paddle-build.yml

name: paddle-build
on:
  push:
    branches:
      - feature/3.1
      - master
    paths:
      - .github/workflows/paddle-build.yml
      - build/capi.patch
  workflow_dispatch:

env:
  PADDLE_VERSION: "3.1.0"
  PADDLE_BRANCH: "v3.1.0"

jobs:
  build-all:
    runs-on: ${{ matrix.os }}

    strategy:
      matrix:
        include:
          - { os: windows-2022,     name: openblas_noavx,       cmake_flags: "-DWITH_AVX=OFF -DWITH_MKL=OFF -DWITH_ONNXRUNTIME=ON" }
          - { os: windows-2022,     name: openblas,             cmake_flags: "-DWITH_MKL=OFF -DWITH_ONNXRUNTIME=ON" }
          - { os: windows-2022,     name: mkldnn,               cmake_flags: "-DWITH_MKL=ON -DWITH_ONNXRUNTIME=ON" }
          - { os: windows-2022,     name: cu118_cudnn89_sm61,   cuda: 11.8.0, cudnn: 8.9.7.29, cmake_flags: "-DWITH_GPU=ON -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN=\"61\"  -DWITH_ONNXRUNTIME=ON"   }
          - { os: windows-2022,     name: cu118_cudnn89_sm75,   cuda: 11.8.0, cudnn: 8.9.7.29, cmake_flags: "-DWITH_GPU=ON -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN=\"75\"  -DWITH_ONNXRUNTIME=ON"   }
          - { os: windows-2022,     name: cu118_cudnn89_sm86,   cuda: 11.8.0, cudnn: 8.9.7.29, cmake_flags: "-DWITH_GPU=ON -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN=\"86\"  -DWITH_ONNXRUNTIME=ON"   }
          - { os: windows-2022,     name: cu118_cudnn89_sm89,   cuda: 11.8.0, cudnn: 8.9.7.29, cmake_flags: "-DWITH_GPU=ON -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN=\"89\"  -DWITH_ONNXRUNTIME=ON"   }
          - { os: windows-2022,     name: cu126_cudnn95_sm61,   cuda: 12.6.3, cudnn: 9.5.1.17, cmake_flags: "-DWITH_GPU=ON -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN=\"61\"  -DWITH_ONNXRUNTIME=ON"   }
          - { os: windows-2022,     name: cu126_cudnn95_sm75,   cuda: 12.6.3, cudnn: 9.5.1.17, cmake_flags: "-DWITH_GPU=ON -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN=\"75\"  -DWITH_ONNXRUNTIME=ON"   }
          - { os: windows-2022,     name: cu126_cudnn95_sm86,   cuda: 12.6.3, cudnn: 9.5.1.17, cmake_flags: "-DWITH_GPU=ON -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN=\"86\"  -DWITH_ONNXRUNTIME=ON"   }
          - { os: windows-2022,     name: cu126_cudnn95_sm89,   cuda: 12.6.3, cudnn: 9.5.1.17, cmake_flags: "-DWITH_GPU=ON -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN=\"89\"  -DWITH_ONNXRUNTIME=ON"   }
          - { os: windows-2022,     name: cu129_cudnn910_sm61,  cuda: 12.9.0, cudnn: 9.10.1.4, cmake_flags: "-DWITH_GPU=ON -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN=\"61\"  -DWITH_ONNXRUNTIME=ON"   }
          - { os: windows-2022,     name: cu129_cudnn910_sm75,  cuda: 12.9.0, cudnn: 9.10.1.4, cmake_flags: "-DWITH_GPU=ON -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN=\"75\"  -DWITH_ONNXRUNTIME=ON"   }
          - { os: windows-2022,     name: cu129_cudnn910_sm86,  cuda: 12.9.0, cudnn: 9.10.1.4, cmake_flags: "-DWITH_GPU=ON -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN=\"86\"  -DWITH_ONNXRUNTIME=ON"   }
          - { os: windows-2022,     name: cu129_cudnn910_sm89,  cuda: 12.9.0, cudnn: 9.10.1.4, cmake_flags: "-DWITH_GPU=ON -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN=\"89\"  -DWITH_ONNXRUNTIME=ON"   }
          - { os: windows-2022,     name: cu129_cudnn910_sm120, cuda: 12.9.0, cudnn: 9.10.1.4, cmake_flags: "-DWITH_GPU=ON -DCUDA_ARCH_NAME=Manual -DCUDA_ARCH_BIN=\"120\" -DWITH_ONNXRUNTIME=ON" }
          - { os: ubuntu-22.04,     name: openblas,             cmake_flags: "-DWITH_MKL=OFF -DWITH_ONNXRUNTIME=ON" }
          - { os: ubuntu-22.04,     name: mkldnn,               cmake_flags: "-DWITH_MKL=ON -DWITH_ONNXRUNTIME=ON" }
          - { os: ubuntu-22.04,     name: openvino,             cmake_flags: "-DWITH_MKL=ON -DWITH_OPENVINO=ON -DWITH_ONNXRUNTIME=ON" }
          - { os: ubuntu-22.04-arm, name: openblas,             cmake_flags: "-DWITH_ARM=ON -DCMAKE_CXX_FLAGS=\"-Wno-error=class-memaccess\"" }
          - { os: macos-13,         name: openblas,             cmake_flags: "-DWITH_ONNXRUNTIME=ON" }
          - { os: macos-15,         name: openblas,             cmake_flags: "-DWITH_ARM=ON" }
    steps:
    - name: Setup dependencies
      shell: bash
      run: |
        if [ "${{ runner.os }}" == "macOS" ]; then
          pip install numpy protobuf wheel ninja --break-system-packages
        else
          pip install numpy protobuf wheel ninja
        fi

