Chiebot-Mirror
/
opencv
mirror of https://github.com/opencv/opencv.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity

added buffered version of gpu::convolve

Browse Source
pull/13383/head
Alexey Spizhevoy 14 years ago
parent e5d1b9eecd
commit be38864dd0
2 changed files with 99 additions and 45 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 29
      modules/gpu/include/opencv2/gpu/gpu.hpp
  2. 115
      modules/gpu/src/imgproc_gpu.cpp

29
modules/gpu/include/opencv2/gpu/gpu.hpp
Unescape View File

@ -656,7 +656,34 @@ namespace cv
//! computes convolution (or cross-correlation) of two images using discrete Fourier transform
//! supports source images of 32FC1 type only
//! result matrix will have 32FC1 type
CV_EXPORTS void convolve(const GpuMat& image, const GpuMat& templ, GpuMat& result, bool ccorr=false);
CV_EXPORTS void convolve(const GpuMat& image, const GpuMat& templ, GpuMat& result,
bool ccorr=false);
struct CV_EXPORTS ConvolveBuf;
//! buffered version
CV_EXPORTS void convolve(const GpuMat& image, const GpuMat& templ, GpuMat& result,
bool ccorr, ConvolveBuf& buf);
struct CV_EXPORTS ConvolveBuf
{
ConvolveBuf() {}
ConvolveBuf(Size image_size, Size templ_size)
{ create(image_size, templ_size); }
void create(Size image_size, Size templ_size);
private:
static Size estimateBlockSize(Size result_size, Size templ_size);
friend void convolve(const GpuMat&, const GpuMat&, GpuMat&, bool, ConvolveBuf&);
Size result_size;
Size block_size;
Size dft_size;
int spect_len;
GpuMat image_spect, templ_spect, result_spect;
GpuMat image_block, templ_block, result_data;
};
//! computes the proximity map for the raster template and the image where the template is searched for
CV_EXPORTS void matchTemplate(const GpuMat& image, const GpuMat& templ, GpuMat& result, int method);

