Implement OpenDataStorageWithProgramIndex partially (#3765)

* Implement OpenDataStorageWithProgramIndex partially

* Was not supposed to change this

ARMeilleure/CodeGen/X86/AssemblerTable.cs

@@ -113,6 +113,7 @@ namespace ARMeilleure.CodeGen.X86
             Add(X86Instruction.Divps,         new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x00000f5e, InstructionFlags.Vex));
             Add(X86Instruction.Divsd,         new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x00000f5e, InstructionFlags.Vex | InstructionFlags.PrefixF2));
             Add(X86Instruction.Divss,         new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x00000f5e, InstructionFlags.Vex | InstructionFlags.PrefixF3));
+            Add(X86Instruction.Gf2p8affineqb, new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f3ace, InstructionFlags.Prefix66));
             Add(X86Instruction.Haddpd,        new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x00000f7c, InstructionFlags.Vex | InstructionFlags.Prefix66));
             Add(X86Instruction.Haddps,        new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x00000f7c, InstructionFlags.Vex | InstructionFlags.PrefixF2));
             Add(X86Instruction.Idiv,          new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x070000f7, InstructionFlags.None));

ARMeilleure/CodeGen/X86/HardwareCapabilities.cs

@@ -20,8 +20,9 @@ namespace ARMeilleure.CodeGen.X86
 
             if (maxNum >= 7)
             {
-                (_, int ebx7, _, _) = X86Base.CpuId(0x00000007, 0x00000000);
+                (_, int ebx7, int ecx7, _) = X86Base.CpuId(0x00000007, 0x00000000);
                 FeatureInfo7Ebx = (FeatureFlags7Ebx)ebx7;
+                FeatureInfo7Ecx = (FeatureFlags7Ecx)ecx7;
             }
         }
 
@@ -54,9 +55,16 @@ namespace ARMeilleure.CodeGen.X86
             Sha = 1 << 29
         }
 
+        [Flags]
+        public enum FeatureFlags7Ecx
+        {
+            Gfni = 1 << 8,
+        }
+
         public static FeatureFlags1Edx FeatureInfo1Edx { get; }
         public static FeatureFlags1Ecx FeatureInfo1Ecx { get; }
         public static FeatureFlags7Ebx FeatureInfo7Ebx { get; } = 0;
+        public static FeatureFlags7Ecx FeatureInfo7Ecx { get; } = 0;
 
         public static bool SupportsSse => FeatureInfo1Edx.HasFlag(FeatureFlags1Edx.Sse);
         public static bool SupportsSse2 => FeatureInfo1Edx.HasFlag(FeatureFlags1Edx.Sse2);
@@ -72,6 +80,7 @@ namespace ARMeilleure.CodeGen.X86
         public static bool SupportsAvx2 => FeatureInfo7Ebx.HasFlag(FeatureFlags7Ebx.Avx2) && SupportsAvx;
         public static bool SupportsF16c => FeatureInfo1Ecx.HasFlag(FeatureFlags1Ecx.F16c);
         public static bool SupportsSha => FeatureInfo7Ebx.HasFlag(FeatureFlags7Ebx.Sha);
+        public static bool SupportsGfni => FeatureInfo7Ecx.HasFlag(FeatureFlags7Ecx.Gfni);
 
         public static bool ForceLegacySse { get; set; }
 

ARMeilleure/CodeGen/X86/IntrinsicTable.cs

@@ -58,6 +58,7 @@ namespace ARMeilleure.CodeGen.X86
             Add(Intrinsic.X86Divps,         new IntrinsicInfo(X86Instruction.Divps,         IntrinsicType.Binary));
             Add(Intrinsic.X86Divsd,         new IntrinsicInfo(X86Instruction.Divsd,         IntrinsicType.Binary));
             Add(Intrinsic.X86Divss,         new IntrinsicInfo(X86Instruction.Divss,         IntrinsicType.Binary));
+            Add(Intrinsic.X86Gf2p8affineqb, new IntrinsicInfo(X86Instruction.Gf2p8affineqb, IntrinsicType.TernaryImm));
             Add(Intrinsic.X86Haddpd,        new IntrinsicInfo(X86Instruction.Haddpd,        IntrinsicType.Binary));
             Add(Intrinsic.X86Haddps,        new IntrinsicInfo(X86Instruction.Haddps,        IntrinsicType.Binary));
             Add(Intrinsic.X86Insertps,      new IntrinsicInfo(X86Instruction.Insertps,      IntrinsicType.TernaryImm));

ARMeilleure/CodeGen/X86/X86Instruction.cs

@@ -54,6 +54,7 @@ namespace ARMeilleure.CodeGen.X86
         Divps,
         Divsd,
         Divss,
+        Gf2p8affineqb,
         Haddpd,
         Haddps,
         Idiv,

ARMeilleure/Instructions/InstEmitSimdArithmetic.cs

@@ -726,7 +726,7 @@ namespace ARMeilleure.Instructions
             {
                 EmitVectorAcrossVectorOpF(context, (op1, op2) =>
                 {
-                    return context.Call(typeof(SoftFloat32).GetMethod(nameof(SoftFloat32.FPMaxNum)), op1, op2);
+                    return EmitSoftFloatCall(context, nameof(SoftFloat32.FPMaxNum), op1, op2);
                 });
             }
         }
@@ -774,7 +774,7 @@ namespace ARMeilleure.Instructions
             {
                 EmitVectorAcrossVectorOpF(context, (op1, op2) =>
                 {
-                    return context.Call(typeof(SoftFloat32).GetMethod(nameof(SoftFloat32.FPMax)), op1, op2);
+                    return EmitSoftFloatCall(context, nameof(SoftFloat32.FPMax), op1, op2);
                 });
             }
         }
@@ -900,7 +900,7 @@ namespace ARMeilleure.Instructions
             {
                 EmitVectorAcrossVectorOpF(context, (op1, op2) =>
                 {
-                    return context.Call(typeof(SoftFloat32).GetMethod(nameof(SoftFloat32.FPMinNum)), op1, op2);
+                    return EmitSoftFloatCall(context, nameof(SoftFloat32.FPMinNum), op1, op2);
                 });
             }
         }
@@ -948,7 +948,7 @@ namespace ARMeilleure.Instructions
             {
                 EmitVectorAcrossVectorOpF(context, (op1, op2) =>
                 {
-                    return context.Call(typeof(SoftFloat32).GetMethod(nameof(SoftFloat32.FPMin)), op1, op2);
+                    return EmitSoftFloatCall(context, nameof(SoftFloat32.FPMin), op1, op2);
                 });
             }
         }
@@ -1633,37 +1633,17 @@ namespace ARMeilleure.Instructions
 
         public static void Frinti_S(ArmEmitterContext context)
         {
-            OpCodeSimd op = (OpCodeSimd)context.CurrOp;
-
             EmitScalarUnaryOpF(context, (op1) =>
             {
-                if (op.Size == 0)
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.RoundF)), op1);
-                }
-                else /* if (op.Size == 1) */
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Round)), op1);
-                }
+                return EmitRoundByRMode(context, op1);
             });
         }
 
         public static void Frinti_V(ArmEmitterContext context)
         {
-            OpCodeSimd op = (OpCodeSimd)context.CurrOp;
-
-            int sizeF = op.Size & 1;
-
             EmitVectorUnaryOpF(context, (op1) =>
             {
-                if (sizeF == 0)
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.RoundF)), op1);
-                }
-                else /* if (sizeF == 1) */
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Round)), op1);
-                }
+                return EmitRoundByRMode(context, op1);
             });
         }
 
@@ -1759,37 +1739,17 @@ namespace ARMeilleure.Instructions
 
         public static void Frintx_S(ArmEmitterContext context)
         {
-            OpCodeSimd op = (OpCodeSimd)context.CurrOp;
-
             EmitScalarUnaryOpF(context, (op1) =>
             {
-                if (op.Size == 0)
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.RoundF)), op1);
-                }
-                else /* if (op.Size == 1) */
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Round)), op1);
-                }
+                return EmitRoundByRMode(context, op1);
             });
         }
 
         public static void Frintx_V(ArmEmitterContext context)
         {
-            OpCodeSimd op = (OpCodeSimd)context.CurrOp;
-
-            int sizeF = op.Size & 1;
-
             EmitVectorUnaryOpF(context, (op1) =>
             {
-                if (sizeF == 0)
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.RoundF)), op1);
-                }
-                else /* if (sizeF == 1) */
-                {
-                    return context.Call(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Round)), op1);
-                }
+                return EmitRoundByRMode(context, op1);
             });
         }
 

ARMeilleure/Instructions/InstEmitSimdCmp32.cs

@@ -3,7 +3,6 @@ using ARMeilleure.IntermediateRepresentation;
 using ARMeilleure.State;
 using ARMeilleure.Translation;
 using System;
-using System.Reflection;
 
 using static ARMeilleure.Instructions.InstEmitHelper;
 using static ARMeilleure.Instructions.InstEmitSimdHelper;
@@ -178,37 +177,20 @@ namespace ARMeilleure.Instructions
 
         private static void EmitCmpOpF32(ArmEmitterContext context, string name, bool zero)
         {
-            Operand one = Const(1);
             if (zero)
             {
                 EmitVectorUnaryOpF32(context, (m) =>
                 {
-                    OperandType type = m.Type;
+                    Operand zeroOp = m.Type == OperandType.FP64 ? ConstF(0.0d) : ConstF(0.0f);
 
-                    if (type == OperandType.FP64)
-                    {
-                        return context.Call(typeof(SoftFloat64).GetMethod(name), m, ConstF(0.0d), one);
-                    }
-                    else
-                    {
-                        return context.Call(typeof(SoftFloat32).GetMethod(name), m, ConstF(0.0f), one);
-                    }
+                    return EmitSoftFloatCallDefaultFpscr(context, name, m, zeroOp);
                 });
             }
             else
             {
                 EmitVectorBinaryOpF32(context, (n, m) =>
                 {
-                    OperandType type = n.Type;
-
-                    if (type == OperandType.FP64)
-                    {
-                        return context.Call(typeof(SoftFloat64).GetMethod(name), n, m, one);
-                    }
-                    else
-                    {
-                        return context.Call(typeof(SoftFloat32).GetMethod(name), n, m, one);
-                    }
+                    return EmitSoftFloatCallDefaultFpscr(context, name, n, m);
                 });
             }
         }
@@ -357,11 +339,7 @@ namespace ARMeilleure.Instructions
                     me = ExtractScalar(context, type, op.Vm);
                 }
 
-                MethodInfo info = sizeF != 0
-                    ? typeof(SoftFloat64).GetMethod(nameof(SoftFloat64.FPCompare))
-                    : typeof(SoftFloat32).GetMethod(nameof(SoftFloat32.FPCompare));
-
-                Operand nzcv = context.Call(info, ne, me, Const(signalNaNs));
+                Operand nzcv = EmitSoftFloatCall(context, nameof(SoftFloat32.FPCompare), ne, me, Const(signalNaNs));
 
                 EmitSetFpscrNzcv(context, nzcv);
             }

ARMeilleure/Instructions/InstEmitSimdCvt.cs

@@ -76,7 +76,9 @@ namespace ARMeilleure.Instructions
                 {
                     Operand ne = context.VectorExtract(OperandType.FP32, GetVec(op.Rn), 0);
 
+                    context.StoreToContext();
                     Operand res = context.Call(typeof(SoftFloat32_16).GetMethod(nameof(SoftFloat32_16.FPConvert)), ne);
+                    context.LoadFromContext();
 
                     res = context.ZeroExtend16(OperandType.I64, res);
 
@@ -98,7 +100,9 @@ namespace ARMeilleure.Instructions
                 {
                     Operand ne = EmitVectorExtractZx(context, op.Rn, 0, 1);
 
+                    context.StoreToContext();
                     Operand res = context.Call(typeof(SoftFloat16_32).GetMethod(nameof(SoftFloat16_32.FPConvert)), ne);
+                    context.LoadFromContext();
 
                     context.Copy(GetVec(op.Rd), context.VectorInsert(context.VectorZero(), res, 0));
                 }
@@ -120,7 +124,9 @@ namespace ARMeilleure.Instructions
                 {
                     Operand ne = context.VectorExtract(OperandType.FP64, GetVec(op.Rn), 0);
 
+                    context.StoreToContext();
                     Operand res = context.Call(typeof(SoftFloat64_16).GetMethod(nameof(SoftFloat64_16.FPConvert)), ne);
+                    context.LoadFromContext();
 
                     res = context.ZeroExtend16(OperandType.I64, res);
 
@@ -143,7 +149,9 @@ namespace ARMeilleure.Instructions
                 {
                     Operand ne = EmitVectorExtractZx(context, op.Rn, 0, 1);
 
+                    context.StoreToContext();
                     Operand res = context.Call(typeof(SoftFloat16_64).GetMethod(nameof(SoftFloat16_64.FPConvert)), ne);
+                    context.LoadFromContext();
 
                     context.Copy(GetVec(op.Rd), context.VectorInsert(context.VectorZero(), res, 0));
                 }
@@ -224,7 +232,9 @@ namespace ARMeilleure.Instructions
                     {
                         Operand ne = EmitVectorExtractZx(context, op.Rn, part + index, 1);
 
+                        context.StoreToContext();
                         Operand e = context.Call(typeof(SoftFloat16_32).GetMethod(nameof(SoftFloat16_32.FPConvert)), ne);
+                        context.LoadFromContext();
 
                         res = context.VectorInsert(res, e, index);
                     }
@@ -333,7 +343,9 @@ namespace ARMeilleure.Instructions
 
                     if (sizeF == 0)
                     {
+                        context.StoreToContext();
                         Operand e = context.Call(typeof(SoftFloat32_16).GetMethod(nameof(SoftFloat32_16.FPConvert)), ne);
+                        context.LoadFromContext();
 
                         e = context.ZeroExtend16(OperandType.I64, e);
 

ARMeilleure/Instructions/InstEmitSimdCvt32.cs

@@ -161,33 +161,14 @@ namespace ARMeilleure.Instructions
                 {
                     Operand toConvert = ExtractScalar(context, floatSize, op.Vm);
 
-                    Operand asInteger;
-
                     // TODO: Fast Path.
                     if (roundWithFpscr)
                     {
-                        MethodInfo info;
-
-                        if (floatSize == OperandType.FP64)
-                        {
-                            info = unsigned
-                                ? typeof(SoftFallback).GetMethod(nameof(SoftFallback.DoubleToUInt32))
-                                : typeof(SoftFallback).GetMethod(nameof(SoftFallback.DoubleToInt32));
-                        }
-                        else
-                        {
-                            info = unsigned
-                                ? typeof(SoftFallback).GetMethod(nameof(SoftFallback.FloatToUInt32))
-                                : typeof(SoftFallback).GetMethod(nameof(SoftFallback.FloatToInt32));
+                        toConvert = EmitRoundByRMode(context, toConvert);
                     }
 
-                        asInteger = context.Call(info, toConvert);
-                    }
-                    else
-                    {
                     // Round towards zero.
-                        asInteger = EmitSaturateFloatToInt(context, toConvert, unsigned);
-                    }
+                    Operand asInteger = EmitSaturateFloatToInt(context, toConvert, unsigned);
 
                     InsertScalar(context, op.Vd, asInteger);
                 }
@@ -271,9 +252,7 @@ namespace ARMeilleure.Instructions
                         break;
                 }
 
-                Operand asInteger;
-
-                asInteger = EmitSaturateFloatToInt(context, toConvert, unsigned);
+                Operand asInteger = EmitSaturateFloatToInt(context, toConvert, unsigned);
 
                 InsertScalar(context, op.Vd, asInteger);
             }
@@ -399,15 +378,9 @@ namespace ARMeilleure.Instructions
         // VRINTX (floating-point).
         public static void Vrintx_S(ArmEmitterContext context)
         {
-            OpCode32SimdS op = (OpCode32SimdS)context.CurrOp;
-
-            bool doubleSize = (op.Size & 1) == 1;
-            string methodName = doubleSize ? nameof(SoftFallback.Round) : nameof(SoftFallback.RoundF);
-
             EmitScalarUnaryOpF32(context, (op1) =>
             {
-                MethodInfo info = typeof(SoftFallback).GetMethod(methodName);
-                return context.Call(info, op1);
+                return EmitRoundByRMode(context, op1);
             });
         }
 

ARMeilleure/Instructions/InstEmitSimdHelper.cs

@@ -243,6 +243,21 @@ namespace ARMeilleure.Instructions
             throw new ArgumentException($"Invalid rounding mode \"{roundMode}\".");
         }
 
+        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);
+        }
+
         public static Operand EmitCountSetBits8(ArmEmitterContext context, Operand op) // "size" is 8 (SIMD&FP Inst.).
         {
             Debug.Assert(op.Type == OperandType.I32 || op.Type == OperandType.I64);
@@ -361,6 +376,54 @@ namespace ARMeilleure.Instructions
             return context.Call(info, n, Const((int)roundMode));
         }
 
+        public static Operand EmitGetRoundingMode(ArmEmitterContext context)
+        {
+            Operand rMode = context.ShiftLeft(GetFpFlag(FPState.RMode1Flag), Const(1));
+                    rMode = context.BitwiseOr(rMode, GetFpFlag(FPState.RMode0Flag));
+
+            return rMode;
+        }
+
+        public static Operand EmitRoundByRMode(ArmEmitterContext context, Operand op)
+        {
+            Debug.Assert(op.Type == OperandType.FP32 || op.Type == OperandType.FP64);
+
+            Operand lbl1 = Label();
+            Operand lbl2 = Label();
+            Operand lbl3 = Label();
+            Operand lblEnd = Label();
+
+            Operand rN = Const((int)FPRoundingMode.ToNearest);
+            Operand rP = Const((int)FPRoundingMode.TowardsPlusInfinity);
+            Operand rM = Const((int)FPRoundingMode.TowardsMinusInfinity);
+
+            Operand res = context.AllocateLocal(op.Type);
+
+            Operand rMode = EmitGetRoundingMode(context);
+
+            context.BranchIf(lbl1, rMode, rN, Comparison.NotEqual);
+            context.Copy(res, EmitRoundMathCall(context, MidpointRounding.ToEven, op));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lbl1);
+            context.BranchIf(lbl2, rMode, rP, Comparison.NotEqual);
+            context.Copy(res, EmitUnaryMathCall(context, nameof(Math.Ceiling), op));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lbl2);
+            context.BranchIf(lbl3, rMode, rM, Comparison.NotEqual);
+            context.Copy(res, EmitUnaryMathCall(context, nameof(Math.Floor), op));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lbl3);
+            context.Copy(res, EmitUnaryMathCall(context, nameof(Math.Truncate), op));
+            context.Branch(lblEnd);
+
+            context.MarkLabel(lblEnd);
+
+            return res;
+        }
+
         public static Operand EmitSoftFloatCall(ArmEmitterContext context, string name, params Operand[] callArgs)
         {
             IOpCodeSimd op = (IOpCodeSimd)context.CurrOp;
@@ -369,7 +432,11 @@ namespace ARMeilleure.Instructions
                 ? typeof(SoftFloat32).GetMethod(name)
                 : typeof(SoftFloat64).GetMethod(name);
 
-            return context.Call(info, callArgs);
+            context.StoreToContext();
+            Operand res = context.Call(info, callArgs);
+            context.LoadFromContext();
+
+            return res;
         }
 
         public static void EmitScalarBinaryOpByElemF(ArmEmitterContext context, Func2I emit)
@@ -1269,7 +1336,7 @@ namespace ARMeilleure.Instructions
 
         public static void EmitSseOrAvxEnterFtzAndDazModesOpF(ArmEmitterContext context, out Operand isTrue)
         {
-            isTrue = context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpcrFz)));
+            isTrue = GetFpFlag(FPState.FzFlag);
 
             Operand lblTrue = Label();
             context.BranchIfFalse(lblTrue, isTrue);
@@ -1281,9 +1348,7 @@ namespace ARMeilleure.Instructions
 
         public static void EmitSseOrAvxExitFtzAndDazModesOpF(ArmEmitterContext context, Operand isTrue = default)
         {
-            isTrue = isTrue == default
-                ? context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpcrFz)))
-                : isTrue;
+            isTrue = isTrue == default ? GetFpFlag(FPState.FzFlag) : isTrue;
 
             Operand lblTrue = Label();
             context.BranchIfFalse(lblTrue, isTrue);
@@ -1552,13 +1617,13 @@ namespace ARMeilleure.Instructions
 
             context.BranchIf(lbl1, op, zeroL, Comparison.LessOrEqual);
             context.Copy(res, maxT);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lbl1);
             context.BranchIf(lblEnd, op, zeroL, Comparison.GreaterOrEqual);
             context.Copy(res, minT);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lblEnd);
@@ -1583,7 +1648,7 @@ namespace ARMeilleure.Instructions
 
             context.BranchIf(lblEnd, op, zeroUL, Comparison.LessOrEqualUI);
             context.Copy(res, maxT);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lblEnd);
@@ -1610,13 +1675,13 @@ namespace ARMeilleure.Instructions
 
             context.BranchIf(lbl1, op, maxT, Comparison.LessOrEqual);
             context.Copy(res, maxT);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lbl1);
             context.BranchIf(lblEnd, op, minT, Comparison.GreaterOrEqual);
             context.Copy(res, minT);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lblEnd);
@@ -1641,7 +1706,7 @@ namespace ARMeilleure.Instructions
 
             context.BranchIf(lblEnd, op, maxT, Comparison.LessOrEqualUI);
             context.Copy(res, maxT);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lblEnd);
@@ -1663,7 +1728,7 @@ namespace ARMeilleure.Instructions
 
             context.BranchIf(lblEnd, op, minL, Comparison.NotEqual);
             context.Copy(res, maxL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lblEnd);
@@ -1691,7 +1756,7 @@ namespace ARMeilleure.Instructions
 
             Operand isPositive = context.ICompareGreaterOrEqual(op1, zeroL);
             context.Copy(res, context.ConditionalSelect(isPositive, maxL, minL));
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lblEnd);
@@ -1713,7 +1778,7 @@ namespace ARMeilleure.Instructions
 
             context.BranchIf(lblEnd, add, op1, Comparison.GreaterOrEqualUI);
             context.Copy(res, maxUL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lblEnd);
@@ -1741,7 +1806,7 @@ namespace ARMeilleure.Instructions
 
             Operand isPositive = context.ICompareGreaterOrEqual(op1, zeroL);
             context.Copy(res, context.ConditionalSelect(isPositive, maxL, minL));
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lblEnd);
@@ -1763,7 +1828,7 @@ namespace ARMeilleure.Instructions
 
             context.BranchIf(lblEnd, op1, op2, Comparison.GreaterOrEqualUI);
             context.Copy(res, zeroL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lblEnd);
@@ -1790,19 +1855,19 @@ namespace ARMeilleure.Instructions
             Operand notOp2AndRes = context.BitwiseAnd(context.BitwiseNot(op2), add);
             context.BranchIf(lblEnd, notOp2AndRes, zeroL, Comparison.GreaterOrEqual);
             context.Copy(res, maxL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lbl1);
             context.BranchIf(lbl2, op2, zeroL, Comparison.Less);
             context.Copy(res, maxL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lbl2);
             context.BranchIf(lblEnd, add, maxL, Comparison.LessOrEqualUI);
             context.Copy(res, maxL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lblEnd);
@@ -1828,14 +1893,14 @@ namespace ARMeilleure.Instructions
             context.BranchIf(lbl1, op1, zeroL, Comparison.Less);
             context.BranchIf(lblEnd, add, op1, Comparison.GreaterOrEqualUI);
             context.Copy(res, maxUL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lbl1);
             context.BranchIf(lblEnd, op2, maxL, Comparison.GreaterUI);
             context.BranchIf(lblEnd, add, zeroL, Comparison.GreaterOrEqual);
             context.Copy(res, zeroL);
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
             context.Branch(lblEnd);
 
             context.MarkLabel(lblEnd);

ARMeilleure/Instructions/InstEmitSimdHelper32.cs

@@ -1181,7 +1181,11 @@ namespace ARMeilleure.Instructions
             Array.Resize(ref callArgs, callArgs.Length + 1);
             callArgs[callArgs.Length - 1] = Const(1);
 
-            return context.Call(info, callArgs);
+            context.StoreToContext();
+            Operand res = context.Call(info, callArgs);
+            context.LoadFromContext();
+
+            return res;
         }
 
         public static Operand EmitVectorExtractSx32(ArmEmitterContext context, int reg, int index, int size)

ARMeilleure/Instructions/InstEmitSimdLogical.cs

@@ -336,8 +336,32 @@ namespace ARMeilleure.Instructions
         {
             OpCodeSimd op = (OpCodeSimd)context.CurrOp;
 
-            Operand res = context.VectorZero();
+            if (Optimizations.UseGfni)
+            {
+                const long bitMatrix =
+                    (0b10000000L << 56) |
+                    (0b01000000L << 48) |
+                    (0b00100000L << 40) |
+                    (0b00010000L << 32) |
+                    (0b00001000L << 24) |
+                    (0b00000100L << 16) |
+                    (0b00000010L <<  8) |
+                    (0b00000001L <<  0);
 
+                Operand vBitMatrix = X86GetAllElements(context, bitMatrix);
+
+                Operand res = context.AddIntrinsic(Intrinsic.X86Gf2p8affineqb, GetVec(op.Rn), vBitMatrix, Const(0));
+
+                if (op.RegisterSize == RegisterSize.Simd64)
+                {
+                    res = context.VectorZeroUpper64(res);
+                }
+
+                context.Copy(GetVec(op.Rd), res);
+            }
+            else
+            {
+                Operand res = context.VectorZero();
                 int elems = op.RegisterSize == RegisterSize.Simd128 ? 16 : 8;
 
                 for (int index = 0; index < elems; index++)
@@ -351,6 +375,7 @@ namespace ARMeilleure.Instructions
 
                 context.Copy(GetVec(op.Rd), res);
             }
+        }
 
         private static Operand EmitReverseBits8Op(ArmEmitterContext context, Operand op)
         {

ARMeilleure/Instructions/InstEmitSimdShift.cs

@@ -88,8 +88,35 @@ namespace ARMeilleure.Instructions
             OpCodeSimdShImm op = (OpCodeSimdShImm)context.CurrOp;
 
             int shift = GetImmShl(op);
+            int eSize = 8 << op.Size;
 
-            if (Optimizations.UseSse2 && op.Size > 0)
+            if (shift >= eSize)
+            {
+                if ((op.RegisterSize == RegisterSize.Simd64))
+                {
+                    Operand res = context.VectorZeroUpper64(GetVec(op.Rd));
+
+                    context.Copy(GetVec(op.Rd), res);
+                }
+            }
+            else if (Optimizations.UseGfni && op.Size == 0)
+            {
+                Operand n = GetVec(op.Rn);
+
+                ulong bitMatrix = X86GetGf2p8LogicalShiftLeft(shift);
+
+                Operand vBitMatrix = X86GetElements(context, bitMatrix, bitMatrix);
+
+                Operand res = context.AddIntrinsic(Intrinsic.X86Gf2p8affineqb, n, vBitMatrix, Const(0));
+
+                if (op.RegisterSize == RegisterSize.Simd64)
+                {
+                    res = context.VectorZeroUpper64(res);
+                }
+
+                context.Copy(GetVec(op.Rd), res);
+            }
+            else if (Optimizations.UseSse2 && op.Size > 0)
             {
                 Operand n = GetVec(op.Rn);
 
@@ -396,10 +423,40 @@ namespace ARMeilleure.Instructions
         {
             OpCodeSimdShImm op = (OpCodeSimdShImm)context.CurrOp;
 
-            if (Optimizations.UseSse2 && op.Size > 0 && op.Size < 3)
-            {
             int shift = GetImmShr(op);
 
+            if (Optimizations.UseGfni && op.Size == 0)
+            {
+                Operand n = GetVec(op.Rn);
+
+                ulong bitMatrix;
+
+                if (shift < 8)
+                {
+                    bitMatrix = X86GetGf2p8LogicalShiftLeft(-shift);
+
+                    // Extend sign-bit
+                    bitMatrix |= 0x8080808080808080UL >> (64 - shift * 8);
+                }
+                else
+                {
+                    // Replicate sign-bit into all bits
+                    bitMatrix = 0x8080808080808080UL;
+                }
+
+                Operand vBitMatrix = X86GetElements(context, bitMatrix, bitMatrix);
+
+                Operand res = context.AddIntrinsic(Intrinsic.X86Gf2p8affineqb, n, vBitMatrix, Const(0));
+
+                if (op.RegisterSize == RegisterSize.Simd64)
+                {
+                    res = context.VectorZeroUpper64(res);
+                }
+
+                context.Copy(GetVec(op.Rd), res);
+            }
+            else if (Optimizations.UseSse2 && op.Size > 0 && op.Size < 3)
+            {
                 Operand n = GetVec(op.Rn);
 
                 Intrinsic sraInst = X86PsraInstruction[op.Size];
@@ -929,10 +986,44 @@ namespace ARMeilleure.Instructions
             OpCodeSimdShImm op = (OpCodeSimdShImm)context.CurrOp;
 
             int shift = GetImmShl(op);
+            int eSize = 8 << op.Size;
 
             ulong mask = shift != 0 ? ulong.MaxValue >> (64 - shift) : 0UL;
 
-            if (Optimizations.UseSse2 && op.Size > 0)
+            if (shift >= eSize)
+            {
+                if ((op.RegisterSize == RegisterSize.Simd64) || scalar)
+                {
+                    Operand res = context.VectorZeroUpper64(GetVec(op.Rd));
+
+                    context.Copy(GetVec(op.Rd), res);
+                }
+            }
+            else if (Optimizations.UseGfni && op.Size == 0)
+            {
+                Operand d = GetVec(op.Rd);
+                Operand n = GetVec(op.Rn);
+
+                ulong bitMatrix = X86GetGf2p8LogicalShiftLeft(shift);
+
+                Operand vBitMatrix = X86GetElements(context, bitMatrix, bitMatrix);
+
+                Operand nShifted = context.AddIntrinsic(Intrinsic.X86Gf2p8affineqb, n, vBitMatrix, Const(0));
+
+                Operand dMask = X86GetAllElements(context, (long)mask * _masks_SliSri[op.Size]);
+
+                Operand dMasked = context.AddIntrinsic(Intrinsic.X86Pand, d, dMask);
+
+                Operand res = context.AddIntrinsic(Intrinsic.X86Por, nShifted, dMasked);
+
+                if ((op.RegisterSize == RegisterSize.Simd64) || scalar)
+                {
+                    res = context.VectorZeroUpper64(res);
+                }
+
+                context.Copy(d, res);
+            }
+            else if (Optimizations.UseSse2 && op.Size > 0)
             {
                 Operand d = GetVec(op.Rd);
                 Operand n = GetVec(op.Rn);
@@ -988,7 +1079,40 @@ namespace ARMeilleure.Instructions
 
             ulong mask = (ulong.MaxValue << (eSize - shift)) & (ulong.MaxValue >> (64 - eSize));
 
