static bool sumTemplate(InputArray _src, UMat & result)
{
int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
int wdepth = CV_32F, wtype = CV_MAKE_TYPE(wdepth, cn);
size_t wgs = ocl::Device::getDefault().maxWorkGroupSize();
int wgs2_aligned = 1;
while (wgs2_aligned < (int)wgs)
wgs2_aligned <<= 1;
wgs2_aligned >>= 1;
char cvt[40];
ocl::Kernel k("calcSum", ocl::imgproc::match_template_oclsrc,
format("-D CALC_SUM -D T=%s -D T1=%s -D WT=%s -D cn=%d -D convertToWT=%s -D WGS=%d -D WGS2_ALIGNED=%d",
ocl::typeToStr(type), ocl::typeToStr(depth), ocl::typeToStr(wtype), cn,
ocl::convertTypeStr(depth, wdepth, cn, cvt),
(int)wgs, wgs2_aligned));
if (k.empty())
return false;
UMat src = _src.getUMat();
result.create(1, 1, CV_32FC1);
ocl::KernelArg srcarg = ocl::KernelArg::ReadOnlyNoSize(src),
resarg = ocl::KernelArg::PtrWriteOnly(result);
k.args(srcarg, src.cols, (int)src.total(), resarg);
size_t globalsize = wgs;
return k.run(1, &globalsize, &wgs, false);
}
static bool ocl_calcAlmostDist2Weight(UMat & almostDist2Weight, int searchWindowSize, int templateWindowSize, FT h, int cn,
int & almostTemplateWindowSizeSqBinShift)
{
const int maxEstimateSumValue = searchWindowSize * searchWindowSize * 255;
int fixedPointMult = std::numeric_limits<int>::max() / maxEstimateSumValue;
int depth = DataType<FT>::depth;
bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
if (depth == CV_64F && !doubleSupport)
return false;
// precalc weight for every possible l2 dist between blocks
// additional optimization of precalced weights to replace division(averaging) by binary shift
CV_Assert(templateWindowSize <= 46340); // sqrt(INT_MAX)
int templateWindowSizeSq = templateWindowSize * templateWindowSize;
almostTemplateWindowSizeSqBinShift = getNearestPowerOf2(templateWindowSizeSq);
FT almostDist2ActualDistMultiplier = (FT)(1 << almostTemplateWindowSizeSqBinShift) / templateWindowSizeSq;
const FT WEIGHT_THRESHOLD = 1e-3f;
int maxDist = 255 * 255 * cn;
int almostMaxDist = (int)(maxDist / almostDist2ActualDistMultiplier + 1);
FT den = 1.0f / (h * h * cn);
almostDist2Weight.create(1, almostMaxDist, CV_32SC1);
ocl::Kernel k("calcAlmostDist2Weight", ocl::photo::nlmeans_oclsrc,
format("-D OP_CALC_WEIGHTS -D FT=%s%s", ocl::typeToStr(depth),
doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
if (k.empty())
return false;
k.args(ocl::KernelArg::PtrWriteOnly(almostDist2Weight), almostMaxDist,
almostDist2ActualDistMultiplier, fixedPointMult, den, WEIGHT_THRESHOLD);
size_t globalsize[1] = { almostMaxDist };
return k.run(1, globalsize, NULL, false);
}
void ConvolveBuf::create(Size image_size, Size templ_size)
{
result_size = Size(image_size.width - templ_size.width + 1,
image_size.height - templ_size.height + 1);
const double blockScale = 4.5;
const int minBlockSize = 256;
block_size.width = cvRound(result_size.width*blockScale);
block_size.width = std::max( block_size.width, minBlockSize - templ_size.width + 1 );
block_size.width = std::min( block_size.width, result_size.width );
block_size.height = cvRound(templ_size.height*blockScale);
block_size.height = std::max( block_size.height, minBlockSize - templ_size.height + 1 );
block_size.height = std::min( block_size.height, result_size.height );
