PerformanceTracker: Use SPSCQueue and atomic to eliminate need for a mutex. Clean up some math.

Source/Core/VideoCommon/PerformanceMetrics.cpp

@@ -48,6 +48,7 @@ void PerformanceMetrics::CountPerformanceMarker(Core::System& system, s64 cycles
 {
   std::unique_lock lock(m_time_lock);
   m_speed_counter.Count();
+  m_speed_counter.UpdateStats();
 
   m_real_times[m_time_index] = Clock::now() - m_time_sleeping;
   m_cpu_times[m_time_index] = system.GetCoreTiming().GetCPUTimePoint(cyclesLate);
@@ -84,6 +85,9 @@ double PerformanceMetrics::GetLastSpeedDenominator() const
 
 void PerformanceMetrics::DrawImGuiStats(const float backbuffer_scale)
 {
+  m_vps_counter.UpdateStats();
+  m_fps_counter.UpdateStats();
+
   const int movable_flag = Config::Get(Config::GFX_MOVABLE_PERFORMANCE_METRICS) ?
                                ImGuiWindowFlags_None :
                                ImGuiWindowFlags_NoMove;

Source/Core/VideoCommon/PerformanceTracker.cpp

@@ -6,12 +6,12 @@
 #include <algorithm>
 #include <cmath>
 #include <iomanip>
-#include <mutex>
 
 #include <implot.h>
 
 #include "Common/CommonTypes.h"
 #include "Common/FileUtil.h"
+#include "Common/MathUtil.h"
 #include "Core/Core.h"
 #include "VideoCommon/VideoConfig.h"
 
@@ -21,14 +21,11 @@ static constexpr u64 MAX_QUALITY_GRAPH_SIZE = 1UL << 8;
 
 PerformanceTracker::PerformanceTracker(const std::optional<std::string> log_name,
                                        const std::optional<DT> sample_window_duration)
-    : m_on_state_changed_handle{Core::AddOnStateChangedCallback([this](Core::State state) {
+    : m_log_name{log_name}, m_sample_window_duration{sample_window_duration}
-        if (state == Core::State::Paused)
-          SetPaused(true);
-        else if (state == Core::State::Running)
-          SetPaused(false);
-      })},
-      m_log_name{log_name}, m_sample_window_duration{sample_window_duration}
 {
+  m_on_state_changed_handle =
+      Core::AddOnStateChangedCallback([this](Core::State state) { m_is_last_time_sane = false; });
+
   Reset();
 }
 
@@ -39,112 +36,110 @@ PerformanceTracker::~PerformanceTracker()
 
 void PerformanceTracker::Reset()
 {
-  std::unique_lock lock{m_mutex};
+  m_raw_dts.Clear();
+  m_dt_queue.clear();
 
   m_dt_total = DT::zero();
-  m_dt_queue.clear();
+  m_last_raw_dt = DT::zero();
   m_last_time = Clock::now();
   m_hz_avg = 0.0;
   m_dt_avg = DT::zero();
-  m_dt_std = std::nullopt;
+  m_dt_std = DT::zero();
+  m_is_last_time_sane = false;
 }
 
 void PerformanceTracker::Count()
 {
-  std::unique_lock lock{m_mutex};
+  const TimePoint current_time{Clock::now()};
 
-  if (m_paused)
+  const DT diff{current_time - m_last_time};
+  m_last_time = current_time;
+
+  if (!m_is_last_time_sane)
+  {
+    m_is_last_time_sane = true;
     return;
+  }
 
-  const DT window{GetSampleWindow()};
+  m_last_raw_dt = diff;
+  m_raw_dts.Push(diff);
+}
 
-  const TimePoint time{Clock::now()};
+void PerformanceTracker::UpdateStats()
-  const DT diff{time - m_last_time};
+{
+  DT diff{};
+  while (m_raw_dts.Pop(diff))
+    HandleRawDt(diff);
 
-  m_last_time = time;
+  // Update Std Dev
+  MathUtil::RunningVariance<double> variance;
+  for (auto dt : m_dt_queue)
+    variance.Push(DT_s(dt).count());
+  m_dt_std = std::chrono::duration_cast<DT>(DT_s(variance.PopulationStandardDeviation()));
+}
+
+void PerformanceTracker::HandleRawDt(DT diff)
+{
+  if (m_dt_queue.size() == MAX_DT_QUEUE_SIZE)
+    PopBack();
 
   PushFront(diff);
 
