[mesa-users] update on mesa platform dependency bugs: natural log function different on Mac 10.6.8 vs 10.8.5

Tomasz Plewa tplewa at fsu.edu
Sun Nov 24 22:54:06 EST 2013


Hi Bill -

I think you both are right. I think it is important to understand the 
root cause of such differences. But I also think the nature does not 
know much about intricacies of the floating point arithmetics. That is, 
model predictions should not depend on the last floating point digit. If 
they do, the model is not robust.

Tomek
--
On 11/24/2013 2:40 PM, Bill Paxton wrote:
> Hi Roni,
>
> Thanks for the input -- but I disagree!
> It will be useful to explore these points in detail, since if you're right, then I'm in deep-deep trouble.
>
>
> On Nov 24, 2013, at 1:09 AM, Roni Waldman wrote:
>
>> Hi Bill,
>> As you probably know, the accuracy of double precision real is 15 significant digits, so the difference you detect between the platforms is expected and unavoidable.
> I disagree
>
>> You can't expect getting the exact same result in a MESA run on different platforms.
> I disagree
>
>> The best you can do is strive to avoid crashes by checking on various platforms and fixing bugs that get caught that way.
> I disagree
>
>
>
> Double precision is a specific format of 64 bits, and I see no reason why the different platforms given identical 64 bit input shouldn't give exactly the same 64 bit output for a well defined math function like log any more than I'd be happy with different answers for multiply or divide -- the answer is known damn it, there is no question about what it should be once you pick a standard rule for round-off (IEEE standards e.g.).   It may be that the operation will take slightly longer to compute in order to get the correct 64 bit answer, and some folks may be willing to give up accuracy for speed, but I certainly am not when it comes to MESA.
>
> In the past I've myself been guilty of waving off such cross platform differences as expected and unavoidable.  But MESA has reached a point where I don't think that it is still a viable option -- I just can't hope to make the code robust and reliable if the results vary across platforms like they do now.    I need to be able to develop and debug on one platform and have high confidence that once it works on that one it will work precisely (bit-for-bit) the same on a reasonable set of other platforms too -- for example, all platforms supported by the mesasdk.
>
> If that's impossible, I'd like to understand why it is at a deeper level than "we've tried, but we've never been able to make it work".
>
> Cheers,
> Bill
>
>
>
>
>
>
>
>
>
>
> ------------------------------------------------------------------------------
> Shape the Mobile Experience: Free Subscription
> Software experts and developers: Be at the forefront of tech innovation.
> Intel(R) Software Adrenaline delivers strategic insight and game-changing
> conversations that shape the rapidly evolving mobile landscape. Sign up now.
> http://pubads.g.doubleclick.net/gampad/clk?id=63431311&iu=/4140/ostg.clktrk
> _______________________________________________
> mesa-users mailing list
> mesa-users at lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/mesa-users
>






More information about the Mesa-users mailing list