Stochasically round on conversions from larger types

[pjaap]

This resolves #79

This enables the missing stochastic rounding calls on both deterministic and stochastic input.

[pjaap]

@milankl Thanks for this nice package. I am researching in the area of numerical analysis with stochastic floats and kind of need this small presented feature :)

I think it just needs a change in the few lines. Test are fine locally. Why are the tests not triggered in this PR? I can have a look if you want :)

edit: Workflow needs approval. This just popped up.

[milankl]

Awesome, thank you!! Will review after CI passes 🥳

[milankl]

Would you mind adding some tests to verify that this is actually the case?

• deterministic -> deterministic is deterministically rounded
• stochastic -> deterministic is deterministically rounded
• deter -> stoch is stoch rounded
• stoch -> stoch also stoch rounded

[pjaap]

Sorry for the late answer. I will add tests and investigate the failure on 1.11.
I am busy this week, so it may take a few days.

[pjaap]

Testing is not so trivial with non-deterministic values :)

I think

deterministic -> deterministic is deterministically rounded

is not in the scope of this package.

stochastic -> deterministic is deterministically rounded

this is already done in the tests. I adjusted the name of the testsets.

deter -> stoch is stoch rounded

this is now implemented. I check 10000 times that the result differs at least once for 100 random numbers :)

stoch -> stoch also stoch rounded

I am not sure what to test. The same test as above with Float64sr?

[giordano]

No need to check that a boolean is equal to another boolean, you can just test its boolean value 🙂

Suggested change

	@test is_stochastic == true
	@test is_stochastic

[pjaap]

Yes, I know. I found it more readable as a user in this particular scenario.
A failing test is more semantic since we get the message

  Expression: is_stochastic == true
   Evaluated: false == true

instead of

  Expression: is_stochastic

I think it is better in the first variant.

But I can change it to the second one if you want.

[milankl]

no need, that's fine by me and I agree with your point

[milankl]

Just saw https://pretalx.com/juliacon-local-paris-2025/talk/UDQLTD/. Very nice, thanks for talking about this!

Sorry I can have a look at this pull request next week and merge. Feel free to ping me in case I forget

[pjaap]

Hey there! I was on vacation until today :) so, ping?

I was looking for an interesting topic for the JuliaCon and your package applied to our PDE solvers seemed perfect :)
Luckily, the talk was approved. Looking forward to it.

[milankl]

This still fails at v1.11 you mind retriggering the CI?

[pjaap]

I added a small commit. But I cannot start or restart CI.