Truncate vertex attribute format if it exceeds stride on MoltenVK (#5094)

* Truncate vertex attribute format if it exceeds stride on MoltenVK

* Fix BGR format

* Move vertex attribute check to pipeline creation to avoid costs

* No need for this to be public

.github/workflows/release.yml

@@ -33,7 +33,7 @@ jobs:
         shell: bash
 
       - name: Create tag
-        uses: actions/github-script@v5
+        uses: actions/github-script@v6
         with:
           script: |
             github.rest.git.createRef({

Directory.Packages.props

@@ -32,7 +32,7 @@
     <PackageVersion Include="OpenTK.Windowing.GraphicsLibraryFramework" Version="4.7.7" />
     <PackageVersion Include="Ryujinx.Audio.OpenAL.Dependencies" Version="1.21.0.1" />
     <PackageVersion Include="Ryujinx.Graphics.Nvdec.Dependencies" Version="5.0.1-build13" />
-    <PackageVersion Include="Ryujinx.Graphics.Vulkan.Dependencies.MoltenVK" Version="1.2.3" />
+    <PackageVersion Include="Ryujinx.Graphics.Vulkan.Dependencies.MoltenVK" Version="1.2.0" />
     <PackageVersion Include="Ryujinx.GtkSharp" Version="3.24.24.59-ryujinx" />
     <PackageVersion Include="Ryujinx.SDL2-CS" Version="2.26.3-build25" />
     <PackageVersion Include="shaderc.net" Version="0.1.0" />

src/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 = 4646;
+        private const uint CodeGenVersion = 5031;
 
         private const string SharedTocFileName = "shared.toc";
         private const string SharedDataFileName = "shared.data";

src/Ryujinx.Graphics.Shader/Instructions/InstEmit.cs

@@ -187,27 +187,6 @@ namespace Ryujinx.Graphics.Shader.Instructions
             context.Config.GpuAccessor.Log("Shader instruction Longjmp is not implemented.");
         }
 
-        public static void P2rR(EmitterContext context)
-        {
-            InstP2rR op = context.GetOp<InstP2rR>();
-
-            context.Config.GpuAccessor.Log("Shader instruction P2rR is not implemented.");
-        }
-
-        public static void P2rI(EmitterContext context)
-        {
-            InstP2rI op = context.GetOp<InstP2rI>();
-
-            context.Config.GpuAccessor.Log("Shader instruction P2rI is not implemented.");
-        }
-
-        public static void P2rC(EmitterContext context)
-        {
-            InstP2rC op = context.GetOp<InstP2rC>();
-
-            context.Config.GpuAccessor.Log("Shader instruction P2rC is not implemented.");
-        }
-
         public static void Pexit(EmitterContext context)
         {
             InstPexit op = context.GetOp<InstPexit>();

src/Ryujinx.Graphics.Shader/Instructions/InstEmitMove.cs

@@ -209,21 +209,15 @@ namespace Ryujinx.Graphics.Shader.Instructions
                 return context.ICompareNotEqual(context.BitwiseAnd(value, Const(1 << bit)), Const(0));
             }
 
-            if (ccpr)
-            {
-                // TODO: Support Register to condition code flags copy.
-                context.Config.GpuAccessor.Log("R2P.CC not implemented.");
-            }
-            else
-            {
+            int count = ccpr ? RegisterConsts.FlagsCount : RegisterConsts.PredsCount;
+            RegisterType type = ccpr ? RegisterType.Flag : RegisterType.Predicate;
             int shift = (int)byteSel * 8;
             
-                for (int bit = 0; bit < RegisterConsts.PredsCount; bit++)
+            for (int bit = 0; bit < count; bit++)
             {
-                    Operand pred = Register(bit, RegisterType.Predicate);
-                    Operand res = context.ConditionalSelect(Test(mask, bit), Test(value, bit + shift), pred);
-                    context.Copy(pred, res);
-                }
+                Operand flag = Register(bit, type);
+                Operand res = context.ConditionalSelect(Test(mask, bit), Test(value, bit + shift), flag);
+                context.Copy(flag, res);
             }
         }
 

src/Ryujinx.Graphics.Shader/Instructions/InstEmitPredicate.cs

@@ -1,7 +1,6 @@
 using Ryujinx.Graphics.Shader.Decoders;
 using Ryujinx.Graphics.Shader.IntermediateRepresentation;
 using Ryujinx.Graphics.Shader.Translation;
-
 using static Ryujinx.Graphics.Shader.Instructions.InstEmitAluHelper;
 using static Ryujinx.Graphics.Shader.Instructions.InstEmitHelper;
 using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
@@ -50,5 +49,68 @@ namespace Ryujinx.Graphics.Shader.Instructions
             context.Copy(Register(op.DestPred, RegisterType.Predicate), p0Res);
             context.Copy(Register(op.DestPredInv, RegisterType.Predicate), p1Res);
         }
+
+        public static void P2rC(EmitterContext context)
+        {
+            InstP2rC op = context.GetOp<InstP2rC>();
+
+            Operand srcA = GetSrcReg(context, op.SrcA);
+            Operand dest = GetSrcReg(context, op.Dest);
+            Operand mask = GetSrcCbuf(context, op.CbufSlot, op.CbufOffset);
+
+            EmitP2r(context, srcA, dest, mask, op.ByteSel, op.Ccpr);
+        }
+
+        public static void P2rI(EmitterContext context)
+        {
+            InstP2rI op = context.GetOp<InstP2rI>();
+
+            Operand srcA = GetSrcReg(context, op.SrcA);
+            Operand dest = GetSrcReg(context, op.Dest);
+            Operand mask = GetSrcImm(context, op.Imm20);
+
+            EmitP2r(context, srcA, dest, mask, op.ByteSel, op.Ccpr);
+        }
+
+        public static void P2rR(EmitterContext context)
+        {
+            InstP2rR op = context.GetOp<InstP2rR>();
+
+            Operand srcA = GetSrcReg(context, op.SrcA);
+            Operand dest = GetSrcReg(context, op.Dest);
+            Operand mask = GetSrcReg(context, op.SrcB);
+
+            EmitP2r(context, srcA, dest, mask, op.ByteSel, op.Ccpr);
+        }
+
+        private static void EmitP2r(
+            EmitterContext context,
+            Operand srcA,
+            Operand dest,
+            Operand mask,
+            ByteSel byteSel,
+            bool ccpr)
+        {
+            int count = ccpr ? RegisterConsts.FlagsCount : RegisterConsts.PredsCount;
+            int shift = (int)byteSel * 8;
+            mask = context.BitwiseAnd(mask, Const(0xff));
+
+            Operand insert = Const(0);
+            for (int i = 0; i < count; i++)
+            {
+                Operand condition = ccpr
+                    ? Register(i, RegisterType.Flag)
+                    : Register(i, RegisterType.Predicate);
+
+                Operand bit = context.ConditionalSelect(condition, Const(1 << (i + shift)), Const(0));
+                insert = context.BitwiseOr(insert, bit);
+            }
+
+            Operand maskShifted = context.ShiftLeft(mask, Const(shift));
+            Operand masked = context.BitwiseAnd(srcA, context.BitwiseNot(maskShifted));
+            Operand res = context.BitwiseOr(masked, context.BitwiseAnd(insert, maskShifted));
+
+            context.Copy(dest, res);
+        }
     }
 }

