Workaround for Intel FrontFacing built-in variable bug (#2540)

Ryujinx.Graphics.GAL/Capabilities.cs

@@ -2,37 +2,44 @@ namespace Ryujinx.Graphics.GAL
 {
     public struct Capabilities
     {
-        public bool SupportsAstcCompression          { get; }
+        public bool HasFrontFacingBug { get; }
-        public bool SupportsImageLoadFormatted       { get; }
+        public bool HasVectorIndexingBug { get; }
-        public bool SupportsMismatchingViewFormat    { get; }
-        public bool SupportsNonConstantTextureOffset { get; }
-        public bool SupportsTextureShadowLod         { get; }
-        public bool SupportsViewportSwizzle          { get; }
 
-        public int   MaximumComputeSharedMemorySize { get; }
+        public bool SupportsAstcCompression { get; }
-        public float MaximumSupportedAnisotropy     { get; }
+        public bool SupportsImageLoadFormatted { get; }
-        public int   StorageBufferOffsetAlignment   { get; }
+        public bool SupportsMismatchingViewFormat { get; }
+        public bool SupportsNonConstantTextureOffset { get; }
+        public bool SupportsTextureShadowLod { get; }
+        public bool SupportsViewportSwizzle { get; }
+
+        public int MaximumComputeSharedMemorySize { get; }
+        public float MaximumSupportedAnisotropy { get; }
+        public int StorageBufferOffsetAlignment { get; }
 
         public Capabilities(
-            bool  supportsAstcCompression,
+            bool hasFrontFacingBug,
-            bool  supportsImageLoadFormatted,
+            bool hasVectorIndexingBug,
-            bool  supportsMismatchingViewFormat,
+            bool supportsAstcCompression,
-            bool  supportsNonConstantTextureOffset,
+            bool supportsImageLoadFormatted,
-            bool  supportsTextureShadowLod,
+            bool supportsMismatchingViewFormat,
-            bool  supportsViewportSwizzle,
+            bool supportsNonConstantTextureOffset,
-            int   maximumComputeSharedMemorySize,
+            bool supportsTextureShadowLod,
+            bool supportsViewportSwizzle,
+            int maximumComputeSharedMemorySize,
             float maximumSupportedAnisotropy,
-            int   storageBufferOffsetAlignment)
+            int storageBufferOffsetAlignment)
         {
-            SupportsAstcCompression          = supportsAstcCompression;
+            HasFrontFacingBug = hasFrontFacingBug;
-            SupportsImageLoadFormatted       = supportsImageLoadFormatted;
+            HasVectorIndexingBug = hasVectorIndexingBug;
-            SupportsMismatchingViewFormat    = supportsMismatchingViewFormat;
+            SupportsAstcCompression = supportsAstcCompression;
+            SupportsImageLoadFormatted = supportsImageLoadFormatted;
+            SupportsMismatchingViewFormat = supportsMismatchingViewFormat;
             SupportsNonConstantTextureOffset = supportsNonConstantTextureOffset;
-            SupportsTextureShadowLod         = supportsTextureShadowLod;
+            SupportsTextureShadowLod = supportsTextureShadowLod;
-            SupportsViewportSwizzle          = supportsViewportSwizzle;
+            SupportsViewportSwizzle = supportsViewportSwizzle;
-            MaximumComputeSharedMemorySize   = maximumComputeSharedMemorySize;
+            MaximumComputeSharedMemorySize = maximumComputeSharedMemorySize;
-            MaximumSupportedAnisotropy       = maximumSupportedAnisotropy;
+            MaximumSupportedAnisotropy = maximumSupportedAnisotropy;
-            StorageBufferOffsetAlignment     = storageBufferOffsetAlignment;
+            StorageBufferOffsetAlignment = storageBufferOffsetAlignment;
         }
     }
 }

Ryujinx.Graphics.Gpu/Shader/CachedGpuAccessor.cs

@@ -9,7 +9,6 @@ namespace Ryujinx.Graphics.Gpu.Shader
 {
     class CachedGpuAccessor : TextureDescriptorCapableGpuAccessor, IGpuAccessor
     {
-        private readonly GpuContext _context;
         private readonly ReadOnlyMemory<byte> _data;
         private readonly ReadOnlyMemory<byte> _cb1Data;
         private readonly GuestGpuAccessorHeader _header;
@@ -28,9 +27,8 @@ namespace Ryujinx.Graphics.Gpu.Shader
             ReadOnlyMemory<byte> data,
             ReadOnlyMemory<byte> cb1Data,
             GuestGpuAccessorHeader header,
-            Dictionary<int, GuestTextureDescriptor> guestTextureDescriptors)
+            IReadOnlyDictionary<int, GuestTextureDescriptor> guestTextureDescriptors) : base(context)
         {
-            _context = context;
             _data = data;
             _cb1Data = cb1Data;
             _header = header;
@@ -136,24 +134,6 @@ namespace Ryujinx.Graphics.Gpu.Shader
             return _header.PrimitiveTopology;
         }
 
