(Re-)Optimizes Zernike computation

[mperrin]

Issue by josePhoenix
Monday Jul 06, 2015 at 19:26 GMT
Originally opened as mperrin/poppy#99

---

Using the LRU cache decorator from Python 3.2+ (and a vendored copy in poppy.vendor.lru_cache for back-compatibility), cache individual Zernike polynomial arrays to speed ZernikeWFE computations.

This adds a test case that generates an input wavefront and compares calling zernike1() with rho and theta arrays (the way previous implementation did, but alas un-cacheable) to a new style where we pass in enough information to reproduce the rho and theta aways within cached_zernike1().

~~The performance improvements ended up being "modest" (that is, ~2x), so there may be further productive profiling to be done.~~ Performance improved by a factor of ~6x for code that used zernike1 a lot.

An obvious next step would be to cache ZernikeWFE phasors with an internal instance attribute and some way to mark it dirty when self.coefficients or self.radius are assigned to (or make those attributes read-only).

---

josePhoenix included the following code: https://github.com/mperrin/poppy/pull/99/commits

[mperrin]

Comment by josePhoenix
Monday Jul 06, 2015 at 20:57 GMT

---

The WIP WFIRST field dependence stuff hits zernike1 ~220 times in a monochromatic calculation, so it's a good realistic test-case to see if we made this any better.

• In the uncached case this is 55 seconds for a 10 wavelength calculation incorporating field dependent aberration.
• All else being equal, caching makes the first calculation ~17 seconds, and subsequent calculations are ~9 seconds.

[mperrin]

Comment by josePhoenix
Tuesday Jul 14, 2015 at 15:31 GMT

---

As discussed, refactored Wavefront.coordinates into static methods for image- and pupil-plane coordinates, then used those in cached_zernike1.