Initial support for shader half float instructions (#507)

Ryujinx.Graphics/Gal/Shader/GlslDecl.cs

@@ -63,6 +63,7 @@ namespace Ryujinx.Graphics.Gal.Shader
         private Dictionary<int, ShaderDeclInfo> m_OutAttributes;
 
         private Dictionary<int, ShaderDeclInfo> m_Gprs;
+        private Dictionary<int, ShaderDeclInfo> m_GprsHalf;
         private Dictionary<int, ShaderDeclInfo> m_Preds;
 
         public IReadOnlyDictionary<ShaderIrOp, ShaderDeclInfo> CbTextures => m_CbTextures;
@@ -75,6 +76,7 @@ namespace Ryujinx.Graphics.Gal.Shader
         public IReadOnlyDictionary<int, ShaderDeclInfo> OutAttributes => m_OutAttributes;
 
         public IReadOnlyDictionary<int, ShaderDeclInfo> Gprs     => m_Gprs;
+        public IReadOnlyDictionary<int, ShaderDeclInfo> GprsHalf => m_GprsHalf;
         public IReadOnlyDictionary<int, ShaderDeclInfo> Preds    => m_Preds;
 
         public GalShaderType ShaderType { get; private set; }
@@ -93,6 +95,7 @@ namespace Ryujinx.Graphics.Gal.Shader
             m_OutAttributes = new Dictionary<int, ShaderDeclInfo>();
 
             m_Gprs     = new Dictionary<int, ShaderDeclInfo>();
+            m_GprsHalf = new Dictionary<int, ShaderDeclInfo>();
             m_Preds    = new Dictionary<int, ShaderDeclInfo>();
         }
 
@@ -147,6 +150,7 @@ namespace Ryujinx.Graphics.Gal.Shader
             Merge(Combined.m_OutAttributes, VpA.m_OutAttributes, VpB.m_OutAttributes);
 
             Merge(Combined.m_Gprs,     VpA.m_Gprs,     VpB.m_Gprs);
+            Merge(Combined.m_GprsHalf, VpA.m_GprsHalf, VpB.m_GprsHalf);
             Merge(Combined.m_Preds,    VpA.m_Preds,    VpB.m_Preds);
 
             //Merge input attributes.
@@ -343,8 +347,21 @@ namespace Ryujinx.Graphics.Gal.Shader
                     {
                         string Name = GetGprName(Gpr.Index);
 
+                        if (Gpr.RegisterSize == ShaderRegisterSize.Single)
+                        {
                             m_Gprs.TryAdd(Gpr.Index, new ShaderDeclInfo(Name, Gpr.Index));
                         }
+                        else if (Gpr.RegisterSize == ShaderRegisterSize.Half)
+                        {
+                            Name += "_h" + Gpr.HalfPart;
+
+                            m_GprsHalf.TryAdd((Gpr.Index << 1) | Gpr.HalfPart, new ShaderDeclInfo(Name, Gpr.Index));
+                        }
+                        else /* if (Gpr.RegisterSize == ShaderRegisterSize.Double) */
+                        {
+                            throw new NotImplementedException("Double types are not supported.");
+                        }
+                    }
                     break;
                 }
 

Ryujinx.Graphics/Gal/Shader/GlslDecompiler.cs

@@ -364,6 +364,7 @@ namespace Ryujinx.Graphics.Gal.Shader
         private void PrintDeclGprs()
         {
             PrintDecls(Decl.Gprs);
+            PrintDecls(Decl.GprsHalf);
         }
 
         private void PrintDeclPreds()
@@ -897,7 +898,23 @@ namespace Ryujinx.Graphics.Gal.Shader
 
         private string GetName(ShaderIrOperGpr Gpr)
         {
-            return Gpr.IsConst ? "0" : GetNameWithSwizzle(Decl.Gprs, Gpr.Index);
+            if (Gpr.IsConst)
+            {
+                return "0";
+            }
+
+            if (Gpr.RegisterSize == ShaderRegisterSize.Single)
+            {
+                return GetNameWithSwizzle(Decl.Gprs, Gpr.Index);
+            }
+            else if (Gpr.RegisterSize == ShaderRegisterSize.Half)
+            {
+                return GetNameWithSwizzle(Decl.GprsHalf, (Gpr.Index << 1) | Gpr.HalfPart);
+            }
+            else /* if (Gpr.RegisterSize == ShaderRegisterSize.Double) */
+            {
+                throw new NotImplementedException("Double types are not supported.");
+            }
         }
 
         private string GetValue(ShaderIrOperImm Imm)

Ryujinx.Graphics/Gal/Shader/ShaderDecodeAlu.cs

@@ -6,6 +6,14 @@ namespace Ryujinx.Graphics.Gal.Shader
 {
     static partial class ShaderDecode
     {
+        private enum HalfOutputType
+        {
+            PackedFp16,
+            Fp32,
+            MergeH0,
+            MergeH1
+        }
+
         public static void Bfe_C(ShaderIrBlock Block, long OpCode, int Position)
         {
             EmitBfe(Block, OpCode, ShaderOper.CR);
@@ -144,6 +152,16 @@ namespace Ryujinx.Graphics.Gal.Shader
             EmitFsetp(Block, OpCode, ShaderOper.RR);
         }
 
+        public static void Hadd2_R(ShaderIrBlock Block, long OpCode, int Position)
+        {
+            EmitBinaryHalfOp(Block, OpCode, ShaderIrInst.Fadd);
+        }
+
+        public static void Hmul2_R(ShaderIrBlock Block, long OpCode, int Position)
+        {
+            EmitBinaryHalfOp(Block, OpCode, ShaderIrInst.Fmul);
+        }
+
         public static void Iadd_C(ShaderIrBlock Block, long OpCode, int Position)
         {
             EmitIadd(Block, OpCode, ShaderOper.CR);
@@ -1041,6 +1059,47 @@ namespace Ryujinx.Graphics.Gal.Shader
             Block.AddNode(OpCode.PredNode(new ShaderIrAsg(P0Node, Op)));
         }
 
+        private static void EmitBinaryHalfOp(ShaderIrBlock Block, long OpCode, ShaderIrInst Inst)
+        {
+            bool AbsB = OpCode.Read(30);
+            bool NegB = OpCode.Read(31);
+            bool Sat  = OpCode.Read(32);
+            bool AbsA = OpCode.Read(44);
+
+            ShaderIrOperGpr[] VecA = OpCode.GprHalfVec8();
+            ShaderIrOperGpr[] VecB = OpCode.GprHalfVec20();
+
+            HalfOutputType OutputType = (HalfOutputType)OpCode.Read(49, 3);
+
+            int Elems = OutputType == HalfOutputType.PackedFp16 ? 2 : 1;
+            int First = OutputType == HalfOutputType.MergeH1    ? 1 : 0;
+
+            for (int Index = First; Index < Elems; Index++)
+            {
+                ShaderIrNode OperA = GetAluFabs    (VecA[Index], AbsA);
+                ShaderIrNode OperB = GetAluFabsFneg(VecB[Index], AbsB, NegB);
+
+                ShaderIrNode Op = new ShaderIrOp(Inst, OperA, OperB);
+
+                ShaderIrOperGpr Dst = GetHalfDst(OpCode, OutputType, Index);
+
+                Block.AddNode(OpCode.PredNode(new ShaderIrAsg(Dst, GetAluFsat(Op, Sat))));
+            }
+        }
+
+        private static ShaderIrOperGpr GetHalfDst(long OpCode, HalfOutputType OutputType, int Index)
+        {
+            switch (OutputType)
+            {
+                case HalfOutputType.PackedFp16: return OpCode.GprHalf0(Index);
+                case HalfOutputType.Fp32:       return OpCode.Gpr0();
+                case HalfOutputType.MergeH0:    return OpCode.GprHalf0(0);
+                case HalfOutputType.MergeH1:    return OpCode.GprHalf0(1);
+            }
+
+            throw new ArgumentException(nameof(OutputType));
+        }
+
         private static void EmitLop(ShaderIrBlock Block, long OpCode, ShaderOper Oper)
         {
             int SubOp = OpCode.Read(41, 3);

Ryujinx.Graphics/Gal/Shader/ShaderDecodeMem.cs

@@ -6,8 +6,6 @@ namespace Ryujinx.Graphics.Gal.Shader
 {
     static partial class ShaderDecode
     {
-        private const int TempRegStart = 0x100;
-
         private const int ____ = 0x0;
         private const int R___ = 0x1;
         private const int _G__ = 0x2;
@@ -149,14 +147,18 @@ namespace Ryujinx.Graphics.Gal.Shader
 
             for (int Index = 0; Index < Coords.Length; Index++)
             {
-                Coords[Index] = OpCode.Gpr8();
+                ShaderIrOperGpr CoordReg = OpCode.Gpr8();
 
