Add frame limiter

#1115

Common/Settings.h

@@ -192,6 +192,7 @@
 	visit(fog_layer2_complexity, 0.055f) \
 	visit(fog_layer2_density_add, 100.0f) \
 	visit(fog_layer2_density_mult, 1.4f) \
+	visit(LimitPerFrameFPS, 0.0f) \
 	visit(water_spec_mult_apt_staircase, 0.035f) \
 	visit(water_spec_mult_strange_area, 0.017f) \
 	visit(water_spec_mult_labyrinth, 0.017f) \
@@ -254,6 +255,7 @@
 	visit(HookDirectSound) \
 	visit(HookWndProc) \
 	visit(LetterSpacing) \
+	visit(LimitPerFrameFPS) \
 	visit(LoadModulesFromMemory) \
 	visit(LockResolution) \
 	visit(NormalFontHeight) \

Common/Utils.cpp

@@ -16,6 +16,7 @@
 
 #define WIN32_LEAN_AND_MEAN
 #include <Windows.h>
+#include <intrin.h>
 #ifdef ISLAUNCHER
 #undef NTDDI_VERSION
 #define NTDDI_VERSION NTDDI_WINXPSP3
@@ -420,6 +421,32 @@ bool GetCoreCount(DWORD& pBits, DWORD& sBits)
 	return false;
 }
 
+void BusyWaitYield(DWORD RemainingMS)
+{
+	static bool supports_pause = []() {
+		int cpu_info[4] = { 0 };
+		__cpuid(cpu_info, 1); // Query CPU features
+		return (cpu_info[3] & (1 << 26)) != 0; // Check for SSE2 support
+		}();
+
+	// If remaining time is very small (e.g., 1 ms or less), use busy-wait with no operations
+	if (RemainingMS < 3 && supports_pause)
+	{
+		// Use _mm_pause or __asm { nop } to prevent unnecessary CPU cycles
+#ifdef YieldProcessor
+		YieldProcessor();
+#else
+		_mm_pause();
+#endif
+	}
+	else
+	{
+		// For larger remaining times, we can relax by yielding to the OS
+		// Sleep(0) yields without consuming CPU excessively
+		Sleep(0); // Let the OS schedule other tasks if there's significant time left
+	}
+}
+
 // Log process affinity
 void LogAffinity()
 {

Common/Utils.h

@@ -17,6 +17,7 @@ bool UpdateMemoryAddress(void *dataAddr, const void *dataBytes, size_t dataSize)
 bool WriteCalltoMemory(BYTE *dataAddr, const void *JMPAddr, DWORD count = 5);
 bool WriteJMPtoMemory(BYTE *dataAddr, const void *JMPAddr, DWORD count = 5);
 DWORD ReplaceMemoryBytes(void *dataSrc, void *dataDest, size_t size, DWORD start, DWORD distance, DWORD count = 0);
+void BusyWaitYield(DWORD RemainingMS);
 void LogAffinity();
 DWORD_PTR GetProcessMask();
 void SetSingleCoreAffinity();

Resources/BuildNo.rc

@@ -1 +1 @@
-#define BUILD_NUMBER 2218 
+#define BUILD_NUMBER 2219 

Wrappers/d3d8/IDirect3D8.cpp

@@ -417,6 +417,10 @@ void SetScreenAndWindowSize()
 	{
 		Logging::Log() << __FUNCTION__ << " Setting display mode: " <<
 			(ScreenMode == WINDOWED ? "Windowed" : ScreenMode == WINDOWED_FULLSCREEN ? "Windowed Fullscreen" : "Exclusive Fullscreen");
+		if (ScreenMode != EXCLUSIVE_FULLSCREEN)
+		{
+			ShowCursor(FALSE);
+		}
 	}
 	LastScreenMode = ScreenMode;
 

Wrappers/d3d8/IDirect3DDevice8.cpp

@@ -20,6 +20,7 @@
 #include <chrono>
 #include <deque>
 #include <ctime>
+#include <numeric>
 #include "Common\Utils.h"
 #include "stb_image.h"
 #include "stb_image_dds.h"
@@ -1374,6 +1375,58 @@ bool m_IDirect3DDevice8::FixPauseMenuOnPresent()
 	return PauseMenuFlag;
 }
 
