什么是浮点和双重比较最有效的方法？
什么是比较两个double float值或两个float值的最有效方法？
简单地做这件事是不正确的：
bool CompareDoubles1 (double A, double B)
{
return A == B;
}
但是像这样：
bool CompareDoubles2 (double A, double B)
{
diff = A - B;
return (diff < EPSILON) && (-diff < EPSILON);
}
似乎浪费处理。
有谁知道更聪明的浮点比较器？
使用任何其他建议时要格外小心。 这一切都取决于上下文。
我花了很长时间来追踪系统中的一个错误，假设a==b如果|ab|<epsilon 。 根本问题是：
如果a==b和b==c则a==c ，算法中隐含的推定。
对于以英寸为单位的线和以密耳（.001英寸）为单位的线使用相同的epsilon。 那是a==b但是1000a!=1000b 。 （这就是为什么AlmostEqual2sComplement要求epsilon或max ULPS）。
对于角度的余弦和线的长度使用相同的epsilon！
使用这种比较函数对集合中的项目进行排序。 （在这种情况下，使用内置的C ++运算符==双打产生正确的结果。）
就像我说的：这一切都取决于上下文和a和b的预期大小。
顺便说一句， std::numeric_limits<double>::epsilon()是“机器epsilon”。 它是1.0和下一个可用double表示的值之间的差值。 我想它可以在比较函数中使用，但只有当期望值小于1时（这是对@ cdv的答案的回应......）
另外，如果你基本上int算术doubles （这里我们用双打来保存在某些情况下INT值）的算术将是正确的。 例如4.0 / 2.0将与1.0 + 1.0相同。 只要你不做导致分数的事情（4.0 / 3.0），或者不要超出整数的大小。
与epsilon值的比较是大多数人所做的（甚至在游戏编程中）。
你应该稍微改变你的实现：
bool AreSame(double a, double b)
{
return fabs(a - b) < EPSILON;
}
编辑：Christer在最近的博客文章中为这个主题添加了一堆很棒的信息。 请享用。
我发现Google C ++测试框架包含一个很好的跨平台基于模板的AlmostEqual2sComplement实现，它可以用于双打和浮动。 鉴于它是根据BSD许可证发布的，只要您保留许可证，在自己的代码中使用它应该没有问题。 我从http://code.google.com/p/googletest/source/browse/trunk/include/gtest/internal/gtest-internal.h中提取了以下代码，并在顶部添加了许可证。
请确保＃定义GTEST_OS_WINDOWS为一些值（或者将其用于适合您的代码库的代码 - 毕竟它是BSD许可的）。
用法示例：
double left = // something
double right = // something
const FloatingPoint<double> lhs(left), rhs(right);
if (lhs.AlmostEquals(rhs)) {
//they're equal!
}
代码如下：
// Copyright 2005, Google Inc.
// All rights reserved.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
//
// * Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
// * Redistributions in binary form must reproduce the above
// copyright notice, this list of conditions and the following disclaimer
// in the documentation and/or other materials provided with the
// distribution.
// * Neither the name of Google Inc. nor the names of its
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
//
// Authors: wan@google.com (Zhanyong Wan), eefacm@gmail.com (Sean Mcafee)
//
// The Google C++ Testing Framework (Google Test)
// This template class serves as a compile-time function from size to
// type. It maps a size in bytes to a primitive type with that
// size. e.g.
//
// TypeWithSize<4>::UInt
//
// is typedef-ed to be unsigned int (unsigned integer made up of 4
// bytes).
//
// Such functionality should belong to STL, but I cannot find it
// there.
//
// Google Test uses this class in the implementation of floating-point
// comparison.
//
// For now it only handles UInt (unsigned int) as that's all Google Test
// needs. Other types can be easily added in the future if need
// arises.
template <size_t size>
class TypeWithSize {
public:
// This prevents the user from using TypeWithSize<N> with incorrect
// values of N.
typedef void UInt;
};
// The specialization for size 4.
template <>
class TypeWithSize<4> {
public:
// unsigned int has size 4 in both gcc and MSVC.
//
// As base/basictypes.h doesn't compile on Windows, we cannot use
// uint32, uint64, and etc here.
typedef int Int;
typedef unsigned int UInt;
};
// The specialization for size 8.
template <>
class TypeWithSize<8> {
public:
#if GTEST_OS_WINDOWS
typedef __int64 Int;
typedef unsigned __int64 UInt;
#else
typedef long long Int; // NOLINT
typedef unsigned long long UInt; // NOLINT
#endif // GTEST_OS_WINDOWS
};
// This template class represents an IEEE floating-point number
// (either single-precision or double-precision, depending on the
// template parameters).
//
// The purpose of this class is to do more sophisticated number
// comparison. (Due to round-off error, etc, it's very unlikely that
// two floating-points will be equal exactly. Hence a naive
// comparison by the == operation often doesn't work.)
//
// Format of IEEE floating-point:
//
// The most-significant bit being the leftmost, an IEEE
// floating-point looks like
//
// sign_bit exponent_bits fraction_bits
//
// Here, sign_bit is a single bit that designates the sign of the
// number.
//
// For float, there are 8 exponent bits and 23 fraction bits.
//
// For double, there are 11 exponent bits and 52 fraction bits.
//
// More details can be found at
// http://en.wikipedia.org/wiki/IEEE_floating-point_standard.
