public static void Mov(EmitterContext context)
{
OpCodeAlu op = (OpCodeAlu)context.CurrOp;
context.Copy(GetDest(context), GetSrcB(context));
}
public static void F2I(EmitterContext context)
{
OpCodeFArith op = (OpCodeFArith)context.CurrOp;
IntegerType intType = (IntegerType)op.RawOpCode.Extract(8, 2);
if (intType == IntegerType.U64)
{
context.Config.PrintLog("Unimplemented 64-bits F2I.");
return;
}
bool isSmallInt = intType <= IntegerType.U16;
FPType floatType = (FPType)op.RawOpCode.Extract(10, 2);
bool isSignedInt = op.RawOpCode.Extract(12);
bool negateB = op.RawOpCode.Extract(45);
bool absoluteB = op.RawOpCode.Extract(49);
if (isSignedInt)
{
intType |= IntegerType.S8;
}
Operand srcB = context.FPAbsNeg(GetSrcB(context, floatType), absoluteB, negateB);
switch (op.RoundingMode)
{
case RoundingMode.ToNearest:
srcB = context.FPRound(srcB);
break;
case RoundingMode.TowardsNegativeInfinity:
srcB = context.FPFloor(srcB);
break;
case RoundingMode.TowardsPositiveInfinity:
srcB = context.FPCeiling(srcB);
break;
case RoundingMode.TowardsZero:
srcB = context.FPTruncate(srcB);
break;
}
if (!isSignedInt)
{
// Negative float to uint cast is undefined, so we clamp
// the value before conversion.
srcB = context.FPMaximum(srcB, ConstF(0));
}
srcB = isSignedInt
? context.FPConvertToS32(srcB)
: context.FPConvertToU32(srcB);
if (isSmallInt)
{
int min = (int)GetIntMin(intType);
int max = (int)GetIntMax(intType);
srcB = isSignedInt
? context.IClampS32(srcB, Const(min), Const(max))
: context.IClampU32(srcB, Const(min), Const(max));
}
Operand dest = GetDest(context);
context.Copy(dest, srcB);
// TODO: CC.
}
public static Operand GetDest(EmitterContext context)
{
return(Register(((IOpCodeRd)context.CurrOp).Rd));
}
public static Operand GetDest2(EmitterContext context)
{
Register rd = ((IOpCodeRd)context.CurrOp).Rd;
return(Register(rd.Index | 1, rd.Type));
}
public static void Sts(EmitterContext context)
{
EmitStore(context, MemoryRegion.Shared);
}
public static Operand GetFromTruthTable(EmitterContext context, Operand srcA, Operand srcB, Operand srcC, int imm)
{
Operand expr = null;
// Handle some simple cases, or cases where
// the KMap would yield poor results (like XORs).
if (imm == 0x96 || imm == 0x69)
{
// XOR (0x96) and XNOR (0x69).
if (imm == 0x69)
{
srcA = context.BitwiseNot(srcA);
}
expr = context.BitwiseExclusiveOr(srcA, srcB);
expr = context.BitwiseExclusiveOr(expr, srcC);
return(expr);
}
else if (imm == 0)
{
// Always false.
return(Const(IrConsts.False));
}
else if (imm == 0xff)
{
// Always true.
return(Const(IrConsts.True));
}
int map;
// Encode into gray code.
map = ((imm >> 0) & 1) << 0;
map |= ((imm >> 1) & 1) << 4;
map |= ((imm >> 2) & 1) << 1;
map |= ((imm >> 3) & 1) << 5;
map |= ((imm >> 4) & 1) << 3;
map |= ((imm >> 5) & 1) << 7;
map |= ((imm >> 6) & 1) << 2;
map |= ((imm >> 7) & 1) << 6;
// Solve KMap, get sum of products.
