Keita, Make sure you ALWAYS use the flag -cuda_synchronize when you run with -log_summary
Otherwise you get misleading numbers. (Which I am guessing you got). Barry On Aug 27, 2010, at 1:28 PM, Keita Teranishi wrote: > Barry, > > I already see a big difference in MatMult routine of > ksp/ksp/examples/tutorials/ex2.c, and I am very happy to try that example > program. > > Thanks, > ================================ > Keita Teranishi > Scientific Library Group > Cray, Inc. > keita at cray.com > ================================ > > > -----Original Message----- > From: petsc-dev-bounces at mcs.anl.gov [mailto:petsc-dev-bounces at > mcs.anl.gov] On Behalf Of Barry Smith > Sent: Friday, August 27, 2010 1:13 PM > To: For users of the development version of PETSc > Subject: Re: [petsc-dev] Problem with petsc-dev > > > You can run, for example, src/snes/examples/tutorials ex19 with the options > -pc_type jacobi -dmmg_nlevels 5 -da_vec_type cuda -da_mat_type aijcuda and > run without those last two options to NOT use the GPU and compare the results > with -log_summary. We'd be interesting in seeing those numbers also. > > Barry > > On Aug 27, 2010, at 1:07 PM, Keita Teranishi wrote: > >> Satish, >> >> Now I got the latest copy using mercurial. Thanks! >> I am going to check the performance with Fermi. Is there any command line >> option available to swith CUSP? Or do I have to apply MatConvet() with PETSc >> function calls? >> >> Thanks, >> ================================ >> Keita Teranishi >> Scientific Library Group >> Cray, Inc. >> keita at cray.com >> ================================ >> >> >> -----Original Message----- >> From: petsc-dev-bounces at mcs.anl.gov [mailto:petsc-dev-bounces at >> mcs.anl.gov] On Behalf Of Satish Balay >> Sent: Friday, August 27, 2010 11:14 AM >> To: For users of the development version of PETSc >> Subject: Re: [petsc-dev] Problem with petsc-dev >> >> On Fri, 27 Aug 2010, Keita Teranishi wrote: >> >>> Satish, >>> >>> Thanks. I do not see any mercurial package for SUSE, let me try if it >>> works. >> >> source install of mercruial is pretty easy >> >> cd mercurial >> python setup.py install --prefix=/foo/bar >> >> [instructions say - use PYTHONPATH - but I like to hardcode it in 'hg' >> script] i.e edit /foo/bar/bin/hg and add the following at the very >> begining [where mercurial uses a different value of pythonXX - >> depending on the python version on your machine] >> >> import sys >> sys.path.insert(0,'/foo/bar/lib/pythonXX/site-packages/') >> >> Satish >