[webgpu] Use subgroup for matmulnbits #23224
guschmue
merged
17 commits into
microsoft:main
from
qjia7:webgpu_subgroup_matmulnbits
[webgpu] Add kernel type to profile info
3b32acc0
relax some limitations
8e0b27f9
[webgpu] Use override shape in shader key
c0dd1db2
subgroup support only for subgroup size = 32
a66dd014
support subgroup with size = 32
d3a196b6
subgroup size = 32, workgroup size = 64
99fe6977
put a/scale data into workgroup memory
0e850483
support subgroup size = 8/16/32/64
6feb70f6
workaround the issue that subgroupShuffle is very slow
55fc6bbe
add sg_size check
128cd7de
Revert "add sg_size check"
0775e4eb
apply subgroup for tile_m > 1
454fd0dc
check subgroup feature
16fc0757
only apply subgroup for intel GPUs
622bdd08
Revert "[webgpu] Use override shape in shader key"
729df58e
Revert "[webgpu] Add kernel type to profile info"
dff1961e
add more limitations to use subgroup
8328e149
qjia7
changed the title
guschmue
approved these changes
on 2025-01-12
Login to write a write a comment.
Login via GitHub