Did I win something for this ? ;)

[maxtarpini]

Jeez what a luck.. a full zero from a random sampler ! 🙂
Jokes a part, there're exactly zero chances for a full of anything in a stochastic simulation.

Repro steps. Open Maya (arnold 6.2.1.0). Put a sphere in the scene. Press F in persp to bring it closer. Add a skydome with any image. Previously you coded a BRDF (even empty) where you just check for zeros in the sample routine from the given default rnds. Discovered this because my code was crashing with a division by zero.

[thiago.ize]

Arnold uses 32-bit floats, which means there are only a finite number of values that can be represented, so the odds are not 1 in infinity of getting exactly zero but roughly 1 in 4 billion. What's more, because we're not trying to get just any random float value but a random number in [0,1), that's 0 all the way to 0.9999999..., the possible values are really more like 1 in a few million.

If you are casting millions of rays, the odds of getting exactly zero goes dramatically up, so it's actually pretty likely this will happen.

If you don't want 0 to happen but are OK with 1, take the random value X and transform it by 1-X. Then you'll have a value that goes from (0,1].

[maxtarpini]

It is true that we use float and not reals, but there are still one billion numbers in the 0-1 floating-point range. Whenever we ask for a random number there is always a one-in-a-billion chance of getting an exact zero, so getting one as long as it's a chance is generally an indication of having a somehow biased sampler. However that's an interesting thing to research, I'll dig with this a bit more formally when time. Thanks for the inputs.

[thiago.ize]

You're right, there are a billion floats in [0,1), but the arnold random number generator produces uniformly distributed values from [0,1) for which there's just a few million values.

Here's a page I found which goes into more details: https://lemire.me/blog/2017/02/28/how-many-floating-point-numbers-are-in-the-interval-01/

even though your computer can represent about 2³⁰ distinct and normal floating-point numbers in [0,1), chances are good that your random generator only produces 2²⁴ distinct floating-point numbers in the interval [0,1)

[maxtarpini]

To some extent it's still debatable if you look at comments there. Btw, that's interesting that there're more fp numbers around zero so for a truly uniform distribution of fp random variates one needs a non uniform distribution of fp numbers. Just to say that's not necessary that the characterization of some random generator properties belong to its final outcome representation (and viceversa).

[maxtarpini]

Eventually let me say I'm super cool with your sampler and that a minor problem is just that I've to use another one as 3 rnd numbers(sample fnc) ain't enough sometime and would be great to have a way to ask for more.