C++ oclMat::depth方法代码示例

// radiusMatchSingle
void cv::ocl::BruteForceMatcher_OCL_base::radiusMatchSingle(const oclMat &query, const oclMat &train,
        oclMat &trainIdx,   oclMat &distance, oclMat &nMatches, float maxDistance, const oclMat &mask)
{
    if (query.empty() || train.empty())
        return;

    // match1 doesn't support signed char type, match2 only support float, hamming support uchar, ushort and int
    int callType = query.depth();
    char cvFuncName[] = "radiusMatchSingle";
    if (callType != 5)
        CV_ERROR(CV_UNSUPPORTED_FORMAT_ERR, "BruteForceMatch OpenCL only support float type query!\n");

    if ((distType == 0 && callType == 1 ) || (distType == 1 && callType != 5) || (distType == 2 && (callType != 0
        || callType != 2 || callType != 4)))
    {
        CV_ERROR(CV_UNSUPPORTED_DEPTH_ERR, "BruteForceMatch OpenCL only support float type query!\n");
    }

    CV_Assert(query.channels() == 1 && query.depth() < CV_64F);
    CV_Assert(train.type() == query.type() && train.cols == query.cols);
    CV_Assert(trainIdx.empty() || (trainIdx.rows == query.rows && trainIdx.size() == distance.size()));

    nMatches.create(1, query.rows, CV_32SC1);
    if (trainIdx.empty())
    {
        trainIdx.create(query.rows, std::max((train.rows/ 100), 10), CV_32SC1);
        distance.create(query.rows, std::max((train.rows/ 100), 10), CV_32FC1);
    }

    nMatches.setTo(Scalar::all(0));

    matchDispatcher(query, train, maxDistance, mask, trainIdx, distance, nMatches, distType);
exit:
    return;
}

///////////////////////////////////////////////////////////////////////////
//////////////////////////////// ConvertTo ////////////////////////////////
///////////////////////////////////////////////////////////////////////////
static void convert_run(const oclMat &src, oclMat &dst, double alpha, double beta)
{
    String kernelName = "convert_to_S";
    std::stringstream idxStr;
    idxStr << src.depth();
    kernelName = kernelName + idxStr.str().c_str();
    float alpha_f = alpha, beta_f = 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 ));
    openCLExecuteKernel(dst.clCxt , &operator_convertTo, kernelName, globalThreads,
                        localThreads, args, dst.oclchannels(), dst.depth());
}

void cv::ocl::distanceToCenters(const oclMat &src, const oclMat &centers, 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 *)&centers.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 *)&centers_step));
    args.push_back(make_pair(sizeof(cl_int), (void *)&src.rows));
    args.push_back(make_pair(sizeof(cl_int), (void *)&centers.rows));

    args.push_back(make_pair(sizeof(cl_int), (void *)&src_offset));
    args.push_back(make_pair(sizeof(cl_int), (void *)&centers_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;
    }
}

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;
    String kernelName = "lkSparse";
    size_t localThreads[3]  = { 8, 8, 1 };
    size_t globalThreads[3] = { 8 * ptcount, 8, 1};
    int cn = I.oclchannels();
    char calcErr = level==0?1:0;

    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 ));

    bool is_cpu = isCpuDevice();
    if (is_cpu)
    {
        openCLExecuteKernel(clCxt, &pyrlk, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), (char*)" -D CPU");
    }
    else
    {
        std::stringstream idxStr;
        idxStr << kernelName << "_C" << I.oclchannels() << "_D" << I.depth();
        cl_kernel kernel = openCLGetKernelFromSource(clCxt, &pyrlk, idxStr.str());
        int wave_size = (int)queryWaveFrontSize(kernel);
        openCLSafeCall(clReleaseKernel(kernel));

        static char opt[32] = {0};
        sprintf(opt, "-D WAVE_SIZE=%d", wave_size);

        openCLExecuteKernel(clCxt, &pyrlk, kernelName, globalThreads, localThreads,
                            args, I.oclchannels(), I.depth(), opt);
    }
    releaseTexture(ITex);
    releaseTexture(JTex);
}

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());
}

void matchUnrolledCached(const oclMat &query, const oclMat &train, const oclMat &/*mask*/,
                         const oclMat &trainIdx, const oclMat &distance, int distType)
{
    cv::ocl::Context *ctx = query.clCxt;
    size_t globalSize[] = {(query.rows + BLOCK_SIZE - 1) / BLOCK_SIZE * BLOCK_SIZE, BLOCK_SIZE, 1};
    size_t localSize[] = {BLOCK_SIZE, BLOCK_SIZE, 1};
    const size_t smemSize = (BLOCK_SIZE * (MAX_DESC_LEN >= 2 * BLOCK_SIZE ? MAX_DESC_LEN : 2 * BLOCK_SIZE) + BLOCK_SIZE * BLOCK_SIZE) * sizeof(int);
    int block_size = BLOCK_SIZE;
    int m_size = MAX_DESC_LEN;
    std::vector< std::pair<size_t, const void *> > args;

    char opt [OPT_SIZE] = "";
    sprintf(opt,
        "-D T=%s -D DIST_TYPE=%d -D BLOCK_SIZE=%d -D MAX_DESC_LEN=%d",
        T_ARR[query.depth()], distType, block_size, m_size);

    if(globalSize[0] != 0)
    {
        args.push_back( std::make_pair( sizeof(cl_mem), (void *)&query.data ));
        args.push_back( std::make_pair( sizeof(cl_mem), (void *)&train.data ));
        //args.push_back( std::make_pair( sizeof(cl_mem), (void *)&mask.data ));
        args.push_back( std::make_pair( sizeof(cl_mem), (void *)&trainIdx.data ));
        args.push_back( std::make_pair( sizeof(cl_mem), (void *)&distance.data ));
        args.push_back( std::make_pair( smemSize, (void *)NULL));
        args.push_back( std::make_pair( sizeof(cl_int), (void *)&query.rows ));
        args.push_back( std::make_pair( sizeof(cl_int), (void *)&query.cols ));
        args.push_back( std::make_pair( sizeof(cl_int), (void *)&train.rows ));
        args.push_back( std::make_pair( sizeof(cl_int), (void *)&train.cols ));
        args.push_back( std::make_pair( sizeof(cl_int), (void *)&query.step ));

