Add support for Vertex Program A and other small shader improvements (#192)

* Add WIP support for Vertex Program A, add the FADD_I32 shader instruction, small fix on FFMA_I encoding, nits

* Add separate subroutines for program A/B, and copy attributes to a temp

* Move finalization code to main

* Add new line after flip uniform on the shader

* Handle possible case where VPB uses an output attribute written by VPA but not available on the vbo

* Address PR feedback

Ryujinx.Graphics/Gal/IGalShader.cs

@@ -6,6 +6,8 @@ namespace Ryujinx.Graphics.Gal
     {
         void Create(IGalMemory Memory, long Key, GalShaderType Type);
 
+        void Create(IGalMemory Memory, long VpAPos, long Key, GalShaderType Type);
+
         IEnumerable<ShaderDeclInfo> GetTextureUsage(long Key);
 
         void SetConstBuffer(long Key, int Cbuf, byte[] Data);

Ryujinx.Graphics/Gal/OpenGL/OGLShader.cs

@@ -97,12 +97,37 @@ namespace Ryujinx.Graphics.Gal.OpenGL
 
         public void Create(IGalMemory Memory, long Key, GalShaderType Type)
         {
-            Stages.GetOrAdd(Key, (Stage) => ShaderStageFactory(Memory, Key, Type));
+            Stages.GetOrAdd(Key, (Stage) => ShaderStageFactory(Memory, Key, 0, false, Type));
         }
 
-        private ShaderStage ShaderStageFactory(IGalMemory Memory, long Position, GalShaderType Type)
+        public void Create(IGalMemory Memory, long VpAPos, long Key, GalShaderType Type)
         {
-            GlslProgram Program = GetGlslProgram(Memory, Position, Type);
+            Stages.GetOrAdd(Key, (Stage) => ShaderStageFactory(Memory, VpAPos, Key, true, Type));
+        }
+
+        private ShaderStage ShaderStageFactory(
+            IGalMemory    Memory,
+            long          Position,
+            long          PositionB,
+            bool          IsDualVp,
+            GalShaderType Type)
+        {
+            GlslProgram Program;
+
+            GlslDecompiler Decompiler = new GlslDecompiler();
+
+            if (IsDualVp)
+            {
+                Program = Decompiler.Decompile(
+                    Memory,
+                    Position  + 0x50,
+                    PositionB + 0x50,
+                    Type);
+            }
+            else
+            {
+                Program = Decompiler.Decompile(Memory, Position + 0x50, Type);
+            }
 
             return new ShaderStage(
                 Type,
@@ -111,13 +136,6 @@ namespace Ryujinx.Graphics.Gal.OpenGL
                 Program.Uniforms);
         }
 
-        private GlslProgram GetGlslProgram(IGalMemory Memory, long Position, GalShaderType Type)
-        {
-            GlslDecompiler Decompiler = new GlslDecompiler();
-
-            return Decompiler.Decompile(Memory, Position + 0x50, Type);
-        }
-
         public IEnumerable<ShaderDeclInfo> GetTextureUsage(long Key)
         {
             if (Stages.TryGetValue(Key, out ShaderStage Stage))

Ryujinx.Graphics/Gal/Shader/GlslDecl.cs

@@ -11,31 +11,41 @@ namespace Ryujinx.Graphics.Gal.Shader
         public const int VertexIdAttr    = 0x2fc;
         public const int GlPositionWAttr = 0x7c;
 
+        public const int MaxUboSize = 1024;
+
+        public const int GlPositionVec4Index = 7;
+
         private const int AttrStartIndex = 8;
         private const int TexStartIndex = 8;
 
         public const string PositionOutAttrName = "position";
 
-        private const string InAttrName  = "in_attr";
+        private const string TextureName = "tex";
-        private const string OutAttrName = "out_attr";
         private const string UniformName = "c";
 
