search

desihub / specter

A toolkit for simulating multi-object spectrographs
Other
8 stars 7 forks source link

avoid at-edge-of-ccd extraction crash #84

Closed sbailey closed 2 years ago

sbailey commented 2 years ago

This PR fixes desihub/desispec#1790 with an extremely corner case extraction bug. If the wavelength solution had been 0.0001 Angstrom different in either direction it would not have crashed. The underlying problem comes from the logic handling whether the requested spot is on, off, or partially on the CCD, and rounding differences in the round trip y -> wavelength -> y . The solution in psf.xyrange was to use psf._xypix instead of psf.xypix to avoid one extra level of "am I on the edge of the CCD?" check.

I tried to add a unit test to catch this, but the trimmed down PSFs in the test suite aren't sensitive to this exact problem so they passed anyway even without the fix. None-the-less, the additional test code does fail with the full PSF from himalayas/exposures/20220221/00123283/psf-z0-00123283.fits if not using the fix in this branch. Having added the test, I left it in there anyway as some partial future insurance against a new bug.

Code to reproduce original failure without requiring full MPI and desi_proc like original report:

cd $CFS/desi/spectro/redux/himalayas
desi_extract_spectra -w 7520.0,9824.0,0.8 --psferr 0.1 --barycentric-correction \
    --nsubbundles 5 --specmin 475 --nspec 25 \
    -i preproc/20220221/00123283/preproc-z0-00123283.fits.gz \
    -p exposures/20220221/00123283/psf-z0-00123283.fits \
    -o $SCRATCH/frame-z0-00123283-no-badcolumn-mask.fits.gz

That fails with specter/main but passes with this branch.