Skip to contentSkip to conversation

perf x86: Change SandyBridge and IvyBrige instructions event to be precise

PATCH In review Jiri Olsa 3 participants 6 messages Feb 20, 2013

Subsystems (2)
lore
Share & tools
lore

Jiri Olsa

author
#0 lore raw

PATCH 1/1 perf x86: Change SandyBridge and IvyBrige instructions event to be precise

Changing the default 'instructions' HW event for SandyBridge and IvyBrige to use the available precise 'instructions' event.

The precise event is defined in Intel SDM as:

INST_RETIRED.ALL - Precise instruction retired event with HW to reduce effect of PEBS shadow in IP distribution

for both SandyBridge and IvyBrige micro-archs. It allows to use precise modifier on 'instructions' event on SandyBridge and IvyBrige micro-archs which is not possible now.

Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Corey Ashford <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Paul Mackerras <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Andi Kleen <[email protected]>
Cc David Ahern <[email protected]>
Cc Stephane Eranian <[email protected]>
---
 arch/x86/kernel/cpu/perf_event_intel.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/kernel/cpu/perf_event_intel.c +6 -0
index 529c893..3bc3592 100644
@@ -2104,6 +2104,9 @@ __init int intel_pmu_init(void)
 		/* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
 		intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
 			X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1);
+		/* INST_RETIRED.ALL to count precise instructions */
+		intel_perfmon_event_map[PERF_COUNT_HW_INSTRUCTIONS] =
+			X86_CONFIG(.event=0xc0, .umask=0x01);
 
 		pr_cont("SandyBridge events, ");
 		break;
@@ -2127,6 +2130,9 @@ __init int intel_pmu_init(void)
 		/* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
 		intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
 			X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
+		/* INST_RETIRED.ALL to count precise instructions */
+		intel_perfmon_event_map[PERF_COUNT_HW_INSTRUCTIONS] =
+			X86_CONFIG(.event=0xc0, .umask=0x01);
 
 		pr_cont("IvyBridge events, ");
 		break;
Signature · 1 line
1.7.11.7
#1 lore raw
Quoted 11 lines · Jiri Olsa

Changing the default 'instructions' HW event for SandyBridge and IvyBrige to use the available precise 'instructions' event.

The precise event is defined in Intel SDM as:

INST_RETIRED.ALL - Precise instruction retired event with HW to reduce effect of PEBS shadow in IP distribution

for both SandyBridge and IvyBrige micro-archs. It allows to use precise modifier on 'instructions' event on SandyBridge and IvyBrige micro-archs which is not possible now.

Actually it's possible, you just have to use it explicitely.

This will break a lot of profiling setups that assume instructions count on the fixed counter and use the 4/8 other counters. Also it will give very bad results with perf stat, which you don't want to run with pebs.

Even worse with your change the fixed instruction counter is not accessible anymore. So pretty much all the standard setups that collect it in addition to other events and fill all the counters would break.

I don't think it should be done by default, but yes explicitely it's quite useful.

My Haswell patchkit (in the "extended version") exposes the precise event as "instructions-p" through sysfs.

-Andi

Signature · 1 line
[email protected] -- Speaking for myself only
#2 lore raw
Quoted 5 lines · Andi Kleen
Quoted 3 lines · Jiri Olsa

Changing the default 'instructions' HW event for SandyBridge and IvyBrige to use the available precise 'instructions' event.

Quoted 9 lines
Quoted 8 lines

The precise event is defined in Intel SDM as:

INST_RETIRED.ALL - Precise instruction retired event with HW to reduce effect of PEBS shadow in IP distribution

for both SandyBridge and IvyBrige micro-archs. It allows to use precise modifier on 'instructions' event on SandyBridge and IvyBrige micro-archs which is not possible now.

There is also the problem that on SNB, PREC_DIST must be taken ALONE to be correct. By that, I mean no other counter must be active. So I don't think it is a good idea.

Also PREC_DIST is marked as constrained to counter 2 only. That in effect, as Andi pointed out, will make the fixed counter for instruction_retired, unused because you've set a umask now.

