DP4MatMulNBits accuracy level 4 matmul for WebGPU EP #23365
guschmue
merged
19 commits into
main
from
user/sushraja/dp4_matmul
sushraja-msft
changed the title
sushraja-msft
changed the title
Initial commit with the dequantizer
5d972348
Enable the matmul
9f200293
Add comment
82236b8a
Fix quantize to support sg_size 8
af2bc783
Rearrange code to keep inputs in registry. On WU driver 179 tps or 2.…
b9df530e
Quantize now supports all subgroup sizes.
283f4b17
removing subgroup stuff from loadSHMA
61c1d607
loadSHMB should now work for all sg_size
05570aaa
Ready for sgsize implementation of perform_16_64_16 matmul
b739ad6f
Remove dequantize and read accuracy level. Not using accuracy level y…
429b7841
Add a no subgroup version, perf on adl 500tk
97a359b4
Ran the linter and added back the accuracy level check.
b690857e
Fix typos
ceddd082
Add additional check that N%16 == 0, because we dont check bounds dur…
23701b9a
Fix last typo
ab389917
Add subgroup check
73ee5d1e
qjia7
commented
on 2025-01-18
Reduce k tile size to 32, to relieve registr memory pressure.
48727757
add f32 support
524da06d
Disable dp4a matmul on macOS, since metal does not appear to have DP4…
d56b9c3d
guschmue
approved these changes
on 2025-01-21
sushraja-msft
changed the title
sushraja-msft
changed the title
Login to write a write a comment.
Login via GitHub