[V2] lavc/qsvdec: expose frame pic_type and key_frame

Message ID 1517305934-23978-1-git-send-email-zhong.li@intel.com
State New
Headers show
Series
  • [V2] lavc/qsvdec: expose frame pic_type and key_frame
Related show

Commit Message

Li, Zhong Jan. 30, 2018, 9:52 a.m.
Currently key_frame and pict_type are unset.
Add an extra param to fetch the picture type from qsv decoder

Signed-off-by: ChaoX A Liu <chaox.a.liu@intel.com>
Signed-off-by: Zhong Li <zhong.li@intel.com>
---
 libavcodec/qsv.c          | 24 ++++++++++++++++++++++++
 libavcodec/qsv_internal.h |  3 +++
 libavcodec/qsvdec.c       |  7 +++++++
 3 files changed, 34 insertions(+)

Comments

Mark Thompson Jan. 30, 2018, 10:41 p.m. | #1
On 30/01/18 09:52, Zhong Li wrote:
> Currently key_frame and pict_type are unset.
> Add an extra param to fetch the picture type from qsv decoder
> 
> Signed-off-by: ChaoX A Liu <chaox.a.liu@intel.com>
> Signed-off-by: Zhong Li <zhong.li@intel.com>
> ---
>  libavcodec/qsv.c          | 24 ++++++++++++++++++++++++
>  libavcodec/qsv_internal.h |  3 +++
>  libavcodec/qsvdec.c       |  7 +++++++
>  3 files changed, 34 insertions(+)
> 
> diff --git a/libavcodec/qsv.c b/libavcodec/qsv.c
> index e78633d..ffe7e6c 100644
> --- a/libavcodec/qsv.c
> +++ b/libavcodec/qsv.c
> @@ -195,6 +195,30 @@ int ff_qsv_find_surface_idx(QSVFramesContext *ctx, QSVFrame *frame)
>      return AVERROR_BUG;
>  }
>  
> +enum AVPictureType ff_qsv_map_pictype(int mfx_pic_type)
> +{
> +    enum AVPictureType type;
> +    switch (mfx_pic_type & 0x7) {
> +        case MFX_FRAMETYPE_I:
> +            if (mfx_pic_type & MFX_FRAMETYPE_S)
> +                type = AV_PICTURE_TYPE_SI;
> +            else
> +                type = AV_PICTURE_TYPE_I;
> +            break;
> +        case MFX_FRAMETYPE_B:
> +            type = AV_PICTURE_TYPE_B;
> +            break;
> +        case MFX_FRAMETYPE_P:
> +             if (mfx_pic_type & MFX_FRAMETYPE_S)
> +                type = AV_PICTURE_TYPE_SP;
> +            else
> +                type = AV_PICTURE_TYPE_P;
> +            break;
> +    }
> +
> +    return type;
> +}
> +
>  static int qsv_load_plugins(mfxSession session, const char *load_plugins,
>                              void *logctx)
>  {
> diff --git a/libavcodec/qsv_internal.h b/libavcodec/qsv_internal.h
> index 975c8de..07ddc59 100644
> --- a/libavcodec/qsv_internal.h
> +++ b/libavcodec/qsv_internal.h
> @@ -48,6 +48,8 @@ typedef struct QSVMid {
>  typedef struct QSVFrame {
>      AVFrame *frame;
>      mfxFrameSurface1 surface;
> +    mfxExtDecodedFrameInfo dec_info;
> +    mfxExtBuffer *ext_param;
>  
>      int queued;
>      int used;
> @@ -83,6 +85,7 @@ int ff_qsv_codec_id_to_mfx(enum AVCodecID codec_id);
>  int ff_qsv_profile_to_mfx(enum AVCodecID codec_id, int profile);
>  
>  int ff_qsv_map_pixfmt(enum AVPixelFormat format, uint32_t *fourcc);
> +enum AVPictureType ff_qsv_map_pictype(int mfx_pic_type);
>  
>  int ff_qsv_init_internal_session(AVCodecContext *avctx, mfxSession *session,
>                                   const char *load_plugins);
> diff --git a/libavcodec/qsvdec.c b/libavcodec/qsvdec.c
> index f31172d..9ee4177 100644
> --- a/libavcodec/qsvdec.c
> +++ b/libavcodec/qsvdec.c
> @@ -232,6 +232,11 @@ static int alloc_frame(AVCodecContext *avctx, QSVContext *q, QSVFrame *frame)
>  
>          frame->surface.Data.MemId = &q->frames_ctx.mids[ret];
>      }
> +    frame->surface.Data.ExtParam = &frame->ext_param;
> +    frame->surface.Data.NumExtParam = 1;
> +    frame->ext_param = (mfxExtBuffer*)&frame->dec_info;
> +    frame->dec_info.Header.BufferId = MFX_EXTBUFF_DECODED_FRAME_INFO;
> +    frame->dec_info.Header.BufferSz = sizeof(frame->dec_info);
>  
>      frame->used = 1;
>  
> @@ -416,6 +421,8 @@ FF_ENABLE_DEPRECATION_WARNINGS
>              outsurf->Info.PicStruct & MFX_PICSTRUCT_FIELD_TFF;
>          frame->interlaced_frame =
>              !(outsurf->Info.PicStruct & MFX_PICSTRUCT_PROGRESSIVE);
> +        frame->pict_type = ff_qsv_map_pictype(out_frame->dec_info.FrameType);
> +        frame->key_frame = !!(out_frame->dec_info.FrameType & MFX_FRAMETYPE_IDR);
>  
>          /* update the surface properties */
>          if (avctx->pix_fmt == AV_PIX_FMT_QSV)
> 

