Convert IntelligibilityEnhancer to multi-threaded mode

webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc

Index: webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
diff --git a/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc b/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
index dc2dcdd2610542a69238a4317c24b7a18286f50e..d2d5edb223f87084f5a557f841b8cad755fc62d0 100644
--- a/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
+++ b/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
@@ -61,19 +61,21 @@ void MapToErbBands(const float* pow,
 }  // namespace
 
 IntelligibilityEnhancer::IntelligibilityEnhancer(int sample_rate_hz,
-                                                 size_t num_render_channels)
+                                                 size_t num_render_channels,
+                                                 size_t num_noise_bins)
     : freqs_(RealFourier::ComplexLength(
           RealFourier::FftOrder(sample_rate_hz * kWindowSizeMs / 1000))),
+      num_noise_bins_(num_noise_bins),
       chunk_length_(static_cast<size_t>(sample_rate_hz * kChunkSizeMs / 1000)),
       bank_size_(GetBankSize(sample_rate_hz, kErbResolution)),
       sample_rate_hz_(sample_rate_hz),
       num_render_channels_(num_render_channels),
       clear_power_estimator_(freqs_, kDecayRate),
-      noise_power_estimator_(
-          new intelligibility::PowerEstimator<float>(freqs_, kDecayRate)),
+      noise_power_estimator_(freqs_, kDecayRate),
       filtered_clear_pow_(bank_size_, 0.f),
-      filtered_noise_pow_(bank_size_, 0.f),
+      filtered_noise_pow_(num_noise_bins, 0.f),

[turaj, 2016/03/07 20:34:35]

I suppose I'm missing a point, |filtered_noise_pow_| and |filtered_clear_pow_|
are the ones used to compute IE gains, right? Then why are they of different
sizes?

[aluebs-webrtc, 2016/03/08 10:53:00]

They can (and have) different sizes, since one is the psd of the speech
calculated internally and the other is the size of the noise estimates 
calculated by the noise suppressor and passed in. This mismatch is fixed when
mapping them to critical bands.

[turaj, 2016/03/08 15:35:04]

Thanks for the explanation. 

       center_freqs_(bank_size_),
+      capture_filter_bank_(CreateErbBank(num_noise_bins)),
       render_filter_bank_(CreateErbBank(freqs_)),
       gains_eq_(bank_size_),
       gain_applier_(freqs_, kMaxRelativeGainChange),
@@ -94,17 +96,19 @@ IntelligibilityEnhancer::IntelligibilityEnhancer(int sample_rate_hz,
   render_mangler_.reset(new LappedTransform(
       num_render_channels_, num_render_channels_, chunk_length_,
       kbd_window.data(), window_size, window_size / 2, this));
+
+  std::vector<float> template_queue_element(num_noise_bins);
+  noise_estimation_queue_.reset(
+      new SwapQueue<std::vector<float>, RenderQueueItemVerifier<float>>(
+          kMaxNumFramesToBuffer, template_queue_element,
+          RenderQueueItemVerifier<float>(num_noise_bins)));
+  noise_estimation_buffer_.resize(num_noise_bins);

[turaj, 2016/03/07 20:34:35]

Any particular reason that |noise_estimation_buffer_| is not initialized in
initializer list?

[aluebs-webrtc, 2016/03/08 10:53:00]

No really. Changed.

 }
 
 void IntelligibilityEnhancer::SetCaptureNoiseEstimate(
     std::vector<float> noise) {
-  if (capture_filter_bank_.size() != bank_size_ ||
-      capture_filter_bank_[0].size() != noise.size()) {
-    capture_filter_bank_ = CreateErbBank(noise.size());
-    noise_power_estimator_.reset(
-        new intelligibility::PowerEstimator<float>(noise.size(), kDecayRate));
-  }
-  noise_power_estimator_->Step(noise.data());
+  RTC_CHECK_EQ(noise.size(), num_noise_bins_);

[peah-webrtc, 2016/03/08 07:02:38]

Do we really want a CHECK here? I think a DCHECK should be preferred.

[hlundin-webrtc, 2016/03/08 10:28:48]

Acknowledged.

[aluebs-webrtc, 2016/03/08 10:53:00]

Agreed. Done.

+  (void)noise_estimation_queue_->Insert(&noise);

[peah-webrtc, 2016/03/08 07:02:38]

Please add a comment on why buffer overflow is acceptable.

[aluebs-webrtc, 2016/03/08 10:53:00]

Done.

 }
 
 void IntelligibilityEnhancer::ProcessRenderAudio(float* const* audio,
@@ -112,6 +116,9 @@ void IntelligibilityEnhancer::ProcessRenderAudio(float* const* audio,
                                                  size_t num_channels) {
   RTC_CHECK_EQ(sample_rate_hz_, sample_rate_hz);
   RTC_CHECK_EQ(num_render_channels_, num_channels);
+  while (noise_estimation_queue_->Remove(&noise_estimation_buffer_)) {
+    noise_power_estimator_.Step(noise_estimation_buffer_.data());
+  }
   is_speech_ = IsSpeech(audio[0]);
   render_mangler_->ProcessChunk(audio, audio);
 }
@@ -127,7 +134,7 @@ void IntelligibilityEnhancer::ProcessAudioBlock(
     clear_power_estimator_.Step(in_block[0]);
   }
   const std::vector<float>& clear_power = clear_power_estimator_.power();
-  const std::vector<float>& noise_power = noise_power_estimator_->power();
+  const std::vector<float>& noise_power = noise_power_estimator_.power();
   MapToErbBands(clear_power.data(), render_filter_bank_,
                 filtered_clear_pow_.data());
   MapToErbBands(noise_power.data(), capture_filter_bank_,