From c5774e328255ddb91fe9f28c27777a96eeeacb90 Mon Sep 17 00:00:00 2001
From: chengduoZH <zhaochengduo@163.com>
Date: Sat, 28 Apr 2018 10:23:44 +0800
Subject: [PATCH] add FLAGS_use_deterministic_algo

---
 paddle/fluid/operators/conv_cudnn_op.cu.cc | 49 ++++++++++++++--------
 python/paddle/fluid/__init__.py            |  8 +++-
 2 files changed, 38 insertions(+), 19 deletions(-)

diff --git a/paddle/fluid/operators/conv_cudnn_op.cu.cc b/paddle/fluid/operators/conv_cudnn_op.cu.cc
index c70e3cc3c9..cf410c3ca1 100644
--- a/paddle/fluid/operators/conv_cudnn_op.cu.cc
+++ b/paddle/fluid/operators/conv_cudnn_op.cu.cc
@@ -20,6 +20,11 @@ limitations under the License. */
 #include "paddle/fluid/platform/cudnn_helper.h"
 #include "paddle/fluid/platform/float16.h"
 
+DEFINE_bool(cudnn_algo_use_autotune, true,
+            "Whether allow using an autotuning algorithm for convolution "
+            "operator. The autotuning algorithm may be non-deterministic. If "
+            "false, the algorithm is deterministic.");
+
 namespace paddle {
 namespace operators {
 
@@ -267,17 +272,23 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
     auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
     auto handle = dev_ctx.cudnn_handle();
     if (input_grad) {
-      PADDLE_ENFORCE(
-          platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
-              handle, cudnn_filter_desc,
-              // dyDesc: Handle to the previously initialized input differential
-              // tensor descriptor.
-              cudnn_output_grad_desc, cudnn_conv_desc,
-              // dxDesc: Handle to the previously initialized output tensor
-              // descriptor.
-              cudnn_input_desc,
-              CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
-              workspace_size_limit, &data_algo));
+      if (FLAGS_cudnn_algo_use_autotune) {
+        PADDLE_ENFORCE(
+            platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
+                handle, cudnn_filter_desc,
+                // dyDesc: Handle to the previously initialized input
+                // differential
+                // tensor descriptor.
+                cudnn_output_grad_desc, cudnn_conv_desc,
+                // dxDesc: Handle to the previously initialized output tensor
+                // descriptor.
+                cudnn_input_desc,
+                CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
+                workspace_size_limit, &data_algo));
+      } else {
+        data_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1;
+      }
+
       PADDLE_ENFORCE(
           platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
               handle, cudnn_filter_desc, cudnn_output_grad_desc,
@@ -286,12 +297,16 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
     }
 
     if (filter_grad) {
-      PADDLE_ENFORCE(
-          platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
-              handle, cudnn_input_desc, cudnn_output_grad_desc, cudnn_conv_desc,
-              cudnn_filter_desc,
-              CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
-              workspace_size_limit, &filter_algo));
+      if (FLAGS_cudnn_algo_use_autotune) {
+        PADDLE_ENFORCE(
+            platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
+                handle, cudnn_input_desc, cudnn_output_grad_desc,
+                cudnn_conv_desc, cudnn_filter_desc,
+                CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
+                workspace_size_limit, &filter_algo));
+      } else {
+        filter_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1;
+      }
 
       PADDLE_ENFORCE(
           platform::dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize(
diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py
index e2502990d5..c940f47478 100644
--- a/python/paddle/fluid/__init__.py
+++ b/python/paddle/fluid/__init__.py
@@ -107,8 +107,12 @@ def __bootstrap__():
     os.environ['OMP_NUM_THREADS'] = str(num_threads)
 
     read_env_flags = [
-        'use_pinned_memory', 'check_nan_inf', 'benchmark', 'warpctc_dir',
-        'eager_delete_scope'
+        'use_pinned_memory',
+        'check_nan_inf',
+        'benchmark',
+        'warpctc_dir',
+        'eager_delete_scope',
+        'cudnn_algo_use_autotune',
     ]
     if core.is_compiled_with_cuda():
         read_env_flags += ['fraction_of_gpu_memory_to_use']
-- 
GitLab