Optimize Texture Binding and Shader Specialization Checks (#3399)

* Changes 1

* Changes 2

* Better ModifiedSequence handling

This should handle PreciseEvents properly, and simplifies a few things.

* Minor changes, remove debug log

* Handle stage.Info being null

Hopefully fixes Catherine crash

* Fix shader specialization fast texture lookup

* Fix some things.

* Address Feedback Part 1

* Make method static.

Ryujinx.Graphics.Gpu/Engine/Compute/ComputeClass.cs

@@ -188,6 +188,9 @@ namespace Ryujinx.Graphics.Gpu.Engine.Compute
             _channel.BufferManager.SetComputeStorageBufferBindings(info.SBuffers);
             _channel.BufferManager.SetComputeUniformBufferBindings(info.CBuffers);
 
+            int maxTextureBinding = -1;
+            int maxImageBinding = -1;
+
             TextureBindingInfo[] textureBindings = _channel.TextureManager.RentComputeTextureBindings(info.Textures.Count);
 
             for (int index = 0; index < info.Textures.Count; index++)
@@ -202,6 +205,11 @@ namespace Ryujinx.Graphics.Gpu.Engine.Compute
                     descriptor.CbufSlot,
                     descriptor.HandleIndex,
                     descriptor.Flags);
+
+                if (descriptor.Binding > maxTextureBinding)
+                {
+                    maxTextureBinding = descriptor.Binding;
+                }
             }
 
             TextureBindingInfo[] imageBindings = _channel.TextureManager.RentComputeImageBindings(info.Images.Count);
@@ -220,9 +228,18 @@ namespace Ryujinx.Graphics.Gpu.Engine.Compute
                     descriptor.CbufSlot,
                     descriptor.HandleIndex,
                     descriptor.Flags);
+
+                if (descriptor.Binding > maxImageBinding)
+                {
+                    maxImageBinding = descriptor.Binding;
+                }
             }
 
-            _channel.TextureManager.CommitComputeBindings();
+            _channel.TextureManager.SetComputeMaxBindings(maxTextureBinding, maxImageBinding);
+
+            // Should never return false for mismatching spec state, since the shader was fetched above.
+            _channel.TextureManager.CommitComputeBindings(cs.SpecializationState); 
+            
             _channel.BufferManager.CommitComputeBindings();
 
             _context.Renderer.Pipeline.DispatchCompute(qmd.CtaRasterWidth, qmd.CtaRasterHeight, qmd.CtaRasterDepth);

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

@@ -201,7 +201,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
             // of the shader for the new state.
             if (_shaderSpecState != null)
             {
-                if (!_shaderSpecState.MatchesGraphics(_channel, GetPoolState(), GetGraphicsState()))
+                if (!_shaderSpecState.MatchesGraphics(_channel, GetPoolState(), GetGraphicsState(), false))
                 {
                     ForceShaderUpdate();
                 }
@@ -275,7 +275,12 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
         {
             UpdateStorageBuffers();
 
-            _channel.TextureManager.CommitGraphicsBindings();
+            if (!_channel.TextureManager.CommitGraphicsBindings(_shaderSpecState))
+            {
+                // Shader must be reloaded.
+                UpdateShaderState();
+            }
+
             _channel.BufferManager.CommitGraphicsBindings();
         }
 
@@ -1150,6 +1155,9 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
                 return;
             }
 
+            int maxTextureBinding = -1;
+            int maxImageBinding = -1;
+
             Span<TextureBindingInfo> textureBindings = _channel.TextureManager.RentGraphicsTextureBindings(stage, info.Textures.Count);
 
             if (info.UsesRtLayer)
@@ -1169,6 +1177,11 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
                     descriptor.CbufSlot,
                     descriptor.HandleIndex,
                     descriptor.Flags);
+
+                if (descriptor.Binding > maxTextureBinding)
+                {
+                    maxTextureBinding = descriptor.Binding;
+                }
             }
 
             TextureBindingInfo[] imageBindings = _channel.TextureManager.RentGraphicsImageBindings(stage, info.Images.Count);
@@ -1187,7 +1200,14 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
                     descriptor.CbufSlot,
                     descriptor.HandleIndex,
                     descriptor.Flags);
+
+                if (descriptor.Binding > maxImageBinding)
+                {
+                    maxImageBinding = descriptor.Binding;
                 }
+            }
+
+            _channel.TextureManager.SetGraphicsMaxBindings(maxTextureBinding, maxImageBinding);
 
             _channel.BufferManager.SetGraphicsStorageBufferBindings(stage, info.SBuffers);
             _channel.BufferManager.SetGraphicsUniformBufferBindings(stage, info.CBuffers);

Ryujinx.Graphics.Gpu/Image/Pool.cs

@@ -1,6 +1,7 @@
 using Ryujinx.Cpu.Tracking;
 using Ryujinx.Graphics.Gpu.Memory;
 using System;
+using System.Runtime.InteropServices;
 
 namespace Ryujinx.Graphics.Gpu.Image
 {
@@ -16,6 +17,7 @@ namespace Ryujinx.Graphics.Gpu.Image
         protected GpuContext Context;
         protected PhysicalMemory PhysicalMemory;
         protected int SequenceNumber;
+        protected int ModifiedSequenceNumber;
 
         protected T1[] Items;
         protected T2[] DescriptorCache;
@@ -41,6 +43,9 @@ namespace Ryujinx.Graphics.Gpu.Image
         private readonly CpuMultiRegionHandle _memoryTracking;
         private readonly Action<ulong, ulong> _modifiedDelegate;
 
+        private int _modifiedSequenceOffset;
+        private bool _modified;
+
         /// <summary>
         /// Creates a new instance of the GPU resource pool.
         /// </summary>
@@ -79,6 +84,16 @@ namespace Ryujinx.Graphics.Gpu.Image
             return PhysicalMemory.Read<T2>(Address + (ulong)id * DescriptorSize);
         }
 
+        /// <summary>
+        /// Gets a reference to the descriptor for a given ID.
+        /// </summary>
+        /// <param name="id">ID of the descriptor. This is effectively a zero-based index</param>
+        /// <returns>A reference to the descriptor</returns>
+        public ref readonly T2 GetDescriptorRef(int id)
+        {
+            return ref MemoryMarshal.Cast<byte, T2>(PhysicalMemory.GetSpan(Address + (ulong)id * DescriptorSize, DescriptorSize))[0];
+        }
+
         /// <summary>
         /// Gets the GPU resource with the given ID.
         /// </summary>
@@ -93,7 +108,13 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// </summary>
         public void SynchronizeMemory()
         {
+            _modified = false;
             _memoryTracking.QueryModified(_modifiedDelegate);
+
+            if (_modified)
+            {
+                UpdateModifiedSequence();
+            }
         }
 
         /// <summary>
@@ -103,6 +124,8 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// <param name="mSize">Size of the modified region</param>
         private void RegionModified(ulong mAddress, ulong mSize)
         {
+            _modified = true;
+
             if (mAddress < Address)
             {
                 mAddress = Address;
@@ -118,6 +141,15 @@ namespace Ryujinx.Graphics.Gpu.Image
             InvalidateRangeImpl(mAddress, mSize);
         }
 
+        /// <summary>
+        /// Updates the modified sequence number using the current sequence number and offset,
+        /// indicating that it has been modified.
+        /// </summary>
+        protected void UpdateModifiedSequence()
+        {
+            ModifiedSequenceNumber = SequenceNumber + _modifiedSequenceOffset;
+        }
+
         /// <summary>
         /// An action to be performed when a precise memory access occurs to this resource.
         /// Makes sure that the dirty flags are checked.
@@ -129,6 +161,16 @@ namespace Ryujinx.Graphics.Gpu.Image
         {
             if (write && Context.SequenceNumber == SequenceNumber)
             {
+                if (ModifiedSequenceNumber == SequenceNumber + _modifiedSequenceOffset)
+                {
+                    // The modified sequence number is offset when PreciseActions occur so that
+                    // users checking it will see an increment and know the pool has changed since
+                    // their last look, even though the main SequenceNumber has not been changed.
+
+                    _modifiedSequenceOffset++;
+                }
+
+                // Force the pool to be checked again the next time it is used.
                 SequenceNumber--;
             }
 

