Chiebot-Mirror
/
opencv
8
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

matrix reduction

Browse Source
pull/258/merge
Vladislav Vinogradov 13 years ago
parent fbf3de43a2
commit e8f9762ef3
3 changed files with 1221 additions and 2169 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 2738
      modules/gpu/src/cuda/matrix_reductions.cu
  2. 638
      modules/gpu/src/matrix_reductions.cpp
  3. 14
      modules/gpu/test/test_core.cpp

2738
modules/gpu/src/cuda/matrix_reductions.cu
View File

File diff suppressed because it is too large Load Diff

638
modules/gpu/src/matrix_reductions.cpp
Unescape View File

@ -204,34 +204,19 @@ double cv::gpu::norm(const GpuMat& src1, const GpuMat& src2, int normType)
////////////////////////////////////////////////////////////////////////
// Sum
namespace cv { namespace gpu { namespace device
namespace sum
{
namespace matrix_reductions
{
namespace sum
{
template <typename T>
void sumCaller(const PtrStepSzb src, PtrStepb buf, double* sum, int cn);
template <typename T>
void sumMultipassCaller(const PtrStepSzb src, PtrStepb buf, double* sum, int cn);
template <typename T>
void absSumCaller(const PtrStepSzb src, PtrStepb buf, double* sum, int cn);
void getBufSize(int cols, int rows, int cn, int& bufcols, int& bufrows);
template <typename T>
void absSumMultipassCaller(const PtrStepSzb src, PtrStepb buf, double* sum, int cn);
template <typename T, int cn>
void run(PtrStepSzb src, void* buf, double* sum);
template <typename T>
void sqrSumCaller(const PtrStepSzb src, PtrStepb buf, double* sum, int cn);
template <typename T, int cn>
void runAbs(PtrStepSzb src, void* buf, double* sum);
template <typename T>
void sqrSumMultipassCaller(const PtrStepSzb src, PtrStepb buf, double* sum, int cn);
void getBufSizeRequired(int cols, int rows, int cn, int& bufcols, int& bufrows);
}
}
}}}
template <typename T, int cn>
void runSqr(PtrStepSzb src, void* buf, double* sum);
}
Scalar cv::gpu::sum(const GpuMat& src)
{
@ -239,159 +224,115 @@ Scalar cv::gpu::sum(const GpuMat& src)
return sum(src, buf);
}
Scalar cv::gpu::sum(const GpuMat& src, GpuMat& buf)
{
using namespace cv::gpu::device::matrix_reductions::sum;
typedef void (*Caller)(const PtrStepSzb, PtrStepb, double*, int);
static Caller multipass_callers[] =
typedef void (*func_t)(PtrStepSzb src, void* buf, double* sum);
static const func_t funcs[7][5] =
{
sumMultipassCaller<unsigned char>, sumMultipassCaller<char>,
sumMultipassCaller<unsigned short>, sumMultipassCaller<short>,
sumMultipassCaller<int>, sumMultipassCaller<float>
{0, ::sum::run<uchar , 1>, ::sum::run<uchar , 2>, ::sum::run<uchar , 3>, ::sum::run<uchar , 4>},
{0, ::sum::run<schar , 1>, ::sum::run<schar , 2>, ::sum::run<schar , 3>, ::sum::run<schar , 4>},
{0, ::sum::run<ushort, 1>, ::sum::run<ushort, 2>, ::sum::run<ushort, 3>, ::sum::run<ushort, 4>},
{0, ::sum::run<short , 1>, ::sum::run<short , 2>, ::sum::run<short , 3>, ::sum::run<short , 4>},
{0, ::sum::run<int , 1>, ::sum::run<int , 2>, ::sum::run<int , 3>, ::sum::run<int , 4>},
{0, ::sum::run<float , 1>, ::sum::run<float , 2>, ::sum::run<float , 3>, ::sum::run<float , 4>},
{0, ::sum::run<double, 1>, ::sum::run<double, 2>, ::sum::run<double, 3>, ::sum::run<double, 4>}
};
static Caller singlepass_callers[] = {
sumCaller<unsigned char>, sumCaller<char>,
sumCaller<unsigned short>, sumCaller<short>,
sumCaller<int>, sumCaller<float>
};
CV_Assert(src.depth() <= CV_32F);
if (src.depth() == CV_64F)
{
if (!TargetArchs::builtWith(NATIVE_DOUBLE) || !DeviceInfo().supports(NATIVE_DOUBLE))
CV_Error(CV_StsUnsupportedFormat, "The device doesn't support double");
}
Size buf_size;
getBufSizeRequired(src.cols, src.rows, src.channels(), buf_size.width, buf_size.height);
::sum::getBufSize(src.cols, src.rows, src.channels(), buf_size.width, buf_size.height);
ensureSizeIsEnough(buf_size, CV_8U, buf);
buf.setTo(Scalar::all(0));
Caller* callers = multipass_callers;
if (TargetArchs::builtWith(GLOBAL_ATOMICS) && DeviceInfo().supports(GLOBAL_ATOMICS))
callers = singlepass_callers;
Caller caller = callers[src.depth()];
const func_t func = funcs[src.depth()][src.channels()];
