[vorbis-dev] Parallelism

[Segher Boessenkool]

> Hello,
> 
> Mmmm!! I do not understand nothing :-? I am encoding MP3 in my K6 II 400 at 
> 2x-3x (aprox) using LAME (CBR), I can get around 5x using Xing codec (But 
> ok, it sounds really awful). In my P200 MMX I can get near 2x with LAME (I 
> had no tested Xing speed in this machine), but you can encode MP3 at 11x in 
> your K5  100 Mhz :-? My CPUs need some steroids :-) What compressor are you 
> using ?

General belief is AMD K5 is much slower on floating point than Intel
Pentium. This is true for pentium optimized code, but certainly not
for K5 optimized code. Optimizing for K5 is easy (well, easier than
for P5/P6); even gcc can do it. Commercial (read: windows) compilers
don't do that good a job; they are totally intel-centric. Running
linux helps a lot: better compiler for the processor. djgpp is no good
either, because it's coff format makes it very difficult to control
cache behaviour.

Most important part nowadays is good use of the cache, and not doing
stupid things (you do _not_ want to see the gcc assembler code for
something like  int a = (int)b, where b is float). Reducing register
pressure is very important on x86 as well. Encoding music is very
data hungry, so 7 general registers and 8 fp registers is not very
much. Most of the time you spend reading memory, and using the cache
in a good way gets even more important.

My encoder is called pem, i wrote it, you can't get it, and lame sounds
better on most samples anyway. I mainly use it to understand psycho
acoustic principles at the moment.

Oh, lame is about 2x realtime on this computer as well.
If you like benchmarks on 'old' pc's, I can bench my 486DX-WB 180MHz :-)

Ciao,

Segher

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/