add depthwise conv hip support (PaddlePaddle#41537)

Author: phlrain

paddle/phi/kernels/gpudnn/conv_grad_kernel.cu

@@ -627,6 +627,39 @@ void Conv3DCudnnGradKernel(const Context& dev_ctx,
                          filter_grad);
 }
 
+template <typename T, typename Context>
+void DepthwiseConvCudnnGradKernel(const Context& dev_ctx,
+                                  const DenseTensor& input,
+                                  const DenseTensor& filter,
+                                  const DenseTensor& out_grad,
+                                  const std::vector<int>& strides,
+                                  const std::vector<int>& paddings,
+                                  const std::string& paddding_algorithm,
+                                  int groups,
+                                  const std::vector<int>& dilations,
+                                  const std::string& data_format,
+                                  bool use_addto,
+                                  int workspace_size_MB,
+                                  bool exhaustive_search,
+                                  DenseTensor* input_grad,
+                                  DenseTensor* filter_grad) {
+  ConvCudnnGradKernel<T>(dev_ctx,
+                         input,
+                         filter,
+                         out_grad,
+                         strides,
+                         paddings,
+                         paddding_algorithm,
+                         groups,
+                         dilations,
+                         data_format,
+                         use_addto,
+                         workspace_size_MB,
+                         exhaustive_search,
+                         input_grad,
+                         filter_grad);
+}
+
 }  // namespace phi
 
 #ifdef PADDLE_WITH_HIP
@@ -643,6 +676,13 @@ PD_REGISTER_KERNEL(conv3d_grad,
                    phi::Conv3DCudnnGradKernel,
                    float,
                    phi::dtype::float16) {}
+
+PD_REGISTER_KERNEL(depthwise_conv2d_grad,
+                   GPUDNN,
+                   ALL_LAYOUT,
+                   phi::DepthwiseConvCudnnGradKernel,
+                   float,
+                   phi::dtype::float16) {}
 #else
 #if CUDNN_VERSION_MIN(8, 1, 0)
 PD_REGISTER_KERNEL(conv2d_grad,

paddle/phi/kernels/gpudnn/conv_kernel.cu

@@ -416,6 +416,35 @@ void Conv3DCudnnKernel(const Context& dev_ctx,
                      out);
 }
 
+template <typename T, typename Context>
+void DepthwiseConvCudnnKernel(const Context& dev_ctx,
+                              const DenseTensor& input,
+                              const DenseTensor& filter,
+                              const std::vector<int>& strides,
+                              const std::vector<int>& paddings,
+                              const std::string& padding_algorithm,
+                              int groups,
+                              const std::vector<int>& dilations,
+                              const std::string& data_format,
+                              bool use_addto,
+                              int workspace_size_MB,
+                              bool exhaustive_search,
+                              DenseTensor* out) {
+  ConvCudnnKernel<T>(dev_ctx,
+                     input,
+                     filter,
+                     strides,
+                     paddings,
+                     padding_algorithm,
+                     groups,
+                     dilations,
+                     data_format,
+                     use_addto,
+                     workspace_size_MB,
+                     exhaustive_search,
+                     out);
+}
+
 }  // namespace phi
 
 #ifdef PADDLE_WITH_HIP
@@ -432,6 +461,14 @@ PD_REGISTER_KERNEL(conv3d,
                    phi::Conv3DCudnnKernel,
                    float,
                    phi::dtype::float16) {}
+
+PD_REGISTER_KERNEL(depthwise_conv2d,
+                   GPUDNN,
+                   ALL_LAYOUT,
+                   phi::DepthwiseConvCudnnKernel,
+                   float,
+                   phi::dtype::float16) {}
+
 #else
 #if CUDNN_VERSION_MIN(8, 1, 0)
 PD_REGISTER_KERNEL(conv2d,