Intel metric fixes and event updates
Ian Rogers
reviewercover Intel metric fixes and event updates
The metric tma_info_pipeline_retire contains uops_retired.slots with perf metric events. Patch 1 fixes this event sorting so that uops_retired.slots isn't made a group leader as that needs to be topdown.slots.
Patch 2 and 3 update the meteorlake and sapphirerapids events.
Patch 4 addresses an issue with event grouping discussed in:
https://lore.kernel.org/lkml/[email protected]/
by adding and altering metric constraints. The constraints avoid groups for metrics where the kernel PMU fails to not open the group (the trigger for the weak group being removed).
Ian Rogers (4):
perf parse-events x86: Avoid sorting uops_retired.slots perf vendor events intel: Update meteorlake to 1.04 perf vendor events intel: Update sapphirerapids to 1.15 perf vendor events intel: Update Icelake+ metric constraints
tools/perf/arch/x86/util/evlist.c | 7 +- tools/perf/arch/x86/util/evsel.c | 7 +- .../arch/x86/alderlake/adl-metrics.json | 11 +- .../arch/x86/alderlaken/adln-metrics.json | 2 + .../arch/x86/icelake/icl-metrics.json | 10 +- .../arch/x86/icelakex/icx-metrics.json | 10 +- tools/perf/pmu-events/arch/x86/mapfile.csv | 4 +- .../pmu-events/arch/x86/meteorlake/cache.json | 165 ++++++++++++++++++ .../arch/x86/meteorlake/floating-point.json | 8 + .../arch/x86/meteorlake/frontend.json | 56 ++++++ .../arch/x86/meteorlake/memory.json | 80 +++++++++ .../pmu-events/arch/x86/meteorlake/other.json | 16 ++ .../arch/x86/meteorlake/pipeline.json | 159 +++++++++++++++++ .../arch/x86/rocketlake/rkl-metrics.json | 10 +- .../arch/x86/sapphirerapids/other.json | 18 ++ .../arch/x86/sapphirerapids/spr-metrics.json | 9 +- .../arch/x86/tigerlake/tgl-metrics.json | 10 +- 17 files changed, 549 insertions(+), 33 deletions(-)
Signature · 1 line
2.41.0.585.gd2178a4bd4-goog
Ian Rogers
reviewerPATCH 1/4 perf parse-events x86: Avoid sorting uops_retired.slots
As topdown.slots may appear as slots it may get confused with uops_retired.slots which is an invalid perf metric event group leader. Special case uops_retired.slots to avoid this confusion.
--- tools/perf/arch/x86/util/evlist.c | 7 ++++--- tools/perf/arch/x86/util/evsel.c | 7 +++---- 2 files changed, 7 insertions(+), 7 deletions(-)
tools/perf/arch/x86/util/evlist.c +4 -3
@@ -75,11 +75,12 @@ int arch_evlist__add_default_attrs(struct evlist *evlist,int arch_evlist__cmp (const struct evsel *lhs,const struct evsel *rhs) { -if (topdown_sys_has_perf_metrics () &&evsel__sys_has_perf_metrics (lhs)) { +if (topdown_sys_has_perf_metrics () && + (arch_evsel__must_be_in_group (lhs) ||arch_evsel__must_be_in_group (rhs))) {/* Ensure the topdown slots comes first. */ -if (strcasestr (lhs->name,"slots" )) +if (strcasestr (lhs->name,"slots" ) && !strcasestr (lhs->name,"uops_retired.slots" ))return -1 ; -if (strcasestr (rhs->name,"slots" )) +if (strcasestr (rhs->name,"slots" ) && !strcasestr (rhs->name,"uops_retired.slots" ))return 1 ;/* Followed by topdown events. */ if (strcasestr (lhs->name,"topdown" ) && !strcasestr (rhs->name,"topdown" ))
tools/perf/arch/x86/util/evsel.c +3 -4
@@ -40,12 +40,11 @@ bool evsel__sys_has_perf_metrics(const struct evsel *evsel)bool arch_evsel__must_be_in_group (const struct evsel *evsel) { -if (!evsel__sys_has_perf_metrics (evsel)) +if (!evsel__sys_has_perf_metrics (evsel) || !evsel->name || +strcasestr (evsel->name,"uops_retired.slots" ))return false; -return evsel->name && - (strcasestr (evsel->name,"slots" ) || -strcasestr (evsel->name,"topdown" )); +return strcasestr (evsel->name,"topdown" ) ||strcasestr (evsel->name,"slots" ); }int arch_evsel__hw_name (struct evsel *evsel,char *bf,size_t size)
Signature · 1 line
2.41.0.585.gd2178a4bd4-goog
Re: [PATCH v1 1/4] perf parse-events x86: Avoid sorting uops_retired.slots
Quoted 4 lines · Ian Rogers
As topdown.slots may appear as slots it may get confused with uops_retired.slots which is an invalid perf metric event group leader. Special case uops_retired.slots to avoid this confusion.
Does any name with format "name.slots" cause the confusion? If so, I don't think we can stop others from naming like the above format.
Is it better to hard code the topdown.slots/slots, rather than uops_retired.slots?
Thanks,
Kan
Quoted 46 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/arch/x86/util/evlist.c | 7 ++++--- tools/perf/arch/x86/util/evsel.c | 7 +++---- 2 files changed, 7 insertions(+), 7 deletions(-)tools/perf/arch/x86/util/evlist.c +4
-3@@ -75,11 +75,12 @@ int arch_evlist__add_default_attrs(struct evlist *evlist,int arch_evlist__cmp (const struct evsel *lhs,const struct evsel *rhs) { -if (topdown_sys_has_perf_metrics () &&evsel__sys_has_perf_metrics (lhs)) { +if (topdown_sys_has_perf_metrics () && + (arch_evsel__must_be_in_group (lhs) ||arch_evsel__must_be_in_group (rhs))) {/* Ensure the topdown slots comes first. */ -if (strcasestr (lhs->name,"slots" )) +if (strcasestr (lhs->name,"slots" ) && !strcasestr (lhs->name,"uops_retired.slots" ))return -1 ; -if (strcasestr (rhs->name,"slots" )) +if (strcasestr (rhs->name,"slots" ) && !strcasestr (rhs->name,"uops_retired.slots" ))return 1 ;/* Followed by topdown events. */ if (strcasestr (lhs->name,"topdown" ) && !strcasestr (rhs->name,"topdown" ))tools/perf/arch/x86/util/evsel.c +3
-4@@ -40,12 +40,11 @@ bool evsel__sys_has_perf_metrics(const struct evsel *evsel)bool arch_evsel__must_be_in_group (const struct evsel *evsel) { -if (!evsel__sys_has_perf_metrics (evsel)) +if (!evsel__sys_has_perf_metrics (evsel) || !evsel->name || +strcasestr (evsel->name,"uops_retired.slots" ))return false; -return evsel->name && - (strcasestr (evsel->name,"slots" ) || -strcasestr (evsel->name,"topdown" )); +return strcasestr (evsel->name,"topdown" ) ||strcasestr (evsel->name,"slots" ); }int arch_evsel__hw_name (struct evsel *evsel,char *bf,size_t size)
Ian Rogers
reviewerRe: [PATCH v1 1/4] perf parse-events x86: Avoid sorting uops_retired.slots
Quoted 9 lines · Kan
Quoted 4 lines · Ian Rogers
As topdown.slots may appear as slots it may get confused with uops_retired.slots which is an invalid perf metric event group leader. Special case uops_retired.slots to avoid this confusion.
