frontend/ir_emitter: Add half-precision opcode variant of FPRSqrtStepFused
This commit is contained in:
parent
e3b2eb57b5
commit
824c551ba2
5 changed files with 57 additions and 40 deletions
|
@ -946,52 +946,54 @@ template<size_t fsize>
|
||||||
static void EmitFPRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
|
static void EmitFPRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
|
||||||
using FPT = mp::unsigned_integer_of_size<fsize>;
|
using FPT = mp::unsigned_integer_of_size<fsize>;
|
||||||
|
|
||||||
if (code.DoesCpuSupport(Xbyak::util::Cpu::tFMA) && code.DoesCpuSupport(Xbyak::util::Cpu::tAVX)) {
|
if constexpr (fsize != 16) {
|
||||||
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
if (code.DoesCpuSupport(Xbyak::util::Cpu::tFMA) && code.DoesCpuSupport(Xbyak::util::Cpu::tAVX)) {
|
||||||
|
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||||
|
|
||||||
Xbyak::Label end, fallback;
|
Xbyak::Label end, fallback;
|
||||||
|
|
||||||
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]);
|
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]);
|
||||||
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
|
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
|
||||||
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
|
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
|
||||||
|
|
||||||
code.vmovaps(result, code.MConst(xword, FP::FPValue<FPT, false, 0, 3>()));
|
code.vmovaps(result, code.MConst(xword, FP::FPValue<FPT, false, 0, 3>()));
|
||||||
FCODE(vfnmadd231s)(result, operand1, operand2);
|
FCODE(vfnmadd231s)(result, operand1, operand2);
|
||||||
|
|
||||||
// Detect if the intermediate result is infinity or NaN or nearly an infinity.
|
// Detect if the intermediate result is infinity or NaN or nearly an infinity.
|
||||||
// Why do we need to care about infinities? This is because x86 doesn't allow us
|
// Why do we need to care about infinities? This is because x86 doesn't allow us
|
||||||
// to fuse the divide-by-two with the rest of the FMA operation. Therefore the
|
// to fuse the divide-by-two with the rest of the FMA operation. Therefore the
|
||||||
// intermediate value may overflow and we would like to handle this case.
|
// intermediate value may overflow and we would like to handle this case.
|
||||||
const Xbyak::Reg32 tmp = ctx.reg_alloc.ScratchGpr().cvt32();
|
const Xbyak::Reg32 tmp = ctx.reg_alloc.ScratchGpr().cvt32();
|
||||||
code.vpextrw(tmp, result, fsize == 32 ? 1 : 3);
|
code.vpextrw(tmp, result, fsize == 32 ? 1 : 3);
|
||||||
code.and_(tmp.cvt16(), fsize == 32 ? 0x7f80 : 0x7ff0);
|
code.and_(tmp.cvt16(), fsize == 32 ? 0x7f80 : 0x7ff0);
|
||||||
code.cmp(tmp.cvt16(), fsize == 32 ? 0x7f00 : 0x7fe0);
|
code.cmp(tmp.cvt16(), fsize == 32 ? 0x7f00 : 0x7fe0);
|
||||||
ctx.reg_alloc.Release(tmp);
|
ctx.reg_alloc.Release(tmp);
|
||||||
|
|
||||||
