Vulkan: Don't flush commands when creating most sync (#4087)

* Vulkan: Don't flush commands when creating most sync

When the WaitForIdle method is called, we create sync as some internal GPU method may read back written buffer data. Some games randomly intersperse compute dispatch into their render passes, which result in this happening an unbounded number of times depending on how many times they run compute.

Creating sync in Vulkan is expensive, as we need to flush the current command buffer so that it can be waited on. We have a limited number of active command buffers due to how we track resource usage, so submitting too many command buffers will force us to wait for them to return to the pool.

This PR allows less "important" sync (things which are less likely to be waited on) to wait on a command buffer's result without submitting it, instead relying on AutoFlush or another, more important sync to flush it later on.

Because of the possibility of us waiting for a command buffer that hasn't submitted yet, any thread needs to be able to force the active command buffer to submit. The ability to do this has been added to the backend multithreading via an "Interrupt", though it is not supported without multithreading.

OpenGL drivers should already be doing something similar so they don't blow up when creating lots of sync, which is why this hasn't been a problem for these games over there.

Improves Vulkan performance on Xenoblade DE, Pokemon Scarlet/Violet, and Zelda BOTW (still another large issue here)

* Add strict argument

This is technically a separate concern from whether the sync is a host syncpoint.

* Remove _interrupted variable

* Actually wait for the invoke

This is required by AMD GPUs, and also may have caused some issues on other GPUs.

* Remove unused using.

* I don't know why it added these ones.

* Address Feedback

* Fix typo

Ryujinx.Graphics.GAL/IRenderer.cs

@@ -22,7 +22,7 @@ namespace Ryujinx.Graphics.GAL
         ISampler CreateSampler(SamplerCreateInfo info);
         ITexture CreateTexture(TextureCreateInfo info, float scale);
 
-        void CreateSync(ulong id);
+        void CreateSync(ulong id, bool strict);
 
         void DeleteBuffer(BufferHandle buffer);
 
@@ -53,6 +53,8 @@ namespace Ryujinx.Graphics.GAL
 
         void Initialize(GraphicsDebugLevel logLevel);
 
+        void SetInterruptAction(Action<Action> interruptAction);
+
         void Screenshot();
     }
 }

Ryujinx.Graphics.GAL/Multithreading/Commands/Renderer/CreateSyncCommand.cs

@@ -4,15 +4,17 @@
     {
         public CommandType CommandType => CommandType.CreateSync;
         private ulong _id;
+        private bool _strict;
 
-        public void Set(ulong id)
+        public void Set(ulong id, bool strict)
         {
             _id = id;
+            _strict = strict;
         }
 
         public static void Run(ref CreateSyncCommand command, ThreadedRenderer threaded, IRenderer renderer)
         {
-            renderer.CreateSync(command._id);
+            renderer.CreateSync(command._id, command._strict);
 
             threaded.Sync.AssignSync(command._id);
         }

Ryujinx.Graphics.GAL/Multithreading/ThreadedRenderer.cs

@@ -29,6 +29,7 @@ namespace Ryujinx.Graphics.GAL.Multithreading
         private int _elementSize;
         private IRenderer _baseRenderer;
         private Thread _gpuThread;
+        private Thread _backendThread;
         private bool _disposed;
         private bool _running;
 
@@ -38,6 +39,7 @@ namespace Ryujinx.Graphics.GAL.Multithreading
         private CircularSpanPool _spanPool;
 
         private ManualResetEventSlim _invokeRun;
+        private AutoResetEvent _interruptRun;
 
         private bool _lastSampleCounterClear = true;
 
@@ -54,6 +56,8 @@ namespace Ryujinx.Graphics.GAL.Multithreading
         private int _refProducerPtr;
         private int _refConsumerPtr;
 
+        private Action _interruptAction;
+
         public event EventHandler<ScreenCaptureImageInfo> ScreenCaptured;
 
         internal BufferMap Buffers { get; }
@@ -73,6 +77,7 @@ namespace Ryujinx.Graphics.GAL.Multithreading
             _baseRenderer = renderer;
 
             renderer.ScreenCaptured += (sender, info) => ScreenCaptured?.Invoke(this, info);
+            renderer.SetInterruptAction(Interrupt);
 
