stuff/Ryujinx
stuff/Ryujinx
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Initial support for double precision shader instructions. (#963)

Browse source 
* Implement DADD, DFMA and DMUL shader instructions

* Rename FP to FP32

* Correct double immediate

* Classic mistake
This commit is contained in:
gdkchan 2020-03-03 11:02:08 -03:00 committed by GitHub
parent 3045c1a186
commit dc97457bf0
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
19 changed files with 428 additions and 184 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
Ryujinx.Graphics.Shader/CodeGen/Glsl/Declarations.cs
View file

@ -189,6 +189,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
{
case VariableType.Bool: return "bool";
case VariableType.F32: return "precise float";
case VariableType.F64: return "double";
case VariableType.S32: return "int";
case VariableType.U32: return "uint";
}

6
Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGen.cs
View file

@ -136,6 +136,9 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
case Instruction.Lod:
return InstGenMemory.Lod(context, operation);
case Instruction.PackDouble2x32:
return InstGenPacking.PackDouble2x32(context, operation);
case Instruction.PackHalf2x16:
return InstGenPacking.PackHalf2x16(context, operation);
@ -154,6 +157,9 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
case Instruction.TextureSize:
return InstGenMemory.TextureSize(context, operation);
case Instruction.UnpackDouble2x32:
return InstGenPacking.UnpackDouble2x32(context, operation);
case Instruction.UnpackHalf2x16:
return InstGenPacking.UnpackHalf2x16(context, operation);
}

4
Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenHelper.cs
View file

@ -50,6 +50,8 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
Add(Instruction.CompareLessU32, InstType.OpBinary, "<", 4);
Add(Instruction.CompareNotEqual, InstType.OpBinaryCom, "!=", 5);
Add(Instruction.ConditionalSelect, InstType.OpTernary, "?:", 12);
Add(Instruction.ConvertFP32ToFP64, InstType.CallUnary, "double");
Add(Instruction.ConvertFP64ToFP32, InstType.CallUnary, "float");
Add(Instruction.ConvertFPToS32, InstType.CallUnary, "int");
Add(Instruction.ConvertFPToU32, InstType.CallUnary, "uint");
Add(Instruction.ConvertS32ToFP, InstType.CallUnary, "float");
@ -83,6 +85,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
Add(Instruction.LogicalOr, InstType.OpBinaryCom, "||", 11);
Add(Instruction.LoopBreak, InstType.OpNullary, "break");
Add(Instruction.LoopContinue, InstType.OpNullary, "continue");
Add(Instruction.PackDouble2x32, InstType.Special);
Add(Instruction.PackHalf2x16, InstType.Special);
Add(Instruction.ShiftLeft, InstType.OpBinary, "<<", 3);
Add(Instruction.ShiftRightS32, InstType.OpBinary, ">>", 3);
@ -113,6 +116,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
Add(Instruction.TextureSample, InstType.Special);
Add(Instruction.TextureSize, InstType.Special);
Add(Instruction.Truncate, InstType.CallUnary, "trunc");
Add(Instruction.UnpackDouble2x32, InstType.Special);
Add(Instruction.UnpackHalf2x16, InstType.Special);
Add(Instruction.VoteAll, InstType.CallUnary, "allInvocationsARB");
Add(Instruction.VoteAllEqual, InstType.CallUnary, "allInvocationsEqualARB");

20
Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenPacking.cs
View file

@ -7,6 +7,17 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
{
static class InstGenPacking
{
public static string PackDouble2x32(CodeGenContext context, AstOperation operation)
{
IAstNode src0 = operation.GetSource(0);
IAstNode src1 = operation.GetSource(1);
string src0Expr = GetSoureExpr(context, src0, GetSrcVarType(operation.Inst, 0));
string src1Expr = GetSoureExpr(context, src1, GetSrcVarType(operation.Inst, 1));
return $"packDouble2x32(uvec2({src0Expr}, {src1Expr}))";
}
public static string PackHalf2x16(CodeGenContext context, AstOperation operation)
{
IAstNode src0 = operation.GetSource(0);
@ -18,6 +29,15 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
return $"packHalf2x16(vec2({src0Expr}, {src1Expr}))";
}
public static string UnpackDouble2x32(CodeGenContext context, AstOperation operation)
{
IAstNode src = operation.GetSource(0);
string srcExpr = GetSoureExpr(context, src, GetSrcVarType(operation.Inst, 0));
return $"unpackDouble2x32({srcExpr}){GetMask(operation.Index)}";
}
public static string UnpackHalf2x16(CodeGenContext context, AstOperation operation)
{
IAstNode src = operation.GetSource(0);

16
Ryujinx.Graphics.Shader/Decoders/DecoderHelper.cs
View file

@ -54,5 +54,21 @@ namespace Ryujinx.Graphics.Shader.Decoders
return BitConverter.Int32BitsToSingle(imm);
}
public static float DecodeD20Immediate(long opCode)
{
long imm = opCode.Extract(20, 19);
bool negate = opCode.Extract(56);
imm <<= 44;
if (negate)
{
imm |= 1L << 63;
}
return (float)BitConverter.Int64BitsToDouble(imm);
}
}
}

10
Ryujinx.Graphics.Shader/Decoders/FPType.cs
View file

@ -1,3 +1,5 @@
using Ryujinx.Graphics.Shader.IntermediateRepresentation;
namespace Ryujinx.Graphics.Shader.Decoders
{
enum FPType
@ -6,4 +8,12 @@ namespace Ryujinx.Graphics.Shader.Decoders
FP32 = 2,
FP64 = 3
}
static class FPTypeExtensions
{
public static Instruction ToInstFPType(this FPType type)
{
return type == FPType.FP64 ? Instruction.FP64 : Instruction.FP32;
}
}
}

