bench_guard.py

#!/usr/bin/env python3
from __future__ import annotations

import argparse
import csv
import statistics
import sys
import time

import torch


def parse_int_list(value: str) -> list[int]:
    return [int(item) for item in value.split(",") if item.strip()]


def sync() -> None:
    torch.cuda.synchronize()


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--lib", required=True)
    parser.add_argument("--device", type=int, default=0)
    parser.add_argument("--inner-loops", type=parse_int_list, default=parse_int_list("0,1,2,4,8,16,32,64"))
    parser.add_argument("--steps", type=int, default=10000)
    parser.add_argument("--warmup", type=int, default=1000)
    parser.add_argument("--rounds", type=int, default=5)
    args = parser.parse_args()

    torch.ops.load_library(args.lib)
    torch.cuda.set_device(args.device)
    tensor = torch.empty(1024, device="cuda")
    op = torch.ops.fastpt_c_overhead_mre.guard_loop

    writer = csv.writer(sys.stdout)
    writer.writerow(
        [
            "section",
            "inner_loops",
            "steps",
            "warmup",
            "rounds",
            "median_step_us",
            "mean_step_us",
            "median_per_guard_us",
        ]
    )

    for inner_loops in args.inner_loops:
        for _ in range(args.warmup):
            op(tensor, inner_loops)
        sync()

        values = []
        for _ in range(args.rounds):
            sync()
            start = time.perf_counter_ns()
            for _ in range(args.steps):
                op(tensor, inner_loops)
            sync()
            stop = time.perf_counter_ns()
            values.append((stop - start) / args.steps / 1000.0)

        median_step = statistics.median(values)
        writer.writerow(
            [
                "guard_loop",
                inner_loops,
                args.steps,
                args.warmup,
                args.rounds,
                f"{median_step:.6f}",
                f"{statistics.mean(values):.6f}",
                f"{median_step / inner_loops:.6f}" if inner_loops else "0.000000",
            ]
        )


if __name__ == "__main__":
    main()