perf stat patches
Ingo,
These patch set :
fixes perf stat output to handle big names like 'L1-d$-prefetch-misses'
Add support for HARDWARE, SOFTWARE and cache events :
perf stat -e all-sw-events
perf stat -e sw-events
perf stat -e all-hw-events
perf stat -e hw-events
perf stat -e all-cache-events
perf stat -e cache-events
The following changes since commit 18fee47ce1ce511b088ddf65f4e5eb700a9f297a:
Ingo Molnar (1):
Merge branch 'perfcounters/urgent'
are available in the git repository at:
git://git.kernel.org/pub/scm/linux/kernel/git/jaswinder/linux-2.6-tip.git master
Jaswinder Singh Rajput (3):
perf stat: fix stat output
perf_counter tools: Add support to set of multiple events in one shot perf_counter tools: Add support for all CACHE events
tools/perf/builtin-stat.c | 11 ++-- tools/perf/util/parse-events.c | 129 +++++++++++++++++++++++++++++++++++++++- 2 files changed, 132 insertions(+), 8 deletions(-)
Complete diff :
tools/perf/builtin-stat.c +5 -6
@@ -32,6 +32,7 @@* Wu Fengguang <[email protected]> * Mike Galbraith <[email protected]> * Paul Mackerras <[email protected]> +* Jaswinder Singh Rajput <[email protected]> * * Released under the GPL v2. (and only v2, not any later version) */
@@ -250,7 +251,7 @@ static void nsec_printout(int counter, u64 *count, u64 *noise) {double msecs = (double )count[0 ] /1000000 ; -fprintf (stderr," %14.6f %-20s" , msecs,event_name (counter)); +fprintf (stderr," %14.6f %-24s" , msecs,event_name (counter));if (attrs[counter].type ==PERF_TYPE_SOFTWARE && attrs[counter].config ==PERF_COUNT_SW_TASK_CLOCK ) {
@@ -264,7 +265,7 @@ static void nsec_printout(int counter, u64 *count, u64 *noise)static void abs_printout (int counter,u64 *count,u64 *noise) { -fprintf (stderr," %14Ld %-20s" , count[0 ],event_name (counter)); +fprintf (stderr," %14Ld %-24s" , count[0 ],event_name (counter)); if (runtime_cycles_avg && attrs[counter].type == PERF_TYPE_HARDWARE &&
@@ -294,7 +295,7 @@ static void print_counter(int counter) scaled = event_scaled_avg[counter];if (scaled ==-1 ) { -fprintf (stderr," %14s %-20s\n" , +fprintf (stderr," %14s %-24s\n" ,"<not counted>" ,event_name (counter));return ; }
@@ -305,8 +306,7 @@ static void print_counter(int counter)abs_printout (counter, count, noise);if (scaled) -fprintf (stderr," (scaled from %.2f%%)" , - (double ) count[2 ] / count[1 ] *100 ); +fprintf (stderr," (%7.2fx scaled)" , (double )count[1 ]/count[2 ]);fprintf (stderr,"\n" ); }
@@ -417,7 +417,6 @@ static void print_stat(int argc, const char **argv)for (counter =0 ; counter < nr_counters; counter++)print_counter (counter); -fprintf (stderr,"\n" );fprintf (stderr," %14.9f seconds time elapsed.\n" , (double )walltime_nsecs_avg/1e9 );
tools/perf/util/parse-events.c +127 -2
153 lines not on this page · raw message
PATCH 1/3 perf stat: fix stat output
Increase size for event name to handle big names like 'L1-d$-prefetch-misses'
Changed scaled as a multiplicative factor is more expressive, than the percentage of from where we scale up.
