Reword the description of the 6GB expand DRAM hack to be less tantalizing (#3870)

* Reword the hack to something less tantalizing

* Address feedback

.github/workflows/build.yml

@@ -52,26 +52,22 @@ jobs:
       - uses: actions/setup-dotnet@v3
         with:
           dotnet-version: 7.0.x
-      - name: Ensure NuGet Source
-        uses: fabriciomurta/ensure-nuget-source@v1
       - name: Get git short hash
         id: git_short_hash
         run: echo "result=$(git rev-parse --short "${{ github.sha }}")" >> $GITHUB_OUTPUT
         shell: bash
-      - name: Clear
-        run: dotnet clean && dotnet nuget locals all --clear
       - name: Build
-        run: dotnet build -c "${{ matrix.configuration }}" /p:Version="${{ env.RYUJINX_BASE_VERSION }}" /p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" /p:ExtraDefineConstants=DISABLE_UPDATER
+        run: dotnet build -c "${{ matrix.configuration }}" -p:Version="${{ env.RYUJINX_BASE_VERSION }}" -p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" -p:ExtraDefineConstants=DISABLE_UPDATER
       - name: Test
         run: dotnet test --no-build -c "${{ matrix.configuration }}"
       - name: Publish Ryujinx
-        run: dotnet publish -c "${{ matrix.configuration }}" -r "${{ matrix.DOTNET_RUNTIME_IDENTIFIER }}" -o ./publish /p:Version="${{ env.RYUJINX_BASE_VERSION }}" /p:DebugType=embedded /p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" /p:ExtraDefineConstants=DISABLE_UPDATER Ryujinx --self-contained
+        run: dotnet publish -c "${{ matrix.configuration }}" -r "${{ matrix.DOTNET_RUNTIME_IDENTIFIER }}" -o ./publish -p:Version="${{ env.RYUJINX_BASE_VERSION }}" -p:DebugType=embedded -p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" -p:ExtraDefineConstants=DISABLE_UPDATER Ryujinx --self-contained true
         if: github.event_name == 'pull_request'
       - name: Publish Ryujinx.Headless.SDL2
-        run: dotnet publish -c "${{ matrix.configuration }}" -r "${{ matrix.DOTNET_RUNTIME_IDENTIFIER }}" -o ./publish_sdl2_headless /p:Version="${{ env.RYUJINX_BASE_VERSION }}" /p:DebugType=embedded /p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" /p:ExtraDefineConstants=DISABLE_UPDATER Ryujinx.Headless.SDL2 --self-contained
+        run: dotnet publish -c "${{ matrix.configuration }}" -r "${{ matrix.DOTNET_RUNTIME_IDENTIFIER }}" -o ./publish_sdl2_headless -p:Version="${{ env.RYUJINX_BASE_VERSION }}" -p:DebugType=embedded -p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" -p:ExtraDefineConstants=DISABLE_UPDATER Ryujinx.Headless.SDL2 --self-contained true
         if: github.event_name == 'pull_request'
       - name: Publish Ryujinx.Ava
-        run: dotnet publish -c "${{ matrix.configuration }}" -r "${{ matrix.DOTNET_RUNTIME_IDENTIFIER }}" -o ./publish_ava /p:Version="1.0.0" /p:DebugType=embedded /p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" /p:ExtraDefineConstants=DISABLE_UPDATER Ryujinx.Ava
+        run: dotnet publish -c "${{ matrix.configuration }}" -r "${{ matrix.DOTNET_RUNTIME_IDENTIFIER }}" -o ./publish_ava -p:Version="${{ env.RYUJINX_BASE_VERSION }}" -p:DebugType=embedded -p:SourceRevisionId="${{ steps.git_short_hash.outputs.result }}" -p:ExtraDefineConstants=DISABLE_UPDATER Ryujinx.Ava --self-contained true
         if: github.event_name == 'pull_request'
       - name: Upload Ryujinx artifact
         uses: actions/upload-artifact@v3

.github/workflows/release.yml

@@ -29,10 +29,6 @@ jobs:
       - uses: actions/setup-dotnet@v3
         with:
           dotnet-version: 7.0.x
-      - name: Ensure NuGet Source
-        uses: fabriciomurta/ensure-nuget-source@v1
-      - name: Clear
-        run: dotnet clean && dotnet nuget locals all --clear
       - name: Get version info
         id: version_info
         run: |
@@ -51,9 +47,9 @@ jobs:
         run: "mkdir release_output"
       - name: Publish Windows
         run: |
-          dotnet publish -c Release -r win10-x64 -o ./publish_windows/publish /p:Version="${{ steps.version_info.outputs.build_version }}" /p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" /p:DebugType=embedded Ryujinx --self-contained
-          dotnet publish -c Release -r win10-x64 -o ./publish_windows_sdl2_headless/publish /p:Version="${{ steps.version_info.outputs.build_version }}" /p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" /p:DebugType=embedded Ryujinx.Headless.SDL2 --self-contained
-          dotnet publish -c Release -r win10-x64 -o ./publish_windows_ava/publish /p:Version="${{ steps.version_info.outputs.build_version }}" /p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" /p:DebugType=embedded Ryujinx.Ava --self-contained
+          dotnet publish -c Release -r win10-x64 -o ./publish_windows/publish -p:Version="${{ steps.version_info.outputs.build_version }}" -p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" -p:DebugType=embedded Ryujinx --self-contained true
+          dotnet publish -c Release -r win10-x64 -o ./publish_windows_sdl2_headless/publish -p:Version="${{ steps.version_info.outputs.build_version }}" -p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" -p:DebugType=embedded Ryujinx.Headless.SDL2 --self-contained true
+          dotnet publish -c Release -r win10-x64 -o ./publish_windows_ava/publish -p:Version="${{ steps.version_info.outputs.build_version }}" -p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" -p:DebugType=embedded Ryujinx.Ava --self-contained true
       - name: Packing Windows builds
         run: |
           pushd publish_windows
@@ -71,9 +67,9 @@ jobs:
 
       - name: Publish Linux
         run: |
-          dotnet publish -c Release -r linux-x64 -o ./publish_linux/publish /p:Version="${{ steps.version_info.outputs.build_version }}" /p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" /p:DebugType=embedded Ryujinx --self-contained
-          dotnet publish -c Release -r linux-x64 -o ./publish_linux_sdl2_headless/publish /p:Version="${{ steps.version_info.outputs.build_version }}" /p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" /p:DebugType=embedded Ryujinx.Headless.SDL2 --self-contained
-          dotnet publish -c Release -r linux-x64 -o ./publish_linux_ava/publish /p:Version="${{ steps.version_info.outputs.build_version }}" /p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" /p:DebugType=embedded Ryujinx.Ava --self-contained
+          dotnet publish -c Release -r linux-x64 -o ./publish_linux/publish -p:Version="${{ steps.version_info.outputs.build_version }}" -p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" -p:DebugType=embedded Ryujinx --self-contained true
+          dotnet publish -c Release -r linux-x64 -o ./publish_linux_sdl2_headless/publish -p:Version="${{ steps.version_info.outputs.build_version }}" -p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" -p:DebugType=embedded Ryujinx.Headless.SDL2 --self-contained true
+          dotnet publish -c Release -r linux-x64 -o ./publish_linux_ava/publish -p:Version="${{ steps.version_info.outputs.build_version }}" -p:SourceRevisionId="${{ steps.version_info.outputs.git_short_hash }}" -p:DebugType=embedded Ryujinx.Ava --self-contained true
 
       - name: Packing Linux builds
         run: |

ARMeilleure/Common/BitUtils.cs

@@ -1,15 +1,11 @@
+using System;
 using System.Numerics;
 
 namespace ARMeilleure.Common
 {
     static class BitUtils
     {
-        private static readonly sbyte[] HbsNibbleLut;
-
-        static BitUtils()
-        {
-            HbsNibbleLut = new sbyte[] { -1, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3 };
-        }
+        private static ReadOnlySpan<sbyte> HbsNibbleLut => new sbyte[] { -1, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3 };
 
         public static long FillWithOnes(int bits)
         {

Ryujinx.Ava/Assets/Locales/en_US.json

@@ -119,7 +119,7 @@
   "SettingsTabSystemAudioBackendSDL2": "SDL2",
   "SettingsTabSystemHacks": "Hacks",
   "SettingsTabSystemHacksNote": " (may cause instability)",
-  "SettingsTabSystemExpandDramSize": "Expand DRAM Size to 6GiB",
+  "SettingsTabSystemExpandDramSize": "Use alternative memory layout (Developers)",
   "SettingsTabSystemIgnoreMissingServices": "Ignore Missing Services",
   "SettingsTabGraphics": "Graphics",
   "SettingsTabGraphicsAPI": "Graphics API",
@@ -440,7 +440,7 @@
   "MemoryManagerSoftwareTooltip": "Use a software page table for address translation. Highest accuracy but slowest performance.",
   "MemoryManagerHostTooltip": "Directly map memory in the host address space. Much faster JIT compilation and execution.",
   "MemoryManagerUnsafeTooltip": "Directly map memory, but do not mask the address within the guest address space before access. Faster, but at the cost of safety. The guest application can access memory from anywhere in Ryujinx, so only run programs you trust with this mode.",
-  "DRamTooltip": "Increases the amount of memory on the emulated system from 4GiB to 6GiB.\n\nThis is only useful for higher-resolution texture packs or 4k resolution mods. Does NOT improve performance.\n\nLeave OFF if unsure.",
+  "DRamTooltip": "Utilizes an alternative MemoryMode layout to mimic a Switch development model.\n\nThis is only useful for higher-resolution texture packs or 4k resolution mods. Does NOT improve performance.\n\nLeave OFF if unsure.",
   "IgnoreMissingServicesTooltip": "Ignores unimplemented Horizon OS services. This may help in bypassing crashes when booting certain games.\n\nLeave OFF if unsure.",
   "GraphicsBackendThreadingTooltip": "Executes graphics backend commands on a second thread.\n\nSpeeds up shader compilation, reduces stuttering, and improves performance on GPU drivers without multithreading support of their own. Slightly better performance on drivers with multithreading.\n\nSet to AUTO if unsure.",
   "GalThreadingTooltip": "Executes graphics backend commands on a second thread.\n\nSpeeds up shader compilation, reduces stuttering, and improves performance on GPU drivers without multithreading support of their own. Slightly better performance on drivers with multithreading.\n\nSet to AUTO if unsure.",

Ryujinx.Ava/Ryujinx.Ava.csproj

@@ -36,7 +36,7 @@
     <PackageReference Include="Silk.NET.Vulkan" Version="2.10.1" />
     <PackageReference Include="Silk.NET.Vulkan.Extensions.EXT" Version="2.10.1" />
     <PackageReference Include="Silk.NET.Vulkan.Extensions.KHR" Version="2.10.1" />
-    <PackageReference Include="SPB" Version="0.0.4-build27" />
+    <PackageReference Include="SPB" Version="0.0.4-build28" />
     <PackageReference Include="SharpZipLib" Version="1.3.3" />
     <PackageReference Include="SixLabors.ImageSharp" Version="1.0.4" />
   </ItemGroup>

Ryujinx.Cpu/IExecutionContext.cs

@@ -27,7 +27,7 @@ namespace Ryujinx.Cpu
         long TpidrroEl0 { get; set; }
 
         /// <summary>
-        /// Processor State register.
+        /// Processor State Register.
         /// </summary>
         uint Pstate { get; set; }
 
@@ -109,4 +109,4 @@ namespace Ryujinx.Cpu
         /// </remarks>
         void StopRunning();
     }
-}
+}

