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

Fix 2.4 ocl Canny.

Browse Source 
This fix is a workaround for current 2.4 branch without introducing an
additional oclMat buffer into CannyBuf object.
Test case is cleaned up.
Volatile keywords in kernels are removed for performance concern.
This commit is contained in:
peng xiao
2013-05-29 14:15:26 +08:00
parent 2ccdf56119
commit d015fa76fa
3 changed files with 45 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
modules/ocl/src/canny.cpp
View File

@@ -87,7 +87,7 @@ void cv::ocl::CannyBuf::create(const Size &image_size, int apperture_size)
filterDY = createDerivFilter_GPU(CV_8U, CV_32S, 0, 1, apperture_size, BORDER_REPLICATE);
}
}
ensureSizeIsEnough(image_size.height + 2, image_size.width + 2, CV_32FC1, edgeBuf);
ensureSizeIsEnough(2 * (image_size.height + 2), image_size.width + 2, CV_32FC1, edgeBuf);
ensureSizeIsEnough(1, image_size.width * image_size.height, CV_16UC2, trackBuf1);
ensureSizeIsEnough(1, image_size.width * image_size.height, CV_16UC2, trackBuf2);
@@ -141,13 +141,16 @@ namespace
void CannyCaller(CannyBuf &buf, oclMat &dst, float low_thresh, float high_thresh)
{
using namespace ::cv::ocl::canny;
calcMap_gpu(buf.dx, buf.dy, buf.edgeBuf, buf.edgeBuf, dst.rows, dst.cols, low_thresh, high_thresh);
oclMat magBuf = buf.edgeBuf(Rect(0, 0, buf.edgeBuf.cols, buf.edgeBuf.rows / 2));
oclMat mapBuf = buf.edgeBuf(Rect(0, buf.edgeBuf.rows / 2, buf.edgeBuf.cols, buf.edgeBuf.rows / 2));
edgesHysteresisLocal_gpu(buf.edgeBuf, buf.trackBuf1, buf.counter, dst.rows, dst.cols);
calcMap_gpu(buf.dx, buf.dy, magBuf, mapBuf, dst.rows, dst.cols, low_thresh, high_thresh);
edgesHysteresisGlobal_gpu(buf.edgeBuf, buf.trackBuf1, buf.trackBuf2, buf.counter, dst.rows, dst.cols);
edgesHysteresisLocal_gpu(mapBuf, buf.trackBuf1, buf.counter, dst.rows, dst.cols);
getEdges_gpu(buf.edgeBuf, dst, dst.rows, dst.cols);
edgesHysteresisGlobal_gpu(mapBuf, buf.trackBuf1, buf.trackBuf2, buf.counter, dst.rows, dst.cols);
getEdges_gpu(mapBuf, dst, dst.rows, dst.cols);
}
}
@@ -172,18 +175,20 @@ void cv::ocl::Canny(const oclMat &src, CannyBuf &buf, oclMat &dst, double low_th
buf.create(src.size(), apperture_size);
buf.edgeBuf.setTo(Scalar::all(0));
oclMat magBuf = buf.edgeBuf(Rect(0, 0, buf.edgeBuf.cols, buf.edgeBuf.rows / 2));
if (apperture_size == 3)
{
calcSobelRowPass_gpu(src, buf.dx_buf, buf.dy_buf, src.rows, src.cols);
calcMagnitude_gpu(buf.dx_buf, buf.dy_buf, buf.dx, buf.dy, buf.edgeBuf, src.rows, src.cols, L2gradient);
calcMagnitude_gpu(buf.dx_buf, buf.dy_buf, buf.dx, buf.dy, magBuf, src.rows, src.cols, L2gradient);
}
else
{
buf.filterDX->apply(src, buf.dx);
buf.filterDY->apply(src, buf.dy);
calcMagnitude_gpu(buf.dx, buf.dy, buf.edgeBuf, src.rows, src.cols, L2gradient);
calcMagnitude_gpu(buf.dx, buf.dy, magBuf, src.rows, src.cols, L2gradient);
}
CannyCaller(buf, dst, static_cast<float>(low_thresh), static_cast<float>(high_thresh));
}
@@ -209,7 +214,10 @@ void cv::ocl::Canny(const oclMat &dx, const oclMat &dy, CannyBuf &buf, oclMat &d
buf.dy = dy;
buf.create(dx.size(), -1);
buf.edgeBuf.setTo(Scalar::all(0));
calcMagnitude_gpu(buf.dx, buf.dy, buf.edgeBuf, dx.rows, dx.cols, L2gradient);
oclMat magBuf = buf.edgeBuf(Rect(0, 0, buf.edgeBuf.cols, buf.edgeBuf.rows / 2));
calcMagnitude_gpu(buf.dx, buf.dy, magBuf, dx.rows, dx.cols, L2gradient);
CannyCaller(buf, dst, static_cast<float>(low_thresh), static_cast<float>(high_thresh));
}
@@ -234,7 +242,7 @@ void canny::calcSobelRowPass_gpu(const oclMat &src, oclMat &dx_buf, oclMat &dy_b
size_t globalThreads[3] = {cols, rows, 1};
size_t localThreads[3] = {16, 16, 1};
openCLExecuteKernel2(clCxt, &imgproc_canny, kernelName, globalThreads, localThreads, args, -1, -1);
openCLExecuteKernel(clCxt, &imgproc_canny, kernelName, globalThreads, localThreads, args, -1, -1);
}
void canny::calcMagnitude_gpu(const oclMat &dx_buf, const oclMat &dy_buf, oclMat &dx, oclMat &dy, oclMat &mag, int rows, int cols, bool L2Grad)
@@ -264,12 +272,8 @@ void canny::calcMagnitude_gpu(const oclMat &dx_buf, const oclMat &dy_buf, oclMat
size_t globalThreads[3] = {cols, rows, 1};
size_t localThreads[3] = {16, 16, 1};
char build_options [15] = "";
if(L2Grad)
{
strcat(build_options, "-D L2GRAD");
}
openCLExecuteKernel2(clCxt, &imgproc_canny, kernelName, globalThreads, localThreads, args, -1, -1, build_options);
const char * build_options = L2Grad ? "-D L2GRAD":"";
openCLExecuteKernel(clCxt, &imgproc_canny, kernelName, globalThreads, localThreads, args, -1, -1, build_options);
}
void canny::calcMagnitude_gpu(const oclMat &dx, const oclMat &dy, oclMat &mag, int rows, int cols, bool L2Grad)
{
@@ -292,12 +296,8 @@ void canny::calcMagnitude_gpu(const oclMat &dx, const oclMat &dy, oclMat &mag, i
size_t globalThreads[3] = {cols, rows, 1};
size_t localThreads[3] = {16, 16, 1};
char build_options [15] = "";
if(L2Grad)
{
strcat(build_options, "-D L2GRAD");
}
openCLExecuteKernel2(clCxt, &imgproc_canny, kernelName, globalThreads, localThreads, args, -1, -1, build_options);
const char * build_options = L2Grad ? "-D L2GRAD":"";
openCLExecuteKernel(clCxt, &imgproc_canny, kernelName, globalThreads, localThreads, args, -1, -1, build_options);
}
void canny::calcMap_gpu(oclMat &dx, oclMat &dy, oclMat &mag, oclMat &map, int rows, int cols, float low_thresh, float high_thresh)
@@ -328,7 +328,7 @@ void canny::calcMap_gpu(oclMat &dx, oclMat &dy, oclMat &mag, oclMat &map, int ro
string kernelName = "calcMap";
size_t localThreads[3] = {16, 16, 1};
openCLExecuteKernel2(clCxt, &imgproc_canny, kernelName, globalThreads, localThreads, args, -1, -1);
openCLExecuteKernel(clCxt, &imgproc_canny, kernelName, globalThreads, localThreads, args, -1, -1);
}
void canny::edgesHysteresisLocal_gpu(oclMat &map, oclMat &st1, void *counter, int rows, int cols)
@@ -348,7 +348,7 @@ void canny::edgesHysteresisLocal_gpu(oclMat &map, oclMat &st1, void *counter, in
size_t globalThreads[3] = {cols, rows, 1};
size_t localThreads[3] = {16, 16, 1};
openCLExecuteKernel2(clCxt, &imgproc_canny, kernelName, globalThreads, localThreads, args, -1, -1);
openCLExecuteKernel(clCxt, &imgproc_canny, kernelName, globalThreads, localThreads, args, -1, -1);
}
void canny::edgesHysteresisGlobal_gpu(oclMat &map, oclMat &st1, oclMat &st2, void *counter, int rows, int cols)
@@ -378,7 +378,7 @@ void canny::edgesHysteresisGlobal_gpu(oclMat &map, oclMat &st1, oclMat &st2, voi
args.push_back( make_pair( sizeof(cl_int), (void *)&map.step));
args.push_back( make_pair( sizeof(cl_int), (void *)&map.offset));
openCLExecuteKernel2(clCxt, &imgproc_canny, kernelName, globalThreads, localThreads, args, -1, -1, DISABLE);
openCLExecuteKernel(clCxt, &imgproc_canny, kernelName, globalThreads, localThreads, args, -1, -1);
openCLSafeCall(clEnqueueReadBuffer(*(cl_command_queue*)getoclCommandQueue(), (cl_mem)counter, 1, 0, sizeof(int), &count, 0, NULL, NULL));
std::swap(st1, st2);
}
@@ -403,5 +403,5 @@ void canny::getEdges_gpu(oclMat &map, oclMat &dst, int rows, int cols)
size_t globalThreads[3] = {cols, rows, 1};
size_t localThreads[3] = {16, 16, 1};
openCLExecuteKernel2(clCxt, &imgproc_canny, kernelName, globalThreads, localThreads, args, -1, -1);
openCLExecuteKernel(clCxt, &imgproc_canny, kernelName, globalThreads, localThreads, args, -1, -1);
}