| |
| """Summarize a tracy ops_perf_results CSV of scripts/prof.py: the LAST forward (segmented on the |
| ingest tilize op), op count, kernel-sum, op-to-op gap sum, time per op code and per ViT/head region. |
| usage: python scripts/prof_summary.py <ops_perf_results.csv> [--out trimmed.csv]""" |
| import csv |
| import sys |
| from collections import OrderedDict, defaultdict |
|
|
| path = sys.argv[1] |
| rows = list(csv.DictReader(open(path))) |
| rows = [r for r in rows if r.get("DEVICE KERNEL DURATION [ns]")] |
| starts = [i for i, r in enumerate(rows) if r["OP CODE"].startswith("Tilize")] |
| if not starts: |
| |
| starts = [i for i, r in enumerate(rows[:-1]) if r["OP CODE"].startswith("GenericOp") |
| and rows[i + 1]["OP CODE"].startswith("MinimalMatmul")] |
| last = rows[starts[-1]:] |
| kern = sum(float(r["DEVICE KERNEL DURATION [ns]"]) for r in last) / 1e3 |
| gaps = [float(r["OP TO OP LATENCY [ns]"] or 0) for r in last][1:] |
| print(f"forwards found: {len(starts)} ops in last forward: {len(last)}") |
| print(f"kernel sum {kern:.1f} us op-to-op gaps (excl. first) {sum(gaps)/1e3:.1f} us") |
| by = OrderedDict() |
| for r in last: |
| k = r["OP CODE"] |
| n, t = by.get(k, (0, 0.0)) |
| by[k] = (n + 1, t + float(r["DEVICE KERNEL DURATION [ns]"]) / 1e3) |
| for k, (n, t) in sorted(by.items(), key=lambda kv: -kv[1][1]): |
| print(f" {k:55s} n={n:4d} {t:9.1f} us") |
| if "--out" in sys.argv: |
| o = sys.argv[sys.argv.index("--out") + 1] |
| keep = ["OP CODE", "CORE COUNT", "MATH FIDELITY", "DEVICE KERNEL DURATION [ns]", "OP TO OP LATENCY [ns]", |
| "INPUT_0_Y_PAD[LOGICAL]", "INPUT_0_X_PAD[LOGICAL]", "INPUT_0_DATATYPE", "INPUT_0_MEMORY", |
| "INPUT_1_Y_PAD[LOGICAL]", "INPUT_1_X_PAD[LOGICAL]", "INPUT_1_DATATYPE", "OUTPUT_0_MEMORY"] |
| with open(o, "w", newline="") as f: |
| w = csv.writer(f) |
| w.writerow(["idx"] + keep) |
| for i, r in enumerate(last): |
| w.writerow([i] + [r.get(k, "") for k in keep]) |
| print("wrote", o) |
|
|
