CoreTiming: Add "Rush Frame Presentation" setting to throttle only once after each presentation for lower input latency.

Source/Core/Core/Config/MainSettings.cpp

@@ -47,6 +47,8 @@ const Info<bool> MAIN_DSP_HLE{{System::Main, "Core", "DSPHLE"}, true};
 const Info<int> MAIN_MAX_FALLBACK{{System::Main, "Core", "MaxFallback"}, 100};
 const Info<int> MAIN_TIMING_VARIANCE{{System::Main, "Core", "TimingVariance"}, 40};
 const Info<bool> MAIN_CORRECT_TIME_DRIFT{{System::Main, "Core", "CorrectTimeDrift"}, false};
+const Info<bool> MAIN_RUSH_FRAME_PRESENTATION{{System::Main, "Core", "RushFramePresentation"},
+                                              false};
 #if defined(ANDROID)
 // Currently enabled by default on Android because the performance boost is really needed.
 constexpr bool DEFAULT_CPU_THREAD = true;

Source/Core/Core/Config/MainSettings.h

@@ -65,6 +65,7 @@ extern const Info<bool> MAIN_DSP_HLE;
 extern const Info<int> MAIN_MAX_FALLBACK;
 extern const Info<int> MAIN_TIMING_VARIANCE;
 extern const Info<bool> MAIN_CORRECT_TIME_DRIFT;
+extern const Info<bool> MAIN_RUSH_FRAME_PRESENTATION;
 extern const Info<bool> MAIN_CPU_THREAD;
 extern const Info<bool> MAIN_SYNC_ON_SKIP_IDLE;
 extern const Info<std::string> MAIN_DEFAULT_ISO;

Source/Core/Core/CoreTiming.cpp

@@ -29,6 +29,7 @@
 #include "VideoCommon/PerformanceMetrics.h"
 #include "VideoCommon/VideoBackendBase.h"
 #include "VideoCommon/VideoConfig.h"
+#include "VideoCommon/VideoEvents.h"
 
 namespace CoreTiming
 {
@@ -113,6 +114,11 @@ void CoreTimingManager::Init()
       ResetThrottle(GetTicks());
     }
   });
+
+  m_throttled_after_presentation = false;
+  m_frame_hook = m_system.GetVideoEvents().after_present_event.Register([this](const PresentInfo&) {
+    m_throttled_after_presentation.store(false, std::memory_order_relaxed);
+  });
 }
 
 void CoreTimingManager::Shutdown()
@@ -124,6 +130,7 @@ void CoreTimingManager::Shutdown()
   ClearPendingEvents();
   UnregisterAllEvents();
   CPUThreadConfigCallback::RemoveConfigChangedCallback(m_registered_config_callback_id);
+  m_frame_hook.reset();
 }
 
 void CoreTimingManager::RefreshConfig()
@@ -134,6 +141,11 @@ void CoreTimingManager::RefreshConfig()
                                                        1.0f);
   m_config_oc_inv_factor = 1.0f / m_config_oc_factor;
   m_config_sync_on_skip_idle = Config::Get(Config::MAIN_SYNC_ON_SKIP_IDLE);
+  m_config_rush_frame_presentation = Config::Get(Config::MAIN_RUSH_FRAME_PRESENTATION);
+
+  // We don't want to skip so much throttling that the audio buffer overfills.
+  m_max_throttle_skip_time =
+      std::chrono::milliseconds{Config::Get(Config::MAIN_AUDIO_BUFFER_SIZE)} / 2;
 
   // A maximum fallback is used to prevent the system from sleeping for
   // too long or going full speed in an attempt to catch up to timings.
@@ -422,6 +434,21 @@ void CoreTimingManager::SleepUntil(TimePoint time_point)
 
 void CoreTimingManager::Throttle(const s64 target_cycle)
 {
+  const TimePoint time = Clock::now();
+
+  const bool already_throttled =
+      m_throttled_after_presentation.exchange(true, std::memory_order_relaxed);
+
+  // If RushFramePresentation is enabled, try to Throttle just once after each presentation.
+  //  This lowers input latency by speeding through to presentation after grabbing input.
+  // Make sure we don't get too far ahead of proper timing though,
+  //  otherwise the emulator unreasonably speeds through loading screens that don't have XFB copies,
+  //  making audio sound terrible.
+  const bool skip_throttle = already_throttled && m_config_rush_frame_presentation &&
+                             ((GetTargetHostTime(target_cycle) - time) < m_max_throttle_skip_time);
+  if (skip_throttle)
+    return;
+
   if (IsSpeedUnlimited())
   {
     ResetThrottle(target_cycle);
@@ -441,8 +468,6 @@ void CoreTimingManager::Throttle(const s64 target_cycle)
 
