|
18 | 18 |
|
19 | 19 | import argparse |
20 | 20 | import re |
| 21 | +import statistics |
21 | 22 |
|
22 | 23 | # Update this if you are changing the layout of the results files |
23 | 24 | INTERNAL_WORKDIR_VERSION = '2.0' |
24 | 25 |
|
| 26 | +def run_iterations(benchmark: Benchmark, env_vars, iters: int, results: dict[str, list[Result]]): |
| 27 | + for iter in range(iters): |
| 28 | + print(f"running {benchmark.name()}, iteration {iter}... ", end='', flush=True) |
| 29 | + bench_results = benchmark.run(env_vars) |
| 30 | + if bench_results is None: |
| 31 | + print(f"did not finish (OK for sycl-bench).") |
| 32 | + break |
| 33 | + |
| 34 | + for bench_result in bench_results: |
| 35 | + # TODO: report failures in markdown/html ? |
| 36 | + if not bench_result.passed: |
| 37 | + print(f"complete ({bench_result.label}: verification FAILED)") |
| 38 | + continue |
| 39 | + |
| 40 | + print(f"complete ({bench_result.label}: {bench_result.value:.3f} {bench_result.unit}).") |
| 41 | + |
| 42 | + bench_result.name = bench_result.label |
| 43 | + bench_result.lower_is_better = benchmark.lower_is_better() |
| 44 | + |
| 45 | + if bench_result.label not in results: |
| 46 | + results[bench_result.label] = [] |
| 47 | + |
| 48 | + results[bench_result.label].append(bench_result) |
| 49 | + |
| 50 | +# https://www.statology.org/modified-z-score/ |
| 51 | +def modified_z_score(values: list[float]) -> list[float]: |
| 52 | + median = statistics.median(values) |
| 53 | + mad = statistics.median([abs(v - median) for v in values]) |
| 54 | + if mad == 0: |
| 55 | + return [0] * len(values) |
| 56 | + return [(0.6745 * (v - median)) / mad for v in values] |
| 57 | + |
| 58 | +def remove_outliers(results: dict[str, list[Result]], threshold: float = 3.5) -> dict[str, list[Result]]: |
| 59 | + new_results = {} |
| 60 | + for key, rlist in results.items(): |
| 61 | + # don't eliminate outliers on first pass |
| 62 | + if len(rlist) <= options.iterations: |
| 63 | + new_results[key] = rlist |
| 64 | + continue |
| 65 | + |
| 66 | + values = [r.value for r in rlist] |
| 67 | + z_scores = modified_z_score(values) |
| 68 | + filtered_rlist = [r for r, z in zip(rlist, z_scores) if abs(z) <= threshold] |
| 69 | + |
| 70 | + if not filtered_rlist: |
| 71 | + new_results[key] = rlist |
| 72 | + else: |
| 73 | + new_results[key] = filtered_rlist |
| 74 | + |
| 75 | + return new_results |
| 76 | + |
| 77 | +def process_results(results: dict[str, list[Result]]) -> tuple[bool, list[Result]]: |
| 78 | + processed: list[Result] = [] |
| 79 | + # technically, we can detect whether result is below or above threshold per |
| 80 | + # individual result. However, we can't repeat benchmark runs with that |
| 81 | + # granularity. So we just reject all results and try again. |
| 82 | + valid_results = True # above stddev threshold |
| 83 | + |
| 84 | + for label, rlist in remove_outliers(results).items(): |
| 85 | + if (len(rlist) == 0): |
| 86 | + continue |
| 87 | + |
| 88 | + if len(rlist) == 1: |
| 89 | + processed.append(rlist[0]) |
| 90 | + continue |
| 91 | + |
| 92 | + values = [r.value for r in rlist] |
| 93 | + |
| 94 | + mean_value = statistics.mean(values) |
| 95 | + stddev = statistics.stdev(values) |
| 96 | + |
| 97 | + threshold = options.stddev_threshold * mean_value |
| 98 | + |
| 99 | + if stddev > threshold: |
| 100 | + print(f"stddev {stddev} above the threshold {threshold} for {label}") |
| 101 | + valid_results = False |
| 102 | + |
| 103 | + rlist.sort(key=lambda res: res.value) |
| 104 | + median_index = len(rlist) // 2 |
| 105 | + median_result = rlist[median_index] |
| 106 | + |
| 107 | + # only override the stddev if not already set |
| 108 | + if median_result.stddev == 0.0: |
| 109 | + median_result.stddev = stddev |
| 110 | + |
| 111 | + processed.append(median_result) |
| 112 | + |
| 113 | + return valid_results, processed |
| 114 | + |
25 | 115 | def main(directory, additional_env_vars, save_name, compare_names, filter): |
26 | 116 | prepare_workdir(directory, INTERNAL_WORKDIR_VERSION) |
27 | 117 |
|
@@ -65,36 +155,14 @@ def main(directory, additional_env_vars, save_name, compare_names, filter): |
65 | 155 | for benchmark in benchmarks: |
66 | 156 | try: |
67 | 157 | merged_env_vars = {**additional_env_vars} |
68 | | - iteration_results = [] |
69 | | - iterations = options.iterations if not benchmark.ignore_iterations() else 1 |