dft_size.width = std::max(getOptimalDFTSize(block_size.width + templ_size.width - 1), 2);
dft_size.height = getOptimalDFTSize(block_size.height + templ_size.height - 1);
if( dft_size.width <= 0 || dft_size.height <= 0 )
CV_Error( CV_StsOutOfRange, "the input arrays are too big" );
// recompute block size
block_size.width = dft_size.width - templ_size.width + 1;
block_size.width = std::min( block_size.width, result_size.width);
block_size.height = dft_size.height - templ_size.height + 1;
block_size.height = std::min( block_size.height, result_size.height );
image_block.create(dft_size, CV_32F);
templ_block.create(dft_size, CV_32F);
result_data.create(dft_size, CV_32F);
image_spect.create(dft_size.height, dft_size.width / 2 + 1, CV_32FC2);
templ_spect.create(dft_size.height, dft_size.width / 2 + 1, CV_32FC2);
result_spect.create(dft_size.height, dft_size.width / 2 + 1, CV_32FC2);
// Use maximum result matrix block size for the estimated DFT block size
block_size.width = std::min(dft_size.width - templ_size.width + 1, result_size.width);
block_size.height = std::min(dft_size.height - templ_size.height + 1, result_size.height);
}
bool BTVL1_Base::ocl_process(InputArrayOfArrays _src, OutputArray _dst, InputArrayOfArrays _forwardMotions,
InputArrayOfArrays _backwardMotions, int baseIdx)
{
std::vector<UMat> & src = *(std::vector<UMat> *)_src.getObj(),
& forwardMotions = *(std::vector<UMat> *)_forwardMotions.getObj(),
& backwardMotions = *(std::vector<UMat> *)_backwardMotions.getObj();
// update blur filter and btv weights
if (!filter_ || blurKernelSize_ != curBlurKernelSize_ || blurSigma_ != curBlurSigma_ || src[0].type() != curSrcType_)
{
filter_ = createGaussianFilter(src[0].type(), Size(blurKernelSize_, blurKernelSize_), blurSigma_);
curBlurKernelSize_ = blurKernelSize_;
curBlurSigma_ = blurSigma_;
curSrcType_ = src[0].type();
}
if (btvWeights_.empty() || btvKernelSize_ != curBtvKernelSize_ || alpha_ != curAlpha_)
{
calcBtvWeights(btvKernelSize_, alpha_, btvWeights_);
Mat(btvWeights_, true).copyTo(ubtvWeights_);
curBtvKernelSize_ = btvKernelSize_;
curAlpha_ = alpha_;
}
// calc high res motions
calcRelativeMotions(forwardMotions, backwardMotions, ulowResForwardMotions_, ulowResBackwardMotions_, baseIdx, src[0].size());
upscaleMotions(ulowResForwardMotions_, uhighResForwardMotions_, scale_);
upscaleMotions(ulowResBackwardMotions_, uhighResBackwardMotions_, scale_);
uforwardMaps_.resize(uhighResForwardMotions_.size());
ubackwardMaps_.resize(uhighResForwardMotions_.size());
for (size_t i = 0; i < uhighResForwardMotions_.size(); ++i)
buildMotionMaps(uhighResForwardMotions_[i], uhighResBackwardMotions_[i], uforwardMaps_[i], ubackwardMaps_[i]);
// initial estimation
const Size lowResSize = src[0].size();
const Size highResSize(lowResSize.width * scale_, lowResSize.height * scale_);
resize(src[baseIdx], uhighRes_, highResSize, 0, 0, INTER_LINEAR); // TODO
// iterations
udiffTerm_.create(highResSize, uhighRes_.type());
ua_.create(highResSize, uhighRes_.type());
ub_.create(highResSize, uhighRes_.type());
uc_.create(lowResSize, uhighRes_.type());
for (int i = 0; i < iterations_; ++i)
{
udiffTerm_.setTo(Scalar::all(0));
for (size_t k = 0; k < src.size(); ++k)
{
// a = M * Ih
