GPU: Further improve vsync handling

src/core/host.cpp

@@ -288,8 +288,7 @@ bool Host::CreateGPUDevice(RenderAPI api, Error* error)
   if (!g_gpu_device || !g_gpu_device->Create(
                          g_settings.gpu_adapter,
                          g_settings.gpu_disable_shader_cache ? std::string_view() : std::string_view(EmuFolders::Cache),
-                         SHADER_CACHE_VERSION, g_settings.gpu_use_debug_device, System::IsHostVSyncEffectivelyEnabled(),
+                         SHADER_CACHE_VERSION, g_settings.gpu_use_debug_device, System::GetEffectiveVSyncMode(),
-                         System::IsHostVSyncEffectivelyEnabled() && !System::IsHostVSyncUsedForTiming(),
                          g_settings.gpu_threaded_presentation, exclusive_fullscreen_control,
                          static_cast<GPUDevice::FeatureMask>(disabled_features), &create_error))
   {

src/core/system.cpp

@@ -122,6 +122,8 @@ static void Throttle(Common::Timer::Value current_time);
 static void UpdatePerformanceCounters();
 static void AccumulatePreFrameSleepTime();
 static void UpdatePreFrameSleepTime();
+static void UpdateDisplayVSync();
+static void UpdateDisplayMaxFPS();
 
 static void SetRewinding(bool enabled);
 static bool SaveRewindState();
@@ -184,7 +186,6 @@ static bool s_turbo_enabled = false;
 static bool s_throttler_enabled = false;
 static bool s_optimal_frame_pacing = false;
 static bool s_pre_frame_sleep = false;
-static bool s_syncing_to_host = false;
 static bool s_syncing_to_host_with_vsync = false;
 static bool s_skip_presenting_duplicate_frames = false;
 static u32 s_skipped_frame_count = 0;
@@ -1263,6 +1264,7 @@ void System::PauseSystem(bool paused)
 
     Host::OnSystemPaused();
     Host::OnIdleStateChanged();
+    UpdateDisplayMaxFPS();
     InvalidateDisplay();
   }
   else
@@ -1277,6 +1279,7 @@ void System::PauseSystem(bool paused)
     Host::OnSystemResumed();
     Host::OnIdleStateChanged();
 
+    UpdateDisplayMaxFPS();
     ResetPerformanceCounters();
     ResetThrottler();
   }
@@ -1819,7 +1822,8 @@ void System::DestroySystem()
   // Restore present-all-frames behavior.
   if (s_keep_gpu_device_on_shutdown && g_gpu_device)
   {
-    g_gpu_device->SetDisplayMaxFPS(0.0f);
+    UpdateDisplayVSync();
+    UpdateDisplayMaxFPS();
   }
   else
   {
@@ -2685,7 +2689,7 @@ float System::GetTargetSpeed()
 
 float System::GetAudioNominalRate()
 {
-  return s_throttler_enabled ? s_target_speed : 1.0f;
+  return (s_throttler_enabled || s_syncing_to_host_with_vsync) ? s_target_speed : 1.0f;
 }
 
 void System::UpdatePerformanceCounters()
@@ -2845,32 +2849,31 @@ void System::UpdateSpeedLimiterState()
                            g_gpu_device->GetWindowInfo().IsSurfaceless(); // surfaceless check for regtest
   s_skip_presenting_duplicate_frames = s_throttler_enabled && g_settings.display_skip_presenting_duplicate_frames;
   s_pre_frame_sleep = s_optimal_frame_pacing && g_settings.display_pre_frame_sleep;
-
-  s_syncing_to_host = false;
   s_syncing_to_host_with_vsync = false;
-  if (g_settings.sync_to_host_refresh_rate &&
+
-      (g_settings.audio_stream_parameters.stretch_mode != AudioStretchMode::Off) && s_target_speed == 1.0f && IsValid())
+  if (const float host_refresh_rate = g_gpu_device->GetWindowInfo().surface_refresh_rate; host_refresh_rate > 0.0f)
   {
-    if (std::optional<float> host_refresh_rate = g_gpu_device->GetHostRefreshRate(); host_refresh_rate.has_value())
+    const float ratio = host_refresh_rate / System::GetThrottleFrequency();
+    const bool can_sync_to_host = (ratio >= 0.95f && ratio <= 1.05f);
+    INFO_LOG("Refresh rate: Host={}hz Guest={}hz Ratio={} - {}", host_refresh_rate, System::GetThrottleFrequency(),
+             ratio, can_sync_to_host ? "can sync" : "can't sync");
+
+    if (can_sync_to_host && g_settings.sync_to_host_refresh_rate && s_target_speed == 1.0f)
     {
-      const float ratio = host_refresh_rate.value() / System::GetThrottleFrequency();
+      s_target_speed = ratio;
-      s_syncing_to_host = (ratio >= 0.95f && ratio <= 1.05f);
+
-      INFO_LOG("Refresh rate: Host={}hz Guest={}hz Ratio={} - {}", host_refresh_rate.value(),
+      // When syncing to host and using vsync, we don't need to sleep.
-               System::GetThrottleFrequency(), ratio, s_syncing_to_host ? "can sync" : "can't sync");
+      s_syncing_to_host_with_vsync = g_settings.display_vsync;
-      if (s_syncing_to_host)
+      if (s_syncing_to_host_with_vsync)
-        s_target_speed *= ratio;
+      {
+        INFO_LOG("Using host vsync for throttling.");
+        s_throttler_enabled = false;
+      }
     }
   }
 
