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..5cc4d6b7c43cc2e146b9710894f029a8f899f533 100644
--- a/webrtc/media/base/videocapturer.h
+++ b/webrtc/media/base/videocapturer.h
@@ -225,15 +225,30 @@ 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.
+
+  // This function also implements timestamp translation/filtering.
+  // |camera_time_ns| is the camera's timestamp for the captured
+  // frame; it is expected to have good accuracy, but it may use an
+  // arbitrary epoch and a small possibly free-running with a frequency
+  // slightly different from the system clock. |system_time_us| is the
+  // monotonic system time (in the same scale as rtc::TimeMicros) when
+  // the frame was captured; the application is expected to read the
+  // system time as soon as possible after frame capture, but it may
+  // suffer scheduling jitter or poor system clock resolution. The
+  // output |translated_camera_time_us| is a combined timestamp,
+  // taking advantage of the supposedly higher accuracy in the camera
+  // timestamp, but using the same epoch and frequency as system time.
   bool AdaptFrame(int width,
                   int height,
-                  int64_t capture_time_ns,
+                  int64_t camera_time_us,
+                  int64_t system_time_us,
                   int* out_width,
                   int* out_height,
                   int* crop_width,
                   int* crop_height,
                   int* crop_x,
-                  int* crop_y);
+                  int* crop_y,
+                  int64_t* translated_camera_time_us);
 
   // Callback attached to SignalFrameCaptured where SignalVideoFrames is called.
   void OnFrameCaptured(VideoCapturer* video_capturer,
@@ -265,6 +280,10 @@ class VideoCapturer : public sigslot::has_slots<>,
   void SetSupportedFormats(const std::vector<VideoFormat>& formats);
   VideoFrameFactory* frame_factory() { return frame_factory_.get(); }
 
+ protected:
+  // Access intended only for testing.

[sprang_webrtc, 2016/06/17 13:10:52]

Can't we use friend class to expose this instead, or do we really hate those?

[nisse-webrtc, 2016/06/17 13:43:01]

I have no strong opinion on that (and not much experience with the C++ "friend"
feature). But to me, a protected non-virtual inline method seems very cheap in
all respects. A friend declaration referencing some class in the testsuite (if
that's what you are suggesting?) seems kind-of ugly.

[sprang_webrtc, 2016/06/17 14:35:38]

Friend class isn't pretty, but exposing and commenting that it's only for test
is no fun either imo. But I'm not religious about it either, so do as you like

+  int64_t clip_bias_us() { return clip_bias_us_; }
+
  private:
   void Construct();
   // Get the distance between the desired format and the supported format.
@@ -284,6 +303,11 @@ class VideoCapturer : public sigslot::has_slots<>,
 
   void UpdateInputSize(int width, int height);
 
+  // Update the estimated offset between camera time and system monotonic time.
+  void UpdateOffset(int64_t camera_time_us, int64_t system_time_us);
+
+  int64_t ClipTimestamp(int64_t filtered_time_us, int64_t system_time_us);
+
   rtc::ThreadChecker thread_checker_;
   std::string id_;
   CaptureState capture_state_;
@@ -310,6 +334,16 @@ 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.
+  int frames_seen_;
+  // Estimated offset between camera time and system monotonic time.
+  int64_t offset_us_;
+
+  // State for timestamp clipping, applied after the filter, to ensure
+  // that translated timestamps are monotonic and not in the future.
+  // Subtracted from the translated timestamps.
+  int64_t clip_bias_us_;
+  rtc::Optional<int64_t> prev_translated_time_us_;
   RTC_DISALLOW_COPY_AND_ASSIGN(VideoCapturer);
 };