On Fri, 5 Apr 2019 09:33:55 -0700 Alyssa Rosenzweig <aly...@rosenzweig.io> wrote:
> > Since one of the primary use cases is to draw pretty graphs of the > > system load [1], this "per-job" information isn't all that relevant (and > > minimal performance overhead is important). And if you want to monitor > > just one application it is usually easiest to ensure that it is the only > > thing running. > > Ah-ha, gotch. I don't know why I didn't put 2 and 2 together, but that > definitely makes sense, yeah :) Boris, thoughts? Nothing to add, except, I had the gut feeling I was doing the wrong choice here, hence the mention to this design decision in my cover letter :-). I'll rework the implementation to have perfmons apply globally (instead of being attached to jobs) and get rid of the perfcnt fence. Note that we could support both per-job and global perfmons and avoid the perf penalty when only global perfmons are used, but I don't think it's worth the extra complexity (not to mention that it makes things even more confusing for userspace users). _______________________________________________ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
