OpenGL: Fix OBS/Overlays again by binding FB before present (#4668)

This seems to have been removed by the Post-Processing PR, but it is required for the display in OBS to be the right way up and properly scaled.

I've tested this with AA and FSR on MK8D and it seems to behave properly. Testing is welcome.

ARMeilleure/CodeGen/Arm64/IntrinsicTable.cs

@@ -226,6 +226,8 @@ namespace ARMeilleure.CodeGen.Arm64
             Add(Intrinsic.Arm64MlsVe,         new IntrinsicInfo(0x2f004000u, IntrinsicType.VectorTernaryRdByElem));
             Add(Intrinsic.Arm64MlsV,          new IntrinsicInfo(0x2e209400u, IntrinsicType.VectorTernaryRd));
             Add(Intrinsic.Arm64MoviV,         new IntrinsicInfo(0x0f000400u, IntrinsicType.VectorMovi));
+            Add(Intrinsic.Arm64MrsFpcr,       new IntrinsicInfo(0xd53b4400u, IntrinsicType.GetRegister));
+            Add(Intrinsic.Arm64MsrFpcr,       new IntrinsicInfo(0xd51b4400u, IntrinsicType.SetRegister));
             Add(Intrinsic.Arm64MrsFpsr,       new IntrinsicInfo(0xd53b4420u, IntrinsicType.GetRegister));
             Add(Intrinsic.Arm64MsrFpsr,       new IntrinsicInfo(0xd51b4420u, IntrinsicType.SetRegister));
             Add(Intrinsic.Arm64MulVe,         new IntrinsicInfo(0x0f008000u, IntrinsicType.VectorBinaryByElem));

ARMeilleure/CodeGen/X86/AssemblerTable.cs

@@ -268,11 +268,13 @@ namespace ARMeilleure.CodeGen.X86
             Add(X86Instruction.Vblendvps,     new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f3a4a, InstructionFlags.Vex | InstructionFlags.Prefix66));
             Add(X86Instruction.Vcvtph2ps,     new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f3813, InstructionFlags.Vex | InstructionFlags.Prefix66));
             Add(X86Instruction.Vcvtps2ph,     new InstructionInfo(0x000f3a1d, BadOp,      BadOp,      BadOp,      BadOp,      InstructionFlags.Vex | InstructionFlags.Prefix66));
+            Add(X86Instruction.Vfmadd231pd,   new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38b8, InstructionFlags.Vex | InstructionFlags.Prefix66 | InstructionFlags.RexW));
             Add(X86Instruction.Vfmadd231ps,   new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38b8, InstructionFlags.Vex | InstructionFlags.Prefix66));
             Add(X86Instruction.Vfmadd231sd,   new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38b9, InstructionFlags.Vex | InstructionFlags.Prefix66 | InstructionFlags.RexW));
             Add(X86Instruction.Vfmadd231ss,   new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38b9, InstructionFlags.Vex | InstructionFlags.Prefix66));
             Add(X86Instruction.Vfmsub231sd,   new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38bb, InstructionFlags.Vex | InstructionFlags.Prefix66 | InstructionFlags.RexW));
             Add(X86Instruction.Vfmsub231ss,   new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38bb, InstructionFlags.Vex | InstructionFlags.Prefix66));
+            Add(X86Instruction.Vfnmadd231pd,  new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38bc, InstructionFlags.Vex | InstructionFlags.Prefix66 | InstructionFlags.RexW));
             Add(X86Instruction.Vfnmadd231ps,  new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38bc, InstructionFlags.Vex | InstructionFlags.Prefix66));
             Add(X86Instruction.Vfnmadd231sd,  new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38bd, InstructionFlags.Vex | InstructionFlags.Prefix66 | InstructionFlags.RexW));
             Add(X86Instruction.Vfnmadd231ss,  new InstructionInfo(BadOp,      BadOp,      BadOp,      BadOp,      0x000f38bd, InstructionFlags.Vex | InstructionFlags.Prefix66));

ARMeilleure/CodeGen/X86/CodeGenerator.cs

@@ -249,10 +249,9 @@ namespace ARMeilleure.CodeGen.X86
                     case IntrinsicType.Mxcsr:
                     {
                         Operand offset = operation.GetSource(0);
-                        Operand bits   = operation.GetSource(1);
 
-                        Debug.Assert(offset.Kind == OperandKind.Constant && bits.Kind == OperandKind.Constant);
+                        Debug.Assert(offset.Kind == OperandKind.Constant);
-                        Debug.Assert(offset.Type == OperandType.I32 && bits.Type == OperandType.I32);
+                        Debug.Assert(offset.Type == OperandType.I32);
 
                         int offs = offset.AsInt32() + context.CallArgsRegionSize;
 
@@ -261,20 +260,22 @@ namespace ARMeilleure.CodeGen.X86
 
                         Debug.Assert(HardwareCapabilities.SupportsSse || HardwareCapabilities.SupportsVexEncoding);
 
-                        context.Assembler.Stmxcsr(memOp);
+                        if (operation.Intrinsic == Intrinsic.X86Ldmxcsr)
-
-                        if (operation.Intrinsic == Intrinsic.X86Mxcsrmb)
                         {
-                            context.Assembler.Or(memOp, bits, OperandType.I32);
+                            Operand bits = operation.GetSource(1);
-                        }
+                            Debug.Assert(bits.Type == OperandType.I32);
-                        else /* if (intrinOp.Intrinsic == Intrinsic.X86Mxcsrub) */
-                        {
-                            Operand notBits = Const(~bits.AsInt32());
-
-                            context.Assembler.And(memOp, notBits, OperandType.I32);
-                        }
 
+                            context.Assembler.Mov(memOp, bits, OperandType.I32);
                             context.Assembler.Ldmxcsr(memOp);
+                        }
+                        else if (operation.Intrinsic == Intrinsic.X86Stmxcsr)
+                        {
+                            Operand dest = operation.Destination;
+                            Debug.Assert(dest.Type == OperandType.I32);
+
+                            context.Assembler.Stmxcsr(memOp);
+                            context.Assembler.Mov(dest, memOp, OperandType.I32);
+                        }
 
                         break;
                     }

ARMeilleure/CodeGen/X86/IntrinsicTable.cs

@@ -60,6 +60,7 @@ namespace ARMeilleure.CodeGen.X86
             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));
+            Add(Intrinsic.X86Ldmxcsr,       new IntrinsicInfo(X86Instruction.None,          IntrinsicType.Mxcsr));
             Add(Intrinsic.X86Maxpd,         new IntrinsicInfo(X86Instruction.Maxpd,         IntrinsicType.Binary));
             Add(Intrinsic.X86Maxps,         new IntrinsicInfo(X86Instruction.Maxps,         IntrinsicType.Binary));
             Add(Intrinsic.X86Maxsd,         new IntrinsicInfo(X86Instruction.Maxsd,         IntrinsicType.Binary));
@@ -75,8 +76,6 @@ namespace ARMeilleure.CodeGen.X86
             Add(Intrinsic.X86Mulps,         new IntrinsicInfo(X86Instruction.Mulps,         IntrinsicType.Binary));
             Add(Intrinsic.X86Mulsd,         new IntrinsicInfo(X86Instruction.Mulsd,         IntrinsicType.Binary));
             Add(Intrinsic.X86Mulss,         new IntrinsicInfo(X86Instruction.Mulss,         IntrinsicType.Binary));
-            Add(Intrinsic.X86Mxcsrmb,       new IntrinsicInfo(X86Instruction.None,          IntrinsicType.Mxcsr)); // Mask bits.
-            Add(Intrinsic.X86Mxcsrub,       new IntrinsicInfo(X86Instruction.None,          IntrinsicType.Mxcsr)); // Unmask bits.
             Add(Intrinsic.X86Paddb,         new IntrinsicInfo(X86Instruction.Paddb,         IntrinsicType.Binary));
             Add(Intrinsic.X86Paddd,         new IntrinsicInfo(X86Instruction.Paddd,         IntrinsicType.Binary));
             Add(Intrinsic.X86Paddq,         new IntrinsicInfo(X86Instruction.Paddq,         IntrinsicType.Binary));
@@ -160,6 +159,7 @@ namespace ARMeilleure.CodeGen.X86
             Add(Intrinsic.X86Sqrtps,        new IntrinsicInfo(X86Instruction.Sqrtps,        IntrinsicType.Unary));
             Add(Intrinsic.X86Sqrtsd,        new IntrinsicInfo(X86Instruction.Sqrtsd,        IntrinsicType.Unary));
             Add(Intrinsic.X86Sqrtss,        new IntrinsicInfo(X86Instruction.Sqrtss,        IntrinsicType.Unary));
+            Add(Intrinsic.X86Stmxcsr,       new IntrinsicInfo(X86Instruction.None,          IntrinsicType.Mxcsr));
             Add(Intrinsic.X86Subpd,         new IntrinsicInfo(X86Instruction.Subpd,         IntrinsicType.Binary));
             Add(Intrinsic.X86Subps,         new IntrinsicInfo(X86Instruction.Subps,         IntrinsicType.Binary));
             Add(Intrinsic.X86Subsd,         new IntrinsicInfo(X86Instruction.Subsd,         IntrinsicType.Binary));
@@ -170,11 +170,13 @@ namespace ARMeilleure.CodeGen.X86
             Add(Intrinsic.X86Unpcklps,      new IntrinsicInfo(X86Instruction.Unpcklps,      IntrinsicType.Binary));
             Add(Intrinsic.X86Vcvtph2ps,     new IntrinsicInfo(X86Instruction.Vcvtph2ps,     IntrinsicType.Unary));
             Add(Intrinsic.X86Vcvtps2ph,     new IntrinsicInfo(X86Instruction.Vcvtps2ph,     IntrinsicType.BinaryImm));
+            Add(Intrinsic.X86Vfmadd231pd,   new IntrinsicInfo(X86Instruction.Vfmadd231pd,   IntrinsicType.Fma));
             Add(Intrinsic.X86Vfmadd231ps,   new IntrinsicInfo(X86Instruction.Vfmadd231ps,   IntrinsicType.Fma));
             Add(Intrinsic.X86Vfmadd231sd,   new IntrinsicInfo(X86Instruction.Vfmadd231sd,   IntrinsicType.Fma));
             Add(Intrinsic.X86Vfmadd231ss,   new IntrinsicInfo(X86Instruction.Vfmadd231ss,   IntrinsicType.Fma));
             Add(Intrinsic.X86Vfmsub231sd,   new IntrinsicInfo(X86Instruction.Vfmsub231sd,   IntrinsicType.Fma));
             Add(Intrinsic.X86Vfmsub231ss,   new IntrinsicInfo(X86Instruction.Vfmsub231ss,   IntrinsicType.Fma));
+            Add(Intrinsic.X86Vfnmadd231pd,  new IntrinsicInfo(X86Instruction.Vfnmadd231pd,  IntrinsicType.Fma));
             Add(Intrinsic.X86Vfnmadd231ps,  new IntrinsicInfo(X86Instruction.Vfnmadd231ps,  IntrinsicType.Fma));
             Add(Intrinsic.X86Vfnmadd231sd,  new IntrinsicInfo(X86Instruction.Vfnmadd231sd,  IntrinsicType.Fma));
             Add(Intrinsic.X86Vfnmadd231ss,  new IntrinsicInfo(X86Instruction.Vfnmadd231ss,  IntrinsicType.Fma));

ARMeilleure/CodeGen/X86/Mxcsr.cs

@@ -0,0 +1,15 @@
+using System;
+
+namespace ARMeilleure.CodeGen.X86
+{
+    [Flags]
+    enum Mxcsr
+    {
+        Ftz = 1 << 15, // Flush To Zero.
+        Rhi = 1 << 14, // Round Mode high bit.
+        Rlo = 1 << 13, // Round Mode low bit.
+        Um = 1 << 11,  // Underflow Mask.
+        Dm = 1 << 8,   // Denormal Mask.
+        Daz = 1 << 6   // Denormals Are Zero.
+    }
+}

ARMeilleure/CodeGen/X86/PreAllocator.cs

@@ -120,12 +120,18 @@ namespace ARMeilleure.CodeGen.X86
                             break;
 
                         case Instruction.Extended:
-                            if (node.Intrinsic == Intrinsic.X86Mxcsrmb || node.Intrinsic == Intrinsic.X86Mxcsrub)
+                            if (node.Intrinsic == Intrinsic.X86Ldmxcsr)
                             {
                                 int stackOffset = stackAlloc.Allocate(OperandType.I32);
 
                                 node.SetSources(new Operand[] { Const(stackOffset), node.GetSource(0) });
                             }
+                            else if (node.Intrinsic == Intrinsic.X86Stmxcsr)
+                            {
+                                int stackOffset = stackAlloc.Allocate(OperandType.I32);
+
+                                node.SetSources(new Operand[] { Const(stackOffset) });
+                            }
                             break;
                     }
                 }

ARMeilleure/CodeGen/X86/X86Instruction.cs

@@ -208,11 +208,13 @@ namespace ARMeilleure.CodeGen.X86
         Vblendvps,
         Vcvtph2ps,
         Vcvtps2ph,
+        Vfmadd231pd,
         Vfmadd231ps,
         Vfmadd231sd,
         Vfmadd231ss,
         Vfmsub231sd,
         Vfmsub231ss,
+        Vfnmadd231pd,
         Vfnmadd231ps,
         Vfnmadd231sd,
         Vfnmadd231ss,

ARMeilleure/Decoders/OpCodeTable.cs

@@ -108,6 +108,13 @@ namespace ARMeilleure.Decoders
             SetA64("11001010xx0xxxxxxxxxxxxxxxxxxxxx", InstName.Eor,             InstEmit.Eor,             OpCodeAluRs.Create);
             SetA64("00010011100xxxxx0xxxxxxxxxxxxxxx", InstName.Extr,            InstEmit.Extr,            OpCodeAluRs.Create);
             SetA64("10010011110xxxxxxxxxxxxxxxxxxxxx", InstName.Extr,            InstEmit.Extr,            OpCodeAluRs.Create);
+            SetA64("11010101000000110010000011011111", InstName.Hint,            InstEmit.Nop,             OpCodeSystem.Create); // Reserved Hint
+            SetA64("11010101000000110010000011111111", InstName.Hint,            InstEmit.Nop,             OpCodeSystem.Create); // Reserved Hint
+            SetA64("110101010000001100100001xxx11111", InstName.Hint,            InstEmit.Nop,             OpCodeSystem.Create); // Reserved Hint
+            SetA64("1101010100000011001000100xx11111", InstName.Hint,            InstEmit.Nop,             OpCodeSystem.Create); // Reserved Hint
+            SetA64("1101010100000011001000101>>11111", InstName.Hint,            InstEmit.Nop,             OpCodeSystem.Create); // Reserved Hint
+            SetA64("110101010000001100100011xxx11111", InstName.Hint,            InstEmit.Nop,             OpCodeSystem.Create); // Reserved Hint
+            SetA64("11010101000000110010>>xxxxx11111", InstName.Hint,            InstEmit.Nop,             OpCodeSystem.Create); // Reserved Hint
             SetA64("11010101000000110011xxxx11011111", InstName.Isb,             InstEmit.Isb,             OpCodeSystem.Create);
             SetA64("xx001000110xxxxx1xxxxxxxxxxxxxxx", InstName.Ldar,            InstEmit.Ldar,            OpCodeMemEx.Create);
             SetA64("1x001000011xxxxx1xxxxxxxxxxxxxxx", InstName.Ldaxp,           InstEmit.Ldaxp,           OpCodeMemEx.Create);

