Add -b, which reads an event specification as a brace list {a,b,c} and
allocates the events as one PMU group. The hardware then counts the
members together, so the counts of the members relate to the same
interval.
Read a group with pmc_group_read instead of reading each member with
pmc_read. A per-member read returns values from different intervals,
which defeats the reason to group the events.
Scale the count of a multiplexed group by the ratio of enabled time to
running time, and cap the ratio at PMC_SCALE_MAX so that a group that
ran for a very short time does not report an absurd estimate. Print
"<not enabled>" or "<not counted>" instead of a scaled value when the
group did not run at all, so that a dead interval is visible rather
than reported as zero. Print the residency of each group next to its
counts.
Co-authored-by: Raghavendra K T <raghavendra.kt@amd.com>
Sponsored by: AMD