-  // When syncing to host and using vsync, we don't need to sleep.
-  s_syncing_to_host_with_vsync = (s_syncing_to_host && IsHostVSyncEffectivelyEnabled());
-  if (s_syncing_to_host_with_vsync)
-  {
-    INFO_LOG("Using host vsync for throttling.");
-    s_throttler_enabled = false;
-  }
-
   VERBOSE_LOG("Target speed: {}%", s_target_speed * 100.0f);
+  VERBOSE_LOG("Preset timing: {}", s_optimal_frame_pacing ? "consistent" : "immediate");
 
   // Update audio output.
   AudioStream* stream = SPU::GetOutputStream();
@@ -2879,35 +2882,60 @@ void System::UpdateSpeedLimiterState()
 
   UpdateThrottlePeriod();
   ResetThrottler();
-  UpdateDisplaySync();
+  UpdateDisplayVSync();
+  UpdateDisplayMaxFPS();
 
   if (g_settings.increase_timer_resolution)
     SetTimerResolutionIncreased(s_throttler_enabled);
 }
 
-void System::UpdateDisplaySync()
+void System::UpdateDisplayVSync()
 {
-  const bool vsync_enabled = IsHostVSyncEffectivelyEnabled();
+  static constexpr std::array<const char*, static_cast<size_t>(GPUVSyncMode::Count)> vsync_modes = {{
-  const float max_display_fps = (s_throttler_enabled || s_syncing_to_host) ? 0.0f : g_settings.display_max_fps;
+    "Disabled",
-  VERBOSE_LOG("VSync: {}{}", vsync_enabled ? "Enabled" : "Disabled",
+    "DoubleBuffered",
+    "TripleBuffered",
+  }};
+
+  // Avoid flipping vsync on and off by manually throttling when vsync is on.
+  const GPUVSyncMode vsync_mode = GetEffectiveVSyncMode();
+  VERBOSE_LOG("VSync: {}{}", vsync_modes[static_cast<size_t>(vsync_mode)],
               s_syncing_to_host_with_vsync ? " (for throttling)" : "");
+
+  g_gpu_device->SetVSyncMode(vsync_mode);
+}
+
+void System::UpdateDisplayMaxFPS()
+{
+  const GPUVSyncMode vsync_mode = GetEffectiveVSyncMode();
+  const float max_display_fps =
+    (!IsPaused() && IsValid() &&
+     (s_target_speed == 0.0f ||
+      (vsync_mode != GPUVSyncMode::Disabled && s_target_speed != 1.0f && !s_syncing_to_host_with_vsync))) ?
+      g_gpu_device->GetWindowInfo().surface_refresh_rate :
+      0.0f;
+
   VERBOSE_LOG("Max display fps: {}", max_display_fps);
-  VERBOSE_LOG("Preset timing: {}", s_optimal_frame_pacing ? "consistent" : "immediate");
-
   g_gpu_device->SetDisplayMaxFPS(max_display_fps);
-  g_gpu_device->SetVSyncEnabled(vsync_enabled, vsync_enabled && !IsHostVSyncUsedForTiming());
 }
 
-bool System::IsHostVSyncEffectivelyEnabled()
+GPUVSyncMode System::GetEffectiveVSyncMode()
 {
-  // Disable vsync if running outside 100%.
+  // Vsync off => always disabled.
-  return (g_settings.display_vsync && (s_state != State::Shutdown && s_state != State::Stopping) &&
+  if (!g_settings.display_vsync)
-          !IsRunningAtNonStandardSpeed());
+    return GPUVSyncMode::Disabled;
-}
 
