binutils Newspaper
OCTOBER 6, 2026
binutils — Under Review — Sentiment 6 / 10

AVX10 instruction naming and spec version feedback

Discussion regarding the naming convention for AVX10 auxiliary instructions and feedback on the ACE specification version.

Christian Ludloff and Jan Beulich discuss the naming convention for new AVX10 auxiliary instructions, with Ludloff suggesting AVX10_AUX_Vn as a potentially clearer alternative to AVX10_Vn_AUX. They also note discrepancies in the version of the ACE (AI Compute Extensions) specification being referenced. The conversation touches on specific instructions like vcvt and vpmovssdb, and the need for direct feedback to the ACE EAG (Ecosystem Advisory Group) on documentation updates.

In the Thread 6 participants
  1. Christian Ludloff proposer

    Questions the naming convention for AVX10 auxiliary instructions and notes the specification version.

    “In fact doc version 1.15 latest ACE spec is 1.16.2 direct feedback on silent doc releases to the EAG also, instead of the AVX10_Vn_AUX insanity, has anyone considered AVX10_AUX_Vn for naming?”
  2. Jan Beulich reviewer

    Comments on the ACE specification version and the proposed naming convention, noting that the 'Vn' might not imply versioning.

    “Interesting, especially when at the top of https://sourceware.org/resource/ai-compute-extensions-ace-specification/ is says "June 15, 2026". Now that it's clear that the Vn here isn't intended to have”
  3. Jiang, Haochen reviewer

    Notes that specification updates might be incomplete and that some comments regarding exception types are unaddressed.

    “I don't know that has been updated either. It seems to me that some of my comments are not addressed. (Here I mean the exception type) I will mention that.”
  4. Jan Beulich reviewer

    Acknowledges Haochen Jiang's point about unaddressed comments and notes that specific instruction issues (FP8 to FP4 convert modrm) appear fixed.

    “I don't know that has been updated either. It seems to me that some of my comments are not addressed. (Here I mean the exception type) I will mention that. It seems to me that FP8 to FP4 convert modrm is fixed, it is a clear bug in document”
  5. Jiang, Haochen reviewer

    Reiterates that some comments remain unaddressed and highlights a specific instruction bug fix.

    “It seems to me that FP8 to FP4 convert modrm is fixed, it is a clear bug in documentation, while FP8 to FP6 still remaining a que”
  6. Christian Ludloff proposer

    Confirms specific instructions ('vcvt', 'vpmovssdb') as outliers and agrees that 'aux1' and 'aux2' are poor naming choices.

    “yep, vcvt{bf,hf}82bf4s and vpmovssdb are the only three "outliers", with dst=r/m yep, aux1 and aux2 are poor names – classic avx512 names did a little better... x86... c'est la vie :)”

Technical Tradeoffs

  • Debates naming conventions for clarity and consistency in instruction sets.
  • Highlights potential issues with specification versioning and documentation accuracy.

In Details

This thread discusses the naming conventions for new AVX10 auxiliary instructions and related AI Compute Extensions (ACE) specification details. Christian Ludloff questions the AVX10_Vn_AUX naming, proposing AVX10_AUX_Vn as potentially more logical. The participants also noted discrepancies in the ACE specification version (1.15 vs. 1.16.2) and discussed specific instruction behaviors like vcvt and vpmovssdb. The conversation highlights the need for clear naming and accurate documentation in evolving instruction set extensions.

For Context
AVX10
A planned extension to the AVX instruction set architecture, aimed at improving AI and compute performance.
ACE
AI Compute Extensions. A set of proposed extensions to the x86 instruction set architecture designed to accelerate artificial intelligence workloads.
EAG
Ecosystem Advisory Group. A group that provides guidance and feedback on technical specifications, likely within the context of processor architecture extensions.
vcvt
A vector instruction likely related to data type conversion within AVX/AVX-512.
vpmovssdb
A vector instruction likely related to packed move operations with scalar doubleword elements, possibly for sparse data.
Filed Under: x86AVXISAdocumentationnaming