3 Support for Intel hybrid events within perf tools.
5 For some Intel platforms, such as AlderLake, which is hybrid platform and
6 it consists of atom cpu and core cpu. Each cpu has dedicated event list.
7 Part of events are available on core cpu, part of events are available
8 on atom cpu and even part of events are available on both.
10 Kernel exports two new cpu pmus via sysfs:
14 The 'cpus' files are created under the directories. For example,
16 cat /sys/devices/cpu_core/cpus
19 cat /sys/devices/cpu_atom/cpus
22 It indicates cpu0-cpu15 are core cpus and cpu16-cpu23 are atom cpus.
24 As before, use perf-list to list the symbolic event.
29 [Fixed Counter: Counts the number of instructions retired. Unit: cpu_atom]
31 [Number of instructions retired. Fixed Counter - architectural event. Unit: cpu_core]
33 The 'Unit: xxx' is added to brief description to indicate which pmu
34 the event is belong to. Same event name but with different pmu can
37 Enable hybrid event with a specific pmu
39 To enable a core only event or atom only event, following syntax is supported:
41 cpu_core/<event name>/
43 cpu_atom/<event name>/
45 For example, count the 'cycles' event on core cpus.
47 perf stat -e cpu_core/cycles/
49 Create two events for one hardware event automatically
51 When creating one event and the event is available on both atom and core,
52 two events are created automatically. One is for atom, the other is for
53 core. Most of hardware events and cache events are available on both
54 cpu_core and cpu_atom.
56 For hardware events, they have pre-defined configs (e.g. 0 for cycles).
57 But on hybrid platform, kernel needs to know where the event comes from
58 (from atom or from core). The original perf event type PERF_TYPE_HARDWARE
59 can't carry pmu information. So now this type is extended to be PMU aware
60 type. The PMU type ID is stored at attr.config[63:32].
62 PMU type ID is retrieved from sysfs.
63 /sys/devices/cpu_atom/type
64 /sys/devices/cpu_core/type
66 The new attr.config layout for PERF_TYPE_HARDWARE:
68 PERF_TYPE_HARDWARE: 0xEEEEEEEE000000AA
72 Cache event is similar. The type PERF_TYPE_HW_CACHE is extended to be
73 PMU aware type. The PMU type ID is stored at attr.config[63:32].
75 The new attr.config layout for PERF_TYPE_HW_CACHE:
77 PERF_TYPE_HW_CACHE: 0xEEEEEEEE00DDCCBB
79 CC: hardware cache op ID
80 DD: hardware cache op result ID
83 When enabling a hardware event without specified pmu, such as,
84 perf stat -e cycles -a (use system-wide in this example), two events
85 are created automatically.
87 ------------------------------------------------------------
91 sample_type IDENTIFIER
92 read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
96 ------------------------------------------------------------
100 ------------------------------------------------------------
104 sample_type IDENTIFIER
105 read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
109 ------------------------------------------------------------
111 type 0 is PERF_TYPE_HARDWARE.
112 0x4 in 0x400000000 indicates it's cpu_core pmu.
113 0x8 in 0x800000000 indicates it's cpu_atom pmu (atom pmu type id is random).
115 The kernel creates 'cycles' (0x400000000) on cpu0-cpu15 (core cpus),
116 and create 'cycles' (0x800000000) on cpu16-cpu23 (atom cpus).
118 For perf-stat result, it displays two events:
120 Performance counter stats for 'system wide':
122 6,744,979 cpu_core/cycles/
123 1,965,552 cpu_atom/cycles/
125 The first 'cycles' is core event, the second 'cycles' is atom event.
129 perf-stat reports the scaled counts for hybrid event and with a percentage
130 displayed. The percentage is the event's running time/enabling time.
132 One example, 'triad_loop' runs on cpu16 (atom core), while we can see the
133 scaled value for core cycles is 160,444,092 and the percentage is 0.47%.
135 perf stat -e cycles \-- taskset -c 16 ./triad_loop
137 As previous, two events are created.
139 ------------------------------------------------------------
143 sample_type IDENTIFIER
144 read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
149 ------------------------------------------------------------
153 ------------------------------------------------------------
157 sample_type IDENTIFIER
158 read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
163 ------------------------------------------------------------
165 Performance counter stats for 'taskset -c 16 ./triad_loop':
167 233,066,666 cpu_core/cycles/ (0.43%)
168 604,097,080 cpu_atom/cycles/ (99.57%)
172 If there is no '-e' specified in perf record, on hybrid platform,
173 it creates two default 'cycles' and adds them to event list. One
174 is for core, the other is for atom.
178 If there is no '-e' specified in perf stat, on hybrid platform,
179 besides of software events, following events are created and
180 added to event list in order.
184 cpu_core/instructions/,
185 cpu_atom/instructions/,
188 cpu_core/branch-misses/,
189 cpu_atom/branch-misses/
191 Of course, both perf-stat and perf-record support to enable
192 hybrid event with a specific pmu.
195 perf stat -e cpu_core/cycles/
196 perf stat -e cpu_atom/cycles/
197 perf stat -e cpu_core/r1a/
198 perf stat -e cpu_atom/L1-icache-loads/
199 perf stat -e cpu_core/cycles/,cpu_atom/instructions/
200 perf stat -e '{cpu_core/cycles/,cpu_core/instructions/}'
202 But '{cpu_core/cycles/,cpu_atom/instructions/}' will return
203 warning and disable grouping, because the pmus in group are
204 not matched (cpu_core vs. cpu_atom).