-  if (m_dt_queue.size() == MAX_DT_QUEUE_SIZE)
+  const DT window{GetSampleWindow()};
-    PopBack();
 
   while (m_dt_total - m_dt_queue.back() >= window)
     PopBack();
 
   // Simple Moving Average Throughout the Window
-  m_dt_avg = m_dt_total / m_dt_queue.size();
+  const DT dt_avg = m_dt_total / m_dt_queue.size();
-  const double hz = DT_s(1.0) / m_dt_avg;
+  const double hz = DT_s(1.0) / dt_avg;
+  m_dt_avg = dt_avg;
 
   // Exponential Moving Average
   const DT_s rc = SAMPLE_RC_RATIO * std::min(window, m_dt_total);
   const double a = 1.0 - std::exp(-(DT_s(diff) / rc));
 
   // Sometimes euler averages can break when the average is inf/nan
-  if (std::isfinite(m_hz_avg))
+  const auto hz_avg = m_hz_avg.load();
-    m_hz_avg += a * (hz - m_hz_avg);
+  if (std::isfinite(hz_avg))
+    m_hz_avg = hz_avg + a * (hz - hz_avg);
   else
     m_hz_avg = hz;
 
-  m_dt_std = std::nullopt;
-
   LogRenderTimeToFile(diff);
 }
 
 DT PerformanceTracker::GetSampleWindow() const
 {
-  // This reads a constant value and thus does not need a mutex
   return m_sample_window_duration.value_or(
       duration_cast<DT>(DT_us{std::max(1, g_ActiveConfig.iPerfSampleUSec)}));
 }
 
 double PerformanceTracker::GetHzAvg() const
 {
-  std::shared_lock lock{m_mutex};
   return m_hz_avg;
 }
 
 DT PerformanceTracker::GetDtAvg() const
 {
-  std::shared_lock lock{m_mutex};
   return m_dt_avg;
 }
 
 DT PerformanceTracker::GetDtStd() const
 {
-  std::unique_lock lock{m_mutex};
+  return m_dt_std;
-
-  if (m_dt_std)
-    return *m_dt_std;
-
-  if (m_dt_queue.empty())
-    return *(m_dt_std = DT::zero());
-
-  double total = 0.0;
-  for (auto dt : m_dt_queue)
-  {
-    double diff = DT_s(dt - m_dt_avg).count();
-    total += diff * diff;
-  }
-
-  // This is a weighted standard deviation
-  return *(m_dt_std = std::chrono::duration_cast<DT>(DT_s(std::sqrt(total / m_dt_queue.size()))));
 }
 
 DT PerformanceTracker::GetLastRawDt() const
 {
-  std::shared_lock lock{m_mutex};
+  return m_last_raw_dt;
-
-  if (m_dt_queue.empty())
-    return DT::zero();
-
-  return m_dt_queue.front();
 }
 
 void PerformanceTracker::ImPlotPlotLines(const char* label) const
 {
-  static std::array<float, MAX_DT_QUEUE_SIZE + 2> x, y;
+  // "quality" graph uses twice as many points.
-
+  static_assert(MAX_QUALITY_GRAPH_SIZE * 2 <= MAX_DT_QUEUE_SIZE);
-  std::shared_lock lock{m_mutex};
+  static std::array<float, MAX_DT_QUEUE_SIZE + 1> x, y;
 
   if (m_dt_queue.empty())
     return;
@@ -152,38 +147,32 @@ void PerformanceTracker::ImPlotPlotLines(const char* label) const
   // Decides if there are too many points to plot using rectangles
   const bool quality = m_dt_queue.size() < MAX_QUALITY_GRAPH_SIZE;
 
-  const DT update_time = Clock::now() - m_last_time;
+  std::size_t point_index = 0;
-  const float predicted_frame_time = DT_ms(std::max(update_time, m_dt_queue.front())).count();
+  const auto add_point = [&](DT dt, DT shift_x, float prev_ms) {
+    const float ms = DT_ms{dt}.count();
 
-  std::size_t points = 0;
     if (quality)
     {
-    x[points] = 0.f;
+      x[point_index] = prev_ms;
-    y[points] = predicted_frame_time;
+      y[point_index] = ms;
-    ++points;
+      ++point_index;
     }
 
-  x[points] = DT_ms(update_time).count();
+    x[point_index] = prev_ms + DT_ms{shift_x}.count();
-  y[points] = predicted_frame_time;
+    y[point_index] = ms;
-  ++points;
+    ++point_index;
+  };
 