-bool System::IsHostVSyncUsedForTiming()
+  // If there's no VM, or we're using vsync for timing, then we always use double-buffered (blocking).
-{
+  if (s_state == State::Shutdown || s_state == State::Stopping || s_syncing_to_host_with_vsync)
-  return (IsHostVSyncEffectivelyEnabled() && s_syncing_to_host_with_vsync);
+    return GPUVSyncMode::DoubleBuffered;
+
+  // For PAL games, we always want to triple buffer, because otherwise we'll be tearing.
+  // Or for when we aren't using sync-to-host-refresh, to avoid dropping frames.
+  // Force vsync off when not running at 100% speed.
+  // We can avoid this by manually throttling when vsync is on (see above).
+  return (s_throttler_enabled && g_gpu_device->GetWindowInfo().surface_refresh_rate == 0.0f) ?
+           GPUVSyncMode::Disabled :
+           GPUVSyncMode::TripleBuffered;
 }
 
 bool System::IsFastForwardEnabled()
@@ -3978,7 +4006,6 @@ void System::CheckForSettingsChanges(const Settings& old_settings)
         g_settings.increase_timer_resolution != old_settings.increase_timer_resolution ||
         g_settings.emulation_speed != old_settings.emulation_speed ||
         g_settings.fast_forward_speed != old_settings.fast_forward_speed ||
-        g_settings.display_max_fps != old_settings.display_max_fps ||
         g_settings.display_optimal_frame_pacing != old_settings.display_optimal_frame_pacing ||
         g_settings.display_skip_presenting_duplicate_frames != old_settings.display_skip_presenting_duplicate_frames ||
         g_settings.display_pre_frame_sleep != old_settings.display_pre_frame_sleep ||
@@ -3999,6 +4026,17 @@ void System::CheckForSettingsChanges(const Settings& old_settings)
 
     PostProcessing::UpdateSettings();
   }
+  else
+  {
+    if (g_gpu_device)
+    {
+      if (g_settings.display_vsync != old_settings.display_vsync)
+      {
+        UpdateDisplayVSync();
+        UpdateDisplayMaxFPS();
+      }
+    }
+  }
 
   if (g_gpu_device)
   {

src/core/system.h

@@ -19,6 +19,8 @@ class Error;
 class SmallStringBase;
 class StateWrapper;
 
+enum class GPUVSyncMode : u8;
+
 class Controller;
 
 struct CheatCode;
@@ -295,9 +297,6 @@ void UpdateThrottlePeriod();
 void ResetThrottler();
 void ResetPerformanceCounters();
 
-/// Resets vsync/max present fps state.
-void UpdateDisplaySync();
-
 // Access controllers for simulating input.
 Controller* GetController(u32 slot);
 void UpdateControllers();
@@ -466,10 +465,7 @@ void ToggleWidescreen();
 bool IsRunningAtNonStandardSpeed();
 
 /// Returns true if vsync should be used.
-bool IsHostVSyncEffectivelyEnabled();
+GPUVSyncMode GetEffectiveVSyncMode();
-
-/// Returns true if vsync is being used for frame pacing.
-bool IsHostVSyncUsedForTiming();
 
 /// Quick switch between software and hardware rendering.
 void ToggleSoftwareRendering();

src/duckstation-qt/qthost.cpp

@@ -1754,7 +1754,7 @@ void EmuThread::run()
       if (g_gpu_device)
       {
         System::PresentDisplay(false, false);
-        if (!g_gpu_device->IsVSyncEnabled())
+        if (!g_gpu_device->IsVSyncModeBlocking())
           g_gpu_device->ThrottlePresentation();
       }
     }

src/duckstation-qt/qtutils.cpp

@@ -7,6 +7,7 @@
 #include "core/system.h"
 
 #include "common/byte_stream.h"
+#include "common/log.h"
 
 #include <QtCore/QCoreApplication>
 #include <QtCore/QMetaObject>
@@ -38,6 +39,8 @@
 #include "common/windows_headers.h"
 #endif
 
