mirrors/dynarmic
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

ir/frontend: Add half-precision opcode for FPVectorMulAdd

Browse source
This commit is contained in:
Lioncash 2019-04-13 01:42:35 -04:00 committed by MerryMage
parent 5f74d25bf7
commit ec6b3ae084
4 changed files with 37 additions and 27 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

60
src/backend/x64/emit_x64_vector_floating_point.cpp
View file

@ -908,44 +908,50 @@ void EmitFPVectorMulAdd(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
}
};
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 xmm_a = ctx.reg_alloc.UseXmm(args[0]);
const Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm xmm_c = ctx.reg_alloc.UseXmm(args[2]);
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm xmm_a = ctx.reg_alloc.UseXmm(args[0]);
const Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm xmm_c = ctx.reg_alloc.UseXmm(args[2]);
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
Xbyak::Label end, fallback;
Xbyak::Label end, fallback;
code.movaps(result, xmm_a);
FCODE(vfmadd231p)(result, xmm_b, xmm_c);
code.movaps(result, xmm_a);
FCODE(vfmadd231p)(result, xmm_b, xmm_c);
code.movaps(tmp, GetNegativeZeroVector<fsize>(code));
code.andnps(tmp, result);
FCODE(vcmpeq_uqp)(tmp, tmp, GetSmallestNormalVector<fsize>(code));
code.vptest(tmp, tmp);
code.jnz(fallback, code.T_NEAR);
code.L(end);
code.movaps(tmp, GetNegativeZeroVector<fsize>(code));
code.andnps(tmp, result);
FCODE(vcmpeq_uqp)(tmp, tmp, GetSmallestNormalVector<fsize>(code));
code.vptest(tmp, tmp);
code.jnz(fallback, code.T_NEAR);
code.L(end);
code.SwitchToFarCode();
code.L(fallback);
code.sub(rsp, 8);
ABI_PushCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
EmitFourOpFallbackWithoutRegAlloc(code, ctx, result, xmm_a, xmm_b, xmm_c, fallback_fn);
ABI_PopCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
code.add(rsp, 8);
code.jmp(end, code.T_NEAR);
code.SwitchToNearCode();
code.SwitchToFarCode();
code.L(fallback);
code.sub(rsp, 8);
ABI_PushCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
EmitFourOpFallbackWithoutRegAlloc(code, ctx, result, xmm_a, xmm_b, xmm_c, 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;
ctx.reg_alloc.DefineValue(inst, result);
return;
}
}
EmitFourOpFallback(code, ctx, inst, fallback_fn);
}
void EmitX64::EmitFPVectorMulAdd16(EmitContext& ctx, IR::Inst* inst) {
EmitFPVectorMulAdd<16>(code, ctx, inst);
}
void EmitX64::EmitFPVectorMulAdd32(EmitContext& ctx, IR::Inst* inst) {
EmitFPVectorMulAdd<32>(code, ctx, inst);
}

2
src/frontend/ir/ir_emitter.cpp
View file

@ -2173,6 +2173,8 @@ U128 IREmitter::FPVectorMul(size_t esize, const U128& a, const U128& b) {
U128 IREmitter::FPVectorMulAdd(size_t esize, const U128& a, const U128& b, const U128& c) {
switch (esize) {
case 16:
return Inst<U128>(Opcode::FPVectorMulAdd16, a, b, c);
case 32:
return Inst<U128>(Opcode::FPVectorMulAdd32, a, b, c);
case 64:

1
src/frontend/ir/microinstruction.cpp
View file

@ -327,6 +327,7 @@ bool Inst::ReadsFromAndWritesToFPSRCumulativeExceptionBits() const {
case Opcode::FPVectorGreaterEqual64:
case Opcode::FPVectorMul32:
case Opcode::FPVectorMul64:
case Opcode::FPVectorMulAdd16:
case Opcode::FPVectorMulAdd32:
case Opcode::FPVectorMulAdd64:
case Opcode::FPVectorPairedAddLower32:

1
src/frontend/ir/opcodes.inc
View file

@ -553,6 +553,7 @@ OPCODE(FPVectorMin32, U128, U128
OPCODE(FPVectorMin64, U128, U128, U128 )
OPCODE(FPVectorMul32, U128, U128, U128 )
OPCODE(FPVectorMul64, U128, U128, U128 )
OPCODE(FPVectorMulAdd16, U128, U128, U128, U128 )
OPCODE(FPVectorMulAdd32, U128, U128, U128, U128 )
OPCODE(FPVectorMulAdd64, U128, U128, U128, U128 )
OPCODE(FPVectorMulX32, U128, U128, U128 )