Computing partials for all segments and triangles at once

[tbenthompson]

At the moment, the segment and triangle partials are being computed one at a time. This is probably pretty slow. It's worth profiling (with something like line_profiler or pyspy) first to check before investing effort to change this, but I'm guessing this is something like 10x slower than it would be to pass all the necessary observation points and source triangles to cutde.disp_matrix. This is basically just standard Python/matlab "vectorize your code" advice.

Concretely, passing only a single triangle at a time removes a lot of the potential for parallelization and for amortizing operations over many source triangles.

As a point of reference, I remember getting something like 5 million TDE source-observation point computations per second running on CPU with cutde. So, if you do a line_profiler run and see performance substantially worse than that, the problem is likely to be the individual calls.

Both here:

celeri/celeri/celeri.py

Line 1125 in dde0941

for i in tqdm(

and here:

celeri/celeri/celeri.py

Line 1851 in dde0941

for i in tqdm(

[brendanjmeade]

I think the challenge that I've seen here is that I do a different map projection for each of the triangles. This essentially means that for each triangle there is a whole new set of observation coordinates because they go through the map projection too. That's not a linear transform so I don't think I can just factor it out. Maybe it's worth considering packaging all of the projected TDEs and projected station coordinates together anyway and doing a lot of unused calculations and then just extracting the subsets we need?

[tbenthompson]

Ah! That's exactly the use case for cutde.halfspace.disp: When you have a bunch of pairs but don't need to calculate all pairs.

https://github.com/tbenthompson/cutde#simple-pair-wise-tdes

[brendanjmeade]

That's a great suggestion and I'll leave this issue open so that I can try it out. Thanks for the idea!