Ryujinx.Graphics.GAL/Capabilities.cs

@@ -21,6 +21,7 @@ namespace Ryujinx.Graphics.GAL
         public readonly bool SupportsFragmentShaderOrderingIntel;
         public readonly bool SupportsGeometryShaderPassthrough;
         public readonly bool SupportsImageLoadFormatted;
+        public readonly bool SupportsLayerVertexTessellation;
         public readonly bool SupportsMismatchingViewFormat;
         public readonly bool SupportsCubemapView;
         public readonly bool SupportsNonConstantTextureOffset;
@@ -55,6 +56,7 @@ namespace Ryujinx.Graphics.GAL
             bool supportsFragmentShaderOrderingIntel,
             bool supportsGeometryShaderPassthrough,
             bool supportsImageLoadFormatted,
+            bool supportsLayerVertexTessellation,
             bool supportsMismatchingViewFormat,
             bool supportsCubemapView,
             bool supportsNonConstantTextureOffset,
@@ -86,6 +88,7 @@ namespace Ryujinx.Graphics.GAL
             SupportsFragmentShaderOrderingIntel = supportsFragmentShaderOrderingIntel;
             SupportsGeometryShaderPassthrough = supportsGeometryShaderPassthrough;
             SupportsImageLoadFormatted = supportsImageLoadFormatted;
+            SupportsLayerVertexTessellation = supportsLayerVertexTessellation;
             SupportsMismatchingViewFormat = supportsMismatchingViewFormat;
             SupportsCubemapView = supportsCubemapView;
             SupportsNonConstantTextureOffset = supportsNonConstantTextureOffset;

Ryujinx.Graphics.Gpu/Constants.cs

@@ -95,5 +95,10 @@ namespace Ryujinx.Graphics.Gpu
         /// Byte alignment for block linear textures
         /// </summary>
         public const int GobAlignment = 64;
+
+        /// <summary>
+        /// Expected byte alignment for storage buffers
+        /// </summary>
+        public const int StorageAlignment = 16;
     }
 }

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

@@ -138,7 +138,8 @@ namespace Ryujinx.Graphics.Gpu.Engine.Compute
                 qmd.CtaThreadDimension1,
                 qmd.CtaThreadDimension2,
                 localMemorySize,
-                sharedMemorySize);
+                sharedMemorySize,
+                _channel.BufferManager.HasUnalignedStorageBuffers);
 
             CachedShaderProgram cs = memoryManager.Physical.ShaderCache.GetComputeShader(_channel, poolState, computeState, shaderGpuVa);
 
@@ -150,6 +151,33 @@ namespace Ryujinx.Graphics.Gpu.Engine.Compute
 
             ShaderProgramInfo info = cs.Shaders[0].Info;
 
+            bool hasUnaligned = _channel.BufferManager.HasUnalignedStorageBuffers;
+
+            for (int index = 0; index < info.SBuffers.Count; index++)
+            {
+                BufferDescriptor sb = info.SBuffers[index];
+
+                ulong sbDescAddress = _channel.BufferManager.GetComputeUniformBufferAddress(0);
+
+                int sbDescOffset = 0x310 + sb.Slot * 0x10;
+
+                sbDescAddress += (ulong)sbDescOffset;
+
+                SbDescriptor sbDescriptor = _channel.MemoryManager.Physical.Read<SbDescriptor>(sbDescAddress);
+
+                _channel.BufferManager.SetComputeStorageBuffer(sb.Slot, sbDescriptor.PackAddress(), (uint)sbDescriptor.Size, sb.Flags);
+            }
+
+            if ((_channel.BufferManager.HasUnalignedStorageBuffers) != hasUnaligned)
+            {
+                // Refetch the shader, as assumptions about storage buffer alignment have changed.
+                cs = memoryManager.Physical.ShaderCache.GetComputeShader(_channel, poolState, computeState, shaderGpuVa);
+
+                _context.Renderer.Pipeline.SetProgram(cs.HostProgram);
+
+                info = cs.Shaders[0].Info;
+            }
+
             for (int index = 0; index < info.CBuffers.Count; index++)
             {
                 BufferDescriptor cb = info.CBuffers[index];
@@ -174,21 +202,6 @@ namespace Ryujinx.Graphics.Gpu.Engine.Compute
                 _channel.BufferManager.SetComputeUniformBuffer(cb.Slot, cbDescriptor.PackAddress(), (uint)cbDescriptor.Size);
             }
 
-            for (int index = 0; index < info.SBuffers.Count; index++)
-            {
-                BufferDescriptor sb = info.SBuffers[index];
-
-                ulong sbDescAddress = _channel.BufferManager.GetComputeUniformBufferAddress(0);
-
-                int sbDescOffset = 0x310 + sb.Slot * 0x10;
-
-                sbDescAddress += (ulong)sbDescOffset;
-
-                SbDescriptor sbDescriptor = _channel.MemoryManager.Physical.Read<SbDescriptor>(sbDescAddress);
-
-                _channel.BufferManager.SetComputeStorageBuffer(sb.Slot, sbDescriptor.PackAddress(), (uint)sbDescriptor.Size, sb.Flags);
-            }
-
             _channel.BufferManager.SetComputeStorageBufferBindings(info.SBuffers);
             _channel.BufferManager.SetComputeUniformBufferBindings(info.CBuffers);
 

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

@@ -293,9 +293,12 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
         /// </summary>
         private void CommitBindings()
         {
+            var buffers = _channel.BufferManager;
+            var hasUnaligned = buffers.HasUnalignedStorageBuffers;
+
             UpdateStorageBuffers();
 
-            if (!_channel.TextureManager.CommitGraphicsBindings(_shaderSpecState))
+            if (!_channel.TextureManager.CommitGraphicsBindings(_shaderSpecState) || (buffers.HasUnalignedStorageBuffers != hasUnaligned))
             {
                 // Shader must be reloaded.
                 UpdateShaderState();
@@ -1361,7 +1364,8 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
                 _state.State.AlphaTestFunc,
                 _state.State.AlphaTestRef,
                 ref attributeTypes,
-                _drawState.HasConstantBufferDrawParameters);
+                _drawState.HasConstantBufferDrawParameters,
+                _channel.BufferManager.HasUnalignedStorageBuffers);
         }
 
         /// <summary>

Ryujinx.Graphics.Gpu/GpuChannel.cs

@@ -67,6 +67,7 @@ namespace Ryujinx.Graphics.Gpu
 
             // Since the memory manager changed, make sure we will get pools from addresses of the new memory manager.
             TextureManager.ReloadPools();
+            MemoryManager.Physical.BufferCache.QueuePrune();
         }
 
         /// <summary>
@@ -77,6 +78,7 @@ namespace Ryujinx.Graphics.Gpu
         private void MemoryUnmappedHandler(object sender, UnmapEventArgs e)
         {
             TextureManager.ReloadPools();
+            MemoryManager.Physical.BufferCache.QueuePrune();
         }
 
         /// <summary>
@@ -131,6 +133,7 @@ namespace Ryujinx.Graphics.Gpu
             {
                 oldMemoryManager.Physical.BufferCache.NotifyBuffersModified -= BufferManager.Rebind;
                 oldMemoryManager.Physical.DecrementReferenceCount();
+                oldMemoryManager.MemoryUnmapped -= MemoryUnmappedHandler;
             }
         }
     }

Ryujinx.Graphics.Gpu/Memory/BufferCache.cs

@@ -28,6 +28,7 @@ namespace Ryujinx.Graphics.Gpu.Memory
 
         private readonly Dictionary<ulong, BufferCacheEntry> _dirtyCache;
         private readonly Dictionary<ulong, BufferCacheEntry> _modifiedCache;
+        private bool _pruneCaches;
 
         public event Action NotifyBuffersModified;
 
@@ -136,6 +137,11 @@ namespace Ryujinx.Graphics.Gpu.Memory
         /// <param name="size">Size in bytes of the buffer</param>
         public void ForceDirty(MemoryManager memoryManager, ulong gpuVa, ulong size)
         {
+            if (_pruneCaches)
+            {
+                Prune();
+            }
+
             if (!_dirtyCache.TryGetValue(gpuVa, out BufferCacheEntry result) ||
                 result.EndGpuAddress < gpuVa + size ||
                 result.UnmappedSequence != result.Buffer.UnmappedSequence)
@@ -158,17 +164,29 @@ namespace Ryujinx.Graphics.Gpu.Memory
         /// <returns>True if modified, false otherwise</returns>
         public bool CheckModified(MemoryManager memoryManager, ulong gpuVa, ulong size, out ulong outAddr)
         {
-            if (!_modifiedCache.TryGetValue(gpuVa, out BufferCacheEntry result) ||
-                result.EndGpuAddress < gpuVa + size ||
-                result.UnmappedSequence != result.Buffer.UnmappedSequence)
+            if (_pruneCaches)
             {
-                ulong address = TranslateAndCreateBuffer(memoryManager, gpuVa, size);
-                result = new BufferCacheEntry(address, gpuVa, GetBuffer(address, size));
-
-                _modifiedCache[gpuVa] = result;
+                Prune();
             }
 
-            outAddr = result.Address;
+            // Align the address to avoid creating too many entries on the quick lookup dictionary.
+            ulong mask = BufferAlignmentMask;
+            ulong alignedGpuVa = gpuVa & (~mask);
+            ulong alignedEndGpuVa = (gpuVa + size + mask) & (~mask);
+
+            size = alignedEndGpuVa - alignedGpuVa;
+
+            if (!_modifiedCache.TryGetValue(alignedGpuVa, out BufferCacheEntry result) ||
+                result.EndGpuAddress < alignedEndGpuVa ||
+                result.UnmappedSequence != result.Buffer.UnmappedSequence)
+            {
+                ulong address = TranslateAndCreateBuffer(memoryManager, alignedGpuVa, size);
+                result = new BufferCacheEntry(address, alignedGpuVa, GetBuffer(address, size));
+
+                _modifiedCache[alignedGpuVa] = result;
+            }
+
+            outAddr = result.Address | (gpuVa & mask);
 
             return result.Buffer.IsModified(result.Address, size);
         }
@@ -435,6 +453,54 @@ namespace Ryujinx.Graphics.Gpu.Memory
             }
         }
 
+        /// <summary>
+        /// Prune any invalid entries from a quick access dictionary.
+        /// </summary>
+        /// <param name="dictionary">Dictionary to prune</param>
+        /// <param name="toDelete">List used to track entries to delete</param>
+        private void Prune(Dictionary<ulong, BufferCacheEntry> dictionary, ref List<ulong> toDelete)
+        {
+            foreach (var entry in dictionary)
+            {
+                if (entry.Value.UnmappedSequence != entry.Value.Buffer.UnmappedSequence)
+                {
+                    (toDelete ??= new()).Add(entry.Key);
+                }
+            }
+
+            if (toDelete != null)
+            {
+                foreach (ulong entry in toDelete)
+                {
+                    dictionary.Remove(entry);
+                }
+            }
+        }
+
+        /// <summary>
+        /// Prune any invalid entries from the quick access dictionaries.
+        /// </summary>
+        private void Prune()
+        {
+            List<ulong> toDelete = null;
+
+            Prune(_dirtyCache, ref toDelete);
+
+            toDelete?.Clear();
+
+            Prune(_modifiedCache, ref toDelete);
+
+            _pruneCaches = false;
+        }
+
+        /// <summary>
+        /// Queues a prune of invalid entries the next time a dictionary cache is accessed.
+        /// </summary>
+        public void QueuePrune()
+        {
+            _pruneCaches = true;
+        }
+
         /// <summary>
         /// Disposes all buffers in the cache.
         /// It's an error to use the buffer manager after disposal.

