Description
Feature gate: #![feature(avx10_target_feature)]
This is a tracking issue for Intel AVX10
target features. The Intel Whitepaper can be found here
AVX10 is a "converged vector ISA", which attempts to simplify the mess that AVX512 was (21 ISAs). AVX10.1 is just a cleanup ISA, which implies the following AVX512 ISAs.
avx512bf16
avx512bitalg
avx512bw
avx512cd
avx512dq
avx512f
avx512fp16
avx512ifma
avx512vbmi
avx512vbmi2
avx512vl
avx512vnni
avx512vpopcntdq
Notably, the ISAs listed below, even though part of AVX512, will not be part of AVX10.1
avx512er
,avx512pf
- Deprecated ISAs, were only available in the KNL and KNM series, neither Rust/LLVM nor GCC supports them anymoreavx5124fmaps
,avx5124vnniw
- Deprecated ISAs, were only available in the KNM series, neither Rust/LLVM nor GCC supports themavx512vp2intersect
- Only Intel CPU to support it was TigerLake, recently AMD has added support in Zen5vaes
,vpclmulqdq
,gfni
- Intel probably wants to keep these separate from other SIMD intrinsics
AVX10.2 introduces a lot of new instructions (the new spec sheet hasn't released yet after the whitepaper, the latest one is this)
This tracking issue is meant to be incremental, i.e. new target features will be added as Intel publishes new versions of AVX10, and old versions will be stabilized.
Public API
The following target features and their associated intrinsics in stdarch
avx10.1
avx10.2
Implementation History:
Unresolved Questions
- None yet.
@rustbot label O-x86-32 O-x86_64 A-target-feature A-SIMD