CoreTiming: Throttle Before Every Event Using Chrono

7 files changed, 225 insertions, 140 deletions

diff --git a/Source/Core/AudioCommon/Mixer.cpp b/Source/Core/AudioCommon/Mixer.cpp
index f71c1e0c83..ae52dc7bb2 100644
--- a/Source/Core/AudioCommon/Mixer.cpp
+++ b/Source/Core/AudioCommon/Mixer.cpp
@@ -162,7 +162,7 @@ unsigned int Mixer::Mix(short* samples, unsigned int num_samples)
   memset(samples, 0, num_samples * 2 * sizeof(short));
 
   // TODO: Determine how emulation speed will be used in audio
-  // const float emulation_speed = std::roundf(g_perf_metrics.GetSpeed()) / 100.f;
+  // const float emulation_speed = g_perf_metrics.GetSpeed();
   const float emulation_speed = m_config_emulation_speed;
   const int timing_variance = m_config_timing_variance;
   if (m_config_audio_stretch)
diff --git a/Source/Core/Core/Core.cpp b/Source/Core/Core/Core.cpp
index 332cc00857..59306619e1 100644
--- a/Source/Core/Core/Core.cpp
+++ b/Source/Core/Core/Core.cpp
@@ -891,9 +891,9 @@ void Callback_NewField()
 
 void UpdateTitle()
 {
-  float FPS = g_perf_metrics.GetFPS();
-  float VPS = g_perf_metrics.GetVPS();
-  float Speed = g_perf_metrics.GetSpeed();
+  const double FPS = g_perf_metrics.GetFPS();
+  const double VPS = g_perf_metrics.GetVPS();
+  const double Speed = 100.0 * g_perf_metrics.GetSpeed();
 
   // Settings are shown the same for both extended and summary info
   const std::string SSettings = fmt::format(
diff --git a/Source/Core/Core/CoreTiming.cpp b/Source/Core/Core/CoreTiming.cpp
index 1d4f085b99..d85db5288d 100644
--- a/Source/Core/Core/CoreTiming.cpp
+++ b/Source/Core/Core/CoreTiming.cpp
@@ -22,6 +22,7 @@
 #include "Core/System.h"
 
 #include "VideoCommon/Fifo.h"
+#include "VideoCommon/PerformanceMetrics.h"
 #include "VideoCommon/VideoBackendBase.h"
 
 namespace CoreTiming
@@ -305,6 +306,8 @@ void CoreTimingManager::Advance()
     Event evt = std::move(m_event_queue.front());
     std::pop_heap(m_event_queue.begin(), m_event_queue.end(), std::greater<Event>());
     m_event_queue.pop_back();
+
+    Throttle(evt.time);
     evt.type->callback(system, evt.userdata, m_globals.global_timer - evt.time);
   }
 
@@ -326,6 +329,58 @@ void CoreTimingManager::Advance()
   PowerPC::CheckExternalExceptions();
 }
 
