Run C++ tests on CI with run_test.py (pytorch#99956)

huydhn · pytorchmergebot · commit 35834a405ca2 · 2023-05-09T21:24:12.000Z
After pytorch#99559, we can now run C++ test with `run_test.py`. Although advance features such as `--import-slow-tests` and `--import-disabled-tests` won't work for now, there will still be a gain in reliability and performance as C++ can now be retried and run in parallel. This covers all C++ tests in the CI including aten, libtorch, and Vulkan C++ tests across all platforms Linux, Windows, MacOS. Notes: * To support C++ test discovery, the env variable `CPP_TESTS_DIR` can be set to where the C++ test binaries is located * Support pytest -k argument via run_test as this is used by pytest-cpp to replace `--gtest-filter` * The XML output is in pytest format, but it's ok now because we don't have slow test or flaky test support for C++ test yet * ~~I need to figure out why conftest.py doesn't work when I invoke pytest directly for C++ test, so `--sc` is not available for C++ tests at the moment. Proper pytest plugin like stepwise works fine though. I'll investigate and fix it in a separate PR~~ Found the cause, `conftest.py` is per directory and needs to be in any arbitrary directory that holds C++ test * Two tests `test_api` and `test_tensorexpr` timed out on ASAN, I suspect that ASAN is now used on top of the python executable, which is slower than running native C++ code. IMO, it's ok to run these tests as before on ASAN for now Pull Request resolved: pytorch#99956 Approved by: https://github.com/clee2000, https://github.com/ZainRizvi
diff --git a/.ci/pytorch/macos-test.sh b/.ci/pytorch/macos-test.sh
@@ -71,13 +71,14 @@ test_libtorch() {
     VERBOSE=1 DEBUG=1 python "$BUILD_LIBTORCH_PY"
     popd
 
-    python tools/download_mnist.py --quiet -d test/cpp/api/mnist
+    MNIST_DIR="${PWD}/test/cpp/api/mnist"
+    python tools/download_mnist.py --quiet -d "${MNIST_DIR}"
 
     # Unfortunately it seems like the test can't load from miniconda3
     # without these paths being set
     export DYLD_LIBRARY_PATH="$DYLD_LIBRARY_PATH:$PWD/miniconda3/lib"
     export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:$PWD/miniconda3/lib"
-    TORCH_CPP_TEST_MNIST_PATH="test/cpp/api/mnist" "$CPP_BUILD"/caffe2/bin/test_api
+    TORCH_CPP_TEST_MNIST_PATH="${MNIST_DIR}" CPP_TESTS_DIR="${CPP_BUILD}/caffe2/bin" python test/run_test.py --cpp --verbose -i cpp/test_api
 
     assert_git_not_dirty
   fi
diff --git a/.ci/pytorch/test.sh b/.ci/pytorch/test.sh
@@ -271,7 +271,7 @@ test_inductor() {
   # docker build uses bdist_wheel which does not work with test_aot_inductor
   # TODO: need a faster way to build
   BUILD_AOT_INDUCTOR_TEST=1 python setup.py develop
-  LD_LIBRARY_PATH="$TORCH_LIB_DIR $TORCH_BIN_DIR"/test_aot_inductor
+  CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="${TORCH_LIB_DIR}" python test/run_test.py --cpp --verbose -i cpp/test_aot_inductor
 }
 
 # "Global" flags for inductor benchmarking controlled by TEST_CONFIG
@@ -547,43 +547,60 @@ test_libtorch() {
     ln -sf "$TORCH_LIB_DIR"/libtbb* "$TORCH_BIN_DIR"
     ln -sf "$TORCH_LIB_DIR"/libnvfuser* "$TORCH_BIN_DIR"
 
+    export CPP_TESTS_DIR="${TORCH_BIN_DIR}"
+
     # Start background download
-    python tools/download_mnist.py --quiet -d test/cpp/api/mnist &
+    MNIST_DIR="${PWD}/test/cpp/api/mnist"
+    python tools/download_mnist.py --quiet -d "${MNIST_DIR}" &
 
-    # Make test_reports directory
-    # NB: the ending test_libtorch must match the current function name for the current
-    # test reporting process to function as expected.
-    TEST_REPORTS_DIR=test/test-reports/cpp-unittest/test_libtorch
-    mkdir -p $TEST_REPORTS_DIR
+    # Prepare the model used by test_jit, the model needs to be in the test directory
+    # to get picked up by run_test
+    pushd test
+    python cpp/jit/tests_setup.py setup
+    popd
 
