imurray / chol-rev Goto Github PK

Hi,
Awsome code!
For the environment in the title, I had to do two small changes to make the system compile:

1. In chol_rev.c line 23, changed int64_t to long long. (I guess int64_t isn't included in stdint.h.)
2. In chol_rev.c line 30, changed dpofrt to DPOFRT. (I guess the linker is case-sensitive.)

Cheers,
David

I suggest adding topics such as cholesky-decomposition, cholesky, automatic-differentiation in the About section.

(This isn't really an issue with the code, just a pointer to some other code in case it's interesting. Feel free to close it immediately!)

We wrote a Cholesky reverse-mode routine in autograd (and based it on one in GPy). The code is in cython, which generates very nice C code and is a bit easier to build with Python, especially since it grabs BLAS function pointers from scipy at runtime and so doesn't have any other compile-time dependencies. I also wrote a loop instead of calling dscal because a modern C compiler can generate code at least as good as some level 1 routines.

I got a bit stuck on writing a routine for the second-order derivative. Smith 1995 doesn't seem to spell out a pure reverse-mode computation for the second derivative.

Your code here looks really nice, and being in Fortran it's not limited to just Python. Having a "LARMPACK" as you describe would indeed be awesome.