[FFmpeg-devel] [PATCH] avformat/mpegts: use a buffer pool for PES payloads

[Marton Balint]

On Thu, 2 Apr 2020, Marton Balint wrote:

>
>
> On Thu, 2 Apr 2020, James Almer wrote:
>
>> On 3/29/2020 6:07 PM, James Almer wrote:
>>> This produces a small speed up in demuxing
>>> 
>>> Signed-off-by: James Almer <jamrial at gmail.com>
>>> ---
>>>  libavformat/mpegts.c | 20 ++++++++++++++++----
>>>  1 file changed, 16 insertions(+), 4 deletions(-)
>>> 
>>> diff --git a/libavformat/mpegts.c b/libavformat/mpegts.c
>>> index 7f56bacb2c..7746a524c4 100644
>>> --- a/libavformat/mpegts.c
>>> +++ b/libavformat/mpegts.c
>>> @@ -157,6 +157,8 @@ struct MpegTSContext {
>>>      int resync_size;
>>>      int merge_pmt_versions;
>>> 
>>> +    AVBufferPool *pool;
>>> +
>>>      /******************************************/
>>>      /* private mpegts data */
>>>      /* scan context */
>>> @@ -1177,8 +1179,7 @@ static int mpegts_push_data(MpegTSFilter *filter,
>>>                          pes->total_size = MAX_PES_PAYLOAD;
>>>
>>>                      /* allocate pes buffer */
>>> -                    pes->buffer = av_buffer_alloc(pes->total_size +
>>> - 
> AV_INPUT_BUFFER_PADDING_SIZE);
>>> +                    pes->buffer = av_buffer_pool_get(ts->pool);
>>>                      if (!pes->buffer)
>>>                          return AVERROR(ENOMEM);
>>> 
>>> @@ -1351,8 +1352,7 @@ skip:
>>>                      if (ret < 0)
>>>                          return ret;
>>>                      pes->total_size = MAX_PES_PAYLOAD;
>>> -                    pes->buffer = av_buffer_alloc(pes->total_size +
>>> - 
> AV_INPUT_BUFFER_PADDING_SIZE);
>>> +                    pes->buffer = av_buffer_pool_get(ts->pool);
>>>                      if (!pes->buffer)
>>>                          return AVERROR(ENOMEM);
>>>                      ts->stop_parse = 1;
>>> @@ -3032,6 +3032,10 @@ static int mpegts_read_header(AVFormatContext *s)
>>>      ts->stream     = s;
>>>      ts->auto_guess = 0;
>>> 
>>> +    ts->pool = av_buffer_pool_init(MAX_PES_PAYLOAD + 
> AV_INPUT_BUFFER_PADDING_SIZE, NULL);
>>> +    if (!ts->pool)
>>> +        return AVERROR(ENOMEM);
>>> +
>>>      if (s->iformat == &ff_mpegts_demuxer) {
>>>          /* normal demux */
>>> 
>>> @@ -3200,6 +3204,8 @@ static void mpegts_free(MpegTSContext *ts)
>>>
>>>      clear_programs(ts);
>>> 
>>> +    av_buffer_pool_uninit(&ts->pool);
>>> +
>>>      for (i = 0; i < NB_PID_MAX; i++)
>>>          if (ts->pids[i])
>>>              mpegts_close_filter(ts, ts->pids[i]);
>>> @@ -3295,6 +3301,12 @@ MpegTSContext 
> *avpriv_mpegts_parse_open(AVFormatContext *s)
>>>      ts->stream = s;
>>>      ts->auto_guess = 1;
>>> 
>>> +    ts->pool = av_buffer_pool_init(MAX_PES_PAYLOAD + 
> AV_INPUT_BUFFER_PADDING_SIZE, NULL);
>>> +    if (!ts->pool) {
>>> +        av_free(ts);
>>> +        return NULL;
>>> +    }
>>> +
>>>      mpegts_open_section_filter(ts, SDT_PID, sdt_cb, ts, 1);
>>>      mpegts_open_section_filter(ts, PAT_PID, pat_cb, ts, 1);
>>>      mpegts_open_section_filter(ts, EIT_PID, eit_cb, ts, 1);
>>
>> Will apply soon if no one objects.
>
> LGTM.

On second thought aint this going to waste too much memory on small 
packets? If packets are queued somewhere it can make a huge difference if 
every buffer is 200kb...

Does the speedup goes away if only unbound packets make use of the buffer 
pool?

Thanks,
Marton