GPU Usage for Quasisep Kernels

[TylerFair]

Hi,

I was interested in PR #210 as I'm working on a multiwavelength transit fitting routine that involves simultaneously detrending each light curve. For reference, with just linear detrending on a GPU, this routine takes a remarkably short time (~20 mins for 2000 LC's on an A100), but implementing a GP detrending such as with the quasisep module encounters the issue where a CPU is required or else its hundreds of times slower.

I know the PR is not actively being worked on but my question is if the current changes in PR #210 could be implemented to boost GPU performance or if a lot of tweaking would be required.

Thanks!

[dfm]

Good question! I think it would be a major research project to get something like #210 working for any real problems. Even if you did the work to get the basic ops working, I'm not very optimistic that you'd get good performance out of the parallel algorithms - these associative scan approaches are theoretically compelling, but I haven't seen great performance in practice!

I don't know enough about your specific project to make concrete suggestions, but in my experience, for many shapes of problems, using the quasisep algorithms on CPU outperforms just about any GP model on GPU!