    - uses: Jimver/cuda-toolkit@v0.2.24
      if: ${{ matrix.cuda != '' }}
      id: cuda-toolkit
      with:
        method: network
        cuda: ${{ matrix.cuda }}
        # Windows sub-packages are kind of different, so we need to handle them separately
        # https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html
        sub-packages: ${{ matrix.os == 'ubuntu-22.04' && 
          '["nvcc"]' || 
          '["nvcc", "cublas", "cublas_dev", "cudart", "cufft", "cufft_dev", "curand", "curand_dev", 
            "cusparse", "cusparse_dev", "cusolver", "cusolver_dev", "nvrtc", "nvrtc_dev", 
            "cuda_profiler_api", "nvjpeg", "nvjpeg_dev", "thrust"]' 
          }}
        non-cuda-sub-packages: ${{ matrix.os == 'ubuntu-22.04' && 
          '["libcublas", "libcublas-dev", "libcufft", "libcufft-dev", "libcurand", "libcurand-dev", 
            "libcusparse", "libcusparse-dev", "libcusolver", "libcusolver-dev", "libnvrtc", 
            "libnvrtc-dev", "profiler_api", "libthrust"]' || 
          '[]' }}
        log-file-suffix: ${{ runner.os }}.txt
    
    - name: Download cuDNN
      if: ${{ matrix.cudnn != '' }}
      shell: pwsh
      run: |
        $cudaMajor = '${{ matrix.cuda }}'.Split('.')[0]
        echo "Downloading cuDNN for CUDA $cudaMajor and cuDNN ${{ matrix.cudnn }}..."
        if ("${{ runner.os }}" -eq "Linux") {
          Invoke-WebRequest -Uri https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/linux-x86_64/cudnn-linux-x86_64-${{ matrix.cudnn }}_cuda$cudaMajor-archive.tar.xz -OutFile cudnn.tar.xz
          mkdir cudnn && tar -xf cudnn.tar.xz --strip-components=1 -C cudnn

          sudo cp cudnn/include/* ${{steps.cuda-toolkit.outputs.CUDA_PATH}}/include/
          sudo cp cudnn/lib/* ${{steps.cuda-toolkit.outputs.CUDA_PATH}}/lib64/
        } elseif ("${{ runner.os }}" -eq "Windows") {
          Invoke-WebRequest -Uri https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/windows-x86_64/cudnn-windows-x86_64-${{ matrix.cudnn }}_cuda$cudaMajor-archive.zip -OutFile cudnn.zip
          mkdir cudnn && tar -xf cudnn.zip --strip-components=1 -C cudnn

          cp -Recurse -Force cudnn\* "${{ steps.cuda-toolkit.outputs.CUDA_PATH }}"
        }
    
        echo "::group:: cuDNN files"
        Get-ChildItem -Recurse cudnn
        echo "::endgroup::"

