd_shuffle_down(sumf[ir1], 16); + } + if (nxpsg >= 16) { + sumf[ir1] += simd_shuffle_down(sumf[ir1], 8); + } + if (nxpsg >= 8) { + sumf[ir1] += simd_shuffle_down(sumf[ir1], 4); + } + if (nxpsg >= 4) { + sumf[ir1] += simd_shuffle_down(sumf[ir1], 2); + } + if (nxpsg >= 2) { + sumf[ir1] += simd_shuffle_down(sumf[ir1], 1); + } + + //sumf[ir1] = simd_sum(sumf[ir1]); + } + + if (tx == 0) { + for (short ir1 = 0; ir1 < r1ptg && i11 + ir1 < args.ne11; ++ir1) { + device float * dst_f32 = (device float *) dst + (uint64_t)i1m*args.ne0*args.ne1 + (uint64_t)(i11 + ir1)*args.ne0; + + if (i01 < args.ne01) { + dst_f32[i01] = sumf[ir1]; + } + } + } +} + +// dispatchers needed for compile-time nxpsg +// epb - elements per quantization block +template +kernel void kernel_mul_mv_ext_q4_f32_disp( + constant ggml_metal_kargs_mul_mv_ext & args, + device const char * src0, + device const char * src1, + device char * dst, + uint3 tgpig[[threadgroup_position_in_grid]], + ushort tiisg[[thread_index_in_simdgroup]], + ushort sgitg[[simdgroup_index_in_threadgroup]]) { + switch (args.nxpsg) { + case 4: kernel_mul_mv_ext_q4_f32_impl<4, r1ptg, q_t, epb/4, deq_t4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break; + case 8: kernel_mul_mv_ext_q4_f32_impl<8, r1ptg, q_t, epb/4, deq_t4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break; + case 16: kernel_mul_mv_ext_q4_f32_impl<16, r1ptg, q_t, epb/4, deq_t4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break; + case 32: kernel_mul_mv_ext_q4_f32_impl<32, r1ptg, q_t, epb/4, deq_t4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break; + } +} + +template +kernel void kernel_mul_mv_ext_q4x4_f32_disp( + constant ggml_metal_kargs_mul_mv_ext & args, + device const char * src0, + device const char * src1, + device char * dst, + uint3 tgpig[[threadgroup_position_in_grid]], + ushort tiisg[[thread_index_in_simdgroup]], + ushort sgitg[[simdgroup_index_in_threadgroup]]) { + switch (args.nxpsg) { + case 4: kernel_mul_mv_ext_q4x4_f32_impl<4, r1ptg, q_t, epb/16, deq_t4x4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break; + case 8: kernel_mul_mv_ext_q4x4_f32_impl<8, r1ptg, q_t, epb/16, deq_t4x4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break; + case 16: kernel_mul_mv_ext_q4x4_f32_impl<16, r1ptg, q_t, epb/16, deq_t4x4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break; + case 32: kernel_mul_mv_ext_q4x4_f32_impl<32, r1ptg, q_t, epb/16, deq_t4x4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break; + } +} + +typedef decltype(kernel_mul_mv_ext_q4_f32_disp <2, block_q8_0, 32, dequantize_q8_0_t4>) mul_mv_ext_q4_f32_t; +typedef decltype(kernel_mul_mv_ext_q4x4_f32_disp<2, block_q4_K, 256, dequantize_q4_K>) mul_mv_ext_q4x4_f32_t; + +template [[host_name("kernel_mul_mv_ext_f16_f32_r1_2")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, half4, 4, dequantize_f16_t4>; +template [[host_name("kernel_mul_mv_ext_f16_f32_r1_3")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, half4, 4, dequantize_f16_t4>; +template [[host_name("kernel_mul_mv_ext_f16_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, half4, 4, dequantize_f16_t4>; +template [[host_name("kernel_mul_mv_ext_f16_f32_r1_5")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, half4, 4, dequantize_f16_t4>; + +template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_2")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_q4_0, 32, dequantize_q4_0_t4>; +template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_3")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_q4_0, 32, dequantize_q4_0_t4>; +template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q4_0, 32, dequantize_q4_0_t4>; +template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_5")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_q4_0, 32, dequantize_q4_0_t4>; + +template [[host_name("kernel_mul_mv_ext_q4_1_f32_r1_2")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_q4_1, 32, dequantize_q4_1_t4>; +template [[host_name("kernel_mul_mv_ext_q4_1_f32_r1_3")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_q4_1, 32, dequantize_q4_1_t4>; +template [[host_name("kernel_mul_mv_ext_q4_1_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q4_1, 32, dequantize_q4_1_t4>; +template [[host_name("kernel_mul_mv_ext_q4_1_f32_r1_5")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_q4_1, 32, dequantize_q4_1_t4>; + +template [[host_name("kernel_mul_mv_ext_q5_0_f32_r1_2")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_q5_0, 32, dequantize_q5_0_t4>; +template [[host_name("kernel_mul_mv_ext_q5_0_f32_r1_3")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_q5_0, 32, dequantize_q5_0_t4>; +template [[host_name("kernel_mul_mv_ext_q5_0_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q5_0, 32, dequantize_q5_0_t4>; +template [[host_name("kernel_mul_mv_ext_q5_0_f32_r1_5")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_q5_0, 32, dequantize_q5_0_t4>; + +template [[host_name("kernel_mul_mv_ext_q5_1_f32_r1_2")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_q5_1, 32, dequantize_q5_1_t4>; +template [[host_name("kernel_mul_mv_ext_q5_1_f32_r1_3")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_q5_1, 32, dequantize_q5_1_t4>; +template [[host_name("kernel_mul_mv_ext_q5_1_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q5_1, 32, dequantize_q5_1_t4>; +template [[host_name("kernel_mul_mv_ext_q5_1_f32_r1_5")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_q5_1, 32, dequantize_q5_1_t4>; + +template [[host_name("kernel_mul_mv_ext_q8_0_f32_r1_2")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_q8_0, 32, dequantize_q8_0_t4>; +template [[host_name("kernel_mul_mv_ext_q8_0_f32_r1_3")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_q8_0, 32, dequantize_q8_0_t4>; +template [[host_name("kernel_mul_mv_ext_q8_0_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q8_0, 32, dequantize_q8_0_t4>; +template [[host_name("kernel_mul_mv_ext_q8_0_f32_r1_5")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_q8_0, 32, dequantize_q8_0_t4>; + +template [[host_name("kernel_mul_mv_ext_iq4_nl_f32_r1_2")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_iq4_nl, 32, dequantize_iq4_nl_t4>; +template [[host_name("kernel_mul_mv_ext_iq4_nl_f32_r1_3")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_iq4_nl, 32, dequantize_iq4_nl_t4>; +template [[host_name("kernel_mul_mv_ext_iq4_nl_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_iq4_nl, 32, dequantize_iq4_nl_t4>; +template [[host_name("kernel_mul_mv_ext_iq4_nl_f32_r1_5")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_iq4_nl, 32, dequantize_iq4_nl_t4>; + +template [[host_name("kernel_mul_mv_ext_q4_K_f32_r1_2")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<2, block_q4_K, 256, dequantize_q4_K>; +template [[host_name("kernel_mul_mv_ext_q4_K_f32_r1_3")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<3, block_q4_K, 256, dequantize_q4_K>; +template [[host_name("kernel_mul_mv_ext_q4_K_f32_r1_4")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<4, block_q4_K, 256, dequantize_q4_K>; +template [[host_name("kernel_mul_mv_ext_q4_K_f32_r1_5")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<5, block_q4_K, 256, dequantize_q4_K>; + +template [[host_name("kernel_mul_mv_ext_q5_K_f32_r1_2")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<2, block_q5_K, 256, dequantize_q5_K>; +template [[host_name("kernel_mul_mv_ext_q5_K_f32_r1_3")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<3, block_q5_K, 256, dequantize_q5_K>; +template [[host_name("kernel_mul_mv_ext_q5_K_f32_r1_4")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<4, block_q5_K, 256, dequantize_q5_K>; +template [[host_name("kernel_mul_mv_ext_q5_K_f32_r1_5")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<5, block_q5_K, 256, dequantize_q5_K>; + +template [[host_name("kernel_mul_mv_ext_q6_K_f32_r1_2")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<2, block_q6_K, 256, dequantize_q6_K>; +template [[host_name("kernel_mul_mv_ext_q6_K_f32_r1_3")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<3, block_q6_K, 256, dequantize_q6_K>; +template [[host_name("kernel_mul_mv_ext_q6_K_f32_r1_4")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<4, block_q6_K, 256, dequantize_q6_K>; +template [[host_name("kernel_mul_mv_ext_q6_K_f32_r1_5")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<5, block_q6_K, 256, dequantize_q6_K>; + #define N_MV_T_T 4 template @@ -2258,6 +2728,79 @@ kernel void kernel_im2col_ext( template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext; template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext; +typedef void (conv_transpose_1d_t)( + device const float * src0, + device const float * src1, + device char * dst, + constant int32_t & IC, + constant int32_t & IL, + constant int32_t & K, + constant int32_t & s0, + constant uint64_t & nb0, + constant uint64_t & nb1, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tgpg[[threadgroups_per_grid]]); + +template +kernel void kernel_conv_transpose_1d( + device const T * src0, + device const float * src1, + device char * dst, + constant int32_t & IC, + constant int32_t & IL, + constant int32_t & K, + constant int32_t & s0, + constant uint64_t & nb0, + constant uint64_t & nb1, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tgpg[[threadgroups_per_grid]]) { + + float v = 0.0f; + + for (int64_t c = 0; c < IC; c++) { + const int32_t kernel_offset = c * tgpg[1] * K + K * tgpig[1]; + const int32_t input_offset = c * IL; + + for (int64_t i = 0; i < IL; i++) { + if (tgpig[0] >= i * s0 && tgpig[0] < i * s0 + K) { + v += src0[kernel_offset + tgpig[0] - i * s0] * src1[input_offset + i]; + } + } + } + + device float * dst_ptr = (device float *) (dst + tgpig[0] * nb0 + tgpig[1] * nb1); + + dst_ptr[0] = v; +} + +template [[host_name("kernel_conv_transpose_1d_f32_f32")]] +kernel void kernel_conv_transpose_1d( + device const float * src0, + device const float * src1, + device char * dst, + constant int32_t & IC, + constant int32_t & IL, + constant int32_t & K, + constant int32_t & s0, + constant uint64_t & nb0, + constant uint64_t & nb1, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tgpg[[threadgroups_per_grid]]); + +template [[host_name("kernel_conv_transpose_1d_f16_f32")]] +kernel void kernel_conv_transpose_1d( + device const half * src0, + device const float * src1, + device char * dst, + constant int32_t & IC, + constant int32_t & IL, + constant int32_t & K, + constant int32_t & s0, + constant uint64_t & nb0, + constant uint64_t & nb1, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tgpg[[threadgroups_per_grid]]); + kernel void kernel_upscale_f32( device const char * src0, device char * dst, @@ -2354,6 +2897,53 @@ kernel void kernel_pad_f32( } } +kernel void kernel_pad_reflect_1d_f32( + device const char * src0, + device char * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant int64_t & ne03, + constant int64_t & ne0, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant uint64_t & nb03, + constant uint64_t & nb0, + constant uint64_t & nb1, + constant uint64_t & nb2, + constant uint64_t & nb3, + constant int32_t & p0, + constant int32_t & p1, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tgpg[[threadgroups_per_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + + const int64_t i3 = tgpig.z; + const int64_t i2 = tgpig.y; + const int64_t i1 = tgpig.x; + + const int64_t i03 = i3; + const int64_t i02 = i2; + const int64_t i01 = i1; + + device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01); + device float * dst_ptr = (device float *) (dst + i3*nb3 + i2*nb2 + i1*nb1); + + if (i1 < ne01 && i2 < ne02 && i3 < ne03) { + for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { + if (i0 < p0) { + dst_ptr[i0] = src0_ptr[p0 - i0]; + } else if (i0 < ne0 - p1) { + dst_ptr[i0] = src0_ptr[i0 - p0]; + } else { + dst_ptr[i0] = src0_ptr[(ne0 - p1 - p0) - (p1 + 1 - (ne0 - i0)) - 1]; + } + } + } +} + kernel void kernel_arange_f32( device char * dst, constant int64_t & ne0, @@ -3337,6 +3927,38 @@ template [[host_name("kernel_flash_attn_ext_vec_q8_0_h256")]] kernel flash_attn_ #undef FA_TYPES +template +kernel void kernel_set( + constant ggml_metal_kargs_set & args, + device const char * src0, + device const char * src1, + device char * dst, + uint3 tgpig[[threadgroup_position_in_grid]], + ushort3 tpitg[[thread_position_in_threadgroup]], + ushort3 ntg[[threads_per_threadgroup]]) { + const int i13 = tgpig[2]; + const int i12 = tgpig[1]; + const int i11 = tgpig[0]; + + const int64_t n = i13*args.ne12*args.ne11*args.ne10 + i12*args.ne11*args.ne10 + i11*args.ne10; + + const int64_t i3 = n / (args.ne12*args.ne11*args.ne10); + const int64_t i2 = (n - i3*args.ne12*args.ne11*args.ne10) / (args.ne11*args.ne10); + const int64_t i1 = (n - i3*args.ne12*args.ne11*args.ne10 - i2*args.ne11*args.ne10) / args.ne10; + + device T * dst_data = (device T *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + args.offs); + + for (int64_t i10 = tpitg.x; i10 < args.ne10; i10 += ntg.x) { + device const T * src = (device T *) (src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11 + i10*args.nb10); + dst_data[i10] = (T) src[0]; + } +} + +typedef decltype(kernel_set) kernel_set_t; + +template [[host_name("kernel_set_f32")]] kernel kernel_set_t kernel_set; +template [[host_name("kernel_set_i32")]] kernel kernel_set_t kernel_set; + template kernel void kernel_cpy( constant ggml_metal_kargs_cpy & args, @@ -5447,12 +6069,12 @@ kernel void kernel_mul_mm( const int im = tgpig.z; // if this block is of 64x32 shape or smaller - short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M; - short n_cols = (args.ne1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (args.ne1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N; + const short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M; + const short n_cols = (args.ne1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (args.ne1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N; // a thread shouldn't load data outside of the matrix - short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1; - short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1; + const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1; + const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1; simdgroup_T8x8 ma[4]; simdgroup_float8x8 mb[2]; @@ -5467,20 +6089,23 @@ kernel void kernel_mul_mm( const int i12 = im%args.ne12; const int i13 = im/args.ne12; - uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03; - short offset1 = il/nl; + const uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03; + const short offset1 = il/nl; + + device const block_q * x = (device const block_q *)(src0 + + args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1; - device const block_q * x = (device const block_q *)(src0 + (r0*BLOCK_SIZE_M + thread_row)*args.nb01 + offset0) + offset1; device const float * y = (device const float *)(src1 + args.nb13*i13 + args.nb12*i12 - + args.nb11*(r1 * BLOCK_SIZE_N + thread_col) + + args.nb11*(r1*BLOCK_SIZE_N + thread_col) + args.nb10*(BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL))); for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) { // load data and store to threadgroup memory T4x4 temp_a; dequantize_func(x, il, temp_a); + threadgroup_barrier(mem_flags::mem_threadgroup); #pragma unroll(16) @@ -5490,44 +6115,46 @@ kernel void kernel_mul_mm( + (tiitg/THREAD_PER_ROW)%8 + (i&7)*8) = temp_a[i/4][i%4]; } - *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL)*8*32 + 8*(tiitg/THREAD_PER_COL)) = *((device float2x4 *) y); + *(threadgroup float2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = *((device float2x4 *) y); il = (il + 2 < nl) ? il + 2 : il % 2; - x = (il < 2) ? x + (2+nl-1)/nl : x; + x = (il < 2) ? x + (2 + nl - 1)/nl : x; y += BLOCK_SIZE_K; threadgroup_barrier(mem_flags::mem_threadgroup); // load matrices from threadgroup memory and conduct outer products - threadgroup T * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2)); - threadgroup float * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2)); + threadgroup const T * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2)); + threadgroup const float * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2)); #pragma unroll(4) - for (short ik = 0; ik < BLOCK_SIZE_K / 8; ik++) { + for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) { #pragma unroll(4) for (short i = 0; i < 4; i++) { simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i); } + simdgroup_barrier(mem_flags::mem_none); + #pragma unroll(2) for (short i = 0; i < 2; i++) { simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i); } - lsma += BLOCK_SIZE_M/SG_MAT_ROW * SG_MAT_SIZE; - lsmb += BLOCK_SIZE_N/SG_MAT_ROW * SG_MAT_SIZE; - #pragma unroll(8) for (short i = 0; i < 8; i++){ simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]); } + + lsma += (BLOCK_SIZE_M/SG_MAT_ROW)*SG_MAT_SIZE; + lsmb += (BLOCK_SIZE_N/SG_MAT_ROW)*SG_MAT_SIZE; } } if ((r0 + 1) * BLOCK_SIZE_M <= args.ne0 && (r1 + 1) * BLOCK_SIZE_N <= args.ne1) { device float * C = (device float *) dst + - (BLOCK_SIZE_M * r0 + 32 * (sgitg & 1)) + \ - (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0; + (BLOCK_SIZE_M * r0 + 32*(sgitg & 1)) + \ + (BLOCK_SIZE_N * r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0; for (short i = 0; i < 8; i++) { simdgroup_store(mc[i], C + 8 * (i%4) + 8 * args.ne0 * (i/4), args.ne0); @@ -5536,7 +6163,7 @@ kernel void kernel_mul_mm( // block is smaller than 64x32, we should avoid writing data outside of the matrix threadgroup_barrier(mem_flags::mem_threadgroup); threadgroup float * temp_str = ((threadgroup float *) shmem) \ - + 32 * (sgitg&1) + (16 * (sgitg>>1))*BLOCK_SIZE_M; + + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M; for (short i = 0; i < 8; i++) { simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M); } diff --git a/ggml/src/ggml-musa/ggml/CMakeLists.txt b/ggml/src/ggml-musa/ggml/CMakeLists.txt index f3c01369205..415b2b2e09c 100644 --- a/ggml/src/ggml-musa/ggml/CMakeLists.txt +++ b/ggml/src/ggml-musa/ggml/CMakeLists.txt @@ -20,6 +20,11 @@ find_package(MUSAToolkit) if (MUSAToolkit_FOUND) message(STATUS "MUSA Toolkit found") + if (NOT DEFINED MUSA_ARCHITECTURES) + set(MUSA_ARCHITECTURES "21;22") + endif() + message(STATUS "Using MUSA architectures: ${MUSA_ARCHITECTURES}") + file(GLOB GGML_HEADERS_MUSA "../ggml-cuda/*.cuh") list(APPEND GGML_HEADERS_MUSA "../../include/ggml-cuda.h") @@ -44,15 +49,17 @@ if (MUSAToolkit_FOUND) set_source_files_properties(${GGML_SOURCES_MUSA} PROPERTIES LANGUAGE CXX) foreach(SOURCE ${GGML_SOURCES_MUSA}) - set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS "-x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22") + set(COMPILE_FLAGS "-x musa -mtgpu") + foreach(ARCH ${MUSA_ARCHITECTURES}) + set(COMPILE_FLAGS "${COMPILE_FLAGS} --cuda-gpu-arch=mp_${ARCH}") + endforeach() + set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS ${COMPILE_FLAGS}) endforeach() - add_library(ggml-musa - ${GGML_HEADERS_MUSA} - ${GGML_SOURCES_MUSA}) - - target_link_libraries(ggml-musa PRIVATE ggml-base) - target_include_directories(ggml-musa PRIVATE . ..) + ggml_add_backend_library(ggml-musa + ${GGML_HEADERS_MUSA} + ${GGML_SOURCES_MUSA} + ) # TODO: do not use CUDA definitions for MUSA target_compile_definitions(ggml PUBLIC GGML_USE_CUDA) diff --git a/ggml/src/ggml-opt.cpp b/ggml/src/ggml-opt.cpp index 040205a31ba..7c3e24103a2 100644 --- a/ggml/src/ggml-opt.cpp +++ b/ggml/src/ggml-opt.cpp @@ -14,51 +14,51 @@ #include struct ggml_opt_dataset { - struct ggml_context * ctx; - ggml_backend_buffer_t buf; - struct ggml_tensor * data; - struct ggml_tensor * labels; + struct ggml_context * ctx = nullptr; + ggml_backend_buffer_t buf = nullptr; + struct ggml_tensor * data = nullptr; + struct ggml_tensor * labels = nullptr; - int64_t ndata; - int64_t ndata_shard; - size_t nbs_data; - size_t nbs_labels; + int64_t ndata = -1; + int64_t ndata_shard = -1; + size_t nbs_data = -1; + size_t nbs_labels = -1; std::vector permutation; }; struct ggml_opt_context { - ggml_backend_sched_t backend_sched; - ggml_cgraph * allocated_graph; - ggml_cgraph * allocated_graph_copy; - struct ggml_context * ctx_static; - struct ggml_context * ctx_static_cpu; - struct ggml_context * ctx_compute; - struct ggml_context * ctx_copy; - ggml_backend_buffer_t buf_static; - ggml_backend_buffer_t buf_static_cpu; + ggml_backend_sched_t backend_sched = nullptr; + ggml_cgraph * allocated_graph = nullptr; + ggml_cgraph * allocated_graph_copy = nullptr; + struct ggml_context * ctx_static = nullptr; + struct ggml_context * ctx_static_cpu = nullptr; + struct ggml_context * ctx_compute = nullptr; + struct ggml_context * ctx_copy = nullptr; + ggml_backend_buffer_t buf_static = nullptr; + ggml_backend_buffer_t buf_static_cpu = nullptr; std::mt19937 rng; - struct ggml_tensor * inputs; - struct ggml_tensor * outputs; - struct ggml_tensor * labels; + struct ggml_tensor * inputs = nullptr; + struct ggml_tensor * outputs = nullptr; + struct ggml_tensor * labels = nullptr; - struct ggml_tensor * loss; - struct ggml_tensor * pred; - struct ggml_tensor * ncorrect; + struct ggml_tensor * loss = nullptr; + struct ggml_tensor * pred = nullptr; + struct ggml_tensor * ncorrect = nullptr; - struct ggml_cgraph * gf; - struct ggml_cgraph * gb_grad; - struct ggml_cgraph * gb_opt; + struct ggml_cgraph * gf = nullptr; + struct ggml_cgraph * gb_grad = nullptr; + struct ggml_cgraph * gb_opt = nullptr; - int64_t iter; - int32_t opt_period; - int32_t opt_i; - bool loss_per_datapoint; + int64_t iter = 1; + int32_t opt_period = 1; + int32_t opt_i = 0; + bool loss_per_datapoint = false; - ggml_opt_get_optimizer_params get_opt_pars; - void * get_opt_pars_ud; - struct ggml_tensor * adamw_params; + ggml_opt_get_optimizer_params get_opt_pars = nullptr; + void * get_opt_pars_ud = nullptr; + struct ggml_tensor * adamw_params = nullptr; }; struct ggml_opt_result { @@ -67,8 +67,8 @@ struct ggml_opt_result { std::vector pred; int64_t ncorrect = 0; - bool loss_per_datapoint = false; - int64_t opt_period = -1; + int64_t opt_period = -1; + bool loss_per_datapoint = false; }; // ====== Dataset ====== @@ -188,11 +188,11 @@ struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * us } struct ggml_opt_params ggml_opt_default_params( - ggml_backend_sched_t backend_sched, - struct ggml_context * ctx_compute, - struct ggml_tensor * inputs, - struct ggml_tensor * outputs, - enum ggml_opt_loss_type loss_type) { + ggml_backend_sched_t backend_sched, + struct ggml_context * ctx_compute, + struct ggml_tensor * inputs, + struct ggml_tensor * outputs, + enum ggml_opt_loss_type loss_type) { return { /*backend_sched =*/ backend_sched, /*ctx_compute =*/ ctx_compute, @@ -237,25 +237,33 @@ static ggml_tensor * map_tensor(std::map & tensor_ return new_tensor; } -static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * graph) { +static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * src) { std::map tensor_map; - ggml_cgraph * new_graph = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); + ggml_cgraph * dst = ggml_new_graph_custom(ctx, src->size, /*grads =*/ true); - for (int i = 0; i < graph->n_leafs; i++) { - ggml_build_forward_expand(new_graph, map_tensor(tensor_map, ctx, graph->leafs[i])); + for (int i = 0; i < src->n_leafs; i++) { + ggml_build_forward_expand(dst, map_tensor(tensor_map, ctx, src->leafs[i])); } - for (int i = 0; i < graph->n_nodes; i++) { - ggml_build_forward_expand(new_graph, map_tensor(tensor_map, ctx, graph->nodes[i])); + GGML_ASSERT(dst->n_leafs == src->n_leafs); + for (int i = 0; i < src->n_nodes; i++) { + ggml_build_forward_expand(dst, map_tensor(tensor_map, ctx, src->nodes[i])); } - for (int i = 0; i < graph->n_nodes; ++i) { - const size_t igrad_src = ggml_hash_find(&graph->visited_hash_set, graph->nodes[i]); - const size_t igrad_dst = ggml_hash_find(&new_graph->visited_hash_set, new_graph->nodes[i]); - graph->grads[igrad_dst] = new_graph->grads[igrad_src]; - graph->grad_accs[igrad_dst] = new_graph->grad_accs[igrad_src]; + GGML_ASSERT(dst->n_nodes == src->n_nodes); + for (int i = 0; i < src->n_nodes; ++i) { + const size_t igrad_src = ggml_hash_find(&src->visited_hash_set, src->nodes[i]); + const size_t igrad_dst = ggml_hash_find(&dst->visited_hash_set, dst->nodes[i]); + + GGML_ASSERT(igrad_src != GGML_HASHSET_FULL); + GGML_ASSERT(ggml_bitset_get(src->visited_hash_set.used, igrad_src)); + GGML_ASSERT(igrad_dst != GGML_HASHSET_FULL); + GGML_ASSERT(ggml_bitset_get(dst->visited_hash_set.used, igrad_dst)); + + dst->grads[igrad_dst] = src->grads[igrad_src]; + dst->grad_accs[igrad_dst] = src->grad_accs[igrad_src]; } - return new_graph; + return dst; } static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph) { @@ -284,18 +292,13 @@ static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) { ggml_opt_context_t result = new struct ggml_opt_context; - result->backend_sched = params.backend_sched; - result->allocated_graph = nullptr; - result->allocated_graph_copy = nullptr; - result->ctx_compute = params.ctx_compute; - result->ctx_copy = nullptr; - result->inputs = params.inputs; - result->outputs = params.outputs; - result->iter = 1; - result->opt_period = params.opt_period; - result->opt_i = 0; - result->get_opt_pars = params.get_opt_pars; - result->get_opt_pars_ud = params.get_opt_pars_ud; + result->backend_sched = params.backend_sched; + result->ctx_compute = params.ctx_compute; + result->inputs = params.inputs; + result->outputs = params.outputs; + result->opt_period = params.opt_period; + result->get_opt_pars = params.get_opt_pars; + result->get_opt_pars_ud = params.get_opt_pars_ud; GGML_ASSERT(result->inputs->data && "the inputs must be allocated statically"); GGML_ASSERT(result->opt_period >= 1); @@ -348,7 +351,6 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) { switch (params.loss_type) { case GGML_OPT_LOSS_TYPE_MEAN: { - result->labels = nullptr; result->loss = ggml_sum(result->ctx_static, result->outputs); ggml_set_name(result->loss, "loss_sum"); const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs)); @@ -358,7 +360,6 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) { break; } case GGML_OPT_LOSS_TYPE_SUM: { - result->labels = nullptr; result->loss = ggml_sum(result->ctx_static, result->outputs); ggml_set_name(result->loss, "loss_sum"); result->loss_per_datapoint = false; @@ -413,14 +414,7 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) { } if (params.build_type == GGML_OPT_BUILD_TYPE_FORWARD) { - result->gb_grad = nullptr; - result->gb_opt = nullptr; - result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0)); - result->buf_static_cpu = nullptr; - - ggml_opt_alloc_graph(result, result->gf); - return result; } @@ -429,14 +423,8 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) { ggml_build_backward_expand(result->ctx_static, result->ctx_compute, result->gb_grad, accumulate); if (params.build_type == GGML_OPT_BUILD_TYPE_GRAD) { - result->gb_opt = nullptr; - result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0)); - result->buf_static_cpu = nullptr; - - ggml_opt_alloc_graph(result, result->gb_grad); ggml_graph_reset(result->gb_grad); - return result; } @@ -466,7 +454,6 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) { result->buf_static_cpu = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx_static_cpu, ggml_backend_cpu_buffer_type()); - ggml_opt_alloc_graph(result, result->gb_opt); ggml_graph_reset(result->gb_opt); return result; diff --git a/ggml/src/ggml-rpc/CMakeLists.txt b/ggml/src/ggml-rpc/CMakeLists.txt index a2d6770eb05..f5acb8ec2cb 100644 --- a/ggml/src/ggml-rpc/CMakeLists.txt +++ b/ggml/src/ggml-rpc/CMakeLists.txt @@ -1,10 +1,8 @@ message(STATUS "Using RPC backend") -add_library(ggml-rpc - ggml-rpc.cpp) - -target_link_libraries(ggml-rpc PRIVATE ggml-base) -target_include_directories(ggml-rpc PRIVATE . ..) +ggml_add_backend_library(ggml-rpc + ggml-rpc.cpp + ) if (WIN32) target_link_libraries(ggml-rpc PRIVATE ws2_32) diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp index 47357daabdf..43108242639 100644 --- a/ggml/src/ggml-rpc/ggml-rpc.cpp +++ b/ggml/src/ggml-rpc/ggml-rpc.cpp @@ -1369,8 +1369,9 @@ static const struct ggml_backend_reg_i ggml_backend_rpc_reg_i = { ggml_backend_reg_t ggml_backend_rpc_reg(void) { static struct ggml_backend_reg ggml_backend_rpc_reg = { - /* .iface = */ ggml_backend_rpc_reg_i, - /* .context = */ NULL, + /* .api_version = */ GGML_BACKEND_API_VERSION, + /* .iface = */ ggml_backend_rpc_reg_i, + /* .context = */ NULL, }; return &ggml_backend_rpc_reg; @@ -1401,3 +1402,5 @@ ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint) { return dev; } + +GGML_BACKEND_DL_IMPL(ggml_backend_rpc_reg) diff --git a/ggml/src/ggml-sycl/CMakeLists.txt b/ggml/src/ggml-sycl/CMakeLists.txt index d1d0ff83d63..3579a311aac 100644 --- a/ggml/src/ggml-sycl/CMakeLists.txt +++ b/ggml/src/ggml-sycl/CMakeLists.txt @@ -16,12 +16,10 @@ endif() message(STATUS "SYCL found") #todo: AOT -add_library(ggml-sycl - ggml-sycl.cpp - ../../include/ggml-sycl.h) - -target_link_libraries(ggml-sycl PRIVATE ggml-base) -target_include_directories(ggml-sycl PRIVATE . ..) +ggml_add_backend_library(ggml-sycl + ggml-sycl.cpp + ../../include/ggml-sycl.h + ) if (GGML_SYCL_F16) if (GGML_SYCL_TARGET STREQUAL "AMD") @@ -70,7 +68,8 @@ else() target_link_libraries(ggml-sycl PRIVATE sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread) elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA") set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda") - target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl) + add_compile_definitions(GGML_SYCL_NVIDIA) + target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl_blas_cublas) elseif (GGML_SYCL_TARGET STREQUAL "AMD") if (NOT GGML_SYCL_DEVICE_ARCH) message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.") diff --git a/ggml/src/ggml-sycl/dpct/helper.hpp b/ggml/src/ggml-sycl/dpct/helper.hpp index c2f28bb4957..d1b5dd87c69 100644 --- a/ggml/src/ggml-sycl/dpct/helper.hpp +++ b/ggml/src/ggml-sycl/dpct/helper.hpp @@ -1689,9 +1689,14 @@ namespace dpct auto data_a = get_memory(a); auto data_b = get_memory(b); auto data_c = get_memory(c); - oneapi::mkl::blas::column_major::gemm( - q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda, - data_b, ldb, beta_value, data_c, ldc); +#ifdef GGML_SYCL_NVIDIA + oneapi::mkl::blas::column_major::gemm(oneapi::mkl::backend_selector{ q }, + a_trans, b_trans, m, n, k, alpha_value, data_a, lda, data_b, ldb, + beta_value, data_c, ldc); +#else + oneapi::mkl::blas::column_major::gemm(q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda, data_b, ldb, + beta_value, data_c, ldc); +#endif } template @@ -1754,14 +1759,22 @@ namespace dpct matrix_info->ld_info[2] = ldc; matrix_info->groupsize_info = batch_size; +#ifdef GGML_SYCL_NVIDIA + sycl::event e = oneapi::mkl::blas::column_major::gemm_batch( + oneapi::mkl::backend_selector{ q }, matrix_info->transpose_info, + matrix_info->transpose_info + 1, matrix_info->size_info, matrix_info->size_info + 1, + matrix_info->size_info + 2, matrix_info->value_info, reinterpret_cast(a), + matrix_info->ld_info, reinterpret_cast(b), matrix_info->ld_info + 1, + matrix_info->value_info + 1, reinterpret_cast(c), matrix_info->ld_info + 2, 1, + &(matrix_info->groupsize_info)); +#else sycl::event e = oneapi::mkl::blas::column_major::gemm_batch( - q, matrix_info->transpose_info, matrix_info->transpose_info + 1, - matrix_info->size_info, matrix_info->size_info + 1, - matrix_info->size_info + 2, matrix_info->value_info, - reinterpret_cast(a), matrix_info->ld_info, - reinterpret_cast(b), matrix_info->ld_info + 1, - matrix_info->value_info + 1, reinterpret_cast(c), + q, matrix_info->transpose_info, matrix_info->transpose_info + 1, matrix_info->size_info, + matrix_info->size_info + 1, matrix_info->size_info + 2, matrix_info->value_info, + reinterpret_cast(a), matrix_info->ld_info, reinterpret_cast(b), + matrix_info->ld_info + 1, matrix_info->value_info + 1, reinterpret_cast(c), matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info)); +#endif q.submit([&](sycl::handler &cgh) { @@ -1783,10 +1796,16 @@ namespace dpct auto data_a = get_memory(a); auto data_b = get_memory(b); auto data_c = get_memory(c); +#ifdef GGML_SYCL_NVIDIA oneapi::mkl::blas::column_major::gemm_batch( - q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda, - stride_a, data_b, ldb, stride_b, beta_value, - data_c, ldc, stride_c, batch_size); + oneapi::mkl::backend_selector{ q }, a_trans, b_trans, m, n, k, + alpha_value, data_a, lda, stride_a, data_b, ldb, stride_b, beta_value, data_c, ldc, stride_c, + batch_size); +#else + oneapi::mkl::blas::column_major::gemm_batch(q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda, + stride_a, data_b, ldb, stride_b, beta_value, data_c, ldc, + stride_c, batch_size); +#endif } } // namespace detail diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp index 255bc64c6ba..135efb521a9 100644 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp @@ -47,7 +47,7 @@ static ggml_sycl_device_info ggml_sycl_init() { info.device_count = dpct::dev_mgr::instance().device_count(); if (info.device_count == 0) { - fprintf(stderr, "%s: failed to initialize " GGML_SYCL_NAME ": %s\n", __func__); + GGML_LOG_ERROR("%s: failed to initialize " GGML_SYCL_NAME ": %s\n", __func__); return info; } @@ -55,16 +55,16 @@ static ggml_sycl_device_info ggml_sycl_init() { int64_t total_vram = 0; #if defined(GGML_SYCL_FORCE_MMQ) - fprintf(stderr, "%s: GGML_SYCL_FORCE_MMQ: yes\n", __func__); + GGML_LOG_INFO("%s: GGML_SYCL_FORCE_MMQ: yes\n", __func__); #else - fprintf(stderr, "%s: GGML_SYCL_FORCE_MMQ: no\n", __func__); + GGML_LOG_INFO("%s: GGML_SYCL_FORCE_MMQ: no\n", __func__); #endif #if defined(SYCL_USE_XMX) - fprintf(stderr, "%s: SYCL_USE_XMX: yes\n", __func__); + GGML_LOG_INFO("%s: SYCL_USE_XMX: yes\n", __func__); #else - fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__); + GGML_LOG_INFO("%s: SYCL_USE_XMX: no\n", __func__); #endif - fprintf(stderr, "%s: found %d " GGML_SYCL_NAME " devices:\n", __func__, info.device_count); + GGML_LOG_INFO("%s: found %d " GGML_SYCL_NAME " devices:\n", __func__, info.device_count); for (int i = 0; i < info.device_count; ++i) { info.devices[i].vmm = 0; @@ -110,7 +110,7 @@ void print_device_detail(int id, sycl::device &device, std::string device_type) auto global_mem_size = prop.get_global_mem_size()/1000000; - fprintf(stderr, "|%2d|%19s|%39s|%7s|%7d|%8d|%5d|%6luM|%21s|\n", id, device_type.c_str(), + GGML_LOG_INFO("|%2d|%19s|%39s|%7s|%7d|%8d|%5d|%6luM|%21s|\n", id, device_type.c_str(), name.c_str(), version.c_str(), prop.get_max_compute_units(), prop.get_max_work_group_size(), prop.get_max_sub_group_size(), global_mem_size, device.get_info().c_str()); @@ -120,19 +120,30 @@ void ggml_backend_sycl_print_sycl_devices() { GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_print_sycl_devices\n"); int device_count = dpct::dev_mgr::instance().device_count(); std::map DeviceNums; - fprintf(stderr, "found %d SYCL devices:\n", device_count); - fprintf(stderr, "| | | | |Max | |Max |Global | |\n"); - fprintf(stderr, "| | | | |compute|Max work|sub |mem | |\n"); - fprintf(stderr, "|ID| Device Type| Name|Version|units |group |group|size | Driver version|\n"); - fprintf(stderr, "|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|\n"); + GGML_LOG_INFO("Found %d SYCL devices:\n", device_count); + + GGML_LOG_INFO( + "| | | | " + " |Max | |Max |Global | |\n"); + GGML_LOG_INFO( + "| | | | " + " |compute|Max work|sub |mem | |\n"); + GGML_LOG_INFO( + "|ID| Device Type| " + "Name|Version|units |group |group|size | Driver version|\n"); + GGML_LOG_INFO( + "|--|-------------------|---------------------------------------|------" + "-|-------|--------|-----|-------|---------------------|\n"); + for (int id = 0; id < device_count; ++id) { - sycl::device device = dpct::dev_mgr::instance().get_device(id); - sycl::backend backend = device.get_backend(); - std::string backend_type = get_device_backend_and_type(device); - int type_id=DeviceNums[backend_type]++; - std::stringstream device_type; - device_type << "[" << backend_type << ":" << std::to_string(type_id) << "]"; - print_device_detail(id, device, device_type.str()); + sycl::device device = dpct::dev_mgr::instance().get_device(id); + sycl::backend backend = device.get_backend(); + std::string backend_type = get_device_backend_and_type(device); + int type_id = DeviceNums[backend_type]++; + std::stringstream device_type; + device_type << "[" << backend_type << ":" << std::to_string(type_id) + << "]"; + print_device_detail(id, device, device_type.str()); } } @@ -154,15 +165,14 @@ static void ggml_check_sycl() try { static bool initialized = false; if (!initialized) { - fprintf(stderr, "[SYCL] call ggml_check_sycl\n"); + GGML_LOG_INFO("[SYCL] call ggml_check_sycl\n"); g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0); - - fprintf(stderr, "%s: GGML_SYCL_DEBUG: %d\n", __func__, g_ggml_sycl_debug); + GGML_LOG_INFO("%s: GGML_SYCL_DEBUG: %d\n", __func__, g_ggml_sycl_debug); #if defined(GGML_SYCL_F16) - fprintf(stderr, "%s: GGML_SYCL_F16: yes\n", __func__); + GGML_LOG_INFO("%s: GGML_SYCL_F16: yes\n", __func__); #else - fprintf(stderr, "%s: GGML_SYCL_F16: no\n", __func__); + GGML_LOG_INFO("%s: GGML_SYCL_F16: no\n", __func__); #endif /* NOT REMOVE, keep it for next optimize for XMX. @@ -180,9 +190,10 @@ static void ggml_check_sycl() try { return; } GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES); - ggml_backend_sycl_print_sycl_devices(); + initialized = true; g_sycl_loaded = true; + ggml_backend_sycl_print_sycl_devices(); } } catch (sycl::exception const &exc) { @@ -205,7 +216,7 @@ inline void check_allow_gpu_index(const int device_index) { __func__, device_index, ggml_sycl_info().device_count - 1); - fprintf(stderr, "%s\n", error_buf); + GGML_LOG_ERROR("%s\n", error_buf); assert(false); } } @@ -475,8 +486,8 @@ ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)sycl::malloc_device( size, *stream))); if (!dev_ptr) { - fprintf(stderr, "%s: can't malloc %lu Bytes memory on device", __func__, size); - return nullptr; + GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device\n", __func__, size); + return nullptr; } ggml_backend_sycl_buffer_context * ctx = new ggml_backend_sycl_buffer_context(buft_ctx->device, dev_ptr, buft_ctx->stream); return ggml_backend_buffer_init(buft, ggml_backend_sycl_buffer_interface, ctx, size); @@ -752,7 +763,7 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer, size, *stream))); if (!buf) { char err_buf[1024]; - snprintf(err_buf, 1023, "%s: can't malloc %lu Bytes memory on device", __func__, size); + snprintf(err_buf, 1023, "%s: can't allocate %lu Bytes of memory on device\n", __func__, size); throw std::runtime_error(err_buf); } // set padding to 0 to avoid possible NaN values @@ -1142,17 +1153,18 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool { CHECK_TRY_ERROR(ptr = (void *)sycl::malloc_device( look_ahead_size, *qptr))); if (!ptr) { - fprintf(stderr, "%s: can't malloc %lu Bytes memory on device", __func__, look_ahead_size); + GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device/GPU\n", __func__, look_ahead_size); return nullptr; } *actual_size = look_ahead_size; pool_size += look_ahead_size; - #ifdef DEBUG_SYCL_MALLOC - fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, id, nnz, +#ifdef DEBUG_SYCL_MALLOC + GGML_LOG_DEBUG("%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, id, nnz, (uint32_t)(max_size/1024/1024), (uint32_t)(g_sycl_pool_size[id]/1024/1024), (uint32_t)(size/1024/1024)); - #endif +#endif + // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg look_ahead_size=%lu, return %p\n", look_ahead_size, ptr); return ptr; } @@ -1166,7 +1178,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool { return; } } - fprintf(stderr, "WARNING: sycl buffer pool full, increase MAX_sycl_BUFFERS\n"); + GGML_LOG_WARN("WARNING: sycl buffer pool full, increase MAX_sycl_BUFFERS\n"); SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, *qptr))); pool_size -= size; } @@ -2437,7 +2449,7 @@ static void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, const ggml_te break; default: // TODO: k-quants - fprintf(stderr, "%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type)); + GGML_LOG_ERROR("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type)); GGML_ABORT("fatal error"); break; } @@ -2561,12 +2573,17 @@ inline void ggml_sycl_op_mul_mat_sycl( const float alpha = 1.0f; const float beta = 0.0f; #if !GGML_SYCL_DNNL +# ifdef GGML_SYCL_NVIDIA SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm( - *stream, oneapi::mkl::transpose::trans, - oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10, - dpct::get_value(&alpha, *stream), src0_ddf_i, ne00, - src1_ddf1_i, ne10, dpct::get_value(&beta, *stream), + oneapi::mkl::backend_selector{ *stream }, oneapi::mkl::transpose::trans, + oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10, dpct::get_value(&alpha, *stream), src0_ddf_i, + ne00, src1_ddf1_i, ne10, dpct::get_value(&beta, *stream), dst_dd_i, ldc))); +# else + SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm( + *stream, oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10, + dpct::get_value(&alpha, *stream), src0_ddf_i, ne00, src1_ddf1_i, ne10, dpct::get_value(&beta, *stream), dst_dd_i, ldc))); +# endif #else auto dnnl_stream = ctx.stream_dnnl(stream); DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i, DnnlGemmWrapper::to_dt(), @@ -3447,8 +3464,15 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q; if (!split && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) { - // KQ single-batch - ggml_sycl_mul_mat_vec_p021(ctx, src0, src1, dst); + // TODO: Refactor and cleanup of mul mat dispatching. + if (src0->ne[3] == 1 && src1->ne[3] == 1) { + // KQ single-batch + // mmv p021 was specific for these dimensions + ggml_sycl_mul_mat_vec_p021(ctx, src0, src1, dst); + } else { + // The kernel from the if path is faster for that specific case, but does not support all mul mats. + ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst); + } } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) { // KQV single-batch ggml_sycl_mul_mat_vec_nc(ctx, src0, src1, dst); @@ -3743,7 +3767,7 @@ static void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor *sr } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) { ggml_cpy_i32_i32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); } else { - fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__, + GGML_LOG_ERROR("%s: unsupported type combination (%s to %s)\n", __func__, ggml_type_name(src0->type), ggml_type_name(src1->type)); GGML_ABORT("fatal error"); } @@ -3818,7 +3842,7 @@ void ggml_sycl_set_main_device(const int main_device) try { dpct::device_info prop; SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info( prop, dpct::dev_mgr::instance().get_device(main_device)))); - fprintf(stderr, "Using device %d (%s) as main device\n", + GGML_LOG_INFO("Using device %d (%s) as main device\n", main_device, prop.get_name()); } } @@ -4165,7 +4189,7 @@ static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_ #endif bool ok = ggml_sycl_compute_forward(*sycl_ctx, node); if (!ok) { - fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op)); + GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op)); } GGML_ASSERT(ok); } @@ -4486,7 +4510,7 @@ static bool ggml_backend_sycl_device_supports_buft(ggml_backend_dev_t dev, ggml_ static int64_t get_op_batch_size(const ggml_tensor * op) { switch (op->op) { case GGML_OP_GET_ROWS: - return op->ne[1]; // this will increse the speed of prefill in test + return 0; case GGML_OP_MUL_MAT: return op->ne[1]; case GGML_OP_MUL_MAT_ID: @@ -4637,16 +4661,17 @@ ggml_backend_reg_t ggml_backend_sycl_reg() { dev_ctx->description = prop.get_name(); ggml_backend_dev_t dev = new ggml_backend_device { - /* .interface = */ ggml_backend_sycl_device_interface, - /* .reg = */ ®, - /* .context = */ dev_ctx + /* .iface = */ ggml_backend_sycl_device_interface, + /* .reg = */ ®, + /* .context = */ dev_ctx }; ctx->devices.push_back(dev); } reg = ggml_backend_reg { - /* .interface = */ ggml_backend_sycl_reg_interface, - /* .context = */ ctx + /* .api_version = */ GGML_BACKEND_API_VERSION, + /* .iface = */ ggml_backend_sycl_reg_interface, + /* .context = */ ctx }; } @@ -4664,7 +4689,7 @@ ggml_backend_t ggml_backend_sycl_init(int device) { ggml_backend_sycl_context * ctx = new ggml_backend_sycl_context(device); if (ctx == nullptr) { - fprintf(stderr, "%s: error: failed to allocate context\n", __func__); + GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__); return nullptr; }; @@ -4678,3 +4703,4 @@ ggml_backend_t ggml_backend_sycl_init(int device) { return sycl_backend; } +GGML_BACKEND_DL_IMPL(ggml_backend_sycl_reg) diff --git a/ggml/src/ggml-sycl/outprod.cpp b/ggml/src/ggml-sycl/outprod.cpp index e61cdc2ca5d..ef9af0b7633 100644 --- a/ggml/src/ggml-sycl/outprod.cpp +++ b/ggml/src/ggml-sycl/outprod.cpp @@ -40,14 +40,14 @@ void ggml_sycl_op_out_prod(ggml_backend_sycl_context& ctx, const ggml_tensor* sr try { // Perform matrix multiplication using oneMKL GEMM - oneapi::mkl::blas::column_major::gemm(*stream, - oneapi::mkl::transpose::nontrans, src1_op, - ne0, ne1, ne01, - alpha, - src0_d, ne00, - src1_d, ldb, - beta, - dst_d, ne0); +#ifdef GGML_SYCL_NVIDIA + oneapi::mkl::blas::column_major::gemm(oneapi::mkl::backend_selector{ *stream }, + oneapi::mkl::transpose::nontrans, src1_op, ne0, ne1, ne01, alpha, src0_d, + ne00, src1_d, ldb, beta, dst_d, ne0); +#else + oneapi::mkl::blas::column_major::gemm(*stream, oneapi::mkl::transpose::nontrans, src1_op, ne0, ne1, ne01, alpha, + src0_d, ne00, src1_d, ldb, beta, dst_d, ne0); +#endif } catch (sycl::exception const& exc) { std::cerr << exc.what() << std::endl; diff --git a/ggml/src/ggml-vulkan/CMakeLists.txt b/ggml/src/ggml-vulkan/CMakeLists.txt index 1e85dd15b7a..ae0485e0425 100644 --- a/ggml/src/ggml-vulkan/CMakeLists.txt +++ b/ggml/src/ggml-vulkan/CMakeLists.txt @@ -3,13 +3,13 @@ find_package(Vulkan COMPONENTS glslc REQUIRED) if (Vulkan_FOUND) message(STATUS "Vulkan found") - add_library(ggml-vulkan - ggml-vulkan.cpp - ../../include/ggml-vulkan.h - ) + ggml_add_backend_library(ggml-vulkan + ggml-vulkan.cpp + ../../include/ggml-vulkan.h + ) - target_link_libraries(ggml-vulkan PRIVATE ggml-base Vulkan::Vulkan) - target_include_directories(ggml-vulkan PRIVATE . .. ${CMAKE_CURRENT_BINARY_DIR}) + target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan) + target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR}) # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build # Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 21fee2f3d75..07b45d6b998 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -158,12 +158,14 @@ struct vk_device_struct { std::string name; uint64_t max_memory_allocation_size; bool fp16; + bool pipeline_robustness; vk::Device device; uint32_t vendor_id; vk_queue compute_queue; vk_queue transfer_queue; bool single_queue; uint32_t subgroup_size; + uint32_t shader_core_count; bool uma; size_t idx; @@ -351,7 +353,45 @@ struct vk_op_unary_push_constants { uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13; uint32_t d_offset; float param1; float param2; + uint32_t ne0_012mp; uint32_t ne0_012L; + uint32_t ne0_01mp; uint32_t ne0_01L; + uint32_t ne0_0mp; uint32_t ne0_0L; + uint32_t ne1_012mp; uint32_t ne1_012L; + uint32_t ne1_01mp; uint32_t ne1_01L; + uint32_t ne1_0mp; uint32_t ne1_0L; }; +static_assert(sizeof(vk_op_unary_push_constants) <= 128, "sizeof(vk_op_unary_push_constants) must be <= 128"); + +// See https://gmplib.org/~tege/divcnst-pldi94.pdf figure 4.1. +// Precompute mp (m' in the paper) and L such that division +// can be computed using a multiply (high 32b of 64b result) +// and a shift: +// +// n/d = (mulhi(n, mp) + n) >> L; +void init_fastdiv_values(uint32_t d, uint32_t &mp, uint32_t &L) +{ + // compute L = ceil(log2(d)); + L = 0; + while (L < 32 && (uint32_t{1} << L) < d) { + L++; + } + + mp = (uint32_t)((uint64_t{1} << 32) * ((uint64_t{1} << L) - d) / d + 1); +} + +template void init_pushconst_fastdiv(T &p) { + static_assert(!std::is_const::value, "unexpected type"); +} + +template <> void init_pushconst_fastdiv(vk_op_unary_push_constants &p) { + // Compute magic values to divide by these six numbers. + init_fastdiv_values(p.ne02*p.ne01*p.ne00, p.ne0_012mp, p.ne0_012L); + init_fastdiv_values(p.ne01*p.ne00, p.ne0_01mp, p.ne0_01L); + init_fastdiv_values(p.ne00, p.ne0_0mp, p.ne0_0L); + init_fastdiv_values(p.ne12*p.ne11*p.ne10, p.ne1_012mp, p.ne1_012L); + init_fastdiv_values(p.ne11*p.ne10, p.ne1_01mp, p.ne1_01L); + init_fastdiv_values(p.ne10, p.ne1_0mp, p.ne1_0L); +} struct vk_op_binary_push_constants { uint32_t ne; @@ -654,7 +694,7 @@ static uint32_t compile_count = 0; static std::mutex compile_count_mutex; static std::condition_variable compile_count_cond; -static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, const std::string name, size_t spv_size, const void* spv_data, const std::string entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array wg_denoms, std::vector specialization_constants, uint32_t align) { +static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, const std::string name, size_t spv_size, const void* spv_data, const std::string entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array wg_denoms, std::vector specialization_constants, uint32_t align, bool disable_robustness) { VK_LOG_DEBUG("ggml_vk_create_pipeline(" << device->name << ", " << name << ", " << entrypoint << ", " << parameter_count << ", " << push_constant_size << ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " << align << ")"); GGML_ASSERT(parameter_count > 0); GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT @@ -724,6 +764,15 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin vk::PipelineCreateFlags(), pipeline_shader_create_info, pipeline->layout); + + vk::PipelineRobustnessCreateInfoEXT rci; + + if (device->pipeline_robustness && disable_robustness) { + rci.storageBuffers = vk::PipelineRobustnessBufferBehaviorEXT::eDisabled; + rci.uniformBuffers = vk::PipelineRobustnessBufferBehaviorEXT::eDisabled; + compute_pipeline_create_info.setPNext(&rci); + } + pipeline->pipeline = device->device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value; { @@ -1221,27 +1270,66 @@ static void ggml_vk_load_shaders(vk_device& device) { std::cerr << "ggml_vulkan: Compiling shaders"; + // some shaders require the subgroup size to be 16 or larger + const uint32_t subgroup_size_16 = std::max(device->subgroup_size, 16u); + // mulmat - std::vector l_warptile, m_warptile, s_warptile, l_warptile_mmq, m_warptile_mmq, s_warptile_mmq; - std::array l_wg_denoms, m_wg_denoms, s_wg_denoms; + std::vector l_warptile, m_warptile, s_warptile, + l_warptile_mmq, m_warptile_mmq, s_warptile_mmq; + std::array l_wg_denoms, m_wg_denoms, s_wg_denoms, + l_mmq_wg_denoms, m_mmq_wg_denoms, s_mmq_wg_denoms; uint32_t l_align, m_align, s_align; l_warptile = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size }; m_warptile = { 128, 64, 64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size }; - s_warptile = { std::max(device->subgroup_size, 16u), 32, 32, 16, 32, 32, 2, 2, 2, device->subgroup_size }; + s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, 2, 2, device->subgroup_size }; l_warptile_mmq = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size }; m_warptile_mmq = { 128, 64, 64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size }; - s_warptile_mmq = { std::max(device->subgroup_size, 16u), 32, 32, 32, 32, 32, 2, 2, 2, device->subgroup_size }; + s_warptile_mmq = { subgroup_size_16, 32, 32, 32, 32, 32, 2, 2, 2, device->subgroup_size }; - l_wg_denoms = {128, 128, 1 }; - m_wg_denoms = { 64, 64, 1 }; - s_wg_denoms = { 32, 32, 1 }; + l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 }; + m_mmq_wg_denoms = m_wg_denoms = { 64, 64, 1 }; + s_mmq_wg_denoms = s_wg_denoms = { 32, 32, 1 }; l_align = 128; m_align = 64; s_align = 32; + // Fallback to smaller sizes if there's not enough shared memory. Given the current shaders + // and tile sizes, this should handle 16KB, 32KB, and 48KB+. + // This logic doesn't explicitly account for the 12KB row_ids in the mul_mat_mat_id shaders. + // But the numbers happen to work out for 32KB shared memory size that when using the medium + // size there's enough room for everything, and we assert for this. + uint32_t shmem_needed = (l_warptile[1] + l_warptile[2]) * (l_warptile[3] + 1) * sizeof(float); + if (shmem_needed > device->properties.limits.maxComputeSharedMemorySize) { + l_warptile = m_warptile; + l_wg_denoms = m_wg_denoms; + shmem_needed = (l_warptile[1] + l_warptile[2]) * (l_warptile[3] + 1) * sizeof(float); + GGML_ASSERT(shmem_needed <= device->properties.limits.maxComputeSharedMemorySize); + } + if (device->properties.limits.maxComputeSharedMemorySize >= 32768) { + // assert mul_mat_mat_id shaders will fit. + GGML_ASSERT(shmem_needed + 3072*4 <= device->properties.limits.maxComputeSharedMemorySize); + } + + shmem_needed = (l_warptile_mmq[1] + l_warptile_mmq[2]) * (l_warptile_mmq[3] + 1) * sizeof(float); + if (shmem_needed > device->properties.limits.maxComputeSharedMemorySize) { + if (device->properties.limits.maxComputeSharedMemorySize == 32768) { + l_warptile_mmq = m_warptile_mmq; + l_mmq_wg_denoms = m_mmq_wg_denoms; + } else { + l_warptile_mmq = s_warptile_mmq; + l_mmq_wg_denoms = s_mmq_wg_denoms; + } + shmem_needed = (l_warptile_mmq[1] + l_warptile_mmq[2]) * (l_warptile_mmq[3] + 1) * sizeof(float); + GGML_ASSERT(shmem_needed <= device->properties.limits.maxComputeSharedMemorySize); + } + if (device->properties.limits.maxComputeSharedMemorySize >= 32768) { + // assert mul_mat_mat_id shaders will fit. + GGML_ASSERT(shmem_needed + 3072*4 <= device->properties.limits.maxComputeSharedMemorySize); + } + device->pipeline_matmul_f32 = std::make_shared(); device->pipeline_matmul_f32_f16 = std::make_shared(); @@ -1261,7 +1349,7 @@ static void ggml_vk_load_shaders(vk_device& device) { device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL] = std::make_shared(); std::vector> compiles; - auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array wg_denoms, const std::vector& specialization_constants, uint32_t align) { + auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array wg_denoms, const std::vector& specialization_constants, uint32_t align, bool disable_robustness = false) { { // wait until fewer than N compiles are in progress uint32_t N = std::max(1u, std::thread::hardware_concurrency()); @@ -1271,7 +1359,7 @@ static void ggml_vk_load_shaders(vk_device& device) { } compile_count++; } - compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint, parameter_count, push_constant_size, wg_denoms, specialization_constants, align)); + compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint, parameter_count, push_constant_size, wg_denoms, specialization_constants, align, disable_robustness)); }; if (device->fp16) { @@ -1289,35 +1377,38 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3); CREATE_MM(pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - - CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4); - CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4); - CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4); - - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + + // If there's not enough shared memory for row_ids and the result tile, don't create these pipelines. + if (device->properties.limits.maxComputeSharedMemorySize >= 32768) { + CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4); + CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4); + CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4); + + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + } #undef CREATE_MM } else { // Create 6 variants, {s,m,l}x{unaligned,aligned} @@ -1334,35 +1425,38 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3); CREATE_MM(pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); - - CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4); - CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4); - CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4); - - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3); + + // If there's not enough shared memory for row_ids and the result tile, don't create these pipelines. + if (device->properties.limits.maxComputeSharedMemorySize >= 32768) { + CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4); + CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4); + CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4); + + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4); + } #undef CREATE_MM } @@ -1370,45 +1464,45 @@ static void ggml_vk_load_shaders(vk_device& device) { // computing two rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0. ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32", mul_mat_vec_f32_f32_f32_len, mul_mat_vec_f32_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32", mul_mat_vec_f16_f32_f32_len, mul_mat_vec_f16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32", mul_mat_vec_f32_f16_f32_len, mul_mat_vec_f32_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32", mul_mat_vec_f16_f16_f32_len, mul_mat_vec_f16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32", mul_mat_vec_id_f32_f32_len, mul_mat_vec_id_f32_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32", mul_mat_vec_id_f16_f32_len, mul_mat_vec_id_f16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true); // dequant shaders ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16", dequant_f32_len, dequant_f32_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1); @@ -1443,7 +1537,7 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q8_0], "get_rows_q8_0_f32", get_rows_q8_0_f32_len, get_rows_q8_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ4_NL], "get_rows_iq4_nl_f32", get_rows_iq4_nl_f32_len, get_rows_iq4_nl_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32, "mul_mat_vec_p021_f16_f32", mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1); @@ -1555,11 +1649,14 @@ static vk_device ggml_vk_get_device(size_t idx) { const std::vector ext_props = device->physical_device.enumerateDeviceExtensionProperties(); bool maintenance4_support = false; + bool sm_builtins = false; // Check if maintenance4 is supported for (const auto& properties : ext_props) { if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) { maintenance4_support = true; + } else if (strcmp("VK_NV_shader_sm_builtins", properties.extensionName) == 0) { + sm_builtins = true; } } @@ -1567,11 +1664,21 @@ static vk_device ggml_vk_get_device(size_t idx) { vk::PhysicalDeviceMaintenance3Properties props3; vk::PhysicalDeviceMaintenance4Properties props4; vk::PhysicalDeviceSubgroupProperties subgroup_props; + vk::PhysicalDeviceShaderSMBuiltinsPropertiesNV sm_props; props2.pNext = &props3; props3.pNext = &subgroup_props; + + VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&subgroup_props; + if (maintenance4_support) { - subgroup_props.pNext = &props4; + last_struct->pNext = (VkBaseOutStructure *)&props4; + last_struct = (VkBaseOutStructure *)&props4; + } + if (sm_builtins) { + last_struct->pNext = (VkBaseOutStructure *)&sm_props; + last_struct = (VkBaseOutStructure *)&sm_props; } + device->physical_device.getProperties2(&props2); device->properties = props2.properties; @@ -1588,15 +1695,23 @@ static vk_device ggml_vk_get_device(size_t idx) { device->vendor_id = device->properties.vendorID; device->subgroup_size = subgroup_props.subgroupSize; device->uma = device->properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu; + if (sm_builtins) { + device->shader_core_count = sm_props.shaderSMCount; + } else { + device->shader_core_count = 0; + } bool fp16_storage = false; bool fp16_compute = false; + bool pipeline_robustness = false; for (const auto& properties : ext_props) { if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) { fp16_storage = true; } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) { fp16_compute = true; + } else if (strcmp("VK_EXT_pipeline_robustness", properties.extensionName) == 0) { + pipeline_robustness = true; } } @@ -1642,10 +1757,22 @@ static vk_device ggml_vk_get_device(size_t idx) { vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES; vk11_features.pNext = &vk12_features; + VkPhysicalDevicePipelineRobustnessFeaturesEXT pl_robustness_features; + pl_robustness_features.pNext = nullptr; + pl_robustness_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_ROBUSTNESS_FEATURES_EXT; + pl_robustness_features.pipelineRobustness = VK_FALSE; + + if (pipeline_robustness) { + vk12_features.pNext = &pl_robustness_features; + device_extensions.push_back("VK_EXT_pipeline_robustness"); + } + vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2); device->fp16 = device->fp16 && vk12_features.shaderFloat16; + device->pipeline_robustness = pl_robustness_features.pipelineRobustness; + if (!vk11_features.storageBuffer16BitAccess) { std::cerr << "ggml_vulkan: device " << GGML_VK_NAME << idx << " does not support 16-bit storage." << std::endl; throw std::runtime_error("Unsupported device"); @@ -2662,15 +2789,25 @@ static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, siz dst->device->device.resetFences({ dst->device->fence }); } -static uint32_t ggml_vk_guess_split_k(int m, int n, int k) { +static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, int m, int n, int k, const vk_pipeline& pipeline) { VK_LOG_DEBUG("ggml_vk_guess_split_k(" << m << ", " << n << ", " << k << ")"); - // if (k > 128 && (m < 128 || n < 128) && m > 2 && n > 2) { - // return 4; - // } - return 1; + uint32_t split_k = 1; + if (ctx->device->shader_core_count != 0 && m >= (int)pipeline->wg_denoms[0] && n >= (int)pipeline->wg_denoms[1]) { + // If k is 'large' and the SMs will fill less than halfway, use split_k. + uint32_t m_tiles = CEIL_DIV(m, pipeline->wg_denoms[0]); + uint32_t n_tiles = CEIL_DIV(n, pipeline->wg_denoms[1]); + if (k >= 2048 && m_tiles * n_tiles < ctx->device->shader_core_count / 2) { + split_k = ctx->device->shader_core_count / (m_tiles * n_tiles); + // Clamp to 2 or 4 + split_k = std::min(split_k, 4u); + if (split_k == 3) { + split_k = 2; + } + } + } - GGML_UNUSED(m); GGML_UNUSED(n); GGML_UNUSED(k); + return split_k; } static vk_pipeline ggml_vk_guess_matmul_pipeline_amd(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) { @@ -2815,13 +2952,14 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context& elements = { ne, 1, 1 }; } - const vk_op_unary_push_constants pc = { + vk_op_unary_push_constants pc = { (uint32_t)ne, (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3], (uint32_t)tensor->nb[0] / tensor_type_size, (uint32_t)tensor->nb[1] / tensor_type_size, (uint32_t)tensor->nb[2] / tensor_type_size, (uint32_t)tensor->nb[3] / tensor_type_size, (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3], 1 , (uint32_t)tensor->ne[0] , (uint32_t)(tensor->ne[0] * tensor->ne[1]) , (uint32_t)(tensor->ne[0] * tensor->ne[1] * tensor->ne[2]), 0, 0.0f, 0.0f, }; + init_pushconst_fastdiv(pc); ggml_vk_sync_buffers(subctx); ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, elements); } @@ -2894,10 +3032,10 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11)); const bool aligned = ne10 == kpad && ne01 > 8 && ne11 > 8; - const uint32_t split_k = ggml_vk_guess_split_k(ne01, ne11, ne10); - vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned); + const uint32_t split_k = ggml_vk_guess_split_k(ctx, ne01, ne11, ne10, pipeline); + const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type); const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type); const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne; @@ -2923,7 +3061,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub if (dryrun) { const uint64_t x_sz_upd = x_sz * ne02 * ne03; const uint64_t y_sz_upd = y_sz * ne12 * ne13; - const uint64_t split_k_size = split_k > 1 ? d_sz * ne12 * ne13 * 4 : 0; + const uint64_t split_k_size = split_k > 1 ? d_sz * ne12 * ne13 * split_k : 0; if ( (qx_needs_dequant && x_sz_upd > ctx->device->max_memory_allocation_size) || (qy_needs_dequant && y_sz_upd > ctx->device->max_memory_allocation_size) || @@ -3190,7 +3328,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& if (ne01 > max_groups_x) { groups_z = 64; - groups_x /= groups_z; + groups_x = CEIL_DIV(groups_x, groups_z); } // compute @@ -3767,7 +3905,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte if (ne01 > max_groups_x) { groups_z = 64; - groups_x /= groups_z; + groups_x = CEIL_DIV(groups_x, groups_z); } // compute @@ -4026,7 +4164,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) { } template -static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, const PC&& pc, bool dryrun = false) { +static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, PC&& pc, bool dryrun = false) { VK_LOG_DEBUG("ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3]; if (src1 != nullptr) { std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3]; @@ -4066,6 +4204,8 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co const uint64_t ned3 = dst->ne[3]; const uint64_t ned = ned0 * ned1; + init_pushconst_fastdiv(pc); + vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, dst, op); if (pipeline == nullptr) { @@ -5605,6 +5745,48 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod } else { compute_ctx = ctx->compute_ctx.lock(); } + } else { + switch (node->op) { + case GGML_OP_REPEAT: + case GGML_OP_ACC: + case GGML_OP_GET_ROWS: + case GGML_OP_ADD: + case GGML_OP_MUL: + case GGML_OP_DIV: + case GGML_OP_CONCAT: + case GGML_OP_UPSCALE: + case GGML_OP_SCALE: + case GGML_OP_SQR: + case GGML_OP_SIN: + case GGML_OP_COS: + case GGML_OP_CLAMP: + case GGML_OP_PAD: + case GGML_OP_CPY: + case GGML_OP_CONT: + case GGML_OP_DUP: + case GGML_OP_NORM: + case GGML_OP_GROUP_NORM: + case GGML_OP_RMS_NORM: + case GGML_OP_UNARY: + case GGML_OP_DIAG_MASK_INF: + case GGML_OP_SOFT_MAX: + case GGML_OP_ROPE: + case GGML_OP_ARGSORT: + case GGML_OP_SUM_ROWS: + case GGML_OP_IM2COL: + case GGML_OP_TIMESTEP_EMBEDDING: + case GGML_OP_POOL_2D: + case GGML_OP_LEAKY_RELU: + { + // These operations all go through ggml_vk_op_f32, so short-circuit and + // do the only thing needed for the dryrun. + vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, node, node->op); + ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1); + return false; + } + default: + break; + } } switch (node->op) { @@ -6334,16 +6516,17 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg bool first_node_in_batch = true; // true if next node will be first node in a batch int submit_node_idx = 0; // index to first node in a batch - // submit work every submit_count node to overlap CPU cmdbuffer generation with GPU execution - constexpr int submit_count = 100; + // Submit work every nodes_per_submit nodes to overlap CPU