Mirror/Ryujinx
3
5
Fork 13
mirror of https://github.com/Ryujinx/Ryujinx.git synced 2024-11-20 00:16:35 +00:00
Code Issues Projects Releases Packages Wiki Activity

Move a few more SIMD instructions to emit CIL directly instead of a method call

Browse source
This commit is contained in:
gdkchan 2018-02-09 17:14:47 -03:00
parent 12e263f8b2
commit ccc9ce1908
7 changed files with 207 additions and 324 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
Ryujinx/Cpu/AOpCodeTable.cs
View file

@ -152,13 +152,13 @@ namespace ChocolArm64
Set("x0011110xx110000000000xxxxxxxxxx", AInstEmit.Fcvtms_S, typeof(AOpCodeSimdCvt));
Set("x0011110xx101000000000xxxxxxxxxx", AInstEmit.Fcvtps_S, typeof(AOpCodeSimdCvt));
Set("x0011110xx111000000000xxxxxxxxxx", AInstEmit.Fcvtzs_S, typeof(AOpCodeSimdCvt));
Set("0x0011101x100001101110xxxxxxxxxx", AInstEmit.Fcvtzs_V, typeof(AOpCodeSimd));
Set("0x0011110>>xxxxx111111xxxxxxxxxx", AInstEmit.Fcvtzs_V_Fix, typeof(AOpCodeSimdShImm));
Set("x0011110xx111001000000xxxxxxxxxx", AInstEmit.Fcvtzu_S, typeof(AOpCodeSimdCvt));
Set("0x1011101x100001101110xxxxxxxxxx", AInstEmit.Fcvtzu_V, typeof(AOpCodeSimd));
Set("0x1011110>>xxxxx111111xxxxxxxxxx", AInstEmit.Fcvtzu_V_Fix, typeof(AOpCodeSimdShImm));
Set("x0011110xx011000xxxxxxxxxxxxxxxx", AInstEmit.Fcvtzs_Fix, typeof(AOpCodeSimdCvt));
Set("0x0011101x100001101110xxxxxxxxxx", AInstEmit.Fcvtzs_V, typeof(AOpCodeSimd));
Set("0x0011110>>xxxxx111111xxxxxxxxxx", AInstEmit.Fcvtzs_V, typeof(AOpCodeSimdShImm));
Set("x0011110xx111001000000xxxxxxxxxx", AInstEmit.Fcvtzu_S, typeof(AOpCodeSimdCvt));
Set("x0011110xx011001xxxxxxxxxxxxxxxx", AInstEmit.Fcvtzu_Fix, typeof(AOpCodeSimdCvt));
Set("0x1011101x100001101110xxxxxxxxxx", AInstEmit.Fcvtzu_V, typeof(AOpCodeSimd));
Set("0x1011110>>xxxxx111111xxxxxxxxxx", AInstEmit.Fcvtzu_V, typeof(AOpCodeSimdShImm));
Set("00011110xx1xxxxx000110xxxxxxxxxx", AInstEmit.Fdiv_S, typeof(AOpCodeSimdReg));
Set("00011111xx0xxxxx0xxxxxxxxxxxxxxx", AInstEmit.Fmadd_S, typeof(AOpCodeSimdReg));
Set("00011110xx1xxxxx010010xxxxxxxxxx", AInstEmit.Fmax_S, typeof(AOpCodeSimdReg));

14
Ryujinx/Cpu/Decoder/AOpCodeSimdShImm.cs
View file

@ -1,26 +1,16 @@
using ChocolArm64.Instruction;
using ChocolArm64.State;
namespace ChocolArm64.Decoder
{
class AOpCodeSimdShImm : AOpCode, IAOpCodeSimd
class AOpCodeSimdShImm : AOpCodeSimd
{
public int Rd { get; private set; }
public int Rn { get; private set; }
public int Imm { get; private set; }
public int Size { get; private set; }
public AOpCodeSimdShImm(AInst Inst, long Position, int OpCode) : base(Inst, Position)
public AOpCodeSimdShImm(AInst Inst, long Position, int OpCode) : base(Inst, Position, OpCode)
{
Rd = (OpCode >> 0) & 0x1f;
Rn = (OpCode >> 5) & 0x1f;
Imm = (OpCode >> 16) & 0x7f;
Size = ABitUtils.HighestBitSet32(Imm >> 3);
RegisterSize = ((OpCode >> 30) & 1) != 0
? ARegisterSize.SIMD128
: ARegisterSize.SIMD64;
}
}
}

