Low-latency AudioInputStream implementation based on WASAPI for Windows.

content/renderer/media/webrtc_audio_device_impl.cc

 // Copyright (c) 2011 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "content/renderer/media/webrtc_audio_device_impl.h"
 
 #include "base/bind.h"
 #include "base/string_util.h"
+#include "content/common/view_messages.h"
 #include "content/renderer/render_thread_impl.h"
 #include "media/audio/audio_util.h"
 
 static const int64 kMillisecondsBetweenProcessCalls = 5000;
 static const char kVersion[] = "WebRTC AudioDevice 1.0.0.Chrome";
 
+static int GetAudioInputHardwareSampleRate() {

[henrika (OOO until Aug 14), 2011/10/18 08:13:14]

I decided to ask for the input sample rate here instead of doing it within the
AudioInputDevice class.

+  static double input_sample_rate = 0;
+  if (!input_sample_rate) {
+    RenderThreadImpl::current()->Send(
+        new ViewHostMsg_GetHardwareInputSampleRate(&input_sample_rate));
+  }
+  return static_cast<int>(input_sample_rate);
+}
+
 WebRtcAudioDeviceImpl::WebRtcAudioDeviceImpl()
     : ref_count_(0),
       render_loop_(base::MessageLoopProxy::current()),
       audio_transport_callback_(NULL),
       input_buffer_size_(0),
       output_buffer_size_(0),
       input_channels_(0),
       output_channels_(0),
       input_sample_rate_(0),
       output_sample_rate_(0),
@@ skipping 40 matching lines @@
 
   // Store the reported audio delay locally.
   output_delay_ms_ = audio_delay_milliseconds;
 
   const int channels = audio_data.size();
   DCHECK_LE(channels, output_channels_);
 
   int samples_per_sec = static_cast<int>(output_sample_rate_);
   if (samples_per_sec == 44100) {
     // Even if the hardware runs at 44.1kHz, we use 44.0 internally.
-    // Can only happen on Mac OS X currently since Windows and Mac
-    // both uses 48kHz.
     samples_per_sec = 44000;
   }
   uint32_t samples_per_10_msec = (samples_per_sec / 100);
   const int bytes_per_10_msec =
       channels * samples_per_10_msec * bytes_per_sample_;
 
   uint32_t num_audio_samples = 0;
   size_t accumulated_audio_samples = 0;
 
   char* audio_byte_buffer = reinterpret_cast<char*>(output_buffer_.get());
@@ skipping 38 matching lines @@
   const int channels = audio_data.size();
   DCHECK_LE(channels, input_channels_);
   uint32_t new_mic_level = 0;
 
   // Interleave, scale, and clip input to int16 and store result in
   // a local byte buffer.
   media::InterleaveFloatToInt16(audio_data,
                                 input_buffer_.get(),
                                 number_of_frames);
 
-  const int samples_per_sec = static_cast<int>(input_sample_rate_);
+  int samples_per_sec = static_cast<int>(input_sample_rate_);
+  if (samples_per_sec == 44100) {
+    // Even if the hardware runs at 44.1kHz, we use 44.0 internally.
+    samples_per_sec = 44000;
+  }
   const int samples_per_10_msec = (samples_per_sec / 100);
   const int bytes_per_10_msec =
       channels * samples_per_10_msec * bytes_per_sample_;
   size_t accumulated_audio_samples = 0;
 
   char* audio_byte_buffer = reinterpret_cast<char*>(input_buffer_.get());
 
   // Write audio samples in blocks of 10 milliseconds to the registered
   // webrtc::AudioTransport sink. Keep writing until our internal byte
   // buffer is empty.
@@ skipping 121 matching lines @@
 
   // Calling Init() multiple times in a row is OK.
   if (initialized_)
     return 0;
 
   DCHECK(!audio_input_device_);
   DCHECK(!audio_output_device_);
   DCHECK(!input_buffer_.get());
   DCHECK(!output_buffer_.get());
 
-  // TODO(henrika): add AudioInputDevice::GetAudioHardwareSampleRate().
-  // Assume that input and output sample rates are identical for now.
-
   // Ask the browser for the default audio output hardware sample-rate.
   // This request is based on a synchronous IPC message.
   int output_sample_rate =
       static_cast<int>(AudioDevice::GetAudioHardwareSampleRate());
-  VLOG(1) << "Audio hardware sample rate: " << output_sample_rate;
+  VLOG(1) << "Audio output hardware sample rate: " << output_sample_rate;
+
+  // Ask the browser for the default audio input hardware sample-rate.
+  // This request is based on a synchronous IPC message.
+  int input_sample_rate = GetAudioInputHardwareSampleRate();

[henrika (OOO until Aug 14), 2011/10/18 08:13:14]

We now ask for the input sample rate as well.