-        /// <summary>
-        /// Queries host storage buffer alignment required.
-        /// </summary>
-        /// <returns>Host storage buffer alignment in bytes</returns>
-        public int QueryStorageBufferOffsetAlignment() => _context.Capabilities.StorageBufferOffsetAlignment;
-
-        /// <summary>
-        /// Queries host support for readable images without a explicit format declaration on the shader.
-        /// </summary>
-        /// <returns>True if formatted image load is supported, false otherwise</returns>
-        public bool QuerySupportsImageLoadFormatted() => _context.Capabilities.SupportsImageLoadFormatted;
-
-        /// <summary>
-        /// Queries host GPU non-constant texture offset support.
-        /// </summary>
-        /// <returns>True if the GPU and driver supports non-constant texture offsets, false otherwise</returns>
-        public bool QuerySupportsNonConstantTextureOffset() => _context.Capabilities.SupportsNonConstantTextureOffset;
-
         /// <summary>
         /// Gets the texture descriptor for a given texture on the pool.
         /// </summary>

Ryujinx.Graphics.Gpu/Shader/GpuAccessor.cs

@@ -7,9 +7,8 @@ namespace Ryujinx.Graphics.Gpu.Shader
     /// <summary>
     /// Represents a GPU state and memory accessor.
     /// </summary>
-    class GpuAccessor : TextureDescriptorCapableGpuAccessor, IGpuAccessor
+    class GpuAccessor : TextureDescriptorCapableGpuAccessor
     {
-        private readonly GpuContext _context;
         private readonly GpuChannel _channel;
         private readonly GpuAccessorState _state;
         private readonly int _stageIndex;
@@ -29,9 +28,8 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// <param name="channel">GPU channel</param>
         /// <param name="state">Current GPU state</param>
         /// <param name="stageIndex">Graphics shader stage index (0 = Vertex, 4 = Fragment)</param>
-        public GpuAccessor(GpuContext context, GpuChannel channel, GpuAccessorState state, int stageIndex)
+        public GpuAccessor(GpuContext context, GpuChannel channel, GpuAccessorState state, int stageIndex) : base(context)
         {
-            _context = context;
             _channel = channel;
             _state = state;
             _stageIndex = stageIndex;
@@ -56,9 +54,8 @@ namespace Ryujinx.Graphics.Gpu.Shader
             int localSizeY,
             int localSizeZ,
             int localMemorySize,
-            int sharedMemorySize)
+            int sharedMemorySize) : base(context)
         {
-            _context = context;
             _channel = channel;
             _state = state;
             _compute = true;
@@ -182,30 +179,6 @@ namespace Ryujinx.Graphics.Gpu.Shader
             };
         }
 
-        /// <summary>
-        /// Queries host storage buffer alignment required.
-        /// </summary>
-        /// <returns>Host storage buffer alignment in bytes</returns>
-        public int QueryStorageBufferOffsetAlignment() => _context.Capabilities.StorageBufferOffsetAlignment;
-
-        /// <summary>
-        /// Queries host support for readable images without a explicit format declaration on the shader.
-        /// </summary>
-        /// <returns>True if formatted image load is supported, false otherwise</returns>
-        public bool QuerySupportsImageLoadFormatted() => _context.Capabilities.SupportsImageLoadFormatted;
-
-        /// <summary>
-        /// Queries host GPU non-constant texture offset support.
-        /// </summary>
-        /// <returns>True if the GPU and driver supports non-constant texture offsets, false otherwise</returns>
-        public bool QuerySupportsNonConstantTextureOffset() => _context.Capabilities.SupportsNonConstantTextureOffset;
-
-        /// <summary>
-        /// Queries host GPU texture shadow LOD support.
-        /// </summary>
-        /// <returns>True if the GPU and driver supports texture shadow LOD, false otherwise</returns>
-        public bool QuerySupportsTextureShadowLod() => _context.Capabilities.SupportsTextureShadowLod;
-
         /// <summary>
         /// Gets the texture descriptor for a given texture on the pool.
         /// </summary>

