rsx: Improve balancing of the offloader thread

- Use two counters to avoid atomic operations
- Yield instead of sleeping because some games are very sensitive to timing

rpcs3/Emu/Cell/SPUThread.cpp

@@ -1797,7 +1797,7 @@ bool spu_thread::process_mfc_cmd()
 					break;
 				}
 
-				thread_ctrl::wait_for(100);
+				thread_ctrl::wait_for(500);
 			}
 
 			if (test_stopped())

rpcs3/Emu/RSX/RSXOffload.cpp

@@ -4,12 +4,20 @@
 #include "Emu/System.h"
 #include "RSXOffload.h"
 
+#include <thread>
+
 namespace rsx
 {
 	// initialization
 	void dma_manager::init()
 	{
 		m_worker_state = thread_state::created;
+		m_enqueued_count.store(0);
+		m_processed_count = 0;
+
+		// Empty work queue in case of stale contents
+		m_work_queue.pop_all();
+
 		thread_ctrl::spawn("RSX offloader", [this]()
 		{
 			if (!g_cfg.video.multithreaded_rsx)
@@ -25,7 +33,7 @@ namespace rsx
 
 			while (m_worker_state != thread_state::finished)
 			{
-				if (m_jobs_count)
+				if (m_enqueued_count.load() != m_processed_count)
 				{
 					for (auto slice = m_work_queue.pop_all(); slice; slice.pop_front())
 					{
@@ -49,16 +57,17 @@ namespace rsx
 							fmt::throw_exception("Unreachable" HERE);
 						}
 
-						m_jobs_count--;
+						++m_processed_count;
 					}
 				}
 				else
 				{
-					thread_ctrl::wait_for(500);
+					// Yield
+					std::this_thread::yield();
 				}
 			}
 
-			m_jobs_count.store(0);
+			m_processed_count = m_enqueued_count.load();
 		});
 	}
 
@@ -71,7 +80,7 @@ namespace rsx
 		}
 		else
 		{
-			++m_jobs_count;
+			++m_enqueued_count;
 			m_work_queue.push(dst, src, length);
 		}
 	}
@@ -84,7 +93,7 @@ namespace rsx
 		}
 		else
 		{
-			++m_jobs_count;
+			++m_enqueued_count;
 			m_work_queue.push(dst, src, length);
 		}
 	}
@@ -99,7 +108,7 @@ namespace rsx
 		}
 		else
 		{
-			++m_jobs_count;
+			++m_enqueued_count;
 			m_work_queue.push(dst, primitive, count);
 		}
 	}
@@ -107,11 +116,14 @@ namespace rsx
 	// Synchronization
 	void dma_manager::sync()
 	{
-		if (g_cfg.video.multithreaded_rsx)
+		if (LIKELY(m_enqueued_count.load() == m_processed_count))
 		{
-			while (m_jobs_count)
-				_mm_lfence();
+			// Nothing to do
+			return;
 		}
+
+		while (m_enqueued_count.load() != m_processed_count)
+			_mm_lfence();
 	}
 
 	void dma_manager::join()

rpcs3/Emu/RSX/RSXOffload.h

@@ -42,8 +42,9 @@ namespace rsx
 		};
 
 		lf_queue<transport_packet> m_work_queue;
-		atomic_t<int> m_jobs_count;
-		thread_state m_worker_state;
+		atomic_t<u64> m_enqueued_count{ 0 };
+		volatile u64 m_processed_count = 0;
+		thread_state m_worker_state = thread_state::detached;
 
 		// TODO: Improved benchmarks here; value determined by profiling on a Ryzen CPU, rounded to the nearest 512 bytes
 		const u32 max_immediate_transfer_size = 3584;

rpcs3/Emu/RSX/RSXThread.cpp

@@ -2241,6 +2241,9 @@ namespace rsx
 		// Fragment constants may have been updated
 		m_graphics_state |= rsx::pipeline_state::fragment_constants_dirty;
 
+		// DMA sync; if you need this, don't use MTRSX
+		// g_dma_manager.sync();
+
 		//TODO: On sync every sub-unit should finish any pending tasks
 		//Might cause zcull lockup due to zombie 'unclaimed reports' which are not forcefully removed currently
 		//verify (HERE), async_tasks_pending.load() == 0;