[GPU] Use non-transposed output shape for gemm onednn

Author: kelvinchoi-intel

src/plugins/intel_gpu/src/graph/impls/onednn/gemm_onednn.cpp

@@ -272,7 +272,25 @@ struct gemm_onednn : typed_primitive_onednn_impl<gemm> {
 
         dnnl::memory::desc in0_md = get_input_memory_desc(in0_dims, in0_dt, in0_fmt, in0_strides);
         dnnl::memory::desc in1_md = get_input_memory_desc(in1_dims, in1_dt, in1_fmt, in1_strides);
-        dnnl::memory::desc out_md(out_dims, out_dt, out_fmt);
+
+        // Get non transposed output dims
+        auto output_order = prim->output_transpose_order;
+        bool b_output_transposed = false;
+        for (int64_t i = 0; i < static_cast<int64_t>(output_order.size()); i++) {
+            if (output_order[i] != i) {
+                b_output_transposed = true;
+                break;
+            }
+        }
+
+        dnnl::memory::dims non_transposed_out_dims = out_dims;
+        if (b_output_transposed) {
+            for (int64_t i = 0; i < static_cast<int64_t>(output_order.size()); i++) {
+                non_transposed_out_dims[i] = out_dims[output_order[i]];
+            }
+        }
+
+        dnnl::memory::desc out_md(non_transposed_out_dims, out_dt, out_fmt);
 
         if (gemm_with_bias) {
             dnnl::memory::desc bias_md(bias_dims, bias_dt, bias_fmt);

src/plugins/intel_gpu/src/graph/impls/onednn/gemm_onednn.hpp

@@ -72,14 +72,6 @@ struct GemmImplementationManager : public ImplementationManager {
         if (gemm_prim->indirect_a || gemm_prim->indirect_b)
             return false;
 
-        // Keep this condition until gemm_onednn supports transposed order of output
-        const int64_t OTO_SIZE = static_cast<int64_t>(gemm_prim->output_transpose_order.size());
-        if (OTO_SIZE > 0 &&
-            !(gemm_prim->output_transpose_order[OTO_SIZE - 2] == (OTO_SIZE - 2) &&
-            gemm_prim->output_transpose_order[OTO_SIZE - 1] == (OTO_SIZE - 1))) {
-            return false;
-        }
-
         return true;
     }
 

src/plugins/intel_gpu/tests/functional/shared_tests_instances/subgraph_tests/transpose_matmul_fusion.cpp

@@ -187,4 +187,4 @@ INSTANTIATE_TEST_SUITE_P(smoke_MatMulTransposeFusion, MatMulTransposeFusionOnGPU
 TEST_P(MatMulTransposeFusionOnGPU, CompareWithRefs){
     run();
 };
-}  // namespace
+}  // namespace