[FFmpeg-devel] [PATCH v2] avcodec/v4l2_m2m_enc: Support changing qmin/qmax

[Andriy Gelman]

On Sat, 01. Feb 22:38, Mark Thompson wrote:
> On 19/01/2020 19:54, Andriy Gelman wrote:
> > From: Andriy Gelman <andriy.gelman at gmail.com>
> > 
> > Hard coded parameters for qmin and qmax are currently used to initialize
> > v4l2_m2m device. This commit uses values from avctx->{qmin,qmax} if they
> > are set.
> > 
> > Signed-off-by: Andriy Gelman <andriy.gelman at gmail.com>
> > ---
> >  libavcodec/v4l2_m2m_enc.c | 33 +++++++++++++++++++++++++++++++--
> >  1 file changed, 31 insertions(+), 2 deletions(-)
> > 
> > diff --git a/libavcodec/v4l2_m2m_enc.c b/libavcodec/v4l2_m2m_enc.c
> > index 8059e3bb48f..318be0d3379 100644
> > --- a/libavcodec/v4l2_m2m_enc.c
> > +++ b/libavcodec/v4l2_m2m_enc.c
> > @@ -31,10 +31,25 @@
> >  #include "v4l2_context.h"
> >  #include "v4l2_m2m.h"
> >  #include "v4l2_fmt.h"
> > +#include "internal.h"
> >  
> >  #define MPEG_CID(x) V4L2_CID_MPEG_VIDEO_##x
> >  #define MPEG_VIDEO(x) V4L2_MPEG_VIDEO_##x
> >  
> > +#define CLIP_MIN_MAX(in, min_val, max_val, name, logctx)    \
> > +    do {                                                    \
> > +        if ((in) < (min_val)) {                             \
> > +            av_log((logctx), AV_LOG_WARNING,                \
> > +                   "Adjusted: " name " (%d)\n", (min_val)); \
> > +            in = min_val;                                   \
> > +        }                                                   \
> > +        if ((in) > (max_val)) {                             \
> > +            av_log((logctx), AV_LOG_WARNING,                \
> > +                   "Adjusted: " name " (%d)\n", (max_val)); \
> > +            (in) = (max_val);                               \
> > +        }                                                   \
> > +    } while (0)
> > +
> >  static inline void v4l2_set_timeperframe(V4L2m2mContext *s, unsigned int num, unsigned int den)
> >  {
> >      struct v4l2_streamparm parm = { 0 };
> > @@ -232,8 +247,15 @@ static int v4l2_prepare_encoder(V4L2m2mContext *s)
> >          return 0;
> >      }
> >  
> > -    if (qmin != avctx->qmin || qmax != avctx->qmax)
> > -        av_log(avctx, AV_LOG_WARNING, "Encoder adjusted: qmin (%d), qmax (%d)\n", qmin, qmax);
> > +    if (avctx->qmin >= 0) {
> > +        CLIP_MIN_MAX(avctx->qmin, qmin, qmax, "qmin", avctx);
> > +        qmin = avctx->qmin;
> > +    }
> > +
> > +    if (avctx->qmax >= 0) {
> > +        CLIP_MIN_MAX(avctx->qmax, qmin, qmax, "qmax", avctx);
> > +        qmax = avctx->qmax;
> > +    }
> >  
> >      v4l2_set_ext_ctrl(s, qmin_cid, qmin, "minimum video quantizer scale");
> >      v4l2_set_ext_ctrl(s, qmax_cid, qmax, "maximum video quantizer scale");
> > @@ -349,6 +371,12 @@ static const AVOption options[] = {
> >      { NULL },
> >  };
> >  
> > +static const AVCodecDefault v4l2_m2m_defaults[] = {
> > +    { "qmin", "-1" },
> > +    { "qmax", "-1" },
> > +    { NULL },
> > +};
> > +
> >  #define M2MENC_CLASS(NAME) \
> >      static const AVClass v4l2_m2m_ ## NAME ## _enc_class = { \
> >          .class_name = #NAME "_v4l2m2m_encoder", \
> > @@ -370,6 +398,7 @@ static const AVOption options[] = {
> >          .send_frame     = v4l2_send_frame, \
> >          .receive_packet = v4l2_receive_packet, \
> >          .close          = v4l2_encode_close, \
> > +        .defaults       = v4l2_m2m_defaults, \
> >          .capabilities   = AV_CODEC_CAP_HARDWARE | AV_CODEC_CAP_DELAY, \
> >          .wrapper_name   = "v4l2m2m", \
> >      };
> > 

> 
> Can we avoid some of the clumsiness around clipping twice in different ways by querying the quantiser values from the encode device here?
> 

thanks, I'll investigate.

> E.g. on s5p-mfc (exynos) I can see:
> 
>           h264_minimum_qp_value 0x00990a61 (int)    : min=0 max=51 step=1 default=1 value=1
>           h264_maximum_qp_value 0x00990a62 (int)    : min=0 max=51 step=1 default=51 value=51
> 
> which matches the expected values for 8-bit H.264, but then for VP8 there is:
> 
>            vpx_minimum_qp_value 0x00990afb (int)    : min=0 max=11 step=1 default=0 value=0
>            vpx_maximum_qp_value 0x00990afc (int)    : min=0 max=127 step=1 default=127 value=127
> 
> which is going to pretty much entirely stop qmin from doing anything useful, and it would be nice to diagnose that.

The max for vpx_maximum_qp_value has been bothering me. The usual qp range for vp8/9 is [0,63]. 

I checked today with a dev on #v4l and he told me that only 6 bits are used in
the driver so it's probably a bug. I'll look into this further. 

On a related note, I actually haven't been able to play a stream that's been
compressed with vp8/9 on the s5p-mfc (odroid xu4 for me). Have you had any
luck? 

Andriy