Support non-constant texture offsets on non-NVIDIA gpus

11 files changed, 319 insertions, 33 deletions

diff --git a/Ryujinx.Graphics.GAL/Capabilities.cs b/Ryujinx.Graphics.GAL/Capabilities.cs
index 8a2dd8b6..f06c3c4b 100644
--- a/Ryujinx.Graphics.GAL/Capabilities.cs
+++ b/Ryujinx.Graphics.GAL/Capabilities.cs
@@ -2,7 +2,8 @@ namespace Ryujinx.Graphics.GAL
 {
     public struct Capabilities
     {
-        public bool SupportsAstcCompression { get; }
+        public bool SupportsAstcCompression          { get; }
+        public bool SupportsNonConstantTextureOffset { get; }
 
         public int MaximumViewportDimensions      { get; }
         public int MaximumComputeSharedMemorySize { get; }
@@ -10,14 +11,16 @@ namespace Ryujinx.Graphics.GAL
 
         public Capabilities(
             bool supportsAstcCompression,
+            bool supportsNonConstantTextureOffset,
             int  maximumViewportDimensions,
             int  maximumComputeSharedMemorySize,
             int  storageBufferOffsetAlignment)
         {
-            SupportsAstcCompression        = supportsAstcCompression;
-            MaximumViewportDimensions      = maximumViewportDimensions;
-            MaximumComputeSharedMemorySize = maximumComputeSharedMemorySize;
-            StorageBufferOffsetAlignment   = storageBufferOffsetAlignment;
+            SupportsAstcCompression          = supportsAstcCompression;
+            SupportsNonConstantTextureOffset = supportsNonConstantTextureOffset;
+            MaximumViewportDimensions        = maximumViewportDimensions;
+            MaximumComputeSharedMemorySize   = maximumComputeSharedMemorySize;
+            StorageBufferOffsetAlignment     = storageBufferOffsetAlignment;
         }
     }
 }
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs b/Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs
index d54c1942..648e073c 100644
--- a/Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs
+++ b/Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs
@@ -355,7 +355,8 @@ namespace Ryujinx.Graphics.Gpu.Shader
             return new ShaderCapabilities(
                 _context.Capabilities.MaximumViewportDimensions,
                 _context.Capabilities.MaximumComputeSharedMemorySize,
-                _context.Capabilities.StorageBufferOffsetAlignment);
+                _context.Capabilities.StorageBufferOffsetAlignment,
+                _context.Capabilities.SupportsNonConstantTextureOffset);
         }
     }
 }
\ No newline at end of file
diff --git a/Ryujinx.Graphics.OpenGL/HwCapabilities.cs b/Ryujinx.Graphics.OpenGL/HwCapabilities.cs
index 7524dc1d..dc147484 100644
--- a/Ryujinx.Graphics.OpenGL/HwCapabilities.cs
+++ b/Ryujinx.Graphics.OpenGL/HwCapabilities.cs
@@ -11,11 +11,14 @@ namespace Ryujinx.Graphics.OpenGL
         private static Lazy<int> _maximumComputeSharedMemorySize = new Lazy<int>(() => GetLimit(All.MaxComputeSharedMemorySize));
         private static Lazy<int> _storageBufferOffsetAlignment   = new Lazy<int>(() => GetLimit(All.ShaderStorageBufferOffsetAlignment));
 
-        public static bool SupportsAstcCompression => _supportsAstcCompression.Value;
+        private static Lazy<bool> _isNvidiaDriver = new Lazy<bool>(() => IsNvidiaDriver());
 
-        public static int MaximumViewportDimensions      => _maximumViewportDimensions.Value;
-        public static int MaximumComputeSharedMemorySize => _maximumComputeSharedMemorySize.Value;
-        public static int StorageBufferOffsetAlignment   => _storageBufferOffsetAlignment.Value;
+        public static bool SupportsAstcCompression          => _supportsAstcCompression.Value;
+        public static bool SupportsNonConstantTextureOffset => _isNvidiaDriver.Value;
+
+        public static int  MaximumViewportDimensions      => _maximumViewportDimensions.Value;
+        public static int  MaximumComputeSharedMemorySize => _maximumComputeSharedMemorySize.Value;
+        public static int  StorageBufferOffsetAlignment   => _storageBufferOffsetAlignment.Value;
 
