You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
What would you like to be added:
For GEMM-FLOPs test with ROCm, pass flags to rocblas-bench specify what random tensor generation is used. --initialization rand_int --initialization hpl
Why is this needed: --initialization rand_int to use simple random tensor generation, which results relatively more 0 values --initialization hpl to use HPL style random tensor generation, which results relatively less 0 values
Without this feature, how does current superbenchmark work:
more 0 values results better benchmark results, which is by default for ROCm before 5.1 release
What would you like to be added:
For GEMM-FLOPs test with ROCm, pass flags to rocblas-bench specify what random tensor generation is used.
--initialization rand_int
--initialization hpl
Why is this needed:
--initialization rand_int
to use simple random tensor generation, which results relatively more 0 values--initialization hpl
to use HPL style random tensor generation, which results relatively less 0 valuesWithout this feature, how does current superbenchmark work:
more 0 values results better benchmark results, which is by default for ROCm before 5.1 release
Components that may involve changes:
GEMM-FLOPs
Brief description of your proposal if any:
reference:
https://github.com/ROCmSoftwarePlatform/rocBLAS/releases/tag/rocm-5.1.0
https://ontrack.amd.com/browse/MSRCHA-325
ROCm/rocBLAS@9e9ced4
The text was updated successfully, but these errors were encountered: