potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 07:43:03 +04:00
Code Packages Releases Wiki Activity

sobel and scharr ocl kernel optimization

Browse Source 
It improves 108%~230% performance in the perf test
with image format "CV_8UC1" and kernel size 3.

Signed-off-by: Li Peng <peng.li@intel.com>
This commit is contained in:
Li Peng
2016-10-27 12:41:40 +08:00
parent acaa1e1b29
commit 8d4a7d3dcc
3 changed files with 319 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/imgproc/src/deriv.cpp
+70
View File
@@ -517,6 +517,66 @@ static bool ipp_sobel(InputArray _src, OutputArray _dst, int ddepth, int dx, int
}
#endif
#ifdef HAVE_OPENCL
namespace cv
{
static bool ocl_sepFilter3x3_8UC1(InputArray _src, OutputArray _dst, int ddepth,
InputArray _kernelX, InputArray _kernelY, double delta, int borderType)
{
const ocl::Device & dev = ocl::Device::getDefault();
int type = _src.type(), sdepth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
if ( !(dev.isIntel() && (type == CV_8UC1) && (ddepth == CV_8U) &&
(_src.offset() == 0) && (_src.step() % 4 == 0) &&
(_src.cols() % 16 == 0) && (_src.rows() % 2 == 0)) )
return false;
Mat kernelX = _kernelX.getMat().reshape(1, 1);
if (kernelX.cols % 2 != 1)
return false;
Mat kernelY = _kernelY.getMat().reshape(1, 1);
if (kernelY.cols % 2 != 1)
return false;
if (ddepth < 0)
ddepth = sdepth;
Size size = _src.size();
size_t globalsize[2] = { 0, 0 };
size_t localsize[2] = { 0, 0 };
globalsize[0] = size.width / 16;
globalsize[1] = size.height / 2;
const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", 0, "BORDER_REFLECT_101" };
char build_opts[1024];
sprintf(build_opts, "-D %s %s%s", borderMap[borderType],
ocl::kernelToStr(kernelX, CV_32F, "KERNEL_MATRIX_X").c_str(),
ocl::kernelToStr(kernelY, CV_32F, "KERNEL_MATRIX_Y").c_str());
ocl::Kernel kernel("sepFilter3x3_8UC1_cols16_rows2", cv::ocl::imgproc::sepFilter3x3_oclsrc, build_opts);
if (kernel.empty())
return false;
UMat src = _src.getUMat();
_dst.create(size, CV_MAKETYPE(ddepth, cn));
if (!(_dst.offset() == 0 && _dst.step() % 4 == 0))
return false;
UMat dst = _dst.getUMat();
int idxArg = kernel.set(0, ocl::KernelArg::PtrReadOnly(src));
idxArg = kernel.set(idxArg, (int)src.step);
idxArg = kernel.set(idxArg, ocl::KernelArg::PtrWriteOnly(dst));
idxArg = kernel.set(idxArg, (int)dst.step);
idxArg = kernel.set(idxArg, (int)dst.rows);
idxArg = kernel.set(idxArg, (int)dst.cols);
idxArg = kernel.set(idxArg, static_cast<float>(delta));
return kernel.run(2, globalsize, (localsize[0] == 0) ? NULL : localsize, false);
}
}
#endif
void cv::Sobel( InputArray _src, OutputArray _dst, int ddepth, int dx, int dy,
int ksize, double scale, double delta, int borderType )
{
@@ -554,6 +614,11 @@ void cv::Sobel( InputArray _src, OutputArray _dst, int ddepth, int dx, int dy,
else
ky *= scale;
}
CV_OCL_RUN(_dst.isUMat() && _src.dims() <= 2 && ksize == 3 &&
(size_t)_src.rows() > ky.total() && (size_t)_src.cols() > kx.total(),
ocl_sepFilter3x3_8UC1(_src, _dst, ddepth, kx, ky, delta, borderType));
sepFilter2D( _src, _dst, ddepth, kx, ky, Point(-1, -1), delta, borderType );
}
@@ -593,6 +658,11 @@ void cv::Scharr( InputArray _src, OutputArray _dst, int ddepth, int dx, int dy,
else
ky *= scale;
}
CV_OCL_RUN(_dst.isUMat() && _src.dims() <= 2 &&
(size_t)_src.rows() > ky.total() && (size_t)_src.cols() > kx.total(),
ocl_sepFilter3x3_8UC1(_src, _dst, ddepth, kx, ky, delta, borderType));
sepFilter2D( _src, _dst, ddepth, kx, ky, Point(-1, -1), delta, borderType );
}