def test_meta_schedule_measure_callback_fail():
@derived_object
class FailingMeasureCallback(PyMeasureCallback):
def apply(
self,
task_scheduler: TaskScheduler,
task_id: int,
measure_candidates: List[MeasureCandidate],
builds: List[BuilderResult],
results: List[RunnerResult],
) -> None:
raise ValueError("test")
measure_callback = FailingMeasureCallback()
with pytest.raises(ValueError, match="test"):
measure_callback.apply(
RoundRobin([], [],
DummyBuilder(),
DummyRunner(),
DummyDatabase(),
max_trials=1),
0,
[MeasureCandidate(Schedule(Matmul), None)],
[BuilderResult("test_build", None)],
[RunnerResult([1.0, 2.1], None)],
)
def test_meta_schedule_replay_func(
TestClass: SearchStrategy): # pylint: disable = invalid-name
num_trials_per_iter = 7
max_trials_per_task = 20
strategy = TestClass(num_trials_per_iter=num_trials_per_iter,
max_trials_per_task=max_trials_per_task)
context = TuneContext(mod=Matmul,
space_generator=ScheduleFn(sch_fn=_schedule_matmul))
context.space_generator.initialize_with_tune_context(context)
spaces = context.space_generator.generate_design_space(context.mod)
strategy.initialize_with_tune_context(context)
strategy.pre_tuning(spaces)
(correct_sch, ) = ScheduleFn(
sch_fn=_schedule_matmul).generate_design_space(Matmul)
num_trials_each_iter: List[int] = []
candidates = strategy.generate_measure_candidates()
while candidates is not None:
num_trials_each_iter.append(len(candidates))
runner_results: List[RunnerResult] = []
for candidate in candidates:
_is_trace_equal(
candidate.sch,
correct_sch,
remove_decisions=(isinstance(strategy, ReplayTrace)),
)
runner_results.append(
RunnerResult(run_secs=[0.11, 0.41, 0.54], error_msg=None))
strategy.notify_runner_results(context, candidates, runner_results)
candidates = strategy.generate_measure_candidates()
strategy.post_tuning()
assert num_trials_each_iter == [7, 7, 6]
def test_meta_schedule_measure_callback():
@derived_object
class FancyMeasureCallback(PyMeasureCallback):
def apply(
self,
task_scheduler: TaskScheduler,
task_id: int,
measure_candidates: List[MeasureCandidate],
builds: List[BuilderResult],
results: List[RunnerResult],
) -> None:
assert len(measure_candidates) == 1
assert_structural_equal(measure_candidates[0].sch.mod, Matmul)
assert (len(builds) == 1 and builds[0].error_msg is None
and builds[0].artifact_path == "test_build")
assert (len(results) == 1 and results[0].error_msg is None
and len(results[0].run_secs) == 2)
measure_callback = FancyMeasureCallback()
measure_callback.apply(
RoundRobin([], [],
DummyBuilder(),
DummyRunner(),
DummyDatabase(),
max_trials=1),
0,
[MeasureCandidate(Schedule(Matmul), None)],
[BuilderResult("test_build", None)],
[RunnerResult([1.0, 2.1], None)],
)
def test_meta_schedule_replay_trace():
num_trials_per_iter = 7
num_trials_total = 20
(example_sch, ) = ScheduleFn(
sch_fn=_schedule_matmul).generate_design_space(Matmul)
replay = ReplayTrace(num_trials_per_iter=num_trials_per_iter,
num_trials_total=num_trials_total)
tune_context = TuneContext(mod=Matmul)
replay.initialize_with_tune_context(tune_context)
num_trials_each_round: List[int] = []
replay.pre_tuning([example_sch])
while True:
candidates = replay.generate_measure_candidates()
if candidates is None:
break
num_trials_each_round.append(len(candidates))
runner_results: List[RunnerResult] = []
for candidate in candidates:
assert _is_trace_equal(candidate.sch, example_sch)
runner_results.append(
RunnerResult(run_secs=[0.5, 0.4, 0.3], error_msg=None))
replay.notify_runner_results(runner_results)
replay.post_tuning()
assert num_trials_each_round == [7, 7, 6]
def test_meta_schedule_measure_callback_fail():
class FailingMeasureCallback(PyMeasureCallback):
def apply(
self,
task_scheduler: TaskScheduler,
task_id: int,
measure_candidates: List[MeasureCandidate],
builds: List[BuilderResult],
results: List[RunnerResult],
) -> None:
raise ValueError("test")
measure_callback = FailingMeasureCallback()
with pytest.raises(ValueError, match="test"):
measure_callback.apply(
TaskScheduler(),
0,
[MeasureCandidate(Schedule(Matmul), None)],
[BuilderResult("test_build", None)],
[RunnerResult([1.0, 2.1], None)],
)
def result(self) -> RunnerResult:
return RunnerResult(
[random.uniform(5, 30) for _ in range(random.randint(1, 10))],
None)
def test_meta_schedule_evolutionary_search_early_stop(
): # pylint: disable = invalid-name]
def _schedule_matmul_empty(sch: Schedule):
return sch
num_trials_per_iter = 10
max_trials_per_task = 100
strategy = EvolutionarySearch(
num_trials_per_iter=num_trials_per_iter,
max_trials_per_task=max_trials_per_task,
population_size=5,
init_measured_ratio=0.1,
init_min_unmeasured=50,
genetic_num_iters=3,
genetic_mutate_prob=0.5,
genetic_max_fail_count=10,
eps_greedy=0.9,
)
context = TuneContext(
mod=Matmul,