265
Ryujinx/Cpu/Instruction/AInstEmitSimd.cs
View file

@ -102,80 +102,10 @@ namespace ChocolArm64.Instruction
public static void Eor_V(AILEmitterCtx Context) => EmitVectorBinaryZx(Context, OpCodes.Xor);
public static void Fadd_V(AILEmitterCtx Context)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
public static void Fadd_V(AILEmitterCtx Context) => EmitVectorBinaryFOp(Context, OpCodes.Add);
Context.EmitLdvec(Op.Rn);
Context.EmitLdvec(Op.Rm);
Context.EmitLdc_I4(Op.SizeF);
ASoftFallback.EmitCall(Context,
nameof(ASoftFallback.Fadd64),
nameof(ASoftFallback.Fadd128));
Context.EmitStvec(Op.Rd);
}
public static void Fcvtzs_V(AILEmitterCtx Context)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
Context.EmitLdvec(Op.Rn);
Context.EmitLdc_I4(0);
Context.EmitLdc_I4(Op.SizeF);
ASoftFallback.EmitCall(Context,
nameof(ASoftFallback.Fcvtzs_V64),
nameof(ASoftFallback.Fcvtzs_V128));
Context.EmitStvec(Op.Rd);
}
public static void Fcvtzs_V_Fix(AILEmitterCtx Context)
{
AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
Context.EmitLdvec(Op.Rn);
Context.EmitLdc_I4((8 << (Op.Size + 1)) - Op.Imm);
Context.EmitLdc_I4(Op.Size - 2);
ASoftFallback.EmitCall(Context,
nameof(ASoftFallback.Fcvtzs_V64),
nameof(ASoftFallback.Fcvtzs_V128));
Context.EmitStvec(Op.Rd);
}
public static void Fcvtzu_V(AILEmitterCtx Context)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
Context.EmitLdvec(Op.Rn);
Context.EmitLdc_I4(0);
Context.EmitLdc_I4(Op.SizeF);
ASoftFallback.EmitCall(Context,
nameof(ASoftFallback.Fcvtzu_V_64),
nameof(ASoftFallback.Fcvtzu_V_128));
Context.EmitStvec(Op.Rd);
}
public static void Fcvtzu_V_Fix(AILEmitterCtx Context)
{
AOpCodeSimdShImm Op = (AOpCodeSimdShImm)Context.CurrOp;
Context.EmitLdvec(Op.Rn);
Context.EmitLdc_I4((8 << (Op.Size + 1)) - Op.Imm);
Context.EmitLdc_I4(Op.Size - 2);
ASoftFallback.EmitCall(Context,
nameof(ASoftFallback.Fcvtzu_V_64),
nameof(ASoftFallback.Fcvtzu_V_128));
Context.EmitStvec(Op.Rd);
}
public static void Fcvtzs_V(AILEmitterCtx Context) => EmitVectorFcvtS(Context);
public static void Fcvtzu_V(AILEmitterCtx Context) => EmitVectorFcvtU(Context);
public static void Fmla_V(AILEmitterCtx Context)
{
@ -224,20 +154,7 @@ namespace ChocolArm64.Instruction
Context.EmitStvec(Op.Rd);
}
public static void Fmul_V(AILEmitterCtx Context)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
Context.EmitLdvec(Op.Rn);
Context.EmitLdvec(Op.Rm);
Context.EmitLdc_I4(Op.SizeF);
ASoftFallback.EmitCall(Context,
nameof(ASoftFallback.Fmul64),
nameof(ASoftFallback.Fmul128));
Context.EmitStvec(Op.Rd);
}
public static void Fmul_V(AILEmitterCtx Context) => EmitVectorBinaryFOp(Context, OpCodes.Mul);
public static void Fmul_Vs(AILEmitterCtx Context)
{
@ -255,20 +172,7 @@ namespace ChocolArm64.Instruction
Context.EmitStvec(Op.Rd);
}
public static void Fsub_V(AILEmitterCtx Context)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
Context.EmitLdvec(Op.Rn);
Context.EmitLdvec(Op.Rm);
Context.EmitLdc_I4(Op.SizeF);
ASoftFallback.EmitCall(Context,
nameof(ASoftFallback.Fsub64),
nameof(ASoftFallback.Fsub128));
Context.EmitStvec(Op.Rd);
}
public static void Fsub_V(AILEmitterCtx Context) => EmitVectorBinaryFOp(Context, OpCodes.Sub);
public static void Ins_Gp(AILEmitterCtx Context)
{
@ -870,6 +774,114 @@ namespace ChocolArm64.Instruction
});
}
private static void EmitVectorFcvtS(AILEmitterCtx Context)
{
EmitVectorCvtOp(Context, CvtDir.Fcvt, true);
}
private static void EmitVectorFcvtU(AILEmitterCtx Context)
{
EmitVectorCvtOp(Context, CvtDir.Fcvt, false);
}
private static void EmitVectorCvtfS(AILEmitterCtx Context)
{
EmitVectorCvtOp(Context, CvtDir.Cvtf, true);
