Model built with Intel oneMKL

* Add windows build with cuBLAS

* Remove error causing lines for cuBLAS on Windows

.github/workflows/build.yml

@@ -236,6 +236,61 @@ jobs:
           name: whisper-blas-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
 
+  windows-cublas:
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        build: [Release]
+        arch: [x64]
+        cublas: [ON]
+        sdl2: [ON]
+        include:
+          - arch: x64
+            s2arc: x64
+          - sdl2: ON
+            s2ver: 2.26.0
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
+
+      - name: Add msbuild to PATH
+        uses: microsoft/setup-msbuild@v1
+
+      - name: Install CUDA Toolkit
+        id: cuda-toolkit
+        uses: Jimver/cuda-toolkit@v0.2.10
+
+      - 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 }}
+          -DWHISPER_CUBLAS=1
+
+      - name: Build
+        run: |
+          cd ./build
+          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
+
+      - name: Copy SDL2.dll
+        if: matrix.sdl2 == 'ON'
+        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
+
+      - name: Upload binaries
+        if: matrix.sdl2 == 'ON'
+        uses: actions/upload-artifact@v1
+        with:
+          name: whisper-cublas-bin-${{ matrix.arch }}
+          path: build/bin/${{ matrix.build }}
+
   emscripten:
     runs-on: ubuntu-latest
 
@@ -282,7 +337,7 @@ jobs:
 
       - 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
 
@@ -298,11 +353,11 @@ jobs:
         with:
           distribution: zulu
           java-version: 17
-      
+
       - name: Setup Android SDK
         uses: android-actions/setup-android@v2
 
       - name: Build
         run: |
           cd examples/whisper.android
-          ./gradlew assembleRelease --no-daemon
+          ./gradlew assembleRelease --no-daemon

.github/workflows/release-deb.yml

@@ -0,0 +1,68 @@
+name: release-deb
+
+on:
+  release:
+    types: [created]
+
+jobs:
+  build:
+    runs-on: ubuntu-20.04
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: Configure
+        run: |
+          set -x -e
+          VERSION=$(echo $GITHUB_REF | cut --delimiter=/ -f 3)
+          ID="whisper-cpp-small_${VERSION}_amd64"
+          
+          echo "PKG_VERSION=$VERSION"     >> $GITHUB_ENV
+          echo "PKG_ID=$ID"               >> $GITHUB_ENV
+
+      - name: Install deps
+        run: |
+          sudo apt install -y --no-install-recommends intel-mkl
+
+      - name: Build
+        run: |
+          cmake -S . -B build-mkl \
+            -DCMAKE_BUILD_TYPE=Release\
+            -DBUILD_SHARED_LIBS=0\
+            -DWHISPER_BLAS=1\
+            -DWHISPER_BLAS_VENDOR=Intel10_64lp
+          cd build-mkl
+          make
+          cd ..
+
+      - name: Create package tree
+        env:
+          GITHUB_REPO: ${{ github.repository }}
+        run: |
+          export ROOT=$PKG_ID/opt/project/whisper.cpp
+          mkdir -p $ROOT/bin
+          mkdir -p $ROOT/share
+          mkdir -p $PKG_ID/DEBIAN
+          
+          cp build-mkl/bin/main $ROOT/bin/whisper 
+          cp -r contrib/debian/control $PKG_ID/DEBIAN/
+          
+          echo "Version: $PKG_VERSION"          >> $PKG_ID/DEBIAN/control
+          echo "Vcs-Git: $GITHUB_REPO"          >> $PKG_ID/DEBIAN/control
+          echo "Vcs-Git-Commit: $GITHUB_SHA"    >> $PKG_ID/DEBIAN/control
+          
+          models/download-ggml-model.sh small
+          build-mkl/bin/quantize models/ggml-small.bin \
+            $ROOT/share/ggml-small-q5_1.bin q5_1
+
+      - name: Create deb package
+        run: |
+          mkdir artifacts
+          dpkg-deb --build --root-owner-group $PKG_ID
+
+      - name: Upload Release Asset
+        uses: xresloader/upload-to-github-release@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          release_id: ${{ github.event.release.id }}
+          file: ${{ env.PKG_ID }}.deb

