static void lkSparse_run(oclMat &I, oclMat &J,
const oclMat &prevPts, oclMat &nextPts, oclMat &status, oclMat& err, bool /*GET_MIN_EIGENVALS*/, int ptcount,
int level, /*dim3 block, */dim3 patch, Size winSize, int iters)
{
Context *clCxt = I.clCxt;
int elemCntPerRow = I.step / I.elemSize();
String kernelName = "lkSparse";
bool isImageSupported = support_image2d();
size_t localThreads[3] = { 8, isImageSupported ? 8 : 32, 1 };
size_t globalThreads[3] = { 8 * ptcount, isImageSupported ? 8 : 32, 1};
int cn = I.oclchannels();
char calcErr = level==0?1:0;
std::vector<std::pair<size_t , const void *> > args;
cl_mem ITex = isImageSupported ? bindTexture(I) : (cl_mem)I.data;
cl_mem JTex = isImageSupported ? bindTexture(J) : (cl_mem)J.data;
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&ITex ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&JTex ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&prevPts.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&prevPts.step ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&nextPts.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&nextPts.step ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&status.data ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&err.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&level ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&I.rows ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&I.cols ));
if (!isImageSupported)
args.push_back( std::make_pair( sizeof(cl_int), (void *)&elemCntPerRow ) );
args.push_back( std::make_pair( sizeof(cl_int), (void *)&patch.x ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&patch.y ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&cn ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&winSize.width ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&winSize.height ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&iters ));
args.push_back( std::make_pair( sizeof(cl_char), (void *)&calcErr ));
if(isImageSupported)
{
std::stringstream idxStr;
idxStr << kernelName.c_str() << "_C" << I.oclchannels() << "_D" << I.depth();
cl_kernel kernel = openCLGetKernelFromSource(clCxt, &pyrlk, idxStr.str().c_str());
int wave_size = (int)queryWaveFrontSize(kernel);
openCLSafeCall(clReleaseKernel(kernel));
static char opt[32] = {0};
sprintf(opt, " -D WAVE_SIZE=%d", wave_size);
openCLExecuteKernel2(clCxt, &pyrlk, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), opt, CLFLUSH);
releaseTexture(ITex);
releaseTexture(JTex);
}
else
openCLExecuteKernel2(clCxt, &pyrlk_no_image, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), CLFLUSH);
}
static void convert_C3C4(const cl_mem &src, oclMat &dst)
{
Context *clCxt = dst.clCxt;
int pixel_end = dst.wholecols * dst.wholerows - 1;
int dstStep_in_pixel = dst.step1() / dst.oclchannels();
const char * const typeMap[] = { "uchar", "char", "ushort", "short", "int", "float", "double" };
std::string buildOptions = format("-D GENTYPE4=%s4", typeMap[dst.depth()]);
std::vector< std::pair<size_t, const void *> > args;
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&src));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&dst.data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&dst.wholecols));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&dst.wholerows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&dstStep_in_pixel));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&pixel_end));
size_t globalThreads[3] = { divUp(dst.wholecols * dst.wholerows, 4), 1, 1 };
#ifdef ANDROID
openCLExecuteKernel(clCxt, &convertC3C4, "convertC3C4", globalThreads, NULL,
args, -1, -1, buildOptions.c_str());
#else
size_t localThreads[3] = { 256, 1, 1 };
openCLExecuteKernel(clCxt, &convertC3C4, "convertC3C4", globalThreads, localThreads,
args, -1, -1, buildOptions.c_str());
#endif
}
///////////////////////////////////////////////////////////////////////////
//////////////////////////////// ConvertTo ////////////////////////////////
///////////////////////////////////////////////////////////////////////////
static void convert_run_cus(const oclMat &src, oclMat &dst, double alpha, double beta)
{
std::string kernelName = "convert_to_S";
std::stringstream idxStr;
idxStr << src.depth();
kernelName += idxStr.str();
float alpha_f = (float)alpha, beta_f = (float)beta;
CV_DbgAssert(src.rows == dst.rows && src.cols == dst.cols);
std::vector<std::pair<size_t , const void *> > args;
size_t localThreads[3] = {16, 16, 1};
size_t globalThreads[3];
globalThreads[0] = (dst.cols + localThreads[0] - 1) / localThreads[0] * localThreads[0];
globalThreads[1] = (dst.rows + localThreads[1] - 1) / localThreads[1] * localThreads[1];
globalThreads[2] = 1;
int dststep_in_pixel = dst.step / dst.elemSize(), dstoffset_in_pixel = dst.offset / dst.elemSize();
int srcstep_in_pixel = src.step / src.elemSize(), srcoffset_in_pixel = src.offset / src.elemSize();
if(dst.type() == CV_8UC1)
{
globalThreads[0] = ((dst.cols + 4) / 4 + localThreads[0]) / localThreads[0] * localThreads[0];
}
args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&src.data ));
args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&dst.data ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&src.cols ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&src.rows ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&srcstep_in_pixel ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&srcoffset_in_pixel ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dststep_in_pixel ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dstoffset_in_pixel ));
args.push_back( std::make_pair( sizeof(cl_float) , (void *)&alpha_f ));
args.push_back( std::make_pair( sizeof(cl_float) , (void *)&beta_f ));
openCLExecuteKernel2(dst.clCxt , &operator_convertTo, kernelName, globalThreads,
localThreads, args, dst.oclchannels(), dst.depth(), CLFLUSH);
}
void convertTo( const oclMat &src, oclMat &dst, int rtype, double alpha, double beta )
{
//cout << "cv::ocl::oclMat::convertTo()" << endl;
bool noScale = fabs(alpha - 1) < std::numeric_limits<double>::epsilon()
&& fabs(beta) < std::numeric_limits<double>::epsilon();
if( rtype < 0 )
rtype = src.type();
else
rtype = CV_MAKETYPE(CV_MAT_DEPTH(rtype), src.oclchannels());
int sdepth = src.depth(), ddepth = CV_MAT_DEPTH(rtype);
if( sdepth == ddepth && noScale )
{
src.copyTo(dst);
return;
}
oclMat temp;
const oclMat *psrc = &src;
if( sdepth != ddepth && psrc == &dst )
psrc = &(temp = src);
dst.create( src.size(), rtype );
convert_run_cus(*psrc, dst, alpha, beta);
}
void cv::ocl::blendLinear(const oclMat &img1, const oclMat &img2, const oclMat &weights1, const oclMat &weights2,
oclMat &result)
{
cv::ocl::Context *ctx = img1.clCxt;
assert(ctx == img2.clCxt && ctx == weights1.clCxt && ctx == weights2.clCxt);
int channels = img1.oclchannels();
int depth = img1.depth();
int rows = img1.rows;
int cols = img1.cols;
int istep = img1.step1();
int wstep = weights1.step1();
size_t globalSize[] = {cols * channels / 4, rows, 1};
size_t localSize[] = {256, 1, 1};
vector< pair<size_t, const void *> > args;
if(globalSize[0] != 0)
{