         private static bool HasExtension(string name)
         {
@@ -36,5 +39,10 @@ namespace Ryujinx.Graphics.OpenGL
         {
             return GL.GetInteger((GetPName)name);
         }
+
+        private static bool IsNvidiaDriver()
+        {
+            return GL.GetString(StringName.Vendor).Equals("NVIDIA Corporation");
+        }
     }
 }
\ No newline at end of file
diff --git a/Ryujinx.Graphics.OpenGL/Renderer.cs b/Ryujinx.Graphics.OpenGL/Renderer.cs
index eec3e320..ac16a37f 100644
--- a/Ryujinx.Graphics.OpenGL/Renderer.cs
+++ b/Ryujinx.Graphics.OpenGL/Renderer.cs
@@ -63,6 +63,7 @@ namespace Ryujinx.Graphics.OpenGL
         {
             return new Capabilities(
                 HwCapabilities.SupportsAstcCompression,
+                HwCapabilities.SupportsNonConstantTextureOffset,
                 HwCapabilities.MaximumViewportDimensions,
                 HwCapabilities.MaximumComputeSharedMemorySize,
                 HwCapabilities.StorageBufferOffsetAlignment);
diff --git a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGen.cs b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGen.cs
index b6cdd7f6..73a71f9e 100644
--- a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGen.cs
+++ b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGen.cs
@@ -133,6 +133,9 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
                     case Instruction.LoadStorage:
                         return InstGenMemory.LoadStorage(context, operation);
 
+                    case Instruction.Lod:
+                        return InstGenMemory.Lod(context, operation);
+
                     case Instruction.PackHalf2x16:
                         return InstGenPacking.PackHalf2x16(context, operation);
 
diff --git a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenHelper.cs b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenHelper.cs
index 2b4ae7f1..ef998fdd 100644
--- a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenHelper.cs
+++ b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenHelper.cs
@@ -73,6 +73,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
             Add(Instruction.LoadLocal,                InstType.Special);
             Add(Instruction.LoadShared,               InstType.Special);
             Add(Instruction.LoadStorage,              InstType.Special);
+            Add(Instruction.Lod,                      InstType.Special);
             Add(Instruction.LogarithmB2,              InstType.CallUnary,      "log2");
             Add(Instruction.LogicalAnd,               InstType.OpBinaryCom,    "&&",              9);
             Add(Instruction.LogicalExclusiveOr,       InstType.OpBinaryCom,    "^^",              10);
diff --git a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenMemory.cs b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenMemory.cs
index ffed4c71..5687ce7e 100644
--- a/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenMemory.cs
+++ b/Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenMemory.cs
@@ -148,6 +148,48 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
             return GetStorageBufferAccessor(indexExpr, offsetExpr, context.Config.Stage);
         }
 
+        public static string Lod(CodeGenContext context, AstOperation operation)
+        {
+            AstTextureOperation texOp = (AstTextureOperation)operation;
+
+            int coordsCount = texOp.Type.GetDimensions();
+
+            bool isBindless = (texOp.Flags & TextureFlags.Bindless) != 0;
+
+            bool isIndexed = (texOp.Type & SamplerType.Indexed) != 0;
+
+            string indexExpr = null;
+
+            if (isIndexed)
+            {
+                indexExpr = GetSoureExpr(context, texOp.GetSource(0), VariableType.S32);
+            }
+
+            string samplerName = OperandManager.GetSamplerName(context.Config.Stage, texOp, indexExpr);
+
+            int coordsIndex = isBindless || isIndexed ? 1 : 0;
+
+            string coordsExpr;
+
+            if (coordsCount > 1)
+            {
+                string[] elems = new string[coordsCount];
+
+                for (int index = 0; index < coordsCount; index++)
+                {
+                    elems[index] = GetSoureExpr(context, texOp.GetSource(coordsIndex + index), VariableType.F32);
+                }
+
+                coordsExpr = "vec" + coordsCount + "(" + string.Join(", ", elems) + ")";
+            }
+            else
+            {
+                coordsExpr = GetSoureExpr(context, texOp.GetSource(coordsIndex), VariableType.F32);
+            }
+
+            return $"textureQueryLod({samplerName}, {coordsExpr}){GetMask(texOp.Index)}";
+        }
+
         public static string StoreLocal(CodeGenContext context, AstOperation operation)
         {
             return StoreLocalOrShared(context, operation, DefaultNames.LocalMemoryName);
@@ -359,15 +401,15 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
                 }
             }
 
-            if (hasDerivatives)
+            if (hasExtraCompareArg)
             {
-                Append(AssembleDerivativesVector(coordsCount)); // dPdx
-                Append(AssembleDerivativesVector(coordsCount)); // dPdy
+                Append(Src(VariableType.F32));
             }
 
