A32: Implement VCVTT, VCVTB (#3710 )

* A32: Implement VCVTT, VCVTB * A32: F16C implementation of VCVTT/VCVTB
A64: Add fast path for Fcvtas_Gp/S/V, Fcvtau_Gp/S/V and Frinta_S/V in… (#3712 )
2022-10-19 02:36:04 +02:00 · 2022-10-19 00:21:33 +00:00
10 changed files with 453 additions and 46 deletions
--- a/ARMeilleure/Decoders/OpCode32SimdCvtTB.cs
+++ b/ARMeilleure/Decoders/OpCode32SimdCvtTB.cs
@@ -0,0 +1,44 @@
+namespace ARMeilleure.Decoders
+{
+    class OpCode32SimdCvtTB : OpCode32, IOpCode32Simd
+    {
+        public int Vd { get; }
+        public int Vm { get; }
+        public bool Op { get; } // Convert to Half / Convert from Half
+        public bool T { get; } // Top / Bottom
+        public int Size { get; } // Double / Single
+
+        public new static OpCode Create(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdCvtTB(inst, address, opCode, false);
+        public static OpCode CreateT32(InstDescriptor inst, ulong address, int opCode) => new OpCode32SimdCvtTB(inst, address, opCode, true);
+
+        public OpCode32SimdCvtTB(InstDescriptor inst, ulong address, int opCode, bool isThumb) : base(inst, address, opCode)
+        {
+            IsThumb = isThumb;
+
+            Op   = ((opCode >> 16) & 0x1) != 0;
+            T    = ((opCode >> 7) & 0x1) != 0;
+            Size = ((opCode >> 8) & 0x1);
+
+            RegisterSize = Size == 1 ? RegisterSize.Int64 : RegisterSize.Int32;
+
+            if (Size == 1)
+            {
+                if (Op)
+                {
+                    Vm = ((opCode >> 1) & 0x10) | ((opCode >> 0) & 0xf);
+                    Vd = ((opCode >> 22) & 0x1) | ((opCode >> 11) & 0x1e);
+                }
+                else
+                {
+                    Vm = ((opCode >> 5) & 0x1) | ((opCode << 1) & 0x1e);
+                    Vd = ((opCode >> 18) & 0x10) | ((opCode >> 12) & 0xf);
+                }
+            }
+            else
+            {
+                Vm = ((opCode >> 5) & 0x1) | ((opCode << 1) & 0x1e);
+                Vd = ((opCode >> 22) & 0x1) | ((opCode >> 11) & 0x1e);
+            }
+        }
+    }
+}
--- a/ARMeilleure/Decoders/OpCodeTable.cs
+++ b/ARMeilleure/Decoders/OpCodeTable.cs
@@ -828,6 +828,7 @@ namespace ARMeilleure.Decoders
            SetVfp("<<<<11101x11110xxxxx101x11x0xxxx", InstName.Vcvt,   InstEmit32.Vcvt_FI,  OpCode32SimdCvtFI.Create,       OpCode32SimdCvtFI.CreateT32); // FP32 to int.
            SetVfp("<<<<11101x111000xxxx101xx1x0xxxx", InstName.Vcvt,   InstEmit32.Vcvt_FI,  OpCode32SimdCvtFI.Create,       OpCode32SimdCvtFI.CreateT32); // Int to FP32.
            SetVfp("111111101x1111xxxxxx101xx1x0xxxx", InstName.Vcvt,   InstEmit32.Vcvt_RM,  OpCode32SimdCvtFI.Create,       OpCode32SimdCvtFI.CreateT32); // The many FP32 to int encodings (fp).
+            SetVfp("<<<<11101x11001xxxxx101xx1x0xxxx", InstName.Vcvt,   InstEmit32.Vcvt_TB,  OpCode32SimdCvtTB.Create,       OpCode32SimdCvtTB.CreateT32);
            SetVfp("<<<<11101x00xxxxxxxx101xx0x0xxxx", InstName.Vdiv,   InstEmit32.Vdiv_S,   OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
            SetVfp("<<<<11101xx0xxxxxxxx1011x0x10000", InstName.Vdup,   InstEmit32.Vdup,     OpCode32SimdDupGP.Create,       OpCode32SimdDupGP.CreateT32);
            SetVfp("<<<<11101x10xxxxxxxx101xx0x0xxxx", InstName.Vfma,   InstEmit32.Vfma_S,   OpCode32SimdRegS.Create,        OpCode32SimdRegS.CreateT32);
--- a/ARMeilleure/Instructions/InstEmitSimdArithmetic.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdArithmetic.cs
@@ -1617,18 +1617,32 @@ namespace ARMeilleure.Instructions

