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

used new device layer for cv::gpu::compare

Browse Source
This commit is contained in:
Vladislav Vinogradov
2013-07-29 17:53:03 +04:00
parent 43e811118b
commit ef9917ecf1
4 changed files with 276 additions and 460 deletions
Download Patch File Download Diff File Expand all files Collapse all files

142
modules/cudaarithm/src/element_operations.cpp
View File

@@ -454,147 +454,9 @@ void cv::cuda::absdiff(InputArray src1, InputArray src2, OutputArray dst, Stream
//////////////////////////////////////////////////////////////////////////////
// compare
namespace arithm
{
void cmpMatEq_v4(PtrStepSz<uint> src1, PtrStepSz<uint> src2, PtrStepSz<uint> dst, cudaStream_t stream);
void cmpMatNe_v4(PtrStepSz<uint> src1, PtrStepSz<uint> src2, PtrStepSz<uint> dst, cudaStream_t stream);
void cmpMatLt_v4(PtrStepSz<uint> src1, PtrStepSz<uint> src2, PtrStepSz<uint> dst, cudaStream_t stream);
void cmpMatLe_v4(PtrStepSz<uint> src1, PtrStepSz<uint> src2, PtrStepSz<uint> dst, cudaStream_t stream);
void cmpMat(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, const GpuMat&, double, Stream& stream, int cmpop);
template <typename T> void cmpMatEq(PtrStepSzb src1, PtrStepSzb src2, PtrStepSzb dst, cudaStream_t stream);
template <typename T> void cmpMatNe(PtrStepSzb src1, PtrStepSzb src2, PtrStepSzb dst, cudaStream_t stream);
template <typename T> void cmpMatLt(PtrStepSzb src1, PtrStepSzb src2, PtrStepSzb dst, cudaStream_t stream);
template <typename T> void cmpMatLe(PtrStepSzb src1, PtrStepSzb src2, PtrStepSzb dst, cudaStream_t stream);
}
static void cmpMat(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, const GpuMat&, double, Stream& _stream, int cmpop)
{
using namespace arithm;
typedef void (*func_t)(PtrStepSzb src1, PtrStepSzb src2, PtrStepSzb dst, cudaStream_t stream);
static const func_t funcs[7][4] =
{
{cmpMatEq<unsigned char> , cmpMatNe<unsigned char> , cmpMatLt<unsigned char> , cmpMatLe<unsigned char> },
{cmpMatEq<signed char> , cmpMatNe<signed char> , cmpMatLt<signed char> , cmpMatLe<signed char> },
{cmpMatEq<unsigned short>, cmpMatNe<unsigned short>, cmpMatLt<unsigned short>, cmpMatLe<unsigned short>},
{cmpMatEq<short> , cmpMatNe<short> , cmpMatLt<short> , cmpMatLe<short> },
{cmpMatEq<int> , cmpMatNe<int> , cmpMatLt<int> , cmpMatLe<int> },
{cmpMatEq<float> , cmpMatNe<float> , cmpMatLt<float> , cmpMatLe<float> },
{cmpMatEq<double> , cmpMatNe<double> , cmpMatLt<double> , cmpMatLe<double> }
};
typedef void (*func_v4_t)(PtrStepSz<uint> src1, PtrStepSz<uint> src2, PtrStepSz<uint> dst, cudaStream_t stream);
static const func_v4_t funcs_v4[] =
{
cmpMatEq_v4, cmpMatNe_v4, cmpMatLt_v4, cmpMatLe_v4
};
const int depth = src1.depth();
const int cn = src1.channels();
cudaStream_t stream = StreamAccessor::getStream(_stream);
static const int codes[] =
{
0, 2, 3, 2, 3, 1
};
const GpuMat* psrc1[] =
{
&src1, &src2, &src2, &src1, &src1, &src1
};
const GpuMat* psrc2[] =
{
&src2, &src1, &src1, &src2, &src2, &src2
};
const int code = codes[cmpop];
PtrStepSzb src1_(src1.rows, src1.cols * cn, psrc1[cmpop]->data, psrc1[cmpop]->step);
PtrStepSzb src2_(src1.rows, src1.cols * cn, psrc2[cmpop]->data, psrc2[cmpop]->step);
PtrStepSzb dst_(src1.rows, src1.cols * cn, dst.data, dst.step);
if (depth == CV_8U && (src1_.cols & 3) == 0)
{
const intptr_t src1ptr = reinterpret_cast<intptr_t>(src1_.data);
const intptr_t src2ptr = reinterpret_cast<intptr_t>(src2_.data);
const intptr_t dstptr = reinterpret_cast<intptr_t>(dst_.data);
