Is it possible to directly map the NvBuffer to a OpenCV GpuMat?

[ysyyork]

Currently, I was able to convert a NvBuffer directly to CPU opencv Mat by using NvBufferMemMap, NvBufferMemSyncForCpu, NvBufferMemUnMap. I was also able to convert NvBuffer to a OpenCV GpuMat by firstly converting the NvBuffer to CUEglFrame and then to OpenCV GpuMat. The second route is a little redundant cus I need to copy data from NvBuffer to CUEglFrame and then to OpenCV GpuMat. I saw there is also a NvBufferMemSyncForDevice function so i’m wondering if I can do it like the first route directly from NvBuffer to Opencv GpuMat. I quickly tried this API but I got the below error

error: (-217:Gpu API call) invalid argument in function 'download'

Not sure if my usage is correct. Is there any example? Thanks.

This is the code snippet:

                void *pdata = NULL;

                NvBufferMemMap(fd, 0, NvBufferMem_Read, &pdata);
                NvBufferMemSyncForDevice(fd, 0, &pdata);
                cv::cuda::GpuMat imgBuf = cv::cuda::GpuMat(512,
                                                  512,
                                                  CV_8UC4, pdata);

                cv::Mat cpuImgBuf;
                bb.download(cpuImgBuf);
                NvBufferMemUnMap(fd, 0, &pdata);

If I change the NvBufferMemSyncForDevice to NvBufferMemSyncForCpu, it works

[DaneLLL]

Hi,
Please refer to the patches:
LibArgus EGLStream to nvivafilter - #14 by DaneLLL
Nano not using GPU with gstreamer/python. Slow FPS, dropped frames - #8 by DaneLLL

[ysyyork]

Hi, I checked both links you shared, but both links are doing NvBuffer → EglFrame → GpuMat route. My question is actually that is there a way to do NvBuffer → GpuMat directly. The reason I asked is that for Cpu Mat, I can use NvBufferMemMap and NvBufferMemSyncForCpu to directly map the NvBuffer to Mat without an intermediate EglFrame representation. But I couldn’t achieve the same when I change to NvBufferMemSyncForDevice. However, I think from the performance perspective, directly converting NvBuffer to GpuMat should be faster than relaying with EglFrame, correct?

[DaneLLL]

Hi,
For mapping to GpuMat, the pointer has to be CUDA pointer, so the CUDA function calls are required. For calling NvBufferMemMap(), can get CPU pointer and map to cv::Mat.

On AGX Xavier, both GPU engine and CPU cores have good capability, so it depends on your use-case. If you need CUDA filters in OpenCV, can use GpuMat. If GPU is at high loading, it is good to use cv::Mat to shift the loading got CPU.

[ysyyork]

OK, got it. thanks!