opencv cuda filter lacks step(stride) information

[yl070360]

System information (version)

• OpenCV => 4.2 - 4.5
• Operating System / Platform => linux
• Compiler => gcc/g++

Detailed description

Steps to reproduce

Hi，I found that the processing of GPUMat in filtering.cpp(line:284) lacks step(stride) information, 
And tried the following modifications, it seems to work normally

filtering.cpp
// C++ code example
void LinearFilter::apply(InputArray _src, OutputArray _dst, Stream& _stream)
{
    GpuMat src = _src.getGpuMat();
    ...
    Point ofs;
    Size wholeSize;
    src.locateROI(wholeSize, ofs);

    // GpuMat srcWhole(wholeSize, src.type(), src.datastart);
    // line:284, this is my modifications test
    GpuMat srcWhole(wholeSize, src.type(), src.datastart, src.step);
    

    func_(srcWhole, ofs.x, ofs.y, dst, kernel_.ptr<float>(),
          kernel_.cols, kernel_.rows, anchor_.x, anchor_.y,
          borderMode_, borderVal_.val, StreamAccessor::getStream(_stream));
}

Issue submission checklist

• I report the issue, it's not a question
• I checked the problem with documentation, FAQ, open issues,
  forum.opencv.org, Stack Overflow, etc and have not found any solution
• I updated to the latest OpenCV version and the issue is still there
• There is reproducer code and related data files: videos, images, onnx, etc

[yl070360]

@cudawarped can you help me

[cudawarped]

@cudawarped can you help me

Can you give me an example of where it previously failed?
Looking at the cpp, src.locateROI(wholeSize, ofs); should return the size of the matrix including step in wholeSize. That is wholesize.width == src.step/sizeof(DATATYPE) and because cudaMallocPitch will be used in the allocation of _src and srcWhole I would expect the step to be identical in both cases. Off the top of my head I can't think of a reason for using wholeSize is instead of src.size() and src.step() apart from because its just been queried, but I wouldn't be suprised if there is one.

[yl070360]

Thanks much for @cudawarped reply!

This is my test case，_src Mat is from NvBufSurface, in which layout (_src.width(1920) != _src.step(2048))(There is an area of invalid data in the right side of the matrix),

I just want to avoid data copying and reuse existing data as much as possible, so the _src is from exist NvBufSurface

following test case, the laplace-results of cpu side and gpu side are not equal, if I modify cpp
“GpuMat srcWhole(wholeSize, src.type(), src.datastart)” to “GpuMat srcWhole(wholeSize, src.type(), src.datastart, src.step)”，the results are equal.

Maybe some mistake in my usage, thanks and waiting for your advice

#define CPU_LAPLACE
#define GPU_LAPLACE
static GstPadProbeReturn nvinfer_pad_buffer_probe (GstPad * pad, GstPadProbeInfo * info, gpointer param) {
    GstBuffer *buf = (GstBuffer *)info->data; GstMapInfo map_info = { 0 };
    gst_buffer_map(buf, &map_info, GST_MAP_READ);
    NvBufSurface *ip_surf = (NvBufSurface *) map_info.data;

    NvDsBatchMeta *batch_meta = gst_buffer_get_nvds_batch_meta(buf);
    for (NvDsMetaList *l_frame = batch_meta->frame_meta_list; l_frame != NULL; l_frame = l_frame->next) {
        NvDsFrameMeta *frame_meta = (NvDsFrameMeta *) (l_frame->data);
#ifdef CPU_LAPLACE
        {
            NvBufSurfaceMap (ip_surf, frame_meta->batch_id, 0, NVBUF_MAP_READ_WRITE);

            cv::Mat in_mat = cv::Mat(ip_surf->surfaceList[frame_meta->batch_id].planeParams.height[0], ip_surf->surfaceList[frame_meta->batch_id].planeParams.width[0], 
                                     CV_8UC1,
                                     ip_surf->surfaceList[frame_meta->batch_id].mappedAddr.addr[0], ip_surf->surfaceList[frame_meta->batch_id].planeParams.pitch[0]);
            cv::Mat imageSobel;
            cv::Laplacian(in_mat, imageSobel, CV_8U);

            cv::Scalar mean, sdv;
            cv::meanStdDev(imageSobel, mean, sdv);
            printf("frame-cpu mean-stddev( %f, %f )\n", mean[0], sdv[0]);
        }
#endif

#ifdef GPU_LAPLACE
        {
            
            cv::cuda::GpuMat in_mat = cv::cuda::GpuMat(ip_surf->surfaceList[frame_meta->batch_id].planeParams.height[0], ip_surf->surfaceList[frame_meta->batch_id].planeParams.width[0], 
                                                       CV_8UC1, 
                                                       (void *)ip_surf->surfaceList[frame_meta->batch_id].dataPtr, ip_surf->surfaceList[frame_meta->batch_id].planeParams.pitch[0]);
            cv::Ptr<cv::cuda::Filter> filter = cv::cuda::createLaplacianFilter(in_mat.type(), -1);
            cv::cuda::GpuMat laplace_mat;
            filter->apply(in_mat, laplace_mat);

            cv::Scalar mean, sdv;
            cv::cuda::meanStdDev(laplace_mat, mean, sdv);

            printf("frame-gpu mean-stddev( %f, %f )\n", mean[0], sdv[0]);   
        }
#endif

    gst_buffer_unmap(buf, &map_info);
}

[cudawarped]

Ignore my previous comment, this looks like a bug and I would submit a PR.

In fact it looks like this should always return the incorrect result but due to another bug in the GpuMat constructor it works unless you allocate the device memory yourself.

[yl070360]

I just modified the code(add src.step in filtering.cpp:284) temporarily in my project.
I would like to pull the new source when the graceful PR submit is OK.
Happy Chinese New Year!

[cudawarped]

I can submit the Pull Request on your behalf but I think if you haven't submited one before, it would be beneficial to do it yourself to understand how the process works. Let me know what you decide.

[yl070360]

I am a beginner of opencv, I am glad to wait and enjoy your PR, thanks very much!