def test_trigonometric_fn(filter_str, trig_op, input_arrays):
# FIXME: Why does archcosh fail on Gen12 discrete graphics card?
if trig_op == "arccosh" and is_gen12(filter_str):
pytest.skip()
a, b = input_arrays
trig_fn = getattr(np, trig_op)
actual = np.empty(shape=a.shape, dtype=a.dtype)
expected = np.empty(shape=a.shape, dtype=a.dtype)
if trig_op == "arctan2":
@njit
def f(a, b):
return trig_fn(a, b)
device = dpctl.SyclDevice(filter_str)
with dpctl.device_context(device), assert_auto_offloading():
actual = f(a, b)
expected = trig_fn(a, b)
else:
@njit
def f(a):
return trig_fn(a)
device = dpctl.SyclDevice(filter_str)
with dpctl.device_context(device), assert_auto_offloading():
actual = f(a)
expected = trig_fn(a)
np.testing.assert_allclose(actual, expected, rtol=1e-5, atol=0)
def test_multiple_prange(self):
@njit
def f(a, b):
# dimensions must be provided as scalar
m, n = a.shape
for i in prange(m):
val = 10
for j in prange(n):
b[i, j] = a[i, j] * val
for i in prange(m):
for j in prange(n):
a[i, j] = a[i, j] * 10
m = 8
n = 8
a = np.ones((m, n))
b = np.ones((m, n))
device = dpctl.SyclDevice("opencl:gpu")
with assert_auto_offloading(
parfor_offloaded=2), dpctl.device_context(device):
f(a, b)
assert np.all(b == 10)
assert np.all(a == 10)
def test_njit(filter_str):
if _helper.platform_not_supported(filter_str):
pytest.skip()
if _helper.skip_test(filter_str):
pytest.skip()
@vectorize(nopython=True)
def axy(a, x, y):
return a * x + y
def f(a0, a1):
return np.cos(axy(a0, np.sin(a1) - 1.0, 1.0))
A = np.random.random(10)
B = np.random.random(10)
device = dpctl.SyclDevice(filter_str)
with dppy.offload_to_sycl_device(device), assert_auto_offloading():
f_njit = njit(f)
expected = f_njit(A, B)
actual = f(A, B)
max_abs_err = expected.sum() - actual.sum()
assert max_abs_err < 1e-5
def test_dpnp_interacting_with_parfor(self):
def f(a, b):
c = np.sum(a)
e = np.add(b, a)
d = c + e
return d
device = dpctl.SyclDevice("opencl:gpu")
with dpctl.device_context(
device), assert_auto_offloading(), dpnp_debug():
njit_f = njit(f)
got = njit_f(self.a, self.b)
expected = f(self.a, self.b)
max_abs_err = got.sum() - expected.sum()
assert max_abs_err < 1e-4
def test_binary_ops(filter_str, binary_op, input_arrays):
a, b = input_arrays
binop = getattr(np, binary_op)
actual = np.empty(shape=a.shape, dtype=a.dtype)
expected = np.empty(shape=a.shape, dtype=a.dtype)
@njit
def f(a, b):
return binop(a, b)
device = dpctl.SyclDevice(filter_str)
with dpctl.device_context(device), assert_auto_offloading():
actual = f(a, b)
expected = binop(a, b)
np.testing.assert_allclose(actual, expected, rtol=1e-5, atol=0)
def test_one_prange(self):
@njit
def f(a, b):
for i in prange(4):
b[i, 0] = a[i, 0] * 10
m = 8
n = 8
a = np.ones((m, n))
b = np.ones((m, n))
device = dpctl.SyclDevice("opencl:gpu")
with assert_auto_offloading(), dpctl.device_context(device):
f(a, b)
for i in range(4):
assert b[i, 0] == a[i, 0] * 10
def test_one_prange(self):
@njit
def f(a, b):
for i in prange(4):
b[i, 0] = a[i, 0] * 10
m = 8