.gitignore

@@ -5,6 +5,7 @@
 .test/
 .vs/
 .vscode/
+.idea/
 .DS_Store
 
 build/
@@ -16,6 +17,7 @@ build-cublas/
 build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
+cmake-build-debug/
 
 /main
 /stream

CMakeLists.txt

@@ -2,10 +2,6 @@ cmake_minimum_required (VERSION 3.0)
 
 project(whisper.cpp VERSION 1.4.1)
 
-if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
-    add_compile_options(/utf-8)
-endif ()
-
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
 
@@ -53,17 +49,21 @@ option(WHISPER_BUILD_EXAMPLES         "whisper: build examples" ${WHISPER_STANDA
 
 option(WHISPER_SDL2                   "whisper: support for libSDL2" OFF)
 
+option(WHISPER_NO_AVX                 "whisper: disable AVX"  OFF)
+option(WHISPER_NO_AVX2                "whisper: disable AVX2" OFF)
+option(WHISPER_NO_FMA                 "whisper: disable FMA"  OFF)
+option(WHISPER_NO_F16C                "whisper: disable F16c" OFF)
+
 if (APPLE)
     option(WHISPER_NO_ACCELERATE         "whisper: disable Accelerate framework" OFF)
-    option(WHISPER_NO_AVX                "whisper: disable AVX" OFF)
-    option(WHISPER_NO_AVX2               "whisper: disable AVX2" OFF)
-    option(WHISPER_NO_FMA                "whisper: disable FMA" OFF)
-
-    option(WHISPER_COREML                "whisper: enable Core ML framework" OFF)
-    option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback" OFF)
+    option(WHISPER_COREML                "whisper: enable Core ML framework"     OFF)
+    option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback"    OFF)
 else()
-    option(WHISPER_OPENBLAS              "whisper: support for OpenBLAS" OFF)
-    option(WHISPER_CUBLAS                "whisper: support for cuBLAS" OFF)
+    option(WHISPER_BLAS                  "whisper: use BLAS libraries"  OFF)
+    option(WHISPER_BLAS_VENDOR           "whisper: BLAS library vendor" Generic)
+    option(WHISPER_OPENBLAS              "whisper: prefer OpenBLAS"     OFF)
+    option(WHISPER_CUBLAS                "whisper: support for cuBLAS"  OFF)
+    option(WHISPER_CLBLAST               "whisper: use CLBlast"         OFF)
 endif()
 
 option(WHISPER_PERF "whisper: enable perf timings" OFF)
@@ -129,19 +129,32 @@ if (APPLE)
 endif()
 
 if (WHISPER_OPENBLAS)
-    find_library(OPENBLAS_LIB
-        NAMES openblas libopenblas
-        )
-    if (OPENBLAS_LIB)
-        message(STATUS "OpenBLAS found")
-
-        set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${OPENBLAS_LIB})
-        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
-    else()
-        message(WARNING "OpenBLAS not found")
-    endif()
+    set(WHISPER_BLAS_VENDOR "OpenBLAS")
+    set(WHISPER_BLAS ON)
 endif()
 
+if (WHISPER_BLAS)
+    if (WHISPER_STATIC)
+        set(BLA_STATIC 1)
+    else()
+        set(BLA_STATIC 0)
+    endif ()
+    set(BLA_VENDOR ${WHISPER_BLAS_VENDOR})
+    set(BLA_SIZEOF_INTEGER 8)
+    find_package(BLAS)
+
+    if(BLAS_FOUND)
+        message(STATUS "BLAS compatible library found")
+        message(STATUS "Libraries ${BLAS_LIBRARIES}")
+        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
+
+        include_directories(${BLAS_INCLUDE_DIRS})
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
+    else()
+        message(WARNING "BLAS library was not found")
+    endif()
+endif ()
+
 if (WHISPER_CUBLAS)
     cmake_minimum_required(VERSION 3.17)
 
@@ -149,7 +162,7 @@ if (WHISPER_CUBLAS)
 
     if (CUDAToolkit_FOUND)
         message(STATUS "cuBLAS found")
-
+        set(CMAKE_CUDA_COMPILER /usr/local/cuda/bin/nvcc)
         enable_language(CUDA)
 
         set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h)
@@ -167,6 +180,21 @@ if (WHISPER_CUBLAS)
     endif()
 endif()
 
+if (WHISPER_CLBLAST)
+    find_package(CLBlast)
+    if (CLBlast_FOUND)
+        message(STATUS "CLBlast found")
+
+        set(GGML_OPENCL_SOURCES ggml-opencl.c ggml-opencl.h)
+
+        add_compile_definitions(GGML_USE_CLBLAST)
+
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} clblast)
+    else()
+        message(WARNING "CLBlast not found")
+    endif()
+endif()
+
 # compiler flags
 
 if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
@@ -274,6 +302,7 @@ add_library(${TARGET}
     ggml.h
     ggml.c
     ${GGML_CUDA_SOURCES}
+    ${GGML_OPENCL_SOURCES}
     whisper.h
     whisper.cpp
     )

Makefile

@@ -171,13 +171,22 @@ ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif
 
+ifdef WHISPER_CLBLAST
+	CFLAGS 		+= -DGGML_USE_CLBLAST
+	LDFLAGS	 	+= -lclblast -lOpenCL
+	WHISPER_OBJ	+= ggml-opencl.o
+	
+ggml-opencl.o: ggml-opencl.c ggml-opencl.h
+	$(CC) $(CFLAGS) -c $< -o $@
+endif
+
 ifdef WHISPER_GPROF
 	CFLAGS   += -pg
 	CXXFLAGS += -pg
 endif
 
 ifneq ($(filter aarch64%,$(UNAME_M)),)
-	CFLAGS += -mcpu=native
+	CFLAGS   += -mcpu=native
 	CXXFLAGS += -mcpu=native
 endif
 
@@ -188,15 +197,18 @@ endif
 
 ifneq ($(filter armv7%,$(UNAME_M)),)
 	# 32-bit ARM, for example on Armbian or possibly raspbian
-	CFLAGS += -mfpu=neon -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+	#CFLAGS   += -mfpu=neon -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
+	#CXXFLAGS += -mfpu=neon -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
 
-	# 64-bit ARM, use these (TODO: auto-detect 64-bit)
-	# CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+	# 64-bit ARM on 32-bit OS, use these (TODO: auto-detect 64-bit)
+	CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
+	CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
 endif
 
 ifneq ($(filter armv8%,$(UNAME_M)),)
 	# Raspberry Pi 4
-	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
+	CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
+	CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
 endif
 
 #

README.md

@@ -20,6 +20,7 @@ High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisp
 - Zero memory allocations at runtime
 - Runs on the CPU
 - [Partial GPU support for NVIDIA via cuBLAS](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
+- [Partial OpenCL GPU support via CLBlast](https://github.com/ggerganov/whisper.cpp#opencl-gpu-support-via-clblast)
 - [C-style API](https://github.com/ggerganov/whisper.cpp/blob/master/whisper.h)
 
