AMD, ARM, Intel metric generation with Python
Ian Rogers
reviewerv9 cover AMD, ARM, Intel metric generation with Python
Metrics in the perf tool come in via json. Json doesn't allow comments, line breaks, etc. making it an inconvenient way to write metrics. Further, it is useful to detect when writing a metric that the event specified is supported within the event json for a model. From the metric python code Event(s) are used, with fallback events provided, if no event is found then an exception is thrown and that can either indicate a failure or an unsupported model. To avoid confusion all the metrics and their metricgroups are prefixed with 'lpm_', where LPM is an abbreviation of Linux Perf Metric. While extra characters aren't ideal, this separates the metrics from other vendor provided metrics.
* The first 14 patches introduce infrastructure and fixes for the addition of metrics written in python for Arm64, AMD Zen and Intel CPUs. The ilist.py and perf python module are fixed to work better with metrics on hybrid architectures.
* The next 9 patches generate additional metrics for AMD zen. Rapl and Idle metrics aren't specific to AMD but are placed here for ease and convenience. Uncore L3 metrics are added along with the majority of core metrics.
* The next 20 patches add additional metrics for Intel. Rapl and Idle metrics aren't specific to Intel but are placed here for ease and convenience. Smi and tsx metrics are added so they can be dropped from the per model json files. There are four uncore sets of metrics and eleven core metrics. Add a CheckPmu function to metric to simplify detecting the presence of hybrid PMUs in events. Metrics with experimental events are flagged as experimental in their description.
* The next 2 patches add additional metrics for Arm64, where the topdown set decomposes yet further. The metrcs primarily use json events, where the json contains architecture standard events. Not all events are in the json, such as for a53 where the events are in sysfs. Workaround this by adding the sysfs events to the metrics but longer-term such events should be added to the json.
* The final patch validates that all events provided to an Event object exist in a json file somewhere. This is to avoid mistakes like unfortunate typos.
This series has benefitted from the input of Leo Yan
<[email protected]>, Sandipan Das <[email protected]>, Thomas Falcon <[email protected]> and Perry Taylor <[email protected]>.
v9. Drop (for now) 4 AMD sets of metrics for additional follow up. Add
reviewed-by tags from Sandipan Das (AMD) and tested-by tags from Thomas Falcon (Intel).
v8. Combine the previous 4 series for clarity. Rebase on top of the
more recent legacy metric and event changes. Make the python more pep8 and pylint compliant.
Foundations:
v6. Fix issue with '\-' escape not being '\\-' (reported-by Sandipan
Das <[email protected]>) which didn't alter the generated json.
https://lore.kernel.org/lkml/[email protected]/
v5. Rebase on top of legacy hardware/cache changes that now generate
events using python:
https://lore.kernel.org/lkml/[email protected]/
the v5 series is:
https://lore.kernel.org/lkml/[email protected]/
v4. Rebase and small Build/Makefile tweak
https://lore.kernel.org/lkml/[email protected]/
v3. Some code tidying, make the input directory a command line
argument, but no other functional or output changes.
https://lore.kernel.org/lkml/[email protected]/
v2. Fixes two type issues in the python code but no functional or
output changes.
https://lore.kernel.org/lkml/[email protected]/v1. https://lore.kernel.org/lkml/[email protected]/
AMD:
v6. Fix issue with '\-' escape not being '\\-' (reported-by Sandipan
Das <[email protected]>) which didn't alter the generated json.
https://lore.kernel.org/lkml/[email protected]/
v5. Rebase. Add uop cache hit/miss rates patch. Prefix all metric
names with lpm_ (short for Linux Perf Metric) so that python generated metrics are clearly namespaced.
https://lore.kernel.org/lkml/[email protected]/
v4. Rebase.
https://lore.kernel.org/lkml/[email protected]/
v3. Some minor code cleanup changes.
https://lore.kernel.org/lkml/[email protected]/
v2. Drop the cycles breakdown in favor of having it as a common
metric, suggested by Kan Liang <[email protected]>.
https://lore.kernel.org/lkml/[email protected]/
v1. https://lore.kernel.org/lkml/[email protected]/
Intel:
v6. Fix issue with '\-' escape not being '\\-' (reported-by Sandipan
Das <[email protected]>) which didn't alter the generated json.
https://lore.kernel.org/lkml/[email protected]/
v5. Rebase. Fix description for smi metric (Kan). Prefix all metric
names with lpm_ (short for Linux Perf Metric) so that python generated metrics are clearly namespaced. Kan requested a namespace in his review:
https://lore.kernel.org/lkml/[email protected]/
The v5 series is:
https://lore.kernel.org/lkml/[email protected]/
v4. Experimental metric descriptions. Add mesh bandwidth metric. Rebase.
https://lore.kernel.org/lkml/[email protected]/
v3. Swap tsx and CheckPMU patches that were in the wrong order. Some
minor code cleanup changes. Drop reference to merged fix for umasks/occ_sel in PCU events and for cstate metrics.
https://lore.kernel.org/lkml/[email protected]/
v2. Drop the cycles breakdown in favor of having it as a common
metric, spelling and other improvements suggested by Kan Liang <[email protected]>.
https://lore.kernel.org/lkml/[email protected]/
v1. https://lore.kernel.org/lkml/[email protected]/
ARM:
v7. Switch a use of cycles to cpu-cycles due to ARM having too many
cycles events.
https://lore.kernel.org/lkml/[email protected]/v6. Fix issue with '\-' escape not being '\\-' (reported-by Sandipan
Das <[email protected]>) which didn't alter the generated json.
https://lore.kernel.org/lkml/[email protected]/
v5. Rebase. Address review comments from Leo Yan
<[email protected]>. Prefix all metric names with lpm_ (short for Linux Perf Metric) so that python generated metrics are clearly namespaced. Use cpu-cycles rather than cycles legacy event for cycles metrics to avoid confusion with ARM PMUs. Add patch that checks events to ensure all possible event names are present in at least one json file.
https://lore.kernel.org/lkml/[email protected]/
v4. Tweak to build dependencies and rebase.
https://lore.kernel.org/lkml/[email protected]/
v3. Some minor code cleanup changes.
https://lore.kernel.org/lkml/[email protected]/
v2. The cycles metrics are now made common and shared with AMD and
Intel, suggested by Kan Liang <[email protected]>. This assumes these patches come after the AMD and Intel sets.
https://lore.kernel.org/lkml/[email protected]/
v1. https://lore.kernel.org/lkml/[email protected]/
Ian Rogers (48):
perf python: Correct copying of metric_leader in an evsel perf ilist: Be tolerant of reading a metric on the wrong CPU perf jevents: Allow multiple metricgroups.json files perf jevents: Update metric constraint support
perf jevents: Add descriptions to metricgroup abstraction perf jevents: Allow metric groups not to be named
perf jevents: Support parsing negative exponents
perf jevents: Term list fix in event parsing
perf jevents: Add threshold expressions to Metric
perf jevents: Move json encoding to its own functions perf jevents: Drop duplicate pending metrics
perf jevents: Skip optional metrics in metric group list perf jevents: Build support for generating metrics from python perf jevents: Add load event json to verify and allow fallbacks perf jevents: Add RAPL event metric for AMD zen models perf jevents: Add idle metric for AMD zen models
perf jevents: Add upc metric for uops per cycle for AMD perf jevents: Add br metric group for branch statistics on AMD perf jevents: Add itlb metric group for AMD
perf jevents: Add dtlb metric group for AMD
perf jevents: Add uncore l3 metric group for AMD
perf jevents: Add load store breakdown metrics ldst for AMD perf jevents: Add context switch metrics for AMD
perf jevents: Add RAPL metrics for all Intel models
perf jevents: Add idle metric for Intel models
perf jevents: Add CheckPmu to see if a PMU is in loaded json events perf jevents: Add smi metric group for Intel models
perf jevents: Mark metrics with experimental events as experimental perf jevents: Add tsx metric group for Intel models
perf jevents: Add br metric group for branch statistics on Intel perf jevents: Add software prefetch (swpf) metric group for Intel perf jevents: Add ports metric group giving utilization on Intel perf jevents: Add L2 metrics for Intel
perf jevents: Add load store breakdown metrics ldst for Intel perf jevents: Add ILP metrics for Intel
perf jevents: Add context switch metrics for Intel
perf jevents: Add FPU metrics for Intel
perf jevents: Add Miss Level Parallelism (MLP) metric for Intel perf jevents: Add mem_bw metric for Intel
perf jevents: Add local/remote "mem" breakdown metrics for Intel perf jevents: Add dir breakdown metrics for Intel
perf jevents: Add C-State metrics from the PCU PMU for Intel perf jevents: Add local/remote miss latency metrics for Intel perf jevents: Add upi_bw metric for Intel
perf jevents: Add mesh bandwidth saturation metric for Intel perf jevents: Add collection of topdown like metrics for arm64 perf jevents: Add cycles breakdown metric for arm64/AMD/Intel perf jevents: Validate that all names given an Event
tools/perf/.gitignore | 5 + tools/perf/Makefile.perf | 2 + tools/perf/pmu-events/Build | 51 +- tools/perf/pmu-events/amd_metrics.py | 491 ++++++++++ tools/perf/pmu-events/arm64_metrics.py | 187 ++++ tools/perf/pmu-events/common_metrics.py | 19 + tools/perf/pmu-events/intel_metrics.py | 1129 +++++++++++++++++++++++ tools/perf/pmu-events/jevents.py | 7 +- tools/perf/pmu-events/metric.py | 256 ++++- tools/perf/pmu-events/metric_test.py | 4 + tools/perf/python/ilist.py | 8 +- tools/perf/util/evsel.c | 1 + tools/perf/util/python.c | 82 +- 13 files changed, 2188 insertions(+), 54 deletions(-) create mode 100755 tools/perf/pmu-events/amd_metrics.py create mode 100755 tools/perf/pmu-events/arm64_metrics.py create mode 100644 tools/perf/pmu-events/common_metrics.py create mode 100755 tools/perf/pmu-events/intel_metrics.py
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 1/48 perf python: Correct copying of metric_leader in an evsel
Ensure the metric_leader is copied and set up correctly. In compute_metric determine the correct metric_leader event to match the requested CPU. Fixes the handling of metrics particularly on hybrid machines.
--- tools/perf/util/evsel.c | 1 + tools/perf/util/python.c | 82 +++++++++++++++++++++++++++++----------- 2 files changed, 61 insertions(+), 22 deletions(-)
tools/perf/util/evsel.c +1 -0
@@ -538,6 +538,7 @@ struct evsel *evsel__clone(struct evsel *dest, struct evsel *orig)#endif evsel->handler = orig->handler; evsel->core.leader = orig->core.leader; + evsel->metric_leader = orig->metric_leader; evsel->max_events = orig->max_events;zfree (&evsel->unit);
tools/perf/util/python.c +60 -22
132 lines not on this page · raw message
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 2/48 perf ilist: Be tolerant of reading a metric on the wrong CPU
In reply to Ian RogersThis happens on hybrid machine metrics. Be tolerant and don't cause the ilist application to crash with an exception.
--- tools/perf/python/ilist.py | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-)
tools/perf/python/ilist.py +6 -2
@@ -77,8 +77,12 @@ class Metric(TreeValue): return perf.parse_metrics(self.metric_name, self.metric_pmu) def value(self, evlist: perf.evlist, evsel: perf.evsel, cpu: int, thread: int) -> float: - val = evlist.compute_metric(self.metric_name, cpu, thread) - return 0 if math.isnan(val) else val + try: + val = evlist.compute_metric(self.metric_name, cpu, thread) + return 0 if math.isnan(val) else val + except: + # Be tolerant of failures to compute metrics on particular CPUs/threads. + return 0 @dataclass
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 3/48 perf jevents: Allow multiple metricgroups.json files
In reply to Ian RogersAllow multiple metricgroups.json files by handling any file ending with metricgroups.json as a metricgroups file.
--- tools/perf/pmu-events/jevents.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
tools/perf/pmu-events/jevents.py +2 -2
@@ -640,7 +640,7 @@ def preprocess_one_file(parents: Sequence[str], item: os.DirEntry) -> None: if not item.is_file() or not item.name.endswith('.json'): return - if item.name == 'metricgroups.json': + if item.name.endswith('metricgroups.json'): metricgroup_descriptions = json.load(open(item.path)) for mgroup in metricgroup_descriptions: assert len(mgroup) > 1, parents
@@ -693,7 +693,7 @@ def process_one_file(parents: Sequence[str], item: os.DirEntry) -> None: # Ignore other directories. If the file name does not have a .json # extension, ignore it. It could be a readme.txt for instance. - if not item.is_file() or not item.name.endswith('.json') or item.name == 'metricgroups.json': + if not item.is_file() or not item.name.endswith('.json') or item.name.endswith('metricgroups.json'): return add_events_table_entries(item, get_topic(item.name))
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 4/48 perf jevents: Update metric constraint support
In reply to Ian RogersPrevious metric constraints were binary, either none or don't group when the NMI watchdog is present. Update to match the definitions in 'enum metric_event_groups' in pmu-events.h.
--- tools/perf/pmu-events/metric.py | 14 ++++++++++---- 1 file changed, 10 insertions(+), 4 deletions(-)
tools/perf/pmu-events/metric.py +10 -4
@@ -4,8 +4,14 @@ import ast import decimal import json import re +from enum import Enum from typing import Dict, List, Optional, Set, Tuple, Union +class MetricConstraint(Enum): + GROUPED_EVENTS = 0 + NO_GROUP_EVENTS = 1 + NO_GROUP_EVENTS_NMI = 2 + NO_GROUP_EVENTS_SMT = 3 class Expression: """Abstract base class of elements in a metric expression."""
@@ -423,14 +429,14 @@ class Metric: groups: Set[str] expr: Expression scale_unit: str - constraint: bool + constraint: MetricConstraint def __init__(self, name: str, description: str, expr: Expression, scale_unit: str, - constraint: bool = False): + constraint: MetricConstraint = MetricConstraint.GROUPED_EVENTS): self.name = name self.description = description self.expr = expr.Simplify()
@@ -464,8 +470,8 @@ class Metric: 'MetricExpr': self.expr.ToPerfJson(), 'ScaleUnit': self.scale_unit } - if self.constraint: - result['MetricConstraint'] = 'NO_NMI_WATCHDOG' + if self.constraint != MetricConstraint.GROUPED_EVENTS: + result['MetricConstraint'] = self.constraint.name return result
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 5/48 perf jevents: Add descriptions to metricgroup abstraction
In reply to Ian RogersAdd a function to recursively generate metric group descriptions.
--- tools/perf/pmu-events/metric.py | 14 ++++++++++++-- 1 file changed, 12 insertions(+), 2 deletions(-)
tools/perf/pmu-events/metric.py +12 -2
@@ -475,6 +475,8 @@ class Metric: return result + def ToMetricGroupDescriptions(self, root: bool = True) -> Dict[str, str]: + return {} class _MetricJsonEncoder(json.JSONEncoder): """Special handling for Metric objects."""
@@ -493,10 +495,12 @@ class MetricGroup: which can facilitate arrangements similar to trees. """ - def __init__(self, name: str, metric_list: List[Union[Metric, - 'MetricGroup']]): + def __init__(self, name: str, + metric_list: List[Union[Metric, 'MetricGroup']], + description: Optional[str] = None): self.name = name self.metric_list = metric_list + self.description = description for metric in metric_list: metric.AddToMetricGroup(self)
@@ -516,6 +520,12 @@ class MetricGroup: def ToPerfJson(self) -> str: return json.dumps(sorted(self.Flatten()), indent=2, cls=_MetricJsonEncoder) + def ToMetricGroupDescriptions(self, root: bool = True) -> Dict[str, str]: + result = {self.name: self.description} if self.description else {} + for x in self.metric_list: + result.update(x.ToMetricGroupDescriptions(False)) + return result + def __str__(self) -> str: return self.ToPerfJson()
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 6/48 perf jevents: Allow metric groups not to be named
In reply to Ian RogersIt can be convenient to have unnamed metric groups for the sake of organizing other metrics and metric groups. An unspecified name shouldn't contribute to the MetricGroup json value, so don't record it.
--- tools/perf/pmu-events/metric.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
tools/perf/pmu-events/metric.py +2 -1
@@ -455,7 +455,8 @@ class Metric: def AddToMetricGroup(self, group): """Callback used when being added to a MetricGroup.""" - self.groups.add(group.name) + if group.name: + self.groups.add(group.name) def Flatten(self) -> Set['Metric']: """Return a leaf metric."""
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 7/48 perf jevents: Support parsing negative exponents
In reply to Ian RogersSupport negative exponents when parsing from a json metric string by making the numbers after the 'e' optional in the 'Event' insertion fix up.
--- tools/perf/pmu-events/metric.py | 2 +- tools/perf/pmu-events/metric_test.py | 4 ++++ 2 files changed, 5 insertions(+), 1 deletion(-)
tools/perf/pmu-events/metric.py +1 -1
@@ -573,7 +573,7 @@ def ParsePerfJson(orig: str) -> Expression: # a double by the Bison parser py = re.sub(r'0Event\(r"[xX]([0-9a-fA-F]*)"\)', r'Event("0x\1")', py) # Convert accidentally converted scientific notation constants back - py = re.sub(r'([0-9]+)Event\(r"(e[0-9]+)"\)', r'\1\2', py) + py = re.sub(r'([0-9]+)Event\(r"(e[0-9]*)"\)', r'\1\2', py) # Convert all the known keywords back from events to just the keyword keywords = ['if', 'else', 'min', 'max', 'd_ratio', 'source_count', 'has_event', 'strcmp_cpuid_str'] for kw in keywords:
tools/perf/pmu-events/metric_test.py +4 -0
@@ -61,6 +61,10 @@ class TestMetricExpressions(unittest.TestCase): after = before self.assertEqual(ParsePerfJson(before).ToPerfJson(), after) + before = r'a + 3e-12 + b' + after = before + self.assertEqual(ParsePerfJson(before).ToPerfJson(), after) + def test_IfElseTests(self): # if-else needs rewriting to Select and back. before = r'Event1 if #smt_on else Event2'
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 8/48 perf jevents: Term list fix in event parsing
In reply to Ian RogersFix events seemingly broken apart at a comma.
--- tools/perf/pmu-events/metric.py | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)
tools/perf/pmu-events/metric.py +6 -1
@@ -568,6 +568,12 @@ def ParsePerfJson(orig: str) -> Expression: r'Event(r"\1")', py) # If it started with a # it should have been a literal, rather than an event name py = re.sub(r'#Event\(r"([^"]*)"\)', r'Literal("#\1")', py) + # Fix events wrongly broken at a ',' + while True: + prev_py = py + py = re.sub(r'Event\(r"([^"]*)"\),Event\(r"([^"]*)"\)', r'Event(r"\1,\2")', py) + if py == prev_py: + break # Convert accidentally converted hex constants ("0Event(r"xDEADBEEF)"") back to a constant, # but keep it wrapped in Event(), otherwise Python drops the 0x prefix and it gets interpreted as # a double by the Bison parser
@@ -586,7 +592,6 @@ def ParsePerfJson(orig: str) -> Expression: parsed = ast.fix_missing_locations(parsed) return _Constify(eval(compile(parsed, orig, 'eval'))) - def RewriteMetricsInTermsOfOthers(metrics: List[Tuple[str, str, Expression]] )-> Dict[Tuple[str, str], Expression]: """Shorten metrics by rewriting in terms of others.
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 9/48 perf jevents: Add threshold expressions to Metric
In reply to Ian RogersAllow threshold expressions for metrics to be generated.
--- tools/perf/pmu-events/metric.py | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)
tools/perf/pmu-events/metric.py +6 -1
@@ -430,13 +430,15 @@ class Metric: expr: Expression scale_unit: str constraint: MetricConstraint + threshold: Optional[Expression] def __init__(self, name: str, description: str, expr: Expression, scale_unit: str, - constraint: MetricConstraint = MetricConstraint.GROUPED_EVENTS): + constraint: MetricConstraint = MetricConstraint.GROUPED_EVENTS, + threshold: Optional[Expression] = None): self.name = name self.description = description self.expr = expr.Simplify()
@@ -447,6 +449,7 @@ class Metric: else: self.scale_unit = f'1{scale_unit}' self.constraint = constraint + self.threshold = threshold self.groups = set() def __lt__(self, other):
@@ -473,6 +476,8 @@ class Metric: } if self.constraint != MetricConstraint.GROUPED_EVENTS: result['MetricConstraint'] = self.constraint.name + if self.threshold: + result['MetricThreshold'] = self.threshold.ToPerfJson() return result
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 10/48 perf jevents: Move json encoding to its own functions
In reply to Ian RogersHave dedicated encode functions rather than having them embedded in MetricGroup. This is to provide some uniformity in the Metric ToXXX routines.
--- tools/perf/pmu-events/metric.py | 34 +++++++++++++++++++++------------ 1 file changed, 22 insertions(+), 12 deletions(-)
tools/perf/pmu-events/metric.py +22 -12
@@ -484,15 +484,6 @@ class Metric: def ToMetricGroupDescriptions(self, root: bool = True) -> Dict[str, str]: return {} -class _MetricJsonEncoder(json.JSONEncoder): - """Special handling for Metric objects.""" - - def default(self, o): - if isinstance(o, Metric): - return o.ToPerfJson() - return json.JSONEncoder.default(self, o) - - class MetricGroup: """A group of metrics.
@@ -523,8 +514,11 @@ class MetricGroup: return result - def ToPerfJson(self) -> str: - return json.dumps(sorted(self.Flatten()), indent=2, cls=_MetricJsonEncoder) + def ToPerfJson(self) -> List[Dict[str, str]]: + result = [] + for x in sorted(self.Flatten()): + result.append(x.ToPerfJson()) + return result def ToMetricGroupDescriptions(self, root: bool = True) -> Dict[str, str]: result = {self.name: self.description} if self.description else {}
@@ -533,7 +527,23 @@ class MetricGroup: return result def __str__(self) -> str: - return self.ToPerfJson() + return str(self.ToPerfJson()) + + +def JsonEncodeMetric(x: MetricGroup): + class MetricJsonEncoder(json.JSONEncoder): + """Special handling for Metric objects.""" + + def default(self, o): + if isinstance(o, Metric) or isinstance(o, MetricGroup): + return o.ToPerfJson() + return json.JSONEncoder.default(self, o) + + return json.dumps(x, indent=2, cls=MetricJsonEncoder) + + +def JsonEncodeMetricGroupDescriptions(x: MetricGroup): + return json.dumps(x.ToMetricGroupDescriptions(), indent=2) class _RewriteIfExpToSelect(ast.NodeTransformer):
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 11/48 perf jevents: Drop duplicate pending metrics
In reply to Ian RogersDrop adding a pending metric if there is an existing one. Ensure the PMUs differ for hybrid systems.
--- tools/perf/pmu-events/jevents.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
tools/perf/pmu-events/jevents.py +2 -1
@@ -501,7 +501,8 @@ def add_events_table_entries(item: os.DirEntry, topic: str) -> None: for e in read_json_events(item.path, topic): if e.name: _pending_events.append(e) - if e.metric_name: + if e.metric_name and not any(e.metric_name == x.metric_name and + e.pmu == x.pmu for x in _pending_metrics): _pending_metrics.append(e)
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 12/48 perf jevents: Skip optional metrics in metric group list
In reply to Ian RogersFor metric groups, skip metrics in the list that are None. This allows functions to better optionally return metrics.
