[gfx1250] Add 256x256 compute-bound GEMM (WIP) - #971

Draft

aoli26 wants to merge 19 commits into

mainfrom

gfx1250/gemm-opt-0805

Commits on Aug 6, 2026

refactor mxscale a8w8 gemm
aoli26
committed
7fa3132
add a8w4 256x256 compute-bound kernel
aoli26
committed
462fd44
relax the k constraint
aoli26
committed
8ab3a28

Commits on Aug 7, 2026

grouped grid rasterization
aoli26
committed
f0be547
add base a4w4 256x256 kernel
aoli26
committed
0a4eb77

Commits on Aug 8, 2026

optimize a8w4 cluster&pipeline
aoli26
committed
4d33be4
optimize a8w4 TDM supply one TDM per two K-tiles
aoli26
committed
3120216
optimize a4w4 cluster
aoli26
committed
a113838

Commits on Aug 9, 2026

optimiza epilogue
aoli26
committed
48a0bb0

Commits on Aug 10, 2026

fix cluster ragged-m
aoli26
committed
9a61bf1
optimzie a8w8 epilogue
aoli26
committed
84f4194
fix num buffer and perm
aoli26
committed
59c124e

Commits on Aug 11, 2026

hoist do_sync and optimize early timeout
aoli26
committed
9c70689
add base a8w8 256x256 kernel
aoli26
committed
f5482e9

Commits on Aug 13, 2026

Merge branch 'main' into gfx1250/gemm-opt-0805
aoli26
committed
449be79
tuning iglp and fix saturation test var
aoli26
committed
ac84c9e

Commits on Aug 14, 2026

tune tdm/ds phase pipeline
aoli26
committed
0919aa7

Commits on Aug 17, 2026

interleave a8w8 LDS fragment loads
aoli26
committed
ce4a71b
128x256 profiles and open up the cluster shape
aoli26
committed
a732527