Skip to contentSkip to conversation

perf vendor events arm64: AmpereOne: Core PMU event update and metrics

PATCH In review Ilkka Koskinen 3 participants 17 messages Aug 3, 2023 – Aug 14, 2023

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover perf vendor events arm64: AmpereOne: Core PMU event update and metrics

AmpereOne Core PMU documentation had unsupported events listed, which have been removed since. In addition, certain STALL_* events have a bug and, thus, the errata number is added for reference. Finally, metrics are added for AmpereOne.

Ilkka Koskinen (4):
perf vendor events arm64: Remove L1D_CACHE_LMISS from AmpereOne list perf vendor events arm64: AmpereOne: Mark affected STALL_* events
impacted by errata
perf vendor events arm64: Add AmpereOne metrics
perf vendor events arm64: AmpereOne: Remove unsupported events

 .../arch/arm64/ampere/ampereone/cache.json    |   3 -
 .../arm64/ampere/ampereone/core-imp-def.json  | 120 ------
 .../arch/arm64/ampere/ampereone/metrics.json  | 362 ++++++++++++++++++
 .../arch/arm64/ampere/ampereone/pipeline.json |  12 +-
 4 files changed, 371 insertions(+), 126 deletions(-)
 create mode 100644 tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json
Signature · 1 line
2.40.1
#1 lore raw

PATCH 1/4 perf vendor events arm64: Remove L1D_CACHE_LMISS from AmpereOne list

amperene/cache.json file tried to include L1D_CACHE_LMISS while it doesn't exist in common-and-microarch.json. While this bug doesn't seem to cause issue in newer kernels with jevents.py script, it prevents building older perf tools with the backported patch.

Fixes a9650b7f6fc0 ("perf vendor events arm64: Add AmpereOne core PMU events")
---
 tools/perf/pmu-events/arch/arm64/ampere/ampereone/cache.json | 3 ---
 1 file changed, 3 deletions(-)
tools/perf/pmu-events/arch/arm64/ampere/ampereone/cache.json +0 -3
index fc0633054211..7a2b7b200f14 100644
@@ -92,9 +92,6 @@
     {
         "ArchStdEvent": "L1D_CACHE_LMISS_RD"
     },
-    {
-        "ArchStdEvent": "L1D_CACHE_LMISS"
-    },
     {
         "ArchStdEvent": "L1I_CACHE_LMISS"
     },
Signature · 1 line
2.40.1
reviewer Reviewed-by on v1 1/4
#2 lore raw

Re: [PATCH 1/4] perf vendor events arm64: Remove L1D_CACHE_LMISS from AmpereOne list

Quoted 4 lines · Ilkka Koskinen

amperene/cache.json file tried to include L1D_CACHE_LMISS while it doesn't exist in common-and-microarch.json. While this bug doesn't seem to cause issue in newer kernels with jevents.py script, it prevents building older perf tools with the backported patch.

jevents.py needs to be improved so it errors on these events which cannot be fixed up, like it used to. I'll look to do that when I get a chance.

Quoted 5 lines
Quoted 18 lines
---
  tools/perf/pmu-events/arch/arm64/ampere/ampereone/cache.json | 3 ---
  1 file changed, 3 deletions(-)
tools/perf/pmu-events/arch/arm64/ampere/ampereone/cache.json +0 -3
index fc0633054211..7a2b7b200f14 100644
@@ -92,9 +92,6 @@
      {
          "ArchStdEvent": "L1D_CACHE_LMISS_RD"
      },
-    {
-        "ArchStdEvent": "L1D_CACHE_LMISS"
-    },
      {
          "ArchStdEvent": "L1I_CACHE_LMISS"
      },
reviewer Reviewed-by
#3 lore raw

Re: [PATCH 1/4] perf vendor events arm64: Remove L1D_CACHE_LMISS from AmpereOne list

Quoted 6 lines · John Garry
Quoted 4 lines · Ilkka Koskinen

amperene/cache.json file tried to include L1D_CACHE_LMISS while it doesn't exist in common-and-microarch.json. While this bug doesn't seem to cause issue in newer kernels with jevents.py script, it prevents building older perf tools with the backported patch.

Fwiw, newer perf tool on old kernel should always be fine. But I understand that if you are trying to build a tree with backports in it...

Quoted 6 lines

jevents.py needs to be improved so it errors on these events which cannot be fixed up, like it used to. I'll look to do that when I get a chance.

Quoted 1 line
Quoted 6 lines
Quoted 4 lines

Thanks,
Ian

Quoted 4 lines
Quoted 4 lines
---
  tools/perf/pmu-events/arch/arm64/ampere/ampereone/cache.json | 3 ---
  1 file changed, 3 deletions(-)
Quoted 15 lines
Quoted 5 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/cache.json +0 -0
index fc0633054211..7a2b7b200f14 100644

tools/perf/pmu-events/arch/arm64/ampere/ampereone/cache.json line 97

Quoted 9 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/cache.json +0 -3
@@ -92,9 +92,6 @@
      {
          "ArchStdEvent": "L1D_CACHE_LMISS_RD"
      },
-    {
-        "ArchStdEvent": "L1D_CACHE_LMISS"
-    },
      {
          "ArchStdEvent": "L1I_CACHE_LMISS"
      },

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#4 lore raw

PATCH 2/4 perf vendor events arm64: AmpereOne: Mark affected STALL_* events impacted by errata

In reply to Ilkka Koskinen

Per errata AC03_CPU_29, STALL_SLOT_FRONTEND, STALL_FRONTEND, and STALL events are not counting as expected. The follow up metrics patch will include correct way to calculate the impacted events.

---
 .../arch/arm64/ampere/ampereone/pipeline.json        | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)
tools/perf/pmu-events/arch/arm64/ampere/ampereone/pipeline.json +9 -3
index f9fae15f7555..711028377f3e 100644
@@ -1,18 +1,24 @@
 [
     {
-        "ArchStdEvent": "STALL_FRONTEND"
+        "ArchStdEvent": "STALL_FRONTEND",
+        "Errata": "Errata AC03_CPU_29",
+        "BriefDescription": "Impacted by errata, use metrics instead -"
     },
     {
         "ArchStdEvent": "STALL_BACKEND"
     },
     {
-        "ArchStdEvent": "STALL"
+        "ArchStdEvent": "STALL",
+        "Errata": "Errata AC03_CPU_29",
+        "BriefDescription": "Impacted by errata, use metrics instead -"
     },
     {
         "ArchStdEvent": "STALL_SLOT_BACKEND"
     },
     {
-        "ArchStdEvent": "STALL_SLOT_FRONTEND"
+        "ArchStdEvent": "STALL_SLOT_FRONTEND",
+        "Errata": "Errata AC03_CPU_29",
+        "BriefDescription": "Impacted by errata, use metrics instead -"
     },
     {
         "ArchStdEvent": "STALL_SLOT"
Signature · 1 line
2.40.1
reviewer Reviewed-by on v1 2/4
#5 lore raw

Re: [PATCH 2/4] perf vendor events arm64: AmpereOne: Mark affected STALL_* events impacted by errata

Quoted 5 lines · Ilkka Koskinen

Per errata AC03_CPU_29, STALL_SLOT_FRONTEND, STALL_FRONTEND, and STALL events are not counting as expected. The follow up metrics patch will include correct way to calculate the impacted events.

Quoted 9 lines
---
  .../arch/arm64/ampere/ampereone/pipeline.json        | 12 +++++++++---
  1 file changed, 9 insertions(+), 3 deletions(-)
tools/perf/pmu-events/arch/arm64/ampere/ampereone/pipeline.json +0 -0
index f9fae15f7555..711028377f3e 100644

tools/perf/pmu-events/arch/arm64/ampere/ampereone/pipeline.json line 5

Quoted 6 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/pipeline.json +3 -1
@@ -1,18 +1,24 @@
  [
      {
-        "ArchStdEvent": "STALL_FRONTEND"
+        "ArchStdEvent": "STALL_FRONTEND",
+        "Errata": "Errata AC03_CPU_29",
+        "BriefDescription": "Impacted by errata, use metrics instead -"

why end with a '-'?

Quoted 21 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/pipeline.json +6 -2
      },
      {
          "ArchStdEvent": "STALL_BACKEND"
      },
      {
-        "ArchStdEvent": "STALL"
+        "ArchStdEvent": "STALL",
+        "Errata": "Errata AC03_CPU_29",
+        "BriefDescription": "Impacted by errata, use metrics instead -"
      },
      {
          "ArchStdEvent": "STALL_SLOT_BACKEND"
      },
      {
-        "ArchStdEvent": "STALL_SLOT_FRONTEND"
+        "ArchStdEvent": "STALL_SLOT_FRONTEND",
+        "Errata": "Errata AC03_CPU_29",
+        "BriefDescription": "Impacted by errata, use metrics instead -"
      },
      {
          "ArchStdEvent": "STALL_SLOT"
#6 lore raw

Re: [PATCH 2/4] perf vendor events arm64: AmpereOne: Mark affected STALL_* events impacted by errata

Quoted 13 lines · John Garry
Quoted 5 lines · Ilkka Koskinen

Per errata AC03_CPU_29, STALL_SLOT_FRONTEND, STALL_FRONTEND, and STALL events are not counting as expected. The follow up metrics patch will include correct way to calculate the impacted events.

Quoted 4 lines
---
  .../arch/arm64/ampere/ampereone/pipeline.json        | 12 +++++++++---
  1 file changed, 9 insertions(+), 3 deletions(-)
Quoted 13 lines

tools/perf/pmu-events/arch/arm64/ampere/ampereone/pipeline.json line 5

Quoted 11 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/pipeline.json +3 -1
index f9fae15f7555..711028377f3e 100644
@@ -1,18 +1,24 @@
  [
      {
-        "ArchStdEvent": "STALL_FRONTEND"
+        "ArchStdEvent": "STALL_FRONTEND",
+        "Errata": "Errata AC03_CPU_29",
+        "BriefDescription": "Impacted by errata, use metrics instead -"

why end with a '-'?

That's a great question! I wish, I remembered why I did that. Looking at other events with Errata field, they use '.'. I can fix those and submit the patchset again.

