By: anon1 (anon.delete@this.anon.com), May 17, 2022 4:47 am
Room: Moderated Discussions
rwessel (rwessel.delete@this.yahoo.com) on May 17, 2022 1:42 am wrote:
> Alternatively, after half a century of supposed ISA and compiler improvements, which have
> largely utterly failed to allow broad vectorization (or the compiler extraction of other
> types of ILP), why anyone would think that this time will be any different. This was,
> in case you don't remember, the explicit promise made for AVX512 as well, for a recent example.
> Heck, AVX512 was supposed to allow broad vectorization of integer code as well.
I am not aware of any time-tested ISA (past or present) with similar features for auto-vectorisation as SVE. I mean, there is RISK-V, but it's even newer and if anything, less tested. AVX512 is not vector-agnostic and the hardware availability is so poor that almost nobody bothers to deal with it.
All in all, it's a bit early to conclude that these approaches have failed. They have never even seriously started.
> Alternatively, after half a century of supposed ISA and compiler improvements, which have
> largely utterly failed to allow broad vectorization (or the compiler extraction of other
> types of ILP), why anyone would think that this time will be any different. This was,
> in case you don't remember, the explicit promise made for AVX512 as well, for a recent example.
> Heck, AVX512 was supposed to allow broad vectorization of integer code as well.
I am not aware of any time-tested ISA (past or present) with similar features for auto-vectorisation as SVE. I mean, there is RISK-V, but it's even newer and if anything, less tested. AVX512 is not vector-agnostic and the hardware availability is so poor that almost nobody bothers to deal with it.
All in all, it's a bit early to conclude that these approaches have failed. They have never even seriously started.
