Feature/deterministic (#11205)

* "fix deterministic"

* "fix ci"

* "fix init"

paddle/fluid/operators/conv_cudnn_op.cu.cc

@@ -20,7 +20,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/cudnn_helper.h"
 #include "paddle/fluid/platform/float16.h"
 
-DEFINE_bool(cudnn_algo_use_autotune, true,
+DEFINE_bool(cudnn_deterministic, true,
             "Whether allow using an autotuning algorithm for convolution "
             "operator. The autotuning algorithm may be non-deterministic. If "
             "false, the algorithm is deterministic.");
@@ -272,7 +272,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
     auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
     auto handle = dev_ctx.cudnn_handle();
     if (input_grad) {
-      if (FLAGS_cudnn_algo_use_autotune) {
+      if (FLAGS_cudnn_deterministic) {
         PADDLE_ENFORCE(
             platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
                 handle, cudnn_filter_desc,
@@ -297,7 +297,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
     }
 
     if (filter_grad) {
-      if (FLAGS_cudnn_algo_use_autotune) {
+      if (FLAGS_cudnn_deterministic) {
         PADDLE_ENFORCE(
             platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
                 handle, cudnn_input_desc, cudnn_output_grad_desc,

paddle/fluid/operators/pool_cudnn_op.cu.cc

@@ -135,7 +135,11 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel<T> {
 
     PoolingMode pooling_mode;
     if (pooling_type == "max") {
+      if (FLAGS_cudnn_deterministic) {
+        pooling_mode = PoolingMode::kMaximumDeterministic;
+      } else {
         pooling_mode = PoolingMode::kMaximum;
+      }
     } else {
       pooling_mode = PoolingMode::kAverage;
     }

paddle/fluid/platform/cudnn_helper.h

@@ -22,6 +22,8 @@ limitations under the License. */
 #include "paddle/fluid/platform/float16.h"
 #include "paddle/fluid/platform/macros.h"
 
+DECLARE_bool(cudnn_deterministic);
+
 namespace paddle {
 namespace platform {
 
@@ -76,8 +78,22 @@ enum class DataLayout {  // Not use
 enum class PoolingMode {
   kMaximum,
   kAverage,
+  kMaximumDeterministic,
 };
 
+inline cudnnPoolingMode_t GetPoolingMode(const PoolingMode& mode) {
+  switch (mode) {
+    case PoolingMode::kMaximumDeterministic:
+      return CUDNN_POOLING_MAX_DETERMINISTIC;
+    case PoolingMode::kAverage:
+      return CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING;
+    case PoolingMode::kMaximum:
+      return CUDNN_POOLING_MAX;
+    default:
+      PADDLE_THROW("Unexpected pooling mode.");
+  }
+}
+
 template <typename T>
 class CudnnDataType;
 
@@ -293,9 +309,7 @@ class ScopedPoolingDescriptor {
     PADDLE_ENFORCE_EQ(kernel.size(), pads.size());
     PADDLE_ENFORCE_EQ(kernel.size(), strides.size());
     PADDLE_ENFORCE(dynload::cudnnSetPoolingNdDescriptor(
-        desc_, (mode == PoolingMode::kMaximum
-                    ? CUDNN_POOLING_MAX
-                    : CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING),
+        desc_, (GetPoolingMode(mode)),
         CUDNN_PROPAGATE_NAN,  // Always propagate nans.
         kernel.size(), kernel.data(), pads.data(), strides.data()));
     return desc_;

python/paddle/fluid/__init__.py

@@ -120,7 +120,7 @@ def __bootstrap__():
     ]
     if core.is_compiled_with_cuda():
         read_env_flags += [
-            'fraction_of_gpu_memory_to_use', 'cudnn_algo_use_autotune'
+            'fraction_of_gpu_memory_to_use', 'cudnn_deterministic'
         ]
     core.init_gflags([sys.argv[0]] +
                      ["--tryfromenv=" + ",".join(read_env_flags)])