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

optimized cv::norm with 2 args

Browse Source
This commit is contained in:
Ilya Lavrenov
2014-06-07 15:51:41 +04:00
parent 1a7a262f74
commit 2040995801
3 changed files with 387 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files

148
modules/core/src/stat.cpp
View File

@@ -469,21 +469,25 @@ template <typename T> Scalar ocl_part_sum(Mat m)
enum { OCL_OP_SUM = 0, OCL_OP_SUM_ABS = 1, OCL_OP_SUM_SQR = 2 };
static bool ocl_sum( InputArray _src, Scalar & res, int sum_op, InputArray _mask = noArray() )
static bool ocl_sum( InputArray _src, Scalar & res, int sum_op, InputArray _mask = noArray(),
InputArray _src2 = noArray(), bool calc2 = false, const Scalar & res2 = Scalar() )
{
CV_Assert(sum_op == OCL_OP_SUM || sum_op == OCL_OP_SUM_ABS || sum_op == OCL_OP_SUM_SQR);
bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0,
haveMask = _mask.kind() != _InputArray::NONE;
const ocl::Device & dev = ocl::Device::getDefault();
bool doubleSupport = dev.doubleFPConfig() > 0,
haveMask = _mask.kind() != _InputArray::NONE,
haveSrc2 = _src2.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);
CV_Assert(!haveSrc2 || _src2.type() == type);
if ( (!doubleSupport && depth == CV_64F) || cn > 4 )
return false;
int dbsize = ocl::Device::getDefault().maxComputeUnits();
size_t wgs = ocl::Device::getDefault().maxWorkGroupSize();
int ngroups = dev.maxComputeUnits(), dbsize = ngroups * (calc2 ? 2 : 1);
size_t wgs = dev.maxWorkGroupSize();
int ddepth = std::max(sum_op == OCL_OP_SUM_SQR ? CV_32F : CV_32S, depth),
dtype = CV_MAKE_TYPE(ddepth, cn);
@@ -497,7 +501,7 @@ 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];
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",
" -D convertToDT=%s -D %s -D WGS=%d -D WGS2_ALIGNED=%d%s%s%s%s -D kercn=%d%s%s%s",
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,
@@ -506,30 +510,49 @@ static bool ocl_sum( InputArray _src, Scalar & res, int sum_op, InputArray _mask
doubleSupport ? " -D DOUBLE_SUPPORT" : "",
haveMask ? " -D HAVE_MASK" : "",
_src.isContinuous() ? " -D HAVE_SRC_CONT" : "",
_mask.isContinuous() ? " -D HAVE_MASK_CONT" : "", kercn);
haveMask && _mask.isContinuous() ? " -D HAVE_MASK_CONT" : "", kercn,
haveSrc2 ? " -D HAVE_SRC2" : "", calc2 ? " -D OP_CALC2" : "",
haveSrc2 && _src2.isContinuous() ? " -D HAVE_SRC2_CONT" : "");
ocl::Kernel k("reduce", ocl::core::reduce_oclsrc, opts);
if (k.empty())
return false;
UMat src = _src.getUMat(), db(1, dbsize, dtype), mask = _mask.getUMat();
UMat src = _src.getUMat(), src2 = _src2.getUMat(),
db(1, dbsize, dtype), mask = _mask.getUMat();
ocl::KernelArg srcarg = ocl::KernelArg::ReadOnlyNoSize(src),
dbarg = ocl::KernelArg::PtrWriteOnly(db),
maskarg = ocl::KernelArg::ReadOnlyNoSize(mask);
maskarg = ocl::KernelArg::ReadOnlyNoSize(mask),
src2arg = ocl::KernelArg::ReadOnlyNoSize(src2);
if (haveMask)
k.args(srcarg, src.cols, (int)src.total(), dbsize, dbarg, maskarg);
{
if (haveSrc2)
k.args(srcarg, src.cols, (int)src.total(), ngroups, dbarg, maskarg, src2arg);
else
k.args(srcarg, src.cols, (int)src.total(), ngroups, dbarg, maskarg);
}
else
k.args(srcarg, src.cols, (int)src.total(), dbsize, dbarg);
{
if (haveSrc2)
k.args(srcarg, src.cols, (int)src.total(), ngroups, dbarg, src2arg);
else
k.args(srcarg, src.cols, (int)src.total(), ngroups, dbarg);
}
size_t globalsize = dbsize * wgs;