+  VLOG(1) << "Audio input hardware sample rate: " << input_sample_rate;
 
   int input_channels = 0;
   int output_channels = 0;
 
   size_t input_buffer_size = 0;
   size_t output_buffer_size = 0;
 
   // For real-time audio (in combination with the webrtc::VoiceEngine) it
   // is convenient to use audio buffers of size N*10ms.
+
 #if defined(OS_WIN)
   if (output_sample_rate != 48000) {
     DLOG(ERROR) << "Only 48kHz sample rate is supported on Windows.";
     return -1;
   }
-  input_channels = 1;
+
+  // Use stereo recording on Windows since low-latency Core Audio (WASAPI)
+  // does not support mono.
+  input_channels = 2;
   output_channels = 1;
-  // Capture side: AUDIO_PCM_LINEAR on Windows is based on a callback-
-  // driven Wave implementation where 3 buffers are used for recording audio.
-  // Each buffer is of the size that we specify here and using more than one
-  // does not increase the delay but only adds robustness against dropping
-  // audio. It might also affect the initial start-up time before callbacks
-  // start to pump. Real-time tests have shown that a buffer size of 10ms
-  // works fine on the capture side.
-  input_buffer_size = 480;
+
+  // Capture side: AUDIO_PCM_LOW_LATENCY is based on the Core Audio (WASAPI)
+  // API which was introduced in Windows Vista. For lower Windows versions,
+  // a callback-driven Wave implementation is used instead. An input buffer
+  // size of 10ms works well for both these implementations.
+
+  // Use different buffer sizes depending on the current hardware sample rate.
+  if (input_sample_rate == 48000) {
+    input_buffer_size = 480;
+  } else {
+    // We do run at 44.1kHz at the actual audio layer, but ask for frames
+    // at 44.0kHz to ensure that we can feed them to the webrtc::VoiceEngine.
+    input_buffer_size = 440;
+  }
+
   // Rendering side: AUDIO_PCM_LOW_LATENCY on Windows is based on a callback-
   // driven Wave implementation where 2 buffers are fed to the audio driver
   // before actual rendering starts. Initial real-time tests have shown that
   // 20ms buffer size (corresponds to ~40ms total delay) is not enough but
   // can lead to buffer underruns. The next even multiple of 10ms is 30ms
   // (<=> ~60ms total delay) and it works fine also under high load.
   output_buffer_size = 3 * 480;
 #elif defined(OS_MACOSX)
   if (output_sample_rate != 48000 && output_sample_rate != 44100) {
     DLOG(ERROR) << "Only 48 and 44.1kHz sample rates are supported on Mac OSX.";
     return -1;
   }
   input_channels = 1;
   output_channels = 1;
+
   // Rendering side: AUDIO_PCM_LOW_LATENCY on Mac OS X is based on a callback-
   // driven Core Audio implementation. Tests have shown that 10ms is a suitable
   // frame size to use, both for 48kHz and 44.1kHz.
   // Capturing side: AUDIO_PCM_LINEAR on Mac OS X uses the Audio Queue Services
   // API which is not well suited for real-time applications since the delay
   // is very high. We set buffer sizes to 10ms for the input side here as well
   // but none of them will work.
   // TODO(henrika): add support for AUDIO_PCM_LOW_LATENCY on the capture side
   // based on the Mac OS X Core Audio API.
 
   // Use different buffer sizes depending on the current hardware sample rate.
   if (output_sample_rate == 48000) {
     input_buffer_size = 480;
     output_buffer_size = 480;
   } else {
     // We do run at 44.1kHz at the actual audio layer, but ask for frames
     // at 44.0kHz to ensure that we can feed them to the webrtc::VoiceEngine.
     input_buffer_size = 440;
     output_buffer_size = 440;
   }
 #elif defined(OS_LINUX)
   if (output_sample_rate != 48000) {
     DLOG(ERROR) << "Only 48kHz sample rate is supported on Linux.";
     return -1;
   }
   input_channels = 1;
   output_channels = 1;
+
   // Based on tests using the current ALSA implementation in Chrome, we have
   // found that the best combination is 20ms on the input side and 10ms on the
   // output side.
   // TODO(henrika): It might be possible to reduce the input buffer
   // size and reduce the delay even more.
   input_buffer_size = 2 * 480;
   output_buffer_size = 480;
 #else
   DLOG(ERROR) << "Unsupported platform";
   return -1;
 #endif
 
   // Store utilized parameters to ensure that we can check them
   // after a successful initialization.
   output_buffer_size_ = output_buffer_size;
   output_channels_ = output_channels;
   output_sample_rate_ = static_cast<double>(output_sample_rate);
 
   input_buffer_size_ = input_buffer_size;
   input_channels_ = input_channels;
-  // TODO(henrika): we use same rate as on output for now.
-  input_sample_rate_ = output_sample_rate_;
+  input_sample_rate_ = input_sample_rate;
 
   // Create and configure the audio capturing client.
   audio_input_device_ = new AudioInputDevice(
-      input_buffer_size, input_channels, output_sample_rate, this, this);
-#if defined(OS_MACOSX)
-  // We create the input device for Mac as well but the performance
-  // will be very bad.
-  DLOG(WARNING) << "Real-time recording is not yet supported on Mac OS X";
-#endif
+      input_buffer_size, input_channels, input_sample_rate, this, this);

[henrika (OOO until Aug 14), 2011/10/18 08:13:14]

Not using same rate as output any longer since we can now ask for native input
sample as well.

 
   // Create and configure the audio rendering client.
   audio_output_device_ = new AudioDevice(
       output_buffer_size, output_channels, output_sample_rate, this);
 
   DCHECK(audio_input_device_);
   DCHECK(audio_output_device_);
 
   // Allocate local audio buffers based on the parameters above.
   // It is assumed that each audio sample contains 16 bits and each
@@ skipping 532 matching lines @@
 }
 
 int32_t WebRtcAudioDeviceImpl::GetLoudspeakerStatus(bool* enabled) const {
   NOTIMPLEMENTED();
   return -1;
 }
 
 void WebRtcAudioDeviceImpl::SetSessionId(int session_id) {
   session_id_ = session_id;
 }