merge changes to solve the slowndown of gpu::gemm

src/targets/gpu/gemm.cpp

@@ -127,7 +127,11 @@ argument miopen_gemm::compute(context& ctx,
         auto alpha_r    = to_rocblas_type(as(op.alpha));
         auto beta_r     = to_rocblas_type(as(op.beta));
         auto to_pointer = [&](auto&& arg) { return to_rocblas_type(as.from(arg.data())); };
-        generic_rocblas_batched_gemm(as,
+        // call the strided implementation only if there are multiple matrices
+        if(batch_num > 1)
+        {
+            generic_rocblas_batched_gemm(
+                as,
                 ctx.get_stream().get_rocblas(),
                 transb ? rocblas_operation_transpose : rocblas_operation_none,
                 transa ? rocblas_operation_transpose : rocblas_operation_none,
@@ -142,11 +146,29 @@ argument miopen_gemm::compute(context& ctx,
                 lda,
                 m * k,
                 &beta_r,
-                                     is_3inputs ? to_pointer(args[3]) : to_pointer(args[2]),
+                to_pointer(args[2]),
                 ldc,
                 m * n,
                 batch_num);
-
+        }
+        else
+        {
+            generic_rocblas_gemm(as,
+                                 ctx.get_stream().get_rocblas(),
+                                 transb ? rocblas_operation_transpose : rocblas_operation_none,
+                                 transa ? rocblas_operation_transpose : rocblas_operation_none,
+                                 n,
+                                 m,
+                                 k,
+                                 &alpha_r,
+                                 to_pointer(args[1]),
+                                 ldb,
+                                 to_pointer(args[0]),
+                                 lda,
+                                 &beta_r,
+                                 to_pointer(args[2]),
+                                 ldc);
+        }
     });
 
     return (is_3inputs ? args[3] : args[2]);