src/Ryujinx.Graphics.Vulkan/FormatTable.cs

@@ -168,5 +168,223 @@ namespace Ryujinx.Graphics.Vulkan
         {
             return _table[(int)format];
         }
+
+        public static int GetAttributeFormatSize(VkFormat format)
+        {
+            switch (format)
+            {
+                case VkFormat.R8Unorm:
+                case VkFormat.R8SNorm:
+                case VkFormat.R8Uint:
+                case VkFormat.R8Sint:
+                case VkFormat.R8Uscaled:
+                case VkFormat.R8Sscaled:
+                    return 1;
+
+                case VkFormat.R8G8Unorm:
+                case VkFormat.R8G8SNorm:
+                case VkFormat.R8G8Uint:
+                case VkFormat.R8G8Sint:
+                case VkFormat.R8G8Uscaled:
+                case VkFormat.R8G8Sscaled:
+                case VkFormat.R16Sfloat:
+                case VkFormat.R16Unorm:
+                case VkFormat.R16SNorm:
+                case VkFormat.R16Uint:
+                case VkFormat.R16Sint:
+                case VkFormat.R16Uscaled:
+                case VkFormat.R16Sscaled:
+                    return 2;
+
+                case VkFormat.R8G8B8Unorm:
+                case VkFormat.R8G8B8SNorm:
+                case VkFormat.R8G8B8Uint:
+                case VkFormat.R8G8B8Sint:
+                case VkFormat.R8G8B8Uscaled:
+                case VkFormat.R8G8B8Sscaled:
+                    return 3;
+
+                case VkFormat.R8G8B8A8Unorm:
+                case VkFormat.R8G8B8A8SNorm:
+                case VkFormat.R8G8B8A8Uint:
+                case VkFormat.R8G8B8A8Sint:
+                case VkFormat.R8G8B8A8Srgb:
+                case VkFormat.R8G8B8A8Uscaled:
+                case VkFormat.R8G8B8A8Sscaled:
+                case VkFormat.B8G8R8A8Unorm:
+                case VkFormat.B8G8R8A8Srgb:
+                case VkFormat.R16G16Sfloat:
+                case VkFormat.R16G16Unorm:
+                case VkFormat.R16G16SNorm:
+                case VkFormat.R16G16Uint:
+                case VkFormat.R16G16Sint:
+                case VkFormat.R16G16Uscaled:
+                case VkFormat.R16G16Sscaled:
+                case VkFormat.R32Sfloat:
+                case VkFormat.R32Uint:
+                case VkFormat.R32Sint:
+                case VkFormat.A2B10G10R10UnormPack32:
+                case VkFormat.A2B10G10R10UintPack32:
+                case VkFormat.B10G11R11UfloatPack32:
+                case VkFormat.E5B9G9R9UfloatPack32:
+                case VkFormat.A2B10G10R10SNormPack32:
+                case VkFormat.A2B10G10R10SintPack32:
+                case VkFormat.A2B10G10R10UscaledPack32:
+                case VkFormat.A2B10G10R10SscaledPack32:
+                    return 4;
+
+                case VkFormat.R16G16B16Sfloat:
+                case VkFormat.R16G16B16Unorm:
+                case VkFormat.R16G16B16SNorm:
+                case VkFormat.R16G16B16Uint:
+                case VkFormat.R16G16B16Sint:
+                case VkFormat.R16G16B16Uscaled:
+                case VkFormat.R16G16B16Sscaled:
+                    return 6;
+
+                case VkFormat.R16G16B16A16Sfloat:
+                case VkFormat.R16G16B16A16Unorm:
+                case VkFormat.R16G16B16A16SNorm:
+                case VkFormat.R16G16B16A16Uint:
+                case VkFormat.R16G16B16A16Sint:
+                case VkFormat.R16G16B16A16Uscaled:
+                case VkFormat.R16G16B16A16Sscaled:
+                case VkFormat.R32G32Sfloat:
+                case VkFormat.R32G32Uint:
+                case VkFormat.R32G32Sint:
+                    return 8;
+
+                case VkFormat.R32G32B32Sfloat:
+                case VkFormat.R32G32B32Uint:
+                case VkFormat.R32G32B32Sint:
+                    return 12;
+
+                case VkFormat.R32G32B32A32Sfloat:
+                case VkFormat.R32G32B32A32Uint:
+                case VkFormat.R32G32B32A32Sint:
+                    return 16;
+            }
+
+            return 1;
+        }
+
+        public static VkFormat DropLastComponent(VkFormat format)
+        {
+            switch (format)
+            {
+                case VkFormat.R8G8Unorm:
+                    return VkFormat.R8Unorm;
+                case VkFormat.R8G8SNorm:
+                    return VkFormat.R8SNorm;
+                case VkFormat.R8G8Uint:
+                    return VkFormat.R8Uint;
+                case VkFormat.R8G8Sint:
+                    return VkFormat.R8Sint;
+                case VkFormat.R8G8Uscaled:
