Prefetch capabilities before spawning translation threads. (#3338)

* Prefetch capabilities before spawning translation threads.

The Backend Multithreading only expects one thread to submit commands at a time. When compiling shaders, the translator may request the host GPU capabilities from the backend. It's possible for a bunch of translators to do this at the same time.

There's a caching mechanism in place so that the capabilities are only fetched once. By triggering this before spawning the thread, the async translation threads no longer try to queue onto the backend queue all at the same time.

The Capabilities do need to be checked from the GPU thread, due to OpenGL needing a context to check them, so it's not possible to call the underlying backend directly.

* Initialize the capabilities when setting the GPU thread + missing call in headless

* Remove private variables

Ryujinx.Graphics.GAL/IPipeline.cs

@@ -94,7 +94,7 @@ namespace Ryujinx.Graphics.GAL
         void SetVertexAttribs(ReadOnlySpan<VertexAttribDescriptor> vertexAttribs);
         void SetVertexBuffers(ReadOnlySpan<VertexBufferDescriptor> vertexBuffers);
 
-        void SetViewports(int first, ReadOnlySpan<Viewport> viewports);
+        void SetViewports(int first, ReadOnlySpan<Viewport> viewports, bool disableTransform);
 
         void TextureBarrier();
         void TextureBarrierTiled();

Ryujinx.Graphics.GAL/Multithreading/Commands/SetViewportsCommand.cs

@@ -9,17 +9,19 @@ namespace Ryujinx.Graphics.GAL.Multithreading.Commands
         public CommandType CommandType => CommandType.SetViewports;
         private int _first;
         private SpanRef<Viewport> _viewports;
+        private bool _disableTransform;
 
-        public void Set(int first, SpanRef<Viewport> viewports)
+        public void Set(int first, SpanRef<Viewport> viewports, bool disableTransform)
         {
             _first = first;
             _viewports = viewports;
+            _disableTransform = disableTransform;
         }
 
         public static void Run(ref SetViewportsCommand command, ThreadedRenderer threaded, IRenderer renderer)
         {
             ReadOnlySpan<Viewport> viewports = command._viewports.Get(threaded);
-            renderer.Pipeline.SetViewports(command._first, viewports);
+            renderer.Pipeline.SetViewports(command._first, viewports, command._disableTransform);
             command._viewports.Dispose(threaded);
         }
     }

Ryujinx.Graphics.GAL/Multithreading/ThreadedPipeline.cs

@@ -304,9 +304,9 @@ namespace Ryujinx.Graphics.GAL.Multithreading
             _renderer.QueueCommand();
         }
 
-        public void SetViewports(int first, ReadOnlySpan<Viewport> viewports)
+        public void SetViewports(int first, ReadOnlySpan<Viewport> viewports, bool disableTransform)
         {
-            _renderer.New<SetViewportsCommand>().Set(first, _renderer.CopySpan(viewports));
+            _renderer.New<SetViewportsCommand>().Set(first, _renderer.CopySpan(viewports), disableTransform);
             _renderer.QueueCommand();
         }
 

Ryujinx.Graphics.GAL/SupportBufferUpdater.cs

@@ -72,6 +72,13 @@ namespace Ryujinx.Graphics.GAL
             UpdateGenericField(SupportBuffer.FragmentIsBgraOffset, data, Data.FragmentIsBgra.ToSpan(), offset, count);
         }
 
+        public void UpdateViewportInverse(Vector4<float> data)
+        {
+            Data.ViewportInverse = data;
+
+            MarkDirty(SupportBuffer.ViewportInverseOffset, SupportBuffer.FieldSize);
+        }
+
         public void Commit()
         {
             if (_startOffset != -1)

Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdater.cs

@@ -113,7 +113,8 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
                     nameof(ThreedClassState.DepthMode),
                     nameof(ThreedClassState.ViewportTransform),
                     nameof(ThreedClassState.ViewportExtents),
-                    nameof(ThreedClassState.YControl)),
+                    nameof(ThreedClassState.YControl),
+                    nameof(ThreedClassState.ViewportTransformEnable)),
 
                 new StateUpdateCallbackEntry(UpdatePolygonMode,
                     nameof(ThreedClassState.PolygonModeFront),
@@ -200,7 +201,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
             // of the shader for the new state.
             if (_shaderSpecState != null)
             {
-                if (!_shaderSpecState.MatchesGraphics(_channel, GetPoolState()))
+                if (!_shaderSpecState.MatchesGraphics(_channel, GetPoolState(), GetGraphicsState()))
                 {
                     ForceShaderUpdate();
                 }
@@ -568,6 +569,8 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
             var yControl = _state.State.YControl;
             var face = _state.State.FaceState;
 
+            bool disableTransform = _state.State.ViewportTransformEnable == 0;
+
             UpdateFrontFace(yControl, face.FrontFace);
             UpdateDepthMode();
 
@@ -577,6 +580,17 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
 
             for (int index = 0; index < Constants.TotalViewports; index++)
             {
+                if (disableTransform)
+                {
+                    ref var scissor = ref _state.State.ScreenScissorState;
+
+                    float rScale = _channel.TextureManager.RenderTargetScale;
+                    var scissorRect = new RectangleF(0, 0, (scissor.X + scissor.Width) * rScale, (scissor.Y + scissor.Height) * rScale);
+
+                    viewports[index] = new Viewport(scissorRect, ViewportSwizzle.PositiveX, ViewportSwizzle.PositiveY, ViewportSwizzle.PositiveZ, ViewportSwizzle.PositiveW, 0, 1);
+                    continue;
+                }
+
                 ref var transform = ref _state.State.ViewportTransform[index];
                 ref var extents = ref _state.State.ViewportExtents[index];
 
@@ -628,7 +642,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
                 viewports[index] = new Viewport(region, swizzleX, swizzleY, swizzleZ, swizzleW, depthNear, depthFar);
             }
 
-            _context.Renderer.Pipeline.SetViewports(0, viewports);
+            _context.Renderer.Pipeline.SetViewports(0, viewports, disableTransform);
         }
 
         /// <summary>
@@ -1194,7 +1208,8 @@ namespace Ryujinx.Graphics.Gpu.Engine.Threed
             return new GpuChannelGraphicsState(
                 _state.State.EarlyZForce,
                 _drawState.Topology,
-                _state.State.TessMode);
+                _state.State.TessMode,
+                _state.State.ViewportTransformEnable == 0);
         }
 
         /// <summary>

Ryujinx.Graphics.Gpu/GpuContext.cs

@@ -82,27 +82,13 @@ namespace Ryujinx.Graphics.Gpu
         /// <summary>
         /// Host hardware capabilities.
         /// </summary>
-        internal ref Capabilities Capabilities
-        {
-            get
-            {
-                if (!_capsLoaded)
-                {
-                    _caps = Renderer.GetCapabilities();
-                    _capsLoaded = true;
-                }
-
-                return ref _caps;
-            }
-        }
+        internal Capabilities Capabilities { get; private set; }
 
         /// <summary>
         /// Event for signalling shader cache loading progress.
         /// </summary>
         public event Action<ShaderCacheState, int, int> ShaderCacheStateChanged;
 
-        private bool _capsLoaded;
-        private Capabilities _caps;
         private Thread _gpuThread;
 
         /// <summary>
@@ -254,6 +240,8 @@ namespace Ryujinx.Graphics.Gpu
         public void SetGpuThread()
         {
             _gpuThread = Thread.CurrentThread;
+
+            Capabilities = Renderer.GetCapabilities();
         }
 
         /// <summary>

Ryujinx.Graphics.Gpu/Shader/Cache/Migration.cs

@@ -166,7 +166,8 @@ namespace Ryujinx.Graphics.Gpu.Shader.Cache
                         GpuChannelGraphicsState graphicsState = new GpuChannelGraphicsState(
                             accessorHeader.StateFlags.HasFlag(GuestGpuStateFlags.EarlyZForce),
                             topology,
-                            tessMode);
+                            tessMode,
+                            false);
 
                         TransformFeedbackDescriptor[] tfdNew = null;
 