Ryujinx.Graphics.Gpu/Image/Sampler.cs

@@ -8,6 +8,11 @@ namespace Ryujinx.Graphics.Gpu.Image
     /// </summary>
     class Sampler : IDisposable
     {
+        /// <summary>
+        /// True if the sampler is disposed, false otherwise.
+        /// </summary>
+        public bool IsDisposed { get; private set; }
+
         /// <summary>
         /// Host sampler object.
         /// </summary>
@@ -101,6 +106,8 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// </summary>
         public void Dispose()
         {
+            IsDisposed = true;
+
             _hostSampler.Dispose();
             _anisoSampler?.Dispose();
         }

Ryujinx.Graphics.Gpu/Image/SamplerPool.cs

@@ -48,6 +48,8 @@ namespace Ryujinx.Graphics.Gpu.Image
                             Items[i] = null;
                         }
                     }
+
+                    UpdateModifiedSequence();
                 }
 
                 SequenceNumber = Context.SequenceNumber;
@@ -71,6 +73,39 @@ namespace Ryujinx.Graphics.Gpu.Image
             return sampler;
         }
 
+        /// <summary>
+        /// Checks if the pool was modified, and returns the last sequence number where a modification was detected.
+        /// </summary>
+        /// <returns>A number that increments each time a modification is detected</returns>
+        public int CheckModified()
+        {
+            if (SequenceNumber != Context.SequenceNumber)
+            {
+                SequenceNumber = Context.SequenceNumber;
+
+                if (_forcedAnisotropy != GraphicsConfig.MaxAnisotropy)
+                {
+                    _forcedAnisotropy = GraphicsConfig.MaxAnisotropy;
+
+                    for (int i = 0; i < Items.Length; i++)
+                    {
+                        if (Items[i] != null)
+                        {
+                            Items[i].Dispose();
+
+                            Items[i] = null;
+                        }
+                    }
+
+                    UpdateModifiedSequence();
+                }
+
+                SynchronizeMemory();
+            }
+
+            return ModifiedSequenceNumber;
+        }
+
         /// <summary>
         /// Implementation of the sampler pool range invalidation.
         /// </summary>

Ryujinx.Graphics.Gpu/Image/Texture.cs

@@ -100,6 +100,11 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// </summary>
         public bool AlwaysFlushOnOverlap { get; private set; }
 
+        /// <summary>
+        /// Increments when the host texture is swapped, or when the texture is removed from all pools.
+        /// </summary>
+        public int InvalidatedSequence { get; private set; }
+
         private int _depth;
         private int _layers;
         public int FirstLayer { get; private set; }
@@ -1407,6 +1412,7 @@ namespace Ryujinx.Graphics.Gpu.Image
             DisposeTextures();
 
             HostTexture = hostTexture;
+            InvalidatedSequence++;
         }
 
         /// <summary>
@@ -1535,6 +1541,8 @@ namespace Ryujinx.Graphics.Gpu.Image
 
                 _poolOwners.Clear();
             }
+
+            InvalidatedSequence++;
         }
 
         /// <summary>

Ryujinx.Graphics.Gpu/Image/TextureBindingsManager.cs

@@ -1,8 +1,12 @@
 using Ryujinx.Common.Logging;
 using Ryujinx.Graphics.GAL;
 using Ryujinx.Graphics.Gpu.Engine.Types;
+using Ryujinx.Graphics.Gpu.Memory;
+using Ryujinx.Graphics.Gpu.Shader;
 using Ryujinx.Graphics.Shader;
 using System;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
 
 namespace Ryujinx.Graphics.Gpu.Image
 {
@@ -35,17 +39,24 @@ namespace Ryujinx.Graphics.Gpu.Image
         {
             public ITexture Texture;
             public ISampler Sampler;
+
+            public int TextureHandle;
+            public int SamplerHandle;
+            public int InvalidatedSequence;
+            public Texture CachedTexture;
+            public Sampler CachedSampler;
         }
 
-        private readonly TextureStatePerStage[][] _textureState;
+        private TextureStatePerStage[] _textureState;
-        private readonly TextureStatePerStage[][] _imageState;
+        private TextureStatePerStage[] _imageState;
 
         private int[] _textureBindingsCount;
         private int[] _imageBindingsCount;
 
-        private int _textureBufferIndex;
+        private int _texturePoolSequence;
+        private int _samplerPoolSequence;
 
-        private bool _rebind;
+        private int _textureBufferIndex;
 
         private readonly float[] _scales;
         private bool _scaleChanged;
@@ -72,8 +83,8 @@ namespace Ryujinx.Graphics.Gpu.Image
             _textureBindings = new TextureBindingInfo[stages][];
             _imageBindings   = new TextureBindingInfo[stages][];
 
-            _textureState = new TextureStatePerStage[stages][];
+            _textureState = new TextureStatePerStage[InitialTextureStateSize];
-            _imageState   = new TextureStatePerStage[stages][];
+            _imageState   = new TextureStatePerStage[InitialImageStateSize];
 
             _textureBindingsCount = new int[stages];
             _imageBindingsCount = new int[stages];
@@ -82,9 +93,6 @@ namespace Ryujinx.Graphics.Gpu.Image
             {
                 _textureBindings[stage] = new TextureBindingInfo[InitialTextureStateSize];
                 _imageBindings[stage] = new TextureBindingInfo[InitialImageStateSize];
-
-                _textureState[stage] = new TextureStatePerStage[InitialTextureStateSize];
-                _imageState[stage] = new TextureStatePerStage[InitialImageStateSize];
             }
         }
 
@@ -99,15 +107,6 @@ namespace Ryujinx.Graphics.Gpu.Image
             if (count > _textureBindings[stage].Length)
             {
                 Array.Resize(ref _textureBindings[stage], count);
-                Array.Resize(ref _textureState[stage], count);
-            }
-
-            int toClear = Math.Max(_textureBindingsCount[stage], count);
-            TextureStatePerStage[] state = _textureState[stage];
-
-            for (int i = 0; i < toClear; i++)
-            {
-                state[i] = new TextureStatePerStage();
             }
 
             _textureBindingsCount[stage] = count;
@@ -126,15 +125,6 @@ namespace Ryujinx.Graphics.Gpu.Image
             if (count > _imageBindings[stage].Length)
             {
                 Array.Resize(ref _imageBindings[stage], count);
-                Array.Resize(ref _imageState[stage], count);
-            }
-
-            int toClear = Math.Max(_imageBindingsCount[stage], count);
-            TextureStatePerStage[] state = _imageState[stage];
-
-            for (int i = 0; i < toClear; i++)
-            {
-                state[i] = new TextureStatePerStage();
             }
 
             _imageBindingsCount[stage] = count;
@@ -142,6 +132,24 @@ namespace Ryujinx.Graphics.Gpu.Image
             return _imageBindings[stage];
         }
 
+        /// <summary>
+        /// Sets the max binding indexes for textures and images.
+        /// </summary>
+        /// <param name="maxTextureBinding">The maximum texture binding</param>
+        /// <param name="maxImageBinding">The maximum image binding</param>
+        public void SetMaxBindings(int maxTextureBinding, int maxImageBinding)
+        {
+            if (maxTextureBinding >= _textureState.Length)
+            {
+                Array.Resize(ref _textureState, maxTextureBinding + 1);
+            }
+
+            if (maxImageBinding >= _imageState.Length)
+            {
+                Array.Resize(ref _imageState, maxImageBinding + 1);
+            }
+        }
+
         /// <summary>
         /// Sets the textures constant buffer index.
         /// The constant buffer specified holds the texture handles.
@@ -323,7 +331,9 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// Ensures that the bindings are visible to the host GPU.
         /// Note: this actually performs the binding using the host graphics API.
         /// </summary>
-        public void CommitBindings()
+        /// <param name="specState">Specialization state for the bound shader</param>
+        /// <returns>True if all bound textures match the current shader specialiation state, false otherwise</returns>
+        public bool CommitBindings(ShaderSpecializationState specState)
         {
             ulong texturePoolAddress = _texturePoolAddress;
 