 Supported platforms:
@@ -311,6 +312,29 @@ make clean
 WHISPER_CUBLAS=1 make -j
 ```
 
+## OpenCL GPU support via CLBlast
+
+For cards and integrated GPUs that support OpenCL, the Encoder processing can be largely offloaded to the GPU through CLBlast. This is especially useful for users with AMD APU's or low end devices for up to ~2x speedup.
+
+First, make sure you have installed `CLBlast` for your OS or Distribution: https://github.com/CNugteren/CLBlast
+
+Now build `whisper.cpp` with CLBlast support:
+
+```
+Makefile:
+cd whisper.cpp
+make clean
+WHISPER_CLBLAST=1 make -j
+
+CMake:
+cd whisper.cpp ; mkdir build ; cd build
+cmake -DWHISPER_CLBLAST=ON  ..
+make clean
+make -j
+cp bin/* ../ 
+```
+
+
 Run all the examples as usual.
 
 ## Limitations

contrib/debian/control

@@ -0,0 +1,5 @@
+Package: whisper-small-cpp
+Architecture: amd64
+Maintainer: Alexey Kharlamov <alexey@kharlamov.biz>
+Description: Whisper Speech to Text Converter
+Depends: libc6 (>= 2.2.1), intel-mkl

ggml.c

@@ -671,35 +671,91 @@ float vmaxvq_f32(float32x4_t v) {
 }
 
 int8x8_t vzip1_s8(int8x8_t a, int8x8_t b) {
-    return vget_low_s8(vcombine_s8(a, b));
+    int8x8_t res;
+
+    res[0] = a[0]; res[1] = b[0];
+    res[2] = a[1]; res[3] = b[1];
+    res[4] = a[2]; res[5] = b[2];
+    res[6] = a[3]; res[7] = b[3];
+
+    return res;
 }
 
 int8x8_t vzip2_s8(int8x8_t a, int8x8_t b) {
-    return vget_high_s8(vcombine_s8(a, b));
+    int8x8_t res;
+
+    res[0] = a[4]; res[1] = b[4];
+    res[2] = a[5]; res[3] = b[5];
+    res[4] = a[6]; res[5] = b[6];
+    res[6] = a[7]; res[7] = b[7];
+
+    return res;
 }
 
 uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
-    return vget_low_u8(vcombine_u8(a, b));
+    uint8x8_t res;
+
+    res[0] = a[0]; res[1] = b[0];
+    res[2] = a[1]; res[3] = b[1];
+    res[4] = a[2]; res[5] = b[2];
+    res[6] = a[3]; res[7] = b[3];
+
+    return res;
 }
 
 uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
-    return vget_high_u8(vcombine_u8(a, b));
+    uint8x8_t res;
+
+    res[0] = a[4]; res[1] = b[4];
+    res[2] = a[5]; res[3] = b[5];
+    res[4] = a[6]; res[5] = b[6];
+    res[6] = a[7]; res[7] = b[7];
+
+    return res;
 }
 
 int8x16_t vzip1q_s8(int8x16_t a, int8x16_t b) {
-    return vcombine_s8(vget_low_s8(a), vget_low_s8(b));
+    int8x16_t res;
+
+    res[0]  = a[0]; res[1]  = b[0]; res[2]  = a[1]; res[3]  = b[1];
+    res[4]  = a[2]; res[5]  = b[2]; res[6]  = a[3]; res[7]  = b[3];
+    res[8]  = a[4]; res[9]  = b[4]; res[10] = a[5]; res[11] = b[5];
+    res[12] = a[6]; res[13] = b[6]; res[14] = a[7]; res[15] = b[7];
+
+    return res;
 }
 
 int8x16_t vzip2q_s8(int8x16_t a, int8x16_t b) {
-    return vcombine_s8(vget_high_s8(a), vget_high_s8(b));
+    int8x16_t res;
+
+    res[0]  = a[8];  res[1]  = b[8];  res[2]  = a[9];  res[3]  = b[9];
+    res[4]  = a[10]; res[5]  = b[10]; res[6]  = a[11]; res[7]  = b[11];
+    res[8]  = a[12]; res[9]  = b[12]; res[10] = a[13]; res[11] = b[13];
+    res[12] = a[14]; res[13] = b[14]; res[14] = a[15]; res[15] = b[15];
+
+    return res;
 }
 
 uint8x16_t vzip1q_u8(uint8x16_t a, uint8x16_t b) {
-    return vcombine_u8(vget_low_u8(a), vget_low_u8(b));
+    uint8x16_t res;
+
+    res[0]  = a[0];  res[1]  = b[0];  res[2]  = a[1];  res[3]  = b[1];
+    res[4]  = a[2];  res[5]  = b[2];  res[6]  = a[3];  res[7]  = b[3];
+    res[8]  = a[4];  res[9]  = b[4];  res[10] = a[5];  res[11] = b[5];
+    res[12] = a[6];  res[13] = b[6];  res[14] = a[7];  res[15] = b[7];
+
+    return res;
 }
 
 uint8x16_t vzip2q_u8(uint8x16_t a, uint8x16_t b) {
-    return vcombine_u8(vget_high_u8(a), vget_high_u8(b));
+    uint8x16_t res;
+
+    res[0]  = a[8];  res[1]  = b[8];  res[2]  = a[9];  res[3]  = b[9];
+    res[4]  = a[10]; res[5]  = b[10]; res[6]  = a[11]; res[7]  = b[11];
+    res[8]  = a[12]; res[9]  = b[12]; res[10] = a[13]; res[11] = b[13];
+    res[12] = a[14]; res[13] = b[14]; res[14] = a[15]; res[15] = b[15];
+
+    return res;
 }
 
 int32x4_t vcvtnq_s32_f32(float32x4_t v) {