Ryujinx.Graphics.Gpu/Shader/GpuAccessorBase.cs

@@ -0,0 +1,55 @@
+namespace Ryujinx.Graphics.Gpu.Shader
+{
+    /// <summary>
+    /// Represents a GPU state and memory accessor.
+    /// </summary>
+    class GpuAccessorBase
+    {
+        private readonly GpuContext _context;
+
+        /// <summary>
+        /// Creates a new instance of the GPU state accessor.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        public GpuAccessorBase(GpuContext context)
+        {
+            _context = context;
+        }
+
+        /// <summary>
+        /// Queries host about the presence of the FrontFacing built-in variable bug.
+        /// </summary>
+        /// <returns>True if the bug is present on the host device used, false otherwise</returns>
+        public bool QueryHostHasFrontFacingBug() => _context.Capabilities.HasFrontFacingBug;
+
+        /// <summary>
+        /// Queries host about the presence of the vector indexing bug.
+        /// </summary>
+        /// <returns>True if the bug is present on the host device used, false otherwise</returns>
+        public bool QueryHostHasVectorIndexingBug() => _context.Capabilities.HasVectorIndexingBug;
+
+        /// <summary>
+        /// Queries host storage buffer alignment required.
+        /// </summary>
+        /// <returns>Host storage buffer alignment in bytes</returns>
+        public int QueryHostStorageBufferOffsetAlignment() => _context.Capabilities.StorageBufferOffsetAlignment;
+
+        /// <summary>
+        /// Queries host support for readable images without a explicit format declaration on the shader.
+        /// </summary>
+        /// <returns>True if formatted image load is supported, false otherwise</returns>
+        public bool QueryHostSupportsImageLoadFormatted() => _context.Capabilities.SupportsImageLoadFormatted;
+
+        /// <summary>
+        /// Queries host GPU non-constant texture offset support.
+        /// </summary>
+        /// <returns>True if the GPU and driver supports non-constant texture offsets, false otherwise</returns>
+        public bool QueryHostSupportsNonConstantTextureOffset() => _context.Capabilities.SupportsNonConstantTextureOffset;
+
+        /// <summary>
+        /// Queries host GPU texture shadow LOD support.
+        /// </summary>
+        /// <returns>True if the GPU and driver supports texture shadow LOD, false otherwise</returns>
+        public bool QueryHostSupportsTextureShadowLod() => _context.Capabilities.SupportsTextureShadowLod;
+    }
+}

Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs

@@ -38,7 +38,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// <summary>
         /// Version of the codegen (to be changed when codegen or guest format change).
         /// </summary>
-        private const ulong ShaderCodeGenVersion = 2538;
+        private const ulong ShaderCodeGenVersion = 2540;
 
         // Progress reporting helpers
         private volatile int _shaderCount;

Ryujinx.Graphics.Gpu/Shader/TextureDescriptorCapableGpuAccessor.cs

@@ -4,8 +4,12 @@ using Ryujinx.Graphics.Shader;
 
 namespace Ryujinx.Graphics.Gpu.Shader
 {
-    abstract class TextureDescriptorCapableGpuAccessor : IGpuAccessor
+    abstract class TextureDescriptorCapableGpuAccessor : GpuAccessorBase, IGpuAccessor
     {
+        public TextureDescriptorCapableGpuAccessor(GpuContext context) : base(context)
+        {
+        }
+
         public abstract T MemoryRead<T>(ulong address) where T : unmanaged;
 
         public abstract ITextureDescriptor GetTextureDescriptor(int handle, int cbufSlot);

Ryujinx.Graphics.OpenGL/Renderer.cs

@@ -99,6 +99,8 @@ namespace Ryujinx.Graphics.OpenGL
         public Capabilities GetCapabilities()
         {
             return new Capabilities(
+                HwCapabilities.Vendor == HwCapabilities.GpuVendor.IntelWindows,
+                HwCapabilities.Vendor == HwCapabilities.GpuVendor.AmdWindows,
                 HwCapabilities.SupportsAstcCompression,
                 HwCapabilities.SupportsImageLoadFormatted,
                 HwCapabilities.SupportsMismatchingViewFormat,

Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenMemory.cs

@@ -157,15 +157,18 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
 
             offsetExpr = Enclose(offsetExpr, src2, Instruction.ShiftRightS32, isLhs: true);
 