--- tools/perf/pmu-events/metric.py | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)
tools/perf/pmu-events/metric.py +5 -3
@@ -493,13 +493,15 @@ class MetricGroup: """ def __init__(self, name: str, - metric_list: List[Union[Metric, 'MetricGroup']], + metric_list: List[Union[Optional[Metric], Optional['MetricGroup']]], description: Optional[str] = None): self.name = name - self.metric_list = metric_list + self.metric_list = [] self.description = description for metric in metric_list: - metric.AddToMetricGroup(self) + if metric: + self.metric_list.append(metric) + metric.AddToMetricGroup(self) def AddToMetricGroup(self, group): """Callback used when a MetricGroup is added into another."""
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 13/48 perf jevents: Build support for generating metrics from python
In reply to Ian RogersGenerate extra-metrics.json and extra-metricgroups.json from python architecture specific scripts. The metrics themselves will be added in later patches.
If a build takes place in tools/perf/ then extra-metrics.json and extra-metricgroups.json are generated in that directory and so added to .gitignore. If there is an OUTPUT directory then the tools/perf/pmu-events/arch files are copied to it so the generated extra-metrics.json and extra-metricgroups.json can be added/generated there.
--- tools/perf/.gitignore | 5 +++ tools/perf/Makefile.perf | 2 + tools/perf/pmu-events/Build | 51 +++++++++++++++++++++++++- tools/perf/pmu-events/amd_metrics.py | 42 +++++++++++++++++++++ tools/perf/pmu-events/arm64_metrics.py | 43 ++++++++++++++++++++++ tools/perf/pmu-events/intel_metrics.py | 42 +++++++++++++++++++++ 6 files changed, 184 insertions(+), 1 deletion(-) create mode 100755 tools/perf/pmu-events/amd_metrics.py create mode 100755 tools/perf/pmu-events/arm64_metrics.py create mode 100755 tools/perf/pmu-events/intel_metrics.py
tools/perf/.gitignore +5 -0
@@ -42,6 +42,11 @@ pmu-events/metric_test.log pmu-events/empty-pmu-events.log pmu-events/test-empty-pmu-events.c *.shellcheck_log +pmu-events/arch/**/extra-metrics.json +pmu-events/arch/**/extra-metricgroups.json +tests/shell/*.shellcheck_log +tests/shell/coresight/*.shellcheck_log +tests/shell/lib/*.shellcheck_log feature/ libapi/ libbpf/
tools/perf/Makefile.perf +2 -0
@@ -1277,6 +1277,8 @@ ifeq ($(OUTPUT),) pmu-events/metric_test.log \ pmu-events/test-empty-pmu-events.c \ pmu-events/empty-pmu-events.log + $(Q)find pmu-events/arch -name 'extra-metrics.json' -delete -o \ + -name 'extra-metricgroups.json' -delete else # When an OUTPUT directory is present, clean up the copied pmu-events/arch directory. $(call QUIET_CLEAN, pmu-events) $(RM) -r $(OUTPUT)pmu-events/arch \ $(OUTPUT)pmu-events/pmu-events.c \
tools/perf/pmu-events/Build +50 -1
63 lines not on this page · raw message
tools/perf/pmu-events/amd_metrics.py new file +42 -0
42 lines not on this page · raw message
tools/perf/pmu-events/arm64_metrics.py new file +43 -0
43 lines not on this page · raw message
tools/perf/pmu-events/intel_metrics.py new file +42 -0
42 lines not on this page · raw message
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 14/48 perf jevents: Add load event json to verify and allow fallbacks
In reply to Ian RogersAdd a LoadEvents function that loads all event json files in a directory. In the Event constructor ensure all events are defined in the event json except for legacy events like "cycles". If the initial event isn't found then legacy_event1 is used, and if that isn't found legacy_event2 is used. This allows a single Event to have multiple event names as models will often rename the same event over time. If the event doesn't exist an exception is raised.
So that references to metrics can be added, add the MetricRef class. This doesn't validate as an event name and so provides an escape hatch for metrics to refer to each other.
--- tools/perf/pmu-events/Build | 12 ++-- tools/perf/pmu-events/amd_metrics.py | 7 ++- tools/perf/pmu-events/arm64_metrics.py | 7 ++- tools/perf/pmu-events/intel_metrics.py | 7 ++- tools/perf/pmu-events/metric.py | 83 +++++++++++++++++++++++++- 5 files changed, 101 insertions(+), 15 deletions(-)
tools/perf/pmu-events/Build +6 -6
@@ -53,11 +53,11 @@ ZEN_METRICGROUPS = $(foreach x,$(ZENS),$(OUTPUT)$(x)/extra-metricgroups.json) $(ZEN_METRICS): pmu-events/amd_metrics.py $(GEN_METRIC_DEPS) $(call rule_mkdir) - $(Q)$(call echo-cmd,gen)$(PYTHON) $< $(call model_name,$@) arch > $@ + $(Q)$(call echo-cmd,gen)$(PYTHON) $< $(call model_name,$@) pmu-events/arch > $@ $(ZEN_METRICGROUPS): pmu-events/amd_metrics.py $(GEN_METRIC_DEPS) $(call rule_mkdir) - $(Q)$(call echo-cmd,gen)$(PYTHON) $< -metricgroups $(call model_name,$@) arch > $@ + $(Q)$(call echo-cmd,gen)$(PYTHON) $< -metricgroups $(call model_name,$@) pmu-events/arch > $@ # Generate ARM Json ARMS = $(shell ls -d pmu-events/arch/arm64/arm/*)
@@ -66,11 +66,11 @@ ARM_METRICGROUPS = $(foreach x,$(ARMS),$(OUTPUT)$(x)/extra-metricgroups.json) $(ARM_METRICS): pmu-events/arm64_metrics.py $(GEN_METRIC_DEPS) $(call rule_mkdir) - $(Q)$(call echo-cmd,gen)$(PYTHON) $< $(call vendor_name,$@) $(call model_name,$@) arch > $@ + $(Q)$(call echo-cmd,gen)$(PYTHON) $< $(call vendor_name,$@) $(call model_name,$@) pmu-events/arch > $@ $(ARM_METRICGROUPS): pmu-events/arm64_metrics.py $(GEN_METRIC_DEPS) $(call rule_mkdir) - $(Q)$(call echo-cmd,gen)$(PYTHON) $< -metricgroups $(call vendor_name,$@) $(call model_name,$@) arch > $@ + $(Q)$(call echo-cmd,gen)$(PYTHON) $< -metricgroups $(call vendor_name,$@) $(call model_name,$@) pmu-events/arch > $@ # Generate Intel Json INTELS = $(shell ls -d pmu-events/arch/x86/*|grep -v amdzen|grep -v mapfile.csv)
@@ -79,11 +79,11 @@ INTEL_METRICGROUPS = $(foreach x,$(INTELS),$(OUTPUT)$(x)/extra-metricgroups.json $(INTEL_METRICS): pmu-events/intel_metrics.py $(GEN_METRIC_DEPS) $(call rule_mkdir) - $(Q)$(call echo-cmd,gen)$(PYTHON) $< $(call model_name,$@) arch > $@ + $(Q)$(call echo-cmd,gen)$(PYTHON) $< $(call model_name,$@) pmu-events/arch > $@ $(INTEL_METRICGROUPS): pmu-events/intel_metrics.py $(GEN_METRIC_DEPS) $(call rule_mkdir) - $(Q)$(call echo-cmd,gen)$(PYTHON) $< -metricgroups $(call model_name,$@) arch > $@ + $(Q)$(call echo-cmd,gen)$(PYTHON) $< -metricgroups $(call model_name,$@) pmu-events/arch > $@ GEN_JSON = $(patsubst %,$(OUTPUT)%,$(JSON)) \ $(LEGACY_CACHE_JSON) \
tools/perf/pmu-events/amd_metrics.py +5 -2
@@ -2,8 +2,8 @@ # SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause) import argparse import os -from metric import ( - JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, MetricGroup) +from metric import (JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, LoadEvents, + MetricGroup) # Global command line arguments. _args = None
@@ -30,6 +30,9 @@ def main() -> None: ) _args = parser.parse_args() + directory = f"{_args.events_path}/x86/{_args.model}/" + LoadEvents(directory) + all_metrics = MetricGroup("", []) if _args.metricgroups:
tools/perf/pmu-events/arm64_metrics.py +5 -2
@@ -2,8 +2,8 @@ # SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause) import argparse import os -from metric import ( - JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, MetricGroup) +from metric import (JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, LoadEvents, + MetricGroup) # Global command line arguments. _args = None
@@ -31,6 +31,9 @@ def main() -> None: ) _args = parser.parse_args() + directory = f"{_args.events_path}/arm64/{_args.vendor}/{_args.model}/" + LoadEvents(directory) + all_metrics = MetricGroup("", []) if _args.metricgroups:
tools/perf/pmu-events/intel_metrics.py +5 -2
@@ -2,8 +2,8 @@ # SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause) import argparse import os -from metric import ( - JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, MetricGroup) +from metric import (JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, LoadEvents, + MetricGroup) # Global command line arguments. _args = None
@@ -30,6 +30,9 @@ def main() -> None: ) _args = parser.parse_args() + directory = f"{_args.events_path}/x86/{_args.model}/" + LoadEvents(directory) + all_metrics = MetricGroup("", []) if _args.metricgroups:
tools/perf/pmu-events/metric.py +80 -3
105 lines not on this page · raw message
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 15/48 perf jevents: Add RAPL event metric for AMD zen models
In reply to Ian RogersAdd power per second metrics based on RAPL.
--- tools/perf/pmu-events/amd_metrics.py | 31 +++++++++++++++++++++++++--- 1 file changed, 28 insertions(+), 3 deletions(-)
tools/perf/pmu-events/amd_metrics.py +28 -3
@@ -1,13 +1,36 @@ #!/usr/bin/env python3 # SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause) import argparse +import math import os -from metric import (JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, LoadEvents, - MetricGroup) +from metric import (d_ratio, has_event, Event, JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, + LoadEvents, Metric, MetricGroup, Select) # Global command line arguments. _args = None +interval_sec = Event("duration_time") + + +def Rapl() -> MetricGroup: + """Processor socket power consumption estimate. + + Use events from the running average power limit (RAPL) driver. + """ + # Watts = joules/second + # Currently only energy-pkg is supported by AMD: + # https://lore.kernel.org/lkml/[email protected]/ + pkg = Event("power/energy\\-pkg/") + cond_pkg = Select(pkg, has_event(pkg), math.nan) + scale = 2.3283064365386962890625e-10 + metrics = [ + Metric("lpm_cpu_power_pkg", "", + d_ratio(cond_pkg * scale, interval_sec), "Watts"), + ] + + return MetricGroup("lpm_cpu_power", metrics, + description="Processor socket power consumption estimates") + def main() -> None: global _args
@@ -33,7 +56,9 @@ def main() -> None: directory = f"{_args.events_path}/x86/{_args.model}/" LoadEvents(directory) - all_metrics = MetricGroup("", []) + all_metrics = MetricGroup("", [ + Rapl(), + ]) if _args.metricgroups: print(JsonEncodeMetricGroupDescriptions(all_metrics))
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 16/48 perf jevents: Add idle metric for AMD zen models
In reply to Ian RogersCompute using the msr PMU the percentage of wallclock cycles where the CPUs are in a low power state.
--- tools/perf/pmu-events/amd_metrics.py | 16 ++++++++++++++-- 1 file changed, 14 insertions(+), 2 deletions(-)
tools/perf/pmu-events/amd_metrics.py +14 -2
@@ -3,8 +3,9 @@ import argparse import math import os -from metric import (d_ratio, has_event, Event, JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, - LoadEvents, Metric, MetricGroup, Select) +from metric import (d_ratio, has_event, max, Event, JsonEncodeMetric, + JsonEncodeMetricGroupDescriptions, LoadEvents, Metric, + MetricGroup, Select) # Global command line arguments. _args = None
@@ -12,6 +13,16 @@ _args = None interval_sec = Event("duration_time") +def Idle() -> Metric: + cyc = Event("msr/mperf/") + tsc = Event("msr/tsc/") + low = max(tsc - cyc, 0) + return Metric( + "lpm_idle", + "Percentage of total wallclock cycles where CPUs are in low power state (C1 or deeper sleep state)", + d_ratio(low, tsc), "100%") + + def Rapl() -> MetricGroup: """Processor socket power consumption estimate.
@@ -57,6 +68,7 @@ def main() -> None: LoadEvents(directory) all_metrics = MetricGroup("", [ + Idle(), Rapl(), ])
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 17/48 perf jevents: Add upc metric for uops per cycle for AMD
In reply to Ian RogersThe metric adjusts for whether or not SMT is on.
--- tools/perf/pmu-events/amd_metrics.py | 22 +++++++++++++++++++--- 1 file changed, 19 insertions(+), 3 deletions(-)
tools/perf/pmu-events/amd_metrics.py +19 -3
@@ -3,14 +3,26 @@ import argparse import math import os +from typing import Optional from metric import (d_ratio, has_event, max, Event, JsonEncodeMetric, - JsonEncodeMetricGroupDescriptions, LoadEvents, Metric, - MetricGroup, Select) + JsonEncodeMetricGroupDescriptions, Literal, LoadEvents, + Metric, MetricGroup, Select) # Global command line arguments. _args = None - +_zen_model: int = 1 interval_sec = Event("duration_time") +ins = Event("instructions") +cycles = Event("cycles") +# Number of CPU cycles scaled for SMT. +smt_cycles = Select(cycles / 2, Literal("#smt_on"), cycles) + + +def AmdUpc() -> Metric: + ops = Event("ex_ret_ops", "ex_ret_cops") + upc = d_ratio(ops, smt_cycles) + return Metric("lpm_upc", "Micro-ops retired per core cycle (higher is better)", + upc, "uops/cycle") def Idle() -> Metric:
@@ -45,6 +57,7 @@ def Rapl() -> MetricGroup: def main() -> None: global _args + global _zen_model def dir_path(path: str) -> str: """Validate path is a directory for argparse."""
@@ -67,7 +80,10 @@ def main() -> None: directory = f"{_args.events_path}/x86/{_args.model}/" LoadEvents(directory) + _zen_model = int(_args.model[6:]) + all_metrics = MetricGroup("", [ + AmdUpc(), Idle(), Rapl(), ])
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Re: [PATCH v9 17/48] perf jevents: Add upc metric for uops per cycle for AMD
Quoted 22 lines · Ian Rogers
The metric adjusts for whether or not SMT is on.
Signed-off-by Ian Rogers <[email protected]>Reviewed-by Sandipan Das <[email protected]>--- tools/perf/pmu-events/amd_metrics.py | 22 +++++++++++++++++++--- 1 file changed, 19 insertions(+), 3 deletions(-)tools/perf/pmu-events/amd_metrics.py +3
-2@@ -3,14 +3,26 @@ import argparse import math import os +from typing import Optional from metric import (d_ratio, has_event, max, Event, JsonEncodeMetric, - JsonEncodeMetricGroupDescriptions, LoadEvents, Metric, - MetricGroup, Select) + JsonEncodeMetricGroupDescriptions, Literal, LoadEvents, + Metric, MetricGroup, Select)
tools/perf/pmu-events/amd_metrics.py line 18
Quoted 10 lines
tools/perf/pmu-events/amd_metrics.py +5
-1# Global command line arguments. _args = None - +_zen_model: int = 1 interval_sec = Event("duration_time") +ins = Event("instructions") +cycles = Event("cycles") +# Number of CPU cycles scaled for SMT. +smt_cycles = Select(cycles / 2, Literal("#smt_on"), cycles)
The CPU architects said that while there are some fairness guarantees, the distribution of resources amongst sibling threads is not always equal. There is also no easy way to determine this.
tools/perf/pmu-events/amd_metrics.py line 25
Quoted 7 lines
tools/perf/pmu-events/amd_metrics.py +7
-0+ + +def AmdUpc() -> Metric: + ops = Event("ex_ret_ops", "ex_ret_cops") + upc = d_ratio(ops, smt_cycles) + return Metric("lpm_upc", "Micro-ops retired per core cycle (higher is better)", + upc, "uops/cycle")
Zen 3 onwards, PMCx0C1, which is used by ex_ret_ops and ex_ret_cops, counts retired macro-ops.
Quoted 22 lines
tools/perf/pmu-events/amd_metrics.py +4
-0def Idle() -> Metric:@@ -45,6 +57,7 @@ def Rapl() -> MetricGroup: def main() -> None: global _args + global _zen_model def dir_path(path: str) -> str: """Validate path is a directory for argparse."""@@ -67,7 +80,10 @@ def main() -> None: directory = f"{_args.events_path}/x86/{_args.model}/" LoadEvents(directory) + _zen_model = int(_args.model[6:]) + all_metrics = MetricGroup("", [ + AmdUpc(), Idle(), Rapl(), ])
Ian Rogers
reviewerPATCH v9 18/48 perf jevents: Add br metric group for branch statistics on AMD
In reply to Ian RogersThe br metric group for branches itself comprises metric groups for total, taken, conditional, fused and far metric groups using json events. The lack of conditional events on anything but zen2 means this category is lacking on zen1, zen3 and zen4.
--- tools/perf/pmu-events/amd_metrics.py | 104 +++++++++++++++++++++++++++ 1 file changed, 104 insertions(+)
tools/perf/pmu-events/amd_metrics.py +104 -0
@@ -18,6 +18,109 @@ cycles = Event("cycles") smt_cycles = Select(cycles / 2, Literal("#smt_on"), cycles) +def AmdBr(): + def Total() -> MetricGroup: + br = Event("ex_ret_brn") + br_m_all = Event("ex_ret_brn_misp") + br_clr = Event("ex_ret_msprd_brnch_instr_dir_msmtch", + "ex_ret_brn_resync") + + br_r = d_ratio(br, interval_sec) + ins_r = d_ratio(ins, br) + misp_r = d_ratio(br_m_all, br) + clr_r = d_ratio(br_clr, interval_sec) + + return MetricGroup("lpm_br_total", [ + Metric("lpm_br_total_retired", + "The number of branch instructions retired per second.", br_r, + "insn/s"), + Metric( + "lpm_br_total_mispred", + "The number of branch instructions retired, of any type, that were " + "not correctly predicted as a percentage of all branch instrucions.", + misp_r, "100%"), + Metric("lpm_br_total_insn_between_branches", + "The number of instructions divided by the number of branches.", + ins_r, "insn"), + Metric("lpm_br_total_insn_fe_resteers", + "The number of resync branches per second.", clr_r, "req/s") + ]) + + def Taken() -> MetricGroup: + br = Event("ex_ret_brn_tkn") + br_m_tk = Event("ex_ret_brn_tkn_misp") + br_r = d_ratio(br, interval_sec) + ins_r = d_ratio(ins, br) + misp_r = d_ratio(br_m_tk, br) + return MetricGroup("lpm_br_taken", [ + Metric("lpm_br_taken_retired", + "The number of taken branches that were retired per second.", + br_r, "insn/s"), + Metric( + "lpm_br_taken_mispred", + "The number of retired taken branch instructions that were " + "mispredicted as a percentage of all taken branches.", misp_r, + "100%"), + Metric( + "lpm_br_taken_insn_between_branches", + "The number of instructions divided by the number of taken branches.", + ins_r, "insn"), + ]) + + def Conditional() -> Optional[MetricGroup]: + global _zen_model + br = Event("ex_ret_cond") + br_r = d_ratio(br, interval_sec) + ins_r = d_ratio(ins, br) + + metrics = [ + Metric("lpm_br_cond_retired", "Retired conditional branch instructions.", + br_r, "insn/s"), + Metric("lpm_br_cond_insn_between_branches", + "The number of instructions divided by the number of conditional " + "branches.", ins_r, "insn"), + ] + if _zen_model == 2: + br_m_cond = Event("ex_ret_cond_misp") + misp_r = d_ratio(br_m_cond, br) + metrics += [ + Metric("lpm_br_cond_mispred", + "Retired conditional branch instructions mispredicted as a " + "percentage of all conditional branches.", misp_r, "100%"), + ] + + return MetricGroup("lpm_br_cond", metrics) + + def Fused() -> MetricGroup: + br = Event("ex_ret_fused_instr", "ex_ret_fus_brnch_inst") + br_r = d_ratio(br, interval_sec) + ins_r = d_ratio(ins, br) + return MetricGroup("lpm_br_cond", [ + Metric("lpm_br_fused_retired", + "Retired fused branch instructions per second.", br_r, "insn/s"), + Metric( + "lpm_br_fused_insn_between_branches", + "The number of instructions divided by the number of fused " + "branches.", ins_r, "insn"), + ]) + + def Far() -> MetricGroup: + br = Event("ex_ret_brn_far") + br_r = d_ratio(br, interval_sec) + ins_r = d_ratio(ins, br) + return MetricGroup("lpm_br_far", [ + Metric("lpm_br_far_retired", "Retired far control transfers per second.", + br_r, "insn/s"), + Metric( + "lpm_br_far_insn_between_branches", + "The number of instructions divided by the number of far branches.", + ins_r, "insn"), + ]) + + return MetricGroup("lpm_br", [Total(), Taken(), Conditional(), Fused(), Far()], + description="breakdown of retired branch instructions") + + def AmdUpc() -> Metric: ops = Event("ex_ret_ops", "ex_ret_cops") upc = d_ratio(ops, smt_cycles)
@@ -83,6 +186,7 @@ def main() -> None: _zen_model = int(_args.model[6:]) all_metrics = MetricGroup("", [ + AmdBr(), AmdUpc(), Idle(), Rapl(),
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Re: [PATCH v9 18/48] perf jevents: Add br metric group for branch statistics on AMD
Quoted 16 lines · Ian Rogers
The br metric group for branches itself comprises metric groups for total, taken, conditional, fused and far metric groups using json events. The lack of conditional events on anything but zen2 means this category is lacking on zen1, zen3 and zen4.
Signed-off-by Ian Rogers <[email protected]>Reviewed-by Sandipan Das <[email protected]>--- tools/perf/pmu-events/amd_metrics.py | 104 +++++++++++++++++++++++++++ 1 file changed, 104 insertions(+)tools/perf/pmu-events/amd_metrics.py +0
-0
tools/perf/pmu-events/amd_metrics.py cycles = Event("cycles") · line 26
Quoted 9 lines
tools/perf/pmu-events/amd_metrics.py +6
-0@@ -18,6 +18,109 @@ cycles = Event("cycles") smt_cycles = Select(cycles / 2, Literal("#smt_on"), cycles) +def AmdBr(): + def Total() -> MetricGroup: + br = Event("ex_ret_brn") + br_m_all = Event("ex_ret_brn_misp") + br_clr = Event("ex_ret_msprd_brnch_instr_dir_msmtch", + "ex_ret_brn_resync")
PMCx1C7, used by ex_ret_msprd_brnch_instr_dir_msmtch, and PMCx0C7, used by ex_ret_brn_resync, are not equivalent. I have been told that the now deprecated PMCx0C7 was meant to count pipeline restarts for reasons other than branch misprediction. Is the intention here to count such restarts or just the ones caused by branch misprediction or both?
