[gmx-users] Problem with CUDA

Szilárd Páll pall.szilard at gmail.com
Fri Apr 6 21:40:27 CEST 2018 

On CSCS Piz Daint I use the following command line (assuming
PrgEnv-gnu) where everything in "[]" is optional and compilation
should work just fine without.

CC=cc CXX=CC cmake ../
-DGMX_SIMD=THE_RIGHT_SIMD_FLAVOR
-DGMX_MPI=ON
-DGMX_GPU=ON
-DCMAKE_PREFIX_PATH=${FFTW_DIR}/.. \
[ -DGMX_FFT_LIBRARY=fftw3
-DGMX_CUDA_TARGET_SM=60
-DGMX_PREFER_STATIC_LIBS=ON
-DBUILD_SHARED_LIBS=OFF
-DGMX_BUILD_MDRUN_ONLY=ON
-DGMX_EXTERNAL_BLAS=OFF -DGMX_EXTERNAL_LAPACK=OFF ]

In fact, other than the "-DCMAKE_PREFIX_PATH=${FFTW_DIR}/.."and
setting the right compiler wrappers, the rest is most of the time
unnecessary for a "vanilla" build.

Cheers,
--
Szilárd


On Fri, Apr 6, 2018 at 8:32 PM, Borchert, Christopher B
ERDC-RDE-ITL-MS Contractor <Christopher.B.Borchert at erdc.dren.mil>
wrote:
> You are trying to give me a hint. :) My cmake args are taken from a co-worker, and the statement syntax is from the CMakeCache.txt file. On a Cray you force cc/CC and all the module libraries/headers should be automatically found. Strangely it didn’t find fftw without help. Regardless I get the fpic error. But you've given me a path to investigate. Thanks.
>
> cmake ..
> -DGMX_GPU=ON
> -DCMAKE_C_COMPILER:FILEPATH=`which cc`
> -DCMAKE_CXX_COMPILER:FILEPATH=`which CC`
> -DGMX_FFT_LIBRARY=fftw3
> -DFFTWF_LIBRARY:FILEPATH=${FFTW_DIR}/libfftw3f.so
> -DFFTWF_INCLUDE_DIR:PATH=$FFTW_INC
>
> /usr/bin/ld: CMakeFiles/libgromacs.dir/mdlib/nbnxn_cuda/libgromacs_generated_nbnxn_cuda.cu.o: relocation R_X86_64_32 against `_Z58nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_prune_cuda11cu_atomdata10cu_nbparam8cu_plistb' can not be used when making a shared object; recompile with -fPIC
>
> Chris
>
> -----Original Message-----
> From: gromacs.org_gmx-users-bounces at maillist.sys.kth.se [mailto:gromacs.org_gmx-users-bounces at maillist.sys.kth.se] On Behalf Of Szilárd Páll
> Sent: Friday, April 06, 2018 12:05 PM
> To: Discussion list for GROMACS users <gmx-users at gromacs.org>
> Cc: gromacs.org_gmx-users at maillist.sys.kth.se
> Subject: Re: [gmx-users] Problem with CUDA
>
> FYI: not even a my vanilla (non-CRAY) local build which does work otherwise succeeds with cmake . -DCUDA_NVCC_FLAGS:STRING="-rdc=true"
> so as I guessed, that's the culprit.
>
> Out of curiosity I wonder: what's the reason for the "inventive" use of CMake options, none of which are needed?
>
> --
> Szilárd
>
>
> On Fri, Apr 6, 2018 at 6:57 PM, Szilárd Páll <pall.szilard at gmail.com> wrote:
>> I think the fpic errors can't be caused by missing rdc=true because
>> the latter refers to the GPU _device_ code, but GROMACS does not need
>> relocatable device code, so that should not be necessary.
