vulkan: opt mul_mat_vecq for mi50 #22933
0cc4m
merged
36 commits into
ggml-org:master
from
chraac:dev-mi50
add hip preset
8baf4250
ggml-vulkan: add half-precision dequantization functions and update m…
1ef9ea1d
add settings configuration and update Vulkan shader for improved iter…
0fad0731
refactor: simplify iter function parameters and optimize loop unrolli…
12c7a06e
refactor: optimize compute_outputs function by using col_stride for c…
61a5fde9
refactor: improve spacing for readability and simplify loop unrolling…
e3432966
Merge branch 'master' into dev-mi50
68b930c9
Revert "refactor: improve spacing for readability and simplify loop u…
42b91145
refactor: enhance mul_q8_1 function for half-precision support and up…
40e43680
Merge branch 'master' into dev-mi50
1bc6a290
feat: add support for subgroup operations on AMD GCN GPUs in Vulkan
b6f7d97f
wip
b2a9d86e
revert unused changes
eb84512d
fix: correct subgroup operations enabling condition for AMD GCN GPUs
78719835
fix: correct type casting in mul_q8_1_hf function for accurate calcul…
e2f43bd3
Potential fix for pull request finding
b9f5db99
fix: update regex for subgroup operations to include Instinct MI seri…
854c6439
wip
bb76c28f
fix: unify cache_b_ds definition and initialization for consistency
ca2f39b9
fix: using extension ... enable clause for fp16
b8b5ea64
Revert "fix: using extension ... enable clause for fp16"
24b26938
fix: enhance support for float16 in Vulkan shader pipelines and relat…
5422311d
rename
3e7843e2
Merge branch 'master' into dev-mi50
df1114a1
Merge branch 'master' into dev-mi50
db328228
Merge branch 'master' into dev-mi50
4d561285
Merge branch 'master' into dev-mi50
a28c9cb6
Revert "rename"
8184b3e2
revert the fp16 changes
f7a675ba
Merge branch 'master' into dev-mi50
d9d44205
refactor: simplify subgroup operations check for AMD GCN 5.0/5.1 GPUs
37c29bc9
si6Och4j
commented
on 2026-06-09
fix: update MI50 device check to use driver version instead of device ID
440eadec
fix: update subgroup operations check for AMD GCN GPUs to exclude Mes…
c64fb820
wip
9a251ae6
chraac
commented
on 2026-06-12
Merge branch 'master' into dev-mi50
364dbac6
fix: remove GCN subgroup operations check for Vulkan device
c6350b80
0cc4m
approved these changes
on 2026-06-23
CISC
approved these changes
on 2026-06-26
Reviewers
Assignees
No one assigned
Labels
Vulkan
ggml
Milestone
No milestone
Login to write a write a comment.
Login via GitHub