+void m_IDirect3DDevice8::LimitFrameRate()
+{
+	// Count the number of frames
+	Counter.FrameCounter++;
+
+	// Get performance frequency if not already cached
+	static LARGE_INTEGER Frequency = {};
+	if (!Frequency.QuadPart)
+	{
+		QueryPerformanceFrequency(&Frequency);
+	}
+	static LONGLONG TicksPerMS = Frequency.QuadPart / 1000;
+
+	// Calculate the delay time in ticks
+	static long double PerFrameFPS = LimitPerFrameFPS ? LimitPerFrameFPS : (SetSixtyFPS ? 59.94 : 29.97);
+	static LONGLONG PreFrameTicks = static_cast<LONGLONG>(static_cast<long double>(Frequency.QuadPart) / PerFrameFPS);
+
+	// Get next tick time
+	LARGE_INTEGER ClickTime = {};
+	QueryPerformanceCounter(&ClickTime);
+	LONGLONG TargetEndTicks = Counter.LastPresentTime.QuadPart;
+	LONGLONG FramesSinceLastCall = ((ClickTime.QuadPart - Counter.LastPresentTime.QuadPart - 1) / PreFrameTicks) + 1;
+	if (Counter.LastPresentTime.QuadPart == 0 || FramesSinceLastCall > 2)
+	{
+		QueryPerformanceCounter(&Counter.LastPresentTime);
+		TargetEndTicks = Counter.LastPresentTime.QuadPart;
+	}
+	else
+	{
+		TargetEndTicks += FramesSinceLastCall * PreFrameTicks;
+	}
+
+	// Wait for time to expire
+	bool DoLoop;
+	do {
+		QueryPerformanceCounter(&ClickTime);
+		LONGLONG RemainingTicks = TargetEndTicks - ClickTime.QuadPart;
+
+		// Check if we still need to wait
+		DoLoop = RemainingTicks > 0;
+
+		if (DoLoop)
+		{
+			// Busy wait until we reach the target time
+			BusyWaitYield(static_cast<DWORD>(RemainingTicks / TicksPerMS));
+		}
+	} while (DoLoop);
+
+	// Update the last present time
+	Counter.LastPresentTime.QuadPart = TargetEndTicks;
+}
+
 HRESULT m_IDirect3DDevice8::Present(CONST RECT* pSourceRect, CONST RECT* pDestRect, HWND hDestWindowOverride, CONST RGNDATA* pDirtyRegion)
 {
 	Logging::LogDebug() << __FUNCTION__;
@@ -1438,6 +1491,11 @@ HRESULT m_IDirect3DDevice8::Present(CONST RECT* pSourceRect, CONST RECT* pDestRe
 		if (ClearScreen)
 		{
 			ClearScreen = false;
+			if (ScreenMode != EXCLUSIVE_FULLSCREEN)
+			{
+				LimitFrameRate();
+			}
+
 			HRESULT hr = ProxyInterface->Present(pSourceRect, pDestRect, hDestWindowOverride, pDirtyRegion);
 
 			if (SUCCEEDED(hr))
@@ -1476,6 +1534,11 @@ HRESULT m_IDirect3DDevice8::Present(CONST RECT* pSourceRect, CONST RECT* pDestRe
 	// Present screen
 	if (!PauseMenuFlag)
 	{
+		if (ScreenMode != EXCLUSIVE_FULLSCREEN)
+		{
+			LimitFrameRate();
+		}
+
 		hr = ProxyInterface->Present(pSourceRect, pDestRect, hDestWindowOverride, pDirtyRegion);
 
 		if (SUCCEEDED(hr))

Wrappers/d3d8/IDirect3DDevice8.h

@@ -208,6 +208,12 @@ class m_IDirect3DDevice8 : public IDirect3DDevice8
 		UINT stream0Stride;
 	};
 
+	// Limit frame rate
+	struct {
+		DWORD FrameCounter = 0;
+		LARGE_INTEGER LastPresentTime = {};
+	} Counter;
+
 	// Helper functions
 	void EnableAntiAliasing();
 	void DisableAntiAliasing();
@@ -224,6 +230,7 @@ class m_IDirect3DDevice8 : public IDirect3DDevice8
 	void CaptureScreenShot();
 	HRESULT CreateDCSurface(EMUSURFACE& surface, LONG Width, LONG Height);
 	void ReleaseDCSurface(EMUSURFACE& surface);
+	void LimitFrameRate();
 
 public:
 	m_IDirect3DDevice8(LPDIRECT3DDEVICE8 pDevice, m_IDirect3D8* pD3D) : ProxyInterface(pDevice), m_pD3D(pD3D)