Benchmark: Micro benchmark - Add float datatype support and other refinements to GPU Stream

docs/user-tutorial/benchmarks/micro-benchmarks.md

@@ -267,20 +267,22 @@ For measurements of peer-to-peer communication performance between AMD GPUs, GPU
 
 #### Introduction
 
-Measure the memory bandwidth of GPU using the STREAM benchmark. The benchmark tests various memory operations including copy, scale, add, and triad for double datatype.
+Measure the memory bandwidth of GPU using the STREAM benchmark. The benchmark tests various memory operations including copy, scale, add, and triad for double and float datatypes.
+
+__Note__: When `--check_data` is enabled, each process allocates 6× `--size` bytes of host memory (data\_buf + check\_buf + 4 validation buffers, e.g. 24 GiB with the default 4 GiB `--size`). Under `default_local_mode` with 8 GPUs this totals ~192 GiB of host RAM. Recommend using a small `--size` such as `1048576` (1 MiB) when `--check_data` is enabled.
 
 #### Metrics
 
 | Metric Name                                                | Unit             | Description                                                                                                                             |
 |------------------------------------------------------------|------------------|-----------------------------------------------------------------------------------------------------------------------------------------|
-| STREAM\_COPY\_double\_gpu\_[0-9]\_buffer\_[0-9]+\_block\_[0-9]+\_bw  | bandwidth (GB/s) | The fp64 memory bandwidth of the GPU for the copy operation with specified buffer size and block size.                         |
-| STREAM\_SCALE\_double\_gpu\_[0-9]\_buffer\_[0-9]+\_block\_[0-9]+\_bw | bandwidth (GB/s) | The fp64 memory bandwidth of the GPU for the scale operation with specified buffer size and block size.                         |
-| STREAM\_ADD\_double\_gpu\_[0-9]\_buffer\_[0-9]+\_block\_[0-9]+\_bw   | bandwidth (GB/s) | The fp64 memory bandwidth of the GPU for the add operation with specified buffer size and block size.                         |
-| STREAM\_TRIAD\_double\_gpu\_[0-9]\_buffer\_[0-9]+\_block\_[0-9]+\_bw | bandwidth (GB/s) | The fp64 memory bandwidth of the GPU for the triad operation with specified buffer size and block size.                         |
-| STREAM\_COPY\_double\_gpu\_[0-9]\_buffer\_[0-9]+\_block\_[0-9]+\_ratio  | Efficiency (%) | The fp64 memory bandwidth efficiency of the GPU for the copy operation with specified buffer size and block size.                         |
-| STREAM\_SCALE\_double\_gpu\_[0-9]\_buffer\_[0-9]+\_block\_[0-9]+\_ratio | Efficiency (%) | The fp64 memory bandwidth efficiency of the GPU for the scale operation with specified buffer size and block size.                         |
-| STREAM\_ADD\_double\_gpu\_[0-9]\_buffer\_[0-9]+\_block\_[0-9]+\_ratio   | Efficiency (%) | The fp64 memory bandwidth efficiency of the GPU for the add operation with specified buffer size and block size.                         |
-| STREAM\_TRIAD\_double\_gpu\_[0-9]\_buffer\_[0-9]+\_block\_[0-9]+\_ratio | Efficiency (%) | The fp64 memory bandwidth efficiency of the GPU for the triad operation with specified buffer size and block size.                         |
+| STREAM\_COPY\_(double\|float)\_buffer\_[0-9]+\_block\_[0-9]+\_bw  | bandwidth (GB/s) | The memory bandwidth of the GPU for the copy operation with the selected data type (double for fp64, float for fp32), for the specified buffer size and block size.                         |
+| STREAM\_SCALE\_(double\|float)\_buffer\_[0-9]+\_block\_[0-9]+\_bw | bandwidth (GB/s) | The memory bandwidth of the GPU for the scale operation with the selected data type (double for fp64, float for fp32), for the specified buffer size and block size.                         |
+| STREAM\_ADD\_(double\|float)\_buffer\_[0-9]+\_block\_[0-9]+\_bw   | bandwidth (GB/s) | The memory bandwidth of the GPU for the add operation with the selected data type (double for fp64, float for fp32), for the specified buffer size and block size.                         |
+| STREAM\_TRIAD\_(double\|float)\_buffer\_[0-9]+\_block\_[0-9]+\_bw | bandwidth (GB/s) | The memory bandwidth of the GPU for the triad operation with the selected data type (double for fp64, float for fp32), for the specified buffer size and block size.                         |
+| STREAM\_COPY\_(double\|float)\_buffer\_[0-9]+\_block\_[0-9]+\_ratio  | Efficiency (%) | The memory bandwidth efficiency of the GPU for the copy operation with the selected data type (double for fp64, float for fp32), for the specified buffer size and block size.                         |
+| STREAM\_SCALE\_(double\|float)\_buffer\_[0-9]+\_block\_[0-9]+\_ratio | Efficiency (%) | The memory bandwidth efficiency of the GPU for the scale operation with the selected data type (double for fp64, float for fp32), for the specified buffer size and block size.                         |
+| STREAM\_ADD\_(double\|float)\_buffer\_[0-9]+\_block\_[0-9]+\_ratio   | Efficiency (%) | The memory bandwidth efficiency of the GPU for the add operation with the selected data type (double for fp64, float for fp32), for the specified buffer size and block size.                         |
+| STREAM\_TRIAD\_(double\|float)\_buffer\_[0-9]+\_block\_[0-9]+\_ratio | Efficiency (%) | The memory bandwidth efficiency of the GPU for the triad operation with the selected data type (double for fp64, float for fp32), for the specified buffer size and block size.                         |
 
