perf vendor events: Add JSON metrics for Arm CMN
Subsystems (4)
v7 cover perf vendor events: Add JSON metrics for Arm CMN
Changes since v6:
- Supplement the omitted EventCode;
- Keep the original way of ConfigCode;
- Supplement the test in empty-pmu-events.c, so that the pmu event test
can succeed when build with NO_JEVENT=1.
- Link: https://lore.kernel.org/all/[email protected]/
Jing Zhang (8):
perf pmu: "Compat" supports matching multiple identifiers perf metric: "Compat" supports matching multiple identifiers perf vendor events: Supplement the omitted EventCode perf jevents: Support more event fields
perf test: Make matching_pmu effective
perf test: Add pmu-event test for "Compat" and new event_field. perf jevents: Add support for Arm CMN PMU aliasing perf vendor events: Add JSON metrics for Arm CMN
.../pmu-events/arch/arm64/arm/cmn/sys/cmn.json | 266 +++++++++++++++++++++ .../pmu-events/arch/arm64/arm/cmn/sys/metric.json | 74 ++++++ .../pmu-events/arch/test/test_soc/sys/uncore.json | 8 + .../pmu-events/arch/x86/alderlake/pipeline.json | 9 + .../pmu-events/arch/x86/alderlaken/pipeline.json | 3 + .../pmu-events/arch/x86/broadwell/pipeline.json | 4 + .../pmu-events/arch/x86/broadwellde/pipeline.json | 4 + .../arch/x86/broadwellde/uncore-cache.json | 2 + .../arch/x86/broadwellde/uncore-interconnect.json | 1 + .../arch/x86/broadwellde/uncore-memory.json | 1 + .../arch/x86/broadwellde/uncore-power.json | 1 + .../pmu-events/arch/x86/broadwellx/pipeline.json | 4 + .../arch/x86/broadwellx/uncore-cache.json | 2 + .../arch/x86/broadwellx/uncore-interconnect.json | 13 + .../arch/x86/broadwellx/uncore-memory.json | 2 + .../arch/x86/broadwellx/uncore-power.json | 1 + .../pmu-events/arch/x86/cascadelakex/pipeline.json | 4 + .../arch/x86/cascadelakex/uncore-cache.json | 2 + .../arch/x86/cascadelakex/uncore-interconnect.json | 1 + .../arch/x86/cascadelakex/uncore-io.json | 1 + .../arch/x86/cascadelakex/uncore-memory.json | 1 + .../arch/x86/cascadelakex/uncore-power.json | 1 + .../pmu-events/arch/x86/elkhartlake/pipeline.json | 2 + .../pmu-events/arch/x86/goldmont/pipeline.json | 3 + .../pmu-events/arch/x86/goldmontplus/pipeline.json | 3 + .../pmu-events/arch/x86/grandridge/pipeline.json | 3 + .../arch/x86/graniterapids/pipeline.json | 4 + .../perf/pmu-events/arch/x86/haswell/pipeline.json | 4 + .../pmu-events/arch/x86/haswellx/pipeline.json | 4 + .../pmu-events/arch/x86/haswellx/uncore-cache.json | 2 + .../arch/x86/haswellx/uncore-interconnect.json | 14 ++ .../arch/x86/haswellx/uncore-memory.json | 2 + .../pmu-events/arch/x86/haswellx/uncore-power.json | 1 + .../perf/pmu-events/arch/x86/icelake/pipeline.json | 4 + .../pmu-events/arch/x86/icelakex/pipeline.json | 4 + .../pmu-events/arch/x86/icelakex/uncore-cache.json | 1 + .../arch/x86/icelakex/uncore-interconnect.json | 1 + .../arch/x86/icelakex/uncore-memory.json | 1 + .../pmu-events/arch/x86/icelakex/uncore-power.json | 1 + .../pmu-events/arch/x86/ivybridge/pipeline.json | 3 + .../perf/pmu-events/arch/x86/ivytown/pipeline.json | 4 + .../pmu-events/arch/x86/ivytown/uncore-cache.json | 2 + .../arch/x86/ivytown/uncore-interconnect.json | 11 + .../pmu-events/arch/x86/ivytown/uncore-memory.json | 1 + .../pmu-events/arch/x86/ivytown/uncore-power.json | 1 + .../pmu-events/arch/x86/jaketown/pipeline.json | 4 + .../pmu-events/arch/x86/jaketown/uncore-cache.json | 2 + .../arch/x86/jaketown/uncore-interconnect.json | 12 + .../arch/x86/jaketown/uncore-memory.json | 1 + .../pmu-events/arch/x86/jaketown/uncore-power.json | 2 + .../arch/x86/knightslanding/pipeline.json | 3 + .../arch/x86/knightslanding/uncore-cache.json | 1 + .../arch/x86/knightslanding/uncore-memory.json | 4 + .../pmu-events/arch/x86/meteorlake/pipeline.json | 8 + .../pmu-events/arch/x86/sandybridge/pipeline.json | 4 + .../arch/x86/sapphirerapids/pipeline.json | 5 + .../pmu-events/arch/x86/sierraforest/pipeline.json | 4 + .../pmu-events/arch/x86/silvermont/pipeline.json | 3 + .../perf/pmu-events/arch/x86/skylake/pipeline.json | 4 + .../pmu-events/arch/x86/skylakex/pipeline.json | 4 + .../pmu-events/arch/x86/skylakex/uncore-cache.json | 2 + .../arch/x86/skylakex/uncore-interconnect.json | 1 + .../pmu-events/arch/x86/skylakex/uncore-io.json | 1 + .../arch/x86/skylakex/uncore-memory.json | 1 + .../pmu-events/arch/x86/skylakex/uncore-power.json | 1 + .../pmu-events/arch/x86/snowridgex/pipeline.json | 2 + .../arch/x86/snowridgex/uncore-cache.json | 1 + .../arch/x86/snowridgex/uncore-interconnect.json | 1 + .../arch/x86/snowridgex/uncore-memory.json | 1 + .../arch/x86/snowridgex/uncore-power.json | 1 + .../pmu-events/arch/x86/tigerlake/pipeline.json | 5 + tools/perf/pmu-events/empty-pmu-events.c | 8 + tools/perf/pmu-events/jevents.py | 21 +- tools/perf/tests/pmu-events.c | 64 ++++- tools/perf/util/metricgroup.c | 2 +- tools/perf/util/pmu.c | 33 ++- tools/perf/util/pmu.h | 1 + 77 files changed, 679 insertions(+), 9 deletions(-) create mode 100644 tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.json create mode 100644 tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.json
Signature · 1 line
1.8.3.1
PATCH v7 1/8 perf pmu: "Compat" supports matching multiple identifiers
The jevent "Compat" is used for uncore PMU alias or metric definitions.
The same PMU driver has different PMU identifiers due to different hardware versions and types, but they may have some common PMU event. Since a Compat value can only match one identifier, when adding the same event alias to PMUs with different identifiers, each identifier needs to be defined once, which is not streamlined enough.
So let "Compat" supports matching multiple identifiers for uncore PMU alias. For example, the Compat value {43401;436*} can match the PMU identifier "43401", that is, CMN600_r0p0, and the PMU identifier with the prefix "436", that is, all CMN650, where "*" is a wildcard. Tokens in Unit field are delimited by ';' with no spaces.
--- tools/perf/util/pmu.c | 33 +++++++++++++++++++++++++++++++-- tools/perf/util/pmu.h | 1 + 2 files changed, 32 insertions(+), 2 deletions(-)
tools/perf/util/pmu.c +31 -2
@@ -776,6 +776,35 @@ static bool pmu_uncore_alias_match(const char *pmu_name, const char *name)return res; } +bool pmu_uncore_identifier_match (const char *id,const char *compat) +{ +char *tmp =NULL , *tok, *str; +bool res; +int n; + +/* +* The strdup() call is necessary here because "compat" is a const str* +* type and cannot be used as an argument to strtok_r(). +*/ + str =strdup (compat); +if (!str) +return false; + + tok =strtok_r (str,";" , &tmp); +for (; tok; tok =strtok_r (NULL ,";" , &tmp)) { + n =strlen (tok); +if ((tok[n -1 ] =='*' && !strncmp (id, tok, n -1 )) || + !strcmp (id, tok)) { + res = true; + goto out; + } + } + res = false; +out: +free (str); +return res; +} +struct pmu_add_cpu_aliases_map_data {struct list_head *head;const char *name;
@@ -847,8 +876,8 @@ static int pmu_add_sys_aliases_iter_fn(const struct pmu_event *pe,if (!pe->compat || !pe->pmu)return 0 ; -if (!strcmp (pmu->id, pe->compat) && -pmu_uncore_alias_match (pe->pmu, pmu->name)) { +if (pmu_uncore_alias_match (pe->pmu, pmu->name) && +pmu_uncore_identifier_match (pmu->id, pe->compat)) { __perf_pmu__new_alias(idata->head, -1, (char *)pe->name, (char *)pe->desc,
tools/perf/util/pmu.h +1 -0
@@ -241,6 +241,7 @@ void pmu_add_cpu_aliases_table(struct list_head *head, struct perf_pmu *pmu,char *perf_pmu__getcpuid (struct perf_pmu *pmu);const struct pmu_events_table *pmu_events_table__find (void );const struct pmu_metrics_table *pmu_metrics_table__find (void ); +bool pmu_uncore_identifier_match (const char *id,const char *compat);void perf_pmu_free_alias (struct perf_pmu_alias *alias);int perf_pmu__convert_scale (const char *scale,char **end,double *sval);
Signature · 1 line
1.8.3.1
Robin Murphy
maintainerRe: [PATCH v7 1/8] perf pmu: "Compat" supports matching multiple identifiers
Quoted 8 lines · Jing Zhang
The jevent "Compat" is used for uncore PMU alias or metric definitions.
The same PMU driver has different PMU identifiers due to different hardware versions and types, but they may have some common PMU event. Since a Compat value can only match one identifier, when adding the same event alias to PMUs with different identifiers, each identifier needs to be defined once, which is not streamlined enough.
Quoted 5 lines
So let "Compat" supports matching multiple identifiers for uncore PMU alias. For example, the Compat value {43401;436*} can match the PMU identifier "43401", that is,
CMN600_r0p0, and the PMU identifier with the prefix "436", that is, all CMN650, where "*" is a wildcard. Tokens in Unit field are delimited by ';' with no spaces.
I wonder is there any possibility of supporting multiple values as a JSON array, rather than a single delimited string? Otherwise, if we're putting restrictions on what characters a driver can expose as an identifier, then I think that really wants explicitly documenting. AFAICT there's currently not even any documentation of the de-facto ABI that it's expected to be a free-form string rather than completely arbitrary binary data.
Thanks,
Robin.
Quoted 63 lines
Signed-off-by Jing Zhang <[email protected]>Reviewed-by John Garry <[email protected]>--- tools/perf/util/pmu.c | 33 +++++++++++++++++++++++++++++++-- tools/perf/util/pmu.h | 1 + 2 files changed, 32 insertions(+), 2 deletions(-)tools/perf/util/pmu.c +31
-2@@ -776,6 +776,35 @@ static bool pmu_uncore_alias_match(const char *pmu_name, const char *name)return res; } +bool pmu_uncore_identifier_match (const char *id,const char *compat) +{ +char *tmp =NULL , *tok, *str; +bool res; +int n; + +/* +* The strdup() call is necessary here because "compat" is a const str* +* type and cannot be used as an argument to strtok_r(). +*/ + str =strdup (compat); +if (!str) +return false; + + tok =strtok_r (str,";" , &tmp); +for (; tok; tok =strtok_r (NULL ,";" , &tmp)) { + n =strlen (tok); +if ((tok[n -1 ] =='*' && !strncmp (id, tok, n -1 )) || + !strcmp (id, tok)) { + res = true; + goto out; + } + } + res = false; +out: +free (str); +return res; +} +struct pmu_add_cpu_aliases_map_data {struct list_head *head;const char *name;@@ -847,8 +876,8 @@ static int pmu_add_sys_aliases_iter_fn(const struct pmu_event *pe,if (!pe->compat || !pe->pmu)return 0 ; -if (!strcmp (pmu->id, pe->compat) && -pmu_uncore_alias_match (pe->pmu, pmu->name)) { +if (pmu_uncore_alias_match (pe->pmu, pmu->name) && +pmu_uncore_identifier_match (pmu->id, pe->compat)) { __perf_pmu__new_alias(idata->head, -1, (char *)pe->name, (char *)pe->desc,tools/perf/util/pmu.h +0
-0
tools/perf/util/pmu.h void pmu_add_cpu_aliases_table(struct list_head *head, struct perf_pmu *pmu, · line 247
Quoted 7 lines
tools/perf/util/pmu.h +1
-0@@ -241,6 +241,7 @@ void pmu_add_cpu_aliases_table(struct list_head *head, struct perf_pmu *pmu,char *perf_pmu__getcpuid (struct perf_pmu *pmu);const struct pmu_events_table *pmu_events_table__find (void );const struct pmu_metrics_table *pmu_metrics_table__find (void ); +bool pmu_uncore_identifier_match (const char *id,const char *compat);void perf_pmu_free_alias (struct perf_pmu_alias *alias);int perf_pmu__convert_scale (const char *scale,char **end,double *sval);
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Re: [PATCH v7 1/8] perf pmu: "Compat" supports matching multiple identifiers
Quoted 9 lines · Robin Murphy
Quoted 8 lines · Jing Zhang
The jevent "Compat" is used for uncore PMU alias or metric definitions.
The same PMU driver has different PMU identifiers due to different hardware versions and types, but they may have some common PMU event. Since a Compat value can only match one identifier, when adding the same event alias to PMUs with different identifiers, each identifier needs to be defined once, which is not streamlined enough.
Quoted 8 lines
Quoted 5 lines
So let "Compat" supports matching multiple identifiers for uncore PMU alias. For example, the Compat value {43401;436*} can match the PMU identifier "43401", that is,
CMN600_r0p0, and the PMU identifier with the prefix "436", that is, all CMN650, where "*" is a wildcard. Tokens in Unit field are delimited by ';' with no spaces.I wonder is there any possibility of supporting multiple values as a JSON array, rather than a single delimited string? Otherwise, if we're putting restrictions on what characters a driver can expose as an identifier, then I think that really wants explicitly documenting. AFAICT there's currently not even any documentation of the de-facto ABI that it's expected to be a free-form string rather than completely arbitrary binary data.
I'm sorry I almost missed this message, as it was in my spam folder.
If we put multiple values as an array, its parsing in jevent.py will become complicated. I agree that we need to document the character restrictions for driver identifier composition. Both Unit and Compat have the same problem, so certain characters need to be restricted in the identifiers and names of drivers. However, it seems that there is no such document.
Thanks,
Jing
Quoted 62 lines
Thanks,
Robin.Quoted 59 lines
Signed-off-by Jing Zhang <[email protected]>Reviewed-by John Garry <[email protected]>---
tools/perf/util/pmu.c| 33 +++++++++++++++++++++++++++++++--tools/perf/util/pmu.h| 1 +2 files changed, 32 insertions(+), 2 deletions(-)tools/perf/util/pmu.c +28
-0@@ -776,6 +776,35 @@ static bool pmu_uncore_alias_match(const char *pmu_name, const char *name) return res; } +bool pmu_uncore_identifier_match(const char *id, const char *compat) +{ + char *tmp = NULL, *tok, *str; + bool res; + int n; + + /* + * The strdup() call is necessary here because "compat" is a const str* + * type and cannot be used as an argument to strtok_r(). + */ + str = strdup(compat); + if (!str) + return false; + + tok = strtok_r(str, ";", &tmp); + for (; tok; tok = strtok_r(NULL, ";", &tmp)) { + n = strlen(tok); + if ((tok[n - 1] == '*' && !strncmp(id, tok, n - 1)) || + !strcmp(id, tok)) { + res = true; + goto out; + } + } + res = false; +out: + free(str); + return res; +} + struct pmu_add_cpu_aliases_map_data { struct list_head *head; const char *name;@@ -847,8 +876,8 @@ static int pmu_add_sys_aliases_iter_fn(const struct pmu_event *pe, if (!pe->compat || !pe->pmu) return 0;- if (!strcmp(pmu->id,
pe->compat) &&
-pmu_uncore_alias_match(pe->pmu, pmu->name)) { + if (pmu_uncore_alias_match(pe->pmu, pmu->name)&& +pmu_uncore_identifier_match(pmu->id, pe->compat)) { __perf_pmu__new_alias(idata->head, -1, (char *)pe->name,
(char *)pe->desc,tools/perf/util/pmu.h +0
-1--- a/tools/perf/util/pmu.h
Quoted 8 lines
tools/perf/util/pmu.h void pmu_add_cpu_aliases_table(struct list_head *head, struct perf_pmu *pmu, · line 246
Quoted 8 lines
+++
b/tools/perf/util/pmu.htools/perf/util/pmu.h +1
-0@@ -241,6 +241,7 @@ void pmu_add_cpu_aliases_table(struct list_head *head, struct perf_pmu *pmu, char *perf_pmu__getcpuid(struct perf_pmu *pmu); const struct pmu_events_table *pmu_events_table__find(void); const struct pmu_metrics_table *pmu_metrics_table__find(void); +bool pmu_uncore_identifier_match (const char *id,const char *compat); void perf_pmu_free_alias(struct perf_pmu_alias *alias); int perf_pmu__convert_scale(const char *scale, char **end, double *sval);
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Ian Rogers
reviewerRe: [PATCH v7 1/8] perf pmu: "Compat" supports matching multiple identifiers
In reply to Jing ZhangQuoted 27 lines · Jing Zhang
The jevent "Compat" is used for uncore PMU alias or metric definitions.
The same PMU driver has different PMU identifiers due to different hardware versions and types, but they may have some common PMU event. Since a Compat value can only match one identifier, when adding the same event alias to PMUs with different identifiers, each identifier needs to be defined once, which is not streamlined enough.
So let "Compat" supports matching multiple identifiers for uncore PMU alias. For example, the Compat value {43401;436*} can match the PMU identifier "43401", that is,
CMN600_r0p0, and the PMU identifier with the prefix "436", that is, all CMN650, where "*" is a wildcard. Tokens in Unit field are delimited by ';' with no spaces.Signed-off-by Jing Zhang <[email protected]>Reviewed-by John Garry <[email protected]>--- tools/perf/util/pmu.c | 33 +++++++++++++++++++++++++++++++-- tools/perf/util/pmu.h | 1 + 2 files changed, 32 insertions(+), 2 deletions(-)tools/perf/util/pmu.c +0
-0
tools/perf/util/pmu.c static bool pmu_uncore_alias_match(const char *pmu_name, const char *name) · line 779
Quoted 4 lines
tools/perf/util/pmu.c +1
-0@@ -776,6 +776,35 @@ static bool pmu_uncore_alias_match(const char *pmu_name, const char *name)return res; } +bool pmu_uncore_identifier_match (const char *id,const char *compat)
static?
tools/perf/util/pmu.c static bool pmu_uncore_alias_match(const char *pmu_name, const char *name) · line 782
Quoted 3 lines
tools/perf/util/pmu.c +3
-0+{ +char *tmp =NULL , *tok, *str; +bool res;
Initialize to false to avoid the goto.
tools/perf/util/pmu.c static bool pmu_uncore_alias_match(const char *pmu_name, const char *name) · line 783
Quoted 1 line
tools/perf/util/pmu.c +1
-0+int n;
Move into the scope of the for loop, to reduce the scope.
Quoted 4 lines
tools/perf/util/pmu.c +4
-0+ +/* +* The strdup() call is necessary here because "compat" is a const str* +* type and cannot be used as an argument to strtok_r().
tools/perf/util/pmu.c static bool pmu_uncore_alias_match(const char *pmu_name, const char *name) · line 797
Quoted 10 lines
tools/perf/util/pmu.c +10
-0+*/ + str =strdup (compat); +if (!str) +return false; + + tok =strtok_r (str,";" , &tmp); +for (; tok; tok =strtok_r (NULL ,";" , &tmp)) { + n =strlen (tok); +if ((tok[n -1 ] =='*' && !strncmp (id, tok, n -1 )) || + !strcmp (id, tok)) {
We use fnmatch for a similar check:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/util/pmu.c?h=perf-tools-next#n1982
tools/perf/util/pmu.c static bool pmu_uncore_alias_match(const char *pmu_name, const char *name) · line 799
Quoted 2 lines
tools/perf/util/pmu.c +2
-0+ res = true; + goto out;
With "res=false;" above this can just be a regular break.
Thanks,
Ian
Quoted 35 lines
tools/perf/util/pmu.c +10
-2+ } + } + res = false; +out: +free (str); +return res; +} +struct pmu_add_cpu_aliases_map_data {struct list_head *head;const char *name;@@ -847,8 +876,8 @@ static int pmu_add_sys_aliases_iter_fn(const struct pmu_event *pe,if (!pe->compat || !pe->pmu)return 0 ; -if (!strcmp (pmu->id, pe->compat) && -pmu_uncore_alias_match (pe->pmu, pmu->name)) { +if (pmu_uncore_alias_match (pe->pmu, pmu->name) && +pmu_uncore_identifier_match (pmu->id, pe->compat)) { __perf_pmu__new_alias(idata->head, -1, (char *)pe->name, (char *)pe->desc,tools/perf/util/pmu.h +1
-1@@ -241,6 +241,7 @@ void pmu_add_cpu_aliases_table(struct list_head *head, struct perf_pmu *pmu,char *perf_pmu__getcpuid (struct perf_pmu *pmu);const struct pmu_events_table *pmu_events_table__find (void );const struct pmu_metrics_table *pmu_metrics_table__find (void ); +bool pmu_uncore_identifier_match (const char *id,const char *compat);void perf_pmu_free_alias (struct perf_pmu_alias *alias);int perf_pmu__convert_scale (const char *scale,char **end,double *sval); --
Quoted 2 lines
1.8.3.1
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Re: [PATCH v7 1/8] perf pmu: "Compat" supports matching multiple identifiers
Quoted 31 lines · Ian Rogers
Quoted 30 lines · Jing Zhang
The jevent "Compat" is used for uncore PMU alias or metric definitions.
The same PMU driver has different PMU identifiers due to different hardware versions and types, but they may have some common PMU event. Since a Compat value can only match one identifier, when adding the same event alias to PMUs with different identifiers, each identifier needs to be defined once, which is not streamlined enough.
So let "Compat" supports matching multiple identifiers for uncore PMU alias. For example, the Compat value {43401;436*} can match the PMU identifier "43401", that is,
CMN600_r0p0, and the PMU identifier with the prefix "436", that is, all CMN650, where "*" is a wildcard. Tokens in Unit field are delimited by ';' with no spaces.Signed-off-by Jing Zhang <[email protected]>Reviewed-by John Garry <[email protected]>--- tools/perf/util/pmu.c | 33 +++++++++++++++++++++++++++++++-- tools/perf/util/pmu.h | 1 + 2 files changed, 32 insertions(+), 2 deletions(-)tools/perf/util/pmu.c +0
-0@@ -776,6 +776,35 @@ static bool pmu_uncore_alias_match(const char *pmu_name, const char *name)return res; }
Quoted 4 lines
tools/perf/util/pmu.c static bool pmu_uncore_alias_match(const char *pmu_name, const char *name) · line 779
Quoted 1 line
tools/perf/util/pmu.c +1
-0+bool pmu_uncore_identifier_match (const char *id,const char *compat)static?
This function needs to be called in utils/metricgroup.c, so it cannot be static.
Quoted 6 lines
tools/perf/util/pmu.c static bool pmu_uncore_alias_match(const char *pmu_name, const char *name) · line 782
Quoted 3 lines
tools/perf/util/pmu.c +3
-0+{ +char *tmp =NULL , *tok, *str; +bool res;Initialize to false to avoid the goto.
ok,no problem.
Quoted 4 lines
tools/perf/util/pmu.c static bool pmu_uncore_alias_match(const char *pmu_name, const char *name) · line 783
Quoted 1 line
tools/perf/util/pmu.c +1
-0+int n;Move into the scope of the for loop, to reduce the scope.
ok
Quoted 9 lines
Quoted 9 lines
tools/perf/util/pmu.c +9
-0+ +/* +* The strdup() call is necessary here because "compat" is a const str* +* type and cannot be used as an argument to strtok_r(). +*/ + str =strdup (compat); +if (!str) +return false; +
Quoted 9 lines
tools/perf/util/pmu.c static bool pmu_uncore_alias_match(const char *pmu_name, const char *name) · line 797
Quoted 5 lines
tools/perf/util/pmu.c +5
-0+ tok =strtok_r (str,";" , &tmp); +for (; tok; tok =strtok_r (NULL ,";" , &tmp)) { + n =strlen (tok); +if ((tok[n -1 ] =='*' && !strncmp (id, tok, n -1 )) || + !strcmp (id, tok)) {We use fnmatch for a similar check:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/util/pmu.c?h=perf-tools-next#n1982
ok
Quoted 5 lines
tools/perf/util/pmu.c static bool pmu_uncore_alias_match(const char *pmu_name, const char *name) · line 799
Quoted 2 lines
tools/perf/util/pmu.c +2
-0+ res = true; + goto out;With "res=false;" above this can just be a regular break.
ok, thank you!
Quoted 36 lines
Thanks,
IanQuoted 33 lines
tools/perf/util/pmu.c +10
-2+ } + } + res = false; +out: +free (str); +return res; +} +struct pmu_add_cpu_aliases_map_data {struct list_head *head;const char *name;@@ -847,8 +876,8 @@ static int pmu_add_sys_aliases_iter_fn(const struct pmu_event *pe,if (!pe->compat || !pe->pmu)return 0 ; -if (!strcmp (pmu->id, pe->compat) && -pmu_uncore_alias_match (pe->pmu, pmu->name)) { +if (pmu_uncore_alias_match (pe->pmu, pmu->name) && +pmu_uncore_identifier_match (pmu->id, pe->compat)) { __perf_pmu__new_alias(idata->head, -1, (char *)pe->name, (char *)pe->desc,tools/perf/util/pmu.h +1
-0@@ -241,6 +241,7 @@ void pmu_add_cpu_aliases_table(struct list_head *head, struct perf_pmu *pmu,char *perf_pmu__getcpuid (struct perf_pmu *pmu);const struct pmu_events_table *pmu_events_table__find (void );const struct pmu_metrics_table *pmu_metrics_table__find (void ); +bool pmu_uncore_identifier_match (const char *id,const char *compat);void perf_pmu_free_alias (struct perf_pmu_alias *alias);
Quoted 4 lines
Quoted 4 lines
tools/perf/util/pmu.h +0
-1int perf_pmu__convert_scale (const char *scale,char **end,double *sval); --1.8.3.1
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
PATCH v7 2/8 perf metric: "Compat" supports matching multiple identifiers
In reply to Jing ZhangThe jevent "Compat" is used for uncore PMU alias or metric definitions.
The same PMU driver has different PMU identifiers due to different hardware versions and types, but they may have some common PMU metric. Since a Compat value can only match one identifier, when adding the same metric to PMUs with different identifiers, each identifier needs to be defined once, which is not streamlined enough.
So let "Compat" supports matching multiple identifiers for uncore PMU metric.
--- tools/perf/util/metricgroup.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/util/metricgroup.c +1 -1
@@ -477,7 +477,7 @@ static int metricgroup__sys_event_iter(const struct pmu_metric *pm,while ((pmu =perf_pmu__scan (pmu))) { -if (!pmu->id ||strcmp (pmu->id, pm->compat)) +if (!pmu->id || !pmu_uncore_identifier_match (pmu->id, pm->compat))continue ;return d->fn (pm, table, d->data);
Signature · 1 line
1.8.3.1
Re: [PATCH v7 2/8] perf metric: "Compat" supports matching multiple identifiers
Quoted 9 lines · Jing Zhang
The jevent "Compat" is used for uncore PMU alias or metric definitions.
The same PMU driver has different PMU identifiers due to different hardware versions and types, but they may have some common PMU metric. Since a Compat value can only match one identifier, when adding the same metric to PMUs with different identifiers, each identifier needs to be defined once, which is not streamlined enough.
Quoted 5 lines
So let "Compat" supports matching multiple identifiers for uncore PMU metric.
Signed-off-by Jing Zhang <[email protected]>Reviewed-by John Garry <[email protected]>
Thanks,
Ian
Quoted 18 lines
--- tools/perf/util/metricgroup.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)tools/perf/util/metricgroup.c +1
-2@@ -477,7 +477,7 @@ static int metricgroup__sys_event_iter(const struct pmu_metric *pm,while ((pmu =perf_pmu__scan (pmu))) { -if (!pmu->id ||strcmp (pmu->id, pm->compat)) +if (!pmu->id || !pmu_uncore_identifier_match (pmu->id, pm->compat))continue ;return d->fn (pm, table, d->data); --
Quoted 2 lines
1.8.3.1
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
PATCH v7 3/8 perf vendor events: Supplement the omitted EventCode
In reply to Jing ZhangIf there is an "event=0" in the event description, the EventCode can be omitted in the JSON file, and jevent.py will automatically fill in "event=0" during parsing.
However, for some events where EventCode and ConfigCode are missing, it is not necessary to automatically fill in "event=0", such as the CMN event description which is typically "type=xxx, eventid=xxx".
Therefore, before modifying jevent.py to prevent it from automatically adding "event=0" by default, it is necessary to fill in all omitted EventCodes first.
