generic-library/opencv
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

moved crossCorr (as NPP_Staging wrapper) into public GPU module part from the internal matchTemplate files

Browse Source
This commit is contained in:
Alexey Spizhevoy
2010-12-22 08:56:16 +00:00
parent f9bcef9003
commit fef06c25b5
5 changed files with 166 additions and 141 deletions
Download Patch File Download Diff File Expand all files Collapse all files

131
modules/gpu/src/imgproc_gpu.cpp
View File

@@ -41,6 +41,7 @@
//M*/
#include "precomp.hpp"
#include <utility>
using namespace cv;
using namespace cv::gpu;
@@ -73,6 +74,7 @@ void cv::gpu::histRange(const GpuMat&, GpuMat&, const GpuMat&) { throw_nogpu();
void cv::gpu::histRange(const GpuMat&, GpuMat*, const GpuMat*) { throw_nogpu(); }
void cv::gpu::cornerHarris(const GpuMat&, GpuMat&, int, int, double, int) { throw_nogpu(); }
void cv::gpu::cornerMinEigenVal(const GpuMat&, GpuMat&, int, int, int) { throw_nogpu(); }
void cv::gpu::crossCorr(const GpuMat&, const GpuMat&, GpuMat&) { throw_nogpu(); }
#else /* !defined (HAVE_CUDA) */
@@ -1062,6 +1064,135 @@ void cv::gpu::cornerMinEigenVal(const GpuMat& src, GpuMat& dst, int blockSize, i
imgproc::cornerMinEigenVal_caller(blockSize, Dx, Dy, dst, gpuBorderType);
}
//////////////////////////////////////////////////////////////////////////////
// crossCorr
namespace
{
// Estimates optimal block size
void crossCorrOptBlockSize(int w, int h, int tw, int th, int& bw, int& bh)
{
int major, minor;
getComputeCapability(getDevice(), major, minor);
int scale = 40;
int bh_min = 1024;
int bw_min = 1024;
// Check whether we use Fermi generation or newer GPU
if (major >= 2)
{
bh_min = 2048;
bw_min = 2048;
}
bw = std::max(tw * scale, bw_min);
bh = std::max(th * scale, bh_min);
bw = std::min(bw, w);
bh = std::min(bh, h);
}
}
namespace cv { namespace gpu { namespace imgproc
{
void multiplyAndNormalizeSpects(int n, float scale, const cufftComplex* a,
const cufftComplex* b, cufftComplex* c);
}}}
void cv::gpu::crossCorr(const GpuMat& image, const GpuMat& templ, GpuMat& result)
{
CV_Assert(image.type() == CV_32F);
CV_Assert(templ.type() == CV_32F);
result.create(image.rows - templ.rows + 1, image.cols - templ.cols + 1, CV_32F);
Size block_size;
crossCorrOptBlockSize(result.cols, result.rows, templ.cols, templ.rows,
block_size.width, block_size.height);
Size dft_size;
dft_size.width = getOptimalDFTSize(block_size.width + templ.cols - 1);
dft_size.height = getOptimalDFTSize(block_size.width + templ.rows - 1);
block_size.width = std::min(dft_size.width - templ.cols + 1, result.cols);
block_size.height = std::min(dft_size.height - templ.rows + 1, result.rows);
cufftReal* image_data;
cufftReal* templ_data;
cufftReal* result_data;
cudaSafeCall(cudaMalloc((void**)&image_data, sizeof(cufftReal) * dft_size.area()));
cudaSafeCall(cudaMalloc((void**)&templ_data, sizeof(cufftReal) * dft_size.area()));
cudaSafeCall(cudaMalloc((void**)&result_data, sizeof(cufftReal) * dft_size.area()));
int spect_len = dft_size.height * (dft_size.width / 2 + 1);
cufftComplex* image_spect;
cufftComplex* templ_spect;
cufftComplex* result_spect;
cudaSafeCall(cudaMalloc((void**)&image_spect, sizeof(cufftComplex) * spect_len));
cudaSafeCall(cudaMalloc((void**)&templ_spect, sizeof(cufftComplex) * spect_len));
cudaSafeCall(cudaMalloc((void**)&result_spect, sizeof(cufftComplex) * spect_len));
cufftHandle planR2C, planC2R;
cufftSafeCall(cufftPlan2d(&planC2R, dft_size.height, dft_size.width, CUFFT_C2R));
cufftSafeCall(cufftPlan2d(&planR2C, dft_size.height, dft_size.width, CUFFT_R2C));
GpuMat templ_roi(templ.size(), CV_32S, templ.data, templ.step);
GpuMat templ_block(dft_size, CV_32S, templ_data, dft_size.width * sizeof(cufftReal));
copyMakeBorder(templ_roi, templ_block, 0, templ_block.rows - templ_roi.rows, 0,
templ_block.cols - templ_roi.cols, 0);
cufftSafeCall(cufftExecR2C(planR2C, templ_data, templ_spect));
GpuMat image_block(dft_size, CV_32S, image_data, dft_size.width * sizeof(cufftReal));
// Process all blocks of the result matrix
for (int y = 0; y < result.rows; y += block_size.height)
{
for (int x = 0; x < result.cols; x += block_size.width)
{
// Locate ROI in the source matrix
Size image_roi_size;
image_roi_size.width = std::min(x + dft_size.width, image.cols) - x;
image_roi_size.height = std::min(y + dft_size.height, image.rows) - y;
GpuMat image_roi(image_roi_size, CV_32S, (void*)(image.ptr<float>(y) + x), image.step);
// Make source image block continous
copyMakeBorder(image_roi, image_block, 0, image_block.rows - image_roi.rows, 0,
image_block.cols - image_roi.cols, 0);
cufftSafeCall(cufftExecR2C(planR2C, image_data, image_spect));
imgproc::multiplyAndNormalizeSpects(spect_len, 1.f / dft_size.area(),
image_spect, templ_spect, result_spect);
cufftSafeCall(cufftExecC2R(planC2R, result_spect, result_data));
// Copy result block into appropriate part of the result matrix.
// We can't compute it inplace as the result of the CUFFT transforms
// is always continous, while the result matrix and its blocks can have gaps.
Size result_roi_size;
result_roi_size.width = std::min(x + block_size.width, result.cols) - x;
result_roi_size.height = std::min(y + block_size.height, result.rows) - y;
GpuMat result_roi(result_roi_size, CV_32F, (void*)(result.ptr<float>(y) + x), result.step);
GpuMat result_block(result_roi_size, CV_32F, result_data, dft_size.width * sizeof(cufftReal));
result_block.copyTo(result_roi);
}
}
cufftSafeCall(cufftDestroy(planR2C));
cufftSafeCall(cufftDestroy(planC2R));
cudaSafeCall(cudaFree(image_spect));
cudaSafeCall(cudaFree(templ_spect));
cudaSafeCall(cudaFree(result_spect));
cudaSafeCall(cudaFree(image_data));
cudaSafeCall(cudaFree(templ_data));
cudaSafeCall(cudaFree(result_data));
}
#endif /* !defined (HAVE_CUDA) */