Also aligned scaled otherwise sometimes it looks like :
384 iTLB-load-misses (4.74x scaled)
452029 branch-loads (8.00x scaled)
5892 branch-load-misses (20.39x scaled)
972315 iTLB-loads (3.24x scaled)Before :
150708 L1-d$-stores (scaled from 23.57%)
428804 L1-d$-prefetches (scaled from 23.47%)
314446 L1-d$-prefetch-misses (scaled from 23.42%)
252626137 L1-i$-loads (scaled from 23.24%)
5297550 dTLB-load-misses (scaled from 23.96%)
106992392 branch-loads (scaled from 23.67%)
5239561 branch-load-misses (scaled from 23.43%)After :
1731713 L1-d$-loads ( 14.25x scaled)
44241 L1-d$-prefetches ( 3.88x scaled)
21076 L1-d$-prefetch-misses ( 3.40x scaled)
5789421 L1-i$-loads ( 3.78x scaled)
29645 dTLB-load-misses ( 2.95x scaled)
461474 branch-loads ( 6.52x scaled)
7493 branch-load-misses ( 26.57x scaled)--- tools/perf/builtin-stat.c | 11 +++++------ 1 files changed, 5 insertions(+), 6 deletions(-)
tools/perf/builtin-stat.c +5 -6
@@ -32,6 +32,7 @@* Wu Fengguang <[email protected]> * Mike Galbraith <[email protected]> * Paul Mackerras <[email protected]> +* Jaswinder Singh Rajput <[email protected]> * * Released under the GPL v2. (and only v2, not any later version) */
@@ -250,7 +251,7 @@ static void nsec_printout(int counter, u64 *count, u64 *noise) {double msecs = (double )count[0 ] /1000000 ; -fprintf (stderr," %14.6f %-20s" , msecs,event_name (counter)); +fprintf (stderr," %14.6f %-24s" , msecs,event_name (counter));if (attrs[counter].type ==PERF_TYPE_SOFTWARE && attrs[counter].config ==PERF_COUNT_SW_TASK_CLOCK ) {
@@ -264,7 +265,7 @@ static void nsec_printout(int counter, u64 *count, u64 *noise)static void abs_printout (int counter,u64 *count,u64 *noise) { -fprintf (stderr," %14Ld %-20s" , count[0 ],event_name (counter)); +fprintf (stderr," %14Ld %-24s" , count[0 ],event_name (counter)); if (runtime_cycles_avg && attrs[counter].type == PERF_TYPE_HARDWARE &&
@@ -294,7 +295,7 @@ static void print_counter(int counter) scaled = event_scaled_avg[counter];if (scaled ==-1 ) { -fprintf (stderr," %14s %-20s\n" , +fprintf (stderr," %14s %-24s\n" ,"<not counted>" ,event_name (counter));return ; }
@@ -305,8 +306,7 @@ static void print_counter(int counter)abs_printout (counter, count, noise);if (scaled) -fprintf (stderr," (scaled from %.2f%%)" , - (double ) count[2 ] / count[1 ] *100 ); +fprintf (stderr," (%7.2fx scaled)" , (double )count[1 ]/count[2 ]);fprintf (stderr,"\n" ); }
@@ -417,7 +417,6 @@ static void print_stat(int argc, const char **argv)for (counter =0 ; counter < nr_counters; counter++)print_counter (counter); -fprintf (stderr,"\n" );fprintf (stderr," %14.9f seconds time elapsed.\n" , (double )walltime_nsecs_avg/1e9 );
Signature · 1 line
1.6.0.6
PATCH 2/3 perf_counter tools: Add support to set of multiple events in one shot
Add support for HARDWARE and SOFTWARE events :
perf stat -e all-sw-events
perf stat -e sw-events
perf stat -e all-hw-events
perf stat -e hw-events
On AMD box :
$ ./perf stat -e hw-events -e all-sw-events -- ls -lR /usr/include/ > /dev/null
Performance counter stats for 'ls -lR /usr/include/':
744418792 cycles # 2027.230 M/sec ( 3.28x scaled)
515314667 instructions # 0.692 IPC ( 3.29x scaled)
247900772 cache-references # 675.093 M/sec ( 1.18x scaled)
3587971 cache-misses # 9.771 M/sec ( 1.18x scaled)
65830547 branches # 179.272 M/sec ( 1.18x scaled)
3743637 branch-misses # 10.195 M/sec ( 1.18x scaled)
<not counted> bus-cycles
367.880756 cpu-clock-msecs
367.209910 task-clock-msecs # 0.990 CPUs
441 page-faults # 0.001 M/sec
441 minor-faults # 0.001 M/sec
0 major-faults # 0.000 M/sec
41 context-switches # 0.000 M/sec
1 CPU-migrations # 0.000 M/sec0.371065298 seconds time elapsed.
--- tools/perf/util/parse-events.c | 66 ++++++++++++++++++++++++++++++++++++++- 1 files changed, 64 insertions(+), 2 deletions(-)
tools/perf/util/parse-events.c +64 -2
@@ -40,6 +40,16 @@ static struct event_symbol event_symbols[] = { { CSW(CPU_MIGRATIONS), "cpu-migrations", "migrations" }, }; +struct event_type_symbol { +char *symbol; +char *alias; +}; + +static struct event_type_symbol event_type_symbols[] = { + [PERF_TYPE_HARDWARE ] = {"hw-events" ,"all-hw-events" , }, + [PERF_TYPE_SOFTWARE ] = {"sw-events" ,"all-sw-events" , }, +}; +#define __PERF_COUNTER_FIELD (config, name) \ ((config & PERF_COUNTER_##name##_MASK) >> PERF_COUNTER_##name##_SHIFT)