double result[4];
caller(src, buf, result, src.channels());
func(src, buf.data, result);
return Scalar(result[0], result[1], result[2], result[3]);
}
Scalar cv::gpu::absSum(const GpuMat& src)
{
GpuMat buf;
return absSum(src, buf);
}
Scalar cv::gpu::absSum(const GpuMat& src, GpuMat& buf)
{
using namespace cv::gpu::device::matrix_reductions::sum;
typedef void (*Caller)(const PtrStepSzb, PtrStepb, double*, int);
static Caller multipass_callers[] =
{
absSumMultipassCaller<unsigned char>, absSumMultipassCaller<char>,
absSumMultipassCaller<unsigned short>, absSumMultipassCaller<short>,
absSumMultipassCaller<int>, absSumMultipassCaller<float>
};
static Caller singlepass_callers[] =
typedef void (*func_t)(PtrStepSzb src, void* buf, double* sum);
static const func_t funcs[7][5] =
{
absSumCaller<unsigned char>, absSumCaller<char>,
absSumCaller<unsigned short>, absSumCaller<short>,
absSumCaller<int>, absSumCaller<float>
{0, ::sum::runAbs<uchar , 1>, ::sum::runAbs<uchar , 2>, ::sum::runAbs<uchar , 3>, ::sum::runAbs<uchar , 4>},
{0, ::sum::runAbs<schar , 1>, ::sum::runAbs<schar , 2>, ::sum::runAbs<schar , 3>, ::sum::runAbs<schar , 4>},
{0, ::sum::runAbs<ushort, 1>, ::sum::runAbs<ushort, 2>, ::sum::runAbs<ushort, 3>, ::sum::runAbs<ushort, 4>},
{0, ::sum::runAbs<short , 1>, ::sum::runAbs<short , 2>, ::sum::runAbs<short , 3>, ::sum::runAbs<short , 4>},
{0, ::sum::runAbs<int , 1>, ::sum::runAbs<int , 2>, ::sum::runAbs<int , 3>, ::sum::runAbs<int , 4>},
{0, ::sum::runAbs<float , 1>, ::sum::runAbs<float , 2>, ::sum::runAbs<float , 3>, ::sum::runAbs<float , 4>},
{0, ::sum::runAbs<double, 1>, ::sum::runAbs<double, 2>, ::sum::runAbs<double, 3>, ::sum::runAbs<double, 4>}
};
CV_Assert(src.depth() <= CV_32F);
Size buf_size;
getBufSizeRequired(src.cols, src.rows, src.channels(), buf_size.width, buf_size.height);
::sum::getBufSize(src.cols, src.rows, src.channels(), buf_size.width, buf_size.height);
ensureSizeIsEnough(buf_size, CV_8U, buf);
buf.setTo(Scalar::all(0));
Caller* callers = multipass_callers;
if (TargetArchs::builtWith(GLOBAL_ATOMICS) && DeviceInfo().supports(GLOBAL_ATOMICS))
callers = singlepass_callers;
Caller caller = callers[src.depth()];
const func_t func = funcs[src.depth()][src.channels()];
double result[4];
caller(src, buf, result, src.channels());
func(src, buf.data, result);
return Scalar(result[0], result[1], result[2], result[3]);
}
Scalar cv::gpu::sqrSum(const GpuMat& src)
{
GpuMat buf;
return sqrSum(src, buf);
}
Scalar cv::gpu::sqrSum(const GpuMat& src, GpuMat& buf)
{
using namespace cv::gpu::device::matrix_reductions::sum;
typedef void (*Caller)(const PtrStepSzb, PtrStepb, double*, int);
static Caller multipass_callers[] =
typedef void (*func_t)(PtrStepSzb src, void* buf, double* sum);
static const func_t funcs[7][5] =
{
sqrSumMultipassCaller<unsigned char>, sqrSumMultipassCaller<char>,
sqrSumMultipassCaller<unsigned short>, sqrSumMultipassCaller<short>,
sqrSumMultipassCaller<int>, sqrSumMultipassCaller<float>
{0, ::sum::runSqr<uchar , 1>, ::sum::runSqr<uchar , 2>, ::sum::runSqr<uchar , 3>, ::sum::runSqr<uchar , 4>},
{0, ::sum::runSqr<schar , 1>, ::sum::runSqr<schar , 2>, ::sum::runSqr<schar , 3>, ::sum::runSqr<schar , 4>},
{0, ::sum::runSqr<ushort, 1>, ::sum::runSqr<ushort, 2>, ::sum::runSqr<ushort, 3>, ::sum::runSqr<ushort, 4>},
{0, ::sum::runSqr<short , 1>, ::sum::runSqr<short , 2>, ::sum::runSqr<short , 3>, ::sum::runSqr<short , 4>},
{0, ::sum::runSqr<int , 1>, ::sum::runSqr<int , 2>, ::sum::runSqr<int , 3>, ::sum::runSqr<int , 4>},
{0, ::sum::runSqr<float , 1>, ::sum::runSqr<float , 2>, ::sum::runSqr<float , 3>, ::sum::runSqr<float , 4>},
{0, ::sum::runSqr<double, 1>, ::sum::runSqr<double, 2>, ::sum::runSqr<double, 3>, ::sum::runSqr<double, 4>}
};
static Caller singlepass_callers[7] =
{
sqrSumCaller<unsigned char>, sqrSumCaller<char>,
sqrSumCaller<unsigned short>, sqrSumCaller<short>,
sqrSumCaller<int>, sqrSumCaller<float>
};
CV_Assert(src.depth() <= CV_32F);
Caller* callers = multipass_callers;
if (TargetArchs::builtWith(GLOBAL_ATOMICS) && DeviceInfo().supports(GLOBAL_ATOMICS))
callers = singlepass_callers;