14
Ryujinx.Graphics.Shader/Decoders/OpCodeDArithImm.cs Normal file
View file

@ -0,0 +1,14 @@
using Ryujinx.Graphics.Shader.Instructions;
namespace Ryujinx.Graphics.Shader.Decoders
{
class OpCodeDArithImm : OpCodeFArith, IOpCodeImmF
{
public float Immediate { get; }
public OpCodeDArithImm(InstEmitter emitter, ulong address, long opCode) : base(emitter, address, opCode)
{
Immediate = DecoderHelper.DecodeD20Immediate(opCode);
}
}
}

10
Ryujinx.Graphics.Shader/Decoders/OpCodeTable.cs
View file

@ -45,6 +45,16 @@ namespace Ryujinx.Graphics.Shader.Decoders
Set("111000110100xx", InstEmit.Brk, typeof(OpCodeBranchPop));
Set("111000100101xx", InstEmit.Brx, typeof(OpCodeBranchIndir));
Set("0101000010100x", InstEmit.Csetp, typeof(OpCodePset));
Set("0100110001110x", InstEmit.Dadd, typeof(OpCodeFArithCbuf));
Set("0011100x01110x", InstEmit.Dadd, typeof(OpCodeDArithImm));
Set("0101110001110x", InstEmit.Dadd, typeof(OpCodeFArithReg));
Set("010010110111xx", InstEmit.Dfma, typeof(OpCodeFArithCbuf));
Set("0011011x0111xx", InstEmit.Dfma, typeof(OpCodeDArithImm));
Set("010100110111xx", InstEmit.Dfma, typeof(OpCodeFArithRegCbuf));
Set("010110110111xx", InstEmit.Dfma, typeof(OpCodeFArithReg));
Set("0100110010000x", InstEmit.Dmul, typeof(OpCodeFArithCbuf));
Set("0011100x10000x", InstEmit.Dmul, typeof(OpCodeDArithImm));
Set("0101110010000x", InstEmit.Dmul, typeof(OpCodeFArithReg));
Set("111000110000xx", InstEmit.Exit, typeof(OpCodeExit));
Set("0100110010101x", InstEmit.F2F, typeof(OpCodeFArithCbuf));
Set("0011100x10101x", InstEmit.F2F, typeof(OpCodeFArithImm));

13
Ryujinx.Graphics.Shader/Instructions/InstEmitAluHelper.cs
View file

@ -85,12 +85,19 @@ namespace Ryujinx.Graphics.Shader.Instructions
context.Copy(GetNF(), context.ICompareLess(dest, Const(0)));
}
public static void SetFPZnFlags(EmitterContext context, Operand dest, bool setCC)
public static void SetFPZnFlags(EmitterContext context, Operand dest, bool setCC, Instruction fpType = Instruction.FP32)
{
if (setCC)
{
context.Copy(GetZF(), context.FPCompareEqual(dest, ConstF(0)));
context.Copy(GetNF(), context.FPCompareLess (dest, ConstF(0)));
Operand zero = ConstF(0);
if (fpType == Instruction.FP64)
{
zero = context.FP32ConvertToFP64(zero);
}
context.Copy(GetZF(), context.FPCompareEqual(dest, zero, fpType));
context.Copy(GetNF(), context.FPCompareLess (dest, zero, fpType));
}
}
}

33
Ryujinx.Graphics.Shader/Instructions/InstEmitConversion.cs
View file

@ -21,31 +21,43 @@ namespace Ryujinx.Graphics.Shader.Instructions
bool negateB = op.RawOpCode.Extract(45);
bool absoluteB = op.RawOpCode.Extract(49);
Operand srcB = context.FPAbsNeg(GetSrcB(context, srcType), absoluteB, negateB);
Operand srcB = context.FPAbsNeg(GetSrcB(context, srcType), absoluteB, negateB, srcType.ToInstFPType());
if (round)
if (round && srcType == dstType)
{
switch (op.RoundingMode)
{
case RoundingMode.ToNearest:
srcB = context.FPRound(srcB);
srcB = context.FPRound(srcB, srcType.ToInstFPType());
break;
case RoundingMode.TowardsNegativeInfinity:
srcB = context.FPFloor(srcB);
srcB = context.FPFloor(srcB, srcType.ToInstFPType());
break;
case RoundingMode.TowardsPositiveInfinity:
srcB = context.FPCeiling(srcB);
srcB = context.FPCeiling(srcB, srcType.ToInstFPType());
break;
case RoundingMode.TowardsZero:
srcB = context.FPTruncate(srcB);
srcB = context.FPTruncate(srcB, srcType.ToInstFPType());
break;
}
}
srcB = context.FPSaturate(srcB, op.Saturate);
// We don't need to handle conversions between FP16 <-> FP32
// since we do FP16 operations as FP32 directly.
// FP16 <-> FP64 conversions are invalid.
if (srcType == FPType.FP32 && dstType == FPType.FP64)
{
srcB = context.FP32ConvertToFP64(srcB);
}
else if (srcType == FPType.FP64 && dstType == FPType.FP32)
{
srcB = context.FP64ConvertToFP32(srcB);
}
srcB = context.FPSaturate(srcB, op.Saturate, dstType.ToInstFPType());
WriteFP(context, dstType, srcB);
@ -229,9 +241,12 @@ namespace Ryujinx.Graphics.Shader.Instructions
{
context.Copy(dest, context.PackHalf2x16(srcB, ConstF(0)));
}
else
else /* if (type == FPType.FP64) */
{
// TODO.
Operand dest2 = GetDest2(context);
context.Copy(dest, context.UnpackDouble2x32Low(srcB));
context.Copy(dest2, context.UnpackDouble2x32High(srcB));
}
}
}

