Skip to content

O
opencv
Commit 6e0d7746 authored by Alexander Alekhin's avatar Alexander Alekhin
Browse files
Options

Merge pull request #2864 from ilya-lavrenov:tapi_boxfilter

parents 07446ec9 c424d360
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 106 additions and 37 deletions
modules/imgproc/src/opencl/boxFilterSmall.cl 0 → 100755
View file @ 6e0d7746
This diff is collapsed.
modules/imgproc/src/smooth.cpp
View file @ 6e0d7746
......@@ -629,12 +629,14 @@ struct ColumnSum<int, ushort> :
#ifdef HAVE_OPENCL
#define DIVUP(total, grain) ((total + grain - 1) / (grain))
#define ROUNDUP(sz, n) ((sz) + (n) - 1 - (((sz) + (n) - 1) % (n)))
static bool ocl_boxFilter( InputArray _src, OutputArray _dst, int ddepth,
Size ksize, Point anchor, int borderType, bool normalize, bool sqr = false )
{
const ocl::Device & dev = ocl::Device::getDefault();
int type = _src.type(), sdepth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type), esz = CV_ELEM_SIZE(type);
bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
bool doubleSupport = dev.doubleFPConfig() > 0;
if (ddepth < 0)
ddepth = sdepth;
......@@ -653,11 +655,12 @@ static bool ocl_boxFilter( InputArray _src, OutputArray _dst, int ddepth,
Size size = _src.size(), wholeSize;
bool isolated = (borderType & BORDER_ISOLATED) != 0;
borderType &= ~BORDER_ISOLATED;
int wdepth = std::max(CV_32F, std::max(ddepth, sdepth));
int wdepth = std::max(CV_32F, std::max(ddepth, sdepth)),
wtype = CV_MAKE_TYPE(wdepth, cn), dtype = CV_MAKE_TYPE(ddepth, cn);
const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", 0, "BORDER_REFLECT_101" };
size_t globalsize[2] = { size.width, size.height };
size_t localsize[2] = { 0, 1 };
size_t localsize_general[2] = { 0, 1 }, * localsize = NULL;
UMat src = _src.getUMat();
if (!isolated)
......@@ -674,46 +677,110 @@ static bool ocl_boxFilter( InputArray _src, OutputArray _dst, int ddepth,
int tryWorkItems = (int)maxWorkItemSizes[0];
ocl::Kernel kernel;
for ( ; ; )
{
int BLOCK_SIZE_X = tryWorkItems, BLOCK_SIZE_Y = std::min(ksize.height * 10, size.height);
while (BLOCK_SIZE_X > 32 && BLOCK_SIZE_X >= ksize.width * 2 && BLOCK_SIZE_X > size.width * 2)
BLOCK_SIZE_X /= 2;
while (BLOCK_SIZE_Y < BLOCK_SIZE_X / 8 && BLOCK_SIZE_Y * computeUnits * 32 < size.height)
BLOCK_SIZE_Y *= 2;
if (ksize.width > BLOCK_SIZE_X || w < ksize.width || h < ksize.height)
if (dev.isIntel() && !(dev.type() & ocl::Device::TYPE_CPU) &&
((ksize.width < 5 && ksize.height < 5 && esz <= 4) ||
(ksize.width == 5 && ksize.height == 5 && cn == 1)))
{
if (w < ksize.width || h < ksize.height)
return false;
char cvt[2][50];
