Matt Borland wrote:
* CR on mandatory operations ... The Intel Library's README is not correctly rounded for all operations either, and states as much (which I quote in the docs). I can add in a line that says results will differ in the last few ULPs. You will find that even between standard library implementations for binary-floating point math results will vary outside of 0.5 ULPs.
To be precise, I'm referring only to "mandatory operations" in IEEE 754-2019 (Clause 5). "Optional/extended operations" are out of scope here. The README says that the library conforms to IEEE 754-2019. By definition, this implies that mandatory operations are CR. If you've observed a non-CR result for a mandatory operation, that's likely a bug, and it would be helpful to report an issue.
Can you share the test harness that you've already built?
I'm happy to share my harness after I take the time to brush it up, but that may take a while. That said, I think crafting a new harness is straightforward and takes minimal effort.
I never put effort into them because the copyright All Rights Reserved IBM precludes them from being checked into the repo or modified like your previous comments call out. I don't really want to spam downloading from Mike's website, but he at least has the disclaimer of reproduced with permission from IBM.
The website (https://www.speleotrove.com/decimal/) says that the testcases are also covered by the ICU license. Regarding upstream bandwidth during CI, we can use caching in GitHub Actions, so we don't need to fetch the files on every run. Once the cache is populated, later jobs restore from the cache instead of downloading, which keeps upstream traffic to a minimum. Regards, Michel