marian/azure-pipelines.yml

# Azure pipelines for Marian NMT
#
# The pipeline need to be added manually to the repository, for example:
# 1. Go to Your repository > Pipelines, click "New pipeline"
# 2. Choose "Azure Repos Git" and a repository
# 3. Choose "Existing Azure Pipelines YAML file" and specify path to this file
# 4. "More actions" > "Save"

# The pipeline CI trigger is set on the branch master only and PR trigger on a
# (non-draft) pull request to any branch
trigger:
- master

pool:
  name: Azure Pipelines

variables:
  CUDA_PATH_WINDOWS: "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA"
  MKL_DIR: "$(Build.SourcesDirectory)/mkl"
  MKL_URL: "https://romang.blob.core.windows.net/mariandev/ci/mkl-2020.1-windows-static.zip"
  VCPKG_COMMIT: 6185aa76504a5025f36754324abf307cc776f3da
  VCPKG_DIR: "$(Build.SourcesDirectory)/vcpkg"
  VCPKG_PACKAGES: "protobuf"
  # The Visual Studio installation directory can be found using:
  #   pushd "C:\Program Files (x86)\Microsoft Visual Studio\Installer\"
  #   for /f "delims=" %%x in ('.\vswhere.exe -latest -property InstallationPath') do set VSPATH=%%x
  #   popd
  VS_PATH: "C:/Program Files (x86)/Microsoft Visual Studio/2019/Enterprise"

stages:
- stage: Build
  jobs:

  ######################################################################
  - job: Windows

    strategy:
      matrix:
        # Windows CPU-only build
        "CPU":
          cuda: false
          cuda_version: ""
        # Windows CPU+GPU build
        "CUDA 10.2":
          cuda: true
          cuda_version: 10.2

    pool:
      vmImage: windows-latest

    steps:
    - checkout: self
      submodules: true

    - pwsh: .\scripts\ci\install_cuda_windows.ps1 "$(cuda_version)"
      displayName: Install CUDA
      condition: eq(variables.cuda, true)

    - pwsh: |
        C:\msys64\usr\bin\wget.exe -nv $(MKL_URL) -O mkl.zip
        Expand-Archive -Force mkl.zip $(MKL_DIR)
      displayName: Download MKL

    ## Cache for vcpkg packages. It does not work yet properly due to linker errors after restoring it.
    #- task: Cache@2
    #  displayName: Cache
    #  inputs:
    #    # Change the first value (v0) to another value to clear the cache
    #    key: 'v0 | "$(VCPKG_PACKAGES)" | vcpkg | "$(Agent.OS)"'
    #    path: $(VCPKG_DIR)

    - pwsh: |
        git clone https://github.com/Microsoft/vcpkg.git $(VCPKG_DIR)
        cd $(VCPKG_DIR)
        git checkout $(VCPKG_COMMIT)
        pushd
        .\bootstrap-vcpkg.bat -disableMetrics
        popd
        # Install packages
        .\vcpkg.exe install --triplet x64-windows-static $(VCPKG_PACKAGES)
        # Clean to make the cache smaller
        Remove-Item $(VCPKG_DIR)\downloads  -Force -Recurse
        Remove-Item $(VCPKG_DIR)\buildtrees -Force -Recurse
      displayName: Prepare vcpkg

