PerformanceTracker: Small cleanup, properly use RunningVariance for the standard deviation

UpdateStats() was calculating the standard deviation from the entirety of m_dt_queue, which has a noticeable performance impact when the frame rates are very high.
On a scene so light it runs at 3000 FPS, UpdateStats() was taking around 5% of the profiling (at default 1s sample window). Now it doesn't.

3 files changed, 73 insertions, 51 deletions

diff --git a/Source/Core/Common/MathUtil.h b/Source/Core/Common/MathUtil.h
index 47440f2239..b96e8f55fe 100644
--- a/Source/Core/Common/MathUtil.h
+++ b/Source/Core/Common/MathUtil.h
@@ -134,9 +134,19 @@ public:
   constexpr void Clear() { *this = {}; }
 
   constexpr void Push(T x) { m_mean = m_mean + (x - m_mean) / ++m_count; }
+  constexpr void Pop(T x)
+  {
+    if (m_count <= 1)
+    {
+      Clear();
+      return;
+    }
+    m_mean = m_mean - (x - m_mean) / --m_count;
+  }
 
   constexpr size_t Count() const { return m_count; }
   constexpr T Mean() const { return m_mean; }
+  constexpr T Total() const { return m_mean * m_count; }
 
 private:
   size_t m_count = 0;
@@ -155,9 +165,22 @@ public:
     m_running_mean.Push(x);
     m_variance += (x - old_mean) * (x - m_running_mean.Mean());
   }
+  constexpr void Pop(T x)
+  {
+    if (m_running_mean.Count() <= 1)
+    {
+      Clear();
+      return;
+    }
+
+    const T old_mean = m_running_mean.Mean();
+    m_running_mean.Pop(x);
+    m_variance -= (x - old_mean) * (x - m_running_mean.Mean());
+  }
 
   constexpr size_t Count() const { return m_running_mean.Count(); }
   constexpr T Mean() const { return m_running_mean.Mean(); }
+  constexpr T Total() const { return m_running_mean.Total(); }
 
   constexpr T Variance() const { return m_variance / (Count() - 1); }
   T StandardDeviation() const { return std::sqrt(Variance()); }
diff --git a/Source/Core/VideoCommon/PerformanceTracker.cpp b/Source/Core/VideoCommon/PerformanceTracker.cpp
index 2ff6a7863b..a70d7607a0 100644
--- a/Source/Core/VideoCommon/PerformanceTracker.cpp
+++ b/Source/Core/VideoCommon/PerformanceTracker.cpp
@@ -31,13 +31,13 @@ void PerformanceTracker::Reset()
   m_raw_dts.Clear();
   m_dt_queue.clear();
 
-  m_dt_total = DT::zero();
-  m_last_raw_dt = DT::zero();
+  m_running_variance.Clear();
+  m_last_raw_dt.store(DT::zero(), std::memory_order_relaxed);
   m_last_time = Clock::now();
-  m_hz_avg = 0.0;
-  m_dt_avg = DT::zero();
-  m_dt_std = DT::zero();
-  m_is_last_time_sane = false;
+  m_hz_avg.store(0.0, std::memory_order_relaxed);
+  m_dt_avg.store(DT::zero(), std::memory_order_relaxed);
+  m_dt_std.store(DT::zero(), std::memory_order_relaxed);
+  m_is_last_time_sane.store(false, std::memory_order_relaxed);
 }
 
 void PerformanceTracker::Count()
@@ -47,58 +47,54 @@ void PerformanceTracker::Count()
   const DT diff{current_time - m_last_time};
   m_last_time = current_time;
 
-  if (!m_is_last_time_sane)
+  if (!m_is_last_time_sane.load(std::memory_order_relaxed))
   {
-    m_is_last_time_sane = true;
+    m_is_last_time_sane.store(true, std::memory_order_relaxed);
     return;
   }
 
-  m_last_raw_dt = diff;
+  m_last_raw_dt.store(diff, std::memory_order_relaxed);
   m_raw_dts.Push(diff);
 }
 
 void PerformanceTracker::UpdateStats()
 {
-  DT diff{};
-  while (m_raw_dts.Pop(diff))
-    HandleRawDt(diff);
-
-  // 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()));
-}
+  if (m_raw_dts.Empty())
+    return;
 
-void PerformanceTracker::HandleRawDt(DT diff)
-{
-  if (m_dt_queue.size() == MAX_DT_QUEUE_SIZE)
-    PopBack();
+  const DT window{GetSampleWindow()};
+  auto hz_avg = m_hz_avg.load(std::memory_order_relaxed);
 
-  PushFront(diff);
+  DT diff{};
+  while (m_raw_dts.Pop(diff))
+  {
+    if (m_dt_queue.size() == MAX_DT_QUEUE_SIZE)
+      PopBack();
 