Quoted 5 lines
Does any name with format "name.slots" cause the confusion? If so, I don't think we can stop others from naming like the above format.
Is it better to hard code the topdown.slots/slots, rather than uops_retired.slots?
So firstly, yet more fringe perf metric event benefits with this silly bit of complexity. The issue with "just" trying to pattern match "topdown.slots" and "slots" is that there may be pmu names in there. So (ignoring case) we get:
slots
topdown.slots
cpu/slots/
cpu/topdown.slots/
cpu_core/slots/
cpu_core/topdown.slots/
but the name can have other junk like modifiers in there and also there's the name= config term, but let's just not think about that breaking stuff. To avoid 6 searches I searched for the known problematic uops_retired.slots, knowing if that's not there then one of the 6 above is the match. Searching for just "slots" or "topdown.slots" isn't good enough as "slots" will get a hit in "uops_retired.slots", how we ended up with this patch in the 1st place. So I ended up using the uops_retired.slots search to reduce complexity in the code and to avoid writing an event parser just to figure out what the name is... Ideally we'd be using the perf_event_attr to do this comparison, but immediately after parse-events when this code runs it hasn't been computed yet. Maybe I can shuffle things around and make this true. Ideally I think parse-events would just be a library that given some event description gives back perf_event_attr and not evsels, but that's yet further work...
Thanks,
Ian
Quoted 50 lines
Thanks,
KanQuoted 46 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/arch/x86/util/evlist.c | 7 ++++--- tools/perf/arch/x86/util/evsel.c | 7 +++---- 2 files changed, 7 insertions(+), 7 deletions(-)tools/perf/arch/x86/util/evlist.c +4
-3@@ -75,11 +75,12 @@ int arch_evlist__add_default_attrs(struct evlist *evlist,int arch_evlist__cmp (const struct evsel *lhs,const struct evsel *rhs) { -if (topdown_sys_has_perf_metrics () &&evsel__sys_has_perf_metrics (lhs)) { +if (topdown_sys_has_perf_metrics () && + (arch_evsel__must_be_in_group (lhs) ||arch_evsel__must_be_in_group (rhs))) {/* Ensure the topdown slots comes first. */ -if (strcasestr (lhs->name,"slots" )) +if (strcasestr (lhs->name,"slots" ) && !strcasestr (lhs->name,"uops_retired.slots" ))return -1 ; -if (strcasestr (rhs->name,"slots" )) +if (strcasestr (rhs->name,"slots" ) && !strcasestr (rhs->name,"uops_retired.slots" ))return 1 ;/* Followed by topdown events. */ if (strcasestr (lhs->name,"topdown" ) && !strcasestr (rhs->name,"topdown" ))tools/perf/arch/x86/util/evsel.c +3
-4@@ -40,12 +40,11 @@ bool evsel__sys_has_perf_metrics(const struct evsel *evsel)bool arch_evsel__must_be_in_group (const struct evsel *evsel) { -if (!evsel__sys_has_perf_metrics (evsel)) +if (!evsel__sys_has_perf_metrics (evsel) || !evsel->name || +strcasestr (evsel->name,"uops_retired.slots" ))return false; -return evsel->name && - (strcasestr (evsel->name,"slots" ) || -strcasestr (evsel->name,"topdown" )); +return strcasestr (evsel->name,"topdown" ) ||strcasestr (evsel->name,"slots" ); }int arch_evsel__hw_name (struct evsel *evsel,char *bf,size_t size)
Re: [PATCH v1 1/4] perf parse-events x86: Avoid sorting uops_retired.slots
Quoted 28 lines · Ian Rogers
Quoted 9 lines · Kan
Quoted 4 lines · Ian Rogers
As topdown.slots may appear as slots it may get confused with uops_retired.slots which is an invalid perf metric event group leader. Special case uops_retired.slots to avoid this confusion.
Quoted 5 lines
Does any name with format "name.slots" cause the confusion? If so, I don't think we can stop others from naming like the above format.
Is it better to hard code the topdown.slots/slots, rather than uops_retired.slots?
So firstly, yet more fringe perf metric event benefits with this silly bit of complexity. The issue with "just" trying to pattern match "topdown.slots" and "slots" is that there may be pmu names in there. So (ignoring case) we get:
slots
topdown.slots
cpu/slots/
cpu/topdown.slots/
cpu_core/slots/
cpu_core/topdown.slots/
Quoted 17 lines
but the name can have other junk like modifiers in there and also there's the name= config term, but let's just not think about that breaking stuff. To avoid 6 searches I searched for the known problematic uops_retired.slots, knowing if that's not there then one of the 6 above is the match. Searching for just "slots" or "topdown.slots" isn't good enough as "slots" will get a hit in "uops_retired.slots", how we ended up with this patch in the 1st place. So I ended up using the uops_retired.slots search to reduce complexity in the code and to avoid writing an event parser just to figure out what the name is... Ideally we'd be using the
perf_event_attrto do this comparison, but immediately after parse-events when this code runs it hasn't been computed yet. Maybe I can shuffle things around and make this true. Ideally I think parse-events would just be a library that given some event description gives backperf_event_attrand not evsels, but that's yet further work...
Thanks for the details.
I don't think we have other events with a similar pattern. For now, special case uops_retired.slots should be good enough.
Thanks,
Kan
Quoted 53 lines
Thanks,
IanQuoted 50 lines
Thanks,
KanQuoted 46 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/arch/x86/util/evlist.c | 7 ++++--- tools/perf/arch/x86/util/evsel.c | 7 +++---- 2 files changed, 7 insertions(+), 7 deletions(-)tools/perf/arch/x86/util/evlist.c +4
-3@@ -75,11 +75,12 @@ int arch_evlist__add_default_attrs(struct evlist *evlist,int arch_evlist__cmp (const struct evsel *lhs,const struct evsel *rhs) { -if (topdown_sys_has_perf_metrics () &&evsel__sys_has_perf_metrics (lhs)) { +if (topdown_sys_has_perf_metrics () && + (arch_evsel__must_be_in_group (lhs) ||arch_evsel__must_be_in_group (rhs))) {/* Ensure the topdown slots comes first. */ -if (strcasestr (lhs->name,"slots" )) +if (strcasestr (lhs->name,"slots" ) && !strcasestr (lhs->name,"uops_retired.slots" ))return -1 ; -if (strcasestr (rhs->name,"slots" )) +if (strcasestr (rhs->name,"slots" ) && !strcasestr (rhs->name,"uops_retired.slots" ))return 1 ;/* Followed by topdown events. */ if (strcasestr (lhs->name,"topdown" ) && !strcasestr (rhs->name,"topdown" ))tools/perf/arch/x86/util/evsel.c +3
-4@@ -40,12 +40,11 @@ bool evsel__sys_has_perf_metrics(const struct evsel *evsel)bool arch_evsel__must_be_in_group (const struct evsel *evsel) { -if (!evsel__sys_has_perf_metrics (evsel)) +if (!evsel__sys_has_perf_metrics (evsel) || !evsel->name || +strcasestr (evsel->name,"uops_retired.slots" ))return false; -return evsel->name && - (strcasestr (evsel->name,"slots" ) || -strcasestr (evsel->name,"topdown" )); +return strcasestr (evsel->name,"topdown" ) ||strcasestr (evsel->name,"slots" ); }int arch_evsel__hw_name (struct evsel *evsel,char *bf,size_t size)
Ian Rogers
reviewerPATCH 2/4 perf vendor events intel: Update meteorlake to 1.04
In reply to Ian Rogers1.04 events were released in:
https://github.com/intel/perfmon/commit/44fe3681501f43fc515577aced8e944b187c8e51
Addition of 51 core events.