+Log_SetChannel(QtUtils);
+
 namespace QtUtils {
 
 QFrame* CreateHorizontalLine(QWidget* parent)
@@ -365,6 +368,11 @@ std::optional<WindowInfo> GetWindowInfoForWidget(QWidget* widget)
   wi.surface_width = static_cast<u32>(static_cast<qreal>(widget->width()) * dpr);
   wi.surface_height = static_cast<u32>(static_cast<qreal>(widget->height()) * dpr);
   wi.surface_scale = static_cast<float>(dpr);
+
+  // Query refresh rate, we need it for sync.
+  wi.surface_refresh_rate = WindowInfo::QueryRefreshRateForWindow(wi).value_or(0.0f);
+  INFO_LOG("Surface refresh rate: {} hz", wi.surface_refresh_rate);
+
   return wi;
 }
 

src/util/d3d11_device.cpp

@@ -201,7 +201,7 @@ u32 D3D11Device::GetSwapChainBufferCount() const
 {
   // With vsync off, we only need two buffers. Same for blocking vsync.
   // With triple buffering, we need three.
-  return (m_vsync_enabled && m_vsync_prefer_triple_buffer) ? 3 : 2;
+  return (m_vsync_mode == GPUVSyncMode::TripleBuffered) ? 3 : 2;
 }
 
 bool D3D11Device::CreateSwapChain()
@@ -357,10 +357,6 @@ bool D3D11Device::CreateSwapChainRTV()
       m_window_info.surface_refresh_rate = static_cast<float>(desc.BufferDesc.RefreshRate.Numerator) /
                                            static_cast<float>(desc.BufferDesc.RefreshRate.Denominator);
     }
-    else
-    {
-      m_window_info.surface_refresh_rate = 0.0f;
-    }
   }
 
   return true;
@@ -592,31 +588,13 @@ void D3D11Device::InvalidateRenderTarget(GPUTexture* t)
     static_cast<D3D11Texture*>(t)->CommitClear(m_context.Get());
 }
 
-std::optional<float> D3D11Device::GetHostRefreshRate()
+void D3D11Device::SetVSyncMode(GPUVSyncMode mode)
 {
-  if (m_swap_chain && m_is_exclusive_fullscreen)
+  if (m_vsync_mode == mode)
-  {
-    DXGI_SWAP_CHAIN_DESC desc;
-    if (SUCCEEDED(m_swap_chain->GetDesc(&desc)) && desc.BufferDesc.RefreshRate.Numerator > 0 &&
-        desc.BufferDesc.RefreshRate.Denominator > 0)
-    {
-      DEV_LOG("using fs rr: {} {}", desc.BufferDesc.RefreshRate.Numerator, desc.BufferDesc.RefreshRate.Denominator);
-      return static_cast<float>(desc.BufferDesc.RefreshRate.Numerator) /
-             static_cast<float>(desc.BufferDesc.RefreshRate.Denominator);
-    }
-  }
-
-  return GPUDevice::GetHostRefreshRate();
-}
-
-void D3D11Device::SetVSyncEnabled(bool enabled, bool prefer_triple_buffer)
-{
-  if (m_vsync_enabled == enabled && m_vsync_prefer_triple_buffer == prefer_triple_buffer)
     return;
 
   const u32 old_buffer_count = GetSwapChainBufferCount();
-  m_vsync_enabled = enabled;
+  m_vsync_mode = mode;
-  m_vsync_prefer_triple_buffer = prefer_triple_buffer;
   if (!m_swap_chain)
     return;
 
@@ -656,7 +634,7 @@ bool D3D11Device::BeginPresent(bool skip_present)
   // This blows our our GPU usage number considerably, so read the timestamp before the final blit
   // in this configuration. It does reduce accuracy a little, but better than seeing 100% all of
   // the time, when it's more like a couple of percent.
-  if (m_vsync_enabled && m_gpu_timing_enabled)
+  if (IsVSyncModeBlocking() && m_gpu_timing_enabled)
     PopTimestampQuery();
 
   static constexpr float clear_color[4] = {0.0f, 0.0f, 0.0f, 1.0f};
@@ -674,12 +652,12 @@ void D3D11Device::EndPresent(bool explicit_present)
   DebugAssert(!explicit_present);
   DebugAssert(m_num_current_render_targets == 0 && !m_current_depth_target);
 
-  if (!m_vsync_enabled && m_gpu_timing_enabled)
+  if (!IsVSyncModeBlocking() && m_gpu_timing_enabled)
     PopTimestampQuery();
 