+                    return VkFormat.R8Uscaled;
+                case VkFormat.R8G8Sscaled:
+                    return VkFormat.R8Sscaled;
+
+                case VkFormat.R8G8B8Unorm:
+                    return VkFormat.R8G8Unorm;
+                case VkFormat.R8G8B8SNorm:
+                    return VkFormat.R8G8SNorm;
+                case VkFormat.R8G8B8Uint:
+                    return VkFormat.R8G8Uint;
+                case VkFormat.R8G8B8Sint:
+                    return VkFormat.R8G8Sint;
+                case VkFormat.R8G8B8Uscaled:
+                    return VkFormat.R8G8Uscaled;
+                case VkFormat.R8G8B8Sscaled:
+                    return VkFormat.R8G8Sscaled;
+
+                case VkFormat.R8G8B8A8Unorm:
+                    return VkFormat.R8G8B8Unorm;
+                case VkFormat.R8G8B8A8SNorm:
+                    return VkFormat.R8G8B8SNorm;
+                case VkFormat.R8G8B8A8Uint:
+                    return VkFormat.R8G8B8Uint;
+                case VkFormat.R8G8B8A8Sint:
+                    return VkFormat.R8G8B8Sint;
+                case VkFormat.R8G8B8A8Srgb:
+                    return VkFormat.R8G8B8Srgb;
+                case VkFormat.R8G8B8A8Uscaled:
+                    return VkFormat.R8G8B8Uscaled;
+                case VkFormat.R8G8B8A8Sscaled:
+                    return VkFormat.R8G8B8Sscaled;
+                case VkFormat.B8G8R8A8Unorm:
+                    return VkFormat.B8G8R8Unorm;
+                case VkFormat.B8G8R8A8Srgb:
+                    return VkFormat.B8G8R8Srgb;
+
+                case VkFormat.R16G16Sfloat:
+                    return VkFormat.R16Sfloat;
+                case VkFormat.R16G16Unorm:
+                    return VkFormat.R16Unorm;
+                case VkFormat.R16G16SNorm:
+                    return VkFormat.R16SNorm;
+                case VkFormat.R16G16Uint:
+                    return VkFormat.R16Uint;
+                case VkFormat.R16G16Sint:
+                    return VkFormat.R16Sint;
+                case VkFormat.R16G16Uscaled:
+                    return VkFormat.R16Uscaled;
+                case VkFormat.R16G16Sscaled:
+                    return VkFormat.R16Sscaled;
+
+                case VkFormat.R16G16B16Sfloat:
+                    return VkFormat.R16G16Sfloat;
+                case VkFormat.R16G16B16Unorm:
+                    return VkFormat.R16G16Unorm;
+                case VkFormat.R16G16B16SNorm:
+                    return VkFormat.R16G16SNorm;
+                case VkFormat.R16G16B16Uint:
+                    return VkFormat.R16G16Uint;
+                case VkFormat.R16G16B16Sint:
+                    return VkFormat.R16G16Sint;
+                case VkFormat.R16G16B16Uscaled:
+                    return VkFormat.R16G16Uscaled;
+                case VkFormat.R16G16B16Sscaled:
+                    return VkFormat.R16G16Sscaled;
+
+                case VkFormat.R16G16B16A16Sfloat:
+                    return VkFormat.R16G16B16Sfloat;
+                case VkFormat.R16G16B16A16Unorm:
+                    return VkFormat.R16G16B16Unorm;
+                case VkFormat.R16G16B16A16SNorm:
+                    return VkFormat.R16G16B16SNorm;
+                case VkFormat.R16G16B16A16Uint:
+                    return VkFormat.R16G16B16Uint;
+                case VkFormat.R16G16B16A16Sint:
+                    return VkFormat.R16G16B16Sint;
+                case VkFormat.R16G16B16A16Uscaled:
+                    return VkFormat.R16G16B16Uscaled;
+                case VkFormat.R16G16B16A16Sscaled:
+                    return VkFormat.R16G16B16Sscaled;
+
+                case VkFormat.R32G32Sfloat:
+                    return VkFormat.R32Sfloat;
+                case VkFormat.R32G32Uint:
+                    return VkFormat.R32Uint;
+                case VkFormat.R32G32Sint:
+                    return VkFormat.R32Sint;
+
+                case VkFormat.R32G32B32Sfloat:
+                    return VkFormat.R32G32Sfloat;
+                case VkFormat.R32G32B32Uint:
+                    return VkFormat.R32G32Uint;
+                case VkFormat.R32G32B32Sint:
+                    return VkFormat.R32G32Sint;
+
+                case VkFormat.R32G32B32A32Sfloat:
+                    return VkFormat.R32G32B32Sfloat;
+                case VkFormat.R32G32B32A32Uint:
+                    return VkFormat.R32G32B32Uint;
+                case VkFormat.R32G32B32A32Sint:
+                    return VkFormat.R32G32B32Sint;
+            }
+
+            return format;
+        }
     }
 }

