IR: Implement FPVector{To,From}Half32
Implement ASIMD VCVT (half) in terms of this instruction. Correct handling of ASIMDStandardValue.
This commit is contained in:
parent
9a23c09c3b
commit
828959caed
7 changed files with 124 additions and 28 deletions
|
@ -5,8 +5,11 @@
|
||||||
|
|
||||||
#pragma once
|
#pragma once
|
||||||
|
|
||||||
|
#include <optional>
|
||||||
|
|
||||||
#include "dynarmic/common/bit_util.h"
|
#include "dynarmic/common/bit_util.h"
|
||||||
#include "dynarmic/common/common_types.h"
|
#include "dynarmic/common/common_types.h"
|
||||||
|
#include "dynarmic/common/fp/rounding_mode.h"
|
||||||
|
|
||||||
namespace Dynarmic::Backend::X64 {
|
namespace Dynarmic::Backend::X64 {
|
||||||
|
|
||||||
|
@ -84,4 +87,19 @@ constexpr u32 FixupLUT(FpFixup src_qnan = FpFixup::A,
|
||||||
return fixup_lut;
|
return fixup_lut;
|
||||||
}
|
}
|
||||||
|
|
||||||
|
constexpr std::optional<int> ConvertRoundingModeToX64Immediate(FP::RoundingMode rounding_mode) {
|
||||||
|
switch (rounding_mode) {
|
||||||
|
case FP::RoundingMode::ToNearest_TieEven:
|
||||||
|
return 0b00;
|
||||||
|
case FP::RoundingMode::TowardsPlusInfinity:
|
||||||
|
return 0b10;
|
||||||
|
case FP::RoundingMode::TowardsMinusInfinity:
|
||||||
|
return 0b01;
|
||||||
|
case FP::RoundingMode::TowardsZero:
|
||||||
|
return 0b11;
|
||||||
|
default:
|
||||||
|
return std::nullopt;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
} // namespace Dynarmic::Backend::X64
|
} // namespace Dynarmic::Backend::X64
|
||||||
|
|
|
@ -80,21 +80,6 @@ constexpr u64 f64_max_u64_lim = 0x43f0000000000000u; // 2^64 as a double (actua
|
||||||
} \
|
} \
|
||||||
}
|
}
|
||||||
|
|
||||||
std::optional<int> ConvertRoundingModeToX64Immediate(FP::RoundingMode rounding_mode) {
|
|
||||||
switch (rounding_mode) {
|
|
||||||
case FP::RoundingMode::ToNearest_TieEven:
|
|
||||||
return 0b00;
|
|
||||||
case FP::RoundingMode::TowardsPlusInfinity:
|
|
||||||
return 0b10;
|
|
||||||
case FP::RoundingMode::TowardsMinusInfinity:
|
|
||||||
return 0b01;
|
|
||||||
case FP::RoundingMode::TowardsZero:
|
|
||||||
return 0b11;
|
|
||||||
default:
|
|
||||||
return std::nullopt;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
template<size_t fsize>
|
template<size_t fsize>
|
||||||
void DenormalsAreZero(BlockOfCode& code, EmitContext& ctx, std::initializer_list<Xbyak::Xmm> to_daz) {
|
void DenormalsAreZero(BlockOfCode& code, EmitContext& ctx, std::initializer_list<Xbyak::Xmm> to_daz) {
|
||||||
if (ctx.FPCR().FZ()) {
|
if (ctx.FPCR().FZ()) {
|
||||||
|
|
|
@ -642,6 +642,49 @@ void EmitX64::EmitFPVectorEqual64(EmitContext& ctx, IR::Inst* inst) {
|
||||||
ctx.reg_alloc.DefineValue(inst, a);
|
ctx.reg_alloc.DefineValue(inst, a);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitFPVectorFromHalf32(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
const auto rounding_mode = static_cast<FP::RoundingMode>(inst->GetArg(1).GetU8());
|
||||||
|
const bool fpcr_controlled = inst->GetArg(2).GetU1();
|
||||||
|
|
||||||
|
if (code.HasHostFeature(HostFeature::F16C) && !ctx.FPCR().AHP() && !ctx.FPCR().FZ16()) {
|