size_t globalsize = ngroups * wgs;
if (k.run(1, &globalsize, &wgs, false))
{
typedef Scalar (*part_sum)(Mat m);
part_sum funcs[3] = { ocl_part_sum<int>, ocl_part_sum<float>, ocl_part_sum<double> },
func = funcs[ddepth - CV_32S];
res = func(db.getMat(ACCESS_READ));
Mat mres = db.getMat(ACCESS_READ);
if (calc2)
const_cast<Scalar &>(res2) = func(mres.colRange(dbsize, dbsize));
res = func(mres.colRange(0, dbsize));
return true;
}
return false;
@@ -1396,18 +1419,21 @@ typedef void (*getMinMaxResFunc)(const Mat & db, double *minVal, double *maxVal,
int *minLoc, int *maxLoc, int gropunum, int cols);
static bool ocl_minMaxIdx( InputArray _src, double* minVal, double* maxVal, int* minLoc, int* maxLoc, InputArray _mask,
int ddepth = -1, bool absValues = false)
int ddepth = -1, bool absValues = false, InputArray _src2 = noArray(), bool calc2 = false)
{
CV_Assert( (_src.channels() == 1 && (_mask.empty() || _mask.type() == CV_8U)) ||
(_src.channels() >= 1 && _mask.empty() && !minLoc && !maxLoc) );
const ocl::Device & dev = ocl::Device::getDefault();
bool doubleSupport = dev.doubleFPConfig() > 0, haveMask = !_mask.empty();
bool doubleSupport = dev.doubleFPConfig() > 0, haveMask = !_mask.empty(),
haveSrc2 = _src2.kind() != _InputArray::NONE;
int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type),
kercn = haveMask ? 1 : std::min(4, ocl::predictOptimalVectorWidth(_src));
if (ddepth < 0)
ddepth = depth;
CV_Assert(!haveSrc2 || _src2.type() == type);
if ((depth == CV_64F || ddepth == CV_64F) && !doubleSupport)
return false;
@@ -1435,7 +1461,7 @@ static bool ocl_minMaxIdx( InputArray _src, double* minVal, double* maxVal, int*
char cvt[40];
String opts = format("-D DEPTH_%d -D srcT1=%s%s -D WGS=%d -D srcT=%s"
" -D WGS2_ALIGNED=%d%s%s%s -D kercn=%d%s%s%s%s"
" -D dstT1=%s -D dstT=%s -D convertToDT=%s%s",
" -D dstT1=%s -D dstT=%s -D convertToDT=%s%s%s%s%s",
depth, ocl::typeToStr(depth), haveMask ? " -D HAVE_MASK" : "", (int)wgs,
ocl::typeToStr(CV_MAKE_TYPE(depth, kercn)), wgs2_aligned,
doubleSupport ? " -D DOUBLE_SUPPORT" : "",
@@ -1444,7 +1470,9 @@ static bool ocl_minMaxIdx( InputArray _src, double* minVal, double* maxVal, int*
needMinVal ? " -D NEED_MINVAL" : "", needMaxVal ? " -D NEED_MAXVAL" : "",
needMinLoc ? " -D NEED_MINLOC" : "", needMaxLoc ? " -D NEED_MAXLOC" : "",
ocl::typeToStr(ddepth), ocl::typeToStr(CV_MAKE_TYPE(ddepth, kercn)),
ocl::convertTypeStr(depth, ddepth, kercn, cvt), absValues ? " -D OP_ABS" : "");
ocl::convertTypeStr(depth, ddepth, kercn, cvt), absValues ? " -D OP_ABS" : "",
haveSrc2 ? " -D HAVE_SRC2" : "", calc2 ? " -D OP_CALC2" : "",
haveSrc2 && _src2.isContinuous() ? " -D HAVE_SRC2_CONT" : "");
ocl::Kernel k("minmaxloc", ocl::core::minmaxloc_oclsrc, opts);
if (k.empty())
@@ -1452,18 +1480,35 @@ static bool ocl_minMaxIdx( InputArray _src, double* minVal, double* maxVal, int*
int esz = CV_ELEM_SIZE(ddepth), esz32s = CV_ELEM_SIZE1(CV_32S),
dbsize = groupnum * ((needMinVal ? esz : 0) + (needMaxVal ? esz : 0) +
(needMinLoc ? esz32s : 0) + (needMaxLoc ? esz32s : 0));
UMat src = _src.getUMat(), db(1, dbsize, CV_8UC1), mask = _mask.getUMat();
(needMinLoc ? esz32s : 0) + (needMaxLoc ? esz32s : 0) +
(calc2 ? esz : 0));
UMat src = _src.getUMat(), src2 = _src2.getUMat(), db(1, dbsize, CV_8UC1), mask = _mask.getUMat();
if (cn > 1)
{
src = src.reshape(1);
src2 = src2.reshape(1);
}
if (!haveMask)