int visited = 0;
for (int index = 0; index < 8 && visited != 0xff; index++)
{
if ((map & (1 << index)) == 0)
{
continue;
}
int mask = 0;
for (int mSize = 4; mSize != 0; mSize >>= 1)
{
mask = RotateLeft4((1 << mSize) - 1, index & 3) << (index & 4);
if ((map & mask) == mask)
{
break;
}
}
// The mask should wrap, if we are on the high row, shift to low etc.
int mask2 = (index & 4) != 0 ? mask >> 4 : mask << 4;
if ((map & mask2) == mask2)
{
mask |= mask2;
}
if ((mask & visited) == mask)
{
continue;
}
bool notA = (mask & 0x33) != 0;
bool notB = (mask & 0x99) != 0;
bool notC = (mask & 0x0f) != 0;
bool aChanges = (mask & 0xcc) != 0 && notA;
bool bChanges = (mask & 0x66) != 0 && notB;
bool cChanges = (mask & 0xf0) != 0 && notC;
Operand localExpr = null;
void And(Operand source)
{
if (localExpr != null)
{
localExpr = context.BitwiseAnd(localExpr, source);
}
else
{
localExpr = source;
}
}
if (!aChanges)
{
And(context.BitwiseNot(srcA, notA));
}
if (!bChanges)
{
And(context.BitwiseNot(srcB, notB));
}
if (!cChanges)
{
And(context.BitwiseNot(srcC, notC));
}
if (expr != null)
{
expr = context.BitwiseOr(expr, localExpr);
}
else
{
expr = localExpr;
}
visited |= mask;
}
return(expr);
}
public static void Depbar(EmitterContext context)
{
}
public static void St(EmitterContext context)
{
EmitStore(context, MemoryRegion.Local);
}
public static Operand GetSrcA(EmitterContext context)
{
return(Register(((IOpCodeRa)context.CurrOp).Ra));
}
public static void Bra(EmitterContext context)
{
EmitBranch(context, context.CurrBlock.Branch.Address);
}
private static void Tex(EmitterContext context, TextureFlags flags)
{
OpCodeTexture op = (OpCodeTexture)context.CurrOp;
bool isBindless = (flags & TextureFlags.Bindless) != 0;
bool intCoords = (flags & TextureFlags.IntCoords) != 0;
if (op.Rd.IsRZ)
{
return;
}
int raIndex = op.Ra.Index;
int rbIndex = op.Rb.Index;
Operand Ra()
{
if (raIndex > RegisterConsts.RegisterZeroIndex)
{
return(Const(0));
}
return(context.Copy(Register(raIndex++, RegisterType.Gpr)));
}
Operand Rb()
{
if (rbIndex > RegisterConsts.RegisterZeroIndex)
{
return(Const(0));
}
return(context.Copy(Register(rbIndex++, RegisterType.Gpr)));
}
Operand arrayIndex = op.IsArray ? Ra() : null;
List <Operand> sourcesList = new List <Operand>();
if (isBindless)
{
sourcesList.Add(Rb());
}
TextureType type = GetTextureType(op.Dimensions);
int coordsCount = type.GetCoordsCount();
for (int index = 0; index < coordsCount; index++)
{
sourcesList.Add(Ra());
}
if (op.IsArray)
{
sourcesList.Add(arrayIndex);
type |= TextureType.Array;
}
bool hasLod = op.LodMode > TextureLodMode.LodZero;
Operand lodValue = hasLod ? Rb() : ConstF(0);
Operand packedOffs = op.HasOffset ? Rb() : null;
if (op.HasDepthCompare)
{
sourcesList.Add(Rb());
type |= TextureType.Shadow;
}
if ((op.LodMode == TextureLodMode.LodZero ||
op.LodMode == TextureLodMode.LodLevel ||
op.LodMode == TextureLodMode.LodLevelA) && !op.IsMultisample)
{
sourcesList.Add(lodValue);
flags |= TextureFlags.LodLevel;
}
if (op.HasOffset)
{
for (int index = 0; index < coordsCount; index++)
{
sourcesList.Add(context.BitfieldExtractS32(packedOffs, Const(index * 4), Const(4)));