>> --
>> Szilárd
>>
>>
>> On Fri, Apr 6, 2018 at 6:33 PM, Borchert, Christopher B
>> ERDC-RDE-ITL-MS Contractor <Christopher.B.Borchert at erdc.dren.mil>
>> wrote:
>>> Thanks Szilárd. My understanding is rdc is nvcc's equivalent of fpic. I get fpic errors without it. In fact I get fpic errors without including fpic explicitly in the C/CXX flags.
>>>
>>> /usr/bin/ld:
>>> CMakeFiles/libgromacs.dir/mdlib/nbnxn_cuda/libgromacs_generated_nbnxn
>>> _cuda.cu.o: relocation R_X86_64_32 against
>>> `_Z58nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_prune_cuda11cu_a
>>> tomdata10cu_nbparam8cu_plistb' can not be used when making a shared
>>> object; recompile with -fPIC
>>>
>>> So I removed boost, avx2, mpi, and dynamic but get the same result. What else should I remove?
>>>
>>> cmake ..
>>> -DCMAKE_VERBOSE_MAKEFILE:BOOL=TRUE
>>> -DCMAKE_C_COMPILER:FILEPATH=`which cc` -DCMAKE_C_FLAGS:STRING=-fPIC
>>> -DCMAKE_CXX_COMPILER:FILEPATH=`which CC`
>>> -DCMAKE_CXX_FLAGS:STRING=-fPIC -DCMAKE_INSTALL_PREFIX:PATH=$PREFIX
>>> -DGMX_FFT_LIBRARY=fftw3
>>> -DFFTWF_LIBRARY:FILEPATH=${FFTW_DIR}/libfftw3f.so
>>> -DFFTWF_INCLUDE_DIR:PATH=$FFTW_INC
>>> -DGMX_GPU=ON
>>> -DCUDA_NVCC_FLAGS:STRING="-rdc=true"
>>>
>>> /opt/cray/pe/craype/2.5.13/bin/CC    -march=core-avx2   -fPIC -std=c++11   -O3 -DNDEBUG -funroll-all-loops -fexcess-precision=fast       CMakeFiles/template.dir/template.cpp.o  -o ../../bin/template -Wl,-rpath,/p/work/borchert/gromacs-2018.1/build/lib ../../lib/libgromacs.so.3.1.0 -fopenmp -lm
>>> ../../lib/libgromacs.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_59_tmpxft_000001e3_00000000_21_pmalloc_cuda_compute_61_cpp1_ii_63d60154'
>>> ../../lib/libgromacs.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_57_tmpxft_0000a64f_00000000_21_pme_spread_compute_61_cpp1_ii_d982d3ad'
>>> ../../lib/libgromacs.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_71_tmpxft_000003a4_00000000_21_cuda_version_information_compute_61_cpp1_ii_8ab8dc1d'
>>> ../../lib/libgromacs.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_58_tmpxft_0000a80b_00000000_21_pme_timings_compute_61_cpp1_ii_75ae0e44'
>>> ../../lib/libgromacs.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_56_tmpxft_0000a10b_00000000_21_pme_3dfft_compute_61_cpp1_ii_79dff388'
>>> ../../lib/libgromacs.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_57_tmpxft_00009bd7_00000000_21_nbnxn_cuda_compute_61_cpp1_ii_f147f02c'
>>> ../../lib/libgromacs.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_50_tmpxft_0000a9c8_00000000_21_pme_compute_61_cpp1_ii_6dbf966c'
>>> ../../lib/libgromacs.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_56_tmpxft_0000a490_00000000_21_pme_solve_compute_61_cpp1_ii_06051a94'
>>> ../../lib/libgromacs.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_56_tmpxft_0000ab85_00000000_21_cudautils_compute_61_cpp1_ii_25933dd5'
>>> ../../lib/libgromacs.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_54_tmpxft_0000aefc_00000000_21_pinning_compute_61_cpp1_ii_5d0f4aae'
