Skip to content

Commit

Permalink
Merge pull request #9970 from JosJuice/jit64-fmaddxx-accurate-nan
Browse files Browse the repository at this point in the history 
Jit64: Fix fmaddXX with accurate NaNs
  • Loading branch information
@lioncash
lioncash committed Jul 29, 2021
2 parents 11d1269 + 3bb4a4e commit 28ee0af
Show file tree
Hide file tree
Showing 3 changed files with 49 additions and 37 deletions.
2 changes: 1 addition & 1 deletion Source/Core/Core/PowerPC/Jit64/Jit.h
View file
Expand Up @@ -126,7 +126,7 @@ class Jit64 : public JitBase, public QuantizedMemoryRoutines
bool duplicate = false);
void FinalizeDoubleResult(Gen::X64Reg output, const Gen::OpArg& input);
void HandleNaNs(UGeckoInstruction inst, Gen::X64Reg xmm_out, Gen::X64Reg xmm_in,
Gen::X64Reg clobber = Gen::XMM0);
Gen::X64Reg clobber);

void MultiplyImmediate(u32 imm, int a, int d, bool overflow);

Expand Down
82 changes: 47 additions & 35 deletions Source/Core/Core/PowerPC/Jit64/Jit_FloatingPoint.cpp
View file
Expand Up @@ -257,7 +257,7 @@ void Jit64::fp_arith(UGeckoInstruction inst)
avx_op(avxOp, sseOp, dest, Rop1, Rop2, packed, reversible);
}

HandleNaNs(inst, Rd, dest);
HandleNaNs(inst, Rd, dest, XMM0);
if (single)
FinalizeSingleResult(Rd, Rd, packed, true);
else
Expand Down Expand Up @@ -345,7 +345,8 @@ void Jit64::fmaddXX(UGeckoInstruction inst)
RegCache::Realize(Ra, Rb, Rc, Rd);
}

X64Reg result_reg = XMM0;
X64Reg scratch_xmm = !use_fma && inst.SUBOP5 == 30 ? XMM1 : XMM0;
X64Reg result_xmm = scratch_xmm == XMM0 ? XMM1 : XMM0;
if (software_fma)
{
for (size_t i = (packed ? 1 : 0); i != std::numeric_limits<size_t>::max(); --i)
Expand Down Expand Up @@ -392,31 +393,35 @@ void Jit64::fmaddXX(UGeckoInstruction inst)
if (packed)
{
MOVSD(Rd, XMM0);
result_reg = Rd;
result_xmm = Rd;
}
else
{
result_xmm = XMM0;
}

if (inst.SUBOP5 == 30 || inst.SUBOP5 == 31) // nmsub, nmadd
XORPD(result_reg, MConst(packed ? psSignBits2 : psSignBits));
XORPD(result_xmm, MConst(packed ? psSignBits2 : psSignBits));
}
else
{
switch (inst.SUBOP5)
{
case 14: // madds0
MOVDDUP(XMM0, Rc);
MOVDDUP(result_xmm, Rc);
if (round_input)
Force25BitPrecision(XMM0, R(XMM0), XMM1);
Force25BitPrecision(result_xmm, R(result_xmm), scratch_xmm);
break;
case 15: // madds1
avx_op(&XEmitter::VSHUFPD, &XEmitter::SHUFPD, XMM0, Rc, Rc, 3);
avx_op(&XEmitter::VSHUFPD, &XEmitter::SHUFPD, result_xmm, Rc, Rc, 3);
if (round_input)
Force25BitPrecision(XMM0, R(XMM0), XMM1);
Force25BitPrecision(result_xmm, R(result_xmm), scratch_xmm);
break;
default:
if (single && round_input)
Force25BitPrecision(XMM0, Rc, XMM1);
Force25BitPrecision(result_xmm, Rc, scratch_xmm);
else
MOVAPD(XMM0, Rc);
MOVAPD(result_xmm, Rc);
break;
}

