add CI test for yitian (#1016)

.github/workflows/pytorch113_yitian.yml

@@ -0,0 +1,75 @@
+name: pytorch1_13-yitian
+on:
+  push:
+    branches: [ main ]
+  pull_request:
+    branches: [ main ]
+  schedule:
+    - cron: '0 17 * * *' # 1:00 am UTC+8:00
+  workflow_dispatch:
+
+jobs:
+  build:
+    if: github.repository == 'alibaba/BladeDISC'
+    # The type of runner that the job will run on
+    runs-on: [self-hosted, yitian-ci]
+    steps:
+    - name: Checkout
+      uses: actions/[email protected]
+    - name: Build Dev Docker
+      run: |
+        set -e
+        source $HOME/.cache/proxy_config
+        git submodule sync
+        git submodule update --depth=1 --init --recursive
+        cp /etc/apt/sources.list .
+        docker build -t disc-dev-cpu-yitian \
+          --build-arg BASEIMAGE=ubuntu:20.04 \
+          -f docker/dev/Dockerfile.aarch64 .
+    - name: Build And Test DISC
+      run: |
+        set -e
+        docker run --rm -t --user $(id -u):$(id -g) \
+          -v $HOME/.cache:$HOME/.cache \
+          -v /etc/passwd:/etc/passwd:ro \
+          -v /etc/group:/etc/group:ro \
+          -v /etc/hosts:/etc/hosts:ro \
+          -v $PWD:/disc \
+          -e GITHUB_WORKFLOW=$GITHUB_WORKFLOW \
+          -e TORCH_BLADE_BUILD_WITH_CUDA_SUPPORT=OFF \
+          -e TORCH_BLADE_CI_BUILD_TORCH_VERSION=1.13.1+aarch64 \
+          -e TORCH_BLADE_BUILD_ON_YITIAN=ON \
+          -w /disc \
+          disc-dev-cpu-yitian \
+          bash ./scripts/ci/test_pytorch_blade.sh
+    - name: Deploy Dev Docker Image
+      if: github.event.ref == 'refs/heads/main'
+      env:
+        ALIYUN_DOCKER_USERNAME: ${{ secrets.ALIYUN_DOCKER_USERNAME }}
+        ALIYUN_DOCKER_PASSWORD: ${{ secrets.ALIYUN_DOCKER_PASSWORD }}
+        DOCKER_USERNAME: ${{ secrets.DOCKER_USERNAME }}
+        DOCKER_PASSWORD: ${{ secrets.DOCKER_PASSWORD }}
+        GITHUB_PULL_REQUEST: ${{ github.event.number }}
+        LOCAL_DEV_DOCKER: disc-dev-cpu-yitian
+        REMOTE_DEV_DOCKER: bladedisc:latest-devel-cpu-yitian
+      run: |
+        set -e
+        echo "Try to deploy runtime docker image..."
+        source $HOME/.cache/proxy_config
+        bash ./scripts/ci/deploy_wrapper.sh
+    - name: Deploy Runtime Docker Image
+      if: github.event.ref == 'refs/heads/main'
+      env:
+        ALIYUN_DOCKER_USERNAME: ${{ secrets.ALIYUN_DOCKER_USERNAME }}
+        ALIYUN_DOCKER_PASSWORD: ${{ secrets.ALIYUN_DOCKER_PASSWORD }}
+        DOCKER_USERNAME: ${{ secrets.DOCKER_USERNAME }}
+        DOCKER_PASSWORD: ${{ secrets.DOCKER_PASSWORD }}
+        GITHUB_PULL_REQUEST: ${{ github.event.number }}
+        RUNTIME_BASEIMAGE: bladedisc/bladedisc:latest-devel-cpu-yitian
+        RUNTIME_DOCKER_FILE: docker/runtime/Dockerfile.pytorch.aarch64
+        REMOTE_RUNTIME_DOCKER: bladedisc:latest-runtime-torch1.13.1-cpu-yitian
+      run: |
+        set -e
+        echo "Try to deploy runtime docker image..."
+        source $HOME/.cache/proxy_config
+        bash ./scripts/ci/deploy_wrapper.sh