+void CoreTimingManager::Throttle(const s64 target_cycle)
+{
+  if (target_cycle <= m_throttle_last_cycle)
+    return;
+
+  const double speed =
+      Core::GetIsThrottlerTempDisabled() ? 0.0 : Config::Get(Config::MAIN_EMULATION_SPEED);
+
+  // Based on number of cycles and emulation speed, increase the target deadline
+  const s64 cycles = target_cycle - m_throttle_last_cycle;
+  m_throttle_last_cycle = target_cycle;
+
+  if (0.0 < speed)
+    m_throttle_deadline += std::chrono::duration_cast<DT>((m_throttle_per_clock * cycles) / speed);
+
+  // 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.
+  const DT max_fallback =
+      std::chrono::duration_cast<DT>(DT_ms(Config::Get(Config::MAIN_TIMING_VARIANCE)));
+
+  const TimePoint time = Clock::now();
+  const TimePoint min_deadline = time - max_fallback;
+  const TimePoint max_deadline = time + max_fallback;
+
+  if (m_throttle_deadline > max_deadline)
+  {
+    m_throttle_deadline = max_deadline;
+  }
+  else if (m_throttle_deadline < min_deadline)
+  {
+    DEBUG_LOG_FMT(COMMON, "System can not to keep up with timings! [relaxing timings by {} us]",
+                  DT_us(min_deadline - m_throttle_deadline).count());
+    m_throttle_deadline = min_deadline;
+  }
+
+  // Only sleep if we are behind the deadline
+  if (time < m_throttle_deadline)
+  {
+    std::this_thread::sleep_until(m_throttle_deadline);
+
+    // Count amount of time sleeping for analytics
+    const TimePoint time_after_sleep = Clock::now();
+    g_perf_metrics.CountThrottleSleep(time_after_sleep - time);
+  }
+}
+
+TimePoint CoreTimingManager::GetCPUTimePoint(s64 cyclesLate) const
+{
+  return TimePoint(
+      std::chrono::duration_cast<DT>(m_throttle_per_clock * (m_globals.global_timer - cyclesLate)));
+}
+
 void CoreTimingManager::LogPendingEvents() const
 {
   auto clone = m_event_queue;
@@ -340,6 +395,8 @@ void CoreTimingManager::LogPendingEvents() const
 // Should only be called from the CPU thread after the PPC clock has changed
 void CoreTimingManager::AdjustEventQueueTimes(u32 new_ppc_clock, u32 old_ppc_clock)
 {
+  m_throttle_per_clock = DT_s(1.0) / new_ppc_clock;
+
   for (Event& ev : m_event_queue)
   {
     const s64 ticks = (ev.time - m_globals.global_timer) * new_ppc_clock / old_ppc_clock;
diff --git a/Source/Core/Core/CoreTiming.h b/Source/Core/Core/CoreTiming.h
index 737f3889b9..a10525cbbb 100644
--- a/Source/Core/Core/CoreTiming.h
+++ b/Source/Core/Core/CoreTiming.h
@@ -140,6 +140,8 @@ public:
   // Directly accessed by the JIT.
   Globals& GetGlobals() { return m_globals; }
 
+  TimePoint GetCPUTimePoint(s64 cyclesLate) const;  // Used by Dolphin Analytics
+
 private:
   Globals m_globals;
 
@@ -173,6 +175,12 @@ private:
   float m_config_oc_inv_factor = 0.0f;
   bool m_config_sync_on_skip_idle = false;
 
+  s64 m_throttle_last_cycle = 0;
+  TimePoint m_throttle_deadline = Clock::now();
+  DT_s m_throttle_per_clock = DT_s();
+
+  void Throttle(const s64 target_cycle);
+
   int DowncountToCycles(int downcount) const;
   int CyclesToDowncount(int cycles) const;
 };
diff --git a/Source/Core/Core/HW/SystemTimers.cpp b/Source/Core/Core/HW/SystemTimers.cpp
index d7908f10e1..de9dc729ae 100644
--- a/Source/Core/Core/HW/SystemTimers.cpp
+++ b/Source/Core/Core/HW/SystemTimers.cpp
@@ -67,6 +67,7 @@ IPC_HLE_PERIOD: For the Wii Remote this is the call schedule:
 #include "Core/PowerPC/PowerPC.h"
 #include "Core/System.h"
 #include "VideoCommon/Fifo.h"
+#include "VideoCommon/PerformanceMetrics.h"
 
 namespace SystemTimers
 {
@@ -77,9 +78,10 @@ CoreTiming::EventType* et_VI;
 CoreTiming::EventType* et_AudioDMA;
 CoreTiming::EventType* et_DSP;
 CoreTiming::EventType* et_IPC_HLE;
+CoreTiming::EventType* et_GPU_sleeper;
+CoreTiming::EventType* et_perf_tracker;
 // PatchEngine updates every 1/60th of a second by default
 CoreTiming::EventType* et_PatchEngine;
-CoreTiming::EventType* et_Throttle;
 
 u32 s_cpu_core_clock = 486000000u;  // 486 mhz (its not 485, stop bugging me!)
 