     # Run JIT cpp tests
-    python test/cpp/jit/tests_setup.py setup
-
     if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
-      "$TORCH_BIN_DIR"/test_jit  --gtest_output=xml:$TEST_REPORTS_DIR/test_jit.xml
-      "$TORCH_BIN_DIR"/nvfuser_tests --gtest_output=xml:$TEST_REPORTS_DIR/nvfuser_tests.xml
+      python test/run_test.py --cpp --verbose -i cpp/test_jit cpp/nvfuser_tests
     else
-      "$TORCH_BIN_DIR"/test_jit  --gtest_filter='-*CUDA' --gtest_output=xml:$TEST_REPORTS_DIR/test_jit.xml
+      # CUDA tests have already been skipped when CUDA is not available
+      python test/run_test.py --cpp --verbose -i cpp/test_jit -k "not CUDA"
     fi
 
     # Run Lazy Tensor cpp tests
     if [[ "$BUILD_ENVIRONMENT" == *cuda* && "$TEST_CONFIG" != *nogpu* ]]; then
-      LTC_TS_CUDA=1 "$TORCH_BIN_DIR"/test_lazy  --gtest_output=xml:$TEST_REPORTS_DIR/test_lazy.xml
+      LTC_TS_CUDA=1 python test/run_test.py --cpp --verbose -i cpp/test_lazy
     else
-      "$TORCH_BIN_DIR"/test_lazy  --gtest_output=xml:$TEST_REPORTS_DIR/test_lazy.xml
+      python test/run_test.py --cpp --verbose -i cpp/test_lazy
     fi
 
-    python test/cpp/jit/tests_setup.py shutdown
+    # Cleaning up test artifacts in the test folder
+    pushd test
+    python cpp/jit/tests_setup.py shutdown
+    popd
 
     # Wait for background download to finish
     wait
-    # Exclude IMethodTest that relies on torch::deploy, which will instead be ran in test_deploy.
-    OMP_NUM_THREADS=2 TORCH_CPP_TEST_MNIST_PATH="test/cpp/api/mnist" "$TORCH_BIN_DIR"/test_api --gtest_filter='-IMethodTest.*' --gtest_output=xml:$TEST_REPORTS_DIR/test_api.xml
-    "$TORCH_BIN_DIR"/test_tensorexpr --gtest_output=xml:$TEST_REPORTS_DIR/test_tensorexpr.xml
+
+    if [[ "$BUILD_ENVIRONMENT" == *asan* ]]; then
+        TEST_REPORTS_DIR=test/test-reports/cpp-unittest/test_libtorch
+        mkdir -p $TEST_REPORTS_DIR
+
+        # TODO: Not quite sure why these tests time out only on ASAN, probably
+        # this is due to the fact that a python executable is used and ASAN
+        # treats that differently
+        OMP_NUM_THREADS=2 TORCH_CPP_TEST_MNIST_PATH="${MNIST_DIR}" "$TORCH_BIN_DIR"/test_api --gtest_filter='-IMethodTest.*' --gtest_output=xml:$TEST_REPORTS_DIR/test_api.xml
+        "$TORCH_BIN_DIR"/test_tensorexpr --gtest_output=xml:$TEST_REPORTS_DIR/test_tensorexpr.xml
+    else
+        # Exclude IMethodTest that relies on torch::deploy, which will instead be ran in test_deploy
+        OMP_NUM_THREADS=2 TORCH_CPP_TEST_MNIST_PATH="${MNIST_DIR}" python test/run_test.py --cpp --verbose -i cpp/test_api -k "not IMethodTest"
+        python test/run_test.py --cpp --verbose -i cpp/test_tensorexpr
+    fi
 
     if [[ "${BUILD_ENVIRONMENT}" != *android* && "${BUILD_ENVIRONMENT}" != *cuda* && "${BUILD_ENVIRONMENT}" != *asan* ]]; then
-      # TODO: Consider to run static_runtime_test from $TORCH_BIN_DIR (may need modify build script)
-      "$BUILD_BIN_DIR"/static_runtime_test --gtest_output=xml:$TEST_REPORTS_DIR/static_runtime_test.xml
+      # NB: This test is not under TORCH_BIN_DIR but under BUILD_BIN_DIR
+      export CPP_TESTS_DIR="${BUILD_BIN_DIR}"
+      python test/run_test.py --cpp --verbose -i cpp/static_runtime_test
     fi
 
     assert_git_not_dirty
@@ -595,26 +612,21 @@ test_aot_compilation() {
   ln -sf "$TORCH_LIB_DIR"/libc10* "$TORCH_BIN_DIR"
   ln -sf "$TORCH_LIB_DIR"/libtorch* "$TORCH_BIN_DIR"
 
