AIJCUSPARSE does not actually implement MatBindToCPU()
AIJCUSPARSE does not actually implement MatBindToCPU() functionality. (The flag gets set if MatBindToCPU() is called, but this is not actually used for anything!) I noticed this looking at GPU usage statistics in -log_view, and confirmed by stepping through code with a debugger and then inspecting the code. (I wish I had been paying more attention to what @stefanozampini has been doing, as he noted this in commit 4e4bbfaa back in September!) I need to implement this ASAP for some performance studies I am doing.
Heads up to @markadams4: Mark, you had mentioned that binding (formerly termed "pinning") to the CPU did not seem to make any difference for coarse GAMG levels on Summit. I assume you were running with AIJCUSPARSE matrices, and this may be (at least in part) why.
@BarrySmith, should this feature go into 'master' (because it's adding new functionality) or do we consider this a bug that should be fixed in 'maint' because we offer a MatPinToCPU() that seems like it should be doing something but has no practical effect?