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 roi, 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 = roi.width * roi.height * sizeof(float);
  27
  28 #if  !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
  29         CudaSafeCall(cudaStreamCreate(&this->stream));
  30 #endif
  31
  32 #if defined(CUFFT) || defined(FFTW)
  33         this->gauss_corr_res = DynMem(cells_size * num_of_scales);
  34         this->data_features = DynMem(cells_size * num_of_feats);
  35
  36         uint width_freq = roi.width / 2 + 1;
  37
  38         this->in_all = cv::Mat(roi.height * num_of_scales, roi.width, CV_32F, this->gauss_corr_res.hostMem());
  39         this->fw_all = cv::Mat(roi.height * num_of_feats, roi.width, CV_32F, this->data_features.hostMem());
  40 #else
  41         uint width_freq = roi.width;
  42
  43         this->in_all = cv::Mat(roi, CV_32F);
  44 #endif
  45
  46         this->data_i_features = DynMem(cells_size * num_of_feats);
  47         this->data_i_1ch = DynMem(cells_size * num_of_scales);
  48
  49         this->ifft2_res = cv::Mat(roi, CV_32FC(num_of_feats), this->data_i_features.hostMem());
  50         this->response = cv::Mat(roi, CV_32FC(num_of_scales), this->data_i_1ch.hostMem());
  51
  52         this->patch_feats.reserve(num_of_feats);
  53
  54 #ifdef CUFFT
  55         this->zf.create(roi.height, width_freq, num_of_feats, num_of_scales, this->stream);
  56         this->kzf.create(roi.height, width_freq, num_of_scales, this->stream);
  57         this->kf.create(roi.height, width_freq, num_of_scales, this->stream);
  58 #else
  59         this->zf.create(roi.height, width_freq, num_of_feats, num_of_scales);
  60         this->kzf.create(roi.height, width_freq, num_of_scales);
  61         this->kf.create(roi.height, width_freq, num_of_scales);
  62 #endif
  63
  64         if (num_of_scales > 1) {
  65             this->max_responses.reserve(num_of_scales);
  66             this->max_locs.reserve(num_of_scales);
  67             this->response_maps.reserve(num_of_scales);
  68         }
  69     }
  70     ThreadCtx(ThreadCtx &&) = default;
  71     ~ThreadCtx()
  72     {
  73 #if  !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
  74         CudaSafeCall(cudaStreamDestroy(this->stream));
  75 #endif
  76     }
  77
  78     const double scale;
  79 #ifdef ASYNC
  80     std::future<void> async_res;
  81 #endif
  82
  83     DynMem xf_sqr_norm, yf_sqr_norm;
  84     std::vector<cv::Mat> patch_feats;
  85
  86     cv::Mat in_all, fw_all, ifft2_res, response;
  87     ComplexMat zf, kzf, kf, xyf;
  88
  89     DynMem data_i_features, data_i_1ch;
  90     // CuFFT and FFTW variables
  91     DynMem gauss_corr_res, data_features;
  92
  93     // CuFFT variables
  94     cudaStream_t stream = nullptr;
  95     ComplexMat model_alphaf, model_xf;
  96
  97     // Variables used during non big batch mode and in big batch mode with ThreadCtx in p_threadctxs in kcf  on zero index.
  98     cv::Point2i max_loc;
  99     double max_val, max_response;
 100
 101     // Big batch variables
 102     // Stores value of responses, location of maximal response and response maps for each scale
 103     std::vector<double> max_responses;
 104     std::vector<cv::Point2i> max_locs;
 105     std::vector<cv::Mat> response_maps;
 106 };
 107
 108 #endif // SCALE_VARS_HPP