115
modules/gpu/src/imgproc_gpu.cpp
Unescape View File

@ -77,7 +77,9 @@ void cv::gpu::cornerMinEigenVal(const GpuMat&, GpuMat&, int, int, int) { throw_n
void cv::gpu::mulSpectrums(const GpuMat&, const GpuMat&, GpuMat&, int, bool) { throw_nogpu(); }
void cv::gpu::mulAndScaleSpectrums(const GpuMat&, const GpuMat&, GpuMat&, int, float, bool) { throw_nogpu(); }
void cv::gpu::dft(const GpuMat&, GpuMat&, Size, int) { throw_nogpu(); }
void cv::gpu::ConvolveBuf::create(Size, Size) { throw_nogpu(); }
void cv::gpu::convolve(const GpuMat&, const GpuMat&, GpuMat&, bool) { throw_nogpu(); }
void cv::gpu::convolve(const GpuMat&, const GpuMat&, GpuMat&, bool, ConvolveBuf&) { throw_nogpu(); }
#else /* !defined (HAVE_CUDA) */
@ -1211,36 +1213,65 @@ void cv::gpu::dft(const GpuMat& src, GpuMat& dst, Size dft_size, int flags)
}
//////////////////////////////////////////////////////////////////////////////
// crossCorr
// convolve
namespace
void cv::gpu::ConvolveBuf::create(Size image_size, Size templ_size)
{
// Estimates optimal block size
void convolveOptBlockSize(int w, int h, int tw, int th, int& bw, int& bh)
{
int major, minor;
getComputeCapability(getDevice(), major, minor);
result_size = Size(image_size.width - templ_size.width + 1,
image_size.height - templ_size.height + 1);
block_size = estimateBlockSize(result_size, templ_size);
dft_size.width = getOptimalDFTSize(block_size.width + templ_size.width - 1);
dft_size.height = getOptimalDFTSize(block_size.width + templ_size.height - 1);
createContinuous(dft_size, CV_32F, image_block);
createContinuous(dft_size, CV_32F, templ_block);
createContinuous(dft_size, CV_32F, result_data);
spect_len = dft_size.height * (dft_size.width / 2 + 1);
createContinuous(1, spect_len, CV_32FC2, image_spect);
createContinuous(1, spect_len, CV_32FC2, templ_spect);
createContinuous(1, spect_len, CV_32FC2, result_spect);
block_size.width = std::min(dft_size.width - templ_size.width + 1, result_size.width);
block_size.height = std::min(dft_size.height - templ_size.height + 1, result_size.height);
}
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;
}
Size cv::gpu::ConvolveBuf::estimateBlockSize(Size result_size, Size templ_size)
{
int major, minor;
getComputeCapability(getDevice(), major, minor);
bw = std::max(tw * scale, bw_min);
bh = std::max(th * scale, bh_min);
bw = std::min(bw, w);
bh = std::min(bh, h);
int scale = 40;
Size bsize_min(1024, 1024);
// Check whether we use Fermi generation or newer GPU
if (major >= 2)
{
bsize_min.width = 2048;
bsize_min.height = 2048;
}
Size bsize(std::max(templ_size.width * scale, bsize_min.width),
std::max(templ_size.height * scale, bsize_min.height));
bsize.width = std::min(bsize.width, result_size.width);
bsize.height = std::min(bsize.height, result_size.height);
return bsize;
}
void cv::gpu::convolve(const GpuMat& image, const GpuMat& templ, GpuMat& result,
bool ccorr)
{
ConvolveBuf buf;
convolve(image, templ, result, ccorr, buf);
}
void cv::gpu::convolve(const GpuMat& image, const GpuMat& templ, GpuMat& result, bool ccorr)
void cv::gpu::convolve(const GpuMat& image, const GpuMat& templ, GpuMat& result,
bool ccorr, ConvolveBuf& buf)
{
StaticAssert<sizeof(float) == sizeof(cufftReal)>::check();
StaticAssert<sizeof(float) * 2 == sizeof(cufftComplex)>::check();
@ -1248,32 +1279,25 @@ void cv::gpu::convolve(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);
buf.create(image.size(), templ.size());
result.create(buf.result_size, CV_32F);
Size block_size;
convolveOptBlockSize(result.cols, result.rows, templ.cols, templ.rows,
block_size.width, block_size.height);
Size& block_size = buf.block_size;
Size& dft_size = buf.dft_size;
int& spect_len = buf.spect_len;
Size dft_size;
dft_size.width = getOptimalDFTSize(block_size.width + templ.cols - 1);
dft_size.height = getOptimalDFTSize(block_size.width + templ.rows - 1);
GpuMat& image_block = buf.image_block;
GpuMat& templ_block = buf.templ_block;
GpuMat& result_data = buf.result_data;
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);
int spect_len = dft_size.height * (dft_size.width / 2 + 1);
GpuMat image_spect = createContinuous(1, spect_len, CV_32FC2);
GpuMat templ_spect = createContinuous(1, spect_len, CV_32FC2);
GpuMat result_spect = createContinuous(1, spect_len, CV_32FC2);
GpuMat& image_spect = buf.image_spect;
GpuMat& templ_spect = buf.templ_spect;
GpuMat& result_spect = buf.result_spect;
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 image_block = createContinuous(dft_size, CV_32F);
GpuMat templ_block = createContinuous(dft_size, CV_32F);
GpuMat result_data = createContinuous(dft_size, CV_32F);
GpuMat templ_roi(templ.size(), CV_32F, templ.data, templ.step);
copyMakeBorder(templ_roi, templ_block, 0, templ_block.rows - templ_roi.rows, 0,
templ_block.cols - templ_roi.cols, 0);
@ -1288,9 +1312,10 @@ void cv::gpu::convolve(const GpuMat& image, const GpuMat& templ, GpuMat& result,
{
Size image_roi_size(std::min(x + dft_size.width, image.cols) - x,
std::min(y + dft_size.height, image.rows) - y);
GpuMat image_roi(image_roi_size, CV_32F, (void*)(image.ptr<float>(y) + x), image.step);
copyMakeBorder(image_roi, image_block, 0, image_block.rows - image_roi.rows, 0,
image_block.cols - image_roi.cols, 0);
GpuMat image_roi(image_roi_size, CV_32F, (void*)(image.ptr<float>(y) + x),
image.step);
copyMakeBorder(image_roi, image_block, 0, image_block.rows - image_roi.rows,
0, image_block.cols - image_roi.cols, 0);
cufftSafeCall(cufftExecR2C(planR2C, image_block.ptr<cufftReal>(),
image_spect.ptr<cufftComplex>()));
@ -1301,8 +1326,10 @@ void cv::gpu::convolve(const GpuMat& image, const GpuMat& templ, GpuMat& result,
Size result_roi_size(std::min(x + block_size.width, result.cols) - x,
std::min(y + block_size.height, result.rows) - y);
GpuMat result_roi(result_roi_size, result.type(), (void*)(result.ptr<float>(y) + x), result.step);
GpuMat result_block(result_roi_size, result_data.type(), result_data.ptr(), result_data.step);
GpuMat result_roi(result_roi_size, result.type(),
(void*)(result.ptr<float>(y) + x), result.step);
GpuMat result_block(result_roi_size, result_data.type(),
result_data.ptr(), result_data.step);
result_block.copyTo(result_roi);
}
}

Write Preview
Loading…
Cancel
Save