ClickHouse/dbms/include/DB/Functions/FunctionsMath.h

497 lines
18 KiB
C++

#pragma once
#include <DB/DataTypes/DataTypesNumberFixed.h>
#include <DB/Functions/IFunction.h>
#include <vectorf128.h>
#include <vectormath_exp.h>
#include <vectormath_trig.h>
namespace DB
{
template <typename Impl>
class FunctionMathNullaryConstFloat64 : public IFunction
{
public:
static constexpr auto name = Impl::name;
static IFunction * create(const Context &) { return new FunctionMathNullaryConstFloat64; }
private:
String getName() const override { return name; }
DataTypePtr getReturnType(const DataTypes & arguments) const override
{
if (arguments.size() != 0)
throw Exception{
"Number of arguments for function " + getName() + "doesn't match: passed "
+ toString(arguments.size()) + ", should be 0",
ErrorCodes::NUMBER_OF_ARGUMENTS_DOESNT_MATCH
};
return new DataTypeFloat64;
}
void execute(Block & block, const ColumnNumbers & arguments, const size_t result) override
{
block.getByPosition(result).column = new ColumnConst<Float64>{
block.rowsInFirstColumn(),
Impl::value
};
}
};
template <typename Impl> class FunctionMathUnaryFloat64 : public IFunction
{
public:
static constexpr auto name = Impl::name;
static IFunction * create(const Context &) { return new FunctionMathUnaryFloat64; }
static_assert(Impl::rows_per_iteration > 0, "Impl must process at least one row per iteration");
private:
String getName() const override { return name; }
DataTypePtr getReturnType(const DataTypes & arguments) const override
{
if (arguments.size() != 1)
throw Exception{
"Number of arguments for function " + getName() + "doesn't match: passed "
+ toString(arguments.size()) + ", should be 1",
ErrorCodes::NUMBER_OF_ARGUMENTS_DOESNT_MATCH
};
const auto check_argument_type = [this] (const IDataType * const arg) {
if (!typeid_cast<const DataTypeUInt8 *>(arg) &&
!typeid_cast<const DataTypeUInt16 *>(arg) &&
!typeid_cast<const DataTypeUInt32 *>(arg) &&
!typeid_cast<const DataTypeUInt64 *>(arg) &&
!typeid_cast<const DataTypeInt8 *>(arg) &&
!typeid_cast<const DataTypeInt16 *>(arg) &&
!typeid_cast<const DataTypeInt32 *>(arg) &&
!typeid_cast<const DataTypeInt64 *>(arg) &&
!typeid_cast<const DataTypeFloat32 *>(arg) &&
!typeid_cast<const DataTypeFloat64 *>(arg))
{
throw Exception{
"Illegal type " + arg->getName() + " of argument of function " + getName(),
ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT
};
}
};
check_argument_type(arguments.front().get());
return new DataTypeFloat64;
}
template <typename FieldType>
bool execute(Block & block, const IColumn * const arg, const size_t result)
{
if (const auto col = typeid_cast<const ColumnVector<FieldType> *>(arg))
{
const auto dst = new ColumnVector<Float64>;
block.getByPosition(result).column = dst;
const auto & src_data = col->getData();
const auto src_size = src_data.size();
auto & dst_data = dst->getData();
dst_data.resize(src_size);
const auto rows_remaining = src_size % Impl::rows_per_iteration;
const auto rows_size = src_size - rows_remaining;
for (size_t i = 0; i < rows_size; i += Impl::rows_per_iteration)
Impl::execute(&src_data[i], &dst_data[i]);
if (rows_remaining != 0)
{
FieldType src_remaining[Impl::rows_per_iteration];
memcpy(src_remaining, &src_data[rows_size], rows_remaining * sizeof(FieldType));
memset(src_remaining + rows_remaining, 0, (Impl::rows_per_iteration - rows_remaining) * sizeof(FieldType));
Float64 dst_remaining[Impl::rows_per_iteration];
Impl::execute(src_remaining, dst_remaining);
memcpy(&dst_data[rows_size], dst_remaining, rows_remaining * sizeof(Float64));