@@ -90,16 +92,6 @@ int s_ipc_hle_period;
 // Custom RTC
 s64 s_localtime_rtc_offset = 0;
 
-// For each emulated milliseconds, what was the real time timestamp (excluding sleep time). This is
-// a "special" ring buffer where we only need to read the first and last value.
-std::array<u64, 1000> s_emu_to_real_time_ring_buffer;
-size_t s_emu_to_real_time_index;
-std::mutex s_emu_to_real_time_mutex;
-
-// How much time was spent sleeping since the emulator started. Note: this does not need to be reset
-// at initialization (or ever), since only the "derivative" of that value really matters.
-u64 s_time_spent_sleeping;
-
 // DSP/CPU timeslicing.
 void DSPCallback(Core::System& system, u64 userdata, s64 cyclesLate)
 {
@@ -132,6 +124,26 @@ void IPC_HLE_UpdateCallback(Core::System& system, u64 userdata, s64 cyclesLate)
   }
 }
 
+void GPUSleepCallback(Core::System& system, u64 userdata, s64 cyclesLate)
+{
+  auto& core_timing = system.GetCoreTiming();
+  system.GetFifo().GpuMaySleep();
+
+  // We want to call GpuMaySleep at about 1000hz so
+  // that the thread can sleep while not doing anything.
+  core_timing.ScheduleEvent(GetTicksPerSecond() / 1000 - cyclesLate, et_GPU_sleeper);
+}
+
+void PerfTrackerCallback(Core::System& system, u64 userdata, s64 cyclesLate)
+{
+  auto& core_timing = system.GetCoreTiming();
+  g_perf_metrics.CountPerformanceMarker(system, cyclesLate);
+
+  // Call this performance tracker again in 1/64th of a second.
+  // The tracker stores 256 values so this will let us summarize the last 4 seconds.
+  core_timing.ScheduleEvent(GetTicksPerSecond() / 64 - cyclesLate, et_perf_tracker);
+}
+
 void VICallback(Core::System& system, u64 userdata, s64 cyclesLate)
 {
   auto& core_timing = system.GetCoreTiming();
@@ -169,50 +181,6 @@ void PatchEngineCallback(Core::System& system, u64 userdata, s64 cycles_late)
 
   system.GetCoreTiming().ScheduleEvent(next_schedule, et_PatchEngine, cycles_pruned);
 }
-
-void ThrottleCallback(Core::System& system, u64 deadline, s64 cyclesLate)
-{
-  // Allow the GPU thread to sleep. Setting this flag here limits the wakeups to 1 kHz.
-  system.GetFifo().GpuMaySleep();
-
-  const u64 time = Common::Timer::NowUs();
-
-  if (deadline == 0)
-    deadline = time;
-
-  const s64 diff = deadline - time;
-  const float emulation_speed = Config::Get(Config::MAIN_EMULATION_SPEED);
-  const bool frame_limiter = emulation_speed > 0.0f && !Core::GetIsThrottlerTempDisabled();
-  u32 next_event = GetTicksPerSecond() / 1000;
-
-  {
-    std::lock_guard lk(s_emu_to_real_time_mutex);
-    s_emu_to_real_time_ring_buffer[s_emu_to_real_time_index] = time - s_time_spent_sleeping;
-    s_emu_to_real_time_index =
-        (s_emu_to_real_time_index + 1) % s_emu_to_real_time_ring_buffer.size();
-  }
-
-  if (frame_limiter)
-  {
-    if (emulation_speed != 1.0f)
-      next_event = u32(next_event * emulation_speed);
-    const s64 max_fallback = Config::Get(Config::MAIN_TIMING_VARIANCE) * 1000;
-    if (std::abs(diff) > max_fallback)
-    {
-      DEBUG_LOG_FMT(COMMON, "system too {}, {} us skipped", diff < 0 ? "slow" : "fast",
-                    std::abs(diff) - max_fallback);
-      deadline = time - max_fallback;
-    }
-    else if (diff > 1000)
-    {
-      Common::SleepCurrentThread(diff / 1000);
-      s_time_spent_sleeping += Common::Timer::NowUs() - time;
-    }
-  }
-  // reschedule 1ms (possibly scaled by emulation_speed) into future on ppc
-  // add 1ms to the deadline
-  system.GetCoreTiming().ScheduleEvent(next_event - cyclesLate, et_Throttle, deadline + 1000);
-}
 }  // namespace
 