}
private static void EmitVectorCvtfU(AILEmitterCtx Context)
{
EmitVectorCvtOp(Context, CvtDir.Cvtf, false);
}
private enum CvtDir
{
Fcvt,
Cvtf
}
private static void EmitVectorCvtOp(AILEmitterCtx Context, CvtDir Dir, bool Signed)
{
AOpCodeSimd Op = (AOpCodeSimd)Context.CurrOp;
int SizeF = Op.Size & 1;
int SizeI = SizeF + 2;
int FBits = 0;
if (Op is AOpCodeSimdShImm OpImm)
{
FBits = (8 << (Op.Size + 1)) - OpImm.Imm;
}
int Bytes = Context.CurrOp.GetBitsCount() >> 3;
for (int Index = 0; Index < (Bytes >> SizeI); Index++)
{
EmitVectorExtractF(Context, Op.Rn, Index, SizeF);
Context.EmitLdc_I4(FBits);
if (Dir == CvtDir.Fcvt)
{
//Float to Integer.
if (SizeF == 0)
{
ASoftFallback.EmitCall(Context, Signed
? nameof(ASoftFallback.SatSingleToInt32)
: nameof(ASoftFallback.SatSingleToUInt32));
}
else if (SizeF == 1)
{
ASoftFallback.EmitCall(Context, Signed
? nameof(ASoftFallback.SatDoubleToInt64)
: nameof(ASoftFallback.SatDoubleToUInt64));
}
}
else if (Dir == CvtDir.Cvtf)
{
//Integer to Float.
//TODO.
}
EmitVectorInsert(Context, Op.Rd, Index, SizeI);
}
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
private static void EmitVectorBinaryFOp(AILEmitterCtx Context, OpCode ILOp)
{
EmitVectorBinaryFOp(Context, () => Context.Emit(ILOp));
}
private static void EmitVectorBinaryFOp(AILEmitterCtx Context, Action Emit)
{
AOpCodeSimdReg Op = (AOpCodeSimdReg)Context.CurrOp;
int SizeF = Op.Size & 1;
int Bytes = Context.CurrOp.GetBitsCount() >> 3;
for (int Index = 0; Index < (Bytes >> SizeF + 2); Index++)
{
EmitVectorExtractF(Context, Op.Rn, Index, SizeF);
EmitVectorExtractF(Context, Op.Rm, Index, SizeF);
Emit();
EmitVectorInsertF(Context, Op.Rd, Index, SizeF);
}
if (Op.RegisterSize == ARegisterSize.SIMD64)
{
EmitVectorZeroUpper(Context, Op.Rd);
}
}
private static void EmitVectorUnarySx(AILEmitterCtx Context, OpCode ILOp)
{
EmitVectorUnarySx(Context, () => Context.Emit(ILOp));
@ -1047,6 +1059,25 @@ namespace ChocolArm64.Instruction
}
}
private static void EmitVectorExtractF(AILEmitterCtx Context, int Reg, int Index, int Size)
{
Context.EmitLdvec(Reg);
Context.EmitLdc_I4(Index);
if (Size == 0)
{
ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorExtractSingle));
}
else if (Size == 1)
{
ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorExtractDouble));
}
else
{
throw new ArgumentOutOfRangeException(nameof(Size));
}
}
private static void EmitVectorExtractSx(AILEmitterCtx Context, int Reg, int Index)
{
EmitVectorExtract(Context, Reg, Index, true);
@ -1070,6 +1101,27 @@ namespace ChocolArm64.Instruction
: nameof(ASoftFallback.ExtractVec));
}
private static void EmitVectorInsertF(AILEmitterCtx Context, int Reg, int Index, int Size)
{
Context.EmitLdvec(Reg);
Context.EmitLdc_I4(Index);
if (Size == 0)
{
ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorInsertSingle));
}
else if (Size == 1)
{
ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorInsertDouble));
}
else
{
throw new ArgumentOutOfRangeException(nameof(Size));
}
Context.EmitStvec(Reg);
}
private static void EmitVectorZeroLower(AILEmitterCtx Context, int Rd)
{
EmitVectorInsert(Context, Rd, 0, 3, 0);
@ -1080,6 +1132,17 @@ namespace ChocolArm64.Instruction
EmitVectorInsert(Context, Rd, 1, 3, 0);
}
private static void EmitVectorInsert(AILEmitterCtx Context, int Reg, int Index, int Size)
{
Context.EmitLdvec(Reg);
Context.EmitLdc_I4(Index);
Context.EmitLdc_I4(Size);
ASoftFallback.EmitCall(Context, nameof(ASoftFallback.VectorInsertInt));
Context.EmitStvec(Reg);
}
private static void EmitVectorInsert(AILEmitterCtx Context, int Reg, int Index, int Size, long Value)
{
Context.EmitLdvec(Reg);