Ryujinx.Graphics.Gpu/Shader/DiskCache/DiskCacheGpuAccessor.cs

@@ -185,6 +185,12 @@ namespace Ryujinx.Graphics.Gpu.Shader.DiskCache
             return _oldSpecState.GraphicsState.EarlyZForce;
         }
 
+        /// <inheritdoc/>
+        public bool QueryViewportTransformDisable()
+        {
+            return _oldSpecState.GraphicsState.ViewportTransformDisable;
+        }
+
         /// <inheritdoc/>
         public void RegisterTexture(int handle, int cbufSlot)
         {

Ryujinx.Graphics.Gpu/Shader/DiskCache/DiskCacheHostStorage.cs

@@ -21,7 +21,7 @@ namespace Ryujinx.Graphics.Gpu.Shader.DiskCache
         private const ushort FileFormatVersionMajor = 1;
         private const ushort FileFormatVersionMinor = 1;
         private const uint FileFormatVersionPacked = ((uint)FileFormatVersionMajor << 16) | FileFormatVersionMinor;
-        private const uint CodeGenVersion = 0;
+        private const uint CodeGenVersion = 1;
 
         private const string SharedTocFileName = "shared.toc";
         private const string SharedDataFileName = "shared.data";

Ryujinx.Graphics.Gpu/Shader/GpuAccessor.cs

@@ -217,6 +217,12 @@ namespace Ryujinx.Graphics.Gpu.Shader
             return _state.GraphicsState.EarlyZForce;
         }
 
+        /// <inheritdoc/>
+        public bool QueryViewportTransformDisable()
+        {
+            return _state.GraphicsState.ViewportTransformDisable;
+        }
+
         /// <inheritdoc/>
         public void RegisterTexture(int handle, int cbufSlot)
         {

Ryujinx.Graphics.Gpu/Shader/GpuChannelGraphicsState.cs

@@ -25,17 +25,24 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// </summary>
         public readonly TessMode TessellationMode;
 
+        /// <summary>
+        /// Indicates whenever the viewport transform is disabled.
+        /// </summary>
+        public readonly bool ViewportTransformDisable;
+
         /// <summary>
         /// Creates a new GPU graphics state.
         /// </summary>
         /// <param name="earlyZForce">Early Z force enable</param>
         /// <param name="topology">Primitive topology</param>
         /// <param name="tessellationMode">Tessellation mode</param>
-        public GpuChannelGraphicsState(bool earlyZForce, PrimitiveTopology topology, TessMode tessellationMode)
+        /// <param name="viewportTransformDisable">Indicates whenever the viewport transform is disabled</param>
+        public GpuChannelGraphicsState(bool earlyZForce, PrimitiveTopology topology, TessMode tessellationMode, bool viewportTransformDisable)
         {
             EarlyZForce = earlyZForce;
             Topology = topology;
             TessellationMode = tessellationMode;
+            ViewportTransformDisable = viewportTransformDisable;
         }
     }
 }

Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs

@@ -249,12 +249,12 @@ namespace Ryujinx.Graphics.Gpu.Shader
             GpuChannelGraphicsState graphicsState,
             ShaderAddresses addresses)
         {
-            if (_gpPrograms.TryGetValue(addresses, out var gpShaders) && IsShaderEqual(channel, poolState, gpShaders, addresses))
+            if (_gpPrograms.TryGetValue(addresses, out var gpShaders) && IsShaderEqual(channel, poolState, graphicsState, gpShaders, addresses))
             {
                 return gpShaders;
             }
 
-            if (_graphicsShaderCache.TryFind(channel, poolState, addresses, out gpShaders, out var cachedGuestCode))
+            if (_graphicsShaderCache.TryFind(channel, poolState, graphicsState, addresses, out gpShaders, out var cachedGuestCode))
             {
                 _gpPrograms[addresses] = gpShaders;
                 return gpShaders;
@@ -429,12 +429,14 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// </summary>
         /// <param name="channel">GPU channel using the shader</param>
         /// <param name="poolState">GPU channel state to verify shader compatibility</param>
+        /// <param name="graphicsState">GPU channel graphics state to verify shader compatibility</param>
         /// <param name="gpShaders">Cached graphics shaders</param>
         /// <param name="addresses">GPU virtual addresses of all enabled shader stages</param>
         /// <returns>True if the code is different, false otherwise</returns>
         private static bool IsShaderEqual(
             GpuChannel channel,
             GpuChannelPoolState poolState,
+            GpuChannelGraphicsState graphicsState,
             CachedShaderProgram gpShaders,
             ShaderAddresses addresses)
         {
@@ -452,7 +454,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
                 }
             }
 
