search

lattice / quda

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

invertMultiShiftQuda should not abort if the source norm is zero. #400

Open detar opened 8 years ago

detar commented 8 years ago

Aborting is wrong. It should simply return a zero solution.

maddyscientist commented 8 years ago

Under what circumstances should this appear in a normal run?

I would suggest we simply return 0 and give a warning then?

detar commented 8 years ago

Hi Kate,

When we run the spectrum analysis code, we often choose sources with support on only the even sites of a single time slice. The MILC code calls for separate multimass inversions on the even and odd sites. The odd site solve then has a zero rhs. Of course, calling QUDA with such a setup is inefficient, because all the data movement is for naught. The alternative is to compute the source norm in the MILC QUDA wrapper, which is also inefficient, because such cases never occur when we are running HMC, so we shouldn't want to be wasting time checking for rare events.

We have not encountered this problem before, because we have not been running such projects on GPUs until now.

Best, Carleton

On 11/25/2015 1:57 PM, maddyscientist wrote:

Under what circumstances should this appear in a normal run?

I would suggest we simply return 0 and give a warning then?

— Reply to this email directly or view it on GitHub https://github.com/lattice/quda/issues/400#issuecomment-159729981.

Carleton DeTar Department of Physics and Astronomy University of Utah

maddyscientist commented 8 years ago

So in QUDA we could check for zero before copying to the GPU to reduce the unnecessary overhead. We can get this into the develop branch in the next few days.

detar commented 8 years ago

Hi Kate,

Don't make this urgent. I have already hacked the MILC wrapper for now to prevent this, so my jobs can run.

Thanks, Carleton

On 11/25/2015 2:26 PM, maddyscientist wrote:

So in QUDA we could check for zero before copying to the GPU to reduce the unnecessary overhead. We can get this into the develop branch in the next few days.

— Reply to this email directly or view it on GitHub https://github.com/lattice/quda/issues/400#issuecomment-159734868.

Carleton DeTar Department of Physics and Astronomy University of Utah

mathiaswagner commented 7 years ago

Close by #633

maddyscientist commented 7 years ago

Not sure what #633 has to do with this issue? #633 was a fix to a bug I inadvertently introduced a few months ago, where a tolerance of 0 resulted in the solver never existing until the limit of precision is reached.

mathiaswagner commented 7 years ago

Argh, Right. I should get a coffee. I reopened the issue.