modules/dnn/src/cuda4dnn/primitives/shuffle_channel.hpp

   1 // This file is part of OpenCV project.
   2 // It is subject to the license terms in the LICENSE file found in the top-level directory
   3 // of this distribution and at http://opencv.org/license.html.
   4
   5 #ifndef OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_SHUFFLE_CHANNEL_HPP
   6 #define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_SHUFFLE_CHANNEL_HPP
   7
   8 #include "../../op_cuda.hpp"
   9
  10 #include "../csl/stream.hpp"
  11 #include "../csl/tensor_ops.hpp"
  12
  13 #include "../kernels/permute.hpp"
  14
  15 #include <opencv2/core.hpp>
  16
  17 #include <vector>
  18 #include <utility>
  19
  20 namespace cv { namespace dnn { namespace cuda4dnn {
  21
  22     template <class T>
  23     class ShuffleChannelOp final : public CUDABackendNode {
  24     public:
  25         using wrapper_type = GetCUDABackendWrapperType<T>;
  26
  27         ShuffleChannelOp(csl::Stream stream_, std::size_t group_)
  28             : stream(std::move(stream_)), group{ group_ } { }
  29
  30         void forward(
  31             const std::vector<cv::Ptr<BackendWrapper>>& inputs,
  32             const std::vector<cv::Ptr<BackendWrapper>>& outputs,
  33             csl::Workspace& workspace) override
  34         {
  35             CV_Assert(inputs.size() == 1 && outputs.size() == 1);
  36
  37             auto input_wrapper = inputs[0].dynamicCast<wrapper_type>();
  38             auto input = input_wrapper->getView();
  39
  40             auto output_wrapper = outputs[0].dynamicCast<wrapper_type>();
  41             auto output = output_wrapper->getSpan();
  42
  43             if (group == 1) {
  44                 /* permute is redundant; check else branch to know why */
  45                 if (input.get() != output.get()) {
  46                     input.reshape_as(output);
  47                     csl::tensor_ops::copy(stream, output, input);
  48                 }
  49             } else {
  50                 const std::size_t permute_input_shape[] = {
  51                    input.get_axis_size(0),
  52                    group,
  53                    input.get_axis_size(1) / group,
  54                    input.get_axis_size(2) * input.get_axis_size(3)
  55                 };
  56
  57                 constexpr std::size_t order[] = { 0, 2, 1, 3 };
  58
  59                 const std::size_t permute_output_shape[] = {
  60                     permute_input_shape[order[0]],
  61                     permute_input_shape[order[1]],
  62                     permute_input_shape[order[2]],
  63                     permute_input_shape[order[3]],
  64                 };
  65
  66                 input.reshape(std::begin(permute_input_shape), std::end(permute_input_shape));
  67                 output.reshape(std::begin(permute_output_shape), std::end(permute_output_shape));
  68                 kernels::permute(stream, output, input, { std::begin(order), std::end(order) });
  69             }
  70         }
  71
  72     private:
  73         csl::Stream stream;
  74         std::size_t group;
  75     };
  76
  77 }}} /* namespace cv::dnn::cuda4dnn */
  78
  79 #endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_SHUFFLE_CHANNEL_HPP */