Implement TBB-based parallelism for C lib

1 files changed, 83 insertions, 58 deletions

diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 4ce575e..2fc5e94 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -225,54 +225,48 @@ jobs:
 
   # Currently only on x86.
   c_tests:
-    name: C Makefile tests
+    name: C tests SIMD=${{ matrix.simd }} TBB=${{ matrix.use_tbb }}
     runs-on: ubuntu-latest
-
+    strategy:
+      fail-fast: false
+      matrix:
+        use_tbb: ["OFF", "ON"]
+        simd: ["x86-intrinsics", "amd64-asm"]
     steps:
     - uses: actions/checkout@v4
-    # Test the intrinsics-based implementations.
-    - run: make -f Makefile.testing test
-      working-directory: ./c
-    - run: make -f Makefile.testing clean && rm blake3_sse2.c
-      working-directory: ./c
-    - run: BLAKE3_NO_SSE2=1 make -f Makefile.testing test
-      working-directory: ./c
-    - run: make -f Makefile.testing clean && rm blake3_sse41.c
-      working-directory: ./c
-    - run: BLAKE3_NO_SSE2=1 BLAKE3_NO_SSE41=1 make -f Makefile.testing test
-      working-directory: ./c
-    - run: make -f Makefile.testing clean && rm blake3_avx2.c
-      working-directory: ./c
-    - run: BLAKE3_NO_SSE2=1 BLAKE3_NO_SSE41=1 BLAKE3_NO_AVX2=1 make -f Makefile.testing test
-      working-directory: ./c
-    - run: make -f Makefile.testing clean && rm blake3_avx512.c
-      working-directory: ./c
-    - run: BLAKE3_NO_SSE2=1 BLAKE3_NO_SSE41=1 BLAKE3_NO_AVX2=1 BLAKE3_NO_AVX512=1 make -f Makefile.testing test
-      working-directory: ./c
-    # Test the assembly implementations.
-    - run: make -f Makefile.testing test_asm
-      working-directory: ./c
-    - run: make -f Makefile.testing clean && rm blake3_sse2_x86-64_unix.S
-      working-directory: ./c
-    - run: BLAKE3_NO_SSE2=1 make -f Makefile.testing test_asm
-      working-directory: ./c
-    - run: make -f Makefile.testing clean && rm blake3_sse41_x86-64_unix.S
-      working-directory: ./c
-    - run: BLAKE3_NO_SSE2=1 BLAKE3_NO_SSE41=1 make -f Makefile.testing test_asm
-      working-directory: ./c
-    - run: make -f Makefile.testing clean && rm blake3_avx2_x86-64_unix.S
-      working-directory: ./c
-    - run: BLAKE3_NO_SSE2=1 BLAKE3_NO_SSE41=1 BLAKE3_NO_AVX2=1 make -f Makefile.testing test_asm
-      working-directory: ./c
-    - run: make -f Makefile.testing clean && rm blake3_avx512_x86-64_unix.S
-      working-directory: ./c
-    - run: BLAKE3_NO_SSE2=1 BLAKE3_NO_SSE41=1 BLAKE3_NO_AVX2=1 BLAKE3_NO_AVX512=1 make -f Makefile.testing test_asm
-      working-directory: ./c
-    # Restore the files we deleted above.
-    - run: git checkout .
-    # Build the example.
-    - run: make -f Makefile.testing example
-      working-directory: ./c
+    - run: |
+        sudo apt-get update
+        sudo apt-get install ninja-build libtbb-dev libtbb12
+    # Test the intrinsics-based and assembly-based implementations.
+    - run: |
+        cmake --fresh -S c -B c/build -G Ninja -DBLAKE3_TESTING=ON -DBLAKE3_TESTING_CI=ON "-DBLAKE3_SIMD_TYPE=${{ matrix.simd }}"
+        cmake --build c/build --target test
+        cat c/build/Testing/Temporary/LastTest.log
+    - run: |
+        cmake --fresh -S c -B c/build -G Ninja -DBLAKE3_TESTING=ON -DBLAKE3_TESTING_CI=ON "-DBLAKE3_SIMD_TYPE=${{ matrix.simd }}" -DBLAKE3_NO_SSE2=1
+        cmake --build c/build --target test
+        cat c/build/Testing/Temporary/LastTest.log
+    - run: |
+        cmake --fresh -S c -B c/build -G Ninja -DBLAKE3_TESTING=ON -DBLAKE3_TESTING_CI=ON "-DBLAKE3_SIMD_TYPE=${{ matrix.simd }}" -DBLAKE3_NO_SSE2=1 -DBLAKE3_NO_SSE41=1
+        cmake --build c/build --target test
+        cat c/build/Testing/Temporary/LastTest.log
+    - run: |
+        cmake --fresh -S c -B c/build -G Ninja -DBLAKE3_TESTING=ON -DBLAKE3_TESTING_CI=ON "-DBLAKE3_SIMD_TYPE=${{ matrix.simd }}" "-DBLAKE3_NO_SSE2=1" "-DBLAKE3_NO_SSE41=1" "-DBLAKE3_NO_AVX2=1"
+        cmake --build c/build --target test
+        cat c/build/Testing/Temporary/LastTest.log
+    - run: |
+        cmake --fresh -S c -B c/build -G Ninja -DBLAKE3_TESTING=ON -DBLAKE3_TESTING_CI=ON "-DBLAKE3_SIMD_TYPE=${{ matrix.simd }}" "-DBLAKE3_NO_SSE2=1" "-DBLAKE3_NO_SSE41=1" "-DBLAKE3_NO_AVX2=1" "-DBLAKE3_NO_AVX512=1"
+        cmake --build c/build --target test
+        cat c/build/Testing/Temporary/LastTest.log
+    # Test with TBB disabled/enabled.
+    - run: |
+        cmake --fresh -S c -B c/build -G Ninja -DBLAKE3_TESTING=ON -DBLAKE3_TESTING_CI=ON "-DBLAKE3_USE_TBB=${{ matrix.use_tbb }}"
+        cmake --build c/build --target test
+        cat c/build/Testing/Temporary/LastTest.log
+    # Build the example with TBB disabled/enabled.
+    - run: |
+        cmake --fresh -S c -B c/build -G Ninja -DBLAKE3_TESTING=ON -DBLAKE3_TESTING_CI=ON -DBLAKE3_EXAMPLES=ON
+        cmake --build c/build --target blake3-example
 
