Enhance AudioSplicer to crossfade marked splice frames.

media/base/audio_splicer.cc

Index: media/base/audio_splicer.cc
diff --git a/media/base/audio_splicer.cc b/media/base/audio_splicer.cc
index 14b4199e0e3389d8d32478fae511333915cfcf12..aa5bad9b16605032f0530e05880c205588b6889b 100644
--- a/media/base/audio_splicer.cc
+++ b/media/base/audio_splicer.cc
@@ -5,12 +5,15 @@
 #include "media/base/audio_splicer.h"
 
 #include <cstdlib>
+#include <deque>
 
 #include "base/logging.h"
 #include "media/base/audio_buffer.h"
+#include "media/base/audio_bus.h"
 #include "media/base/audio_decoder_config.h"
 #include "media/base/audio_timestamp_helper.h"
 #include "media/base/buffers.h"
+#include "media/base/vector_math.h"
 
 namespace media {
 
@@ -20,22 +23,104 @@ namespace media {
 // roughly represents the duration of 2 compressed AAC or MP3 frames.
 static const int kMaxTimeDeltaInMilliseconds = 50;
 
-AudioSplicer::AudioSplicer(int samples_per_second)
-    : output_timestamp_helper_(samples_per_second),
-      min_gap_size_(2),
-      received_end_of_stream_(false) {
+// Minimum gap size needed before the splicer will take action to
+// fill a gap. This avoids periodically inserting and then dropping samples
+// when the buffer timestamps are slightly off because of timestamp rounding
+// in the source content. Unit is frames.
+static const int kMinGapSize = 2;
+
+// The number of milliseconds to crossfade before trimming when buffers overlap.
+static const int kCrossfadeDurationInMilliseconds = 5;
+
+// AudioBuffer::TrimStart() is not as accurate as the timestamp helper, so
+// manually adjust the duration and timestamp after trimming.
+static void AccurateTrimStart(int frames_to_trim,
+                              const scoped_refptr<AudioBuffer> buffer,
+                              const AudioTimestampHelper& timestamp_helper) {
+  buffer->TrimStart(frames_to_trim);
+  buffer->set_timestamp(timestamp_helper.GetTimestamp());
+  buffer->set_duration(
+      timestamp_helper.GetFrameDuration(buffer->frame_count()));
 }
 
-AudioSplicer::~AudioSplicer() {
+// AudioBuffer::TrimEnd() is not as accurate as the timestamp helper, so
+// manually adjust the duration after trimming.
+static void AccurateTrimEnd(int frames_to_trim,
+                            const scoped_refptr<AudioBuffer> buffer,
+                            const AudioTimestampHelper& timestamp_helper) {
+  DCHECK(buffer->timestamp() == timestamp_helper.GetTimestamp());
+  buffer->TrimEnd(frames_to_trim);
+  buffer->set_duration(
+      timestamp_helper.GetFrameDuration(buffer->frame_count()));
 }
 
-void AudioSplicer::Reset() {
-  output_timestamp_helper_.SetBaseTimestamp(kNoTimestamp());
+class AudioStreamSanitizer {
+ public:
+  explicit AudioStreamSanitizer(int samples_per_second);
+  ~AudioStreamSanitizer();
+
+  // Resets the sanitizer state by clearing the output buffers queue, and
+  // resetting the timestamp helper.
+  void Reset();
+
+  // Similar to Reset(), but initializes the timestamp helper with the given
+  // parameters.
+  void ResetTimestampState(int64 frame_count, base::TimeDelta base_timestamp);
+
+  // Adds a new buffer full of samples or end of stream buffer to the splicer.
+  // Returns true if the buffer was accepted. False is returned if an error
+  // occurred.
+  bool AddInput(const scoped_refptr<AudioBuffer>& input);
+
+  // Returns true if the sanitizer has a buffer to return.
+  bool HasNextBuffer() const;
+
+  // Removes the next buffer from the output buffer queue and returns it; should
+  // only be called if HasNextBuffer() returns true.
+  scoped_refptr<AudioBuffer> GetNextBuffer();
+
+  // Returns the total frame count of all buffers available for output.
+  int GetFrameCount() const;
+
+  // Returns the duration of all buffers added to the output queue thus far.
+  base::TimeDelta GetDuration() const;
+
+  const AudioTimestampHelper& timestamp_helper() {
+    return output_timestamp_helper_;
+  }
+
+ private:
+  void AddOutputBuffer(const scoped_refptr<AudioBuffer>& buffer);
+
+  AudioTimestampHelper output_timestamp_helper_;
+  bool received_end_of_stream_;
+
+  typedef std::deque<scoped_refptr<AudioBuffer> > BufferQueue;
+  BufferQueue output_buffers_;
+
+  DISALLOW_ASSIGN(AudioStreamSanitizer);
+};
+
+AudioStreamSanitizer::AudioStreamSanitizer(int samples_per_second)
+    : output_timestamp_helper_(samples_per_second),
+      received_end_of_stream_(false) {}
+
+AudioStreamSanitizer::~AudioStreamSanitizer() {}
+
+void AudioStreamSanitizer::Reset() {
+  ResetTimestampState(0, kNoTimestamp());
+}
+
+void AudioStreamSanitizer::ResetTimestampState(int64 frame_count,
+                                               base::TimeDelta base_timestamp) {
   output_buffers_.clear();
   received_end_of_stream_ = false;
+  output_timestamp_helper_.SetBaseTimestamp(base_timestamp);
+  if (frame_count > 0)
+    output_timestamp_helper_.AddFrames(frame_count);
 }
 