@@ -237,6 +247,49 @@ parse_generic_hw_symbols(const char *str, struct perf_counter_attr *attr)return 0 ; } +static int set_multiple_events (unsigned int type) +{ +struct perf_counter_attr attr; +int i; + +switch (type) { +case PERF_TYPE_HARDWARE : +case PERF_TYPE_SOFTWARE : +for (i =0 ; i <ARRAY_SIZE (event_symbols); i++) { +if (event_symbols[i].type == type) { +memset (&attr,0 ,sizeof (attr)); + attr.type = event_symbols[i].type; + attr.config = event_symbols[i].config; + attrs[nr_counters] = attr; + nr_counters++; + } + } +break ; + +default : +return -1 ; + } + +/* +* parse_events() is assuming that only single event will be set, +* but we are setting multiple events so we need to return magical 1 +*/ +return 1 ; +} + +static int check_type_events (const char *str,unsigned int i) +{ +if (!strncmp (str, event_type_symbols[i].symbol, +strlen (event_type_symbols[i].symbol))) +return 1 ; + +if (strlen (event_type_symbols[i].alias)) +if (!strncmp (str, event_type_symbols[i].alias, +strlen (event_type_symbols[i].alias))) +return 1 ; +return 0 ; +} + static int check_events(const char *str, unsigned int i) { if (!strncmp(str, event_symbols[i].symbol,
@@ -288,6 +341,12 @@ static int parse_event_symbols(const char *str, struct perf_counter_attr *attr)return 0 ; } +for (i =0 ; i <ARRAY_SIZE (event_type_symbols); i++) { +if (check_type_events (str, i)) { +return set_multiple_events (i); + } + } +for (i =0 ; i <ARRAY_SIZE (event_symbols); i++) {if (check_events (str, i)) { attr->type = event_symbols[i].type;
@@ -314,8 +373,11 @@ again:if (ret <0 )return ret; - attrs[nr_counters] = attr; - nr_counters++; +/* No need to set attrs and increment counter when already set */ +if (ret ==0 ) { + attrs[nr_counters] = attr; + nr_counters++; + } str =strstr (str,"," );if (str) {
Signature · 1 line
1.6.0.6
PATCH 3/3 perf_counter tools: Add support for all CACHE events
Add support for all CACHE events :
perf stat -e all-cache-events
perf stat -e cache-events
On AMD box (<not-counted> events are not available for AMD):
$./perf stat -e all-cache-events -- ls -lR /usr/include/ > /dev/null
Performance counter stats for 'ls -lR /usr/include/':
246786934 L1-d$-loads ( 4.64x scaled)
936899 L1-d$-load-misses ( 4.53x scaled)
138961 L1-d$-stores ( 4.46x scaled)
<not counted> L1-d$-store-misses
348659 L1-d$-prefetches ( 4.41x scaled)
236550 L1-d$-prefetch-misses ( 4.41x scaled)
248192242 L1-i$-loads ( 4.46x scaled)
3805771 L1-i$-load-misses ( 4.46x scaled)
334292 L1-d$-prefetches ( 4.46x scaled)
239715 L1-d$-prefetch-misses ( 4.47x scaled)
4966124 LLC-loads ( 4.47x scaled)
531900 LLC-load-misses ( 4.47x scaled)
5605759 LLC-stores ( 4.47x scaled)
<not counted> LLC-store-misses
<not counted> LLC-prefetches
<not counted> LLC-prefetch-misses
253681838 dTLB-loads ( 4.48x scaled)
4634809 dTLB-load-misses ( 4.49x scaled)
<not counted> dTLB-stores
<not counted> dTLB-store-misses
<not counted> dTLB-prefetches
<not counted> dTLB-prefetch-misses
253610942 iTLB-loads ( 4.51x scaled)
3271 iTLB-load-misses ( 4.56x scaled)
105697493 branch-loads ( 4.61x scaled)
5136856 branch-load-misses ( 4.66x scaled)0.375218449 seconds time elapsed.
--- tools/perf/util/parse-events.c | 67 ++++++++++++++++++++++++++++++++++++++- 1 files changed, 65 insertions(+), 2 deletions(-)
tools/perf/util/parse-events.c +65 -2
@@ -48,6 +48,58 @@ struct event_type_symbol {static struct event_type_symbol event_type_symbols[] = { [PERF_TYPE_HARDWARE ] = {"hw-events" ,"all-hw-events" , }, [PERF_TYPE_SOFTWARE ] = {"sw-events" ,"all-sw-events" , }, + [PERF_TYPE_TRACEPOINT ] = {"" ,"" , }, + [PERF_TYPE_HW_CACHE ] = {"cache-events" ,"all-cache-events" , }, + [PERF_TYPE_RAW ] = {"" ,"" , }, +}; + +struct event_cache_symbol { +u8 type; +u64 config; +}; + +#define CHCACHE (x, y, z) \ +.type = PERF_TYPE_HW_CACHE, \ +.config = (PERF_COUNT_HW_CACHE_##x | (PERF_COUNT_HW_CACHE_OP_##y << 8) |\ + (PERF_COUNT_HW_CACHE_RESULT_##z << 16)) + +/* +* Generalized Hardware cache counters events +* L1I is READ and PREFETCH only +* ITLB and BPU is READ only +*/ +static struct event_cache_symbol event_cache_symbols[] = { + {CHCACHE (L1D ,READ ,ACCESS ) }, + {CHCACHE (L1D ,READ ,MISS ) }, + {CHCACHE (L1D ,WRITE ,ACCESS ) }, + {CHCACHE (L1D ,WRITE ,MISS ) }, + {CHCACHE (L1D ,PREFETCH ,ACCESS ) }, + {CHCACHE (L1D ,PREFETCH ,MISS ) }, + + {CHCACHE (L1I ,READ ,ACCESS ) }, + {CHCACHE (L1I ,READ ,MISS ) }, + {CHCACHE (L1D ,PREFETCH ,ACCESS ) }, + {CHCACHE (L1D ,PREFETCH ,MISS ) }, + + {CHCACHE (LL ,READ ,ACCESS ) }, + {CHCACHE (LL ,READ ,MISS ) }, + {CHCACHE (LL ,WRITE ,ACCESS ) }, + {CHCACHE (LL ,WRITE ,MISS ) }, + {CHCACHE (LL ,PREFETCH ,ACCESS ) }, + {CHCACHE (LL ,PREFETCH ,MISS ) }, + + {CHCACHE (DTLB ,READ ,ACCESS ) }, + {CHCACHE (DTLB ,READ ,MISS ) }, + {CHCACHE (DTLB ,WRITE ,ACCESS ) }, + {CHCACHE (DTLB ,WRITE ,MISS ) }, + {CHCACHE (DTLB ,PREFETCH ,ACCESS ) }, + {CHCACHE (DTLB ,PREFETCH ,MISS ) }, + + {CHCACHE (ITLB ,READ ,ACCESS ) }, + {CHCACHE (ITLB ,READ ,MISS ) }, + + {CHCACHE (BPU ,READ ,ACCESS ) }, + {CHCACHE (BPU ,READ ,MISS ) }, };#define __PERF_COUNTER_FIELD (config, name) \
@@ -266,6 +318,16 @@ static int set_multiple_events(unsigned int type) }break ; +case PERF_TYPE_HW_CACHE : +for (i =0 ; i <ARRAY_SIZE (event_cache_symbols); i++) { +memset (&attr,0 ,sizeof (attr)); + attr.type = event_cache_symbols[i].type; + attr.config = event_cache_symbols[i].config; + attrs[nr_counters] = attr; + nr_counters++; + } +break ; +default :return -1 ; }
@@ -279,9 +341,10 @@ static int set_multiple_events(unsigned int type) static int check_type_events(const char *str, unsigned int i) { - if (!strncmp(str, event_type_symbols[i].symbol, + if (strlen(event_type_symbols[i].symbol)) + if (!strncmp(str, event_type_symbols[i].symbol, strlen(event_type_symbols[i].symbol))) - return 1; + return 1; if (strlen(event_type_symbols[i].alias)) if (!strncmp(str, event_type_symbols[i].alias,
Signature · 1 line
1.6.0.6
Re: [PATCH 2/3 -tip] perf_counter tools: Add support to set of multiple events in one shot
In reply to Jaswinder Singh RajputQuoted 9 lines · Jaswinder Singh Rajput
Add support for HARDWARE and SOFTWARE events :
perf stat -e all-sw-events
perf stat -e sw-events
perf stat -e all-hw-events
perf stat -e hw-events+static struct
event_type_symbolevent_type_symbols[] = {
+ [PERF_TYPE_HARDWARE] = { "hw-events", "all-hw-events", },
+ [PERF_TYPE_SOFTWARE] = { "sw-events", "all-sw-events", },
Hm, this really just special-cases and open-codes these. The better solution is what i suggested in my review of your prior patches:
regex pattern matching.
Ingo
Re: [PATCH 2/3 -tip] perf_counter tools: Add support to set of multiple events in one shot
Quoted 8 lines · Ingo Molnar
Quoted 5 lines · Jaswinder Singh Rajput
Add support for HARDWARE and SOFTWARE events :
perf stat -e all-sw-events
perf stat -e sw-events
perf stat -e all-hw-events
perf stat -e hw-events
Quoted 8 lines
Quoted 3 lines
+static struct
event_type_symbolevent_type_symbols[] = {
+ [PERF_TYPE_HARDWARE] = { "hw-events", "all-hw-events", },
+ [PERF_TYPE_SOFTWARE] = { "sw-events", "all-sw-events", },Hm, this really just special-cases and open-codes these. The better solution is what i suggested in my review of your prior patches:
regex pattern matching.
parse_events() is responsible to parse events for 'perf stat -e' and it is parsing by parse_event_symbol()
If you want to use regex pattern matching then either we should make some another option or if we need to rewrite parse_event_symbol to use regex pattern matching which will be applicable to all the events.
Let me know which option you are looking for.
Thanks,
--
JSR
Re: [PATCH 2/3 -tip] perf_counter tools: Add support to set of multiple events in one shot
Quoted 18 lines · Jaswinder Singh Rajput
Quoted 8 lines · Ingo Molnar
Quoted 5 lines · Jaswinder Singh Rajput
Add support for HARDWARE and SOFTWARE events :
perf stat -e all-sw-events
perf stat -e sw-events
perf stat -e all-hw-events
perf stat -e hw-eventsQuoted 8 lines
Quoted 3 lines
+static struct
event_type_symbolevent_type_symbols[] = {
+ [PERF_TYPE_HARDWARE] = { "hw-events", "all-hw-events", },
+ [PERF_TYPE_SOFTWARE] = { "sw-events", "all-sw-events", },Hm, this really just special-cases and open-codes these. The better solution is what i suggested in my review of your prior patches:
regex pattern matching.
Quoted 7 lines
parse_events()is responsible to parse events for 'perf stat -e' and it is parsing byparse_event_symbol()If you want to use regex pattern matching then either we should make some another option or if we need to rewrite
parse_event_symbolto use regex pattern matching which will be applicable to all the events.
