From 2229f9b70ef7575952a8e698e6de0014c5b8f530 Mon Sep 17 00:00:00 2001
From: Muzychenko Andrey <33288308+k4zmu2a@users.noreply.github.com>
Date: Thu, 18 Nov 2021 17:58:53 +0300
Subject: [PATCH] Added hybrid sleep/spin wait mode.

---
 SpaceCadetPinball/options.cpp |  2 ++
 SpaceCadetPinball/options.h   |  1 +
 SpaceCadetPinball/winmain.cpp | 45 ++++++++++++++++++++++++++++++++---
 SpaceCadetPinball/winmain.h   | 28 ++++++++++++++++++++++
 4 files changed, 73 insertions(+), 3 deletions(-)

diff --git a/SpaceCadetPinball/options.cpp b/SpaceCadetPinball/options.cpp
index d78850a..733b432 100644
--- a/SpaceCadetPinball/options.cpp
+++ b/SpaceCadetPinball/options.cpp
@@ -99,6 +99,7 @@ void options::init()
 	Options.UncappedUpdatesPerSecond = get_int("Uncapped Updates Per Second", false);
 	Options.SoundChannels = get_int("Sound Channels", DefSoundChannels);
 	Options.SoundChannels = std::min(MaxSoundChannels, std::max(MinSoundChannels, Options.SoundChannels));
+	Options.HybridSleep = get_int("HybridSleep", false);
 
 	winmain::UpdateFrameRate();
 
@@ -130,6 +131,7 @@ void options::uninit()
 	set_int("ShowMenu", Options.ShowMenu);
 	set_int("Uncapped Updates Per Second", Options.UncappedUpdatesPerSecond);
 	set_int("Sound Channels", Options.SoundChannels);
+	set_int("HybridSleep", Options.HybridSleep);
 }
 
 
diff --git a/SpaceCadetPinball/options.h b/SpaceCadetPinball/options.h
index 5578cb6..2ea0505 100644
--- a/SpaceCadetPinball/options.h
+++ b/SpaceCadetPinball/options.h
@@ -73,6 +73,7 @@ struct optionsStruct
 	bool ShowMenu;
 	bool UncappedUpdatesPerSecond;
 	int SoundChannels;
+	bool HybridSleep;
 };
 
 struct ControlRef
diff --git a/SpaceCadetPinball/winmain.cpp b/SpaceCadetPinball/winmain.cpp
index c98f4ac..e302c59 100644
--- a/SpaceCadetPinball/winmain.cpp
+++ b/SpaceCadetPinball/winmain.cpp
@@ -42,6 +42,8 @@ std::string winmain::FpsDetails;
 double winmain::UpdateToFrameRatio;
 winmain::DurationMs winmain::TargetFrameTime;
 optionsStruct& winmain::Options = options::Options;
+winmain::DurationMs winmain::SpinThreshold = DurationMs(0.005);
+WelfordState winmain::SleepState{};
 
 int winmain::WinMain(LPCSTR lpCmdLine)
 {
@@ -115,7 +117,8 @@ int winmain::WinMain(LPCSTR lpCmdLine)
 
 	// If HW fails, fallback to SW SDL renderer.
 	SDL_Renderer* renderer = nullptr;
-	for (int i = 0; i < 2 && !renderer; i++)
+	auto swOffset = strstr(lpCmdLine, "-sw") != nullptr ? 1 : 0;
+	for (int i = swOffset; i < 2 && !renderer; i++)
 	{
 		Renderer = renderer = SDL_CreateRenderer
 		(
@@ -288,7 +291,10 @@ int winmain::WinMain(LPCSTR lpCmdLine)
 			TimePoint frameEnd;
 			if (targetTimeDelta > DurationMs::zero() && !Options.UncappedUpdatesPerSecond)
 			{
-				std::this_thread::sleep_for(targetTimeDelta);
+				if (Options.HybridSleep)
+					HybridSleep(targetTimeDelta);
+				else
+					std::this_thread::sleep_for(targetTimeDelta);
 				frameEnd = Clock::now();
 			}
 			else
@@ -519,6 +525,12 @@ void winmain::RenderUi()
 				{
 					Options.UncappedUpdatesPerSecond ^= true;
 				}
+				if (ImGui::MenuItem("Precise Sleep", nullptr, Options.HybridSleep))
+				{
+					Options.HybridSleep ^= true;
+					SleepState = WelfordState{};
+					SpinThreshold = DurationMs::zero();
+				}
 