ARMeilleure/Instructions/InstEmitSimdArithmetic.cs

@@ -614,8 +614,6 @@ namespace ARMeilleure.Instructions
                 EmitSse2VectorPairwiseOpF(context, (op1, op2) =>
                 {
                     return EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
-                    {
-                        return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
                     {
                         IOpCodeSimd op = (IOpCodeSimd)context.CurrOp;
 
@@ -623,7 +621,6 @@ namespace ARMeilleure.Instructions
 
                         return context.AddIntrinsic(addInst, op1, op2);
                     }, scalar: false, op1, op2);
-                    }, scalar: false, op1, op2);
                 });
             }
             else
@@ -696,17 +693,33 @@ namespace ARMeilleure.Instructions
                 Operand n = GetVec(op.Rn);
                 Operand m = GetVec(op.Rm);
 
+                Operand res;
+
                 if (op.Size == 0)
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmadd231ss, a, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Addss, a, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper96(res));
                 }
                 else /* if (op.Size == 1) */
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmadd231sd, a, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Addsd, a, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper64(res));
                 }
@@ -729,11 +742,8 @@ namespace ARMeilleure.Instructions
             else if (Optimizations.FastFP && Optimizations.UseSse41)
             {
                 EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
-                {
-                    return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
                 {
                     return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: true);
-                    }, scalar: true, op1, op2);
                 }, scalar: true);
             }
             else
@@ -754,11 +764,8 @@ namespace ARMeilleure.Instructions
             else if (Optimizations.FastFP && Optimizations.UseSse41)
             {
                 EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
-                {
-                    return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
                 {
                     return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: true);
-                    }, scalar: false, op1, op2);
                 }, scalar: false);
             }
             else
@@ -885,12 +892,9 @@ namespace ARMeilleure.Instructions
                 EmitSse2VectorPairwiseOpF(context, (op1, op2) =>
                 {
                     return EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
-                    {
-                        return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
                     {
                         return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: true);
                     }, scalar: false, op1, op2);
-                    }, scalar: false, op1, op2);
                 });
             }
             else
@@ -913,12 +917,9 @@ namespace ARMeilleure.Instructions
                 EmitSse2VectorAcrossVectorOpF(context, (op1, op2) =>
                 {
                     return EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
-                    {
-                        return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
                     {
                         return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: true);
                     }, scalar: false, op1, op2);
-                    }, scalar: false, op1, op2);
                 });
             }
             else
@@ -939,11 +940,8 @@ namespace ARMeilleure.Instructions
             else if (Optimizations.FastFP && Optimizations.UseSse41)
             {
                 EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
-                {
-                    return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
                 {
                     return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: false);
-                    }, scalar: true, op1, op2);
                 }, scalar: true);
             }
             else
@@ -964,11 +962,8 @@ namespace ARMeilleure.Instructions
             else if (Optimizations.FastFP && Optimizations.UseSse41)
             {
                 EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
-                {
-                    return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
                 {
                     return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: false);
-                    }, scalar: false, op1, op2);
                 }, scalar: false);
             }
             else
@@ -1095,12 +1090,9 @@ namespace ARMeilleure.Instructions
                 EmitSse2VectorPairwiseOpF(context, (op1, op2) =>
                 {
                     return EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
-                    {
-                        return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
                     {
                         return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: false);
                     }, scalar: false, op1, op2);
-                    }, scalar: false, op1, op2);
                 });
             }
             else
@@ -1123,12 +1115,9 @@ namespace ARMeilleure.Instructions
                 EmitSse2VectorAcrossVectorOpF(context, (op1, op2) =>
                 {
                     return EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
-                    {
-                        return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
                     {
                         return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: false);
                     }, scalar: false, op1, op2);
-                    }, scalar: false, op1, op2);
                 });
             }
             else
@@ -1146,6 +1135,37 @@ namespace ARMeilleure.Instructions
             {
                 InstEmitSimdHelperArm64.EmitScalarTernaryOpFRdByElem(context, Intrinsic.Arm64FmlaSe);
             }
+            else if (Optimizations.UseFma)
+            {
+                OpCodeSimdRegElemF op = (OpCodeSimdRegElemF)context.CurrOp;
+
+                Operand d = GetVec(op.Rd);
+                Operand n = GetVec(op.Rn);
+                Operand m = GetVec(op.Rm);
+
+                int sizeF = op.Size & 1;
+
+                if (sizeF == 0)
+                {
+                    int shuffleMask = op.Index | op.Index << 2 | op.Index << 4 | op.Index << 6;
+
+                    Operand res = context.AddIntrinsic(Intrinsic.X86Shufps, m, m, Const(shuffleMask));
+
+                    res = context.AddIntrinsic(Intrinsic.X86Vfmadd231ss, d, n, res);
+
+                    context.Copy(d, context.VectorZeroUpper96(res));
+                }
+                else /* if (sizeF == 1) */
+                {
+                    int shuffleMask = op.Index | op.Index << 1;
+
+                    Operand res = context.AddIntrinsic(Intrinsic.X86Shufpd, m, m, Const(shuffleMask));
+
+                    res = context.AddIntrinsic(Intrinsic.X86Vfmadd231sd, d, n, res);
+
+                    context.Copy(d, context.VectorZeroUpper64(res));
+                }
+            }
             else
             {
                 EmitScalarTernaryOpByElemF(context, (op1, op2, op3) =>
@@ -1171,11 +1191,19 @@ namespace ARMeilleure.Instructions
 
                 int sizeF = op.Size & 1;
 
+                Operand res;
+
                 if (sizeF == 0)
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
+                    if (Optimizations.UseFma)
-
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmadd231ps, d, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Addps, d, res);
+                    }
 
                     if (op.RegisterSize == RegisterSize.Simd64)
                     {
@@ -1186,9 +1214,15 @@ namespace ARMeilleure.Instructions
                 }
                 else /* if (sizeF == 1) */
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
+                    if (Optimizations.UseFma)
-
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmadd231pd, d, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Addpd, d, res);
+                    }
 
                     context.Copy(d, res);
                 }
@@ -1224,8 +1258,15 @@ namespace ARMeilleure.Instructions
 
                     Operand res = context.AddIntrinsic(Intrinsic.X86Shufps, m, m, Const(shuffleMask));
 
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmadd231ps, d, n, res);
+                    }
+                    else
+                    {
                         res = context.AddIntrinsic(Intrinsic.X86Mulps, n, res);
                         res = context.AddIntrinsic(Intrinsic.X86Addps, d, res);
+                    }
 
                     if (op.RegisterSize == RegisterSize.Simd64)
                     {
@@ -1240,8 +1281,15 @@ namespace ARMeilleure.Instructions
 
                     Operand res = context.AddIntrinsic(Intrinsic.X86Shufpd, m, m, Const(shuffleMask));
 
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmadd231pd, d, n, res);
+                    }
+                    else
+                    {
                         res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, res);
                         res = context.AddIntrinsic(Intrinsic.X86Addpd, d, res);
+                    }
 
                     context.Copy(d, res);
                 }
@@ -1261,6 +1309,37 @@ namespace ARMeilleure.Instructions
             {
                 InstEmitSimdHelperArm64.EmitScalarTernaryOpFRdByElem(context, Intrinsic.Arm64FmlsSe);
             }
+            else if (Optimizations.UseFma)
+            {
+                OpCodeSimdRegElemF op = (OpCodeSimdRegElemF)context.CurrOp;
+
+                Operand d = GetVec(op.Rd);
+                Operand n = GetVec(op.Rn);
+                Operand m = GetVec(op.Rm);
+
+                int sizeF = op.Size & 1;
+
+                if (sizeF == 0)
+                {
+                    int shuffleMask = op.Index | op.Index << 2 | op.Index << 4 | op.Index << 6;
+
+                    Operand res = context.AddIntrinsic(Intrinsic.X86Shufps, m, m, Const(shuffleMask));
+
+                    res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ss, d, n, res);
+
+                    context.Copy(d, context.VectorZeroUpper96(res));
+                }
+                else /* if (sizeF == 1) */
+                {
+                    int shuffleMask = op.Index | op.Index << 1;
+
+                    Operand res = context.AddIntrinsic(Intrinsic.X86Shufpd, m, m, Const(shuffleMask));
+
+                    res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231sd, d, n, res);
+
+                    context.Copy(d, context.VectorZeroUpper64(res));
+                }
+            }
             else
             {
                 EmitScalarTernaryOpByElemF(context, (op1, op2, op3) =>
@@ -1286,11 +1365,19 @@ namespace ARMeilleure.Instructions
 
                 int sizeF = op.Size & 1;
 
+                Operand res;
+
                 if (sizeF == 0)
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
+                    if (Optimizations.UseFma)
-
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ps, d, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Subps, d, res);
+                    }
 
                     if (op.RegisterSize == RegisterSize.Simd64)
                     {
@@ -1301,9 +1388,15 @@ namespace ARMeilleure.Instructions
                 }
                 else /* if (sizeF == 1) */
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
+                    if (Optimizations.UseFma)
-
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231pd, d, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Subpd, d, res);
+                    }
 
                     context.Copy(d, res);
                 }
@@ -1339,8 +1432,15 @@ namespace ARMeilleure.Instructions
 
                     Operand res = context.AddIntrinsic(Intrinsic.X86Shufps, m, m, Const(shuffleMask));
 
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ps, d, n, res);
+                    }
+                    else
+                    {
                         res = context.AddIntrinsic(Intrinsic.X86Mulps, n, res);
                         res = context.AddIntrinsic(Intrinsic.X86Subps, d, res);
+                    }
 
                     if (op.RegisterSize == RegisterSize.Simd64)
                     {
@@ -1355,8 +1455,15 @@ namespace ARMeilleure.Instructions
 
                     Operand res = context.AddIntrinsic(Intrinsic.X86Shufpd, m, m, Const(shuffleMask));
 
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231pd, d, n, res);
+                    }
+                    else
+                    {
                         res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, res);
                         res = context.AddIntrinsic(Intrinsic.X86Subpd, d, res);
+                    }
 
                     context.Copy(d, res);
                 }
@@ -1385,17 +1492,33 @@ namespace ARMeilleure.Instructions
                 Operand n = GetVec(op.Rn);
                 Operand m = GetVec(op.Rm);
 
+                Operand res;
+
                 if (op.Size == 0)
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ss, a, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Subss, a, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper96(res));
                 }
                 else /* if (op.Size == 1) */
                 {
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231sd, a, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Subsd, a, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper64(res));
                 }
@@ -1669,25 +1792,39 @@ namespace ARMeilleure.Instructions
                 Operand n = GetVec(op.Rn);
                 Operand m = GetVec(op.Rm);
 
+                Operand res;
+
                 if (op.Size == 0)
+                {
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmsub231ss, a, n, m);
+                    }
+                    else
                     {
                         Operand mask = X86GetScalar(context, -0f);
-
                         Operand aNeg = context.AddIntrinsic(Intrinsic.X86Xorps, mask, a);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Subss, aNeg, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper96(res));
                 }
                 else /* if (op.Size == 1) */
+                {
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmsub231sd, a, n, m);
+                    }
+                    else
                     {
                         Operand mask = X86GetScalar(context, -0d);
-
                         Operand aNeg = context.AddIntrinsic(Intrinsic.X86Xorpd, mask, a);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Subsd, aNeg, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper64(res));
                 }
@@ -1716,25 +1853,39 @@ namespace ARMeilleure.Instructions
                 Operand n = GetVec(op.Rn);
                 Operand m = GetVec(op.Rm);
 
+                Operand res;
+
                 if (op.Size == 0)
+                {
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmsub231ss, a, n, m);
+                    }
+                    else
                     {
                         Operand mask = X86GetScalar(context, -0f);
-
                         Operand aNeg = context.AddIntrinsic(Intrinsic.X86Xorps, mask, a);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Addss, aNeg, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper96(res));
                 }
                 else /* if (op.Size == 1) */
+                {
+                    if (Optimizations.UseFma)
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfmsub231sd, a, n, m);
+                    }
+                    else
                     {
                         Operand mask = X86GetScalar(context, -0d);
-
                         Operand aNeg = context.AddIntrinsic(Intrinsic.X86Xorpd, mask, a);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
+                        res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Addsd, aNeg, res);
+                    }
 
                     context.Copy(d, context.VectorZeroUpper64(res));
                 }
@@ -1830,13 +1981,22 @@ namespace ARMeilleure.Instructions
 
                 int sizeF = op.Size & 1;
 
+                Operand res;
+
                 if (sizeF == 0)
                 {
                     Operand mask = X86GetScalar(context, 2f);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
+                    if (Optimizations.UseFma)
-
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ss, mask, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Subss, mask, res);
+                    }
+
                     res = EmitSse41RecipStepSelectOpF(context, n, m, res, mask, scalar: true, sizeF);
 
                     context.Copy(GetVec(op.Rd), context.VectorZeroUpper96(res));
@@ -1845,9 +2005,16 @@ namespace ARMeilleure.Instructions
                 {
                     Operand mask = X86GetScalar(context, 2d);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
+                    if (Optimizations.UseFma)
-
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231sd, mask, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Subsd, mask, res);
+                    }
+
                     res = EmitSse41RecipStepSelectOpF(context, n, m, res, mask, scalar: true, sizeF);
 
                     context.Copy(GetVec(op.Rd), context.VectorZeroUpper64(res));
@@ -1877,14 +2044,23 @@ namespace ARMeilleure.Instructions
 
                 int sizeF = op.Size & 1;
 
+                Operand res;
+
                 if (sizeF == 0)
                 {
                     Operand mask = X86GetAllElements(context, 2f);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
+                    if (Optimizations.UseFma)
-                    res = EmitSse41RecipStepSelectOpF(context, n, m, res, mask, scalar: false, sizeF);
+                    {
-
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ps, mask, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Subps, mask, res);
+                    }
+
+                    res = EmitSse41RecipStepSelectOpF(context, n, m, res, mask, scalar: false, sizeF);
 
                     if (op.RegisterSize == RegisterSize.Simd64)
                     {
@@ -1897,10 +2073,17 @@ namespace ARMeilleure.Instructions
                 {
                     Operand mask = X86GetAllElements(context, 2d);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
+                    if (Optimizations.UseFma)
-                    res = EmitSse41RecipStepSelectOpF(context, n, m, res, mask, scalar: false, sizeF);
+                    {
-
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231pd, mask, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Subpd, mask, res);
+                    }
+
+                    res = EmitSse41RecipStepSelectOpF(context, n, m, res, mask, scalar: false, sizeF);
 
                     context.Copy(GetVec(op.Rd), res);
                 }
@@ -2113,21 +2296,33 @@ namespace ARMeilleure.Instructions
 
         public static void Frintx_S(ArmEmitterContext context)
         {
-            // TODO Arm64: Fast path. Should we set host FPCR?
+            if (Optimizations.UseAdvSimd)
+            {
+                InstEmitSimdHelperArm64.EmitScalarUnaryOpF(context, Intrinsic.Arm64FrintxS);
+            }
+            else
+            {
                 EmitScalarUnaryOpF(context, (op1) =>
                 {
                     return EmitRoundByRMode(context, op1);
                 });
             }
+        }
 
         public static void Frintx_V(ArmEmitterContext context)
         {
-            // TODO Arm64: Fast path. Should we set host FPCR?
+            if (Optimizations.UseAdvSimd)
+            {
+                InstEmitSimdHelperArm64.EmitVectorUnaryOpF(context, Intrinsic.Arm64FrintxV);
+            }
+            else
+            {
                 EmitVectorUnaryOpF(context, (op1) =>
                 {
                     return EmitRoundByRMode(context, op1);
                 });
             }
+        }
 
         public static void Frintz_S(ArmEmitterContext context)
         {
@@ -2237,15 +2432,24 @@ namespace ARMeilleure.Instructions
 
                 int sizeF = op.Size & 1;
 