}
flags |= TextureFlags.Offset;
}
if (op.LodMode == TextureLodMode.LodBias ||
op.LodMode == TextureLodMode.LodBiasA)
{
sourcesList.Add(lodValue);
flags |= TextureFlags.LodBias;
}
if (op.IsMultisample)
{
sourcesList.Add(Rb());
type |= TextureType.Multisample;
}
Operand[] sources = sourcesList.ToArray();
int rdIndex = op.Rd.Index;
Operand GetDest()
{
if (rdIndex > RegisterConsts.RegisterZeroIndex)
{
return(Const(0));
}
return(Register(rdIndex++, RegisterType.Gpr));
}
int handle = !isBindless ? op.Immediate : 0;
for (int compMask = op.ComponentMask, compIndex = 0; compMask != 0; compMask >>= 1, compIndex++)
{
if ((compMask & 1) != 0)
{
Operand dest = GetDest();
TextureOperation operation = new TextureOperation(
Instruction.TextureSample,
type,
flags,
handle,
compIndex,
dest,
sources);
context.Add(operation);
}
}
}
private static void Txq(EmitterContext context, bool bindless)
{
OpCodeTex op = (OpCodeTex)context.CurrOp;
if (op.Rd.IsRZ)
{
return;
}
TextureProperty property = (TextureProperty)op.RawOpCode.Extract(22, 6);
// TODO: Validate and use property.
Instruction inst = Instruction.TextureSize;
TextureType type = TextureType.Texture2D;
TextureFlags flags = bindless ? TextureFlags.Bindless : TextureFlags.None;
int raIndex = op.Ra.Index;
Operand Ra()
{
if (raIndex > RegisterConsts.RegisterZeroIndex)
{
return(Const(0));
}
return(context.Copy(Register(raIndex++, RegisterType.Gpr)));
}
List <Operand> sourcesList = new List <Operand>();
if (bindless)
{
sourcesList.Add(Ra());
}
sourcesList.Add(Ra());
Operand[] sources = sourcesList.ToArray();
int rdIndex = op.Rd.Index;
Operand GetDest()
{
if (rdIndex > RegisterConsts.RegisterZeroIndex)
{
return(Const(0));
}
return(Register(rdIndex++, RegisterType.Gpr));
}
int handle = !bindless ? op.Immediate : 0;
for (int compMask = op.ComponentMask, compIndex = 0; compMask != 0; compMask >>= 1, compIndex++)
{
if ((compMask & 1) != 0)
{
Operand dest = GetDest();
TextureOperation operation = new TextureOperation(
inst,
type,
flags,
handle,
compIndex,
dest,
sources);
context.Add(operation);
}
}
}
public static void Txq_B(EmitterContext context)
{
Txq(context, bindless: true);
}
public static void Txq(EmitterContext context)
{
Txq(context, bindless: false);
}
public static void Kil(EmitterContext context)
{
context.Discard();
}
public static void Lds(EmitterContext context)
{
EmitLoad(context, MemoryRegion.Shared);
}
public static void Nop(EmitterContext context)
{
}
public static void Stg(EmitterContext context)
{
EmitStoreGlobal(context);
}
public static void Ssy(EmitterContext context)
{
EmitPbkOrSsy(context);
}
private static Operand EmitAtomicOp(
EmitterContext context,
Instruction mr,
AtomicOp op,
ReductionType type,
Operand addrLow,
Operand addrHigh,
Operand value)
{
Operand res = Const(0);
switch (op)
{
case AtomicOp.Add:
if (type == ReductionType.S32 || type == ReductionType.U32)
{
res = context.AtomicAdd(mr, addrLow, addrHigh, value);
}
else
{
context.Config.GpuAccessor.Log($"Invalid reduction type: {type}.");
}
break;
case AtomicOp.BitwiseAnd:
if (type == ReductionType.S32 || type == ReductionType.U32)
{
res = context.AtomicAnd(mr, addrLow, addrHigh, value);
}
else
{
context.Config.GpuAccessor.Log($"Invalid reduction type: {type}.");
}
break;
case AtomicOp.BitwiseExclusiveOr:
if (type == ReductionType.S32 || type == ReductionType.U32)
{
res = context.AtomicXor(mr, addrLow, addrHigh, value);
}
else
{
context.Config.GpuAccessor.Log($"Invalid reduction type: {type}.");
}
break;
case AtomicOp.BitwiseOr:
if (type == ReductionType.S32 || type == ReductionType.U32)
{
res = context.AtomicOr(mr, addrLow, addrHigh, value);
}
else
{
context.Config.GpuAccessor.Log($"Invalid reduction type: {type}.");
}
break;
case AtomicOp.Maximum:
if (type == ReductionType.S32)
{
res = context.AtomicMaxS32(mr, addrLow, addrHigh, value);
}
else if (type == ReductionType.U32)
{
res = context.AtomicMaxU32(mr, addrLow, addrHigh, value);
}
else
{
context.Config.GpuAccessor.Log($"Invalid reduction type: {type}.");
}
break;
case AtomicOp.Minimum:
if (type == ReductionType.S32)
{
res = context.AtomicMinS32(mr, addrLow, addrHigh, value);
}
else if (type == ReductionType.U32)
{
res = context.AtomicMinU32(mr, addrLow, addrHigh, value);
}
else
{
context.Config.GpuAccessor.Log($"Invalid reduction type: {type}.");
}
break;
}
return(res);
}
public static void Sync(EmitterContext context)
{
EmitBrkOrSync(context);
}
public static Operand SignExtendTo32(EmitterContext context, Operand src, int srcBits)
{
return(context.BitfieldExtractS32(src, Const(0), Const(srcBits)));
}
public static void Al2p(EmitterContext context)
{
InstAl2p op = context.GetOp <InstAl2p>();
context.Copy(GetDest(op.Dest), context.IAdd(GetSrcReg(context, op.SrcA), Const(op.Imm11)));
}
public static Operand ZeroExtendTo32(EmitterContext context, Operand src, int srcBits)
{
int mask = (int)(0xffffffffu >> (32 - srcBits));
return(context.BitwiseAnd(src, Const(mask)));
}
public static void OutC(EmitterContext context)
{
InstOutC op = context.GetOp <InstOutC>();
EmitOut(context, op.OutType.HasFlag(OutType.Emit), op.OutType.HasFlag(OutType.Cut));
}
public static void Mov(EmitterContext context)
{
context.Copy(GetDest(context), GetSrcB(context));
}
public static void Ld(EmitterContext context)
{
EmitLoad(context, MemoryRegion.Local);
}
private static void EmitF2I(
EmitterContext context,
DstFmt srcType,
IDstFmt dstType,
RoundMode2 roundingMode,
Operand src,
int rd,
bool absolute,
bool negate)
{
if (dstType == IDstFmt.U64)
{
context.Config.GpuAccessor.Log("Unimplemented 64-bits F2I.");
}
Instruction fpType = srcType.ToInstFPType();
bool isSignedInt = dstType == IDstFmt.S16 || dstType == IDstFmt.S32 || dstType == IDstFmt.S64;
bool isSmallInt = dstType == IDstFmt.U16 || dstType == IDstFmt.S16;
Operand srcB = context.FPAbsNeg(src, absolute, negate, fpType);
srcB = roundingMode switch
{
RoundMode2.Round => context.FPRound(srcB, fpType),
RoundMode2.Floor => context.FPFloor(srcB, fpType),
RoundMode2.Ceil => context.FPCeiling(srcB, fpType),
RoundMode2.Trunc => context.FPTruncate(srcB, fpType),
_ => srcB
};
if (!isSignedInt)
{
// Negative float to uint cast is undefined, so we clamp the value before conversion.