}
return true;
}
else if (const auto col = typeid_cast<const ColumnConst<FieldType> *>(arg))
{
const FieldType src[Impl::rows_per_iteration] { col->getData() };
Float64 dst[Impl::rows_per_iteration];
Impl::execute(src, dst);
block.getByPosition(result).column = new ColumnConst<Float64>{col->size(), dst[0]};
return true;
}
return false;
}
void execute(Block & block, const ColumnNumbers & arguments, const size_t result) override
{
const auto arg = block.getByPosition(arguments[0]).column.get();
if (!execute<UInt8>(block, arg, result) &&
!execute<UInt16>(block, arg, result) &&
!execute<UInt32>(block, arg, result) &&
!execute<UInt64>(block, arg, result) &&
!execute<Int8>(block, arg, result) &&
!execute<Int16>(block, arg, result) &&
!execute<Int32>(block, arg, result) &&
!execute<Int64>(block, arg, result) &&
!execute<Float32>(block, arg, result) &&
!execute<Float64>(block, arg, result))
{
throw Exception{
"Illegal column " + arg->getName() + " of argument of function " + getName(),
ErrorCodes::ILLEGAL_COLUMN
};
}
}
};
template <typename Name, Vec2d(&Function)(const Vec2d &)>
struct UnaryFunctionVectorized
{
static constexpr auto name = Name::name;
static constexpr auto rows_per_iteration = 2;
template <typename T>
static void execute(const T * const src, Float64 * const dst)
{
const auto & result = Function(Vec2d(src[0], src[1]));
dst[0] = result[0];
dst[1] = result[1];
}
};
template <typename Name, Float64(&Function)(Float64)>
struct UnaryFunctionPlain
{
static constexpr auto name = Name::name;
static constexpr auto rows_per_iteration = 1;
template <typename T>
static void execute(const T * const src, Float64 * const dst)
{
dst[0] = static_cast<Float64>(Function(static_cast<Float64>(src[0])));
}
};
template <typename Impl> class FunctionMathBinaryFloat64 : public IFunction
{
public:
static constexpr auto name = Impl::name;
static IFunction * create(const Context &) { return new FunctionMathBinaryFloat64; }
static_assert(Impl::rows_per_iteration > 0, "Impl must process at least one row per iteration");
private:
String getName() const override { return name; }
DataTypePtr getReturnType(const DataTypes & arguments) const override
{
if (arguments.size() != 2)
throw Exception{
"Number of arguments for function " + getName() + "doesn't match: passed "
+ toString(arguments.size()) + ", should be 2",
ErrorCodes::NUMBER_OF_ARGUMENTS_DOESNT_MATCH
};
const auto check_argument_type = [this] (const IDataType * const arg) {
if (!typeid_cast<const DataTypeUInt8 *>(arg) &&
!typeid_cast<const DataTypeUInt16 *>(arg) &&
!typeid_cast<const DataTypeUInt32 *>(arg) &&
!typeid_cast<const DataTypeUInt64 *>(arg) &&
!typeid_cast<const DataTypeInt8 *>(arg) &&
!typeid_cast<const DataTypeInt16 *>(arg) &&
!typeid_cast<const DataTypeInt32 *>(arg) &&
!typeid_cast<const DataTypeInt64 *>(arg) &&
!typeid_cast<const DataTypeFloat32 *>(arg) &&
!typeid_cast<const DataTypeFloat64 *>(arg))
{
throw Exception{
"Illegal type " + arg->getName() + " of argument of function " + getName(),
ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT
};
}
};
check_argument_type(arguments.front().get());
check_argument_type(arguments.back().get());
return new DataTypeFloat64;
}
template <typename LeftType, typename RightType>
bool executeRight(Block & block, const size_t result, const ColumnConst<LeftType> * const left_arg,
const IColumn * const right_arg)
{
if (const auto right_arg_typed = typeid_cast<const ColumnVector<RightType> *>(right_arg))
{
const auto dst = new ColumnVector<Float64>;
