def objective(trial):
args = suggest_func(trial)
max_total_time = optimize_config.max_total_time_per_trial
try:
perf = profiler.benchmark(target_func,
args,
max_duration=max_total_time)
return perf.gpu_times.mean()
except Exception as e:
if isinstance(e, ignore_error):
return math.inf
else:
raise e
def test_multigpu_routine(self):
with mock.patch('time.perf_counter',
mock.Mock(side_effect=[2.4, 3.8, 3.8] * 10)):
with mock.patch('cupy.cuda.get_elapsed_time',
mock.Mock(return_value=2500)):
mock_func = mock.Mock()
mock_func.__name__ = 'test_name_xxx'
x = cupy.testing.shaped_random((2, 3), cupy, 'int32')
y = cupy.testing.shaped_random((2, 3), cupy, 'int32')
assert mock_func.call_count == 0
perf = profiler.benchmark(
mock_func, (x, y), n_repeat=10, n_warmup=3, devices=(0, 1))
assert perf.name == 'test_name_xxx'
assert mock_func.call_count == 13
assert perf.cpu_times.shape == (10,)
assert perf.gpu_times.shape == (2, 10,)
assert (perf.cpu_times == 1.4).all()
assert (perf.gpu_times == 2.5).all()
def test_benchmark_max_duration(self):
with mock.patch('time.perf_counter',
mock.Mock(side_effect=[1., 2., 2.] * 6)):
with mock.patch('cupy.cuda.get_elapsed_time',
mock.Mock(return_value=2500)):
mock_func = mock.Mock()
mock_func.__name__ = 'test_name_xxx'
x = cupy.testing.shaped_random((2, 3), cupy, 'int32')
y = cupy.testing.shaped_random((2, 3), cupy, 'int32')
assert mock_func.call_count == 0
perf = profiler.benchmark(
mock_func, (x, y), n_warmup=3, max_duration=2.5)
assert perf.name == 'test_name_xxx'
assert mock_func.call_count == 6
assert perf.cpu_times.shape == (3,)
assert perf.gpu_times.shape == (1, 3)
assert (perf.cpu_times == 1.).all()
assert (perf.gpu_times == 2.5).all()
def test_benchmark_kwargs(self):
x = cupy.random.rand(5)
profiler.benchmark(
cupy.nonzero, kwargs={'a': x}, n_repeat=1, n_warmup=1)
def launch_test(n, ts):
flow = CavityFlow(n, ts)
print(benchmark(flow.compute, n_repeat=5, n_warmup=2))