   // DirectX has no concept of tear-or-sync. I guess if we measured times ourselves, we could implement it.
-  if (m_vsync_enabled)
+  if (IsVSyncModeBlocking())
-    m_swap_chain->Present(BoolToUInt32(1), 0);
+    m_swap_chain->Present(1, 0);
   else if (m_using_allow_tearing) // Disabled or VRR, VRR requires the allow tearing flag :/
     m_swap_chain->Present(0, DXGI_PRESENT_ALLOW_TEARING);
   else

src/util/d3d11_device.h

@@ -96,8 +96,7 @@ public:
   void DrawIndexed(u32 index_count, u32 base_index, u32 base_vertex) override;
   void DrawIndexedWithBarrier(u32 index_count, u32 base_index, u32 base_vertex, DrawBarrier type) override;
 
-  std::optional<float> GetHostRefreshRate() override;
+  void SetVSyncMode(GPUVSyncMode mode) override;
-  void SetVSyncEnabled(bool enabled, bool prefer_triple_buffer) override;
 
   bool SetGPUTimingEnabled(bool enabled) override;
   float GetAndResetAccumulatedGPUTime() override;

src/util/d3d12_device.cpp

@@ -801,7 +801,7 @@ u32 D3D12Device::GetSwapChainBufferCount() const
 {
   // With vsync off, we only need two buffers. Same for blocking vsync.
   // With triple buffering, we need three.
-  return (m_vsync_enabled && m_vsync_prefer_triple_buffer) ? 3 : 2;
+  return (m_vsync_mode == GPUVSyncMode::TripleBuffered) ? 3 : 2;
 }
 
 bool D3D12Device::CreateSwapChain()
@@ -942,10 +942,6 @@ bool D3D12Device::CreateSwapChainRTV()
       m_window_info.surface_refresh_rate = static_cast<float>(desc.BufferDesc.RefreshRate.Numerator) /
                                            static_cast<float>(desc.BufferDesc.RefreshRate.Denominator);
     }
-    else
-    {
-      m_window_info.surface_refresh_rate = 0.0f;
-    }
   }
 
   m_current_swap_chain_buffer = 0;
@@ -1088,31 +1084,13 @@ std::string D3D12Device::GetDriverInfo() const
   return ret;
 }
 
-std::optional<float> D3D12Device::GetHostRefreshRate()
+void D3D12Device::SetVSyncMode(GPUVSyncMode mode)
 {
-  if (m_swap_chain && m_is_exclusive_fullscreen)
+  if (m_vsync_mode == mode)
-  {
-    DXGI_SWAP_CHAIN_DESC desc;
-    if (SUCCEEDED(m_swap_chain->GetDesc(&desc)) && desc.BufferDesc.RefreshRate.Numerator > 0 &&
-        desc.BufferDesc.RefreshRate.Denominator > 0)
-    {
-      DEV_LOG("using fs rr: {} {}", desc.BufferDesc.RefreshRate.Numerator, desc.BufferDesc.RefreshRate.Denominator);
-      return static_cast<float>(desc.BufferDesc.RefreshRate.Numerator) /
-             static_cast<float>(desc.BufferDesc.RefreshRate.Denominator);
-    }
-  }
-
-  return GPUDevice::GetHostRefreshRate();
-}
-
-void D3D12Device::SetVSyncEnabled(bool enabled, bool prefer_triple_buffer)
-{
-  if (m_vsync_enabled == enabled && m_vsync_prefer_triple_buffer == prefer_triple_buffer)
     return;
 
   const u32 old_buffer_count = GetSwapChainBufferCount();
-  m_vsync_enabled = enabled;
+  m_vsync_mode = mode;
-  m_vsync_prefer_triple_buffer = prefer_triple_buffer;
   if (!m_swap_chain)
     return;
 
@@ -1181,8 +1159,8 @@ void D3D12Device::SubmitPresent()
   DebugAssert(m_swap_chain);
 
   // DirectX has no concept of tear-or-sync. I guess if we measured times ourselves, we could implement it.
-  if (m_vsync_enabled)
+  if (IsVSyncModeBlocking())
-    m_swap_chain->Present(BoolToUInt32(1), 0);
+    m_swap_chain->Present(1, 0);
   else if (m_using_allow_tearing) // Disabled or VRR, VRR requires the allow tearing flag :/
     m_swap_chain->Present(0, DXGI_PRESENT_ALLOW_TEARING);
   else

src/util/d3d12_device.h

@@ -118,8 +118,7 @@ public:
   void DrawIndexed(u32 index_count, u32 base_index, u32 base_vertex) override;
   void DrawIndexedWithBarrier(u32 index_count, u32 base_index, u32 base_vertex, DrawBarrier type) override;
 
