From: Stephane E. <er...@hp...> - 2003-10-15 00:42:51
|
John, On Wed, Oct 15, 2003 at 12:55:51AM +0100, John Levon wrote: > On Tue, Oct 14, 2003 at 04:28:55PM -0700, John Levon wrote: > > > Currently we work against pfmlib v3 (not yet released) but I think we > have to drop that dependency - instead of separating out unit mask vs. > event, pfmlib has a big list of all possible combinations. It seems > broken for the bitmask-style events anyway. > Using a big list is more convenient, and the user can quickly identify an event just by looking at the name. But again, you are free to use whather support library you want. > Also pfmlib only seems to support user vs. kernel counting across all > the counters; I forget if ia64 has per-counter support for that, but I > think that it does. > It is possible to have per-event priv levels with libpfm. There is a global setting used when nothing was specified for a specific event: pfmlib_input_param_t.pfp_events[i].plm > It shouldn't be too much work to drop the dependency - basically we have > to encode the value ourselves like the other arches, but in userspace. > We already have the counter allocation stuff anyway. Keep in mind that Itanium 2 processors have about 400 events. Of course, not all of them may matter for the kind of measurements you are doing with your user level tools. In addition, there are some very complex sets of constraints between events that you would have to take care of. -- -Stephane |