magicl VS numcl

(Link didn't work for me)

https://github.com/quil-lang/magicl/blob/master/src/high-lev...

use MAGICL. [1] It is optionally and transparently accelerated by BLAS/LAPACK.

[1] https://github.com/quil-lang/magicl/blob/master/doc/high-lev...

Maybe have a look at how magicl does this?

This is essentially the first option, which is already supported by MAGICL by loading MAGICL/EXT-LAPACK [1].

[1] https://github.com/quil-lang/magicl#extensions

I think what the magicl team has done is brilliant - allowing multiple implementations is awesome.

Second up is magicl, especially useful if performance is a concern. This might not be as extensive as numcl, but it's been battle tested in the industry over the last decade or so. Because this uses generic functions, so long as you are using not-very-small arrays, performance should not be a concern for you. And even if you are, you could write your own functions that use the low-level functions that magicl's backends define. Otherwise performance can be at par with numpy.

This Dev How-To describes (I hope in enough detail) how to add these specialized routines to MAGICL.

I agree... and I do don't want be the person who has not rallied. I just took a look at guicho's issue from 2019. And here, you yourself have admitted that the high level interface is less than ideal and needs more work. However, the very point that magicl is an industry standard could imply that potentially radical backward-incompatible changes can be hard. But, honestly, I want to discuss this, time permitting!

Common Lisp is a great language to build new tools for data science, but currently has pretty awful library support existing data science workflows. Common Lisp is sorely lacking in high-quality statistics, plotting, and sparse arrays. There’s been a long work-in-progress library to bring flexible and high-performance linear algebra to Lisp, but it needs more contributors.

Is it me or numcl is faster than magicl? Matrix multiplication on magicl with pure lisp backend is

Interesting, I will, thanks! I am aware of numcl for CL, but I don't think it is "there" yet :).

The first that comes to mind is numcl. This works if (i) performance is not seriously a concern, (ii) you are not annoyed by julia-like JIT/JAOT compilation delays, (iii) copy-based slicing won't be a performance issue for you. To be fair, limitation (i) might be overcome by writing a better (simd-based) backend for numcl. numcl is fast, it compiles to fairly good code, but simd can boost the performance by another 4-8 times or so.

am I missing something that already exists (numcl / Alexandria / core language, etc?)

>Either implement numpy equivalent on your own or half of your code is data massaging data between libraries

I haven't tested this but here you go:

https://github.com/numcl/numcl

Personally, I've been relying on the stream-based method using py4cl/2, mostly because I did not - and perhaps do not - have the knowledge and time to dig into the CFFI based method. The limitation is that this would get you less than 10000 python interactions per second. That is sufficient if you will be running a long running python task - and I have successfully run trivial ML programs using it, but any intensive array processing gets in the way. For this later task, there are a few emerging libraries like numcl and array-operations without SIMD (yet), and numericals using SIMD. For reasons mentioned on the readme, I recently cooked up dense-arrays. This has interchangeable backends and can also use cl-cuda. But barring that, the developer overhead of actually setting up native-CFFI ecosystem is still too high, and I'm back to py4cl/2 for tasks beyond array processing.

numcl - Numpy clone in Common Lisp. [LGPL3][9].

[3] https://github.com/numcl/numcl