Preload command speedup, Texture/buffer data flush, blit shader fix (#30)

* Move encoder state to be tied to command buffer, so preload and background cbs have their own encoder state

* Texture buffer/data flush, blit shader fix

src/Ryujinx.Graphics.Metal/BackgroundResources.cs

@@ -27,6 +27,7 @@ namespace Ryujinx.Graphics.Metal
             {
                 MTLCommandQueue queue = _renderer.BackgroundQueue;
                 _pool = new CommandBufferPool(queue);
+                _pool.Initialize(null); // TODO: Proper encoder factory for background render/compute
             }
 
             return _pool;

src/Ryujinx.Graphics.Metal/BufferHolder.cs

@@ -162,7 +162,7 @@ namespace Ryujinx.Graphics.Metal
             throw new InvalidOperationException("The buffer is not mapped.");
         }
 
-        public unsafe void SetData(int offset, ReadOnlySpan<byte> data, CommandBufferScoped? cbs = null, Action endRenderPass = null, bool allowCbsWait = true)
+        public unsafe void SetData(int offset, ReadOnlySpan<byte> data, CommandBufferScoped? cbs = null, bool allowCbsWait = true)
         {
             int dataSize = Math.Min(data.Length, Size - offset);
             if (dataSize == 0)
@@ -199,12 +199,11 @@ namespace Ryujinx.Graphics.Metal
                 // This avoids ending and beginning render passes on each buffer data upload.
 
                 cbs = _pipeline.PreloadCbs;
-                endRenderPass = null;
             }
 
             if (allowCbsWait)
             {
-                _renderer.BufferManager.StagingBuffer.PushData(_renderer.CommandBufferPool, cbs, endRenderPass, this, offset, data);
+                _renderer.BufferManager.StagingBuffer.PushData(_renderer.CommandBufferPool, cbs, this, offset, data);
             }
             else
             {
@@ -214,7 +213,7 @@ namespace Ryujinx.Graphics.Metal
                     cbs = _renderer.CommandBufferPool.Rent();
                 }
 
-                if (!_renderer.BufferManager.StagingBuffer.TryPushData(cbs.Value, endRenderPass, this, offset, data))
+                if (!_renderer.BufferManager.StagingBuffer.TryPushData(cbs.Value, this, offset, data))
                 {
                     // Need to do a slow upload.
                     BufferHolder srcHolder = _renderer.BufferManager.Create(dataSize);
@@ -223,7 +222,7 @@ namespace Ryujinx.Graphics.Metal
                     var srcBuffer = srcHolder.GetBuffer();
                     var dstBuffer = this.GetBuffer(true);
 
-                    Copy(_pipeline, cbs.Value, srcBuffer, dstBuffer, 0, offset, dataSize);
+                    Copy(cbs.Value, srcBuffer, dstBuffer, 0, offset, dataSize);
 
                     srcHolder.Dispose();
                 }
@@ -255,7 +254,6 @@ namespace Ryujinx.Graphics.Metal
         }
 
         public static void Copy(
-            Pipeline pipeline,
             CommandBufferScoped cbs,
             Auto<DisposableBuffer> src,
             Auto<DisposableBuffer> dst,
@@ -267,7 +265,7 @@ namespace Ryujinx.Graphics.Metal
             var srcBuffer = registerSrcUsage ? src.Get(cbs, srcOffset, size).Value : src.GetUnsafe().Value;
             var dstbuffer = dst.Get(cbs, dstOffset, size, true).Value;
 
-            pipeline.GetOrCreateBlitEncoder().CopyFromBuffer(
+            cbs.Encoders.EnsureBlitEncoder().CopyFromBuffer(
                 srcBuffer,
                 (ulong)srcOffset,
                 dstbuffer,

src/Ryujinx.Graphics.Metal/BufferManager.cs

@@ -176,14 +176,14 @@ namespace Ryujinx.Graphics.Metal
 
         public void SetData<T>(BufferHandle handle, int offset, ReadOnlySpan<T> data) where T : unmanaged
         {
-            SetData(handle, offset, MemoryMarshal.Cast<T, byte>(data), null, null);
+            SetData(handle, offset, MemoryMarshal.Cast<T, byte>(data), null);
         }
 
-        public void SetData(BufferHandle handle, int offset, ReadOnlySpan<byte> data, CommandBufferScoped? cbs, Action endRenderPass)
+        public void SetData(BufferHandle handle, int offset, ReadOnlySpan<byte> data, CommandBufferScoped? cbs)
         {
             if (TryGetBuffer(handle, out var holder))
             {
-                holder.SetData(offset, data, cbs, endRenderPass);
+                holder.SetData(offset, data, cbs);
             }
         }
 