@@ -331,10 +341,38 @@ namespace Ryujinx.Graphics.Gpu.Image
                 ? _texturePoolCache.FindOrCreate(_channel, texturePoolAddress, _texturePoolMaximumId)
                 : null;
 
+            // Check if the texture pool has been modified since bindings were last committed.
+            // If it wasn't, then it's possible to avoid looking up textures again when the handle remains the same.
+            bool poolModified = false;
+
+            if (texturePool != null)
+            {
+                int texturePoolSequence = texturePool.CheckModified();
+
+                if (_texturePoolSequence != texturePoolSequence)
+                {
+                    poolModified = true;
+                    _texturePoolSequence = texturePoolSequence;
+                }
+            }
+
+            if (_samplerPool != null)
+            {
+                int samplerPoolSequence = _samplerPool.CheckModified();
+
+                if (_samplerPoolSequence != samplerPoolSequence)
+                {
+                    poolModified = true;
+                    _samplerPoolSequence = samplerPoolSequence;
+                }
+            }
+
+            bool specStateMatches = true;
+
             if (_isCompute)
             {
-                CommitTextureBindings(texturePool, ShaderStage.Compute, 0);
+                specStateMatches &= CommitTextureBindings(texturePool, ShaderStage.Compute, 0, poolModified, specState);
-                CommitImageBindings  (texturePool, ShaderStage.Compute, 0);
+                specStateMatches &= CommitImageBindings(texturePool, ShaderStage.Compute, 0, poolModified, specState);
             }
             else
             {
@@ -342,14 +380,57 @@ namespace Ryujinx.Graphics.Gpu.Image
                 {
                     int stageIndex = (int)stage - 1;
 
-                    CommitTextureBindings(texturePool, stage, stageIndex);
+                    specStateMatches &= CommitTextureBindings(texturePool, stage, stageIndex, poolModified, specState);
-                    CommitImageBindings  (texturePool, stage, stageIndex);
+                    specStateMatches &= CommitImageBindings(texturePool, stage, stageIndex, poolModified, specState);
                 }
             }
 
             CommitRenderScale();
 
-            _rebind = false;
+            return specStateMatches;
+        }
+
+        /// <summary>
+        /// Fetch the constant buffers used for a texture to cache.
+        /// </summary>
+        /// <param name="stageIndex">Stage index of the constant buffer</param>
+        /// <param name="cachedTextureBufferIndex">The currently cached texture buffer index</param>
+        /// <param name="cachedSamplerBufferIndex">The currently cached sampler buffer index</param>
+        /// <param name="cachedTextureBuffer">The currently cached texture buffer data</param>
+        /// <param name="cachedSamplerBuffer">The currently cached sampler buffer data</param>
+        /// <param name="textureBufferIndex">The new texture buffer index</param>
+        /// <param name="samplerBufferIndex">The new sampler buffer index</param>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private void UpdateCachedBuffer(
+            int stageIndex,
+            ref int cachedTextureBufferIndex,
+            ref int cachedSamplerBufferIndex,
+            ref ReadOnlySpan<int> cachedTextureBuffer,
+            ref ReadOnlySpan<int> cachedSamplerBuffer,
+            int textureBufferIndex,
+            int samplerBufferIndex)
+        {
+            if (textureBufferIndex != cachedTextureBufferIndex)
+            {
+                ref BufferBounds bounds = ref _channel.BufferManager.GetUniformBufferBounds(_isCompute, stageIndex, textureBufferIndex);
+
+                cachedTextureBuffer = MemoryMarshal.Cast<byte, int>(_channel.MemoryManager.Physical.GetSpan(bounds.Address, (int)bounds.Size));
+                cachedTextureBufferIndex = textureBufferIndex;
+
+                if (samplerBufferIndex == textureBufferIndex)
+                {
+                    cachedSamplerBuffer = cachedTextureBuffer;
+                    cachedSamplerBufferIndex = samplerBufferIndex;
+                }
+            }
+
+            if (samplerBufferIndex != cachedSamplerBufferIndex)
+            {
+                ref BufferBounds bounds = ref _channel.BufferManager.GetUniformBufferBounds(_isCompute, stageIndex, samplerBufferIndex);
+
+                cachedSamplerBuffer = MemoryMarshal.Cast<byte, int>(_channel.MemoryManager.Physical.GetSpan(bounds.Address, (int)bounds.Size));
+                cachedSamplerBufferIndex = samplerBufferIndex;
+            }
         }
 
         /// <summary>
@@ -358,13 +439,16 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// </summary>
         /// <param name="pool">The current texture pool</param>
         /// <param name="stage">The shader stage using the textures to be bound</param>
-        /// <param name="stageIndex">The stage number of the specified shader stage</param>
+        /// <param name="stageIndex">The stage number of the specified shader stage</param
-        private void CommitTextureBindings(TexturePool pool, ShaderStage stage, int stageIndex)
+        /// <param name="poolModified">True if either the texture or sampler pool was modified, false otherwise</param>
+        /// <param name="specState">Specialization state for the bound shader</param>
+        /// <returns>True if all bound textures match the current shader specialiation state, false otherwise</returns>
+        private bool CommitTextureBindings(TexturePool pool, ShaderStage stage, int stageIndex, bool poolModified, ShaderSpecializationState specState)
         {
             int textureCount = _textureBindingsCount[stageIndex];
             if (textureCount == 0)
             {
-                return;
+                return true;
             }
 
             var samplerPool = _samplerPool;
@@ -372,17 +456,26 @@ namespace Ryujinx.Graphics.Gpu.Image
             if (pool == null)
             {
                 Logger.Error?.Print(LogClass.Gpu, $"Shader stage \"{stage}\" uses textures, but texture pool was not set.");
-                return;
+                return true;
             }
 
+            bool specStateMatches = true;
+
+            int cachedTextureBufferIndex = -1;
+            int cachedSamplerBufferIndex = -1;
+            ReadOnlySpan<int> cachedTextureBuffer = Span<int>.Empty;
+            ReadOnlySpan<int> cachedSamplerBuffer = Span<int>.Empty;
+
             for (int index = 0; index < textureCount; index++)
             {
                 TextureBindingInfo bindingInfo = _textureBindings[stageIndex][index];
 
                 (int textureBufferIndex, int samplerBufferIndex) = TextureHandle.UnpackSlots(bindingInfo.CbufSlot, _textureBufferIndex);
 
-                int packedId = ReadPackedId(stageIndex, bindingInfo.Handle, textureBufferIndex, samplerBufferIndex);
+                UpdateCachedBuffer(stageIndex, ref cachedTextureBufferIndex, ref cachedSamplerBufferIndex, ref cachedTextureBuffer, ref cachedSamplerBuffer, textureBufferIndex, samplerBufferIndex);
-                int textureId = UnpackTextureId(packedId);
+
+                int packedId = TextureHandle.ReadPackedId(bindingInfo.Handle, cachedTextureBuffer, cachedSamplerBuffer);
+                int textureId = TextureHandle.UnpackTextureId(packedId);
                 int samplerId;
 
                 if (_samplerIndex == SamplerIndex.ViaHeaderIndex)
@@ -391,10 +484,30 @@ namespace Ryujinx.Graphics.Gpu.Image
                 }
                 else
                 {
-                    samplerId = UnpackSamplerId(packedId);
+                    samplerId = TextureHandle.UnpackSamplerId(packedId);
                 }
 
