From 23e122f3048d61e39c01ecfb22c2d88b0466d323 Mon Sep 17 00:00:00 2001
From: Yufeng Li <liyufeng1987@gmail.com>
Date: Fri, 7 Dec 2018 10:24:57 -0800
Subject: [PATCH] Run backward gru in main thread for bidirection (#118)

---
 .../core/providers/cpu/rnn/deep_cpu_gru.cc    | 25 +++++++------------
 1 file changed, 9 insertions(+), 16 deletions(-)

diff --git a/onnxruntime/core/providers/cpu/rnn/deep_cpu_gru.cc b/onnxruntime/core/providers/cpu/rnn/deep_cpu_gru.cc
index 9be1d1ffd0..3ff4584820 100644
--- a/onnxruntime/core/providers/cpu/rnn/deep_cpu_gru.cc
+++ b/onnxruntime/core/providers/cpu/rnn/deep_cpu_gru.cc
@@ -374,21 +374,6 @@ 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);
 
-    // run backward first as it needs an extra reverse sequence operation
-    std::packaged_task<void()> task_bw{
-        [&]() {
-          std::unique_ptr<detail::UniDirectionalGru<T>> bw = std::make_unique<detail::UniDirectionalGru<T>>(
-              alloc, logger,
-              seq_length, batch_size, input_size, hidden_size_, linear_before_reset_, Direction::kReverse,
-              bias_2, initial_hidden_2,
-              activation_funcs_.Entries()[2],
-              activation_funcs_.Entries()[3],
-              clip_, ttp_);
-          bw->Compute(input, sequence_lens_span, num_directions_, input_weights_2, recurrent_weights_2, output_2, hidden_output_2);
-        }};
-    auto task_results_bw = task_bw.get_future();
-    ttp_.RunTask(std::move(task_bw));
-
     std::packaged_task<void()> task_fw{
         [&]() {
           std::unique_ptr<detail::UniDirectionalGru<T>> fw = std::make_unique<detail::UniDirectionalGru<T>>(
@@ -403,8 +388,16 @@ Status DeepCpuGruOp::ComputeImpl(OpKernelContext& context) const {
     auto task_results_fw = task_fw.get_future();
     ttp_.RunTask(std::move(task_fw));
 
+    std::unique_ptr<detail::UniDirectionalGru<T>> bw = std::make_unique<detail::UniDirectionalGru<T>>(
+        alloc, logger,
+        seq_length, batch_size, input_size, hidden_size_, linear_before_reset_, Direction::kReverse,
+        bias_2, initial_hidden_2,
+        activation_funcs_.Entries()[2],
+        activation_funcs_.Entries()[3],
+        clip_, ttp_);
+    bw->Compute(input, sequence_lens_span, num_directions_, input_weights_2, recurrent_weights_2, output_2, hidden_output_2);
+
     task_results_fw.get();
-    task_results_bw.get();
   } else {
     std::unique_ptr<detail::UniDirectionalGru<T>> gru_p = std::make_unique<detail::UniDirectionalGru<T>>(
         alloc, logger,