[FFmpeg-devel] [PATCH 1/3] cbs_h2645: Avoid memcpy when splitting fragment

[Mark Thompson]

On 21/11/18 18:34, Andreas Rheinhardt wrote:
> Now memcpy is avoided for NAL units that don't contain 0x03 escape
> characters.
> 
> Improves performance of cbs_h2645_fragment_add_nals from 36940
> decicycles to 6364 decicycles based on 8 runs with a 5.1 Mb/s H.264
> sample (262144 runs each).
> 
> Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt at googlemail.com>
> ---
>  libavcodec/cbs_h2645.c   | 29 +++++++++++++++++------------
>  libavcodec/h2645_parse.h |  5 +++++
>  2 files changed, 22 insertions(+), 12 deletions(-)
> 
> diff --git a/libavcodec/cbs_h2645.c b/libavcodec/cbs_h2645.c
> index a2d0170e97..6846cad0bb 100644
> --- a/libavcodec/cbs_h2645.c
> +++ b/libavcodec/cbs_h2645.c
> @@ -532,24 +532,29 @@ static int cbs_h2645_fragment_add_nals(CodedBitstreamContext *ctx,
>      for (i = 0; i < packet->nb_nals; i++) {
>          const H2645NAL *nal = &packet->nals[i];
>          size_t size = nal->size;
> -        uint8_t *data;
> -
>          // Remove trailing zeroes.
>          while (size > 0 && nal->data[size - 1] == 0)
>              --size;
>          av_assert0(size > 0);
>  
> -        data = av_malloc(size + AV_INPUT_BUFFER_PADDING_SIZE);
> -        if (!data)
> -            return AVERROR(ENOMEM);
> -        memcpy(data, nal->data, size);
> -        memset(data + size, 0, AV_INPUT_BUFFER_PADDING_SIZE);
> +        if (nal->data == nal->raw_data) {
> +            err = ff_cbs_insert_unit_data(ctx, frag, -1, nal->type,
> +                                (uint8_t*)nal->data, size, frag->data_ref);
> +            if (err < 0)
> +                return err;
> +        } else {
> +            uint8_t *data = av_malloc(size + AV_INPUT_BUFFER_PADDING_SIZE);
> +            if (!data)
> +                return AVERROR(ENOMEM);
> +            memcpy(data, nal->data, size);
> +            memset(data + size, 0, AV_INPUT_BUFFER_PADDING_SIZE);
>  
> -        err = ff_cbs_insert_unit_data(ctx, frag, -1, nal->type,
> -                                      data, size, NULL);
> -        if (err < 0) {
> -            av_freep(&data);
> -            return err;
> +            err = ff_cbs_insert_unit_data(ctx, frag, -1, nal->type,
> +                                          data, size, NULL);
> +            if (err < 0) {
> +                av_freep(&data);
> +                return err;
> +            }
>          }
>      }
>  
> diff --git a/libavcodec/h2645_parse.h b/libavcodec/h2645_parse.h
> index 2e29ad26cb..a0a5ca5868 100644
> --- a/libavcodec/h2645_parse.h
> +++ b/libavcodec/h2645_parse.h
> @@ -86,6 +86,11 @@ int ff_h2645_extract_rbsp(const uint8_t *src, int length, H2645RBSP *rbsp,
>  
>  /**
>   * Split an input packet into NAL units.
> + * If data == raw_data holds true for a NAL unit of the returned pkt, then
> + * said NAL unit does not contain any emulation_prevention_three_byte and
> + * the data is contained in the input buffer pointed to by buf.
> + * Otherwise, the unescaped data is part of the rbsp_buffer described by the
> + * packet's H2645RBSP.
>   */
>  int ff_h2645_packet_split(H2645Packet *pkt, const uint8_t *buf, int length,
>                            void *logctx, int is_nalff, int nal_length_size,
> 

LGTM, tested, applied.

Thanks,

- Mark