Size buf_size;
getBufSizeRequired(src.cols, src.rows, src.channels(), buf_size.width, buf_size.height);
::sum::getBufSize(src.cols, src.rows, src.channels(), buf_size.width, buf_size.height);
ensureSizeIsEnough(buf_size, CV_8U, buf);
buf.setTo(Scalar::all(0));
Caller caller = callers[src.depth()];
const func_t func = funcs[src.depth()][src.channels()];
double result[4];
caller(src, buf, result, src.channels());
func(src, buf.data, result);
return Scalar(result[0], result[1], result[2], result[3]);
}
////////////////////////////////////////////////////////////////////////
// Find min or max
// minMax
namespace cv { namespace gpu { namespace device
namespace minMax
{
namespace matrix_reductions
{
namespace minmax
{
void getBufSizeRequired(int cols, int rows, int elem_size, int& bufcols, int& bufrows);
template <typename T>
void minMaxCaller(const PtrStepSzb src, double* minval, double* maxval, PtrStepb buf);
template <typename T>
void minMaxMaskCaller(const PtrStepSzb src, const PtrStepb mask, double* minval, double* maxval, PtrStepb buf);
template <typename T>
void minMaxMultipassCaller(const PtrStepSzb src, double* minval, double* maxval, PtrStepb buf);
template <typename T>
void minMaxMaskMultipassCaller(const PtrStepSzb src, const PtrStepb mask, double* minval, double* maxval, PtrStepb buf);
}
}
}}}
void getBufSize(int cols, int rows, int& bufcols, int& bufrows);
template <typename T>
void run(const PtrStepSzb src, const PtrStepb mask, double* minval, double* maxval, PtrStepb buf);
}
void cv::gpu::minMax(const GpuMat& src, double* minVal, double* maxVal, const GpuMat& mask)
{
@ -399,45 +340,22 @@ void cv::gpu::minMax(const GpuMat& src, double* minVal, double* maxVal, const Gp
minMax(src, minVal, maxVal, mask, buf);
}
void cv::gpu::minMax(const GpuMat& src, double* minVal, double* maxVal, const GpuMat& mask, GpuMat& buf)
{
using namespace ::cv::gpu::device::matrix_reductions::minmax;
typedef void (*Caller)(const PtrStepSzb, double*, double*, PtrStepb);
typedef void (*MaskedCaller)(const PtrStepSzb, const PtrStepb, double*, double*, PtrStepb);
static Caller multipass_callers[] =
{
minMaxMultipassCaller<unsigned char>, minMaxMultipassCaller<char>,
minMaxMultipassCaller<unsigned short>, minMaxMultipassCaller<short>,
minMaxMultipassCaller<int>, minMaxMultipassCaller<float>, 0
};
static Caller singlepass_callers[] =
{
minMaxCaller<unsigned char>, minMaxCaller<char>,
minMaxCaller<unsigned short>, minMaxCaller<short>,
minMaxCaller<int>, minMaxCaller<float>, minMaxCaller<double>
};
static MaskedCaller masked_multipass_callers[] =
{
minMaxMaskMultipassCaller<unsigned char>, minMaxMaskMultipassCaller<char>,
minMaxMaskMultipassCaller<unsigned short>, minMaxMaskMultipassCaller<short>,
minMaxMaskMultipassCaller<int>, minMaxMaskMultipassCaller<float>, 0
};
static MaskedCaller masked_singlepass_callers[] =
typedef void (*func_t)(const PtrStepSzb src, const PtrStepb mask, double* minval, double* maxval, PtrStepb buf);
static const func_t funcs[] =
{
minMaxMaskCaller<unsigned char>, minMaxMaskCaller<char>,
minMaxMaskCaller<unsigned short>, minMaxMaskCaller<short>,
minMaxMaskCaller<int>, minMaxMaskCaller<float>, minMaxMaskCaller<double>
::minMax::run<uchar>,
::minMax::run<schar>,
::minMax::run<ushort>,
::minMax::run<short>,
::minMax::run<int>,
::minMax::run<float>,
::minMax::run<double>
};
CV_Assert(src.depth() <= CV_64F);
CV_Assert(src.channels() == 1);
CV_Assert(mask.empty() || (mask.type() == CV_8U && src.size() == mask.size()));
CV_Assert( src.channels() == 1 );
CV_Assert( mask.empty() || (mask.size() == src.size() && mask.type() == CV_8U) );
if (src.depth() == CV_64F)
{
@ -445,66 +363,26 @@ void cv::gpu::minMax(const GpuMat& src, double* minVal, double* maxVal, const Gp
CV_Error(CV_StsUnsupportedFormat, "The device doesn't support double");
}
double minVal_; if (!minVal) minVal = &minVal_;
double maxVal_; if (!maxVal) maxVal = &maxVal_;
Size buf_size;
getBufSizeRequired(src.cols, src.rows, static_cast<int>(src.elemSize()), buf_size.width, buf_size.height);
::minMax::getBufSize(src.cols, src.rows, buf_size.width, buf_size.height);