Cheers, Ilkka

Quoted 24 lines
Quoted 22 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/pipeline.json +6 -2
      },
      {
          "ArchStdEvent": "STALL_BACKEND"
      },
      {
-        "ArchStdEvent": "STALL"
+        "ArchStdEvent": "STALL",
+        "Errata": "Errata AC03_CPU_29",
+        "BriefDescription": "Impacted by errata, use metrics instead -"
      },
      {
          "ArchStdEvent": "STALL_SLOT_BACKEND"
      },
      {
-        "ArchStdEvent": "STALL_SLOT_FRONTEND"
+        "ArchStdEvent": "STALL_SLOT_FRONTEND",
+        "Errata": "Errata AC03_CPU_29",
+        "BriefDescription": "Impacted by errata, use metrics instead -"
      },
      {
          "ArchStdEvent": "STALL_SLOT"
#7 lore raw

Re: [PATCH 2/4] perf vendor events arm64: AmpereOne: Mark affected STALL_* events impacted by errata

Quoted 29 lines · Ilkka Koskinen
Quoted 13 lines · John Garry
Quoted 5 lines · Ilkka Koskinen

Per errata AC03_CPU_29, STALL_SLOT_FRONTEND, STALL_FRONTEND, and STALL events are not counting as expected. The follow up metrics patch will include correct way to calculate the impacted events.

Quoted 4 lines
---
  .../arch/arm64/ampere/ampereone/pipeline.json        | 12 +++++++++---
  1 file changed, 9 insertions(+), 3 deletions(-)
Quoted 13 lines

tools/perf/pmu-events/arch/arm64/ampere/ampereone/pipeline.json line 5

Quoted 11 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/pipeline.json +3 -1
index f9fae15f7555..711028377f3e 100644
@@ -1,18 +1,24 @@
  [
      {
-        "ArchStdEvent": "STALL_FRONTEND"
+        "ArchStdEvent": "STALL_FRONTEND",
+        "Errata": "Errata AC03_CPU_29",
+        "BriefDescription": "Impacted by errata, use metrics instead -"

why end with a '-'?

Quoted 1 line

That's a great question! I wish, I remembered why I did that. Looking at other events with Errata field, they use '.'. I can fix those and submit the patchset again.

John, I took a look at it and it seems that if I use '.', perf will remove it. Here are two examples:

metrics.json:
"BriefDescription": "Impacted by errata, use metrics instead."

# perf list
[Impacted by errata, use metrics instead Spec update: Errata AC03_CPU_29]

vs.

"BriefDescription": "Impacted by errata, use metrics instead -"

# perf list

Impacted by errata, use metrics instead - Spec update: Errata AC03_CPU_29

---

So, I believe, I probably used the hyphen to make the 'perf list' output a little "prettier". Well, I just change to dot and it's fine.

Cheers, Ilkka

Quoted 25 lines
Quoted 24 lines
Quoted 22 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/pipeline.json +6 -2
      },
      {
          "ArchStdEvent": "STALL_BACKEND"
      },
      {
-        "ArchStdEvent": "STALL"
+        "ArchStdEvent": "STALL",
+        "Errata": "Errata AC03_CPU_29",
+        "BriefDescription": "Impacted by errata, use metrics instead -"
      },
      {
          "ArchStdEvent": "STALL_SLOT_BACKEND"
      },
      {
-        "ArchStdEvent": "STALL_SLOT_FRONTEND"
+        "ArchStdEvent": "STALL_SLOT_FRONTEND",
+        "Errata": "Errata AC03_CPU_29",
+        "BriefDescription": "Impacted by errata, use metrics instead -"
      },
      {
          "ArchStdEvent": "STALL_SLOT"
reviewer
#8 lore raw

Re: [PATCH 2/4] perf vendor events arm64: AmpereOne: Mark affected STALL_* events impacted by errata

Quoted 12 lines · Ilkka Koskinen
Quoted 3 lines
Quoted 1 line

why end with a '-'?

That's a great question! I wish, I remembered why I did that. Looking at other events with Errata field, they use '.'. I can fix those and submit the patchset again.

John, I took a look at it and it seems that if I use '.', perf will remove it. Here are two examples:

metrics.json:
"BriefDescription": "Impacted by errata, use metrics instead."

# perf list
[Impacted by errata, use metrics instead Spec update: Errata AC03_CPU_29]
Quoted 7 lines

vs.

"BriefDescription": "Impacted by errata, use metrics instead -"

# perf list

Impacted by errata, use metrics instead - Spec update: Errata AC03_CPU_29

ah, ok. I am not sure where the '.' removal is coming from.

Quoted 4 lines

---

So, I believe, I probably used the hyphen to make the 'perf list' output a little "prettier". Well, I just change to dot and it's fine.

Fine. Or we could consider changing the 'perf list' formatting code to add this '-' or similar automatically.

Thanks,
John

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#9 lore raw

PATCH 3/4 perf vendor events arm64: Add AmpereOne metrics

In reply to Ilkka Koskinen

This patch adds AmpereOne metrics. The metrics also work around the issue related to some of the events.