 u32 GetTicksPerSecond()
@@ -268,27 +236,7 @@ s64 GetLocalTimeRTCOffset()
 
 double GetEstimatedEmulationPerformance()
 {
-  u64 ts_now, ts_before;  // In microseconds
-  {
-    std::lock_guard lk(s_emu_to_real_time_mutex);
-    size_t index_now = s_emu_to_real_time_index == 0 ? s_emu_to_real_time_ring_buffer.size() - 1 :
-                                                       s_emu_to_real_time_index - 1;
-    size_t index_before = s_emu_to_real_time_index;
-
-    ts_now = s_emu_to_real_time_ring_buffer[index_now];
-    ts_before = s_emu_to_real_time_ring_buffer[index_before];
-  }
-
-  if (ts_before == 0)
-  {
-    // Not enough data yet to estimate. We could technically provide an estimate based on a shorter
-    // time horizon, but it's not really worth it.
-    return 1.0;
-  }
-
-  u64 delta_us = ts_now - ts_before;
-  double emulated_us = s_emu_to_real_time_ring_buffer.size() * 1000.0;  // For each emulated ms.
-  return delta_us == 0 ? DBL_MAX : emulated_us / delta_us;
+  return g_perf_metrics.GetMaxSpeed();
 }
 
 // split from Init to break a circular dependency between VideoInterface::Init and
@@ -345,20 +293,20 @@ void Init()
   et_DSP = core_timing.RegisterEvent("DSPCallback", DSPCallback);
   et_AudioDMA = core_timing.RegisterEvent("AudioDMACallback", AudioDMACallback);
   et_IPC_HLE = core_timing.RegisterEvent("IPC_HLE_UpdateCallback", IPC_HLE_UpdateCallback);
+  et_GPU_sleeper = core_timing.RegisterEvent("GPUSleeper", GPUSleepCallback);
+  et_perf_tracker = core_timing.RegisterEvent("PerfTracker", PerfTrackerCallback);
   et_PatchEngine = core_timing.RegisterEvent("PatchEngine", PatchEngineCallback);
-  et_Throttle = core_timing.RegisterEvent("Throttle", ThrottleCallback);
 
+  core_timing.ScheduleEvent(0, et_perf_tracker);
+  core_timing.ScheduleEvent(0, et_GPU_sleeper);
   core_timing.ScheduleEvent(VideoInterface::GetTicksPerHalfLine(), et_VI);
   core_timing.ScheduleEvent(0, et_DSP);
   core_timing.ScheduleEvent(GetAudioDMACallbackPeriod(), et_AudioDMA);
-  core_timing.ScheduleEvent(0, et_Throttle, 0);
 
   core_timing.ScheduleEvent(VideoInterface::GetTicksPerField(), et_PatchEngine);
 
   if (SConfig::GetInstance().bWii)
     core_timing.ScheduleEvent(s_ipc_hle_period, et_IPC_HLE);
-
-  s_emu_to_real_time_ring_buffer.fill(0);
 }
 
 void Shutdown()
diff --git a/Source/Core/VideoCommon/PerformanceMetrics.cpp b/Source/Core/VideoCommon/PerformanceMetrics.cpp
index 23e35623cb..b343af0f4f 100644
--- a/Source/Core/VideoCommon/PerformanceMetrics.cpp
+++ b/Source/Core/VideoCommon/PerformanceMetrics.cpp
@@ -3,10 +3,14 @@
 
