Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

sycl : Optimize Q3_K mul_mat by reorder (#23725)

Browse Source
This commit is contained in:
Neo Zhang
2026-06-01 09:50:55 +03:00
committed by GitHub
parent af6528e6df
commit 44e211cecf
7 changed files with 340 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-sycl/convert.cpp
+23 -2
View File
@@ -107,6 +107,19 @@ static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int64_t k,
#endif
}
template <typename dst_t>
static void dequantize_row_q3_K_sycl_reorder(const void *vx, dst_t *y, const int64_t k,
dpct::queue_ptr stream) {
const int64_t nb = k / QK_K;
dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
stream->parallel_for(
sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
[=](sycl::nd_item<3> item_ct1) {
dequantize_block_q3_K_reorder(vx, y, item_ct1, nb);
});
}
template <typename dst_t>
static void dequantize_row_q4_0_sycl(const void *vx, dst_t *y, const int64_t k,
dpct::queue_ptr stream) {
@@ -652,7 +665,11 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
case GGML_TYPE_Q2_K:
return dequantize_row_q2_K_sycl;
case GGML_TYPE_Q3_K:
return dequantize_row_q3_K_sycl;
if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
return dequantize_row_q3_K_sycl_reorder;
} else {
return dequantize_row_q3_K_sycl;
}
case GGML_TYPE_Q4_K:
if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
return dequantize_row_q4_K_sycl_reorder;
@@ -730,7 +747,11 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
case GGML_TYPE_Q2_K:
return dequantize_row_q2_K_sycl;
case GGML_TYPE_Q3_K:
return dequantize_row_q3_K_sycl;
if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
return dequantize_row_q3_K_sycl_reorder;
} else {
return dequantize_row_q3_K_sycl;
}
case GGML_TYPE_Q4_K:
if (dst->src[0]->extra &&
((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) {