src/Ryujinx.Graphics.Metal/CommandBufferEncoder.cs

@@ -0,0 +1,170 @@
+using Ryujinx.Graphics.Metal;
+using SharpMetal.Metal;
+using System;
+using System.Runtime.CompilerServices;
+using System.Runtime.Versioning;
+
+interface IEncoderFactory
+{
+    MTLRenderCommandEncoder CreateRenderCommandEncoder();
+    MTLComputeCommandEncoder CreateComputeCommandEncoder();
+}
+
+/// <summary>
+/// Tracks active encoder object for a command buffer.
+/// </summary>
+[SupportedOSPlatform("macos")]
+class CommandBufferEncoder
+{
+    public EncoderType CurrentEncoderType { get; private set; } = EncoderType.None;
+
+    public MTLBlitCommandEncoder BlitEncoder => new MTLBlitCommandEncoder(CurrentEncoder.Value);
+
+    public MTLComputeCommandEncoder ComputeEncoder => new MTLComputeCommandEncoder(CurrentEncoder.Value);
+
+    public MTLRenderCommandEncoder RenderEncoder => new MTLRenderCommandEncoder(CurrentEncoder.Value);
+
+    internal MTLCommandEncoder? CurrentEncoder { get; private set; }
+
+    private MTLCommandBuffer _commandBuffer;
+    private IEncoderFactory _encoderFactory;
+
+    public void Initialize(MTLCommandBuffer commandBuffer, IEncoderFactory encoderFactory)
+    {
+        _commandBuffer = commandBuffer;
+        _encoderFactory = encoderFactory;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public MTLRenderCommandEncoder EnsureRenderEncoder()
+    {
+        if (CurrentEncoderType != EncoderType.Render)
+        {
+            return BeginRenderPass();
+        }
+
+        return RenderEncoder;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public MTLBlitCommandEncoder EnsureBlitEncoder()
+    {
+        if (CurrentEncoderType != EncoderType.Blit)
+        {
+            return BeginBlitPass();
+        }
+
+        return BlitEncoder;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public MTLComputeCommandEncoder EnsureComputeEncoder()
+    {
+        if (CurrentEncoderType != EncoderType.Compute)
+        {
+            return BeginComputePass();
+        }
+
+        return ComputeEncoder;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public bool TryGetRenderEncoder(out MTLRenderCommandEncoder encoder)
+    {
+        if (CurrentEncoderType != EncoderType.Render)
+        {
+            encoder = default;
+            return false;
+        }
+
+        encoder = RenderEncoder;
+        return true;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public bool TryGetBlitEncoder(out MTLBlitCommandEncoder encoder)
+    {
+        if (CurrentEncoderType != EncoderType.Blit)
+        {
+            encoder = default;
+            return false;
+        }
+
+        encoder = BlitEncoder;
+        return true;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public bool TryGetComputeEncoder(out MTLComputeCommandEncoder encoder)
+    {
+        if (CurrentEncoderType != EncoderType.Compute)
+        {
+            encoder = default;
+            return false;
+        }
+
+        encoder = ComputeEncoder;
+        return true;
+    }
+
+    public void EndCurrentPass()
+    {
+        if (CurrentEncoder != null)
+        {
+            switch (CurrentEncoderType)
+            {
+                case EncoderType.Blit:
+                    BlitEncoder.EndEncoding();
+                    CurrentEncoder = null;
+                    break;
+                case EncoderType.Compute:
+                    ComputeEncoder.EndEncoding();
+                    CurrentEncoder = null;
+                    break;
+                case EncoderType.Render:
+                    RenderEncoder.EndEncoding();
+                    CurrentEncoder = null;
+                    break;
+                default:
+                    throw new InvalidOperationException();
+            }
+
+            CurrentEncoderType = EncoderType.None;
+        }
+    }
+
+    private MTLRenderCommandEncoder BeginRenderPass()
+    {
+        EndCurrentPass();
+
+        var renderCommandEncoder = _encoderFactory.CreateRenderCommandEncoder();
+
+        CurrentEncoder = renderCommandEncoder;
+        CurrentEncoderType = EncoderType.Render;
+
+        return renderCommandEncoder;
+    }
+
+    private MTLBlitCommandEncoder BeginBlitPass()
+    {
+        EndCurrentPass();
+
+        var descriptor = new MTLBlitPassDescriptor();
+        var blitCommandEncoder = _commandBuffer.BlitCommandEncoder(descriptor);
+
+        CurrentEncoder = blitCommandEncoder;
+        CurrentEncoderType = EncoderType.Blit;
+        return blitCommandEncoder;
+    }
+
+    private MTLComputeCommandEncoder BeginComputePass()
+    {
+        EndCurrentPass();
+
+        var computeCommandEncoder = _encoderFactory.CreateComputeCommandEncoder();
+
+        CurrentEncoder = computeCommandEncoder;
+        CurrentEncoderType = EncoderType.Compute;
+        return computeCommandEncoder;
+    }
+}

src/Ryujinx.Graphics.Metal/CommandBufferPool.cs

@@ -3,6 +3,7 @@ using System;
 using System.Collections.Generic;
 using System.Diagnostics;
 using System.Runtime.Versioning;
+using System.Threading;
 
 namespace Ryujinx.Graphics.Metal
 {
@@ -14,6 +15,10 @@ namespace Ryujinx.Graphics.Metal
         private readonly int _totalCommandBuffers;
         private readonly int _totalCommandBuffersMask;
         private readonly MTLCommandQueue _queue;
+        private readonly Thread _owner;
+        private IEncoderFactory _defaultEncoderFactory;
+
+        public bool OwnedByCurrentThread => _owner == Thread.CurrentThread;
 
         [SupportedOSPlatform("macos")]
         private struct ReservedCommandBuffer
@@ -22,22 +27,28 @@ namespace Ryujinx.Graphics.Metal
             public bool InConsumption;
             public int SubmissionCount;
             public MTLCommandBuffer CommandBuffer;
+            public CommandBufferEncoder Encoders;
             public FenceHolder Fence;
 
             public List<IAuto> Dependants;
             public List<MultiFenceHolder> Waitables;
 
-            public void Reinitialize(MTLCommandQueue queue)
+            public void Reinitialize(MTLCommandQueue queue, IEncoderFactory stateManager)
             {
                 CommandBuffer = queue.CommandBuffer();
+
+                Encoders.Initialize(CommandBuffer, stateManager);
             }
 
-            public void Initialize(MTLCommandQueue queue)
+            public void Initialize(MTLCommandQueue queue, IEncoderFactory stateManager)
             {
                 CommandBuffer = queue.CommandBuffer();
 