cmdbuffer generation with GPU execution. + // Start with a smaller count to get work submitted right away, and increase it after each submit. + int nodes_per_submit = 20; int submitted_nodes = 0; + int submit_count = 0; for (int i = 0; i < cgraph->n_nodes; i++) { if (first_node_in_batch) { submit_node_idx = i; } - bool submit = (submitted_nodes >= submit_count) || (i == last_node); - + bool submit = (submitted_nodes >= nodes_per_submit) || (i == last_node); bool enqueued = ggml_vk_build_graph(ctx, cgraph->nodes[i], i, cgraph->nodes[submit_node_idx], submit_node_idx, false, i == last_node, submit); @@ -6360,6 +6543,15 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg if (submit) { first_node_in_batch = true; submitted_nodes = 0; + switch (submit_count) { + case 0: + nodes_per_submit = 50; + break; + default: + nodes_per_submit = 100; + break; + } + submit_count++; } } @@ -6516,6 +6708,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: { + ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context; + if (op->op == GGML_OP_MUL_MAT_ID && + ggml_vk_get_device(ctx->device)->properties.limits.maxComputeSharedMemorySize < 32768) { + // If there's not enough shared memory for row_ids and the result tile, fallback to CPU + return false; + } switch (op->src[0]->type) { case GGML_TYPE_F32: case GGML_TYPE_F16: @@ -6713,8 +6911,9 @@ static const struct ggml_backend_reg_i ggml_backend_vk_reg_i = { ggml_backend_reg_t ggml_backend_vk_reg() { static ggml_backend_reg reg = { - /* .iface = */ ggml_backend_vk_reg_i, - /* .context = */ nullptr, + /* .api_version = */ GGML_BACKEND_API_VERSION, + /* .iface = */ ggml_backend_vk_reg_i, + /* .context = */ nullptr, }; return ® @@ -7131,7 +7330,7 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { const int32_t max_period = tensor->op_params[1]; tensor_clone = ggml_timestep_embedding(ggml_ctx, src0_clone, dim, max_period); } else if (tensor->op == GGML_OP_POOL_2D) { - enum ggml_op_pool op = static_cast(dst->op_params[0]); + enum ggml_op_pool op = static_cast(tensor->op_params[0]); const int32_t k0 = tensor->op_params[1]; const int32_t k1 = tensor->op_params[2]; const int32_t s0 = tensor->op_params[3]; @@ -7340,3 +7539,5 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) { VK_LOG_DEBUG("END ggml_vk_check_results_1(" << tensor->name << ")"); } #endif + +GGML_BACKEND_DL_IMPL(ggml_backend_vk_reg) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt b/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt index 10075db3377..51c78b7d229 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +++ b/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt @@ -3,5 +3,5 @@ find_package (Threads REQUIRED) set(TARGET vulkan-shaders-gen) add_executable(${TARGET} vulkan-shaders-gen.cpp) install(TARGETS ${TARGET} RUNTIME) -target_compile_features(${TARGET} PRIVATE cxx_std_11) +target_compile_features(${TARGET} PRIVATE cxx_std_17) target_link_libraries(vulkan-shaders-gen PUBLIC Threads::Threads) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp index d5b989735bc..5fc1ba4ad36 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp @@ -2,6 +2,15 @@ #extension GL_EXT_shader_explicit_arithmetic_types_int8 : require #endif +#include "types.comp" + +#if defined(A_TYPE_PACKED16) +layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];}; +#endif +#if defined(A_TYPE_PACKED32) +layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];}; +#endif + #if defined(DATA_A_F32) vec2 dequantize(uint ib, uint iqs, uint a_offset) { return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]); @@ -20,6 +29,11 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { const uint vui = uint(data_a[a_offset + ib].qs[iqs]); return (vec2(vui & 0xF, vui >> 4) - 8.0f) * d; } +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const float d = float(data_a_packed16[a_offset + ib].d); + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return (vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, (vui >> 12) & 0xF) - 8.0f) * d; +} #endif #if defined(DATA_A_Q4_1) @@ -29,6 +43,12 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { const uint vui = uint(data_a[a_offset + ib].qs[iqs]); return vec2(vui & 0xF, vui >> 4) * d + m; } +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const float d = float(data_a_packed16[a_offset + ib].d); + const float m = float(data_a_packed16[a_offset + ib].m); + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, (vui >> 12) & 0xF) * d + m; +} #endif #if defined(DATA_A_Q5_0) @@ -39,6 +59,14 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { const uint vui = uint(data_a[a_offset + ib].qs[iqs]); return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d; } +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const float d = float(data_a_packed16[a_offset + ib].d); + const uint uint_qh = uint(data_a_packed16[a_offset + ib].qh[1]) << 16 | data_a_packed16[a_offset + ib].qh[0]; + const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10); + const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10); + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return (vec4(((vui >> 0) & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, ((vui >> 12) & 0xF) | qh1.y) - 16.0f) * d; +} #endif #if defined(DATA_A_Q5_1) @@ -50,6 +78,15 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { const uint vui = uint(data_a[a_offset + ib].qs[iqs]); return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m; } +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const float d = float(data_a_packed16[a_offset + ib].d); + const float m = float(data_a_packed16[a_offset + ib].m); + const uint uint_qh = data_a_packed16[a_offset + ib].qh; + const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10); + const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10); + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return vec4(((vui >> 0) & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, ((vui >> 12) & 0xF) | qh1.y) * d + m; +} #endif #if defined(DATA_A_Q8_0) @@ -57,6 +94,12 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { const float d = float(data_a[a_offset + ib].d); return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])) * d; } +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const float d = float(data_a_packed16[a_offset + ib].d); + uint32_t v0 = data_a_packed16[a_offset + ib].qs[iqs/2]; + uint32_t v1 = data_a_packed16[a_offset + ib].qs[iqs/2 + 1]; + return vec4(int8_t(v0 & 0xFF), int8_t((v0 >> 8) & 0xFF), int8_t(v1 & 0xFF), int8_t((v1 >> 8) & 0xFF)) * d; +} #endif #if defined(DATA_A_IQ4_NL) @@ -65,4 +108,9 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { const uint vui = uint(data_a[a_offset + ib].qs[iqs]); return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d; } +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const float d = float(data_a_packed16[a_offset + ib].d); + const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]); + return vec4(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[(vui >> 4) & 0xF], kvalues_iq4nl[(vui >> 8) & 0xF], kvalues_iq4nl[(vui >> 12) & 0xF]) * d; +} #endif diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp index 34ef3da30b8..8de14fc03f1 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp @@ -10,6 +10,8 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; void main() { const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64; + init_iq4nl_shmem(); + const uint tid = gl_LocalInvocationID.x % 64; const uint il = tid/32; const uint ir = tid%32; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.comp b/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.comp index 4e1fa3af3ad..ab7c9d7eb1b 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.comp @@ -8,6 +8,13 @@ layout (push_constant) uniform parameter uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13; uint d_offset; float param1; float param2; + + uint ne0_012mp; uint ne0_012L; + uint ne0_01mp; uint ne0_01L; + uint ne0_0mp; uint ne0_0L; + uint ne1_012mp; uint ne1_012L; + uint ne1_01mp; uint ne1_01L; + uint ne1_0mp; uint ne1_0L; } p; layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; @@ -17,22 +24,30 @@ uint get_idx() { return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x; } +// see init_fastdiv_values in ggml-vulkan.cpp +uint fastdiv(uint n, uint mp, uint L) { + uint msbs, lsbs; + // msbs = mulhi(n, mp) + umulExtended(n, mp, msbs, lsbs); + return (msbs + n) >> L; +} + uint src0_idx(uint idx) { - const uint i03 = idx / (p.ne02*p.ne01*p.ne00); + const uint i03 = fastdiv(idx, p.ne0_012mp, p.ne0_012L); const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00; - const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00); + const uint i02 = fastdiv(idx - i03_offset, p.ne0_01mp, p.ne0_01L); const uint i02_offset = i02*p.ne01*p.ne00; - const uint i01 = (idx - i03_offset - i02_offset) / p.ne00; + const uint i01 = fastdiv(idx - i03_offset - i02_offset, p.ne0_0mp, p.ne0_0L); const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00; return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00; } uint dst_idx(uint idx) { - const uint i13 = idx / (p.ne12*p.ne11*p.ne10); + const uint i13 = fastdiv(idx, p.ne1_012mp, p.ne1_012L); const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10; - const uint i12 = (idx - i13_offset) / (p.ne11*p.ne10); + const uint i12 = fastdiv(idx - i13_offset, p.ne1_01mp, p.ne1_01L); const uint i12_offset = i12*p.ne11*p.ne10; - const uint i11 = (idx - i13_offset - i12_offset) / p.ne10; + const uint i11 = fastdiv(idx - i13_offset - i12_offset, p.ne1_0mp, p.ne1_0L); const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10; return i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + i10*p.nb10; } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp index 8d30b63c165..7f608315b68 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp @@ -12,6 +12,10 @@ void main() { const uint i11 = (gl_GlobalInvocationID.z)/p.ne12; const uint i12 = (gl_GlobalInvocationID.z)%p.ne12; +#if defined(DATA_A_IQ4_NL) + init_iq4nl_shmem(); +#endif + if (i00 >= p.ne00) { return; } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp index 5ad9b28daff..b6a0d564549 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp @@ -19,7 +19,7 @@ void main() { const uint tid = gl_LocalInvocationID.x; const uint start = gl_WorkGroupID.x * group_size + tid; - const uint end = start + group_size; + const uint end = (gl_WorkGroupID.x + 1) * group_size; tmp[tid] = 0.0f; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp index 825b91031f5..4c64fd47af7 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp @@ -5,7 +5,9 @@ layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; layout (binding = 0) readonly buffer A {float data_a[];}; +layout (binding = 0) readonly buffer A4 {vec4 data_a4[];}; layout (binding = 1) writeonly buffer D {float data_d[];}; +layout (binding = 1) writeonly buffer D4 {vec4 data_d4[];}; layout (push_constant) uniform parameter { uint ne; @@ -13,17 +15,34 @@ layout (push_constant) uniform parameter { } p; void main() { - const uint idx = gl_GlobalInvocationID.x; + // Each invocation handles four consecutive components + const uint idx = gl_GlobalInvocationID.x * 4; if (idx >= p.ne) { return; } - float result = 0.0f; + // Check if all four components are in bounds and aligned, + // then use vector loads + if (idx + 3 < p.ne && (p.ne % 4) == 0) { + vec4 result = vec4(0.0f); - [[unroll]] for (uint i = 0; i < p.k_num; i++) { - result += data_a[i * p.ne + idx]; - } + [[unroll]] for (uint i = 0; i < p.k_num; i++) { + result += data_a4[(i * p.ne + idx) / 4]; + } + + data_d4[idx / 4] = result; + } else { + [[unroll]] for (uint j = 0; j < 4; ++j) { + if (idx + j < p.ne) { + float result = 0.0f; - data_d[idx] = result; + [[unroll]] for (uint i = 0; i < p.k_num; i++) { + result += data_a[i * p.ne + idx + j]; + } + + data_d[idx + j] = result; + } + } + } } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp index 970aac6ef72..2d5b8e46613 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp @@ -3,7 +3,7 @@ #ifdef FLOAT16 #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require #endif -#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require +#extension GL_EXT_shader_explicit_arithmetic_types : require #include "mul_mat_vec_base.comp" @@ -12,16 +12,48 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; layout (constant_id = 0) const uint BLOCK_SIZE = 32; layout (constant_id = 1) const uint NUM_ROWS = 1; +#if !defined(DATA_A_F32) && !defined(DATA_A_F16) +#define K_PER_ITER 8 +#else +#define K_PER_ITER 2 +#endif + + uint a_offset, b_offset, d_offset, y_offset; shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE]; void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter) { - const uint col = i*BLOCK_SIZE + 2*tid; + const uint col = i*BLOCK_SIZE + K_PER_ITER*tid; const uint iqs = (col%QUANT_K)/QUANT_R; // quant index const uint iybs = col - col%QUANT_K; // y block start index +#if K_PER_ITER == 8 +#if QUANT_R == 2 + B_TYPE_VEC4 bv02 = data_b_v4[(b_offset + iybs + iqs) / 4]; + B_TYPE_VEC4 bv13 = data_b_v4[(b_offset + iybs + iqs + y_offset) / 4]; + FLOAT_TYPE b0 = FLOAT_TYPE(bv02.x); + FLOAT_TYPE b1 = FLOAT_TYPE(bv13.x); + FLOAT_TYPE b2 = FLOAT_TYPE(bv02.y); + FLOAT_TYPE b3 = FLOAT_TYPE(bv13.y); + FLOAT_TYPE b4 = FLOAT_TYPE(bv02.z); + FLOAT_TYPE b5 = FLOAT_TYPE(bv13.z); + FLOAT_TYPE b6 = FLOAT_TYPE(bv02.w); + FLOAT_TYPE b7 = FLOAT_TYPE(bv13.w); +#else + B_TYPE_VEC4 bv0 = data_b_v4[(b_offset + iybs + iqs) / 4]; + B_TYPE_VEC4 bv1 = data_b_v4[(b_offset + iybs + iqs) / 4 + 1]; + FLOAT_TYPE b0 = FLOAT_TYPE(bv0.x); + FLOAT_TYPE b1 = FLOAT_TYPE(bv0.y); + FLOAT_TYPE b2 = FLOAT_TYPE(bv0.z); + FLOAT_TYPE b3 = FLOAT_TYPE(bv0.w); + FLOAT_TYPE b4 = FLOAT_TYPE(bv1.x); + FLOAT_TYPE b5 = FLOAT_TYPE(bv1.y); + FLOAT_TYPE b6 = FLOAT_TYPE(bv1.z); + FLOAT_TYPE b7 = FLOAT_TYPE(bv1.w); +#endif +#else // Check if the second of the pair of elements is OOB, and don't fetch B or // accumulate it. We still fetch a pair of elements for A, which is fine for // quantized formats since they'll be within the same block. We should @@ -34,9 +66,24 @@ void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_ if (!OOB) { b1 = FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]); } +#endif [[unroll]] for (uint n = 0; n < num_rows; ++n) { const uint ib = ((first_row + n)*p.ncols + col)/QUANT_K; // block index +#if K_PER_ITER == 8 + const vec4 v = dequantize4(ib, iqs, a_offset); + const vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset); + + // matrix multiplication + temp[n] = fma(FLOAT_TYPE(v.x), b0, temp[n]); + temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]); + temp[n] = fma(FLOAT_TYPE(v.z), b2, temp[n]); + temp[n] = fma(FLOAT_TYPE(v.w), b3, temp[n]); + temp[n] = fma(FLOAT_TYPE(v2.x), b4, temp[n]); + temp[n] = fma(FLOAT_TYPE(v2.y), b5, temp[n]); + temp[n] = fma(FLOAT_TYPE(v2.z), b6, temp[n]); + temp[n] = fma(FLOAT_TYPE(v2.w), b7, temp[n]); +#else const vec2 v = dequantize(ib, iqs, a_offset); // matrix multiplication @@ -44,6 +91,7 @@ void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_ if (!OOB) { temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]); } +#endif } } @@ -61,22 +109,33 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { temp[i] = FLOAT_TYPE(0); } - const int unroll_count = 8; - - const uint num_iters = (p.ncols >= 2*tid) ? ((p.ncols - 2*tid + BLOCK_SIZE - 1) / BLOCK_SIZE) : 0; - const uint unrolled_iters = num_iters & ~(2*unroll_count - 1); + uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE); + if (num_iters * K_PER_ITER * BLOCK_SIZE + K_PER_ITER*tid < p.ncols) { + num_iters++; + } + int unroll_count = 4; + uint unrolled_iters = num_iters & ~(unroll_count - 1); uint i = 0; while (i < unrolled_iters) { // Manually partially unroll the loop [[unroll]] for (uint k = 0; k < unroll_count; ++k) { - iter(temp, first_row, num_rows, tid, i, false); - i += 2; + iter(temp, first_row, num_rows, tid, i*K_PER_ITER, false); + i++; + } + } + unroll_count = 2; + unrolled_iters = num_iters & ~(unroll_count - 1); + while (i < unrolled_iters) { + // Manually partially unroll the loop + [[unroll]] for (uint k = 0; k < unroll_count; ++k) { + iter(temp, first_row, num_rows, tid, i*K_PER_ITER, false); + i++; } } while (i < num_iters) { - iter(temp, first_row, num_rows, tid, i, true); - i += 2; + iter(temp, first_row, num_rows, tid, i*K_PER_ITER, true); + i++; } // sum up partial sums and write back result @@ -102,10 +161,17 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { void main() { const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); +#if defined(DATA_A_IQ4_NL) + init_iq4nl_shmem(); +#endif + // do NUM_ROWS at a time, unless there aren't enough remaining rows if (first_row + NUM_ROWS <= p.stride_d) { compute_outputs(first_row, NUM_ROWS); } else { + if (first_row >= p.stride_d) { + return; + } compute_outputs(first_row, p.stride_d - first_row); } } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp index 5920bc93641..2ec1af5c755 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp @@ -12,6 +12,9 @@ layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; layout (binding = 1) readonly buffer B {B_TYPE data_b[];}; +layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];}; +layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];}; + layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; #ifdef MUL_MAT_ID layout (binding = 3) readonly buffer IDS {int data_ids[];}; @@ -49,13 +52,16 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) { #endif #ifndef MUL_MAT_ID - const uint i13 = batch_idx / p.ne12; - const uint i12 = batch_idx % p.ne12; + uint batch_idx_a = 0; + if (batch_idx != 0) { + const uint i13 = batch_idx / p.ne12; + const uint i12 = batch_idx % p.ne12; - const uint i03 = i13 / p.broadcast3; - const uint i02 = i12 / p.broadcast2; + const uint i03 = i13 / p.broadcast3; + const uint i02 = i12 / p.broadcast2; - const uint batch_idx_a = i03 * p.ne02 + i02; + batch_idx_a = i03 * p.ne02 + i02; + } #else const uint expert_id = data_ids[expert_idx]; #endif diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp index ec8eadcd582..fcf02210ea6 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp @@ -1,4 +1,5 @@ #version 450 +#extension GL_EXT_shader_explicit_arithmetic_types : require #include "mul_mat_vec_base.comp" @@ -9,6 +10,10 @@ shared FLOAT_TYPE tmp[32]; void main() { const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; + if (row >= p.stride_d) { + return; + } + uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); @@ -28,38 +33,67 @@ void main() { const uint s_offset = 8*v_im; const uint y_offset = 128*v_im + l0; - tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp + FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { const uint y_idx = i * QUANT_K + y_offset; - const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x); - const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y); + f16vec2 d = data_a[ib0 + i].d; + const FLOAT_TYPE dall = d.x; + const FLOAT_TYPE dmin = d.y; + + B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0]; + B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8]; + B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16]; + B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24]; + B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32]; + B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40]; + B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48]; + B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56]; + + uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0]; + uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1]; + + uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F; + uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F; + uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F; + uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F; + + uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32)); + uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32)); + uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32)); + uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32)); + + uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0]; + uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]; + uvec2 qs0 = uvec2(unpack8(qs0_u16)); + uvec2 qs16 = uvec2(unpack8(qs16_u16)); FLOAT_TYPE sum1 = FLOAT_TYPE(0.0); FLOAT_TYPE sum2 = FLOAT_TYPE(0.0); - for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { - sum1 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 0]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 0) & 3), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 0) & 3), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 2) & 3), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 2) & 3), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 4) & 3), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 4) & 3), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 6) & 3), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 6) & 3), sum1)))))))); - sum2 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 0]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 0] >> 4) & 0xF), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 1] >> 4) & 0xF), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 2] >> 4) & 0xF), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 3] >> 4) & 0xF), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 4] >> 4) & 0xF), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 5] >> 4) & 0xF), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 6] >> 4) & 0xF), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 7] >> 4) & 0xF), sum2)))))))); + [[unroll]] for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { + sum1 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l] >> 0) & 3), + fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3), + fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l] >> 2) & 3), + fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3), + fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l] >> 4) & 3), + fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3), + fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l] >> 6) & 3), + fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1)))))))); + sum2 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_hi4[0]), + fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_hi4[1]), + fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_hi4[2]), + fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_hi4[3]), + fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_hi4[0]), + fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_hi4[1]), + fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_hi4[2]), + fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2)))))))); } - const uint tmp_idx = 16 * ix + tid; - tmp[tmp_idx] = fma(dall, sum1, fma(-dmin, sum2, tmp[tmp_idx])); + temp = fma(dall, sum1, fma(-dmin, sum2, temp)); } + tmp[gl_LocalInvocationID.x] = temp; + // sum up partial sums and write back result barrier(); [[unroll]] for (uint s = 16; s > 0; s >>= 1) { diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp index 3ca4ad85a5c..723fadde0d5 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp @@ -1,4 +1,5 @@ #version 450 +#extension GL_EXT_shader_explicit_arithmetic_types : require #include "mul_mat_vec_base.comp" @@ -9,6 +10,10 @@ shared FLOAT_TYPE tmp[32]; void main() { const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; + if (row >= p.stride_d) { + return; + } + uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); @@ -29,7 +34,7 @@ void main() { const uint q_offset = 32*v_im + l0; const uint y_offset = 128*v_im + l0; - tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp + FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp const uint s_shift = 4 * v_im; @@ -38,21 +43,44 @@ void main() { const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d); + B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0]; + B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8]; + B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16]; + B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24]; + B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32]; + B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40]; + B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48]; + B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56]; + + uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0]; + uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1]; + uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2]; + uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3]; + uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4]; + uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5]; + u8vec2 s0 = unpack8(s0_16); + u8vec2 s2 = unpack8(s2_16); + u8vec2 s4 = unpack8(s4_16); + u8vec2 s6 = unpack8(s6_16); + u8vec2 s8 = unpack8(s8_16); + u8vec2 s10 = unpack8(s10_16); + FLOAT_TYPE sum = FLOAT_TYPE(0.0); for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { - sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 0]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[0] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 0)) != 0) ? 0 : 4)), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[2] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 1)) != 0) ? 0 : 4)), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[4] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 2)) != 0) ? 0 : 4)), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[6] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 3)) != 0) ? 0 : 4)), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[1] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[3] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[5] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[7] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum)))))))); + sum = fma(FLOAT_TYPE(b0[l]) * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 0)) != 0) ? 0 : 4)), + fma(FLOAT_TYPE(b32[l]) * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 1)) != 0) ? 0 : 4)), + fma(FLOAT_TYPE(b64[l]) * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 2)) != 0) ? 0 : 4)), + fma(FLOAT_TYPE(b96[l]) * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 3)) != 0) ? 0 : 4)), + fma(FLOAT_TYPE(b16[l]) * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)), + fma(FLOAT_TYPE(b48[l]) * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)), + fma(FLOAT_TYPE(b80[l]) * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)), + fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum)))))))); } - const uint tmp_idx = 16 * ix + tid; - tmp[tmp_idx] = fma(d, sum, tmp[tmp_idx]); + temp = fma(d, sum, temp); } + tmp[gl_LocalInvocationID.x] = temp; + // sum up partial sums and write back result barrier(); [[unroll]] for (uint s = 16; s > 0; s >>= 1) { diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp index b7c9b722d12..5846f2e86f3 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp @@ -8,30 +8,14 @@ layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in; shared FLOAT_TYPE tmp[32]; -// Declare aliased versions of A and B bindings that can use 16b/32b loads for -// the quantized values, and vec4 loads for B. -struct block_q4_K_u32 -{ - f16vec2 d; - uint32_t scales[3*QUANT_K/64/4]; - uint32_t qs[QUANT_K/2/4]; -}; - -struct block_q4_K_u16 -{ - f16vec2 d; - uint16_t scales[3*QUANT_K/64/2]; - uint16_t qs[QUANT_K/2/2]; -}; - -layout (binding = 0) readonly buffer A_u32 {block_q4_K_u32 data_a_u32[];}; -layout (binding = 0) readonly buffer A_u16 {block_q4_K_u16 data_a_u16[];}; -layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];}; - // This shader assumes K_QUANTS_PER_ITERATION == 2 for alignment of loads void main() { const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; + if (row >= p.stride_d) { + return; + } + uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); @@ -64,9 +48,9 @@ void main() { const FLOAT_TYPE dall = FLOAT_TYPE(d.x); const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); - uint32_t scale0_u32 = data_a_u16[ib0 + i].scales[v_im ]; - uint32_t scale4_u32 = data_a_u16[ib0 + i].scales[v_im + 2]; - uint32_t scale8_u32 = data_a_u16[ib0 + i].scales[v_im + 4]; + uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; + uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; + uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; uvec4 scale0 = uvec4(unpack8(scale0_u32)); uvec4 scale4 = uvec4(unpack8(scale4_u32)); uvec4 scale8 = uvec4(unpack8(scale8_u32)); @@ -80,8 +64,8 @@ void main() { const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2)); const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2)); - uint32_t qs0_u32 = data_a_u32[ib0 + i].qs[q_offset / 4]; - uint32_t qs64_u32 = data_a_u32[ib0 + i].qs[q_offset / 4 + 16]; + uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4]; + uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16]; uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F; uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp index 2306785af42..b455cbd31ec 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp @@ -1,5 +1,7 @@ #version 450 +#extension GL_EXT_shader_explicit_arithmetic_types : require + #include "mul_mat_vec_base.comp" layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in; @@ -9,6 +11,10 @@ shared FLOAT_TYPE tmp[32]; void main() { const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; + if (row >= p.stride_d) { + return; + } + uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); @@ -28,73 +34,113 @@ void main() { const uint q_offset = 32*v_im + l0; const uint y_offset = 64*v_im + l0; - const uint8_t hm1 = uint8_t(1 << (2*v_im)); - const uint8_t hm2 = uint8_t(hm1 << 4); - - tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp + FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += 2) { const uint y1_idx = i * QUANT_K + y_offset; const uint y2_idx = y1_idx + 128; - const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x); - const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y); - - const uint8_t sc0 = uint8_t( data_a[ib0 + i].scales[v_im * 2 ] & 0x3f); - const uint8_t sc1 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 1] & 0x3f); - const uint8_t sc2 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 4] & 0x3f); - const uint8_t sc3 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 5] & 0x3f); - const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 ] & 0xc0) >> 2)); - const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2)); - const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2)); - const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2)); - - const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset ] & 0xf); - const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] & 0xf); - const uint8_t q4_2 = uint8_t(data_a[ib0 + i].qs[q_offset + 16] & 0xf); - const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset + 17] & 0xf); - const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset ] >> 4); - const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] >> 4); - const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 16] >> 4); - const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 17] >> 4); - const uint8_t q4_8 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf); - const uint8_t q4_9 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf); - const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] & 0xf); - const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] & 0xf); - const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] >> 4); - const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] >> 4); - const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] >> 4); - const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] >> 4); + f16vec2 d = data_a[ib0 + i].d; + const FLOAT_TYPE dall = FLOAT_TYPE(d.x); + const FLOAT_TYPE dmin = FLOAT_TYPE(d.y); + + uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ]; + uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2]; + uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4]; + uvec4 scale0 = uvec4(unpack8(scale0_u32)); + uvec4 scale4 = uvec4(unpack8(scale4_u32)); + uvec4 scale8 = uvec4(unpack8(scale8_u32)); + + const uint32_t sc0 = ( scale0.x & 0x3f); + const uint32_t sc1 = ( scale0.y & 0x3f); + const uint32_t sc2 = ( scale4.x & 0x3f); + const uint32_t sc3 = ( scale4.y & 0x3f); + const uint32_t sc4 = (( scale8.x & 0x0f) | ((scale0.x & 0xc0) >> 2)); + const uint32_t sc5 = (( scale8.y & 0x0f) | ((scale0.y & 0xc0) >> 2)); + const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2)); + const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2)); + + uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16); + uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16); + + uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F; + uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F; + uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F; + uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F; + + uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8])); + + uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4; + uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3; + uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010) << 0; + uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1; + + qs0_16_u32_lo4 += qs0_16_lo4_offset16; + qs0_16_u32_hi4 += qs0_16_hi4_offset16; + qs64_80_u32_lo4 += qs64_80_lo4_offset16; + qs64_80_u32_hi4 += qs64_80_hi4_offset16; + + uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4)); + uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4)); + uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4)); + uvec4 qs64_80_hi4 = uvec4(unpack8(qs64_80_u32_hi4)); + + const uint32_t q4_0 = qs0_16_lo4.x; + const uint32_t q4_1 = qs0_16_lo4.y; + const uint32_t q4_2 = qs0_16_lo4.z; + const uint32_t q4_3 = qs0_16_lo4.w; + const uint32_t q4_4 = qs0_16_hi4.x; + const uint32_t q4_5 = qs0_16_hi4.y; + const uint32_t q4_6 = qs0_16_hi4.z; + const uint32_t q4_7 = qs0_16_hi4.w; + const uint32_t q4_8 = qs64_80_lo4.x; + const uint32_t q4_9 = qs64_80_lo4.y; + const uint32_t q4_10 = qs64_80_lo4.z; + const uint32_t q4_11 = qs64_80_lo4.w; + const uint32_t q4_12 = qs64_80_hi4.x; + const uint32_t q4_13 = qs64_80_hi4.y; + const uint32_t q4_14 = qs64_80_hi4.z; + const uint32_t q4_15 = qs64_80_hi4.w; + + B_TYPE_VEC2 by10 = data_b_v2[(b_offset + y1_idx) / 2]; + B_TYPE_VEC2 by116 = data_b_v2[(b_offset + y1_idx) / 2 + 8]; + B_TYPE_VEC2 by132 = data_b_v2[(b_offset + y1_idx) / 2 + 16]; + B_TYPE_VEC2 by148 = data_b_v2[(b_offset + y1_idx) / 2 + 24]; + B_TYPE_VEC2 by20 = data_b_v2[(b_offset + y2_idx) / 2]; + B_TYPE_VEC2 by216 = data_b_v2[(b_offset + y2_idx) / 2 + 8]; + B_TYPE_VEC2 by232 = data_b_v2[(b_offset + y2_idx) / 2 + 16]; + B_TYPE_VEC2 by248 = data_b_v2[(b_offset + y2_idx) / 2 + 24]; const FLOAT_TYPE sx = - fma(FLOAT_TYPE(data_b[b_offset + y1_idx ]), (q4_0 + (((data_a[ib0 + i].qh[l0 ] & hm1) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), (q4_1 + (((data_a[ib0 + i].qh[l0 + 1] & hm1) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 16]), (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0)), - FLOAT_TYPE(data_b[b_offset + y1_idx + 17]) * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0))))); + fma(FLOAT_TYPE(by10.x), q4_0, + fma(FLOAT_TYPE(by10.y), q4_1, + fma(FLOAT_TYPE(by116.x), q4_2, + FLOAT_TYPE(by116.y) * q4_3))); const FLOAT_TYPE sy = - fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), (q4_4 + (((data_a[ib0 + i].qh[l0 ] & (hm1 << 1)) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), (q4_5 + (((data_a[ib0 + i].qh[l0 + 1] & (hm1 << 1)) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 48]), (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0)), - FLOAT_TYPE(data_b[b_offset + y1_idx + 49]) * (q4_7 + (((data_a[ib0 + i].qh[l0 + 17] & (hm1 << 1)) != 0) ? 