suspected bug: different results on x86 vs apple arm

[faddat]

This is the same floating point math test. The only difference is that it is being run on an ubuntu instance and on an M1.

Ubuntu fails because I used the result from my m1.

• https://github.com/CosmWasm/wasmvm/actions/runs/12732434887/job/35487631744?pr=582

• https://github.com/CosmWasm/wasmvm/actions/runs/12732434887/job/35487631959?pr=582

My understanding is that the output should be exactly the same, regardless of platform. Is that correct?

[ncruces]

Do you have source for the Wasm? Also, even if you don't have source for the Wasm, some minimal Go code to load the blob and show the difference would be helpful (obviously, the larger the blob, the worst that is).

A repo (or a gist) with a MCVE is always helpful.

Also, how was the Wasm compiled? Does it use SIMD? Is it possible the difference is in SIMD code?

In terms of assurances of the output being equal, I'm reasonably sure it should be equal, if it's using IEEE 754 floats. I'm less sure if it is using SIMD.

[anuraaga]

Also one thing to keep in mind is NaN isn't deterministic in Wasm.

https://github.com/WebAssembly/design/blob/main/Nondeterminism.md

Not sure but found that perhaps it is being used in the test case. Would be good to confirm it (though I would be surprised if Wasmer handled them more deterministically than Wazero).

https://github.com/CosmWasm/cosmwasm/blob/main/contracts/floaty/src/floats.rs#L27
https://github.com/CosmWasm/cosmwasm/blob/main/contracts/floaty/src/instructions.rs#L381
https://github.com/CosmWasm/wasmvm/blob/0b829226597bdbab9a2278c6a457ba452fb1b8a7/internal/api/lib_test.go#L1409

If NaN is indeed an issue, one approach could be to check the result IsNan and replace it with a constant NaN representation within the unit test.

[ncruces]

@faddat if this is indeed an issue, it's something we'd want to fix. But we need more info/help.

[mathetake]

yep, floating point is non deterministic and it's clearly documented in the spec (and each CPU has different behavior as it's undefined!) - unless you give us the real repro which doesn't involve NaN or Inf, i am closing

[faddat]

Ok, thank you all for your feedback!

I have been afk for a few days, and I'll poke around and come back.

[ncruces]

#2365 should not change anything, unless (again) there is a NaN (which is non-deterministic), or some SIMD that is allowed to have reduced precision, or of course, some other form of non-determinism (a clock, whatever).

If you have a discrepancy to report, between interpreter and compiler or across platforms, please do so. But it should be (as much as possible) a reduced test case, which is not supposed to be non-deterministic or unspecified to begin with.