120 lines
		
	
	
		
			3.7 KiB
		
	
	
	
		
			V
		
	
	
			
		
		
	
	
			120 lines
		
	
	
		
			3.7 KiB
		
	
	
	
		
			V
		
	
	
// Copyright (c) 2019-2022 Alexander Medvednikov. All rights reserved.
 | 
						|
// Use of this source code is governed by an MIT license
 | 
						|
// that can be found in the LICENSE file.
 | 
						|
module fractions
 | 
						|
 | 
						|
import math
 | 
						|
 | 
						|
const (
 | 
						|
	default_eps    = 1.0e-4
 | 
						|
	max_iterations = 50
 | 
						|
	zero           = fraction(0, 1)
 | 
						|
)
 | 
						|
 | 
						|
// ------------------------------------------------------------------------
 | 
						|
// Unwrapped evaluation methods for fast evaluation of continued fractions.
 | 
						|
// ------------------------------------------------------------------------
 | 
						|
// We need these functions because the evaluation of continued fractions
 | 
						|
// always has to be done from the end. Also, the numerator-denominator pairs
 | 
						|
// are generated from front to end. This means building a result from a
 | 
						|
// previous one isn't possible. So we need unrolled versions to ensure that
 | 
						|
// we don't take too much of a performance penalty by calling eval_cf
 | 
						|
// several times.
 | 
						|
// ------------------------------------------------------------------------
 | 
						|
// eval_1 returns the result of evaluating a continued fraction series of length 1
 | 
						|
fn eval_1(whole i64, d []i64) Fraction {
 | 
						|
	return fraction(whole * d[0] + 1, d[0])
 | 
						|
}
 | 
						|
 | 
						|
// eval_2 returns the result of evaluating a continued fraction series of length 2
 | 
						|
fn eval_2(whole i64, d []i64) Fraction {
 | 
						|
	den := d[0] * d[1] + 1
 | 
						|
	return fraction(whole * den + d[1], den)
 | 
						|
}
 | 
						|
 | 
						|
// eval_3 returns the result of evaluating a continued fraction series of length 3
 | 
						|
fn eval_3(whole i64, d []i64) Fraction {
 | 
						|
	d1d2_plus_n2 := d[1] * d[2] + 1
 | 
						|
	den := d[0] * d1d2_plus_n2 + d[2]
 | 
						|
	return fraction(whole * den + d1d2_plus_n2, den)
 | 
						|
}
 | 
						|
 | 
						|
// eval_cf evaluates a continued fraction series and returns a Fraction.
 | 
						|
fn eval_cf(whole i64, den []i64) Fraction {
 | 
						|
	count := den.len
 | 
						|
	// Offload some small-scale calculations
 | 
						|
	// to dedicated functions
 | 
						|
	match count {
 | 
						|
		1 {
 | 
						|
			return eval_1(whole, den)
 | 
						|
		}
 | 
						|
		2 {
 | 
						|
			return eval_2(whole, den)
 | 
						|
		}
 | 
						|
		3 {
 | 
						|
			return eval_3(whole, den)
 | 
						|
		}
 | 
						|
		else {
 | 
						|
			last := count - 1
 | 
						|
			mut n := i64(1)
 | 
						|
			mut d := den[last]
 | 
						|
			// The calculations are done from back to front
 | 
						|
			for index := count - 2; index >= 0; index-- {
 | 
						|
				t := d
 | 
						|
				d = den[index] * d + n
 | 
						|
				n = t
 | 
						|
			}
 | 
						|
			return fraction(d * whole + n, d)
 | 
						|
		}
 | 
						|
	}
 | 
						|
}
 | 
						|
 | 
						|
// approximate returns a Fraction that approcimates the given value to
 | 
						|
// within the default epsilon value (1.0e-4). This means the result will
 | 
						|
// be accurate to 3 places after the decimal.
 | 
						|
pub fn approximate(val f64) Fraction {
 | 
						|
	return approximate_with_eps(val, fractions.default_eps)
 | 
						|
}
 | 
						|
 | 
						|
// approximate_with_eps returns a Fraction
 | 
						|
pub fn approximate_with_eps(val f64, eps f64) Fraction {
 | 
						|
	if val == 0.0 {
 | 
						|
		return fractions.zero
 | 
						|
	}
 | 
						|
	if eps < 0.0 {
 | 
						|
		panic('Epsilon value cannot be negative.')
 | 
						|
	}
 | 
						|
	if math.fabs(val) > math.max_i64 {
 | 
						|
		panic('Value out of range.')
 | 
						|
	}
 | 
						|
	// The integer part is separated first. Then we process the fractional
 | 
						|
	// part to generate numerators and denominators in tandem.
 | 
						|
	whole := i64(val)
 | 
						|
	mut frac := val - f64(whole)
 | 
						|
	// Quick exit for integers
 | 
						|
	if frac == 0.0 {
 | 
						|
		return fraction(whole, 1)
 | 
						|
	}
 | 
						|
	mut d := []i64{}
 | 
						|
	mut partial := fractions.zero
 | 
						|
	// We must complete the approximation within the maximum number of
 | 
						|
	// itertations allowed. If we can't panic.
 | 
						|
	// Empirically tested: the hardest constant to approximate is the
 | 
						|
	// golden ratio (math.phi) and for f64s, it only needs 38 iterations.
 | 
						|
	for _ in 0 .. fractions.max_iterations {
 | 
						|
		// We calculate the reciprocal. That's why the numerator is
 | 
						|
		// always 1.
 | 
						|
		frac = 1.0 / frac
 | 
						|
		den := i64(frac)
 | 
						|
		d << den
 | 
						|
		// eval_cf is called often so it needs to be performant
 | 
						|
		partial = eval_cf(whole, d)
 | 
						|
		// Check if we're done
 | 
						|
		if math.fabs(val - partial.f64()) < eps {
 | 
						|
			return partial
 | 
						|
		}
 | 
						|
		frac -= f64(den)
 | 
						|
	}
 | 
						|
	panic("Couldn't converge. Please create an issue on https://github.com/vlang/v")
 | 
						|
}
 |