def test_state_from_csv_invalid_format():
buf = StringIO("abcdef\n")
reader = CompilerEnvStateReader(buf)
with pytest.raises(
ValueError,
match=r"Expected 4 columns in the first row of CSV: \['abcdef'\]"):
next(iter(reader))
def test_compiler_env_state_reader_no_header():
buf = StringIO("benchmark://cbench-v0/foo,2.0,5.0,-a -b -c\n")
reader = CompilerEnvStateReader(buf)
assert list(reader) == [
CompilerEnvState(
benchmark="benchmark://cbench-v0/foo",
walltime=5,
commandline="-a -b -c",
reward=2,
)
]
def test_compiler_env_state_reader_with_header_out_of_order_columns():
buf = StringIO("commandline,reward,benchmark,walltime\n"
"-a -b -c,2.0,benchmark://cbench-v0/foo,5.0\n")
reader = CompilerEnvStateReader(buf)
assert list(reader) == [
CompilerEnvState(
benchmark="benchmark://cbench-v0/foo",
walltime=5,
commandline="-a -b -c",
reward=2,
)
]
def test_state_serialize_deserialize_equality_no_reward():
original_state = CompilerEnvState(benchmark="benchmark://cbench-v0/foo",
walltime=100,
commandline="-a -b -c")
buf = StringIO()
CompilerEnvStateWriter(buf).write_state(original_state)
buf.seek(0) # Rewind the buffer for reading.
state_from_csv = next(iter(CompilerEnvStateReader(buf)))
assert state_from_csv.benchmark == "benchmark://cbench-v0/foo"
assert state_from_csv.walltime == 100
assert state_from_csv.reward is None
assert state_from_csv.commandline == "-a -b -c"
def test_state_serialize_deserialize_equality(env: LlvmEnv):
env.reset(benchmark="cbench-v1/crc32")
env.episode_reward = 10
state = env.state
assert state.reward == 10
buf = StringIO()
CompilerEnvStateWriter(buf).write_state(state)
buf.seek(0) # Rewind the buffer for reading.
state_from_csv = next(iter(CompilerEnvStateReader(buf)))
assert state_from_csv.reward == 10
assert state == state_from_csv
def test_compiler_env_state_reader_header_only():
buf = StringIO("benchmark,reward,walltime,commandline\n")
reader = CompilerEnvStateReader(buf)
assert list(reader) == []
def test_compiler_env_state_reader_empty_input():
buf = StringIO("")
reader = CompilerEnvStateReader(buf)
assert list(reader) == []
def main(argv):
assert len(argv) == 1, f"Unknown args: {argv[:1]}"
assert FLAGS.n > 0, "n must be > 0"
with gym.make("llvm-ic-v0") as env:
# Stream verbose CompilerGym logs to file.
logger = logging.getLogger("compiler_gym")
logger.setLevel(logging.DEBUG)
log_handler = logging.FileHandler(FLAGS.leaderboard_logfile)
logger.addHandler(log_handler)
logger.propagate = False
print(f"Writing results to {FLAGS.leaderboard_results}")
print(f"Writing logs to {FLAGS.leaderboard_logfile}")
# Build the list of benchmarks to evaluate.
benchmarks = env.datasets[FLAGS.test_dataset].benchmark_uris()
if FLAGS.max_benchmarks:
benchmarks = islice(benchmarks, FLAGS.max_benchmarks)
benchmarks = list(benchmarks)
# Repeat the searches for the requested number of iterations.
benchmarks *= FLAGS.n
total_count = len(benchmarks)
# If we are resuming from a previous job, read the states that have
# already been proccessed and remove those benchmarks from the list
# of benchmarks to evaluate.
init_states = []
if FLAGS.resume and Path(FLAGS.leaderboard_results).is_file():
with CompilerEnvStateReader(open(
FLAGS.leaderboard_results)) as reader:
for state in reader:
init_states.append(state)
if state.benchmark in benchmarks:
benchmarks.remove(state.benchmark)
# Run the benchmark loop in background so that we can asynchronously
# log progress.
worker = _EvalPolicyWorker(env, benchmarks, policy, init_states)
worker.start()
timer = Timer().reset()
try:
print(f"=== Evaluating policy on "
f"{humanize.intcomma(total_count)} "
f"{FLAGS.test_dataset} benchmarks ==="
"\n\n" # Blank lines will be filled below
)
while worker.is_alive():
done_count = len(worker.states)
remaining_count = total_count - done_count
time = timer.time
gmean_reward = geometric_mean(
[s.reward for s in worker.states])
mean_walltime = (arithmetic_mean(
[s.walltime for s in worker.states]) or time)
print(
"\r\033[2A"
"\033[K"
f"Runtime: {humanize_duration_hms(time)}. "
f"Estimated completion: {humanize_duration_hms(mean_walltime * remaining_count)}. "
f"Completed: {humanize.intcomma(done_count)} / {humanize.intcomma(total_count)} "
f"({done_count / total_count:.1%})."