16 : 0))))); + fma(FLOAT_TYPE(by132.x), q4_4, + fma(FLOAT_TYPE(by132.y), q4_5, + fma(FLOAT_TYPE(by148.x), q4_6, + FLOAT_TYPE(by148.y) * q4_7))); const FLOAT_TYPE sz = - fma(FLOAT_TYPE(data_b[b_offset + y2_idx ]), (q4_8 + (((data_a[ib0 + i].qh[l0 ] & hm2) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), (q4_9 + (((data_a[ib0 + i].qh[l0 + 1] & hm2) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 16]), (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0)), - FLOAT_TYPE(data_b[b_offset + y2_idx + 17]) * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0))))); + fma(FLOAT_TYPE(by20.x), q4_8, + fma(FLOAT_TYPE(by20.y), q4_9, + fma(FLOAT_TYPE(by216.x), q4_10, + FLOAT_TYPE(by216.y) * q4_11))); const FLOAT_TYPE sw = - fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), (q4_12 + (((data_a[ib0 + i].qh[l0 ] & (hm2 << 1)) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), (q4_13 + (((data_a[ib0 + i].qh[l0 + 1] & (hm2 << 1)) != 0) ? 16 : 0)), - fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 48]), (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0)), - FLOAT_TYPE(data_b[b_offset + y2_idx + 49]) * (q4_15 + (((data_a[ib0 + i].qh[l0 + 17] & (hm2 << 1)) != 0) ? 16 : 0))))); + fma(FLOAT_TYPE(by232.x), q4_12, + fma(FLOAT_TYPE(by232.y), q4_13, + fma(FLOAT_TYPE(by248.x), q4_14, + FLOAT_TYPE(by248.y) * q4_15))); const FLOAT_TYPE smin = - fma(FLOAT_TYPE(data_b[b_offset + y1_idx ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 17]), sc2, - fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 49]), sc3, - fma(FLOAT_TYPE(data_b[b_offset + y2_idx ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 17]), sc6, - (FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 49])) * sc7))); - const uint tmp_idx = 16 * ix + tid; - tmp[tmp_idx] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, tmp[tmp_idx])); + fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2, + fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3, + fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6, + (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7))); + temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp)); } + tmp[gl_LocalInvocationID.x] = temp; + // sum up partial sums and write back result barrier(); [[unroll]] for (uint s = 16; s > 0; s >>= 1) { diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp index 95c286eeb17..760aff85499 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp @@ -1,76 +1,107 @@ #version 450 +#extension GL_EXT_shader_explicit_arithmetic_types : require + #include "mul_mat_vec_base.comp" -layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +layout (constant_id = 0) const uint BLOCK_SIZE = 32; -shared FLOAT_TYPE tmp[32]; +shared FLOAT_TYPE tmp[BLOCK_SIZE]; void main() { const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z; + if (row >= p.stride_d) { + return; + } + uint a_offset, b_offset, d_offset; get_offsets(a_offset, b_offset, d_offset); const uint num_blocks_per_row = p.ncols / QUANT_K; const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row; - const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16 - const uint ix = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1 + // 16 threads are used to process each block + const uint it_size = gl_WorkGroupSize.x/16; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid%16; // 0...16 + const uint ix = tid/16; - const uint step = 16/K_QUANTS_PER_ITERATION; // 16 or 8 + const uint step = 8; - const uint v_im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... - const uint v_in = tid - step*v_im; // 0...15 or 0...7 + const uint v_im = itid/step; // 0 or 1. 0 computes 0..., 1 computes 128... + const uint v_in = itid - step*v_im; // 0...15 or 0...7 -#if K_QUANTS_PER_ITERATION == 1 - const uint l0 = v_in; // 0...15 - const uint is = 0; -#else const uint l0 = 4 * v_in; // 0, 4, 8, ..., 28 const uint is = v_in / 4; -#endif const uint ql_offset = 64*v_im + l0; const uint qh_offset = 32*v_im + l0; const uint s_offset = 8*v_im + is; const uint y_offset = 128*v_im + l0; - tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp + FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp - [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) { const uint y_idx = i * QUANT_K + y_offset; const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d); -#if K_QUANTS_PER_ITERATION == 1 - const uint tmp_idx = 16 * ix + tid; - tmp[tmp_idx] = fma(FLOAT_TYPE(data_b[b_offset + y_idx + 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x03) << 4)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x0c) << 2)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 0] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x30) >> 0)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0xc0) >> 2)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32), tmp[tmp_idx])))))))); -#else + FLOAT_TYPE scales[4]; + scales[0] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]); + scales[1] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]); + scales[2] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]); + scales[3] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]); + + uint32_t ql0_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16); + uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16); + + uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F; + uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F; + uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F; + uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F; + + uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16); + uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4; + uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2; + uint32_t qh4_u32 = (qh_u32 & 0x30303030) << 0; + uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2; + + uint32_t q0_u32 = ql0_u32_lo4 | qh0_u32; + uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32; + uint32_t q2_u32 = ql0_u32_hi4 | qh4_u32; + uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32; + + uvec4 q0 = uvec4(unpack8(q0_u32)); + uvec4 q1 = uvec4(unpack8(q1_u32)); + uvec4 q2 = uvec4(unpack8(q2_u32)); + uvec4 q3 = uvec4(unpack8(q3_u32)); + + B_TYPE_VEC4 by0 = data_b_v4[(b_offset + y_idx) / 4]; + B_TYPE_VEC4 by32 = data_b_v4[(b_offset + y_idx) / 4 + 8]; + B_TYPE_VEC4 by64 = data_b_v4[(b_offset + y_idx) / 4 + 16]; + B_TYPE_VEC4 by96 = data_b_v4[(b_offset + y_idx) / 4 + 24]; + FLOAT_TYPE sum = FLOAT_TYPE(0.0); [[unroll]] for (int l = 0; l < 4; ++l) { - sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32), - fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32), sum)))); + sum = fma(FLOAT_TYPE(by0[l]) * scales[0], FLOAT_TYPE(int8_t(q0[l]) - 32), + fma(FLOAT_TYPE(by32[l]) * scales[1], FLOAT_TYPE(int8_t(q1[l]) - 32), + fma(FLOAT_TYPE(by64[l]) * scales[2], FLOAT_TYPE(int8_t(q2[l]) - 32), + fma(FLOAT_TYPE(by96[l]) * scales[3], FLOAT_TYPE(int8_t(q3[l]) - 32), sum)))); } - tmp[16 * ix + tid] += sum; -#endif + temp += sum * d; } + tmp[gl_LocalInvocationID.x] = temp; // sum up partial sums and write back result + barrier(); - [[unroll]] for (uint s = 16; s > 0; s >>= 1) { + [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) { if (tid < s) { tmp[tid] += tmp[tid + s]; - } + } barrier(); } if (tid == 0) { diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp index fffdd18189d..2ff5c430519 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp @@ -75,6 +75,10 @@ shared u16vec2 row_ids[3072]; #endif void main() { +#if defined(DATA_A_IQ4_NL) + init_iq4nl_shmem(); +#endif + #ifdef MUL_MAT_ID const uint expert_idx = gl_GlobalInvocationID.z; #else diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp b/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp index f9727679ec5..6e20b6411ca 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp @@ -73,7 +73,9 @@ void soft_max(uint num_iters) { FLOAT_TYPE v = a * p.scale + slope * b; - max_val = max(max_val, v); + if (col < p.KX) { + max_val = max(max_val, v); + } if (idx < DATA_CACHE_SIZE) { data_cache[idx] = v; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp index 21dce72fc7d..eecc47f3a97 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp @@ -1,6 +1,8 @@ -#if !defined(DATA_A_F32) && !defined(DATA_A_F16) -#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require -#endif + +#if !defined(GGML_TYPES_COMP) +#define GGML_TYPES_COMP + +#extension GL_EXT_shader_explicit_arithmetic_types : require #if defined(DATA_A_F32) #define QUANT_K 1 @@ -28,24 +30,29 @@ #endif #endif -#if defined(DATA_A_Q4_0) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 32 -#define QUANT_R 2 +#define QUANT_K_Q4_0 32 +#define QUANT_R_Q4_0 2 struct block_q4_0 { float16_t d; uint8_t qs[16]; }; +struct block_q4_0_packed16 +{ + float16_t d; + uint16_t qs[16/2]; +}; +#if defined(DATA_A_Q4_0) +#define QUANT_K QUANT_K_Q4_0 +#define QUANT_R QUANT_R_Q4_0 #define A_TYPE block_q4_0 +#define A_TYPE_PACKED16 block_q4_0_packed16 #endif -#if defined(DATA_A_Q4_1) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 32 -#define QUANT_R 2 +#define QUANT_K_Q4_1 32 +#define QUANT_R_Q4_1 2 struct block_q4_1 { @@ -54,14 +61,22 @@ struct block_q4_1 uint8_t qs[16]; }; +struct block_q4_1_packed16 +{ + float16_t d; + float16_t m; + uint16_t qs[16/2]; +}; + +#if defined(DATA_A_Q4_1) +#define QUANT_K QUANT_K_Q4_1 +#define QUANT_R QUANT_R_Q4_1 #define A_TYPE block_q4_1 +#define A_TYPE_PACKED16 block_q4_1_packed16 #endif -#if defined(DATA_A_Q5_0) -#extension GL_EXT_shader_16bit_storage : require -#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require -#define QUANT_K 32 -#define QUANT_R 2 +#define QUANT_K_Q5_0 32 +#define QUANT_R_Q5_0 2 struct block_q5_0 { @@ -70,14 +85,22 @@ struct block_q5_0 uint8_t qs[16]; }; +struct block_q5_0_packed16 +{ + float16_t d; + uint16_t qh[2]; + uint16_t qs[16/2]; +}; + +#if defined(DATA_A_Q5_0) +#define QUANT_K QUANT_K_Q5_0 +#define QUANT_R QUANT_R_Q5_0 #define A_TYPE block_q5_0 +#define A_TYPE_PACKED16 block_q5_0_packed16 #endif -#if defined(DATA_A_Q5_1) -#extension GL_EXT_shader_16bit_storage : require -#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require -#define QUANT_K 32 -#define QUANT_R 2 +#define QUANT_K_Q5_1 32 +#define QUANT_R_Q5_1 2 struct block_q5_1 { @@ -87,114 +110,214 @@ struct block_q5_1 uint8_t qs[16]; }; +struct block_q5_1_packed16 +{ + float16_t d; + float16_t m; + uint qh; + uint16_t qs[16/2]; +}; + +#if defined(DATA_A_Q5_1) +#define QUANT_K QUANT_K_Q5_1 +#define QUANT_R QUANT_R_Q5_1 #define A_TYPE block_q5_1 +#define A_TYPE_PACKED16 block_q5_1_packed16 #endif -#if defined(DATA_A_Q8_0) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 32 -#define QUANT_R 1 +#define QUANT_K_Q8_0 32 +#define QUANT_R_Q8_0 1 struct block_q8_0 { float16_t d; int8_t qs[32]; }; +struct block_q8_0_packed16 +{ + float16_t d; + uint16_t qs[32/2]; +}; +#if defined(DATA_A_Q8_0) +#define QUANT_K QUANT_K_Q8_0 +#define QUANT_R QUANT_R_Q8_0 #define A_TYPE block_q8_0 +#define A_TYPE_PACKED16 block_q8_0_packed16 #endif // K-quants -#if defined(DATA_A_Q2_K) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 256 +#define QUANT_K_Q2_K 256 struct block_q2_K { - uint8_t scales[QUANT_K/16]; - uint8_t qs[QUANT_K/4]; + uint8_t scales[QUANT_K_Q2_K/16]; + uint8_t qs[QUANT_K_Q2_K/4]; + f16vec2 d; +}; + +struct block_q2_K_packed16 +{ + uint16_t scales[QUANT_K_Q2_K/16/2]; + uint16_t qs[QUANT_K_Q2_K/4/2]; + f16vec2 d; +}; + +struct block_q2_K_packed32 +{ + uint32_t scales[QUANT_K_Q2_K/16/4]; + uint32_t qs[QUANT_K_Q2_K/4/4]; f16vec2 d; }; +#if defined(DATA_A_Q2_K) +#define QUANT_K QUANT_K_Q2_K #define A_TYPE block_q2_K +#define A_TYPE_PACKED16 block_q2_K_packed16 +#define A_TYPE_PACKED32 block_q2_K_packed32 #endif -#if defined(DATA_A_Q3_K) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 256 +#define QUANT_K_Q3_K 256 struct block_q3_K { - uint8_t hmask[QUANT_K/8]; - uint8_t qs[QUANT_K/4]; + uint8_t hmask[QUANT_K_Q3_K/8]; + uint8_t qs[QUANT_K_Q3_K/4]; uint8_t scales[12]; float16_t d; }; +struct block_q3_K_packed16 +{ + uint16_t hmask[QUANT_K_Q3_K/8/2]; + uint16_t qs[QUANT_K_Q3_K/4/2]; + uint16_t scales[12/2]; + float16_t d; +}; + +#if defined(DATA_A_Q3_K) +#define QUANT_K QUANT_K_Q3_K #define A_TYPE block_q3_K +#define A_TYPE_PACKED16 block_q3_K_packed16 #endif -#if defined(DATA_A_Q4_K) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 256 +#define QUANT_K_Q4_K 256 struct block_q4_K { f16vec2 d; - uint8_t scales[3*QUANT_K/64]; - uint8_t qs[QUANT_K/2]; + uint8_t scales[3*QUANT_K_Q4_K/64]; + uint8_t qs[QUANT_K_Q4_K/2]; +}; + +struct block_q4_K_packed16 +{ + f16vec2 d; + uint16_t scales[3*QUANT_K_Q4_K/64/2]; + uint16_t qs[QUANT_K_Q4_K/2/2]; }; +struct block_q4_K_packed32 +{ + f16vec2 d; + uint32_t scales[3*QUANT_K_Q4_K/64/4]; + uint32_t qs[QUANT_K_Q4_K/2/4]; +}; + +#if defined(DATA_A_Q4_K) +#define QUANT_K QUANT_K_Q4_K #define A_TYPE block_q4_K +#define A_TYPE_PACKED16 block_q4_K_packed16 +#define A_TYPE_PACKED32 block_q4_K_packed32 #endif -#if defined(DATA_A_Q5_K) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 256 +#define QUANT_K_Q5_K 256 struct block_q5_K { f16vec2 d; uint8_t scales[12]; - uint8_t qh[QUANT_K/8]; - uint8_t qs[QUANT_K/2]; + uint8_t qh[QUANT_K_Q5_K/8]; + uint8_t qs[QUANT_K_Q5_K/2]; }; +struct block_q5_K_packed16 +{ + f16vec2 d; + uint16_t scales[12/2]; + uint16_t qh[QUANT_K_Q5_K/8/2]; + uint16_t qs[QUANT_K_Q5_K/2/2]; +}; + +#if defined(DATA_A_Q5_K) +#define QUANT_K QUANT_K_Q5_K #define A_TYPE block_q5_K +#define A_TYPE_PACKED16 block_q5_K_packed16 #endif -#if defined(DATA_A_Q6_K) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 256 +#define QUANT_K_Q6_K 256 struct block_q6_K { - uint8_t ql[QUANT_K/2]; - uint8_t qh[QUANT_K/4]; - int8_t scales[QUANT_K/16]; + uint8_t ql[QUANT_K_Q6_K/2]; + uint8_t qh[QUANT_K_Q6_K/4]; + int8_t scales[QUANT_K_Q6_K/16]; float16_t d; }; +struct block_q6_K_packed16 +{ + uint16_t ql[QUANT_K_Q6_K/2/2]; + uint16_t qh[QUANT_K_Q6_K/4/2]; + int8_t scales[QUANT_K_Q6_K/16]; + float16_t d; +}; + +#if defined(DATA_A_Q6_K) +#define QUANT_K QUANT_K_Q6_K #define A_TYPE block_q6_K +#define A_TYPE_PACKED16 block_q6_K_packed16 #endif // IQuants -#if defined(DATA_A_IQ4_NL) -#extension GL_EXT_shader_16bit_storage : require -#define QUANT_K 32 -#define QUANT_R 2 +#define QUANT_K_IQ4_NL 32 +#define