Add the autotune version for the backward passes. It is currently

disabled by default, and can be enabled through the env-var
"TF_CUDNN_USE_AUTOTUNE=1". It will eventually be turned on by default.

The following is the benchmarks with large enough changes.

Benchmark                          Base (ns)  New (ns) Improvement
------------------------------------------------------------------
BM_ConvFloatFwdGPU_conv13            3810933   2167784    +43.1%
BM_ConvFloatFwdGPU_conv23            4173607   2450503    +41.3%
BM_ConvFloatFwdGPU_conv54           26731131   7098361    +73.4%
BM_ConvFloatBkInGPU_conv1            1496407   1039979    +30.5%
BM_ConvFloatBkInGPU_conv2            1501744    999774    +33.4%
BM_ConvFloatBkInGPU_conv12           6826426    968258    +85.8%
BM_ConvFloatBkFilterGPU_conv13       3852185   2110649    +45.2%
BM_ConvFloatBkInGPU_conv15           7011109    910837    +87.0%
BM_ConvFloatBkInGPU_conv17           2724054   1930013    +29.1%
BM_ConvFloatBkInGPU_conv18           2940634   1846089    +37.2%
BM_ConvFloatBkInGPU_conv19           2995599   1853970    +38.1%
BM_ConvFloatBkInGPU_conv22           2685772   1940984    +27.7%
BM_ConvFloatBkInGPU_conv24           2343034   1519468    +35.1%
BM_ConvFloatBkInGPU_conv27           2339471   1516779    +35.2%
BM_ConvFloatBkFilterGPU_conv28       3091452   1880773    +39.2%
BM_ConvFloatBkInGPU_conv31           1265237   1120846    +11.4%
BM_ConvFloatBkInGPU_conv46           3346414   2070659    +38.1%
BM_ConvFloatBkFilterGPU_conv52      20677347  14342254    +30.6%
BM_ConvFloatBkInGPU_conv54          13291278  10495521    +21.0%
Change: 122067373

1 files changed, 9 insertions, 0 deletions

diff --git a/tensorflow/stream_executor/stream_executor_pimpl.h b/tensorflow/stream_executor/stream_executor_pimpl.h
index 31b110a8e0..e424411143 100644
--- a/tensorflow/stream_executor/stream_executor_pimpl.h
+++ b/tensorflow/stream_executor/stream_executor_pimpl.h
@@ -341,6 +341,15 @@ class StreamExecutor {
   // Get the list of supported algorithms for the forward convolution opeartion.
   bool GetConvolveAlgorithms(std::vector<dnn::AlgorithmType> *out_algorithms);
 
+  // Get the list of supported algorithms for the backward convolution on data.
+  bool GetConvolveBackwardDataAlgorithms(
+      std::vector<dnn::AlgorithmType> *out_algorithms);
+
+  // Get the list of supported algorithms for the backward convolution on the
+  // filter.
+  bool GetConvolveBackwardFilterAlgorithms(
+      std::vector<dnn::AlgorithmType> *out_algorithms);
+
   // Returns the device ordinal that this StreamExecutor was initialized with.
   // Meaningless before initialization.
   int device_ordinal() const { return device_ordinal_; }