block.getByPosition(result).column = dst;
LeftType left_src_data[Impl::rows_per_iteration];
std::fill(std::begin(left_src_data), std::end(left_src_data), left_arg->getData());
const auto & right_src_data = right_arg_typed->getData();
const auto src_size = right_src_data.size();
auto & dst_data = dst->getData();
dst_data.resize(src_size);
const auto rows_remaining = src_size % Impl::rows_per_iteration;
const auto rows_size = src_size - rows_remaining;
for (size_t i = 0; i < rows_size; i += Impl::rows_per_iteration)
Impl::execute(left_src_data, &right_src_data[i], &dst_data[i]);
if (rows_remaining != 0)
{
RightType right_src_remaining[Impl::rows_per_iteration];
memcpy(right_src_remaining, &right_src_data[rows_size], rows_remaining * sizeof(RightType));
memset(right_src_remaining + rows_remaining, 0, (Impl::rows_per_iteration - rows_remaining) * sizeof(RightType));
Float64 dst_remaining[Impl::rows_per_iteration];
Impl::execute(left_src_data, right_src_remaining, dst_remaining);
memcpy(&dst_data[rows_size], dst_remaining, rows_remaining * sizeof(Float64));
}
return true;
}
else if (const auto right_arg_typed = typeid_cast<const ColumnConst<RightType> *>(right_arg))
{
const LeftType left_src[Impl::rows_per_iteration] { left_arg->getData() };
const RightType right_src[Impl::rows_per_iteration] { right_arg_typed->getData() };
Float64 dst[Impl::rows_per_iteration];
Impl::execute(left_src, right_src, dst);
block.getByPosition(result).column = new ColumnConst<Float64>{left_arg->size(), dst[0]};
return true;
}
return false;
}
template <typename LeftType, typename RightType>
bool executeRight(Block & block, const size_t result, const ColumnVector<LeftType> * const left_arg,
const IColumn * const right_arg)
{
if (const auto right_arg_typed = typeid_cast<const ColumnVector<RightType> *>(right_arg))
{
const auto dst = new ColumnVector<Float64>;
block.getByPosition(result).column = dst;
const auto & left_src_data = left_arg->getData();
const auto & right_src_data = right_arg_typed->getData();
const auto src_size = left_src_data.size();
auto & dst_data = dst->getData();
dst_data.resize(src_size);
const auto rows_remaining = src_size % Impl::rows_per_iteration;
const auto rows_size = src_size - rows_remaining;
for (size_t i = 0; i < rows_size; i += Impl::rows_per_iteration)
Impl::execute(&left_src_data[i], &right_src_data[i], &dst_data[i]);
if (rows_remaining != 0)
{
LeftType left_src_remaining[Impl::rows_per_iteration];
memcpy(left_src_remaining, &left_src_data[rows_size], rows_remaining * sizeof(LeftType));
memset(left_src_remaining + rows_remaining, 0, (Impl::rows_per_iteration - rows_remaining) * sizeof(LeftType));
RightType right_src_remaining[Impl::rows_per_iteration];
memcpy(right_src_remaining, &right_src_data[rows_size], rows_remaining * sizeof(RightType));
memset(right_src_remaining + rows_remaining, 0, (Impl::rows_per_iteration - rows_remaining) * sizeof(RightType));
Float64 dst_remaining[Impl::rows_per_iteration];
Impl::execute(left_src_remaining, right_src_remaining, dst_remaining);
memcpy(&dst_data[rows_size], dst_remaining, rows_remaining * sizeof(Float64));
}
return true;
}
else if (const auto right_arg_typed = typeid_cast<const ColumnConst<RightType> *>(right_arg))
{
const auto dst = new ColumnVector<Float64>;
block.getByPosition(result).column = dst;
const auto & left_src_data = left_arg->getData();
RightType right_src_data[Impl::rows_per_iteration];
std::fill(std::begin(right_src_data), std::end(right_src_data), right_arg_typed->getData());