-            if (hasExtraCompareArg)
+            if (hasDerivatives)
             {
-                Append(Src(VariableType.F32));
+                Append(AssembleDerivativesVector(coordsCount)); // dPdx
+                Append(AssembleDerivativesVector(coordsCount)); // dPdy
             }
 
             if (isMultisample)
@@ -446,11 +488,13 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
 
             string samplerName = OperandManager.GetSamplerName(context.Config.Stage, texOp, indexExpr);
 
-            IAstNode src0 = operation.GetSource(isBindless || isIndexed ? 1 : 0);
+            int lodSrcIndex = isBindless || isIndexed ? 1 : 0;
+
+            IAstNode lod = operation.GetSource(lodSrcIndex);
 
-            string src0Expr = GetSoureExpr(context, src0, GetSrcVarType(operation.Inst, 0));
+            string lodExpr = GetSoureExpr(context, lod, GetSrcVarType(operation.Inst, lodSrcIndex));
 
-            return $"textureSize({samplerName}, {src0Expr}){GetMask(texOp.Index)}";
+            return $"textureSize({samplerName}, {lodExpr}){GetMask(texOp.Index)}";
         }
 
         private static string GetStorageBufferAccessor(string slotExpr, string offsetExpr, ShaderStage stage)
diff --git a/Ryujinx.Graphics.Shader/IntermediateRepresentation/Instruction.cs b/Ryujinx.Graphics.Shader/IntermediateRepresentation/Instruction.cs
index 5f0407c2..bffdd0fa 100644
--- a/Ryujinx.Graphics.Shader/IntermediateRepresentation/Instruction.cs
+++ b/Ryujinx.Graphics.Shader/IntermediateRepresentation/Instruction.cs
@@ -71,6 +71,7 @@ namespace Ryujinx.Graphics.Shader.IntermediateRepresentation
         LoadLocal,
         LoadShared,
         LoadStorage,
+        Lod,
         LogarithmB2,
         LogicalAnd,
         LogicalExclusiveOr,
diff --git a/Ryujinx.Graphics.Shader/ShaderCapabilities.cs b/Ryujinx.Graphics.Shader/ShaderCapabilities.cs
index 8e0c95e9..809481b5 100644
--- a/Ryujinx.Graphics.Shader/ShaderCapabilities.cs
+++ b/Ryujinx.Graphics.Shader/ShaderCapabilities.cs
@@ -3,22 +3,25 @@ namespace Ryujinx.Graphics.Shader
     public struct ShaderCapabilities
     {
         // Initialize with default values for Maxwell.
-        private static readonly ShaderCapabilities _default = new ShaderCapabilities(32768, 49152, 16);
+        private static readonly ShaderCapabilities _default = new ShaderCapabilities(0x8000, 0xc000, 16, true);
 
         public static ShaderCapabilities Default => _default;
 
-        public int MaximumViewportDimensions      { get; }
-        public int MaximumComputeSharedMemorySize { get; }
-        public int StorageBufferOffsetAlignment   { get; }
+        public int  MaximumViewportDimensions        { get; }
+        public int  MaximumComputeSharedMemorySize   { get; }
+        public int  StorageBufferOffsetAlignment     { get; }
+        public bool SupportsNonConstantTextureOffset { get; }
 
