Speex dev - Jul 2007 - Speex optimization and 12 bits conversion for 12 bits ADC

Hi?Jean-Marc?

Thanks for your suggestions very much!
> > I am porting speex on ARM7TDMI, I have done some optimization, the
> > result is that the encoder and decoder need about 60 MCPS for 5.96kbps
> > bitrate and complexity 0. Can someone give me informtion about Speex
> > optimization on ARM7?
>
> That's quite good. A few suggestions here:
> 1) Don't use complexity 0 unless you're really desperate (the CPU
gain
> over complexity 1 is marginal and there's a cost in quality)
> 2) Have you enabled the current ARM4 assembly bits?
> 3) There are a couple functions that could definitely benefit from
> assembly optimisations. They used to be optimised for AMR4, but I had to
> remove that when I optimised the C version. Those are filter_mem16(),
> iir_mem16() and fir_mem16(). Let me know if you're going to do that
> because the best way to do it is not to start from the current C code.
>
Yes, I have enabled the the current ARM4 assembly bits. I have tried
to optimize these functions(filter_mem16(),iir_mem16() and
fir_mem16()) in C code, but the effect is very small, and I am going
to do assembly optimization for these functions. Do you have another C
code about these functions for assembly optimisations? Can you give me
some information and suggestions?
> > Another question, my ADC and DAC are 12 bits, but Speex codec is
> > 16bits, Did someone try to modify speex to 12 bits? I think if I
> > modify speex to 12 bits, the computation power will be reduced, is it
> > right?
>
> Wrong. Speex would still need 16 bits internally. I also recommend you
> only shift by 3 so that you're sure never to have clipping problems at
> 16 bits.
>
My ADC is 12 bits, and the data from 12 bits ADC just has 12-bit
accuracy, means that I need fill the low 4 bit with zero to make the
input data is 16 bit, I means that Some computation may be simplified
so that to reduce computation power because the just 12 bit is useful.
And the DAC is also 12 bit, it means that I just need the high 12 bit
and discard the low 4 bit after decoer?

Thanks.

Best Regards,
Cheng Rong

> Yes, I have enabled the the current ARM4 assembly bits. I have tried
> to optimize these functions(filter_mem16(),iir_mem16() and
> fir_mem16()) in C code, but the effect is very small, and I am going
> to do assembly optimization for these functions. Do you have another C
> code about these functions for assembly optimisations? Can you give me
> some information and suggestions?
Basically, if either of these apply to your CPU:
1) A store is more costly than a load (e.g. you have write-through cache)
2) You have a MAC instruction where it's complicated to load the
accumulator every time
then you should consider implementing filter_mem16 similarly to the
commented version of filter_mem2() at the end of filters_bfin.h.
>> > Another question, my ADC and DAC are 12 bits, but Speex codec is
>> > 16bits, Did someone try to modify speex to 12 bits? I think if I
>> > modify speex to 12 bits, the computation power will be reduced, is
it
>> > right?
>>
>> Wrong. Speex would still need 16 bits internally. I also recommend you
>> only shift by 3 so that you're sure never to have clipping problems
at
>> 16 bits.
> 
> My ADC is 12 bits, and the data from 12 bits ADC just has 12-bit
> accuracy, means that I need fill the low 4 bit with zero to make the
> input data is 16 bit, I means that Some computation may be simplified
> so that to reduce computation power because the just 12 bit is useful.
> And the DAC is also 12 bit, it means that I just need the high 12 bit
> and discard the low 4 bit after decoer?
That's what I said. It's not going to work. It's not because your
input
signal is 12-bit deep that suddenly the LPC and LSP coefs will accept
being quantised in 12 bits. Also, as I said in my previous email, I'd
recommend shifting your 12-bit data only by 3 (not by 4) when converting
to 16-bit. That'll give you a bit of headroom.

Cheers,

	Jean-Marc

Hello,
 What is meant by crude delay in PESQ quality testing?
And what is its tolerant limit for good quality.
rgds,
Sharan