-            return gpShaders.SpecializationState.MatchesGraphics(channel, poolState);
+            return gpShaders.SpecializationState.MatchesGraphics(channel, poolState, graphicsState);
         }
 
         /// <summary>

Ryujinx.Graphics.Gpu/Shader/ShaderCacheHashTable.cs

@@ -208,6 +208,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// </remarks>
         /// <param name="channel">GPU channel</param>
         /// <param name="poolState">Texture pool state</param>
+        /// <param name="graphicsState">Graphics state</param>
         /// <param name="addresses">Guest addresses of the shaders to find</param>
         /// <param name="program">Cached host program for the given state, if found</param>
         /// <param name="guestCode">Cached guest code, if any found</param>
@@ -215,6 +216,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
         public bool TryFind(
             GpuChannel channel,
             GpuChannelPoolState poolState,
+            GpuChannelGraphicsState graphicsState,
             ShaderAddresses addresses,
             out CachedShaderProgram program,
             out CachedGraphicsGuestCode guestCode)
@@ -234,7 +236,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
 
             if (found && _shaderPrograms.TryGetValue(idTable, out ShaderSpecializationList specList))
             {
-                return specList.TryFindForGraphics(channel, poolState, out program);
+                return specList.TryFindForGraphics(channel, poolState, graphicsState, out program);
             }
 
             return false;

Ryujinx.Graphics.Gpu/Shader/ShaderSpecializationList.cs

@@ -24,13 +24,18 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// </summary>
         /// <param name="channel">GPU channel</param>
         /// <param name="poolState">Texture pool state</param>
+        /// <param name="graphicsState">Graphics state</param>
         /// <param name="program">Cached program, if found</param>
         /// <returns>True if a compatible program is found, false otherwise</returns>
-        public bool TryFindForGraphics(GpuChannel channel, GpuChannelPoolState poolState, out CachedShaderProgram program)
+        public bool TryFindForGraphics(
+            GpuChannel channel,
+            GpuChannelPoolState poolState,
+            GpuChannelGraphicsState graphicsState,
+            out CachedShaderProgram program)
         {
             foreach (var entry in _entries)
             {
-                if (entry.SpecializationState.MatchesGraphics(channel, poolState))
+                if (entry.SpecializationState.MatchesGraphics(channel, poolState, graphicsState))
                 {
                     program = entry;
                     return true;

Ryujinx.Graphics.Gpu/Shader/ShaderSpecializationState.cs

@@ -395,9 +395,15 @@ namespace Ryujinx.Graphics.Gpu.Shader
         /// </summary>
         /// <param name="channel">GPU channel</param>
         /// <param name="poolState">Texture pool state</param>
+        /// <param name="graphicsState">Graphics state</param>
         /// <returns>True if the state matches, false otherwise</returns>
-        public bool MatchesGraphics(GpuChannel channel, GpuChannelPoolState poolState)
+        public bool MatchesGraphics(GpuChannel channel, GpuChannelPoolState poolState, GpuChannelGraphicsState graphicsState)
         {
+            if (graphicsState.ViewportTransformDisable != GraphicsState.ViewportTransformDisable)
+            {
+                return false;
+            }
+
             return Matches(channel, poolState, isCompute: false);
         }
 

Ryujinx.Graphics.OpenGL/Pipeline.cs

@@ -1266,7 +1266,7 @@ namespace Ryujinx.Graphics.OpenGL
             _vertexArray.SetVertexBuffers(vertexBuffers);
         }
 
-        public void SetViewports(int first, ReadOnlySpan<Viewport> viewports)
+        public void SetViewports(int first, ReadOnlySpan<Viewport> viewports, bool disableTransform)
         {
             Array.Resize(ref _viewportArray, viewports.Length * 4);
             Array.Resize(ref _depthRangeArray, viewports.Length * 2);
@@ -1305,6 +1305,19 @@ namespace Ryujinx.Graphics.OpenGL
 