+                Operand res;
+
                 if (sizeF == 0)
                 {
                     Operand maskHalf    = X86GetScalar(context, 0.5f);
                     Operand maskThree   = X86GetScalar(context, 3f);
                     Operand maskOneHalf = X86GetScalar(context, 1.5f);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
+                    if (Optimizations.UseFma)
-
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ss, maskThree, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulss, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Subss, maskThree, res);
+                    }
+
                     res = context.AddIntrinsic(Intrinsic.X86Mulss, maskHalf, res);
                     res = EmitSse41RecipStepSelectOpF(context, n, m, res, maskOneHalf, scalar: true, sizeF);
 
@@ -2257,9 +2461,16 @@ namespace ARMeilleure.Instructions
                     Operand maskThree   = X86GetScalar(context, 3d);
                     Operand maskOneHalf = X86GetScalar(context, 1.5d);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
+                    if (Optimizations.UseFma)
-
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231sd, maskThree, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulsd, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Subsd, maskThree, res);
+                    }
+
                     res = context.AddIntrinsic(Intrinsic.X86Mulsd, maskHalf, res);
                     res = EmitSse41RecipStepSelectOpF(context, n, m, res, maskOneHalf, scalar: true, sizeF);
 
@@ -2290,15 +2501,24 @@ namespace ARMeilleure.Instructions
 
                 int sizeF = op.Size & 1;
 
+                Operand res;
+
                 if (sizeF == 0)
                 {
                     Operand maskHalf    = X86GetAllElements(context, 0.5f);
                     Operand maskThree   = X86GetAllElements(context, 3f);
                     Operand maskOneHalf = X86GetAllElements(context, 1.5f);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
+                    if (Optimizations.UseFma)
-
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231ps, maskThree, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulps, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Subps, maskThree, res);
+                    }
+
                     res = context.AddIntrinsic(Intrinsic.X86Mulps, maskHalf,  res);
                     res = EmitSse41RecipStepSelectOpF(context, n, m, res, maskOneHalf, scalar: false, sizeF);
 
@@ -2315,9 +2535,16 @@ namespace ARMeilleure.Instructions
                     Operand maskThree   = X86GetAllElements(context, 3d);
                     Operand maskOneHalf = X86GetAllElements(context, 1.5d);
 
-                    Operand res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
+                    if (Optimizations.UseFma)
-
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Vfnmadd231pd, maskThree, n, m);
+                    }
+                    else
+                    {
+                        res = context.AddIntrinsic(Intrinsic.X86Mulpd, n, m);
                         res = context.AddIntrinsic(Intrinsic.X86Subpd, maskThree, res);
+                    }
+
                     res = context.AddIntrinsic(Intrinsic.X86Mulpd, maskHalf,  res);
                     res = EmitSse41RecipStepSelectOpF(context, n, m, res, maskOneHalf, scalar: false, sizeF);
 
@@ -4728,53 +4955,6 @@ namespace ARMeilleure.Instructions
             }
         }
 
-        public static Operand EmitSseOrAvxHandleFzModeOpF(
-            ArmEmitterContext context,
-            Func2I emit,
-            bool scalar,
-            Operand n = default,
-            Operand m = default)
-        {
-            Operand nCopy = n == default ? context.Copy(GetVec(((OpCodeSimdReg)context.CurrOp).Rn)) : n;
-            Operand mCopy = m == default ? context.Copy(GetVec(((OpCodeSimdReg)context.CurrOp).Rm)) : m;
-
-            EmitSseOrAvxEnterFtzAndDazModesOpF(context, out Operand isTrue);
-
-            Operand res = emit(nCopy, mCopy);
-
-            EmitSseOrAvxExitFtzAndDazModesOpF(context, isTrue);
-
-            if (n != default || m != default)
-            {
-                return res;
-            }
-
-            int sizeF = ((IOpCodeSimd)context.CurrOp).Size & 1;
-
-            if (sizeF == 0)
-            {
-                if (scalar)
-                {
-                    res = context.VectorZeroUpper96(res);
-                }
-                else if (((OpCodeSimdReg)context.CurrOp).RegisterSize == RegisterSize.Simd64)
-                {
-                    res = context.VectorZeroUpper64(res);
-                }
-            }
-            else /* if (sizeF == 1) */
-            {
-                if (scalar)
-                {
-                    res = context.VectorZeroUpper64(res);
-                }
-            }
-
-            context.Copy(GetVec(((OpCodeSimdReg)context.CurrOp).Rd), res);
-
-            return default;
-        }
-
         private static Operand EmitSse2VectorMaxMinOpF(ArmEmitterContext context, Operand n, Operand m, bool isMax)
         {
             IOpCodeSimd op = (IOpCodeSimd)context.CurrOp;
@@ -4833,11 +5013,8 @@ namespace ARMeilleure.Instructions
                 mCopy = context.AddIntrinsic(Intrinsic.X86Blendvps, mCopy, negInfMask, mMask);
 
                 Operand res = EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
-                {
-                    return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
                 {
                     return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: isMaxNum);
-                    }, scalar: scalar, op1, op2);
                 }, scalar: scalar, nCopy, mCopy);
 
                 if (n != default || m != default)
@@ -4871,11 +5048,8 @@ namespace ARMeilleure.Instructions
                 mCopy = context.AddIntrinsic(Intrinsic.X86Blendvpd, mCopy, negInfMask, mMask);
 
                 Operand res = EmitSse41ProcessNaNsOpF(context, (op1, op2) =>
-                {
-                    return EmitSseOrAvxHandleFzModeOpF(context, (op1, op2) =>
                 {
                     return EmitSse2VectorMaxMinOpF(context, op1, op2, isMax: isMaxNum);
-                    }, scalar: scalar, op1, op2);
                 }, scalar: scalar, nCopy, mCopy);
 
                 if (n != default || m != default)

ARMeilleure/Instructions/InstEmitSimdCvt32.cs

@@ -356,9 +356,11 @@ namespace ARMeilleure.Instructions
                         ? typeof(SoftFloat64_16).GetMethod(nameof(SoftFloat64_16.FPConvert))
                         : typeof(SoftFloat32_16).GetMethod(nameof(SoftFloat32_16.FPConvert));
 
+                    context.ExitArmFpMode();
                     context.StoreToContext();
                     Operand res = context.Call(method, src);
                     context.LoadFromContext();
+                    context.EnterArmFpMode();
 
                     InsertScalar16(context, op.Vd, op.T, res);
                 }
@@ -372,9 +374,11 @@ namespace ARMeilleure.Instructions
                         ? typeof(SoftFloat16_64).GetMethod(nameof(SoftFloat16_64.FPConvert))
                         : typeof(SoftFloat16_32).GetMethod(nameof(SoftFloat16_32.FPConvert));
 
+                    context.ExitArmFpMode();
                     context.StoreToContext();
                     Operand res = context.Call(method, src);
                     context.LoadFromContext();
+                    context.EnterArmFpMode();
 
                     InsertScalar(context, op.Vd, res);
                 }
@@ -541,12 +545,19 @@ namespace ARMeilleure.Instructions
 
         // VRINTX (floating-point).
         public static void Vrintx_S(ArmEmitterContext context)
+        {
+            if (Optimizations.UseAdvSimd)
+            {
+                InstEmitSimdHelper32Arm64.EmitScalarUnaryOpF32(context, Intrinsic.Arm64FrintxS);
+            }
+            else
             {
                 EmitScalarUnaryOpF32(context, (op1) =>
                 {
                     return EmitRoundByRMode(context, op1);
                 });
             }
+        }
 
         private static Operand EmitFPConvert(ArmEmitterContext context, Operand value, OperandType type, bool signed)
         {

ARMeilleure/Instructions/InstEmitSimdHelper.cs

@@ -1,3 +1,4 @@
+using ARMeilleure.CodeGen.X86;
 using ARMeilleure.Decoders;
 using ARMeilleure.IntermediateRepresentation;
 using ARMeilleure.State;
@@ -158,6 +159,75 @@ namespace ARMeilleure.Instructions
         };
 #endregion
 
+        public static void EnterArmFpMode(EmitterContext context, Func<FPState, Operand> getFpFlag)
+        {
+            if (Optimizations.UseSse2)
+            {
+                Operand mxcsr = context.AddIntrinsicInt(Intrinsic.X86Stmxcsr);
+
+                Operand fzTrue = getFpFlag(FPState.FzFlag);
+                Operand r0True = getFpFlag(FPState.RMode0Flag);
+                Operand r1True = getFpFlag(FPState.RMode1Flag);
+
+                mxcsr = context.BitwiseAnd(mxcsr, Const(~(int)(Mxcsr.Ftz | Mxcsr.Daz | Mxcsr.Rhi | Mxcsr.Rlo)));
+
+                mxcsr = context.BitwiseOr(mxcsr, context.ConditionalSelect(fzTrue, Const((int)(Mxcsr.Ftz | Mxcsr.Daz | Mxcsr.Um | Mxcsr.Dm)), Const(0)));
+
+                // X86 round modes in order: nearest, negative, positive, zero
+                // ARM round modes in order: nearest, positive, negative, zero
+                // Read the bits backwards to correct this.
+
+                mxcsr = context.BitwiseOr(mxcsr, context.ConditionalSelect(r0True, Const((int)Mxcsr.Rhi), Const(0)));
+                mxcsr = context.BitwiseOr(mxcsr, context.ConditionalSelect(r1True, Const((int)Mxcsr.Rlo), Const(0)));
+
+                context.AddIntrinsicNoRet(Intrinsic.X86Ldmxcsr, mxcsr);
+            }
+            else if (Optimizations.UseAdvSimd)
+            {
+                Operand fpcr = context.AddIntrinsicInt(Intrinsic.Arm64MrsFpcr);
+
+                Operand fzTrue = getFpFlag(FPState.FzFlag);
+                Operand r0True = getFpFlag(FPState.RMode0Flag);
+                Operand r1True = getFpFlag(FPState.RMode1Flag);
+
+                fpcr = context.BitwiseAnd(fpcr, Const(~(int)(FPCR.Fz | FPCR.RMode0 | FPCR.RMode1)));
+
+                fpcr = context.BitwiseOr(fpcr, context.ConditionalSelect(fzTrue, Const((int)FPCR.Fz), Const(0)));
+                fpcr = context.BitwiseOr(fpcr, context.ConditionalSelect(r0True, Const((int)FPCR.RMode0), Const(0)));
+                fpcr = context.BitwiseOr(fpcr, context.ConditionalSelect(r1True, Const((int)FPCR.RMode1), Const(0)));
+
+                context.AddIntrinsicNoRet(Intrinsic.Arm64MsrFpcr, fpcr);
+
+                // TODO: Restore FPSR
+            }
+        }
+
+        public static void ExitArmFpMode(EmitterContext context, Action<FPState, Operand> setFpFlag)
+        {
+            if (Optimizations.UseSse2)
+            {
+                Operand mxcsr = context.AddIntrinsicInt(Intrinsic.X86Stmxcsr);
+
+                // Unset round mode (to nearest) and ftz.
+                mxcsr = context.BitwiseAnd(mxcsr, Const(~(int)(Mxcsr.Ftz | Mxcsr.Daz | Mxcsr.Rhi | Mxcsr.Rlo)));
+
+                context.AddIntrinsicNoRet(Intrinsic.X86Ldmxcsr, mxcsr);
+
+                // Status flags would be stored here if they were used.
+            }
+            else if (Optimizations.UseAdvSimd)
+            {
+                Operand fpcr = context.AddIntrinsicInt(Intrinsic.Arm64MrsFpcr);
+
+                // Unset round mode (to nearest) and fz.
+                fpcr = context.BitwiseAnd(fpcr, Const(~(int)(FPCR.Fz | FPCR.RMode0 | FPCR.RMode1)));
+
+                context.AddIntrinsicNoRet(Intrinsic.Arm64MsrFpcr, fpcr);
+
+                // TODO: Store FPSR
+            }
+        }
+
         public static int GetImmShl(OpCodeSimdShImm op)
         {
             return op.Imm - (8 << op.Size);
@@ -465,9 +535,11 @@ namespace ARMeilleure.Instructions
                 ? typeof(SoftFloat32).GetMethod(name)
                 : typeof(SoftFloat64).GetMethod(name);
 
+            context.ExitArmFpMode();
             context.StoreToContext();
             Operand res = context.Call(info, callArgs);
             context.LoadFromContext();
+            context.EnterArmFpMode();
 
             return res;
         }
@@ -1358,39 +1430,6 @@ namespace ARMeilleure.Instructions
             }
         }
 
-        [Flags]
-        public enum Mxcsr
-        {
-            Ftz = 1 << 15, // Flush To Zero.
-            Um  = 1 << 11, // Underflow Mask.
-            Dm  = 1 << 8,  // Denormal Mask.
-            Daz = 1 << 6   // Denormals Are Zero.
-        }
-
-        public static void EmitSseOrAvxEnterFtzAndDazModesOpF(ArmEmitterContext context, out Operand isTrue)
-        {
-            isTrue = GetFpFlag(FPState.FzFlag);
-
-            Operand lblTrue = Label();
-            context.BranchIfFalse(lblTrue, isTrue);
-
-            context.AddIntrinsicNoRet(Intrinsic.X86Mxcsrmb, Const((int)(Mxcsr.Ftz | Mxcsr.Um | Mxcsr.Dm | Mxcsr.Daz)));
-
-            context.MarkLabel(lblTrue);
-        }
-
-        public static void EmitSseOrAvxExitFtzAndDazModesOpF(ArmEmitterContext context, Operand isTrue = default)
-        {
-            isTrue = isTrue == default ? GetFpFlag(FPState.FzFlag) : isTrue;
-
-            Operand lblTrue = Label();
-            context.BranchIfFalse(lblTrue, isTrue);
-
-            context.AddIntrinsicNoRet(Intrinsic.X86Mxcsrub, Const((int)(Mxcsr.Ftz | Mxcsr.Daz)));
-
-            context.MarkLabel(lblTrue);
-        }
-
         public enum CmpCondition
         {
             // Legacy Sse.

