VPI pycuda interoperability - VPI_ERROR_INTERNAL: Can't perform shared mapping

Hello,
I am using a pipeline that includes VPI for preprocessing images and I use TensorRT for inference. To interact with the CUDA memory in the inference part, I use pycuda. Separately, each part runs without errors, however, when both VPI and pycuda come together, when bringing the ImageData into the CPU memory with either of the options described here, the above mentioned error arises. VPI_ERROR_INTERNAL: Can’t perform shared mapping. Does anyone have a suggestion?

1 Like

Hi,

Which memory type do you use for allocation?
If you want to access the buffer with both CPU and GPU, please try unified memory:

https://documen.tician.de/pycuda/driver.html#managed-memory

Please noted that Jetson doesn’t support concurrent access.
You will need to make sure the previous job is finished before accessing it with another processor.

Thanks.