+            var config = context.Config;
+            bool indexElement = !config.GpuAccessor.QueryHostHasVectorIndexingBug();
+
             if (src1 is AstOperand oper && oper.Type == OperandType.Constant)
             {
-                bool cbIndexable = context.Config.UsedFeatures.HasFlag(Translation.FeatureFlags.CbIndexing);
+                bool cbIndexable = config.UsedFeatures.HasFlag(Translation.FeatureFlags.CbIndexing);
-                return OperandManager.GetConstantBufferName(oper.Value, offsetExpr, context.Config.Stage, cbIndexable);
+                return OperandManager.GetConstantBufferName(oper.Value, offsetExpr, config.Stage, cbIndexable, indexElement);
             }
             else
             {
                 string slotExpr = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 0));
-                return OperandManager.GetConstantBufferName(slotExpr, offsetExpr, context.Config.Stage);
+                return OperandManager.GetConstantBufferName(slotExpr, offsetExpr, config.Stage, indexElement);
             }
         }
 
@@ -314,7 +317,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
 
             // 2D Array and Cube shadow samplers with LOD level or bias requires an extension.
             // If the extension is not supported, just remove the LOD parameter.
-            if (isArray && isShadow && (is2D || isCube) && !context.Config.GpuAccessor.QuerySupportsTextureShadowLod())
+            if (isArray && isShadow && (is2D || isCube) && !context.Config.GpuAccessor.QueryHostSupportsTextureShadowLod())
             {
                 hasLodBias = false;
                 hasLodLevel = false;
@@ -322,7 +325,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
 
             // Cube shadow samplers with LOD level requires an extension.
             // If the extension is not supported, just remove the LOD level parameter.
-            if (isShadow && isCube && !context.Config.GpuAccessor.QuerySupportsTextureShadowLod())
+            if (isShadow && isCube && !context.Config.GpuAccessor.QueryHostSupportsTextureShadowLod())
             {
                 hasLodLevel = false;
             }

Ryujinx.Graphics.Shader/CodeGen/Glsl/OperandManager.cs

@@ -117,8 +117,13 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
             return $"{GetUbName(stage, slot, cbIndexable)}[{offset >> 2}].{GetSwizzleMask(offset & 3)}";
         }
 
-        private static string GetVec4Indexed(string vectorName, string indexExpr)
+        private static string GetVec4Indexed(string vectorName, string indexExpr, bool indexElement)
         {
+            if (indexElement)
+            {
+                return $"{vectorName}[{indexExpr}]";
+            }
+
             string result = $"{vectorName}.x";
             for (int i = 1; i < 4; i++)
             {
@@ -127,14 +132,14 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
             return $"({result})";
         }
 
-        public static string GetConstantBufferName(int slot, string offsetExpr, ShaderStage stage, bool cbIndexable)
+        public static string GetConstantBufferName(int slot, string offsetExpr, ShaderStage stage, bool cbIndexable, bool indexElement)
         {
-            return GetVec4Indexed(GetUbName(stage, slot, cbIndexable) + $"[{offsetExpr} >> 2]", offsetExpr + " & 3");
+            return GetVec4Indexed(GetUbName(stage, slot, cbIndexable) + $"[{offsetExpr} >> 2]", offsetExpr + " & 3", indexElement);
         }
 
-        public static string GetConstantBufferName(string slotExpr, string offsetExpr, ShaderStage stage)
+        public static string GetConstantBufferName(string slotExpr, string offsetExpr, ShaderStage stage, bool indexElement)
         {
-            return GetVec4Indexed(GetUbName(stage, slotExpr) + $"[{offsetExpr} >> 2]", offsetExpr + " & 3");
+            return GetVec4Indexed(GetUbName(stage, slotExpr) + $"[{offsetExpr} >> 2]", offsetExpr + " & 3", indexElement);
         }
 
         public static string GetOutAttributeName(AstOperand attr, ShaderConfig config)
@@ -198,6 +203,15 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                             case AttributeConsts.PositionY: return $"(gl_FragCoord.y / {DefaultNames.SupportBlockRenderScaleName}[0])";
                             case AttributeConsts.PositionZ: return "gl_FragCoord.z";
                             case AttributeConsts.PositionW: return "gl_FragCoord.w";
+
+                            case AttributeConsts.FrontFacing:
+                                if (config.GpuAccessor.QueryHostHasFrontFacingBug())
+                                {
+                                    // This is required for Intel on Windows, gl_FrontFacing sometimes returns incorrect
+                                    // (flipped) values. Doing this seems to fix it.
+                                    return "(-floatBitsToInt(float(gl_FrontFacing)) < 0)";
+                                }
+                                break;
                         }
                     }
 

