Scryer-prolog: Please consider providing Unum or Posit arithmetic

[triska]

In his book, The End of Error, John Gustafson describes a new number format called Unum with many interesting properties.

Since then, Prof. Gustafson has also designed several alternative formats that are worth providing. Please see these links:

http://www.johngustafson.net/unums.html

https://posithub.org/about

https://posithub.org/khub_community

https://users.rust-lang.org/t/create-rust-posit-arithmetic-library-based-on-21-february-2017-posit-arithmetic-by-john-l-gustafson-beating-floating-point-at-its-own-game/10077

This would be a great number format to have available in a Prolog implementation, and a quite unique selling point. In fact, the idea is similar to number formats that were available in previous Prolog implementations and designs (Prolog IV).

I hope that you find these number formats interesting, and consider providing them at some point!

[mthom]

It would be a unique selling point indeed! I remember Willi Riha from the Shen community mentioning unums a while back. Creating a Rust implementation of unums could be a fun project onto itself. Chances are fair that one exists already. I will look into it. Thanks!

[mthom]

Is The End of Error still worth a purchase, or have unums been eclipsed by posits? I can't find much on posits.

[triska]

To the first question: In my opinion, the book is well worth purchasing, due to the many examples and nice illustrations of floating point issues alone already. It is very nicely done.

As to the other part: I have seen posits being described as "a hardware-friendly version of unums":

https://www.johndcook.com/blog/2018/04/11/anatomy-of-a-posit-number/

I think an important takeaway of this is that ideally, several different number formats can be nicely supported, with rational numbers as a very important application, and unums possibly as a quite unique attraction.

[jmlMetaswitch]

https://www.nextplatform.com/2019/07/08/new-approach-could-sink-floating-point-computation/ gives a good overview.

[XVilka]

There is also softposit crate to build on.

[mthom]

It surprises me that none of the unum/posit projects I've seen experiment with GPU parallelism. Posits/unums are designed with parallelism in mind, no?

[UWN]

GPU parallelism heavily depends on (the absence of) state (like all the tiny flags in IEEE that make switching more cumbersome) - unums remove that state.

Currently it seems that Gustafson explores the design space around unums. Unums were much closer to traditional IEEEs than what posits are

[UWN]

... but in any case, it is probably more promising to first get regular Prolog (including partial lists) and constraints up and running.