--- tools/perf/pmu-events/arch/x86/mapfile.csv | 2 +- .../pmu-events/arch/x86/meteorlake/cache.json | 165 ++++++++++++++++++ .../arch/x86/meteorlake/floating-point.json | 8 + .../arch/x86/meteorlake/frontend.json | 56 ++++++ .../arch/x86/meteorlake/memory.json | 80 +++++++++ .../pmu-events/arch/x86/meteorlake/other.json | 16 ++ .../arch/x86/meteorlake/pipeline.json | 159 +++++++++++++++++ 7 files changed, 485 insertions(+), 1 deletion(-)
tools/perf/pmu-events/arch/x86/mapfile.csv +1 -1
@@ -19,7 +19,7 @@ GenuineIntel-6-3A,v24,ivybridge,core GenuineIntel-6-3E,v23,ivytown,core GenuineIntel-6-2D,v23,jaketown,core GenuineIntel-6-(57|85),v10,knightslanding,core -GenuineIntel-6-A[AC],v1.03,meteorlake,core +GenuineIntel-6-A[AC],v1.04,meteorlake,core GenuineIntel-6-1[AEF],v3,nehalemep,core GenuineIntel-6-2E,v3,nehalemex,core GenuineIntel-6-A7,v1.01,rocketlake,core
tools/perf/pmu-events/arch/x86/meteorlake/cache.json +165 -0
207 lines not on this page · raw message
tools/perf/pmu-events/arch/x86/meteorlake/floating-point.json +8 -0
@@ -41,6 +41,14 @@ "UMask": "0x2", "Unit": "cpu_core" }, + { + "BriefDescription": "FP_ARITH_DISPATCHED.PORT_5", + "EventCode": "0xb3", + "EventName": "FP_ARITH_DISPATCHED.PORT_5", + "SampleAfterValue": "2000003", + "UMask": "0x4", + "Unit": "cpu_core" + }, { "BriefDescription": "Counts number of SSE/AVX computational 128-bit packed double precision floating-point instructions retired; some instructions will count twice as noted below. Each count represents 2 computation operations, one for each element. Applies to SSE* and AVX* packed double precision floating-point instructions: ADD SUB HADD HSUB SUBADD MUL DIV MIN MAX SQRT DPP FM(N)ADD/SUB. DPP and FM(N)ADD/SUB instructions count twice as they perform 2 calculations per element.", "EventCode": "0xc7",
tools/perf/pmu-events/arch/x86/meteorlake/frontend.json +56 -0
80 lines not on this page · raw message
tools/perf/pmu-events/arch/x86/meteorlake/memory.json +80 -0
103 lines not on this page · raw message
tools/perf/pmu-events/arch/x86/meteorlake/other.json +16 -0
@@ -1,4 +1,12 @@ [ + { + "BriefDescription": "ASSISTS.PAGE_FAULT", + "EventCode": "0xc1", + "EventName": "ASSISTS.PAGE_FAULT", + "SampleAfterValue": "1000003", + "UMask": "0x8", + "Unit": "cpu_core" + }, { "BriefDescription": "Counts streaming stores that have any type of response.", "EventCode": "0x2A,0x2B",
@@ -30,6 +38,14 @@ "UMask": "0x7", "Unit": "cpu_core" }, + { + "BriefDescription": "RS.EMPTY_RESOURCE", + "EventCode": "0xa5", + "EventName": "RS.EMPTY_RESOURCE", + "SampleAfterValue": "1000003", + "UMask": "0x1", + "Unit": "cpu_core" + }, { "BriefDescription": "Counts the number of issue slots in a UMWAIT or TPAUSE instruction where no uop issues due to the instruction putting the CPU into the C0.1 activity state. For Tremont, UMWAIT and TPAUSE will only put the CPU into C0.1 activity state (not C0.2 activity state)", "EventCode": "0x75",
tools/perf/pmu-events/arch/x86/meteorlake/pipeline.json +159 -0
237 lines not on this page · raw message
Signature · 1 line
2.41.0.585.gd2178a4bd4-goog
Ian Rogers
reviewerPATCH 3/4 perf vendor events intel: Update sapphirerapids to 1.15
In reply to Ian Rogers1.15 events were released in:
https://github.com/intel/perfmon/commit/76dfb81a1148ec049fd9caae9c62529404da63df
Adds the events OCR.DEMAND_DATA_RD.LOCAL_SOCKET_PMM and OCR.DEMAND_DATA_RD.PMM.
--- tools/perf/pmu-events/arch/x86/mapfile.csv | 2 +- .../arch/x86/sapphirerapids/other.json | 18 ++++++++++++++++++ 2 files changed, 19 insertions(+), 1 deletion(-)
tools/perf/pmu-events/arch/x86/mapfile.csv +1 -1
@@ -24,7 +24,7 @@ GenuineIntel-6-1[AEF],v3,nehalemep,core GenuineIntel-6-2E,v3,nehalemex,core GenuineIntel-6-A7,v1.01,rocketlake,core GenuineIntel-6-2A,v19,sandybridge,core -GenuineIntel-6-(8F|CF),v1.14,sapphirerapids,core +GenuineIntel-6-(8F|CF),v1.15,sapphirerapids,core GenuineIntel-6-AF,v1.00,sierraforest,core GenuineIntel-6-(37|4A|4C|4D|5A),v15,silvermont,core GenuineIntel-6-(4E|5E|8E|9E|A5|A6),v57,skylake,core
tools/perf/pmu-events/arch/x86/sapphirerapids/other.json +18 -0
@@ -76,6 +76,24 @@ "SampleAfterValue": "100003", "UMask": "0x1" }, + { + "BriefDescription": "Counts demand data reads that were supplied by PMM attached to this socket, whether or not in Sub NUMA Cluster(SNC) Mode. In SNC Mode counts PMM accesses that are controlled by the close or distant SNC Cluster.", + "EventCode": "0x2A,0x2B", + "EventName": "OCR.DEMAND_DATA_RD.LOCAL_SOCKET_PMM", + "MSRIndex": "0x1a6,0x1a7", + "MSRValue": "0x700C00001", + "SampleAfterValue": "100003", + "UMask": "0x1" + }, + { + "BriefDescription": "Counts demand data reads that were supplied by PMM.", + "EventCode": "0x2A,0x2B", + "EventName": "OCR.DEMAND_DATA_RD.PMM", + "MSRIndex": "0x1a6,0x1a7", + "MSRValue": "0x703C00001", + "SampleAfterValue": "100003", + "UMask": "0x1" + }, { "BriefDescription": "Counts demand data reads that were supplied by DRAM attached to another socket.", "EventCode": "0x2A,0x2B",
Signature · 1 line
2.41.0.585.gd2178a4bd4-goog
Ian Rogers
reviewerPATCH 4/4 perf vendor events intel: Update Icelake+ metric constraints
In reply to Ian RogersAvoid grouping events especially in cases where the kernel's PMU driver fails to not open the events, causing the events to report back as "<not counted>".