Ryujinx.Graphics.Gpu/Memory/BufferManager.cs

@@ -17,6 +17,9 @@ namespace Ryujinx.Graphics.Gpu.Memory
         private readonly GpuContext _context;
         private readonly GpuChannel _channel;
 
+        private int _unalignedStorageBuffers;
+        public bool HasUnalignedStorageBuffers => _unalignedStorageBuffers > 0;
+
         private IndexBuffer _indexBuffer;
         private readonly VertexBuffer[] _vertexBuffers;
         private readonly BufferBounds[] _transformFeedbackBuffers;
@@ -38,6 +41,11 @@ namespace Ryujinx.Graphics.Gpu.Memory
             /// </summary>
             public BufferBounds[] Buffers { get; }
 
+            /// <summary>
+            /// Flag indicating if this binding is unaligned.
+            /// </summary>
+            public bool[] Unaligned { get; }
+
             /// <summary>
             /// Total amount of buffers used on the shader.
             /// </summary>
@@ -51,6 +59,7 @@ namespace Ryujinx.Graphics.Gpu.Memory
             {
                 Bindings = new BufferDescriptor[count];
                 Buffers = new BufferBounds[count];
+                Unaligned = new bool[count];
             }
 
             /// <summary>
@@ -202,6 +211,31 @@ namespace Ryujinx.Graphics.Gpu.Memory
             _transformFeedbackBuffersDirty = true;
         }
 
+        /// <summary>
+        /// Records the alignment of a storage buffer.
+        /// Unaligned storage buffers disable some optimizations on the shader.
+        /// </summary>
+        /// <param name="buffers">The binding list to modify</param>
+        /// <param name="index">Index of the storage buffer</param>
+        /// <param name="gpuVa">Start GPU virtual address of the buffer</param>
+        private void RecordStorageAlignment(BuffersPerStage buffers, int index, ulong gpuVa)
+        {
+            bool unaligned = (gpuVa & (Constants.StorageAlignment - 1)) != 0;
+
+            if (unaligned || HasUnalignedStorageBuffers)
+            {
+                // Check if the alignment changed for this binding.
+
+                ref bool currentUnaligned = ref buffers.Unaligned[index];
+
+                if (currentUnaligned != unaligned)
+                {
+                    currentUnaligned = unaligned;
+                    _unalignedStorageBuffers += unaligned ? 1 : -1;
+                }
+            }
+        }
+
         /// <summary>
         /// Sets a storage buffer on the compute pipeline.
         /// Storage buffers can be read and written to on shaders.
@@ -214,6 +248,8 @@ namespace Ryujinx.Graphics.Gpu.Memory
         {
             size += gpuVa & ((ulong)_context.Capabilities.StorageBufferOffsetAlignment - 1);
 
+            RecordStorageAlignment(_cpStorageBuffers, index, gpuVa);
+
             gpuVa = BitUtils.AlignDown(gpuVa, _context.Capabilities.StorageBufferOffsetAlignment);
 
             ulong address = _channel.MemoryManager.Physical.BufferCache.TranslateAndCreateBuffer(_channel.MemoryManager, gpuVa, size);
@@ -234,17 +270,21 @@ namespace Ryujinx.Graphics.Gpu.Memory
         {
             size += gpuVa & ((ulong)_context.Capabilities.StorageBufferOffsetAlignment - 1);
 
+            BuffersPerStage buffers = _gpStorageBuffers[stage];
+
+            RecordStorageAlignment(buffers, index, gpuVa);
+
             gpuVa = BitUtils.AlignDown(gpuVa, _context.Capabilities.StorageBufferOffsetAlignment);
 
             ulong address = _channel.MemoryManager.Physical.BufferCache.TranslateAndCreateBuffer(_channel.MemoryManager, gpuVa, size);
 
-            if (_gpStorageBuffers[stage].Buffers[index].Address != address ||
-                _gpStorageBuffers[stage].Buffers[index].Size != size)
+            if (buffers.Buffers[index].Address != address ||
+                buffers.Buffers[index].Size != size)
             {
                 _gpStorageBuffersDirty = true;
             }
 
-            _gpStorageBuffers[stage].SetBounds(index, address, size, flags);
+            buffers.SetBounds(index, address, size, flags);
         }
 
         /// <summary>

Ryujinx.Graphics.Gpu/Memory/BufferModifiedRangeList.cs

@@ -325,13 +325,15 @@ namespace Ryujinx.Graphics.Gpu.Memory
         public void ReregisterRanges(Action<ulong, ulong> rangeAction)
         {
             ref var ranges = ref ThreadStaticArray<BufferModifiedRange>.Get();
+            int count;
 
             // Range list must be consistent for this operation.
             lock (_lock)
             {
-                if (ranges.Length < Count)
+                count = Count;
+                if (ranges.Length < count)
                 {
-                    Array.Resize(ref ranges, Count);
+                    Array.Resize(ref ranges, count);
                 }
 
                 int i = 0;
@@ -342,7 +344,7 @@ namespace Ryujinx.Graphics.Gpu.Memory
             }
 
             ulong currentSync = _context.SyncNumber;
-            for (int i = 0; i < Count; i++)
+            for (int i = 0; i < count; i++)
             {
                 BufferModifiedRange range = ranges[i];
                 if (range.SyncNumber != currentSync)

Ryujinx.Graphics.Gpu/Shader/ComputeShaderCacheHashTable.cs

@@ -36,6 +36,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// </summary>
         /// <param name="channel">GPU channel</param>
         /// <param name="poolState">Texture pool state</param>
+        /// <param name="computeState">Compute state</param>
         /// <param name="gpuVa">GPU virtual address of the compute shader</param>
         /// <param name="program">Cached host program for the given state, if found</param>
         /// <param name="cachedGuestCode">Cached guest code, if any found</param>
@@ -43,6 +44,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
         public bool TryFind(
             GpuChannel channel,
             GpuChannelPoolState poolState,
+            GpuChannelComputeState computeState,
             ulong gpuVa,
             out CachedShaderProgram program,
             out byte[] cachedGuestCode)
@@ -50,7 +52,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
             program = null;
             ShaderCodeAccessor codeAccessor = new ShaderCodeAccessor(channel.MemoryManager, gpuVa);
             bool hasSpecList = _cache.TryFindItem(codeAccessor, out var specList, out cachedGuestCode);
-            return hasSpecList && specList.TryFindForCompute(channel, poolState, out program);
+            return hasSpecList && specList.TryFindForCompute(channel, poolState, computeState, out program);
         }
 
         /// <summary>

Ryujinx.Graphics.Gpu/Shader/DiskCache/DiskCacheGpuAccessor.cs

@@ -225,6 +225,12 @@ namespace Ryujinx.Graphics.Gpu.Shader.DiskCache
             return _oldSpecState.GraphicsState.EarlyZForce;
         }
 
+        /// <inheritdoc/>
+        public bool QueryHasUnalignedStorageBuffer()
+        {
+            return _oldSpecState.GraphicsState.HasUnalignedStorageBuffer || _oldSpecState.ComputeState.HasUnalignedStorageBuffer;
+        }
+
         /// <inheritdoc/>
         public bool QueryViewportTransformDisable()
         {

Ryujinx.Graphics.Gpu/Shader/DiskCache/DiskCacheHostStorage.cs

@@ -22,7 +22,7 @@ namespace Ryujinx.Graphics.Gpu.Shader.DiskCache
         private const ushort FileFormatVersionMajor = 1;
         private const ushort FileFormatVersionMinor = 2;
         private const uint FileFormatVersionPacked = ((uint)FileFormatVersionMajor << 16) | FileFormatVersionMinor;
-        private const uint CodeGenVersion = 3747;
+        private const uint CodeGenVersion = 3868;
 
         private const string SharedTocFileName = "shared.toc";
         private const string SharedDataFileName = "shared.data";
@@ -379,7 +379,7 @@ namespace Ryujinx.Graphics.Gpu.Shader.DiskCache
 
                             if (context.Capabilities.Api == TargetApi.Vulkan)
                             {
-                                ShaderSource[] shaderSources = ShaderBinarySerializer.Unpack(shaders, hostCode, isCompute);
+                                ShaderSource[] shaderSources = ShaderBinarySerializer.Unpack(shaders, hostCode);
 
                                 hostProgram = context.Renderer.CreateProgram(shaderSources, shaderInfo);
                             }

Ryujinx.Graphics.Gpu/Shader/DiskCache/ParallelDiskCacheLoader.cs

@@ -636,6 +636,8 @@ namespace Ryujinx.Graphics.Gpu.Shader.DiskCache
             CachedShaderStage[] shaders = new CachedShaderStage[guestShaders.Length];
             List<ShaderProgram> translatedStages = new List<ShaderProgram>();
 
+            TranslatorContext previousStage = null;
+
             for (int stageIndex = 0; stageIndex < Constants.ShaderStages; stageIndex++)
             {
                 TranslatorContext currentStage = translatorContexts[stageIndex + 1];
@@ -668,6 +670,16 @@ namespace Ryujinx.Graphics.Gpu.Shader.DiskCache
                     {
                         translatedStages.Add(program);
                     }
+
+                    previousStage = currentStage;
+                }
+                else if (
+                    previousStage != null &&
+                    previousStage.LayerOutputWritten &&
+                    stageIndex == 3 &&
+                    !_context.Capabilities.SupportsLayerVertexTessellation)
+                {
+                    translatedStages.Add(previousStage.GenerateGeometryPassthrough());
                 }
             }
 

Ryujinx.Graphics.Gpu/Shader/DiskCache/ShaderBinarySerializer.cs

@@ -1,5 +1,7 @@
 using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Shader;
 using Ryujinx.Graphics.Shader.Translation;
+using System;
 using System.Collections.Generic;
 using System.IO;
 
@@ -12,8 +14,11 @@ namespace Ryujinx.Graphics.Gpu.Shader.DiskCache
             using MemoryStream output = new MemoryStream();
             using BinaryWriter writer = new BinaryWriter(output);
 
+            writer.Write(sources.Length);
+
             for (int i = 0; i < sources.Length; i++)
             {
+                writer.Write((int)sources[i].Stage);
                 writer.Write(sources[i].BinaryCode.Length);
                 writer.Write(sources[i].BinaryCode);
             }
@@ -21,29 +26,40 @@ namespace Ryujinx.Graphics.Gpu.Shader.DiskCache
             return output.ToArray();
         }
 
-        public static ShaderSource[] Unpack(CachedShaderStage[] stages, byte[] code, bool compute)
+        public static ShaderSource[] Unpack(CachedShaderStage[] stages, byte[] code)
         {
             using MemoryStream input = new MemoryStream(code);
             using BinaryReader reader = new BinaryReader(input);
 
             List<ShaderSource> output = new List<ShaderSource>();
 
-            for (int i = compute ? 0 : 1; i < stages.Length; i++)
+            int count = reader.ReadInt32();
+
+            for (int i = 0; i < count; i++)
             {
-                CachedShaderStage stage = stages[i];
-
-                if (stage == null)
-                {
-                    continue;
-                }
-
+                ShaderStage stage = (ShaderStage)reader.ReadInt32();
                 int binaryCodeLength = reader.ReadInt32();
                 byte[] binaryCode = reader.ReadBytes(binaryCodeLength);
 
-                output.Add(new ShaderSource(binaryCode, ShaderCache.GetBindings(stage.Info), stage.Info.Stage, TargetLanguage.Spirv));
+                output.Add(new ShaderSource(binaryCode, GetBindings(stages, stage), stage, TargetLanguage.Spirv));
             }
 
             return output.ToArray();
         }
+
+        private static ShaderBindings GetBindings(CachedShaderStage[] stages, ShaderStage stage)
+        {
+            for (int i = 0; i < stages.Length; i++)
+            {
+                CachedShaderStage currentStage = stages[i];
+
+                if (currentStage != null && currentStage.Info.Stage == stage && currentStage.Info != null)
+                {
+                    return ShaderCache.GetBindings(currentStage.Info);
+                }
+            }
+
+            return new ShaderBindings(Array.Empty<int>(), Array.Empty<int>(), Array.Empty<int>(), Array.Empty<int>());
+        }
     }
 }