||||||
|
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||||
|
|
||||||
|
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
|
||||||
|
const Xbyak::Xmm value = ctx.reg_alloc.UseXmm(args[0]);
|
||||||
|
|
||||||
|
code.vcvtph2ps(result, value);
|
||||||
|
ForceToDefaultNaN<32>(code, ctx.FPCR(fpcr_controlled), result);
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(inst, result);
|
||||||
|
return;
|
||||||
|
}
|
||||||
|
|
||||||
|
using rounding_list = mp::list<
|
||||||
|
mp::lift_value<FP::RoundingMode::ToNearest_TieEven>,
|
||||||
|
mp::lift_value<FP::RoundingMode::TowardsPlusInfinity>,
|
||||||
|
mp::lift_value<FP::RoundingMode::TowardsMinusInfinity>,
|
||||||
|
mp::lift_value<FP::RoundingMode::TowardsZero>,
|
||||||
|
mp::lift_value<FP::RoundingMode::ToNearest_TieAwayFromZero>>;
|
||||||
|
|
||||||
|
static const auto lut = Common::GenerateLookupTableFromList(
|
||||||
|
[](auto arg) {
|
||||||
|
return std::pair{
|
||||||
|
mp::lower_to_tuple_v<decltype(arg)>,
|
||||||
|
Common::FptrCast(
|
||||||
|
[](VectorArray<u32>& output, const VectorArray<u16>& input, FP::FPCR fpcr, FP::FPSR& fpsr) {
|
||||||
|
constexpr auto t = mp::lower_to_tuple_v<decltype(arg)>;
|
||||||
|
constexpr FP::RoundingMode rounding_mode = std::get<0>(t);
|
||||||
|
|
||||||
|
for (size_t i = 0; i < output.size(); ++i) {
|
||||||
|
output[i] = FP::FPConvert<u32, u16>(input[i], fpcr, rounding_mode, fpsr);
|
||||||
|
}
|
||||||
|
})};
|
||||||
|
},
|
||||||
|
mp::cartesian_product<rounding_list>{});
|
||||||
|
|
||||||
|
EmitTwoOpFallback<2>(code, ctx, inst, lut.at(std::make_tuple(rounding_mode)));
|
||||||
|
}
|
||||||
|
|
||||||
void EmitX64::EmitFPVectorFromSignedFixed32(EmitContext& ctx, IR::Inst* inst) {
|
void EmitX64::EmitFPVectorFromSignedFixed32(EmitContext& ctx, IR::Inst* inst) {
|
||||||
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||||
const Xbyak::Xmm xmm = ctx.reg_alloc.UseScratchXmm(args[0]);
|
const Xbyak::Xmm xmm = ctx.reg_alloc.UseScratchXmm(args[0]);
|
||||||
|
@ -1607,6 +1650,53 @@ void EmitX64::EmitFPVectorSub64(EmitContext& ctx, IR::Inst* inst) {
|
||||||
EmitThreeOpVectorOperation<64, DefaultIndexer>(code, ctx, inst, &Xbyak::CodeGenerator::subpd);
|
EmitThreeOpVectorOperation<64, DefaultIndexer>(code, ctx, inst, &Xbyak::CodeGenerator::subpd);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitFPVectorToHalf32(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
const auto rounding_mode = static_cast<FP::RoundingMode>(inst->GetArg(1).GetU8());
|
||||||
|
const bool fpcr_controlled = inst->GetArg(2).GetU1();
|
||||||
|
|
||||||
|
if (code.HasHostFeature(HostFeature::F16C) && !ctx.FPCR().AHP() && !ctx.FPCR().FZ16()) {
|
||||||
|
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||||
|
const auto round_imm = ConvertRoundingModeToX64Immediate(rounding_mode);
|
||||||
|
|
||||||
|
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
|
||||||
|
|
||||||
|
ForceToDefaultNaN<32>(code, ctx.FPCR(fpcr_controlled), result);
|
||||||
|
code.vcvtps2ph(result, result, static_cast<u8>(*round_imm));
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(inst, result);