Quoted 108 lines
tools/perf/pmu-events/amd_metrics.py +98
-0+ + br_r = d_ratio(br, interval_sec) + ins_r = d_ratio(ins, br) + misp_r = d_ratio(br_m_all, br) + clr_r = d_ratio(br_clr, interval_sec) + + return MetricGroup("lpm_br_total", [ + Metric("lpm_br_total_retired", + "The number of branch instructions retired per second.", br_r, + "insn/s"), + Metric( + "lpm_br_total_mispred", + "The number of branch instructions retired, of any type, that were " + "not correctly predicted as a percentage of all branch instrucions.", + misp_r, "100%"), + Metric("lpm_br_total_insn_between_branches", + "The number of instructions divided by the number of branches.", + ins_r, "insn"), + Metric("lpm_br_total_insn_fe_resteers", + "The number of resync branches per second.", clr_r, "req/s") + ]) + + def Taken() -> MetricGroup: + br = Event("ex_ret_brn_tkn") + br_m_tk = Event("ex_ret_brn_tkn_misp") + br_r = d_ratio(br, interval_sec) + ins_r = d_ratio(ins, br) + misp_r = d_ratio(br_m_tk, br) + return MetricGroup("lpm_br_taken", [ + Metric("lpm_br_taken_retired", + "The number of taken branches that were retired per second.", + br_r, "insn/s"), + Metric( + "lpm_br_taken_mispred", + "The number of retired taken branch instructions that were " + "mispredicted as a percentage of all taken branches.", misp_r, + "100%"), + Metric( + "lpm_br_taken_insn_between_branches", + "The number of instructions divided by the number of taken branches.", + ins_r, "insn"), + ]) + + def Conditional() -> Optional[MetricGroup]: + global _zen_model + br = Event("ex_ret_cond") + br_r = d_ratio(br, interval_sec) + ins_r = d_ratio(ins, br) + + metrics = [ + Metric("lpm_br_cond_retired", "Retired conditional branch instructions.", + br_r, "insn/s"), + Metric("lpm_br_cond_insn_between_branches", + "The number of instructions divided by the number of conditional " + "branches.", ins_r, "insn"), + ] + if _zen_model == 2: + br_m_cond = Event("ex_ret_cond_misp") + misp_r = d_ratio(br_m_cond, br) + metrics += [ + Metric("lpm_br_cond_mispred", + "Retired conditional branch instructions mispredicted as a " + "percentage of all conditional branches.", misp_r, "100%"), + ] + + return MetricGroup("lpm_br_cond", metrics) + + def Fused() -> MetricGroup: + br = Event("ex_ret_fused_instr", "ex_ret_fus_brnch_inst") + br_r = d_ratio(br, interval_sec) + ins_r = d_ratio(ins, br) + return MetricGroup("lpm_br_cond", [ + Metric("lpm_br_fused_retired", + "Retired fused branch instructions per second.", br_r, "insn/s"), + Metric( + "lpm_br_fused_insn_between_branches", + "The number of instructions divided by the number of fused " + "branches.", ins_r, "insn"), + ]) + + def Far() -> MetricGroup: + br = Event("ex_ret_brn_far") + br_r = d_ratio(br, interval_sec) + ins_r = d_ratio(ins, br) + return MetricGroup("lpm_br_far", [ + Metric("lpm_br_far_retired", "Retired far control transfers per second.", + br_r, "insn/s"), + Metric( + "lpm_br_far_insn_between_branches", + "The number of instructions divided by the number of far branches.", + ins_r, "insn"), + ]) + + return MetricGroup("lpm_br", [Total(), Taken(), Conditional(), Fused(), Far()], + description="breakdown of retired branch instructions") + + def AmdUpc() -> Metric: ops = Event("ex_ret_ops", "ex_ret_cops") upc = d_ratio(ops, smt_cycles)@@ -83,6 +186,7 @@ def main() -> None: _zen_model = int(_args.model[6:]) all_metrics = MetricGroup("", [ + AmdBr(), AmdUpc(), Idle(), Rapl(),
Ian Rogers
reviewerPATCH v9 19/48 perf jevents: Add itlb metric group for AMD
In reply to Ian RogersAdd metrics that give an overview and details of the l1 itlb (zen1, zen2, zen3) and l2 itlb (all zens).
--- tools/perf/pmu-events/amd_metrics.py | 49 ++++++++++++++++++++++++++++ 1 file changed, 49 insertions(+)
tools/perf/pmu-events/amd_metrics.py +49 -0
@@ -121,6 +121,54 @@ def AmdBr(): description="breakdown of retired branch instructions") +def AmdItlb(): + global _zen_model + l2h = Event("bp_l1_tlb_miss_l2_tlb_hit", "bp_l1_tlb_miss_l2_hit") + l2m = Event("l2_itlb_misses") + l2r = l2h + l2m + + itlb_l1_mg = None + l1m = l2r + if _zen_model <= 3: + l1r = Event("ic_fw32") + l1h = max(l1r - l1m, 0) + itlb_l1_mg = MetricGroup("lpm_itlb_l1", [ + Metric("lpm_itlb_l1_hits", + "L1 ITLB hits as a perecentage of L1 ITLB accesses.", + d_ratio(l1h, l1h + l1m), "100%"), + Metric("lpm_itlb_l1_miss", + "L1 ITLB misses as a perecentage of L1 ITLB accesses.", + d_ratio(l1m, l1h + l1m), "100%"), + Metric("lpm_itlb_l1_reqs", + "The number of 32B fetch windows transferred from IC pipe to DE " + "instruction decoder per second.", d_ratio( + l1r, interval_sec), + "windows/sec"), + ]) + + return MetricGroup("lpm_itlb", [ + MetricGroup("lpm_itlb_ov", [ + Metric("lpm_itlb_ov_insn_bt_l1_miss", + "Number of instructions between l1 misses", d_ratio( + ins, l1m), "insns"), + Metric("lpm_itlb_ov_insn_bt_l2_miss", + "Number of instructions between l2 misses", d_ratio( + ins, l2m), "insns"), + ]), + itlb_l1_mg, + MetricGroup("lpm_itlb_l2", [ + Metric("lpm_itlb_l2_hits", + "L2 ITLB hits as a percentage of all L2 ITLB accesses.", + d_ratio(l2h, l2r), "100%"), + Metric("lpm_itlb_l2_miss", + "L2 ITLB misses as a percentage of all L2 ITLB accesses.", + d_ratio(l2m, l2r), "100%"), + Metric("lpm_itlb_l2_reqs", "ITLB accesses per second.", + d_ratio(l2r, interval_sec), "accesses/sec"), + ]), + ], description="Instruction TLB breakdown") + + def AmdUpc() -> Metric: ops = Event("ex_ret_ops", "ex_ret_cops") upc = d_ratio(ops, smt_cycles)
@@ -187,6 +235,7 @@ def main() -> None: all_metrics = MetricGroup("", [ AmdBr(), + AmdItlb(), AmdUpc(), Idle(), Rapl(),
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 20/48 perf jevents: Add dtlb metric group for AMD
In reply to Ian RogersAdd metrics that give an overview and details of the dtlb (zen1, zen2, zen3).
--- tools/perf/pmu-events/amd_metrics.py | 111 +++++++++++++++++++++++++++ 1 file changed, 111 insertions(+)
tools/perf/pmu-events/amd_metrics.py +111 -0
@@ -121,6 +121,116 @@ def AmdBr(): description="breakdown of retired branch instructions") +def AmdDtlb() -> Optional[MetricGroup]: + global _zen_model + if _zen_model >= 4: + return None + + d_dat = Event("ls_dc_accesses") if _zen_model <= 3 else None + d_h4k = Event("ls_l1_d_tlb_miss.tlb_reload_4k_l2_hit") + d_hcoal = Event( + "ls_l1_d_tlb_miss.tlb_reload_coalesced_page_hit") if _zen_model >= 2 else 0 + d_h2m = Event("ls_l1_d_tlb_miss.tlb_reload_2m_l2_hit") + d_h1g = Event("ls_l1_d_tlb_miss.tlb_reload_1g_l2_hit") + + d_m4k = Event("ls_l1_d_tlb_miss.tlb_reload_4k_l2_miss") + d_mcoal = Event( + "ls_l1_d_tlb_miss.tlb_reload_coalesced_page_miss") if _zen_model >= 2 else 0 + d_m2m = Event("ls_l1_d_tlb_miss.tlb_reload_2m_l2_miss") + d_m1g = Event("ls_l1_d_tlb_miss.tlb_reload_1g_l2_miss") + + d_w0 = Event("ls_tablewalker.dc_type0") if _zen_model <= 3 else None + d_w1 = Event("ls_tablewalker.dc_type1") if _zen_model <= 3 else None + walks = d_w0 + d_w1 + walks_r = d_ratio(walks, interval_sec) + ins_w = d_ratio(ins, walks) + l1 = d_dat + l1_r = d_ratio(l1, interval_sec) + l2_hits = d_h4k + d_hcoal + d_h2m + d_h1g + l2_miss = d_m4k + d_mcoal + d_m2m + d_m1g + l2_r = d_ratio(l2_hits + l2_miss, interval_sec) + l1_miss = l2_hits + l2_miss + walks + l1_hits = max(l1 - l1_miss, 0) + ins_l = d_ratio(ins, l1_miss) + + return MetricGroup("lpm_dtlb", [ + MetricGroup("lpm_dtlb_ov", [ + Metric("lpm_dtlb_ov_insn_bt_l1_miss", + "DTLB overview: instructions between l1 misses.", ins_l, + "insns"), + Metric("lpm_dtlb_ov_insn_bt_walks", + "DTLB overview: instructions between dtlb page table walks.", + ins_w, "insns"), + ]), + MetricGroup("lpm_dtlb_l1", [ + Metric("lpm_dtlb_l1_hits", + "DTLB L1 hits as percentage of all DTLB L1 accesses.", + d_ratio(l1_hits, l1), "100%"), + Metric("lpm_dtlb_l1_miss", + "DTLB L1 misses as percentage of all DTLB L1 accesses.", + d_ratio(l1_miss, l1), "100%"), + Metric("lpm_dtlb_l1_reqs", "DTLB L1 accesses per second.", l1_r, + "insns/s"), + ]), + MetricGroup("lpm_dtlb_l2", [ + Metric("lpm_dtlb_l2_hits", + "DTLB L2 hits as percentage of all DTLB L2 accesses.", + d_ratio(l2_hits, l2_hits + l2_miss), "100%"), + Metric("lpm_dtlb_l2_miss", + "DTLB L2 misses as percentage of all DTLB L2 accesses.", + d_ratio(l2_miss, l2_hits + l2_miss), "100%"), + Metric("lpm_dtlb_l2_reqs", "DTLB L2 accesses per second.", l2_r, + "insns/s"), + MetricGroup("lpm_dtlb_l2_4kb", [ + Metric( + "lpm_dtlb_l2_4kb_hits", + "DTLB L2 4kb page size hits as percentage of all DTLB L2 4kb " + "accesses.", d_ratio(d_h4k, d_h4k + d_m4k), "100%"), + Metric( + "lpm_dtlb_l2_4kb_miss", + "DTLB L2 4kb page size misses as percentage of all DTLB L2 4kb" + "accesses.", d_ratio(d_m4k, d_h4k + d_m4k), "100%") + ]), + MetricGroup("lpm_dtlb_l2_coalesced", [ + Metric( + "lpm_dtlb_l2_coal_hits", + "DTLB L2 coalesced page (16kb) hits as percentage of all DTLB " + "L2 coalesced accesses.", d_ratio(d_hcoal, + d_hcoal + d_mcoal), "100%"), + Metric( + "lpm_dtlb_l2_coal_miss", + "DTLB L2 coalesced page (16kb) misses as percentage of all " + "DTLB L2 coalesced accesses.", + d_ratio(d_mcoal, d_hcoal + d_mcoal), "100%") + ]), + MetricGroup("lpm_dtlb_l2_2mb", [ + Metric( + "lpm_dtlb_l2_2mb_hits", + "DTLB L2 2mb page size hits as percentage of all DTLB L2 2mb " + "accesses.", d_ratio(d_h2m, d_h2m + d_m2m), "100%"), + Metric( + "lpm_dtlb_l2_2mb_miss", + "DTLB L2 2mb page size misses as percentage of all DTLB L2 " + "accesses.", d_ratio(d_m2m, d_h2m + d_m2m), "100%") + ]), + MetricGroup("lpm_dtlb_l2_1g", [ + Metric( + "lpm_dtlb_l2_1g_hits", + "DTLB L2 1gb page size hits as percentage of all DTLB L2 1gb " + "accesses.", d_ratio(d_h1g, d_h1g + d_m1g), "100%"), + Metric( + "lpm_dtlb_l2_1g_miss", + "DTLB L2 1gb page size misses as percentage of all DTLB L2 " + "1gb accesses.", d_ratio(d_m1g, d_h1g + d_m1g), "100%") + ]), + ]), + MetricGroup("lpm_dtlb_walks", [ + Metric("lpm_dtlb_walks_reqs", "DTLB page table walks per second.", + walks_r, "walks/s"), + ]), + ], description="Data TLB metrics") + + def AmdItlb(): global _zen_model l2h = Event("bp_l1_tlb_miss_l2_tlb_hit", "bp_l1_tlb_miss_l2_hit")
@@ -235,6 +345,7 @@ def main() -> None: all_metrics = MetricGroup("", [ AmdBr(), + AmdDtlb(), AmdItlb(), AmdUpc(), Idle(),
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 21/48 perf jevents: Add uncore l3 metric group for AMD
In reply to Ian RogersMetrics use the amd_l3 PMU for access/miss/hit information.
--- tools/perf/pmu-events/amd_metrics.py | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+)
tools/perf/pmu-events/amd_metrics.py +19 -0
@@ -316,6 +316,24 @@ def Rapl() -> MetricGroup: description="Processor socket power consumption estimates") +def UncoreL3(): + acc = Event("l3_lookup_state.all_coherent_accesses_to_l3", + "l3_lookup_state.all_l3_req_typs") + miss = Event("l3_lookup_state.l3_miss", + "l3_comb_clstr_state.request_miss") + acc = max(acc, miss) + hits = acc - miss + + return MetricGroup("lpm_l3", [ + Metric("lpm_l3_accesses", "L3 victim cache accesses", + d_ratio(acc, interval_sec), "accesses/sec"), + Metric("lpm_l3_hits", "L3 victim cache hit rate", + d_ratio(hits, acc), "100%"), + Metric("lpm_l3_miss", "L3 victim cache miss rate", d_ratio(miss, acc), + "100%"), + ], description="L3 cache breakdown per CCX") + + def main() -> None: global _args global _zen_model
@@ -350,6 +368,7 @@ def main() -> None: AmdUpc(), Idle(), Rapl(), + UncoreL3(), ]) if _args.metricgroups:
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 22/48 perf jevents: Add load store breakdown metrics ldst for AMD
In reply to Ian RogersGive breakdown of number of instructions. Use the counter mask (cmask) to show the number of cycles taken to retire the instructions.
--- tools/perf/pmu-events/amd_metrics.py | 75 ++++++++++++++++++++++++++++ 1 file changed, 75 insertions(+)
tools/perf/pmu-events/amd_metrics.py +75 -0
@@ -279,6 +279,80 @@ def AmdItlb(): ], description="Instruction TLB breakdown") +def AmdLdSt() -> MetricGroup: + ldst_ld = Event("ls_dispatch.ld_dispatch") + ldst_st = Event("ls_dispatch.store_dispatch") + ldst_ldc1 = Event(f"{ldst_ld}/cmask=1/") + ldst_stc1 = Event(f"{ldst_st}/cmask=1/") + ldst_ldc2 = Event(f"{ldst_ld}/cmask=2/") + ldst_stc2 = Event(f"{ldst_st}/cmask=2/") + ldst_ldc3 = Event(f"{ldst_ld}/cmask=3/") + ldst_stc3 = Event(f"{ldst_st}/cmask=3/") + ldst_cyc = Event("ls_not_halted_cyc") + + ld_rate = d_ratio(ldst_ld, interval_sec) + st_rate = d_ratio(ldst_st, interval_sec) + + ld_v1 = max(ldst_ldc1 - ldst_ldc2, 0) + ld_v2 = max(ldst_ldc2 - ldst_ldc3, 0) + ld_v3 = ldst_ldc3 + + st_v1 = max(ldst_stc1 - ldst_stc2, 0) + st_v2 = max(ldst_stc2 - ldst_stc3, 0) + st_v3 = ldst_stc3 + + return MetricGroup("lpm_ldst", [ + MetricGroup("lpm_ldst_total", [ + Metric("lpm_ldst_total_ld", "Number of loads dispatched per second.", + ld_rate, "insns/sec"), + Metric("lpm_ldst_total_st", "Number of stores dispatched per second.", + st_rate, "insns/sec"), + ]), + MetricGroup("lpm_ldst_percent_insn", [ + Metric("lpm_ldst_percent_insn_ld", + "Load instructions as a percentage of all instructions.", + d_ratio(ldst_ld, ins), "100%"), + Metric("lpm_ldst_percent_insn_st", + "Store instructions as a percentage of all instructions.", + d_ratio(ldst_st, ins), "100%"), + ]), + MetricGroup("lpm_ldst_ret_loads_per_cycle", [ + Metric( + "lpm_ldst_ret_loads_per_cycle_1", + "Load instructions retiring in 1 cycle as a percentage of all " + "unhalted cycles.", d_ratio(ld_v1, ldst_cyc), "100%"), + Metric( + "lpm_ldst_ret_loads_per_cycle_2", + "Load instructions retiring in 2 cycles as a percentage of all " + "unhalted cycles.", d_ratio(ld_v2, ldst_cyc), "100%"), + Metric( + "lpm_ldst_ret_loads_per_cycle_3", + "Load instructions retiring in 3 or more cycles as a percentage" + "of all unhalted cycles.", d_ratio(ld_v3, ldst_cyc), "100%"), + ]), + MetricGroup("lpm_ldst_ret_stores_per_cycle", [ + Metric( + "lpm_ldst_ret_stores_per_cycle_1", + "Store instructions retiring in 1 cycle as a percentage of all " + "unhalted cycles.", d_ratio(st_v1, ldst_cyc), "100%"), + Metric( + "lpm_ldst_ret_stores_per_cycle_2", + "Store instructions retiring in 2 cycles as a percentage of all " + "unhalted cycles.", d_ratio(st_v2, ldst_cyc), "100%"), + Metric( + "lpm_ldst_ret_stores_per_cycle_3", + "Store instructions retiring in 3 or more cycles as a percentage" + "of all unhalted cycles.", d_ratio(st_v3, ldst_cyc), "100%"), + ]), + MetricGroup("lpm_ldst_insn_bt", [ + Metric("lpm_ldst_insn_bt_ld", "Number of instructions between loads.", + d_ratio(ins, ldst_ld), "insns"), + Metric("lpm_ldst_insn_bt_st", "Number of instructions between stores.", + d_ratio(ins, ldst_st), "insns"), + ]) + ], description="Breakdown of load/store instructions") + + def AmdUpc() -> Metric: ops = Event("ex_ret_ops", "ex_ret_cops") upc = d_ratio(ops, smt_cycles)
@@ -365,6 +439,7 @@ def main() -> None: AmdBr(), AmdDtlb(), AmdItlb(), + AmdLdSt(), AmdUpc(), Idle(), Rapl(),
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Re: [PATCH v9 22/48] perf jevents: Add load store breakdown metrics ldst for AMD
Quoted 72 lines · Ian Rogers
Give breakdown of number of instructions. Use the counter mask (cmask) to show the number of cycles taken to retire the instructions.
Signed-off-by Ian Rogers <[email protected]>Reviewed-by Sandipan Das <[email protected]>--- tools/perf/pmu-events/amd_metrics.py | 75 ++++++++++++++++++++++++++++ 1 file changed, 75 insertions(+)tools/perf/pmu-events/amd_metrics.py +55
-0@@ -279,6 +279,80 @@ def AmdItlb(): ], description="Instruction TLB breakdown") +def AmdLdSt() -> MetricGroup: + ldst_ld = Event("ls_dispatch.ld_dispatch") + ldst_st = Event("ls_dispatch.store_dispatch") + ldst_ldc1 = Event(f"{ldst_ld}/cmask=1/") + ldst_stc1 = Event(f"{ldst_st}/cmask=1/") + ldst_ldc2 = Event(f"{ldst_ld}/cmask=2/") + ldst_stc2 = Event(f"{ldst_st}/cmask=2/") + ldst_ldc3 = Event(f"{ldst_ld}/cmask=3/") + ldst_stc3 = Event(f"{ldst_st}/cmask=3/") + ldst_cyc = Event("ls_not_halted_cyc") + + ld_rate = d_ratio(ldst_ld, interval_sec) + st_rate = d_ratio(ldst_st, interval_sec) + + ld_v1 = max(ldst_ldc1 - ldst_ldc2, 0) + ld_v2 = max(ldst_ldc2 - ldst_ldc3, 0) + ld_v3 = ldst_ldc3 + + st_v1 = max(ldst_stc1 - ldst_stc2, 0) + st_v2 = max(ldst_stc2 - ldst_stc3, 0) + st_v3 = ldst_stc3 + + return MetricGroup("lpm_ldst", [ + MetricGroup("lpm_ldst_total", [ + Metric("lpm_ldst_total_ld", "Number of loads dispatched per second.", + ld_rate, "insns/sec"), + Metric("lpm_ldst_total_st", "Number of stores dispatched per second.", + st_rate, "insns/sec"), + ]), + MetricGroup("lpm_ldst_percent_insn", [ + Metric("lpm_ldst_percent_insn_ld", + "Load instructions as a percentage of all instructions.", + d_ratio(ldst_ld, ins), "100%"), + Metric("lpm_ldst_percent_insn_st", + "Store instructions as a percentage of all instructions.", + d_ratio(ldst_st, ins), "100%"), + ]), + MetricGroup("lpm_ldst_ret_loads_per_cycle", [ + Metric( + "lpm_ldst_ret_loads_per_cycle_1", + "Load instructions retiring in 1 cycle as a percentage of all " + "unhalted cycles.", d_ratio(ld_v1, ldst_cyc), "100%"), + Metric( + "lpm_ldst_ret_loads_per_cycle_2", + "Load instructions retiring in 2 cycles as a percentage of all " + "unhalted cycles.", d_ratio(ld_v2, ldst_cyc), "100%"), + Metric( + "lpm_ldst_ret_loads_per_cycle_3", + "Load instructions retiring in 3 or more cycles as a percentage" + "of all unhalted cycles.", d_ratio(ld_v3, ldst_cyc), "100%"), + ]), + MetricGroup("lpm_ldst_ret_stores_per_cycle", [ + Metric( + "lpm_ldst_ret_stores_per_cycle_1", + "Store instructions retiring in 1 cycle as a percentage of all "
tools/perf/pmu-events/amd_metrics.py def AmdItlb(): · line 346
Quoted 10 lines
tools/perf/pmu-events/amd_metrics.py +10
-0+ "unhalted cycles.", d_ratio(st_v1, ldst_cyc), "100%"), + Metric( + "lpm_ldst_ret_stores_per_cycle_2", + "Store instructions retiring in 2 cycles as a percentage of all " + "unhalted cycles.", d_ratio(st_v2, ldst_cyc), "100%"), + Metric( + "lpm_ldst_ret_stores_per_cycle_3", + "Store instructions retiring in 3 or more cycles as a percentage" + "of all unhalted cycles.", d_ratio(st_v3, ldst_cyc), "100%"), + ]),
A subset of dispatched loads and stores do not retire. So PMCx029, which is used by ldst_ld and ldst_st above, does not provide the number of retired loads and stores. There is currently no event which counts retired loads and stores :(
Quoted 20 lines
tools/perf/pmu-events/amd_metrics.py +10
-0+ MetricGroup("lpm_ldst_insn_bt", [ + Metric("lpm_ldst_insn_bt_ld", "Number of instructions between loads.", + d_ratio(ins, ldst_ld), "insns"), + Metric("lpm_ldst_insn_bt_st", "Number of instructions between stores.", + d_ratio(ins, ldst_st), "insns"), + ]) + ], description="Breakdown of load/store instructions") + + def AmdUpc() -> Metric: ops = Event("ex_ret_ops", "ex_ret_cops") upc = d_ratio(ops, smt_cycles)@@ -365,6 +439,7 @@ def main() -> None: AmdBr(), AmdDtlb(), AmdItlb(), + AmdLdSt(), AmdUpc(), Idle(), Rapl(),
Ian Rogers
reviewerPATCH v9 23/48 perf jevents: Add context switch metrics for AMD
In reply to Ian RogersMetrics break down context switches for different kinds of instruction.