-  # Make test_reports directory
-  # NB: the ending test_libtorch must match the current function name for the current
-  # test reporting process to function as expected.
-  TEST_REPORTS_DIR=test/test-reports/cpp-unittest/test_aot_compilation
-  mkdir -p $TEST_REPORTS_DIR
-  if [ -f "$TORCH_BIN_DIR"/test_mobile_nnc ]; then "$TORCH_BIN_DIR"/test_mobile_nnc --gtest_output=xml:$TEST_REPORTS_DIR/test_mobile_nnc.xml; fi
-  # shellcheck source=test/mobile/nnc/test_aot_compile.sh
-  if [ -f "$TORCH_BIN_DIR"/aot_model_compiler_test ]; then source test/mobile/nnc/test_aot_compile.sh; fi
+  if [ -f "$TORCH_BIN_DIR"/test_mobile_nnc ]; then
+    CPP_TESTS_DIR="${TORCH_BIN_DIR}" python test/run_test.py --cpp --verbose -i cpp/test_mobile_nnc
+  fi
+
+  if [ -f "$TORCH_BIN_DIR"/aot_model_compiler_test ]; then
+    source test/mobile/nnc/test_aot_compile.sh
+  fi
 }
 
 test_vulkan() {
   if [[ "$BUILD_ENVIRONMENT" == *vulkan* ]]; then
     ln -sf "$TORCH_LIB_DIR"/libtorch* "$TORCH_TEST_DIR"
     ln -sf "$TORCH_LIB_DIR"/libc10* "$TORCH_TEST_DIR"
     export VK_ICD_FILENAMES=/var/lib/jenkins/swiftshader/swiftshader/build/Linux/vk_swiftshader_icd.json
-    # NB: the ending test_vulkan must match the current function name for the current
-    # test reporting process to function as expected.
-    TEST_REPORTS_DIR=test/test-reports/cpp-vulkan/test_vulkan
-    mkdir -p $TEST_REPORTS_DIR
-    LD_LIBRARY_PATH=/var/lib/jenkins/swiftshader/swiftshader/build/Linux/ "$TORCH_TEST_DIR"/vulkan_api_test --gtest_output=xml:$TEST_REPORTS_DIR/vulkan_test.xml
+    CPP_TESTS_DIR="${TORCH_TEST_DIR}" LD_LIBRARY_PATH=/var/lib/jenkins/swiftshader/swiftshader/build/Linux/ python test/run_test.py --cpp --verbose -i cpp/vulkan_api_test
   fi
 }
 
@@ -631,22 +643,24 @@ test_distributed() {
     echo "Testing distributed C++ tests"
     ln -sf "$TORCH_LIB_DIR"/libtorch* "$TORCH_BIN_DIR"
     ln -sf "$TORCH_LIB_DIR"/libc10* "$TORCH_BIN_DIR"
-    # NB: the ending test_distributed must match the current function name for the current
-    # test reporting process to function as expected.
-    TEST_REPORTS_DIR=test/test-reports/cpp-distributed/test_distributed
-    mkdir -p $TEST_REPORTS_DIR
-    "$TORCH_BIN_DIR"/FileStoreTest --gtest_output=xml:$TEST_REPORTS_DIR/FileStoreTest.xml
-    "$TORCH_BIN_DIR"/HashStoreTest --gtest_output=xml:$TEST_REPORTS_DIR/HashStoreTest.xml
-    "$TORCH_BIN_DIR"/TCPStoreTest --gtest_output=xml:$TEST_REPORTS_DIR/TCPStoreTest.xml
+
+    export CPP_TESTS_DIR="${TORCH_BIN_DIR}"
+    # These are distributed tests, so let's continue running them sequentially here to avoid
+    # any surprise
+    python test/run_test.py --cpp --verbose -i cpp/FileStoreTest
+    python test/run_test.py --cpp --verbose -i cpp/HashStoreTest
+    python test/run_test.py --cpp --verbose -i cpp/TCPStoreTest
 
     MPIEXEC=$(command -v mpiexec)
     if [[ -n "$MPIEXEC" ]]; then
+      # NB: mpiexec only works directly with the C++ test binary here
       MPICMD="${MPIEXEC} -np 2 $TORCH_BIN_DIR/ProcessGroupMPITest"
       eval "$MPICMD"
     fi
-    "$TORCH_BIN_DIR"/ProcessGroupGlooTest --gtest_output=xml:$TEST_REPORTS_DIR/ProcessGroupGlooTest.xml
-    "$TORCH_BIN_DIR"/ProcessGroupNCCLTest --gtest_output=xml:$TEST_REPORTS_DIR/ProcessGroupNCCLTest.xml
-    "$TORCH_BIN_DIR"/ProcessGroupNCCLErrorsTest --gtest_output=xml:$TEST_REPORTS_DIR/ProcessGroupNCCLErrorsTest.xml
+
+    python test/run_test.py --cpp --verbose -i cpp/ProcessGroupGlooTest
+    python test/run_test.py --cpp --verbose -i cpp/ProcessGroupNCCLTest
+    python test/run_test.py --cpp --verbose -i cpp/ProcessGroupNCCLErrorsTest
   fi
 }
 
@@ -658,9 +672,8 @@ test_rpc() {
     ln -sf "$TORCH_LIB_DIR"/libtorch* "$TORCH_BIN_DIR"
     ln -sf "$TORCH_LIB_DIR"/libc10* "$TORCH_BIN_DIR"
     ln -sf "$TORCH_LIB_DIR"/libtbb* "$TORCH_BIN_DIR"
-    TEST_REPORTS_DIR=test/test-reports/cpp-rpc/test_rpc
-    mkdir -p $TEST_REPORTS_DIR
-    "$TORCH_BIN_DIR"/test_cpp_rpc --gtest_output=xml:$TEST_REPORTS_DIR/test_cpp_rpc.xml
+
+    CPP_TESTS_DIR="${TORCH_BIN_DIR}" python test/run_test.py --cpp --verbose -i cpp/test_cpp_rpc
   fi
 }
 
diff --git a/.ci/pytorch/win-test-helpers/test_libtorch.bat b/.ci/pytorch/win-test-helpers/test_libtorch.bat
@@ -5,14 +5,16 @@ if "%USE_CUDA%" == "0" IF NOT "%CUDA_VERSION%" == "cpu" exit /b 0
 call %SCRIPT_HELPERS_DIR%\setup_pytorch_env.bat
 if errorlevel 1 exit /b 1
 
-cd %TMP_DIR_WIN%\build\torch\bin
-set TEST_OUT_DIR=%~dp0\..\..\..\test\test-reports\cpp-unittest
-md %TEST_OUT_DIR%
+:: Save the current working directory so that we can go back there
+set CWD=%cd%
+
+set CPP_TESTS_DIR=%TMP_DIR_WIN%\build\torch\bin
 set PATH=C:\Program Files\NVIDIA Corporation\NvToolsExt\bin\x64;%TMP_DIR_WIN%\build\torch\lib;%PATH%
 
-set TEST_API_OUT_DIR=%TEST_OUT_DIR%\test_api
-md %TEST_API_OUT_DIR%
-test_api.exe --gtest_filter="-IntegrationTest.MNIST*" --gtest_output=xml:%TEST_API_OUT_DIR%\test_api.xml
+set TORCH_CPP_TEST_MNIST_PATH=%CWD%\test\cpp\api\mnist
+python tools\download_mnist.py --quiet -d %TORCH_CPP_TEST_MNIST_PATH%
+
+python test\run_test.py --cpp --verbose -i cpp/test_api
 if errorlevel 1 exit /b 1
 if not errorlevel 0 exit /b 1
 
@@ -25,6 +27,10 @@ for /r "." %%a in (*.exe) do (
 goto :eof
 
 :libtorch_check
+
+cd %CWD%
+set CPP_TESTS_DIR=%TMP_DIR_WIN%\build\torch\test
+
 :: Skip verify_api_visibility as it a compile level test
 if "%~1" == "verify_api_visibility" goto :eof
 
@@ -42,12 +48,12 @@ if "%~1" == "utility_ops_gpu_test" goto :eof
 
 echo Running "%~2"
 if "%~1" == "c10_intrusive_ptr_benchmark" (
+  :: NB: This is not a gtest executable file, thus couldn't be handled by pytest-cpp
   call "%~2"
   goto :eof
 )
-:: Differentiating the test report directories is crucial for test time reporting.
-md %TEST_OUT_DIR%\%~n2
-call "%~2" --gtest_output=xml:%TEST_OUT_DIR%\%~n2\%~1.xml
+
+python test\run_test.py --cpp --verbose -i "cpp/%~1"
 if errorlevel 1 (
   echo %1 failed with exit code %errorlevel%
   exit /b 1
diff --git a/aten/tools/run_tests.sh b/aten/tools/run_tests.sh
@@ -3,75 +3,78 @@ set -x
 set -e
 
 VALGRIND_SUP="${PWD}/`dirname $0`/valgrind.sup"
-pushd $1
+export CPP_TESTS_DIR=$1
 
 VALGRIND=${VALGRIND:=ON}
-./basic
-./atest
-./scalar_test
-./broadcast_test
-./wrapdim_test
-./apply_utils_test
-./dlconvertor_test
-./native_test
-./scalar_tensor_test
+python test/run_test.py --cpp --verbose -i \
+  cpp/basic \
+  cpp/atest \
+  cpp/scalar_test \
+  cpp/broadcast_test \
+  cpp/wrapdim_test \
+  cpp/apply_utils_test \
+  cpp/dlconvertor_test \
+  cpp/native_test \
+  cpp/scalar_tensor_test \
+  cpp/undefined_tensor_test \
+  cpp/extension_backend_test \
+  cpp/lazy_tensor_test \
+  cpp/tensor_iterator_test \
+  cpp/Dimname_test \
+  cpp/Dict_test \
+  cpp/NamedTensor_test \
+  cpp/cpu_generator_test \
+  cpp/legacy_vmap_test \
+  cpp/operators_test
+
 if [[ -x ./tensor_interop_test ]]; then
-  ./tensor_interop_test
+  python test/run_test.py --cpp --verbose -i cpp/tensor_interop_test
 fi
-./undefined_tensor_test
-./extension_backend_test
-./lazy_tensor_test
-./tensor_iterator_test
-./Dimname_test
-./Dict_test
-./NamedTensor_test
-./cpu_generator_test
-./legacy_vmap_test
-./operators_test
 if [[ -x ./cudnn_test ]]; then
-  ./cudnn_test
+  python test/run_test.py --cpp --verbose -i cpp/cudnn_test
 fi
 if [[ -x ./cuda_generator_test ]]; then
-  ./cuda_generator_test
+  python test/run_test.py --cpp --verbose -i cpp/cuda_generator_test
 fi
 if [[ -x ./apply_test ]]; then
-  ./apply_test
+  python test/run_test.py --cpp --verbose -i cpp/apply_test
 fi
 if [[ -x ./stream_test ]]; then
-  ./stream_test
+  python test/run_test.py --cpp --verbose -i cpp/stream_test
 fi
 if [[ -x ./cuda_half_test ]]; then
-  ./cuda_half_test
+  python test/run_test.py --cpp --verbose -i cpp/cuda_half_test
 fi
 if [[ -x ./cuda_vectorized_test ]]; then
-  ./cuda_vectorized_test
+  python test/run_test.py --cpp --verbose -i cpp/cuda_vectorized_test
 fi
 if [[ -x ./cuda_distributions_test ]]; then
-  ./cuda_distributions_test
+  python test/run_test.py --cpp --verbose -i cpp/cuda_distributions_test
 fi
 if [[ -x ./cuda_optional_test ]]; then
-  ./cuda_optional_test
+  python test/run_test.py --cpp --verbose -i cpp/cuda_optional_test
 fi
 if [[ -x ./cuda_tensor_interop_test ]]; then
-  ./cuda_tensor_interop_test
+  python test/run_test.py --cpp --verbose -i cpp/cuda_tensor_interop_test
 fi
 if [[ -x ./cuda_complex_test ]]; then
-  ./cuda_complex_test
+  python test/run_test.py --cpp --verbose -i cpp/cuda_complex_test
 fi
 if [[ -x ./cuda_complex_math_test ]]; then
-  ./cuda_complex_math_test
+  python test/run_test.py --cpp --verbose -i cpp/cuda_complex_math_test
 fi
 if [[ -x ./cuda_cub_test ]]; then
-  ./cuda_cub_test
+  python test/run_test.py --cpp --verbose -i cpp/cuda_cub_test
 fi
 if [[ -x ./cuda_atomic_ops_test ]]; then
-  ./cuda_atomic_ops_test
+  python test/run_test.py --cpp --verbose -i cpp/cuda_atomic_ops_test
 fi
+
 if [ "$VALGRIND" == "ON" ]; then
-  valgrind --suppressions="$VALGRIND_SUP" --error-exitcode=1 ./basic --gtest_filter='-*CUDA'
+  # NB: As these tests are invoked by valgrind, let's leave them for now as it's
+  # unclear if valgrind -> python -> gtest would work
+  valgrind --suppressions="$VALGRIND_SUP" --error-exitcode=1 "${CPP_TESTS_DIR}/basic" --gtest_filter='-*CUDA'
   if [[ -x ./tensor_interop_test ]]; then
-    valgrind --suppressions="$VALGRIND_SUP" --error-exitcode=1 ./tensor_interop_test
+    valgrind --suppressions="$VALGRIND_SUP" --error-exitcode=1 "${CPP_TESTS_DIR}/tensor_interop_test"
   fi
 fi
-
-popd
diff --git a/test/run_test.py b/test/run_test.py
