search

WebAssembly / relaxed-simd

Relax the strict determinism requirements of SIMD operations.
Other
43 stars 9 forks source link

Poll on relaxed mode #105

Open ngzhian opened 2 years ago

ngzhian commented 2 years ago

Dan presented a relaxed mode proposal today.

Please share comments and ask questions here, and vote using emoji reactions on this comment:

For: 👍 Neutral: 👀 Against: 👎

sunfishcode commented 2 years ago

Link to the proposal slides: https://sunfishcode.github.io/RelaxedMathMode.pdf

conrad-watt commented 2 years ago

Are we also still going to do a poll regarding the status of bfloat16_dot?

EDIT: IMO the story for the relaxed mode proposal is cleaner if we drop this instruction.

tlively commented 2 years ago

Relaxed mode idea sounds good to me. For the presentation to the larger CG, I think it will be helpful to be clearer on exactly what is meant by "mode," i.e. how it will be specified, documented by consumers, and used in practice.

+1 to having a separate poll on bfloat16_dot.

dtig commented 2 years ago

There's already one here: https://github.com/WebAssembly/relaxed-simd/issues/77#issuecomment-1226191806

Do we want to reuse that poll or add another one on the same issue?

tlively commented 2 years ago

A new poll as a follow-on to this one makes sense because the outcome of this poll may change the calculus on whether it makes sense to include bfloat16_dot. I believe we also still needed to have some follow-up discussion about what a deterministic, canonical lowering for bfloat16_dot might be under the relaxed mode proposal.

ngzhian commented 2 years ago

Overwhelming positive reception for relaxed mode! I started #106 for dot product.