Linux/MacOS: Greatly improve performance (#370)

std::unordered_set is implemented as a flat hashtable on libstdc++ which makes clearing expensive due to invoking memset on the entire table. To get the best performance across all platforms this replaces the unordered_set with a custom high-performance sparse bitset

src/Cafe/HW/Latte/Core/LatteBufferCache.cpp

@@ -1005,8 +1005,67 @@ void LatteBufferCache_getStats(uint32& heapSize, uint32& allocationSize, uint32&
 }
 
 FSpinlock g_spinlockDCFlushQueue;
-std::unordered_set<uint32>* g_DCFlushQueue = new std::unordered_set<uint32>(); // queued pages
+
-std::unordered_set<uint32>* g_DCFlushQueueAlternate = new std::unordered_set<uint32>();
+class SparseBitset
+{
+	static inline constexpr size_t TABLE_MASK = 0xFF;
+
+public:
+	bool Empty() const
+	{
+		return m_numNonEmptyVectors == 0;
+	}
+
+	void Set(uint32 index)
+	{
+		auto& v = m_bits[index & TABLE_MASK];
+		if (std::find(v.cbegin(), v.cend(), index) != v.end())
+			return;
+		if (v.empty())
+		{
+			m_nonEmptyVectors[m_numNonEmptyVectors] = &v;
+			m_numNonEmptyVectors++;
+		}
+		v.emplace_back(index);
+	}
+
+	template<typename TFunc>
+	void ForAllAndClear(TFunc callbackFunc)
+	{
+		auto vCurrent = m_nonEmptyVectors + 0;
+		auto vEnd = m_nonEmptyVectors + m_numNonEmptyVectors;
+		while (vCurrent < vEnd)
+		{
+			std::vector<uint32>* vec = *vCurrent;
+			vCurrent++;
+			for (const auto& it : *vec)
+				callbackFunc(it);
+			vec->clear();
+		}
+		m_numNonEmptyVectors = 0;
+	}
+
+	void Clear()
+	{
+		auto vCurrent = m_nonEmptyVectors + 0;
+		auto vEnd = m_nonEmptyVectors + m_numNonEmptyVectors;
+		while (vCurrent < vEnd)
+		{
+			std::vector<uint32>* vec = *vCurrent;
+			vCurrent++;
+			vec->clear();
+		}
+		m_numNonEmptyVectors = 0;
+	}
+
+private:
+	std::vector<uint32> m_bits[TABLE_MASK + 1];
+	std::vector<uint32>* m_nonEmptyVectors[TABLE_MASK + 1];
+	size_t m_numNonEmptyVectors{ 0 };
+};
+
+SparseBitset* s_DCFlushQueue = new SparseBitset();
+SparseBitset* s_DCFlushQueueAlternate = new SparseBitset();
 
 void LatteBufferCache_notifyDCFlush(MPTR address, uint32 size)
 {
@@ -1017,20 +1076,18 @@ void LatteBufferCache_notifyDCFlush(MPTR address, uint32 size)
 	uint32 lastPage = (address + size - 1) / CACHE_PAGE_SIZE;
 	g_spinlockDCFlushQueue.acquire();
 	for (uint32 i = firstPage; i <= lastPage; i++)
-		g_DCFlushQueue->emplace(i);
+		s_DCFlushQueue->Set(i);
 	g_spinlockDCFlushQueue.release();
 }
 
 void LatteBufferCache_processDCFlushQueue()
 {
-	if (g_DCFlushQueue->empty()) // accessing this outside of the lock is technically undefined/unsafe behavior but on all known implementations this is fine and we can avoid the spinlock
+	if (s_DCFlushQueue->Empty()) // quick check to avoid locking if there is no work to do
 		return;
 	g_spinlockDCFlushQueue.acquire();
-	std::swap(g_DCFlushQueue, g_DCFlushQueueAlternate);
+	std::swap(s_DCFlushQueue, s_DCFlushQueueAlternate);
 	g_spinlockDCFlushQueue.release();
-	for (auto& itr : *g_DCFlushQueueAlternate)
+	s_DCFlushQueueAlternate->ForAllAndClear([](uint32 index) {LatteBufferCache_invalidatePage(index * CACHE_PAGE_SIZE); });
-		LatteBufferCache_invalidatePage(itr * CACHE_PAGE_SIZE);
-	g_DCFlushQueueAlternate->clear();
 }
 
