Small speedups - Githubissues

sbailey commented 6 years ago

This PR includes simple three O(1%) speedups that I found while reviewing how GaussHermitePSF._xypix in search of larger speedups.

Keep self._polyparams as a dict instead of an astropy.fits.Header for faster element lookups
Cache the minimum and maximum wavelength per fiber so that it doesn't have to be recalculated for every call to xypix to see if we are off the edge of the CCD.
In PSF.projection_matrix, fill a transposed version of A (faster) and then transpose it once at the end.
- there are likely better ways than filling A and then converting to a csr sparse matrix. This could be an area for future study.

lastephey commented 6 years ago

testing in progress

lastephey commented 6 years ago

Typical test case on Haswell runs in about ~4:40, this ran in about ~ 4:14. Limited statistics, only tested once.

Transposing only once at the end seems to be a good strategy.

weaverba137 commented 3 years ago

Is there a reason the branch small-speedups has not been deleted?

desihub / specter