Benchmarks: Add Benchmark - Add onnx model benchmarks based on docker image. (#227)

Add RocmOnnxModelBenchmark class to run benchmarks packaged in superbench/benchmark:rocm4.3.1-onnxruntime1.9.0

examples/benchmarks/rocm_onnxruntime_model_benchmark.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+"""Docker benchmark example for onnxruntime models.
+
+Commands to run:
+  python3 examples/benchmarks/rocm_onnxruntime_model_benchmark.py
+"""
+
+from superbench.benchmarks import BenchmarkRegistry, Framework, Platform
+from superbench.common.utils import logger
+
+if __name__ == '__main__':
+    context = BenchmarkRegistry.create_benchmark_context(
+        'ort-models', platform=Platform.ROCM, framework=Framework.ONNXRUNTIME
+    )
+
+    benchmark = BenchmarkRegistry.launch_benchmark(context)
+    if benchmark:
+        logger.info(
+            'benchmark: {}, return code: {}, result: {}'.format(
+                benchmark.name, benchmark.return_code, benchmark.result
+            )
+        )

superbench/benchmarks/docker_benchmarks/__init__.py

@@ -5,5 +5,6 @@
 
 from superbench.benchmarks.docker_benchmarks.docker_base import DockerBenchmark, CudaDockerBenchmark, \
     RocmDockerBenchmark
+from superbench.benchmarks.docker_benchmarks.rocm_onnxruntime_performance import RocmOnnxRuntimeModelBenchmark
 
-__all__ = ['DockerBenchmark', 'CudaDockerBenchmark', 'RocmDockerBenchmark']
+__all__ = ['DockerBenchmark', 'CudaDockerBenchmark', 'RocmDockerBenchmark', 'RocmOnnxRuntimeModelBenchmark']

superbench/benchmarks/docker_benchmarks/rocm_onnxruntime_performance.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+"""Module of the onnxruntime E2E model benchmarks.
+
+Including:
+  bert-large-uncased ngpu=1
+  bert-large-uncased ngpu=8
+  distilbert-base-uncased ngpu=1
+  distilbert-base-uncased ngpu=8
+  gpt2 ngpu=1
+  gpt2 ngpu=8
+  facebook/bart-large ngpu=1
+  facebook/bart-large ngpu=8
+  roberta-large ngpu=1
+  roberta-large ngpu=8
+"""
+
+from superbench.common.utils import logger
+from superbench.benchmarks import BenchmarkRegistry, Platform
+from superbench.benchmarks.docker_benchmarks.docker_base import RocmDockerBenchmark
+
+
+class RocmOnnxRuntimeModelBenchmark(RocmDockerBenchmark):
+    """The onnxruntime E2E model benchmark class."""
+    def __init__(self, name, parameters=''):
+        """Constructor.
+
+        Args:
+            name (str): benchmark name.
+            parameters (str): benchmark parameters.
+        """
+        super().__init__(name, parameters)
+
+        # Image uri of the current docker-benchmark.
+        self._image_uri = 'superbench/benchmark:rocm4.3.1-onnxruntime1.9.0'
+
+        # Image digest of the current docker-benchmark.
+        self._digest = 'f5e6c832e3cdcbba9820c619bb30fa47ca7117aa7f2c15944d17e6983d37ab9a'
+
+        # Container name of the current docker-benchmark.
+        self._container_name = 'rocm-onnxruntime-model-benchmarks'
+
+        # Entrypoint option of the current docker-benchmark.
+        self._entrypoint = '/stage/onnxruntime-training-examples/huggingface/azureml/run_benchmark.sh'
+
+        # CMD option of the current docker-benchmark.
+        self._cmd = None
+
+    def _process_raw_result(self, cmd_idx, raw_output):
+        """Function to parse raw results and save the summarized results.
+
+          self._result.add_raw_data() and self._result.add_result() need to be called to save the results.
+
+        Args:
+            cmd_idx (int): the index of command corresponding with the raw_output.
+            raw_output (str): raw output string of the micro-benchmark.
+
+        Return:
+            True if the raw output string is valid and result can be extracted.
+        """
+        self._result.add_raw_data('raw_output', raw_output)
+
+        content = raw_output.splitlines(False)
+        try:
+            name_prefix = '__superbench__ begin '
+            value_prefix = '    "samples_per_second": '
+            model_name = None
+            for line in content:
+                if name_prefix in line:
+                    model_name = line[len(name_prefix):]
+                    for char in ['-', ' ', '=', '/']:
+                        model_name = model_name.replace(char, '_')
+                elif value_prefix in line and model_name is not None:
+                    throughput = float(line[len(value_prefix):])
+                    self._result.add_result(model_name, throughput)
+                    model_name = None
+        except BaseException as e:
+            logger.error(
+                'The result format is invalid - round: {}, benchmark: {}, message: {}.'.format(
+                    self._curr_run_index, self._name, str(e)
+                )
+            )
+            return False
+
+        return True
+
+
+BenchmarkRegistry.register_benchmark('onnxruntime-ort-models', RocmOnnxRuntimeModelBenchmark, platform=Platform.ROCM)