 #include "VideoCommon/PerformanceMetrics.h"
 
+#include <mutex>
+
 #include <imgui.h>
 #include <implot.h>
 
+#include "Core/CoreTiming.h"
 #include "Core/HW/VideoInterface.h"
+#include "Core/System.h"
 #include "VideoCommon/VideoConfig.h"
 
 PerformanceMetrics g_perf_metrics;
@@ -16,6 +20,10 @@ void PerformanceMetrics::Reset()
   m_fps_counter.Reset();
   m_vps_counter.Reset();
   m_speed_counter.Reset();
+
+  m_time_sleeping = DT::zero();
+  m_real_times.fill(Clock::now());
+  m_cpu_times.fill(Core::System::GetInstance().GetCoreTiming().GetCPUTimePoint(0));
 }
 
 void PerformanceMetrics::CountFrame()
@@ -29,6 +37,20 @@ void PerformanceMetrics::CountVBlank()
   m_speed_counter.Count();
 }
 
+void PerformanceMetrics::CountThrottleSleep(DT sleep)
+{
+  std::unique_lock lock(m_time_lock);
+  m_time_sleeping += sleep;
+}
+
+void PerformanceMetrics::CountPerformanceMarker(Core::System& system, s64 cyclesLate)
+{
+  std::unique_lock lock(m_time_lock);
+  m_real_times[m_time_index] = Clock::now() - m_time_sleeping;
+  m_cpu_times[m_time_index] = system.GetCoreTiming().GetCPUTimePoint(cyclesLate);
+  m_time_index += 1;
+}
+
 double PerformanceMetrics::GetFPS() const
 {
   return m_fps_counter.GetHzAvg();
@@ -41,7 +63,14 @@ double PerformanceMetrics::GetVPS() const
 
 double PerformanceMetrics::GetSpeed() const
 {
-  return 100.0 * m_speed_counter.GetHzAvg() / VideoInterface::GetTargetRefreshRate();
+  return m_speed_counter.GetHzAvg() / VideoInterface::GetTargetRefreshRate();
+}
+
+double PerformanceMetrics::GetMaxSpeed() const
+{
+  std::shared_lock lock(m_time_lock);
+  return DT_s(m_cpu_times[u8(m_time_index - 1)] - m_cpu_times[m_time_index]) /
+         DT_s(m_real_times[u8(m_time_index - 1)] - m_real_times[m_time_index]);
 }
 
 double PerformanceMetrics::GetLastSpeedDenominator() const
@@ -49,7 +78,7 @@ double PerformanceMetrics::GetLastSpeedDenominator() const
   return DT_s(m_speed_counter.GetLastRawDt()).count() * VideoInterface::GetTargetRefreshRate();
 }
 
-void PerformanceMetrics::DrawImGuiStats(const float backbuffer_scale) const
+void PerformanceMetrics::DrawImGuiStats(const float backbuffer_scale)
 {
   const float bg_alpha = 0.7f;
   const auto imgui_flags = ImGuiWindowFlags_NoTitleBar | ImGuiWindowFlags_NoInputs |
@@ -65,9 +94,9 @@ void PerformanceMetrics::DrawImGuiStats(const float backbuffer_scale) const
   float r = 0.0f, g = 1.0f, b = 1.0f;
   if (g_ActiveConfig.bShowSpeedColors)
   {
-    r = 1.0 - (speed - 80.0) / 20.0;
-    g = speed / 80.0;
-    b = (speed - 90.0) / 10.0;
+    r = 1.0 - (speed - 0.8) / 0.2;
+    g = speed / 0.8;
+    b = (speed - 0.9) / 0.1;
   }
 
   const float window_padding = 8.f * backbuffer_scale;
@@ -79,6 +108,8 @@ void PerformanceMetrics::DrawImGuiStats(const float backbuffer_scale) const
   const float graph_height =
       std::min(200.f * backbuffer_scale, ImGui::GetIO().DisplaySize.y - 85.f * backbuffer_scale);
 
