frontend/ir_emitter: Add half-precision opcode variant of FPVectorRSqrtStepFused
This commit is contained in:
parent
824c551ba2
commit
96356fac93
4 changed files with 45 additions and 35 deletions
|
@ -1273,6 +1273,7 @@ static void EmitRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
|
|||
}
|
||||
};
|
||||
|
||||
if constexpr (fsize != 16) {
|
||||
if (code.DoesCpuSupport(Xbyak::util::Cpu::tFMA) && code.DoesCpuSupport(Xbyak::util::Cpu::tAVX)) {
|
||||
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||
|
||||
|
@ -1314,10 +1315,15 @@ static void EmitRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
|
|||
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);
|
||||
}
|
||||
|
|
|
@ -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:
|
||||
|
|
|
@ -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:
|
||||
|
|
|
@ -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 )
|
||||
|
|
Loading…
Reference in a new issue