This update comes from:
https://github.com/intel/perfmon/pull/94
Fixes issues reported with patch:
https://lore.kernel.org/lkml/[email protected]/
--- .../pmu-events/arch/x86/alderlake/adl-metrics.json | 11 +++++++---- .../pmu-events/arch/x86/alderlaken/adln-metrics.json | 2 ++ .../perf/pmu-events/arch/x86/icelake/icl-metrics.json | 10 ++++++---- .../pmu-events/arch/x86/icelakex/icx-metrics.json | 10 ++++++---- .../pmu-events/arch/x86/rocketlake/rkl-metrics.json | 10 ++++++---- .../arch/x86/sapphirerapids/spr-metrics.json | 9 +++++---- .../pmu-events/arch/x86/tigerlake/tgl-metrics.json | 10 ++++++---- 7 files changed, 38 insertions(+), 24 deletions(-)
tools/perf/pmu-events/arch/x86/alderlake/adl-metrics.json +7 -4
@@ -558,6 +558,7 @@ }, { "BriefDescription": "Counts the number of cycles a core is stalled due to a demand load which hit in the Last Level Cache (LLC) or other core with HITE/F/M.", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "cpu_atom@MEM_BOUND_STALLS.LOAD_LLC_HIT@ / tma_info_core_clks - max((cpu_atom@MEM_BOUND_STALLS.LOAD@ - cpu_atom@LD_HEAD.L1_MISS_AT_RET@) / tma_info_core_clks, 0) * cpu_atom@MEM_BOUND_STALLS.LOAD_LLC_HIT@ / cpu_atom@MEM_BOUND_STALLS.LOAD@", "MetricGroup": "TopdownL3;tma_L3_group;tma_memory_bound_group", "MetricName": "tma_l3_bound",
@@ -800,6 +801,7 @@ }, { "BriefDescription": "Counts the number of cycles that the oldest load of the load buffer is stalled at retirement due to a store forward block.", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "LD_HEAD.ST_ADDR_AT_RET / tma_info_core_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_l1_bound_group", "MetricName": "tma_store_fwd_blk",
@@ -1058,7 +1060,6 @@ }, { "BriefDescription": "This metric represents overall arithmetic floating-point (FP) operations fraction the CPU has executed (retired)", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "tma_x87_use + tma_fp_scalar + tma_fp_vector", "MetricGroup": "HPC;TopdownL3;tma_L3_group;tma_light_operations_group", "MetricName": "tma_fp_arith",
@@ -1230,6 +1231,7 @@ }, { "BriefDescription": "Total pipeline cost of Instruction Cache misses - subset of the Big_Code Bottleneck", + "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "100 * (tma_fetch_latency * tma_icache_misses / (tma_branch_resteers + tma_dsb_switches + tma_icache_misses + tma_itlb_misses + tma_lcp + tma_ms_switches))", "MetricGroup": "Fed;FetchLat;IcMiss;tma_issueFL", "MetricName": "tma_info_botlnk_l2_ic_misses",
@@ -1267,6 +1269,7 @@ }, { "BriefDescription": "Total pipeline cost of (external) Memory Bandwidth related bottlenecks", + "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "100 * tma_memory_bound * (tma_dram_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_store_bound) * (tma_mem_bandwidth / (tma_mem_bandwidth + tma_mem_latency)) + tma_l3_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_store_bound) * (tma_sq_full / (tma_contested_accesses + tma_data_sharing + tma_l3_hit_latency + tma_sq_full))) + tma_l1_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_store_bound) * (tma_fb_full / (tma_dtlb_load + tma_fb_full + tma_lock_latency + tma_split_loads + tma_store_fwd_blk))", "MetricGroup": "Mem;MemoryBW;Offcore;tma_issueBW", "MetricName": "tma_info_bottleneck_memory_bandwidth",
@@ -1355,7 +1358,6 @@ }, { "BriefDescription": "Floating Point Operations Per Cycle", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "(cpu_core@FP_ARITH_INST_RETIRED.SCALAR_SINGLE@ + cpu_core@FP_ARITH_INST_RETIRED.SCALAR_DOUBLE@ + 2 * cpu_core@FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE@ + 4 * (cpu_core@FP_ARITH_INST_RETIRED.128B_PACKED_SINGLE@ + cpu_core@FP_ARITH_INST_RETIRED.256B_PACKED_DOUBLE@) + 8 * cpu_core@FP_ARITH_INST_RETIRED.256B_PACKED_SINGLE@) / tma_info_core_core_clks", "MetricGroup": "Flops;Ret", "MetricName": "tma_info_core_flopc",
@@ -1363,7 +1365,6 @@ }, { "BriefDescription": "Actual per-core usage of the Floating Point non-X87 execution units (regardless of precision or vector-width)", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "(cpu_core@FP_ARITH_DISPATCHED.PORT_0@ + cpu_core@FP_ARITH_DISPATCHED.PORT_1@ + cpu_core@FP_ARITH_DISPATCHED.PORT_5@) / (2 * tma_info_core_core_clks)", "MetricGroup": "Cor;Flops;HPC", "MetricName": "tma_info_core_fp_arith_utilization",
@@ -1769,7 +1770,6 @@ }, { "BriefDescription": "Average number of Uops retired in cycles where at least one uop has retired.", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "tma_retiring * tma_info_thread_slots / cpu_core@UOPS_RETIRED.SLOTS\\,cmask\\=1@", "MetricGroup": "Pipeline;Ret", "MetricName": "tma_info_pipeline_retire",
@@ -2002,6 +2002,7 @@ }, { "BriefDescription": "This metric estimates how often the CPU was stalled due to loads accesses to L3 cache or contended with a sibling Core", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "(cpu_core@MEMORY_ACTIVITY.STALLS_L2_MISS@ - cpu_core@MEMORY_ACTIVITY.STALLS_L3_MISS@) / tma_info_thread_clks", "MetricGroup": "CacheMisses;MemoryBound;TmaL3mem;TopdownL3;tma_L3_group;tma_memory_bound_group", "MetricName": "tma_l3_bound",
@@ -2375,6 +2376,7 @@ }, { "BriefDescription": "This metric represents rate of split store accesses", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "MEM_INST_RETIRED.SPLIT_STORES / tma_info_core_core_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_issueSpSt;tma_store_bound_group", "MetricName": "tma_split_stores",
@@ -2405,6 +2407,7 @@ }, { "BriefDescription": "This metric roughly estimates fraction of cycles when the memory subsystem had loads blocked since they could not forward data from earlier (in program order) overlapping stores", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "13 * cpu_core@LD_BLOCKS.STORE_FORWARD@ / tma_info_thread_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_l1_bound_group", "MetricName": "tma_store_fwd_blk",
tools/perf/pmu-events/arch/x86/alderlaken/adln-metrics.json +2 -0
@@ -466,6 +466,7 @@ }, { "BriefDescription": "Counts the number of cycles a core is stalled due to a demand load which hit in the Last Level Cache (LLC) or other core with HITE/F/M.", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "MEM_BOUND_STALLS.LOAD_LLC_HIT / tma_info_core_clks - max((MEM_BOUND_STALLS.LOAD - LD_HEAD.L1_MISS_AT_RET) / tma_info_core_clks, 0) * MEM_BOUND_STALLS.LOAD_LLC_HIT / MEM_BOUND_STALLS.LOAD", "MetricGroup": "TopdownL3;tma_L3_group;tma_memory_bound_group", "MetricName": "tma_l3_bound",
