5 CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
6 cudaErrorCheck(cublasCreate(&cublas));
9 cufftHandle cuFFT::create_plan_fwd(uint howmany) const
12 int n[] = {(int)m_height, (int)m_width};
13 int idist = m_height * m_width, odist = m_height * (m_width / 2 + 1);
14 int istride = 1, ostride = 1;
15 int *inembed = n, onembed[] = {(int)m_height, (int)m_width / 2 + 1};
18 cudaErrorCheck(cufftPlanMany(&plan, rank, n, inembed, istride, idist, onembed, ostride, odist, CUFFT_R2C, howmany));
19 cudaErrorCheck(cufftSetStream(plan, cudaStreamPerThread));
23 cufftHandle cuFFT::create_plan_inv(uint howmany) const
26 int n[] = {(int)m_height, (int)m_width};
27 int idist = m_height * (m_width / 2 + 1), odist = m_height * m_width;
28 int istride = 1, ostride = 1;
29 int inembed[] = {(int)m_height, (int)m_width / 2 + 1}, *onembed = n;
32 cudaErrorCheck(cufftPlanMany(&plan, rank, n, inembed, istride, idist, onembed, ostride, odist, CUFFT_C2R, howmany));
33 cudaErrorCheck(cufftSetStream(plan, cudaStreamPerThread));
38 void cuFFT::init(unsigned width, unsigned height, unsigned num_of_feats, unsigned num_of_scales)
40 Fft::init(width, height, num_of_feats, num_of_scales);
42 std::cout << "FFT: cuFFT" << std::endl;
44 plan_f = create_plan_fwd(1);
45 plan_fw = create_plan_fwd(m_num_of_feats);
46 plan_i_1ch = create_plan_inv(1);
49 plan_f_all_scales = create_plan_fwd(m_num_of_scales);
50 plan_fw_all_scales = create_plan_fwd(m_num_of_scales * m_num_of_feats);
51 plan_i_all_scales = create_plan_inv(m_num_of_scales);
55 void cuFFT::set_window(const MatDynMem &window)
57 Fft::set_window(window);
61 void cuFFT::forward(const MatScales &real_input, ComplexMat &complex_result)
63 Fft::forward(real_input, complex_result);
64 auto in = static_cast<cufftReal *>(const_cast<MatScales&>(real_input).deviceMem());
66 if (real_input.size[0] == 1)
67 cudaErrorCheck(cufftExecR2C(plan_f, in, complex_result.get_p_data()));
70 cudaErrorCheck(cufftExecR2C(plan_f_all_scales, in, complex_result.get_p_data()));
74 void cuFFT::forward_window(MatScaleFeats &feat, ComplexMat &complex_result, MatScaleFeats &temp)
76 Fft::forward_window(feat, complex_result, temp);
78 cufftReal *temp_data = temp.deviceMem();
79 uint n_scales = feat.size[0];
81 for (uint s = 0; s < n_scales; ++s) {
82 for (uint ch = 0; ch < uint(feat.size[1]); ++ch) {
83 cv::Mat feat_plane = feat.plane(s, ch);
84 cv::Mat temp_plane = temp.plane(s, ch);
85 temp_plane = feat_plane.mul(m_window);
90 cudaErrorCheck(cufftExecR2C(plan_fw, temp_data, complex_result.get_p_data()));
93 cudaErrorCheck(cufftExecR2C(plan_fw_all_scales, temp_data, complex_result.get_p_data()));
97 void cuFFT::inverse(ComplexMat &complex_input, MatScales &real_result)
99 Fft::inverse(complex_input, real_result);
101 uint n_channels = complex_input.n_channels;
102 cufftComplex *in = reinterpret_cast<cufftComplex *>(complex_input.get_p_data());
103 cufftReal *out = real_result.deviceMem();
104 float alpha = 1.0 / (m_width * m_height);
107 cudaErrorCheck(cufftExecC2R(plan_i_1ch, in, out));
109 cudaErrorCheck(cufftExecC2R(plan_i_all_scales, in, out));
111 // TODO: Investigate whether this scalling is needed or not
112 cudaErrorCheck(cublasSscal(cublas, real_result.total(), &alpha, out, 1));
117 cudaErrorCheck(cublasDestroy(cublas));
119 cudaErrorCheck(cufftDestroy(plan_f));
120 cudaErrorCheck(cufftDestroy(plan_fw));
121 cudaErrorCheck(cufftDestroy(plan_i_1ch));
124 cudaErrorCheck(cufftDestroy(plan_f_all_scales));
125 cudaErrorCheck(cufftDestroy(plan_fw_all_scales));
126 cudaErrorCheck(cufftDestroy(plan_i_all_scales));