[torotonnato]

Of course, it's uniform. I always swap the two names randomly hehe. I understand your code. The difference with the snippet I posted (which, btw, is not mine) is that the difference between successive floats is always a constant, i.e. there is no rounding. It's fast, has 23 bits of entropy and it's suited for constrained platforms/small demos [0], since it can be coded with very few asm instructions.

I'm not saying that your solution hasn't its merits, well done.

[0] The author of the code is a known member of the demoscene

[orlp]

Oh, I know very well who iq is :) And for a visual demo it might make perfect sense to sacrifice completeness of a solution for speed / code size. But to me, "pick a random number from {k/2^23 : 0 <= k < 2^23, k ∈ ℕ}" and "pick a random number uniformly at random from [0, 1)" are not the same.

All operations (add/sub, mul/div, sin, exp, log, etc...), with IEEE754 floats are defined using real arithmetic, and then assumed to round to the nearest representable floating point value. I don't see why random number generation, at least as a default before micro-optimizing, should be any different.

[torotonnato]

You are right, I was uncomfortable with the uneven density of the floats approaching zero, but it's just an inherent property of their representation. Have a nice day :)