args.push_back( make_pair( sizeof(cl_mem), (void *)&result.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&img1.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&img2.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&weights1.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&weights2.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&rows ));
args.push_back( make_pair( sizeof(cl_int), (void *)&cols ));
args.push_back( make_pair( sizeof(cl_int), (void *)&istep ));
args.push_back( make_pair( sizeof(cl_int), (void *)&wstep ));
std::string kernelName = "BlendLinear";
openCLExecuteKernel(ctx, &blend_linear, kernelName, globalSize, localSize, args, channels, depth);
}
}
static void lkSparse_run(oclMat &I, oclMat &J,
const oclMat &prevPts, oclMat &nextPts, oclMat &status, oclMat& err, bool /*GET_MIN_EIGENVALS*/, int ptcount,
int level, /*dim3 block, */dim3 patch, Size winSize, int iters)
{
Context *clCxt = I.clCxt;
int elemCntPerRow = I.step / I.elemSize();
std::string kernelName = "lkSparse";
size_t localThreads[3] = { 8, 8, 1 };
size_t globalThreads[3] = { 8 * ptcount, 8, 1};
int cn = I.oclchannels();
char calcErr;
if (level == 0)
{
calcErr = 1;
}
else
{
calcErr = 0;
}
std::vector<std::pair<size_t , const void *> > args;
cl_mem ITex = bindTexture(I);
cl_mem JTex = bindTexture(J);
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&ITex ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&JTex ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&prevPts.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&prevPts.step ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&nextPts.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&nextPts.step ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&status.data ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&err.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&level ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&I.rows ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&I.cols ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&patch.x ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&patch.y ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&cn ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&winSize.width ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&winSize.height ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&iters ));
args.push_back( std::make_pair( sizeof(cl_char), (void *)&calcErr ));
try
{
openCLExecuteKernel2(clCxt, &pyrlk, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), CLFLUSH);
}
catch(Exception&)
{
printf("Warning: The image2d_t is not supported by the device. Using alternative method!\n");
releaseTexture(ITex);
releaseTexture(JTex);
ITex = (cl_mem)I.data;
JTex = (cl_mem)J.data;
localThreads[1] = globalThreads[1] = 32;
args.insert( args.begin()+11, std::make_pair( sizeof(cl_int), (void *)&elemCntPerRow ) );
openCLExecuteKernel2(clCxt, &pyrlk_no_image, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), CLFLUSH);
}
}
void cv::ocl::MOG2::getBackgroundImage(oclMat& backgroundImage) const
{
using namespace cv::ocl::device::mog;
backgroundImage.create(frameSize_, frameType_);
cv::ocl::device::mog::getBackgroundImage2_ocl(backgroundImage.oclchannels(), bgmodelUsedModes_, weight_, mean_, backgroundImage, nmixtures_);
}
static void lkSparse_run(oclMat &I, oclMat &J,
const oclMat &prevPts, oclMat &nextPts, oclMat &status, oclMat& err, bool /*GET_MIN_EIGENVALS*/, int ptcount,
int level, /*dim3 block, */dim3 patch, Size winSize, int iters)
{
Context *clCxt = I.clCxt;
int elemCntPerRow = I.step / I.elemSize();
String kernelName = "lkSparse";
bool isImageSupported = support_image2d();
size_t localThreads[3] = { 8, isImageSupported ? 8 : 32, 1 };
size_t globalThreads[3] = { 8 * ptcount, isImageSupported ? 8 : 32, 1};
int cn = I.oclchannels();
char calcErr;
if (level == 0)
{
calcErr = 1;
}
else
{
calcErr = 0;
}
std::vector<std::pair<size_t , const void *> > args;
cl_mem ITex = isImageSupported ? bindTexture(I) : (cl_mem)I.data;
cl_mem JTex = isImageSupported ? bindTexture(J) : (cl_mem)J.data;
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&ITex ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&JTex ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&prevPts.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&prevPts.step ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&nextPts.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&nextPts.step ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&status.data ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&err.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&level ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&I.rows ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&I.cols ));
if (!isImageSupported)
args.push_back( std::make_pair( sizeof(cl_int), (void *)&elemCntPerRow ) );
args.push_back( std::make_pair( sizeof(cl_int), (void *)&patch.x ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&patch.y ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&cn ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&winSize.width ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&winSize.height ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&iters ));
args.push_back( std::make_pair( sizeof(cl_char), (void *)&calcErr ));
if(isImageSupported)
{
openCLExecuteKernel2(clCxt, &pyrlk, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), CLFLUSH);
releaseTexture(ITex);
releaseTexture(JTex);
}
else
{
openCLExecuteKernel2(clCxt, &pyrlk_no_image, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), CLFLUSH);
}
}
void cv::ocl::distanceToCenters(const oclMat &src, const oclMat ¢ers, Mat &dists, Mat &labels, int distType)
{
CV_Assert(src.cols * src.channels() == centers.cols * centers.channels());
CV_Assert(src.depth() == CV_32F && centers.depth() == CV_32F);
CV_Assert(distType == NORM_L1 || distType == NORM_L2SQR);
dists.create(src.rows, 1, CV_32FC1);
labels.create(src.rows, 1, CV_32SC1);
std::stringstream build_opt_ss;
build_opt_ss << (distType == NORM_L1 ? "-D L1_DIST" : "-D L2SQR_DIST");
int src_step = src.step / src.elemSize1();
int centers_step = centers.step / centers.elemSize1();
int feature_width = centers.cols * centers.oclchannels();
int src_offset = src.offset / src.elemSize1();
int centers_offset = centers.offset / centers.elemSize1();
int all_dist_count = src.rows * centers.rows;
oclMat all_dist(1, all_dist_count, CV_32FC1);
vector<pair<size_t, const void *> > args;
args.push_back(make_pair(sizeof(cl_mem), (void *)&src.data));
args.push_back(make_pair(sizeof(cl_mem), (void *)¢ers.data));
args.push_back(make_pair(sizeof(cl_mem), (void *)&all_dist.data));
args.push_back(make_pair(sizeof(cl_int), (void *)&feature_width));
args.push_back(make_pair(sizeof(cl_int), (void *)&src_step));
args.push_back(make_pair(sizeof(cl_int), (void *)¢ers_step));
args.push_back(make_pair(sizeof(cl_int), (void *)&src.rows));
args.push_back(make_pair(sizeof(cl_int), (void *)¢ers.rows));
args.push_back(make_pair(sizeof(cl_int), (void *)&src_offset));
args.push_back(make_pair(sizeof(cl_int), (void *)¢ers_offset));
size_t globalThreads[3] = { all_dist_count, 1, 1 };
openCLExecuteKernel(Context::getContext(), &kmeans_kernel,
"distanceToCenters", globalThreads, NULL, args, -1, -1, build_opt_ss.str().c_str());
Mat all_dist_cpu;
