Re: [hatari-devel] FPU trig functions emulation accuracy?

[ Thread Index | Date Index | More lists.tuxfamily.org/hatari-devel Archives ]


Le 06/12/2020 à 12:15, Eero Tamminen a écrit :
Hi Toni,

On 12/6/20 10:04 AM, Toni Wilen wrote:
...
Assuming this is using softfloat mode: it returns same results as
68040/060 FPSP which does not return exactly same result as 6888x. FPSP
is more accurate.

6888x internal logaritmic/trigonometic algorithms are not known except
that they are CORDIC based but that does not help much.

Differences are very minimal, no program cares.

I added partial FPU support to my tester tool few months ago which can
report expected value and returned value and in my tests last few bits
of (trig/log) results are usually different when comparing between FPSP
and 6888x.

Other (non-trig/log) instructions are accurate.

Ok, so it's already fixed to work as well as it
could. And remaining marginal difference would
anyway cause problems for FPU using programs if
there happened to one that relies on few last bits
of trig/log accuracy, as it differs between FPU
types.

Great, thanks!

Nicolas, should this be documented somehow in
Hatari documentation?


Hi

yes, maybe a small note could be added to tell that although result is not 100% identical to real 6888x fpu, it's still between 99% and 100% in most cases.

Nicolas




Mail converted by MHonArc 2.6.19+ http://listengine.tuxfamily.org/