--- tools/perf/pmu-events/arch/x86/alderlake/pipeline.json | 9 +++++++++ tools/perf/pmu-events/arch/x86/alderlaken/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/broadwell/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/broadwellde/pipeline.json | 4 ++++ .../perf/pmu-events/arch/x86/broadwellde/uncore-cache.json | 2 ++ .../arch/x86/broadwellde/uncore-interconnect.json | 1 + .../pmu-events/arch/x86/broadwellde/uncore-memory.json | 1 + .../perf/pmu-events/arch/x86/broadwellde/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/broadwellx/pipeline.json | 4 ++++ .../perf/pmu-events/arch/x86/broadwellx/uncore-cache.json | 2 ++ .../arch/x86/broadwellx/uncore-interconnect.json | 13 +++++++++++++ .../perf/pmu-events/arch/x86/broadwellx/uncore-memory.json | 2 ++ .../perf/pmu-events/arch/x86/broadwellx/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/cascadelakex/pipeline.json | 4 ++++ .../pmu-events/arch/x86/cascadelakex/uncore-cache.json | 2 ++ .../arch/x86/cascadelakex/uncore-interconnect.json | 1 + tools/perf/pmu-events/arch/x86/cascadelakex/uncore-io.json | 1 + .../pmu-events/arch/x86/cascadelakex/uncore-memory.json | 1 + .../pmu-events/arch/x86/cascadelakex/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/elkhartlake/pipeline.json | 2 ++ tools/perf/pmu-events/arch/x86/goldmont/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/goldmontplus/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/grandridge/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/graniterapids/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/haswell/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/haswellx/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/haswellx/uncore-cache.json | 2 ++ .../pmu-events/arch/x86/haswellx/uncore-interconnect.json | 14 ++++++++++++++ tools/perf/pmu-events/arch/x86/haswellx/uncore-memory.json | 2 ++ tools/perf/pmu-events/arch/x86/haswellx/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/icelake/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/icelakex/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/icelakex/uncore-cache.json | 1 + .../pmu-events/arch/x86/icelakex/uncore-interconnect.json | 1 + tools/perf/pmu-events/arch/x86/icelakex/uncore-memory.json | 1 + tools/perf/pmu-events/arch/x86/icelakex/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/ivybridge/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/ivytown/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/ivytown/uncore-cache.json | 2 ++ .../pmu-events/arch/x86/ivytown/uncore-interconnect.json | 11 +++++++++++ tools/perf/pmu-events/arch/x86/ivytown/uncore-memory.json | 1 + tools/perf/pmu-events/arch/x86/ivytown/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/jaketown/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/jaketown/uncore-cache.json | 2 ++ .../pmu-events/arch/x86/jaketown/uncore-interconnect.json | 12 ++++++++++++ tools/perf/pmu-events/arch/x86/jaketown/uncore-memory.json | 1 + tools/perf/pmu-events/arch/x86/jaketown/uncore-power.json | 2 ++ .../perf/pmu-events/arch/x86/knightslanding/pipeline.json | 3 +++ .../pmu-events/arch/x86/knightslanding/uncore-cache.json | 1 + .../pmu-events/arch/x86/knightslanding/uncore-memory.json | 4 ++++ tools/perf/pmu-events/arch/x86/meteorlake/pipeline.json | 8 ++++++++ tools/perf/pmu-events/arch/x86/sandybridge/pipeline.json | 4 ++++ .../perf/pmu-events/arch/x86/sapphirerapids/pipeline.json | 5 +++++ tools/perf/pmu-events/arch/x86/sierraforest/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/silvermont/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/skylake/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/skylakex/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/skylakex/uncore-cache.json | 2 ++ .../pmu-events/arch/x86/skylakex/uncore-interconnect.json | 1 + tools/perf/pmu-events/arch/x86/skylakex/uncore-io.json | 1 + tools/perf/pmu-events/arch/x86/skylakex/uncore-memory.json | 1 + tools/perf/pmu-events/arch/x86/skylakex/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/snowridgex/pipeline.json | 2 ++ .../perf/pmu-events/arch/x86/snowridgex/uncore-cache.json | 1 + .../arch/x86/snowridgex/uncore-interconnect.json | 1 + .../perf/pmu-events/arch/x86/snowridgex/uncore-memory.json | 1 + .../perf/pmu-events/arch/x86/snowridgex/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/tigerlake/pipeline.json | 5 +++++ 68 files changed, 211 insertions(+)
tools/perf/pmu-events/arch/x86/alderlake/pipeline.json +3 -0
@@ -489,6 +489,7 @@ }, { "BriefDescription": "Counts the number of unhalted core clock cycles. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1.", "SampleAfterValue": "2000003",
@@ -550,6 +551,7 @@ }, { "BriefDescription": "Counts the number of unhalted reference clock cycles at TSC frequency. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is not affected by core frequency changes and increments at a fixed frequency that is also used for the Time Stamp Counter (TSC). This event uses fixed counter 2.", "SampleAfterValue": "2000003",
@@ -558,6 +560,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.",
"SampleAfterValue": "2000003",
@@ -584,6 +587,7 @@
},
{
"BriefDescription": "Counts the number of unhalted core clock cycles. (Fixed event)",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1.", "SampleAfterValue": "2000003",
@@ -592,6 +596,7 @@
},
{
"BriefDescription": "Core cycles when the thread is not in halt state",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -743,6 +748,7 @@
},
{
"BriefDescription": "Counts the total number of instructions retired. (Fixed event)",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.ANY",
"PEBS": "1",
"PublicDescription": "Counts the total number of instructions that retired. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. This event continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0.",
@@ -752,6 +758,7 @@
},
{
"BriefDescription": "Number of instructions retired. Fixed Counter - architectural event",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.ANY",
"PEBS": "1",
"PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",
@@ -796,6 +803,7 @@
},
{
"BriefDescription": "Precise instruction retired with PEBS precise-distribution",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.PREC_DIST",
"PEBS": "1",
"PublicDescription": "A version of INST_RETIRED that allows for a precise distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR++) feature to fix bias in how retired instructions get sampled. Use on Fixed Counter 0.",
@@ -1160,6 +1168,7 @@
},
{
"BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event",+ "EventCode": "0x0",
"EventName": "TOPDOWN.SLOTS",
"PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",
tools/perf/pmu-events/arch/x86/alderlaken/pipeline.json +3 -0
@@ -211,6 +211,7 @@ }, { "BriefDescription": "Counts the number of unhalted core clock cycles. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1.", "SampleAfterValue": "2000003",
@@ -225,6 +226,7 @@ }, { "BriefDescription": "Counts the number of unhalted reference clock cycles at TSC frequency. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is not affected by core frequency changes and increments at a fixed frequency that is also used for the Time Stamp Counter (TSC). This event uses fixed counter 2.", "SampleAfterValue": "2000003",
@@ -240,6 +242,7 @@ }, { "BriefDescription": "Counts the number of unhalted core clock cycles. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1.", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/broadwell/pipeline.json +1 -0
@@ -336,6 +336,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. \nNote: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. This event is clocked by base clock (100 Mhz) on Sandy Bridge. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appea r 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.",
"SampleAfterValue": "2000003",
@@ -359,6 +360,7 @@
},
{
"BriefDescription": "Core cycles when the thread is not in halt state",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -366,6 +368,7 @@
},
{
"AnyThread": "1",+ "EventCode": "0x0",
"BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY",
"SampleAfterValue": "2000003",
@@ -514,6 +517,7 @@
},
{
"BriefDescription": "Instructions retired from execution.",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.ANY",
"PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. \nNotes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. \nCounting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/broadwellde/pipeline.json +1 -0
@@ -336,6 +336,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. \nNote: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. This event is clocked by base clock (100 Mhz) on Sandy Bridge. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appea r 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.",
"SampleAfterValue": "2000003",
@@ -359,6 +360,7 @@
},
{
"BriefDescription": "Core cycles when the thread is not in halt state",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -367,6 +369,7 @@
{
"AnyThread": "1",
"BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.",
+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD_ANY",
"SampleAfterValue": "2000003",
"UMask": "0x2"
@@ -514,6 +517,7 @@
},
{
"BriefDescription": "Instructions retired from execution.",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.ANY",
"PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. \nNotes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. \nCounting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/broadwellde/uncore-cache.json +2 -0
@@ -8,6 +8,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"
@@ -1501,6 +1502,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",
tools/perf/pmu-events/arch/x86/broadwellde/uncore-interconnect.json +1 -0
@@ -19,6 +19,7 @@ }, { "BriefDescription": "Clocks in the IRP", + "EventCode": "0x0", "EventName": "UNC_I_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Number of clocks in the IRP.",
tools/perf/pmu-events/arch/x86/broadwellde/uncore-memory.json +1 -0
@@ -131,6 +131,7 @@ }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_DCLOCKTICKS", "PerPkg": "1", "Unit": "iMC"
tools/perf/pmu-events/arch/x86/broadwellde/uncore-power.json +1 -0
@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",
tools/perf/pmu-events/arch/x86/broadwellx/pipeline.json +1 -0
@@ -336,6 +336,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. \nNote: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. This event is clocked by base clock (100 Mhz) on Sandy Bridge. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appea r 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.",
"SampleAfterValue": "2000003",
@@ -359,6 +360,7 @@
},
{
"BriefDescription": "Core cycles when the thread is not in halt state",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -367,6 +369,7 @@
{
"AnyThread": "1",
"BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.",
+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD_ANY",
"SampleAfterValue": "2000003",
"UMask": "0x2"
@@ -514,6 +517,7 @@
},
{
"BriefDescription": "Instructions retired from execution.",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.ANY",
"PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. \nNotes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. \nCounting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/broadwellx/uncore-cache.json +2 -0
@@ -183,6 +183,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"
@@ -1689,6 +1690,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",
tools/perf/pmu-events/arch/x86/broadwellx/uncore-interconnect.json +1 -0
@@ -1,6 +1,7 @@ [ { "BriefDescription": "Number of non data (control) flits transmitted . Derived from unc_q_txl_flits_g0.non_data", + "EventCode": "0x0", "EventName": "QPI_CTL_BANDWIDTH_TX", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calc
ulate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",
@@ -10,6 +11,7 @@
},
{
"BriefDescription": "Number of data flits transmitted . Derived from unc_q_txl_flits_g0.data",+ "EventCode": "0x0",
"EventName": "QPI_DATA_BANDWIDTH_TX",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calc
ulate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",
@@ -37,6 +39,7 @@
},
{
"BriefDescription": "Clocks in the IRP",+ "EventCode": "0x0",
"EventName": "UNC_I_CLOCKTICKS",
"PerPkg": "1",
"PublicDescription": "Number of clocks in the IRP.",
@@ -1400,6 +1403,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 0; Data Tx Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G0.DATA",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calc
ulate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",
@@ -1408,6 +1412,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 0; Non-Data protocol Tx Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G0.NON_DATA",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calc
ulate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",
@@ -1416,6 +1421,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; DRS Flits (both Header and Data)",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.DRS",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency.",
@@ -1424,6 +1430,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; DRS Data Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.DRS_DATA",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of data flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the data flits (not the header).",
@@ -1432,6 +1439,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; DRS Header Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.DRS_NONDATA",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of protocol flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the header flits (not the data). This includes extended headers.",
@@ -1440,6 +1448,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; HOM Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.HOM",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of flits transmitted over QPI on the home channel.",
@@ -1448,6 +1457,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; HOM Non-Request Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.HOM_NONREQ",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of non-request flits transmitted over QPI on the home channel. These are most commonly snoop responses, and this event can be used as a proxy for that.",
@@ -1456,6 +1466,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; HOM Request Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.HOM_REQ",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of data request transmitted over QPI on the home channel. This basically counts the number of remote memory requests transmitted over QPI. In conjunction with the local read count in the Home Agent, one can calculate the number of LLC Misses.",
@@ -1464,6 +1475,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; SNP Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.SNP",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of snoop request flits transmitted over QPI. These requests are contained in the snoop channel. This does not include snoop responses, which are transmitted on the home channel.",
@@ -3162,6 +3174,7 @@
},
{
"BriefDescription": "Uncore Clocks",
+ "EventCode": "0x0",
"EventName": "UNC_S_CLOCKTICKS",
"PerPkg": "1",
"Unit": "SBOX"tools/perf/pmu-events/arch/x86/broadwellx/uncore-memory.json +2 -0
@@ -158,12 +158,14 @@ }, { "BriefDescription": "Clockticks in the Memory Controller using one of the programmable counters", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS_P", "PerPkg": "1", "Unit": "iMC" }, { "BriefDescription": "This event is deprecated. Refer to new event UNC_M_CLOCKTICKS_P", + "EventCode": "0x0", "EventName": "UNC_M_DCLOCKTICKS", "PerPkg": "1", "Unit": "iMC"
tools/perf/pmu-events/arch/x86/broadwellx/uncore-power.json +1 -0
@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",
tools/perf/pmu-events/arch/x86/cascadelakex/pipeline.json +1 -0
@@ -191,6 +191,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overfl ow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.",
"SampleAfterValue": "2000003",
@@ -222,6 +223,7 @@
},
{
"BriefDescription": "Core cycles when the thread is not in halt state",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -230,6 +232,7 @@
{
"AnyThread": "1",
"BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.",
+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD_ANY",
"SampleAfterValue": "2000003",
"UMask": "0x2"
@@ -369,6 +372,7 @@
},
{
"BriefDescription": "Instructions retired from execution.",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.ANY",
"PublicDescription": "Counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, Counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. Counting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/cascadelakex/uncore-cache.json +2 -0
@@ -512,6 +512,7 @@ }, { "BriefDescription": "Uncore cache clock ticks", + "EventCode": "0x0", "EventName": "UNC_CHA_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts clockticks of the clock controlling the uncore caching and home agent (CHA).",
@@ -5792,6 +5793,7 @@ }, { "BriefDescription": "This event is deprecated. Refer to new event UNC_CHA_CLOCKTICKS", + "EventCode": "0x0", "Deprecated": "1", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1",
tools/perf/pmu-events/arch/x86/cascadelakex/uncore-interconnect.json +1 -0
@@ -1090,6 +1090,7 @@ }, { "BriefDescription": "Cycles - at UCLK", + "EventCode": "0x0", "EventName": "UNC_M2M_CLOCKTICKS", "PerPkg": "1", "Unit": "M2M"
tools/perf/pmu-events/arch/x86/cascadelakex/uncore-io.json +1 -0
@@ -1271,6 +1271,7 @@ }, { "BriefDescription": "Counting disabled", + "EventCode": "0x0", "EventName": "UNC_IIO_NOTHING", "PerPkg": "1", "Unit": "IIO"
tools/perf/pmu-events/arch/x86/cascadelakex/uncore-memory.json +1 -0
@@ -167,6 +167,7 @@ }, { "BriefDescription": "Memory controller clock ticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts clockticks of the fixed frequency clock of the memory controller using one of the programmable counters.",
tools/perf/pmu-events/arch/x86/cascadelakex/uncore-power.json +1 -0
@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",
tools/perf/pmu-events/arch/x86/elkhartlake/pipeline.json +2 -0
@@ -150,6 +150,7 @@ }, { "BriefDescription": "Counts the number of unhalted reference clock cycles at TSC frequency. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is not affected by core frequency changes and increments at a fixed frequency that is also used for the Time Stamp Counter (TSC). This event uses fixed counter 2.", "SampleAfterValue": "2000003",
@@ -179,6 +180,7 @@ }, { "BriefDescription": "Counts the total number of instructions retired. (Fixed event)", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the total number of instructions that retired. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. This event continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0.",
tools/perf/pmu-events/arch/x86/goldmont/pipeline.json +3 -0
@@ -143,6 +143,7 @@ }, { "BriefDescription": "Core cycles when core is not halted (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. In mobile systems the core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",
@@ -165,6 +166,7 @@ }, { "BriefDescription": "Reference cycles when core is not halted (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. In mobile systems the core frequency may change from time. This event is not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. This event uses fixed counter 2. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",
@@ -187,6 +189,7 @@ }, { "BriefDescription": "Instructions retired (Fixed event)", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "Counts the number of instructions that retire execution. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. The counter continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/goldmontplus/pipeline.json +3 -0
@@ -143,6 +143,7 @@ }, { "BriefDescription": "Core cycles when core is not halted (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. In mobile systems the core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",
@@ -165,6 +166,7 @@ }, { "BriefDescription": "Reference cycles when core is not halted (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. In mobile systems the core frequency may change from time. This event is not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. This event uses fixed counter 2. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",
@@ -187,6 +189,7 @@ }, { "BriefDescription": "Instructions retired (Fixed event)", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "2", "PublicDescription": "Counts the number of instructions that retire execution. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. The counter continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0. You cannot collect a PEBs record for this event.",
tools/perf/pmu-events/arch/x86/grandridge/pipeline.json +3 -0
@@ -29,6 +29,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3"
@@ -43,6 +44,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "SampleAfterValue": "2000003", "UMask": "0x2"
@@ -55,6 +57,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of instructions retired", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/graniterapids/pipeline.json +1 -0
@@ -17,6 +17,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.",
"SampleAfterValue": "2000003",
@@ -32,6 +33,7 @@
},
{
"BriefDescription": "Core cycles when the thread is not in halt state",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -46,6 +48,7 @@
},
{
"BriefDescription": "Number of instructions retired. Fixed Counter - architectural event",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.ANY",
"PEBS": "1",
"PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",
@@ -78,6 +81,7 @@
},
{
"BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event",+ "EventCode": "0x0",
"EventName": "TOPDOWN.SLOTS",
"PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",
tools/perf/pmu-events/arch/x86/haswell/pipeline.json +4 -0
@@ -303,6 +303,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state.", "SampleAfterValue": "2000003",
@@ -327,6 +328,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of thread cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. The core frequency may change from time to time due to power or thermal throttling.", "SampleAfterValue": "2000003",
@@ -335,6 +337,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"
@@ -436,6 +439,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "Errata": "HSD140, HSD143", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. INST_RETIRED.ANY is counted by a designated fixed counter, leaving the programmable counters available for other events. Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.",
tools/perf/pmu-events/arch/x86/haswellx/pipeline.json +4 -0
@@ -303,6 +303,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state.", "SampleAfterValue": "2000003",
@@ -327,6 +328,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of thread cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. The core frequency may change from time to time due to power or thermal throttling.", "SampleAfterValue": "2000003",
@@ -335,6 +337,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"
@@ -436,6 +439,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "Errata": "HSD140, HSD143", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. INST_RETIRED.ANY is counted by a designated fixed counter, leaving the programmable counters available for other events. Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.",
tools/perf/pmu-events/arch/x86/haswellx/uncore-cache.json +2 -0
@@ -183,6 +183,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"
@@ -1698,6 +1699,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",
tools/perf/pmu-events/arch/x86/haswellx/uncore-interconnect.json +1 -0
@@ -1,6 +1,7 @@ [ { "BriefDescription": "Number of non data (control) flits transmitted . Derived from unc_q_txl_flits_g0.non_data", + "EventCode": "0x0", "EventName": "QPI_CTL_BANDWIDTH_TX", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calc
ulate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",
@@ -10,6 +11,7 @@
},
{
"BriefDescription": "Number of data flits transmitted . Derived from unc_q_txl_flits_g0.data",+ "EventCode": "0x0",
"EventName": "QPI_DATA_BANDWIDTH_TX",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calc
ulate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",
@@ -37,6 +39,7 @@
},
{
"BriefDescription": "Clocks in the IRP",+ "EventCode": "0x0",
"EventName": "UNC_I_CLOCKTICKS",
"PerPkg": "1",
"PublicDescription": "Number of clocks in the IRP.",
@@ -1401,6 +1404,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 0; Data Tx Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G0.DATA",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calc
ulate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",
@@ -1409,6 +1413,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 0; Non-Data protocol Tx Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G0.NON_DATA",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calc
ulate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",
@@ -1417,6 +1422,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; DRS Flits (both Header and Data)",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.DRS",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency.",
@@ -1425,6 +1431,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; DRS Data Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.DRS_DATA",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of data flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the data flits (not the header).",
@@ -1433,6 +1440,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; DRS Header Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.DRS_NONDATA",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of protocol flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the header flits (not the data). This includes extended headers.",
@@ -1441,6 +1449,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; HOM Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.HOM",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of flits transmitted over QPI on the home channel.",
@@ -1449,6 +1458,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; HOM Non-Request Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.HOM_NONREQ",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of non-request flits transmitted over QPI on the home channel. These are most commonly snoop responses, and this event can be used as a proxy for that.",
@@ -1457,6 +1467,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; HOM Request Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.HOM_REQ",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of data request transmitted over QPI on the home channel. This basically counts the number of remote memory requests transmitted over QPI. In conjunction with the local read count in the Home Agent, one can calculate the number of LLC Misses.",
@@ -1465,6 +1476,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; SNP Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.SNP",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of snoop request flits transmitted over QPI. These requests are contained in the snoop channel. This does not include snoop responses, which are transmitted on the home channel.",
@@ -3136,6 +3148,7 @@
},
{
"BriefDescription": "Uncore Clocks",
+ "EventCode": "0x0",
"EventName": "UNC_S_CLOCKTICKS",
"PerPkg": "1",
"Unit": "SBOX"
@@ -3823,6 +3836,7 @@
},
{
"BriefDescription": "UNC_U_CLOCKTICKS",
+ "EventCode": "0x0",
"EventName": "UNC_U_CLOCKTICKS",
"PerPkg": "1",
"Unit": "UBOX"tools/perf/pmu-events/arch/x86/haswellx/uncore-memory.json +2 -0
@@ -151,12 +151,14 @@ }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "Unit": "iMC" }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_DCLOCKTICKS", "PerPkg": "1", "Unit": "iMC"
tools/perf/pmu-events/arch/x86/haswellx/uncore-power.json +1 -0
@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 800 MHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",
tools/perf/pmu-events/arch/x86/icelake/pipeline.json +1 -0
@@ -193,6 +193,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.",
"SampleAfterValue": "2000003",
@@ -208,6 +209,7 @@
},
{
"BriefDescription": "Core cycles when the thread is not in halt state",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -359,6 +361,7 @@
},
{
"BriefDescription": "Precise instruction retired event with a reduced effect of PEBS shadow in IP distribution",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.PREC_DIST",
"PEBS": "1",
"PublicDescription": "A version of INST_RETIRED that allows for a more unbiased distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR) feature to mitigate some bias in how retired instructions get sampled. Use on Fixed Counter 0.",
@@ -562,6 +565,7 @@
},
{
"BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event",+ "EventCode": "0x0",
"EventName": "TOPDOWN.SLOTS",
"PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",
tools/perf/pmu-events/arch/x86/icelakex/pipeline.json +1 -0
@@ -193,6 +193,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.",
"SampleAfterValue": "2000003",
@@ -208,6 +209,7 @@
},
{
"BriefDescription": "Core cycles when the thread is not in halt state",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -359,6 +361,7 @@
},
{
"BriefDescription": "Precise instruction retired event with a reduced effect of PEBS shadow in IP distribution",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.PREC_DIST",
"PEBS": "1",
"PublicDescription": "A version of INST_RETIRED that allows for a more unbiased distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR) feature to mitigate some bias in how retired instructions get sampled. Use on Fixed Counter 0.",
@@ -544,6 +547,7 @@
},
{
"BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event",+ "EventCode": "0x0",
"EventName": "TOPDOWN.SLOTS",
"PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",
tools/perf/pmu-events/arch/x86/icelakex/uncore-cache.json +1 -0
@@ -892,6 +892,7 @@ }, { "BriefDescription": "Clockticks of the uncore caching and home agent (CHA)", + "EventCode": "0x0", "EventName": "UNC_CHA_CLOCKTICKS", "PerPkg": "1", "Unit": "CHA"
tools/perf/pmu-events/arch/x86/icelakex/uncore-interconnect.json +1 -0
@@ -1419,6 +1419,7 @@ }, { "BriefDescription": "Clockticks of the mesh to memory (M2M)", + "EventCode": "0x0", "EventName": "UNC_M2M_CLOCKTICKS", "PerPkg": "1", "Unit": "M2M"
tools/perf/pmu-events/arch/x86/icelakex/uncore-memory.json +1 -0
@@ -100,6 +100,7 @@ }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "Unit": "iMC"
tools/perf/pmu-events/arch/x86/icelakex/uncore-power.json +1 -0
@@ -1,6 +1,7 @@ [ { "BriefDescription": "Clockticks of the power control unit (PCU)", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Clockticks of the power control unit (PCU) : The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",
tools/perf/pmu-events/arch/x86/ivybridge/pipeline.json +3 -0
@@ -326,6 +326,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3"
@@ -348,6 +349,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "SampleAfterValue": "2000003", "UMask": "0x2"
@@ -355,6 +357,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "PublicDescription": "Core cycles when at least one thread on the physical core is not in halt state.", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/ivytown/pipeline.json +4 -0
@@ -326,6 +326,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3"
@@ -348,6 +349,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "SampleAfterValue": "2000003", "UMask": "0x2"
@@ -355,6 +357,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "PublicDescription": "Core cycles when at least one thread on the physical core is not in halt state.", "SampleAfterValue": "2000003",
@@ -510,6 +513,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "SampleAfterValue": "2000003", "UMask": "0x1"
tools/perf/pmu-events/arch/x86/ivytown/uncore-cache.json +2 -0
@@ -1,6 +1,7 @@ [ { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"
@@ -1533,6 +1534,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",
tools/perf/pmu-events/arch/x86/ivytown/uncore-interconnect.json +2 -0
@@ -109,6 +109,7 @@ }, { "BriefDescription": "Clocks in the IRP", + "EventCode": "0x0", "EventName": "UNC_I_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Number of clocks in the IRP.",
@@ -1522,6 +1523,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Data Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calc
ulate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",
@@ -1530,6 +1532,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 0; Non-Data protocol Tx Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G0.NON_DATA",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calc
ulate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",
@@ -1538,6 +1541,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; DRS Flits (both Header and Data)",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.DRS",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency.",
@@ -1546,6 +1550,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; DRS Data Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.DRS_DATA",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of data flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the data flits (not the header).",
@@ -1554,6 +1559,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; DRS Header Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.DRS_NONDATA",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of protocol flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the header flits (not the data). This includes extended headers.",
@@ -1562,6 +1568,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; HOM Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.HOM",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of flits transmitted over QPI on the home channel.",
@@ -1570,6 +1577,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; HOM Non-Request Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.HOM_NONREQ",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of non-request flits transmitted over QPI on the home channel. These are most commonly snoop responses, and this event can be used as a proxy for that.",
@@ -1578,6 +1586,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; HOM Request Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.HOM_REQ",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of data request transmitted over QPI on the home channel. This basically counts the number of remote memory requests transmitted over QPI. In conjunction with the local read count in the Home Agent, one can calculate the number of LLC Misses.",
@@ -1586,6 +1595,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; SNP Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.SNP",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual
data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of snoop request flits transmitted over QPI. These requests are contained in the snoop channel. This does not include snoop responses, which are transmitted on the home channel.",
@@ -3104,6 +3114,7 @@
},
{
"EventName": "UNC_U_CLOCKTICKS",
+ "EventCode": "0x0",
"PerPkg": "1",
"Unit": "UBOX"
},tools/perf/pmu-events/arch/x86/ivytown/uncore-memory.json +1 -0
@@ -131,6 +131,7 @@ }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_DCLOCKTICKS", "PerPkg": "1", "Unit": "iMC"
tools/perf/pmu-events/arch/x86/ivytown/uncore-power.json +1 -0
@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 800 MHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",
tools/perf/pmu-events/arch/x86/jaketown/pipeline.json +4 -0
@@ -329,6 +329,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -351,6 +352,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -359,6 +361,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"
@@ -432,6 +435,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers.", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/jaketown/uncore-cache.json +2 -0
@@ -1,6 +1,7 @@ [ { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"
@@ -863,6 +864,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",
tools/perf/pmu-events/arch/x86/jaketown/uncore-interconnect.json +2 -0
@@ -109,6 +109,7 @@ }, { "BriefDescription": "Clocks in the IRP", + "EventCode": "0x0", "EventName": "UNC_I_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Number of clocks in the IRP.",
@@ -847,6 +848,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Data Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other infor
mation. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.",
@@ -855,6 +857,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 0; Idle and Null Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G0.IDLE",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other infor
mation. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.",
@@ -863,6 +866,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 0; Non-Data protocol Tx Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G0.NON_DATA",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other infor
mation. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.",
@@ -871,6 +875,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; DRS Flits (both Header and Data)",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.DRS",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 6
4 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",
@@ -879,6 +884,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; DRS Data Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.DRS_DATA",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 6
4 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",
@@ -887,6 +893,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; DRS Header Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.DRS_NONDATA",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 6
4 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",
@@ -895,6 +902,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; HOM Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.HOM",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 6
4 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",
@@ -903,6 +911,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; HOM Non-Request Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.HOM_NONREQ",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 6
4 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",
@@ -911,6 +920,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; HOM Request Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.HOM_REQ",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 6
4 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",
@@ -919,6 +929,7 @@
},
{
"BriefDescription": "Flits Transferred - Group 1; SNP Flits",
+ "EventCode": "0x0",
"EventName": "UNC_Q_TxL_FLITS_G1.SNP",
"PerPkg": "1",
"PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 6
4 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",
@@ -1576,6 +1587,7 @@
},
{
"EventName": "UNC_U_CLOCKTICKS",
+ "EventCode": "0x0",