Ryujinx.Graphics.Gpu/Shader/GpuAccessor.cs

@@ -145,6 +145,12 @@ namespace Ryujinx.Graphics.Gpu.Shader
             return _state.GraphicsState.HasConstantBufferDrawParameters;
         }
 
+        /// <inheritdoc/>
+        public bool QueryHasUnalignedStorageBuffer()
+        {
+            return _state.GraphicsState.HasUnalignedStorageBuffer || _state.ComputeState.HasUnalignedStorageBuffer;
+        }
+
         /// <inheritdoc/>
         public InputTopology QueryPrimitiveTopology()
         {

Ryujinx.Graphics.Gpu/Shader/GpuAccessorBase.cs

@@ -128,6 +128,8 @@ namespace Ryujinx.Graphics.Gpu.Shader
 
         public bool QueryHostSupportsImageLoadFormatted() => _context.Capabilities.SupportsImageLoadFormatted;
 
+        public bool QueryHostSupportsLayerVertexTessellation() => _context.Capabilities.SupportsLayerVertexTessellation;
+
         public bool QueryHostSupportsNonConstantTextureOffset() => _context.Capabilities.SupportsNonConstantTextureOffset;
 
         public bool QueryHostSupportsShaderBallot() => _context.Capabilities.SupportsShaderBallot;

Ryujinx.Graphics.Gpu/Shader/GpuChannelComputeState.cs

@@ -32,6 +32,11 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// </summary>
         public readonly int SharedMemorySize;
 
+        /// <summary>
+        /// Indicates that any storage buffer use is unaligned.
+        /// </summary>
+        public readonly bool HasUnalignedStorageBuffer;
+
         /// <summary>
         /// Creates a new GPU compute state.
         /// </summary>
@@ -40,18 +45,21 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// <param name="localSizeZ">Local group size Z of the compute shader</param>
         /// <param name="localMemorySize">Local memory size of the compute shader</param>
         /// <param name="sharedMemorySize">Shared memory size of the compute shader</param>
+        /// <param name="hasUnalignedStorageBuffer">Indicates that any storage buffer use is unaligned</param>
         public GpuChannelComputeState(
             int localSizeX,
             int localSizeY,
             int localSizeZ,
             int localMemorySize,
-            int sharedMemorySize)
+            int sharedMemorySize,
+            bool hasUnalignedStorageBuffer)
         {
             LocalSizeX = localSizeX;
             LocalSizeY = localSizeY;
             LocalSizeZ = localSizeZ;
             LocalMemorySize = localMemorySize;
             SharedMemorySize = sharedMemorySize;
+            HasUnalignedStorageBuffer = hasUnalignedStorageBuffer;
         }
     }
 }

Ryujinx.Graphics.Gpu/Shader/GpuChannelGraphicsState.cs

@@ -82,6 +82,11 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// </summary>
         public readonly bool HasConstantBufferDrawParameters;
 
+        /// <summary>
+        /// Indicates that any storage buffer use is unaligned.
+        /// </summary>
+        public readonly bool HasUnalignedStorageBuffer;
+
         /// <summary>
         /// Creates a new GPU graphics state.
         /// </summary>
@@ -99,6 +104,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// <param name="alphaTestReference">When alpha test is enabled, indicates the value to compare with the fragment output alpha</param>
         /// <param name="attributeTypes">Type of the vertex attributes consumed by the shader</param>
         /// <param name="hasConstantBufferDrawParameters">Indicates that the draw is writing the base vertex, base instance and draw index to Constant Buffer 0</param>
+        /// <param name="hasUnalignedStorageBuffer">Indicates that any storage buffer use is unaligned</param>
         public GpuChannelGraphicsState(
             bool earlyZForce,
             PrimitiveTopology topology,
@@ -113,7 +119,8 @@ namespace Ryujinx.Graphics.Gpu.Shader
             CompareOp alphaTestCompare,
             float alphaTestReference,
             ref Array32<AttributeType> attributeTypes,
-            bool hasConstantBufferDrawParameters)
+            bool hasConstantBufferDrawParameters,
+            bool hasUnalignedStorageBuffer)
         {
             EarlyZForce = earlyZForce;
             Topology = topology;
@@ -129,6 +136,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
             AlphaTestReference = alphaTestReference;
             AttributeTypes = attributeTypes;
             HasConstantBufferDrawParameters = hasConstantBufferDrawParameters;
+            HasUnalignedStorageBuffer = hasUnalignedStorageBuffer;
         }
     }
 }

Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs

@@ -203,12 +203,12 @@ namespace Ryujinx.Graphics.Gpu.Shader
             GpuChannelComputeState computeState,
             ulong gpuVa)
         {
-            if (_cpPrograms.TryGetValue(gpuVa, out var cpShader) && IsShaderEqual(channel, poolState, cpShader, gpuVa))
+            if (_cpPrograms.TryGetValue(gpuVa, out var cpShader) && IsShaderEqual(channel, poolState, computeState, cpShader, gpuVa))
             {
                 return cpShader;
             }
 
-            if (_computeShaderCache.TryFind(channel, poolState, gpuVa, out cpShader, out byte[] cachedGuestCode))
+            if (_computeShaderCache.TryFind(channel, poolState, computeState, gpuVa, out cpShader, out byte[] cachedGuestCode))
             {
                 _cpPrograms[gpuVa] = cpShader;
                 return cpShader;
@@ -356,6 +356,8 @@ namespace Ryujinx.Graphics.Gpu.Shader
             CachedShaderStage[] shaders = new CachedShaderStage[Constants.ShaderStages + 1];
             List<ShaderSource> shaderSources = new List<ShaderSource>();
 
+            TranslatorContext previousStage = null;
+
             for (int stageIndex = 0; stageIndex < Constants.ShaderStages; stageIndex++)
             {
                 TranslatorContext currentStage = translatorContexts[stageIndex + 1];
@@ -392,6 +394,16 @@ namespace Ryujinx.Graphics.Gpu.Shader
                     {
                         shaderSources.Add(CreateShaderSource(program));
                     }
+
+                    previousStage = currentStage;
+                }
+                else if (
+                    previousStage != null &&
+                    previousStage.LayerOutputWritten &&
+                    stageIndex == 3 &&
+                    !_context.Capabilities.SupportsLayerVertexTessellation)
+                {
+                    shaderSources.Add(CreateShaderSource(previousStage.GenerateGeometryPassthrough()));
                 }
             }
 
@@ -473,18 +485,20 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// </summary>
         /// <param name="channel">GPU channel using the shader</param>
         /// <param name="poolState">GPU channel state to verify shader compatibility</param>
+        /// <param name="computeState">GPU channel compute state to verify shader compatibility</param>
         /// <param name="cpShader">Cached compute shader</param>
         /// <param name="gpuVa">GPU virtual address of the shader code in memory</param>
         /// <returns>True if the code is different, false otherwise</returns>
         private static bool IsShaderEqual(
             GpuChannel channel,
             GpuChannelPoolState poolState,
+            GpuChannelComputeState computeState,
             CachedShaderProgram cpShader,
             ulong gpuVa)
         {
             if (IsShaderEqual(channel.MemoryManager, cpShader.Shaders[0], gpuVa))
             {
-                return cpShader.SpecializationState.MatchesCompute(channel, poolState, true);
+                return cpShader.SpecializationState.MatchesCompute(channel, poolState, computeState, true);
             }
 
             return false;

Ryujinx.Graphics.Gpu/Shader/ShaderSpecializationList.cs

@@ -53,13 +53,14 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// </summary>
         /// <param name="channel">GPU channel</param>
         /// <param name="poolState">Texture pool state</param>
+        /// <param name="computeState">Compute state</param>
         /// <param name="program">Cached program, if found</param>
         /// <returns>True if a compatible program is found, false otherwise</returns>
-        public bool TryFindForCompute(GpuChannel channel, GpuChannelPoolState poolState, out CachedShaderProgram program)
+        public bool TryFindForCompute(GpuChannel channel, GpuChannelPoolState poolState, GpuChannelComputeState computeState, out CachedShaderProgram program)
         {
             foreach (var entry in _entries)
             {
-                if (entry.SpecializationState.MatchesCompute(channel, poolState, true))
+                if (entry.SpecializationState.MatchesCompute(channel, poolState, computeState, true))
                 {
                     program = entry;
                     return true;

Ryujinx.Graphics.Gpu/Shader/ShaderSpecializationState.cs

@@ -531,6 +531,11 @@ namespace Ryujinx.Graphics.Gpu.Shader
                 return false;
             }
 
+            if (graphicsState.HasUnalignedStorageBuffer != GraphicsState.HasUnalignedStorageBuffer)
+            {
+                return false;
+            }
+
             return Matches(channel, poolState, checkTextures, isCompute: false);
         }
 
@@ -539,10 +544,16 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// </summary>
         /// <param name="channel">GPU channel</param>
         /// <param name="poolState">Texture pool state</param>
+        /// <param name="computeState">Compute 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, bool checkTextures)
+        public bool MatchesCompute(GpuChannel channel, GpuChannelPoolState poolState, GpuChannelComputeState computeState, bool checkTextures)
         {
+            if (computeState.HasUnalignedStorageBuffer != ComputeState.HasUnalignedStorageBuffer)
+            {
+                return false;
+            }
+
             return Matches(channel, poolState, checkTextures, isCompute: true);
         }
 

Ryujinx.Graphics.Nvdec.Vp9/Detokenize.cs

@@ -63,7 +63,7 @@ namespace Ryujinx.Graphics.Nvdec.Vp9
             short dqv = dq[0];
             ReadOnlySpan<byte> cat6Prob = (xd.Bd == 12)
                 ? Luts.Vp9Cat6ProbHigh12
-                : (xd.Bd == 10) ? new ReadOnlySpan<byte>(Luts.Vp9Cat6ProbHigh12).Slice(2) : Luts.Vp9Cat6Prob;
+                : (xd.Bd == 10) ? Luts.Vp9Cat6ProbHigh12.Slice(2) : Luts.Vp9Cat6Prob;
             int cat6Bits = (xd.Bd == 12) ? 18 : (xd.Bd == 10) ? 16 : 14;
             // Keep value, range, and count as locals.  The compiler produces better
             // results with the locals than using r directly.