@@ -682,6 +683,7 @@ }, { "BriefDescription": "Counts the number of cycles that the oldest load of the load buffer is stalled at retirement due to a store forward block.", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "LD_HEAD.ST_ADDR_AT_RET / tma_info_core_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_l1_bound_group", "MetricName": "tma_store_fwd_blk",
tools/perf/pmu-events/arch/x86/icelake/icl-metrics.json +6 -4
@@ -85,6 +85,7 @@ }, { "BriefDescription": "This metric estimates how often memory load accesses were aliased by preceding stores (in program order) with a 4K address offset", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "LD_BLOCKS_PARTIAL.ADDRESS_ALIAS / tma_info_thread_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_l1_bound_group", "MetricName": "tma_4k_aliasing",
@@ -319,7 +320,6 @@ }, { "BriefDescription": "This metric represents overall arithmetic floating-point (FP) operations fraction the CPU has executed (retired)", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "tma_x87_use + tma_fp_scalar + tma_fp_vector", "MetricGroup": "HPC;TopdownL3;tma_L3_group;tma_light_operations_group", "MetricName": "tma_fp_arith",
@@ -464,6 +464,7 @@ }, { "BriefDescription": "Total pipeline cost of Instruction Cache misses - subset of the Big_Code Bottleneck", + "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "100 * (tma_fetch_latency * tma_icache_misses / (tma_branch_resteers + tma_dsb_switches + tma_icache_misses + tma_itlb_misses + tma_lcp + tma_ms_switches))", "MetricGroup": "Fed;FetchLat;IcMiss;tma_issueFL", "MetricName": "tma_info_botlnk_l2_ic_misses",
@@ -497,6 +498,7 @@ }, { "BriefDescription": "Total pipeline cost of (external) Memory Bandwidth related bottlenecks", + "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "100 * tma_memory_bound * (tma_dram_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_store_bound) * (tma_mem_bandwidth / (tma_mem_bandwidth + tma_mem_latency)) + tma_l3_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_store_bound) * (tma_sq_full / (tma_contested_accesses + tma_data_sharing + tma_l3_hit_latency + tma_sq_full))) + tma_l1_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_store_bound) * (tma_fb_full / (tma_4k_aliasing + tma_dtlb_load + tma_fb_full + tma_lock_latency + tma_split_loads + tma_store_fwd_blk))", "MetricGroup": "Mem;MemoryBW;Offcore;tma_issueBW", "MetricName": "tma_info_bottleneck_memory_bandwidth",
@@ -574,14 +576,12 @@ }, { "BriefDescription": "Floating Point Operations Per Cycle", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "(cpu@FP_ARITH_INST_RETIRED.SCALAR_SINGLE\\,umask\\=0x03@ + 2 * FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE + 4 * cpu@FP_ARITH_INST_RETIRED.128B_PACKED_SINGLE\\,umask\\=0x18@ + 8 * cpu@FP_ARITH_INST_RETIRED.256B_PACKED_SINGLE\\,umask\\=0x60@ + 16 * FP_ARITH_INST_RETIRED.512B_PACKED_SINGLE) / tma_info_core_core_clks", "MetricGroup": "Flops;Ret", "MetricName": "tma_info_core_flopc" }, { "BriefDescription": "Actual per-core usage of the Floating Point non-X87 execution units (regardless of precision or vector-width)", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "(cpu@FP_ARITH_INST_RETIRED.SCALAR_SINGLE\\,umask\\=0x03@ + cpu@FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE\\,umask\\=0xfc@) / (2 * tma_info_core_core_clks)", "MetricGroup": "Cor;Flops;HPC", "MetricName": "tma_info_core_fp_arith_utilization",
@@ -927,7 +927,6 @@ }, { "BriefDescription": "Average number of Uops retired in cycles where at least one uop has retired.", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "tma_retiring * tma_info_thread_slots / cpu@UOPS_RETIRED.SLOTS\\,cmask\\=1@", "MetricGroup": "Pipeline;Ret", "MetricName": "tma_info_pipeline_retire"
@@ -1100,6 +1099,7 @@ }, { "BriefDescription": "This metric estimates how often the CPU was stalled due to loads accesses to L3 cache or contended with a sibling Core", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "(CYCLE_ACTIVITY.STALLS_L2_MISS - CYCLE_ACTIVITY.STALLS_L3_MISS) / tma_info_thread_clks", "MetricGroup": "CacheMisses;MemoryBound;TmaL3mem;TopdownL3;tma_L3_group;tma_memory_bound_group", "MetricName": "tma_l3_bound",
@@ -1419,6 +1419,7 @@ }, { "BriefDescription": "This metric represents rate of split store accesses", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "MEM_INST_RETIRED.SPLIT_STORES / tma_info_core_core_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_issueSpSt;tma_store_bound_group", "MetricName": "tma_split_stores",
@@ -1446,6 +1447,7 @@ }, { "BriefDescription": "This metric roughly estimates fraction of cycles when the memory subsystem had loads blocked since they could not forward data from earlier (in program order) overlapping stores", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "13 * LD_BLOCKS.STORE_FORWARD / tma_info_thread_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_l1_bound_group", "MetricName": "tma_store_fwd_blk",
tools/perf/pmu-events/arch/x86/icelakex/icx-metrics.json +6 -4
@@ -289,6 +289,7 @@ }, { "BriefDescription": "This metric estimates how often memory load accesses were aliased by preceding stores (in program order) with a 4K address offset", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "LD_BLOCKS_PARTIAL.ADDRESS_ALIAS / tma_info_thread_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_l1_bound_group", "MetricName": "tma_4k_aliasing",
@@ -523,7 +524,6 @@ }, { "BriefDescription": "This metric represents overall arithmetic floating-point (FP) operations fraction the CPU has executed (retired)", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "tma_x87_use + tma_fp_scalar + tma_fp_vector", "MetricGroup": "HPC;TopdownL3;tma_L3_group;tma_light_operations_group", "MetricName": "tma_fp_arith",
@@ -668,6 +668,7 @@ }, { "BriefDescription": "Total pipeline cost of Instruction Cache misses - subset of the Big_Code Bottleneck", + "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "100 * (tma_fetch_latency * tma_icache_misses / (tma_branch_resteers + tma_dsb_switches + tma_icache_misses + tma_itlb_misses + tma_lcp + tma_ms_switches))", "MetricGroup": "Fed;FetchLat;IcMiss;tma_issueFL", "MetricName": "tma_info_botlnk_l2_ic_misses",
