SPU: Add time stretched audio output

duckstation.sln

@@ -102,7 +102,6 @@ Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "fmt", "dep\fmt\fmt.vcxproj"
 EndProject
 Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "util", "src\util\util.vcxproj", "{57F6206D-F264-4B07-BAF8-11B9BBE1F455}"
 	ProjectSection(ProjectDependencies) = postProject
-		{39F0ADFF-3A84-470D-9CF0-CA49E164F2F3} = {39F0ADFF-3A84-470D-9CF0-CA49E164F2F3}
 		{425D6C99-D1C8-43C2-B8AC-4D7B1D941017} = {425D6C99-D1C8-43C2-B8AC-4D7B1D941017}
 		{751D9F62-881C-454E-BCE8-CB9CF5F1D22F} = {751D9F62-881C-454E-BCE8-CB9CF5F1D22F}
 		{EE054E08-3799-4A59-A422-18259C105FFD} = {EE054E08-3799-4A59-A422-18259C105FFD}

src/core/host.h

@@ -13,6 +13,7 @@
 
 struct WindowInfo;
 enum class AudioBackend : u8;
+enum class AudioStretchMode : u8;
 class AudioStream;
 class CDImage;
 
@@ -77,7 +78,8 @@ std::optional<std::time_t> GetResourceFileTimestamp(const char* filename);
 TinyString TranslateString(const char* context, const char* str, const char* disambiguation = nullptr, int n = -1);
 std::string TranslateStdString(const char* context, const char* str, const char* disambiguation = nullptr, int n = -1);
 
-std::unique_ptr<AudioStream> CreateAudioStream(AudioBackend backend);
+std::unique_ptr<AudioStream> CreateAudioStream(AudioBackend backend, u32 sample_rate, u32 channels, u32 buffer_ms,
+                                               u32 latency_ms, AudioStretchMode stretch);
 
 /// Returns the scale of OSD elements.
 float GetOSDScale();

src/core/settings.cpp

@@ -275,10 +275,15 @@ void Settings::Load(SettingsInterface& si)
   audio_backend =
     ParseAudioBackend(si.GetStringValue("Audio", "Backend", GetAudioBackendName(DEFAULT_AUDIO_BACKEND)).c_str())
       .value_or(DEFAULT_AUDIO_BACKEND);
-  audio_output_volume = si.GetIntValue("Audio", "OutputVolume", 100);
+  audio_stretch_mode =
-  audio_fast_forward_volume = si.GetIntValue("Audio", "FastForwardVolume", 100);
+    AudioStream::ParseStretchMode(
-  audio_buffer_size = si.GetIntValue("Audio", "BufferSize", DEFAULT_AUDIO_BUFFER_SIZE);
+      si.GetStringValue("Audio", "StretchMode", AudioStream::GetStretchModeName(DEFAULT_AUDIO_STRETCH_MODE)).c_str())
-  audio_resampling = si.GetBoolValue("Audio", "Resampling", true);
+      .value_or(DEFAULT_AUDIO_STRETCH_MODE);
+  audio_output_latency_ms = si.GetUIntValue("Audio", "OutputLatencyMS", DEFAULT_AUDIO_OUTPUT_LATENCY_MS);
+  audio_buffer_ms = si.GetUIntValue("Audio", "BufferMS", DEFAULT_AUDIO_BUFFER_MS);
+  audio_output_volume = si.GetUIntValue("Audio", "OutputVolume", 100);
+  audio_fast_forward_volume = si.GetUIntValue("Audio", "FastForwardVolume", 100);
+
   audio_output_muted = si.GetBoolValue("Audio", "OutputMuted", false);
   audio_sync_enabled = si.GetBoolValue("Audio", "Sync", true);
   audio_dump_on_boot = si.GetBoolValue("Audio", "DumpOnBoot", false);
@@ -472,10 +477,11 @@ void Settings::Save(SettingsInterface& si) const
   si.SetIntValue("CDROM", "SeekSpeedup", cdrom_seek_speedup);
 
   si.SetStringValue("Audio", "Backend", GetAudioBackendName(audio_backend));
-  si.SetIntValue("Audio", "OutputVolume", audio_output_volume);
+  si.SetStringValue("Audio", "StretchMode", AudioStream::GetStretchModeName(audio_stretch_mode));
-  si.SetIntValue("Audio", "FastForwardVolume", audio_fast_forward_volume);
+  si.SetUIntValue("Audio", "BufferMS", audio_buffer_ms);
-  si.SetIntValue("Audio", "BufferSize", audio_buffer_size);
+  si.SetUIntValue("Audio", "OutputLatencyMS", audio_output_latency_ms);
-  si.SetBoolValue("Audio", "Resampling", audio_resampling);
+  si.SetUIntValue("Audio", "OutputVolume", audio_output_volume);
+  si.SetUIntValue("Audio", "FastForwardVolume", audio_fast_forward_volume);
   si.SetBoolValue("Audio", "OutputMuted", audio_output_muted);
   si.SetBoolValue("Audio", "Sync", audio_sync_enabled);
   si.SetBoolValue("Audio", "DumpOnBoot", audio_dump_on_boot);

src/core/settings.h

@@ -3,6 +3,7 @@
 #include "common/settings_interface.h"
 #include "common/string.h"
 #include "types.h"
+#include "util/audio_stream.h"
 #include <array>
 #include <optional>
 #include <string>
@@ -142,10 +143,11 @@ struct Settings
   u32 cdrom_seek_speedup = 1;
 
   AudioBackend audio_backend = DEFAULT_AUDIO_BACKEND;
-  s32 audio_output_volume = 100;
+  AudioStretchMode audio_stretch_mode = DEFAULT_AUDIO_STRETCH_MODE;
-  s32 audio_fast_forward_volume = 100;
+  u32 audio_output_latency_ms = DEFAULT_AUDIO_OUTPUT_LATENCY_MS;
-  u32 audio_buffer_size = DEFAULT_AUDIO_BUFFER_SIZE;
+  u32 audio_buffer_ms = DEFAULT_AUDIO_BUFFER_MS;
-  bool audio_resampling = true;
+  u32 audio_output_volume = 100;
+  u32 audio_fast_forward_volume = 100;
   bool audio_output_muted = false;
   bool audio_sync_enabled = true;
   bool audio_dump_on_boot = false;
@@ -400,7 +402,9 @@ struct Settings
 
   static constexpr LOGLEVEL DEFAULT_LOG_LEVEL = LOGLEVEL_INFO;
 
-  static constexpr u32 DEFAULT_AUDIO_BUFFER_SIZE = 2048;
+  static constexpr u32 DEFAULT_AUDIO_BUFFER_MS = 50;
+  static constexpr u32 DEFAULT_AUDIO_OUTPUT_LATENCY_MS = 20;
+  static constexpr AudioStretchMode DEFAULT_AUDIO_STRETCH_MODE = AudioStretchMode::TimeStretch;
 
   // Enable console logging by default on Linux platforms.
 #if defined(__linux__) && !defined(__ANDROID__)

src/core/spu.cpp

@@ -30,8 +30,7 @@ void SPU::Initialize()
     "SPU Transfer", TRANSFER_TICKS_PER_HALFWORD, TRANSFER_TICKS_PER_HALFWORD,
     [](void* param, TickCount ticks, TickCount ticks_late) { static_cast<SPU*>(param)->ExecuteTransfer(ticks); }, this,
     false);
-  m_null_audio_stream = AudioStream::CreateNullAudioStream();
+  m_null_audio_stream = AudioStream::CreateNullStream(SAMPLE_RATE, NUM_CHANNELS, g_settings.audio_buffer_ms);
-  m_null_audio_stream->Reconfigure(SAMPLE_RATE, SAMPLE_RATE, NUM_CHANNELS, Settings::DEFAULT_AUDIO_BUFFER_SIZE);
 
   CreateOutputStream();
   Reset();
@@ -39,22 +38,23 @@ void SPU::Initialize()
 
 void SPU::CreateOutputStream()
 {
-  Log_InfoPrintf("Creating '%s' audio stream, sample rate = %u, channels = %u, buffer size = %u",
+  Log_InfoPrintf(
-                 Settings::GetAudioBackendName(g_settings.audio_backend), SAMPLE_RATE, NUM_CHANNELS,
+    "Creating '%s' audio stream, sample rate = %u, channels = %u, buffer = %u, latency = %u, stretching = %s",
-                 g_settings.audio_buffer_size);
+    Settings::GetAudioBackendName(g_settings.audio_backend), SAMPLE_RATE, NUM_CHANNELS, g_settings.audio_buffer_ms,
+    g_settings.audio_output_latency_ms, AudioStream::GetStretchModeName(g_settings.audio_stretch_mode));
 
-  m_audio_stream = Host::CreateAudioStream(g_settings.audio_backend);
+  m_audio_stream =
-
+    Host::CreateAudioStream(g_settings.audio_backend, SAMPLE_RATE, NUM_CHANNELS, g_settings.audio_buffer_ms,
-  if (!m_audio_stream ||
+                            g_settings.audio_output_latency_ms, g_settings.audio_stretch_mode);
-      !m_audio_stream->Reconfigure(SAMPLE_RATE, SAMPLE_RATE, NUM_CHANNELS, g_settings.audio_buffer_size))
+  if (!m_audio_stream)
   {
     Host::ReportErrorAsync("Error", "Failed to create or configure audio stream, falling back to null output.");
     m_audio_stream.reset();
-    m_audio_stream = AudioStream::CreateNullAudioStream();
+    m_audio_stream = AudioStream::CreateNullStream(SAMPLE_RATE, NUM_CHANNELS, g_settings.audio_buffer_ms);
-    m_audio_stream->Reconfigure(SAMPLE_RATE, SAMPLE_RATE, NUM_CHANNELS, g_settings.audio_buffer_size);
   }
 
   m_audio_stream->SetOutputVolume(System::GetAudioOutputVolume());
+  m_audio_stream->SetPaused(System::IsPaused());
 }
 
 void SPU::RecreateOutputStream()
@@ -77,7 +77,7 @@ void SPU::Shutdown()
   m_tick_event.reset();
   m_transfer_event.reset();
   m_dump_writer.reset();
-  m_audio_stream = nullptr;
+  m_audio_stream.reset();
 }
 
 void SPU::Reset()

src/core/system.cpp

@@ -923,9 +923,7 @@ void System::PauseSystem(bool paused)
     return;
 
   SetState(paused ? State::Paused : State::Running);
-  if (!paused)
+  g_spu.GetOutputStream()->SetPaused(paused);
-    g_spu.GetOutputStream()->EmptyBuffers();
-  g_spu.GetOutputStream()->PauseOutput(paused);
 
   if (paused)
   {
@@ -1179,7 +1177,7 @@ bool System::BootSystem(SystemBootParameters parameters)
   // Good to go.
   Host::OnSystemStarted();
   UpdateSoftwareCursor();
-  g_spu.GetOutputStream()->PauseOutput(false);
+  g_spu.GetOutputStream()->SetPaused(false);
 
   // Initial state must be set before loading state.
   s_state =
@@ -1813,7 +1811,7 @@ bool System::DoLoadState(ByteStream* state, bool force_software_renderer, bool u
   if (s_state == State::Starting)
     s_state = State::Running;
 
-  g_spu.GetOutputStream()->EmptyBuffers();
+  g_spu.GetOutputStream()->EmptyBuffer();
   ResetPerformanceCounters();
   ResetThrottler();
   return true;
@@ -2035,14 +2033,6 @@ void System::ResetThrottler()
 
 void System::Throttle()
 {
-  // Reset the throttler on audio buffer overflow, so we don't end up out of phase.
-  if (g_spu.GetOutputStream()->DidUnderflow() && s_target_speed >= 1.0f)
-  {
-    Log_VerbosePrintf("Audio buffer underflowed, resetting throttler");
-    ResetThrottler();
-    return;
-  }
-
   // Allow variance of up to 40ms either way.
 #ifndef __ANDROID__
   static constexpr double MAX_VARIANCE_TIME_NS = 40 * 1000000;
@@ -2181,7 +2171,8 @@ void System::UpdateSpeedLimiterState()
   m_display_all_frames = !m_throttler_enabled || g_settings.display_all_frames;
 
   bool syncing_to_host = false;
-  if (g_settings.sync_to_host_refresh_rate && g_settings.audio_resampling && target_speed == 1.0f && IsRunning())
+  if (g_settings.sync_to_host_refresh_rate && (g_settings.audio_stretch_mode != AudioStretchMode::Off) &&
+      target_speed == 1.0f && IsValid())
   {
     float host_refresh_rate;
     if (g_host_display->GetHostRefreshRate(&host_refresh_rate))
@@ -2212,21 +2203,18 @@ void System::UpdateSpeedLimiterState()
     UpdateThrottlePeriod();
     ResetThrottler();
 
-    const u32 input_sample_rate = (target_speed == 0.0f || !g_settings.audio_resampling) ?
-                                    SPU::SAMPLE_RATE :
-                                    static_cast<u32>(static_cast<float>(SPU::SAMPLE_RATE) * target_speed);
-    Log_InfoPrintf("Audio input sample rate: %u hz", input_sample_rate);
-
     AudioStream* stream = g_spu.GetOutputStream();
-    stream->SetInputSampleRate(input_sample_rate);
-    stream->SetWaitForBufferFill(true);
-
     if (g_settings.audio_fast_forward_volume != g_settings.audio_output_volume)
       stream->SetOutputVolume(GetAudioOutputVolume());
 
-    stream->SetSync(audio_sync_enabled);
+    // Adjust nominal rate when resampling, or syncing to host.
-    if (audio_sync_enabled)
+    const bool rate_adjust =
-      stream->EmptyBuffers();
+      (syncing_to_host || g_settings.audio_stretch_mode == AudioStretchMode::Resample) && target_speed > 0.0f;
+    stream->SetNominalRate(rate_adjust ? target_speed : 1.0f);
+
+    // stream->SetSync(audio_sync_enabled);
+    // if (audio_sync_enabled)
+    // stream->EmptyBuffer();
   }
 
   g_host_display->SetDisplayMaxFPS(max_display_fps);
@@ -3034,8 +3022,7 @@ void System::CheckForSettingsChanges(const Settings& old_settings)
       UpdateOverclock();
     }
 
-    if (g_settings.audio_backend != old_settings.audio_backend ||
+    if (g_settings.audio_backend != old_settings.audio_backend)
-        g_settings.audio_buffer_size != old_settings.audio_buffer_size)
     {
       if (g_settings.audio_backend != old_settings.audio_backend)
       {
@@ -3044,7 +3031,15 @@ void System::CheckForSettingsChanges(const Settings& old_settings)
       }
 
       g_spu.RecreateOutputStream();
-      g_spu.GetOutputStream()->PauseOutput(IsPaused());
+    }
+    if (g_settings.audio_stretch_mode != old_settings.audio_stretch_mode)
+      g_spu.GetOutputStream()->SetStretchMode(g_settings.audio_stretch_mode);
+    if (g_settings.audio_buffer_ms != old_settings.audio_buffer_ms ||
+        g_settings.audio_output_latency_ms != old_settings.audio_output_latency_ms ||
+        g_settings.audio_stretch_mode != old_settings.audio_stretch_mode)
+    {
+      g_spu.RecreateOutputStream();
+      UpdateSpeedLimiterState();
     }
 
     if (g_settings.emulation_speed != old_settings.emulation_speed)
@@ -3169,7 +3164,6 @@ void System::CheckForSettingsChanges(const Settings& old_settings)
     g_dma.SetHaltTicks(g_settings.dma_halt_ticks);
 
     if (g_settings.audio_backend != old_settings.audio_backend ||
-        g_settings.audio_buffer_size != old_settings.audio_buffer_size ||
         g_settings.video_sync_enabled != old_settings.video_sync_enabled ||
         g_settings.audio_sync_enabled != old_settings.audio_sync_enabled ||
         g_settings.increase_timer_resolution != old_settings.increase_timer_resolution ||
@@ -3177,7 +3171,6 @@ void System::CheckForSettingsChanges(const Settings& old_settings)
         g_settings.fast_forward_speed != old_settings.fast_forward_speed ||
         g_settings.display_max_fps != old_settings.display_max_fps ||
         g_settings.display_all_frames != old_settings.display_all_frames ||
-        g_settings.audio_resampling != old_settings.audio_resampling ||
         g_settings.sync_to_host_refresh_rate != old_settings.sync_to_host_refresh_rate)
     {
       UpdateSpeedLimiterState();

src/duckstation-qt/audiosettingswidget.cpp

@@ -20,14 +20,24 @@ AudioSettingsWidget::AudioSettingsWidget(SettingsDialog* dialog, QWidget* parent
   SettingWidgetBinder::BindWidgetToEnumSetting(sif, m_ui.audioBackend, "Audio", "Backend", &Settings::ParseAudioBackend,
                                                &Settings::GetAudioBackendName, Settings::DEFAULT_AUDIO_BACKEND);
   SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.syncToOutput, "Audio", "Sync", true);
-  SettingWidgetBinder::BindWidgetToIntSetting(sif, m_ui.bufferSize, "Audio", "BufferSize",
+  SettingWidgetBinder::BindWidgetToEnumSetting(sif, m_ui.stretchMode, "Audio", "StretchMode",
-                                              Settings::DEFAULT_AUDIO_BUFFER_SIZE);
+                                               &AudioStream::ParseStretchMode, &AudioStream::GetStretchModeName,
+                                               Settings::DEFAULT_AUDIO_STRETCH_MODE);
+  SettingWidgetBinder::BindWidgetToIntSetting(sif, m_ui.bufferMS, "Audio", "BufferMS",
+                                              Settings::DEFAULT_AUDIO_BUFFER_MS);
+  SettingWidgetBinder::BindWidgetToIntSetting(sif, m_ui.outputLatencyMS, "Audio", "OutputLatencyMS",
+                                              Settings::DEFAULT_AUDIO_OUTPUT_LATENCY_MS);
   SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.startDumpingOnBoot, "Audio", "DumpOnBoot", false);
   SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.muteCDAudio, "CDROM", "MuteCDAudio", false);
-  SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.resampling, "Audio", "Resampling", true);
 
-  connect(m_ui.bufferSize, &QSlider::valueChanged, this, &AudioSettingsWidget::updateBufferingLabel);
+  m_ui.outputLatencyMinimal->setChecked(m_ui.outputLatencyMS->value() == 0);
-  updateBufferingLabel();
+  m_ui.outputLatencyMS->setEnabled(m_ui.outputLatencyMinimal->isChecked());
+  m_ui.driver->setEnabled(false);
+
+  connect(m_ui.bufferMS, &QSlider::valueChanged, this, &AudioSettingsWidget::updateLatencyLabel);
+  connect(m_ui.outputLatencyMS, &QSlider::valueChanged, this, &AudioSettingsWidget::updateLatencyLabel);
+  connect(m_ui.outputLatencyMinimal, &QCheckBox::toggled, this, &AudioSettingsWidget::onMinimalOutputLatencyChecked);
+  updateLatencyLabel();
 
   // for per-game, just use the normal path, since it needs to re-read/apply
   if (!dialog->isPerGameSettings())
@@ -53,7 +63,7 @@ AudioSettingsWidget::AudioSettingsWidget(SettingsDialog* dialog, QWidget* parent
        "lowest latency, if you encounter issues, try the SDL backend. The null backend disables all host audio "
        "output."));
   dialog->registerWidgetHelp(
-    m_ui.bufferSize, tr("Buffer Size"), QStringLiteral("2048"),
+    m_ui.outputLatencyMS, tr("Output Latency"), QStringLiteral("50 ms"),
     tr("The buffer size determines the size of the chunks of audio which will be pulled by the "
        "host. Smaller values reduce the output latency, but may cause hitches if the emulation "
        "speed is inconsistent. Note that the Cubeb backend uses smaller chunks regardless of "
@@ -75,27 +85,31 @@ AudioSettingsWidget::AudioSettingsWidget(SettingsDialog* dialog, QWidget* parent
                              tr("Forcibly mutes both CD-DA and XA audio from the CD-ROM. Can be used to disable "
                                 "background music in some games."));
   dialog->registerWidgetHelp(
-    m_ui.resampling, tr("Resampling"), tr("Checked"),
+    m_ui.stretchMode, tr("Stretch Mode"), tr("Time Stretching"),
-    tr("When running outside of 100% speed, resamples audio from the target speed instead of dropping frames. Produces "
+    tr("When running outside of 100% speed, adjusts the tempo on audio instead of dropping frames. Produces "
        "much nicer fast forward/slowdown audio at a small cost to performance."));
 }
 
 AudioSettingsWidget::~AudioSettingsWidget() = default;
 
-void AudioSettingsWidget::updateBufferingLabel()
+void AudioSettingsWidget::updateLatencyLabel()
 {
-  constexpr float step = 128;
+  const u32 output_latency_ms = static_cast<u32>(m_ui.outputLatencyMS->value());
-  const u32 actual_buffer_size =
+  const u32 output_latency_frames = AudioStream::GetBufferSizeForMS(SPU::SAMPLE_RATE, output_latency_ms);
-    static_cast<u32>(std::round(static_cast<float>(m_ui.bufferSize->value()) / step) * step);
+  const u32 buffer_ms = static_cast<u32>(m_ui.bufferMS->value());
-  if (static_cast<u32>(m_ui.bufferSize->value()) != actual_buffer_size)
+  const u32 buffer_frames = AudioStream::GetBufferSizeForMS(SPU::SAMPLE_RATE, buffer_ms);
+  if (output_latency_ms > 0)
   {
-    m_ui.bufferSize->setValue(static_cast<int>(actual_buffer_size));
+    m_ui.bufferingLabel->setText(tr("Maximum Latency: %1 frames / %2 ms (%3ms buffer + %5ms output)")
-    return;
+                                   .arg(buffer_frames + output_latency_frames)
+                                   .arg(buffer_ms + output_latency_ms)
+                                   .arg(buffer_ms)
+                                   .arg(output_latency_ms));
+  }
+  else
+  {
+    m_ui.bufferingLabel->setText(tr("Maximum Latency: %1 frames / %2 ms").arg(buffer_frames).arg(buffer_ms));
   }
-
-  const float max_latency = AudioStream::GetMaxLatency(SPU::SAMPLE_RATE, actual_buffer_size);
-  m_ui.bufferingLabel->setText(tr("Maximum Latency: %n frames (%1ms)", "", actual_buffer_size)
-                                 .arg(static_cast<double>(max_latency) * 1000.0, 0, 'f', 2));
 }
 
 void AudioSettingsWidget::updateVolumeLabel()
@@ -104,9 +118,21 @@ void AudioSettingsWidget::updateVolumeLabel()
   m_ui.fastForwardVolumeLabel->setText(tr("%1%").arg(m_ui.fastForwardVolume->value()));
 }
 
+void AudioSettingsWidget::onMinimalOutputLatencyChecked(bool new_value)
+{
+  const u32 value = new_value ? 0u : Settings::DEFAULT_AUDIO_OUTPUT_LATENCY_MS;
+  m_dialog->setIntSettingValue("Audio", "OutputLatencyMS", value);
+  QSignalBlocker sb(m_ui.outputLatencyMS);
+  m_ui.outputLatencyMS->setValue(value);
+  m_ui.outputLatencyMS->setEnabled(!new_value);
+  updateLatencyLabel();
+}
+
 void AudioSettingsWidget::onOutputVolumeChanged(int new_value)
 {
-  m_dialog->setIntSettingValue("Audio", "OutputVolume", new_value);
+  // only called for base settings
+  DebugAssert(!m_dialog->isPerGameSettings());
+  Host::SetBaseIntSettingValue("Audio", "OutputVolume", new_value);
   g_emu_thread->setAudioOutputVolume(new_value, m_ui.fastForwardVolume->value());
 
   updateVolumeLabel();
@@ -114,7 +140,9 @@ void AudioSettingsWidget::onOutputVolumeChanged(int new_value)
 
 void AudioSettingsWidget::onFastForwardVolumeChanged(int new_value)
 {
-  m_dialog->setIntSettingValue("Audio", "FastForwardVolume", new_value);
+  // only called for base settings
+  DebugAssert(!m_dialog->isPerGameSettings());
+  Host::SetBaseIntSettingValue("Audio", "FastForwardVolume", new_value);
   g_emu_thread->setAudioOutputVolume(m_ui.volume->value(), new_value);
 
   updateVolumeLabel();
@@ -122,7 +150,10 @@ void AudioSettingsWidget::onFastForwardVolumeChanged(int new_value)
 
 void AudioSettingsWidget::onOutputMutedChanged(int new_state)
 {
+  // only called for base settings
+  DebugAssert(!m_dialog->isPerGameSettings());
+
   const bool muted = (new_state != 0);
-  m_dialog->setBoolSettingValue("Audio", "OutputMuted", muted);
+  Host::SetBaseBoolSettingValue("Audio", "OutputMuted", muted);
   g_emu_thread->setAudioOutputMuted(muted);
 }

src/duckstation-qt/audiosettingswidget.h

@@ -15,8 +15,9 @@ public:
   ~AudioSettingsWidget();
 
 private Q_SLOTS:
-  void updateBufferingLabel();
+  void updateLatencyLabel();
   void updateVolumeLabel();
+  void onMinimalOutputLatencyChecked(bool new_value);
   void onOutputVolumeChanged(int new_value);
   void onFastForwardVolumeChanged(int new_value);
   void onOutputMutedChanged(int new_state);

src/duckstation-qt/audiosettingswidget.ui

@@ -6,8 +6,8 @@
    <rect>
     <x>0</x>
     <y>0</y>
-    <width>502</width>
+    <width>516</width>
-    <height>312</height>
+    <height>435</height>
    </rect>
   </property>
   <property name="windowTitle">
@@ -31,7 +31,21 @@
      <property name="title">
       <string>Configuration</string>
      </property>
-     <layout class="QFormLayout" name="formLayout">
+     <layout class="QGridLayout" name="gridLayout">
+      <item row="3" column="0">
+       <widget class="QLabel" name="label_2">
+        <property name="text">
+         <string>Buffer Size:</string>
+        </property>
+       </widget>
+      </item>
+      <item row="7" column="0" colspan="2">
+       <widget class="QCheckBox" name="startDumpingOnBoot">
+        <property name="text">
+         <string>Start Dumping On Boot</string>
+        </property>
+       </widget>
+      </item>
       <item row="0" column="0">
        <widget class="QLabel" name="label">
         <property name="text">
@@ -39,55 +53,87 @@
         </property>
        </widget>
       </item>
+      <item row="4" column="1">
+       <layout class="QHBoxLayout" name="horizontalLayout_3">
+        <item>
+         <widget class="QSlider" name="outputLatencyMS">
+          <property name="maximum">
+           <number>500</number>
+          </property>
+          <property name="orientation">
+           <enum>Qt::Horizontal</enum>
+          </property>
+          <property name="tickPosition">
+           <enum>QSlider::TicksBothSides</enum>
+          </property>
+          <property name="tickInterval">
+           <number>20</number>
+          </property>
+         </widget>
+        </item>
+        <item>
+         <widget class="QCheckBox" name="outputLatencyMinimal">
+          <property name="text">
+           <string>Minimal</string>
+          </property>
+         </widget>
+        </item>
+       </layout>
+      </item>
+      <item row="1" column="1">
+       <widget class="QComboBox" name="driver"/>
+      </item>
+      <item row="2" column="1">
+       <widget class="QComboBox" name="stretchMode">
+        <item>
+         <property name="text">
+          <string>Off (Noisy)</string>
+         </property>
+        </item>
+        <item>
+         <property name="text">
+          <string>Resampling (Pitch Shift)</string>
+         </property>
+        </item>
+        <item>
+         <property name="text">
+          <string>Time Stretch (Tempo Change, Best Sound)</string>
+         </property>
+        </item>
+       </widget>
+      </item>
+      <item row="4" column="0">
+       <widget class="QLabel" name="label_5">
+        <property name="text">
+         <string>Output Latency:</string>
+        </property>
+       </widget>
+      </item>
       <item row="0" column="1">
        <widget class="QComboBox" name="audioBackend"/>
       </item>
       <item row="1" column="0">
-       <widget class="QLabel" name="label_2">
+       <widget class="QLabel" name="label_6">
         <property name="text">
-         <string>Buffer Size:</string>
+         <string>Driver:</string>
-        </property>
-       </widget>
-      </item>
-      <item row="1" column="1">
-       <widget class="QSlider" name="bufferSize">
-        <property name="minimum">
-         <number>1024</number>
-        </property>
-        <property name="maximum">
-         <number>8192</number>
-        </property>
-        <property name="singleStep">
-         <number>128</number>
-        </property>
-        <property name="pageStep">
-         <number>1024</number>
-        </property>
-        <property name="orientation">
-         <enum>Qt::Horizontal</enum>
-        </property>
-        <property name="tickPosition">
-         <enum>QSlider::TicksBothSides</enum>
-        </property>
-        <property name="tickInterval">
-         <number>1024</number>
         </property>
        </widget>
       </item>
       <item row="2" column="0">
-       <spacer name="horizontalSpacer">
+       <widget class="QLabel" name="label_7">
-        <property name="orientation">
+        <property name="text">
-         <enum>Qt::Horizontal</enum>
+         <string>Stretch Mode:</string>
         </property>
-        <property name="sizeHint" stdset="0">
+       </widget>
-         <size>
-          <width>40</width>
-          <height>20</height>
-         </size>
-        </property>
-       </spacer>
       </item>
-      <item row="2" column="1">
+      <item row="6" column="0" colspan="2">
+       <widget class="QCheckBox" name="syncToOutput">
+        <property name="text">
+         <string>Sync To Output</string>
+        </property>
+       </widget>
+      </item>
+      <item row="5" column="0" colspan="2">
        <widget class="QLabel" name="bufferingLabel">
         <property name="text">
          <string>Maximum latency: 0 frames (0.00ms)</string>
@@ -97,24 +143,31 @@
         </property>
        </widget>
       </item>
-      <item row="3" column="0" colspan="2">
+      <item row="3" column="1">
-       <widget class="QCheckBox" name="syncToOutput">
+       <widget class="QSlider" name="bufferMS">
-        <property name="text">
+        <property name="minimum">
-         <string>Sync To Output</string>
+         <number>15</number>
         </property>
-       </widget>
+        <property name="maximum">
-      </item>
+         <number>500</number>
-      <item row="4" column="0" colspan="2">
-       <widget class="QCheckBox" name="resampling">
-        <property name="text">
-         <string>Resampling</string>
         </property>
-       </widget>
+        <property name="singleStep">
-      </item>
+         <number>1</number>
-      <item row="5" column="0" colspan="2">
+        </property>
-       <widget class="QCheckBox" name="startDumpingOnBoot">
+        <property name="pageStep">
-        <property name="text">
+         <number>5</number>
-         <string>Start Dumping On Boot</string>
+        </property>
+        <property name="value">
+         <number>50</number>
+        </property>
+        <property name="orientation">
+         <enum>Qt::Horizontal</enum>
+        </property>
+        <property name="tickPosition">
+         <enum>QSlider::TicksBothSides</enum>
+        </property>
+        <property name="tickInterval">
+         <number>20</number>
         </property>
        </widget>
       </item>

src/frontend-common/common_host.cpp

@@ -66,14 +66,6 @@
 #include <mmsystem.h>
 #endif
 
-namespace FrontendCommon {
-
-#ifdef _WIN32
-std::unique_ptr<AudioStream> CreateXAudio2AudioStream();
-#endif
-
-} // namespace FrontendCommon
-
 Log_SetChannel(CommonHostInterface);
 
 namespace CommonHost {
@@ -148,26 +140,27 @@ void CommonHost::ReleaseHostDisplayResources()
   //
 }
 
