generate random numbers in a kernel

[Claudio Martella]

I've been trying for some time to generate random numbers in numba inside of a kernel (e.g., inside a vectorize). I've tried using numpy.random.uniform, i've tried using cuda.np.random.uniform, i've tried random.uniform, but all complain of undefined types in the nopython frontend.

Is there a way to generate random numbers ala curadom in cudalib?

something like:

@vectorize("float64(float64)", target="cuda") # could be target="parallel" too

def get_random(a):

    return a + numpy.random.uniform(0, 1)

# in principle in my app they won't be zeros of couse

data = numpy.zeros(10000000, dtype=float64)

get_random(data)

Thanks,

/CM

[Stanley Seibert]

Hi,

We don't yet have a built-in random number generator for the CUDA target.  The curand device functions that NVIDIA provides for CUDA C are not available at the LLVM level where we generate CUDA code.  (We've asked them to put these functions into their LLVM bitcode library, so hopefully they will in the future.)  Fixing this is on our todo list.

In the mean time, Siu has an xorshift implementation that you can use with Numba:

https://gist.github.com/sklam/2b9574b51c6bd2e6cf9e

(We don't want to support this RNG officially because there may be some problems with xorshift that we aren't aware of.)

--
You received this message because you are subscribed to the Google Groups "Numba Public Discussion - Public" group.
To unsubscribe from this group and stop receiving emails from it, send an email to numba-users...@continuum.io.
To post to this group, send email to numba...@continuum.io.
To view this discussion on the web visit https://groups.google.com/a/continuum.io/d/msgid/numba-users/1ef5145c-e7c8-48a8-bc0a-8e03645899c7%40continuum.io.
For more options, visit https://groups.google.com/a/continuum.io/d/optout.

[Claudio Martella]

Thanks for the quick answer. xorshift* does indeed look like a good approach. One question about the usage of that implementation (thanks for it):

how am I supposed to initialize states and id? it is not completely clear to me how to use them, for example, if i want to add a random number to each element of an array (simplest example).

[rall...@cfa.harvard.edu]

Hi Stanley,

Thanks for linking to the example, but it is terribly unclear how to use it. What is states and what is id?

Ryan

[Siu Kwan Lam]

I added an example here: https://gist.github.com/sklam/2b9574b51c6bd2e6cf9e#gistcomment-1900718

To view this discussion on the web visit https://groups.google.com/a/continuum.io/d/msgid/numba-users/2d326284-e47e-4850-be4d-ff97fc2b212c%40continuum.io.

For more options, visit https://groups.google.com/a/continuum.io/d/optout.

--

Siu Kwan Lam

Software Engineer

Continuum Analytics

[Kevin Sheppard]

The most common approach is to read from random or urandom (or WinCryptProvider).  If you need more bits, say if you have a very large state, you can use 64 bits from system RNGs plus splitmix64 to generate more than enough to generate an initial state for an xorshift RNG. 

To view this discussion on the web visit https://groups.google.com/a/continuum.io/d/msgid/numba-users/CAHjo7%3DXm%3Dp3mjytFSpDFsYN7nXdU_xwH62%3DPR0PPZU0uiV92bQ%40mail.gmail.com.