ir: Add opcodes for vector paired maximum and minimums

For the time being, we can just do a naive implementation which avoids
falling back to the interpreter a bit. Horizontal operations aren't
necessarily x86 SIMD's forte anyways.
This commit is contained in:
Lioncash 2018-07-29 16:08:31 -04:00 committed by MerryMage
parent 43344c5400
commit 463b9a3d02
4 changed files with 167 additions and 0 deletions

View file

@ -1866,6 +1866,101 @@ void EmitX64::EmitVectorPairedAddUnsignedWiden32(EmitContext& ctx, IR::Inst* ins
ctx.reg_alloc.DefineValue(inst, a); ctx.reg_alloc.DefineValue(inst, a);
} }
template <typename T, typename Function>
static void PairedOperation(VectorArray<T>& result, const VectorArray<T>& x, const VectorArray<T>& y, Function fn) {
const size_t range = x.size() / 2;
for (size_t i = 0; i < range; i++) {
result[i] = fn(x[2 * i], x[2 * i + 1]);
}
for (size_t i = 0; i < range; i++) {
result[range + i] = fn(y[2 * i], y[2 * i + 1]);
}
}
template <typename T>
static void PairedMax(VectorArray<T>& result, const VectorArray<T>& x, const VectorArray<T>& y) {
PairedOperation(result, x, y, [](auto a, auto b) { return std::max(a, b); });
}
template <typename T>
static void PairedMin(VectorArray<T>& result, const VectorArray<T>& x, const VectorArray<T>& y) {
PairedOperation(result, x, y, [](auto a, auto b) { return std::min(a, b); });
}
void EmitX64::EmitVectorPairedMaxS8(EmitContext& ctx, IR::Inst* inst) {
EmitTwoArgumentFallback(code, ctx, inst, [](VectorArray<s8>& result, const VectorArray<s8>& a, const VectorArray<s8>& b) {
PairedMax(result, a, b);
});
}
void EmitX64::EmitVectorPairedMaxS16(EmitContext& ctx, IR::Inst* inst) {
EmitTwoArgumentFallback(code, ctx, inst, [](VectorArray<s16>& result, const VectorArray<s16>& a, const VectorArray<s16>& b) {
PairedMax(result, a, b);
});
}
void EmitX64::EmitVectorPairedMaxS32(EmitContext& ctx, IR::Inst* inst) {
EmitTwoArgumentFallback(code, ctx, inst, [](VectorArray<s32>& result, const VectorArray<s32>& a, const VectorArray<s32>& b) {
PairedMax(result, a, b);
});
}
void EmitX64::EmitVectorPairedMaxU8(EmitContext& ctx, IR::Inst* inst) {
EmitTwoArgumentFallback(code, ctx, inst, [](VectorArray<u8>& result, const VectorArray<u8>& a, const VectorArray<u8>& b) {
PairedMax(result, a, b);
});
}
void EmitX64::EmitVectorPairedMaxU16(EmitContext& ctx, IR::Inst* inst) {
EmitTwoArgumentFallback(code, ctx, inst, [](VectorArray<u16>& result, const VectorArray<u16>& a, const VectorArray<u16>& b) {
PairedMax(result, a, b);
});
}
void EmitX64::EmitVectorPairedMaxU32(EmitContext& ctx, IR::Inst* inst) {
EmitTwoArgumentFallback(code, ctx, inst, [](VectorArray<u32>& result, const VectorArray<u32>& a, const VectorArray<u32>& b) {
PairedMax(result, a, b);
});
}
void EmitX64::EmitVectorPairedMinS8(EmitContext& ctx, IR::Inst* inst) {
EmitTwoArgumentFallback(code, ctx, inst, [](VectorArray<s8>& result, const VectorArray<s8>& a, const VectorArray<s8>& b) {
PairedMin(result, a, b);
});
}
void EmitX64::EmitVectorPairedMinS16(EmitContext& ctx, IR::Inst* inst) {
EmitTwoArgumentFallback(code, ctx, inst, [](VectorArray<s16>& result, const VectorArray<s16>& a, const VectorArray<s16>& b) {
PairedMin(result, a, b);
});
}
void EmitX64::EmitVectorPairedMinS32(EmitContext& ctx, IR::Inst* inst) {
EmitTwoArgumentFallback(code, ctx, inst, [](VectorArray<s32>& result, const VectorArray<s32>& a, const VectorArray<s32>& b) {
PairedMin(result, a, b);
});
}
void EmitX64::EmitVectorPairedMinU8(EmitContext& ctx, IR::Inst* inst) {
EmitTwoArgumentFallback(code, ctx, inst, [](VectorArray<u8>& result, const VectorArray<u8>& a, const VectorArray<u8>& b) {
PairedMin(result, a, b);
});
}
void EmitX64::EmitVectorPairedMinU16(EmitContext& ctx, IR::Inst* inst) {
EmitTwoArgumentFallback(code, ctx, inst, [](VectorArray<u16>& result, const VectorArray<u16>& a, const VectorArray<u16>& b) {
PairedMin(result, a, b);
});
}
void EmitX64::EmitVectorPairedMinU32(EmitContext& ctx, IR::Inst* inst) {
EmitTwoArgumentFallback(code, ctx, inst, [](VectorArray<u32>& result, const VectorArray<u32>& a, const VectorArray<u32>& b) {
PairedMin(result, a, b);
});
}
template <typename D, typename T> template <typename D, typename T>
static D PolynomialMultiply(T lhs, T rhs) { static D PolynomialMultiply(T lhs, T rhs) {
constexpr size_t bit_size = Common::BitSize<T>(); constexpr size_t bit_size = Common::BitSize<T>();