ARMeilleure/Instructions/InstEmitSimdHelper32.cs

@@ -1197,9 +1197,11 @@ namespace ARMeilleure.Instructions
             Array.Resize(ref callArgs, callArgs.Length + 1);
             callArgs[callArgs.Length - 1] = Const(1);
 
+            context.ExitArmFpMode();
             context.StoreToContext();
             Operand res = context.Call(info, callArgs);
             context.LoadFromContext();
+            context.EnterArmFpMode();
 
             return res;
         }

ARMeilleure/Instructions/InstEmitSystem.cs

@@ -192,6 +192,8 @@ namespace ARMeilleure.Instructions
                     SetFpFlag(context, (FPState)flag, context.BitwiseAnd(context.ShiftRightUI(fpcr, Const(flag)), Const(1)));
                 }
             }
+
+            context.UpdateArmFpMode();
         }
 
         private static void EmitSetFpsr(ArmEmitterContext context)
@@ -210,6 +212,8 @@ namespace ARMeilleure.Instructions
                     SetFpFlag(context, (FPState)flag, context.BitwiseAnd(context.ShiftRightUI(fpsr, Const(flag)), Const(1)));
                 }
             }
+
+            context.UpdateArmFpMode();
         }
     }
 }

ARMeilleure/Instructions/InstEmitSystem32.cs

@@ -321,6 +321,8 @@ namespace ARMeilleure.Instructions
                     SetFpFlag(context, (FPState)flag, context.BitwiseAnd(context.ShiftRightUI(fpscr, Const(flag)), Const(1)));
                 }
             }
+
+            context.UpdateArmFpMode();
         }
     }
 }

ARMeilleure/IntermediateRepresentation/Intrinsic.cs

@@ -53,6 +53,7 @@ namespace ARMeilleure.IntermediateRepresentation
         X86Haddpd,
         X86Haddps,
         X86Insertps,
+        X86Ldmxcsr,
         X86Maxpd,
         X86Maxps,
         X86Maxsd,
@@ -68,8 +69,6 @@ namespace ARMeilleure.IntermediateRepresentation
         X86Mulps,
         X86Mulsd,
         X86Mulss,
-        X86Mxcsrmb,
-        X86Mxcsrub,
         X86Paddb,
         X86Paddd,
         X86Paddq,
@@ -153,6 +152,7 @@ namespace ARMeilleure.IntermediateRepresentation
         X86Sqrtps,
         X86Sqrtsd,
         X86Sqrtss,
+        X86Stmxcsr,
         X86Subpd,
         X86Subps,
         X86Subsd,
@@ -163,11 +163,13 @@ namespace ARMeilleure.IntermediateRepresentation
         X86Unpcklps,
         X86Vcvtph2ps,
         X86Vcvtps2ph,
+        X86Vfmadd231pd,
         X86Vfmadd231ps,
         X86Vfmadd231sd,
         X86Vfmadd231ss,
         X86Vfmsub231sd,
         X86Vfmsub231ss,
+        X86Vfnmadd231pd,
         X86Vfnmadd231ps,
         X86Vfnmadd231sd,
         X86Vfnmadd231ss,
@@ -394,6 +396,8 @@ namespace ARMeilleure.IntermediateRepresentation
         Arm64MlsVe,
         Arm64MlsV,
         Arm64MoviV,
+        Arm64MrsFpcr,
+        Arm64MsrFpcr,
         Arm64MrsFpsr,
         Arm64MsrFpsr,
         Arm64MulVe,

ARMeilleure/Translation/ArmEmitterContext.cs

@@ -188,6 +188,21 @@ namespace ARMeilleure.Translation
             }
         }
 
+        public void EnterArmFpMode()
+        {
+            InstEmitSimdHelper.EnterArmFpMode(this, InstEmitHelper.GetFpFlag);
+        }
+
+        public void UpdateArmFpMode()
+        {
+            EnterArmFpMode();
+        }
+
+        public void ExitArmFpMode()
+        {
+            InstEmitSimdHelper.ExitArmFpMode(this, (flag, value) => InstEmitHelper.SetFpFlag(this, flag, value));
+        }
+
         public Operand TryGetComparisonResult(Condition condition)
         {
             if (_optOpLastCompare == null || _optOpLastCompare != _optOpLastFlagSet)

ARMeilleure/Translation/DispatcherFunction.cs

@@ -3,4 +3,5 @@
 namespace ARMeilleure.Translation
 {
     delegate void DispatcherFunction(IntPtr nativeContext, ulong startAddress);
+    delegate ulong WrapperFunction(IntPtr nativeContext, ulong startAddress);
 }

ARMeilleure/Translation/PTC/Ptc.cs

@@ -30,7 +30,7 @@ namespace ARMeilleure.Translation.PTC
         private const string OuterHeaderMagicString = "PTCohd\0\0";
         private const string InnerHeaderMagicString = "PTCihd\0\0";
 
-        private const uint InternalVersion = 4485; //! To be incremented manually for each change to the ARMeilleure project.
+        private const uint InternalVersion = 4626; //! To be incremented manually for each change to the ARMeilleure project.
 
         private const string ActualDir = "0";
         private const string BackupDir = "1";

ARMeilleure/Translation/TranslatedFunction.cs

@@ -25,5 +25,10 @@ namespace ARMeilleure.Translation
         {
             return _func(context.NativeContextPtr);
         }
+
+        public ulong Execute(WrapperFunction dispatcher, State.ExecutionContext context)
+        {
+            return dispatcher(context.NativeContextPtr, (ulong)FuncPointer);
+        }
     }
 }

ARMeilleure/Translation/Translator.cs

@@ -183,7 +183,7 @@ namespace ARMeilleure.Translation
 
             Statistics.StartTimer();
 
-            ulong nextAddr = func.Execute(context);
+            ulong nextAddr = func.Execute(Stubs.ContextWrapper, context);
 
             Statistics.StopTimer(address);
 
@@ -194,7 +194,7 @@ namespace ARMeilleure.Translation
         {
             TranslatedFunction func = Translate(address, context.ExecutionMode, highCq: false, singleStep: true);
 
-            address = func.Execute(context);
+            address = func.Execute(Stubs.ContextWrapper, context);
 
             EnqueueForDeletion(address, func);
 

ARMeilleure/Translation/TranslatorStubs.cs

@@ -21,6 +21,7 @@ namespace ARMeilleure.Translation
         private readonly Translator _translator;
         private readonly Lazy<IntPtr> _dispatchStub;
         private readonly Lazy<DispatcherFunction> _dispatchLoop;
+        private readonly Lazy<WrapperFunction> _contextWrapper;
 
         /// <summary>
         /// Gets the dispatch stub.
@@ -64,6 +65,20 @@ namespace ARMeilleure.Translation
             }
         }
 
+        /// <summary>
+        /// Gets the context wrapper function.
+        /// </summary>
+        /// <exception cref="ObjectDisposedException"><see cref="TranslatorStubs"/> instance was disposed</exception>
+        public WrapperFunction ContextWrapper
+        {
+            get
+            {
+                ObjectDisposedException.ThrowIf(_disposed, this);
+
+                return _contextWrapper.Value;
+            }
+        }
+
         /// <summary>
         /// Initializes a new instance of the <see cref="TranslatorStubs"/> class with the specified
         /// <see cref="Translator"/> instance.
@@ -77,6 +92,7 @@ namespace ARMeilleure.Translation
             _translator = translator;
             _dispatchStub = new(GenerateDispatchStub, isThreadSafe: true);
             _dispatchLoop = new(GenerateDispatchLoop, isThreadSafe: true);
+            _contextWrapper = new(GenerateContextWrapper, isThreadSafe: true);
         }
 
         /// <summary>
@@ -202,6 +218,32 @@ namespace ARMeilleure.Translation
             return Marshal.GetFunctionPointerForDelegate(func);
         }
 
+        /// <summary>
+        /// Emits code that syncs FP state before executing guest code, or returns it to normal.
+        /// </summary>
+        /// <param name="context">Emitter context for the method</param>
+        /// <param name="nativeContext">Pointer to the native context</param>
+        /// <param name="enter">True if entering guest code, false otherwise</param>
+        private void EmitSyncFpContext(EmitterContext context, Operand nativeContext, bool enter)
+        {
+            if (enter)
+            {
+                InstEmitSimdHelper.EnterArmFpMode(context, (flag) =>
+                {
+                    Operand flagAddress = context.Add(nativeContext, Const((ulong)NativeContext.GetRegisterOffset(new Register((int)flag, RegisterType.FpFlag))));
+                    return context.Load(OperandType.I32, flagAddress);
+                });
+            }
+            else
+            {
+                InstEmitSimdHelper.ExitArmFpMode(context, (flag, value) =>
+                {
+                    Operand flagAddress = context.Add(nativeContext, Const((ulong)NativeContext.GetRegisterOffset(new Register((int)flag, RegisterType.FpFlag))));
+                    context.Store(flagAddress, value);
+                });
+            }
+        }
+
         /// <summary>
         /// Generates a <see cref="DispatchLoop"/> function.
         /// </summary>
@@ -221,6 +263,8 @@ namespace ARMeilleure.Translation
             Operand runningAddress = context.Add(nativeContext, Const((ulong)NativeContext.GetRunningOffset()));
             Operand dispatchAddress = context.Add(nativeContext, Const((ulong)NativeContext.GetDispatchAddressOffset()));
 
+            EmitSyncFpContext(context, nativeContext, true);
+
             context.MarkLabel(beginLbl);
             context.Store(dispatchAddress, guestAddress);
             context.Copy(guestAddress, context.Call(Const((ulong)DispatchStub), OperandType.I64, nativeContext));
@@ -229,6 +273,9 @@ namespace ARMeilleure.Translation
             context.Branch(beginLbl);
 
             context.MarkLabel(endLbl);
+
+            EmitSyncFpContext(context, nativeContext, false);
+
             context.Return();
 
             var cfg = context.GetControlFlowGraph();
@@ -237,5 +284,29 @@ namespace ARMeilleure.Translation
 
             return Compiler.Compile(cfg, argTypes, retType, CompilerOptions.HighCq, RuntimeInformation.ProcessArchitecture).Map<DispatcherFunction>();
         }
+
+        /// <summary>
+        /// Generates a <see cref="ContextWrapper"/> function.
+        /// </summary>
+        /// <returns><see cref="ContextWrapper"/> function</returns>
+        private WrapperFunction GenerateContextWrapper()
+        {
+            var context = new EmitterContext();
+
+            Operand nativeContext = context.LoadArgument(OperandType.I64, 0);
+            Operand guestMethod = context.LoadArgument(OperandType.I64, 1);
+
+            EmitSyncFpContext(context, nativeContext, true);
+            Operand returnValue = context.Call(guestMethod, OperandType.I64, nativeContext);
+            EmitSyncFpContext(context, nativeContext, false);
+
+            context.Return(returnValue);
+
+            var cfg = context.GetControlFlowGraph();
+            var retType = OperandType.I64;
+            var argTypes = new[] { OperandType.I64, OperandType.I64 };
+
+            return Compiler.Compile(cfg, argTypes, retType, CompilerOptions.HighCq, RuntimeInformation.ProcessArchitecture).Map<WrapperFunction>();
+        }
     }
 }

ARMeilleure/Translation/TranslatorTestMethods.cs

@@ -0,0 +1,148 @@
+using ARMeilleure.CodeGen.X86;
+using ARMeilleure.IntermediateRepresentation;
+using ARMeilleure.State;
+using ARMeilleure.Translation;
+using System;
+using System.Runtime.InteropServices;
+using static ARMeilleure.IntermediateRepresentation.Operand.Factory;
+
+namespace ARMeilleure.Translation
+{
+    public static class TranslatorTestMethods
+    {
+        public delegate int FpFlagsPInvokeTest(IntPtr managedMethod);
+
+        private static bool SetPlatformFtz(EmitterContext context, bool ftz)
+        {
+            if (Optimizations.UseSse2)
+            {
+                Operand mxcsr = context.AddIntrinsicInt(Intrinsic.X86Stmxcsr);
+
+                if (ftz)
+                {
+                    mxcsr = context.BitwiseOr(mxcsr, Const((int)(Mxcsr.Ftz | Mxcsr.Um | Mxcsr.Dm)));
+                }
+                else
+                {
+                    mxcsr = context.BitwiseAnd(mxcsr, Const(~(int)Mxcsr.Ftz));
+                }
+
+                context.AddIntrinsicNoRet(Intrinsic.X86Ldmxcsr, mxcsr);
+
+                return true;
+            }
+            else if (Optimizations.UseAdvSimd)
+            {
+                Operand fpcr = context.AddIntrinsicInt(Intrinsic.Arm64MrsFpcr);
+
+                if (ftz)
+                {
+                    fpcr = context.BitwiseOr(fpcr, Const((int)FPCR.Fz));
+                }
+                else
+                {
+                    fpcr = context.BitwiseAnd(fpcr, Const(~(int)FPCR.Fz));
+                }
+
+                context.AddIntrinsicNoRet(Intrinsic.Arm64MsrFpcr, fpcr);
+
+                return true;
+            }
+            else
+            {
+                return false;
+            }
+        }
+
+        private static Operand FpBitsToInt(EmitterContext context, Operand fp)
+        {
+            Operand vec = context.VectorInsert(context.VectorZero(), fp, 0);
+            return context.VectorExtract(OperandType.I32, vec, 0);
+        }
+
+        public static FpFlagsPInvokeTest GenerateFpFlagsPInvokeTest()
+        {
+            EmitterContext context = new EmitterContext();
+
+            Operand methodAddress = context.Copy(context.LoadArgument(OperandType.I64, 0));
+
+            // Verify that default dotnet fp state does not flush to zero.
+            // This is required for SoftFloat to function.
+
+            // Denormal + zero != 0
+
+            Operand denormal = ConstF(BitConverter.Int32BitsToSingle(1)); // 1.40129846432e-45
+            Operand zeroF = ConstF(0f);
+            Operand zero = Const(0);
+
+            Operand result = context.Add(zeroF, denormal);
+
+            // Must not be zero.
+
+            Operand correct1Label = Label();
+
+            context.BranchIfFalse(correct1Label, context.ICompareEqual(FpBitsToInt(context, result), zero));
+
+            context.Return(Const(1));
+
+            context.MarkLabel(correct1Label);
+
+            // Set flush to zero flag. If unsupported by the backend, just return true.
+
+            if (!SetPlatformFtz(context, true))
+            {
+                context.Return(Const(0));
+            }
+
+            // Denormal + zero == 0
+
+            Operand resultFz = context.Add(zeroF, denormal);
+
+            // Must equal zero.
+
+            Operand correct2Label = Label();
+
+            context.BranchIfTrue(correct2Label, context.ICompareEqual(FpBitsToInt(context, resultFz), zero));
+
+            SetPlatformFtz(context, false);
+
+            context.Return(Const(2));
+
+            context.MarkLabel(correct2Label);
+
+            // Call a managed method. This method should not change Fz state. 
+
+            context.Call(methodAddress, OperandType.None);
+
+            // Denormal + zero == 0
+
+            Operand resultFz2 = context.Add(zeroF, denormal);
+
+            // Must equal zero.
+
+            Operand correct3Label = Label();
+
+            context.BranchIfTrue(correct3Label, context.ICompareEqual(FpBitsToInt(context, resultFz2), zero));
+
+            SetPlatformFtz(context, false);
+
+            context.Return(Const(3));
+
+            context.MarkLabel(correct3Label);
+
+            // Success.
+
+            SetPlatformFtz(context, false);
+
+            context.Return(Const(0));
+
+            // Compile and return the function.
+
+            ControlFlowGraph cfg = context.GetControlFlowGraph();
+
+            OperandType[] argTypes = new OperandType[] { OperandType.I64 };
+
+            return Compiler.Compile(cfg, argTypes, OperandType.I32, CompilerOptions.HighCq, RuntimeInformation.ProcessArchitecture).Map<FpFlagsPInvokeTest>();
+        }
+    }
+}

