mirrors/ryujinx
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

A32: Implement VCVTT, VCVTB (#3710)

Browse source 
* A32: Implement VCVTT, VCVTB

* A32: F16C implementation of VCVTT/VCVTB
This commit is contained in:
merry 2022-10-19 01:36:04 +01:00 • committed by GitHub
parent 5af8ce7c38
commit 8d41402fa6
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
5 changed files with 210 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

62
ARMeilleure/Instructions/InstEmitSimdCvt32.cs
View file

@ -261,6 +261,68 @@ namespace ARMeilleure.Instructions
}
}
public static void Vcvt_TB(ArmEmitterContext context)
{
OpCode32SimdCvtTB op = (OpCode32SimdCvtTB)context.CurrOp;
if (Optimizations.UseF16c)
{
Debug.Assert(!Optimizations.ForceLegacySse);
if (op.Op)
{
Operand res = ExtractScalar(context, op.Size == 1 ? OperandType.FP64 : OperandType.FP32, op.Vm);
if (op.Size == 1)
{
res = context.AddIntrinsic(Intrinsic.X86Cvtsd2ss, context.VectorZero(), res);
}
res = context.AddIntrinsic(Intrinsic.X86Vcvtps2ph, res, Const(X86GetRoundControl(FPRoundingMode.ToNearest)));
res = context.VectorExtract16(res, 0);
InsertScalar16(context, op.Vd, op.T, res);
}
else
{
Operand res = context.VectorCreateScalar(ExtractScalar16(context, op.Vm, op.T));
res = context.AddIntrinsic(Intrinsic.X86Vcvtph2ps, res);
if (op.Size == 1)
{
res = context.AddIntrinsic(Intrinsic.X86Cvtss2sd, context.VectorZero(), res);
}
res = context.VectorExtract(op.Size == 1 ? OperandType.I64 : OperandType.I32, res, 0);
InsertScalar(context, op.Vd, res);
}
}
else
{
if (op.Op)
{
// Convert to half
Operand src = ExtractScalar(context, op.Size == 1 ? OperandType.FP64 : OperandType.FP32, op.Vm);
MethodInfo method = op.Size == 1
? typeof(SoftFloat64_16).GetMethod(nameof(SoftFloat64_16.FPConvert))
: typeof(SoftFloat32_16).GetMethod(nameof(SoftFloat32_16.FPConvert));
Operand res = context.Call(method, src);
InsertScalar16(context, op.Vd, op.T, res);
}
else
{
// Convert from half
Operand src = ExtractScalar16(context, op.Vm, op.T);
MethodInfo method = op.Size == 1
? typeof(SoftFloat16_64).GetMethod(nameof(SoftFloat16_64.FPConvert))
: typeof(SoftFloat16_32).GetMethod(nameof(SoftFloat16_32.FPConvert));
Operand res = context.Call(method, src);
InsertScalar(context, op.Vd, res);
}
}
}
// VRINTA/M/N/P (floating-point).
public static void Vrint_RM(ArmEmitterContext context)
{

16
ARMeilleure/Instructions/InstEmitSimdHelper32.cs
View file

@ -70,6 +70,22 @@ namespace ARMeilleure.Instructions
context.Copy(vec, insert);
}
public static Operand ExtractScalar16(ArmEmitterContext context, int reg, bool top)
{
return context.VectorExtract16(GetVecA32(reg >> 2), ((reg & 3) << 1) | (top ? 1 : 0));
}
public static void InsertScalar16(ArmEmitterContext context, int reg, bool top, Operand value)
{
Debug.Assert(value.Type == OperandType.FP32 || value.Type == OperandType.I32);
Operand vec, insert;
vec = GetVecA32(reg >> 2);
insert = context.VectorInsert16(vec, value, ((reg & 3) << 1) | (top ? 1 : 0));
context.Copy(vec, insert);
}
public static Operand ExtractElement(ArmEmitterContext context, int reg, int size, bool signed)
{
return EmitVectorExtract32(context, reg >> (4 - size), reg & ((16 >> size) - 1), size, signed);