         public ShaderCapabilities(
-            int maximumViewportDimensions,
-            int maximumComputeSharedMemorySize,
-            int storageBufferOffsetAlignment)
+            int  maximumViewportDimensions,
+            int  maximumComputeSharedMemorySize,
+            int  storageBufferOffsetAlignment,
+            bool supportsNonConstantTextureOffset)
         {
-            MaximumViewportDimensions      = maximumViewportDimensions;
-            MaximumComputeSharedMemorySize = maximumComputeSharedMemorySize;
-            StorageBufferOffsetAlignment   = storageBufferOffsetAlignment;
+            MaximumViewportDimensions        = maximumViewportDimensions;
+            MaximumComputeSharedMemorySize   = maximumComputeSharedMemorySize;
+            StorageBufferOffsetAlignment     = storageBufferOffsetAlignment;
+            SupportsNonConstantTextureOffset = supportsNonConstantTextureOffset;
         }
     }
 }
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Shader/StructuredIr/InstructionInfo.cs b/Ryujinx.Graphics.Shader/StructuredIr/InstructionInfo.cs
index 9614b659..c276959a 100644
--- a/Ryujinx.Graphics.Shader/StructuredIr/InstructionInfo.cs
+++ b/Ryujinx.Graphics.Shader/StructuredIr/InstructionInfo.cs
@@ -85,6 +85,7 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
             Add(Instruction.LoadLocal,                VariableType.U32,    VariableType.S32);
             Add(Instruction.LoadShared,               VariableType.U32,    VariableType.S32);
             Add(Instruction.LoadStorage,              VariableType.U32,    VariableType.S32,    VariableType.S32);
+            Add(Instruction.Lod,                      VariableType.F32);
             Add(Instruction.LogarithmB2,              VariableType.Scalar, VariableType.Scalar);
             Add(Instruction.LogicalAnd,               VariableType.Bool,   VariableType.Bool,   VariableType.Bool);
             Add(Instruction.LogicalExclusiveOr,       VariableType.Bool,   VariableType.Bool,   VariableType.Bool);
@@ -139,9 +140,11 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
         {
             // TODO: Return correct type depending on source index,
             // that can improve the decompiler output.
-            if (inst == Instruction.TextureSample ||
-                inst == Instruction.ImageLoad     ||
-                inst == Instruction.ImageStore)
+            if (
+                inst == Instruction.ImageLoad  ||
+                inst == Instruction.ImageStore ||
+                inst == Instruction.Lod        ||
+                inst == Instruction.TextureSample)
             {
                 return VariableType.F32;
             }
diff --git a/Ryujinx.Graphics.Shader/Translation/Lowering.cs b/Ryujinx.Graphics.Shader/Translation/Lowering.cs
index 9a17dd83..1cd1df37 100644
--- a/Ryujinx.Graphics.Shader/Translation/Lowering.cs
+++ b/Ryujinx.Graphics.Shader/Translation/Lowering.cs
@@ -1,5 +1,6 @@
 using Ryujinx.Graphics.Shader.IntermediateRepresentation;
 using System.Collections.Generic;
+using System.Diagnostics;
 
 using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
 using static Ryujinx.Graphics.Shader.Translation.GlobalMemory;
@@ -23,13 +24,18 @@ namespace Ryujinx.Graphics.Shader.Translation
 
                     if (UsesGlobalMemory(operation.Inst))
                     {
-                        node = LowerGlobal(node, config);
+                        node = RewriteGlobalAccess(node, config);
+                    }
+
+                    if (!config.Capabilities.SupportsNonConstantTextureOffset && operation.Inst == Instruction.TextureSample)
+                    {
+                        node = RewriteTextureSample(node);
                     }
                 }
             }
         }
 
-        private static LinkedListNode<INode> LowerGlobal(LinkedListNode<INode> node, ShaderConfig config)
+        private static LinkedListNode<INode> RewriteGlobalAccess(LinkedListNode<INode> node, ShaderConfig config)
         {
             Operation operation = (Operation)node.Value;
 