As i mentioned it before, i think the most intuitive solution is to extend the --event syntax with regex patterns. No new option - just richer -e syntax.
We could have this syntax:
hw-cpu-cycles
hw-instructions
hw-cache-references
hw-cache-misses
hw-branch-instructions
hw-branch-misses
hw-bus-cycles
sw-cpu-clock
sw-task-clock
sw-page-faults
sw-minor-faults
sw-major-faults
sw-context-switches
sw-cpu-migrations
regex patterns like:
hw-*
sw-*
*
the first one would select all hardware events - the second all software events - the third all events in general. But other regex patterns make sense too, like:
*branch*
*cache*
*fault*
And as the number of generic events increases, so will regex patterns become more and more useful.
Ingo
Re: [PATCH 2/3 -tip] perf_counter tools: Add support to set of multiple events in one shot
Quoted 60 lines · Ingo Molnar
Quoted 18 lines · Jaswinder Singh Rajput
Quoted 8 lines · Ingo Molnar
Quoted 5 lines · Jaswinder Singh Rajput
Add support for HARDWARE and SOFTWARE events :
perf stat -e all-sw-events
perf stat -e sw-events
perf stat -e all-hw-events
perf stat -e hw-eventsQuoted 8 lines
Quoted 3 lines
+static struct
event_type_symbolevent_type_symbols[] = {
+ [PERF_TYPE_HARDWARE] = { "hw-events", "all-hw-events", },
+ [PERF_TYPE_SOFTWARE] = { "sw-events", "all-sw-events", },Hm, this really just special-cases and open-codes these. The better solution is what i suggested in my review of your prior patches:
regex pattern matching.Quoted 7 lines
parse_events()is responsible to parse events for 'perf stat -e' and it is parsing byparse_event_symbol()If you want to use regex pattern matching then either we should make some another option or if we need to rewrite
parse_event_symbolto use regex pattern matching which will be applicable to all the events.As i mentioned it before, i think the most intuitive solution is to extend the --event syntax with regex patterns. No new option - just richer -e syntax.
We could have this syntax:
hw-cpu-cycles
hw-instructions
hw-cache-references
hw-cache-misses
hw-branch-instructions
hw-branch-misses
hw-bus-cyclessw-cpu-clock
sw-task-clock
sw-page-faults
sw-minor-faults
sw-major-faults
sw-context-switches
sw-cpu-migrationsregex patterns like:
hw-*
sw-*
*the first one would select all hardware events - the second all software events - the third all events in general. But other regex patterns make sense too, like:
Quoted 7 lines
*branch*
*cache*
*fault*And as the number of generic events increases, so will regex patterns become more and more useful.
This is a good approach, but this should be done separately because these will effect all the events.
Currently I send patches to support multiple events in one shot which is adding new feature.
And you telling to use regex wildcard for all the events in parse_event_symbol() which is enhancement and should be done in series of patches.
Do not you think you are mixing 2 things. I request you to accept "multiple events in one shot" patches. And then we will do enhancement.
Thanks,
--
JSR
Re: [PATCH 2/3 -tip] perf_counter tools: Add support to set of multiple events in one shot
Quoted 72 lines · Jaswinder Singh Rajput
Quoted 60 lines · Ingo Molnar
Quoted 18 lines · Jaswinder Singh Rajput
Quoted 8 lines · Ingo Molnar
Quoted 5 lines · Jaswinder Singh Rajput
Add support for HARDWARE and SOFTWARE events :
perf stat -e all-sw-events
perf stat -e sw-events
perf stat -e all-hw-events
perf stat -e hw-eventsQuoted 8 lines
Quoted 3 lines
+static struct
event_type_symbolevent_type_symbols[] = {
+ [PERF_TYPE_HARDWARE] = { "hw-events", "all-hw-events", },
+ [PERF_TYPE_SOFTWARE] = { "sw-events", "all-sw-events", },Hm, this really just special-cases and open-codes these. The better solution is what i suggested in my review of your prior patches:
regex pattern matching.Quoted 7 lines
parse_events()is responsible to parse events for 'perf stat -e' and it is parsing byparse_event_symbol()If you want to use regex pattern matching then either we should make some another option or if we need to rewrite
parse_event_symbolto use regex pattern matching which will be applicable to all the events.As i mentioned it before, i think the most intuitive solution is to extend the --event syntax with regex patterns. No new option - just richer -e syntax.
We could have this syntax:
hw-cpu-cycles
hw-instructions
hw-cache-references
hw-cache-misses
hw-branch-instructions
hw-branch-misses
hw-bus-cyclessw-cpu-clock
sw-task-clock
sw-page-faults
sw-minor-faults
sw-major-faults
sw-context-switches
sw-cpu-migrationsregex patterns like:
hw-*
sw-*
*the first one would select all hardware events - the second all software events - the third all events in general. But other regex patterns make sense too, like:
Quoted 7 lines
*branch*
*cache*
*fault*And as the number of generic events increases, so will regex patterns become more and more useful.
This is a good approach, but this should be done separately because these will effect all the events.
Quoted 6 lines
Currently I send patches to support multiple events in one shot which is adding new feature.
And you telling to use regex wildcard for all the events in
parse_event_symbol()which is enhancement and should be done in series of patches.
The above patterns i suggested _already cover_ 'multiple events'.
We might define further aliases like:
all := "*"
all-sw := "sw-*"but it should all be in terms of patterns and regular expressions, not via some hardcoded special-case thing as your posted patches did.
Ingo
Re: [PATCH 2/3 -tip] perf_counter tools: Add support to set of multiple events in one shot
Quoted 11 lines · Ingo Molnar
The above patterns i suggested
_alreadycover_'multiple events'.We might define further aliases like:
all := "*" all-sw := "sw-*"but it should all be in terms of patterns and regular expressions, not via some hardcoded special-case thing as your posted patches did.
It seems to me very confusing and needs lot of book-keeping and need to rewrite whole tools/perf/util/parse-events.c because :
* means all perf_event_types :
PERF_TYPE_HARDWARE,
PERF_TYPE_SOFTWARE,
PERF_TYPE_TRACEPOINT,
PERF_TYPE_HW_CACHE,
PERF_TYPE_RAW
hw-* means all hardware events :
PERF_TYPE_HARDWARE,
PERF_TYPE_HW_CACHE,
PERF_TYPE_RAW
sw-* means all software events :
PERF_TYPE_SOFTWARE,
PERF_TYPE_TRACEPOINT
*cache* means all cache based events :
PERF_COUNT_CACHE_REFERENCES, /* Generalized H/W */
PERF_COUNT_CACHE_MISSES, /* Generalized H/W */
PERF_TYPE_HW_CACHE, /* Generalized Cache */
*write* means all write based events :
(L1D, WRITE, ACCESS), (L1D, WRITE, MISS), (LL, WRITE, ACCESS), (LL, WRITE, MISS), (DTLB, WRITE, ACCESS), (DTLB, WRITE, MISS)
Please let me know why it looks complex to me, is it really complex or I am going in wrong direction.
Thanks,
--
JSR
Re: [PATCH 2/3 -tip] perf_counter tools: Add support to set of multiple events in one shot
Quoted 46 lines · Jaswinder Singh Rajput
Quoted 11 lines · Ingo Molnar
The above patterns i suggested
_alreadycover_'multiple events'.We might define further aliases like:
all := "*" all-sw := "sw-*"but it should all be in terms of patterns and regular expressions, not via some hardcoded special-case thing as your posted patches did.
It seems to me very confusing and needs lot of book-keeping and need to rewrite whole
tools/perf/util/parse-events.cbecause :* means all
perf_event_types:
PERF_TYPE_HARDWARE,
PERF_TYPE_SOFTWARE,
PERF_TYPE_TRACEPOINT,
PERF_TYPE_HW_CACHE,
PERF_TYPE_RAWhw-* means all hardware events :
PERF_TYPE_HARDWARE,
PERF_TYPE_HW_CACHE,
PERF_TYPE_RAWsw-* means all software events :
PERF_TYPE_SOFTWARE,
PERF_TYPE_TRACEPOINT*cache* means all cache based events :
PERF_COUNT_CACHE_REFERENCES, /* Generalized H/W */
PERF_COUNT_CACHE_MISSES, /* Generalized H/W */
PERF_TYPE_HW_CACHE, /* Generalized Cache */*write* means all write based events :
(L1D, WRITE, ACCESS), (L1D, WRITE, MISS), (LL, WRITE, ACCESS), (LL, WRITE, MISS), (DTLB, WRITE, ACCESS), (DTLB, WRITE, MISS)
Quoted 2 lines
Please let me know why it looks complex to me, is it really complex or I am going in wrong direction.
It would certainly need some reorganization of the code but the end result would be more flexible and other places could use it too, for example:
perf test -e hw-*
would test all (known) hardware counters.
Ingo
Re: [PATCH 2/3 -tip] perf_counter tools: Add support to set of multiple events in one shot
Quoted 51 lines · Ingo Molnar
Quoted 46 lines · Jaswinder Singh Rajput
Quoted 11 lines · Ingo Molnar
The above patterns i suggested
_alreadycover_'multiple events'.We might define further aliases like:
all := "*" all-sw := "sw-*"but it should all be in terms of patterns and regular expressions, not via some hardcoded special-case thing as your posted patches did.
It seems to me very confusing and needs lot of book-keeping and need to rewrite whole
tools/perf/util/parse-events.cbecause :* means all
perf_event_types:
PERF_TYPE_HARDWARE,
PERF_TYPE_SOFTWARE,
PERF_TYPE_TRACEPOINT,
PERF_TYPE_HW_CACHE,
PERF_TYPE_RAWhw-* means all hardware events :
PERF_TYPE_HARDWARE,
PERF_TYPE_HW_CACHE,
PERF_TYPE_RAWsw-* means all software events :
PERF_TYPE_SOFTWARE,
PERF_TYPE_TRACEPOINT*cache* means all cache based events :
PERF_COUNT_CACHE_REFERENCES, /* Generalized H/W */
PERF_COUNT_CACHE_MISSES, /* Generalized H/W */
PERF_TYPE_HW_CACHE, /* Generalized Cache */*write* means all write based events :
(L1D, WRITE, ACCESS), (L1D, WRITE, MISS), (LL, WRITE, ACCESS), (LL, WRITE, MISS), (DTLB, WRITE, ACCESS), (DTLB, WRITE, MISS)Quoted 2 lines
Please let me know why it looks complex to me, is it really complex or I am going in wrong direction.
