perf(pa tile): readable tile-programming PA decode kernel, matches/beats production#825

Merged

coderfeli merged 78 commits into

mainfrom

Jul 24, 2026

Commits on Jun 29, 2026

Add tile-programming PA-decode reference kernel

fsx950223
and
claude
committed
Make PA-decode tile output accumulator register-resident

fsx950223
and
claude
committed
Distribute PA-decode softmax phase-2 across all 64 lanes

fsx950223
and
claude
committed

Commits on Jun 30, 2026

Reorient PA-decode tile to token=M (production) layout

fsx950223
and
claude
committed
Add K prefetch at occupancy 3 via TLOOP softmax (pa_decode_ps structure)

fsx950223
and
claude
committed
Hoist V load before QK while keeping K prefetch (hide both loads)

fsx950223
and
claude
committed

Commits on Jul 3, 2026

Commits on Jul 6, 2026

perf(pa tile): epilogue vectorization, HW exp2, V page-index prefetch, LDS bank-conflict fix

fsx950223
and
claude
committed

Commits on Jul 8, 2026

perf(pa tile): head_dim 64/128 support, match K/V cache layout to pa_decode_ps_kernel, register-carry m/l
fsx950223
committed

Commits on Jul 9, 2026

perf(pa tile): bf16 query, scalar KV-scale via buffer_load, closer match to pa_decode_ps_kernel's ISA

fsx950223
and
claude
committed

Commits on Jul 10, 2026

Commits on Jul 15, 2026

Commits on Jul 16, 2026

Commits on Jul 17, 2026

Commits on Jul 21, 2026

Commits on Jul 22, 2026

Commits on Jul 23, 2026

Merge remote-tracking branch 'origin/main' into pr-825
Felix Li
committed