Allow multiple concurrent Decode() requests in VideoDecoder interface.

media/filters/gpu_video_decoder.h

Index: media/filters/gpu_video_decoder.h
diff --git a/media/filters/gpu_video_decoder.h b/media/filters/gpu_video_decoder.h
index 41742e91b42bad90fd878563ade31f016998c00c..f050ffc7e802453f8c5963bf0bebf119bd3bbb2d 100644
--- a/media/filters/gpu_video_decoder.h
+++ b/media/filters/gpu_video_decoder.h
@@ -51,6 +51,7 @@ class MEDIA_EXPORT GpuVideoDecoder
   virtual bool HasAlpha() const OVERRIDE;
   virtual bool NeedsBitstreamConversion() const OVERRIDE;
   virtual bool CanReadWithoutStalling() const OVERRIDE;
+  virtual int GetMaxDecodeRequests() const OVERRIDE;
 
   // VideoDecodeAccelerator::Client implementation.
   virtual void NotifyInitializeDone() OVERRIDE;
@@ -93,16 +94,7 @@ class MEDIA_EXPORT GpuVideoDecoder
 
   typedef std::map<int32, PictureBuffer> PictureBufferMap;
 
-  // Return true if more decode work can be piled on to the VDA.
-  bool CanMoreDecodeWorkBeDone();
-
-  // Enqueue a frame for later delivery (or drop it on the floor if a
-  // vda->Reset() is in progress) and trigger out-of-line delivery of the oldest
-  // ready frame to the client if there is a pending read.  A NULL |frame|
-  // merely triggers delivery, and requires the ready_video_frames_ queue not be
-  // empty.
-  void EnqueueFrameAndTriggerFrameDelivery(
-      const scoped_refptr<VideoFrame>& frame);
+  void DeliverFrame( const scoped_refptr<VideoFrame>& frame);
 
   // Static method is to allow it to run even after GVD is deleted.
   static void ReleaseMailbox(
@@ -142,9 +134,12 @@ class MEDIA_EXPORT GpuVideoDecoder
   // occurs.
   scoped_ptr<VideoDecodeAccelerator> vda_;
 
-  // Callbacks that are !is_null() only during their respective operation being
+  // Callbacks for the pending Decode requests (not more that
+  // GetMaxDecodeRequests()).
+  std::list<DecodeCB> pending_decode_callbacks_;

[xhwang, 2014/04/17 01:06:47]

Usually decoders have decoding delay, e.g. for a decoding delay of 3:

input  0 1 2 3 4 5 6 ...
output       0 1 2 3 ...

Assuming kMaxInFlightDecodes > 3 (see my question to fischman@ below), it seems
we'll always have at least 3 pending decode_cb in pending_decode_callbacks_. The
result of this is that the effective max number of in-flight decodes are dropped
to kMaxInFlightDecodes - 3. This will be bad if decoding is slow and we really
want to have kMaxInFlightDecodes decodes in parallel.

[xhwang, 2014/04/17 01:21:07]

I think this can be fixed if VDA always returns something even if it cannot
produce an output frame. With that, we can return kNotEnoughData (which is
created for this very purpose), to DS so that DS can schedule another Decode
call.

Another possible solution is that in GVD, we always return kNotEnoughData for
the first decoding_delay Decode() calls. But I guess we don't know
decoding_delay in advance.

Another possible solution is that maybe we can keep a permanent decode_cb_ in
VideoDecoder (not a one-off cb passed in Decode() call). Whenever a decoder
generates a decoded frame, it fires decode_cb_ to return the frame. This will
actually work nicely for AudioDecoders where one input can result in multiple
outputs. But this would be a big change and I need to think more about it...

[Sergey Ulanov, 2014/04/23 02:44:21]

I think Ami responded to this over e-mail. Decoder is supposed to return each
frame as soon as it is encoded.

On 2014/04/17 01:06:47, xhwang wrote:

+
+  // Callback that is !is_null() only during their respective operation being
   // asynchronously executed.
-  DecodeCB pending_decode_cb_;
   base::Closure pending_reset_cb_;
 
   State state_;