void cv::gpu::BroxOpticalFlow::operator ()(const GpuMat& frame0, const GpuMat& frame1, GpuMat& u, GpuMat& v, Stream& s)
{
ncvSetDebugOutputHandler(outputHandler);
CV_Assert(frame0.type() == CV_32FC1);
CV_Assert(frame1.size() == frame0.size() && frame1.type() == frame0.type());
u.create(frame0.size(), CV_32FC1);
v.create(frame0.size(), CV_32FC1);
cudaDeviceProp devProp;
cudaSafeCall( cudaGetDeviceProperties(&devProp, getDevice()) );
NCVBroxOpticalFlowDescriptor desc;
desc.alpha = alpha;
desc.gamma = gamma;
desc.scale_factor = scale_factor;
desc.number_of_inner_iterations = inner_iterations;
desc.number_of_outer_iterations = outer_iterations;
desc.number_of_solver_iterations = solver_iterations;
NCVMemSegment frame0MemSeg;
frame0MemSeg.begin.memtype = NCVMemoryTypeDevice;
frame0MemSeg.begin.ptr = const_cast<uchar*>(frame0.data);
frame0MemSeg.size = frame0.step * frame0.rows;
NCVMemSegment frame1MemSeg;
frame1MemSeg.begin.memtype = NCVMemoryTypeDevice;
frame1MemSeg.begin.ptr = const_cast<uchar*>(frame1.data);
frame1MemSeg.size = frame1.step * frame1.rows;
NCVMemSegment uMemSeg;
uMemSeg.begin.memtype = NCVMemoryTypeDevice;
uMemSeg.begin.ptr = u.ptr();
uMemSeg.size = u.step * u.rows;
NCVMemSegment vMemSeg;
vMemSeg.begin.memtype = NCVMemoryTypeDevice;
vMemSeg.begin.ptr = v.ptr();
vMemSeg.size = v.step * v.rows;
NCVMatrixReuse<Ncv32f> frame0Mat(frame0MemSeg, static_cast<Ncv32u>(devProp.textureAlignment), frame0.cols, frame0.rows, static_cast<Ncv32u>(frame0.step));
NCVMatrixReuse<Ncv32f> frame1Mat(frame1MemSeg, static_cast<Ncv32u>(devProp.textureAlignment), frame1.cols, frame1.rows, static_cast<Ncv32u>(frame1.step));
NCVMatrixReuse<Ncv32f> uMat(uMemSeg, static_cast<Ncv32u>(devProp.textureAlignment), u.cols, u.rows, static_cast<Ncv32u>(u.step));
NCVMatrixReuse<Ncv32f> vMat(vMemSeg, static_cast<Ncv32u>(devProp.textureAlignment), v.cols, v.rows, static_cast<Ncv32u>(v.step));
cudaStream_t stream = StreamAccessor::getStream(s);
size_t bufSize = getBufSize(desc, frame0Mat, frame1Mat, uMat, vMat, devProp);
ensureSizeIsEnough(1, static_cast<int>(bufSize), CV_8UC1, buf);
NCVMemStackAllocator gpuAllocator(NCVMemoryTypeDevice, bufSize, static_cast<Ncv32u>(devProp.textureAlignment), buf.ptr());
ncvSafeCall( NCVBroxOpticalFlow(desc, gpuAllocator, frame0Mat, frame1Mat, uMat, vMat, stream) );
}
void cv::gpu::equalizeHist(const GpuMat& src, GpuMat& dst, GpuMat& hist, GpuMat& buf, Stream& s)
{
CV_Assert(src.type() == CV_8UC1);
dst.create(src.size(), src.type());
int intBufSize;
nppSafeCall( nppsIntegralGetBufferSize_32s(256, &intBufSize) );
ensureSizeIsEnough(1, intBufSize + 256 * sizeof(int), CV_8UC1, buf);
GpuMat intBuf(1, intBufSize, CV_8UC1, buf.ptr());
GpuMat lut(1, 256, CV_32S, buf.ptr() + intBufSize);
calcHist(src, hist, s);
cudaStream_t stream = StreamAccessor::getStream(s);
NppStreamHandler h(stream);
nppSafeCall( nppsIntegral_32s(hist.ptr<Npp32s>(), lut.ptr<Npp32s>(), 256, intBuf.ptr<Npp8u>()) );
hist::equalizeHist(src, dst, lut.ptr<int>(), stream);
}
int cv::gpu::FAST_GPU::getKeyPoints(GpuMat& keypoints)
{
using namespace cv::gpu::cudev::fast;
if (count_ == 0)
return 0;
ensureSizeIsEnough(ROWS_COUNT, count_, CV_32FC1, keypoints);
if (nonmaxSupression)
