Run all backends compatible with CPU in CI

soda-inria · Sep 14, 2023 · 5d9c3a0 · 5d9c3a0
1 parent 2fabb18
commit 5d9c3a0
Showing 1 changed file with 90 additions and 12 deletions.
diff --git a/.github/workflows/test_kmeans_cpu_benchmarks.yaml b/.github/workflows/test_kmeans_cpu_benchmarks.yaml
@@ -9,6 +9,12 @@ env:
   _SCIKIT_LEARN_COMMIT: 2ccfc8c4bdf66db005d7681757b4145842944fb9
   _BENCHOPT_COMMIT: fc4d7fe6492dacca3d6cb5d512a3a9b3d5f84df5
   _SKLEARN_NUMBA_DPEX_COMMIT: e6e901757fced52abee8e1750a5f0004d35b0f3e
+  _SKLEARN_PYTORCH_ENGINE_COMMIT: e962dea262ba4d6e9873fced395e815a02a17fd0
+  _KMEANS_DPCPP_COMMIT: dceaa6418463c7921ba56dfcc80e3aeda0bcd5c3
+
+  _ONEAPI_INSTALLER_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/992857b9-624c-45de-9701-f6445d845359
+  _ONEAPI_INSTALL_BINARY_NAME: l_BaseKit_p_2023.2.0.49397.sh
+  _ONEAPI_COMPONENTS: "intel.oneapi.lin.dpcpp-cpp-compiler:intel.oneapi.lin.mkl.devel"
 
 
 # TODO: setup runners with compatible gpus and also test gpu benchmarks
@@ -19,39 +25,102 @@ jobs:
     container: jjerphan/numba_dpex_dev:latest
 
     steps:
-    - name: Install build pre-requisites
+    - name: Cache scikit-learn and benchopt wheels
+      id: wheels-cache
+      uses: actions/cache@v3
+      with:
+        path: |
+          scikit-learn/dist/
+          benchopt/dist/
+        key: sklearn-benchopt-${{ env._SCIKIT_LEARN_COMMIT}}
+
+    - if: ${{ steps.wheels-cache.outputs.cache-hit != 'true' }}
+      name: Install build pre-requisites
       run: |
         apt-get update --quiet
         apt-get install -y build-essential python3-dev
         pip install cython numpy scipy joblib threadpoolctl build wheel
 
-    - name: Build scikit-learn`
+    - if: ${{ steps.wheels-cache.outputs.cache-hit != 'true' }}
+      name: Build scikit-learn
       run: |
-        git clone https://github.com/fcharras/scikit-learn -b "feature/engine-api"
+        git clone https://github.com/fcharras/scikit-learn.git -b "feature/engine-api"
         cd scikit-learn
         git checkout $_SCIKIT_LEARN_COMMIT
         python setup.py bdist_wheel
 
     # NB: building benchopt from source will not be needed after next benchopt release
-    - name: Build benchopt
+    - if: ${{ steps.wheels-cache.outputs.cache-hit != 'true' }}
+      name: Build benchopt
       run: |
-        git clone https://github.com/benchopt/benchopt -b main
+        git clone https://github.com/benchopt/benchopt.git -b main
         cd benchopt
         git checkout $_BENCHOPT_COMMIT
         python -m build --wheel --outdir dist/
 
     - name: Upload wheels
       uses: actions/upload-artifact@v3
       with:
-        name: wheels
+        name: binaries
         path: |
           benchopt/dist/
           scikit-learn/dist/
         retention-days: 14
 