                 Dependants = new List<IAuto>();
                 Waitables = new List<MultiFenceHolder>();
+                Encoders = new CommandBufferEncoder();
+
+                Encoders.Initialize(CommandBuffer, stateManager);
             }
         }
 
@@ -51,6 +62,7 @@ namespace Ryujinx.Graphics.Metal
         public CommandBufferPool(MTLCommandQueue queue)
         {
             _queue = queue;
+            _owner = Thread.CurrentThread;
 
             _totalCommandBuffers = MaxCommandBuffers;
             _totalCommandBuffersMask = _totalCommandBuffers - 1;
@@ -60,10 +72,15 @@ namespace Ryujinx.Graphics.Metal
             _queuedIndexes = new int[_totalCommandBuffers];
             _queuedIndexesPtr = 0;
             _queuedCount = 0;
+        }
+
+        public void Initialize(IEncoderFactory encoderFactory)
+        {
+            _defaultEncoderFactory = encoderFactory;
 
             for (int i = 0; i < _totalCommandBuffers; i++)
             {
-                _commandBuffers[i].Initialize(_queue);
+                _commandBuffers[i].Initialize(_queue, _defaultEncoderFactory);
                 WaitAndDecrementRef(i);
             }
         }
@@ -194,7 +211,7 @@ namespace Ryujinx.Graphics.Metal
 
                         _inUseCount++;
 
-                        return new CommandBufferScoped(this, entry.CommandBuffer, cursor);
+                        return new CommandBufferScoped(this, entry.CommandBuffer, entry.Encoders, cursor);
                     }
 
                     cursor = (cursor + 1) & _totalCommandBuffersMask;
@@ -206,6 +223,9 @@ namespace Ryujinx.Graphics.Metal
 
         public void Return(CommandBufferScoped cbs)
         {
+            // Ensure the encoder is committed.
+            cbs.Encoders.EndCurrentPass();
+
             lock (_commandBuffers)
             {
                 int cbIndex = cbs.CommandBufferIndex;
@@ -223,7 +243,7 @@ namespace Ryujinx.Graphics.Metal
                 commandBuffer.Commit();
 
                 // Replace entry with new MTLCommandBuffer
-                entry.Reinitialize(_queue);
+                entry.Reinitialize(_queue, _defaultEncoderFactory);
 
                 int ptr = (_queuedIndexesPtr + _queuedCount) % _totalCommandBuffers;
                 _queuedIndexes[ptr] = cbIndex;

src/Ryujinx.Graphics.Metal/CommandBufferScoped.cs

@@ -9,12 +9,14 @@ namespace Ryujinx.Graphics.Metal
     {
         private readonly CommandBufferPool _pool;
         public MTLCommandBuffer CommandBuffer { get; }
+        public CommandBufferEncoder Encoders { get; }
         public int CommandBufferIndex { get; }
 
-        public CommandBufferScoped(CommandBufferPool pool, MTLCommandBuffer commandBuffer, int commandBufferIndex)
+        public CommandBufferScoped(CommandBufferPool pool, MTLCommandBuffer commandBuffer, CommandBufferEncoder encoders, int commandBufferIndex)
         {
             _pool = pool;
             CommandBuffer = commandBuffer;
+            Encoders = encoders;
             CommandBufferIndex = commandBufferIndex;
         }
 

src/Ryujinx.Graphics.Metal/EncoderStateManager.cs

@@ -581,9 +581,8 @@ namespace Ryujinx.Graphics.Metal
             _currentState.DepthClipMode = clamp ? MTLDepthClipMode.Clamp : MTLDepthClipMode.Clip;
 
             // Inline update
-            if (_pipeline.CurrentEncoderType == EncoderType.Render && _pipeline.CurrentEncoder != null)
+            if (_pipeline.Encoders.TryGetRenderEncoder(out MTLRenderCommandEncoder renderCommandEncoder))
             {
-                var renderCommandEncoder = new MTLRenderCommandEncoder(_pipeline.CurrentEncoder.Value);
                 SetDepthClamp(renderCommandEncoder);
                 return;
             }
@@ -600,9 +599,8 @@ namespace Ryujinx.Graphics.Metal
             _currentState.Clamp = clamp;
 
             // Inline update
-            if (_pipeline.CurrentEncoderType == EncoderType.Render && _pipeline.CurrentEncoder != null)
+            if (_pipeline.Encoders.TryGetRenderEncoder(out MTLRenderCommandEncoder renderCommandEncoder))
             {
-                var renderCommandEncoder = new MTLRenderCommandEncoder(_pipeline.CurrentEncoder.Value);
                 SetDepthBias(renderCommandEncoder);
                 return;
             }
@@ -632,9 +630,8 @@ namespace Ryujinx.Graphics.Metal
             }
 
             // Inline update
-            if (_pipeline.CurrentEncoderType == EncoderType.Render && _pipeline.CurrentEncoder != null)
+            if (_pipeline.Encoders.TryGetRenderEncoder(out MTLRenderCommandEncoder renderCommandEncoder))
             {
-                var renderCommandEncoder = new MTLRenderCommandEncoder(_pipeline.CurrentEncoder.Value);
                 SetScissors(renderCommandEncoder);
                 return;
             }
@@ -669,9 +666,8 @@ namespace Ryujinx.Graphics.Metal
             }
 
             // Inline update
-            if (_pipeline.CurrentEncoderType == EncoderType.Render && _pipeline.CurrentEncoder != null)
+            if (_pipeline.Encoders.TryGetRenderEncoder(out MTLRenderCommandEncoder renderCommandEncoder))
             {
-                var renderCommandEncoder = new MTLRenderCommandEncoder(_pipeline.CurrentEncoder.Value);
                 SetViewports(renderCommandEncoder);
                 return;
             }
@@ -688,9 +684,8 @@ namespace Ryujinx.Graphics.Metal
             UpdatePipelineVertexState(_currentState.VertexBuffers, _currentState.VertexAttribs);
 
