[FFmpeg-cvslog] r12286 - trunk/libavcodec/ac3dec.c

[Ivan Kalvachev]

On Sat, Mar 1, 2008 at 5:48 PM, jbr <subversion at mplayerhq.hu> wrote:
> Author: jbr
>  Date: Sat Mar  1 16:48:18 2008
>  New Revision: 12286
>
>  Log:
>  pre-calculate downmix coefficient sums. 2% faster 5.1-to-stereo downmixing.
>
>  Modified:
>    trunk/libavcodec/ac3dec.c
>
>  Modified: trunk/libavcodec/ac3dec.c
>  ==============================================================================
>  --- trunk/libavcodec/ac3dec.c   (original)
>  +++ trunk/libavcodec/ac3dec.c   Sat Mar  1 16:48:18 2008
>  @@ -154,6 +154,7 @@ typedef struct {
>      int center_mix_level;                   ///< Center mix level index
>      int surround_mix_level;                 ///< Surround mix level index
>      float downmix_coeffs[AC3_MAX_CHANNELS][2];  ///< stereo downmix coefficients
>  +    float downmix_coeff_sum[2];             ///< sum of downmix coeffs for each output channel
>      float dynamic_range[2];                 ///< dynamic range
>      int   cpl_coords[AC3_MAX_CHANNELS][18]; ///< coupling coordinates
>      int   num_cpl_bands;                    ///< number of coupling bands
>  @@ -402,6 +403,12 @@ static void set_downmix_coeffs(AC3Decode
>          int nf = s->channel_mode - 4;
>          s->downmix_coeffs[nf][0] = s->downmix_coeffs[nf+1][1] = smix;
>      }
>  +
>  +    s->downmix_coeff_sum[0] = s->downmix_coeff_sum[1] = 0.0f;
>  +    for(i=0; i<s->fbw_channels; i++) {
>  +        s->downmix_coeff_sum[0] += s->downmix_coeffs[i][0];
>  +        s->downmix_coeff_sum[1] += s->downmix_coeffs[i][1];
>  +    }
>   }
>
>   /**
>  @@ -732,18 +739,16 @@ static inline void do_imdct(AC3DecodeCon
>   static void ac3_downmix(AC3DecodeContext *s)
>   {
>      int i, j;
>  -    float v0, v1, s0, s1;
>  +    float v0, v1;
>
>      for(i=0; i<256; i++) {
>  -        v0 = v1 = s0 = s1 = 0.0f;
>  +        v0 = v1 = 0.0f;
>          for(j=0; j<s->fbw_channels; j++) {
>              v0 += s->output[j][i] * s->downmix_coeffs[j][0];
>              v1 += s->output[j][i] * s->downmix_coeffs[j][1];
>  -            s0 += s->downmix_coeffs[j][0];
>  -            s1 += s->downmix_coeffs[j][1];
>          }
>  -        v0 /= s0;
>  -        v1 /= s1;
>  +        v0 /= s->downmix_coeff_sum[0];
>  +        v1 /= s->downmix_coeff_sum[1];

You know, multiply is usually faster than division, so if these sums
are constant, you can probably calculate reciprocals and multiply in
the loop.
{sr0=1/s0;...; v0*=sr0};