EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

vulkan: mul_mat_id coopmat2 optimizations (#15546)

Browse Source 
* vulkan: mul_mat_id coopmat2 optimizations

Add a path for when the tile fits in BN/2, similar to what we have for mul_mat.

Only call fetch_scales/store_scales once per QUANT_K block, and once at the
beginning in case start_k is not aligned.

* Also add a path for BN/4 - worth a couple more percent
This commit is contained in:
Jeff Bolz
2025-08-31 02:06:43 -05:00
committed by GitHub
parent 5c16b9c87d
commit c37052ab4d
2 changed files with 93 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@@ -2225,7 +2225,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
s_mmq_wg_denoms_k = { 32, 64, 1 };
// spec constants and tile sizes for quant matmul_id
l_warptile_mmqid = { 256, 128, 128, 16, 0, device->subgroup_size };
l_warptile_mmqid = { 256, 128, 128, 16, 1, device->subgroup_size };
m_warptile_mmqid = { 256, 128, 64, 16, 0, device->subgroup_size };
s_warptile_mmqid = { 256, 128, 64, 16, 0, device->subgroup_size };
l_mmqid_wg_denoms = { 128, 128, 1 };