@@ -117,5 +123,217 @@ namespace Ryujinx.Graphics.Shader.Translation
 
             return node;
         }
+
+        private static LinkedListNode<INode> RewriteTextureSample(LinkedListNode<INode> node)
+        {
+            TextureOperation texOp = (TextureOperation)node.Value;
+
+            bool hasOffset  = (texOp.Flags & TextureFlags.Offset)  != 0;
+            bool hasOffsets = (texOp.Flags & TextureFlags.Offsets) != 0;
+
+            if (!(hasOffset || hasOffsets))
+            {
+                return node;
+            }
+
+            bool isBindless     = (texOp.Flags & TextureFlags.Bindless)    != 0;
+            bool isGather       = (texOp.Flags & TextureFlags.Gather)      != 0;
+            bool hasDerivatives = (texOp.Flags & TextureFlags.Derivatives) != 0;
+            bool hasLodBias     = (texOp.Flags & TextureFlags.LodBias)     != 0;
+            bool hasLodLevel    = (texOp.Flags & TextureFlags.LodLevel)    != 0;
+
+            bool isArray       = (texOp.Type & SamplerType.Array)       != 0;
+            bool isIndexed     = (texOp.Type & SamplerType.Indexed)     != 0;
+            bool isMultisample = (texOp.Type & SamplerType.Multisample) != 0;
+            bool isShadow      = (texOp.Type & SamplerType.Shadow)      != 0;
+
+            int coordsCount = texOp.Type.GetDimensions();
+
+            int offsetsCount = coordsCount * (hasOffsets ? 4 : 1);
+
+            Operand[] offsets = new Operand[offsetsCount];
+            Operand[] sources = new Operand[texOp.SourcesCount - offsetsCount];
+
+            int srcIndex = 0;
+            int dstIndex = 0;
+
+            int copyCount = 0;
+
+            if (isBindless || isIndexed)
+            {
+                copyCount++;
+            }
+
+            Operand[] lodSources = new Operand[copyCount + coordsCount];
+
+            for (int index = 0; index < lodSources.Length; index++)
+            {
+                lodSources[index] = texOp.GetSource(index);
+            }
+
+            copyCount += coordsCount;
+
+            if (isArray)
+            {
+                copyCount++;
+            }
+
+            if (isShadow)
+            {
+                copyCount++;
+            }
+
+            if (hasDerivatives)
+            {
+                copyCount += coordsCount * 2;
+            }
+
+            if (isMultisample)
+            {
+                copyCount++;
+            }
+            else if (hasLodLevel)
+            {
+                copyCount++;
+            }
+
+            for (int index = 0; index < copyCount; index++)
+            {
+                sources[dstIndex++] = texOp.GetSource(srcIndex++);
+            }
+
+            bool areAllOffsetsConstant = true;
+
+            for (int index = 0; index < offsetsCount; index++)
+            {
+                Operand offset = texOp.GetSource(srcIndex++);
+
+                areAllOffsetsConstant &= offset.Type == OperandType.Constant;
+
+                offsets[index] = offset;
+            }
+
+            if (areAllOffsetsConstant)
+            {
+                return node;
+            }
+
+            if (hasLodBias)
+            {
+               sources[dstIndex++] = texOp.GetSource(srcIndex++);
+            }
+
+            if (isGather && !isShadow)
+            {
+               sources[dstIndex++] = texOp.GetSource(srcIndex++);
+            }
+
+            Operand Int(Operand value)
+            {
+                Operand res = Local();
+
+                node.List.AddBefore(node, new Operation(Instruction.ConvertFPToS32, res, value));
+
+                return res;
+            }
+
+            Operand Float(Operand value)
+            {
+                Operand res = Local();
+
+                node.List.AddBefore(node, new Operation(Instruction.ConvertS32ToFP, res, value));
+
+                return res;
+            }
+
+            Operand lod = Local();
+
+            node.List.AddBefore(node, new TextureOperation(
+                Instruction.Lod,
+                texOp.Type,
+                texOp.Flags,
+                texOp.Handle,
+                1,
+                lod,
+                lodSources));
+
+            int coordsIndex = isBindless || isIndexed ? 1 : 0;
+
+            for (int index = 0; index < coordsCount; index++)
+            {
+                Operand coordSize = Local();
+
+                Operand[] texSizeSources;
+
+                if (isBindless || isIndexed)
+                {
+                    texSizeSources = new Operand[] { sources[0], Int(lod) };
+                }
+                else
+                {
+                    texSizeSources = new Operand[] { Int(lod) };
+                }
+
+                node.List.AddBefore(node, new TextureOperation(
+                    Instruction.TextureSize,
+                    texOp.Type,
+                    texOp.Flags,
+                    texOp.Handle,
+                    index,
+                    coordSize,
+                    texSizeSources));
+
+                Operand offset = Local();
+
+                Operand intOffset = offsets[index + (hasOffsets ? texOp.Index * coordsCount : 0)];
+
+                node.List.AddBefore(node, new Operation(Instruction.FP | Instruction.Divide, offset, Float(intOffset), Float(coordSize)));
+
+                Operand source = sources[coordsIndex + index];
+
+                Operand coordPlusOffset = Local();
+
+                node.List.AddBefore(node, new Operation(Instruction.FP | Instruction.Add, coordPlusOffset, source, offset));
+
+                sources[coordsIndex + index] = coordPlusOffset;
+            }
+
+            int componentIndex;
+
+            if (isGather && !isShadow)
+            {
+                Operand gatherComponent = sources[dstIndex - 1];
+
+                Debug.Assert(gatherComponent.Type == OperandType.Constant);
+
+                componentIndex = gatherComponent.Value;
+            }
+            else
+            {
+                componentIndex = texOp.Index;
+            }
+
+            TextureOperation newTexOp = new TextureOperation(
+                Instruction.TextureSample,
+                texOp.Type,
+                texOp.Flags & ~(TextureFlags.Offset | TextureFlags.Offsets),
+                texOp.Handle,
+                componentIndex,
+                texOp.Dest,
+                sources);
+
+            for (int index = 0; index < texOp.SourcesCount; index++)
+            {
+                texOp.SetSource(index, null);
+            }
+
+            LinkedListNode<INode> oldNode = node;
+
+            node = node.List.AddBefore(node, newTexOp);
+
+            node.List.Remove(oldNode);
+
+            return node;
+        }
     }
 }
\ No newline at end of file