-  std::optional<float> GetHostRefreshRate() override;
+  void SetVSyncMode(GPUVSyncMode mode) override;
-  void SetVSyncEnabled(bool enabled, bool prefer_triple_buffer) override;
 
   bool SetGPUTimingEnabled(bool enabled) override;
   float GetAndResetAccumulatedGPUTime() override;

src/util/gpu_device.cpp

@@ -276,11 +276,10 @@ bool GPUDevice::IsSameRenderAPI(RenderAPI lhs, RenderAPI rhs)
 }
 
 bool GPUDevice::Create(std::string_view adapter, std::string_view shader_cache_path, u32 shader_cache_version,
-                       bool debug_device, bool vsync, bool vsync_prefer_triple_buffer, bool threaded_presentation,
+                       bool debug_device, GPUVSyncMode vsync, bool threaded_presentation,
                        std::optional<bool> exclusive_fullscreen_control, FeatureMask disabled_features, Error* error)
 {
-  m_vsync_enabled = vsync;
+  m_vsync_mode = vsync;
-  m_vsync_prefer_triple_buffer = vsync_prefer_triple_buffer;
   m_debug_device = debug_device;
 
   if (!AcquireWindow(true))
@@ -591,11 +590,6 @@ void GPUDevice::RenderImGui()
   }
 }
 
-void GPUDevice::SetVSyncEnabled(bool enabled, bool prefer_triple_buffer)
-{
-  m_vsync_enabled = enabled;
-}
-
 void GPUDevice::UploadVertexBuffer(const void* vertices, u32 vertex_size, u32 vertex_count, u32* base_vertex)
 {
   void* map;
@@ -1053,14 +1047,6 @@ void GPUDevice::ThrottlePresentation()
   Common::Timer::SleepUntil(m_last_frame_displayed_time, false);
 }
 
-std::optional<float> GPUDevice::GetHostRefreshRate()
-{
-  if (m_window_info.surface_refresh_rate > 0.0f)
-    return m_window_info.surface_refresh_rate;
-
-  return WindowInfo::QueryRefreshRateForWindow(m_window_info);
-}
-
 bool GPUDevice::SetGPUTimingEnabled(bool enabled)
 {
   return false;

src/util/gpu_device.h

@@ -36,6 +36,14 @@ enum class RenderAPI : u32
   Metal
 };
 
+enum class GPUVSyncMode : u8
+{
+  Disabled,
+  DoubleBuffered,
+  TripleBuffered,
+  Count
+};
+
 class GPUSampler
 {
 public:
@@ -573,8 +581,8 @@ public:
   virtual RenderAPI GetRenderAPI() const = 0;
 
   bool Create(std::string_view adapter, std::string_view shader_cache_path, u32 shader_cache_version, bool debug_device,
-              bool vsync, bool vsync_prefer_triple_buffer, bool threaded_presentation,
+              GPUVSyncMode vsync, bool threaded_presentation, std::optional<bool> exclusive_fullscreen_control,
-              std::optional<bool> exclusive_fullscreen_control, FeatureMask disabled_features, Error* error);
+              FeatureMask disabled_features, Error* error);
   void Destroy();
 
   virtual bool HasSurface() const = 0;
@@ -672,8 +680,9 @@ public:
   /// Renders ImGui screen elements. Call before EndPresent().
   void RenderImGui();
 
-  ALWAYS_INLINE bool IsVSyncEnabled() const { return m_vsync_enabled; }
+  ALWAYS_INLINE GPUVSyncMode GetVSyncMode() const { return m_vsync_mode; }
-  virtual void SetVSyncEnabled(bool enabled, bool prefer_triple_buffer);
+  ALWAYS_INLINE bool IsVSyncModeBlocking() const { return (m_vsync_mode >= GPUVSyncMode::DoubleBuffered); }
+  virtual void SetVSyncMode(GPUVSyncMode mode) = 0;
 
   ALWAYS_INLINE bool IsDebugDevice() const { return m_debug_device; }
   ALWAYS_INLINE size_t GetVRAMUsage() const { return s_total_vram_usage; }
@@ -689,8 +698,6 @@ public:
 
   virtual bool SupportsTextureFormat(GPUTexture::Format format) const = 0;
 
-  virtual std::optional<float> GetHostRefreshRate();
-
   /// Enables/disables GPU frame timing.
   virtual bool SetGPUTimingEnabled(bool enabled);
 
@@ -793,8 +800,7 @@ private:
 protected:
   static Statistics s_stats;
 