.github/workflows/tf280_yitian.yml

@@ -0,0 +1,74 @@
+name: tf2_8-yitian
+on:
+  push:
+    branches: [ main ]
+  pull_request:
+    branches: [ main ]
+  schedule:
+    - cron: '0 17 * * *' # 1:00 am UTC+8:00
+  workflow_dispatch:
+
+jobs:
+  build:
+    if: github.repository == 'alibaba/BladeDISC'
+    # The type of runner that the job will run on
+    runs-on: [self-hosted, yitian-ci]
+    steps:
+    - name: Checkout
+      uses: actions/[email protected]
+    - name: Build Dev Docker
+      run: |
+        set -e
+        source $HOME/.cache/proxy_config
+        git submodule sync
+        git submodule update --depth=1 --init --recursive
+        cp /etc/apt/sources.list .
+        docker build -t disc-dev-cpu-yitian \
+          --build-arg BASEIMAGE=ubuntu:20.04 \
+          --build-arg DISC_HOST_TF_VERSION="tensorflow-aarch64==2.8" \
+          -f docker/dev/Dockerfile.aarch64 .
+    - name: Build And Test DISC
+      run: |
+        set -e
+        docker run --rm -t --user $(id -u):$(id -g) \
+          -v $HOME/.cache:$HOME/.cache \
+          -v /etc/passwd:/etc/passwd:ro \
+          -v /etc/group:/etc/group:ro \
+          -v /etc/hosts:/etc/hosts:ro \
+          -v $PWD:/disc \
+          -e GITHUB_WORKFLOW=$GITHUB_WORKFLOW \
+          -h=`hostname` \
+          -w /disc \
+          disc-dev-cpu-yitian \
+          bash ./scripts/ci/build_and_test.sh --cpu-only --target_cpu_arch="armv8.6-a-sve2"
+    - name: Deploy Dev Docker Image
+      if: github.event.ref == 'refs/heads/main'
+      env:
+        ALIYUN_DOCKER_USERNAME: ${{ secrets.ALIYUN_DOCKER_USERNAME }}
+        ALIYUN_DOCKER_PASSWORD: ${{ secrets.ALIYUN_DOCKER_PASSWORD }}
+        DOCKER_USERNAME: ${{ secrets.DOCKER_USERNAME }}
+        DOCKER_PASSWORD: ${{ secrets.DOCKER_PASSWORD }}
+        GITHUB_PULL_REQUEST: ${{ github.event.number }}
+        LOCAL_DEV_DOCKER: disc-dev-cpu-yitian
+        REMOTE_DEV_DOCKER: bladedisc:latest-devel-cpu-yitian
+      run: |
+        set -e
+        echo "Try to deploy runtime docker image..."
+        source $HOME/.cache/proxy_config
+        bash ./scripts/ci/deploy_wrapper.sh
+    - name: Deploy Runtime Docker Image
+      if: github.event.ref == 'refs/heads/main'
+      env:
+        ALIYUN_DOCKER_USERNAME: ${{ secrets.ALIYUN_DOCKER_USERNAME }}
+        ALIYUN_DOCKER_PASSWORD: ${{ secrets.ALIYUN_DOCKER_PASSWORD }}
+        DOCKER_USERNAME: ${{ secrets.DOCKER_USERNAME }}
+        DOCKER_PASSWORD: ${{ secrets.DOCKER_PASSWORD }}
+        GITHUB_PULL_REQUEST: ${{ github.event.number }}
+        RUNTIME_BASEIMAGE: bladedisc/bladedisc:latest-devel-cpu-yitian
+        RUNTIME_DOCKER_FILE: docker/runtime/Dockerfile.tf.aarch64
+        REMOTE_RUNTIME_DOCKER: bladedisc:latest-runtime-tensorflow2.8-cpu-yitian
+      run: |
+        set -e
+        echo "Try to deploy runtime docker image..."
+        source $HOME/.cache/proxy_config
+        bash ./scripts/ci/deploy_wrapper.sh

pytorch_blade/scripts/build_pytorch_blade.sh

@@ -45,6 +45,10 @@ function ci_build() {
       COMMON_SETUP_ARGS="--platform_alibaba"
     fi
 
+    if [ "$TORCH_BLADE_BUILD_ON_YITIAN" = "ON" ]; then
+      COMMON_SETUP_ARGS+=" --target_cpu_arch='armv8.6-a-sve2'"
+    fi
+
     if [ "$TORCH_BLADE_BUILD_WITH_CUDA_SUPPORT" = "ON"  ]; then
       export TORCH_BLADE_BUILD_TENSORRT=${TORCH_BLADE_BUILD_TENSORRT:-ON}
       export TORCH_BLADE_BUILD_TENSORRT_STATIC=${TORCH_BLADE_BUILD_TENSORRT_STATIC:-OFF}

pytorch_blade/tests/disc/ops/test_conv_ops.py

@@ -13,8 +13,9 @@
 import torch_blade
 import unittest
 
-from tests.disc.testing_base import DiscTestCase, skipTorchLE
+from tests.disc.testing_base import DiscTestCase, skipTorchLE, skipIfOnYitian
 
+@skipIfOnYitian()
 @skipTorchLE("1.6.1")
 class TestMlirConvolution(DiscTestCase):
     def _test_conv(self, conv_func, inp_test_data=None):

pytorch_blade/tests/disc/testing_base.py

@@ -25,6 +25,8 @@
 from torch_blade.quantization import is_available as is_quantization_available
 from torch_blade.testing.common_utils import TestCase
 
