#include "fft_cufft.h"
+cuFFT::cuFFT()
+{
+ CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
+ cudaErrorCheck(cublasCreate(&cublas));
+ cudaErrorCheck(cublasSetStream(cublas, cudaStreamPerThread));
+}
+
+cufftHandle cuFFT::create_plan_fwd(uint howmany) const
+{
+ 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};
+
+ cufftHandle plan;
+ cudaErrorCheck(cufftPlanMany(&plan, rank, n, inembed, istride, idist, onembed, ostride, odist, CUFFT_R2C, howmany));
+ cudaErrorCheck(cufftSetStream(plan, cudaStreamPerThread));
+ return plan;
+}
+
+cufftHandle cuFFT::create_plan_inv(uint howmany) const
+{
+ int rank = 2;
+ int n[] = {(int)m_height, (int)m_width};
+ int idist = m_height * (m_width / 2 + 1), odist = m_height * m_width;
+ int istride = 1, ostride = 1;
+ int inembed[] = {(int)m_height, (int)m_width / 2 + 1}, *onembed = n;
+
+ cufftHandle plan;
+ cudaErrorCheck(cufftPlanMany(&plan, rank, n, inembed, istride, idist, onembed, ostride, odist, CUFFT_C2R, howmany));
+ cudaErrorCheck(cufftSetStream(plan, cudaStreamPerThread));
+ return plan;
+}
+
+
void cuFFT::init(unsigned width, unsigned height, unsigned num_of_feats, unsigned num_of_scales)
{
- m_width = width;
- m_height = height;
- m_num_of_feats = num_of_feats;
- m_num_of_scales = num_of_scales;
+ Fft::init(width, height, num_of_feats, num_of_scales);
std::cout << "FFT: cuFFT" << std::endl;
- // FFT forward one scale
- {
- CufftErrorCheck(cufftPlan2d(&plan_f, int(m_height), int(m_width), CUFFT_R2C));
- }
-#ifdef BIG_BATCH
- // FFT forward all scales
- if (m_num_of_scales > 1 && BIG_BATCH_MODE) {
- 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 istride = 1, ostride = 1;
- int *inembed = n, onembed[] = {(int)m_height, (int)m_width / 2 + 1};
-
- CufftErrorCheck(cufftPlanMany(&plan_f_all_scales, rank, n, inembed, istride, idist, onembed, ostride, odist,
- CUFFT_R2C, howmany));
- CufftErrorCheck(cufftSetStream(plan_f_all_scales, cudaStreamPerThread));
- }
-#endif
- // FFT forward window one scale
- {
- int rank = 2;
- int n[] = {int(m_height), int(m_width)};
- int howmany = int(m_num_of_feats);
- int idist = int(m_height * m_width), odist = int(m_height * (m_width / 2 + 1));
- int istride = 1, ostride = 1;
- int *inembed = n, onembed[] = {int(m_height), int(m_width / 2 + 1)};
-
- CufftErrorCheck(
- cufftPlanMany(&plan_fw, rank, n, inembed, istride, idist, onembed, ostride, odist, CUFFT_R2C, howmany));
- CufftErrorCheck(cufftSetStream(plan_fw, cudaStreamPerThread));
- }
-#ifdef BIG_BATCH
- // FFT forward window all scales all feats
- if (m_num_of_scales > 1 && BIG_BATCH_MODE) {
- 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 = n, onembed[] = {(int)m_height, (int)m_width / 2 + 1};
-
- CufftErrorCheck(cufftPlanMany(&plan_fw_all_scales, rank, n, inembed, istride, idist, onembed, ostride, odist,
- CUFFT_R2C, howmany));
- CufftErrorCheck(cufftSetStream(plan_fw_all_scales, cudaStreamPerThread));
- }
-#endif
- // FFT inverse one scale
- {
- int rank = 2;
- int n[] = {int(m_height), int(m_width)};
- int howmany = int(m_num_of_feats);
