def _preprocess(self):
if self.xp == numpy:
start = time.time()
self._running_stack.append(start)
else:
start = cuda.Event()
stop = cuda.Event()
start.record()
self._running_stack.append((start, stop))
self._depth += 1
def _preprocess(self):
if self.xp is numpy:
start = _get_time()
self._running_stack.append(start)
else:
assert self.xp is cuda.cupy
start = cuda.Event()
stop = cuda.Event()
start.record()
self._running_stack.append((start, stop))
self._depth += 1
def __init__(self, stream=None, compute_stream=None):
self._stream = stream
self.compute_stream = compute_stream
self._device = None
self._conveyor = collections.defaultdict(self._get_conveyor)
if compute_stream is not None:
# * event1 prevents a CPU thread to update arrays that might be
# still being used by GPU kernels.
# * event2 prevents a GPU kernel to read arrays that might be
# still being transferred to GPU.
self._event1 = cuda.Event()
self._event2 = cuda.Event()
self._sync_get = False
else:
self._sync_get = True
