Assertions for floating-point comparisons

[ghost]

It would be nice if testworks provide at least one method to compare floating-point numbers

[waywardmonkeys]

I need this as well now.

Previously, @niko2342 had found https://randomascii.wordpress.com/2012/02/25/comparing-floating-point-numbers-2012-edition/ which seems like a good start.

[waywardmonkeys]

To compare ULPs, you can use decode-single-float or decode-double-float from dylan-extensions:

let desired-lo = decode-double-float(desired);
let calc-lo = decode-double-float(calc);
assert-true(abs(as(<integer>, desired-lo) - as(<integer>, calc-lo)) < 2)

[waywardmonkeys]

@housel pointed me at this code from Gwydion Dylan:

define constant $max-ulps = 3.0d0;

define method almost-equal
    (x :: <float>, y :: <real>)
 => (result :: <boolean>)
  if (y.zero?)
    abs(x) < scale-float($max-ulps, 1 - float-digits(x));
  else
    let (x-sig, x-exp, x-sign) = decode-float(x);
    let scaled-diff = abs(scale-float(x, - x-exp) - scale-float(y, - x-exp));
    let ulp-diff = scale-float(scaled-diff, float-digits(x) - 1);

    ulp-diff < $max-ulps;
  end if;
end method almost-equal;