Merge pull request #3144 from ElenaGvozdeva:ocl_morphSmall

This commit is contained in:
Vadim Pisarevsky 2014-08-22 12:14:06 +00:00
commit e7539bd2c8
3 changed files with 11 additions and 17 deletions

View File

@ -1434,7 +1434,7 @@ static bool ocl_morphSmall( InputArray _src, OutputArray _dst, InputArray _kerne
"-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=srcT -D dstT1=srcT1 -D WT=%s -D WT1=%s "
"-D convertToWT=%s -D convertToDstT=%s -D PROCESS_ELEM_=%s -D %s%s",
"-D convertToWT=%s -D convertToDstT=%s -D PX_LOAD_FLOAT_VEC_CONV=convert_%s -D PROCESS_ELEM_=%s -D %s%s",
cn, anchor.x, anchor.y, ksize.width, ksize.height,
pxLoadVecSize, pxLoadNumPixels, depth,
pxPerWorkItemX, pxPerWorkItemY, privDataWidth, borderMap[borderType],
@ -1445,6 +1445,7 @@ static bool ocl_morphSmall( InputArray _src, OutputArray _dst, InputArray _kerne
haveExtraMat ? ocl::typeToStr(wdepth):"srcT1",//to prevent overflow - WT1
haveExtraMat ? ocl::convertTypeStr(depth, wdepth, cn, cvt[0]) : "noconvert",//to prevent overflow - src to WT
haveExtraMat ? ocl::convertTypeStr(wdepth, depth, cn, cvt[1]) : "noconvert",//to prevent overflow - WT to dst
ocl::typeToStr(CV_MAKE_TYPE(haveExtraMat ? wdepth : depth, pxLoadVecSize)), //PX_LOAD_FLOAT_VEC_CONV
processing.c_str(), op2str[op],
actual_op == op ? "" : cv::format(" -D %s", op2str[actual_op]).c_str());
@ -1531,7 +1532,11 @@ static bool ocl_morphOp(InputArray _src, OutputArray _dst, InputArray _kernel,
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)) &&
(iterations == 1) && cn == 1)
(iterations == 1)
#if defined __APPLE__
&& cn == 1
#endif
)
{
if (ocl_morphSmall(_src, _dst, kernel, anchor, borderType, op, actual_op, _extraMat))
return true;

View File

@ -164,18 +164,6 @@ inline bool isBorder(const struct RectCoords bounds, int2 coord, int numPixels)
#define vload1(OFFSET, PTR) (*(PTR + OFFSET))
#define PX_LOAD_VEC_TYPE CAT(srcT1, PX_LOAD_VEC_SIZE)
#define PX_LOAD_FLOAT_VEC_TYPE CAT(WT1, PX_LOAD_VEC_SIZE)
#if PX_LOAD_VEC_SIZE == 1
#define PX_LOAD_FLOAT_VEC_CONV (float)
#elif PX_LOAD_VEC_SIZE == 2
#define PX_LOAD_FLOAT_VEC_CONV convert_float2
#elif PX_LOAD_VEC_SIZE == 3
#define PX_LOAD_FLOAT_VEC_CONV convert_float3
#elif PX_LOAD_VEC_SIZE == 4
#define PX_LOAD_FLOAT_VEC_CONV convert_float4
#endif
//#define PX_LOAD_FLOAT_VEC_CONV CAT(convert_, PX_LOAD_FLOAT_VEC_TYPE)
#define PX_LOAD CAT(vload, PX_LOAD_VEC_SIZE)

View File

@ -720,7 +720,7 @@ static bool ocl_boxFilter( InputArray _src, OutputArray _dst, int ddepth,
"-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 -D OP_BOX_FILTER",
"-D convertToWT=%s -D convertToDstT=%s%s%s -D PX_LOAD_FLOAT_VEC_CONV=convert_%s -D OP_BOX_FILTER",
cn, anchor.x, anchor.y, ksize.width, ksize.height,
pxLoadVecSize, pxLoadNumPixels,
pxPerWorkItemX, pxPerWorkItemY, privDataWidth, borderMap[borderType],
@ -730,8 +730,9 @@ static bool ocl_boxFilter( InputArray _src, OutputArray _dst, int ddepth,
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" : "");
normalize ? " -D NORMALIZE" : "", sqr ? " -D SQR" : "",
ocl::typeToStr(CV_MAKE_TYPE(wdepth, pxLoadVecSize)) //PX_LOAD_FLOAT_VEC_CONV
);
if (!kernel.create("filterSmall", cv::ocl::imgproc::filterSmall_oclsrc, build_options))