Implement buffer texture alignment

src/Ryujinx.Graphics.GAL/Capabilities.cs

@@ -32,6 +32,7 @@ namespace Ryujinx.Graphics.GAL
         public readonly bool SupportsGeometryShader;
         public readonly bool SupportsGeometryShaderPassthrough;
         public readonly bool SupportsTransformFeedback;
+        public readonly bool SupportsImageBufferPixelAlignment;
         public readonly bool SupportsImageLoadFormatted;
         public readonly bool SupportsLayerVertexTessellation;
         public readonly bool SupportsMismatchingViewFormat;
@@ -43,6 +44,7 @@ namespace Ryujinx.Graphics.GAL
         public readonly bool SupportsShaderBarrierDivergence;
         public readonly bool SupportsShaderFloat64;
         public readonly bool SupportsShaderNonUniformIndexing;
+        public readonly bool SupportsTextureBufferPixelAlignment;
         public readonly bool SupportsTextureGatherOffsets;
         public readonly bool SupportsTextureShadowLod;
         public readonly bool SupportsVertexStoreAndAtomics;
@@ -101,6 +103,7 @@ namespace Ryujinx.Graphics.GAL
             bool supportsGeometryShader,
             bool supportsGeometryShaderPassthrough,
             bool supportsTransformFeedback,
+            bool supportsImageBufferPixelAlignment,
             bool supportsImageLoadFormatted,
             bool supportsLayerVertexTessellation,
             bool supportsMismatchingViewFormat,
@@ -112,6 +115,7 @@ namespace Ryujinx.Graphics.GAL
             bool supportsShaderBarrierDivergence,
             bool supportsShaderFloat64,
             bool supportsShaderNonUniformIndexing,
+            bool supportsTextureBufferPixelAlignment,
             bool supportsTextureGatherOffsets,
             bool supportsTextureShadowLod,
             bool supportsVertexStoreAndAtomics,
@@ -164,6 +168,7 @@ namespace Ryujinx.Graphics.GAL
             SupportsGeometryShader = supportsGeometryShader;
             SupportsGeometryShaderPassthrough = supportsGeometryShaderPassthrough;
             SupportsTransformFeedback = supportsTransformFeedback;
+            SupportsImageBufferPixelAlignment = supportsImageBufferPixelAlignment;
             SupportsImageLoadFormatted = supportsImageLoadFormatted;
             SupportsLayerVertexTessellation = supportsLayerVertexTessellation;
             SupportsMismatchingViewFormat = supportsMismatchingViewFormat;
@@ -175,6 +180,7 @@ namespace Ryujinx.Graphics.GAL
             SupportsShaderBarrierDivergence = supportsShaderBarrierDivergence;
             SupportsShaderFloat64 = supportsShaderFloat64;
             SupportsShaderNonUniformIndexing = supportsShaderNonUniformIndexing;
+            SupportsTextureBufferPixelAlignment = supportsTextureBufferPixelAlignment;
             SupportsTextureGatherOffsets = supportsTextureGatherOffsets;
             SupportsTextureShadowLod = supportsTextureShadowLod;
             SupportsVertexStoreAndAtomics = supportsVertexStoreAndAtomics;

src/Ryujinx.Graphics.Gpu/Image/TextureBindingsManager.cs

@@ -4,6 +4,7 @@ using Ryujinx.Graphics.Gpu.Engine.Types;
 using Ryujinx.Graphics.Gpu.Memory;
 using Ryujinx.Graphics.Gpu.Shader;
 using Ryujinx.Graphics.Shader;
+using Ryujinx.Memory.Range;
 using System;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
@@ -67,6 +68,9 @@ namespace Ryujinx.Graphics.Gpu.Image
 
         private int _lastFragmentTotal;
 
+        private readonly int _bufferTextureAlignment;
+        private readonly int _bufferImageAlignment;
+
         /// <summary>
         /// Constructs a new instance of the texture bindings manager.
         /// </summary>
@@ -108,6 +112,10 @@ namespace Ryujinx.Graphics.Gpu.Image
             }
 
             _textureCounts = [];
+
+            int alignment = context.Capabilities.TextureBufferOffsetAlignment;
+            _bufferTextureAlignment = context.Capabilities.SupportsTextureBufferPixelAlignment ? 0 : alignment;
+            _bufferImageAlignment = context.Capabilities.SupportsImageBufferPixelAlignment ? 0 : alignment;
         }
 
