search

QMCPACK / miniqmc

QMCPACK miniapp: a simplified real space QMC code for algorithm development, performance portability testing, and computer science experiments
Other
26 stars 35 forks source link

swap cublas batched getrf/getri for cusolver getrf/getrs, drastically… #191

Closed lshulen closed 5 years ago

lshulen commented 5 years ago

… improve initialization on gpus

prckent commented 5 years ago

Post-deadline, should we do this in mainline QMCPACK?

lshulen commented 5 years ago

It is worth looking into. Some of this may be mitigated by the fact that mainline qmcpack may actually be doing batched evaluation. However, for ~5000 electron problems, switching from the batched version to the cusolver version could be the difference between 10 s spent under the initialization timer and ~200 s spent under that timer. It’s a relatively easy switch to make, so it would be worth looking at.

Luke

From: "Paul R. C. Kent" notifications@github.com Reply-To: QMCPACK/miniqmc reply@reply.github.com Date: Wednesday, September 26, 2018 at 2:15 PM To: QMCPACK/miniqmc miniqmc@noreply.github.com Cc: "Shulenburger, Luke" lshulen@sandia.gov, State change state_change@noreply.github.com Subject: [EXTERNAL] Re: [QMCPACK/miniqmc] swap cublas batched getrf/getri for cusolver getrf/getrs, drastically… (#191)

Post-deadline, should we do this in mainline QMCPACK?

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHubhttps://github.com/QMCPACK/miniqmc/pull/191#issuecomment-424854499, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AXMsFVBZDRl3HJAElapCA1wA9npZYMtbks5ue-BQgaJpZM4W7Vce.