search

lattice / quda

QUDA is a library for performing calculations in lattice QCD on GPUs.
https://lattice.github.io/quda
Other
286 stars 94 forks source link

assert failed destNode != QMP_comm_get_node_number(comm) #75

Closed chrisschroeder closed 12 years ago

chrisschroeder commented 12 years ago

Hi, all.

I am trying to run dslash_test for the current (gb2b5915) multi-GPU Wilson dslash and getting the error

assert failed destNode != QMP_comm_get_node_number(comm).

From what I can tell, ala face_qmp.cpp, QUDA is calling QMP_declare_send_relative() for all directions even when I am only using 2 nodes and only have more than 1 node in one direction, while QMP is asserting that this is not done.

Am I misunderstanding? Should I use an older version of QMP (older than qmp-2.4.0-a4)?

Thanks, Chris

rbabich commented 12 years ago

This is a known issue. The assert was added in QMP 2.4.0-a[something], and I think James Osborn is planning to revert it (at our request) in the next release. QMP 2.3.1 should work.

I'm a little surprised you're seeing this, though, unless you're giving dslash_test the "--partition" flag to explicitly turn on comms in directions with no neighbors.

maddyscientist commented 12 years ago

I have not seen this error before. On the other hand, I have not used qmp-2.4.0-a4 before. Please check with 2.3.1 and see if the error remains. From what I recall, 2.4.0-alpha is the work of James Osborn, contains a couple of controversial features, and isn't fully ratified by others who use QMP (I could be wrong on this however).

maddyscientist commented 12 years ago

Moving to 0.4.1 milestone.