__init__.py 2.21 KB
Newer Older
1
2
3
4
5
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT License.

"""A module containing all the micro-benchmarks."""

6
from superbench.benchmarks.micro_benchmarks.micro_base import MicroBenchmark, MicroBenchmarkWithInvoke
7
from superbench.benchmarks.micro_benchmarks.sharding_matmul import ShardingMatmul
8
from superbench.benchmarks.micro_benchmarks.computation_communication_overlap import ComputationCommunicationOverlap
9
from superbench.benchmarks.micro_benchmarks.kernel_launch_overhead import KernelLaunch
10
from superbench.benchmarks.micro_benchmarks.cublas_function import CublasBenchmark
11
from superbench.benchmarks.micro_benchmarks.cudnn_function import CudnnBenchmark
12
13
from superbench.benchmarks.micro_benchmarks.gemm_flops_performance_base import GemmFlopsBenchmark
from superbench.benchmarks.micro_benchmarks.cuda_gemm_flops_performance import CudaGemmFlopsBenchmark
14
from superbench.benchmarks.micro_benchmarks.memory_bw_performance_base import MemBwBenchmark
15
from superbench.benchmarks.micro_benchmarks.cuda_memory_bw_performance import CudaMemBwBenchmark
16
from superbench.benchmarks.micro_benchmarks.disk_performance import DiskBenchmark
17
from superbench.benchmarks.micro_benchmarks.ib_loopback_performance import IBLoopbackBenchmark
18
from superbench.benchmarks.micro_benchmarks.cuda_nccl_bw_performance import CudaNcclBwBenchmark
19
from superbench.benchmarks.micro_benchmarks.rocm_memory_bw_performance import RocmMemBwBenchmark
20
from superbench.benchmarks.micro_benchmarks.rocm_gemm_flops_performance import RocmGemmFlopsBenchmark
21
from superbench.benchmarks.micro_benchmarks.gpu_copy_bw_performance import GpuCopyBwBenchmark
22
from superbench.benchmarks.micro_benchmarks.tcp_connectivity import TCPConnectivityBenchmark
23
from superbench.benchmarks.micro_benchmarks.gpcnet_performance import GPCNetBenchmark
24

25
__all__ = [
26
    'MicroBenchmark', 'MicroBenchmarkWithInvoke', 'ShardingMatmul', 'ComputationCommunicationOverlap', 'KernelLaunch',
27
    'CublasBenchmark', 'CudnnBenchmark', 'GemmFlopsBenchmark', 'CudaGemmFlopsBenchmark', 'MemBwBenchmark',
28
    'CudaMemBwBenchmark', 'DiskBenchmark', 'IBLoopbackBenchmark', 'CudaNcclBwBenchmark', 'RocmMemBwBenchmark',
29
    'RocmGemmFlopsBenchmark', 'GpuCopyBwBenchmark', 'TCPConnectivityBenchmark', 'GPCNetBenchmark'
30
]