"PerPkg": "1",
"Unit": "UBOX"
},tools/perf/pmu-events/arch/x86/jaketown/uncore-memory.json +1 -0
@@ -65,6 +65,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Uncore Fixed Counter - uclks",
tools/perf/pmu-events/arch/x86/jaketown/uncore-power.json +2 -0
@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 800 MHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",
@@ -216,6 +217,7 @@ }, { "BriefDescription": "Cycles spent changing Frequency", + "EventCode": "0x0", "EventName": "UNC_P_FREQ_TRANS_CYCLES", "PerPkg": "1", "PublicDescription": "Counts the number of cycles when the system is changing frequency. This can not be filtered by thread ID. One can also use it with the occupancy counter that monitors number of threads in C0 to estimate the performance impact that frequency transitions had on the system.",
tools/perf/pmu-events/arch/x86/knightslanding/pipeline.json +3 -0
@@ -150,12 +150,14 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3" }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter", "SampleAfterValue": "2000003",
@@ -177,6 +179,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of instructions retired", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions that retire. For instructions that consist of multiple micro-ops, this event counts exactly once, as the last micro-op of the instruction retires. The event continues counting while instructions retire, including during interrupt service routines caused by hardware interrupts, faults or traps.", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/knightslanding/uncore-cache.json +1 -0
@@ -3246,6 +3246,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_H_U_CLOCKTICKS", "PerPkg": "1", "Unit": "CHA"
tools/perf/pmu-events/arch/x86/knightslanding/uncore-memory.json +4 -0
@@ -41,6 +41,7 @@ }, { "BriefDescription": "ECLK count", + "EventCode": "0x0", "EventName": "UNC_E_E_CLOCKTICKS", "PerPkg": "1", "Unit": "EDC_ECLK"
@@ -55,6 +56,7 @@ }, { "BriefDescription": "UCLK count", + "EventCode": "0x0", "EventName": "UNC_E_U_CLOCKTICKS", "PerPkg": "1", "Unit": "EDC_UCLK"
@@ -93,12 +95,14 @@ }, { "BriefDescription": "DCLK count", + "EventCode": "0x0", "EventName": "UNC_M_D_CLOCKTICKS", "PerPkg": "1", "Unit": "iMC_DCLK" }, { "BriefDescription": "UCLK count", + "EventCode": "0x0", "EventName": "UNC_M_U_CLOCKTICKS", "PerPkg": "1", "Unit": "iMC_UCLK"
tools/perf/pmu-events/arch/x86/meteorlake/pipeline.json +3 -0
@@ -37,6 +37,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "SampleAfterValue": "2000003", "UMask": "0x2",
@@ -51,6 +52,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3",
@@ -58,6 +60,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.",
"SampleAfterValue": "2000003",
@@ -75,6 +78,7 @@
},
{
"BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles",
+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"SampleAfterValue": "2000003",
"UMask": "0x2",
@@ -82,6 +86,7 @@
},
{
"BriefDescription": "Core cycles when the thread is not in halt state",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -105,6 +110,7 @@
},
{
"BriefDescription": "Fixed Counter: Counts the number of instructions retired",
+ "EventCode": "0x0",
"EventName": "INST_RETIRED.ANY",
"PEBS": "1",
"SampleAfterValue": "2000003",
@@ -113,6 +119,7 @@
},
{
"BriefDescription": "Number of instructions retired. Fixed Counter - architectural event",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.ANY",
"PEBS": "1",
"PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",
@@ -157,6 +164,7 @@
},
{
"BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event",+ "EventCode": "0x0",
"EventName": "TOPDOWN.SLOTS",
"PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",
tools/perf/pmu-events/arch/x86/sandybridge/pipeline.json +4 -0
@@ -337,6 +337,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -359,6 +360,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -367,6 +369,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"
@@ -440,6 +443,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers.", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/sapphirerapids/pipeline.json +1 -0
@@ -284,6 +284,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.",
"SampleAfterValue": "2000003",
@@ -299,6 +300,7 @@
},
{
"BriefDescription": "Core cycles when the thread is not in halt state",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -426,6 +428,7 @@
},
{
"BriefDescription": "Number of instructions retired. Fixed Counter - architectural event",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.ANY",
"PEBS": "1",
"PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",
@@ -457,6 +460,7 @@
},
{
"BriefDescription": "Precise instruction retired with PEBS precise-distribution",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.PREC_DIST",
"PEBS": "1",
"PublicDescription": "A version of INST_RETIRED that allows for a precise distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR++) feature to fix bias in how retired instructions get sampled. Use on Fixed Counter 0.",
@@ -719,6 +723,7 @@
},
{
"BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event",+ "EventCode": "0x0",
"EventName": "TOPDOWN.SLOTS",
"PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",
tools/perf/pmu-events/arch/x86/sierraforest/pipeline.json +4 -0
@@ -17,6 +17,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "SampleAfterValue": "2000003", "UMask": "0x2"
@@ -29,6 +30,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3"
@@ -43,6 +45,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "SampleAfterValue": "2000003", "UMask": "0x2"
@@ -55,6 +58,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of instructions retired", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/silvermont/pipeline.json +1 -0
@@ -143,6 +143,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. In systems with a constant core frequency, this event can give you a measurement of the elapsed time while the core was not in halt state by dividing the event count by the core frequency. This event is architecturally defined and is a designated fixed counter. CPU_CLK_UNHALTED.CORE and CPU_CLK_UNHALTED.CORE_P use the core frequency which may change from time to time. CPU_CLK_UNHALTE.REF_TSC and CPU_CLK_UNHALTED.REF are not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. The fixed events are CPU_CLK_UNHALTED.CORE and CPU_CLK_UNHALTED.REF_TSC and the programmable events are C PU_CLK_UNHALTED.CORE_P and CPU_CLK_UNHALTED.REF.",
"SampleAfterValue": "2000003",
@@ -165,6 +166,7 @@
},
{
"BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.REF_TSC",
"PublicDescription": "Counts the number of reference cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time. This event is not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. Divide this event count by core frequency to determine the elapsed time while the core was not in halt state. Divide this event count by core frequency to determine the elapsed time while the core was not in halt state. This event is architecturally defined and is a designated fixed counter. CPU_CLK_UNHALTED.CORE and CPU_CLK_UNHALTED.CORE_P use the core frequency which may change from time to time. CPU_CLK_UNHALTE.REF_TSC and CPU_CLK_UNHALTED.REF are not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. The fixed events are
CPU_CLK_UNHALTED.CORE and CPU_CLK_UNHALTED.REF_TSC and the programmable events are CPU_CLK_UNHALTED.CORE_P and CPU_CLK_UNHALTED.REF.",
"SampleAfterValue": "2000003",
@@ -180,6 +182,7 @@
},
{
"BriefDescription": "Fixed Counter: Counts the number of instructions retired",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.ANY",
"PublicDescription": "This event counts the number of instructions that retire. For instructions that consist of multiple micro-ops, this event counts exactly once, as the last micro-op of the instruction retires. The event continues counting while instructions retire, including during interrupt service routines caused by hardware interrupts, faults or traps. Background: Modern microprocessors employ extensive pipelining and speculative techniques. Since sometimes an instruction is started but never completed, the notion of \"retirement\" is introduced. A retired instruction is one that commits its states. Or stated differently, an instruction might be abandoned at some point. No instruction is truly finished until it retires. This counter measures the number of completed instructions. The fixed event is INST_RETIRED.ANY and the programmable event is INST_RETIRED.ANY_P.", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/skylake/pipeline.json +1 -0
@@ -182,6 +182,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overfl ow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.",
"SampleAfterValue": "2000003",
@@ -213,6 +214,7 @@
},
{
"BriefDescription": "Core cycles when the thread is not in halt state",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -221,6 +223,7 @@
{
"AnyThread": "1",
"BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.",
+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD_ANY",
"SampleAfterValue": "2000003",
"UMask": "0x2"
@@ -360,6 +363,7 @@
},
{
"BriefDescription": "Instructions retired from execution.",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.ANY",
"PublicDescription": "Counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, Counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. Counting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/skylakex/pipeline.json +1 -0
@@ -191,6 +191,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overfl ow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.",
"SampleAfterValue": "2000003",
@@ -222,6 +223,7 @@
},
{
"BriefDescription": "Core cycles when the thread is not in halt state",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -230,6 +232,7 @@
{
"AnyThread": "1",
"BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.",
+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD_ANY",
"SampleAfterValue": "2000003",
"UMask": "0x2"
@@ -369,6 +372,7 @@
},
{
"BriefDescription": "Instructions retired from execution.",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.ANY",
"PublicDescription": "Counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, Counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. Counting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",
tools/perf/pmu-events/arch/x86/skylakex/uncore-cache.json +2 -0
@@ -460,6 +460,7 @@ }, { "BriefDescription": "Clockticks of the uncore caching & home agent (CHA)", + "EventCode": "0x0", "EventName": "UNC_CHA_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts clockticks of the clock controlling the uncore caching and home agent (CHA).",
@@ -5678,6 +5679,7 @@ { "BriefDescription": "This event is deprecated. Refer to new event UNC_CHA_CLOCKTICKS", "Deprecated": "1", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CHA"
tools/perf/pmu-events/arch/x86/skylakex/uncore-interconnect.json +1 -0
@@ -1090,6 +1090,7 @@ }, { "BriefDescription": "Cycles - at UCLK", + "EventCode": "0x0", "EventName": "UNC_M2M_CLOCKTICKS", "PerPkg": "1", "Unit": "M2M"
tools/perf/pmu-events/arch/x86/skylakex/uncore-io.json +1 -0
@@ -1271,6 +1271,7 @@ }, { "BriefDescription": "Counting disabled", + "EventCode": "0x0", "EventName": "UNC_IIO_NOTHING", "PerPkg": "1", "Unit": "IIO"
tools/perf/pmu-events/arch/x86/skylakex/uncore-memory.json +1 -0
@@ -167,6 +167,7 @@ }, { "BriefDescription": "Memory controller clock ticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts clockticks of the fixed frequency clock of the memory controller using one of the programmable counters.",
tools/perf/pmu-events/arch/x86/skylakex/uncore-power.json +1 -0
@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",
tools/perf/pmu-events/arch/x86/snowridgex/pipeline.json +2 -0
@@ -150,6 +150,7 @@ }, { "BriefDescription": "Counts the number of unhalted reference clock cycles at TSC frequency. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is not affected by core frequency changes and increments at a fixed frequency that is also used for the Time Stamp Counter (TSC). This event uses fixed counter 2.", "SampleAfterValue": "2000003",
@@ -179,6 +180,7 @@ }, { "BriefDescription": "Counts the total number of instructions retired. (Fixed event)", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the total number of instructions that retired. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. This event continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0.",
tools/perf/pmu-events/arch/x86/snowridgex/uncore-cache.json +1 -0
@@ -872,6 +872,7 @@ }, { "BriefDescription": "Uncore cache clock ticks", + "EventCode": "0x0", "EventName": "UNC_CHA_CLOCKTICKS", "PerPkg": "1", "Unit": "CHA"
tools/perf/pmu-events/arch/x86/snowridgex/uncore-interconnect.json +1 -0
@@ -1419,6 +1419,7 @@ }, { "BriefDescription": "Clockticks of the mesh to memory (M2M)", + "EventCode": "0x0", "EventName": "UNC_M2M_CLOCKTICKS", "PerPkg": "1", "Unit": "M2M"
tools/perf/pmu-events/arch/x86/snowridgex/uncore-memory.json +1 -0
@@ -120,6 +120,7 @@ }, { "BriefDescription": "Memory controller clock ticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Clockticks of the integrated memory controller (IMC)",
tools/perf/pmu-events/arch/x86/snowridgex/uncore-power.json +1 -0
@@ -1,6 +1,7 @@ [ { "BriefDescription": "Clockticks of the power control unit (PCU)", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "Unit": "PCU"
tools/perf/pmu-events/arch/x86/tigerlake/pipeline.json +1 -0
@@ -193,6 +193,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.",
"SampleAfterValue": "2000003",
@@ -208,6 +209,7 @@
},
{
"BriefDescription": "Core cycles when the thread is not in halt state",+ "EventCode": "0x0",
"EventName": "CPU_CLK_UNHALTED.THREAD",
"PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",
@@ -352,6 +354,7 @@
},
{
"BriefDescription": "Number of instructions retired. Fixed Counter - architectural event",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.ANY",
"PEBS": "1",
"PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",
@@ -377,6 +380,7 @@
},
{
"BriefDescription": "Precise instruction retired event with a reduced effect of PEBS shadow in IP distribution",+ "EventCode": "0x0",
"EventName": "INST_RETIRED.PREC_DIST",
"PEBS": "1",
"PublicDescription": "A version of INST_RETIRED that allows for a more unbiased distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR) feature to mitigate some bias in how retired instructions get sampled. Use on Fixed Counter 0.",
@@ -569,6 +573,7 @@
},
{
"BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event",+ "EventCode": "0x0",
"EventName": "TOPDOWN.SLOTS",
"PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",
Signature · 1 line
1.8.3.1
Ian Rogers
reviewerRe: [PATCH v7 3/8] perf vendor events: Supplement the omitted EventCode
Quoted 9 lines · Jing Zhang
If there is an "event=0" in the event description, the EventCode can be omitted in the JSON file, and
jevent.pywill automatically fill in "event=0" during parsing.However, for some events where EventCode and ConfigCode are missing, it is not necessary to automatically fill in "event=0", such as the CMN event description which is typically "type=xxx, eventid=xxx".
Quoted 5 lines
Therefore, before modifying
jevent.pyto prevent it from automatically adding "event=0" by default, it is necessary to fill in all omitted EventCodes first.Signed-off-by Jing Zhang <[email protected]>
As these files are generated, the generator script needs updating. However, I don't think this change makes sense as the event=0 is overwritten in the case of an arch_std event:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/pmu-events/jevents.py?h=perf-tools-next#n369
So yes event=0 was filled in, but it was then overwritten.
Thanks,
Ian
Quoted 2,030 lines
--- tools/perf/pmu-events/arch/x86/alderlake/pipeline.json | 9 +++++++++ tools/perf/pmu-events/arch/x86/alderlaken/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/broadwell/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/broadwellde/pipeline.json | 4 ++++ .../perf/pmu-events/arch/x86/broadwellde/uncore-cache.json | 2 ++ .../arch/x86/broadwellde/uncore-interconnect.json | 1 + .../pmu-events/arch/x86/broadwellde/uncore-memory.json | 1 + .../perf/pmu-events/arch/x86/broadwellde/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/broadwellx/pipeline.json | 4 ++++ .../perf/pmu-events/arch/x86/broadwellx/uncore-cache.json | 2 ++ .../arch/x86/broadwellx/uncore-interconnect.json | 13 +++++++++++++ .../perf/pmu-events/arch/x86/broadwellx/uncore-memory.json | 2 ++ .../perf/pmu-events/arch/x86/broadwellx/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/cascadelakex/pipeline.json | 4 ++++ .../pmu-events/arch/x86/cascadelakex/uncore-cache.json | 2 ++ .../arch/x86/cascadelakex/uncore-interconnect.json | 1 + tools/perf/pmu-events/arch/x86/cascadelakex/uncore-io.json | 1 + .../pmu-events/arch/x86/cascadelakex/uncore-memory.json | 1 + .../pmu-events/arch/x86/cascadelakex/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/elkhartlake/pipeline.json | 2 ++ tools/perf/pmu-events/arch/x86/goldmont/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/goldmontplus/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/grandridge/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/graniterapids/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/haswell/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/haswellx/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/haswellx/uncore-cache.json | 2 ++ .../pmu-events/arch/x86/haswellx/uncore-interconnect.json | 14 ++++++++++++++ tools/perf/pmu-events/arch/x86/haswellx/uncore-memory.json | 2 ++ tools/perf/pmu-events/arch/x86/haswellx/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/icelake/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/icelakex/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/icelakex/uncore-cache.json | 1 + .../pmu-events/arch/x86/icelakex/uncore-interconnect.json | 1 + tools/perf/pmu-events/arch/x86/icelakex/uncore-memory.json | 1 + tools/perf/pmu-events/arch/x86/icelakex/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/ivybridge/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/ivytown/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/ivytown/uncore-cache.json | 2 ++ .../pmu-events/arch/x86/ivytown/uncore-interconnect.json | 11 +++++++++++ tools/perf/pmu-events/arch/x86/ivytown/uncore-memory.json | 1 + tools/perf/pmu-events/arch/x86/ivytown/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/jaketown/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/jaketown/uncore-cache.json | 2 ++ .../pmu-events/arch/x86/jaketown/uncore-interconnect.json | 12 ++++++++++++ tools/perf/pmu-events/arch/x86/jaketown/uncore-memory.json | 1 + tools/perf/pmu-events/arch/x86/jaketown/uncore-power.json | 2 ++ .../perf/pmu-events/arch/x86/knightslanding/pipeline.json | 3 +++ .../pmu-events/arch/x86/knightslanding/uncore-cache.json | 1 + .../pmu-events/arch/x86/knightslanding/uncore-memory.json | 4 ++++ tools/perf/pmu-events/arch/x86/meteorlake/pipeline.json | 8 ++++++++ tools/perf/pmu-events/arch/x86/sandybridge/pipeline.json | 4 ++++ .../perf/pmu-events/arch/x86/sapphirerapids/pipeline.json | 5 +++++ tools/perf/pmu-events/arch/x86/sierraforest/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/silvermont/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/skylake/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/skylakex/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/skylakex/uncore-cache.json | 2 ++ .../pmu-events/arch/x86/skylakex/uncore-interconnect.json | 1 + tools/perf/pmu-events/arch/x86/skylakex/uncore-io.json | 1 + tools/perf/pmu-events/arch/x86/skylakex/uncore-memory.json | 1 + tools/perf/pmu-events/arch/x86/skylakex/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/snowridgex/pipeline.json | 2 ++ .../perf/pmu-events/arch/x86/snowridgex/uncore-cache.json | 1 + .../arch/x86/snowridgex/uncore-interconnect.json | 1 + .../perf/pmu-events/arch/x86/snowridgex/uncore-memory.json | 1 + .../perf/pmu-events/arch/x86/snowridgex/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/tigerlake/pipeline.json | 5 +++++ 68 files changed, 211 insertions(+)tools/perf/pmu-events/arch/x86/alderlake/pipeline.json +9
-0@@ -489,6 +489,7 @@ }, { "BriefDescription": "Counts the number of unhalted core clock cycles. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1.", "SampleAfterValue": "2000003",@@ -550,6 +551,7 @@ }, { "BriefDescription": "Counts the number of unhalted reference clock cycles at TSC frequency. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is not affected by core frequency changes and increments at a fixed frequency that is also used for the Time Stamp Counter (TSC). This event uses fixed counter 2.", "SampleAfterValue": "2000003",@@ -558,6 +560,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -584,6 +587,7 @@ }, { "BriefDescription": "Counts the number of unhalted core clock cycles. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1.", "SampleAfterValue": "2000003",@@ -592,6 +596,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -743,6 +748,7 @@ }, { "BriefDescription": "Counts the total number of instructions retired. (Fixed event)", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the total number of instructions that retired. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. This event continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0.",@@ -752,6 +758,7 @@ }, { "BriefDescription": "Number of instructions retired. Fixed Counter - architectural event", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",@@ -796,6 +803,7 @@ }, { "BriefDescription": "Precise instruction retired with PEBS precise-distribution", + "EventCode": "0x0", "EventName": "INST_RETIRED.PREC_DIST", "PEBS": "1", "PublicDescription": "A version of INST_RETIRED that allows for a precise distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR++) feature to fix bias in how retired instructions get sampled. Use on Fixed Counter 0.",@@ -1160,6 +1168,7 @@ }, { "BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event", + "EventCode": "0x0", "EventName": "TOPDOWN.SLOTS", "PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",tools/perf/pmu-events/arch/x86/alderlaken/pipeline.json +3
-0@@ -211,6 +211,7 @@ }, { "BriefDescription": "Counts the number of unhalted core clock cycles. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1.", "SampleAfterValue": "2000003",@@ -225,6 +226,7 @@ }, { "BriefDescription": "Counts the number of unhalted reference clock cycles at TSC frequency. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is not affected by core frequency changes and increments at a fixed frequency that is also used for the Time Stamp Counter (TSC). This event uses fixed counter 2.", "SampleAfterValue": "2000003",@@ -240,6 +242,7 @@ }, { "BriefDescription": "Counts the number of unhalted core clock cycles. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/broadwell/pipeline.json +4
-0@@ -336,6 +336,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. \nNote: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. This event is clocked by base clock (100 Mhz) on Sandy Bridge. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -359,6 +360,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -366,6 +368,7 @@ }, { "AnyThread": "1", + "EventCode": "0x0", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003",@@ -514,6 +517,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. \nNotes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. \nCounting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/broadwellde/pipeline.json +4
-0@@ -336,6 +336,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. \nNote: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. This event is clocked by base clock (100 Mhz) on Sandy Bridge. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -359,6 +360,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -367,6 +369,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -514,6 +517,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. \nNotes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. \nCounting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/broadwellde/uncore-cache.json +2
-0@@ -8,6 +8,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"@@ -1501,6 +1502,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",tools/perf/pmu-events/arch/x86/broadwellde/uncore-interconnect.json +1
-0@@ -19,6 +19,7 @@ }, { "BriefDescription": "Clocks in the IRP", + "EventCode": "0x0", "EventName": "UNC_I_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Number of clocks in the IRP.",tools/perf/pmu-events/arch/x86/broadwellde/uncore-memory.json +1
-0@@ -131,6 +131,7 @@ }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_DCLOCKTICKS", "PerPkg": "1", "Unit": "iMC"tools/perf/pmu-events/arch/x86/broadwellde/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",tools/perf/pmu-events/arch/x86/broadwellx/pipeline.json +4
-0@@ -336,6 +336,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. \nNote: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. This event is clocked by base clock (100 Mhz) on Sandy Bridge. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -359,6 +360,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -367,6 +369,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -514,6 +517,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. \nNotes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. \nCounting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/broadwellx/uncore-cache.json +2
-0@@ -183,6 +183,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"@@ -1689,6 +1690,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",tools/perf/pmu-events/arch/x86/broadwellx/uncore-interconnect.json +13
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "Number of non data (control) flits transmitted . Derived from unc_q_txl_flits_g0.non_data", + "EventCode": "0x0", "EventName": "QPI_CTL_BANDWIDTH_TX", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",@@ -10,6 +11,7 @@ }, { "BriefDescription": "Number of data flits transmitted . Derived from unc_q_txl_flits_g0.data", + "EventCode": "0x0", "EventName": "QPI_DATA_BANDWIDTH_TX", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",@@ -37,6 +39,7 @@ }, { "BriefDescription": "Clocks in the IRP", + "EventCode": "0x0", "EventName": "UNC_I_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Number of clocks in the IRP.",@@ -1400,6 +1403,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Data Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",@@ -1408,6 +1412,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Non-Data protocol Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.NON_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",@@ -1416,6 +1421,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Flits (both Header and Data)", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency.",@@ -1424,6 +1430,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Data Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of data flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the data flits (not the header).",@@ -1432,6 +1439,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Header Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_NONDATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of protocol flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the header flits (not the data). This includes extended headers.",@@ -1440,6 +1448,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of flits transmitted over QPI on the home channel.",@@ -1448,6 +1457,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Non-Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_NONREQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of non-request flits transmitted over QPI on the home channel. These are most commonly snoop responses, and this event can be used as a proxy for that.",@@ -1456,6 +1466,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_REQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of data request transmitted over QPI on the home channel. This basically counts the number of remote memory requests transmitted over QPI. In conjunction with the local read count in the Home Agent, one can calculate the number of LLC Misses.",@@ -1464,6 +1475,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; SNP Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.SNP", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of snoop request flits transmitted over QPI. These requests are contained in the snoop channel. This does not include snoop responses, which are transmitted on the home channel.",@@ -3162,6 +3174,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_S_CLOCKTICKS", "PerPkg": "1", "Unit": "SBOX"tools/perf/pmu-events/arch/x86/broadwellx/uncore-memory.json +2
-0@@ -158,12 +158,14 @@ }, { "BriefDescription": "Clockticks in the Memory Controller using one of the programmable counters", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS_P", "PerPkg": "1", "Unit": "iMC" }, { "BriefDescription": "This event is deprecated. Refer to new event UNC_M_CLOCKTICKS_P", + "EventCode": "0x0", "EventName": "UNC_M_DCLOCKTICKS", "PerPkg": "1", "Unit": "iMC"tools/perf/pmu-events/arch/x86/broadwellx/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",tools/perf/pmu-events/arch/x86/cascadelakex/pipeline.json +4
-0@@ -191,6 +191,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -222,6 +223,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -230,6 +232,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -369,6 +372,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "Counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, Counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. Counting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/cascadelakex/uncore-cache.json +2
-0@@ -512,6 +512,7 @@ }, { "BriefDescription": "Uncore cache clock ticks", + "EventCode": "0x0", "EventName": "UNC_CHA_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts clockticks of the clock controlling the uncore caching and home agent (CHA).",@@ -5792,6 +5793,7 @@ }, { "BriefDescription": "This event is deprecated. Refer to new event UNC_CHA_CLOCKTICKS", + "EventCode": "0x0", "Deprecated": "1", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1",tools/perf/pmu-events/arch/x86/cascadelakex/uncore-interconnect.json +1
-0@@ -1090,6 +1090,7 @@ }, { "BriefDescription": "Cycles - at UCLK", + "EventCode": "0x0", "EventName": "UNC_M2M_CLOCKTICKS", "PerPkg": "1", "Unit": "M2M"tools/perf/pmu-events/arch/x86/cascadelakex/uncore-io.json +1
-0@@ -1271,6 +1271,7 @@ }, { "BriefDescription": "Counting disabled", + "EventCode": "0x0", "EventName": "UNC_IIO_NOTHING", "PerPkg": "1", "Unit": "IIO"tools/perf/pmu-events/arch/x86/cascadelakex/uncore-memory.json +1
-0@@ -167,6 +167,7 @@ }, { "BriefDescription": "Memory controller clock ticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts clockticks of the fixed frequency clock of the memory controller using one of the programmable counters.",tools/perf/pmu-events/arch/x86/cascadelakex/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",tools/perf/pmu-events/arch/x86/elkhartlake/pipeline.json +2
-0@@ -150,6 +150,7 @@ }, { "BriefDescription": "Counts the number of unhalted reference clock cycles at TSC frequency. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is not affected by core frequency changes and increments at a fixed frequency that is also used for the Time Stamp Counter (TSC). This event uses fixed counter 2.", "SampleAfterValue": "2000003",@@ -179,6 +180,7 @@ }, { "BriefDescription": "Counts the total number of instructions retired. (Fixed event)", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the total number of instructions that retired. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. This event continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0.",tools/perf/pmu-events/arch/x86/goldmont/pipeline.json +3
-0@@ -143,6 +143,7 @@ }, { "BriefDescription": "Core cycles when core is not halted (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. In mobile systems the core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",@@ -165,6 +166,7 @@ }, { "BriefDescription": "Reference cycles when core is not halted (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. In mobile systems the core frequency may change from time. This event is not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. This event uses fixed counter 2. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",@@ -187,6 +189,7 @@ }, { "BriefDescription": "Instructions retired (Fixed event)", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "Counts the number of instructions that retire execution. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. The counter continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/goldmontplus/pipeline.json +3
-0@@ -143,6 +143,7 @@ }, { "BriefDescription": "Core cycles when core is not halted (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. In mobile systems the core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",@@ -165,6 +166,7 @@ }, { "BriefDescription": "Reference cycles when core is not halted (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. In mobile systems the core frequency may change from time. This event is not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. This event uses fixed counter 2. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",@@ -187,6 +189,7 @@ }, { "BriefDescription": "Instructions retired (Fixed event)", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "2", "PublicDescription": "Counts the number of instructions that retire execution. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. The counter continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0. You cannot collect a PEBs record for this event.",tools/perf/pmu-events/arch/x86/grandridge/pipeline.json +3
-0@@ -29,6 +29,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3"@@ -43,6 +44,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -55,6 +57,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of instructions retired", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/graniterapids/pipeline.json +4
-0@@ -17,6 +17,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -32,6 +33,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -46,6 +48,7 @@ }, { "BriefDescription": "Number of instructions retired. Fixed Counter - architectural event", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",@@ -78,6 +81,7 @@ }, { "BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event", + "EventCode": "0x0", "EventName": "TOPDOWN.SLOTS", "PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",tools/perf/pmu-events/arch/x86/haswell/pipeline.json +4
-0@@ -303,6 +303,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state.", "SampleAfterValue": "2000003",@@ -327,6 +328,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of thread cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. The core frequency may change from time to time due to power or thermal throttling.", "SampleAfterValue": "2000003",@@ -335,6 +337,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -436,6 +439,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "Errata": "HSD140, HSD143", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. INST_RETIRED.ANY is counted by a designated fixed counter, leaving the programmable counters available for other events. Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.",tools/perf/pmu-events/arch/x86/haswellx/pipeline.json +4
-0@@ -303,6 +303,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state.", "SampleAfterValue": "2000003",@@ -327,6 +328,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of thread cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. The core frequency may change from time to time due to power or thermal throttling.", "SampleAfterValue": "2000003",@@ -335,6 +337,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -436,6 +439,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "Errata": "HSD140, HSD143", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. INST_RETIRED.ANY is counted by a designated fixed counter, leaving the programmable counters available for other events. Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.",tools/perf/pmu-events/arch/x86/haswellx/uncore-cache.json +2
-0@@ -183,6 +183,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"@@ -1698,6 +1699,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",tools/perf/pmu-events/arch/x86/haswellx/uncore-interconnect.json +14
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "Number of non data (control) flits transmitted . Derived from unc_q_txl_flits_g0.non_data", + "EventCode": "0x0", "EventName": "QPI_CTL_BANDWIDTH_TX", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",@@ -10,6 +11,7 @@ }, { "BriefDescription": "Number of data flits transmitted . Derived from unc_q_txl_flits_g0.data", + "EventCode": "0x0", "EventName": "QPI_DATA_BANDWIDTH_TX", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",@@ -37,6 +39,7 @@ }, { "BriefDescription": "Clocks in the IRP", + "EventCode": "0x0", "EventName": "UNC_I_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Number of clocks in the IRP.",@@ -1401,6 +1404,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Data Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",@@ -1409,6 +1413,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Non-Data protocol Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.NON_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",@@ -1417,6 +1422,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Flits (both Header and Data)", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency.",@@ -1425,6 +1431,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Data Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of data flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the data flits (not the header).",@@ -1433,6 +1440,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Header Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_NONDATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of protocol flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the header flits (not the data). This includes extended headers.",@@ -1441,6 +1449,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of flits transmitted over QPI on the home channel.",@@ -1449,6 +1458,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Non-Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_NONREQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of non-request flits transmitted over QPI on the home channel. These are most commonly snoop responses, and this event can be used as a proxy for that.",@@ -1457,6 +1467,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_REQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of data request transmitted over QPI on the home channel. This basically counts the number of remote memory requests transmitted over QPI. In conjunction with the local read count in the Home Agent, one can calculate the number of LLC Misses.",@@ -1465,6 +1476,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; SNP Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.SNP", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of snoop request flits transmitted over QPI. These requests are contained in the snoop channel. This does not include snoop responses, which are transmitted on the home channel.",@@ -3136,6 +3148,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_S_CLOCKTICKS", "PerPkg": "1", "Unit": "SBOX"@@ -3823,6 +3836,7 @@ }, { "BriefDescription": "UNC_U_CLOCKTICKS", + "EventCode": "0x0", "EventName": "UNC_U_CLOCKTICKS", "PerPkg": "1", "Unit": "UBOX"tools/perf/pmu-events/arch/x86/haswellx/uncore-memory.json +2
