Vorbis dev - Jan 2005 - Questions regarding OGG implementation on DSP

As to tremor fixed-point implementation, i got several questions, and hopefully
some one can help me out.
Thanks a lot!

1. For DSP chips with small and limited memory, how much memory do we really
need for
saving all codebooks before we start to decode audio packets? Can the bit-rate
give us some clues?
Is there some simple guidelines for this? 

2. For CRC error protection, it seems to me that we need to get all the pages,
compute the CRC,
and then compare the result to the one for that packet. That means a huge input
buffer in some cases.
So, ignoring the CRC words seems a reasonable compromise for portable devices, 
since it is anyway a error detection, not correction. We cannot do anything
about it except skipping that
packet. Is this assumption ok?

3. Any further information/papers/books about how Vorbis gets audio samples
compressed?
Well, the Vorbis I specification does give me some clues, but not very clear to
me.


-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://lists.xiph.org/pipermail/vorbis-dev/attachments/20050131/a1a3c6b7/attachment.html

Hi!
On Monday 31 January 2005 13:06, Yuh-Chin Chang wrote:
> As to tremor fixed-point implementation, i got several questions, and
> hopefully some one can help me out. Thanks a lot!
>
> 1. For DSP chips with small and limited memory, how much memory do we
> really need for saving all codebooks before we start to decode audio
> packets? Can the bit-rate give us some clues? Is there some simple
> guidelines for this?
I just tried this:
download svn tremor version.
compiled and did 'make example' as well
ran:
 valgrind -v --tool=massif ./ivorbisfile_example < testfile.ogg >
/dev/null
this has created a very nice "memory usage over time" diagram.
in my case the maximum memory usage was abt 140kByte.
ogginfo testfile.ogg shows:
Vorbis headers parsed for stream 1, information follows...
Version: 0
Vendor: Xiph.Org libVorbis I 20040629
Channels: 2
Rate: 44100

Nominal bitrate: 45.000000 kb/s
Upper bitrate not set
Lower bitrate not set
Vorbis stream 1:
        Total data length: 2236995 bytes
        Playback length: 6m:19s
        Average bitrate: 47.157508 kbps
Logical stream 1 ended

Well, ok, this (valgrind) works only on x86 platforms, 
but you'll get a feeling which part uses how much memory.
>
> 2. For CRC error protection, it seems to me that we need to get all the
> pages, compute the CRC, and then compare the result to the one for that
You need one page (i.e. header + all packets) for the crc
check
> packet. That means a huge input buffer in some cases. So, ignoring the CRC
the reference encoder produces about 4kByte / page, but you cannot assume this
is true in general.
see libogg: framing.c function ogg_stream_pageout (line
~449)
> words seems a reasonable compromise for portable devices, since it is
> anyway a error detection, not correction. We cannot do anything about it
> except skipping that packet. Is this assumption ok?
yes. But if you decode corrupted packets you will probably get some noise
(sounds you don't want to) in your output.
the memory consumption for one page is much less than for the
codebooks.
>
> 3. Any further information/papers/books about how Vorbis gets audio samples
> compressed? Well, the Vorbis I specification does give me some clues, but
> not very clear to me.
Feel free to ask questions :-) 
There's very few information about the encoder in written form (or am I
wrong?)


BTW (a bit off topic):  
in "ivorbisfile_example.c" line 24+25 
#include <vorbis/ivorbiscodec.h>
#include <vorbis/ivorbisfile.h>
just before compiling a symbolic link is created: ln -fs . vorbis
why not changing the example lines to 
#include <tremor/ivorbiscodec.h>
#include <tremor/ivorbisfile.h>
and it will compile without "ln -fs"


Dominik

On Mon, 31 Jan 2005 20:06:01 +0800, Yuh-Chin Chang
<yuhchin@ms63.hinet.net> wrote:
>  
>  
> As to tremor fixed-point implementation, i got several questions, and
> hopefully some one can help me out. 
> Thanks a lot! 
>   
> 1. For DSP chips with small and limited memory, how much memory do we
really
> need for 
> saving all codebooks before we start to decode audio packets? Can the
> bit-rate give us some clues? 
> Is there some simple guidelines for this? 
This depends mostly on what decoder you use. The Tremor-lowmem version
is specifically tuned to use as little memory as possible. I believe
it typically requires somewhere around 40-60 kB total heap+stack for
decoding.

It's always possible to create pathological streams that require much
more memory, but it's generally considered reasonable for
memory-limited devices to refuse to play those.
>   
> 2. For CRC error protection, it seems to me that we need to get all the
> pages, compute the CRC, 
> and then compare the result to the one for that packet. That means a huge
> input buffer in some cases. 
> So, ignoring the CRC words seems a reasonable compromise for portable
> devices, 
> since it is anyway a error detection, not correction. We cannot do anything
> about it except skipping that 
> packet. Is this assumption ok? 
You only need a single page (NOT a packet), which will typically be
approximately 4 kB or less, which is a pretty small fraction of the
total memory required.

Not reading in the whole page at once would be quite a lot more
development effort for very little gain, plus you'd have the
disadvantage of not detecting corrupt data.
>   
> 3. Any further information/papers/books about how Vorbis gets audio samples
> compressed? 
> Well, the Vorbis I specification does give me some clues, but not very
clear
> to me. 
There's little written about the specific techniques the encoder uses
(though there's plenty of information about the general approaches
taken in the literature), and the specification only dictates what a
decoder needs to do. If you have more specific questions, we should be
able to help.

Mike