Substantial Python loop overhead for VBBL
kmzzhang opened this issue · comments
Currently, the binary lens light curve is calculated in a python loop where each point is evaluated separately.
https://github.com/rpoleski/MulensModel/blob/master/source/MulensModel/magnificationcurve.py#L421
When using VBBL as the finite source method, the python loop overhead can slow things down by up to ~7 times compared to VBBL's own python wrapper, where the loop occurs in C++. This most apparent when VBBL is used for the full light curve (which itself decides automatically whether to trigger full FS calculation). Perhaps could aggregate points that use VBBL and move the loop into C++? I considered making a pull request but I realize this may involve refracting larger parts of the code.
@kmzzhang We are about to do a major refactor for Version 3, so we can add this to the list. In particular, we are talking about having subclasses for models, which could include creating MagnificationCurve subclasses such as MagnificationCurveVBBL()
We don't have to wait till v3. This seems relatively easy thing. I'm guessing we did it on the epoch-by-epoch basis because it's easier to pass a specific number of floats between python on C++, rather than arrays of unknown sizes.
Thanks @kmzzhang for bringing that up!
You're welcome. I believe this also applies to 2L1S point source when the SG12 solver in VBBL is used.
Since the finite source method is specified in time intervals in MulensModel, perhaps one way to refractor the code is to aggregate epochs by those intervals and calculate the magnifications?
Yes, that's what I plan to do.
@kmzzhang Can you please share how you pass pointers/vectors of floats between python and C++? I see that there are different approaches to PyArg_Parse*() functions and would prefer not to re-invent the wheel, if you have already done so.
@rpoleski Apologies for the late reply. I don't have a particular way of doing this, but perhaps you could use Valario's way of making python wrappers: https://github.com/valboz/VBBinaryLensing/tree/master/VBBinaryLensing/lib. He has a python wrapper for the BinaryLightCurve function that takes array of times (ts). One could easily modify this function (and its wrapper) to take arrays of source locations y1s and y2s instead of times. Everything else would be the same.