mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

vulkan: Fix Windows performance regression on Intel GPU BF16 workloads for Xe2 and newer (#22461)

Browse Source 
* refactor

* Use l_warptile only when coopamt is available for BF16
This commit is contained in:
Masato Nakasaka
2026-05-12 03:15:34 -07:00
committed by GitHub
parent 706fbd8ab6
commit ef93e98d01
1 changed files with 8 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+8 -13
View File
@@ -4260,11 +4260,6 @@ static void ggml_vk_load_shaders(vk_device& device) {
m_wg_denoms = { 64, 64, 1 }; m_wg_denoms = { 64, 64, 1 };
s_wg_denoms = { 32, 32, 1 }; s_wg_denoms = { 32, 32, 1 };
if (device->vendor_id == VK_VENDOR_ID_INTEL && device->architecture == INTEL_XE2) {
// Xe2/Xe3 - bf16 warptile performance tuning
l_warptile = { 512, 128, 128, 16, subgroup_size_8, 32, 2, 4, 4, 1, subgroup_size_8 };
}
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, , 0); CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, , 0);
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0); CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
} }
@@ -5689,19 +5684,19 @@ static vk_device ggml_vk_get_device(size_t idx) {
device->mul_mat_id_m[i] = true; device->mul_mat_id_m[i] = true;
device->mul_mat_id_s[i] = true; device->mul_mat_id_s[i] = true;
break; break;
case VK_VENDOR_ID_INTEL: case VK_VENDOR_ID_INTEL: {
if (!device->coopmat_support || device->architecture != INTEL_XE2) { // Current Windows driver does not expose BF16 support.
device->mul_mat_l[i] = false; // We only want to use l_warptile if coopmat is available and is Xe2+
device->mul_mat_id_l[i] = false; const bool xe2_with_coopmat = device->coopmat_support && device->architecture == INTEL_XE2;
} else { const bool use_l_warptile = (i == GGML_TYPE_BF16) ? (device->coopmat_bf16_support && xe2_with_coopmat) : xe2_with_coopmat;
device->mul_mat_l[i] = true; // if coopmat & XE2+, allow large matmul warptile config for Intel device->mul_mat_l[i] = use_l_warptile;
device->mul_mat_id_l[i] = true; device->mul_mat_id_l[i] = use_l_warptile;
}
device->mul_mat_m[i] = true; device->mul_mat_m[i] = true;
device->mul_mat_s[i] = true; device->mul_mat_s[i] = true;
device->mul_mat_id_m[i] = true; device->mul_mat_id_m[i] = true;
device->mul_mat_id_s[i] = true; device->mul_mat_id_s[i] = true;
break; break;
}
case VK_VENDOR_ID_APPLE: case VK_VENDOR_ID_APPLE:
device->mul_mat_l[i] = false; device->mul_mat_l[i] = false;
device->mul_mat_m[i] = true; device->mul_mat_m[i] = true;