   # Note that this jobs builds AArch64 binaries from an x86_64 host.
   build_apple_silicon:
@@ -323,28 +317,59 @@ jobs:
           cd /work
           ~/.cargo/bin/cargo test --features prefer_intrinsics
 
-  # CMake build test (Library only), current macOS/Linux only.
+  # CMake build test (Library only).
   cmake_current_build:
-    name: CMake ${{ matrix.os }} ${{ matrix.compiler }}
+    name: CMake ${{ matrix.os }} CC=${{ matrix.toolchain.cc }} CXX=${{ matrix.toolchain.cxx }} TBB=${{ matrix.use_tbb }}
     runs-on: ${{ matrix.os }}
     strategy:
       fail-fast: false
       matrix:
-        os: ["ubuntu-latest", "macOS-latest", "windows-latest"]
-        compiler: [gcc, clang, cl]
+        cmakeVersion: [latest]
+        ninjaVersion: [latest]
+        os: [ubuntu-latest, macOS-latest, windows-latest]
+        toolchain: [
+          { cc: cl, cxx: cl },
+          { cc: clang, cxx: clang++ },
+          { cc: clang-cl, cxx: clang-cl },
+          { cc: gcc, cxx: g++ },
+        ]
+        use_tbb: [OFF, ON]
         exclude:
-          - os: windows-latest
-            compiler: gcc
-          - os: ubuntu-latest
-            compiler: msvc
           - os: macOS-latest
-            compiler: msvc
+            toolchain: { cc: cl, cxx: cl }
+          - os: macOS-latest
+            toolchain: { cc: clang-cl, cxx: clang-cl }
+          - os: ubuntu-latest
+            toolchain: { cc: cl, cxx: cl }
+          - os: ubuntu-latest
+            toolchain: { cc: clang-cl, cxx: clang-cl }
+          - os: windows-latest
+            toolchain: { cc: clang, cxx: clang++ }
+            use_tbb: ON
+          - os: windows-latest
+            toolchain: { cc: gcc, cxx: g++ }
+            use_tbb: ON
     steps:
       - uses: actions/checkout@v4
-      - name: CMake generation
-        run: cmake -S c -B c/build -DCMAKE_INSTALL_PREFIX=${{github.workspace}}/target
-      - name: CMake build / install
-        run: cmake --build c/build --target install
+      - uses: lukka/get-cmake@5f6e04f5267c8133f1273bf2103583fc72c46b17
+        with:
+          cmakeVersion: ${{ matrix.cmakeVersion }}
+          ninjaVersion: ${{ matrix.ninjaVersion }}
+      - if: matrix.os == 'macOS-latest'
+        name: Install dependencies on macOS
+        run: |
+          brew update
+          brew install tbb
+      - if: matrix.os == 'ubuntu-latest'
+        name: Install dependencies on Linux
+        run: |
+          sudo apt-get update
+          sudo apt-get install libtbb-dev libtbb12
+      - name: CMake generation, build, install
+        run: |
+          ${{ matrix.os != 'windows-latest' || '& "C:/Program Files/Microsoft Visual Studio/2022/Enterprise/Common7/Tools/Launch-VsDevShell.ps1" -Arch amd64 -SkipAutomaticLocation' }}
+          cmake -S c -B c/build -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=${{ github.workspace }}/target -DCMAKE_C_COMPILER=${{ matrix.toolchain.cc }} -DCMAKE_CXX_COMPILER=${{ matrix.toolchain.cxx }} -DBLAKE3_USE_TBB=${{ matrix.use_tbb }} -DBLAKE3_FETCH_TBB=${{ matrix.os == 'windows-latest' && 'YES' || 'NO' }} -DBLAKE3_EXAMPLES=ON
+          cmake --build c/build --target install
   cmake_3-9_build:
     name: CMake 3.9.6 ubuntu-latest
     runs-on: ubuntu-latest