-                Texture texture = pool.Get(textureId);
+                ref TextureStatePerStage state = ref _textureState[bindingInfo.Binding];
+
+                if (!poolModified &&
+                    state.TextureHandle == textureId &&
+                    state.SamplerHandle == samplerId &&
+                    state.CachedTexture != null &&
+                    state.CachedTexture.InvalidatedSequence == state.InvalidatedSequence &&
+                    state.CachedSampler?.IsDisposed != true)
+                {
+                    // The texture is already bound.
+                    state.CachedTexture.SynchronizeMemory();
+
+                    continue;
+                }
+
+                state.TextureHandle = textureId;
+                state.SamplerHandle = samplerId;
+
+                ref readonly TextureDescriptor descriptor = ref pool.GetForBinding(textureId, out Texture texture);
+
+                specStateMatches &= specState.MatchesTexture(stage, index, descriptor);
 
                 ITexture hostTexture = texture?.GetTargetTexture(bindingInfo.Target);
 
@@ -407,30 +520,36 @@ namespace Ryujinx.Graphics.Gpu.Image
                 }
                 else
                 {
-                    if (_textureState[stageIndex][index].Texture != hostTexture || _rebind)
+                    if (state.Texture != hostTexture)
                     {
                         if (UpdateScale(texture, bindingInfo, index, stage))
                         {
                             hostTexture = texture?.GetTargetTexture(bindingInfo.Target);
                         }
 
-                        _textureState[stageIndex][index].Texture = hostTexture;
+                        state.Texture = hostTexture;
 
                         _context.Renderer.Pipeline.SetTexture(bindingInfo.Binding, hostTexture);
                     }
 
                     Sampler sampler = samplerPool?.Get(samplerId);
+                    state.CachedSampler = sampler;
 
                     ISampler hostSampler = sampler?.GetHostSampler(texture);
 
-                    if (_textureState[stageIndex][index].Sampler != hostSampler || _rebind)
+                    if (state.Sampler != hostSampler)
                     {
-                        _textureState[stageIndex][index].Sampler = hostSampler;
+                        state.Sampler = hostSampler;
 
                         _context.Renderer.Pipeline.SetSampler(bindingInfo.Binding, hostSampler);
                     }
+
+                    state.CachedTexture = texture;
+                    state.InvalidatedSequence = texture?.InvalidatedSequence ?? 0;
                 }
             }
+
+            return specStateMatches;
         }
 
         /// <summary>
@@ -440,38 +559,72 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// <param name="pool">The current texture pool</param>
         /// <param name="stage">The shader stage using the textures to be bound</param>
         /// <param name="stageIndex">The stage number of the specified shader stage</param>
-        private void CommitImageBindings(TexturePool pool, ShaderStage stage, int stageIndex)
+        /// <param name="poolModified">True if either the texture or sampler pool was modified, false otherwise</param>
+        /// <param name="specState">Specialization state for the bound shader</param>
+        /// <returns>True if all bound images match the current shader specialiation state, false otherwise</returns>
+        private bool CommitImageBindings(TexturePool pool, ShaderStage stage, int stageIndex, bool poolModified, ShaderSpecializationState specState)
         {
             int imageCount = _imageBindingsCount[stageIndex];
             if (imageCount == 0)
             {
-                return;
+                return true;
             }
 
             if (pool == null)
             {
                 Logger.Error?.Print(LogClass.Gpu, $"Shader stage \"{stage}\" uses images, but texture pool was not set.");
-                return;
+                return true;
             }
 
             // Scales for images appear after the texture ones.
             int baseScaleIndex = _textureBindingsCount[stageIndex];
 
+            int cachedTextureBufferIndex = -1;
+            int cachedSamplerBufferIndex = -1;
+            ReadOnlySpan<int> cachedTextureBuffer = Span<int>.Empty;
+            ReadOnlySpan<int> cachedSamplerBuffer = Span<int>.Empty;
+
+            bool specStateMatches = true;
+
             for (int index = 0; index < imageCount; index++)
             {
                 TextureBindingInfo bindingInfo = _imageBindings[stageIndex][index];
 
                 (int textureBufferIndex, int samplerBufferIndex) = TextureHandle.UnpackSlots(bindingInfo.CbufSlot, _textureBufferIndex);
 
-                int packedId = ReadPackedId(stageIndex, bindingInfo.Handle, textureBufferIndex, samplerBufferIndex);
+                UpdateCachedBuffer(stageIndex, ref cachedTextureBufferIndex, ref cachedSamplerBufferIndex, ref cachedTextureBuffer, ref cachedSamplerBuffer, textureBufferIndex, samplerBufferIndex);
-                int textureId = UnpackTextureId(packedId);
 
-                Texture texture = pool.Get(textureId);
+                int packedId = TextureHandle.ReadPackedId(bindingInfo.Handle, cachedTextureBuffer, cachedSamplerBuffer);
+                int textureId = TextureHandle.UnpackTextureId(packedId);
 
-                ITexture hostTexture = texture?.GetTargetTexture(bindingInfo.Target);
+                ref TextureStatePerStage state = ref _imageState[bindingInfo.Binding];
 
                 bool isStore = bindingInfo.Flags.HasFlag(TextureUsageFlags.ImageStore);
 
+                if (!poolModified &&
+                    state.TextureHandle == textureId &&
+                    state.CachedTexture != null &&
+                    state.CachedTexture.InvalidatedSequence == state.InvalidatedSequence)
+                {
+                    // The texture is already bound.
+                    state.CachedTexture.SynchronizeMemory();
+
+                    if (isStore)
+                    {
+                        state.CachedTexture?.SignalModified();
+                    }
+
+                    continue;
+                }
+
+                state.TextureHandle = textureId;
+
+                ref readonly TextureDescriptor descriptor = ref pool.GetForBinding(textureId, out Texture texture);
+
+                specStateMatches &= specState.MatchesImage(stage, index, descriptor);
+
+                ITexture hostTexture = texture?.GetTargetTexture(bindingInfo.Target);
+
                 if (hostTexture != null && texture.Target == Target.TextureBuffer)
                 {
                     // Ensure that the buffer texture is using the correct buffer as storage.
@@ -494,14 +647,14 @@ namespace Ryujinx.Graphics.Gpu.Image
                         texture?.SignalModified();
                     }
 
-                    if (_imageState[stageIndex][index].Texture != hostTexture || _rebind)
+                    if (state.Texture != hostTexture)
                     {
                         if (UpdateScale(texture, bindingInfo, baseScaleIndex + index, stage))
                         {
                             hostTexture = texture?.GetTargetTexture(bindingInfo.Target);
                         }
 
-                        _imageState[stageIndex][index].Texture = hostTexture;
+                        state.Texture = hostTexture;
 
                         Format format = bindingInfo.Format;
 
@@ -512,8 +665,13 @@ namespace Ryujinx.Graphics.Gpu.Image
 
                         _context.Renderer.Pipeline.SetImage(bindingInfo.Binding, hostTexture, format);
                     }
+
+                    state.CachedTexture = texture;
+                    state.InvalidatedSequence = texture?.InvalidatedSequence ?? 0;
                 }
             }
+
+            return specStateMatches;
         }
 
         /// <summary>
@@ -537,7 +695,7 @@ namespace Ryujinx.Graphics.Gpu.Image
             (int textureBufferIndex, int samplerBufferIndex) = TextureHandle.UnpackSlots(cbufSlot, bufferIndex);
 
             int packedId = ReadPackedId(stageIndex, handle, textureBufferIndex, samplerBufferIndex);
-            int textureId = UnpackTextureId(packedId);
+            int textureId = TextureHandle.UnpackTextureId(packedId);
 
             ulong poolAddress = _channel.MemoryManager.Translate(poolGpuVa);
 
@@ -555,6 +713,7 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// <param name="textureBufferIndex">Index of the constant buffer holding the texture handles</param>
         /// <param name="samplerBufferIndex">Index of the constant buffer holding the sampler handles</param>
         /// <returns>The packed texture and sampler ID (the real texture handle)</returns>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         private int ReadPackedId(int stageIndex, int wordOffset, int textureBufferIndex, int samplerBufferIndex)
         {
             (int textureWordOffset, int samplerWordOffset, TextureHandleType handleType) = TextureHandle.UnpackOffsets(wordOffset);
@@ -590,32 +749,13 @@ namespace Ryujinx.Graphics.Gpu.Image
             return handle;
         }
 
