Implement timestamp translation/filter in VideoCapturer.

webrtc/media/base/videocapturer.h

Index: webrtc/media/base/videocapturer.h
diff --git a/webrtc/media/base/videocapturer.h b/webrtc/media/base/videocapturer.h
index 329ba086b90dfe1139c6f83cdd41b291af38540f..08e07e3b35cbca3e495759d5f7e180eaa9b7a281 100644
--- a/webrtc/media/base/videocapturer.h
+++ b/webrtc/media/base/videocapturer.h
@@ -225,6 +225,9 @@ class VideoCapturer : public sigslot::has_slots<>,
   // Reports the appropriate frame size after adaptation. Returns true
   // if a frame is wanted. Returns false if there are no interested
   // sinks, or if the VideoAdapter decides to drop the frame.
+  // |capture_time_ns| is the camera's timestamp, using an arbitrary
+  // offset and possibly slightly off frequency. |time_us| is the
+  // timestamp translated into the same scale as rtc::TimeMicros.

[pthatcher1, 2016/05/27 22:47:44]

Then why not call it capture_time_us?  

And why require passing both in?  Shouldn't passing one in be sufficient?

[nisse-webrtc, 2016/05/30 08:57:29]

The nanosecond unit on the input argument, |capture_time_ns|, is kept for now
only to make the change a little smaller; it would be more consistent to use
microsecond unit here too.

The output argument is that time translated into rtc::TimeMicros timescale.

To eliminate the call to rtc::TimeMicros from the VideoCapturer class (using
global state complicates testing), I'm considering adding a second input
argument, which is the current rtc::TimeMicros, or, preferably, the value as
rtc::TimeMicros read as soon as possibly when the application gets the frame.

To recap, the model here is that the capture timestamp has pretty good accuracy,
in its own unknown timescale, while reading rtc::Time may exhibit jitter of
several ms due to scheduling or poor accuracy in the OS clock.

   bool AdaptFrame(int width,
                   int height,
                   int64_t capture_time_ns,
@@ -233,7 +236,8 @@ class VideoCapturer : public sigslot::has_slots<>,
                   int* crop_width,
                   int* crop_height,
                   int* crop_x,
-                  int* crop_y);
+                  int* crop_y,
+                  int64_t* time_us);
 
   // Callback attached to SignalFrameCaptured where SignalVideoFrames is called.
   void OnFrameCaptured(VideoCapturer* video_capturer,
@@ -284,6 +288,8 @@ class VideoCapturer : public sigslot::has_slots<>,
 
   void UpdateInputSize(int width, int height);
 
+  void UpdateOffset(int64_t capture_time_us);

[pthatcher1, 2016/05/27 22:47:44]

This method could use a comment explaining what it does.  It's not obvious
looking at it.

[nisse-webrtc, 2016/05/30 08:57:29]

Since it's a private method, I put the explanatory comment in the implementation
file. I'm adding a short comment in the header too.

+
   rtc::ThreadChecker thread_checker_;
   std::string id_;
   CaptureState capture_state_;
@@ -310,6 +316,10 @@ class VideoCapturer : public sigslot::has_slots<>,
   // Whether capturer should apply rotation to the frame before signaling it.
   bool apply_rotation_;
 
+  // State for the timestamp translation.
+  unsigned frames_seen_;
+  int64_t offset_us_;

[pthatcher1, 2016/05/27 22:47:44]

This could use a little more clarity on what it means.

[nisse-webrtc, 2016/05/30 08:57:29]

I'm adding a short comment.

+
   RTC_DISALLOW_COPY_AND_ASSIGN(VideoCapturer);
 };