+  const bool stack_vertically = !g_ActiveConfig.bShowGraphs;
+
   ImGui::PushStyleVar(ImGuiStyleVar_WindowBorderSize, 0.f);
   ImGui::PushStyleVar(ImGuiStyleVar_WindowRounding, 14.f * backbuffer_scale);
   if (g_ActiveConfig.bShowGraphs)
@@ -93,40 +124,39 @@ void PerformanceMetrics::DrawImGuiStats(const float backbuffer_scale) const
 
     if (ImGui::Begin("PerformanceGraphs", nullptr, imgui_flags))
     {
-      const static int num_ticks = 17;
-      const static double tick_marks[num_ticks] = {0.0,
-                                                   1000.0 / 360.0,
-                                                   1000.0 / 240.0,
-                                                   1000.0 / 180.0,
-                                                   1000.0 / 120.0,
-                                                   1000.0 / 90.00,
-                                                   1000.0 / 59.94,
-                                                   1000.0 / 40.00,
-                                                   1000.0 / 29.97,
-                                                   1000.0 / 24.00,
-                                                   1000.0 / 20.00,
-                                                   1000.0 / 15.00,
-                                                   1000.0 / 10.00,
-                                                   1000.0 / 5.000,
-                                                   1000.0 / 2.000,
-                                                   1000.0,
-                                                   2000.0};
-
-      const DT vblank_time = m_vps_counter.GetDtAvg() + m_vps_counter.GetDtStd();
-      const DT frame_time = m_fps_counter.GetDtAvg() + m_fps_counter.GetDtStd();
+      static constexpr std::size_t num_ticks = 17;
+      static constexpr std::array<double, num_ticks> tick_marks = {0.0,
+                                                                   1000.0 / 360.0,
+                                                                   1000.0 / 240.0,
+                                                                   1000.0 / 180.0,
+                                                                   1000.0 / 120.0,
+                                                                   1000.0 / 90.00,
+                                                                   1000.0 / 59.94,
+                                                                   1000.0 / 40.00,
+                                                                   1000.0 / 29.97,
+                                                                   1000.0 / 24.00,
+                                                                   1000.0 / 20.00,
+                                                                   1000.0 / 15.00,
+                                                                   1000.0 / 10.00,
+                                                                   1000.0 / 5.000,
+                                                                   1000.0 / 2.000,
+                                                                   1000.0,
+                                                                   2000.0};
+
+      const DT vblank_time = m_vps_counter.GetDtAvg() + 2 * m_vps_counter.GetDtStd();
+      const DT frame_time = m_fps_counter.GetDtAvg() + 2 * m_fps_counter.GetDtStd();
       const double target_max_time = DT_ms(vblank_time + frame_time).count();
       const double a =
-          std::max(0.0, 1.0 - std::exp(-4000.0 * m_vps_counter.GetLastRawDt().count() /
-                                       DT_ms(m_vps_counter.GetSampleWindow()).count()));
+          std::max(0.0, 1.0 - std::exp(-4.0 * (DT_s(m_vps_counter.GetLastRawDt()) /
+                                               DT_s(m_vps_counter.GetSampleWindow()))));
 
-      static double max_time = 0.0;
-      if (std::isfinite(max_time))
-        max_time += a * (target_max_time - max_time);
+      if (std::isfinite(m_graph_max_time))
+        m_graph_max_time += a * (target_max_time - m_graph_max_time);
       else
-        max_time = target_max_time;
+        m_graph_max_time = target_max_time;
 
