I was talking about emulation mode.
It is a pity that I cannot use it with pycuda. For I have implemented
a part of my algorithm in cuda now, and wanted to see how much faster
it is. Right now I can only compare it with my previous implementation
using python and rpy. And this will always be slower than anything
wrapped around some C stuff. So a quick way to see what the cuda stuff
actually did, was using the emulation mode.
However, for me a real benchmark is not required. I was just curious.
I am just using this algorithm to learn stuff about cuda, because I
see some real potential in it.
And it is always better to learn new stuff for things you need.
My first Python/R implementation took 1.5 hours and now it is done in
3 seconds. And that with only an hour or 2 coding/learning. I know for
sure that I could have slashed a big chunk from the 1.5 hours by just
doing the computation in C instead of R, but then I woulnd't be
learning cuda :)

But thanks for explaining that it is not possible nor that it would be
a good benchmark.

Cheers,

Willem Ligtenberg

On Mon, Apr 20, 2009 at 11:38, Ahmed Fasih <[email protected]> wrote:
> I think you are talking about emulation mode, which depends on the
> "other" CUDA API that PyCUDA is not written in. It is not supported in
> PyCUDA: http://tiker.net/pipermail/pycuda_tiker.net/2008-December/000063.html
>
> Furthermore, emulation mode is no way to benchmark speedups; this
> would be unacceptable for a computing journal or conference (or by
> anyone else :P). Frequently the most trivial CPU implementation of an
> algorithm is faster than emulation mode, especially for large
> problems. To compare CPU and GPU performance, you find the fastest CPU
> implementation (leveraging all the features of modern processors such
> as SIMD, multi-threading, maximum cache hits, etc.) and compare your
> GPU code to that.
>
> On Mon, Apr 20, 2009 at 5:14 AM, Willem Ligtenberg
> <[email protected]> wrote:
>> Hi,
>>
>> I know cuda has a compile option that lets you run the code on CPU
>> instead of GPU (and then uses multicore as well). How do I make my
>> pycuda script run on the CPUinstead of the GPU? This may also help me
>> with specifying how much faster it goes.
>>
>> Thanks in advance,
>>
>> Willem Ligtenberg
>>
>> _______________________________________________
>> PyCuda mailing list
>> [email protected]
>> http://tiker.net/mailman/listinfo/pycuda_tiker.net
>>
>

_______________________________________________
PyCuda mailing list
[email protected]
http://tiker.net/mailman/listinfo/pycuda_tiker.net

Reply via email to