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

Added implementation and test for the GPU version of subtract, multiply, divide, transpose, absdiff, threshold, compare, meanStdDev, norm, based on NPP.

Browse Source
This commit is contained in:
Vladislav Vinogradov
2010-09-13 14:30:09 +00:00
parent 88a7a8f567
commit 37d39bd9de
6 changed files with 706 additions and 194 deletions
Download Patch File Download Diff File Expand all files Collapse all files

209
modules/gpu/src/arithm.cpp
View File

@@ -49,44 +49,211 @@ using namespace std;
#if !defined (HAVE_CUDA)
void cv::gpu::add(const GpuMat& src1, const GpuMat& src2, GpuMat& dst) { throw_nogpu(); }
void cv::gpu::subtract(const GpuMat& src1, const GpuMat& src2, GpuMat& dst) { throw_nogpu(); }
void cv::gpu::multiply(const GpuMat& src1, const GpuMat& src2, GpuMat& dst) { throw_nogpu(); }
void cv::gpu::divide(const GpuMat& src1, const GpuMat& src2, GpuMat& dst) { throw_nogpu(); }
void cv::gpu::transpose(const GpuMat& src1, GpuMat& dst) { throw_nogpu(); }
void cv::gpu::absdiff(const GpuMat& src1, const GpuMat& src2, GpuMat& dst) { throw_nogpu(); }
double cv::gpu::threshold(const GpuMat& src, GpuMat& dst, double thresh, double maxVal, int thresholdType) { throw_nogpu(); return 0.0; }
void cv::gpu::compare(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, int cmpop) { throw_nogpu(); }
void cv::gpu::meanStdDev(const GpuMat& mtx, Scalar& mean, Scalar& stddev) { throw_nogpu(); }
double cv::gpu::norm(const GpuMat& src1, int normType) { throw_nogpu(); return 0.0; }
double cv::gpu::norm(const GpuMat& src1, const GpuMat& src2, int normType) { throw_nogpu(); return 0.0; }
#else /* !defined (HAVE_CUDA) */
namespace
{
typedef NppStatus (*npp_binary_func_8u_scale_t)(const Npp8u* pSrc1, int nSrc1Step, const Npp8u* pSrc2, int nSrc2Step, Npp8u* pDst, int nDstStep,
NppiSize oSizeROI, int nScaleFactor);
typedef NppStatus (*npp_binary_func_32f_t)(const Npp32f* pSrc1, int nSrc1Step, const Npp32f* pSrc2, int nSrc2Step, Npp32f* pDst,
int nDstStep, NppiSize oSizeROI);
void nppFuncCaller(const GpuMat& src1, const GpuMat& src2, GpuMat& dst,
npp_binary_func_8u_scale_t npp_func_8uc1, npp_binary_func_8u_scale_t npp_func_8uc4, npp_binary_func_32f_t npp_func_32fc1)
{
CV_DbgAssert(src1.size() == src2.size() && src1.type() == src2.type());
CV_Assert(src1.type() == CV_8UC1 || src1.type() == CV_8UC4 || src1.type() == CV_32FC1);
dst.create( src1.size(), src1.type() );
NppiSize sz;
sz.width = src1.cols;
sz.height = src1.rows;
if (src1.depth() == CV_8U)
{
if (src1.channels() == 1)
{
npp_func_8uc1((const Npp8u*)src1.ptr<char>(), src1.step,
(const Npp8u*)src2.ptr<char>(), src2.step,
(Npp8u*)dst.ptr<char>(), dst.step, sz, 0);
}
else
{