Ryujinx.Graphics.Nvdec.Vp9/Luts.cs

@@ -1,14 +1,12 @@
 using Ryujinx.Common.Memory;
 using Ryujinx.Graphics.Nvdec.Vp9.Types;
+using System;
 
 namespace Ryujinx.Graphics.Nvdec.Vp9
 {
     internal static class Luts
     {
-        public static readonly byte[] SizeGroupLookup = new byte[]
-        {
-            0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 3
-        };
+        public static ReadOnlySpan<byte> SizeGroupLookup => new byte[] { 0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 3 };
 
         public static readonly BlockSize[][] SubsizeLookup = new BlockSize[][]
         {
@@ -1070,18 +1068,18 @@ namespace Ryujinx.Graphics.Nvdec.Vp9
             -(sbyte)MvClassType.MvClass10,
         };
 
-        public static readonly sbyte[] Vp9MvFPTree = new sbyte[] { -0, 2, -1, 4, -2, -3 };
+        public static ReadOnlySpan<sbyte> Vp9MvFPTree => new sbyte[] { -0, 2, -1, 4, -2, -3 };
 
         // Entropy
 
-        public static readonly byte[] Vp9Cat1Prob = new byte[] { 159 };
-        public static readonly byte[] Vp9Cat2Prob = new byte[] { 165, 145 };
-        public static readonly byte[] Vp9Cat3Prob = new byte[] { 173, 148, 140 };
-        public static readonly byte[] Vp9Cat4Prob = new byte[] { 176, 155, 140, 135 };
-        public static readonly byte[] Vp9Cat5Prob = new byte[] { 180, 157, 141, 134, 130 };
-        public static readonly byte[] Vp9Cat6Prob = new byte[] { 254, 254, 254, 252, 249, 243, 230, 196, 177, 153, 140, 133, 130, 129 };
+        public static ReadOnlySpan<byte> Vp9Cat1Prob => new byte[] { 159 };
+        public static ReadOnlySpan<byte> Vp9Cat2Prob => new byte[] { 165, 145 };
+        public static ReadOnlySpan<byte> Vp9Cat3Prob => new byte[] { 173, 148, 140 };
+        public static ReadOnlySpan<byte> Vp9Cat4Prob => new byte[] { 176, 155, 140, 135 };
+        public static ReadOnlySpan<byte> Vp9Cat5Prob => new byte[] { 180, 157, 141, 134, 130 };
+        public static ReadOnlySpan<byte> Vp9Cat6Prob => new byte[] { 254, 254, 254, 252, 249, 243, 230, 196, 177, 153, 140, 133, 130, 129 };
 
-        public static readonly byte[] Vp9Cat6ProbHigh12 = new byte[]
+        public static ReadOnlySpan<byte> Vp9Cat6ProbHigh12 => new byte[]
         {
             255, 255, 255, 255, 254, 254, 54, 252, 249, 243, 230, 196, 177, 153, 140, 133, 130, 129
         };
@@ -1131,12 +1129,12 @@ namespace Ryujinx.Graphics.Nvdec.Vp9
             5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
         };
 
-        private static readonly byte[] Vp9CoefbandTrans4X4 = new byte[]
+        private static ReadOnlySpan<byte> Vp9CoefbandTrans4X4 => new byte[]
         {
             0, 1, 1, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 5, 5, 5,
         };
 
-        public static byte[] get_band_translate(TxSize txSize)
+        public static ReadOnlySpan<byte> get_band_translate(TxSize txSize)
         {
             return txSize == TxSize.Tx4x4 ? Vp9CoefbandTrans4X4 : Vp9CoefbandTrans8X8Plus;
         }

Ryujinx.Graphics.OpenGL/HwCapabilities.cs

@@ -18,6 +18,7 @@ namespace Ryujinx.Graphics.OpenGL
         private static readonly Lazy<bool> _supportsQuads                        = new Lazy<bool>(SupportsQuadsCheck);
         private static readonly Lazy<bool> _supportsSeamlessCubemapPerTexture    = new Lazy<bool>(() => HasExtension("GL_ARB_seamless_cubemap_per_texture"));
         private static readonly Lazy<bool> _supportsShaderBallot                 = new Lazy<bool>(() => HasExtension("GL_ARB_shader_ballot"));
+        private static readonly Lazy<bool> _supportsShaderViewportLayerArray     = new Lazy<bool>(() => HasExtension("GL_ARB_shader_viewport_layer_array"));
         private static readonly Lazy<bool> _supportsTextureCompressionBptc       = new Lazy<bool>(() => HasExtension("GL_EXT_texture_compression_bptc"));
         private static readonly Lazy<bool> _supportsTextureCompressionRgtc       = new Lazy<bool>(() => HasExtension("GL_EXT_texture_compression_rgtc"));
         private static readonly Lazy<bool> _supportsTextureCompressionS3tc       = new Lazy<bool>(() => HasExtension("GL_EXT_texture_compression_s3tc"));
@@ -61,6 +62,7 @@ namespace Ryujinx.Graphics.OpenGL
         public static bool SupportsQuads                        => _supportsQuads.Value;
         public static bool SupportsSeamlessCubemapPerTexture    => _supportsSeamlessCubemapPerTexture.Value;
         public static bool SupportsShaderBallot                 => _supportsShaderBallot.Value;
+        public static bool SupportsShaderViewportLayerArray     => _supportsShaderViewportLayerArray.Value;
         public static bool SupportsTextureCompressionBptc       => _supportsTextureCompressionBptc.Value;
         public static bool SupportsTextureCompressionRgtc       => _supportsTextureCompressionRgtc.Value;
         public static bool SupportsTextureCompressionS3tc       => _supportsTextureCompressionS3tc.Value;

Ryujinx.Graphics.OpenGL/OpenGLRenderer.cs

@@ -117,12 +117,13 @@ namespace Ryujinx.Graphics.OpenGL
                 supportsFragmentShaderOrderingIntel: HwCapabilities.SupportsFragmentShaderOrdering,
                 supportsGeometryShaderPassthrough: HwCapabilities.SupportsGeometryShaderPassthrough,
                 supportsImageLoadFormatted: HwCapabilities.SupportsImageLoadFormatted,
+                supportsLayerVertexTessellation: HwCapabilities.SupportsShaderViewportLayerArray,
                 supportsMismatchingViewFormat: HwCapabilities.SupportsMismatchingViewFormat,
                 supportsCubemapView: true,
                 supportsNonConstantTextureOffset: HwCapabilities.SupportsNonConstantTextureOffset,
                 supportsShaderBallot: HwCapabilities.SupportsShaderBallot,
                 supportsTextureShadowLod: HwCapabilities.SupportsTextureShadowLod,
-                supportsViewportIndex: true,
+                supportsViewportIndex: HwCapabilities.SupportsShaderViewportLayerArray,
                 supportsViewportSwizzle: HwCapabilities.SupportsViewportSwizzle,
                 supportsIndirectParameters: HwCapabilities.SupportsIndirectParameters,
                 maximumUniformBuffersPerStage: 13, // TODO: Avoid hardcoding those limits here and get from driver?

Ryujinx.Graphics.Shader/CodeGen/Spirv/Instructions.cs

@@ -1829,7 +1829,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
 
                 if (texOp.Index < 2 || (type & SamplerType.Mask) == SamplerType.Texture3D)
                 {
-                    result = ScalingHelpers.ApplyUnscaling(context, texOp, result, isBindless, isIndexed);
+                    result = ScalingHelpers.ApplyUnscaling(context, texOp.WithType(type), result, isBindless, isIndexed);
                 }
 
                 return new OperationResult(AggregateType.S32, result);

Ryujinx.Graphics.Shader/Constants.cs

@@ -10,5 +10,7 @@ namespace Ryujinx.Graphics.Shader
         public const int NvnBaseVertexByteOffset = 0x640;
         public const int NvnBaseInstanceByteOffset = 0x644;
         public const int NvnDrawIndexByteOffset = 0x648;
+
+        public const int StorageAlignment = 16;
     }
 }

Ryujinx.Graphics.Shader/IGpuAccessor.cs

@@ -177,6 +177,15 @@ namespace Ryujinx.Graphics.Shader
             return false;
         }
 
+        /// <summary>
+        /// Queries whenever the current draw uses unaligned storage buffer addresses.
+        /// </summary>
+        /// <returns>True if any storage buffer address is not aligned to 16 bytes, false otherwise</returns>
+        bool QueryHasUnalignedStorageBuffer()
+        {
+            return false;
+        }
+
         /// <summary>
         /// Queries host about the presence of the FrontFacing built-in variable bug.
         /// </summary>
@@ -249,6 +258,15 @@ namespace Ryujinx.Graphics.Shader
             return true;
         }
 
+        /// <summary>
+        /// Queries host support for writes to Layer from vertex or tessellation shader stages.
+        /// </summary>
+        /// <returns>True if writes to layer from vertex or tessellation are supported, false otherwise</returns>
+        bool QueryHostSupportsLayerVertexTessellation()
+        {
+            return true;
+        }
+
         /// <summary>
         /// Queries host GPU non-constant texture offset support.
         /// </summary>

Ryujinx.Graphics.Shader/Instructions/InstEmitAttribute.cs

@@ -278,13 +278,21 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
         private static int FixedFuncToUserAttribute(ShaderConfig config, int attr, bool isOutput)
         {
-            if (attr >= AttributeConsts.FrontColorDiffuseR && attr < AttributeConsts.ClipDistance0)
+            bool supportsLayerFromVertexOrTess = config.GpuAccessor.QueryHostSupportsLayerVertexTessellation();
+            int fixedStartAttr = supportsLayerFromVertexOrTess ? 0 : 1;
+
+            if (attr == AttributeConsts.Layer && config.Stage != ShaderStage.Geometry && !supportsLayerFromVertexOrTess)
             {
-                attr = FixedFuncToUserAttribute(config, attr, AttributeConsts.FrontColorDiffuseR, 0, isOutput);
+                attr = FixedFuncToUserAttribute(config, attr, AttributeConsts.Layer, 0, isOutput);
+                config.SetLayerOutputAttribute(attr);
+            }
+            else if (attr >= AttributeConsts.FrontColorDiffuseR && attr < AttributeConsts.ClipDistance0)
+            {
+                attr = FixedFuncToUserAttribute(config, attr, AttributeConsts.FrontColorDiffuseR, fixedStartAttr, isOutput);
             }
             else if (attr >= AttributeConsts.TexCoordBase && attr < AttributeConsts.TexCoordEnd)
             {
-                attr = FixedFuncToUserAttribute(config, attr, AttributeConsts.TexCoordBase, 4, isOutput);
+                attr = FixedFuncToUserAttribute(config, attr, AttributeConsts.TexCoordBase, fixedStartAttr + 4, isOutput);
             }
 
             return attr;

Ryujinx.Graphics.Shader/StructuredIr/AstTextureOperation.cs

@@ -27,5 +27,10 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
             CbufSlot = cbufSlot;
             Handle = handle;
         }
+
+        public AstTextureOperation WithType(SamplerType type)
+        {
+            return new AstTextureOperation(Inst, type, Format, Flags, CbufSlot, Handle, Index);
+        }
     }
 }

Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs

@@ -34,7 +34,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
                         // we can guess which storage buffer it is accessing.
                         // We can then replace the global memory access with a storage
                         // buffer access.
-                        node = ReplaceGlobalWithStorage(node, config, storageIndex);
+                        node = ReplaceGlobalWithStorage(block, node, config, storageIndex);
                     }
                     else if (config.Stage == ShaderStage.Compute && operation.Inst == Instruction.LoadGlobal)
                     {
@@ -54,7 +54,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
             }
         }
 
-        private static LinkedListNode<INode> ReplaceGlobalWithStorage(LinkedListNode<INode> node, ShaderConfig config, int storageIndex)
+        private static LinkedListNode<INode> ReplaceGlobalWithStorage(BasicBlock block, LinkedListNode<INode> node, ShaderConfig config, int storageIndex)
         {
             Operation operation = (Operation)node.Value;
 
@@ -64,42 +64,10 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
 
             config.SetUsedStorageBuffer(storageIndex, isWrite);
 
-            Operand GetStorageOffset()
-            {
-                Operand addrLow = operation.GetSource(0);
-
-                Operand baseAddrLow = Cbuf(0, GetStorageCbOffset(config.Stage, storageIndex));
-
-                Operand baseAddrTrunc = Local();
-
-                Operand alignMask = Const(-config.GpuAccessor.QueryHostStorageBufferOffsetAlignment());
-
-                Operation andOp = new Operation(Instruction.BitwiseAnd, baseAddrTrunc, baseAddrLow, alignMask);
-
-                node.List.AddBefore(node, andOp);
-
-                Operand byteOffset = Local();
-                Operation subOp = new Operation(Instruction.Subtract, byteOffset, addrLow, baseAddrTrunc);
-
-                node.List.AddBefore(node, subOp);
-
-                if (isStg16Or8)
-                {
-                    return byteOffset;
-                }
-
-                Operand wordOffset = Local();
-                Operation shrOp = new Operation(Instruction.ShiftRightU32, wordOffset, byteOffset, Const(2));
-
-                node.List.AddBefore(node, shrOp);
-
-                return wordOffset;
-            }
-
             Operand[] sources = new Operand[operation.SourcesCount];
 
             sources[0] = Const(storageIndex);
-            sources[1] = GetStorageOffset();
+            sources[1] = GetStorageOffset(block, node, config, storageIndex, operation.GetSource(0), isStg16Or8);
 
             for (int index = 2; index < operation.SourcesCount; index++)
             {
@@ -144,6 +112,170 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
             return node;
         }
 