-bool AudioSplicer::AddInput(const scoped_refptr<AudioBuffer>& input) {
+bool AudioStreamSanitizer::AddInput(const scoped_refptr<AudioBuffer>& input) {
   DCHECK(!received_end_of_stream_ || input->end_of_stream());
 
   if (input->end_of_stream()) {
@@ -56,9 +141,10 @@ bool AudioSplicer::AddInput(const scoped_refptr<AudioBuffer>& input) {
     return false;
   }
 
-  base::TimeDelta timestamp = input->timestamp();
-  base::TimeDelta expected_timestamp = output_timestamp_helper_.GetTimestamp();
-  base::TimeDelta delta = timestamp - expected_timestamp;
+  const base::TimeDelta timestamp = input->timestamp();
+  const base::TimeDelta expected_timestamp =
+      output_timestamp_helper_.GetTimestamp();
+  const base::TimeDelta delta = timestamp - expected_timestamp;
 
   if (std::abs(delta.InMilliseconds()) > kMaxTimeDeltaInMilliseconds) {
     DVLOG(1) << "Timestamp delta too large: " << delta.InMicroseconds() << "us";
@@ -69,7 +155,7 @@ bool AudioSplicer::AddInput(const scoped_refptr<AudioBuffer>& input) {
   if (delta != base::TimeDelta())
     frames_to_fill = output_timestamp_helper_.GetFramesToTarget(timestamp);
 
-  if (frames_to_fill == 0 || std::abs(frames_to_fill) < min_gap_size_) {
+  if (frames_to_fill == 0 || std::abs(frames_to_fill) < kMinGapSize) {
     AddOutputBuffer(input);
     return true;
   }
@@ -92,11 +178,16 @@ bool AudioSplicer::AddInput(const scoped_refptr<AudioBuffer>& input) {
     return true;
   }
 
-  int frames_to_skip = -frames_to_fill;
-
+  // Overlapping buffers marked as splice frames are handled by AudioSplicer,
+  // but decoder and demuxer quirks may sometimes produce overlapping samples
+  // which need to be sanitized.
+  //
+  // A crossfade can't be done here because only the current buffer is available
+  // at this point, not previous buffers.
   DVLOG(1) << "Overlap detected @ " << expected_timestamp.InMicroseconds()
-           << " us: "  << -delta.InMicroseconds() << " us";
+           << " us: " << -delta.InMicroseconds() << " us";
 
+  const int frames_to_skip = -frames_to_fill;
   if (input->frame_count() <= frames_to_skip) {
     DVLOG(1) << "Dropping whole buffer";
     return true;
@@ -104,27 +195,277 @@ bool AudioSplicer::AddInput(const scoped_refptr<AudioBuffer>& input) {
 
   // Copy the trailing samples that do not overlap samples already output
   // into a new buffer. Add this new buffer to the output queue.
-  //
-  // TODO(acolwell): Implement a cross-fade here so the transition is less

[acolwell GONE FROM CHROMIUM, 2014/02/28 18:50:27]

nit: I think this comment should stay. For the "badly encoded stream" case it
would be nice to have a simple crossfade to make the transition a little less
jarring especially if there is a step function type change in loudness.

