[llvm-dev] Floating point variance in the test suite

[Hubert Tong via llvm-dev]

On Thu, Jun 24, 2021 at 10:15 AM Kaylor, Andrew <andrew.kaylor at intel.com>
wrote:

> > We've hit at least one failure when porting because the hashing means
> that miniscule differences in standard library implementations are not
> tolerated.
>
>
>
> How did you address that?
>

We're still in the process of trying to address it. First by turning off
the hashing. I am not sure the "implement portable replacement for standard
library function" strategy is appropriate in the case we're dealing with.
We're working to get better traces.

It currently looks like small differences keep growing, so this may be a
"reduce number of passes" situation.


> I would expect there to be many small differences between standard library
> implementations and across different platforms.
>
>
>
> Wherever the language standard doesn’t require correctly rounded results
> there is a potential for small variations. I think this problem becomes
> more prominent with tests that execute on GPU and similar hardware.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20210625/0e6a08fe/attachment.html>