vulkan: int8 coopmat1 matmul implementation for AMD RDNA3 and RDNA4 - #27952

Open

0cc4m wants to merge 50 commits into

0cc4m/vulkan-coopmat-int8

Commits on Aug 29, 2026

vulkan: add int8 coopmat quantized matmul shader
0cc4m
committed
13f6d8e
apply scales inline
0cc4m
committed
e086b9c
use scalar sums
0cc4m
committed
5f034d6
probe and directly access coopmat values instead of going through shmem

0cc4m
and
pwilkin
committed
1625608
add q8_0 support
0cc4m
committed
513f436
add BK_STEP to shader, default to 2

0cc4m
and
pwilkin
committed
b121ef3
use larger workgroups

0cc4m
and
pwilkin
committed
2a878d9
double buffering

0cc4m
and
pwilkin
committed
5553b89
preload scales
0cc4m
committed
b453be3
coopmat load first, then wmma
0cc4m
committed
14a6080
use float for scales
0cc4m
committed
95d14c9
add faster RDNA int->float conversion
0cc4m
committed
ad015d6
revert load reordering and scale pre-loading
0cc4m
committed
31b9bc6
workgroup scheduling for cache proximity
0cc4m
committed
2335676
clean up
0cc4m
committed
9f069f4
Revert "revert load reordering and scale pre-loading"
0cc4m
committed
15b3ce0
use wave32
0cc4m
committed
57b4276
increase large tile size
0cc4m
committed
a9e0e89
Revert "increase large tile size"
0cc4m
committed
ef5c629
restructure for vgpr use
0cc4m
committed
2b9c533
skip computation for inactive tiles
0cc4m
committed
57dd3f8
only force subgroup size 32 on AMD RDNA
0cc4m
committed
ab6e84b
use BK_STEP 4
0cc4m
committed
42a6269
fix compilation
0cc4m
committed
1825270
move quant-specific prefetch function out of main file
0cc4m
committed
eb46b8c
add q4_1, q5_0, q5_1 support
0cc4m
committed
d2a5ac6
restructure mmq cm1 functions
0cc4m
committed
ed294e8
enable mul_mat_id support
0cc4m
committed
fa37db9
fix segfault
0cc4m
committed
d8ed372
fix mul_mat_id bug
0cc4m
committed
3cd7643
support iq4_nl and mxfp4
0cc4m
committed
dc08fa9
remove elem row/col fast path, invalid for RDNA4
0cc4m
committed
ba8402e
use shmem arrays for LUTs
0cc4m
committed
5037f02
use 4-byte loads where possible
0cc4m
committed
ab94ffd
add q3_k, q4_k, q5_k, q6_k and nvfp4 support
0cc4m
committed
4695f10
fix l warptile
0cc4m
committed
22f588c
improve performance
0cc4m
committed
3985f64
improve performance
0cc4m
committed
092661d
improvements
0cc4m
committed
eeeaced
dedup b scales
0cc4m
committed
2ba1d22
merge shmem arrays
0cc4m
committed
9ef5996
undo uint8_t, gate to RDNA3/4
0cc4m
committed
f252045
add RDNA4 architecture, use for hardcoded coopmat elem thread access, set BK_STEP back to 4
0cc4m
committed
0eab519
improve offset application
0cc4m
committed
c30a874
clean up
0cc4m
committed
6421fbe
fix iq4_nl and nvfp4 performance
0cc4m
committed
04d7ef7
rdna4 tuning
0cc4m
committed
584ae05
use BK_STEP 2 on MUL_MAT_ID
0cc4m
committed
e85ff84
adapt to upstream changes
0cc4m
committed
8b194a1
fix shmem support function, clean up comments
0cc4m
committed
965e571