[gmx-users] Build problem with 4.6beta1

[Justin Lemkul]

On 11/29/12 10:25 PM, Szilárd Páll wrote:
> On Fri, Nov 30, 2012 at 3:20 AM, Justin Lemkul <jalemkul at vt.edu> wrote:
>
>>
>> Hooray for being the first to report a problem with the beta :)
>>
>> We have a cluster at our university that provides us with access to some
>> CPU-only nodes and some CPU-GPU nodes.  I'm having problems with getting
>> 4.6beta1 to build, and I suspect the issue is related to GPU detection.
>>
>> Here are some specifics:
>>
>> FFTW 3.3.3
>> CMake 2.8.10
>> gcc 4.3.4
>> CUDA 4.0
>> 64-bit Linux on AMD hardware
>> GPU nodes have Tesla C2050 cards
>>
>
> Note that on AMD you might have to use less threads per process than the
> default and therefore more than one process per GPU: set -gpu_id=00 for two
> processes in a node to both use the first GPU. Additionally, you might want
> to use a newer gcc and CUDA, both will provide performance improvements.
>

Thanks for the tips.  I may be able to get the sysadmins to install a newer gcc, 
but I've already been told that CUDA can't be upgraded past 4.0, for a variety 
of technical reasons that I quite honestly don't understand.

>
>>
>> Commands:
>>
>> cmake ../gromacs-4.6-beta1 -DCMAKE_INSTALL_PREFIX=/home/**
>> jalemkul/software/gromacs-**46beta1 -DGMX_X11=OFF -DGMX_GPU=ON
>> -DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DCMAKE_PREFIX_PATH=/home/
>> **jalemkul/software/fftw-3.3.3/
>>
>> The first step runs alright, but two things to note:
>>
>> 1. It doesn't detect any GPU, which is correct because I'm on a head node
>> and not a compute node:
>>
>> ...-- Looking for NVIDIA GPUs present in the system
>> -- Could not detect NVIDIA GPUs
>> -- Found CUDA: /cm/shared/apps/cuda40/**toolkit/4.0.17 (found suitable
>> version "4.0", minimum required is "3.2")
>> ...
>>
>
> That is just an informative message. We try to detect GPUs on the build
> machine to suggest users compiling with GPU acceleration enabled.
>
>

OK, I just wanted to make sure this wouldn't be an issue.

>>
>> 2. It says FFTW isn't detected, but it actually is:
>>
>> ...
>> -- checking for module 'fftw3f'
>> --   package 'fftw3f' not found
>> -- Looking for fftwf_plan_r2r_1d in /home/jalemkul/ATHENA/**
>> software/fftw-3.3.3/lib/**libfftw3f.so
>> -- Looking for fftwf_plan_r2r_1d in /home/jalemkul/ATHENA/**
>> software/fftw-3.3.3/lib/**libfftw3f.so - found
>> -- Looking for fftwf_have_simd_avx in /home/jalemkul/ATHENA/**
>> software/fftw-3.3.3/lib/**libfftw3f.so
>> -- Looking for fftwf_have_simd_avx in /home/jalemkul/ATHENA/**
>> software/fftw-3.3.3/lib/**libfftw3f.so - not found
>> -- Looking for fftwf_have_simd_sse2 in /home/jalemkul/ATHENA/**
>> software/fftw-3.3.3/lib/**libfftw3f.so
>> -- Looking for fftwf_have_simd_sse2 in /home/jalemkul/ATHENA/**
>> software/fftw-3.3.3/lib/**libfftw3f.so - found
>> ...
>>
>> Upon running "make," I get an immediate failure:
>>
>> [  0%] Building NVCC (Device) object src/gmxlib/gpu_utils/**
>> CMakeFiles/gpu_utils.dir//./**gpu_utils_generated_gpu_utils.**cu.o
>> nvcc fatal   : redefinition of argument 'compiler-bindir'
>> CMake Error at gpu_utils_generated_gpu_utils.**cu.o.cmake:206 (message):
>>    Error generating
>>
>> /home/jalemkul/gmxbuild/src/**gmxlib/gpu_utils/CMakeFiles/**
>> gpu_utils.dir//./gpu_utils_**generated_gpu_utils.cu.o
>>
>>
>> make[2]: *** [src/gmxlib/gpu_utils/**CMakeFiles/gpu_utils.dir/./**
>> gpu_utils_generated_gpu_utils.**cu.o] Error 1
>> make[1]: *** [src/gmxlib/gpu_utils/**CMakeFiles/gpu_utils.dir/all] Error 2
>> make: *** [all] Error 2
>>
>
> That's bad and I don't exactly know why does it happen, it looks like the
> nvcc host compiler gets double-set. Could you put your CMakeCache.txt of
> this configuration into pastbin and link it (attachments are blocked, I
> think)?
>
> One option to get it work is to remove the "-ccbin=/path/to/your/gcc" form
> the CUDA_NVCC_FLAGS cache variable and re-configure.
>

I can try to play around with it a little more later today, though it seems like 
simply using a pre-2.8.10 version of CMake does the trick.

-Justin

-- 
========================================

Justin A. Lemkul, Ph.D.
Research Scientist
Department of Biochemistry
Virginia Tech
Blacksburg, VA
jalemkul[at]vt.edu | (540) 231-9080
http://www.bevanlab.biochem.vt.edu/Pages/Personal/justin

========================================