Chameleon build fails when FxT is enabled
Hello,
When building the master of Chameleon:
cmake .. -DCHAMELEON_USE_MPI=ON -DCHAMELEON_ENABLE_EXAMPLE=OFF -DCHAMELEON_ENABLE_TESTING=ON -DBLA_VENDOR=Intel10_64lp_seq -DCHAMELEON_USE_MPI_DATATYPES=ON
with StarPU built with FxT:
../configure --prefix=/home/philippe/dev/builds/starpu --disable-opencl --disable-cuda --disable-mlr --enable-blas-lib=mkl --with-mkl-cflags=-I/usr/include/mkl --with-mkl-ldflags="-lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lpthread -lm -ldl" --enable-nmad --with-fxt
(I guess nmad or OpenMPI makes no difference here), I get warnings during Chameleon compilation:
[ 48%] Building C object runtime/starpu/CMakeFiles/chameleon_starpu.dir/control/runtime_context.c.o
In file included from /home/philippe/dev/chameleon/runtime/starpu/control/runtime_async.c:21:
/home/philippe/dev/chameleon/build/runtime/starpu/include/chameleon_starpu.h:104: warning: "starpu_insert_task" redefined
#define starpu_insert_task starpu_mpi_insert_task
In file included from /home/philippe/dev/builds/starpu/include/starpu/1.3/starpu.h:66,
from /home/philippe/dev/builds/starpu/include/starpu/1.3/starpu_mpi.h:20,
from /home/philippe/dev/chameleon/build/runtime/starpu/include/chameleon_starpu.h:48,
from /home/philippe/dev/chameleon/runtime/starpu/control/runtime_async.c:21:
/home/philippe/dev/builds/starpu/include/starpu/1.3/starpu_task_util.h:424: note: this is the location of the previous definition
#define starpu_insert_task(cl, ...) \
[... and a lot of other warnings of the same kind ...]
Then during the execution:
mpirun -np 1 -DSTARPU_FXT_TRACE=0 -DSTARPU_NCPU=2 ./testing/chameleon_stesting -o potrf --mtxfmt=1 -H
a communicator is NULL / corrupted.
The commit responsible for this in StarPU is 18b4a1e0 and Chameleon redifines the involved function (https://gitlab.inria.fr/solverstack/chameleon/-/blob/master/runtime/starpu/include/chameleon_starpu.h.in#L104).
Unfortunately, I have no idea how to fix it.
CC @faverge @fpruvost (maybe the bug has to be fixed in Chameleon)
If you try to reproduce the bug, have also in mind this bug solverstack/chameleon#99 (closed), please. :)
Edited by Philippe SWARTVAGHER