def test_eager_reward_space(env: CompilerEnv):
env.eager_reward_space = "IrInstructionCount"
assert env.eager_reward_space == "IrInstructionCount"
env.eager_reward_space = None
assert env.eager_reward_space is None
invalid = "invalid value"
with pytest.raises(LookupError) as ctx:
env.eager_reward_space = invalid
assert str(ctx.value) == f"Reward space not found: {invalid}"
def test_eager_reward(env: CompilerEnv):
env.eager_reward_space = "codesize"
env.reset()
observation, reward, done, info = env.step(0)
assert observation is None
assert reward == 0
assert not done
def test_step(env: CompilerEnv, observation_space: str, reward_space: str):
"""Request every combination of observation and reward in a fresh environment."""
env.eager_reward_space = None
env.eager_observation_space = None
env.reset(benchmark="cBench-v0/crc32")
observation = env.observation[observation_space]
assert observation is not None
reward = env.reward[reward_space]
assert reward is not None
def test_step(env: CompilerEnv, action_name: str):
"""Run each action on a single benchmark."""
env.eager_reward_space = "IrInstructionCount"
env.eager_observation_space = "Autophase"
env.reset(benchmark="cBench-v0/crc32")
observation, reward, done, info = env.step(
env.action_space.from_string(action_name)
)
if done:
assert observation is None
assert reward is None
else:
assert isinstance(observation, np.ndarray)
assert observation.shape == (56,)
assert reward < 0
def test_service_env_dies_reset(env: CompilerEnv):
env.eager_observation_space = "Autophase"
env.eager_reward_space = "IrInstructionCount"
env.reset("cBench-v0/crc32")
# Kill the service.
env.service.close()
# Check that the environment doesn't fall over.
observation, reward, done, _ = env.step(0)
assert done
assert observation is None
assert reward is None
# Reset the environment and check that it works.
env.reset(benchmark="cBench-v0/crc32")
observation, reward, done, _ = env.step(0)
assert not done
assert observation is not None
assert reward is not None
def replay_actions_from_logs(env: CompilerEnv,
logdir: Path,
benchmark=None) -> None:
best_actions_path = logdir / logs.BEST_ACTIONS_NAME
meta_path = logdir / logs.METADATA_NAME
assert best_actions_path.is_file(), f"File not found: {best_actions_path}"
assert meta_path.is_file(), f"File not found: {meta_path}"
with open(meta_path, "rb") as f:
meta = json.load(f)
with open(best_actions_path) as f:
actions = [l.strip() for l in f.readlines() if l.strip()]
benchmark = benchmark or meta["benchmark"]
env.eager_reward_space = meta["reward"]
env.reset(benchmark=benchmark)
replay_actions(env, actions, logdir)
env.close()
def test_invalid_eager_reward_space(env: CompilerEnv, ):
with pytest.raises(LookupError):
env.eager_reward_space = 100