        public static void Frinta_S(ArmEmitterContext context)
        {
-            EmitScalarUnaryOpF(context, (op1) =>
+            if (Optimizations.UseSse41)
            {
-                return EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1);
-            });
+                EmitSse41ScalarRoundOpF(context, FPRoundingMode.ToNearestAway);
+            }
+            else
+            {
+                EmitScalarUnaryOpF(context, (op1) =>
+                {
+                    return EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1);
+                });
+            }
        }

        public static void Frinta_V(ArmEmitterContext context)
        {
-            EmitVectorUnaryOpF(context, (op1) =>
+            if (Optimizations.UseSse41)
            {
-                return EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1);
-            });
+                EmitSse41VectorRoundOpF(context, FPRoundingMode.ToNearestAway);
+            }
+            else
+            {
+                EmitVectorUnaryOpF(context, (op1) =>
+                {
+                    return EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1);
+                });
+            }
        }

        public static void Frinti_S(ArmEmitterContext context)
@@ -3516,9 +3530,18 @@ namespace ARMeilleure.Instructions

            Operand n = GetVec(op.Rn);

-            Intrinsic inst = (op.Size & 1) != 0 ? Intrinsic.X86Roundsd : Intrinsic.X86Roundss;
+            Operand res;

-            Operand res = context.AddIntrinsic(inst, n, Const(X86GetRoundControl(roundMode)));
+            if (roundMode != FPRoundingMode.ToNearestAway)
+            {
+                Intrinsic inst = (op.Size & 1) != 0 ? Intrinsic.X86Roundsd : Intrinsic.X86Roundss;
+
+                res = context.AddIntrinsic(inst, n, Const(X86GetRoundControl(roundMode)));
+            }
+            else
+            {
+                res = EmitSse41RoundToNearestWithTiesToAwayOpF(context, n, scalar: true);
+            }

            if ((op.Size & 1) != 0)
            {
@@ -3538,9 +3561,18 @@ namespace ARMeilleure.Instructions

            Operand n = GetVec(op.Rn);

-            Intrinsic inst = (op.Size & 1) != 0 ? Intrinsic.X86Roundpd : Intrinsic.X86Roundps;
+            Operand res;

-            Operand res = context.AddIntrinsic(inst, n, Const(X86GetRoundControl(roundMode)));
+            if (roundMode != FPRoundingMode.ToNearestAway)
+            {
+                Intrinsic inst = (op.Size & 1) != 0 ? Intrinsic.X86Roundpd : Intrinsic.X86Roundps;
+
+                res = context.AddIntrinsic(inst, n, Const(X86GetRoundControl(roundMode)));
+            }
+            else
+            {
+                res = EmitSse41RoundToNearestWithTiesToAwayOpF(context, n, scalar: false);
+            }

            if (op.RegisterSize == RegisterSize.Simd64)
            {
--- a/ARMeilleure/Instructions/InstEmitSimdCvt.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdCvt.cs
@@ -164,32 +164,74 @@ namespace ARMeilleure.Instructions

        public static void Fcvtas_Gp(ArmEmitterContext context)
        {
-            EmitFcvt_s_Gp(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1));
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvts_Gp(context, FPRoundingMode.ToNearestAway, isFixed: false);
+            }
+            else
+            {
+                EmitFcvt_s_Gp(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1));
+            }
        }

        public static void Fcvtas_S(ArmEmitterContext context)
        {
-            EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: true, scalar: true);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41FcvtsOpF(context, FPRoundingMode.ToNearestAway, scalar: true);
+            }
+            else
+            {
+                EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: true, scalar: true);
+            }
        }

        public static void Fcvtas_V(ArmEmitterContext context)
        {
-            EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: true, scalar: false);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41FcvtsOpF(context, FPRoundingMode.ToNearestAway, scalar: false);
+            }
+            else
+            {
+                EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: true, scalar: false);
+            }
        }