npp_func_8uc4((const Npp8u*)src1.ptr<char>(), src1.step,
(const Npp8u*)src2.ptr<char>(), src2.step,
(Npp8u*)dst.ptr<char>(), dst.step, sz, 0);
}
}
else //if (src1.depth() == CV_32F)
{
npp_func_32fc1((const Npp32f*)src1.ptr<float>(), src1.step,
(const Npp32f*)src2.ptr<float>(), src2.step,
(Npp32f*)dst.ptr<float>(), dst.step, sz);
}
}
}
void cv::gpu::add(const GpuMat& src1, const GpuMat& src2, GpuMat& dst)
{
nppFuncCaller(src1, src2, dst, nppiAdd_8u_C1RSfs, nppiAdd_8u_C4RSfs, nppiAdd_32f_C1R);
}
void cv::gpu::subtract(const GpuMat& src1, const GpuMat& src2, GpuMat& dst)
{
nppFuncCaller(src2, src1, dst, nppiSub_8u_C1RSfs, nppiSub_8u_C4RSfs, nppiSub_32f_C1R);
}
void cv::gpu::multiply(const GpuMat& src1, const GpuMat& src2, GpuMat& dst)
{
nppFuncCaller(src1, src2, dst, nppiMul_8u_C1RSfs, nppiMul_8u_C4RSfs, nppiMul_32f_C1R);
}
void cv::gpu::divide(const GpuMat& src1, const GpuMat& src2, GpuMat& dst)
{
nppFuncCaller(src2, src1, dst, nppiDiv_8u_C1RSfs, nppiDiv_8u_C4RSfs, nppiDiv_32f_C1R);
}
void cv::gpu::transpose(const GpuMat& src, GpuMat& dst)
{
CV_Assert(src.type() == CV_8UC1);
dst.create( src.cols, src.rows, src.type() );
NppiSize sz;
sz.width = src.cols;
sz.height = src.rows;
nppiTranspose_8u_C1R((const Npp8u*)src.ptr<char>(), src.step, (Npp8u*)dst.ptr<char>(), dst.step, sz);
}
void cv::gpu::absdiff(const GpuMat& src1, const GpuMat& src2, GpuMat& dst)
{
CV_DbgAssert(src1.size() == src2.size() && src1.type() == src2.type());
CV_Assert((src1.depth() == CV_8U || src1.depth() == CV_32F) && src1.channels() == 1);
dst.create( src1.size(), src1.type() );
CV_Assert(src1.size() == src2.size() && src1.type() == src2.type());
int nChannels = src1.channels();
CV_DbgAssert((src1.depth() == CV_8U && nChannels == 1 || nChannels == 4) ||
(src1.depth() == CV_32F && nChannels == 1));
NppiSize sz;
sz.width = src1.cols;
sz.height = src1.rows;
if (src1.depth() == CV_8U)
{
if (nChannels == 1)
{
nppiAdd_8u_C1RSfs((const Npp8u*)src1.ptr<char>(), src1.step,
(const Npp8u*)src2.ptr<char>(), src2.step,
(Npp8u*)dst.ptr<char>(), dst.step, sz, 0);
}
else
{
nppiAdd_8u_C4RSfs((const Npp8u*)src1.ptr<char>(), src1.step,
(const Npp8u*)src2.ptr<char>(), src2.step,
(Npp8u*)dst.ptr<char>(), dst.step, sz, 0);
}
nppiAbsDiff_8u_C1R((const Npp8u*)src1.ptr<char>(), src1.step,
(const Npp8u*)src2.ptr<char>(), src2.step,
(Npp8u*)dst.ptr<char>(), dst.step, sz);
}
else //if (src1.depth() == CV_32F)
{
nppiAdd_32f_C1R((const Npp32f*)src1.ptr<float>(), src1.step,
(const Npp32f*)src2.ptr<float>(), src2.step,
(Npp32f*)dst.ptr<float>(), dst.step, sz);
nppiAbsDiff_32f_C1R((const Npp32f*)src1.ptr<float>(), src1.step,
(const Npp32f*)src2.ptr<float>(), src2.step,
(Npp32f*)dst.ptr<float>(), dst.step, sz);
}
}
