Hi,
I'd also like to say that I *strongly* recommend GNU compilers for use
with the CUDA code. I have seen issues crop up with Intel+CUDA that do
not happen with GNU+CUDA. Specifically, when performing GPU validation
of pmemd.cuda via multiple long-time (~24 hour) runs on GPUs I have
gotten 2 different final energy values with Intel compilers (i.e. out
of 10 runs some will end up with one value and some with another one),
but the same exact value each time with GNU compilers. This is with
the same exact hardware, Amber version, etc - only difference is
compilers. I suspect it has something to do with overly-aggressive
optimization on the Intel side but haven't dug too deeply.
-Dan
PS - IMHO Intel compilers are still better for CPU-only runs.
On Wed, Jul 5, 2017 at 9:27 AM, Hector A. Baldoni <hbaldoni.unsl.edu.ar> wrote:
> Dear FabrÃcio,
>
> Well ... there is a trick, but apply it by your own risk. Also check all
> the test are passed:
>
> At the line:
> #if (__ICC != 1500 && __ICC != 1600) || !defined(__GNUC__) ||
> !defined(__LP64__)
>
> in the file:
> /usr/local/cuda-X.X/include/host_config.h
>
> change it by:
> #if (__ICC != 1600 && __ICC != 1700) || !defined(__GNUC__) ||
> !defined(__LP64__)
>
> ... and happy hack !!!
>
> Regards,
> Hector.
>
>
>> On Wed, Jul 05, 2017, FabrÃcio Bracht wrote:
>>
>>> Hello, I've installed the intel parallel suite of compilers on my
>>> machine
>>> and tried to compile amber but the CUDA installation pops an error at
>>> the
>>> end of the make install command.
>>>
>>> #error -- unsupported ICC configuration! Only ICC 15.0 and ICC 16.0 on
>>> Linux x86_64 are supported!
>>>
>>> The icc version I am using is 17.0.4
>>> Is there a fix to it? (other than using icc 16)
>>
>> There probably is some fix, but there is little point in using icc for a
>> CUDA
>> compile. Do you have the GNU compiler collection available? That is what
>> is
>> recommended. (All the real work in pmemd.cuda is done on the GPU, so
>> there is
>> no advantage in using a compiler optimized for Intel CPU's, vs a "generic"
>> compiler like gcc.)
>>
>> ....dac
>>
>> p.s.: going out beyond my comfort level here, but i think the error
>> reported above is generated by nvidia code, not by Amber.
>>
>>
>> _______________________________________________
>> AMBER mailing list
>> AMBER.ambermd.org
>> http://lists.ambermd.org/mailman/listinfo/amber
>>
>
>
> --------------------------------------
> Dr. Hector A. Baldoni
> Profesor Adjunto (FQByF-UNSL)
> Investigador Adjunto (IMASL-CONICET)
> Area de Quimica General e Inorganica
> Universidad Nacional de San Luis
> Chacabuco 917 (D5700BWS)
> San Luis - Argentina
> hbaldoni at unsl dot edu dot ar
> Tel.:+54-(0)266-4520300 ext. 6157
> --------------------------------------
> -.First thinghs first, but not necesarily in that order.-
>
> -.The theoretical approach these days can predict/show almost anything,
> but the ultimate truth lies in the experiment.-
>
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
--
-------------------------
Daniel R. Roe
Laboratory of Computational Biology
National Institutes of Health, NHLBI
5635 Fishers Ln, Rm T900
Rockville MD, 20852
https://www.lobos.nih.gov/lcb
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Jul 05 2017 - 07:00:02 PDT