return nonmaxSupression_gpu(kpLoc_.ptr<short2>(), count_, score_, keypoints.ptr<short2>(LOCATION_ROW), keypoints.ptr<float>(RESPONSE_ROW));
GpuMat locRow(1, count_, kpLoc_.type(), keypoints.ptr(0));
kpLoc_.colRange(0, count_).copyTo(locRow);
keypoints.row(1).setTo(Scalar::all(0));
return count_;
}
int cv::gpu::FAST_GPU::getKeyPoints(GpuMat& keypoints)
{
using namespace cv::gpu::device::fast;
if (!TargetArchs::builtWith(GLOBAL_ATOMICS) || !DeviceInfo().supports(GLOBAL_ATOMICS))
CV_Error(CV_StsNotImplemented, "The device doesn't support global atomics");
if (count_ == 0)
return 0;
ensureSizeIsEnough(ROWS_COUNT, count_, CV_32FC1, keypoints);
if (nonmaxSupression)
return nonmaxSupression_gpu(kpLoc_.ptr<short2>(), count_, score_, keypoints.ptr<short2>(LOCATION_ROW), keypoints.ptr<float>(RESPONSE_ROW));
GpuMat locRow(1, count_, kpLoc_.type(), keypoints.ptr(0));
kpLoc_.colRange(0, count_).copyTo(locRow);
keypoints.row(1).setTo(Scalar::all(0));
return count_;
}
void cv::gpu::PyrLKOpticalFlow::sparse(const GpuMat& prevImg, const GpuMat& nextImg, const GpuMat& prevPts, GpuMat& nextPts, GpuMat& status, GpuMat* err)
{
using namespace cv::gpu::device::pyrlk;
if (prevPts.empty())
{
nextPts.release();
status.release();
if (err) err->release();
return;
}
dim3 block, patch;
calcPatchSize(winSize, block, patch, isDeviceArch11_);
CV_Assert(prevImg.type() == CV_8UC1 || prevImg.type() == CV_8UC3 || prevImg.type() == CV_8UC4);
CV_Assert(prevImg.size() == nextImg.size() && prevImg.type() == nextImg.type());
CV_Assert(maxLevel >= 0);
CV_Assert(winSize.width > 2 && winSize.height > 2);
CV_Assert(patch.x > 0 && patch.x < 6 && patch.y > 0 && patch.y < 6);
CV_Assert(prevPts.rows == 1 && prevPts.type() == CV_32FC2);
if (useInitialFlow)
CV_Assert(nextPts.size() == prevPts.size() && nextPts.type() == CV_32FC2);
else
ensureSizeIsEnough(1, prevPts.cols, prevPts.type(), nextPts);
GpuMat temp1 = (useInitialFlow ? nextPts : prevPts).reshape(1);
GpuMat temp2 = nextPts.reshape(1);
multiply(temp1, Scalar::all(1.0 / (1 << maxLevel) / 2.0), temp2);
ensureSizeIsEnough(1, prevPts.cols, CV_8UC1, status);
status.setTo(Scalar::all(1));
if (err)
ensureSizeIsEnough(1, prevPts.cols, CV_32FC1, *err);
// build the image pyramids.
prevPyr_.resize(maxLevel + 1);
nextPyr_.resize(maxLevel + 1);
int cn = prevImg.channels();
if (cn == 1 || cn == 4)
{
prevImg.convertTo(prevPyr_[0], CV_32F);
nextImg.convertTo(nextPyr_[0], CV_32F);
}
else
{
cvtColor(prevImg, dx_calcBuf_, COLOR_BGR2BGRA);
dx_calcBuf_.convertTo(prevPyr_[0], CV_32F);
cvtColor(nextImg, dx_calcBuf_, COLOR_BGR2BGRA);
dx_calcBuf_.convertTo(nextPyr_[0], CV_32F);
}
for (int level = 1; level <= maxLevel; ++level)
{
pyrDown(prevPyr_[level - 1], prevPyr_[level]);
pyrDown(nextPyr_[level - 1], nextPyr_[level]);
}
loadConstants(make_int2(winSize.width, winSize.height), iters);
for (int level = maxLevel; level >= 0; level--)
{
if (cn == 1)
{
lkSparse1_gpu(prevPyr_[level], nextPyr_[level],
prevPts.ptr<float2>(), nextPts.ptr<float2>(), status.ptr(), level == 0 && err ? err->ptr<float>() : 0, prevPts.cols,
level, block, patch);
}
else
{
lkSparse4_gpu(prevPyr_[level], nextPyr_[level],
prevPts.ptr<float2>(), nextPts.ptr<float2>(), status.ptr(), level == 0 && err ? err->ptr<float>() : 0, prevPts.cols,
level, block, patch);
}
}
}
