Update ggml-vulkan.cpp

This commit is contained in:
Eve 2024-12-07 21:24:34 +00:00 committed by GitHub
parent 4a185ad39b
commit 984d4707f3
No known key found for this signature in database
GPG key ID: B5690EEEBB952194

View file

@ -1783,7 +1783,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
}
// mul mat vec
// computing four rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0
// computing four rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0.
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32", mul_mat_vec_f32_f32_f32_len, mul_mat_vec_f32_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32", mul_mat_vec_f16_f32_f32_len, mul_mat_vec_f16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {4, 1, 1}, {device->subgroup_size, 4}, 1, true);