struct BufferPSNR                                     // Optimized GPU versions
{   // Data allocations are very expensive on GPU. Use a buffer to solve: allocate once reuse later.
    gpu::GpuMat gI1, gI2, gs, t1,t2;

    gpu::GpuMat buf;
};

double getPSNR_GPU(const Mat& I1, const Mat& I2)
{
    gpu::GpuMat gI1, gI2, gs, t1,t2;

    gI1.upload(I1);
    gI2.upload(I2);

    gI1.convertTo(t1, CV_32F);
    gI2.convertTo(t2, CV_32F);

    gpu::absdiff(t1.reshape(1), t2.reshape(1), gs);
    gpu::multiply(gs, gs, gs);

    Scalar s = gpu::sum(gs);
    double sse = s.val[0] + s.val[1] + s.val[2];

    if( sse <= 1e-10) // for small values return zero
        return 0;
    else
    {
        double  mse =sse /(double)(gI1.channels() * I1.total());
        double psnr = 10.0*log10((255*255)/mse);
        return psnr;
    }
}