SwitchEmu/dynarmic
1
0
Fork 0
mirror of https://git.suyu.dev/suyu/dynarmic.git synced 2026-03-05 10:16:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

frontend/ir_emitter: Add half-precision->fixed-point opcodes

Browse Source
This commit is contained in:
Lioncash
2019-04-15 00:20:37 -04:00
committed by MerryMage
parent 4ecfbc14de
commit 604f39f00a
6 changed files with 270 additions and 163 deletions
Download Patch File Download Diff File Expand all files Collapse all files

141
src/backend/x64/emit_x64_floating_point.cpp
View File

@@ -1222,74 +1222,77 @@ static void EmitFPToFixed(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
const size_t fbits = args[1].GetImmediateU8();
const auto rounding_mode = static_cast<FP::RoundingMode>(args[2].GetImmediateU8());
const auto round_imm = ConvertRoundingModeToX64Immediate(rounding_mode);
if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41) && round_imm){
const Xbyak::Xmm src = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm scratch = ctx.reg_alloc.ScratchXmm();
const Xbyak::Reg64 result = ctx.reg_alloc.ScratchGpr().cvt64();
if constexpr (fsize != 16) {
const auto round_imm = ConvertRoundingModeToX64Immediate(rounding_mode);
if constexpr (fsize == 64) {
if (fbits != 0) {
const u64 scale_factor = static_cast<u64>((fbits + 1023) << 52);
code.mulsd(src, code.MConst(xword, scale_factor));
}
if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41) && round_imm){
const Xbyak::Xmm src = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm scratch = ctx.reg_alloc.ScratchXmm();
const Xbyak::Reg64 result = ctx.reg_alloc.ScratchGpr().cvt64();
code.roundsd(src, src, *round_imm);
} else {
if (fbits != 0) {
const u32 scale_factor = static_cast<u32>((fbits + 127) << 23);
code.mulss(src, code.MConst(xword, scale_factor));
}
if constexpr (fsize == 64) {
if (fbits != 0) {
const u64 scale_factor = static_cast<u64>((fbits + 1023) << 52);
code.mulsd(src, code.MConst(xword, scale_factor));
}
code.roundss(src, src, *round_imm);
code.cvtss2sd(src, src);
}
ZeroIfNaN<64>(code, src, scratch);
if constexpr (isize == 64) {
Xbyak::Label saturate_max, end;
if (unsigned_) {
code.maxsd(src, code.MConst(xword, f64_min_u64));
}
code.movsd(scratch, code.MConst(xword, unsigned_ ? f64_max_u64_lim : f64_max_s64_lim));
code.comisd(scratch, src);
code.jna(saturate_max, code.T_NEAR);
if (unsigned_) {
Xbyak::Label below_max;
code.movsd(scratch, code.MConst(xword, f64_max_s64_lim));
code.comisd(src, scratch);
code.jb(below_max);
code.subsd(src, scratch);
code.cvttsd2si(result, src);
code.btc(result, 63);
code.jmp(end);
code.L(below_max);
}
code.cvttsd2si(result, src); // 64 bit gpr
code.L(end);
code.SwitchToFarCode();
code.L(saturate_max);
code.mov(result, unsigned_ ? 0xFFFF'FFFF'FFFF'FFFF : 0x7FFF'FFFF'FFFF'FFFF);
code.jmp(end, code.T_NEAR);
code.SwitchToNearCode();
} else {
code.minsd(src, code.MConst(xword, unsigned_ ? f64_max_u32 : f64_max_s32));
if (unsigned_) {
code.maxsd(src, code.MConst(xword, f64_min_u32));
code.cvttsd2si(result, src); // 64 bit gpr
code.roundsd(src, src, *round_imm);
} else {
code.cvttsd2si(result.cvt32(), src);
if (fbits != 0) {
const u32 scale_factor = static_cast<u32>((fbits + 127) << 23);
code.mulss(src, code.MConst(xword, scale_factor));
}
code.roundss(src, src, *round_imm);
code.cvtss2sd(src, src);
}
ZeroIfNaN<64>(code, src, scratch);
if constexpr (isize == 64) {
Xbyak::Label saturate_max, end;
if (unsigned_) {
code.maxsd(src, code.MConst(xword, f64_min_u64));
}
code.movsd(scratch, code.MConst(xword, unsigned_ ? f64_max_u64_lim : f64_max_s64_lim));
code.comisd(scratch, src);
code.jna(saturate_max, code.T_NEAR);
if (unsigned_) {
Xbyak::Label below_max;
code.movsd(scratch, code.MConst(xword, f64_max_s64_lim));
code.comisd(src, scratch);
code.jb(below_max);
code.subsd(src, scratch);
code.cvttsd2si(result, src);
code.btc(result, 63);
code.jmp(end);
code.L(below_max);
}
code.cvttsd2si(result, src); // 64 bit gpr
code.L(end);
code.SwitchToFarCode();
code.L(saturate_max);
code.mov(result, unsigned_ ? 0xFFFF'FFFF'FFFF'FFFF : 0x7FFF'FFFF'FFFF'FFFF);
code.jmp(end, code.T_NEAR);
code.SwitchToNearCode();
} else {
code.minsd(src, code.MConst(xword, unsigned_ ? f64_max_u32 : f64_max_s32));
if (unsigned_) {
code.maxsd(src, code.MConst(xword, f64_min_u32));
code.cvttsd2si(result, src); // 64 bit gpr
} else {
code.cvttsd2si(result.cvt32(), src);
}
}
ctx.reg_alloc.DefineValue(inst, result);
return;
}
ctx.reg_alloc.DefineValue(inst, result);
return;
}
using fbits_list = mp::vllift<std::make_index_sequence<isize + 1>>;
@@ -1345,6 +1348,22 @@ void EmitX64::EmitFPDoubleToFixedU64(EmitContext& ctx, IR::Inst* inst) {
EmitFPToFixed<64, true, 64>(code, ctx, inst);
}
void EmitX64::EmitFPHalfToFixedS32(EmitContext& ctx, IR::Inst* inst) {
EmitFPToFixed<16, false, 32>(code, ctx, inst);
}
void EmitX64::EmitFPHalfToFixedS64(EmitContext& ctx, IR::Inst* inst) {
EmitFPToFixed<16, false, 64>(code, ctx, inst);
}
void EmitX64::EmitFPHalfToFixedU32(EmitContext& ctx, IR::Inst* inst) {
EmitFPToFixed<16, true, 32>(code, ctx, inst);
}
void EmitX64::EmitFPHalfToFixedU64(EmitContext& ctx, IR::Inst* inst) {
EmitFPToFixed<16, true, 64>(code, ctx, inst);
}
void EmitX64::EmitFPSingleToFixedS32(EmitContext& ctx, IR::Inst* inst) {
EmitFPToFixed<32, false, 32>(code, ctx, inst);
}