             // Inline update
-            if (_pipeline.CurrentEncoderType == EncoderType.Render && _pipeline.CurrentEncoder != null)
+            if (_pipeline.Encoders.TryGetRenderEncoder(out MTLRenderCommandEncoder renderCommandEncoder))
             {
-                var renderCommandEncoder = new MTLRenderCommandEncoder(_pipeline.CurrentEncoder.Value);
                 SetVertexBuffers(renderCommandEncoder, _currentState.VertexBuffers);
                 return;
             }
@@ -755,9 +750,8 @@ namespace Ryujinx.Graphics.Metal
             _currentState.CullBoth = face == Face.FrontAndBack;
 
             // Inline update
-            if (_pipeline.CurrentEncoderType == EncoderType.Render && _pipeline.CurrentEncoder != null)
+            if (_pipeline.Encoders.TryGetRenderEncoder(out MTLRenderCommandEncoder renderCommandEncoder))
             {
-                var renderCommandEncoder = new MTLRenderCommandEncoder(_pipeline.CurrentEncoder.Value);
                 SetCullMode(renderCommandEncoder);
                 SetScissors(renderCommandEncoder);
                 return;
@@ -778,9 +772,8 @@ namespace Ryujinx.Graphics.Metal
             _currentState.Winding = frontFace.Convert();
 
             // Inline update
-            if (_pipeline.CurrentEncoderType == EncoderType.Render && _pipeline.CurrentEncoder != null)
+            if (_pipeline.Encoders.TryGetRenderEncoder(out MTLRenderCommandEncoder renderCommandEncoder))
             {
-                var renderCommandEncoder = new MTLRenderCommandEncoder(_pipeline.CurrentEncoder.Value);
                 SetFrontFace(renderCommandEncoder);
                 return;
             }
@@ -795,9 +788,8 @@ namespace Ryujinx.Graphics.Metal
             _currentState.BackRefValue = backRef;
 
             // Inline update
-            if (_pipeline.CurrentEncoderType == EncoderType.Render && _pipeline.CurrentEncoder != null)
+            if (_pipeline.Encoders.TryGetRenderEncoder(out MTLRenderCommandEncoder renderCommandEncoder))
             {
-                var renderCommandEncoder = new MTLRenderCommandEncoder(_pipeline.CurrentEncoder.Value);
                 SetStencilRefValue(renderCommandEncoder);
             }
 

src/Ryujinx.Graphics.Metal/HelperShader.cs

@@ -277,7 +277,7 @@ namespace Ryujinx.Graphics.Metal
             DirtyFlags clearFlags = DirtyFlags.All & (~DirtyFlags.Scissors);
 
             // Save current state
-            EncoderState originalState = _pipeline.SwapState(_helperShaderState, clearFlags);
+            EncoderState originalState = _pipeline.SwapState(_helperShaderState, clearFlags, false);
 
             // Inherit some state without fully recreating render pipeline.
             RenderTargetCopy save = _helperShaderState.InheritForClear(originalState, false, index);
@@ -312,7 +312,7 @@ namespace Ryujinx.Graphics.Metal
             _pipeline.Draw(4, 1, 0, 0);
 
             // Restore previous state
-            _pipeline.SwapState(null, clearFlags);
+            _pipeline.SwapState(null, clearFlags, false);
 
             _helperShaderState.Restore(save);
         }
@@ -330,7 +330,7 @@ namespace Ryujinx.Graphics.Metal
             var helperScissors = _helperShaderState.Scissors;
 
             // Save current state
-            EncoderState originalState = _pipeline.SwapState(_helperShaderState, clearFlags);
+            EncoderState originalState = _pipeline.SwapState(_helperShaderState, clearFlags, false);
 
             // Inherit some state without fully recreating render pipeline.
             RenderTargetCopy save = _helperShaderState.InheritForClear(originalState, true);
@@ -365,7 +365,7 @@ namespace Ryujinx.Graphics.Metal
             _pipeline.SetStencilTest(CreateStencilTestDescriptor(false));
 
             // Restore previous state
-            _pipeline.SwapState(null, clearFlags);
+            _pipeline.SwapState(null, clearFlags, false);
 
             _helperShaderState.Restore(save);
         }

src/Ryujinx.Graphics.Metal/MetalRenderer.cs

@@ -67,7 +67,10 @@ namespace Ryujinx.Graphics.Metal
 
         public void BackgroundContextAction(Action action, bool alwaysBackground = false)
         {
-            Logger.Warning?.Print(LogClass.Gpu, "Not Implemented!");
+            // GetData methods should be thread safe, so we can call this directly.
+            // Texture copy (scaled) may also happen in here, so that should also be thread safe.
+
+            action();
         }
 
         public BufferHandle CreateBuffer(int size, BufferAccess access)
@@ -221,7 +224,7 @@ namespace Ryujinx.Graphics.Metal
 
         public void SetBufferData(BufferHandle buffer, int offset, ReadOnlySpan<byte> data)
         {
-            BufferManager.SetData(buffer, offset, data, _pipeline.Cbs, _pipeline.EndRenderPassDelegate);
+            BufferManager.SetData(buffer, offset, data, _pipeline.Cbs);
         }
 
         public void UpdateCounters()

src/Ryujinx.Graphics.Metal/PersistentFlushBuffer.cs

@@ -1,3 +1,4 @@
+using Ryujinx.Graphics.GAL;
 using System;
 using System.Runtime.Versioning;
 