-std::unique_ptr<AudioStream> Host::CreateAudioStream(AudioBackend backend)
+std::unique_ptr<AudioStream> Host::CreateAudioStream(AudioBackend backend, u32 sample_rate, u32 channels, u32 buffer_ms,
+                                                     u32 latency_ms, AudioStretchMode stretch)
 {
   switch (backend)
   {
     case AudioBackend::Null:
-      return AudioStream::CreateNullAudioStream();
+      return AudioStream::CreateNullStream(sample_rate, channels, buffer_ms);
 
 #ifndef _UWP
     case AudioBackend::Cubeb:
-      return CubebAudioStream::Create();
+      return CommonHost::CreateCubebAudioStream(sample_rate, channels, buffer_ms, latency_ms, stretch);
 #endif
 
 #ifdef _WIN32
     case AudioBackend::XAudio2:
-      return FrontendCommon::CreateXAudio2AudioStream();
+      return CommonHost::CreateXAudio2Stream(sample_rate, channels, buffer_ms, latency_ms, stretch);
 #endif
 
 #ifdef WITH_SDL2
     case AudioBackend::SDL:
-      return SDLAudioStream::Create();
+      return CommonHost::CreateSDLAudioStream(sample_rate, channels, buffer_ms, latency_ms, stretch);
 #endif
 
     default:
@@ -927,7 +920,7 @@ DEFINE_HOTKEY("AudioMute", TRANSLATABLE("Hotkeys", "Audio"), TRANSLATABLE("Hotke
   {
     g_settings.audio_output_muted = !g_settings.audio_output_muted;
     const s32 volume = System::GetAudioOutputVolume();
-    g_spu.GetOutputStream()->SetOutputVolume(volume);
+    // g_spu.GetOutputStream()->SetOutputVolume(volume);
     if (g_settings.audio_output_muted)
     {
       Host::AddKeyedOSDMessage("AudioControlHotkey", Host::TranslateStdString("OSDMessage", "Volume: Muted"), 2.0f);
@@ -959,7 +952,7 @@ DEFINE_HOTKEY("AudioVolumeUp", TRANSLATABLE("Hotkeys", "Audio"), TRANSLATABLE("H
     const s32 volume = std::min<s32>(System::GetAudioOutputVolume() + 10, 100);
     g_settings.audio_output_volume = volume;
     g_settings.audio_fast_forward_volume = volume;
-    g_spu.GetOutputStream()->SetOutputVolume(volume);
+    // g_spu.GetOutputStream()->SetOutputVolume(volume);
     Host::AddKeyedFormattedOSDMessage("AudioControlHotkey", 2.0f, Host::TranslateString("OSDMessage", "Volume: %d%%"),
                                       volume);
   }
@@ -973,7 +966,7 @@ DEFINE_HOTKEY("AudioVolumeDown", TRANSLATABLE("Hotkeys", "Audio"), TRANSLATABLE(
                   const s32 volume = std::max<s32>(System::GetAudioOutputVolume() - 10, 0);
                   g_settings.audio_output_volume = volume;
                   g_settings.audio_fast_forward_volume = volume;
-                  g_spu.GetOutputStream()->SetOutputVolume(volume);
+                  // g_spu.GetOutputStream()->SetOutputVolume(volume);
                   Host::AddKeyedFormattedOSDMessage("AudioControlHotkey", 2.0f,
                                                     Host::TranslateString("OSDMessage", "Volume: %d%%"), volume);
                 }

src/frontend-common/common_host.h

@@ -1,9 +1,13 @@
 #pragma once
 #include "core/system.h"
+#include <memory>
 #include <mutex>
 
 class SettingsInterface;
 
+class AudioStream;
+enum class AudioStretchMode : u8;
+
 namespace CommonHost {
 /// Initializes configuration.
 void UpdateLogSettings();
@@ -25,6 +29,19 @@ void OnGameChanged(const std::string& disc_path, const std::string& game_serial,
 void PumpMessagesOnCPUThread();
 bool CreateHostDisplayResources();
 void ReleaseHostDisplayResources();
+
+#ifdef _WIN32
+std::unique_ptr<AudioStream> CreateXAudio2Stream(u32 sample_rate, u32 channels, u32 buffer_ms, u32 latency_ms,
+                                                 AudioStretchMode stretch);
+#endif
+#ifdef WITH_SDL2
+std::unique_ptr<AudioStream> CreateSDLAudioStream(u32 sample_rate, u32 channels, u32 buffer_ms, u32 latency_ms,
+                                                  AudioStretchMode stretch);
+#endif
+#ifndef _UWP
+std::unique_ptr<AudioStream> CreateCubebAudioStream(u32 sample_rate, u32 channels, u32 buffer_ms, u32 latency_ms,
+                                                    AudioStretchMode stretch);
+#endif
 } // namespace CommonHost
 
 namespace ImGuiManager {

src/frontend-common/cubeb_audio_stream.cpp

@@ -1,6 +1,11 @@
 #include "cubeb_audio_stream.h"
 #include "common/assert.h"
 #include "common/log.h"
+#include "common/string_util.h"
+#include "common_host.h"
+#include "core/host.h"
+#include "core/host_settings.h"
+#include "cubeb/cubeb.h"
 Log_SetChannel(CubebAudioStream);
 
 #ifdef _WIN32
@@ -9,154 +14,188 @@ Log_SetChannel(CubebAudioStream);
 #pragma comment(lib, "Ole32.lib")
 #endif
 
-CubebAudioStream::CubebAudioStream() = default;
+static void StateCallback(cubeb_stream* stream, void* user_ptr, cubeb_state state);
+
+CubebAudioStream::CubebAudioStream(u32 sample_rate, u32 channels, u32 buffer_ms, AudioStretchMode stretch)
+  : AudioStream(sample_rate, channels, buffer_ms, stretch)
+{
+}
 
 CubebAudioStream::~CubebAudioStream()
 {
-  if (IsOpen())
+  DestroyContextAndStream();
-    CubebAudioStream::CloseDevice();
 }
 
-bool CubebAudioStream::OpenDevice()
+void CubebAudioStream::LogCallback(const char* fmt, ...)
 {
-  Assert(!IsOpen());
+  std::va_list ap;
+  va_start(ap, fmt);
+  std::string msg(StringUtil::StdStringFromFormatV(fmt, ap));
+  va_end(ap);
+  Log_DevPrintf("(Cubeb): %s", msg.c_str());
+}
 
+void CubebAudioStream::DestroyContextAndStream()
+{
+  if (stream)
+  {
+    cubeb_stream_stop(stream);
+    cubeb_stream_destroy(stream);
+    stream = nullptr;
+  }
+
+  if (m_context)
+  {
+    cubeb_destroy(m_context);
+    m_context = nullptr;
+  }
+
+#ifdef _WIN32
+  if (m_com_initialized_by_us)
+  {
+    CoUninitialize();
+    m_com_initialized_by_us = false;
+  }
+#endif
+}
+
+bool CubebAudioStream::Initialize(u32 latency_ms)
+{
 #ifdef _WIN32
   HRESULT hr = CoInitializeEx(nullptr, COINIT_MULTITHREADED);
   m_com_initialized_by_us = SUCCEEDED(hr);
-  if (FAILED(hr) && hr != RPC_E_CHANGED_MODE && hr != S_FALSE)
+  if (FAILED(hr) && hr != RPC_E_CHANGED_MODE)
   {
-    Log_ErrorPrintf("Failed to initialize COM");
+    Host::ReportErrorAsync("Error", "Failed to initialize COM for Cubeb");
     return false;
   }
 #endif
 
-  int rv = cubeb_init(&m_cubeb_context, "DuckStation", nullptr);
+  cubeb_set_log_callback(CUBEB_LOG_NORMAL, LogCallback);
+
+  std::string backend(Host::GetStringSettingValue("Audio", "CubebBackend"));
+
+  int rv = cubeb_init(&m_context, "DuckStation", backend.empty() ? nullptr : backend.c_str());
   if (rv != CUBEB_OK)
   {
-    Log_ErrorPrintf("Could not initialize cubeb context: %d", rv);
+    Host::ReportFormattedErrorAsync("Error", "Could not initialize cubeb context: %d", rv);
     return false;
   }
 
   cubeb_stream_params params = {};
   params.format = CUBEB_SAMPLE_S16LE;
-  params.rate = m_output_sample_rate;
+  params.rate = m_sample_rate;
   params.channels = m_channels;
   params.layout = CUBEB_LAYOUT_UNDEFINED;
-  params.prefs = CUBEB_STREAM_PREF_PERSIST;
+  params.prefs = CUBEB_STREAM_PREF_NONE;
 
-  u32 latency_frames = 0;
+  u32 latency_frames = GetBufferSizeForMS(m_sample_rate, (latency_ms == 0) ? m_buffer_ms : latency_ms);
-  rv = cubeb_get_min_latency(m_cubeb_context, &params, &latency_frames);
+  u32 min_latency_frames = 0;
+  rv = cubeb_get_min_latency(m_context, &params, &min_latency_frames);
   if (rv == CUBEB_ERROR_NOT_SUPPORTED)
   {
-    Log_WarningPrintf("Cubeb backend does not support latency queries, using buffer size of %u.", m_buffer_size);
+    Log_DevPrintf("(Cubeb) Cubeb backend does not support latency queries, using latency of %d ms (%u frames).",
-    latency_frames = m_buffer_size;
+                  m_buffer_ms, latency_frames);
   }
   else
   {
     if (rv != CUBEB_OK)
     {
-      Log_ErrorPrintf("Could not get minimum latency: %d", rv);
+      Log_ErrorPrintf("(Cubeb) Could not get minimum latency: %d", rv);
-      DestroyContext();
+      DestroyContextAndStream();
       return false;
     }
 
-    Log_InfoPrintf("Minimum latency in frames: %u", latency_frames);
+    const u32 minimum_latency_ms = GetMSForBufferSize(m_sample_rate, min_latency_frames);
-    if (latency_frames > m_buffer_size)
+    Log_DevPrintf("(Cubeb) Minimum latency: %u ms (%u audio frames)", minimum_latency_ms, min_latency_frames);
+    if (latency_ms == 0)
     {
-      Log_WarningPrintf("Minimum latency is above buffer size: %u vs %u, adjusting to compensate.", latency_frames,
+      // use minimum
-                        m_buffer_size);
+      latency_frames = min_latency_frames;
-
-      if (!SetBufferSize(latency_frames))
-      {
-        Log_ErrorPrintf("Failed to set new buffer size of %u frames", latency_frames);
-        DestroyContext();
-        return false;
-      }
     }
-    else
+    else if (minimum_latency_ms > latency_ms)
     {
-      latency_frames = m_buffer_size;
+      Log_WarningPrintf("(Cubeb) Minimum latency is above requested latency: %u vs %u, adjusting to compensate.",
+                        min_latency_frames, latency_frames);
+      latency_frames = min_latency_frames;
     }
   }
 
-  char stream_name[32];
+  BaseInitialize();
-  std::snprintf(stream_name, sizeof(stream_name), "AudioStream_%p", this);
+  m_volume = 100;
+  m_paused = false;
 
-  rv = cubeb_stream_init(m_cubeb_context, &m_cubeb_stream, stream_name, nullptr, nullptr, nullptr, &params,
+  char stream_name[32];
-                         latency_frames, DataCallback, StateCallback, this);
+  std::snprintf(stream_name, sizeof(stream_name), "%p", this);
+
+  rv = cubeb_stream_init(m_context, &stream, stream_name, nullptr, nullptr, nullptr, &params, latency_frames,
+                         &CubebAudioStream::DataCallback, StateCallback, this);
   if (rv != CUBEB_OK)
   {
-    Log_ErrorPrintf("Could not create stream: %d", rv);
+    Log_ErrorPrintf("(Cubeb) Could not create stream: %d", rv);
-    DestroyContext();
+    DestroyContextAndStream();
+    return false;
+  }
+
+  rv = cubeb_stream_start(stream);
+  if (rv != CUBEB_OK)
+  {
+    Log_ErrorPrintf("(Cubeb) Could not start stream: %d", rv);
+    DestroyContextAndStream();
     return false;
   }
 
-  cubeb_stream_set_volume(m_cubeb_stream, static_cast<float>(m_output_volume) / 100.0f);
   return true;
 }
 
-void CubebAudioStream::PauseDevice(bool paused)
+void StateCallback(cubeb_stream* stream, void* user_ptr, cubeb_state state)
 {
-  if (paused == m_paused)
+  // noop
+}
+
+long CubebAudioStream::DataCallback(cubeb_stream* stm, void* user_ptr, const void* input_buffer, void* output_buffer,
+                                    long nframes)
+{
+  static_cast<CubebAudioStream*>(user_ptr)->ReadFrames(static_cast<s16*>(output_buffer), static_cast<u32>(nframes));
+  return nframes;
+}
+
+void CubebAudioStream::SetPaused(bool paused)
+{
+  if (paused == m_paused || !stream)
     return;
 
