optimize the calling of rocmblas function.

src/targets/gpu/gemm.cpp

@@ -117,6 +117,9 @@ argument miopen_gemm::compute(context& ctx,
         auto alpha_r    = to_rocblas_type(as(alpha));
         auto beta_r     = to_rocblas_type(as(beta));
         auto to_pointer = [&](auto&& arg) { return to_rocblas_type(as.from(arg.data())); };
+        // 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,
@@ -136,7 +139,25 @@ argument miopen_gemm::compute(context& ctx,
                                         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 args[2];