-0@@ -151,12 +151,14 @@ }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "Unit": "iMC" }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_DCLOCKTICKS", "PerPkg": "1", "Unit": "iMC"tools/perf/pmu-events/arch/x86/haswellx/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 800 MHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",tools/perf/pmu-events/arch/x86/icelake/pipeline.json +4
-0@@ -193,6 +193,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -208,6 +209,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -359,6 +361,7 @@ }, { "BriefDescription": "Precise instruction retired event with a reduced effect of PEBS shadow in IP distribution", + "EventCode": "0x0", "EventName": "INST_RETIRED.PREC_DIST", "PEBS": "1", "PublicDescription": "A version of INST_RETIRED that allows for a more unbiased distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR) feature to mitigate some bias in how retired instructions get sampled. Use on Fixed Counter 0.",@@ -562,6 +565,7 @@ }, { "BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event", + "EventCode": "0x0", "EventName": "TOPDOWN.SLOTS", "PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",tools/perf/pmu-events/arch/x86/icelakex/pipeline.json +4
-0@@ -193,6 +193,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -208,6 +209,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -359,6 +361,7 @@ }, { "BriefDescription": "Precise instruction retired event with a reduced effect of PEBS shadow in IP distribution", + "EventCode": "0x0", "EventName": "INST_RETIRED.PREC_DIST", "PEBS": "1", "PublicDescription": "A version of INST_RETIRED that allows for a more unbiased distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR) feature to mitigate some bias in how retired instructions get sampled. Use on Fixed Counter 0.",@@ -544,6 +547,7 @@ }, { "BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event", + "EventCode": "0x0", "EventName": "TOPDOWN.SLOTS", "PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",tools/perf/pmu-events/arch/x86/icelakex/uncore-cache.json +1
-0@@ -892,6 +892,7 @@ }, { "BriefDescription": "Clockticks of the uncore caching and home agent (CHA)", + "EventCode": "0x0", "EventName": "UNC_CHA_CLOCKTICKS", "PerPkg": "1", "Unit": "CHA"tools/perf/pmu-events/arch/x86/icelakex/uncore-interconnect.json +1
-0@@ -1419,6 +1419,7 @@ }, { "BriefDescription": "Clockticks of the mesh to memory (M2M)", + "EventCode": "0x0", "EventName": "UNC_M2M_CLOCKTICKS", "PerPkg": "1", "Unit": "M2M"tools/perf/pmu-events/arch/x86/icelakex/uncore-memory.json +1
-0@@ -100,6 +100,7 @@ }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "Unit": "iMC"tools/perf/pmu-events/arch/x86/icelakex/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "Clockticks of the power control unit (PCU)", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Clockticks of the power control unit (PCU) : The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",tools/perf/pmu-events/arch/x86/ivybridge/pipeline.json +3
-0@@ -326,6 +326,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3"@@ -348,6 +349,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -355,6 +357,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "PublicDescription": "Core cycles when at least one thread on the physical core is not in halt state.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/ivytown/pipeline.json +4
-0@@ -326,6 +326,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3"@@ -348,6 +349,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -355,6 +357,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "PublicDescription": "Core cycles when at least one thread on the physical core is not in halt state.", "SampleAfterValue": "2000003",@@ -510,6 +513,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "SampleAfterValue": "2000003", "UMask": "0x1"tools/perf/pmu-events/arch/x86/ivytown/uncore-cache.json +2
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"@@ -1533,6 +1534,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",tools/perf/pmu-events/arch/x86/ivytown/uncore-interconnect.json +11
-0@@ -109,6 +109,7 @@ }, { "BriefDescription": "Clocks in the IRP", + "EventCode": "0x0", "EventName": "UNC_I_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Number of clocks in the IRP.",@@ -1522,6 +1523,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Data Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",@@ -1530,6 +1532,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Non-Data protocol Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.NON_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",@@ -1538,6 +1541,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Flits (both Header and Data)", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency.",@@ -1546,6 +1550,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Data Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of data flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the data flits (not the header).",@@ -1554,6 +1559,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Header Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_NONDATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of protocol flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the header flits (not the data). This includes extended headers.",@@ -1562,6 +1568,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of flits transmitted over QPI on the home channel.",@@ -1570,6 +1577,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Non-Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_NONREQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of non-request flits transmitted over QPI on the home channel. These are most commonly snoop responses, and this event can be used as a proxy for that.",@@ -1578,6 +1586,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_REQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of data request transmitted over QPI on the home channel. This basically counts the number of remote memory requests transmitted over QPI. In conjunction with the local read count in the Home Agent, one can calculate the number of LLC Misses.",@@ -1586,6 +1595,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; SNP Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.SNP", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of snoop request flits transmitted over QPI. These requests are contained in the snoop channel. This does not include snoop responses, which are transmitted on the home channel.",@@ -3104,6 +3114,7 @@ }, { "EventName": "UNC_U_CLOCKTICKS", + "EventCode": "0x0", "PerPkg": "1", "Unit": "UBOX" },tools/perf/pmu-events/arch/x86/ivytown/uncore-memory.json +1
-0@@ -131,6 +131,7 @@ }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_DCLOCKTICKS", "PerPkg": "1", "Unit": "iMC"tools/perf/pmu-events/arch/x86/ivytown/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 800 MHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",tools/perf/pmu-events/arch/x86/jaketown/pipeline.json +4
-0@@ -329,6 +329,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -351,6 +352,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -359,6 +361,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -432,6 +435,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/jaketown/uncore-cache.json +2
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"@@ -863,6 +864,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",tools/perf/pmu-events/arch/x86/jaketown/uncore-interconnect.json +12
-0@@ -109,6 +109,7 @@ }, { "BriefDescription": "Clocks in the IRP", + "EventCode": "0x0", "EventName": "UNC_I_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Number of clocks in the IRP.",@@ -847,6 +848,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Data Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.",@@ -855,6 +857,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Idle and Null Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.IDLE", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.",@@ -863,6 +866,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Non-Data protocol Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.NON_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.",@@ -871,6 +875,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Flits (both Header and Data)", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",@@ -879,6 +884,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Data Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",@@ -887,6 +893,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Header Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_NONDATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",@@ -895,6 +902,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",@@ -903,6 +911,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Non-Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_NONREQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",@@ -911,6 +920,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_REQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",@@ -919,6 +929,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; SNP Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.SNP", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",@@ -1576,6 +1587,7 @@ }, { "EventName": "UNC_U_CLOCKTICKS", + "EventCode": "0x0", "PerPkg": "1", "Unit": "UBOX" },tools/perf/pmu-events/arch/x86/jaketown/uncore-memory.json +1
-0@@ -65,6 +65,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Uncore Fixed Counter - uclks",tools/perf/pmu-events/arch/x86/jaketown/uncore-power.json +2
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 800 MHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",@@ -216,6 +217,7 @@ }, { "BriefDescription": "Cycles spent changing Frequency", + "EventCode": "0x0", "EventName": "UNC_P_FREQ_TRANS_CYCLES", "PerPkg": "1", "PublicDescription": "Counts the number of cycles when the system is changing frequency. This can not be filtered by thread ID. One can also use it with the occupancy counter that monitors number of threads in C0 to estimate the performance impact that frequency transitions had on the system.",tools/perf/pmu-events/arch/x86/knightslanding/pipeline.json +3
-0@@ -150,12 +150,14 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3" }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter", "SampleAfterValue": "2000003",@@ -177,6 +179,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of instructions retired", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions that retire. For instructions that consist of multiple micro-ops, this event counts exactly once, as the last micro-op of the instruction retires. The event continues counting while instructions retire, including during interrupt service routines caused by hardware interrupts, faults or traps.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/knightslanding/uncore-cache.json +1
-0@@ -3246,6 +3246,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_H_U_CLOCKTICKS", "PerPkg": "1", "Unit": "CHA"tools/perf/pmu-events/arch/x86/knightslanding/uncore-memory.json +4
-0@@ -41,6 +41,7 @@ }, { "BriefDescription": "ECLK count", + "EventCode": "0x0", "EventName": "UNC_E_E_CLOCKTICKS", "PerPkg": "1", "Unit": "EDC_ECLK"@@ -55,6 +56,7 @@ }, { "BriefDescription": "UCLK count", + "EventCode": "0x0", "EventName": "UNC_E_U_CLOCKTICKS", "PerPkg": "1", "Unit": "EDC_UCLK"@@ -93,12 +95,14 @@ }, { "BriefDescription": "DCLK count", + "EventCode": "0x0", "EventName": "UNC_M_D_CLOCKTICKS", "PerPkg": "1", "Unit": "iMC_DCLK" }, { "BriefDescription": "UCLK count", + "EventCode": "0x0", "EventName": "UNC_M_U_CLOCKTICKS", "PerPkg": "1", "Unit": "iMC_UCLK"tools/perf/pmu-events/arch/x86/meteorlake/pipeline.json +8
-0@@ -37,6 +37,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "SampleAfterValue": "2000003", "UMask": "0x2",@@ -51,6 +52,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3",@@ -58,6 +60,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -75,6 +78,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "SampleAfterValue": "2000003", "UMask": "0x2",@@ -82,6 +86,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -105,6 +110,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of instructions retired", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "SampleAfterValue": "2000003",@@ -113,6 +119,7 @@ }, { "BriefDescription": "Number of instructions retired. Fixed Counter - architectural event", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",@@ -157,6 +164,7 @@ }, { "BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event", + "EventCode": "0x0", "EventName": "TOPDOWN.SLOTS", "PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",tools/perf/pmu-events/arch/x86/sandybridge/pipeline.json +4
-0@@ -337,6 +337,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -359,6 +360,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -367,6 +369,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -440,6 +443,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/sapphirerapids/pipeline.json +5
-0@@ -284,6 +284,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -299,6 +300,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -426,6 +428,7 @@ }, { "BriefDescription": "Number of instructions retired. Fixed Counter - architectural event", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",@@ -457,6 +460,7 @@ }, { "BriefDescription": "Precise instruction retired with PEBS precise-distribution", + "EventCode": "0x0", "EventName": "INST_RETIRED.PREC_DIST", "PEBS": "1", "PublicDescription": "A version of INST_RETIRED that allows for a precise distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR++) feature to fix bias in how retired instructions get sampled. Use on Fixed Counter 0.",@@ -719,6 +723,7 @@ }, { "BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event", + "EventCode": "0x0", "EventName": "TOPDOWN.SLOTS", "PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",tools/perf/pmu-events/arch/x86/sierraforest/pipeline.json +4
-0@@ -17,6 +17,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -29,6 +30,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3"@@ -43,6 +45,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -55,6 +58,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of instructions retired", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/silvermont/pipeline.json +3
-0@@ -143,6 +143,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. In systems with a constant core frequency, this event can give you a measurement of the elapsed time while the core was not in halt state by dividing the event count by the core frequency. This event is architecturally defined and is a designated fixed counter. CPU_CLK_UNHALTED.CORE and CPU_CLK_UNHALTED.CORE_P use the core frequency which may change from time to time. CPU_CLK_UNHALTE.REF_TSC and CPU_CLK_UNHALTED.REF are not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. The fixed events are CPU_CLK_UNHALTED.CORE and CPU_CLK_UNHALTED.REF_TSC and the programmable events are CPU_CLK_UNHALTED.CORE_P and CPU_CLK_UNHALTED.REF.", "SampleAfterValue": "2000003",@@ -165,6 +166,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time. This event is not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. Divide this event count by core frequency to determine the elapsed time while the core was not in halt state. Divide this event count by core frequency to determine the elapsed time while the core was not in halt state. This event is architecturally defined and is a designated fixed counter. CPU_CLK_UNHALTED.CORE and CPU_CLK_UNHALTED.CORE_P use the core frequency which may change from time to time. CPU_CLK_UNHALTE.REF_TSC and CPU_CLK_UNHALTED.REF are not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. The fixed events are CPU_CLK_UNHALTED.CORE and CPU_CLK_UNHALTED.REF_TSC and the programmable events are CPU_CLK_UNHALTED.CORE_P and CPU_CLK_UNHALTED.REF.", "SampleAfterValue": "2000003",@@ -180,6 +182,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of instructions retired", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions that retire. For instructions that consist of multiple micro-ops, this event counts exactly once, as the last micro-op of the instruction retires. The event continues counting while instructions retire, including during interrupt service routines caused by hardware interrupts, faults or traps. Background: Modern microprocessors employ extensive pipelining and speculative techniques. Since sometimes an instruction is started but never completed, the notion of \"retirement\" is introduced. A retired instruction is one that commits its states. Or stated differently, an instruction might be abandoned at some point. No instruction is truly finished until it retires. This counter measures the number of completed instructions. The fixed event is INST_RETIRED.ANY and the programmable event is INST_RETIRED.ANY_P.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/skylake/pipeline.json +4
-0@@ -182,6 +182,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -213,6 +214,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -221,6 +223,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -360,6 +363,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "Counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, Counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. Counting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/skylakex/pipeline.json +4
-0@@ -191,6 +191,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -222,6 +223,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -230,6 +232,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -369,6 +372,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "Counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, Counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. Counting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/skylakex/uncore-cache.json +2
-0@@ -460,6 +460,7 @@ }, { "BriefDescription": "Clockticks of the uncore caching & home agent (CHA)", + "EventCode": "0x0", "EventName": "UNC_CHA_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts clockticks of the clock controlling the uncore caching and home agent (CHA).",@@ -5678,6 +5679,7 @@ { "BriefDescription": "This event is deprecated. Refer to new event UNC_CHA_CLOCKTICKS", "Deprecated": "1", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CHA"tools/perf/pmu-events/arch/x86/skylakex/uncore-interconnect.json +1
-0@@ -1090,6 +1090,7 @@ }, { "BriefDescription": "Cycles - at UCLK", + "EventCode": "0x0", "EventName": "UNC_M2M_CLOCKTICKS", "PerPkg": "1", "Unit": "M2M"tools/perf/pmu-events/arch/x86/skylakex/uncore-io.json +1
-0@@ -1271,6 +1271,7 @@ }, { "BriefDescription": "Counting disabled", + "EventCode": "0x0", "EventName": "UNC_IIO_NOTHING", "PerPkg": "1", "Unit": "IIO"tools/perf/pmu-events/arch/x86/skylakex/uncore-memory.json +1
-0@@ -167,6 +167,7 @@ }, { "BriefDescription": "Memory controller clock ticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts clockticks of the fixed frequency clock of the memory controller using one of the programmable counters.",tools/perf/pmu-events/arch/x86/skylakex/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",tools/perf/pmu-events/arch/x86/snowridgex/pipeline.json +2
-0@@ -150,6 +150,7 @@ }, { "BriefDescription": "Counts the number of unhalted reference clock cycles at TSC frequency. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is not affected by core frequency changes and increments at a fixed frequency that is also used for the Time Stamp Counter (TSC). This event uses fixed counter 2.", "SampleAfterValue": "2000003",@@ -179,6 +180,7 @@ }, { "BriefDescription": "Counts the total number of instructions retired. (Fixed event)", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the total number of instructions that retired. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. This event continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0.",tools/perf/pmu-events/arch/x86/snowridgex/uncore-cache.json +1
-0@@ -872,6 +872,7 @@ }, { "BriefDescription": "Uncore cache clock ticks", + "EventCode": "0x0", "EventName": "UNC_CHA_CLOCKTICKS", "PerPkg": "1", "Unit": "CHA"tools/perf/pmu-events/arch/x86/snowridgex/uncore-interconnect.json +1
-0@@ -1419,6 +1419,7 @@ }, { "BriefDescription": "Clockticks of the mesh to memory (M2M)", + "EventCode": "0x0", "EventName": "UNC_M2M_CLOCKTICKS", "PerPkg": "1", "Unit": "M2M"tools/perf/pmu-events/arch/x86/snowridgex/uncore-memory.json +1
-0@@ -120,6 +120,7 @@ }, { "BriefDescription": "Memory controller clock ticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Clockticks of the integrated memory controller (IMC)",tools/perf/pmu-events/arch/x86/snowridgex/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "Clockticks of the power control unit (PCU)", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "Unit": "PCU"tools/perf/pmu-events/arch/x86/tigerlake/pipeline.json +5
-1@@ -193,6 +193,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -208,6 +209,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -352,6 +354,7 @@ }, { "BriefDescription": "Number of instructions retired. Fixed Counter - architectural event", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",@@ -377,6 +380,7 @@ }, { "BriefDescription": "Precise instruction retired event with a reduced effect of PEBS shadow in IP distribution", + "EventCode": "0x0", "EventName": "INST_RETIRED.PREC_DIST", "PEBS": "1", "PublicDescription": "A version of INST_RETIRED that allows for a more unbiased distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR) feature to mitigate some bias in how retired instructions get sampled. Use on Fixed Counter 0.",@@ -569,6 +573,7 @@ }, { "BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event", + "EventCode": "0x0", "EventName": "TOPDOWN.SLOTS", "PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003", --1.8.3.1
Re: [PATCH v7 3/8] perf vendor events: Supplement the omitted EventCode
Quoted 14 lines · Ian Rogers
Quoted 13 lines · Jing Zhang
If there is an "event=0" in the event description, the EventCode can be omitted in the JSON file, and
jevent.pywill automatically fill in "event=0" during parsing.However, for some events where EventCode and ConfigCode are missing, it is not necessary to automatically fill in "event=0", such as the CMN event description which is typically "type=xxx, eventid=xxx".
Therefore, before modifying
jevent.pyto prevent it from automatically adding "event=0" by default, it is necessary to fill in all omitted EventCodes first.
Quoted 8 lines
Quoted 1 line
Signed-off-by Jing Zhang <[email protected]>As these files are generated, the generator script needs updating. However, I don't think this change makes sense as the event=0 is overwritten in the case of an
arch_stdevent:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/pmu-events/jevents.py?h=perf-tools-next#n369
So yes event=0 was filled in, but it was then overwritten.
Yes, arch_std_event will indeed be overwritten, but the events I added EventCode are not arch_std_event, and the x86 architecture does not define arch_std_event. I used JEVENT_ARCH=all to compile and check, there is no problem. In patch 4, I modified jevent.py, no longer adding "event=0" by default, and comparing the pmu_event.c before and after, they are consistent.
Thanks,
Jing
Quoted 2,033 lines
Thanks,
IanQuoted 2,030 lines
--- tools/perf/pmu-events/arch/x86/alderlake/pipeline.json | 9 +++++++++ tools/perf/pmu-events/arch/x86/alderlaken/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/broadwell/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/broadwellde/pipeline.json | 4 ++++ .../perf/pmu-events/arch/x86/broadwellde/uncore-cache.json | 2 ++ .../arch/x86/broadwellde/uncore-interconnect.json | 1 + .../pmu-events/arch/x86/broadwellde/uncore-memory.json | 1 + .../perf/pmu-events/arch/x86/broadwellde/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/broadwellx/pipeline.json | 4 ++++ .../perf/pmu-events/arch/x86/broadwellx/uncore-cache.json | 2 ++ .../arch/x86/broadwellx/uncore-interconnect.json | 13 +++++++++++++ .../perf/pmu-events/arch/x86/broadwellx/uncore-memory.json | 2 ++ .../perf/pmu-events/arch/x86/broadwellx/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/cascadelakex/pipeline.json | 4 ++++ .../pmu-events/arch/x86/cascadelakex/uncore-cache.json | 2 ++ .../arch/x86/cascadelakex/uncore-interconnect.json | 1 + tools/perf/pmu-events/arch/x86/cascadelakex/uncore-io.json | 1 + .../pmu-events/arch/x86/cascadelakex/uncore-memory.json | 1 + .../pmu-events/arch/x86/cascadelakex/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/elkhartlake/pipeline.json | 2 ++ tools/perf/pmu-events/arch/x86/goldmont/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/goldmontplus/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/grandridge/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/graniterapids/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/haswell/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/haswellx/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/haswellx/uncore-cache.json | 2 ++ .../pmu-events/arch/x86/haswellx/uncore-interconnect.json | 14 ++++++++++++++ tools/perf/pmu-events/arch/x86/haswellx/uncore-memory.json | 2 ++ tools/perf/pmu-events/arch/x86/haswellx/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/icelake/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/icelakex/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/icelakex/uncore-cache.json | 1 + .../pmu-events/arch/x86/icelakex/uncore-interconnect.json | 1 + tools/perf/pmu-events/arch/x86/icelakex/uncore-memory.json | 1 + tools/perf/pmu-events/arch/x86/icelakex/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/ivybridge/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/ivytown/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/ivytown/uncore-cache.json | 2 ++ .../pmu-events/arch/x86/ivytown/uncore-interconnect.json | 11 +++++++++++ tools/perf/pmu-events/arch/x86/ivytown/uncore-memory.json | 1 + tools/perf/pmu-events/arch/x86/ivytown/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/jaketown/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/jaketown/uncore-cache.json | 2 ++ .../pmu-events/arch/x86/jaketown/uncore-interconnect.json | 12 ++++++++++++ tools/perf/pmu-events/arch/x86/jaketown/uncore-memory.json | 1 + tools/perf/pmu-events/arch/x86/jaketown/uncore-power.json | 2 ++ .../perf/pmu-events/arch/x86/knightslanding/pipeline.json | 3 +++ .../pmu-events/arch/x86/knightslanding/uncore-cache.json | 1 + .../pmu-events/arch/x86/knightslanding/uncore-memory.json | 4 ++++ tools/perf/pmu-events/arch/x86/meteorlake/pipeline.json | 8 ++++++++ tools/perf/pmu-events/arch/x86/sandybridge/pipeline.json | 4 ++++ .../perf/pmu-events/arch/x86/sapphirerapids/pipeline.json | 5 +++++ tools/perf/pmu-events/arch/x86/sierraforest/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/silvermont/pipeline.json | 3 +++ tools/perf/pmu-events/arch/x86/skylake/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/skylakex/pipeline.json | 4 ++++ tools/perf/pmu-events/arch/x86/skylakex/uncore-cache.json | 2 ++ .../pmu-events/arch/x86/skylakex/uncore-interconnect.json | 1 + tools/perf/pmu-events/arch/x86/skylakex/uncore-io.json | 1 + tools/perf/pmu-events/arch/x86/skylakex/uncore-memory.json | 1 + tools/perf/pmu-events/arch/x86/skylakex/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/snowridgex/pipeline.json | 2 ++ .../perf/pmu-events/arch/x86/snowridgex/uncore-cache.json | 1 + .../arch/x86/snowridgex/uncore-interconnect.json | 1 + .../perf/pmu-events/arch/x86/snowridgex/uncore-memory.json | 1 + .../perf/pmu-events/arch/x86/snowridgex/uncore-power.json | 1 + tools/perf/pmu-events/arch/x86/tigerlake/pipeline.json | 5 +++++ 68 files changed, 211 insertions(+)tools/perf/pmu-events/arch/x86/alderlake/pipeline.json +9
-0@@ -489,6 +489,7 @@ }, { "BriefDescription": "Counts the number of unhalted core clock cycles. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1.", "SampleAfterValue": "2000003",@@ -550,6 +551,7 @@ }, { "BriefDescription": "Counts the number of unhalted reference clock cycles at TSC frequency. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is not affected by core frequency changes and increments at a fixed frequency that is also used for the Time Stamp Counter (TSC). This event uses fixed counter 2.", "SampleAfterValue": "2000003",@@ -558,6 +560,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -584,6 +587,7 @@ }, { "BriefDescription": "Counts the number of unhalted core clock cycles. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1.", "SampleAfterValue": "2000003",@@ -592,6 +596,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -743,6 +748,7 @@ }, { "BriefDescription": "Counts the total number of instructions retired. (Fixed event)", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the total number of instructions that retired. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. This event continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0.",@@ -752,6 +758,7 @@ }, { "BriefDescription": "Number of instructions retired. Fixed Counter - architectural event", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",@@ -796,6 +803,7 @@ }, { "BriefDescription": "Precise instruction retired with PEBS precise-distribution", + "EventCode": "0x0", "EventName": "INST_RETIRED.PREC_DIST", "PEBS": "1", "PublicDescription": "A version of INST_RETIRED that allows for a precise distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR++) feature to fix bias in how retired instructions get sampled. Use on Fixed Counter 0.",@@ -1160,6 +1168,7 @@ }, { "BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event", + "EventCode": "0x0", "EventName": "TOPDOWN.SLOTS", "PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",tools/perf/pmu-events/arch/x86/alderlaken/pipeline.json +3
-0@@ -211,6 +211,7 @@ }, { "BriefDescription": "Counts the number of unhalted core clock cycles. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1.", "SampleAfterValue": "2000003",@@ -225,6 +226,7 @@ }, { "BriefDescription": "Counts the number of unhalted reference clock cycles at TSC frequency. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is not affected by core frequency changes and increments at a fixed frequency that is also used for the Time Stamp Counter (TSC). This event uses fixed counter 2.", "SampleAfterValue": "2000003",@@ -240,6 +242,7 @@ }, { "BriefDescription": "Counts the number of unhalted core clock cycles. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/broadwell/pipeline.json +4
-0@@ -336,6 +336,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. \nNote: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. This event is clocked by base clock (100 Mhz) on Sandy Bridge. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -359,6 +360,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -366,6 +368,7 @@ }, { "AnyThread": "1", + "EventCode": "0x0", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003",@@ -514,6 +517,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. \nNotes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. \nCounting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/broadwellde/pipeline.json +4
-0@@ -336,6 +336,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. \nNote: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. This event is clocked by base clock (100 Mhz) on Sandy Bridge. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -359,6 +360,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -367,6 +369,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -514,6 +517,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. \nNotes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. \nCounting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/broadwellde/uncore-cache.json +2
-0@@ -8,6 +8,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"@@ -1501,6 +1502,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",tools/perf/pmu-events/arch/x86/broadwellde/uncore-interconnect.json +1
-0@@ -19,6 +19,7 @@ }, { "BriefDescription": "Clocks in the IRP", + "EventCode": "0x0", "EventName": "UNC_I_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Number of clocks in the IRP.",tools/perf/pmu-events/arch/x86/broadwellde/uncore-memory.json +1
-0@@ -131,6 +131,7 @@ }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_DCLOCKTICKS", "PerPkg": "1", "Unit": "iMC"tools/perf/pmu-events/arch/x86/broadwellde/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",tools/perf/pmu-events/arch/x86/broadwellx/pipeline.json +4
-0@@ -336,6 +336,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. \nNote: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. This event is clocked by base clock (100 Mhz) on Sandy Bridge. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -359,6 +360,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -367,6 +369,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -514,6 +517,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. \nNotes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. \nCounting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/broadwellx/uncore-cache.json +2
-0@@ -183,6 +183,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"@@ -1689,6 +1690,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",tools/perf/pmu-events/arch/x86/broadwellx/uncore-interconnect.json +13
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "Number of non data (control) flits transmitted . Derived from unc_q_txl_flits_g0.non_data", + "EventCode": "0x0", "EventName": "QPI_CTL_BANDWIDTH_TX", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",@@ -10,6 +11,7 @@ }, { "BriefDescription": "Number of data flits transmitted . Derived from unc_q_txl_flits_g0.data", + "EventCode": "0x0", "EventName": "QPI_DATA_BANDWIDTH_TX", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",@@ -37,6 +39,7 @@ }, { "BriefDescription": "Clocks in the IRP", + "EventCode": "0x0", "EventName": "UNC_I_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Number of clocks in the IRP.",@@ -1400,6 +1403,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Data Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",@@ -1408,6 +1412,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Non-Data protocol Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.NON_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",@@ -1416,6 +1421,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Flits (both Header and Data)", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency.",@@ -1424,6 +1430,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Data Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of data flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the data flits (not the header).",@@ -1432,6 +1439,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Header Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_NONDATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of protocol flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the header flits (not the data). This includes extended headers.",@@ -1440,6 +1448,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of flits transmitted over QPI on the home channel.",@@ -1448,6 +1457,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Non-Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_NONREQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of non-request flits transmitted over QPI on the home channel. These are most commonly snoop responses, and this event can be used as a proxy for that.",@@ -1456,6 +1466,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_REQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of data request transmitted over QPI on the home channel. This basically counts the number of remote memory requests transmitted over QPI. In conjunction with the local read count in the Home Agent, one can calculate the number of LLC Misses.",@@ -1464,6 +1475,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; SNP Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.SNP", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of snoop request flits transmitted over QPI. These requests are contained in the snoop channel. This does not include snoop responses, which are transmitted on the home channel.",@@ -3162,6 +3174,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_S_CLOCKTICKS", "PerPkg": "1", "Unit": "SBOX"tools/perf/pmu-events/arch/x86/broadwellx/uncore-memory.json +2
-0@@ -158,12 +158,14 @@ }, { "BriefDescription": "Clockticks in the Memory Controller using one of the programmable counters", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS_P", "PerPkg": "1", "Unit": "iMC" }, { "BriefDescription": "This event is deprecated. Refer to new event UNC_M_CLOCKTICKS_P", + "EventCode": "0x0", "EventName": "UNC_M_DCLOCKTICKS", "PerPkg": "1", "Unit": "iMC"tools/perf/pmu-events/arch/x86/broadwellx/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",tools/perf/pmu-events/arch/x86/cascadelakex/pipeline.json +4
-0@@ -191,6 +191,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -222,6 +223,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -230,6 +232,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -369,6 +372,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "Counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, Counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. Counting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/cascadelakex/uncore-cache.json +2
-0@@ -512,6 +512,7 @@ }, { "BriefDescription": "Uncore cache clock ticks", + "EventCode": "0x0", "EventName": "UNC_CHA_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts clockticks of the clock controlling the uncore caching and home agent (CHA).",@@ -5792,6 +5793,7 @@ }, { "BriefDescription": "This event is deprecated. Refer to new event UNC_CHA_CLOCKTICKS", + "EventCode": "0x0", "Deprecated": "1", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1",tools/perf/pmu-events/arch/x86/cascadelakex/uncore-interconnect.json +1
-0@@ -1090,6 +1090,7 @@ }, { "BriefDescription": "Cycles - at UCLK", + "EventCode": "0x0", "EventName": "UNC_M2M_CLOCKTICKS", "PerPkg": "1", "Unit": "M2M"tools/perf/pmu-events/arch/x86/cascadelakex/uncore-io.json +1
-0@@ -1271,6 +1271,7 @@ }, { "BriefDescription": "Counting disabled", + "EventCode": "0x0", "EventName": "UNC_IIO_NOTHING", "PerPkg": "1", "Unit": "IIO"tools/perf/pmu-events/arch/x86/cascadelakex/uncore-memory.json +1
-0@@ -167,6 +167,7 @@ }, { "BriefDescription": "Memory controller clock ticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts clockticks of the fixed frequency clock of the memory controller using one of the programmable counters.",tools/perf/pmu-events/arch/x86/cascadelakex/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",tools/perf/pmu-events/arch/x86/elkhartlake/pipeline.json +2