-                Coords[Index].Index += Index;
+                CoordReg.Index += Index;
 
-                if (Coords[Index].Index > ShaderIrOperGpr.ZRIndex)
+                if (!CoordReg.IsValidRegister)
                 {
-                    Coords[Index].Index = ShaderIrOperGpr.ZRIndex;
+                    CoordReg.Index = ShaderIrOperGpr.ZRIndex;
                 }
+
+                Coords[Index] = ShaderIrOperGpr.MakeTemporary(Index);
+
+                Block.AddNode(new ShaderIrAsg(Coords[Index], CoordReg));
             }
 
             int ChMask = OpCode.Read(31, 0xf);
@@ -167,17 +169,6 @@ namespace Ryujinx.Graphics.Gal.Shader
 
             ShaderIrInst Inst = GprHandle ? ShaderIrInst.Texb : ShaderIrInst.Texs;
 
-            for (int Ch = 0; Ch < 4; Ch++)
-            {
-                ShaderIrOperGpr Dst = new ShaderIrOperGpr(TempRegStart + Ch);
-
-                ShaderIrMetaTex Meta = new ShaderIrMetaTex(Ch);
-
-                ShaderIrOp Op = new ShaderIrOp(Inst, Coords[0], Coords[1], OperC, Meta);
-
-                Block.AddNode(OpCode.PredNode(new ShaderIrAsg(Dst, Op)));
-            }
-
             int RegInc = 0;
 
             for (int Ch = 0; Ch < 4; Ch++)
@@ -187,18 +178,20 @@ namespace Ryujinx.Graphics.Gal.Shader
                     continue;
                 }
 
-                ShaderIrOperGpr Src = new ShaderIrOperGpr(TempRegStart + Ch);
-
                 ShaderIrOperGpr Dst = OpCode.Gpr0();
 
                 Dst.Index += RegInc++;
 
-                if (Dst.Index >= ShaderIrOperGpr.ZRIndex)
+                if (!Dst.IsValidRegister || Dst.IsConst)
                 {
                     continue;
                 }
 
-                Block.AddNode(OpCode.PredNode(new ShaderIrAsg(Dst, Src)));
+                ShaderIrMetaTex Meta = new ShaderIrMetaTex(Ch);
+
+                ShaderIrOp Op = new ShaderIrOp(Inst, Coords[0], Coords[1], OperC, Meta);
+
+                Block.AddNode(OpCode.PredNode(new ShaderIrAsg(Dst, Op)));
             }
         }
 
@@ -215,57 +208,81 @@ namespace Ryujinx.Graphics.Gal.Shader
         private static void EmitTexs(ShaderIrBlock Block, long OpCode, ShaderIrInst Inst)
         {
             //TODO: Support other formats.
-            ShaderIrNode OperA = OpCode.Gpr8();
-            ShaderIrNode OperB = OpCode.Gpr20();
-            ShaderIrNode OperC = OpCode.Imm13_36();
-
             int LutIndex;
 
-            LutIndex  = OpCode.Gpr0 ().Index != ShaderIrOperGpr.ZRIndex ? 1 : 0;
+            LutIndex  = !OpCode.Gpr0().IsConst  ? 1 : 0;
-            LutIndex |= OpCode.Gpr28().Index != ShaderIrOperGpr.ZRIndex ? 2 : 0;
+            LutIndex |= !OpCode.Gpr28().IsConst ? 2 : 0;
 
             if (LutIndex == 0)
             {
-                //Both registers are RZ, color is not written anywhere.
+                //Both destination registers are RZ, do nothing.
-                //So, the intruction is basically a no-op.
                 return;
             }
 
-            int ChMask = MaskLut[LutIndex, OpCode.Read(50, 7)];
+            bool Fp16 = !OpCode.Read(59);
 
-            for (int Ch = 0; Ch < 4; Ch++)
+            int DstIncrement = 0;
-            {
-                ShaderIrOperGpr Dst = new ShaderIrOperGpr(TempRegStart + Ch);
-
-                ShaderIrMetaTex Meta = new ShaderIrMetaTex(Ch);
-
-                ShaderIrOp Op = new ShaderIrOp(Inst, OperA, OperB, OperC, Meta);
-
-                Block.AddNode(OpCode.PredNode(new ShaderIrAsg(Dst, Op)));
-            }
-
-            int RegInc = 0;
 
             ShaderIrOperGpr GetDst()
             {
                 ShaderIrOperGpr Dst;
 
+                if (Fp16)
+                {
+                    //FP16 mode, two components are packed on the two
+                    //halfs of a 32-bits register, as two half-float values.
+                    int HalfPart = DstIncrement & 1;
+
+                    switch (LutIndex)
+                    {
+                        case 1: Dst = OpCode.GprHalf0(HalfPart);  break;
+                        case 2: Dst = OpCode.GprHalf28(HalfPart); break;
+                        case 3: Dst = (DstIncrement >> 1) != 0
+                            ? OpCode.GprHalf28(HalfPart)
+                            : OpCode.GprHalf0(HalfPart); break;
+
+                        default: throw new InvalidOperationException();
+                    }
+                }
+                else
+                {
+                    //32-bits mode, each component uses one register.
+                    //Two components uses two consecutive registers.
                     switch (LutIndex)
                     {
                         case 1: Dst = OpCode.Gpr0();  break;
                         case 2: Dst = OpCode.Gpr28(); break;
-                    case 3: Dst = (RegInc >> 1) != 0
+                        case 3: Dst = (DstIncrement >> 1) != 0
                             ? OpCode.Gpr28()
-                        : OpCode.Gpr0 (); break;
+                            : OpCode.Gpr0(); break;
 
                         default: throw new InvalidOperationException();
                     }
 
-                Dst.Index += RegInc++ & 1;
+                    Dst.Index += DstIncrement & 1;
+                }
+
+                DstIncrement++;
 
                 return Dst;
             }
 
+            int ChMask = MaskLut[LutIndex, OpCode.Read(50, 7)];
+
+            if (ChMask == 0)
+            {
+                //All channels are disabled, do nothing.
+                return;
+            }
+
+            ShaderIrNode OperC = OpCode.Imm13_36();
+
+            ShaderIrOperGpr Coord0 = ShaderIrOperGpr.MakeTemporary(0);
+            ShaderIrOperGpr Coord1 = ShaderIrOperGpr.MakeTemporary(1);
+
+            Block.AddNode(new ShaderIrAsg(Coord0, OpCode.Gpr8()));
+            Block.AddNode(new ShaderIrAsg(Coord1, OpCode.Gpr20()));
+
             for (int Ch = 0; Ch < 4; Ch++)
             {
                 if (!IsChannelUsed(ChMask, Ch))
@@ -273,13 +290,15 @@ namespace Ryujinx.Graphics.Gal.Shader
                     continue;
                 }
 
-                ShaderIrOperGpr Src = new ShaderIrOperGpr(TempRegStart + Ch);
+                ShaderIrMetaTex Meta = new ShaderIrMetaTex(Ch);
+
+                ShaderIrOp Op = new ShaderIrOp(Inst, Coord0, Coord1, OperC, Meta);
 
                 ShaderIrOperGpr Dst = GetDst();
 
-                if (Dst.Index != ShaderIrOperGpr.ZRIndex)
+                if (Dst.IsValidRegister && !Dst.IsConst)
                 {
-                    Block.AddNode(OpCode.PredNode(new ShaderIrAsg(Dst, Src)));
+                    Block.AddNode(OpCode.PredNode(new ShaderIrAsg(Dst, Op)));
                 }
             }
         }

Ryujinx.Graphics/Gal/Shader/ShaderDecodeOpCode.cs

@@ -75,6 +75,49 @@ namespace Ryujinx.Graphics.Gal.Shader
             return new ShaderIrOperGpr(OpCode.Read(28, 0xff));
         }
 