         /// <summary>
@@ -521,10 +529,12 @@ namespace Ryujinx.Graphics.Gpu.Image
 
                 if (hostTexture != null && texture.Target == Target.TextureBuffer)
                 {
+                    MultiRange range = GetAlignedBufferTextureRange(texture, index, stageIndex, false);
+
                     // Ensure that the buffer texture is using the correct buffer as storage.
                     // Buffers are frequently re-created to accommodate larger data, so we need to re-bind
                     // to ensure we're not using a old buffer that was already deleted.
-                    _channel.BufferManager.SetBufferTextureStorage(stage, hostTexture, texture.Range, bindingInfo, false);
+                    _channel.BufferManager.SetBufferTextureStorage(stage, hostTexture, range, bindingInfo, false);
 
                     // Cache is not used for buffer texture, it must always rebind.
                     state.CachedTexture = null;
@@ -659,7 +669,9 @@ namespace Ryujinx.Graphics.Gpu.Image
                     // Buffers are frequently re-created to accommodate larger data, so we need to re-bind
                     // to ensure we're not using a old buffer that was already deleted.
 
-                    _channel.BufferManager.SetBufferTextureStorage(stage, hostTexture, texture.Range, bindingInfo, true);
+                    MultiRange range = GetAlignedBufferTextureRange(texture, index, stageIndex, true);
+
+                    _channel.BufferManager.SetBufferTextureStorage(stage, hostTexture, range, bindingInfo, true);
 
                     // Cache is not used for buffer texture, it must always rebind.
                     state.CachedTexture = null;
@@ -692,6 +704,61 @@ namespace Ryujinx.Graphics.Gpu.Image
             return specStateMatches;
         }
 
+        /// <summary>
+        /// Gets the aligned address of the texture according to the host requirements.
+        /// </summary>
+        /// <param name="texture">Texture to have its address aligned</param>
+        /// <param name="index">Index of the texture in the shader</param>
+        /// <param name="stageIndex">Index of the shader stage</param>
+        /// <param name="isImage">True if the texture is bound as image, false for sampled textures</param>
+        /// <returns>Aligned address and size of the buffer texture</returns>
+        private MultiRange GetAlignedBufferTextureRange(Texture texture, int index, int stageIndex, bool isImage)
+        {
+            MultiRange range = texture.Range;
+            int alignment = isImage ? _bufferImageAlignment : _bufferTextureAlignment;
+
+            if (alignment != 0)
+            {
+                MemoryRange firstRange = range.GetSubRange(0);
+                ulong misalign = firstRange.Address & ((ulong)alignment - 1);
+                int offset = misalign != 0 ? (int)misalign / texture.Info.FormatInfo.BytesPerPixel : 0;
+
+                if (misalign != 0)
+                {
+                    firstRange = new MemoryRange(firstRange.Address - misalign, firstRange.Size + misalign);
+
+                    if (range.Count > 1)
+                    {
+                        MemoryRange[] ranges = new MemoryRange[range.Count];
+
+                        ranges[0] = firstRange;
+
+                        for (int i = 1; i < range.Count; i++)
+                        {
+                            ranges[i] = range.GetSubRange(i);
+                        }
+
+                        range = new MultiRange(ranges);
+                    }
+                    else
+                    {
+                        range = new MultiRange(firstRange.Address, firstRange.Size);
+                    }
+                }
+
+                if (isImage)
+                {
+                    _context.SupportBufferUpdater.UpdateBufferImageOffset(stageIndex, index, offset);
+                }
+                else
+                {
+                    _context.SupportBufferUpdater.UpdateBufferTextureOffset(stageIndex, index, offset);
+                }
+            }
+
+            return range;
+        }
+
         /// <summary>
         /// Gets the texture descriptor for a given texture handle.
         /// </summary>

src/Ryujinx.Graphics.Gpu/Memory/SupportBufferUpdater.cs

@@ -129,6 +129,38 @@ namespace Ryujinx.Graphics.Gpu.Memory
             }
         }
 
