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
|
#!/usr/bin/env python3
"""the benchmark suite behind `make bench`, `make bench-save` and `make bench-check`.
usage: tests/bench/suite.py run [--zy PATH] [--out FILE]
tests/bench/suite.py check [--zy PATH] [--baseline FILE] [--out FILE]
it wraps tests/bench/bench.sh, which times the kernels of tests/programs/bench.zc
(collatz, sieve, digits, fib) plain and o3 on each backend it can build, and adds the
256-word pointer sweep of tests/bench/sweep.zc, checked (plain column) and unbounded,
timed here on build/bin/zy the same way: the slope between 1 and N repetitions, the
median of TRIALS slopes. bench.sh itself is run as it is (it needs bc).
one row per backend and kernel. columns: plain ms and o3 ms per repetition; checks, the
number of bounds checks the compiler keeps (from --explain-checks, which does not exist
yet: '-'); peak heap words and gc mark time, which the runtime does not expose: '-'.
check compares with the baseline (tests/bench.baseline): each of the plain and o3
columns fails on its own when it is more than TOLERANCE (10%) slower, and a row fails
when both sides have a checks count and it rose. when anything looks slower the whole
measurement is repeated, up to RETRIES more times, and each cell keeps its best figure,
so one noisy run does not fail the check.
"""
import argparse
import os
import statistics
import subprocess
import sys
import time
from pathlib import Path
ROOT = Path(__file__).resolve().parent.parent.parent
TRIALS = int(os.environ.get("TRIALS", "5"))
MIN_S = float(os.environ.get("MIN_S", "0.4"))
TOLERANCE = float(os.environ.get("TOLERANCE", "0.10"))
RETRIES = int(os.environ.get("RETRIES", "1"))
# the rows this script times itself: name, source, kernel letter (bench.sh has the rest)
SWEEPS = [
("zy:sweep.checked", "tests/bench/sweep.zc", "c"),
("zy:sweep.unbounded", "tests/bench/sweep.zc", "u"),
]
COLUMNS = ["program", "plain_ms", "o3_ms", "checks", "peak_heap_words", "gc_mark_ms"]
KERNELS = ("collatz", "sieve", "digits", "fib")
def bench_sh():
"""runs tests/bench/bench.sh and returns its rows: (backend:kernel, plain ms, o3 ms)"""
r = subprocess.run(["bash", str(ROOT / "tests/bench/bench.sh")], cwd=ROOT, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
if r.returncode != 0:
sys.exit("bench: tests/bench/bench.sh failed:\n" + r.stdout[-2000:] + r.stderr[-2000:])
rows = []
for line in r.stdout.splitlines():
w = line.split()
# backend names have spaces ("x86-64 jit"): the kernel is the fourth field from the end
if len(w) >= 5 and w[-4] in KERNELS and w[-1].endswith("x"):
backend = "-".join(w[:-4])
rows.append((backend + ":" + w[-4], float(w[-3]), float(w[-2])))
if not rows:
sys.exit("bench: no rows in the output of tests/bench/bench.sh:\n" + r.stdout[-2000:])
return rows
def run_once(zy, src, letter, reps):
data = src + b"`" + letter.encode() + str(reps).encode()
t = time.perf_counter()
r = subprocess.run([zy], input=data, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
dt = time.perf_counter() - t
if r.returncode != 0:
sys.exit("bench: %s %s failed: %s" % (zy, letter, r.stderr.decode(errors="replace")[:300]))
return dt
def slope_ms(zy, path, letter):
src = (ROOT / path).read_bytes()
n = 4
while True:
one = min(run_once(zy, src, letter, 1) for _ in range(2))
many = run_once(zy, src, letter, n)
if many - one >= MIN_S or n >= 1 << 16:
break
n *= 4
slopes = []
for _ in range(TRIALS):
one = run_once(zy, src, letter, 1)
many = run_once(zy, src, letter, n)
slopes.append((many - one) / (n - 1) * 1000)
return statistics.median(slopes)
def fmt_row(cells):
return "%-26s %10s %10s %8s %16s %11s" % tuple(cells)
def num(x):
return "-" if x is None else "%.3f" % x
def write_table(path, rows):