-  int rv = paused ? cubeb_stream_stop(m_cubeb_stream) : cubeb_stream_start(m_cubeb_stream);
+  const int rv = paused ? cubeb_stream_stop(stream) : cubeb_stream_start(stream);
   if (rv != CUBEB_OK)
   {
-    Log_ErrorPrintf("cubeb_stream_%s failed: %d", paused ? "stop" : "start", rv);
+    Log_ErrorPrintf("Could not %s stream: %d", paused ? "pause" : "resume", rv);
     return;
   }
 
   m_paused = paused;
 }
 
-void CubebAudioStream::CloseDevice()
-{
-  Assert(IsOpen());
-
-  if (!m_paused)
-  {
-    cubeb_stream_stop(m_cubeb_stream);
-    m_paused = true;
-  }
-
-  cubeb_stream_destroy(m_cubeb_stream);
-  m_cubeb_stream = nullptr;
-
-  DestroyContext();
-}
-
-long CubebAudioStream::DataCallback(cubeb_stream* stm, void* user_ptr, const void* input_buffer, void* output_buffer,
-                                    long nframes)
-{
-  CubebAudioStream* const this_ptr = static_cast<CubebAudioStream*>(user_ptr);
-  this_ptr->ReadFrames(reinterpret_cast<SampleType*>(output_buffer), static_cast<u32>(nframes), false);
-  return nframes;
-}
-
-void CubebAudioStream::StateCallback(cubeb_stream* stream, void* user_ptr, cubeb_state state) {}
-
-void CubebAudioStream::FramesAvailable() {}
-
 void CubebAudioStream::SetOutputVolume(u32 volume)
 {
-  AudioStream::SetOutputVolume(volume);
+  if (volume == m_volume)
-  cubeb_stream_set_volume(m_cubeb_stream, static_cast<float>(m_output_volume) / 100.0f);
+    return;
+
+  int rv = cubeb_stream_set_volume(stream, static_cast<float>(volume) / 100.0f);
+  if (rv != CUBEB_OK)
+  {
+    Log_ErrorPrintf("cubeb_stream_set_volume() failed: %d", rv);
+    return;
+  }
+
+  m_volume = volume;
 }
 
-void CubebAudioStream::DestroyContext()
+std::unique_ptr<AudioStream> CommonHost::CreateCubebAudioStream(u32 sample_rate, u32 channels, u32 buffer_ms,
+                                                                u32 latency_ms, AudioStretchMode stretch)
 {
-  cubeb_destroy(m_cubeb_context);
+  std::unique_ptr<CubebAudioStream> stream(
-  m_cubeb_context = nullptr;
+    std::make_unique<CubebAudioStream>(sample_rate, channels, buffer_ms, stretch));
-
+  if (!stream->Initialize(latency_ms))
-#ifdef _WIN32
+    stream.reset();
-  if (m_com_initialized_by_us)
+  return stream;
-    CoUninitialize();
-#endif
-}
-
-std::unique_ptr<AudioStream> CubebAudioStream::Create()
-{
-  return std::make_unique<CubebAudioStream>();
 }

src/frontend-common/cubeb_audio_stream.h

@@ -1,34 +1,30 @@
 #pragma once
-#include "cubeb/cubeb.h"
 #include "util/audio_stream.h"
 #include <cstdint>
 
-class CubebAudioStream final : public AudioStream
+struct cubeb;
+struct cubeb_stream;
+
+class CubebAudioStream : public AudioStream
 {
 public:
-  CubebAudioStream();
+  CubebAudioStream(u32 sample_rate, u32 channels, u32 buffer_ms, AudioStretchMode stretch);
   ~CubebAudioStream();
 
-  static std::unique_ptr<AudioStream> Create();
+  void SetPaused(bool paused) override;
-
-protected:
-  bool IsOpen() const { return m_cubeb_stream != nullptr; }
-
-  bool OpenDevice() override;
-  void PauseDevice(bool paused) override;
-  void CloseDevice() override;
-  void FramesAvailable() override;
   void SetOutputVolume(u32 volume) override;
 
-  void DestroyContext();
+  bool Initialize(u32 latency_ms);
 
+private:
+  static void LogCallback(const char* fmt, ...);
   static long DataCallback(cubeb_stream* stm, void* user_ptr, const void* input_buffer, void* output_buffer,
                            long nframes);
-  static void StateCallback(cubeb_stream* stream, void* user_ptr, cubeb_state state);
 
-  cubeb* m_cubeb_context = nullptr;
+  void DestroyContextAndStream();
-  cubeb_stream* m_cubeb_stream = nullptr;
+
-  bool m_paused = true;
+  cubeb* m_context = nullptr;
+  cubeb_stream* stream = nullptr;
 
 #ifdef _WIN32
   bool m_com_initialized_by_us = false;

src/frontend-common/fullscreen_ui.cpp

@@ -3152,18 +3152,18 @@ void FullscreenUI::DrawAudioSettingsPage()
                   "The audio backend determines how frames produced by the emulator are submitted to the host.",
                   "Audio", "Backend", Settings::DEFAULT_AUDIO_BACKEND, &Settings::ParseAudioBackend,
                   &Settings::GetAudioBackendName, &Settings::GetAudioBackendDisplayName, AudioBackend::Count);
-  DrawIntRangeSetting("Buffer Size",
+  DrawIntRangeSetting("Latency",
                       "The buffer size determines the size of the chunks of audio which will be pulled by the host.",
-                      "Audio", "BufferSize", Settings::DEFAULT_AUDIO_BUFFER_SIZE, 1024, 8192, "%d Frames");
+                      "Audio", "Latency", Settings::DEFAULT_AUDIO_BUFFER_MS, 10, 500, "%d ms");
 
   DrawToggleSetting("Sync To Output",
                     "Throttles the emulation speed based on the audio backend pulling audio "
                     "frames. Enable to reduce the chances of crackling.",
                     "Audio", "Sync", true);
   DrawToggleSetting(
-    "Resampling",
+    "Time Stretching",
-    "When running outside of 100% speed, resamples audio from the target speed instead of dropping frames.", "Audio",
+    "When running outside of 100% speed, adjusts tempo on audio from the target speed instead of dropping frames.",
-    "Resampling", true);
+    "Audio", "TimeStretching", true);
 
   EndMenuButtons();
 }

src/frontend-common/imgui_overlays.cpp

@@ -13,6 +13,7 @@
 #include "core/host_display.h"
 #include "core/host_settings.h"
 #include "core/settings.h"
+#include "core/spu.h"
 #include "core/system.h"
 #include "fmt/chrono.h"
 #include "fmt/format.h"
@@ -23,6 +24,7 @@
 #include "imgui_internal.h"
 #include "imgui_manager.h"
 #include "input_manager.h"
+#include "util/audio_stream.h"
 #include <atomic>
 #include <chrono>
 #include <cmath>
@@ -172,6 +174,16 @@ void ImGuiManager::DrawPerformanceOverlay()
         FormatProcessorStat(text, System::GetSWThreadUsage(), System::GetSWThreadAverageTime());
         DRAW_LINE(fixed_font, text, IM_COL32(255, 255, 255, 255));
       }
+
+#if 0
+      {
+        AudioStream* stream = g_spu.GetOutputStream();
+        const u32 frames = stream->GetBufferedFramesRelaxed();
+        text.Clear();
+        text.Fmt("Audio: {:<4u}f/{:<3u}ms", frames, AudioStream::GetMSForBufferSize(stream->GetSampleRate(), frames));
+        DRAW_LINE(fixed_font, text, IM_COL32(255, 255, 255, 255));
+      }
+#endif
     }
 
     if (g_settings.display_show_status_indicators)

src/frontend-common/sdl_audio_stream.cpp

@@ -1,11 +1,15 @@
 #include "sdl_audio_stream.h"
 #include "common/assert.h"
 #include "common/log.h"
+#include "common_host.h"
 #include "sdl_initializer.h"
 #include <SDL.h>
 Log_SetChannel(SDLAudioStream);
 
-SDLAudioStream::SDLAudioStream() = default;
+SDLAudioStream::SDLAudioStream(u32 sample_rate, u32 channels, u32 buffer_ms, AudioStretchMode stretch)
+  : AudioStream(sample_rate, channels, buffer_ms, stretch)
+{
+}
 
 SDLAudioStream::~SDLAudioStream()
 {
@@ -13,12 +17,16 @@ SDLAudioStream::~SDLAudioStream()
     SDLAudioStream::CloseDevice();
 }
 
-std::unique_ptr<SDLAudioStream> SDLAudioStream::Create()
+std::unique_ptr<AudioStream> CommonHost::CreateSDLAudioStream(u32 sample_rate, u32 channels, u32 buffer_ms,
+                                                              u32 latency_ms, AudioStretchMode stretch)
 {
-  return std::make_unique<SDLAudioStream>();
+  std::unique_ptr<SDLAudioStream> stream(std::make_unique<SDLAudioStream>(sample_rate, channels, buffer_ms, stretch));
+  if (!stream->OpenDevice(latency_ms))
+    stream.reset();
+  return stream;
 }
 
-bool SDLAudioStream::OpenDevice()
+bool SDLAudioStream::OpenDevice(u32 latency_ms)
 {
   DebugAssert(!IsOpen());
 
@@ -31,22 +39,15 @@ bool SDLAudioStream::OpenDevice()
   }
 
   SDL_AudioSpec spec = {};
-  spec.freq = m_output_sample_rate;
+  spec.freq = m_sample_rate;
   spec.channels = static_cast<Uint8>(m_channels);
   spec.format = AUDIO_S16;
-  spec.samples = static_cast<Uint16>(m_buffer_size);
+  spec.samples = static_cast<Uint16>(GetBufferSizeForMS(m_sample_rate, (latency_ms == 0) ? m_buffer_ms : latency_ms));
   spec.callback = AudioCallback;
   spec.userdata = static_cast<void*>(this);
 
   SDL_AudioSpec obtained_spec = {};
-
+  m_device_id = SDL_OpenAudioDevice(nullptr, 0, &spec, &obtained_spec, SDL_AUDIO_ALLOW_SAMPLES_CHANGE);
-#ifdef SDL_AUDIO_ALLOW_SAMPLES_CHANGE
-  const u32 allowed_change_flags = SDL_AUDIO_ALLOW_SAMPLES_CHANGE;
-#else
-  const u32 allowed_change_flags = 0;
-#endif
-
-  m_device_id = SDL_OpenAudioDevice(nullptr, 0, &spec, &obtained_spec, allowed_change_flags);
   if (m_device_id == 0)
   {
     Log_ErrorPrintf("SDL_OpenAudioDevice() failed: %s", SDL_GetError());
@@ -54,25 +55,23 @@ bool SDLAudioStream::OpenDevice()
     return false;
   }
 
-  if (obtained_spec.samples > spec.samples)
+  Log_DevPrintf("Requested %u frame buffer, got %u frame buffer", spec.samples, obtained_spec.samples);
-  {
-    Log_WarningPrintf("Requested buffer size %u, got buffer size %u. Adjusting to compensate.", spec.samples,
-                      obtained_spec.samples);
 
-    if (!SetBufferSize(obtained_spec.samples))
+  BaseInitialize();
-    {
+  m_volume = 100;
-      Log_ErrorPrintf("Failed to set new buffer size of %u", obtained_spec.samples);
+  m_paused = false;
-      CloseDevice();
+  SDL_PauseAudioDevice(m_device_id, 0);
-      return false;
-    }
-  }
 
   return true;
 }
 
-void SDLAudioStream::PauseDevice(bool paused)
+void SDLAudioStream::SetPaused(bool paused)
 {
+  if (m_paused == paused)
+    return;
+
   SDL_PauseAudioDevice(m_device_id, paused ? 1 : 0);
+  m_paused = paused;
 }
 
 void SDLAudioStream::CloseDevice()
@@ -87,7 +86,13 @@ void SDLAudioStream::AudioCallback(void* userdata, uint8_t* stream, int len)
   SDLAudioStream* const this_ptr = static_cast<SDLAudioStream*>(userdata);
   const u32 num_frames = len / sizeof(SampleType) / this_ptr->m_channels;
 
-  this_ptr->ReadFrames(reinterpret_cast<SampleType*>(stream), num_frames, true);
+  this_ptr->ReadFrames(reinterpret_cast<SampleType*>(stream), num_frames);
 }
 
-void SDLAudioStream::FramesAvailable() {}
+void SDLAudioStream::SetOutputVolume(u32 volume)
+{
+  if (m_volume == volume)
+    return;
+
+  Panic("Fixme");
+}

src/frontend-common/sdl_audio_stream.h

@@ -5,19 +5,18 @@
 class SDLAudioStream final : public AudioStream
 {
 public:
-  SDLAudioStream();
+  SDLAudioStream(u32 sample_rate, u32 channels, u32 buffer_ms, AudioStretchMode stretch);
   ~SDLAudioStream();
 
-  static std::unique_ptr<SDLAudioStream> Create();
+  void SetPaused(bool paused) override;
+  void SetOutputVolume(u32 volume) override;
+
+  bool OpenDevice(u32 latency_ms);
+  void CloseDevice();
 
 protected:
   ALWAYS_INLINE bool IsOpen() const { return (m_device_id != 0); }
 
-  bool OpenDevice() override;
-  void PauseDevice(bool paused) override;
-  void CloseDevice() override;
-  void FramesAvailable() override;
-
   static void AudioCallback(void* userdata, uint8_t* stream, int len);
 
   u32 m_device_id = 0;

src/frontend-common/xaudio2_audio_stream.cpp

@@ -1,6 +1,7 @@
 #include "xaudio2_audio_stream.h"
 #include "common/assert.h"
 #include "common/log.h"
+#include "common_host.h"
 #include <VersionHelpers.h>
 #include <xaudio2.h>
 Log_SetChannel(XAudio2AudioStream);
@@ -9,12 +10,15 @@ Log_SetChannel(XAudio2AudioStream);
 #pragma comment(lib, "xaudio2.lib")
 #endif
 
-XAudio2AudioStream::XAudio2AudioStream() = default;
+XAudio2AudioStream::XAudio2AudioStream(u32 sample_rate, u32 channels, u32 buffer_ms, AudioStretchMode stretch)
+  : AudioStream(sample_rate, channels, buffer_ms, stretch)
+{
+}
 
 XAudio2AudioStream::~XAudio2AudioStream()
 {
   if (IsOpen())
-    XAudio2AudioStream::CloseDevice();
+    CloseDevice();
 
 #if WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_DESKTOP)
   if (m_xaudio2_library)
@@ -25,8 +29,20 @@ XAudio2AudioStream::~XAudio2AudioStream()
 #endif
 }
 
-bool XAudio2AudioStream::Initialize()
+std::unique_ptr<AudioStream> CommonHost::CreateXAudio2Stream(u32 sample_rate, u32 channels, u32 buffer_ms,
+                                                             u32 latency_ms, AudioStretchMode stretch)
 {
+  std::unique_ptr<XAudio2AudioStream> stream(
+    std::make_unique<XAudio2AudioStream>(sample_rate, channels, buffer_ms, stretch));
+  if (!stream->OpenDevice(latency_ms))
+    stream.reset();
+  return stream;
+}
+
+bool XAudio2AudioStream::OpenDevice(u32 latency_ms)
+{
+  DebugAssert(!IsOpen());
+
 #if WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_DESKTOP)
   m_xaudio2_library = LoadLibraryW(XAUDIO2_DLL_W);
   if (!m_xaudio2_library)
@@ -36,13 +52,6 @@ bool XAudio2AudioStream::Initialize()
   }
 #endif
 