--- tools/perf/pmu-events/amd_metrics.py | 33 ++++++++++++++++++++++++++++ 1 file changed, 33 insertions(+)
tools/perf/pmu-events/amd_metrics.py +33 -0
@@ -121,6 +121,38 @@ def AmdBr(): description="breakdown of retired branch instructions") +def AmdCtxSw() -> MetricGroup: + cs = Event("context\\-switches") + metrics = [ + Metric("lpm_cs_rate", "Context switches per second", + d_ratio(cs, interval_sec), "ctxsw/s") + ] + + ev = Event("instructions") + metrics.append(Metric("lpm_cs_instr", "Instructions per context switch", + d_ratio(ev, cs), "instr/cs")) + + ev = Event("cycles") + metrics.append(Metric("lpm_cs_cycles", "Cycles per context switch", + d_ratio(ev, cs), "cycles/cs")) + + ev = Event("ls_dispatch.ld_dispatch") + metrics.append(Metric("lpm_cs_loads", "Loads per context switch", + d_ratio(ev, cs), "loads/cs")) + + ev = Event("ls_dispatch.store_dispatch") + metrics.append(Metric("lpm_cs_stores", "Stores per context switch", + d_ratio(ev, cs), "stores/cs")) + + ev = Event("ex_ret_brn_tkn") + metrics.append(Metric("lpm_cs_br_taken", "Branches taken per context switch", + d_ratio(ev, cs), "br_taken/cs")) + + return MetricGroup("lpm_cs", metrics, + description=("Number of context switches per second, instructions " + "retired & core cycles between context switches")) + + def AmdDtlb() -> Optional[MetricGroup]: global _zen_model if _zen_model >= 4:
@@ -437,6 +469,7 @@ def main() -> None: all_metrics = MetricGroup("", [ AmdBr(), + AmdCtxSw(), AmdDtlb(), AmdItlb(), AmdLdSt(),
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 24/48 perf jevents: Add RAPL metrics for all Intel models
In reply to Ian RogersAdd a 'cpu_power' metric group that computes the power consumption from RAPL events if they are present.
--- tools/perf/pmu-events/intel_metrics.py | 44 ++++++++++++++++++++++++-- 1 file changed, 41 insertions(+), 3 deletions(-)
tools/perf/pmu-events/intel_metrics.py +41 -3
@@ -1,12 +1,48 @@ #!/usr/bin/env python3 # SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause) import argparse +import math import os -from metric import (JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, LoadEvents, - MetricGroup) +from metric import (d_ratio, has_event, Event, JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, + LoadEvents, Metric, MetricGroup, Select) # Global command line arguments. _args = None +interval_sec = Event("duration_time") + + +def Rapl() -> MetricGroup: + """Processor power consumption estimate. + + Use events from the running average power limit (RAPL) driver. + """ + # Watts = joules/second + pkg = Event("power/energy\\-pkg/") + cond_pkg = Select(pkg, has_event(pkg), math.nan) + cores = Event("power/energy\\-cores/") + cond_cores = Select(cores, has_event(cores), math.nan) + ram = Event("power/energy\\-ram/") + cond_ram = Select(ram, has_event(ram), math.nan) + gpu = Event("power/energy\\-gpu/") + cond_gpu = Select(gpu, has_event(gpu), math.nan) + psys = Event("power/energy\\-psys/") + cond_psys = Select(psys, has_event(psys), math.nan) + scale = 2.3283064365386962890625e-10 + metrics = [ + Metric("lpm_cpu_power_pkg", "", + d_ratio(cond_pkg * scale, interval_sec), "Watts"), + Metric("lpm_cpu_power_cores", "", + d_ratio(cond_cores * scale, interval_sec), "Watts"), + Metric("lpm_cpu_power_ram", "", + d_ratio(cond_ram * scale, interval_sec), "Watts"), + Metric("lpm_cpu_power_gpu", "", + d_ratio(cond_gpu * scale, interval_sec), "Watts"), + Metric("lpm_cpu_power_psys", "", + d_ratio(cond_psys * scale, interval_sec), "Watts"), + ] + + return MetricGroup("lpm_cpu_power", metrics, + description="Running Average Power Limit (RAPL) power consumption estimates") def main() -> None:
@@ -33,7 +69,9 @@ def main() -> None: directory = f"{_args.events_path}/x86/{_args.model}/" LoadEvents(directory) - all_metrics = MetricGroup("", []) + all_metrics = MetricGroup("", [ + Rapl(), + ]) if _args.metricgroups: print(JsonEncodeMetricGroupDescriptions(all_metrics))
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 25/48 perf jevents: Add idle metric for Intel models
In reply to Ian RogersCompute using the msr PMU the percentage of wallclock cycles where the CPUs are in a low power state.
--- tools/perf/pmu-events/intel_metrics.py | 16 ++++++++++++++-- 1 file changed, 14 insertions(+), 2 deletions(-)
tools/perf/pmu-events/intel_metrics.py +14 -2
@@ -3,14 +3,25 @@ import argparse import math import os -from metric import (d_ratio, has_event, Event, JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, - LoadEvents, Metric, MetricGroup, Select) +from metric import (d_ratio, has_event, max, Event, JsonEncodeMetric, + JsonEncodeMetricGroupDescriptions, LoadEvents, Metric, + MetricGroup, Select) # Global command line arguments. _args = None interval_sec = Event("duration_time") +def Idle() -> Metric: + cyc = Event("msr/mperf/") + tsc = Event("msr/tsc/") + low = max(tsc - cyc, 0) + return Metric( + "lpm_idle", + "Percentage of total wallclock cycles where CPUs are in low power state (C1 or deeper sleep state)", + d_ratio(low, tsc), "100%") + + def Rapl() -> MetricGroup: """Processor power consumption estimate.
@@ -70,6 +81,7 @@ def main() -> None: LoadEvents(directory) all_metrics = MetricGroup("", [ + Idle(), Rapl(), ])
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 26/48 perf jevents: Add CheckPmu to see if a PMU is in loaded json events
In reply to Ian RogersCheckPmu can be used to determine if hybrid events are present, allowing for hybrid conditional metrics/events/pmus to be premised on the json files rather than hard coded tables.
--- tools/perf/pmu-events/metric.py | 8 ++++++++ 1 file changed, 8 insertions(+)
tools/perf/pmu-events/metric.py +8 -0
@@ -8,10 +8,12 @@ import re from enum import Enum from typing import Dict, List, Optional, Set, Tuple, Union +all_pmus = set() all_events = set() def LoadEvents(directory: str) -> None: """Populate a global set of all known events for the purpose of validating Event names""" + global all_pmus global all_events all_events = { "context\\-switches",
@@ -26,6 +28,8 @@ def LoadEvents(directory: str) -> None: if filename.endswith(".json"): try: for x in json.load(open(f"{directory}/{filename}")): + if "Unit" in x: + all_pmus.add(x["Unit"]) if "EventName" in x: all_events.add(x["EventName"]) elif "ArchStdEvent" in x:
@@ -36,6 +40,10 @@ def LoadEvents(directory: str) -> None: pass +def CheckPmu(name: str) -> bool: + return name in all_pmus + + def CheckEvent(name: str) -> bool: """Check the event name exists in the set of all loaded events""" global all_events
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 27/48 perf jevents: Add smi metric group for Intel models
In reply to Ian RogersAllow duplicated metric to be dropped from json files.
--- tools/perf/pmu-events/intel_metrics.py | 24 ++++++++++++++++++++++-- 1 file changed, 22 insertions(+), 2 deletions(-)
tools/perf/pmu-events/intel_metrics.py +22 -2
@@ -3,9 +3,9 @@ import argparse import math import os -from metric import (d_ratio, has_event, max, Event, JsonEncodeMetric, +from metric import (d_ratio, has_event, max, CheckPmu, Event, JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, LoadEvents, Metric, - MetricGroup, Select) + MetricGroup, MetricRef, Select) # Global command line arguments. _args = None
@@ -56,6 +56,25 @@ def Rapl() -> MetricGroup: description="Running Average Power Limit (RAPL) power consumption estimates") +def Smi() -> MetricGroup: + pmu = "<cpu_core or cpu_atom>" if CheckPmu("cpu_core") else "cpu" + aperf = Event('msr/aperf/') + cycles = Event('cycles') + smi_num = Event('msr/smi/') + smi_cycles = Select(Select((aperf - cycles) / aperf, smi_num > 0, 0), + has_event(aperf), + 0) + return MetricGroup('smi', [ + Metric('smi_num', 'Number of SMI interrupts.', + Select(smi_num, has_event(smi_num), 0), 'SMI#'), + # Note, the smi_cycles "Event" is really a reference to the metric. + Metric('smi_cycles', + 'Percentage of cycles spent in System Management Interrupts. ' + f'Requires /sys/bus/event_source/devices/{pmu}/freeze_on_smi to be 1.', + smi_cycles, '100%', threshold=(MetricRef('smi_cycles') > 0.10)) + ], description='System Management Interrupt metrics') + + def main() -> None: global _args
@@ -83,6 +102,7 @@ def main() -> None: all_metrics = MetricGroup("", [ Idle(), Rapl(), + Smi(), ]) if _args.metricgroups:
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 28/48 perf jevents: Mark metrics with experimental events as experimental
In reply to Ian RogersWhen metrics are made with experimental events it is desirable the metric description also carries this information in case of metric inaccuracies.
--- tools/perf/pmu-events/metric.py | 44 +++++++++++++++++++++++++++++++++ 1 file changed, 44 insertions(+)
tools/perf/pmu-events/metric.py +44 -0
@@ -10,11 +10,13 @@ from typing import Dict, List, Optional, Set, Tuple, Union all_pmus = set() all_events = set() +experimental_events = set() def LoadEvents(directory: str) -> None: """Populate a global set of all known events for the purpose of validating Event names""" global all_pmus global all_events + global experimental_events all_events = { "context\\-switches", "cpu\\-cycles",
@@ -32,6 +34,8 @@ def LoadEvents(directory: str) -> None: all_pmus.add(x["Unit"]) if "EventName" in x: all_events.add(x["EventName"]) + if "Experimental" in x and x["Experimental"] == "1": + experimental_events.add(x["EventName"]) elif "ArchStdEvent" in x: all_events.add(x["ArchStdEvent"]) except json.decoder.JSONDecodeError:
@@ -61,6 +65,18 @@ def CheckEvent(name: str) -> bool: return name in all_events +def IsExperimentalEvent(name: str) -> bool: + global experimental_events + if ':' in name: + # Remove trailing modifier. + name = name[:name.find(':')] + elif '/' in name: + # Name could begin with a PMU or an event, for now assume it is not experimental. + return False + + return name in experimental_events + + class MetricConstraint(Enum): GROUPED_EVENTS = 0 NO_GROUP_EVENTS = 1
@@ -82,6 +98,10 @@ class Expression: """Returns a simplified version of self.""" raise NotImplementedError() + def HasExperimentalEvents(self) -> bool: + """Are experimental events used in the expression?""" + raise NotImplementedError() + def Equals(self, other) -> bool: """Returns true when two expressions are the same.""" raise NotImplementedError()
@@ -249,6 +269,9 @@ class Operator(Expression): return Operator(self.operator, lhs, rhs) + def HasExperimentalEvents(self) -> bool: + return self.lhs.HasExperimentalEvents() or self.rhs.HasExperimentalEvents() + def Equals(self, other: Expression) -> bool: if isinstance(other, Operator): return self.operator == other.operator and self.lhs.Equals(
@@ -297,6 +320,10 @@ class Select(Expression): return Select(true_val, cond, false_val) + def HasExperimentalEvents(self) -> bool: + return (self.cond.HasExperimentalEvents() or self.true_val.HasExperimentalEvents() or + self.false_val.HasExperimentalEvents()) + def Equals(self, other: Expression) -> bool: if isinstance(other, Select): return self.cond.Equals(other.cond) and self.false_val.Equals(
@@ -345,6 +372,9 @@ class Function(Expression): return Function(self.fn, lhs, rhs) + def HasExperimentalEvents(self) -> bool: + return self.lhs.HasExperimentalEvents() or (self.rhs and self.rhs.HasExperimentalEvents()) + def Equals(self, other: Expression) -> bool: if isinstance(other, Function): result = self.fn == other.fn and self.lhs.Equals(other.lhs)
@@ -384,6 +414,9 @@ class Event(Expression): global all_events raise Exception(f"No event {error} in:\n{all_events}") + def HasExperimentalEvents(self) -> bool: + return IsExperimentalEvent(self.name) + def ToPerfJson(self): result = re.sub('/', '@', self.name) return result
@@ -416,6 +449,9 @@ class MetricRef(Expression): def Simplify(self) -> Expression: return self + def HasExperimentalEvents(self) -> bool: + return False + def Equals(self, other: Expression) -> bool: return isinstance(other, MetricRef) and self.name == other.name
@@ -443,6 +479,9 @@ class Constant(Expression): def Simplify(self) -> Expression: return self + def HasExperimentalEvents(self) -> bool: + return False + def Equals(self, other: Expression) -> bool: return isinstance(other, Constant) and self.value == other.value
@@ -465,6 +504,9 @@ class Literal(Expression): def Simplify(self) -> Expression: return self + def HasExperimentalEvents(self) -> bool: + return False + def Equals(self, other: Expression) -> bool: return isinstance(other, Literal) and self.value == other.value
@@ -527,6 +569,8 @@ class Metric: self.name = name self.description = description self.expr = expr.Simplify() + if self.expr.HasExperimentalEvents(): + self.description += " (metric should be considered experimental as it contains experimental events)." # Workraound valid_only_metric hiding certain metrics based on unit. scale_unit = scale_unit.replace('/sec', ' per sec') if scale_unit[0].isdigit():
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 29/48 perf jevents: Add tsx metric group for Intel models
In reply to Ian RogersAllow duplicated metric to be dropped from json files. Detect when TSX is supported by a model by using the json events, use sysfs events at runtime as hypervisors, etc. may disable TSX.
Add CheckPmu to metric to determine if which PMUs have been associated with the loaded events.
--- tools/perf/pmu-events/intel_metrics.py | 50 ++++++++++++++++++++++++++ 1 file changed, 50 insertions(+)
tools/perf/pmu-events/intel_metrics.py +50 -0
@@ -3,6 +3,7 @@ import argparse import math import os +from typing import Optional from metric import (d_ratio, has_event, max, CheckPmu, Event, JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, LoadEvents, Metric, MetricGroup, MetricRef, Select)
@@ -75,6 +76,54 @@ def Smi() -> MetricGroup: ], description='System Management Interrupt metrics') +def Tsx() -> Optional[MetricGroup]: + pmu = "cpu_core" if CheckPmu("cpu_core") else "cpu" + cycles = Event('cycles') + cycles_in_tx = Event(f'{pmu}/cycles\\-t/') + cycles_in_tx_cp = Event(f'{pmu}/cycles\\-ct/') + try: + # Test if the tsx event is present in the json, prefer the + # sysfs version so that we can detect its presence at runtime. + transaction_start = Event("RTM_RETIRED.START") + transaction_start = Event(f'{pmu}/tx\\-start/') + except: + return None + + elision_start = None + try: + # Elision start isn't supported by all models, but we'll not + # generate the tsx_cycles_per_elision metric in that + # case. Again, prefer the sysfs encoding of the event. + elision_start = Event("HLE_RETIRED.START") + elision_start = Event(f'{pmu}/el\\-start/') + except: + pass + + return MetricGroup('transaction', [ + Metric('tsx_transactional_cycles', + 'Percentage of cycles within a transaction region.', + Select(cycles_in_tx / cycles, has_event(cycles_in_tx), 0), + '100%'), + Metric('tsx_aborted_cycles', 'Percentage of cycles in aborted transactions.', + Select(max(cycles_in_tx - cycles_in_tx_cp, 0) / cycles, + has_event(cycles_in_tx), + 0), + '100%'), + Metric('tsx_cycles_per_transaction', + 'Number of cycles within a transaction divided by the number of transactions.', + Select(cycles_in_tx / transaction_start, + has_event(cycles_in_tx), + 0), + "cycles / transaction"), + Metric('tsx_cycles_per_elision', + 'Number of cycles within a transaction divided by the number of elisions.', + Select(cycles_in_tx / elision_start, + has_event(elision_start), + 0), + "cycles / elision") if elision_start else None, + ], description="Breakdown of transactional memory statistics") + + def main() -> None: global _args
@@ -103,6 +152,7 @@ def main() -> None: Idle(), Rapl(), Smi(), + Tsx(), ]) if _args.metricgroups:
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 30/48 perf jevents: Add br metric group for branch statistics on Intel
In reply to Ian RogersThe br metric group for branches itself comprises metric groups for total, taken, conditional, fused and far metric groups using json events. Conditional taken and not taken metrics are specific to Icelake and later generations, so the presence of the event is used to determine whether the metric should exist.
--- tools/perf/pmu-events/intel_metrics.py | 138 +++++++++++++++++++++++++ 1 file changed, 138 insertions(+)
tools/perf/pmu-events/intel_metrics.py +138 -0
@@ -124,6 +124,143 @@ def Tsx() -> Optional[MetricGroup]: ], description="Breakdown of transactional memory statistics") +def IntelBr(): + ins = Event("instructions") + + def Total() -> MetricGroup: + br_all = Event("BR_INST_RETIRED.ALL_BRANCHES", "BR_INST_RETIRED.ANY") + br_m_all = Event("BR_MISP_RETIRED.ALL_BRANCHES", + "BR_INST_RETIRED.MISPRED", + "BR_MISP_EXEC.ANY") + br_clr = None + try: + br_clr = Event("BACLEARS.ANY", "BACLEARS.ALL") + except: + pass + + br_r = d_ratio(br_all, interval_sec) + ins_r = d_ratio(ins, br_all) + misp_r = d_ratio(br_m_all, br_all) + clr_r = d_ratio(br_clr, interval_sec) if br_clr else None + + return MetricGroup("lpm_br_total", [ + Metric("lpm_br_total_retired", + "The number of branch instructions retired per second.", br_r, + "insn/s"), + Metric( + "lpm_br_total_mispred", + "The number of branch instructions retired, of any type, that were " + "not correctly predicted as a percentage of all branch instrucions.", + misp_r, "100%"), + Metric("lpm_br_total_insn_between_branches", + "The number of instructions divided by the number of branches.", + ins_r, "insn"), + Metric("lpm_br_total_insn_fe_resteers", + "The number of resync branches per second.", clr_r, "req/s" + ) if clr_r else None + ]) + + def Taken() -> MetricGroup: + br_all = Event("BR_INST_RETIRED.ALL_BRANCHES", "BR_INST_RETIRED.ANY") + br_m_tk = None + try: + br_m_tk = Event("BR_MISP_RETIRED.NEAR_TAKEN", + "BR_MISP_RETIRED.TAKEN_JCC", + "BR_INST_RETIRED.MISPRED_TAKEN") + except: + pass + br_r = d_ratio(br_all, interval_sec) + ins_r = d_ratio(ins, br_all) + misp_r = d_ratio(br_m_tk, br_all) if br_m_tk else None + return MetricGroup("lpm_br_taken", [ + Metric("lpm_br_taken_retired", + "The number of taken branches that were retired per second.", + br_r, "insn/s"), + Metric( + "lpm_br_taken_mispred", + "The number of retired taken branch instructions that were " + "mispredicted as a percentage of all taken branches.", misp_r, + "100%") if misp_r else None, + Metric( + "lpm_br_taken_insn_between_branches", + "The number of instructions divided by the number of taken branches.", + ins_r, "insn"), + ]) + + def Conditional() -> Optional[MetricGroup]: + try: + br_cond = Event("BR_INST_RETIRED.COND", + "BR_INST_RETIRED.CONDITIONAL", + "BR_INST_RETIRED.TAKEN_JCC") + br_m_cond = Event("BR_MISP_RETIRED.COND", + "BR_MISP_RETIRED.CONDITIONAL", + "BR_MISP_RETIRED.TAKEN_JCC") + except: + return None + + br_cond_nt = None + br_m_cond_nt = None + try: + br_cond_nt = Event("BR_INST_RETIRED.COND_NTAKEN") + br_m_cond_nt = Event("BR_MISP_RETIRED.COND_NTAKEN") + except: + pass + br_r = d_ratio(br_cond, interval_sec) + ins_r = d_ratio(ins, br_cond) + misp_r = d_ratio(br_m_cond, br_cond) + taken_metrics = [ + Metric("lpm_br_cond_retired", "Retired conditional branch instructions.", + br_r, "insn/s"), + Metric("lpm_br_cond_insn_between_branches", + "The number of instructions divided by the number of conditional " + "branches.", ins_r, "insn"), + Metric("lpm_br_cond_mispred", + "Retired conditional branch instructions mispredicted as a " + "percentage of all conditional branches.", misp_r, "100%"), + ] + if not br_m_cond_nt: + return MetricGroup("lpm_br_cond", taken_metrics) + + br_r = d_ratio(br_cond_nt, interval_sec) + ins_r = d_ratio(ins, br_cond_nt) + misp_r = d_ratio(br_m_cond_nt, br_cond_nt) + + not_taken_metrics = [ + Metric("lpm_br_cond_retired", "Retired conditional not taken branch instructions.", + br_r, "insn/s"), + Metric("lpm_br_cond_insn_between_branches", + "The number of instructions divided by the number of not taken conditional " + "branches.", ins_r, "insn"), + Metric("lpm_br_cond_mispred", + "Retired not taken conditional branch instructions mispredicted as a " + "percentage of all not taken conditional branches.", misp_r, "100%"), + ] + return MetricGroup("lpm_br_cond", [ + MetricGroup("lpm_br_cond_nt", not_taken_metrics), + MetricGroup("lpm_br_cond_tkn", taken_metrics), + ]) + + def Far() -> Optional[MetricGroup]: + try: + br_far = Event("BR_INST_RETIRED.FAR_BRANCH") + except: + return None + + br_r = d_ratio(br_far, interval_sec) + ins_r = d_ratio(ins, br_far) + return MetricGroup("lpm_br_far", [ + Metric("lpm_br_far_retired", "Retired far control transfers per second.", + br_r, "insn/s"), + Metric( + "lpm_br_far_insn_between_branches", + "The number of instructions divided by the number of far branches.", + ins_r, "insn"), + ]) + + return MetricGroup("lpm_br", [Total(), Taken(), Conditional(), Far()], + description="breakdown of retired branch instructions") + + def main() -> None: global _args
@@ -153,6 +290,7 @@ def main() -> None: Rapl(), Smi(), Tsx(), + IntelBr(), ]) if _args.metricgroups:
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 31/48 perf jevents: Add software prefetch (swpf) metric group for Intel
In reply to Ian RogersAdd metrics that breakdown software prefetch instruction use.
--- tools/perf/pmu-events/intel_metrics.py | 66 ++++++++++++++++++++++++++ 1 file changed, 66 insertions(+)
tools/perf/pmu-events/intel_metrics.py +66 -0
@@ -261,6 +261,71 @@ def IntelBr(): description="breakdown of retired branch instructions") +def IntelSwpf() -> Optional[MetricGroup]: + ins = Event("instructions") + try: + s_ld = Event("MEM_INST_RETIRED.ALL_LOADS", + "MEM_UOPS_RETIRED.ALL_LOADS") + s_nta = Event("SW_PREFETCH_ACCESS.NTA") + s_t0 = Event("SW_PREFETCH_ACCESS.T0") + s_t1 = Event("SW_PREFETCH_ACCESS.T1_T2") + s_w = Event("SW_PREFETCH_ACCESS.PREFETCHW") + except: + return None + + all_sw = s_nta + s_t0 + s_t1 + s_w + swp_r = d_ratio(all_sw, interval_sec) + ins_r = d_ratio(ins, all_sw) + ld_r = d_ratio(s_ld, all_sw) + + return MetricGroup("lpm_swpf", [ + MetricGroup("lpm_swpf_totals", [ + Metric("lpm_swpf_totals_exec", "Software prefetch instructions per second", + swp_r, "swpf/s"), + Metric("lpm_swpf_totals_insn_per_pf", + "Average number of instructions between software prefetches", + ins_r, "insn/swpf"), + Metric("lpm_swpf_totals_loads_per_pf", + "Average number of loads between software prefetches", + ld_r, "loads/swpf"), + ]), + MetricGroup("lpm_swpf_bkdwn", [ + MetricGroup("lpm_swpf_bkdwn_nta", [ + Metric("lpm_swpf_bkdwn_nta_per_swpf", + "Software prefetch NTA instructions as a percent of all prefetch instructions", + d_ratio(s_nta, all_sw), "100%"), + Metric("lpm_swpf_bkdwn_nta_rate", + "Software prefetch NTA instructions per second", + d_ratio(s_nta, interval_sec), "insn/s"), + ]), + MetricGroup("lpm_swpf_bkdwn_t0", [ + Metric("lpm_swpf_bkdwn_t0_per_swpf", + "Software prefetch T0 instructions as a percent of all prefetch instructions", + d_ratio(s_t0, all_sw), "100%"), + Metric("lpm_swpf_bkdwn_t0_rate", + "Software prefetch T0 instructions per second", + d_ratio(s_t0, interval_sec), "insn/s"), + ]), + MetricGroup("lpm_swpf_bkdwn_t1_t2", [ + Metric("lpm_swpf_bkdwn_t1_t2_per_swpf", + "Software prefetch T1 or T2 instructions as a percent of all prefetch instructions", + d_ratio(s_t1, all_sw), "100%"), + Metric("lpm_swpf_bkdwn_t1_t2_rate", + "Software prefetch T1 or T2 instructions per second", + d_ratio(s_t1, interval_sec), "insn/s"), + ]), + MetricGroup("lpm_swpf_bkdwn_w", [ + Metric("lpm_swpf_bkdwn_w_per_swpf", + "Software prefetch W instructions as a percent of all prefetch instructions", + d_ratio(s_w, all_sw), "100%"), + Metric("lpm_swpf_bkdwn_w_rate", + "Software prefetch W instructions per second", + d_ratio(s_w, interval_sec), "insn/s"), + ]), + ]), + ], description="Software prefetch instruction breakdown") + + def main() -> None: global _args
@@ -291,6 +356,7 @@ def main() -> None: Smi(), Tsx(), IntelBr(), + IntelSwpf(), ]) if _args.metricgroups:
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 32/48 perf jevents: Add ports metric group giving utilization on Intel
In reply to Ian RogersThe ports metric group contains a metric for each port giving its utilization as a ratio of cycles. The metrics are created by looking for UOPS_DISPATCHED.PORT events.
--- tools/perf/pmu-events/intel_metrics.py | 35 ++++++++++++++++++++++++-- 1 file changed, 33 insertions(+), 2 deletions(-)
tools/perf/pmu-events/intel_metrics.py +33 -2
@@ -1,12 +1,14 @@ #!/usr/bin/env python3 # SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause) import argparse +import json import math import os +import re from typing import Optional from metric import (d_ratio, has_event, max, CheckPmu, Event, JsonEncodeMetric, - JsonEncodeMetricGroupDescriptions, LoadEvents, Metric, - MetricGroup, MetricRef, Select) + JsonEncodeMetricGroupDescriptions, Literal, LoadEvents, + Metric, MetricGroup, MetricRef, Select) # Global command line arguments. _args = None
@@ -261,6 +263,34 @@ def IntelBr(): description="breakdown of retired branch instructions") +def IntelPorts() -> Optional[MetricGroup]: + pipeline_events = json.load( + open(f"{_args.events_path}/x86/{_args.model}/pipeline.json")) + + core_cycles = Event("CPU_CLK_UNHALTED.THREAD_P_ANY", + "CPU_CLK_UNHALTED.DISTRIBUTED", + "cycles") + # Number of CPU cycles scaled for SMT. + smt_cycles = Select(core_cycles / 2, Literal("#smt_on"), core_cycles) + + metrics = [] + for x in pipeline_events: + if "EventName" in x and re.search("^UOPS_DISPATCHED.PORT", x["EventName"]): + name = x["EventName"] + port = re.search(r"(PORT_[0-9].*)", name).group(0).lower() + if name.endswith("_CORE"): + cyc = core_cycles + else: + cyc = smt_cycles + metrics.append(Metric(f"lpm_{port}", f"{port} utilization (higher is better)", + d_ratio(Event(name), cyc), "100%")) + if len(metrics) == 0: + return None + + return MetricGroup("lpm_ports", metrics, "functional unit (port) utilization -- " + "fraction of cycles each port is utilized (higher is better)") + + def IntelSwpf() -> Optional[MetricGroup]: ins = Event("instructions") try:
@@ -356,6 +386,7 @@ def main() -> None: Smi(), Tsx(), IntelBr(), + IntelPorts(), IntelSwpf(), ])
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 33/48 perf jevents: Add L2 metrics for Intel
In reply to Ian RogersGive a breakdown of various L2 counters as metrics, including totals, reads, hardware prefetcher, RFO, code and evictions.
--- tools/perf/pmu-events/intel_metrics.py | 170 +++++++++++++++++++++++++ 1 file changed, 170 insertions(+)
tools/perf/pmu-events/intel_metrics.py +170 -0
@@ -263,6 +263,175 @@ def IntelBr(): description="breakdown of retired branch instructions") +def IntelL2() -> Optional[MetricGroup]: + try: + DC_HIT = Event("L2_RQSTS.DEMAND_DATA_RD_HIT") + except: + return None + try: + DC_MISS = Event("L2_RQSTS.DEMAND_DATA_RD_MISS") + l2_dmnd_miss = DC_MISS + l2_dmnd_rd_all = DC_MISS + DC_HIT + except: + DC_ALL = Event("L2_RQSTS.ALL_DEMAND_DATA_RD") + l2_dmnd_miss = DC_ALL - DC_HIT + l2_dmnd_rd_all = DC_ALL + l2_dmnd_mrate = d_ratio(l2_dmnd_miss, interval_sec) + l2_dmnd_rrate = d_ratio(l2_dmnd_rd_all, interval_sec) + + DC_PFH = None + DC_PFM = None + l2_pf_all = None + l2_pf_mrate = None + l2_pf_rrate = None + try: + DC_PFH = Event("L2_RQSTS.PF_HIT") + DC_PFM = Event("L2_RQSTS.PF_MISS") + l2_pf_all = DC_PFH + DC_PFM + l2_pf_mrate = d_ratio(DC_PFM, interval_sec) + l2_pf_rrate = d_ratio(l2_pf_all, interval_sec) + except: + pass + + DC_RFOH = None + DC_RFOM = None + l2_rfo_all = None + l2_rfo_mrate = None + l2_rfo_rrate = None + try: + DC_RFOH = Event("L2_RQSTS.RFO_HIT") + DC_RFOM = Event("L2_RQSTS.RFO_MISS") + l2_rfo_all = DC_RFOH + DC_RFOM + l2_rfo_mrate = d_ratio(DC_RFOM, interval_sec) + l2_rfo_rrate = d_ratio(l2_rfo_all, interval_sec) + except: + pass + + DC_CH = None + try: + DC_CH = Event("L2_RQSTS.CODE_RD_HIT") + except: + pass + DC_CM = Event("L2_RQSTS.CODE_RD_MISS") + DC_IN = Event("L2_LINES_IN.ALL") + DC_OUT_NS = None + DC_OUT_S = None + l2_lines_out = None + l2_out_rate = None + wbn = None + isd = None + try: + DC_OUT_NS = Event("L2_LINES_OUT.NON_SILENT", + "L2_LINES_OUT.DEMAND_DIRTY", + "L2_LINES_IN.S") + DC_OUT_S = Event("L2_LINES_OUT.SILENT", + "L2_LINES_OUT.DEMAND_CLEAN", + "L2_LINES_IN.I") + if DC_OUT_S.name == "L2_LINES_OUT.SILENT" and ( + args.model.startswith("skylake") or + args.model == "cascadelakex"): + DC_OUT_S.name = "L2_LINES_OUT.SILENT/any/" + # bring is back to per-CPU + l2_s = Select(DC_OUT_S / 2, Literal("#smt_on"), DC_OUT_S) + l2_ns = DC_OUT_NS + l2_lines_out = l2_s + l2_ns + l2_out_rate = d_ratio(l2_lines_out, interval_sec) + nlr = max(l2_ns - DC_WB_U - DC_WB_D, 0) + wbn = d_ratio(nlr, interval_sec) + isd = d_ratio(l2_s, interval_sec) + except: + pass + DC_OUT_U = None + l2_pf_useless = None + l2_useless_rate = None + try: + DC_OUT_U = Event("L2_LINES_OUT.USELESS_HWPF") + l2_pf_useless = DC_OUT_U + l2_useless_rate = d_ratio(l2_pf_useless, interval_sec) + except: + pass + DC_WB_U = None + DC_WB_D = None + wbu = None + wbd = None + try: + DC_WB_U = Event("IDI_MISC.WB_UPGRADE") + DC_WB_D = Event("IDI_MISC.WB_DOWNGRADE") + wbu = d_ratio(DC_WB_U, interval_sec) + wbd = d_ratio(DC_WB_D, interval_sec) + except: + pass + + l2_lines_in = DC_IN + l2_code_all = (DC_CH + DC_CM) if DC_CH else None + l2_code_rate = d_ratio(l2_code_all, interval_sec) if DC_CH else None + l2_code_miss_rate = d_ratio(DC_CM, interval_sec) + l2_in_rate = d_ratio(l2_lines_in, interval_sec) + + return MetricGroup("lpm_l2", [ + MetricGroup("lpm_l2_totals", [ + Metric("lpm_l2_totals_in", "L2 cache total in per second", + l2_in_rate, "In/s"), + Metric("lpm_l2_totals_out", "L2 cache total out per second", + l2_out_rate, "Out/s") if l2_out_rate else None, + ]), + MetricGroup("lpm_l2_rd", [ + Metric("lpm_l2_rd_hits", "L2 cache data read hits", + d_ratio(DC_HIT, l2_dmnd_rd_all), "100%"), + Metric("lpm_l2_rd_hits", "L2 cache data read hits", + d_ratio(l2_dmnd_miss, l2_dmnd_rd_all), "100%"), + Metric("lpm_l2_rd_requests", "L2 cache data read requests per second", + l2_dmnd_rrate, "requests/s"), + Metric("lpm_l2_rd_misses", "L2 cache data read misses per second", + l2_dmnd_mrate, "misses/s"), + ]), + MetricGroup("lpm_l2_hwpf", [ + Metric("lpm_l2_hwpf_hits", "L2 cache hardware prefetcher hits", + d_ratio(DC_PFH, l2_pf_all), "100%"), + Metric("lpm_l2_hwpf_misses", "L2 cache hardware prefetcher misses", + d_ratio(DC_PFM, l2_pf_all), "100%"), + Metric("lpm_l2_hwpf_useless", "L2 cache hardware prefetcher useless prefetches per second", + l2_useless_rate, "100%") if l2_useless_rate else None, + Metric("lpm_l2_hwpf_requests", "L2 cache hardware prefetcher requests per second", + l2_pf_rrate, "100%"), + Metric("lpm_l2_hwpf_misses", "L2 cache hardware prefetcher misses per second", + l2_pf_mrate, "100%"), + ]) if DC_PFH else None, + MetricGroup("lpm_l2_rfo", [ + Metric("lpm_l2_rfo_hits", "L2 cache request for ownership (RFO) hits", + d_ratio(DC_RFOH, l2_rfo_all), "100%"), + Metric("lpm_l2_rfo_misses", "L2 cache request for ownership (RFO) misses", + d_ratio(DC_RFOM, l2_rfo_all), "100%"), + Metric("lpm_l2_rfo_requests", "L2 cache request for ownership (RFO) requests per second", + l2_rfo_rrate, "requests/s"), + Metric("lpm_l2_rfo_misses", "L2 cache request for ownership (RFO) misses per second", + l2_rfo_mrate, "misses/s"), + ]) if DC_RFOH else None, + MetricGroup("lpm_l2_code", [ + Metric("lpm_l2_code_hits", "L2 cache code hits", + d_ratio(DC_CH, l2_code_all), "100%") if DC_CH else None, + Metric("lpm_l2_code_misses", "L2 cache code misses", + d_ratio(DC_CM, l2_code_all), "100%") if DC_CH else None, + Metric("lpm_l2_code_requests", "L2 cache code requests per second", + l2_code_rate, "requests/s") if DC_CH else None, + Metric("lpm_l2_code_misses", "L2 cache code misses per second", + l2_code_miss_rate, "misses/s"), + ]), + MetricGroup("lpm_l2_evict", [ + MetricGroup("lpm_l2_evict_mef_lines", [ + Metric("lpm_l2_evict_mef_lines_l3_hot_lru", "L2 evictions M/E/F lines L3 hot LRU per second", + wbu, "HotLRU/s") if wbu else None, + Metric("lpm_l2_evict_mef_lines_l3_norm_lru", "L2 evictions M/E/F lines L3 normal LRU per second", + wbn, "NormLRU/s") if wbn else None, + Metric("lpm_l2_evict_mef_lines_dropped", "L2 evictions M/E/F lines dropped per second", + wbd, "dropped/s") if wbd else None, + Metric("lpm_l2_evict_is_lines_dropped", "L2 evictions I/S lines dropped per second", + isd, "dropped/s") if isd else None, + ]), + ]), + ], description="L2 data cache analysis") + + def IntelPorts() -> Optional[MetricGroup]: pipeline_events = json.load( open(f"{_args.events_path}/x86/{_args.model}/pipeline.json"))
@@ -386,6 +555,7 @@ def main() -> None: Smi(), Tsx(), IntelBr(), + IntelL2(), IntelPorts(), IntelSwpf(), ])
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 34/48 perf jevents: Add load store breakdown metrics ldst for Intel
In reply to Ian RogersGive breakdown of number of instructions. Use the counter mask (cmask) to show the number of cycles taken to retire the instructions.
--- tools/perf/pmu-events/intel_metrics.py | 87 +++++++++++++++++++++++++- 1 file changed, 86 insertions(+), 1 deletion(-)
tools/perf/pmu-events/intel_metrics.py +86 -1
@@ -8,7 +8,7 @@ import re from typing import Optional from metric import (d_ratio, has_event, max, CheckPmu, Event, JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, Literal, LoadEvents, - Metric, MetricGroup, MetricRef, Select) + Metric, MetricConstraint, MetricGroup, MetricRef, Select) # Global command line arguments. _args = None
@@ -525,6 +525,90 @@ def IntelSwpf() -> Optional[MetricGroup]: ], description="Software prefetch instruction breakdown") +def IntelLdSt() -> Optional[MetricGroup]: + if _args.model in [ + "bonnell", + "nehalemep", + "nehalemex", + "westmereep-dp", + "westmereep-sp", + "westmereex", + ]: + return None + LDST_LD = Event("MEM_INST_RETIRED.ALL_LOADS", "MEM_UOPS_RETIRED.ALL_LOADS") + LDST_ST = Event("MEM_INST_RETIRED.ALL_STORES", + "MEM_UOPS_RETIRED.ALL_STORES") + LDST_LDC1 = Event(f"{LDST_LD.name}/cmask=1/") + LDST_STC1 = Event(f"{LDST_ST.name}/cmask=1/") + LDST_LDC2 = Event(f"{LDST_LD.name}/cmask=2/") + LDST_STC2 = Event(f"{LDST_ST.name}/cmask=2/") + LDST_LDC3 = Event(f"{LDST_LD.name}/cmask=3/") + LDST_STC3 = Event(f"{LDST_ST.name}/cmask=3/") + ins = Event("instructions") + LDST_CYC = Event("CPU_CLK_UNHALTED.THREAD", + "CPU_CLK_UNHALTED.CORE_P", + "CPU_CLK_UNHALTED.THREAD_P") + LDST_PRE = None + try: + LDST_PRE = Event("LOAD_HIT_PREFETCH.SWPF", "LOAD_HIT_PRE.SW_PF") + except: + pass + LDST_AT = None + try: + LDST_AT = Event("MEM_INST_RETIRED.LOCK_LOADS") + except: + pass + cyc = LDST_CYC + + ld_rate = d_ratio(LDST_LD, interval_sec) + st_rate = d_ratio(LDST_ST, interval_sec) + pf_rate = d_ratio(LDST_PRE, interval_sec) if LDST_PRE else None + at_rate = d_ratio(LDST_AT, interval_sec) if LDST_AT else None + + ldst_ret_constraint = MetricConstraint.GROUPED_EVENTS + if LDST_LD.name == "MEM_UOPS_RETIRED.ALL_LOADS": + ldst_ret_constraint = MetricConstraint.NO_GROUP_EVENTS_NMI + + return MetricGroup("lpm_ldst", [ + MetricGroup("lpm_ldst_total", [ + Metric("lpm_ldst_total_loads", "Load/store instructions total loads", + ld_rate, "loads"), + Metric("lpm_ldst_total_stores", "Load/store instructions total stores", + st_rate, "stores"), + ]), + MetricGroup("lpm_ldst_prcnt", [ + Metric("lpm_ldst_prcnt_loads", "Percent of all instructions that are loads", + d_ratio(LDST_LD, ins), "100%"), + Metric("lpm_ldst_prcnt_stores", "Percent of all instructions that are stores", + d_ratio(LDST_ST, ins), "100%"), + ]), + MetricGroup("lpm_ldst_ret_lds", [ + Metric("lpm_ldst_ret_lds_1", "Retired loads in 1 cycle", + d_ratio(max(LDST_LDC1 - LDST_LDC2, 0), cyc), "100%", + constraint=ldst_ret_constraint), + Metric("lpm_ldst_ret_lds_2", "Retired loads in 2 cycles", + d_ratio(max(LDST_LDC2 - LDST_LDC3, 0), cyc), "100%", + constraint=ldst_ret_constraint), + Metric("lpm_ldst_ret_lds_3", "Retired loads in 3 or more cycles", + d_ratio(LDST_LDC3, cyc), "100%"), + ]), + MetricGroup("lpm_ldst_ret_sts", [ + Metric("lpm_ldst_ret_sts_1", "Retired stores in 1 cycle", + d_ratio(max(LDST_STC1 - LDST_STC2, 0), cyc), "100%", + constraint=ldst_ret_constraint), + Metric("lpm_ldst_ret_sts_2", "Retired stores in 2 cycles", + d_ratio(max(LDST_STC2 - LDST_STC3, 0), cyc), "100%", + constraint=ldst_ret_constraint), + Metric("lpm_ldst_ret_sts_3", "Retired stores in 3 more cycles", + d_ratio(LDST_STC3, cyc), "100%"), + ]), + Metric("lpm_ldst_ld_hit_swpf", "Load hit software prefetches per second", + pf_rate, "swpf/s") if pf_rate else None, + Metric("lpm_ldst_atomic_lds", "Atomic loads per second", + at_rate, "loads/s") if at_rate else None, + ], description="Breakdown of load/store instructions") + + def main() -> None: global _args
@@ -556,6 +640,7 @@ def main() -> None: Tsx(), IntelBr(), IntelL2(), + IntelLdSt(), IntelPorts(), IntelSwpf(), ])
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 35/48 perf jevents: Add ILP metrics for Intel
In reply to Ian RogersUse the counter mask (cmask) to see how many cycles an instruction takes to retire. Present as a set of ILP metrics.
--- tools/perf/pmu-events/intel_metrics.py | 40 ++++++++++++++++++++++++++ 1 file changed, 40 insertions(+)
tools/perf/pmu-events/intel_metrics.py +40 -0
@@ -263,6 +263,45 @@ def IntelBr(): description="breakdown of retired branch instructions") +def IntelIlp() -> MetricGroup: + tsc = Event("msr/tsc/") + c0 = Event("msr/mperf/") + low = tsc - c0 + inst_ret = Event("INST_RETIRED.ANY_P") + inst_ret_c = [Event(f"{inst_ret.name}/cmask={x}/") for x in range(1, 6)] + core_cycles = Event("CPU_CLK_UNHALTED.THREAD_P_ANY", + "CPU_CLK_UNHALTED.DISTRIBUTED", + "cycles") + ilp = [d_ratio(max(inst_ret_c[x] - inst_ret_c[x + 1], 0), core_cycles) + for x in range(0, 4)] + ilp.append(d_ratio(inst_ret_c[4], core_cycles)) + ilp0 = 1 + for x in ilp: + ilp0 -= x + return MetricGroup("lpm_ilp", [ + Metric("lpm_ilp_idle", "Lower power cycles as a percentage of all cycles", + d_ratio(low, tsc), "100%"), + Metric("lpm_ilp_inst_ret_0", + "Instructions retired in 0 cycles as a percentage of all cycles", + ilp0, "100%"), + Metric("lpm_ilp_inst_ret_1", + "Instructions retired in 1 cycles as a percentage of all cycles", + ilp[0], "100%"), + Metric("lpm_ilp_inst_ret_2", + "Instructions retired in 2 cycles as a percentage of all cycles", + ilp[1], "100%"), + Metric("lpm_ilp_inst_ret_3", + "Instructions retired in 3 cycles as a percentage of all cycles", + ilp[2], "100%"), + Metric("lpm_ilp_inst_ret_4", + "Instructions retired in 4 cycles as a percentage of all cycles", + ilp[3], "100%"), + Metric("lpm_ilp_inst_ret_5", + "Instructions retired in 5 or more cycles as a percentage of all cycles", + ilp[4], "100%"), + ]) + + def IntelL2() -> Optional[MetricGroup]: try: DC_HIT = Event("L2_RQSTS.DEMAND_DATA_RD_HIT")
@@ -639,6 +678,7 @@ def main() -> None: Smi(), Tsx(), IntelBr(), + IntelIlp(), IntelL2(), IntelLdSt(), IntelPorts(),
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 36/48 perf jevents: Add context switch metrics for Intel
In reply to Ian RogersMetrics break down context switches for different kinds of instruction.
--- tools/perf/pmu-events/intel_metrics.py | 58 ++++++++++++++++++++++++++ 1 file changed, 58 insertions(+)
tools/perf/pmu-events/intel_metrics.py +58 -0
@@ -263,6 +263,63 @@ def IntelBr(): description="breakdown of retired branch instructions") +def IntelCtxSw() -> MetricGroup: + cs = Event("context\\-switches") + metrics = [ + Metric("lpm_cs_rate", "Context switches per second", + d_ratio(cs, interval_sec), "ctxsw/s") + ] + + ev = Event("instructions") + metrics.append(Metric("lpm_cs_instr", "Instructions per context switch", + d_ratio(ev, cs), "instr/cs")) + + ev = Event("cycles") + metrics.append(Metric("lpm_cs_cycles", "Cycles per context switch", + d_ratio(ev, cs), "cycles/cs")) + + try: + ev = Event("MEM_INST_RETIRED.ALL_LOADS", "MEM_UOPS_RETIRED.ALL_LOADS") + metrics.append(Metric("lpm_cs_loads", "Loads per context switch", + d_ratio(ev, cs), "loads/cs")) + except: + pass + + try: + ev = Event("MEM_INST_RETIRED.ALL_STORES", + "MEM_UOPS_RETIRED.ALL_STORES") + metrics.append(Metric("lpm_cs_stores", "Stores per context switch", + d_ratio(ev, cs), "stores/cs")) + except: + pass + + try: + ev = Event("BR_INST_RETIRED.NEAR_TAKEN", "BR_INST_RETIRED.TAKEN_JCC") + metrics.append(Metric("lpm_cs_br_taken", "Branches taken per context switch", + d_ratio(ev, cs), "br_taken/cs")) + except: + pass + + try: + l2_misses = (Event("L2_RQSTS.DEMAND_DATA_RD_MISS") + + Event("L2_RQSTS.RFO_MISS") + + Event("L2_RQSTS.CODE_RD_MISS")) + try: + l2_misses += Event("L2_RQSTS.HWPF_MISS", + "L2_RQSTS.L2_PF_MISS", "L2_RQSTS.PF_MISS") + except: + pass + + metrics.append(Metric("lpm_cs_l2_misses", "L2 misses per context switch", + d_ratio(l2_misses, cs), "l2_misses/cs")) + except: + pass + + return MetricGroup("lpm_cs", metrics, + description=("Number of context switches per second, instructions " + "retired & core cycles between context switches")) + + def IntelIlp() -> MetricGroup: tsc = Event("msr/tsc/") c0 = Event("msr/mperf/")
@@ -678,6 +735,7 @@ def main() -> None: Smi(), Tsx(), IntelBr(), + IntelCtxSw(), IntelIlp(), IntelL2(), IntelLdSt(),
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 37/48 perf jevents: Add FPU metrics for Intel
In reply to Ian RogersMetrics break down of floating point operations.
--- tools/perf/pmu-events/intel_metrics.py | 97 ++++++++++++++++++++++++++ 1 file changed, 97 insertions(+)
tools/perf/pmu-events/intel_metrics.py +97 -0
@@ -320,6 +320,102 @@ def IntelCtxSw() -> MetricGroup: "retired & core cycles between context switches")) +def IntelFpu() -> Optional[MetricGroup]: + cyc = Event("cycles") + try: + s_64 = Event("FP_ARITH_INST_RETIRED.SCALAR_SINGLE", + "SIMD_INST_RETIRED.SCALAR_SINGLE") + except: + return None + d_64 = Event("FP_ARITH_INST_RETIRED.SCALAR_DOUBLE", + "SIMD_INST_RETIRED.SCALAR_DOUBLE") + s_128 = Event("FP_ARITH_INST_RETIRED.128B_PACKED_SINGLE", + "SIMD_INST_RETIRED.PACKED_SINGLE") + + flop = s_64 + d_64 + 4 * s_128 + + d_128 = None + s_256 = None + d_256 = None + s_512 = None + d_512 = None + try: + d_128 = Event("FP_ARITH_INST_RETIRED.128B_PACKED_DOUBLE") + flop += 2 * d_128 + s_256 = Event("FP_ARITH_INST_RETIRED.256B_PACKED_SINGLE") + flop += 8 * s_256 + d_256 = Event("FP_ARITH_INST_RETIRED.256B_PACKED_DOUBLE") + flop += 4 * d_256 + s_512 = Event("FP_ARITH_INST_RETIRED.512B_PACKED_SINGLE") + flop += 16 * s_512 + d_512 = Event("FP_ARITH_INST_RETIRED.512B_PACKED_DOUBLE") + flop += 8 * d_512 + except: + pass + + f_assist = Event("ASSISTS.FP", "FP_ASSIST.ANY", "FP_ASSIST.S") + if f_assist in [ + "ASSISTS.FP", + "FP_ASSIST.S", + ]: + f_assist += "/cmask=1/" + + flop_r = d_ratio(flop, interval_sec) + flop_c = d_ratio(flop, cyc) + nmi_constraint = MetricConstraint.GROUPED_EVENTS + if f_assist.name == "ASSISTS.FP": # Icelake+ + nmi_constraint = MetricConstraint.NO_GROUP_EVENTS_NMI + + def FpuMetrics(group: str, fl: Optional[Event], mult: int, desc: str) -> Optional[MetricGroup]: + if not fl: + return None + + f = fl * mult + fl_r = d_ratio(f, interval_sec) + r_s = d_ratio(fl, interval_sec) + return MetricGroup(group, [ + Metric(f"{group}_of_total", desc + " floating point operations per second", + d_ratio(f, flop), "100%"), + Metric(f"{group}_flops", desc + " floating point operations per second", + fl_r, "flops/s"), + Metric(f"{group}_ops", desc + " operations per second", + r_s, "ops/s"), + ]) + + return MetricGroup("lpm_fpu", [ + MetricGroup("lpm_fpu_total", [ + Metric("lpm_fpu_total_flops", "Floating point operations per second", + flop_r, "flops/s"), + Metric("lpm_fpu_total_flopc", "Floating point operations per cycle", + flop_c, "flops/cycle", constraint=nmi_constraint), + ]), + MetricGroup("lpm_fpu_64", [ + FpuMetrics("lpm_fpu_64_single", s_64, 1, "64-bit single"), + FpuMetrics("lpm_fpu_64_double", d_64, 1, "64-bit double"), + ]), + MetricGroup("lpm_fpu_128", [ + FpuMetrics("lpm_fpu_128_single", s_128, + 4, "128-bit packed single"), + FpuMetrics("lpm_fpu_128_double", d_128, + 2, "128-bit packed double"), + ]), + MetricGroup("lpm_fpu_256", [ + FpuMetrics("lpm_fpu_256_single", s_256, + 8, "128-bit packed single"), + FpuMetrics("lpm_fpu_256_double", d_256, + 4, "128-bit packed double"), + ]), + MetricGroup("lpm_fpu_512", [ + FpuMetrics("lpm_fpu_512_single", s_512, + 16, "128-bit packed single"), + FpuMetrics("lpm_fpu_512_double", d_512, + 8, "128-bit packed double"), + ]), + Metric("lpm_fpu_assists", "FP assists as a percentage of cycles", + d_ratio(f_assist, cyc), "100%"), + ]) + + def IntelIlp() -> MetricGroup: tsc = Event("msr/tsc/") c0 = Event("msr/mperf/")
@@ -736,6 +832,7 @@ def main() -> None: Tsx(), IntelBr(), IntelCtxSw(), + IntelFpu(), IntelIlp(), IntelL2(), IntelLdSt(),
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 38/48 perf jevents: Add Miss Level Parallelism (MLP) metric for Intel
In reply to Ian RogersNumber of outstanding load misses per cycle.
--- tools/perf/pmu-events/intel_metrics.py | 15 +++++++++++++++ 1 file changed, 15 insertions(+)
tools/perf/pmu-events/intel_metrics.py +15 -0
@@ -624,6 +624,20 @@ def IntelL2() -> Optional[MetricGroup]: ], description="L2 data cache analysis") +def IntelMlp() -> Optional[Metric]: + try: + l1d = Event("L1D_PEND_MISS.PENDING") + l1dc = Event("L1D_PEND_MISS.PENDING_CYCLES") + except: + return None + + l1dc = Select(l1dc / 2, Literal("#smt_on"), l1dc) + ml = d_ratio(l1d, l1dc) + return Metric("lpm_mlp", + "Miss level parallelism - number of outstanding load misses per cycle (higher is better)", + ml, "load_miss_pending/cycle") + + def IntelPorts() -> Optional[MetricGroup]: pipeline_events = json.load( open(f"{_args.events_path}/x86/{_args.model}/pipeline.json"))
@@ -836,6 +850,7 @@ def main() -> None: IntelIlp(), IntelL2(), IntelLdSt(), + IntelMlp(), IntelPorts(), IntelSwpf(), ])
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 39/48 perf jevents: Add mem_bw metric for Intel
In reply to Ian RogersBreak down memory bandwidth using uncore counters. For many models this matches the memory_bandwidth_* metrics, but these metrics aren't made available on all models. Add support for free running counters. Query the event json when determining which what events/counters are available.
--- tools/perf/pmu-events/intel_metrics.py | 62 ++++++++++++++++++++++++++ 1 file changed, 62 insertions(+)
tools/perf/pmu-events/intel_metrics.py +62 -0
@@ -815,6 +815,67 @@ def IntelLdSt() -> Optional[MetricGroup]: ], description="Breakdown of load/store instructions") +def UncoreMemBw() -> Optional[MetricGroup]: + mem_events = [] + try: + mem_events = json.load(open(f"{os.path.dirname(os.path.realpath(__file__))}" + f"/arch/x86/{args.model}/uncore-memory.json")) + except: + pass + + ddr_rds = 0 + ddr_wrs = 0 + ddr_total = 0 + for x in mem_events: + if "EventName" in x: + name = x["EventName"] + if re.search("^UNC_MC[0-9]+_RDCAS_COUNT_FREERUN", name): + ddr_rds += Event(name) + elif re.search("^UNC_MC[0-9]+_WRCAS_COUNT_FREERUN", name): + ddr_wrs += Event(name) + # elif re.search("^UNC_MC[0-9]+_TOTAL_REQCOUNT_FREERUN", name): + # ddr_total += Event(name) + + if ddr_rds == 0: + try: + ddr_rds = Event("UNC_M_CAS_COUNT.RD") + ddr_wrs = Event("UNC_M_CAS_COUNT.WR") + except: + return None + + ddr_total = ddr_rds + ddr_wrs + + pmm_rds = 0 + pmm_wrs = 0 + try: + pmm_rds = Event("UNC_M_PMM_RPQ_INSERTS") + pmm_wrs = Event("UNC_M_PMM_WPQ_INSERTS") + except: + pass + + pmm_total = pmm_rds + pmm_wrs + + scale = 64 / 1_000_000 + return MetricGroup("lpm_mem_bw", [ + MetricGroup("lpm_mem_bw_ddr", [ + Metric("lpm_mem_bw_ddr_read", "DDR memory read bandwidth", + d_ratio(ddr_rds, interval_sec), f"{scale}MB/s"), + Metric("lpm_mem_bw_ddr_write", "DDR memory write bandwidth", + d_ratio(ddr_wrs, interval_sec), f"{scale}MB/s"), + Metric("lpm_mem_bw_ddr_total", "DDR memory write bandwidth", + d_ratio(ddr_total, interval_sec), f"{scale}MB/s"), + ], description="DDR Memory Bandwidth"), + MetricGroup("lpm_mem_bw_pmm", [ + Metric("lpm_mem_bw_pmm_read", "PMM memory read bandwidth", + d_ratio(pmm_rds, interval_sec), f"{scale}MB/s"), + Metric("lpm_mem_bw_pmm_write", "PMM memory write bandwidth", + d_ratio(pmm_wrs, interval_sec), f"{scale}MB/s"), + Metric("lpm_mem_bw_pmm_total", "PMM memory write bandwidth", + d_ratio(pmm_total, interval_sec), f"{scale}MB/s"), + ], description="PMM Memory Bandwidth") if pmm_rds != 0 else None, + ], description="Memory Bandwidth") + + def main() -> None: global _args
@@ -853,6 +914,7 @@ def main() -> None: IntelMlp(), IntelPorts(), IntelSwpf(), + UncoreMemBw(), ]) if _args.metricgroups:
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 40/48 perf jevents: Add local/remote "mem" breakdown metrics for Intel
In reply to Ian RogersBreakdown local and remote memory bandwidth, read and writes. The implementation uses the HA and CHA PMUs present in server models broadwellde, broadwellx cascadelakex, emeraldrapids, haswellx, icelakex, ivytown, sapphirerapids and skylakex.
--- tools/perf/pmu-events/intel_metrics.py | 31 ++++++++++++++++++++++++++ 1 file changed, 31 insertions(+)
tools/perf/pmu-events/intel_metrics.py +31 -0
@@ -815,6 +815,36 @@ def IntelLdSt() -> Optional[MetricGroup]: ], description="Breakdown of load/store instructions") +def UncoreMem() -> Optional[MetricGroup]: + try: + loc_rds = Event("UNC_CHA_REQUESTS.READS_LOCAL", + "UNC_H_REQUESTS.READS_LOCAL") + rem_rds = Event("UNC_CHA_REQUESTS.READS_REMOTE", + "UNC_H_REQUESTS.READS_REMOTE") + loc_wrs = Event("UNC_CHA_REQUESTS.WRITES_LOCAL", + "UNC_H_REQUESTS.WRITES_LOCAL") + rem_wrs = Event("UNC_CHA_REQUESTS.WRITES_REMOTE", + "UNC_H_REQUESTS.WRITES_REMOTE") + except: + return None + + scale = 64 / 1_000_000 + return MetricGroup("lpm_mem", [ + MetricGroup("lpm_mem_local", [ + Metric("lpm_mem_local_read", "Local memory read bandwidth not including directory updates", + d_ratio(loc_rds, interval_sec), f"{scale}MB/s"), + Metric("lpm_mem_local_write", "Local memory write bandwidth not including directory updates", + d_ratio(loc_wrs, interval_sec), f"{scale}MB/s"), + ]), + MetricGroup("lpm_mem_remote", [ + Metric("lpm_mem_remote_read", "Remote memory read bandwidth not including directory updates", + d_ratio(rem_rds, interval_sec), f"{scale}MB/s"), + Metric("lpm_mem_remote_write", "Remote memory write bandwidth not including directory updates", + d_ratio(rem_wrs, interval_sec), f"{scale}MB/s"), + ]), + ], description="Memory Bandwidth breakdown local vs. remote (remote requests in). directory updates not included") + + def UncoreMemBw() -> Optional[MetricGroup]: mem_events = [] try:
@@ -914,6 +944,7 @@ def main() -> None: IntelMlp(), IntelPorts(), IntelSwpf(), + UncoreMem(), UncoreMemBw(), ])
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 41/48 perf jevents: Add dir breakdown metrics for Intel
In reply to Ian RogersBreakdown directory hit, misses and requests. The implementation uses the M2M and CHA PMUs present in server models broadwellde, broadwellx cascadelakex, emeraldrapids, icelakex, sapphirerapids and skylakex.
--- tools/perf/pmu-events/intel_metrics.py | 36 ++++++++++++++++++++++++++ 1 file changed, 36 insertions(+)
tools/perf/pmu-events/intel_metrics.py +36 -0
@@ -815,6 +815,41 @@ def IntelLdSt() -> Optional[MetricGroup]: ], description="Breakdown of load/store instructions") +def UncoreDir() -> Optional[MetricGroup]: + try: + m2m_upd = Event("UNC_M2M_DIRECTORY_UPDATE.ANY") + m2m_hits = Event("UNC_M2M_DIRECTORY_HIT.DIRTY_I") + # Turn the umask into a ANY rather than DIRTY_I filter. + m2m_hits.name += "/umask=0xFF,name=UNC_M2M_DIRECTORY_HIT.ANY/" + m2m_miss = Event("UNC_M2M_DIRECTORY_MISS.DIRTY_I") + # Turn the umask into a ANY rather than DIRTY_I filter. + m2m_miss.name += "/umask=0xFF,name=UNC_M2M_DIRECTORY_MISS.ANY/" + cha_upd = Event("UNC_CHA_DIR_UPDATE.HA") + # Turn the umask into a ANY rather than HA filter. + cha_upd.name += "/umask=3,name=UNC_CHA_DIR_UPDATE.ANY/" + except: + return None + + m2m_total = m2m_hits + m2m_miss + upd = m2m_upd + cha_upd # in cache lines + upd_r = upd / interval_sec + look_r = m2m_total / interval_sec + + scale = 64 / 1_000_000 # Cache lines to MB + return MetricGroup("lpm_dir", [ + Metric("lpm_dir_lookup_rate", "", + d_ratio(m2m_total, interval_sec), "requests/s"), + Metric("lpm_dir_lookup_hits", "", + d_ratio(m2m_hits, m2m_total), "100%"), + Metric("lpm_dir_lookup_misses", "", + d_ratio(m2m_miss, m2m_total), "100%"), + Metric("lpm_dir_update_requests", "", + d_ratio(m2m_upd + cha_upd, interval_sec), "requests/s"), + Metric("lpm_dir_update_bw", "", + d_ratio(m2m_upd + cha_upd, interval_sec), f"{scale}MB/s"), + ]) + + def UncoreMem() -> Optional[MetricGroup]: try: loc_rds = Event("UNC_CHA_REQUESTS.READS_LOCAL",
@@ -944,6 +979,7 @@ def main() -> None: IntelMlp(), IntelPorts(), IntelSwpf(), + UncoreDir(), UncoreMem(), UncoreMemBw(), ])
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 42/48 perf jevents: Add C-State metrics from the PCU PMU for Intel
In reply to Ian RogersUse occupancy events fixed in:
https://lore.kernel.org/lkml/[email protected]/
Metrics are at the socket level referring to cores, not hyperthreads.
--- tools/perf/pmu-events/intel_metrics.py | 30 ++++++++++++++++++++++++++ 1 file changed, 30 insertions(+)
tools/perf/pmu-events/intel_metrics.py +30 -0
@@ -815,6 +815,35 @@ def IntelLdSt() -> Optional[MetricGroup]: ], description="Breakdown of load/store instructions") +def UncoreCState() -> Optional[MetricGroup]: + try: + pcu_ticks = Event("UNC_P_CLOCKTICKS") + c0 = Event("UNC_P_POWER_STATE_OCCUPANCY.CORES_C0") + c3 = Event("UNC_P_POWER_STATE_OCCUPANCY.CORES_C3") + c6 = Event("UNC_P_POWER_STATE_OCCUPANCY.CORES_C6") + except: + return None + + num_cores = Literal("#num_cores") / Literal("#num_packages") + + max_cycles = pcu_ticks * num_cores + total_cycles = c0 + c3 + c6 + + # remove fused-off cores which show up in C6/C7. + c6 = Select(max(c6 - (total_cycles - max_cycles), 0), + total_cycles > max_cycles, + c6) + + return MetricGroup("lpm_cstate", [ + Metric("lpm_cstate_c0", "C-State cores in C0/C1", + d_ratio(c0, pcu_ticks), "cores"), + Metric("lpm_cstate_c3", "C-State cores in C3", + d_ratio(c3, pcu_ticks), "cores"), + Metric("lpm_cstate_c6", "C-State cores in C6/C7", + d_ratio(c6, pcu_ticks), "cores"), + ]) + + def UncoreDir() -> Optional[MetricGroup]: try: m2m_upd = Event("UNC_M2M_DIRECTORY_UPDATE.ANY")
@@ -979,6 +1008,7 @@ def main() -> None: IntelMlp(), IntelPorts(), IntelSwpf(), + UncoreCState(), UncoreDir(), UncoreMem(), UncoreMemBw(),
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 43/48 perf jevents: Add local/remote miss latency metrics for Intel
In reply to Ian RogersDerive from CBOX/CHA occupancy and inserts the average latency as is provided in Intel's uncore performance monitoring reference.
--- tools/perf/pmu-events/intel_metrics.py | 70 ++++++++++++++++++++++++-- 1 file changed, 67 insertions(+), 3 deletions(-)
tools/perf/pmu-events/intel_metrics.py +67 -3
@@ -6,9 +6,10 @@ import math import os import re from typing import Optional -from metric import (d_ratio, has_event, max, CheckPmu, Event, JsonEncodeMetric, - JsonEncodeMetricGroupDescriptions, Literal, LoadEvents, - Metric, MetricConstraint, MetricGroup, MetricRef, Select) +from metric import (d_ratio, has_event, max, source_count, CheckPmu, Event, + JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, + Literal, LoadEvents, Metric, MetricConstraint, MetricGroup, + MetricRef, Select) # Global command line arguments. _args = None
@@ -624,6 +625,68 @@ def IntelL2() -> Optional[MetricGroup]: ], description="L2 data cache analysis") +def IntelMissLat() -> Optional[MetricGroup]: + try: + ticks = Event("UNC_CHA_CLOCKTICKS", "UNC_C_CLOCKTICKS") + data_rd_loc_occ = Event("UNC_CHA_TOR_OCCUPANCY.IA_MISS_DRD_LOCAL", + "UNC_CHA_TOR_OCCUPANCY.IA_MISS", + "UNC_C_TOR_OCCUPANCY.MISS_LOCAL_OPCODE", + "UNC_C_TOR_OCCUPANCY.MISS_OPCODE") + data_rd_loc_ins = Event("UNC_CHA_TOR_INSERTS.IA_MISS_DRD_LOCAL", + "UNC_CHA_TOR_INSERTS.IA_MISS", + "UNC_C_TOR_INSERTS.MISS_LOCAL_OPCODE", + "UNC_C_TOR_INSERTS.MISS_OPCODE") + data_rd_rem_occ = Event("UNC_CHA_TOR_OCCUPANCY.IA_MISS_DRD_REMOTE", + "UNC_CHA_TOR_OCCUPANCY.IA_MISS", + "UNC_C_TOR_OCCUPANCY.MISS_REMOTE_OPCODE", + "UNC_C_TOR_OCCUPANCY.NID_MISS_OPCODE") + data_rd_rem_ins = Event("UNC_CHA_TOR_INSERTS.IA_MISS_DRD_REMOTE", + "UNC_CHA_TOR_INSERTS.IA_MISS", + "UNC_C_TOR_INSERTS.MISS_REMOTE_OPCODE", + "UNC_C_TOR_INSERTS.NID_MISS_OPCODE") + except: + return None + + if (data_rd_loc_occ.name == "UNC_C_TOR_OCCUPANCY.MISS_LOCAL_OPCODE" or + data_rd_loc_occ.name == "UNC_C_TOR_OCCUPANCY.MISS_OPCODE"): + data_rd = 0x182 + for e in [data_rd_loc_occ, data_rd_loc_ins, data_rd_rem_occ, data_rd_rem_ins]: + e.name += f"/filter_opc={hex(data_rd)}/" + elif data_rd_loc_occ.name == "UNC_CHA_TOR_OCCUPANCY.IA_MISS": + # Demand Data Read - Full cache-line read requests from core for + # lines to be cached in S or E, typically for data + demand_data_rd = 0x202 + # LLC Prefetch Data - Uncore will first look up the line in the + # LLC; for a cache hit, the LRU will be updated, on a miss, the + # DRd will be initiated + llc_prefetch_data = 0x25a + local_filter = (f"/filter_opc0={hex(demand_data_rd)}," + f"filter_opc1={hex(llc_prefetch_data)}," + "filter_loc,filter_nm,filter_not_nm/") + remote_filter = (f"/filter_opc0={hex(demand_data_rd)}," + f"filter_opc1={hex(llc_prefetch_data)}," + "filter_rem,filter_nm,filter_not_nm/") + for e in [data_rd_loc_occ, data_rd_loc_ins]: + e.name += local_filter + for e in [data_rd_rem_occ, data_rd_rem_ins]: + e.name += remote_filter + else: + assert data_rd_loc_occ.name == "UNC_CHA_TOR_OCCUPANCY.IA_MISS_DRD_LOCAL", data_rd_loc_occ + + ticks_per_cha = ticks / source_count(data_rd_loc_ins) + loc_lat = interval_sec * 1e9 * data_rd_loc_occ / \ + (ticks_per_cha * data_rd_loc_ins) + ticks_per_cha = ticks / source_count(data_rd_rem_ins) + rem_lat = interval_sec * 1e9 * data_rd_rem_occ / \ + (ticks_per_cha * data_rd_rem_ins) + return MetricGroup("lpm_miss_lat", [ + Metric("lpm_miss_lat_loc", "Local to a socket miss latency in nanoseconds", + loc_lat, "ns"), + Metric("lpm_miss_lat_rem", "Remote to a socket miss latency in nanoseconds", + rem_lat, "ns"), + ]) + + def IntelMlp() -> Optional[Metric]: try: l1d = Event("L1D_PEND_MISS.PENDING")
@@ -1005,6 +1068,7 @@ def main() -> None: IntelIlp(), IntelL2(), IntelLdSt(), + IntelMissLat(), IntelMlp(), IntelPorts(), IntelSwpf(),
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 44/48 perf jevents: Add upi_bw metric for Intel
In reply to Ian RogersBreak down UPI read and write bandwidth using uncore_upi counters.
