refine conv cudnn enforce (#12353)

* refine conv cudnn enforce

* update

* update all cudnn ops

* fix

paddle/fluid/operators/conv_cudnn_op.cu.cc

@@ -77,7 +77,7 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
     // cudnn 7 can support groups, no need to do it mannually
     // FIXME(typhoonzero): find a better way to disable groups
     // rather than setting it to 1.
-    PADDLE_ENFORCE(platform::dynload::cudnnSetConvolutionGroupCount(
+    CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionGroupCount(
         cudnn_conv_desc, groups));
     groups = 1;
 #endif
@@ -129,7 +129,7 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
     auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
     auto handle = dev_ctx.cudnn_handle();
 
-    PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm(
+    CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm(
         handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc,
         cudnn_output_desc, CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
         workspace_size_limit, &algo));
@@ -140,18 +140,18 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
     if (dev_ctx.GetComputeCapability() >= 70 &&
         std::type_index(typeid(T)) ==
             std::type_index(typeid(platform::float16))) {
-      PADDLE_ENFORCE(platform::dynload::cudnnSetConvolutionMathType(
+      CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionMathType(
           cudnn_conv_desc, CUDNN_TENSOR_OP_MATH));
       // Currently tensor core is only enabled using this algo
       algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM;
     } else {
-      PADDLE_ENFORCE(platform::dynload::cudnnSetConvolutionMathType(
+      CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionMathType(
           cudnn_conv_desc, CUDNN_DEFAULT_MATH));
     }
 #endif
 
     // get workspace size able to allocate
-    PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardWorkspaceSize(
+    CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardWorkspaceSize(
         handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc,
         cudnn_output_desc, algo, &workspace_size_in_bytes));
     // It is possible for float16 on Volta GPU to allocate more memory than
@@ -165,7 +165,7 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
     // ------------------- cudnn conv forward ---------------------
     ScalingParamType<T> alpha = 1.0f, beta = 0.0f;
     for (int i = 0; i < groups; i++) {
-      PADDLE_ENFORCE(platform::dynload::cudnnConvolutionForward(
+      CUDNN_ENFORCE(platform::dynload::cudnnConvolutionForward(
           handle, &alpha, cudnn_input_desc, input_data + i * group_offset_in,
           cudnn_filter_desc, filter_data + i * group_offset_filter,
           cudnn_conv_desc, algo, cudnn_workspace, workspace_size_in_bytes,
@@ -218,7 +218,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
     // cudnn 7 can support groups, no need to do it mannually
     // FIXME(typhoonzero): find a better way to disable groups
     // rather than setting it to 1.
-    PADDLE_ENFORCE(platform::dynload::cudnnSetConvolutionGroupCount(
+    CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionGroupCount(
         cudnn_conv_desc, groups));
     groups = 1;
 #endif
@@ -273,7 +273,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
     auto handle = dev_ctx.cudnn_handle();
     if (input_grad) {
       if (FLAGS_cudnn_deterministic) {
-        PADDLE_ENFORCE(
+        CUDNN_ENFORCE(
             platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
                 handle, cudnn_filter_desc,
                 // dyDesc: Handle to the previously initialized input
@@ -289,7 +289,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
         data_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1;
       }
 
-      PADDLE_ENFORCE(
+      CUDNN_ENFORCE(
           platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
               handle, cudnn_filter_desc, cudnn_output_grad_desc,
               cudnn_conv_desc, cudnn_input_desc, data_algo, &tmp_size));
@@ -298,7 +298,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
 
     if (filter_grad) {
       if (FLAGS_cudnn_deterministic) {
-        PADDLE_ENFORCE(
+        CUDNN_ENFORCE(
             platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
                 handle, cudnn_input_desc, cudnn_output_grad_desc,
                 cudnn_conv_desc, cudnn_filter_desc,
@@ -308,7 +308,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
         filter_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1;
       }
 
-      PADDLE_ENFORCE(
+      CUDNN_ENFORCE(
           platform::dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize(
               handle, cudnn_input_desc, cudnn_output_grad_desc, cudnn_conv_desc,
               cudnn_filter_desc, filter_algo, &tmp_size));
@@ -326,7 +326,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
       // Because beta is zero, it is unnecessary to reset input_grad.
 
       for (int i = 0; i < groups; i++) {
-        PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardData(
+        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardData(
             handle, &alpha, cudnn_filter_desc,
             filter_data + i * group_offset_filter, cudnn_output_grad_desc,
             output_grad_data + i * group_offset_out, cudnn_conv_desc, data_algo,
@@ -339,7 +339,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
       T* filter_grad_data = filter_grad->mutable_data<T>(ctx.GetPlace());
       // Because beta is zero, it is unnecessary to reset filter_grad.
       for (int i = 0; i < groups; i++) {
-        PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter(
+        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter(
             handle, &alpha, cudnn_input_desc, input_data + i * group_offset_in,
             cudnn_output_grad_desc, output_grad_data + i * group_offset_out,
             cudnn_conv_desc, filter_algo, cudnn_workspace,

paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc

@@ -87,7 +87,7 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel<T> {
     auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
     auto handle = dev_ctx.cudnn_handle();
     // Get the algorithm
-    PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
+    CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
         handle, cudnn_filter_desc, cudnn_input_desc, cudnn_conv_desc,
         // dxDesc: Handle to the previously initialized output tensor
         // descriptor.
@@ -95,7 +95,7 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel<T> {
         workspace_size_limit, &algo));
 
