Andrey Prokopenko
Andrey Prokopenko
> I am a bit surprised by the variability of the construction timings in the results you posted. Is it something you have been seeing before? Something else was running...
> The culprit seems to be not specifying `-fp-model=precise`. `Release` with `-fp-model=precise` works. `RelWithDebInfo` with `-fp-model=precise` fails.
> `Release` with `-fp-model=precise` works. > `RelWithDebInfo` with `-fp-model=precise` fails. @masterleinad Is this something you can look at?
Related to kokkos/kokkos#6743
Should this not be a user's choice through a parameter? I.e., make some parameter ("fix zero pivots"?) true by default and allow a user to set it. If it's false,...
> ILUT always fixes zero pivots if you give it threshold parameters. Why? This it not a textbook ILUT, is it?
I tried to see anything in ArborX. I ran DBSCAN driver with 4.6.0 and 4.7.0 on Frontier (AMD gpus), and the differences are in the noise. We only use 1D...
I think in general it is a good idea. The challenge is that there are few good update options. If we assume that the number of primitives does not change,...
I did a quick and pretty dirty implementation of the update procedure that I was talking about (available [here](https://github.com/aprokop/ArborX/tree/bvh_update)). I ran the standard HACC problem on MI250X, first calling the...
> I might have missed the forest for the trees here 😄. Applying the prior bias "construction slow, reuse good" was not necessarily applicable. Now understanding better the profiling results...