[FFmpeg-devel] Make ff_acelp_lsp2lpc() more generic

[Benjamin Larsson]

Michael Niedermayer skrev 2010-06-16 17:41:
> On Wed, Jun 16, 2010 at 02:36:44PM +0100, M?ns Rullg?rd wrote:
>    
>> Mohamed Naufal<naufal11 at gmail.com>  writes:
>>
>>      
>>> Hi
>>>
>>> The attached patch allows using the function by both G.729&  G.723.1.
>>>
>>> -void ff_acelp_lsp2lpc(int16_t* lp, const int16_t* lsp, int lp_half_order)
>>> +void ff_acelp_lsp2lpc(int16_t* lp, const int16_t* lsp, int lp_half_order, int shift, int rounder, int sign)
>>>   {
>>>       int i;
>>>       int f1[lp_half_order+1]; // (3.22)
>>> @@ -103,9 +103,10 @@ void ff_acelp_lsp2lpc(int16_t* lp, const int16_t* lsp, int lp_half_order)
>>>           int ff1 = f1[i] + f1[i-1]; // (3.22)
>>>           int ff2 = f2[i] - f2[i-1]; // (3.22)
>>>
>>> -        ff1 += 1<<  10; // for rounding
>>> -        lp[i]    = (ff1 + ff2)>>  11; // divide by 2 and (3.22) ->  (3.12)
>>> -        lp[(lp_half_order<<  1) + 1 - i] = (ff1 - ff2)>>  11; // divide by 2 and (3.22) ->  (3.12)
>>> +        ff1 += rounder; // for rounding
>>> +        // divide by 2 and scale
>>> +        lp[i]    =                         av_clip_int16(sign * av_clip_int16((ff1 + ff2)>>  (shift + 1)));
>>> +        lp[(lp_half_order<<  1) + 1 - i] = av_clip_int16(sign * av_clip_int16((ff1 - ff2)>>  (shift + 1)));
>>>        
>> That clipping will probably have a severe performance impact.  Did you
>> benchmark this?
>>      
> also, some of the cliping is clearly unneeded
> and we need to find out of what remains is needed
>    
>

I told naufal to focus on completing the decoder. I see 4 options. 
Duplicate the  function, add a g723.1 flag, add more function arguments 
like the patch or just ignore bit exactness.

I think the best and least complex solution is to add a g723.1 flag and 
add a conditional in the loop, "if (g723.1) do this; else do that". This 
code is executed every subframe (4 times per frame I think). So I think 
a conditional would be an ok compromise. It is not possible to benchmark 
the slowdown for the g729 decoder as the complete decoder never got merged.

MvH
Benjamin Larsson