-            if (Optimizations.UseSse2 && op.Size > 0)
+            if (shift >= eSize)
+            {
+                if ((op.RegisterSize == RegisterSize.Simd64) || scalar)
+                {
+                    Operand res = context.VectorZeroUpper64(GetVec(op.Rd));
+
+                    context.Copy(GetVec(op.Rd), res);
+                }
+            }
+            else if (Optimizations.UseGfni && op.Size == 0)
+            {
+                Operand d = GetVec(op.Rd);
+                Operand n = GetVec(op.Rn);
+
+                ulong bitMatrix = X86GetGf2p8LogicalShiftLeft(-shift);
+
+                Operand vBitMatrix = X86GetElements(context, bitMatrix, bitMatrix);
+
+                Operand nShifted = context.AddIntrinsic(Intrinsic.X86Gf2p8affineqb, n, vBitMatrix, Const(0));
+
+                Operand dMask = X86GetAllElements(context, (long)mask * _masks_SliSri[op.Size]);
+
+                Operand dMasked = context.AddIntrinsic(Intrinsic.X86Pand, d, dMask);
+
+                Operand res = context.AddIntrinsic(Intrinsic.X86Por, nShifted, dMasked);
+
+                if ((op.RegisterSize == RegisterSize.Simd64) || scalar)
+                {
+                    res = context.VectorZeroUpper64(res);
+                }
+
+                context.Copy(d, res);
+            }
+            else if (Optimizations.UseSse2 && op.Size > 0)
             {
                 Operand d = GetVec(op.Rd);
                 Operand n = GetVec(op.Rn);

ARMeilleure/Instructions/InstEmitSimdShift32.cs

@@ -1,5 +1,6 @@
 using ARMeilleure.Decoders;
 using ARMeilleure.IntermediateRepresentation;
+using ARMeilleure.State;
 using ARMeilleure.Translation;
 using System;
 using System.Diagnostics;
@@ -378,7 +379,7 @@ namespace ARMeilleure.Instructions
 
             context.BranchIfFalse(lblNoSat, context.BitwiseOr(gt, lt));
 
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
+            SetFpFlag(context, FPState.QcFlag, Const(1));
 
             context.MarkLabel(lblNoSat);
 

ARMeilleure/Instructions/InstEmitSystem.cs

@@ -31,8 +31,8 @@ namespace ARMeilleure.Instructions
                 case 0b11_011_0000_0000_001: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetCtrEl0)); break;
                 case 0b11_011_0000_0000_111: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetDczidEl0)); break;
                 case 0b11_011_0100_0010_000: EmitGetNzcv(context); return;
-                case 0b11_011_0100_0100_000: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpcr)); break;
-                case 0b11_011_0100_0100_001: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpsr)); break;
+                case 0b11_011_0100_0100_000: EmitGetFpcr(context); return;
+                case 0b11_011_0100_0100_001: EmitGetFpsr(context); return;
                 case 0b11_011_1101_0000_010: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetTpidrEl0)); break;
                 case 0b11_011_1101_0000_011: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetTpidrroEl0)); break;
                 case 0b11_011_1110_0000_000: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetCntfrqEl0)); break;
@@ -54,8 +54,8 @@ namespace ARMeilleure.Instructions
             switch (GetPackedId(op))
             {
                 case 0b11_011_0100_0010_000: EmitSetNzcv(context); return;
-                case 0b11_011_0100_0100_000: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpcr));     break;
-                case 0b11_011_0100_0100_001: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsr));     break;
+                case 0b11_011_0100_0100_000: EmitSetFpcr(context); return;
+                case 0b11_011_0100_0100_001: EmitSetFpsr(context); return;
                 case 0b11_011_1101_0000_010: info = typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetTpidrEl0)); break;
 
                 default: throw new NotImplementedException($"Unknown MSR 0x{op.RawOpCode:X8} at 0x{op.Address:X16}.");
@@ -121,39 +121,91 @@ namespace ARMeilleure.Instructions
         {
             OpCodeSystem op = (OpCodeSystem)context.CurrOp;
 
-            Operand vSh = context.ShiftLeft(GetFlag(PState.VFlag), Const((int)PState.VFlag));
-            Operand cSh = context.ShiftLeft(GetFlag(PState.CFlag), Const((int)PState.CFlag));
-            Operand zSh = context.ShiftLeft(GetFlag(PState.ZFlag), Const((int)PState.ZFlag));
-            Operand nSh = context.ShiftLeft(GetFlag(PState.NFlag), Const((int)PState.NFlag));
+            Operand nzcv = context.ShiftLeft(GetFlag(PState.VFlag), Const((int)PState.VFlag));
+            nzcv = context.BitwiseOr(nzcv, context.ShiftLeft(GetFlag(PState.CFlag), Const((int)PState.CFlag)));
+            nzcv = context.BitwiseOr(nzcv, context.ShiftLeft(GetFlag(PState.ZFlag), Const((int)PState.ZFlag)));
+            nzcv = context.BitwiseOr(nzcv, context.ShiftLeft(GetFlag(PState.NFlag), Const((int)PState.NFlag)));
 
-            Operand nzcvSh = context.BitwiseOr(context.BitwiseOr(nSh, zSh), context.BitwiseOr(cSh, vSh));
+            SetIntOrZR(context, op.Rt, nzcv);
+        }
 
-            SetIntOrZR(context, op.Rt, nzcvSh);
+        private static void EmitGetFpcr(ArmEmitterContext context)
+        {
+            OpCodeSystem op = (OpCodeSystem)context.CurrOp;
+
+            Operand fpcr = Const(0);
+
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                if (FPCR.Mask.HasFlag((FPCR)(1u << flag)))
+                {
+                    fpcr = context.BitwiseOr(fpcr, context.ShiftLeft(GetFpFlag((FPState)flag), Const(flag)));
+                }
+            }
+
+            SetIntOrZR(context, op.Rt, fpcr);
+        }
+
+        private static void EmitGetFpsr(ArmEmitterContext context)
+        {
+            OpCodeSystem op = (OpCodeSystem)context.CurrOp;
+
+            Operand fpsr = Const(0);
+
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                if (FPSR.Mask.HasFlag((FPSR)(1u << flag)))
+                {
+                    fpsr = context.BitwiseOr(fpsr, context.ShiftLeft(GetFpFlag((FPState)flag), Const(flag)));
+                }
+            }
+
+            SetIntOrZR(context, op.Rt, fpsr);
         }
 
         private static void EmitSetNzcv(ArmEmitterContext context)
         {
             OpCodeSystem op = (OpCodeSystem)context.CurrOp;
 
-            Operand t = GetIntOrZR(context, op.Rt);
-                    t = context.ConvertI64ToI32(t);
+            Operand nzcv = GetIntOrZR(context, op.Rt);
+                    nzcv = context.ConvertI64ToI32(nzcv);
 
-            Operand v = context.ShiftRightUI(t, Const((int)PState.VFlag));
-                    v = context.BitwiseAnd  (v, Const(1));
+            SetFlag(context, PState.VFlag, context.BitwiseAnd(context.ShiftRightUI(nzcv, Const((int)PState.VFlag)), Const(1)));
+            SetFlag(context, PState.CFlag, context.BitwiseAnd(context.ShiftRightUI(nzcv, Const((int)PState.CFlag)), Const(1)));
+            SetFlag(context, PState.ZFlag, context.BitwiseAnd(context.ShiftRightUI(nzcv, Const((int)PState.ZFlag)), Const(1)));
+            SetFlag(context, PState.NFlag, context.BitwiseAnd(context.ShiftRightUI(nzcv, Const((int)PState.NFlag)), Const(1)));
+        }
 
-            Operand c = context.ShiftRightUI(t, Const((int)PState.CFlag));
-                    c = context.BitwiseAnd  (c, Const(1));
+        private static void EmitSetFpcr(ArmEmitterContext context)
+        {
+            OpCodeSystem op = (OpCodeSystem)context.CurrOp;
 
-            Operand z = context.ShiftRightUI(t, Const((int)PState.ZFlag));
-                    z = context.BitwiseAnd  (z, Const(1));
+            Operand fpcr = GetIntOrZR(context, op.Rt);
+                    fpcr = context.ConvertI64ToI32(fpcr);
 
-            Operand n = context.ShiftRightUI(t, Const((int)PState.NFlag));
-                    n = context.BitwiseAnd  (n, Const(1));
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                if (FPCR.Mask.HasFlag((FPCR)(1u << flag)))
+                {
+                    SetFpFlag(context, (FPState)flag, context.BitwiseAnd(context.ShiftRightUI(fpcr, Const(flag)), Const(1)));
+                }
+            }
+        }
 
-            SetFlag(context, PState.VFlag, v);
-            SetFlag(context, PState.CFlag, c);
-            SetFlag(context, PState.ZFlag, z);
-            SetFlag(context, PState.NFlag, n);
+        private static void EmitSetFpsr(ArmEmitterContext context)
+        {
+            OpCodeSystem op = (OpCodeSystem)context.CurrOp;
+
+            Operand fpsr = GetIntOrZR(context, op.Rt);
+                    fpsr = context.ConvertI64ToI32(fpsr);
+
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                if (FPSR.Mask.HasFlag((FPSR)(1u << flag)))
+                {
+                    SetFpFlag(context, (FPState)flag, context.BitwiseAnd(context.ShiftRightUI(fpsr, Const(flag)), Const(1)));
+                }
+            }
         }
     }
 }

ARMeilleure/Instructions/InstEmitSystem32.cs

@@ -169,14 +169,11 @@ namespace ARMeilleure.Instructions
             }
             else
             {
-                Operand vSh = context.ShiftLeft(GetFlag(PState.VFlag), Const((int)PState.VFlag));
-                Operand cSh = context.ShiftLeft(GetFlag(PState.CFlag), Const((int)PState.CFlag));
-                Operand zSh = context.ShiftLeft(GetFlag(PState.ZFlag), Const((int)PState.ZFlag));
-                Operand nSh = context.ShiftLeft(GetFlag(PState.NFlag), Const((int)PState.NFlag));
-                Operand qSh = context.ShiftLeft(GetFlag(PState.QFlag), Const((int)PState.QFlag));
-
-                Operand spsr = context.BitwiseOr(context.BitwiseOr(nSh, zSh), context.BitwiseOr(cSh, vSh));
-                spsr = context.BitwiseOr(spsr, qSh);
+                Operand spsr = context.ShiftLeft(GetFlag(PState.VFlag), Const((int)PState.VFlag));
+                spsr = context.BitwiseOr(spsr, context.ShiftLeft(GetFlag(PState.CFlag), Const((int)PState.CFlag)));
+                spsr = context.BitwiseOr(spsr, context.ShiftLeft(GetFlag(PState.ZFlag), Const((int)PState.ZFlag)));
+                spsr = context.BitwiseOr(spsr, context.ShiftLeft(GetFlag(PState.NFlag), Const((int)PState.NFlag)));
+                spsr = context.BitwiseOr(spsr, context.ShiftLeft(GetFlag(PState.QFlag), Const((int)PState.QFlag)));
 
                 // TODO: Remaining flags.
 
@@ -200,8 +197,7 @@ namespace ARMeilleure.Instructions
 
                     EmitSetNzcv(context, value);
 
-                    Operand q = context.ShiftRightUI(value, Const((int)PState.QFlag));
-                    q = context.BitwiseAnd(q, Const(1));
+                    Operand q = context.BitwiseAnd(context.ShiftRightUI(value, Const((int)PState.QFlag)), Const(1));
 
                     SetFlag(context, PState.QFlag, q);
                 }
@@ -284,17 +280,10 @@ namespace ARMeilleure.Instructions
 
         private static void EmitSetNzcv(ArmEmitterContext context, Operand t)
         {
-            Operand v = context.ShiftRightUI(t, Const((int)PState.VFlag));
-            v = context.BitwiseAnd(v, Const(1));
-
-            Operand c = context.ShiftRightUI(t, Const((int)PState.CFlag));
-            c = context.BitwiseAnd(c, Const(1));
-
-            Operand z = context.ShiftRightUI(t, Const((int)PState.ZFlag));
-            z = context.BitwiseAnd(z, Const(1));
-
-            Operand n = context.ShiftRightUI(t, Const((int)PState.NFlag));
-            n = context.BitwiseAnd(n, Const(1));
+            Operand v = context.BitwiseAnd(context.ShiftRightUI(t, Const((int)PState.VFlag)), Const(1));
+            Operand c = context.BitwiseAnd(context.ShiftRightUI(t, Const((int)PState.CFlag)), Const(1));
+            Operand z = context.BitwiseAnd(context.ShiftRightUI(t, Const((int)PState.ZFlag)), Const(1));
+            Operand n = context.BitwiseAnd(context.ShiftRightUI(t, Const((int)PState.NFlag)), Const(1));
 
             SetFlag(context, PState.VFlag, v);
             SetFlag(context, PState.CFlag, c);
@@ -306,42 +295,32 @@ namespace ARMeilleure.Instructions
         {
             OpCode32SimdSpecial op = (OpCode32SimdSpecial)context.CurrOp;
 
-            Operand vSh = context.ShiftLeft(GetFpFlag(FPState.VFlag), Const((int)FPState.VFlag));
-            Operand cSh = context.ShiftLeft(GetFpFlag(FPState.CFlag), Const((int)FPState.CFlag));
-            Operand zSh = context.ShiftLeft(GetFpFlag(FPState.ZFlag), Const((int)FPState.ZFlag));
-            Operand nSh = context.ShiftLeft(GetFpFlag(FPState.NFlag), Const((int)FPState.NFlag));
+            Operand fpscr = Const(0);
 
-            Operand nzcvSh = context.BitwiseOr(context.BitwiseOr(nSh, zSh), context.BitwiseOr(cSh, vSh));
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                if (FPSCR.Mask.HasFlag((FPSCR)(1u << flag)))
+                {
+                    fpscr = context.BitwiseOr(fpscr, context.ShiftLeft(GetFpFlag((FPState)flag), Const(flag)));
+                }
+            }
 
-            Operand fpscr = context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpscr)));
-
-            SetIntA32(context, op.Rt, context.BitwiseOr(nzcvSh, fpscr));
+            SetIntA32(context, op.Rt, fpscr);
         }
 
         private static void EmitSetFpscr(ArmEmitterContext context)
         {
             OpCode32SimdSpecial op = (OpCode32SimdSpecial)context.CurrOp;
 
-            Operand t = GetIntA32(context, op.Rt);
+            Operand fpscr = GetIntA32(context, op.Rt);
 
-            Operand v = context.ShiftRightUI(t, Const((int)FPState.VFlag));
-            v = context.BitwiseAnd(v, Const(1));
-
-            Operand c = context.ShiftRightUI(t, Const((int)FPState.CFlag));
-            c = context.BitwiseAnd(c, Const(1));
-
-            Operand z = context.ShiftRightUI(t, Const((int)FPState.ZFlag));
-            z = context.BitwiseAnd(z, Const(1));
-
-            Operand n = context.ShiftRightUI(t, Const((int)FPState.NFlag));
-            n = context.BitwiseAnd(n, Const(1));
-
-            SetFpFlag(context, FPState.VFlag, v);
-            SetFpFlag(context, FPState.CFlag, c);
-            SetFpFlag(context, FPState.ZFlag, z);
-            SetFpFlag(context, FPState.NFlag, n);
-
-            context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpscr)), t);
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                if (FPSCR.Mask.HasFlag((FPSCR)(1u << flag)))
+                {
+                    SetFpFlag(context, (FPState)flag, context.BitwiseAnd(context.ShiftRightUI(fpscr, Const(flag)), Const(1)));
+                }
+            }
         }
     }
 }

ARMeilleure/Instructions/NativeInterface.cs

@@ -72,29 +72,6 @@ namespace ARMeilleure.Instructions
             return (ulong)GetContext().DczidEl0;
         }
 
-        public static ulong GetFpcr()
-        {
-            return (ulong)GetContext().Fpcr;
-        }
-
-        public static bool GetFpcrFz()
-        {
-            return (GetContext().Fpcr & FPCR.Fz) != 0;
-        }
-
-        public static ulong GetFpsr()
-        {
-            return (ulong)GetContext().Fpsr;
-        }
-
-        public static uint GetFpscr()
-        {
-            ExecutionContext context = GetContext();
-
-            return (uint)(context.Fpsr & FPSR.A32Mask & ~FPSR.Nzcv) |
-                   (uint)(context.Fpcr & FPCR.A32Mask);
-        }
-
         public static ulong GetTpidrEl0()
         {
             return (ulong)GetContext().TpidrEl0;
@@ -130,29 +107,6 @@ namespace ARMeilleure.Instructions
             return GetContext().CntvctEl0;
         }
 
-        public static void SetFpcr(ulong value)
-        {
-            GetContext().Fpcr = (FPCR)value;
-        }
-
-        public static void SetFpsr(ulong value)
-        {
-            GetContext().Fpsr = (FPSR)value;
-        }
-
-        public static void SetFpsrQc()
-        {
-            GetContext().Fpsr |= FPSR.Qc;
-        }
-
-        public static void SetFpscr(uint fpscr)
-        {
-            ExecutionContext context = GetContext();
-
-            context.Fpsr = FPSR.A32Mask & (FPSR)fpscr;
-            context.Fpcr = FPCR.A32Mask & (FPCR)fpscr;
-        }
-
         public static void SetTpidrEl0(ulong value)
         {
             GetContext().TpidrEl0 = (long)value;

ARMeilleure/Instructions/SoftFallback.cs

@@ -91,76 +91,6 @@ namespace ARMeilleure.Instructions
         }
 #endregion
 
-#region "Rounding"
-        public static double Round(double value)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            FPRoundingMode roundMode = context.Fpcr.GetRoundingMode();
-
-            if (roundMode == FPRoundingMode.ToNearest)
-            {
-                return Math.Round(value); // even
-            }
-            else if (roundMode == FPRoundingMode.TowardsPlusInfinity)
-            {
-                return Math.Ceiling(value);
-            }
-            else if (roundMode == FPRoundingMode.TowardsMinusInfinity)
-            {
-                return Math.Floor(value);
-            }
-            else /* if (roundMode == FPRoundingMode.TowardsZero) */
-            {
-                return Math.Truncate(value);
-            }
-        }
-
-        public static float RoundF(float value)
-        {
-            ExecutionContext context = NativeInterface.GetContext();
-
-            FPRoundingMode roundMode = context.Fpcr.GetRoundingMode();
-
-            if (roundMode == FPRoundingMode.ToNearest)
-            {
-                return MathF.Round(value); // even
-            }
-            else if (roundMode == FPRoundingMode.TowardsPlusInfinity)
-            {
-                return MathF.Ceiling(value);
-            }
-            else if (roundMode == FPRoundingMode.TowardsMinusInfinity)
-            {
-                return MathF.Floor(value);
-            }
-            else /* if (roundMode == FPRoundingMode.TowardsZero) */
-            {
-                return MathF.Truncate(value);
-            }
-        }
-
-        public static int FloatToInt32(float value)
-        {
-            return SatF32ToS32(RoundF(value));
-        }
-
-        public static int DoubleToInt32(double value)
-        {
-            return SatF64ToS32(Round(value));
-        }
-
-        public static uint FloatToUInt32(float value)
-        {
-            return SatF32ToU32(RoundF(value));
-        }
-
-        public static uint DoubleToUInt32(double value)
-        {
-            return SatF64ToU32(Round(value));
-        }
-#endregion
-
 #region "Saturation"
         public static int SatF32ToS32(float value)
         {

ARMeilleure/Instructions/SoftFloat.cs

@@ -12,8 +12,8 @@ namespace ARMeilleure.Instructions
             RecipSqrtEstimateTable = BuildRecipSqrtEstimateTable();
         }
 
-        internal static readonly byte[] RecipEstimateTable;
-        internal static readonly byte[] RecipSqrtEstimateTable;
+        public static readonly byte[] RecipEstimateTable;
+        public static readonly byte[] RecipSqrtEstimateTable;
 
         private static byte[] BuildRecipEstimateTable()
         {
@@ -94,6 +94,13 @@ namespace ARMeilleure.Instructions
                 context.Fpsr |= (FPSR)(1 << (int)exc);
             }
         }
+
+        public static FPRoundingMode GetRoundingMode(this FPCR fpcr)
+        {
+            const int RModeShift = 22;
+
+            return (FPRoundingMode)(((uint)fpcr >> RModeShift) & 3u);
+        }
     }
 
     static class SoftFloat16

ARMeilleure/IntermediateRepresentation/Intrinsic.cs

@@ -47,6 +47,7 @@ namespace ARMeilleure.IntermediateRepresentation
         X86Divps,
         X86Divsd,
         X86Divss,
+        X86Gf2p8affineqb,
         X86Haddpd,
         X86Haddps,
         X86Insertps,

ARMeilleure/Optimizations.cs

@@ -22,6 +22,7 @@ namespace ARMeilleure
         public static bool UseAesniIfAvailable     { get; set; } = true;
         public static bool UsePclmulqdqIfAvailable { get; set; } = true;
         public static bool UseShaIfAvailable       { get; set; } = true;
+        public static bool UseGfniIfAvailable      { get; set; } = true;
 
         public static bool ForceLegacySse
         {
@@ -42,5 +43,6 @@ namespace ARMeilleure
         internal static bool UseAesni     => UseAesniIfAvailable     && HardwareCapabilities.SupportsAesni;
         internal static bool UsePclmulqdq => UsePclmulqdqIfAvailable && HardwareCapabilities.SupportsPclmulqdq;
         internal static bool UseSha       => UseShaIfAvailable       && HardwareCapabilities.SupportsSha;
+        internal static bool UseGfni      => UseGfniIfAvailable      && HardwareCapabilities.SupportsGfni;
     }
 }

ARMeilleure/State/ExecutionContext.cs

@@ -36,10 +36,25 @@ namespace ARMeilleure.State
             set => _nativeContext.SetPstate(value);
         }
 
-        public FPCR Fpcr { get; set; }
-        public FPSR Fpsr { get; set; }
+        public FPSR Fpsr
+        {
+            get => (FPSR)_nativeContext.GetFPState((uint)FPSR.Mask);
+            set => _nativeContext.SetFPState((uint)value, (uint)FPSR.Mask);
+        }
+
+        public FPCR Fpcr
+        {
+            get => (FPCR)_nativeContext.GetFPState((uint)FPCR.Mask);
+            set => _nativeContext.SetFPState((uint)value, (uint)FPCR.Mask);
+        }
         public FPCR StandardFpcrValue => (Fpcr & (FPCR.Ahp)) | FPCR.Dn | FPCR.Fz;
 
+        public FPSCR Fpscr
+        {
+            get => (FPSCR)_nativeContext.GetFPState((uint)FPSCR.Mask);
+            set => _nativeContext.SetFPState((uint)value, (uint)FPSCR.Mask);
+        }
+
         public bool IsAarch32 { get; set; }
 
         internal ExecutionMode ExecutionMode

ARMeilleure/State/FPCR.cs

@@ -5,21 +5,18 @@ namespace ARMeilleure.State
     [Flags]
     public enum FPCR : uint
     {
+        Ioe = 1u << 8,
+        Dze = 1u << 9,
+        Ofe = 1u << 10,
         Ufe = 1u << 11,
+        Ixe = 1u << 12,
+        Ide = 1u << 15,
+        RMode0 = 1u << 22,
+        RMode1 = 1u << 23,
         Fz  = 1u << 24,
         Dn  = 1u << 25,
         Ahp = 1u << 26,
 
-        A32Mask = 0x07FF9F00u
-    }
-
-    public static class FPCRExtensions
-    {
-        private const int RModeShift = 22;
-
-        public static FPRoundingMode GetRoundingMode(this FPCR fpcr)
-        {
-            return (FPRoundingMode)(((int)fpcr >> RModeShift) & 3);
-        }
+        Mask = Ahp | Dn | Fz | RMode1 | RMode0 | Ide | Ixe | Ufe | Ofe | Dze | Ioe // 0x07C09F00u
     }
 }

ARMeilleure/State/FPSCR.cs

@@ -0,0 +1,15 @@
+using System;
+
+namespace ARMeilleure.State
+{
+    [Flags]
+    public enum FPSCR : uint
+    {
+        V = 1u << 28,
+        C = 1u << 29,
+        Z = 1u << 30,
+        N = 1u << 31,
+
+        Mask = N | Z | C | V | FPSR.Mask | FPCR.Mask // 0xFFC09F9Fu
+    }
+}

ARMeilleure/State/FPSR.cs

@@ -5,11 +5,14 @@ namespace ARMeilleure.State
     [Flags]
     public enum FPSR : uint
     {
+        Ioc = 1u << 0,
+        Dzc = 1u << 1,
+        Ofc = 1u << 2,
         Ufc = 1u << 3,
+        Ixc = 1u << 4,
+        Idc = 1u << 7,
         Qc = 1u << 27,
 
-        Nzcv = (1u << 31) | (1u << 30) | (1u << 29) | (1u << 28),
-
-        A32Mask = 0xF800009Fu
+        Mask = Qc | Idc | Ixc | Ufc | Ofc | Dzc | Ioc // 0x0800009Fu
     }
 }

ARMeilleure/State/FPState.cs

@@ -2,9 +2,30 @@
 {
     public enum FPState
     {
+        // FPSR Flags.
+        IocFlag = 0,
+        DzcFlag = 1,
+        OfcFlag = 2,
+        UfcFlag = 3,
+        IxcFlag = 4,
+        IdcFlag = 7,
+        QcFlag = 27,
         VFlag = 28,
         CFlag = 29,
         ZFlag = 30,
-        NFlag = 31
+        NFlag = 31,
+
+        // FPCR Flags.
+        IoeFlag = 8,
+        DzeFlag = 9,
+        OfeFlag = 10,
+        UfeFlag = 11,
+        IxeFlag = 12,
+        IdeFlag = 15,
+        RMode0Flag = 22,
+        RMode1Flag = 23,
+        FzFlag = 24,
+        DnFlag = 25,
+        AhpFlag = 26
     }
 }

ARMeilleure/State/NativeContext.cs

@@ -140,6 +140,34 @@ namespace ARMeilleure.State
             GetStorage().FpFlags[(int)flag] = value ? 1u : 0u;
         }
 
+        public unsafe uint GetFPState(uint mask = uint.MaxValue)
+        {
+            uint value = 0;
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                uint bit = 1u << flag;
+
+                if ((mask & bit) == bit)
+                {
+                    value |= GetStorage().FpFlags[flag] != 0 ? bit : 0u;
+                }
+            }
+            return value;
+        }
+
+        public unsafe void SetFPState(uint value, uint mask = uint.MaxValue)
+        {
+            for (int flag = 0; flag < RegisterConsts.FpFlagsCount; flag++)
+            {
+                uint bit = 1u << flag;
+
+                if ((mask & bit) == bit)
+                {
+                    GetStorage().FpFlags[flag] = (value & bit) == bit ? 1u : 0u;
+                }
+            }
+        }
+
         public int GetCounter() => GetStorage().Counter;
         public void SetCounter(int value) => GetStorage().Counter = value;
 

ARMeilleure/Translation/Delegates.cs

@@ -109,10 +109,6 @@ namespace ARMeilleure.Translation
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetCntvctEl0)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetCtrEl0)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetDczidEl0)));
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpcr)));
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpcrFz)));
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpscr))); // A32 only.
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFpsr)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetFunctionAddress)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.InvalidateCacheLine)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.GetTpidrroEl0)));
@@ -124,10 +120,6 @@ namespace ARMeilleure.Translation
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.ReadUInt32)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.ReadUInt64)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.ReadVector128)));
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpcr)));
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpscr))); // A32 only.
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsr)));
-            SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetFpsrQc)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetTpidrEl0)));
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SetTpidrEl032))); // A32 only.
             SetDelegateInfo(typeof(NativeInterface).GetMethod(nameof(NativeInterface.SignalMemoryTracking)));
@@ -151,12 +143,8 @@ namespace ARMeilleure.Translation
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Crc32w)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Crc32x)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Decrypt)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.DoubleToInt32))); // A32 only.
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.DoubleToUInt32))); // A32 only.
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Encrypt)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.FixedRotate)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.FloatToInt32))); // A32 only.
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.FloatToUInt32))); // A32 only.
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.HashChoose)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.HashLower)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.HashMajority)));
@@ -165,8 +153,6 @@ namespace ARMeilleure.Translation
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.InverseMixColumns)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.MixColumns)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.PolynomialMult64_128)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.Round)));
-            SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.RoundF)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SatF32ToS32)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SatF32ToS64)));
             SetDelegateInfo(typeof(SoftFallback).GetMethod(nameof(SoftFallback.SatF32ToU32)));

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 = 3700; //! To be incremented manually for each change to the ARMeilleure project.
+        private const uint InternalVersion = 3710; //! To be incremented manually for each change to the ARMeilleure project.
 
         private const string ActualDir = "0";
         private const string BackupDir = "1";
@@ -951,7 +951,8 @@ namespace ARMeilleure.Translation.PTC
             return new FeatureInfo(
                 (uint)HardwareCapabilities.FeatureInfo1Ecx,
                 (uint)HardwareCapabilities.FeatureInfo1Edx,
-                (uint)HardwareCapabilities.FeatureInfo7Ebx);
+                (uint)HardwareCapabilities.FeatureInfo7Ebx,
+                (uint)HardwareCapabilities.FeatureInfo7Ecx);
         }
 
         private static byte GetMemoryManagerMode()
@@ -971,7 +972,7 @@ namespace ARMeilleure.Translation.PTC
             return osPlatform;
         }
 
-        [StructLayout(LayoutKind.Sequential, Pack = 1/*, Size = 54*/)]
+        [StructLayout(LayoutKind.Sequential, Pack = 1/*, Size = 58*/)]
         private struct OuterHeader
         {
             public ulong Magic;
@@ -1002,8 +1003,8 @@ namespace ARMeilleure.Translation.PTC
             }
         }
 
-        [StructLayout(LayoutKind.Sequential, Pack = 1/*, Size = 12*/)]
-        private record struct FeatureInfo(uint FeatureInfo0, uint FeatureInfo1, uint FeatureInfo2);
+        [StructLayout(LayoutKind.Sequential, Pack = 1/*, Size = 16*/)]
+        private record struct FeatureInfo(uint FeatureInfo0, uint FeatureInfo1, uint FeatureInfo2, uint FeatureInfo3);
 
         [StructLayout(LayoutKind.Sequential, Pack = 1/*, Size = 128*/)]
         private struct InnerHeader

Ryujinx.Ava/AppHost.cs

@@ -58,6 +58,8 @@ namespace Ryujinx.Ava
         private const float MaxResolutionScale = 4.0f; // Max resolution hotkeys can scale to before wrapping.
         private const int TargetFps = 60;
 
