WARNING: Number of MPI greater than GPUs / Floating point exception

mirix miro... at gmail.com
Fri Dec 14 11:28:56 CET 2012


UPDATE: If I remove -D__DBCSR_CUDA the warning disappears but not the 
floating point exception...

On Friday, December 14, 2012 11:06:20 AM UTC+1, mirix wrote:
>
> Hello,
>
> I am still struggling to compile CP2K in my laptop both with MPI and CUDA 
> support. 
>
> The laptop has a Core2 Duo processor, 4GB RAM and a Nvidia 8800M GLX card.
>
> The operating system is 64-bit Debian Sid. I am using GCC 4.6 (except for 
> NVCC, in which case I have to use GCC 4.4 because I am using CUDA version 
> 4.0).
>
> I am able to compile the code. Before adding -D__DBCSR_CUDA, -lhwloc and 
> -lstdc++ I was even able to run a few tests successfully. Most of them 
> would not run, though.
>
> After adding those options, the code still compiles but I get the 
> following error:
>
> ---------------------
> mirix at sopranix:~/Software/cp2k/cp2k-2.3/tests/QS$ 
> /home/mirix/Software/cp2k/cp2k-2.3/exe/Linux-x86-64-GNU-OpenMPI-CUDA/cp2k.popt
> WARNING: Number of MPI greater than GPUs
>
> Floating point exception
> ---------------------
>
> This is my arch file:
>
> PERL     = perl
> CC       = mpicc
> CPP      = cpp
> NVCC     = /usr/local/cuda/bin/nvcc
> FC       = mpif90
> LD       = mpif90
> AR       = ar -r
> #DFLAGS   = -D__GFORTRAN -D__CUDA -D__FFTSG -D__FFTSGL -D__FFTW3 
> -D__CUBLASDP
> #DFLAGS   = -D__GFORTRAN -D__CUDA -D__FFTSG -D__FFTSGL -D__FFTW3 
> -D__CUBLASDP -DDGEMM=cuda_d_gemm -Ddgemm=cuda_d_gemm -DDSYMM=gpu_d_symm 
> -Ddsymm=gpu_d_symm
> #DFLAGS   = -D__GFORTRAN -D__FFTSG -D__FFTSGL -D__FFTW3 -D__CUBLASDP 
> -DDGEMM=cuda_d_gemm -Ddgemm=cuda_d_gemm -DDSYMM=gpu_d_symm 
> -Ddsymm=gpu_d_symm -DDSYEVD=cuda_lapack_dsyevd  -Ddsyevd=cuda_lapack_dsyevd
>
> DFLAGSBASE   = -D__GFORTRAN -D__PARALLEL -D__SCALAPACK -D__BLACS -D__FFTW3 
> -D__LIBINT -D__CUDAPW -D__FFTCU -D__FFTSGL -D__FFTSG -D__CUBLASDP 
> -D__DBCSR_CUDA -D__FFTW3 -D__HWLOC
> DFLAGS   = $(DFLAGSBASE) -DDGEMM=cuda_d_gemm -Ddgemm=cuda_d_gemm 
> -DDSYMM=gpu_d_symm -Ddsymm=gpu_d_symm
>
> DFLAGSCU  = $(DFLAGSBASE) -Dcpu_d_gemm=dgemm -Dcpu_d_symm=dsymm
>
> #DFLAGS   = -D__GFORTRAN -D__CUDA -D__FFTCU -D__FFTSGL -D__FFTSG
> #DFLAGSCU = $(DFLAGS)
>
> HWTOPO = yes
> HWLOC = /usr/include/hwloc/
> CFLAGS   =  $(DFLAGS) -I$(HWLOC) -O2 
> CPPFLAGS = -traditional -C $(DFLAGS) -P
> CPPFLAGSCU = -traditional -C $(DFLAGSCU) -P
> NVFLAGS     = $(DFLAGS) -gencode=arch=compute_11,code=\"sm_11,compute_11\" 
> --compiler-bindir=/usr/local/cuda/bin/gcc
> FCFLAGS  = $(DFLAGS) -O2
> FCFLAGSCU = $(DFLAGSCU) -O2
> LDFLAGS  = $(FCFLAGS)
> CUDAPATH = /usr/local/cuda/lib64
> ATLASPATH = /usr/local/atlas/lib
> LIBINTPATH = /usr/local/libint/1.1.4-stable/lib
> LIBS     = -L$(CUDAPATH)/libcudart.so\
> $(CUDAPATH)/libcufft.so\
> $(CUDAPATH)/libcublas.so\
> /home/mirix/Software/cp2k/scalapack-2.0.2/libscalapack.a\
> $(ATLASPATH)/liblapack.a\
> $(ATLASPATH)/libcblas.a\
> $(ATLASPATH)/libf77blas.a\
> $(ATLASPATH)/libptcblas.a\
> $(ATLASPATH)/libptf77blas.a\
> $(ATLASPATH)/libatlas.a\
> $(LIBINTPATH)/libderiv.a\
> $(LIBINTPATH)/libr12.a\
> $(LIBINTPATH)/libint.a\
> /usr/local/lib/libfftw3f.a\
> /usr/lib/x86_64-linux-gnu/libhwloc.so.5.0.1 -lhwloc\
> -lstdc++ -lpthread -lrt
>
> OBJECTS_ARCHITECTURE = machine_gfortran.o
>
> cuda_cublas.o: cuda_cublas.F
>     $(CPP) $(CPPFLAGSCU) $< > cuda_cublas.f90
>     $(FC) -c $(FCFLAGSCU) cuda_cublas.f90
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20121214/db0f76a9/attachment.html>


More information about the CP2K-user mailing list