---
 .../arch/arm64/ampere/ampereone/metrics.json  | 362 ++++++++++++++++++
 1 file changed, 362 insertions(+)
 create mode 100644 tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +362 -0
new file mode 100644
index 000000000000..1e7e8901a445
@@ -0,0 +1,362 @@
+[
+    {
+	"MetricExpr": "BR_MIS_PRED / BR_PRED",
+	"BriefDescription": "Branch predictor misprediction rate. May not count branches that are never resolved because they are in the misprediction shadow of an earlier branch",
+	"MetricGroup": "Branch Prediction",
+	"MetricName": "Misprediction"
+    },
+    {
+	"MetricExpr": "BR_MIS_PRED_RETIRED / BR_RETIRED",
+	"BriefDescription": "Branch predictor misprediction rate",
+	"MetricGroup": "Branch Prediction",
+	"MetricName": "Misprediction (retired)"
+    },
+    {
+	"MetricExpr": "BUS_ACCESS / ( BUS_CYCLES * 1)",
+	"BriefDescription": "Core-to-uncore bus utilization",
+	"MetricGroup": "Bus",
+	"MetricName": "Bus utilization"
+    },
+    {
+	"MetricExpr": "L1D_CACHE_REFILL / L1D_CACHE",
+	"BriefDescription": "L1D cache miss rate",
+	"MetricGroup": "Cache",
+	"MetricName": "L1D cache miss"
+    },
+    {
+	"MetricExpr": "L1D_CACHE_LMISS_RD / L1D_CACHE_RD",
+	"BriefDescription": "L1D cache read miss rate",
+	"MetricGroup": "Cache",
+	"MetricName": "L1D cache read miss"
+    },
+    {
+	"MetricExpr": "L1I_CACHE_REFILL / L1I_CACHE",
+	"BriefDescription": "L1I cache miss rate",
+	"MetricGroup": "Cache",
+	"MetricName": "L1I cache miss"
+    },
+    {
+	"MetricExpr": "L2D_CACHE_REFILL / L2D_CACHE",
+	"BriefDescription": "L2 cache miss rate",
+	"MetricGroup": "Cache",
+	"MetricName": "L2 cache miss"
+    },
+    {
+	"MetricExpr": "L1I_CACHE_LMISS / L1I_CACHE",
+	"BriefDescription": "L1I cache read miss rate",
+	"MetricGroup": "Cache",
+	"MetricName": "L1I cache read miss"
+    },
+    {
+	"MetricExpr": "L2D_CACHE_LMISS_RD / L2D_CACHE_RD",
+	"BriefDescription": "L2 cache read miss rate",
+	"MetricGroup": "Cache",
+	"MetricName": "L2 cache read miss"
+    },
+    {
+	"MetricExpr": "(L1D_CACHE_LMISS_RD * 1000) / INST_RETIRED",
+	"BriefDescription": "Misses per thousand instructions (data)",
+	"MetricGroup": "Cache",
+	"MetricName": "MPKI data"
+    },
+    {
+	"MetricExpr": "(L1I_CACHE_LMISS * 1000) / INST_RETIRED",
+	"BriefDescription": "Misses per thousand instructions (instruction)",
+	"MetricGroup": "Cache",
+	"MetricName": "MPKI instruction"
+    },
+    {
+	"MetricExpr": "ASE_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of advanced SIMD data processing operations (excluding DP_SPEC/LD_SPEC) operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "ASE mix"
+    },
+    {
+	"MetricExpr": "CRYPTO_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of crypto data processing operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "Crypto mix"
+    },
+    {
+	"MetricExpr": "VFP_SPEC / (duration_time *1000000000)",
+	"BriefDescription": "Giga-floating point operations per second",
+	"MetricGroup": "Instruction",
+	"MetricName": "GFLOPS_ISSUED"
+    },
+    {
+	"MetricExpr": "DP_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of integer data processing operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "Integer mix"
+    },
+    {
+	"MetricExpr": "INST_RETIRED / CPU_CYCLES",
+	"BriefDescription": "Instructions per cycle",
+	"MetricGroup": "Instruction",
+	"MetricName": "IPC"
+    },
+    {
+	"MetricExpr": "LD_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of load operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "Load mix"
+    },
+    {
+	"MetricExpr": "LDST_SPEC/ OP_SPEC",
+	"BriefDescription": "Proportion of load & store operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "Load-store mix"
+    },
+    {
+	"MetricExpr": "INST_RETIRED / (duration_time * 1000000)",
+	"BriefDescription": "Millions of instructions per second",
+	"MetricGroup": "Instruction",
+	"MetricName": "MIPS_RETIRED"
+    },
+    {
+	"MetricExpr": "INST_SPEC / (duration_time * 1000000)",
+	"BriefDescription": "Millions of instructions per second",
+	"MetricGroup": "Instruction",
+	"MetricName": "MIPS_UTILIZATION"
+    },
+    {
+	"MetricExpr": "PC_WRITE_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of software change of PC operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "PC write mix"
+    },
+    {
+	"MetricExpr": "ST_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of store operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "Store mix"
+    },
+    {
+	"MetricExpr": "VFP_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of FP operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "VFP mix"
+    },
+    {
+	"MetricExpr": "1 - (OP_RETIRED/ (CPU_CYCLES * 4))",
+	"BriefDescription": "Proportion of slots lost",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "CPU lost"
+    },
+    {
+	"MetricExpr": "OP_RETIRED/ (CPU_CYCLES * 4)",
+	"BriefDescription": "Proportion of slots retiring",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "CPU utilization"
+    },
+    {
+	"MetricExpr": "OP_RETIRED - OP_SPEC",
+	"BriefDescription": "Operations lost due to misspeculation",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "Operations lost"
+    },
+    {
+	"MetricExpr": "1 - (OP_RETIRED / OP_SPEC)",
+	"BriefDescription": "Proportion of operations lost",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "Operations lost (ratio)"
+    },
+    {
+	"MetricExpr": "OP_RETIRED / OP_SPEC",
+	"BriefDescription": "Proportion of operations retired",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "Operations retired"
+    },
+    {
+	"MetricExpr": "STALL_BACKEND_CACHE / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no operations issued to backend and cache miss",
+	"MetricGroup": "Stall",
+	"MetricName": "Stall backend cache cycles"
+    },
+    {
+	"MetricExpr": "STALL_BACKEND_RESOURCE / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no operations issued to backend and resource full",
+	"MetricGroup": "Stall",
+	"MetricName": "Stall backend resource cycles"
+    },
+    {
+	"MetricExpr": "STALL_BACKEND_TLB / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no operations issued to backend and TLB miss",
+	"MetricGroup": "Stall",
+	"MetricName": "Stall backend tlb cycles"
+    },
+    {
+	"MetricExpr": "STALL_FRONTEND_CACHE / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no ops delivered from frontend and cache miss",
+	"MetricGroup": "Stall",
+	"MetricName": "Stall frontend cache cycles"
+    },
+    {
+	"MetricExpr": "STALL_FRONTEND_TLB / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no ops delivered from frontend and TLB miss",
+	"MetricGroup": "Stall",
+	"MetricName": "Stall frontend tlb cycles"
+    },
+    {
+	"MetricExpr": "DTLB_WALK / L1D_TLB",
+	"BriefDescription": "D-side walk per d-side translation request",
+	"MetricGroup": "TLB",
+	"MetricName": "DTLB walks"
+    },
+    {
+	"MetricExpr": "ITLB_WALK / L1I_TLB",
+	"BriefDescription": "I-side walk per i-side translation request",
+	"MetricGroup": "TLB",
+	"MetricName": "ITLB walks"
+    },
+    {
+        "MetricExpr": "STALL_SLOT_BACKEND / (CPU_CYCLES * 4)",
+        "BriefDescription": "Fraction of slots backend bound",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "backend"
+    },
+    {
+        "MetricExpr": "1 - (retiring + lost + backend)",
+        "BriefDescription": "Fraction of slots frontend bound",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "frontend"
+    },
+    {
+        "MetricExpr": "((OP_SPEC - OP_RETIRED) / (CPU_CYCLES * 4))",
+        "BriefDescription": "Fraction of slots lost due to misspeculation",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "lost"
+    },
+    {
+        "MetricExpr": "(OP_RETIRED / (CPU_CYCLES * 4))",
+        "BriefDescription": "Fraction of slots retiring, useful work",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "retiring"
+    },
+    {
+        "MetricExpr": "backend - backend_memory",
+        "BriefDescription": "Fraction of slots the CPU was stalled due to backend non-memory subsystem issues",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "backend_core"
+    },
+    {
+        "MetricExpr": "(STALL_BACKEND_TLB + STALL_BACKEND_CACHE + STALL_BACKEND_MEM) / CPU_CYCLES ",
+        "BriefDescription": "Fraction of slots the CPU was stalled due to backend memory subsystem issues (cache/tlb miss)",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "backend_memory"
+    },
+    {
+        "MetricExpr": " (BR_MIS_PRED_RETIRED / GPC_FLUSH) * lost",
+        "BriefDescription": "Fraction of slots lost due to branch misprediciton",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "branch_mispredict"
+    },
+    {
+        "MetricExpr": "frontend - frontend_latency",
+        "BriefDescription": "Fraction of slots the CPU did not dispatch at full bandwidth - able to dispatch partial slots only (1, 2, or 3 uops)",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "frontend_bandwidth"
+    },
+    {
+        "MetricExpr": "(STALL_FRONTEND - ((STALL_SLOT_FRONTEND - (frontend * CPU_CYCLES * 4)) / 4)) / CPU_CYCLES",
+        "BriefDescription": "Fraction of slots the CPU was stalled due to frontend latency issues (cache/tlb miss); nothing to dispatch",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "frontend_latency"
+    },
+    {
+        "MetricExpr": "lost - branch_mispredict",
+        "BriefDescription": "Fraction of slots lost due to other/non-branch misprediction misspeculation",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "other_clears"
+    },
+    {
+        "MetricExpr": "(IXU_NUM_UOPS_ISSUED + FSU_ISSUED) / (CPU_CYCLES * 6)",
+        "BriefDescription": "Fraction of execute slots utilized",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "pipe_utilization"
+    },
+    {
+        "MetricExpr": "STALL_BACKEND_MEM / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to data L2 cache miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "d_cache_l2_miss"
+    },
+    {
+        "MetricExpr": "STALL_BACKEND_CACHE / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to data cache miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "d_cache_miss"
+    },
+    {
+        "MetricExpr": "STALL_BACKEND_TLB / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to data TLB miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "d_tlb_miss"
+    },
+    {
+        "MetricExpr": "FSU_ISSUED / (CPU_CYCLES * 2)",
+        "BriefDescription": "Fraction of FSU execute slots utilized",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "fsu_pipe_utilization"
+    },
+    {
+        "MetricExpr": "STALL_FRONTEND_CACHE / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to instruction cache miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "i_cache_miss"
+    },
+    {
+        "MetricExpr": " STALL_FRONTEND_TLB / CPU_CYCLES ",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to instruction TLB miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "i_tlb_miss"
+    },
+    {
+        "MetricExpr": "IXU_NUM_UOPS_ISSUED / (CPU_CYCLES / 4)",
+        "BriefDescription": "Fraction of IXU execute slots utilized",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "ixu_pipe_utilization"
+    },
+    {
+        "MetricExpr": "IDR_STALL_FLUSH / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to flush recovery",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "recovery"
+    },
+    {
+        "MetricExpr": "STALL_BACKEND_RESOURCE / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to core resource shortage",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "resource"
+    },
+    {
+        "MetricExpr": "IDR_STALL_FSU_SCHED / CPU_CYCLES ",
+        "BriefDescription": "Fraction of cycles the CPU was stalled and FSU was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_fsu_sched"
+    },
+    {
+        "MetricExpr": "IDR_STALL_IXU_SCHED / CPU_CYCLES ",
+        "BriefDescription": "Fraction of cycles the CPU was stalled and IXU was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_ixu_sched"
+    },
+    {
+        "MetricExpr": "IDR_STALL_LOB_ID / CPU_CYCLES ",
+        "BriefDescription": "Fraction of cycles the CPU was stalled and LOB was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_lob_id"
+    },
+    {
+        "MetricExpr": "IDR_STALL_ROB_ID / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled and ROB was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_rob_id"
+    },
+    {
+        "MetricExpr": "IDR_STALL_SOB_ID / CPU_CYCLES ",
+        "BriefDescription": "Fraction of cycles the CPU was stalled and SOB was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_sob_id"
+    }
+]
Signature · 1 line
2.40.1
reviewer
#10 lore raw

Re: [PATCH 3/4] perf vendor events arm64: Add AmpereOne metrics

Quoted 2 lines · Ilkka Koskinen

This patch adds AmpereOne metrics. The metrics also work around the issue related to some of the events.

Just curious, are these events/metrics described in some publically-available document?

Quoted 1 line
#11 lore raw

Re: [PATCH 3/4] perf vendor events arm64: Add AmpereOne metrics

Hi John

Quoted 5 lines · John Garry
Quoted 2 lines · Ilkka Koskinen

This patch adds AmpereOne metrics. The metrics also work around the issue related to some of the events.

Just curious, are these events/metrics described in some publically-available document?

I quickly checked that and there are a spreadsheet and a document available, which list the supported PMUs, their events and metrics in the customer connect website but that requires registering.

Cheers, Ilkka

reviewer
#12 lore raw

Re: [PATCH 3/4] perf vendor events arm64: Add AmpereOne metrics

Quoted 7 lines · Ilkka Koskinen

Hi John

Quoted 3 lines · John Garry
Quoted 2 lines · Ilkka Koskinen

This patch adds AmpereOne metrics. The metrics also work around the issue related to some of the events.

Would these events be any metrics added which are not a "Topdown"? I guess no, since there are many, but I just don't know.

Quoted 5 lines
Quoted 2 lines

Just curious, are these events/metrics described in some publically-available document?

I quickly checked that and there are a spreadsheet and a document available, which list the supported PMUs, their events and metrics in the customer connect website but that requires registering.

OK, thanks for the info. I ask is it always worthwhile mentioning a link in the changelog if publicly available.

Just a few minor comments:

Quoted 8 lines · Ilkka Koskinen

This patch adds AmpereOne metrics. The metrics also work around the issue related to some of the events.

---
  .../arch/arm64/ampere/ampereone/metrics.json  | 362 ++++++++++++++++++
  1 file changed, 362 insertions(+)

...