const auto src_size = left_src_data.size();
auto & dst_data = dst->getData();
dst_data.resize(src_size);
const auto rows_remaining = src_size % Impl::rows_per_iteration;
const auto rows_size = src_size - rows_remaining;
for (size_t i = 0; i < rows_size; i += Impl::rows_per_iteration)
Impl::execute(&left_src_data[i], right_src_data, &dst_data[i]);
if (rows_remaining != 0)
{
LeftType left_src_remaining[Impl::rows_per_iteration];
memcpy(left_src_remaining, &left_src_data[rows_size], rows_remaining * sizeof(LeftType));
memset(left_src_remaining + rows_remaining, 0, (Impl::rows_per_iteration - rows_remaining) * sizeof(LeftType));
Float64 dst_remaining[Impl::rows_per_iteration];
Impl::execute(left_src_remaining, right_src_data, dst_remaining);
memcpy(&dst_data[rows_size], dst_remaining, rows_remaining * sizeof(Float64));
}
return true;
}
return false;
}
template <typename LeftType, template <typename> class LeftColumnType>
bool executeLeftImpl(Block & block, const ColumnNumbers & arguments, const size_t result,
const IColumn * const left_arg)
{
if (const auto left_arg_typed = typeid_cast<const LeftColumnType<LeftType> *>(left_arg))
{
const auto right_arg = block.getByPosition(arguments[1]).column.get();
if (executeRight<LeftType, UInt8>(block, result, left_arg_typed, right_arg) ||
executeRight<LeftType, UInt16>(block, result, left_arg_typed, right_arg) ||
executeRight<LeftType, UInt32>(block, result, left_arg_typed, right_arg) ||
executeRight<LeftType, UInt64>(block, result, left_arg_typed, right_arg) ||
executeRight<LeftType, Int8>(block, result, left_arg_typed, right_arg) ||
executeRight<LeftType, Int16>(block, result, left_arg_typed, right_arg) ||
executeRight<LeftType, Int32>(block, result, left_arg_typed, right_arg) ||
executeRight<LeftType, Int64>(block, result, left_arg_typed, right_arg) ||
executeRight<LeftType, Float32>(block, result, left_arg_typed, right_arg) ||
executeRight<LeftType, Float64>(block, result, left_arg_typed, right_arg))
{
return true;
}
else
{
throw Exception{
"Illegal column " + block.getByPosition(arguments[1]).column->getName() +
" of second argument of function " + getName(),
ErrorCodes::ILLEGAL_COLUMN
};
}
}
return false;
}
template <typename LeftType>
bool executeLeft(Block & block, const ColumnNumbers & arguments, const size_t result,
const IColumn * const left_arg)
{
if (executeLeftImpl<LeftType, ColumnVector>(block, arguments, result, left_arg) ||
executeLeftImpl<LeftType, ColumnConst>(block, arguments, result, left_arg))
return true;
return false;
}
void execute(Block & block, const ColumnNumbers & arguments, const size_t result) override
{
const auto left_arg = block.getByPosition(arguments[0]).column.get();
if (!executeLeft<UInt8>(block, arguments, result, left_arg) &&
!executeLeft<UInt16>(block, arguments, result, left_arg) &&
!executeLeft<UInt32>(block, arguments, result, left_arg) &&
!executeLeft<UInt64>(block, arguments, result, left_arg) &&
!executeLeft<Int8>(block, arguments, result, left_arg) &&
!executeLeft<Int16>(block, arguments, result, left_arg) &&
!executeLeft<Int32>(block, arguments, result, left_arg) &&
!executeLeft<Int64>(block, arguments, result, left_arg) &&
!executeLeft<Float32>(block, arguments, result, left_arg) &&
!executeLeft<Float64>(block, arguments, result, left_arg))
{
throw Exception{
"Illegal column " + left_arg->getName() + " of argument of function " + getName(),
ErrorCodes::ILLEGAL_COLUMN
};
}
}
};
template <typename Name, Vec2d(&Function)(const Vec2d &, const Vec2d &)>
struct BinaryFunctionVectorized
{
