libcruft-util/float.cpp

#include "float.hpp"

#include <cmath>

using namespace std;


/* Constructors */

template <unsigned int E, unsigned int S>
ieee_float<E, S>::ieee_float (void)
{ ; }


template <unsigned int E, unsigned int S>
ieee_float<E, S>::ieee_float (floating_t _floating):
    m_floating (_floating)
{ ; }


template <unsigned int E, unsigned int S>
ieee_float<E, S>::ieee_float (const ieee_float &rhs):
    m_bits (rhs.m_bits)
{ ; }


/* Classifiers */

template <unsigned int E, unsigned int S>
bool
ieee_float<E, S>::is_zero (void) const {
    return m_components.exponent    == 0 &&
           m_components.significand == 0;
}

template <unsigned int E, unsigned int S>
bool
ieee_float<E, S>::is_subnormal (void) const {
    return m_components.exponent    == 0 &&
           m_components.significand != 0;
}

template <unsigned int E, unsigned int S>
bool
ieee_float<E, S>::is_inifinity (void) const {
    return m_components.exponent    == (1 << EXPONENT_BITS) - 1 &&
           m_components.significand == 0;
}

template <unsigned int E, unsigned int S>
bool
ieee_float<E, S>::is_nan (void) const {
    return m_components.exponent    == (1 << EXPONENT_BITS) - 1 &&
           m_components.significand != 0;
}

template <unsigned int E, unsigned int S>
bool
ieee_float<E, S>::operator==(floating_t _floating) const {
    // TODO: This method really shouldn't be generated if there's no
    // representative native floating point type. But I'm sick of
    // C++'s template bullshit for tonight.
    CHECK (bits_type<TOTAL_BITS>::has_floating);

    union {
        floating_t _floating;
        uint_t     _uint;
    } convertor;

    convertor._floating = _floating;

    return m_bits == convertor._uint;
}


#include <iostream>


//template <unsigned int E, unsigned int S>
//bool
//ieee_float<E, S>::almost_equal (floating_t a,
//                                floating_t b) {
//    // Static cast to avoid integer casting warnings when using uint16_t for half
//    static const floating_t epsilon = static_cast<floating_t> (0.001);
//    const floating_t diff = static_cast<floating_t> (std::fabs (a - b));
//
//    // * Use an exact equality first so that infinities are not indirectly compared. This would generate NaNs in the diff.
//    // * Do not use gte or lte. This stops an infinite from making infinities on both sides of the inequality.
//    return exactly_equal (a, b)           ||
//           diff < epsilon * std::fabs (a) ||
//           diff < epsilon * std::fabs (b);
//}


template <unsigned int E, unsigned int S>
bool
ieee_float<E, S>::almost_equal (floating_t a,
                                floating_t b)
{
    return almost_equal (a, b, 128u);
}


// Based on the cynus `AlmostEqual2sComplement`
template <unsigned int E, unsigned int S>
bool
ieee_float<E, S>::almost_equal (floating_t _a,
                                floating_t _b,
                                unsigned ulps)
{
    // Ensure ULPs is small enough that the default NaNs won't compare as
    // equal to anything else.
    CHECK_LE (ulps, 4 * 1024 * 1024u);

    union {
        floating_t f;
        sint_t     s;
        uint_t     u;
    } a, b;

    a.f = _a;
    b.f = _b;

    // Special case the NaNs early so simplify diffs
    if (std::isnan (a.f) || std::isnan (b.f))
        return false;

    // Early out, as identity comparisons are reasonably common
    if (a.s == b.s)
        return true;

    // Re-base negative floats to be continuous against +ve/-ve 0
    static const union {
        floating_t f;
        sint_t     s;
    } NEG_ZERO { -floating_t {0} };

    if (a.s < 0)
        a.s = NEG_ZERO.s - a.s;
    if (b.s < 0)
        b.s = NEG_ZERO.s - b.s;

    // Calculate ULP difference, but don't use abs(a.s - b.s) as it may cause
    // signed overflow
    uint_t diff = a.u > b.u ? a.u - b.u : b.u - a.u;
    return diff <= ulps;
}


//-----------------------------------------------------------------------------
template class ieee_float< 5,  10>; // ieee_half
template class ieee_float< 8,  23>; // ieee_single;
template class ieee_float<11,  52>; // ieee_double;