+        /// <summary>
+        /// Updates the offset used for accessing buffer textures that are not aligned to the host requirements.
+        /// </summary>
+        /// <param name="stageIndex">Index of the shader stage where the texture is used</param>
+        /// <param name="bindingIndex">Index of the texture binding in the shader</param>
+        /// <param name="offset">Offset for the misaligned part of the address</param>
+        public void UpdateBufferTextureOffset(int stageIndex, int bindingIndex, int offset)
+        {
+            if (_data.BufferTextureOffset[stageIndex][bindingIndex].X != offset)
+            {
+                _data.BufferTextureOffset[stageIndex][bindingIndex].X = offset;
+                int index = stageIndex * SupportBuffer.TextureCount + bindingIndex;
+                MarkDirty(SupportBuffer.BufferTextureOffsetOffset + index * sizeof(int) * 4, sizeof(int));
+            }
+        }
+
+        /// <summary>
+        /// Updates the offset used for accessing buffer images that are not aligned to the host requirements.
+        /// </summary>
+        /// <param name="stageIndex">Index of the shader stage where the image is used</param>
+        /// <param name="bindingIndex">Index of the image binding in the shader</param>
+        /// <param name="offset">Offset for the misaligned part of the address</param>
+        public void UpdateBufferImageOffset(int stageIndex, int bindingIndex, int offset)
+        {
+            if (_data.BufferTextureOffset[stageIndex][bindingIndex].Y != offset)
+            {
+                _data.BufferTextureOffset[stageIndex][bindingIndex].Y = offset;
+                int index = stageIndex * SupportBuffer.TextureCount + bindingIndex;
+                MarkDirty(SupportBuffer.BufferTextureOffsetOffset + index * sizeof(int) * 4 + sizeof(int), sizeof(int));
+            }
+        }
+
         /// <summary>
         /// Sets whether the format of a given render target is a BGRA format.
         /// </summary>

src/Ryujinx.Graphics.Gpu/Shader/GpuAccessorBase.cs

@@ -207,6 +207,10 @@ namespace Ryujinx.Graphics.Gpu.Shader
 
         public bool QueryHostSupportsBgraFormat() => _context.Capabilities.SupportsBgraFormat;
 
+        public bool QueryHostSupportsBufferImagePixelAlignment() => _context.Capabilities.SupportsImageBufferPixelAlignment;
+
+        public bool QueryHostSupportsBufferTexturePixelAlignment() => _context.Capabilities.SupportsTextureBufferPixelAlignment;
+
         public bool QueryHostSupportsFragmentShaderInterlock() => _context.Capabilities.SupportsFragmentShaderInterlock;
 
         public bool QueryHostSupportsFragmentShaderOrderingIntel() => _context.Capabilities.SupportsFragmentShaderOrderingIntel;

src/Ryujinx.Graphics.OpenGL/OpenGLRenderer.cs

@@ -174,6 +174,7 @@ namespace Ryujinx.Graphics.OpenGL
                 supportsGeometryShader: true,
                 supportsGeometryShaderPassthrough: HwCapabilities.SupportsGeometryShaderPassthrough,
                 supportsTransformFeedback: true,
+                supportsImageBufferPixelAlignment: false,
                 supportsImageLoadFormatted: HwCapabilities.SupportsImageLoadFormatted,
                 supportsLayerVertexTessellation: HwCapabilities.SupportsShaderViewportLayerArray,
                 supportsMismatchingViewFormat: HwCapabilities.SupportsMismatchingViewFormat,
@@ -185,6 +186,7 @@ namespace Ryujinx.Graphics.OpenGL
                 supportsShaderBarrierDivergence: !(intelWindows || intelUnix),
                 supportsShaderFloat64: true,
                 supportsShaderNonUniformIndexing: false,
+                supportsTextureBufferPixelAlignment: false,
                 supportsTextureGatherOffsets: true,
                 supportsTextureShadowLod: HwCapabilities.SupportsTextureShadowLod,
                 supportsVertexStoreAndAtomics: true,

src/Ryujinx.Graphics.Shader/IGpuAccessor.cs

@@ -234,6 +234,24 @@ namespace Ryujinx.Graphics.Shader
             return true;
         }
 