Ryujinx.Ava/UI/Applet/AvaHostUiHandler.cs

@@ -53,6 +53,8 @@ namespace Ryujinx.Ava.UI.Applet
 
                     bool opened = false;
 
+                    _parent.Activate();
+
                     UserResult response = await ContentDialogHelper.ShowDeferredContentDialog(_parent,
                        title,
                        message,

Ryujinx.Graphics.OpenGL/Window.cs

@@ -226,6 +226,7 @@ namespace Ryujinx.Graphics.OpenGL
             // Set clip control, viewport and the framebuffer to the output to placate overlays and OBS capture.
             GL.ClipControl(ClipOrigin.LowerLeft, ClipDepthMode.NegativeOneToOne);
             GL.Viewport(0, 0, _width, _height);
+            GL.BindFramebuffer(FramebufferTarget.Framebuffer, drawFramebuffer);
 
             swapBuffersCallback();
 

Ryujinx.Graphics.Vulkan/MemoryAllocator.cs

@@ -9,21 +9,18 @@ namespace Ryujinx.Graphics.Vulkan
         private ulong MaxDeviceMemoryUsageEstimate = 16UL * 1024 * 1024 * 1024;
 
         private readonly Vk _api;
-        private readonly PhysicalDevice _physicalDevice;
+        private readonly VulkanPhysicalDevice _physicalDevice;
         private readonly Device _device;
         private readonly List<MemoryAllocatorBlockList> _blockLists;
         private readonly int _blockAlignment;
-        private readonly PhysicalDeviceMemoryProperties _physicalDeviceMemoryProperties;
 
-        public MemoryAllocator(Vk api, PhysicalDevice physicalDevice, Device device, uint maxMemoryAllocationCount)
+        public MemoryAllocator(Vk api, VulkanPhysicalDevice physicalDevice, Device device)
         {
             _api = api;
             _physicalDevice = physicalDevice;
             _device = device;
             _blockLists = new List<MemoryAllocatorBlockList>();
-            _blockAlignment = (int)Math.Min(int.MaxValue, MaxDeviceMemoryUsageEstimate / (ulong)maxMemoryAllocationCount);
+            _blockAlignment = (int)Math.Min(int.MaxValue, MaxDeviceMemoryUsageEstimate / (ulong)_physicalDevice.PhysicalDeviceProperties.Limits.MaxMemoryAllocationCount);
-
-            _api.GetPhysicalDeviceMemoryProperties(_physicalDevice, out _physicalDeviceMemoryProperties);
         }
 
         public MemoryAllocation AllocateDeviceMemory(
@@ -64,9 +61,9 @@ namespace Ryujinx.Graphics.Vulkan
             uint memoryTypeBits,
             MemoryPropertyFlags flags)
         {
-            for (int i = 0; i < _physicalDeviceMemoryProperties.MemoryTypeCount; i++)
+            for (int i = 0; i < _physicalDevice.PhysicalDeviceMemoryProperties.MemoryTypeCount; i++)
             {
-                var type = _physicalDeviceMemoryProperties.MemoryTypes[i];
+                var type = _physicalDevice.PhysicalDeviceMemoryProperties.MemoryTypes[i];
 
                 if ((memoryTypeBits & (1 << i)) != 0)
                 {
@@ -80,15 +77,11 @@ namespace Ryujinx.Graphics.Vulkan
             return -1;
         }
 
-        public static bool IsDeviceMemoryShared(Vk api, PhysicalDevice physicalDevice)
+        public static bool IsDeviceMemoryShared(VulkanPhysicalDevice physicalDevice)
         {
-            // The device is regarded as having shared memory if all heaps have the device local bit.
+            for (int i = 0; i < physicalDevice.PhysicalDeviceMemoryProperties.MemoryHeapCount; i++)
-
-            api.GetPhysicalDeviceMemoryProperties(physicalDevice, out var properties);
-
-            for (int i = 0; i < properties.MemoryHeapCount; i++)
             {
-                if (!properties.MemoryHeaps[i].Flags.HasFlag(MemoryHeapFlags.DeviceLocalBit))
+                if (!physicalDevice.PhysicalDeviceMemoryProperties.MemoryHeaps[i].Flags.HasFlag(MemoryHeapFlags.DeviceLocalBit))
                 {
                     return false;
                 }

Ryujinx.Graphics.Vulkan/VulkanInitialization.cs

@@ -47,36 +47,24 @@ namespace Ryujinx.Graphics.Vulkan
             KhrSwapchain.ExtensionName
         };
 
-        internal static Instance CreateInstance(Vk api, GraphicsDebugLevel logLevel, string[] requiredExtensions)
+        internal static VulkanInstance CreateInstance(Vk api, GraphicsDebugLevel logLevel, string[] requiredExtensions)
         {
             var enabledLayers = new List<string>();
 
+            var instanceExtensions = VulkanInstance.GetInstanceExtensions(api);
+            var instanceLayers = VulkanInstance.GetInstanceLayers(api);
+
             void AddAvailableLayer(string layerName)
             {
-                uint layerPropertiesCount;
+                if (instanceLayers.Contains(layerName))
-
-                api.EnumerateInstanceLayerProperties(&layerPropertiesCount, null).ThrowOnError();
-
-                LayerProperties[] layerProperties = new LayerProperties[layerPropertiesCount];
-
-                fixed (LayerProperties* pLayerProperties = layerProperties)
-                {
-                    api.EnumerateInstanceLayerProperties(&layerPropertiesCount, layerProperties).ThrowOnError();
-
-                    for (int i = 0; i < layerPropertiesCount; i++)
-                    {
-                        string currentLayerName = Marshal.PtrToStringAnsi((IntPtr)pLayerProperties[i].LayerName);
-
-                        if (currentLayerName == layerName)
                 {
                     enabledLayers.Add(layerName);
-                            return;
                 }
-                    }
+                else
-                }
+                {
-
                     Logger.Warning?.Print(LogClass.Gpu, $"Missing layer {layerName}");
                 }
+            }
 
             if (logLevel != GraphicsDebugLevel.None)
             {
@@ -85,7 +73,7 @@ namespace Ryujinx.Graphics.Vulkan
 
             var enabledExtensions = requiredExtensions;
 
-            if (api.IsInstanceExtensionPresent("VK_EXT_debug_utils"))
+            if (instanceExtensions.Contains("VK_EXT_debug_utils"))
             {
                 enabledExtensions = enabledExtensions.Append(ExtDebugUtils.ExtensionName).ToArray();
             }
@@ -124,7 +112,7 @@ namespace Ryujinx.Graphics.Vulkan
                 EnabledLayerCount = (uint)enabledLayers.Count
             };
 
-            api.CreateInstance(in instanceCreateInfo, null, out var instance).ThrowOnError();
+            Result result = VulkanInstance.Create(api, ref instanceCreateInfo, out var instance);
 
             Marshal.FreeHGlobal(appName);
 
@@ -138,21 +126,14 @@ namespace Ryujinx.Graphics.Vulkan
                 Marshal.FreeHGlobal(ppEnabledLayers[i]);
             }
 
+            result.ThrowOnError();
+
             return instance;
         }
 
-        internal static PhysicalDevice FindSuitablePhysicalDevice(Vk api, Instance instance, SurfaceKHR surface, string preferredGpuId)
+        internal static VulkanPhysicalDevice FindSuitablePhysicalDevice(Vk api, VulkanInstance instance, SurfaceKHR surface, string preferredGpuId)
         {
-            uint physicalDeviceCount;
+            instance.EnumeratePhysicalDevices(out var physicalDevices).ThrowOnError();
-
-            api.EnumeratePhysicalDevices(instance, &physicalDeviceCount, null).ThrowOnError();
-
-            PhysicalDevice[] physicalDevices = new PhysicalDevice[physicalDeviceCount];
-
-            fixed (PhysicalDevice* pPhysicalDevices = physicalDevices)
-            {
-                api.EnumeratePhysicalDevices(instance, &physicalDeviceCount, pPhysicalDevices).ThrowOnError();
-            }
 
             // First we try to pick the the user preferred GPU.
             for (int i = 0; i < physicalDevices.Length; i++)
@@ -198,76 +179,41 @@ namespace Ryujinx.Graphics.Vulkan
                 EnabledLayerCount = 0
             };
 
-            api.CreateInstance(in instanceCreateInfo, null, out var instance).ThrowOnError();
+            Result result = VulkanInstance.Create(api, ref instanceCreateInfo, out var rawInstance);
-
-            // We ensure that vkEnumerateInstanceVersion is present (added in 1.1).
-            // If the instance doesn't support it, no device is going to be 1.1 compatible.
-            if (api.GetInstanceProcAddr(instance, "vkEnumerateInstanceVersion") == IntPtr.Zero)
-            {
-                api.DestroyInstance(instance, null);
-
-                return Array.Empty<DeviceInfo>();
-            }
-
-            // We currently assume that the instance is compatible with Vulkan 1.2
-            // TODO: Remove this once we relax our initialization codepaths.
-            uint instanceApiVerison = 0;
-            api.EnumerateInstanceVersion(ref instanceApiVerison).ThrowOnError();
-
-            if (instanceApiVerison < MinimalInstanceVulkanVersion)
-            {
-                api.DestroyInstance(instance, null);
-
-                return Array.Empty<DeviceInfo>();
-            }
 
             Marshal.FreeHGlobal(appName);
 
-            uint physicalDeviceCount;
+            result.ThrowOnError();
 
-            api.EnumeratePhysicalDevices(instance, &physicalDeviceCount, null).ThrowOnError();
+            using VulkanInstance instance = rawInstance;
 
-            PhysicalDevice[] physicalDevices = new PhysicalDevice[physicalDeviceCount];
+            // We currently assume that the instance is compatible with Vulkan 1.2
-
+            // TODO: Remove this once we relax our initialization codepaths.
-            fixed (PhysicalDevice* pPhysicalDevices = physicalDevices)
+            if (instance.InstanceVersion < MinimalInstanceVulkanVersion)
             {
-                api.EnumeratePhysicalDevices(instance, &physicalDeviceCount, pPhysicalDevices).ThrowOnError();
+                return Array.Empty<DeviceInfo>();
             }
 
-            DeviceInfo[] devices = new DeviceInfo[physicalDevices.Length];
+            instance.EnumeratePhysicalDevices(out VulkanPhysicalDevice[] physicalDevices).ThrowOnError();
 
-            for (int i = 0; i < physicalDevices.Length; i++)
+            List<DeviceInfo> deviceInfos = new List<DeviceInfo>();
+
+            foreach (VulkanPhysicalDevice physicalDevice in physicalDevices)
             {
-                var physicalDevice = physicalDevices[i];
+                if (physicalDevice.PhysicalDeviceProperties.ApiVersion < MinimalVulkanVersion)
-                api.GetPhysicalDeviceProperties(physicalDevice, out var properties);
-
-                if (properties.ApiVersion < MinimalVulkanVersion)
                 {
                     continue;
                 }
 
-                devices[i] = new DeviceInfo(
+                deviceInfos.Add(physicalDevice.ToDeviceInfo());
-                    StringFromIdPair(properties.VendorID, properties.DeviceID),
-                    VendorUtils.GetNameFromId(properties.VendorID),
-                    Marshal.PtrToStringAnsi((IntPtr)properties.DeviceName),
-                    properties.DeviceType == PhysicalDeviceType.DiscreteGpu);
             }
 
-            api.DestroyInstance(instance, null);
+            return deviceInfos.ToArray();
-
-            return devices;
         }
 
-        public static string StringFromIdPair(uint vendorId, uint deviceId)
+        private static bool IsPreferredAndSuitableDevice(Vk api, VulkanPhysicalDevice physicalDevice, SurfaceKHR surface, string preferredGpuId)
         {
-            return $"0x{vendorId:X}_0x{deviceId:X}";
+            if (physicalDevice.Id != preferredGpuId)
-        }
-
-        private static bool IsPreferredAndSuitableDevice(Vk api, PhysicalDevice physicalDevice, SurfaceKHR surface, string preferredGpuId)
-        {
-            api.GetPhysicalDeviceProperties(physicalDevice, out var properties);
-
-            if (StringFromIdPair(properties.VendorID, properties.DeviceID) != preferredGpuId)
             {
                 return false;
             }
@@ -275,68 +221,47 @@ namespace Ryujinx.Graphics.Vulkan
             return IsSuitableDevice(api, physicalDevice, surface);
         }
 
-        private static bool IsSuitableDevice(Vk api, PhysicalDevice physicalDevice, SurfaceKHR surface)
+        private static bool IsSuitableDevice(Vk api, VulkanPhysicalDevice physicalDevice, SurfaceKHR surface)
         {
             int extensionMatches = 0;
-            uint propertiesCount;
 
-            api.EnumerateDeviceExtensionProperties(physicalDevice, (byte*)null, &propertiesCount, null).ThrowOnError();
+            foreach (string requiredExtension in _requiredExtensions)
-
-            ExtensionProperties[] extensionProperties = new ExtensionProperties[propertiesCount];
-
-            fixed (ExtensionProperties* pExtensionProperties = extensionProperties)
             {
-                api.EnumerateDeviceExtensionProperties(physicalDevice, (byte*)null, &propertiesCount, pExtensionProperties).ThrowOnError();
+                if (physicalDevice.IsDeviceExtensionPresent(requiredExtension))
-
-                for (int i = 0; i < propertiesCount; i++)
-                {
-                    string extensionName = Marshal.PtrToStringAnsi((IntPtr)pExtensionProperties[i].ExtensionName);
-
-                    if (_requiredExtensions.Contains(extensionName))
                 {
                     extensionMatches++;
                 }
             }
-            }
 
             return extensionMatches == _requiredExtensions.Length && FindSuitableQueueFamily(api, physicalDevice, surface, out _) != InvalidIndex;
         }
 
-        internal static uint FindSuitableQueueFamily(Vk api, PhysicalDevice physicalDevice, SurfaceKHR surface, out uint queueCount)
+        internal static uint FindSuitableQueueFamily(Vk api, VulkanPhysicalDevice physicalDevice, SurfaceKHR surface, out uint queueCount)
         {
             const QueueFlags RequiredFlags = QueueFlags.GraphicsBit | QueueFlags.ComputeBit;
 
             var khrSurface = new KhrSurface(api.Context);
 
-            uint propertiesCount;
+            for (uint index = 0; index < physicalDevice.QueueFamilyProperties.Length; index++)
-
-            api.GetPhysicalDeviceQueueFamilyProperties(physicalDevice, &propertiesCount, null);
-
-            QueueFamilyProperties[] properties = new QueueFamilyProperties[propertiesCount];
-
-            fixed (QueueFamilyProperties* pProperties = properties)
             {
-                api.GetPhysicalDeviceQueueFamilyProperties(physicalDevice, &propertiesCount, pProperties);
+                ref QueueFamilyProperties property = ref physicalDevice.QueueFamilyProperties[index];
-            }
 
-            for (uint index = 0; index < propertiesCount; index++)
+                khrSurface.GetPhysicalDeviceSurfaceSupport(physicalDevice.PhysicalDevice, index, surface, out var surfaceSupported).ThrowOnError();
+
+                if (property.QueueFlags.HasFlag(RequiredFlags) && surfaceSupported)
                 {
-                var queueFlags = properties[index].QueueFlags;
+                    queueCount = property.QueueCount;
 