--- tools/perf/pmu-events/intel_metrics.py | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+)
tools/perf/pmu-events/intel_metrics.py +22 -0
@@ -1033,6 +1033,27 @@ def UncoreMemBw() -> Optional[MetricGroup]: ], description="Memory Bandwidth") +def UncoreUpiBw() -> Optional[MetricGroup]: + try: + upi_rds = Event("UNC_UPI_RxL_FLITS.ALL_DATA") + upi_wrs = Event("UNC_UPI_TxL_FLITS.ALL_DATA") + except: + return None + + upi_total = upi_rds + upi_wrs + + # From "Uncore Performance Monitoring": When measuring the amount of + # bandwidth consumed by transmission of the data (i.e. NOT including + # the header), it should be .ALL_DATA / 9 * 64B. + scale = (64 / 9) / 1_000_000 + return MetricGroup("lpm_upi_bw", [ + Metric("lpm_upi_bw_read", "UPI read bandwidth", + d_ratio(upi_rds, interval_sec), f"{scale}MB/s"), + Metric("lpm_upi_bw_write", "DDR memory write bandwidth", + d_ratio(upi_wrs, interval_sec), f"{scale}MB/s"), + ], description="UPI Bandwidth") + + def main() -> None: global _args
@@ -1076,6 +1097,7 @@ def main() -> None: UncoreDir(), UncoreMem(), UncoreMemBw(), + UncoreUpiBw(), ]) if _args.metricgroups:
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 45/48 perf jevents: Add mesh bandwidth saturation metric for Intel
In reply to Ian RogersMemory bandwidth saturation from CBOX/CHA events present in broadwellde, broadwellx, cascadelakex, haswellx, icelakex, skylakex and snowridgex.
--- tools/perf/pmu-events/intel_metrics.py | 17 +++++++++++++++++ 1 file changed, 17 insertions(+)
tools/perf/pmu-events/intel_metrics.py +17 -0
@@ -1033,6 +1033,22 @@ def UncoreMemBw() -> Optional[MetricGroup]: ], description="Memory Bandwidth") +def UncoreMemSat() -> Optional[Metric]: + try: + clocks = Event("UNC_CHA_CLOCKTICKS", "UNC_C_CLOCKTICKS") + sat = Event("UNC_CHA_DISTRESS_ASSERTED.VERT", "UNC_CHA_FAST_ASSERTED.VERT", + "UNC_C_FAST_ASSERTED") + except: + return None + + desc = ("Mesh Bandwidth saturation (% CBOX cycles with FAST signal asserted, " + "include QPI bandwidth saturation), lower is better") + if "UNC_CHA_" in sat.name: + desc = ("Mesh Bandwidth saturation (% CHA cycles with FAST signal asserted, " + "include UPI bandwidth saturation), lower is better") + return Metric("lpm_mem_sat", desc, d_ratio(sat, clocks), "100%") + + def UncoreUpiBw() -> Optional[MetricGroup]: try: upi_rds = Event("UNC_UPI_RxL_FLITS.ALL_DATA")
@@ -1097,6 +1113,7 @@ def main() -> None: UncoreDir(), UncoreMem(), UncoreMemBw(), + UncoreMemSat(), UncoreUpiBw(), ])
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 46/48 perf jevents: Add collection of topdown like metrics for arm64
In reply to Ian RogersMetrics are created using legacy, common and recommended events. As events may be missing a TryEvent function will give None if an event is missing. To workaround missing JSON events for cortex-a53, sysfs encodings are used.
---
An earlier review of this patch by Leo Yan is here:
https://lore.kernel.org/lkml/[email protected]/
Hopefully all corrections were made.
--- tools/perf/pmu-events/arm64_metrics.py | 145 ++++++++++++++++++++++++- 1 file changed, 142 insertions(+), 3 deletions(-)
tools/perf/pmu-events/arm64_metrics.py +142 -3
@@ -2,13 +2,150 @@ # SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause) import argparse import os -from metric import (JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, LoadEvents, - MetricGroup) +from typing import Optional +from metric import (d_ratio, Event, JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, + LoadEvents, Metric, MetricGroup) # Global command line arguments. _args = None +def Arm64Topdown() -> MetricGroup: + """Returns a MetricGroup representing ARM64 topdown like metrics.""" + def TryEvent(name: str) -> Optional[Event]: + # Skip an event if not in the json files. + try: + return Event(name) + except: + return None + # ARM models like a53 lack JSON for INST_RETIRED but have the + # architetural standard event in sysfs. Use the PMU name to identify + # the sysfs event. + pmu_name = f'armv8_{_args.model.replace("-", "_")}' + ins = Event("instructions") + ins_ret = Event("INST_RETIRED", f"{pmu_name}/inst_retired/") + cycles = Event("cpu\\-cycles") + stall_fe = TryEvent("STALL_FRONTEND") + stall_be = TryEvent("STALL_BACKEND") + br_ret = TryEvent("BR_RETIRED") + br_mp_ret = TryEvent("BR_MIS_PRED_RETIRED") + dtlb_walk = TryEvent("DTLB_WALK") + itlb_walk = TryEvent("ITLB_WALK") + l1d_tlb = TryEvent("L1D_TLB") + l1i_tlb = TryEvent("L1I_TLB") + l1d_refill = Event("L1D_CACHE_REFILL", f"{pmu_name}/l1d_cache_refill/") + l2d_refill = Event("L2D_CACHE_REFILL", f"{pmu_name}/l2d_cache_refill/") + l1i_refill = Event("L1I_CACHE_REFILL", f"{pmu_name}/l1i_cache_refill/") + l1d_access = Event("L1D_CACHE", f"{pmu_name}/l1d_cache/") + l2d_access = Event("L2D_CACHE", f"{pmu_name}/l2d_cache/") + llc_access = TryEvent("LL_CACHE_RD") + l1i_access = Event("L1I_CACHE", f"{pmu_name}/l1i_cache/") + llc_miss_rd = TryEvent("LL_CACHE_MISS_RD") + ase_spec = TryEvent("ASE_SPEC") + ld_spec = TryEvent("LD_SPEC") + st_spec = TryEvent("ST_SPEC") + vfp_spec = TryEvent("VFP_SPEC") + dp_spec = TryEvent("DP_SPEC") + br_immed_spec = TryEvent("BR_IMMED_SPEC") + br_indirect_spec = TryEvent("BR_INDIRECT_SPEC") + br_ret_spec = TryEvent("BR_RETURN_SPEC") + crypto_spec = TryEvent("CRYPTO_SPEC") + inst_spec = TryEvent("INST_SPEC") + return MetricGroup("lpm_topdown", [ + MetricGroup("lpm_topdown_tl", [ + Metric("lpm_topdown_tl_ipc", "Instructions per cycle", d_ratio( + ins, cycles), "insn/cycle"), + Metric("lpm_topdown_tl_stall_fe_rate", "Frontend stalls to all cycles", + d_ratio(stall_fe, cycles), "100%") if stall_fe else None, + Metric("lpm_topdown_tl_stall_be_rate", "Backend stalls to all cycles", + d_ratio(stall_be, cycles), "100%") if stall_be else None, + ]), + MetricGroup("lpm_topdown_fe_bound", [ + MetricGroup("lpm_topdown_fe_br", [ + Metric("lpm_topdown_fe_br_mp_per_insn", + "Branch mispredicts per instruction retired", + d_ratio(br_mp_ret, ins_ret), "br/insn") if br_mp_ret else None, + Metric("lpm_topdown_fe_br_ins_rate", + "Branches per instruction retired", d_ratio( + br_ret, ins_ret), "100%") if br_ret else None, + Metric("lpm_topdown_fe_br_mispredict", + "Branch mispredicts per branch instruction", + d_ratio(br_mp_ret, br_ret), "100%") if (br_mp_ret and br_ret) else None, + ]), + MetricGroup("lpm_topdown_fe_itlb", [ + Metric("lpm_topdown_fe_itlb_walks", "Itlb walks per insn", + d_ratio(itlb_walk, ins_ret), "walk/insn"), + Metric("lpm_topdown_fe_itlb_walk_rate", "Itlb walks per L1I TLB access", + d_ratio(itlb_walk, l1i_tlb) if l1i_tlb else None, "100%"), + ]) if itlb_walk else None, + MetricGroup("lpm_topdown_fe_icache", [ + Metric("lpm_topdown_fe_icache_l1i_per_insn", + "L1I cache refills per instruction", + d_ratio(l1i_refill, ins_ret), "l1i/insn"), + Metric("lpm_topdown_fe_icache_l1i_miss_rate", + "L1I cache refills per L1I cache access", + d_ratio(l1i_refill, l1i_access), "100%"), + ]), + ]), + MetricGroup("lpm_topdown_be_bound", [ + MetricGroup("lpm_topdown_be_dtlb", [ + Metric("lpm_topdown_be_dtlb_walks", "Dtlb walks per instruction", + d_ratio(dtlb_walk, ins_ret), "walk/insn"), + Metric("lpm_topdown_be_dtlb_walk_rate", "Dtlb walks per L1D TLB access", + d_ratio(dtlb_walk, l1d_tlb) if l1d_tlb else None, "100%"), + ]) if dtlb_walk else None, + MetricGroup("lpm_topdown_be_mix", [ + Metric("lpm_topdown_be_mix_ld", "Percentage of load instructions", + d_ratio(ld_spec, inst_spec), "100%") if ld_spec else None, + Metric("lpm_topdown_be_mix_st", "Percentage of store instructions", + d_ratio(st_spec, inst_spec), "100%") if st_spec else None, + Metric("lpm_topdown_be_mix_simd", "Percentage of SIMD instructions", + d_ratio(ase_spec, inst_spec), "100%") if ase_spec else None, + Metric("lpm_topdown_be_mix_fp", + "Percentage of floating point instructions", + d_ratio(vfp_spec, inst_spec), "100%") if vfp_spec else None, + Metric("lpm_topdown_be_mix_dp", + "Percentage of data processing instructions", + d_ratio(dp_spec, inst_spec), "100%") if dp_spec else None, + Metric("lpm_topdown_be_mix_crypto", + "Percentage of data processing instructions", + d_ratio(crypto_spec, inst_spec), "100%") if crypto_spec else None, + Metric( + "lpm_topdown_be_mix_br", "Percentage of branch instructions", + d_ratio(br_immed_spec + br_indirect_spec + br_ret_spec, + inst_spec), "100%") if br_immed_spec and br_indirect_spec and br_ret_spec else None, + ], description="Breakdown of instructions by type. Counts include both useful and wasted speculative instructions" + ) if inst_spec else None, + MetricGroup("lpm_topdown_be_dcache", [ + MetricGroup("lpm_topdown_be_dcache_l1", [ + Metric("lpm_topdown_be_dcache_l1_per_insn", + "L1D cache refills per instruction", + d_ratio(l1d_refill, ins_ret), "refills/insn"), + Metric("lpm_topdown_be_dcache_l1_miss_rate", + "L1D cache refills per L1D cache access", + d_ratio(l1d_refill, l1d_access), "100%") + ]), + MetricGroup("lpm_topdown_be_dcache_l2", [ + Metric("lpm_topdown_be_dcache_l2_per_insn", + "L2D cache refills per instruction", + d_ratio(l2d_refill, ins_ret), "refills/insn"), + Metric("lpm_topdown_be_dcache_l2_miss_rate", + "L2D cache refills per L2D cache access", + d_ratio(l2d_refill, l2d_access), "100%") + ]), + MetricGroup("lpm_topdown_be_dcache_llc", [ + Metric("lpm_topdown_be_dcache_llc_per_insn", + "Last level cache misses per instruction", + d_ratio(llc_miss_rd, ins_ret), "miss/insn"), + Metric("lpm_topdown_be_dcache_llc_miss_rate", + "Last level cache misses per last level cache access", + d_ratio(llc_miss_rd, llc_access), "100%") + ]) if llc_miss_rd and llc_access else None, + ]), + ]), + ]) + + def main() -> None: global _args
@@ -34,7 +171,9 @@ def main() -> None: directory = f"{_args.events_path}/arm64/{_args.vendor}/{_args.model}/" LoadEvents(directory) - all_metrics = MetricGroup("", []) + all_metrics = MetricGroup("", [ + Arm64Topdown(), + ]) if _args.metricgroups: print(JsonEncodeMetricGroupDescriptions(all_metrics))
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
James Clark
reviewerRe: [PATCH v9 46/48] perf jevents: Add collection of topdown like metrics for arm64
Quoted 14 lines · Ian Rogers
Metrics are created using legacy, common and recommended events. As events may be missing a TryEvent function will give None if an event is missing. To workaround missing JSON events for cortex-a53, sysfs encodings are used.
Signed-off-by Ian Rogers <[email protected]>---
An earlier review of this patch by Leo Yan is here:
https://lore.kernel.org/lkml/[email protected]/
Hopefully all corrections were made.--- tools/perf/pmu-events/arm64_metrics.py | 145 ++++++++++++++++++++++++- 1 file changed, 142 insertions(+), 3 deletions(-)
[...]
Quoted 17 lines
tools/perf/pmu-events/arm64_metrics.py +17
-0+ MetricGroup("lpm_topdown_be_bound", [ + MetricGroup("lpm_topdown_be_dtlb", [ + Metric("lpm_topdown_be_dtlb_walks", "Dtlb walks per instruction", + d_ratio(dtlb_walk, ins_ret), "walk/insn"), + Metric("lpm_topdown_be_dtlb_walk_rate", "Dtlb walks per L1D TLB access", + d_ratio(dtlb_walk, l1d_tlb) if l1d_tlb else None, "100%"), + ]) if dtlb_walk else None, + MetricGroup("lpm_topdown_be_mix", [ + Metric("lpm_topdown_be_mix_ld", "Percentage of load instructions", + d_ratio(ld_spec, inst_spec), "100%") if ld_spec else None, + Metric("lpm_topdown_be_mix_st", "Percentage of store instructions", + d_ratio(st_spec, inst_spec), "100%") if st_spec else None, + Metric("lpm_topdown_be_mix_simd", "Percentage of SIMD instructions", + d_ratio(ase_spec, inst_spec), "100%") if ase_spec else None, + Metric("lpm_topdown_be_mix_fp", + "Percentage of floating point instructions", + d_ratio(vfp_spec, inst_spec), "100%") if vfp_spec else None,
tools/perf/pmu-events/arm64_metrics.py line 116
Quoted 10 lines
tools/perf/pmu-events/arm64_metrics.py +10
-0+ Metric("lpm_topdown_be_mix_dp", + "Percentage of data processing instructions", + d_ratio(dp_spec, inst_spec), "100%") if dp_spec else None, + Metric("lpm_topdown_be_mix_crypto", + "Percentage of data processing instructions", + d_ratio(crypto_spec, inst_spec), "100%") if crypto_spec else None, + Metric( + "lpm_topdown_be_mix_br", "Percentage of branch instructions", + d_ratio(br_immed_spec + br_indirect_spec + br_ret_spec, + inst_spec), "100%") if br_immed_spec and br_indirect_spec and br_ret_spec else None,
Hi Ian,
I've been trying to engage with the team that's publishing the metrics in Arm [1] to see if there was any chance in getting some unity between these new metrics and their existing json ones. The feedback from them was that the decision to only publish metrics for certain cores is deliberate and there is no plan to change anything. The metrics there are well tested, known to be working, and usually contain workarounds for specific issues. They don't want to do "Arm wide" common metrics for existing cores as they believe it has more potential to mislead people than help.
I'm commenting on this "lpm_topdown_be_mix_br" as one example, that the equivalent Arm metric "branch_percentage" excludes br_ret_spec because br_indirect_spec also counts returns. Or on neoverse-n3 it's "PC_WRITE_SPEC / INST_SPEC".
I see that you've prefixed all the metrics so the names won't clash from Kan's feedback [2]. But it makes me wonder if at some point some kind of alias list could be implemented to override the generated metrics with hand written json ones. But by that point why not just use the same names? The Arm metric team's feedback was that there isn't really an industry standard for naming, and that differences between architectures would make it almost impossible to standardise anyway in their opinion. But here we're adding duplicate metrics with different names, where the new ones are known to have issues. It's not a great user experience IMO, but at the same time missing old cores from the Arm metrics isn't a great user experience either. I actually don't have a solution, other than to say I tried to get them to consider more unified naming.
I also have to say that I do still agree with Andi's old feedback [3] that the existing json was good enough, and maybe this isn't the right direction, although it's not very useful feedback at this point. I thought I had replied to that thread long ago, but must not have pressed send, sorry about that.
[1]: https://gitlab.arm.com/telemetry-solution/telemetry-solution/-/tree/main/data
[2]: https://lore.kernel.org/lkml/[email protected]
[3]: https://lore.kernel.org/lkml/ZeJJyCmXO9GxpDiF@tassilo/
Ian Rogers
reviewerRe: [PATCH v9 46/48] perf jevents: Add collection of topdown like metrics for arm64
Quoted 47 lines · James Clark
Quoted 14 lines · Ian Rogers
Metrics are created using legacy, common and recommended events. As events may be missing a TryEvent function will give None if an event is missing. To workaround missing JSON events for cortex-a53, sysfs encodings are used.
Signed-off-by Ian Rogers <[email protected]>---
An earlier review of this patch by Leo Yan is here:
https://lore.kernel.org/lkml/[email protected]/
Hopefully all corrections were made.--- tools/perf/pmu-events/arm64_metrics.py | 145 ++++++++++++++++++++++++- 1 file changed, 142 insertions(+), 3 deletions(-)[...]
Quoted 17 lines
tools/perf/pmu-events/arm64_metrics.py +17
-0+ MetricGroup("lpm_topdown_be_bound", [ + MetricGroup("lpm_topdown_be_dtlb", [ + Metric("lpm_topdown_be_dtlb_walks", "Dtlb walks per instruction", + d_ratio(dtlb_walk, ins_ret), "walk/insn"), + Metric("lpm_topdown_be_dtlb_walk_rate", "Dtlb walks per L1D TLB access", + d_ratio(dtlb_walk, l1d_tlb) if l1d_tlb else None, "100%"), + ]) if dtlb_walk else None, + MetricGroup("lpm_topdown_be_mix", [ + Metric("lpm_topdown_be_mix_ld", "Percentage of load instructions", + d_ratio(ld_spec, inst_spec), "100%") if ld_spec else None, + Metric("lpm_topdown_be_mix_st", "Percentage of store instructions", + d_ratio(st_spec, inst_spec), "100%") if st_spec else None, + Metric("lpm_topdown_be_mix_simd", "Percentage of SIMD instructions", + d_ratio(ase_spec, inst_spec), "100%") if ase_spec else None, + Metric("lpm_topdown_be_mix_fp", + "Percentage of floating point instructions", + d_ratio(vfp_spec, inst_spec), "100%") if vfp_spec else None,tools/perf/pmu-events/arm64_metrics.py line 116
Quoted 10 lines
tools/perf/pmu-events/arm64_metrics.py +10
-0+ Metric("lpm_topdown_be_mix_dp", + "Percentage of data processing instructions", + d_ratio(dp_spec, inst_spec), "100%") if dp_spec else None, + Metric("lpm_topdown_be_mix_crypto", + "Percentage of data processing instructions", + d_ratio(crypto_spec, inst_spec), "100%") if crypto_spec else None, + Metric( + "lpm_topdown_be_mix_br", "Percentage of branch instructions", + d_ratio(br_immed_spec + br_indirect_spec + br_ret_spec, + inst_spec), "100%") if br_immed_spec and br_indirect_spec and br_ret_spec else None,Hi Ian,
Quoted 9 lines
I've been trying to engage with the team that's publishing the metrics in Arm [1] to see if there was any chance in getting some unity between these new metrics and their existing json ones. The feedback from them was that the decision to only publish metrics for certain cores is deliberate and there is no plan to change anything. The metrics there are well tested, known to be working, and usually contain workarounds for specific issues. They don't want to do "Arm wide" common metrics for existing cores as they believe it has more potential to mislead people than help.
So this is sad, but I'll drop the patch from the series so as not to delay things and keep carrying it in Google's tree. Just looking in tools/perf/pmu-events/arch/arm64/arm there are 20 ARM models of which only neoverse models (5 of the 20) have metrics. Could ARM's metric people step up to fill the void? Models like cortex-a76 are actively sold in the Raspberry Pi 5 and yet lack metrics.
I think there has to be a rule at some point of, "don't let perfect be the enemy of good." There's no implication that ARM should maintain these metrics and they be perfect just as there isn't an implication that ARM should maintain the legacy metrics like "stalled_cycles_per_instruction":
https://web.git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/pmu-events/arch/common/common/metrics.json?h=perf-tools-next#n49
I'm guessing the cycles breakdown:
https://lore.kernel.org/lkml/[email protected]/
is okay and will keep that for ARM.
Quoted 4 lines
I'm commenting on this "
lpm_topdown_be_mix_br" as one example, that the equivalent Arm metric "branch_percentage" excludesbr_ret_specbecausebr_indirect_specalso counts returns. Or on neoverse-n3 it's "PC_WRITE_SPEC/INST_SPEC".
This is the value in upstreaming metrics like this, to bug fix. This is what has happened with the AMD and Intel metrics. I'm happy we can deliver more metrics to users on those CPUs.
Quoted 7 lines
I see that you've prefixed all the metrics so the names won't clash from Kan's feedback [2]. But it makes me wonder if at some point some kind of alias list could be implemented to override the generated metrics with hand written json ones. But by that point why not just use the same names? The Arm metric team's feedback was that there isn't really an industry standard for naming, and that differences between architectures would make it almost impossible to standardise anyway in their opinion.
So naming is always a challenge. One solution here is the ilist application. When doing the legacy event reorganization I remember you arguing that legacy events should be the norm and not the exception, but it is because of all the model quirks precisely why that doesn't work. By working through the quirks with cross platform metrics, that's value to users and not misleading. If it does mislead then that's a bug, let's fix it. Presenting users with no data isn't a fix nor being particularly helpful.
Quoted 5 lines
But here we're adding duplicate metrics with different names, where the new ones are known to have issues. It's not a great user experience IMO, but at the same time missing old cores from the Arm metrics isn't a great user experience either. I actually don't have a solution, other than to say I tried to get them to consider more unified naming.
So the lpm_ metrics are on top of whatever a vendor wants to add. There are often more than one way to compute a metric, such as memory controller counters vs l3 cache, on Intel an lpm_ metric may use uncore counters while a tma_ metric uses the cache. I don't know if sticking "ARM doesn't support this" in all the ARM lpm_ metric descriptions would mitigate your metric creators' concerns, it is something implied by Linux's licensing. We do highlight metrics that contain experimental events, such as on Intel, should be considered similarly experimental.
Quoted 5 lines
I also have to say that I do still agree with Andi's old feedback [3] that the existing json was good enough, and maybe this isn't the right direction, although it's not very useful feedback at this point. I thought I had replied to that thread long ago, but must not have pressed send, sorry about that.
So having handwritten long metrics in json it's horrid, Having been there I wouldn't want to be doing more of it. No comments, no line breaks, huge potential for typos, peculiar rules on when commas are allowed (so removing a line breaks parsing), .. This is why we have make_legacy_cache.py
https://web.git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/pmu-events/make_legacy_cache.py?h=perf-tools-next
writing 1216 legacy cache event descriptions (7266 lines of json) vs 129 lines of python. I'm going to be on team python all day long. In terms of the Linux build, I don't think there's a reasonable alternative language.
Thanks,
Ian
James Clark
reviewerRe: [PATCH v9 46/48] perf jevents: Add collection of topdown like metrics for arm64
Quoted 58 lines · Ian Rogers
Quoted 47 lines · James Clark
Quoted 14 lines · Ian Rogers
Metrics are created using legacy, common and recommended events. As events may be missing a TryEvent function will give None if an event is missing. To workaround missing JSON events for cortex-a53, sysfs encodings are used.