+        /// <summary>
+        /// Queries host support for buffer image access with the buffer offset aligned to a single pixel rather than a fixed alignment.
+        /// </summary>
+        /// <returns>True if the host supports buffer image access with pixel alignment, false otherwise</returns>
+        bool QueryHostSupportsBufferImagePixelAlignment()
+        {
+            return true;
+        }
+
+        /// <summary>
+        /// Queries host support for buffer texture access with the buffer offset aligned to a single pixel rather than a fixed alignment.
+        /// </summary>
+        /// <returns>True if the host supports buffer texture access with pixel alignment, false otherwise</returns>
+        bool QueryHostSupportsBufferTexturePixelAlignment()
+        {
+            return true;
+        }
+
         /// <summary>
         /// Queries host support for fragment shader ordering critical sections on the shader code.
         /// </summary>

src/Ryujinx.Graphics.Shader/SupportBuffer.cs

@@ -24,6 +24,7 @@ namespace Ryujinx.Graphics.Shader
         RenderScale,
         TfeOffset,
         TfeVertexCount,
+        BufferTextureOffset,
     }
 
     public struct SupportBuffer
@@ -42,10 +43,13 @@ namespace Ryujinx.Graphics.Shader
         public static readonly int ComputeRenderScaleOffset;
         public static readonly int TfeOffsetOffset;
         public static readonly int TfeVertexCountOffset;
+        public static readonly int BufferTextureOffsetOffset;
 
         public const int FragmentIsBgraCount = 8;
+        public const int TextureCount = 64;
+
         // One for the render target, 64 for the textures, and 8 for the images.
-        public const int RenderScaleMaxCount = 1 + 64 + 8;
+        public const int RenderScaleMaxCount = 1 + TextureCount + 8;
 
         private static int OffsetOf<T>(ref SupportBuffer storage, ref T target)
         {
@@ -68,6 +72,7 @@ namespace Ryujinx.Graphics.Shader
             ComputeRenderScaleOffset = GraphicsRenderScaleOffset + FieldSize;
             TfeOffsetOffset = OffsetOf(ref instance, ref instance.TfeOffset);
             TfeVertexCountOffset = OffsetOf(ref instance, ref instance.TfeVertexCount);
+            BufferTextureOffsetOffset = OffsetOf(ref instance, ref instance.BufferTextureOffset);
         }
 
         internal static StructureType GetStructureType()
@@ -80,7 +85,8 @@ namespace Ryujinx.Graphics.Shader
                 new StructureField(AggregateType.S32, "frag_scale_count"),
                 new StructureField(AggregateType.Array | AggregateType.FP32, "render_scale", RenderScaleMaxCount),
                 new StructureField(AggregateType.Vector4 | AggregateType.S32, "tfe_offset"),
-                new StructureField(AggregateType.S32, "tfe_vertex_count")
+                new StructureField(AggregateType.S32, "tfe_vertex_count"),
+                new StructureField(AggregateType.Array | AggregateType.Vector2 | AggregateType.S32, "buffer_texture_offset")
             ]);
         }
 
@@ -95,5 +101,7 @@ namespace Ryujinx.Graphics.Shader
 
         public Vector4<int> TfeOffset;
         public Vector4<int> TfeVertexCount;
+
+        public Array5<Array64<Vector4<int>>> BufferTextureOffset;
     }
 }

src/Ryujinx.Graphics.Shader/Translation/Transforms/TexturePass.cs

@@ -1,4 +1,5 @@
 using Ryujinx.Graphics.Shader.IntermediateRepresentation;
+using Ryujinx.Graphics.Shader.Translation.Optimizations;
 using System.Collections.Generic;
 using System.Linq;
 using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
@@ -27,7 +28,19 @@ namespace Ryujinx.Graphics.Shader.Translation.Transforms
 
                     if (texOp.Type == SamplerType.TextureBuffer && !context.GpuAccessor.QueryHostSupportsSnormBufferTextureFormat())
                     {
-                        node = InsertSnormNormalization(node, context.ResourceManager, context.GpuAccessor);
+                        if (!context.GpuAccessor.QueryHostSupportsSnormBufferTextureFormat())
+                        {
+                            node = InsertSnormNormalization(node, context.ResourceManager, context.GpuAccessor);
+                        }
+
+                        if (!context.GpuAccessor.QueryHostSupportsBufferTexturePixelAlignment())
+                        {
+                            node = InsertCoordOffset(node, context.ResourceManager, context.Stage, isImage: false);
+                        }
+                    }
+                    else if (!context.GpuAccessor.QueryHostSupportsBufferTexturePixelAlignment() && texOp.Inst.IsImage())
+                    {
+                        node = InsertCoordOffset(node, context.ResourceManager, context.Stage, isImage: true);
                     }
                 }
             }
@@ -278,6 +291,87 @@ namespace Ryujinx.Graphics.Shader.Translation.Transforms
             return node;
         }
 