code.jae(fallback, code.T_NEAR);
|
code.jae(fallback, code.T_NEAR);
|
||||||
|
|
||||||
FCODE(vmuls)(result, result, code.MConst(xword, FP::FPValue<FPT, false, -1, 1>()));
|
FCODE(vmuls)(result, result, code.MConst(xword, FP::FPValue<FPT, false, -1, 1>()));
|
||||||
code.L(end);
|
code.L(end);
|
||||||
|
|
||||||
code.SwitchToFarCode();
|
code.SwitchToFarCode();
|
||||||
code.L(fallback);
|
code.L(fallback);
|
||||||
|
|
||||||
code.sub(rsp, 8);
|
code.sub(rsp, 8);
|
||||||
ABI_PushCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
|
ABI_PushCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
|
||||||
code.movq(code.ABI_PARAM1, operand1);
|
code.movq(code.ABI_PARAM1, operand1);
|
||||||
code.movq(code.ABI_PARAM2, operand2);
|
code.movq(code.ABI_PARAM2, operand2);
|
||||||
code.mov(code.ABI_PARAM3.cvt32(), ctx.FPCR().Value());
|
code.mov(code.ABI_PARAM3.cvt32(), ctx.FPCR().Value());
|
||||||
code.lea(code.ABI_PARAM4, code.ptr[code.r15 + code.GetJitStateInfo().offsetof_fpsr_exc]);
|
code.lea(code.ABI_PARAM4, code.ptr[code.r15 + code.GetJitStateInfo().offsetof_fpsr_exc]);
|
||||||
code.CallFunction(&FP::FPRSqrtStepFused<FPT>);
|
code.CallFunction(&FP::FPRSqrtStepFused<FPT>);
|
||||||
code.movq(result, code.ABI_RETURN);
|
code.movq(result, code.ABI_RETURN);
|
||||||
ABI_PopCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
|
ABI_PopCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
|
||||||
code.add(rsp, 8);
|
code.add(rsp, 8);
|
||||||
|
|
||||||
code.jmp(end, code.T_NEAR);
|
code.jmp(end, code.T_NEAR);
|
||||||
code.SwitchToNearCode();
|
code.SwitchToNearCode();
|
||||||
|
|
||||||
ctx.reg_alloc.DefineValue(inst, result);
|
ctx.reg_alloc.DefineValue(inst, result);
|
||||||
return;
|
return;
|
||||||
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||||
|
@ -1001,6 +1003,10 @@ static void EmitFPRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst*
|
||||||
code.CallFunction(&FP::FPRSqrtStepFused<FPT>);
|
code.CallFunction(&FP::FPRSqrtStepFused<FPT>);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitFPRSqrtStepFused16(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitFPRSqrtStepFused<16>(code, ctx, inst);
|
||||||
|
}
|
||||||
|
|
||||||
void EmitX64::EmitFPRSqrtStepFused32(EmitContext& ctx, IR::Inst* inst) {
|
void EmitX64::EmitFPRSqrtStepFused32(EmitContext& ctx, IR::Inst* inst) {
|
||||||
EmitFPRSqrtStepFused<32>(code, ctx, inst);
|
EmitFPRSqrtStepFused<32>(code, ctx, inst);
|
||||||
}
|
}
|
||||||
|
|
|
@ -1997,11 +1997,20 @@ U16U32U64 IREmitter::FPRSqrtEstimate(const U16U32U64& a) {
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
U32U64 IREmitter::FPRSqrtStepFused(const U32U64& a, const U32U64& b) {
|
U16U32U64 IREmitter::FPRSqrtStepFused(const U16U32U64& a, const U16U32U64& b) {
|
||||||
if (a.GetType() == Type::U32) {
|
ASSERT(a.GetType() == b.GetType());
|
||||||
|
|
||||||