70 | | - for iter in range(iterations): |
71 | | - print(f"running {benchmark.name()}, iteration {iter}... ", end='', flush=True) |
72 | | - bench_results = benchmark.run(merged_env_vars) |
73 | | - if bench_results is not None: |
74 | | - for bench_result in bench_results: |
75 | | - if bench_result.passed: |
76 | | - print(f"complete ({bench_result.label}: {bench_result.value:.3f} {bench_result.unit}).") |
77 | | - else: |
78 | | - print(f"complete ({bench_result.label}: verification FAILED)") |
79 | | - iteration_results.append(bench_result) |
80 | | - else: |
81 | | - print(f"did not finish (OK for sycl-bench).") |
| 158 | + intermediate_results: dict[str, list[Result]] = {} |
| 159 | + processed: list[Result] = [] |
| 160 | + for _ in range(5): |
| 161 | + run_iterations(benchmark, merged_env_vars, options.iterations, intermediate_results) |
| 162 | + valid, processed = process_results(intermediate_results) |
| 163 | + if valid: |
82 | 164 | break |
83 | | - |
84 | | - if len(iteration_results) == 0: |
85 | | - continue |
86 | | - |
87 | | - for label in set([result.label for result in iteration_results]): |
88 | | - label_results = [result for result in iteration_results if result.label == label and result.passed == True] |
89 | | - if len(label_results) > 0: |
90 | | - label_results.sort(key=lambda res: res.value) |
91 | | - median_index = len(label_results) // 2 |
92 | | - median_result = label_results[median_index] |
93 | | - |
94 | | - median_result.name = label |
95 | | - median_result.lower_is_better = benchmark.lower_is_better() |
96 | | - |
97 | | - results.append(median_result) |
| 165 | + results += processed |
98 | 166 | except Exception as e: |
99 | 167 | if options.exit_on_failure: |
100 | 168 | raise e |
@@ -164,14 +232,15 @@ def validate_and_parse_env_args(env_args): |
164 | 232 | parser.add_argument("--env", type=str, help='Use env variable for a benchmark run.', action="append", default=[]) |
165 | 233 | parser.add_argument("--save", type=str, help='Save the results for comparison under a specified name.') |
166 | 234 | parser.add_argument("--compare", type=str, help='Compare results against previously saved data.', action="append", default=["baseline"]) |
167 | | - parser.add_argument("--iterations", type=int, help='Number of times to run each benchmark to select a median value.', default=5) |
168 | | - parser.add_argument("--timeout", type=int, help='Timeout for individual benchmarks in seconds.', default=600) |
| 235 | + parser.add_argument("--iterations", type=int, help='Number of times to run each benchmark to select a median value.', default=options.iterations) |
| 236 | + parser.add_argument("--stddev-threshold", type=float, help='If stddev % is above this threshold, rerun all iterations', default=options.stddev_threshold) |
| 237 | + parser.add_argument("--timeout", type=int, help='Timeout for individual benchmarks in seconds.', default=options.timeout) |
169 | 238 | parser.add_argument("--filter", type=str, help='Regex pattern to filter benchmarks by name.', default=None) |
170 | | - parser.add_argument("--epsilon", type=float, help='Threshold to consider change of performance significant', default=0.005) |
| 239 | + parser.add_argument("--epsilon", type=float, help='Threshold to consider change of performance significant', default=options.epsilon) |
171 | 240 | parser.add_argument("--verbose", help='Print output of all the commands.', action="store_true") |
172 | 241 | parser.add_argument("--exit-on-failure", help='Exit on first failure.', action="store_true") |
173 | 242 | parser.add_argument("--compare-type", type=str, choices=[e.value for e in Compare], help='Compare results against previously saved data.', default=Compare.LATEST.value) |
174 | | - parser.add_argument("--compare-max", type=int, help='How many results to read for comparisions', default=10) |
| 243 | + parser.add_argument("--compare-max", type=int, help='How many results to read for comparisions', default=options.compare_max) |
175 | 244 | parser.add_argument("--output-html", help='Create HTML output', action="store_true", default=False) |
176 | 245 | parser.add_argument("--output-markdown", help='Create Markdown output', action="store_true", default=True) |
177 | 246 | parser.add_argument("--dry-run", help='Do not run any actual benchmarks', action="store_true", default=False) |
|
0 commit comments