[Tremor] question about IMDCT on 24-bits fixed-point DSP

[jim luo]

>hi, Jim,
>From VorbisI spec, the residue value precision could be +-140dB (48bit).
>So for IMDCT, I personally don't think 24bit precision is enough.
>(right shift by 2 is just the simplest way to keep 24bit precision).
>
>I think, if MIPs are enough for you, you may try double-precision for IMDCT.
>That helps gain precision in lost of MIPs and memory.
>
>Good luck to your project!
>
>//by the way: are you using Freescale DSPs?
>
>Regards,
>SSP
>
Yes, I noticed that in vorbis spec, but 2^26 is the maximum that I see in current 
bitstream, and I don't want to sacrifice more Mips or memory for precision.
	It is so bad! :(
re btw: I'm just evaluate the possibility on 24 bits precision, still don't decision to 
use which DSP. 
>>From: "jim luo"
>>To: "tremor" <tremor at xiph.org>
>>Subject: [Tremor] question about  IMDCT on 24-bits fixed-point DSP
>>Date: Wed, 11 May 2005 21:25:24 +0800
>>
>>Dear All:
>>	I am porting Tremor to my 24-bits fixed-point DSP, But meet the accuracy 
>>problem during IMDCT:
>>several samples before IMDCT are 26 bits integer value, and can not be 
>>stored in 24-bits intenal
>>RAM, so I add some code before mdct_backword(), right shift all the 
>>samples_before_IMDCT value by 2 bits, to avoid overflow.
>>	but the PCM samples are deviated from that gotten from original tremor, 
>>the max difference between them are 6.
>>	I think my technique to process 26 bits value is unproper. Do you have any 
>>more clever means to process the IMDCT with several input samples large 
>>than 2^24 on 24 bits fixed-point DSP, and then get more accuracy?
>>	If any, please discuss on it! thanks
>>												Jim
>>
>>
>>
>>_______________________________________________
>>Tremor mailing list
>>Tremor at xiph.org
>>http://lists.xiph.org/mailman/listinfo/tremor
>
>_________________________________________________________________
>FREE pop-up blocking with the new MSN Toolbar - get it now! 
>http://toolbar.msn.click-url.com/go/onm00200415ave/direct/01/
>