[Ffmpeg-devel] [PATCH] Vorbis decoder

Balatoni Denes dbalatoni
Thu May 12 18:13:39 CEST 2005


Hi!

Let's hope it won't be a flame war.

2005. m?jus 12. 17.10-kor Mike Melanson ezeket a bolcs gondolatokat fogalmazta 
meg:
> Rich Felker wrote:
> > I don't believe float is ever faster than a _proper_ int
>
> 	The issue, per my understanding, is that these perceptual audio
> decoders need to work with numbers that are larger than 32 bits in
> precision. You either use a float/double and use the CPU's FPU
> instructions or you break the 64-bit number up into 2 32-bit numbers and
> do a little more work to perform, e.g., a multiplication (vs. a single
> FMUL instruction for the floating point data type).
>
> 	CPUs without FPU benefit from pure integer implementation since it is
> faster than having to do FP ops in software with integers.

It is a somewhat twisted idea that a perceptual audio codec with 16 bits 
precission would need so many bits. In the case of Vorbis the spec. allows 48 
bits of precision, and in fact tremor uses an optimized floating point 
implementation (the idea: many numbers share the same exponent). In practice 
however (looking at the current encoder) 32 bits are more than enough for the 
decoder, but this could change. 

And to Rich, integer and floating point operations at least on the new 
(desktop) intel cpus are about the same speed - but integer ops can only be
efficiently paralellized up to 16 bits of precission.

bye
Denes

-- 
- Use the Source Luke ! -





More information about the ffmpeg-devel mailing list