Just wondering, what's the performance for initialising a large pyopencl.array in PyPy?
I'm asking because numpy.empty has abysmal performance on PyPy when you get to large dimensions:

CPython:
>>> timeit('np.empty((5,), dtype=np.float64)', setup='import numpy as np', number=1e6)
0.45080995559692383
>>> timeit('np.empty((500000,), dtype=np.float64)', setup='import numpy as np', number=1e6)
0.5748040676116943

PyPy:
>>>> timeit('np.empty((5,), dtype=np.float64)', setup='import numpy as np', number=1e6)
0.20131182670593262
>>>> timeit('np.empty((500000,), dtype=np.float64)', setup='import numpy as np', number=1e3)
1.2475969791412354

PyPy is 2000 times slower than CPython when allocating 4MB worth of memory O_O


On 31 May 2014 19:18, Andreas Kloeckner <lists@informa.tiker.net> wrote:
Tomasz Rybak <tomasz.rybak@post.pl> writes:
> First - general comment.
> I quite like an idea of changing PyOpenCL to work with PyPy.
> Can I assume that this is just for PyOpenCL, and PyCUDA stays
> Boost (and thus CPython) only for now?

Yep, I have no plans currently to port PyCUDA to cffi. If someone
decided to do that, I'd be supportive, but for now only PyOpenCL will
support PyPy.

> I do not know yet how packaging will look like for package
> that both supports CPython and PyPy - I'll need to check that.

To be safe, I think pyopencl-bpl and pyopencl-cffi should be two
different packages, perhaps with a virtual package for dependencies to
use. pyopencl-cffi works on both CPython and PyPy, so it'd be nice if
the package exposed that. The two packages should conflict with each
other because they both supply a module called "pyopencl".

> As for testing in the wild - if I'm able to build PyPy
> version I hope to be able to upload it to Debian experimental,
> so interested Debian users could test it. I think that
> some separate versioning for PyPy version will help here.

Great. PyPy's numpy isn't packaged yet, and PyPy 2.3 (which fixes a lot
of bugs, thus making all of this possible) is only a recent addition to
the Debian archive.

[1] https://bitbucket.org/pypy/numpy

Andreas


_______________________________________________
PyOpenCL mailing list
PyOpenCL@tiker.net
http://lists.tiker.net/listinfo/pyopencl