Scramble a floating point number?

Question

I need a repeatable pseudo-random function from floats in [0,1] to floats in [0,1]. I.e. given a 32-bit IEEE float, return a "different" one (as random as possible, given the 24 bits of mantissa). It has to be repeatable, so keeping tons of internal state is out. And unfortunately it has to work with only 32-bit int and single-float math (no doubles and not even 32x32=64bit multiply, though I could emulate that if needed -- basically it needs to work on older CUDA hardware). The better the randomness the better, of course, within these rather severe limitations. Anyone have any ideas?

(I've been through Park-Miller, which requires 64-bit int math, and the CUDA version of Park-Miller which requires doubles, Mersenne Twisters which have lots of internal state, and a few other things which didn't work.)

Answer 1

Best I understand the requirements, a hash accomplishes the desired functionality. Re-interprete the float input as an integer, apply the hash function to produce an integer approximately uniformly distributed in [0,2^32), then multiply this integer by 2^-32 to convert the resulting integer back to a float roughly uniformly distributed in [0,1]. One suitable hash function which does not require multiplication is Bob Jenkin's mix(), which can be found here: http://www.burtleburtle.net/bob/hash/doobs.html.

To re-interpret the bits of a float as an integer and vice versa, there are two choices in CUDA. Use intrinsics, or use C++-style reinterpretation casts:

float f;
int i;
i = __float_as_int(f);
f = __int_as_float(i);
i = reinterpret_cast<int&>(f);
f = reinterpret_cast<float&>(i);

So as a self-contained function, the entire process might look something like this:

/* transform float in [0,1] into a different float in [0,1] */
float scramble_float (float f)
{
    unsigned int magic1 = 0x96f563ae; /* number of your choice */
    unsigned int magic2 = 0xb93c7563; /* number of your choice */
    unsigned int j;
    j = reinterpret_cast<unsigned int &>(f);
    mix (magic1, magic2, j);
    return 2.3283064365386963e-10f * j;
}

Answer 2

The NVIDIA CUDA Toolkit includes a library called CURAND that I believe fits your requirements: it produces repeatable results (assuming you start with the same seed), works on the GPU, supports 32-bit floats and ints, and should work on older GPUs. It also supports multiple pseudo- and quasi-random generation algorithms and distributions.

[Note: a problem with using the C library rand() function (other than that it does not run in CUDA on the device) is that on Windows, rand() only returns a 16-bit value, and thus any float created by division by RAND_MAX has only 16 random bits of precision. What's more, on linux/mac it returns a 32-bit value so code that uses it is not numerically portable.]