I don't think it buys you much. People read too much into what PEBS gives you. I can show you examples where PEBS is worse than no PEBS.

Quoted 22 lines

Actually it's possible, you just have to use it explicitely.

This will break a lot of profiling setups that assume instructions count on the fixed counter and use the 4/8 other counters. Also it will give very bad results with perf stat, which you don't want to run with pebs.

Even worse with your change the fixed instruction counter is not accessible anymore. So pretty much all the standard setups that collect it in addition to other events and fill all the counters would break.

I don't think it should be done by default, but yes explicitely it's quite useful.

My Haswell patchkit (in the "extended version") exposes the precise event as "instructions-p" through sysfs.

-Andi

--
[email protected] -- Speaking for myself only

author
#3 lore raw
Quoted 19 lines · Stephane Eranian
Quoted 5 lines · Andi Kleen
Quoted 3 lines · Jiri Olsa

Changing the default 'instructions' HW event for SandyBridge and IvyBrige to use the available precise 'instructions' event.

Quoted 9 lines
Quoted 8 lines

The precise event is defined in Intel SDM as:

INST_RETIRED.ALL - Precise instruction retired event with HW to reduce effect of PEBS shadow in IP distribution

for both SandyBridge and IvyBrige micro-archs. It allows to use precise modifier on 'instructions' event on SandyBridge and IvyBrige micro-archs which is not possible now.

There is also the problem that on SNB, PREC_DIST must be taken ALONE to be correct. By that, I mean no other counter must be active. So I don't think it is a good idea.

Quoted 7 lines

Also PREC_DIST is marked as constrained to counter 2 only. That in effect, as Andi pointed out, will make the fixed counter for instruction_retired, unused because you've set a umask now.

I don't think it buys you much. People read too much into what PEBS gives you. I can show you examples where PEBS is worse than no PEBS.

well, we sort of advertise we have precise event via precise modifier and it confused me I was not able to use it on IvyBridge..

maybe we could export precise alias in sysfs as well which we could check & use any time there's 'p' modifier specified

not pretty.. meeds more thinking ;-)

thanks,
jirka

#4 lore raw
In reply to Stephane Eranian
Quoted 3 lines

I don't think it buys you much. People read too much into what PEBS gives you. I can show you examples where PEBS is worse than no PEBS.

Nothing is perfect in sampling, but the precise instructions event is definitely better in a range of situations. It avoids a lot of shadow effects.

I think it's a good choice for a lot of situations, although on Ivy Bridge without EventingRip the BB walking is still fairly expensive.

It would be a good default for the "user only asks for instructions for sampling and nothing else" situation, especially on Haswell which has EventingRip. But that's not straight forward to detect.

-Andi

Signature · 1 line
[email protected] -- Speaking for myself only.
author
#5 lore raw
In reply to Andi Kleen
Quoted 14 lines · Andi Kleen
Quoted 11 lines · Jiri Olsa

Changing the default 'instructions' HW event for SandyBridge and IvyBrige to use the available precise 'instructions' event.

The precise event is defined in Intel SDM as:

INST_RETIRED.ALL - Precise instruction retired event with HW to reduce effect of PEBS shadow in IP distribution

for both SandyBridge and IvyBrige micro-archs. It allows to use precise modifier on 'instructions' event on SandyBridge and IvyBrige micro-archs which is not possible now.

Quoted 6 lines

Actually it's possible, you just have to use it explicitely.

This will break a lot of profiling setups that assume instructions count on the fixed counter and use the 4/8 other counters. Also it will give very bad results with perf stat, which you don't want to run with pebs.

right, I haven't realized that..

Quoted 3 lines

Even worse with your change the fixed instruction counter is not accessible anymore. So pretty much all the standard setups

still could be used explicitly via cpu/event=0xc0/

Quoted 8 lines

that collect it in addition to other events and fill all the counters would break.

I don't think it should be done by default, but yes explicitely it's quite useful.

My Haswell patchkit (in the "extended version") exposes the precise event as "instructions-p" through sysfs.

will check

thanks,
jirka