-        private const string GprName     = "gpr";
+        private const string AttrName    = "attr";
-        private const string PredName    = "pred";
+        private const string InAttrName  = "in_"  + AttrName;
-        private const string TextureName = "tex";
+        private const string OutAttrName = "out_" + AttrName;
+
+        private const string GprName  = "gpr";
+        private const string PredName = "pred";
 
         public const string FragmentOutputName = "FragColor";
 
         public const string FlipUniformName = "flip";
 
+        public const string ProgramName  = "program";
+        public const string ProgramAName = ProgramName + "_a";
+        public const string ProgramBName = ProgramName + "_b";
+
         private string[] StagePrefixes = new string[] { "vp", "tcp", "tep", "gp", "fp" };
 
         private string StagePrefix;
 
         private Dictionary<int, ShaderDeclInfo> m_Textures;
-
         private Dictionary<int, ShaderDeclInfo> m_Uniforms;
 
+        private Dictionary<int, ShaderDeclInfo> m_Attributes;
         private Dictionary<int, ShaderDeclInfo> m_InAttributes;
         private Dictionary<int, ShaderDeclInfo> m_OutAttributes;
 
@@ -43,9 +53,9 @@ namespace Ryujinx.Graphics.Gal.Shader
         private Dictionary<int, ShaderDeclInfo> m_Preds;
 
         public IReadOnlyDictionary<int, ShaderDeclInfo> Textures => m_Textures;
-
         public IReadOnlyDictionary<int, ShaderDeclInfo> Uniforms => m_Uniforms;
 
+        public IReadOnlyDictionary<int, ShaderDeclInfo> Attributes    => m_Attributes;
         public IReadOnlyDictionary<int, ShaderDeclInfo> InAttributes  => m_InAttributes;
         public IReadOnlyDictionary<int, ShaderDeclInfo> OutAttributes => m_OutAttributes;
 
@@ -54,31 +64,28 @@ namespace Ryujinx.Graphics.Gal.Shader
 
         public GalShaderType ShaderType { get; private set; }
 
-        public GlslDecl(ShaderIrBlock[] Blocks, GalShaderType ShaderType)
+        private GlslDecl(GalShaderType ShaderType)
         {
             this.ShaderType = ShaderType;
 
-            StagePrefix = StagePrefixes[(int)ShaderType] + "_";
-
             m_Uniforms = new Dictionary<int, ShaderDeclInfo>();
-
             m_Textures = new Dictionary<int, ShaderDeclInfo>();
 
+            m_Attributes    = new Dictionary<int, ShaderDeclInfo>();
             m_InAttributes  = new Dictionary<int, ShaderDeclInfo>();
             m_OutAttributes = new Dictionary<int, ShaderDeclInfo>();
 
             m_Gprs  = new Dictionary<int, ShaderDeclInfo>();
             m_Preds = new Dictionary<int, ShaderDeclInfo>();
+        }
+
+        public GlslDecl(ShaderIrBlock[] Blocks, GalShaderType ShaderType) : this(ShaderType)
+        {
+            StagePrefix = StagePrefixes[(int)ShaderType] + "_";
 
             if (ShaderType == GalShaderType.Fragment)
             {
                 m_Gprs.Add(0, new ShaderDeclInfo(FragmentOutputName, 0, 0, 4));
-
-                m_InAttributes.Add(7, new ShaderDeclInfo(PositionOutAttrName, -1, 0, 4));
-            }
-            else
-            {
-                m_OutAttributes.Add(7, new ShaderDeclInfo("gl_Position", -1, 0, 4));
             }
 
             foreach (ShaderIrBlock Block in Blocks)
@@ -90,6 +97,57 @@ namespace Ryujinx.Graphics.Gal.Shader
             }
         }
 