Quoted 22 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +22 -0
+    {
+	"MetricExpr": "CRYPTO_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of crypto data processing operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "Crypto mix"
+    },
+    {
+	"MetricExpr": "VFP_SPEC / (duration_time *1000000000)",
+	"BriefDescription": "Giga-floating point operations per second",
+	"MetricGroup": "Instruction",
+	"MetricName": "GFLOPS_ISSUED"
+    },
+    {
+	"MetricExpr": "DP_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of integer data processing operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "Integer mix"
+    },
+    {
+	"MetricExpr": "INST_RETIRED / CPU_CYCLES",
+	"BriefDescription": "Instructions per cycle",
+	"MetricGroup": "Instruction",

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json line 105

Quoted 10 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +10 -0
+	"MetricName": "IPC"
+    },
+    {
+	"MetricExpr": "LD_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of load operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "Load mix"
+    },
+    {
+	"MetricExpr": "LDST_SPEC/ OP_SPEC",

mega nit: missing whitespace before '/'

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json line 111

Quoted 6 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +6 -0
+	"BriefDescription": "Proportion of load & store operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "Load-store mix"
+    },
+    {
+	"MetricExpr": "INST_RETIRED / (duration_time * 1000000)",

I think that we may use 1e6 here for shorthand - it helps avoid mistakes with too few or many '0's :)

Quoted 94 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +94 -0
+	"BriefDescription": "Millions of instructions per second",
+	"MetricGroup": "Instruction",
+	"MetricName": "MIPS_RETIRED"
+    },
+    {
+	"MetricExpr": "INST_SPEC / (duration_time * 1000000)",
+	"BriefDescription": "Millions of instructions per second",
+	"MetricGroup": "Instruction",
+	"MetricName": "MIPS_UTILIZATION"
+    },
+    {
+	"MetricExpr": "PC_WRITE_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of software change of PC operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "PC write mix"
+    },
+    {
+	"MetricExpr": "ST_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of store operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "Store mix"
+    },
+    {
+	"MetricExpr": "VFP_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of FP operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "VFP mix"
+    },
+    {
+	"MetricExpr": "1 - (OP_RETIRED/ (CPU_CYCLES * 4))",
+	"BriefDescription": "Proportion of slots lost",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "CPU lost"
+    },
+    {
+	"MetricExpr": "OP_RETIRED/ (CPU_CYCLES * 4)",
+	"BriefDescription": "Proportion of slots retiring",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "CPU utilization"
+    },
+    {
+	"MetricExpr": "OP_RETIRED - OP_SPEC",
+	"BriefDescription": "Operations lost due to misspeculation",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "Operations lost"
+    },
+    {
+	"MetricExpr": "1 - (OP_RETIRED / OP_SPEC)",
+	"BriefDescription": "Proportion of operations lost",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "Operations lost (ratio)"
+    },
+    {
+	"MetricExpr": "OP_RETIRED / OP_SPEC",
+	"BriefDescription": "Proportion of operations retired",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "Operations retired"
+    },
+    {
+	"MetricExpr": "STALL_BACKEND_CACHE / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no operations issued to backend and cache miss",
+	"MetricGroup": "Stall",
+	"MetricName": "Stall backend cache cycles"
+    },
+    {
+	"MetricExpr": "STALL_BACKEND_RESOURCE / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no operations issued to backend and resource full",
+	"MetricGroup": "Stall",
+	"MetricName": "Stall backend resource cycles"
+    },
+    {
+	"MetricExpr": "STALL_BACKEND_TLB / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no operations issued to backend and TLB miss",
+	"MetricGroup": "Stall",
+	"MetricName": "Stall backend tlb cycles"
+    },
+    {
+	"MetricExpr": "STALL_FRONTEND_CACHE / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no ops delivered from frontend and cache miss",
+	"MetricGroup": "Stall",
+	"MetricName": "Stall frontend cache cycles"
+    },
+    {
+	"MetricExpr": "STALL_FRONTEND_TLB / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no ops delivered from frontend and TLB miss",
+	"MetricGroup": "Stall",
+	"MetricName": "Stall frontend tlb cycles"
+    },
+    {
+	"MetricExpr": "DTLB_WALK / L1D_TLB",
+	"BriefDescription": "D-side walk per d-side translation request",
+	"MetricGroup": "TLB",
+	"MetricName": "DTLB walks"
+    },

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json line 215

Quoted 10 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +10 -0
+    {
+	"MetricExpr": "ITLB_WALK / L1I_TLB",
+	"BriefDescription": "I-side walk per i-side translation request",
+	"MetricGroup": "TLB",
+	"MetricName": "ITLB walks"
+    },
+    {
+        "MetricExpr": "STALL_SLOT_BACKEND / (CPU_CYCLES * 4)",
+        "BriefDescription": "Fraction of slots backend bound",
+        "MetricGroup": "TopDownL1",

@Ian, should this be "Default;TopDownL1"?

Quoted 1 line

+ "MetricName": "backend"

How about use consistent names with other other archs and arm64 platforms, like "backend_bound"? I did not check all names, but please consider this.

If 'perf topdown' is ever supported for arm64, we would prob rely on metricgroups, so would need use a fixed standard name here. Note that x86 uses custom kernel events for this instead.

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json line 222

Quoted 6 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +6 -0
+    },
+    {
+        "MetricExpr": "1 - (retiring + lost + backend)",
+        "BriefDescription": "Fraction of slots frontend bound",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "frontend"

As above, it would be "frontend_bound"

Quoted 8 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +8 -0
+    },
+    {
+        "MetricExpr": "((OP_SPEC - OP_RETIRED) / (CPU_CYCLES * 4))",
+        "BriefDescription": "Fraction of slots lost due to misspeculation",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "lost"
+    },
+    {
#13 lore raw

Re: [PATCH 3/4] perf vendor events arm64: Add AmpereOne metrics

Quoted 12 lines · John Garry
Quoted 7 lines · Ilkka Koskinen

Hi John

Quoted 3 lines · John Garry
Quoted 2 lines · Ilkka Koskinen

This patch adds AmpereOne metrics. The metrics also work around the issue related to some of the events.

Would these events be any metrics added which are not a "Topdown"? I guess no, since there are many, but I just don't know.

Quoted 1 line
Quoted 1 line
Quoted 6 lines
Quoted 4 lines
Quoted 1 line

Just curious, are these events/metrics described in some publically-available document?

I quickly checked that and there are a spreadsheet and a document available, which list the supported PMUs, their events and metrics in the customer connect website but that requires registering.

OK, thanks for the info. I ask is it always worthwhile mentioning a link in the changelog if publicly available.

I can certainly add a comment that the events are available at the customer connect website.

Quoted 44 lines

Just a few minor comments:

Quoted 8 lines · Ilkka Koskinen

This patch adds AmpereOne metrics. The metrics also work around the issue related to some of the events.

---
  .../arch/arm64/ampere/ampereone/metrics.json  | 362 ++++++++++++++++++
  1 file changed, 362 insertions(+)

...

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json

Quoted 3 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +3 -0
+    {
+	"MetricExpr": "CRYPTO_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of crypto data processing 

operations",

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json

Quoted 12 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +12 -0
+	"MetricGroup": "Instruction",
+	"MetricName": "Crypto mix"
+    },
+    {
+	"MetricExpr": "VFP_SPEC / (duration_time *1000000000)",
+	"BriefDescription": "Giga-floating point operations per second",
+	"MetricGroup": "Instruction",
+	"MetricName": "GFLOPS_ISSUED"
+    },
+    {
+	"MetricExpr": "DP_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of integer data processing 

operations",

Quoted 11 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +11 -0
+	"MetricGroup": "Instruction",
+	"MetricName": "Integer mix"
+    },
+    {
+	"MetricExpr": "INST_RETIRED / CPU_CYCLES",
+	"BriefDescription": "Instructions per cycle",
+	"MetricGroup": "Instruction",
+	"MetricName": "IPC"
+    },
+    {
+	"MetricExpr": "LD_SPEC / OP_SPEC",
Quoted 8 lines

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json line 105

Quoted 6 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +6 -0
+	"BriefDescription": "Proportion of load operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "Load mix"
+    },
+    {
+	"MetricExpr": "LDST_SPEC/ OP_SPEC",

mega nit: missing whitespace before '/'

I'll fix it.

Quoted 9 lines

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json line 111

Quoted 7 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +6 -0
+	"BriefDescription": "Proportion of load & store operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "Load-store mix"
+    },
+    {
+	"MetricExpr": "INST_RETIRED / (duration_time * 1000000)",

I think that we may use 1e6 here for shorthand - it helps avoid mistakes with too few or many '0's :)

Oh, that's great. I don't think anyone needed to use those in arm64 and I guess I didn't realize to take a look at other architectures. I'll change all the numbers.

Quoted 113 lines
Quoted 52 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +51 -0
+	"BriefDescription": "Millions of instructions per second",
+	"MetricGroup": "Instruction",
+	"MetricName": "MIPS_RETIRED"
+    },
+    {
+	"MetricExpr": "INST_SPEC / (duration_time * 1000000)",
+	"BriefDescription": "Millions of instructions per second",
+	"MetricGroup": "Instruction",
+	"MetricName": "MIPS_UTILIZATION"
+    },
+    {
+	"MetricExpr": "PC_WRITE_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of software change of PC operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "PC write mix"
+    },
+    {
+	"MetricExpr": "ST_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of store operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "Store mix"
+    },
+    {
+	"MetricExpr": "VFP_SPEC / OP_SPEC",
+	"BriefDescription": "Proportion of FP operations",
+	"MetricGroup": "Instruction",
+	"MetricName": "VFP mix"
+    },
+    {
+	"MetricExpr": "1 - (OP_RETIRED/ (CPU_CYCLES * 4))",
+	"BriefDescription": "Proportion of slots lost",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "CPU lost"
+    },
+    {
+	"MetricExpr": "OP_RETIRED/ (CPU_CYCLES * 4)",
+	"BriefDescription": "Proportion of slots retiring",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "CPU utilization"
+    },
+    {
+	"MetricExpr": "OP_RETIRED - OP_SPEC",
+	"BriefDescription": "Operations lost due to misspeculation",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "Operations lost"
+    },
+    {
+	"MetricExpr": "1 - (OP_RETIRED / OP_SPEC)",
+	"BriefDescription": "Proportion of operations lost",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "Operations lost (ratio)"

