Implement CPU FCVT Half <-> Double conversion variants (#3439)

* Half <-> Double conversion support

* Add tests, fast path and deduplicate SoftFloat code

* PPTC version
This commit is contained in:
gdkchan 2022-07-06 08:40:31 -03:00 committed by GitHub
parent b46b63e06a
commit f7ef6364b7
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
5 changed files with 729 additions and 406 deletions

View file

@ -105,11 +105,48 @@ namespace ARMeilleure.Instructions
}
else if (op.Size == 1 && op.Opc == 3) // Double -> Half.
{
throw new NotImplementedException("Double-precision to half-precision.");
if (Optimizations.UseF16c)
{
Debug.Assert(!Optimizations.ForceLegacySse);
Operand n = GetVec(op.Rn);
Operand res = context.AddIntrinsic(Intrinsic.X86Cvtsd2ss, context.VectorZero(), n);
res = context.AddIntrinsic(Intrinsic.X86Vcvtps2ph, res, Const(X86GetRoundControl(FPRoundingMode.ToNearest)));
context.Copy(GetVec(op.Rd), res);
}
else
{
Operand ne = context.VectorExtract(OperandType.FP64, GetVec(op.Rn), 0);
Operand res = context.Call(typeof(SoftFloat64_16).GetMethod(nameof(SoftFloat64_16.FPConvert)), ne);
res = context.ZeroExtend16(OperandType.I64, res);
context.Copy(GetVec(op.Rd), EmitVectorInsert(context, context.VectorZero(), res, 0, 1));
}
}
else if (op.Size == 3 && op.Opc == 1) // Double -> Half.
else if (op.Size == 3 && op.Opc == 1) // Half -> Double.
{
throw new NotImplementedException("Half-precision to double-precision.");
if (Optimizations.UseF16c)
{
Operand n = GetVec(op.Rn);
Operand res = context.AddIntrinsic(Intrinsic.X86Vcvtph2ps, GetVec(op.Rn));
res = context.AddIntrinsic(Intrinsic.X86Cvtss2sd, context.VectorZero(), res);
res = context.VectorZeroUpper64(res);
context.Copy(GetVec(op.Rd), res);
}
else
{
Operand ne = EmitVectorExtractZx(context, op.Rn, 0, 1);
Operand res = context.Call(typeof(SoftFloat16_64).GetMethod(nameof(SoftFloat16_64.FPConvert)), ne);
context.Copy(GetVec(op.Rd), context.VectorInsert(context.VectorZero(), res, 0));
}
}
else // Invalid encoding.
{