-0@@ -150,6 +150,7 @@ }, { "BriefDescription": "Counts the number of unhalted reference clock cycles at TSC frequency. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is not affected by core frequency changes and increments at a fixed frequency that is also used for the Time Stamp Counter (TSC). This event uses fixed counter 2.", "SampleAfterValue": "2000003",@@ -179,6 +180,7 @@ }, { "BriefDescription": "Counts the total number of instructions retired. (Fixed event)", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the total number of instructions that retired. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. This event continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0.",tools/perf/pmu-events/arch/x86/goldmont/pipeline.json +3
-0@@ -143,6 +143,7 @@ }, { "BriefDescription": "Core cycles when core is not halted (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. In mobile systems the core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",@@ -165,6 +166,7 @@ }, { "BriefDescription": "Reference cycles when core is not halted (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. In mobile systems the core frequency may change from time. This event is not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. This event uses fixed counter 2. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",@@ -187,6 +189,7 @@ }, { "BriefDescription": "Instructions retired (Fixed event)", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "Counts the number of instructions that retire execution. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. The counter continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/goldmontplus/pipeline.json +3
-0@@ -143,6 +143,7 @@ }, { "BriefDescription": "Core cycles when core is not halted (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. In mobile systems the core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. This event uses fixed counter 1. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",@@ -165,6 +166,7 @@ }, { "BriefDescription": "Reference cycles when core is not halted (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. In mobile systems the core frequency may change from time. This event is not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. This event uses fixed counter 2. You cannot collect a PEBs record for this event.", "SampleAfterValue": "2000003",@@ -187,6 +189,7 @@ }, { "BriefDescription": "Instructions retired (Fixed event)", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "2", "PublicDescription": "Counts the number of instructions that retire execution. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. The counter continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0. You cannot collect a PEBs record for this event.",tools/perf/pmu-events/arch/x86/grandridge/pipeline.json +3
-0@@ -29,6 +29,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3"@@ -43,6 +44,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -55,6 +57,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of instructions retired", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/graniterapids/pipeline.json +4
-0@@ -17,6 +17,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -32,6 +33,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -46,6 +48,7 @@ }, { "BriefDescription": "Number of instructions retired. Fixed Counter - architectural event", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",@@ -78,6 +81,7 @@ }, { "BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event", + "EventCode": "0x0", "EventName": "TOPDOWN.SLOTS", "PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",tools/perf/pmu-events/arch/x86/haswell/pipeline.json +4
-0@@ -303,6 +303,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state.", "SampleAfterValue": "2000003",@@ -327,6 +328,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of thread cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. The core frequency may change from time to time due to power or thermal throttling.", "SampleAfterValue": "2000003",@@ -335,6 +337,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -436,6 +439,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "Errata": "HSD140, HSD143", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. INST_RETIRED.ANY is counted by a designated fixed counter, leaving the programmable counters available for other events. Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.",tools/perf/pmu-events/arch/x86/haswellx/pipeline.json +4
-0@@ -303,6 +303,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state.", "SampleAfterValue": "2000003",@@ -327,6 +328,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of thread cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. The core frequency may change from time to time due to power or thermal throttling.", "SampleAfterValue": "2000003",@@ -335,6 +337,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -436,6 +439,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "Errata": "HSD140, HSD143", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. INST_RETIRED.ANY is counted by a designated fixed counter, leaving the programmable counters available for other events. Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.",tools/perf/pmu-events/arch/x86/haswellx/uncore-cache.json +2
-0@@ -183,6 +183,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"@@ -1698,6 +1699,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",tools/perf/pmu-events/arch/x86/haswellx/uncore-interconnect.json +14
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "Number of non data (control) flits transmitted . Derived from unc_q_txl_flits_g0.non_data", + "EventCode": "0x0", "EventName": "QPI_CTL_BANDWIDTH_TX", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",@@ -10,6 +11,7 @@ }, { "BriefDescription": "Number of data flits transmitted . Derived from unc_q_txl_flits_g0.data", + "EventCode": "0x0", "EventName": "QPI_DATA_BANDWIDTH_TX", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",@@ -37,6 +39,7 @@ }, { "BriefDescription": "Clocks in the IRP", + "EventCode": "0x0", "EventName": "UNC_I_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Number of clocks in the IRP.",@@ -1401,6 +1404,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Data Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",@@ -1409,6 +1413,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Non-Data protocol Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.NON_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",@@ -1417,6 +1422,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Flits (both Header and Data)", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency.",@@ -1425,6 +1431,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Data Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of data flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the data flits (not the header).",@@ -1433,6 +1440,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Header Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_NONDATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of protocol flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the header flits (not the data). This includes extended headers.",@@ -1441,6 +1449,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of flits transmitted over QPI on the home channel.",@@ -1449,6 +1458,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Non-Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_NONREQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of non-request flits transmitted over QPI on the home channel. These are most commonly snoop responses, and this event can be used as a proxy for that.",@@ -1457,6 +1467,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_REQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of data request transmitted over QPI on the home channel. This basically counts the number of remote memory requests transmitted over QPI. In conjunction with the local read count in the Home Agent, one can calculate the number of LLC Misses.",@@ -1465,6 +1476,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; SNP Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.SNP", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of snoop request flits transmitted over QPI. These requests are contained in the snoop channel. This does not include snoop responses, which are transmitted on the home channel.",@@ -3136,6 +3148,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_S_CLOCKTICKS", "PerPkg": "1", "Unit": "SBOX"@@ -3823,6 +3836,7 @@ }, { "BriefDescription": "UNC_U_CLOCKTICKS", + "EventCode": "0x0", "EventName": "UNC_U_CLOCKTICKS", "PerPkg": "1", "Unit": "UBOX"tools/perf/pmu-events/arch/x86/haswellx/uncore-memory.json +2
-0@@ -151,12 +151,14 @@ }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "Unit": "iMC" }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_DCLOCKTICKS", "PerPkg": "1", "Unit": "iMC"tools/perf/pmu-events/arch/x86/haswellx/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 800 MHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",tools/perf/pmu-events/arch/x86/icelake/pipeline.json +4
-0@@ -193,6 +193,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -208,6 +209,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -359,6 +361,7 @@ }, { "BriefDescription": "Precise instruction retired event with a reduced effect of PEBS shadow in IP distribution", + "EventCode": "0x0", "EventName": "INST_RETIRED.PREC_DIST", "PEBS": "1", "PublicDescription": "A version of INST_RETIRED that allows for a more unbiased distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR) feature to mitigate some bias in how retired instructions get sampled. Use on Fixed Counter 0.",@@ -562,6 +565,7 @@ }, { "BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event", + "EventCode": "0x0", "EventName": "TOPDOWN.SLOTS", "PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",tools/perf/pmu-events/arch/x86/icelakex/pipeline.json +4
-0@@ -193,6 +193,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -208,6 +209,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -359,6 +361,7 @@ }, { "BriefDescription": "Precise instruction retired event with a reduced effect of PEBS shadow in IP distribution", + "EventCode": "0x0", "EventName": "INST_RETIRED.PREC_DIST", "PEBS": "1", "PublicDescription": "A version of INST_RETIRED that allows for a more unbiased distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR) feature to mitigate some bias in how retired instructions get sampled. Use on Fixed Counter 0.",@@ -544,6 +547,7 @@ }, { "BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event", + "EventCode": "0x0", "EventName": "TOPDOWN.SLOTS", "PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",tools/perf/pmu-events/arch/x86/icelakex/uncore-cache.json +1
-0@@ -892,6 +892,7 @@ }, { "BriefDescription": "Clockticks of the uncore caching and home agent (CHA)", + "EventCode": "0x0", "EventName": "UNC_CHA_CLOCKTICKS", "PerPkg": "1", "Unit": "CHA"tools/perf/pmu-events/arch/x86/icelakex/uncore-interconnect.json +1
-0@@ -1419,6 +1419,7 @@ }, { "BriefDescription": "Clockticks of the mesh to memory (M2M)", + "EventCode": "0x0", "EventName": "UNC_M2M_CLOCKTICKS", "PerPkg": "1", "Unit": "M2M"tools/perf/pmu-events/arch/x86/icelakex/uncore-memory.json +1
-0@@ -100,6 +100,7 @@ }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "Unit": "iMC"tools/perf/pmu-events/arch/x86/icelakex/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "Clockticks of the power control unit (PCU)", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Clockticks of the power control unit (PCU) : The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",tools/perf/pmu-events/arch/x86/ivybridge/pipeline.json +3
-0@@ -326,6 +326,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3"@@ -348,6 +349,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -355,6 +357,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "PublicDescription": "Core cycles when at least one thread on the physical core is not in halt state.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/ivytown/pipeline.json +4
-0@@ -326,6 +326,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3"@@ -348,6 +349,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -355,6 +357,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "PublicDescription": "Core cycles when at least one thread on the physical core is not in halt state.", "SampleAfterValue": "2000003",@@ -510,6 +513,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "SampleAfterValue": "2000003", "UMask": "0x1"tools/perf/pmu-events/arch/x86/ivytown/uncore-cache.json +2
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"@@ -1533,6 +1534,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",tools/perf/pmu-events/arch/x86/ivytown/uncore-interconnect.json +11
-0@@ -109,6 +109,7 @@ }, { "BriefDescription": "Clocks in the IRP", + "EventCode": "0x0", "EventName": "UNC_I_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Number of clocks in the IRP.",@@ -1522,6 +1523,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Data Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of data flits transmitted over QPI. Each flit contains 64b of data. This includes both DRS and NCB data flits (coherent and non-coherent). This can be used to calculate the data bandwidth of the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This does not include the header flits that go in data packets.",@@ -1530,6 +1532,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Non-Data protocol Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.NON_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.; Number of non-NULL non-data flits transmitted across QPI. This basically tracks the protocol overhead on the QPI link. One can get a good picture of the QPI-link characteristics by evaluating the protocol flits, data flits, and idle/null flits. This includes the header flits for data packets.",@@ -1538,6 +1541,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Flits (both Header and Data)", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency.",@@ -1546,6 +1550,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Data Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of data flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the data flits (not the header).",@@ -1554,6 +1559,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Header Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_NONDATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the total number of protocol flits transmitted over QPI on the DRS (Data Response) channel. DRS flits are used to transmit data with coherency. This does not count data flits transmitted over the NCB channel which transmits non-coherent data. This includes only the header flits (not the data). This includes extended headers.",@@ -1562,6 +1568,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of flits transmitted over QPI on the home channel.",@@ -1570,6 +1577,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Non-Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_NONREQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of non-request flits transmitted over QPI on the home channel. These are most commonly snoop responses, and this event can be used as a proxy for that.",@@ -1578,6 +1586,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_REQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of data request transmitted over QPI on the home channel. This basically counts the number of remote memory requests transmitted over QPI. In conjunction with the local read count in the Home Agent, one can calculate the number of LLC Misses.",@@ -1586,6 +1595,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; SNP Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.SNP", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three groups that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each flit is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four fits, each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI speed (for example, 8.0 GT/s), the transfers here refer to fits. Therefore, in L0, the system will transfer 1 flit at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as data bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual data and an additional 16 bits of other information. To calculate data bandwidth, one should therefore do: data flits * 8B / time.; Counts the number of snoop request flits transmitted over QPI. These requests are contained in the snoop channel. This does not include snoop responses, which are transmitted on the home channel.",@@ -3104,6 +3114,7 @@ }, { "EventName": "UNC_U_CLOCKTICKS", + "EventCode": "0x0", "PerPkg": "1", "Unit": "UBOX" },tools/perf/pmu-events/arch/x86/ivytown/uncore-memory.json +1
-0@@ -131,6 +131,7 @@ }, { "BriefDescription": "DRAM Clockticks", + "EventCode": "0x0", "EventName": "UNC_M_DCLOCKTICKS", "PerPkg": "1", "Unit": "iMC"tools/perf/pmu-events/arch/x86/ivytown/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 800 MHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",tools/perf/pmu-events/arch/x86/jaketown/pipeline.json +4
-0@@ -329,6 +329,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -351,6 +352,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -359,6 +361,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -432,6 +435,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/jaketown/uncore-cache.json +2
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CBOX"@@ -863,6 +864,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_H_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts the number of uclks in the HA. This will be slightly different than the count in the Ubox because of enable/freeze delays. The HA is on the other side of the die from the fixed Ubox uclk counter, so the drift could be somewhat larger than in units that are closer like the QPI Agent.",tools/perf/pmu-events/arch/x86/jaketown/uncore-interconnect.json +12
-0@@ -109,6 +109,7 @@ }, { "BriefDescription": "Clocks in the IRP", + "EventCode": "0x0", "EventName": "UNC_I_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Number of clocks in the IRP.",@@ -847,6 +848,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Data Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.",@@ -855,6 +857,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Idle and Null Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.IDLE", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.",@@ -863,6 +866,7 @@ }, { "BriefDescription": "Flits Transferred - Group 0; Non-Data protocol Tx Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G0.NON_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. It includes filters for Idle, protocol, and Data Flits. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time (for L0) or 4B instead of 8B for L0p.",@@ -871,6 +875,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Flits (both Header and Data)", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",@@ -879,6 +884,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Data Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_DATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",@@ -887,6 +893,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; DRS Header Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.DRS_NONDATA", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",@@ -895,6 +902,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",@@ -903,6 +911,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Non-Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_NONREQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",@@ -911,6 +920,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; HOM Request Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.HOM_REQ", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",@@ -919,6 +929,7 @@ }, { "BriefDescription": "Flits Transferred - Group 1; SNP Flits", + "EventCode": "0x0", "EventName": "UNC_Q_TxL_FLITS_G1.SNP", "PerPkg": "1", "PublicDescription": "Counts the number of flits transmitted across the QPI Link. This is one of three 'groups' that allow us to track flits. It includes filters for SNP, HOM, and DRS message classes. Each 'flit' is made up of 80 bits of information (in addition to some ECC data). In full-width (L0) mode, flits are made up of four 'fits', each of which contains 20 bits of data (along with some additional ECC data). In half-width (L0p) mode, the fits are only 10 bits, and therefore it takes twice as many fits to transmit a flit. When one talks about QPI 'speed' (for example, 8.0 GT/s), the 'transfers' here refer to 'fits'. Therefore, in L0, the system will transfer 1 'flit' at the rate of 1/4th the QPI speed. One can calculate the bandwidth of the link by taking: flits*80b/time. Note that this is not the same as 'data' bandwidth. For example, when we are transferring a 64B cacheline across QPI, we will break it into 9 flits -- 1 with header information and 8 with 64 bits of actual 'data' and an additional 16 bits of other information. To calculate 'data' bandwidth, one should therefore do: data flits * 8B / time.",@@ -1576,6 +1587,7 @@ }, { "EventName": "UNC_U_CLOCKTICKS", + "EventCode": "0x0", "PerPkg": "1", "Unit": "UBOX" },tools/perf/pmu-events/arch/x86/jaketown/uncore-memory.json +1
-0@@ -65,6 +65,7 @@ }, { "BriefDescription": "uclks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Uncore Fixed Counter - uclks",tools/perf/pmu-events/arch/x86/jaketown/uncore-power.json +2
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 800 MHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",@@ -216,6 +217,7 @@ }, { "BriefDescription": "Cycles spent changing Frequency", + "EventCode": "0x0", "EventName": "UNC_P_FREQ_TRANS_CYCLES", "PerPkg": "1", "PublicDescription": "Counts the number of cycles when the system is changing frequency. This can not be filtered by thread ID. One can also use it with the occupancy counter that monitors number of threads in C0 to estimate the performance impact that frequency transitions had on the system.",tools/perf/pmu-events/arch/x86/knightslanding/pipeline.json +3
-0@@ -150,12 +150,14 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3" }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter", "SampleAfterValue": "2000003",@@ -177,6 +179,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of instructions retired", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions that retire. For instructions that consist of multiple micro-ops, this event counts exactly once, as the last micro-op of the instruction retires. The event continues counting while instructions retire, including during interrupt service routines caused by hardware interrupts, faults or traps.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/knightslanding/uncore-cache.json +1
-0@@ -3246,6 +3246,7 @@ }, { "BriefDescription": "Uncore Clocks", + "EventCode": "0x0", "EventName": "UNC_H_U_CLOCKTICKS", "PerPkg": "1", "Unit": "CHA"tools/perf/pmu-events/arch/x86/knightslanding/uncore-memory.json +4
-0@@ -41,6 +41,7 @@ }, { "BriefDescription": "ECLK count", + "EventCode": "0x0", "EventName": "UNC_E_E_CLOCKTICKS", "PerPkg": "1", "Unit": "EDC_ECLK"@@ -55,6 +56,7 @@ }, { "BriefDescription": "UCLK count", + "EventCode": "0x0", "EventName": "UNC_E_U_CLOCKTICKS", "PerPkg": "1", "Unit": "EDC_UCLK"@@ -93,12 +95,14 @@ }, { "BriefDescription": "DCLK count", + "EventCode": "0x0", "EventName": "UNC_M_D_CLOCKTICKS", "PerPkg": "1", "Unit": "iMC_DCLK" }, { "BriefDescription": "UCLK count", + "EventCode": "0x0", "EventName": "UNC_M_U_CLOCKTICKS", "PerPkg": "1", "Unit": "iMC_UCLK"tools/perf/pmu-events/arch/x86/meteorlake/pipeline.json +8
-0@@ -37,6 +37,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "SampleAfterValue": "2000003", "UMask": "0x2",@@ -51,6 +52,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3",@@ -58,6 +60,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -75,6 +78,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "SampleAfterValue": "2000003", "UMask": "0x2",@@ -82,6 +86,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -105,6 +110,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of instructions retired", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "SampleAfterValue": "2000003",@@ -113,6 +119,7 @@ }, { "BriefDescription": "Number of instructions retired. Fixed Counter - architectural event", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",@@ -157,6 +164,7 @@ }, { "BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event", + "EventCode": "0x0", "EventName": "TOPDOWN.SLOTS", "PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",tools/perf/pmu-events/arch/x86/sandybridge/pipeline.json +4
-0@@ -337,6 +337,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "This event counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -359,6 +360,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "This event counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -367,6 +369,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -440,6 +443,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, this event counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/sapphirerapids/pipeline.json +5
-0@@ -284,6 +284,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -299,6 +300,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -426,6 +428,7 @@ }, { "BriefDescription": "Number of instructions retired. Fixed Counter - architectural event", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",@@ -457,6 +460,7 @@ }, { "BriefDescription": "Precise instruction retired with PEBS precise-distribution", + "EventCode": "0x0", "EventName": "INST_RETIRED.PREC_DIST", "PEBS": "1", "PublicDescription": "A version of INST_RETIRED that allows for a precise distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR++) feature to fix bias in how retired instructions get sampled. Use on Fixed Counter 0.",@@ -719,6 +723,7 @@ }, { "BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event", + "EventCode": "0x0", "EventName": "TOPDOWN.SLOTS", "PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003",tools/perf/pmu-events/arch/x86/sierraforest/pipeline.json +4
-0@@ -17,6 +17,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -29,6 +30,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "SampleAfterValue": "2000003", "UMask": "0x3"@@ -43,6 +45,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -55,6 +58,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of instructions retired", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/silvermont/pipeline.json +3
-0@@ -143,6 +143,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted core clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.CORE", "PublicDescription": "Counts the number of core cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time. For this reason this event may have a changing ratio with regards to time. In systems with a constant core frequency, this event can give you a measurement of the elapsed time while the core was not in halt state by dividing the event count by the core frequency. This event is architecturally defined and is a designated fixed counter. CPU_CLK_UNHALTED.CORE and CPU_CLK_UNHALTED.CORE_P use the core frequency which may change from time to time. CPU_CLK_UNHALTE.REF_TSC and CPU_CLK_UNHALTED.REF are not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. The fixed events are CPU_CLK_UNHALTED.CORE and CPU_CLK_UNHALTED.REF_TSC and the programmable events are CPU_CLK_UNHALTED.CORE_P and CPU_CLK_UNHALTED.REF.", "SampleAfterValue": "2000003",@@ -165,6 +166,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of unhalted reference clock cycles", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles while the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time. This event is not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. Divide this event count by core frequency to determine the elapsed time while the core was not in halt state. Divide this event count by core frequency to determine the elapsed time while the core was not in halt state. This event is architecturally defined and is a designated fixed counter. CPU_CLK_UNHALTED.CORE and CPU_CLK_UNHALTED.CORE_P use the core frequency which may change from time to time. CPU_CLK_UNHALTE.REF_TSC and CPU_CLK_UNHALTED.REF are not affected by core frequency changes but counts as if the core is running at the maximum frequency all the time. The fixed events are CPU_CLK_UNHALTED.CORE and CPU_CLK_UNHALTED.REF_TSC and the programmable events are CPU_CLK_UNHALTED.CORE_P and CPU_CLK_UNHALTED.REF.", "SampleAfterValue": "2000003",@@ -180,6 +182,7 @@ }, { "BriefDescription": "Fixed Counter: Counts the number of instructions retired", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "This event counts the number of instructions that retire. For instructions that consist of multiple micro-ops, this event counts exactly once, as the last micro-op of the instruction retires. The event continues counting while instructions retire, including during interrupt service routines caused by hardware interrupts, faults or traps. Background: Modern microprocessors employ extensive pipelining and speculative techniques. Since sometimes an instruction is started but never completed, the notion of \"retirement\" is introduced. A retired instruction is one that commits its states. Or stated differently, an instruction might be abandoned at some point. No instruction is truly finished until it retires. This counter measures the number of completed instructions. The fixed event is INST_RETIRED.ANY and the programmable event is INST_RETIRED.ANY_P.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/skylake/pipeline.json +4
-0@@ -182,6 +182,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -213,6 +214,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -221,6 +223,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -360,6 +363,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "Counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, Counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. Counting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/skylakex/pipeline.json +4
-0@@ -191,6 +191,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -222,6 +223,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -230,6 +232,7 @@ { "AnyThread": "1", "BriefDescription": "Core cycles when at least one thread on the physical core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD_ANY", "SampleAfterValue": "2000003", "UMask": "0x2"@@ -369,6 +372,7 @@ }, { "BriefDescription": "Instructions retired from execution.", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PublicDescription": "Counts the number of instructions retired from execution. For instructions that consist of multiple micro-ops, Counts the retirement of the last micro-op of the instruction. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter, leaving the four (eight when Hyperthreading is disabled) programmable counters available for other events. INST_RETIRED.ANY_P is counted by a programmable counter and it is an architectural performance event. Counting: Faulting executions of GETSEC/VM entry/VM Exit/MWait will not count as retired instructions.", "SampleAfterValue": "2000003",tools/perf/pmu-events/arch/x86/skylakex/uncore-cache.json +2
-0@@ -460,6 +460,7 @@ }, { "BriefDescription": "Clockticks of the uncore caching & home agent (CHA)", + "EventCode": "0x0", "EventName": "UNC_CHA_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts clockticks of the clock controlling the uncore caching and home agent (CHA).",@@ -5678,6 +5679,7 @@ { "BriefDescription": "This event is deprecated. Refer to new event UNC_CHA_CLOCKTICKS", "Deprecated": "1", + "EventCode": "0x0", "EventName": "UNC_C_CLOCKTICKS", "PerPkg": "1", "Unit": "CHA"tools/perf/pmu-events/arch/x86/skylakex/uncore-interconnect.json +1
-0@@ -1090,6 +1090,7 @@ }, { "BriefDescription": "Cycles - at UCLK", + "EventCode": "0x0", "EventName": "UNC_M2M_CLOCKTICKS", "PerPkg": "1", "Unit": "M2M"tools/perf/pmu-events/arch/x86/skylakex/uncore-io.json +1
-0@@ -1271,6 +1271,7 @@ }, { "BriefDescription": "Counting disabled", + "EventCode": "0x0", "EventName": "UNC_IIO_NOTHING", "PerPkg": "1", "Unit": "IIO"tools/perf/pmu-events/arch/x86/skylakex/uncore-memory.json +1
-0@@ -167,6 +167,7 @@ }, { "BriefDescription": "Memory controller clock ticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Counts clockticks of the fixed frequency clock of the memory controller using one of the programmable counters.",tools/perf/pmu-events/arch/x86/skylakex/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "pclk Cycles", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "The PCU runs off a fixed 1 GHz clock. This event counts the number of pclk cycles measured while the counter was enabled. The pclk, like the Memory Controller's dclk, counts at a constant rate making it a good measure of actual wall time.",tools/perf/pmu-events/arch/x86/snowridgex/pipeline.json +2
-0@@ -150,6 +150,7 @@ }, { "BriefDescription": "Counts the number of unhalted reference clock cycles at TSC frequency. (Fixed event)", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles that the core is not in a halt state. The core enters the halt state when it is running the HLT instruction. This event is not affected by core frequency changes and increments at a fixed frequency that is also used for the Time Stamp Counter (TSC). This event uses fixed counter 2.", "SampleAfterValue": "2000003",@@ -179,6 +180,7 @@ }, { "BriefDescription": "Counts the total number of instructions retired. (Fixed event)", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the total number of instructions that retired. For instructions that consist of multiple uops, this event counts the retirement of the last uop of the instruction. This event continues counting during hardware interrupts, traps, and inside interrupt handlers. This event uses fixed counter 0.",tools/perf/pmu-events/arch/x86/snowridgex/uncore-cache.json +1
-0@@ -872,6 +872,7 @@ }, { "BriefDescription": "Uncore cache clock ticks", + "EventCode": "0x0", "EventName": "UNC_CHA_CLOCKTICKS", "PerPkg": "1", "Unit": "CHA"tools/perf/pmu-events/arch/x86/snowridgex/uncore-interconnect.json +1
-0@@ -1419,6 +1419,7 @@ }, { "BriefDescription": "Clockticks of the mesh to memory (M2M)", + "EventCode": "0x0", "EventName": "UNC_M2M_CLOCKTICKS", "PerPkg": "1", "Unit": "M2M"tools/perf/pmu-events/arch/x86/snowridgex/uncore-memory.json +1
-0@@ -120,6 +120,7 @@ }, { "BriefDescription": "Memory controller clock ticks", + "EventCode": "0x0", "EventName": "UNC_M_CLOCKTICKS", "PerPkg": "1", "PublicDescription": "Clockticks of the integrated memory controller (IMC)",tools/perf/pmu-events/arch/x86/snowridgex/uncore-power.json +1
-0@@ -1,6 +1,7 @@ [ { "BriefDescription": "Clockticks of the power control unit (PCU)", + "EventCode": "0x0", "EventName": "UNC_P_CLOCKTICKS", "PerPkg": "1", "Unit": "PCU"tools/perf/pmu-events/arch/x86/tigerlake/pipeline.json +5
-1@@ -193,6 +193,7 @@ }, { "BriefDescription": "Reference cycles when the core is not in halt state.", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.REF_TSC", "PublicDescription": "Counts the number of reference cycles when the core is not in a halt state. The core enters the halt state when it is running the HLT instruction or the MWAIT instruction. This event is not affected by core frequency changes (for example, P states, TM2 transitions) but has the same incrementing frequency as the time stamp counter. This event can approximate elapsed time while the core was not in a halt state. This event has a constant ratio with the CPU_CLK_UNHALTED.REF_XCLK event. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events. Note: On all current platforms this event stops counting during 'throttling (TM)' states duty off periods the processor is 'halted'. The counter update is done at a lower clock rate then the core clock the overflow status bit for this counter may appear 'sticky'. After the counter has overflowed and software clears the overflow status bit and resets the counter to less than MAX. The reset value to the counter is not clocked immediately so the overflow status bit will flip 'high (1)' and generate another PMI (if enabled) after which the reset value gets clocked into the counter. Therefore, software will get the interrupt, read the overflow status bit '1 for bit 34 while the counter value is less than MAX. Software should ignore this case.", "SampleAfterValue": "2000003",@@ -208,6 +209,7 @@ }, { "BriefDescription": "Core cycles when the thread is not in halt state", + "EventCode": "0x0", "EventName": "CPU_CLK_UNHALTED.THREAD", "PublicDescription": "Counts the number of core cycles while the thread is not in a halt state. The thread enters the halt state when it is running the HLT instruction. This event is a component in many key event ratios. The core frequency may change from time to time due to transitions associated with Enhanced Intel SpeedStep Technology or TM2. For this reason this event may have a changing ratio with regards to time. When the core frequency is constant, this event can approximate elapsed time while the core was not in the halt state. It is counted on a dedicated fixed counter, leaving the eight programmable counters available for other events.", "SampleAfterValue": "2000003",@@ -352,6 +354,7 @@ }, { "BriefDescription": "Number of instructions retired. Fixed Counter - architectural event", + "EventCode": "0x0", "EventName": "INST_RETIRED.ANY", "PEBS": "1", "PublicDescription": "Counts the number of X86 instructions retired - an Architectural PerfMon event. Counting continues during hardware interrupts, traps, and inside interrupt handlers. Notes: INST_RETIRED.ANY is counted by a designated fixed counter freeing up programmable counters to count other events. INST_RETIRED.ANY_P is counted by a programmable counter.",@@ -377,6 +380,7 @@ }, { "BriefDescription": "Precise instruction retired event with a reduced effect of PEBS shadow in IP distribution", + "EventCode": "0x0", "EventName": "INST_RETIRED.PREC_DIST", "PEBS": "1", "PublicDescription": "A version of INST_RETIRED that allows for a more unbiased distribution of samples across instructions retired. It utilizes the Precise Distribution of Instructions Retired (PDIR) feature to mitigate some bias in how retired instructions get sampled. Use on Fixed Counter 0.",@@ -569,6 +573,7 @@ }, { "BriefDescription": "TMA slots available for an unhalted logical processor. Fixed counter - architectural event", + "EventCode": "0x0", "EventName": "TOPDOWN.SLOTS", "PublicDescription": "Number of available slots for an unhalted logical processor. The event increments by machine-width of the narrowest pipeline as employed by the Top-down Microarchitecture Analysis method (TMA). The count is distributed among unhalted logical processors (hyper-threads) who share the same physical core. Software can use this event as the denominator for the top-level metrics of the TMA method. This architectural event is counted on a designated fixed counter (Fixed Counter 3).", "SampleAfterValue": "10000003", --1.8.3.1
PATCH v7 4/8 perf jevents: Support more event fields
In reply to Jing ZhangThe previous code assumes an event has either an "event=" or "config" field at the beginning. For CMN neither of these may be present, as an event is typically "type=xx,eventid=xxx".
