Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-20 01:31:31 +02:00
Code Issues Packages Projects Releases Wiki Activity

increase large tile size

Browse Source
This commit is contained in:
Ruben Ortlam
2026-08-24 12:41:33 +02:00
parent 870c77988b
commit 7fabc25c5e
1 changed files with 8 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+8 -4
View File
@@ -4190,6 +4190,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
std::array<uint32_t, 3> l_wg_denoms, m_wg_denoms, s_wg_denoms,
l_mmq_wg_denoms, m_mmq_wg_denoms, s_mmq_wg_denoms,
l_mmq_wg_denoms_k, m_mmq_wg_denoms_k, s_mmq_wg_denoms_k,
l_mmq_cm1_wg_denoms, m_mmq_cm1_wg_denoms, s_mmq_cm1_wg_denoms,
l_mmqid_wg_denoms, m_mmqid_wg_denoms, s_mmqid_wg_denoms;
uint32_t l_align, m_align, s_align;
@@ -4278,8 +4279,8 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
// Coopmat int8 cm1 shader uses larger workgroups for better occupancy
// Force wave32 for VOPD on RDNA3+
l_warptile_mmq_cm1_int = { 512, 128, 128, 32, 32, 32, 2, itm_l, itn_l, itk_l, 32 };
m_warptile_mmq_cm1_int = { 128, 64, 64, 32, 32, 32, 2, itm_m, itn_m, itk_m, 32 };
l_warptile_mmq_cm1_int = { 640, 128, 160, 32, 32, 32, 2, itm_l, itn_l, itk_l, 32 };
m_warptile_mmq_cm1_int = { 256, 128, 64, 32, 32, 32, 2, itm_m, itn_m, itk_m, 32 };
s_warptile_mmq_cm1_int = { 32, 32, 32, 32, 32, 32, 2, itm_s, itn_s, itk_s, 32 };
// K-quants use even more registers, mitigate by setting WMITER to 1
@@ -4321,6 +4322,9 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 };
m_mmq_wg_denoms = m_wg_denoms = { 64, 64, 1 };
s_mmq_wg_denoms = s_wg_denoms = { 32, 32, 1 };
l_mmq_cm1_wg_denoms = {128, 160, 1 };
m_mmq_cm1_wg_denoms = {128, 64, 1 };
s_mmq_cm1_wg_denoms = { 32, 32, 1 };
l_align = 128;
m_align = 64;
s_align = 32;
@@ -4808,8 +4812,8 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
}
if (device->coopmat_int_support) {
CREATE_MMQ2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_0], matmul_q4_0_q8_1, mmq_wg_denoms, warptile_mmq_cm1_int, vk_mat_mat_push_constants, 3, );
CREATE_MMQ2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q8_0], matmul_q8_0_q8_1, mmq_wg_denoms, warptile_mmq_cm1_int, vk_mat_mat_push_constants, 3, );
CREATE_MMQ2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_0], matmul_q4_0_q8_1, mmq_cm1_wg_denoms, warptile_mmq_cm1_int, vk_mat_mat_push_constants, 3, );
CREATE_MMQ2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q8_0], matmul_q8_0_q8_1, mmq_cm1_wg_denoms, warptile_mmq_cm1_int, vk_mat_mat_push_constants, 3, );
}
GGML_ASSERT(device->subgroup_ballot);