[DFT] Allow the descriptor to be modified and recommitted

[FMarno]

Description

Allows the descriptor class for DFTs to be modified and recommitted. Some backends will likely need to regenerate the descriptor completely, but this will allow for those that don't.

This builds on top of #259

Checklist

All Submissions

• Do all unit tests pass locally? Attach a log.
• Have you formatted the code using clang-format?

New features

• Have you provided motivation for adding a new feature?
• Have you added relevant tests?

[hjabird]

I think the biggest challenge is what to do in situations that the spec doesn't cover - eg. changing queue or backend for the descriptor.

[FMarno]

Log of the test runs. My machine only supports single precision and through opencl, but I hope this demonstrates the descriptor tests are passing.
multi_commit_result.txt

[Rbiessy]

Next time this needs to be rebased you could fix the typo recommited -> recommitted in the commit description.

[raphael-egan]

Maybe it's worth adding a recommit_values test which would change the queue between commit calls?

[FMarno]

I've added a tests to show that there is a wait on the old queue here 088fa08
I was unable to prove that work is submitted to the new queue or not submitted to the old.

• I don't believe there is any direct way to query if a queue has not completed tasks
• I don't believe there is a way to query if an event is associated with a task.
• When an command group function object is submitted to the queue, I don't believe there are any guarantees that is will complete before slow work on other queues.

Because of all this I can't think of how to test that reliably.

[raphael-egan]

Thank you for that. I think you're right for every bullet point you mentioned and I'm sorry if my original suggestion generated complications/misunderstanding on the testing scope. Those changes look fine to me.
To be clearer, my original concern was about a use case like

1. create a descriptor desc;
2. create a queue q1 and commit desc to q1;
3. delete q1;
4. create a queue q2 and commit desc to q2;
5. submit a compute task to desc.
   I am not 100% sure this works as expected for the closed-source mkl. But that is something we can/should check on our side as that would be an issue for us to fix anyways. We may consider this thread resolved.

[FMarno]

Ok, that's interesting. I think this should work because objects like sycl::queue are kept alive by the SYCL runtime with reference counting semantics. I guess I could have a test that follows the steps you've detailed, waits on q2 and then shows that the event returned from the task is complete. It's not perfect but its something. I'm about to finish for the day but I'll do this sometime tomorrow.

[FMarno]

I've added a test for that in 8d0a608

[raphael-egan]

Thank you for your work!

[lhuot]

@FMarno conflicts need to be resolved before I can merge this. Thanks!

[FMarno]

In the last rebase I made some changes to compute_inplace.hpp and compute_out_of_place.hpp which meant the same descriptor was used twice with a recommit in the middle to update the strides.

Here is evidence of the passing tests. I still only have access to single precision and the OpenCL backend -
recommit.txt