Correction of the AEC3 underrun behavior and minor other corrections

webrtc/modules/audio_processing/aec3/render_delay_buffer.cc

Index: webrtc/modules/audio_processing/aec3/render_delay_buffer.cc
diff --git a/webrtc/modules/audio_processing/aec3/render_delay_buffer.cc b/webrtc/modules/audio_processing/aec3/render_delay_buffer.cc
index a816b4c7eb96866729ec47b264c0696f511c1c83..aa72e3bb6cbedb6d3cd5200adfa4e3a5f8989dd8 100644
--- a/webrtc/modules/audio_processing/aec3/render_delay_buffer.cc
+++ b/webrtc/modules/audio_processing/aec3/render_delay_buffer.cc
@@ -126,6 +126,11 @@ void RenderDelayBufferImpl::Reset() {
   downsampled_render_buffer_.buffer.fill(0.f);
   fft_buffer_.Clear();
   api_call_jitter_buffer_.Reset();
+  for (auto& c : buffer_) {

[peah-webrtc, 2017/04/06 13:39:08]

After a buffer delay reset (due to a too long buffer delay for instance
occurring during clockdrift), the missing zeroing of the buffer caused the
filter to diverge initially. Setting the buffer to zero solves that.

+    for (auto& b : c) {
+      std::fill(b.begin(), b.end(), 0.f);
+    }
+  }
 }
 
 bool RenderDelayBufferImpl::Insert(std::vector<std::vector<float>>* block) {