#include "fft_cufft.h"
-#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
-inline void gpuAssert(cudaError_t code, char *file, int line, bool abort=true)
+cuFFT::cuFFT()
{
- if (code != cudaSuccess)
- {
- fprintf(stderr,"GPUassert: %s %s %d\n", cudaGetErrorString(code), file, line);
- if (abort) exit(code);
- }
+ CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
+ cudaErrorCheck(cublasCreate(&cublas));
}
-cuFFT::cuFFT(): m_num_of_streams(4)
-{}
-void cuFFT::init(unsigned width, unsigned height, unsigned num_of_feats, unsigned num_of_scales)
+cufftHandle cuFFT::create_plan_fwd(uint howmany) const
{
- m_width = width;
- m_height = height;
- m_num_of_feats = num_of_feats;
- m_num_of_scales = num_of_scales;
-
- std::cout << "FFT: cuFFT" << std::endl;
+ int rank = 2;
+ int n[] = {(int)m_height, (int)m_width};
+ int idist = m_height * m_width, odist = m_height * (m_width / 2 + 1);
+ int istride = 1, ostride = 1;
+ int *inembed = n, onembed[] = {(int)m_height, (int)m_width / 2 + 1};
- for (unsigned i = 0; i < m_num_of_streams; i++) gpuErrchk(cudaStreamCreate(&streams[i]));
+ cufftHandle plan;
+ cudaErrorCheck(cufftPlanMany(&plan, rank, n, inembed, istride, idist, onembed, ostride, odist, CUFFT_R2C, howmany));
+ cudaErrorCheck(cufftSetStream(plan, cudaStreamPerThread));
+ return plan;
+}
- //FFT forward one scale
- {
- cufftPlan2d(&plan_f, m_height, m_width, CUFFT_C2R);
- }
- //FFT forward all scales
- if(m_num_of_scales > 1)
- {
+cufftHandle cuFFT::create_plan_inv(uint howmany) const
+{
int rank = 2;
int n[] = {(int)m_height, (int)m_width};
- int howmany = m_num_of_scales;
- int idist = m_height*m_width, odist = m_height*(m_width/2+1);
+ int idist = m_height * (m_width / 2 + 1), odist = m_height * m_width;
int istride = 1, ostride = 1;
- int *inembed = NULL, *onembed = NULL;
+ int inembed[] = {(int)m_height, (int)m_width / 2 + 1}, *onembed = n;
- cufftPlanMany(&plan_f_all_scales, rank, n,
- inembed, istride, idist,
- onembed, ostride, odist,
- CUFFT_R2C, howmany);
- }
- //FFT forward window one scale
- {
- int rank = 2;
- int n[] = {(int)m_height, (int)m_width};
- int howmany = m_num_of_feats;
- int idist = m_height*m_width, odist = m_height*(m_width/2+1);
- int istride = 1, ostride = 1;
- int *inembed = NULL, *onembed = NULL;
-
- cufftPlanMany(&plan_fw, rank, n,
- inembed, istride, idist,
- onembed, ostride, odist,
- CUFFT_R2C, howmany);
- }
- //FFT forward window all scales all feats
- if(m_num_of_scales > 1)
- {
- int rank = 2;
- int n[] = {(int)m_height, (int)m_width};
- int howmany = m_num_of_scales*m_num_of_feats;
- int idist = m_height*m_width, odist = m_height*(m_width/2+1);
- int istride = 1, ostride = 1;
- int *inembed = NULL, *onembed = NULL;
-
- cufftPlanMany(&plan_fw_all_scales, rank, n,
- inembed, istride, idist,
- onembed, ostride, odist,
- CUFFT_R2C, howmany);
- }
- //FFT inverse one scale
- {
- int rank = 2;
- int n[] = {(int)m_height, (int)m_width};
- int howmany = m_num_of_feats;
- int idist = m_height*(m_width/2+1), odist = 1;
- int istride = 1, ostride = m_num_of_feats;
- int inembed[] = {(int)m_height, (int)m_width/2+1}, *onembed = n;
-
- cufftPlanMany(&plan_i_features, rank, n,
- inembed, istride, idist,
- onembed, ostride, odist,
- CUFFT_C2R, howmany);
- }
- //FFT inverse all scales