-  bool m_vsync_enabled = false;
+  GPUVSyncMode m_vsync_mode = GPUVSyncMode::Disabled;
-  bool m_vsync_prefer_triple_buffer = false;
   bool m_gpu_timing_enabled = false;
   bool m_debug_device = false;
 };

src/util/metal_device.h

@@ -260,12 +260,10 @@ public:
   void DrawIndexed(u32 index_count, u32 base_index, u32 base_vertex) override;
   void DrawIndexedWithBarrier(u32 index_count, u32 base_index, u32 base_vertex, DrawBarrier type) override;
 
-  std::optional<float> GetHostRefreshRate() override;
-
   bool SetGPUTimingEnabled(bool enabled) override;
   float GetAndResetAccumulatedGPUTime() override;
 
-  void SetVSyncEnabled(bool enabled, bool prefer_triple_buffer) override;
+  void SetVSyncMode(GPUVSyncMode mode) override;
 
   bool BeginPresent(bool skip_present) override;
   void EndPresent(bool explicit_submit) override;

src/util/metal_device.mm

@@ -115,20 +115,14 @@ bool MetalDevice::HasSurface() const
   return (m_layer != nil);
 }
 
-std::optional<float> MetalDevice::GetHostRefreshRate()
+void MetalDevice::SetVSyncMode(GPUVSyncMode mode)
 {
-  return GPUDevice::GetHostRefreshRate();
+  if (m_vsync_mode == mode)
-}
-
-void MetalDevice::SetVSyncEnabled(bool enabled, bool prefer_triple_buffer)
-{
-  if (m_vsync_enabled == enabled && m_vsync_prefer_triple_buffer == prefer_triple_buffer)
     return;
 
-  m_vsync_enabled = enabled;
+  m_vsync_mode = mode;
-  m_vsync_prefer_triple_buffer = prefer_triple_buffer;
   if (m_layer != nil)
-    [m_layer setDisplaySyncEnabled:enabled];
+    [m_layer setDisplaySyncEnabled:m_vsync_mode >= GPUVSyncMode::DoubleBuffered];
 }
 
 bool MetalDevice::CreateDevice(std::string_view adapter, bool threaded_presentation,
@@ -402,7 +396,7 @@ bool MetalDevice::CreateLayer()
       }
     });
 
-    [m_layer setDisplaySyncEnabled:m_vsync_enabled];
+    [m_layer setDisplaySyncEnabled:m_vsync_mode >= GPUVSyncMode::DoubleBuffered];
 
     DebugAssert(m_layer_pass_desc == nil);
     m_layer_pass_desc = [[MTLRenderPassDescriptor renderPassDescriptor] retain];

src/util/opengl_device.cpp

@@ -238,13 +238,12 @@ void OpenGLDevice::InsertDebugMessage(const char* msg)
 #endif
 }
 
-void OpenGLDevice::SetVSyncEnabled(bool enabled, bool prefer_triple_buffer)
+void OpenGLDevice::SetVSyncMode(GPUVSyncMode mode)
 {
-  if (m_vsync_enabled == enabled && m_vsync_prefer_triple_buffer == prefer_triple_buffer)
+  if (m_vsync_mode == mode)
     return;
 
-  m_vsync_enabled = enabled;
+  m_vsync_mode = mode;
-  m_vsync_prefer_triple_buffer = prefer_triple_buffer;
   SetSwapInterval();
 }
 
@@ -584,7 +583,7 @@ void OpenGLDevice::SetSwapInterval()
     return;
 
   // Window framebuffer has to be bound to call SetSwapInterval.
-  const s32 interval = m_vsync_enabled ? 1 : 0;
+  const s32 interval = (m_vsync_mode >= GPUVSyncMode::DoubleBuffered) ? 1 : 0;
   GLint current_fbo = 0;
   glGetIntegerv(GL_DRAW_FRAMEBUFFER_BINDING, &current_fbo);
   glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);

src/util/opengl_device.h

@@ -100,7 +100,7 @@ public:
   void DrawIndexed(u32 index_count, u32 base_index, u32 base_vertex) override;
   void DrawIndexedWithBarrier(u32 index_count, u32 base_index, u32 base_vertex, DrawBarrier type) override;
 
