Poll on list or set determinism and deterministic FMA

[ngzhian]

Please vote on the replies below with respective emoji reactions.

[ngzhian]

Question 1:

1. List determinism (current status) 😄
2. Set determinism (implies deterministic FMA) ❤️
3. Neutral 🚀

[ngzhian]

Question 2:

1. Add deterministic FMA 👍
2. Do not add deterministic FMA 👎
3. Neutral (or it depends on the outcome of Q1) 🚀

[penzn]

Background: #86, #44, #5. #71.

[zeux]

@ngzhian Just to double check... since there's no neutral option, people who are neutral on either decision should not leave any emoji, correct?

[penzn]

@ngzhian Just to double check... since there's no neutral option, people who are neutral on either decision should not leave any emoji, correct?

Maybe we need a neutral, we had this problem with bfloat16 vote 😉

[ngzhian]

@ngzhian Just to double check... since there's no neutral option, people who are neutral on either decision should not leave any emoji, correct?

Good point, I'll add a neural option to both questions.

[conrad-watt]

Apologies for complicating things, but I would be (slightly) against adding deterministic fma if we voted for list non-determinism, so I've voted "neutral" in the second poll (edit: although maybe I should vote against?).

[Maratyszcza]

I’d like to point out that QFMA is not unique in that we may want to specialize code on its implementation, and if we go with set-non determinism, in addition to FMA we should add Sign Select, PSHUFB-style Swizzle, float/double to int conversions which produce INT32_MIN for out-of-range inputs, PMADDUSWB-flavor 2-wide integer dot product, and four variants of 4-wide integer dot product with accumulation.

[rossberg]

I'd like to remark that the question of introducing list non-determinism has repercussions on Wasm at large, far beyond this proposal. So ultimately, it is a question that the full CG will have to weigh carefully. So it might make sense to at least think about a plan B.

[conrad-watt]

Just to note again, I'm personally fine with list non-determinism, and I won't advocate for set non-determinism in the meeting if this comes at the expense of the phase 4 poll. I think the differences between the two are pretty slim, and I'm sympathetic to the argument that leaving a gap between spec and implementation practice (i.e. implementations are expected to always be "list"-style) is a hazard.

I think the bravery of this proposal comes from its facilitation of implementation-sniffing/specialised code, and this is true whether or not we go with list or set non-determinism (e.g. qfma testing would exist in both). I can imagine CG members wanting to pass comment on this general point (given potential feelings that Wasm/Web code should be platform-agnostic), but I worry that having a bikeshedding discussion on list vs set won't get to the heart of the debate.

[sunfishcode]

In case it's of interest, this question of set nondeterminism vs list nondeterminism is the subject of the earlier fp_env proposal. It provides the consistent-within-a-run property while avoiding baking in a concept of list nondeterminism into the language semantics.

[penzn]

There is definitely a philosophical difference here, but I don't think that either way would make a practical difference with tooling or user code. Given that I prefer the set approach because it is closer to the core spec, but list can work as well.