Enable MacOS ARM64 and Windows CUDA builds

.github/actions/build-plugin/action.yaml

@@ -8,6 +8,10 @@ inputs:
     description: 'Build configuration'
     required: false
     default: 'RelWithDebInfo'
+  cublas:
+    description: 'Enable cuBLAS'
+    required: false
+    default: 'cpu'
   codesign:
     description: 'Enable codesigning (macOS only)'
     required: false
@@ -78,6 +82,7 @@ runs:
         $BuildArgs = @{
           Target = '${{ inputs.target }}'
           Configuration = '${{ inputs.config }}'
+          Cublas = '${{ inputs.cublas }}'
         }
 
         .github/scripts/Build-Windows.ps1 @BuildArgs

.github/actions/package-plugin/action.yaml

@@ -8,6 +8,10 @@ inputs:
     description: 'Build configuration'
     required: false
     default: 'RelWithDebInfo'
+  cublas:
+    description: 'Enable cuBLAS'
+    required: false
+    default: 'cpu'
   codesign:
     description: 'Enable codesigning (macOS only)'
     required: false
@@ -108,6 +112,7 @@ runs:
         $PackageArgs = @{
           Target = '${{ inputs.target }}'
           Configuration = '${{ inputs.config }}'
+          Cublas = '${{ inputs.cublas }}'
         }
 
         if ( '${{ inputs.package }}' -eq 'true' ) {

.github/scripts/.build.zsh

@@ -230,7 +230,15 @@ ${_usage_host:-}"
           -DCODESIGN_IDENTITY=${CODESIGN_IDENT:--}
         )
 
-        cmake_build_args+=(--preset ${_preset} --parallel --config ${config} -- ONLY_ACTIVE_ARCH=NO -arch arm64 -arch x86_64)
+        cmake_build_args+=(--preset ${_preset} --parallel --config ${config} -- ONLY_ACTIVE_ARCH=YES)
+        # check the MACOS_ARCH env var to determine the build architecture
+        if [[ -n ${MACOS_ARCH} ]] {
+          cmake_build_args+=(-arch ${MACOS_ARCH})
+        } else {
+          # error out
+          log_error "No MACOS_ARCH environment variable set. Please set it to the desired architecture."
+          exit 2
+        }
         cmake_install_args+=(build_macos --config ${config} --prefix "${project_root}/release/${config}")
 
         local -a xcbeautify_opts=()

.github/scripts/.package.zsh

@@ -168,8 +168,10 @@ ${_usage_host:-}"
 
   if [[ ${host_os} == macos ]] {
     autoload -Uz check_packages read_codesign read_codesign_installer read_codesign_pass
+    # get the arch from MACOS_ARCH env var
+    local -r macos_arch=${MACOS_ARCH:-$(uname -m)}
 
-    local output_name="${product_name}-${product_version}-${host_os}-universal"
+    local output_name="${product_name}-${product_version}-${host_os}-${macos_arch}"
 
     if [[ ! -d ${project_root}/release/${config}/${product_name}.plugin ]] {
       log_error 'No release artifact found. Run the build script or the CMake install procedure first.'

.github/scripts/Build-Windows.ps1

@@ -4,6 +4,8 @@ param(
     [string] $Target = 'x64',
     [ValidateSet('Debug', 'RelWithDebInfo', 'Release', 'MinSizeRel')]
     [string] $Configuration = 'RelWithDebInfo',
+    [ValidateSet('cpu', '12.2.0', '11.8.0')]
+    [string] $Cublas = 'cpu',
     [switch] $SkipAll,
     [switch] $SkipBuild,
     [switch] $SkipDeps,
@@ -77,6 +79,13 @@ function Build {
             '--preset', $Preset
         )
 
+        if ( $Cublas -ne 'cpu' ) {
+            $CmakeArgs += @(
+                '-DLOCALVOCAL_WITH_CUDA=ON',
+                "-DCUDA_TOOLKIT_ROOT_DIR=$Env:CUDA_TOOLKIT_ROOT_DIR"
+            )
+        }
+
         $CmakeBuildArgs += @(
             '--build'
             '--preset', $Preset