View file

@ -1226,6 +1226,62 @@ U128 IREmitter::VectorPairedAddUnsignedWiden(size_t original_esize, const U128&
return {}; return {};
} }
U128 IREmitter::VectorPairedMaxSigned(size_t esize, const U128& a, const U128& b) {
switch (esize) {
case 8:
return Inst<U128>(Opcode::VectorPairedMaxS8, a, b);
case 16:
return Inst<U128>(Opcode::VectorPairedMaxS16, a, b);
case 32:
return Inst<U128>(Opcode::VectorPairedMaxS32, a, b);
default:
UNREACHABLE();
return {};
}
}
U128 IREmitter::VectorPairedMaxUnsigned(size_t esize, const U128& a, const U128& b) {
switch (esize) {
case 8:
return Inst<U128>(Opcode::VectorPairedMaxU8, a, b);
case 16:
return Inst<U128>(Opcode::VectorPairedMaxU16, a, b);
case 32:
return Inst<U128>(Opcode::VectorPairedMaxU32, a, b);
default:
UNREACHABLE();
return {};
}
}
U128 IREmitter::VectorPairedMinSigned(size_t esize, const U128& a, const U128& b) {
switch (esize) {
case 8:
return Inst<U128>(Opcode::VectorPairedMinS8, a, b);
case 16:
return Inst<U128>(Opcode::VectorPairedMinS16, a, b);
case 32:
return Inst<U128>(Opcode::VectorPairedMinS32, a, b);
default:
UNREACHABLE();
return {};
}
}
U128 IREmitter::VectorPairedMinUnsigned(size_t esize, const U128& a, const U128& b) {
switch (esize) {
case 8:
return Inst<U128>(Opcode::VectorPairedMinU8, a, b);
case 16:
return Inst<U128>(Opcode::VectorPairedMinU16, a, b);
case 32:
return Inst<U128>(Opcode::VectorPairedMinU32, a, b);
default:
UNREACHABLE();
return {};
}
}
U128 IREmitter::VectorPolynomialMultiply(const U128& a, const U128& b) { U128 IREmitter::VectorPolynomialMultiply(const U128& a, const U128& b) {
return Inst<U128>(Opcode::VectorPolynomialMultiply8, a, b); return Inst<U128>(Opcode::VectorPolynomialMultiply8, a, b);
} }

