Conversation

Notices

1. Embed this notice
   Paul ‮etnomailgaT (paul@soylent.green)'s status on Monday, 18-Nov-2024 07:03:21 JST Paul ‮etnomailgaT

   • Adam Chalmers

   *EDIT* We got a lot of responses and it helped a ton! My suspect is now libm; thank you SO MUCH for the replies, I love all of you weirdos ❤️

   Original post:

   Anyone with an M1,M2,M3 CPU running a **Linux**, could you be my hero and run the following rust code and reply with what it ran as?

   https://github.com/adamchalmers/fp-repro

   @adam_chal found that there was a difference in the returned values on his mac and intel

   So far, my boxes have returned -3.4047613719613308; but on an M* Mac, it's -3.40476137196133

   • Embed this notice
     pinskia (pinskia@hachyderm.io)'s status on Monday, 18-Nov-2024 07:03:20 JST pinskia

     in reply to

     • Adam Chalmers

     @paul @adam_chal I suspect the difference is because on x86_64 the default libc does not use FMA (since it is not part of x86_64v1) while on aarch64 (since FMA is part of armv8-a) will use FMA.
     Especially when it is 2up difference.
     Fused Multiple Add can make a huge difference since the addition is done in infinite precision after the multiply and only rounded afterwards.

   • Embed this notice
     pinskia (pinskia@hachyderm.io)'s status on Monday, 18-Nov-2024 07:34:03 JST pinskia

     in reply to

     • Adam Chalmers

     @paul @adam_chal well there are different implementations of tan too. Some have a slow path included and some don't. Even different versions of glibc are different. see https://sourceware.org/bugzilla/show_bug.cgi?id=15267 for talking about the slow path.
     I am not 100% sure FMA is the only difference here though. But it does change the ULP so does the slow path in some cases.

   • Embed this notice
     Paul ‮etnomailgaT (paul@soylent.green)'s status on Monday, 18-Nov-2024 07:34:05 JST Paul ‮etnomailgaT

     in reply to

     • Adam Chalmers
     • pinskia

     @pinskia @adam_chal this is an astonishing good and helpful insight, thank you!

     (Edit thought about it more and I think you're probably right!)