search

wavefunction91 / IntegratorXX

Reusable DFT Grids for the Masses
BSD 3-Clause "New" or "Revised" License
13 stars 9 forks source link

Extend tests to cover a larger range of quadrature points and rules #52

Closed susilehtola closed 1 year ago

susilehtola commented 1 year ago

Instead of checking all tests from 2 to 100 points, I decided to extend the range of checking the orders, while still keeping the time to generate the reference data at a reasonable amount and the number of test cases below the roof.

The key here is to include both odd and even numbered rules, since they tend to behave differently. This is easily achieved by using an odd increment: 1, 3, 5, 7, 9, ...

The Chebyshev tests are failing since #42 is not yet solved.

susilehtola commented 1 year ago

@wavefunction91 I've rebased to master, and implemented tests for Chebyshev1 and Chebyshev2 which pass.

If it would please you, please set your workstation to generating the Legendre and Lobatto rules. Maybe you can also think about ways to parallellize the generation? Generating the tests takes quite a long time... Fortunately it only needs to be done once, though.

wavefunction91 commented 1 year ago

Running now, re parallelization, should be easy enough, but I'm not great at python. Would definitely accept a PR on it. It should be done overnight though, not a huge priority.

wavefunction91 commented 1 year ago

NVM, figured out how to parallelize it. Will check this in and PR the new tests / parallel generator

wavefunction91 commented 1 year ago

After having a workstation work on this, in parallel, for over a few days. I think the requested test granularity is too fine. At n over ~500, things start taking 5-10 hours to run

Generating test for GaussLegendre with 812 points ... (51311.454052632)

I fail to see the utility of testing to this granularity, I can see maybe testing every 100 - 200 after n = 500. but even then, we should have a demonstrable need for why we should expect the code to break down to add it as a unit test.

susilehtola commented 1 year ago

Oh wow. Well, one can toy with the two parameters to get a good compromise between granularity and the total number of test cases.

I was mainly interested to see how far the rules stay accurate, since my benchmarks with https://github.com/JFurness1/AtomicOrbitals/ using SciPy's roots_legendre show pretty poor results already with a hundred or so points.

wavefunction91 commented 1 year ago

@susilehtola Gauss-Lobatto is running now

wavefunction91 commented 1 year ago

@susilehtola is this g2g?

susilehtola commented 1 year ago

@susilehtola is this g2g?

I think so