static constexpr auto name = Name::name;
static constexpr auto rows_per_iteration = 2;
template <typename T1, typename T2>
static void execute(const T1 * const src_left, const T2 * const src_right, Float64 * const dst)
{
const auto & result = Function(Vec2d(src_left[0], src_left[1]), Vec2d(src_right[0], src_right[1]));
dst[0] = result[0];
dst[1] = result[1];
}
};
struct EImpl
{
static constexpr auto name = "e";
static constexpr auto value = 2.7182818284590452353602874713526624977572470;
};
struct PiImpl
{
static constexpr auto name = "pi";
static constexpr auto value = 3.1415926535897932384626433832795028841971693;
};
struct ExpName { static constexpr auto name = "exp"; };
struct LogName { static constexpr auto name = "log"; };
struct Exp2Name { static constexpr auto name = "exp2"; };
struct Log2Name { static constexpr auto name = "log2"; };
struct Exp10Name { static constexpr auto name = "exp10"; };
struct Log10Name { static constexpr auto name = "log10"; };
struct SqrtName { static constexpr auto name = "sqrt"; };
struct CbrtName { static constexpr auto name = "cbrt"; };
struct SinName { static constexpr auto name = "sin"; };
struct CosName { static constexpr auto name = "cos"; };
struct TanName { static constexpr auto name = "tan"; };
struct AsinName { static constexpr auto name = "asin"; };
struct AcosName { static constexpr auto name = "acos"; };
struct AtanName { static constexpr auto name = "atan"; };
struct ErfName { static constexpr auto name = "erf"; };
struct ErfcName { static constexpr auto name = "erfc"; };
struct LGammaName { static constexpr auto name = "lgamma"; };
struct TGammaName { static constexpr auto name = "tgamma"; };
struct PowName { static constexpr auto name = "pow"; };
using FunctionE = FunctionMathNullaryConstFloat64<EImpl>;
using FunctionPi = FunctionMathNullaryConstFloat64<PiImpl>;
using FunctionExp = FunctionMathUnaryFloat64<UnaryFunctionVectorized<ExpName, exp>>;
using FunctionLog = FunctionMathUnaryFloat64<UnaryFunctionVectorized<LogName, log>>;
using FunctionExp2 = FunctionMathUnaryFloat64<UnaryFunctionVectorized<Exp2Name, exp2>>;
using FunctionLog2 = FunctionMathUnaryFloat64<UnaryFunctionVectorized<Log2Name, log2>>;
using FunctionExp10 = FunctionMathUnaryFloat64<UnaryFunctionVectorized<Exp10Name, exp10>>;
using FunctionLog10 = FunctionMathUnaryFloat64<UnaryFunctionVectorized<Log10Name, log10>>;
using FunctionSqrt = FunctionMathUnaryFloat64<UnaryFunctionVectorized<SqrtName, sqrt>>;
using FunctionCbrt = FunctionMathUnaryFloat64<UnaryFunctionVectorized<CbrtName, Power_rational<1, 3>::pow>>;
using FunctionSin = FunctionMathUnaryFloat64<UnaryFunctionVectorized<SinName, sin>>;
using FunctionCos = FunctionMathUnaryFloat64<UnaryFunctionVectorized<CosName, cos>>;
using FunctionTan = FunctionMathUnaryFloat64<UnaryFunctionVectorized<TanName, tan>>;
using FunctionAsin = FunctionMathUnaryFloat64<UnaryFunctionVectorized<AsinName, asin>>;
using FunctionAcos = FunctionMathUnaryFloat64<UnaryFunctionVectorized<AcosName, acos>>;
using FunctionAtan = FunctionMathUnaryFloat64<UnaryFunctionVectorized<AtanName, atan>>;
using FunctionErf = FunctionMathUnaryFloat64<UnaryFunctionPlain<ErfName, std::erf>>;
using FunctionErfc = FunctionMathUnaryFloat64<UnaryFunctionPlain<ErfcName, std::erfc>>;
using FunctionLGamma = FunctionMathUnaryFloat64<UnaryFunctionPlain<LGammaName, std::lgamma>>;
using FunctionTGamma = FunctionMathUnaryFloat64<UnaryFunctionPlain<TGammaName, std::tgamma>>;
using FunctionPow = FunctionMathBinaryFloat64<BinaryFunctionVectorized<PowName, pow>>;
}