View file

@ -242,6 +242,10 @@ public:
U128 VectorPairedAddLower(size_t esize, const U128& a, const U128& b); U128 VectorPairedAddLower(size_t esize, const U128& a, const U128& b);
U128 VectorPairedAddSignedWiden(size_t original_esize, const U128& a); U128 VectorPairedAddSignedWiden(size_t original_esize, const U128& a);
U128 VectorPairedAddUnsignedWiden(size_t original_esize, const U128& a); U128 VectorPairedAddUnsignedWiden(size_t original_esize, const U128& a);
U128 VectorPairedMaxSigned(size_t esize, const U128& a, const U128& b);
U128 VectorPairedMaxUnsigned(size_t esize, const U128& a, const U128& b);
U128 VectorPairedMinSigned(size_t esize, const U128& a, const U128& b);
U128 VectorPairedMinUnsigned(size_t esize, const U128& a, const U128& b);
U128 VectorPolynomialMultiply(const U128& a, const U128& b); U128 VectorPolynomialMultiply(const U128& a, const U128& b);
U128 VectorPolynomialMultiplyLong(size_t esize, const U128& a, const U128& b); U128 VectorPolynomialMultiplyLong(size_t esize, const U128& a, const U128& b);
U128 VectorPopulationCount(const U128& a); U128 VectorPopulationCount(const U128& a);

View file

@ -338,6 +338,18 @@ OPCODE(VectorPairedAdd8, T::U128, T::U128,
OPCODE(VectorPairedAdd16, T::U128, T::U128, T::U128 ) OPCODE(VectorPairedAdd16, T::U128, T::U128, T::U128 )
OPCODE(VectorPairedAdd32, T::U128, T::U128, T::U128 ) OPCODE(VectorPairedAdd32, T::U128, T::U128, T::U128 )
OPCODE(VectorPairedAdd64, T::U128, T::U128, T::U128 ) OPCODE(VectorPairedAdd64, T::U128, T::U128, T::U128 )
OPCODE(VectorPairedMaxS8, T::U128, T::U128, T::U128 )
OPCODE(VectorPairedMaxS16, T::U128, T::U128, T::U128 )
OPCODE(VectorPairedMaxS32, T::U128, T::U128, T::U128 )
OPCODE(VectorPairedMaxU8, T::U128, T::U128, T::U128 )
OPCODE(VectorPairedMaxU16, T::U128, T::U128, T::U128 )
OPCODE(VectorPairedMaxU32, T::U128, T::U128, T::U128 )
OPCODE(VectorPairedMinS8, T::U128, T::U128, T::U128 )
OPCODE(VectorPairedMinS16, T::U128, T::U128, T::U128 )
OPCODE(VectorPairedMinS32, T::U128, T::U128, T::U128 )
OPCODE(VectorPairedMinU8, T::U128, T::U128, T::U128 )
OPCODE(VectorPairedMinU16, T::U128, T::U128, T::U128 )
OPCODE(VectorPairedMinU32, T::U128, T::U128, T::U128 )
OPCODE(VectorPolynomialMultiply8, T::U128, T::U128, T::U128 ) OPCODE(VectorPolynomialMultiply8, T::U128, T::U128, T::U128 )
OPCODE(VectorPolynomialMultiplyLong8, T::U128, T::U128, T::U128 ) OPCODE(VectorPolynomialMultiplyLong8, T::U128, T::U128, T::U128 )
OPCODE(VectorPolynomialMultiplyLong64, T::U128, T::U128, T::U128 ) OPCODE(VectorPolynomialMultiplyLong64, T::U128, T::U128, T::U128 )