def final_reward(self, env: LlvmEnv, runtime_count: int = 30) -> float:
"""Compute the final reward of the environment.
Note that this may modify the environment state. You should call
:code:`reset()` before continuing to use the environment after this.
"""
# Reapply the environment state in a retry loop.
actions = list(env.actions)
env.reset()
for i in range(1, 5 + 1):
_, _, done, info = env.multistep(actions)
if not done:
break
logger.warning(
"Attempt %d to apply actions during final reward failed: %s",
i,
info.get("error_details"),
)
else:
raise ValueError("Failed to replay environment's actions")
if self.value == OptimizationTarget.CODESIZE:
return env.observation.IrInstructionCountOz() / max(
env.observation.IrInstructionCount(), 1
)
if self.value == OptimizationTarget.BINSIZE:
return env.observation.ObjectTextSizeOz() / max(
env.observation.ObjectTextSizeBytes(), 1
)
if self.value == OptimizationTarget.RUNTIME:
with _RUNTIME_LOCK:
with compiler_gym.make("llvm-v0", benchmark=env.benchmark) as new_env:
new_env.reset()
new_env.runtime_observation_count = runtime_count
new_env.runtime_warmup_count = 0
new_env.apply(env.state)
final_runtimes = new_env.observation.Runtime()
assert len(final_runtimes) == runtime_count
new_env.reset()
new_env.send_param("llvm.apply_baseline_optimizations", "-O3")
o3_runtimes = new_env.observation.Runtime()
assert len(o3_runtimes) == runtime_count
logger.debug("O3 runtimes: %s", o3_runtimes)
logger.debug("Final runtimes: %s", final_runtimes)
speedup = np.median(o3_runtimes) / max(np.median(final_runtimes), 1e-12)
logger.debug("Speedup: %.4f", speedup)
return speedup
assert False, f"Unknown OptimizationTarget: {self.value}"
def test_fork_regression_test(env: LlvmEnv, test: ForkRegressionTest):
"""Run the fork regression test:
1. Initialize an environment.
2. Apply a "pre_fork" sequence of actions.
3. Create a fork of the environment.
4. Apply a "post_fork" sequence of actions in both the fork and parent.
5. Verify that the environment states have gone out of sync.
"""
env.reward_space = test.reward_space
env.reset(test.benchmark)
pre_fork = [env.action_space[f] for f in test.pre_fork.split()]
post_fork = [env.action_space[f] for f in test.post_fork.split()]
_, _, done, info = env.multistep(pre_fork)
assert not done, info
with env.fork() as fkd:
assert env.state == fkd.state # Sanity check
env.multistep(post_fork)
fkd.multistep(post_fork)
# Verify that the environment states no longer line up.
assert env.state != fkd.state