+        private const float VolumeDelta = 0.05f;
+
         private static readonly Cursor InvisibleCursor = new Cursor(StandardCursorType.None);        
 
         private readonly long _ticksPerFrame;
@@ -73,6 +75,7 @@ namespace Ryujinx.Ava
         private bool _isStopped;
         private bool _isActive;
         private long _lastCursorMoveTime;
+        private float _newVolume;
         private long _ticks = 0;
 
         private KeyboardHotkeyState _prevHotkeyState;
@@ -1003,6 +1006,18 @@ namespace Ryujinx.Ava
                             GraphicsConfig.ResScale =
                             (MaxResolutionScale + GraphicsConfig.ResScale - 2) % MaxResolutionScale + 1;
                             break;
+                        case KeyboardHotkeyState.VolumeUp:
+                            _newVolume = MathF.Round((Device.GetVolume() + VolumeDelta), 2);
+                            Device.SetVolume(_newVolume);
+
+                            _parent.ViewModel.Volume = Device.GetVolume();
+                            break;
+                        case KeyboardHotkeyState.VolumeDown:
+                            _newVolume = MathF.Round((Device.GetVolume() - VolumeDelta), 2);
+                            Device.SetVolume(_newVolume);
+
+                            _parent.ViewModel.Volume = Device.GetVolume();
+                            break;
                         case KeyboardHotkeyState.None:
                             (_keyboardInterface as AvaloniaKeyboard).Clear();
                             break;
@@ -1068,6 +1083,14 @@ namespace Ryujinx.Ava
             {
                 state = KeyboardHotkeyState.ResScaleDown;
             }
+            else if (_keyboardInterface.IsPressed((Key)ConfigurationState.Instance.Hid.Hotkeys.Value.VolumeUp))
+            {
+                state = KeyboardHotkeyState.VolumeUp;
+            }
+            else if (_keyboardInterface.IsPressed((Key)ConfigurationState.Instance.Hid.Hotkeys.Value.VolumeDown))
+            {
+                state = KeyboardHotkeyState.VolumeDown;
+            }
 
             return state;
         }

Ryujinx.Ava/Assets/Locales/en_US.json

@@ -579,7 +579,7 @@
   "SettingsTabHotkeysResScaleUpHotkey": "Increase resolution:",
   "SettingsTabHotkeysResScaleDownHotkey": "Decrease resolution:",
   "UserProfilesName": "Name:",
-  "UserProfilesUserId" : "User Id:",
+  "UserProfilesUserId": "User Id:",
   "SettingsTabGraphicsBackend": "Graphics Backend",
   "SettingsTabGraphicsBackendTooltip": "Graphics Backend to use",
   "SettingsEnableTextureRecompression": "Enable Texture Recompression",
@@ -589,5 +589,8 @@
   "SettingsAppRequiredRestartMessage": "Ryujinx Restart Required",
   "SettingsGpuBackendRestartMessage": "Graphics Backend or Gpu settings have been modified. This will require a restart to be applied",
   "SettingsGpuBackendRestartSubMessage": "Do you want to restart now?",
+  "RyujinxUpdaterMessage": "Do you want to update Ryujinx to the latest version?",
+  "SettingsTabHotkeysVolumeUpHotkey": "Increase Volume:",
+  "SettingsTabHotkeysVolumeDownHotkey": "Decrease Volume:",
   "VolumeShort": "Vol"
 }

Ryujinx.Ava/Common/KeyboardHotkeyState.cs

@@ -9,6 +9,8 @@
         Pause,
         ToggleMute,
         ResScaleUp,
-        ResScaleDown
+        ResScaleDown,
+        VolumeUp,
+        VolumeDown
     }
 }

Ryujinx.Ava/Modules/Updater/Updater.cs

@@ -1,4 +1,6 @@
+using Avalonia.Controls;
 using Avalonia.Threading;
+using FluentAvalonia.UI.Controls;
 using ICSharpCode.SharpZipLib.GZip;
 using ICSharpCode.SharpZipLib.Tar;
 using ICSharpCode.SharpZipLib.Zip;
@@ -11,11 +13,13 @@ using Ryujinx.Common;
 using Ryujinx.Common.Logging;
 using System;
 using System.Collections.Generic;
+using System.Diagnostics;
 using System.IO;
 using System.Linq;
 using System.Net;
 using System.Net.Http;
 using System.Net.NetworkInformation;
+using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
 using System.Text;
 using System.Threading;
@@ -40,6 +44,8 @@ namespace Ryujinx.Modules
 
         private static readonly string[] WindowsDependencyDirs = Array.Empty<string>();
 
+        public static bool UpdateSuccessful { get; private set; }
+
         public static async Task BeginParse(MainWindow mainWindow, bool showVersionUpToDate)
         {
             if (Running)
@@ -198,11 +204,18 @@ namespace Ryujinx.Modules
                     _buildSize = -1;
                 }
             }
+
             Dispatcher.UIThread.Post(async () =>
             {
                 // Show a message asking the user if they want to update
-                UpdaterWindow updateDialog = new(mainWindow, newVersion, _buildUrl);
-                await updateDialog.ShowDialog(mainWindow);
+                var shouldUpdate = await ContentDialogHelper.CreateChoiceDialog(LocaleManager.Instance["RyujinxUpdater"],
+                    LocaleManager.Instance["RyujinxUpdaterMessage"],
+                    $"{Program.Version} -> {newVersion}");
+
+                if (shouldUpdate)
+                {
+                    UpdateRyujinx(mainWindow, _buildUrl);
+                }
             });
         }
 
@@ -216,8 +229,10 @@ namespace Ryujinx.Modules
             return result;
         }
 
-        public static void UpdateRyujinx(UpdaterWindow updateDialog, string downloadUrl)
+        public static async void UpdateRyujinx(Window parent, string downloadUrl)
         {
+            UpdateSuccessful = false;
+
             // Empty update dir, although it shouldn't ever have anything inside it
             if (Directory.Exists(UpdateDir))
             {
@@ -228,25 +243,56 @@ namespace Ryujinx.Modules
 
             string updateFile = Path.Combine(UpdateDir, "update.bin");
 
-            // Download the update .zip
-            updateDialog.MainText.Text = LocaleManager.Instance["UpdaterDownloading"];
-            updateDialog.ProgressBar.Value = 0;
-            updateDialog.ProgressBar.Maximum = 100;
+            var taskDialog = new TaskDialog()
+            {
+                Header = LocaleManager.Instance["RyujinxUpdater"],
+                SubHeader = LocaleManager.Instance["UpdaterDownloading"],
+                IconSource = new SymbolIconSource { Symbol = Symbol.Download },
+                Buttons = { },
+                ShowProgressBar = true
+            };
 
-            Task.Run(() =>
+            taskDialog.XamlRoot = parent;
+
+            taskDialog.Opened += (s, e) =>
             {
                 if (_buildSize >= 0)
                 {
-                    DoUpdateWithMultipleThreads(updateDialog, downloadUrl, updateFile);
+                    DoUpdateWithMultipleThreads(taskDialog, downloadUrl, updateFile);
                 }
                 else
                 {
-                    DoUpdateWithSingleThread(updateDialog, downloadUrl, updateFile);
+                    DoUpdateWithSingleThread(taskDialog, downloadUrl, updateFile);
                 }
-            });
+            };
+
+            await taskDialog.ShowAsync(true);
+
+            if (UpdateSuccessful)
+            {
+                var shouldRestart = await ContentDialogHelper.CreateChoiceDialog(LocaleManager.Instance["RyujinxUpdater"],
+                    LocaleManager.Instance["DialogUpdaterCompleteMessage"],
+                    LocaleManager.Instance["DialogUpdaterRestartMessage"]);
+
+                if (shouldRestart)
+                {
+                    string ryuName = Path.GetFileName(Environment.ProcessPath);
+                    string ryuExe = Path.Combine(AppDomain.CurrentDomain.BaseDirectory, ryuName);
+                    string ryuArg = string.Join(" ", Environment.GetCommandLineArgs().Skip(1).ToArray());
+
+                    if (!OperatingSystem.IsWindows())
+                    {
+                        chmod(ryuExe, Convert.ToUInt32("0777", 8));
                     }
 
-        private static void DoUpdateWithMultipleThreads(UpdaterWindow updateDialog, string downloadUrl, string updateFile)
+                    Process.Start(ryuExe, ryuArg);
+
+                    Environment.Exit(0);
+                }
+            }
+        }
+
+        private static void DoUpdateWithMultipleThreads(TaskDialog taskDialog, string downloadUrl, string updateFile)
         {
             // Multi-Threaded Updater
             long chunkSize = _buildSize / ConnectionCount;
@@ -290,7 +336,7 @@ namespace Ryujinx.Modules
                         Interlocked.Exchange(ref progressPercentage[index], args.ProgressPercentage);
                         Interlocked.Add(ref totalProgressPercentage, args.ProgressPercentage);
 
-                        updateDialog.ProgressBar.Value = totalProgressPercentage / ConnectionCount;
+                        taskDialog.SetProgressBarState(totalProgressPercentage / ConnectionCount, TaskDialogProgressState.Normal);
                     };
 
                     client.DownloadDataCompleted += (_, args) =>
@@ -301,6 +347,8 @@ namespace Ryujinx.Modules
                         {
                             webClients[index].Dispose();
 
+                            taskDialog.Hide();
+
                             return;
                         }
 
@@ -320,14 +368,14 @@ namespace Ryujinx.Modules
 
                             try
                             {
-                                InstallUpdate(updateDialog, updateFile);
+                                InstallUpdate(taskDialog, updateFile);
                             }
                             catch (Exception e)
                             {
                                 Logger.Warning?.Print(LogClass.Application, e.Message);
                                 Logger.Warning?.Print(LogClass.Application, "Multi-Threaded update failed, falling back to single-threaded updater.");
 
-                                DoUpdateWithSingleThread(updateDialog, downloadUrl, updateFile);
+                                DoUpdateWithSingleThread(taskDialog, downloadUrl, updateFile);
 
                                 return;
                             }
@@ -348,7 +396,7 @@ namespace Ryujinx.Modules
                             webClients[j].CancelAsync();
                         }
 
-                        DoUpdateWithSingleThread(updateDialog, downloadUrl, updateFile);
+                        DoUpdateWithSingleThread(taskDialog, downloadUrl, updateFile);
 
                         return;
                     }
@@ -356,7 +404,7 @@ namespace Ryujinx.Modules
             }
         }
 
-        private static void DoUpdateWithSingleThreadWorker(UpdaterWindow updateDialog, string downloadUrl, string updateFile)
+        private static void DoUpdateWithSingleThreadWorker(TaskDialog taskDialog, string downloadUrl, string updateFile)
         {
             using (HttpClient client = new HttpClient())
             {
@@ -384,19 +432,26 @@ namespace Ryujinx.Modules
 
                             byteWritten += readSize;
 
-                            updateDialog.ProgressBar.Value = ((double)byteWritten / totalBytes) * 100;
+                            taskDialog.SetProgressBarState(GetPercentage(byteWritten, totalBytes), TaskDialogProgressState.Normal);
+
                             updateFileStream.Write(buffer, 0, readSize);
                         }
                     }
                 }
 
-                InstallUpdate(updateDialog, updateFile);
+                InstallUpdate(taskDialog, updateFile);
             }
         }
 
-        private static void DoUpdateWithSingleThread(UpdaterWindow updateDialog, string downloadUrl, string updateFile)
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private static double GetPercentage(double value, double max)
         {
-            Thread worker = new Thread(() => DoUpdateWithSingleThreadWorker(updateDialog, downloadUrl, updateFile));
+            return max == 0 ? 0 : value / max * 100;
+        }
+
+        private static void DoUpdateWithSingleThread(TaskDialog taskDialog, string downloadUrl, string updateFile)
+        {
+            Thread worker = new Thread(() => DoUpdateWithSingleThreadWorker(taskDialog, downloadUrl, updateFile));
             worker.Name = "Updater.SingleThreadWorker";
             worker.Start();
         }
@@ -414,11 +469,11 @@ namespace Ryujinx.Modules
             }
         }
 
-        private static async void InstallUpdate(UpdaterWindow updateDialog, string updateFile)
+        private static async void InstallUpdate(TaskDialog taskDialog, string updateFile)
         {
             // Extract Update
-            updateDialog.MainText.Text = LocaleManager.Instance["UpdaterExtracting"];
-            updateDialog.ProgressBar.Value = 0;
+            taskDialog.SubHeader = LocaleManager.Instance["UpdaterExtracting"];
+            taskDialog.SetProgressBarState(0, TaskDialogProgressState.Normal);
 
             if (OperatingSystem.IsLinux())
             {
@@ -426,8 +481,6 @@ namespace Ryujinx.Modules
                 using (Stream gzipStream = new GZipInputStream(inStream))
                 using (TarInputStream tarStream = new TarInputStream(gzipStream, Encoding.ASCII))
                 {
-                    updateDialog.ProgressBar.Maximum = inStream.Length;
-
                     await Task.Run(() =>
                     {
                         TarEntry tarEntry;
@@ -450,12 +503,12 @@ namespace Ryujinx.Modules
 
                             Dispatcher.UIThread.Post(() =>
                             {
-                                updateDialog.ProgressBar.Value += entry.Size;
+                                taskDialog.SetProgressBarState(GetPercentage(entry.Size, inStream.Length), TaskDialogProgressState.Normal);
                             });
                         }
                     });
 
-                    updateDialog.ProgressBar.Value = inStream.Length;
+                    taskDialog.SetProgressBarState(100, TaskDialogProgressState.Normal);
                 }
             }
             else
@@ -463,12 +516,12 @@ namespace Ryujinx.Modules
                 using (Stream inStream = File.OpenRead(updateFile))
                 using (ZipFile zipFile = new ZipFile(inStream))
                 {
-                    updateDialog.ProgressBar.Maximum = zipFile.Count;
-
                     await Task.Run(() =>
                     {
+                        double count = 0;
                         foreach (ZipEntry zipEntry in zipFile)
                         {
+                            count++;
                             if (zipEntry.IsDirectory) continue;
 
                             string outPath = Path.Combine(UpdateDir, zipEntry.Name);
@@ -485,7 +538,7 @@ namespace Ryujinx.Modules
 
                             Dispatcher.UIThread.Post(() =>
                             {
-                                updateDialog.ProgressBar.Value++;
+                                taskDialog.SetProgressBarState(GetPercentage(count, zipFile.Count), TaskDialogProgressState.Normal);
                             });
                         }
                     });
@@ -497,22 +550,23 @@ namespace Ryujinx.Modules
 
             List<string> allFiles = EnumerateFilesToDelete().ToList();
 
-            updateDialog.MainText.Text = LocaleManager.Instance["UpdaterRenaming"];
-            updateDialog.ProgressBar.Value = 0;
-            updateDialog.ProgressBar.Maximum = allFiles.Count;
+            taskDialog.SubHeader = LocaleManager.Instance["UpdaterRenaming"];
+            taskDialog.SetProgressBarState(0, TaskDialogProgressState.Normal);
 
             // Replace old files
             await Task.Run(() =>
             {
+                double count = 0;
                 foreach (string file in allFiles)
                 {
+                    count++;
                     try
                     {
                         File.Move(file, file + ".ryuold");
 
                         Dispatcher.UIThread.Post(() =>
                         {
-                            updateDialog.ProgressBar.Value++;
+                            taskDialog.SetProgressBarState(GetPercentage(count, allFiles.Count), TaskDialogProgressState.Normal);
                         });
                     }
                     catch
@@ -523,23 +577,20 @@ namespace Ryujinx.Modules
 
                 Dispatcher.UIThread.Post(() =>
                 {
-                    updateDialog.MainText.Text = LocaleManager.Instance["UpdaterAddingFiles"];
-                    updateDialog.ProgressBar.Value = 0;
-                    updateDialog.ProgressBar.Maximum = Directory.GetFiles(UpdatePublishDir, "*", SearchOption.AllDirectories).Length;
+                    taskDialog.SubHeader = LocaleManager.Instance["UpdaterAddingFiles"];
+                    taskDialog.SetProgressBarState(0, TaskDialogProgressState.Normal);
                 });
 
-                MoveAllFilesOver(UpdatePublishDir, HomeDir, updateDialog);
+                MoveAllFilesOver(UpdatePublishDir, HomeDir, taskDialog);
             });
 
             Directory.Delete(UpdateDir, true);
 
             SetUnixPermissions();
 
-            updateDialog.MainText.Text = LocaleManager.Instance["DialogUpdaterCompleteMessage"];
-            updateDialog.SecondaryText.Text = LocaleManager.Instance["DialogUpdaterRestartMessage"];
+            UpdateSuccessful = true;
 
-            updateDialog.ProgressBar.IsVisible = false;
-            updateDialog.ButtonBox.IsVisible = true;
+            taskDialog.Hide();
         }
 
 #pragma warning disable CS4014 // Because this call is not awaited, execution of the current method continues before the call is completed
@@ -618,8 +669,9 @@ namespace Ryujinx.Modules
             return files;
         }
 
-        private static void MoveAllFilesOver(string root, string dest, UpdaterWindow dialog)
+        private static void MoveAllFilesOver(string root, string dest, TaskDialog taskDialog)
         {
+            var total = Directory.GetFiles(root, "*", SearchOption.AllDirectories).Length;
             foreach (string directory in Directory.GetDirectories(root))
             {
                 string dirName = Path.GetFileName(directory);
@@ -629,16 +681,18 @@ namespace Ryujinx.Modules
                     Directory.CreateDirectory(Path.Combine(dest, dirName));
                 }
 
-                MoveAllFilesOver(directory, Path.Combine(dest, dirName), dialog);
+                MoveAllFilesOver(directory, Path.Combine(dest, dirName), taskDialog);
             }
 
+            double count = 0;
             foreach (string file in Directory.GetFiles(root))
             {
+                count++;
                 File.Move(file, Path.Combine(dest, Path.GetFileName(file)), true);
 
                 Dispatcher.UIThread.InvokeAsync(() =>
                 {
-                    dialog.ProgressBar.Value++;
+                    taskDialog.SetProgressBarState(GetPercentage(count, total), TaskDialogProgressState.Normal);
                 });
             }
         }

Ryujinx.Ava/Ui/Windows/AboutWindow.axaml

@@ -6,7 +6,6 @@
     xmlns:locale="clr-namespace:Ryujinx.Ava.Common.Locale"
     xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006"
     xmlns:window="clr-namespace:Ryujinx.Ava.Ui.Windows"
-    Title="Ryujinx - About"
     Width="850"
     Height="550"
     MinWidth="500"

Ryujinx.Ava/Ui/Windows/SettingsWindow.axaml

@@ -257,6 +257,22 @@
                                     TextAlignment="Center" />
                             </ToggleButton>
                         </StackPanel>
+                        <StackPanel Margin="10,0,0,0" Orientation="Horizontal">
+                            <TextBlock VerticalAlignment="Center" Text="{locale:Locale SettingsTabHotkeysVolumeUpHotkey}" Width="230" />
+                            <ToggleButton Width="90" Height="27" Checked="Button_Checked" Unchecked="Button_Unchecked">
+                                <TextBlock
+                                    Text="{Binding KeyboardHotkeys.VolumeUp, Mode=TwoWay, Converter={StaticResource Key}}"
+                                    TextAlignment="Center" />
+                            </ToggleButton>
+                        </StackPanel>
+                        <StackPanel Margin="10,0,0,0" Orientation="Horizontal">
+                            <TextBlock VerticalAlignment="Center" Text="{locale:Locale SettingsTabHotkeysVolumeDownHotkey}" Width="230" />
+                            <ToggleButton Width="90" Height="27" Checked="Button_Checked" Unchecked="Button_Unchecked">
+                                <TextBlock
+                                    Text="{Binding KeyboardHotkeys.VolumeDown, Mode=TwoWay, Converter={StaticResource Key}}"
+                                    TextAlignment="Center" />
+                            </ToggleButton>
+                        </StackPanel>
                     </StackPanel>
                 </Border>
             </ScrollViewer>

Ryujinx.Ava/Ui/Windows/UpdaterWindow.axaml

@@ -1,66 +0,0 @@
-<window:StyleableWindow
-    x:Class="Ryujinx.Ava.Ui.Windows.UpdaterWindow"
-    xmlns="https://github.com/avaloniaui"
-    xmlns:x="http://schemas.microsoft.com/winfx/2006/xaml"
-    xmlns:d="http://schemas.microsoft.com/expression/blend/2008"
-    xmlns:locale="clr-namespace:Ryujinx.Ava.Common.Locale"
-    xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006"
-    xmlns:window="clr-namespace:Ryujinx.Ava.Ui.Windows"
-    Title="Ryujinx Updater"
-    Width="500"
-    Height="500"
-    MinWidth="500"
-    MinHeight="500"
-    d:DesignHeight="350"
-    d:DesignWidth="400"
-    CanResize="False"
-    SizeToContent="Height"
-    WindowStartupLocation="CenterOwner"
-    mc:Ignorable="d">
-    <Grid
-        Margin="20"
-        HorizontalAlignment="Stretch"
-        VerticalAlignment="Stretch">
-        <Grid.RowDefinitions>
-            <RowDefinition />
-            <RowDefinition />
-            <RowDefinition />
-            <RowDefinition />
-            <RowDefinition />
-        </Grid.RowDefinitions>
-        <TextBlock
-            Name="MainText"
-            Grid.Row="1"
-            Height="20"
-            HorizontalAlignment="Stretch"
-            TextAlignment="Center" />
-        <TextBlock
-            Name="SecondaryText"
-            Grid.Row="2"
-            Height="20"
-            HorizontalAlignment="Stretch"
-            TextAlignment="Center" />
-        <ProgressBar
-            Name="ProgressBar"
-            Grid.Row="3"
-            Margin="20"
-            HorizontalAlignment="Stretch"
-            IsVisible="False"
-            Maximum="100"
-            Minimum="0" />
-        <StackPanel
-            Name="ButtonBox"
-            Grid.Row="4"
-            HorizontalAlignment="Right"
-            IsVisible="False"
-            Orientation="Horizontal"
-            Spacing="20">
-            <Button MinWidth="50" Command="{Binding YesPressed}">
-                <TextBlock Text="{locale:Locale InputDialogYes}" TextAlignment="Center" />
-            </Button>
-            <Button MinWidth="50" Command="{Binding NoPressed}">
-                <TextBlock Text="{locale:Locale InputDialogNo}" TextAlignment="Center" />
-            </Button>
-        </StackPanel>
-    </Grid>
-</window:StyleableWindow>

Ryujinx.Ava/Ui/Windows/UpdaterWindow.axaml.cs

@@ -1,73 +0,0 @@
-using Ryujinx.Ava.Common.Locale;
-using Ryujinx.Modules;
-using System;
-using System.Diagnostics;
-using System.IO;
-using System.Linq;
-using System.Runtime.InteropServices;
-
-namespace Ryujinx.Ava.Ui.Windows
-{
-    public partial class UpdaterWindow : StyleableWindow
-    {
-        private readonly string _buildUrl;
-        private readonly MainWindow _mainWindow;
-        private readonly Version _newVersion;
-        private bool _restartQuery;
-
-        public UpdaterWindow()
-        {
-            DataContext = this;
-
-            InitializeComponent();
-
-            Title = LocaleManager.Instance["RyujinxUpdater"];
-        }
-
-        public UpdaterWindow(MainWindow mainWindow, Version newVersion, string buildUrl) : this()
-        {
-            _mainWindow = mainWindow;
-            _newVersion = newVersion;
-            _buildUrl = buildUrl;
-        }
-
-        [DllImport("libc", SetLastError = true)]
-        private static extern int chmod(string path, uint mode);
-
-        public void YesPressed()
-        {
-            if (_restartQuery)
-            {
-                string ryuName = OperatingSystem.IsWindows() ? "Ryujinx.Ava.exe" : "Ryujinx.Ava";
-                string ryuExe = Path.Combine(AppDomain.CurrentDomain.BaseDirectory, ryuName);
-                string ryuArg = string.Join(" ", Environment.GetCommandLineArgs().AsEnumerable().Skip(1).ToArray());
-
-                if (!OperatingSystem.IsWindows())
-                {
-                    chmod(ryuExe, 0777);
-                }
-
-                Process.Start(ryuExe, ryuArg);
-
-                Environment.Exit(0);
-            }
-            else
-            {
-                ButtonBox.IsVisible = false;
-                ProgressBar.IsVisible = true;
-
-                SecondaryText.Text = "";
-                _restartQuery = true;
-
-                Updater.UpdateRyujinx(this, _buildUrl);
-            }
-        }
-
-        public void NoPressed()
-        {
-            _mainWindow.UpdateMenuItem.IsEnabled = true;
-
-            Close();
-        }
-    }
-}

Ryujinx.Common/Configuration/Hid/KeyboardHotkeys.cs

@@ -9,5 +9,7 @@
         public Key ToggleMute { get; set; }
         public Key ResScaleUp { get; set; }
         public Key ResScaleDown { get; set; }
+        public Key VolumeUp { get; set; }
+        public Key VolumeDown { get; set; }
     }
 }

Ryujinx.Common/Logging/LogClass.cs

@@ -30,6 +30,7 @@ namespace Ryujinx.Common.Logging
         ServiceBsd,
         ServiceBtm,
         ServiceCaps,
+        ServiceFatal,
         ServiceFriend,
         ServiceFs,
         ServiceHid,

Ryujinx.Common/Memory/SpanOrArray.cs

@@ -0,0 +1,89 @@
+using System;
+
+namespace Ryujinx.Common.Memory
+{
+    /// <summary>
+    /// A struct that can represent both a Span and Array.
+    /// This is useful to keep the Array representation when possible to avoid copies.
+    /// </summary>
+    /// <typeparam name="T">Element Type</typeparam>
+    public ref struct SpanOrArray<T> where T : unmanaged
+    {
+        public readonly T[] Array;
+        public readonly ReadOnlySpan<T> Span;
+
+        /// <summary>
+        /// Create a new SpanOrArray from an array.
+        /// </summary>
+        /// <param name="array">Array to store</param>
+        public SpanOrArray(T[] array)
+        {
+            Array = array;
+            Span = ReadOnlySpan<T>.Empty;
+        }
+
+        /// <summary>
+        /// Create a new SpanOrArray from a readonly span.
+        /// </summary>
+        /// <param name="array">Span to store</param>
+        public SpanOrArray(ReadOnlySpan<T> span)
+        {
+            Array = null;
+            Span = span;
+        }
+
+        /// <summary>
+        /// Return the contained array, or convert the span if necessary.
+        /// </summary>
+        /// <returns>An array containing the data</returns>
+        public T[] ToArray()
+        {
+            return Array ?? Span.ToArray();
+        }
+
+        /// <summary>
+        /// Return a ReadOnlySpan from either the array or ReadOnlySpan.
+        /// </summary>
+        /// <returns>A ReadOnlySpan containing the data</returns>
+        public ReadOnlySpan<T> AsSpan()
+        {
+            return Array ?? Span;
+        }
+
+        /// <summary>
+        /// Cast an array to a SpanOrArray.
+        /// </summary>
+        /// <param name="array">Source array</param>
+        public static implicit operator SpanOrArray<T>(T[] array)
+        {
+            return new SpanOrArray<T>(array);
+        }
+
+        /// <summary>
+        /// Cast a ReadOnlySpan to a SpanOrArray.
+        /// </summary>
+        /// <param name="span">Source ReadOnlySpan</param>
+        public static implicit operator SpanOrArray<T>(ReadOnlySpan<T> span)
+        {
+            return new SpanOrArray<T>(span);
+        }
+
+        /// <summary>
+        /// Cast a Span to a SpanOrArray.
+        /// </summary>
+        /// <param name="span">Source Span</param>
+        public static implicit operator SpanOrArray<T>(Span<T> span)
+        {
+            return new SpanOrArray<T>(span);
+        }
+
+        /// <summary>
+        /// Cast a SpanOrArray to a ReadOnlySpan
+        /// </summary>
+        /// <param name="spanOrArray">Source SpanOrArray</param>
+        public static implicit operator ReadOnlySpan<T>(SpanOrArray<T> spanOrArray)
+        {
+            return spanOrArray.AsSpan();
+        }
+    }
+}

Ryujinx.Graphics.GAL/ITexture.cs

@@ -1,3 +1,4 @@
+using Ryujinx.Common.Memory;
 using System;
 
 namespace Ryujinx.Graphics.GAL
@@ -17,9 +18,9 @@ namespace Ryujinx.Graphics.GAL
         ReadOnlySpan<byte> GetData();
         ReadOnlySpan<byte> GetData(int layer, int level);
 
-        void SetData(ReadOnlySpan<byte> data);
-        void SetData(ReadOnlySpan<byte> data, int layer, int level);
-        void SetData(ReadOnlySpan<byte> data, int layer, int level, Rectangle<int> region);
+        void SetData(SpanOrArray<byte> data);
+        void SetData(SpanOrArray<byte> data, int layer, int level);
+        void SetData(SpanOrArray<byte> data, int layer, int level, Rectangle<int> region);
         void SetStorage(BufferRange buffer);
         void Release();
     }

Ryujinx.Graphics.GAL/Multithreading/Resources/ThreadedTexture.cs

@@ -1,4 +1,5 @@
-using Ryujinx.Graphics.GAL.Multithreading.Commands.Texture;
+using Ryujinx.Common.Memory;
+using Ryujinx.Graphics.GAL.Multithreading.Commands.Texture;
 using Ryujinx.Graphics.GAL.Multithreading.Model;
 using System;
 
@@ -107,19 +108,19 @@ namespace Ryujinx.Graphics.GAL.Multithreading.Resources
             }
         }
 
-        public void SetData(ReadOnlySpan<byte> data)
+        public void SetData(SpanOrArray<byte> data)
         {
             _renderer.New<TextureSetDataCommand>().Set(Ref(this), Ref(data.ToArray()));
             _renderer.QueueCommand();
         }
 
-        public void SetData(ReadOnlySpan<byte> data, int layer, int level)
+        public void SetData(SpanOrArray<byte> data, int layer, int level)
         {
             _renderer.New<TextureSetDataSliceCommand>().Set(Ref(this), Ref(data.ToArray()), layer, level);
             _renderer.QueueCommand();
         }
 
-        public void SetData(ReadOnlySpan<byte> data, int layer, int level, Rectangle<int> region)
+        public void SetData(SpanOrArray<byte> data, int layer, int level, Rectangle<int> region)
         {
             _renderer.New<TextureSetDataSliceRegionCommand>().Set(Ref(this), Ref(data.ToArray()), layer, level, region);
             _renderer.QueueCommand();

Ryujinx.Graphics.Gpu/Engine/Dma/DmaClass.cs

@@ -229,7 +229,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.Dma
 
                     if (target != null)
                     {
-                        ReadOnlySpan<byte> data;
+                        byte[] data;
                         if (srcLinear)
                         {
                             data = LayoutConverter.ConvertLinearStridedToLinear(

Ryujinx.Graphics.Gpu/Image/Texture.cs

@@ -1,5 +1,6 @@
 using Ryujinx.Common;
 using Ryujinx.Common.Logging;
+using Ryujinx.Common.Memory;
 using Ryujinx.Graphics.GAL;
 using Ryujinx.Graphics.Gpu.Memory;
 using Ryujinx.Graphics.Texture;
@@ -136,11 +137,6 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// </summary>
         public LinkedListNode<Texture> CacheNode { get; set; }
 
-        /// <summary>
-        /// Event to fire when texture data is disposed.
-        /// </summary>
-        public event Action<Texture> Disposed;
-
         /// <summary>
         /// Physical memory ranges where the texture data is located.
         /// </summary>
@@ -720,9 +716,9 @@ namespace Ryujinx.Graphics.Gpu.Image
                 }
             }
 
-            data = ConvertToHostCompatibleFormat(data);
+            SpanOrArray<byte> result = ConvertToHostCompatibleFormat(data);
 
-            HostTexture.SetData(data);
+            HostTexture.SetData(result);
 
             _hasData = true;
         }
@@ -731,7 +727,7 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// Uploads new texture data to the host GPU.
         /// </summary>
         /// <param name="data">New data</param>
-        public void SetData(ReadOnlySpan<byte> data)
+        public void SetData(SpanOrArray<byte> data)
         {
             BlacklistScale();
 
@@ -750,7 +746,7 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// <param name="data">New data</param>
         /// <param name="layer">Target layer</param>
         /// <param name="level">Target level</param>
-        public void SetData(ReadOnlySpan<byte> data, int layer, int level)
+        public void SetData(SpanOrArray<byte> data, int layer, int level)
         {
             BlacklistScale();
 
@@ -786,7 +782,7 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// <param name="level">Mip level to convert</param>
         /// <param name="single">True to convert a single slice</param>
         /// <returns>Converted data</returns>
-        public ReadOnlySpan<byte> ConvertToHostCompatibleFormat(ReadOnlySpan<byte> data, int level = 0, bool single = false)
+        public SpanOrArray<byte> ConvertToHostCompatibleFormat(ReadOnlySpan<byte> data, int level = 0, bool single = false)
         {
             int width = Info.Width;
             int height = Info.Height;
@@ -799,9 +795,11 @@ namespace Ryujinx.Graphics.Gpu.Image
             height = Math.Max(height >> level, 1);
             depth = Math.Max(depth >> level, 1);
 
