Hacker News Reader

Comment by PaulHoule | Hacker News Reader

Even if you are using python, R or java you are still best off linking to matrix subroutines in fortran if you want to solve or diagonalize, get eigenvalues and such. This is not just about raw performance but also correctness, ecology of roundoff errors, etc.

View on HN

For Python, at least, the most popular numeric computation library (NumPy) uses a fair bit of Fortran under the hood. I'm not certain whether the matrix operations do, but it certainly wouldn't surprise me.

Doesn't NumPy just use BLAS and LAPACK for matrix operations?

The vast majority of BLAS and LAPACK are written in Fortran (at least, the ones provided by netlib.org). Github also reports that OpenBLAS is around 50% Fortran, so there's quite a bit there too [1]. Usually when you call BLAS functions from C you're really calling a cBLAS wrapper to some Fortran function. This is generally why you see -lgfortran -lf77blas when you compile programs that require numerical tools.

[1] https://github.com/xianyi/OpenBLAS

At least in OpenBLAS, most (all?) of the BLAS routines are C functions that call kernels written in assembly, but many of the LAPACK routines are the Netlib reference Fortran versions.

Yet, I root for D and its GLAS library [0]. It is a young project, but it beats the usual suspects for single-threaded matrix-multiplication. Stuff like "solve or diagonalize, get eigenvalues and such" is not written yet. It is very promising for a one-man project, though. We'll see in 10 years, I guess.

Fortran certainly has a mythical reputation for performance, but it should be challenged every once in a while.

[0] http://blog.mir.dlang.io/glas/benchmark/openblas/2016/09/23/...

Reply on news.ycombinator.com