def timeit(self, number=1000000):
with common.set_torch_threads(self._task_spec.num_threads):
# Warmup
self._timer.timeit(number=max(int(number // 100), 1))
return common.Measurement(
number_per_run=number,
raw_times=[self._timer.timeit(number=number)],
task_spec=self._task_spec)
def timeit(self, number: int = 1000000) -> common.Measurement:
"""Mirrors the semantics of timeit.Timer.timeit().
Execute the main statement (`stmt`) `number` times.
https://docs.python.org/3/library/timeit.html#timeit.Timer.timeit
"""
with common.set_torch_threads(self._task_spec.num_threads):
# Warmup
self._timeit(number=max(int(number // 100), 2))
return common.Measurement(number_per_run=number,
raw_times=[self._timeit(number=number)],
task_spec=self._task_spec)
def _estimate_block_size(self, min_run_time: float) -> int:
with common.set_torch_threads(self._task_spec.num_threads):
# Estimate the block size needed for measurement to be negligible
# compared to the inner loop. This also serves as a warmup.
overhead = np.median([self._timer.timeit(0) for _ in range(5)])
number = 1
while True:
time_taken = self._timer.timeit(number)
relative_overhead = overhead / time_taken
if relative_overhead <= 1e-4 and time_taken >= min_run_time / 1000:
break
if time_taken > min_run_time:
break
number *= 10
return number
def _estimate_block_size(self, min_run_time: float) -> int:
with common.set_torch_threads(self._task_spec.num_threads):
# Estimate the block size needed for measurement to be negligible
# compared to the inner loop. This also serves as a warmup.
overhead = torch.tensor([self._timeit(0)
for _ in range(5)]).median().item()
number = 1
while True:
time_taken = self._timeit(number)
relative_overhead = overhead / time_taken
if relative_overhead <= 1e-4 and time_taken >= min_run_time / 1000:
break
if time_taken > min_run_time:
break
# Avoid overflow in C++ pybind11 interface
if number * 10 > 2147483647:
break
number *= 10
return number
def _threaded_measurement_loop(
self,
number: int,
time_hook: Callable[[], float],
stop_hook: Callable[[List[float]], bool],
min_run_time: float,
max_run_time: Optional[float] = None,
callback: Optional[Callable[[int, float], NoReturn]] = None):
total_time = 0.0
can_stop = False
times: List[float] = []
with common.set_torch_threads(self._task_spec.num_threads):
while (total_time < min_run_time) or (not can_stop):
time_spent = time_hook()
times.append(time_spent)
total_time += time_spent
if callback:
callback(number, time_spent)
can_stop = stop_hook(times)
if max_run_time and total_time > max_run_time:
break
return times