space_generator=ScheduleFn(sch_fn=_schedule_matmul_empty),
mutator_probs={
DummyMutator(): 1.0,
},
target=tvm.target.Target("llvm"),
num_threads=1, # because we are using a mutator from the python side
)
_scheduler = RoundRobin(
tasks=[context],
task_weights=[1.0],
builder=ms.builder.LocalBuilder(),
runner=ms.runner.LocalRunner(),
database=DummyDatabase(),
cost_model=ms.cost_model.RandomModel(),
measure_callbacks=[],
max_trials=1,
)
context.space_generator.initialize_with_tune_context(context)
spaces = context.space_generator.generate_design_space(context.mod)
strategy.initialize_with_tune_context(context)
strategy.pre_tuning(spaces)
(correct_sch, ) = ScheduleFn(
sch_fn=_schedule_matmul).generate_design_space(Matmul)
num_trials_each_iter: List[int] = []
candidates = strategy.generate_measure_candidates()
while candidates is not None:
num_trials_each_iter.append(len(candidates))
runner_results: List[RunnerResult] = []
for candidate in candidates:
_is_trace_equal(
candidate.sch,
correct_sch,
remove_decisions=(isinstance(strategy, ReplayTrace)),
)
runner_results.append(
RunnerResult(run_secs=[0.11, 0.41, 0.54], error_msg=None))
strategy.notify_runner_results(context, candidates, runner_results)
candidates = strategy.generate_measure_candidates()
strategy.post_tuning()
assert num_trials_each_iter == [1, 0, 0, 0, 0]
del _scheduler
def _dummy_result(num_samples: int = 4, max_run_sec: int = 10):
return RunnerResult(list(np.random.rand(num_samples) * max_run_sec + 1e-6), None)
def test_meta_schedule_evolutionary_search(
): # pylint: disable = invalid-name]
@derived_object
class DummyMutator(PyMutator):
"""Dummy Mutator for testing"""
def initialize_with_tune_context(self, context: "TuneContext") -> None:
pass
def apply(self, trace: Trace, _) -> Optional[Trace]:
return Trace(trace.insts, {})
@derived_object
class DummyDatabase(PyDatabase):
"""Dummy Database for testing"""
def __init__(self):
super().__init__()
self.records = []
self.workload_reg = []
def has_workload(self, mod: IRModule) -> bool:
for workload in self.workload_reg:
if tvm.ir.structural_equal(workload.mod, mod):
return True
return False
def commit_tuning_record(self, record: TuningRecord) -> None:
self.records.append(record)
def commit_workload(self, mod: IRModule) -> Workload:
for workload in self.workload_reg:
if tvm.ir.structural_equal(workload.mod, mod):
return workload
workload = Workload(mod)
self.workload_reg.append(workload)
return workload
def get_top_k(self, workload: Workload,
top_k: int) -> List[TuningRecord]:
return list(
filter(
lambda x: x.workload == workload,
sorted(self.records,
key=lambda x: sum(x.run_secs) / len(x.run_secs)),
))[:int(top_k)]
def __len__(self) -> int:
return len(self.records)
def print_results(self) -> None:
print("\n".join([str(r) for r in self.records]))
num_trials_per_iter = 10
num_trials_total = 100
strategy = EvolutionarySearch(
num_trials_per_iter=num_trials_per_iter,
num_trials_total=num_trials_total,
population_size=5,
init_measured_ratio=0.1,
init_min_unmeasured=50,
genetic_num_iters=3,
genetic_mutate_prob=0.5,
genetic_max_fail_count=10,
eps_greedy=0.9,
)
context = TuneContext(
mod=Matmul,
space_generator=ScheduleFn(sch_fn=_schedule_matmul),
mutator_probs={
DummyMutator(): 1.0,
},
target=tvm.target.Target("llvm"),
num_threads=1, # because we are using a mutator from the python side
)
_scheduler = RoundRobin(
tasks=[context],
builder=LocalBuilder(),
runner=LocalRunner(),
database=DummyDatabase(),
cost_model=RandomModel(),
measure_callbacks=[],
)
context.space_generator.initialize_with_tune_context(context)
spaces = context.space_generator.generate_design_space(context.mod)
strategy.initialize_with_tune_context(context)
strategy.pre_tuning(spaces)
(correct_sch, ) = ScheduleFn(
sch_fn=_schedule_matmul).generate_design_space(Matmul)
num_trials_each_iter: List[int] = []
candidates = strategy.generate_measure_candidates()
while candidates is not None:
num_trials_each_iter.append(len(candidates))
runner_results: List[RunnerResult] = []
for candidate in candidates:
_is_trace_equal(
candidate.sch,
correct_sch,
remove_decisions=(isinstance(strategy, ReplayTrace)),
)
runner_results.append(
RunnerResult(run_secs=[0.11, 0.41, 0.54], error_msg=None))
strategy.notify_runner_results(context, candidates, runner_results)
candidates = strategy.generate_measure_candidates()
strategy.post_tuning()
print(num_trials_each_iter)
correct_count = 10 # For each iteration except the last one
assert num_trials_each_iter == [correct_count] * (
num_trials_total // correct_count) + (
[num_trials_total %
correct_count] if num_trials_total % correct_count != 0 else [])
del _scheduler
def result(self) -> RunnerResult:
run_secs = [
random.uniform(5, 30) for _ in range(random.randint(1, 10))
]
return RunnerResult(run_secs, None)