+            SpanOrArray<byte> result;
+
             if (Info.IsLinear)
             {
-                data = LayoutConverter.ConvertLinearStridedToLinear(
+                result = LayoutConverter.ConvertLinearStridedToLinear(
                     width,
                     height,
                     Info.FormatInfo.BlockWidth,
@@ -813,7 +811,7 @@ namespace Ryujinx.Graphics.Gpu.Image
             }
             else
             {
-                data = LayoutConverter.ConvertBlockLinearToLinear(
+                result = LayoutConverter.ConvertBlockLinearToLinear(
                     width,
                     height,
                     depth,
@@ -836,7 +834,7 @@ namespace Ryujinx.Graphics.Gpu.Image
             if (!_context.Capabilities.SupportsAstcCompression && Format.IsAstc())
             {
                 if (!AstcDecoder.TryDecodeToRgba8P(
-                    data.ToArray(),
+                    result.ToArray(),
                     Info.FormatInfo.BlockWidth,
                     Info.FormatInfo.BlockHeight,
                     width,
@@ -856,11 +854,11 @@ namespace Ryujinx.Graphics.Gpu.Image
                     decoded = BCnEncoder.EncodeBC7(decoded, width, height, depth, levels, layers);
                 }
 
-                data = decoded;
+                result = decoded;
             }
             else if (!_context.Capabilities.SupportsR4G4Format && Format == Format.R4G4Unorm)
             {
-                data = PixelConverter.ConvertR4G4ToR4G4B4A4(data);
+                result = PixelConverter.ConvertR4G4ToR4G4B4A4(result);
             }
             else if (!TextureCompatibility.HostSupportsBcFormat(Format, Target, _context.Capabilities))
             {
@@ -868,36 +866,36 @@ namespace Ryujinx.Graphics.Gpu.Image
                 {
                     case Format.Bc1RgbaSrgb:
                     case Format.Bc1RgbaUnorm:
-                        data = BCnDecoder.DecodeBC1(data, width, height, depth, levels, layers);
+                        result = BCnDecoder.DecodeBC1(result, width, height, depth, levels, layers);
                         break;
                     case Format.Bc2Srgb:
                     case Format.Bc2Unorm:
-                        data = BCnDecoder.DecodeBC2(data, width, height, depth, levels, layers);
+                        result = BCnDecoder.DecodeBC2(result, width, height, depth, levels, layers);
                         break;
                     case Format.Bc3Srgb:
                     case Format.Bc3Unorm:
-                        data = BCnDecoder.DecodeBC3(data, width, height, depth, levels, layers);
+                        result = BCnDecoder.DecodeBC3(result, width, height, depth, levels, layers);
                         break;
                     case Format.Bc4Snorm:
                     case Format.Bc4Unorm:
-                        data = BCnDecoder.DecodeBC4(data, width, height, depth, levels, layers, Format == Format.Bc4Snorm);
+                        result = BCnDecoder.DecodeBC4(result, width, height, depth, levels, layers, Format == Format.Bc4Snorm);
                         break;
                     case Format.Bc5Snorm:
                     case Format.Bc5Unorm:
-                        data = BCnDecoder.DecodeBC5(data, width, height, depth, levels, layers, Format == Format.Bc5Snorm);
+                        result = BCnDecoder.DecodeBC5(result, width, height, depth, levels, layers, Format == Format.Bc5Snorm);
                         break;
                     case Format.Bc6HSfloat:
                     case Format.Bc6HUfloat:
-                        data = BCnDecoder.DecodeBC6(data, width, height, depth, levels, layers, Format == Format.Bc6HSfloat);
+                        result = BCnDecoder.DecodeBC6(result, width, height, depth, levels, layers, Format == Format.Bc6HSfloat);
                         break;
                     case Format.Bc7Srgb:
                     case Format.Bc7Unorm:
-                        data = BCnDecoder.DecodeBC7(data, width, height, depth, levels, layers);
+                        result = BCnDecoder.DecodeBC7(result, width, height, depth, levels, layers);
                         break;
                 }
             }
 
-            return data;
+            return result;
         }
 
         /// <summary>
@@ -1445,7 +1443,6 @@ namespace Ryujinx.Graphics.Gpu.Image
             DisposeTextures();
 
             HostTexture = hostTexture;
-            InvalidatedSequence++;
         }
 
         /// <summary>
@@ -1600,6 +1597,8 @@ namespace Ryujinx.Graphics.Gpu.Image
         /// </summary>
         private void DisposeTextures()
         {
+            InvalidatedSequence++;
+
             _currentData = null;
             HostTexture.Release();
 
@@ -1634,8 +1633,6 @@ namespace Ryujinx.Graphics.Gpu.Image
         {
             DisposeTextures();
 
-            Disposed?.Invoke(this);
-
             if (Group.Storage == this)
             {
                 Group.Dispose();

Ryujinx.Graphics.Gpu/Image/TextureGroup.cs

@@ -1,4 +1,5 @@
-using Ryujinx.Cpu.Tracking;
+using Ryujinx.Common.Memory;
+using Ryujinx.Cpu.Tracking;
 using Ryujinx.Graphics.GAL;
 using Ryujinx.Graphics.Gpu.Memory;
 using Ryujinx.Graphics.Texture;
@@ -348,9 +349,9 @@ namespace Ryujinx.Graphics.Gpu.Image
 
                             ReadOnlySpan<byte> data = _physicalMemory.GetSpan(Storage.Range.GetSlice((ulong)offset, (ulong)size));
 
-                            data = Storage.ConvertToHostCompatibleFormat(data, info.BaseLevel, true);
+                            SpanOrArray<byte> result = Storage.ConvertToHostCompatibleFormat(data, info.BaseLevel, true);
 
-                            Storage.SetData(data, info.BaseLayer, info.BaseLevel);
+                            Storage.SetData(result, info.BaseLayer, info.BaseLevel);
 
                             offsetIndex++;
                         }

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

@@ -22,7 +22,7 @@ namespace Ryujinx.Graphics.Gpu.Shader.DiskCache
         private const ushort FileFormatVersionMajor = 1;
         private const ushort FileFormatVersionMinor = 2;
         private const uint FileFormatVersionPacked = ((uint)FileFormatVersionMajor << 16) | FileFormatVersionMinor;
-        private const uint CodeGenVersion = 3697;
+        private const uint CodeGenVersion = 3759;
 
         private const string SharedTocFileName = "shared.toc";
         private const string SharedDataFileName = "shared.data";

Ryujinx.Graphics.Gpu/Window.cs

@@ -123,7 +123,8 @@ namespace Ryujinx.Graphics.Gpu
         /// <param name="releaseCallback">Texture release callback</param>
         /// <param name="userObj">User defined object passed to the release callback</param>
         /// <exception cref="ArgumentException">Thrown when <paramref name="pid"/> is invalid</exception>
-        public void EnqueueFrameThreadSafe(
+        /// <returns>True if the frame was added to the queue, false otherwise</returns>
+        public bool EnqueueFrameThreadSafe(
             ulong                      pid,
             ulong                      address,
             int                        width,
@@ -140,7 +141,7 @@ namespace Ryujinx.Graphics.Gpu
         {
             if (!_context.PhysicalMemoryRegistry.TryGetValue(pid, out var physicalMemory))
             {
-                throw new ArgumentException("The PID is invalid or the process was not registered", nameof(pid));
+                return false;
             }
 
             FormatInfo formatInfo = new FormatInfo(format, 1, 1, bytesPerPixel, 4);
@@ -184,6 +185,8 @@ namespace Ryujinx.Graphics.Gpu
                 acquireCallback,
                 releaseCallback,
                 userObj));
+
+            return true;
         }
 
         /// <summary>

Ryujinx.Graphics.OpenGL/Image/IntermmediatePool.cs

@@ -1,98 +0,0 @@
-using Ryujinx.Graphics.GAL;
-using System;
-using System.Collections.Generic;
-
-namespace Ryujinx.Graphics.OpenGL.Image
-{
-    class IntermmediatePool : IDisposable
-    {
-        private readonly OpenGLRenderer _renderer;
-        private readonly List<TextureView> _entries;
-
-        public IntermmediatePool(OpenGLRenderer renderer)
-        {
-            _renderer = renderer;
-            _entries = new List<TextureView>();
-        }
-
-        public TextureView GetOrCreateWithAtLeast(
-            Target target,
-            int blockWidth,
-            int blockHeight,
-            int bytesPerPixel,
-            Format format,
-            int width,
-            int height,
-            int depth,
-            int levels)
-        {
-            TextureView entry;
-
-            for (int i = 0; i < _entries.Count; i++)
-            {
-                entry = _entries[i];
-
-                if (entry.Target == target && entry.Format == format)
-                {
-                    if (entry.Width < width || entry.Height < height || entry.Info.Depth < depth || entry.Info.Levels < levels)
-                    {
-                        width = Math.Max(width, entry.Width);
-                        height = Math.Max(height, entry.Height);
-                        depth = Math.Max(depth, entry.Info.Depth);
-                        levels = Math.Max(levels, entry.Info.Levels);
-
-                        entry.Dispose();
-                        entry = CreateNew(target, blockWidth, blockHeight, bytesPerPixel, format, width, height, depth, levels);
-                        _entries[i] = entry;
-                    }
-
-                    return entry;
-                }
-            }
-
-            entry = CreateNew(target, blockWidth, blockHeight, bytesPerPixel, format, width, height, depth, levels);
-            _entries.Add(entry);
-
-            return entry;
-        }
-
-        private TextureView CreateNew(
-            Target target,
-            int blockWidth,
-            int blockHeight,
-            int bytesPerPixel,
-            Format format,
-            int width,
-            int height,
-            int depth,
-            int levels)
-        {
-            return (TextureView)_renderer.CreateTexture(new TextureCreateInfo(
-                width,
-                height,
-                depth,
-                levels,
-                1,
-                blockWidth,
-                blockHeight,
-                bytesPerPixel,
-                format,
-                DepthStencilMode.Depth,
-                target,
-                SwizzleComponent.Red,
-                SwizzleComponent.Green,
-                SwizzleComponent.Blue,
-                SwizzleComponent.Alpha), 1f);
-        }
-
-        public void Dispose()
-        {
-            foreach (TextureView entry in _entries)
-            {
-                entry.Dispose();
-            }
-
-            _entries.Clear();
-        }
-    }
-}

Ryujinx.Graphics.OpenGL/Image/TextureBuffer.cs

@@ -1,4 +1,5 @@
 using OpenTK.Graphics.OpenGL;
+using Ryujinx.Common.Memory;
 using Ryujinx.Graphics.GAL;
 using System;
 
@@ -48,17 +49,19 @@ namespace Ryujinx.Graphics.OpenGL.Image
             return GetData();
         }
 
-        public void SetData(ReadOnlySpan<byte> data)
+        public void SetData(SpanOrArray<byte> data)
         {
-            Buffer.SetData(_buffer, _bufferOffset, data.Slice(0, Math.Min(data.Length, _bufferSize)));
+            var dataSpan = data.AsSpan();
+
+            Buffer.SetData(_buffer, _bufferOffset, dataSpan.Slice(0, Math.Min(dataSpan.Length, _bufferSize)));
         }
 
-        public void SetData(ReadOnlySpan<byte> data, int layer, int level)
+        public void SetData(SpanOrArray<byte> data, int layer, int level)
         {
             throw new NotSupportedException();
         }
 
-        public void SetData(ReadOnlySpan<byte> data, int layer, int level, Rectangle<int> region)
+        public void SetData(SpanOrArray<byte> data, int layer, int level, Rectangle<int> region)
         {
             throw new NotSupportedException();
         }

Ryujinx.Graphics.OpenGL/Image/TextureCopy.cs

@@ -9,8 +9,6 @@ namespace Ryujinx.Graphics.OpenGL.Image
     {
         private readonly OpenGLRenderer _renderer;
 
-        public IntermmediatePool IntermmediatePool { get; }
-
         private int _srcFramebuffer;
         private int _dstFramebuffer;
 
@@ -20,7 +18,6 @@ namespace Ryujinx.Graphics.OpenGL.Image
         public TextureCopy(OpenGLRenderer renderer)
         {
             _renderer = renderer;
-            IntermmediatePool = new IntermmediatePool(renderer);
         }
 
         public void Copy(
@@ -517,8 +514,6 @@ namespace Ryujinx.Graphics.OpenGL.Image
 
                 _copyPboHandle = 0;
             }
-
-            IntermmediatePool.Dispose();
         }
     }
 }

Ryujinx.Graphics.OpenGL/Image/TextureCopyMS.cs

@@ -0,0 +1,276 @@
+using OpenTK.Graphics.OpenGL;
+using Ryujinx.Graphics.GAL;
+using System;
+using System.Numerics;
+
+namespace Ryujinx.Graphics.OpenGL.Image
+{
+    class TextureCopyMS
+    {
+        private const string ComputeShaderMSToNonMS = @"#version 450 core
+
+layout (binding = 0, $FORMAT$) uniform uimage2DMS imgIn;
+layout (binding = 1, $FORMAT$) uniform uimage2D imgOut;
+
+layout (local_size_x = 32, local_size_y = 32, local_size_z = 1) in;
+
+void main()
+{
+    uvec2 coords = gl_GlobalInvocationID.xy;
+    ivec2 imageSz = imageSize(imgOut);
+    if (int(coords.x) >= imageSz.x || int(coords.y) >= imageSz.y)
+    {
+        return;
+    }
+    int inSamples = imageSamples(imgIn);
+    int samplesInXLog2 = 0;
+    int samplesInYLog2 = 0;
+    switch (inSamples)
+    {
+        case 2:
+            samplesInXLog2 = 1;
+            break;
+        case 4:
+            samplesInXLog2 = 1;
+            samplesInYLog2 = 1;
+            break;
+        case 8:
+            samplesInXLog2 = 2;
+            samplesInYLog2 = 1;
+            break;
+        case 16:
+            samplesInXLog2 = 2;
+            samplesInYLog2 = 2;
+            break;
+    }
+    int samplesInX = 1 << samplesInXLog2;
+    int samplesInY = 1 << samplesInYLog2;
+    int sampleIdx = (int(coords.x) & (samplesInX - 1)) | ((int(coords.y) & (samplesInY - 1)) << samplesInXLog2);
+    uvec4 value = imageLoad(imgIn, ivec2(int(coords.x) >> samplesInXLog2, int(coords.y) >> samplesInYLog2), sampleIdx);
+    imageStore(imgOut, ivec2(coords), value);
+}";
+
+        private const string ComputeShaderNonMSToMS = @"#version 450 core
+
+layout (binding = 0, $FORMAT$) uniform uimage2D imgIn;
+layout (binding = 1, $FORMAT$) uniform uimage2DMS imgOut;
+
+layout (local_size_x = 32, local_size_y = 32, local_size_z = 1) in;
+
+void main()
+{
+    uvec2 coords = gl_GlobalInvocationID.xy;
+    ivec2 imageSz = imageSize(imgIn);
+    if (int(coords.x) >= imageSz.x || int(coords.y) >= imageSz.y)
+    {
+        return;
+    }
+    int outSamples = imageSamples(imgOut);
+    int samplesInXLog2 = 0;
+    int samplesInYLog2 = 0;
+    switch (outSamples)
+    {
+        case 2:
+            samplesInXLog2 = 1;
+            break;
+        case 4:
+            samplesInXLog2 = 1;
+            samplesInYLog2 = 1;
+            break;
+        case 8:
+            samplesInXLog2 = 2;
+            samplesInYLog2 = 1;
+            break;
+        case 16:
+            samplesInXLog2 = 2;
+            samplesInYLog2 = 2;
+            break;
+    }
+    int samplesInX = 1 << samplesInXLog2;
+    int samplesInY = 1 << samplesInYLog2;
+    int sampleIdx = (int(coords.x) & (samplesInX - 1)) | ((int(coords.y) & (samplesInY - 1)) << samplesInXLog2);
+    uvec4 value = imageLoad(imgIn, ivec2(coords));
+    imageStore(imgOut, ivec2(int(coords.x) >> samplesInXLog2, int(coords.y) >> samplesInYLog2), sampleIdx, value);
+}";
+
+        private readonly OpenGLRenderer _renderer;
+        private int[] _msToNonMSProgramHandles;
+        private int[] _nonMSToMSProgramHandles;
+
+        public TextureCopyMS(OpenGLRenderer renderer)
+        {
+            _renderer = renderer;
+            _msToNonMSProgramHandles = new int[5];
+            _nonMSToMSProgramHandles = new int[5];
+        }
+
+        public void CopyMSToNonMS(ITextureInfo src, ITextureInfo dst, int srcLayer, int dstLayer, int depth)
+        {
+            TextureCreateInfo srcInfo = src.Info;
+            TextureCreateInfo dstInfo = dst.Info;
+
+            int srcHandle = CreateViewIfNeeded(src);
+            int dstHandle = CreateViewIfNeeded(dst);
+
+            int dstWidth = dstInfo.Width;
+            int dstHeight = dstInfo.Height;
+
+            GL.UseProgram(GetMSToNonMSShader(srcInfo.BytesPerPixel));
+
+            for (int z = 0; z < depth; z++)
+            {
+                GL.BindImageTexture(0, srcHandle, 0, false, srcLayer + z, TextureAccess.ReadOnly, GetFormat(srcInfo.BytesPerPixel));
+                GL.BindImageTexture(1, dstHandle, 0, false, dstLayer + z, TextureAccess.WriteOnly, GetFormat(dstInfo.BytesPerPixel));
+
+                GL.DispatchCompute((dstWidth + 31) / 32, (dstHeight + 31) / 32, 1);
+            }
+
+            Pipeline pipeline = (Pipeline)_renderer.Pipeline;
+
+            pipeline.RestoreProgram();
+            pipeline.RestoreImages1And2();
+
+            DestroyViewIfNeeded(src, srcHandle);
+            DestroyViewIfNeeded(dst, dstHandle);
+        }
+
+        public void CopyNonMSToMS(ITextureInfo src, ITextureInfo dst, int srcLayer, int dstLayer, int depth)
+        {
+            TextureCreateInfo srcInfo = src.Info;
+            TextureCreateInfo dstInfo = dst.Info;
+
+            int srcHandle = CreateViewIfNeeded(src);
+            int dstHandle = CreateViewIfNeeded(dst);
+
+            int srcWidth = srcInfo.Width;
+            int srcHeight = srcInfo.Height;
+
+            GL.UseProgram(GetNonMSToMSShader(srcInfo.BytesPerPixel));
+
+            for (int z = 0; z < depth; z++)
+            {
+                GL.BindImageTexture(0, srcHandle, 0, false, srcLayer + z, TextureAccess.ReadOnly, GetFormat(srcInfo.BytesPerPixel));
+                GL.BindImageTexture(1, dstHandle, 0, false, dstLayer + z, TextureAccess.WriteOnly, GetFormat(dstInfo.BytesPerPixel));
+
+                GL.DispatchCompute((srcWidth + 31) / 32, (srcHeight + 31) / 32, 1);
+            }
+
+            Pipeline pipeline = (Pipeline)_renderer.Pipeline;
+
+            pipeline.RestoreProgram();
+            pipeline.RestoreImages1And2();
+
+            DestroyViewIfNeeded(src, srcHandle);
+            DestroyViewIfNeeded(dst, dstHandle);
+        }
+
+        private static SizedInternalFormat GetFormat(int bytesPerPixel)
+        {
+            return bytesPerPixel switch
+            {
+                1 => SizedInternalFormat.R8ui,
+                2 => SizedInternalFormat.R16ui,
+                4 => SizedInternalFormat.R32ui,
+                8 => SizedInternalFormat.Rg32ui,
+                16 => SizedInternalFormat.Rgba32ui,
+                _ => throw new ArgumentException($"Invalid bytes per pixel {bytesPerPixel}.")
+            };
+        }
+
+        private static int CreateViewIfNeeded(ITextureInfo texture)
+        {
+            // Binding sRGB textures as images doesn't work on NVIDIA,
+            // we need to create and bind a RGBA view for it to work.
+            if (texture.Info.Format == Format.R8G8B8A8Srgb)
+            {
+                int handle = GL.GenTexture();
+
+                GL.TextureView(
+                    handle,
+                    texture.Info.Target.Convert(),
+                    texture.Storage.Handle,
+                    PixelInternalFormat.Rgba8,
+                    texture.FirstLevel,
+                    1,
+                    texture.FirstLayer,
+                    texture.Info.GetLayers());
+
+                return handle;
+            }
+
+            return texture.Handle;
+        }
+
+        private static void DestroyViewIfNeeded(ITextureInfo info, int handle)
+        {
+            if (info.Handle != handle)
+            {
+                GL.DeleteTexture(handle);
+            }
+        }
+
+        private int GetMSToNonMSShader(int bytesPerPixel)
+        {
+            return GetShader(ComputeShaderMSToNonMS, _msToNonMSProgramHandles, bytesPerPixel);
+        }
+
+        private int GetNonMSToMSShader(int bytesPerPixel)
+        {
+            return GetShader(ComputeShaderNonMSToMS, _nonMSToMSProgramHandles, bytesPerPixel);
+        }
+
+        private int GetShader(string code, int[] programHandles, int bytesPerPixel)
+        {
+            int index = BitOperations.Log2((uint)bytesPerPixel);
+
+            if (programHandles[index] == 0)
+            {
+                int csHandle = GL.CreateShader(ShaderType.ComputeShader);
+
+                string format = new[] { "r8ui", "r16ui", "r32ui", "rg32ui", "rgba32ui" }[index];
+
+                GL.ShaderSource(csHandle, code.Replace("$FORMAT$", format));
+                GL.CompileShader(csHandle);
+
+                int programHandle = GL.CreateProgram();
+
+                GL.AttachShader(programHandle, csHandle);
+                GL.LinkProgram(programHandle);
+                GL.DetachShader(programHandle, csHandle);
+                GL.DeleteShader(csHandle);
+
+                GL.GetProgram(programHandle, GetProgramParameterName.LinkStatus, out int status);
+
+                if (status == 0)
+                {
+                    throw new Exception(GL.GetProgramInfoLog(programHandle));
+                }
+
+                programHandles[index] = programHandle;
+            }
+
+            return programHandles[index];
+        }
+
+        public void Dispose()
+        {
+            for (int i = 0; i < _msToNonMSProgramHandles.Length; i++)
+            {
+                if (_msToNonMSProgramHandles[i] != 0)
+                {
+                    GL.DeleteProgram(_msToNonMSProgramHandles[i]);
+                    _msToNonMSProgramHandles[i] = 0;
+                }
+            }
+
+            for (int i = 0; i < _nonMSToMSProgramHandles.Length; i++)
+            {
+                if (_nonMSToMSProgramHandles[i] != 0)
+                {
+                    GL.DeleteProgram(_nonMSToMSProgramHandles[i]);
+                    _nonMSToMSProgramHandles[i] = 0;
+                }
+            }
+        }
+    }
+}

Ryujinx.Graphics.OpenGL/Image/TextureView.cs

@@ -1,5 +1,6 @@
 using OpenTK.Graphics.OpenGL;
 using Ryujinx.Common;
+using Ryujinx.Common.Memory;
 using Ryujinx.Graphics.GAL;
 using System;
 
@@ -116,28 +117,15 @@ namespace Ryujinx.Graphics.OpenGL.Image
         {
             TextureView destinationView = (TextureView)destination;
 
-            if (destinationView.Target.IsMultisample() || Target.IsMultisample())
+            if (!destinationView.Target.IsMultisample() && Target.IsMultisample())
             {
-                Extents2D srcRegion = new Extents2D(0, 0, Width, Height);
-                Extents2D dstRegion = new Extents2D(0, 0, destinationView.Width, destinationView.Height);
-
-                TextureView intermmediate = _renderer.TextureCopy.IntermmediatePool.GetOrCreateWithAtLeast(
-                    GetIntermmediateTarget(Target),
-                    Info.BlockWidth,
-                    Info.BlockHeight,
-                    Info.BytesPerPixel,
-                    Format,
-                    Width,
-                    Height,
-                    Info.Depth,
-                    Info.Levels);
-
-                GL.Disable(EnableCap.FramebufferSrgb);
-
-                _renderer.TextureCopy.Copy(this, intermmediate, srcRegion, srcRegion, true);
-                _renderer.TextureCopy.Copy(intermmediate, destinationView, srcRegion, dstRegion, true, 0, firstLayer, 0, firstLevel);
-
-                GL.Enable(EnableCap.FramebufferSrgb);
+                int layers = Math.Min(Info.GetLayers(), destinationView.Info.GetLayers() - firstLayer);
+                _renderer.TextureCopyMS.CopyMSToNonMS(this, destinationView, 0, firstLayer, layers);
+            }
+            else if (destinationView.Target.IsMultisample() && !Target.IsMultisample())
+            {
+                int layers = Math.Min(Info.GetLayers(), destinationView.Info.GetLayers() - firstLayer);
+                _renderer.TextureCopyMS.CopyNonMSToMS(this, destinationView, 0, firstLayer, layers);
             }
             else
             {
@@ -149,28 +137,13 @@ namespace Ryujinx.Graphics.OpenGL.Image
         {
             TextureView destinationView = (TextureView)destination;
 
-            if (destinationView.Target.IsMultisample() || Target.IsMultisample())
+            if (!destinationView.Target.IsMultisample() && Target.IsMultisample())
             {
-                Extents2D srcRegion = new Extents2D(0, 0, Width, Height);
-                Extents2D dstRegion = new Extents2D(0, 0, destinationView.Width, destinationView.Height);
-
-                TextureView intermmediate = _renderer.TextureCopy.IntermmediatePool.GetOrCreateWithAtLeast(
-                    GetIntermmediateTarget(Target),
-                    Info.BlockWidth,
-                    Info.BlockHeight,
-                    Info.BytesPerPixel,
-                    Format,
-                    Math.Max(1, Width >> srcLevel),
-                    Math.Max(1, Height >> srcLevel),
-                    1,
-                    1);
-
-                GL.Disable(EnableCap.FramebufferSrgb);
-
-                _renderer.TextureCopy.Copy(this, intermmediate, srcRegion, srcRegion, true, srcLayer, 0, srcLevel, 0, 1, 1);
-                _renderer.TextureCopy.Copy(intermmediate, destinationView, srcRegion, dstRegion, true, 0, dstLayer, 0, dstLevel, 1, 1);
-
-                GL.Enable(EnableCap.FramebufferSrgb);
+                _renderer.TextureCopyMS.CopyMSToNonMS(this, destinationView, srcLayer, dstLayer,1);
+            }
+            else if (destinationView.Target.IsMultisample() && !Target.IsMultisample())
+            {
+                _renderer.TextureCopyMS.CopyNonMSToMS(this, destinationView, srcLayer, dstLayer, 1);
             }
             else
             {
@@ -178,17 +151,6 @@ namespace Ryujinx.Graphics.OpenGL.Image
             }
         }
 
-        private static Target GetIntermmediateTarget(Target srcTarget)
-        {
-            return srcTarget switch
-            {
-                Target.Texture2D => Target.Texture2DMultisample,
-                Target.Texture2DArray => Target.Texture2DMultisampleArray,
-                Target.Texture2DMultisampleArray => Target.Texture2DArray,
-                _ => Target.Texture2D
-            };
-        }
-
         public void CopyTo(ITexture destination, Extents2D srcRegion, Extents2D dstRegion, bool linearFilter)
         {
             _renderer.TextureCopy.Copy(this, (TextureView)destination, srcRegion, dstRegion, linearFilter);
@@ -356,32 +318,36 @@ namespace Ryujinx.Graphics.OpenGL.Image
             }
         }
 
-        public void SetData(ReadOnlySpan<byte> data)
+        public void SetData(SpanOrArray<byte> data)
         {
+            var dataSpan = data.AsSpan();
+
             if (Format == Format.S8UintD24Unorm)
             {
-                data = FormatConverter.ConvertS8D24ToD24S8(data);
+                dataSpan = FormatConverter.ConvertS8D24ToD24S8(dataSpan);
             }
 
             unsafe
             {
-                fixed (byte* ptr = data)
+                fixed (byte* ptr = dataSpan)
                 {
-                    ReadFrom((IntPtr)ptr, data.Length);
+                    ReadFrom((IntPtr)ptr, dataSpan.Length);
                 }
             }
         }
 
-        public void SetData(ReadOnlySpan<byte> data, int layer, int level)
+        public void SetData(SpanOrArray<byte> data, int layer, int level)
         {
+            var dataSpan = data.AsSpan();
+
             if (Format == Format.S8UintD24Unorm)
             {
-                data = FormatConverter.ConvertS8D24ToD24S8(data);
+                dataSpan = FormatConverter.ConvertS8D24ToD24S8(dataSpan);
             }
 
             unsafe
             {
-                fixed (byte* ptr = data)
+                fixed (byte* ptr = dataSpan)
                 {
                     int width = Math.Max(Info.Width >> level, 1);
                     int height = Math.Max(Info.Height >> level, 1);
@@ -391,11 +357,13 @@ namespace Ryujinx.Graphics.OpenGL.Image
             }
         }
 
-        public void SetData(ReadOnlySpan<byte> data, int layer, int level, Rectangle<int> region)
+        public void SetData(SpanOrArray<byte> data, int layer, int level, Rectangle<int> region)
         {
+            var dataSpan = data.AsSpan();
+
             if (Format == Format.S8UintD24Unorm)
             {
-                data = FormatConverter.ConvertS8D24ToD24S8(data);
+                dataSpan = FormatConverter.ConvertS8D24ToD24S8(dataSpan);
             }
 
             int wInBlocks = BitUtils.DivRoundUp(region.Width, Info.BlockWidth);
@@ -403,7 +371,7 @@ namespace Ryujinx.Graphics.OpenGL.Image
 
             unsafe
             {
-                fixed (byte* ptr = data)
+                fixed (byte* ptr = dataSpan)
                 {
                     ReadFrom2D(
                         (IntPtr)ptr,

Ryujinx.Graphics.OpenGL/OpenGLRenderer.cs

@@ -24,6 +24,7 @@ namespace Ryujinx.Graphics.OpenGL
         private TextureCopy _textureCopy;
         private TextureCopy _backgroundTextureCopy;
         internal TextureCopy TextureCopy => BackgroundContextWorker.InBackground ? _backgroundTextureCopy : _textureCopy;
+        internal TextureCopyMS TextureCopyMS { get; }
 
         private Sync _sync;
 
@@ -48,6 +49,7 @@ namespace Ryujinx.Graphics.OpenGL
             _window = new Window(this);
             _textureCopy = new TextureCopy(this);
             _backgroundTextureCopy = new TextureCopy(this);
+            TextureCopyMS = new TextureCopyMS(this);
             _sync = new Sync();
             PersistentBuffers = new PersistentBuffers();
             ResourcePool = new ResourcePool();
@@ -211,6 +213,7 @@ namespace Ryujinx.Graphics.OpenGL
         {
             _textureCopy.Dispose();
             _backgroundTextureCopy.Dispose();
+            TextureCopyMS.Dispose();
             PersistentBuffers.Dispose();
             ResourcePool.Dispose();
             _pipeline.Dispose();

Ryujinx.Graphics.OpenGL/Pipeline.cs

@@ -11,6 +11,8 @@ namespace Ryujinx.Graphics.OpenGL
 {
     class Pipeline : IPipeline, IDisposable
     {
+        private const int SavedImages = 2;
+
         private readonly DrawTextureEmulation _drawTexture;
 
         internal ulong DrawCount { get; private set; }
@@ -46,6 +48,7 @@ namespace Ryujinx.Graphics.OpenGL
         private Vector4<float>[] _renderScale = new Vector4<float>[73];
         private int _fragmentScaleCount;
 
