File size: 10,175 Bytes
ca4edfb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 | # SPDX-FileCopyrightText: © 2023 Tenstorrent USA, Inc.
# SPDX-License-Identifier: Apache-2.0
import csv
import re
import time
from os import listdir
from os.path import isfile, join
import git
from loguru import logger
from models.perf.benchmarking_utils import BenchmarkData, BenchmarkProfiler
today = time.strftime("%Y_%m_%d")
def merge_perf_files(fname, perf_fname, expected_cols):
mypath = "./"
csvfiles = [
f
for f in listdir(mypath)
if isfile(join(mypath, f)) and re.match(f"{perf_fname}_.*_{today}.csv", f) is not None
]
repo = git.Repo(search_parent_directories=True)
merge_res = open(fname, "w")
if not repo.head.is_detached:
merge_res.write(f"branch: {repo.active_branch} \n")
else:
merge_res.write(f"branch: detached HEAD \n")
merge_res.write(f"hash: {repo.head.object.hexsha} \n")
cols = ", ".join(expected_cols)
merge_res.write(f"{cols} \n")
csvfiles.sort()
for csvfile in csvfiles:
row_name = csvfile.replace("perf_", "")
row_name = row_name.replace(f"{today}", "")
row_name = row_name.replace(".csv", "")
f = open(f"./{csvfile}", "r")
f.readline()
row = f.readline().strip().strip("\n")
merge_res.write(f"{row}\n")
merge_res.close()
def process_perf_results(fname, expected_cols):
with open(fname) as file:
merge_res = csv.reader(file, skipinitialspace=True)
logger.info(next(merge_res)[0].strip())
logger.info(next(merge_res)[0].strip())
cols = next(merge_res)
cols = [c.strip() for c in cols]
assert len(expected_cols) == len(cols), "Mismatch between expected and actual columns in perf CSV"
for expected_c, c in zip(expected_cols, cols):
assert expected_c == c, f"Expected column {expected_c} but instead got {c}"
merge_res = list(merge_res)
return cols, merge_res
def check_perf_results(
fname, expected_cols, check_cols, table_width=110, float_precision=4, lower_threshold=0.80, upper_threshold=1.0
):
red_color_code = "\033[91m"
green_color_code = "\033[92m"
reset_color_code = "\033[0m"
bold_code = "\033[1m"
underline_code = "\033[4m"
fail_label = "[FAIL]"
pass_label = "[PASS]"
cols, merge_res = process_perf_results(fname, expected_cols)
visited_models = []
slow_measured = {col: [] for col in check_cols}
logger.info(f"Checking perf results for {fname}...")
print("=" * table_width)
print(f"{bold_code}PERFORMANCE RESULTS{reset_color_code}")
print("=" * table_width)
for models_info in merge_res:
dict_info = {name: value for name, value in zip(cols, models_info)}
model_name = f"{dict_info['Model']}_{dict_info['Setting']}"
visited_models.append(model_name)
print(f"\n{bold_code}{underline_code}Model: {dict_info['Model']}{reset_color_code}")
print(f"{bold_code}Setting: {dict_info['Setting']}{reset_color_code}")
print("-" * 60)
highlight = False
for col in check_cols:
model_expected_col = float(dict_info[f"Expected {col}"])
model_measured_col = float(dict_info[col])
low_tolerance = (
model_expected_col * lower_threshold
) # fail if model is too fast so we can update the expected values
high_tolerance = model_expected_col * upper_threshold # upper bound for regression
if model_measured_col > high_tolerance:
slow_measured[col].append((model_name, model_measured_col, model_expected_col))
regression_pct = ((model_measured_col - model_expected_col) / model_expected_col) * 100
print(
f" FAIL {col}: {red_color_code}{model_measured_col:.{float_precision}f} > {high_tolerance:.{float_precision}f} ({regression_pct:.1f}% worse than expected - performance regression){reset_color_code}"
)
highlight = True
elif model_measured_col < low_tolerance:
slow_measured[col].append((model_name, model_measured_col, model_expected_col))
improvement_pct = ((model_expected_col - model_measured_col) / model_expected_col) * 100
print(
f" FAIL {col}: {red_color_code}{model_measured_col:.{float_precision}f} < {low_tolerance:.{float_precision}f} ({improvement_pct:.1f}% better than expected - update baseline){reset_color_code}"
)
highlight = True
else:
print(
f" PASS {col}: {model_measured_col:.{float_precision}f} (within tolerance: {low_tolerance:.{float_precision}f} - {high_tolerance:.{float_precision}f})"
)
if highlight:
print(f" {red_color_code}This model has performance issues{reset_color_code}")
else:
print(f" {green_color_code}All performance targets met{reset_color_code}")
print("\n" + "=" * table_width)
print(f"{bold_code}SUMMARY{reset_color_code}")
print("=" * table_width)
all_passed = True
for col in slow_measured:
if slow_measured[col]:
all_passed = False
print(f"{red_color_code}{fail_label} The following models failed '{col}' checks:{reset_color_code}")
for model_name, measured, expected in slow_measured[col]:
tolerance_range = f"[{expected * lower_threshold:.{float_precision}f}, {expected * upper_threshold:.{float_precision}f}]"
if measured > expected * upper_threshold:
regression_pct = ((measured - expected) / expected) * 100
status = f"{regression_pct:.1f}% regression"
else:
improvement_pct = ((expected - measured) / expected) * 100
status = f"{improvement_pct:.1f}% improvement"
print(
f" - {model_name}: measured={measured:.{float_precision}f}, expected={expected:.{float_precision}f}, tolerance={tolerance_range} ({status})"
)
else:
print(f"{green_color_code}{pass_label} All models meet expected '{col}' performance.{reset_color_code}")
print("=" * table_width)
if all_passed:
print(f"{green_color_code}All performance checks passed!{reset_color_code}")
else:
print(f"{red_color_code}Some performance checks failed!{reset_color_code}")
print("=" * table_width)
if any(len(slow) > 0 for slow in slow_measured.values()):
raise Exception(
"Performance checks failed - see above for models with regressions or significant improvements requiring baseline updates"
)
def prep_perf_report(
model_name: str,
batch_size: int,
inference_and_compile_time: float,
inference_time: float,
expected_compile_time: float,
expected_inference_time: float,
comments: str,
inference_time_cpu: float = None,
):
today = time.strftime("%Y_%m_%d")
def write_dict_to_file(csv_path, dict_res):
columns = ", ".join([str(d) for d in dict_res.keys()])
values = ", ".join([d for d in dict_res.values()])
with open(csv_path, "w") as csvfile:
csvfile.write(columns)
csvfile.write("\n")
csvfile.write(values)
compile_time = inference_and_compile_time - inference_time
device_throughput = "{:.4f}".format(batch_size * (1 / inference_time))
cpu_throughput = batch_size * (1 / inference_time_cpu) if inference_time_cpu else "unknown"
cpu_throughput = "{:.4f}".format(cpu_throughput) if not isinstance(cpu_throughput, str) else cpu_throughput
dict_res = {
"Model": model_name,
"Setting": comments,
"Batch": str(batch_size),
"First Run (sec)": "{:.2f}".format(inference_and_compile_time),
"Second Run (sec)": "{:.2f}".format(inference_time),
"Compile Time (sec)": "{:.2f}".format(compile_time),
"Expected Compile Time (sec)": "{:.2f}".format(expected_compile_time),
"Inference Time (sec)": "{:.4f}".format(inference_time),
"Expected Inference Time (sec)": "{:.4f}".format(expected_inference_time),
"Throughput (batch*inf/sec)": device_throughput,
"Inference Time CPU (sec)": "{:.4f}".format(inference_time_cpu) if inference_time_cpu else "unknown",
"Throughput CPU (batch*inf/sec)": cpu_throughput,
}
model_name = model_name.replace("/", "_")
csv_file = f"perf_{model_name}_{comments}_{today}.csv"
write_dict_to_file(csv_file, dict_res)
# Dummy profiler to satisfy BenchmarkData's requirements
profiler = BenchmarkProfiler()
profiler.start("run")
profiler.end("run")
step_name = "end_to_end_perf"
profiler.start(step_name)
profiler.end(step_name)
benchmark_data = BenchmarkData()
# Add measurements for the key performance metrics
# Convert string values back to float for measurements
benchmark_data.add_measurement(profiler, 0, step_name, "inference_time_s", inference_time)
benchmark_data.add_measurement(profiler, 0, step_name, "compile_time_s", compile_time)
benchmark_data.add_measurement(profiler, 0, step_name, "throughput_iter_per_s", float(device_throughput))
benchmark_data.add_measurement(profiler, 0, step_name, "first_run_s", inference_and_compile_time)
benchmark_data.add_measurement(profiler, 0, step_name, "expected_inference_time_s", expected_inference_time)
benchmark_data.add_measurement(profiler, 0, step_name, "expected_compile_time_s", expected_compile_time)
if inference_time_cpu is not None:
benchmark_data.add_measurement(profiler, 0, step_name, "inference_time_cpu_s", inference_time_cpu)
benchmark_data.add_measurement(
profiler,
0,
step_name,
"throughput_cpu_iter_per_s",
float(cpu_throughput) if cpu_throughput != "unknown" else 0.0,
)
benchmark_data.save_partial_run_json(
profiler,
run_type="end_to_end_perf",
ml_model_name=model_name,
batch_size=batch_size,
)
|