def test_meta_schedule_task_scheduler_override_next_task_id_only(): # pylint: disable=invalid-name
num_trials_per_iter = 6
max_trials_per_task = 101
tasks = [
TuneContext(
MatmulModule,
target=tvm.target.Target("llvm"),
space_generator=ScheduleFn(sch_fn=_schedule_matmul),
search_strategy=ReplayTrace(num_trials_per_iter,
max_trials_per_task),
task_name="Matmul",
rand_state=42,
),
TuneContext(
MatmulReluModule,
target=tvm.target.Target("llvm"),
space_generator=ScheduleFn(sch_fn=_schedule_matmul),
search_strategy=ReplayTrace(num_trials_per_iter,
max_trials_per_task),
task_name="MatmulRelu",
rand_state=0xDEADBEEF,
),
TuneContext(
BatchMatmulModule,
target=tvm.target.Target("llvm"),
space_generator=ScheduleFn(sch_fn=_schedule_batch_matmul),
search_strategy=ReplayTrace(num_trials_per_iter,
max_trials_per_task),
task_name="BatchMatmul",
rand_state=0x114514,
),
]
database = DummyDatabase()
scheduler = MyTaskScheduler(
tasks,
DummyBuilder(),
DummyRunner(),
database,
measure_callbacks=[
measure_callback.AddToDatabase(),
],
max_trials=max_trials_per_task * len(tasks),
)
scheduler.tune()
assert len(database) == max_trials_per_task * len(tasks)
for task in tasks:
assert (len(
database.get_top_k(
database.commit_workload(task.mod),
100000,
)) == max_trials_per_task)
def test_meta_schedule_task_scheduler_multiple():
num_trials_per_iter = 6
max_trials_per_task = 101
tasks = [
TuneContext(
MatmulModule,
target=tvm.target.Target("llvm"),
space_generator=ScheduleFn(sch_fn=_schedule_matmul),
search_strategy=ReplayTrace(num_trials_per_iter,
max_trials_per_task),
task_name="Matmul",
rand_state=42,
),
TuneContext(
MatmulReluModule,
target=tvm.target.Target("llvm"),
space_generator=ScheduleFn(sch_fn=_schedule_matmul),
search_strategy=ReplayTrace(num_trials_per_iter,
max_trials_per_task),
task_name="MatmulRelu",
rand_state=0xDEADBEEF,
),
TuneContext(
BatchMatmulModule,
target=tvm.target.Target("llvm"),
space_generator=ScheduleFn(sch_fn=_schedule_batch_matmul),
search_strategy=ReplayTrace(num_trials_per_iter,
max_trials_per_task),
task_name="BatchMatmul",
rand_state=0x114514,
),
]
database = DummyDatabase()
round_robin = RoundRobin(
tasks,
[1.0],
DummyBuilder(),
DummyRunner(),
database,
measure_callbacks=[measure_callback.AddToDatabase()],
max_trials=max_trials_per_task * len(tasks),
)
round_robin.tune()
assert len(database) == max_trials_per_task * len(tasks)
for task in tasks:
assert (len(
database.get_top_k(
database.commit_workload(task.mod),
100000,
)) == max_trials_per_task)
def test_meta_schedule_task_scheduler_single():
num_trials_per_iter = 3
max_trials_per_task = 10
sch_fn = ScheduleFn(sch_fn=_schedule_matmul)
replay = ReplayTrace(num_trials_per_iter, max_trials_per_task)
task = TuneContext(
MatmulModule,
target=tvm.target.Target("llvm"),
space_generator=sch_fn,
search_strategy=replay,
task_name="Test",
rand_state=42,
)
database = DummyDatabase()
round_robin = RoundRobin(
[task],
[1.0],
DummyBuilder(),
DummyRunner(),
database,
measure_callbacks=[measure_callback.AddToDatabase()],
max_trials=max_trials_per_task,
)
round_robin.tune()
assert len(database) == max_trials_per_task
def test_meta_schedule_task_scheduler_multiple():
num_trials_per_iter = 6
num_trials_total = 101
tasks = [
TuneContext(
MatmulModule,
target=tvm.target.Target("llvm"),
space_generator=ScheduleFn(sch_fn=_schedule_matmul),