    - script: |
        :: Load VS environment
        call "$(VS_PATH)/VC/Auxiliary/Build/vcvarsall.bat" x64
        :: Create build directory
        mkdir build
        cd build
        :: Run CMake
        cmake .. -G Ninja ^
          -DCMAKE_BUILD_TYPE="Debug" ^
          -DCMAKE_C_COMPILER="cl.exe" ^
          -DCMAKE_CXX_COMPILER="cl.exe" ^
          -DCMAKE_MAKE_PROGRAM="ninja.exe" ^
          -DCMAKE_TOOLCHAIN_FILE="$(VCPKG_DIR)\scripts\buildsystems\vcpkg.cmake" ^
          -DVCPKG_TARGET_TRIPLET="x64-windows-static" ^
          ^
          -DBOOST_ROOT="$(BOOST_ROOT)" ^
          -DBOOST_INCLUDEDIR="$(BOOST_ROOT)/include" ^
          -DBOOST_LIBRARYDIR="$(BOOST_ROOT)/lib" ^
          -DOPENSSL_USE_STATIC_LIBS="TRUE" ^
          -DOPENSSL_MSVC_STATIC_RT="TRUE" ^
          ^
          -DCOMPILE_CPU="TRUE" ^
          -DCOMPILE_CUDA="$(cuda)" ^
          -DCOMPILE_EXAMPLES="FALSE" ^
          -DCOMPILE_SERVER="TRUE" ^
          -DCOMPILE_TESTS="TRUE" ^
          ^
          -DUSE_CUDNN="FALSE" ^
          -DUSE_FBGEMM="TRUE" ^
          -DUSE_MPI="FALSE" ^
          -DUSE_NCCL="FALSE" ^
          -DUSE_SENTENCEPIECE="TRUE" ^
          -DUSE_STATIC_LIBS="TRUE"
      displayName: Configure CMake
      env:
        # Set envvars so that CMake can find the installed packages
        MKLROOT: $(MKL_DIR)
        CUDA_PATH: $(CUDA_PATH_WINDOWS)/v$(cuda_version)
        # Boost is pre-installed on Azure/GitHub-hosted Windows runners
        # https://github.com/actions/virtual-environments/blob/main/images/win/Windows2019-Readme.md#boost
        BOOST_ROOT: $(BOOST_ROOT_1_72_0)

    - script: |
        call "$(VS_PATH)/VC/Auxiliary/Build/vcvarsall.bat" x64
        ninja
      displayName: Compile
      workingDirectory: build

    - script: |
        call "$(VS_PATH)/VC/Auxiliary/Build/vcvarsall.bat" x64
        ctest
      displayName: Run unit tests
      workingDirectory: build
      condition: eq(variables.cuda, false)

    # Note that versions from Marian executables will not be printed for CUDA builds
    - script: |
        .\marian.exe --version
        .\marian-decoder.exe --version
        .\marian-scorer.exe --version
        .\marian-server.exe --version
        .\spm_encode.exe --version
      displayName: Print versions
      workingDirectory: build

  ######################################################################
  - job: BuildUbuntu
    displayName: Ubuntu
    timeoutInMinutes: 90

    strategy:
      matrix:
        ################################################################
        # Ubuntu CPU-only build
        "CPU-only":
          image: ubuntu-18.04
          boost: true
          cpu: true
          gpu: false
          cuda: ""
          gcc: 7
          unit_tests: true
          examples: false
          static: true
        # Ubuntu GPU-only build
        "GPU-only":
          image: ubuntu-18.04
          boost: true
          cpu: false
          gpu: true
          cuda: 10.2
          gcc: 7
          unit_tests: false
          examples: false
          static: false
        ################################################################
        # Ubuntu 20.04 supports CUDA 11+
        #
        # CPU is disabled because FBGEMM + GCC 9+ do not compile on machines with
        # avx512_vnni, see https://github.com/marian-nmt/marian-dev/issues/709
        #
        "20.04 CUDA 11.1 gcc-9":
          image: ubuntu-20.04
          boost: false      # ubuntu-20.04 does not have Boost pre-installed yet
          cpu: false        # the used fbgemm does not compile with gcc 9+
          gpu: true
          cuda: 11.1
          gcc: 9
          unit_tests: false # disable unit tests to minimize compilation time
          examples: false   # disable examples to minimize compilation time
          static: false
        ################################################################
        # Ubuntu 18.04 supports CUDA 10.1+
        "18.04 CUDA 10.2 gcc-8":
          image: ubuntu-18.04
          boost: true
          cpu: true
          gpu: true
          cuda: 10.2
          gcc: 8
          unit_tests: true
          examples: true
          static: true
        ################################################################
        # Ubuntu 16.04 supports CUDA 8+
        "16.04 CUDA 9.2 gcc-7":
          image: ubuntu-16.04
          boost: true
          cpu: true
          gpu: true
          cuda: 9.2
          gcc: 7
          unit_tests: true
          examples: true
          static: false

    pool:
      vmImage: $(image)

    steps:
    - checkout: self
      submodules: true

      # The following packages are already installed on Azure-hosted runners: build-essential openssl libssl-dev
      # No need to install libprotobuf{17,10,9v5} on Ubuntu {20,18,16}.04 because it is installed together with libprotobuf-dev
    - bash: sudo apt-get install -y libgoogle-perftools-dev libprotobuf-dev protobuf-compiler
      displayName: Install packages

