比较两个双精度浮点数或两个浮点数最有效的方法是什么?

简单地这样做是不正确的:

bool CompareDoubles1 (double A, double B)
{
   return A == B;
}

比如:

bool CompareDoubles2 (double A, double B) 
{
   diff = A - B;
   return (diff < EPSILON) && (-diff < EPSILON);
}

似乎是浪费加工。

有人知道更聪明的浮点比较器吗?


当前回答

我对任何涉及浮点减法的答案都非常谨慎(例如,fabs(a-b) < epsilon)。首先,浮点数在更大的量级上变得更稀疏,在足够大的量级上,当间隔大于时,您可能只需要做a == b。其次,减去两个非常接近的浮点数(因为您正在寻找接近相等的浮点数)正是您得到灾难性抵消的方式。

虽然不能移植,但我认为grom的答案在避免这些问题方面做得最好。

其他回答

你写的代码有bug:

return (diff < EPSILON) && (-diff > EPSILON);

正确的代码应该是:

return (diff < EPSILON) && (diff > -EPSILON);

(…是的,这是不同的)

我想知道晶圆厂是否会让你在某些情况下失去懒惰的评价。我会说这取决于编译器。你可能想两种都试试。如果它们在平均水平上是相等的,则采用晶圆厂实现。

如果你有一些关于两个浮点数中哪一个比另一个更大的信息,你可以根据比较的顺序来更好地利用惰性求值。

最后,通过内联这个函数可能会得到更好的结果。不过不太可能有太大改善……

编辑:OJ,谢谢你纠正你的代码。我相应地删除了我的评论

有关更深入的方法,请参阅比较浮点数。以下是该链接的代码片段:

// Usable AlmostEqual function    
bool AlmostEqual2sComplement(float A, float B, int maxUlps)    
{    
    // Make sure maxUlps is non-negative and small enough that the    
    // default NAN won't compare as equal to anything.    
    assert(maxUlps > 0 && maxUlps < 4 * 1024 * 1024);    
    int aInt = *(int*)&A;    
    // Make aInt lexicographically ordered as a twos-complement int    
    if (aInt < 0)    
        aInt = 0x80000000 - aInt;    
    // Make bInt lexicographically ordered as a twos-complement int    
    int bInt = *(int*)&B;    
    if (bInt < 0)    
        bInt = 0x80000000 - bInt;    
    int intDiff = abs(aInt - bInt);    
    if (intDiff <= maxUlps)    
        return true;    
    return false;    
}

I found that the Google C++ Testing Framework contains a nice cross-platform template-based implementation of AlmostEqual2sComplement which works on both doubles and floats. Given that it is released under the BSD license, using it in your own code should be no problem, as long as you retain the license. I extracted the below code from http://code.google.com/p/googletest/source/browse/trunk/include/gtest/internal/gtest-internal.h https://github.com/google/googletest/blob/master/googletest/include/gtest/internal/gtest-internal.h and added the license on top.

一定要将GTEST_OS_WINDOWS定义为某个值(或者将使用它的代码更改为适合您的代码库的代码-毕竟它是BSD许可的)。

使用的例子:

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年前写的。它可能仍然有效,代码也很好,但有些人发现了改进。最好从谷歌Test源代码中获得最新版本的AlmostEquals,而不是我粘贴在这里的那个。

在这个版本中,你可以检查,这些数字之间的差异并不比某些分数(比如,0.0001%)更大:

bool floatApproximatelyEquals(const float a, const float b) {
    if (b == 0.) return a == 0.; // preventing division by zero
    return abs(1. - a / b) < 1e-6;
}

请注意Sneftel关于浮动可能的分数限制的评论。

还要注意的是,它不同于使用绝对的epsilon的方法——这里你不需要担心“数量级”——数字可能是,比如说1e100,或者1e-100,它们总是会被一致地比较,而且你不必为每一种情况更新epsilon。

我的课程是基于之前发布的答案。非常类似于谷歌的代码,但我使用了一个偏差,将所有NaN值推到0xFF000000以上。这样可以更快地检查NaN。

这段代码是为了演示概念,而不是通用的解决方案。谷歌的代码已经展示了如何计算所有平台特定的值,我不想复制所有这些。我对这段代码做了有限的测试。

typedef unsigned int   U32;
//  Float           Memory          Bias (unsigned)
//  -----           ------          ---------------
//   NaN            0xFFFFFFFF      0xFF800001
//   NaN            0xFF800001      0xFFFFFFFF
//  -Infinity       0xFF800000      0x00000000 ---
//  -3.40282e+038   0xFF7FFFFF      0x00000001    |
//  -1.40130e-045   0x80000001      0x7F7FFFFF    |
//  -0.0            0x80000000      0x7F800000    |--- Valid <= 0xFF000000.
//   0.0            0x00000000      0x7F800000    |    NaN > 0xFF000000
//   1.40130e-045   0x00000001      0x7F800001    |
//   3.40282e+038   0x7F7FFFFF      0xFEFFFFFF    |
//   Infinity       0x7F800000      0xFF000000 ---
//   NaN            0x7F800001      0xFF000001
//   NaN            0x7FFFFFFF      0xFF7FFFFF
//
//   Either value of NaN returns false.
//   -Infinity and +Infinity are not "close".
//   -0 and +0 are equal.
//
class CompareFloat{
public:
    union{
        float     m_f32;
        U32       m_u32;
    };
    static bool   CompareFloat::IsClose( float A, float B, U32 unitsDelta = 4 )
                  {
                      U32    a = CompareFloat::GetBiased( A );
                      U32    b = CompareFloat::GetBiased( B );

                      if ( (a > 0xFF000000) || (b > 0xFF000000) )
                      {
                          return( false );
                      }
                      return( (static_cast<U32>(abs( a - b ))) < unitsDelta );
                  }
    protected:
    static U32    CompareFloat::GetBiased( float f )
                  {
                      U32    r = ((CompareFloat*)&f)->m_u32;

                      if ( r & 0x80000000 )
                      {
                          return( ~r - 0x007FFFFF );
                      }
                      return( r + 0x7F800000 );
                  }
};