Fixed bary_corr issue in and refactor of extract.py to remove repeated code #943
Conversation
|
I showed the straight sum of differenced pixels, here I show the more useful sum of the abs for all pixels: Results are unsurprisingly higher: r1 z1 but keep in mind this is a sum over 500 * 2881 = 1440500 differences. |
|
I'm still a bit mystified about why there are differences at all, but I agree that they are very small. I also checked the flux difference normalized by the errors, and the max differences are 0.004 sigma. I also ran just 2 bundles and got identical fluxes. So it's a bit odd, but I think this is good enough to merge. Side note: there is a bug in the MPI extractions that predates this PR, where it writes out 500 spectra regardless of the --nspec parameter. Laurie, Mark, and I are working on another extraction refactor in prep for GPUs so we can leave this for now. |
Addressed issue #941 . Also goes ahead and refactors extract.py to remove repeated code.
Tested using different numbers of tasks and cpus-per-task (
srun ... -n -c).Tested with night 20200314, exposure 55445.
Barycentric correction is applied appropriately now. Below I show examples in cam=z1.
Normal
-n 20 -c 2:Using
-n 2 -c 32:No mpi:
The two mpi versions have bitwise-identical data.
The non-mpi version does differ to a minor extent, I believe this is due to the precision loss in the read/write operations in the mpi version which writes temp files before consolidation.
Below I show the sum of all differences in a given HDU (over all values in the HDU.data).