 void LatteBufferCache_notifyDrawDone()

src/audio/CubebAPI.cpp

@@ -167,25 +167,11 @@ void CubebAPI::SetVolume(sint32 volume)
 
 bool CubebAPI::InitializeStatic()
 {
-#if BOOST_OS_WINDOWS
-	s_com_initialized = (SUCCEEDED(CoInitializeEx(nullptr, COINIT_MULTITHREADED)));
-#endif
-
 	if (cubeb_init(&s_context, "Cemu Cubeb", nullptr))
 	{
 		cemuLog_force("can't create cubeb audio api");
-
-#if BOOST_OS_WINDOWS
-		if (s_com_initialized)
-		{
-			CoUninitialize();
-			s_com_initialized = false;
-		}
-#endif
-
 		return false;
 	}
-
 	return true;
 }
 
@@ -193,10 +179,6 @@ void CubebAPI::Destroy()
 {
 	if (s_context)
 		cubeb_destroy(s_context);
-#if BOOST_OS_WINDOWS
-	if (s_com_initialized)
-		CoUninitialize();
-#endif
 }
 
 std::vector<IAudioAPI::DeviceDescriptionPtr> CubebAPI::GetDevices()

src/audio/CubebAPI.h

@@ -41,7 +41,6 @@ public:
 	static void Destroy();
 
 private:
-	inline static bool s_com_initialized = false;
 	inline static cubeb* s_context = nullptr;
 
 	cubeb_stream* m_stream = nullptr;

src/audio/XAudio27API.cpp

@@ -5,7 +5,6 @@
 static_assert(IAudioAPI::kBlockCount < XAUDIO2_MAX_QUEUED_BUFFERS, "too many xaudio2 buffers");
 
 HMODULE XAudio27API::s_xaudio_dll = nullptr;
-bool XAudio27API::s_com_initialized = false;
 std::unique_ptr<IXAudio2, XAudio27API::XAudioDeleter> XAudio27API::s_xaudio;
 
 XAudio27API::XAudio27API(uint32 device_id, uint32 samplerate, uint32 channels, uint32 samples_per_block, uint32 bits_per_sample)
@@ -115,8 +114,6 @@ bool XAudio27API::InitializeStatic()
 	if (s_xaudio)
 		return true;
 
-	s_com_initialized = (SUCCEEDED(CoInitializeEx(nullptr, COINIT_MULTITHREADED | COINIT_DISABLE_OLE1DDE)));
-
 #ifdef _DEBUG
 	s_xaudio_dll = LoadLibraryExW(L"XAudioD2_7.DLL", nullptr, LOAD_LIBRARY_SEARCH_SYSTEM32);
 	if(!s_xaudio_dll)
@@ -142,9 +139,6 @@ bool XAudio27API::InitializeStatic()
 		if (s_xaudio_dll)
 			FreeLibrary(s_xaudio_dll);
 
-		if (s_com_initialized)
-			CoUninitialize();
-
 		return false;
 	}
 }
@@ -155,9 +149,6 @@ void XAudio27API::Destroy()
 
 	if (s_xaudio_dll)
 		FreeLibrary(s_xaudio_dll);
-
-	if (s_com_initialized)
-		CoUninitialize();
 }
 
 std::vector<XAudio27API::DeviceDescriptionPtr> XAudio27API::GetDevices()

src/audio/XAudio27API.h

@@ -58,7 +58,6 @@ private:
 	};
 
 	static HMODULE s_xaudio_dll;
-	static bool s_com_initialized;
 	static std::unique_ptr<IXAudio2, XAudioDeleter> s_xaudio;
 
 	std::unique_ptr<IXAudio2, XAudioDeleter> m_xaudio;

src/audio/XAudio2API.cpp

@@ -23,7 +23,6 @@ static const GUID DEVINTERFACE_AUDIO_RENDER_GUID = { 0xe6327cad, 0xdcec, 0x4949,
 static_assert(IAudioAPI::kBlockCount < XAUDIO2_MAX_QUEUED_BUFFERS, "too many xaudio2 buffers");
 
