[Ffmpeg-devel] [PATCH] another vorbis optimization

[Luca Barbato]

Loren Merritt wrote:
> Now with optimized windowing too, for a total of 15% speedup.

Wonderful =)

>  static int vorbis_parse_audio_packet(vorbis_context *vc) {
>      GetBitContext *gb=&vc->gb;
> @@ -1490,6 +1507,7 @@
>      uint_fast8_t res_num=0;
>      int_fast16_t retlen=0;
>      uint_fast16_t saved_start=0;

what about moving bias somewhere else (like in the vorbis_context) and
spare a branch?

> +    int bias= (vc->dsp.float_to_int16 == ff_float_to_int16_c) ? 385 : 0;
>  
>      if (get_bits1(gb)) {
>          av_log(vc->avccontext, AV_LOG_ERROR, "Not a Vorbis I audio packet.\n");

[...]

>                  saved_start=(vc->blocksize_1-vc->blocksize_0)/4;
>                  buf += vc->blocksize_1/2;
> -                for(i=0;i<saved_start;++i) {
> -                    saved[i]=buf[i];
> +                if(vc->dsp.float_to_int16 != ff_float_to_int16_c) {

as ugly as could be: what about having 2 bias values in the context:
 scalar [385, 0], vector [0,(15<<23)] and spare some other branches?

> +                    for(i=0; i<saved_start; i++)
> +                        ((uint32_t*)saved)[i] = ((uint32_t*)buf)[i] + (15<<23);
> +                } else {
> +                    for(i=0; i<saved_start; i++)
> +                        saved[i] = buf[i];
>                  }

[....]

> +        int_fast32_t tmp = ((int32_t*)src)[i];
> +        if(tmp & 0xf0000){

Is it really slower now?

> +//          tmp = (0x43c0ffff - tmp)>>31; //ask gcc devs why this is slower
> +            if(tmp > 0x43c0ffff) tmp = 0xFFFF;
> +            else                 tmp = 0;
> +        }


-- 

Luca Barbato

Gentoo/linux Gentoo/PPC
http://dev.gentoo.org/~lu_zero