[FFmpeg-devel] [PATCH 2/2] vf_blend: Use memcpy when opacity is 0

[Paul B Mahol]

On 2/14/16, Timothy Gu <timothygu99 at gmail.com> wrote:
> For xor mode, 1.8x faster on Haswell.
> ---
>  libavfilter/vf_blend.c | 6 ++++++
>  1 file changed, 6 insertions(+)
>
> diff --git a/libavfilter/vf_blend.c b/libavfilter/vf_blend.c
> index c24013d..599084f 100644
> --- a/libavfilter/vf_blend.c
> +++ b/libavfilter/vf_blend.c
> @@ -130,6 +130,7 @@ static void blend_copy ## src(const uint8_t *top,
> ptrdiff_t top_linesize,    \
>  }
>
>  COPY(top)
> +COPY(bottom)
>
>  #undef COPY
>
> @@ -512,6 +513,7 @@ static int config_output(AVFilterLink *outlink)
>          case BLEND_MULTIPLY128:param->blend = is_16bit ?
> blend_multiply128_16bit: blend_multiply128_8bit;break;
>          case BLEND_NEGATION:   param->blend = is_16bit ?
> blend_negation_16bit   : blend_negation_8bit;   break;
>          case BLEND_NORMAL:     param->blend = param->opacity == 1 ?
> blend_copytop :
> +                                              param->opacity == 0 ?
> blend_copybottom :
>                                                is_16bit ? blend_normal_16bit
>     : blend_normal_8bit;     break;
>          case BLEND_OR:         param->blend = is_16bit ? blend_or_16bit
>     : blend_or_8bit;         break;
>          case BLEND_OVERLAY:    param->blend = is_16bit ?
> blend_overlay_16bit    : blend_overlay_8bit;    break;
> @@ -528,6 +530,10 @@ static int config_output(AVFilterLink *outlink)
>          if (ARCH_X86)
>              ff_blend_init_x86(param, is_16bit);
>
> +        if (param->opacity == 0 && param->mode != BLEND_NORMAL) {
> +            param->blend = blend_copytop;
> +        }
> +
>          if (s->all_expr && !param->expr_str) {
>              param->expr_str = av_strdup(s->all_expr);
>              if (!param->expr_str)

lgtm if bitexact for all modes.