const bool isAllAligned = (src1ptr & 31) == 0 && (src2ptr & 31) == 0 && (dstptr & 31) == 0;
if (isAllAligned)
{
const int vcols = src1_.cols >> 2;
funcs_v4[code](PtrStepSz<unsigned int>(src1_.rows, vcols, (unsigned int*) src1_.data, src1_.step),
PtrStepSz<unsigned int>(src1_.rows, vcols, (unsigned int*) src2_.data, src2_.step),
PtrStepSz<unsigned int>(src1_.rows, vcols, (unsigned int*) dst_.data, dst_.step),
stream);
return;
}
}
const func_t func = funcs[depth][code];
func(src1_, src2_, dst_, stream);
}
namespace arithm
{
template <typename T> void cmpScalarEq(PtrStepSzb src, int cn, double val[4], PtrStepSzb dst, cudaStream_t stream);
template <typename T> void cmpScalarNe(PtrStepSzb src, int cn, double val[4], PtrStepSzb dst, cudaStream_t stream);
template <typename T> void cmpScalarLt(PtrStepSzb src, int cn, double val[4], PtrStepSzb dst, cudaStream_t stream);
template <typename T> void cmpScalarLe(PtrStepSzb src, int cn, double val[4], PtrStepSzb dst, cudaStream_t stream);
template <typename T> void cmpScalarGt(PtrStepSzb src, int cn, double val[4], PtrStepSzb dst, cudaStream_t stream);
template <typename T> void cmpScalarGe(PtrStepSzb src, int cn, double val[4], PtrStepSzb dst, cudaStream_t stream);
}
namespace
{
template <typename T> void castScalar(Scalar& sc)
{
sc.val[0] = saturate_cast<T>(sc.val[0]);
sc.val[1] = saturate_cast<T>(sc.val[1]);
sc.val[2] = saturate_cast<T>(sc.val[2]);
sc.val[3] = saturate_cast<T>(sc.val[3]);
}
}
static void cmpScalar(const GpuMat& src, Scalar val, bool inv, GpuMat& dst, const GpuMat&, double, Stream& stream, int cmpop)
{
using namespace arithm;
typedef void (*func_t)(PtrStepSzb src, int cn, double val[4], PtrStepSzb dst, cudaStream_t stream);
static const func_t funcs[7][6] =
{
{cmpScalarEq<unsigned char> , cmpScalarGt<unsigned char> , cmpScalarGe<unsigned char> , cmpScalarLt<unsigned char> , cmpScalarLe<unsigned char> , cmpScalarNe<unsigned char> },
{cmpScalarEq<signed char> , cmpScalarGt<signed char> , cmpScalarGe<signed char> , cmpScalarLt<signed char> , cmpScalarLe<signed char> , cmpScalarNe<signed char> },
{cmpScalarEq<unsigned short>, cmpScalarGt<unsigned short>, cmpScalarGe<unsigned short>, cmpScalarLt<unsigned short>, cmpScalarLe<unsigned short>, cmpScalarNe<unsigned short>},
{cmpScalarEq<short> , cmpScalarGt<short> , cmpScalarGe<short> , cmpScalarLt<short> , cmpScalarLe<short> , cmpScalarNe<short> },
{cmpScalarEq<int> , cmpScalarGt<int> , cmpScalarGe<int> , cmpScalarLt<int> , cmpScalarLe<int> , cmpScalarNe<int> },
{cmpScalarEq<float> , cmpScalarGt<float> , cmpScalarGe<float> , cmpScalarLt<float> , cmpScalarLe<float> , cmpScalarNe<float> },
{cmpScalarEq<double> , cmpScalarGt<double> , cmpScalarGe<double> , cmpScalarLt<double> , cmpScalarLe<double> , cmpScalarNe<double> }
};
typedef void (*cast_func_t)(Scalar& sc);
static const cast_func_t cast_func[] =
{
castScalar<unsigned char>, castScalar<signed char>, castScalar<unsigned short>, castScalar<short>, castScalar<int>, castScalar<float>, castScalar<double>
};
if (inv)
{
// src1 is a scalar; swap it with src2
cmpop = cmpop == CMP_LT ? CMP_GT : cmpop == CMP_LE ? CMP_GE :
cmpop == CMP_GE ? CMP_LE : cmpop == CMP_GT ? CMP_LT : cmpop;
}
const int depth = src.depth();
const int cn = src.channels();
cast_func[depth](val);
funcs[depth][cmpop](src, cn, val.val, dst, StreamAccessor::getStream(stream));
}
void cmpScalar(const GpuMat& src, Scalar val, bool inv, GpuMat& dst, const GpuMat&, double, Stream& stream, int cmpop);
void cv::cuda::compare(InputArray src1, InputArray src2, OutputArray dst, int cmpop, Stream& stream)
{