double cv::gpu::threshold(const GpuMat& src, GpuMat& dst, double thresh, double /*maxVal*/, int thresholdType)
{
CV_Assert(src.type() == CV_32FC1 && thresholdType == THRESH_TRUNC);
dst.create( src.size(), src.type() );
NppiSize sz;
sz.width = src.cols;
sz.height = src.rows;
nppiThreshold_32f_C1R((const Npp32f*)src.ptr<float>(), src.step,
(Npp32f*)dst.ptr<float>(), dst.step, sz, (Npp32f)thresh, NPP_CMP_GREATER);
return thresh;
}
void cv::gpu::compare(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, int cmpop)
{
CV_DbgAssert(src1.size() == src2.size() && src1.type() == src2.type());
CV_Assert((src1.type() == CV_8UC4 || src1.type() == CV_32FC1) && cmpop != CMP_NE);
dst.create( src1.size(), CV_8UC1 );
static const NppCmpOp nppCmpOp[] = { NPP_CMP_EQ, NPP_CMP_GREATER, NPP_CMP_GREATER_EQ, NPP_CMP_LESS, NPP_CMP_LESS_EQ };
NppiSize sz;
sz.width = src1.cols;
sz.height = src1.rows;
if (src1.depth() == CV_8U)
{
nppiCompare_8u_C4R((const Npp8u*)src1.ptr<char>(), src1.step,
(const Npp8u*)src2.ptr<char>(), src2.step,
(Npp8u*)dst.ptr<char>(), dst.step, sz, nppCmpOp[cmpop]);
}
else //if (src1.depth() == CV_32F)
{
nppiCompare_32f_C1R((const Npp32f*)src1.ptr<float>(), src1.step,
(const Npp32f*)src2.ptr<float>(), src2.step,
(Npp8u*)dst.ptr<char>(), dst.step, sz, nppCmpOp[cmpop]);
}
}
void cv::gpu::meanStdDev(const GpuMat& src, Scalar& mean, Scalar& stddev)
{
CV_Assert(src.type() == CV_8UC1);
NppiSize sz;
sz.width = src.cols;
sz.height = src.rows;
nppiMean_StdDev_8u_C1R((const Npp8u*)src.ptr<char>(), src.step, sz, mean.val, stddev.val);
}
double cv::gpu::norm(const GpuMat& src1, int normType)
{
return norm(src1, GpuMat(src1.size(), src1.type(), Scalar::all(0.0)), normType);
}
double cv::gpu::norm(const GpuMat& src1, const GpuMat& src2, int normType)
{
CV_DbgAssert(src1.size() == src2.size() && src1.type() == src2.type());
CV_Assert((src1.type() == CV_8UC1) && (normType == NORM_INF || normType == NORM_L1 || normType == NORM_L2));
typedef NppStatus (*npp_norm_diff_func_t)(const Npp8u* pSrc1, int nSrcStep1, const Npp8u* pSrc2, int nSrcStep2,
NppiSize oSizeROI, Npp64f* pRetVal);
static const npp_norm_diff_func_t npp_norm_diff_func[] = {nppiNormDiff_Inf_8u_C1R, nppiNormDiff_L1_8u_C1R, nppiNormDiff_L2_8u_C1R};
NppiSize sz;
sz.width = src1.cols;
sz.height = src1.rows;
int funcIdx = normType >> 1;
Npp64f retVal[3];
npp_norm_diff_func[funcIdx]((const Npp8u*)src1.ptr<char>(), src1.step,
(const Npp8u*)src2.ptr<char>(), src2.step,
sz, retVal);
return retVal[0];
}
#endif /* !defined (HAVE_CUDA) */

2
modules/gpu/src/precomp.hpp
View File

@@ -55,7 +55,7 @@
#include <vector>
#include "opencv2/gpu/gpu.hpp"
#include "opencv2/imgproc/types_c.h"
#include "opencv2/imgproc/imgproc.hpp"
#if defined(HAVE_CUDA)