News for hackers. Stuff that matters.
Re:Mastering Dyalog APL (2009) [pdf] by raphlinus on Thursday February 24, 2022 @11:25PM (#30463713)
Thanks for this explanation, it helps. I got into APL about 40 years ago (using an IBM 2741 terminal connected to a 360 if reconstruction of memory serves), but have only recently picked it up again, and not very deeply. From the kinds of things you're saying, the semantics line up somewhat with GPU-friendly parallelism, but not entirely.

If my (newly refreshed) understanding is correct, you can implement f⍣≡ as a per-thread fixpoint if you know that f is elementwise independent, which it seems like you'd be able to determine at least some of the time by static analysis. But I can also see that would be brittle, and if you fell back to do a dispatch per iteration, performance would tank.

Re:Mastering Dyalog APL (2009) [pdf] by moonchild on Thursday February 24, 2022 @11:40PM (#30463779)
> If my (newly refreshed) understanding is correct, you can implement f⍣≡ as a per-thread fixpoint if you know that f is elementwise independent, which it seems like you'd be able to determine at least some of the time by static analysis

Hmm, yeah, that would work. Though you would also need to account for side effects in f.

> semantics line up somewhat with GPU-friendly parallelism, but not entirely

An APL machine was proposed[0]. I say it is the GPUs that are wrong! :)

https://www.slac.stanford.edu/pubs/slacreports/reports07/sla...