-                khrSurface.GetPhysicalDeviceSurfaceSupport(physicalDevice, index, surface, out var surfaceSupported).ThrowOnError();
-
-                if (queueFlags.HasFlag(RequiredFlags) && surfaceSupported)
-                {
-                    queueCount = properties[index].QueueCount;
                     return index;
                 }
             }
 
             queueCount = 0;
+
             return InvalidIndex;
         }
 
-        public static Device CreateDevice(Vk api, PhysicalDevice physicalDevice, uint queueFamilyIndex, string[] supportedExtensions, uint queueCount)
+        internal static Device CreateDevice(Vk api, VulkanPhysicalDevice physicalDevice, uint queueFamilyIndex, uint queueCount)
         {
             if (queueCount > QueuesCount)
             {
@@ -358,8 +283,7 @@ namespace Ryujinx.Graphics.Vulkan
                 PQueuePriorities = queuePriorities
             };
 
-            api.GetPhysicalDeviceProperties(physicalDevice, out var properties);
+            bool useRobustBufferAccess = VendorUtils.FromId(physicalDevice.PhysicalDeviceProperties.VendorID) == Vendor.Nvidia;
-            bool useRobustBufferAccess = VendorUtils.FromId(properties.VendorID) == Vendor.Nvidia;
 
             PhysicalDeviceFeatures2 features2 = new PhysicalDeviceFeatures2()
             {
@@ -380,7 +304,7 @@ namespace Ryujinx.Graphics.Vulkan
                 PNext = features2.PNext
             };
 
-            if (supportedExtensions.Contains("VK_EXT_custom_border_color"))
+            if (physicalDevice.IsDeviceExtensionPresent("VK_EXT_custom_border_color"))
             {
                 features2.PNext = &supportedFeaturesCustomBorderColor;
             }
@@ -391,7 +315,7 @@ namespace Ryujinx.Graphics.Vulkan
                 PNext = features2.PNext
             };
 
-            if (supportedExtensions.Contains("VK_EXT_primitive_topology_list_restart"))
+            if (physicalDevice.IsDeviceExtensionPresent("VK_EXT_primitive_topology_list_restart"))
             {
                 features2.PNext = &supportedFeaturesPrimitiveTopologyListRestart;
             }
@@ -402,7 +326,7 @@ namespace Ryujinx.Graphics.Vulkan
                 PNext = features2.PNext
             };
 
-            if (supportedExtensions.Contains(ExtTransformFeedback.ExtensionName))
+            if (physicalDevice.IsDeviceExtensionPresent(ExtTransformFeedback.ExtensionName))
             {
                 features2.PNext = &supportedFeaturesTransformFeedback;
             }
@@ -412,14 +336,14 @@ namespace Ryujinx.Graphics.Vulkan
                 SType = StructureType.PhysicalDeviceRobustness2FeaturesExt
             };
 
-            if (supportedExtensions.Contains("VK_EXT_robustness2"))
+            if (physicalDevice.IsDeviceExtensionPresent("VK_EXT_robustness2"))
             {
                 supportedFeaturesRobustness2.PNext = features2.PNext;
 
                 features2.PNext = &supportedFeaturesRobustness2;
             }
 
-            api.GetPhysicalDeviceFeatures2(physicalDevice, &features2);
+            api.GetPhysicalDeviceFeatures2(physicalDevice.PhysicalDevice, &features2);
 
             var supportedFeatures = features2.Features;
 
@@ -452,7 +376,7 @@ namespace Ryujinx.Graphics.Vulkan
 
             PhysicalDeviceTransformFeedbackFeaturesEXT featuresTransformFeedback;
 
-            if (supportedExtensions.Contains(ExtTransformFeedback.ExtensionName))
+            if (physicalDevice.IsDeviceExtensionPresent(ExtTransformFeedback.ExtensionName))
             {
                 featuresTransformFeedback = new PhysicalDeviceTransformFeedbackFeaturesEXT()
                 {
@@ -466,7 +390,7 @@ namespace Ryujinx.Graphics.Vulkan
 
             PhysicalDevicePrimitiveTopologyListRestartFeaturesEXT featuresPrimitiveTopologyListRestart;
 
-            if (supportedExtensions.Contains("VK_EXT_primitive_topology_list_restart"))
+            if (physicalDevice.IsDeviceExtensionPresent("VK_EXT_primitive_topology_list_restart"))
             {
                 featuresPrimitiveTopologyListRestart = new PhysicalDevicePrimitiveTopologyListRestartFeaturesEXT()
                 {
@@ -481,7 +405,7 @@ namespace Ryujinx.Graphics.Vulkan
 
             PhysicalDeviceRobustness2FeaturesEXT featuresRobustness2;
 
-            if (supportedExtensions.Contains("VK_EXT_robustness2"))
+            if (physicalDevice.IsDeviceExtensionPresent("VK_EXT_robustness2"))
             {
                 featuresRobustness2 = new PhysicalDeviceRobustness2FeaturesEXT()
                 {
@@ -497,7 +421,7 @@ namespace Ryujinx.Graphics.Vulkan
             {
                 SType = StructureType.PhysicalDeviceExtendedDynamicStateFeaturesExt,
                 PNext = pExtendedFeatures,
-                ExtendedDynamicState = supportedExtensions.Contains(ExtExtendedDynamicState.ExtensionName)
+                ExtendedDynamicState = physicalDevice.IsDeviceExtensionPresent(ExtExtendedDynamicState.ExtensionName)
             };
 
             pExtendedFeatures = &featuresExtendedDynamicState;
@@ -515,16 +439,16 @@ namespace Ryujinx.Graphics.Vulkan
             {
                 SType = StructureType.PhysicalDeviceVulkan12Features,
                 PNext = pExtendedFeatures,
-                DescriptorIndexing = supportedExtensions.Contains("VK_EXT_descriptor_indexing"),
+                DescriptorIndexing = physicalDevice.IsDeviceExtensionPresent("VK_EXT_descriptor_indexing"),
-                DrawIndirectCount = supportedExtensions.Contains(KhrDrawIndirectCount.ExtensionName),
+                DrawIndirectCount = physicalDevice.IsDeviceExtensionPresent(KhrDrawIndirectCount.ExtensionName),
-                UniformBufferStandardLayout = supportedExtensions.Contains("VK_KHR_uniform_buffer_standard_layout")
+                UniformBufferStandardLayout = physicalDevice.IsDeviceExtensionPresent("VK_KHR_uniform_buffer_standard_layout")
             };
 
             pExtendedFeatures = &featuresVk12;
 
             PhysicalDeviceIndexTypeUint8FeaturesEXT featuresIndexU8;
 
-            if (supportedExtensions.Contains("VK_EXT_index_type_uint8"))
+            if (physicalDevice.IsDeviceExtensionPresent("VK_EXT_index_type_uint8"))
             {
                 featuresIndexU8 = new PhysicalDeviceIndexTypeUint8FeaturesEXT()
                 {
@@ -538,7 +462,7 @@ namespace Ryujinx.Graphics.Vulkan
 
             PhysicalDeviceFragmentShaderInterlockFeaturesEXT featuresFragmentShaderInterlock;
 
-            if (supportedExtensions.Contains("VK_EXT_fragment_shader_interlock"))
+            if (physicalDevice.IsDeviceExtensionPresent("VK_EXT_fragment_shader_interlock"))
             {
                 featuresFragmentShaderInterlock = new PhysicalDeviceFragmentShaderInterlockFeaturesEXT()
                 {
@@ -552,7 +476,7 @@ namespace Ryujinx.Graphics.Vulkan
 
             PhysicalDeviceSubgroupSizeControlFeaturesEXT featuresSubgroupSizeControl;
 
-            if (supportedExtensions.Contains("VK_EXT_subgroup_size_control"))
+            if (physicalDevice.IsDeviceExtensionPresent("VK_EXT_subgroup_size_control"))
             {
                 featuresSubgroupSizeControl = new PhysicalDeviceSubgroupSizeControlFeaturesEXT()
                 {
@@ -566,7 +490,7 @@ namespace Ryujinx.Graphics.Vulkan
 
             PhysicalDeviceCustomBorderColorFeaturesEXT featuresCustomBorderColor;
 
-            if (supportedExtensions.Contains("VK_EXT_custom_border_color") &&
+            if (physicalDevice.IsDeviceExtensionPresent("VK_EXT_custom_border_color") &&
                 supportedFeaturesCustomBorderColor.CustomBorderColors &&
                 supportedFeaturesCustomBorderColor.CustomBorderColorWithoutFormat)
             {
@@ -581,7 +505,7 @@ namespace Ryujinx.Graphics.Vulkan
                 pExtendedFeatures = &featuresCustomBorderColor;
             }
 
-            var enabledExtensions = _requiredExtensions.Union(_desirableExtensions.Intersect(supportedExtensions)).ToArray();
+            var enabledExtensions = _requiredExtensions.Union(_desirableExtensions.Intersect(physicalDevice.DeviceExtensions)).ToArray();
 
             IntPtr* ppEnabledExtensions = stackalloc IntPtr[enabledExtensions.Length];
 
@@ -601,7 +525,7 @@ namespace Ryujinx.Graphics.Vulkan
                 PEnabledFeatures = &features
             };
 
-            api.CreateDevice(physicalDevice, in deviceCreateInfo, null, out var device).ThrowOnError();
+            api.CreateDevice(physicalDevice.PhysicalDevice, in deviceCreateInfo, null, out var device).ThrowOnError();
 
             for (int i = 0; i < enabledExtensions.Length; i++)
             {
@@ -610,21 +534,5 @@ namespace Ryujinx.Graphics.Vulkan
 
             return device;
         }
-
-        public static string[] GetSupportedExtensions(Vk api, PhysicalDevice physicalDevice)
-        {
-            uint propertiesCount;
-
-            api.EnumerateDeviceExtensionProperties(physicalDevice, (byte*)null, &propertiesCount, null).ThrowOnError();
-
-            ExtensionProperties[] extensionProperties = new ExtensionProperties[propertiesCount];
-
-            fixed (ExtensionProperties* pExtensionProperties = extensionProperties)
-            {
-                api.EnumerateDeviceExtensionProperties(physicalDevice, (byte*)null, &propertiesCount, pExtensionProperties).ThrowOnError();
-            }
-
-            return extensionProperties.Select(x => Marshal.PtrToStringAnsi((IntPtr)x.ExtensionName)).ToArray();
-        }
     }
 }

Ryujinx.Graphics.Vulkan/VulkanInstance.cs

@@ -0,0 +1,127 @@
+using Ryujinx.Common.Utilities;
+using Silk.NET.Core;
+using Silk.NET.Vulkan;
+using System;
+using System.Collections.Generic;
+using System.Collections.Immutable;
+using System.Linq;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Vulkan
+{
+    class VulkanInstance : IDisposable
+    {
+        private readonly Vk _api;
+        public readonly Instance Instance;
+        public readonly Version32 InstanceVersion;
+
+        private bool _disposed;
+
+        private VulkanInstance(Vk api, Instance instance)
+        {
+            _api = api;
+            Instance = instance;
+
+            if (api.GetInstanceProcAddr(instance, "vkEnumerateInstanceVersion") == IntPtr.Zero)
+            {
+                InstanceVersion = Vk.Version10;
+            }
+            else
+            {
+                uint rawInstanceVersion = 0;
+
+                if (api.EnumerateInstanceVersion(ref rawInstanceVersion) != Result.Success)
+                {
+                    rawInstanceVersion = Vk.Version11.Value;
+                }
+
+                InstanceVersion = (Version32)rawInstanceVersion;
+            }
+        }
+
+        public static Result Create(Vk api, ref InstanceCreateInfo createInfo, out VulkanInstance instance)
+        {
+            instance = null;
+
+            Instance rawInstance = default;
+
+            Result result = api.CreateInstance(SpanHelpers.AsReadOnlySpan(ref createInfo), ReadOnlySpan<AllocationCallbacks>.Empty, SpanHelpers.AsSpan(ref rawInstance));
+
+            if (result == Result.Success)
+            {
+                instance = new VulkanInstance(api, rawInstance);
+            }
+
+            return result;
+        }
+
+        public Result EnumeratePhysicalDevices(out VulkanPhysicalDevice[] physicalDevices)
+        {
+            physicalDevices = null;
+
+            uint physicalDeviceCount = 0;
+
+            Result result = _api.EnumeratePhysicalDevices(Instance, SpanHelpers.AsSpan(ref physicalDeviceCount), Span<PhysicalDevice>.Empty);
+
+            if (result != Result.Success)
+            {
+                return result;
+            }
+
+            PhysicalDevice[] rawPhysicalDevices = new PhysicalDevice[physicalDeviceCount];
+
+            result = _api.EnumeratePhysicalDevices(Instance, SpanHelpers.AsSpan(ref physicalDeviceCount), rawPhysicalDevices);
+
+            if (result != Result.Success)
+            {
+                return result;
+            }
+
+            physicalDevices = rawPhysicalDevices.Select(x => new VulkanPhysicalDevice(_api, x)).ToArray();
+
+            return Result.Success;
+        }
+
+        public static IReadOnlySet<string> GetInstanceExtensions(Vk api)
+        {
+            uint propertiesCount = 0;
+
+            api.EnumerateInstanceExtensionProperties(ReadOnlySpan<byte>.Empty, SpanHelpers.AsSpan(ref propertiesCount), Span<ExtensionProperties>.Empty).ThrowOnError();
+
+            ExtensionProperties[] extensionProperties = new ExtensionProperties[propertiesCount];
+
+            api.EnumerateInstanceExtensionProperties(ReadOnlySpan<byte>.Empty, SpanHelpers.AsSpan(ref propertiesCount), extensionProperties).ThrowOnError();
+
+            unsafe
+            {
+                return extensionProperties.Select(x => Marshal.PtrToStringAnsi((IntPtr)x.ExtensionName)).ToImmutableHashSet();
+            }
+        }
+
+        public static IReadOnlySet<string> GetInstanceLayers(Vk api)
+        {
+            uint propertiesCount = 0;
+
+            api.EnumerateInstanceLayerProperties(SpanHelpers.AsSpan(ref propertiesCount), Span<LayerProperties>.Empty).ThrowOnError();
+
+            LayerProperties[] layerProperties = new LayerProperties[propertiesCount];
+
+            api.EnumerateInstanceLayerProperties(SpanHelpers.AsSpan(ref propertiesCount), layerProperties).ThrowOnError();
+
+            unsafe
+            {
+                return layerProperties.Select(x => Marshal.PtrToStringAnsi((IntPtr)x.LayerName)).ToImmutableHashSet();
+            }
+        }
+
+        public void Dispose()
+        {
+            if (!_disposed)
+            {
+                _api.DestroyInstance(Instance, ReadOnlySpan<AllocationCallbacks>.Empty);
+
+                _disposed = true;
+            }
+        }
+    }
+}

