parse_perf_metrics.py

Go to the documentation of this file.

# ===========================================
# parse_perf_metrics.py
# ===========================================
 

 
 
from pipeline.utils import safe_div
import polars as pl
import sys
 
trials = int(sys.argv[2])
 
df = pl.read_csv(sys.argv[1], has_header=False)
df.columns = ["value", "col1", "event", "timestamp", "cpu%", "derived", "label"]
 
field_map = {
    "cycles": "cycles:u",
    "instr": "instructions:u",
    "cache_loads": "cache-references:u",
    "cache_miss": "cache-misses:u",
    "l1_loads": "L1-dcache-loads:u",
    "l1_misses": "L1-dcache-load-misses:u",
    "l2_loads": "NA",
    "l2_misses": "NA",
    "l3_loads": "NA",
   "l3_misses": "NA",
    "tlb_loads": "dTLB-loads:u",
    "tlb_misses": "dTLB-load-misses:u",
    "branch_instr": "branch-instructions:u",
    "branch_misses": "branch-misses:u",
}
 
event_to_key = {v: k for k, v in field_map.items() if v != "NA"}
filtered = df.filter(pl.col("event").is_in(event_to_key.keys()))
 
# Detect non-numeric values (e.g. "<not supported>", "N/A")
to_clean = (
    pl.when(pl.col("value").cast(pl.Float64, strict=False).is_null())
      .then(pl.lit("NA"))
      .otherwise(pl.col("value").cast(pl.Utf8))  # stringify numeric
)
 
filtered = filtered.with_columns([to_clean.alias("clean_value")])
 
values = {key: "NA" for key in field_map}
 
for row in filtered.iter_rows(named=True):
    cli_key = event_to_key[row["event"]]
    values[cli_key] = row["clean_value"]
 
values["ipc"] = safe_div(values["instr"], values["cycles"])
values["miss_per_trial"] = safe_div(values["cache_miss"], trials)
values["cycles_per_trial"] = safe_div(values["cycles"], trials)
 
# Ordered output for clean downstream piping
ordered_keys = [
    "cycles", "instr", "ipc",
    "cache_loads", "cache_miss",
    "l1_loads", "l1_misses",
    "l2_loads", "l2_misses",
    "l3_loads", "l3_misses",
    "tlb_loads", "tlb_misses",
    "branch_instr", "branch_misses",
    "miss_per_trial", "cycles_per_trial"
]
 
# Debugging output
def debug_print(values: dict):
    for k in ordered_keys:
        print(f"[DEBUG] {k} = {values.get(k)}", file=sys.stderr)
 
 
print(" ".join([
    f"{k.upper()}={values[k]}"
    for k in ordered_keys
]))