@@ -44,7 +45,7 @@ namespace Ryujinx.Graphics.Metal
 
                 if (srcBuffer.TryIncrementReferenceCount())
                 {
-                    BufferHolder.Copy(_pipeline, cbs, srcBuffer, dstBuffer, offset, 0, size, registerSrcUsage: false);
+                    BufferHolder.Copy(cbs, srcBuffer, dstBuffer, offset, 0, size, registerSrcUsage: false);
                 }
                 else
                 {
@@ -58,6 +59,40 @@ namespace Ryujinx.Graphics.Metal
             return flushStorage.GetDataStorage(0, size);
         }
 
+        public Span<byte> GetTextureData(CommandBufferPool cbp, Texture view, int size)
+        {
+            TextureCreateInfo info = view.Info;
+
+            var flushStorage = ResizeIfNeeded(size);
+
+            using (var cbs = cbp.Rent())
+            {
+                var buffer = flushStorage.GetBuffer().Get(cbs).Value;
+                var image = view.GetHandle();
+
+                view.CopyFromOrToBuffer(cbs, buffer, image, size, true, 0, 0, info.GetLayers(), info.Levels, singleSlice: false);
+            }
+
+            flushStorage.WaitForFences();
+            return flushStorage.GetDataStorage(0, size);
+        }
+
+        public Span<byte> GetTextureData(CommandBufferPool cbp, Texture view, int size, int layer, int level)
+        {
+            var flushStorage = ResizeIfNeeded(size);
+
+            using (var cbs = cbp.Rent())
+            {
+                var buffer = flushStorage.GetBuffer().Get(cbs).Value;
+                var image = view.GetHandle();
+
+                view.CopyFromOrToBuffer(cbs, buffer, image, size, true, layer, level, 1, 1, singleSlice: true);
+            }
+
+            flushStorage.WaitForFences();
+            return flushStorage.GetDataStorage(0, size);
+        }
+
         public void Dispose()
         {
             _flushStorage.Dispose();

src/Ryujinx.Graphics.Metal/Pipeline.cs

@@ -18,7 +18,7 @@ namespace Ryujinx.Graphics.Metal
     }
 
     [SupportedOSPlatform("macos")]
-    class Pipeline : IPipeline, IDisposable
+    class Pipeline : IPipeline, IEncoderFactory, IDisposable
     {
         private const ulong MinByteWeightForFlush = 256 * 1024 * 1024; // MiB
 
@@ -27,7 +27,6 @@ namespace Ryujinx.Graphics.Metal
         private EncoderStateManager _encoderStateManager;
         private ulong _byteWeight;
 
-        public readonly Action EndRenderPassDelegate;
         public MTLCommandBuffer CommandBuffer;
 
         public IndexBufferPattern QuadsToTrisPattern;
@@ -35,8 +34,8 @@ namespace Ryujinx.Graphics.Metal
 
         internal CommandBufferScoped? PreloadCbs { get; private set; }
         internal CommandBufferScoped Cbs { get; private set; }
-        internal MTLCommandEncoder? CurrentEncoder { get; private set; }
+        internal CommandBufferEncoder Encoders => Cbs.Encoders;
-        internal EncoderType CurrentEncoderType { get; private set; } = EncoderType.None;
+        internal EncoderType CurrentEncoderType => Encoders.CurrentEncoderType;
         internal bool RenderPassActive { get; private set; }
 
         public Pipeline(MTLDevice device, MetalRenderer renderer)
@@ -44,7 +43,7 @@ namespace Ryujinx.Graphics.Metal
             _device = device;
             _renderer = renderer;
 
-            EndRenderPassDelegate = EndCurrentPass;
+            renderer.CommandBufferPool.Initialize(this);
 
             CommandBuffer = (Cbs = _renderer.CommandBufferPool.Rent()).CommandBuffer;
         }
@@ -57,8 +56,13 @@ namespace Ryujinx.Graphics.Metal
             TriFanToTrisPattern = new IndexBufferPattern(_renderer, 3, 3, 2, [int.MinValue, -1, 0], 1, true);
         }
 
-        public EncoderState SwapState(EncoderState state, DirtyFlags flags = DirtyFlags.All)
+        public EncoderState SwapState(EncoderState state, DirtyFlags flags = DirtyFlags.All, bool endRenderPass = true)
         {
+            if (endRenderPass && CurrentEncoderType == EncoderType.Render)
+            {
+                EndCurrentPass();
+            }
+
             return _encoderStateManager.SwapState(state, flags);
         }
 
@@ -79,15 +83,7 @@ namespace Ryujinx.Graphics.Metal
 
         public MTLRenderCommandEncoder GetOrCreateRenderEncoder(bool forDraw = false)
         {
-            MTLRenderCommandEncoder renderCommandEncoder;
+            MTLRenderCommandEncoder renderCommandEncoder = Cbs.Encoders.EnsureRenderEncoder();
-            if (CurrentEncoder == null || CurrentEncoderType != EncoderType.Render)
-            {
-                renderCommandEncoder = BeginRenderPass();
-            }
-            else
-            {
-                renderCommandEncoder = new MTLRenderCommandEncoder(CurrentEncoder.Value);
-            }
 
             if (forDraw)
             {
@@ -99,28 +95,12 @@ namespace Ryujinx.Graphics.Metal
 
         public MTLBlitCommandEncoder GetOrCreateBlitEncoder()
         {
-            if (CurrentEncoder != null)
+            return Cbs.Encoders.EnsureBlitEncoder();
-            {
-                if (CurrentEncoderType == EncoderType.Blit)
-                {
-                    return new MTLBlitCommandEncoder(CurrentEncoder.Value);
-                }
-            }
-
-            return BeginBlitPass();
         }
 
