Open bd-j opened 5 years ago
might be worth looking at RectBivariateSpline instead of SmoothBivariateSpline cc @stacchella
We now use RectBivariateSpline which has better behavior. Timing tests compared to the pixel computation kernel remain to be done.
Right now generating a proposal for a single source (even for the GPU) involves evaluating several bicubic splines and their derivatives using a loop over scipy.SmoothBivariateSpline objects in the CPU code. This is expensive (~150 microseconds for ten Sersic radii) and should be sped up to make sure proposals can be generated fast enough for the CPU.