+        public static GlslDecl Merge(GlslDecl VpA, GlslDecl VpB)
+        {
+            GlslDecl Combined = new GlslDecl(GalShaderType.Vertex);
+
+            Merge(Combined.m_Textures, VpA.m_Textures, VpB.m_Textures);
+            Merge(Combined.m_Uniforms, VpA.m_Uniforms, VpB.m_Uniforms);
+
+            Merge(Combined.m_Attributes,    VpA.m_Attributes,    VpB.m_Attributes);
+            Merge(Combined.m_OutAttributes, VpA.m_OutAttributes, VpB.m_OutAttributes);
+
+            Merge(Combined.m_Gprs,  VpA.m_Gprs,  VpB.m_Gprs);
+            Merge(Combined.m_Preds, VpA.m_Preds, VpB.m_Preds);
+
+            //Merge input attributes.
+            foreach (KeyValuePair<int, ShaderDeclInfo> KV in VpA.m_InAttributes)
+            {
+                Combined.m_InAttributes.TryAdd(KV.Key, KV.Value);
+            }
+
+            foreach (KeyValuePair<int, ShaderDeclInfo> KV in VpB.m_InAttributes)
+            {
+                //If Vertex Program A already writes to this attribute,
+                //then we don't need to add it as an input attribute since
+                //Vertex Program A will already have written to it anyway,
+                //and there's no guarantee that there is an input attribute
+                //for this slot.
+                if (!VpA.m_OutAttributes.ContainsKey(KV.Key))
+                {
+                    Combined.m_InAttributes.TryAdd(KV.Key, KV.Value);
+                }
+            }
+
+            return Combined;
+        }
+
+        private static void Merge(
+            Dictionary<int, ShaderDeclInfo> C,
+            Dictionary<int, ShaderDeclInfo> A,
+            Dictionary<int, ShaderDeclInfo> B)
+        {
+            foreach (KeyValuePair<int, ShaderDeclInfo> KV in A)
+            {
+                C.TryAdd(KV.Key, KV.Value);
+            }
+
+            foreach (KeyValuePair<int, ShaderDeclInfo> KV in B)
+            {
+                C.TryAdd(KV.Key, KV.Value);
+            }
+        }
+
         private void Traverse(ShaderIrNode Parent, ShaderIrNode Node)
         {
             switch (Node)
@@ -133,28 +191,14 @@ namespace Ryujinx.Graphics.Gal.Shader
 
                 case ShaderIrOperCbuf Cbuf:
                 {
-                    if (m_Uniforms.TryGetValue(Cbuf.Index, out ShaderDeclInfo DeclInfo))
+                    if (!m_Uniforms.ContainsKey(Cbuf.Index))
-                    {
-                        DeclInfo.SetCbufOffs(Cbuf.Pos);
-                    }
-                    else
                     {
                         string Name = StagePrefix + UniformName + Cbuf.Index;
 
-                        DeclInfo = new ShaderDeclInfo(Name, Cbuf.Pos, Cbuf.Index);
+                        ShaderDeclInfo DeclInfo = new ShaderDeclInfo(Name, Cbuf.Pos, Cbuf.Index);
 
                         m_Uniforms.Add(Cbuf.Index, DeclInfo);
                     }
-
-                    if (Cbuf.Offs != null)
-                    {
-                        //The constant buffer is being accessed as an array,
-                        //we have no way to know the max element it may access in this case.
-                        //Here, we just assume the array size with arbitrary values.
-                        //TODO: Find a better solution for this.
-                        DeclInfo.SetCbufOffs(Cbuf.Pos + 15);
-                    }
-
                     break;
                 }
 
@@ -172,6 +216,11 @@ namespace Ryujinx.Graphics.Gal.Shader
 
                     int GlslIndex = Index - AttrStartIndex;
 
+                    if (GlslIndex < 0)
+                    {
+                        return;
+                    }
+
                     ShaderDeclInfo DeclInfo;
 
                     if (Parent is ShaderIrAsg Asg && Asg.Dst == Node)
@@ -195,6 +244,12 @@ namespace Ryujinx.Graphics.Gal.Shader
 
                     DeclInfo.Enlarge(Elem + 1);
 
