Re: [AMBER-Developers] problems in buildling pbsa.cuda

From: Ruxi Qi <ruxiq.uci.edu>
Date: Thu, 6 Apr 2017 23:38:11 -0700

Hi All,

I found the real cause. It is the <cuda_runtime.h> included in our cuda
wrapper code (not cusp) that triggered the ICC version checking in
host_config.h. In contrast, pmemd.cuda doesn't use this header, thus
won't trigger the runtime compiler checking. A quick "fix" is just
removing the <cuda_runtime.h> inclusion in our code as long as we use
nvcc to compile.

Two concerns:

1) As shown by those OS/Compiler/CUDA compatibility matrices from
Nvidia, allowing the use of the compiler not officially supported
(tested) by CUDA is of risk. I don't know to what extent this effect
will be, maybe not a big deal?

2) For now, we can fix this compiler error either through a) Add a logic
checking on the icc version to configure as suggested previously, or b)
I remove the <cuda_rutime.h> inclusion in our code - this will need
another round combination tests. Also just mention again, user manually
commenting out the #error line in host_config.h will also be a
workaround (not recommended?).

Thanks!

Ruxi



On 04/06/2017 07:55 PM, Daniel Roe wrote:
> Hi,
>
> On Thu, Apr 6, 2017 at 6:59 PM, Ruxi Qi <ruxiq.uci.edu> wrote:
>> 1. I once encountered the same error. Indeed, this error is not because
>> of linking CUSP library. It's nvcc that complains the outdated ICC
>> version. Per the Nvidia official documents, CUDA 7.5 only supports ICC
>> 15.0.0, as shown bellow:
> This is not really the case, at least for Intel compilers version 14.
> I was just able to compile pmemd.cuda from Amber 16 and
> cpptraj/pmemd.cuda from amber17-with-patches with Intel V 14 compilers
> and CUDA 7.5. I think this issue is really CUSP-related since as Brent
> stated he was previously able to compile pmemd.cuda from Amber 16 with
> Intel V12 compilers and CUDA 7.5.
>
>> 3. Adding a --no-pbsa.cuda option is good for safety's sake in this
>> release. I agree with Hai that we may put it built as default, since it
>> generally follows the Nvidia CUDA installation rule.
> Probably the best solution would be to add logic to 'configure' that
> will disable pbsa.cuda if the Intel compiler version is not high
> enough. May need to do the same thing for GNU compilers if there is a
> similar issue.
>
> -Dan
>


_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Fri Apr 07 2017 - 00:00:02 PDT
Custom Search