             Pipeline = new ThreadedPipeline(this, renderer.Pipeline);
             Window = new ThreadedWindow(this, renderer);
@@ -82,6 +87,7 @@ namespace Ryujinx.Graphics.GAL.Multithreading
 
             _galWorkAvailable = new ManualResetEventSlim(false);
             _invokeRun = new ManualResetEventSlim();
+            _interruptRun = new AutoResetEvent(false);
             _spanPool = new CircularSpanPool(this, SpanPoolBytes);
             SpanPool = _spanPool;
 
@@ -95,6 +101,8 @@ namespace Ryujinx.Graphics.GAL.Multithreading
         {
             _running = true;
 
+            _backendThread = Thread.CurrentThread;
+
             _gpuThread = new Thread(() => {
                 gpuLoop();
                 _running = false;
@@ -116,10 +124,18 @@ namespace Ryujinx.Graphics.GAL.Multithreading
                 _galWorkAvailable.Wait();
                 _galWorkAvailable.Reset();
 
+                if (Volatile.Read(ref _interruptAction) != null)
+                {
+                    _interruptAction();
+                    _interruptRun.Set();
+
+                    Interlocked.Exchange(ref _interruptAction, null);
+                }
+
                 // The other thread can only increase the command count.
                 // We can assume that if it is above 0, it will stay there or get higher.
 
-                while (_commandCount > 0)
+                while (_commandCount > 0 && Volatile.Read(ref _interruptAction) == null)
                 {
                     int commandPtr = _consumerPtr;
 
@@ -281,10 +297,10 @@ namespace Ryujinx.Graphics.GAL.Multithreading
             return sampler;
         }
 
-        public void CreateSync(ulong id)
+        public void CreateSync(ulong id, bool strict)
         {
             Sync.CreateSyncHandle(id);
-            New<CreateSyncCommand>().Set(id);
+            New<CreateSyncCommand>().Set(id, strict);
             QueueCommand();
         }
 
@@ -421,6 +437,30 @@ namespace Ryujinx.Graphics.GAL.Multithreading
             _baseRenderer.WaitSync(id);
         }
 
+        private void Interrupt(Action action)
+        {
+            // Interrupt the backend thread from any external thread and invoke the given action.
+
+            if (Thread.CurrentThread == _backendThread)
+            {
+                // If this is called from the backend thread, the action can run immediately.
+                action();
+            }
+            else
+            {
+                while (Interlocked.CompareExchange(ref _interruptAction, action, null) != null) { }
+
+                _galWorkAvailable.Set();
+
+                _interruptRun.WaitOne();
+            }
+        }
+
+        public void SetInterruptAction(Action<Action> interruptAction)
+        {
+            // Threaded renderer ignores given interrupt action, as it provides its own to the child renderer.
+        }
+
         public void Dispose()
         {
             // Dispose must happen from the render thread, after all commands have completed.
@@ -440,6 +480,7 @@ namespace Ryujinx.Graphics.GAL.Multithreading
             _frameComplete.Dispose();
             _galWorkAvailable.Dispose();
             _invokeRun.Dispose();
+            _interruptRun.Dispose();
 
             Sync.Dispose();
         }

Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoClass.cs

@@ -59,7 +59,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
             if (_createSyncPending)
             {
                 _createSyncPending = false;
-                _context.CreateHostSyncIfNeeded(false);
+                _context.CreateHostSyncIfNeeded(false, false);
             }
         }
 
@@ -157,7 +157,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
             }
             else if (operation == SyncpointbOperation.Incr)
             {
-                _context.CreateHostSyncIfNeeded(true);
+                _context.CreateHostSyncIfNeeded(true, true);
                 _context.Synchronization.IncrementSyncpoint(syncpointId);
             }
 
@@ -184,7 +184,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
         {
             _context.Renderer.Pipeline.CommandBufferBarrier();
 