   TimePoint target_time = CalculateTargetHostTimeInternal(target_cycle);
 
-  const TimePoint time = Clock::now();
-
   const TimePoint min_target = time - m_max_fallback;
 
   // "Correct Time Drift" setting prevents timing relaxing.

Source/Core/Core/CoreTiming.h

@@ -207,6 +207,7 @@ private:
   float m_config_oc_factor = 1.0f;
   float m_config_oc_inv_factor = 1.0f;
   bool m_config_sync_on_skip_idle = false;
+  bool m_config_rush_frame_presentation = false;
 
   s64 m_throttle_reference_cycle = 0;
   TimePoint m_throttle_reference_time = Clock::now();
@@ -232,6 +233,11 @@ private:
   Common::PrecisionTimer m_precision_gpu_timer;
 
   Common::EventHook m_core_state_changed_hook;
+  Common::EventHook m_frame_hook;
+
+  // Used to optionally minimize throttling for improving input latency.
+  std::atomic_bool m_throttled_after_presentation = false;
+  DT m_max_throttle_skip_time{};
 };
 
 }  // namespace CoreTiming

Source/Core/DolphinQt/Settings/AdvancedPane.cpp

@@ -114,6 +114,19 @@ void AdvancedPane::CreateLayout()
          "<br><br><dolphin_emphasis>If unsure, leave this unchecked.</dolphin_emphasis>"));
   timing_group_layout->addWidget(correct_time_drift);
 
+  auto* const rush_frame_presentation =
+      // i18n: "Rush" is a verb
+      new ConfigBool{tr("Rush Frame Presentation"), Config::MAIN_RUSH_FRAME_PRESENTATION};
+  rush_frame_presentation->SetDescription(
+      tr("Limits throttling between input and frame output,"
+         " speeding through emulation to reach presentation,"
+         " displaying sooner, and thus reducing input latency."
+         "<br><br>This will generally make frame pacing worse."
+         "<br>This setting can work either with or without Immediately Present XFB."
+         "<br>An Audio Buffer Size of at least 80 ms is recommended to ensure full effect."
+         "<br><br><dolphin_emphasis>If unsure, leave this unchecked.</dolphin_emphasis>"));
+  timing_group_layout->addWidget(rush_frame_presentation);
+
   // Make all labels the same width, so that the sliders are aligned.
   const QFontMetrics font_metrics{font()};
   const int label_width = font_metrics.boundingRect(QStringLiteral(" 500% (000.00 VPS)")).width();

Source/Core/VideoCommon/Present.cpp

@@ -5,6 +5,7 @@
 
 #include "Common/ChunkFile.h"
 #include "Core/Config/GraphicsSettings.h"
+#include "Core/Config/MainSettings.h"
 #include "Core/CoreTiming.h"
 #include "Core/HW/VideoInterface.h"
 #include "Core/Host.h"
@@ -201,7 +202,13 @@ void Presenter::ViSwap(u32 xfb_addr, u32 fb_width, u32 fb_stride, u32 fb_height,
 
   if (!is_duplicate || !g_ActiveConfig.bSkipPresentingDuplicateXFBs)
   {
+    // If RushFramePresentation is enabled, ignore the proper time to present as soon as possible.
+    // The goal is to achieve the lowest possible input latency.
+    if (Config::Get(Config::MAIN_RUSH_FRAME_PRESENTATION))
+      Present();
+    else
       Present(presentation_time);
+
     ProcessFrameDumping(ticks);
 
     video_events.after_present_event.Trigger(present_info);