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

optimization of cv::accumulate**

Browse Source
This commit is contained in:
Ilya Lavrenov
2014-07-09 19:00:33 +04:00
parent 4b84d5d2c1
commit a350b76738
2 changed files with 59 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
modules/imgproc/src/accum.cpp
View File

@@ -369,11 +369,17 @@ static bool ocl_accumulate( InputArray _src, InputArray _src2, InputOutputArray
CV_Assert(op_type == ACCUMULATE || op_type == ACCUMULATE_SQUARE ||
op_type == ACCUMULATE_PRODUCT || op_type == ACCUMULATE_WEIGHTED);
int stype = _src.type(), cn = CV_MAT_CN(stype);
int sdepth = CV_MAT_DEPTH(stype), ddepth = _dst.depth();
const ocl::Device & dev = ocl::Device::getDefault();
int vectorWidths[] = { 4, 4, 2, 2, 1, 1, 1, -1 };
int stype = _src.type(), sdepth = CV_MAT_DEPTH(stype), cn = CV_MAT_CN(stype), ddepth = _dst.depth();
int pcn = std::max(vectorWidths[sdepth], vectorWidths[ddepth]), sesz = CV_ELEM_SIZE(sdepth) * pcn,
desz = CV_ELEM_SIZE(ddepth) * pcn, rowsPerWI = dev.isIntel() ? 4 : 1;
bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0,
haveMask = !_mask.empty();
bool doubleSupport = dev.doubleFPConfig() > 0, haveMask = !_mask.empty(),
usepcn = _src.offset() % sesz == 0 && _src.step() % sesz == 0 && (_src.cols() * cn) % pcn == 0 &&
_src2.offset() % desz == 0 && _src2.step() % desz == 0 &&
_dst.offset() % pcn == 0 && _dst.step() % desz == 0 && !haveMask;
int kercn = usepcn ? pcn : haveMask ? cn : 1;
if (!doubleSupport && (sdepth == CV_64F || ddepth == CV_64F))
return false;
@@ -381,11 +387,13 @@ static bool ocl_accumulate( InputArray _src, InputArray _src2, InputOutputArray
const char * const opMap[4] = { "ACCUMULATE", "ACCUMULATE_SQUARE", "ACCUMULATE_PRODUCT",
"ACCUMULATE_WEIGHTED" };
char cvt[40];
ocl::Kernel k("accumulate", ocl::imgproc::accumulate_oclsrc,
format("-D %s%s -D srcT=%s -D cn=%d -D dstT=%s%s",
format("-D %s%s -D srcT1=%s -D cn=%d -D dstT1=%s%s -D rowsPerWI=%d -D convertToDT=%s",
opMap[op_type], haveMask ? " -D HAVE_MASK" : "",
ocl::typeToStr(sdepth), cn, ocl::typeToStr(ddepth),
doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
ocl::typeToStr(sdepth), kercn, ocl::typeToStr(ddepth),
doubleSupport ? " -D DOUBLE_SUPPORT" : "", rowsPerWI,
ocl::convertTypeStr(sdepth, ddepth, 1, cvt)));
if (k.empty())
return false;
@@ -393,7 +401,7 @@ static bool ocl_accumulate( InputArray _src, InputArray _src2, InputOutputArray
ocl::KernelArg srcarg = ocl::KernelArg::ReadOnlyNoSize(src),
src2arg = ocl::KernelArg::ReadOnlyNoSize(src2),
dstarg = ocl::KernelArg::ReadWrite(dst),
dstarg = ocl::KernelArg::ReadWrite(dst, cn, kercn),
maskarg = ocl::KernelArg::ReadOnlyNoSize(mask);
int argidx = k.set(0, srcarg);
@@ -410,7 +418,7 @@ static bool ocl_accumulate( InputArray _src, InputArray _src2, InputOutputArray
if (haveMask)
k.set(argidx, maskarg);
size_t globalsize[2] = { src.cols, src.rows };
size_t globalsize[2] = { src.cols * cn / kercn, (src.rows + rowsPerWI - 1) / rowsPerWI };
return k.run(2, globalsize, NULL, false);
}