Re: [codec] it MUST NOT exceed 1275 bytes?

[Gregory Maxwell <[e-mail]>]

Benjamin M. Schwartz [[e-mail]] wrote:
>> 2) What are the calculations behind 1275?
>"The maximum representable size is 255*4+255=1275 bytes."
> Because the first N-1 frames in a packet cannot have a size greater than
> 1275, it would be very strange if this were permitted for the final frame.
> An unbounded frame size would also unreasonably increase the minimum
> computational performance required of a conformant decoder.

As Ben and the draft point out, the actual limit comes from the
scheme used to encoded lengths in the multiple packed frames
case.

It's also worth pointing out that 1275 is high enough that the codec is 
at the point where it is often _unable_ to use more rate (and would just
be coding zeros) due to a multitude of internal complications and tradeoffs. 

It's also low enough that the frames will by supportable on common networks
even with a fair amount of higher layer overhead without needing
fragmentation.

In the CELT modes the maximum precision used is limited by the codebook
size (memory trade-off) and band splitting depth (computation trade-off), so
higher rates couldn't be usefully supported without non-trivial worst case
complexity/memory increases for implementations.

And at 510kbit/sec we're at nearly twice the highest ABXable
bitrate for any killer sample I'm currently aware of... leaving plenty of
just-in-case headroom for VBR encoders to throw lots of bits at
challenging sections.