Fix NumPy copy functions when using GPU backend - #229
Conversation
followed the other functions' way of copying
|
@Sichao25 , @cwsmith , Could you please add these changes into your spack pull request spack/spack-packages#5792 after adding this in a new version? Otherwise I will need to create a different pull request with this which is time consuming since the spack people take time to merge. |
There was a problem hiding this comment.
Please feel free to let me know if there's a better way of testing this.
|
Why tests are not run on the CUDA workflows? May I run the tests there? |
|
/runtests |
|
@Fuad-HH It looks like the github hosted cuda builds completed without issue (note, these don't run tests as github doesn't provide gpus). I just queued the self-hosted builds+tests. |
That is fine with me. @Sichao25 would you mind creating the new 'patch' version and including this? @Fuad-HH @Sichao25 Are there other python interface changes coming soon we should include in the spack PR (it seems like it can/will be merged soon as we have some help). |
|
Test Results:
|
jacobmerson
left a comment
There was a problem hiding this comment.
I only added the comment to one place, but all of these require two copies of the data. I'm wondering if we can somehow do this with a direct deep_copy style call from the numpy buffer. That may require wrapping in a Kokkos<unmanaged>
|
No problem. I'll create the patch version once the PR is merged. |
Sichao25
left a comment
There was a problem hiding this comment.
No further questions here. This reminds me I should remove or update the related/duplicate functions in pcms as well.
|
@Sichao25, @Fuad-HH just made an issue for that here: SCOREC/pcms#351 If we want to reuse the Omega_h functionality in PCMS (which is beneficial), we need to make sure to include the |
|
@cwsmith, Could you please run the tests again? |
|
/runtests |
|
Test Results:
|
|
@cwsmith this looks good to me. I had asked Fuad to add tests for strided numpy arrays which he did. |
Uh oh!
There was an error while loading. Please reload this page.