-        /// <summary>
-        /// Unpacks the texture ID from the real texture handle.
-        /// </summary>
-        /// <param name="packedId">The real texture handle</param>
-        /// <returns>The texture ID</returns>
-        private static int UnpackTextureId(int packedId)
-        {
-            return (packedId >> 0) & 0xfffff;
-        }
-
-        /// <summary>
-        /// Unpacks the sampler ID from the real texture handle.
-        /// </summary>
-        /// <param name="packedId">The real texture handle</param>
-        /// <returns>The sampler ID</returns>
-        private static int UnpackSamplerId(int packedId)
-        {
-            return (packedId >> 20) & 0xfff;
-        }
-
         /// <summary>
         /// Force all bound textures and images to be rebound the next time CommitBindings is called.
         /// </summary>
         public void Rebind()
         {
-            _rebind = true;
+            Array.Clear(_textureState);
+            Array.Clear(_imageState);
         }
 
         /// <summary>

Ryujinx.Graphics.Gpu/Image/TextureManager.cs

@@ -1,5 +1,6 @@
 using Ryujinx.Graphics.GAL;
 using Ryujinx.Graphics.Gpu.Engine.Types;
+using Ryujinx.Graphics.Gpu.Shader;
 using System;
 
 namespace Ryujinx.Graphics.Gpu.Image
@@ -10,9 +11,11 @@ namespace Ryujinx.Graphics.Gpu.Image
     class TextureManager : IDisposable
     {
         private readonly GpuContext _context;
+        private readonly GpuChannel _channel;
 
         private readonly TextureBindingsManager _cpBindingsManager;
         private readonly TextureBindingsManager _gpBindingsManager;
+        private readonly TexturePoolCache _texturePoolCache;
 
         private readonly Texture[] _rtColors;
         private readonly ITexture[] _rtHostColors;
@@ -35,6 +38,7 @@ namespace Ryujinx.Graphics.Gpu.Image
         public TextureManager(GpuContext context, GpuChannel channel)
         {
             _context = context;
+            _channel = channel;
 
             TexturePoolCache texturePoolCache = new TexturePoolCache(context);
 
@@ -43,6 +47,7 @@ namespace Ryujinx.Graphics.Gpu.Image
 
             _cpBindingsManager = new TextureBindingsManager(context, channel, texturePoolCache, scales, isCompute: true);
             _gpBindingsManager = new TextureBindingsManager(context, channel, texturePoolCache, scales, isCompute: false);
+            _texturePoolCache = texturePoolCache;
 
             _rtColors = new Texture[Constants.TotalRenderTargets];
             _rtHostColors = new ITexture[Constants.TotalRenderTargets];
@@ -99,6 +104,16 @@ namespace Ryujinx.Graphics.Gpu.Image
             _cpBindingsManager.SetTextureBufferIndex(index);
         }
 
+        /// <summary>
+        /// Sets the max binding indexes on the compute pipeline.
+        /// </summary>
+        /// <param name="maxTextureBinding">The maximum texture binding</param>
+        /// <param name="maxImageBinding">The maximum image binding</param>
+        public void SetComputeMaxBindings(int maxTextureBinding, int maxImageBinding)
+        {
+            _cpBindingsManager.SetMaxBindings(maxTextureBinding, maxImageBinding);
+        }
+
         /// <summary>
         /// Sets the texture constant buffer index on the graphics pipeline.
         /// </summary>
@@ -108,6 +123,16 @@ namespace Ryujinx.Graphics.Gpu.Image
             _gpBindingsManager.SetTextureBufferIndex(index);
         }
 
+        /// <summary>
+        /// Sets the max binding indexes on the graphics pipeline.
+        /// </summary>
+        /// <param name="maxTextureBinding">The maximum texture binding</param>
+        /// <param name="maxImageBinding">The maximum image binding</param>
+        public void SetGraphicsMaxBindings(int maxTextureBinding, int maxImageBinding)
+        {
+            _gpBindingsManager.SetMaxBindings(maxTextureBinding, maxImageBinding);
+        }
+
         /// <summary>
         /// Sets the current sampler pool on the compute pipeline.
         /// </summary>
@@ -335,25 +360,48 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// <summary>
         /// Commits bindings on the compute pipeline.
         /// </summary>
-        public void CommitComputeBindings()
+        /// <param name="specState">Specialization state for the bound shader</param>
+        /// <returns>True if all bound textures match the current shader specialization state, false otherwise</returns>
+        public bool CommitComputeBindings(ShaderSpecializationState specState)
         {
             // Every time we switch between graphics and compute work,
             // we must rebind everything.
             // Since compute work happens less often, we always do that
             // before and after the compute dispatch.
             _cpBindingsManager.Rebind();
-            _cpBindingsManager.CommitBindings();
+            bool result = _cpBindingsManager.CommitBindings(specState);
             _gpBindingsManager.Rebind();
+
+            return result;
         }
 
         /// <summary>
         /// Commits bindings on the graphics pipeline.
         /// </summary>
-        public void CommitGraphicsBindings()
+        /// <param name="specState">Specialization state for the bound shader</param>
+        /// <returns>True if all bound textures match the current shader specialization state, false otherwise</returns>
+        public bool CommitGraphicsBindings(ShaderSpecializationState specState)
         {
-            _gpBindingsManager.CommitBindings();
+            bool result = _gpBindingsManager.CommitBindings(specState);
 
             UpdateRenderTargets();
+
+            return result;
+        }
+
+        /// <summary>
+        /// Returns a texture pool from the cache, with the given address and maximum id.
+        /// </summary>
+        /// <param name="poolGpuVa">GPU virtual address of the texture pool</param>
+        /// <param name="maximumId">Maximum ID of the texture pool</param>
+        /// <returns>The texture pool</returns>
+        public TexturePool GetTexturePool(ulong poolGpuVa, int maximumId)
+        {
+            ulong poolAddress = _channel.MemoryManager.Translate(poolGpuVa);
+
+            TexturePool texturePool = _texturePoolCache.FindOrCreate(_channel, poolAddress, maximumId);
+
+            return texturePool;
         }
 
         /// <summary>

Ryujinx.Graphics.Gpu/Image/TexturePool.cs

@@ -14,6 +14,7 @@ namespace Ryujinx.Graphics.Gpu.Image
     {
         private readonly GpuChannel _channel;
         private readonly ConcurrentQueue<Texture> _dereferenceQueue = new ConcurrentQueue<Texture>();
+        private TextureDescriptor _defaultDescriptor;
 
         /// <summary>
         /// Intrusive linked list node used on the texture pool cache.
@@ -32,6 +33,62 @@ namespace Ryujinx.Graphics.Gpu.Image
             _channel = channel;
         }
 
+        /// <summary>
+        /// Gets the texture descripor and texture with the given ID with no bounds check or synchronization.
+        /// </summary>
+        /// <param name="id">ID of the texture. This is effectively a zero-based index</param>
+        /// <param name="texture">The texture with the given ID</param>
+        /// <returns>The texture descriptor with the given ID</returns>
+        private ref readonly TextureDescriptor GetInternal(int id, out Texture texture)
+        {
+            texture = Items[id];
+
+            ref readonly TextureDescriptor descriptor = ref GetDescriptorRef(id);
+
+            if (texture == null)
+            {
+                TextureInfo info = GetInfo(descriptor, out int layerSize);
+
+                ProcessDereferenceQueue();
+
+                texture = PhysicalMemory.TextureCache.FindOrCreateTexture(_channel.MemoryManager, TextureSearchFlags.ForSampler, info, layerSize);
+
+                // If this happens, then the texture address is invalid, we can't add it to the cache.
+                if (texture == null)
+                {
+                    return ref descriptor;
+                }
+
+                texture.IncrementReferenceCount(this, id);
+
+                Items[id] = texture;
+
+                DescriptorCache[id] = descriptor;
+            }
+            else
+            {
+                if (texture.ChangedSize)
+                {
+                    // Texture changed size at one point - it may be a different size than the sampler expects.
+                    // This can be triggered when the size is changed by a size hint on copy or draw, but the texture has been sampled before.
+
+                    int baseLevel = descriptor.UnpackBaseLevel();
+                    int width = Math.Max(1, descriptor.UnpackWidth() >> baseLevel);
+                    int height = Math.Max(1, descriptor.UnpackHeight() >> baseLevel);
+
+                    if (texture.Info.Width != width || texture.Info.Height != height)
+                    {
+                        texture.ChangeSize(width, height, texture.Info.DepthOrLayers);
+                    }
+                }
+
+                // Memory is automatically synchronized on texture creation.
+                texture.SynchronizeMemory();
+            }
+
+            return ref descriptor;
+        }
+
         /// <summary>
         /// Gets the texture with the given ID.
         /// </summary>
@@ -51,56 +108,49 @@ namespace Ryujinx.Graphics.Gpu.Image
                 SynchronizeMemory();
             }
 
