def test_lin_comb_diff(ctx_factory, arg_type):
ctx = ctx_factory()
dev, = ctx.devices
if not has_double_support(dev):
if arg_type in (np.float64, np.complex128):
pytest.skip('Device does not support double.')
n = 100000
a_np = (np.random.randn(n)).astype(arg_type)
b_np = (np.random.randn(n)).astype(arg_type)
c_np = (np.random.randn(n) * 10).astype(arg_type)
queue = cl.CommandQueue(ctx)
a_g = cl.array.to_device(queue, a_np)
b_g = cl.array.to_device(queue, b_np)
c_g = cl.array.to_device(queue, c_np)
res_g = cl.array.empty_like(a_g)
lin_comb_diff = lin_comb_diff_kernel(ctx, arg_type, arg_type,
arg_type, np.float32, 2)
gs, ls = get_group_sizes(n, dev, lin_comb_diff)
evt = run_elwise_kernel(lin_comb_diff, queue, gs, ls, n, [],
res_g, c_g, a_g, b_g, 2, 3)
evt.wait()
# Check on GPU with PyOpenCL Array:
assert np.linalg.norm((res_g - (c_g + 2 * a_g + 3 * b_g)).get()) <= 2e-4
# Check on CPU with Numpy:
res_np = res_g.get()
assert np.linalg.norm(res_np - (c_np + 2 * a_np + 3 * b_np)) <= 2e-4
def main2():
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
dev = queue.device
knl = _get_wave_kernel(ctx)
gs, ls = get_group_sizes(len_x * 2, dev, knl)
def f(t, y_in, y_out, wait_for=None):
return run_elwise_kernel(knl, queue, gs, ls, len_x * 2, wait_for,
y_out, y_in, h_x, len_x)
xs = np.arange(len_x) * np.pi / (len_x - 1)
y0 = np.r_[(np.sin(xs) + np.sin(xs * 2) + np.sin(xs * 3)
+ np.sin(xs * 4) + np.sin(xs * 5)) / 5,
np.zeros(len_x)].astype(np.float32)
# y0 += np.r_[np.zeros(len_x),
# [(min((i / len_x) - 0.4, 0.5 - (i / len_x)) * 20
# if 0.4 < (i / len_x) < 0.5 else 0)
# for i in range(len_x)]].astype(np.float32)
y0 += np.r_[np.zeros(len_x),
[((i / len_x) - 0.2 if 0.15 < (i / len_x) < 0.25 else 0) * 20
for i in range(len_x)]].astype(np.float32)
# y0 = np.r_[[(1 if 0.4 < (i / len_x) < 0.5 else 0)
# for i in range(len_x)],
# np.zeros(len_x)].astype(np.float32)
y0 += np.r_[[(1 if 0.75 < (i / len_x) < 0.85 else 0)
for i in range(len_x)],
np.zeros(len_x)].astype(np.float32)
res, evt = solve_ode(t0, t1, h, y0, f, queue)
print('queued')
evt.wait()
print('finished')
res_np = [a.get() for a in res]
def test_bloch(ctx_factory):
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
dev = queue.device
# if dev.type != cl.device_type.GPU:
# pytest.skip('Only use GPU')
knl = _get_bloch_kernel(ctx)
t0 = 0
t1 = 550
h = 0.02
len_x = 512
t_x = 1
slope = 0.05
# slope = 0
gs, ls = get_group_sizes(len_x, dev, knl)
def f(t, y_in, y_out, wait_for=None):
return run_elwise_kernel(knl, queue, gs, ls, len_x, wait_for,
y_out, y_in, t_x, slope, len_x)
y0 = np.zeros(len_x).astype(np.complex64)
# y0[int(len_x / 2)] = 1
# y0[int(len_x * 2 / 5)] = np.sqrt(2) / 2
# y0[int(len_x * 3 / 5)] = np.sqrt(2) / 2
y0[int(len_x * 2 / 5)] = 1 / np.sqrt(3)
y0[int(len_x / 2)] = 1 / np.sqrt(3)
y0[int(len_x * 3 / 5)] = 1 / np.sqrt(3)
print('start')
start_evt = cl.UserEvent(ctx)
res, evt = solve_ode(t0, t1, h, y0, f, queue, wait_for=[start_evt])
start_evt.set_status(cl.command_execution_status.COMPLETE)
print('wait')
evt.wait()
print('done')
dn = 19
res_np = [np.abs(a.get()) for a in res[::dn]]