Ryujinx.Graphics.Vulkan/VulkanPhysicalDevice.cs

@@ -0,0 +1,70 @@
+using Ryujinx.Common.Utilities;
+using Ryujinx.Graphics.GAL;
+using Silk.NET.Vulkan;
+using System;
+using System.Collections.Generic;
+using System.Collections.Immutable;
+using System.Linq;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Vulkan
+{
+    readonly struct VulkanPhysicalDevice
+    {
+        public readonly PhysicalDevice PhysicalDevice;
+        public readonly PhysicalDeviceFeatures PhysicalDeviceFeatures;
+        public readonly PhysicalDeviceProperties PhysicalDeviceProperties;
+        public readonly PhysicalDeviceMemoryProperties PhysicalDeviceMemoryProperties;
+        public readonly QueueFamilyProperties[] QueueFamilyProperties;
+        public readonly string DeviceName;
+        public readonly IReadOnlySet<string> DeviceExtensions;
+
+        public VulkanPhysicalDevice(Vk api, PhysicalDevice physicalDevice)
+        {
+            PhysicalDevice = physicalDevice;
+            PhysicalDeviceFeatures = api.GetPhysicalDeviceFeature(PhysicalDevice);
+
+            api.GetPhysicalDeviceProperties(PhysicalDevice, out var physicalDeviceProperties);
+            PhysicalDeviceProperties = physicalDeviceProperties;
+
+            api.GetPhysicalDeviceMemoryProperties(PhysicalDevice, out PhysicalDeviceMemoryProperties);
+
+            unsafe
+            {
+                DeviceName = Marshal.PtrToStringAnsi((IntPtr)physicalDeviceProperties.DeviceName);
+            }
+
+            uint propertiesCount = 0;
+
+            api.GetPhysicalDeviceQueueFamilyProperties(physicalDevice, SpanHelpers.AsSpan(ref propertiesCount), Span<QueueFamilyProperties>.Empty);
+
+            QueueFamilyProperties = new QueueFamilyProperties[propertiesCount];
+
+            api.GetPhysicalDeviceQueueFamilyProperties(physicalDevice, SpanHelpers.AsSpan(ref propertiesCount), QueueFamilyProperties);
+
+            api.EnumerateDeviceExtensionProperties(PhysicalDevice, Span<byte>.Empty, SpanHelpers.AsSpan(ref propertiesCount), Span<ExtensionProperties>.Empty).ThrowOnError();
+
+            ExtensionProperties[] extensionProperties = new ExtensionProperties[propertiesCount];
+
+            api.EnumerateDeviceExtensionProperties(PhysicalDevice, Span<byte>.Empty, SpanHelpers.AsSpan(ref propertiesCount), extensionProperties).ThrowOnError();
+
+            unsafe
+            {
+                DeviceExtensions = extensionProperties.Select(x => Marshal.PtrToStringAnsi((IntPtr)x.ExtensionName)).ToImmutableHashSet();
+            }
+        }
+
+        public string Id => $"0x{PhysicalDeviceProperties.VendorID:X}_0x{PhysicalDeviceProperties.DeviceID:X}";
+
+        public bool IsDeviceExtensionPresent(string extension) => DeviceExtensions.Contains(extension);
+
+        public DeviceInfo ToDeviceInfo()
+        {
+            return new DeviceInfo(
+                Id,
+                VendorUtils.GetNameFromId(PhysicalDeviceProperties.VendorID),
+                DeviceName,
+                PhysicalDeviceProperties.DeviceType == PhysicalDeviceType.DiscreteGpu);
+        }
+    }
+}

Ryujinx.Graphics.Vulkan/VulkanRenderer.cs

@@ -17,9 +17,9 @@ namespace Ryujinx.Graphics.Vulkan
 {
     public sealed class VulkanRenderer : IRenderer
     {
-        private Instance _instance;
+        private VulkanInstance _instance;
         private SurfaceKHR _surface;
-        private PhysicalDevice _physicalDevice;
+        private VulkanPhysicalDevice _physicalDevice;
         private Device _device;
         private WindowBase _window;
 
@@ -106,33 +106,31 @@ namespace Ryujinx.Graphics.Vulkan
             }
         }
 
-        private unsafe void LoadFeatures(string[] supportedExtensions, uint maxQueueCount, uint queueFamilyIndex)
+        private unsafe void LoadFeatures(uint maxQueueCount, uint queueFamilyIndex)
         {
-            FormatCapabilities = new FormatCapabilities(Api, _physicalDevice);
+            FormatCapabilities = new FormatCapabilities(Api, _physicalDevice.PhysicalDevice);
 
-            var supportedFeatures = Api.GetPhysicalDeviceFeature(_physicalDevice);
+            if (Api.TryGetDeviceExtension(_instance.Instance, _device, out ExtConditionalRendering conditionalRenderingApi))
-
-            if (Api.TryGetDeviceExtension(_instance, _device, out ExtConditionalRendering conditionalRenderingApi))
             {
                 ConditionalRenderingApi = conditionalRenderingApi;
             }
 
-            if (Api.TryGetDeviceExtension(_instance, _device, out ExtExtendedDynamicState extendedDynamicStateApi))
+            if (Api.TryGetDeviceExtension(_instance.Instance, _device, out ExtExtendedDynamicState extendedDynamicStateApi))
             {
                 ExtendedDynamicStateApi = extendedDynamicStateApi;
             }
 
-            if (Api.TryGetDeviceExtension(_instance, _device, out KhrPushDescriptor pushDescriptorApi))
+            if (Api.TryGetDeviceExtension(_instance.Instance, _device, out KhrPushDescriptor pushDescriptorApi))
             {
                 PushDescriptorApi = pushDescriptorApi;
             }
 
-            if (Api.TryGetDeviceExtension(_instance, _device, out ExtTransformFeedback transformFeedbackApi))
+            if (Api.TryGetDeviceExtension(_instance.Instance, _device, out ExtTransformFeedback transformFeedbackApi))
             {
                 TransformFeedbackApi = transformFeedbackApi;
             }
 
-            if (Api.TryGetDeviceExtension(_instance, _device, out KhrDrawIndirectCount drawIndirectCountApi))
+            if (Api.TryGetDeviceExtension(_instance.Instance, _device, out KhrDrawIndirectCount drawIndirectCountApi))
             {
                 DrawIndirectCountApi = drawIndirectCountApi;
             }
@@ -154,7 +152,7 @@ namespace Ryujinx.Graphics.Vulkan
                 SType = StructureType.PhysicalDeviceBlendOperationAdvancedPropertiesExt
             };
 
-            bool supportsBlendOperationAdvanced = supportedExtensions.Contains("VK_EXT_blend_operation_advanced");
+            bool supportsBlendOperationAdvanced = _physicalDevice.IsDeviceExtensionPresent("VK_EXT_blend_operation_advanced");
 
             if (supportsBlendOperationAdvanced)
             {
@@ -167,14 +165,14 @@ namespace Ryujinx.Graphics.Vulkan
                 SType = StructureType.PhysicalDeviceSubgroupSizeControlPropertiesExt
             };
 
-            bool supportsSubgroupSizeControl = supportedExtensions.Contains("VK_EXT_subgroup_size_control");
+            bool supportsSubgroupSizeControl = _physicalDevice.IsDeviceExtensionPresent("VK_EXT_subgroup_size_control");
 
             if (supportsSubgroupSizeControl)
             {
                 properties2.PNext = &propertiesSubgroupSizeControl;
             }
 
-            bool supportsTransformFeedback = supportedExtensions.Contains(ExtTransformFeedback.ExtensionName);
+            bool supportsTransformFeedback = _physicalDevice.IsDeviceExtensionPresent(ExtTransformFeedback.ExtensionName);
 
             PhysicalDeviceTransformFeedbackPropertiesEXT propertiesTransformFeedback = new PhysicalDeviceTransformFeedbackPropertiesEXT()
             {
@@ -222,30 +220,30 @@ namespace Ryujinx.Graphics.Vulkan
                 SType = StructureType.PhysicalDevicePortabilitySubsetFeaturesKhr
             };
 
-            if (supportedExtensions.Contains("VK_EXT_primitive_topology_list_restart"))
+            if (_physicalDevice.IsDeviceExtensionPresent("VK_EXT_primitive_topology_list_restart"))
             {
                 features2.PNext = &featuresPrimitiveTopologyListRestart;
             }
 
-            if (supportedExtensions.Contains("VK_EXT_robustness2"))
+            if (_physicalDevice.IsDeviceExtensionPresent("VK_EXT_robustness2"))
             {
                 featuresRobustness2.PNext = features2.PNext;
                 features2.PNext = &featuresRobustness2;
             }
 
-            if (supportedExtensions.Contains("VK_KHR_shader_float16_int8"))
+            if (_physicalDevice.IsDeviceExtensionPresent("VK_KHR_shader_float16_int8"))
             {
                 featuresShaderInt8.PNext = features2.PNext;
                 features2.PNext = &featuresShaderInt8;
             }
 
-            if (supportedExtensions.Contains("VK_EXT_custom_border_color"))
+            if (_physicalDevice.IsDeviceExtensionPresent("VK_EXT_custom_border_color"))
             {
                 featuresCustomBorderColor.PNext = features2.PNext;
                 features2.PNext = &featuresCustomBorderColor;
             }
 
-            bool usePortability = supportedExtensions.Contains("VK_KHR_portability_subset");
+            bool usePortability = _physicalDevice.IsDeviceExtensionPresent("VK_KHR_portability_subset");
 
             if (usePortability)
             {
@@ -256,8 +254,8 @@ namespace Ryujinx.Graphics.Vulkan
                 features2.PNext = &featuresPortabilitySubset;
             }
 
-            Api.GetPhysicalDeviceProperties2(_physicalDevice, &properties2);
+            Api.GetPhysicalDeviceProperties2(_physicalDevice.PhysicalDevice, &properties2);
-            Api.GetPhysicalDeviceFeatures2(_physicalDevice, &features2);
+            Api.GetPhysicalDeviceFeatures2(_physicalDevice.PhysicalDevice, &features2);
 
             var portabilityFlags = PortabilitySubsetFlags.None;
             uint vertexBufferAlignment = 1;
@@ -272,7 +270,7 @@ namespace Ryujinx.Graphics.Vulkan
                 portabilityFlags |= featuresPortabilitySubset.SamplerMipLodBias ? 0 : PortabilitySubsetFlags.NoLodBias;
             }
 
-            bool supportsCustomBorderColor = supportedExtensions.Contains("VK_EXT_custom_border_color") &&
+            bool supportsCustomBorderColor = _physicalDevice.IsDeviceExtensionPresent("VK_EXT_custom_border_color") &&
                                              featuresCustomBorderColor.CustomBorderColors &&
                                              featuresCustomBorderColor.CustomBorderColorWithoutFormat;
 
@@ -284,30 +282,30 @@ namespace Ryujinx.Graphics.Vulkan
                 properties.Limits.FramebufferStencilSampleCounts;
 
             Capabilities = new HardwareCapabilities(
-                supportedExtensions.Contains("VK_EXT_index_type_uint8"),
+                _physicalDevice.IsDeviceExtensionPresent("VK_EXT_index_type_uint8"),
                 supportsCustomBorderColor,
                 supportsBlendOperationAdvanced,
                 propertiesBlendOperationAdvanced.AdvancedBlendCorrelatedOverlap,
                 propertiesBlendOperationAdvanced.AdvancedBlendNonPremultipliedSrcColor,
                 propertiesBlendOperationAdvanced.AdvancedBlendNonPremultipliedDstColor,
-                supportedExtensions.Contains(KhrDrawIndirectCount.ExtensionName),
+                _physicalDevice.IsDeviceExtensionPresent(KhrDrawIndirectCount.ExtensionName),
-                supportedExtensions.Contains("VK_EXT_fragment_shader_interlock"),
+                _physicalDevice.IsDeviceExtensionPresent("VK_EXT_fragment_shader_interlock"),
-                supportedExtensions.Contains("VK_NV_geometry_shader_passthrough"),
+                _physicalDevice.IsDeviceExtensionPresent("VK_NV_geometry_shader_passthrough"),
                 supportsSubgroupSizeControl,
                 featuresShaderInt8.ShaderInt8,
-                supportedExtensions.Contains("VK_EXT_shader_stencil_export"),
+                _physicalDevice.IsDeviceExtensionPresent("VK_EXT_shader_stencil_export"),
-                supportedExtensions.Contains(ExtConditionalRendering.ExtensionName),
+                _physicalDevice.IsDeviceExtensionPresent(ExtConditionalRendering.ExtensionName),
-                supportedExtensions.Contains(ExtExtendedDynamicState.ExtensionName),
+                _physicalDevice.IsDeviceExtensionPresent(ExtExtendedDynamicState.ExtensionName),
                 features2.Features.MultiViewport,
                 featuresRobustness2.NullDescriptor || IsMoltenVk,
-                supportedExtensions.Contains(KhrPushDescriptor.ExtensionName),
+                _physicalDevice.IsDeviceExtensionPresent(KhrPushDescriptor.ExtensionName),
                 featuresPrimitiveTopologyListRestart.PrimitiveTopologyListRestart,
                 featuresPrimitiveTopologyListRestart.PrimitiveTopologyPatchListRestart,
                 supportsTransformFeedback,
                 propertiesTransformFeedback.TransformFeedbackQueries,
                 features2.Features.OcclusionQueryPrecise,
-                supportedFeatures.PipelineStatisticsQuery,
+                _physicalDevice.PhysicalDeviceFeatures.PipelineStatisticsQuery,
-                supportedFeatures.GeometryShader,
+                _physicalDevice.PhysicalDeviceFeatures.GeometryShader,
                 propertiesSubgroupSizeControl.MinSubgroupSize,
                 propertiesSubgroupSizeControl.MaxSubgroupSize,
                 propertiesSubgroupSizeControl.RequiredSubgroupSizeStages,
@@ -315,9 +313,9 @@ namespace Ryujinx.Graphics.Vulkan
                 portabilityFlags,
                 vertexBufferAlignment);
 
-            IsSharedMemory = MemoryAllocator.IsDeviceMemoryShared(Api, _physicalDevice);
+            IsSharedMemory = MemoryAllocator.IsDeviceMemoryShared(_physicalDevice);
 
-            MemoryAllocator = new MemoryAllocator(Api, _physicalDevice, _device, properties.Limits.MaxMemoryAllocationCount);
+            MemoryAllocator = new MemoryAllocator(Api, _physicalDevice, _device);
 
             CommandBufferPool = new CommandBufferPool(Api, _device, Queue, QueueLock, queueFamilyIndex);
 
@@ -345,22 +343,21 @@ namespace Ryujinx.Graphics.Vulkan
             Api = api;
 
             _instance = VulkanInitialization.CreateInstance(api, logLevel, _getRequiredExtensions());
-            _debugMessenger = new VulkanDebugMessenger(api, _instance, logLevel);
+            _debugMessenger = new VulkanDebugMessenger(api, _instance.Instance, logLevel);
 
-            if (api.TryGetInstanceExtension(_instance, out KhrSurface surfaceApi))
+            if (api.TryGetInstanceExtension(_instance.Instance, out KhrSurface surfaceApi))
             {
                 SurfaceApi = surfaceApi;
             }
 
-            _surface = _getSurface(_instance, api);
+            _surface = _getSurface(_instance.Instance, api);
             _physicalDevice = VulkanInitialization.FindSuitablePhysicalDevice(api, _instance, _surface, _preferredGpuId);
 
             var queueFamilyIndex = VulkanInitialization.FindSuitableQueueFamily(api, _physicalDevice, _surface, out uint maxQueueCount);
-            var supportedExtensions = VulkanInitialization.GetSupportedExtensions(api, _physicalDevice);
 
-            _device = VulkanInitialization.CreateDevice(api, _physicalDevice, queueFamilyIndex, supportedExtensions, maxQueueCount);
+            _device = VulkanInitialization.CreateDevice(api, _physicalDevice, queueFamilyIndex, maxQueueCount);
 