+                    if (!m_Attributes.ContainsKey(Index))
+                    {
+                        DeclInfo = new ShaderDeclInfo(AttrName + GlslIndex, GlslIndex, 0, 4);
+
+                        m_Attributes.Add(Index, DeclInfo);
+                    }
                     break;
                 }
 
@@ -224,7 +279,10 @@ namespace Ryujinx.Graphics.Gal.Shader
 
         private bool HasName(Dictionary<int, ShaderDeclInfo> Decls, int Index)
         {
-            int VecIndex = Index >> 2;
+            //This is used to check if the dictionary already contains
+            //a entry for a vector at a given index position.
+            //Used to enable turning gprs into vectors.
+            int VecIndex = Index & ~3;
 
             if (Decls.TryGetValue(VecIndex, out ShaderDeclInfo DeclInfo))
             {

Ryujinx.Graphics/Gal/Shader/GlslDecompiler.cs

@@ -25,7 +25,7 @@ namespace Ryujinx.Graphics.Gal.Shader
 
         private GlslDecl Decl;
 
-        private ShaderIrBlock[] Blocks;
+        private ShaderIrBlock[] Blocks, BlocksB;
 
         private StringBuilder SB;
 
@@ -104,24 +104,63 @@ namespace Ryujinx.Graphics.Gal.Shader
             };
         }
 
+        public GlslProgram Decompile(
+            IGalMemory    Memory,
+            long          VpAPosition,
+            long          VpBPosition,
+            GalShaderType ShaderType)
+        {
+            Blocks  = ShaderDecoder.Decode(Memory, VpAPosition);
+            BlocksB = ShaderDecoder.Decode(Memory, VpBPosition);
+
+            GlslDecl DeclVpA = new GlslDecl(Blocks,  ShaderType);
+            GlslDecl DeclVpB = new GlslDecl(BlocksB, ShaderType);
+
+            Decl = GlslDecl.Merge(DeclVpA, DeclVpB);
+
+            return Decompile();
+        }
+
         public GlslProgram Decompile(IGalMemory Memory, long Position, GalShaderType ShaderType)
         {
-            Blocks = ShaderDecoder.Decode(Memory, Position);
+            Blocks  = ShaderDecoder.Decode(Memory, Position);
+            BlocksB = null;
 
             Decl = new GlslDecl(Blocks, ShaderType);
 
+            return Decompile();
+        }
+
+        private GlslProgram Decompile()
+        {
             SB = new StringBuilder();
 
             SB.AppendLine("#version 410 core");
 
             PrintDeclTextures();
             PrintDeclUniforms();
+            PrintDeclAttributes();
             PrintDeclInAttributes();
             PrintDeclOutAttributes();
             PrintDeclGprs();
             PrintDeclPreds();
 
-            PrintBlockScope(Blocks[0], null, null, "void main()", IdentationStr);
+            if (BlocksB != null)
+            {
+                PrintBlockScope(Blocks[0],  null, null, "void " + GlslDecl.ProgramAName + "()", IdentationStr);
+
+                SB.AppendLine();
+
+                PrintBlockScope(BlocksB[0], null, null, "void " + GlslDecl.ProgramBName + "()", IdentationStr);
+            }
+            else
+            {
+                PrintBlockScope(Blocks[0],  null, null, "void " + GlslDecl.ProgramName + "()", IdentationStr);
+            }
+
+            SB.AppendLine();
+
+            PrintMain();
 
             string GlslCode = SB.ToString();
 
@@ -143,11 +182,15 @@ namespace Ryujinx.Graphics.Gal.Shader
                 SB.AppendLine("uniform vec2 " + GlslDecl.FlipUniformName + ";");
             }
 
