update

paddle/fluid/operators/batch_norm_op.cc

@@ -83,7 +83,7 @@ class BatchNormOp : public framework::OperatorWithKernel {
 
  protected:
   framework::OpKernelType GetExpectedKernelType(
-      const ExecutionContext &ctx) const override {
+      const framework::ExecutionContext &ctx) const override {
     auto input_data_type =
         framework::ToDataType(ctx.Input<Tensor>("X")->type());
     // For float or float16 input tensor, the type of the scale, bias, mean,

paddle/fluid/operators/batch_norm_op.cu.cc

@@ -28,7 +28,7 @@ using DataLayout = framework::DataLayout;
 template <typename T>
 using CudnnDataType = platform::CudnnDataType<T>;
 template <typename T>
-using ScalingParamType = typename CudnnDataType<T>::ScalingParamType;
+using BatchNormParamType = typename CudnnDataType<T>::BatchNormParamType;
 
 void ExtractNCWHD(const framework::DDim &dims, const DataLayout &data_layout,
                   int *N, int *C, int *H, int *W, int *D) {
@@ -122,15 +122,16 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
 
     // alloc memory
     y->mutable_data<T>(ctx.GetPlace());
-    mean_out->mutable_data<ScalingParamType<T>>(ctx.GetPlace());
-    variance_out->mutable_data<ScalingParamType<T>>(ctx.GetPlace());
-    saved_mean->mutable_data<ScalingParamType<T>>(ctx.GetPlace());
-    saved_variance->mutable_data<ScalingParamType<T>>(ctx.GetPlace());
+    mean_out->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
+    variance_out->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
+    saved_mean->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
+    saved_variance->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
 
     auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
-    math::SetConstant<platform::CUDADeviceContext, ScalingParamType<T>> functor;
-    functor(dev_ctx, saved_mean, static_cast<ScalingParamType<T>>(0));
-    functor(dev_ctx, saved_variance, static_cast<ScalingParamType<T>>(0));
+    math::SetConstant<platform::CUDADeviceContext, BatchNormParamType<T>>
+        functor;
+    functor(dev_ctx, saved_mean, static_cast<BatchNormParamType<T>>(0));
+    functor(dev_ctx, saved_variance, static_cast<BatchNormParamType<T>>(0));
 
     auto handle = dev_ctx.cudnn_handle();
 
@@ -151,10 +152,10 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
           CUDNN_BATCHNORM_SPATIAL, CudnnDataType<T>::kOne(),
           CudnnDataType<T>::kZero(), data_desc_, x->template data<T>(),
           data_desc_, y->template mutable_data<T>(ctx.GetPlace()),
-          bn_param_desc_, scale->template data<ScalingParamType<T>>(),
-          bias->template data<ScalingParamType<T>>(),
-          est_mean->template data<ScalingParamType<T>>(),
-          est_var->template data<ScalingParamType<T>>(), epsilon));
+          bn_param_desc_, scale->template data<BatchNormParamType<T>>(),
+          bias->template data<BatchNormParamType<T>>(),
+          est_mean->template data<BatchNormParamType<T>>(),
+          est_var->template data<BatchNormParamType<T>>(), epsilon));
     } else {
       // Run training mode.
       // obtain running mean and running inv var, and see if we need to
@@ -165,14 +166,15 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
           handle, mode_, CudnnDataType<T>::kOne(), CudnnDataType<T>::kZero(),
           data_desc_, x->template data<T>(), data_desc_,
           y->template mutable_data<T>(ctx.GetPlace()), bn_param_desc_,
-          scale->template data<ScalingParamType<T>>(),
-          bias->template data<ScalingParamType<T>>(), this_factor,
-          mean_out->template mutable_data<ScalingParamType<T>>(ctx.GetPlace()),
-          variance_out->template mutable_data<ScalingParamType<T>>(
+          scale->template data<BatchNormParamType<T>>(),
+          bias->template data<BatchNormParamType<T>>(), this_factor,
+          mean_out->template mutable_data<BatchNormParamType<T>>(
               ctx.GetPlace()),
-          epsilon, saved_mean->template mutable_data<ScalingParamType<T>>(
+          variance_out->template mutable_data<BatchNormParamType<T>>(
               ctx.GetPlace()),
-          saved_variance->template mutable_data<ScalingParamType<T>>(
+          epsilon, saved_mean->template mutable_data<BatchNormParamType<T>>(
+                       ctx.GetPlace()),
+          saved_variance->template mutable_data<BatchNormParamType<T>>(
               ctx.GetPlace())));
     }
 

paddle/fluid/platform/cudnn_helper.h

@@ -86,7 +86,8 @@ class CudnnDataType<float16> {
  public:
   static const cudnnDataType_t type = CUDNN_DATA_HALF;
   // The scaling param type is float for HALF and FLOAT tensors
-  typedef const float ScalingParamType;
+  using ScalingParamType = const float;
+  using BatchNormParamType = float;
   static ScalingParamType* kOne() {
     static ScalingParamType v = 1.0;
     return &v;
@@ -101,7 +102,8 @@ template <>
 class CudnnDataType<float> {
  public:
   static const cudnnDataType_t type = CUDNN_DATA_FLOAT;
-  typedef const float ScalingParamType;
+  using ScalingParamType = const float;
+  using BatchNormParamType = float;
   static ScalingParamType* kOne() {
     static ScalingParamType v = 1.0;
     return &v;
@@ -116,7 +118,8 @@ template <>
 class CudnnDataType<double> {
  public:
   static const cudnnDataType_t type = CUDNN_DATA_DOUBLE;
-  typedef const double ScalingParamType;
+  using ScalingParamType = const double;
+  using BatchNormParamType = double;
   static ScalingParamType* kOne() {
     static ScalingParamType v = 1.0;
     return &v;