Move gl_Layer from vertex to geometry if GPU does not support it on vertex (#3866)

* Move gl_Layer from vertex to geometry if GPU does not support it on vertex

* Shader cache version bump

* PR feedback

ARMeilleure/Common/BitUtils.cs

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

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) ||
+            if (_pruneCaches)
-                result.EndGpuAddress < gpuVa + size ||
-                result.UnmappedSequence != result.Buffer.UnmappedSequence)
             {
-                ulong address = TranslateAndCreateBuffer(memoryManager, gpuVa, size);
+                Prune();
-                result = new BufferCacheEntry(address, gpuVa, GetBuffer(address, size));
-
-                _modifiedCache[gpuVa] = result;
             }
 
-            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/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/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 = 3848;
+        private const uint CodeGenVersion = 3866;
 
         private const string SharedTocFileName = "shared.toc";
         private const string SharedDataFileName = "shared.data";

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/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/ShaderCache.cs

@@ -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()));
                 }
             }
 

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[]
+        public static ReadOnlySpan<byte> SizeGroupLookup => new byte[] { 0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 3 };
-        {
-            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 ReadOnlySpan<byte> Vp9Cat1Prob => new byte[] { 159 };
-        public static readonly byte[] Vp9Cat2Prob = new byte[] { 165, 145 };
+        public static ReadOnlySpan<byte> Vp9Cat2Prob => new byte[] { 165, 145 };
-        public static readonly byte[] Vp9Cat3Prob = new byte[] { 173, 148, 140 };
+        public static ReadOnlySpan<byte> Vp9Cat3Prob => new byte[] { 173, 148, 140 };
-        public static readonly byte[] Vp9Cat4Prob = new byte[] { 176, 155, 140, 135 };
+        public static ReadOnlySpan<byte> Vp9Cat4Prob => new byte[] { 176, 155, 140, 135 };
-        public static readonly byte[] Vp9Cat5Prob = new byte[] { 180, 157, 141, 134, 130 };
+        public static ReadOnlySpan<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> 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/IGpuAccessor.cs

@@ -258,6 +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/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(
+            var info = config.CreateProgramInfo();
-                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);
 
             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 long _defaultSaveDataSize        = 200000000;
-        private ulong _defaultJournalSaveDataSize = 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();
+            long        saveDataSize  = context.RequestData.ReadInt64();
-            ulong        journalSize  = context.RequestData.ReadUInt64();
+            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.