From d7c2a04bce764764647008ae31be0c35aab75152 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Alberto=20Cabrera=20P=C3=A9rez?= Date: Fri, 29 Nov 2024 09:49:43 +0000 Subject: [PATCH] sycl : Reroute permuted mul_mats through oneMKL (llama/10408) This PR fixes the failing MUL_MAT tests for the sycl backend. --- ggml/src/ggml-sycl/ggml-sycl.cpp | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp index b6392ed8..aabcdc22 100644 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp @@ -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) { - // KQ single-batch - ggml_sycl_mul_mat_vec_p021(ctx, src0, src1, dst); + // 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);