//
// Template parameter:
//
// RawType: the raw floating-point type (either float or double)
template <typename RawType>
class FloatingPoint {
public:
// Defines the unsigned integer type that has the same size as the
// floating point number.
typedef typename TypeWithSize<sizeof(RawType)>::UInt Bits;
// Constants.
// # of bits in a number.
static const size_t kBitCount = 8*sizeof(RawType);
// # of fraction bits in a number.
static const size_t kFractionBitCount =
std::numeric_limits<RawType>::digits - 1;
// # of exponent bits in a number.
static const size_t kExponentBitCount = kBitCount - 1 - kFractionBitCount;
// The mask for the sign bit.
static const Bits kSignBitMask = static_cast<Bits>(1) << (kBitCount - 1);
// The mask for the fraction bits.
static const Bits kFractionBitMask =
~static_cast<Bits>(0) >> (kExponentBitCount + 1);
// The mask for the exponent bits.
static const Bits kExponentBitMask = ~(kSignBitMask | kFractionBitMask);
// How many ULP's (Units in the Last Place) we want to tolerate when
// comparing two numbers. The larger the value, the more error we
// allow. A 0 value means that two numbers must be exactly the same
// to be considered equal.
//
// The maximum error of a single floating-point operation is 0.5
// units in the last place. On Intel CPU's, all floating-point
// calculations are done with 80-bit precision, while double has 64
// bits. Therefore, 4 should be enough for ordinary use.
//
// See the following article for more details on ULP:
// http://www.cygnus-software.com/papers/comparingfloats/comparingfloats.htm.
static const size_t kMaxUlps = 4;
// Constructs a FloatingPoint from a raw floating-point number.
//
// On an Intel CPU, passing a non-normalized NAN (Not a Number)
// around may change its bits, although the new value is guaranteed
// to be also a NAN. Therefore, don't expect this constructor to
// preserve the bits in x when x is a NAN.
explicit FloatingPoint(const RawType& x) { u_.value_ = x; }
// Static methods
// Reinterprets a bit pattern as a floating-point number.
//
// This function is needed to test the AlmostEquals() method.
static RawType ReinterpretBits(const Bits bits) {
FloatingPoint fp(0);
fp.u_.bits_ = bits;
return fp.u_.value_;
}
// Returns the floating-point number that represent positive infinity.
static RawType Infinity() {
return ReinterpretBits(kExponentBitMask);
}
// Non-static methods
// Returns the bits that represents this number.
const Bits &bits() const { return u_.bits_; }
// Returns the exponent bits of this number.
Bits exponent_bits() const { return kExponentBitMask & u_.bits_; }
// Returns the fraction bits of this number.
Bits fraction_bits() const { return kFractionBitMask & u_.bits_; }
// Returns the sign bit of this number.
Bits sign_bit() const { return kSignBitMask & u_.bits_; }
// Returns true iff this is NAN (not a number).
bool is_nan() const {
// It's a NAN if the exponent bits are all ones and the fraction
// bits are not entirely zeros.
return (exponent_bits() == kExponentBitMask) && (fraction_bits() != 0);
}
// Returns true iff this number is at most kMaxUlps ULP's away from
// rhs. In particular, this function:
//
// - returns false if either number is (or both are) NAN.
// - treats really large numbers as almost equal to infinity.
// - thinks +0.0 and -0.0 are 0 DLP's apart.
bool AlmostEquals(const FloatingPoint& rhs) const {
// The IEEE standard says that any comparison operation involving
// a NAN must return false.
if (is_nan() || rhs.is_nan()) return false;
return DistanceBetweenSignAndMagnitudeNumbers(u_.bits_, rhs.u_.bits_)
<= kMaxUlps;
}
private:
// The data type used to store the actual floating-point number.
union FloatingPointUnion {
RawType value_; // The raw floating-point number.
Bits bits_; // The bits that represent the number.
};
// Converts an integer from the sign-and-magnitude representation to
// the biased representation. More precisely, let N be 2 to the
// power of (kBitCount - 1), an integer x is represented by the
// unsigned number x + N.
//
// For instance,
//
// -N + 1 (the most negative number representable using
// sign-and-magnitude) is represented by 1;
// 0 is represented by N; and
// N - 1 (the biggest number representable using
// sign-and-magnitude) is represented by 2N - 1.
//
// Read http://en.wikipedia.org/wiki/Signed_number_representations
// for more details on signed number representations.
static Bits SignAndMagnitudeToBiased(const Bits &sam) {
if (kSignBitMask & sam) {
// sam represents a negative number.
return ~sam + 1;
} else {
// sam represents a positive number.
return kSignBitMask | sam;
}
}
// Given two numbers in the sign-and-magnitude representation,
// returns the distance between them as an unsigned number.
static Bits DistanceBetweenSignAndMagnitudeNumbers(const Bits &sam1,
const Bits &sam2) {
const Bits biased1 = SignAndMagnitudeToBiased(sam1);
const Bits biased2 = SignAndMagnitudeToBiased(sam2);
return (biased1 >= biased2) ? (biased1 - biased2) : (biased2 - biased1);
}
FloatingPointUnion u_;
};
编辑：这篇文章是4岁。 它可能仍然有效，代码很好，但有些人找到了改进。 最好从Google Test源代码中获取最新版本的AlmostEquals ，而不是我在这里粘贴的那个。
上一篇: What is the most effective way for float and double comparison?