Skip to content

ggml-vulkan : tune L-tile warp micro-dimension for RDNA3 iGPUs #28611

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Troncooooo wants to merge 2 commits into
base: master
Choose a base branch
from
Open

ggml-vulkan : tune L-tile warp micro-dimension for RDNA3 iGPUs #28611

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
60b099c
ggml-vulkan : tune L-tile warp micro-dimension for RDNA3 iGPUs
Troncooooo Sep 6, 2026
c76d7d8
Merge branch 'ggml-org:master' into ggml-vk-rdna3-igpu-ltile
Troncooooo Sep 8, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 12 additions & 0 deletions ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4525,6 +4525,18 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
l_warptile = { 256, 128, 128, 16, mm_warp_8, 64, 2, tm_m, tn_m, tk_m, mm_warp_8 };
l_warptile_mmq = l_warptile_mmq_int = { 256, 128, 128, 32, mm_warp_8, 64, 2, tm_m, tn_m, tk_m, mm_warp_8 };
l_warptile_mmq_int_k = { 256, 128, 128, 32, mm_warp_16, 64, 1, 4, 2, 1, mm_warp_16 };
if (device->architecture == AMD_RDNA3 && device->uma) {
// RDNA3 integrated GPU (e.g. Radeon 780M, shared system memory): the 64x64 warp
// micro-tile is LDS/register limited on the KHR coopmat path; halving the
// M micro-dimension (WM 64 -> 32) substantially speeds up prompt processing,
// while token generation (mul_mat_vec path) is unaffected. Measured on
// RADV PHOENIX (gfx1151) with llama-bench -p 512 -n 64 -r 3:
// Qwen3.8-9B-Q4_K_M: pp512 250.5 -> 404.4 t/s (+61%); tg64 14.42 -> 14.43
// Qwen3.8-27B-UD-Q4_K_M: pp512 72.4 -> 117.5 t/s (+62%); tg64 4.73 -> 4.73
// (WN 64 -> 32 is equally good; WMITER 2 -> 4/8 is within noise.)
l_warptile = { 256, 128, 128, 16, 32, 64, 2, tm_m, tn_m, tk_m, mm_warp_8 };
l_warptile_mmq = l_warptile_mmq_int = { 256, 128, 128, 32, 32, 64, 2, tm_m, tn_m, tk_m, mm_warp_8 };
}
} else if (device->vendor_id == VK_VENDOR_ID_INTEL && device->coopmat_support) {
// Xe2/Xe3 with coopmat enabled - warptile performance tuning
l_warptile = { 512, 128, 128, 16, mm_warp_8, 32, 2, tm_m, tn_m, tk_m, mm_warp_8 };
Expand Down