+            SB.AppendLine();
+
             foreach (ShaderDeclInfo DeclInfo in Decl.Uniforms.Values.OrderBy(DeclKeySelector))
             {
                 SB.AppendLine($"layout (std140) uniform {DeclInfo.Name} {{");
-                SB.AppendLine($"{IdentationStr}vec4 {DeclInfo.Name}_data[{DeclInfo.Index / 4 + 1}];");
+
-                SB.AppendLine($"}};");
+                SB.AppendLine($"{IdentationStr}vec4 {DeclInfo.Name}_data[{GlslDecl.MaxUboSize}];");
+
+                SB.AppendLine("};");
             }
 
             if (Decl.Uniforms.Count > 0)
@@ -156,6 +199,11 @@ namespace Ryujinx.Graphics.Gal.Shader
             }
         }
 
+        private void PrintDeclAttributes()
+        {
+            PrintDecls(Decl.Attributes);
+        }
+
         private void PrintDeclInAttributes()
         {
             if (Decl.ShaderType == GalShaderType.Fragment)
@@ -244,6 +292,59 @@ namespace Ryujinx.Graphics.Gal.Shader
             return ElemTypes[DeclInfo.Size - 1] + " " + DeclInfo.Name;
         }
 
+        private void PrintMain()
+        {
+            SB.AppendLine("void main() {");
+
+            foreach (KeyValuePair<int, ShaderDeclInfo> KV in Decl.InAttributes)
+            {
+                if (!Decl.Attributes.TryGetValue(KV.Key, out ShaderDeclInfo Attr))
+                {
+                    continue;
+                }
+
+                ShaderDeclInfo DeclInfo = KV.Value;
+
+                string Swizzle = ".xyzw".Substring(0, DeclInfo.Size + 1);
+
+                SB.AppendLine(IdentationStr + Attr.Name + Swizzle + " = " + DeclInfo.Name + ";");
+            }
+
+            if (BlocksB != null)
+            {
+                SB.AppendLine(IdentationStr + GlslDecl.ProgramAName + "();");
+                SB.AppendLine(IdentationStr + GlslDecl.ProgramBName + "();");
+            }
+            else
+            {
+                SB.AppendLine(IdentationStr + GlslDecl.ProgramName + "();");
+            }
+
+            foreach (KeyValuePair<int, ShaderDeclInfo> KV in Decl.OutAttributes)
+            {
+                if (!Decl.Attributes.TryGetValue(KV.Key, out ShaderDeclInfo Attr))
+                {
+                    continue;
+                }
+
+                ShaderDeclInfo DeclInfo = KV.Value;
+
+                string Swizzle = ".xyzw".Substring(0, DeclInfo.Size + 1);
+
+                SB.AppendLine(IdentationStr + DeclInfo.Name + " = " + Attr.Name + Swizzle + ";");
+            }
+
+            if (Decl.ShaderType == GalShaderType.Vertex)
+            {
+                SB.AppendLine(IdentationStr + "gl_Position.xy *= " + GlslDecl.FlipUniformName + ";");
+
+                SB.AppendLine(IdentationStr + GlslDecl.PositionOutAttrName + " = gl_Position;");
+                SB.AppendLine(IdentationStr + GlslDecl.PositionOutAttrName + ".w = 1;");
+            }
+
+            SB.AppendLine("}");
+        }
+
         private void PrintBlockScope(
             ShaderIrBlock Block,
             ShaderIrBlock EndBlock,
@@ -383,18 +484,6 @@ namespace Ryujinx.Graphics.Gal.Shader
 
                         continue;
                     }
-                    else if (Op.Inst == ShaderIrInst.Exit)
-                    {
-                        //Do everything that needs to be done before
-                        //the shader ends here.
-                        if (Decl.ShaderType == GalShaderType.Vertex)
-                        {
-                            SB.AppendLine(Identation + "gl_Position.xy *= flip;");
-
-                            SB.AppendLine(Identation + GlslDecl.PositionOutAttrName + " = gl_Position;");
-                            SB.AppendLine(Identation + GlslDecl.PositionOutAttrName + ".w = 1;");
-                        }
-                    }
 