k.args(ocl::KernelArg::ReadOnlyNoSize(src), src.cols, (int)src.total(),
groupnum, ocl::KernelArg::PtrWriteOnly(db));
if (haveSrc2)
{
if (!haveMask)
k.args(ocl::KernelArg::ReadOnlyNoSize(src), src.cols, (int)src.total(),
groupnum, ocl::KernelArg::PtrWriteOnly(db), ocl::KernelArg::ReadOnlyNoSize(src2));
else
k.args(ocl::KernelArg::ReadOnlyNoSize(src), src.cols, (int)src.total(),
groupnum, ocl::KernelArg::PtrWriteOnly(db), ocl::KernelArg::ReadOnlyNoSize(mask),
ocl::KernelArg::ReadOnlyNoSize(src2));
}
else
k.args(ocl::KernelArg::ReadOnlyNoSize(src), src.cols, (int)src.total(),
groupnum, ocl::KernelArg::PtrWriteOnly(db), ocl::KernelArg::ReadOnlyNoSize(mask));
{
if (!haveMask)
k.args(ocl::KernelArg::ReadOnlyNoSize(src), src.cols, (int)src.total(),
groupnum, ocl::KernelArg::PtrWriteOnly(db));
else
k.args(ocl::KernelArg::ReadOnlyNoSize(src), src.cols, (int)src.total(),
groupnum, ocl::KernelArg::PtrWriteOnly(db), ocl::KernelArg::ReadOnlyNoSize(mask));
}
size_t globalsize = groupnum * wgs;
if (!k.run(1, &globalsize, &wgs, false))
@@ -2498,38 +2543,45 @@ namespace cv {
static bool ocl_norm( InputArray _src1, InputArray _src2, int normType, InputArray _mask, double & result )
{
const ocl::Device & d = ocl::Device::getDefault();
int type = _src1.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type), rowsPerWI = d.isIntel() ? 4 : 1;
bool doubleSupport = d.doubleFPConfig() > 0;
bool relative = (normType & NORM_RELATIVE) != 0;
Scalar sc1, sc2;
int type = _src1.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
bool relative = (normType & NORM_RELATIVE) != 0,
normsum = normType == NORM_L1 || normType == NORM_L2 || normType == NORM_L2SQR;
normType &= ~NORM_RELATIVE;
if ( !(normType == NORM_INF || normType == NORM_L1 || normType == NORM_L2 || normType == NORM_L2SQR) ||
(!doubleSupport && depth == CV_64F))
if ( !(normType == NORM_INF || normsum) )
return false;
int wdepth = std::max(CV_32S, depth);
char cvt[50];
ocl::Kernel k("KF", ocl::core::arithm_oclsrc,
format("-D BINARY_OP -D OP_ABSDIFF -D dstT=%s -D workT=dstT -D srcT1=%s -D srcT2=srcT1"
" -D convertToDT=%s -D convertToWT1=convertToDT -D convertToWT2=convertToDT -D rowsPerWI=%d%s",
ocl::typeToStr(wdepth), ocl::typeToStr(depth),
ocl::convertTypeStr(depth, wdepth, 1, cvt), rowsPerWI,
doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
if (k.empty())
return false;
if (normsum)
{
if (!ocl_sum(_src1, sc1, normType == NORM_L2 || normType == NORM_L2SQR ?
OCL_OP_SUM_SQR : OCL_OP_SUM, _mask, _src2, relative, sc2))
return false;
}
else
{
if (!ocl_minMaxIdx(_src1, NULL, &result, NULL, NULL, _mask, std::max(CV_32S, depth),
false, _src2, relative))
return false;
}
UMat src1 = _src1.getUMat(), src2 = _src2.getUMat(), diff(src1.size(), CV_MAKE_TYPE(wdepth, cn));
k.args(ocl::KernelArg::ReadOnlyNoSize(src1), ocl::KernelArg::ReadOnlyNoSize(src2),
ocl::KernelArg::WriteOnly(diff, cn));
double s2 = 0;
for (int i = 0; i < cn; ++i)
{
result += sc1[i];
if (relative)
s2 += sc2[i];
}
size_t globalsize[2] = { diff.cols * cn, (diff.rows + rowsPerWI - 1) / rowsPerWI };
if (!k.run(2, globalsize, NULL, false))
return false;
if (normType == NORM_L2)
{
result = std::sqrt(result);
if (relative)
s2 = std::sqrt(s2);
}
result = cv::norm(diff, normType, _mask);
if (relative)
result /= cv::norm(src2, normType, _mask) + DBL_EPSILON;
result /= (s2 + DBL_EPSILON);
return true;
}