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json

Quoted 10 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +10 -0
+    },
+    {
+	"MetricExpr": "OP_RETIRED / OP_SPEC",
+	"BriefDescription": "Proportion of operations retired",
+	"MetricGroup": "Speculation / TDA",
+	"MetricName": "Operations retired"
+    },
+    {
+	"MetricExpr": "STALL_BACKEND_CACHE / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no operations 

issued to backend and cache miss",

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json

Quoted 6 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +6 -0
+	"MetricGroup": "Stall",
+	"MetricName": "Stall backend cache cycles"
+    },
+    {
+	"MetricExpr": "STALL_BACKEND_RESOURCE / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no operations 

issued to backend and resource full",

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json

Quoted 6 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +6 -0
+	"MetricGroup": "Stall",
+	"MetricName": "Stall backend resource cycles"
+    },
+    {
+	"MetricExpr": "STALL_BACKEND_TLB / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no operations 

issued to backend and TLB miss",

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json

Quoted 6 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +6 -0
+	"MetricGroup": "Stall",
+	"MetricName": "Stall backend tlb cycles"
+    },
+    {
+	"MetricExpr": "STALL_FRONTEND_CACHE / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no ops 

delivered from frontend and cache miss",

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json

Quoted 6 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +6 -0
+	"MetricGroup": "Stall",
+	"MetricName": "Stall frontend cache cycles"
+    },
+    {
+	"MetricExpr": "STALL_FRONTEND_TLB / CPU_CYCLES",
+	"BriefDescription": "Proportion of cycles stalled and no ops 

delivered from frontend and TLB miss",

Quoted 9 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +9 -0
+	"MetricGroup": "Stall",
+	"MetricName": "Stall frontend tlb cycles"
+    },
+    {
+	"MetricExpr": "DTLB_WALK / L1D_TLB",
+	"BriefDescription": "D-side walk per d-side translation request",
+	"MetricGroup": "TLB",
+	"MetricName": "DTLB walks"
+    },

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json line 215

Quoted 10 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +10 -0
+    {
+	"MetricExpr": "ITLB_WALK / L1I_TLB",
+	"BriefDescription": "I-side walk per i-side translation request",
+	"MetricGroup": "TLB",
+	"MetricName": "ITLB walks"
+    },
+    {
+        "MetricExpr": "STALL_SLOT_BACKEND / (CPU_CYCLES * 4)",
+        "BriefDescription": "Fraction of slots backend bound",
+        "MetricGroup": "TopDownL1",

@Ian, should this be "Default;TopDownL1"?

Quoted 5 lines
Quoted 1 line

+ "MetricName": "backend"

How about use consistent names with other other archs and arm64 platforms, like "backend_bound"? I did not check all names, but please consider this.

If 'perf topdown' is ever supported for arm64, we would prob rely on metricgroups, so would need use a fixed standard name here. Note that x86 uses custom kernel events for this instead.

That's an excellent point. I'll reach out to our architect and we'll change the names and groups in the patch and the document to be more consistent to the existing ones.

Quoted 8 lines

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json line 222

Quoted 6 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +6 -0
+    },
+    {
+        "MetricExpr": "1 - (retiring + lost + backend)",
+        "BriefDescription": "Fraction of slots frontend bound",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "frontend"

As above, it would be "frontend_bound"

I'll fix it.

Cheers, Ilkka

Quoted 12 lines

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json

Quoted 5 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +4 -0
+    },
+    {
+        "MetricExpr": "((OP_SPEC - OP_RETIRED) / (CPU_CYCLES * 4))",
+        "BriefDescription": "Fraction of slots lost due to 

misspeculation",

Quoted 4 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +4 -0
+        "MetricGroup": "TopDownL1",
+        "MetricName": "lost"
+    },
+    {
reviewer
#14 lore raw

Re: [PATCH 3/4] perf vendor events arm64: Add AmpereOne metrics

In reply to Ilkka Koskinen
Quoted 48 lines · Ilkka Koskinen

This patch adds AmpereOne metrics. The metrics also work around the issue related to some of the events.

---
 .../arch/arm64/ampere/ampereone/metrics.json  | 362 ++++++++++++++++++
 1 file changed, 362 insertions(+)
 create mode 100644 tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +32 -0
new file mode 100644
index 000000000000..1e7e8901a445
@@ -0,0 +1,362 @@
+[
+    {
+       "MetricExpr": "BR_MIS_PRED / BR_PRED",
+       "BriefDescription": "Branch predictor misprediction rate. May not count branches that are never resolved because they are in the misprediction shadow of an earlier branch",
+       "MetricGroup": "Branch Prediction",
+       "MetricName": "Misprediction"
+    },
+    {
+       "MetricExpr": "BR_MIS_PRED_RETIRED / BR_RETIRED",
+       "BriefDescription": "Branch predictor misprediction rate",
+       "MetricGroup": "Branch Prediction",
+       "MetricName": "Misprediction (retired)"
+    },
+    {
+       "MetricExpr": "BUS_ACCESS / ( BUS_CYCLES * 1)",
+       "BriefDescription": "Core-to-uncore bus utilization",
+       "MetricGroup": "Bus",
+       "MetricName": "Bus utilization"
+    },
+    {
+       "MetricExpr": "L1D_CACHE_REFILL / L1D_CACHE",
+       "BriefDescription": "L1D cache miss rate",
+       "MetricGroup": "Cache",
+       "MetricName": "L1D cache miss"
+    },
+    {
+       "MetricExpr": "L1D_CACHE_LMISS_RD / L1D_CACHE_RD",
+       "BriefDescription": "L1D cache read miss rate",
+       "MetricGroup": "Cache",
+       "MetricName": "L1D cache read miss"
+    },
+    {

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json line 42

Quoted 10 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +10 -0
+       "MetricExpr": "L1I_CACHE_REFILL / L1I_CACHE",
+       "BriefDescription": "L1I cache miss rate",
+       "MetricGroup": "Cache",
+       "MetricName": "L1I cache miss"
+    },
+    {
+       "MetricExpr": "L2D_CACHE_REFILL / L2D_CACHE",
+       "BriefDescription": "L2 cache miss rate",
+       "MetricGroup": "Cache",
+       "MetricName": "L2 cache miss"

I'm a bit concerned with spaces in metric names. There's logic to replace/rewrite metrics in terms of each other:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/pmu-events/metric.py?h=perf-tools-next#n561
but spaces in rewritten metric name would break this as spaces are separators.