Quoted 8 lines
It would certainly need some reorganization of the code but the end result would be more flexible and other places could use it too, for example:
perf test -e hw-*
would test all (known) hardware counters.
Its true.
Can you please verify that the assumptions I made above are correct.
Thanks,
--
JSR
Re: [PATCH 2/3 -tip] perf_counter tools: Add support to set of multiple events in one shot
Quoted 58 lines · Jaswinder Singh Rajput
Quoted 57 lines · Ingo Molnar
Quoted 46 lines · Jaswinder Singh Rajput
Quoted 11 lines · Ingo Molnar
The above patterns i suggested
_alreadycover_'multiple events'.We might define further aliases like:
all := "*" all-sw := "sw-*"but it should all be in terms of patterns and regular expressions, not via some hardcoded special-case thing as your posted patches did.
It seems to me very confusing and needs lot of book-keeping and need to rewrite whole
tools/perf/util/parse-events.cbecause :* means all
perf_event_types:
PERF_TYPE_HARDWARE,
PERF_TYPE_SOFTWARE,
PERF_TYPE_TRACEPOINT,
PERF_TYPE_HW_CACHE,
PERF_TYPE_RAWhw-* means all hardware events :
PERF_TYPE_HARDWARE,
PERF_TYPE_HW_CACHE,
PERF_TYPE_RAWsw-* means all software events :
PERF_TYPE_SOFTWARE,
PERF_TYPE_TRACEPOINT*cache* means all cache based events :
PERF_COUNT_CACHE_REFERENCES, /* Generalized H/W */
PERF_COUNT_CACHE_MISSES, /* Generalized H/W */
PERF_TYPE_HW_CACHE, /* Generalized Cache */*write* means all write based events :
(L1D, WRITE, ACCESS), (L1D, WRITE, MISS), (LL, WRITE, ACCESS), (LL, WRITE, MISS), (DTLB, WRITE, ACCESS), (DTLB, WRITE, MISS)Quoted 2 lines
Please let me know why it looks complex to me, is it really complex or I am going in wrong direction.
It would certainly need some reorganization of the code but the end result would be more flexible and other places could use it too, for example:
perf test -e hw-*
Quoted 7 lines
Quoted 2 lines
would test all (known) hardware counters.
Its true.
Can you please verify that the assumptions I made above are correct.
Well, the right way to approach this is to assign each event a "full name" and a list of aliases/shortcuts (like we have now), and then do pattern matching on the full name.
So we'd have full/long event names like:
hw-cycles
hw-instructions
hw-l1-cache-load-misses
sw-minor-page-faults
...
to implement regex patterns over these, no event specific knowledge should be put into the pattern matching engine itself - it just blindly goes over the full names as strings.
As long as the full names are unique and structured well, this will work fine. The only non-trivial piece of restructuring is to make it easy for the pattern matching engine to iterate over all events. Right now they are in separate tables - perhaps they should be collected into a single table or so.
Ingo
[tip:perfcounters/urgent] perf stat: Improve output
In reply to Jaswinder Singh RajputCommit-ID: 6e750a8fc009fd0ae98704525d1d8e80d60e8cc9 Gitweb: http://git.kernel.org/tip/6e750a8fc009fd0ae98704525d1d8e80d60e8cc9 Author: Jaswinder Singh Rajput <[email protected]> AuthorDate: Sat, 27 Jun 2009 03:02:07 +0530 Committer: Ingo Molnar <[email protected]> CommitDate: Sat, 27 Jun 2009 18:39:41 +0200
perf stat: Improve output
Increase size for event name to handle bigger names like 'L1-d$-prefetch-misses'
Changed scaled counters from percentage to a multiplicative factor because the latter is more expressive.
