Optimized memory access by using stride pattern

This commit is contained in:
Alexander Karsakov
2014-05-16 10:22:03 +04:00
parent 4612b4b827
commit 8ecb8c6a3d
2 changed files with 32 additions and 17 deletions

View File

@@ -833,9 +833,12 @@ static bool ocl_threshold( InputArray _src, OutputArray _dst, double & thresh, d
const char * const thresholdMap[] = { "THRESH_BINARY", "THRESH_BINARY_INV", "THRESH_TRUNC",
"THRESH_TOZERO", "THRESH_TOZERO_INV" };
ocl::Device dev = ocl::Device::getDefault();
int stride_size = dev.isIntel() && (dev.type() & ocl::Device::TYPE_GPU) ? 4 : 1;
ocl::Kernel k("threshold", ocl::imgproc::threshold_oclsrc,
format("-D %s -D T=%s -D T1=%s%s", thresholdMap[thresh_type],
ocl::typeToStr(ktype), ocl::typeToStr(depth),
format("-D %s -D T=%s -D T1=%s -D STRIDE_SIZE=%d%s", thresholdMap[thresh_type],
ocl::typeToStr(ktype), ocl::typeToStr(depth), stride_size,
doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
if (k.empty())
return false;
@@ -856,6 +859,7 @@ static bool ocl_threshold( InputArray _src, OutputArray _dst, double & thresh, d
ocl::KernelArg::Constant(Mat(1, 1, depth, Scalar::all(min_val))));
size_t globalsize[2] = { dst.cols * cn / kercn, dst.rows };
globalsize[1] = (globalsize[1] + stride_size - 1) / stride_size;
return k.run(2, globalsize, NULL, false);
}