Add test_amd_mi250 CI job including ROCm matmul tests. (#17293)
Progress on https://github.com/iree-org/iree/issues/17159
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index cc79e39..93616ee 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -390,6 +390,45 @@
./build_tools/scripts/check_vulkan.sh
./build_tools/cmake/ctest_all.sh ${BUILD_DIR}"
+ test_amd_mi250:
+ needs: [setup, build_all]
+ if: contains(fromJson(needs.setup.outputs.enabled-jobs), 'test_amd_mi250')
+ env:
+ BUILD_DIR: build-tests
+ INSTALL_DIR: ${{ needs.build_all.outputs.install-dir }}
+ INSTALL_DIR_ARCHIVE: ${{ needs.build_all.outputs.install-dir-archive }}
+ INSTALL_DIR_GCS_URL: ${{ needs.build_all.outputs.install-dir-gcs-url }}
+ IREE_CPU_DISABLE: 1
+ IREE_VULKAN_DISABLE: 1
+ IREE_CUDA_DISABLE: 1
+ IREE_HIP_DISABLE: 0
+ IREE_HIP_TEST_TARGET_CHIP: "gfx90a"
+ runs-on: nodai-amdgpu-mi250-x86-64
+ steps:
+ - name: "Checking out repository"
+ uses: actions/checkout@ac593985615ec2ede58e132d2e21d2b1cbd6127c # v3.3.0
+ - name: "Checking out runtime submodules"
+ run: ./build_tools/scripts/git/update_runtime_submodules.sh
+ - name: "Downloading install dir archive"
+ run: wget "${INSTALL_DIR_GCS_URL}" -O "${INSTALL_DIR_ARCHIVE}"
+ - name: "Extracting install directory"
+ run: tar -xf "${INSTALL_DIR_ARCHIVE}"
+ - name: "Building tests"
+ run: |
+ ./build_tools/pkgci/build_tests_using_package.sh ${INSTALL_DIR}
+ - name: "Running GPU tests"
+ env:
+ IREE_CTEST_LABEL_REGEX: ^requires-gpu|^driver=hip$
+ IREE_NVIDIA_SM80_TESTS_DISABLE: 1
+ IREE_MULTI_DEVICE_TESTS_DISABLE: 0
+ IREE_AMD_RDNA3_TESTS_DISABLE: 1
+ IREE_NVIDIA_GPU_TESTS_DISABLE: 0
+ IREE_CUDA_DISABLE: 1
+ IREE_CPU_DISABLE: 1
+ IREE_HIP_DISABLE: 0
+ run: |
+ ./build_tools/cmake/ctest_all.sh ${BUILD_DIR}
+
################################## Subsets ###################################
# Jobs that build some subset of IREE
##############################################################################
@@ -1076,6 +1115,7 @@
# Accelerators
- test_nvidia_gpu
- test_nvidia_a100
+ - test_amd_mi250
# Subsets
- build_test_runtime
diff --git a/build_tools/pkgci/build_tests_using_package.sh b/build_tools/pkgci/build_tests_using_package.sh
index cd5ad13..522c0c3 100755
--- a/build_tools/pkgci/build_tests_using_package.sh
+++ b/build_tools/pkgci/build_tests_using_package.sh
@@ -39,6 +39,7 @@
export IREE_METAL_DISABLE="${IREE_METAL_DISABLE:-1}"
export IREE_CUDA_DISABLE="${IREE_CUDA_DISABLE:-1}"
export IREE_HIP_DISABLE="${IREE_HIP_DISABLE:-1}"
+export IREE_HIP_TEST_TARGET_CHIP="${IREE_HIP_TEST_TARGET_CHIP:-}"