229
Ryujinx/Cpu/Instruction/ASoftFallback.cs
View file

@ -365,122 +365,6 @@ namespace ChocolArm64.Instruction
return Res;
}
public static AVec Fadd64(AVec LHS, AVec RHS, int Size)
{
return Fadd(LHS, RHS, Size, 2);
}
public static AVec Fadd128(AVec LHS, AVec RHS, int Size)
{
return Fadd(LHS, RHS, Size, 4);
}
private static AVec Fadd(AVec LHS, AVec RHS, int Size, int Bytes)
{
AVec Res = new AVec();
int Elems = Bytes >> Size;
if (Size == 0)
{
for (int Index = 0; Index < Elems; Index++)
{
float L = LHS.ExtractSingle(Index);
float R = RHS.ExtractSingle(Index);
Res = AVec.InsertSingle(Res, Index, L + R);
}
}
else
{
for (int Index = 0; Index < Elems; Index++)
{
double L = LHS.ExtractDouble(Index);
double R = RHS.ExtractDouble(Index);
Res = AVec.InsertDouble(Res, Index, L + R);
}
}
return Res;
}
public static AVec Fcvtzs_V64(AVec Vector, int FBits, int Size)
{
return Fcvtzs_V(Vector, FBits, Size, 2);
}
public static AVec Fcvtzs_V128(AVec Vector, int FBits, int Size)
{
return Fcvtzs_V(Vector, FBits, Size, 4);
}
private static AVec Fcvtzs_V(AVec Vector, int FBits, int Size, int Bytes)
{
AVec Res = new AVec();
int Elems = Bytes >> Size;
if (Size == 0)
{
for (int Index = 0; Index < Elems; Index++)
{
float Value = Vector.ExtractSingle(Index);
Res = InsertSVec(Res, Index, Size + 2, SatSingleToInt32(Value, FBits));
}
}
else
{
for (int Index = 0; Index < Elems; Index++)
{
double Value = Vector.ExtractDouble(Index);
Res = InsertSVec(Res, Index, Size + 2, SatDoubleToInt64(Value, FBits));
}
}
return Res;
}
public static AVec Fcvtzu_V_64(AVec Vector, int FBits, int Size)
{
return Fcvtzu_V(Vector, FBits, Size, 2);
}
public static AVec Fcvtzu_V_128(AVec Vector, int FBits, int Size)
{
return Fcvtzu_V(Vector, FBits, Size, 4);
}
private static AVec Fcvtzu_V(AVec Vector, int FBits, int Size, int Bytes)
{
AVec Res = new AVec();
int Elems = Bytes >> Size;
if (Size == 0)
{
for (int Index = 0; Index < Elems; Index++)
{
float Value = Vector.ExtractSingle(Index);
Res = InsertVec(Res, Index, Size + 2, SatSingleToUInt32(Value, FBits));
}
}
else
{
for (int Index = 0; Index < Elems; Index++)
{
double Value = Vector.ExtractDouble(Index);
Res = InsertVec(Res, Index, Size + 2, SatDoubleToUInt64(Value, FBits));
}
}
return Res;
}
public static AVec Fmla64(AVec Res, AVec LHS, AVec RHS, int Size)
{
return Fmla(Res, LHS, RHS, Size, 2);
@ -568,46 +452,6 @@ namespace ChocolArm64.Instruction
return InsertVec(new AVec(), Elem, Size, Value);
}
public static AVec Fmul64(AVec LHS, AVec RHS, int Size)
{
return Fmul(LHS, RHS, Size, 2);
}
public static AVec Fmul128(AVec LHS, AVec RHS, int Size)
{
return Fmul(LHS, RHS, Size, 4);
}
private static AVec Fmul(AVec LHS, AVec RHS, int Size, int Bytes)
{
AVec Res = new AVec();
int Elems = Bytes >> Size;
if (Size == 0)
{
for (int Index = 0; Index < Elems; Index++)
{
float L = LHS.ExtractSingle(Index);
float R = RHS.ExtractSingle(Index);
Res = AVec.InsertSingle(Res, Index, L * R);
}
}
else
{
