generic-library/opencv
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge pull request #2801 from ilya-lavrenov:tapi_reduction

Browse Source
This commit is contained in:
Alexander Alekhin
2014-06-06 18:38:21 +04:00
committed by OpenCV Buildbot
parent 3f485d58bb c52a77b90f
commit f30301d171
3 changed files with 223 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
modules/core/src/stat.cpp
View File

@@ -473,8 +473,11 @@ static bool ocl_sum( InputArray _src, Scalar & res, int sum_op, InputArray _mask
{
CV_Assert(sum_op == OCL_OP_SUM || sum_op == OCL_OP_SUM_ABS || sum_op == OCL_OP_SUM_SQR);
int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0,
haveMask = _mask.kind() != _InputArray::NONE;
int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type),
kercn = cn == 1 && !haveMask ? ocl::predictOptimalVectorWidth(_src) : 1,
mcn = std::max(cn, kercn);
if ( (!doubleSupport && depth == CV_64F) || cn > 4 )
return false;
@@ -484,7 +487,6 @@ static bool ocl_sum( InputArray _src, Scalar & res, int sum_op, InputArray _mask
int ddepth = std::max(sum_op == OCL_OP_SUM_SQR ? CV_32F : CV_32S, depth),
dtype = CV_MAKE_TYPE(ddepth, cn);
bool haveMask = _mask.kind() != _InputArray::NONE;
CV_Assert(!haveMask || _mask.type() == CV_8UC1);
int wgs2_aligned = 1;
@@ -494,15 +496,19 @@ static bool ocl_sum( InputArray _src, Scalar & res, int sum_op, InputArray _mask
static const char * const opMap[3] = { "OP_SUM", "OP_SUM_ABS", "OP_SUM_SQR" };
char cvt[40];
ocl::Kernel k("reduce", ocl::core::reduce_oclsrc,
format("-D srcT=%s -D srcT1=%s -D dstT=%s -D dstT1=%s -D ddepth=%d -D cn=%d"
" -D convertToDT=%s -D %s -D WGS=%d -D WGS2_ALIGNED=%d%s%s",
ocl::typeToStr(type), ocl::typeToStr(depth),
ocl::typeToStr(dtype), ocl::typeToStr(ddepth), ddepth, cn,
ocl::convertTypeStr(depth, ddepth, cn, cvt),
String opts = format("-D srcT=%s -D srcT1=%s -D dstT=%s -D dstTK=%s -D dstT1=%s -D ddepth=%d -D cn=%d"
" -D convertToDT=%s -D %s -D WGS=%d -D WGS2_ALIGNED=%d%s%s%s%s -D kercn=%d",
ocl::typeToStr(CV_MAKE_TYPE(depth, mcn)), ocl::typeToStr(depth),
ocl::typeToStr(dtype), ocl::typeToStr(CV_MAKE_TYPE(ddepth, mcn)),
ocl::typeToStr(ddepth), ddepth, cn,
ocl::convertTypeStr(depth, ddepth, mcn, cvt),
opMap[sum_op], (int)wgs, wgs2_aligned,
doubleSupport ? " -D DOUBLE_SUPPORT" : "",
haveMask ? " -D HAVE_MASK" : ""));
haveMask ? " -D HAVE_MASK" : "",
_src.isContinuous() ? " -D HAVE_SRC_CONT" : "",
_mask.isContinuous() ? " -D HAVE_MASK_CONT" : "", kercn);
ocl::Kernel k("reduce", ocl::core::reduce_oclsrc, opts);
if (k.empty())
return false;
@@ -643,7 +649,7 @@ namespace cv {
static bool ocl_countNonZero( InputArray _src, int & res )
{
int type = _src.type(), depth = CV_MAT_DEPTH(type);
int type = _src.type(), depth = CV_MAT_DEPTH(type), kercn = ocl::predictOptimalVectorWidth(_src);
bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
if (depth == CV_64F && !doubleSupport)
@@ -658,9 +664,12 @@ static bool ocl_countNonZero( InputArray _src, int & res )
wgs2_aligned >>= 1;
ocl::Kernel k("reduce", ocl::core::reduce_oclsrc,
format("-D srcT=%s -D OP_COUNT_NON_ZERO -D WGS=%d -D WGS2_ALIGNED=%d%s",
ocl::typeToStr(type), (int)wgs,
wgs2_aligned, doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
format("-D srcT=%s -D srcT1=%s -D cn=1 -D OP_COUNT_NON_ZERO"
" -D WGS=%d -D kercn=%d -D WGS2_ALIGNED=%d%s%s",
ocl::typeToStr(CV_MAKE_TYPE(depth, kercn)),
ocl::typeToStr(depth), (int)wgs, kercn,
wgs2_aligned, doubleSupport ? " -D DOUBLE_SUPPORT" : "",
_src.isContinuous() ? " -D HAVE_SRC_CONT" : ""));
if (k.empty())
return false;
@@ -1349,7 +1358,7 @@ static bool ocl_minMaxIdx( InputArray _src, double* minVal, double* maxVal, int*
CV_Assert( (_src.channels() == 1 && (_mask.empty() || _mask.type() == CV_8U)) ||
(_src.channels() >= 1 && _mask.empty() && !minLoc && !maxLoc) );
int type = _src.type(), depth = CV_MAT_DEPTH(type);
int type = _src.type(), depth = CV_MAT_DEPTH(type), kercn = 1;
bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
if (depth == CV_64F && !doubleSupport)
@@ -1363,9 +1372,12 @@ static bool ocl_minMaxIdx( InputArray _src, double* minVal, double* maxVal, int*
wgs2_aligned <<= 1;
wgs2_aligned >>= 1;
String opts = format("-D DEPTH_%d -D srcT=%s -D OP_MIN_MAX_LOC%s -D WGS=%d -D WGS2_ALIGNED=%d%s",
String opts = format("-D DEPTH_%d -D srcT=%s -D OP_MIN_MAX_LOC%s -D WGS=%d"
" -D WGS2_ALIGNED=%d%s%s%s -D kercn=%d",
depth, ocl::typeToStr(depth), _mask.empty() ? "" : "_MASK", (int)wgs,
wgs2_aligned, doubleSupport ? " -D DOUBLE_SUPPORT" : "");
wgs2_aligned, doubleSupport ? " -D DOUBLE_SUPPORT" : "",
_src.isContinuous() ? " -D HAVE_SRC_CONT" : "",
_mask.isContinuous() ? " -D HAVE_MASK_CONT" : "", kercn);
ocl::Kernel k("reduce", ocl::core::reduce_oclsrc, opts);
if (k.empty())
@@ -2090,9 +2102,11 @@ static bool ocl_norm( InputArray _src, int normType, InputArray _mask, double &
ocl::Kernel k("reduce", ocl::core::reduce_oclsrc,
format("-D OP_NORM_INF_MASK -D HAVE_MASK -D DEPTH_%d"
" -D srcT=%s -D srcT1=%s -D WGS=%d -D cn=%d -D WGS2_ALIGNED=%d%s",
" -D srcT=%s -D srcT1=%s -D WGS=%d -D cn=%d -D WGS2_ALIGNED=%d%s%s%s",
depth, ocl::typeToStr(type), ocl::typeToStr(depth),
wgs, cn, wgs2_aligned, doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
wgs, cn, wgs2_aligned, doubleSupport ? " -D DOUBLE_SUPPORT" : "",
src.isContinuous() ? " -D HAVE_CONT_SRC" : "",
_mask.isContinuous() ? " -D HAVE_MASK_CONT" : ""));
if (k.empty())
return false;