# Set cmake options based on disabled features.
declare -a cmake_config_options=()
@@ -58,6 +59,9 @@
if (( IREE_HIP_DISABLE == 1 )); then
cmake_config_options+=("-DIREE_HAL_DRIVER_HIP=OFF")
fi
+if [[ -n "${IREE_HIP_TEST_TARGET_CHIP}" ]]; then
+ cmake_config_options+=("-DIREE_HIP_TEST_TARGET_CHIP=${IREE_HIP_TEST_TARGET_CHIP}")
+fi
###############################################################################
# Build the runtime and compile 'test deps' #
diff --git a/runtime/src/iree/hal/drivers/hip/cts/CMakeLists.txt b/runtime/src/iree/hal/drivers/hip/cts/CMakeLists.txt
index e8cadc2..5d7c109 100644
--- a/runtime/src/iree/hal/drivers/hip/cts/CMakeLists.txt
+++ b/runtime/src/iree/hal/drivers/hip/cts/CMakeLists.txt
@@ -37,6 +37,7 @@
"event"
LABELS
driver=hip
+ requires-gpu-amd
)
iree_hal_cts_test_suite(
@@ -59,6 +60,12 @@
DEPS
iree::hal::drivers::hip::registration
EXCLUDED_TESTS
+ # These tests fail with:
+ # UNAVAILABLE; missing hipDrvGraphAddMemcpyNode symbol;
+ # cannot use graph-based command buffer
+ "command_buffer"
+ "command_buffer_dispatch"
+ "file"
# HAL event is unimplemented for now.
"event"
LABELS
diff --git a/runtime/src/iree/hal/drivers/hip/registration/CMakeLists.txt b/runtime/src/iree/hal/drivers/hip/registration/CMakeLists.txt
index 318412e..cf51242 100644
--- a/runtime/src/iree/hal/drivers/hip/registration/CMakeLists.txt
+++ b/runtime/src/iree/hal/drivers/hip/registration/CMakeLists.txt
@@ -14,6 +14,7 @@
DEPS
iree::base
iree::base::core_headers
+ iree::base::internal::flags
iree::hal::drivers::hip
iree::hal
DEFINES
diff --git a/tests/e2e/matmul/BUILD.bazel b/tests/e2e/matmul/BUILD.bazel
index b6fe7b8..ef1948a 100644
--- a/tests/e2e/matmul/BUILD.bazel
+++ b/tests/e2e/matmul/BUILD.bazel
@@ -521,60 +521,6 @@
##
###########################################################################
-# Testing CDNA3 + matrix core path.
-iree_generated_e2e_runner_test(
- name = "e2e_matmul_rocm_f16_large_cdna3_mfma",
- compiler_flags = [
- "--iree-rocm-target-chip=gfx942",
- ],
- generator = ":generate_e2e_matmul_tests",
- generator_args = [
- "--lhs_rhs_type=f16",
- "--acc_type=f32",
- "--shapes=gpu_large_aligned",
- "--compilation_info=LLVMGPUVectorDistributeMFMA",
- ],
- tags = [
- "noasan",
- "nomsan",
- "notsan",
- "noubsan",
- "requires-gpu-cdna3",
- ],
- target_backends_and_drivers = [
- ("rocm", "hip"),
- ],
- test_runner = "//tools/testing/e2e:iree-e2e-matmul-test",
- test_type = "matmul",
-)
-
-iree_generated_e2e_runner_test(
- name = "e2e_matmul_rocm_f16_large_cdna3_mfma_tb",
- compiler_flags = [
- "--iree-rocm-target-chip=gfx942",
- ],
- generator = ":generate_e2e_matmul_tests",
- generator_args = [
- "--lhs_rhs_type=f16",
- "--acc_type=f32",
- "--transpose_rhs",
- "--shapes=gpu_large_aligned",
- "--compilation_info=LLVMGPUVectorDistributeMFMA",
- ],
- tags = [
- "noasan",
- "nomsan",
- "notsan",
- "noubsan",
- "requires-gpu-cdna3",
- ],
- target_backends_and_drivers = [
- ("rocm", "hip"),
- ],
- test_runner = "//tools/testing/e2e:iree-e2e-matmul-test",
- test_type = "matmul",
-)
