On 2026-02-15 15:14, andrew drakeford via SG14 wrote:

Hi all,

Following the recent SG14 discussion on deterministic parallel reduction, I’ve now written up the proposal as a formal WG21 draft.

The paper focuses on expression-structure determinism for parallel reductions: it defines a canonical pairwise reduction tree parameterised by a lane count L. Implementations (threads, SIMD, GPU) must produce results as-if evaluating that fixed tree.

This extends the determinism guarantee of std::accumulate (fixed evaluation structure) into parallel and vectorised contexts, without imposing associativity or commutativity requirements.

Draft paper (HTML):
https://andyd123.github.io/canonical-reduce/generated/DxxxxR0.html

As discussed, the reduction kernel follows the “iterated pairwise summation” approach described in:

Dalton, Wang, Blainey — “Fast, Accurate Summation of Floating-Point Numbers” (IBM Research)
https://research.ibm.com/publications/fast-accurate-summation-of-floating-point-numbers

I would particularly welcome feedback from SG14 on:

Whether the semantic contract is clear and appropriately scoped

Any concerns regarding interaction with existing parallel algorithms

Whether the lane-parameterised formulation is the right abstraction boundary

If there are no major objections, I intend to move this toward LEWG for initial direction.

Would it be worthwhile to float it past SG1 for feedback as well? They might just rubber-stamp it, but it would probably be a good idea to at least ensure they're aware of it.

Best regards,
Andrew

_______________________________________________
SG14 mailing list
SG14@lists.isocpp.org
https://lists.isocpp.org/mailman/listinfo.cgi/sg14