@@ -701,6 +702,7 @@ }, { "BriefDescription": "Total pipeline cost of (external) Memory Bandwidth related bottlenecks", + "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "100 * tma_memory_bound * (tma_dram_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_pmm_bound + tma_store_bound) * (tma_mem_bandwidth / (tma_mem_bandwidth + tma_mem_latency)) + tma_l3_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_pmm_bound + tma_store_bound) * (tma_sq_full / (tma_contested_accesses + tma_data_sharing + tma_l3_hit_latency + tma_sq_full))) + tma_l1_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_pmm_bound + tma_store_bound) * (tma_fb_full / (tma_4k_aliasing + tma_dtlb_load + tma_fb_full + tma_lock_latency + tma_split_loads + tma_store_fwd_blk))", "MetricGroup": "Mem;MemoryBW;Offcore;tma_issueBW", "MetricName": "tma_info_bottleneck_memory_bandwidth",
@@ -778,14 +780,12 @@ }, { "BriefDescription": "Floating Point Operations Per Cycle", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "(cpu@FP_ARITH_INST_RETIRED.SCALAR_SINGLE\\,umask\\=0x03@ + 2 * FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE + 4 * cpu@FP_ARITH_INST_RETIRED.128B_PACKED_SINGLE\\,umask\\=0x18@ + 8 * cpu@FP_ARITH_INST_RETIRED.256B_PACKED_SINGLE\\,umask\\=0x60@ + 16 * FP_ARITH_INST_RETIRED.512B_PACKED_SINGLE) / tma_info_core_core_clks", "MetricGroup": "Flops;Ret", "MetricName": "tma_info_core_flopc" }, { "BriefDescription": "Actual per-core usage of the Floating Point non-X87 execution units (regardless of precision or vector-width)", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "(cpu@FP_ARITH_INST_RETIRED.SCALAR_SINGLE\\,umask\\=0x03@ + cpu@FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE\\,umask\\=0xfc@) / (2 * tma_info_core_core_clks)", "MetricGroup": "Cor;Flops;HPC", "MetricName": "tma_info_core_fp_arith_utilization",
@@ -1144,7 +1144,6 @@ }, { "BriefDescription": "Average number of Uops retired in cycles where at least one uop has retired.", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "tma_retiring * tma_info_thread_slots / cpu@UOPS_RETIRED.SLOTS\\,cmask\\=1@", "MetricGroup": "Pipeline;Ret", "MetricName": "tma_info_pipeline_retire"
@@ -1369,6 +1368,7 @@ }, { "BriefDescription": "This metric estimates how often the CPU was stalled due to loads accesses to L3 cache or contended with a sibling Core", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "(CYCLE_ACTIVITY.STALLS_L2_MISS - CYCLE_ACTIVITY.STALLS_L3_MISS) / tma_info_thread_clks", "MetricGroup": "CacheMisses;MemoryBound;TmaL3mem;TopdownL3;tma_L3_group;tma_memory_bound_group", "MetricName": "tma_l3_bound",
@@ -1715,6 +1715,7 @@ }, { "BriefDescription": "This metric represents rate of split store accesses", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "MEM_INST_RETIRED.SPLIT_STORES / tma_info_core_core_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_issueSpSt;tma_store_bound_group", "MetricName": "tma_split_stores",
@@ -1742,6 +1743,7 @@ }, { "BriefDescription": "This metric roughly estimates fraction of cycles when the memory subsystem had loads blocked since they could not forward data from earlier (in program order) overlapping stores", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "13 * LD_BLOCKS.STORE_FORWARD / tma_info_thread_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_l1_bound_group", "MetricName": "tma_store_fwd_blk",
tools/perf/pmu-events/arch/x86/rocketlake/rkl-metrics.json +6 -4
@@ -85,6 +85,7 @@ }, { "BriefDescription": "This metric estimates how often memory load accesses were aliased by preceding stores (in program order) with a 4K address offset", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "LD_BLOCKS_PARTIAL.ADDRESS_ALIAS / tma_info_thread_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_l1_bound_group", "MetricName": "tma_4k_aliasing",
@@ -319,7 +320,6 @@ }, { "BriefDescription": "This metric represents overall arithmetic floating-point (FP) operations fraction the CPU has executed (retired)", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "tma_x87_use + tma_fp_scalar + tma_fp_vector", "MetricGroup": "HPC;TopdownL3;tma_L3_group;tma_light_operations_group", "MetricName": "tma_fp_arith",
@@ -464,6 +464,7 @@ }, { "BriefDescription": "Total pipeline cost of Instruction Cache misses - subset of the Big_Code Bottleneck", + "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "100 * (tma_fetch_latency * tma_icache_misses / (tma_branch_resteers + tma_dsb_switches + tma_icache_misses + tma_itlb_misses + tma_lcp + tma_ms_switches))", "MetricGroup": "Fed;FetchLat;IcMiss;tma_issueFL", "MetricName": "tma_info_botlnk_l2_ic_misses",
@@ -497,6 +498,7 @@ }, { "BriefDescription": "Total pipeline cost of (external) Memory Bandwidth related bottlenecks", + "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "100 * tma_memory_bound * (tma_dram_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_store_bound) * (tma_mem_bandwidth / (tma_mem_bandwidth + tma_mem_latency)) + tma_l3_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_store_bound) * (tma_sq_full / (tma_contested_accesses + tma_data_sharing + tma_l3_hit_latency + tma_sq_full))) + tma_l1_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_store_bound) * (tma_fb_full / (tma_4k_aliasing + tma_dtlb_load + tma_fb_full + tma_lock_latency + tma_split_loads + tma_store_fwd_blk))", "MetricGroup": "Mem;MemoryBW;Offcore;tma_issueBW", "MetricName": "tma_info_bottleneck_memory_bandwidth",
@@ -574,14 +576,12 @@ }, { "BriefDescription": "Floating Point Operations Per Cycle", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "(cpu@FP_ARITH_INST_RETIRED.SCALAR_SINGLE\\,umask\\=0x03@ + 2 * FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE + 4 * cpu@FP_ARITH_INST_RETIRED.128B_PACKED_SINGLE\\,umask\\=0x18@ + 8 * cpu@FP_ARITH_INST_RETIRED.256B_PACKED_SINGLE\\,umask\\=0x60@ + 16 * FP_ARITH_INST_RETIRED.512B_PACKED_SINGLE) / tma_info_core_core_clks", "MetricGroup": "Flops;Ret", "MetricName": "tma_info_core_flopc" }, { "BriefDescription": "Actual per-core usage of the Floating Point non-X87 execution units (regardless of precision or vector-width)", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "(cpu@FP_ARITH_INST_RETIRED.SCALAR_SINGLE\\,umask\\=0x03@ + cpu@FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE\\,umask\\=0xfc@) / (2 * tma_info_core_core_clks)", "MetricGroup": "Cor;Flops;HPC", "MetricName": "tma_info_core_fp_arith_utilization",
@@ -933,7 +933,6 @@ }, { "BriefDescription": "Average number of Uops retired in cycles where at least one uop has retired.", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "tma_retiring * tma_info_thread_slots / cpu@UOPS_RETIRED.SLOTS\\,cmask\\=1@", "MetricGroup": "Pipeline;Ret", "MetricName": "tma_info_pipeline_retire"