lines = ["# tinyvm bench: ms per repetition, from tests/bench/bench.sh and tests/bench/suite.py; '-' where a figure is not available",
fmt_row(COLUMNS)]
for r in rows:
lines.append(fmt_row([r["program"], num(r["plain"]), num(r["o3"]), r["checks"], r["peak"], r["gc"]]))
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text("\n".join(lines) + "\n")
def read_table(path):
rows = {}
f = lambda x: None if x == "-" else float(x)
for line in path.read_text().splitlines():
if not line.strip() or line.startswith("#") or line.split()[0] == "program":
continue
w = line.split()
rows[w[0]] = {"program": w[0], "plain": f(w[1]), "o3": f(w[2]), "checks": w[3], "peak": w[4], "gc": w[5]}
return rows
def run_all(zy):
rows = []
for name, plain, o3 in bench_sh():
print("bench: %-26s %10.3f %10.3f ms" % (name, plain, o3), file=sys.stderr, flush=True)
rows.append({"program": name, "plain": plain, "o3": o3, "checks": "-", "peak": "-", "gc": "-"})
for name, path, letter in SWEEPS:
ms = slope_ms(zy, path, letter)
print("bench: %-26s %10.3f ms" % (name, ms), file=sys.stderr, flush=True)
rows.append({"program": name, "plain": ms, "o3": None, "checks": "-", "peak": "-", "gc": "-"})
return rows
def slower(now, base):
return now is not None and base is not None and base > 0 and now > base * (1 + TOLERANCE)
def pct(now, base):
return "-" if now is None or base is None or base <= 0 else "%+.1f%%" % ((now - base) / base * 100)
def main():
ap = argparse.ArgumentParser(description="tinyvm benchmark suite")
ap.add_argument("mode", choices=["run", "check"])
ap.add_argument("--zy", default=str(ROOT / "build/bin/zy"))
ap.add_argument("--out", default=str(ROOT / "build/bench.txt"))
ap.add_argument("--baseline", default=str(ROOT / "tests/bench.baseline"))
a = ap.parse_args()
rows = run_all(a.zy)
if a.mode == "run":
write_table(Path(a.out), rows)
print(Path(a.out).read_text(), end="")
return 0
base_path = Path(a.baseline)
if not base_path.exists():
sys.exit("bench: no baseline at %s (run make bench-save)" % base_path)
base = read_table(base_path)
tries = 0
while tries < RETRIES and any(slower(r[c], base[r["program"]][c]) for r in rows if r["program"] in base for c in ("plain", "o3")):
tries += 1
print("bench: something looks slower; measuring again (%d of %d)" % (tries, RETRIES), file=sys.stderr, flush=True)
again = {r["program"]: r for r in run_all(a.zy)}
for r in rows:
g = again.get(r["program"])
if g:
for c in ("plain", "o3"):
if r[c] is not None and g[c] is not None:
r[c] = min(r[c], g[c])
bad = 0
print("%-26s %9s %9s %8s %9s %9s %8s %6s %6s %s" % ("program", "base pl", "now pl", "change", "base o3", "now o3", "change", "b chk", "n chk", "verdict"))
names = set()
for r in rows:
names.add(r["program"])
b = base.get(r["program"])
if not b:
print("%-26s %9s %9s %8s %9s %9s %8s %6s %6s new" % (r["program"], "-", num(r["plain"]), "-", "-", num(r["o3"]), "-", "-", r["checks"]))
continue
why = [c for c in ("plain", "o3") if slower(r[c], b[c])]
verdict = ", ".join(c + " SLOWER" for c in why)
if b["checks"] != "-" and r["checks"] != "-" and int(r["checks"]) > int(b["checks"]):
verdict = (verdict + ", " if verdict else "") + "MORE CHECKS"
if verdict:
bad += 1
print("%-26s %9s %9s %8s %9s %9s %8s %6s %6s %s" % (r["program"], num(b["plain"]), num(r["plain"]), pct(r["plain"], b["plain"]),
num(b["o3"]), num(r["o3"]), pct(r["o3"], b["o3"]), b["checks"], r["checks"], verdict or "ok"))
for name in base:
if name not in names:
print("%-26s %9s %9s gone" % (name, num(base[name]["plain"]), "-"))
write_table(Path(a.out), rows)
print("bench-check: %s (%d of %d rows over the limit; tolerance %d%% per column)" % ("FAIL" if bad else "ok", bad, len(rows), round(TOLERANCE * 100)))
return 1 if bad else 0
if __name__ == "__main__":
sys.exit(main())
|