-  return true;
-}
-
-bool XAudio2AudioStream::OpenDevice()
-{
-  DebugAssert(!IsOpen());
-
   HRESULT hr;
 #if WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_DESKTOP)
   using PFNXAUDIO2CREATE =
@@ -70,7 +79,7 @@ bool XAudio2AudioStream::OpenDevice()
     return false;
   }
 
-  hr = m_xaudio->CreateMasteringVoice(&m_mastering_voice, m_channels, m_output_sample_rate, 0, nullptr);
+  hr = m_xaudio->CreateMasteringVoice(&m_mastering_voice, m_channels, m_sample_rate, 0, nullptr);
   if (FAILED(hr))
   {
     Log_ErrorPrintf("CreateMasteringVoice() failed: %08X", hr);
@@ -79,10 +88,10 @@ bool XAudio2AudioStream::OpenDevice()
 
   WAVEFORMATEX wf = {};
   wf.cbSize = sizeof(wf);
-  wf.nAvgBytesPerSec = m_output_sample_rate * m_channels * sizeof(s16);
+  wf.nAvgBytesPerSec = m_sample_rate * m_channels * sizeof(s16);
   wf.nBlockAlign = static_cast<WORD>(sizeof(s16) * m_channels);
   wf.nChannels = static_cast<WORD>(m_channels);
-  wf.nSamplesPerSec = m_output_sample_rate;
+  wf.nSamplesPerSec = m_sample_rate;
   wf.wBitsPerSample = sizeof(s16) * 8;
   wf.wFormatTag = WAVE_FORMAT_PCM;
   hr = m_xaudio->CreateSourceVoice(&m_source_voice, &wf, 0, 1.0f, this);
@@ -99,13 +108,27 @@ bool XAudio2AudioStream::OpenDevice()
     return false;
   }
 
+  m_enqueue_buffer_size = std::max<u32>(INTERNAL_BUFFER_SIZE, GetBufferSizeForMS(m_sample_rate, latency_ms));
+  Log_DevPrintf("Allocating %u buffers of %u frames", NUM_BUFFERS, m_enqueue_buffer_size);
   for (u32 i = 0; i < NUM_BUFFERS; i++)
-    m_buffers[i] = std::make_unique<SampleType[]>(m_buffer_size * m_channels);
+    m_enqueue_buffers[i] = std::make_unique<SampleType[]>(m_enqueue_buffer_size * m_channels);
 
+  BaseInitialize();
+  m_volume = 100;
+  m_paused = false;
+
+  hr = m_source_voice->Start(0, 0);
+  if (FAILED(hr))
+  {
+    Log_ErrorPrintf("Start() failed: %08X", hr);
+    return false;
+  }
+
+  EnqueueBuffer();
   return true;
 }
 
-void XAudio2AudioStream::PauseDevice(bool paused)
+void XAudio2AudioStream::SetPaused(bool paused)
 {
   if (m_paused == paused)
     return;
@@ -124,6 +147,9 @@ void XAudio2AudioStream::PauseDevice(bool paused)
   }
 
   m_paused = paused;
+
+  if (!m_buffer_enqueued)
+    EnqueueBuffer();
 }
 
 void XAudio2AudioStream::CloseDevice()
@@ -139,29 +165,20 @@ void XAudio2AudioStream::CloseDevice()
   m_source_voice = nullptr;
   m_mastering_voice = nullptr;
   m_xaudio.Reset();
-  m_buffers = {};
+  m_enqueue_buffers = {};
   m_current_buffer = 0;
   m_paused = true;
 }
 
-void XAudio2AudioStream::FramesAvailable()
-{
-  if (!m_buffer_enqueued)
-  {
-    m_buffer_enqueued = true;
-    EnqueueBuffer();
-  }
-}
-
 void XAudio2AudioStream::EnqueueBuffer()
 {
-  SampleType* samples = m_buffers[m_current_buffer].get();
+  SampleType* samples = m_enqueue_buffers[m_current_buffer].get();
-  ReadFrames(samples, m_buffer_size, false);
+  ReadFrames(samples, m_enqueue_buffer_size);
 
   const XAUDIO2_BUFFER buf = {
-    static_cast<UINT32>(0),                                        // flags
+    static_cast<UINT32>(0),                                                // flags
-    static_cast<UINT32>(sizeof(s16) * m_channels * m_buffer_size), // bytes
+    static_cast<UINT32>(sizeof(s16) * m_channels * m_enqueue_buffer_size), // bytes
-    reinterpret_cast<const BYTE*>(samples)                         // data
+    reinterpret_cast<const BYTE*>(samples)                                 // data
   };
 
   HRESULT hr = m_source_voice->SubmitSourceBuffer(&buf, nullptr);
@@ -173,10 +190,14 @@ void XAudio2AudioStream::EnqueueBuffer()
 
 void XAudio2AudioStream::SetOutputVolume(u32 volume)
 {
-  AudioStream::SetOutputVolume(volume);
+  HRESULT hr = m_mastering_voice->SetVolume(static_cast<float>(m_volume) / 100.0f);
-  HRESULT hr = m_mastering_voice->SetVolume(static_cast<float>(m_output_volume) / 100.0f);
   if (FAILED(hr))
+  {
     Log_ErrorPrintf("SetVolume() failed: %08X", hr);
+    return;
+  }
+
+  m_volume = volume;
 }
 
 void __stdcall XAudio2AudioStream::OnVoiceProcessingPassStart(UINT32 BytesRequired) {}
@@ -195,16 +216,3 @@ void __stdcall XAudio2AudioStream::OnBufferEnd(void* pBufferContext)
 void __stdcall XAudio2AudioStream::OnLoopEnd(void* pBufferContext) {}
 
 void __stdcall XAudio2AudioStream::OnVoiceError(void* pBufferContext, HRESULT Error) {}
-
-namespace FrontendCommon {
-
-std::unique_ptr<AudioStream> CreateXAudio2AudioStream()
-{
-  std::unique_ptr<XAudio2AudioStream> stream = std::make_unique<XAudio2AudioStream>();
-  if (!stream->Initialize())
-    return {};
-
-  return stream;
-}
-
-} // namespace FrontendCommon

src/frontend-common/xaudio2_audio_stream.h

@@ -14,45 +14,42 @@
 class XAudio2AudioStream final : public AudioStream, private IXAudio2VoiceCallback
 {
 public:
-  XAudio2AudioStream();
+  XAudio2AudioStream(u32 sample_rate, u32 channels, u32 buffer_ms, AudioStretchMode stretch);
   ~XAudio2AudioStream();
 
-  bool Initialize();
+  void SetPaused(bool paused) override;
-
   void SetOutputVolume(u32 volume) override;
 
-protected:
+  bool OpenDevice(u32 latency_ms);
+  void CloseDevice();
+  void EnqueueBuffer();
+
+private:
   enum : u32
   {
-    NUM_BUFFERS = 2
+    NUM_BUFFERS = 2,
+    INTERNAL_BUFFER_SIZE = 512,
   };
 
   ALWAYS_INLINE bool IsOpen() const { return static_cast<bool>(m_xaudio); }
 
-  bool OpenDevice() override;
-  void PauseDevice(bool paused) override;
-  void CloseDevice() override;
-  void FramesAvailable() override;
-
   // Inherited via IXAudio2VoiceCallback
-  virtual void __stdcall OnVoiceProcessingPassStart(UINT32 BytesRequired) override;
+  void __stdcall OnVoiceProcessingPassStart(UINT32 BytesRequired) override;
-  virtual void __stdcall OnVoiceProcessingPassEnd(void) override;
+  void __stdcall OnVoiceProcessingPassEnd(void) override;
-  virtual void __stdcall OnStreamEnd(void) override;
+  void __stdcall OnStreamEnd(void) override;
-  virtual void __stdcall OnBufferStart(void* pBufferContext) override;
+  void __stdcall OnBufferStart(void* pBufferContext) override;
-  virtual void __stdcall OnBufferEnd(void* pBufferContext) override;
+  void __stdcall OnBufferEnd(void* pBufferContext) override;
-  virtual void __stdcall OnLoopEnd(void* pBufferContext) override;
+  void __stdcall OnLoopEnd(void* pBufferContext) override;
-  virtual void __stdcall OnVoiceError(void* pBufferContext, HRESULT Error) override;
+  void __stdcall OnVoiceError(void* pBufferContext, HRESULT Error) override;
-
-  void EnqueueBuffer();
 
   Microsoft::WRL::ComPtr<IXAudio2> m_xaudio;
   IXAudio2MasteringVoice* m_mastering_voice = nullptr;
   IXAudio2SourceVoice* m_source_voice = nullptr;
 
-  std::array<std::unique_ptr<SampleType[]>, NUM_BUFFERS> m_buffers;
+  std::array<std::unique_ptr<SampleType[]>, NUM_BUFFERS> m_enqueue_buffers;
+  u32 m_enqueue_buffer_size = 0;
   u32 m_current_buffer = 0;
   bool m_buffer_enqueued = false;
-  bool m_paused = true;
 
 #if WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_DESKTOP)
   HMODULE m_xaudio2_library = {};

src/util/CMakeLists.txt

@@ -27,8 +27,6 @@ add_library(util
   iso_reader.h
   jit_code_buffer.cpp
   jit_code_buffer.h
-  null_audio_stream.cpp
-  null_audio_stream.h
   memory_arena.cpp
   memory_arena.h
   page_fault_handler.cpp
@@ -44,4 +42,4 @@ add_library(util
 target_include_directories(util PRIVATE "${CMAKE_CURRENT_SOURCE_DIR}/..")
 target_include_directories(util PUBLIC "${CMAKE_CURRENT_SOURCE_DIR}/..")
 target_link_libraries(util PUBLIC common simpleini)
-target_link_libraries(util PRIVATE libchdr samplerate zlib)
+target_link_libraries(util PRIVATE libchdr samplerate zlib soundtouch)

src/util/audio_stream.cpp

@@ -1,387 +1,615 @@
 #include "audio_stream.h"
-#include "assert.h"
+#include "SoundTouch.h"
+#include "common/align.h"
+#include "common/assert.h"
 #include "common/log.h"
-#include "samplerate.h"
+#include "common/make_array.h"
+#include "common/timer.h"
 #include <algorithm>
+#include <cmath>
 #include <cstring>
 Log_SetChannel(AudioStream);
 
-AudioStream::AudioStream() = default;
+#if defined(_M_ARM64)
+#include <arm64_neon.h>
+#elif defined(__aarch64__)
+#include <arm_neon.h>
+#elif defined(_M_IX86) || defined(_M_AMD64)
+#include <emmintrin.h>
+#endif
+
+static constexpr bool LOG_TIMESTRETCH_STATS = false;
+
+AudioStream::AudioStream(u32 sample_rate, u32 channels, u32 buffer_ms, AudioStretchMode stretch)
+  : m_sample_rate(sample_rate), m_channels(channels), m_buffer_ms(buffer_ms), m_stretch_mode(stretch)
+{
+}
 
 AudioStream::~AudioStream()
 {
-  DestroyResampler();
+  DestroyBuffer();
 }
 
-bool AudioStream::Reconfigure(u32 input_sample_rate /* = DefaultInputSampleRate */,
+std::unique_ptr<AudioStream> AudioStream::CreateNullStream(u32 sample_rate, u32 channels, u32 buffer_ms)
-                              u32 output_sample_rate /* = DefaultOutputSampleRate */, u32 channels /* = 1 */,
-                              u32 buffer_size /* = DefaultBufferSize */)
 {
-  std::unique_lock<std::mutex> buffer_lock(m_buffer_mutex);
+  return std::unique_ptr<AudioStream>(new AudioStream(sample_rate, channels, buffer_ms, AudioStretchMode::Off));
-  std::unique_lock<std::mutex> resampler_Lock(m_resampler_mutex);
+}
 
-  DestroyResampler();
+u32 AudioStream::GetAlignedBufferSize(u32 size)
-  if (IsDeviceOpen())
+{
-    CloseDevice();
+  static_assert(Common::IsPow2(CHUNK_SIZE));
+  return Common::AlignUpPow2(size, CHUNK_SIZE);
+}
 
-  m_output_sample_rate = output_sample_rate;
+u32 AudioStream::GetBufferSizeForMS(u32 sample_rate, u32 ms)
-  m_channels = channels;
+{
-  m_buffer_size = buffer_size;
+  return GetAlignedBufferSize((ms * sample_rate) / 1000u);
-  m_buffer_filling.store(m_wait_for_buffer_fill);
+}
-  m_output_paused = true;
 
-  if (!SetBufferSize(buffer_size))
+u32 AudioStream::GetMSForBufferSize(u32 sample_rate, u32 buffer_size)
-    return false;
+{
+  buffer_size = GetAlignedBufferSize(buffer_size);
+  return (buffer_size * 1000u) / sample_rate;
+}
 
-  if (!OpenDevice())
+static constexpr const auto s_stretch_mode_names = make_array("None", "Resample", "TimeStretch");
+
+const char* AudioStream::GetStretchModeName(AudioStretchMode mode)
+{
+  return (static_cast<u32>(mode) < s_stretch_mode_names.size()) ? s_stretch_mode_names[static_cast<u32>(mode)] : "";
+}
+
+std::optional<AudioStretchMode> AudioStream::ParseStretchMode(const char* name)
+{
+  for (u8 i = 0; i < static_cast<u8>(AudioStretchMode::Count); i++)
   {
-    LockedEmptyBuffers();
+    if (std::strcmp(name, s_stretch_mode_names[i]) == 0)
-    m_buffer_size = 0;
+      return static_cast<AudioStretchMode>(i);
-    m_output_sample_rate = 0;
-    m_channels = 0;
-    return false;
   }
 
-  CreateResampler();
+  return std::nullopt;
-  InternalSetInputSampleRate(input_sample_rate);
-
-  return true;
 }
 
-void AudioStream::SetInputSampleRate(u32 sample_rate)
+u32 AudioStream::GetBufferedFramesRelaxed() const
 {
-  std::unique_lock<std::mutex> buffer_lock(m_buffer_mutex);
+  const u32 rpos = m_rpos.load(std::memory_order_relaxed);
-  std::unique_lock<std::mutex> resampler_lock(m_resampler_mutex);
+  const u32 wpos = m_wpos.load(std::memory_order_relaxed);
-
+  return (wpos + m_buffer_size - rpos) % m_buffer_size;
-  InternalSetInputSampleRate(sample_rate);
 }
 
