Re: Problems with Octave random number generation, part 1

[Daniel J Sebald]

On 12/18/19 12:39 AM, Daniel J Sebald wrote:
> On 12/17/19 2:27 PM, Rik wrote:
> > Maintainers,
[snip]
> > Given that IEEE single-precision floating point has only 24 bits for the
> > mantissa, I think something like this would work better
> >
> >    static float randu32 (void)
> >    {
> >      uint32_t i;
> >
> >      do
> >        {
> >          i = randi32 () & static_cast<uint32_t> (0xFFFFFF);
> >        }
> >      while (i == 0 || i == 0xFFFFFF);
> >
> >      return i * (1.0f / 16777216.0f);
> >    }
> >
> > In my testing, this does not produce exact values of 1, nor does it map
> > multiple integer values to the same floating point value.  Is anyone 
> > aware
> > of problems with this algorithm?
>
> Creative, but mapping out that whole upper level of pseudo-random 
> numbers doesn't feel good.  It may be as good as anything though 
> because, yeah, we could change the name of the function to randu24() and 
> it would have the same meaning because of the imposed range of (0,1). In 
> some way, that range must implicitly toss away the exponent and the 
> floating point bit.  It's effectively now fractional point, not floating 
> point.

In fact, the range [1:0xFFFFFE] could go directly into the mantissa 
rather than multiply or divide.  I.e., think of 1/16777216 as 2^-24 or 
right bit shift by 24.  Is there a C macro for loading a mantissa directly?

Dan