-            if (api.TryGetDeviceExtension(_instance, _device, out KhrSwapchain swapchainApi))
+            if (api.TryGetDeviceExtension(_instance.Instance, _device, out KhrSwapchain swapchainApi))
             {
                 SwapchainApi = swapchainApi;
             }
@@ -369,9 +366,9 @@ namespace Ryujinx.Graphics.Vulkan
             Queue = queue;
             QueueLock = new object();
 
-            LoadFeatures(supportedExtensions, maxQueueCount, queueFamilyIndex);
+            LoadFeatures(maxQueueCount, queueFamilyIndex);
 
-            _window = new Window(this, _surface, _physicalDevice, _device);
+            _window = new Window(this, _surface, _physicalDevice.PhysicalDevice, _device);
 
             _initialized = true;
         }
@@ -536,10 +533,9 @@ namespace Ryujinx.Graphics.Vulkan
                 PNext = &featuresVk12
             };
 
-            Api.GetPhysicalDeviceFeatures2(_physicalDevice, &features2);
+            Api.GetPhysicalDeviceFeatures2(_physicalDevice.PhysicalDevice, &features2);
-            Api.GetPhysicalDeviceProperties(_physicalDevice, out var properties);
 
-            var limits = properties.Limits;
+            var limits = _physicalDevice.PhysicalDeviceProperties.Limits;
 
             return new Capabilities(
                 api: TargetApi.Vulkan,
@@ -623,7 +619,7 @@ namespace Ryujinx.Graphics.Vulkan
 
         private unsafe void PrintGpuInformation()
         {
-            Api.GetPhysicalDeviceProperties(_physicalDevice, out var properties);
+            var properties = _physicalDevice.PhysicalDeviceProperties;
 
             string vendorName = VendorUtils.GetNameFromId(properties.VendorID);
 
@@ -807,14 +803,14 @@ namespace Ryujinx.Graphics.Vulkan
                 sampler.Dispose();
             }
 
-            SurfaceApi.DestroySurface(_instance, _surface, null);
+            SurfaceApi.DestroySurface(_instance.Instance, _surface, null);
 
             Api.DestroyDevice(_device, null);
 
             _debugMessenger.Dispose();
 
             // Last step destroy the instance
-            Api.DestroyInstance(_instance, null);
+            _instance.Dispose();
         }
     }
 }

Ryujinx.HLE/HOS/Services/Hid/Types/SharedMemory/Common/AtomicStorage.cs

@@ -2,7 +2,7 @@
 
 namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Common
 {
-    struct AtomicStorage<T> where T: unmanaged
+    struct AtomicStorage<T> where T: unmanaged, ISampledDataStruct
     {
         public ulong SamplingNumber;
         public T Object;
@@ -14,9 +14,9 @@ namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Common
 
         public void SetObject(ref T obj)
         {
-            ISampledData samplingProvider = obj as ISampledData;
+            ulong samplingNumber = ISampledDataStruct.GetSamplingNumber(ref obj);
 
-            Interlocked.Exchange(ref SamplingNumber, samplingProvider.SamplingNumber);
+            Interlocked.Exchange(ref SamplingNumber, samplingNumber);
 
             Thread.MemoryBarrier();
 

Ryujinx.HLE/HOS/Services/Hid/Types/SharedMemory/Common/ISampledData.cs

@@ -1,7 +0,0 @@
-namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Common
-{
-    interface ISampledData
-    {
-        ulong SamplingNumber { get; }
-    }
-}

Ryujinx.HLE/HOS/Services/Hid/Types/SharedMemory/Common/ISampledDataStruct.cs

@@ -0,0 +1,65 @@
+using System;
+using System.Buffers.Binary;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Common
+{
+    /// <summary>
+    /// This is a "marker interface" to add some compile-time safety to a convention-based optimization.
+    ///
+    /// Any struct implementing this interface should:
+    ///   - use <c>StructLayoutAttribute</c> (and related attributes) to explicity control how the struct is laid out in memory.
+    ///   - ensure that the method <c>ISampledDataStruct.GetSamplingNumberFieldOffset()</c> correctly returns the offset, in bytes,
+    ///     to the ulong "Sampling Number" field within the struct. Most types have it as the first field, so the default offset is 0.
+    /// 
+    /// Example:
+    /// 
+    /// <c>
+    ///         [StructLayout(LayoutKind.Sequential, Pack = 8)]
+    ///         struct DebugPadState : ISampledDataStruct
+    ///         {
+    ///             public ulong SamplingNumber;    // 1st field, so no need to add special handling to GetSamplingNumberFieldOffset()
+    ///             // other members...
+    ///         }
+    ///
+    ///         [StructLayout(LayoutKind.Sequential, Pack = 8)]
+    ///         struct SixAxisSensorState : ISampledDataStruct
+    ///         {
+    ///             public ulong DeltaTime;
+    ///             public ulong SamplingNumber;    // Not the first field - needs special handling in GetSamplingNumberFieldOffset()
+    ///             // other members...
+    ///         }
+    /// </c>
+    /// </summary>
+    internal interface ISampledDataStruct
+    {
+        // No Instance Members - marker interface only
+
+        public static ulong GetSamplingNumber<T>(ref T sampledDataStruct) where T : unmanaged, ISampledDataStruct
+        {
+            ReadOnlySpan<T> structSpan = MemoryMarshal.CreateReadOnlySpan(ref sampledDataStruct, 1);
+
+            ReadOnlySpan<byte> byteSpan = MemoryMarshal.Cast<T, byte>(structSpan);
+
+            int fieldOffset = GetSamplingNumberFieldOffset(ref sampledDataStruct);
+
+            if (fieldOffset > 0)
+            {
+                byteSpan = byteSpan.Slice(fieldOffset);
+            }
+
+            ulong value = BinaryPrimitives.ReadUInt64LittleEndian(byteSpan);
+
+            return value;
+        }
+
+        private static int GetSamplingNumberFieldOffset<T>(ref T sampledDataStruct) where T : unmanaged, ISampledDataStruct
+        {
+            return sampledDataStruct switch
+            {
+                Npad.SixAxisSensorState _ => sizeof(ulong),
+                _ => 0
+            };
+        }
+    }
+}

Ryujinx.HLE/HOS/Services/Hid/Types/SharedMemory/Common/RingLifo.cs

@@ -5,7 +5,7 @@ using System.Threading;
 
 namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Common
 {
-    struct RingLifo<T> where T: unmanaged
+    struct RingLifo<T> where T: unmanaged, ISampledDataStruct
     {
         private const ulong MaxEntries = 17;
 

Ryujinx.HLE/HOS/Services/Hid/Types/SharedMemory/DebugPad/DebugPadState.cs

@@ -1,15 +1,15 @@
 using Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Common;
+using System.Runtime.InteropServices;
 
 namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.DebugPad
 {
-    struct DebugPadState : ISampledData
+    [StructLayout(LayoutKind.Sequential, Pack = 1)]
+    struct DebugPadState : ISampledDataStruct
     {
         public ulong SamplingNumber;
         public DebugPadAttribute Attributes;
         public DebugPadButton Buttons;
         public AnalogStickState AnalogStickR;
         public AnalogStickState AnalogStickL;
-
-        ulong ISampledData.SamplingNumber => SamplingNumber;
     }
 }

Ryujinx.HLE/HOS/Services/Hid/Types/SharedMemory/Keyboard/KeyboardModifier.cs

@@ -2,9 +2,8 @@
 
 namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Keyboard
 {
-    // TODO: This seems entirely wrong
     [Flags]
-    enum KeyboardModifier : uint
+    enum KeyboardModifier : ulong
     {
         None = 0,
         Control = 1 << 0,

Ryujinx.HLE/HOS/Services/Hid/Types/SharedMemory/Keyboard/KeyboardState.cs

@@ -1,13 +1,13 @@
 using Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Common;
+using System.Runtime.InteropServices;
 
 namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Keyboard
 {
-    struct KeyboardState : ISampledData
+    [StructLayout(LayoutKind.Sequential, Pack = 1)]
+    struct KeyboardState : ISampledDataStruct
     {
         public ulong SamplingNumber;
         public KeyboardModifier Modifiers;
         public KeyboardKey Keys;
-
-        ulong ISampledData.SamplingNumber => SamplingNumber;
     }
 }

Ryujinx.HLE/HOS/Services/Hid/Types/SharedMemory/Mouse/MouseState.cs

@@ -1,8 +1,10 @@
 using Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Common;
+using System.Runtime.InteropServices;
 
 namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Mouse
 {
-    struct MouseState : ISampledData
+    [StructLayout(LayoutKind.Sequential, Pack = 1)]
+    struct MouseState : ISampledDataStruct
     {
         public ulong SamplingNumber;
         public int X;
@@ -13,7 +15,5 @@ namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Mouse
         public int WheelDeltaY;
         public MouseButton Buttons;
         public MouseAttribute Attributes;
-
-        ulong ISampledData.SamplingNumber => SamplingNumber;
     }
 }

Ryujinx.HLE/HOS/Services/Hid/Types/SharedMemory/Npad/NpadCommonState.cs

@@ -1,8 +1,10 @@
 using Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Common;
+using System.Runtime.InteropServices;
 
 namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Npad
 {
-    struct NpadCommonState : ISampledData
+    [StructLayout(LayoutKind.Sequential, Pack = 1)]
+    struct NpadCommonState : ISampledDataStruct
     {
         public ulong SamplingNumber;
         public NpadButton Buttons;
@@ -10,7 +12,5 @@ namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Npad
         public AnalogStickState AnalogStickR;
         public NpadAttribute Attributes;
         private uint _reserved;
-
-        ulong ISampledData.SamplingNumber => SamplingNumber;
     }
 }

Ryujinx.HLE/HOS/Services/Hid/Types/SharedMemory/Npad/NpadGcTriggerState.cs

@@ -1,15 +1,15 @@
 using Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Common;
+using System.Runtime.InteropServices;
 
 namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Npad
 {
-    struct NpadGcTriggerState : ISampledData
+    [StructLayout(LayoutKind.Sequential, Pack = 1)]
+    struct NpadGcTriggerState : ISampledDataStruct
     {
 #pragma warning disable CS0649
         public ulong SamplingNumber;
         public uint TriggerL;
         public uint TriggerR;
 #pragma warning restore CS0649
-
-        ulong ISampledData.SamplingNumber => SamplingNumber;
     }
 }

Ryujinx.HLE/HOS/Services/Hid/Types/SharedMemory/Npad/SixAxisSensorState.cs

@@ -1,9 +1,11 @@
 using Ryujinx.Common.Memory;
 using Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Common;
+using System.Runtime.InteropServices;
 
 namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Npad
 {
-    struct SixAxisSensorState : ISampledData
+    [StructLayout(LayoutKind.Sequential, Pack = 1)]
+    struct SixAxisSensorState : ISampledDataStruct
     {
         public ulong DeltaTime;
         public ulong SamplingNumber;
@@ -13,7 +15,5 @@ namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Npad
         public Array9<float> Direction;
         public SixAxisSensorAttribute Attributes;
         private uint _reserved;
-
-        ulong ISampledData.SamplingNumber => SamplingNumber;
     }
 }

Ryujinx.HLE/HOS/Services/Hid/Types/SharedMemory/TouchScreen/TouchScreenState.cs

@@ -1,15 +1,15 @@
 using Ryujinx.Common.Memory;
 using Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.Common;
+using System.Runtime.InteropServices;
 
 namespace Ryujinx.HLE.HOS.Services.Hid.Types.SharedMemory.TouchScreen
 {
-    struct TouchScreenState : ISampledData
+    [StructLayout(LayoutKind.Sequential, Pack = 1)]
+    struct TouchScreenState : ISampledDataStruct
     {
         public ulong SamplingNumber;
         public int TouchesCount;
         private int _reserved;
         public Array16<TouchState> Touches;
-
-        ulong ISampledData.SamplingNumber => SamplingNumber;
     }
 }

Ryujinx.Tests/Cpu/EnvironmentTests.cs

@@ -0,0 +1,91 @@
+using ARMeilleure.Translation;
+using NUnit.Framework;
+using Ryujinx.Cpu.Jit;
+using Ryujinx.Tests.Memory;
+using System;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Tests.Cpu
+{
+    internal class EnvironmentTests
+    {
+        private static Translator _translator;
+
+        private void EnsureTranslator()
+        {
+            // Create a translator, as one is needed to register the signal handler or emit methods.
+            _translator ??= new Translator(new JitMemoryAllocator(), new MockMemoryManager(), true);
+        }
+
+        [MethodImpl(MethodImplOptions.NoInlining | MethodImplOptions.NoOptimization)]
+        private float GetDenormal()
+        {
+            return BitConverter.Int32BitsToSingle(1);
+        }
+
+        [MethodImpl(MethodImplOptions.NoInlining | MethodImplOptions.NoOptimization)]
+        private float GetZero()
+        {
+            return BitConverter.Int32BitsToSingle(0);
+        }
+
+        /// <summary>
+        /// This test ensures that managed methods do not reset floating point control flags.
+        /// This is used to avoid changing control flags when running methods that don't require it, such as SVC calls, software memory...
+        /// </summary>
+        [Test]
+        public void FpFlagsPInvoke()
+        {
+            EnsureTranslator();
+
+            // Subnormal results are not flushed to zero by default.
+            // This operation should not be allowed to do constant propagation, hence the methods that explicitly disallow inlining.
+            Assert.AreNotEqual(GetDenormal() + GetZero(), 0f);
+
+            bool methodCalled = false;
+            bool isFz = false;
+
+            var managedMethod = () =>
+            {
+                // Floating point math should not modify fp flags.
+                float test = 2f * 3.5f;
+
+                if (test < 4f)
+                {
+                    throw new System.Exception("Sanity check.");
+                }
+
+                isFz = GetDenormal() + GetZero() == 0f;
+
+                try
+                {
+                    if (test >= 4f)
+                    {
+                        throw new System.Exception("Always throws.");
+                    }
+                }
+                catch
+                {
+                    // Exception handling should not modify fp flags.
+
+                    methodCalled = true;
+                }
+            };
+
+            var method = TranslatorTestMethods.GenerateFpFlagsPInvokeTest();
+
+            // This method sets flush-to-zero and then calls the managed method.
+            // Before and after setting the flags, it ensures subnormal addition works as expected.
+            // It returns a positive result if any tests fail, and 0 on success (or if the platform cannot change FP flags)
+            int result = method(Marshal.GetFunctionPointerForDelegate(managedMethod));
+
+            // Subnormal results are not flushed to zero by default, which we should have returned to exiting the method.
+            Assert.AreNotEqual(GetDenormal() + GetZero(), 0f);
+
+            Assert.True(result == 0);
+            Assert.True(methodCalled);
+            Assert.True(isFz);
+        }
+    }
+}

Ryujinx/Ui/GLRenderer.cs

@@ -119,7 +119,7 @@ namespace Ryujinx.Ui
             }
             catch (Exception) { }
 
-            Device.DisposeGpu();
+            Device?.DisposeGpu();
             NpadManager.Dispose();
 
             // Unbind context and destroy everything
@@ -129,7 +129,7 @@ namespace Ryujinx.Ui
             }
             catch (Exception) { }
 
-            _openGLContext.Dispose();
+            _openGLContext?.Dispose();
         }
     }
 }