diff options
| author |  Yangzihao Wang <yangzihao@google.com> | 2017-05-28 14:20:19 -0700 |
|---|---|---|
| committer |  TensorFlower Gardener <gardener@tensorflow.org> | 2017-05-28 14:23:23 -0700 |
| commit | e78e5ec8a8c862e65b6a194e9caea377120d7207 (patch) | |
| tree | ef713546bfdacced118056167b49de0c4c2f1a36 /tensorflow/core/kernels/conv_grad_input_ops.cc | |
| parent | 3f9b69a50f40154f6078e1610ce7d3afa94bd07c (diff) | |
Set winograd nofused flag to be true by default.
Disable winograd nonfused conv for certain input params to avoid a known bug in cuDNNv5 and cuDNNv6.
PiperOrigin-RevId: 157352847
Diffstat (limited to 'tensorflow/core/kernels/conv_grad_input_ops.cc')
| -rw-r--r-- | tensorflow/core/kernels/conv_grad_input_ops.cc | 3 |
1 files changed, 2 insertions, 1 deletions
diff --git a/tensorflow/core/kernels/conv_grad_input_ops.cc b/tensorflow/core/kernels/conv_grad_input_ops.cc index 9a50431a2f..a94b1bea4b 100644 --- a/tensorflow/core/kernels/conv_grad_input_ops.cc +++ b/tensorflow/core/kernels/conv_grad_input_ops.cc @@ -856,7 +856,8 @@ class Conv2DSlowBackpropInputOp : public OpKernel { if (cudnn_use_autotune_ && !AutoTuneConvBwdData::GetInstance()->Find( conv_parameters, &algorithm_config)) { std::vector<AlgorithmType> algorithms; - CHECK(stream->parent()->GetConvolveBackwardDataAlgorithms(&algorithms)); + CHECK(stream->parent()->GetConvolveBackwardDataAlgorithms( + conv_parameters.ShouldIncludeWinogradNonfusedAlgo<T>(), &algorithms)); ProfileResult best_result; ProfileResult best_result_no_scratch; for (auto profile_algorithm : algorithms) { |