Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-19 01:04:55 +02:00
Code Issues Packages Projects Releases Wiki Activity

use wave32

Browse Source
This commit is contained in:
Ruben Ortlam
2026-08-24 12:24:21 +02:00
parent b5251372d5
commit 870c77988b
1 changed files with 7 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+7 -6
View File
@@ -4277,9 +4277,10 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
s_warptile_mmq_int = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, itm_s, itn_s, itk_s, subgroup_size_8 };
// Coopmat int8 cm1 shader uses larger workgroups for better occupancy
l_warptile_mmq_cm1_int = { subgroup_size_8 * 8, 128, 128, 32, 64, 32, 2, itm_l, itn_l, itk_l, subgroup_size_8 };
m_warptile_mmq_cm1_int = { subgroup_size_8 * 4, 64, 64, 32, 32, 32, 2, itm_m, itn_m, itk_m, subgroup_size_8 };
s_warptile_mmq_cm1_int = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, itm_s, itn_s, itk_s, subgroup_size_8 };
// 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 };
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
l_warptile_mmq_int_k = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 1, 4, 4, 1, subgroup_size_8 };
@@ -4743,11 +4744,11 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
#define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
if (device->mul_mat ## ID ## _l[TYPE]) \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, true); \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, true, 32); \
if (device->mul_mat ## ID ## _m[TYPE]) \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, true); \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, true, 32); \
if (device->mul_mat ## ID ## _s[TYPE]) \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, true); \
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, true, 32); \
// Create 2 variants, {f16,f32} accumulator
#define CREATE_MM2(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \