Adam Paszke 6d6c8c2e6c [Mosaic GPU] Write a higher-level tcgen05.mma helper reusing WGMMA implementation
Hopper and Blackwell MMA instructions can share a lot of the same logic, which is
why I ended up splitting out a large fraction of WGMMA implementation into a common
utility. This should be an NFC for WGMMA, but it allows us to concisely implement
unrolling of MMAs of different sizes into a number of tcgen05.mma instructions.

PiperOrigin-RevId: 723544349
2025-02-05 09:38:29 -08:00
..
2024-12-29 13:06:19 +00:00
2024-12-18 19:38:37 -08:00