Thanks,
Ian

Quoted 321 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +320 -1
+    },
+    {
+       "MetricExpr": "L1I_CACHE_LMISS / L1I_CACHE",
+       "BriefDescription": "L1I cache read miss rate",
+       "MetricGroup": "Cache",
+       "MetricName": "L1I cache read miss"
+    },
+    {
+       "MetricExpr": "L2D_CACHE_LMISS_RD / L2D_CACHE_RD",
+       "BriefDescription": "L2 cache read miss rate",
+       "MetricGroup": "Cache",
+       "MetricName": "L2 cache read miss"
+    },
+    {
+       "MetricExpr": "(L1D_CACHE_LMISS_RD * 1000) / INST_RETIRED",
+       "BriefDescription": "Misses per thousand instructions (data)",
+       "MetricGroup": "Cache",
+       "MetricName": "MPKI data"
+    },
+    {
+       "MetricExpr": "(L1I_CACHE_LMISS * 1000) / INST_RETIRED",
+       "BriefDescription": "Misses per thousand instructions (instruction)",
+       "MetricGroup": "Cache",
+       "MetricName": "MPKI instruction"
+    },
+    {
+       "MetricExpr": "ASE_SPEC / OP_SPEC",
+       "BriefDescription": "Proportion of advanced SIMD data processing operations (excluding DP_SPEC/LD_SPEC) operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "ASE mix"
+    },
+    {
+       "MetricExpr": "CRYPTO_SPEC / OP_SPEC",
+       "BriefDescription": "Proportion of crypto data processing operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "Crypto mix"
+    },
+    {
+       "MetricExpr": "VFP_SPEC / (duration_time *1000000000)",
+       "BriefDescription": "Giga-floating point operations per second",
+       "MetricGroup": "Instruction",
+       "MetricName": "GFLOPS_ISSUED"
+    },
+    {
+       "MetricExpr": "DP_SPEC / OP_SPEC",
+       "BriefDescription": "Proportion of integer data processing operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "Integer mix"
+    },
+    {
+       "MetricExpr": "INST_RETIRED / CPU_CYCLES",
+       "BriefDescription": "Instructions per cycle",
+       "MetricGroup": "Instruction",
+       "MetricName": "IPC"
+    },
+    {
+       "MetricExpr": "LD_SPEC / OP_SPEC",
+       "BriefDescription": "Proportion of load operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "Load mix"
+    },
+    {
+       "MetricExpr": "LDST_SPEC/ OP_SPEC",
+       "BriefDescription": "Proportion of load & store operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "Load-store mix"
+    },
+    {
+       "MetricExpr": "INST_RETIRED / (duration_time * 1000000)",
+       "BriefDescription": "Millions of instructions per second",
+       "MetricGroup": "Instruction",
+       "MetricName": "MIPS_RETIRED"
+    },
+    {
+       "MetricExpr": "INST_SPEC / (duration_time * 1000000)",
+       "BriefDescription": "Millions of instructions per second",
+       "MetricGroup": "Instruction",
+       "MetricName": "MIPS_UTILIZATION"
+    },
+    {
+       "MetricExpr": "PC_WRITE_SPEC / OP_SPEC",
+       "BriefDescription": "Proportion of software change of PC operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "PC write mix"
+    },
+    {
+       "MetricExpr": "ST_SPEC / OP_SPEC",
+       "BriefDescription": "Proportion of store operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "Store mix"
+    },
+    {
+       "MetricExpr": "VFP_SPEC / OP_SPEC",
+       "BriefDescription": "Proportion of FP operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "VFP mix"
+    },
+    {
+       "MetricExpr": "1 - (OP_RETIRED/ (CPU_CYCLES * 4))",
+       "BriefDescription": "Proportion of slots lost",
+       "MetricGroup": "Speculation / TDA",
+       "MetricName": "CPU lost"
+    },
+    {
+       "MetricExpr": "OP_RETIRED/ (CPU_CYCLES * 4)",
+       "BriefDescription": "Proportion of slots retiring",
+       "MetricGroup": "Speculation / TDA",
+       "MetricName": "CPU utilization"
+    },
+    {
+       "MetricExpr": "OP_RETIRED - OP_SPEC",
+       "BriefDescription": "Operations lost due to misspeculation",
+       "MetricGroup": "Speculation / TDA",
+       "MetricName": "Operations lost"
+    },
+    {
+       "MetricExpr": "1 - (OP_RETIRED / OP_SPEC)",
+       "BriefDescription": "Proportion of operations lost",
+       "MetricGroup": "Speculation / TDA",
+       "MetricName": "Operations lost (ratio)"
+    },
+    {
+       "MetricExpr": "OP_RETIRED / OP_SPEC",
+       "BriefDescription": "Proportion of operations retired",
+       "MetricGroup": "Speculation / TDA",
+       "MetricName": "Operations retired"
+    },
+    {
+       "MetricExpr": "STALL_BACKEND_CACHE / CPU_CYCLES",
+       "BriefDescription": "Proportion of cycles stalled and no operations issued to backend and cache miss",
+       "MetricGroup": "Stall",
+       "MetricName": "Stall backend cache cycles"
+    },
+    {
+       "MetricExpr": "STALL_BACKEND_RESOURCE / CPU_CYCLES",
+       "BriefDescription": "Proportion of cycles stalled and no operations issued to backend and resource full",
+       "MetricGroup": "Stall",
+       "MetricName": "Stall backend resource cycles"
+    },
+    {
+       "MetricExpr": "STALL_BACKEND_TLB / CPU_CYCLES",
+       "BriefDescription": "Proportion of cycles stalled and no operations issued to backend and TLB miss",
+       "MetricGroup": "Stall",
+       "MetricName": "Stall backend tlb cycles"
+    },
+    {
+       "MetricExpr": "STALL_FRONTEND_CACHE / CPU_CYCLES",
+       "BriefDescription": "Proportion of cycles stalled and no ops delivered from frontend and cache miss",
+       "MetricGroup": "Stall",
+       "MetricName": "Stall frontend cache cycles"
+    },
+    {
+       "MetricExpr": "STALL_FRONTEND_TLB / CPU_CYCLES",
+       "BriefDescription": "Proportion of cycles stalled and no ops delivered from frontend and TLB miss",
+       "MetricGroup": "Stall",
+       "MetricName": "Stall frontend tlb cycles"
+    },
+    {
+       "MetricExpr": "DTLB_WALK / L1D_TLB",
+       "BriefDescription": "D-side walk per d-side translation request",
+       "MetricGroup": "TLB",
+       "MetricName": "DTLB walks"
+    },
+    {
+       "MetricExpr": "ITLB_WALK / L1I_TLB",
+       "BriefDescription": "I-side walk per i-side translation request",
+       "MetricGroup": "TLB",
+       "MetricName": "ITLB walks"
+    },
+    {
+        "MetricExpr": "STALL_SLOT_BACKEND / (CPU_CYCLES * 4)",
+        "BriefDescription": "Fraction of slots backend bound",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "backend"
+    },
+    {
+        "MetricExpr": "1 - (retiring + lost + backend)",
+        "BriefDescription": "Fraction of slots frontend bound",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "frontend"
+    },
+    {
+        "MetricExpr": "((OP_SPEC - OP_RETIRED) / (CPU_CYCLES * 4))",
+        "BriefDescription": "Fraction of slots lost due to misspeculation",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "lost"
+    },
+    {
+        "MetricExpr": "(OP_RETIRED / (CPU_CYCLES * 4))",
+        "BriefDescription": "Fraction of slots retiring, useful work",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "retiring"
+    },
+    {
+        "MetricExpr": "backend - backend_memory",
+        "BriefDescription": "Fraction of slots the CPU was stalled due to backend non-memory subsystem issues",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "backend_core"
+    },
+    {
+        "MetricExpr": "(STALL_BACKEND_TLB + STALL_BACKEND_CACHE + STALL_BACKEND_MEM) / CPU_CYCLES ",
+        "BriefDescription": "Fraction of slots the CPU was stalled due to backend memory subsystem issues (cache/tlb miss)",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "backend_memory"
+    },
+    {
+        "MetricExpr": " (BR_MIS_PRED_RETIRED / GPC_FLUSH) * lost",
+        "BriefDescription": "Fraction of slots lost due to branch misprediciton",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "branch_mispredict"
+    },
+    {
+        "MetricExpr": "frontend - frontend_latency",
+        "BriefDescription": "Fraction of slots the CPU did not dispatch at full bandwidth - able to dispatch partial slots only (1, 2, or 3 uops)",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "frontend_bandwidth"
+    },
+    {
+        "MetricExpr": "(STALL_FRONTEND - ((STALL_SLOT_FRONTEND - (frontend * CPU_CYCLES * 4)) / 4)) / CPU_CYCLES",
+        "BriefDescription": "Fraction of slots the CPU was stalled due to frontend latency issues (cache/tlb miss); nothing to dispatch",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "frontend_latency"
+    },
+    {
+        "MetricExpr": "lost - branch_mispredict",
+        "BriefDescription": "Fraction of slots lost due to other/non-branch misprediction misspeculation",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "other_clears"
+    },
+    {
+        "MetricExpr": "(IXU_NUM_UOPS_ISSUED + FSU_ISSUED) / (CPU_CYCLES * 6)",
+        "BriefDescription": "Fraction of execute slots utilized",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "pipe_utilization"
+    },
+    {
+        "MetricExpr": "STALL_BACKEND_MEM / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to data L2 cache miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "d_cache_l2_miss"
+    },
+    {
+        "MetricExpr": "STALL_BACKEND_CACHE / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to data cache miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "d_cache_miss"
+    },
+    {
+        "MetricExpr": "STALL_BACKEND_TLB / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to data TLB miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "d_tlb_miss"
+    },
+    {
+        "MetricExpr": "FSU_ISSUED / (CPU_CYCLES * 2)",
+        "BriefDescription": "Fraction of FSU execute slots utilized",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "fsu_pipe_utilization"
+    },
+    {
+        "MetricExpr": "STALL_FRONTEND_CACHE / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to instruction cache miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "i_cache_miss"
+    },
+    {
+        "MetricExpr": " STALL_FRONTEND_TLB / CPU_CYCLES ",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to instruction TLB miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "i_tlb_miss"
+    },
+    {
+        "MetricExpr": "IXU_NUM_UOPS_ISSUED / (CPU_CYCLES / 4)",
+        "BriefDescription": "Fraction of IXU execute slots utilized",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "ixu_pipe_utilization"
+    },
+    {
+        "MetricExpr": "IDR_STALL_FLUSH / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to flush recovery",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "recovery"
+    },
+    {
+        "MetricExpr": "STALL_BACKEND_RESOURCE / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to core resource shortage",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "resource"
+    },
+    {
+        "MetricExpr": "IDR_STALL_FSU_SCHED / CPU_CYCLES ",
+        "BriefDescription": "Fraction of cycles the CPU was stalled and FSU was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_fsu_sched"
+    },
+    {
+        "MetricExpr": "IDR_STALL_IXU_SCHED / CPU_CYCLES ",
+        "BriefDescription": "Fraction of cycles the CPU was stalled and IXU was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_ixu_sched"
+    },
+    {
+        "MetricExpr": "IDR_STALL_LOB_ID / CPU_CYCLES ",
+        "BriefDescription": "Fraction of cycles the CPU was stalled and LOB was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_lob_id"
+    },
+    {
+        "MetricExpr": "IDR_STALL_ROB_ID / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled and ROB was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_rob_id"
+    },
+    {
+        "MetricExpr": "IDR_STALL_SOB_ID / CPU_CYCLES ",
+        "BriefDescription": "Fraction of cycles the CPU was stalled and SOB was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_sob_id"
+    }
+]
--
Quoted 2 lines

2.40.1

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#15 lore raw

Re: [PATCH 3/4] perf vendor events arm64: Add AmpereOne metrics

Quoted 66 lines · Ian Rogers
Quoted 48 lines · Ilkka Koskinen

This patch adds AmpereOne metrics. The metrics also work around the issue related to some of the events.

---
 .../arch/arm64/ampere/ampereone/metrics.json  | 362 ++++++++++++++++++
 1 file changed, 362 insertions(+)
 create mode 100644 tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +32 -0
new file mode 100644
index 000000000000..1e7e8901a445
@@ -0,0 +1,362 @@
+[
+    {
+       "MetricExpr": "BR_MIS_PRED / BR_PRED",
+       "BriefDescription": "Branch predictor misprediction rate. May not count branches that are never resolved because they are in the misprediction shadow of an earlier branch",
+       "MetricGroup": "Branch Prediction",
+       "MetricName": "Misprediction"
+    },
+    {
+       "MetricExpr": "BR_MIS_PRED_RETIRED / BR_RETIRED",
+       "BriefDescription": "Branch predictor misprediction rate",
+       "MetricGroup": "Branch Prediction",
+       "MetricName": "Misprediction (retired)"
+    },
+    {
+       "MetricExpr": "BUS_ACCESS / ( BUS_CYCLES * 1)",
+       "BriefDescription": "Core-to-uncore bus utilization",
+       "MetricGroup": "Bus",
+       "MetricName": "Bus utilization"
+    },
+    {
+       "MetricExpr": "L1D_CACHE_REFILL / L1D_CACHE",
+       "BriefDescription": "L1D cache miss rate",
+       "MetricGroup": "Cache",
+       "MetricName": "L1D cache miss"
+    },
+    {
+       "MetricExpr": "L1D_CACHE_LMISS_RD / L1D_CACHE_RD",
+       "BriefDescription": "L1D cache read miss rate",
+       "MetricGroup": "Cache",
+       "MetricName": "L1D cache read miss"
+    },
+    {

tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json line 42

Quoted 10 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +10 -0
+       "MetricExpr": "L1I_CACHE_REFILL / L1I_CACHE",
+       "BriefDescription": "L1I cache miss rate",
+       "MetricGroup": "Cache",
+       "MetricName": "L1I cache miss"
+    },
+    {
+       "MetricExpr": "L2D_CACHE_REFILL / L2D_CACHE",
+       "BriefDescription": "L2 cache miss rate",
+       "MetricGroup": "Cache",
+       "MetricName": "L2 cache miss"

I'm a bit concerned with spaces in metric names. There's logic to replace/rewrite metrics in terms of each other:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/pmu-events/metric.py?h=perf-tools-next#n561
but spaces in rewritten metric name would break this as spaces are separators.

