Stephane - > -----Original Message----- > From: Stephane Eranian [mailto:eran...@google.com] > Sent: Friday, January 15, 2010 2:42 AM > To: Dan Terpstra > Cc: perfmon2-devel@lists.sourceforge.net; jag...@eecs.utk.edu > Subject: Re: Nehalem and PFMLIB_NHM_UMASK_NCOMBO > > On Fri, Jan 15, 2010 at 2:38 AM, Dan Terpstra <terps...@eecs.utk.edu> > wrote: > > Yes. It actually did count without error and produce a result of 0. > > We didn't exactly track the error handling to see if an error was > returned > > that PAPI didn't process (always possible :)), but no obvious errors > > appeared. Our original assumption was that the second unit mask, > > SSE_SINGLE_PRECISION, was simply overwriting the first and causing a > silent > > error. > > Is that with libpfm4 or libpfm4? > libpfm3
> > The question is, if these masks *can* be combined for FP_COMP_OPS_EXE, > what > > other events can combine masks? And are the newly imposed restrictions > > actually overly cautious? > > The obvious case where the restrictions is necessary is when unit masks > values > overlap, e.g., 0x5 and 0x4. But for all others, it is far less clear > and I believe it may > have as to do with the way this is actually implemented in HW. > So would the better triage solution be to roll back to the previous version and (re)expose the possibility of illegal combinations or just change the flag on COMP_OPS_EXE to enable FP counting combinations and run the risk of silent failures for other potentially useful combinations? - d > >> -----Original Message----- > >> From: Stephane Eranian [mailto:eran...@google.com] > >> Sent: Thursday, January 14, 2010 5:18 PM > >> To: Dan Terpstra > >> Cc: perfmon2-devel@lists.sourceforge.net; jag...@eecs.utk.edu > >> Subject: Re: Nehalem and PFMLIB_NHM_UMASK_NCOMBO > >> > >> Dan, > >> > >> I made that change recently because I learned that on NHM, not all unit > >> masks > >> can be combined. In fact, it seems only unit masks of events counting > >> cycles > >> apparently can. > >> > >> But it may be that FP_COMP is an exception I missed. > >> > >> But something looks strange to me in your results. If you pass 2 unit > >> masks, > >> libpfm should return an error (pfm_dispatch_events or > >> pfm_get_perf_event_encoding). > >> But you say you're getting zero count, which leads me to believe, > >> libpfm did not return > >> an error. > >> > >> On Thu, Jan 14, 2010 at 11:08 PM, Dan Terpstra <terps...@eecs.utk.edu> > >> wrote: > >> > Stephane - > >> > > >> > I imported the latest sources from libpfm the other day in > preparation > >> for > >> > the PAPI-C release. Today our regression tests showed 0 counts for > >> > PAPI_FP_OPS. In tracing this problem we discovered that the latest > >> > intel_corei7_events.h has a whole bunch of new > PFMLIB_NHM_UMASK_NCOMBO > >> flags > >> > set, including on FP_COMP_OPS_EXE. I don't know about the other > >> instances, > >> > but I know that UMASKS can be combined for that event. Our definition > of > >> > PAPI_FP_OPS is: > >> > > >> > FP_COMP_OPS_EXE:SSE_DOUBLE_PRECISION:SSE_SINGLE_PRECISION > >> > > >> > This produces 0 counts. > >> > > >> > If we clear that flag, we get the proper # of counts. > >> > > >> > If we use: FP_COMP_OPS_EXE:SSE_DOUBLE_PRECISION we get the right > answer; > >> > > >> > If we use: FP_COMP_OPS_EXE:SSE_SINGLE_PRECISION we get 0. > >> > > >> > > >> > > >> > I hesitate to change just this one flag without knowing the reasoning > >> behind > >> > all these restrictions. But I can't release PAPI if it doesn't count > >> > floating point. The other concern is that even though we passed in > two > >> > UMASKs to an event that supposedly could only handle one, we didn't > see > >> an > >> > error message. Is that PAPI's fault or libpfm's? > >> > > >> > > >> > > >> > Thanks, > >> > > >> > - dan > >> > > >> > > >> > >> > >> ------------------------------------------------------------------------------ Throughout its 18-year history, RSA Conference consistently attracts the world's best and brightest in the field, creating opportunities for Conference attendees to learn about information security's most important issues through interactions with peers, luminaries and emerging and established companies. http://p.sf.net/sfu/rsaconf-dev2dev _______________________________________________ perfmon2-devel mailing list perfmon2-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/perfmon2-devel