 				if (changed)
 				{
@@ -947,7 +959,9 @@ void winmain::RenderFrameTimeDialog()
 		auto target = static_cast<float>(TargetFrameTime.count());
 		auto scale = 1 / (64 / 2 / target);
 
-		ImGui::Text("Target frame time:%03.04fms, 1px:%03.04fms", target, scale);
+		auto spin = Options.HybridSleep ? static_cast<float>(SpinThreshold.count()) : 0;
+		ImGui::Text("Target frame time:%03.04fms, 1px:%03.04fms, SpinThreshold:%03.04fms",
+			target, scale, spin);
 		gfr_display->BlitToTexture();
 		auto region = ImGui::GetContentRegionAvail();
 		ImGui::Image(gfr_display->Texture, region);
@@ -955,3 +969,28 @@ void winmain::RenderFrameTimeDialog()
 	ImGui::End();
 	ImGui::PopStyleVar();
 }
+
+void winmain::HybridSleep(DurationMs sleepTarget)
+{	
+	static constexpr double StdDevFactor = 0.5;
+
+	// This nice concept is from https://blat-blatnik.github.io/computerBear/making-accurate-sleep-function/
+	// Sacrifices some CPU time for smaller frame time jitter
+	while (sleepTarget > SpinThreshold)
+	{
+		auto start = Clock::now();
+		std::this_thread::sleep_for(DurationMs(1));
+		auto end = Clock::now();
+
+		auto actualDuration = DurationMs(end - start);
+		sleepTarget -= actualDuration;
+
+		// Update expected sleep duration using Welford's online algorithm
+		// With bad timer, this will run away to 100% spin
+		SleepState.Advance(actualDuration.count());
+		SpinThreshold = DurationMs(SleepState.mean + SleepState.GetStdDev() * StdDevFactor);
+	}
+
+	// spin lock
+	for (auto start = Clock::now(); DurationMs(Clock::now() - start) < sleepTarget;);
+}
diff --git a/SpaceCadetPinball/winmain.h b/SpaceCadetPinball/winmain.h
index c6ab002..289d142 100644
--- a/SpaceCadetPinball/winmain.h
+++ b/SpaceCadetPinball/winmain.h
@@ -33,6 +33,31 @@ struct SdlPerformanceClock
 	}
 };
 
+struct WelfordState
+{
+	double mean;
+	double M2;
+	int64_t count;
+
+	WelfordState() : mean(0.005), M2(0), count(1)
+	{
+
+	}
+
+	void Advance(double newValue)
+	{
+		++count;
+		auto delta = newValue - mean;
+		mean += delta / count;
+		M2 += delta * (newValue - mean); //M2n = M2n-1 + (Xn - AvgXn-1) * (Xn - AvgXn)
+	}
+
+	double GetStdDev() const
+	{
+		return std::sqrt(M2 / (count - 1)); // Sn^2 = M2n / (n - 1)
+	}
+};
+
 class winmain
 {
 	using Clock = SdlPerformanceClock; // Or std::chrono::steady_clock.
@@ -75,7 +100,10 @@ private:
 	static double UpdateToFrameRatio;
 	static DurationMs TargetFrameTime;
 	static struct optionsStruct& Options;
+	static DurationMs SpinThreshold;
+	static WelfordState SleepState;
 
 	static void RenderUi();
 	static void RenderFrameTimeDialog();
+	static void HybridSleep(DurationMs seconds);
 };