-            Texture texture = Items[id];
+            GetInternal(id, out Texture texture);
-
-            if (texture == null)
-            {
-                TextureDescriptor descriptor = GetDescriptor(id);
-
-                TextureInfo info = GetInfo(descriptor, out int layerSize);
-
-                ProcessDereferenceQueue();
-
-                texture = PhysicalMemory.TextureCache.FindOrCreateTexture(_channel.MemoryManager, TextureSearchFlags.ForSampler, info, layerSize);
-
-                // If this happens, then the texture address is invalid, we can't add it to the cache.
-                if (texture == null)
-                {
-                    return null;
-                }
-
-                texture.IncrementReferenceCount(this, id);
-
-                Items[id] = texture;
-
-                DescriptorCache[id] = descriptor;
-            }
-            else
-            {
-                if (texture.ChangedSize)
-                {
-                    // Texture changed size at one point - it may be a different size than the sampler expects.
-                    // This can be triggered when the size is changed by a size hint on copy or draw, but the texture has been sampled before.
-
-                    TextureDescriptor descriptor = GetDescriptor(id);
-
-                    int baseLevel = descriptor.UnpackBaseLevel();
-                    int width = Math.Max(1, descriptor.UnpackWidth() >> baseLevel);
-                    int height = Math.Max(1, descriptor.UnpackHeight() >> baseLevel);
-
-                    if (texture.Info.Width != width || texture.Info.Height != height)
-                    {
-                        texture.ChangeSize(width, height, texture.Info.DepthOrLayers);
-                    }
-                }
-
-                // Memory is automatically synchronized on texture creation.
-                texture.SynchronizeMemory();
-            }
 
             return texture;
         }
 
+        /// <summary>
+        /// Gets the texture descriptor and texture with the given ID.
+        /// </summary>
+        /// <remarks>
+        /// This method assumes that the pool has been manually synchronized before doing binding.
+        /// </remarks>
+        /// <param name="id">ID of the texture. This is effectively a zero-based index</param>
+        /// <param name="texture">The texture with the given ID</param>
+        /// <returns>The texture descriptor with the given ID</returns>
+        public ref readonly TextureDescriptor GetForBinding(int id, out Texture texture)
+        {
+            if ((uint)id >= Items.Length)
+            {
+                texture = null;
+                return ref _defaultDescriptor;
+            }
+
+            // When getting for binding, assume the pool has already been synchronized.
+
+            return ref GetInternal(id, out texture);
+        }
+
+        /// <summary>
+        /// Checks if the pool was modified, and returns the last sequence number where a modification was detected.
+        /// </summary>
+        /// <returns>A number that increments each time a modification is detected</returns>
+        public int CheckModified()
+        {
+            if (SequenceNumber != Context.SequenceNumber)
+            {
+                SequenceNumber = Context.SequenceNumber;
+
+                SynchronizeMemory();
+            }
+
+            return ModifiedSequenceNumber;
+        }
+
         /// <summary>
         /// Forcibly remove a texture from this pool's items.
         /// If deferred, the dereference will be queued to occur on the render thread.
@@ -175,7 +225,7 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// <param name="descriptor">The texture descriptor</param>
         /// <param name="layerSize">Layer size for textures using a sub-range of mipmap levels, otherwise 0</param>
         /// <returns>The texture information</returns>
-        private TextureInfo GetInfo(TextureDescriptor descriptor, out int layerSize)
+        private TextureInfo GetInfo(in TextureDescriptor descriptor, out int layerSize)
         {
             int depthOrLayers = descriptor.UnpackDepth();
             int levels        = descriptor.UnpackLevels();

Ryujinx.Graphics.Gpu/Memory/BufferManager.cs

@@ -378,6 +378,25 @@ namespace Ryujinx.Graphics.Gpu.Memory
             return _gpUniformBuffers[stage].Buffers[index].Address;
         }
 
+        /// <summary>
+        /// Gets the bounds of the uniform buffer currently bound at the given index.
+        /// </summary>
+        /// <param name="isCompute">Indicates whenever the uniform is requested by the 3D or compute engine</param>
+        /// <param name="stage">Index of the shader stage, if the uniform is for the 3D engine</param>
+        /// <param name="index">Index of the uniform buffer binding</param>
+        /// <returns>The uniform buffer bounds, or an undefined value if the buffer is not currently bound</returns>
+        public ref BufferBounds GetUniformBufferBounds(bool isCompute, int stage, int index)
+        {
+            if (isCompute)
+            {
+                return ref _cpUniformBuffers.Buffers[index];
+            }
+            else
+            {
+                return ref _gpUniformBuffers[stage].Buffers[index];
+            }
+        }
+
         /// <summary>
         /// Ensures that the compute engine bindings are visible to the host GPU.
         /// Note: this actually performs the binding using the host graphics API.

Ryujinx.Graphics.Gpu/Shader/CachedShaderProgram.cs

@@ -35,6 +35,8 @@ namespace Ryujinx.Graphics.Gpu.Shader
             HostProgram = hostProgram;
             SpecializationState = specializationState;
             Shaders = shaders;
+
+            SpecializationState.Prepare(shaders);
         }
 
         /// <summary>

Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs

@@ -418,7 +418,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
         {
             if (IsShaderEqual(channel.MemoryManager, cpShader.Shaders[0], gpuVa))
             {
-                return cpShader.SpecializationState.MatchesCompute(channel, poolState);
+                return cpShader.SpecializationState.MatchesCompute(channel, poolState, true);
             }
 
             return false;
@@ -454,7 +454,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
                 }
             }
 
-            return gpShaders.SpecializationState.MatchesGraphics(channel, poolState, graphicsState);
+            return gpShaders.SpecializationState.MatchesGraphics(channel, poolState, graphicsState, true);
         }
 
         /// <summary>

Ryujinx.Graphics.Gpu/Shader/ShaderSpecializationList.cs

@@ -35,7 +35,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
         {
             foreach (var entry in _entries)
             {
-                if (entry.SpecializationState.MatchesGraphics(channel, poolState, graphicsState))
+                if (entry.SpecializationState.MatchesGraphics(channel, poolState, graphicsState, true))
                 {
                     program = entry;
                     return true;
@@ -57,7 +57,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
         {
             foreach (var entry in _entries)
             {
-                if (entry.SpecializationState.MatchesCompute(channel, poolState))
+                if (entry.SpecializationState.MatchesCompute(channel, poolState, true))
                 {
                     program = entry;
                     return true;