Quoted 2 lines

Thanks,
Ian

After John's reply of using the same metric group names as the existing ones, I decided to go through all the metric names and groups with our architect and use the same or at least similar names as all the other architectures have done. So, there shouldn't be spaces in the next version any more.

Cheers, Ilkka

Quoted 316 lines
Quoted 316 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +315 -0
+    },
+    {
+       "MetricExpr": "L1I_CACHE_LMISS / L1I_CACHE",
+       "BriefDescription": "L1I cache read miss rate",
+       "MetricGroup": "Cache",
+       "MetricName": "L1I cache read miss"
+    },
+    {
+       "MetricExpr": "L2D_CACHE_LMISS_RD / L2D_CACHE_RD",
+       "BriefDescription": "L2 cache read miss rate",
+       "MetricGroup": "Cache",
+       "MetricName": "L2 cache read miss"
+    },
+    {
+       "MetricExpr": "(L1D_CACHE_LMISS_RD * 1000) / INST_RETIRED",
+       "BriefDescription": "Misses per thousand instructions (data)",
+       "MetricGroup": "Cache",
+       "MetricName": "MPKI data"
+    },
+    {
+       "MetricExpr": "(L1I_CACHE_LMISS * 1000) / INST_RETIRED",
+       "BriefDescription": "Misses per thousand instructions (instruction)",
+       "MetricGroup": "Cache",
+       "MetricName": "MPKI instruction"
+    },
+    {
+       "MetricExpr": "ASE_SPEC / OP_SPEC",
+       "BriefDescription": "Proportion of advanced SIMD data processing operations (excluding DP_SPEC/LD_SPEC) operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "ASE mix"
+    },
+    {
+       "MetricExpr": "CRYPTO_SPEC / OP_SPEC",
+       "BriefDescription": "Proportion of crypto data processing operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "Crypto mix"
+    },
+    {
+       "MetricExpr": "VFP_SPEC / (duration_time *1000000000)",
+       "BriefDescription": "Giga-floating point operations per second",
+       "MetricGroup": "Instruction",
+       "MetricName": "GFLOPS_ISSUED"
+    },
+    {
+       "MetricExpr": "DP_SPEC / OP_SPEC",
+       "BriefDescription": "Proportion of integer data processing operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "Integer mix"
+    },
+    {
+       "MetricExpr": "INST_RETIRED / CPU_CYCLES",
+       "BriefDescription": "Instructions per cycle",
+       "MetricGroup": "Instruction",
+       "MetricName": "IPC"
+    },
+    {
+       "MetricExpr": "LD_SPEC / OP_SPEC",
+       "BriefDescription": "Proportion of load operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "Load mix"
+    },
+    {
+       "MetricExpr": "LDST_SPEC/ OP_SPEC",
+       "BriefDescription": "Proportion of load & store operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "Load-store mix"
+    },
+    {
+       "MetricExpr": "INST_RETIRED / (duration_time * 1000000)",
+       "BriefDescription": "Millions of instructions per second",
+       "MetricGroup": "Instruction",
+       "MetricName": "MIPS_RETIRED"
+    },
+    {
+       "MetricExpr": "INST_SPEC / (duration_time * 1000000)",
+       "BriefDescription": "Millions of instructions per second",
+       "MetricGroup": "Instruction",
+       "MetricName": "MIPS_UTILIZATION"
+    },
+    {
+       "MetricExpr": "PC_WRITE_SPEC / OP_SPEC",
+       "BriefDescription": "Proportion of software change of PC operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "PC write mix"
+    },
+    {
+       "MetricExpr": "ST_SPEC / OP_SPEC",
+       "BriefDescription": "Proportion of store operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "Store mix"
+    },
+    {
+       "MetricExpr": "VFP_SPEC / OP_SPEC",
+       "BriefDescription": "Proportion of FP operations",
+       "MetricGroup": "Instruction",
+       "MetricName": "VFP mix"
+    },
+    {
+       "MetricExpr": "1 - (OP_RETIRED/ (CPU_CYCLES * 4))",
+       "BriefDescription": "Proportion of slots lost",
+       "MetricGroup": "Speculation / TDA",
+       "MetricName": "CPU lost"
+    },
+    {
+       "MetricExpr": "OP_RETIRED/ (CPU_CYCLES * 4)",
+       "BriefDescription": "Proportion of slots retiring",
+       "MetricGroup": "Speculation / TDA",
+       "MetricName": "CPU utilization"
+    },
+    {
+       "MetricExpr": "OP_RETIRED - OP_SPEC",
+       "BriefDescription": "Operations lost due to misspeculation",
+       "MetricGroup": "Speculation / TDA",
+       "MetricName": "Operations lost"
+    },
+    {
+       "MetricExpr": "1 - (OP_RETIRED / OP_SPEC)",
+       "BriefDescription": "Proportion of operations lost",
+       "MetricGroup": "Speculation / TDA",
+       "MetricName": "Operations lost (ratio)"
+    },
+    {
+       "MetricExpr": "OP_RETIRED / OP_SPEC",
+       "BriefDescription": "Proportion of operations retired",
+       "MetricGroup": "Speculation / TDA",
+       "MetricName": "Operations retired"
+    },
+    {
+       "MetricExpr": "STALL_BACKEND_CACHE / CPU_CYCLES",
+       "BriefDescription": "Proportion of cycles stalled and no operations issued to backend and cache miss",
+       "MetricGroup": "Stall",
+       "MetricName": "Stall backend cache cycles"
+    },
+    {
+       "MetricExpr": "STALL_BACKEND_RESOURCE / CPU_CYCLES",
+       "BriefDescription": "Proportion of cycles stalled and no operations issued to backend and resource full",
+       "MetricGroup": "Stall",
+       "MetricName": "Stall backend resource cycles"
+    },
+    {
+       "MetricExpr": "STALL_BACKEND_TLB / CPU_CYCLES",
+       "BriefDescription": "Proportion of cycles stalled and no operations issued to backend and TLB miss",
+       "MetricGroup": "Stall",
+       "MetricName": "Stall backend tlb cycles"
+    },
+    {
+       "MetricExpr": "STALL_FRONTEND_CACHE / CPU_CYCLES",
+       "BriefDescription": "Proportion of cycles stalled and no ops delivered from frontend and cache miss",
+       "MetricGroup": "Stall",
+       "MetricName": "Stall frontend cache cycles"
+    },
+    {
+       "MetricExpr": "STALL_FRONTEND_TLB / CPU_CYCLES",
+       "BriefDescription": "Proportion of cycles stalled and no ops delivered from frontend and TLB miss",
+       "MetricGroup": "Stall",
+       "MetricName": "Stall frontend tlb cycles"
+    },
+    {
+       "MetricExpr": "DTLB_WALK / L1D_TLB",
+       "BriefDescription": "D-side walk per d-side translation request",
+       "MetricGroup": "TLB",
+       "MetricName": "DTLB walks"
+    },
+    {
+       "MetricExpr": "ITLB_WALK / L1I_TLB",
+       "BriefDescription": "I-side walk per i-side translation request",
+       "MetricGroup": "TLB",
+       "MetricName": "ITLB walks"
+    },
+    {
+        "MetricExpr": "STALL_SLOT_BACKEND / (CPU_CYCLES * 4)",
+        "BriefDescription": "Fraction of slots backend bound",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "backend"
+    },
+    {
+        "MetricExpr": "1 - (retiring + lost + backend)",
+        "BriefDescription": "Fraction of slots frontend bound",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "frontend"
+    },
+    {
+        "MetricExpr": "((OP_SPEC - OP_RETIRED) / (CPU_CYCLES * 4))",
+        "BriefDescription": "Fraction of slots lost due to misspeculation",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "lost"
+    },
+    {
+        "MetricExpr": "(OP_RETIRED / (CPU_CYCLES * 4))",
+        "BriefDescription": "Fraction of slots retiring, useful work",
+        "MetricGroup": "TopDownL1",
+        "MetricName": "retiring"
+    },
+    {
+        "MetricExpr": "backend - backend_memory",
+        "BriefDescription": "Fraction of slots the CPU was stalled due to backend non-memory subsystem issues",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "backend_core"
+    },
+    {
+        "MetricExpr": "(STALL_BACKEND_TLB + STALL_BACKEND_CACHE + STALL_BACKEND_MEM) / CPU_CYCLES ",
+        "BriefDescription": "Fraction of slots the CPU was stalled due to backend memory subsystem issues (cache/tlb miss)",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "backend_memory"
+    },
+    {
+        "MetricExpr": " (BR_MIS_PRED_RETIRED / GPC_FLUSH) * lost",
+        "BriefDescription": "Fraction of slots lost due to branch misprediciton",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "branch_mispredict"
+    },
+    {
+        "MetricExpr": "frontend - frontend_latency",
+        "BriefDescription": "Fraction of slots the CPU did not dispatch at full bandwidth - able to dispatch partial slots only (1, 2, or 3 uops)",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "frontend_bandwidth"
+    },
+    {
+        "MetricExpr": "(STALL_FRONTEND - ((STALL_SLOT_FRONTEND - (frontend * CPU_CYCLES * 4)) / 4)) / CPU_CYCLES",
+        "BriefDescription": "Fraction of slots the CPU was stalled due to frontend latency issues (cache/tlb miss); nothing to dispatch",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "frontend_latency"
+    },
+    {
+        "MetricExpr": "lost - branch_mispredict",
+        "BriefDescription": "Fraction of slots lost due to other/non-branch misprediction misspeculation",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "other_clears"
+    },
+    {
+        "MetricExpr": "(IXU_NUM_UOPS_ISSUED + FSU_ISSUED) / (CPU_CYCLES * 6)",
+        "BriefDescription": "Fraction of execute slots utilized",
+        "MetricGroup": "TopDownL2",
+        "MetricName": "pipe_utilization"
+    },
+    {
+        "MetricExpr": "STALL_BACKEND_MEM / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to data L2 cache miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "d_cache_l2_miss"
+    },
+    {
+        "MetricExpr": "STALL_BACKEND_CACHE / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to data cache miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "d_cache_miss"
+    },
+    {
+        "MetricExpr": "STALL_BACKEND_TLB / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to data TLB miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "d_tlb_miss"
+    },
+    {
+        "MetricExpr": "FSU_ISSUED / (CPU_CYCLES * 2)",
+        "BriefDescription": "Fraction of FSU execute slots utilized",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "fsu_pipe_utilization"
+    },
+    {
+        "MetricExpr": "STALL_FRONTEND_CACHE / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to instruction cache miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "i_cache_miss"
+    },
+    {
+        "MetricExpr": " STALL_FRONTEND_TLB / CPU_CYCLES ",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to instruction TLB miss",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "i_tlb_miss"
+    },
+    {
+        "MetricExpr": "IXU_NUM_UOPS_ISSUED / (CPU_CYCLES / 4)",
+        "BriefDescription": "Fraction of IXU execute slots utilized",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "ixu_pipe_utilization"
+    },
+    {
+        "MetricExpr": "IDR_STALL_FLUSH / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to flush recovery",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "recovery"
+    },
+    {
+        "MetricExpr": "STALL_BACKEND_RESOURCE / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled due to core resource shortage",
+        "MetricGroup": "TopDownL3",
+        "MetricName": "resource"
+    },
+    {
+        "MetricExpr": "IDR_STALL_FSU_SCHED / CPU_CYCLES ",
+        "BriefDescription": "Fraction of cycles the CPU was stalled and FSU was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_fsu_sched"
+    },
+    {
+        "MetricExpr": "IDR_STALL_IXU_SCHED / CPU_CYCLES ",
+        "BriefDescription": "Fraction of cycles the CPU was stalled and IXU was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_ixu_sched"
+    },
+    {
+        "MetricExpr": "IDR_STALL_LOB_ID / CPU_CYCLES ",
+        "BriefDescription": "Fraction of cycles the CPU was stalled and LOB was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_lob_id"
+    },
+    {
+        "MetricExpr": "IDR_STALL_ROB_ID / CPU_CYCLES",
+        "BriefDescription": "Fraction of cycles the CPU was stalled and ROB was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_rob_id"
+    },
+    {
+        "MetricExpr": "IDR_STALL_SOB_ID / CPU_CYCLES ",
Quoted 9 lines
Quoted 8 lines
tools/perf/pmu-events/arch/arm64/ampere/ampereone/metrics.json new file +5 -1
+        "BriefDescription": "Fraction of cycles the CPU was stalled and SOB was full",
+        "MetricGroup": "TopDownL4",
+        "MetricName": "stall_sob_id"
+    }
+]
--