search_strategy=ReplayTrace(num_trials_per_iter, num_trials_total),
task_name="Matmul",
rand_state=42,
),
TuneContext(
MatmulReluModule,
target=tvm.target.Target("llvm"),
space_generator=ScheduleFn(sch_fn=_schedule_matmul),
search_strategy=ReplayTrace(num_trials_per_iter, num_trials_total),
task_name="MatmulRelu",
rand_state=0xDEADBEEF,
),
TuneContext(
BatchMatmulModule,
target=tvm.target.Target("llvm"),
space_generator=ScheduleFn(sch_fn=_schedule_batch_matmul),
search_strategy=ReplayTrace(num_trials_per_iter, num_trials_total),
task_name="BatchMatmul",
rand_state=0x114514,
),
]
database = DummyDatabase()
round_robin = RoundRobin(tasks, DummyBuilder(), DummyRunner(), database)
round_robin.tune()
assert len(database) == num_trials_total * len(tasks)
print(database.workload_reg)
for task in tasks:
assert len(database.get_top_k(database.commit_workload(task.mod),
1e9)) == num_trials_total
def test_meta_schedule_task_scheduler_single():
num_trials_per_iter = 3
num_trials_total = 10
sch_fn = ScheduleFn(sch_fn=_schedule_matmul)
replay = ReplayTrace(num_trials_per_iter, num_trials_total)
task = TuneContext(
MatmulModule,
target=tvm.target.Target("llvm"),
space_generator=sch_fn,
search_strategy=replay,
task_name="Test",
rand_state=42,
)
database = DummyDatabase()
round_robin = RoundRobin([task], DummyBuilder(), DummyRunner(), database)
round_robin.tune()
assert len(database) == num_trials_total
def test_meta_schedule_task_scheduler_override_next_task_id_only():
class MyTaskScheduler(PyTaskScheduler):
done = set()
def next_task_id(self) -> int:
while len(self.done) != len(tasks):
x = random.randint(0, len(tasks) - 1)
task = tasks[x]
if not task.is_stopped:
"""Calling base func via following route:
Python side:
PyTaskScheduler does not have `_is_task_running`
Call TaskScheduler's `is_task_running`, which calls ffi
C++ side:
The ffi calls TaskScheduler's `is_task_running`
But it is overridden in PyTaskScheduler
PyTaskScheduler checks if the function is overridden in python
If not, it returns the TaskScheduler's vtable, calling
TaskScheduler::IsTaskRunning
"""
if self._is_task_running(x):
# Same Here
self._join_running_task(x)
return x
else:
self.done.add(x)
return -1
num_trials_per_iter = 6
num_trials_total = 101
tasks = [
TuneContext(
MatmulModule,
target=tvm.target.Target("llvm"),
space_generator=ScheduleFn(sch_fn=_schedule_matmul),
search_strategy=ReplayTrace(num_trials_per_iter, num_trials_total),
task_name="Matmul",
rand_state=42,
),
TuneContext(
MatmulReluModule,
target=tvm.target.Target("llvm"),
space_generator=ScheduleFn(sch_fn=_schedule_matmul),
search_strategy=ReplayTrace(num_trials_per_iter, num_trials_total),
task_name="MatmulRelu",
rand_state=0xDEADBEEF,
),
TuneContext(
BatchMatmulModule,
target=tvm.target.Target("llvm"),
space_generator=ScheduleFn(sch_fn=_schedule_batch_matmul),
search_strategy=ReplayTrace(num_trials_per_iter, num_trials_total),
task_name="BatchMatmul",
rand_state=0x114514,
),
]
database = DummyDatabase()
scheduler = MyTaskScheduler(tasks, DummyBuilder(), DummyRunner(), database)
scheduler.tune()
assert len(database) == num_trials_total * len(tasks)
for task in tasks:
assert len(database.get_top_k(database.commit_workload(task.mod),
1e9)) == num_trials_total
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]
