[C++] Prospects of SVE based optimizations

[chiranmoyf]

Are SIMD optimizations with regard to SVE acceptable by the community?
I see that there are many functions utilizing NEON while benchmarking but barely any that use SVE.
There are lots of functions using, and in turn utilizing SSE, AVX and NEON, but not for SVE.

[kou]

It's acceptable if we can maintain. See also: #46002

[kou]

I don't know...
If macos-14 or macos-15 support SVE, we have it.

[lidavidm]

I don't think M-series processors support SVE (not even M4, according to Wikipedia, which otherwise supports ARMv9). It's only on Fujitsu's A64X and relatively recent ARM designs.

It appears Linux ARM runners from GitHub should support it, being either Neoverse V1 or Neoverse V2, but apparently they're missing support: https://github.com/orgs/community/discussions/19197#discussioncomment-12315925

[lidavidm]

It's frustratingly hard to find documentation on exact processors but it looks like "maybe in the future" is the answer, at least for GitHub hosted runners.

[chiranmoyf]

I don't think M-series processors support SVE (not even M4, according to Wikipedia, which otherwise supports ARMv9). It's only on Fujitsu's A64X and relatively recent ARM designs.

Not just Fujitsu A64X, the Graviton series from AWS also supports SVE.

It appears Linux ARM runners from GitHub should support it, being either Neoverse V1 or Neoverse V2, but apparently they're missing support:

Thank you for the information about the CI. For now, we will try to look into possible optimisations with SVE, and the community can make a decision based on the utility

[kou]

Note that we need not only CI but also at least one maintainer for maintain.
If we don't have any maintainer, we need to remove it eventually. For example: https://lists.apache.org/thread/f5ndxf0o6v4wbzgoy6nsrj4xkn34f380

[pitrou]

Are SIMD optimizations with regard to SVE acceptable by the community?

Hand-written SIMD optimizations should almost always use xsimd.

I see that there are many functions utilizing NEON while benchmarking but barely any that use SVE.

How did you come to this conclusion ("many functions utilizing NEON"?). NEON is enabled by default when compiling for ARM/ARM64, but we have very few hand-written optimizations. So this might simply be code vectorized by the compiler.