|
switch (a.GetType()) {
|
||||||
|
case Type::U16:
|
||||||
|
return Inst<U16>(Opcode::FPRSqrtStepFused16, a, b);
|
||||||
|
case Type::U32:
|
||||||
return Inst<U32>(Opcode::FPRSqrtStepFused32, a, b);
|
return Inst<U32>(Opcode::FPRSqrtStepFused32, a, b);
|
||||||
|
case Type::U64:
|
||||||
|
return Inst<U64>(Opcode::FPRSqrtStepFused64, a, b);
|
||||||
|
default:
|
||||||
|
UNREACHABLE();
|
||||||
|
return U16U32U64{};
|
||||||
}
|
}
|
||||||
return Inst<U64>(Opcode::FPRSqrtStepFused64, a, b);
|
|
||||||
}
|
}
|
||||||
|
|
||||||
U32U64 IREmitter::FPSqrt(const U32U64& a) {
|
U32U64 IREmitter::FPSqrt(const U32U64& a) {
|
||||||
|
|
|
@ -310,7 +310,7 @@ public:
|
||||||
U16U32U64 FPRecipStepFused(const U16U32U64& a, const U16U32U64& b);
|
U16U32U64 FPRecipStepFused(const U16U32U64& a, const U16U32U64& b);
|
||||||
U16U32U64 FPRoundInt(const U16U32U64& a, FP::RoundingMode rounding, bool exact);
|
U16U32U64 FPRoundInt(const U16U32U64& a, FP::RoundingMode rounding, bool exact);
|
||||||
U16U32U64 FPRSqrtEstimate(const U16U32U64& a);
|
U16U32U64 FPRSqrtEstimate(const U16U32U64& a);
|
||||||
U32U64 FPRSqrtStepFused(const U32U64& a, const U32U64& b);
|
U16U32U64 FPRSqrtStepFused(const U16U32U64& a, const U16U32U64& b);
|
||||||
U32U64 FPSqrt(const U32U64& a);
|
U32U64 FPSqrt(const U32U64& a);
|
||||||
U32U64 FPSub(const U32U64& a, const U32U64& b, bool fpcr_controlled);
|
U32U64 FPSub(const U32U64& a, const U32U64& b, bool fpcr_controlled);
|
||||||
U16 FPDoubleToHalf(const U64& a, FP::RoundingMode rounding);
|
U16 FPDoubleToHalf(const U64& a, FP::RoundingMode rounding);
|
||||||
|
|
|
@ -287,6 +287,7 @@ bool Inst::ReadsFromAndWritesToFPSRCumulativeExceptionBits() const {
|
||||||
case Opcode::FPRSqrtEstimate16:
|
case Opcode::FPRSqrtEstimate16:
|
||||||
case Opcode::FPRSqrtEstimate32:
|
case Opcode::FPRSqrtEstimate32:
|
||||||
case Opcode::FPRSqrtEstimate64:
|
case Opcode::FPRSqrtEstimate64:
|
||||||
|
case Opcode::FPRSqrtStepFused16:
|
||||||
case Opcode::FPRSqrtStepFused32:
|
case Opcode::FPRSqrtStepFused32:
|
||||||
case Opcode::FPRSqrtStepFused64:
|
case Opcode::FPRSqrtStepFused64:
|
||||||
case Opcode::FPSqrt32:
|
case Opcode::FPSqrt32:
|
||||||
|
|
|
@ -506,6 +506,7 @@ OPCODE(FPRoundInt64, U64, U64,
|
||||||
OPCODE(FPRSqrtEstimate16, U16, U16 )
|
OPCODE(FPRSqrtEstimate16, U16, U16 )
|
||||||
OPCODE(FPRSqrtEstimate32, U32, U32 )
|
OPCODE(FPRSqrtEstimate32, U32, U32 )
|
||||||
OPCODE(FPRSqrtEstimate64, U64, U64 )
|
OPCODE(FPRSqrtEstimate64, U64, U64 )
|
||||||
|
OPCODE(FPRSqrtStepFused16, U16, U16, U16 )
|
||||||
OPCODE(FPRSqrtStepFused32, U32, U32, U32 )
|
OPCODE(FPRSqrtStepFused32, U32, U32, U32 )
|
||||||
OPCODE(FPRSqrtStepFused64, U64, U64, U64 )
|
OPCODE(FPRSqrtStepFused64, U64, U64, U64 )
|
||||||
OPCODE(FPSqrt32, U32, U32 )
|
OPCODE(FPSqrt32, U32, U32 )
|
||||||
|
|
Loading…
Reference in a new issue