[Ffmpeg-devel] int vs. float: Hard Numbers

[Luca Barbato]

Mike Melanson wrote:
> 
>     It's worth noting that this type of CPU is at the heart of VIA's
> mini-ITX boards. Do you know one of the main uses for this form factor?
> Media center PCs.
>

There are better suited cpu... Like PPC...

I think that benchmark is relatively pointless since you should issue
more than one sum per cpucycle....

At least on decent cpus.

So the point is:

- how many integer instructions I could process per cycle?
- how many float instruction I could process per cycle?
- how well they mix each other/ mix with SIMD etc?

The answers are in the cpu PEM most times.

Having the floating point unit as fast as the integer unit and
completely separated could be ok (per cycle then you can perform N
integer, M floating instruction, Q SIMD instructions).

That way you can archive the best performarce using everything.

float+mmx ...

lu

-- 

Luca Barbato

Gentoo/linux Developer		Gentoo/PPC Operational Manager
http://dev.gentoo.org/~lu_zero