             GL.ViewportArray(first, viewports.Length, viewportArray);
             GL.DepthRangeArray(first, viewports.Length, depthRangeArray);
+
+            float disableTransformF = disableTransform ? 1.0f : 0.0f;
+            if (_supportBuffer.Data.ViewportInverse.W != disableTransformF || disableTransform)
+            {
+                float scale = _renderScale[0].X;
+                _supportBuffer.UpdateViewportInverse(new Vector4<float>
+                {
+                    X = scale * 2f / viewports[first].Region.Width,
+                    Y = scale * 2f / viewports[first].Region.Height,
+                    Z = 1,
+                    W = disableTransformF
+                });
+            }
         }
 
         public void TextureBarrier()

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

@@ -249,7 +249,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                         context.AppendLine();
                     }
                 }
-                else if (isFragment)
+                else if (isFragment || context.Config.Stage == ShaderStage.Vertex)
                 {
                     DeclareSupportUniformBlock(context, context.Config.Stage, 0);
                 }
@@ -615,8 +615,10 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
 
         private static void DeclareSupportUniformBlock(CodeGenContext context, ShaderStage stage, int scaleElements)
         {
-            bool isFragment = stage == ShaderStage.Fragment;
-            if (!isFragment && scaleElements == 0)
+            bool needsSupportBlock = stage == ShaderStage.Fragment || 
+                (context.Config.LastInVertexPipeline && context.Config.GpuAccessor.QueryViewportTransformDisable());
+
+            if (!needsSupportBlock && scaleElements == 0)
             {
                 return;
             }
@@ -630,6 +632,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                 case ShaderStage.Vertex:
                     context.AppendLine($"uint {DefaultNames.SupportBlockAlphaTestName};");
                     context.AppendLine($"bool {DefaultNames.SupportBlockIsBgraName}[{SupportBuffer.FragmentIsBgraCount}];");
+                    context.AppendLine($"vec4 {DefaultNames.SupportBlockViewportInverse};");
                     context.AppendLine($"int {DefaultNames.SupportBlockFragmentScaleCount};");
                     break;
                 case ShaderStage.Compute:

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

@@ -18,6 +18,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
         public const string SupportBlockName = "support_block";
         public const string SupportBlockAlphaTestName = "s_alpha_test";
         public const string SupportBlockIsBgraName = "s_is_bgra";
+        public const string SupportBlockViewportInverse = "s_viewport_inverse";
         public const string SupportBlockFragmentScaleCount = "s_frag_scale_count";
         public const string SupportBlockRenderScaleName = "s_render_scale";
 

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

@@ -84,7 +84,10 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
             { AttributeConsts.FragmentOutputIsBgraBase + 16, new BuiltInAttribute($"{DefaultNames.SupportBlockIsBgraName}[4]",  VariableType.Bool) },
             { AttributeConsts.FragmentOutputIsBgraBase + 20, new BuiltInAttribute($"{DefaultNames.SupportBlockIsBgraName}[5]",  VariableType.Bool) },
             { AttributeConsts.FragmentOutputIsBgraBase + 24, new BuiltInAttribute($"{DefaultNames.SupportBlockIsBgraName}[6]",  VariableType.Bool) },
-            { AttributeConsts.FragmentOutputIsBgraBase + 28, new BuiltInAttribute($"{DefaultNames.SupportBlockIsBgraName}[7]",  VariableType.Bool) }
+            { AttributeConsts.FragmentOutputIsBgraBase + 28, new BuiltInAttribute($"{DefaultNames.SupportBlockIsBgraName}[7]",  VariableType.Bool) },
+
+            { AttributeConsts.SupportBlockViewInverseX,  new BuiltInAttribute($"{DefaultNames.SupportBlockViewportInverse}.x",  VariableType.F32) },
+            { AttributeConsts.SupportBlockViewInverseY,  new BuiltInAttribute($"{DefaultNames.SupportBlockViewportInverse}.y",  VariableType.F32) }
         };
 
         private Dictionary<AstOperand, string> _locals;

Ryujinx.Graphics.Shader/IGpuAccessor.cs

@@ -329,6 +329,15 @@ namespace Ryujinx.Graphics.Shader
             return false;
         }
 
+        /// <summary>
+        /// Queries if host state disables the viewport transform.
+        /// </summary>
+        /// <returns>True if the viewport transform is disabled</returns>
+        bool QueryViewportTransformDisable()
+        {
+            return false;
+        }
+
         /// <summary>
         /// Registers a texture used by the shader.
         /// </summary>

Ryujinx.Graphics.Shader/Instructions/InstEmitAttribute.cs

@@ -206,7 +206,33 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
             if (emit)
             {
-                context.EmitVertex();
+                if (context.Config.LastInVertexPipeline)
+                {
+                    context.PrepareForVertexReturn(out var tempXLocal, out var tempYLocal, out var tempZLocal);
+
+                    context.EmitVertex();
+
+                    // Restore output position value before transformation.
+
+                    if (tempXLocal != null)
+                    {
+                        context.Copy(Attribute(AttributeConsts.PositionX), tempXLocal);
+                    }
+
+                    if (tempYLocal != null)
+                    {
+                        context.Copy(Attribute(AttributeConsts.PositionY), tempYLocal);
+                    }
+
+                    if (tempZLocal != null)
+                    {
+                        context.Copy(Attribute(AttributeConsts.PositionZ), tempZLocal);
+                    }
+                }
+                else
+                {
+                    context.EmitVertex();
+                }
             }
 
             if (cut)

Ryujinx.Graphics.Shader/SupportBuffer.cs

@@ -18,6 +18,7 @@ namespace Ryujinx.Graphics.Shader
 
         public static int FragmentAlphaTestOffset;
         public static int FragmentIsBgraOffset;
+        public static int ViewportInverseOffset;
         public static int FragmentRenderScaleCountOffset;
         public static int GraphicsRenderScaleOffset;
         public static int ComputeRenderScaleOffset;
@@ -40,6 +41,7 @@ namespace Ryujinx.Graphics.Shader
 
             FragmentAlphaTestOffset = OffsetOf(ref instance, ref instance.FragmentAlphaTest);
             FragmentIsBgraOffset = OffsetOf(ref instance, ref instance.FragmentIsBgra);
+            ViewportInverseOffset = OffsetOf(ref instance, ref instance.ViewportInverse);
             FragmentRenderScaleCountOffset = OffsetOf(ref instance, ref instance.FragmentRenderScaleCount);
             GraphicsRenderScaleOffset = OffsetOf(ref instance, ref instance.RenderScale);
             ComputeRenderScaleOffset = GraphicsRenderScaleOffset + FieldSize;
@@ -47,6 +49,7 @@ namespace Ryujinx.Graphics.Shader
 
         public Vector4<int> FragmentAlphaTest;
         public Array8<Vector4<int>> FragmentIsBgra;
+        public Vector4<float> ViewportInverse;
         public Vector4<int> FragmentRenderScaleCount;
 
         // Render scale max count: 1 + 32 + 8. First scale is fragment output scale, others are textures/image inputs.

Ryujinx.Graphics.Shader/Translation/AttributeConsts.cs

@@ -67,6 +67,9 @@ namespace Ryujinx.Graphics.Shader.Translation
         public const int FragmentOutputIsBgraBase = 0x1000100;
         public const int FragmentOutputIsBgraEnd  = FragmentOutputIsBgraBase + 8 * 4;
 
+        public const int SupportBlockViewInverseX = 0x1000200;
+        public const int SupportBlockViewInverseY = 0x1000204;
+
         public const int ThreadIdX = 0x2000000;
         public const int ThreadIdY = 0x2000004;
         public const int ThreadIdZ = 0x2000008;

Ryujinx.Graphics.Shader/Translation/EmitterContext.cs

@@ -154,9 +154,56 @@ namespace Ryujinx.Graphics.Shader.Translation
             return label;
         }
 
+        public void PrepareForVertexReturn()
+        {
+            if (Config.GpuAccessor.QueryViewportTransformDisable())
+            {
+                Operand x = Attribute(AttributeConsts.PositionX | AttributeConsts.LoadOutputMask);
+                Operand y = Attribute(AttributeConsts.PositionY | AttributeConsts.LoadOutputMask);
+                Operand xScale = Attribute(AttributeConsts.SupportBlockViewInverseX);
+                Operand yScale = Attribute(AttributeConsts.SupportBlockViewInverseY);
+                Operand negativeOne = ConstF(-1.0f);
+
+                this.Copy(Attribute(AttributeConsts.PositionX), this.FPFusedMultiplyAdd(x, xScale, negativeOne));
+                this.Copy(Attribute(AttributeConsts.PositionY), this.FPFusedMultiplyAdd(y, yScale, negativeOne));
+            }
+        }
+
+        public void PrepareForVertexReturn(out Operand oldXLocal, out Operand oldYLocal, out Operand oldZLocal)
+        {
+            if (Config.GpuAccessor.QueryViewportTransformDisable())
+            {
+                oldXLocal = Local();
+                this.Copy(oldXLocal, Attribute(AttributeConsts.PositionX | AttributeConsts.LoadOutputMask));
+                oldYLocal = Local();
+                this.Copy(oldYLocal, Attribute(AttributeConsts.PositionY | AttributeConsts.LoadOutputMask));
+            }
+            else
+            {
+                oldXLocal = null;
+                oldYLocal = null;
+            }
+
+            // Will be used by Vulkan backend for depth mode emulation.
+            oldZLocal = null;
+
+            PrepareForVertexReturn();
+        }
+
         public void PrepareForReturn()
         {
-            if (!IsNonMain && Config.Stage == ShaderStage.Fragment)
+            if (IsNonMain)
+            {
+                return;
+            }
+
+            if (Config.LastInVertexPipeline &&
+                (Config.Stage == ShaderStage.Vertex || Config.Stage == ShaderStage.TessellationEvaluation) &&
+                (Config.Options.Flags & TranslationFlags.VertexA) == 0)
+            {
+                PrepareForVertexReturn();
+            }
+            else if (Config.Stage == ShaderStage.Fragment)
             {
                 if (Config.OmapDepth)
                 {

Ryujinx.Graphics.Shader/Translation/ShaderConfig.cs

@@ -14,6 +14,7 @@ namespace Ryujinx.Graphics.Shader.Translation
         public ShaderStage Stage { get; }
 
         public bool GpPassthrough { get; }
+        public bool LastInVertexPipeline { get; private set; }
 
         public int ThreadsPerInputPrimitive { get; }
 
@@ -135,6 +136,7 @@ namespace Ryujinx.Graphics.Shader.Translation
             OmapSampleMask           = header.OmapSampleMask;
             OmapDepth                = header.OmapDepth;
             TransformFeedbackEnabled = gpuAccessor.QueryTransformFeedbackEnabled();
+            LastInVertexPipeline     = header.Stage < ShaderStage.Fragment;
         }
 
         public int GetDepthRegister()
@@ -274,6 +276,11 @@ namespace Ryujinx.Graphics.Shader.Translation
             NextInputAttributesPerPatchComponents = config.ThisInputAttributesPerPatchComponents;
             NextUsesFixedFuncAttributes = config.UsedFeatures.HasFlag(FeatureFlags.FixedFuncAttr);
             MergeOutputUserAttributes(config.UsedInputAttributes, config.UsedInputAttributesPerPatch);
+
+            if (config.Stage != ShaderStage.Fragment)
+            {
+                LastInVertexPipeline = false;
+            }
         }
 
         public void MergeOutputUserAttributes(int mask, int maskPerPatch)

Ryujinx.Headless.SDL2/WindowBase.cs

@@ -164,6 +164,7 @@ namespace Ryujinx.Headless.SDL2
 
             Device.Gpu.Renderer.RunLoop(() =>
             {
+                Device.Gpu.SetGpuThread();
                 Device.Gpu.InitializeShaderCache(_gpuCancellationTokenSource.Token);
                 Translator.IsReadyForTranslation.Set();