         public MTLComputeCommandEncoder GetOrCreateComputeEncoder(bool forDispatch = false)
         {
-            MTLComputeCommandEncoder computeCommandEncoder;
+            MTLComputeCommandEncoder computeCommandEncoder = Cbs.Encoders.EnsureComputeEncoder();
-            if (CurrentEncoder == null || CurrentEncoderType != EncoderType.Compute)
-            {
-                computeCommandEncoder = BeginComputePass();
-            }
-            else
-            {
-                computeCommandEncoder = new MTLComputeCommandEncoder(CurrentEncoder.Value);
-            }
 
             if (forDispatch)
             {
@@ -132,65 +112,17 @@ namespace Ryujinx.Graphics.Metal
 
         public void EndCurrentPass()
         {
-            if (CurrentEncoder != null)
+            Cbs.Encoders.EndCurrentPass();
-            {
-                switch (CurrentEncoderType)
-                {
-                    case EncoderType.Blit:
-                        new MTLBlitCommandEncoder(CurrentEncoder.Value).EndEncoding();
-                        CurrentEncoder = null;
-                        break;
-                    case EncoderType.Compute:
-                        new MTLComputeCommandEncoder(CurrentEncoder.Value).EndEncoding();
-                        CurrentEncoder = null;
-                        break;
-                    case EncoderType.Render:
-                        new MTLRenderCommandEncoder(CurrentEncoder.Value).EndEncoding();
-                        CurrentEncoder = null;
-                        RenderPassActive = false;
-                        break;
-                    default:
-                        throw new ArgumentOutOfRangeException();
         }
 
-                CurrentEncoderType = EncoderType.None;
+        public MTLRenderCommandEncoder CreateRenderCommandEncoder()
-            }
-        }
-
-        private MTLRenderCommandEncoder BeginRenderPass()
         {
-            EndCurrentPass();
+            return _encoderStateManager.CreateRenderCommandEncoder();
-
-            var renderCommandEncoder = _encoderStateManager.CreateRenderCommandEncoder();
-
-            CurrentEncoder = renderCommandEncoder;
-            CurrentEncoderType = EncoderType.Render;
-            RenderPassActive = true;
-
-            return renderCommandEncoder;
         }
 
-        private MTLBlitCommandEncoder BeginBlitPass()
+        public MTLComputeCommandEncoder CreateComputeCommandEncoder()
         {
-            EndCurrentPass();
+            return _encoderStateManager.CreateComputeCommandEncoder();
-
-            var descriptor = new MTLBlitPassDescriptor();
-            var blitCommandEncoder = Cbs.CommandBuffer.BlitCommandEncoder(descriptor);
-
-            CurrentEncoder = blitCommandEncoder;
-            CurrentEncoderType = EncoderType.Blit;
-            return blitCommandEncoder;
-        }
-
-        private MTLComputeCommandEncoder BeginComputePass()
-        {
-            EndCurrentPass();
-
-            var computeCommandEncoder = _encoderStateManager.CreateComputeCommandEncoder();
-
-            CurrentEncoder = computeCommandEncoder;
-            CurrentEncoderType = EncoderType.Compute;
-            return computeCommandEncoder;
         }
 
         public void Present(CAMetalDrawable drawable, Texture src, Extents2D srcRegion, Extents2D dstRegion, bool isLinear)
@@ -279,19 +211,15 @@ namespace Ryujinx.Graphics.Metal
             {
                 case EncoderType.Render:
                     {
-                        var renderCommandEncoder = GetOrCreateRenderEncoder();
-
                         var scope = MTLBarrierScope.Buffers | MTLBarrierScope.Textures | MTLBarrierScope.RenderTargets;
                         MTLRenderStages stages = MTLRenderStages.RenderStageVertex | MTLRenderStages.RenderStageFragment;
-                        renderCommandEncoder.MemoryBarrier(scope, stages, stages);
+                        Encoders.RenderEncoder.MemoryBarrier(scope, stages, stages);
                         break;
                     }
                 case EncoderType.Compute:
                     {
-                        var computeCommandEncoder = GetOrCreateComputeEncoder();
-
                         var scope = MTLBarrierScope.Buffers | MTLBarrierScope.Textures | MTLBarrierScope.RenderTargets;;
-                        computeCommandEncoder.MemoryBarrier(scope);
+                        Encoders.ComputeEncoder.MemoryBarrier(scope);
                         break;
                     }
             }
@@ -353,7 +281,7 @@ namespace Ryujinx.Graphics.Metal
             var srcBuffer = _renderer.BufferManager.GetBuffer(src, srcOffset, size, false);
             var dstBuffer = _renderer.BufferManager.GetBuffer(dst, dstOffset, size, true);
 
-            BufferHolder.Copy(this, Cbs, srcBuffer, dstBuffer, srcOffset, dstOffset, size);
+            BufferHolder.Copy(Cbs, srcBuffer, dstBuffer, srcOffset, dstOffset, size);
         }
 
         public void DispatchCompute(int groupsX, int groupsY, int groupsZ)
@@ -709,9 +637,7 @@ namespace Ryujinx.Graphics.Metal
         {
             if (CurrentEncoderType == EncoderType.Render)
             {
-                var renderCommandEncoder = GetOrCreateRenderEncoder();
+                Encoders.RenderEncoder.MemoryBarrier(MTLBarrierScope.Textures, MTLRenderStages.RenderStageFragment, MTLRenderStages.RenderStageFragment);
-
-                renderCommandEncoder.MemoryBarrier(MTLBarrierScope.Textures, MTLRenderStages.RenderStageFragment, MTLRenderStages.RenderStageFragment);
             }
         }
 

src/Ryujinx.Graphics.Metal/StagingBuffer.cs

@@ -63,15 +63,13 @@ namespace Ryujinx.Graphics.Metal
             _resourceAlignment = Constants.MinResourceAlignment;
         }
 
-        public void PushData(CommandBufferPool cbp, CommandBufferScoped? cbs, Action endRenderPass, BufferHolder dst, int dstOffset, ReadOnlySpan<byte> data)
+        public void PushData(CommandBufferPool cbp, CommandBufferScoped? cbs, BufferHolder dst, int dstOffset, ReadOnlySpan<byte> data)
         {
             bool isRender = cbs != null;
             CommandBufferScoped scoped = cbs ?? cbp.Rent();
 
             // Must push all data to the buffer. If it can't fit, split it up.
 