+        private (TextureBase, Format)[] _images;
         private TextureBase _unit0Texture;
         private Sampler _unit0Sampler;
 
@@ -78,6 +81,8 @@ namespace Ryujinx.Graphics.OpenGL
             _fragmentOutputMap = uint.MaxValue;
             _componentMasks = uint.MaxValue;
 
+            _images = new (TextureBase, Format)[SavedImages];
+
             var defaultScale = new Vector4<float> { X = 1f, Y = 0f, Z = 0f, W = 0f };
             new Span<Vector4<float>>(_renderScale).Fill(defaultScale);
 
@@ -907,6 +912,11 @@ namespace Ryujinx.Graphics.OpenGL
 
         public void SetImage(int binding, ITexture texture, Format imageFormat)
         {
+            if ((uint)binding < SavedImages)
+            {
+                _images[binding] = (texture as TextureBase, imageFormat);
+            }
+
             if (texture == null)
             {
                 return;
@@ -1608,6 +1618,32 @@ namespace Ryujinx.Graphics.OpenGL
             }
         }
 
+        public void RestoreProgram()
+        {
+            _program?.Bind();
+        }
+
+        public void RestoreImages1And2()
+        {
+            for (int i = 0; i < SavedImages; i++)
+            {
+                (TextureBase texBase, Format imageFormat) = _images[i];
+
+                if (texBase != null)
+                {
+                    SizedInternalFormat format = FormatTable.GetImageFormat(imageFormat);
+
+                    if (format != 0)
+                    {
+                        GL.BindImageTexture(i, texBase.Handle, 0, true, 0, TextureAccess.ReadWrite, format);
+                        continue;
+                    }
+                }
+
+                GL.BindImageTexture(i, 0, 0, true, 0, TextureAccess.ReadWrite, SizedInternalFormat.Rgba8);
+            }
+        }
+
         public bool TryHostConditionalRendering(ICounterEvent value, ulong compare, bool isEqual)
         {
             if (value is CounterQueueEvent)

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

@@ -2,6 +2,7 @@ using Ryujinx.Common;
 using Ryujinx.Graphics.Shader.StructuredIr;
 using Ryujinx.Graphics.Shader.Translation;
 using System;
+using System.Collections.Generic;
 using System.Linq;
 using System.Numerics;
 
@@ -163,9 +164,17 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                 }
                 else if (context.Config.Stage == ShaderStage.TessellationEvaluation)
                 {
+                    bool tessCw = context.Config.GpuAccessor.QueryTessCw();
+
+                    if (context.Config.Options.TargetApi == TargetApi.Vulkan)
+                    {
+                        // We invert the front face on Vulkan backend, so we need to do that here aswell.
+                        tessCw = !tessCw;
+                    }
+
                     string patchType = context.Config.GpuAccessor.QueryTessPatchType().ToGlsl();
                     string spacing = context.Config.GpuAccessor.QueryTessSpacing().ToGlsl();
-                    string windingOrder = context.Config.GpuAccessor.QueryTessCw() ? "cw" : "ccw";
+                    string windingOrder = tessCw ? "cw" : "ccw";
 
                     context.AppendLine($"layout ({patchType}, {spacing}, {windingOrder}) in;");
                     context.AppendLine();
@@ -185,14 +194,14 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                     context.AppendLine();
                 }
 
-                if (context.Config.UsedInputAttributesPerPatch != 0)
+                if (context.Config.UsedInputAttributesPerPatch.Count != 0)
                 {
                     DeclareInputAttributesPerPatch(context, context.Config.UsedInputAttributesPerPatch);
 
                     context.AppendLine();
                 }
 
-                if (context.Config.UsedOutputAttributesPerPatch != 0)
+                if (context.Config.UsedOutputAttributesPerPatch.Count != 0)
                 {
                     DeclareUsedOutputAttributesPerPatch(context, context.Config.UsedOutputAttributesPerPatch);
 
@@ -248,7 +257,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
 
                     DeclareSupportUniformBlock(context, context.Config.Stage, scaleElements);
 
-                    if (context.Config.UsedFeatures.HasFlag(FeatureFlags.IntegerSampling))
+                    if (context.Config.UsedFeatures.HasFlag(FeatureFlags.IntegerSampling) && scaleElements != 0)
                     {
                         AppendHelperFunction(context, $"Ryujinx.Graphics.Shader/CodeGen/Glsl/HelperFunctions/TexelFetchScale_{stage}.glsl");
                         context.AppendLine();
@@ -509,13 +518,11 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
             }
         }
 
-        private static void DeclareInputAttributesPerPatch(CodeGenContext context, int usedAttributes)
+        private static void DeclareInputAttributesPerPatch(CodeGenContext context, HashSet<int> attrs)
         {
-            while (usedAttributes != 0)
+            foreach (int attr in attrs.OrderBy(x => x))
             {
-                int index = BitOperations.TrailingZeroCount(usedAttributes);
-                DeclareInputAttributePerPatch(context, index);
-                usedAttributes &= ~(1 << index);
+                DeclareInputAttributePerPatch(context, attr);
             }
         }
 
@@ -566,16 +573,10 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
 
         private static void DeclareInputAttributePerPatch(CodeGenContext context, int attr)
         {
-            string layout = string.Empty;
-
-            if (context.Config.Options.TargetApi == TargetApi.Vulkan)
-            {
-                layout = $"layout (location = {32 + attr}) ";
-            }
-
+            int location = context.Config.GetPerPatchAttributeLocation(attr);
             string name = $"{DefaultNames.PerPatchAttributePrefix}{attr}";
 
-            context.AppendLine($"{layout}patch in vec4 {name};");
+            context.AppendLine($"layout (location = {location}) patch in vec4 {name};");
         }
 
         private static void DeclareOutputAttributes(CodeGenContext context, StructuredProgramInfo info)
@@ -624,28 +625,20 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
             }
         }
 
-        private static void DeclareUsedOutputAttributesPerPatch(CodeGenContext context, int usedAttributes)
+        private static void DeclareUsedOutputAttributesPerPatch(CodeGenContext context, HashSet<int> attrs)
         {
-            while (usedAttributes != 0)
+            foreach (int attr in attrs.OrderBy(x => x))
             {
-                int index = BitOperations.TrailingZeroCount(usedAttributes);
-                DeclareOutputAttributePerPatch(context, index);
-                usedAttributes &= ~(1 << index);
+                DeclareOutputAttributePerPatch(context, attr);
             }
         }
 
         private static void DeclareOutputAttributePerPatch(CodeGenContext context, int attr)
         {
-            string layout = string.Empty;
-
-            if (context.Config.Options.TargetApi == TargetApi.Vulkan)
-            {
-                layout = $"layout (location = {32 + attr}) ";
-            }
-
+            int location = context.Config.GetPerPatchAttributeLocation(attr);
             string name = $"{DefaultNames.PerPatchAttributePrefix}{attr}";
 
-            context.AppendLine($"{layout}patch out vec4 {name};");
+            context.AppendLine($"layout (location = {location}) patch out vec4 {name};");
         }
 
         private static void DeclareSupportUniformBlock(CodeGenContext context, ShaderStage stage, int scaleElements)

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

@@ -28,12 +28,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
 
         private static Dictionary<int, BuiltInAttribute> _builtInAttributes = new Dictionary<int, BuiltInAttribute>()
         {
-            { AttributeConsts.TessLevelOuter0, new BuiltInAttribute("gl_TessLevelOuter[0]", VariableType.F32)  },
-            { AttributeConsts.TessLevelOuter1, new BuiltInAttribute("gl_TessLevelOuter[1]", VariableType.F32)  },
-            { AttributeConsts.TessLevelOuter2, new BuiltInAttribute("gl_TessLevelOuter[2]", VariableType.F32)  },
-            { AttributeConsts.TessLevelOuter3, new BuiltInAttribute("gl_TessLevelOuter[3]", VariableType.F32)  },
-            { AttributeConsts.TessLevelInner0, new BuiltInAttribute("gl_TessLevelInner[0]", VariableType.F32)  },
-            { AttributeConsts.TessLevelInner1, new BuiltInAttribute("gl_TessLevelInner[1]", VariableType.F32)  },
             { AttributeConsts.Layer,         new BuiltInAttribute("gl_Layer",           VariableType.S32)  },
             { AttributeConsts.PointSize,     new BuiltInAttribute("gl_PointSize",       VariableType.F32)  },
             { AttributeConsts.PositionX,     new BuiltInAttribute("gl_Position.x",      VariableType.F32)  },
@@ -170,7 +164,29 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
             value &= AttributeConsts.Mask & ~3;
             char swzMask = GetSwizzleMask((value >> 2) & 3);
 
-            if (value >= AttributeConsts.UserAttributeBase && value < AttributeConsts.UserAttributeEnd)
+            if (perPatch)
+            {
+                if (value >= AttributeConsts.UserAttributePerPatchBase && value < AttributeConsts.UserAttributePerPatchEnd)
+                {
+                    value -= AttributeConsts.UserAttributePerPatchBase;
+
+                    return $"{DefaultNames.PerPatchAttributePrefix}{(value >> 4)}.{swzMask}";
+                }
+                else if (value < AttributeConsts.UserAttributePerPatchBase)
+                {
+                    return value switch
+                    {
+                        AttributeConsts.TessLevelOuter0 => "gl_TessLevelOuter[0]",
+                        AttributeConsts.TessLevelOuter1 => "gl_TessLevelOuter[1]",
+                        AttributeConsts.TessLevelOuter2 => "gl_TessLevelOuter[2]",
+                        AttributeConsts.TessLevelOuter3 => "gl_TessLevelOuter[3]",
+                        AttributeConsts.TessLevelInner0 => "gl_TessLevelInner[0]",
+                        AttributeConsts.TessLevelInner1 => "gl_TessLevelInner[1]",
+                        _ => null
+                    };
+                }
+            }
+            else if (value >= AttributeConsts.UserAttributeBase && value < AttributeConsts.UserAttributeEnd)
             {
                 value -= AttributeConsts.UserAttributeBase;
 
@@ -180,11 +196,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
 
                 bool indexable = config.UsedFeatures.HasFlag(isOutAttr ? FeatureFlags.OaIndexing : FeatureFlags.IaIndexing);
 
-                if (!indexable && perPatch)
-                {
-                    prefix = DefaultNames.PerPatchAttributePrefix;
-                }
-
                 if (indexable)
                 {
                     string name = prefix;
@@ -202,7 +213,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                 {
                     string name = $"{prefix}{(value >> 4)}_{swzMask}";
 
-                    if (!perPatch && AttributeInfo.IsArrayAttributeGlsl(config.Stage, isOutAttr))
+                    if (AttributeInfo.IsArrayAttributeGlsl(config.Stage, isOutAttr))
                     {
                         name += isOutAttr ? "[gl_InvocationID]" : $"[{indexExpr}]";
                     }
@@ -213,7 +224,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
                 {
                     string name = $"{prefix}{(value >> 4)}";
 
-                    if (!perPatch && AttributeInfo.IsArrayAttributeGlsl(config.Stage, isOutAttr))
+                    if (AttributeInfo.IsArrayAttributeGlsl(config.Stage, isOutAttr))
                     {
                         name += isOutAttr ? "[gl_InvocationID]" : $"[{indexExpr}]";
                     }
@@ -277,7 +288,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
 
                     string name = builtInAttr.Name;
 
-                    if (!perPatch && AttributeInfo.IsArrayAttributeGlsl(config.Stage, isOutAttr) && AttributeInfo.IsArrayBuiltIn(value))
+                    if (AttributeInfo.IsArrayAttributeGlsl(config.Stage, isOutAttr) && AttributeInfo.IsArrayBuiltIn(value))
                     {
                         name = isOutAttr ? $"gl_out[gl_InvocationID].{name}" : $"gl_in[{indexExpr}].{name}";
                     }

Ryujinx.Graphics.Shader/CodeGen/Spirv/CodeGenContext.cs

@@ -382,17 +382,13 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
         public Instruction GetAttributePerPatchElemPointer(int attr, bool isOutAttr, out AggregateType elemType)
         {
             var storageClass = isOutAttr ? StorageClass.Output : StorageClass.Input;
-            var attrInfo = AttributeInfo.From(Config, attr, isOutAttr);
+            var attrInfo = AttributeInfo.FromPatch(Config, attr, isOutAttr);
 
             int attrOffset = attrInfo.BaseValue;
-            Instruction ioVariable;
-
-            bool isUserAttr = attr >= AttributeConsts.UserAttributeBase && attr < AttributeConsts.UserAttributeEnd;
+            Instruction ioVariable = isOutAttr ? OutputsPerPatch[attrOffset] : InputsPerPatch[attrOffset];
 
             elemType = attrInfo.Type & AggregateType.ElementTypeMask;
 
-            ioVariable = isOutAttr ? OutputsPerPatch[attrOffset] : InputsPerPatch[attrOffset];
-
             if ((attrInfo.Type & (AggregateType.Array | AggregateType.Vector)) == 0)
             {
                 return ioVariable;
@@ -404,7 +400,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
 
         public Instruction GetAttributePerPatch(AggregateType type, int attr, bool isOutAttr)
         {
-            if (!AttributeInfo.Validate(Config, attr, isOutAttr: false))
+            if (!AttributeInfo.ValidatePerPatch(Config, attr, isOutAttr: false))
             {
                 return GetConstant(type, new AstOperand(IrOperandType.Constant, 0));
             }

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

@@ -403,7 +403,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
 
             foreach (int attr in inputs)
             {
-                if (!AttributeInfo.Validate(context.Config, attr, isOutAttr: false))
+                if (!AttributeInfo.Validate(context.Config, attr, isOutAttr: false, perPatch))
                 {
                     continue;
                 }
@@ -459,7 +459,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
 
             foreach (int attr in outputs)
             {
-                if (!AttributeInfo.Validate(context.Config, attr, isOutAttr: true))
+                if (!AttributeInfo.Validate(context.Config, attr, isOutAttr: true, perPatch))
                 {
                     continue;
                 }
@@ -519,7 +519,9 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
                 ? (isOutAttr ? context.OutputsPerPatch : context.InputsPerPatch)
                 : (isOutAttr ? context.Outputs : context.Inputs);
 
-            var attrInfo = AttributeInfo.From(context.Config, attr, isOutAttr);
+            var attrInfo = perPatch
+                ? AttributeInfo.FromPatch(context.Config, attr, isOutAttr)
+                : AttributeInfo.From(context.Config, attr, isOutAttr);
 
             if (dict.ContainsKey(attrInfo.BaseValue))
             {
@@ -544,11 +546,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
             var spvType = context.TypePointer(storageClass, attrType);
             var spvVar = context.Variable(spvType, storageClass);
 
-            if (perPatch)
-            {
-                context.Decorate(spvVar, Decoration.Patch);
-            }
-
             if (builtInPassthrough)
             {
                 context.Decorate(spvVar, Decoration.PassthroughNV);
@@ -556,6 +553,11 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
 
             if (attrInfo.IsBuiltin)
             {
+                if (perPatch)
+                {
+                    context.Decorate(spvVar, Decoration.Patch);
+                }
+
                 context.Decorate(spvVar, Decoration.BuiltIn, (LiteralInteger)GetBuiltIn(context, attrInfo.BaseValue));
 
                 if (context.Config.TransformFeedbackEnabled && context.Config.LastInVertexPipeline && isOutAttr)
@@ -569,6 +571,14 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
                     }
                 }
             }
+            else if (perPatch)
+            {
+                context.Decorate(spvVar, Decoration.Patch);
+
+                int location = context.Config.GetPerPatchAttributeLocation((attr - AttributeConsts.UserAttributePerPatchBase) / 16);
+
+                context.Decorate(spvVar, Decoration.Location, (LiteralInteger)location);
+            }
             else if (isUserAttr)
             {
                 int location = (attr - AttributeConsts.UserAttributeBase) / 16;

Ryujinx.Graphics.Shader/CodeGen/Spirv/Instructions.cs

@@ -882,7 +882,8 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
             if (src2 is AstOperand operand && operand.Type == OperandType.Constant)
             {
                 int attrOffset = (baseAttr.Value & AttributeConsts.Mask) + (operand.Value << 2);
-                return new OperationResult(resultType, context.GetAttribute(resultType, attrOffset, isOutAttr: false, index));
+                bool isOutAttr = (baseAttr.Value & AttributeConsts.LoadOutputMask) != 0;
+                return new OperationResult(resultType, context.GetAttribute(resultType, attrOffset, isOutAttr, index));
             }
             else
             {

Ryujinx.Graphics.Shader/CodeGen/Spirv/SpirvGenerator.cs

@@ -191,7 +191,15 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
                             break;
                     }
 
-                    if (context.Config.GpuAccessor.QueryTessCw())
+                    bool tessCw = context.Config.GpuAccessor.QueryTessCw();
+
+                    if (context.Config.Options.TargetApi == TargetApi.Vulkan)
+                    {
+                        // We invert the front face on Vulkan backend, so we need to do that here aswell.
+                        tessCw = !tessCw;
+                    }
+
+                    if (tessCw)
                     {
                         context.AddExecutionMode(spvFunc, ExecutionMode.VertexOrderCw);
                     }
@@ -374,10 +382,11 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
                         context.Store(context.GetLocalPointer(dest), source);
                     }
                     else if (dest.Type == OperandType.Attribute || dest.Type == OperandType.AttributePerPatch)
-                    {
-                        if (AttributeInfo.Validate(context.Config, dest.Value, isOutAttr: true))
                     {
                         bool perPatch = dest.Type == OperandType.AttributePerPatch;
+
+                        if (AttributeInfo.Validate(context.Config, dest.Value, isOutAttr: true, perPatch))
+                        {
                             AggregateType elemType;
 
                             var elemPointer = perPatch

Ryujinx.Graphics.Shader/Decoders/Decoder.cs

@@ -306,18 +306,36 @@ namespace Ryujinx.Graphics.Shader.Decoders
                 for (int elemIndex = 0; elemIndex < count; elemIndex++)
                 {
                     int attr = offset + elemIndex * 4;
-                    if (attr >= AttributeConsts.UserAttributeBase && attr < AttributeConsts.UserAttributeEnd)
+
+                    if (perPatch)
+                    {
+                        if (attr >= AttributeConsts.UserAttributePerPatchBase && attr < AttributeConsts.UserAttributePerPatchEnd)
+                        {
+                            int userAttr = attr - AttributeConsts.UserAttributePerPatchBase;
+                            int index = userAttr / 16;
+
+                            if (isStore)
+                            {
+                                config.SetOutputUserAttributePerPatch(index);
+                            }
+                            else
+                            {
+                                config.SetInputUserAttributePerPatch(index);
+                            }
+                        }
+                    }
+                    else if (attr >= AttributeConsts.UserAttributeBase && attr < AttributeConsts.UserAttributeEnd)
                     {
                         int userAttr = attr - AttributeConsts.UserAttributeBase;
                         int index = userAttr / 16;
 
                         if (isStore)
                         {
-                            config.SetOutputUserAttribute(index, perPatch);
+                            config.SetOutputUserAttribute(index);
                         }
                         else
                         {
-                            config.SetInputUserAttribute(index, (userAttr >> 2) & 3, perPatch);
+                            config.SetInputUserAttribute(index, (userAttr >> 2) & 3);
                         }
                     }
 
@@ -359,6 +377,8 @@ namespace Ryujinx.Graphics.Shader.Decoders
 
                 if (lastOp.Name == InstName.Brx && block.Successors.Count == (hasNext ? 1 : 0))
                 {
+                    HashSet<ulong> visited = new HashSet<ulong>();
+
                     InstBrx opBrx = new InstBrx(lastOp.RawOpCode);
                     ulong baseOffset = lastOp.GetAbsoluteAddress();
 
@@ -374,12 +394,17 @@ namespace Ryujinx.Graphics.Shader.Decoders
                     for (int i = 0; i < cbOffsetsCount; i++)
                     {
                         uint targetOffset = config.ConstantBuffer1Read(cbBaseOffset + i * 4);
-                        Block target = getBlock(baseOffset + targetOffset);
+                        ulong targetAddress = baseOffset + targetOffset;
+
+                        if (visited.Add(targetAddress))
+                        {
+                            Block target = getBlock(targetAddress);
                             target.Predecessors.Add(block);
                             block.Successors.Add(target);
                         }
                     }
                 }
+            }
 
             return hasNewTarget;
         }

Ryujinx.Graphics.Shader/Decoders/InstTable.cs

@@ -279,16 +279,16 @@ namespace Ryujinx.Graphics.Shader.Decoders
             Add("1110101110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuatomB2,    InstEmit.SuatomB2,    InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc);
             Add("1110101011010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuatomCasB,  InstEmit.SuatomCasB,  InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.Rc  | InstProps.SPd);
             Add("1110101x1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuatomCas,   InstEmit.SuatomCas,   InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.SPd);
-            Add("111010110001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuldDB,      InstEmit.SuldDB,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc  | InstProps.SPd | InstProps.TexB);
+            Add("1110101100010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuldDB,      InstEmit.SuldDB,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc  | InstProps.SPd | InstProps.TexB);
             Add("1110101100011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuldD,       InstEmit.SuldD,       InstProps.Rd  | InstProps.Ra  | InstProps.SPd | InstProps.Tex);
-            Add("11101011000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuldB,       InstEmit.SuldB,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc  | InstProps.SPd | InstProps.TexB);
-            Add("11101011000x1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Suld,        InstEmit.Suld,        InstProps.Rd  | InstProps.Ra  | InstProps.SPd | InstProps.Tex);
-            Add("111010110101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuredB,      InstEmit.SuredB,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
+            Add("1110101100000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuldB,       InstEmit.SuldB,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc  | InstProps.SPd | InstProps.TexB);
+            Add("1110101100001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Suld,        InstEmit.Suld,        InstProps.Rd  | InstProps.Ra  | InstProps.SPd | InstProps.Tex);
+            Add("1110101101010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SuredB,      InstEmit.SuredB,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc);
             Add("1110101101011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Sured,       InstEmit.Sured,       InstProps.Rd  | InstProps.Ra);
-            Add("111010110011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SustDB,      InstEmit.SustDB,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc  | InstProps.TexB);
+            Add("1110101100110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SustDB,      InstEmit.SustDB,      InstProps.Rd  | InstProps.Ra  | InstProps.Rc  | InstProps.TexB);
             Add("1110101100111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SustD,       InstEmit.SustD,       InstProps.Rd  | InstProps.Ra  | InstProps.Tex);
-            Add("11101011001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SustB,       InstEmit.SustB,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc  | InstProps.TexB);
-            Add("11101011001x1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Sust,        InstEmit.Sust,        InstProps.Rd  | InstProps.Ra  | InstProps.Tex);
+            Add("1110101100100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.SustB,       InstEmit.SustB,       InstProps.Rd  | InstProps.Ra  | InstProps.Rc  | InstProps.TexB);
+            Add("1110101100101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Sust,        InstEmit.Sust,        InstProps.Rd  | InstProps.Ra  | InstProps.Tex);
             Add("1111000011111xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Sync,        InstEmit.Sync,        InstProps.Bra);
             Add("11000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Tex,         InstEmit.Tex,         InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.TPd | InstProps.Tex);
             Add("1101111010xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.TexB,        InstEmit.TexB,        InstProps.Rd  | InstProps.Ra  | InstProps.Rb  | InstProps.TPd | InstProps.TexB);

Ryujinx.Graphics.Shader/Instructions/InstEmitAttribute.cs

@@ -46,7 +46,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
                     context.FlagAttributeRead(offset);
 
-                    if (op.O)
+                    if (op.O && CanLoadOutput(offset))
                     {
                         offset |= AttributeConsts.LoadOutputMask;
                     }
@@ -61,7 +61,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
                     context.FlagAttributeRead(offset);
 
-                    if (op.O)
+                    if (op.O && CanLoadOutput(offset))
                     {
                         offset |= AttributeConsts.LoadOutputMask;
                     }
@@ -241,6 +241,11 @@ namespace Ryujinx.Graphics.Shader.Instructions
             }
         }
 
+        private static bool CanLoadOutput(int attr)
+        {
+            return attr != AttributeConsts.TessCoordX && attr != AttributeConsts.TessCoordY;
+        }
+
         private static bool TryFixedFuncToUserAttributeIpa(EmitterContext context, int attr, out Operand selectedAttr)
         {
             if (attr >= AttributeConsts.FrontColorDiffuseR && attr < AttributeConsts.BackColorDiffuseR)

Ryujinx.Graphics.Shader/Instructions/InstEmitFlowControl.cs

@@ -41,20 +41,77 @@ namespace Ryujinx.Graphics.Shader.Instructions
 
             Operand address = context.IAdd(Register(op.SrcA, RegisterType.Gpr), Const(offset));
 
-            // Sorting the target addresses in descending order improves the code,
-            // since it will always check the most distant targets first, then the
-            // near ones. This can be easily transformed into if/else statements.
-            var sortedTargets = context.CurrBlock.Successors.Skip(startIndex).OrderByDescending(x => x.Address);
+            var targets = context.CurrBlock.Successors.Skip(startIndex);
 
-            Block lastTarget = sortedTargets.LastOrDefault();
+            bool allTargetsSinglePred = true;
+            int total = context.CurrBlock.Successors.Count - startIndex;
+            int count = 0;
 
-            foreach (Block possibleTarget in sortedTargets)
+            foreach (var target in targets.OrderBy(x => x.Address))
             {
-                Operand label = context.GetLabel(possibleTarget.Address);
-
-                if (possibleTarget != lastTarget)
+                if (++count < total && (target.Predecessors.Count > 1 || target.Address <= context.CurrBlock.Address))
                 {
-                    context.BranchIfTrue(label, context.ICompareEqual(address, Const((int)possibleTarget.Address)));
+                    allTargetsSinglePred = false;
+                    break;
+                }
+            }
+
+            if (allTargetsSinglePred)
+            {
+                // Chain blocks, each target block will check if the BRX target address
+                // matches its own address, if not, it jumps to the next target which will do the same check,
+                // until it reaches the last possible target, which executed unconditionally.
+                // We can only do this if the BRX block is the only predecessor of all target blocks.
+                // Additionally, this is not supported for blocks located before the current block,
+                // since it will be too late to insert a label, but this is something that can be improved
+                // in the future if necessary.
+
+                var sortedTargets = targets.OrderBy(x => x.Address);
+
+                Block currentTarget = null;
+                ulong firstTargetAddress = 0;
+
+                foreach (Block nextTarget in sortedTargets)
+                {
+                    if (currentTarget != null)
+                    {
+                        if (currentTarget.Address != nextTarget.Address)
+                        {
+                            context.SetBrxTarget(currentTarget.Address, address, (int)currentTarget.Address, nextTarget.Address);
+                        }
+                    }
+                    else
+                    {
+                        firstTargetAddress = nextTarget.Address;
+                    }
+
+                    currentTarget = nextTarget;
+                }
+
+                context.Branch(context.GetLabel(firstTargetAddress));
+            }
+            else
+            {
+                // Emit the branches sequentially.
+                // This generates slightly worse code, but should work for all cases.
+
+                var sortedTargets = targets.OrderByDescending(x => x.Address);
+                ulong lastTargetAddress = ulong.MaxValue;
+
+                count = 0;
+
+                foreach (Block target in sortedTargets)
+                {
+                    Operand label = context.GetLabel(target.Address);
+
+                    if (++count < total)
+                    {
+                        if (target.Address != lastTargetAddress)
+                        {
+                            context.BranchIfTrue(label, context.ICompareEqual(address, Const((int)target.Address)));
+                        }
+
+                        lastTargetAddress = target.Address;
                     }
                     else
                     {
@@ -62,6 +119,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
                     }
                 }
             }
+        }
 
         public static void Cal(EmitterContext context)
         {

Ryujinx.Graphics.Shader/Instructions/InstEmitSurface.cs

@@ -123,7 +123,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
         {
             InstSuldB op = context.GetOp<InstSuldB>();
 
-            EmitSuld(context, op.CacheOp, op.Dim, 0, 0, op.Rgba, op.SrcA, op.Dest, 0, useComponents: true, false, isBindless: true);
+            EmitSuld(context, op.CacheOp, op.Dim, 0, 0, op.Rgba, op.SrcA, op.Dest, op.SrcC, useComponents: true, false, isBindless: true);
         }
 
         public static void Suld(EmitterContext context)

Ryujinx.Graphics.Shader/StructuredIr/StructuredProgram.cs

@@ -97,9 +97,17 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
                 if (src1.Type == OperandType.Constant && src2.Type == OperandType.Constant)
                 {
                     int attrOffset = (src1.Value & AttributeConsts.Mask) + (src2.Value << 2);
+
+                    if ((src1.Value & AttributeConsts.LoadOutputMask) != 0)
+                    {
+                        context.Info.Outputs.Add(attrOffset);
+                    }
+                    else
+                    {
                         context.Info.Inputs.Add(attrOffset);
                     }
                 }
+            }
 
             int sourcesCount = operation.SourcesCount;
             int outDestsCount = operation.DestsCount != 0 ? operation.DestsCount - 1 : 0;

Ryujinx.Graphics.Shader/Translation/AttributeConsts.cs

@@ -54,6 +54,9 @@ namespace Ryujinx.Graphics.Shader.Translation
         public const int UserAttributeBase   = 0x80;
         public const int UserAttributeEnd    = UserAttributeBase + UserAttributesCount * 16;
 
+        public const int UserAttributePerPatchBase = 0x18;
+        public const int UserAttributePerPatchEnd  = 0x200;
+
         public const int LoadOutputMask = 1 << 30;
         public const int Mask = 0x3fffffff;
 

Ryujinx.Graphics.Shader/Translation/AttributeInfo.cs

@@ -4,14 +4,8 @@ namespace Ryujinx.Graphics.Shader.Translation
 {
     struct AttributeInfo
     {
-        private static readonly Dictionary<int, AttributeInfo> BuiltInAttributes = new Dictionary<int, AttributeInfo>()
+        private static readonly Dictionary<int, AttributeInfo> _builtInAttributes = new Dictionary<int, AttributeInfo>()
         {
-            { AttributeConsts.TessLevelOuter0, new AttributeInfo(AttributeConsts.TessLevelOuter0, 0, 4, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.TessLevelOuter1, new AttributeInfo(AttributeConsts.TessLevelOuter0, 1, 4, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.TessLevelOuter2, new AttributeInfo(AttributeConsts.TessLevelOuter0, 2, 4, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.TessLevelOuter3, new AttributeInfo(AttributeConsts.TessLevelOuter0, 3, 4, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.TessLevelInner0, new AttributeInfo(AttributeConsts.TessLevelInner0, 0, 2, AggregateType.Array  | AggregateType.FP32) },
-            { AttributeConsts.TessLevelInner1, new AttributeInfo(AttributeConsts.TessLevelInner0, 1, 2, AggregateType.Array  | AggregateType.FP32) },
             { AttributeConsts.Layer,         new AttributeInfo(AttributeConsts.Layer,         0, 1, AggregateType.S32) },
             { AttributeConsts.ViewportIndex, new AttributeInfo(AttributeConsts.ViewportIndex, 0, 1, AggregateType.S32) },
             { AttributeConsts.PointSize,     new AttributeInfo(AttributeConsts.PointSize,     0, 1, AggregateType.FP32) },
@@ -29,8 +23,8 @@ namespace Ryujinx.Graphics.Shader.Translation
             { AttributeConsts.ClipDistance7, new AttributeInfo(AttributeConsts.ClipDistance0, 7, 8, AggregateType.Array  | AggregateType.FP32) },
             { AttributeConsts.PointCoordX,   new AttributeInfo(AttributeConsts.PointCoordX,   0, 2, AggregateType.Vector | AggregateType.FP32) },
             { AttributeConsts.PointCoordY,   new AttributeInfo(AttributeConsts.PointCoordX,   1, 2, AggregateType.Vector | AggregateType.FP32) },
-            { AttributeConsts.TessCoordX,      new AttributeInfo(AttributeConsts.TessCoordX,      0, 2, AggregateType.Vector | AggregateType.FP32) },
-            { AttributeConsts.TessCoordY,      new AttributeInfo(AttributeConsts.TessCoordX,      1, 2, AggregateType.Vector | AggregateType.FP32) },
+            { AttributeConsts.TessCoordX,    new AttributeInfo(AttributeConsts.TessCoordX,    0, 3, AggregateType.Vector | AggregateType.FP32) },
+            { AttributeConsts.TessCoordY,    new AttributeInfo(AttributeConsts.TessCoordX,    1, 3, AggregateType.Vector | AggregateType.FP32) },
             { AttributeConsts.InstanceId,    new AttributeInfo(AttributeConsts.InstanceId,    0, 1, AggregateType.S32) },
             { AttributeConsts.VertexId,      new AttributeInfo(AttributeConsts.VertexId,      0, 1, AggregateType.S32) },
             { AttributeConsts.FrontFacing,   new AttributeInfo(AttributeConsts.FrontFacing,   0, 1, AggregateType.Bool) },
@@ -55,6 +49,16 @@ namespace Ryujinx.Graphics.Shader.Translation
             { AttributeConsts.LtMask,              new AttributeInfo(AttributeConsts.LtMask,              0, 4, AggregateType.Vector | AggregateType.U32) },
         };
 