Expand All @@ -426,33 +431,33 @@ void Jit64::fmaddXX(UGeckoInstruction inst)
{
case 28: // msub
if (packed)
VFMSUB132PD(XMM0, Rb.GetSimpleReg(), Ra);
VFMSUB132PD(result_xmm, Rb.GetSimpleReg(), Ra);
else
VFMSUB132SD(XMM0, Rb.GetSimpleReg(), Ra);
VFMSUB132SD(result_xmm, Rb.GetSimpleReg(), Ra);
break;
case 14: // madds0
case 15: // madds1
case 29: // madd
if (packed)
VFMADD132PD(XMM0, Rb.GetSimpleReg(), Ra);
VFMADD132PD(result_xmm, Rb.GetSimpleReg(), Ra);
else
VFMADD132SD(XMM0, Rb.GetSimpleReg(), Ra);
VFMADD132SD(result_xmm, Rb.GetSimpleReg(), Ra);
break;
// PowerPC and x86 define NMADD/NMSUB differently
// x86: D = -A*C (+/-) B
// PPC: D = -(A*C (+/-) B)
// so we have to swap them; the ADD/SUB here isn't a typo.
case 30: // nmsub
if (packed)
VFNMADD132PD(XMM0, Rb.GetSimpleReg(), Ra);
VFNMADD132PD(result_xmm, Rb.GetSimpleReg(), Ra);
else
VFNMADD132SD(XMM0, Rb.GetSimpleReg(), Ra);
VFNMADD132SD(result_xmm, Rb.GetSimpleReg(), Ra);
break;
case 31: // nmadd
if (packed)
VFNMSUB132PD(XMM0, Rb.GetSimpleReg(), Ra);
VFNMSUB132PD(result_xmm, Rb.GetSimpleReg(), Ra);
else
VFNMSUB132SD(XMM0, Rb.GetSimpleReg(), Ra);
VFNMSUB132SD(result_xmm, Rb.GetSimpleReg(), Ra);
break;
}
}
Expand All @@ -462,52 +467,59 @@ void Jit64::fmaddXX(UGeckoInstruction inst)
{
// We implement nmsub a little differently ((b - a*c) instead of -(a*c - b)),
// so handle it separately.
MOVAPD(XMM1, Rb);
MOVAPD(scratch_xmm, Rb);
if (packed)
{
MULPD(XMM0, Ra);
SUBPD(XMM1, R(XMM0));
MULPD(result_xmm, Ra);
SUBPD(scratch_xmm, R(result_xmm));
}
else
{
MULSD(XMM0, Ra);
SUBSD(XMM1, R(XMM0));
MULSD(result_xmm, Ra);
SUBSD(scratch_xmm, R(result_xmm));
}
result_reg = XMM1;
result_xmm = scratch_xmm;
}
else
{
if (packed)
{
MULPD(XMM0, Ra);
MULPD(result_xmm, Ra);
if (inst.SUBOP5 == 28) // msub
SUBPD(XMM0, Rb);
SUBPD(result_xmm, Rb);
else //(n)madd(s[01])
ADDPD(XMM0, Rb);
ADDPD(result_xmm, Rb);
}
else
{
MULSD(XMM0, Ra);
MULSD(result_xmm, Ra);
if (inst.SUBOP5 == 28)
SUBSD(XMM0, Rb);
SUBSD(result_xmm, Rb);
else
ADDSD(XMM0, Rb);
ADDSD(result_xmm, Rb);
}
if (inst.SUBOP5 == 31) // nmadd
XORPD(XMM0, MConst(packed ? psSignBits2 : psSignBits));
XORPD(result_xmm, MConst(packed ? psSignBits2 : psSignBits));
}
}
}

if (SConfig::GetInstance().bAccurateNaNs && result_xmm == XMM0)
{
// HandleNaNs needs to clobber XMM0
MOVAPD(XMM1, R(result_xmm));
result_xmm = XMM1;
}

if (single)
{
HandleNaNs(inst, result_reg, result_reg, result_reg == XMM1 ? XMM0 : XMM1);
FinalizeSingleResult(Rd, R(result_reg), packed, true);
HandleNaNs(inst, result_xmm, result_xmm, XMM0);
FinalizeSingleResult(Rd, R(result_xmm), packed, true);
}
else
{
HandleNaNs(inst, result_reg, result_reg, XMM1);
FinalizeDoubleResult(Rd, R(result_reg));
HandleNaNs(inst, result_xmm, result_xmm, XMM0);
FinalizeDoubleResult(Rd, R(result_xmm));
}
}

Expand Down
2 changes: 1 addition & 1 deletion Source/Core/Core/PowerPC/Jit64/Jit_Paired.cpp
View file
Expand Up @@ -109,7 +109,7 @@ void Jit64::ps_muls(UGeckoInstruction inst)
if (round_input)
Force25BitPrecision(XMM1, R(XMM1), XMM0);
MULPD(XMM1, Ra);
HandleNaNs(inst, Rd, XMM1);
HandleNaNs(inst, Rd, XMM1, XMM0);
FinalizeSingleResult(Rd, Rd);
}

Expand Down

0 comments on commit 28ee0af

Please sign in to comment.