superbench/config/amd_mi100_hpe.yaml

@@ -81,6 +81,12 @@ superbench:
         mem_type:
           - dtoh
           - htod
+    ort-models:
+      enable: false
+      modes:
+        - name: local
+      frameworks:
+        - onnxruntime
     gpt_models:
       <<: *default_pytorch_mode
       models:

superbench/config/amd_mi100_z53.yaml

@@ -82,6 +82,12 @@ superbench:
         mem_type:
           - dtoh
           - htod
+    ort-models:
+      enable: false
+      modes:
+        - name: local
+      frameworks:
+        - onnxruntime
     gpt_models:
       <<: *default_pytorch_mode
       models:

tests/benchmarks/docker_benchmarks/test_rocm_onnxruntime_performance.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+"""Tests for RocmOnnxRuntimeModelBenchmark modules."""
+
+from superbench.benchmarks import BenchmarkRegistry, BenchmarkType, Platform, ReturnCode
+from superbench.benchmarks.result import BenchmarkResult
+
+
+def test_rocm_onnxruntime_performance():
+    """Test onnxruntime model benchmark."""
+    benchmark_name = 'onnxruntime-ort-models'
+    (benchmark_class,
+     predefine_params) = BenchmarkRegistry._BenchmarkRegistry__select_benchmark(benchmark_name, Platform.ROCM)
+    assert (benchmark_class)
+    benchmark = benchmark_class(benchmark_name)
+    assert (benchmark._benchmark_type == BenchmarkType.DOCKER)
+    assert (benchmark._image_uri == 'superbench/benchmark:rocm4.3.1-onnxruntime1.9.0')
+    assert (benchmark._container_name == 'rocm-onnxruntime-model-benchmarks')
+    assert (benchmark._entrypoint == '/stage/onnxruntime-training-examples/huggingface/azureml/run_benchmark.sh')
+    assert (benchmark._cmd is None)
+    benchmark._result = BenchmarkResult(benchmark._name, benchmark._benchmark_type, ReturnCode.SUCCESS)
+
+    raw_output = """
+__superbench__ begin bert-large-uncased ngpu=1
+    "samples_per_second": 21.829
+__superbench__ begin bert-large-uncased ngpu=8
+    "samples_per_second": 147.181
+__superbench__ begin distilbert-base-uncased ngpu=1
+    "samples_per_second": 126.827
+__superbench__ begin distilbert-base-uncased ngpu=8
+    "samples_per_second": 966.796
+__superbench__ begin gpt2 ngpu=1
+    "samples_per_second": 20.46
+__superbench__ begin gpt2 ngpu=8
+    "samples_per_second": 151.089
+__superbench__ begin facebook/bart-large ngpu=1
+    "samples_per_second": 66.171
+__superbench__ begin facebook/bart-large ngpu=8
+    "samples_per_second": 370.343
+__superbench__ begin roberta-large ngpu=1
+    "samples_per_second": 37.103
+__superbench__ begin roberta-large ngpu=8
+    "samples_per_second": 274.455
+"""
+    assert (benchmark._process_raw_result(0, raw_output))
+    assert (benchmark.result['bert_large_uncased_ngpu_1'][0] == 21.829)
+    assert (benchmark.result['bert_large_uncased_ngpu_8'][0] == 147.181)
+    assert (benchmark.result['distilbert_base_uncased_ngpu_1'][0] == 126.827)
+    assert (benchmark.result['distilbert_base_uncased_ngpu_8'][0] == 966.796)
+    assert (benchmark.result['gpt2_ngpu_1'][0] == 20.46)
+    assert (benchmark.result['gpt2_ngpu_8'][0] == 151.089)
+    assert (benchmark.result['facebook_bart_large_ngpu_1'][0] == 66.171)
+    assert (benchmark.result['facebook_bart_large_ngpu_8'][0] == 370.343)
+    assert (benchmark.result['roberta_large_ngpu_1'][0] == 37.103)
+    assert (benchmark.result['roberta_large_ngpu_8'][0] == 274.455)