Hacker News new | past | comments | ask | show | jobs | submit login

> SIMD ISAs usually do not provide the integer division; the only known exception is RISC-V Vector Extension

It's kind of funny to read "the only known exception is..." in this context. What would an unknown exception be - an ISA that accidentally implements this but that the author believes nobody is aware of yet?

More seriously, I actually don't understand the intended meaning here. I assume the author means "out of all the ISAs I know"? What is that set of ISAs?




Some SIMD ISAs with integer division:

- Arm SVE, though only for 32-bit and 64-bit element types: https://developer.arm.com/architectures/instruction-sets/int...

- loongarch64 LSX/LASX: https://jia.je/unofficial-loongarch-intrinsics-guide/viewer/...

- MRISC32 (though that's somewhat obvious as basically everything in it is shared between scalar and vector).


Practically, could the expression “only know exception” mean anything other than “known by me?” I mean, it is clearly possible for an exception to exist, on account of the existing known exception, so they can’t know that more exceptions don’t exist out there somewhere.

I dunno. I think it is a more meaningful statement if we know more about the author; if we assume that they are very well informed, I guess we would assume that the fact that they don’t know about something is really meaningful. In the case of a blog post where most of us don’t know the author, it is hard to infer much. But at least it tells us why they decided to do the thing.


If a SIMD ISA exists, someone must know that it exists, because definitionally we only apply the term "SIMD ISA" to things that were consciously created to be such. So we could simply check every such example. Saying "only known example" is indeed silly.

But e.g. in mathematics, if we say that "almost every member of set X has property Y; the only known exception is Z" then there absolutely could be more exceptions, even if we pool the knowledge of every mathematician. It isn't necessary that X is finite, or even enumerable. It could be possible for exceptions other than Z to exist even though every other member of the set that we know about has the property. It could be possible to prove that there are at most finitely many exceptions in an infinite set, and only know of Z but not be able to rule out the possibility of more exceptions than that.

We don't even need to appeal to infinities. For example, there are problems in discrete math where nobody has found the exact answer (which necessarily is integer, by the construction of the problem) but we can prove upper and lower bounds. Suppose we find a problem where the known bounds are very tight (but not exact) and the bounded value is positive. Now, construct a set of integers ranging from 0 up to (proven upper bound + 1) inclusive... you can probably see where this is going.

The latter doesn't apply to SIMD ISAs, because we know all the interesting (big hand-wave!) properties of all of them rather precisely - since they're designed to have those properties.


The Convex C1 [0] and for a newer example NEC SX Aurora [1] both also support vector integer division.

[0] https://bitsavers.org/pdf/convex/080-000120-000_CONVEX_Archi...

[1] https://ftp.libre-soc.org/NEC_SX_Aurora_TSUBASA_VectorEngine...


> I assume the author means "out of all the ISAs I know"?

Out of all the ISAs where they know whether it provides integer division or not.


Yeah but my point is that as a reader I'm trying to figure out which ISAs actually don't provide this (vs. which ISAs the author lacks knowledge of), and I still don't know what those are. The sentence looks like it's supposed to tell me, but it doesn't.


You could come up with an ISA that implements it and it wouldn't be "known". Maybe that helps?




Consider applying for YC's Spring batch! Applications are open till Feb 11.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: