[PyCUDA] PyCUDA and fast user-generated code

[Andreas Kloeckner]

On Wed, 10 Nov 2010 22:10:55 -0000, "Daniel White" <twinbee42 at skytopia.com> wrote:
> Hi Tomasz and Andreas,
> 
> Thanks both for the info. I've considered OpenCL, but from what I've
> seen, it's 

<CL myth debunking below>

- not quite as mature as CUDA

To make up for that, it avoids some of CUDA's design mistakes. (e.g.
context as thread-global state) Also, the CL drivers have been through a
few revisions, and they've come a long way. Lastly, CL builds on the
same infrastructure as CUDA (in Nvidia's case) and thus benefits from
CUDA's maturity where it matters most.

- slower

Not true. A colleague of mine (Tim Warburton) was able to get equivalent
code running faster on CL than on CUDA, on the same hardware. (Perhaps
the LLVM-based compiler is better?  Who knows.)

- more difficult to use

If you compare the example programs at
http://documen.tician.de/pyopencl/
and
http://documen.tician.de/pycuda/
it's plain to see that that's not the case.

> though that might have changed in these past few months). I love the
> principle of the open standard. I just wish CUDA could become like that.
> I may also hold off until Microsoft release their GPGPU language.

MS have released a GPGPU language, it's called DirectCompute, and mainly
for use with DirectX in games. (Were they planning on releasing another
one?)

I'm not meaning to tell you that you should use CL, I'm just trying to
make sure you base your decision on accurate information. :)

Andreas
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
URL: <http://lists.tiker.net/pipermail/pycuda/attachments/20101110/f87b42f5/attachment.pgp>