Nondeterminism in Futhark

A closer look at Nondeterminism in Futhark

What Nondeterminism Means in Futhark

This post explains three terms used when talking about numerical nondeterminism: Exact (same result every run), Replay (same result if you rerun with the same schedule/machine), and Approx (potentially nondeterministic). The nondeterminism discussed is about floating-point rounding, not shared-state concurrency. An example is parallel summation: different grouping or chunking changes the order of addition and thus the result. Atomic updates can introduce further variability even on the same machine.

Where nondeterminism actually arises in Futhark

Two main sources are described. First, language abuse: using #[unsafe] can disable safety checks and permit arbitrary memory reads/writes; this can be avoided by passing --safe to the compiler or simply not using unsafe code. Second, parallel constructs: scatter requires that index lists have no duplicates (or duplicates must correspond to identical values), but this is not checked at runtime; violating it makes write order unpredictable and can even produce torn writes. Reduce and other parallel combinators require associative (sometimes commutative) operators; the language does not check associativity, so using non-associative operators (notably floating-point addition) can lead to nonsensical or nondeterministic results.

Who will find this discussion useful

This writeup is aimed at Futhark users who care about numerical reproducibility and parallel correctness: people implementing data-parallel algorithms, those using scatter and reduce, and programmers moving from serial to parallel numeric code. It highlights pitfalls experienced and newer users commonly make and points out which issues are programmer responsibility versus inherent to floating-point arithmetic.

Questions people ask about Nondeterminism in Futhark

Does Futhark check for duplicate indices in scatter or for associativity in reduce?

No, scatter duplicate checks are not performed at runtime and associativity of operators is not checked; both are left to the programmer.

Can I avoid unsafe memory behavior in Futhark?

Yes, the post notes you can pass --safe to the compiler or avoid using the #[unsafe] attribute to prevent disabling safety checks.

Is the nondeterminism discussed due to race conditions?

No, the nondeterminism discussed comes from floating-point rounding and ordering of reductions or atomic combines, not classic shared-state race conditions.

Nondeterminism in Futhark was shared on NexURLs by NexURLs on November 7, 2026, in Tech & AI.