hwpplayer1/suyu
1
0
Fork 0
forked from suyu/suyu
Code Pull requests Activity

shader: Refactor atomic_operations_global_memory

Browse source
This commit is contained in:
ameerj 2021-04-21 21:34:51 -04:00
parent 427951d6fe
commit 7a9dc78398
1 changed files with 36 additions and 44 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

80
src/shader_recompiler/frontend/maxwell/translate/impl/atomic_operations_global_memory.cpp
View file

@ -154,69 +154,61 @@ void StoreResult(TranslatorVisitor& v, IR::Reg dest_reg, const IR::Value& result
break; break;
} }
} }
IR::Value ApplyAtomOp(TranslatorVisitor& v, IR::Reg operand_reg, const IR::U64& offset,
AtomSize size, AtomOp op) {
switch (size) {
case AtomSize::U32:
case AtomSize::S32:
return ApplyIntegerAtomOp(v.ir, offset, v.X(operand_reg), op, size == AtomSize::S32);
case AtomSize::U64:
case AtomSize::S64:
return ApplyIntegerAtomOp(v.ir, offset, v.L(operand_reg), op, size == AtomSize::S64);
case AtomSize::F32:
return ApplyFpAtomOp(v.ir, offset, v.F(operand_reg), op, size);
case AtomSize::F16x2: {
return ApplyFpAtomOp(v.ir, offset, v.ir.UnpackFloat2x16(v.X(operand_reg)), op, size);
}
default:
throw NotImplementedException("Atom Size {}", size);
}
}
void GlobalAtomic(TranslatorVisitor& v, IR::Reg dest_reg, IR::Reg operand_reg,
const IR::U64& offset, AtomSize size, AtomOp op, bool write_dest) {
IR::Value result;
if (AtomOpNotApplicable(size, op)) {
result = LoadGlobal(v.ir, offset, size);
} else {
result = ApplyAtomOp(v, operand_reg, offset, size, op);
}
if (write_dest) {
StoreResult(v, dest_reg, result, size);
}
}
} // Anonymous namespace } // Anonymous namespace
void TranslatorVisitor::ATOM(u64 insn) { void TranslatorVisitor::ATOM(u64 insn) {
union { union {
u64 raw; u64 raw;
BitField<0, 8, IR::Reg> dest_reg; BitField<0, 8, IR::Reg> dest_reg;
BitField<8, 8, IR::Reg> addr_reg; BitField<20, 8, IR::Reg> operand_reg;
BitField<20, 8, IR::Reg> src_reg_b;
BitField<49, 3, AtomSize> size; BitField<49, 3, AtomSize> size;
BitField<52, 4, AtomOp> op; BitField<52, 4, AtomOp> op;
} const atom{insn}; } const atom{insn};
const bool size_64{atom.size == AtomSize::U64 || atom.size == AtomSize::S64};
const bool is_signed{atom.size == AtomSize::S32 || atom.size == AtomSize::S64};
const bool is_integer{atom.size != AtomSize::F32 && atom.size != AtomSize::F16x2};
const IR::U64 offset{AtomOffset(*this, insn)}; const IR::U64 offset{AtomOffset(*this, insn)};
IR::Value result; GlobalAtomic(*this, atom.dest_reg, atom.operand_reg, offset, atom.size, atom.op, true);
if (AtomOpNotApplicable(atom.size, atom.op)) {
result = LoadGlobal(ir, offset, atom.size);
} else if (!is_integer) {
if (atom.size == AtomSize::F32) {
result = ApplyFpAtomOp(ir, offset, F(atom.src_reg_b), atom.op, atom.size);
} else {
const IR::Value src_b{ir.UnpackFloat2x16(X(atom.src_reg_b))};
result = ApplyFpAtomOp(ir, offset, src_b, atom.op, atom.size);
}
} else if (size_64) {
result = ApplyIntegerAtomOp(ir, offset, L(atom.src_reg_b), atom.op, is_signed);
} else {
result = ApplyIntegerAtomOp(ir, offset, X(atom.src_reg_b), atom.op, is_signed);
}
StoreResult(*this, atom.dest_reg, result, atom.size);
} }
void TranslatorVisitor::RED(u64 insn) { void TranslatorVisitor::RED(u64 insn) {
union { union {
u64 raw; u64 raw;
BitField<0, 8, IR::Reg> src_reg_b; BitField<0, 8, IR::Reg> operand_reg;
BitField<8, 8, IR::Reg> addr_reg;
BitField<20, 3, AtomSize> size; BitField<20, 3, AtomSize> size;
BitField<23, 3, AtomOp> op; BitField<23, 3, AtomOp> op;
} const red{insn}; } const red{insn};
if (AtomOpNotApplicable(red.size, red.op)) {
return;
}
const bool size_64{red.size == AtomSize::U64 || red.size == AtomSize::S64};
const bool is_signed{red.size == AtomSize::S32 || red.size == AtomSize::S64};
const bool is_integer{red.size != AtomSize::F32 && red.size != AtomSize::F16x2};
const IR::U64 offset{AtomOffset(*this, insn)}; const IR::U64 offset{AtomOffset(*this, insn)};
if (!is_integer) { GlobalAtomic(*this, IR::Reg::RZ, red.operand_reg, offset, red.size, red.op, true);
if (red.size == AtomSize::F32) {
ApplyFpAtomOp(ir, offset, F(red.src_reg_b), red.op, red.size);
} else {
const IR::Value src_b{ir.UnpackFloat2x16(X(red.src_reg_b))};
ApplyFpAtomOp(ir, offset, src_b, red.op, red.size);
}
} else if (size_64) {
ApplyIntegerAtomOp(ir, offset, L(red.src_reg_b), red.op, is_signed);
} else {
ApplyIntegerAtomOp(ir, offset, X(red.src_reg_b), red.op, is_signed);
}
} }
} // namespace Shader::Maxwell } // namespace Shader::Maxwell