remap(uhighRes_, ua_, ubackwardMaps_[k], noArray(), INTER_NEAREST);
// b = HM * Ih
GaussianBlur(ua_, ub_, Size(blurKernelSize_, blurKernelSize_), blurSigma_);
// c = DHM * Ih
resize(ub_, uc_, lowResSize, 0, 0, INTER_NEAREST);
diffSign(src[k], uc_, uc_);
// a = Dt * diff
upscale(uc_, ua_, scale_);
// b = HtDt * diff
GaussianBlur(ua_, ub_, Size(blurKernelSize_, blurKernelSize_), blurSigma_);
// a = MtHtDt * diff
remap(ub_, ua_, uforwardMaps_[k], noArray(), INTER_NEAREST);
add(udiffTerm_, ua_, udiffTerm_);
}
if (lambda_ > 0)
{
calcBtvRegularization(uhighRes_, uregTerm_, btvKernelSize_, btvWeights_, ubtvWeights_);
addWeighted(udiffTerm_, 1.0, uregTerm_, -lambda_, 0.0, udiffTerm_);
}
addWeighted(uhighRes_, 1.0, udiffTerm_, tau_, 0.0, uhighRes_);
}
Rect inner(btvKernelSize_, btvKernelSize_, uhighRes_.cols - 2 * btvKernelSize_, uhighRes_.rows - 2 * btvKernelSize_);
uhighRes_(inner).copyTo(_dst);
return true;
}
bool SURF_OCL::computeDescriptors(const UMat &keypoints, OutputArray _descriptors)
{
int dsize = params->descriptorSize();
int nFeatures = keypoints.cols;
if (nFeatures == 0)
{
_descriptors.release();
return true;
}
_descriptors.create(nFeatures, dsize, CV_32F);
UMat descriptors;
if( _descriptors.isUMat() )
descriptors = _descriptors.getUMat();
else
descriptors.create(nFeatures, dsize, CV_32F);
ocl::Kernel kerCalcDesc, kerNormDesc;
if( dsize == 64 )
{
kerCalcDesc.create("SURF_computeDescriptors64", ocl::xfeatures2d::surf_oclsrc, kerOpts);
kerNormDesc.create("SURF_normalizeDescriptors64", ocl::xfeatures2d::surf_oclsrc, kerOpts);
}
else
{
CV_Assert(dsize == 128);
kerCalcDesc.create("SURF_computeDescriptors128", ocl::xfeatures2d::surf_oclsrc, kerOpts);
kerNormDesc.create("SURF_normalizeDescriptors128", ocl::xfeatures2d::surf_oclsrc, kerOpts);
}
size_t localThreads[] = {6, 6};
size_t globalThreads[] = {nFeatures*localThreads[0], localThreads[1]};
if(haveImageSupport)
{
kerCalcDesc.args(imgTex,
img_rows, img_cols,
ocl::KernelArg::ReadOnlyNoSize(keypoints),
ocl::KernelArg::WriteOnlyNoSize(descriptors));
}
else
{
kerCalcDesc.args(ocl::KernelArg::ReadOnlyNoSize(img),
img_rows, img_cols,
ocl::KernelArg::ReadOnlyNoSize(keypoints),
ocl::KernelArg::WriteOnlyNoSize(descriptors));
}
if(!kerCalcDesc.run(2, globalThreads, localThreads, true))
return false;
size_t localThreads_n[] = {dsize, 1};
size_t globalThreads_n[] = {nFeatures*localThreads_n[0], localThreads_n[1]};
globalThreads[0] = nFeatures * localThreads[0];
globalThreads[1] = localThreads[1];
bool ok = kerNormDesc.args(ocl::KernelArg::ReadWriteNoSize(descriptors)).
run(2, globalThreads_n, localThreads_n, true);
if(ok && !_descriptors.isUMat())
descriptors.copyTo(_descriptors);
return ok;
}
static bool ocl_Canny(InputArray _src, const UMat& dx_, const UMat& dy_, OutputArray _dst, float low_thresh, float high_thresh,
int aperture_size, bool L2gradient, int cn, const Size & size)
{
CV_INSTRUMENT_REGION_OPENCL()
UMat map;
const ocl::Device &dev = ocl::Device::getDefault();
int max_wg_size = (int)dev.maxWorkGroupSize();
int lSizeX = 32;
int lSizeY = max_wg_size / 32;
if (lSizeY == 0)
{
lSizeX = 16;
lSizeY = max_wg_size / 16;
}
if (lSizeY == 0)
{
lSizeY = 1;
}
if (aperture_size == 7)
{
low_thresh = low_thresh / 16.0f;
high_thresh = high_thresh / 16.0f;