-void AudioStream::SetWaitForBufferFill(bool enabled)
+void AudioStream::ReadFrames(s16* bData, u32 nFrames)
 {
-  std::unique_lock<std::mutex> buffer_lock(m_buffer_mutex);
+  const u32 available_frames = GetBufferedFramesRelaxed();
-  m_wait_for_buffer_fill = enabled;
+  u32 frames_to_read = nFrames;
-  if (enabled && m_buffer.IsEmpty())
+  u32 silence_frames = 0;
-    m_buffer_filling.store(true);
+
+  if (m_filling)
+  {
+    u32 toFill = m_buffer_size / ((m_stretch_mode != AudioStretchMode::TimeStretch) ? 32 : 400);
+    toFill = GetAlignedBufferSize(toFill);
+
+    if (available_frames < toFill)
+    {
+      silence_frames = nFrames;
+      frames_to_read = 0;
+    }
+    else
+    {
+      m_filling = false;
+      Log_VerbosePrintf("Underrun compensation done (%d frames buffered)", toFill);
+    }
+  }
+  else if (available_frames < nFrames)
+  {
+    silence_frames = nFrames - available_frames;
+    frames_to_read = available_frames;
+    m_filling = true;
+
+    if (m_stretch_mode == AudioStretchMode::TimeStretch)
+      StretchUnderrun();
+  }
+
+  if (frames_to_read > 0)
+  {
+    u32 rpos = m_rpos.load(std::memory_order_acquire);
+
+    u32 end = m_buffer_size - rpos;
+    if (end > frames_to_read)
+      end = frames_to_read;
+
+    // towards the end of the buffer
+    if (end > 0)
+    {
+      std::memcpy(bData, &m_buffer[rpos], sizeof(s32) * end);
+      rpos += end;
+      rpos = (rpos == m_buffer_size) ? 0 : rpos;
+    }
+
+    // after wrapping around
+    const u32 start = frames_to_read - end;
+    if (start > 0)
+    {
+      std::memcpy(&bData[end * 2], &m_buffer[0], sizeof(s32) * start);
+      rpos = start;
+    }
+
+    m_rpos.store(rpos, std::memory_order_release);
+  }
+
+  // TODO: Bring back the crappy resampler?
+  if (silence_frames > 0)
+    std::memset(bData + frames_to_read, 0, sizeof(s32) * silence_frames);
 }
 
-void AudioStream::InternalSetInputSampleRate(u32 sample_rate)
+void AudioStream::InternalWriteFrames(s32* bData, u32 nSamples)
 {
-  if (m_input_sample_rate == sample_rate)
+  const u32 free = m_buffer_size - GetBufferedFramesRelaxed();
+  if (free <= nSamples)
+  {
+    if (m_stretch_mode == AudioStretchMode::TimeStretch)
+    {
+      StretchOverrun();
+    }
+    else
+    {
+      Log_DebugPrintf("Buffer overrun, chunk dropped");
+      return;
+    }
+  }
+
+  u32 wpos = m_wpos.load(std::memory_order_acquire);
+
+  // wrapping around the end of the buffer?
+  if ((m_buffer_size - wpos) <= nSamples)
+  {
+    // needs to be written in two parts
+    const u32 end = m_buffer_size - wpos;
+    const u32 start = nSamples - end;
+
+    // start is zero when this chunk reaches exactly the end
+    std::memcpy(&m_buffer[wpos], bData, end * sizeof(s32));
+    if (start > 0)
+      std::memcpy(&m_buffer[0], bData + end, start * sizeof(s32));
+
+    wpos = start;
+  }
+  else
+  {
+    // no split
+    std::memcpy(&m_buffer[wpos], bData, nSamples * sizeof(s32));
+    wpos += nSamples;
+  }
+
+  m_wpos.store(wpos, std::memory_order_release);
+}
+
+void AudioStream::BaseInitialize()
+{
+  AllocateBuffer();
+  StretchAllocate();
+}
+
+void AudioStream::AllocateBuffer()
+{
+  // use a larger buffer when time stretching, since we need more input
+  const u32 multplier =
+    (m_stretch_mode == AudioStretchMode::TimeStretch) ? 16 : ((m_stretch_mode == AudioStretchMode::Off) ? 1 : 2);
+  m_buffer_size = GetAlignedBufferSize(((m_buffer_ms * multplier) * m_sample_rate) / 1000);
+  m_target_buffer_size = GetAlignedBufferSize((m_sample_rate * m_buffer_ms) / 1000u);
+  m_buffer = std::unique_ptr<s32[]>(new s32[m_buffer_size]);
+  Log_DevPrintf("Allocated buffer of %u frames for buffer of %u ms [stretch %s, target size %u].", m_buffer_size,
+                m_buffer_ms, GetStretchModeName(m_stretch_mode), m_target_buffer_size);
+}
+
+void AudioStream::DestroyBuffer()
+{
+  m_buffer.reset();
+  m_buffer_size = 0;
+  m_wpos.store(0, std::memory_order_release);
+  m_rpos.store(0, std::memory_order_release);
+}
+
+void AudioStream::EmptyBuffer()
+{
+  if (m_stretch_mode != AudioStretchMode::Off)
+  {
+    m_soundtouch->clear();
+    if (m_stretch_mode == AudioStretchMode::TimeStretch)
+      m_soundtouch->setTempo(m_nominal_rate);
+  }
+
+  m_wpos.store(m_rpos.load(std::memory_order_acquire), std::memory_order_release);
+}
+
+void AudioStream::SetNominalRate(float tempo)
+{
+  m_nominal_rate = tempo;
+  if (m_stretch_mode == AudioStretchMode::Resample)
+    m_soundtouch->setRate(tempo);
+}
+
+void AudioStream::SetStretchMode(AudioStretchMode mode)
+{
+  if (m_stretch_mode == mode)
     return;
 
-  m_input_sample_rate = sample_rate;
+  // can't resize the buffers while paused
-  m_resampler_ratio = static_cast<double>(m_output_sample_rate) / static_cast<double>(sample_rate);
+  bool paused = m_paused;
-  src_set_ratio(static_cast<SRC_STATE*>(m_resampler_state), m_resampler_ratio);
+  if (!paused)
-  ResetResampler();
+    SetPaused(true);
+
+  DestroyBuffer();
+  StretchDestroy();
+  m_stretch_mode = mode;
+
+  AllocateBuffer();
+  if (m_stretch_mode != AudioStretchMode::Off)
+    StretchAllocate();
+
+  if (!paused)
+    SetPaused(false);
+}
+
+void AudioStream::SetPaused(bool paused)
+{
+  m_paused = paused;
 }
 
 void AudioStream::SetOutputVolume(u32 volume)
 {
-  std::unique_lock<std::mutex> lock(m_buffer_mutex);
+  m_volume = volume;
-  m_output_volume = volume;
-}
-
-void AudioStream::PauseOutput(bool paused)
-{
-  if (m_output_paused == paused)
-    return;
-
-  PauseDevice(paused);
-  m_output_paused = paused;
-
-  // Empty buffers on pause.
-  if (paused)
-    EmptyBuffers();
-}
-
-void AudioStream::Shutdown()
-{
-  if (!IsDeviceOpen())
-    return;
-
-  CloseDevice();
-  EmptyBuffers();
-  m_buffer_size = 0;
-  m_output_sample_rate = 0;
-  m_channels = 0;
-  m_output_paused = true;
 }
 
 void AudioStream::BeginWrite(SampleType** buffer_ptr, u32* num_frames)
 {
-  m_buffer_mutex.lock();
+  // TODO: Write directly to buffer when not using stretching.
-
+  *buffer_ptr = reinterpret_cast<s16*>(&m_staging_buffer[m_staging_buffer_pos]);
-  const u32 requested_frames = std::min(*num_frames, m_buffer_size);
+  *num_frames = CHUNK_SIZE - m_staging_buffer_pos;
-  EnsureBuffer(requested_frames * m_channels);
-
-  *buffer_ptr = m_buffer.GetWritePointer();
-  *num_frames = std::min(m_buffer_size, m_buffer.GetContiguousSpace() / m_channels);
 }
 
 void AudioStream::WriteFrames(const SampleType* frames, u32 num_frames)
 {
-  Assert(num_frames <= m_buffer_size);
+  Panic("not implemented");
-  const u32 num_samples = num_frames * m_channels;
-  {
-    std::unique_lock<std::mutex> lock(m_buffer_mutex);
-    EnsureBuffer(num_samples);
-    m_buffer.PushRange(frames, num_samples);
-  }
-
-  FramesAvailable();
 }
 
 void AudioStream::EndWrite(u32 num_frames)
 {
-  m_buffer.AdvanceTail(num_frames * m_channels);
+  // don't bother committing anything when muted
-  if (m_buffer_filling.load())
+  if (m_volume == 0)
-  {
-    if ((m_buffer.GetSize() / m_channels) >= m_buffer_size)
-      m_buffer_filling.store(false);
-  }
-  m_buffer_mutex.unlock();
-  FramesAvailable();
-}
-
-float AudioStream::GetMaxLatency(u32 sample_rate, u32 buffer_size)
-{
-  return (static_cast<float>(buffer_size) / static_cast<float>(sample_rate));
-}
-
-bool AudioStream::SetBufferSize(u32 buffer_size)
-{
-  const u32 buffer_size_in_samples = buffer_size * m_channels;
-  const u32 max_samples = buffer_size_in_samples * 2u;
-  if (max_samples > m_buffer.GetCapacity())
-    return false;
-
-  m_buffer_size = buffer_size;
-  m_max_samples = max_samples;
-  return true;
-}
-
-u32 AudioStream::GetSamplesAvailable() const
-{
-  // TODO: Use atomic loads
-  u32 available_samples;
-  {
-    std::unique_lock<std::mutex> lock(m_buffer_mutex);
-    available_samples = m_buffer.GetSize();
-  }
-
-  return available_samples / m_channels;
-}
-
-u32 AudioStream::GetSamplesAvailableLocked() const
-{
-  return m_buffer.GetSize() / m_channels;
-}
-
-void AudioStream::ReadFrames(SampleType* samples, u32 num_frames, bool apply_volume)
-{
-  const u32 total_samples = num_frames * m_channels;
-  u32 samples_copied = 0;
-  std::unique_lock<std::mutex> buffer_lock(m_buffer_mutex);
-  if (!m_buffer_filling.load())
-  {
-    if (m_input_sample_rate == m_output_sample_rate)
-    {
-      samples_copied = std::min(m_buffer.GetSize(), total_samples);
-      if (samples_copied > 0)
-        m_buffer.PopRange(samples, samples_copied);
-
-      ReleaseBufferLock(std::move(buffer_lock));
-    }
-    else
-    {
-      if (m_resampled_buffer.GetSize() < total_samples)
-        ResampleInput(std::move(buffer_lock));
-      else
-        ReleaseBufferLock(std::move(buffer_lock));
-
-      samples_copied = std::min(m_resampled_buffer.GetSize(), total_samples);
-      if (samples_copied > 0)
-        m_resampled_buffer.PopRange(samples, samples_copied);
-    }
-  }
-  else
-  {
-    ReleaseBufferLock(std::move(buffer_lock));
-  }
-
-  if (samples_copied < total_samples)
-  {
-    if (samples_copied > 0)
-    {
-      m_resample_buffer.resize(samples_copied);
-      std::memcpy(m_resample_buffer.data(), samples, sizeof(SampleType) * samples_copied);
-
-      // super basic resampler - spread the input samples evenly across the output samples. will sound like ass and have
-      // aliasing, but better than popping by inserting silence.
-      const u32 increment =
-        static_cast<u32>(65536.0f * (static_cast<float>(samples_copied / m_channels) / static_cast<float>(num_frames)));
-
-      SampleType* out_ptr = samples;
-      const SampleType* resample_ptr = m_resample_buffer.data();
-      const u32 copy_stride = sizeof(SampleType) * m_channels;
-      u32 resample_subpos = 0;
-      for (u32 i = 0; i < num_frames; i++)
-      {
-        std::memcpy(out_ptr, resample_ptr, copy_stride);
-        out_ptr += m_channels;
-
-        resample_subpos += increment;
-        resample_ptr += (resample_subpos >> 16) * m_channels;
-        resample_subpos %= 65536u;
-      }
-
-      Log_VerbosePrintf("Audio buffer underflow, resampled %u frames to %u", samples_copied / m_channels, num_frames);
-      m_underflow_flag.store(true);
-    }
-    else
-    {
-      // read nothing, so zero-fill
-      std::memset(samples, 0, sizeof(SampleType) * total_samples);
-      Log_VerbosePrintf("Audio buffer underflow with no samples, added %u frames silence", num_frames);
-      m_underflow_flag.store(true);
-    }
-
-    m_buffer_filling.store(m_wait_for_buffer_fill);
-  }
-
-  if (apply_volume && m_output_volume != FullVolume)
-  {
-    SampleType* current_ptr = samples;
-    const SampleType* end_ptr = samples + (num_frames * m_channels);
-    while (current_ptr != end_ptr)
-    {
-      *current_ptr = ApplyVolume(*current_ptr, m_output_volume);
-      current_ptr++;
-    }
-  }
-}
-
-void AudioStream::EnsureBuffer(u32 size)
-{
-  DebugAssert(size <= (m_buffer_size * m_channels));
-  if (GetBufferSpace() >= size)
     return;
 
-  if (m_sync)
+  m_staging_buffer_pos += num_frames;
-  {
+  DebugAssert(m_staging_buffer_pos <= CHUNK_SIZE);
-    std::unique_lock<std::mutex> lock(m_buffer_mutex, std::adopt_lock);
+  if (m_staging_buffer_pos < CHUNK_SIZE)
-    m_buffer_draining_cv.wait(lock, [this, size]() { return GetBufferSpace() >= size; });
+    return;
-    lock.release();
+
-  }
+  m_staging_buffer_pos = 0;
+
+  if (m_stretch_mode != AudioStretchMode::Off)
+    StretchWrite();
   else
+    InternalWriteFrames(m_staging_buffer.data(), CHUNK_SIZE);
+}
+
+static constexpr float S16_TO_FLOAT = 1.0f / 32767.0f;
+static constexpr float FLOAT_TO_S16 = 32767.0f;
+
+#if defined(_M_ARM64) || defined(__aarch64__)
+
+static void S16ChunkToFloat(const s32* src, float* dst)
+{
+  static_assert((AudioStream::CHUNK_SIZE % 4) == 0);
+  constexpr u32 iterations = AudioStream::CHUNK_SIZE / 4;
+
+  const float32x4_t S16_TO_FLOAT_V = vdupq_n_f32(S16_TO_FLOAT);
+
+  for (u32 i = 0; i < iterations; i++)
   {
-    m_buffer.Remove(size);
+    const int16x8_t sv = vreinterpretq_s16_s32(vld1q_s32(src));
+    src += 4;
+
+    int32x4_t iv1 = vreinterpretq_s32_s16(vzip1q_s16(sv, sv)); // [0, 0, 1, 1, 2, 2, 3, 3]
+    int32x4_t iv2 = vreinterpretq_s32_s16(vzip2q_s16(sv, sv)); // [4, 4, 5, 5, 6, 6, 7, 7]
+    iv1 = vshrq_n_s32(iv1, 16);                                // [0, 1, 2, 3]
+    iv2 = vshrq_n_s32(iv2, 16);                                // [4, 5, 6, 7]
+    float32x4_t fv1 = vcvtq_f32_s32(iv1);                      // [f0, f1, f2, f3]
+    float32x4_t fv2 = vcvtq_f32_s32(iv2);                      // [f4, f5, f6, f7]
+    fv1 = vmulq_f32(fv1, S16_TO_FLOAT_V);
+    fv2 = vmulq_f32(fv2, S16_TO_FLOAT_V);
+
+    vst1q_f32(dst + 0, fv1);
+    vst1q_f32(dst + 4, fv2);
+    dst += 8;
   }
 }
 
