happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

use 2 ELEM_PER_THREAD for AMD/Intel

This commit is contained in:
Jeff Bolz 2026-01-01 14:13:31 -06:00
parent 878b9685de
commit 3adb5f63c4
1 changed files with 2 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -4172,7 +4172,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
ggml_vk_create_pipeline(device, device->pipeline_cumsum_f32, "cumsum_f32", cumsum_f32_len, cumsum_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { 256, device->subgroup_size, 4 }, 1, true, true, device->subgroup_size);
const uint32_t CUMSUM_ELEM_PER_THREAD = device->vendor_id == VK_VENDOR_ID_NVIDIA ? 4 : 2;
ggml_vk_create_pipeline(device, device->pipeline_cumsum_f32, "cumsum_f32", cumsum_f32_len, cumsum_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { 256, device->subgroup_size, CUMSUM_ELEM_PER_THREAD }, 1, true, true, device->subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_cumsum_small_f32, "cumsum_f32", cumsum_f32_len, cumsum_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { 128, device->subgroup_size, 1 }, 1, true, true, device->subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_cumsum_multipass1_f32, "cumsum_multipass1_f32", cumsum_multipass1_f32_len, cumsum_multipass1_f32_data, "main", 3, sizeof(vk_op_sum_rows_push_constants), {256, 1, 1}, { 256, device->subgroup_size }, 1, true, true, device->subgroup_size);
ggml_vk_create_pipeline(device, device->pipeline_cumsum_multipass2_f32, "cumsum_multipass2_f32", cumsum_multipass2_f32_len, cumsum_multipass2_f32_data, "main", 3, sizeof(vk_op_sum_rows_push_constants), {256, 1, 1}, { 256, device->subgroup_size }, 1, true, true, device->subgroup_size);