        String kernelName = "BruteForceMatch_UnrollMatch";

        openCLExecuteKernel(ctx, &brute_force_match, kernelName, globalSize, localSize, args, -1, -1, opt);
    }
}

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);
}

// knn match
void cv::ocl::BruteForceMatcher_OCL_base::knnMatchSingle(const oclMat &query, const oclMat &train, oclMat &trainIdx,
        oclMat &distance, oclMat &allDist, int k, const oclMat &mask)
{
    if (query.empty() || train.empty())
        return;

    CV_Assert(query.channels() == 1 && query.depth() < CV_64F);
    CV_Assert(train.type() == query.type() && train.cols == query.cols);

    const int nQuery = query.rows;
    const int nTrain = train.rows;

    if (k == 2)
    {
        ensureSizeIsEnough(1, nQuery, CV_32SC2, trainIdx);
        ensureSizeIsEnough(1, nQuery, CV_32FC2, distance);
    }
    else
    {
        ensureSizeIsEnough(nQuery, k, CV_32S, trainIdx);
        ensureSizeIsEnough(nQuery, k, CV_32F, distance);
        ensureSizeIsEnough(nQuery, nTrain, CV_32FC1, allDist);
    }

    trainIdx.setTo(Scalar::all(-1));

    kmatchDispatcher(query, train, k, mask, trainIdx, distance, allDist, distType);

    return;
}

static void toHSV_caller(const oclMat &src, oclMat &dst, int bidx, const std::string & kernelName,
                           const std::string & additionalOptions = std::string(),
                           const oclMat & data1 = oclMat(), const oclMat & data2 = oclMat())
{
    int src_offset = src.offset / src.elemSize1(), src_step = src.step1();
    int dst_offset = dst.offset / dst.elemSize1(), dst_step = dst.step1();

    std::string build_options = format("-D DEPTH_%d -D scn=%d -D bidx=%d", src.depth(), src.oclchannels(), bidx);
    if (!additionalOptions.empty())
        build_options += additionalOptions;

    vector<pair<size_t , const void *> > args;
    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 *)&src_step));
    args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_step));
    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_int) , (void *)&src_offset ));
    args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_offset ));

    if (!data1.empty())
        args.push_back( make_pair( sizeof(cl_mem) , (void *)&data1.data ));
    if (!data2.empty())
        args.push_back( make_pair( sizeof(cl_mem) , (void *)&data2.data ));

   size_t gt[3] = { dst.cols, dst.rows, 1 };
#ifdef ANDROID
    size_t lt[3] = { 16, 10, 1 };
#else
    size_t lt[3] = { 16, 16, 1 };
#endif
    openCLExecuteKernel(src.clCxt, &cvt_color, kernelName.c_str(), gt, lt, args, -1, -1, build_options.c_str());
}

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
}

void cv::ocl::oclMat::convertTo( oclMat &dst, int rtype, double alpha, double beta ) const
{
    if (!clCxt->supportsFeature(FEATURE_CL_DOUBLE) &&
            (depth() == CV_64F || dst.depth() == CV_64F))
    {
        CV_Error(Error::OpenCLDoubleNotSupported, "Selected device doesn't support double");
        return;
    }

    bool noScale = fabs(alpha - 1) < std::numeric_limits<double>::epsilon()
                   && fabs(beta) < std::numeric_limits<double>::epsilon();

    if( rtype < 0 )
        rtype = type();
    else
        rtype = CV_MAKETYPE(CV_MAT_DEPTH(rtype), channels());

    int sdepth = depth(), ddepth = CV_MAT_DEPTH(rtype);
    if( sdepth == ddepth && noScale )
    {
        copyTo(dst);
        return;
    }

    oclMat temp;
    const oclMat *psrc = this;
    if( sdepth != ddepth && psrc == &dst )
        psrc = &(temp = *this);

    dst.create( size(), rtype );
    convert_run(*psrc, dst, alpha, beta);
}

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);
    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();

    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()]);

    oclMat m(Mat(1, 1, dst.type(), scalar));
    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 ));
    args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&mask.data ));
    args.push_back( std::make_pair( sizeof(cl_int) , (void *)&mask.step ));
    args.push_back( std::make_pair( sizeof(cl_int) , (void *)&mask.offset ));
    openCLExecuteKernel(dst.clCxt , &operator_setToM, kernelName, globalThreads,
                        localThreads, args, -1, -1, buildOptions.c_str());
}

static void pyrDown_cus(const oclMat &src, oclMat &dst)
{
    CV_Assert(src.depth() <= CV_32F && src.channels() <= 4);

    dst.create((src.rows + 1) / 2, (src.cols + 1) / 2, src.type());
    pyrdown_run_cus(src, dst);
}

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);
    }
}