ShuriZma
/
rpcs3
mirror of https://github.com/RPCS3/rpcs3.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

PPU LLVM: Improve VMADDFP and VNMSUBFP 

- Use native FMA to emulate VMADDFP, with a fallback for processors that don't support FMA
- Use native FMA to emulate VNMSUBFP as well, but note that it differs from the emulated path with regards to negative zero
This commit is contained in:
Malcolm Jestadt 2020-03-18 02:43:46 -04:00 committed by Ivan
parent b11e8f8b8d
commit 0bfdc1f62e
1 changed files with 37 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
rpcs3/Emu/Cell/PPUTranslator.cpp
View File

@ -902,8 +902,24 @@ void PPUTranslator::VLOGEFP(ppu_opcode_t op)
void PPUTranslator::VMADDFP(ppu_opcode_t op)
{
const auto acb = GetVrs(VrType::vf, op.va, op.vc, op.vb);
SetVr(op.vd, m_ir->CreateFAdd(m_ir->CreateFMul(acb[0], acb[1]), acb[2]));
if (m_use_fma)
{
const auto acb = GetVrs(VrType::vf, op.va, op.vc, op.vb);
SetVr(op.vd, m_ir->CreateCall(get_intrinsic<f32[4]>(llvm::Intrinsic::fma), { acb[0], acb[1], acb[2] }));
return;
}
// Emulated FMA via double precision
auto a = get_vr<f32[4]>(op.va);
auto b = get_vr<f32[4]>(op.vb);
auto c = get_vr<f32[4]>(op.vc);
const auto xa = m_ir->CreateFPExt(a.value, get_type<f64[4]>());
const auto xb = m_ir->CreateFPExt(b.value, get_type<f64[4]>());
const auto xc = m_ir->CreateFPExt(c.value, get_type<f64[4]>());
const auto xr = m_ir->CreateCall(get_intrinsic<f64[4]>(llvm::Intrinsic::fmuladd), {xa, xc, xb});
SetVr(op.vd, m_ir->CreateFPTrunc(xr, get_type<f32[4]>()));
}
void PPUTranslator::VMAXFP(ppu_opcode_t op)
@ -1182,8 +1198,25 @@ void PPUTranslator::VMULOUH(ppu_opcode_t op)
void PPUTranslator::VNMSUBFP(ppu_opcode_t op)
{
const auto acb = GetVrs(VrType::vf, op.va, op.vc, op.vb);
SetVr(op.vd, m_ir->CreateFNeg(m_ir->CreateFSub(m_ir->CreateFMul(acb[0], acb[1]), acb[2])));
// Differs from the emulated path with regards to negative zero
if (m_use_fma)
{
const auto acb = GetVrs(VrType::vf, op.va, op.vc, op.vb);
SetVr(op.vd, m_ir->CreateFNeg(m_ir->CreateCall(get_intrinsic<f32[4]>(llvm::Intrinsic::fma), { acb[0], acb[1], m_ir->CreateFNeg(acb[2]) })));
return;
}
// Emulated FMA via double precision
auto a = get_vr<f32[4]>(op.va);
auto b = get_vr<f32[4]>(op.vb);
auto c = get_vr<f32[4]>(op.vc);
const auto xa = m_ir->CreateFPExt(a.value, get_type<f64[4]>());
const auto xb = m_ir->CreateFPExt(b.value, get_type<f64[4]>());
const auto xc = m_ir->CreateFPExt(c.value, get_type<f64[4]>());
const auto xr = m_ir->CreateFNeg(m_ir->CreateFSub(m_ir->CreateFMul(xa, xc), xb));
SetVr(op.vd, m_ir->CreateFPTrunc(xr, get_type<f32[4]>()));
}
void PPUTranslator::VNOR(ppu_opcode_t op)