n = 8
a = np.ones((m, n))
b = np.ones((m, n))
device = dpctl.SyclDevice("opencl:gpu")
with assert_auto_offloading(), dppy.offload_to_sycl_device(device):
f(a, b)
for i in range(4):
self.assertTrue(b[i, 0] == a[i, 0] * 10)
def test_unary_ops(filter_str, unary_op, input_arrays):
if skip_test(filter_str):
pytest.skip()
a = input_arrays[0]
uop = getattr(np, unary_op)
actual = np.empty(shape=a.shape, dtype=a.dtype)
expected = np.empty(shape=a.shape, dtype=a.dtype)
@njit
def f(a):
return uop(a)
device = dpctl.SyclDevice(filter_str)
with dppy.offload_to_sycl_device(device), assert_auto_offloading():
actual = f(a)
expected = uop(a)
np.testing.assert_allclose(actual, expected, rtol=1e-5, atol=0)
def test_njit(filter_str):
@vectorize(nopython=True)
def axy(a, x, y):
return a * x + y
def f(a0, a1):
return np.cos(axy(a0, np.sin(a1) - 1.0, 1.0))
A = np.random.random(10)
B = np.random.random(10)
device = dpctl.SyclDevice(filter_str)
with dpctl.device_context(device), assert_auto_offloading():
f_njit = njit(f)
expected = f_njit(A, B)
actual = f(A, B)
max_abs_err = expected.sum() - actual.sum()
assert max_abs_err < 1e-5
def test_nested_prange(self):
@njit
def f(a, b):
# dimensions must be provided as scalar
m, n = a.shape
for i in prange(m):
for j in prange(n):
b[i, j] = a[i, j] * 10
m = 8
n = 8
a = np.ones((m, n))
b = np.ones((m, n))
device = dpctl.SyclDevice("opencl:gpu")
with assert_auto_offloading(), dppy.offload_to_sycl_device(device):
f(a, b)
self.assertTrue(np.all(b == 10))
def test_unary_ops(filter_str, unary_op, input_arrays):
# FIXME: Why does sign fail on Gen12 discrete graphics card?
skip_ops = ["sign", "log", "log2", "log10", "expm1"]
if unary_op in skip_ops and is_gen12(filter_str):
pytest.skip()
a = input_arrays[0]
uop = getattr(np, unary_op)
actual = np.empty(shape=a.shape, dtype=a.dtype)
expected = np.empty(shape=a.shape, dtype=a.dtype)
@njit
def f(a):
return uop(a)
device = dpctl.SyclDevice(filter_str)
with dpctl.device_context(device), assert_auto_offloading():
actual = f(a)
expected = uop(a)
np.testing.assert_allclose(actual, expected, rtol=1e-5, atol=0)
def test_2d_arrays(self):
def prange_example():
n = 10
a = np.ones((n, n), dtype=np.float64)
b = np.ones((n, n), dtype=np.float64)
c = np.ones((n, n), dtype=np.float64)
for i in prange(n // 2):
a[i] = b[i] + c[i]
return a
jitted = njit(prange_example)
device = dpctl.SyclDevice("opencl:gpu")
with assert_auto_offloading(
parfor_offloaded=2), dpctl.device_context(device):
jitted_res = jitted()
res = prange_example()
np.testing.assert_equal(res, jitted_res)
def test_three_prange(self):
@njit
def f(a, b):
# dimensions must be provided as scalar
m, n, o = a.shape
for i in prange(m):
val = 10
for j in prange(n):
constant = 2
for k in prange(o):
b[i, j, k] = a[i, j, k] * (val + constant)
m = 8
n = 8
o = 8
a = np.ones((m, n, o))
b = np.ones((m, n, o))
device = dpctl.SyclDevice("opencl:gpu")
with assert_auto_offloading(
parfor_offloaded=1), dpctl.device_context(device):
f(a, b)
assert np.all(b == 12)
