llama.cpp
vulkan: optimize conv2d and implement coopmat1 support
#22620
Merged

vulkan: optimize conv2d and implement coopmat1 support #22620

0cc4m merged 8 commits into ggml-org:master from jeffbolznv:conv2d_opts_cm1
jeffbolznv
jeffbolznv vulkan: add CONV_SHAPE_64x128 for medium-K conv2d
f03632ec
jeffbolznv vulkan: skip conv2d bounds checks when shapes align with tile sizes
6b8052d3
jeffbolznv vulkan: use WG_SIZE=128 for CONV_SHAPE_64x32 conv2d
44bcf25a
jeffbolznv vulkan: stage cm2 conv2d accumulator through shmem before global store
bb856b42
jeffbolznv vulkan: add coopmat1 conv2d path
701effc6
jeffbolznv fallback when using too much shared memory. clean up comments
676ac3a9
jeffbolznv jeffbolznv requested a review 55 days ago
github-actions github-actions added Vulkan
github-actions github-actions added ggml
jeffbolznv Require 16x16x16 and subgroup size 32 or 64
ad41a1dc
0cc4m
0cc4m
0cc4m approved these changes on 2026-05-17
jeffbolznv check whether shared memory is sufficient before overwriting conv2d p…
7b9aa522
0cc4m
0cc4m approved these changes on 2026-05-26
0cc4m
CISC
CISC approved these changes on 2026-05-26
0cc4m 0cc4m merged 7799d31e into master 31 days ago

Login to write a write a comment.

Login via GitHub

Reviewers
Assignees
No one assigned
Labels
Milestone