     // get workspace size able to allocate
-    PADDLE_ENFORCE(
+    CUDNN_ENFORCE(
         platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
             handle, cudnn_filter_desc, cudnn_input_desc, cudnn_conv_desc,
             cudnn_output_desc, algo, &workspace_size_in_bytes));
@@ -110,7 +110,7 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel<T> {
     int filter_offset = filter->numel() / groups;
     T alpha = 1.0f, beta = 0.0f;
     for (int g = 0; g < groups; g++) {
-      PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardData(
+      CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardData(
           handle, &alpha, cudnn_filter_desc, filter_data + filter_offset * g,
           cudnn_input_desc, input_data + input_offset * g, cudnn_conv_desc,
           algo, cudnn_workspace, workspace_size_in_bytes, &beta,
@@ -178,11 +178,11 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
     auto handle = dev_ctx.cudnn_handle();
     if (input_grad) {
       // choose backward algorithm for data
-      PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm(
+      CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm(
           handle, cudnn_output_desc, cudnn_filter_desc, cudnn_conv_desc,
           cudnn_input_desc, CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
           workspace_size_limit, &data_algo));
-      PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardWorkspaceSize(
+      CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardWorkspaceSize(
           handle, cudnn_output_desc, cudnn_filter_desc, cudnn_conv_desc,
           cudnn_input_desc, data_algo, &fwd_ws_size));
       workspace_size_in_bytes = std::max(workspace_size_in_bytes, fwd_ws_size);
@@ -190,7 +190,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
 
     if (filter_grad) {
       // choose backward algorithm for filter
-      PADDLE_ENFORCE(
+      CUDNN_ENFORCE(
           platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
               handle, cudnn_output_desc, cudnn_input_desc, cudnn_conv_desc,
               cudnn_filter_desc,
@@ -198,7 +198,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
               workspace_size_limit, &filter_algo));
 
       // get workspace for backwards filter algorithm
-      PADDLE_ENFORCE(
+      CUDNN_ENFORCE(
           platform::dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize(
               handle, cudnn_output_desc, cudnn_input_desc, cudnn_conv_desc,
               cudnn_filter_desc, filter_algo, &bwd_filter_ws_size));
@@ -222,7 +222,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
       T* input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
       // Because beta is zero, it is unnecessary to reset input_grad.
       for (int g = 0; g < groups; g++) {
-        PADDLE_ENFORCE(platform::dynload::cudnnConvolutionForward(
+        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionForward(
             handle, &alpha, cudnn_output_desc,
             output_grad_data + output_grad_offset * g, cudnn_filter_desc,
             filter_data + filter_offset * g, cudnn_conv_desc, data_algo,
@@ -237,7 +237,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
       // Because beta is zero, it is unnecessary to reset filter_grad.
       // Gradient with respect to the filter
       for (int g = 0; g < groups; g++) {
-        PADDLE_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter(
+        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter(
             handle, &alpha, cudnn_output_desc,
             output_grad_data + output_grad_offset * g, cudnn_input_desc,
             input_data + input_offset * g, cudnn_conv_desc, filter_algo,

paddle/fluid/operators/math/softmax.cu

@@ -52,7 +52,7 @@ void SoftmaxCUDNNFunctor<T>::operator()(
       xDesc.descriptor<T>(layout, cudnn_tensor_dims);
   cudnnTensorDescriptor_t cudnn_y_desc =
       xDesc.descriptor<T>(layout, cudnn_tensor_dims);
-  PADDLE_ENFORCE(platform::dynload::cudnnSoftmaxForward(
+  CUDNN_ENFORCE(platform::dynload::cudnnSoftmaxForward(
       context.cudnn_handle(), CUDNN_SOFTMAX_ACCURATE,
       CUDNN_SOFTMAX_MODE_INSTANCE, CudnnDataType<T>::kOne(), cudnn_x_desc,
       X->data<T>(), CudnnDataType<T>::kZero(), cudnn_y_desc,
@@ -83,7 +83,7 @@ void SoftmaxGradCUDNNFunctor<T>::operator()(
       dxDesc.descriptor<T>(layout, cudnn_tensor_dims);
   cudnnTensorDescriptor_t cudnn_ygrad_desc =
       dyDesc.descriptor<T>(layout, cudnn_tensor_dims);
-  PADDLE_ENFORCE(platform::dynload::cudnnSoftmaxBackward(
+  CUDNN_ENFORCE(platform::dynload::cudnnSoftmaxBackward(
       context.cudnn_handle(), CUDNN_SOFTMAX_ACCURATE,
       CUDNN_SOFTMAX_MODE_INSTANCE, CudnnDataType<T>::kOne(), cudnn_y_desc,
       Y->data<T>(), cudnn_ygrad_desc, YGrad->data<T>(),

paddle/fluid/operators/pool_cudnn_op.cu.cc

@@ -81,7 +81,7 @@ class PoolCUDNNOpKernel : public framework::OpKernel<T> {
     // ------------------- cudnn pool algorithm ---------------------
     auto handle = ctx.cuda_device_context().cudnn_handle();
     ScalingParamType<T> alpha = 1.0f, beta = 0.0f;
-    PADDLE_ENFORCE(platform::dynload::cudnnPoolingForward(
+    CUDNN_ENFORCE(platform::dynload::cudnnPoolingForward(
         handle, cudnn_pool_desc, &alpha, cudnn_input_desc, input_data, &beta,
         cudnn_output_desc, output_data));
   }
@@ -154,7 +154,7 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel<T> {
       T *input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
       // Because beta is zero, it is unnecessary to reset input_grad.
 
-      PADDLE_ENFORCE(platform::dynload::cudnnPoolingBackward(
+      CUDNN_ENFORCE(platform::dynload::cudnnPoolingBackward(
           handle, cudnn_pool_desc, &alpha, cudnn_output_desc, output_data,
           cudnn_output_desc, output_grad_data, cudnn_input_desc, input_data,
           &beta, cudnn_input_desc, input_grad_data));