Repository navigation
Assign device for calculating minimum distances as GPU, if available. - #43
messerlyra wants to merge 1 commit into
Conversation
messerlyra
left a comment
messerlyra
left a comment
There was a problem hiding this comment.
Not sure how, but hippynn_interface.py had a duplicate call of calculate_min_dists. The first call still used the old format where kwargs were not named explicitly. Removed that call and included the device specification in the second call.
|
You must trim the database (database.arr_dict[curK] = database.arr_dict[curK][min_dists_flag]) before setting natoms, otherwise the natoms array will be of dimention miss-match with the remainder of the database. |
In ASE, empty cells made with a defined cell default to pbc=False. In simple_condensed_phase_builder_task, this has been updated to explicitly enforce pbcs. This also includes edits from pull request #43. The second database cleaning has been removed from hippynn_inerface and the gpu device has been added for neighbor finding.
|
Edits fixed and included in #44 |
|
@bnebgen-LANL OK, now I get what your comment means. You were answering my question about why we call calc_min_dists twice. Basically, we need to trim the array based on distances before calculating n_atoms. That makes sense to have calc_min_dists before n_atoms. It appears that PR #44 removes the second distance trimming, but it still calls calc_min_dists a second time, which seems unnecessary since min_dists is never used again. Also, the other key part of this PR was to set device=device_cuda in calc_min_dists. But it appears this gets resolved in PR #44 . Is that correct? |
Calculate_min_dists was defaulting to CPU. Now explicitly assigned to GPU.
Also, removing duplicate call of calculate_min_dists