[DaleCurtis, 2014/02/28 21:14:26]

I'll leave it, but it's kind of impossible to implement given how the splicer is
used, so it doesn't seem like an effective comment.

-  // jarring.
-  input->TrimStart(frames_to_skip);
+  AccurateTrimStart(frames_to_skip, input, output_timestamp_helper_);
   AddOutputBuffer(input);
   return true;
 }
 
-bool AudioSplicer::HasNextBuffer() const {
+bool AudioStreamSanitizer::HasNextBuffer() const {
   return !output_buffers_.empty();
 }
 
-scoped_refptr<AudioBuffer> AudioSplicer::GetNextBuffer() {
+scoped_refptr<AudioBuffer> AudioStreamSanitizer::GetNextBuffer() {
   scoped_refptr<AudioBuffer> ret = output_buffers_.front();
   output_buffers_.pop_front();
   return ret;
 }
 
-void AudioSplicer::AddOutputBuffer(const scoped_refptr<AudioBuffer>& buffer) {
+void AudioStreamSanitizer::AddOutputBuffer(
+    const scoped_refptr<AudioBuffer>& buffer) {
   output_timestamp_helper_.AddFrames(buffer->frame_count());
   output_buffers_.push_back(buffer);
 }
 
+int AudioStreamSanitizer::GetFrameCount() const {
+  int frame_count = 0;
+  for (BufferQueue::const_iterator it = output_buffers_.begin();
+       it != output_buffers_.end(); ++it) {
+    frame_count += (*it)->frame_count();
+  }
+  return frame_count;
+}
+
+base::TimeDelta AudioStreamSanitizer::GetDuration() const {
+  DCHECK(output_timestamp_helper_.base_timestamp() != kNoTimestamp());
+  return output_timestamp_helper_.GetTimestamp() -
+         output_timestamp_helper_.base_timestamp();
+}
+
+AudioSplicer::AudioSplicer(int samples_per_second)
+    : max_crossfade_duration_(
+          base::TimeDelta::FromMilliseconds(kCrossfadeDurationInMilliseconds)),
+      splice_timestamp_(kNoTimestamp()),
+      output_sanitizer_(new AudioStreamSanitizer(samples_per_second)),
+      pre_splice_sanitizer_(new AudioStreamSanitizer(samples_per_second)),
+      post_splice_sanitizer_(new AudioStreamSanitizer(samples_per_second)) {}
+
+AudioSplicer::~AudioSplicer() {}
+
+void AudioSplicer::Reset() {
+  output_sanitizer_->Reset();
+  pre_splice_sanitizer_->Reset();
+  post_splice_sanitizer_->Reset();
+  splice_timestamp_ = kNoTimestamp();
+}
+
+bool AudioSplicer::AddInput(const scoped_refptr<AudioBuffer>& input) {
+  // If we're not processing a splice, add the input to the output queue.
+  if (splice_timestamp_ == kNoTimestamp()) {
+    DCHECK(!pre_splice_sanitizer_->HasNextBuffer());
+    DCHECK(!post_splice_sanitizer_->HasNextBuffer());
+    return output_sanitizer_->AddInput(input);
+  }
+
+  // If we're still receiving buffers before the splice point figure out which
+  // sanitizer (if any) to put them in.
+  if (!post_splice_sanitizer_->HasNextBuffer()) {
+    DCHECK(!input->end_of_stream());
+
+    // If the provided buffer is entirely before the splice point it can also be
+    // added to the output queue.
+    if (input->timestamp() + input->duration() < splice_timestamp_) {
+      DCHECK(!pre_splice_sanitizer_->HasNextBuffer());
+      return output_sanitizer_->AddInput(input);
+    }
+
+    // If we've encountered the first pre splice buffer, reset the pre splice
+    // sanitizer based on |output_sanitizer_|.  This is done so that gaps and
+    // overlaps between buffers across the sanitizers are accounted for prior
+    // to calculating crossfade.
+    if (!pre_splice_sanitizer_->HasNextBuffer()) {
+      pre_splice_sanitizer_->ResetTimestampState(
+          output_sanitizer_->timestamp_helper().frame_count(),
+          output_sanitizer_->timestamp_helper().base_timestamp());
+    }
+
+    // If we're processing a splice and the input buffer does not overlap any of
+    // the existing buffers, append it to the splice queue for processing.
+    if (!pre_splice_sanitizer_->HasNextBuffer() ||
+        input->timestamp() != splice_timestamp_) {
+      return pre_splice_sanitizer_->AddInput(input);
+    }
+
+    // We've received the first overlapping buffer.
+  }
+
+  // At this point we have all the fade out preroll buffers from the decoder.
+  // We now need to wait until we have enough data to perform the crossfade (or
+  // we receive an end of stream).
+  if (!post_splice_sanitizer_->AddInput(input))
+    return false;
+
+  if (!input->end_of_stream() &&
+      post_splice_sanitizer_->GetDuration() < max_crossfade_duration_) {
+    return true;
+  }
+
+  // Transfer out preroll buffers involved in the splice, drop those not.  Since
+  // we don't want to care what format the AudioBuffers are in, we need to use
+  // an intermediary AudioBus to convert the data to float.
+  scoped_ptr<AudioBus> pre_splice_bus = ExtractCrossfadeFromPreSplice();
+
+  // Allocate output buffer for crossfade.
+  scoped_refptr<AudioBuffer> crossfade_buffer =
+      AudioBuffer::CreateBuffer(kSampleFormatPlanarF32,
+                                pre_splice_bus->channels(),
+                                pre_splice_bus->frames());
+
+  // Use the calculated timestamp and duration to ensure there's no extra gaps
+  // or overlaps to process when adding the buffer to |output_sanitizer_|.
+  const AudioTimestampHelper& output_ts_helper =
+      output_sanitizer_->timestamp_helper();
+  crossfade_buffer->set_timestamp(output_ts_helper.GetTimestamp());
+  crossfade_buffer->set_duration(
+      output_ts_helper.GetFrameDuration(pre_splice_bus->frames()));
+
+  // AudioBuffer::ReadFrames() only allows output into an AudioBus, so wrap
+  // our AudioBuffer in one so we can avoid extra data copies.
+  scoped_ptr<AudioBus> crossfade_bus_wrapper =
+      AudioBus::CreateWrapper(crossfade_buffer->channel_count());