ensureSizeIsEnough(buf_size, CV_8U, buf);
if (mask.empty())
{
Caller* callers = multipass_callers;
if (TargetArchs::builtWith(GLOBAL_ATOMICS) && DeviceInfo().supports(GLOBAL_ATOMICS))
callers = singlepass_callers;
Caller caller = callers[src.type()];
CV_Assert(caller != 0);
caller(src, minVal, maxVal, buf);
}
else
{
MaskedCaller* callers = masked_multipass_callers;
if (TargetArchs::builtWith(GLOBAL_ATOMICS) && DeviceInfo().supports(GLOBAL_ATOMICS))
callers = masked_singlepass_callers;
const func_t func = funcs[src.depth()];
MaskedCaller caller = callers[src.type()];
CV_Assert(caller != 0);
caller(src, mask, minVal, maxVal, buf);
}
double temp1, temp2;
func(src, mask, minVal ? minVal : &temp1, maxVal ? maxVal : &temp2, buf);
}
////////////////////////////////////////////////////////////////////////
// Locate min and max
// minMaxLoc
namespace cv { namespace gpu { namespace device
namespace minMaxLoc
{
namespace matrix_reductions
{
namespace minmaxloc
{
void getBufSizeRequired(int cols, int rows, int elem_size, int& b1cols,
int& b1rows, int& b2cols, int& b2rows);
template <typename T>
void minMaxLocCaller(const PtrStepSzb src, double* minval, double* maxval,
int minloc[2], int maxloc[2], PtrStepb valBuf, PtrStepb locBuf);
template <typename T>
void minMaxLocMaskCaller(const PtrStepSzb src, const PtrStepb mask, double* minval, double* maxval,
int minloc[2], int maxloc[2], PtrStepb valBuf, PtrStepb locBuf);
template <typename T>
void minMaxLocMultipassCaller(const PtrStepSzb src, double* minval, double* maxval,
int minloc[2], int maxloc[2], PtrStepb valBuf, PtrStepb locBuf);
void getBufSize(int cols, int rows, size_t elem_size, int& b1cols, int& b1rows, int& b2cols, int& b2rows);
template <typename T>
void minMaxLocMaskMultipassCaller(const PtrStepSzb src, const PtrStepb mask, double* minval, double* maxval,
int minloc[2], int maxloc[2], PtrStepb valBuf, PtrStepb locBuf);
}
}
}}}
template <typename T>
void run(const PtrStepSzb src, const PtrStepb mask, double* minval, double* maxval, int* minloc, int* maxloc, PtrStepb valbuf, PtrStep<unsigned int> locbuf);
}
void cv::gpu::minMaxLoc(const GpuMat& src, double* minVal, double* maxVal, Point* minLoc, Point* maxLoc, const GpuMat& mask)
{
@ -515,42 +393,20 @@ void cv::gpu::minMaxLoc(const GpuMat& src, double* minVal, double* maxVal, Point
void cv::gpu::minMaxLoc(const GpuMat& src, double* minVal, double* maxVal, Point* minLoc, Point* maxLoc,
const GpuMat& mask, GpuMat& valBuf, GpuMat& locBuf)
{
using namespace ::cv::gpu::device::matrix_reductions::minmaxloc;
typedef void (*Caller)(const PtrStepSzb, double*, double*, int[2], int[2], PtrStepb, PtrStepb);
typedef void (*MaskedCaller)(const PtrStepSzb, const PtrStepb, double*, double*, int[2], int[2], PtrStepb, PtrStepb);
static Caller multipass_callers[] =
{
minMaxLocMultipassCaller<unsigned char>, minMaxLocMultipassCaller<char>,
minMaxLocMultipassCaller<unsigned short>, minMaxLocMultipassCaller<short>,
minMaxLocMultipassCaller<int>, minMaxLocMultipassCaller<float>, 0
};
static Caller singlepass_callers[] =
typedef void (*func_t)(const PtrStepSzb src, const PtrStepb mask, double* minval, double* maxval, int* minloc, int* maxloc, PtrStepb valbuf, PtrStep<unsigned int> locbuf);
static const func_t funcs[] =
{
minMaxLocCaller<unsigned char>, minMaxLocCaller<char>,
minMaxLocCaller<unsigned short>, minMaxLocCaller<short>,
minMaxLocCaller<int>, minMaxLocCaller<float>, minMaxLocCaller<double>
::minMaxLoc::run<uchar>,
::minMaxLoc::run<schar>,
::minMaxLoc::run<ushort>,
::minMaxLoc::run<short>,
::minMaxLoc::run<int>,
::minMaxLoc::run<float>,
::minMaxLoc::run<double>
};
static MaskedCaller masked_multipass_callers[] =
{
minMaxLocMaskMultipassCaller<unsigned char>, minMaxLocMaskMultipassCaller<char>,
minMaxLocMaskMultipassCaller<unsigned short>, minMaxLocMaskMultipassCaller<short>,
minMaxLocMaskMultipassCaller<int>, minMaxLocMaskMultipassCaller<float>, 0
};
static MaskedCaller masked_singlepass_callers[] =
{
minMaxLocMaskCaller<unsigned char>, minMaxLocMaskCaller<char>,
minMaxLocMaskCaller<unsigned short>, minMaxLocMaskCaller<short>,