        public static void Fcvtau_Gp(ArmEmitterContext context)
        {
-            EmitFcvt_u_Gp(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1));
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41Fcvtu_Gp(context, FPRoundingMode.ToNearestAway, isFixed: false);
+            }
+            else
+            {
+                EmitFcvt_u_Gp(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1));
+            }
        }

        public static void Fcvtau_S(ArmEmitterContext context)
        {
-            EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: false, scalar: true);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41FcvtuOpF(context, FPRoundingMode.ToNearestAway, scalar: true);
+            }
+            else
+            {
+                EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: false, scalar: true);
+            }
        }

        public static void Fcvtau_V(ArmEmitterContext context)
        {
-            EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: false, scalar: false);
+            if (Optimizations.UseSse41)
+            {
+                EmitSse41FcvtuOpF(context, FPRoundingMode.ToNearestAway, scalar: false);
+            }
+            else
+            {
+                EmitFcvt(context, (op1) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, op1), signed: false, scalar: false);
+            }
        }

        public static void Fcvtl_V(ArmEmitterContext context)
@@ -1223,7 +1265,14 @@ namespace ARMeilleure.Instructions
                    nRes = context.AddIntrinsic(Intrinsic.X86Mulps, nRes, fpScaledMask);
                }

-                nRes = context.AddIntrinsic(Intrinsic.X86Roundps, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundps, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar);
+                }

                Operand nInt = context.AddIntrinsic(Intrinsic.X86Cvtps2dq, nRes);

@@ -1265,7 +1314,14 @@ namespace ARMeilleure.Instructions
                    nRes = context.AddIntrinsic(Intrinsic.X86Mulpd, nRes, fpScaledMask);
                }

-                nRes = context.AddIntrinsic(Intrinsic.X86Roundpd, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundpd, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar);
+                }

                Operand nLong = EmitSse2CvtDoubleToInt64OpF(context, nRes, scalar);

@@ -1314,7 +1370,14 @@ namespace ARMeilleure.Instructions
                    nRes = context.AddIntrinsic(Intrinsic.X86Mulps, nRes, fpScaledMask);
                }

-                nRes = context.AddIntrinsic(Intrinsic.X86Roundps, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundps, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar);
+                }

                Operand zero = context.VectorZero();

@@ -1369,7 +1432,14 @@ namespace ARMeilleure.Instructions
                    nRes = context.AddIntrinsic(Intrinsic.X86Mulpd, nRes, fpScaledMask);
                }

-                nRes = context.AddIntrinsic(Intrinsic.X86Roundpd, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundpd, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar);
+                }

                Operand zero = context.VectorZero();

@@ -1424,7 +1494,14 @@ namespace ARMeilleure.Instructions
                    nRes = context.AddIntrinsic(Intrinsic.X86Mulss, nRes, fpScaledMask);
                }

-                nRes = context.AddIntrinsic(Intrinsic.X86Roundss, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundss, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar: true);
+                }

                Operand nIntOrLong = op.RegisterSize == RegisterSize.Int32
                    ? context.AddIntrinsicInt (Intrinsic.X86Cvtss2si, nRes)
@@ -1464,7 +1541,14 @@ namespace ARMeilleure.Instructions
                    nRes = context.AddIntrinsic(Intrinsic.X86Mulsd, nRes, fpScaledMask);
                }

-                nRes = context.AddIntrinsic(Intrinsic.X86Roundsd, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundsd, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar: true);
+                }

                Operand nIntOrLong = op.RegisterSize == RegisterSize.Int32
                    ? context.AddIntrinsicInt (Intrinsic.X86Cvtsd2si, nRes)
@@ -1512,7 +1596,14 @@ namespace ARMeilleure.Instructions
                    nRes = context.AddIntrinsic(Intrinsic.X86Mulss, nRes, fpScaledMask);
                }

-                nRes = context.AddIntrinsic(Intrinsic.X86Roundss, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundss, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar: true);
+                }

                Operand zero = context.VectorZero();

@@ -1567,7 +1658,14 @@ namespace ARMeilleure.Instructions
                    nRes = context.AddIntrinsic(Intrinsic.X86Mulsd, nRes, fpScaledMask);
                }

-                nRes = context.AddIntrinsic(Intrinsic.X86Roundsd, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundsd, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar: true);
+                }

                Operand zero = context.VectorZero();

