[CP2K:3828] Re: Compile cp2k with CUDA on mac osx
urban.b... at gmail.com
Mon May 21 09:30:49 CEST 2012
which defines did you use in compiling? Since the error you report is
an out-of-memory error, I assume you used both __CUDAPW and
__DBCSR_CUDA. Unfortunately, these are slightly incompatible withut a
modified input file because one reserves most of the memory for itself,
leaving the other one without any memory. Try setting the
GLOBAL/CUDA/MEMORY (in kiB) to some sane value. I would guess around
20000 and then adjust as necessary from there: if you get a similar out
of memery message the value is too big. If it's too small you will
probably get some other message.
Also note that using multiple cards from one process is unsupported.
You should use 2 MPI processes per box to take advantage of both cards.
On Fri, 2012-05-18 at 08:40 -0700, Samuel Lamphier wrote:
> All, after a number of modifications, cp2k with CUDA support was
> compiled with no errors, however, upon testing I received the
> following memory error:
> CUDA Error: out of memory
> ASSERTION FAILED: 1.EQ. 0
> error in host_mem_alloc_i at line 166 with error type -1
> message: Could not allocate host pinned memory
> 2 error in host_mem_alloc_i at line 166
> 1 called from dbcsr_init_lib
> libdbcsr| Abnormal program termination, stopped by process number 6
> application called MPI_Abort(MPI_COMM_WORLD, 1) - process 6
> 0:Return code = 0, signaled with Interrupt
> 1:Return code = 0, signaled with Interrupt
> 2:Return code = 0, signaled with Interrupt
> 3:Return code = 0, signaled with Interrupt
> 4:Return code = 0, signaled with Interrupt
> 5:Return code = 0, signaled with Interrupt
> 6:Return code = 1
> 7:Return code = 0, signaled with Interrupt
> Does anyone have any idea what would cause this, I have 2 NVIDIA
> quadro 4000 mac pro and 24 GB of ram on the system (MacPro5.1 Snow
> Lepord 10.6.8).
> On May 15, 3:40 am, Urban Borštnik <urban.b... at gmail.com> wrote:
> > Hi,
> > On Mon, 2012-05-14 at 12:00 -0700, Samuel Lamphier wrote:
> > > Hello all,
> > > I am trying to compile cp2k with CUDA on a macpro. I have 2 nvidia
> > > quadro 4000 cards for GPU computation. However, when trying to
> > > compile, the compiler is looking for -lrt which is the Posix real time
> > > library as I understand. However... this is not supported on a mac.
> > > How can I work around this?
> > You can edit the file cuda_tools/dbcsr_cuda_timing to remove calls to
> > the realtime funcions, and then not link to librt.
> > Cheers,
> > Urban.
> > > Thanks in advance
More information about the CP2K-user