for (int Index = 0; Index < Elems; Index++)
{
double L = LHS.ExtractDouble(Index);
double R = RHS.ExtractDouble(Index);
Res = AVec.InsertDouble(Res, Index, L * R);
}
}
return Res;
}
public static AVec Fmul_Ve64(AVec LHS, AVec RHS, int SIdx, int Size)
{
return Fmul_Ve(LHS, RHS, SIdx, Size, 2);
@ -650,46 +494,6 @@ namespace ChocolArm64.Instruction
return Res;
}
public static AVec Fsub64(AVec LHS, AVec RHS, int Size)
{
return Fsub(LHS, RHS, Size, 2);
}
public static AVec Fsub128(AVec LHS, AVec RHS, int Size)
{
return Fsub(LHS, RHS, Size, 4);
}
private static AVec Fsub(AVec LHS, AVec RHS, int Size, int Bytes)
{
AVec Res = new AVec();
int Elems = Bytes >> Size;
if (Size == 0)
{
for (int Index = 0; Index < Elems; Index++)
{
float L = LHS.ExtractSingle(Index);
float R = RHS.ExtractSingle(Index);
Res = AVec.InsertSingle(Res, Index, L - R);
}
}
else
{
for (int Index = 0; Index < Elems; Index++)
{
double L = LHS.ExtractDouble(Index);
double R = RHS.ExtractDouble(Index);
Res = AVec.InsertDouble(Res, Index, L - R);
}
}
return Res;
}
public static AVec Ins_Gp(AVec Res, ulong Value, int Elem, int Size)
{
return InsertVec(Res, Elem, Size, Value);
@ -1126,6 +930,39 @@ namespace ChocolArm64.Instruction
throw new ArgumentOutOfRangeException(nameof(Size));
}
public static float VectorExtractSingle(AVec Vector, int Index)
{
return Vector.ExtractSingle(Index);
}
public static double VectorExtractDouble(AVec Vector, int Index)
{
return Vector.ExtractDouble(Index);
}
public static AVec VectorInsertSingle(float Value, AVec Vector, int Index)
{
return AVec.InsertSingle(Vector, Index, Value);
}
public static AVec VectorInsertDouble(double Value, AVec Vector, int Index)
{
return AVec.InsertDouble(Vector, Index, Value);
}
public static AVec VectorInsertInt(ulong Value, AVec Vector, int Index, int Size)
{
switch (Size)
{
case 0: return AVec.InsertByte (Vector, Index, (byte)Value);
case 1: return AVec.InsertUInt16(Vector, Index, (ushort)Value);
case 2: return AVec.InsertUInt32(Vector, Index, (uint)Value);
case 3: return AVec.InsertUInt64(Vector, Index, (ulong)Value);
}
throw new ArgumentOutOfRangeException(nameof(Size));
}
public static AVec InsertVec(AVec Vector, int Index, int Size, ulong Value)
{
switch (Size)

2
Ryujinx/Cpu/Memory/AMemoryMgr.cs
View file

@ -1,5 +1,3 @@
using ChocolArm64.Exceptions;
using System;
using System.Runtime.CompilerServices;
namespace ChocolArm64.Memory

2
Ryujinx/Logging.cs
View file

@ -1,8 +1,6 @@
using System;
using System.Diagnostics;
using System.IO;
using System.Reflection;
using System.Text;
namespace Ryujinx
{

7
Ryujinx/OsHle/Objects/AudIAudioOut.cs
View file

@ -1,15 +1,12 @@
using ChocolArm64.Memory;
using OpenTK.Audio;
using OpenTK.Audio.OpenAL;
using Ryujinx.OsHle.Handles;
using Ryujinx.OsHle.Ipc;
using static Ryujinx.OsHle.Objects.ObjHelper;
using System;
using System.Collections.Generic;
using System.IO;
using OpenTK.Audio;
using OpenTK.Audio.OpenAL;
namespace Ryujinx.OsHle.Objects
{
class AudIAudioOut