如果我理解documentation correctly,我们应该能够使用ldexp
来恢复分解为签名尾数的浮点数和frexp
指数。我一直无法做到这一点。请考虑以下代码:
#include <cmath>
#include <iostream>
#include <limits>
template <typename T>
void float_info() {
std::cout << "max=" << std::numeric_limits<T>::max() <<
", max_exp=" << std::numeric_limits<T>::max_exponent <<
", max_10_exp=" << std::numeric_limits<T>::max_exponent10 <<
", min=" << std::numeric_limits<T>::min() <<
", min_exp=" << std::numeric_limits<T>::min_exponent <<
", min_10_exp=" << std::numeric_limits<T>::min_exponent10 <<
", dig=" << std::numeric_limits<T>::digits10 <<
", mant_dig=" << std::numeric_limits<T>::digits <<
", epsilon=" << std::numeric_limits<T>::epsilon() <<
", radix=" << std::numeric_limits<T>::radix <<
", rounds=" << std::numeric_limits<T>::round_style << std::endl;
}
template <typename T>
void compare(T a, T b) {
std::cout << a << " " << b << " (" <<
(a != b ? "un" : "") << "equal)" << std::endl;
}
template<typename T>
void test_ldexp() {
float_info<T>();
T x = 1 + std::numeric_limits<T>::epsilon();
T y = ldexp(x, 0);
int exponent;
T mantissa = frexp(x, &exponent);
T z = ldexp(mantissa, exponent);
compare(x, y);
compare(x, z);
std::cout << std::endl;
}
int main() {
std::cout.precision(25);
test_ldexp<float>();
test_ldexp<double>();
test_ldexp<long double>();
}
在Ubuntu 14.04.3 LTS上使用g++
(版本4.8.4)编译时,输出为:
max=3.402823466385288598117042e+38, max_exp=128, max_10_exp=38,
min=1.175494350822287507968737e-38, min_exp=-125, min_10_exp=-37, dig=6,
mant_dig=24, epsilon=1.1920928955078125e-07, radix=2, rounds=1
1.00000011920928955078125 1.00000011920928955078125 (equal)
1.00000011920928955078125 1.00000011920928955078125 (equal)
max=1.797693134862315708145274e+308, max_exp=1024, max_10_exp=308,
min=2.225073858507201383090233e-308, min_exp=-1021, min_10_exp=-307, dig=15,
mant_dig=53, epsilon=2.220446049250313080847263e-16, radix=2, rounds=1
1.000000000000000222044605 1.000000000000000222044605 (equal)
1.000000000000000222044605 1.000000000000000222044605 (equal)
max=1.189731495357231765021264e+4932, max_exp=16384, max_10_exp=4932,
min=3.362103143112093506262678e-4932, min_exp=-16381, min_10_exp=-4931, dig=18,
mant_dig=64, epsilon=1.084202172485504434007453e-19, radix=2, rounds=1
1.00000000000000000010842 1 (unequal)
1.00000000000000000010842 1 (unequal)
使用long double
时,我们似乎通过使用x
分解frexpr
来丢失某些内容。如果我使用python3
(版本3.4.3)运行以下脚本,我可以实现我期望的行为。
import math
import sys
def compare(a, b):
print('{a} {b} ({pre}equal)'.format(a=a, b=b,
pre='un' if a != b else ''))
x = 1 + sys.float_info.epsilon
mantissa, exponent = math.frexp(x)
print(sys.float_info)
compare(x, math.ldexp(x, 0))
compare(x, math.ldexp(mantissa, exponent))
输出结果为:
sys.float_info(max=1.7976931348623157e+308, max_exp=1024, max_10_exp=308,
min=2.2250738585072014e-308, min_exp=-1021, min_10_exp=-307, dig=15,
mant_dig=53, epsilon=2.220446049250313e-16, radix=2, rounds=1)
1.0000000000000002 1.0000000000000002 (equal)
1.0000000000000002 1.0000000000000002 (equal)
请注意,这仅使用double
s。
我尝试阅读cmath
标头文件,以了解frexpr
和ldexpr
的实施方式,但我无法理解它。发生了什么事?
答案 0 :(得分:1)
修改您的C ++程序,包括typeinfo
标题,并在compare
函数调用test_ldexp
之前插入以下行。
std::cout << "types:" << std::endl;
std::cout << " x : " << typeid(x).name() << std::endl;
std::cout << " mantissa : " << typeid(frexp(x, &exponent)).name()
<< std::endl;
std::cout << " ldexp(...): " << typeid(ldexp(x, 0)).name() << std::endl;
std::cout << " ldexp(...): " << typeid(ldexp(mantissa, exponent)).name()
<< std::endl;
输出现在是:
max=3.402823466385288598117042e+38, max_exp=128, max_10_exp=38,
min=1.175494350822287507968737e-38, min_exp=-125, min_10_exp=-37, dig=6,
mant_dig=24, epsilon=1.1920928955078125e-07, radix=2, rounds=1
types:
x : f
mantissa : d
ldexp(...): d
ldexp(...): d
1.00000011920928955078125 1.00000011920928955078125 (equal)
1.00000011920928955078125 1.00000011920928955078125 (equal)
max=1.797693134862315708145274e+308, max_exp=1024, max_10_exp=308,
min=2.225073858507201383090233e-308, min_exp=-1021, min_10_exp=-307, dig=15,
mant_dig=53, epsilon=2.220446049250313080847263e-16, radix=2, rounds=1
types:
x : d
mantissa : d
ldexp(...): d
ldexp(...): d
1.000000000000000222044605 1.000000000000000222044605 (equal)
1.000000000000000222044605 1.000000000000000222044605 (equal)
max=1.189731495357231765021264e+4932, max_exp=16384, max_10_exp=4932,
min=3.362103143112093506262678e-4932, min_exp=-16381, min_10_exp=-4931, dig=18,
mant_dig=64, epsilon=1.084202172485504434007453e-19, radix=2, rounds=1
types:
x : e
mantissa : d
ldexp(...): d
ldexp(...): d
1.00000000000000000010842 1 (unequal)
1.00000000000000000010842 1 (unequal)
无论您输入什么类型, frexpr
和ldexpr
都会返回double
!您似乎正在使用math.h
中定义的功能(请参阅here和here),而不是cmath
中定义的功能。通过拨打frexpr
和ldexpr
来接听您对std::frexpr
和std::ldexpr
的来电,您的代码将按照您的预期运作。