      # https://software.intel.com/content/www/us/en/develop/articles/installing-intel-free-libs-and-python-apt-repo.html
    - bash: |
        wget -qO- "https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS-2019.PUB" | sudo apt-key add -
        sudo sh -c "echo deb https://apt.repos.intel.com/mkl all main > /etc/apt/sources.list.d/intel-mkl.list"
        sudo apt-get update -o Dir::Etc::sourcelist="/etc/apt/sources.list.d/intel-mkl.list"
        sudo apt-get install -y --no-install-recommends intel-mkl-64bit-2020.0-088
      displayName: Install MKL
      condition: eq(variables.cpu, true)

      # The script simplifies installation of different versions of CUDA
    - bash: ./scripts/ci/install_cuda_ubuntu.sh $(cuda)
      displayName: Install CUDA
      condition: eq(variables.gpu, true)

    # Boost is already installed on Azure-hosted runners in a non-standard location
    # https://github.com/actions/virtual-environments/issues/687#issuecomment-610471671
    - bash: |
        mkdir -p build
        cd build
        CC=/usr/bin/gcc-$(gcc) CXX=/usr/bin/g++-$(gcc) CUDAHOSTCXX=/usr/bin/g++-$(gcc) \
        cmake .. \
          -DCOMPILE_CPU=$(cpu) \
          -DCOMPILE_CUDA=$(gpu) \
          -DCOMPILE_EXAMPLES=$(examples) \
          -DCOMPILE_SERVER=$(boost) \
          -DCOMPILE_TESTS=$(unit_tests) \
          -DUSE_FBGEMM=$(cpu) \
          -DUSE_SENTENCEPIECE=on \
          -DUSE_STATIC_LIBS=$(static) \
          -DBOOST_ROOT=$BOOST_ROOT_1_72_0 \
          -DBOOST_INCLUDEDIR=$BOOST_ROOT_1_72_0/include \
          -DBOOST_LIBRARYDIR=$BOOST_ROOT_1_72_0/lib \
          -DBoost_ARCHITECTURE=-x64 \
          -DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-$(cuda)
      displayName: Configure CMake

    # Clean build/src/ to safe disk space on Azure-hosted VMs and stay below the 10GB limit
    - bash: |
        make -j3
        rm -rf src/
      displayName: Compile
      workingDirectory: build

    # Unit tests are run only for CPU-only builds because Azure-hosted runners do not have GPUs
    # TODO: add an option to CMake to compile unit tests only for CPU
    - bash: make test
      displayName: Run unit tests
      workingDirectory: build
      condition: and(eq(variables.unit_tests, true), eq(variables.gpu, false))

    - bash: |
        ./marian --version
        ./marian-decoder --version
        ./marian-scorer --version
        ./spm_encode --version
      displayName: Print versions
      workingDirectory: build

  ######################################################################
  - job: BuildUbuntuMinimal
    displayName: Ubuntu CPU+GPU gcc-5 cmake 3.5

    pool:
      vmImage: ubuntu-16.04

    steps:
    - checkout: self
      submodules: true

    # The script simplifies installation of different versions of CUDA.
    # Ubuntu 16.04 on Azure-hosted VMs have GCC 5.5 as gcc-5, which is not compatible with CUDA 9.
    # Downgrading to GCC 5.4 (the default gcc on Ubuntu 16.04) would be more work...
    - bash: ./scripts/ci/install_cuda_ubuntu.sh "10.0"
      displayName: Install CUDA

    # CMake 3.5.1 is the minimum version supported
    - bash: |
        wget -nv https://cmake.org/files/v3.5/cmake-3.5.1-Linux-x86_64.tar.gz
        tar zxf cmake-3.5.1-Linux-x86_64.tar.gz
        ./cmake-3.5.1-Linux-x86_64/bin/cmake --version
      displayName: Download CMake