-      const double total_frame_time = std::max(DT_ms(m_fps_counter.GetSampleWindow()).count(),
-                                               DT_ms(m_vps_counter.GetSampleWindow()).count());
+      const double total_frame_time =
+          DT_ms(std::max(m_fps_counter.GetSampleWindow(), m_vps_counter.GetSampleWindow())).count();
 
       if (ImPlot::BeginPlot("PerformanceGraphs", ImVec2(-1.0, -1.0),
                             ImPlotFlags_NoFrame | ImPlotFlags_NoTitle | ImPlotFlags_NoMenus))
@@ -141,8 +171,8 @@ void PerformanceMetrics::DrawImGuiStats(const float backbuffer_scale) const
                           ImPlotAxisFlags_Lock | ImPlotAxisFlags_Invert | ImPlotAxisFlags_NoLabel |
                               ImPlotAxisFlags_NoHighlight);
         ImPlot::SetupAxisFormat(ImAxis_Y1, "%.1f");
-        ImPlot::SetupAxisTicks(ImAxis_Y1, tick_marks, num_ticks);
-        ImPlot::SetupAxesLimits(0, total_frame_time, 0, max_time, ImGuiCond_Always);
+        ImPlot::SetupAxisTicks(ImAxis_Y1, tick_marks.data(), num_ticks);
+        ImPlot::SetupAxesLimits(0, total_frame_time, 0, m_graph_max_time, ImGuiCond_Always);
         ImPlot::SetupLegend(ImPlotLocation_SouthEast, ImPlotLegendFlags_None);
         m_vps_counter.ImPlotPlotLines("V-Blank (ms)");
         m_fps_counter.ImPlotPlotLines("Frame (ms)");
@@ -155,14 +185,42 @@ void PerformanceMetrics::DrawImGuiStats(const float backbuffer_scale) const
     }
   }
 
-  if (g_ActiveConfig.bShowFPS || g_ActiveConfig.bShowFTimes)
+  if (g_ActiveConfig.bShowSpeed)
   {
     // Position in the top-right corner of the screen.
+    float window_height = 47.f * backbuffer_scale;
+
+    ImGui::SetNextWindowPos(ImVec2(window_x, window_y), ImGuiCond_Always, ImVec2(1.0f, 0.0f));
+    ImGui::SetNextWindowSize(ImVec2(window_width, window_height));
+    ImGui::SetNextWindowBgAlpha(bg_alpha);
+
+    if (stack_vertically)
+      window_y += window_height + window_padding;
+    else
+      window_x -= window_width + window_padding;
+
+    if (ImGui::Begin("SpeedStats", nullptr, imgui_flags))
+    {
+      ImGui::TextColored(ImVec4(r, g, b, 1.0f), "Speed:%4.0lf%%", 100.0 * speed);
+      ImGui::TextColored(ImVec4(r, g, b, 1.0f), "Max:%6.0lf%%", 100.0 * GetMaxSpeed());
+      ImGui::End();
+    }
+  }
+
+  if (g_ActiveConfig.bShowFPS || g_ActiveConfig.bShowFTimes)
+  {
     int count = g_ActiveConfig.bShowFPS + 2 * g_ActiveConfig.bShowFTimes;
+    float window_height = (12.f + 17.f * count) * backbuffer_scale;
+
+    // Position in the top-right corner of the screen.
     ImGui::SetNextWindowPos(ImVec2(window_x, window_y), ImGuiCond_Always, ImVec2(1.0f, 0.0f));
-    ImGui::SetNextWindowSize(ImVec2(window_width, (12.f + 17.f * count) * backbuffer_scale));
+    ImGui::SetNextWindowSize(ImVec2(window_width, window_height));
     ImGui::SetNextWindowBgAlpha(bg_alpha);
-    window_x -= window_width + window_padding;
+
+    if (stack_vertically)
+      window_y += window_height + window_padding;
+    else
+      window_x -= window_width + window_padding;
 
