Repository navigation

feat(rocke): async K loop, cshuffle, fp16 split-K and two-stage split-K for gfx1250 WMMA wgrad - #14832

Merged

cgmillette merged 22 commits into

developfrom

users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit

Nov 6, 2026

Merged

feat(rocke): async K loop, cshuffle, fp16 split-K and two-stage split-K for gfx1250 WMMA wgrad#14832
cgmillette merged 22 commits into
developfrom
users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit

Commits on Nov 1, 2026
- feat(rocke): async-to-LDS K loop and cshuffle epilogue for gfx1250 WMMA wgrad
  
  jakpiase
  and
  claude
  committed
  746a9e2
- fix(rocke): lower the packed fp16 atomic add as a monotonic atomicrmw
  
  jakpiase
  and
  claude
  committed
  5e47821
- feat(rocke): split fp16 wgrad automatically on gfx1250 again
  
  jakpiase
  and
  claude
  committed
  9c87577
- fix(rocke): rank async-DMA wgrad cache entries by their recorded chunk widths
  
  jakpiase
  and
  claude
  committed
  2f7afdc
- feat(rocke): two-stage (f32 workspace) split-K wgrad on gfx1250 WMMA
  
  jakpiase
  and
  claude
  committed
  537bf56
- feat(rocke): AOT-cache the gfx1250 WMMA wgrad cshuffle, two-stage and async variants
  
  jakpiase
  and
  claude
  committed
  363f522
- fix(rocke): gate the wgrad async K loop on global_load_async_to_lds in the job count
  
  jakpiase
  and
  claude
  committed
  d02b387
- fix(rocke): prune async-DMA cache kernels apart from synchronous ones
  
  jakpiase
  and
  claude
  committed
  4d5e200
- fix(rocke): cap the automatic gfx1250 fp16 wgrad split-K at the atomic's accuracy bound
  
  jakpiase
  and
  claude
  committed
  6e49bda
- fix(rocke): unsigned range check and tile-typed zero-fill in the async LDS loader
  
  jakpiase
  and
  claude
  committed
  cc12d0f
Commits on Nov 2, 2026
- Merge branch 'develop' into users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit
  jakpiase
  authored
  a69123a
Commits on Nov 3, 2026
Commits on Nov 4, 2026
Commits on Nov 5, 2026
- Merge branch 'develop' into users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit
  jakpiase
  authored
  b4a7d2a
- Merge remote-tracking branch 'origin/develop' into users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit
  jakpiase
  committed
  1f9d175

Navigation Menu

Repository navigation

feat(rocke): async K loop, cshuffle, fp16 split-K and two-stage split-K for gfx1250 WMMA wgrad - #14832

feat(rocke): async K loop, cshuffle, fp16 split-K and two-stage split-K for gfx1250 WMMA wgrad#14832
cgmillette merged 22 commits into
developfrom
users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit

Commits on Nov 1, 2026

feat(rocke): async-to-LDS K loop and cshuffle epilogue for gfx1250 WMMA wgrad

fix(rocke): lower the packed fp16 atomic add as a monotonic atomicrmw

feat(rocke): split fp16 wgrad automatically on gfx1250 again

fix(rocke): rank async-DMA wgrad cache entries by their recorded chunk widths

feat(rocke): two-stage (f32 workspace) split-K wgrad on gfx1250 WMMA

feat(rocke): AOT-cache the gfx1250 WMMA wgrad cshuffle, two-stage and async variants

fix(rocke): gate the wgrad async K loop on global_load_async_to_lds in the job count

fix(rocke): prune async-DMA cache kernels apart from synchronous ones

fix(rocke): cap the automatic gfx1250 fp16 wgrad split-K at the atomic's accuracy bound

fix(rocke): unsigned range check and tile-typed zero-fill in the async LDS loader

Commits on Nov 2, 2026

Merge branch 'develop' into users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit

Commits on Nov 3, 2026

Merge remote-tracking branch 'origin/develop' into users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit

Merge remote-tracking branch 'origin/develop' into users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit

Merge remote-tracking branch 'origin/develop' into users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit

fix(rocke): reject explicit wgrad A/B load-width caps on the async K loop in C++

Commits on Nov 4, 2026

Merge branch 'develop' into users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit

Merge branch 'develop' into users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit

fix(rocke): apply explicit wgrad A/B load-width caps in the C++ builder

test(rocke): pin the smaller of the auto width and the wgrad A/B cap

Merge remote-tracking branch 'origin/develop' into users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit

Commits on Nov 5, 2026

Merge branch 'develop' into users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit

Merge remote-tracking branch 'origin/develop' into users/jakpiase/rocke/gfx1250-wgrad-wmma-implicit