Hi Dave and All,
Here are some clarifications regarding this intel compiler error.
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 It's nvcc that complains the outdated ICC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:
And for comparison, CUDA-8.0 only supports ICC 15 & 16, as shown bellow:
2. Many CUDA applications DO work fine with other and even lower ICC
versions. The way to bypass this is to manually comment out the
error-reporting line in the /usr/local/CUDA-7.5/include/host_config.h,
which is a macro check of __ICC. However, this may put the program
running at unknown risk as said by Nvidia folks.
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.
4. Before 3 is available, Brent, could you also try compiling a CUDA-7.5
sample, say in /usr/local/cuda-7.5/samples, with the same environment?
And probably also try compiling Amber16 as well for double check. :)
Thanks!
Ruxi
On 4/6/17 13:45, David A Case wrote:
> On Thu, Apr 06, 2017, Brent Krueger wrote:
>
>> However, I am now unable to build with CUDA. I have CUDA-7.5 and Intel
>> compilers v12.0 (which worked fine for amber16 last year). When building
>> the CUDA code I get an error message suggesting that I need to have Intel
>> compilers v15.0 or newer (full message below). The error looks to me to be
>> coming from the cuda tree itself. Since this is the same cuda with the
>> same Intel I used before, I'm confused by this. Any thoughts about why
>> this has crept up now when it didn't with amber16?
> Why it comes up now is simple: for the first time, AmberTools now has a cuda
> component: the brand-new pbsa.cuda code. This in terms makes use of the
> NVIDIA cusp library (unused by pmemd.cuda), and your error is coming from
> trying to compile the cusp library:
>
>> /usr/local/cuda/bin/nvcc -gencode arch=compute_20,code=sm_20 -gencode
>> arch=compute_30,code=sm_30 -gencode arch=compute_50,code=sm_50 -gencode
>> arch=compute_52,code=sm_52 -gencode arch=compute_53,code=sm_53
>> -use_fast_math -O3 -ccbin icpc -I../cusplibrary-0.6.0-rc -o
>> cuda_cg_wrapper.o -c cuda_cg_wrapper.cu -DCUSP -DPCG -DJacobi -DDIA
> ^^^^^^^^^^^^^^^^^^
>
>> /usr/local/cuda/bin/..//include/host_config.h(79): catastrophic error:
>> #error directive: -- unsupported ICC configuration! Only ICC 15.0 on Linux
>> x86_64 is supported!
>
> Given the pbsa.cuda is so new, and various combinations of compilers
> have not been extensively tested, I think we really need a "--no-pbsa.cuda"
> option to configure2, to just skip building that. It's also a question
> about what the default should be: try to build pbsa.cuda by default, or skip
> it by default?
>
> Comments/suggestions are welcome, especiall from the UCI folks.
>
> ....dac
>
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Thu Apr 06 2017 - 16:00:02 PDT