 HMODULE XAudio2API::s_xaudio_dll = nullptr;
-bool XAudio2API::s_com_initialized = false;
 std::vector<XAudio2API::DeviceDescriptionPtr> XAudio2API::s_devices;
 
 XAudio2API::XAudio2API(std::wstring device_id, uint32 samplerate, uint32 channels, uint32 samples_per_block, uint32 bits_per_sample)
@@ -143,8 +142,6 @@ bool XAudio2API::InitializeStatic()
 {
 	if (s_xaudio_dll)
 		return true;
-	
-	s_com_initialized = (SUCCEEDED(CoInitializeEx(nullptr, COINIT_MULTITHREADED)));
 
 	// win 10
 	s_xaudio_dll = LoadLibraryEx(XAUDIO2_DLL, nullptr, LOAD_LIBRARY_SEARCH_SYSTEM32);
@@ -166,9 +163,6 @@ bool XAudio2API::InitializeStatic()
 		if (s_xaudio_dll)
 			FreeLibrary(s_xaudio_dll);
 
-		if (s_com_initialized)
-			CoUninitialize();
-
 		return false;
 	}
 }
@@ -177,20 +171,12 @@ void XAudio2API::Destroy()
 {
 	if (s_xaudio_dll)
 		FreeLibrary(s_xaudio_dll);
-
-	if (s_com_initialized)
-		CoUninitialize();
 }
 
 const std::vector<XAudio2API::DeviceDescriptionPtr>& XAudio2API::RefreshDevices()
 {
-	// this function must be called from the same thread as we called CoInitializeEx
 	s_devices.clear();
 
-	HRESULT r = CoInitializeEx(nullptr, COINIT_MULTITHREADED | COINIT_DISABLE_OLE1DDE);
-	if (r != RPC_E_CHANGED_MODE && FAILED(r))
-		return s_devices;
-
 	try
 	{
 		struct IWbemLocator *wbem_locator = nullptr;

src/audio/XAudio2API.h

@@ -59,7 +59,6 @@ private:
 	};
 
 	static HMODULE s_xaudio_dll;
-	static bool s_com_initialized;
 	static std::vector<DeviceDescriptionPtr> s_devices;
 
 	std::unique_ptr<IXAudio2, XAudioDeleter> m_xaudio;

src/main.cpp

@@ -323,10 +323,10 @@ void HandlePostUpdate()
 			fs::remove(filename, ec);
 		}
 #else
-		while( fs::exists(filename) )
+		while (fs::exists(filename))
 		{
 			std::error_code ec;
-			fs::remove(filename, ec);		
+			fs::remove(filename, ec);
 			std::this_thread::sleep_for(std::chrono::milliseconds(1000));
 		}
 #endif
@@ -338,8 +338,10 @@ void ToolShaderCacheMerger();
 #if BOOST_OS_WINDOWS
 
 // entrypoint for release builds
-int wWinMain( _In_ HINSTANCE hInstance, _In_opt_ HINSTANCE hPrevInstance, _In_ LPTSTR lpCmdLine, _In_ int nShowCmd)
+int wWinMain(_In_ HINSTANCE hInstance, _In_opt_ HINSTANCE hPrevInstance, _In_ LPTSTR lpCmdLine, _In_ int nShowCmd)
 {
+	if (FAILED(CoInitializeEx(nullptr, COINIT_MULTITHREADED | COINIT_DISABLE_OLE1DDE)))
+		cemuLog_log(LogType::Force, "CoInitializeEx() failed");
 	SDL_SetMainReady();
 	if (!LaunchSettings::HandleCommandline(lpCmdLine))
 		return 0;
@@ -350,6 +352,8 @@ int wWinMain( _In_ HINSTANCE hInstance, _In_opt_ HINSTANCE hPrevInstance, _In_ L
 // entrypoint for debug builds with console
 int main(int argc, char* argv[])
 {
+	if (FAILED(CoInitializeEx(nullptr, COINIT_MULTITHREADED | COINIT_DISABLE_OLE1DDE)))
+		cemuLog_log(LogType::Force, "CoInitializeEx() failed");
 	SDL_SetMainReady();
 	if (!LaunchSettings::HandleCommandline(argc, argv))
 		return 0;