Skip to content

WIP - Add Arm SVE2 builds - #752

Merged
solidpixel merged 1 commit into
mainfrom
opt-sve2
Nov 12, 2026
Merged

solidpixel merged 1 commit into
mainfrom
opt-sve2

Conversation

@solidpixel

@solidpixel solidpixel commented Nov 11, 2026 •

Copy link
Copy Markdown
Contributor

This PR adds Arm SVE2 support, and as a second commit (draft) tests whether moving the 4-wide version to use native SVE predicates is actually useful.

Both changes have been tested on QEMU and pass, but have not yet been tested on hardware to measure performance.

The predicate change is causing an internal compiler fault in GCC-14 (latest in Ubuntu repo), but is fine with GCC-15.3.

Fixes #740

@solidpixel

solidpixel commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor Author

Predicate change makes no performance difference on hardware, so can be reverted to avoid the ICE in GCC.

@solidpixel

solidpixel commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor Author

SVE128/256 performance signed off on Graviton 3.

  • No change in this patch (as expected).
  • Get 1.2x scaling moving from 5.7 to latest 6.x dev.
  • Get 1.2x scaling moving from SVE_128 6.x dev to SVE_256 6.x dev.

No performance advantage to using the SVE2 wider table gathers on Graviton 4, but it only has a 128-bit SIMD implementation, so presumably takes two passes. Rejecting the SVE2_128, and wont commit only SVE2_256 because there is no hardware to validate it on currently.

@solidpixel
solidpixel force-pushed the opt-sve2 branch 2 times, most recently from 7bdbd74 to 57199df Compare November 12, 2026 13:18
@solidpixel
solidpixel merged commit d26fcda into main Nov 12, 2026
21 checks passed
@solidpixel
solidpixel deleted the opt-sve2 branch November 12, 2026 13:33
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Add Arm SVE2 backend

1 participant