    - name: Checkout Paddle repo
      uses: actions/checkout@v4
      with:
        repository: PaddlePaddle/Paddle
        ref: ${{ env.PADDLE_BRANCH }}
        path: paddle-src
        fetch-depth: 1

    - uses: ilammy/msvc-dev-cmd@v1
      if: ${{ runner.os == 'Windows' }}
      with:
        toolset: '14.29'

    - name: Apply patches
      shell: pwsh
      run: |
        cd ./paddle-src
        Invoke-WebRequest -Uri https://raw.githubusercontent.com/sdcb/PaddleSharp/refs/heads/feature/3.1/build/capi.patch -OutFile capi.patch
        git apply --ignore-whitespace capi.patch
        rm capi.patch

    - name: Configure with CMake
      shell: pwsh
      run: |
        mkdir ./paddle-src/build
        mkdir ./paddle-src/build/paddle/phi/ops/yaml/inconsistent
        mkdir ./paddle-src/build/paddle/phi/ops/yaml/legacy

        cd ./paddle-src/build

        cmake .. -GNinja `
          ${{ matrix.cmake_flags }} `
          -DON_INFER=ON `
          -DWITH_PYTHON=OFF `
          -DWITH_UNITY_BUILD=ON

    - name: Build with Ninja
      run: |
        cd ./paddle-src/build
        ninja all

    - name: Show build files
      shell: bash
      run: |
        ls -lR ./paddle-src/build

    - name: Upload paddle_inference_install_dir
      uses: actions/upload-artifact@v4
      with:
        name: ${{ matrix.os }}_cpp_${{ matrix.name }}
        path: ./paddle-src/build/paddle_inference_install_dir/*
    
    - name: Patch RPATH / install-name for Paddle
      if: ${{ runner.os != 'Windows' }}
      shell: bash
      run: |    
        case "${{ runner.os }}" in    
          Linux)
            echo "Before:"
            LIB=./paddle-src/build/paddle_inference_c_install_dir/paddle/lib/libpaddle_inference_c.so
            patchelf --print-rpath $LIB || true    
            patchelf --set-rpath '$ORIGIN' $LIB
    
            echo "After :"
            patchelf --print-rpath $LIB

            if find . -type f -name 'libmklml_intel.so' -print -quit | grep -q .; then
              echo "libmklml_intel.so found, patching libpaddle_inference_c.so..."
              patchelf --add-needed libmklml_intel.so ./paddle-src/build/paddle_inference_c_install_dir/paddle/lib/libpaddle_inference_c.so
            else
              echo "libmklml_intel.so not found, skip patchelf."
            fi
            ;;
    
          macOS)
            dylib=./paddle-src/build/paddle_inference_c_install_dir/paddle/lib/libpaddle_inference_c.dylib
    
            echo "Before:"
            otool -l "$dylib" | grep -A2 LC_RPATH || true
    
            install_name_tool -id @rpath/$(basename "$dylib") "$dylib"
            install_name_tool -add_rpath @loader_path "$dylib" 2>/dev/null || true
    
            echo "After :"
            otool -l "$dylib" | grep -A2 LC_RPATH
            ;;
    
          *)
            echo "no rpath patch needed for ${{ matrix.rid }}"
            ;;
        esac
    
        echo "RPATH / LC_RPATH patched successfully."

    - name: Post process paddle_inference_install_dir
      shell: bash
      run: |
        cp ./paddle-src/build/paddle_inference_install_dir/paddle/lib/*common.* ./paddle-src/build/paddle_inference_c_install_dir/paddle/lib/
        cp ./paddle-src/build/paddle_inference_install_dir/paddle/lib/*phi* ./paddle-src/build/paddle_inference_c_install_dir/paddle/lib/ || true
        ls -lR ./paddle-src/build/paddle_inference_install_dir

    - name: Upload paddle_inference_c_install_dir
      uses: actions/upload-artifact@v4
      with:
        name: ${{ matrix.os }}_c_${{ matrix.name }}
        path: ./paddle-src/build/paddle_inference_c_install_dir/*