[FFmpeg-devel] [PATCH 13/14] mmal: add option copy_frame to support retrieving sw frames w/o copy

[wm4]

On Sat, 16 Dec 2017 20:26:56 +0800
Wang Bin <wbsecg1 at gmail.com> wrote:

> 2017-12-16 19:47 GMT+08:00 wm4 <nfxjfg at googlemail.com>:
> > On Sat, 16 Dec 2017 13:48:05 +0800
> > Wang Bin <wbsecg1 at gmail.com> wrote:
> >  
> >> 2017-12-16 2:50 GMT+08:00 wm4 <nfxjfg at googlemail.com>:  
> >> > On Fri, 15 Dec 2017 15:05:50 +0800
> >> > wbsecg1 at gmail.com wrote:
> >> >  
> >> >> From: wang-bin <wbsecg1 at gmail.com>
> >> >>
> >> >> mmal buffer->data is already in host memory. AFAIK decoders implemented in omx must
> >> >> be configured to output frames to either memory or something directly used by renderer,
> >> >> for example mediacodec surface, mmal buffer and omxil eglimage.
> >> >> test result: big buck bunny 1080p fps increases from about 100 to 110 if copy_frame is
> >> >> turned off
> >> >> ---
> >> >>  libavcodec/mmaldec.c | 31 +++++++++++++++++++++++--------
> >> >>  1 file changed, 23 insertions(+), 8 deletions(-)
> >> >>
> >> >> diff --git a/libavcodec/mmaldec.c b/libavcodec/mmaldec.c
> >> >> index c1cfb09283..9cd6c6558f 100644
> >> >> --- a/libavcodec/mmaldec.c
> >> >> +++ b/libavcodec/mmaldec.c
> >> >> @@ -69,6 +69,7 @@ typedef struct MMALDecodeContext {
> >> >>      AVClass *av_class;
> >> >>      int extra_buffers;
> >> >>      int extra_decoder_buffers;
> >> >> +    int copy_frame;
> >> >>
> >> >>      MMAL_COMPONENT_T *decoder;
> >> >>      MMAL_QUEUE_T *queue_decoded_frames;
> >> >> @@ -139,7 +140,6 @@ static int ffmmal_set_ref(AVFrame *frame, FFPoolRef *pool,
> >> >>      atomic_fetch_add_explicit(&ref->pool->refcount, 1, memory_order_relaxed);
> >> >>      mmal_buffer_header_acquire(buffer);
> >> >>
> >> >> -    frame->format = AV_PIX_FMT_MMAL;
> >> >>      frame->data[3] = (uint8_t *)ref->buffer;
> >> >>      return 0;
> >> >>  }
> >> >> @@ -650,20 +650,34 @@ static int ffmal_copy_frame(AVCodecContext *avctx,  AVFrame *frame,
> >> >>
> >> >>          if ((ret = ffmmal_set_ref(frame, ctx->pool_out, buffer)) < 0)
> >> >>              goto done;
> >> >> +        frame->format = AV_PIX_FMT_MMAL;
> >> >>      } else {
> >> >>          int w = FFALIGN(avctx->width, 32);
> >> >>          int h = FFALIGN(avctx->height, 16);
> >> >>          uint8_t *src[4];
> >> >>          int linesize[4];
> >> >>
> >> >> -        if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
> >> >> -            goto done;
> >> >> +        if (ctx->copy_frame) {
> >> >> +            if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
> >> >> +                goto done;
> >> >>
> >> >> -        av_image_fill_arrays(src, linesize,
> >> >> -                             buffer->data + buffer->type->video.offset[0],
> >> >> -                             avctx->pix_fmt, w, h, 1);
> >> >> -        av_image_copy(frame->data, frame->linesize, src, linesize,
> >> >> -                      avctx->pix_fmt, avctx->width, avctx->height);
> >> >> +            av_image_fill_arrays(src, linesize,
> >> >> +                                buffer->data + buffer->type->video.offset[0],
> >> >> +                                avctx->pix_fmt, w, h, 1);
> >> >> +            av_image_copy(frame->data, frame->linesize, src, linesize,
> >> >> +                        avctx->pix_fmt, avctx->width, avctx->height);
> >> >> +        } else {
> >> >> +            if ((ret = ff_decode_frame_props(avctx, frame)) < 0)
> >> >> +                goto done;
> >> >> +            /* buffer->type->video.offset/pitch[i]; is always 0 */
> >> >> +            av_image_fill_arrays(src, linesize,
> >> >> +                                buffer->data + buffer->type->video.offset[0],
> >> >> +                                avctx->pix_fmt, w, h, 1);
> >> >> +            if ((ret = ffmmal_set_ref(frame, ctx->pool_out, buffer)) < 0)
> >> >> +                goto done;
> >> >> +            memcpy(frame->data, src, sizeof(src));
> >> >> +            memcpy(frame->linesize, linesize, sizeof(linesize));
> >> >> +        }
> >> >>      }
> >> >>
> >> >>      frame->pts = buffer->pts == MMAL_TIME_UNKNOWN ? AV_NOPTS_VALUE : buffer->pts;
> >> >> @@ -842,6 +856,7 @@ AVHWAccel ff_wmv3_mmal_hwaccel = {
> >> >>  static const AVOption options[]={
> >> >>      {"extra_buffers", "extra buffers", offsetof(MMALDecodeContext, extra_buffers), AV_OPT_TYPE_INT, {.i64 = 10}, 0, 256, 0},
> >> >>      {"extra_decoder_buffers", "extra MMAL internal buffered frames", offsetof(MMALDecodeContext, extra_decoder_buffers), AV_OPT_TYPE_INT, {.i64 = 10}, 0, 256, 0},
> >> >> +    {"copy_frame", "copy deocded data to avframe", offsetof(MMALDecodeContext, copy_frame), AV_OPT_TYPE_BOOL, {.i64 = 1}, 0, 256, 0},
> >> >>      {NULL}
> >> >>  };
> >> >>  
> >> >
> >> > Didn't check too closely what exactly the patch does, but adding an
> >> > option for it sounds very wrong. The user select in the get_format
> >> > callback whether a GPU surface is output (MMAL pixfmt), or software.  
> >>
> >> Avoid copying data from mmal buffer->data to avframe data. Instead,
> >> just fill strides and address of each plane in avframe, and add a
> >> reference to mmal buffer.  
> >
> > Does it make sure to keep the mmal buffer pool alive then? Otherwise a
> > decoded AVFrame would become invalid after closing and destroying the
> > decoder.  
> 
> Yes. ffmmal_set_ref() is called like hw pixfmt code and mmal buffer's
> reference is increased. otherwise the displayed frame is weired
> because the buffer may be invalid as you say. I tested it in mpv.
> 
> > Why would this need a new option?  
> 
> For testing purposes. You can compare the performance with the option.

Fine then I guess. It could still be surprising to the API user that
holding a frame will keep a large chunk of memory alive (the whole
pool), but I guess that's unavoidable with mmal's design. The API user
could still copy the frame manually if it becomes a problem.