- if(m_num_of_scales > 1)
- {
- int rank = 2;
- int n[] = {(int)m_height, (int)m_width};
- int howmany = m_num_of_feats*m_num_of_scales;
- int idist = m_height*(m_width/2+1), odist = 1;
- int istride = 1, ostride = m_num_of_feats*m_num_of_scales;
- int inembed[] = {(int)m_height, (int)m_width/2+1}, *onembed = n;
-
- cufftPlanMany(&plan_i_features_all_scales, rank, n,
- inembed, istride, idist,
- onembed, ostride, odist,
- CUFFT_C2R, howmany);
- }
- //FFT inver one channel one scale
- {
- int rank = 2;
- int n[] = {(int)m_height, (int)m_width};
- int howmany = 1;
- int idist = m_height*(m_width/2+1), odist = 1;
- int istride = 1, ostride = 1;
- int inembed[] = {(int)m_height, (int)m_width/2+1}, *onembed = n;
-
- cufftPlanMany(&plan_i_1ch, rank, n,
- inembed, istride, idist,
- onembed, ostride, odist,
- CUFFT_C2R, howmany);
- }
- //FFT inver one channel all scales
- if(m_num_of_scales > 1)
- {
- int rank = 2;
- int n[] = {(int)m_height, (int)m_width};
- int howmany = m_num_of_scales;
- int idist = m_height*(m_width/2+1), odist = 1;
- int istride = 1, ostride = m_num_of_scales;
- int inembed[] = {(int)m_height, (int)m_width/2+1}, *onembed = n;
-
- cufftPlanMany(&plan_i_1ch_all_scales, rank, n,
- inembed, istride, idist,
- onembed, ostride, odist,
- CUFFT_C2R, howmany);
- }
+ cufftHandle plan;
+ cudaErrorCheck(cufftPlanMany(&plan, rank, n, inembed, istride, idist, onembed, ostride, odist, CUFFT_C2R, howmany));
+ cudaErrorCheck(cufftSetStream(plan, cudaStreamPerThread));
+ return plan;
}
-void cuFFT::set_window(const cv::Mat &window)
+
+void cuFFT::init(unsigned width, unsigned height, unsigned num_of_feats, unsigned num_of_scales)
{
- m_window = window;
+ Fft::init(width, height, num_of_feats, num_of_scales);
+
+ std::cout << "FFT: cuFFT" << std::endl;
+
+ plan_f = create_plan_fwd(1);
+ plan_fw = create_plan_fwd(m_num_of_feats);
+ plan_i_1ch = create_plan_inv(1);
+
+#ifdef BIG_BATCH
+ plan_f_all_scales = create_plan_fwd(m_num_of_scales);
+ plan_fw_all_scales = create_plan_fwd(m_num_of_scales * m_num_of_feats);
+ plan_i_all_scales = create_plan_inv(m_num_of_scales);
+#endif
}
-ComplexMat cuFFT::forward(const cv::Mat &input)
+void cuFFT::set_window(const MatDynMem &window)
{
- ComplexMat complex_result;
- if(input.rows == (int)(m_height*m_num_of_scales)){
- complex_result.create(m_height, m_width / 2 + 1, m_num_of_scales);
- cufftExecR2C(plan_f_all_scales, reinterpret_cast<float*>(input.data),
- reinterpret_cast<cufftComplex*>(complex_result.get_p_data()));
- } else {
- complex_result.create(m_height, m_width / 2 + 1, 1);
- cufftExecR2C(plan_f, reinterpret_cast<float*>(input.data),
- reinterpret_cast<cufftComplex*>(complex_result.get_p_data()));
- }
- return complex_result;
+ Fft::set_window(window);
+ m_window = window;
}
-ComplexMat cuFFT::forward_window(const std::vector<cv::Mat> &input)
+void cuFFT::forward(const MatScales &real_input, ComplexMat &complex_result)
{
- int n_channels = input.size();
- cv::Mat in_all(m_height * n_channels, m_width, CV_32F);
- for (int i = 0; i < n_channels; ++i) {
- cv::Mat in_roi(in_all, cv::Rect(0, i*m_height, m_width, m_height));
- in_roi = input[i].mul(m_window);
- }
- ComplexMat result;
- if(n_channels > (int) m_num_of_feats)
- result.create(m_height, m_width/2 + 1, n_channels,m_num_of_scales);