-  void SetVSyncEnabled(bool enabled, bool prefer_triple_buffer) override;
+  void SetVSyncMode(GPUVSyncMode mode) override;
 
   bool BeginPresent(bool skip_present) override;
   void EndPresent(bool explicit_present) override;

src/util/platform_misc.h

@@ -23,4 +23,10 @@ bool PlaySoundAsync(const char* path);
 namespace Host {
 /// Return the current window handle. Needed for DInput.
 std::optional<WindowInfo> GetTopLevelWindowInfo();
-} // namespace Host
+} // namespace Host
+
+// TODO: Move all the other Cocoa stuff in here.
+namespace CocoaTools {
+/// Returns the refresh rate of the display the window is placed on.
+std::optional<float> GetViewRefreshRate(const WindowInfo& wi);
+}

src/util/platform_misc_mac.mm

@@ -147,3 +147,24 @@ void CocoaTools::DestroyMetalLayer(WindowInfo* wi)
   [view setWantsLayer:NO];
   [layer release];
 }
+
+std::optional<float> CocoaTools::GetViewRefreshRate(const WindowInfo& wi)
+{
+  if (![NSThread isMainThread])
+  {
+    std::optional<float> ret;
+    dispatch_sync(dispatch_get_main_queue(), [&ret, wi]{ ret = GetViewRefreshRate(wi); });
+    return ret;
+  }
+
+  std::optional<float> ret;
+  NSView* const view = (__bridge NSView*)wi.window_handle;
+  const u32 did = [[[[[view window] screen] deviceDescription] valueForKey:@"NSScreenNumber"] unsignedIntValue];
+  if (CGDisplayModeRef mode = CGDisplayCopyDisplayMode(did))
+  {
+    ret = CGDisplayModeGetRefreshRate(mode);
+    CGDisplayModeRelease(mode);
+  }
+
+  return ret;
+}

src/util/vulkan_device.cpp

@@ -440,8 +440,13 @@ VkPresentModeKHR VulkanDevice::SelectPresentMode() const
 {
   // Use mailbox/triple buffering for "normal" vsync, due to PAL refresh rate mismatch.
   // Otherwise, use FIFO when syncing to host, because we don't want to return early.
-  return m_vsync_enabled ? (m_vsync_prefer_triple_buffer ? VK_PRESENT_MODE_MAILBOX_KHR : VK_PRESENT_MODE_FIFO_KHR) :
+  static constexpr std::array<VkPresentModeKHR, static_cast<size_t>(GPUVSyncMode::Count)> modes = {{
-                           VK_PRESENT_MODE_IMMEDIATE_KHR;
+    VK_PRESENT_MODE_IMMEDIATE_KHR, // Disabled
+    VK_PRESENT_MODE_FIFO_KHR,      // DoubleBuffered
+    VK_PRESENT_MODE_MAILBOX_KHR,   // TripleBuffered
+  }};
+
+  return modes[static_cast<size_t>(m_vsync_mode)];
 }
 
 bool VulkanDevice::CreateDevice(VkSurfaceKHR surface, bool enable_validation_layer)
@@ -2344,13 +2349,12 @@ std::string VulkanDevice::GetDriverInfo() const
   return ret;
 }
 
-void VulkanDevice::SetVSyncEnabled(bool enabled, bool prefer_triple_buffer)
+void VulkanDevice::SetVSyncMode(GPUVSyncMode mode)
 {
-  if (m_vsync_enabled == enabled && m_vsync_prefer_triple_buffer == prefer_triple_buffer)
+  if (m_vsync_mode == mode)
     return;
 
-  m_vsync_enabled = enabled;
+  m_vsync_mode = mode;
-  m_vsync_prefer_triple_buffer = prefer_triple_buffer;
   if (!m_swap_chain)
     return;
 

src/util/vulkan_device.h

@@ -129,7 +129,7 @@ public:
   bool SetGPUTimingEnabled(bool enabled) override;
   float GetAndResetAccumulatedGPUTime() override;
 
-  void SetVSyncEnabled(bool enabled, bool prefer_triple_buffer) override;
+  void SetVSyncMode(GPUVSyncMode mode) override;
 
   bool BeginPresent(bool skip_present) override;
   void EndPresent(bool explicit_present) override;

src/util/window_info.cpp

@@ -160,6 +160,18 @@ std::optional<float> WindowInfo::QueryRefreshRateForWindow(const WindowInfo& wi)
   return ret;
 }
 
+#elif defined(__APPLE__)
+
+#include "util/platform_misc.h"
+
+std::optional<float> WindowInfo::QueryRefreshRateForWindow(const WindowInfo& wi)
+{
+  if (wi.type == WindowInfo::Type::MacOS)
+    return CocoaTools::GetViewRefreshRate(wi);
+
+  return std::nullopt;
+}
+
 #else
 
 #ifdef ENABLE_X11