+  // Rightmost point.
+  const auto update_time = Clock::now() - m_last_time;
+  const auto predicted_frame_time = std::max(update_time, m_dt_queue.front());
+  add_point(predicted_frame_time, DT{}, 0);
+
+  // Other points, right to left.
   for (auto dt : m_dt_queue)
-  {
+    add_point(dt, dt, x[point_index - 1]);
-    const float frame_time_ms = DT_ms(dt).count();
 
-    if (quality)
+  ImPlot::PlotLine(label, x.data(), y.data(), static_cast<int>(point_index));
-    {
-      x[points] = x[points - 1];
-      y[points] = frame_time_ms;
-      ++points;
-    }
-
-    x[points] = x[points - 1] + frame_time_ms;
-    y[points] = frame_time_ms;
-    ++points;
-  }
-
-  ImPlot::PlotLine(label, x.data(), y.data(), static_cast<int>(points));
 }
 
 void PerformanceTracker::PushFront(DT value)
@@ -211,18 +200,3 @@ void PerformanceTracker::LogRenderTimeToFile(DT val)
 
   m_bench_file << std::fixed << std::setprecision(8) << DT_ms(val).count() << std::endl;
 }
-
-void PerformanceTracker::SetPaused(bool paused)
-{
-  std::unique_lock lock{m_mutex};
-
-  m_paused = paused;
-  if (m_paused)
-  {
-    m_last_time = TimePoint::max();
-  }
-  else
-  {
-    m_last_time = Clock::now();
-  }
-}

Source/Core/VideoCommon/PerformanceTracker.h

@@ -3,12 +3,13 @@
 
 #pragma once
 
+#include <atomic>
 #include <deque>
 #include <fstream>
 #include <optional>
-#include <shared_mutex>
 
 #include "Common/CommonTypes.h"
+#include "Common/SPSCQueue.h"
 
 class PerformanceTracker
 {
@@ -22,30 +23,30 @@ public:
   PerformanceTracker(PerformanceTracker&&) = delete;
   PerformanceTracker& operator=(PerformanceTracker&&) = delete;
 
-  // Functions for recording performance information
   void Reset();
+
+  // Calls must come from the same thread.
+  // UpdateStats is expected to be called regularly to empty the SPSC queue.
+  void UpdateStats();
+  void ImPlotPlotLines(const char* label) const;
+
+  // May call from any thread, but not concurrently, not that you'd want to..
   void Count();
 
-  // Functions for reading performance information
+  // May call from any thread.
   DT GetSampleWindow() const;
-
   double GetHzAvg() const;
-
   DT GetDtAvg() const;
   DT GetDtStd() const;
-
   DT GetLastRawDt() const;
 
-  void ImPlotPlotLines(const char* label) const;
-
 private:
   void LogRenderTimeToFile(DT val);
-  void SetPaused(bool paused);
 
+  void HandleRawDt(DT value);
   void PushFront(DT value);
   void PopBack();
 
-  bool m_paused = false;
   int m_on_state_changed_handle;
 
   // Name of log file and file stream
@@ -54,6 +55,12 @@ private:
 
   // Last time Count() was called
   TimePoint m_last_time;
+  std::atomic<bool> m_is_last_time_sane = false;
+
+  // Push'd from Count()
+  //  and Pop'd from UpdateStats()
+  Common::SPSCQueue<DT, false> m_raw_dts;
+  std::atomic<DT> m_last_raw_dt = DT::zero();
 
   // Amount of time to sample dt's over (defaults to config)
   const std::optional<DT> m_sample_window_duration;
@@ -63,12 +70,7 @@ private:
   std::deque<DT> m_dt_queue;
 
   // Average rate/time throughout the window
-  DT m_dt_avg = DT::zero();  // Uses Moving Average
+  std::atomic<DT> m_dt_avg = DT::zero();  // Uses Moving Average
-  double m_hz_avg = 0.0;     // Uses Moving Average + Euler Average
+  std::atomic<double> m_hz_avg = 0.0;     // Uses Moving Average + Euler Average
-
+  std::atomic<DT> m_dt_std = DT::zero();
-  // Used to initialize this on demand instead of on every Count()
-  mutable std::optional<DT> m_dt_std = std::nullopt;
-
-  // Used to enable thread safety with the performance tracker
-  mutable std::shared_mutex m_mutex;
 };