|
||||||
|
return;
|
||||||
|
}
|
||||||
|
|
||||||
|
using rounding_list = mp::list<
|
||||||
|
mp::lift_value<FP::RoundingMode::ToNearest_TieEven>,
|
||||||
|
mp::lift_value<FP::RoundingMode::TowardsPlusInfinity>,
|
||||||
|
mp::lift_value<FP::RoundingMode::TowardsMinusInfinity>,
|
||||||
|
mp::lift_value<FP::RoundingMode::TowardsZero>,
|
||||||
|
mp::lift_value<FP::RoundingMode::ToNearest_TieAwayFromZero>>;
|
||||||
|
|
||||||
|
static const auto lut = Common::GenerateLookupTableFromList(
|
||||||
|
[](auto arg) {
|
||||||
|
return std::pair{
|
||||||
|
mp::lower_to_tuple_v<decltype(arg)>,
|
||||||
|
Common::FptrCast(
|
||||||
|
[](VectorArray<u16>& output, const VectorArray<u32>& input, FP::FPCR fpcr, FP::FPSR& fpsr) {
|
||||||
|
constexpr auto t = mp::lower_to_tuple_v<decltype(arg)>;
|
||||||
|
constexpr FP::RoundingMode rounding_mode = std::get<0>(t);
|
||||||
|
|
||||||
|
for (size_t i = 0; i < output.size(); ++i) {
|
||||||
|
if (i < input.size()) {
|
||||||
|
output[i] = FP::FPConvert<u16, u32>(input[i], fpcr, rounding_mode, fpsr);
|
||||||
|
} else {
|
||||||
|
output[i] = 0;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
})};
|
||||||
|
},
|
||||||
|
mp::cartesian_product<rounding_list>{});
|
||||||
|
|
||||||
|
EmitTwoOpFallback<2>(code, ctx, inst, lut.at(std::make_tuple(rounding_mode)));
|
||||||
|
}
|
||||||
|
|
||||||
template<size_t fsize, bool unsigned_>
|
template<size_t fsize, bool unsigned_>
|
||||||
void EmitFPVectorToFixed(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
|
void EmitFPVectorToFixed(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
|
||||||
using FPT = mp::unsigned_integer_of_size<fsize>;
|
using FPT = mp::unsigned_integer_of_size<fsize>;
|
||||||
|
|
|
@ -620,24 +620,13 @@ bool TranslatorVisitor::asimd_VCVT_half(bool D, size_t sz, size_t Vd, bool half_
|
||||||
}
|
}
|
||||||
|
|
||||||
const size_t esize = 8U << sz;
|
const size_t esize = 8U << sz;
|
||||||
const size_t num_elements = 4;
|
|
||||||
const auto rounding_mode = FP::RoundingMode::ToNearest_TieEven; // StandardFPSCRValue().RMode
|
const auto rounding_mode = FP::RoundingMode::ToNearest_TieEven; // StandardFPSCRValue().RMode
|
||||||
const auto d = ToVector(half_to_single, Vd, D);
|
const auto d = ToVector(half_to_single, Vd, D);
|
||||||
const auto m = ToVector(!half_to_single, Vm, M);
|
const auto m = ToVector(!half_to_single, Vm, M);
|
||||||
|
|
||||||
const auto operand = ir.GetVector(m);
|
const auto operand = ir.GetVector(m);
|
||||||
IR::U128 result = ir.ZeroVector();
|
const IR::U128 result = half_to_single ? ir.FPVectorFromHalf(esize * 2, operand, rounding_mode, false)
|
||||||
for (size_t i = 0; i < num_elements; i++) {
|
: ir.FPVectorToHalf(esize * 2, operand, rounding_mode, false);
|
||||||
if (half_to_single) {
|
|
||||||
const IR::U16 old_element = ir.VectorGetElement(esize, operand, i);
|
|
||||||
const IR::U32 new_element = ir.FPHalfToSingle(old_element, rounding_mode);
|
|
||||||
result = ir.VectorSetElement(esize * 2, result, i, new_element);
|
|
||||||
} else {
|
|
||||||
