Qun Lin c3228aaf0d [ck] support remap 32x32 warp tile to 16x16 ...

This PR remap 32x32 warp tile to 16x16 warp tile for all CK kernels in wave32. the logic is same with ROCm/composable_kernel#3421. and the most change is in device classes.

To reduece the instance build time, VGPR estimation is implemented in ~10 gridwise classes. and to pass all test in CI, several tests are minor adjusted.

2026-01-12 14:30:09 +08:00

..

ck

[ck] support remap 32x32 warp tile to 16x16

2026-01-12 14:30:09 +08:00

ck_tile

Dlejeune/ck tile 2d multiple reductions (#3147)

2026-01-09 11:16:37 +01:00

rapidjson

Update pre-commit to fixed versions, run remod for ck_tile (#2895)

2025-10-16 15:29:17 -07:00