2.40.1

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#16 lore raw

PATCH 4/4 perf vendor events arm64: AmpereOne: Remove unsupported events

In reply to Ilkka Koskinen

Some of the events included in the ampereone/core-imp-def are not supported on AmpereOne, remove them.

---
 .../arm64/ampere/ampereone/core-imp-def.json  | 120 ------------------
 1 file changed, 120 deletions(-)
tools/perf/pmu-events/arch/arm64/ampere/ampereone/core-imp-def.json +0 -120
index 95c30243f2b2..88b23b85e33c 100644
@@ -533,66 +533,6 @@
         "EventName": "MMU_D_OTB_ALLOC",
         "BriefDescription": "L2D OTB allocate"
     },
-    {
-        "PublicDescription": "DTLB Translation cache hit on S1L2 walk cache entry",
-        "EventCode": "0xD801",
-        "EventName": "MMU_D_TRANS_CACHE_HIT_S1L2_WALK",
-        "BriefDescription": "DTLB Translation cache hit on S1L2 walk cache entry"
-    },
-    {
-        "PublicDescription": "DTLB Translation cache hit on S1L1 walk cache entry",
-        "EventCode": "0xD802",
-        "EventName": "MMU_D_TRANS_CACHE_HIT_S1L1_WALK",
-        "BriefDescription": "DTLB Translation cache hit on S1L1 walk cache entry"
-    },
-    {
-        "PublicDescription": "DTLB Translation cache hit on S1L0 walk cache entry",
-        "EventCode": "0xD803",
-        "EventName": "MMU_D_TRANS_CACHE_HIT_S1L0_WALK",
-        "BriefDescription": "DTLB Translation cache hit on S1L0 walk cache entry"
-    },
-    {
-        "PublicDescription": "DTLB Translation cache hit on S2L2 walk cache entry",
-        "EventCode": "0xD804",
-        "EventName": "MMU_D_TRANS_CACHE_HIT_S2L2_WALK",
-        "BriefDescription": "DTLB Translation cache hit on S2L2 walk cache entry"
-    },
-    {
-        "PublicDescription": "DTLB Translation cache hit on S2L1 walk cache entry",
-        "EventCode": "0xD805",
-        "EventName": "MMU_D_TRANS_CACHE_HIT_S2L1_WALK",
-        "BriefDescription": "DTLB Translation cache hit on S2L1 walk cache entry"
-    },
-    {
-        "PublicDescription": "DTLB Translation cache hit on S2L0 walk cache entry",
-        "EventCode": "0xD806",
-        "EventName": "MMU_D_TRANS_CACHE_HIT_S2L0_WALK",
-        "BriefDescription": "DTLB Translation cache hit on S2L0 walk cache entry"
-    },
-    {
-        "PublicDescription": "D-side S1 Page walk cache lookup",
-        "EventCode": "0xD807",
-        "EventName": "MMU_D_S1_WALK_CACHE_LOOKUP",
-        "BriefDescription": "D-side S1 Page walk cache lookup"
-    },
-    {
-        "PublicDescription": "D-side S1 Page walk cache refill",
-        "EventCode": "0xD808",
-        "EventName": "MMU_D_S1_WALK_CACHE_REFILL",
-        "BriefDescription": "D-side S1 Page walk cache refill"
-    },
-    {
-        "PublicDescription": "D-side S2 Page walk cache lookup",
-        "EventCode": "0xD809",
-        "EventName": "MMU_D_S2_WALK_CACHE_LOOKUP",
-        "BriefDescription": "D-side S2 Page walk cache lookup"
-    },
-    {
-        "PublicDescription": "D-side S2 Page walk cache refill",
-        "EventCode": "0xD80A",
-        "EventName": "MMU_D_S2_WALK_CACHE_REFILL",
-        "BriefDescription": "D-side S2 Page walk cache refill"
-    },
     {
         "PublicDescription": "D-side Stage1 tablewalk fault",
         "EventCode": "0xD80B",
@@ -617,66 +557,6 @@
         "EventName": "MMU_I_OTB_ALLOC",
         "BriefDescription": "L2I OTB allocate"
     },
-    {
-        "PublicDescription": "ITLB Translation cache hit on S1L2 walk cache entry",
-        "EventCode": "0xD901",
-        "EventName": "MMU_I_TRANS_CACHE_HIT_S1L2_WALK",
-        "BriefDescription": "ITLB Translation cache hit on S1L2 walk cache entry"
-    },
-    {
-        "PublicDescription": "ITLB Translation cache hit on S1L1 walk cache entry",
-        "EventCode": "0xD902",
-        "EventName": "MMU_I_TRANS_CACHE_HIT_S1L1_WALK",
-        "BriefDescription": "ITLB Translation cache hit on S1L1 walk cache entry"
-    },
-    {
-        "PublicDescription": "ITLB Translation cache hit on S1L0 walk cache entry",
-        "EventCode": "0xD903",
-        "EventName": "MMU_I_TRANS_CACHE_HIT_S1L0_WALK",
-        "BriefDescription": "ITLB Translation cache hit on S1L0 walk cache entry"
-    },
-    {
-        "PublicDescription": "ITLB Translation cache hit on S2L2 walk cache entry",
-        "EventCode": "0xD904",
-        "EventName": "MMU_I_TRANS_CACHE_HIT_S2L2_WALK",
-        "BriefDescription": "ITLB Translation cache hit on S2L2 walk cache entry"
-    },
-    {
-        "PublicDescription": "ITLB Translation cache hit on S2L1 walk cache entry",
-        "EventCode": "0xD905",
-        "EventName": "MMU_I_TRANS_CACHE_HIT_S2L1_WALK",
-        "BriefDescription": "ITLB Translation cache hit on S2L1 walk cache entry"
-    },
-    {
-        "PublicDescription": "ITLB Translation cache hit on S2L0 walk cache entry",
-        "EventCode": "0xD906",
-        "EventName": "MMU_I_TRANS_CACHE_HIT_S2L0_WALK",
-        "BriefDescription": "ITLB Translation cache hit on S2L0 walk cache entry"
-    },
-    {
-        "PublicDescription": "I-side S1 Page walk cache lookup",
-        "EventCode": "0xD907",
-        "EventName": "MMU_I_S1_WALK_CACHE_LOOKUP",
-        "BriefDescription": "I-side S1 Page walk cache lookup"
-    },
-    {
-        "PublicDescription": "I-side S1 Page walk cache refill",
-        "EventCode": "0xD908",
-        "EventName": "MMU_I_S1_WALK_CACHE_REFILL",
-        "BriefDescription": "I-side S1 Page walk cache refill"
-    },
-    {
-        "PublicDescription": "I-side S2 Page walk cache lookup",
-        "EventCode": "0xD909",
-        "EventName": "MMU_I_S2_WALK_CACHE_LOOKUP",
-        "BriefDescription": "I-side S2 Page walk cache lookup"
-    },
-    {
-        "PublicDescription": "I-side S2 Page walk cache refill",
-        "EventCode": "0xD90A",
-        "EventName": "MMU_I_S2_WALK_CACHE_REFILL",
-        "BriefDescription": "I-side S2 Page walk cache refill"
-    },
     {
         "PublicDescription": "I-side Stage1 tablewalk fault",
         "EventCode": "0xD90B",
Signature · 1 line
2.40.1