-  const DT window{GetSampleWindow()};
+    PushFront(diff);
 
-  while (m_dt_total - m_dt_queue.back() >= window)
-    PopBack();
+    while (DT(static_cast<long>(m_running_variance.Total())) - m_dt_queue.back() >= window)
+      PopBack();
 
-  // Simple Moving Average Throughout the Window
-  const DT dt_avg = m_dt_total / m_dt_queue.size();
-  const double hz = DT_s(1.0) / dt_avg;
-  m_dt_avg = dt_avg;
+    // Simple Moving Average Throughout the Window
+    const double hz = DT_s(1.0) / DT(static_cast<long>(m_running_variance.Mean()));
 
-  // 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));
+    // Exponential Moving Average
+    const DT_s rc = SAMPLE_RC_RATIO * DT(static_cast<long>(m_running_variance.Total()));
+    const double a = 1.0 - std::exp(-(DT_s(diff) / rc));
 
-  // Sometimes euler averages can break when the average is inf/nan
-  const auto hz_avg = m_hz_avg.load();
-  if (std::isfinite(hz_avg))
-    m_hz_avg = std::lerp(hz_avg, hz, a);
-  else
-    m_hz_avg = hz;
+    // Sometimes euler averages can break when the average is inf/nan
+    if (std::isfinite(hz_avg))
+      hz_avg = std::lerp(hz_avg, hz, a);
+    else
+      hz_avg = hz;
 
-  LogRenderTimeToFile(diff);
+    LogRenderTimeToFile(diff);
+  }
+  m_dt_avg.store(DT(static_cast<long>(m_running_variance.Mean())), std::memory_order_relaxed);
+  m_hz_avg.store(hz_avg, std::memory_order_relaxed);
+  m_dt_std.store(DT(static_cast<long>(m_running_variance.PopulationStandardDeviation())),
+                 std::memory_order_relaxed);
 }
 
 DT PerformanceTracker::GetSampleWindow() const
@@ -109,27 +105,27 @@ DT PerformanceTracker::GetSampleWindow() const
 
 double PerformanceTracker::GetHzAvg() const
 {
-  return m_hz_avg;
+  return m_hz_avg.load(std::memory_order_relaxed);
 }
 
 DT PerformanceTracker::GetDtAvg() const
 {
-  return m_dt_avg;
+  return m_dt_avg.load(std::memory_order_relaxed);
 }
 
 DT PerformanceTracker::GetDtStd() const
 {
-  return m_dt_std;
+  return m_dt_std.load(std::memory_order_relaxed);
 }
 
 DT PerformanceTracker::GetLastRawDt() const
 {
-  return m_last_raw_dt;
+  return m_last_raw_dt.load(std::memory_order_relaxed);
 }
 
 void PerformanceTracker::InvalidateLastTime()
 {
-  m_is_last_time_sane = false;
+  m_is_last_time_sane.store(false, std::memory_order_relaxed);
 }
 
 void PerformanceTracker::ImPlotPlotLines(const char* label) const
@@ -175,12 +171,12 @@ void PerformanceTracker::ImPlotPlotLines(const char* label) const
 void PerformanceTracker::PushFront(DT value)
 {
   m_dt_queue.push_front(value);
-  m_dt_total += value;
+  m_running_variance.Push(value.count());
 }
 
 void PerformanceTracker::PopBack()
 {
-  m_dt_total -= m_dt_queue.back();
+  m_running_variance.Pop(m_dt_queue.back().count());
   m_dt_queue.pop_back();
 }
 
diff --git a/Source/Core/VideoCommon/PerformanceTracker.h b/Source/Core/VideoCommon/PerformanceTracker.h
index e210a8497c..508cfdd960 100644
--- a/Source/Core/VideoCommon/PerformanceTracker.h
+++ b/Source/Core/VideoCommon/PerformanceTracker.h
@@ -44,7 +44,6 @@ public:
 private:
   void LogRenderTimeToFile(DT val);
 
-  void HandleRawDt(DT value);
   void PushFront(DT value);
   void PopBack();
 
@@ -64,8 +63,12 @@ private:
   // Amount of time to sample dt's over (defaults to config)
   const std::optional<DT> m_sample_window_duration;
 
-  // Queue + Running Total used to calculate average dt
-  DT m_dt_total = DT::zero();
+  // For calculating the average dt and standard deviation
+  // Using `double` because the alternative DT::rep (ns) risks causing an overflow, since
+  // RunningVariance stores the current std squared.
+  // (Even though it's basically impossible to achieve in practice, it'd need a
+  // more-than-a-minute long stutter followed by only almost-zero frametimes.)
+  MathUtil::RunningVariance<double> m_running_variance{};
   std::deque<DT> m_dt_queue;
 
   // Average rate/time throughout the window