174
src/backend/x64/emit_x64_vector_floating_point.cpp
View File

@@ -1355,98 +1355,100 @@ void EmitFPVectorToFixed(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
// TODO: AVX512 implementation
if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41) && rounding != FP::RoundingMode::ToNearest_TieAwayFromZero) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
if constexpr (fsize != 16) {
if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41) && rounding != FP::RoundingMode::ToNearest_TieAwayFromZero) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm src = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm src = ctx.reg_alloc.UseScratchXmm(args[0]);
const int round_imm = [&]{
switch (rounding) {
case FP::RoundingMode::ToNearest_TieEven:
default:
return 0b00;
case FP::RoundingMode::TowardsPlusInfinity:
return 0b10;
case FP::RoundingMode::TowardsMinusInfinity:
return 0b01;
case FP::RoundingMode::TowardsZero:
return 0b11;
const int round_imm = [&]{
switch (rounding) {
case FP::RoundingMode::ToNearest_TieEven:
default:
return 0b00;
case FP::RoundingMode::TowardsPlusInfinity:
return 0b10;
case FP::RoundingMode::TowardsMinusInfinity:
return 0b01;
case FP::RoundingMode::TowardsZero:
return 0b11;
}
}();
const auto perform_conversion = [&code, &ctx](const Xbyak::Xmm& src) {
// MSVC doesn't allow us to use a [&] capture, so we have to do this instead.
(void)ctx;
if constexpr (fsize == 32) {
code.cvttps2dq(src, src);
} else {
const Xbyak::Reg64 hi = ctx.reg_alloc.ScratchGpr();
const Xbyak::Reg64 lo = ctx.reg_alloc.ScratchGpr();
code.cvttsd2si(lo, src);
code.punpckhqdq(src, src);
code.cvttsd2si(hi, src);
code.movq(src, lo);
code.pinsrq(src, hi, 1);
ctx.reg_alloc.Release(hi);
ctx.reg_alloc.Release(lo);
}
};
if (fbits != 0) {
const u64 scale_factor = fsize == 32
? static_cast<u64>(fbits + 127) << 23
: static_cast<u64>(fbits + 1023) << 52;
FCODE(mulp)(src, GetVectorOf<fsize>(code, scale_factor));
}
}();
const auto perform_conversion = [&code, &ctx](const Xbyak::Xmm& src) {
// MSVC doesn't allow us to use a [&] capture, so we have to do this instead.
(void)ctx;
FCODE(roundp)(src, src, static_cast<u8>(round_imm));
ZeroIfNaN<fsize>(code, src);
if constexpr (fsize == 32) {
code.cvttps2dq(src, src);
constexpr u64 float_upper_limit_signed = fsize == 32 ? 0x4f000000 : 0x43e0000000000000;
[[maybe_unused]] constexpr u64 float_upper_limit_unsigned = fsize == 32 ? 0x4f800000 : 0x43f0000000000000;
if constexpr (unsigned_) {
// Zero is minimum
code.xorps(xmm0, xmm0);
FCODE(cmplep)(xmm0, src);
FCODE(andp)(src, xmm0);
// Will we exceed unsigned range?
const Xbyak::Xmm exceed_unsigned = ctx.reg_alloc.ScratchXmm();
code.movaps(exceed_unsigned, GetVectorOf<fsize, float_upper_limit_unsigned>(code));
FCODE(cmplep)(exceed_unsigned, src);
// Will be exceed signed range?
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
code.movaps(tmp, GetVectorOf<fsize, float_upper_limit_signed>(code));
code.movaps(xmm0, tmp);
FCODE(cmplep)(xmm0, src);
FCODE(andp)(tmp, xmm0);
FCODE(subp)(src, tmp);
perform_conversion(src);
if constexpr (fsize == 32) {
code.pslld(xmm0, 31);
} else {
code.psllq(xmm0, 63);
}
FCODE(orp)(src, xmm0);
// Saturate to max
FCODE(orp)(src, exceed_unsigned);