204
Ryujinx.Graphics.Shader/Instructions/InstEmitFArith.cs
View file

@ -11,53 +11,12 @@ namespace Ryujinx.Graphics.Shader.Instructions
{
static partial class InstEmit
{
public static void Fadd(EmitterContext context)
{
IOpCodeFArith op = (IOpCodeFArith)context.CurrOp;
public static void Dadd(EmitterContext context) => EmitFPAdd(context, Instruction.FP64);
public static void Dfma(EmitterContext context) => EmitFPFma(context, Instruction.FP64);
public static void Dmul(EmitterContext context) => EmitFPMultiply(context, Instruction.FP64);
bool absoluteA = op.AbsoluteA, absoluteB, negateA, negateB;
if (op is OpCodeFArithImm32)
{
negateB = op.RawOpCode.Extract(53);
negateA = op.RawOpCode.Extract(56);
absoluteB = op.RawOpCode.Extract(57);
}
else
{
negateB = op.RawOpCode.Extract(45);
negateA = op.RawOpCode.Extract(48);
absoluteB = op.RawOpCode.Extract(49);
}
Operand srcA = context.FPAbsNeg(GetSrcA(context), absoluteA, negateA);
Operand srcB = context.FPAbsNeg(GetSrcB(context), absoluteB, negateB);
Operand dest = GetDest(context);
context.Copy(dest, context.FPSaturate(context.FPAdd(srcA, srcB), op.Saturate));
SetFPZnFlags(context, dest, op.SetCondCode);
}
public static void Ffma(EmitterContext context)
{
IOpCodeFArith op = (IOpCodeFArith)context.CurrOp;
bool negateB = op.RawOpCode.Extract(48);
bool negateC = op.RawOpCode.Extract(49);
Operand srcA = GetSrcA(context);
Operand srcB = context.FPNegate(GetSrcB(context), negateB);
Operand srcC = context.FPNegate(GetSrcC(context), negateC);
Operand dest = GetDest(context);
context.Copy(dest, context.FPSaturate(context.FPFusedMultiplyAdd(srcA, srcB, srcC), op.Saturate));
SetFPZnFlags(context, dest, op.SetCondCode);
}
public static void Fadd(EmitterContext context) => EmitFPAdd(context, Instruction.FP32);
public static void Ffma(EmitterContext context) => EmitFPFma(context, Instruction.FP32);
public static void Ffma32i(EmitterContext context)
{
@ -103,40 +62,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
SetFPZnFlags(context, dest, op.SetCondCode);
}
public static void Fmul(EmitterContext context)
{
IOpCodeFArith op = (IOpCodeFArith)context.CurrOp;
bool isImm32 = op is OpCodeFArithImm32;
bool negateB = !isImm32 && op.RawOpCode.Extract(48);
Operand srcA = GetSrcA(context);
Operand srcB = context.FPNegate(GetSrcB(context), negateB);
switch (op.Scale)
{
case FPMultiplyScale.None: break;
case FPMultiplyScale.Divide2: srcA = context.FPDivide (srcA, ConstF(2)); break;
case FPMultiplyScale.Divide4: srcA = context.FPDivide (srcA, ConstF(4)); break;
case FPMultiplyScale.Divide8: srcA = context.FPDivide (srcA, ConstF(8)); break;
case FPMultiplyScale.Multiply2: srcA = context.FPMultiply(srcA, ConstF(2)); break;
case FPMultiplyScale.Multiply4: srcA = context.FPMultiply(srcA, ConstF(4)); break;
case FPMultiplyScale.Multiply8: srcA = context.FPMultiply(srcA, ConstF(8)); break;
default: break; //TODO: Warning.
}
Operand dest = GetDest(context);
bool saturate = isImm32 ? op.RawOpCode.Extract(55) : op.Saturate;
context.Copy(dest, context.FPSaturate(context.FPMultiply(srcA, srcB), saturate));
SetFPZnFlags(context, dest, op.SetCondCode);
}
public static void Fmul(EmitterContext context) => EmitFPMultiply(context, Instruction.FP32);
public static void Fset(EmitterContext context)
{
@ -406,6 +332,107 @@ namespace Ryujinx.Graphics.Shader.Instructions
context.Copy(GetDest(context), context.FPSaturate(res, op.Saturate));
}
private static void EmitFPAdd(EmitterContext context, Instruction fpType)
{
IOpCodeFArith op = (IOpCodeFArith)context.CurrOp;
bool isFP64 = fpType == Instruction.FP64;
bool absoluteA = op.AbsoluteA, absoluteB, negateA, negateB;
if (op is OpCodeFArithImm32)
{
negateB = op.RawOpCode.Extract(53);
negateA = op.RawOpCode.Extract(56);
absoluteB = op.RawOpCode.Extract(57);
}
else
{
negateB = op.RawOpCode.Extract(45);
negateA = op.RawOpCode.Extract(48);
absoluteB = op.RawOpCode.Extract(49);
}
Operand srcA = context.FPAbsNeg(GetSrcA(context, isFP64), absoluteA, negateA, fpType);
Operand srcB = context.FPAbsNeg(GetSrcB(context, isFP64), absoluteB, negateB, fpType);
Operand res = context.FPSaturate(context.FPAdd(srcA, srcB, fpType), op.Saturate, fpType);
SetDest(context, res, isFP64);
SetFPZnFlags(context, res, op.SetCondCode, fpType);
}
private static void EmitFPFma(EmitterContext context, Instruction fpType)
{
IOpCodeFArith op = (IOpCodeFArith)context.CurrOp;
bool isFP64 = fpType == Instruction.FP64;
bool negateB = op.RawOpCode.Extract(48);
bool negateC = op.RawOpCode.Extract(49);
Operand srcA = GetSrcA(context, isFP64);
Operand srcB = context.FPNegate(GetSrcB(context, isFP64), negateB, fpType);
Operand srcC = context.FPNegate(GetSrcC(context, isFP64), negateC, fpType);
Operand res = context.FPSaturate(context.FPFusedMultiplyAdd(srcA, srcB, srcC, fpType), op.Saturate, fpType);
SetDest(context, res, isFP64);
SetFPZnFlags(context, res, op.SetCondCode, fpType);
}
private static void EmitFPMultiply(EmitterContext context, Instruction fpType)
{
IOpCodeFArith op = (IOpCodeFArith)context.CurrOp;
bool isFP64 = fpType == Instruction.FP64;
bool isImm32 = op is OpCodeFArithImm32;
bool negateB = !isImm32 && op.RawOpCode.Extract(48);
Operand srcA = GetSrcA(context, isFP64);
Operand srcB = context.FPNegate(GetSrcB(context, isFP64), negateB, fpType);
if (op.Scale != FPMultiplyScale.None)
{
Operand scale = op.Scale switch
{
FPMultiplyScale.Divide2 => ConstF(0.5f),
FPMultiplyScale.Divide4 => ConstF(0.25f),
FPMultiplyScale.Divide8 => ConstF(0.125f),
FPMultiplyScale.Multiply2 => ConstF(2f),
FPMultiplyScale.Multiply4 => ConstF(4f),
FPMultiplyScale.Multiply8 => ConstF(8f),
_ => ConstF(1) // Invalid, behave as if it had no scale.
};
if (scale.AsFloat() == 1)
{
context.Config.PrintLog($"Invalid FP multiply scale \"{op.Scale}\".");
}
if (isFP64)
{
scale = context.FP32ConvertToFP64(scale);
}
srcA = context.FPMultiply(srcA, scale, fpType);
}
bool saturate = isImm32 ? op.RawOpCode.Extract(55) : op.Saturate;
Operand res = context.FPSaturate(context.FPMultiply(srcA, srcB, fpType), saturate, fpType);
SetDest(context, res, isFP64);
SetFPZnFlags(context, res, op.SetCondCode, fpType);
}
private static Operand GetFPComparison(
EmitterContext context,
Condition cond,
@ -447,7 +474,7 @@ namespace Ryujinx.Graphics.Shader.Instructions
default: throw new InvalidOperationException($"Unexpected condition \"{cond}\".");
}
res = context.Add(inst | Instruction.FP, Local(), srcA, srcB);
res = context.Add(inst | Instruction.FP32, Local(), srcA, srcB);
if ((cond & Condition.Nan) != 0)
{
@ -483,5 +510,20 @@ namespace Ryujinx.Graphics.Shader.Instructions
return FPAbsNeg(context, operands, false, op.NegateC);
}
private static void SetDest(EmitterContext context, Operand value, bool isFP64)
{
if (isFP64)
{
IOpCodeRd op = (IOpCodeRd)context.CurrOp;
context.Copy(Register(op.Rd.Index, op.Rd.Type), context.UnpackDouble2x32Low(value));
context.Copy(Register(op.Rd.Index | 1, op.Rd.Type), context.UnpackDouble2x32High(value));
}
else
{
context.Copy(GetDest(context), value);
}
}
}
}

96
Ryujinx.Graphics.Shader/Instructions/InstEmitHelper.cs
View file

@ -34,9 +34,25 @@ namespace Ryujinx.Graphics.Shader.Instructions
return Register(((IOpCodeRd)context.CurrOp).Rd);
}
public static Operand GetSrcA(EmitterContext context)
public static Operand GetDest2(EmitterContext context)
{
return Register(((IOpCodeRa)context.CurrOp).Ra);
Register rd = ((IOpCodeRd)context.CurrOp).Rd;
return Register(rd.Index | 1, rd.Type);
}
public static Operand GetSrcA(EmitterContext context, bool isFP64 = false)
{
IOpCodeRa op = (IOpCodeRa)context.CurrOp;
if (isFP64)
{
return context.PackDouble2x32(Register(op.Ra.Index, op.Ra.Type), Register(op.Ra.Index | 1, op.Ra.Type));
}
else
{
return Register(op.Ra);
}
}
public static Operand GetSrcB(EmitterContext context, FPType floatType)
@ -53,46 +69,78 @@ namespace Ryujinx.Graphics.Shader.Instructions
}
else if (floatType == FPType.FP64)
{
// TODO: Double floating-point type support.
return GetSrcB(context, true);
}
context.Config.PrintLog($"Invalid floating point type: {floatType}.");
return ConstF(0);
throw new ArgumentException($"Invalid floating point type \"{floatType}\".");
}
public static Operand GetSrcB(EmitterContext context)
public static Operand GetSrcB(EmitterContext context, bool isFP64 = false)
{
switch (context.CurrOp)
if (isFP64)
{
case IOpCodeCbuf op:
return Cbuf(op.Slot, op.Offset);
switch (context.CurrOp)
{
case IOpCodeCbuf op:
return context.PackDouble2x32(Cbuf(op.Slot, op.Offset), Cbuf(op.Slot, op.Offset + 1));
case IOpCodeImm op:
return Const(op.Immediate);
case IOpCodeImmF op:
return context.FP32ConvertToFP64(ConstF(op.Immediate));
case IOpCodeImmF op:
return ConstF(op.Immediate);
case IOpCodeReg op:
return context.PackDouble2x32(Register(op.Rb.Index, op.Rb.Type), Register(op.Rb.Index | 1, op.Rb.Type));
case IOpCodeReg op:
return Register(op.Rb);
case IOpCodeRegCbuf op:
return context.PackDouble2x32(Register(op.Rc.Index, op.Rc.Type), Register(op.Rc.Index | 1, op.Rc.Type));
}
}
else
{
switch (context.CurrOp)
{
case IOpCodeCbuf op:
return Cbuf(op.Slot, op.Offset);
case IOpCodeRegCbuf op:
return Register(op.Rc);
case IOpCodeImm op:
return Const(op.Immediate);
case IOpCodeImmF op:
return ConstF(op.Immediate);
case IOpCodeReg op:
return Register(op.Rb);
case IOpCodeRegCbuf op:
return Register(op.Rc);
}
}
throw new InvalidOperationException($"Unexpected opcode type \"{context.CurrOp.GetType().Name}\".");
}
public static Operand GetSrcC(EmitterContext context)
public static Operand GetSrcC(EmitterContext context, bool isFP64 = false)
{
switch (context.CurrOp)
if (isFP64)
{
case IOpCodeRegCbuf op:
return Cbuf(op.Slot, op.Offset);
switch (context.CurrOp)
{
case IOpCodeRegCbuf op:
return context.PackDouble2x32(Cbuf(op.Slot, op.Offset), Cbuf(op.Slot, op.Offset + 1));
case IOpCodeRc op:
return Register(op.Rc);
case IOpCodeRc op:
return context.PackDouble2x32(Register(op.Rc.Index, op.Rc.Type), Register(op.Rc.Index | 1, op.Rc.Type));
}
}
else
{
switch (context.CurrOp)
{
case IOpCodeRegCbuf op:
return Cbuf(op.Slot, op.Offset);
case IOpCodeRc op:
return Register(op.Rc);
}
}
throw new InvalidOperationException($"Unexpected opcode type \"{context.CurrOp.GetType().Name}\".");

7
Ryujinx.Graphics.Shader/IntermediateRepresentation/Instruction.cs
View file

@ -46,6 +46,8 @@ namespace Ryujinx.Graphics.Shader.IntermediateRepresentation
CompareLessU32,
CompareNotEqual,
ConditionalSelect,
ConvertFP32ToFP64,
ConvertFP64ToFP32,
ConvertFPToS32,
ConvertFPToU32,
ConvertS32ToFP,
@ -122,9 +124,10 @@ namespace Ryujinx.Graphics.Shader.IntermediateRepresentation
Count,
FP = 1 << 16,
FP32 = 1 << 16,
FP64 = 1 << 17,
MrShift = 17,
MrShift = 18,
MrGlobal = 0 << MrShift,
MrShared = 1 << MrShift,

29
Ryujinx.Graphics.Shader/StructuredIr/InstructionInfo.cs
View file

@ -49,7 +49,7 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
Add(Instruction.BitwiseOr, VariableType.Int, VariableType.Int, VariableType.Int);
Add(Instruction.BranchIfTrue, VariableType.None, VariableType.Bool);
Add(Instruction.BranchIfFalse, VariableType.None, VariableType.Bool);
Add(Instruction.Ceiling, VariableType.F32, VariableType.F32, VariableType.F32);
Add(Instruction.Ceiling, VariableType.Scalar, VariableType.Scalar, VariableType.Scalar);
Add(Instruction.Clamp, VariableType.Scalar, VariableType.Scalar, VariableType.Scalar, VariableType.Scalar);
Add(Instruction.ClampU32, VariableType.U32, VariableType.U32, VariableType.U32, VariableType.U32);
Add(Instruction.CompareEqual, VariableType.Bool, VariableType.Scalar, VariableType.Scalar);
@ -63,6 +63,8 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
Add(Instruction.CompareLessU32, VariableType.Bool, VariableType.U32, VariableType.U32);
Add(Instruction.CompareNotEqual, VariableType.Bool, VariableType.Scalar, VariableType.Scalar);
Add(Instruction.ConditionalSelect, VariableType.Scalar, VariableType.Bool, VariableType.Scalar, VariableType.Scalar);
Add(Instruction.ConvertFP32ToFP64, VariableType.F64, VariableType.F32);
Add(Instruction.ConvertFP64ToFP32, VariableType.F32, VariableType.F64);
Add(Instruction.ConvertFPToS32, VariableType.S32, VariableType.F32);
Add(Instruction.ConvertFPToU32, VariableType.U32, VariableType.F32);
Add(Instruction.ConvertS32ToFP, VariableType.F32, VariableType.S32);
@ -74,8 +76,8 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
Add(Instruction.ExponentB2, VariableType.Scalar, VariableType.Scalar);
Add(Instruction.FindFirstSetS32, VariableType.S32, VariableType.S32);
Add(Instruction.FindFirstSetU32, VariableType.S32, VariableType.U32);
Add(Instruction.Floor, VariableType.F32, VariableType.F32);
Add(Instruction.FusedMultiplyAdd, VariableType.F32, VariableType.F32, VariableType.F32, VariableType.F32);
Add(Instruction.Floor, VariableType.Scalar, VariableType.Scalar);
Add(Instruction.FusedMultiplyAdd, VariableType.Scalar, VariableType.Scalar, VariableType.Scalar, VariableType.Scalar);
Add(Instruction.ImageLoad, VariableType.F32);
Add(Instruction.ImageStore, VariableType.None);
Add(Instruction.IsNan, VariableType.Bool, VariableType.F32);
@ -106,9 +108,10 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
Add(Instruction.MultiplyHighS32, VariableType.S32, VariableType.S32, VariableType.S32);
Add(Instruction.MultiplyHighU32, VariableType.U32, VariableType.U32, VariableType.U32);
Add(Instruction.Negate, VariableType.Scalar, VariableType.Scalar);
Add(Instruction.PackDouble2x32, VariableType.F64, VariableType.U32, VariableType.U32);
Add(Instruction.PackHalf2x16, VariableType.U32, VariableType.F32, VariableType.F32);
Add(Instruction.ReciprocalSquareRoot, VariableType.Scalar, VariableType.Scalar);
Add(Instruction.Round, VariableType.F32, VariableType.F32);
Add(Instruction.Round, VariableType.Scalar, VariableType.Scalar);
Add(Instruction.Sine, VariableType.Scalar, VariableType.Scalar);
Add(Instruction.SquareRoot, VariableType.Scalar, VariableType.Scalar);
Add(Instruction.StoreGlobal, VariableType.None, VariableType.S32, VariableType.S32, VariableType.U32);
@ -119,7 +122,8 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
Add(Instruction.SwizzleAdd, VariableType.F32, VariableType.F32, VariableType.F32, VariableType.S32);
Add(Instruction.TextureSample, VariableType.F32);
Add(Instruction.TextureSize, VariableType.S32, VariableType.S32, VariableType.S32);
Add(Instruction.Truncate, VariableType.F32, VariableType.F32);
Add(Instruction.Truncate, VariableType.Scalar, VariableType.Scalar);
Add(Instruction.UnpackDouble2x32, VariableType.U32, VariableType.F64);
Add(Instruction.UnpackHalf2x16, VariableType.F32, VariableType.U32);
Add(Instruction.VoteAll, VariableType.Bool, VariableType.Bool);
Add(Instruction.VoteAllEqual, VariableType.Bool, VariableType.Bool);
@ -155,9 +159,18 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
{
if (type == VariableType.Scalar)
{
return (inst & Instruction.FP) != 0
? VariableType.F32
: VariableType.S32;
if ((inst & Instruction.FP32) != 0)
{
return VariableType.F32;
}
else if ((inst & Instruction.FP64) != 0)
{
return VariableType.F64;
}
else
{
return VariableType.S32;
}
}
else if (type == VariableType.Int)
{