+        private static Operand GetStorageOffset(
+            BasicBlock block,
+            LinkedListNode<INode> node,
+            ShaderConfig config,
+            int storageIndex,
+            Operand addrLow,
+            bool isStg16Or8)
+        {
+            int baseAddressCbOffset = GetStorageCbOffset(config.Stage, storageIndex);
+
+            bool storageAligned = !(config.GpuAccessor.QueryHasUnalignedStorageBuffer() || config.GpuAccessor.QueryHostStorageBufferOffsetAlignment() > Constants.StorageAlignment);
+
+            (Operand byteOffset, int constantOffset) = storageAligned ?
+                GetStorageOffset(block, Utils.FindLastOperation(addrLow, block), baseAddressCbOffset) :
+                (null, 0);
+
+            if (byteOffset == null)
+            {
+                Operand baseAddrLow = Cbuf(0, baseAddressCbOffset);
+                Operand baseAddrTrunc = Local();
+
+                Operand alignMask = Const(-config.GpuAccessor.QueryHostStorageBufferOffsetAlignment());
+
+                Operation andOp = new Operation(Instruction.BitwiseAnd, baseAddrTrunc, baseAddrLow, alignMask);
+
+                node.List.AddBefore(node, andOp);
+
+                Operand offset = Local();
+                Operation subOp = new Operation(Instruction.Subtract, offset, addrLow, baseAddrTrunc);
+
+                node.List.AddBefore(node, subOp);
+
+                byteOffset = offset;
+            }
+            else if (constantOffset != 0)
+            {
+                Operand offset = Local();
+                Operation addOp = new Operation(Instruction.Add, offset, byteOffset, Const(constantOffset));
+
+                node.List.AddBefore(node, addOp);
+
+                byteOffset = offset;
+            }
+
+            if (byteOffset != null)
+            {
+                ReplaceAddressAlignment(node.List, addrLow, byteOffset, constantOffset);
+            }
+
+            if (isStg16Or8)
+            {
+                return byteOffset;
+            }
+
+            Operand wordOffset = Local();
+            Operation shrOp = new Operation(Instruction.ShiftRightU32, wordOffset, byteOffset, Const(2));
+
+            node.List.AddBefore(node, shrOp);
+
+            return wordOffset;
+        }
+
+        private static bool IsCb0Offset(Operand operand, int offset)
+        {
+            return operand.Type == OperandType.ConstantBuffer && operand.GetCbufSlot() == 0 && operand.GetCbufOffset() == offset;
+        }
+
+        private static void ReplaceAddressAlignment(LinkedList<INode> list, Operand address, Operand byteOffset, int constantOffset)
+        {
+            // When we emit 16/8-bit LDG, we add extra code to determine the address alignment.
+            // Eliminate the storage buffer base address from this too, leaving only the byte offset.
+
+            foreach (INode useNode in address.UseOps)
+            {
+                if (useNode is Operation op && op.Inst == Instruction.BitwiseAnd)
+                {
+                    Operand src1 = op.GetSource(0);
+                    Operand src2 = op.GetSource(1);
+
+                    int addressIndex = -1;
+
+                    if (src1 == address && src2.Type == OperandType.Constant && src2.Value == 3)
+                    {
+                        addressIndex = 0;
+                    }
+                    else if (src2 == address && src1.Type == OperandType.Constant && src1.Value == 3)
+                    {
+                        addressIndex = 1;
+                    }
+
+                    if (addressIndex != -1)
+                    {
+                        LinkedListNode<INode> node = list.Find(op);
+
+                        // Add offset calculation before the use. Needs to be on the same block.
+                        if (node != null)
+                        {
+                            Operand offset = Local();
+                            Operation addOp = new Operation(Instruction.Add, offset, byteOffset, Const(constantOffset));
+                            list.AddBefore(node, addOp);
+
+                            op.SetSource(addressIndex, offset);
+                        }
+                    }
+                }
+            }
+        }
+
+        private static (Operand, int) GetStorageOffset(BasicBlock block, Operand address, int baseAddressCbOffset)
+        {
+            if (IsCb0Offset(address, baseAddressCbOffset))
+            {
+                // Direct offset: zero.
+                return (Const(0), 0);
+            }
+
+            (address, int constantOffset) = GetStorageConstantOffset(block, address);
+
+            address = Utils.FindLastOperation(address, block);
+
+            if (IsCb0Offset(address, baseAddressCbOffset))
+            {
+                // Only constant offset
+                return (Const(0), constantOffset);
+            }
+
+            if (!(address.AsgOp is Operation offsetAdd) || offsetAdd.Inst != Instruction.Add)
+            {
+                return (null, 0);
+            }
+
+            Operand src1 = offsetAdd.GetSource(0);
+            Operand src2 = Utils.FindLastOperation(offsetAdd.GetSource(1), block);
+
+            if (IsCb0Offset(src2, baseAddressCbOffset))
+            {
+                return (src1, constantOffset);
+            }
+            else if (IsCb0Offset(src1, baseAddressCbOffset))
+            {
+                return (src2, constantOffset);
+            }
+
+            return (null, 0);
+        }
+
+        private static (Operand, int) GetStorageConstantOffset(BasicBlock block, Operand address)
+        {
+            if (!(address.AsgOp is Operation offsetAdd) || offsetAdd.Inst != Instruction.Add)
+            {
+                return (address, 0);
+            }
+
+            Operand src1 = offsetAdd.GetSource(0);
+            Operand src2 = offsetAdd.GetSource(1);
+
+            if (src2.Type != OperandType.Constant)
+            {
+                return (address, 0);
+            }
+
+            return (src1, src2.Value);
+        }
+
         private static LinkedListNode<INode> ReplaceLdgWithLdc(LinkedListNode<INode> node, ShaderConfig config, int storageIndex)
         {
             Operation operation = (Operation)node.Value;
@@ -165,7 +297,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
                 Operand byteOffset = Local();
                 Operand wordOffset = Local();
 
-                Operation subOp = new Operation(Instruction.Subtract,      byteOffset, addrLow, baseAddrTrunc);
+                Operation subOp = new Operation(Instruction.Subtract, byteOffset, addrLow, baseAddrTrunc);
                 Operation shrOp = new Operation(Instruction.ShiftRightU32, wordOffset, byteOffset, Const(2));
 
                 node.List.AddBefore(node, subOp);
@@ -260,7 +392,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
         {
             if (operand.Type == OperandType.ConstantBuffer)
             {
-                int slot   = operand.GetCbufSlot();
+                int slot = operand.GetCbufSlot();
                 int offset = operand.GetCbufOffset();
 
                 if (slot == 0 && offset >= sbStart && offset < sbEnd)

Ryujinx.Graphics.Shader/Translation/ShaderConfig.cs

@@ -48,6 +48,9 @@ namespace Ryujinx.Graphics.Shader.Translation
 
         public int Cb1DataSize { get; private set; }
 
+        public bool LayerOutputWritten { get; private set; }
+        public int LayerOutputAttribute { get; private set; }
+
         public bool NextUsesFixedFuncAttributes { get; private set; }
         public int UsedInputAttributes { get; private set; }
         public int UsedOutputAttributes { get; private set; }
@@ -131,6 +134,20 @@ namespace Ryujinx.Graphics.Shader.Translation
             _usedImages   = new Dictionary<TextureInfo, TextureMeta>();
         }
 
+        public ShaderConfig(
+            ShaderStage stage,
+            OutputTopology outputTopology,
+            int maxOutputVertices,
+            IGpuAccessor gpuAccessor,
+            TranslationOptions options) : this(gpuAccessor, options)
+        {
+            Stage                    = stage;
+            ThreadsPerInputPrimitive = 1;
+            OutputTopology           = outputTopology;
+            MaxOutputVertices        = maxOutputVertices;
+            TransformFeedbackEnabled = gpuAccessor.QueryTransformFeedbackEnabled();
+        }
+
         public ShaderConfig(ShaderHeader header, IGpuAccessor gpuAccessor, TranslationOptions options) : this(gpuAccessor, options)
         {
             Stage                    = header.Stage;
@@ -240,6 +257,12 @@ namespace Ryujinx.Graphics.Shader.Translation
             }
         }
 
+        public void SetLayerOutputAttribute(int attr)
+        {
+            LayerOutputWritten = true;
+            LayerOutputAttribute = attr;
+        }
+
         public void SetInputUserAttributeFixedFunc(int index)
         {
             UsedInputAttributes |= 1 << index;
@@ -694,5 +717,20 @@ namespace Ryujinx.Graphics.Shader.Translation
         {
             return FindDescriptorIndex(GetImageDescriptors(), texOp);
         }
+
+        public ShaderProgramInfo CreateProgramInfo()
+        {
+            return new ShaderProgramInfo(
+                GetConstantBufferDescriptors(),
+                GetStorageBufferDescriptors(),
+                GetTextureDescriptors(),
+                GetImageDescriptors(),
+                Stage,
+                UsedFeatures.HasFlag(FeatureFlags.InstanceId),
+                UsedFeatures.HasFlag(FeatureFlags.DrawParameters),
+                UsedFeatures.HasFlag(FeatureFlags.RtLayer),
+                ClipDistancesWritten,
+                OmapTargets);
+        }
     }
 }

Ryujinx.Graphics.Shader/Translation/Translator.cs

@@ -79,17 +79,7 @@ namespace Ryujinx.Graphics.Shader.Translation
 
             var sInfo = StructuredProgram.MakeStructuredProgram(funcs, config);
 
-            var info = new ShaderProgramInfo(
-                config.GetConstantBufferDescriptors(),
-                config.GetStorageBufferDescriptors(),
-                config.GetTextureDescriptors(),
-                config.GetImageDescriptors(),
-                config.Stage,
-                config.UsedFeatures.HasFlag(FeatureFlags.InstanceId),
-                config.UsedFeatures.HasFlag(FeatureFlags.DrawParameters),
-                config.UsedFeatures.HasFlag(FeatureFlags.RtLayer),
-                config.ClipDistancesWritten,
-                config.OmapTargets);
+            var info = config.CreateProgramInfo();
 
             return config.Options.TargetLanguage switch
             {

Ryujinx.Graphics.Shader/Translation/TranslatorContext.cs

@@ -1,7 +1,12 @@
-using Ryujinx.Graphics.Shader.Decoders;
+using Ryujinx.Graphics.Shader.CodeGen.Glsl;
+using Ryujinx.Graphics.Shader.CodeGen.Spirv;
+using Ryujinx.Graphics.Shader.Decoders;
 using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.StructuredIr;
+using System;
 using System.Collections.Generic;
 using System.Linq;
+using System.Numerics;
 
 using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
 using static Ryujinx.Graphics.Shader.Translation.Translator;
@@ -18,6 +23,7 @@ namespace Ryujinx.Graphics.Shader.Translation
         public ShaderStage Stage => _config.Stage;
         public int Size => _config.Size;
         public int Cb1DataSize => _config.Cb1DataSize;
+        public bool LayerOutputWritten => _config.LayerOutputWritten;
 
         public IGpuAccessor GpuAccessor => _config.GpuAccessor;
 
