Skip to content

Commit 75da350

Browse files
0cc4m0cc4m
committed
vulkan: tune mul_mat_vecq performance for Intel
1 parent 14ccd94 commit 75da350

File tree

1 file changed

+12
-11
lines changed

1 file changed

+12
-11
lines changed

‎ggml/src/ggml-vulkan/ggml-vulkan.cpp‎

Lines changed: 12 additions & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -2973,18 +2973,19 @@ static void ggml_vk_load_shaders(vk_device& device) {
29732973

29742974
#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
29752975
if (device->integer_dot_product) {
2976-
if (device->subgroup_arithmetic && device->subgroup_require_full_support && device->vendor_id != VK_VENDOR_ID_INTEL) {
2977-
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_q8_1_f32_subgroup_len, mul_mat_vec_q4_0_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true, true);
2978-
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_q8_1_f32_subgroup_len, mul_mat_vec_q4_1_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true, true);
2979-
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_q8_1_f32_subgroup_len, mul_mat_vec_q5_0_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true, true);
2980-
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_q8_1_f32_subgroup_len, mul_mat_vec_q5_1_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true, true);
2981-
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_q8_1_f32_subgroup_len, mul_mat_vec_q8_0_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true, true);
2976+
const uint32_t subgroup_size = (device->subgroup_size_control && device->vendor_id == VK_VENDOR_ID_INTEL) ? device->subgroup_min_size : device->subgroup_size;
2977+
if (device->subgroup_arithmetic && device->subgroup_require_full_support) {
2978+
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_q8_1_f32_subgroup_len, mul_mat_vec_q4_0_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size, 2*rm_stdq, i+1}, 1, true, true, subgroup_size);
2979+
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_q8_1_f32_subgroup_len, mul_mat_vec_q4_1_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size, 2*rm_stdq, i+1}, 1, true, true, subgroup_size);
2980+
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_q8_1_f32_subgroup_len, mul_mat_vec_q5_0_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size, 2*rm_stdq, i+1}, 1, true, true, subgroup_size);
2981+
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_q8_1_f32_subgroup_len, mul_mat_vec_q5_1_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size, 2*rm_stdq, i+1}, 1, true, true, subgroup_size);
2982+
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_q8_1_f32_subgroup_len, mul_mat_vec_q8_0_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {subgroup_size, 1*rm_stdq, i+1}, 1, true, true, subgroup_size);
29822983
} else {
2983-
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_q8_1_f32_len, mul_mat_vec_q4_0_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
2984-
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_q8_1_f32_len, mul_mat_vec_q4_1_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
2985-
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_q8_1_f32_len, mul_mat_vec_q5_0_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
2986-
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_q8_1_f32_len, mul_mat_vec_q5_1_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
2987-
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_q8_1_f32_len, mul_mat_vec_q8_0_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true);
2984+
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_q8_1_f32_len, mul_mat_vec_q4_0_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size, 2*rm_stdq, i+1}, 1, true);
2985+
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_q8_1_f32_len, mul_mat_vec_q4_1_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size, 2*rm_stdq, i+1}, 1, true);
2986+
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_q8_1_f32_len, mul_mat_vec_q5_0_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size, 2*rm_stdq, i+1}, 1, true);
2987+
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_q8_1_f32_len, mul_mat_vec_q5_1_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size, 2*rm_stdq, i+1}, 1, true);
2988+
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_q8_1_f32_len, mul_mat_vec_q8_0_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {subgroup_size, 1*rm_stdq, i+1}, 1, true);
29882989
}
29892990
}
29902991
#endif // GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT

0 commit comments

Comments
 (0)