src/Ryujinx.Graphics.Vulkan/HardwareCapabilities.cs

@@ -28,6 +28,7 @@ namespace Ryujinx.Graphics.Vulkan
         public readonly bool SupportsSubgroupSizeControl;
         public readonly bool SupportsShaderInt8;
         public readonly bool SupportsShaderStencilExport;
+        public readonly bool SupportsShaderStorageImageMultisample;
         public readonly bool SupportsConditionalRendering;
         public readonly bool SupportsExtendedDynamicState;
         public readonly bool SupportsMultiView;
@@ -63,6 +64,7 @@ namespace Ryujinx.Graphics.Vulkan
             bool supportsSubgroupSizeControl,
             bool supportsShaderInt8,
             bool supportsShaderStencilExport,
+            bool supportsShaderStorageImageMultisample,
             bool supportsConditionalRendering,
             bool supportsExtendedDynamicState,
             bool supportsMultiView,
@@ -97,6 +99,7 @@ namespace Ryujinx.Graphics.Vulkan
             SupportsSubgroupSizeControl = supportsSubgroupSizeControl;
             SupportsShaderInt8 = supportsShaderInt8;
             SupportsShaderStencilExport = supportsShaderStencilExport;
+            SupportsShaderStorageImageMultisample = supportsShaderStorageImageMultisample;
             SupportsConditionalRendering = supportsConditionalRendering;
             SupportsExtendedDynamicState = supportsExtendedDynamicState;
             SupportsMultiView = supportsMultiView;

src/Ryujinx.Graphics.Vulkan/PipelineState.cs

@@ -1,4 +1,5 @@
-using Silk.NET.Vulkan;
+using Ryujinx.Common.Memory;
+using Silk.NET.Vulkan;
 using System;
 
 namespace Ryujinx.Graphics.Vulkan
@@ -308,6 +309,8 @@ namespace Ryujinx.Graphics.Vulkan
         public PipelineLayout PipelineLayout;
         public SpecData SpecializationData;
 
+        private Array32<VertexInputAttributeDescription> _vertexAttributeDescriptions2;
+
         public void Initialize()
         {
             Stages = new NativeArray<PipelineShaderStageCreateInfo>(Constants.MaxShaderStages);
@@ -400,7 +403,15 @@ namespace Ryujinx.Graphics.Vulkan
 
             Pipeline pipelineHandle = default;
 
+            bool isMoltenVk = gd.IsMoltenVk;
+
+            if (isMoltenVk)
+            {
+                UpdateVertexAttributeDescriptions();
+            }
+
             fixed (VertexInputAttributeDescription* pVertexAttributeDescriptions = &Internal.VertexAttributeDescriptions[0])
+            fixed (VertexInputAttributeDescription* pVertexAttributeDescriptions2 = &_vertexAttributeDescriptions2[0])
             fixed (VertexInputBindingDescription* pVertexBindingDescriptions = &Internal.VertexBindingDescriptions[0])
             fixed (Viewport* pViewports = &Internal.Viewports[0])
             fixed (Rect2D* pScissors = &Internal.Scissors[0])
@@ -410,7 +421,7 @@ namespace Ryujinx.Graphics.Vulkan
                 {
                     SType = StructureType.PipelineVertexInputStateCreateInfo,
                     VertexAttributeDescriptionCount = VertexAttributeDescriptionsCount,
-                    PVertexAttributeDescriptions = pVertexAttributeDescriptions,
+                    PVertexAttributeDescriptions = isMoltenVk ? pVertexAttributeDescriptions2 : pVertexAttributeDescriptions,
                     VertexBindingDescriptionCount = VertexBindingDescriptionsCount,
                     PVertexBindingDescriptions = pVertexBindingDescriptions
                 };
@@ -612,6 +623,40 @@ namespace Ryujinx.Graphics.Vulkan
             }
         }
 
