frontend/ir_emitter: Add half-precision opcode for FPVectorRecipStepFused
This commit is contained in:
parent
6da0411111
commit
5d5c9f149f
4 changed files with 34 additions and 24 deletions
|
@ -1110,41 +1110,47 @@ static void EmitRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
|
||||||
}
|
}
|
||||||
};
|
};
|
||||||
|
|
||||||
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);
|
||||||
|
|
||||||
const Xbyak::Xmm result = 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 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 tmp = ctx.reg_alloc.ScratchXmm();
|
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
|
||||||
|
|
||||||
Xbyak::Label end, fallback;
|
Xbyak::Label end, fallback;
|
||||||
|
|
||||||
code.movaps(result, GetVectorOf<fsize, false, 0, 2>(code));
|
code.movaps(result, GetVectorOf<fsize, false, 0, 2>(code));
|
||||||
FCODE(vfnmadd231p)(result, operand1, operand2);
|
FCODE(vfnmadd231p)(result, operand1, operand2);
|
||||||
|
|
||||||
FCODE(vcmpunordp)(tmp, result, result);
|
FCODE(vcmpunordp)(tmp, result, result);
|
||||||
code.vptest(tmp, tmp);
|
code.vptest(tmp, tmp);
|
||||||
code.jnz(fallback, code.T_NEAR);
|
code.jnz(fallback, code.T_NEAR);
|
||||||
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()));
|
||||||
EmitThreeOpFallbackWithoutRegAlloc(code, ctx, result, operand1, operand2, fallback_fn);
|
EmitThreeOpFallbackWithoutRegAlloc(code, ctx, result, operand1, operand2, fallback_fn);
|
||||||
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;
|
||||||
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
EmitThreeOpFallback(code, ctx, inst, fallback_fn);
|
EmitThreeOpFallback(code, ctx, inst, fallback_fn);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitFPVectorRecipStepFused16(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitRecipStepFused<16>(code, ctx, inst);
|
||||||
|
}
|
||||||
|
|
||||||
void EmitX64::EmitFPVectorRecipStepFused32(EmitContext& ctx, IR::Inst* inst) {
|
void EmitX64::EmitFPVectorRecipStepFused32(EmitContext& ctx, IR::Inst* inst) {
|
||||||
EmitRecipStepFused<32>(code, ctx, inst);
|
EmitRecipStepFused<32>(code, ctx, inst);
|
||||||
}
|
}
|
||||||
|
|
|
@ -2277,6 +2277,8 @@ U128 IREmitter::FPVectorRecipEstimate(size_t esize, const U128& a) {
|
||||||
|
|
||||||
U128 IREmitter::FPVectorRecipStepFused(size_t esize, const U128& a, const U128& b) {
|
U128 IREmitter::FPVectorRecipStepFused(size_t esize, const U128& a, const U128& b) {
|
||||||
switch (esize) {
|
switch (esize) {
|
||||||
|
case 16:
|
||||||
|
return Inst<U128>(Opcode::FPVectorRecipStepFused16, a, b);
|
||||||
case 32:
|
case 32:
|
||||||
return Inst<U128>(Opcode::FPVectorRecipStepFused32, a, b);
|
return Inst<U128>(Opcode::FPVectorRecipStepFused32, a, b);
|
||||||
case 64:
|
case 64:
|
||||||
|
|
|
@ -338,6 +338,7 @@ bool Inst::ReadsFromAndWritesToFPSRCumulativeExceptionBits() const {
|
||||||
case Opcode::FPVectorPairedAdd64:
|
case Opcode::FPVectorPairedAdd64:
|
||||||
case Opcode::FPVectorRecipEstimate32:
|
case Opcode::FPVectorRecipEstimate32:
|
||||||
case Opcode::FPVectorRecipEstimate64:
|
case Opcode::FPVectorRecipEstimate64:
|
||||||
|
case Opcode::FPVectorRecipStepFused16:
|
||||||
case Opcode::FPVectorRecipStepFused32:
|
case Opcode::FPVectorRecipStepFused32:
|
||||||
case Opcode::FPVectorRecipStepFused64:
|
case Opcode::FPVectorRecipStepFused64:
|
||||||
case Opcode::FPVectorRoundInt16:
|
case Opcode::FPVectorRoundInt16:
|
||||||
|
|
|
@ -573,6 +573,7 @@ OPCODE(FPVectorPairedAddLower32, U128, U128
|
||||||
OPCODE(FPVectorPairedAddLower64, U128, U128, U128 )
|
OPCODE(FPVectorPairedAddLower64, U128, U128, U128 )
|
||||||
OPCODE(FPVectorRecipEstimate32, U128, U128 )
|
OPCODE(FPVectorRecipEstimate32, U128, U128 )
|
||||||
OPCODE(FPVectorRecipEstimate64, U128, U128 )
|
OPCODE(FPVectorRecipEstimate64, U128, U128 )
|
||||||
|
OPCODE(FPVectorRecipStepFused16, U128, U128, U128 )
|
||||||
OPCODE(FPVectorRecipStepFused32, U128, U128, U128 )
|
OPCODE(FPVectorRecipStepFused32, U128, U128, U128 )
|
||||||
OPCODE(FPVectorRecipStepFused64, U128, U128, U128 )
|
OPCODE(FPVectorRecipStepFused64, U128, U128, U128 )
|
||||||
OPCODE(FPVectorRoundInt16, U128, U128, U8, U1 )
|
OPCODE(FPVectorRoundInt16, U128, U128, U8, U1 )
|
||||||
|
|
Loading…
Reference in a new issue