     if (ImGui::Begin("FPSStats", nullptr, imgui_flags))
     {
@@ -181,12 +239,18 @@ void PerformanceMetrics::DrawImGuiStats(const float backbuffer_scale) const
 
   if (g_ActiveConfig.bShowVPS || g_ActiveConfig.bShowVTimes)
   {
-    // Position in the top-right corner of the screen.
     int count = g_ActiveConfig.bShowVPS + 2 * g_ActiveConfig.bShowVTimes;
+    float window_height = (12.f + 17.f * count) * backbuffer_scale;
+
+    // Position in the top-right corner of the screen.
     ImGui::SetNextWindowPos(ImVec2(window_x, window_y), ImGuiCond_Always, ImVec2(1.0f, 0.0f));
     ImGui::SetNextWindowSize(ImVec2(window_width, (12.f + 17.f * count) * backbuffer_scale));
     ImGui::SetNextWindowBgAlpha(bg_alpha);
-    window_x -= window_width + window_padding;
+
+    if (stack_vertically)
+      window_y += window_height + window_padding;
+    else
+      window_x -= window_width + window_padding;
 
     if (ImGui::Begin("VPSStats", nullptr, imgui_flags))
     {
@@ -203,19 +267,5 @@ void PerformanceMetrics::DrawImGuiStats(const float backbuffer_scale) const
     }
   }
 
-  if (g_ActiveConfig.bShowSpeed)
-  {
-    // Position in the top-right corner of the screen.
-    ImGui::SetNextWindowPos(ImVec2(window_x, window_y), ImGuiCond_Always, ImVec2(1.0f, 0.0f));
-    ImGui::SetNextWindowSize(ImVec2(window_width, 29.f * backbuffer_scale));
-    ImGui::SetNextWindowBgAlpha(bg_alpha);
-
-    if (ImGui::Begin("SpeedStats", nullptr, imgui_flags))
-    {
-      ImGui::TextColored(ImVec4(r, g, b, 1.0f), "Speed:%4.0lf%%", speed);
-      ImGui::End();
-    }
-  }
-
   ImGui::PopStyleVar(2);
 }
diff --git a/Source/Core/VideoCommon/PerformanceMetrics.h b/Source/Core/VideoCommon/PerformanceMetrics.h
index eb6f8ecb00..009c27fd3f 100644
--- a/Source/Core/VideoCommon/PerformanceMetrics.h
+++ b/Source/Core/VideoCommon/PerformanceMetrics.h
@@ -3,8 +3,17 @@
 
 #pragma once
 
+#include <array>
+#include <shared_mutex>
+
+#include "Common/CommonTypes.h"
 #include "VideoCommon/PerformanceTracker.h"
 
+namespace Core
+{
+class System;
+}
+
 class PerformanceMetrics
 {
 public:
@@ -21,20 +30,33 @@ public:
   void CountFrame();
   void CountVBlank();
 
+  void CountThrottleSleep(DT sleep);
+  void CountPerformanceMarker(Core::System& system, s64 cyclesLate);
+
   // Getter Functions
   double GetFPS() const;
   double GetVPS() const;
   double GetSpeed() const;
+  double GetMaxSpeed() const;
 
   double GetLastSpeedDenominator() const;
 
   // ImGui Functions
-  void DrawImGuiStats(const float backbuffer_scale) const;
+  void DrawImGuiStats(const float backbuffer_scale);
 
 private:
   PerformanceTracker m_fps_counter{"render_times.txt"};
   PerformanceTracker m_vps_counter{"vblank_times.txt"};
   PerformanceTracker m_speed_counter{std::nullopt, 500000};
+
+  double m_graph_max_time = 0.0;
+
+  mutable std::shared_mutex m_time_lock;
+
+  u8 m_time_index = 0;
+  std::array<TimePoint, 256> m_real_times;
+  std::array<TimePoint, 256> m_cpu_times;
+  DT m_time_sleeping;
 };
 
 extern PerformanceMetrics g_perf_metrics;