-            _context.CreateHostSyncIfNeeded(false);
+            _context.CreateHostSyncIfNeeded(false, true);
         }
 
         /// <summary>

Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClass.cs

@@ -250,7 +250,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
             uint syncpointId = (uint)argument & 0xFFFF;
 
             _context.AdvanceSequence();
-            _context.CreateHostSyncIfNeeded(true);
+            _context.CreateHostSyncIfNeeded(true, true);
             _context.Renderer.UpdateCounters(); // Poll the query counters, the game may want an updated result.
             _context.Synchronization.IncrementSyncpoint(syncpointId);
         }

Ryujinx.Graphics.Gpu/GpuContext.cs

@@ -316,7 +316,8 @@ namespace Ryujinx.Graphics.Gpu
         /// If no actions are present, a host sync object is not created.
         /// </summary>
         /// <param name="syncpoint">True if host sync is being created by a syncpoint</param>
-        public void CreateHostSyncIfNeeded(bool syncpoint)
+        /// <param name="strict">True if the sync should signal as soon as possible</param>
+        public void CreateHostSyncIfNeeded(bool syncpoint, bool strict)
         {
             if (BufferMigrations.Count > 0)
             {
@@ -337,7 +338,7 @@ namespace Ryujinx.Graphics.Gpu
 
             if (_pendingSync || (syncpoint && SyncpointActions.Count > 0))
             {
-                Renderer.CreateSync(SyncNumber);
+                Renderer.CreateSync(SyncNumber, strict);
 
                 SyncNumber++;
 

Ryujinx.Graphics.OpenGL/OpenGLRenderer.cs

@@ -232,7 +232,7 @@ namespace Ryujinx.Graphics.OpenGL
             return new Program(programBinary, hasFragmentShader, info.FragmentOutputMap);
         }
 
-        public void CreateSync(ulong id)
+        public void CreateSync(ulong id, bool strict)
         {
             _sync.Create(id);
         }
@@ -247,6 +247,11 @@ namespace Ryujinx.Graphics.OpenGL
             return _sync.GetCurrent();
         }
 
+        public void SetInterruptAction(Action<Action> interruptAction)
+        {
+            // Currently no need for an interrupt action.
+        }
+
         public void Screenshot()
         {
             _window.ScreenCaptureRequested = true;

Ryujinx.Graphics.Vulkan/CommandBufferPool.cs

@@ -116,6 +116,22 @@ namespace Ryujinx.Graphics.Vulkan
             }
         }
 
+        public void AddInUseWaitable(MultiFenceHolder waitable)
+        {
+            lock (_commandBuffers)
+            {
+                for (int i = 0; i < _totalCommandBuffers; i++)
+                {
+                    ref var entry = ref _commandBuffers[i];
+
+                    if (entry.InUse)
+                    {
+                        AddWaitable(i, waitable);
+                    }
+                }
+            }
+        }
+
         public void AddDependency(int cbIndex, CommandBufferScoped dependencyCbs)
         {
             Debug.Assert(_commandBuffers[cbIndex].InUse);

Ryujinx.Graphics.Vulkan/PipelineFull.cs

@@ -227,6 +227,7 @@ namespace Ryujinx.Graphics.Vulkan
             }
 
             CommandBuffer = (Cbs = Gd.CommandBufferPool.ReturnAndRent(Cbs)).CommandBuffer;
+            Gd.RegisterFlush();
 
             // Restore per-command buffer state.
 