+        private static readonly Dictionary<int, AttributeInfo> _builtInAttributesPerPatch = new Dictionary<int, AttributeInfo>()
+        {
+            { AttributeConsts.TessLevelOuter0, new AttributeInfo(AttributeConsts.TessLevelOuter0, 0, 4, AggregateType.Array | AggregateType.FP32) },
+            { AttributeConsts.TessLevelOuter1, new AttributeInfo(AttributeConsts.TessLevelOuter0, 1, 4, AggregateType.Array | AggregateType.FP32) },
+            { AttributeConsts.TessLevelOuter2, new AttributeInfo(AttributeConsts.TessLevelOuter0, 2, 4, AggregateType.Array | AggregateType.FP32) },
+            { AttributeConsts.TessLevelOuter3, new AttributeInfo(AttributeConsts.TessLevelOuter0, 3, 4, AggregateType.Array | AggregateType.FP32) },
+            { AttributeConsts.TessLevelInner0, new AttributeInfo(AttributeConsts.TessLevelInner0, 0, 2, AggregateType.Array | AggregateType.FP32) },
+            { AttributeConsts.TessLevelInner1, new AttributeInfo(AttributeConsts.TessLevelInner0, 1, 2, AggregateType.Array | AggregateType.FP32) },
+        };
+
         public int BaseValue { get; }
         public int Value { get; }
         public int Length { get; }
@@ -76,6 +80,11 @@ namespace Ryujinx.Graphics.Shader.Translation
             return (Value - BaseValue) / 4;
         }
 
+        public static bool Validate(ShaderConfig config, int value, bool isOutAttr, bool perPatch)
+        {
+            return perPatch ? ValidatePerPatch(config, value, isOutAttr) : Validate(config, value, isOutAttr);
+        }
+
         public static bool Validate(ShaderConfig config, int value, bool isOutAttr)
         {
             if (value == AttributeConsts.ViewportIndex && !config.GpuAccessor.QueryHostSupportsViewportIndex())
@@ -86,6 +95,11 @@ namespace Ryujinx.Graphics.Shader.Translation
             return From(config, value, isOutAttr).IsValid;
         }
 
+        public static bool ValidatePerPatch(ShaderConfig config, int value, bool isOutAttr)
+        {
+            return FromPatch(config, value, isOutAttr).IsValid;
+        }
+
         public static AttributeInfo From(ShaderConfig config, int value, bool isOutAttr)
         {
             value &= ~3;
@@ -115,7 +129,24 @@ namespace Ryujinx.Graphics.Shader.Translation
             {
                 return new AttributeInfo(value, 0, 1, AggregateType.FP32);
             }
-            else if (BuiltInAttributes.TryGetValue(value, out AttributeInfo info))
+            else if (_builtInAttributes.TryGetValue(value, out AttributeInfo info))
+            {
+                return info;
+            }
+
+            return new AttributeInfo(value, 0, 0, AggregateType.Invalid);
+        }
+
+        public static AttributeInfo FromPatch(ShaderConfig config, int value, bool isOutAttr)
+        {
+            value &= ~3;
+
+            if (value >= AttributeConsts.UserAttributePerPatchBase && value < AttributeConsts.UserAttributePerPatchEnd)
+            {
+                int offset = (value - AttributeConsts.UserAttributePerPatchBase) & 0xf;
+                return new AttributeInfo(value - offset, offset >> 2, 4, AggregateType.Vector | AggregateType.FP32, false);
+            }
+            else if (_builtInAttributesPerPatch.TryGetValue(value, out AttributeInfo info))
             {
                 return info;
             }

Ryujinx.Graphics.Shader/Translation/EmitterContext.cs

@@ -21,8 +21,33 @@ namespace Ryujinx.Graphics.Shader.Translation
 
         public int OperationsCount => _operations.Count;
 
+        private struct BrxTarget
+        {
+            public readonly Operand Selector;
+            public readonly int ExpectedValue;
+            public readonly ulong NextTargetAddress;
+
+            public BrxTarget(Operand selector, int expectedValue, ulong nextTargetAddress)
+            {
+                Selector = selector;
+                ExpectedValue = expectedValue;
+                NextTargetAddress = nextTargetAddress;
+            }
+        }
+
+        private class BlockLabel
+        {
+            public readonly Operand Label;
+            public BrxTarget BrxTarget;
+
+            public BlockLabel(Operand label)
+            {
+                Label = label;
+            }
+        }
+
         private readonly List<Operation> _operations;
-        private readonly Dictionary<ulong, Operand> _labels;
+        private readonly Dictionary<ulong, BlockLabel> _labels;
 
         public EmitterContext(DecodedProgram program, ShaderConfig config, bool isNonMain)
         {
@@ -30,7 +55,7 @@ namespace Ryujinx.Graphics.Shader.Translation
             Config = config;
             IsNonMain = isNonMain;
             _operations = new List<Operation>();
-            _labels = new Dictionary<ulong, Operand>();
+            _labels = new Dictionary<ulong, BlockLabel>();
 
             EmitStart();
         }
@@ -158,14 +183,40 @@ namespace Ryujinx.Graphics.Shader.Translation
 
         public Operand GetLabel(ulong address)
         {
-            if (!_labels.TryGetValue(address, out Operand label))
-            {
-                label = Label();
-
-                _labels.Add(address, label);
+            return EnsureBlockLabel(address).Label;
         }
 
-            return label;
+        public void SetBrxTarget(ulong address, Operand selector, int targetValue, ulong nextTargetAddress)
+        {
+            BlockLabel blockLabel = EnsureBlockLabel(address);
+            Debug.Assert(blockLabel.BrxTarget.Selector == null);
+            blockLabel.BrxTarget = new BrxTarget(selector, targetValue, nextTargetAddress);
+        }
+
+        public void EnterBlock(ulong address)
+        {
+            BlockLabel blockLabel = EnsureBlockLabel(address);
+
+            MarkLabel(blockLabel.Label);
+
+            BrxTarget brxTarget = blockLabel.BrxTarget;
+
+            if (brxTarget.Selector != null)
+            {
+                this.BranchIfFalse(GetLabel(brxTarget.NextTargetAddress), this.ICompareEqual(brxTarget.Selector, Const(brxTarget.ExpectedValue)));
+            }
+        }
+
+        private BlockLabel EnsureBlockLabel(ulong address)
+        {
+            if (!_labels.TryGetValue(address, out BlockLabel blockLabel))
+            {
+                blockLabel = new BlockLabel(Label());
+
+                _labels.Add(address, blockLabel);
+            }
+
+            return blockLabel;
         }
 
         public void PrepareForVertexReturn()
@@ -261,7 +312,7 @@ namespace Ryujinx.Graphics.Shader.Translation
                         {
                             int index = BitOperations.TrailingZeroCount(passthroughAttributes);
                             WriteOutput(AttributeConsts.UserAttributeBase + index * 16, primIndex);
-                            Config.SetOutputUserAttribute(index, perPatch: false);
+                            Config.SetOutputUserAttribute(index);
                             passthroughAttributes &= ~(1 << index);
                         }
 
@@ -364,7 +415,7 @@ namespace Ryujinx.Graphics.Shader.Translation
                     bool targetEnabled = (Config.OmapTargets & (0xf << (rtIndex * 4))) != 0;
                     if (targetEnabled)
                     {
-                        Config.SetOutputUserAttribute(rtIndex, perPatch: false);
+                        Config.SetOutputUserAttribute(rtIndex);
                         regIndexBase += 4;
                     }
                 }

Ryujinx.Graphics.Shader/Translation/ShaderConfig.cs

@@ -50,16 +50,16 @@ namespace Ryujinx.Graphics.Shader.Translation
         public bool NextUsesFixedFuncAttributes { get; private set; }
         public int UsedInputAttributes { get; private set; }
         public int UsedOutputAttributes { get; private set; }
-        public int UsedInputAttributesPerPatch { get; private set; }
-        public int UsedOutputAttributesPerPatch { get; private set; }
+        public HashSet<int> UsedInputAttributesPerPatch { get; }
+        public HashSet<int> UsedOutputAttributesPerPatch { get; }
+        public HashSet<int> NextUsedInputAttributesPerPatch { get; private set; }
         public int PassthroughAttributes { get; private set; }
         private int _nextUsedInputAttributes;
         private int _thisUsedInputAttributes;
+        private Dictionary<int, int> _perPatchAttributeLocations;
 
         public UInt128 NextInputAttributesComponents { get; private set; }
         public UInt128 ThisInputAttributesComponents { get; private set; }
-        public UInt128 NextInputAttributesPerPatchComponents { get; private set; }
-        public UInt128 ThisInputAttributesPerPatchComponents { get; private set; }
 
         private int _usedConstantBuffers;
         private int _usedStorageBuffers;
@@ -122,6 +122,10 @@ namespace Ryujinx.Graphics.Shader.Translation
             Stage       = ShaderStage.Compute;
             GpuAccessor = gpuAccessor;
             Options     = options;
+
+            UsedInputAttributesPerPatch  = new HashSet<int>();
+            UsedOutputAttributesPerPatch = new HashSet<int>();
+
             _usedTextures = new Dictionary<TextureInfo, TextureMeta>();
             _usedImages   = new Dictionary<TextureInfo, TextureMeta>();
         }
@@ -244,14 +248,7 @@ namespace Ryujinx.Graphics.Shader.Translation
             UsedOutputAttributes |= 1 << index;
         }
 
-        public void SetInputUserAttribute(int index, int component, bool perPatch)
-        {
-            if (perPatch)
-            {
-                UsedInputAttributesPerPatch |= 1 << index;
-                ThisInputAttributesPerPatchComponents |= UInt128.Pow2(index * 4 + component);
-            }
-            else
+        public void SetInputUserAttribute(int index, int component)
         {
             int mask = 1 << index;
 
@@ -259,34 +256,63 @@ namespace Ryujinx.Graphics.Shader.Translation
             _thisUsedInputAttributes |= mask;
             ThisInputAttributesComponents |= UInt128.Pow2(index * 4 + component);
         }
+
+        public void SetInputUserAttributePerPatch(int index)
+        {
+            UsedInputAttributesPerPatch.Add(index);
         }
 
-        public void SetOutputUserAttribute(int index, bool perPatch)
-        {
-            if (perPatch)
-            {
-                UsedOutputAttributesPerPatch |= 1 << index;
-            }
-            else
+        public void SetOutputUserAttribute(int index)
         {
             UsedOutputAttributes |= 1 << index;
         }
+
+        public void SetOutputUserAttributePerPatch(int index)
+        {
+            UsedOutputAttributesPerPatch.Add(index);
         }
 
         public void MergeFromtNextStage(ShaderConfig config)
         {
             NextInputAttributesComponents = config.ThisInputAttributesComponents;
-            NextInputAttributesPerPatchComponents = config.ThisInputAttributesPerPatchComponents;
+            NextUsedInputAttributesPerPatch = config.UsedInputAttributesPerPatch;
             NextUsesFixedFuncAttributes = config.UsedFeatures.HasFlag(FeatureFlags.FixedFuncAttr);
             MergeOutputUserAttributes(config.UsedInputAttributes, config.UsedInputAttributesPerPatch);
 
+            if (UsedOutputAttributesPerPatch.Count != 0)
+            {
+                // Regular and per-patch input/output locations can't overlap,
+                // so we must assign on our location using unused regular input/output locations.
+
+                Dictionary<int, int> locationsMap = new Dictionary<int, int>();
+
+                int freeMask = ~UsedOutputAttributes;
+
+                foreach (int attr in UsedOutputAttributesPerPatch)
+                {
+                    int location = BitOperations.TrailingZeroCount(freeMask);
+                    if (location == 32)
+                    {
+                        config.GpuAccessor.Log($"No enough free locations for patch input/output 0x{attr:X}.");
+                        break;
+                    }
+
+                    locationsMap.Add(attr, location);
+                    freeMask &= ~(1 << location);
+                }
+
+                // Both stages must agree on the locations, so use the same "map" for both.
+                _perPatchAttributeLocations = locationsMap;
+                config._perPatchAttributeLocations = locationsMap;
+            }
+
             if (config.Stage != ShaderStage.Fragment)
             {
                 LastInVertexPipeline = false;
             }
         }
 
-        public void MergeOutputUserAttributes(int mask, int maskPerPatch)
+        public void MergeOutputUserAttributes(int mask, IEnumerable<int> perPatch)
         {
             _nextUsedInputAttributes = mask;
 
@@ -297,10 +323,20 @@ namespace Ryujinx.Graphics.Shader.Translation
             else
             {
                 UsedOutputAttributes |= mask;
-                UsedOutputAttributesPerPatch |= maskPerPatch;
+                UsedOutputAttributesPerPatch.UnionWith(perPatch);
             }
         }
 
+        public int GetPerPatchAttributeLocation(int index)
+        {
+            if (_perPatchAttributeLocations == null || !_perPatchAttributeLocations.TryGetValue(index, out int location))
+            {
+                return index;
+            }
+
+            return location;
+        }
+
         public bool IsUsedOutputAttribute(int attr)
         {
             // The check for fixed function attributes on the next stage is conservative,

Ryujinx.Graphics.Shader/Translation/Translator.cs

@@ -162,7 +162,7 @@ namespace Ryujinx.Graphics.Shader.Translation
                 {
                     context.CurrBlock = block;
 
-                    context.MarkLabel(context.GetLabel(block.Address));
+                    context.EnterBlock(block.Address);
 
                     EmitOps(context, block);
                 }
@@ -204,14 +204,12 @@ namespace Ryujinx.Graphics.Shader.Translation
                 InitializeOutputComponent(context, AttributeConsts.UserAttributeBase + index * 4, perPatch: false);
             }
 
-            UInt128 usedAttributesPerPatch = context.Config.NextInputAttributesPerPatchComponents;
-            while (usedAttributesPerPatch != UInt128.Zero)
+            if (context.Config.NextUsedInputAttributesPerPatch != null)
             {
-                int index = usedAttributesPerPatch.TrailingZeroCount();
-
-                InitializeOutputComponent(context, AttributeConsts.UserAttributeBase + index * 4, perPatch: true);
-
-                usedAttributesPerPatch &= ~UInt128.Pow2(index);
+                foreach (int vecIndex in context.Config.NextUsedInputAttributesPerPatch.OrderBy(x => x))
+                {
+                    InitializeOutput(context, AttributeConsts.UserAttributePerPatchBase + vecIndex * 16, perPatch: true);
+                }
             }
 
             if (config.NextUsesFixedFuncAttributes)
@@ -236,7 +234,7 @@ namespace Ryujinx.Graphics.Shader.Translation
             for (int c = 0; c < 4; c++)
             {
                 int attrOffset = baseAttr + c * 4;
-                context.Copy(perPatch ? AttributePerPatch(attrOffset) : Attribute(attrOffset), ConstF(c == 3 ? 1f : 0f));
+                InitializeOutputComponent(context, attrOffset, perPatch);
             }
         }
 

Ryujinx.Graphics.Shader/Translation/TranslatorContext.cs

@@ -1,6 +1,7 @@
 using Ryujinx.Graphics.Shader.Decoders;
 using Ryujinx.Graphics.Shader.IntermediateRepresentation;
 using System.Collections.Generic;
+using System.Linq;
 
 using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
 using static Ryujinx.Graphics.Shader.Translation.Translator;
@@ -137,7 +138,7 @@ namespace Ryujinx.Graphics.Shader.Translation
 
             if (other != null)
             {
-                other._config.MergeOutputUserAttributes(_config.UsedOutputAttributes, 0);
+                other._config.MergeOutputUserAttributes(_config.UsedOutputAttributes, Enumerable.Empty<int>());
 
                 FunctionCode[] otherCode = EmitShader(other._program, other._config, initializeOutputs: true, out int aStart);
 

Ryujinx.Graphics.Texture/LayoutConverter.cs

@@ -93,7 +93,7 @@ namespace Ryujinx.Graphics.Texture
             };
         }
 
-        public static Span<byte> ConvertBlockLinearToLinear(
+        public static byte[] ConvertBlockLinearToLinear(
             int width,
             int height,
             int depth,
@@ -119,7 +119,7 @@ namespace Ryujinx.Graphics.Texture
                 blockHeight,
                 bytesPerPixel);
 
-            Span<byte> output = new byte[outSize];
+            byte[] output = new byte[outSize];
 
             int outOffs = 0;
 
@@ -246,7 +246,7 @@ namespace Ryujinx.Graphics.Texture
             return output;
         }
 
-        public static Span<byte> ConvertLinearStridedToLinear(
+        public static byte[] ConvertLinearStridedToLinear(
             int width,
             int height,
             int blockWidth,
@@ -262,14 +262,15 @@ namespace Ryujinx.Graphics.Texture
             int outStride = BitUtils.AlignUp(w * bytesPerPixel, HostStrideAlignment);
             lineSize = Math.Min(lineSize, outStride);
 
-            Span<byte> output = new byte[h * outStride];
+            byte[] output = new byte[h * outStride];
+            Span<byte> outSpan = output;
 
             int outOffs = 0;
             int inOffs = 0;
 
             for (int y = 0; y < h; y++)
             {
-                data.Slice(inOffs, lineSize).CopyTo(output.Slice(outOffs, lineSize));
+                data.Slice(inOffs, lineSize).CopyTo(outSpan.Slice(outOffs, lineSize));
 
                 inOffs += stride;
                 outOffs += outStride;

Ryujinx.Graphics.Vulkan/BufferHolder.cs

@@ -109,12 +109,34 @@ namespace Ryujinx.Graphics.Vulkan
         {
             if (isWrite)
             {
-                _cachedConvertedBuffers.Clear();
+                SignalWrite(0, Size);
             }
 
             return _buffer;
         }
 
+        public Auto<DisposableBuffer> GetBuffer(CommandBuffer commandBuffer, int offset, int size, bool isWrite = false)
+        {
+            if (isWrite)
+            {
+                SignalWrite(offset, size);
+            }
+
+            return _buffer;
+        }
+
+        public void SignalWrite(int offset, int size)
+        {
+            if (offset == 0 && size == Size)
+            {
+                _cachedConvertedBuffers.Clear();
+            }
+            else
+            {
+                _cachedConvertedBuffers.ClearRange(offset, size);
+            }
+        }
+
         public BufferHandle GetHandle()
         {
             var handle = _bufferHandle;
@@ -183,6 +205,8 @@ namespace Ryujinx.Graphics.Vulkan
 
                     data.Slice(0, dataSize).CopyTo(new Span<byte>((void*)(_map + offset), dataSize));
 
+                    SignalWrite(offset, dataSize);
+
                     return;
                 }
             }
@@ -240,7 +264,7 @@ namespace Ryujinx.Graphics.Vulkan
 
             endRenderPass?.Invoke();
 
-            var dstBuffer = GetBuffer(cbs.CommandBuffer, true).Get(cbs, dstOffset, data.Length).Value;
+            var dstBuffer = GetBuffer(cbs.CommandBuffer, dstOffset, data.Length, true).Get(cbs, dstOffset, data.Length).Value;
 
             InsertBufferBarrier(
                 _gd,
@@ -362,16 +386,36 @@ namespace Ryujinx.Graphics.Vulkan
             _waitable.WaitForFences(_gd.Api, _device, offset, size);
         }
 
+        private bool BoundToRange(int offset, ref int size)
+        {
+            if (offset >= Size)
+            {
+                return false;
+            }
+
+            size = Math.Min(Size - offset, size);
+
+            return true;
+        }
+
         public Auto<DisposableBuffer> GetBufferI8ToI16(CommandBufferScoped cbs, int offset, int size)
         {
-            var key = new I8ToI16CacheKey();
+            if (!BoundToRange(offset, ref size))
+            {
+                return null;
+            }
+
+            var key = new I8ToI16CacheKey(_gd);
 
             if (!_cachedConvertedBuffers.TryGetValue(offset, size, key, out var holder))
             {
                 holder = _gd.BufferManager.Create(_gd, (size * 2 + 3) & ~3);
 
+                _gd.PipelineInternal.EndRenderPass();
                 _gd.HelperShader.ConvertI8ToI16(_gd, cbs, this, holder, offset, size);
 
+                key.SetBuffer(holder.GetBuffer());
+
                 _cachedConvertedBuffers.Add(offset, size, key, holder);
             }
 
@@ -380,6 +424,11 @@ namespace Ryujinx.Graphics.Vulkan
 
         public Auto<DisposableBuffer> GetAlignedVertexBuffer(CommandBufferScoped cbs, int offset, int size, int stride, int alignment)
         {
+            if (!BoundToRange(offset, ref size))
+            {
+                return null;
+            }
+
             var key = new AlignedVertexBufferCacheKey(_gd, stride, alignment);
 
             if (!_cachedConvertedBuffers.TryGetValue(offset, size, key, out var holder))
@@ -388,6 +437,7 @@ namespace Ryujinx.Graphics.Vulkan
 
                 holder = _gd.BufferManager.Create(_gd, (size / stride) * alignedStride);
 
+                _gd.PipelineInternal.EndRenderPass();
                 _gd.HelperShader.ChangeStride(_gd, cbs, this, holder, offset, size, stride, alignedStride);
 
                 key.SetBuffer(holder.GetBuffer());
@@ -398,6 +448,36 @@ namespace Ryujinx.Graphics.Vulkan
             return holder.GetBuffer();
         }
 
+        public Auto<DisposableBuffer> GetBufferTopologyConversion(CommandBufferScoped cbs, int offset, int size, IndexBufferPattern pattern, int indexSize)
+        {
+            if (!BoundToRange(offset, ref size))
+            {
+                return null;
+            }
+
+            var key = new TopologyConversionCacheKey(_gd, pattern, indexSize);
+
+            if (!_cachedConvertedBuffers.TryGetValue(offset, size, key, out var holder))
+            {
+                // The destination index size is always I32.
+
+                int indexCount = size / indexSize;
+
+                int convertedCount = pattern.GetConvertedCount(indexCount);
+
+                holder = _gd.BufferManager.Create(_gd, convertedCount * 4);
+
+                _gd.PipelineInternal.EndRenderPass();
+                _gd.HelperShader.ConvertIndexBuffer(_gd, cbs, this, holder, pattern, indexSize, offset, indexCount);
+
+                key.SetBuffer(holder.GetBuffer());
+
+                _cachedConvertedBuffers.Add(offset, size, key, holder);
+            }
+
+            return holder.GetBuffer();
+        }
+
         public void Dispose()
         {
             _gd.PipelineInternal?.FlushCommandsIfWeightExceeding(_buffer, (ulong)Size);

Ryujinx.Graphics.Vulkan/BufferManager.cs

@@ -38,6 +38,8 @@ namespace Ryujinx.Graphics.Vulkan
 
         private readonly IdList<BufferHolder> _buffers;
 
+        public int BufferCount { get; private set; }
+
         public StagingBuffer StagingBuffer { get; }
 
         public BufferManager(VulkanRenderer gd, PhysicalDevice physicalDevice, Device device)
@@ -56,6 +58,8 @@ namespace Ryujinx.Graphics.Vulkan
                 return BufferHandle.Null;
             }
 
+            BufferCount++;
+
             ulong handle64 = (uint)_buffers.Add(holder);
 
             return Unsafe.As<ulong, BufferHandle>(ref handle64);
@@ -120,6 +124,16 @@ namespace Ryujinx.Graphics.Vulkan
             return null;
         }
 
+        public Auto<DisposableBuffer> GetBuffer(CommandBuffer commandBuffer, BufferHandle handle, int offset, int size, bool isWrite)
+        {
+            if (TryGetBuffer(handle, out var holder))
+            {
+                return holder.GetBuffer(commandBuffer, offset, size, isWrite);
+            }
+
+            return null;
+        }
+
         public Auto<DisposableBuffer> GetBufferI8ToI16(CommandBufferScoped cbs, BufferHandle handle, int offset, int size)
         {
             if (TryGetBuffer(handle, out var holder))
@@ -140,6 +154,16 @@ namespace Ryujinx.Graphics.Vulkan
             return null;
         }
 
+        public Auto<DisposableBuffer> GetBufferTopologyConversion(CommandBufferScoped cbs, BufferHandle handle, int offset, int size, IndexBufferPattern pattern, int indexSize)
+        {
+            if (TryGetBuffer(handle, out var holder))
+            {
+                return holder.GetBufferTopologyConversion(cbs, offset, size, pattern, indexSize);
+            }
+
+            return null;
+        }
+
         public Auto<DisposableBuffer> GetBuffer(CommandBuffer commandBuffer, BufferHandle handle, bool isWrite, out int size)
         {
             if (TryGetBuffer(handle, out var holder))

Ryujinx.Graphics.Vulkan/BufferState.cs

@@ -1,5 +1,4 @@
-using Silk.NET.Vulkan;
-using System;
+using System;
 
 namespace Ryujinx.Graphics.Vulkan
 {
@@ -9,38 +8,17 @@ namespace Ryujinx.Graphics.Vulkan
 
         private readonly int _offset;
         private readonly int _size;
-        private readonly IndexType _type;
 
         private readonly Auto<DisposableBuffer> _buffer;
 
-        public BufferState(Auto<DisposableBuffer> buffer, int offset, int size, IndexType type)
-        {
-            _buffer = buffer;
-
-            _offset = offset;
-            _size = size;
-            _type = type;
-            buffer?.IncrementReferenceCount();
-        }
-
         public BufferState(Auto<DisposableBuffer> buffer, int offset, int size)
         {
             _buffer = buffer;
-
             _offset = offset;
             _size = size;
-            _type = IndexType.Uint16;
             buffer?.IncrementReferenceCount();
         }
 
-        public void BindIndexBuffer(Vk api, CommandBufferScoped cbs)
-        {
-            if (_buffer != null)
-            {
-                api.CmdBindIndexBuffer(cbs.CommandBuffer, _buffer.Get(cbs, _offset, _size).Value, (ulong)_offset, _type);
-            }
-        }
-
         public void BindTransformFeedbackBuffer(VulkanRenderer gd, CommandBufferScoped cbs, uint binding)
         {
             if (_buffer != null)

Ryujinx.Graphics.Vulkan/BufferUsageBitmap.cs

@@ -24,6 +24,11 @@
 
         public void Add(int cbIndex, int offset, int size)
         {
+            if (size == 0)
+            {
+                return;
+            }
+
             // Some usages can be out of bounds (vertex buffer on amd), so bound if necessary.
             if (offset + size > _size)
             {
@@ -39,6 +44,11 @@
 
         public bool OverlapsWith(int cbIndex, int offset, int size)
         {
+            if (size == 0)
+            {
+                return false;
+            }
+
             int cbBase = cbIndex * _bitsPerCb;
             int start = cbBase + offset / _granularity;
             int end = cbBase + (offset + size - 1) / _granularity;

Ryujinx.Graphics.Vulkan/CacheByRange.cs

@@ -10,14 +10,30 @@ namespace Ryujinx.Graphics.Vulkan
 
     struct I8ToI16CacheKey : ICacheKey
     {
-        public I8ToI16CacheKey() { }
+        // Used to notify the pipeline that bindings have invalidated on dispose.
+        private readonly VulkanRenderer _gd;
+        private Auto<DisposableBuffer> _buffer;
+
+        public I8ToI16CacheKey(VulkanRenderer gd)
+        {
+            _gd = gd;
+            _buffer = null;
+        }
 
         public bool KeyEqual(ICacheKey other)
         {
             return other is I8ToI16CacheKey;
         }
 
-        public void Dispose() { }
+        public void SetBuffer(Auto<DisposableBuffer> buffer)
+        {
+            _buffer = buffer;
+        }
+
+        public void Dispose()
+        {
+            _gd.PipelineInternal.DirtyIndexBuffer(_buffer);
+        }
     }
 
     struct AlignedVertexBufferCacheKey : ICacheKey
@@ -55,6 +71,41 @@ namespace Ryujinx.Graphics.Vulkan
         }
     }
 
+    struct TopologyConversionCacheKey : ICacheKey
+    {
+        private IndexBufferPattern _pattern;
+        private int _indexSize;
+
+        // Used to notify the pipeline that bindings have invalidated on dispose.
+        private readonly VulkanRenderer _gd;
+        private Auto<DisposableBuffer> _buffer;
+
+        public TopologyConversionCacheKey(VulkanRenderer gd, IndexBufferPattern pattern, int indexSize)
+        {
+            _gd = gd;
+            _pattern = pattern;
+            _indexSize = indexSize;
+            _buffer = null;
+        }
+
+        public bool KeyEqual(ICacheKey other)
+        {
+            return other is TopologyConversionCacheKey entry &&
+                entry._pattern == _pattern &&
+                entry._indexSize == _indexSize;
+        }
+
+        public void SetBuffer(Auto<DisposableBuffer> buffer)
+        {
+            _buffer = buffer;
+        }
+
+        public void Dispose()
+        {
+            _gd.PipelineInternal.DirtyIndexBuffer(_buffer);
+        }
+    }
+
     struct CacheByRange<T> where T : IDisposable
     {
         private struct Entry
@@ -114,6 +165,44 @@ namespace Ryujinx.Graphics.Vulkan
             }
         }
 
+        public void ClearRange(int offset, int size)
+        {
+            if (_ranges != null && _ranges.Count > 0)
+            {
+                int end = offset + size;
+
+                List<ulong> toRemove = null;
+
+                foreach (KeyValuePair<ulong, List<Entry>> range in _ranges)
+                {
+                    (int rOffset, int rSize) = UnpackRange(range.Key);
+
+                    int rEnd = rOffset + rSize;
+
+                    if (rEnd > offset && rOffset < end)
+                    {
+                        List<Entry> entries = range.Value;
+
+                        foreach (Entry entry in entries)
+                        {
+                            entry.Key.Dispose();
+                            entry.Value.Dispose();
+                        }
+
+                        (toRemove ??= new List<ulong>()).Add(range.Key);
+                    }
+                }
+
+                if (toRemove != null)
+                {
+                    foreach (ulong range in toRemove)
+                    {
+                        _ranges.Remove(range);
+                    }
+                }
+            }
+        }
+
         private List<Entry> GetEntries(int offset, int size)
         {
             if (_ranges == null)
@@ -138,6 +227,11 @@ namespace Ryujinx.Graphics.Vulkan
             return (uint)offset | ((ulong)size << 32);
         }
 
+        private static (int offset, int size) UnpackRange(ulong range)
+        {
+            return ((int)range, (int)(range >> 32));
+        }
+
         public void Dispose()
         {
             Clear();

Ryujinx.Graphics.Vulkan/EnumConversion.cs

@@ -2,6 +2,7 @@
 using Ryujinx.Graphics.GAL;
 using Ryujinx.Graphics.Shader;
 using Silk.NET.Vulkan;
+using System;
 
 namespace Ryujinx.Graphics.Vulkan
 {
@@ -179,8 +180,8 @@ namespace Ryujinx.Graphics.Vulkan
                 GAL.PrimitiveTopology.TrianglesAdjacency => Silk.NET.Vulkan.PrimitiveTopology.TriangleListWithAdjacency,
                 GAL.PrimitiveTopology.TriangleStripAdjacency => Silk.NET.Vulkan.PrimitiveTopology.TriangleStripWithAdjacency,
                 GAL.PrimitiveTopology.Patches => Silk.NET.Vulkan.PrimitiveTopology.PatchList,
-                GAL.PrimitiveTopology.Quads => Silk.NET.Vulkan.PrimitiveTopology.TriangleFan, // Emulated with triangle fans
-                GAL.PrimitiveTopology.QuadStrip => Silk.NET.Vulkan.PrimitiveTopology.TriangleStrip, // Emulated with triangle strips
+                GAL.PrimitiveTopology.Quads => throw new NotSupportedException("Quad topology is not available in Vulkan."),
+                GAL.PrimitiveTopology.QuadStrip => throw new NotSupportedException("QuadStrip topology is not available in Vulkan."),
                 _ => LogInvalidAndReturn(topology, nameof(GAL.PrimitiveTopology), Silk.NET.Vulkan.PrimitiveTopology.TriangleList)
             };
         }

Ryujinx.Graphics.Vulkan/HelperShader.cs

@@ -4,6 +4,7 @@ using Ryujinx.Graphics.Shader.Translation;
 using Ryujinx.Graphics.Vulkan.Shaders;
 using Silk.NET.Vulkan;
 using System;
+using System.Collections.Generic;
 using VkFormat = Silk.NET.Vulkan.Format;
 
 namespace Ryujinx.Graphics.Vulkan
@@ -399,6 +400,86 @@ namespace Ryujinx.Graphics.Vulkan
                 newSize);
         }
 
+        public unsafe void ConvertIndexBuffer(VulkanRenderer gd,
+            CommandBufferScoped cbs,
+            BufferHolder src,
+            BufferHolder dst,
+            IndexBufferPattern pattern,
+            int indexSize,
+            int srcOffset,
+            int indexCount)
+        {
+            int convertedCount = pattern.GetConvertedCount(indexCount);
+            int outputIndexSize = 4;
+
+            // TODO: Do this with a compute shader?
+            var srcBuffer = src.GetBuffer().Get(cbs, srcOffset, indexCount * indexSize).Value;
+            var dstBuffer = dst.GetBuffer().Get(cbs, 0, convertedCount * outputIndexSize).Value;
+
+            gd.Api.CmdFillBuffer(cbs.CommandBuffer, dstBuffer, 0, Vk.WholeSize, 0);
+
+            var bufferCopy = new List<BufferCopy>();
+            int outputOffset = 0;
+
+            // Try to merge copies of adjacent indices to reduce copy count.
+            int sequenceStart = 0;
+            int sequenceLength = 0;
+
+            foreach (var index in pattern.GetIndexMapping(indexCount))
+            {
+                if (sequenceLength > 0)
+                {
+                    if (index == sequenceStart + sequenceLength && indexSize == outputIndexSize)
+                    {
+                        sequenceLength++;
+                        continue;
+                    }
+
+                    // Commit the copy so far.
+                    bufferCopy.Add(new BufferCopy((ulong)(srcOffset + sequenceStart * indexSize), (ulong)outputOffset, (ulong)(indexSize * sequenceLength)));
+                    outputOffset += outputIndexSize * sequenceLength;
+                }
+
+                sequenceStart = index;
+                sequenceLength = 1;
+            }
+
+            if (sequenceLength > 0)
+            {
+                // Commit final pending copy.
+                bufferCopy.Add(new BufferCopy((ulong)(srcOffset + sequenceStart * indexSize), (ulong)outputOffset, (ulong)(indexSize * sequenceLength)));
+            }
+
+            var bufferCopyArray = bufferCopy.ToArray();
+
+            BufferHolder.InsertBufferBarrier(
+                gd,
+                cbs.CommandBuffer,
+                dstBuffer,
+                BufferHolder.DefaultAccessFlags,
+                AccessFlags.AccessTransferWriteBit,
+                PipelineStageFlags.PipelineStageAllCommandsBit,
+                PipelineStageFlags.PipelineStageTransferBit,
+                0,
+                convertedCount * outputIndexSize);
+
+            fixed (BufferCopy* pBufferCopy = bufferCopyArray)
+            {
+                gd.Api.CmdCopyBuffer(cbs.CommandBuffer, srcBuffer, dstBuffer, (uint)bufferCopyArray.Length, pBufferCopy);
+            }
+
+            BufferHolder.InsertBufferBarrier(
+                gd,
+                cbs.CommandBuffer,
+                dstBuffer,
+                AccessFlags.AccessTransferWriteBit,
+                BufferHolder.DefaultAccessFlags,
+                PipelineStageFlags.PipelineStageTransferBit,
+                PipelineStageFlags.PipelineStageAllCommandsBit,
+                0,
+                convertedCount * outputIndexSize);
+        }
+
         protected virtual void Dispose(bool disposing)
         {
             if (disposing)

Ryujinx.Graphics.Vulkan/IndexBufferPattern.cs

@@ -0,0 +1,139 @@
+using Ryujinx.Graphics.GAL;
+using System;
+using System.Collections.Generic;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Vulkan
+{
+    internal class IndexBufferPattern : IDisposable
+    {
+        public int PrimitiveVertices { get; }
+        public int PrimitiveVerticesOut { get; }
+        public int BaseIndex { get; }
+        public int[] OffsetIndex { get; }
+        public int IndexStride { get; }
+        public bool RepeatStart { get; }
+
+        private VulkanRenderer _gd;
+        private int _currentSize;
+        private BufferHandle _repeatingBuffer;
+
+        public IndexBufferPattern(VulkanRenderer gd,
+            int primitiveVertices,
+            int primitiveVerticesOut,
+            int baseIndex,
+            int[] offsetIndex,
+            int indexStride,
+            bool repeatStart)
+        {
+            PrimitiveVertices = primitiveVertices;
+            PrimitiveVerticesOut = primitiveVerticesOut;
+            BaseIndex = baseIndex;
+            OffsetIndex = offsetIndex;
+            IndexStride = indexStride;
+            RepeatStart = repeatStart;
+
+            _gd = gd;
+        }
+
+        public int GetPrimitiveCount(int vertexCount)
+        {
+            return Math.Max(0, (vertexCount - BaseIndex) / IndexStride);
+        }
+
+        public int GetConvertedCount(int indexCount)
+        {
+            int primitiveCount = GetPrimitiveCount(indexCount);
+            return primitiveCount * OffsetIndex.Length;
+        }
+
+        public IEnumerable<int> GetIndexMapping(int indexCount)
+        {
+            int primitiveCount = GetPrimitiveCount(indexCount);
+            int index = BaseIndex;
+
+            for (int i = 0; i < primitiveCount; i++)
+            {
+                if (RepeatStart)
+                {
+                    // Used for triangle fan
+                    yield return 0;
+                }
+
+                for (int j = RepeatStart ? 1 : 0; j < OffsetIndex.Length; j++)
+                {
+                    yield return index + OffsetIndex[j];
+                }
+
+                index += IndexStride;
+            }
+        }
+
+        public BufferHandle GetRepeatingBuffer(int vertexCount, out int indexCount)
+        {
+            int primitiveCount = GetPrimitiveCount(vertexCount);
+            indexCount = primitiveCount * PrimitiveVerticesOut;
+
+            int expectedSize = primitiveCount * OffsetIndex.Length;
+
+            if (expectedSize <= _currentSize && _repeatingBuffer != BufferHandle.Null)
+            {
+                return _repeatingBuffer;
+            }
+
+            // Expand the repeating pattern to the number of requested primitives.
+            BufferHandle newBuffer = _gd.CreateBuffer(expectedSize * sizeof(int));
+
+            // Copy the old data to the new one.
+            if (_repeatingBuffer != BufferHandle.Null)
+            {
+                _gd.Pipeline.CopyBuffer(_repeatingBuffer, newBuffer, 0, 0, _currentSize * sizeof(int));
+                _gd.DeleteBuffer(_repeatingBuffer);
+            }
+
+            _repeatingBuffer = newBuffer;
+
+            // Add the additional repeats on top.
+            int newPrimitives = primitiveCount;
+            int oldPrimitives = (_currentSize) / OffsetIndex.Length;
+
+            int[] newData;
+
+            newPrimitives -= oldPrimitives;
+            newData = new int[expectedSize - _currentSize];
+
+            int outOffset = 0;
+            int index = oldPrimitives * IndexStride + BaseIndex;
+
+            for (int i = 0; i < newPrimitives; i++)
+            {
+                if (RepeatStart)
+                {
+                    // Used for triangle fan
+                    newData[outOffset++] = 0;
+                }
+
+                for (int j = RepeatStart ? 1 : 0; j < OffsetIndex.Length; j++)
+                {
+                    newData[outOffset++] = index + OffsetIndex[j];
+                }
+
+                index += IndexStride;
+            }
+
+            _gd.SetBufferData(newBuffer, _currentSize * sizeof(int), MemoryMarshal.Cast<int, byte>(newData));
+            _currentSize = expectedSize;
+
+            return newBuffer;
+        }
+
+        public void Dispose()
+        {
+            if (_repeatingBuffer != BufferHandle.Null)
+            {
+                _gd.DeleteBuffer(_repeatingBuffer);
+                _repeatingBuffer = BufferHandle.Null;
+            }
+        }
+    }
+}

Ryujinx.Graphics.Vulkan/IndexBufferState.cs

@@ -0,0 +1,102 @@
+using Silk.NET.Vulkan;
+using System;
+
+namespace Ryujinx.Graphics.Vulkan
+{
+    internal struct IndexBufferState
+    {
+        public static IndexBufferState Null => new IndexBufferState(GAL.BufferHandle.Null, 0, 0);
+
+        private readonly int _offset;
+        private readonly int _size;
+        private readonly IndexType _type;
+
+        private readonly GAL.BufferHandle _handle;
+        private Auto<DisposableBuffer> _buffer;
+
+        public IndexBufferState(GAL.BufferHandle handle, int offset, int size, IndexType type)
+        {
+            _handle = handle;
+            _offset = offset;
+            _size = size;
+            _type = type;
+            _buffer = null;
+        }
+
+        public IndexBufferState(GAL.BufferHandle handle, int offset, int size)
+        {
+            _handle = handle;
+            _offset = offset;
+            _size = size;
+            _type = IndexType.Uint16;
+            _buffer = null;
+        }
+
+        public void BindIndexBuffer(VulkanRenderer gd, CommandBufferScoped cbs)
+        {
+            Auto<DisposableBuffer> autoBuffer;
+            int offset, size;
+            IndexType type = _type;
+
+            if (_type == IndexType.Uint8Ext && !gd.Capabilities.SupportsIndexTypeUint8)
+            {
+                // Index type is not supported. Convert to I16.
+                autoBuffer = gd.BufferManager.GetBufferI8ToI16(cbs, _handle, _offset, _size);
+
+                type = IndexType.Uint16;
+                offset = 0;
+                size = _size * 2;
+            }
+            else
+            {
+                autoBuffer = gd.BufferManager.GetBuffer(cbs.CommandBuffer, _handle, false, out int bufferSize);
+
+                if (_offset >= bufferSize)
+                {
+                    autoBuffer = null;
+                }
+
+                offset = _offset;
+                size = _size;
+            }
+
+            _buffer = autoBuffer;
+
+            if (autoBuffer != null)
+            {
+                gd.Api.CmdBindIndexBuffer(cbs.CommandBuffer, autoBuffer.Get(cbs, offset, size).Value, (ulong)offset, type);
+            }
+        }
+
+        public void BindConvertedIndexBuffer(VulkanRenderer gd, CommandBufferScoped cbs, int firstIndex, int indexCount, int convertedCount, IndexBufferPattern pattern)
+        {
+            Auto<DisposableBuffer> autoBuffer;
+
+            // Convert the index buffer using the given pattern.
+            int indexSize = _type switch
+            {
+                IndexType.Uint32 => 4,
+                IndexType.Uint16 => 2,
+                _ => 1,
+            };
+
+            int firstIndexOffset = firstIndex * indexSize;
+
+            autoBuffer = gd.BufferManager.GetBufferTopologyConversion(cbs, _handle, _offset + firstIndexOffset, indexCount * indexSize, pattern, indexSize);
+
+            int size = convertedCount * 4;
+
+            _buffer = autoBuffer;
+
+            if (autoBuffer != null)
+            {
+                gd.Api.CmdBindIndexBuffer(cbs.CommandBuffer, autoBuffer.Get(cbs, 0, size).Value, 0, IndexType.Uint32);
+            }
+        }
+
+        public bool BoundEquals(Auto<DisposableBuffer> buffer)
+        {
+            return _buffer == buffer;
+        }
+    }
+}

Ryujinx.Graphics.Vulkan/PipelineBase.cs

@@ -21,7 +21,7 @@ namespace Ryujinx.Graphics.Vulkan
 
         protected readonly AutoFlushCounter AutoFlush;
 
-        private PipelineDynamicState _dynamicState;
+        protected PipelineDynamicState DynamicState;
         private PipelineState _newState;
         private bool _stateDirty;
         private GAL.PrimitiveTopology _topology;
@@ -51,13 +51,16 @@ namespace Ryujinx.Graphics.Vulkan
 
         private readonly DescriptorSetUpdater _descriptorSetUpdater;
 
-        private BufferState _indexBuffer;
+        private IndexBufferState _indexBuffer;
+        private IndexBufferPattern _indexBufferPattern;
         private readonly BufferState[] _transformFeedbackBuffers;
         private readonly VertexBufferState[] _vertexBuffers;
         private ulong _vertexBuffersDirty;
         protected Rectangle<int> ClearScissor;
 
         public SupportBufferUpdater SupportBufferUpdater;
+        public IndexBufferPattern QuadsToTrisPattern;
+        public IndexBufferPattern TriFanToTrisPattern;
 
         private bool _needsIndexBufferRebind;
         private bool _needsTransformFeedbackBuffersRebind;
@@ -65,6 +68,8 @@ namespace Ryujinx.Graphics.Vulkan
         private bool _tfEnabled;
         private bool _tfActive;
 
+        private PipelineColorBlendAttachmentState[] _storedBlend;
+
         public ulong DrawCount { get; private set; }
 
         public unsafe PipelineBase(VulkanRenderer gd, Device device)
@@ -90,7 +95,7 @@ namespace Ryujinx.Graphics.Vulkan
 
             using var emptyVb = gd.BufferManager.Create(gd, EmptyVbSize);
             emptyVb.SetData(0, new byte[EmptyVbSize]);
-            _vertexBuffers[0] = new VertexBufferState(emptyVb.GetBuffer(), 0, EmptyVbSize, 0);
+            _vertexBuffers[0] = new VertexBufferState(emptyVb.GetBuffer(), 0, 0, EmptyVbSize, 0);
             _vertexBuffersDirty = ulong.MaxValue >> (64 - _vertexBuffers.Length);
 
             ClearScissor = new Rectangle<int>(0, 0, 0xffff, 0xffff);
@@ -101,12 +106,17 @@ namespace Ryujinx.Graphics.Vulkan
             _newState.Initialize();
             _newState.LineWidth = 1f;
             _newState.SamplesCount = 1;
+
+            _storedBlend = new PipelineColorBlendAttachmentState[8];
         }
 
         public void Initialize()
         {
             SupportBufferUpdater = new SupportBufferUpdater(Gd);
             SupportBufferUpdater.UpdateRenderScale(_renderScale, 0, SupportBuffer.RenderScaleMaxCount);
+
+            QuadsToTrisPattern = new IndexBufferPattern(Gd, 4, 6, 0, new[] { 0, 1, 2, 0, 2, 3 }, 4, false);
+            TriFanToTrisPattern = new IndexBufferPattern(Gd, 3, 3, 2, new[] { int.MinValue, -1, 0 }, 1, true);
         }
 
         public unsafe void Barrier()
@@ -140,7 +150,7 @@ namespace Ryujinx.Graphics.Vulkan
         {
             EndRenderPass();
 
-            var dst = Gd.BufferManager.GetBuffer(CommandBuffer, destination, true).Get(Cbs, offset, size).Value;
+            var dst = Gd.BufferManager.GetBuffer(CommandBuffer, destination, offset, size, true).Get(Cbs, offset, size).Value;
 
             BufferHolder.InsertBufferBarrier(
                 Gd,
@@ -228,8 +238,8 @@ namespace Ryujinx.Graphics.Vulkan
         {
             EndRenderPass();
 
-            var src = Gd.BufferManager.GetBuffer(CommandBuffer, source, false);
-            var dst = Gd.BufferManager.GetBuffer(CommandBuffer, destination, true);
+            var src = Gd.BufferManager.GetBuffer(CommandBuffer, source, srcOffset, size, false);
+            var dst = Gd.BufferManager.GetBuffer(CommandBuffer, destination, dstOffset, size, true);
 
             BufferHolder.Copy(Gd, Cbs, src, dst, srcOffset, dstOffset, size);
         }
@@ -245,6 +255,14 @@ namespace Ryujinx.Graphics.Vulkan
             }
         }
 
+        public void DirtyIndexBuffer(Auto<DisposableBuffer> buffer)
+        {
+            if (_indexBuffer.BoundEquals(buffer))
+            {
+                _needsIndexBufferRebind = true;
+            }
+        }
+
         public void DispatchCompute(int groupsX, int groupsY, int groupsZ)
         {
             if (!_program.IsLinked)
@@ -267,24 +285,59 @@ namespace Ryujinx.Graphics.Vulkan
 
             RecreatePipelineIfNeeded(PipelineBindPoint.Graphics);
             BeginRenderPass();
-            ResumeTransformFeedbackInternal();
             DrawCount++;
 
-            if (_topology == GAL.PrimitiveTopology.Quads)
+            if (Gd.TopologyUnsupported(_topology))
             {
-                int quadsCount = vertexCount / 4;
+                // Temporarily bind a conversion pattern as an index buffer.
+                _needsIndexBufferRebind = true;
 
-                for (int i = 0; i < quadsCount; i++)
+                IndexBufferPattern pattern = _topology switch
                 {
-                    Gd.Api.CmdDraw(CommandBuffer, 4, (uint)instanceCount, (uint)(firstVertex + i * 4), (uint)firstInstance);
-                }
+                    GAL.PrimitiveTopology.Quads => QuadsToTrisPattern,
+                    GAL.PrimitiveTopology.TriangleFan => TriFanToTrisPattern,
+                    _ => throw new NotSupportedException($"Unsupported topology: {_topology}")
+                };
+
+                BufferHandle handle = pattern.GetRepeatingBuffer(vertexCount, out int indexCount);
+                var buffer = Gd.BufferManager.GetBuffer(CommandBuffer, handle, false);
+
+                Gd.Api.CmdBindIndexBuffer(CommandBuffer, buffer.Get(Cbs, 0, indexCount * sizeof(int)).Value, 0, Silk.NET.Vulkan.IndexType.Uint32);
+
+                BeginRenderPass(); // May have been interrupted to set buffer data.
+                ResumeTransformFeedbackInternal();
+
+                Gd.Api.CmdDrawIndexed(CommandBuffer, (uint)indexCount, (uint)instanceCount, 0, firstVertex, (uint)firstInstance);
             }
             else
             {
+                ResumeTransformFeedbackInternal();
+
                 Gd.Api.CmdDraw(CommandBuffer, (uint)vertexCount, (uint)instanceCount, (uint)firstVertex, (uint)firstInstance);
             }
         }
 
+        private void UpdateIndexBufferPattern()
+        {
+            IndexBufferPattern pattern = null;
+
+            if (Gd.TopologyUnsupported(_topology))
+            {
+                pattern = _topology switch
+                {
+                    GAL.PrimitiveTopology.Quads => QuadsToTrisPattern,
+                    GAL.PrimitiveTopology.TriangleFan => TriFanToTrisPattern,
+                    _ => throw new NotSupportedException($"Unsupported topology: {_topology}")
+                };
+            }
+
+            if (_indexBufferPattern != pattern)
+            {
+                _indexBufferPattern = pattern;
+                _needsIndexBufferRebind = true;
+            }
+        }
+
         public void DrawIndexed(int indexCount, int instanceCount, int firstIndex, int firstVertex, int firstInstance)
         {
             if (!_program.IsLinked)
@@ -292,22 +345,34 @@ namespace Ryujinx.Graphics.Vulkan
                 return;
             }
 
+            UpdateIndexBufferPattern();
             RecreatePipelineIfNeeded(PipelineBindPoint.Graphics);
             BeginRenderPass();
-            ResumeTransformFeedbackInternal();
             DrawCount++;
 
-            if (_topology == GAL.PrimitiveTopology.Quads)
+            if (_indexBufferPattern != null)
             {
-                int quadsCount = indexCount / 4;
+                // Convert the index buffer into a supported topology.
+                IndexBufferPattern pattern = _indexBufferPattern;
 
-                for (int i = 0; i < quadsCount; i++)
+                int convertedCount = pattern.GetConvertedCount(indexCount);
+
+                if (_needsIndexBufferRebind)
                 {
-                    Gd.Api.CmdDrawIndexed(CommandBuffer, 4, (uint)instanceCount, (uint)(firstIndex + i * 4), firstVertex, (uint)firstInstance);
+                    _indexBuffer.BindConvertedIndexBuffer(Gd, Cbs, firstIndex, indexCount, convertedCount, pattern);
+
+                    _needsIndexBufferRebind = false;
                 }
+
+                BeginRenderPass(); // May have been interrupted to set buffer data.
+                ResumeTransformFeedbackInternal();
+
+                Gd.Api.CmdDrawIndexed(CommandBuffer, (uint)convertedCount, (uint)instanceCount, 0, firstVertex, (uint)firstInstance);
             }
             else
             {
+                ResumeTransformFeedbackInternal();
+
                 Gd.Api.CmdDrawIndexed(CommandBuffer, (uint)indexCount, (uint)instanceCount, (uint)firstIndex, firstVertex, (uint)firstInstance);
             }
         }
@@ -323,7 +388,7 @@ namespace Ryujinx.Graphics.Vulkan
                 var oldDepthTestEnable = _newState.DepthTestEnable;
                 var oldDepthWriteEnable = _newState.DepthWriteEnable;
                 var oldTopology = _newState.Topology;
-                var oldViewports = _dynamicState.Viewports;
+                var oldViewports = DynamicState.Viewports;
                 var oldViewportsCount = _newState.ViewportsCount;
 
                 _newState.CullMode = CullModeFlags.CullModeNone;
@@ -346,9 +411,9 @@ namespace Ryujinx.Graphics.Vulkan
                 _newState.DepthWriteEnable = oldDepthWriteEnable;
                 _newState.Topology = oldTopology;
 
-                _dynamicState.Viewports = oldViewports;
-                _dynamicState.ViewportsCount = (int)oldViewportsCount;
-                _dynamicState.SetViewportsDirty();
+                DynamicState.Viewports = oldViewports;
+                DynamicState.ViewportsCount = (int)oldViewportsCount;
+                DynamicState.SetViewportsDirty();
 
                 _newState.ViewportsCount = oldViewportsCount;
                 SignalStateChange();
@@ -383,8 +448,13 @@ namespace Ryujinx.Graphics.Vulkan
             ResumeTransformFeedbackInternal();
             DrawCount++;
 
-            var buffer = Gd.BufferManager.GetBuffer(CommandBuffer, indirectBuffer.Handle, true).Get(Cbs, indirectBuffer.Offset, indirectBuffer.Size).Value;
-            var countBuffer = Gd.BufferManager.GetBuffer(CommandBuffer, parameterBuffer.Handle, true).Get(Cbs, parameterBuffer.Offset, parameterBuffer.Size).Value;
+            var buffer = Gd.BufferManager
+                .GetBuffer(CommandBuffer, indirectBuffer.Handle, indirectBuffer.Offset, indirectBuffer.Size, true)
+                .Get(Cbs, indirectBuffer.Offset, indirectBuffer.Size).Value;
+
+            var countBuffer = Gd.BufferManager
+                .GetBuffer(CommandBuffer, parameterBuffer.Handle, parameterBuffer.Offset, parameterBuffer.Size, true)
+                .Get(Cbs, parameterBuffer.Offset, parameterBuffer.Size).Value;
 
             Gd.DrawIndirectCountApi.CmdDrawIndirectCount(
                 CommandBuffer,
@@ -413,8 +483,13 @@ namespace Ryujinx.Graphics.Vulkan
             ResumeTransformFeedbackInternal();
             DrawCount++;
 
-            var buffer = Gd.BufferManager.GetBuffer(CommandBuffer, indirectBuffer.Handle, true).Get(Cbs, indirectBuffer.Offset, indirectBuffer.Size).Value;
-            var countBuffer = Gd.BufferManager.GetBuffer(CommandBuffer, parameterBuffer.Handle, true).Get(Cbs, parameterBuffer.Offset, parameterBuffer.Size).Value;
+            var buffer = Gd.BufferManager
+                .GetBuffer(CommandBuffer, indirectBuffer.Handle, parameterBuffer.Offset, parameterBuffer.Size, true)
+                .Get(Cbs, indirectBuffer.Offset, indirectBuffer.Size).Value;
+
+            var countBuffer = Gd.BufferManager
+                .GetBuffer(CommandBuffer, parameterBuffer.Handle, parameterBuffer.Offset, parameterBuffer.Size, true)
+                .Get(Cbs, parameterBuffer.Offset, parameterBuffer.Size).Value;
 
             Gd.DrawIndirectCountApi.CmdDrawIndexedIndirectCount(
                 CommandBuffer,
@@ -437,6 +512,8 @@ namespace Ryujinx.Graphics.Vulkan
         {
             ref var vkBlend = ref _newState.Internal.ColorBlendAttachmentState[index];
 
+            if (blend.Enable)
+            {
                 vkBlend.BlendEnable = blend.Enable;
                 vkBlend.SrcColorBlendFactor = blend.ColorSrcFactor.Convert();
                 vkBlend.DstColorBlendFactor = blend.ColorDstFactor.Convert();
@@ -444,6 +521,19 @@ namespace Ryujinx.Graphics.Vulkan
                 vkBlend.SrcAlphaBlendFactor = blend.AlphaSrcFactor.Convert();
                 vkBlend.DstAlphaBlendFactor = blend.AlphaDstFactor.Convert();
                 vkBlend.AlphaBlendOp = blend.AlphaOp.Convert();
+            }
+            else
+            {
+                vkBlend = new PipelineColorBlendAttachmentState(
+                    colorWriteMask: vkBlend.ColorWriteMask);
+            }
+
+            if (vkBlend.ColorWriteMask == 0)
+            {
+                _storedBlend[index] = vkBlend;
+
+                vkBlend = new PipelineColorBlendAttachmentState();
+            }
 
             _newState.BlendConstantR = blend.BlendConstant.Red;
             _newState.BlendConstantG = blend.BlendConstant.Green;
@@ -455,7 +545,7 @@ namespace Ryujinx.Graphics.Vulkan
 
         public void SetDepthBias(PolygonModeMask enables, float factor, float units, float clamp)
         {
-            _dynamicState.SetDepthBias(factor, units, clamp);
+            DynamicState.SetDepthBias(factor, units, clamp);
 
             _newState.DepthBiasEnable = enables != 0;
             SignalStateChange();
@@ -500,34 +590,16 @@ namespace Ryujinx.Graphics.Vulkan
 
         public void SetIndexBuffer(BufferRange buffer, GAL.IndexType type)
         {
-            _indexBuffer.Dispose();
-
             if (buffer.Handle != BufferHandle.Null)
             {
-                Auto<DisposableBuffer> ib = null;
-                int offset = buffer.Offset;
-                int size = buffer.Size;
-
-                if (type == GAL.IndexType.UByte && !Gd.Capabilities.SupportsIndexTypeUint8)
-                {
-                    ib = Gd.BufferManager.GetBufferI8ToI16(Cbs, buffer.Handle, offset, size);
-                    offset = 0;
-                    size *= 2;
-                    type = GAL.IndexType.UShort;
+                _indexBuffer = new IndexBufferState(buffer.Handle, buffer.Offset, buffer.Size, type.Convert());
             }
             else
             {
-                    ib = Gd.BufferManager.GetBuffer(CommandBuffer, buffer.Handle, false);
+                _indexBuffer = IndexBufferState.Null;
             }
 
-                _indexBuffer = new BufferState(ib, offset, size, type.Convert());
-            }
-            else
-            {
-                _indexBuffer = BufferState.Null;
-            }
-
-            _indexBuffer.BindIndexBuffer(Gd.Api, Cbs);
+            _needsIndexBufferRebind = true;
         }
 
         public void SetLineParameters(float width, bool smooth)
@@ -584,7 +656,7 @@ namespace Ryujinx.Graphics.Vulkan
         {
             _topology = topology;
 
-            var vkTopology = topology.Convert();
+            var vkTopology = Gd.TopologyRemap(topology).Convert();
 
             _newState.Topology = vkTopology;
 
@@ -626,8 +698,25 @@ namespace Ryujinx.Graphics.Vulkan
             for (int i = 0; i < count; i++)
             {
                 ref var vkBlend = ref _newState.Internal.ColorBlendAttachmentState[i];
+                var newMask = (ColorComponentFlags)componentMask[i];
 
-                vkBlend.ColorWriteMask = (ColorComponentFlags)componentMask[i];
+                // When color write mask is 0, remove all blend state to help the pipeline cache.
+                // Restore it when the mask becomes non-zero.
+                if (vkBlend.ColorWriteMask != newMask)
+                {
+                    if (newMask == 0)
+                    {
+                        _storedBlend[i] = vkBlend;
+
+                        vkBlend = new PipelineColorBlendAttachmentState();
+                    }
+                    else if (vkBlend.ColorWriteMask == 0)
+                    {
+                        vkBlend = _storedBlend[i];
+                    }
+                }
+
+                vkBlend.ColorWriteMask = newMask;
 
                 if (componentMask[i] != 0)
                 {
@@ -674,10 +763,10 @@ namespace Ryujinx.Graphics.Vulkan
                 var offset = new Offset2D(region.X, region.Y);
                 var extent = new Extent2D((uint)region.Width, (uint)region.Height);
 
-                _dynamicState.SetScissor(i, new Rect2D(offset, extent));
+                DynamicState.SetScissor(i, new Rect2D(offset, extent));
             }
 
-            _dynamicState.ScissorsCount = count;
+            DynamicState.ScissorsCount = count;
 
