1
0
mirror of https://github.com/opencv/opencv.git synced 2026-07-29 15:23:05 +04:00

minor optimization of cv::LUT

This commit is contained in:
Ilya Lavrenov
2014-07-02 17:06:34 +04:00
parent 15197eb09b
commit 10c772fa7f
2 changed files with 76 additions and 76 deletions
+5 -9
View File
@@ -1729,22 +1729,18 @@ static bool ocl_LUT(InputArray _src, InputArray _lut, OutputArray _dst)
UMat src = _src.getUMat(), lut = _lut.getUMat();
_dst.create(src.size(), CV_MAKETYPE(ddepth, dcn));
UMat dst = _dst.getUMat();
bool bAligned = (1 == lcn) && (0 == (src.offset % 4)) && (0 == ((dcn * src.cols) % 4));
// dst.cols == src.cols by params of dst.create
int kercn = lcn == 1 ? std::min(4, ocl::predictOptimalVectorWidth(_dst)) : dcn;
ocl::Kernel k("LUT", ocl::core::lut_oclsrc,
format("-D dcn=%d -D lcn=%d -D srcT=%s -D dstT=%s", bAligned ? 4 : dcn, lcn,
ocl::typeToStr(src.depth()), ocl::memopTypeToStr(ddepth)
));
format("-D dcn=%d -D lcn=%d -D srcT=%s -D dstT=%s", kercn, lcn,
ocl::typeToStr(src.depth()), ocl::memopTypeToStr(ddepth)));
if (k.empty())
return false;
int cols = bAligned ? dcn * dst.cols / 4 : dst.cols;
k.args(ocl::KernelArg::ReadOnlyNoSize(src), ocl::KernelArg::ReadOnlyNoSize(lut),
ocl::KernelArg::WriteOnlyNoSize(dst), dst.rows, cols);
ocl::KernelArg::WriteOnly(dst, dcn, kercn));
size_t globalSize[2] = { cols, (dst.rows + 3) / 4 };
size_t globalSize[2] = { dst.cols * dcn / kercn, (dst.rows + 3) / 4 };
return k.run(2, globalSize, NULL, false);
}