2
Ryujinx.Graphics.Shader/StructuredIr/StructuredProgram.cs
View file

@ -104,7 +104,7 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
if (isCondSel && type == VariableType.F32)
{
inst |= Instruction.FP;
inst |= Instruction.FP32;
}
dest.VarType = type;

1
Ryujinx.Graphics.Shader/StructuredIr/VariableType.cs
View file

@ -7,6 +7,7 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
Scalar,
Int,
F32,
F64,
S32,
U32
}

116
Ryujinx.Graphics.Shader/Translation/EmitterContextInsts.cs
View file

@ -181,44 +181,54 @@ namespace Ryujinx.Graphics.Shader.Translation
return context.Add(Instruction.FindFirstSetU32, Local(), a);
}
public static Operand FPAbsNeg(this EmitterContext context, Operand a, bool abs, bool neg)
public static Operand FP32ConvertToFP64(this EmitterContext context, Operand a)
{
return context.FPNegate(context.FPAbsolute(a, abs), neg);
return context.Add(Instruction.ConvertFP32ToFP64, Local(), a);
}
public static Operand FPAbsolute(this EmitterContext context, Operand a, bool abs)
public static Operand FP64ConvertToFP32(this EmitterContext context, Operand a)
{
return context.Add(Instruction.ConvertFP64ToFP32, Local(), a);
}
public static Operand FPAbsNeg(this EmitterContext context, Operand a, bool abs, bool neg, Instruction fpType = Instruction.FP32)
{
return context.FPNegate(context.FPAbsolute(a, abs, fpType), neg, fpType);
}
public static Operand FPAbsolute(this EmitterContext context, Operand a, bool abs, Instruction fpType = Instruction.FP32)
{
if (abs)
{
a = context.FPAbsolute(a);
a = context.FPAbsolute(a, fpType);
}
return a;
}
public static Operand FPAbsolute(this EmitterContext context, Operand a)
public static Operand FPAbsolute(this EmitterContext context, Operand a, Instruction fpType = Instruction.FP32)
{
return context.Add(Instruction.FP | Instruction.Absolute, Local(), a);
return context.Add(fpType | Instruction.Absolute, Local(), a);
}
public static Operand FPAdd(this EmitterContext context, Operand a, Operand b)
public static Operand FPAdd(this EmitterContext context, Operand a, Operand b, Instruction fpType = Instruction.FP32)
{
return context.Add(Instruction.FP | Instruction.Add, Local(), a, b);
return context.Add(fpType | Instruction.Add, Local(), a, b);
}
public static Operand FPCeiling(this EmitterContext context, Operand a)
public static Operand FPCeiling(this EmitterContext context, Operand a, Instruction fpType = Instruction.FP32)
{
return context.Add(Instruction.FP | Instruction.Ceiling, Local(), a);
return context.Add(fpType | Instruction.Ceiling, Local(), a);
}
public static Operand FPCompareEqual(this EmitterContext context, Operand a, Operand b)
public static Operand FPCompareEqual(this EmitterContext context, Operand a, Operand b, Instruction fpType = Instruction.FP32)
{
return context.Add(Instruction.FP | Instruction.CompareEqual, Local(), a, b);
return context.Add(fpType | Instruction.CompareEqual, Local(), a, b);
}
public static Operand FPCompareLess(this EmitterContext context, Operand a, Operand b)
public static Operand FPCompareLess(this EmitterContext context, Operand a, Operand b, Instruction fpType = Instruction.FP32)
{
return context.Add(Instruction.FP | Instruction.CompareLess, Local(), a, b);
return context.Add(fpType | Instruction.CompareLess, Local(), a, b);
}
public static Operand FPConvertToS32(this EmitterContext context, Operand a)
@ -233,62 +243,62 @@ namespace Ryujinx.Graphics.Shader.Translation
public static Operand FPCosine(this EmitterContext context, Operand a)
{
return context.Add(Instruction.FP | Instruction.Cosine, Local(), a);
return context.Add(Instruction.FP32 | Instruction.Cosine, Local(), a);
}
public static Operand FPDivide(this EmitterContext context, Operand a, Operand b)
{
return context.Add(Instruction.FP | Instruction.Divide, Local(), a, b);
return context.Add(Instruction.FP32 | Instruction.Divide, Local(), a, b);
}
public static Operand FPExponentB2(this EmitterContext context, Operand a)
{
return context.Add(Instruction.FP | Instruction.ExponentB2, Local(), a);
return context.Add(Instruction.FP32 | Instruction.ExponentB2, Local(), a);
}
public static Operand FPFloor(this EmitterContext context, Operand a)
public static Operand FPFloor(this EmitterContext context, Operand a, Instruction fpType = Instruction.FP32)
{
return context.Add(Instruction.FP | Instruction.Floor, Local(), a);
return context.Add(fpType | Instruction.Floor, Local(), a);
}
public static Operand FPFusedMultiplyAdd(this EmitterContext context, Operand a, Operand b, Operand c)
public static Operand FPFusedMultiplyAdd(this EmitterContext context, Operand a, Operand b, Operand c, Instruction fpType = Instruction.FP32)
{
return context.Add(Instruction.FusedMultiplyAdd, Local(), a, b, c);
return context.Add(fpType | Instruction.FusedMultiplyAdd, Local(), a, b, c);
}
public static Operand FPLogarithmB2(this EmitterContext context, Operand a)
{
