GS: Use semaphores with a fast userspace path

common/Threading.h

@@ -230,4 +230,35 @@ namespace Threading
 		/// Should be called by the worker thread if it restarts after dying
 		void Reset();
 	};
+
+	/// A semaphore that definitely has a fast userspace path
+	class UserspaceSemaphore
+	{
+		KernelSemaphore m_sema;
+		std::atomic<uint32_t> m_counter{0};
+
+	public:
+		UserspaceSemaphore() = default;
+		~UserspaceSemaphore() = default;
+
+		void Post()
+		{
+			if (m_counter.fetch_add(1, std::memory_order_release) < 0)
+				m_sema.Post();
+		}
+
+		void Wait()
+		{
+			if (m_counter.fetch_sub(1, std::memory_order_acquire) <= 0)
+				m_sema.Wait();
+		}
+
+		bool TryWait()
+		{
+			int32_t counter = m_counter.load(std::memory_order_relaxed);
+			while (counter > 0 && !m_counter.compare_exchange_weak(counter, counter - 1, std::memory_order_acquire, std::memory_order_relaxed))
+				;
+			return counter > 0;
+		}
+	};
 } // namespace Threading

pcsx2/GS.h

@@ -342,8 +342,8 @@ public:
 	std::mutex m_mtx_RingBufferBusy2; // Gets released on semaXGkick waiting...
 	std::mutex m_mtx_WaitGS;
 	Threading::WorkSema m_sem_event;
-	Threading::KernelSemaphore m_sem_OnRingReset;
-	Threading::KernelSemaphore m_sem_Vsync;
+	Threading::UserspaceSemaphore m_sem_OnRingReset;
+	Threading::UserspaceSemaphore m_sem_Vsync;
 
 	// used to keep multiple threads from sending packets to the ringbuffer concurrently.
 	// (currently not used or implemented -- is a planned feature for a future threaded VU1)
@@ -369,7 +369,7 @@ public:
 	std::atomic_bool m_open_flag{false};
 	std::atomic_bool m_shutdown_flag{false};
 	std::atomic_bool m_run_idle_flag{false};
-	Threading::KernelSemaphore m_open_or_close_done;
+	Threading::UserspaceSemaphore m_open_or_close_done;
 
 public:
 	SysMtgsThread();

pcsx2/MTVU.h

@@ -45,7 +45,7 @@ class VU_Thread final {
 public:
 	alignas(16)  vifStruct        vif;
 	alignas(16)  VIFregisters     vifRegs;
-	Threading::KernelSemaphore semaXGkick;
+	Threading::UserspaceSemaphore semaXGkick;
 	std::atomic<unsigned int> vuCycles[4]; // Used for VU cycle stealing hack
 	u32 vuCycleIdx;  // Used for VU cycle stealing hack
 	u32 vuFBRST;