-void AudioStream::DropFrames(u32 count)
+static void FloatChunkToS16(s32* dst, const float* src, uint size)
 {
-  std::unique_lock<std::mutex> lock(m_buffer_mutex);
+  static_assert((AudioStream::CHUNK_SIZE % 4) == 0);
-  m_buffer.Remove(count);
+  constexpr u32 iterations = AudioStream::CHUNK_SIZE / 4;
-}
 
-void AudioStream::EmptyBuffers()
+  const float32x4_t FLOAT_TO_S16_V = vdupq_n_f32(FLOAT_TO_S16);
-{
-  std::unique_lock<std::mutex> lock(m_buffer_mutex);
-  std::unique_lock<std::mutex> resampler_lock(m_resampler_mutex);
-  LockedEmptyBuffers();
-}
 
-void AudioStream::LockedEmptyBuffers()
+  for (u32 i = 0; i < iterations; i++)
-{
-  m_buffer.Clear();
-  m_underflow_flag.store(false);
-  m_buffer_filling.store(m_wait_for_buffer_fill);
-  ResetResampler();
-}
-
-void AudioStream::CreateResampler()
-{
-  m_resampler_state = src_new(SRC_SINC_MEDIUM_QUALITY, static_cast<int>(m_channels), nullptr);
-  if (!m_resampler_state)
-    Panic("Failed to allocate resampler");
-}
-
-void AudioStream::DestroyResampler()
-{
-  if (m_resampler_state)
   {
-    src_delete(static_cast<SRC_STATE*>(m_resampler_state));
+    float32x4_t fv1 = vld1q_s32(src + 0);
-    m_resampler_state = nullptr;
+    float32x4_t fv2 = vld1q_s32(src + 4);
+    src += 8;
+
+    fv1 = vmulq_f32(fv1, FLOAT_TO_S16_V);
+    fv2 = vmulq_f32(fv2, FLOAT_TO_S16_V);
+    int32x4_t iv1 = vcvtq_s32_f32(fv1);
+    int32x4_t iv2 = vcvtq_s32_f32(fv2);
+
+    int16x8_t iv = vcombine_s16(vqmovn_s32(iv1), vqmovn_s32(iv2));
+    vst1q_s32(dst, vreinterpretq_s32_s16(iv));
+    dst += 4;
   }
 }
 
-void AudioStream::ResetResampler()
+#elif defined(_M_IX86) || defined(_M_AMD64)
+
+static void S16ChunkToFloat(const s32* src, float* dst)
 {
-  m_resampled_buffer.Clear();
+  static_assert((AudioStream::CHUNK_SIZE % 4) == 0);
-  m_resample_in_buffer.clear();
+  constexpr u32 iterations = AudioStream::CHUNK_SIZE / 4;
-  m_resample_out_buffer.clear();
+
-  src_reset(static_cast<SRC_STATE*>(m_resampler_state));
+  const __m128 S16_TO_FLOAT_V = _mm_set1_ps(S16_TO_FLOAT);
+
+  for (u32 i = 0; i < iterations; i++)
+  {
+    const __m128i sv = _mm_load_si128(reinterpret_cast<const __m128i*>(src));
+    src += 4;
+
+    __m128i iv1 = _mm_unpacklo_epi16(sv, sv); // [0, 0, 1, 1, 2, 2, 3, 3]
+    __m128i iv2 = _mm_unpackhi_epi16(sv, sv); // [4, 4, 5, 5, 6, 6, 7, 7]
+    iv1 = _mm_srai_epi32(iv1, 16);            // [0, 1, 2, 3]
+    iv2 = _mm_srai_epi32(iv2, 16);            // [4, 5, 6, 7]
+    __m128 fv1 = _mm_cvtepi32_ps(iv1);        // [f0, f1, f2, f3]
+    __m128 fv2 = _mm_cvtepi32_ps(iv2);        // [f4, f5, f6, f7]
+    fv1 = _mm_mul_ps(fv1, S16_TO_FLOAT_V);
+    fv2 = _mm_mul_ps(fv2, S16_TO_FLOAT_V);
+
+    _mm_store_ps(dst + 0, fv1);
+    _mm_store_ps(dst + 4, fv2);
+    dst += 8;
+  }
 }
 
-void AudioStream::ResampleInput(std::unique_lock<std::mutex> buffer_lock)
+static void FloatChunkToS16(s32* dst, const float* src, uint size)
 {
-  std::unique_lock<std::mutex> resampler_lock(m_resampler_mutex);
+  static_assert((AudioStream::CHUNK_SIZE % 4) == 0);
+  constexpr u32 iterations = AudioStream::CHUNK_SIZE / 4;
 
-  const u32 input_space_from_output = (m_resampled_buffer.GetSpace() * m_output_sample_rate) / m_input_sample_rate;
+  const __m128 FLOAT_TO_S16_V = _mm_set1_ps(FLOAT_TO_S16);
-  u32 remaining = std::min(m_buffer.GetSize(), input_space_from_output);
+
-  if (m_resample_in_buffer.size() < remaining)
+  for (u32 i = 0; i < iterations; i++)
   {
-    remaining -= static_cast<u32>(m_resample_in_buffer.size());
+    __m128 fv1 = _mm_load_ps(src + 0);
-    m_resample_in_buffer.reserve(m_resample_in_buffer.size() + remaining);
+    __m128 fv2 = _mm_load_ps(src + 4);
-    while (remaining > 0)
+    src += 8;
+
+    fv1 = _mm_mul_ps(fv1, FLOAT_TO_S16_V);
+    fv2 = _mm_mul_ps(fv2, FLOAT_TO_S16_V);
+    __m128i iv1 = _mm_cvtps_epi32(fv1);
+    __m128i iv2 = _mm_cvtps_epi32(fv2);
+
+    __m128i iv = _mm_packs_epi32(iv1, iv2);
+    _mm_store_si128(reinterpret_cast<__m128i*>(dst), iv);
+    dst += 4;
+  }
+}
+
+#else
+
+static void S16ChunkToFloat(const s32* src, float* dst)
+{
+  for (uint i = 0; i < AudioStream::CHUNK_SIZE; ++i)
+  {
+    *(dst++) = static_cast<float>(static_cast<s16>((u32)*src)) / 32767.0f;
+    *(dst++) = static_cast<float>(static_cast<s16>(((u32)*src) >> 16)) / 32767.0f;
+    src++;
+  }
+}
+
+static void FloatChunkToS16(s32* dst, const float* src, uint size)
+{
+  for (uint i = 0; i < size; ++i)
+  {
+    const s16 left = static_cast<s16>((*(src++) * 32767.0f));
+    const s16 right = static_cast<s16>((*(src++) * 32767.0f));
+    *(dst++) = (static_cast<u32>(left) & 0xFFFFu) | (static_cast<u32>(right) << 16);
+  }
+}
+#endif
+
+// Time stretching algorithm based on PCSX2 implementation.
+
+template<class T>
+ALWAYS_INLINE static bool IsInRange(const T& val, const T& min, const T& max)
+{
+  return (min <= val && val <= max);
+}
+
+void AudioStream::StretchAllocate()
+{
+  if (m_stretch_mode == AudioStretchMode::Off)
+    return;
+
+  m_soundtouch = std::make_unique<soundtouch::SoundTouch>();
+  m_soundtouch->setSampleRate(m_sample_rate);
+  m_soundtouch->setChannels(m_channels);
+
+  m_soundtouch->setSetting(SETTING_USE_QUICKSEEK, 0);
+  m_soundtouch->setSetting(SETTING_USE_AA_FILTER, 0);
+
+  m_soundtouch->setSetting(SETTING_SEQUENCE_MS, 30);
+  m_soundtouch->setSetting(SETTING_SEEKWINDOW_MS, 20);
+  m_soundtouch->setSetting(SETTING_OVERLAP_MS, 10);
+
+  if (m_stretch_mode == AudioStretchMode::Resample)
+    m_soundtouch->setRate(m_nominal_rate);
+  else
+    m_soundtouch->setTempo(m_nominal_rate);
+
+  m_stretch_reset = STRETCH_RESET_THRESHOLD;
+  m_stretch_inactive = false;
+  m_stretch_ok_count = 0;
+  m_dynamic_target_usage = 0.0f;
+  m_average_position = 0;
+  m_average_available = 0;
+
+  m_staging_buffer_pos = 0;
+}
+
+void AudioStream::StretchDestroy()
+{
+  m_soundtouch.reset();
+}
+
+void AudioStream::StretchWrite()
+{
+  S16ChunkToFloat(m_staging_buffer.data(), m_float_buffer.data());
+
+  m_soundtouch->putSamples(m_float_buffer.data(), CHUNK_SIZE);
+
+  int tempProgress;
+  while (tempProgress = m_soundtouch->receiveSamples((float*)m_float_buffer.data(), CHUNK_SIZE), tempProgress != 0)
+  {
+    FloatChunkToS16(m_staging_buffer.data(), m_float_buffer.data(), tempProgress);
+    InternalWriteFrames(m_staging_buffer.data(), tempProgress);
+  }
+
+  if (m_stretch_mode == AudioStretchMode::TimeStretch)
+    UpdateStretchTempo();
+}
+
+float AudioStream::AddAndGetAverageTempo(float val)
+{
+  if (m_stretch_reset >= STRETCH_RESET_THRESHOLD)
+    m_average_available = 0;
+  if (m_average_available < AVERAGING_BUFFER_SIZE)
+    m_average_available++;
+
+  m_average_fullness[m_average_position] = val;
+  m_average_position = (m_average_position + 1U) % AVERAGING_BUFFER_SIZE;
+
+  const u32 actual_window = std::min<u32>(m_average_available, AVERAGING_WINDOW);
+  const u32 first_index = (m_average_position - actual_window + AVERAGING_BUFFER_SIZE) % AVERAGING_BUFFER_SIZE;
+
+  float sum = 0;
+  for (u32 i = first_index; i < first_index + actual_window; i++)
+    sum += m_average_fullness[i % AVERAGING_BUFFER_SIZE];
+  sum = sum / actual_window;
+
+  return (sum != 0.0f) ? sum : 1.0f;
+}
+
+void AudioStream::UpdateStretchTempo()
+{
+  static constexpr float MIN_TEMPO = 0.05f;
+  static constexpr float MAX_TEMPO = 50.0f;
+
+  // Which range we will run in 1:1 mode for.
+  static constexpr float INACTIVE_GOOD_FACTOR = 1.04f;
+  static constexpr float INACTIVE_BAD_FACTOR = 1.2f;
+  static constexpr u32 INACTIVE_MIN_OK_COUNT = 50;
+  static constexpr u32 COMPENSATION_DIVIDER = 100;
+
+  float base_target_usage = static_cast<float>(m_target_buffer_size) * m_nominal_rate;
+
+  // state vars
+  if (m_stretch_reset >= STRETCH_RESET_THRESHOLD)
+  {
+    Log_VerbosePrintf("___ Stretcher is being reset.");
+    m_stretch_inactive = false;
+    m_stretch_ok_count = 0;
+    m_dynamic_target_usage = base_target_usage;
+  }
+
+  const u32 ibuffer_usage = GetBufferedFramesRelaxed();
+  float buffer_usage = static_cast<float>(ibuffer_usage);
+  float tempo = buffer_usage / m_dynamic_target_usage;
+  tempo = AddAndGetAverageTempo(tempo);
+
+  // Dampening when we get close to target.
+  if (tempo < 2.0f)
+    tempo = std::sqrt(tempo);
+
+  tempo = std::clamp(tempo, MIN_TEMPO, MAX_TEMPO);
+
+  if (tempo < 1.0f)
+    base_target_usage /= std::sqrt(tempo);
+
+  m_dynamic_target_usage +=
+    static_cast<float>(base_target_usage / tempo - m_dynamic_target_usage) / static_cast<float>(COMPENSATION_DIVIDER);
+  if (IsInRange(tempo, 0.9f, 1.1f) &&
+      IsInRange(m_dynamic_target_usage, base_target_usage * 0.9f, base_target_usage * 1.1f))
+  {
+    m_dynamic_target_usage = base_target_usage;
+  }
+
+  if (!m_stretch_inactive)
+  {
+    if (IsInRange(tempo, 1.0f / INACTIVE_GOOD_FACTOR, INACTIVE_GOOD_FACTOR))
+      m_stretch_ok_count++;
+    else
+      m_stretch_ok_count = 0;
+
+    if (m_stretch_ok_count >= INACTIVE_MIN_OK_COUNT)
     {
-      const u32 read_len = std::min(m_buffer.GetContiguousSize(), remaining);
+      Log_VerbosePrintf("=== Stretcher is now inactive.");
-      const size_t old_pos = m_resample_in_buffer.size();
+      m_stretch_inactive = true;
-      m_resample_in_buffer.resize(m_resample_in_buffer.size() + read_len);
-      src_short_to_float_array(m_buffer.GetReadPointer(), m_resample_in_buffer.data() + old_pos,
-                               static_cast<int>(read_len));
-      m_buffer.Remove(read_len);
-      remaining -= read_len;
     }
   }
-
+  else if (!IsInRange(tempo, 1.0f / INACTIVE_BAD_FACTOR, INACTIVE_BAD_FACTOR))
-  ReleaseBufferLock(std::move(buffer_lock));
-
-  const u32 potential_output_size =
-    (static_cast<u32>(m_resample_in_buffer.size()) * m_input_sample_rate) / m_output_sample_rate;
-  const u32 output_size = std::min(potential_output_size, m_resampled_buffer.GetSpace());
-  m_resample_out_buffer.resize(output_size);
-
-  SRC_DATA sd = {};
-  sd.data_in = m_resample_in_buffer.data();
-  sd.data_out = m_resample_out_buffer.data();
-  sd.input_frames = static_cast<u32>(m_resample_in_buffer.size()) / m_channels;
-  sd.output_frames = output_size / m_channels;
-  sd.src_ratio = m_resampler_ratio;
-
-  const int error = src_process(static_cast<SRC_STATE*>(m_resampler_state), &sd);
-  if (error)
   {
-    Log_ErrorPrintf("Resampler error %d", error);
+    Log_VerbosePrintf("~~~ Stretcher is now active @ tempo %f.", tempo);
-    m_resample_in_buffer.clear();
+    m_stretch_inactive = false;
-    m_resample_out_buffer.clear();
+    m_stretch_ok_count = 0;
-    return;
   }
 
-  m_resample_in_buffer.erase(m_resample_in_buffer.begin(),
+  if (m_stretch_inactive)
-                             m_resample_in_buffer.begin() + (static_cast<u32>(sd.input_frames_used) * m_channels));
+    tempo = m_nominal_rate;
 