return context.Add(Instruction.FP | Instruction.LogarithmB2, Local(), a);
return context.Add(Instruction.FP32 | Instruction.LogarithmB2, Local(), a);
}
public static Operand FPMaximum(this EmitterContext context, Operand a, Operand b)
{
return context.Add(Instruction.FP | Instruction.Maximum, Local(), a, b);
return context.Add(Instruction.FP32 | Instruction.Maximum, Local(), a, b);
}
public static Operand FPMinimum(this EmitterContext context, Operand a, Operand b)
{
return context.Add(Instruction.FP | Instruction.Minimum, Local(), a, b);
return context.Add(Instruction.FP32 | Instruction.Minimum, Local(), a, b);
}
public static Operand FPMultiply(this EmitterContext context, Operand a, Operand b)
public static Operand FPMultiply(this EmitterContext context, Operand a, Operand b, Instruction fpType = Instruction.FP32)
{
return context.Add(Instruction.FP | Instruction.Multiply, Local(), a, b);
return context.Add(fpType | Instruction.Multiply, Local(), a, b);
}
public static Operand FPNegate(this EmitterContext context, Operand a, bool neg)
public static Operand FPNegate(this EmitterContext context, Operand a, bool neg, Instruction fpType = Instruction.FP32)
{
if (neg)
{
a = context.FPNegate(a);
a = context.FPNegate(a, fpType);
}
return a;
}
public static Operand FPNegate(this EmitterContext context, Operand a)
public static Operand FPNegate(this EmitterContext context, Operand a, Instruction fpType = Instruction.FP32)
{
return context.Add(Instruction.FP | Instruction.Negate, Local(), a);
return context.Add(fpType | Instruction.Negate, Local(), a);
}
public static Operand FPReciprocal(this EmitterContext context, Operand a)
@ -298,42 +308,42 @@ namespace Ryujinx.Graphics.Shader.Translation
public static Operand FPReciprocalSquareRoot(this EmitterContext context, Operand a)
{
return context.Add(Instruction.FP | Instruction.ReciprocalSquareRoot, Local(), a);
return context.Add(Instruction.FP32 | Instruction.ReciprocalSquareRoot, Local(), a);
}
public static Operand FPRound(this EmitterContext context, Operand a)
public static Operand FPRound(this EmitterContext context, Operand a, Instruction fpType = Instruction.FP32)
{
return context.Add(Instruction.FP | Instruction.Round, Local(), a);
return context.Add(fpType | Instruction.Round, Local(), a);
}
public static Operand FPSaturate(this EmitterContext context, Operand a, bool sat)
public static Operand FPSaturate(this EmitterContext context, Operand a, bool sat, Instruction fpType = Instruction.FP32)
{
if (sat)
{
a = context.FPSaturate(a);
a = context.FPSaturate(a, fpType);
}
return a;
}
public static Operand FPSaturate(this EmitterContext context, Operand a)
public static Operand FPSaturate(this EmitterContext context, Operand a, Instruction fpType = Instruction.FP32)
{
return context.Add(Instruction.FP | Instruction.Clamp, Local(), a, ConstF(0), ConstF(1));
return context.Add(fpType | Instruction.Clamp, Local(), a, ConstF(0), ConstF(1));
}
public static Operand FPSine(this EmitterContext context, Operand a)
{
return context.Add(Instruction.FP | Instruction.Sine, Local(), a);
return context.Add(Instruction.FP32 | Instruction.Sine, Local(), a);
}
public static Operand FPSquareRoot(this EmitterContext context, Operand a)
{
return context.Add(Instruction.FP | Instruction.SquareRoot, Local(), a);
return context.Add(Instruction.FP32 | Instruction.SquareRoot, Local(), a);
}
public static Operand FPTruncate(this EmitterContext context, Operand a)
public static Operand FPTruncate(this EmitterContext context, Operand a, Instruction fpType = Instruction.FP32)
{
return context.Add(Instruction.Truncate, Local(), a);
return context.Add(fpType | Instruction.Truncate, Local(), a);
}
public static Operand FPSwizzleAdd(this EmitterContext context, Operand a, Operand b, int mask)
@ -501,6 +511,11 @@ namespace Ryujinx.Graphics.Shader.Translation
return context.Add(Instruction.MultiplyHighU32, Local(), a, b);
}
public static Operand PackDouble2x32(this EmitterContext context, Operand a, Operand b)
{
return context.Add(Instruction.PackDouble2x32, Local(), a, b);
}
public static Operand PackHalf2x16(this EmitterContext context, Operand a, Operand b)
{
return context.Add(Instruction.PackHalf2x16, Local(), a, b);
@ -563,6 +578,25 @@ namespace Ryujinx.Graphics.Shader.Translation
return context.Add(Instruction.StoreShared, null, a, b);
}
public static Operand UnpackDouble2x32High(this EmitterContext context, Operand a)
{
return UnpackDouble2x32(context, a, 1);
}
public static Operand UnpackDouble2x32Low(this EmitterContext context, Operand a)
{
return UnpackDouble2x32(context, a, 0);
}
private static Operand UnpackDouble2x32(this EmitterContext context, Operand a, int index)
{
Operand dest = Local();
context.Add(new Operation(Instruction.UnpackDouble2x32, index, dest, a));
return dest;
}
public static Operand UnpackHalf2x16High(this EmitterContext context, Operand a)
{
return UnpackHalf2x16(context, a, 1);