all_dist.download(all_dist_cpu);
for (int i = 0; i < src.rows; ++i)
{
Point p;
double minVal;
Rect roi(i * centers.rows, 0, centers.rows, 1);
Mat hdr(all_dist_cpu, roi);
cv::minMaxLoc(hdr, &minVal, NULL, &p);
dists.at<float>(i, 0) = static_cast<float>(minVal);
labels.at<int>(i, 0) = p.x;
}
}
void cv::ocl::MOG2::operator()(const oclMat& frame, oclMat& fgmask, float learningRate)
{
using namespace cv::ocl::device::mog;
int ch = frame.oclchannels();
int work_ch = ch;
if (nframes_ == 0 || learningRate >= 1.0f || frame.size() != frameSize_ || work_ch != mean_.oclchannels())
initialize(frame.size(), frame.type());
fgmask.create(frameSize_, CV_8UC1);
fgmask.setTo(cv::Scalar::all(0));
++nframes_;
learningRate = learningRate >= 0.0f && nframes_ > 1 ? learningRate : 1.0f / std::min(2 * nframes_, history);
CV_Assert(learningRate >= 0.0f);
mog2_ocl(frame, frame.oclchannels(), fgmask, bgmodelUsedModes_, weight_, variance_, mean_, learningRate, -learningRate * fCT, bShadowDetection, nmixtures_);
}
static void set_to_withoutmask_run(const oclMat &dst, const Scalar &scalar, String kernelName)
{
std::vector<std::pair<size_t , const void *> > args;
size_t localThreads[3] = {16, 16, 1};
size_t globalThreads[3] = { dst.cols, dst.rows, 1 };
int step_in_pixel = dst.step / dst.elemSize(), offset_in_pixel = dst.offset / dst.elemSize();
if (dst.type() == CV_8UC1)
globalThreads[0] = ((dst.cols + 4) / 4 + localThreads[0] - 1) / localThreads[0] * localThreads[0];
const char * const typeMap[] = { "uchar", "char", "ushort", "short", "int", "float", "double" };
const char channelMap[] = { ' ', ' ', '2', '4', '4' };
std::string buildOptions = format("-D GENTYPE=%s%c", typeMap[dst.depth()], channelMap[dst.channels()]);
Mat mat(1, 1, dst.type(), scalar);
#ifdef CL_VERSION_1_2
// this enables backwards portability to
// run on OpenCL 1.1 platform if library binaries are compiled with OpenCL 1.2 support
if (Context::getContext()->supportsFeature(Context::CL_VER_1_2) &&
dst.offset == 0 && dst.cols == dst.wholecols)
{
const int sizeofMap[][7] =
{
{ sizeof(cl_uchar) , sizeof(cl_char) , sizeof(cl_ushort) , sizeof(cl_short) , sizeof(cl_int) , sizeof(cl_float) , sizeof(cl_double) },
{ sizeof(cl_uchar2), sizeof(cl_char2), sizeof(cl_ushort2), sizeof(cl_short2), sizeof(cl_int2), sizeof(cl_float2), sizeof(cl_double2) },
{ 0 , 0 , 0 , 0 , 0 , 0 , 0 },
{ sizeof(cl_uchar4), sizeof(cl_char4), sizeof(cl_ushort4), sizeof(cl_short4), sizeof(cl_int4), sizeof(cl_float4), sizeof(cl_double4) },
};
int sizeofGeneric = sizeofMap[dst.oclchannels() - 1][dst.depth()];
clEnqueueFillBuffer((cl_command_queue)dst.clCxt->oclCommandQueue(),
(cl_mem)dst.data, (void*)mat.data, sizeofGeneric,
0, dst.step * dst.rows, 0, NULL, NULL);
}
else
#endif
{
oclMat m(mat);
args.push_back( std::make_pair( sizeof(cl_mem) , (void*)&m.data ));
args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&dst.data ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dst.cols ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dst.rows ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&step_in_pixel ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&offset_in_pixel ));
openCLExecuteKernel(dst.clCxt , &operator_setTo, kernelName, globalThreads,
localThreads, args, -1, -1, buildOptions.c_str());
}
}
void cv::ocl::blendLinear(const oclMat &src1, const oclMat &src2, const oclMat &weights1, const oclMat &weights2,
oclMat &dst)
{
CV_Assert(src1.depth() <= CV_32F);
CV_Assert(src1.size() == src2.size() && src1.type() == src2.type());
CV_Assert(weights1.size() == weights2.size() && weights1.size() == src1.size() &&
weights1.type() == CV_32FC1 && weights2.type() == CV_32FC1);
dst.create(src1.size(), src1.type());
size_t globalSize[] = { (size_t)dst.cols, (size_t)dst.rows, 1};
size_t localSize[] = { 16, 16, 1 };
int depth = dst.depth(), ocn = dst.oclchannels();
int src1_step = src1.step / src1.elemSize(), src1_offset = src1.offset / src1.elemSize();
int src2_step = src2.step / src2.elemSize(), src2_offset = src2.offset / src2.elemSize();
int weight1_step = weights1.step / weights1.elemSize(), weight1_offset = weights1.offset / weights1.elemSize();
int weight2_step = weights2.step / weights2.elemSize(), weight2_offset = weights2.offset / weights2.elemSize();
int dst_step = dst.step / dst.elemSize(), dst_offset = dst.offset / dst.elemSize();
const char * const channelMap[] = { "", "", "2", "4", "4" };
const char * const typeMap[] = { "uchar", "char", "ushort", "short", "int", "float", "double" };
std::string buildOptions = format("-D T=%s%s -D convertToT=convert_%s%s%s -D FT=float%s -D convertToFT=convert_float%s",
typeMap[depth], channelMap[ocn], typeMap[depth], channelMap[ocn],
depth >= CV_32S ? "" : "_sat_rte", channelMap[ocn], channelMap[ocn]);
vector< pair<size_t, const void *> > args;
args.push_back( make_pair( sizeof(cl_mem), (void *)&src1.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&src1_offset ));
args.push_back( make_pair( sizeof(cl_int), (void *)&src1_step ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&src2.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&src2_offset ));
args.push_back( make_pair( sizeof(cl_int), (void *)&src2_step ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&weights1.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&weight1_offset ));
args.push_back( make_pair( sizeof(cl_int), (void *)&weight1_step ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&weights2.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&weight2_offset ));
args.push_back( make_pair( sizeof(cl_int), (void *)&weight2_step ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&dst.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&dst_offset ));
args.push_back( make_pair( sizeof(cl_int), (void *)&dst_step ));
args.push_back( make_pair( sizeof(cl_int), (void *)&dst.rows ));
args.push_back( make_pair( sizeof(cl_int), (void *)&dst.cols ));
openCLExecuteKernel(src1.clCxt, &blend_linear, "blendLinear", globalSize, localSize, args,
-1, -1, buildOptions.c_str());
}
////////////////////////////////////////////////////////////////////////
// convert_C4C3
static void convert_C4C3(const oclMat &src, cl_mem &dst)
{
int srcStep_in_pixel = src.step1() / src.oclchannels();
int pixel_end = src.wholecols * src.wholerows - 1;
Context *clCxt = src.clCxt;
string kernelName = "convertC4C3";
char compile_option[32];
switch(src.depth())
{
case 0:
sprintf(compile_option, "-D GENTYPE4=uchar4");
break;
case 1:
sprintf(compile_option, "-D GENTYPE4=char4");
break;
case 2:
sprintf(compile_option, "-D GENTYPE4=ushort4");
break;
case 3:
sprintf(compile_option, "-D GENTYPE4=short4");
break;
case 4:
sprintf(compile_option, "-D GENTYPE4=int4");
break;
case 5:
sprintf(compile_option, "-D GENTYPE4=float4");
break;
case 6:
sprintf(compile_option, "-D GENTYPE4=double4");
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unknown depth");
}
vector< pair<size_t, const void *> > args;
args.push_back( make_pair( sizeof(cl_mem), (void *)&src.data));
args.push_back( make_pair( sizeof(cl_mem), (void *)&dst));
args.push_back( make_pair( sizeof(cl_int), (void *)&src.wholecols));
args.push_back( make_pair( sizeof(cl_int), (void *)&src.wholerows));