@@ -1126,6 +1125,7 @@ }, { "BriefDescription": "This metric estimates how often the CPU was stalled due to loads accesses to L3 cache or contended with a sibling Core", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "(CYCLE_ACTIVITY.STALLS_L2_MISS - CYCLE_ACTIVITY.STALLS_L3_MISS) / tma_info_thread_clks", "MetricGroup": "CacheMisses;MemoryBound;TmaL3mem;TopdownL3;tma_L3_group;tma_memory_bound_group", "MetricName": "tma_l3_bound",
@@ -1445,6 +1445,7 @@ }, { "BriefDescription": "This metric represents rate of split store accesses", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "MEM_INST_RETIRED.SPLIT_STORES / tma_info_core_core_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_issueSpSt;tma_store_bound_group", "MetricName": "tma_split_stores",
@@ -1472,6 +1473,7 @@ }, { "BriefDescription": "This metric roughly estimates fraction of cycles when the memory subsystem had loads blocked since they could not forward data from earlier (in program order) overlapping stores", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "13 * LD_BLOCKS.STORE_FORWARD / tma_info_thread_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_l1_bound_group", "MetricName": "tma_store_fwd_blk",
tools/perf/pmu-events/arch/x86/sapphirerapids/spr-metrics.json +5 -4
@@ -553,7 +553,6 @@ }, { "BriefDescription": "This metric represents overall arithmetic floating-point (FP) operations fraction the CPU has executed (retired)", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "tma_x87_use + tma_fp_scalar + tma_fp_vector + tma_fp_amx", "MetricGroup": "HPC;TopdownL3;tma_L3_group;tma_light_operations_group", "MetricName": "tma_fp_arith",
@@ -717,6 +716,7 @@ }, { "BriefDescription": "Total pipeline cost of Instruction Cache misses - subset of the Big_Code Bottleneck", + "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "100 * (tma_fetch_latency * tma_icache_misses / (tma_branch_resteers + tma_dsb_switches + tma_icache_misses + tma_itlb_misses + tma_lcp + tma_ms_switches))", "MetricGroup": "Fed;FetchLat;IcMiss;tma_issueFL", "MetricName": "tma_info_botlnk_l2_ic_misses",
@@ -750,6 +750,7 @@ }, { "BriefDescription": "Total pipeline cost of (external) Memory Bandwidth related bottlenecks", + "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "100 * tma_memory_bound * (tma_dram_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_pmm_bound + tma_store_bound) * (tma_mem_bandwidth / (tma_mem_bandwidth + tma_mem_latency)) + tma_l3_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_pmm_bound + tma_store_bound) * (tma_sq_full / (tma_contested_accesses + tma_data_sharing + tma_l3_hit_latency + tma_sq_full))) + tma_l1_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_pmm_bound + tma_store_bound) * (tma_fb_full / (tma_dtlb_load + tma_fb_full + tma_lock_latency + tma_split_loads + tma_store_fwd_blk))", "MetricGroup": "Mem;MemoryBW;Offcore;tma_issueBW", "MetricName": "tma_info_bottleneck_memory_bandwidth",
@@ -827,14 +828,12 @@ }, { "BriefDescription": "Floating Point Operations Per Cycle", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "cpu@FP_ARITH_INST_RETIRED.SCALAR_SINGLE\\,umask\\=0x03@ + FP_ARITH_INST_RETIRED2.SCALAR_HALF + 2 * (FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE + FP_ARITH_INST_RETIRED2.COMPLEX_SCALAR_HALF) + 4 * cpu@FP_ARITH_INST_RETIRED.128B_PACKED_SINGLE\\,umask\\=0x18@ + 8 * (FP_ARITH_INST_RETIRED2.128B_PACKED_HALF + cpu@FP_ARITH_INST_RETIRED.256B_PACKED_SINGLE\\,umask\\=0x60@) + 16 * (FP_ARITH_INST_RETIRED2.256B_PACKED_HALF + FP_ARITH_INST_RETIRED.512B_PACKED_SINGLE) + 32 * FP_ARITH_INST_RETIRED2.512B_PACKED_HALF + 4 * AMX_OPS_RETIRED.BF16", "MetricGroup": "Flops;Ret", "MetricName": "tma_info_core_flopc" }, { "BriefDescription": "Actual per-core usage of the Floating Point non-X87 execution units (regardless of precision or vector-width)", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "(FP_ARITH_DISPATCHED.PORT_0 + FP_ARITH_DISPATCHED.PORT_1 + FP_ARITH_DISPATCHED.PORT_5) / (2 * tma_info_core_core_clks)", "MetricGroup": "Cor;Flops;HPC", "MetricName": "tma_info_core_fp_arith_utilization",
@@ -1216,7 +1215,6 @@ }, { "BriefDescription": "Average number of Uops retired in cycles where at least one uop has retired.", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "tma_retiring * tma_info_thread_slots / cpu@UOPS_RETIRED.SLOTS\\,cmask\\=1@", "MetricGroup": "Pipeline;Ret", "MetricName": "tma_info_pipeline_retire"
@@ -1467,6 +1465,7 @@ }, { "BriefDescription": "This metric estimates how often the CPU was stalled due to loads accesses to L3 cache or contended with a sibling Core", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "(MEMORY_ACTIVITY.STALLS_L2_MISS - MEMORY_ACTIVITY.STALLS_L3_MISS) / tma_info_thread_clks", "MetricGroup": "CacheMisses;MemoryBound;TmaL3mem;TopdownL3;tma_L3_group;tma_memory_bound_group", "MetricName": "tma_l3_bound",
@@ -1841,6 +1840,7 @@ }, { "BriefDescription": "This metric represents rate of split store accesses", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "MEM_INST_RETIRED.SPLIT_STORES / tma_info_core_core_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_issueSpSt;tma_store_bound_group", "MetricName": "tma_split_stores",
@@ -1868,6 +1868,7 @@ }, { "BriefDescription": "This metric roughly estimates fraction of cycles when the memory subsystem had loads blocked since they could not forward data from earlier (in program order) overlapping stores", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "13 * LD_BLOCKS.STORE_FORWARD / tma_info_thread_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_l1_bound_group", "MetricName": "tma_store_fwd_blk",
tools/perf/pmu-events/arch/x86/tigerlake/tgl-metrics.json +6 -4
@@ -79,6 +79,7 @@ }, { "BriefDescription": "This metric estimates how often memory load accesses were aliased by preceding stores (in program order) with a 4K address offset", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "LD_BLOCKS_PARTIAL.ADDRESS_ALIAS / tma_info_thread_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_l1_bound_group", "MetricName": "tma_4k_aliasing",
@@ -313,7 +314,6 @@ }, { "BriefDescription": "This metric represents overall arithmetic floating-point (FP) operations fraction the CPU has executed (retired)", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "tma_x87_use + tma_fp_scalar + tma_fp_vector", "MetricGroup": "HPC;TopdownL3;tma_L3_group;tma_light_operations_group", "MetricName": "tma_fp_arith",
