frontend/ir_emitter: Add half-precision opcode variant of FPVectorRSqrtStepFused

This commit is contained in:
Lioncash 2019-04-14 21:01:00 -04:00 committed by MerryMage
parent 824c551ba2
commit 96356fac93
4 changed files with 45 additions and 35 deletions

View file

@ -1273,51 +1273,57 @@ static void EmitRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
}
};
if (code.DoesCpuSupport(Xbyak::util::Cpu::tFMA) && code.DoesCpuSupport(Xbyak::util::Cpu::tAVX)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
if constexpr (fsize != 16) {
if (code.DoesCpuSupport(Xbyak::util::Cpu::tFMA) && code.DoesCpuSupport(Xbyak::util::Cpu::tAVX)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm mask = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm mask = ctx.reg_alloc.ScratchXmm();
Xbyak::Label end, fallback;
Xbyak::Label end, fallback;
code.vmovaps(result, GetVectorOf<fsize, false, 0, 3>(code));
FCODE(vfnmadd231p)(result, operand1, operand2);
code.vmovaps(result, GetVectorOf<fsize, false, 0, 3>(code));
FCODE(vfnmadd231p)(result, operand1, operand2);
// An explanation for this is given in EmitFPRSqrtStepFused.
code.vmovaps(mask, GetVectorOf<fsize, fsize == 32 ? 0x7f000000 : 0x7fe0000000000000>(code));
FCODE(vandp)(tmp, result, mask);
if constexpr (fsize == 32) {
code.vpcmpeqd(tmp, tmp, mask);
} else {
code.vpcmpeqq(tmp, tmp, mask);
// An explanation for this is given in EmitFPRSqrtStepFused.
code.vmovaps(mask, GetVectorOf<fsize, fsize == 32 ? 0x7f000000 : 0x7fe0000000000000>(code));
FCODE(vandp)(tmp, result, mask);
if constexpr (fsize == 32) {
code.vpcmpeqd(tmp, tmp, mask);
} else {
code.vpcmpeqq(tmp, tmp, mask);
}
code.ptest(tmp, tmp);
code.jnz(fallback, code.T_NEAR);
FCODE(vmulp)(result, result, GetVectorOf<fsize, false, -1, 1>(code));
code.L(end);
code.SwitchToFarCode();
code.L(fallback);
code.sub(rsp, 8);
ABI_PushCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
EmitThreeOpFallbackWithoutRegAlloc(code, ctx, result, operand1, operand2, fallback_fn);
ABI_PopCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
code.add(rsp, 8);
code.jmp(end, code.T_NEAR);
code.SwitchToNearCode();
ctx.reg_alloc.DefineValue(inst, result);
return;
}
code.ptest(tmp, tmp);
code.jnz(fallback, code.T_NEAR);
FCODE(vmulp)(result, result, GetVectorOf<fsize, false, -1, 1>(code));
code.L(end);
code.SwitchToFarCode();
code.L(fallback);
code.sub(rsp, 8);
ABI_PushCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
EmitThreeOpFallbackWithoutRegAlloc(code, ctx, result, operand1, operand2, fallback_fn);
ABI_PopCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
code.add(rsp, 8);
code.jmp(end, code.T_NEAR);
code.SwitchToNearCode();
ctx.reg_alloc.DefineValue(inst, result);
return;
}
EmitThreeOpFallback(code, ctx, inst, fallback_fn);
}
void EmitX64::EmitFPVectorRSqrtStepFused16(EmitContext& ctx, IR::Inst* inst) {
EmitRSqrtStepFused<16>(code, ctx, inst);
}
void EmitX64::EmitFPVectorRSqrtStepFused32(EmitContext& ctx, IR::Inst* inst) {
EmitRSqrtStepFused<32>(code, ctx, inst);
}

View file

@ -2344,6 +2344,8 @@ U128 IREmitter::FPVectorRSqrtEstimate(size_t esize, const U128& a) {
U128 IREmitter::FPVectorRSqrtStepFused(size_t esize, const U128& a, const U128& b) {
switch (esize) {
case 16:
return Inst<U128>(Opcode::FPVectorRSqrtStepFused16, a, b);
case 32:
return Inst<U128>(Opcode::FPVectorRSqrtStepFused32, a, b);
case 64:

View file

@ -351,6 +351,7 @@ bool Inst::ReadsFromAndWritesToFPSRCumulativeExceptionBits() const {
case Opcode::FPVectorRSqrtEstimate16:
case Opcode::FPVectorRSqrtEstimate32:
case Opcode::FPVectorRSqrtEstimate64:
case Opcode::FPVectorRSqrtStepFused16:
case Opcode::FPVectorRSqrtStepFused32:
case Opcode::FPVectorRSqrtStepFused64:
case Opcode::FPVectorSqrt32:

View file

@ -586,6 +586,7 @@ OPCODE(FPVectorRoundInt64, U128, U128
OPCODE(FPVectorRSqrtEstimate16, U128, U128 )
OPCODE(FPVectorRSqrtEstimate32, U128, U128 )
OPCODE(FPVectorRSqrtEstimate64, U128, U128 )
OPCODE(FPVectorRSqrtStepFused16, U128, U128, U128 )
OPCODE(FPVectorRSqrtStepFused32, U128, U128, U128 )
OPCODE(FPVectorRSqrtStepFused64, U128, U128, U128 )
OPCODE(FPVectorSqrt32, U128, U128 )