Ryujinx.Graphics.Gpu/Shader/ShaderSpecializationState.cs

@@ -1,9 +1,14 @@
 using Ryujinx.Common.Memory;
+using Ryujinx.Graphics.Gpu.Image;
+using Ryujinx.Graphics.Gpu.Memory;
 using Ryujinx.Graphics.Gpu.Shader.DiskCache;
 using Ryujinx.Graphics.Shader;
 using System;
 using System.Collections.Generic;
+using System.Linq;
 using System.Numerics;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
 
 namespace Ryujinx.Graphics.Gpu.Shader
 {
@@ -158,6 +163,9 @@ namespace Ryujinx.Graphics.Gpu.Shader
         }
 
         private readonly Dictionary<TextureKey, Box<TextureSpecializationState>> _textureSpecialization;
+        private KeyValuePair<TextureKey, Box<TextureSpecializationState>>[] _allTextures;
+        private Box<TextureSpecializationState>[][] _textureByBinding;
+        private Box<TextureSpecializationState>[][] _imageByBinding;
 
         /// <summary>
         /// Creates a new instance of the shader specialization state.
@@ -194,6 +202,48 @@ namespace Ryujinx.Graphics.Gpu.Shader
             }
         }
 
+        /// <summary>
+        /// Prepare the shader specialization state for quick binding lookups.
+        /// </summary>
+        /// <param name="stages">The shader stages</param>
+        public void Prepare(CachedShaderStage[] stages)
+        {
+            _allTextures = _textureSpecialization.ToArray();
+
+            _textureByBinding = new Box<TextureSpecializationState>[stages.Length][];
+            _imageByBinding = new Box<TextureSpecializationState>[stages.Length][];
+
+            for (int i = 0; i < stages.Length; i++)
+            {
+                CachedShaderStage stage = stages[i];
+                if (stage?.Info != null)
+                {
+                    var textures = stage.Info.Textures;
+                    var images = stage.Info.Images;
+
+                    var texBindings = new Box<TextureSpecializationState>[textures.Count];
+                    var imageBindings = new Box<TextureSpecializationState>[images.Count];
+
+                    int stageIndex = Math.Max(i - 1, 0); // Don't count VertexA for looking up spec state. No-Op for compute.
+
+                    for (int j = 0; j < textures.Count; j++)
+                    {
+                        var texture = textures[j];
+                        texBindings[j] = GetTextureSpecState(stageIndex, texture.HandleIndex, texture.CbufSlot);
+                    }
+
+                    for (int j = 0; j < images.Count; j++)
+                    {
+                        var image = images[j];
+                        imageBindings[j] = GetTextureSpecState(stageIndex, image.HandleIndex, image.CbufSlot);
+                    }
+
+                    _textureByBinding[i] = texBindings;
+                    _imageByBinding[i] = imageBindings;
+                }
+            }
+        }
+
         /// <summary>
         /// Indicates that the shader accesses the early Z force state.
         /// </summary>
@@ -396,15 +446,16 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// <param name="channel">GPU channel</param>
         /// <param name="poolState">Texture pool state</param>
         /// <param name="graphicsState">Graphics state</param>
+        /// <param name="checkTextures">Indicates whether texture descriptors should be checked</param>
         /// <returns>True if the state matches, false otherwise</returns>
-        public bool MatchesGraphics(GpuChannel channel, GpuChannelPoolState poolState, GpuChannelGraphicsState graphicsState)
+        public bool MatchesGraphics(GpuChannel channel, GpuChannelPoolState poolState, GpuChannelGraphicsState graphicsState, bool checkTextures)
         {
             if (graphicsState.ViewportTransformDisable != GraphicsState.ViewportTransformDisable)
             {
                 return false;
             }
 
-            return Matches(channel, poolState, isCompute: false);
+            return Matches(channel, poolState, checkTextures, isCompute: false);
         }
 
         /// <summary>
@@ -412,10 +463,64 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// </summary>
         /// <param name="channel">GPU channel</param>
         /// <param name="poolState">Texture pool state</param>
+        /// <param name="checkTextures">Indicates whether texture descriptors should be checked</param>
         /// <returns>True if the state matches, false otherwise</returns>
-        public bool MatchesCompute(GpuChannel channel, GpuChannelPoolState poolState)
+        public bool MatchesCompute(GpuChannel channel, GpuChannelPoolState poolState, bool checkTextures)
         {
-            return Matches(channel, poolState, isCompute: true);
+            return Matches(channel, poolState, checkTextures, isCompute: true);
+        }
+
+        /// <summary>
+        /// Fetch the constant buffers used for a texture to cache.
+        /// </summary>
+        /// <param name="channel">GPU channel</param>
+        /// <param name="isCompute">Indicates whenever the check is requested by the 3D or compute engine</param>
+        /// <param name="cachedTextureBufferIndex">The currently cached texture buffer index</param>
+        /// <param name="cachedSamplerBufferIndex">The currently cached sampler buffer index</param>
+        /// <param name="cachedTextureBuffer">The currently cached texture buffer data</param>
+        /// <param name="cachedSamplerBuffer">The currently cached sampler buffer data</param>
+        /// <param name="cachedStageIndex">The currently cached stage</param>
+        /// <param name="textureBufferIndex">The new texture buffer index</param>
+        /// <param name="samplerBufferIndex">The new sampler buffer index</param>
+        /// <param name="stageIndex">Stage index of the constant buffer</param>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private static void UpdateCachedBuffer(
+            GpuChannel channel,
+            bool isCompute,
+            ref int cachedTextureBufferIndex,
+            ref int cachedSamplerBufferIndex,
+            ref ReadOnlySpan<int> cachedTextureBuffer,
+            ref ReadOnlySpan<int> cachedSamplerBuffer,
+            ref int cachedStageIndex,
+            int textureBufferIndex,
+            int samplerBufferIndex,
+            int stageIndex)
+        {
+            bool stageChange = stageIndex != cachedStageIndex;
+
+            if (stageChange || textureBufferIndex != cachedTextureBufferIndex)
+            {
+                ref BufferBounds bounds = ref channel.BufferManager.GetUniformBufferBounds(isCompute, stageIndex, textureBufferIndex);
+
+                cachedTextureBuffer = MemoryMarshal.Cast<byte, int>(channel.MemoryManager.Physical.GetSpan(bounds.Address, (int)bounds.Size));
+                cachedTextureBufferIndex = textureBufferIndex;
+
+                if (samplerBufferIndex == textureBufferIndex)
+                {
+                    cachedSamplerBuffer = cachedTextureBuffer;
+                    cachedSamplerBufferIndex = samplerBufferIndex;
+                }
+            }
+
+            if (stageChange || samplerBufferIndex != cachedSamplerBufferIndex)
+            {
+                ref BufferBounds bounds = ref channel.BufferManager.GetUniformBufferBounds(isCompute, stageIndex, samplerBufferIndex);
+
+                cachedSamplerBuffer = MemoryMarshal.Cast<byte, int>(channel.MemoryManager.Physical.GetSpan(bounds.Address, (int)bounds.Size));
+                cachedSamplerBufferIndex = samplerBufferIndex;
+            }
+
+            cachedStageIndex = stageIndex;
         }
 
         /// <summary>
@@ -423,9 +528,10 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// </summary>
         /// <param name="channel">GPU channel</param>
         /// <param name="poolState">Texture pool state</param>
+        /// <param name="checkTextures">Indicates whether texture descriptors should be checked</param>
         /// <param name="isCompute">Indicates whenever the check is requested by the 3D or compute engine</param>
         /// <returns>True if the state matches, false otherwise</returns>
-        private bool Matches(GpuChannel channel, GpuChannelPoolState poolState, bool isCompute)
+        private bool Matches(GpuChannel channel, GpuChannelPoolState poolState, bool checkTextures, bool isCompute)
         {
             int constantBufferUsePerStageMask = _constantBufferUsePerStage;
 
@@ -445,55 +551,60 @@ namespace Ryujinx.Graphics.Gpu.Shader
                 constantBufferUsePerStageMask &= ~(1 << index);
             }
 
-            foreach (var kv in _textureSpecialization)
+            if (checkTextures)
+            {
+                TexturePool pool = channel.TextureManager.GetTexturePool(poolState.TexturePoolGpuVa, poolState.TexturePoolMaximumId);
+
+                int cachedTextureBufferIndex = -1;
+                int cachedSamplerBufferIndex = -1;
+                int cachedStageIndex = -1;
+                ReadOnlySpan<int> cachedTextureBuffer = Span<int>.Empty;
+                ReadOnlySpan<int> cachedSamplerBuffer = Span<int>.Empty;
+
+                foreach (var kv in _allTextures)
                 {
                     TextureKey textureKey = kv.Key;
 