}
if (L2gradient)
{
low_thresh = std::min(32767.0f, low_thresh);
high_thresh = std::min(32767.0f, high_thresh);
if (low_thresh > 0)
low_thresh *= low_thresh;
if (high_thresh > 0)
high_thresh *= high_thresh;
}
int low = cvFloor(low_thresh), high = cvFloor(high_thresh);
if (!useCustomDeriv &&
aperture_size == 3 && !_src.isSubmatrix())
{
/*
stage1_with_sobel:
Sobel operator
Calc magnitudes
Non maxima suppression
Double thresholding
*/
char cvt[40];
ocl::Kernel with_sobel("stage1_with_sobel", ocl::imgproc::canny_oclsrc,
format("-D WITH_SOBEL -D cn=%d -D TYPE=%s -D convert_floatN=%s -D floatN=%s -D GRP_SIZEX=%d -D GRP_SIZEY=%d%s",
cn, ocl::memopTypeToStr(_src.depth()),
ocl::convertTypeStr(_src.depth(), CV_32F, cn, cvt),
ocl::typeToStr(CV_MAKE_TYPE(CV_32F, cn)),
lSizeX, lSizeY,
L2gradient ? " -D L2GRAD" : ""));
if (with_sobel.empty())
return false;
UMat src = _src.getUMat();
map.create(size, CV_32S);
with_sobel.args(ocl::KernelArg::ReadOnly(src),
ocl::KernelArg::WriteOnlyNoSize(map),
(float) low, (float) high);
size_t globalsize[2] = { (size_t)size.width, (size_t)size.height },
localsize[2] = { (size_t)lSizeX, (size_t)lSizeY };
if (!with_sobel.run(2, globalsize, localsize, false))
return false;
}
else
{
/*
stage1_without_sobel:
Calc magnitudes
Non maxima suppression
Double thresholding
*/
double scale = 1.0;
if (aperture_size == 7)
{
scale = 1 / 16.0;
}
UMat dx, dy;
if (!useCustomDeriv)
{
Sobel(_src, dx, CV_16S, 1, 0, aperture_size, scale, 0, BORDER_REPLICATE);
Sobel(_src, dy, CV_16S, 0, 1, aperture_size, scale, 0, BORDER_REPLICATE);
}
else
{
dx = dx_;
dy = dy_;
}
ocl::Kernel without_sobel("stage1_without_sobel", ocl::imgproc::canny_oclsrc,
format("-D WITHOUT_SOBEL -D cn=%d -D GRP_SIZEX=%d -D GRP_SIZEY=%d%s",
cn, lSizeX, lSizeY, L2gradient ? " -D L2GRAD" : ""));
if (without_sobel.empty())
return false;
map.create(size, CV_32S);
without_sobel.args(ocl::KernelArg::ReadOnlyNoSize(dx), ocl::KernelArg::ReadOnlyNoSize(dy),
ocl::KernelArg::WriteOnly(map),
low, high);
size_t globalsize[2] = { (size_t)size.width, (size_t)size.height },
localsize[2] = { (size_t)lSizeX, (size_t)lSizeY };
if (!without_sobel.run(2, globalsize, localsize, false))
return false;
}
int PIX_PER_WI = 8;
/*
stage2:
hysteresis (add weak edges if they are connected with strong edges)
*/
int sizey = lSizeY / PIX_PER_WI;
if (sizey == 0)
sizey = 1;
size_t globalsize[2] = { (size_t)size.width, ((size_t)size.height + PIX_PER_WI - 1) / PIX_PER_WI }, localsize[2] = { (size_t)lSizeX, (size_t)sizey };
ocl::Kernel edgesHysteresis("stage2_hysteresis", ocl::imgproc::canny_oclsrc,
format("-D STAGE2 -D PIX_PER_WI=%d -D LOCAL_X=%d -D LOCAL_Y=%d",
PIX_PER_WI, lSizeX, sizey));
if (edgesHysteresis.empty())
return false;
edgesHysteresis.args(ocl::KernelArg::ReadWrite(map));
if (!edgesHysteresis.run(2, globalsize, localsize, false))
return false;
// get edges
ocl::Kernel getEdgesKernel("getEdges", ocl::imgproc::canny_oclsrc,
format("-D GET_EDGES -D PIX_PER_WI=%d", PIX_PER_WI));
if (getEdgesKernel.empty())
return false;
_dst.create(size, CV_8UC1);
UMat dst = _dst.getUMat();
getEdgesKernel.args(ocl::KernelArg::ReadOnly(map), ocl::KernelArg::WriteOnlyNoSize(dst));
return getEdgesKernel.run(2, globalsize, NULL, false);
}