args.push_back( make_pair( sizeof(cl_int), (void *)&srcStep_in_pixel));
args.push_back( make_pair( sizeof(cl_int), (void *)&pixel_end));
size_t globalThreads[3] = {((src.wholecols * src.wholerows + 3) / 4 + 255) / 256 * 256, 1, 1};
size_t localThreads[3] = {256, 1, 1};
openCLExecuteKernel(clCxt, &convertC3C4, kernelName, globalThreads, localThreads, args, -1, -1, compile_option);
}
static void split(const oclMat &mat_src, oclMat *mat_dst)
{
CV_Assert(mat_dst);
int depth = mat_src.depth();
int num_channels = mat_src.oclchannels();
Size size = mat_src.size();
if(num_channels == 1)
{
mat_src.copyTo(mat_dst[0]);
return;
}
int i;
for(i = 0; i < num_channels; i++)
mat_dst[i].create(size, CV_MAKETYPE(depth, 1));
split_vector_run(mat_src, mat_dst);
}
///////////////////////////////////////////////////////////////////////////
////////////////////////////////// CopyTo /////////////////////////////////
///////////////////////////////////////////////////////////////////////////
static void copy_to_with_mask(const oclMat &src, oclMat &dst, const oclMat &mask, string kernelName)
{
CV_DbgAssert( dst.rows == mask.rows && dst.cols == mask.cols &&
src.rows == dst.rows && src.cols == dst.cols
&& mask.type() == CV_8UC1);
vector<pair<size_t , const void *> > args;
std::string string_types[4][7] = {{"uchar", "char", "ushort", "short", "int", "float", "double"},
{"uchar2", "char2", "ushort2", "short2", "int2", "float2", "double2"},
{"uchar3", "char3", "ushort3", "short3", "int3", "float3", "double3"},
{"uchar4", "char4", "ushort4", "short4", "int4", "float4", "double4"}
};
char compile_option[32];
sprintf(compile_option, "-D GENTYPE=%s", string_types[dst.oclchannels() - 1][dst.depth()].c_str());
size_t localThreads[3] = {16, 16, 1};
size_t globalThreads[3];
globalThreads[0] = divUp(dst.cols, localThreads[0]) * localThreads[0];
globalThreads[1] = divUp(dst.rows, localThreads[1]) * localThreads[1];
globalThreads[2] = 1;
int dststep_in_pixel = dst.step / dst.elemSize(), dstoffset_in_pixel = dst.offset / dst.elemSize();
int srcstep_in_pixel = src.step / src.elemSize(), srcoffset_in_pixel = src.offset / src.elemSize();
args.push_back( make_pair( sizeof(cl_mem) , (void *)&src.data ));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst.data ));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&mask.data ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src.cols ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src.rows ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&srcstep_in_pixel ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&srcoffset_in_pixel ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dststep_in_pixel ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dstoffset_in_pixel ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&mask.step ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&mask.offset ));
openCLExecuteKernel(dst.clCxt , &operator_copyToM, kernelName, globalThreads,
localThreads, args, -1, -1, compile_option);
}
static void convert_run(const oclMat &src, oclMat &dst, double alpha, double beta)
{
String kernelName = "convert_to";
float alpha_f = alpha, beta_f = beta;
int sdepth = src.depth(), ddepth = dst.depth();
int sstep1 = (int)src.step1(), dstep1 = (int)dst.step1();
int cols1 = src.cols * src.oclchannels();
char buildOptions[150], convertString[50];
const char * typeMap[] = { "uchar", "char", "ushort", "short", "int", "float", "double" };
sprintf(convertString, "convert_%s_sat_rte", typeMap[ddepth]);
sprintf(buildOptions, "-D srcT=%s -D dstT=%s -D convertToDstType=%s", typeMap[sdepth],
typeMap[ddepth], CV_32F == ddepth || ddepth == CV_64F ? "" : convertString);
CV_DbgAssert(src.rows == dst.rows && src.cols == dst.cols);
std::vector<std::pair<size_t , const void *> > args;
size_t localThreads[3] = { 16, 16, 1 };
size_t globalThreads[3] = { divUp(cols1, localThreads[0]) * localThreads[0],
divUp(dst.rows, localThreads[1]) * localThreads[1], 1
};
int doffset1 = dst.offset / dst.elemSize1();
int soffset1 = src.offset / src.elemSize1();
args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&src.data ));
args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&dst.data ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&cols1 ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&src.rows ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&sstep1 ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&soffset1 ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dstep1 ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&doffset1 ));
args.push_back( std::make_pair( sizeof(cl_float) , (void *)&alpha_f ));
args.push_back( std::make_pair( sizeof(cl_float) , (void *)&beta_f ));
openCLExecuteKernel(dst.clCxt , &operator_convertTo, kernelName, globalThreads,
localThreads, args, -1, -1, buildOptions);
}
static void convert_C4C3(const oclMat &src, cl_mem &dst)
{
int srcStep_in_pixel = src.step1() / src.oclchannels();
int pixel_end = src.wholecols * src.wholerows - 1;
Context *clCxt = src.clCxt;
const char * const typeMap[] = { "uchar", "char", "ushort", "short", "int", "float", "double" };
std::string buildOptions = format("-D GENTYPE4=%s4", typeMap[src.depth()]);
std::vector< std::pair<size_t, const void *> > args;
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&src.data));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&dst));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&src.wholecols));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&src.wholerows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&srcStep_in_pixel));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&pixel_end));
size_t globalThreads[3] = { divUp(src.wholecols * src.wholerows, 4), 1, 1};
size_t localThreads[3] = { 256, 1, 1 };
openCLExecuteKernel(clCxt, &convertC3C4, "convertC4C3", globalThreads, localThreads, args, -1, -1, buildOptions.c_str());
}
void cv::ocl::PyrLKOpticalFlow::sparse(const oclMat &prevImg, const oclMat &nextImg, const oclMat &prevPts, oclMat &nextPts, oclMat &status, oclMat *err)
{
if (prevPts.empty())
{
nextPts.release();
status.release();
//if (err) err->release();
return;
}
derivLambda = std::min(std::max(derivLambda, 0.0), 1.0);
iters = std::min(std::max(iters, 0), 100);
const int cn = prevImg.oclchannels();
dim3 block, patch;
calcPatchSize(winSize, cn, block, patch, isDeviceArch11_);
CV_Assert(derivLambda >= 0);
CV_Assert(maxLevel >= 0 && winSize.width > 2 && winSize.height > 2);
CV_Assert(prevImg.size() == nextImg.size() && prevImg.type() == nextImg.type());
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);
oclMat temp1 = (useInitialFlow ? nextPts : prevPts).reshape(1);
oclMat temp2 = nextPts.reshape(1);
//oclMat scalar(temp1.rows, temp1.cols, temp1.type(), Scalar(1.0f / (1 << maxLevel) / 2.0f));
multiply_cus(temp1, temp2, 1.0f / (1 << maxLevel) / 2.0f);
//::multiply(temp1, 1.0f / (1 << maxLevel) / 2.0f, temp2);
ensureSizeIsEnough(1, prevPts.cols, CV_8UC1, status);
//status.setTo(Scalar::all(1));
setTo(status, Scalar::all(1));
bool errMat = false;
if (!err)
{
err = new oclMat(1, prevPts.cols, CV_32FC1);
errMat = true;
}
else
ensureSizeIsEnough(1, prevPts.cols, CV_32FC1, *err);
//ensureSizeIsEnough(1, prevPts.cols, CV_32FC1, err);
// build the image pyramids.
prevPyr_.resize(maxLevel + 1);
nextPyr_.resize(maxLevel + 1);
if (cn == 1 || cn == 4)
{
//prevImg.convertTo(prevPyr_[0], CV_32F);
//nextImg.convertTo(nextPyr_[0], CV_32F);
convertTo(prevImg, prevPyr_[0], CV_32F);
convertTo(nextImg, nextPyr_[0], CV_32F);
}
else
{
//oclMat buf_;
// cvtColor(prevImg, buf_, COLOR_BGR2BGRA);
// buf_.convertTo(prevPyr_[0], CV_32F);
// cvtColor(nextImg, buf_, COLOR_BGR2BGRA);
// buf_.convertTo(nextPyr_[0], CV_32F);
}
for (int level = 1; level <= maxLevel; ++level)
{
pyrDown_cus(prevPyr_[level - 1], prevPyr_[level]);
pyrDown_cus(nextPyr_[level - 1], nextPyr_[level]);
}
// dI/dx ~ Ix, dI/dy ~ Iy
for (int level = maxLevel; level >= 0; level--)
{
lkSparse_run(prevPyr_[level], nextPyr_[level],
prevPts, nextPts, status, *err, getMinEigenVals, prevPts.cols,
level, /*block, */patch, winSize, iters);
}
clFinish(prevImg.clCxt->impl->clCmdQueue);
if(errMat)
delete err;
}
static void split_vector_run(const oclMat &mat_src, oclMat *mat_dst)
{
if(!mat_src.clCxt->supportsFeature(FEATURE_CL_DOUBLE) && mat_src.type() == CV_64F)
{
CV_Error(Error::GpuNotSupported, "Selected device don't support double\r\n");
return;
}
Context *clCxt = mat_src.clCxt;
int channels = mat_src.oclchannels();
int depth = mat_src.depth();
String kernelName = "split_vector";
int vector_lengths[4][7] = {{0, 0, 0, 0, 0, 0, 0},
{4, 4, 2, 2, 1, 1, 1},
{4, 4, 2, 2 , 1, 1, 1},
{4, 4, 2, 2, 1, 1, 1}
};
size_t vector_length = vector_lengths[channels - 1][mat_dst[0].depth()];
int max_offset_cols = 0;
for(int i = 0; i < channels; i++)
{
int offset_cols = (mat_dst[i].offset / mat_dst[i].elemSize()) & (vector_length - 1);
if(max_offset_cols < offset_cols)
max_offset_cols = offset_cols;
}
int cols = vector_length == 1 ? divUp(mat_src.cols, vector_length)
: divUp(mat_src.cols + max_offset_cols, vector_length);
size_t localThreads[3] = { 64, 4, 1 };
size_t globalThreads[3] = { cols, mat_src.rows, 1 };
int dst_step1 = mat_dst[0].cols * mat_dst[0].elemSize();
std::vector<std::pair<size_t , const void *> > args;
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&mat_src.data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_src.step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_src.offset));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&mat_dst[0].data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_dst[0].step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_dst[0].offset));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&mat_dst[1].data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_dst[1].step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_dst[1].offset));
if(channels >= 3)
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&mat_dst[2].data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_dst[2].step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_dst[2].offset));
}
if(channels >= 4)
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&mat_dst[3].data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_dst[3].step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_dst[3].offset));
}
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_src.rows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&cols));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&dst_step1));
openCLExecuteKernel(clCxt, &split_mat, kernelName, globalThreads, localThreads, args, channels, depth);
}
static void pyrdown_run_cus(const oclMat &src, const oclMat &dst)
{
CV_Assert(src.type() == dst.type());
CV_Assert(src.depth() != CV_8S);
Context *clCxt = src.clCxt;
std::string kernelName = "pyrDown";
size_t localThreads[3] = { 256, 1, 1 };
size_t globalThreads[3] = { src.cols, dst.rows, 1};
std::vector<std::pair<size_t , const void *> > args;
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&src.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&src.step ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&src.rows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&src.cols));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&dst.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&dst.step ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&dst.cols));
openCLExecuteKernel2(clCxt, &pyr_down, kernelName, globalThreads, localThreads, args, src.oclchannels(), src.depth(), CLFLUSH);
}
///////////////////////////////////////////////////////////////////////////
//////////////////////////////// setTo ////////////////////////////////////
///////////////////////////////////////////////////////////////////////////
//oclMat &operator = (const Scalar &s)
//{
// //cout << "cv::ocl::oclMat::=" << endl;
// setTo(s);
// return *this;
//}
static void set_to_withoutmask_run_cus(const oclMat &dst, const Scalar &scalar, std::string kernelName)
{
std::vector<std::pair<size_t , const void *> > args;
size_t localThreads[3] = {16, 16, 1};
size_t globalThreads[3];
globalThreads[0] = (dst.cols + localThreads[0] - 1) / localThreads[0] * localThreads[0];
globalThreads[1] = (dst.rows + localThreads[1] - 1) / localThreads[1] * localThreads[1];
globalThreads[2] = 1;
int step_in_pixel = dst.step / dst.elemSize(), offset_in_pixel = dst.offset / dst.elemSize();
if(dst.type() == CV_8UC1)
{
globalThreads[0] = ((dst.cols + 4) / 4 + localThreads[0] - 1) / localThreads[0] * localThreads[0];
}
char compile_option[32];
union sc
{
cl_uchar4 uval;
cl_char4 cval;
cl_ushort4 usval;
cl_short4 shval;
cl_int4 ival;
cl_float4 fval;
cl_double4 dval;
} val;
switch(dst.depth())
{
case 0:
val.uval.s[0] = saturate_cast<uchar>(scalar.val[0]);
val.uval.s[1] = saturate_cast<uchar>(scalar.val[1]);
val.uval.s[2] = saturate_cast<uchar>(scalar.val[2]);
val.uval.s[3] = saturate_cast<uchar>(scalar.val[3]);
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=uchar");
args.push_back( std::make_pair( sizeof(cl_uchar) , (void *)&val.uval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=uchar4");
args.push_back( std::make_pair( sizeof(cl_uchar4) , (void *)&val.uval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case 1:
val.cval.s[0] = saturate_cast<char>(scalar.val[0]);
val.cval.s[1] = saturate_cast<char>(scalar.val[1]);
val.cval.s[2] = saturate_cast<char>(scalar.val[2]);
val.cval.s[3] = saturate_cast<char>(scalar.val[3]);
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=char");
args.push_back( std::make_pair( sizeof(cl_char) , (void *)&val.cval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=char4");
args.push_back( std::make_pair( sizeof(cl_char4) , (void *)&val.cval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case 2:
val.usval.s[0] = saturate_cast<ushort>(scalar.val[0]);
val.usval.s[1] = saturate_cast<ushort>(scalar.val[1]);
val.usval.s[2] = saturate_cast<ushort>(scalar.val[2]);
val.usval.s[3] = saturate_cast<ushort>(scalar.val[3]);
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=ushort");
args.push_back( std::make_pair( sizeof(cl_ushort) , (void *)&val.usval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=ushort4");
args.push_back( std::make_pair( sizeof(cl_ushort4) , (void *)&val.usval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case 3:
val.shval.s[0] = saturate_cast<short>(scalar.val[0]);
val.shval.s[1] = saturate_cast<short>(scalar.val[1]);
val.shval.s[2] = saturate_cast<short>(scalar.val[2]);
val.shval.s[3] = saturate_cast<short>(scalar.val[3]);
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=short");
args.push_back( std::make_pair( sizeof(cl_short) , (void *)&val.shval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=short4");
args.push_back( std::make_pair( sizeof(cl_short4) , (void *)&val.shval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case 4:
val.ival.s[0] = saturate_cast<int>(scalar.val[0]);
val.ival.s[1] = saturate_cast<int>(scalar.val[1]);
val.ival.s[2] = saturate_cast<int>(scalar.val[2]);
val.ival.s[3] = saturate_cast<int>(scalar.val[3]);
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=int");
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&val.ival.s[0] ));
break;
case 2:
sprintf(compile_option, "-D GENTYPE=int2");
cl_int2 i2val;
i2val.s[0] = val.ival.s[0];
i2val.s[1] = val.ival.s[1];
args.push_back( std::make_pair( sizeof(cl_int2) , (void *)&i2val ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=int4");
args.push_back( std::make_pair( sizeof(cl_int4) , (void *)&val.ival ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case 5:
val.fval.s[0] = (float)scalar.val[0];
val.fval.s[1] = (float)scalar.val[1];
val.fval.s[2] = (float)scalar.val[2];
val.fval.s[3] = (float)scalar.val[3];
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=float");
args.push_back( std::make_pair( sizeof(cl_float) , (void *)&val.fval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=float4");
args.push_back( std::make_pair( sizeof(cl_float4) , (void *)&val.fval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case 6:
val.dval.s[0] = scalar.val[0];
val.dval.s[1] = scalar.val[1];
val.dval.s[2] = scalar.val[2];
val.dval.s[3] = scalar.val[3];
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=double");
args.push_back( std::make_pair( sizeof(cl_double) , (void *)&val.dval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=double4");
args.push_back( std::make_pair( sizeof(cl_double4) , (void *)&val.dval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unknown depth");
}
#ifdef CL_VERSION_1_2
if(dst.offset == 0 && dst.cols == dst.wholecols)
{
clEnqueueFillBuffer(dst.clCxt->impl->clCmdQueue, (cl_mem)dst.data, args[0].second, args[0].first, 0, dst.step * dst.rows, 0, NULL, NULL);
}
else
{
args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&dst.data ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dst.cols ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dst.rows ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&step_in_pixel ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&offset_in_pixel));
openCLExecuteKernel2(dst.clCxt , &operator_setTo, kernelName, globalThreads,
localThreads, args, -1, -1, compile_option, CLFLUSH);
}
#else
args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&dst.data ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dst.cols ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dst.rows ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&step_in_pixel ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&offset_in_pixel));
openCLExecuteKernel2(dst.clCxt , &operator_setTo, kernelName, globalThreads,
localThreads, args, -1, -1, compile_option, CLFLUSH);
#endif
}
Moments ocl_moments(oclMat& src, bool binary) //for image
{
CV_Assert(src.oclchannels() == 1);
if(src.type() == CV_64FC1 && !Context::getContext()->supportsFeature(FEATURE_CL_DOUBLE))
{
CV_Error(CV_StsUnsupportedFormat, "Moments - double is not supported by your GPU!");
}
if(binary)
{
oclMat mask;
if(src.type() != CV_8UC1)
{
src.convertTo(mask, CV_8UC1);
}
oclMat src8u(src.size(), CV_8UC1);
src8u.setTo(Scalar(255), mask);
src = src8u;
}
const int TILE_SIZE = 256;
CvMoments mom;
memset(&mom, 0, sizeof(mom));
cv::Size size = src.size();
int blockx, blocky;
blockx = (size.width + TILE_SIZE - 1)/TILE_SIZE;
blocky = (size.height + TILE_SIZE - 1)/TILE_SIZE;
oclMat dst_m;
int tile_height = TILE_SIZE;
size_t localThreads[3] = {1, tile_height, 1};
size_t globalThreads[3] = {blockx, size.height, 1};
if(Context::getContext()->supportsFeature(FEATURE_CL_DOUBLE))
{
dst_m.create(blocky * 10, blockx, CV_64FC1);
}else
{
dst_m.create(blocky * 10, blockx, CV_32FC1);
}
int src_step = (int)(src.step/src.elemSize());
int dstm_step = (int)(dst_m.step/dst_m.elemSize());
std::vector<std::pair<size_t , const void *> > args,args_sum;
args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&src.data ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&src.rows ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&src.cols ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&src_step ));
args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&dst_m.data ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dst_m.cols ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dstm_step ));
int binary_;
if(binary)
binary_ = 1;
else
binary_ = 0;
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&binary_));
char builOption[128];
if(binary || src.type() == CV_8UC1)
{
snprintf(builOption, 128, "-D CV_8UC1");
}else if(src.type() == CV_16UC1)
{
snprintf(builOption, 128, "-D CV_16UC1");
}else if(src.type() == CV_16SC1)
{
snprintf(builOption, 128, "-D CV_16SC1");
}else if(src.type() == CV_32FC1)
{
snprintf(builOption, 128, "-D CV_32FC1");
}else if(src.type() == CV_64FC1)
{
snprintf(builOption, 128, "-D CV_64FC1");
}else
{
CV_Error( CV_StsUnsupportedFormat, "" );
}
openCLExecuteKernel(Context::getContext(), &moments, "CvMoments", globalThreads, localThreads, args, -1, -1, builOption);
Mat tmp(dst_m);
tmp.convertTo(tmp, CV_64FC1);
double tmp_m[10] = {0};
for(int j = 0; j < tmp.rows; j += 10)
{
for(int i = 0; i < tmp.cols; i++)
{
tmp_m[0] += tmp.at<double>(j, i);
tmp_m[1] += tmp.at<double>(j + 1, i);
tmp_m[2] += tmp.at<double>(j + 2, i);
tmp_m[3] += tmp.at<double>(j + 3, i);
tmp_m[4] += tmp.at<double>(j + 4, i);
tmp_m[5] += tmp.at<double>(j + 5, i);
tmp_m[6] += tmp.at<double>(j + 6, i);
tmp_m[7] += tmp.at<double>(j + 7, i);
tmp_m[8] += tmp.at<double>(j + 8, i);
tmp_m[9] += tmp.at<double>(j + 9, i);
}
}
mom.m00 = tmp_m[0];
mom.m10 = tmp_m[1];
mom.m01 = tmp_m[2];
mom.m20 = tmp_m[3];
mom.m11 = tmp_m[4];
mom.m02 = tmp_m[5];
mom.m30 = tmp_m[6];
mom.m21 = tmp_m[7];
mom.m12 = tmp_m[8];
mom.m03 = tmp_m[9];
icvCompleteMomentState( &mom );
return mom;
}
static void merge_vector_run(const oclMat *mat_src, size_t n, oclMat &mat_dst)
{
if(!mat_dst.clCxt->supportsFeature(FEATURE_CL_DOUBLE) && mat_dst.type() == CV_64F)
{
CV_Error(Error::GpuNotSupported, "Selected device don't support double\r\n");
return;
}
Context *clCxt = mat_dst.clCxt;
int channels = mat_dst.oclchannels();
int depth = mat_dst.depth();
String kernelName = "merge_vector";
int vector_lengths[4][7] = {{0, 0, 0, 0, 0, 0, 0},
{2, 2, 1, 1, 1, 1, 1},
{4, 4, 2, 2 , 1, 1, 1},
{1, 1, 1, 1, 1, 1, 1}
};
size_t vector_length = vector_lengths[channels - 1][depth];
int offset_cols = (mat_dst.offset / mat_dst.elemSize()) & (vector_length - 1);
int cols = divUp(mat_dst.cols + offset_cols, vector_length);
size_t localThreads[3] = { 64, 4, 1 };
size_t globalThreads[3] = { cols, mat_dst.rows, 1 };
int dst_step1 = mat_dst.cols * mat_dst.elemSize();
std::vector<std::pair<size_t , const void *> > args;
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&mat_dst.data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_dst.step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_dst.offset));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&mat_src[0].data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_src[0].step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_src[0].offset));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&mat_src[1].data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_src[1].step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_src[1].offset));
if(channels == 4)
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&mat_src[2].data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_src[2].step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_src[2].offset));
if(n == 3)
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&mat_src[2].data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_src[2].step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_src[2].offset));
}
else if( n == 4)
{
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&mat_src[3].data));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_src[3].step));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_src[3].offset));
}
}
args.push_back( std::make_pair( sizeof(cl_int), (void *)&mat_dst.rows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&cols));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&dst_step1));
openCLExecuteKernel(clCxt, &merge_mat, kernelName, globalThreads, localThreads, args, channels, depth);
}
static void lkSparse_run(oclMat &I, oclMat &J,
const oclMat &prevPts, oclMat &nextPts, oclMat &status, oclMat& err, bool /*GET_MIN_EIGENVALS*/, int ptcount,
int level, /*dim3 block, */dim3 patch, Size winSize, int iters)
{
Context *clCxt = I.clCxt;
char platform[256] = {0};
