Scientists are using artificial intelligence and large language models to rewrite old code in modern languages

Uncategorized
102

The problem, as someone pointed out earlier, is the lack of connection and development between the standard committee and a compiler to implement all the features that are proposed or accepted.

So we are left with the vendors adopting and implementing things, this is why do concurrent is implemented differently in the Nvidia, amd, cray, and gnu compilers.

Fortran can do GPUs but portability across vendors is not a given. Kokkos attempts this and mostly succeeds because C++ got premier support during the first big GPU push.

103

I think that sums up the issue. A language without good handling of unstructured data is not a modern language, period. As far as I know, LFortran is the only compiler that does provide some implentation of a hashmap, although not standardized.

General purpose languages won.

1 Like
104

Not rewrite but modernize enough to provide a bind(C) interface and a Python ctypes binding. Claude struggled with understanding fortran rules (for example regarding common block declarations). I would say we managed to get it done the 5th time, but it required frequent corrections. It was an old code from '82 with dozens of common variables and LLM was very good at following all the dependencies, but tight guidance over architecture of the solution was needed. “Autonomous coding agents” are incovenient to me for one reason: they make decisions I disagree with requiring rollback and waste of tokens, I prefer much tighter control and checking their every step.

105

Here’s another of my AI-assisted C++ to modern Fortran conversions: Degenerate Conic | Brouwer-Lyddane Elements

5 Likes
106

Beautiful. I can see you are using your own pyplot-fortran, which is constructing a Python string and evaluating it in Python (right?).

I figured, let’s translate Matplotlib to Fortran, to be more consistent with this thread:

It is 100% pure Fortran, the usual pyplot-style API, and it has svg, png, pdf (and even gif!) backends, the backends are all in pure Fortran, they generate the svg text, or png/pdf/gif binary. I embed the glyphs of a font directly into Fortran sources, support latex-style math rendering, etc.

I needed something to use in the JupyterLite notebooks that LFortran now supports, so since it runs in the browser and must be fast, I needed to generate svg/png directly in pure Fortran.

Anyway, it was super easy to do this thanks to AI, so I think we can now have the full ecosystem from plotting, file formats everything in pure Fortran.

3 Likes