+        private static ShaderIrOperGpr[] GprHalfVec8(this long OpCode)
+        {
+            return GetGprHalfVec2(OpCode.Read(8, 0xff), OpCode.Read(47, 3));
+        }
+
+        private static ShaderIrOperGpr[] GprHalfVec20(this long OpCode)
+        {
+            return GetGprHalfVec2(OpCode.Read(20, 0xff), OpCode.Read(28, 3));
+        }
+
+        private static ShaderIrOperGpr[] GetGprHalfVec2(int Gpr, int Mask)
+        {
+            if (Mask == 1)
+            {
+                //This value is used for FP32, the whole 32-bits register
+                //is used as each element on the vector.
+                return new ShaderIrOperGpr[]
+                {
+                    new ShaderIrOperGpr(Gpr),
+                    new ShaderIrOperGpr(Gpr)
+                };
+            }
+
+            ShaderIrOperGpr Low  = new ShaderIrOperGpr(Gpr, 0);
+            ShaderIrOperGpr High = new ShaderIrOperGpr(Gpr, 1);
+
+            return new ShaderIrOperGpr[]
+            {
+                (Mask & 1) != 0 ? High : Low,
+                (Mask & 2) != 0 ? High : Low
+            };
+        }
+
+        private static ShaderIrOperGpr GprHalf0(this long OpCode, int HalfPart)
+        {
+            return new ShaderIrOperGpr(OpCode.Read(0, 0xff), HalfPart);
+        }
+
+        private static ShaderIrOperGpr GprHalf28(this long OpCode, int HalfPart)
+        {
+            return new ShaderIrOperGpr(OpCode.Read(28, 0xff), HalfPart);
+        }
+
         private static ShaderIrOperImm Imm5_39(this long OpCode)
         {
             return new ShaderIrOperImm(OpCode.Read(39, 0x1f));

Ryujinx.Graphics/Gal/Shader/ShaderIrOperGpr.cs

@@ -6,13 +6,26 @@ namespace Ryujinx.Graphics.Gal.Shader
 
         public bool IsConst => Index == ZRIndex;
 
-        public bool IsValidRegister => (Index <= ZRIndex);
+        public bool IsValidRegister => (uint)Index <= ZRIndex;
 
         public int Index    { get; set; }
+        public int HalfPart { get; set; }
+
+        public ShaderRegisterSize RegisterSize { get; private set; }
 
         public ShaderIrOperGpr(int Index)
         {
             this.Index = Index;
+
+            RegisterSize = ShaderRegisterSize.Single;
+        }
+
+        public ShaderIrOperGpr(int Index, int HalfPart)
+        {
+            this.Index    = Index;
+            this.HalfPart = HalfPart;
+
+            RegisterSize = ShaderRegisterSize.Half;
         }
 
         public static ShaderIrOperGpr MakeTemporary(int Index = 0)

Ryujinx.Graphics/Gal/Shader/ShaderOpCodeTable.cs

@@ -58,6 +58,8 @@ namespace Ryujinx.Graphics.Gal.Shader
             Set("010010111011xx", ShaderDecode.Fsetp_C);
             Set("0011011x1011xx", ShaderDecode.Fsetp_I);
             Set("010110111011xx", ShaderDecode.Fsetp_R);
+            Set("0101110100010x", ShaderDecode.Hadd2_R);
+            Set("0101110100001x", ShaderDecode.Hmul2_R);
             Set("0100110010111x", ShaderDecode.I2f_C);
             Set("0011100x10111x", ShaderDecode.I2f_I);
             Set("0101110010111x", ShaderDecode.I2f_R);
@@ -118,7 +120,7 @@ namespace Ryujinx.Graphics.Gal.Shader
             Set("110000xxxx111x", ShaderDecode.Tex);
             Set("1101111010111x", ShaderDecode.Tex_B);
             Set("1101111101001x", ShaderDecode.Texq);
-            Set("1101100xxxxxxx", ShaderDecode.Texs);
+            Set("1101x00xxxxxxx", ShaderDecode.Texs);
             Set("1101101xxxxxxx", ShaderDecode.Tlds);
             Set("01011111xxxxxx", ShaderDecode.Vmad);
             Set("0100111xxxxxxx", ShaderDecode.Xmad_CR);

Ryujinx.Graphics/Gal/Shader/ShaderRegisterSize.cs

@@ -0,0 +1,9 @@
+namespace Ryujinx.Graphics.Gal.Shader
+{
+    enum ShaderRegisterSize
+    {
+        Half,
+        Single,
+        Double
+    }
+}

Ryujinx.Graphics/Gal/ShaderDumper.cs

@@ -39,7 +39,7 @@ namespace Ryujinx.Graphics.Gal
                 ulong Instruction = 0;
 
                 //Dump until a NOP instruction is found
-                while ((Instruction >> 52 & 0xfff8) != 0x50b0)
+                while ((Instruction >> 48 & 0xfff8) != 0x50b0)
                 {
                     uint Word0 = (uint)Memory.ReadInt32(Position + 0x50 + Offset + 0);
                     uint Word1 = (uint)Memory.ReadInt32(Position + 0x50 + Offset + 4);