Vulkan: Cache delegate for EndRenderPass (#6132)

This prevents a small allocation each time this method is called. This is a top 3 SOH allocation during gameplay in most games, and eliminating it is pretty free.

.github/dependabot.yml

@@ -13,7 +13,7 @@ updates:
 
   - package-ecosystem: nuget
     directory: /
-    open-pull-requests-limit: 5
+    open-pull-requests-limit: 10
     schedule:
       interval: daily
     labels:
@@ -22,3 +22,19 @@ updates:
       - marysaka
     commit-message:
       prefix: nuget
+    groups:
+      Avalonia:
+        patterns:
+          - "*Avalonia*"
+      Silk.NET:
+        patterns:
+          - "Silk.NET*"
+      OpenTK:
+        patterns:
+          - "OpenTK*"
+      SixLabors:
+        patterns:
+          - "SixLabors*"
+      NUnit:
+        patterns:
+          - "NUnit*"

src/Ryujinx.Cpu/AppleHv/HvCodePatcher.cs

@@ -1,62 +0,0 @@
-using System;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-
-namespace Ryujinx.Cpu.AppleHv
-{
-    static class HvCodePatcher
-    {
-        private const uint XMask = 0x3f808000u;
-        private const uint XValue = 0x8000000u;
-
-        private const uint ZrIndex = 31u;
-
-        public static void RewriteUnorderedExclusiveInstructions(Span<byte> code)
-        {
-            Span<uint> codeUint = MemoryMarshal.Cast<byte, uint>(code);
-            Span<Vector128<uint>> codeVector = MemoryMarshal.Cast<byte, Vector128<uint>>(code);
-
-            Vector128<uint> mask = Vector128.Create(XMask);
-            Vector128<uint> value = Vector128.Create(XValue);
-
-            for (int index = 0; index < codeVector.Length; index++)
-            {
-                Vector128<uint> v = codeVector[index];
-
-                if (Vector128.EqualsAny(Vector128.BitwiseAnd(v, mask), value))
-                {
-                    int baseIndex = index * 4;
-
-                    for (int instIndex = baseIndex; instIndex < baseIndex + 4; instIndex++)
-                    {
-                        ref uint inst = ref codeUint[instIndex];
-
-                        if ((inst & XMask) != XValue)
-                        {
-                            continue;
-                        }
-
-                        bool isPair = (inst & (1u << 21)) != 0;
-                        bool isLoad = (inst & (1u << 22)) != 0;
-
-                        uint rt2 = (inst >> 10) & 0x1fu;
-                        uint rs = (inst >> 16) & 0x1fu;
-
-                        if (isLoad && rs != ZrIndex)
-                        {
-                            continue;
-                        }
-
-                        if (!isPair && rt2 != ZrIndex)
-                        {
-                            continue;
-                        }
-
-                        // Set the ordered flag.
-                        inst |= 1u << 15;
-                    }
-                }
-            }
-        }
-    }
-}

src/Ryujinx.Cpu/AppleHv/HvMemoryManager.cs

@@ -724,18 +724,6 @@ namespace Ryujinx.Cpu.AppleHv
         /// <inheritdoc/>
         public void Reprotect(ulong va, ulong size, MemoryPermission protection)
         {
-            if (protection.HasFlag(MemoryPermission.Execute))
-            {
-                // Some applications use unordered exclusive memory access instructions
-                // where it is not valid to do so, leading to memory re-ordering that
-                // makes the code behave incorrectly on some CPUs.
-                // To work around this, we force all such accesses to be ordered.
-
-                using WritableRegion writableRegion = GetWritableRegion(va, (int)size);
-
-                HvCodePatcher.RewriteUnorderedExclusiveInstructions(writableRegion.Memory.Span);
-            }
-
             // TODO
         }
 

src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawManager.cs

@@ -103,6 +103,8 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
         /// <param name="argument">Method call argument</param>
         public void DrawEnd(ThreedClass engine, int argument)
         {
+            _drawState.DrawUsesEngineState = true;
+
             DrawEnd(
                 engine,
                 _state.State.IndexBufferState.First,
@@ -205,10 +207,6 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
             }
             else
             {
-#pragma warning disable IDE0059 // Remove unnecessary value assignment
-                var drawState = _state.State.VertexBufferDrawState;
-#pragma warning restore IDE0059
-
                 DrawImpl(engine, drawVertexCount, 1, 0, drawFirstVertex, firstInstance, indexed: false);
             }
 
@@ -379,6 +377,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
             bool oldDrawIndexed = _drawState.DrawIndexed;
 
             _drawState.DrawIndexed = true;
+            _drawState.DrawUsesEngineState = false;
             engine.ForceStateDirty(IndexBufferCountMethodOffset * 4);
 
             DrawEnd(engine, firstIndex, indexCount, 0, 0);
@@ -424,6 +423,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
             bool oldDrawIndexed = _drawState.DrawIndexed;
 
             _drawState.DrawIndexed = false;
+            _drawState.DrawUsesEngineState = false;
             engine.ForceStateDirty(VertexBufferFirstMethodOffset * 4);
 
             DrawEnd(engine, 0, 0, firstVertex, vertexCount);
@@ -544,6 +544,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
             _state.State.FirstInstance = (uint)firstInstance;
 
             _drawState.DrawIndexed = indexed;
+            _drawState.DrawUsesEngineState = true;
             _currentSpecState.SetHasConstantBufferDrawParameters(true);
 
             engine.UpdateState();
@@ -676,6 +677,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
 
             _drawState.DrawIndexed = indexed;
             _drawState.DrawIndirect = true;
+            _drawState.DrawUsesEngineState = true;
             _currentSpecState.SetHasConstantBufferDrawParameters(true);
 
             engine.UpdateState();

src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawState.cs

@@ -38,6 +38,11 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
         /// </summary>
         public bool DrawIndirect;
 
+        /// <summary>
+        /// Indicates that the draw is using the draw parameters on the 3D engine state, rather than inline parameters submitted with the draw command.
+        /// </summary>
+        public bool DrawUsesEngineState;
+
         /// <summary>
         /// Indicates if any of the currently used vertex shaders reads the instance ID.
         /// </summary>
@@ -48,11 +53,6 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
         /// </summary>
         public bool IsAnyVbInstanced;
 
-        /// <summary>
-        /// Indicates that the draw is writing the base vertex, base instance and draw index to Constant Buffer 0.
-        /// </summary>
-        public bool HasConstantBufferDrawParameters;
-
         /// <summary>
         /// Primitive topology for the next draw.
         /// </summary>

src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdater.cs

@@ -47,7 +47,8 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
         private uint _vbEnableMask;
 
         private bool _prevDrawIndexed;
-        private readonly bool _prevDrawIndirect;
+        private bool _prevDrawIndirect;
+        private bool _prevDrawUsesEngineState;
         private IndexType _prevIndexType;
         private uint _prevFirstVertex;
         private bool _prevTfEnable;
@@ -236,7 +237,9 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
             // method when doing indexed draws, so we need to make sure
             // to update the vertex buffers if we are doing a regular
             // draw after a indexed one and vice-versa.
-            if (_drawState.DrawIndexed != _prevDrawIndexed)
+            // Some draws also do not update the engine state, so it is possible for it
+            // to not be dirty even if the vertex counts or other state changed. We need to force it to be dirty in this case.
+            if (_drawState.DrawIndexed != _prevDrawIndexed || _drawState.DrawUsesEngineState != _prevDrawUsesEngineState)
             {
                 _updateTracker.ForceDirty(VertexBufferStateIndex);
 
@@ -251,6 +254,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
                 }
 
                 _prevDrawIndexed = _drawState.DrawIndexed;
+                _prevDrawUsesEngineState = _drawState.DrawUsesEngineState;
             }
 
             // Some draw parameters are used to restrict the vertex buffer size,
@@ -260,6 +264,8 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
             if (_drawState.DrawIndirect != _prevDrawIndirect)
             {
                 _updateTracker.ForceDirty(VertexBufferStateIndex);
+
+                _prevDrawIndirect = _drawState.DrawIndirect;
             }
 
             // In some cases, the index type is also used to guess the

src/Ryujinx.Graphics.Vulkan/BufferHolder.cs

@@ -520,7 +520,7 @@ namespace Ryujinx.Graphics.Vulkan
 
                     if (_gd.PipelineInternal.CurrentCommandBuffer.CommandBuffer.Handle == cbs.CommandBuffer.Handle)
                     {
-                        SetData(rangeOffset, _pendingData.AsSpan(rangeOffset, rangeSize), cbs, _gd.PipelineInternal.EndRenderPass, false);
+                        SetData(rangeOffset, _pendingData.AsSpan(rangeOffset, rangeSize), cbs, _gd.PipelineInternal.EndRenderPassDelegate, false);
                     }
                     else
                     {

src/Ryujinx.Graphics.Vulkan/PipelineBase.cs

@@ -30,6 +30,7 @@ namespace Ryujinx.Graphics.Vulkan
         public readonly PipelineCache PipelineCache;
 
         public readonly AutoFlushCounter AutoFlush;
+        public readonly Action EndRenderPassDelegate;
 
         protected PipelineDynamicState DynamicState;
         private PipelineState _newState;
@@ -92,6 +93,7 @@ namespace Ryujinx.Graphics.Vulkan
             Device = device;
 
             AutoFlush = new AutoFlushCounter(gd);
+            EndRenderPassDelegate = EndRenderPass;
 
             var pipelineCacheCreateInfo = new PipelineCacheCreateInfo
             {

src/Ryujinx.Graphics.Vulkan/VulkanRenderer.cs

@@ -784,7 +784,7 @@ namespace Ryujinx.Graphics.Vulkan
 
         public void SetBufferData(BufferHandle buffer, int offset, ReadOnlySpan<byte> data)
         {
-            BufferManager.SetData(buffer, offset, data, _pipeline.CurrentCommandBuffer, _pipeline.EndRenderPass);
+            BufferManager.SetData(buffer, offset, data, _pipeline.CurrentCommandBuffer, _pipeline.EndRenderPassDelegate);
         }
 
         public void UpdateCounters()