[FFmpeg-devel] [PATCH] VP8: simplify splitmv parsing

[David Conrad]

On Jun 26, 2010, at 8:22 PM, Ronald S. Bultje wrote:

> Hi,
> 
> related to the previous patch, this patch simplifies the splitMV
> reading since we don't need to spread out all MVs across 4x4 subblocks
> (except if they are actually coded that way). This probably leads to a
> modest speedup but I didn't test that.
> 
> Ronald

> Index: ffmpeg-svn/libavcodec/vp8.c
> ===================================================================
> --- ffmpeg-svn.orig/libavcodec/vp8.c	2010-06-26 14:22:50.000000000 -0400
> +++ ffmpeg-svn/libavcodec/vp8.c	2010-06-26 16:28:45.000000000 -0400
> @@ -590,44 +590,54 @@
>  
>  /**
>   * Split motion vector prediction, 16.4.
> + * @returns the number of motion vectors parsed (2, 4 or 16)
>   */
> -static void decode_splitmvs(VP8Context    *s,  VP56RangeCoder *c,
> +static int decode_splitmvs(VP8Context    *s,  VP56RangeCoder *c,
>                              VP8Macroblock *mb, VP56mv         *base_mv)
>  {
>      int part_idx = mb->partitioning =
>          vp8_rac_get_tree(c, vp8_mbsplit_tree, vp8_mbsplit_prob);
>      int n, num = vp8_mbsplit_count[part_idx];
> -    VP56mv part_mv[16];
> +    const uint8_t *mbsplits = vp8_mbsplits[part_idx],
> +                  *firstidx = vp8_mbfirstidx[part_idx];
>  
>      for (n = 0; n < num; n++) {
> -        int k = vp8_mbfirstidx[part_idx][n];
> -        const VP56mv *left  = (k & 3) ? &mb->bmv[k - 1] : &mb[-1].bmv[k + 3],
> -                     *above = (k > 3) ? &mb->bmv[k - 4] : &mb[-s->mb_stride].bmv[k + 12];
> -        const uint8_t *submv_prob = get_submv_prob(left, above);
> +        int k = firstidx[n];
> +        const VP56mv *left, *above;
> +        const uint8_t *submv_prob;
> +
> +        if (!(k & 3)) {
> +            VP8Macroblock *left_mb = &mb[-1];
> +            left = &left_mb->bmv[vp8_mbsplits[left_mb->partitioning][k + 3]];
> +        } else
> +            left  = &mb->bmv[mbsplits[k - 1]];
> +        if (k <= 3) {
> +            VP8Macroblock *above_mb = &mb[-s->mb_stride];
> +            above = &above_mb->bmv[vp8_mbsplits[above_mb->partitioning][k + 12]];
> +        } else
> +            above = &mb->bmv[mbsplits[k - 4]];
> +
> +        submv_prob = get_submv_prob(left, above);

I was thinking about splitting out bmv to its own array like in the mpeg decoders partly to make this cleaner (maybe). The idea being to switch(partitioning) here and fill_rect, though I'm not sure which would be faster or if that method would even be less LoC.

So this is ok for now.