Comparison with other computations

rodluger / Limbdark.jl

Analytical transit light curves for limb darkened stars

MIT License

14 stars 4 forks source link

Comparison with other computations #9

Closed ericagol closed 5 years ago

ericagol commented 6 years ago

[x] Compare accuracy with Mandel & Agol computation.
[x] Compare accuracy and speed with Gimenez' code.
[x] Compare accuracy and speed with Kreidberg's quadratic limb-darkening (both unbinned, and integrated with adaptive binning).
[x] Compare with non-linear limb-darkening law (to what order do we need to go to achieve same accuracy of non-linear limb-darkening?).
[x] Compare with Pàl’s quadratic limb darkening with analytic derivatives. (Right now test_transit_poly_struct.jl matches ntiq-fortran which was compiled with -O3; both take 0.5 seconds to compute 10^6 models, with the same impact parameters for both. We may be able to save some time by reusing elliptic integral computations.)
[ ] Compare with Maxted's power 2 limb-darkening law.
[x] Comparison with EXOFAST.

rodluger commented 6 years ago

@ericagol Here's how your code compares to batman: you're killing it!

@ericagol: Is this with the Python version or the Julia version? @rodluger: This is with the Julia version. The starry implementation is a bit slower than batman because of the overhead from all the bells and whistles. @rodluger: Oops, I screwed up on this one. See my post below...

rodluger commented 6 years ago

I'll code up the comparison to the nonlinear model next.

ericagol commented 6 years ago

@rodluger I already did one version of the comparison to the nonlinear model - that's what the occult_nonlinear.jl does. But, it would probably be good to explore a range of parameters.

ericagol commented 6 years ago

And, batman doesn't do derivatives.

rodluger commented 6 years ago

@ericagol I screwed up on the above plot: I was computing the batman transit 10 times, but forgot to divide by 10 when taking the average. So your code is actually about 5 times slower than batman. But you're right, you're computing derivatives. I tried setting the gradient switch to false when instantiating the transit instance, but that didn't change the timing, so I think you're still doing all the math in the background.

rodluger commented 6 years ago

Also, when we both compute gradients, you are a factor of ~3 faster than me. I need to look into why this is.

ericagol commented 6 years ago

@rodluger Yes, I am working on turning off the gradient computation (although I probably won't get to that today). That's too bad, but I'm glad you figured that out. It is also worth checking if there is multi-threading in batman; that explained the speedup I had found when using the BLAS gemmv routine.

ericagol commented 5 years ago

Going to skip comparison with Maxted's power-2 limb-darkening.