>>> ../../lib/libgromacs.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_56_tmpxft_0000ad42_00000000_21_gpu_utils_compute_61_cpp1_ii_70828085'
>>> ../../lib/libgromacs.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_57_tmpxft_0000a2d0_00000000_21_pme_gather_compute_61_cpp1_ii_a7a2f9c7'
>>> ../../lib/libgromacs.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_67_tmpxft_00009f4e_00000000_21_nbnxn_cuda_data_mgmt_compute_61_cpp1_ii_a1eafeba'
>>>
>>> Chris
>>>
>>> -----Original Message-----
>>> From: gromacs.org_gmx-users-bounces at maillist.sys.kth.se
>>> [mailto:gromacs.org_gmx-users-bounces at maillist.sys.kth.se] On Behalf
>>> Of Szilárd Páll
>>> Sent: Friday, April 06, 2018 10:17 AM
>>> To: Discussion list for GROMACS users <gmx-users at gromacs.org>
>>> Cc: gromacs.org_gmx-users at maillist.sys.kth.se
>>> Subject: Re: [gmx-users] Problem with CUDA
>>>
>>> Hi,
>>>
>>> What is the reason for using the custom CMake options? What's the -rdc=true for -- I don't think it's needed and it can very well be causing the issue. Have you tried to actually do an as-vanilla-as-possible build?
>>>
>>> --
>>> Szilárd
>>>
>>>
>>> On Thu, Apr 5, 2018 at 6:52 PM, Borchert, Christopher B
>>> ERDC-RDE-ITL-MS Contractor <Christopher.B.Borchert at erdc.dren.mil>
>>> wrote:
>>>> Hello. I'm taking a working build from a co-worker and trying to add GPU support on a Cray XC. CMake works but make fails. Both 2016 and 2018 die at the same point -- can't find gromac's own routines.
>>>>
>>>> 2016.5:
>>>> /opt/cray/pe/craype/2.5.13/bin/CC    -march=core-avx2   -O2 -fPIC -dynamic -std=c++0x   -O3 -DNDEBUG -funroll-all-loops -fexcess-precision=fast      -dynamic CMakeFiles/template.dir/template.cpp.o  -o ../../bin/template -Wl,-rpath,/p/work/cots/gromacs-2016.5/build/lib:/opt/nvidia/cudatoolkit8.0/8.0.54_2.3.12_g180d272-2.2/lib64/stubs -dynamic ../../lib/libgromacs_mpi.so.2.5.0 -fopenmp -lcudart /opt/nvidia/cudatoolkit8.0/8.0.54_2.3.12_g180d272-2.2/lib64/stubs/libnvidia-ml.so -lhwloc -lz -ldl -lrt -lm -lfftw3f
>>>> ../../lib/libgromacs_mpi.so.2.5.0: undefined reference to `__cudaRegisterLinkedBinary_59_tmpxft_0001bc78_00000000_21_pmalloc_cuda_compute_61_cpp1_ii_63d60154'
>>>> ../../lib/libgromacs_mpi.so.2.5.0: undefined reference to `__cudaRegisterLinkedBinary_56_tmpxft_0001bac2_00000000_21_gpu_utils_compute_61_cpp1_ii_d70ebee0'
>>>> ../../lib/libgromacs_mpi.so.2.5.0: undefined reference to `__cudaRegisterLinkedBinary_56_tmpxft_0001b90b_00000000_21_cudautils_compute_61_cpp1_ii_24d20763'
>>>> ../../lib/libgromacs_mpi.so.2.5.0: undefined reference to `__cudaRegisterLinkedBinary_71_tmpxft_0001c016_00000000_21_cuda_version_information_compute_61_cpp1_ii_e35285be'
>>>> ../../lib/libgromacs_mpi.so.2.5.0: undefined reference to `__cudaRegisterLinkedBinary_57_tmpxft_0001b592_00000000_21_nbnxn_cuda_compute_61_cpp1_ii_6e47f057'
