vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

sycl: Reroute permuted mul_mats through oneMKL

Browse source
This commit is contained in:
Alberto Cabrera 2024-11-19 14:15:41 +00:00
parent 3ee6382d48
commit 005c267481
1 changed files with 9 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

7
ggml/src/ggml-sycl/ggml-sycl.cpp
View file

@ -3447,8 +3447,15 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
if (!split && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
// TODO: Refactor and cleanup of mul mat dispatching.
if (src0->ne[3] == 1 && src1->ne[3] == 1) {
// KQ single-batch
// mmv p021 was specific for these dimensions
ggml_sycl_mul_mat_vec_p021(ctx, src0, src1, dst);
} else {
// The kernel from the if path is faster for that specific case, but does not support all mul mats.
ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
}
} else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
// KQV single-batch
ggml_sycl_mul_mat_vec_nc(ctx, src0, src1, dst);