Re: Why ParallelTable does not use CUDA
- To: mathgroup at smc.vnet.net
- Subject: [mg114330] Re: Why ParallelTable does not use CUDA
- From: Zach Bjornson <bjornson at stanford.edu>
- Date: Wed, 1 Dec 2010 02:11:55 -0500 (EST)
- References: <ico25c$niu$1@smc.vnet.net> <id2ek3$d4i$1@smc.vnet.net>
You have CUDAMap available, so you can do
CUDAMap[Sin,1.0*Range[0,1000,0.1]]
My guess is that CUDA's limitations make something like CUDATable too
prone to upset users who want it to do more than the name implies it is
capable of. That is, there are severe memory limitations, varied support
for/handling of doubles and limited support for branching and
conditional structures. I'd venture to say that everything reasonably
possible in a "CUDATable" right now is doable with CUDAMap. As evident
when comparing other languages that have CUDA interfaces, for the most
part it's still better to vectorize things, and that's clearly the
approach Wolfram has stuck to so far.
Zach -- (admittedly with only limited experience in GPU computing)
On 11/30/2010 1:02 AM, foice wrote:
> On Nov 26, 11:29 am, foice<franceschini.robe... at gmail.com> wrote:
>> I was wandering why should I write a complicated and unnecessary CUDA
>> kernel to do something a number of times and put the result in a
>> table. We already have a thing like ParallelTalble ... it seems
>> however that ParallelTable uses the CPUs that one has available in his
>> system and not the GPU ... I do not see the point of having CUDA if
>> this is not even used by the internal parallel functions ...
>>
>> ParallelTable[Sin[x],{x,0,1000,0.1},CUDA->True] would be great!!!
>
> it's a shame parallel still means cpu parallel and a shame that nobody
> answers :(
>