>>>> ../../lib/libgromacs_mpi.so.2.5.0: undefined reference to `__cudaRegisterLinkedBinary_67_tmpxft_0001b754_00000000_21_nbnxn_cuda_data_mgmt_compute_61_cpp1_ii_a1eafeba'
>>>> collect2: error: ld returned 1 exit status
>>>>
>>>> 2018.1:
>>>> /opt/cray/pe/craype/2.5.13/bin/CC    -march=core-avx2   -O2 -fPIC -dynamic -std=c++11   -O3 -DNDEBUG -funroll-all-loops -fexcess-precision=fast      -dynamic CMakeFiles/template.dir/template.cpp.o  -o ../../bin/template -Wl,-rpath,/p/work/cots/gromacs-2018.1/build/lib ../../lib/libgromacs_mpi.so.3.1.0 -fopenmp -lm
>>>> ../../lib/libgromacs_mpi.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_56_tmpxft_000068a5_00000000_21_pme_3dfft_compute_61_cpp1_ii_79dff388'
>>>> ../../lib/libgromacs_mpi.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_67_tmpxft_00006621_00000000_21_nbnxn_cuda_data_mgmt_compute_61_cpp1_ii_a1eafeba'
>>>> ../../lib/libgromacs_mpi.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_57_tmpxft_00006f47_00000000_21_pme_spread_compute_61_cpp1_ii_d982d3ad'
>>>> ../../lib/libgromacs_mpi.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_56_tmpxft_00006d70_00000000_21_pme_solve_compute_61_cpp1_ii_06051a94'
>>>> ../../lib/libgromacs_mpi.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_59_tmpxft_00008da7_00000000_21_pmalloc_cuda_compute_61_cpp1_ii_63d60154'
>>>> ../../lib/libgromacs_mpi.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_50_tmpxft_00007930_00000000_21_pme_compute_61_cpp1_ii_6dbf966c'
>>>> ../../lib/libgromacs_mpi.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_58_tmpxft_00007382_00000000_21_pme_timings_compute_61_cpp1_ii_75ae0e44'
>>>> ../../lib/libgromacs_mpi.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_57_tmpxft_00006b11_00000000_21_pme_gather_compute_61_cpp1_ii_a7a2f9c7'
>>>> ../../lib/libgromacs_mpi.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_56_tmpxft_00007f9f_00000000_21_cudautils_compute_61_cpp1_ii_25933dd5'
>>>> ../../lib/libgromacs_mpi.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_54_tmpxft_000088f9_00000000_21_pinning_compute_61_cpp1_ii_5d0f4aae'
>>>> ../../lib/libgromacs_mpi.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_57_tmpxft_000039b7_00000000_21_nbnxn_cuda_compute_61_cpp1_ii_f147f02c'
>>>> ../../lib/libgromacs_mpi.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_71_tmpxft_000091d4_00000000_21_cuda_version_information_compute_61_cpp1_ii_8ab8dc1d'
>>>> ../../lib/libgromacs_mpi.so.3.1.0: undefined reference to `__cudaRegisterLinkedBinary_56_tmpxft_00008407_00000000_21_gpu_utils_compute_61_cpp1_ii_70828085'
>>>> collect2: error: ld returned 1 exit status
>>>>
>>>> BUILD INSTRUCTIONS:
>>>> module swap PrgEnv-cray PrgEnv-gnu
>>>> module swap gcc gcc/5.3.0
>>>> export CRAYPE_LINK_TYPE=dynamic
>>>>
>>>> module load cudatoolkit/8.0.54_2.3.12_g180d272-2.2
>>>> module load cmake/gcc-6.3.0/3.7.2
>>>> module load fftw/3.3.4.11
>>>> export BOOST_DIR=/app/unsupported/boost/1.64.0-gcc-6.3.0
