[FFmpeg-devel] avcodec/libx264: update notes to explain the scale chosen for ROI encoding

Submitted by Guo, Yejun on Jan. 29, 2019, 10:14 a.m.

Details

Message ID 1548756871-6918-1-git-send-email-yejun.guo@intel.com
State New
Headers show

Commit Message

Guo, Yejun Jan. 29, 2019, 10:14 a.m.
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
---
 libavcodec/libx264.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

Comments

Mark Thompson Feb. 4, 2019, 9:33 a.m.
On 29/01/2019 10:14, Guo, Yejun wrote:
> Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
> ---
>  libavcodec/libx264.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/libavcodec/libx264.c b/libavcodec/libx264.c
> index a3493f3..8c96728 100644
> --- a/libavcodec/libx264.c
> +++ b/libavcodec/libx264.c
> @@ -383,7 +383,9 @@ static int X264_frame(AVCodecContext *ctx, AVPacket *pkt, const AVFrame *frame,
>                          qoffset = roi->qoffset.num * 1.0f / roi->qoffset.den;
>                          qoffset = av_clipf(qoffset, -1.0f, 1.0f);
>  
> -                        // 25 is a number that I think it is a possible proper scale value.
> +                        /* qp range of x264 is from 0 to 51, just choose 25 as the scale value,
> +                         * so the range of final qoffset is [-25.0, 25.0].
> +                         */
>                          qoffset = qoffset * 25;
>  
>                          for (int y = starty; y < endy; y++) {
> 

I'm not understanding where this number has come from at all.  Is 25 purely arbitrary?  Why does the effect vary by bit depth?

From the docs explaining the best/worst behaviour on the range, I would expect this to map 0 to 0 and 1 to 3 + 6 * bit_depth (positive or negative).

- Mark
Guo, Yejun Feb. 11, 2019, 8:15 a.m.
> -----Original Message-----

> From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On Behalf

> Of Mark Thompson

> Sent: Monday, February 04, 2019 5:34 PM

> To: ffmpeg-devel@ffmpeg.org

> Subject: Re: [FFmpeg-devel] [PATCH] avcodec/libx264: update notes to

> explain the scale chosen for ROI encoding

> 

> On 29/01/2019 10:14, Guo, Yejun wrote:

> > Signed-off-by: Guo, Yejun <yejun.guo@intel.com>

> > ---

> >  libavcodec/libx264.c | 4 +++-

> >  1 file changed, 3 insertions(+), 1 deletion(-)

> >

> > diff --git a/libavcodec/libx264.c b/libavcodec/libx264.c

> > index a3493f3..8c96728 100644

> > --- a/libavcodec/libx264.c

> > +++ b/libavcodec/libx264.c

> > @@ -383,7 +383,9 @@ static int X264_frame(AVCodecContext *ctx,

> AVPacket *pkt, const AVFrame *frame,

> >                          qoffset = roi->qoffset.num * 1.0f / roi->qoffset.den;

> >                          qoffset = av_clipf(qoffset, -1.0f, 1.0f);

> >

> > -                        // 25 is a number that I think it is a possible proper scale value.

> > +                        /* qp range of x264 is from 0 to 51, just choose 25 as the scale

> value,

> > +                         * so the range of final qoffset is [-25.0, 25.0].

> > +                         */

> >                          qoffset = qoffset * 25;

> >

> >                          for (int y = starty; y < endy; y++) {

> >

> 

> I'm not understanding where this number has come from at all.  Is 25 purely


From H264 spec: SliceQPY is in the range of −QpBdOffsetY to +51, inclusive. And lots of documents/papers says the qp range of H264 is [0, 51].
So, I chose 25 (equal with 51/2) roughly to make qpoffset to be [-25.0, 25.0]. 

> arbitrary?  Why does the effect vary by bit depth?


imo, 51 has nothing to do with bit depth, not quite understand your concern.

> 

> From the docs explaining the best/worst behaviour on the range, I would

> expect this to map 0 to 0 and 1 to 3 + 6 * bit_depth (positive or negative).


for example if bit_depth=8, do you mean map -1.0 to -51, and map 1.0 to 51?
for example if bit_depth=10, do you mean map -1.0 to -63, and map 1.0 to 63?

If so, I'm not sure if it is better or not.

Imho, there is still a long way from qpoffset to the final quantization value, I would think a rough qpoffset (with range [-25, 25]) is enough. Anyway, I'm open to use whatever better method.

> 

> - Mark

> _______________________________________________

> ffmpeg-devel mailing list

> ffmpeg-devel@ffmpeg.org

> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Patch hide | download patch | download mbox

diff --git a/libavcodec/libx264.c b/libavcodec/libx264.c
index a3493f3..8c96728 100644
--- a/libavcodec/libx264.c
+++ b/libavcodec/libx264.c
@@ -383,7 +383,9 @@  static int X264_frame(AVCodecContext *ctx, AVPacket *pkt, const AVFrame *frame,
                         qoffset = roi->qoffset.num * 1.0f / roi->qoffset.den;
                         qoffset = av_clipf(qoffset, -1.0f, 1.0f);
 
-                        // 25 is a number that I think it is a possible proper scale value.
+                        /* qp range of x264 is from 0 to 51, just choose 25 as the scale value,
+                         * so the range of final qoffset is [-25.0, 25.0].
+                         */
                         qoffset = qoffset * 25;
 
                         for (int y = starty; y < endy; y++) {