Switch AudioRendererAlgorithm to buffer based on time.

It doesn't make sense to buffer based on a frame count when the
amount of time that data represent varies with sample rate; it's
the opposite of what we want.  I.e., a 48kHz playback will exhaust
this buffer faster than a 44.1kHz playback.

Instead, switch to time based initial capacity using the sample
rate provided during initialization.  This provides a greater
resilience against underflow when video decoding is overloading
the media thread.

At 48kHz we were previously buffering ~171ms and at 44.1kHz we were
buffering ~186ms of audio data.  I've chosen a nice round 200ms for
simplicity and since it works well on an Android One device.

BUG=519108
TEST=media_unittests still pass.

Review URL: https://codereview.chromium.org/1278733004

Cr-Commit-Position: refs/heads/master@{#343279}

3 files changed, 19 insertions, 14 deletions

diff --git a/media/filters/audio_renderer_algorithm.cc b/media/filters/audio_renderer_algorithm.cc
index 2d2bfbf..a370124 100644
--- a/media/filters/audio_renderer_algorithm.cc
+++ b/media/filters/audio_renderer_algorithm.cc
@@ -61,20 +61,14 @@ static const int kWsolaSearchIntervalMs = 30;
 // The maximum size in seconds for the |audio_buffer_|. Arbitrarily determined.
 static const int kMaxCapacityInSeconds = 3;
 
-// The starting size in frames for |audio_buffer_|. Previous usage maintained a
-// queue of 16 AudioBuffers, each of 512 frames. This worked well, so we
-// maintain this number of frames.
-static const int kStartingBufferSizeInFrames = 16 * 512;
-
-static_assert(kStartingBufferSizeInFrames <
-              (kMaxCapacityInSeconds * limits::kMinSampleRate),
-              "max capacity smaller than starting buffer size");
+// The minimum size in ms for the |audio_buffer_|. Arbitrarily determined.
+static const int kStartingCapacityInMs = 200;
 
 AudioRendererAlgorithm::AudioRendererAlgorithm()
     : channels_(0),
       samples_per_second_(0),
       muted_partial_frame_(0),
-      capacity_(kStartingBufferSizeInFrames),
+      capacity_(0),
       output_time_(0.0),
       search_block_center_offset_(0),
       search_block_index_(0),
@@ -92,8 +86,9 @@ void AudioRendererAlgorithm::Initialize(const AudioParameters& params) {
 
   channels_ = params.channels();
   samples_per_second_ = params.sample_rate();
-  num_candidate_blocks_ = (kWsolaSearchIntervalMs * samples_per_second_) / 1000;
-  ola_window_size_ = kOlaWindowSizeMs * samples_per_second_ / 1000;
+  capacity_ = ConvertMillisecondsToFrames(kStartingCapacityInMs);
+  num_candidate_blocks_ = ConvertMillisecondsToFrames(kWsolaSearchIntervalMs);
+  ola_window_size_ = ConvertMillisecondsToFrames(kOlaWindowSizeMs);
 
   // Make sure window size in an even number.
   ola_window_size_ += ola_window_size_ & 1;
@@ -206,9 +201,9 @@ void AudioRendererAlgorithm::FlushBuffers() {
   wsola_output_->Zero();
   num_complete_frames_ = 0;
 
-  // Reset |capacity_| so growth triggered by underflows doesn't penalize
-  // seek time.
-  capacity_ = kStartingBufferSizeInFrames;
+  // Reset |capacity_| so growth triggered by underflows doesn't penalize seek
+  // time.
+  capacity_ = ConvertMillisecondsToFrames(kStartingCapacityInMs);
 }
 
 void AudioRendererAlgorithm::EnqueueBuffer(
@@ -235,6 +230,11 @@ bool AudioRendererAlgorithm::CanPerformWsola() const {
       search_block_index_ + search_block_size <= frames;
 }
 
+int AudioRendererAlgorithm::ConvertMillisecondsToFrames(int ms) const {
+  return ms * (samples_per_second_ /
+               static_cast<double>(base::Time::kMillisecondsPerSecond));
+}
+
 bool AudioRendererAlgorithm::RunOneWsolaIteration(double playback_rate) {
   if (!CanPerformWsola())
     return false;
diff --git a/media/filters/audio_renderer_algorithm.h b/media/filters/audio_renderer_algorithm.h
index 2005bfe..940e7e7 100644
--- a/media/filters/audio_renderer_algorithm.h
+++ b/media/filters/audio_renderer_algorithm.h
@@ -118,6 +118,9 @@ class MEDIA_EXPORT AudioRendererAlgorithm {
   // Do we have enough data to perform one round of WSOLA?
   bool CanPerformWsola() const;
 
+  // Converts a time in milliseconds to frames using |samples_per_second_|.
+  int ConvertMillisecondsToFrames(int ms) const;
+
   // Number of channels in audio stream.
   int channels_;
 
diff --git a/media/filters/audio_renderer_algorithm_unittest.cc b/media/filters/audio_renderer_algorithm_unittest.cc
index 003cd51..82c59675 100644
--- a/media/filters/audio_renderer_algorithm_unittest.cc
+++ b/media/filters/audio_renderer_algorithm_unittest.cc
@@ -666,6 +666,7 @@ TEST_F(AudioRendererAlgorithmTest, FillBufferOffset) {
     ASSERT_EQ(kHalfSize, frames_filled);
     ASSERT_TRUE(VerifyAudioData(bus.get(), 0, kHalfSize, 0));
     ASSERT_FALSE(VerifyAudioData(bus.get(), kHalfSize, kHalfSize, 0));
+    FillAlgorithmQueue();
   }
 
   const float kMutedRates[] = {5.0f, 0.25f};
@@ -679,6 +680,7 @@ TEST_F(AudioRendererAlgorithmTest, FillBufferOffset) {
     ASSERT_EQ(kHalfSize, frames_filled);
     ASSERT_FALSE(VerifyAudioData(bus.get(), 0, kHalfSize, 0));
     ASSERT_TRUE(VerifyAudioData(bus.get(), kHalfSize, kHalfSize, 0));
+    FillAlgorithmQueue();
   }
 }