|
| 1 | +#!/usr/bin/env python3.13 |
| 2 | +""" |
| 3 | +Tesseract Decoder Benchmarker |
| 4 | +
|
| 5 | +This script automates the process of benchmarking the Tesseract decoder using hyperfine. |
| 6 | +It compares the performance of your current working directory against a baseline revision. |
| 7 | +
|
| 8 | +Basic Usage: |
| 9 | + Run the benchmarker with default settings (compares current directory against 'main'): |
| 10 | + $ ./benchmarking/benchmark.py |
| 11 | +
|
| 12 | + Run a quick benchmark (minimal shots and runs, useful for sanity checking before a long run): |
| 13 | + $ ./benchmarking/benchmark.py -q |
| 14 | +
|
| 15 | + Compare against a specific baseline revision (e.g., a specific commit or branch): |
| 16 | + $ ./benchmarking/benchmark.py -b my-feature-branch |
| 17 | +
|
| 18 | + Filter circuits by group name (e.g., only run 'surface_code' circuits) See circuits.json for available groups: |
| 19 | + $ ./benchmarking/benchmark.py -g surface_code |
| 20 | +
|
| 21 | +Benchmarking Multiple Changes: |
| 22 | + You can benchmark multiple working directories simultaneously against the baseline. |
| 23 | + This is useful if you have several different implementations across different |
| 24 | + directories that you want to compare side-by-side in a single run. |
| 25 | + |
| 26 | + To set up additional directories for your changes: |
| 27 | + - Using git: Create a new worktree. |
| 28 | + $ git worktree add ../path-to-experiment1 <branch-or-commit> |
| 29 | + - Using jj (jujutsu): Add a new workspace. |
| 30 | + $ jj workspace add ../path-to-experiment1 -r <revision> |
| 31 | + |
| 32 | + Use the -d or --dir flag for each additional directory you want to include: |
| 33 | + $ ./benchmarking/benchmark.py -d ../path-to-experiment1 -d ../path-to-experiment2 |
| 34 | + |
| 35 | + You can also provide a label for the plot by using the format label=path: |
| 36 | + $ ./benchmarking/benchmark.py -d "experiment1=../path-to-experiment1" |
| 37 | + |
| 38 | + This will benchmark the baseline, the current working directory, and the two |
| 39 | + extra directories specified, providing a single cohesive report. |
| 40 | +
|
| 41 | +Command Line Flags: |
| 42 | + -b, --baseline <rev> : Specify baseline revision (default: main). Can be a branch or commit. |
| 43 | + -d, --dir <lbl=path> : Add extra working directories to benchmark against. Format: path or label=path. Can be specified multiple times. |
| 44 | + -q, --quick : Enable quick mode (fewer shots, warmup rounds, and runs). Useful for testing. |
| 45 | + -g, --group <name> : Filter circuits to benchmark by group name (e.g. 'surface_code'). |
| 46 | + --skip-build : Skip the bazel build step (assuming binaries are already built). |
| 47 | + --loop : Continuously loop the benchmarks. Take a step away from your computer, and grab a Nuka Cola. |
| 48 | +""" |
| 49 | + |
| 50 | +import argparse |
| 51 | +import contextlib |
| 52 | +import json |
| 53 | +import logging |
| 54 | +import shutil |
| 55 | +import subprocess |
| 56 | +import sys |
| 57 | +import time |
| 58 | +from datetime import datetime |
| 59 | +from pathlib import Path |
| 60 | +from zoneinfo import ZoneInfo |
| 61 | +import plotting |
| 62 | +import workspace |
| 63 | + |
| 64 | +# Configure logging with LA timezone |
| 65 | +class Formatter(logging.Formatter): |
| 66 | + def converter(self, timestamp): |
| 67 | + dt = datetime.fromtimestamp(timestamp, tz=ZoneInfo('America/Los_Angeles')) |
| 68 | + return dt.timetuple() |
| 69 | + |
| 70 | +logger = logging.getLogger() |
| 71 | +logger.setLevel(logging.INFO) |
| 72 | +handler = logging.StreamHandler(sys.stdout) |
| 73 | +formatter = Formatter('[%(asctime)s] %(message)s', datefmt='%Y-%m-%d %H:%M:%S') |
| 74 | +handler.setFormatter(formatter) |
| 75 | +logger.addHandler(handler) |
| 76 | + |
| 77 | +def print_batch_summary(json_output_files: list[Path], circuit_names: list[str]) -> None: |
