[FFmpeg-devel] [PATCH 6/7] utvideoenc: align mangled buffer starts.

[Derek Buitenhuis]

On 22/08/2012 10:26 AM, Michael Niedermayer wrote:
> This is essential for fast SIMD accesses.
> The same should be done with the predict output.
> 
> Signed-off-by: Michael Niedermayer <michaelni at gmx.at>
> ---
>  libavcodec/utvideo.h    |    1 +
>  libavcodec/utvideoenc.c |   14 ++++++++------
>  2 files changed, 9 insertions(+), 6 deletions(-)

Is this a preemptive measure for future SIMD, or does it have direct
benefits in this patchset? Just curious, as I'm a n00b at SIMD.

> diff --git a/libavcodec/utvideo.h b/libavcodec/utvideo.h
> index 2441375..135e927 100644
> --- a/libavcodec/utvideo.h
> +++ b/libavcodec/utvideo.h
> @@ -75,6 +75,7 @@ typedef struct UtvideoContext {
>      int      interlaced;
>      int      frame_pred;
>  
> +    int slice_stride;
>      uint8_t *slice_bits, *slice_buffer[4];
>      int      slice_bits_size;

Please align with the rest of the declarations here.

>      for(i=0; i<c->planes; i++) {
> -        c->slice_buffer[i] = av_malloc(avctx->width * (avctx->height+1) +
> +        c->slice_stride = FFALIGN(avctx->width, 32);
> +        c->slice_buffer[i] = av_malloc(c->slice_stride * (avctx->height+2) +
>                                         FF_INPUT_BUFFER_PADDING_SIZE);

(avctx->height+2) -> (avctx->height + 2)

 
> -static void mangle_rgb_planes(uint8_t *dst[4], uint8_t *src, int step, int stride, int width,
> +static void mangle_rgb_planes(uint8_t *dst[4], int dst_stride, uint8_t *src, int step, int stride, int width,
>                                int height)

80 cols.

>      int i, j;
> -    int k=width;
> +    int k = 2*dst_stride;

See my comment from the previous patch. Also, 2*dst_stride -> 2 * dst_stride.

> -        mangle_rgb_planes(c->slice_buffer, pic->data[0], c->planes, pic->linesize[0], width,
> +        mangle_rgb_planes(c->slice_buffer, c->slice_stride, pic->data[0], c->planes, pic->linesize[0], width,
>                            height);

80 cols.


> -            ret = encode_plane(avctx, c->slice_buffer[i] + width,
> -                               c->slice_buffer[i], width,
> +            ret = encode_plane(avctx, c->slice_buffer[i] + 2*c->slice_stride,
> +                               c->slice_buffer[i], c->slice_stride,
>                                 width, height, &pb);

2*c->slice_stride -> 2 * c->slice_stride.

- Derek