             _newState.ScissorsCount = (uint)count;
             SignalStateChange();
@@ -685,7 +774,7 @@ namespace Ryujinx.Graphics.Vulkan
 
         public void SetStencilTest(StencilTestDescriptor stencilTest)
         {
-            _dynamicState.SetStencilMasks(
+            DynamicState.SetStencilMasks(
                 (uint)stencilTest.BackFuncMask,
                 (uint)stencilTest.BackMask,
                 (uint)stencilTest.BackFuncRef,
@@ -734,7 +823,8 @@ namespace Ryujinx.Graphics.Vulkan
 
                 if (range.Handle != BufferHandle.Null)
                 {
-                    _transformFeedbackBuffers[i] = new BufferState(Gd.BufferManager.GetBuffer(CommandBuffer, range.Handle, true), range.Offset, range.Size);
+                    _transformFeedbackBuffers[i] = 
+                        new BufferState(Gd.BufferManager.GetBuffer(CommandBuffer, range.Handle, range.Offset, range.Size, true), range.Offset, range.Size);
                     _transformFeedbackBuffers[i].BindTransformFeedbackBuffer(Gd, Cbs, (uint)i);
                 }
                 else
@@ -896,7 +986,7 @@ namespace Ryujinx.Graphics.Vulkan
             {
                 var viewport = viewports[i];
 
-                _dynamicState.SetViewport(i, new Silk.NET.Vulkan.Viewport(
+                DynamicState.SetViewport(i, new Silk.NET.Vulkan.Viewport(
                     viewport.Region.X,
                     viewport.Region.Y,
                     viewport.Region.Width == 0f ? 1f : viewport.Region.Width,
@@ -905,7 +995,7 @@ namespace Ryujinx.Graphics.Vulkan
                     Clamp(viewport.DepthFar)));
             }
 
-            _dynamicState.ViewportsCount = count;
+            DynamicState.ViewportsCount = count;
 
             float disableTransformF = disableTransform ? 1.0f : 0.0f;
             if (SupportBufferUpdater.Data.ViewportInverse.W != disableTransformF || disableTransform)
@@ -984,7 +1074,7 @@ namespace Ryujinx.Graphics.Vulkan
             _vertexBuffersDirty = ulong.MaxValue >> (64 - _vertexBuffers.Length);
 
             _descriptorSetUpdater.SignalCommandBufferChange();
-            _dynamicState.ForceAllDirty();
+            DynamicState.ForceAllDirty();
             _currentPipelineHandle = 0;
         }
 
@@ -1122,14 +1212,14 @@ namespace Ryujinx.Graphics.Vulkan
 
         private void RecreatePipelineIfNeeded(PipelineBindPoint pbp)
         {
-            _dynamicState.ReplayIfDirty(Gd.Api, CommandBuffer);
+            DynamicState.ReplayIfDirty(Gd.Api, CommandBuffer);
 
             // Commit changes to the support buffer before drawing.
             SupportBufferUpdater.Commit();
 
-            if (_needsIndexBufferRebind)
+            if (_needsIndexBufferRebind && _indexBufferPattern == null)
             {
-                _indexBuffer.BindIndexBuffer(Gd.Api, Cbs);
+                _indexBuffer.BindIndexBuffer(Gd, Cbs);
                 _needsIndexBufferRebind = false;
             }
 
@@ -1153,7 +1243,7 @@ namespace Ryujinx.Graphics.Vulkan
 
                     _vertexBuffers[i].BindVertexBuffer(Gd, Cbs, (uint)i, ref _newState);
 
-                    _vertexBuffersDirty &= ~(1u << i);
+                    _vertexBuffersDirty &= ~(1UL << i);
                 }
             }
 
@@ -1265,7 +1355,6 @@ namespace Ryujinx.Graphics.Vulkan
             {
                 _renderPass?.Dispose();
                 _framebuffer?.Dispose();
-                _indexBuffer.Dispose();
                 _newState.Dispose();
                 _descriptorSetUpdater.Dispose();
 

Ryujinx.Graphics.Vulkan/PipelineConverter.cs

@@ -199,7 +199,7 @@ namespace Ryujinx.Graphics.Vulkan
 
             pipeline.StencilTestEnable = state.StencilTest.TestEnable;
 
-            pipeline.Topology = state.Topology.Convert();
+            pipeline.Topology = gd.TopologyRemap(state.Topology).Convert();
 
             int vaCount = Math.Min(Constants.MaxVertexAttributes, state.VertexAttribCount);
             int vbCount = Math.Min(Constants.MaxVertexBuffers, state.VertexBufferCount);
@@ -257,6 +257,8 @@ namespace Ryujinx.Graphics.Vulkan
             {
                 var blend = state.BlendDescriptors[i];
 
+                if (blend.Enable && state.ColorWriteMask[i] != 0)
+                {
                     pipeline.Internal.ColorBlendAttachmentState[i] = new PipelineColorBlendAttachmentState(
                         blend.Enable,
                         blend.ColorSrcFactor.Convert(),
@@ -267,6 +269,12 @@ namespace Ryujinx.Graphics.Vulkan
                         blend.AlphaOp.Convert(),
                         (ColorComponentFlags)state.ColorWriteMask[i]);
                 }
+                else
+                {
+                    pipeline.Internal.ColorBlendAttachmentState[i] = new PipelineColorBlendAttachmentState(
+                        colorWriteMask: (ColorComponentFlags)state.ColorWriteMask[i]);
+                }
+            }
 
             int maxAttachmentIndex = 0;
             for (int i = 0; i < 8; i++)

Ryujinx.Graphics.Vulkan/PipelineFull.cs

@@ -204,6 +204,8 @@ namespace Ryujinx.Graphics.Vulkan
             }
 
             SignalCommandBufferChange();
+
+            DynamicState.ReplayIfDirty(Gd.Api, CommandBuffer);
         }
 
         public void FlushCommandsImpl()

Ryujinx.Graphics.Vulkan/SamplerHolder.cs

@@ -64,6 +64,7 @@ namespace Ryujinx.Graphics.Vulkan
                 };
 
                 samplerCreateInfo.PNext = &customBorderColor;
+                samplerCreateInfo.BorderColor = BorderColor.FloatCustomExt;
             }
 
             gd.Api.CreateSampler(device, samplerCreateInfo, null, out var sampler).ThrowOnError();

Ryujinx.Graphics.Vulkan/Shader.cs

@@ -9,17 +9,19 @@ using System.Threading.Tasks;
 
 namespace Ryujinx.Graphics.Vulkan
 {
-    class Shader
+    class Shader : IDisposable
     {
         // The shaderc.net dependency's Options constructor and dispose are not thread safe.
         // Take this lock when using them.
         private static object _shaderOptionsLock = new object();
 
+        private static readonly IntPtr _ptrMainEntryPointName = Marshal.StringToHGlobalAnsi("main");
+
         private readonly Vk _api;
         private readonly Device _device;
         private readonly ShaderStageFlags _stage;
 
-        private IntPtr _entryPointName;
+        private bool _disposed;
         private ShaderModule _module;
 
         public ShaderStageFlags StageFlags => _stage;
@@ -39,7 +41,6 @@ namespace Ryujinx.Graphics.Vulkan
             CompileStatus = ProgramLinkStatus.Incomplete;
 
             _stage = shaderSource.Stage.Convert();
-            _entryPointName = Marshal.StringToHGlobalAnsi("main");
 
             CompileTask = Task.Run(() =>
             {
@@ -145,7 +146,7 @@ namespace Ryujinx.Graphics.Vulkan
                 SType = StructureType.PipelineShaderStageCreateInfo,
                 Stage = _stage,
                 Module = _module,
-                PName = (byte*)_entryPointName
+                PName = (byte*)_ptrMainEntryPointName
             };
         }
 
@@ -156,11 +157,10 @@ namespace Ryujinx.Graphics.Vulkan
 
         public unsafe void Dispose()
         {
-            if (_entryPointName != IntPtr.Zero)
+            if (!_disposed)
             {
                 _api.DestroyShaderModule(_device, _module, null);
-                Marshal.FreeHGlobal(_entryPointName);
-                _entryPointName = IntPtr.Zero;
+                _disposed = true;
             }
         }
     }

Ryujinx.Graphics.Vulkan/StagingBuffer.cs

@@ -87,7 +87,7 @@ namespace Ryujinx.Graphics.Vulkan
         private void PushDataImpl(CommandBufferScoped cbs, BufferHolder dst, int dstOffset, ReadOnlySpan<byte> data)
         {
             var srcBuffer = _buffer.GetBuffer();
-            var dstBuffer = dst.GetBuffer();
+            var dstBuffer = dst.GetBuffer(cbs.CommandBuffer, dstOffset, data.Length, true);
 
             int offset = _freeOffset;
             int capacity = BufferSize - offset;

Ryujinx.Graphics.Vulkan/TextureBuffer.cs

@@ -1,4 +1,5 @@
-using Ryujinx.Graphics.GAL;
+using Ryujinx.Common.Memory;
+using Ryujinx.Graphics.GAL;
 using Silk.NET.Vulkan;
 using System;
 using System.Collections.Generic;
@@ -16,6 +17,8 @@ namespace Ryujinx.Graphics.Vulkan
         private Auto<DisposableBufferView> _bufferView;
         private Dictionary<GAL.Format, Auto<DisposableBufferView>> _selfManagedViews;
 
+        private int _bufferCount;
+
         public int Width { get; }
         public int Height { get; }
 
@@ -88,17 +91,17 @@ namespace Ryujinx.Graphics.Vulkan
             _bufferView = null;
         }
 
-        public void SetData(ReadOnlySpan<byte> data)
+        public void SetData(SpanOrArray<byte> data)
         {
             _gd.SetBufferData(_bufferHandle, _offset, data);
         }
 
-        public void SetData(ReadOnlySpan<byte> data, int layer, int level)
+        public void SetData(SpanOrArray<byte> data, int layer, int level)
         {
             throw new NotSupportedException();
         }
 
-        public void SetData(ReadOnlySpan<byte> data, int layer, int level, Rectangle<int> region)
+        public void SetData(SpanOrArray<byte> data, int layer, int level, Rectangle<int> region)
         {
             throw new NotSupportedException();
         }
@@ -107,7 +110,8 @@ namespace Ryujinx.Graphics.Vulkan
         {
             if (_bufferHandle == buffer.Handle &&
                 _offset == buffer.Offset &&
-                _size == buffer.Size)
+                _size == buffer.Size &&
+                _bufferCount == _gd.BufferManager.BufferCount)
             {
                 return;
             }
@@ -115,6 +119,7 @@ namespace Ryujinx.Graphics.Vulkan
             _bufferHandle = buffer.Handle;
             _offset = buffer.Offset;
             _size = buffer.Size;
+            _bufferCount = _gd.BufferManager.BufferCount;
 
             ReleaseImpl();
         }

Ryujinx.Graphics.Vulkan/TextureView.cs

@@ -1,4 +1,5 @@
-using Ryujinx.Graphics.GAL;
+using Ryujinx.Common.Memory;
+using Ryujinx.Graphics.GAL;
 using Silk.NET.Vulkan;
 using System;
 using System.Collections.Generic;
@@ -873,17 +874,17 @@ namespace Ryujinx.Graphics.Vulkan
             return GetDataFromBuffer(result, size, result);
         }
 
-        public void SetData(ReadOnlySpan<byte> data)
+        public void SetData(SpanOrArray<byte> data)
         {
             SetData(data, 0, 0, Info.GetLayers(), Info.Levels, singleSlice: false);
         }
 
-        public void SetData(ReadOnlySpan<byte> data, int layer, int level)
+        public void SetData(SpanOrArray<byte> data, int layer, int level)
         {
             SetData(data, layer, level, 1, 1, singleSlice: true);
         }
 
-        public void SetData(ReadOnlySpan<byte> data, int layer, int level, Rectangle<int> region)
+        public void SetData(SpanOrArray<byte> data, int layer, int level, Rectangle<int> region)
         {
             SetData(data, layer, level, 1, 1, singleSlice: true, region);
         }

Ryujinx.Graphics.Vulkan/VertexBufferState.cs

@@ -57,9 +57,13 @@ namespace Ryujinx.Graphics.Vulkan
                 if (gd.NeedsVertexBufferAlignment(AttributeScalarAlignment, out int alignment) && (_stride % alignment) != 0)
                 {
                     autoBuffer = gd.BufferManager.GetAlignedVertexBuffer(cbs, _handle, _offset, _size, _stride, alignment);
-                    int stride = (_stride + (alignment - 1)) & -alignment;
 
-                    var buffer = autoBuffer.Get(cbs, _offset, _size).Value;
+                    if (autoBuffer != null)
+                    {
+                        int stride = (_stride + (alignment - 1)) & -alignment;
+                        int newSize = (_size / _stride) * stride;
+
+                        var buffer = autoBuffer.Get(cbs, 0, newSize).Value;
 
                         if (gd.Capabilities.SupportsExtendedDynamicState)
                         {
@@ -69,7 +73,7 @@ namespace Ryujinx.Graphics.Vulkan
                                 1,
                                 buffer,
                                 0,
-                            (ulong)(_size / _stride) * (ulong)stride,
+                                (ulong)newSize,
                                 (ulong)stride);
                         }
                         else
@@ -78,16 +82,23 @@ namespace Ryujinx.Graphics.Vulkan
                         }
 
                         _buffer = autoBuffer;
-                    state.Internal.VertexBindingDescriptions[DescriptorIndex].Stride = (uint)stride;
+                    }
+
+                    state.Internal.VertexBindingDescriptions[DescriptorIndex].Stride = (uint)_stride;
 
                     return;
                 }
                 else
                 {
-                    autoBuffer = gd.BufferManager.GetBuffer(cbs.CommandBuffer, _handle, false, out int _);
+                    autoBuffer = gd.BufferManager.GetBuffer(cbs.CommandBuffer, _handle, false, out int size);
 
                     // The original stride must be reapplied in case it was rewritten.
                     state.Internal.VertexBindingDescriptions[DescriptorIndex].Stride = (uint)_stride;
+
+                    if (_offset >= size)
+                    {
+                        autoBuffer = null;
+                    }
                 }
             }
 

Ryujinx.Graphics.Vulkan/VulkanRenderer.cs

@@ -471,6 +471,25 @@ namespace Ryujinx.Graphics.Vulkan
             Logger.Notice.Print(LogClass.Gpu, $"{GpuVendor} {GpuRenderer} ({GpuVersion})");
         }
 
+        public GAL.PrimitiveTopology TopologyRemap(GAL.PrimitiveTopology topology)
+        {
+            return topology switch
+            {
+                GAL.PrimitiveTopology.Quads => GAL.PrimitiveTopology.Triangles,
+                GAL.PrimitiveTopology.QuadStrip => GAL.PrimitiveTopology.TriangleStrip,
+                _ => topology
+            };
+        }
+
+        public bool TopologyUnsupported(GAL.PrimitiveTopology topology)
+        {
+            return topology switch
+            {
+                GAL.PrimitiveTopology.Quads => true,
+                _ => false
+            };
+        }
+
         public void Initialize(GraphicsDebugLevel logLevel)
         {
             SetupContext(logLevel);

Ryujinx.HLE/HOS/Services/Account/Acc/AccountManager.cs

@@ -22,7 +22,8 @@ namespace Ryujinx.HLE.HOS.Services.Account.Acc
         // outside of the AccountManager.
         private readonly HorizonClient _horizonClient;
 
-        private ConcurrentDictionary<string, UserProfile> _profiles;
+        private readonly ConcurrentDictionary<string, UserProfile> _profiles;
+        private UserProfile[] _storedOpenedUsers;
 
         public UserProfile LastOpenedUser { get; private set; }
 
@@ -31,6 +32,7 @@ namespace Ryujinx.HLE.HOS.Services.Account.Acc
             _horizonClient = horizonClient;
 
             _profiles = new ConcurrentDictionary<string, UserProfile>();
+            _storedOpenedUsers = Array.Empty<UserProfile>();
 
             _accountSaveDataManager = new AccountSaveDataManager(_profiles);
 
@@ -221,6 +223,16 @@ namespace Ryujinx.HLE.HOS.Services.Account.Acc
             return _profiles.Values.Where(x => x.AccountState == AccountState.Open);
         }
 
+        internal IEnumerable<UserProfile> GetStoredOpenedUsers()
+        {
+            return _storedOpenedUsers;
+        }
+
+        internal void StoreOpenedUsers()
+        {
+            _storedOpenedUsers = _profiles.Values.Where(x => x.AccountState == AccountState.Open).ToArray();
+        }
+
         internal UserProfile GetFirst()
         {
             return _profiles.First().Value;

Ryujinx.HLE/HOS/Services/Account/Acc/AccountService/ManagerServer.cs

@@ -162,7 +162,7 @@ namespace Ryujinx.HLE.HOS.Services.Account.Acc.AccountService
 
         public ResultCode StoreOpenContext(ServiceCtx context)
         {
-            Logger.Stub?.PrintStub(LogClass.ServiceAcc);
+            context.Device.System.AccountManager.StoreOpenedUsers();
 
             return ResultCode.Success;
         }

Ryujinx.HLE/HOS/Services/Account/Acc/ApplicationServiceServer.cs

@@ -201,6 +201,11 @@ namespace Ryujinx.HLE.HOS.Services.Account.Acc
             return ResultCode.Success;
         }
 
+        public ResultCode ListOpenContextStoredUsers(ServiceCtx context)
+        {
+            return WriteUserList(context, context.Device.System.AccountManager.GetStoredOpenedUsers());
+        }
+
         public ResultCode ListQualifiedUsers(ServiceCtx context)
         {
             // TODO: Determine how users are "qualified". We assume all users are "qualified" for now.

Ryujinx.HLE/HOS/Services/Account/Acc/IAccountServiceForApplication.cs

@@ -1,5 +1,4 @@
 using Ryujinx.Common.Logging;
-using Ryujinx.Cpu;
 using Ryujinx.HLE.HOS.Services.Account.Acc.AccountService;
 using Ryujinx.HLE.HOS.Services.Arp;
 
@@ -139,20 +138,21 @@ namespace Ryujinx.HLE.HOS.Services.Account.Acc
             return _applicationServiceServer.ClearSaveDataThumbnail(context);
         }
 
+        [CommandHipc(130)] // 5.0.0+
+        // LoadOpenContext(nn::account::Uid)
+        public ResultCode LoadOpenContext(ServiceCtx context)
+        {
+            Logger.Stub?.PrintStub(LogClass.ServiceAcc);
+
+            return ResultCode.Success;
+        }
+
+        [CommandHipc(60)] // 5.0.0-5.1.0
         [CommandHipc(131)] // 6.0.0+
         // ListOpenContextStoredUsers() -> array<nn::account::Uid, 0xa>
         public ResultCode ListOpenContextStoredUsers(ServiceCtx context)
         {
-            ulong outputPosition = context.Request.RecvListBuff[0].Position;
-            ulong outputSize     = context.Request.RecvListBuff[0].Size;
-
-            MemoryHelper.FillWithZeros(context.Memory, outputPosition, (int)outputSize);
-
-            // TODO: This seems to write stored userids of the OpenContext in the buffer. We needs to determine them.
-            
-            Logger.Stub?.PrintStub(LogClass.ServiceAcc);
-
-            return ResultCode.Success;
+            return _applicationServiceServer.ListOpenContextStoredUsers(context);
         }
 
         [CommandHipc(141)] // 6.0.0+

Ryujinx.HLE/HOS/Services/Bcat/ServiceCreator/IDeliveryCacheDirectoryService.cs

@@ -38,19 +38,18 @@ namespace Ryujinx.HLE.HOS.Services.Bcat.ServiceCreator
         // Read() -> (u32, buffer<nn::bcat::DeliveryCacheDirectoryEntry, 6>)
         public ResultCode Read(ServiceCtx context)
         {
-            ulong position = context.Request.ReceiveBuff[0].Position;
-            ulong size = context.Request.ReceiveBuff[0].Size;
+            ulong bufferAddress = context.Request.ReceiveBuff[0].Position;
+            ulong bufferLen = context.Request.ReceiveBuff[0].Size;
 
-            byte[] data = new byte[size];
-
-            Result result = _base.Get.Read(out int entriesRead, MemoryMarshal.Cast<byte, DeliveryCacheDirectoryEntry>(data));
-
-            context.Memory.Write(position, data);
+            using (var region = context.Memory.GetWritableRegion(bufferAddress, (int)bufferLen, true))
+            {
+                Result result = _base.Get.Read(out int entriesRead, MemoryMarshal.Cast<byte, DeliveryCacheDirectoryEntry>(region.Memory.Span));
 
                 context.ResponseData.Write(entriesRead);
 
                 return (ResultCode)result.Value;
             }
+        }
 
         [CommandHipc(2)]
         // GetCount() -> u32

Ryujinx.HLE/HOS/Services/Bcat/ServiceCreator/IDeliveryCacheFileService.cs

@@ -38,21 +38,20 @@ namespace Ryujinx.HLE.HOS.Services.Bcat.ServiceCreator
         // Read(u64) -> (u64, buffer<bytes, 6>)
         public ResultCode Read(ServiceCtx context)
         {
-            ulong position = context.Request.ReceiveBuff[0].Position;
-            ulong size = context.Request.ReceiveBuff[0].Size;
+            ulong bufferAddress = context.Request.ReceiveBuff[0].Position;
+            ulong bufferLen = context.Request.ReceiveBuff[0].Size;
 
             long offset = context.RequestData.ReadInt64();
 
-            byte[] data = new byte[size];
-
-            Result result = _base.Get.Read(out long bytesRead, offset, data);
-
-            context.Memory.Write(position, data);
+            using (var region = context.Memory.GetWritableRegion(bufferAddress, (int)bufferLen, true))
+            {
+                Result result = _base.Get.Read(out long bytesRead, offset, region.Memory.Span);
 
                 context.ResponseData.Write(bytesRead);
 
                 return (ResultCode)result.Value;
             }
+        }
 
         [CommandHipc(2)]
         // GetSize() -> u64

Ryujinx.HLE/HOS/Services/Bcat/ServiceCreator/IDeliveryCacheStorageService.cs

@@ -50,19 +50,18 @@ namespace Ryujinx.HLE.HOS.Services.Bcat.ServiceCreator
         // EnumerateDeliveryCacheDirectory() -> (u32, buffer<nn::bcat::DirectoryName, 6>)
         public ResultCode EnumerateDeliveryCacheDirectory(ServiceCtx context)
         {
-            ulong position = context.Request.ReceiveBuff[0].Position;
-            ulong size = context.Request.ReceiveBuff[0].Size;
+            ulong bufferAddress = context.Request.ReceiveBuff[0].Position;
+            ulong bufferLen = context.Request.ReceiveBuff[0].Size;
 
-            byte[] data = new byte[size];
-
-            Result result = _base.Get.EnumerateDeliveryCacheDirectory(out int count, MemoryMarshal.Cast<byte, DirectoryName>(data));
-
-            context.Memory.Write(position, data);
+            using (var region = context.Memory.GetWritableRegion(bufferAddress, (int)bufferLen, true))
+            {
+                Result result = _base.Get.EnumerateDeliveryCacheDirectory(out int count, MemoryMarshal.Cast<byte, DirectoryName>(region.Memory.Span));
 
                 context.ResponseData.Write(count);
 
                 return (ResultCode)result.Value;
             }
+        }
 
         protected override void Dispose(bool isDisposing)
         {

Ryujinx.HLE/HOS/Services/Fatal/IService.cs

@@ -1,8 +1,147 @@
-namespace Ryujinx.HLE.HOS.Services.Fatal
+using Ryujinx.Common.Logging;
+using Ryujinx.HLE.HOS.Services.Fatal.Types;
+using System;
+using System.Runtime.InteropServices;
+using System.Text;
+
+namespace Ryujinx.HLE.HOS.Services.Fatal
 {
     [Service("fatal:u")]
     class IService : IpcService
     {
         public IService(ServiceCtx context) { }
+
+        [CommandHipc(0)]
+        // ThrowFatal(u64 result_code, u64 pid)
+        public ResultCode ThrowFatal(ServiceCtx context)
+        {
+            ResultCode resultCode = (ResultCode)context.RequestData.ReadUInt64();
+            ulong      pid        = context.Request.HandleDesc.PId;
+
+            return ThrowFatalWithCpuContextImpl(context, resultCode, pid, FatalPolicy.ErrorReportAndErrorScreen, null);
+        }
+
+        [CommandHipc(1)]
+        // ThrowFatalWithPolicy(u64 result_code, u32 fatal_policy, u64 pid)
+        public ResultCode ThrowFatalWithPolicy(ServiceCtx context)
+        {
+            ResultCode  resultCode  = (ResultCode)context.RequestData.ReadUInt64();
+            FatalPolicy fatalPolicy = (FatalPolicy)context.RequestData.ReadUInt32();
+            ulong       pid         = context.Request.HandleDesc.PId;
+
+            return ThrowFatalWithCpuContextImpl(context, resultCode, pid, fatalPolicy, null);
+        }
+
+        [CommandHipc(2)]
+        // ThrowFatalWithCpuContext(u64 result_code, u32 fatal_policy, u64 pid, buffer<bytes, 0x15> cpu_context)
+        public ResultCode ThrowFatalWithCpuContext(ServiceCtx context)
+        {
+            ResultCode  resultCode  = (ResultCode)context.RequestData.ReadUInt64();
+            FatalPolicy fatalPolicy = (FatalPolicy)context.RequestData.ReadUInt32();
+            ulong       pid         = context.Request.HandleDesc.PId;
+
+            ulong cpuContextPosition = context.Request.SendBuff[0].Position;
+            ulong cpuContextSize     = context.Request.SendBuff[0].Size;
+
+            ReadOnlySpan<byte> cpuContextData = context.Memory.GetSpan(cpuContextPosition, (int)cpuContextSize);
+
+            return ThrowFatalWithCpuContextImpl(context, resultCode, pid, fatalPolicy, cpuContextData);
+        }
+
+        private ResultCode ThrowFatalWithCpuContextImpl(ServiceCtx context, ResultCode resultCode, ulong pid, FatalPolicy fatalPolicy, ReadOnlySpan<byte> cpuContext)
+        {
+            StringBuilder errorReport = new StringBuilder();
+
+            errorReport.AppendLine();
+            errorReport.AppendLine("ErrorReport log:");
+
+            errorReport.AppendLine($"\tTitleId: {context.Device.Application.TitleId:x16}");
+            errorReport.AppendLine($"\tPid: {pid}");
+            errorReport.AppendLine($"\tResultCode: {((int)resultCode & 0x1FF) + 2000}-{((int)resultCode >> 9) & 0x3FFF:d4}");
+            errorReport.AppendLine($"\tFatalPolicy: {fatalPolicy}");
+
+            if (cpuContext != null)
+            {
+                errorReport.AppendLine("CPU Context:");
+
+                if (context.Device.Application.TitleIs64Bit)
+                {
+                    CpuContext64 cpuContext64 = MemoryMarshal.Cast<byte, CpuContext64>(cpuContext)[0];
+
+                    errorReport.AppendLine($"\tStartAddress: 0x{cpuContext64.StartAddress:x16}");
+                    errorReport.AppendLine($"\tRegisterSetFlags: {cpuContext64.RegisterSetFlags}");
+
+                    if (cpuContext64.StackTraceSize > 0)
+                    {
+                        errorReport.AppendLine("\tStackTrace:");
+
+                        for (int i = 0; i < cpuContext64.StackTraceSize; i++)
+                        {
+                            errorReport.AppendLine($"\t\t0x{cpuContext64.StackTrace[i]:x16}");
+                        }
+                    }
+
+                    errorReport.AppendLine("\tRegisters:");
+
+                    for (int i = 0; i < cpuContext64.X.Length; i++)
+                    {
+                        errorReport.AppendLine($"\t\tX[{i:d2}]:\t0x{cpuContext64.X[i]:x16}");
+                    }
+
+                    errorReport.AppendLine();
+                    errorReport.AppendLine($"\t\tFP:\t0x{cpuContext64.FP:x16}");
+                    errorReport.AppendLine($"\t\tLR:\t0x{cpuContext64.LR:x16}");
+                    errorReport.AppendLine($"\t\tSP:\t0x{cpuContext64.SP:x16}");
+                    errorReport.AppendLine($"\t\tPC:\t0x{cpuContext64.PC:x16}");
+                    errorReport.AppendLine($"\t\tPState:\t0x{cpuContext64.PState:x16}");
+                    errorReport.AppendLine($"\t\tAfsr0:\t0x{cpuContext64.Afsr0:x16}");
+                    errorReport.AppendLine($"\t\tAfsr1:\t0x{cpuContext64.Afsr1:x16}");
+                    errorReport.AppendLine($"\t\tEsr:\t0x{cpuContext64.Esr:x16}");
+                    errorReport.AppendLine($"\t\tFar:\t0x{cpuContext64.Far:x16}");
+                }
+                else
+                {
+                    CpuContext32 cpuContext32 = MemoryMarshal.Cast<byte, CpuContext32>(cpuContext)[0];
+
+                    errorReport.AppendLine($"\tStartAddress: 0x{cpuContext32.StartAddress:16}");
+                    errorReport.AppendLine($"\tRegisterSetFlags: {cpuContext32.RegisterSetFlags}");
+
+                    if (cpuContext32.StackTraceSize > 0)
+                    {
+                        errorReport.AppendLine("\tStackTrace:");
+
+                        for (int i = 0; i < cpuContext32.StackTraceSize; i++)
+                        {
+                            errorReport.AppendLine($"\t\t0x{cpuContext32.StackTrace[i]:x16}");
+                        }
+                    }
+
+                    errorReport.AppendLine("\tRegisters:");
+
+                    for (int i = 0; i < cpuContext32.X.Length; i++)
+                    {
+                        errorReport.AppendLine($"\t\tX[{i:d2}]:\t0x{cpuContext32.X[i]:x16}");
+                    }
+
+                    errorReport.AppendLine();
+                    errorReport.AppendLine($"\t\tFP:\t0x{cpuContext32.FP:x16}");
+                    errorReport.AppendLine($"\t\tFP:\t0x{cpuContext32.IP:x16}");
+                    errorReport.AppendLine($"\t\tSP:\t0x{cpuContext32.SP:x16}");
+                    errorReport.AppendLine($"\t\tLR:\t0x{cpuContext32.LR:x16}");
+                    errorReport.AppendLine($"\t\tPC:\t0x{cpuContext32.PC:x16}");
+                    errorReport.AppendLine($"\t\tPState:\t0x{cpuContext32.PState:x16}");
+                    errorReport.AppendLine($"\t\tAfsr0:\t0x{cpuContext32.Afsr0:x16}");
+                    errorReport.AppendLine($"\t\tAfsr1:\t0x{cpuContext32.Afsr1:x16}");
+                    errorReport.AppendLine($"\t\tEsr:\t0x{cpuContext32.Esr:x16}");
+                    errorReport.AppendLine($"\t\tFar:\t0x{cpuContext32.Far:x16}");
+                }
+            }
+
+            Logger.Info?.Print(LogClass.ServiceFatal, errorReport.ToString());
+
+            context.Device.System.KernelContext.Syscall.Break((ulong)resultCode);
+
+            return ResultCode.Success;
+        }
     }
 }

Ryujinx.HLE/HOS/Services/Fatal/Types/CpuContext32.cs

@@ -0,0 +1,25 @@
+using Ryujinx.Common.Memory;
+
+namespace Ryujinx.HLE.HOS.Services.Fatal.Types
+{
+    public struct CpuContext32
+    {
+        public Array11<uint> X;
+        public uint FP;
+        public uint IP;
+        public uint SP;
+        public uint LR;
+        public uint PC;
+
+        public uint PState;
+        public uint Afsr0;
+        public uint Afsr1;
+        public uint Esr;
+        public uint Far;
+
+        public Array32<uint> StackTrace;
+        public uint StackTraceSize;
+        public uint StartAddress;
+        public uint RegisterSetFlags;
+    }
+}