                     SB.AppendLine(Identation + GetSrcExpr(Op, true) + ";");
                 }
@@ -546,11 +635,12 @@ namespace Ryujinx.Graphics.Gal.Shader
 
         private string GetOutAbufName(ShaderIrOperAbuf Abuf)
         {
-            return GetName(Decl.OutAttributes, Abuf);
+            return GetAttrTempName(Abuf);
         }
 
         private string GetName(ShaderIrOperAbuf Abuf)
         {
+            //Handle special scalar read-only attributes here.
             if (Decl.ShaderType == GalShaderType.Vertex)
             {
                 switch (Abuf.Offs)
@@ -569,20 +659,33 @@ namespace Ryujinx.Graphics.Gal.Shader
                 }
             }
 
-            return GetName(Decl.InAttributes, Abuf);
+            return GetAttrTempName(Abuf);
         }
 
-        private string GetName(IReadOnlyDictionary<int, ShaderDeclInfo> Dict, ShaderIrOperAbuf Abuf)
+        private string GetAttrTempName(ShaderIrOperAbuf Abuf)
         {
             int Index =  Abuf.Offs >> 4;
             int Elem  = (Abuf.Offs >> 2) & 3;
 
-            if (!Dict.TryGetValue(Index, out ShaderDeclInfo DeclInfo))
+            string Swizzle = "." + GetAttrSwizzle(Elem);
+
+            if (!Decl.Attributes.TryGetValue(Index, out ShaderDeclInfo DeclInfo))
             {
+                //Handle special vec4 attributes here
+                //(for example, index 7 is aways gl_Position).
+                if (Index == GlslDecl.GlPositionVec4Index)
+                {
+                    string Name =
+                        Decl.ShaderType != GalShaderType.Vertex &&
+                        Decl.ShaderType != GalShaderType.Geometry ? GlslDecl.PositionOutAttrName : "gl_Position";
+
+                    return Name + Swizzle;
+                }
+
                 throw new InvalidOperationException();
             }
 
-            return DeclInfo.Size > 1 ? DeclInfo.Name + "." + GetAttrSwizzle(Elem) : DeclInfo.Name;
+            return DeclInfo.Name + Swizzle;
         }
 
         private string GetName(ShaderIrOperGpr Gpr)

Ryujinx.Graphics/Gal/Shader/ShaderDecodeAlu.cs

@@ -31,6 +31,24 @@ namespace Ryujinx.Graphics.Gal.Shader
             EmitAluBinaryF(Block, OpCode, ShaderOper.Immf, ShaderIrInst.Fadd);
         }
 
+        public static void Fadd_I32(ShaderIrBlock Block, long OpCode)
+        {
+            ShaderIrNode OperA = GetOperGpr8     (OpCode);
+            ShaderIrNode OperB = GetOperImmf32_20(OpCode);
+
+            bool NegB = ((OpCode >> 53) & 1) != 0;
+            bool AbsA = ((OpCode >> 54) & 1) != 0;
+            bool NegA = ((OpCode >> 56) & 1) != 0;
+            bool AbsB = ((OpCode >> 57) & 1) != 0;
+
+            OperA = GetAluFabsFneg(OperA, AbsA, NegA);
+            OperB = GetAluFabsFneg(OperB, AbsB, NegB);
+
+            ShaderIrOp Op = new ShaderIrOp(ShaderIrInst.Fadd, OperA, OperB);
+
+            Block.AddNode(GetPredNode(new ShaderIrAsg(GetOperGpr0(OpCode), Op), OpCode));
+        }
+
         public static void Fadd_R(ShaderIrBlock Block, long OpCode)
         {
             EmitAluBinaryF(Block, OpCode, ShaderOper.RR, ShaderIrInst.Fadd);

Ryujinx.Graphics/Gal/Shader/ShaderDecodeFlow.cs

@@ -24,7 +24,14 @@ namespace Ryujinx.Graphics.Gal.Shader
 
         public static void Exit(ShaderIrBlock Block, long OpCode)
         {
-            Block.AddNode(GetPredNode(new ShaderIrOp(ShaderIrInst.Exit), OpCode));
+            int CCode = (int)OpCode & 0x1f;
+
+            //TODO: Figure out what the other condition codes mean...
+            if (CCode == 0xf)
+            {
+                Block.AddNode(GetPredNode(new ShaderIrOp(ShaderIrInst.Exit), OpCode));
+            }
+
         }
 
         public static void Kil(ShaderIrBlock Block, long OpCode)

