modules/dnn/src/cuda4dnn/primitives/reorg.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_REORG_HPP
   6 #define OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_REORG_HPP
   7
   8 #include "../../op_cuda.hpp"
   9
  10 #include "../csl/stream.hpp"
  11 #include "../kernels/permute.hpp"
  12
  13 #include <opencv2/core.hpp>
  14
  15 #include <vector>
  16 #include <utility>
  17
  18 namespace cv { namespace dnn { namespace cuda4dnn {
  19
  20     template <class T>
  21     class ReorgOp final : public CUDABackendNode {
  22     public:
  23         using wrapper_type = GetCUDABackendWrapperType<T>;
  24
  25         ReorgOp(csl::Stream stream_, std::size_t stride_)
  26             : stream(std::move(stream_)), stride{ stride_ } { }
  27
  28         void forward(
  29             const std::vector<cv::Ptr<BackendWrapper>>& inputs,
  30             const std::vector<cv::Ptr<BackendWrapper>>& outputs,
  31             csl::Workspace& workspace) override
  32         {
  33             CV_Assert(inputs.size() == 1 && outputs.size() == 1);
  34
  35             auto input_wrapper = inputs[0].dynamicCast<wrapper_type>();
  36             auto input = input_wrapper->getView();
  37
  38             auto output_wrapper = outputs[0].dynamicCast<wrapper_type>();
  39             auto output = output_wrapper->getSpan();
  40
  41             const std::size_t permute_input_shape[] = {
  42                input.get_axis_size(0),
  43                input.get_axis_size(1) * input.get_axis_size(2) / (stride * stride),
  44                stride,
  45                input.get_axis_size(3),
  46                stride
  47             };
  48
  49             constexpr std::size_t order[] = { 0, 2, 4, 1, 3 };
  50
  51             const std::size_t permute_output_shape[] = {
  52                 permute_input_shape[order[0]],
  53                 permute_input_shape[order[1]],
  54                 permute_input_shape[order[2]],
  55                 permute_input_shape[order[3]],
  56                 permute_input_shape[order[4]]
  57             };
  58
  59             input.unsqueeze();
  60             input.reshape(std::begin(permute_input_shape), std::end(permute_input_shape));
  61
  62             output.unsqueeze();
  63             output.reshape(std::begin(permute_output_shape), std::end(permute_output_shape));
  64
  65             kernels::permute(stream, output, input, { std::begin(order), std::end(order) });
  66         }
  67
  68     private:
  69         csl::Stream stream;
  70         std::size_t stride;
  71     };
  72
  73 }}} /* namespace cv::dnn::cuda4dnn */
  74
  75 #endif /* OPENCV_DNN_SRC_CUDA4DNN_PRIMITIVES_REORG_HPP */