>>>>
>>>> export PREFIX=/app/unsupported/gromacs/201x.x
>>>> mkdir $PREFIX
>>>>
>>>> cmake ..  \
>>>>  -DCMAKE_VERBOSE_MAKEFILE:BOOL=TRUE \
>>>> -DCMAKE_C_COMPILER:FILEPATH=`which cc` \
>>>>  -DCMAKE_C_FLAGS:STRING="-O2 -fPIC -dynamic" \
>>>> -DCMAKE_CXX_COMPILER:FILEPATH=`which CC` \
>>>>  -DCMAKE_CXX_FLAGS:STRING="-O2 -fPIC -dynamic" \
>>>> -DCMAKE_INSTALL_PREFIX:PATH=$PREFIX \
>>>>  -DGMX_FFT_LIBRARY=fftw3 \
>>>>  -DCMAKE_EXE_LINKER_FLAGS:STRING=-dynamic \
>>>>  -DGMX_SIMD=AVX2_256 \
>>>>  -DFFTWF_LIBRARY:FILEPATH=${FFTW_DIR}/libfftw3f.so \
>>>> -DFFTWF_INCLUDE_DIR:PATH=$FFTW_INC \  -DBoost_DIR:PATH=$BOOST_DIR \
>>>> -DBoost_INCLUDE_DIR:PATH=${BOOST_DIR}/include \
>>>> -DBoost_LIBRARY_DIR:PATH=${BOOST_DIR}/lib \  -DGMX_MPI=ON \
>>>> -DGMX_GPU=ON \  -DCUDA_NVCC_FLAGS:STRING="-rdc=true" \
>>>> -DCUDA_USE_STATIC_CUDA_RUNTIME:BOOL=OFF
>>>>
>>>> make 2>&1 | tee make.log
>>>>
>>>> Thanks,
>>>> Chris
>>>>
>>>> --
>>>> Gromacs Users mailing list
>>>>
>>>> * Please search the archive at BlockedBlockedhttp://www.gromacs.org/Support/Mailing_Lists/GMX-Users_ListBlockedBlocked before posting!
>>>>
>>>> * Can't post? Read
>>>> BlockedBlockedhttp://www.gromacs.org/Support/Mailing_ListsBlockedBlo
>>>> cked
>>>>
>>>> * For (un)subscribe requests visit
>>>> BlockedBlockedhttps://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-usersBlockedBlocked or send a mail to gmx-users-request at gromacs.org.
>>> --
>>> Gromacs Users mailing list
>>>
>>> * Please search the archive at BlockedBlockedhttp://www.gromacs.org/Support/Mailing_Lists/GMX-Users_ListBlockedBlocked before posting!
>>>
>>> * Can't post? Read
>>> BlockedBlockedhttp://www.gromacs.org/Support/Mailing_ListsBlockedBloc
>>> ked
>>>
>>> * For (un)subscribe requests visit
>>> BlockedBlockedhttps://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-usersBlockedBlocked or send a mail to gmx-users-request at gromacs.org.
>>> --
>>> Gromacs Users mailing list
>>>
>>> * Please search the archive at Blockedhttp://www.gromacs.org/Support/Mailing_Lists/GMX-Users_ListBlocked before posting!
>>>
>>> * Can't post? Read
>>> Blockedhttp://www.gromacs.org/Support/Mailing_ListsBlocked
>>>
>>> * For (un)subscribe requests visit
>>> Blockedhttps://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-usersBlocked or send a mail to gmx-users-request at gromacs.org.
> --
> Gromacs Users mailing list
>
> * Please search the archive at Blockedhttp://www.gromacs.org/Support/Mailing_Lists/GMX-Users_ListBlocked before posting!
>
> * Can't post? Read Blockedhttp://www.gromacs.org/Support/Mailing_ListsBlocked
>
> * For (un)subscribe requests visit
> Blockedhttps://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-usersBlocked or send a mail to gmx-users-request at gromacs.org.
> --
> Gromacs Users mailing list
>
> * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-request at gromacs.org.

More information about the gromacs.org_gmx-users mailing list