Re: [mpg123-devel] fixed point decoders

[Taihei M. <tm...@ma...>]

On 2009/05/30, at 5:51, Taihei Monma wrote:

> On 2009/05/29, at 23:40, Thomas Orgis wrote:
>
>> Well, I somehow believe that our current simple engine, while not  
>> being optimized for accuracy, should work better that it does now.
>> It works fine to produce audible output, just this issue with  
>> relatively silent tracks...
>> I don't think that we really need to implement an integer engine  
>> that rivals MAD in it's core territory (thinking about arches with  
>> 32bit native integers, at least). Something that works efficiently  
>> on ARM machines (preferrably more efficient than MAD) and has  
>> reasonable quality would be fine with me.
>> Anyhow, so far folks on Sony PSP or nintendoDS seem to be happy...  
>> though, I wonder about getting a proper mp3 player on my Palm Treo  
>> phone. RealPlayer sucks.
>
> OK, now I've finished investigation, and integer decoder (for  
> layer3, no work for layer 1/2) works better. The problems are:
> - The current double-to-real scalefactor (32768.0) is too small for  
> gainpow2 table. This causes a corruption (truncated to zero) of  
> dequantized samples in some cases. But using larger constant scale  
> factor doesn't fix the problem because the values in the table have  
> a wide distribution. So for small values, we should use larger scale  
> factor, and for large values, we should use smaller scale factor.  
> Fortunately gainpow2 is a constant table, so we can pre-calculate  
> these adaptive scale factors. Finally this improves the RMS error in  
> the compliance test to the order of 10e-4.
>
> - The scalefactor is too large for ispow table. The maximum value in  
> the table is about 165,516, so the value multiplied by 32768  
> surpasses the limit of 32-bit signed integer.
>
> The prototype code is working fine in my machine. I think it is  
> still not accurate enough, but should be better than now. Hopefully  
> I'll commit it tomorrow...

Now fixes are on the trunk. Layer 1/2 decoder (was totally broken) is  
also fixed.

Thanks,
Taihei Monma