[FFmpeg-devel] [PATCH] updated LGPL AC-3 decoder

[Justin Ruggles]

Michael Niedermayer wrote:
> Hi
> 
> On Sun, May 06, 2007 at 06:49:00PM -0400, Justin Ruggles wrote:
> [...]
> 
>>>[...]
>>>
>>>
>>>>+    /* convert float to 16-bit integer */
>>>>+    for(ch=0; ch<ctx->out_channels; ch++) {
>>>>+        for(i=0; i<256; i++) {
>>>>+            ctx->output[ch][i] = ctx->output[ch][i] * ctx->mul_bias +
>>>>+                                 ctx->add_bias;
>>>>+        }
>>>>+        ctx->dsp.float_to_int16(ctx->int_output[ch], ctx->output[ch], 256);
>>>
>>>
>>>this isnt exactly ideal
>>>the bias should be applied in some more efficient way like multiplying some
>>>scaling coefficients, allthough i dunno if thats possible in AC3 ...
>>
>>I changed it so that the bias is multiplied in before the IMDCT along
>>with the dialogue normalization and dynamic range compression.
>>
>>On top of your suggestions, I also did several other changes.  I added
>>dialogue normalization, put in some references to the specs in various
>>places, fixed some bugs, simplified a few things, added more checks for
>>error conditions, and removed some unused variables.
>>
>>I still can't figure out why the 5.1 decoding is so much slower than
>>liba52.  My guess is the downmixing, but it might be other things as
>>well.  Stereo decoding is pretty much the same speed on my system as
>>liba52, but 5.1 decoding takes twice as long...
> 
> 
> you mean 5.1 with downmixing to stereo or true 5.1 ?
> 
> the downmixing should be done before the imdct, that way you dont need to
> do the imdct on 6 channels but rather just on 2 (i didnt check where the
> downmix is done exactly but its just IIRC ...)
> 
> 
> [...]
> 
> 
>>+    /* the rest of the bsi. read twice for dual mono mode. */
>>+    for(i=0; i<=(ctx->acmod == AC3_ACMOD_DUALMONO); i++) {
>>+        ctx->dialnorm[i] = dialnorm_tbl[get_bits(gb, 5)]; // dialogue normalization
>>+        if(get_bits1(gb))
>>+            skip_bits(gb, 8); // skip compression gain
>>+        if(get_bits1(gb))
>>+            skip_bits(gb, 8); // skip language code
>>+        if(get_bits1(gb))
>>+            skip_bits(gb, 7); // skip audio production information
>>+    } while(i--);
> 
> 
> hmm for(){}while(); ?

yeah, you haven't heard of that construct? ;)

really though, mans is right, i missed that one when i converted it from
a do{}while() to a for().

> and patch looks ok

great.  now...

Should I apply it all at once or try to go back to the original SoC code
and do incremental changes?  I looked back at the original, and my best
guess is that I changed about 75% of the code.  I could do it either way
really.

-Justin