const IR::U32 old_element = ir.VectorGetElement(esize * 2, operand, i);
|
|
||||||
const IR::U16 new_element = ir.FPSingleToHalf(old_element, rounding_mode);
|
|
||||||
result = ir.VectorSetElement(esize, result, i, new_element);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
ir.SetVector(d, result);
|
ir.SetVector(d, result);
|
||||||
return true;
|
return true;
|
||||||
}
|
}
|
||||||
|
|
|
@ -2404,6 +2404,11 @@ U128 IREmitter::FPVectorEqual(size_t esize, const U128& a, const U128& b, bool f
|
||||||
UNREACHABLE();
|
UNREACHABLE();
|
||||||
}
|
}
|
||||||
|
|
||||||
|
U128 IREmitter::FPVectorFromHalf(size_t esize, const U128& a, FP::RoundingMode rounding, bool fpcr_controlled) {
|
||||||
|
ASSERT(esize == 32);
|
||||||
|
return Inst<U128>(Opcode::FPVectorFromHalf32, a, Imm8(static_cast<u8>(rounding)), Imm1(fpcr_controlled));
|
||||||
|
}
|
||||||
|
|
||||||
U128 IREmitter::FPVectorFromSignedFixed(size_t esize, const U128& a, size_t fbits, FP::RoundingMode rounding, bool fpcr_controlled) {
|
U128 IREmitter::FPVectorFromSignedFixed(size_t esize, const U128& a, size_t fbits, FP::RoundingMode rounding, bool fpcr_controlled) {
|
||||||
ASSERT(fbits <= esize);
|
ASSERT(fbits <= esize);
|
||||||
switch (esize) {
|
switch (esize) {
|
||||||
|
@ -2613,6 +2618,11 @@ U128 IREmitter::FPVectorSub(size_t esize, const U128& a, const U128& b, bool fpc
|
||||||
UNREACHABLE();
|
UNREACHABLE();
|
||||||
}
|
}
|
||||||
|
|
||||||
|
U128 IREmitter::FPVectorToHalf(size_t esize, const U128& a, FP::RoundingMode rounding, bool fpcr_controlled) {
|
||||||
|
ASSERT(esize == 32);
|
||||||
|
return Inst<U128>(Opcode::FPVectorToHalf32, a, Imm8(static_cast<u8>(rounding)), Imm1(fpcr_controlled));
|
||||||
|
}
|
||||||
|
|
||||||
U128 IREmitter::FPVectorToSignedFixed(size_t esize, const U128& a, size_t fbits, FP::RoundingMode rounding, bool fpcr_controlled) {
|
U128 IREmitter::FPVectorToSignedFixed(size_t esize, const U128& a, size_t fbits, FP::RoundingMode rounding, bool fpcr_controlled) {
|
||||||
ASSERT(fbits <= esize);
|
ASSERT(fbits <= esize);
|
||||||
|
|
||||||
|
|
|
@ -370,6 +370,7 @@ public:
|
||||||
U128 FPVectorAdd(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
U128 FPVectorAdd(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
||||||
U128 FPVectorDiv(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
U128 FPVectorDiv(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
||||||
U128 FPVectorEqual(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
U128 FPVectorEqual(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
||||||
|
U128 FPVectorFromHalf(size_t esize, const U128& a, FP::RoundingMode rounding, bool fpcr_controlled = true);
|
||||||
U128 FPVectorFromSignedFixed(size_t esize, const U128& a, size_t fbits, FP::RoundingMode rounding, bool fpcr_controlled = true);
|
U128 FPVectorFromSignedFixed(size_t esize, const U128& a, size_t fbits, FP::RoundingMode rounding, bool fpcr_controlled = true);
|
||||||
U128 FPVectorFromUnsignedFixed(size_t esize, const U128& a, size_t fbits, FP::RoundingMode rounding, bool fpcr_controlled = true);
|