-  run_sklearn_numba_dpex_benchmarks:
-    name: Run sklearn_numba_dpex benchmarks on CPU
-    needs: build_scikit_learn_and_benchopt
+  build_kmeans_dpcpp:
+    name: Build kmeans_dpcpp
+    runs-on: ubuntu-latest
+    container: jjerphan/numba_dpex_dev:latest
+
+    steps:
+    - name: Cache kmeans_dpcpp build
+      id: kmeans-dpcpp-cache
+      uses: actions/cache@v3
+      with:
+        path: kmeans_dpcpp/
+        key: kmeans-dpcpp-${{ env._KMEANS_DPCPP_COMMIT}}
+
+    - if: ${{ steps.kmeans-dpcpp-cache.outputs.cache-hit != 'true' }}
+      name: Install build pre-requisites
+      run: |
+        apt-get update --quiet
+        apt-get install -y --no-install-recommends gawk
+        pip install cmake ninja packaging setuptools distro scikit-build numpy
+
+    - if: ${{ steps.kmeans-dpcpp-cache.outputs.cache-hit != 'true' }}
+      name: Install necessary oneapi basekit components
+      run: |
+        mkdir -p ./opt/intel/oneapi $RUNNER_TEMP/intel/cache $RUNNER_TEMP/intel/download $RUNNER_TEMP/intel/log
+        wget -P $RUNNER_TEMP/intel/download $_ONEAPI_INSTALLER_URL/$_ONEAPI_INSTALL_BINARY_NAME
+        chmod +x $RUNNER_TEMP/intel/download/$_ONEAPI_INSTALL_BINARY_NAME
+        $RUNNER_TEMP/intel/download/$_ONEAPI_INSTALL_BINARY_NAME -a -s --eula accept \
+          --action install --components $_ONEAPI_COMPONENTS  \
+          --install-dir ./opt/intel/oneapi --log-dir $RUNNER_TEMP/intel/log \
+          --download-cache $RUNNER_TEMP/intel/cache \
+
+    # HACK: `kmeans_dpcpp` packaging is broken, just run `develop` to build the
+    # C++ extensions, then later on add it explicitly to PYTHONPATH
+    - if: ${{ steps.kmeans-dpcpp-cache.outputs.cache-hit != 'true' }}
+      name: build kmeans_dpcpp
+      shell: bash
+      run: |
+        source ./opt/intel/oneapi/setvars.sh
+        git clone https://github.com/oleksandr-pavlyk/kmeans_dpcpp.git -b main
+        cd kmeans_dpcpp
+        git checkout $_KMEANS_DPCPP_COMMIT
+        python setup.py develop -- -DCMAKE_CXX_COMPILER:PATH=$(which icpx) -DDPCTL_MODULE_PATH=$(python -m dpctl --cmakedir)
+
+    - name: Upload kmeans_dpcpp build files
+      uses: actions/upload-artifact@v3
+      with:
+        name: binaries
+        path: kmeans_dpcpp*/
+        retention-days: 14
+
+
+  run_benchmarks_on_cpu:
+    name: Run benchmarks on cpu
+    needs: [build_scikit_learn_and_benchopt, build_kmeans_dpcpp]
     runs-on: ubuntu-latest
     container: jjerphan/numba_dpex_dev:latest
 
@@ -62,21 +131,30 @@ jobs:
     - name: download scikit-learn and benchopt wheels
       uses: actions/download-artifact@v3
       with:
-        name: wheels
+        name: binaries
 
+    # TODO: instead of installing all dependencies of all solvers in the same
+    # environment, might be more cautious to have one solver per environment ?
     - name: install all dependencies
       run: |
-        git clone https://github.com/soda-inria/sklearn-numba-dpex -b main
+        git clone https://github.com/soda-inria/sklearn-numba-dpex.git -b main
         cd sklearn-numba-dpex
         git checkout $_SKLEARN_NUMBA_DPEX_COMMIT
+        cd ../
+        git clone https://github.com/soda-inria/sklearn-pytorch-engine.git -b main
+        cd sklearn-pytorch-engine
+        git checkout $_SKLEARN_PYTORCH_ENGINE_COMMIT
+        cd ../
         # TODO: it seems that non-editable install of sklearn-numba-dpex doesn't work ?
         # Fix and update the install command here to avoid having to git clone manually
-        pip install ../scikit-learn/dist/*.whl ../benchopt/dist/*.whl -e .
+        pip install ./scikit-learn/dist/*.whl ./benchopt/dist/*.whl scikit-learn-intelex -e ./sklearn-numba-dpex/ -e ./sklearn-pytorch-engine
+        pip install torch --index-url https://download.pytorch.org/whl/cpu
 
     - name: Check oneapi default device
       run: python -c "import dpctl; dpctl.select_default_device().print_device_info()"
 
     - name: Run k-means benchmarks
       run: |
+        export PYTHONPATH=$PYTHONPATH:$(realpath ./kmeans_dpcpp/)
         cd benchmarks/kmeans
         benchopt run --no-plot -l -d simulated[n_samples=1000,n_features=14]