If EventCode and ConfigCode is not added in the alias JSON file, the event description will add "event=0" by default. So, even if the event field is added "eventid=xxx" and "type=xxx", the CMN events final parsing result will be "event=0, eventid=xxx, type=xxx".
Therefore, when EventCode and ConfigCode are missing in JSON, "event=0" is no longer added by default. And add EventIdCode and Type to the event field.
I compared pmu_event.c before and after compiling with JEVENT_ARCH=all, they are consistent.
--- tools/perf/pmu-events/jevents.py | 20 ++++++++++++++++---- 1 file changed, 16 insertions(+), 4 deletions(-)
tools/perf/pmu-events/jevents.py +16 -4
@@ -275,11 +275,14 @@ class JsonEvent: } return table[unit] if unit in table else f'uncore_{unit.lower()}' - eventcode = 0 + eventcode = None if 'EventCode' in jd: eventcode = int(jd['EventCode'].split(',', 1)[0], 0) if 'ExtSel' in jd: - eventcode |= int(jd['ExtSel']) << 8 + if eventcode is None: + eventcode = int(jd['ExtSel']) << 8 + else: + eventcode |= int(jd['ExtSel']) << 8 configcode = int(jd['ConfigCode'], 0) if 'ConfigCode' in jd else None self.name = jd['EventName'].lower() if 'EventName' in jd else None self.topic = ''
@@ -317,7 +320,11 @@ class JsonEvent: if precise and self.desc and '(Precise Event)' not in self.desc: extra_desc += ' (Must be precise)' if precise == '2' else (' (Precise ' 'event)') - event = f'config={llx(configcode)}' if configcode is not None else f'event={llx(eventcode)}' + event = None + if eventcode is not None: + event = f'event={llx(eventcode)}' + elif configcode is not None: + event = f'config={llx(configcode)}' event_fields = [ ('AnyThread', 'any='), ('PortMask', 'ch_mask='),
@@ -327,10 +334,15 @@ class JsonEvent: ('Invert', 'inv='), ('SampleAfterValue', 'period='), ('UMask', 'umask='), + ('NodeType', 'type='), + ('EventIdCode', 'eventid='), ] for key, value in event_fields: if key in jd and jd[key] != '0': - event += ',' + value + jd[key] + if event: + event += ',' + value + jd[key] + else: + event = value + jd[key] if filter: event += f',{filter}' if msr:
Signature · 1 line
1.8.3.1
Robin Murphy
maintainerRe: [PATCH v7 4/8] perf jevents: Support more event fields
Quoted 56 lines · Jing Zhang
The previous code assumes an event has either an "event=" or "config" field at the beginning. For CMN neither of these may be present, as an event is typically "type=xx,eventid=xxx".
If EventCode and ConfigCode is not added in the alias JSON file, the event description will add "event=0" by default. So, even if the event field is added "eventid=xxx" and "type=xxx", the CMN events final parsing result will be "event=0, eventid=xxx, type=xxx".
Therefore, when EventCode and ConfigCode are missing in JSON, "event=0" is no longer added by default. And add EventIdCode and Type to the event field.
I compared
pmu_event.cbefore and after compiling with JEVENT_ARCH=all, they are consistent.Signed-off-by Jing Zhang <[email protected]>--- tools/perf/pmu-events/jevents.py | 20 ++++++++++++++++---- 1 file changed, 16 insertions(+), 4 deletions(-)tools/perf/pmu-events/jevents.py +10
-3@@ -275,11 +275,14 @@ class JsonEvent: } return table[unit] if unit in table else f'uncore_{unit.lower()}' - eventcode = 0 + eventcode = None if 'EventCode' in jd: eventcode = int(jd['EventCode'].split(',', 1)[0], 0) if 'ExtSel' in jd: - eventcode |= int(jd['ExtSel']) << 8 + if eventcode is None: + eventcode = int(jd['ExtSel']) << 8 + else: + eventcode |= int(jd['ExtSel']) << 8 configcode = int(jd['ConfigCode'], 0) if 'ConfigCode' in jd else None self.name = jd['EventName'].lower() if 'EventName' in jd else None self.topic = ''@@ -317,7 +320,11 @@ class JsonEvent: if precise and self.desc and '(Precise Event)' not in self.desc: extra_desc += ' (Must be precise)' if precise == '2' else (' (Precise ' 'event)') - event = f'config={llx(configcode)}' if configcode is not None else f'event={llx(eventcode)}' + event = None + if eventcode is not None: + event = f'event={llx(eventcode)}' + elif configcode is not None: + event = f'config={llx(configcode)}' event_fields = [ ('AnyThread', 'any='), ('PortMask', 'ch_mask='),
tools/perf/pmu-events/jevents.py class JsonEvent: · line 338
Quoted 5 lines
tools/perf/pmu-events/jevents.py +2
-0@@ -327,10 +334,15 @@ class JsonEvent: ('Invert', 'inv='), ('SampleAfterValue', 'period='), ('UMask', 'umask='), + ('NodeType', 'type='), + ('EventIdCode', 'eventid='),
FWIW, this smells like another brewing scalability problem, given that these are entirely driver-specific. Not sure off-hand how feasible it might be, but my instinct says that a neat solution would be to encode them right in the JSON, e.g.:
"FormatAttr": { "type": 0x5 }such that jevents should then only really need to consider whether an event is defined in terms of a raw "ConfigCode", one or more "FormatAttr"s which it can then parse dynamically, or reasonable special cases like "EventCode" (given how "event" is one of the most commonly used formats).
Thanks,
Robin.
tools/perf/pmu-events/jevents.py class JsonEvent: · line 348
Quoted 11 lines
tools/perf/pmu-events/jevents.py +4
-1] for key, value in event_fields: if key in jd and jd[key] != '0': - event += ',' + value + jd[key] + if event: + event += ',' + value + jd[key] + else: + event = value + jd[key] if filter: event += f',{filter}' if msr:
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Ian Rogers
reviewerRe: [PATCH v7 4/8] perf jevents: Support more event fields
Quoted 69 lines · Robin Murphy
Quoted 56 lines · Jing Zhang
The previous code assumes an event has either an "event=" or "config" field at the beginning. For CMN neither of these may be present, as an event is typically "type=xx,eventid=xxx".
If EventCode and ConfigCode is not added in the alias JSON file, the event description will add "event=0" by default. So, even if the event field is added "eventid=xxx" and "type=xxx", the CMN events final parsing result will be "event=0, eventid=xxx, type=xxx".
Therefore, when EventCode and ConfigCode are missing in JSON, "event=0" is no longer added by default. And add EventIdCode and Type to the event field.
I compared
pmu_event.cbefore and after compiling with JEVENT_ARCH=all, they are consistent.Signed-off-by Jing Zhang <[email protected]>--- tools/perf/pmu-events/jevents.py | 20 ++++++++++++++++---- 1 file changed, 16 insertions(+), 4 deletions(-)tools/perf/pmu-events/jevents.py +10
-3@@ -275,11 +275,14 @@ class JsonEvent: } return table[unit] if unit in table else f'uncore_{unit.lower()}' - eventcode = 0 + eventcode = None if 'EventCode' in jd: eventcode = int(jd['EventCode'].split(',', 1)[0], 0) if 'ExtSel' in jd: - eventcode |= int(jd['ExtSel']) << 8 + if eventcode is None: + eventcode = int(jd['ExtSel']) << 8 + else: + eventcode |= int(jd['ExtSel']) << 8 configcode = int(jd['ConfigCode'], 0) if 'ConfigCode' in jd else None self.name = jd['EventName'].lower() if 'EventName' in jd else None self.topic = ''@@ -317,7 +320,11 @@ class JsonEvent: if precise and self.desc and '(Precise Event)' not in self.desc: extra_desc += ' (Must be precise)' if precise == '2' else (' (Precise ' 'event)') - event = f'config={llx(configcode)}' if configcode is not None else f'event={llx(eventcode)}' + event = None + if eventcode is not None: + event = f'event={llx(eventcode)}' + elif configcode is not None: + event = f'config={llx(configcode)}' event_fields = [ ('AnyThread', 'any='), ('PortMask', 'ch_mask='),tools/perf/pmu-events/jevents.py class JsonEvent: · line 338
Quoted 5 lines
tools/perf/pmu-events/jevents.py +2
-0@@ -327,10 +334,15 @@ class JsonEvent: ('Invert', 'inv='), ('SampleAfterValue', 'period='), ('UMask', 'umask='), + ('NodeType', 'type='), + ('EventIdCode', 'eventid='),FWIW, this smells like another brewing scalability problem, given that these are entirely driver-specific. Not sure off-hand how feasible it might be, but my instinct says that a neat solution would be to encode them right in the JSON, e.g.:
Quoted 7 lines
"FormatAttr": { "type": 0x5 }such that jevents should then only really need to consider whether an event is defined in terms of a raw "ConfigCode", one or more "FormatAttr"s which it can then parse dynamically, or reasonable special cases like "EventCode" (given how "event" is one of the most commonly used formats).
Hi Robin,
I'm not sure about scalability but I think it is a problem that we encode names into the event that should correspond with formats, but we don't test that the PMU driver in question supports the format. If we tar up sysfs directories I think we can check/test this and it makes sense for the perf tool to be able to parse a sysfs type directory structure.
I think the hard coded names and matching to dictionary entries is suboptimal, we should really be able to know the formats, abstract this, etc. The code is deliberately done this way so that we could migrate away from a legacy C version of this code and generate an identical pmu-events.c. As much as is possible the python code matches the C code, but now the transition has happened there is no reason to maintain this behavior.
Thanks,
Ian
Quoted 14 lines
Thanks,
Robin.Quoted 11 lines
tools/perf/pmu-events/jevents.py +4
-1] for key, value in event_fields: if key in jd and jd[key] != '0': - event += ',' + value + jd[key] + if event: + event += ',' + value + jd[key] + else: + event = value + jd[key] if filter: event += f',{filter}' if msr:
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
PATCH v7 5/8 perf test: Make matching_pmu effective
In reply to Jing ZhangThe perf_pmu_test_event.matching_pmu didn't work. No matter what its value is, it does not affect the test results. So let matching_pmu be used for matching perf_pmu_test_pmu.pmu.name.
--- tools/perf/tests/pmu-events.c | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-)
tools/perf/tests/pmu-events.c +7 -2
@@ -238,7 +238,7 @@ struct perf_pmu_test_pmu { }, .alias_str = "event=0x2b", .alias_long_desc = "ddr write-cycles event. Unit: uncore_sys_ddr_pmu ", - .matching_pmu = "uncore_sys_ddr_pmu", + .matching_pmu = "uncore_sys_ddr_pmu0", };static const struct perf_pmu_test_event sys_ccn_pmu_read_cycles = {
@@ -252,7 +252,7 @@ struct perf_pmu_test_pmu { }, .alias_str = "config=0x2c", .alias_long_desc = "ccn read-cycles event. Unit: uncore_sys_ccn_pmu ", - .matching_pmu = "uncore_sys_ccn_pmu", + .matching_pmu = "uncore_sys_ccn_pmu4", };static const struct perf_pmu_test_event *sys_events[] = {
@@ -599,6 +599,11 @@ static int __test_uncore_pmu_event_aliases(struct perf_pmu_test_pmu *test_pmu)struct pmu_event const *event = &test_event->event;if (!strcmp (event->name, alias->name)) { +if (strcmp (pmu_name, test_event->matching_pmu)) { +pr_debug ("testing aliases uncore PMU %s: mismatched matching_pmu, %s vs %s\n" , + pmu_name, test_event->matching_pmu, pmu_name); +continue ; + }if (compare_alias_to_test_event (alias, test_event, pmu_name)) {
Signature · 1 line
1.8.3.1
Ian Rogers
reviewerRe: [PATCH v7 5/8] perf test: Make matching_pmu effective
Quoted 4 lines · Jing Zhang
The perf_pmu_test_event.matching_pmu didn't work. No matter what its value is, it does not affect the test results. So let
matching_pmube used for matching perf_pmu_test_pmu.pmu.name.
Could you rebase this onto the latest perf-tools-next, I'd like to test this.
Thanks,
Ian
Quoted 41 lines
Signed-off-by Jing Zhang <[email protected]>Reviewed-by John Garry <[email protected]>--- tools/perf/tests/pmu-events.c | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-)tools/perf/tests/pmu-events.c +7
-3@@ -238,7 +238,7 @@ struct perf_pmu_test_pmu { }, .alias_str = "event=0x2b", .alias_long_desc = "ddr write-cycles event. Unit: uncore_sys_ddr_pmu ", - .matching_pmu = "uncore_sys_ddr_pmu", + .matching_pmu = "uncore_sys_ddr_pmu0", };static const struct perf_pmu_test_event sys_ccn_pmu_read_cycles = {@@ -252,7 +252,7 @@ struct perf_pmu_test_pmu { }, .alias_str = "config=0x2c", .alias_long_desc = "ccn read-cycles event. Unit: uncore_sys_ccn_pmu ", - .matching_pmu = "uncore_sys_ccn_pmu", + .matching_pmu = "uncore_sys_ccn_pmu4", };static const struct perf_pmu_test_event *sys_events[] = {@@ -599,6 +599,11 @@ static int __test_uncore_pmu_event_aliases(struct perf_pmu_test_pmu *test_pmu)struct pmu_event const *event = &test_event->event;if (!strcmp (event->name, alias->name)) { +if (strcmp (pmu_name, test_event->matching_pmu)) { +pr_debug ("testing aliases uncore PMU %s: mismatched matching_pmu, %s vs %s\n" , + pmu_name, test_event->matching_pmu, pmu_name); +continue ; + }if (compare_alias_to_test_event (alias, test_event, pmu_name)) { --
Quoted 2 lines
1.8.3.1
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Re: [PATCH v7 5/8] perf test: Make matching_pmu effective
Quoted 8 lines · Ian Rogers
Quoted 4 lines · Jing Zhang
The perf_pmu_test_event.matching_pmu didn't work. No matter what its value is, it does not affect the test results. So let
matching_pmube used for matching perf_pmu_test_pmu.pmu.name.Could you rebase this onto the latest perf-tools-next, I'd like to test this.
Ok, I will rebase it onto the latest perf-tools-next in next version.
Thanks,
Jing
Quoted 34 lines
Thanks,
IanQuoted 31 lines
Signed-off-by Jing Zhang <[email protected]>Reviewed-by John Garry <[email protected]>--- tools/perf/tests/pmu-events.c | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-)tools/perf/tests/pmu-events.c +2
-2@@ -238,7 +238,7 @@ struct perf_pmu_test_pmu { }, .alias_str = "event=0x2b", .alias_long_desc = "ddr write-cycles event. Unit: uncore_sys_ddr_pmu ", - .matching_pmu = "uncore_sys_ddr_pmu", + .matching_pmu = "uncore_sys_ddr_pmu0", };static const struct perf_pmu_test_event sys_ccn_pmu_read_cycles = {@@ -252,7 +252,7 @@ struct perf_pmu_test_pmu { }, .alias_str = "config=0x2c", .alias_long_desc = "ccn read-cycles event. Unit: uncore_sys_ccn_pmu ", - .matching_pmu = "uncore_sys_ccn_pmu", + .matching_pmu = "uncore_sys_ccn_pmu4", };static const struct perf_pmu_test_event *sys_events[] = {@@ -599,6 +599,11 @@ static int __test_uncore_pmu_event_aliases(struct perf_pmu_test_pmu *test_pmu)struct pmu_event const *event = &test_event->event;
Quoted 12 lines
Quoted 10 lines
tools/perf/tests/pmu-events.c +5
-1if (!strcmp (event->name, alias->name)) { +if (strcmp (pmu_name, test_event->matching_pmu)) { +pr_debug ("testing aliases uncore PMU %s: mismatched matching_pmu, %s vs %s\n" , + pmu_name, test_event->matching_pmu, pmu_name); +continue ; + }if (compare_alias_to_test_event (alias, test_event, pmu_name)) { --Quoted 2 lines
1.8.3.1
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
PATCH v7 6/8 perf test: Add pmu-event test for "Compat" and new event_field.
In reply to Jing ZhangAdd new event test for uncore system event which is used to verify the functionality of "Compat" matching multiple identifiers and the new event fields "EventIdCode" and "Type".
--- .../pmu-events/arch/test/test_soc/sys/uncore.json | 8 ++++ tools/perf/pmu-events/empty-pmu-events.c | 8 ++++ tools/perf/tests/pmu-events.c | 55 ++++++++++++++++++++++ 3 files changed, 71 insertions(+)
tools/perf/pmu-events/arch/test/test_soc/sys/uncore.json +8 -0
@@ -12,5 +12,13 @@ "EventName": "sys_ccn_pmu.read_cycles", "Unit": "sys_ccn_pmu", "Compat": "0x01" + }, + { + "BriefDescription": "Counts total cache misses in first lookup result (high priority).", + "NodeType": "0x05", + "EventIdCode": "0x01", + "EventName": "sys_cmn_pmu.hnf_cache_miss", + "Unit": "sys_cmn_pmu", + "Compat": "434*;436*;43c*;43a01" } ]
tools/perf/pmu-events/empty-pmu-events.c +8 -0
@@ -245,6 +245,14 @@ struct pmu_events_map { .pmu = "uncore_sys_ccn_pmu", }, { + .name = "sys_cmn_pmu.hnf_cache_miss", + .event = "type=0x05,eventid=0x01", + .desc = "Counts total cache misses in first lookup result (high priority). Unit: uncore_sys_cmn_pmu ", + .compat = "434*;436*;43c*;43a01", + .topic = "uncore", + .pmu = "uncore_sys_cmn_pmu", + }, + { .name = 0, .event = 0, .desc = 0,
tools/perf/tests/pmu-events.c +55 -0
70 lines not on this page · raw message
Signature · 1 line
1.8.3.1
Ian Rogers
reviewerRe: [PATCH v7 6/8] perf test: Add pmu-event test for "Compat" and new event_field.
Quoted 6 lines · Jing Zhang
Add new event test for uncore system event which is used to verify the functionality of "Compat" matching multiple identifiers and the new event fields "EventIdCode" and "Type".
Signed-off-by Jing Zhang <[email protected]>
Thanks for the tests! I've no issue with them beside the already mentioned ';'. This will need updating for:
https://lore.kernel.org/lkml/[email protected]/
https://lore.kernel.org/lkml/[email protected]/
Thanks,
Ian
Quoted 120 lines
--- .../pmu-events/arch/test/test_soc/sys/uncore.json | 8 ++++ tools/perf/pmu-events/empty-pmu-events.c | 8 ++++ tools/perf/tests/pmu-events.c | 55 ++++++++++++++++++++++ 3 files changed, 71 insertions(+)tools/perf/pmu-events/arch/test/test_soc/sys/uncore.json +8
-0@@ -12,5 +12,13 @@ "EventName": "sys_ccn_pmu.read_cycles", "Unit": "sys_ccn_pmu", "Compat": "0x01" + }, + { + "BriefDescription": "Counts total cache misses in first lookup result (high priority).", + "NodeType": "0x05", + "EventIdCode": "0x01", + "EventName": "sys_cmn_pmu.hnf_cache_miss", + "Unit": "sys_cmn_pmu", + "Compat": "434*;436*;43c*;43a01" } ]tools/perf/pmu-events/empty-pmu-events.c +8
-0@@ -245,6 +245,14 @@ struct pmu_events_map { .pmu = "uncore_sys_ccn_pmu", }, { + .name = "sys_cmn_pmu.hnf_cache_miss", + .event = "type=0x05,eventid=0x01", + .desc = "Counts total cache misses in first lookup result (high priority). Unit: uncore_sys_cmn_pmu ", + .compat = "434*;436*;43c*;43a01", + .topic = "uncore", + .pmu = "uncore_sys_cmn_pmu", + }, + { .name = 0, .event = 0, .desc = 0,tools/perf/tests/pmu-events.c +55
-1@@ -255,9 +255,24 @@ struct perf_pmu_test_pmu { .matching_pmu = "uncore_sys_ccn_pmu4", }; +static const struct perf_pmu_test_event sys_cmn_pmu_hnf_cache_miss = { + .event = { + .name ="sys_cmn_pmu.hnf_cache_miss" , + .event ="type=0x05,eventid=0x01" , + .desc ="Counts total cache misses in first lookup result (high priority). Unit: uncore_sys_cmn_pmu " , + .topic ="uncore" , + .pmu ="uncore_sys_cmn_pmu" , + .compat ="434*;436*;43c*;43a01" , + }, + .alias_str ="type=0x5,eventid=0x1" , + .alias_long_desc ="Counts total cache misses in first lookup result (high priority). Unit: uncore_sys_cmn_pmu " , + .matching_pmu ="uncore_sys_cmn_pmu0" , +}; +static const struct perf_pmu_test_event *sys_events[] = { &sys_ddr_pmu_write_cycles, &sys_ccn_pmu_read_cycles, + &sys_cmn_pmu_hnf_cache_miss,NULL };@@ -704,6 +719,46 @@ static int __test_uncore_pmu_event_aliases(struct perf_pmu_test_pmu *test_pmu) &sys_ccn_pmu_read_cycles, }, }, + { + .pmu = { + .name = (char *)"uncore_sys_cmn_pmu0", + .is_uncore = 1, + .id = (char *)"43401", + }, + .aliases = { + &sys_cmn_pmu_hnf_cache_miss, + }, + }, + { + .pmu = { + .name = (char *)"uncore_sys_cmn_pmu0", + .is_uncore = 1, + .id = (char *)"43602", + }, + .aliases = { + &sys_cmn_pmu_hnf_cache_miss, + }, + }, + { + .pmu = { + .name = (char *)"uncore_sys_cmn_pmu0", + .is_uncore = 1, + .id = (char *)"43c03", + }, + .aliases = { + &sys_cmn_pmu_hnf_cache_miss, + }, + }, + { + .pmu = { + .name = (char *)"uncore_sys_cmn_pmu0", + .is_uncore = 1, + .id = (char *)"43a01", + }, + .aliases = { + &sys_cmn_pmu_hnf_cache_miss, + }, + } }; /* Test that aliases generated are as expected */ --
Quoted 2 lines
1.8.3.1
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Re: [PATCH v7 6/8] perf test: Add pmu-event test for "Compat" and new event_field.
Quoted 6 lines · Ian Rogers
Quoted 5 lines · Jing Zhang
Add new event test for uncore system event which is used to verify the functionality of "Compat" matching multiple identifiers and the new event fields "EventIdCode" and "Type".
Quoted 7 lines
Quoted 1 line
Signed-off-by Jing Zhang <[email protected]>Thanks for the tests! I've no issue with them beside the already mentioned ';'. This will need updating for:
https://lore.kernel.org/lkml/[email protected]/
https://lore.kernel.org/lkml/[email protected]/
Ok, will do.
Thanks,
Jing
Quoted 121 lines
Thanks,
IanQuoted 118 lines
--- .../pmu-events/arch/test/test_soc/sys/uncore.json | 8 ++++ tools/perf/pmu-events/empty-pmu-events.c | 8 ++++ tools/perf/tests/pmu-events.c | 55 ++++++++++++++++++++++ 3 files changed, 71 insertions(+)tools/perf/pmu-events/arch/test/test_soc/sys/uncore.json +8
-0@@ -12,5 +12,13 @@ "EventName": "sys_ccn_pmu.read_cycles", "Unit": "sys_ccn_pmu", "Compat": "0x01" + }, + { + "BriefDescription": "Counts total cache misses in first lookup result (high priority).", + "NodeType": "0x05", + "EventIdCode": "0x01", + "EventName": "sys_cmn_pmu.hnf_cache_miss", + "Unit": "sys_cmn_pmu", + "Compat": "434*;436*;43c*;43a01" } ]tools/perf/pmu-events/empty-pmu-events.c +8
-0@@ -245,6 +245,14 @@ struct pmu_events_map { .pmu = "uncore_sys_ccn_pmu", }, { + .name = "sys_cmn_pmu.hnf_cache_miss", + .event = "type=0x05,eventid=0x01", + .desc = "Counts total cache misses in first lookup result (high priority). Unit: uncore_sys_cmn_pmu ", + .compat = "434*;436*;43c*;43a01", + .topic = "uncore", + .pmu = "uncore_sys_cmn_pmu", + }, + { .name = 0, .event = 0, .desc = 0,tools/perf/tests/pmu-events.c +55
-0@@ -255,9 +255,24 @@ struct perf_pmu_test_pmu { .matching_pmu = "uncore_sys_ccn_pmu4", }; +static const struct perf_pmu_test_event sys_cmn_pmu_hnf_cache_miss = { + .event = { + .name ="sys_cmn_pmu.hnf_cache_miss" , + .event ="type=0x05,eventid=0x01" , + .desc ="Counts total cache misses in first lookup result (high priority). Unit: uncore_sys_cmn_pmu " , + .topic ="uncore" , + .pmu ="uncore_sys_cmn_pmu" , + .compat ="434*;436*;43c*;43a01" , + }, + .alias_str ="type=0x5,eventid=0x1" , + .alias_long_desc ="Counts total cache misses in first lookup result (high priority). Unit: uncore_sys_cmn_pmu " , + .matching_pmu ="uncore_sys_cmn_pmu0" , +}; +static const struct perf_pmu_test_event *sys_events[] = { &sys_ddr_pmu_write_cycles, &sys_ccn_pmu_read_cycles, + &sys_cmn_pmu_hnf_cache_miss,NULL };@@ -704,6 +719,46 @@ static int __test_uncore_pmu_event_aliases(struct perf_pmu_test_pmu *test_pmu) &sys_ccn_pmu_read_cycles, }, }, + { + .pmu = { + .name = (char *)"uncore_sys_cmn_pmu0", + .is_uncore = 1, + .id = (char *)"43401", + }, + .aliases = { + &sys_cmn_pmu_hnf_cache_miss, + }, + }, + { + .pmu = { + .name = (char *)"uncore_sys_cmn_pmu0", + .is_uncore = 1, + .id = (char *)"43602", + }, + .aliases = { + &sys_cmn_pmu_hnf_cache_miss, + }, + }, + { + .pmu = { + .name = (char *)"uncore_sys_cmn_pmu0", + .is_uncore = 1, + .id = (char *)"43c03", + }, + .aliases = { + &sys_cmn_pmu_hnf_cache_miss, + }, + }, + { + .pmu = { + .name = (char *)"uncore_sys_cmn_pmu0", + .is_uncore = 1, + .id = (char *)"43a01", + }, + .aliases = { + &sys_cmn_pmu_hnf_cache_miss, + }, + } };
Quoted 4 lines
Quoted 4 lines
tools/perf/tests/pmu-events.c +0
-1/* Test that aliases generated are as expected */ --1.8.3.1
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
PATCH v7 7/8 perf jevents: Add support for Arm CMN PMU aliasing
In reply to Jing ZhangCurrently just add aliases for part of Arm CMN PMU events which are general and compatible for any SoC and CMN-ANY.
