HOW-TO: Building a parallel and CUDA-enabled CP2K binary in a 64-bit Linux box with GNU compilers

[labicia]

We have followed the instructions of the given link but found the following 
problem:

NO COMPILATION due to undefined reference to "cuda_d_gemm"

We are using CUDA Version 5.0.

Removing "cuda_d_gemm" we are able to compile, obtaining a speedup of 150% 
in the test with 128 water molecules with respect to no GPU usage.
Is it possible to get better results improving the compilation options?

Moreover, we have some concern regarding the usage of both flags:
 "__CUDAPW and __FFTSGL" 
together with " __CUBLASDP".
In the README of CP2K/cuda_tools it seems that first couple of flags should 
be used instead of  " __CUBLASDP".

Any advice is mostly welcome. We are moving our first steps in the usage of 
CUDA.

Thanks a lot!
Marco and Andrea



Il giorno sabato 1 dicembre 2012 00:16:56 UTC+1, mirix ha scritto:
>
> I hope that this how-to will save some time to people willing to build 
> CP2K with GNU compilers:
>
>
> http://prosciens.com/logbook/2012/11/30/notes-on-building-parallel-and-cuda-enabled-cp2k-in-a-linux-box-with-gnu-compilers/
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20130109/78de90ed/attachment.htm>