Ali Hassani eee0cab26c Stamp out 1x1x1 clusters, 128x256 CTA shape (#1665) ...

Adds 128x256 tile shapes to FP16/BF16 and FP8 generators.
Also adds 1x1x1 clusters to all existing FP16/BF16/FP8 generators.

NOTE: it is important to set kernel filter (--kernels /
CUTLASS_LIBRARY_KERNELS) to a non empty string and skip pruning to get
all of the new configurations.

If profiling exhaustively, they can be set to `*`.

Number of CUTLASS 3.X GEMMs before this commit: 2868
Number of CUTLASS 3.X GEMMs after this commit: 4016

Co-authored-by: Ali Hassani <ahassani@nvidia.com>

2024-07-31 20:22:29 -04:00

284 KiB

Raw Blame History

View Raw