diff --git a/onnxruntime/core/providers/cpu/rnn/deep_cpu_gru.cc b/onnxruntime/core/providers/cpu/rnn/deep_cpu_gru.cc
index e34f4c8848..64fe3cec62 100644
--- a/onnxruntime/core/providers/cpu/rnn/deep_cpu_gru.cc
+++ b/onnxruntime/core/providers/cpu/rnn/deep_cpu_gru.cc
@@ -383,15 +383,14 @@ Status DeepCpuGruOp::ComputeImpl(OpKernelContext& context) const {
     gsl::span<T> hidden_output_2 = hidden_output.subspan(hidden_output_size_per_direction,
                                                          hidden_output_size_per_direction);
 
-#ifndef USE_MKLDNN
-
+#if defined(USE_MLAS) && !defined(USE_OPENMP)
 #ifdef USE_EIGEN_THREADPOOL
     auto fn =
 #else
     std::packaged_task<void()> task_fw{
 #endif
         [&]() {
-#endif  // ! USE_MKLDNN
+#endif  // USE_MLAS && ! USE_OPENMP
           std::unique_ptr<detail::UniDirectionalGru<T>> fw = std::make_unique<detail::UniDirectionalGru<T>>(
               alloc, logger,
               seq_length, batch_size, input_size, hidden_size_, linear_before_reset_, Direction::kForward,
@@ -400,8 +399,8 @@ Status DeepCpuGruOp::ComputeImpl(OpKernelContext& context) const {
               activation_funcs_.Entries()[1],
               clip_, ttp_);
           fw->Compute(input, sequence_lens_span, num_directions_, input_weights_1, recurrent_weights_1, output_1, hidden_output_1);
-#ifndef USE_MKLDNN
 
+#if defined(USE_MLAS) && !defined(USE_OPENMP)
 #ifndef USE_EIGEN_THREADPOOL
         }
   };
@@ -420,8 +419,7 @@ Status DeepCpuGruOp::ComputeImpl(OpKernelContext& context) const {
       cv.notify_one();
     });
 #endif  // USE_EIGEN_THREADPOOL
-
-#endif  // ! USE_MKLDNN
+#endif  // USE_MLAS && ! USE_OPENMP
 
   std::unique_ptr<detail::UniDirectionalGru<T>> bw = std::make_unique<detail::UniDirectionalGru<T>>(
       alloc, logger,
@@ -432,15 +430,14 @@ Status DeepCpuGruOp::ComputeImpl(OpKernelContext& context) const {
       clip_, ttp_);
   bw->Compute(input, sequence_lens_span, num_directions_, input_weights_2, recurrent_weights_2, output_2, hidden_output_2);
 
-#ifndef USE_MKLDNN
+#if defined(USE_MLAS) && !defined(USE_OPENMP)
 #ifdef USE_EIGEN_THREADPOOL
   auto ul = std::unique_lock<OrtMutex>(lock);
-    if (!done) cv.wait(ul);
+  if (!done) cv.wait(ul);
 #else
     task_results_fw.get();
 #endif
-
-#endif  // ! USE_MKLDNN
+#endif  // USE_MLAS && ! USE_OPENMP
 }  // namespace onnxruntime
 else {
   std::unique_ptr<detail::UniDirectionalGru<T>> gru_p = std::make_unique<detail::UniDirectionalGru<T>>(