parse_perf_metrics.py File Reference

CLI flag generator from perf CSV output (perf stat -x,) More...

Go to the source code of this file.

Namespaces
namespace	pipeline
namespace	pipeline.parse_perf_metrics

Functions
	pipeline.parse_perf_metrics.debug_print (dict values)

Variables
	pipeline.parse_perf_metrics.trials = int(sys.argv[2])
	pipeline.parse_perf_metrics.df = pl.read_csv(sys.argv[1], has_header=False)
	pipeline.parse_perf_metrics.columns
dict	pipeline.parse_perf_metrics.field_map
dict	pipeline.parse_perf_metrics.event_to_key = {v: k for k, v in field_map.items() if v != "NA"}
	pipeline.parse_perf_metrics.filtered = df.filter(pl.col("event").is_in(event_to_key.keys()))
tuple	pipeline.parse_perf_metrics.to_clean
dict	pipeline.parse_perf_metrics.values = {key: "NA" for key in field_map}
	pipeline.parse_perf_metrics.named
dict	pipeline.parse_perf_metrics.cli_key = event_to_key[row["event"]]
list	pipeline.parse_perf_metrics.ordered_keys

Detailed Description

CLI flag generator from perf CSV output (perf stat -x,)

Parses Linux perf stat logs in CSV format and extracts a fixed set of performance metrics. Outputs these metrics as --key value shell arguments for downstream use in pipeline scripts or shell evaluation.

Metrics are mapped from perf event names (e.g., "cycles:u") into canonical CLI keys (e.g., CYCLES, IPC). Unsupported metrics are filled as "NA". Derived values like IPC and misses per trial are computed inline using safe arithmetic fallbacks.

Output Format

CYCLES=123456789 INSTR=123456 IPC=1.23 ...

Printed as a single line, space-separated key-value flags, suitable for eval.

Example

$ eval $(python3 parse_perf_metrics.py perf_SIMD.csv 1000000)
$ echo $CYCLES

Usage

python3 parse_perf_metrics.py <perf_log.csv> <num_trials>

Arguments

• <perf_log.csv> — Path to perf CSV file (from perf stat -x,)
• <num_trials> — Number of simulation trials (used for normalization)

Note

• L2/L3 metrics are set to "NA" unless enabled manually via raw PMU events.
• Designed for use via eval in shell pipelines or programmatically via subprocess.
• Will safely skip unsupported perf fields and calculate derived metrics (e.g., IPC, miss rate).

Definition in file parse_perf_metrics.py.