U128 FPVectorFromUnsignedFixed(size_t esize, const U128& a, size_t fbits, FP::RoundingMode rounding, bool fpcr_controlled = true);
|
||||||
U128 FPVectorGreater(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
U128 FPVectorGreater(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
||||||
|
@ -389,6 +390,7 @@ public:
|
||||||
U128 FPVectorRSqrtStepFused(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
U128 FPVectorRSqrtStepFused(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
||||||
U128 FPVectorSqrt(size_t esize, const U128& a, bool fpcr_controlled = true);
|
U128 FPVectorSqrt(size_t esize, const U128& a, bool fpcr_controlled = true);
|
||||||
U128 FPVectorSub(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
U128 FPVectorSub(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
||||||
|
U128 FPVectorToHalf(size_t esize, const U128& a, FP::RoundingMode rounding, bool fpcr_controlled = true);
|
||||||
U128 FPVectorToSignedFixed(size_t esize, const U128& a, size_t fbits, FP::RoundingMode rounding, bool fpcr_controlled = true);
|
U128 FPVectorToSignedFixed(size_t esize, const U128& a, size_t fbits, FP::RoundingMode rounding, bool fpcr_controlled = true);
|
||||||
U128 FPVectorToUnsignedFixed(size_t esize, const U128& a, size_t fbits, FP::RoundingMode rounding, bool fpcr_controlled = true);
|
U128 FPVectorToUnsignedFixed(size_t esize, const U128& a, size_t fbits, FP::RoundingMode rounding, bool fpcr_controlled = true);
|
||||||
|
|
||||||
|
|
|
@ -613,6 +613,7 @@ OPCODE(FPVectorDiv64, U128, U128
|
||||||
OPCODE(FPVectorEqual16, U128, U128, U128, U1 )
|
OPCODE(FPVectorEqual16, U128, U128, U128, U1 )
|
||||||
OPCODE(FPVectorEqual32, U128, U128, U128, U1 )
|
OPCODE(FPVectorEqual32, U128, U128, U128, U1 )
|
||||||
OPCODE(FPVectorEqual64, U128, U128, U128, U1 )
|
OPCODE(FPVectorEqual64, U128, U128, U128, U1 )
|
||||||
|
OPCODE(FPVectorFromHalf32, U128, U128, U8, U1 )
|
||||||
OPCODE(FPVectorFromSignedFixed32, U128, U128, U8, U8, U1 )
|
OPCODE(FPVectorFromSignedFixed32, U128, U128, U8, U8, U1 )
|
||||||
OPCODE(FPVectorFromSignedFixed64, U128, U128, U8, U8, U1 )
|
OPCODE(FPVectorFromSignedFixed64, U128, U128, U8, U8, U1 )
|
||||||
OPCODE(FPVectorFromUnsignedFixed32, U128, U128, U8, U8, U1 )
|
OPCODE(FPVectorFromUnsignedFixed32, U128, U128, U8, U8, U1 )
|
||||||
|
@ -658,6 +659,7 @@ OPCODE(FPVectorSqrt32, U128, U128
|
||||||
OPCODE(FPVectorSqrt64, U128, U128, U1 )
|
OPCODE(FPVectorSqrt64, U128, U128, U1 )
|
||||||
OPCODE(FPVectorSub32, U128, U128, U128, U1 )
|
OPCODE(FPVectorSub32, U128, U128, U128, U1 )
|
||||||
OPCODE(FPVectorSub64, U128, U128, U128, U1 )
|
OPCODE(FPVectorSub64, U128, U128, U128, U1 )
|
||||||
|
OPCODE(FPVectorToHalf32, U128, U128, U8, U1 )
|
||||||
OPCODE(FPVectorToSignedFixed16, U128, U128, U8, U8, U1 )
|
OPCODE(FPVectorToSignedFixed16, U128, U128, U8, U8, U1 )
|
||||||
OPCODE(FPVectorToSignedFixed32, U128, U128, U8, U8, U1 )
|
OPCODE(FPVectorToSignedFixed32, U128, U128, U8, U8, U1 )
|
||||||
OPCODE(FPVectorToSignedFixed64, U128, U128, U8, U8, U1 )
|
OPCODE(FPVectorToSignedFixed64, U128, U128, U8, U8, U1 )
|
||||||
|
|
Loading…
Reference in a new issue