Ryujinx.Graphics.Shader/IGpuAccessor.cs

@@ -49,6 +49,36 @@
             return 0;
         }
 
+        bool QueryHostHasFrontFacingBug()
+        {
+            return false;
+        }
+
+        bool QueryHostHasVectorIndexingBug()
+        {
+            return false;
+        }
+
+        int QueryHostStorageBufferOffsetAlignment()
+        {
+            return 16;
+        }
+
+        bool QueryHostSupportsImageLoadFormatted()
+        {
+            return true;
+        }
+
+        bool QueryHostSupportsNonConstantTextureOffset()
+        {
+            return true;
+        }
+
+        bool QueryHostSupportsTextureShadowLod()
+        {
+            return true;
+        }
+
         bool QueryIsTextureBuffer(int handle, int cbufSlot = -1)
         {
             return false;
@@ -64,26 +94,6 @@
             return InputTopology.Points;
         }
 
-        int QueryStorageBufferOffsetAlignment()
-        {
-            return 16;
-        }
-
-        bool QuerySupportsImageLoadFormatted()
-        {
-            return true;
-        }
-
-        bool QuerySupportsNonConstantTextureOffset()
-        {
-            return true;
-        }
-
-        bool QuerySupportsTextureShadowLod()
-        {
-            return true;
-        }
-
         TextureFormat QueryTextureFormat(int handle, int cbufSlot = -1)
         {
             return TextureFormat.R8G8B8A8Unorm;

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

@@ -71,7 +71,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
 
                 Operand baseAddrTrunc = Local();
 
-                Operand alignMask = Const(-config.GpuAccessor.QueryStorageBufferOffsetAlignment());
+                Operand alignMask = Const(-config.GpuAccessor.QueryHostStorageBufferOffsetAlignment());
 
                 Operation andOp = new Operation(Instruction.BitwiseAnd, baseAddrTrunc, baseAddrLow, alignMask);
 
@@ -142,7 +142,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
 
                 Operand baseAddrTrunc = Local();
 
-                Operand alignMask = Const(-config.GpuAccessor.QueryStorageBufferOffsetAlignment());
+                Operand alignMask = Const(-config.GpuAccessor.QueryHostStorageBufferOffsetAlignment());
 
                 Operation andOp = new Operation(Instruction.BitwiseAnd, baseAddrTrunc, baseAddrLow, alignMask);
 

Ryujinx.Graphics.Shader/Translation/Rewriter.cs

@@ -93,7 +93,7 @@ namespace Ryujinx.Graphics.Shader.Translation
                 sbSlot        = PrependOperation(Instruction.ConditionalSelect, inRange, Const(slot), sbSlot);
             }
 
-            Operand alignMask = Const(-config.GpuAccessor.QueryStorageBufferOffsetAlignment());
+            Operand alignMask = Const(-config.GpuAccessor.QueryHostStorageBufferOffsetAlignment());
 
             Operand baseAddrTrunc = PrependOperation(Instruction.BitwiseAnd,    sbBaseAddrLow, alignMask);
             Operand byteOffset    = PrependOperation(Instruction.Subtract,      addrLow, baseAddrTrunc);
@@ -145,7 +145,7 @@ namespace Ryujinx.Graphics.Shader.Translation
             bool hasOffset  = (texOp.Flags & TextureFlags.Offset)  != 0;
             bool hasOffsets = (texOp.Flags & TextureFlags.Offsets) != 0;
 
-            bool hasInvalidOffset = (hasOffset || hasOffsets) && !config.GpuAccessor.QuerySupportsNonConstantTextureOffset();
+            bool hasInvalidOffset = (hasOffset || hasOffsets) && !config.GpuAccessor.QueryHostSupportsNonConstantTextureOffset();
 
             bool isBindless = (texOp.Flags & TextureFlags.Bindless) != 0;
 

Ryujinx.Graphics.Shader/Translation/ShaderConfig.cs

@@ -145,7 +145,7 @@ namespace Ryujinx.Graphics.Shader.Translation
         {
             // When the formatted load extension is supported, we don't need to
             // specify a format, we can just declare it without a format and the GPU will handle it.
-            if (GpuAccessor.QuerySupportsImageLoadFormatted())
+            if (GpuAccessor.QueryHostSupportsImageLoadFormatted())
             {
                 return TextureFormat.Unknown;
             }