[Draft] Add Multigrid support #606
Conversation
| @@ -1920,6 +2227,13 @@ def _apply_lineshape_LDM( | |||
| if ``"min-RMS"`` weights optimized by analytical minimization of the RMS-error. | |||
| Otherwise, weights equal to their relative position in the grid. | |||
|
|
|||
| Other Parameters | |||
| ---------------- | |||
| wavenumber_group: (int, int), or ``None`` | |||
| abscoeff_common.append(abscoeff_grid) | ||
|
|
||
| # HACK STOP POINT For RADIS Community paper. Plot different abscoeff for all grids: | ||
| if False: |
|
For my understanding, is this basically what SPEARS does? if not, how is it different? |
|
It is different from SPEARS. SPEARS basically builds one grid per line, and combines them ; therefore the grid is irregular which is not compatible with the DIT/DCS method where we need a regular grid to apply the convolution. Quoting our paper draft :
|
save some memory on very large ranges / low pressures (different from DIT Sparse 2021 algorithm where we apply masks on the grid of spectra after the DIT step. Here the sparse operation happens before DIT. ) next : multi grids
implement multigrid by adding the corresponding lineshapes add tests for overlapping-ranges, fix the function (wip) 1st version of multi-grid boundary effects corrections (when gluing multiple grids) Add fast interpolation of grids (which was the new limiting factor) Results : 2* faster than normal on an optimization=None example
(runs and returns a spectrum, but not all grids are aggregated yet)
(accuracy ~ok, still some interpolation problems) , speed is increased up to x10 in Sparse LDM mode (for cases with low number of lines where LDM_Distribute_lines is not the limiting factor), and slower with pure LDM
…/ wstep too large fix adaptative grid in non-eq calculations
…ion, make a numba version, make a Cython version numba version is the new default : ~ 10x faster
…computation of all spectral arrays after a SpectrumFactory calculation
…ization=None version without adaptative grid)
erwanp
force-pushed
the
2208/add-multigrid
branch
from
bc59427
to
e5be046
Compare
|
@dcmvdbekerom I updated the test conditions in the main PR tests ; so far I see mostly an improvement over a regular grid ; and no penalty in the worse conditions (i.e. dense spectra like our favorite 4.2 µm CO2 band). |
Multigrid support in RADIS
tl;dr
optimisation=None) & regular grid output : 4x speed-up in CO HITRAN exampleoptimization=simple)self._multisparsegridand other hidden variables)(wstep, truncation)parameters of each grid (and the number of grids. Maybe 2 are enough)optimization=None)This will be particularly useful for "isolated-lines" spectra, i.e. :
👉 Combined with DIT (for "dense spectra") and sparse-DIT (for "dense-by-part spectra") it should make RADIS the ultimate, fastest code for every kind of spectra.
Other changes:
get_overlapping_rangesFull-range calculations, low number of lines
With Optimization=None (no LDM) : x4
CO-HITRAN, 500 - 10,000 cm-1
Code : see at the bottom
Here are the performance details :
We see a x25 speed-up in the Voigt broadening part, and an overall x4 speedup in the full Lineshape broadening step (including aggregation of lines & interpolation of grids).
(Fig.1 13/08/23):
Grids are automatically refined around the line centers . Each grid is regular (i.e. constant wstep) which (will) allow the use of the DIT Algorithm. (Fig 2.)
Same, zoomed-out to see the multiple (3) and discontinued (multi-groups) grids (Fig 3.)
Code:
With Sparse LDM : x4
@dcmvdbekerom same calculations with LDM now : x6 speed-up ; and x10 if looking only at the line broadening part (Fig 5.)
Code : same as above, with the following changes:
Note : the example above was ran with CO HITRAN, with limited number of lines. For CO HITEMP below we see that the multigrid accelerates the LDM_convolve step, but does not change the LDM_Distribute_lines step (obviously)
With (non-sparse) LDM : x4
CO HITEMP version (same as above)
Acceleration by a factor of 5 (but non-sparse LDM is still very slow for such a large range)
Appendix - Adaptative grid in worse conditions
(non-sparse) LDM in dense regions : x1
Adaptative grid is developed for isolated-lines spectra. Below we deal with a very dense spectral region, i.E. the 4.2 µm (2300 cm-1) band of CO2
Adaptative grid doesn't accelerate the calculations, but it doesn't slow it down neither