[vorbis-dev] CVS troubles

[Segher Boessenkool]

Hi,

> Ah, my profiling is old (MDCT will be bottleneck in decode though).  Also the
> psych stuff is in flux.  Optimization may end up be wasted effort until I
> smnail it all down (you'll notice the code suddently gets very messy in
> analysis and psy.c).

I'll optimize mdct. Did it before, so shouldn't be too hard.

> 
> > Library routines like cos, sin, atan, log, exp are not shown in a normal
> > gprof listing. Ouch. cos was using a _lot_ of time.
> 
> I profile with -lc_p and -lm_p.  But I did do my profiling before the lpc/lsp 
> was in place.  BTW, I just committed some new LPC code that uses lookups for 
> most trig.

Did you get rid of the 1.0 / hypot(real, imag); in lpc_magnitude() as well?
It just _eats_ my cpu :-(

> 
> > > problem there is using gcc/egc as the compiler.  On IRIX, the MIPSpro CC 
> > > geenrates code 4x faster than GCC and on Solaris, SunSoft CC generates code 
> > > 2.5x faster.  I'd assume the same is true on Intel but I don't have any 
> > > commercial compilers to bench against on x86.
> > 
> > On intel, on my mdct algorithm, gcc is about 2x faster than visual c.
> 
> !! I'm surprised.  Maybe I'm missing some obvious trick to get better 
> performance...
> 

I guess MS is missing the obvious trick... cache.
Smaller code is almost always faster, 'cause a cache miss costs you _a lot_
of cycles. Too bad :-)

**Segher

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

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