LUMIERA.clone/src/lib/util-quant.hpp
Ichthyostega 418a5691ea Library: relocate integer-log2 and make it constexpr
This highly optimised function was introduced about one year ago
for handling of denomals with rational values (fractions), as
an interim solution until we'll switch to C++20.

Since this function uses an unrolled loop and basically
just does a logarithmic search for the highest set bit,
it can just be declared constexpr. Moreover, it is now
relocated into one of the basic utility headers

Remark: the primary "competitor" is the ilogb(double),
which can exploit hardware acceleration. For 64bit integers,
the ilog2() is only marginally faster according to my own
repeated invocation benchmarks.
2023-11-21 19:39:18 +01:00

206 lines
6.2 KiB
C++

/*
UTIL-QUANT.hpp - helper functions to deal with quantisation and comparison
Copyright (C) Lumiera.org
2011, Hermann Vosseler <Ichthyostega@web.de>
This program is free software; you can redistribute it and/or
modify it under the terms of the GNU General Public License as
published by the Free Software Foundation; either version 2 of
the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
*/
/** @file util-quant.hpp
** Utilities for quantisation (grid alignment) and comparisons.
*/
#ifndef LIB_UTIL_QUANT_H
#define LIB_UTIL_QUANT_H
#include <cstdlib>
#include <climits>
#include <cfloat>
#include <cmath>
namespace util {
/** helper to treat int or long division uniformly */
template<typename I>
struct IDiv
{
I quot;
I rem;
IDiv (I num, I den)
: quot(num/den)
, rem(num - quot*den)
{ }
};
template<>
struct IDiv<int>
: div_t
{
IDiv<int> (int num, int den)
: div_t(div (num,den))
{ }
};
template<>
struct IDiv<long>
: ldiv_t
{
IDiv<long> (long num, long den)
: ldiv_t(ldiv (num,den))
{ }
};
template<>
struct IDiv<long long>
: lldiv_t
{
IDiv<long long> (long long num, long long den)
: lldiv_t(lldiv (num,den))
{ }
};
template<typename I>
inline IDiv<I>
iDiv (I num, I den) ///< support type inference and auto typing...
{
return IDiv<I>{num,den};
}
/** floor function for integer arithmetics.
* Unlike the built-in integer division, this function
* always rounds towards the _next smaller integer,_
* even for negative numbers.
* @warning floor on doubles performs way better
* @see UtilFloordiv_test
*/
template<typename I>
inline I
floordiv (I num, I den)
{
if (0 < (num^den))
return num/den;
else
{ // truncate similar to floor()
IDiv<I> res(num,den);
return (res.rem)? res.quot-1 // negative results truncated towards next smaller int
: res.quot; //..unless the division result not truncated at all
}
}
/** scale wrapping operation.
* Quantises the numerator value into the scale given by the denominator.
* Unlike the built-in integer division, this function always rounds towards
* the _next smaller integer_ and also relates the remainder (=modulo) to
* this next lower scale grid point.
* @return quotient and remainder packed into a struct
* @see UtilFloorwarp_test
*/
template<typename I>
inline IDiv<I>
floorwrap (I num, I den)
{
IDiv<I> res(num,den);
if (0 > (num^den) && res.rem)
{ // negative results
// wrapped similar to floor()
--res.quot;
res.rem = den - (-res.rem);
}
return res;
}
/**
* epsilon comparison of doubles.
* @remarks Floating point calculations are only accurate up to a certain degree,
* and we need to adjust for the magnitude of the involved numbers, since
* floating point numbers are scaled by the exponent. Moreover, we need
* to be careful with very small numbers (close to zero), where calculating
* the difference could yield coarse grained 'subnormal' values.
* @param ulp number of grid steps to allow for difference (default = 2).
* Here, a 'grid step' is the smallest difference to 1.0 which can be
* represented in floating point ('units in the last place')
* @warning don't use this for comparison against zero, rather use an absolute epsilon then.
* @see https://randomascii.wordpress.com/2012/02/25/comparing-floating-point-numbers-2012-edition/
* @see http://en.cppreference.com/w/cpp/types/numeric_limits/epsilon
* @see https://en.wikipedia.org/wiki/Unit_in_the_last_place
*/
inline bool
almostEqual (double d1, double d2, unsigned int ulp =2)
{
using std::fabs;
return fabs (d1-d2) < DBL_EPSILON * fabs (d1+d2) * ulp
|| fabs (d1-d2) < DBL_MIN; // special treatment for subnormal results
}
/**
* Integral binary logarithm (disregarding fractional part)
* @return index of the largest bit set in `num`; -1 for `num==0`
* @todo C++20 will provide `std::bit_width(i)` — run a microbenchmark!
* @remark The implementation uses an unrolled loop to break down the given number
* in a logarithmic search, subtracting away the larger powers of 2 first.
* Explained 10/2021 by user «[ToddLehman]» in this [stackoverflow].
* @note Microbenchmarks indicate that this function and `std::ilogb(double)` perform
* in the same order of magnitude (which is surprising). This function gets
* slightly faster for smaller data types. The naive bitshift-count implementation
* is always significantly slower (8 times for int64_t, 1.6 times for int8_t)
* @see Rational_test::verify_intLog2()
* @see ZoomWindow_test
*
* [ToddLehman]: https://stackoverflow.com/users/267551/todd-lehman
* [stackoverflow]: https://stackoverflow.com/a/24748637 "How to do an integer log2()"
*/
template<typename I>
inline constexpr int
ilog2 (I num)
{
if (num <= 0)
return -1;
const I MAX_POW = sizeof(I)*CHAR_BIT - 1;
int logB{0};
auto remove_power = [&](I pow) constexpr
{
if (pow > MAX_POW) return;
if (num >= I{1} << pow)
{
logB += pow;
num >>= pow;
}
};
remove_power(32);
remove_power(16);
remove_power (8);
remove_power (4);
remove_power (2);
remove_power (1);
return logB;
}
} // namespace util
#endif /*UTIL_QUANT_H*/