6
Ryujinx.Graphics.Shader/Translation/Lowering.cs
View file

@ -304,7 +304,7 @@ namespace Ryujinx.Graphics.Shader.Translation
Operand coordNormalized = Local();
node.List.AddBefore(node, new Operation(Instruction.FP | Instruction.Divide, coordNormalized, source, Float(coordSize)));
node.List.AddBefore(node, new Operation(Instruction.FP32 | Instruction.Divide, coordNormalized, source, Float(coordSize)));
sources[coordsIndex + index] = coordNormalized;
}
@ -375,13 +375,13 @@ namespace Ryujinx.Graphics.Shader.Translation
Operand intOffset = offsets[index + (hasOffsets ? texOp.Index * coordsCount : 0)];
node.List.AddBefore(node, new Operation(Instruction.FP | Instruction.Divide, offset, Float(intOffset), Float(coordSize)));
node.List.AddBefore(node, new Operation(Instruction.FP32 | Instruction.Divide, offset, Float(intOffset), Float(coordSize)));
Operand source = sources[coordsIndex + index];
Operand coordPlusOffset = Local();
node.List.AddBefore(node, new Operation(Instruction.FP | Instruction.Add, coordPlusOffset, source, offset));
node.List.AddBefore(node, new Operation(Instruction.FP32 | Instruction.Add, coordPlusOffset, source, offset));
sources[coordsIndex + index] = coordPlusOffset;
}

24
Ryujinx.Graphics.Shader/Translation/Optimizations/ConstantFolding.cs
View file

@ -101,51 +101,51 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
EvaluateBinary(operation, (x, y) => y != 0 ? x / y : 0);
break;
case Instruction.FP | Instruction.Add:
case Instruction.FP32 | Instruction.Add:
EvaluateFPBinary(operation, (x, y) => x + y);
break;
case Instruction.FP | Instruction.Clamp:
case Instruction.FP32 | Instruction.Clamp:
EvaluateFPTernary(operation, (x, y, z) => Math.Clamp(x, y, z));
break;
case Instruction.FP | Instruction.CompareEqual:
case Instruction.FP32 | Instruction.CompareEqual:
EvaluateFPBinary(operation, (x, y) => x == y);
break;
case Instruction.FP | Instruction.CompareGreater:
case Instruction.FP32 | Instruction.CompareGreater:
EvaluateFPBinary(operation, (x, y) => x > y);
break;
case Instruction.FP | Instruction.CompareGreaterOrEqual:
case Instruction.FP32 | Instruction.CompareGreaterOrEqual:
EvaluateFPBinary(operation, (x, y) => x >= y);
break;
case Instruction.FP | Instruction.CompareLess:
case Instruction.FP32 | Instruction.CompareLess:
EvaluateFPBinary(operation, (x, y) => x < y);
break;
case Instruction.FP | Instruction.CompareLessOrEqual:
case Instruction.FP32 | Instruction.CompareLessOrEqual:
EvaluateFPBinary(operation, (x, y) => x <= y);
break;
case Instruction.FP | Instruction.CompareNotEqual:
case Instruction.FP32 | Instruction.CompareNotEqual:
EvaluateFPBinary(operation, (x, y) => x != y);
break;
case Instruction.FP | Instruction.Divide:
case Instruction.FP32 | Instruction.Divide:
EvaluateFPBinary(operation, (x, y) => x / y);
break;
case Instruction.FP | Instruction.Multiply:
case Instruction.FP32 | Instruction.Multiply:
EvaluateFPBinary(operation, (x, y) => x * y);
break;
case Instruction.FP | Instruction.Negate:
case Instruction.FP32 | Instruction.Negate:
EvaluateFPUnary(operation, (x) => -x);
break;
case Instruction.FP | Instruction.Subtract:
case Instruction.FP32 | Instruction.Subtract:
EvaluateFPBinary(operation, (x, y) => x - y);
break;