Re-use custom call output buffer

[kkHuang-amd]

I am doing JAX custom call implementation and refer to this page tutorial "https://jax.readthedocs.io/en/latest/Custom_Operation_for_GPUs.html".

I require to accumulate the output buffer continuously.

But the output buffer is automatically generated according to the abstract function.

This means I must copy the current output into the new generated output buffer to let the new output have the previous output result.

This will generate extra copy overhead.

Is it possible to keep using the first created output buffer to avoid the extra copy?

Thanks