-            endRenderPass?.Invoke();
-
             while (data.Length > 0)
             {
                 if (_freeSize < data.Length)
@@ -122,14 +120,14 @@ namespace Ryujinx.Graphics.Metal
                 _buffer.SetDataUnchecked(offset, data[..capacity]);
                 _buffer.SetDataUnchecked(0, data[capacity..]);
 
-                BufferHolder.Copy(_pipeline, cbs, srcBuffer, dstBuffer, offset, dstOffset, capacity);
+                BufferHolder.Copy(cbs, srcBuffer, dstBuffer, offset, dstOffset, capacity);
-                BufferHolder.Copy(_pipeline, cbs, srcBuffer, dstBuffer, 0, dstOffset + capacity, data.Length - capacity);
+                BufferHolder.Copy(cbs, srcBuffer, dstBuffer, 0, dstOffset + capacity, data.Length - capacity);
             }
             else
             {
                 _buffer.SetDataUnchecked(offset, data);
 
-                BufferHolder.Copy(_pipeline, cbs, srcBuffer, dstBuffer, offset, dstOffset, data.Length);
+                BufferHolder.Copy(cbs, srcBuffer, dstBuffer, offset, dstOffset, data.Length);
             }
 
             _freeOffset = (offset + data.Length) & (BufferSize - 1);
@@ -139,7 +137,7 @@ namespace Ryujinx.Graphics.Metal
             _pendingCopies.Enqueue(new PendingCopy(cbs.GetFence(), data.Length));
         }
 
-        public bool TryPushData(CommandBufferScoped cbs, Action endRenderPass, BufferHolder dst, int dstOffset, ReadOnlySpan<byte> data)
+        public bool TryPushData(CommandBufferScoped cbs, BufferHolder dst, int dstOffset, ReadOnlySpan<byte> data)
         {
             if (data.Length > BufferSize)
             {
@@ -156,8 +154,6 @@ namespace Ryujinx.Graphics.Metal
                 }
             }
 
-            endRenderPass?.Invoke();
-
             PushDataImpl(cbs, dst, dstOffset, data);
 
             return true;

src/Ryujinx.Graphics.Metal/Texture.cs

@@ -1,3 +1,4 @@
+using Ryujinx.Common.Logging;
 using Ryujinx.Graphics.GAL;
 using SharpMetal.Foundation;
 using SharpMetal.Metal;
@@ -94,6 +95,13 @@ namespace Ryujinx.Graphics.Metal
 
         public void CopyTo(ITexture destination, int firstLayer, int firstLevel)
         {
+            if (!_renderer.CommandBufferPool.OwnedByCurrentThread)
+            {
+                Logger.Warning?.PrintMsg(LogClass.Gpu, "Metal doesn't currently support scaled blit on background thread.");
+
+                return;
+            }
+
             var blitCommandEncoder = _pipeline.GetOrCreateBlitEncoder();
 
             if (destination is Texture destinationTexture)
@@ -202,48 +210,92 @@ namespace Ryujinx.Graphics.Metal
             return new Texture(_device, _renderer, _pipeline, info, _mtlTexture, firstLayer, firstLevel);
         }
 
-        public PinnedSpan<byte> GetData()
+        private int GetBufferDataLength(int size)
         {
-            var blitCommandEncoder = _pipeline.GetOrCreateBlitEncoder();
+            // TODO: D32S8 conversion
 
-            ulong length = 0;
+            return size;
-
-            for (int level = 0; level < Info.Levels; level++)
-            {
-                length += (ulong)Info.GetMipSize(level);
         }
 
-            unsafe
+        private ReadOnlySpan<byte> GetDataFromBuffer(ReadOnlySpan<byte> storage, int size, Span<byte> output)
         {
-                var mtlBuffer = _device.NewBuffer(length, MTLResourceOptions.ResourceStorageModeShared);
+            // TODO: D32S8 conversion
+
+            return storage;
+        }
+
+        public void CopyFromOrToBuffer(
+            CommandBufferScoped cbs,
+            MTLBuffer buffer,
+            MTLTexture image,
+            int size,
+            bool to,
+            int dstLayer,
+            int dstLevel,
+            int dstLayers,
+            int dstLevels,
+            bool singleSlice,
+            int offset = 0,
+            int stride = 0)
+        {
+            MTLBlitCommandEncoder blitCommandEncoder = cbs.Encoders.EnsureBlitEncoder();
 
-                int width = Info.Width;
-                int height = Info.Height;
-                int depth = Info.Depth;
-                int levels = Info.GetLevelsClamped();
-                int layers = Info.GetLayers();
             bool is3D = Info.Target == Target.Texture3D;
+            int width = Math.Max(1, Info.Width >> dstLevel);
+            int height = Math.Max(1, Info.Height >> dstLevel);
+            int depth = is3D && !singleSlice ? Math.Max(1, Info.Depth >> dstLevel) : 1;
+            int layers = dstLayers;
+            int levels = dstLevels;
 
-                int offset = 0;
+            for (int oLevel = 0; oLevel < levels; oLevel++)
-
-                for (int level = 0; level < levels; level++)
             {
+                int level = oLevel + dstLevel;
                 int mipSize = Info.GetMipSize2D(level);
-                    int endOffset = offset + mipSize;
 