- int idist = int(m_height * (m_width / 2 + 1)), odist = 1;
- int istride = 1, ostride = int(m_num_of_feats);
- int inembed[] = {int(m_height), int(m_width / 2 + 1)}, *onembed = n;
-
- CufftErrorCheck(cufftPlanMany(&plan_i_features, rank, n, inembed, istride, idist, onembed, ostride, odist,
- CUFFT_C2R, howmany));
- CufftErrorCheck(cufftSetStream(plan_i_features, cudaStreamPerThread));
- }
- // FFT inverse all scales
+ 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
- if (m_num_of_scales > 1 && BIG_BATCH_MODE) {
- 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;
-
- CufftErrorCheck(cufftPlanMany(&plan_i_features_all_scales, rank, n, inembed, istride, idist, onembed, ostride,
- odist, CUFFT_C2R, howmany));
- CufftErrorCheck(cufftSetStream(plan_i_features_all_scales, cudaStreamPerThread));
- }
-#endif
- // FFT inverse one channel one scale
- {
- int rank = 2;
- int n[] = {int(m_height), int(m_width)};
- int howmany = 1;
- int idist = int(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;
-
- CufftErrorCheck(
- cufftPlanMany(&plan_i_1ch, rank, n, inembed, istride, idist, onembed, ostride, odist, CUFFT_C2R, howmany));
- CufftErrorCheck(cufftSetStream(plan_i_1ch, cudaStreamPerThread));
- }
-#ifdef BIG_BATCH
- // FFT inverse one channel all scales
- if (m_num_of_scales > 1 && BIG_BATCH_MODE) {
- 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;
-
- CufftErrorCheck(cufftPlanMany(&plan_i_1ch_all_scales, rank, n, inembed, istride, idist, onembed, ostride, odist,
- CUFFT_C2R, howmany));
- CufftErrorCheck(cufftSetStream(plan_i_1ch_all_scales, cudaStreamPerThread));
- }
+ 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
}
void cuFFT::set_window(const MatDynMem &window)
{
+ Fft::set_window(window);
m_window = window;
}
-void cuFFT::forward(const cv::Mat &real_input, ComplexMat &complex_result, float *real_input_arr)
+void cuFFT::forward(const MatScales &real_input, ComplexMat &complex_result)
{
- if (BIG_BATCH_MODE && real_input.rows == int(m_height * m_num_of_scales)) {
- CufftErrorCheck(cufftExecR2C(plan_f_all_scales, reinterpret_cast<cufftReal *>(real_input_arr),
- complex_result.get_p_data()));
- } else {
- NORMAL_OMP_CRITICAL
- {
- CufftErrorCheck(
- cufftExecR2C(plan_f, reinterpret_cast<cufftReal *>(real_input_arr), complex_result.get_p_data()));
- cudaStreamSynchronize(cudaStreamPerThread);
- }
- }
- return;
+ 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
+ cudaErrorCheck(cufftExecR2C(plan_f_all_scales, in, complex_result.get_dev_data()));
+#endif
}
-void cuFFT::forward_window(MatDynMem &patch_feats_in, ComplexMat & complex_result, MatDynMem &tmp)
+void cuFFT::forward_window(MatScaleFeats &feat, ComplexMat &complex_result, MatScaleFeats &temp)
{
- int n_channels = int(patch_feats.size());
+ Fft::forward_window(feat, complex_result, temp);
- if (n_channels > int(m_num_of_feats)) {
- for (uint i = 0; i < uint(n_channels); ++i) {
- cv::Mat in_roi(fw_all, cv::Rect(0, int(i * m_height), int(m_width), int(m_height)));
- in_roi = patch_feats[i].mul(m_window);
- }
- CufftErrorCheck(cufftExecR2C(plan_fw_all_scales, reinterpret_cast<cufftReal *>(real_input_arr),
- complex_result.get_p_data()));
- } else {
- for (uint i = 0; i < uint(n_channels); ++i) {
- cv::Mat in_roi(fw_all, cv::Rect(0, int(i * m_height), int(m_width), int(m_height)));
- in_roi = patch_feats[i].mul(m_window);
- }
- NORMAL_OMP_CRITICAL
- {
- CufftErrorCheck(
- cufftExecR2C(plan_fw, reinterpret_cast<cufftReal *>(real_input_arr), complex_result.get_p_data()));
- cudaStreamSynchronize(cudaStreamPerThread);
+ 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;
+
+ 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
}
-void cuFFT::inverse(ComplexMat & complex_input, MatDynMem & real_result)
+void cuFFT::inverse(ComplexMat &complex_input, MatScales &real_result)
{
- int n_channels = complex_input.n_channels;
- cufftComplex *in = reinterpret_cast<cufftComplex *>(complex_input.get_p_data());
-
- if (n_channels == 1) {
- NORMAL_OMP_CRITICAL
- {
- CufftErrorCheck(cufftExecC2R(plan_i_1ch, in, reinterpret_cast<cufftReal *>(real_result_arr)));
- cudaStreamSynchronize(cudaStreamPerThread);
- }
- real_result = real_result / (m_width * m_height);
- return;
- } else if (n_channels == int(m_num_of_scales)) {
- CufftErrorCheck(cufftExecC2R(plan_i_1ch_all_scales, in, reinterpret_cast<cufftReal *>(real_result_arr)));
- cudaStreamSynchronize(cudaStreamPerThread);
-
- real_result = real_result / (m_width * m_height);
- return;
- } else if (n_channels == int(m_num_of_feats) * int(m_num_of_scales)) {
- CufftErrorCheck(cufftExecC2R(plan_i_features_all_scales, in, reinterpret_cast<cufftReal *>(real_result_arr)));
- cudaStreamSynchronize(cudaStreamPerThread);
- return;
- }
- NORMAL_OMP_CRITICAL
- {
- CufftErrorCheck(cufftExecC2R(plan_i_features, in, reinterpret_cast<cufftReal *>(real_result_arr)));
-#if defined(OPENMP) && !defined(BIG_BATCH)
- CudaSafeCall(cudaStreamSynchronize(cudaStreamPerThread));
+ Fft::inverse(complex_input, real_result);
+
+ 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
- }
- return;
+ cudaErrorCheck(cublasSscal(cublas, real_result.total(), &alpha, out, 1));
+ // The result is a cv::Mat, which will be accesses by CPU, so we
+ // must synchronize with the GPU here
+ CudaSafeCall(cudaStreamSynchronize(cudaStreamPerThread));
}
cuFFT::~cuFFT()
{
- CufftErrorCheck(cufftDestroy(plan_f));
- CufftErrorCheck(cufftDestroy(plan_fw));
- CufftErrorCheck(cufftDestroy(plan_i_1ch));
- CufftErrorCheck(cufftDestroy(plan_i_features));
-
- if (BIG_BATCH_MODE) {
- CufftErrorCheck(cufftDestroy(plan_f_all_scales));
- CufftErrorCheck(cufftDestroy(plan_fw_all_scales));
- CufftErrorCheck(cufftDestroy(plan_i_1ch_all_scales));
- CufftErrorCheck(cufftDestroy(plan_i_features_all_scales));
- }
+ cudaErrorCheck(cublasDestroy(cublas));
+
+ cudaErrorCheck(cufftDestroy(plan_f));
+ cudaErrorCheck(cufftDestroy(plan_fw));
+ cudaErrorCheck(cufftDestroy(plan_i_1ch));
+
+#ifdef BIG_BATCH
+ cudaErrorCheck(cufftDestroy(plan_f_all_scales));
+ cudaErrorCheck(cufftDestroy(plan_fw_all_scales));
+ cudaErrorCheck(cufftDestroy(plan_i_all_scales));
+#endif
}
projects / hercules2020 / kcf.git / blobdiff