gh-977: Port remaining functions in fields.py

[connoraird]

Description

Ports the functions left in fields.py which haven't yet been ported

Closes: #977

Changelog entry

Added: Full support for the array-api to fields.py

Checks

• Is your code passing linting?
• Is your code passing tests?
• Have you added additional tests (if required)?
• Have you modified/extended the documentation (if required)?
• Have you added a one-liner changelog entry above (if required)?

[paddyroddy]

Current regressions seem quite large

test_generate_grf[numpy-None-False] (0001_8fcf233) - Field 'mean' has failed PercentageRegressionCheck: 25.512178944 > 5.000000000
test_generate_grf[numpy-None-True] (0001_8fcf233) - Field 'mean' has failed PercentageRegressionCheck: 26.795189419 > 5.000000000
test_generate_grf[numpy-1-False] (0001_8fcf233) - Field 'mean' has failed PercentageRegressionCheck: 11.350133326 > 5.000000000
test_generate_grf[numpy-1-True] (0001_8fcf233) - Field 'mean' has failed PercentageRegressionCheck: 10.812015623 > 5.000000000
test_generate[numpy-None] (0001_8fcf233) - Field 'mean' has failed PercentageRegressionCheck: 42.996966146 > 5.000000000
test_generate[numpy-1] (0001_8fcf233) - Field 'mean' has failed PercentageRegressionCheck: 116.517332491 > 5.000000000

[paddyroddy]

Please can you make a comment on the regressions?

[paddyroddy]

This PR looks like it will have regressions to should have the needs-2-reviewers label

[paddyroddy]

Okay, this looks in a good place now. Five things for me:

• Need new line in docstring to match numpydoc
• Move yield comment
• Fill in checklist
• Add need-2-reviewers label + Nicolas (if regressions)
• Comment the regressions and check if okay

[connoraird]

Current regressions appear to only be in generate. That may make sense as generate indirectly calls a lot of the functions which have been ported so the total regressions would accumulate.

--------------------------------------------------------------------------------
Performance has regressed:
	test_generate[numpy-None] (0001_20ea738) - Field 'mean' has failed PercentageRegressionCheck: 30.930359092 > 5.000000000
	test_generate[numpy-1] (0001_20ea738) - Field 'mean' has failed PercentageRegressionCheck: 68.876285110 > 5.000000000
--------------------------------------------------------------------------------

[paddyroddy]

Current regressions appear to only be in generate

generate has been a consistent problem. Looks like it calls cltovar, getcl, _generate_grf. getcl, _generate_grf both have benchmark tests. Turns out cltovar is actually from transformcl. It appears to be array compatible in itself.

Does this mean the regression is actually coming from the for loop within generate?

[connoraird]

Turns out cltovar is actually from transformcl. It appears to be array compatible in itself.

It looks like it has been ported but it actually doesn't work with array_api_strict due to this error TypeError: Only floating-point dtypes are allowed in __truediv__. This is because this is the code which produces an ell.dtype of array_api_strict.int64

xp = array_namespace(cl)
ell = xp.arange(cl.shape[-1])
return xp.sum((2 * ell + 1) / (4 * xp.pi) * cl, axis=-1)

A solution would be the following change.

   xp = array_namespace(cl)
- ell = xp.arange(cl.shape[-1])
+ ell = xp.arange(cl.shape[-1], dtype=xp.float64)
   return xp.sum((2 * ell + 1) / (4 * xp.pi) * cl, axis=-1)

Can we make this change? I've raised a PR to fix this - glass-dev/transformcl#6

Does this mean the regression is actually coming from the for loop within generate?

It could be, however, I think my point of the regressions from the other functions accumulating is still possibly true. Even if the other functions don't fail the regression tests, they could be regressing by 4%. then the total regression of all of them could add up to more than 5%. Or is that not how percentages work? 🤔

[paddyroddy]

Can we make this change? I've raised a PR to fix this - glass-dev/transformcl#6

Yes, I don't see why not

It could be, however, I think my point of the regressions from the other functions accumulating is still possibly true. Even if the other functions don't fail the regression tests, they could be regressing by 4%. then the total regression of all of them could add up to more than 5%. Or is that not how percentages work? 🤔

I could buy that argument

[connoraird]

Still seems to just be the known generate regression and a potential bit of noise.

--------------------------------------------------------------------------------
Performance has regressed:
	test_iternorm_k_0[numpy] (0001_4c89818) - Field 'mean' has failed PercentageRegressionCheck: 5.224544781 > 5.000000000
	test_generate[numpy-None] (0001_4c89818) - Field 'mean' has failed PercentageRegressionCheck: 33.989695641 > 5.000000000
	test_generate[numpy-1] (0001_4c89818) - Field 'mean' has failed PercentageRegressionCheck: 68.915726817 > 5.000000000
--------------------------------------------------------------------------------

[paddyroddy]

Making an executive decision to accept that this regression is inevitable for now. In the example notebook, the matter fields and created once using generate. They then are iterated over within a for loop. The net effect of generate is much smaller than functions like shear_from_convergence, uniform_positions, etc.

cc: @ntessore