Does this do the right thing for H.265?  (Can't test right now.)

I'm asking that because I know that frame types are still a problem for H.265 encoding, because not all IRAPs are IDR - when CRA frames are generated they aren't marked as key.  The FrameType field (as used here) is not sufficient in that case, but I never sorted out exactly what does contain the necessary information.

- Mark
Li, Zhong Jan. 31, 2018, 8:11 a.m. | #2
> > @@ -416,6 +421,8 @@ FF_ENABLE_DEPRECATION_WARNINGS

> >              outsurf->Info.PicStruct & MFX_PICSTRUCT_FIELD_TFF;

> >          frame->interlaced_frame =

> >              !(outsurf->Info.PicStruct &

> MFX_PICSTRUCT_PROGRESSIVE);

> > +        frame->pict_type =

> ff_qsv_map_pictype(out_frame->dec_info.FrameType);

> > +        frame->key_frame = !!(out_frame->dec_info.FrameType &

> > + MFX_FRAMETYPE_IDR);

> >

> >          /* update the surface properties */

> >          if (avctx->pix_fmt == AV_PIX_FMT_QSV)

> >

>

> Does this do the right thing for H.265?  (Can't test right now.)



Testing should use a HEVC clips



> I'm asking that because I know that frame types are still a problem for H.265

> encoding, because not all IRAPs are IDR - when CRA frames are generated

> they aren't marked as key.  The FrameType field (as used here) is not

> sufficient in that case, but I never sorted out exactly what does contain the

> necessary information.

>

> - Mark



Exactly, only IDR type is exposed from FrameType here, and it is set by parsing slice_header (https://github.com/Intel-Media-SDK/MediaSDK/blob/master/_studio/shared/umc/codec/h265_dec/src/umc_h265_bitstream_headers.cpp#L1428 )

But the problem is (I am not familiar with HEVC as well as H264, so please correct me if I am wrong), should CRA frames be treated as key frames?

See the description as blew (source http://ieeexplore.ieee.org/document/6324417/?part=undefined%7Cfig4#fig4 )

[cid:image003.jpg@01D39AAE.23C1D950]



Here some leading pictures such as B30 can’t be decoded when random access from I28.

As my understanding, all frames behand (in display order) key frame should be decodable.

IDR it can make sure this, but CRA can’t. So at least not all of the CRA frames can be treated as key frames.

Looks some people agree with me (refer Jeanlf’s comment on https://github.com/gpac/gpac/issues/144)
Li, Zhong Jan. 31, 2018, 8:16 a.m. | #3
Sorry it is interrupted, I mean could you specify which hevc clip should be tested?

As your comment, maybe we need to find a HEVC clip with CRA frames to test.

From: Li, Zhong
Sent: Wednesday, January 31, 2018 4:11 PM
To: libav-devel@libav.org
Subject: RE: [libav-devel] [PATCH V2] lavc/qsvdec: expose frame pic_type and key_frame


> > @@ -416,6 +421,8 @@ FF_ENABLE_DEPRECATION_WARNINGS

> >              outsurf->Info.PicStruct & MFX_PICSTRUCT_FIELD_TFF;

> >          frame->interlaced_frame =

> >              !(outsurf->Info.PicStruct &

> MFX_PICSTRUCT_PROGRESSIVE);

> > +        frame->pict_type =

> ff_qsv_map_pictype(out_frame->dec_info.FrameType);

> > +        frame->key_frame = !!(out_frame->dec_info.FrameType &

> > + MFX_FRAMETYPE_IDR);

> >

> >          /* update the surface properties */

> >          if (avctx->pix_fmt == AV_PIX_FMT_QSV)

> >

>

> Does this do the right thing for H.265?  (Can't test right now.)



Testing should use a HEVC clips
Mark Thompson Jan. 31, 2018, 9:40 a.m. | #4
On 31/01/18 08:11, Li, Zhong wrote:
>>> @@ -416,6 +421,8 @@ FF_ENABLE_DEPRECATION_WARNINGS
> 
>>>              outsurf->Info.PicStruct & MFX_PICSTRUCT_FIELD_TFF;
> 
>>>          frame->interlaced_frame =
> 
>>>              !(outsurf->Info.PicStruct &
> 
>> MFX_PICSTRUCT_PROGRESSIVE);
> 
>>> +        frame->pict_type =
> 
>> ff_qsv_map_pictype(out_frame->dec_info.FrameType);
> 
>>> +        frame->key_frame = !!(out_frame->dec_info.FrameType &
> 
>>> + MFX_FRAMETYPE_IDR);
> 
>>>
> 
>>>          /* update the surface properties */
> 
>>>          if (avctx->pix_fmt == AV_PIX_FMT_QSV)
> 
>>>
> 
>>
> 
>> Does this do the right thing for H.265?  (Can't test right now.)
> 
> 
> 
> Testing should use a HEVC clips
> 
> 
> 
>> I'm asking that because I know that frame types are still a problem for H.265
> 
>> encoding, because not all IRAPs are IDR - when CRA frames are generated
> 
>> they aren't marked as key.  The FrameType field (as used here) is not
> 
>> sufficient in that case, but I never sorted out exactly what does contain the
> 
>> necessary information.
> 
>>
> 
>> - Mark
> 
> 
> 
> Exactly, only IDR type is exposed from FrameType here, and it is set by parsing slice_header (https://github.com/Intel-Media-SDK/MediaSDK/blob/master/_studio/shared/umc/codec/h265_dec/src/umc_h265_bitstream_headers.cpp#L1428 )
> 
> But the problem is (I am not familiar with HEVC as well as H264, so please correct me if I am wrong), should CRA frames be treated as key frames?

Yes, they have to be.  IDR, CRA and BLA frames are all IRAP pictures, so they are key frames because seeking can always move to that point in the stream.

> See the description as blew (source http://ieeexplore.ieee.org/document/6324417/?part=undefined%7Cfig4#fig4 )
> 
> [cid:image003.jpg@01D39AAE.23C1D950]
> 
> 
> 
> Here some leading pictures such as B30 can’t be decoded when random access from I28.
> 
> As my understanding, all frames behand (in display order) key frame should be decodable.

Yes, all frames after the IRAP in display order should be decodable.  However, B30 is not after I28 in display order - it be marked as RASL to indicate that it should be skipped when it is found in decode order.

> IDR it can make sure this, but CRA can’t. So at least not all of the CRA frames can be treated as key frames.
> 
> Looks some people agree with me (refer Jeanlf’s comment on https://github.com/gpac/gpac/issues/144)
I'm not familiar with the precise meaning of "sync sample" in this case, so I'm not sure.  If it means the same thing as IRAP (a point at which decoding of a stream can always start correctly without any pictures which precede it in decode order), then the commenter is wrong.


On 31/01/18 08:16, Li, Zhong wrote:
> As your comment, maybe we need to find a HEVC clip with CRA frames to test.

Try the conformance sample RAP_A_docomo_4.bit (you can find it in fate/hevc-conformance).  That contains CRA and RASL frames (and no IDR frames at all).