Also aligned the scaling factor, otherwise sometimes it looks like:
384 iTLB-load-misses (4.74x scaled)
452029 branch-loads (8.00x scaled)
5892 branch-load-misses (20.39x scaled)
972315 iTLB-loads (3.24x scaled)Before:
150708 L1-d$-stores (scaled from 23.57%)
428804 L1-d$-prefetches (scaled from 23.47%)
314446 L1-d$-prefetch-misses (scaled from 23.42%)
252626137 L1-i$-loads (scaled from 23.24%)
5297550 dTLB-load-misses (scaled from 23.96%)
106992392 branch-loads (scaled from 23.67%)
5239561 branch-load-misses (scaled from 23.43%)After:
1731713 L1-d$-loads ( 14.25x scaled)
44241 L1-d$-prefetches ( 3.88x scaled)
21076 L1-d$-prefetch-misses ( 3.40x scaled)
5789421 L1-i$-loads ( 3.78x scaled)
29645 dTLB-load-misses ( 2.95x scaled)
461474 branch-loads ( 6.52x scaled)
7493 branch-load-misses ( 26.57x scaled)LKML-Reference: <[email protected]>
--- tools/perf/builtin-stat.c | 11 +++++------ 1 files changed, 5 insertions(+), 6 deletions(-)
tools/perf/builtin-stat.c +5 -6
@@ -32,6 +32,7 @@* Wu Fengguang <[email protected]> * Mike Galbraith <[email protected]> * Paul Mackerras <[email protected]> +* Jaswinder Singh Rajput <[email protected]> * * Released under the GPL v2. (and only v2, not any later version) */
@@ -251,7 +252,7 @@ static void nsec_printout(int counter, u64 *count, u64 *noise) {double msecs = (double )count[0 ] /1000000 ; -fprintf (stderr," %14.6f %-20s" , msecs,event_name (counter)); +fprintf (stderr," %14.6f %-24s" , msecs,event_name (counter));if (attrs[counter].type ==PERF_TYPE_SOFTWARE && attrs[counter].config ==PERF_COUNT_SW_TASK_CLOCK ) {
@@ -265,7 +266,7 @@ static void nsec_printout(int counter, u64 *count, u64 *noise)static void abs_printout (int counter,u64 *count,u64 *noise) { -fprintf (stderr," %14Ld %-20s" , count[0 ],event_name (counter)); +fprintf (stderr," %14Ld %-24s" , count[0 ],event_name (counter)); if (runtime_cycles_avg && attrs[counter].type == PERF_TYPE_HARDWARE &&
@@ -295,7 +296,7 @@ static void print_counter(int counter) scaled = event_scaled_avg[counter];if (scaled ==-1 ) { -fprintf (stderr," %14s %-20s\n" , +fprintf (stderr," %14s %-24s\n" ,"<not counted>" ,event_name (counter));return ; }
@@ -306,8 +307,7 @@ static void print_counter(int counter)abs_printout (counter, count, noise);if (scaled) -fprintf (stderr," (scaled from %.2f%%)" , - (double ) count[2 ] / count[1 ] *100 ); +fprintf (stderr," (%7.2fx scaled)" , (double )count[1 ]/count[2 ]);fprintf (stderr,"\n" ); }
@@ -421,7 +421,6 @@ static void print_stat(int argc, const char **argv)for (counter =0 ; counter < nr_counters; counter++)print_counter (counter); -fprintf (stderr,"\n" );fprintf (stderr," %14.9f seconds time elapsed" , (double )walltime_nsecs_avg/1e9 );
Peter Zijlstra
maintainerRe: [tip:perfcounters/urgent] perf stat: Improve output
Quoted 9 lines · tip-bot for Jaswinder Singh Rajput
Before:
150708 L1-d$-stores (scaled from 23.57%) 428804 L1-d$-prefetches (scaled from 23.47%) 314446 L1-d$-prefetch-misses (scaled from 23.42%) 252626137 L1-i$-loads (scaled from 23.24%) 5297550 dTLB-load-misses (scaled from 23.96%) 106992392 branch-loads (scaled from 23.67%) 5239561 branch-load-misses (scaled from 23.43%)
Quoted 8 lines
After:
1731713 L1-d$-loads ( 14.25x scaled) 44241 L1-d$-prefetches ( 3.88x scaled) 21076 L1-d$-prefetch-misses ( 3.40x scaled) 5789421 L1-i$-loads ( 3.78x scaled) 29645 dTLB-load-misses ( 2.95x scaled) 461474 branch-loads ( 6.52x scaled) 7493 branch-load-misses ( 26.57x scaled)
I really rather like the before better. Its far easier to get a feel for the overload factor from a fraction that from this multiplier.
Re: [tip:perfcounters/urgent] perf stat: Improve output
Quoted 20 lines · Peter Zijlstra
Quoted 9 lines · tip-bot for Jaswinder Singh Rajput
Before:
150708 L1-d$-stores (scaled from 23.57%) 428804 L1-d$-prefetches (scaled from 23.47%) 314446 L1-d$-prefetch-misses (scaled from 23.42%) 252626137 L1-i$-loads (scaled from 23.24%) 5297550 dTLB-load-misses (scaled from 23.96%) 106992392 branch-loads (scaled from 23.67%) 5239561 branch-load-misses (scaled from 23.43%)Quoted 8 lines
After:
1731713 L1-d$-loads ( 14.25x scaled) 44241 L1-d$-prefetches ( 3.88x scaled) 21076 L1-d$-prefetch-misses ( 3.40x scaled) 5789421 L1-i$-loads ( 3.78x scaled) 29645 dTLB-load-misses ( 2.95x scaled) 461474 branch-loads ( 6.52x scaled) 7493 branch-load-misses ( 26.57x scaled)
Quoted 3 lines
I really rather like the before better. Its far easier to get a feel for the overload factor from a fraction that from this multiplier.
ok - i reverted this portion. I liked the multiplicator a bit more - but you seem to have a strong preference for the percentage unit.
Ingo