"Compat" value "434*;436*;43c*;43a*" means it is compatible with all CMN600/CMN650/CMN700/Ci700, which can be obtained from commit 7819e05a0dce ("perf/arm-cmn: Revamp model detection").
The arm-cmn PMU events got from:
[0] https://developer.arm.com/documentation/100180/0302/?lang=en
[1] https://developer.arm.com/documentation/101408/0100/?lang=en
[2] https://developer.arm.com/documentation/102308/0302/?lang=en
[3] https://developer.arm.com/documentation/101569/0300/?lang=en
--- .../pmu-events/arch/arm64/arm/cmn/sys/cmn.json | 266 +++++++++++++++++++++ tools/perf/pmu-events/jevents.py | 1 + 2 files changed, 267 insertions(+) create mode 100644 tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.json
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.json new file +266 -0
@@ -0,0 +1,266 @@ +[ + { + "EventName": "hnf_cache_miss", + "EventIdCode": "0x1", + "NodeType": "0x5", + "BriefDescription": "Counts total cache misses in first lookup result (high priority).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_slc_sf_cache_access", + "EventIdCode": "0x2", + "NodeType": "0x5", + "BriefDescription": "Counts number of cache accesses in first access (high priority).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_cache_fill", + "EventIdCode": "0x3", + "NodeType": "0x5", + "BriefDescription": "Counts total allocations in HN SLC (all cache line allocations to SLC).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_pocq_retry", + "EventIdCode": "0x4", + "NodeType": "0x5", + "BriefDescription": "Counts number of retried requests.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_pocq_reqs_recvd", + "EventIdCode": "0x5", + "NodeType": "0x5", + "BriefDescription": "Counts number of requests that HN receives.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_sf_hit", + "EventIdCode": "0x6", + "NodeType": "0x5", + "BriefDescription": "Counts number of SF hits.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_sf_evictions", + "EventIdCode": "0x7", + "NodeType": "0x5", + "BriefDescription": "Counts number of SF eviction cache invalidations initiated.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_dir_snoops_sent", + "EventIdCode": "0x8", + "NodeType": "0x5", + "BriefDescription": "Counts number of directed snoops sent (not including SF back invalidation).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_brd_snoops_sent", + "EventIdCode": "0x9", + "NodeType": "0x5", + "BriefDescription": "Counts number of multicast snoops sent (not including SF back invalidation).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_slc_eviction", + "EventIdCode": "0xa", + "NodeType": "0x5", + "BriefDescription": "Counts number of SLC evictions (dirty only).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_slc_fill_invalid_way", + "EventIdCode": "0xb", + "NodeType": "0x5", + "BriefDescription": "Counts number of SLC fills to an invalid way.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_mc_retries", + "EventIdCode": "0xc", + "NodeType": "0x5", + "BriefDescription": "Counts number of retried transactions by the MC.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_mc_reqs", + "EventIdCode": "0xd", + "NodeType": "0x5", + "BriefDescription": "Counts number of requests that are sent to MC.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_qos_hh_retry", + "EventIdCode": "0xe", + "NodeType": "0x5", + "BriefDescription": "Counts number of times a HighHigh priority request is protocolretried at the HN‑F.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_s0_rdata_beats", + "EventIdCode": "0x1", + "NodeType": "0xa", + "BriefDescription": "Number of RData beats (RVALID and RREADY) dispatched on port 0. This event measures the read bandwidth, including CMO responses.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_s1_rdata_beats", + "EventIdCode": "0x2", + "NodeType": "0xa", + "BriefDescription": "Number of RData beats (RVALID and RREADY) dispatched on port 1. This event measures the read bandwidth, including CMO responses.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_s2_rdata_beats", + "EventIdCode": "0x3", + "NodeType": "0xa", + "BriefDescription": "Number of RData beats (RVALID and RREADY) dispatched on port 2. This event measures the read bandwidth, including CMO responses.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_rxdat_flits", + "EventIdCode": "0x4", + "NodeType": "0xa", + "BriefDescription": "Number of RXDAT flits received. This event measures the true read data bandwidth, excluding CMOs.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_txdat_flits", + "EventIdCode": "0x5", + "NodeType": "0xa", + "BriefDescription": "Number of TXDAT flits dispatched. This event measures the write bandwidth.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_txreq_flits_total", + "EventIdCode": "0x6", + "NodeType": "0xa", + "BriefDescription": "Number of TXREQ flits dispatched. This event measures the total request bandwidth.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_txreq_flits_retried", + "EventIdCode": "0x7", + "NodeType": "0xa", + "BriefDescription": "Number of retried TXREQ flits dispatched. This event measures the retry rate.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_txrsp_retryack", + "EventIdCode": "0x4", + "NodeType": "0x7", + "BriefDescription": "Number of RXREQ flits dispatched. This event is a measure of the retry rate.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_txdat_flitv", + "EventIdCode": "0x5", + "NodeType": "0x7", + "BriefDescription": "Number of TXDAT flits dispatched from XP to SBSX. This event is a measure of the write bandwidth.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_arvalid_no_arready", + "EventIdCode": "0x21", + "NodeType": "0x7", + "BriefDescription": "Number of cycles the SBSX bridge is stalled because of backpressure on AR channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_awvalid_no_awready", + "EventIdCode": "0x22", + "NodeType": "0x7", + "BriefDescription": "Number of cycles the SBSX bridge is stalled because of backpressure on AW channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_wvalid_no_wready", + "EventIdCode": "0x23", + "NodeType": "0x7", + "BriefDescription": "Number of cycles the SBSX bridge is stalled because of backpressure on W channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_txrsp_retryack", + "EventIdCode": "0x2a", + "NodeType": "0x4", + "BriefDescription": "Number of RXREQ flits dispatched. This event is a measure of the retry rate.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_arvalid_no_arready", + "EventIdCode": "0x2b", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the HN-I bridge is stalled because of backpressure on AR channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_arready_no_arvalid", + "EventIdCode": "0x2c", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the AR channel is waiting for new requests from HN-I bridge.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_awvalid_no_awready", + "EventIdCode": "0x2d", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the HN-I bridge is stalled because of backpressure on AW channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_awready_no_awvalid", + "EventIdCode": "0x2e", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the AW channel is waiting for new requests from HN-I bridge.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_wvalid_no_wready", + "EventIdCode": "0x2f", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the HN-I bridge is stalled because of backpressure on W channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_txdat_stall", + "EventIdCode": "0x30", + "NodeType": "0x4", + "BriefDescription": "TXDAT valid but no link credit available.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + } +]
tools/perf/pmu-events/jevents.py +1 -0
@@ -272,6 +272,7 @@ class JsonEvent: 'DFPMC': 'amd_df', 'cpu_core': 'cpu_core', 'cpu_atom': 'cpu_atom', + 'arm_cmn': 'arm_cmn', } return table[unit] if unit in table else f'uncore_{unit.lower()}'
Signature · 1 line
1.8.3.1
Robin Murphy
maintainerRe: [PATCH v7 7/8] perf jevents: Add support for Arm CMN PMU aliasing
Quoted 27 lines · Jing Zhang
Currently just add aliases for part of Arm CMN PMU events which are general and compatible for any SoC and CMN-ANY.
"Compat" value "434*;436*;43c*;43a*" means it is compatible with all CMN600/CMN650/CMN700/Ci700, which can be obtained from commit
7819e05a0dce("perf/arm-cmn: Revamp model detection").The arm-cmn PMU events got from:
[0] https://developer.arm.com/documentation/100180/0302/?lang=en
[1] https://developer.arm.com/documentation/101408/0100/?lang=en
[2] https://developer.arm.com/documentation/102308/0302/?lang=en
[3] https://developer.arm.com/documentation/101569/0300/?lang=enSigned-off-by Jing Zhang <[email protected]>Reviewed-by John Garry <[email protected]>--- .../pmu-events/arch/arm64/arm/cmn/sys/cmn.json | 266 +++++++++++++++++++++ tools/perf/pmu-events/jevents.py | 1 + 2 files changed, 267 insertions(+)create mode 100644
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.jsontools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.json new file +0
-0
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.json line 5
Quoted 5 lines
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.json new file +5
-0@@ -0,0 +1,266 @@ +[ + { + "EventName": "hnf_cache_miss", + "EventIdCode": "0x1", + "NodeType": "0x5",
Given my other comment, I also think there would be no harm in just having these as:
"ConfigCode" : "0x10005"
if you'd rather make life easier to begin with, then be able to come back and improve things later. IMO it doesn't affect the readability of the important values *all* that much, since it's not like they're tighly packed together in oddly-aligned bitfields.
Thanks,
Robin.
Quoted 266 lines
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.json new file +261
-0+ "BriefDescription": "Counts total cache misses in first lookup result (high priority).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_slc_sf_cache_access", + "EventIdCode": "0x2", + "NodeType": "0x5", + "BriefDescription": "Counts number of cache accesses in first access (high priority).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_cache_fill", + "EventIdCode": "0x3", + "NodeType": "0x5", + "BriefDescription": "Counts total allocations in HN SLC (all cache line allocations to SLC).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_pocq_retry", + "EventIdCode": "0x4", + "NodeType": "0x5", + "BriefDescription": "Counts number of retried requests.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_pocq_reqs_recvd", + "EventIdCode": "0x5", + "NodeType": "0x5", + "BriefDescription": "Counts number of requests that HN receives.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_sf_hit", + "EventIdCode": "0x6", + "NodeType": "0x5", + "BriefDescription": "Counts number of SF hits.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_sf_evictions", + "EventIdCode": "0x7", + "NodeType": "0x5", + "BriefDescription": "Counts number of SF eviction cache invalidations initiated.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_dir_snoops_sent", + "EventIdCode": "0x8", + "NodeType": "0x5", + "BriefDescription": "Counts number of directed snoops sent (not including SF back invalidation).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_brd_snoops_sent", + "EventIdCode": "0x9", + "NodeType": "0x5", + "BriefDescription": "Counts number of multicast snoops sent (not including SF back invalidation).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_slc_eviction", + "EventIdCode": "0xa", + "NodeType": "0x5", + "BriefDescription": "Counts number of SLC evictions (dirty only).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_slc_fill_invalid_way", + "EventIdCode": "0xb", + "NodeType": "0x5", + "BriefDescription": "Counts number of SLC fills to an invalid way.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_mc_retries", + "EventIdCode": "0xc", + "NodeType": "0x5", + "BriefDescription": "Counts number of retried transactions by the MC.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_mc_reqs", + "EventIdCode": "0xd", + "NodeType": "0x5", + "BriefDescription": "Counts number of requests that are sent to MC.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_qos_hh_retry", + "EventIdCode": "0xe", + "NodeType": "0x5", + "BriefDescription": "Counts number of times a HighHigh priority request is protocolretried at the HN‑F.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_s0_rdata_beats", + "EventIdCode": "0x1", + "NodeType": "0xa", + "BriefDescription": "Number of RData beats (RVALID and RREADY) dispatched on port 0. This event measures the read bandwidth, including CMO responses.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_s1_rdata_beats", + "EventIdCode": "0x2", + "NodeType": "0xa", + "BriefDescription": "Number of RData beats (RVALID and RREADY) dispatched on port 1. This event measures the read bandwidth, including CMO responses.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_s2_rdata_beats", + "EventIdCode": "0x3", + "NodeType": "0xa", + "BriefDescription": "Number of RData beats (RVALID and RREADY) dispatched on port 2. This event measures the read bandwidth, including CMO responses.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_rxdat_flits", + "EventIdCode": "0x4", + "NodeType": "0xa", + "BriefDescription": "Number of RXDAT flits received. This event measures the true read data bandwidth, excluding CMOs.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_txdat_flits", + "EventIdCode": "0x5", + "NodeType": "0xa", + "BriefDescription": "Number of TXDAT flits dispatched. This event measures the write bandwidth.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_txreq_flits_total", + "EventIdCode": "0x6", + "NodeType": "0xa", + "BriefDescription": "Number of TXREQ flits dispatched. This event measures the total request bandwidth.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_txreq_flits_retried", + "EventIdCode": "0x7", + "NodeType": "0xa", + "BriefDescription": "Number of retried TXREQ flits dispatched. This event measures the retry rate.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_txrsp_retryack", + "EventIdCode": "0x4", + "NodeType": "0x7", + "BriefDescription": "Number of RXREQ flits dispatched. This event is a measure of the retry rate.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_txdat_flitv", + "EventIdCode": "0x5", + "NodeType": "0x7", + "BriefDescription": "Number of TXDAT flits dispatched from XP to SBSX. This event is a measure of the write bandwidth.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_arvalid_no_arready", + "EventIdCode": "0x21", + "NodeType": "0x7", + "BriefDescription": "Number of cycles the SBSX bridge is stalled because of backpressure on AR channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_awvalid_no_awready", + "EventIdCode": "0x22", + "NodeType": "0x7", + "BriefDescription": "Number of cycles the SBSX bridge is stalled because of backpressure on AW channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_wvalid_no_wready", + "EventIdCode": "0x23", + "NodeType": "0x7", + "BriefDescription": "Number of cycles the SBSX bridge is stalled because of backpressure on W channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_txrsp_retryack", + "EventIdCode": "0x2a", + "NodeType": "0x4", + "BriefDescription": "Number of RXREQ flits dispatched. This event is a measure of the retry rate.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_arvalid_no_arready", + "EventIdCode": "0x2b", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the HN-I bridge is stalled because of backpressure on AR channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_arready_no_arvalid", + "EventIdCode": "0x2c", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the AR channel is waiting for new requests from HN-I bridge.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_awvalid_no_awready", + "EventIdCode": "0x2d", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the HN-I bridge is stalled because of backpressure on AW channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_awready_no_awvalid", + "EventIdCode": "0x2e", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the AW channel is waiting for new requests from HN-I bridge.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_wvalid_no_wready", + "EventIdCode": "0x2f", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the HN-I bridge is stalled because of backpressure on W channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_txdat_stall", + "EventIdCode": "0x30", + "NodeType": "0x4", + "BriefDescription": "TXDAT valid but no link credit available.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + } +]tools/perf/pmu-events/jevents.py +0
-0
tools/perf/pmu-events/jevents.py class JsonEvent: · line 278
Quoted 7 lines
tools/perf/pmu-events/jevents.py +1
-0@@ -272,6 +272,7 @@ class JsonEvent: 'DFPMC': 'amd_df', 'cpu_core': 'cpu_core', 'cpu_atom': 'cpu_atom', + 'arm_cmn': 'arm_cmn', } return table[unit] if unit in table else f'uncore_{unit.lower()}'
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Re: [PATCH v7 7/8] perf jevents: Add support for Arm CMN PMU aliasing
Quoted 38 lines · Robin Murphy
Quoted 27 lines · Jing Zhang
Currently just add aliases for part of Arm CMN PMU events which are general and compatible for any SoC and CMN-ANY.
"Compat" value "434*;436*;43c*;43a*" means it is compatible with all CMN600/CMN650/CMN700/Ci700, which can be obtained from commit
7819e05a0dce("perf/arm-cmn: Revamp model detection").The arm-cmn PMU events got from:
[0] https://developer.arm.com/documentation/100180/0302/?lang=en
[1] https://developer.arm.com/documentation/101408/0100/?lang=en
[2] https://developer.arm.com/documentation/102308/0302/?lang=en
[3] https://developer.arm.com/documentation/101569/0300/?lang=enSigned-off-by Jing Zhang <[email protected]>Reviewed-by John Garry <[email protected]>---
.../pmu-events/arch/arm64/arm/cmn/sys/cmn.json | 266 +++++++++++++++++++++ tools/perf/pmu-events/jevents.py | 1 +2 files changed, 267 insertions(+)create mode 100644
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.jsontools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.json new file +0
-0tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.json line 5
Quoted 5 lines
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.json new file +5
-0@@ -0,0 +1,266 @@ +[ + { + "EventName": "hnf_cache_miss", + "EventIdCode": "0x1", + "NodeType": "0x5",Given my other comment, I also think there would be no harm in just having these as:
"ConfigCode" : "0x10005"
Quoted 2 lines
if you'd rather make life easier to begin with, then be able to come back and improve things later. IMO it doesn't affect the readability of the important values *all* that much, since it's not like they're tighly packed together in oddly-aligned bitfields.
Thanks a lot! that's a great suggestion. I hope to merge it in v6.6 first in the current way, and then I will improve it in CMN driver and perf tools later.
Thanks,
Jing
Quoted 268 lines
Thanks,
Robin.Quoted 265 lines
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.json new file +261
-0+ "BriefDescription": "Counts total cache misses in first lookup result (high priority).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_slc_sf_cache_access", + "EventIdCode": "0x2", + "NodeType": "0x5", + "BriefDescription": "Counts number of cache accesses in first access (high priority).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_cache_fill", + "EventIdCode": "0x3", + "NodeType": "0x5", + "BriefDescription": "Counts total allocations in HN SLC (all cache line allocations to SLC).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_pocq_retry", + "EventIdCode": "0x4", + "NodeType": "0x5", + "BriefDescription": "Counts number of retried requests.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_pocq_reqs_recvd", + "EventIdCode": "0x5", + "NodeType": "0x5", + "BriefDescription": "Counts number of requests that HN receives.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_sf_hit", + "EventIdCode": "0x6", + "NodeType": "0x5", + "BriefDescription": "Counts number of SF hits.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_sf_evictions", + "EventIdCode": "0x7", + "NodeType": "0x5", + "BriefDescription": "Counts number of SF eviction cache invalidations initiated.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_dir_snoops_sent", + "EventIdCode": "0x8", + "NodeType": "0x5", + "BriefDescription": "Counts number of directed snoops sent (not including SF back invalidation).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_brd_snoops_sent", + "EventIdCode": "0x9", + "NodeType": "0x5", + "BriefDescription": "Counts number of multicast snoops sent (not including SF back invalidation).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_slc_eviction", + "EventIdCode": "0xa", + "NodeType": "0x5", + "BriefDescription": "Counts number of SLC evictions (dirty only).", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_slc_fill_invalid_way", + "EventIdCode": "0xb", + "NodeType": "0x5", + "BriefDescription": "Counts number of SLC fills to an invalid way.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_mc_retries", + "EventIdCode": "0xc", + "NodeType": "0x5", + "BriefDescription": "Counts number of retried transactions by the MC.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_mc_reqs", + "EventIdCode": "0xd", + "NodeType": "0x5", + "BriefDescription": "Counts number of requests that are sent to MC.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hnf_qos_hh_retry", + "EventIdCode": "0xe", + "NodeType": "0x5", + "BriefDescription": "Counts number of times a HighHigh priority request is protocolretried at the HN‑F.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_s0_rdata_beats", + "EventIdCode": "0x1", + "NodeType": "0xa", + "BriefDescription": "Number of RData beats (RVALID and RREADY) dispatched on port 0. This event measures the read bandwidth, including CMO responses.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_s1_rdata_beats", + "EventIdCode": "0x2", + "NodeType": "0xa", + "BriefDescription": "Number of RData beats (RVALID and RREADY) dispatched on port 1. This event measures the read bandwidth, including CMO responses.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_s2_rdata_beats", + "EventIdCode": "0x3", + "NodeType": "0xa", + "BriefDescription": "Number of RData beats (RVALID and RREADY) dispatched on port 2. This event measures the read bandwidth, including CMO responses.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_rxdat_flits", + "EventIdCode": "0x4", + "NodeType": "0xa", + "BriefDescription": "Number of RXDAT flits received. This event measures the true read data bandwidth, excluding CMOs.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_txdat_flits", + "EventIdCode": "0x5", + "NodeType": "0xa", + "BriefDescription": "Number of TXDAT flits dispatched. This event measures the write bandwidth.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_txreq_flits_total", + "EventIdCode": "0x6", + "NodeType": "0xa", + "BriefDescription": "Number of TXREQ flits dispatched. This event measures the total request bandwidth.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "rnid_txreq_flits_retried", + "EventIdCode": "0x7", + "NodeType": "0xa", + "BriefDescription": "Number of retried TXREQ flits dispatched. This event measures the retry rate.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_txrsp_retryack", + "EventIdCode": "0x4", + "NodeType": "0x7", + "BriefDescription": "Number of RXREQ flits dispatched. This event is a measure of the retry rate.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_txdat_flitv", + "EventIdCode": "0x5", + "NodeType": "0x7", + "BriefDescription": "Number of TXDAT flits dispatched from XP to SBSX. This event is a measure of the write bandwidth.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_arvalid_no_arready", + "EventIdCode": "0x21", + "NodeType": "0x7", + "BriefDescription": "Number of cycles the SBSX bridge is stalled because of backpressure on AR channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_awvalid_no_awready", + "EventIdCode": "0x22", + "NodeType": "0x7", + "BriefDescription": "Number of cycles the SBSX bridge is stalled because of backpressure on AW channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "sbsx_wvalid_no_wready", + "EventIdCode": "0x23", + "NodeType": "0x7", + "BriefDescription": "Number of cycles the SBSX bridge is stalled because of backpressure on W channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_txrsp_retryack", + "EventIdCode": "0x2a", + "NodeType": "0x4", + "BriefDescription": "Number of RXREQ flits dispatched. This event is a measure of the retry rate.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_arvalid_no_arready", + "EventIdCode": "0x2b", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the HN-I bridge is stalled because of backpressure on AR channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_arready_no_arvalid", + "EventIdCode": "0x2c", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the AR channel is waiting for new requests from HN-I bridge.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_awvalid_no_awready", + "EventIdCode": "0x2d", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the HN-I bridge is stalled because of backpressure on AW channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_awready_no_awvalid", + "EventIdCode": "0x2e", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the AW channel is waiting for new requests from HN-I bridge.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_wvalid_no_wready", + "EventIdCode": "0x2f", + "NodeType": "0x4", + "BriefDescription": "Number of cycles the HN-I bridge is stalled because of backpressure on W channel.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "EventName": "hni_txdat_stall", + "EventIdCode": "0x30", + "NodeType": "0x4", + "BriefDescription": "TXDAT valid but no link credit available.", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + } +]tools/perf/pmu-events/jevents.py +0
-0
Quoted 8 lines
Quoted 8 lines
tools/perf/pmu-events/jevents.py +1
-0@@ -272,6 +272,7 @@ class JsonEvent: 'DFPMC': 'amd_df', 'cpu_core': 'cpu_core', 'cpu_atom': 'cpu_atom', + 'arm_cmn': 'arm_cmn', }return table[unit] if unit in table else f'uncore_{unit.lower()}'
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
PATCH v7 8/8 perf vendor events: Add JSON metrics for Arm CMN
In reply to Jing ZhangAdd JSON metrics for Arm CMN. Currently just add part of CMN PMU metrics which are general and compatible for any SoC with CMN-ANY.
--- .../pmu-events/arch/arm64/arm/cmn/sys/metric.json | 74 ++++++++++++++++++++++ 1 file changed, 74 insertions(+) create mode 100644 tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.json
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.json new file +74 -0
@@ -0,0 +1,74 @@ +[ + { + "MetricName": "slc_miss_rate", + "BriefDescription": "The system level cache miss rate.", + "MetricGroup": "cmn", + "MetricExpr": "hnf_cache_miss / hnf_slc_sf_cache_access", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "hnf_message_retry_rate", + "BriefDescription": "HN-F message retry rate indicates whether a lack of credits is causing the bottlenecks.", + "MetricGroup": "cmn", + "MetricExpr": "hnf_pocq_retry / hnf_pocq_reqs_recvd", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "sf_hit_rate", + "BriefDescription": "Snoop filter hit rate can be used to measure the snoop filter efficiency.", + "MetricGroup": "cmn", + "MetricExpr": "hnf_sf_hit / hnf_slc_sf_cache_access", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "mc_message_retry_rate", + "BriefDescription": "The memory controller request retries rate indicates whether the memory controller is the bottleneck.", + "MetricGroup": "cmn", + "MetricExpr": "hnf_mc_retries / hnf_mc_reqs", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "rni_actual_read_bandwidth.all", + "BriefDescription": "This event measure the actual bandwidth that RN-I bridge sends to the interconnect.", + "MetricGroup": "cmn", + "MetricExpr": "rnid_rxdat_flits * 32 / 1e6 / duration_time", + "ScaleUnit": "1MB/s", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "rni_actual_write_bandwidth.all", + "BriefDescription": "This event measures the actual write bandwidth at RN-I bridges.", + "MetricGroup": "cmn", + "MetricExpr": "rnid_txdat_flits * 32 / 1e6 / duration_time", + "ScaleUnit": "1MB/s", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "rni_retry_rate", + "BriefDescription": "RN-I bridge retry rate indicates whether the memory controller is the bottleneck.", + "MetricGroup": "cmn", + "MetricExpr": "rnid_txreq_flits_retried / rnid_txreq_flits_total", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "sbsx_actual_write_bandwidth.all", + "BriefDescription": "sbsx actual write bandwidth.", + "MetricGroup": "cmn", + "MetricExpr": "sbsx_txdat_flitv * 32 / 1e6 / duration_time", + "ScaleUnit": "1MB/s", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + } +]
Signature · 1 line
1.8.3.1
Ian Rogers
reviewerQuoted 17 lines · Jing Zhang
Add JSON metrics for Arm CMN. Currently just add part of CMN PMU metrics which are general and compatible for any SoC with CMN-ANY.
Signed-off-by Jing Zhang <[email protected]>Reviewed-by John Garry <[email protected]>--- .../pmu-events/arch/arm64/arm/cmn/sys/metric.json | 74 ++++++++++++++++++++++ 1 file changed, 74 insertions(+) create mode 100644 tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.jsontools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.json new file +0
-0
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.json line 9
Quoted 9 lines
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.json new file +9
-0@@ -0,0 +1,74 @@ +[ + { + "MetricName": "slc_miss_rate", + "BriefDescription": "The system level cache miss rate.", + "MetricGroup": "cmn", + "MetricExpr": "hnf_cache_miss / hnf_slc_sf_cache_access", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*"
Here a ';' is used as a separator, but for "Unit" ',' is used as a separator. Is there a reason for the inconsistency?
Thanks,
Ian
Quoted 66 lines
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.json new file +65
-1+ }, + { + "MetricName": "hnf_message_retry_rate", + "BriefDescription": "HN-F message retry rate indicates whether a lack of credits is causing the bottlenecks.", + "MetricGroup": "cmn", + "MetricExpr": "hnf_pocq_retry / hnf_pocq_reqs_recvd", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "sf_hit_rate", + "BriefDescription": "Snoop filter hit rate can be used to measure the snoop filter efficiency.", + "MetricGroup": "cmn", + "MetricExpr": "hnf_sf_hit / hnf_slc_sf_cache_access", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "mc_message_retry_rate", + "BriefDescription": "The memory controller request retries rate indicates whether the memory controller is the bottleneck.", + "MetricGroup": "cmn", + "MetricExpr": "hnf_mc_retries / hnf_mc_reqs", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "rni_actual_read_bandwidth.all", + "BriefDescription": "This event measure the actual bandwidth that RN-I bridge sends to the interconnect.", + "MetricGroup": "cmn", + "MetricExpr": "rnid_rxdat_flits * 32 / 1e6 / duration_time", + "ScaleUnit": "1MB/s", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "rni_actual_write_bandwidth.all", + "BriefDescription": "This event measures the actual write bandwidth at RN-I bridges.", + "MetricGroup": "cmn", + "MetricExpr": "rnid_txdat_flits * 32 / 1e6 / duration_time", + "ScaleUnit": "1MB/s", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "rni_retry_rate", + "BriefDescription": "RN-I bridge retry rate indicates whether the memory controller is the bottleneck.", + "MetricGroup": "cmn", + "MetricExpr": "rnid_txreq_flits_retried / rnid_txreq_flits_total", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "sbsx_actual_write_bandwidth.all", + "BriefDescription": "sbsx actual write bandwidth.", + "MetricGroup": "cmn", + "MetricExpr": "sbsx_txdat_flitv * 32 / 1e6 / duration_time", + "ScaleUnit": "1MB/s", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + } +] --
Quoted 2 lines
1.8.3.1
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Quoted 22 lines · Ian Rogers
Quoted 21 lines · Jing Zhang
Add JSON metrics for Arm CMN. Currently just add part of CMN PMU metrics which are general and compatible for any SoC with CMN-ANY.
Signed-off-by Jing Zhang <[email protected]>Reviewed-by John Garry <[email protected]>--- .../pmu-events/arch/arm64/arm/cmn/sys/metric.json | 74 ++++++++++++++++++++++ 1 file changed, 74 insertions(+) create mode 100644 tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.jsontools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.json new file +4
-0@@ -0,0 +1,74 @@ +[ + { + "MetricName": "slc_miss_rate", + "BriefDescription": "The system level cache miss rate.",
Quoted 9 lines
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.json line 9
Quoted 5 lines
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.json new file +5
-0+ "MetricGroup": "cmn", + "MetricExpr": "hnf_cache_miss / hnf_slc_sf_cache_access", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*"Here a ';' is used as a separator, but for "Unit" ',' is used as a separator. Is there a reason for the inconsistency?
John and I have previously discussed this issue, and in fact, I deliberately used ';'.