@@ -149,5 +155,94 @@ namespace Ryujinx.Graphics.Shader.Translation
 
             return Translator.Translate(code, _config);
         }
+
+        public ShaderProgram GenerateGeometryPassthrough()
+        {
+            int outputAttributesMask = _config.UsedOutputAttributes;
+            int layerOutputAttr = _config.LayerOutputAttribute;
+
+            OutputTopology outputTopology;
+            int maxOutputVertices;
+
+            switch (GpuAccessor.QueryPrimitiveTopology())
+            {
+                case InputTopology.Points:
+                    outputTopology = OutputTopology.PointList;
+                    maxOutputVertices = 1;
+                    break;
+                case InputTopology.Lines:
+                case InputTopology.LinesAdjacency:
+                    outputTopology = OutputTopology.LineStrip;
+                    maxOutputVertices = 2;
+                    break;
+                default:
+                    outputTopology = OutputTopology.TriangleStrip;
+                    maxOutputVertices = 3;
+                    break;
+            }
+
+            ShaderConfig config = new ShaderConfig(ShaderStage.Geometry, outputTopology, maxOutputVertices, GpuAccessor, _config.Options);
+
+            EmitterContext context = new EmitterContext(default, config, false);
+
+            for (int v = 0; v < maxOutputVertices; v++)
+            {
+                int outAttrsMask = outputAttributesMask;
+
+                while (outAttrsMask != 0)
+                {
+                    int attrIndex = BitOperations.TrailingZeroCount(outAttrsMask);
+
+                    outAttrsMask &= ~(1 << attrIndex);
+
+                    for (int c = 0; c < 4; c++)
+                    {
+                        int attr = AttributeConsts.UserAttributeBase + attrIndex * 16 + c * 4;
+
+                        Operand value = context.LoadAttribute(Const(attr), Const(0), Const(v));
+
+                        if (attr == layerOutputAttr)
+                        {
+                            context.Copy(Attribute(AttributeConsts.Layer), value);
+                        }
+                        else
+                        {
+                            context.Copy(Attribute(attr), value);
+                            config.SetOutputUserAttribute(attrIndex);
+                        }
+
+                        config.SetInputUserAttribute(attrIndex, c);
+                    }
+                }
+
+                for (int c = 0; c < 4; c++)
+                {
+                    int attr = AttributeConsts.PositionX + c * 4;
+
+                    Operand value = context.LoadAttribute(Const(attr), Const(0), Const(v));
+
+                    context.Copy(Attribute(attr), value);
+                }
+
+                context.EmitVertex();
+            }
+
+            context.EndPrimitive();
+
+            var operations = context.GetOperations();
+            var cfg = ControlFlowGraph.Create(operations);
+            var function = new Function(cfg.Blocks, "main", false, 0, 0);
+
+            var sInfo = StructuredProgram.MakeStructuredProgram(new[] { function }, config);
+
+            var info = config.CreateProgramInfo();
+
+            return config.Options.TargetLanguage switch
+            {
+                TargetLanguage.Glsl => new ShaderProgram(info, TargetLanguage.Glsl, GlslGenerator.Generate(sInfo, config)),
+                TargetLanguage.Spirv => new ShaderProgram(info, TargetLanguage.Spirv, SpirvGenerator.Generate(sInfo, config)),
+                _ => throw new NotImplementedException(config.Options.TargetLanguage.ToString())
+            };
+        }
     }
 }

Ryujinx.Graphics.Vulkan/BufferHolder.cs

@@ -210,7 +210,10 @@ namespace Ryujinx.Graphics.Vulkan
                 }
             }
 
-            if (cbs != null && !(_buffer.HasCommandBufferDependency(cbs.Value) && _waitable.IsBufferRangeInUse(cbs.Value.CommandBufferIndex, offset, dataSize)))
+            if (cbs != null &&
+                _gd.PipelineInternal.RenderPassActive &&
+                !(_buffer.HasCommandBufferDependency(cbs.Value) &&
+                _waitable.IsBufferRangeInUse(cbs.Value.CommandBufferIndex, offset, dataSize)))
             {
                 // If the buffer hasn't been used on the command buffer yet, try to preload the data.
                 // This avoids ending and beginning render passes on each buffer data upload.

Ryujinx.Graphics.Vulkan/DescriptorSetUpdater.cs

@@ -130,6 +130,12 @@ namespace Ryujinx.Graphics.Vulkan
                 1f));
         }
 
+        public void Initialize()
+        {
+            Span<byte> dummyTextureData = stackalloc byte[4];
+            _dummyTexture.SetData(dummyTextureData);
+        }
+
         public void SetProgram(ShaderCollection program)
         {
             _program = program;

Ryujinx.Graphics.Vulkan/PipelineBase.cs

@@ -49,7 +49,6 @@ namespace Ryujinx.Graphics.Vulkan
         private Auto<DisposableFramebuffer> _framebuffer;
         private Auto<DisposableRenderPass> _renderPass;
         private int _writtenAttachmentCount;
-        private bool _renderPassActive;
 
         private readonly DescriptorSetUpdater _descriptorSetUpdater;
 
@@ -73,6 +72,7 @@ namespace Ryujinx.Graphics.Vulkan
         private PipelineColorBlendAttachmentState[] _storedBlend;
 
         public ulong DrawCount { get; private set; }
+        public bool RenderPassActive { get; private set; }
 
         public unsafe PipelineBase(VulkanRenderer gd, Device device)
         {
@@ -114,6 +114,8 @@ namespace Ryujinx.Graphics.Vulkan
 
         public void Initialize()
         {
+            _descriptorSetUpdater.Initialize();
+
             SupportBufferUpdater = new SupportBufferUpdater(Gd);
             SupportBufferUpdater.UpdateRenderScale(_renderScale, 0, SupportBuffer.RenderScaleMaxCount);
 
@@ -838,6 +840,11 @@ namespace Ryujinx.Graphics.Vulkan
             stages.CopyTo(_newState.Stages.AsSpan().Slice(0, stages.Length));
 
             SignalStateChange();
+
+            if (_program.IsCompute)
+            {
+                EndRenderPass();
+            }
         }
 
         public void Specialize<T>(in T data) where T : unmanaged
@@ -1451,7 +1458,7 @@ namespace Ryujinx.Graphics.Vulkan
 
         private unsafe void BeginRenderPass()
         {
-            if (!_renderPassActive)
+            if (!RenderPassActive)
             {
                 var renderArea = new Rect2D(null, new Extent2D(FramebufferParams.Width, FramebufferParams.Height));
                 var clearValue = new ClearValue();
@@ -1467,18 +1474,18 @@ namespace Ryujinx.Graphics.Vulkan
                 };
 
                 Gd.Api.CmdBeginRenderPass(CommandBuffer, renderPassBeginInfo, SubpassContents.Inline);
-                _renderPassActive = true;
+                RenderPassActive = true;
             }
         }
 
         public void EndRenderPass()
         {
-            if (_renderPassActive)
+            if (RenderPassActive)
             {
                 PauseTransformFeedbackInternal();
                 Gd.Api.CmdEndRenderPass(CommandBuffer);
                 SignalRenderPassEnd();
-                _renderPassActive = false;
+                RenderPassActive = false;
             }
         }
 

Ryujinx.Graphics.Vulkan/ShaderCollection.cs

@@ -19,6 +19,7 @@ namespace Ryujinx.Graphics.Vulkan
 
         public bool HasMinimalLayout { get; }
         public bool UsePushDescriptors { get; }
+        public bool IsCompute { get; }
 
         public uint Stages { get; }
 
@@ -47,7 +48,6 @@ namespace Ryujinx.Graphics.Vulkan
         private VulkanRenderer _gd;
         private Device _device;
         private bool _initialized;
-        private bool _isCompute;
 
         private ProgramPipelineState _state;
         private DisposableRenderPass _dummyRenderPass;
@@ -91,7 +91,7 @@ namespace Ryujinx.Graphics.Vulkan
 
                 if (shader.StageFlags == ShaderStageFlags.ShaderStageComputeBit)
                 {
-                    _isCompute = true;
+                    IsCompute = true;
                 }
 
                 internalShaders[i] = shader;
@@ -163,7 +163,7 @@ namespace Ryujinx.Graphics.Vulkan
 
             try
             {
-                if (_isCompute)
+                if (IsCompute)
                 {
                     CreateBackgroundComputePipeline();
                 }

Ryujinx.Graphics.Vulkan/VulkanRenderer.cs

@@ -396,6 +396,7 @@ namespace Ryujinx.Graphics.Vulkan
                 supportsFragmentShaderOrderingIntel: false,
                 supportsGeometryShaderPassthrough: Capabilities.SupportsGeometryShaderPassthrough,
                 supportsImageLoadFormatted: features2.Features.ShaderStorageImageReadWithoutFormat,
+                supportsLayerVertexTessellation: featuresVk12.ShaderOutputLayer,
                 supportsMismatchingViewFormat: true,
                 supportsCubemapView: !IsAmdGcn,
                 supportsNonConstantTextureOffset: false,

Ryujinx.HLE/HOS/Services/Am/AppletOE/ApplicationProxyService/ApplicationProxy/IApplicationFunctions.cs

@@ -26,8 +26,8 @@ namespace Ryujinx.HLE.HOS.Services.Am.AppletOE.ApplicationProxyService.Applicati
 {
     class IApplicationFunctions : IpcService
     {
-        private ulong _defaultSaveDataSize        = 200000000;
-        private ulong _defaultJournalSaveDataSize = 200000000;
+        private long _defaultSaveDataSize        = 200000000;
+        private long _defaultJournalSaveDataSize = 200000000;
 
         private KEvent _gpuErrorDetectedSystemEvent;
         private KEvent _friendInvitationStorageChannelEvent;
@@ -203,13 +203,13 @@ namespace Ryujinx.HLE.HOS.Services.Am.AppletOE.ApplicationProxyService.Applicati
         }
 
         [CommandHipc(25)] // 3.0.0+
-        // ExtendSaveData(u8 save_data_type, nn::account::Uid, u64 save_size, u64 journal_size) -> u64 result_code
+        // ExtendSaveData(u8 save_data_type, nn::account::Uid, s64 save_size, s64 journal_size) -> u64 result_code
         public ResultCode ExtendSaveData(ServiceCtx context)
         {
             SaveDataType saveDataType = (SaveDataType)context.RequestData.ReadUInt64();
             Uid          userId       = context.RequestData.ReadStruct<Uid>();
-            ulong        saveDataSize = context.RequestData.ReadUInt64();
-            ulong        journalSize  = context.RequestData.ReadUInt64();
+            long        saveDataSize  = context.RequestData.ReadInt64();
+            long        journalSize   = context.RequestData.ReadInt64();
 
             // NOTE: Service calls nn::fs::ExtendApplicationSaveData.
             //       Since LibHac currently doesn't support this method, we can stub it for now.
@@ -225,7 +225,7 @@ namespace Ryujinx.HLE.HOS.Services.Am.AppletOE.ApplicationProxyService.Applicati
         }
 
         [CommandHipc(26)] // 3.0.0+
-        // GetSaveDataSize(u8 save_data_type, nn::account::Uid) -> (u64 save_size, u64 journal_size)
+        // GetSaveDataSize(u8 save_data_type, nn::account::Uid) -> (s64 save_size, s64 journal_size)
         public ResultCode GetSaveDataSize(ServiceCtx context)
         {
             SaveDataType saveDataType = (SaveDataType)context.RequestData.ReadUInt64();
@@ -268,6 +268,23 @@ namespace Ryujinx.HLE.HOS.Services.Am.AppletOE.ApplicationProxyService.Applicati
             return ResultCode.Success;
         }
 