.github/scripts/Package-Windows.ps1

@@ -4,6 +4,8 @@ param(
     [string] $Target = 'x64',
     [ValidateSet('Debug', 'RelWithDebInfo', 'Release', 'MinSizeRel')]
     [string] $Configuration = 'RelWithDebInfo',
+    [ValidateSet('cpu', '12.2.0', '11.8.0')]
+    [string] $Cublas = 'cpu',
     [switch] $BuildInstaller,
     [switch] $SkipDeps
 )
@@ -47,8 +49,9 @@ function Package {
     $BuildSpec = Get-Content -Path ${BuildSpecFile} -Raw | ConvertFrom-Json
     $ProductName = $BuildSpec.name
     $ProductVersion = $BuildSpec.version
+    $CudaName = "cuda${Cublas}"
 
-    $OutputName = "${ProductName}-${ProductVersion}-windows-${Target}"
+    $OutputName = "${ProductName}-${ProductVersion}-windows-${Target}-${CudaName}"
 
     if ( ! $SkipDeps ) {
         Install-BuildDependencies -WingetFile "${ScriptHome}/.Wingetfile"

.github/workflows/build-project.yaml

@@ -60,6 +60,9 @@ jobs:
     name: Build for macOS 🍏
     runs-on: macos-13
     needs: check-event
+    strategy:
+      matrix:
+        architecture: [x86_64, arm64]
     defaults:
       run:
         shell: zsh --no-rcs --errexit --pipefail {0}
@@ -114,15 +117,17 @@ jobs:
       - name: Build Plugin 🧱
         uses: ./.github/actions/build-plugin
         with:
-          target: macos-universal
+          target: macos-${{ matrix.architecture }}
           config: ${{ needs.check-event.outputs.config }}
           codesign: ${{ fromJSON(needs.check-event.outputs.codesign) }}
           codesignIdent: ${{ steps.codesign.outputs.codesignIdent }}
+        env:
+          MACOS_ARCH: ${{ matrix.architecture }}
 
       - name: Package Plugin 📀
         uses: ./.github/actions/package-plugin
         with:
-          target: macos-universal
+          target: macos-${{ matrix.architecture }}
           config: ${{ needs.check-event.outputs.config }}
           package: ${{ fromJSON(needs.check-event.outputs.package) }}
           codesign: ${{ fromJSON(needs.check-event.outputs.codesign) && fromJSON(steps.codesign.outputs.haveCodesignIdent) }}
@@ -132,19 +137,21 @@ jobs:
           notarize: ${{ fromJSON(needs.check-event.outputs.notarize) && fromJSON(steps.codesign.outputs.haveNotarizationUser) }}
           codesignUser: ${{ secrets.MACOS_NOTARIZATION_USERNAME }}
           codesignPass: ${{ secrets.MACOS_NOTARIZATION_PASSWORD }}
+        env:
+          MACOS_ARCH: ${{ matrix.architecture }}
 
       - name: Upload Artifacts 📡
         uses: actions/upload-artifact@v3
         with:
-          name: ${{ steps.setup.outputs.pluginName }}-${{ steps.setup.outputs.pluginVersion }}-macos-universal-${{ needs.check-event.outputs.commitHash }}
-          path: ${{ github.workspace }}/release/${{ steps.setup.outputs.pluginName }}-${{ steps.setup.outputs.pluginVersion }}-macos-universal.*
+          name: ${{ steps.setup.outputs.pluginName }}-${{ steps.setup.outputs.pluginVersion }}-macos-${{ matrix.architecture }}-${{ needs.check-event.outputs.commitHash }}
+          path: ${{ github.workspace }}/release/${{ steps.setup.outputs.pluginName }}-${{ steps.setup.outputs.pluginVersion }}-macos-${{ matrix.architecture }}.*
 
       - name: Upload Debug Symbol Artifacts 🪲
         uses: actions/upload-artifact@v3
         if: ${{ needs.check-event.outputs.config == 'Release' }}
         with:
-          name: ${{ steps.setup.outputs.pluginName }}-${{ steps.setup.outputs.pluginVersion }}-macos-universal-${{ needs.check-event.outputs.commitHash }}-dSYMs
-          path: ${{ github.workspace }}/release/${{ steps.setup.outputs.pluginName }}-${{ steps.setup.outputs.pluginVersion }}-macos-universal-dSYMs.*
+          name: ${{ steps.setup.outputs.pluginName }}-${{ steps.setup.outputs.pluginVersion }}-macos-${{ matrix.architecture }}-${{ needs.check-event.outputs.commitHash }}-dSYMs
+          path: ${{ github.workspace }}/release/${{ steps.setup.outputs.pluginName }}-${{ steps.setup.outputs.pluginVersion }}-macos-${{ matrix.architecture }}-dSYMs.*
 
   ubuntu-build:
     name: Build for Ubuntu 🐧
@@ -218,6 +225,9 @@ jobs:
     name: Build for Windows 🪟
     runs-on: windows-2022
     needs: check-event
+    strategy:
+      matrix:
+        cublas: [cpu, 12.2.0, 11.8.0]
     defaults:
       run:
         shell: pwsh
@@ -227,6 +237,20 @@ jobs:
           submodules: recursive
           fetch-depth: 0
 
+      - name: Install CUDA Toolkit
+        if: ${{ matrix.cublas != 'cpu' }}
+        id: cuda-toolkit
+        uses: Jimver/cuda-toolkit@v0.2.11
+        with:
+          cuda: '${{ matrix.cublas }}'
+
+      - name: Set CUDA_TOOLKIT_ROOT_DIR if CUDA is installed
+        if: ${{ matrix.cublas != 'cpu' }}
+        run: |
+          ls "$env:CUDA_PATH\bin"
+          ls -d 2 "$env:CUDA_PATH\lib"
+          "CUDA_TOOLKIT_ROOT_DIR=$env:CUDA_PATH" >> $env:GITHUB_ENV
+
       - name: Set Up Environment 🔧
         id: setup
         run: |
@@ -247,16 +271,18 @@ jobs:
         with:
           target: x64
           config: ${{ needs.check-event.outputs.config }}
+          cublas: ${{ matrix.cublas }}
 
       - name: Package Plugin 📀
         uses: ./.github/actions/package-plugin
         with:
           target: x64
           config: ${{ needs.check-event.outputs.config }}
           package: ${{ fromJSON(needs.check-event.outputs.package) }}
+          cublas: ${{ matrix.cublas }}
 
       - name: Upload Artifacts 📡
         uses: actions/upload-artifact@v3
         with:
-          name: ${{ steps.setup.outputs.pluginName }}-${{ steps.setup.outputs.pluginVersion }}-windows-x64-${{ needs.check-event.outputs.commitHash }}
+          name: ${{ steps.setup.outputs.pluginName }}-${{ steps.setup.outputs.pluginVersion }}-windows-x64-${{ matrix.cublas }}-${{ needs.check-event.outputs.commitHash }}
           path: ${{ github.workspace }}/release/${{ steps.setup.outputs.pluginName }}-${{ steps.setup.outputs.pluginVersion }}-windows-x64*.*

cmake/BuildWhispercpp.cmake

@@ -17,8 +17,20 @@ if(UNIX AND NOT APPLE)
                                     -DWHISPER_NO_AVX2=ON)
 endif()
 if(APPLE)
-  # disable Metal on MacOS as it hurts performance right now
-  set(WHISPER_ADDITIONAL_CMAKE_ARGS -DWHISPER_METAL=OFF)
+  # check the "MACOS_ARCH" env var to figure out if this is x86 or arm64
+  if(NOT DEFINED ENV{MACOS_ARCH})
+    message(FATAL_ERROR "The MACOS_ARCH environment variable is not set. Please set it to either `x86` or `arm64`")
+  endif(NOT DEFINED ENV{MACOS_ARCH})
+  set(CMAKE_OSX_ARCHITECTURES_ "$ENV{MACOS_ARCH}")
+  if($ENV{MACOS_ARCH} STREQUAL "x86_64")
+    set(WHISPER_ADDITIONAL_CMAKE_ARGS -DWHISPER_METAL=OFF -DWHISPER_COREML=OFF -DWHISPER_NO_AVX=OFF
+                                      -DWHISPER_NO_AVX2=OFF -DWHISPER_NO_F16C=OFF)
+  else()
+    set(WHISPER_ADDITIONAL_CMAKE_ARGS -DWHISPER_METAL=OFF -DWHISPER_COREML=OFF -DWHISPER_NO_AVX=ON -DWHISPER_NO_AVX2=ON
+                                      -DWHISPER_NO_F16C=ON -DWHISPER_NO_FMA=ON)
+  endif()
+  set(WHISPER_EXTRA_CXX_FLAGS
+      "-Wno-shorten-64-to-32 -Wno-unused-parameter -Wno-unused-function -Wno-unguarded-availability-new")
 endif()
 
 if(WIN32)