Benchmarks: Add Test - Add tests for matmul and sharding-matmul benchmarks. (#41)

* add tests for matmul and sharding-matmul benchmarks.

* add decorator for sharding_matmul tests.

* add __init__.py for utils of benchmarks tests.

* disable GPU tests for CPU platform validation.

* fix typo
Co-authored-by: Guoshuai Zhao <guzhao@microsoft.com>
Co-authored-by: Peng Cheng <chengpeng5555@outlook.com>
Co-authored-by: Yifan Xiong <yifan.xiong@microsoft.com>

test.dockerfile

@@ -26,4 +26,4 @@ RUN python3 -m pip install --upgrade pip setuptools && \
 RUN python3 setup.py lint
 
 # Test code
-RUN python3 setup.py test
+RUN SB_TEST_CUDA=0 python3 setup.py test

tests/benchmarks/__init__.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+"""Utilities module for benchmarks tests."""

tests/benchmarks/micro_benchmarks/test_matmul.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+"""Tests for matmul benchmark."""
+
+from tests.helper import decorator
+from superbench.benchmarks import BenchmarkRegistry, Platform, Framework, BenchmarkType, ReturnCode
+from superbench.benchmarks.micro_benchmarks.sharding_matmul import ShardingMode
+
+
+@decorator.cuda_test
+@decorator.pytorch_test
+def test_pytorch_matmul():
+    """Test pytorch-matmul benchmark."""
+    context = BenchmarkRegistry.create_benchmark_context(
+        'matmul', platform=Platform.CUDA, parameters='--run_count 2 --num_steps 20', framework=Framework.PYTORCH
+    )
+
+    assert (BenchmarkRegistry.is_benchmark_context_valid(context))
+
+    benchmark = BenchmarkRegistry.launch_benchmark(context)
+
+    # Check basic information.
+    assert (benchmark)
+    assert (benchmark.name == 'pytorch-matmul')
+    assert (benchmark.type == BenchmarkType.MICRO)
+
+    # Check predefined parameters of sharding-matmul benchmark.
+    assert (benchmark._args.mode == [ShardingMode.NOSHARDING])
+
+    # Check parameters specified in BenchmarkContext.
+    assert (benchmark._args.run_count == 2)
+    assert (benchmark._args.num_steps == 20)
+
+    # Check results and metrics.
+    assert (benchmark.run_count == 2)
+    assert (benchmark.return_code == ReturnCode.SUCCESS)
+    assert (len(benchmark.raw_data['nosharding']) == benchmark.run_count)
+    assert (len(benchmark.raw_data['nosharding'][0]) == benchmark._args.num_steps)
+    assert (len(benchmark.result['nosharding']) == benchmark.run_count)

tests/benchmarks/micro_benchmarks/test_sharding_matmul.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+"""Tests for sharding-matmul benchmark."""
+
+import tests.benchmarks.utils as utils
+from tests.helper import decorator
+from superbench.benchmarks import BenchmarkRegistry, Platform, Framework, BenchmarkType, ReturnCode
+from superbench.benchmarks.micro_benchmarks.sharding_matmul import ShardingMatmul, ShardingMode
+
+
+@decorator.cuda_test
+@decorator.pytorch_test
+def test_pytorch_sharding_matmul():
+    """Test pytorch-sharding-matmul benchmark."""
+    context = BenchmarkRegistry.create_benchmark_context(
+        'sharding-matmul',
+        platform=Platform.CUDA,
+        parameters='--run_count 2 --num_steps 20',
+        framework=Framework.PYTORCH
+    )
+
+    assert (BenchmarkRegistry.is_benchmark_context_valid(context))
+
+    utils.setup_simulated_ddp_distributed_env()
+    benchmark = BenchmarkRegistry.launch_benchmark(context)
+
+    # Check basic information.
+    assert (benchmark)
+    assert (isinstance(benchmark, ShardingMatmul))
+    assert (benchmark.name == 'pytorch-sharding-matmul')
+    assert (benchmark.type == BenchmarkType.MICRO)
+
+    # Check predefined parameters of sharding-matmul benchmark.
+    assert (benchmark._args.mode == [ShardingMode.ALLREDUCE, ShardingMode.ALLGATHER])
+
+    # Check parameters specified in BenchmarkContext.
+    assert (benchmark._args.run_count == 2)
+    assert (benchmark._args.num_steps == 20)
+
+    # Check results and metrics.
+    assert (benchmark.run_count == 2)
+    assert (benchmark.return_code == ReturnCode.SUCCESS)
+    for metric in ['allreduce', 'allgather']:
+        assert (len(benchmark.raw_data[metric]) == benchmark.run_count)
+        assert (len(benchmark.raw_data[metric][0]) == benchmark._args.num_steps)
+        assert (len(benchmark.result[metric]) == benchmark.run_count)
+
+    utils.clean_simulated_ddp_distributed_env()

tests/benchmarks/utils.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+"""Utilities for benchmark tests."""
+
+import os
+
+
+def setup_simulated_ddp_distributed_env():
+    """Function to setup the simulated DDP distributed envionment variables."""
+    os.environ['WORLD_SIZE'] = '1'
+    os.environ['RANK'] = '0'
+    os.environ['LOCAL_RANK'] = '0'
+    os.environ['MASTER_ADDR'] = 'localhost'
+    os.environ['MASTER_PORT'] = '12345'
+
+
+def clean_simulated_ddp_distributed_env():
+    """Function to clean up the simulated DDP distributed envionment variables."""
+    os.environ.pop('WORLD_SIZE')
+    os.environ.pop('RANK')
+    os.environ.pop('LOCAL_RANK')
+    os.environ.pop('MASTER_ADDR')
+    os.environ.pop('MASTER_PORT')