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
  25         uint cells_size = roi.width * roi.height * sizeof(float);
  26
  27 #if  !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
  28         CudaSafeCall(cudaStreamCreate(&this->stream));
  29 #endif
  30
  31 #if defined(CUFFT) || defined(FFTW)
  32         this->gauss_corr_res = DynMem(cells_size * num_of_scales);
  33         this->data_features = DynMem(cells_size * num_of_feats);
  34
  35         uint width_freq = roi.width / 2 + 1;
  36
  37         this->in_all = cv::Mat(roi.height * num_of_scales, roi.width, CV_32F, this->gauss_corr_res.hostMem());
  38         this->fw_all = cv::Mat(roi.height * num_of_feats, roi.width, CV_32F, this->data_features.hostMem());
  39 #else
  40         uint width_freq = roi.width;
  41
  42         this->in_all = cv::Mat(roi, CV_32F);
  43 #endif
  44
  45         this->data_i_features = DynMem(cells_size * num_of_feats);
  46         this->data_i_1ch = DynMem(cells_size * num_of_scales);
  47
  48         this->ifft2_res = cv::Mat(roi, CV_32FC(num_of_feats), this->data_i_features.hostMem());
  49         this->response = cv::Mat(roi, CV_32FC(num_of_scales), this->data_i_1ch.hostMem());
  50
  51 #ifdef CUFFT
  52         this->zf.create(roi.height, width_freq, num_of_feats, num_of_scales, this->stream);
  53         this->kzf.create(roi.height, width_freq, num_of_scales, this->stream);
  54         this->kf.create(roi.height, width_freq, num_of_scales, this->stream);
  55 #else
  56         this->zf.create(roi.height, width_freq, num_of_feats, num_of_scales);
  57         this->kzf.create(roi.height, width_freq, num_of_scales);
  58         this->kf.create(roi.height, width_freq, num_of_scales);
  59 #endif
  60
  61 #ifdef BIG_BATCH
  62         if (num_of_scales > 1) {
  63             this->max_responses.reserve(num_of_scales);
  64             this->max_locs.reserve(num_of_scales);
  65             this->response_maps.reserve(num_of_scales);
  66         }
  67 #endif
  68     }
  69     ThreadCtx(ThreadCtx &&) = default;
  70     ~ThreadCtx()
  71     {
  72 #if  !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
  73         CudaSafeCall(cudaStreamDestroy(this->stream));
  74 #endif
  75     }
  76
  77     const double scale;
  78 #ifdef ASYNC
  79     std::future<void> async_res;
  80 #endif
  81
  82     DynMem xf_sqr_norm, yf_sqr_norm;
  83
  84     cv::Mat in_all, fw_all, ifft2_res, response;
  85     ComplexMat zf, kzf, kf, xyf;
  86
  87     DynMem data_i_features, data_i_1ch;
  88     // CuFFT and FFTW variables
  89     DynMem gauss_corr_res, data_features;
  90
  91     // CuFFT variables
  92     cudaStream_t stream = nullptr;
  93     ComplexMat model_alphaf, model_xf;
  94
  95     // Variables used during non big batch mode and in big batch mode with ThreadCtx in p_threadctxs in kcf  on zero index.
  96     cv::Point2i max_loc;
  97     double max_val, max_response;
  98
  99 #ifdef BIG_BATCH
 100     // Stores value of responses, location of maximal response and response maps for each scale
 101     std::vector<double> max_responses;
 102     std::vector<cv::Point2i> max_locs;
 103     std::vector<cv::Mat> response_maps;
 104 #endif
 105 };
 106
 107 #endif // SCALE_VARS_HPP