-                    for (int layer = 0; layer < layers; layer++)
+                int mipSizeLevel = GetBufferDataLength(is3D && !singleSlice
+                    ? Info.GetMipSize(level)
+                    : mipSize * dstLayers);
+
+                int endOffset = offset + mipSizeLevel;
+
+                if ((uint)endOffset > (uint)size)
+                {
+                    break;
+                }
+
+                for (int oLayer = 0; oLayer < layers; oLayer++)
+                {
+                    int layer = !is3D ? dstLayer + oLayer : 0;
+                    int z = is3D ? dstLayer + oLayer : 0;
+
+                    if (to)
                     {
                         blitCommandEncoder.CopyFromTexture(
-                            _mtlTexture,
+                            image,
                             (ulong)layer,
                             (ulong)level,
-                            new MTLOrigin(),
+                            new MTLOrigin { z = (ulong)z },
-                            new MTLSize { width = (ulong)width, height = (ulong)height, depth = is3D ? (ulong)depth : 1 },
+                            new MTLSize { width = (ulong)width, height = (ulong)height, depth = 1 },
-                            mtlBuffer,
+                            buffer,
                             (ulong)offset,
                             (ulong)Info.GetMipStride(level),
                             (ulong)mipSize
                         );
+                    }
+                    else
+                    {
+                        blitCommandEncoder.CopyFromBuffer(
+                            buffer,
+                            (ulong)offset,
+                            (ulong)Info.GetMipStride(level),
+                            (ulong)mipSize,
+                            new MTLSize { width = (ulong)width, height = (ulong)height, depth = 1 },
+                            image,
+                            (ulong)(layer + oLayer),
+                            (ulong)level,
+                            new MTLOrigin { z = (ulong)z }
+                        );
+                    }
 
                     offset += mipSize;
                 }
@@ -251,49 +303,64 @@ namespace Ryujinx.Graphics.Metal
                 width = Math.Max(1, width >> 1);
                 height = Math.Max(1, height >> 1);
 
-                    if (is3D)
+                if (Info.Target == Target.Texture3D)
                 {
                     depth = Math.Max(1, depth >> 1);
                 }
             }
-
-                // TODO: wait
-
-                return new PinnedSpan<byte>(mtlBuffer.Contents.ToPointer(), (int)length, () => mtlBuffer.Dispose());
         }
+
+        private ReadOnlySpan<byte> GetData(CommandBufferPool cbp, PersistentFlushBuffer flushBuffer)
+        {
+            int size = 0;
+
+            for (int level = 0; level < Info.Levels; level++)
+            {
+                size += Info.GetMipSize(level);
+            }
+
+            size = GetBufferDataLength(size);
+
+            Span<byte> result = flushBuffer.GetTextureData(cbp, this, size);
+
+            return GetDataFromBuffer(result, size, result);
+        }
+
+        private ReadOnlySpan<byte> GetData(CommandBufferPool cbp, PersistentFlushBuffer flushBuffer, int layer, int level)
+        {
+            int size = GetBufferDataLength(Info.GetMipSize(level));
+
+            Span<byte> result = flushBuffer.GetTextureData(cbp, this, size, layer, level);
+
+            return GetDataFromBuffer(result, size, result);
+        }
+
+        public PinnedSpan<byte> GetData()
+        {
+            BackgroundResource resources = _renderer.BackgroundResources.Get();
+
+            if (_renderer.CommandBufferPool.OwnedByCurrentThread)
+            {
+                _renderer.FlushAllCommands();
+
+                return PinnedSpan<byte>.UnsafeFromSpan(GetData(_renderer.CommandBufferPool, resources.GetFlushBuffer()));
+            }
+
+            return PinnedSpan<byte>.UnsafeFromSpan(GetData(resources.GetPool(), resources.GetFlushBuffer()));
         }
 
         public PinnedSpan<byte> GetData(int layer, int level)
         {
-            var blitCommandEncoder = _pipeline.GetOrCreateBlitEncoder();
+            BackgroundResource resources = _renderer.BackgroundResources.Get();
 
-            ulong bytesPerRow = (ulong)Info.GetMipStride(level);
+            if (_renderer.CommandBufferPool.OwnedByCurrentThread)
-            ulong length = bytesPerRow * (ulong)Info.Height;
-            ulong bytesPerImage = 0;
-            if (_mtlTexture.TextureType == MTLTextureType.Type3D)
             {
-                bytesPerImage = length;
+                _renderer.FlushAllCommands();
+
+                return PinnedSpan<byte>.UnsafeFromSpan(GetData(_renderer.CommandBufferPool, resources.GetFlushBuffer(), layer, level));
             }
 
-            unsafe
+            return PinnedSpan<byte>.UnsafeFromSpan(GetData(resources.GetPool(), resources.GetFlushBuffer(), layer, level));
-            {
-
-                var mtlBuffer = _device.NewBuffer(length, MTLResourceOptions.ResourceStorageModeShared);
-
-                blitCommandEncoder.CopyFromTexture(
-                    _mtlTexture,
-                    (ulong)layer,
-                    (ulong)level,
-                    new MTLOrigin(),
-                    new MTLSize { width = _mtlTexture.Width, height = _mtlTexture.Height, depth = _mtlTexture.Depth },
-                    mtlBuffer,
-                    0,
-                    bytesPerRow,
-                    bytesPerImage
-                );
-
-                return new PinnedSpan<byte>(mtlBuffer.Contents.ToPointer(), (int)length, () => mtlBuffer.Dispose());
-            }
         }
 
         public void SetData(IMemoryOwner<byte> data)