--- a/ARMeilleure/Instructions/InstEmitSimdCvt32.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdCvt32.cs
@@ -203,6 +203,9 @@ namespace ARMeilleure.Instructions
            FPRoundingMode roundMode;
            switch (rm)
            {
+                case 0b00:
+                    roundMode = FPRoundingMode.ToNearestAway;
+                    break;
                case 0b01:
                    roundMode = FPRoundingMode.ToNearest;
                    break;
@@ -228,7 +231,7 @@ namespace ARMeilleure.Instructions
            bool unsigned = op.Opc == 0;
            int rm = op.Opc2 & 3;

-            if (Optimizations.UseSse41 && rm != 0b00)
+            if (Optimizations.UseSse41)
            {
                EmitSse41ConvertInt32(context, RMToRoundMode(rm), !unsigned);
            }
@@ -258,6 +261,68 @@ namespace ARMeilleure.Instructions
            }
        }

+        public static void Vcvt_TB(ArmEmitterContext context)
+        {
+            OpCode32SimdCvtTB op = (OpCode32SimdCvtTB)context.CurrOp;
+
+            if (Optimizations.UseF16c)
+            {
+                Debug.Assert(!Optimizations.ForceLegacySse);
+
+                if (op.Op)
+                {
+                    Operand res = ExtractScalar(context, op.Size == 1 ? OperandType.FP64 : OperandType.FP32, op.Vm);
+                    if (op.Size == 1)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Cvtsd2ss, context.VectorZero(), res);
+                    }
+                    res = context.AddIntrinsic(Intrinsic.X86Vcvtps2ph, res, Const(X86GetRoundControl(FPRoundingMode.ToNearest)));
+                    res = context.VectorExtract16(res, 0);
+                    InsertScalar16(context, op.Vd, op.T, res);
+                }
+                else
+                {
+                    Operand res = context.VectorCreateScalar(ExtractScalar16(context, op.Vm, op.T));
+                    res = context.AddIntrinsic(Intrinsic.X86Vcvtph2ps, res);
+                    if (op.Size == 1)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Cvtss2sd, context.VectorZero(), res);
+                    }
+                    res = context.VectorExtract(op.Size == 1 ? OperandType.I64 : OperandType.I32, res, 0);
+                    InsertScalar(context, op.Vd, res);
+                }
+            }
+            else
+            {
+                if (op.Op)
+                {
+                    // Convert to half
+
+                    Operand src = ExtractScalar(context, op.Size == 1 ? OperandType.FP64 : OperandType.FP32, op.Vm);
+
+                    MethodInfo method = op.Size == 1
+                        ? typeof(SoftFloat64_16).GetMethod(nameof(SoftFloat64_16.FPConvert))
+                        : typeof(SoftFloat32_16).GetMethod(nameof(SoftFloat32_16.FPConvert));
+                    Operand res = context.Call(method, src);
+
+                    InsertScalar16(context, op.Vd, op.T, res);
+                }
+                else
+                {
+                    // Convert from half
+
+                    Operand src = ExtractScalar16(context, op.Vm, op.T);
+
+                    MethodInfo method = op.Size == 1
+                        ? typeof(SoftFloat16_64).GetMethod(nameof(SoftFloat16_64.FPConvert))
+                        : typeof(SoftFloat16_32).GetMethod(nameof(SoftFloat16_32.FPConvert));
+                    Operand res = context.Call(method, src);
+
+                    InsertScalar(context, op.Vd, res);
+                }
+            }
+        }
+
        // VRINTA/M/N/P (floating-point).
        public static void Vrint_RM(ArmEmitterContext context)
        {
@@ -267,15 +332,21 @@ namespace ARMeilleure.Instructions

            int rm = op.Opc2 & 3;

-            if (Optimizations.UseSse2 && rm != 0b00)
+            if (Optimizations.UseSse41)
            {
                EmitScalarUnaryOpSimd32(context, (m) =>
                {
-                    Intrinsic inst = (op.Size & 1) == 0 ? Intrinsic.X86Roundss : Intrinsic.X86Roundsd;
-
                    FPRoundingMode roundMode = RMToRoundMode(rm);