[acolwell GONE FROM CHROMIUM, 2014/02/28 18:50:27]

nit: Please move this and the following 5 lines into a helper method. I think it
makes this section less readable.

[DaleCurtis, 2014/02/28 21:14:26]

Done.

+  crossfade_bus_wrapper->set_frames(crossfade_buffer->frame_count());
+  for (int ch = 0; ch < crossfade_buffer->channel_count(); ++ch) {
+    crossfade_bus_wrapper->SetChannelData(
+        ch, reinterpret_cast<float*>(crossfade_buffer->channel_data()[ch]));
+  }
+
+  // Insert the crossfade buffer into the output queue now so post splice
+  // buffers can be added in processing order.  We will still modify the buffer
+  // during the crossfade step.
+  CHECK(output_sanitizer_->AddInput(crossfade_buffer));
+  DCHECK_EQ(crossfade_buffer->frame_count(), crossfade_bus_wrapper->frames());
+
+  ExtractCrossfadeFromPostSplice(crossfade_bus_wrapper.get());
+
+  // Crossfade the audio into |crossfade_buffer|.
+  for (int ch = 0; ch < crossfade_bus_wrapper->channels(); ++ch) {
+    vector_math::Crossfade(pre_splice_bus->channel(ch),
+                           pre_splice_bus->frames(),
+                           crossfade_bus_wrapper->channel(ch));
+  }
+
+  // Clear the splice timestamp so new splices can be accepted.
+  splice_timestamp_ = kNoTimestamp();
+  return true;
+}
+
+bool AudioSplicer::HasNextBuffer() const {
+  return output_sanitizer_->HasNextBuffer();
+}
+
+scoped_refptr<AudioBuffer> AudioSplicer::GetNextBuffer() {
+  return output_sanitizer_->GetNextBuffer();
+}
+
+void AudioSplicer::SetSpliceTimestamp(base::TimeDelta splice_timestamp) {
+  DCHECK(splice_timestamp != kNoTimestamp());
+  if (splice_timestamp_ == splice_timestamp)
+    return;
+
+  // TODO(dalecurtis): We may need the concept of a future_splice_timestamp_ to
+  // handle cases where another splice comes in before we've received 5ms of
+  // data from the last one.  Leave this as a CHECK for now to figure out if
+  // this case is possible.
+  CHECK(splice_timestamp_ == kNoTimestamp());
+  splice_timestamp_ = splice_timestamp;
+}
+
+scoped_ptr<AudioBus> AudioSplicer::ExtractCrossfadeFromPreSplice() {
+  const AudioTimestampHelper& output_ts_helper =
+      output_sanitizer_->timestamp_helper();
+
+  // Ensure |output_sanitizer_| has a valid base timestamp so we can use it for
+  // timestamp calculations.
+  if (output_ts_helper.base_timestamp() == kNoTimestamp()) {
+    output_sanitizer_->ResetTimestampState(
+        0, pre_splice_sanitizer_->timestamp_helper().base_timestamp());
+  }
+
+  int frames_before_splice =
+      output_ts_helper.GetFramesToTarget(splice_timestamp_);
+
+  // Determine crossfade frame count based on available frames in each splicer
+  // and capping to the maximum crossfade duration.
+  const int max_crossfade_frame_count =
+      output_ts_helper.GetFramesToTarget(splice_timestamp_ +
+                                         max_crossfade_duration_) -
+      frames_before_splice;
+  const int frames_to_crossfade = std::min(
+      max_crossfade_frame_count,