Signed-off-by Ian Rogers <[email protected]>---
An earlier review of this patch by Leo Yan is here:
https://lore.kernel.org/lkml/[email protected]/
Hopefully all corrections were made.--- tools/perf/pmu-events/arm64_metrics.py | 145 ++++++++++++++++++++++++- 1 file changed, 142 insertions(+), 3 deletions(-)[...]
Quoted 17 lines
tools/perf/pmu-events/arm64_metrics.py +17
-0+ MetricGroup("lpm_topdown_be_bound", [ + MetricGroup("lpm_topdown_be_dtlb", [ + Metric("lpm_topdown_be_dtlb_walks", "Dtlb walks per instruction", + d_ratio(dtlb_walk, ins_ret), "walk/insn"), + Metric("lpm_topdown_be_dtlb_walk_rate", "Dtlb walks per L1D TLB access", + d_ratio(dtlb_walk, l1d_tlb) if l1d_tlb else None, "100%"), + ]) if dtlb_walk else None, + MetricGroup("lpm_topdown_be_mix", [ + Metric("lpm_topdown_be_mix_ld", "Percentage of load instructions", + d_ratio(ld_spec, inst_spec), "100%") if ld_spec else None, + Metric("lpm_topdown_be_mix_st", "Percentage of store instructions", + d_ratio(st_spec, inst_spec), "100%") if st_spec else None, + Metric("lpm_topdown_be_mix_simd", "Percentage of SIMD instructions", + d_ratio(ase_spec, inst_spec), "100%") if ase_spec else None, + Metric("lpm_topdown_be_mix_fp", + "Percentage of floating point instructions", + d_ratio(vfp_spec, inst_spec), "100%") if vfp_spec else None,tools/perf/pmu-events/arm64_metrics.py line 116
Quoted 10 lines
tools/perf/pmu-events/arm64_metrics.py +10
-0+ Metric("lpm_topdown_be_mix_dp", + "Percentage of data processing instructions", + d_ratio(dp_spec, inst_spec), "100%") if dp_spec else None, + Metric("lpm_topdown_be_mix_crypto", + "Percentage of data processing instructions", + d_ratio(crypto_spec, inst_spec), "100%") if crypto_spec else None, + Metric( + "lpm_topdown_be_mix_br", "Percentage of branch instructions", + d_ratio(br_immed_spec + br_indirect_spec + br_ret_spec, + inst_spec), "100%") if br_immed_spec and br_indirect_spec and br_ret_spec else None,Hi Ian,
Quoted 9 lines
I've been trying to engage with the team that's publishing the metrics in Arm [1] to see if there was any chance in getting some unity between these new metrics and their existing json ones. The feedback from them was that the decision to only publish metrics for certain cores is deliberate and there is no plan to change anything. The metrics there are well tested, known to be working, and usually contain workarounds for specific issues. They don't want to do "Arm wide" common metrics for existing cores as they believe it has more potential to mislead people than help.
Quoted 6 lines
So this is sad, but I'll drop the patch from the series so as not to delay things and keep carrying it in Google's tree. Just looking in
tools/perf/pmu-events/arch/arm64/armthere are 20 ARM models of which only neoverse models (5 of the 20) have metrics. Could ARM's metric people step up to fill the void? Models like cortex-a76 are actively sold in the Raspberry Pi 5 and yet lack metrics.
I don't know of any plan to do so, although it's obviously possible. I think if some kind of official request came through some customer (including Google) then it could happen. But it looks like nobody is really asking for it so it wasn't done and that's how the decision was made.
Quoted 8 lines
I think there has to be a rule at some point of, "don't let perfect be the enemy of good." There's no implication that ARM should maintain these metrics and they be perfect just as there isn't an implication that ARM should maintain the legacy metrics like "
stalled_cycles_per_instruction":
https://web.git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/pmu-events/arch/common/common/metrics.json?h=perf-tools-next#n49
I kind of agree, I'm not saying to definitely not add these metrics. Just making sure that you're aware of the potential for some issues and conflicts going forwards.
As you say, the bugs can always be fixed. It's just that now we have to do it in two places.
Quoted 4 lines
I'm guessing the cycles breakdown:
https://lore.kernel.org/lkml/[email protected]/
is okay and will keep that for ARM.
Yes this one looks fine.
Quoted 32 lines
Quoted 4 lines
I'm commenting on this "
lpm_topdown_be_mix_br" as one example, that the equivalent Arm metric "branch_percentage" excludesbr_ret_specbecausebr_indirect_specalso counts returns. Or on neoverse-n3 it's "PC_WRITE_SPEC/INST_SPEC".This is the value in upstreaming metrics like this, to bug fix. This is what has happened with the AMD and Intel metrics. I'm happy we can deliver more metrics to users on those CPUs.
Quoted 7 lines
I see that you've prefixed all the metrics so the names won't clash from Kan's feedback [2]. But it makes me wonder if at some point some kind of alias list could be implemented to override the generated metrics with hand written json ones. But by that point why not just use the same names? The Arm metric team's feedback was that there isn't really an industry standard for naming, and that differences between architectures would make it almost impossible to standardise anyway in their opinion.
So naming is always a challenge. One solution here is the ilist application. When doing the legacy event reorganization I remember you arguing that legacy events should be the norm and not the exception, but it is because of all the model quirks precisely why that doesn't work. By working through the quirks with cross platform metrics, that's value to users and not misleading. If it does mislead then that's a bug, let's fix it. Presenting users with no data isn't a fix nor being particularly helpful.
Quoted 5 lines
But here we're adding duplicate metrics with different names, where the new ones are known to have issues. It's not a great user experience IMO, but at the same time missing old cores from the Arm metrics isn't a great user experience either. I actually don't have a solution, other than to say I tried to get them to consider more unified naming.
Quoted 6 lines
So the
lpm_metrics are on top of whatever a vendor wants to add. There are often more than one way to compute a metric, such as memory controller counters vs l3 cache, on Intel anlpm_metric may use uncore counters while atma_metric uses the cache. I don't know if sticking "ARM doesn't support this" in all the ARMlpm_metric descriptions would mitigate your metric creators' concerns, it is
Probably not necessary to do that no, but maybe some kind of header or documentation about what "lpm_" stands for would be good. I don't know if we'd expect users to know about the subtle differences or what lpm stands for?
Quoted 21 lines
something implied by Linux's licensing. We do highlight metrics that contain experimental events, such as on Intel, should be considered similarly experimental.
Quoted 5 lines
I also have to say that I do still agree with Andi's old feedback [3] that the existing json was good enough, and maybe this isn't the right direction, although it's not very useful feedback at this point. I thought I had replied to that thread long ago, but must not have pressed send, sorry about that.
So having handwritten long metrics in json it's horrid, Having been there I wouldn't want to be doing more of it. No comments, no line breaks, huge potential for typos, peculiar rules on when commas are allowed (so removing a line breaks parsing), .. This is why we have
make_legacy_cache.py
https://web.git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/pmu-events/make_legacy_cache.py?h=perf-tools-next
writing 1216 legacy cache event descriptions (7266 lines of json) vs 129 lines of python. I'm going to be on team python all day long. In terms of the Linux build, I don't think there's a reasonable alternative language.
Quoted 3 lines
Thanks,
Ian
Comments and line breaks could have been worked around by bodging the json parser or moving to yaml or something. But yes having a big formula in a one line string with conditionals wasn't that great. The biggest problem I think this fixes is the fact that it fills in missing metrics.
It does however make it harder to autogenerate using any formulas published elsewhere. Whether that's actually an issue or not I don't know because these are going to be changing much less frequently than something like the event names that are published in json (and there's fewer metrics than events as well).
Ian Rogers
reviewerPATCH v9 47/48 perf jevents: Add cycles breakdown metric for arm64/AMD/Intel
In reply to Ian RogersBreakdown cycles to user, kernel and guest. Add a common_metrics.py file for such metrics.
--- tools/perf/pmu-events/Build | 2 +- tools/perf/pmu-events/amd_metrics.py | 2 ++ tools/perf/pmu-events/arm64_metrics.py | 2 ++ tools/perf/pmu-events/common_metrics.py | 19 +++++++++++++++++++ tools/perf/pmu-events/intel_metrics.py | 2 ++ 5 files changed, 26 insertions(+), 1 deletion(-) create mode 100644 tools/perf/pmu-events/common_metrics.py
tools/perf/pmu-events/Build +1 -1
@@ -44,7 +44,7 @@ $(LEGACY_CACHE_JSON): $(LEGACY_CACHE_PY) $(call rule_mkdir) $(Q)$(call echo-cmd,gen)$(PYTHON) $(LEGACY_CACHE_PY) > $@ -GEN_METRIC_DEPS := pmu-events/metric.py +GEN_METRIC_DEPS := pmu-events/metric.py pmu-events/common_metrics.py # Generate AMD Json ZENS = $(shell ls -d pmu-events/arch/x86/amdzen*)
tools/perf/pmu-events/amd_metrics.py +2 -0
@@ -4,6 +4,7 @@ import argparse import math import os from typing import Optional +from common_metrics import Cycles from metric import (d_ratio, has_event, max, Event, JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, Literal, LoadEvents, Metric, MetricGroup, Select)
@@ -474,6 +475,7 @@ def main() -> None: AmdItlb(), AmdLdSt(), AmdUpc(), + Cycles(), Idle(), Rapl(), UncoreL3(),
tools/perf/pmu-events/arm64_metrics.py +2 -0
@@ -3,6 +3,7 @@ import argparse import os from typing import Optional +from common_metrics import Cycles from metric import (d_ratio, Event, JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, LoadEvents, Metric, MetricGroup)
@@ -173,6 +174,7 @@ def main() -> None: all_metrics = MetricGroup("", [ Arm64Topdown(), + Cycles(), ]) if _args.metricgroups:
tools/perf/pmu-events/common_metrics.py new file +19 -0
@@ -0,0 +1,19 @@ +# SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause) +from metric import (d_ratio, Event, Metric, MetricGroup) + + +def Cycles() -> MetricGroup: + cyc_k = Event("cpu\\-cycles:kHh") # exclude user and guest + cyc_g = Event("cpu\\-cycles:G") # exclude host + cyc_u = Event("cpu\\-cycles:uH") # exclude kernel, hypervisor and guest + cyc = cyc_k + cyc_g + cyc_u + + return MetricGroup("lpm_cycles", [ + Metric("lpm_cycles_total", "Total number of cycles", cyc, "cycles"), + Metric("lpm_cycles_user", "User cycles as a percentage of all cycles", + d_ratio(cyc_u, cyc), "100%"), + Metric("lpm_cycles_kernel", "Kernel cycles as a percentage of all cycles", + d_ratio(cyc_k, cyc), "100%"), + Metric("lpm_cycles_guest", "Hypervisor guest cycles as a percentage of all cycles", + d_ratio(cyc_g, cyc), "100%"), + ], description="cycles breakdown per privilege level (users, kernel, guest)")
tools/perf/pmu-events/intel_metrics.py +2 -0
@@ -6,6 +6,7 @@ import math import os import re from typing import Optional +from common_metrics import Cycles from metric import (d_ratio, has_event, max, source_count, CheckPmu, Event, JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, Literal, LoadEvents, Metric, MetricConstraint, MetricGroup,
@@ -1095,6 +1096,7 @@ def main() -> None: LoadEvents(directory) all_metrics = MetricGroup("", [ + Cycles(), Idle(), Rapl(), Smi(),
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Ian Rogers
reviewerPATCH v9 48/48 perf jevents: Validate that all names given an Event
In reply to Ian RogersValidate they exist in a json file from one directory found from one directory above the model's json directory. This avoids broken fallback encodings being created.
--- tools/perf/pmu-events/metric.py | 36 +++++++++++++++++++++++++++++++++ 1 file changed, 36 insertions(+)
tools/perf/pmu-events/metric.py +36 -0
@@ -11,12 +11,14 @@ from typing import Dict, List, Optional, Set, Tuple, Union all_pmus = set() all_events = set() experimental_events = set() +all_events_all_models = set() def LoadEvents(directory: str) -> None: """Populate a global set of all known events for the purpose of validating Event names""" global all_pmus global all_events global experimental_events + global all_events_all_models all_events = { "context\\-switches", "cpu\\-cycles",
@@ -42,6 +44,20 @@ def LoadEvents(directory: str) -> None: # The generated directory may be the same as the input, which # causes partial json files. Ignore errors. pass + all_events_all_models = all_events.copy() + for root, dirs, files in os.walk(directory + ".."): + for filename in files: + if filename.endswith(".json"): + try: + for x in json.load(open(f"{root}/{filename}")): + if "EventName" in x: + all_events_all_models.add(x["EventName"]) + elif "ArchStdEvent" in x: + all_events_all_models.add(x["ArchStdEvent"]) + except json.decoder.JSONDecodeError: + # The generated directory may be the same as the input, which + # causes partial json files. Ignore errors. + pass def CheckPmu(name: str) -> bool:
@@ -64,6 +80,25 @@ def CheckEvent(name: str) -> bool: return name in all_events +def CheckEveryEvent(*names: str) -> None: + """Check all the events exist in at least one json file""" + global all_events_all_models + if len(all_events_all_models) == 0: + assert len(names) == 1, f"Cannot determine valid events in {names}" + # No events loaded so assume any event is good. + return + + for name in names: + # Remove trailing modifier. + if ':' in name: + name = name[:name.find(':')] + elif '/' in name: + name = name[:name.find('/')] + if any([name.startswith(x) for x in ['amd', 'arm', 'cpu', 'msr', 'power']]): + continue + if name not in all_events_all_models: + raise Exception(f"Is {name} a named json event?") + def IsExperimentalEvent(name: str) -> bool: global experimental_events
@@ -403,6 +438,7 @@ class Event(Expression): def __init__(self, *args: str): error = "" + CheckEveryEvent(*args) for name in args: if CheckEvent(name): self.name = _FixEscapes(name)
Signature · 1 line
2.52.0.158.g65b55ccf14-goog
Quoted 12 lines · Ian Rogers
Metrics in the perf tool come in via json. Json doesn't allow comments, line breaks, etc. making it an inconvenient way to write metrics. Further, it is useful to detect when writing a metric that the event specified is supported within the event json for a model. From the metric python code Event(s) are used, with fallback events provided, if no event is found then an exception is thrown and that can either indicate a failure or an unsupported model. To avoid confusion all the metrics and their metricgroups are prefixed with '
lpm_', where LPM is an abbreviation of Linux Perf Metric. While extra characters aren't ideal, this separates the metrics from other vendor provided metrics.
Quoted 4 lines
* The first 14 patches introduce infrastructure and fixes for the addition of metrics written in python for Arm64, AMD Zen and Intel CPUs. The
ilist.pyand perf python module are fixed to work better with metrics on hybrid architectures.
I've applied the first 12 patches to perf-tools-next, thanks! Namhyung
Quoted 215 lines
* The next 9 patches generate additional metrics for AMD zen. Rapl and Idle metrics aren't specific to AMD but are placed here for ease and convenience. Uncore L3 metrics are added along with the majority of core metrics.
* The next 20 patches add additional metrics for Intel. Rapl and Idle metrics aren't specific to Intel but are placed here for ease and convenience. Smi and tsx metrics are added so they can be dropped from the per model json files. There are four uncore sets of metrics and eleven core metrics. Add a CheckPmu function to metric to simplify detecting the presence of hybrid PMUs in events. Metrics with experimental events are flagged as experimental in their description.
* The next 2 patches add additional metrics for Arm64, where the topdown set decomposes yet further. The metrcs primarily use json events, where the json contains architecture standard events. Not all events are in the json, such as for a53 where the events are in sysfs. Workaround this by adding the sysfs events to the metrics but longer-term such events should be added to the json.
* The final patch validates that all events provided to an Event object exist in a json file somewhere. This is to avoid mistakes like unfortunate typos.
This series has benefitted from the input of Leo Yan
<[email protected]>, Sandipan Das <[email protected]>, Thomas Falcon <[email protected]> and Perry Taylor <[email protected]>.v9. Drop (for now) 4 AMD sets of metrics for additional follow up. Add
reviewed-by tags from Sandipan Das (AMD) and tested-by tags from Thomas Falcon (Intel).v8. Combine the previous 4 series for clarity. Rebase on top of the
more recent legacy metric and event changes. Make the python more pep8 and pylint compliant.Foundations:
v6. Fix issue with '\-' escape not being '\\-' (reported-by Sandipan
Das <[email protected]>) which didn't alter the generated json.
https://lore.kernel.org/lkml/[email protected]/v5. Rebase on top of legacy hardware/cache changes that now generate
events using python:
https://lore.kernel.org/lkml/[email protected]/
the v5 series is:
https://lore.kernel.org/lkml/[email protected]/v4. Rebase and small Build/Makefile tweak
https://lore.kernel.org/lkml/[email protected]/v3. Some code tidying, make the input directory a command line
argument, but no other functional or output changes.
https://lore.kernel.org/lkml/[email protected]/v2. Fixes two type issues in the python code but no functional or
output changes. https://lore.kernel.org/lkml/[email protected]/v1. https://lore.kernel.org/lkml/[email protected]/
AMD:
v6. Fix issue with '\-' escape not being '\\-' (reported-by Sandipan
Das <[email protected]>) which didn't alter the generated json.
https://lore.kernel.org/lkml/[email protected]/v5. Rebase. Add uop cache hit/miss rates patch. Prefix all metric
names withlpm_(short for Linux Perf Metric) so that python generated metrics are clearly namespaced.
https://lore.kernel.org/lkml/[email protected]/v4. Rebase.
https://lore.kernel.org/lkml/[email protected]/v3. Some minor code cleanup changes.
https://lore.kernel.org/lkml/[email protected]/v2. Drop the cycles breakdown in favor of having it as a common
metric, suggested by Kan Liang <[email protected]>.
https://lore.kernel.org/lkml/[email protected]/v1. https://lore.kernel.org/lkml/[email protected]/
Intel:
v6. Fix issue with '\-' escape not being '\\-' (reported-by Sandipan
Das <[email protected]>) which didn't alter the generated json.
https://lore.kernel.org/lkml/[email protected]/v5. Rebase. Fix description for smi metric (Kan). Prefix all metric
names withlpm_(short for Linux Perf Metric) so that python generated metrics are clearly namespaced. Kan requested a namespace in his review:
https://lore.kernel.org/lkml/[email protected]/
The v5 series is:
https://lore.kernel.org/lkml/[email protected]/v4. Experimental metric descriptions. Add mesh bandwidth metric. Rebase.
https://lore.kernel.org/lkml/[email protected]/v3. Swap tsx and CheckPMU patches that were in the wrong order. Some
minor code cleanup changes. Drop reference to merged fix for umasks/occ_sel in PCU events and for cstate metrics.
https://lore.kernel.org/lkml/[email protected]/v2. Drop the cycles breakdown in favor of having it as a common
metric, spelling and other improvements suggested by Kan Liang <[email protected]>.
https://lore.kernel.org/lkml/[email protected]/v1. https://lore.kernel.org/lkml/[email protected]/
ARM:
v7. Switch a use of cycles to cpu-cycles due to ARM having too manycycles events. https://lore.kernel.org/lkml/[email protected]/v6. Fix issue with '\-' escape not being '\\-' (reported-by Sandipan
Das <[email protected]>) which didn't alter the generated json.
https://lore.kernel.org/lkml/[email protected]/v5. Rebase. Address review comments from Leo Yan
<[email protected]>. Prefix all metric names withlpm_(short for Linux Perf Metric) so that python generated metrics are clearly namespaced. Use cpu-cycles rather than cycles legacy event for cycles metrics to avoid confusion with ARM PMUs. Add patch that checks events to ensure all possible event names are present in at least one json file.
https://lore.kernel.org/lkml/[email protected]/v4. Tweak to build dependencies and rebase.
https://lore.kernel.org/lkml/[email protected]/v3. Some minor code cleanup changes.
https://lore.kernel.org/lkml/[email protected]/v2. The cycles metrics are now made common and shared with AMD and
Intel, suggested by Kan Liang <[email protected]>. This assumes these patches come after the AMD and Intel sets.
https://lore.kernel.org/lkml/[email protected]/v1. https://lore.kernel.org/lkml/[email protected]/
Ian Rogers (48):
perf python: Correct copying ofmetric_leaderin an evsel perf ilist: Be tolerant of reading a metric on the wrong CPU perf jevents: Allow multiple metricgroups.json files perf jevents: Update metric constraint support
perf jevents: Add descriptions to metricgroup abstraction perf jevents: Allow metric groups not to be named
perf jevents: Support parsing negative exponents
perf jevents: Term list fix in event parsing
perf jevents: Add threshold expressions to Metric
perf jevents: Move json encoding to its own functions perf jevents: Drop duplicate pending metrics
perf jevents: Skip optional metrics in metric group list perf jevents: Build support for generating metrics from python perf jevents: Add load event json to verify and allow fallbacks perf jevents: Add RAPL event metric for AMD zen models perf jevents: Add idle metric for AMD zen models
perf jevents: Add upc metric for uops per cycle for AMD perf jevents: Add br metric group for branch statistics on AMD perf jevents: Add itlb metric group for AMD
perf jevents: Add dtlb metric group for AMD
perf jevents: Add uncore l3 metric group for AMD
perf jevents: Add load store breakdown metrics ldst for AMD perf jevents: Add context switch metrics for AMD
perf jevents: Add RAPL metrics for all Intel models
perf jevents: Add idle metric for Intel models
perf jevents: Add CheckPmu to see if a PMU is in loaded json events perf jevents: Add smi metric group for Intel models
perf jevents: Mark metrics with experimental events as experimental perf jevents: Add tsx metric group for Intel models
perf jevents: Add br metric group for branch statistics on Intel perf jevents: Add software prefetch (swpf) metric group for Intel perf jevents: Add ports metric group giving utilization on Intel perf jevents: Add L2 metrics for Intel
perf jevents: Add load store breakdown metrics ldst for Intel perf jevents: Add ILP metrics for Intel
perf jevents: Add context switch metrics for Intel
perf jevents: Add FPU metrics for Intel
perf jevents: Add Miss Level Parallelism (MLP) metric for Intel perf jevents: Addmem_bwmetric for Intel
perf jevents: Add local/remote "mem" breakdown metrics for Intel perf jevents: Add dir breakdown metrics for Intel
perf jevents: Add C-State metrics from the PCU PMU for Intel perf jevents: Add local/remote miss latency metrics for Intel perf jevents: Addupi_bwmetric for Intel
perf jevents: Add mesh bandwidth saturation metric for Intel perf jevents: Add collection of topdown like metrics for arm64 perf jevents: Add cycles breakdown metric for arm64/AMD/Intel perf jevents: Validate that all names given an Eventtools/perf/.gitignore | 5 + tools/perf/Makefile.perf | 2 + tools/perf/pmu-events/Build | 51 +- tools/perf/pmu-events/amd_metrics.py | 491 ++++++++++ tools/perf/pmu-events/arm64_metrics.py | 187 ++++ tools/perf/pmu-events/common_metrics.py | 19 + tools/perf/pmu-events/intel_metrics.py | 1129 +++++++++++++++++++++++ tools/perf/pmu-events/jevents.py | 7 +- tools/perf/pmu-events/metric.py | 256 ++++- tools/perf/pmu-events/metric_test.py | 4 + tools/perf/python/ilist.py | 8 +- tools/perf/util/evsel.c | 1 + tools/perf/util/python.c | 82 +- 13 files changed, 2188 insertions(+), 54 deletions(-) create mode 100755 tools/perf/pmu-events/amd_metrics.py create mode 100755 tools/perf/pmu-events/arm64_metrics.py create mode 100644 tools/perf/pmu-events/common_metrics.py create mode 100755 tools/perf/pmu-events/intel_metrics.pySignature · 1 line
2.52.0.158.g65b55ccf14-goog