+        private static LinkedListNode<INode> InsertCoordOffset(LinkedListNode<INode> node, ResourceManager resourceManager, ShaderStage stage, bool isImage)
+        {
+            // Some GPUs have fixed alignment requirements for buffer textures.
+            // For those cases, we bind the aligned buffer offset, and apply the remaining offset on the shader.
+
+            TextureOperation texOp = (TextureOperation)node.Value;
+
+            if ((texOp.Type & SamplerType.Mask) != SamplerType.TextureBuffer)
+            {
+                return node;
+            }
+
+            Operand[] sources = new Operand[texOp.SourcesCount];
+
+            for (int i = 0; i < texOp.SourcesCount; i++)
+            {
+                sources[i] = texOp.GetSource(i);
+            }
+
+            bool isBindless = (texOp.Flags & TextureFlags.Bindless) != 0;
+            bool isIndexed = resourceManager.IsArrayOfTexturesOrImages(texOp.Binding, isImage);
+
+            int coordsIndex = isBindless || isIndexed ? 1 : 0;
+
+            Operand[] dests = new Operand[texOp.DestsCount];
+
+            for (int i = 0; i < texOp.DestsCount; i++)
+            {
+                dests[i] = texOp.GetDest(i);
+            }
+
+            LinkedListNode<INode> oldNode = node;
+
+            Operand source = sources[coordsIndex];
+            Operand offset = Local();
+            Operand coordPlusOffset = Local();
+
+            int stageIndex = stage switch
+            {
+                ShaderStage.TessellationControl => 1,
+                ShaderStage.TessellationEvaluation => 2,
+                ShaderStage.Geometry => 3,
+                ShaderStage.Fragment => 4,
+                _ => 0,
+            };
+
+            int bindingIndex = isImage
+                ? resourceManager.FindImageDescriptorIndex(texOp.Binding)
+                : resourceManager.FindTextureDescriptorIndex(texOp.Binding);
+
+            node.List.AddBefore(node, new Operation(
+                Instruction.Load,
+                StorageKind.ConstantBuffer,
+                offset,
+                Const(SupportBuffer.Binding),
+                Const((int)SupportBufferField.BufferTextureOffset),
+                Const(stageIndex * SupportBuffer.TextureCount + bindingIndex),
+                Const(isImage ? 1 : 0)));
+
+            node.List.AddBefore(node, new Operation(Instruction.Add, coordPlusOffset, source, offset));
+
+            sources[coordsIndex] = coordPlusOffset;
+
+            TextureOperation newTexOp = new(
+                texOp.Inst,
+                texOp.Type,
+                texOp.Format,
+                texOp.Flags,
+                texOp.Set,
+                texOp.Binding,
+                texOp.Index,
+                dests,
+                sources);
+
+            node = node.List.AddBefore(node, newTexOp);
+
+            Utils.DeleteNode(oldNode, texOp);
+
+            return node;
+        }
+
         private static LinkedListNode<INode> InsertConstOffsets(LinkedListNode<INode> node, ResourceManager resourceManager, IGpuAccessor gpuAccessor, ShaderStage stage)
         {
             // Non-constant texture offsets are not allowed (according to the spec),

src/Ryujinx.Graphics.Vulkan/VulkanRenderer.cs

@@ -784,6 +784,7 @@ namespace Ryujinx.Graphics.Vulkan
                 supportsGeometryShader: Capabilities.SupportsGeometryShader,
                 supportsGeometryShaderPassthrough: Capabilities.SupportsGeometryShaderPassthrough,
                 supportsTransformFeedback: Capabilities.SupportsTransformFeedback,
+                supportsImageBufferPixelAlignment: false,
                 supportsImageLoadFormatted: features2.Features.ShaderStorageImageReadWithoutFormat,
                 supportsLayerVertexTessellation: featuresVk12.ShaderOutputLayer,
                 supportsMismatchingViewFormat: true,
@@ -797,6 +798,7 @@ namespace Ryujinx.Graphics.Vulkan
                 supportsShaderNonUniformIndexing:
                     featuresVk12.ShaderSampledImageArrayNonUniformIndexing &&
                     featuresVk12.ShaderStorageImageArrayNonUniformIndexing,
+                supportsTextureBufferPixelAlignment: false,
                 supportsTextureGatherOffsets: features2.Features.ShaderImageGatherExtended,
                 supportsTextureShadowLod: false,
                 supportsVertexStoreAndAtomics: features2.Features.VertexPipelineStoresAndAtomics,