String opts = format("-D LOCAL_SIZE_X=%d -D BLOCK_SIZE_Y=%d -D ST=%s -D DT=%s -D WT=%s -D convertToDT=%s -D convertToWT=%s"
" -D ANCHOR_X=%d -D ANCHOR_Y=%d -D KERNEL_SIZE_X=%d -D KERNEL_SIZE_Y=%d -D %s%s%s%s%s"
" -D ST1=%s -D DT1=%s -D cn=%d",
BLOCK_SIZE_X, BLOCK_SIZE_Y, ocl::typeToStr(type), ocl::typeToStr(CV_MAKE_TYPE(ddepth, cn)),
ocl::typeToStr(CV_MAKE_TYPE(wdepth, cn)),
ocl::convertTypeStr(wdepth, ddepth, cn, cvt[0]),
ocl::convertTypeStr(sdepth, wdepth, cn, cvt[1]),
anchor.x, anchor.y, ksize.width, ksize.height, borderMap[borderType],
isolated ? " -D BORDER_ISOLATED" : "", doubleSupport ? " -D DOUBLE_SUPPORT" : "",
normalize ? " -D NORMALIZE" : "", sqr ? " -D SQR" : "",
ocl::typeToStr(sdepth), ocl::typeToStr(ddepth), cn);
localsize[0] = BLOCK_SIZE_X;
globalsize[0] = DIVUP(size.width, BLOCK_SIZE_X - (ksize.width - 1)) * BLOCK_SIZE_X;
globalsize[1] = DIVUP(size.height, BLOCK_SIZE_Y);
kernel.create("boxFilter", cv::ocl::imgproc::boxFilter_oclsrc, opts);
if (kernel.empty())
return false;
// Figure out what vector size to use for loading the pixels.
int pxLoadNumPixels = cn != 1 || size.width % 4 ? 1 : 4;
int pxLoadVecSize = cn * pxLoadNumPixels;
size_t kernelWorkGroupSize = kernel.workGroupSize();
if (localsize[0] <= kernelWorkGroupSize)
break;
if (BLOCK_SIZE_X < (int)kernelWorkGroupSize)
// Figure out how many pixels per work item to compute in X and Y
// directions. Too many and we run out of registers.
int pxPerWorkItemX = 1, pxPerWorkItemY = 1;
if (cn <= 2 && ksize.width <= 4 && ksize.height <= 4)
{
pxPerWorkItemX = size.width % 8 ? size.width % 4 ? size.width % 2 ? 1 : 2 : 4 : 8;
pxPerWorkItemY = size.height % 2 ? 1 : 2;
}
else if (cn < 4 || (ksize.width <= 4 && ksize.height <= 4))
{
pxPerWorkItemX = size.width % 2 ? 1 : 2;
pxPerWorkItemY = size.height % 2 ? 1 : 2;
}
globalsize[0] = size.width / pxPerWorkItemX;
globalsize[1] = size.height / pxPerWorkItemY;
// Need some padding in the private array for pixels
int privDataWidth = ROUNDUP(pxPerWorkItemX + ksize.width - 1, pxLoadNumPixels);
// Make the global size a nice round number so the runtime can pick
// from reasonable choices for the workgroup size
const int wgRound = 256;
globalsize[0] = ROUNDUP(globalsize[0], wgRound);
char build_options[1024], cvt[2][40];
sprintf(build_options, "-D cn=%d "
"-D ANCHOR_X=%d -D ANCHOR_Y=%d -D KERNEL_SIZE_X=%d -D KERNEL_SIZE_Y=%d "
"-D PX_LOAD_VEC_SIZE=%d -D PX_LOAD_NUM_PX=%d "
"-D PX_PER_WI_X=%d -D PX_PER_WI_Y=%d -D PRIV_DATA_WIDTH=%d -D %s -D %s "
"-D PX_LOAD_X_ITERATIONS=%d -D PX_LOAD_Y_ITERATIONS=%d "
"-D srcT=%s -D srcT1=%s -D dstT=%s -D dstT1=%s -D WT=%s -D WT1=%s "
"-D convertToWT=%s -D convertToDstT=%s%s%s",
cn, anchor.x, anchor.y, ksize.width, ksize.height,