-
# Testing RDNA3 + matrix core path.
iree_generated_e2e_runner_test(
name = "e2e_matmul_rocm_f16_large_rdna3_wmma",
diff --git a/tests/e2e/matmul/CMakeLists.txt b/tests/e2e/matmul/CMakeLists.txt
index cd4f225..f7a4fe9 100644
--- a/tests/e2e/matmul/CMakeLists.txt
+++ b/tests/e2e/matmul/CMakeLists.txt
@@ -1814,63 +1814,6 @@
iree_generated_e2e_runner_test(
NAME
- e2e_matmul_rocm_f16_large_cdna3_mfma
- TEST_TYPE
- matmul
- GENERATOR
- "generate_e2e_matmul_tests.py"
- GENERATOR_ARGS
- "--lhs_rhs_type=f16"
- "--acc_type=f32"
- "--shapes=gpu_large_aligned"
- "--compilation_info=LLVMGPUVectorDistributeMFMA"
- TEST_RUNNER
- iree_tools_testing_e2e_iree-e2e-matmul-test
- TARGET_BACKENDS
- "rocm"
- DRIVERS
- "hip"
- COMPILER_FLAGS
- "--iree-rocm-target-chip=gfx942"
- LABELS
- "noasan"
- "nomsan"
- "notsan"
- "noubsan"
- "requires-gpu-cdna3"
-)
-
-iree_generated_e2e_runner_test(
- NAME
- e2e_matmul_rocm_f16_large_cdna3_mfma_tb
- TEST_TYPE
- matmul
- GENERATOR
- "generate_e2e_matmul_tests.py"
- GENERATOR_ARGS
- "--lhs_rhs_type=f16"
- "--acc_type=f32"
- "--transpose_rhs"
- "--shapes=gpu_large_aligned"
- "--compilation_info=LLVMGPUVectorDistributeMFMA"
- TEST_RUNNER
- iree_tools_testing_e2e_iree-e2e-matmul-test
- TARGET_BACKENDS
- "rocm"
- DRIVERS
- "hip"
- COMPILER_FLAGS
- "--iree-rocm-target-chip=gfx942"
- LABELS
- "noasan"
- "nomsan"
- "notsan"
- "noubsan"
- "requires-gpu-cdna3"
-)
-
-iree_generated_e2e_runner_test(
- NAME
e2e_matmul_rocm_f16_large_rdna3_wmma
TEST_TYPE
matmul
@@ -2098,3 +2041,68 @@
)
### BAZEL_TO_CMAKE_PRESERVES_ALL_CONTENT_BELOW_THIS_LINE ###
+
+if(IREE_HIP_TEST_TARGET_CHIP MATCHES "^gfx")
+
+unset(IREE_HIP_TEST_COMPILER_FLAGS)
+list(APPEND IREE_HIP_TEST_COMPILER_FLAGS
+ "--iree-rocm-target-chip=${IREE_HIP_TEST_TARGET_CHIP}"
+)
+
+iree_generated_e2e_runner_test(
+ NAME
+ e2e_matmul_rocm_f16_large_cdna3_mfma
+ TEST_TYPE
+ matmul
+ GENERATOR
+ "generate_e2e_matmul_tests.py"
+ GENERATOR_ARGS
+ "--lhs_rhs_type=f16"
+ "--acc_type=f32"
+ "--shapes=gpu_large_aligned"
+ "--compilation_info=LLVMGPUVectorDistributeMFMA"
+ TEST_RUNNER
+ iree_tools_testing_e2e_iree-e2e-matmul-test
+ TARGET_BACKENDS
+ "rocm"
+ DRIVERS
+ "hip"
+ COMPILER_FLAGS
+ ${IREE_HIP_TEST_COMPILER_FLAGS}
+ LABELS
+ "noasan"
+ "nomsan"
+ "notsan"
+ "noubsan"
+ "requires-gpu-cdna3"
+)
+
+iree_generated_e2e_runner_test(
+ NAME
+ e2e_matmul_rocm_f16_large_cdna3_mfma_tb
+ TEST_TYPE
+ matmul
+ GENERATOR
+ "generate_e2e_matmul_tests.py"
+ GENERATOR_ARGS
+ "--lhs_rhs_type=f16"
+ "--acc_type=f32"
+ "--transpose_rhs"
+ "--shapes=gpu_large_aligned"
+ "--compilation_info=LLVMGPUVectorDistributeMFMA"
+ TEST_RUNNER
+ iree_tools_testing_e2e_iree-e2e-matmul-test
+ TARGET_BACKENDS
+ "rocm"
+ DRIVERS
+ "hip"
+ COMPILER_FLAGS
+ ${IREE_HIP_TEST_COMPILER_FLAGS}
+ LABELS
+ "noasan"
+ "nomsan"
+ "notsan"
+ "noubsan"
+ "requires-gpu-cdna3"
+)
+endif(IREE_HIP_TEST_TARGET_CHIP MATCHES "^gfx")
