Is the mapped EGLImage object from dmabuff_fd reusable?

[calvin0328]

Hi,

Currently I’m working on the workflow where the camera frames are streaming out and I will do some CUDA processing on it then display it into screen. The buffer is allocated into DMA buffer using NvBufSurface. According to the sample codes of cuda_postprocess from 12_v4l2_camera_cuda, and Handle_EGLImage from NvCudaProc, I can get the device pointer from EGLImage wrapped with the dmabuff_fd. However, I overserved that while receiving every frame, I need to:

• Get NvBufSurface from fd,
• Map NvBufSurface to EGLImage
• Register the EGLImage with CUDA
• Register resource
• CUDA processing
• Unregister resource
• Unmap the EGLImage

I have to repeat those steps for every frame. I was wondering whether those steps can be reusable or not? For example, I allocate the dma buffer, and mapped into EGL image and register the corresponding resources beforehand, and store them as member variables. During the processing, I just reuse the pre-allocated/registered variables, and do the processing. Upon deconstructor, I deallocate, unregister or unmap those resources. I guess it would be saving a lot of latency. I have tested there is no issue at all, no memory leak, no frame artefacts, but I didn’t find any official reference of such kind of usage, so I was wondering whether there is any potential issue with those operations? Should I avoid it? Is there any method to avoid those repeated steps?

Thanks in advance!

Best,
Wenhai.

[DaneLLL]

Hi,
Your method looks good. It should work well to do map/register in initializing NvBufSurfaces, and unmap/unregister in destroying the NvBufSurfaces.

[calvin0328]

Hi Dane,

Thanks for quickly getting back! Let me reorganize the bullet points as below:

• Get NvBufSurface from fd,
• Map NvBufSurface to EGLImage
• Register the EGLImage with CUDA
• Register resource
  while(new frame available & !quit)
  • CUDA_processing()
• Unregister resource
• Unmap the EGLImage

Is this workflow still safe?

[DaneLLL]

Hi,
The workflow looks fine. Please give it a try. If you hit certain issue, please share a patch on either jetson_multimedia_api sample for replicating the issue. We will set up and check further.

[calvin0328]

Thanks a lot!