@@ -458,6 +458,7 @@ }, { "BriefDescription": "Total pipeline cost of Instruction Cache misses - subset of the Big_Code Bottleneck", + "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "100 * (tma_fetch_latency * tma_icache_misses / (tma_branch_resteers + tma_dsb_switches + tma_icache_misses + tma_itlb_misses + tma_lcp + tma_ms_switches))", "MetricGroup": "Fed;FetchLat;IcMiss;tma_issueFL", "MetricName": "tma_info_botlnk_l2_ic_misses",
@@ -491,6 +492,7 @@ }, { "BriefDescription": "Total pipeline cost of (external) Memory Bandwidth related bottlenecks", + "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "100 * tma_memory_bound * (tma_dram_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_store_bound) * (tma_mem_bandwidth / (tma_mem_bandwidth + tma_mem_latency)) + tma_l3_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_store_bound) * (tma_sq_full / (tma_contested_accesses + tma_data_sharing + tma_l3_hit_latency + tma_sq_full))) + tma_l1_bound / (tma_dram_bound + tma_l1_bound + tma_l2_bound + tma_l3_bound + tma_store_bound) * (tma_fb_full / (tma_4k_aliasing + tma_dtlb_load + tma_fb_full + tma_lock_latency + tma_split_loads + tma_store_fwd_blk))", "MetricGroup": "Mem;MemoryBW;Offcore;tma_issueBW", "MetricName": "tma_info_bottleneck_memory_bandwidth",
@@ -568,14 +570,12 @@ }, { "BriefDescription": "Floating Point Operations Per Cycle", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "(cpu@FP_ARITH_INST_RETIRED.SCALAR_SINGLE\\,umask\\=0x03@ + 2 * FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE + 4 * cpu@FP_ARITH_INST_RETIRED.128B_PACKED_SINGLE\\,umask\\=0x18@ + 8 * cpu@FP_ARITH_INST_RETIRED.256B_PACKED_SINGLE\\,umask\\=0x60@ + 16 * FP_ARITH_INST_RETIRED.512B_PACKED_SINGLE) / tma_info_core_core_clks", "MetricGroup": "Flops;Ret", "MetricName": "tma_info_core_flopc" }, { "BriefDescription": "Actual per-core usage of the Floating Point non-X87 execution units (regardless of precision or vector-width)", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "(cpu@FP_ARITH_INST_RETIRED.SCALAR_SINGLE\\,umask\\=0x03@ + cpu@FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE\\,umask\\=0xfc@) / (2 * tma_info_core_core_clks)", "MetricGroup": "Cor;Flops;HPC", "MetricName": "tma_info_core_fp_arith_utilization",
@@ -927,7 +927,6 @@ }, { "BriefDescription": "Average number of Uops retired in cycles where at least one uop has retired.", - "MetricConstraint": "NO_GROUP_EVENTS", "MetricExpr": "tma_retiring * tma_info_thread_slots / cpu@UOPS_RETIRED.SLOTS\\,cmask\\=1@", "MetricGroup": "Pipeline;Ret", "MetricName": "tma_info_pipeline_retire"
@@ -1114,6 +1113,7 @@ }, { "BriefDescription": "This metric estimates how often the CPU was stalled due to loads accesses to L3 cache or contended with a sibling Core", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "(CYCLE_ACTIVITY.STALLS_L2_MISS - CYCLE_ACTIVITY.STALLS_L3_MISS) / tma_info_thread_clks", "MetricGroup": "CacheMisses;MemoryBound;TmaL3mem;TopdownL3;tma_L3_group;tma_memory_bound_group", "MetricName": "tma_l3_bound",
@@ -1433,6 +1433,7 @@ }, { "BriefDescription": "This metric represents rate of split store accesses", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "MEM_INST_RETIRED.SPLIT_STORES / tma_info_core_core_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_issueSpSt;tma_store_bound_group", "MetricName": "tma_split_stores",
@@ -1460,6 +1461,7 @@ }, { "BriefDescription": "This metric roughly estimates fraction of cycles when the memory subsystem had loads blocked since they could not forward data from earlier (in program order) overlapping stores", + "MetricConstraint": "NO_GROUP_EVENTS_NMI", "MetricExpr": "13 * LD_BLOCKS.STORE_FORWARD / tma_info_thread_clks", "MetricGroup": "TopdownL4;tma_L4_group;tma_l1_bound_group", "MetricName": "tma_store_fwd_blk",
Signature · 1 line
2.41.0.585.gd2178a4bd4-goog
Quoted 13 lines · Ian Rogers
The metric
tma_info_pipeline_retirecontains uops_retired.slots with perf metric events. Patch 1 fixes this event sorting so that uops_retired.slots isn't made a group leader as that needs to be topdown.slots.Patch 2 and 3 update the meteorlake and sapphirerapids events.
Patch 4 addresses an issue with event grouping discussed in:
https://lore.kernel.org/lkml/[email protected]/
by adding and altering metric constraints. The constraints avoid groups for metrics where the kernel PMU fails to not open the group (the trigger for the weak group being removed).
Quoted 6 lines
Ian Rogers (4):
perf parse-events x86: Avoid sorting uops_retired.slots perf vendor events intel: Update meteorlake to 1.04 perf vendor events intel: Update sapphirerapids to 1.15 perf vendor events intel: Update Icelake+ metric constraints
Thanks Ian.
Thanks,
Kan
Quoted 19 lines
tools/perf/arch/x86/util/evlist.c | 7 +- tools/perf/arch/x86/util/evsel.c | 7 +- .../arch/x86/alderlake/adl-metrics.json | 11 +- .../arch/x86/alderlaken/adln-metrics.json | 2 + .../arch/x86/icelake/icl-metrics.json | 10 +- .../arch/x86/icelakex/icx-metrics.json | 10 +- tools/perf/pmu-events/arch/x86/mapfile.csv | 4 +- .../pmu-events/arch/x86/meteorlake/cache.json | 165 ++++++++++++++++++ .../arch/x86/meteorlake/floating-point.json | 8 + .../arch/x86/meteorlake/frontend.json | 56 ++++++ .../arch/x86/meteorlake/memory.json | 80 +++++++++ .../pmu-events/arch/x86/meteorlake/other.json | 16 ++ .../arch/x86/meteorlake/pipeline.json | 159 +++++++++++++++++ .../arch/x86/rocketlake/rkl-metrics.json | 10 +- .../arch/x86/sapphirerapids/other.json | 18 ++ .../arch/x86/sapphirerapids/spr-metrics.json | 9 +- .../arch/x86/tigerlake/tgl-metrics.json | 10 +- 17 files changed, 549 insertions(+), 33 deletions(-)
Em Tue, Aug 01, 2023 at 01:48:45PM -0400, Liang, Kan escreveu:
Quoted 23 lines · Ian Rogers
Quoted 13 lines
The metric
tma_info_pipeline_retirecontains uops_retired.slots with perf metric events. Patch 1 fixes this event sorting so that uops_retired.slots isn't made a group leader as that needs to be topdown.slots.Patch 2 and 3 update the meteorlake and sapphirerapids events.
Patch 4 addresses an issue with event grouping discussed in:
https://lore.kernel.org/lkml/[email protected]/
by adding and altering metric constraints. The constraints avoid groups for metrics where the kernel PMU fails to not open the group (the trigger for the weak group being removed).Quoted 6 lines
Ian Rogers (4):
perf parse-events x86: Avoid sorting uops_retired.slots perf vendor events intel: Update meteorlake to 1.04 perf vendor events intel: Update sapphirerapids to 1.15 perf vendor events intel: Update Icelake+ metric constraints
Quoted 3 lines
Thanks Ian.
Reviewed-by Kan Liang <[email protected]>
Thanks, applied.
- Arnaldo