-                    return context.AddIntrinsic(inst, m, Const(X86GetRoundControl(roundMode)));
+                    if (roundMode != FPRoundingMode.ToNearestAway)
+                    {
+                        Intrinsic inst = (op.Size & 1) == 0 ? Intrinsic.X86Roundss : Intrinsic.X86Roundsd;
+                        return context.AddIntrinsic(inst, m, Const(X86GetRoundControl(roundMode)));
+                    }
+                    else
+                    {
+                        return EmitSse41RoundToNearestWithTiesToAwayOpF(context, m, scalar: true);
+                    }
                });
            }
            else
@@ -305,7 +376,17 @@ namespace ARMeilleure.Instructions
        // VRINTA (vector).
        public static void Vrinta_V(ArmEmitterContext context)
        {
-            EmitVectorUnaryOpF32(context, (m) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, m));
+            if (Optimizations.UseSse41)
+            {
+                EmitVectorUnaryOpSimd32(context, (m) =>
+                {
+                    return EmitSse41RoundToNearestWithTiesToAwayOpF(context, m, scalar: false);
+                });
+            }
+            else
+            {
+                EmitVectorUnaryOpF32(context, (m) => EmitRoundMathCall(context, MidpointRounding.AwayFromZero, m));
+            }
        }

        // VRINTM (vector).
@@ -413,7 +494,14 @@ namespace ARMeilleure.Instructions
                Operand nRes = context.AddIntrinsic(Intrinsic.X86Cmpss, n, n, Const((int)CmpCondition.OrderedQ));
                nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, n);

-                nRes = context.AddIntrinsic(Intrinsic.X86Roundss, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundss, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar: true);
+                }

                Operand zero = context.VectorZero();

@@ -464,7 +552,14 @@ namespace ARMeilleure.Instructions
                Operand nRes = context.AddIntrinsic(Intrinsic.X86Cmpsd, n, n, Const((int)CmpCondition.OrderedQ));
                nRes = context.AddIntrinsic(Intrinsic.X86Pand, nRes, n);

-                nRes = context.AddIntrinsic(Intrinsic.X86Roundsd, nRes, Const(X86GetRoundControl(roundMode)));
+                if (roundMode != FPRoundingMode.ToNearestAway)
+                {
+                    nRes = context.AddIntrinsic(Intrinsic.X86Roundsd, nRes, Const(X86GetRoundControl(roundMode)));
+                }
+                else
+                {
+                    nRes = EmitSse41RoundToNearestWithTiesToAwayOpF(context, nRes, scalar: true);
+                }

                Operand zero = context.VectorZero();

--- a/ARMeilleure/Instructions/InstEmitSimdHelper.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdHelper.cs
@@ -33,6 +33,14 @@ namespace ARMeilleure.Instructions
        };

        public static readonly long ZeroMask = 128L << 56 | 128L << 48 | 128L << 40 | 128L << 32 | 128L << 24 | 128L << 16 | 128L << 8 | 128L << 0;
+
+        public static ulong X86GetGf2p8LogicalShiftLeft(int shift)
+        {
+            ulong identity = (0b00000001UL << 56) | (0b00000010UL << 48) | (0b00000100UL << 40) | (0b00001000UL << 32) |
+                             (0b00010000UL << 24) | (0b00100000UL << 16) | (0b01000000UL <<  8) | (0b10000000UL <<  0);
+
+            return shift >= 0 ? identity >> (shift * 8) : identity << (-shift * 8);
+        }
 #endregion

 #region "X86 SSE Intrinsics"
@@ -243,19 +251,44 @@ namespace ARMeilleure.Instructions
            throw new ArgumentException($"Invalid rounding mode \"{roundMode}\".");
        }

-        public static ulong X86GetGf2p8LogicalShiftLeft(int shift)
+        public static Operand EmitSse41RoundToNearestWithTiesToAwayOpF(ArmEmitterContext context, Operand n, bool scalar)
        {
-            ulong identity = 
-                (0b00000001UL << 56) |
-                (0b00000010UL << 48) |
-                (0b00000100UL << 40) |
-                (0b00001000UL << 32) |
-                (0b00010000UL << 24) |
-                (0b00100000UL << 16) |
-                (0b01000000UL <<  8) |
-                (0b10000000UL <<  0);
+            Debug.Assert(n.Type == OperandType.V128);

