Add test filter standardization for rocshmem

build_tools/github_actions/fetch_test_configurations.py

@@ -446,6 +446,20 @@ def _build_container_options(job_config: dict, platform: str) -> dict:
         # Architectures that we have multi GPU setup for testing
         "multi_gpu": {"linux": ["gfx94X-dcgpu", "gfx950-dcgpu"]},
     },
+    # rocSHMEM tests
+    "rocshmem": {
+        "job_name": "rocshmem",
+        "fetch_artifact_args": "--rocshmem --tests",
+        "timeout_minutes": 30,
+        "test_script": f"python {_get_script_path('test_runner.py')}",
+        "platform": ["linux"],
+        "total_shards_dict": {
+            "linux": 1,
+        },
+        # rocSHMEM functional/unit tests launch via mpirun with RANKS 2..64, so
+        # they need a multi-GPU runner (same setup as rccl).
+        "multi_gpu": {"linux": ["gfx94X-dcgpu", "gfx950-dcgpu"]},
+    },
     # rocprofiler-sdk tests
     "rocprofiler-sdk": {
         "job_name": "rocprofiler-sdk",

build_tools/github_actions/tests/fetch_test_configurations_test.py

@@ -283,9 +283,13 @@ def fake_select_weighted_label(labels_config, context_name):
 
             rccl = next(j for j in components if j["job_name"] == "rccl")
             self.assertEqual(rccl["multi_gpu_runner"], "linux-mi300-mgpu-a")
-            # Verify select_weighted_label was called
-            self.assertEqual(len(selected_labels), 1)
-            self.assertEqual(selected_labels[0][1], "gfx94x-multi-gpu")
+            # Verify select_weighted_label was used for multi-GPU jobs. There may
+            # be more than one multi-GPU job (e.g. rccl and rocshmem), so assert
+            # it was called at least once and always with the multi-gpu context.
+            self.assertGreaterEqual(len(selected_labels), 1)
+            self.assertTrue(
+                all(ctx == "gfx94x-multi-gpu" for _, ctx in selected_labels)
+            )
         finally:
             fetch_test_configurations.select_weighted_label = (
                 original_select_weighted_label

build_tools/install_rocm_from_artifacts.py

@@ -33,6 +33,7 @@
     [--prim | --no-prim]
     [--rand | --no-rand]
     [--rccl | --no-rccl]
+    [--rocshmem | --no-rocshmem]
     [--rocdecode | --no-rocdecode]
     [--rocjpeg | --no-rocjpeg]
     [--rocjitsu | --no-rocjitsu]
@@ -358,6 +359,7 @@ def retrieve_artifacts_by_run_id(args):
             args.mpi,
             args.rand,
             args.rccl,
+            args.rocshmem,
             args.rocdecode,
             args.rocjpeg,
             args.rocjitsu,
@@ -447,6 +449,16 @@ def retrieve_artifacts_by_run_id(args):
             extra_artifacts.append("rand")
         if args.rccl:
             extra_artifacts.append("rccl")
+        if args.rocshmem:
+            extra_artifacts.append("rocshmem")
+            # The functional test binary (bin/rocshmem_functional_tests) and
+            # bin/rocshmem_info live in the _run component; the install-tree
+            # CTestTestfile.cmake references them via relative paths.
+            argv.append("rocshmem_run")
+            # rocSHMEM tests launch via mpirun and link against TheRock's
+            # vendored OpenMPI, so pull it (with its run component for mpiexec).
+            extra_artifacts.append("openmpi")
+            argv.append("openmpi_run")
         if args.rocprofiler_sdk:
             extra_artifacts.append("rocprofiler-sdk")
             extra_artifacts.append("aqlprofile")
@@ -788,6 +800,13 @@ def main(argv):
         action=argparse.BooleanOptionalAction,
     )
 
+    artifacts_group.add_argument(
+        "--rocshmem",
+        default=False,
+        help="Include 'rocshmem' artifacts",
+        action=argparse.BooleanOptionalAction,
+    )
+
     artifacts_group.add_argument(
         "--mpi",
         default=False,

comm-libs/CMakeLists.txt

@@ -153,6 +153,10 @@ if(THEROCK_ENABLE_ROCSHMEM)
       "-DEXPLICIT_ROCM_VERSION=${ROCM_MAJOR_VERSION}.${ROCM_MINOR_VERSION}.${ROCM_PATCH_VERSION}"
       -DBUILD_FUNCTIONAL_TESTS=${THEROCK_BUILD_TESTING}
       -DBUILD_UNIT_TESTS=${THEROCK_BUILD_TESTING}
+      # Register functional/unit tests with CTest and install the relocatable
+      # bin/rocshmem/**/CTestTestfile.cmake files (with quick/standard/
+      # comprehensive/full tier labels) that test_runner.py drives via ctest.
+      -DBUILD_CTESTS=${THEROCK_BUILD_TESTING}
       -DSOURCE_DIR=${_source_dir}
       -DPATCHELF=${PATCHELF}
       -DPython3_EXECUTABLE=${Python3_EXECUTABLE}

comm-libs/artifact-rocshmem.toml

@@ -21,4 +21,19 @@ include = [
 ]
 # Tests
 [components.test."comm-libs/rocshmem/stage"]
+include = [
+  # Install-tree CTest files: test_runner.py runs `ctest --test-dir bin/rocshmem`,
+  # whose root CTestTestfile.cmake subdirs into tests/functional and tests/unit.
+  "bin/rocshmem/**",
+  # Unit test binary (the functional test binary and rocshmem_info come from the
+  # run component, fetched via rocshmem_run).
+  "bin/rocshmem_unit_tests",
+  # CTest wrapper scripts referenced (by relative path) from the install-tree
+  # CTestTestfile.cmake, plus the retry helper.
+  "share/rocshmem/test_wrapper.sh",
+  "share/rocshmem/unit_test_wrapper.sh",
+  "share/rocshmem/ctest_with_retry.sh",
+  # HSACOs used by the device_bitcode functional tests (quick tier).
+  "share/rocshmem/*.hsaco",
+]
 optional = true