Jun Liu
AFAIK Navi21 only support one warpSize. Should we enforce it for all kernels? ref to #1407 @atamazov
@JehandadKhan how should we maintain the doc at https://rocmsoftwareplatform.github.io/MIOpen/doc/html/index.html please? It seems that it was last updated for 2.11
@qianfengz The Workaround should be sufficient for ROCm 5.1. Could you implement the full "fix" as suggested above as a follow-up? Thanks!
> @junliume I can do this, if you wish Yes it would be much appreciated :)
It looks like these numbers are troublesome: > MIOpen(HIP): miopenStatus_t miopenGetTensorDescriptorSize(miopenTensorDescriptor_t, int *){ MIOpen(HIP): tensorDesc = 32, 192, 27, 27 MIOpen(HIP): size = 0 MIOpen(HIP): } MIOpen(HIP): miopenStatus_t miopenGet4dTensorDescriptorLengths(miopenTensorDescriptor_t, int...
@atamazov Tensordesc is correct while the NCHW numbers are interpreted incorrectly, (if this is the case) unlikely it won't be caught sooner I guess. > MIOpen(HIP): tensorDesc = 32, 192,...
@carlushuang @atamazov and @ce1adon any recommendations for enhancement or to fix this verification issue?
> You only see errors in "-F 2" is because by running backward only you are missing info calculated from forward. @DrizztDoUrden could we follow up a bit on the...
@JehandadKhan could we provide some way of on-site tuning for the user?
@xinlipn please help watch the CI and progress of the PR. Thanks!