Ryujinx.Graphics/Gal/Shader/ShaderOpCodeTable.cs

@@ -39,14 +39,15 @@ namespace Ryujinx.Graphics.Gal.Shader
             Set("0101110010110x", ShaderDecode.F2i_R);
             Set("0100110001011x", ShaderDecode.Fadd_C);
             Set("0011100x01011x", ShaderDecode.Fadd_I);
+            Set("000010xxxxxxxx", ShaderDecode.Fadd_I32);
             Set("0101110001011x", ShaderDecode.Fadd_R);
             Set("010010011xxxxx", ShaderDecode.Ffma_CR);
-            Set("001100101xxxxx", ShaderDecode.Ffma_I);
+            Set("0011001x1xxxxx", ShaderDecode.Ffma_I);
             Set("010100011xxxxx", ShaderDecode.Ffma_RC);
             Set("010110011xxxxx", ShaderDecode.Ffma_RR);
-            Set("00011110xxxxxx", ShaderDecode.Fmul_I32);
             Set("0100110001101x", ShaderDecode.Fmul_C);
             Set("0011100x01101x", ShaderDecode.Fmul_I);
+            Set("00011110xxxxxx", ShaderDecode.Fmul_I32);
             Set("0101110001101x", ShaderDecode.Fmul_R);
             Set("0100110001100x", ShaderDecode.Fmnmx_C);
             Set("0011100x01100x", ShaderDecode.Fmnmx_I);

Ryujinx.Graphics/Gal/ShaderDeclInfo.cs

@@ -23,13 +23,5 @@ namespace Ryujinx.Graphics.Gal
                 Size = NewSize;
             }
         }
-
-        internal void SetCbufOffs(int Offs)
-        {
-            if (Index < Offs)
-            {
-                Index = Offs;
-            }
-        }
     }
 }

Ryujinx.HLE/Gpu/Engines/NvGpuEngine3d.cs

@@ -126,7 +126,33 @@ namespace Ryujinx.HLE.Gpu.Engines
 
             long BasePosition = MakeInt64From2xInt32(NvGpuEngine3dReg.ShaderAddress);
 
-            for (int Index = 0; Index < 6; Index++)
+            int Index = 1;
+
+            int VpAControl = ReadRegister(NvGpuEngine3dReg.ShaderNControl);
+
+            bool VpAEnable = (VpAControl & 1) != 0;
+
+            if (VpAEnable)
+            {
+                //Note: The maxwell supports 2 vertex programs, usually
+                //only VP B is used, but in some cases VP A is also used.
+                //In this case, it seems to function as an extra vertex
+                //shader stage.
+                //The graphics abstraction layer has a special overload for this
+                //case, which should merge the two shaders into one vertex shader.
+                int VpAOffset = ReadRegister(NvGpuEngine3dReg.ShaderNOffset);
+                int VpBOffset = ReadRegister(NvGpuEngine3dReg.ShaderNOffset + 0x10);
+
+                long VpAPos = BasePosition + (uint)VpAOffset;
+                long VpBPos = BasePosition + (uint)VpBOffset;
+
+                Gpu.Renderer.Shader.Create(Vmm, VpAPos, VpBPos, GalShaderType.Vertex);
+                Gpu.Renderer.Shader.Bind(VpBPos);
+
+                Index = 2;
+            }
+
+            for (; Index < 6; Index++)
             {
                 int Control = ReadRegister(NvGpuEngine3dReg.ShaderNControl + Index * 0x10);
                 int Offset  = ReadRegister(NvGpuEngine3dReg.ShaderNOffset  + Index * 0x10);