cl_platform_id pid;
clGetDeviceInfo(clCxt->impl->devices, CL_DEVICE_PLATFORM, sizeof(pid), &pid, NULL);
clGetPlatformInfo(pid, CL_PLATFORM_NAME, 256, platform, NULL);
std::string namestr = platform;
bool isImageSupported = true;
if(namestr.find("NVIDIA")!=string::npos || namestr.find("Intel")!=string::npos)
isImageSupported = false;
int elemCntPerRow = I.step / I.elemSize();
string kernelName = "lkSparse";
size_t localThreads[3] = { 8, isImageSupported?8:32, 1 };
size_t globalThreads[3] = { 8 * ptcount, isImageSupported?8:32, 1};
int cn = I.oclchannels();
char calcErr;
if (level == 0)
{
calcErr = 1;
}
else
{
calcErr = 0;
}
vector<pair<size_t , const void *> > args;
cl_mem ITex;
cl_mem JTex;
if (isImageSupported)
{
ITex = bindTexture(I, I.depth(), cn);
JTex = bindTexture(J, J.depth(), cn);
}
else
{
ITex = (cl_mem)I.data;
JTex = (cl_mem)J.data;
}
args.push_back( make_pair( sizeof(cl_mem), (void *)&ITex ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&JTex ));
//cl_mem clmD = clCreateBuffer(clCxt, CL_MEM_READ_WRITE, ptcount * sizeof(float), NULL, NULL);
args.push_back( make_pair( sizeof(cl_mem), (void *)&prevPts.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&prevPts.step ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&nextPts.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&nextPts.step ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&status.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&err.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&level ));
args.push_back( make_pair( sizeof(cl_int), (void *)&I.rows ));
args.push_back( make_pair( sizeof(cl_int), (void *)&I.cols ));
if (!isImageSupported)
{
args.push_back( make_pair( sizeof(cl_int), (void *)&elemCntPerRow ) );
}
args.push_back( make_pair( sizeof(cl_int), (void *)&patch.x ));
args.push_back( make_pair( sizeof(cl_int), (void *)&patch.y ));
args.push_back( make_pair( sizeof(cl_int), (void *)&cn ));
args.push_back( make_pair( sizeof(cl_int), (void *)&winSize.width ));
args.push_back( make_pair( sizeof(cl_int), (void *)&winSize.height ));
args.push_back( make_pair( sizeof(cl_int), (void *)&iters ));
args.push_back( make_pair( sizeof(cl_char), (void *)&calcErr ));
//args.push_back( make_pair( sizeof(cl_char), (void *)&GET_MIN_EIGENVALS ));
if (isImageSupported)
{
openCLExecuteKernel2(clCxt, &pyrlk, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), CLFLUSH);
releaseTexture(ITex);
releaseTexture(JTex);
}
else
{
//printf("Warning: The image2d_t is not supported by the device. Using alternative method!\n");
openCLExecuteKernel2(clCxt, &pyrlk_no_image, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), CLFLUSH);
}
}
static void set_to_withmask_run(const oclMat &dst, const Scalar &scalar, const oclMat &mask, string kernelName)
{
CV_DbgAssert( dst.rows == mask.rows && dst.cols == mask.cols);
vector<pair<size_t , const void *> > args;
size_t localThreads[3] = {16, 16, 1};
size_t globalThreads[3];
globalThreads[0] = (dst.cols + localThreads[0] - 1) / localThreads[0] * localThreads[0];
globalThreads[1] = (dst.rows + localThreads[1] - 1) / localThreads[1] * localThreads[1];
globalThreads[2] = 1;
int step_in_pixel = dst.step / dst.elemSize(), offset_in_pixel = dst.offset / dst.elemSize();
char compile_option[32];
union sc
{
cl_uchar4 uval;
cl_char4 cval;
cl_ushort4 usval;
cl_short4 shval;
cl_int4 ival;
cl_float4 fval;
cl_double4 dval;
} val;
switch(dst.depth())
{
case CV_8U:
val.uval.s[0] = saturate_cast<uchar>(scalar.val[0]);
val.uval.s[1] = saturate_cast<uchar>(scalar.val[1]);
val.uval.s[2] = saturate_cast<uchar>(scalar.val[2]);
val.uval.s[3] = saturate_cast<uchar>(scalar.val[3]);
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=uchar");
args.push_back( make_pair( sizeof(cl_uchar) , (void *)&val.uval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=uchar4");
args.push_back( make_pair( sizeof(cl_uchar4) , (void *)&val.uval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case CV_8S:
val.cval.s[0] = saturate_cast<char>(scalar.val[0]);
val.cval.s[1] = saturate_cast<char>(scalar.val[1]);
val.cval.s[2] = saturate_cast<char>(scalar.val[2]);
val.cval.s[3] = saturate_cast<char>(scalar.val[3]);
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=char");
args.push_back( make_pair( sizeof(cl_char) , (void *)&val.cval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=char4");
args.push_back( make_pair( sizeof(cl_char4) , (void *)&val.cval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case CV_16U:
val.usval.s[0] = saturate_cast<ushort>(scalar.val[0]);
val.usval.s[1] = saturate_cast<ushort>(scalar.val[1]);
val.usval.s[2] = saturate_cast<ushort>(scalar.val[2]);
val.usval.s[3] = saturate_cast<ushort>(scalar.val[3]);
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=ushort");
args.push_back( make_pair( sizeof(cl_ushort) , (void *)&val.usval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=ushort4");
args.push_back( make_pair( sizeof(cl_ushort4) , (void *)&val.usval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case CV_16S:
val.shval.s[0] = saturate_cast<short>(scalar.val[0]);
val.shval.s[1] = saturate_cast<short>(scalar.val[1]);
val.shval.s[2] = saturate_cast<short>(scalar.val[2]);
val.shval.s[3] = saturate_cast<short>(scalar.val[3]);
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=short");
args.push_back( make_pair( sizeof(cl_short) , (void *)&val.shval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=short4");
args.push_back( make_pair( sizeof(cl_short4) , (void *)&val.shval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case CV_32S:
val.ival.s[0] = saturate_cast<int>(scalar.val[0]);
val.ival.s[1] = saturate_cast<int>(scalar.val[1]);
val.ival.s[2] = saturate_cast<int>(scalar.val[2]);
val.ival.s[3] = saturate_cast<int>(scalar.val[3]);
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=int");
args.push_back( make_pair( sizeof(cl_int) , (void *)&val.ival.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=int4");
args.push_back( make_pair( sizeof(cl_int4) , (void *)&val.ival ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case CV_32F:
val.fval.s[0] = scalar.val[0];
val.fval.s[1] = scalar.val[1];
val.fval.s[2] = scalar.val[2];
val.fval.s[3] = scalar.val[3];
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=float");
args.push_back( make_pair( sizeof(cl_float) , (void *)&val.fval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=float4");
args.push_back( make_pair( sizeof(cl_float4) , (void *)&val.fval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case CV_64F:
val.dval.s[0] = scalar.val[0];
val.dval.s[1] = scalar.val[1];
val.dval.s[2] = scalar.val[2];
val.dval.s[3] = scalar.val[3];
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=double");
args.push_back( make_pair( sizeof(cl_double) , (void *)&val.dval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=double4");
args.push_back( make_pair( sizeof(cl_double4) , (void *)&val.dval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unknown depth");
}
args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst.data ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.cols ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.rows ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&step_in_pixel ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&offset_in_pixel ));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&mask.data ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&mask.step ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&mask.offset ));
openCLExecuteKernel(dst.clCxt , &operator_setToM, kernelName, globalThreads,
localThreads, args, -1, -1, compile_option);
}
static void lkDense_run(oclMat &I, oclMat &J, oclMat &u, oclMat &v,
oclMat &prevU, oclMat &prevV, oclMat *err, Size winSize, int iters)
{
Context *clCxt = I.clCxt;
bool isImageSupported = clCxt->impl->devName.find("Intel(R) HD Graphics") == std::string::npos;
int elemCntPerRow = I.step / I.elemSize();
std::string kernelName = "lkDense";
size_t localThreads[3] = { 16, 16, 1 };
size_t globalThreads[3] = { I.cols, I.rows, 1};
bool calcErr;
if (err)
{
calcErr = true;
}
else
{
calcErr = false;
