Re: [hatari-devel] Functional tester for 68881/2 FPUs

[Nicolas Pomarède <npomarede@xxxxxxxxxxxx>]

Le 03/11/2016 à 18:51, Douglas Little a écrit :
> I just posted this up on AF today - a fidelity/correctness test for FPUs.
>
> https://dl.dropboxusercontent.com/u/12947585/fputest.zip
>
> Obviously it fails some tests on Hatari because emulation is virtualized
> a bit (?) by handing calculations to the host FPU. I'm posting this here
> in case it turns out to be useful for improving emulation or finding
> bugs. I noticed some time ago that rounding modes are not correct in
> Hatari but not sure about the rest - if those are just fine precision
> differences or actual bugs.
>
> At the very least it might be used for regression checking since you can
> record expectations from a revision of the code and then test against
> them using newer builds.
>
> If its any use I'll link source for it (it's a small asm program with
> some tables).
>
> BFN
>
> Doug.

Hi

thanks, that's certainly very helpful for Hatari tests.

Regarding the precision / rounding difference, WinUAE's fpu contains 
some code to set the host fpu operating mode closer to motorola's fpu 
(by using specifing asm code path for intel cpu for example)

It's not used yet, but I plan to enable this later (for now, Hatari 
doesn't use that much of specific intel/arm/others asm optimisations in 
WinUAE).

Toni, is it possible to get an exact behaviour like 68881/2 fpu with 
those asm calls, or did you measure a percentage of accuracy compared to 
default settings of the host's FPU ?


Nicolas