[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[Discuss-gnuradio] R: Re: THESIS with CUDA
From: |
address@hidden |
Subject: |
[Discuss-gnuradio] R: Re: THESIS with CUDA |
Date: |
Sun, 25 Jan 2015 18:53:41 +0100 (CET) |
Thanks for your response.
>----Messaggio originale----
>Da: address@hidden
>Data: 25-gen-2015 18.29
>A: <address@hidden>
>Ogg: Re: [Discuss-gnuradio] THESIS with CUDA
>
>Hey Marco,
>
>good luck for your project!
>
>FYI, there have been quite a few projects using GPPs in the past. I
>recommend going through the slides from all of the GRCons, you'll find
>something there (at the very least from the first GRCon).
>Also, there's gr-theano, which also uses CUDA, and for inspiration,
>having a look at gr-fosphor might be a good idea.
>
>For high-performance GR flow graphs, the overhead for ingress/egress has
>always been the problem. If you have ideas for modifying GR itself, it
>might be worth contacting the "Accelerators" working group, which
>discusses these matters.
Ye,I can imagine. Also because there are noticeable difference between
different CUDA devices. And with OpenCL it can go worse, because there are
devices of different categories too.
One of my possible idea in order to reduce it, is to merge different blocks on
the same kernel:serial,"concatenating code",or in "parallel",where each warp
elaborate different block; In this way is possible to use shared memory(which
is very small).
But it is complicated(PTX code,while OpenCL can compile more easily; different
schedule policy) and I will try it at the end,after time spent over more normal
version,like multiple kernel communicating via global memory on-card. Latency
will be my hell.
>
>Cheers,
>M
>
[Prev in Thread] |
Current Thread |
[Next in Thread] |
- [Discuss-gnuradio] R: Re: THESIS with CUDA,
address@hidden <=