[CP2K:1877] Re: Energy is NaN with CUFFT

[Ondrej Marsalek]

Hi again,

> hi,
>
> i don't see this. just recompiled on the tesla machine at uiuc.
>
> as usual, please provide the smallest possible input that reproduces
> it,
> as well as the compiler flags and hardware specs that you were using.
> otherwise it is impossible to track it down.

It happens only in parallel runs. I get the same kind of trouble with
the Ar.inp example (as well as my own inputs) when run in parallel. In
serial, it all seems to work. Can anyone else reproduce this?

I use a single node with two dual core Opterons, OpenMPI 1.3 and the
Intel compiler, version 10.1.018. The arch file used is below.

Best,
Ondrej


NVCC     = nvcc
CC       = mpicc
CPP      =
FC       = mpif90
LD       = mpif90
AR       = ar -r
DFLAGS   = -D__INTEL -D__FFTSG -D__parallel -D__BLACS -D__SCALAPACK
-D__CUDA -D__FFTSGL
CPPFLAGS =
FCFLAGS  = $(DFLAGS) -I$(INCLUDE) -fpp -free -O2 -unroll -pc64
-heap-arrays 64 -march=pentium3 -mtune=core2
NVFLAGS  = $(DFLAGS)
LDFLAGS  = $(FCFLAGS) -static-intel

LIBS    = -L$(LIBRARY_PATH)        \
          -lmkl_scalapack_lp64     \
          -lmkl_blacs_openmpi_lp64 \
          -lmkl_intel_lp64         \
          -lmkl_sequential         \
          -lmkl_core               \
          -L/usr/local/cuda/lib    \
          -lcudart                 \
          -lcufft                  \
          -lcublas                 \
          -lstdc++

OBJECTS_ARCHITECTURE = machine_intel.o