Alan cleanup #234
Alan cleanup #234
Conversation
jvshields
requested review from
andrewfullard,
DeerWhale and
Knights-Templars
Correct. You used to specify a range that you wanted to search around for all lines, which wasn't great because you need a wide range for large lines which was wasteful for small ones. Now it's adaptive, though it does do a minimum search region of +/- 10 spectral pixels. Though this makes me realize that perhaps the better way to do this is to use the d_nu that already has to be calculated and maybe impose a minimum search region of 1 angstrom or so. EDIT: I implemented this. |
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
| < broadening_range[frequency_index] | ||
| for line_index in numba.prange(len(line_nus)): | ||
| line_nu = line_nus[line_index] | ||
| thread_id = numba.get_thread_id() |
There was a problem hiding this comment.
Should doublecheck if numba.get_thread_id() is safe/reliable to be the index.
This should massively improve the speed of large simulations. Rather than calculating the difference between each the frequency of each line and each frequency grid point, this does a binary search on the wavelength grid to find the closest frequency array index of the line, and then calculates how many indices out we should consider the line for.
Currently for this I'm just doing (gamma + doppler width) * 20 * alpha_of_line/frequency_bin_width, or 10 bins to each side of the line. This should probably be thought out more, but looks like it gives a good first answer to consider each line out to where appropriate.
This also implements parallelization of this calculation.
As proof of implementation, simulations that used to be far too big to fit in memory on my laptop (6e5 wavelength points with 3e5 lines) now run in <5 minutes.
It was hard to get good time estimates on moria before this was made, but simulations that could hang for between 30 minutes and an hour and a half when they had to use swap memory can now finish in ~5 to 10 minutes and don't appear to have memory problems.