When only host is updated, the device is out of date (and vice versa)

[pelesh]

In update methods in vector and matrix classes it is probably safer to set both update flags at the same time - i.e if cpu data gets updated, gpu data automatically should switch to NOT updated.

_Originally posted by @kswirydo in https://github.com/ORNL/ReSolve/pull/124#discussion_r1429345023_

[pelesh]

After looking at possible use cases, I think it is better if update methods set flags only for memory spaces they updated. A user may run updates for for host and device in separate calls and bring them both to same up-to-date values. In this case "automatic" assumption that the other memory space is out of date would be untrue.

What I suggest is that

• updateData and updateValues functions set "updated" flags only for the memory space they act on (update).
• It is users responsibility to indicate if the host and device memory are out of sync, using setUpdated and setNotUpdated functions.

A possible action here would be to allow setNotUpdated function to set specific memory space as not current. Other functionality is already available in ReSolve.

@maksud, please chime in, as well.

[pelesh]

The alternative to my earlier proposal could be to do exactly as @kswirydo suggested -- once device memory is updated alone, the host memory is automatically flagged as out of date. In that case, it would be user's responsibility to sync host and data using syncData methods in vector and matrix classes when needed. The syncData method would need to check the destination is out of date and the source is updated, and send warning/error message otherwise.