Closed rbabich closed 11 years ago
Ok, I have reproduced this now. Ron, I think you introduced the bug when you altered the makefile to always enable texture objects on Kepler. You have introduced a "USE_TEXTURE_OBJECTS" macro, but haven't updated the library to use this uniformly.
Thanks for fixing that, but I'm still getting this error with the latest master (MPI build).
Ron, what type of system are you testing on?
This is on blast (GTX 480, Harpertown Xeon). I guess you're not seeing it?
Nope, I've never repo-ed this problem. I did have a problem with the macro definition, but I didn't get this partition error. Can you paste your configure params?
I just noticed that I have Dslash profiling turned on. Maybe that's it.
./configure \
--enable-cpu-arch=x86_64 \
--enable-gpu-arch=sm_20 \
--disable-twisted-mass-dirac \
--disable-domain-wall-dirac \
--with-cuda=/usr/local/cuda \
--enable-dslash-profiling \
--enable-multi-gpu \
--with-mpi=/usr/lib64/openmpi/1.4-gcc
Can you try without the profiling enabled then? I was actually considering removing the profiling since nvvp does such a good job now, and it gets in the way of the rewrite required to use GPU Direct RDMA / P2P.
That fixed it. Now to figure out why...
We could just delete the profiling and call it a day.
No objections here.
I'm closing this issue with commit 074a130119fc2af953a9b1e0f5936d6dc5b9f851. The underlying cause is still there, but this is addressed with #95.
I'm guessing that this was introduced by Mike's comms merge. With an MPI build, running "dslash_test --partition 8" gives