LLM inference in C/C++
C++ ggml
sycl : pass device into mul_mat_uses_onednn (#31014)
Titaniumtownf64a49819ccb
diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cppindex ce9ef7dd23..ed0b272806 100644--- a/ggml/src/ggml-sycl/ggml-sycl.cpp+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp@@ -6868,7 +6868,7 @@ static bool ggml_sycl_graph_update_required(ggml_sycl_graph * graph, ggml_cgraph// Reports if ggml_sycl_mul_mat() gives this node to oneDNN. oneDNN chains the submission on events// made before recording started, which SYCL graphs do not allow.-static bool mul_mat_uses_onednn(ggml_tensor * dst) {+static bool mul_mat_uses_onednn(int device, ggml_tensor * dst) {#if GGML_SYCL_DNNLif (!g_ggml_sycl_enable_dnn) {return false;@@ -6894,7 +6894,7 @@ static bool mul_mat_uses_onednn(ggml_tensor * dst) {// ggml_sycl_op_mul_mat_sycl() is the last resort of the dispatch and the only user of oneDNN.// The reorder decision that ggml_sycl_mul_mat() makes between DMMV and MMVQ is left out here:// neither uses a library.- if (can_use_dequantize_mul_mat_vec(src0, src1, dst) || can_use_mul_mat_vec_q(src0, src1, dst) ||+ if (can_use_dequantize_mul_mat_vec(src0, src1, dst) || can_use_mul_mat_vec_q(device, src0, src1, dst) ||can_use_mul_mat_q(src0, src1, dst)) {return false;}@@ -6916,6 +6916,7 @@ static bool mul_mat_uses_onednn(ggml_tensor * dst) {const int64_t gemm_flops = src0->ne[1] * src1->ne[1] * src0->ne[0];return gemm_flops >= SYCL_F32_DNNL_MIN_GEMM_FLOPS;#else+ GGML_UNUSED(device);GGML_UNUSED(dst);return false;#endif@@ -6978,7 +6979,7 @@ static bool check_graph_compatibility(ggml_backend_sycl_context * ctx, ggml_cgrareturn false;}#ifndef GGML_SYCL_GRAPH_DNNL_SUPPORTED- if (mul_mat_uses_onednn(node)) {+ if (mul_mat_uses_onednn(ctx->device, node)) {GGML_LOG_DEBUG("%s: disabling SYCL graphs due to %s using oneDNN\n", __func__,ggml_op_name(node_op));return false;