} else {
const Xbyak::Reg64 hi = ctx.reg_alloc.ScratchGpr();
const Xbyak::Reg64 lo = ctx.reg_alloc.ScratchGpr();
constexpr u64 integer_max = static_cast<FPT>(std::numeric_limits<std::conditional_t<unsigned_, FPT, std::make_signed_t<FPT>>>::max());
code.cvttsd2si(lo, src);
code.punpckhqdq(src, src);
code.cvttsd2si(hi, src);
code.movq(src, lo);
code.pinsrq(src, hi, 1);
ctx.reg_alloc.Release(hi);
ctx.reg_alloc.Release(lo);
code.movaps(xmm0, GetVectorOf<fsize, float_upper_limit_signed>(code));
FCODE(cmplep)(xmm0, src);
perform_conversion(src);
FCODE(blendvp)(src, GetVectorOf<fsize, integer_max>(code));
}
};
if (fbits != 0) {
const u64 scale_factor = fsize == 32
? static_cast<u64>(fbits + 127) << 23
: static_cast<u64>(fbits + 1023) << 52;
FCODE(mulp)(src, GetVectorOf<fsize>(code, scale_factor));
ctx.reg_alloc.DefineValue(inst, src);
return;
}
FCODE(roundp)(src, src, static_cast<u8>(round_imm));
ZeroIfNaN<fsize>(code, src);
constexpr u64 float_upper_limit_signed = fsize == 32 ? 0x4f000000 : 0x43e0000000000000;
[[maybe_unused]] constexpr u64 float_upper_limit_unsigned = fsize == 32 ? 0x4f800000 : 0x43f0000000000000;
if constexpr (unsigned_) {
// Zero is minimum
code.xorps(xmm0, xmm0);
FCODE(cmplep)(xmm0, src);
FCODE(andp)(src, xmm0);
// Will we exceed unsigned range?
const Xbyak::Xmm exceed_unsigned = ctx.reg_alloc.ScratchXmm();
code.movaps(exceed_unsigned, GetVectorOf<fsize, float_upper_limit_unsigned>(code));
FCODE(cmplep)(exceed_unsigned, src);
// Will be exceed signed range?
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
code.movaps(tmp, GetVectorOf<fsize, float_upper_limit_signed>(code));
code.movaps(xmm0, tmp);
FCODE(cmplep)(xmm0, src);
FCODE(andp)(tmp, xmm0);
FCODE(subp)(src, tmp);
perform_conversion(src);
if constexpr (fsize == 32) {
code.pslld(xmm0, 31);
} else {
code.psllq(xmm0, 63);
}
FCODE(orp)(src, xmm0);
// Saturate to max
FCODE(orp)(src, exceed_unsigned);
} else {
constexpr u64 integer_max = static_cast<FPT>(std::numeric_limits<std::conditional_t<unsigned_, FPT, std::make_signed_t<FPT>>>::max());
code.movaps(xmm0, GetVectorOf<fsize, float_upper_limit_signed>(code));
FCODE(cmplep)(xmm0, src);
perform_conversion(src);
FCODE(blendvp)(src, GetVectorOf<fsize, integer_max>(code));
}
ctx.reg_alloc.DefineValue(inst, src);
return;
}
using fbits_list = mp::vllift<std::make_index_sequence<fsize + 1>>;
@@ -1483,6 +1485,10 @@ void EmitFPVectorToFixed(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
EmitTwoOpFallback(code, ctx, inst, lut.at(std::make_tuple(fbits, rounding)));
}
void EmitX64::EmitFPVectorToSignedFixed16(EmitContext& ctx, IR::Inst* inst) {
EmitFPVectorToFixed<16, false>(code, ctx, inst);
}
void EmitX64::EmitFPVectorToSignedFixed32(EmitContext& ctx, IR::Inst* inst) {
EmitFPVectorToFixed<32, false>(code, ctx, inst);
}
@@ -1491,6 +1497,10 @@ void EmitX64::EmitFPVectorToSignedFixed64(EmitContext& ctx, IR::Inst* inst) {
EmitFPVectorToFixed<64, false>(code, ctx, inst);
}
void EmitX64::EmitFPVectorToUnsignedFixed16(EmitContext& ctx, IR::Inst* inst) {
EmitFPVectorToFixed<16, true>(code, ctx, inst);
}
void EmitX64::EmitFPVectorToUnsignedFixed32(EmitContext& ctx, IR::Inst* inst) {
EmitFPVectorToFixed<32, true>(code, ctx, inst);
}