                     (int textureBufferIndex, int samplerBufferIndex) = TextureHandle.UnpackSlots(textureKey.CbufSlot, poolState.TextureBufferIndex);
 
-                ulong textureCbAddress;
+                    UpdateCachedBuffer(channel,
-                ulong samplerCbAddress;
+                        isCompute,
+                        ref cachedTextureBufferIndex,
+                        ref cachedSamplerBufferIndex,
+                        ref cachedTextureBuffer,
+                        ref cachedSamplerBuffer,
+                        ref cachedStageIndex,
+                        textureBufferIndex,
+                        samplerBufferIndex,
+                        textureKey.StageIndex);
 
-                if (isCompute)
+                    int packedId = TextureHandle.ReadPackedId(textureKey.Handle, cachedTextureBuffer, cachedSamplerBuffer);
+
+                    int textureId = TextureHandle.UnpackTextureId(packedId);
+
+                    ref readonly Image.TextureDescriptor descriptor = ref pool.GetDescriptorRef(textureId);
+
+                    if (!MatchesTexture(kv.Value, descriptor))
                     {
-                    textureCbAddress = channel.BufferManager.GetComputeUniformBufferAddress(textureBufferIndex);
+                        return false;
-                    samplerCbAddress = channel.BufferManager.GetComputeUniformBufferAddress(samplerBufferIndex);
+                    }
                 }
-                else
-                {
-                    textureCbAddress = channel.BufferManager.GetGraphicsUniformBufferAddress(textureKey.StageIndex, textureBufferIndex);
-                    samplerCbAddress = channel.BufferManager.GetGraphicsUniformBufferAddress(textureKey.StageIndex, samplerBufferIndex);
             }
 
-                if (!channel.MemoryManager.Physical.IsMapped(textureCbAddress) || !channel.MemoryManager.Physical.IsMapped(samplerCbAddress))
+            return true;
-                {
-                    continue;
         }
 
-                Image.TextureDescriptor descriptor;
+        /// <summary>
-
+        /// Checks if the recorded texture state matches the given texture descriptor.
-                if (isCompute)
+        /// </summary>
+        /// <param name="specializationState">Texture specialization state</param>
+        /// <param name="descriptor">Texture descriptor</param>
+        /// <returns>True if the state matches, false otherwise</returns>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private bool MatchesTexture(Box<TextureSpecializationState> specializationState, in Image.TextureDescriptor descriptor)
         {
-                    descriptor = channel.TextureManager.GetComputeTextureDescriptor(
+            if (specializationState != null)
-                        poolState.TexturePoolGpuVa,
-                        poolState.TextureBufferIndex,
-                        poolState.TexturePoolMaximumId,
-                        textureKey.Handle,
-                        textureKey.CbufSlot);
-                }
-                else
             {
-                    descriptor = channel.TextureManager.GetGraphicsTextureDescriptor(
-                        poolState.TexturePoolGpuVa,
-                        poolState.TextureBufferIndex,
-                        poolState.TexturePoolMaximumId,
-                        textureKey.StageIndex,
-                        textureKey.Handle,
-                        textureKey.CbufSlot);
-                }
-
-                Box<TextureSpecializationState> specializationState = kv.Value;
-
                 if (specializationState.Value.QueriedFlags.HasFlag(QueriedTextureStateFlags.CoordNormalized) &&
                     specializationState.Value.CoordNormalized != descriptor.UnpackTextureCoordNormalized())
                 {
@@ -504,6 +615,34 @@ namespace Ryujinx.Graphics.Gpu.Shader
             return true;
         }
 
+        /// <summary>
+        /// Checks if the recorded texture state for a given texture binding matches a texture descriptor.
+        /// </summary>
+        /// <param name="stage">The shader stage</param>
+        /// <param name="index">The texture index</param>
+        /// <param name="descriptor">Texture descriptor</param>
+        /// <returns>True if the state matches, false otherwise</returns>
+        public bool MatchesTexture(ShaderStage stage, int index, in Image.TextureDescriptor descriptor)
+        {
+            Box<TextureSpecializationState> specializationState = _textureByBinding[(int)stage][index];
+
+            return MatchesTexture(specializationState, descriptor);
+        }
+
+        /// <summary>
+        /// Checks if the recorded texture state for a given image binding matches a texture descriptor.
+        /// </summary>
+        /// <param name="stage">The shader stage</param>
+        /// <param name="index">The texture index</param>
+        /// <param name="descriptor">Texture descriptor</param>
+        /// <returns>True if the state matches, false otherwise</returns>
+        public bool MatchesImage(ShaderStage stage, int index, in Image.TextureDescriptor descriptor)
+        {
+            Box<TextureSpecializationState> specializationState = _imageByBinding[(int)stage][index];
+
+            return MatchesTexture(specializationState, descriptor);
+        }
+
         /// <summary>
         /// Reads shader specialization state that has been serialized.
         /// </summary>

Ryujinx.Graphics.Shader/TextureHandle.cs

@@ -1,3 +1,4 @@
+using System;
 using System.Runtime.CompilerServices;
 
 namespace Ryujinx.Graphics.Shader
@@ -50,5 +51,63 @@ namespace Ryujinx.Graphics.Shader
         {
             return (handle & 0x3fff, (handle >> 14) & 0x3fff, (TextureHandleType)((uint)handle >> 28));
         }
+
+        /// <summary>
+        /// Unpacks the texture ID from the real texture handle.
+        /// </summary>
+        /// <param name="packedId">The real texture handle</param>
+        /// <returns>The texture ID</returns>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static int UnpackTextureId(int packedId)
+        {
+            return (packedId >> 0) & 0xfffff;
+        }
+
+        /// <summary>
+        /// Unpacks the sampler ID from the real texture handle.
+        /// </summary>
+        /// <param name="packedId">The real texture handle</param>
+        /// <returns>The sampler ID</returns>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static int UnpackSamplerId(int packedId)
+        {
+            return (packedId >> 20) & 0xfff;
+        }
+
+        /// <summary>
+        /// Reads a packed texture and sampler ID (basically, the real texture handle)
+        /// from a given texture/sampler constant buffer.
+        /// </summary>
+        /// <param name="wordOffset">A word offset of the handle on the buffer (the "fake" shader handle)</param>
+        /// <param name="cachedTextureBuffer">The constant buffer to fetch texture IDs from</param>
+        /// <param name="cachedSamplerBuffer">The constant buffer to fetch sampler IDs from</param>
+        /// <returns>The packed texture and sampler ID (the real texture handle)</returns>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static int ReadPackedId(int wordOffset, ReadOnlySpan<int> cachedTextureBuffer, ReadOnlySpan<int> cachedSamplerBuffer)
+        {
+            (int textureWordOffset, int samplerWordOffset, TextureHandleType handleType) = UnpackOffsets(wordOffset);
+
+            int handle = cachedTextureBuffer[textureWordOffset];
+
+            // The "wordOffset" (which is really the immediate value used on texture instructions on the shader)
+            // is a 13-bit value. However, in order to also support separate samplers and textures (which uses
+            // bindless textures on the shader), we extend it with another value on the higher 16 bits with
+            // another offset for the sampler.
+            // The shader translator has code to detect separate texture and sampler uses with a bindless texture,
+            // turn that into a regular texture access and produce those special handles with values on the higher 16 bits.
+            if (handleType != TextureHandleType.CombinedSampler)
+            {
+                int samplerHandle = cachedSamplerBuffer[samplerWordOffset];
+
+                if (handleType == TextureHandleType.SeparateSamplerId)
+                {
+                    samplerHandle <<= 20;
+                }
+
+                handle |= samplerHandle;
+            }
+
+            return handle;
+        }
     }
 }