 ### `ib-loopback`
 

examples/benchmarks/gpu_stream.py

@@ -12,7 +12,7 @@
 
 if __name__ == '__main__':
     context = BenchmarkRegistry.create_benchmark_context(
-        'gpu-stream', platform=Platform.CUDA, parameters='--num_warm_up 1 --num_loops 10'
+        'gpu-stream', platform=Platform.CUDA, parameters='--num_warm_up 1 --num_loops 10 --data_type double'
     )
     # For ROCm environment, please specify the benchmark name and the platform as the following.
     # context = BenchmarkRegistry.create_benchmark_context(

superbench/benchmarks/micro_benchmarks/gpu_stream.py

@@ -51,10 +51,21 @@ def add_parser_arguments(self):
             help='Number of data buffer copies performed.',
         )
 
+        self._parser.add_argument(
+            '--data_type',
+            type=str,
+            default='double',
+            choices=['float', 'double'],
+            required=False,
+            help='Data type of the buffer elements.',
+        )
+
         self._parser.add_argument(
             '--check_data',
             action='store_true',
-            help='Enable data checking',
+            help='Enable data checking. Note: allocates 6x --size bytes of host memory per process '
+            '(data_buf + check_buf + 4 validation buffers, e.g. 24 GiB with default 4 GiB --size). '
+            'Recommend using a small --size such as 1048576 (1 MiB) when this flag is enabled.',
         )
 
     def _preprocess(self):
@@ -68,8 +79,8 @@ def _preprocess(self):
 
         self.__bin_path = os.path.join(self._args.bin_dir, self._bin_name)
 
-        args = '--size %d --num_warm_up %d --num_loops %d ' % (
-            self._args.size, self._args.num_warm_up, self._args.num_loops
+        args = '--size %d --num_warm_up %d --num_loops %d --data_type %s' % (
+            self._args.size, self._args.num_warm_up, self._args.num_loops, self._args.data_type
         )
 
         if self._args.check_data:

superbench/benchmarks/micro_benchmarks/gpu_stream/CMakeLists.txt

@@ -29,7 +29,7 @@ message(STATUS "Found CUDA: " ${CUDAToolkit_VERSION})
 
 # Source files
 set(SOURCES
-    gpu_stream_test.cpp
+    gpu_stream_main.cpp
     gpu_stream_utils.cpp
     gpu_stream.cu
     gpu_stream_kernels.cu
@@ -38,7 +38,8 @@ set(SOURCES
 include(../cuda_common.cmake)
 add_executable(gpu_stream ${SOURCES})
 set_property(TARGET gpu_stream PROPERTY CUDA_ARCHITECTURES ${NVCC_ARCHS_SUPPORTED})
+target_compile_definitions(gpu_stream PRIVATE _GNU_SOURCE)
 target_include_directories(gpu_stream PRIVATE ${CUDAToolkit_INCLUDE_DIRS})
-target_link_libraries(gpu_stream numa ${NVML_LIBRARY})
+target_link_libraries(gpu_stream numa ${NVML_LIBRARY} ${CMAKE_DL_LIBS})
 
 install(TARGETS gpu_stream RUNTIME DESTINATION bin)