Conversation
- Option check_same=true by default now, as the dry-run, then allocating, then re-running approach checks if the same number of neighbors are found for each target, in each search. - For searches without a dry-run, this check would previously always fail, so this allows a disable of that check.
- If Compadre_USE_ARBORX is not set to on, then default back to nanoflann search routine
…for floats - Added SEARCH_SCALAR_EPS conditional on which neighbor search routine is being used (1e-7 for ArborX and 1e-14 for Nanoflann) because ArborX searches using floats and Nanoflann searches using doubles. - Modified tests and search routine to make the various casts explicit.
…changes API) - ArborX neighbor search will ignore all do_dry_run bool flags and do full neighbor search with no recursive calls - Nanoflann is broken until next commit when dry-run/real-run is built into recursive calls
- Switches itself from dry-run to storing results (with resizing) all through a single call - Changed size_t returns from search methods to void
- Replaces internals of 2D searches for Nanoflann
- Replaces two definitions with one for both radius and KNN search - Removed parameter do_dry_run from all rank 2 layout searches
- Changed a type to a decltype
|
Overall, I think the code in But this are small potatoes. |
|
@kuberry Just a quick question: do you plan to use ArborX path for any real applications? Or is the plan right now to simply have it in if a need arises? I would be really interested to see the benefits if its the former, and see what can be improved. |
Compadre is a Trilinos package as well, so it is hard to say who all of the external users are. It is used in a number of projects at Sandia. Many of them that relate to circuit simulation are calling it in serial, however (query a single current for a voltage, query a single current for a voltage, ....). |
|
Is there any desire to resurrect this effort? We added a lot of new feature to ArborX since we last talked, including multi-dimensional points with user-specified precision. |
Add ArborX as an option for neighbor search, enabling GPU accelerated neighbor search.
What has been done:
What remains:
This PR changes the API for neighbor search with respect to the location of
is_dry_runbecomingdo_dry_run, getting a default value of true (never to be changed manually by a user).