minMaxLocMaskCaller<int>, minMaxLocMaskCaller<float>, minMaxLocMaskCaller<double>
};
CV_Assert(src.depth() <= CV_64F);
CV_Assert(src.channels() == 1);
CV_Assert(mask.empty() || (mask.type() == CV_8U && src.size() == mask.size()));
CV_Assert( src.channels() == 1 );
CV_Assert( mask.empty() || (mask.size() == src.size() && mask.type() == CV_8U) );
if (src.depth() == CV_64F)
{
@ -558,61 +414,28 @@ void cv::gpu::minMaxLoc(const GpuMat& src, double* minVal, double* maxVal, Point
CV_Error(CV_StsUnsupportedFormat, "The device doesn't support double");
}
double minVal_; if (!minVal) minVal = &minVal_;
double maxVal_; if (!maxVal) maxVal = &maxVal_;
int minLoc_[2];
int maxLoc_[2];
Size valbuf_size, locbuf_size;
getBufSizeRequired(src.cols, src.rows, static_cast<int>(src.elemSize()), valbuf_size.width,
valbuf_size.height, locbuf_size.width, locbuf_size.height);
::minMaxLoc::getBufSize(src.cols, src.rows, src.elemSize(), valbuf_size.width, valbuf_size.height, locbuf_size.width, locbuf_size.height);
ensureSizeIsEnough(valbuf_size, CV_8U, valBuf);
ensureSizeIsEnough(locbuf_size, CV_8U, locBuf);
if (mask.empty())
{
Caller* callers = multipass_callers;
if (TargetArchs::builtWith(GLOBAL_ATOMICS) && DeviceInfo().supports(GLOBAL_ATOMICS))
callers = singlepass_callers;
Caller caller = callers[src.type()];
CV_Assert(caller != 0);
caller(src, minVal, maxVal, minLoc_, maxLoc_, valBuf, locBuf);
}
else
{
MaskedCaller* callers = masked_multipass_callers;
if (TargetArchs::builtWith(GLOBAL_ATOMICS) && DeviceInfo().supports(GLOBAL_ATOMICS))
callers = masked_singlepass_callers;
MaskedCaller caller = callers[src.type()];
CV_Assert(caller != 0);
caller(src, mask, minVal, maxVal, minLoc_, maxLoc_, valBuf, locBuf);
}
const func_t func = funcs[src.depth()];
if (minLoc) { minLoc->x = minLoc_[0]; minLoc->y = minLoc_[1]; }
if (maxLoc) { maxLoc->x = maxLoc_[0]; maxLoc->y = maxLoc_[1]; }
double temp1, temp2;
Point temp3, temp4;
func(src, mask, minVal ? minVal : &temp1, maxVal ? maxVal : &temp2, minLoc ? &minLoc->x : &temp3.x, maxLoc ? &maxLoc->x : &temp4.x, valBuf, locBuf);
}
//////////////////////////////////////////////////////////////////////////////
// Count non-zero elements
// countNonZero
namespace cv { namespace gpu { namespace device
namespace countNonZero
{
namespace matrix_reductions
{
namespace countnonzero
{
void getBufSizeRequired(int cols, int rows, int& bufcols, int& bufrows);
void getBufSize(int cols, int rows, int& bufcols, int& bufrows);
template <typename T>
int countNonZeroCaller(const PtrStepSzb src, PtrStepb buf);
template <typename T>
int countNonZeroMultipassCaller(const PtrStepSzb src, PtrStepb buf);
}
}
}}}
template <typename T>
int run(const PtrStepSzb src, PtrStep<unsigned int> buf);
}
int cv::gpu::countNonZero(const GpuMat& src)
{
@ -620,27 +443,20 @@ int cv::gpu::countNonZero(const GpuMat& src)
return countNonZero(src, buf);
}
int cv::gpu::countNonZero(const GpuMat& src, GpuMat& buf)
{
using namespace ::cv::gpu::device::matrix_reductions::countnonzero;
typedef int (*Caller)(const PtrStepSzb src, PtrStepb buf);
static Caller multipass_callers[7] =
typedef int (*func_t)(const PtrStepSzb src, PtrStep<unsigned int> buf);
static const func_t funcs[] =
{
countNonZeroMultipassCaller<unsigned char>, countNonZeroMultipassCaller<char>,
countNonZeroMultipassCaller<unsigned short>, countNonZeroMultipassCaller<short>,
countNonZeroMultipassCaller<int>, countNonZeroMultipassCaller<float>, 0
::countNonZero::run<uchar>,
::countNonZero::run<schar>,
::countNonZero::run<ushort>,
::countNonZero::run<short>,
::countNonZero::run<int>,
::countNonZero::run<float>,
::countNonZero::run<double>
};
static Caller singlepass_callers[7] =
{
countNonZeroCaller<unsigned char>, countNonZeroCaller<char>,
countNonZeroCaller<unsigned short>, countNonZeroCaller<short>,
countNonZeroCaller<int>, countNonZeroCaller<float>, countNonZeroCaller<double> };
CV_Assert(src.depth() <= CV_64F);
CV_Assert(src.channels() == 1);
if (src.depth() == CV_64F)
@ -650,168 +466,190 @@ int cv::gpu::countNonZero(const GpuMat& src, GpuMat& buf)
}
Size buf_size;
getBufSizeRequired(src.cols, src.rows, buf_size.width, buf_size.height);
::countNonZero::getBufSize(src.cols, src.rows, buf_size.width, buf_size.height);
ensureSizeIsEnough(buf_size, CV_8U, buf);
Caller* callers = multipass_callers;
if (TargetArchs::builtWith(GLOBAL_ATOMICS) && DeviceInfo().supports(GLOBAL_ATOMICS))
callers = singlepass_callers;
const func_t func = funcs[src.depth()];
Caller caller = callers[src.type()];
CV_Assert(caller != 0);
return caller(src, buf);
return func(src, buf);
}
//////////////////////////////////////////////////////////////////////////////
// reduce
namespace cv { namespace gpu { namespace device
namespace reduce
{
namespace matrix_reductions
{
template <typename T, typename S, typename D> void reduceRows_gpu(const PtrStepSzb& src, const PtrStepSzb& dst, int reduceOp, cudaStream_t stream);
template <typename T, typename S, typename D> void reduceCols_gpu(const PtrStepSzb& src, int cn, const PtrStepSzb& dst, int reduceOp, cudaStream_t stream);
}
}}}
template <typename T, typename S, typename D>
void rows(PtrStepSzb src, void* dst, int op, cudaStream_t stream);
template <typename T, typename S, typename D>
void cols(PtrStepSzb src, void* dst, int cn, int op, cudaStream_t stream);
}
void cv::gpu::reduce(const GpuMat& src, GpuMat& dst, int dim, int reduceOp, int dtype, Stream& stream)
{
using namespace ::cv::gpu::device::matrix_reductions;
CV_Assert(src.depth() <= CV_32F && src.channels() <= 4 && dtype <= CV_32F);
CV_Assert(dim == 0 || dim == 1);
CV_Assert(reduceOp == CV_REDUCE_SUM || reduceOp == CV_REDUCE_AVG || reduceOp == CV_REDUCE_MAX || reduceOp == CV_REDUCE_MIN);
CV_Assert( src.channels() <= 4 );
CV_Assert( dim == 0 || dim == 1 );
CV_Assert( reduceOp == CV_REDUCE_SUM || reduceOp == CV_REDUCE_AVG || reduceOp == CV_REDUCE_MAX || reduceOp == CV_REDUCE_MIN );
if (dtype < 0)
dtype = src.depth();
dst.create(1, dim == 0 ? src.cols : src.rows, CV_MAKETYPE(dtype, src.channels()));
dst.create(1, dim == 0 ? src.cols : src.rows, CV_MAKE_TYPE(CV_MAT_DEPTH(dtype), src.channels()));
if (dim == 0)
{
typedef void (*caller_t)(const PtrStepSzb& src, const PtrStepSzb& dst, int reduceOp, cudaStream_t stream);
static const caller_t callers[6][6] =
typedef void (*func_t)(PtrStepSzb src, void* dst, int op, cudaStream_t stream);
static const func_t funcs[7][7] =
{
{
reduceRows_gpu<unsigned char, int, unsigned char>,
0/*reduceRows_gpu<unsigned char, int, signed char>*/,
0/*reduceRows_gpu<unsigned char, int, unsigned short>*/,
0/*reduceRows_gpu<unsigned char, int, short>*/,
reduceRows_gpu<unsigned char, int, int>,
reduceRows_gpu<unsigned char, int, float>
::reduce::rows<unsigned char, int, unsigned char>,
0/*::reduce::rows<unsigned char, int, signed char>*/,
0/*::reduce::rows<unsigned char, int, unsigned short>*/,
0/*::reduce::rows<unsigned char, int, short>*/,
::reduce::rows<unsigned char, int, int>,
::reduce::rows<unsigned char, float, float>,
::reduce::rows<unsigned char, double, double>
},
{
0/*::reduce::rows<signed char, int, unsigned char>*/,
0/*::reduce::rows<signed char, int, signed char>*/,
0/*::reduce::rows<signed char, int, unsigned short>*/,
0/*::reduce::rows<signed char, int, short>*/,
0/*::reduce::rows<signed char, int, int>*/,
0/*::reduce::rows<signed char, float, float>*/,
0/*::reduce::rows<signed char, double, double>*/
},
{
0/*reduceRows_gpu<signed char, int, unsigned char>*/,
0/*reduceRows_gpu<signed char, int, signed char>*/,
0/*reduceRows_gpu<signed char, int, unsigned short>*/,
0/*reduceRows_gpu<signed char, int, short>*/,
0/*reduceRows_gpu<signed char, int, int>*/,
0/*reduceRows_gpu<signed char, int, float>*/
0/*::reduce::rows<unsigned short, int, unsigned char>*/,
0/*::reduce::rows<unsigned short, int, signed char>*/,
::reduce::rows<unsigned short, int, unsigned short>,
0/*::reduce::rows<unsigned short, int, short>*/,
::reduce::rows<unsigned short, int, int>,
::reduce::rows<unsigned short, float, float>,
::reduce::rows<unsigned short, double, double>
},
{
0/*reduceRows_gpu<unsigned short, int, unsigned char>*/,
0/*reduceRows_gpu<unsigned short, int, signed char>*/,
reduceRows_gpu<unsigned short, int, unsigned short>,