-            return shift >= 0 ? identity >> (shift * 8) : identity << (-shift * 8);
+            Operand nCopy = context.Copy(n);
+
+            Operand rC = Const(X86GetRoundControl(FPRoundingMode.TowardsZero));
+
+            IOpCodeSimd op = (IOpCodeSimd)context.CurrOp;
+
+            if ((op.Size & 1) == 0)
+            {
+                Operand signMask = scalar ? X86GetScalar(context, int.MinValue) : X86GetAllElements(context, int.MinValue);
+                        signMask = context.AddIntrinsic(Intrinsic.X86Pand, signMask, nCopy);
+
+                // 0x3EFFFFFF == BitConverter.SingleToInt32Bits(0.5f) - 1
+                Operand valueMask = scalar ? X86GetScalar(context, 0x3EFFFFFF) : X86GetAllElements(context, 0x3EFFFFFF);
+                        valueMask = context.AddIntrinsic(Intrinsic.X86Por, valueMask, signMask);
+
+                nCopy = context.AddIntrinsic(scalar ? Intrinsic.X86Addss : Intrinsic.X86Addps, nCopy, valueMask);
+
+                nCopy = context.AddIntrinsic(scalar ? Intrinsic.X86Roundss : Intrinsic.X86Roundps, nCopy, rC);
+            }
+            else
+            {
+                Operand signMask = scalar ? X86GetScalar(context, long.MinValue) : X86GetAllElements(context, long.MinValue);
+                        signMask = context.AddIntrinsic(Intrinsic.X86Pand, signMask, nCopy);
+
+                // 0x3FDFFFFFFFFFFFFFL == BitConverter.DoubleToInt64Bits(0.5d) - 1L
+                Operand valueMask = scalar ? X86GetScalar(context, 0x3FDFFFFFFFFFFFFFL) : X86GetAllElements(context, 0x3FDFFFFFFFFFFFFFL);
+                        valueMask = context.AddIntrinsic(Intrinsic.X86Por, valueMask, signMask);
+
+                nCopy = context.AddIntrinsic(scalar ? Intrinsic.X86Addsd : Intrinsic.X86Addpd, nCopy, valueMask);
+
+                nCopy = context.AddIntrinsic(scalar ? Intrinsic.X86Roundsd : Intrinsic.X86Roundpd, nCopy, rC);
+            }
+
+            return nCopy;
        }

        public static Operand EmitCountSetBits8(ArmEmitterContext context, Operand op) // "size" is 8 (SIMD&FP Inst.).
--- a/ARMeilleure/Instructions/InstEmitSimdHelper32.cs
+++ b/ARMeilleure/Instructions/InstEmitSimdHelper32.cs
@@ -70,6 +70,22 @@ namespace ARMeilleure.Instructions
            context.Copy(vec, insert);
        }

+        public static Operand ExtractScalar16(ArmEmitterContext context, int reg, bool top)
+        {
+            return context.VectorExtract16(GetVecA32(reg >> 2), ((reg & 3) << 1) | (top ? 1 : 0));
+        }
+
+        public static void InsertScalar16(ArmEmitterContext context, int reg, bool top, Operand value)
+        {
+            Debug.Assert(value.Type == OperandType.FP32 || value.Type == OperandType.I32);
+
+            Operand vec, insert;
+            vec = GetVecA32(reg >> 2);
+            insert = context.VectorInsert16(vec, value, ((reg & 3) << 1) | (top ? 1 : 0));
+
+            context.Copy(vec, insert);
+        }
+
        public static Operand ExtractElement(ArmEmitterContext context, int reg, int size, bool signed)
        {
            return EmitVectorExtract32(context, reg >> (4 - size), reg & ((16 >> size) - 1), size, signed);
--- a/ARMeilleure/State/FPRoundingMode.cs
+++ b/ARMeilleure/State/FPRoundingMode.cs
@@ -2,9 +2,10 @@ namespace ARMeilleure.State
 {
    public enum FPRoundingMode
    {
-        ToNearest            = 0,
+        ToNearest            = 0, // With ties to even.
        TowardsPlusInfinity  = 1,
        TowardsMinusInfinity = 2,
-        TowardsZero          = 3
+        TowardsZero          = 3,
+        ToNearestAway        = 4 // With ties to away.
    }
 }
--- a/ARMeilleure/Translation/PTC/Ptc.cs
+++ b/ARMeilleure/Translation/PTC/Ptc.cs
@@ -27,7 +27,7 @@ namespace ARMeilleure.Translation.PTC
        private const string OuterHeaderMagicString = "PTCohd\0\0";
        private const string InnerHeaderMagicString = "PTCihd\0\0";

