rsx: Trivial non-blocking display synchronization

rpcs3/Emu/Cell/lv2/sys_rsx.cpp

@@ -498,9 +498,8 @@ error_code sys_rsx_context_attribute(u32 context_id, u32 package_id, u64 a3, u64
 		driverInfo.head[a3].lastQueuedBufferId = static_cast<u32>(a4);
 		driverInfo.head[a3].flipFlags |= 0x40000000 | (1 << a4);
 
-		render->send_event(0, SYS_RSX_EVENT_QUEUE_BASE << a3, 0);
-
 		render->on_frame_end(static_cast<u32>(a4));
+		render->send_event(0, SYS_RSX_EVENT_QUEUE_BASE << a3, 0);
 	}
 	break;
 

rpcs3/Emu/RSX/RSXThread.cpp

@@ -3099,6 +3099,18 @@ namespace rsx
 			Emu.Pause();
 		}
 
+		if (false && wait_for_flip_sema) // Breaks framepacing
+		{
+			const auto& value = vm::_ref<RsxSemaphore>(device_addr + 0x30).val;
+			while (value != flip_sema_wait_val && !test_stopped())
+			{
+				_mm_pause();
+			}
+
+			wait_for_flip_sema = false;
+			m_eng_interrupt_mask |= rsx::display_interrupt;
+		}
+
 		if (zcull_ctrl->has_pending())
 		{
 			// NOTE: This is a workaround for buggy games.
@@ -3195,7 +3207,7 @@ namespace rsx
 		case frame_limit_type::_50: limit = 50.; break;
 		case frame_limit_type::_60: limit = 60.; break;
 		case frame_limit_type::_30: limit = 30.; break;
-		case frame_limit_type::_auto: limit = 0.; break; // Handled in RSX semaphore_acquire
+		case frame_limit_type::_auto: limit = 0.; break;
 		default:
 			break;
 		}
@@ -3231,6 +3243,18 @@ namespace rsx
 				}
 			}
 		}
+		else if (wait_for_flip_sema)
+		{
+			const auto& value = vm::_ref<RsxSemaphore>(device_addr + 0x30).val;
+			if (value != flip_sema_wait_val)
+			{
+				// Not yet signaled, handle it later
+				async_flip_requested |= flip_request::emu_requested;
+				return;
+			}
+
+			wait_for_flip_sema = false;
+		}
 
 		int_flip_index++;
 

rpcs3/Emu/RSX/RSXThread.h

@@ -650,6 +650,9 @@ namespace rsx
 		atomic_t<u64> vblank_count{0};
 		bool capture_current_frame = false;
 
+		bool wait_for_flip_sema = false;
+		u32 flip_sema_wait_val = 0;
+
 	public:
 		atomic_t<bool> sync_point_request = false;
 		bool in_begin_end = false;

rpcs3/Emu/RSX/rsx_methods.cpp

@@ -104,8 +104,13 @@ namespace rsx
 				rsx->flush_fifo();
 			}
 
-			if (addr == rsx->device_addr + 0x30 && g_cfg.video.frame_limit == frame_limit_type::none)
+			if (addr == rsx->device_addr + 0x30)
 			{
+				if (g_cfg.video.frame_limit == frame_limit_type::_auto)
+				{
+					rsx->flip_sema_wait_val = arg;
+					rsx->wait_for_flip_sema = (sema != arg);
+				}
 				return;
 			}
 
@@ -1733,7 +1738,7 @@ namespace rsx
 		template<u32 index>
 		struct driver_flip
 		{
-			static void impl(thread* /*rsx*/, u32 /*reg*/, u32 arg)
+			static void impl(thread* rsx, u32 /*reg*/, u32 arg)
 			{
 				sys_rsx_context_attribute(0x55555555, 0x102, index, arg, 0, 0);
 			}
@@ -1742,7 +1747,7 @@ namespace rsx
 		template<u32 index>
 		struct queue_flip
 		{
-			static void impl(thread* /*rsx*/, u32 /*reg*/, u32 arg)
+			static void impl(thread* rsx, u32 /*reg*/, u32 arg)
 			{
 				sys_rsx_context_attribute(0x55555555, 0x103, index, arg, 0, 0);
 			}