Repository navigation

perf(hip): prompt attention tile keeps its probabilities in registers and fits 6 waves, same bits - #1788

Merged

ashhart merged 7 commits into

ashhart:mainfrom

AbrahamCete:perf/hip-prefill-attention

Nov 1, 2026

Merged

perf(hip): prompt attention tile keeps its probabilities in registers and fits 6 waves, same bits#1788
ashhart merged 7 commits into
ashhart:mainfrom
AbrahamCete:perf/hip-prefill-attention

Commits on Nov 1, 2026
- test(hip): pin the paged prompt tile's output bits on 27B shapes
  
  AbrahamCete
  authored and
  ashhart
  committed
  44917d6
- perf(hip): skip the masks and range guards on whole prompt tiles
  
  AbrahamCete
  authored and
  ashhart
  committed
  de0b764
- perf(hip): reduce the prompt tile's softmax rows with DPP butterflies
  
  AbrahamCete
  authored and
  ashhart
  committed
  709184e
- perf(hip): prefetch the prompt tile's next key and value chunks
  
  AbrahamCete
  authored and
  ashhart
  committed
  751f80f
- perf(hip): keep the prompt tile's probabilities in registers
  
  AbrahamCete
  authored and
  ashhart
  committed
  139d047
- test(hip): register the prompt attention tiles with the kernel contract
  
  AbrahamCete
  authored and
  ashhart
  committed
  5c3ba89
- changelog: #1788 under Unreleased
  ashhart
  committed
  7217ce7