Operand c0 = srcType == DstFmt.F64 ? context.PackDouble2x32(0.0) : ConstF(0);
srcB = context.FPMaximum(srcB, c0, fpType);
}
if (srcType == DstFmt.F64)
{
srcB = isSignedInt
? context.FP64ConvertToS32(srcB)
: context.FP64ConvertToU32(srcB);
}
else
{
srcB = isSignedInt
? context.FP32ConvertToS32(srcB)
: context.FP32ConvertToU32(srcB);
}
if (isSmallInt)
{
int min = (int)GetIntMin(dstType);
int max = (int)GetIntMax(dstType);
srcB = isSignedInt
? context.IClampS32(srcB, Const(min), Const(max))
: context.IClampU32(srcB, Const(min), Const(max));
}
Operand dest = GetDest(rd);
context.Copy(dest, srcB);
// TODO: CC.
}
public static void Ldg(EmitterContext context)
{
EmitLoadGlobal(context);
}
public abstract void EmitObject(LogicalObject obj, EmitterContext context);
public static void Texs(EmitterContext context)
{
OpCodeTextureScalar op = (OpCodeTextureScalar)context.CurrOp;
if (op.Rd0.IsRZ && op.Rd1.IsRZ)
{
return;
}
List <Operand> sourcesList = new List <Operand>();
int raIndex = op.Ra.Index;
int rbIndex = op.Rb.Index;
Operand Ra()
{
if (raIndex > RegisterConsts.RegisterZeroIndex)
{
return(Const(0));
}
return(context.Copy(Register(raIndex++, RegisterType.Gpr)));
}
Operand Rb()
{
if (rbIndex > RegisterConsts.RegisterZeroIndex)
{
return(Const(0));
}
return(context.Copy(Register(rbIndex++, RegisterType.Gpr)));
}
void AddTextureOffset(int coordsCount, int stride, int size)
{
Operand packedOffs = Rb();
for (int index = 0; index < coordsCount; index++)
{
sourcesList.Add(context.BitfieldExtractS32(packedOffs, Const(index * stride), Const(size)));
}
}
TextureType type;
TextureFlags flags;
if (op is OpCodeTexs texsOp)
{
type = GetTextureType(texsOp.Type);
flags = GetTextureFlags(texsOp.Type);
if ((type & TextureType.Array) != 0)
{
Operand arrayIndex = Ra();
sourcesList.Add(Ra());
sourcesList.Add(Rb());
sourcesList.Add(arrayIndex);
if ((type & TextureType.Shadow) != 0)
{
sourcesList.Add(Rb());
}
if ((flags & TextureFlags.LodLevel) != 0)
{
sourcesList.Add(ConstF(0));
}
}
else
{
switch (texsOp.Type)
{
case TextureScalarType.Texture1DLodZero:
sourcesList.Add(Ra());
break;
case TextureScalarType.Texture2D:
sourcesList.Add(Ra());
sourcesList.Add(Rb());
break;
case TextureScalarType.Texture2DLodZero:
sourcesList.Add(Ra());
sourcesList.Add(Rb());
sourcesList.Add(ConstF(0));
break;
case TextureScalarType.Texture2DLodLevel:
case TextureScalarType.Texture2DDepthCompare:
case TextureScalarType.Texture3D:
case TextureScalarType.TextureCube:
sourcesList.Add(Ra());
sourcesList.Add(Ra());
sourcesList.Add(Rb());
break;
case TextureScalarType.Texture2DLodZeroDepthCompare:
case TextureScalarType.Texture3DLodZero:
sourcesList.Add(Ra());
sourcesList.Add(Ra());
sourcesList.Add(Rb());
sourcesList.Add(ConstF(0));
break;
case TextureScalarType.Texture2DLodLevelDepthCompare:
case TextureScalarType.TextureCubeLodLevel:
sourcesList.Add(Ra());
sourcesList.Add(Ra());
sourcesList.Add(Rb());