+ Fft::forward(real_input, complex_result);
+ auto in = static_cast<cufftReal *>(const_cast<MatScales&>(real_input).deviceMem());
+
+ if (real_input.size[0] == 1)
+ cudaErrorCheck(cufftExecR2C(plan_f, in, complex_result.get_dev_data()));
+#ifdef BIG_BATCH
else
- result.create(m_height, m_width/2 + 1, n_channels);
+ cudaErrorCheck(cufftExecR2C(plan_f_all_scales, in, complex_result.get_dev_data()));
+#endif
+}
- float *in = reinterpret_cast<float*>(in_all.data);
- cufftComplex *out = reinterpret_cast<cufftComplex*>(result.get_p_data());
+void cuFFT::forward_window(MatScaleFeats &feat, ComplexMat &complex_result, MatScaleFeats &temp)
+{
+ Fft::forward_window(feat, complex_result, temp);
- if (n_channels <= (int) m_num_of_feats)
- cufftExecR2C(plan_fw, in, out);
- else
- cufftExecR2C(plan_fw_all_scales, in, out);
+ cufftReal *temp_data = temp.deviceMem();
+ uint n_scales = feat.size[0];
+
+ for (uint s = 0; s < n_scales; ++s) {
+ for (uint ch = 0; ch < uint(feat.size[1]); ++ch) {
+ cv::Mat feat_plane = feat.plane(s, ch);
+ cv::Mat temp_plane = temp.plane(s, ch);
+ temp_plane = feat_plane.mul(m_window);
+ }
+ }
- return result;
+ if (n_scales == 1)
+ cudaErrorCheck(cufftExecR2C(plan_fw, temp_data, complex_result.get_dev_data()));
+#ifdef BIG_BATCH
+ else
+ cudaErrorCheck(cufftExecR2C(plan_fw_all_scales, temp_data, complex_result.get_dev_data()));
+#endif
+ CudaSafeCall(cudaStreamSynchronize(cudaStreamPerThread));
}
-cv::Mat cuFFT::inverse(const ComplexMat &inputf)
+void cuFFT::inverse(ComplexMat &complex_input, MatScales &real_result)
{
- int n_channels = inputf.n_channels;
- cv::Mat real_result(m_height, m_width, CV_32FC(n_channels));
- cufftComplex *in = reinterpret_cast<cufftComplex*>(inputf.get_p_data());
- float *out = reinterpret_cast<float*>(real_result.data);
-
- if(n_channels == 1)
- cufftExecC2R(plan_i_1ch, in, out);
- else if(n_channels == (int) m_num_of_scales)
- cufftExecC2R(plan_i_1ch_all_scales, in, out);
- else if(n_channels == (int) m_num_of_feats * (int) m_num_of_scales)
- cufftExecC2R(plan_i_features_all_scales, in, out);
- else
- cufftExecC2R(plan_i_features, in, out);
+ Fft::inverse(complex_input, real_result);
- return real_result/(m_width*m_height);
+ uint n_channels = complex_input.n_channels;
+ cufftComplex *in = reinterpret_cast<cufftComplex *>(complex_input.get_dev_data());
+ cufftReal *out = real_result.deviceMem();
+ float alpha = 1.0 / (m_width * m_height);
+
+ if (n_channels == 1)
+ cudaErrorCheck(cufftExecC2R(plan_i_1ch, in, out));
+#ifdef BIG_BATCH
+ else
+ cudaErrorCheck(cufftExecC2R(plan_i_all_scales, in, out));
+#endif
+ // TODO: Investigate whether this scalling is needed or not
+ cudaErrorCheck(cublasSscal(cublas, real_result.total(), &alpha, out, 1));
+ CudaSafeCall(cudaStreamSynchronize(cudaStreamPerThread));
}
cuFFT::~cuFFT()
{
+ cudaErrorCheck(cublasDestroy(cublas));
- for(unsigned i = 0; i < m_num_of_streams; i++) gpuErrchk(cudaStreamDestroy(streams[i]));
+ cudaErrorCheck(cufftDestroy(plan_f));
+ cudaErrorCheck(cufftDestroy(plan_fw));
+ cudaErrorCheck(cufftDestroy(plan_i_1ch));
- cudaDeviceReset();
+#ifdef BIG_BATCH
+ cudaErrorCheck(cufftDestroy(plan_f_all_scales));
+ cudaErrorCheck(cufftDestroy(plan_fw_all_scales));
+ cudaErrorCheck(cufftDestroy(plan_i_all_scales));
+#endif
}