Keegan Owsley <keegano(a)gmail.com> writes:
I've just slapped together a patch to pycuda that
makes most elementwise
operations work with noncontiguous arrays. There are a bunch of hacks in
there, and the code needs some reorg before it's ready to be considered for
upstream (I made these changes while learning the pycuda codebase, so
there's a bunch of crud that can be cleaned out), but I figure I might as
well put it out there in its current state and see what you guys think.
It's also not extremely well-tested (I have no idea if it interferes with
skcuda, for example), but all of the main functions appear to work.
You can check out the code at https://bitbucket.org/owsleyk_omega/pycuda
Briefly, this works by adding new parameters into elementwise kernels that
describe the stride and shape of your arrays, then using a function that
computes the location in memory from the stride, shape, and index.
Elementwise kernel ops are modified so that they use the proper indexing.
See an example of a kernel that's generated below:
Thanks for putting this together and sharing it! I have one main
question about this, regarding performance:
Modulo (especially variable-denominator modulo) has a habit of being
fantastically slow on GPUs. Could you time contiguous
vs. noncontiguous for various levels of "gappiness" and number of
axes? I'm asking this because I'd be OK with a 50% slowdown, but not
necessarily a factor of 5 slowdown on actual GPU hardware.