CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-18 11:46:58 +00:00
Code Issues Packages Projects Releases Wiki Activity

Revert "sycl: add usage of enqueue_functions extension (#14244)" (#15910)

Browse Source 
* Revert "sycl: add usage of enqueue_functions extension (#14244)"

This reverts commit 8308f98c7f.

* fix missed revert code, format the code
This commit is contained in:
Neo Zhang Jianyu
2025-09-12 09:15:12 +08:00
committed by GitHub
parent 360d6533db
commit 704d90c987
20 changed files with 845 additions and 674 deletions
Download Patch File Download Diff File Expand all files Collapse all files

140
ggml/src/ggml-sycl/mmq.cpp
View File

@@ -1818,7 +1818,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_qs_q4_0_acc_ct1(
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<float, 1> tile_x_d_q4_0_acc_ct1(
@@ -1829,8 +1829,9 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q4_0<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -1852,7 +1853,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_qs_q4_0_acc_ct1(
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<float, 1> tile_x_d_q4_0_acc_ct1(
@@ -1863,8 +1864,9 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q4_0<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -1931,7 +1933,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_qs_q4_1_acc_ct1(
sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh);
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_1_acc_ct1(
@@ -1942,8 +1944,9 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q4_1<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -1965,7 +1968,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_qs_q4_1_acc_ct1(
sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh);
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_1_acc_ct1(
@@ -1976,8 +1979,9 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q4_1<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2044,7 +2048,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_ql_q5_0_acc_ct1(
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<float, 1> tile_x_d_q5_0_acc_ct1(
@@ -2055,8 +2059,9 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q5_0<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2078,7 +2083,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_ql_q5_0_acc_ct1(
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<float, 1> tile_x_d_q5_0_acc_ct1(
@@ -2089,8 +2094,9 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q5_0<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2157,7 +2163,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_ql_q5_1_acc_ct1(
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_1_acc_ct1(
@@ -2168,8 +2174,9 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q5_1<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2191,7 +2198,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_ql_q5_1_acc_ct1(
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_1_acc_ct1(
@@ -2202,8 +2209,9 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q5_1<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2270,7 +2278,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_qs_q8_0_acc_ct1(
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<float, 1> tile_x_d_q8_0_acc_ct1(
@@ -2281,8 +2289,9 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q8_0<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2304,7 +2313,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_qs_q8_0_acc_ct1(
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<float, 1> tile_x_d_q8_0_acc_ct1(
@@ -2315,8 +2324,9 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q8_0<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2383,7 +2393,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_ql_q2_K_acc_ct1(
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q2_K_acc_ct1(
@@ -2396,8 +2406,9 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q2_K<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2420,7 +2431,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_ql_q2_K_acc_ct1(
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q2_K_acc_ct1(
@@ -2433,8 +2444,9 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q2_K<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2504,7 +2516,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_ql_q3_K_acc_ct1(
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q3_K_acc_ct1(
@@ -2519,8 +2531,9 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q3_K<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2544,7 +2557,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_ql_q3_K_acc_ct1(
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q3_K_acc_ct1(
@@ -2559,8 +2572,9 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q3_K<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2630,7 +2644,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_ql_q4_K_acc_ct1(
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_K_acc_ct1(
@@ -2643,8 +2657,9 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q4_K<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2667,7 +2682,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_ql_q4_K_acc_ct1(
sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_K_acc_ct1(
@@ -2680,8 +2695,9 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q4_K<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2749,7 +2765,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_ql_q5_K_acc_ct1(
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_K_acc_ct1(
@@ -2762,8 +2778,9 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q5_K<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2786,7 +2803,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_ql_q5_K_acc_ct1(
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_K_acc_ct1(
@@ -2799,8 +2816,9 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q5_K<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2868,7 +2886,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_ql_acc_ct1(
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<sycl::half2, 1> tile_x_dm_acc_ct1(
@@ -2881,8 +2899,9 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q6_K<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,
@@ -2905,7 +2924,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
dpct::has_capability_or_fail(stream->get_device(),
{sycl::aspect::fp16});
sycl_launch(stream, [&](sycl::handler & cgh) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<int, 1> tile_x_ql_acc_ct1(
sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
sycl::local_accessor<sycl::half2, 1> tile_x_dm_acc_ct1(
@@ -2918,8 +2937,9 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
sycl_parallel_for(
cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
mul_mat_q6_K<need_check>(
vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
nrows_dst, item_ct1,