0/*reduceRows_gpu<unsigned short, int, short>*/,
reduceRows_gpu<unsigned short, int, int>,
reduceRows_gpu<unsigned short, int, float>
0/*::reduce::rows<short, int, unsigned char>*/,
0/*::reduce::rows<short, int, signed char>*/,
0/*::reduce::rows<short, int, unsigned short>*/,
::reduce::rows<short, int, short>,
::reduce::rows<short, int, int>,
::reduce::rows<short, float, float>,
::reduce::rows<short, double, double>
},
{
0/*reduceRows_gpu<short, int, unsigned char>*/,
0/*reduceRows_gpu<short, int, signed char>*/,
0/*reduceRows_gpu<short, int, unsigned short>*/,
reduceRows_gpu<short, int, short>,
reduceRows_gpu<short, int, int>,
reduceRows_gpu<short, int, float>
0/*::reduce::rows<int, int, unsigned char>*/,
0/*::reduce::rows<int, int, signed char>*/,
0/*::reduce::rows<int, int, unsigned short>*/,
0/*::reduce::rows<int, int, short>*/,
::reduce::rows<int, int, int>,
::reduce::rows<int, float, float>,
::reduce::rows<int, double, double>
},
{
0/*reduceRows_gpu<int, int, unsigned char>*/,
0/*reduceRows_gpu<int, int, signed char>*/,
0/*reduceRows_gpu<int, int, unsigned short>*/,
0/*reduceRows_gpu<int, int, short>*/,
reduceRows_gpu<int, int, int>,
reduceRows_gpu<int, int, float>
0/*::reduce::rows<float, float, unsigned char>*/,
0/*::reduce::rows<float, float, signed char>*/,
0/*::reduce::rows<float, float, unsigned short>*/,
0/*::reduce::rows<float, float, short>*/,
0/*::reduce::rows<float, float, int>*/,
::reduce::rows<float, float, float>,
::reduce::rows<float, double, double>
},
{
0/*reduceRows_gpu<float, float, unsigned char>*/,
0/*reduceRows_gpu<float, float, signed char>*/,
0/*reduceRows_gpu<float, float, unsigned short>*/,
0/*reduceRows_gpu<float, float, short>*/,
0/*reduceRows_gpu<float, float, int>*/,
reduceRows_gpu<float, float, float>
0/*::reduce::rows<double, double, unsigned char>*/,
0/*::reduce::rows<double, double, signed char>*/,
0/*::reduce::rows<double, double, unsigned short>*/,
0/*::reduce::rows<double, double, short>*/,
0/*::reduce::rows<double, double, int>*/,
0/*::reduce::rows<double, double, float>*/,
::reduce::rows<double, double, double>
}
};
const caller_t func = callers[src.depth()][dst.depth()];
const func_t func = funcs[src.depth()][dst.depth()];
if (!func)
CV_Error(CV_StsUnsupportedFormat, "Unsupported combination of input and output array formats");
func(src.reshape(1), dst.reshape(1), reduceOp, StreamAccessor::getStream(stream));
func(src.reshape(1), dst.data, reduceOp, StreamAccessor::getStream(stream));
}
else
{
typedef void (*caller_t)(const PtrStepSzb& src, int cn, const PtrStepSzb& dst, int reduceOp, cudaStream_t stream);
static const caller_t callers[6][6] =
typedef void (*func_t)(PtrStepSzb src, void* dst, int cn, int op, cudaStream_t stream);
static const func_t funcs[7][7] =
{
{
reduceCols_gpu<unsigned char, int, unsigned char>,
0/*reduceCols_gpu<unsigned char, int, signed char>*/,
0/*reduceCols_gpu<unsigned char, int, unsigned short>*/,
0/*reduceCols_gpu<unsigned char, int, short>*/,
reduceCols_gpu<unsigned char, int, int>,
reduceCols_gpu<unsigned char, int, float>
::reduce::cols<unsigned char, int, unsigned char>,
0/*::reduce::cols<unsigned char, int, signed char>*/,
0/*::reduce::cols<unsigned char, int, unsigned short>*/,
0/*::reduce::cols<unsigned char, int, short>*/,
::reduce::cols<unsigned char, int, int>,
::reduce::cols<unsigned char, float, float>,
::reduce::cols<unsigned char, double, double>
},
{
0/*::reduce::cols<signed char, int, unsigned char>*/,
0/*::reduce::cols<signed char, int, signed char>*/,
0/*::reduce::cols<signed char, int, unsigned short>*/,
0/*::reduce::cols<signed char, int, short>*/,
0/*::reduce::cols<signed char, int, int>*/,
0/*::reduce::cols<signed char, float, float>*/,
0/*::reduce::cols<signed char, double, double>*/
},
{
0/*reduceCols_gpu<signed char, int, unsigned char>*/,
0/*reduceCols_gpu<signed char, int, signed char>*/,
0/*reduceCols_gpu<signed char, int, unsigned short>*/,
0/*reduceCols_gpu<signed char, int, short>*/,
0/*reduceCols_gpu<signed char, int, int>*/,
0/*reduceCols_gpu<signed char, int, float>*/
0/*::reduce::cols<unsigned short, int, unsigned char>*/,
