[MEncoder-users] Re: Now mmx seg fault

[Guillaume POIRIER]

Hi,

On 6/13/06, larrystotler at netscape.net <larrystotler at netscape.net> wrote:
> -----Original Message-----
> From: Dominik 'Rathann' Mierzejewski <dominik at rangers.eu.org>
>
> >Who told you that? My amd64 supports MMX just fine.
>
> >From Wikipedia:
>
> "Since SSE and SSE2 are generally faster than, and duplicate most of
> the features of, the traditional x87 instructions, MMX, and 3DNow!, the
> latter are redundant under AMD64."

This is true. redundant != takes over

> "SSE2 replaces the x87 instruction set's IEEE 80-bit precision, with
> the choice of either IEEE 32-bit or 64-bit floating-point math"

This sentence is wrong. This what it should read:

> "SSE2 replaces either IEEE 32-bit or 64-bit floating-point math instruction of x87. IEEE 80-bit precision, are still only suppored through x87".

> There was an article about how they still support MMX, but it's through
> the use of SSEx instead of actually having MMX.  I can't find the link
> tho.  MMX is not as useful as SSE anyway.  I'm not sure what this means
>  from a programmer's point of view.  I would assume that the MMX calls
> are handled transparently, but I dunno.  I'm not a programmer.

No one really knows (besides Hammer's authors) how iSSE and MMX are
really threated, but since according to AMD's software optimization
guide, iSSE instructions take twice as much time to execture that the
corresponding MMX instructions, it's a safe bet to say that iSSE are
somewhat treated as 2 MMX instructions.

Real world test shows that Hammer benefits more from MMX-tuned code
than SSE-tuned code, except in 64 bits more where the extra regs allow
to write smarter algorithms in SSE.

Guillaume
-- 
"Success consists of going from failure to failure without loss of enthusiasm."
 -- Winston Churchill