+        private void UpdateVertexAttributeDescriptions()
+        {
+            // Vertex attributes exceeding the stride are invalid.
+            // In metal, they cause glitches with the vertex shader fetching incorrect values.
+            // To work around this, we reduce the format to something that doesn't exceed the stride if possible.
+            // The assumption is that the exceeding components are not actually accessed on the shader.
+
+            for (int index = 0; index < VertexAttributeDescriptionsCount; index++)
+            {
+                var attribute = Internal.VertexAttributeDescriptions[index];
+                ref var vb = ref Internal.VertexBindingDescriptions[(int)attribute.Binding];
+
+                Format format = attribute.Format;
+
+                while (vb.Stride != 0 && attribute.Offset + FormatTable.GetAttributeFormatSize(format) > vb.Stride)
+                {
+                    Format newFormat = FormatTable.DropLastComponent(format);
+
+                    if (newFormat == format)
+                    {
+                        // That case means we failed to find a format that fits within the stride,
+                        // so just restore the original format and give up.
+                        format = attribute.Format;
+                        break;
+                    }
+
+                    format = newFormat;
+                }
+
+                attribute.Format = format;
+                _vertexAttributeDescriptions2[index] = attribute;
+            }
+        }
+
         public void Dispose()
         {
             Stages.Dispose();

src/Ryujinx.Graphics.Vulkan/TextureStorage.cs

@@ -80,7 +80,7 @@ namespace Ryujinx.Graphics.Vulkan
 
             var sampleCountFlags = ConvertToSampleCountFlags(gd.Capabilities.SupportedSampleCounts, (uint)info.Samples);
 
-            var usage = GetImageUsageFromFormat(info.Format);
+            var usage = GetImageUsage(info.Format, info.Target, gd.Capabilities.SupportsShaderStorageImageMultisample);
 
             var flags = ImageCreateFlags.CreateMutableFormatBit;
 
@@ -293,7 +293,7 @@ namespace Ryujinx.Graphics.Vulkan
             }
         }
 
-        public static ImageUsageFlags GetImageUsageFromFormat(GAL.Format format)
+        public static ImageUsageFlags GetImageUsage(GAL.Format format, Target target, bool supportsMsStorage)
         {
             var usage = DefaultUsageFlags;
 
@@ -306,7 +306,7 @@ namespace Ryujinx.Graphics.Vulkan
                 usage |= ImageUsageFlags.ColorAttachmentBit;
             }
 
-            if (format.IsImageCompatible())
+            if (format.IsImageCompatible() && (supportsMsStorage || !target.IsMultisample()))
             {
                 usage |= ImageUsageFlags.StorageBit;
             }

src/Ryujinx.Graphics.Vulkan/TextureView.cs

@@ -54,7 +54,7 @@ namespace Ryujinx.Graphics.Vulkan
             gd.Textures.Add(this);
 
             var format = _gd.FormatCapabilities.ConvertToVkFormat(info.Format);
-            var usage = TextureStorage.GetImageUsageFromFormat(info.Format);
+            var usage = TextureStorage.GetImageUsage(info.Format, info.Target, gd.Capabilities.SupportsShaderStorageImageMultisample);
             var levels = (uint)info.Levels;
             var layers = (uint)info.GetLayers();
 

src/Ryujinx.Graphics.Vulkan/VulkanRenderer.cs

@@ -295,6 +295,7 @@ namespace Ryujinx.Graphics.Vulkan
                 supportsSubgroupSizeControl,
                 featuresShaderInt8.ShaderInt8,
                 _physicalDevice.IsDeviceExtensionPresent("VK_EXT_shader_stencil_export"),
+                features2.Features.ShaderStorageImageMultisample,
                 _physicalDevice.IsDeviceExtensionPresent(ExtConditionalRendering.ExtensionName),
                 _physicalDevice.IsDeviceExtensionPresent(ExtExtendedDynamicState.ExtensionName),
                 features2.Features.MultiViewport,