For low ell, the current 3j computation might be alright, but the error grows exponentially, and should be nonsense for ell~50. The Rasch and Yu algorithm could be made quite fast, I think, so as long as the storage isn't killer, I could just do that. It's certainly fine for ell<30 (couple hundred kb). Starting to get ugly for ell<=64, and pretty bad for Barry's case (ell<=80; ~30MB).
My newer spherical package has a Wigner3jCalculator object that should be faster and accurate to far higher ell values (probably at least 1000). I won't be implementing that method in this package.
For low ell, the current 3j computation might be alright, but the error grows exponentially, and should be nonsense for ell~50. The Rasch and Yu algorithm could be made quite fast, I think, so as long as the storage isn't killer, I could just do that. It's certainly fine for ell<30 (couple hundred kb). Starting to get ugly for ell<=64, and pretty bad for Barry's case (ell<=80; ~30MB).