Ryujinx.Graphics.Vulkan/SyncManager.cs

@@ -11,7 +11,13 @@ namespace Ryujinx.Graphics.Vulkan
         {
             public ulong ID;
             public MultiFenceHolder Waitable;
+            public ulong FlushId;
             public bool Signalled;
+
+            public bool NeedsFlush(ulong currentFlushId)
+            {
+                return (long)(FlushId - currentFlushId) >= 0;
+            }
         }
 
         private ulong _firstHandle = 0;
@@ -19,6 +25,7 @@ namespace Ryujinx.Graphics.Vulkan
         private readonly VulkanRenderer _gd;
         private readonly Device _device;
         private List<SyncHandle> _handles;
+        private ulong FlushId;
 
         public SyncManager(VulkanRenderer gd, Device device)
         {
@@ -27,17 +34,33 @@ namespace Ryujinx.Graphics.Vulkan
             _handles = new List<SyncHandle>();
         }
 
-        public void Create(ulong id)
+        public void RegisterFlush()
         {
-            MultiFenceHolder waitable = new MultiFenceHolder();
+            FlushId++;
+        }
 
+        public void Create(ulong id, bool strict)
+        {
+            ulong flushId = FlushId;
+            MultiFenceHolder waitable = new MultiFenceHolder();
+            if (strict || _gd.InterruptAction == null)
+            {
                 _gd.FlushAllCommands();
                 _gd.CommandBufferPool.AddWaitable(waitable);
+            }
+            else
+            {
+                // Don't flush commands, instead wait for the current command buffer to finish.
+                // If this sync is waited on before the command buffer is submitted, interrupt the gpu thread and flush it manually.
+
+                _gd.CommandBufferPool.AddInUseWaitable(waitable);
+            }
 
             SyncHandle handle = new SyncHandle
             {
                 ID = id,
-                Waitable = waitable
+                Waitable = waitable,
+                FlushId = flushId
             };
 
             lock (_handles)
@@ -107,6 +130,17 @@ namespace Ryujinx.Graphics.Vulkan
                         return;
                     }
 
+                    if (result.NeedsFlush(FlushId))
+                    {
+                        _gd.InterruptAction(() =>
+                        {
+                            if (result.NeedsFlush(FlushId))
+                            {
+                                _gd.FlushAllCommands();
+                            }
+                        });
+                    }
+
                     bool signaled = result.Signalled || result.Waitable.WaitForFences(_gd.Api, _device, 1000000000);
                     if (!signaled)
                     {
@@ -132,7 +166,7 @@ namespace Ryujinx.Graphics.Vulkan
                     first = _handles.FirstOrDefault();
                 }
 
-                if (first == null) break;
+                if (first == null || first.NeedsFlush(FlushId)) break;
 
                 bool signaled = first.Waitable.WaitForFences(_gd.Api, _device, 0);
                 if (signaled)

Ryujinx.Graphics.Vulkan/VulkanRenderer.cs

@@ -48,6 +48,7 @@ namespace Ryujinx.Graphics.Vulkan
         internal DescriptorSetManager DescriptorSetManager { get; private set; }
         internal PipelineLayoutCache PipelineLayoutCache { get; private set; }
         internal BackgroundResources BackgroundResources { get; private set; }
+        internal Action<Action> InterruptAction { get; private set; }
 
         internal BufferManager BufferManager { get; private set; }
 
@@ -354,6 +355,11 @@ namespace Ryujinx.Graphics.Vulkan
             _pipeline?.FlushCommandsImpl();
         }
 
+        internal void RegisterFlush()
+        {
+            _syncManager.RegisterFlush();
+        }
+
         public ReadOnlySpan<byte> GetBufferData(BufferHandle buffer, int offset, int size)
         {
             return BufferManager.GetData(buffer, offset, size);
@@ -593,9 +599,9 @@ namespace Ryujinx.Graphics.Vulkan
             action();
         }
 
-        public void CreateSync(ulong id)
+        public void CreateSync(ulong id, bool strict)
         {
-            _syncManager.Create(id);
+            _syncManager.Create(id, strict);
         }
 
         public IProgram LoadProgramBinary(byte[] programBinary, bool isFragment, ShaderInfo info)
@@ -613,6 +619,11 @@ namespace Ryujinx.Graphics.Vulkan
             return _syncManager.GetCurrent();
         }
 
+        public void SetInterruptAction(Action<Action> interruptAction)
+        {
+            InterruptAction = interruptAction;
+        }
+
         public void Screenshot()
         {
             _window.ScreenCaptureRequested = true;