}
cl_mem ITex;
cl_mem JTex;
if (isImageSupported)
{
ITex = bindTexture(I);
JTex = bindTexture(J);
}
else
{
ITex = (cl_mem)I.data;
JTex = (cl_mem)J.data;
}
//int2 halfWin = {(winSize.width - 1) / 2, (winSize.height - 1) / 2};
//const int patchWidth = 16 + 2 * halfWin.x;
//const int patchHeight = 16 + 2 * halfWin.y;
//size_t smem_size = 3 * patchWidth * patchHeight * sizeof(int);
std::vector<std::pair<size_t , const void *> > args;
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&ITex ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&JTex ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&u.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&u.step ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&v.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&v.step ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&prevU.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&prevU.step ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&prevV.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&prevV.step ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&I.rows ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&I.cols ));
//args.push_back( std::make_pair( sizeof(cl_mem), (void *)&(*err).data ));
//args.push_back( std::make_pair( sizeof(cl_int), (void *)&(*err).step ));
if (!isImageSupported)
{
args.push_back( std::make_pair( sizeof(cl_int), (void *)&elemCntPerRow ) );
}
args.push_back( std::make_pair( sizeof(cl_int), (void *)&winSize.width ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&winSize.height ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&iters ));
args.push_back( std::make_pair( sizeof(cl_char), (void *)&calcErr ));
if (isImageSupported)
{
openCLExecuteKernel2(clCxt, &pyrlk, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), CLFLUSH);
releaseTexture(ITex);
releaseTexture(JTex);
}
else
{
//printf("Warning: The image2d_t is not supported by the device. Using alternative method!\n");
openCLExecuteKernel2(clCxt, &pyrlk_no_image, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), CLFLUSH);
}
}
//////////////////////////////////////////////////////////////////////////////
/////////////////////// add subtract multiply divide /////////////////////////
//////////////////////////////////////////////////////////////////////////////
static void pyrdown_run(const oclMat &src, const oclMat &dst)
{
CV_Assert(src.type() == dst.type());
CV_Assert(src.depth() != CV_8S);
Context *clCxt = src.clCxt;
//int channels = dst.channels();
//int depth = dst.depth();
String kernelName = "pyrDown";
//int vector_lengths[4][7] = {{4, 0, 4, 4, 1, 1, 1},
// {4, 0, 4, 4, 1, 1, 1},
// {4, 0, 4, 4, 1, 1, 1},
// {4, 0, 4, 4, 1, 1, 1}
//};
//size_t vector_length = vector_lengths[channels-1][depth];
//int offset_cols = (dst.offset / dst.elemSize1()) & (vector_length - 1);
size_t localThreads[3] = { 256, 1, 1 };
size_t globalThreads[3] = { src.cols, dst.rows, 1};
//int dst_step1 = dst.cols * dst.elemSize();
std::vector<std::pair<size_t , const void *> > args;
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&src.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&src.step ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&src.rows));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&src.cols));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&dst.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&dst.step ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&dst.cols));
openCLExecuteKernel(clCxt, &pyr_down, kernelName, globalThreads, localThreads, args, src.oclchannels(), src.depth());
}
static void lkDense_run(oclMat &I, oclMat &J, oclMat &u, oclMat &v,
oclMat &prevU, oclMat &prevV, oclMat *err, Size winSize, int iters)
{
Context *clCxt = I.clCxt;
bool isImageSupported = support_image2d();
int elemCntPerRow = I.step / I.elemSize();
String kernelName = "lkDense";
size_t localThreads[3] = { 16, 16, 1 };
size_t globalThreads[3] = { I.cols, I.rows, 1};
bool calcErr;
if (err)
{
calcErr = true;
}
else
{
calcErr = false;
}
cl_mem ITex;
cl_mem JTex;
if (isImageSupported)
{
ITex = bindTexture(I);
JTex = bindTexture(J);
}
else
{
ITex = (cl_mem)I.data;
JTex = (cl_mem)J.data;
}
std::vector<std::pair<size_t , const void *> > args;
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&ITex ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&JTex ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&u.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&u.step ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&v.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&v.step ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&prevU.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&prevU.step ));
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&prevV.data ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&prevV.step ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&I.rows ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&I.cols ));
if (!isImageSupported)
args.push_back( std::make_pair( sizeof(cl_int), (void *)&elemCntPerRow ) );
args.push_back( std::make_pair( sizeof(cl_int), (void *)&winSize.width ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&winSize.height ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&iters ));
args.push_back( std::make_pair( sizeof(cl_char), (void *)&calcErr ));
if (isImageSupported)
{
openCLExecuteKernel2(clCxt, &pyrlk, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), CLFLUSH);
releaseTexture(ITex);
releaseTexture(JTex);
}
else
openCLExecuteKernel2(clCxt, &pyrlk_no_image, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), CLFLUSH);
}
cl_mem bindTexture(const oclMat &mat)
{
cl_mem texture;
cl_image_format format;
int err;
int depth = mat.depth();
int channels = mat.oclchannels();
switch(depth)
{
case CV_8U:
format.image_channel_data_type = CL_UNSIGNED_INT8;
break;
case CV_32S:
format.image_channel_data_type = CL_UNSIGNED_INT32;
break;
case CV_32F:
format.image_channel_data_type = CL_FLOAT;
break;
default:
CV_Error(-1, "Image forma is not supported");
break;
}
switch(channels)
{
case 1:
format.image_channel_order = CL_R;
break;
case 3:
format.image_channel_order = CL_RGB;
break;
case 4:
format.image_channel_order = CL_RGBA;
break;
default:
CV_Error(-1, "Image format is not supported");
break;
}
#ifdef CL_VERSION_1_2
//this enables backwards portability to
//run on OpenCL 1.1 platform if library binaries are compiled with OpenCL 1.2 support
if(Context::getContext()->supportsFeature(Context::CL_VER_1_2))
{
cl_image_desc desc;
desc.image_type = CL_MEM_OBJECT_IMAGE2D;
desc.image_width = mat.cols;
desc.image_height = mat.rows;
desc.image_depth = 0;
desc.image_array_size = 1;
desc.image_row_pitch = 0;
desc.image_slice_pitch = 0;
desc.buffer = NULL;
desc.num_mip_levels = 0;
desc.num_samples = 0;
texture = clCreateImage((cl_context)mat.clCxt->oclContext(), CL_MEM_READ_WRITE, &format, &desc, NULL, &err);
}
else
#endif
{
texture = clCreateImage2D(
(cl_context)mat.clCxt->oclContext(),
CL_MEM_READ_WRITE,
&format,
mat.cols,
mat.rows,
0,
NULL,
&err);
}
size_t origin[] = { 0, 0, 0 };
size_t region[] = { mat.cols, mat.rows, 1 };
cl_mem devData;
if (mat.cols * mat.elemSize() != mat.step)
{
devData = clCreateBuffer((cl_context)mat.clCxt->oclContext(), CL_MEM_READ_ONLY, mat.cols * mat.rows
* mat.elemSize(), NULL, NULL);
const size_t regin[3] = {mat.cols * mat.elemSize(), mat.rows, 1};
clEnqueueCopyBufferRect((cl_command_queue)mat.clCxt->oclCommandQueue(), (cl_mem)mat.data, devData, origin, origin,
regin, mat.step, 0, mat.cols * mat.elemSize(), 0, 0, NULL, NULL);
clFlush((cl_command_queue)mat.clCxt->oclCommandQueue());
}
else
{
devData = (cl_mem)mat.data;
}
clEnqueueCopyBufferToImage((cl_command_queue)mat.clCxt->oclCommandQueue(), devData, texture, 0, origin, region, 0, NULL, 0);
if ((mat.cols * mat.elemSize() != mat.step))
{
clFlush((cl_command_queue)mat.clCxt->oclCommandQueue());
clReleaseMemObject(devData);
}
openCLSafeCall(err);
return texture;
}
void cv::ocl::split(const oclMat &src, std::vector<oclMat> &dst)
{
dst.resize(src.oclchannels());
if(src.oclchannels() > 0)
split_merge::split(src, &dst[0]);
}
