Can GPU play a role in data processing?

[wangshuaicumt]

Hello,

During the installation of sweets, I replaced isce3 with isce3-cuda in the environment file and made sure that the CUDA drivers and dependencies are correctly configured (nvidia-smi recognizes the GPU). However, when I run sweets, I notice that the GPU is not utilized. I also did not find any explicit option in the sweets config files or documentation to enable or configure GPU usage.

I would like to ask:

When sweets calls ISCE3, does it automatically detect and use the GPU, or is there additional configuration required in the config files or runtime parameters?

How can I confirm that a specific sweets module is actually running on the GPU rather than falling back to the CPU?

If sweets is not yet adapted for isce3-cuda, are there any plans to support this in the future? Or are there any recommended workarounds to enable GPU acceleration at this time?

Thank you very much for your help. I look forward to your reply!

[scottstanie]

Hi, thanks for trying it out.
I think this could be an issue raise on the COMPASS repo (which is what sweets uses to make the geocoded SLCs).

COMPASS makes the runconfigs for the stack of SLCS here: https://github.com/opera-adt/COMPASS/blob/main/src/compass/s1_geocode_stack.py#L225
It doesn't include the GPU option at the moment, and the default is False

But also, I'm not sure that setting it to true would make it work, as there's no gpu referenced in the main geocode slc workflow

There's an earlier script which does use it:
https://github.com/opera-adt/COMPASS/blob/487c7eb7461faaf4719f0342b7757cdcf61401e1/src/compass/s1_rdr2geo.py#L53-L61
but that may be out of date for the latest isce3 versions.
then once those changes are in, we could add the use_gpu option to sweets here: https://github.com/isce-framework/sweets/blob/main/src/sweets/_geocode_slcs.py