Merge pull request #2851 from ilya-lavrenov:tapi_set_identity

This commit is contained in:
Alexander Alekhin
2014-07-11 10:00:15 +00:00
2 changed files with 48 additions and 13 deletions

View File

@@ -2758,21 +2758,30 @@ namespace cv {
static bool ocl_setIdentity( InputOutputArray _m, const Scalar& s )
{
int type = _m.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type),
sctype = CV_MAKE_TYPE(depth, cn == 3 ? 4 : cn),
int type = _m.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type), kercn = cn;
if (cn == 1)
{
kercn = std::min(ocl::predictOptimalVectorWidth(_m), 4);
if (kercn != 4)
kercn = 1;
}
int sctype = CV_MAKE_TYPE(depth, cn == 3 ? 4 : cn),
rowsPerWI = ocl::Device::getDefault().isIntel() ? 4 : 1;
ocl::Kernel k("setIdentity", ocl::core::set_identity_oclsrc,
format("-D T=%s -D T1=%s -D cn=%d -D ST=%s", ocl::memopTypeToStr(type),
ocl::memopTypeToStr(depth), cn, ocl::memopTypeToStr(sctype)));
format("-D T=%s -D T1=%s -D cn=%d -D ST=%s -D kercn=%d -D rowsPerWI=%d",
ocl::memopTypeToStr(CV_MAKE_TYPE(depth, kercn)),
ocl::memopTypeToStr(depth), cn,
ocl::memopTypeToStr(sctype),
kercn, rowsPerWI));
if (k.empty())
return false;
UMat m = _m.getUMat();
k.args(ocl::KernelArg::WriteOnly(m), ocl::KernelArg::Constant(Mat(1, 1, sctype, s)),
rowsPerWI);
k.args(ocl::KernelArg::WriteOnly(m, cn, kercn),
ocl::KernelArg::Constant(Mat(1, 1, sctype, s)));
size_t globalsize[2] = { m.cols, (m.rows + rowsPerWI - 1) / rowsPerWI };
size_t globalsize[2] = { m.cols * cn / kercn, (m.rows + rowsPerWI - 1) / rowsPerWI };
return k.run(2, globalsize, NULL, false);
}