+def skipIfOnYitian():
+    return unittest.skipIf(os.popen("lscpu").read().find("svebf16") != -1, "Yitian bug was not fix")
 
 def skipIfNoDISC():
     return unittest.skipIf(not is_available(), "DISC support was not built")

scripts/ci/build_and_test.sh

@@ -28,12 +28,12 @@ fi
 # cleanup build cache
 (rm -rf build \
   && rm -rf tao/build \
-  && cd tao && bazel clean --expunge && cd .. \
-  && cd tf_community && bazel clean --expunge)
+  && cd tao && bazel clean --expunge && cd ..\
+  && cd tao_compiler && bazel clean --expunge)
 
-python ${ENTRY} ${VENV_PATH} -s configure --bridge-gcc default --compiler-gcc default ${CPU_ONLY} ${ROCM} ${DCU} ${ROCM_PATH}
-python ${ENTRY} ${VENV_PATH} -s build_tao_bridge ${CPU_ONLY} ${ROCM} ${DCU} ${ROCM_PATH}
-python ${ENTRY} ${VENV_PATH} -s build_tao_compiler ${CPU_ONLY} ${ROCM} ${DCU} ${ROCM_PATH}
+python ${ENTRY} ${VENV_PATH} -s configure --bridge-gcc default --compiler-gcc default ${CPU_ONLY} ${ROCM} ${DCU} ${ROCM_PATH} ${TARGET_CPU_ARCH}
+python ${ENTRY} ${VENV_PATH} -s build_tao_bridge ${CPU_ONLY} ${ROCM} ${DCU} ${ROCM_PATH} ${TARGET_CPU_ARCH}
+python ${ENTRY} ${VENV_PATH} -s build_tao_compiler ${CPU_ONLY} ${ROCM} ${DCU} ${ROCM_PATH} ${TARGET_CPU_ARCH}
 if [[ -z "$ROCM" ]] && [[ -z "$DCU" ]]; then
   python ${ENTRY} ${VENV_PATH} -s test_tao_bridge_cpp ${CPU_ONLY} ${ROCM} ${DCU} ${ROCM_PATH}
   python ${ENTRY} ${VENV_PATH} -s test_tao_bridge_py ${CPU_ONLY} ${ROCM} ${DCU} ${ROCM_PATH}

scripts/ci/parse_args.sh

@@ -26,6 +26,10 @@ while [[ $# -gt 0 ]]; do
       VENV_PATH="$2"
       shift 2
       ;;
+    --target_cpu_arch)
+      TARGET_CPU_ARCH="--target_cpu_arch $2"
+      shift 2
+      ;;
     --rocm)
       ROCM="--rocm"
       shift

scripts/ci/test_pytorch_blade.sh

@@ -45,7 +45,7 @@ if [[ -f ~/.cache/proxy_config ]]; then
 fi
 
 # cleanup build cache
-(cd tf_community && bazel clean --expunge)
+(cd tao_compiler && bazel clean --expunge)
 
 # note(yancey.yx): using virtualenv to avoid permission issue on workflow actions CI,
 if [ $TORCH_BLADE_CI_BUILD_TORCH_VERSION = "ngc" ]; then

scripts/python/common_setup.py

@@ -261,11 +261,17 @@ def mkl_install_dir(root):
 def acl_root_dir(root):
     return os.path.join(mkldnn_build_dir(root), 'acl', 'ComputeLibrary')
 
+def extra_acl_patch_dir(root):
+    if root is None:
+        root = get_source_root_dir()
+    return os.path.join(root, "third_party", "bazel", "acl")
+
 def config_mkldnn(root, args):
     build_dir = mkldnn_build_dir(root)
     ensure_empty_dir(build_dir, clear_hidden=False)
     mkl_dir = mkl_install_dir(root)
     acl_dir = acl_root_dir(root)
+    acl_patch_dir = extra_acl_patch_dir(root)
     ensure_empty_dir(mkl_dir, clear_hidden=False)
     ensure_empty_dir(acl_dir, clear_hidden=False)
     if args.x86:
@@ -293,11 +299,19 @@ def config_mkldnn(root, args):
               ACL_DIR={}
               git clone --branch v22.02 --depth 1 $ACL_REPO $ACL_DIR
               cd $ACL_DIR
+              EXTRA_ACL_PATCH_DIR={}
+              for file in $EXTRA_ACL_PATCH_DIR/acl_*.patch
+              do
+                  if [[ $file == *makefile* ]]; then
+                      continue
+                  fi
+                  patch -p1 < $file
+              done
 
               scons --silent $MAKE_NP Werror=0 debug=0 neon=1 opencl=0 openmp=1 embed_kernels=0 os=linux arch={} build=native extra_cxx_flags="-fPIC"
 
               exit $?
-            '''.format(acl_dir, arch)
+            '''.format(acl_dir, acl_patch_dir, arch)
             execute(cmd)
             # a workaround for static linking
             execute('rm -f build/*.so')