QUANT_R_IQ4_NL 2 struct block_iq4_nl { float16_t d; - uint8_t qs[QUANT_K/2]; + uint8_t qs[QUANT_K_IQ4_NL/2]; }; -#define A_TYPE block_iq4_nl +struct block_iq4_nl_packed16 +{ + float16_t d; + uint16_t qs[QUANT_K_IQ4_NL/2/2]; +}; + +#if defined(DATA_A_IQ4_NL) -const int8_t kvalues_iq4nl[16] = { +const int8_t kvalues_iq4nl_const[16] = { int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10), int8_t(1), int8_t(13), int8_t(25), int8_t(38), int8_t(53), int8_t(69), int8_t(89), int8_t(113) }; + +shared FLOAT_TYPE kvalues_iq4nl[16]; + +void init_iq4nl_shmem() +{ + // copy the table into shared memory and sync + if (gl_LocalInvocationIndex.x < 16) { + kvalues_iq4nl[gl_LocalInvocationIndex.x] = FLOAT_TYPE(kvalues_iq4nl_const[gl_LocalInvocationIndex.x]); + } + barrier(); +} + +#define QUANT_K QUANT_K_IQ4_NL +#define QUANT_R QUANT_R_IQ4_NL +#define A_TYPE block_iq4_nl +#define A_TYPE_PACKED16 block_iq4_nl_packed16 #endif + +#endif // !defined(GGML_TYPES_COMP) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp index fe3e4cb39ff..5c317b68b7e 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -317,10 +317,10 @@ void process_shaders() { std::string data_a_key = "DATA_A_" + to_uppercase(tname); std::string shader = (string_ends_with(tname, "_k")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp"; - string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}})); - string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}})); + string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}})); + string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}})); - string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}})); + string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}})); // Dequant shaders if (tname != "f16") { @@ -331,11 +331,11 @@ void process_shaders() { shader = (tname == "f32" || tname == "f16") ? "get_rows.comp" : "get_rows_quant.comp"; if (tname == "f16") { - string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); + string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}})); } else { - string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}}); + string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}})); } - string_to_spv("get_rows_" + tname + "_f32", shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}}); + string_to_spv("get_rows_" + tname + "_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}})); } } @@ -474,9 +474,15 @@ void write_output_files() { int main(int argc, char** argv) { std::map args; - for (int i = 1; i < argc; i += 2) { - if (i + 1 < argc) { - args[argv[i]] = argv[i + 1]; + for (int i = 1; i < argc; ++i) { + std::string arg = argv[i]; + if (arg.rfind("--", 0) == 0) { + if (i + 1 < argc && argv[i + 1][0] != '-') { + args[arg] = argv[i + 1]; + ++i; + } else { + args[arg] = ""; + } } } diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index ee72a173e07..2c338dee556 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -831,6 +831,15 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = { .to_float = (ggml_to_float_t) dequantize_row_tq2_0, .from_float_ref = (ggml_from_float_t) quantize_row_tq2_0_ref, }, + [GGML_TYPE_IQ4_NL_4_4] = { + .type_name = "iq4_nl_4x4", + .blck_size = QK4_NL, + .blck_size_interleave = 4, + .type_size = sizeof(block_iq4_nl), + .is_quantized = true, + .to_float = NULL, + .from_float_ref = NULL, + }, }; const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type) { @@ -941,6 +950,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "POOL_2D_BACK", "UPSCALE", "PAD", + "PAD_REFLECT_1D", "ARANGE", "TIMESTEP_EMBEDDING", "ARGSORT", @@ -974,7 +984,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "OPT_STEP_ADAMW", }; -static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81"); +static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -1036,6 +1046,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "pool_2d_back(x)", "upscale(x)", "pad(x)", + "pad_reflect_1d(x)", "arange(start, stop, step)", "timestep_embedding(timesteps, dim, max_period)", "argsort(x)", @@ -1069,7 +1080,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "adamw(x)", }; -static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81"); +static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82"); static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); @@ -2255,6 +2266,7 @@ struct ggml_tensor * ggml_argmax( struct ggml_context * ctx, struct ggml_tensor * a) { GGML_ASSERT(ggml_is_matrix(a)); + GGML_ASSERT(a->ne[0] <= INT32_MAX); struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, a->ne[1]); @@ -4087,6 +4099,37 @@ struct ggml_tensor * ggml_pad( return result; } +// ggml_pad_reflect_1d + +struct ggml_tensor * ggml_pad_reflect_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int p0, + int p1) { + GGML_ASSERT(p0 >= 0); + GGML_ASSERT(p1 >= 0); + + GGML_ASSERT(p0 < a->ne[0]); // padding length on each size must be less than the + GGML_ASSERT(p1 < a->ne[0]); // existing length of the dimension being padded + + GGML_ASSERT(ggml_is_contiguous(a)); + GGML_ASSERT(a->type == GGML_TYPE_F32); + + struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, + a->ne[0] + p0 + p1, + a->ne[1], + a->ne[2], + a->ne[3]); + + int32_t params[] = { p0, p1 }; + ggml_set_op_params(result, params, sizeof(params)); + + result->op = GGML_OP_PAD_REFLECT_1D; + result->src[0] = a; + + return result; +} + // ggml_arange struct ggml_tensor * ggml_arange( @@ -4138,6 +4181,7 @@ struct ggml_tensor * ggml_argsort( struct ggml_context * ctx, struct ggml_tensor * a, enum ggml_sort_order order) { + GGML_ASSERT(a->ne[0] <= INT32_MAX); struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne); ggml_set_op_params_i32(result, 0, (int32_t) order); @@ -5019,8 +5063,10 @@ static void ggml_hash_map_free(struct hash_map * map) { } // utility functions to change gradients -// if a is in acc_table, modify gradients in-place and mark result as gradient accumulator -// else if a is in zero_table, replace a +// isrc is the index of tensor in cgraph->visited_has_set.keys +// the corresponding gradient (accumulators) are also at position isrc +// if tensor has a gradient accumulator, modify that accumulator in-place +// else if there is no gradient for tensor, set the corresponding value // else, just add/subtract/etc. the gradients static void ggml_add_or_set( @@ -5028,11 +5074,14 @@ static void ggml_add_or_set( struct ggml_cgraph * cgraph, size_t isrc, struct ggml_tensor * tensor) { + struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc]; + GGML_ASSERT(src); if (cgraph->grads[isrc]) { - cgraph->grads[isrc] = ggml_add_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]); + cgraph->grads[isrc] = ggml_add_impl(ctx, cgraph->grads[isrc], tensor, /*inplace =*/ cgraph->grad_accs[isrc]); } else { cgraph->grads[isrc] = tensor; } + ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name); ggml_build_forward_expand(cgraph, cgraph->grads[isrc]); } @@ -5040,18 +5089,20 @@ static void ggml_acc_or_set( struct ggml_context * ctx, struct ggml_cgraph * cgraph, size_t isrc, - struct ggml_tensor * src, struct ggml_tensor * tensor, const size_t nb1, const size_t nb2, const size_t nb3, const size_t offset) { + struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc]; + GGML_ASSERT(src); if (cgraph->grads[isrc]) { cgraph->grads[isrc] = ggml_acc_impl(ctx, cgraph->grads[isrc], tensor, nb1, nb2, nb3, offset, cgraph->grad_accs[isrc]); } else { struct ggml_tensor * a_zero = ggml_scale(ctx, src, 0.0f); // FIXME this is going to produce NaN if a contains inf/NaN cgraph->grads[isrc] = ggml_acc_impl(ctx, a_zero, tensor, nb1, nb2, nb3, offset, false); } + ggml_format_name(cgraph->grads[isrc], "grad for %s", cgraph->visited_hash_set.keys[isrc]->name); ggml_build_forward_expand(cgraph, cgraph->grads[isrc]); } @@ -5059,13 +5110,15 @@ static void ggml_add1_or_set( struct ggml_context * ctx, struct ggml_cgraph * cgraph, size_t isrc, - struct ggml_tensor * src, struct ggml_tensor * tensor) { + struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc]; + GGML_ASSERT(src); if (cgraph->grads[isrc]) { cgraph->grads[isrc] = ggml_add1_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]); } else { cgraph->grads[isrc] = ggml_repeat(ctx, tensor, src); } + ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name); ggml_build_forward_expand(cgraph, cgraph->grads[isrc]); } @@ -5074,11 +5127,14 @@ static void ggml_sub_or_set( struct ggml_cgraph * cgraph, size_t isrc, struct ggml_tensor * tensor) { + struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc]; + GGML_ASSERT(src); if (cgraph->grads[isrc]) { cgraph->grads[isrc] = ggml_sub_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]); } else { cgraph->grads[isrc] = ggml_neg(ctx, tensor); } + ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name); ggml_build_forward_expand(cgraph, cgraph->grads[isrc]); } @@ -5095,12 +5151,12 @@ static void ggml_compute_backward( struct ggml_tensor * src1 = tensor->src[1]; struct ggml_tensor * src2 = tensor->src[2]; struct ggml_hash_set * hash_set = &cgraph->visited_hash_set; - const size_t isrc0 = ggml_hash_find(hash_set, src0); - const size_t isrc1 = ggml_hash_find(hash_set, src1); - const size_t isrc2 = ggml_hash_find(hash_set, src2); - const bool src0_needs_grads = isrc0 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc0) && grads_needed[isrc0]; - const bool src1_needs_grads = isrc1 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc1) && grads_needed[isrc1]; - const bool src2_needs_grads = isrc2 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc2) && grads_needed[isrc2]; + const size_t isrc0 = src0 ? ggml_hash_find(hash_set, src0) : (size_t) -1; + const size_t isrc1 = src1 ? ggml_hash_find(hash_set, src1) : (size_t) -1; + const size_t isrc2 = src2 ? ggml_hash_find(hash_set, src2) : (size_t) -1; + const bool src0_needs_grads = src0 && isrc0 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc0) && grads_needed[isrc0]; + const bool src1_needs_grads = src1 && isrc1 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc1) && grads_needed[isrc1]; + const bool src2_needs_grads = src2 && isrc2 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc2) && grads_needed[isrc2]; switch (tensor->op) { case GGML_OP_DUP: { @@ -5200,7 +5256,7 @@ static void ggml_compute_backward( } break; case GGML_OP_SUM: { if (src0_needs_grads) { - ggml_add1_or_set(ctx, cgraph, isrc0, src0, grad); + ggml_add1_or_set(ctx, cgraph, isrc0, grad); } } break; case GGML_OP_SUM_ROWS: { @@ -5210,7 +5266,7 @@ static void ggml_compute_backward( } break; case GGML_OP_MEAN: { if (src0_needs_grads) { - ggml_add1_or_set(ctx, cgraph, isrc0, src0, ggml_scale_impl(ctx, grad, 1.0f/src0->ne[0], false)); + ggml_add1_or_set(ctx, cgraph, isrc0, ggml_scale_impl(ctx, grad, 1.0f/src0->ne[0], false)); } } break; case GGML_OP_REPEAT: { @@ -5363,7 +5419,7 @@ static void ggml_compute_backward( nb3 = (nb3 / n0) * ng; } - ggml_acc_or_set(ctx, cgraph, isrc0, src0, grad, nb1, nb2, nb3, offset); + ggml_acc_or_set(ctx, cgraph, isrc0, grad, nb1, nb2, nb3, offset); } } break; case GGML_OP_PERMUTE: { @@ -5597,10 +5653,9 @@ void ggml_build_backward_expand( const int n_nodes_f = cgraph->n_nodes; - const size_t hash_size = ggml_hash_size(2*cgraph->size); - memset(cgraph->grads, 0, hash_size*sizeof(struct ggml_tensor *)); - memset(cgraph->grad_accs, 0, hash_size*sizeof(struct ggml_tensor *)); - bool * grads_needed = calloc(hash_size, sizeof(bool)); + memset(cgraph->grads, 0, cgraph->visited_hash_set.size*sizeof(struct ggml_tensor *)); + memset(cgraph->grad_accs, 0, cgraph->visited_hash_set.size*sizeof(struct ggml_tensor *)); + bool * grads_needed = calloc(cgraph->visited_hash_set.size, sizeof(bool)); { bool any_params = false; @@ -5621,7 +5676,7 @@ void ggml_build_backward_expand( continue; } - bool node_needs_grad = node->flags & GGML_TENSOR_FLAG_PARAM; + bool node_needs_grad = (node->flags & GGML_TENSOR_FLAG_PARAM) || (node->flags & GGML_TENSOR_FLAG_LOSS); bool ignore_src[GGML_MAX_SRC] = {false}; switch (node->op) { // gradients in node->src[0] for one reason or another have no effect on output gradients @@ -5638,7 +5693,7 @@ void ggml_build_backward_expand( } break; // gradients in node->src[1] for one reason or another have no effect on output gradients - case GGML_OP_CPY: // gradients in CPY target are irrelevant + case GGML_OP_CPY: // gradients in CPY target are irrelevant case GGML_OP_GET_ROWS: // row indices not differentiable case GGML_OP_GET_ROWS_BACK: // same as for GET_ROWS case GGML_OP_ROPE: // positions not differentiable @@ -5665,9 +5720,12 @@ void ggml_build_backward_expand( node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE); const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node); + GGML_ASSERT(igrad != GGML_HASHSET_FULL); + GGML_ASSERT(ggml_bitset_get(cgraph->visited_hash_set.used, igrad)); if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) { - cgraph->grads[igrad] = ggml_dup_tensor(ctx_static, node); - cgraph->grad_accs[igrad] = cgraph->grads[igrad]; + cgraph->grad_accs[igrad] = ggml_dup_tensor(ctx_static, node); + cgraph->grads[igrad] = cgraph->grad_accs[igrad]; + ggml_format_name(cgraph->grad_accs[igrad], "grad acc for %s", node->name); } grads_needed[igrad] = true; } @@ -5761,15 +5819,15 @@ struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) { struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1) { struct ggml_cgraph cgraph = { - /*.size =*/ 0, - /*.n_nodes =*/ i1 - i0, - /*.n_leafs =*/ 0, - /*.nodes =*/ cgraph0->nodes + i0, - /*.grads =*/ cgraph0->grads ? cgraph0->grads + i0 : NULL, - /*.grad_accs =*/ cgraph0->grad_accs ? cgraph0->grad_accs + i0 : NULL, - /*.leafs =*/ NULL, - /*.hash_table =*/ { 0, NULL, NULL }, - /*.order =*/ cgraph0->order, + /*.size =*/ 0, + /*.n_nodes =*/ i1 - i0, + /*.n_leafs =*/ 0, + /*.nodes =*/ cgraph0->nodes + i0, + /*.grads =*/ NULL, // gradients would need visited_hash_set + /*.grad_accs =*/ NULL, + /*.leafs =*/ NULL, + /*.visited_hash_set =*/ { 0, NULL, NULL }, + /*.order =*/ cgraph0->order, }; return cgraph; @@ -5799,12 +5857,22 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) { } } + if (dst->grads) { + memset(dst->grads, 0, dst->visited_hash_set.size*sizeof(struct ggml_tensor *)); + memset(dst->grad_accs, 0, dst->visited_hash_set.size*sizeof(struct ggml_tensor *)); + } if (src->grads) { GGML_ASSERT(dst->grads != NULL); GGML_ASSERT(dst->grad_accs != NULL); for (int i = 0; i < src->n_nodes; ++i) { const size_t igrad_src = ggml_hash_find(&src->visited_hash_set, src->nodes[i]); const size_t igrad_dst = ggml_hash_find(&dst->visited_hash_set, dst->nodes[i]); + + GGML_ASSERT(igrad_src != GGML_HASHSET_FULL); + GGML_ASSERT(ggml_bitset_get(src->visited_hash_set.used, igrad_src)); + GGML_ASSERT(igrad_dst != GGML_HASHSET_FULL); + GGML_ASSERT(ggml_bitset_get(dst->visited_hash_set.used, igrad_dst)); + dst->grads[igrad_dst] = src->grads[igrad_src]; dst->grad_accs[igrad_dst] = src->grad_accs[igrad_src]; } @@ -5839,12 +5907,8 @@ void ggml_graph_reset(struct ggml_cgraph * cgraph) { if (node->op == GGML_OP_OPT_STEP_ADAMW) { // clear momenta - if (node->src[2]->data) { - ggml_set_zero(node->src[2]); - } - if (node->src[3]->data) { - ggml_set_zero(node->src[3]); - } + ggml_set_zero(node->src[2]); + ggml_set_zero(node->src[3]); } // initial gradients of loss should be 1, 0 otherwise @@ -7549,3 +7613,26 @@ void ggml_log_set(ggml_log_callback log_callback, void * user_data) { g_logger_state.log_callback = log_callback ? log_callback : ggml_log_callback_default; g_logger_state.log_callback_user_data = user_data; } + +void ggml_threadpool_params_init(struct ggml_threadpool_params * p, int n_threads) { + p->n_threads = n_threads; + p->prio = 0; // default priority (usually means normal or inherited) + p->poll = 50; // hybrid-polling enabled + p->strict_cpu = false; // no strict placement (all threads share same cpumask) + p->paused = false; // threads are ready to go + memset(p->cpumask, 0, GGML_MAX_N_THREADS); // all-zero means use the default affinity (usually inherited) +} + +struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads) { + struct ggml_threadpool_params p; + ggml_threadpool_params_init(&p, n_threads); + return p; +} + +bool ggml_threadpool_params_match(const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1) { + if (p0->n_threads != p1->n_threads ) return false; + if (p0->prio != p1->prio ) return false; + if (p0->poll != p1->poll ) return false; + if (p0->strict_cpu != p1->strict_cpu ) return false; + return memcmp(p0->cpumask, p1->cpumask, GGML_MAX_N_THREADS) == 0; +} diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index e9bd2dbb05e..cd023029711 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -2884dd72fea8922910fe53387c3d17ab928d3a8e +a5960e80d3e65ce6ff18f90315ab96f63cf9c4cc