- Mark
Li, Zhong Feb. 1, 2018, 7:48 a.m. | #5
> On 31/01/18 08:16, Li, Zhong wrote:
> > As your comment, maybe we need to find a HEVC clip with CRA frames to
> test.
> 
> Try the conformance sample RAP_A_docomo_4.bit (you can find it in
> fate/hevc-conformance).  That contains CRA and RASL frames (and no IDR
> frames at all).
> 
> - Mark

Tested it. Yup, it is an issue just like your inference, the CRA frames aren't set as key frame for QSV. But software codec set as key frames.
I haven't find where MSDK expose this, maybe need file an issue to MSDK.

Patch

diff --git a/libavcodec/qsv.c b/libavcodec/qsv.c
index e78633d..ffe7e6c 100644
--- a/libavcodec/qsv.c
+++ b/libavcodec/qsv.c
@@ -195,6 +195,30 @@  int ff_qsv_find_surface_idx(QSVFramesContext *ctx, QSVFrame *frame)
     return AVERROR_BUG;
 }
 
+enum AVPictureType ff_qsv_map_pictype(int mfx_pic_type)
+{
+    enum AVPictureType type;
+    switch (mfx_pic_type & 0x7) {
+        case MFX_FRAMETYPE_I:
+            if (mfx_pic_type & MFX_FRAMETYPE_S)
+                type = AV_PICTURE_TYPE_SI;
+            else
+                type = AV_PICTURE_TYPE_I;
+            break;
+        case MFX_FRAMETYPE_B:
+            type = AV_PICTURE_TYPE_B;
+            break;
+        case MFX_FRAMETYPE_P:
+             if (mfx_pic_type & MFX_FRAMETYPE_S)
+                type = AV_PICTURE_TYPE_SP;
+            else
+                type = AV_PICTURE_TYPE_P;
+            break;
+    }
+
+    return type;
+}
+
 static int qsv_load_plugins(mfxSession session, const char *load_plugins,
                             void *logctx)
 {
diff --git a/libavcodec/qsv_internal.h b/libavcodec/qsv_internal.h
index 975c8de..07ddc59 100644
--- a/libavcodec/qsv_internal.h
+++ b/libavcodec/qsv_internal.h
@@ -48,6 +48,8 @@  typedef struct QSVMid {
 typedef struct QSVFrame {
     AVFrame *frame;
     mfxFrameSurface1 surface;
+    mfxExtDecodedFrameInfo dec_info;
+    mfxExtBuffer *ext_param;
 
     int queued;
     int used;
@@ -83,6 +85,7 @@  int ff_qsv_codec_id_to_mfx(enum AVCodecID codec_id);
 int ff_qsv_profile_to_mfx(enum AVCodecID codec_id, int profile);
 
 int ff_qsv_map_pixfmt(enum AVPixelFormat format, uint32_t *fourcc);
+enum AVPictureType ff_qsv_map_pictype(int mfx_pic_type);
 
 int ff_qsv_init_internal_session(AVCodecContext *avctx, mfxSession *session,
                                  const char *load_plugins);
diff --git a/libavcodec/qsvdec.c b/libavcodec/qsvdec.c
index f31172d..9ee4177 100644
--- a/libavcodec/qsvdec.c
+++ b/libavcodec/qsvdec.c
@@ -232,6 +232,11 @@  static int alloc_frame(AVCodecContext *avctx, QSVContext *q, QSVFrame *frame)
 
         frame->surface.Data.MemId = &q->frames_ctx.mids[ret];
     }
+    frame->surface.Data.ExtParam = &frame->ext_param;
+    frame->surface.Data.NumExtParam = 1;
+    frame->ext_param = (mfxExtBuffer*)&frame->dec_info;
+    frame->dec_info.Header.BufferId = MFX_EXTBUFF_DECODED_FRAME_INFO;
+    frame->dec_info.Header.BufferSz = sizeof(frame->dec_info);
 
     frame->used = 1;
 
@@ -416,6 +421,8 @@  FF_ENABLE_DEPRECATION_WARNINGS
             outsurf->Info.PicStruct & MFX_PICSTRUCT_FIELD_TFF;
         frame->interlaced_frame =
             !(outsurf->Info.PicStruct & MFX_PICSTRUCT_PROGRESSIVE);
+        frame->pict_type = ff_qsv_map_pictype(out_frame->dec_info.FrameType);
+        frame->key_frame = !!(out_frame->dec_info.FrameType & MFX_FRAMETYPE_IDR);
 
         /* update the surface properties */
         if (avctx->pix_fmt == AV_PIX_FMT_QSV)