ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

vulkan: Add N/2 and N/4 optimized paths in coopmat2 shader (#12312)

Browse source
This commit is contained in:
Jeff Bolz 2025-03-17 09:26:18 -05:00 • committed by GitHub
parent cf2270e4d3
commit 484a8ab513
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 72 additions and 31 deletions
Download patch file Download diff file Expand all files Collapse all files

24
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file

@ -1597,33 +1597,33 @@ static void ggml_vk_load_shaders(vk_device& device) {
uint32_t l_align, m_align, s_align;
if (device->coopmat2) {
// spec constants and tile sizes for non-quant matmul/matmul_id
l_warptile = { 256, 128, 256, 64 };
m_warptile = { 256, 128, 128, 64 };
s_warptile = { 128, 64, 64, 64 };
l_warptile = { 256, 128, 256, 64, 1 };
m_warptile = { 256, 128, 128, 64, 0 };
s_warptile = { 128, 64, 64, 64, 0 };
l_wg_denoms = {128, 256, 1 };
m_wg_denoms = {128, 128, 1 };
s_wg_denoms = { 64, 64, 1 };
// spec constants and tile sizes for quant matmul (non-Qi_K)
l_warptile_mmq = { 256, 128, 256, 64 };
m_warptile_mmq = { 256, 128, 128, 64 };
s_warptile_mmq = { 256, 32, 64, 128 };
l_warptile_mmq = { 256, 128, 256, 64, 1 };
m_warptile_mmq = { 256, 128, 128, 64, 1 };
s_warptile_mmq = { 256, 32, 64, 128, 0 };
l_mmq_wg_denoms = { 128, 256, 1 };
m_mmq_wg_denoms = { 128, 128, 1 };
s_mmq_wg_denoms = { 32, 64, 1 };
// spec constants and tile sizes for quant matmul (Qi_K)
l_warptile_mmq_k = { 256, 64, 128, 64 };
m_warptile_mmq_k = { 256, 32, 64, 64 };
s_warptile_mmq_k = { 256, 32, 32, 128 };
l_warptile_mmq_k = { 256, 64, 128, 64, 1 };
m_warptile_mmq_k = { 256, 32, 64, 64, 0 };
s_warptile_mmq_k = { 256, 32, 32, 128, 0 };
l_mmq_wg_denoms_k = { 64, 128, 1 };
m_mmq_wg_denoms_k = { 32, 64, 1 };
s_mmq_wg_denoms_k = { 32, 32, 1 };
// spec constants and tile sizes for quant matmul_id
l_warptile_mmqid = { 256, 128, 64, 16 };
m_warptile_mmqid = { 256, 128, 64, 16 };
s_warptile_mmqid = { 256, 128, 64, 16 };
l_warptile_mmqid = { 256, 128, 64, 16, 0 };
m_warptile_mmqid = { 256, 128, 64, 16, 0 };
s_warptile_mmqid = { 256, 128, 64, 16, 0 };
l_mmqid_wg_denoms = { 128, 64, 1 };
m_mmqid_wg_denoms = { 128, 64, 1 };
s_mmqid_wg_denoms = { 128, 64, 1 };