Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-18 08:44:49 +02:00
Code Issues Packages Projects Releases Wiki Activity

use larger workgroups

Browse Source
This commit is contained in:
Ruben Ortlam
2026-08-24 09:57:40 +02:00
parent 1a97c2c54d
commit 39f4463627
2 changed files with 12 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+8 -2
View File
@@ -4181,6 +4181,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
l_warptile_id, m_warptile_id, s_warptile_id,
l_warptile_mmq, m_warptile_mmq, s_warptile_mmq,
l_warptile_mmq_int, m_warptile_mmq_int, s_warptile_mmq_int,
l_warptile_mmq_cm1_int, m_warptile_mmq_cm1_int, s_warptile_mmq_cm1_int,
l_warptile_mmq_int_k, m_warptile_mmq_int_k, s_warptile_mmq_int_k,
l_warptile_mmq_k, m_warptile_mmq_k, s_warptile_mmq_k,
l_warptile_mmqid, m_warptile_mmqid, s_warptile_mmqid,
@@ -4275,6 +4276,11 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
m_warptile_mmq_int = { 128, 64, 64, 32, subgroup_size_8, 32, 2, itm_m, itn_m, itk_m, subgroup_size_8 };
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 };
// 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 };
m_warptile_mmq_int_k = { 128, 64, 64, 32, subgroup_size_8, 32, 1, 2, 2, 1, subgroup_size_8 };
@@ -4801,8 +4807,8 @@ 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_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_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_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, );
}
GGML_ASSERT(device->subgroup_ballot);
ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_cm1.comp
+4 -4
View File
@@ -67,11 +67,11 @@ layout (push_constant) uniform parameter
#endif
} p;
layout (constant_id = 0) const uint BLOCK_SIZE = 64;
layout (constant_id = 1) const uint BM = 64;
layout (constant_id = 2) const uint BN = 64;
layout (constant_id = 0) const uint BLOCK_SIZE = 256;
layout (constant_id = 1) const uint BM = 128;
layout (constant_id = 2) const uint BN = 128;
// layout (constant_id = 3) const uint BK = 32;
layout (constant_id = 4) const uint WM = 32;
layout (constant_id = 4) const uint WM = 64;
layout (constant_id = 5) const uint WN = 32;
layout (constant_id = 6) const uint WMITER = 2;
layout (constant_id = 7) const uint TM = 16;