| 78 | + logger.info("===================================================") |
| 79 | + logger.info(">>> BATCH RUN SUMMARY") |
| 80 | + logger.info("===================================================") |
| 81 | + |
| 82 | + for json_file, c_name in zip(json_output_files, circuit_names): |
| 83 | + if Path(json_file).exists(): |
| 84 | + try: |
| 85 | + with open(json_file, 'r') as f: |
| 86 | + results_data = json.load(f) |
| 87 | + |
| 88 | + results_list = results_data.get('results', []) |
| 89 | + if len(results_list) >= 2: |
| 90 | + baseline_mean = results_list[0].get('mean') |
| 91 | + pwd_mean = results_list[1].get('mean') |
| 92 | + |
| 93 | + if baseline_mean is not None and pwd_mean is not None and pwd_mean > 0: |
| 94 | + speedup = baseline_mean / pwd_mean |
| 95 | + logger.info(f"Circuit: {c_name}") |
| 96 | + logger.info(f" Baseline Mean: {baseline_mean:.4f} s") |
| 97 | + logger.info(f" PWD Mean: {pwd_mean:.4f} s") |
| 98 | + logger.info(f" Speedup: {speedup:.4f}x") |
| 99 | + logger.info("---------------------------------------------------") |
| 100 | + except Exception as e: |
| 101 | + logger.error(f"Failed to parse or summarize {json_file}: {e}") |
| 102 | + |
| 103 | +def run_benchmark_batch(args: argparse.Namespace, workspaces: list[str | Path], workspace_names: list[str]) -> None: |
| 104 | + logger.info("===================================================") |
| 105 | + logger.info(">>> STARTING NEW BATCH RUN SEQUENCE") |
| 106 | + logger.info("===================================================") |
| 107 | + |
| 108 | + |
| 109 | + if args.quick: |
| 110 | + logger.info(f">>> Quick mode enabled: Reduced shots ({args.sample_num_shots}), warmup ({args.warmup_rounds}), and runs ({args.num_runs}).") |
| 111 | + |
| 112 | + la_tz = ZoneInfo('America/Los_Angeles') |
| 113 | + timestamp = datetime.now(la_tz).strftime('%Y-%m-%d_%H_%M') |
| 114 | + result_dir = Path(f"benchmarking/results/{timestamp}_{args.num_runs}") |
| 115 | + |
| 116 | + logger.info(f">>> Output directory: {result_dir}") |
| 117 | + (result_dir / "benchmark_json").mkdir(parents=True, exist_ok=True) |
| 118 | + (result_dir / "benchmark_whiskers").mkdir(parents=True, exist_ok=True) |
| 119 | + |
| 120 | + try: |
| 121 | + with open("benchmarking/circuits.json", 'r') as f: |
| 122 | + circuits_data = json.load(f) |
| 123 | + except (FileNotFoundError, json.JSONDecodeError) as e: |
| 124 | + logger.error(f"Failed to load circuits JSON: {e}") |
| 125 | + sys.exit(1) |
| 126 | + |
| 127 | + if args.group: |
| 128 | + logger.info(f">>> Filtering circuits by group: {args.group}") |
| 129 | + circuits = [c for c in circuits_data if c.get('group') == args.group] |
| 130 | + else: |
| 131 | + circuits = circuits_data |
| 132 | + |
| 133 | + json_output_files = [] |
| 134 | + circuit_names = [] |
| 135 | + |
| 136 | + tesseract_args = [ |
| 137 | + "--sample-num-shots", str(args.sample_num_shots), |
| 138 | + "--print-stats", "--threads", "48", "--beam", "5", |
| 139 | + "--no-revisit-dets", "--num-det-orders", "1", |
| 140 | + "--pqlimit", "100000", "--sample-seed", "123456" |
| 141 | + ] |
| 142 | + |
| 143 | + for circuit in circuits: |
| 144 | + c_name = circuit['name'] |
| 145 | + c_path = circuit['path'] |
| 146 | + |
| 147 | + json_file = result_dir / "benchmark_json" / f"results_{c_name}.json" |
| 148 | + whisker_file = result_dir / "benchmark_whiskers" / f"results_{c_name}.png" |
| 149 | + |
| 150 | + json_output_files.append(json_file) |
| 151 | + circuit_names.append(c_name) |
| 152 | + |
| 153 | + logger.info("---------------------------------------------------") |
| 154 | + logger.info(f">>> BENCHMARKING CIRCUIT: {c_name}") |
| 155 | + logger.info(f">>> Path: {c_path}") |
| 156 | + |
| 157 | + hyperfine_cmd = [ |
| 158 | + "hyperfine", |
| 159 | + "--warmup", str(args.warmup_rounds), |
| 160 | + "--runs", str(args.num_runs), |
| 161 | + "--export-json", str(json_file) |
