Apple upgrades Lapack

[dada_dave]

For those interested in scientific computing, like @leman @theorist9, and especially those that use directly or indirectly the Accelerate framework Apple has finally upgraded and bug fixed lapack.

It’s now version 3.9.x. From what I understand this is the first update in a very long while.

https://mobile.Twitter or X not allowed/stephentyrone/status/1640740889835327491

 

[Andropov]

What's the current state of the Accelerate framework? I'm curious, I haven't really used it much and all I know is that maintainers at Numpy had an horrendous opinion of it (Accelerate is a possible backend for Numpy).

 

[dada_dave]

What's the current state of the Accelerate framework? I'm curious, I haven't really used it much and all I know is that maintainers at Numpy had an horrendous opinion of it (Accelerate is a possible backend for Numpy).

To be honest as a whole I’m not sure myself. However, I believe that at least part of the complaints about Accelerate were wildly out of date APIs like Lapack. So this is good news on that front.

 

[leman]

What's the current state of the Accelerate framework? I'm curious, I haven't really used it much and all I know is that maintainers at Numpy had an horrendous opinion of it (Accelerate is a possible backend for Numpy).

Some of it is very good (I like Apple's simd library for example), some of it was outdated and practically unmaintained (BLAS, LAPACK). This is also the reason why no scientific package recommends linking agains Accelerate.

But in the recent years Apple started to pour more resources into these things. They did some maintenance on these components somewhere around Big Sur (if I remember correctly), and now a big patch in 13.3. They are still a bit behind the official spec (I think they implement the API of 2021), but it's definitely much better than being a decade behind.

Apple's numerical code is usually of exceptionally high quality (if they are trying, that is), and it's good to see them taking scientific computing seriously again. If they have fixed the issues with the previous implementation, switching to Accelerate might bring massive improvements for matrix-heavy code. AMX is no joke.

 

[theorist9]

I wonder if it would be worth rerunning some of the Mathematica benchmarks.

 

[Andropov]

Apparently because of this LAPACK version update, SciPy is re-adding Accelerate backend support for macOS 13.3.

 

[leman]

Apparently because of this LAPACK version update, SciPy is re-adding Accelerate backend support for macOS 13.3.

Would you have a link for this? Can’t find anything on GitHub…

 

[Andropov]

Would you have a link for this? Can’t find anything on GitHub…

Sorry, not yet there. I read this comment from a Numpy/SciPy maintainer and thought it was ready (I hadn't had my morning coffee yet...) but what he says instead is that it's "highly likely" that Accelerate support will be re-added. But Accelerate was dropped (in 2018) mainly because its version of LAPACK was from 2009. The current version of LAPACK as provided via Accelerate is 3.9.1 (from 2021, as you say) and SciPy's minimum required version of LAPACK is 3.7.1 so it should be enough.