0/*::reduce::cols<unsigned short, int, signed char>*/,
::reduce::cols<unsigned short, int, unsigned short>,
0/*::reduce::cols<unsigned short, int, short>*/,
::reduce::cols<unsigned short, int, int>,
::reduce::cols<unsigned short, float, float>,
::reduce::cols<unsigned short, double, double>
},
{
0/*reduceCols_gpu<unsigned short, int, unsigned char>*/,
0/*reduceCols_gpu<unsigned short, int, signed char>*/,
reduceCols_gpu<unsigned short, int, unsigned short>,
0/*reduceCols_gpu<unsigned short, int, short>*/,
reduceCols_gpu<unsigned short, int, int>,
reduceCols_gpu<unsigned short, int, float>
0/*::reduce::cols<short, int, unsigned char>*/,
0/*::reduce::cols<short, int, signed char>*/,
0/*::reduce::cols<short, int, unsigned short>*/,
::reduce::cols<short, int, short>,
::reduce::cols<short, int, int>,
::reduce::cols<short, float, float>,
::reduce::cols<short, double, double>
},
{
0/*reduceCols_gpu<short, int, unsigned char>*/,
0/*reduceCols_gpu<short, int, signed char>*/,
0/*reduceCols_gpu<short, int, unsigned short>*/,
reduceCols_gpu<short, int, short>,
reduceCols_gpu<short, int, int>,
reduceCols_gpu<short, int, float>
0/*::reduce::cols<int, int, unsigned char>*/,
0/*::reduce::cols<int, int, signed char>*/,
0/*::reduce::cols<int, int, unsigned short>*/,
0/*::reduce::cols<int, int, short>*/,
::reduce::cols<int, int, int>,
::reduce::cols<int, float, float>,
::reduce::cols<int, double, double>
},
{
0/*reduceCols_gpu<int, int, unsigned char>*/,
0/*reduceCols_gpu<int, int, signed char>*/,
0/*reduceCols_gpu<int, int, unsigned short>*/,
0/*reduceCols_gpu<int, int, short>*/,
reduceCols_gpu<int, int, int>,
reduceCols_gpu<int, int, float>
0/*::reduce::cols<float, float, unsigned char>*/,
0/*::reduce::cols<float, float, signed char>*/,
0/*::reduce::cols<float, float, unsigned short>*/,
0/*::reduce::cols<float, float, short>*/,
0/*::reduce::cols<float, float, int>*/,
::reduce::cols<float, float, float>,
::reduce::cols<float, double, double>
},
{
0/*reduceCols_gpu<float, unsigned char>*/,
0/*reduceCols_gpu<float, signed char>*/,
0/*reduceCols_gpu<float, unsigned short>*/,
0/*reduceCols_gpu<float, short>*/,
0/*reduceCols_gpu<float, int>*/,
reduceCols_gpu<float, float, float>
0/*::reduce::cols<double, double, unsigned char>*/,
0/*::reduce::cols<double, double, signed char>*/,
0/*::reduce::cols<double, double, unsigned short>*/,
0/*::reduce::cols<double, double, short>*/,
0/*::reduce::cols<double, double, int>*/,
0/*::reduce::cols<double, double, float>*/,
::reduce::cols<double, double, double>
}
};
const caller_t func = callers[src.depth()][dst.depth()];
const func_t func = funcs[src.depth()][dst.depth()];
if (!func)
CV_Error(CV_StsUnsupportedFormat, "Unsupported combination of input and output array formats");
func(src, src.channels(), dst, reduceOp, StreamAccessor::getStream(stream));
func(src, dst.data, src.channels(), reduceOp, StreamAccessor::getStream(stream));
}
}

14
modules/gpu/test/test_core.cpp
Unescape View File

@ -2982,7 +2982,7 @@ TEST_P(Sum, Sqr)
INSTANTIATE_TEST_CASE_P(GPU_Core, Sum, testing::Combine(
ALL_DEVICES,
DIFFERENT_SIZES,
TYPES(CV_8U, CV_32F, 1, 4),
TYPES(CV_8U, CV_64F, 1, 4),
WHOLE_SUBMAT));
////////////////////////////////////////////////////////////////////////////////
@ -3351,7 +3351,14 @@ PARAM_TEST_CASE(Reduce, cv::gpu::DeviceInfo, cv::Size, MatDepth, Channels, Reduc
cv::gpu::setDevice(devInfo.deviceID());
type = CV_MAKE_TYPE(depth, channels);
dst_depth = (reduceOp == CV_REDUCE_MAX || reduceOp == CV_REDUCE_MIN) ? depth : CV_32F;
if (reduceOp == CV_REDUCE_MAX || reduceOp == CV_REDUCE_MIN)
dst_depth = depth;
else if (reduceOp == CV_REDUCE_SUM)
dst_depth = depth == CV_8U ? CV_32S : depth < CV_64F ? CV_32F : depth;
else
dst_depth = depth < CV_32F ? CV_32F : depth;
dst_type = CV_MAKE_TYPE(dst_depth, channels);
}
@ -3392,7 +3399,8 @@ INSTANTIATE_TEST_CASE_P(GPU_Core, Reduce, testing::Combine(
testing::Values(MatDepth(CV_8U),
MatDepth(CV_16U),
MatDepth(CV_16S),
MatDepth(CV_32F)),
MatDepth(CV_32F),
MatDepth(CV_64F)),
ALL_CHANNELS,
ALL_REDUCE_CODES,
WHOLE_SUBMAT));

Write Preview
Loading…
Cancel
Save