    # GCC 5 is the minimum version supported
    - bash: |
        /usr/bin/gcc-5 --version
        mkdir -p build
        cd build
        CC=/usr/bin/gcc-5 CXX=/usr/bin/g++-5 CUDAHOSTCXX=/usr/bin/g++-5 \
        ../cmake-3.5.1-Linux-x86_64/bin/cmake .. \
          -DCOMPILE_CPU=on \
          -DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-10.0
      displayName: Configure CMake

    - bash: make -j3
      displayName: Compile
      workingDirectory: build

    - bash: |
        ./marian --version
        ./marian-decoder --version
        ./marian-scorer --version
      displayName: Print versions
      workingDirectory: build

  ######################################################################
  - job: BuildMacOS
    displayName: macOS CPU clang

    pool:
      vmImage: macos-latest

    steps:
    - checkout: self
      submodules: true

    - bash: brew install openblas protobuf
      displayName: Install packages

    # Openblas location is exported explicitly because openblas is keg-only, which means it was not symlinked into /usr/local/.
    # CMake cannot find BLAS on Azure runners if Marian is being compiled statically, hence USE_STATIC_LIBS=off
    - bash: |
        export LDFLAGS="-L/usr/local/opt/openblas/lib"
        export CPPFLAGS="-I/usr/local/opt/openblas/include"
        mkdir -p build
        cd build
        cmake .. \
          -DCOMPILE_CPU=on \
          -DCOMPILE_CUDA=off \
          -DCOMPILE_EXAMPLES=on \
          -DCOMPILE_SERVER=on \
          -DCOMPILE_TESTS=on \
          -DUSE_FBGEMM=on \
          -DUSE_SENTENCEPIECE=on \
          -DUSE_STATIC_LIBS=off
      displayName: Configure CMake

    - bash: make -j2
      displayName: Compile
      workingDirectory: build

    - bash: make test
      displayName: Run unit tests
      workingDirectory: build

    - bash: |
        ./marian --version
        ./marian-decoder --version
        ./marian-scorer --version
        ./spm_encode --version
      displayName: Print versions
      workingDirectory: build

  ######################################################################
  - job: BuildInstall
    displayName: Linux CPU library install

    pool:
      vmImage: ubuntu-18.04

    steps:
    - checkout: self
      submodules: true

      # The following packages are already installed on Azure-hosted runners: build-essential openssl libssl-dev
      # No need to install libprotobuf{17,10,9v5} on Ubuntu {20,18,16}.04 because it is installed together with libprotobuf-dev
    - bash: sudo apt-get install -y libgoogle-perftools-dev libprotobuf-dev protobuf-compiler
      displayName: Install packages

      # https://software.intel.com/content/www/us/en/develop/articles/installing-intel-free-libs-and-python-apt-repo.html
    - bash: |
        wget -qO- "https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS-2019.PUB" | sudo apt-key add -
        sudo sh -c "echo deb https://apt.repos.intel.com/mkl all main > /etc/apt/sources.list.d/intel-mkl.list"
        sudo apt-get update -o Dir::Etc::sourcelist="/etc/apt/sources.list.d/intel-mkl.list"
        sudo apt-get install -y --no-install-recommends intel-mkl-64bit-2020.0-088
      displayName: Install MKL

    - bash: /usr/bin/gcc-7 --version
      displayName: Print GCC version

    - bash: |
        mkdir -p install
        mkdir -p build
        cd build
        CC=/usr/bin/gcc-7 CXX=/usr/bin/g++-7 \
        cmake .. \
          -DCMAKE_INSTALL_PREFIX=../install \
          -DCMAKE_BUILD_TYPE=slim \
          -DCOMPILE_LIBRARY_ONLY=on \
          -DCOMPILE_CUDA=off \
          -DGENERATE_MARIAN_INSTALL_TARGETS=on \
          -DUSE_FBGEMM=on \
          -DUSE_SENTENCEPIECE=on
      displayName: Configure CMake

    - bash: make -j3 install
      displayName: Compile & install
      workingDirectory: build

    - bash: |
        test -e lib/libmarian.a
        test -e lib/libfbgemm.a
        test -e lib/libsentencepiece.a
        ls -lah *
      displayName: Check targets
      workingDirectory: install