diff options
author | Xiaoqiang Zheng <zhengxq@google.com> | 2016-05-11 09:28:36 -0800 |
---|---|---|
committer | TensorFlower Gardener <gardener@tensorflow.org> | 2016-05-11 10:32:20 -0700 |
commit | 939ede027be73ecafcc422371afe27dceccc720d (patch) | |
tree | 2ac4b3f35f4e0744d5993271a01185e6bcc8905e /tensorflow/stream_executor/stream_executor_pimpl.h | |
parent | 1f8fe742e11de53ccbb34d9fa540302156cb1655 (diff) |
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
Diffstat (limited to 'tensorflow/stream_executor/stream_executor_pimpl.h')
-rw-r--r-- | tensorflow/stream_executor/stream_executor_pimpl.h | 9 |
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_; } |