7 #include "complexmat.cuh"
9 #include "complexmat.hpp"
11 // For compatibility reasons between CuFFT and FFTW, OpenCVfft versions.
12 typedef int *cudaStream_t;
18 ThreadCtx(cv::Size windows_size, uint cell_size, uint num_of_feats, uint num_of_scales = 1)
20 this->xf_sqr_norm = DynMem(num_of_scales * sizeof(float));
21 this->yf_sqr_norm = DynMem(sizeof(float));
22 this->patch_feats.reserve(uint(num_of_feats));
25 ((uint(windows_size.width) / cell_size) * (uint(windows_size.height) / cell_size)) * sizeof(float);
27 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
28 CudaSafeCall(cudaStreamCreate(&this->stream));
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);
35 uint width_freq = (uint(windows_size.width) / cell_size) / 2 + 1;
37 this->in_all = cv::Mat(windows_size.height / int(cell_size) * int(num_of_scales),
38 windows_size.width / int(cell_size), CV_32F, this->gauss_corr_res.hostMem());
40 this->fw_all = cv::Mat((windows_size.height / int(cell_size)) * int(num_of_feats),
41 windows_size.width / int(cell_size), CV_32F, this->data_features.hostMem());
43 uint width_freq = uint(windows_size.width) / cell_size;
45 this->in_all = cv::Mat((windows_size.height / int(cell_size)), windows_size.width / int(cell_size), CV_32F);
48 this->data_i_features = DynMem(cells_size * num_of_feats);
49 this->data_i_1ch = DynMem(cells_size * num_of_scales);
51 this->ifft2_res = cv::Mat(windows_size.height / int(cell_size), windows_size.width / int(cell_size),
52 CV_32FC(int(num_of_feats)), this->data_i_features.hostMem());
54 this->response = cv::Mat(windows_size.height / int(cell_size), windows_size.width / int(cell_size),
55 CV_32FC(int(num_of_scales)), this->data_i_1ch.hostMem());
57 this->patch_feats.reserve(num_of_feats);
60 this->zf.create(uint(windows_size.height) / cell_size, width_freq, num_of_feats, num_of_scales, this->stream);
61 this->kzf.create(uint(windows_size.height) / cell_size, width_freq, num_of_scales, this->stream);
62 this->kf.create(uint(windows_size.height) / cell_size, width_freq, num_of_scales, this->stream);
64 this->zf.create(uint(windows_size.height) / cell_size, width_freq, num_of_feats, num_of_scales);
65 this->kzf.create(uint(windows_size.height) / cell_size, width_freq, num_of_scales);
66 this->kf.create(uint(windows_size.height) / cell_size, width_freq, num_of_scales);
69 if (num_of_scales > 1) {
70 this->max_responses.reserve(uint(num_of_scales));
71 this->max_locs.reserve(uint(num_of_scales));
72 this->response_maps.reserve(uint(num_of_scales));
78 #if !defined(BIG_BATCH) && defined(CUFFT) && (defined(ASYNC) || defined(OPENMP))
79 CudaSafeCall(cudaStreamDestroy(this->stream));
83 DynMem xf_sqr_norm, yf_sqr_norm;
84 std::vector<cv::Mat> patch_feats;
86 cv::Mat in_all, fw_all, ifft2_res, response;
87 ComplexMat zf, kzf, kf, xyf;
89 DynMem data_i_features, data_i_1ch;
90 // CuFFT and FFTW variables
91 DynMem gauss_corr_res, data_features;
94 cudaStream_t stream = nullptr;
95 ComplexMat model_alphaf, model_xf;
97 // Variables used during non big batch mode and in big batch mode with ThreadCtx in p_threadctxs in kcf on zero index.
99 double max_val, max_response;
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;
108 #endif // SCALE_VARS_HPP