Zero copy with tensorflow

Hi there,

I use some network which requires a lot of CPU <-> GPU transfers. AFAIK, CPU and GPU have unified (common) memory, so using a zero-copy technique is possible.

I’ve read a topic about TX1, in which the answer states that a modification to a tensorflow sources is needed.
This article gives a drection in what modifications should be done.

But I already have a specific implementation by nvidia included in JetPack (I use v4.2).

Does this version use zero-copy?
Is it planned to implement such an improvement in JetPack?
Is it possible to get sources to modify it enabling zero-copy?

Does anyone succeed in implementing zero-copy on Jetson AGX Xavier?

Thanks in advance for answers.


Sorry that the implementation must be applied in the TensorFlow source.
It’s recommended to file this request to the TensorFlow team directly.

For an alternative, you can try our Deepstream SDK which implements zero-copy technique between camera and inference instead: