X-Git-Url: http://rtime.felk.cvut.cz/gitweb/hercules2020/kcf.git/blobdiff_plain/99a5220a74c18aa8f13938f09cd9b6bafd3878e1..87b7b7429bc2d4e7bf7aa278b2a6e5c6afc08d45:/src/fft_cufft.h

diff --git a/src/fft_cufft.h b/src/fft_cufft.h
index 44c3556..4241c06 100644
--- a/src/fft_cufft.h
+++ b/src/fft_cufft.h
@@ -1,44 +1,38 @@
-
 #ifndef FFT_CUDA_H
 #define FFT_CUDA_H
 
-#include "fft.h"
-#include "cuda/cuda_error_check.cuh"
-
-#if CV_MAJOR_VERSION == 2
-  #include <opencv2/gpu/gpu.hpp>
-  #define CUDA cv::gpu
-#else
-  #include "opencv2/opencv.hpp"
-  #define CUDA cv::cuda
-#endif
-
 #include <cufft.h>
 #include <cuda_runtime.h>
+#include <cublas_v2.h>
 
-struct Scale_vars;
+#include "fft.h"
+#include "cuda_error_check.hpp"
+#include "pragmas.h"
+
+struct ThreadCtx;
 
 class cuFFT : public Fft
 {
 public:
-    void init(unsigned width, unsigned height, unsigned num_of_feats, unsigned num_of_scales, bool big_batch_mode) override;
-    void set_window(const cv::Mat & window) override;
-    ComplexMat forward(const cv::Mat & input) override;
-    void forward(Scale_vars & vars) override;
-    void forward_raw(Scale_vars & vars, bool all_scales) override;
-    ComplexMat forward_window(const std::vector<cv::Mat> & input) override;
-    void forward_window(Scale_vars & vars) override;
-    cv::Mat inverse(const ComplexMat & input) override;
-    void inverse(Scale_vars & vars) override;
-    float* inverse_raw(const ComplexMat & input) override;
-    ~cuFFT() override;
+    cuFFT();
+    void init(unsigned width, unsigned height, unsigned num_of_feats, unsigned num_of_scales);
+    void set_window(const MatDynMem &window);
+    void forward(const MatScales &real_input, ComplexMat &complex_result);
+    void forward_window(MatScaleFeats &patch_feats_in, ComplexMat &complex_result, MatScaleFeats &tmp);
+    void inverse(ComplexMat &complex_input, MatScales &real_result);
+    ~cuFFT();
+
+protected:
+    cufftHandle create_plan_fwd(uint howmany) const;
+    cufftHandle create_plan_inv(uint howmany) const;
+
 private:
     cv::Mat m_window;
-    unsigned m_width, m_height, m_num_of_feats, m_num_of_scales;
-    bool m_big_batch_mode;
-    cufftHandle plan_f, plan_f_all_scales, plan_fw, plan_fw_all_scales, plan_i_features,
-     plan_i_features_all_scales, plan_i_1ch, plan_i_1ch_all_scales;
-    float *data_f = nullptr, *data_f_all_scales = nullptr, *data_fw = nullptr, *data_fw_d = nullptr, *data_fw_all_scales = nullptr, *data_fw_all_scales_d = nullptr, *data_i_features = nullptr, *data_i_features_d = nullptr, *data_i_features_all_scales = nullptr, *data_i_features_all_scales_d = nullptr, *data_i_1ch = nullptr, *data_i_1ch_d = nullptr, *data_i_1ch_all_scales = nullptr, *data_i_1ch_all_scales_d = nullptr;
+    cufftHandle plan_f, plan_fw, plan_i_1ch;
+#ifdef BIG_BATCH
+    cufftHandle plan_f_all_scales, plan_fw_all_scales, plan_i_all_scales;
+#endif
+    cublasHandle_t cublas;
 };
 
 #endif // FFT_CUDA_H