[QST] kInternalError while increasing warp count in older SIMT GEMM kernels.

[Shreya-gaur]

What is your question?

Internal CUTLASS error is observed, when I try increasing the warp count for kernel "cutlass_simt_hgemm_256x128_8x2_nt_align1" to values other than default 4x2x1 (by changing the warpshape accordingly in the generated kernel). In the cutlass_profiler, this presents itself as a Disposition failure. These are the values I tried:

1. 4x4x1
2. 8x2x1
3. 8x4x1
4. 16x2x1

How should I debug this? Is there any proper documentation on how older SIMT GEMM kernels work?

We want more warps scheduled per sub-core, any insights into how one can achieve this apart from just making the warp tiles smaller?

PS: When I try increasing the warp count for kernel "cutlass_simt_sgemm_128x128_8x2_nt_align1", everything works okay.

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[github-actions]

This issue has been labeled inactive-90d due to no recent activity in the past 90 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.