-        private const uint InternalVersion = 3710; //! To be incremented manually for each change to the ARMeilleure project.
+        private const uint InternalVersion = 3713; //! To be incremented manually for each change to the ARMeilleure project.

        private const string ActualDir = "0";
        private const string BackupDir = "1";
--- a/Ryujinx.Tests/Cpu/CpuTestSimdCvt32.cs
+++ b/Ryujinx.Tests/Cpu/CpuTestSimdCvt32.cs
@@ -339,6 +339,93 @@ namespace Ryujinx.Tests.Cpu

            CompareAgainstUnicorn();
        }
+
+        [Explicit]
+        [Test, Pairwise, Description("VCVT<top>.F16.F32 <Sd>, <Dm>")]
+        public void Vcvt_F32_F16([Values(0u, 1u, 2u, 3u)] uint rd,
+                                 [Values(0u, 1u, 2u, 3u)] uint rm,
+                                 [ValueSource(nameof(_1S_))] [Random(RndCnt)] uint s0,
+                                 [ValueSource(nameof(_1S_))] [Random(RndCnt)] uint s1,
+                                 [ValueSource(nameof(_1S_))] [Random(RndCnt)] uint s2,
+                                 [ValueSource(nameof(_1S_))] [Random(RndCnt)] uint s3,
+                                 [Values] bool top)
+        {
+            uint opcode = 0xeeb30a40; // VCVTB.F16.F32 S0, D0
+
+            if (top)
+            {
+                opcode |= 1 << 7;
+            }
+
+            opcode |= ((rd & 0x1e) << 11) | ((rd & 0x1) << 22);
+            opcode |= ((rm & 0x1e) >> 1) | ((rm & 0x1) << 5);
+
+            V128 v0 = MakeVectorE0E1E2E3(s0, s1, s2, s3);
+
+            SingleOpcode(opcode, v0: v0);
+
+            CompareAgainstUnicorn();
+        }
+
+        [Explicit]
+        [Test, Pairwise, Description("VCVT<top>.F16.F64 <Sd>, <Dm>")]
+        public void Vcvt_F64_F16([Values(0u, 1u, 2u, 3u)] uint rd,
+                                 [Values(0u, 1u)] uint rm,
+                                 [ValueSource(nameof(_1D_F_))] ulong d0,
+                                 [ValueSource(nameof(_1D_F_))] ulong d1,
+                                 [Values] bool top)
+        {
+            uint opcode = 0xeeb30b40; // VCVTB.F16.F64 S0, D0
+
+            if (top)
+            {
+                opcode |= 1 << 7;
+            }
+
+            opcode |= ((rd & 0x1e) << 11) | ((rd & 0x1) << 22);
+            opcode |= ((rm & 0xf) << 0) | ((rm & 0x10) << 1);
+
+            V128 v0 = MakeVectorE0E1(d0, d1);
+
+            SingleOpcode(opcode, v0: v0);
+
+            CompareAgainstUnicorn();
+        }
+
+        [Explicit]
+        [Test, Pairwise, Description("VCVT<top>.F<size>.F16 <Vd>, <Sm>")]
+        public void Vcvt_F16_Fx([Values(0u, 1u, 2u, 3u)] uint rd,
+                                 [Values(0u, 1u, 2u, 3u)] uint rm,
+                                 [ValueSource(nameof(_1D_F_))] ulong d0,
+                                 [ValueSource(nameof(_1D_F_))] ulong d1,
+                                 [Values] bool top,
+                                 [Values] bool sz)
+        {
+            uint opcode = 0xeeb20a40; // VCVTB.F32.F16 S0, S0
+
+            if (top)
+            {
+                opcode |= 1 << 7;
+            }
+
+            if (sz)
+            {
+                opcode |= 1 << 8;
+                opcode |= ((rd & 0xf) << 12) | ((rd & 0x10) << 18);
+            }
+            else
+            {
+                opcode |= ((rd & 0x1e) << 11) | ((rd & 0x1) << 22);
+            }
+
+            opcode |= ((rm & 0xf) << 0) | ((rm & 0x10) << 1);
+
+            V128 v0 = MakeVectorE0E1(d0, d1);
+
+            SingleOpcode(opcode, v0: v0);
+
+            CompareAgainstUnicorn();
+        }
 #endif
    }
 }