I use a semicolon instead of a comma because I want to distinguish it from the function of the comma in "Unit" and avoid confusion between the use of commas in "Unit" and "Compat". Because in Unit, commas act as wildcards, and in “Compat”, the semicolon means "or". So I think semicolons are more appropriate.
Thanks,
Jing
Quoted 71 lines
Thanks,
IanQuoted 68 lines
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.json new file +65
-1+ }, + { + "MetricName": "hnf_message_retry_rate", + "BriefDescription": "HN-F message retry rate indicates whether a lack of credits is causing the bottlenecks.", + "MetricGroup": "cmn", + "MetricExpr": "hnf_pocq_retry / hnf_pocq_reqs_recvd", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "sf_hit_rate", + "BriefDescription": "Snoop filter hit rate can be used to measure the snoop filter efficiency.", + "MetricGroup": "cmn", + "MetricExpr": "hnf_sf_hit / hnf_slc_sf_cache_access", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "mc_message_retry_rate", + "BriefDescription": "The memory controller request retries rate indicates whether the memory controller is the bottleneck.", + "MetricGroup": "cmn", + "MetricExpr": "hnf_mc_retries / hnf_mc_reqs", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "rni_actual_read_bandwidth.all", + "BriefDescription": "This event measure the actual bandwidth that RN-I bridge sends to the interconnect.", + "MetricGroup": "cmn", + "MetricExpr": "rnid_rxdat_flits * 32 / 1e6 / duration_time", + "ScaleUnit": "1MB/s", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "rni_actual_write_bandwidth.all", + "BriefDescription": "This event measures the actual write bandwidth at RN-I bridges.", + "MetricGroup": "cmn", + "MetricExpr": "rnid_txdat_flits * 32 / 1e6 / duration_time", + "ScaleUnit": "1MB/s", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "rni_retry_rate", + "BriefDescription": "RN-I bridge retry rate indicates whether the memory controller is the bottleneck.", + "MetricGroup": "cmn", + "MetricExpr": "rnid_txreq_flits_retried / rnid_txreq_flits_total", + "ScaleUnit": "100%", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + }, + { + "MetricName": "sbsx_actual_write_bandwidth.all", + "BriefDescription": "sbsx actual write bandwidth.", + "MetricGroup": "cmn", + "MetricExpr": "sbsx_txdat_flitv * 32 / 1e6 / duration_time", + "ScaleUnit": "1MB/s", + "Unit": "arm_cmn", + "Compat": "434*;436*;43c*;43a*" + } +] --1.8.3.1
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
John Garry
reviewerOn 21/08/2023 09:36, Jing Zhang wrote:
I'm hoping that Ian can check the outstanding patches here, but I'll also have a look.
Quoted 95 lines
Changes since v6:
- Supplement the omitted EventCode;
- Keep the original way of ConfigCode;
- Supplement the test inempty-pmu-events.c, so that the pmu event test
can succeed when build with NO_JEVENT=1.
- Link: https://urldefense.com/v3/__https://lore.kernel.org/all/[email protected]/__;!!ACWV5N9M2RV99hQ!Lakh7Lf-6l6Hovm4Tt5S5pqV1xHm-LAW2ICVl6gLONlN4CNk-BvyADBfjwQe5yQQj89yMK7s7rSjMNHCyFsIUxnHXg$Jing Zhang (8):
perf pmu: "Compat" supports matching multiple identifiers perf metric: "Compat" supports matching multiple identifiers perf vendor events: Supplement the omitted EventCode perf jevents: Support more event fields
perf test: Makematching_pmueffective
perf test: Add pmu-event test for "Compat" and newevent_field. perf jevents: Add support for Arm CMN PMU aliasing perf vendor events: Add JSON metrics for Arm CMN.../pmu-events/arch/arm64/arm/cmn/sys/cmn.json | 266 +++++++++++++++++++++ .../pmu-events/arch/arm64/arm/cmn/sys/metric.json | 74 ++++++ .../pmu-events/arch/test/test_soc/sys/uncore.json | 8 + .../pmu-events/arch/x86/alderlake/pipeline.json | 9 + .../pmu-events/arch/x86/alderlaken/pipeline.json | 3 + .../pmu-events/arch/x86/broadwell/pipeline.json | 4 + .../pmu-events/arch/x86/broadwellde/pipeline.json | 4 + .../arch/x86/broadwellde/uncore-cache.json | 2 + .../arch/x86/broadwellde/uncore-interconnect.json | 1 + .../arch/x86/broadwellde/uncore-memory.json | 1 + .../arch/x86/broadwellde/uncore-power.json | 1 + .../pmu-events/arch/x86/broadwellx/pipeline.json | 4 + .../arch/x86/broadwellx/uncore-cache.json | 2 + .../arch/x86/broadwellx/uncore-interconnect.json | 13 + .../arch/x86/broadwellx/uncore-memory.json | 2 + .../arch/x86/broadwellx/uncore-power.json | 1 + .../pmu-events/arch/x86/cascadelakex/pipeline.json | 4 + .../arch/x86/cascadelakex/uncore-cache.json | 2 + .../arch/x86/cascadelakex/uncore-interconnect.json | 1 + .../arch/x86/cascadelakex/uncore-io.json | 1 + .../arch/x86/cascadelakex/uncore-memory.json | 1 + .../arch/x86/cascadelakex/uncore-power.json | 1 + .../pmu-events/arch/x86/elkhartlake/pipeline.json | 2 + .../pmu-events/arch/x86/goldmont/pipeline.json | 3 + .../pmu-events/arch/x86/goldmontplus/pipeline.json | 3 + .../pmu-events/arch/x86/grandridge/pipeline.json | 3 + .../arch/x86/graniterapids/pipeline.json | 4 + .../perf/pmu-events/arch/x86/haswell/pipeline.json | 4 + .../pmu-events/arch/x86/haswellx/pipeline.json | 4 + .../pmu-events/arch/x86/haswellx/uncore-cache.json | 2 + .../arch/x86/haswellx/uncore-interconnect.json | 14 ++ .../arch/x86/haswellx/uncore-memory.json | 2 + .../pmu-events/arch/x86/haswellx/uncore-power.json | 1 + .../perf/pmu-events/arch/x86/icelake/pipeline.json | 4 + .../pmu-events/arch/x86/icelakex/pipeline.json | 4 + .../pmu-events/arch/x86/icelakex/uncore-cache.json | 1 + .../arch/x86/icelakex/uncore-interconnect.json | 1 + .../arch/x86/icelakex/uncore-memory.json | 1 + .../pmu-events/arch/x86/icelakex/uncore-power.json | 1 + .../pmu-events/arch/x86/ivybridge/pipeline.json | 3 + .../perf/pmu-events/arch/x86/ivytown/pipeline.json | 4 + .../pmu-events/arch/x86/ivytown/uncore-cache.json | 2 + .../arch/x86/ivytown/uncore-interconnect.json | 11 + .../pmu-events/arch/x86/ivytown/uncore-memory.json | 1 + .../pmu-events/arch/x86/ivytown/uncore-power.json | 1 + .../pmu-events/arch/x86/jaketown/pipeline.json | 4 + .../pmu-events/arch/x86/jaketown/uncore-cache.json | 2 + .../arch/x86/jaketown/uncore-interconnect.json | 12 + .../arch/x86/jaketown/uncore-memory.json | 1 + .../pmu-events/arch/x86/jaketown/uncore-power.json | 2 + .../arch/x86/knightslanding/pipeline.json | 3 + .../arch/x86/knightslanding/uncore-cache.json | 1 + .../arch/x86/knightslanding/uncore-memory.json | 4 + .../pmu-events/arch/x86/meteorlake/pipeline.json | 8 + .../pmu-events/arch/x86/sandybridge/pipeline.json | 4 + .../arch/x86/sapphirerapids/pipeline.json | 5 + .../pmu-events/arch/x86/sierraforest/pipeline.json | 4 + .../pmu-events/arch/x86/silvermont/pipeline.json | 3 + .../perf/pmu-events/arch/x86/skylake/pipeline.json | 4 + .../pmu-events/arch/x86/skylakex/pipeline.json | 4 + .../pmu-events/arch/x86/skylakex/uncore-cache.json | 2 + .../arch/x86/skylakex/uncore-interconnect.json | 1 + .../pmu-events/arch/x86/skylakex/uncore-io.json | 1 + .../arch/x86/skylakex/uncore-memory.json | 1 + .../pmu-events/arch/x86/skylakex/uncore-power.json | 1 + .../pmu-events/arch/x86/snowridgex/pipeline.json | 2 + .../arch/x86/snowridgex/uncore-cache.json | 1 + .../arch/x86/snowridgex/uncore-interconnect.json | 1 + .../arch/x86/snowridgex/uncore-memory.json | 1 + .../arch/x86/snowridgex/uncore-power.json | 1 + .../pmu-events/arch/x86/tigerlake/pipeline.json | 5 + tools/perf/pmu-events/empty-pmu-events.c | 8 + tools/perf/pmu-events/jevents.py | 21 +- tools/perf/tests/pmu-events.c | 64 ++++- tools/perf/util/metricgroup.c | 2 +- tools/perf/util/pmu.c | 33 ++- tools/perf/util/pmu.h | 1 + 77 files changed, 679 insertions(+), 9 deletions(-)
Quoted 3 lines
create mode 100644
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.jsoncreate mode 100644tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.json
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Quoted 4 lines · John Garry
On 21/08/2023 09:36, Jing Zhang wrote:
I'm hoping that Ian can check the outstanding patches here, but I'll also have a look.
Thanks! I haven't added your tag to patch 6 for now, because I made a modification and added the code in empty-pmu-events.c. Looking forward to your review.
Thanks,
Jing
Quoted 90 lines
Quoted 90 lines
Changes since v6:
- Supplement the omitted EventCode;
- Keep the original way of ConfigCode;
- Supplement the test inempty-pmu-events.c, so that the pmu event test
can succeed when build with NO_JEVENT=1.
- Link: https://urldefense.com/v3/__https://lore.kernel.org/all/[email protected]/__;!!ACWV5N9M2RV99hQ!Lakh7Lf-6l6Hovm4Tt5S5pqV1xHm-LAW2ICVl6gLONlN4CNk-BvyADBfjwQe5yQQj89yMK7s7rSjMNHCyFsIUxnHXg$Jing Zhang (8):
perf pmu: "Compat" supports matching multiple identifiers perf metric: "Compat" supports matching multiple identifiers perf vendor events: Supplement the omitted EventCode perf jevents: Support more event fields
perf test: Makematching_pmueffective
perf test: Add pmu-event test for "Compat" and newevent_field. perf jevents: Add support for Arm CMN PMU aliasing perf vendor events: Add JSON metrics for Arm CMN.../pmu-events/arch/arm64/arm/cmn/sys/cmn.json | 266 +++++++++++++++++++++ .../pmu-events/arch/arm64/arm/cmn/sys/metric.json | 74 ++++++ .../pmu-events/arch/test/test_soc/sys/uncore.json | 8 + .../pmu-events/arch/x86/alderlake/pipeline.json | 9 + .../pmu-events/arch/x86/alderlaken/pipeline.json | 3 + .../pmu-events/arch/x86/broadwell/pipeline.json | 4 + .../pmu-events/arch/x86/broadwellde/pipeline.json | 4 + .../arch/x86/broadwellde/uncore-cache.json | 2 + .../arch/x86/broadwellde/uncore-interconnect.json | 1 + .../arch/x86/broadwellde/uncore-memory.json | 1 + .../arch/x86/broadwellde/uncore-power.json | 1 + .../pmu-events/arch/x86/broadwellx/pipeline.json | 4 + .../arch/x86/broadwellx/uncore-cache.json | 2 + .../arch/x86/broadwellx/uncore-interconnect.json | 13 + .../arch/x86/broadwellx/uncore-memory.json | 2 + .../arch/x86/broadwellx/uncore-power.json | 1 +
.../pmu-events/arch/x86/cascadelakex/pipeline.json| 4 + .../arch/x86/cascadelakex/uncore-cache.json | 2 +.../arch/x86/cascadelakex/uncore-interconnect.json| 1 + .../arch/x86/cascadelakex/uncore-io.json | 1 + .../arch/x86/cascadelakex/uncore-memory.json | 1 + .../arch/x86/cascadelakex/uncore-power.json | 1 + .../pmu-events/arch/x86/elkhartlake/pipeline.json | 2 + .../pmu-events/arch/x86/goldmont/pipeline.json | 3 +.../pmu-events/arch/x86/goldmontplus/pipeline.json| 3 + .../pmu-events/arch/x86/grandridge/pipeline.json | 3 + .../arch/x86/graniterapids/pipeline.json | 4 +.../perf/pmu-events/arch/x86/haswell/pipeline.json| 4 + .../pmu-events/arch/x86/haswellx/pipeline.json | 4 +.../pmu-events/arch/x86/haswellx/uncore-cache.json| 2 + .../arch/x86/haswellx/uncore-interconnect.json | 14 ++ .../arch/x86/haswellx/uncore-memory.json | 2 +.../pmu-events/arch/x86/haswellx/uncore-power.json| 1 +.../perf/pmu-events/arch/x86/icelake/pipeline.json| 4 + .../pmu-events/arch/x86/icelakex/pipeline.json | 4 +.../pmu-events/arch/x86/icelakex/uncore-cache.json| 1 + .../arch/x86/icelakex/uncore-interconnect.json | 1 + .../arch/x86/icelakex/uncore-memory.json | 1 +.../pmu-events/arch/x86/icelakex/uncore-power.json| 1 + .../pmu-events/arch/x86/ivybridge/pipeline.json | 3 +.../perf/pmu-events/arch/x86/ivytown/pipeline.json| 4 + .../pmu-events/arch/x86/ivytown/uncore-cache.json | 2 + .../arch/x86/ivytown/uncore-interconnect.json | 11 +.../pmu-events/arch/x86/ivytown/uncore-memory.json| 1 + .../pmu-events/arch/x86/ivytown/uncore-power.json | 1 + .../pmu-events/arch/x86/jaketown/pipeline.json | 4 +.../pmu-events/arch/x86/jaketown/uncore-cache.json| 2 + .../arch/x86/jaketown/uncore-interconnect.json | 12 + .../arch/x86/jaketown/uncore-memory.json | 1 +.../pmu-events/arch/x86/jaketown/uncore-power.json| 2 + .../arch/x86/knightslanding/pipeline.json | 3 + .../arch/x86/knightslanding/uncore-cache.json | 1 + .../arch/x86/knightslanding/uncore-memory.json | 4 + .../pmu-events/arch/x86/meteorlake/pipeline.json | 8 + .../pmu-events/arch/x86/sandybridge/pipeline.json | 4 + .../arch/x86/sapphirerapids/pipeline.json | 5 +.../pmu-events/arch/x86/sierraforest/pipeline.json| 4 + .../pmu-events/arch/x86/silvermont/pipeline.json | 3 +.../perf/pmu-events/arch/x86/skylake/pipeline.json| 4 + .../pmu-events/arch/x86/skylakex/pipeline.json | 4 +.../pmu-events/arch/x86/skylakex/uncore-cache.json| 2 + .../arch/x86/skylakex/uncore-interconnect.json | 1 + .../pmu-events/arch/x86/skylakex/uncore-io.json | 1 + .../arch/x86/skylakex/uncore-memory.json | 1 +.../pmu-events/arch/x86/skylakex/uncore-power.json| 1 + .../pmu-events/arch/x86/snowridgex/pipeline.json | 2 + .../arch/x86/snowridgex/uncore-cache.json | 1 + .../arch/x86/snowridgex/uncore-interconnect.json | 1 + .../arch/x86/snowridgex/uncore-memory.json | 1 + .../arch/x86/snowridgex/uncore-power.json | 1 + .../pmu-events/arch/x86/tigerlake/pipeline.json | 5 + tools/perf/pmu-events/empty-pmu-events.c | 8 + tools/perf/pmu-events/jevents.py | 21 +-
Quoted 8 lines
Quoted 8 lines
tools/perf/tests/pmu-events.c | 64 ++++- tools/perf/util/metricgroup.c | 2 +- tools/perf/util/pmu.c | 33 ++- tools/perf/util/pmu.h | 1 +
77 files changed, 679 insertions(+), 9 deletions(-)create mode 100644
tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.jsoncreate mode 100644tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.json
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Arnaldo Carvalho de Melo
maintainerEm Mon, Aug 21, 2023 at 04:36:09PM +0800, Jing Zhang escreveu:
Quoted 6 lines
Changes since v6:
- Supplement the omitted EventCode;
- Keep the original way of ConfigCode;
- Supplement the test inempty-pmu-events.c, so that the pmu event test
can succeed when build with NO_JEVENT=1.
- Link: https://lore.kernel.org/all/[email protected]/
Waiting for a v8, from what I saw there are several more review comments to address, right?
- Arnaldo
Quoted 94 lines
Jing Zhang (8):
perf pmu: "Compat" supports matching multiple identifiers perf metric: "Compat" supports matching multiple identifiers perf vendor events: Supplement the omitted EventCode perf jevents: Support more event fields
perf test: Makematching_pmueffective
perf test: Add pmu-event test for "Compat" and newevent_field. perf jevents: Add support for Arm CMN PMU aliasing perf vendor events: Add JSON metrics for Arm CMN.../pmu-events/arch/arm64/arm/cmn/sys/cmn.json | 266 +++++++++++++++++++++ .../pmu-events/arch/arm64/arm/cmn/sys/metric.json | 74 ++++++ .../pmu-events/arch/test/test_soc/sys/uncore.json | 8 + .../pmu-events/arch/x86/alderlake/pipeline.json | 9 + .../pmu-events/arch/x86/alderlaken/pipeline.json | 3 + .../pmu-events/arch/x86/broadwell/pipeline.json | 4 + .../pmu-events/arch/x86/broadwellde/pipeline.json | 4 + .../arch/x86/broadwellde/uncore-cache.json | 2 + .../arch/x86/broadwellde/uncore-interconnect.json | 1 + .../arch/x86/broadwellde/uncore-memory.json | 1 + .../arch/x86/broadwellde/uncore-power.json | 1 + .../pmu-events/arch/x86/broadwellx/pipeline.json | 4 + .../arch/x86/broadwellx/uncore-cache.json | 2 + .../arch/x86/broadwellx/uncore-interconnect.json | 13 + .../arch/x86/broadwellx/uncore-memory.json | 2 + .../arch/x86/broadwellx/uncore-power.json | 1 + .../pmu-events/arch/x86/cascadelakex/pipeline.json | 4 + .../arch/x86/cascadelakex/uncore-cache.json | 2 + .../arch/x86/cascadelakex/uncore-interconnect.json | 1 + .../arch/x86/cascadelakex/uncore-io.json | 1 + .../arch/x86/cascadelakex/uncore-memory.json | 1 + .../arch/x86/cascadelakex/uncore-power.json | 1 + .../pmu-events/arch/x86/elkhartlake/pipeline.json | 2 + .../pmu-events/arch/x86/goldmont/pipeline.json | 3 + .../pmu-events/arch/x86/goldmontplus/pipeline.json | 3 + .../pmu-events/arch/x86/grandridge/pipeline.json | 3 + .../arch/x86/graniterapids/pipeline.json | 4 + .../perf/pmu-events/arch/x86/haswell/pipeline.json | 4 + .../pmu-events/arch/x86/haswellx/pipeline.json | 4 + .../pmu-events/arch/x86/haswellx/uncore-cache.json | 2 + .../arch/x86/haswellx/uncore-interconnect.json | 14 ++ .../arch/x86/haswellx/uncore-memory.json | 2 + .../pmu-events/arch/x86/haswellx/uncore-power.json | 1 + .../perf/pmu-events/arch/x86/icelake/pipeline.json | 4 + .../pmu-events/arch/x86/icelakex/pipeline.json | 4 + .../pmu-events/arch/x86/icelakex/uncore-cache.json | 1 + .../arch/x86/icelakex/uncore-interconnect.json | 1 + .../arch/x86/icelakex/uncore-memory.json | 1 + .../pmu-events/arch/x86/icelakex/uncore-power.json | 1 + .../pmu-events/arch/x86/ivybridge/pipeline.json | 3 + .../perf/pmu-events/arch/x86/ivytown/pipeline.json | 4 + .../pmu-events/arch/x86/ivytown/uncore-cache.json | 2 + .../arch/x86/ivytown/uncore-interconnect.json | 11 + .../pmu-events/arch/x86/ivytown/uncore-memory.json | 1 + .../pmu-events/arch/x86/ivytown/uncore-power.json | 1 + .../pmu-events/arch/x86/jaketown/pipeline.json | 4 + .../pmu-events/arch/x86/jaketown/uncore-cache.json | 2 + .../arch/x86/jaketown/uncore-interconnect.json | 12 + .../arch/x86/jaketown/uncore-memory.json | 1 + .../pmu-events/arch/x86/jaketown/uncore-power.json | 2 + .../arch/x86/knightslanding/pipeline.json | 3 + .../arch/x86/knightslanding/uncore-cache.json | 1 + .../arch/x86/knightslanding/uncore-memory.json | 4 + .../pmu-events/arch/x86/meteorlake/pipeline.json | 8 + .../pmu-events/arch/x86/sandybridge/pipeline.json | 4 + .../arch/x86/sapphirerapids/pipeline.json | 5 + .../pmu-events/arch/x86/sierraforest/pipeline.json | 4 + .../pmu-events/arch/x86/silvermont/pipeline.json | 3 + .../perf/pmu-events/arch/x86/skylake/pipeline.json | 4 + .../pmu-events/arch/x86/skylakex/pipeline.json | 4 + .../pmu-events/arch/x86/skylakex/uncore-cache.json | 2 + .../arch/x86/skylakex/uncore-interconnect.json | 1 + .../pmu-events/arch/x86/skylakex/uncore-io.json | 1 + .../arch/x86/skylakex/uncore-memory.json | 1 + .../pmu-events/arch/x86/skylakex/uncore-power.json | 1 + .../pmu-events/arch/x86/snowridgex/pipeline.json | 2 + .../arch/x86/snowridgex/uncore-cache.json | 1 + .../arch/x86/snowridgex/uncore-interconnect.json | 1 + .../arch/x86/snowridgex/uncore-memory.json | 1 + .../arch/x86/snowridgex/uncore-power.json | 1 + .../pmu-events/arch/x86/tigerlake/pipeline.json | 5 + tools/perf/pmu-events/empty-pmu-events.c | 8 + tools/perf/pmu-events/jevents.py | 21 +- tools/perf/tests/pmu-events.c | 64 ++++- tools/perf/util/metricgroup.c | 2 +- tools/perf/util/pmu.c | 33 ++- tools/perf/util/pmu.h | 1 + 77 files changed, 679 insertions(+), 9 deletions(-) create mode 100644 tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.json create mode 100644 tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.jsonSignature · 1 line
1.8.3.1
Signature · 6 lines
- Arnaldo _______________________________________________ linux-arm-kernel mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Quoted 11 lines · Arnaldo Carvalho de Melo
Em Mon, Aug 21, 2023 at 04:36:09PM +0800, Jing Zhang escreveu:
Quoted 6 lines
Changes since v6:
- Supplement the omitted EventCode;
- Keep the original way of ConfigCode;
- Supplement the test inempty-pmu-events.c, so that the pmu event test
can succeed when build with NO_JEVENT=1.
- Link: https://lore.kernel.org/all/[email protected]/Waiting for a v8, from what I saw there are several more review comments to address, right?
Yes, I will send v8 soon.
Thanks,
Jing
Quoted 97 lines
- Arnaldo
Quoted 94 lines
Jing Zhang (8):
perf pmu: "Compat" supports matching multiple identifiers perf metric: "Compat" supports matching multiple identifiers perf vendor events: Supplement the omitted EventCode perf jevents: Support more event fields
perf test: Makematching_pmueffective
perf test: Add pmu-event test for "Compat" and newevent_field. perf jevents: Add support for Arm CMN PMU aliasing perf vendor events: Add JSON metrics for Arm CMN.../pmu-events/arch/arm64/arm/cmn/sys/cmn.json | 266 +++++++++++++++++++++ .../pmu-events/arch/arm64/arm/cmn/sys/metric.json | 74 ++++++ .../pmu-events/arch/test/test_soc/sys/uncore.json | 8 + .../pmu-events/arch/x86/alderlake/pipeline.json | 9 + .../pmu-events/arch/x86/alderlaken/pipeline.json | 3 + .../pmu-events/arch/x86/broadwell/pipeline.json | 4 + .../pmu-events/arch/x86/broadwellde/pipeline.json | 4 + .../arch/x86/broadwellde/uncore-cache.json | 2 + .../arch/x86/broadwellde/uncore-interconnect.json | 1 + .../arch/x86/broadwellde/uncore-memory.json | 1 + .../arch/x86/broadwellde/uncore-power.json | 1 + .../pmu-events/arch/x86/broadwellx/pipeline.json | 4 + .../arch/x86/broadwellx/uncore-cache.json | 2 + .../arch/x86/broadwellx/uncore-interconnect.json | 13 + .../arch/x86/broadwellx/uncore-memory.json | 2 + .../arch/x86/broadwellx/uncore-power.json | 1 + .../pmu-events/arch/x86/cascadelakex/pipeline.json | 4 + .../arch/x86/cascadelakex/uncore-cache.json | 2 + .../arch/x86/cascadelakex/uncore-interconnect.json | 1 + .../arch/x86/cascadelakex/uncore-io.json | 1 + .../arch/x86/cascadelakex/uncore-memory.json | 1 + .../arch/x86/cascadelakex/uncore-power.json | 1 + .../pmu-events/arch/x86/elkhartlake/pipeline.json | 2 + .../pmu-events/arch/x86/goldmont/pipeline.json | 3 + .../pmu-events/arch/x86/goldmontplus/pipeline.json | 3 + .../pmu-events/arch/x86/grandridge/pipeline.json | 3 + .../arch/x86/graniterapids/pipeline.json | 4 + .../perf/pmu-events/arch/x86/haswell/pipeline.json | 4 + .../pmu-events/arch/x86/haswellx/pipeline.json | 4 + .../pmu-events/arch/x86/haswellx/uncore-cache.json | 2 + .../arch/x86/haswellx/uncore-interconnect.json | 14 ++ .../arch/x86/haswellx/uncore-memory.json | 2 + .../pmu-events/arch/x86/haswellx/uncore-power.json | 1 + .../perf/pmu-events/arch/x86/icelake/pipeline.json | 4 + .../pmu-events/arch/x86/icelakex/pipeline.json | 4 + .../pmu-events/arch/x86/icelakex/uncore-cache.json | 1 + .../arch/x86/icelakex/uncore-interconnect.json | 1 + .../arch/x86/icelakex/uncore-memory.json | 1 + .../pmu-events/arch/x86/icelakex/uncore-power.json | 1 + .../pmu-events/arch/x86/ivybridge/pipeline.json | 3 + .../perf/pmu-events/arch/x86/ivytown/pipeline.json | 4 + .../pmu-events/arch/x86/ivytown/uncore-cache.json | 2 + .../arch/x86/ivytown/uncore-interconnect.json | 11 + .../pmu-events/arch/x86/ivytown/uncore-memory.json | 1 + .../pmu-events/arch/x86/ivytown/uncore-power.json | 1 + .../pmu-events/arch/x86/jaketown/pipeline.json | 4 + .../pmu-events/arch/x86/jaketown/uncore-cache.json | 2 + .../arch/x86/jaketown/uncore-interconnect.json | 12 + .../arch/x86/jaketown/uncore-memory.json | 1 + .../pmu-events/arch/x86/jaketown/uncore-power.json | 2 + .../arch/x86/knightslanding/pipeline.json | 3 + .../arch/x86/knightslanding/uncore-cache.json | 1 + .../arch/x86/knightslanding/uncore-memory.json | 4 + .../pmu-events/arch/x86/meteorlake/pipeline.json | 8 + .../pmu-events/arch/x86/sandybridge/pipeline.json | 4 + .../arch/x86/sapphirerapids/pipeline.json | 5 + .../pmu-events/arch/x86/sierraforest/pipeline.json | 4 + .../pmu-events/arch/x86/silvermont/pipeline.json | 3 + .../perf/pmu-events/arch/x86/skylake/pipeline.json | 4 + .../pmu-events/arch/x86/skylakex/pipeline.json | 4 + .../pmu-events/arch/x86/skylakex/uncore-cache.json | 2 + .../arch/x86/skylakex/uncore-interconnect.json | 1 + .../pmu-events/arch/x86/skylakex/uncore-io.json | 1 + .../arch/x86/skylakex/uncore-memory.json | 1 + .../pmu-events/arch/x86/skylakex/uncore-power.json | 1 + .../pmu-events/arch/x86/snowridgex/pipeline.json | 2 + .../arch/x86/snowridgex/uncore-cache.json | 1 + .../arch/x86/snowridgex/uncore-interconnect.json | 1 + .../arch/x86/snowridgex/uncore-memory.json | 1 + .../arch/x86/snowridgex/uncore-power.json | 1 + .../pmu-events/arch/x86/tigerlake/pipeline.json | 5 + tools/perf/pmu-events/empty-pmu-events.c | 8 + tools/perf/pmu-events/jevents.py | 21 +- tools/perf/tests/pmu-events.c | 64 ++++- tools/perf/util/metricgroup.c | 2 +- tools/perf/util/pmu.c | 33 ++- tools/perf/util/pmu.h | 1 + 77 files changed, 679 insertions(+), 9 deletions(-) create mode 100644 tools/perf/pmu-events/arch/arm64/arm/cmn/sys/cmn.json create mode 100644 tools/perf/pmu-events/arch/arm64/arm/cmn/sys/metric.jsonSignature · 1 line
1.8.3.1