| 162 | + ] |
| 163 | + |
| 164 | + for name, d in zip(workspace_names, workspaces): |
| 165 | + hyperfine_cmd.extend(["-n", name]) |
| 166 | + |
| 167 | + binary_path = Path(d) / "bazel-bin" / "src" / "tesseract" |
| 168 | + if str(d) == ".": |
| 169 | + binary_path = Path("bazel-bin") / "src" / "tesseract" |
| 170 | + |
| 171 | + cmd_for_binary = f"{binary_path} --circuit '{c_path}' " + " ".join(tesseract_args) |
| 172 | + hyperfine_cmd.append(cmd_for_binary) |
| 173 | + |
| 174 | + workspace.run_cmd(hyperfine_cmd) |
| 175 | + |
| 176 | + plotting.plot_benchmark_results(json_file=str(json_file), labels=workspace_names, output_file=str(whisker_file)) |
| 177 | + |
| 178 | + print_batch_summary(json_output_files, circuit_names) |
| 179 | + logger.info(f">>> Batch Run Complete! Results saved in: {result_dir}") |
| 180 | + |
| 181 | +def main() -> None: |
| 182 | + parser = argparse.ArgumentParser(description="Benchmark tesseract decoder using hyperfine.") |
| 183 | + parser.add_argument("-b", "--baseline", default="main", help="Specify baseline revision (default: main)") |
| 184 | + parser.add_argument("-d", "--dir", action="append", default=[], help="Add extra working directories to benchmark against. Format: path or label=path. Can be specified multiple times.") |
| 185 | + parser.add_argument("--skip-build", action="store_true", help="Skip the bazel build step") |
| 186 | + parser.add_argument("--loop", action="store_true", help="Loop runs rather than running once.") |
| 187 | + |
| 188 | + parser.add_argument("-q", "--quick", action="store_true", help="Enable quick mode (fewer shots/runs)") |
| 189 | + parser.add_argument("-g", "--group", default="", help="Filter circuits by group name") |
| 190 | + parser.add_argument("--shots", type=int, default=5000, help="Override the default sample-num-shots (mutually exclusive with -q)") |
| 191 | + parser.add_argument("--warmup", type=int, default=15, help="Override the default warmup-rounds (mutually exclusive with -q)") |
| 192 | + parser.add_argument("--runs", type=int, default=50, help="Override the default num-runs (mutually exclusive with -q)") |
| 193 | + |
| 194 | + args = parser.parse_args() |
| 195 | + |
| 196 | + if args.quick and (args.shots != 5000 or args.warmup != 15 or args.runs != 50): |
| 197 | + parser.error("-q/--quick cannot be used with --shots, --warmup, or --runs") |
| 198 | + |
| 199 | + args.sample_num_shots = 500 if args.quick else args.shots |
| 200 | + args.warmup_rounds = 1 if args.quick else args.warmup |
| 201 | + args.num_runs = 2 if args.quick else args.runs |
| 202 | + |
| 203 | + baseline_dir = "../baseline_bench_tmp" |
| 204 | + vcs = workspace.check_vcs() |
| 205 | + if not vcs: |
| 206 | + logger.error("Error: Neither a jj nor git repository detected.") |
| 207 | + sys.exit(1) |
| 208 | + with workspace.managed_baseline(baseline_dir, args.baseline, vcs): |
| 209 | + extra_workspaces = [] |
| 210 | + extra_names = [] |
| 211 | + for d in args.dir: |
| 212 | + if '=' in d: |
| 213 | + lbl, pth = d.split('=', 1) |
| 214 | + extra_names.append(lbl) |
| 215 | + extra_workspaces.append(pth) |
| 216 | + else: |
| 217 | + extra_names.append(Path(d).name) |
| 218 | + extra_workspaces.append(d) |
| 219 | + |
| 220 | + workspaces = [baseline_dir, "."] + extra_workspaces |
| 221 | + workspace_names = ["baseline", "pwd"] + extra_names |
| 222 | + |
| 223 | + workspace.build_all(workspaces, args.skip_build) |
| 224 | + |
| 225 | + if args.loop: |
| 226 | + while True: |
| 227 | + run_benchmark_batch(args, workspaces, workspace_names) |
| 228 | + logger.info(">>> Restarting in 5 seconds... (Press Ctrl+C to stop)") |
| 229 | + time.sleep(5) |
| 230 | + workspace.build_all(workspaces, args.skip_build) |
| 231 | + else: |
| 232 | + run_benchmark_batch(args, workspaces, workspace_names) |
| 233 | + |
| 234 | +if __name__ == "__main__": |
| 235 | + main() |
0 commit comments