+        [CommandHipc(28)] // 11.0.0+
+        // GetSaveDataSizeMax() -> (s64 save_size_max, s64 journal_size_max)
+        public ResultCode GetSaveDataSizeMax(ServiceCtx context)
+        {
+            // NOTE: We are currently using a stub for GetSaveDataSize() which returns the default values.
+            //       For this method we shouldn't return anything lower than that, but since we aren't interacting
+            //       with fs to get the actual sizes, we return the default values here as well.
+            //       This also helps in case ExtendSaveData() has been executed and the default values were modified.
+
+            context.ResponseData.Write(_defaultSaveDataSize);
+            context.ResponseData.Write(_defaultJournalSaveDataSize);
+
+            Logger.Stub?.PrintStub(LogClass.ServiceAm);
+
+            return ResultCode.Success;
+        }
+
         [CommandHipc(30)]
         // BeginBlockingHomeButtonShortAndLongPressed()
         public ResultCode BeginBlockingHomeButtonShortAndLongPressed(ServiceCtx context)

Ryujinx.Memory/Tracking/MultiRegionHandle.cs

@@ -25,6 +25,7 @@ namespace Ryujinx.Memory.Tracking
 
         private int _sequenceNumber;
         private BitMap _sequenceNumberBitmap;
+        private BitMap _dirtyCheckedBitmap;
         private int _uncheckedHandles;
 
         public bool Dirty { get; private set; } = true;
@@ -36,6 +37,7 @@ namespace Ryujinx.Memory.Tracking
 
             _dirtyBitmap = new ConcurrentBitmap(_handles.Length, true);
             _sequenceNumberBitmap = new BitMap(_handles.Length);
+            _dirtyCheckedBitmap = new BitMap(_handles.Length);
 
             int i = 0;
 
@@ -246,16 +248,18 @@ namespace Ryujinx.Memory.Tracking
         }
 
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        private void ParseDirtyBits(long dirtyBits, long mask, int index, long[] seqMasks, ref int baseBit, ref int prevHandle, ref ulong rgStart, ref ulong rgSize, Action<ulong, ulong> modifiedAction)
+        private void ParseDirtyBits(long dirtyBits, long mask, int index, long[] seqMasks, long[] checkMasks, ref int baseBit, ref int prevHandle, ref ulong rgStart, ref ulong rgSize, Action<ulong, ulong> modifiedAction)
         {
             long seqMask = mask & ~seqMasks[index];
+            long checkMask = (~dirtyBits) & seqMask;
             dirtyBits &= seqMask;
 
             while (dirtyBits != 0)
             {
                 int bit = BitOperations.TrailingZeroCount(dirtyBits);
+                long bitValue = 1L << bit;
 
-                dirtyBits &= ~(1L << bit);
+                dirtyBits &= ~bitValue;
 
                 int handleIndex = baseBit + bit;
 
@@ -273,11 +277,14 @@ namespace Ryujinx.Memory.Tracking
                 }
 
                 rgSize += handle.Size;
-                handle.Reprotect();
+                handle.Reprotect(false, (checkMasks[index] & bitValue) == 0);
+
+                checkMasks[index] &= ~bitValue;
 
                 prevHandle = handleIndex;
             }
 
+            checkMasks[index] |= checkMask;
             seqMasks[index] |= mask;
             _uncheckedHandles -= BitOperations.PopCount((ulong)seqMask);
 
@@ -328,6 +335,7 @@ namespace Ryujinx.Memory.Tracking
             ulong rgSize = 0;
 
             long[] seqMasks = _sequenceNumberBitmap.Masks;
+            long[] checkedMasks = _dirtyCheckedBitmap.Masks;
             long[] masks = _dirtyBitmap.Masks;
 
             int startIndex = startHandle >> ConcurrentBitmap.IntShift;
@@ -345,20 +353,20 @@ namespace Ryujinx.Memory.Tracking
 
             if (startIndex == endIndex)
             {
-                ParseDirtyBits(startValue, startMask & endMask, startIndex, seqMasks, ref baseBit, ref prevHandle, ref rgStart, ref rgSize, modifiedAction);
+                ParseDirtyBits(startValue, startMask & endMask, startIndex, seqMasks, checkedMasks, ref baseBit, ref prevHandle, ref rgStart, ref rgSize, modifiedAction);
             }
             else
             {
-                ParseDirtyBits(startValue, startMask, startIndex, seqMasks, ref baseBit, ref prevHandle, ref rgStart, ref rgSize, modifiedAction);
+                ParseDirtyBits(startValue, startMask, startIndex, seqMasks, checkedMasks, ref baseBit, ref prevHandle, ref rgStart, ref rgSize, modifiedAction);
 
                 for (int i = startIndex + 1; i < endIndex; i++)
                 {
-                    ParseDirtyBits(Volatile.Read(ref masks[i]), -1L, i, seqMasks, ref baseBit, ref prevHandle, ref rgStart, ref rgSize, modifiedAction);
+                    ParseDirtyBits(Volatile.Read(ref masks[i]), -1L, i, seqMasks, checkedMasks, ref baseBit, ref prevHandle, ref rgStart, ref rgSize, modifiedAction);
                 }
 
                 long endValue = Volatile.Read(ref masks[endIndex]);
 
-                ParseDirtyBits(endValue, endMask, endIndex, seqMasks, ref baseBit, ref prevHandle, ref rgStart, ref rgSize, modifiedAction);
+                ParseDirtyBits(endValue, endMask, endIndex, seqMasks, checkedMasks, ref baseBit, ref prevHandle, ref rgStart, ref rgSize, modifiedAction);
             }
 
             if (rgSize != 0)

Ryujinx.Memory/Tracking/RegionHandle.cs

@@ -263,15 +263,15 @@ namespace Ryujinx.Memory.Tracking
         /// </summary>
         public void ForceDirty()
         {
-            _checkCount++;
-
             Dirty = true;
         }
 
         /// <summary>
         /// Consume the dirty flag for this handle, and reprotect so it can be set on the next write.
         /// </summary>
-        public void Reprotect(bool asDirty = false)
+        /// <param name="asDirty">True if the handle should be reprotected as dirty, rather than have it cleared</param>
+        /// <param name="consecutiveCheck">True if this reprotect is the result of consecutive dirty checks</param>
+        public void Reprotect(bool asDirty, bool consecutiveCheck = false)
         {
             if (_volatile) return;
 
@@ -296,7 +296,7 @@ namespace Ryujinx.Memory.Tracking
             }
             else if (!asDirty)
             {
-                if (_checkCount > 0 && _checkCount < CheckCountForInfrequent)
+                if (consecutiveCheck || (_checkCount > 0 && _checkCount < CheckCountForInfrequent))
                 {
                     if (++_volatileCount >= VolatileThreshold && _preAction == null)
                     {
@@ -313,6 +313,15 @@ namespace Ryujinx.Memory.Tracking
             }
         }
 
+        /// <summary>
+        /// Consume the dirty flag for this handle, and reprotect so it can be set on the next write.
+        /// </summary>
+        /// <param name="asDirty">True if the handle should be reprotected as dirty, rather than have it cleared</param>
+        public void Reprotect(bool asDirty = false)
+        {
+            Reprotect(asDirty, false);
+        }
+
         /// <summary>
         /// Register an action to perform when the tracked region is read or written.
         /// The action is automatically removed after it runs.

Ryujinx/Program.cs

@@ -29,7 +29,11 @@ namespace Ryujinx
 
         public static string ConfigurationPath { get; set; }
 
-        [DllImport("libX11")]
+        public static string CommandLineProfile { get; set; }
+
+        private const string X11LibraryName = "libX11";
+
+        [DllImport(X11LibraryName)]
         private extern static int XInitThreads();
 
         [DllImport("user32.dll", SetLastError = true)]
@@ -37,6 +41,30 @@ namespace Ryujinx
 
         private const uint MB_ICONWARNING = 0x30;
 
+        static Program()
+        {
+            if (OperatingSystem.IsLinux())
+            {
+                NativeLibrary.SetDllImportResolver(typeof(Program).Assembly, (name, assembly, path) =>
+                {
+                    if (name != X11LibraryName)
+                    {
+                        return IntPtr.Zero;
+                    }
+
+                    if (!NativeLibrary.TryLoad("libX11.so.6", assembly, path, out IntPtr result))
+                    {
+                        if (!NativeLibrary.TryLoad("libX11.so", assembly, path, out result))
+                        {
+                            return IntPtr.Zero;
+                        }
+                    }
+
+                    return result;
+                });
+            }
+        }
+
         static void Main(string[] args)
         {
             Version = ReleaseInformations.GetVersion();

Ryujinx/Ryujinx.csproj

@@ -25,7 +25,7 @@
     <PackageReference Include="Ryujinx.Graphics.Nvdec.Dependencies" Version="5.0.1-build10" Condition="'$(RuntimeIdentifier)' != 'linux-x64' AND '$(RuntimeIdentifier)' != 'osx-x64'" />
     <PackageReference Include="Ryujinx.Audio.OpenAL.Dependencies" Version="1.21.0.1" Condition="'$(RuntimeIdentifier)' != 'linux-x64' AND '$(RuntimeIdentifier)' != 'osx-x64'" />
     <PackageReference Include="OpenTK.Graphics" Version="4.7.2" />
-    <PackageReference Include="SPB" Version="0.0.4-build27" />
+    <PackageReference Include="SPB" Version="0.0.4-build28" />
     <PackageReference Include="SharpZipLib" Version="1.3.3" />
     <PackageReference Include="SixLabors.ImageSharp" Version="1.0.4" />
   </ItemGroup>

Ryujinx/Ui/MainWindow.cs

@@ -148,10 +148,10 @@ namespace Ryujinx.Ui
 
             // Apply custom theme if needed.
             ThemeHelper.ApplyTheme();
-
+            Gdk.Monitor monitor = Display.GetMonitor(0);
             // Sets overridden fields.
-            int monitorWidth  = Display.PrimaryMonitor.Geometry.Width  * Display.PrimaryMonitor.ScaleFactor;
-            int monitorHeight = Display.PrimaryMonitor.Geometry.Height * Display.PrimaryMonitor.ScaleFactor;
+            int monitorWidth  = monitor.Geometry.Width  * monitor.ScaleFactor;
+            int monitorHeight = monitor.Geometry.Height * monitor.ScaleFactor;
 
             DefaultWidth  = monitorWidth  < 1280 ? monitorWidth  : 1280;
             DefaultHeight = monitorHeight < 760  ? monitorHeight : 760;

Ryujinx/Ui/Windows/SettingsWindow.glade

@@ -1749,11 +1749,11 @@
                                 <property name="orientation">vertical</property>
                                 <child>
                                   <object class="GtkCheckButton" id="_expandRamToggle">
-                                    <property name="label" translatable="yes">Expand DRAM Size to 6GiB</property>
+                                    <property name="label" translatable="yes">Use alternative memory layout (Developers)</property>
                                     <property name="visible">True</property>
                                     <property name="can-focus">True</property>
                                     <property name="receives-default">False</property>
-                                    <property name="tooltip-text" translatable="yes">Increases the amount of memory on the emulated system from 4GiB to 6GiB.&#13;This is only useful for higher-resolution texture packs or 4k resolution mods. Does NOT improve performance.&#13;Leave OFF if unsure.</property>
+                                    <property name="tooltip-text" translatable="yes">Utilizes an alternative MemoryMode layout to mimic a Switch development model.&#13;This is only useful for higher-resolution texture packs or 4k resolution mods. Does NOT improve performance.&#13;Leave OFF if unsure.</property>
                                     <property name="halign">start</property>
                                     <property name="margin-top">5</property>
                                     <property name="margin-bottom">5</property>