+      std::min(pre_splice_sanitizer_->GetFrameCount() - frames_before_splice,
+               post_splice_sanitizer_->GetFrameCount()));
+
+  int frames_read = 0;
+  scoped_ptr<AudioBus> output_bus;
+  while (pre_splice_sanitizer_->HasNextBuffer() &&
+         frames_read < frames_to_crossfade) {
+    scoped_refptr<AudioBuffer> preroll = pre_splice_sanitizer_->GetNextBuffer();
+
+    // We don't know the channel count until we see the first buffer, so wait
+    // until the first buffer to allocate the output AudioBus.
+    if (!output_bus) {
+      output_bus =
+          AudioBus::Create(preroll->channel_count(), frames_to_crossfade);
+    }
+
+    // There may be enough of a gap introduced during decoding such that an
+    // entire buffer exists before the splice point.
+    if (frames_before_splice >= preroll->frame_count()) {
+      frames_before_splice -= preroll->frame_count();
+      CHECK(output_sanitizer_->AddInput(preroll));
+      continue;
+    }
+
+    const int frames_to_read =
+        std::min(preroll->frame_count() - frames_before_splice,
+                 output_bus->frames() - frames_read);
+    preroll->ReadFrames(
+        frames_to_read, frames_before_splice, frames_read, output_bus.get());
+    frames_read += frames_to_read;
+
+    // If only part of the buffer was consumed, trim it appropriately and stick
+    // it into the output queue.
+    if (frames_before_splice) {
+      AccurateTrimEnd(preroll->frame_count() - frames_before_splice,
+                      preroll,
+                      output_ts_helper);
+      CHECK(output_sanitizer_->AddInput(preroll));
+      frames_before_splice = 0;
+    }
+  }
+
+  // All necessary buffers have been processed, it's safe to reset.
+  pre_splice_sanitizer_->Reset();
+  DCHECK_EQ(output_bus->frames(), frames_read);
+  DCHECK_EQ(output_ts_helper.GetFramesToTarget(splice_timestamp_), 0);
+  return output_bus.Pass();
+}
+
+void AudioSplicer::ExtractCrossfadeFromPostSplice(AudioBus* output_bus) {
+  int frames_read = 0;
+  while (post_splice_sanitizer_->HasNextBuffer() &&
+         frames_read < output_bus->frames()) {
+    scoped_refptr<AudioBuffer> postroll =
+        post_splice_sanitizer_->GetNextBuffer();
+    const int frames_to_read =
+        std::min(postroll->frame_count(), output_bus->frames() - frames_read);
+    postroll->ReadFrames(frames_to_read, 0, frames_read, output_bus);
+    frames_read += frames_to_read;
+
+    // If only part of the buffer was consumed, trim it appropriately and stick
+    // it into the output queue.
+    if (frames_to_read < postroll->frame_count()) {
+      AccurateTrimStart(
+          frames_to_read, postroll, output_sanitizer_->timestamp_helper());
+      CHECK(output_sanitizer_->AddInput(postroll));
+    }
+  }
+
+  DCHECK_EQ(output_bus->frames(), frames_read);
+
+  // Transfer all remaining buffers out and reset once empty.
+  while (post_splice_sanitizer_->HasNextBuffer())
+    CHECK(output_sanitizer_->AddInput(post_splice_sanitizer_->GetNextBuffer()));
+  post_splice_sanitizer_->Reset();
+}
+
 }  // namespace media