-  const float* write_ptr = m_resample_out_buffer.data();
+  if constexpr (LOG_TIMESTRETCH_STATS)
-  remaining = static_cast<u32>(sd.output_frames_gen) * m_channels;
-  while (remaining > 0)
   {
-    const u32 samples_to_write = std::min(m_resampled_buffer.GetContiguousSpace(), remaining);
+    static int iterations = 0;
-    src_float_to_short_array(write_ptr, m_resampled_buffer.GetWritePointer(), static_cast<int>(samples_to_write));
+    static u64 last_log_time = 0;
-    m_resampled_buffer.AdvanceTail(samples_to_write);
+
-    write_ptr += samples_to_write;
+    const u64 now = Common::Timer::GetCurrentValue();
-    remaining -= samples_to_write;
+
+    if (Common::Timer::ConvertValueToSeconds(now - last_log_time) > 1.0f)
+    {
+      Log_VerbosePrintf("buffers: %4u ms (%3.0f%%), tempo: %f, comp: %2.3f, iters: %d, reset:%d",
+                        (ibuffer_usage * 1000u) / m_sample_rate, 100.0f * buffer_usage / base_target_usage, tempo,
+                        m_dynamic_target_usage / base_target_usage, iterations, m_stretch_reset);
+
+      last_log_time = now;
+      iterations = 0;
+    }
+
+    iterations++;
   }
-  m_resample_out_buffer.erase(m_resample_out_buffer.begin(),
+
-                              m_resample_out_buffer.begin() + (static_cast<u32>(sd.output_frames_gen) * m_channels));
+  m_soundtouch->setTempo(tempo);
+
+  if (m_stretch_reset >= STRETCH_RESET_THRESHOLD)
+    m_stretch_reset = 0;
+}
+
+void AudioStream::StretchUnderrun()
+{
+  // Didn't produce enough frames in time.
+  m_stretch_reset++;
+}
+
+void AudioStream::StretchOverrun()
+{
+  // Produced more frames than can fit in the buffer.
+  m_stretch_reset++;
+
+  // Drop two packets to give the time stretcher a bit more time to slow things down.
+  const u32 discard = CHUNK_SIZE * 2;
+  m_rpos.store((m_rpos.load(std::memory_order_acquire) + discard) % m_buffer_size, std::memory_order_release);
 }

src/util/audio_stream.h

@@ -1,13 +1,26 @@
 #pragma once
-#include "common/fifo_queue.h"
 #include "common/types.h"
+#include <array>
 #include <atomic>
-#include <condition_variable>
 #include <memory>
-#include <mutex>
+#include <optional>
-#include <vector>
 
-// Uses signed 16-bits samples.
+#ifdef _MSC_VER
+#pragma warning(push)
+#pragma warning(disable : 4324) // warning C4324: structure was padded due to alignment specifier
+#endif
+
+namespace soundtouch {
+class SoundTouch;
+}
+
+enum class AudioStretchMode : u8
+{
+  Off,
+  Resample,
+  TimeStretch,
+  Count
+};
 
 class AudioStream
 {
@@ -16,111 +29,116 @@ public:
 
   enum : u32
   {
-    DefaultInputSampleRate = 44100,
+    CHUNK_SIZE = 64,
-    DefaultOutputSampleRate = 44100,
+    MAX_CHANNELS = 2
-    DefaultBufferSize = 2048,
-    MaxSamples = 32768,
-    FullVolume = 100
   };
 
-  AudioStream();
+public:
   virtual ~AudioStream();
 
-  u32 GetOutputSampleRate() const { return m_output_sample_rate; }
+  static u32 GetAlignedBufferSize(u32 size);
-  u32 GetChannels() const { return m_channels; }
+  static u32 GetBufferSizeForMS(u32 sample_rate, u32 ms);
-  u32 GetBufferSize() const { return m_buffer_size; }
+  static u32 GetMSForBufferSize(u32 sample_rate, u32 buffer_size);
-  s32 GetOutputVolume() const { return m_output_volume; }
-  bool IsSyncing() const { return m_sync; }
 
-  bool Reconfigure(u32 input_sample_rate = DefaultInputSampleRate, u32 output_sample_rate = DefaultOutputSampleRate,
+  static const char* GetStretchModeName(AudioStretchMode mode);
-                   u32 channels = 1, u32 buffer_size = DefaultBufferSize);
+  static std::optional<AudioStretchMode> ParseStretchMode(const char* name);
-  void SetSync(bool enable) { m_sync = enable; }
 
-  void SetInputSampleRate(u32 sample_rate);
+  ALWAYS_INLINE u32 GetSampleRate() const { return m_sample_rate; }
-  void SetWaitForBufferFill(bool enabled);
+  ALWAYS_INLINE u32 GetChannels() const { return m_channels; }
+  ALWAYS_INLINE u32 GetBufferSize() const { return m_buffer_size; }
+  ALWAYS_INLINE u32 GetTargetBufferSize() const { return m_target_buffer_size; }
+  ALWAYS_INLINE u32 GetOutputVolume() const { return m_volume; }
+  ALWAYS_INLINE float GetNominalTempo() const { return m_nominal_rate; }
+  ALWAYS_INLINE bool IsPaused() const { return m_paused; }
+
+  u32 GetBufferedFramesRelaxed() const;
+
+  /// Temporarily pauses the stream, preventing it from requesting data.
+  virtual void SetPaused(bool paused);
 
   virtual void SetOutputVolume(u32 volume);
 
-  void PauseOutput(bool paused);
-  void EmptyBuffers();
-
-  void Shutdown();
-
   void BeginWrite(SampleType** buffer_ptr, u32* num_frames);
   void WriteFrames(const SampleType* frames, u32 num_frames);
   void EndWrite(u32 num_frames);
 
-  bool DidUnderflow()
+  void EmptyBuffer();
-  {
-    bool expected = true;
-    return m_underflow_flag.compare_exchange_strong(expected, false);
-  }
 
-  static std::unique_ptr<AudioStream> CreateNullAudioStream();
+  /// Nominal rate is used for both resampling and timestretching, input samples are assumed to be this amount faster
+  /// than the sample rate.
+  void SetNominalRate(float tempo);
 
-  // Latency computation - returns values in seconds
+  void SetStretchMode(AudioStretchMode mode);
-  static float GetMaxLatency(u32 sample_rate, u32 buffer_size);
+
+  static std::unique_ptr<AudioStream> CreateNullStream(u32 sample_rate, u32 channels, u32 buffer_ms);
 
 protected:
-  virtual bool OpenDevice() = 0;
+  AudioStream(u32 sample_rate, u32 channels, u32 buffer_ms, AudioStretchMode stretch);
-  virtual void PauseDevice(bool paused) = 0;
+  void BaseInitialize();
-  virtual void CloseDevice() = 0;
-  virtual void FramesAvailable() = 0;
 
-  ALWAYS_INLINE static SampleType ApplyVolume(SampleType sample, u32 volume)
+  void ReadFrames(s16* bData, u32 nSamples);
-  {
-    return s16((s32(sample) * s32(volume)) / 100);
-  }
 
-  ALWAYS_INLINE u32 GetBufferSpace() const { return (m_max_samples - m_buffer.GetSize()); }
+  u32 m_sample_rate = 0;
-  ALWAYS_INLINE void ReleaseBufferLock(std::unique_lock<std::mutex> lock)
-  {
-    // lock is released implicitly by destruction
-    m_buffer_draining_cv.notify_one();
-  }
-
-  bool SetBufferSize(u32 buffer_size);
-  bool IsDeviceOpen() const { return (m_output_sample_rate > 0); }
-
-  void EnsureBuffer(u32 size);
-  void LockedEmptyBuffers();
-  u32 GetSamplesAvailable() const;
-  u32 GetSamplesAvailableLocked() const;
-  void ReadFrames(SampleType* samples, u32 num_frames, bool apply_volume);
-  void DropFrames(u32 count);
-
-  void CreateResampler();
-  void DestroyResampler();
-  void ResetResampler();
-  void InternalSetInputSampleRate(u32 sample_rate);
-  void ResampleInput(std::unique_lock<std::mutex> buffer_lock);
-
-  u32 m_input_sample_rate = 0;
-  u32 m_output_sample_rate = 0;
   u32 m_channels = 0;
+  u32 m_buffer_ms = 0;
+  u32 m_volume = 0;
+
+  AudioStretchMode m_stretch_mode = AudioStretchMode::Off;
+  bool m_stretch_inactive = false;
+  bool m_filling = false;
+  bool m_paused = false;
+
+private:
+  enum : u32
+  {
+    AVERAGING_BUFFER_SIZE = 256,
+    AVERAGING_WINDOW = 50,
+    STRETCH_RESET_THRESHOLD = 5,
+    TARGET_IPS = 691,
+  };
+
+  void AllocateBuffer();
+  void DestroyBuffer();
+
+  void InternalWriteFrames(s32* bData, u32 nFrames);
+
+  void StretchAllocate();
+  void StretchDestroy();
+  void StretchWrite();
+  void StretchUnderrun();
+  void StretchOverrun();
+
+  float AddAndGetAverageTempo(float val);
+  void UpdateStretchTempo();
+
   u32 m_buffer_size = 0;
+  std::unique_ptr<s32[]> m_buffer;
 
-  // volume, 0-100
+  std::atomic<u32> m_rpos{0};
-  u32 m_output_volume = FullVolume;
+  std::atomic<u32> m_wpos{0};
 
-  HeapFIFOQueue<SampleType, MaxSamples> m_buffer;
+  std::unique_ptr<soundtouch::SoundTouch> m_soundtouch;
-  mutable std::mutex m_buffer_mutex;
-  std::condition_variable m_buffer_draining_cv;
-  std::vector<SampleType> m_resample_buffer;
 
-  std::atomic_bool m_underflow_flag{false};
+  u32 m_target_buffer_size = 0;
-  std::atomic_bool m_buffer_filling{false};
+  u32 m_stretch_reset = STRETCH_RESET_THRESHOLD;
-  u32 m_max_samples = 0;
 
-  bool m_output_paused = true;
+  u32 m_stretch_ok_count = 0;
-  bool m_sync = true;
+  float m_nominal_rate = 1.0f;
-  bool m_wait_for_buffer_fill = false;
+  float m_dynamic_target_usage = 0.0f;
 
-  // Resampling
+  u32 m_average_position = 0;
-  double m_resampler_ratio = 1.0;
+  u32 m_average_available = 0;
-  void* m_resampler_state = nullptr;
+  u32 m_staging_buffer_pos = 0;
-  std::mutex m_resampler_mutex;
+
-  HeapFIFOQueue<SampleType, MaxSamples> m_resampled_buffer;
+  std::array<float, AVERAGING_BUFFER_SIZE> m_average_fullness = {};
-  std::vector<float> m_resample_in_buffer;
+
-  std::vector<float> m_resample_out_buffer;
+  // temporary staging buffer, used for timestretching
+  alignas(16) std::array<s32, CHUNK_SIZE> m_staging_buffer;
+
+  // float buffer, soundtouch only accepts float samples as input
+  alignas(16) std::array<float, CHUNK_SIZE * MAX_CHANNELS> m_float_buffer;
 };
+
+#ifdef _MSC_VER
+#pragma warning(pop)
+#endif

src/util/null_audio_stream.cpp

@@ -1,25 +0,0 @@
-#include "null_audio_stream.h"
-
-NullAudioStream::NullAudioStream() = default;
-
-NullAudioStream::~NullAudioStream() = default;
-
-bool NullAudioStream::OpenDevice()
-{
-  return true;
-}
-
-void NullAudioStream::PauseDevice(bool paused) {}
-
-void NullAudioStream::CloseDevice() {}
-
-void NullAudioStream::FramesAvailable()
-{
-  // drop any buffer as soon as they're available
-  DropFrames(GetSamplesAvailable());
-}
-
-std::unique_ptr<AudioStream> AudioStream::CreateNullAudioStream()
-{
-  return std::make_unique<NullAudioStream>();
-}

src/util/null_audio_stream.h

@@ -1,15 +0,0 @@
-#pragma once
-#include "audio_stream.h"
-
-class NullAudioStream final : public AudioStream
-{
-public:
-  NullAudioStream();
-  ~NullAudioStream();
-
-protected:
-  bool OpenDevice() override;
-  void PauseDevice(bool paused) override;
-  void CloseDevice() override;
-  void FramesAvailable() override;
-};

src/util/util.props

@@ -4,13 +4,15 @@
 
   <ItemDefinitionGroup>
     <ClCompile>
-      <AdditionalIncludeDirectories>$(SolutionDir)dep\simpleini\include;$(SolutionDir)dep\libsamplerate\include;$(SolutionDir)dep\libchdr\include;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <PreprocessorDefinitions>%(PreprocessorDefinitions);SOUNDTOUCH_FLOAT_SAMPLES;SOUNDTOUCH_ALLOW_SSE</PreprocessorDefinitions>
+      <PreprocessorDefinitions Condition="'$(Platform)'=='ARM64'">%(PreprocessorDefinitions);SOUNDTOUCH_USE_NEON</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(SolutionDir)dep\soundtouch\include;$(SolutionDir)dep\simpleini\include;$(SolutionDir)dep\libchdr\include;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
     </ClCompile>
   </ItemDefinitionGroup>
 
   <ItemDefinitionGroup>
     <Link>
-      <AdditionalDependencies>$(RootBuildDir)simpleini\simpleini.lib;$(RootBuildDir)libchdr\libchdr.lib;$(RootBuildDir)libsamplerate\libsamplerate.lib;%(AdditionalDependencies)</AdditionalDependencies>
+      <AdditionalDependencies>$(RootBuildDir)soundtouch\soundtouch.lib;$(RootBuildDir)simpleini\simpleini.lib;$(RootBuildDir)libchdr\libchdr.lib;%(AdditionalDependencies)</AdditionalDependencies>
     </Link>
   </ItemDefinitionGroup>
 </Project>

src/util/util.vcxproj

@@ -9,7 +9,6 @@
     <ClInclude Include="ini_settings_interface.h" />
     <ClInclude Include="iso_reader.h" />
     <ClInclude Include="jit_code_buffer.h" />
-    <ClInclude Include="null_audio_stream.h" />
     <ClInclude Include="pbp_types.h" />
     <ClInclude Include="memory_arena.h" />
     <ClInclude Include="page_fault_handler.h" />
@@ -38,7 +37,6 @@
     <ClCompile Include="iso_reader.cpp" />
     <ClCompile Include="jit_code_buffer.cpp" />
     <ClCompile Include="cd_subchannel_replacement.cpp" />
-    <ClCompile Include="null_audio_stream.cpp" />
     <ClCompile Include="shiftjis.cpp" />
     <ClCompile Include="memory_arena.cpp" />
     <ClCompile Include="page_fault_handler.cpp" />

src/util/util.vcxproj.filters

@@ -8,7 +8,6 @@
     <ClInclude Include="iso_reader.h" />
     <ClInclude Include="cd_image.h" />
     <ClInclude Include="cd_subchannel_replacement.h" />
-    <ClInclude Include="null_audio_stream.h" />
     <ClInclude Include="wav_writer.h" />
     <ClInclude Include="cd_image_hasher.h" />
     <ClInclude Include="shiftjis.h" />
@@ -28,7 +27,6 @@
     <ClCompile Include="cd_image_bin.cpp" />
     <ClCompile Include="iso_reader.cpp" />
     <ClCompile Include="cd_subchannel_replacement.cpp" />
-    <ClCompile Include="null_audio_stream.cpp" />
     <ClCompile Include="cd_image_chd.cpp" />
     <ClCompile Include="wav_writer.cpp" />
     <ClCompile Include="cd_image_hasher.cpp" />