sourcesList.Add(Rb());
break;
}
}
}
else if (op is OpCodeTlds tldsOp)
{
type = GetTextureType(tldsOp.Type);
flags = GetTextureFlags(tldsOp.Type) | TextureFlags.IntCoords;
switch (tldsOp.Type)
{
case TexelLoadScalarType.Texture1DLodZero:
sourcesList.Add(Ra());
sourcesList.Add(Const(0));
break;
case TexelLoadScalarType.Texture1DLodLevel:
sourcesList.Add(Ra());
sourcesList.Add(Rb());
break;
case TexelLoadScalarType.Texture2DLodZero:
sourcesList.Add(Ra());
sourcesList.Add(Rb());
sourcesList.Add(Const(0));
break;
case TexelLoadScalarType.Texture2DLodZeroOffset:
sourcesList.Add(Ra());
sourcesList.Add(Ra());
sourcesList.Add(Const(0));
break;
case TexelLoadScalarType.Texture2DLodZeroMultisample:
case TexelLoadScalarType.Texture2DLodLevel:
case TexelLoadScalarType.Texture2DLodLevelOffset:
sourcesList.Add(Ra());
sourcesList.Add(Ra());
sourcesList.Add(Rb());
break;
case TexelLoadScalarType.Texture3DLodZero:
sourcesList.Add(Ra());
sourcesList.Add(Ra());
sourcesList.Add(Rb());
sourcesList.Add(Const(0));
break;
case TexelLoadScalarType.Texture2DArrayLodZero:
sourcesList.Add(Rb());
sourcesList.Add(Rb());
sourcesList.Add(Ra());
sourcesList.Add(Const(0));
break;
}
if ((flags & TextureFlags.Offset) != 0)
{
AddTextureOffset(type.GetCoordsCount(), 4, 4);
}
}
else if (op is OpCodeTld4s tld4sOp)
{
if (!(tld4sOp.HasDepthCompare || tld4sOp.HasOffset))
{
sourcesList.Add(Ra());
sourcesList.Add(Rb());
}
else
{
sourcesList.Add(Ra());
sourcesList.Add(Ra());
}
type = TextureType.Texture2D;
flags = TextureFlags.Gather;
if (tld4sOp.HasDepthCompare)
{
sourcesList.Add(Rb());
type |= TextureType.Shadow;
}
if (tld4sOp.HasOffset)
{
AddTextureOffset(type.GetCoordsCount(), 8, 6);
flags |= TextureFlags.Offset;
}
sourcesList.Add(Const(tld4sOp.GatherCompIndex));
}
else
{
throw new InvalidOperationException($"Invalid opcode type \"{op.GetType().Name}\".");
}
Operand[] sources = sourcesList.ToArray();
Operand[] rd0 = new Operand[2] {
ConstF(0), ConstF(0)
};
Operand[] rd1 = new Operand[2] {
ConstF(0), ConstF(0)
};
int destIncrement = 0;
Operand GetDest()
{
int high = destIncrement >> 1;
int low = destIncrement & 1;
destIncrement++;
if (op.IsFp16)
{
return(high != 0
? (rd1[low] = Local())
: (rd0[low] = Local()));
}
else
{
int rdIndex = high != 0 ? op.Rd1.Index : op.Rd0.Index;
if (rdIndex < RegisterConsts.RegisterZeroIndex)
{
rdIndex += low;
}
return(Register(rdIndex, RegisterType.Gpr));
}
}
int handle = op.Immediate;
for (int compMask = op.ComponentMask, compIndex = 0; compMask != 0; compMask >>= 1, compIndex++)
{
if ((compMask & 1) != 0)
{
Operand dest = GetDest();
TextureOperation operation = new TextureOperation(
Instruction.TextureSample,
type,
flags,
handle,
compIndex,
dest,
sources);
context.Add(operation);
}
}
if (op.IsFp16)
{
context.Copy(Register(op.Rd0), context.PackHalf2x16(rd0[0], rd0[1]));
context.Copy(Register(op.Rd1), context.PackHalf2x16(rd1[0], rd1[1]));
}
}