"\n\033[K"
f"Current mean walltime: {mean_walltime:.3f}s / benchmark."
"\n\033[K"
f"Current geomean reward: {gmean_reward:.4f}.",
flush=True,
end="",
)
sleep(1)
except KeyboardInterrupt:
print("\nkeyboard interrupt", flush=True)
worker.alive = False
# User interrupt, don't validate.
FLAGS.validate = False
if FLAGS.validate:
FLAGS.env = "llvm-ic-v0"
validate(["argv0", FLAGS.leaderboard_results])
def main(argv):
"""Main entry point."""
try:
states = list(CompilerEnvStateReader.read_paths(argv[1:]))
except ValueError as e:
print(e, file=sys.stderr)
sys.exit(1)
if not states:
print(
"No inputs to validate. Pass a CSV file path as an argument, or "
"use - to read from stdin.",
file=sys.stderr,
)
sys.exit(1)
# Send the states off for validation
if FLAGS.debug_force_valid:
validation_results = (
ValidationResult(
state=state,
reward_validated=True,
actions_replay_failed=False,
reward_validation_failed=False,
benchmark_semantics_validated=False,
benchmark_semantics_validation_failed=False,
walltime=0,
)
for state in states
)
else:
validation_results = validate_states(
env_from_flags,
states,
nproc=FLAGS.nproc,
inorder=FLAGS.inorder,
)
# Determine the name of the reward space.
with env_from_flags() as env:
if FLAGS.reward_aggregation == "geomean":
def reward_aggregation(a):
return geometric_mean(np.clip(a, 0, None))
reward_aggregation_name = "Geometric mean"
elif FLAGS.reward_aggregation == "mean":
reward_aggregation = arithmetic_mean
reward_aggregation_name = "Mean"
else:
raise app.UsageError(
f"Unknown aggregation type: '{FLAGS.reward_aggregation}'"
)
if env.reward_space:
reward_name = f"{reward_aggregation_name} {env.reward_space.id}"
else:
reward_name = ""
# Determine the maximum column width required for printing tabular output.
max_state_name_length = max(
len(s)
for s in [state_name(s) for s in states]
+ [
"Mean inference walltime",
reward_name,
]
)
name_col_width = min(max_state_name_length + 2, 78)
error_count = 0
rewards = []
walltimes = []
if FLAGS.summary_only:
def intermediate_print(*args, **kwargs):
del args
del kwargs
else:
intermediate_print = print
def progress_message(i):
intermediate_print(
f"{i} remaining {plural(i, 'state', 'states')} to validate ... ",
end="",
flush=True,
)
progress_message(len(states))
result_dicts = []
def dump_result_dicst_to_json():
with open(FLAGS.validation_logfile, "w") as f:
json.dump(result_dicts, f)
for i, result in enumerate(validation_results, start=1):
intermediate_print("\r\033[K", to_string(result, name_col_width), sep="")
progress_message(len(states) - i)
result_dicts.append(result.dict())
if not result.okay():
error_count += 1
elif result.reward_validated and not result.reward_validation_failed:
rewards.append(result.state.reward)
walltimes.append(result.state.walltime)
if not i % 10:
dump_result_dicst_to_json()
dump_result_dicst_to_json()
# Print a summary footer.
intermediate_print("\r\033[K----", "-" * name_col_width, "-----------", sep="")
print(f"Number of validated results: {emph(len(walltimes))} of {len(states)}")
walltime_mean = f"{arithmetic_mean(walltimes):.3f}s"
walltime_std = f"{stdev(walltimes):.3f}s"
print(
f"Mean walltime per benchmark: {emph(walltime_mean)} "
f"(std: {emph(walltime_std)})"
)
reward = f"{reward_aggregation(rewards):.3f}"
reward_std = f"{stdev(rewards):.3f}"
print(f"{reward_name}: {emph(reward)} " f"(std: {emph(reward_std)})")
if error_count:
sys.exit(1)
def test_state_from_csv_invalid_benchmark_uri():
buf = StringIO("benchmark,reward,walltime,commandline\n"
"invalid-uri,2.0,5.0,-a -b -c\n")
reader = CompilerEnvStateReader(buf)
with pytest.raises(ValueError, match="string does not match regex"):
next(iter(reader))
