src/threadctx.hpp

   1 #ifndef SCALE_VARS_HPP
   2 #define SCALE_VARS_HPP
   3
   4 #include <future>
   5 #include "dynmem.hpp"
   6
   7 #ifdef CUFFT
   8 #include "complexmat.cuh"
   9 #else
  10 #include "complexmat.hpp"
  11 #ifndef CUFFTW
  12 // For compatibility reasons between CuFFT and FFTW, OpenCVfft versions.
  13 typedef int *cudaStream_t;
  14 #endif
  15 #endif
  16
  17 struct ThreadCtx {
  18   public:
  19     ThreadCtx(cv::Size windows_size, uint cell_size, uint num_of_feats, double scale, uint num_of_scales)
  20         : scale(scale)
  21     {
  22         this->xf_sqr_norm = DynMem(num_of_scales * sizeof(float));
  23         this->yf_sqr_norm = DynMem(sizeof(float));
  24         this->patch_feats.reserve(uint(num_of_feats));
  25
  26         uint cells_size =
  27             ((uint(windows_size.width) / cell_size) * (uint(windows_size.height) / cell_size)) * sizeof(float);
  28
  29 #if  !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
  30         CudaSafeCall(cudaStreamCreate(&this->stream));
  31 #endif
  32
  33 #if defined(CUFFT) || defined(FFTW)
  34         this->gauss_corr_res = DynMem(cells_size * num_of_scales);
  35         this->data_features = DynMem(cells_size * num_of_feats);
  36
  37         uint width_freq = (uint(windows_size.width) / cell_size) / 2 + 1;
  38
  39         this->in_all = cv::Mat(windows_size.height / int(cell_size) * int(num_of_scales),
  40                                windows_size.width / int(cell_size), CV_32F, this->gauss_corr_res.hostMem());
  41
  42         this->fw_all = cv::Mat((windows_size.height / int(cell_size)) * int(num_of_feats),
  43                                windows_size.width / int(cell_size), CV_32F, this->data_features.hostMem());
  44 #else
  45         uint width_freq = uint(windows_size.width) / cell_size;
  46
  47         this->in_all = cv::Mat((windows_size.height / int(cell_size)), windows_size.width / int(cell_size), CV_32F);
  48 #endif
  49
  50         this->data_i_features = DynMem(cells_size * num_of_feats);
  51         this->data_i_1ch = DynMem(cells_size * num_of_scales);
  52
  53         this->ifft2_res = cv::Mat(windows_size.height / int(cell_size), windows_size.width / int(cell_size),
  54                                   CV_32FC(int(num_of_feats)), this->data_i_features.hostMem());
  55
  56         this->response = cv::Mat(windows_size.height / int(cell_size), windows_size.width / int(cell_size),
  57                                  CV_32FC(int(num_of_scales)), this->data_i_1ch.hostMem());
  58
  59         this->patch_feats.reserve(num_of_feats);
  60
  61 #ifdef CUFFT
  62         this->zf.create(uint(windows_size.height) / cell_size, width_freq, num_of_feats, num_of_scales, this->stream);
  63         this->kzf.create(uint(windows_size.height) / cell_size, width_freq, num_of_scales, this->stream);
  64         this->kf.create(uint(windows_size.height) / cell_size, width_freq, num_of_scales, this->stream);
  65 #else
  66         this->zf.create(uint(windows_size.height) / cell_size, width_freq, num_of_feats, num_of_scales);
  67         this->kzf.create(uint(windows_size.height) / cell_size, width_freq, num_of_scales);
  68         this->kf.create(uint(windows_size.height) / cell_size, width_freq, num_of_scales);
  69 #endif
  70
  71         if (num_of_scales > 1) {
  72             this->max_responses.reserve(uint(num_of_scales));
  73             this->max_locs.reserve(uint(num_of_scales));
  74             this->response_maps.reserve(uint(num_of_scales));
  75         }
  76     }
  77     ThreadCtx(ThreadCtx &&) = default;
  78     ~ThreadCtx()
  79     {
  80 #if  !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
  81         CudaSafeCall(cudaStreamDestroy(this->stream));
  82 #endif
  83     }
  84
  85     const double scale;
  86 #ifdef ASYNC
  87     std::future<void> async_res;
  88 #endif
  89
  90     DynMem xf_sqr_norm, yf_sqr_norm;
  91     std::vector<cv::Mat> patch_feats;
  92
  93     cv::Mat in_all, fw_all, ifft2_res, response;
  94     ComplexMat zf, kzf, kf, xyf;
  95
  96     DynMem data_i_features, data_i_1ch;
  97     // CuFFT and FFTW variables
  98     DynMem gauss_corr_res, data_features;
  99
 100     // CuFFT variables
 101     cudaStream_t stream = nullptr;
 102     ComplexMat model_alphaf, model_xf;
 103
 104     // Variables used during non big batch mode and in big batch mode with ThreadCtx in p_threadctxs in kcf  on zero index.
 105     cv::Point2i max_loc;
 106     double max_val, max_response;
 107
 108     // Big batch variables
 109     // Stores value of responses, location of maximal response and response maps for each scale
 110     std::vector<double> max_responses;
 111     std::vector<cv::Point2i> max_locs;
 112     std::vector<cv::Mat> response_maps;
 113 };
 114
 115 #endif // SCALE_VARS_HPP