happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

modify warptile tuning for bf16 matmuls to fix regression (m_warptile to l_warptile)

This commit is contained in:
Viraj Wadhwa 2025-12-30 13:07:54 -07:00
parent dc147928ea
commit 328116e62f
1 changed files with 7 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -3528,6 +3528,13 @@ static void ggml_vk_load_shaders(vk_device& device) {
m_wg_denoms = { 64, 64, 1 };
s_wg_denoms = { 32, 32, 1 };
if (device->vendor_id == VK_VENDOR_ID_INTEL) {
if (device->architecture == INTEL_XE2) {
// Xe2/Xe3 - bf16 warptile performance tuning
l_warptile = { 512, 128, 128, 16, subgroup_size_8, 32, 2, 4, 4, 1, subgroup_size_8 };
}
}
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, , 0);
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id, 0);
}