pxLoadVecSize, pxLoadNumPixels,
pxPerWorkItemX, pxPerWorkItemY, privDataWidth, borderMap[borderType],
isolated ? "BORDER_ISOLATED" : "NO_BORDER_ISOLATED",
privDataWidth / pxLoadNumPixels, pxPerWorkItemY + ksize.height - 1,
ocl::typeToStr(type), ocl::typeToStr(sdepth), ocl::typeToStr(dtype),
ocl::typeToStr(ddepth), ocl::typeToStr(wtype), ocl::typeToStr(wdepth),
ocl::convertTypeStr(sdepth, wdepth, cn, cvt[0]),
ocl::convertTypeStr(wdepth, ddepth, cn, cvt[1]),
normalize ? " -D NORMALIZE" : "", sqr ? " -D SQR" : "");
if (!kernel.create("boxFilterSmall", cv::ocl::imgproc::boxFilterSmall_oclsrc, build_options))
return false;
}
else
{
localsize = localsize_general;
for ( ; ; )
{
int BLOCK_SIZE_X = tryWorkItems, BLOCK_SIZE_Y = std::min(ksize.height * 10, size.height);
while (BLOCK_SIZE_X > 32 && BLOCK_SIZE_X >= ksize.width * 2 && BLOCK_SIZE_X > size.width * 2)
BLOCK_SIZE_X /= 2;
while (BLOCK_SIZE_Y < BLOCK_SIZE_X / 8 && BLOCK_SIZE_Y * computeUnits * 32 < size.height)
BLOCK_SIZE_Y *= 2;
if (ksize.width > BLOCK_SIZE_X || w < ksize.width || h < ksize.height)
return false;
char cvt[2][50];
String opts = format("-D LOCAL_SIZE_X=%d -D BLOCK_SIZE_Y=%d -D ST=%s -D DT=%s -D WT=%s -D convertToDT=%s -D convertToWT=%s"
" -D ANCHOR_X=%d -D ANCHOR_Y=%d -D KERNEL_SIZE_X=%d -D KERNEL_SIZE_Y=%d -D %s%s%s%s%s"
" -D ST1=%s -D DT1=%s -D cn=%d",
BLOCK_SIZE_X, BLOCK_SIZE_Y, ocl::typeToStr(type), ocl::typeToStr(CV_MAKE_TYPE(ddepth, cn)),
ocl::typeToStr(CV_MAKE_TYPE(wdepth, cn)),
ocl::convertTypeStr(wdepth, ddepth, cn, cvt[0]),
ocl::convertTypeStr(sdepth, wdepth, cn, cvt[1]),
anchor.x, anchor.y, ksize.width, ksize.height, borderMap[borderType],
isolated ? " -D BORDER_ISOLATED" : "", doubleSupport ? " -D DOUBLE_SUPPORT" : "",
normalize ? " -D NORMALIZE" : "", sqr ? " -D SQR" : "",
ocl::typeToStr(sdepth), ocl::typeToStr(ddepth), cn);
localsize[0] = BLOCK_SIZE_X;
globalsize[0] = DIVUP(size.width, BLOCK_SIZE_X - (ksize.width - 1)) * BLOCK_SIZE_X;
globalsize[1] = DIVUP(size.height, BLOCK_SIZE_Y);
kernel.create("boxFilter", cv::ocl::imgproc::boxFilter_oclsrc, opts);
if (kernel.empty())
return false;
size_t kernelWorkGroupSize = kernel.workGroupSize();
if (localsize[0] <= kernelWorkGroupSize)
break;
if (BLOCK_SIZE_X < (int)kernelWorkGroupSize)
return false;
tryWorkItems = (int)kernelWorkGroupSize;
tryWorkItems = (int)kernelWorkGroupSize;
}
}
_dst.create(size, CV_MAKETYPE(ddepth, cn));
......@@ -736,6 +803,8 @@ static bool ocl_boxFilter( InputArray _src, OutputArray _dst, int ddepth,
return kernel.run(2, globalsize, localsize, false);
}
#undef ROUNDUP
#endif
}
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment