def test_save_different_dtype_unallocated(self):
devices = ['cpu']
if torch.cuda.is_available():
devices.append('cuda')
def save_load_check(a, b):
with io.BytesIO() as f:
torch.save([a, b], f)
f.seek(0)
a_loaded, b_loaded = torch.load(f)
self.assertEqual(a, a_loaded)
self.assertEqual(b, b_loaded)
for device, dtype in product(
devices,
all_types_and_complex_and(torch.half, torch.bfloat16,
torch.bool)):
a = torch.tensor([], dtype=dtype, device=device)
for other_dtype in all_types_and_complex_and(
torch.half, torch.bfloat16, torch.bool):
s = torch._TypedStorage(wrap_storage=a.storage()._untyped(),
dtype=other_dtype)
save_load_check(a, s)
save_load_check(a.storage(), s)
b = torch.tensor([], dtype=other_dtype, device=device)
save_load_check(a, b)
def get_supported_dtypes(op, sample_inputs_fn, device_type):
# Returns the supported dtypes for the given operator and device_type pair.
assert device_type in ['cpu', 'cuda']
if not TEST_CUDA and device_type == 'cuda':
warnings.warn(
"WARNING: CUDA is not available, empty_dtypes dispatch will be returned!"
)
return _dynamic_dispatch_dtypes(())
supported_dtypes = set()
for dtype in all_types_and_complex_and(torch.bool, torch.bfloat16,
torch.half):
try:
samples = sample_inputs_fn(op, device_type, dtype, False)
except RuntimeError:
# If `sample_inputs_fn` doesn't support sampling for a given
# `dtype`, we assume that the `dtype` is not supported.
# We raise a warning, so that user knows that this was the case
# and can investigate if there was an issue with the `sample_inputs_fn`.
warnings.warn(
f"WARNING: Unable to generate sample for device:{device_type} and dtype:{dtype}"
)
continue
# We assume the dtype is supported
# only if all samples pass for the given dtype.
supported = True
for sample in samples:
try:
op(sample.input, *sample.args, **sample.kwargs)
except RuntimeError as re:
# dtype is not supported
supported = False
break
if supported:
supported_dtypes.add(dtype)
return _dynamic_dispatch_dtypes(supported_dtypes)
class TestForeach(TestCase):
@property
def is_cuda(self):
return self.device_type == 'cuda'
# note(mkozuki): It might be the case that the expected number of `cudaLaunchKernel`s
# is greater than 1 once foreach functions internally separate their input `TensorList`s by
# devices & dtypes into vectors of tensors.
def _get_funcs(self, op, n_expected_cudaLaunchKernels: int):
return (
ForeachFuncWrapper(op.method_variant,
n_expected_cudaLaunchKernels),
RegularFuncWrapper(op.ref),
ForeachFuncWrapper(op.inplace_variant,
n_expected_cudaLaunchKernels),
RegularFuncWrapper(op.ref_inplace),
)
def _binary_test(self,
dtype,
op,
ref,
inputs,
is_fastpath,
is_inplace,
*,
alpha=None):
ref_inputs = [[t.clone().detach() for t in inputs[0]], inputs[1]
] if is_inplace else inputs
try:
actual = op(inputs, self.is_cuda, is_fastpath)
except RuntimeError as e:
with self.assertRaisesRegex(type(e), re.escape(str(e))):
ref(ref_inputs)
else:
expected = ref(ref_inputs)
self.assertEqual(actual, expected)
if alpha is not None:
kwargs = {'alpha': alpha}
ref_inputs = inputs
try:
actual = op(inputs, self.is_cuda, is_fastpath, **kwargs)
except RuntimeError as e:
with self.assertRaisesRegex(type(e), re.escape(str(e))):
ref(ref_inputs, **kwargs)
else:
expected = ref(ref_inputs, **kwargs)
if dtype in (torch.float16, torch.bfloat16) and TEST_WITH_ROCM:
self.assertEqual(expected,
actual,
atol=1.e-3,
rtol=default_tolerances(dtype)[0])
else:
self.assertEqual(expected, actual)
def _test_binary_op_tensorlists(self, device, dtype, opinfo, N,
is_fastpath, disable_fastpath):
n_expected_cudaLaunchKernels = N if disable_fastpath else 1
op, ref, inplace_op, inplace_ref = self._get_funcs(
opinfo, n_expected_cudaLaunchKernels)
inputs = [
opinfo.sample_inputs(device,
dtype,
N,
noncontiguous=not is_fastpath),
opinfo.sample_inputs(device,
dtype,
N,
noncontiguous=not is_fastpath),
]
self._binary_test(dtype,
op,
ref,
inputs,
is_fastpath,
is_inplace=False)
self._binary_test(dtype,
inplace_op,
inplace_ref,
inputs,
is_fastpath,
is_inplace=True)
if opinfo.supports_alpha_param:
alpha = None
if dtype in integral_types():
alpha = 3
elif dtype.is_complex:
alpha = complex(3, 3)
else:
alpha = 3.14
self._binary_test(dtype,
op,
ref,
inputs,
is_fastpath,
is_inplace=False,
alpha=alpha)
self._binary_test(dtype,
inplace_op,
inplace_ref,
inputs,
is_fastpath,
is_inplace=True,
alpha=alpha)
# Tests of implicit broadcasting
# When sizes of tensors don't match, foreach functions are supposed to choose slow path
# even if this methods's argument `is_fastpath` is True.
# `cudaLaunchKernel` will be equal to `N`. For assert in `ForeachFuncWrapper` to pass,
# we pass `is_fastpath and disable_fastpath` to `_binary_test`'s argument of is_fastpath.
# as n_expected_cudaLaunchKernels is N if disable_fastpath.
inputs = [
opinfo.sample_inputs(device,
dtype,
N,
noncontiguous=not is_fastpath),
[
make_tensor((N - i, 1),
device=device,
dtype=dtype,
noncontiguous=not is_fastpath) for i in range(N)
],
]
self._binary_test(dtype,
op,
ref,
inputs,
is_fastpath and disable_fastpath,
is_inplace=False)
self._binary_test(dtype,
inplace_op,
inplace_ref,
inputs,
is_fastpath and disable_fastpath,
is_inplace=True)
@skipMeta
@ops(foreach_binary_op_db)
def test_binary_op_tensorlists_fastpath(self, device, dtype, op):
for N in N_values:
disable_fastpath = op.ref == torch.div and dtype in integral_types_and(
torch.bool)
if op.ref == torch.add and dtype == torch.bool:
disable_fastpath = True
self._test_binary_op_tensorlists(device, dtype, op, N, True,
disable_fastpath)
@ops(foreach_binary_op_db)
def test_binary_op_tensorlists_slowpath(self, device, dtype, op):
for N in N_values:
self._test_binary_op_tensorlists(device, dtype, op, N, False,
False)
def _test_binary_op_scalar(self, device, dtype, opinfo, N, scalar,
is_fastpath, disable_fastpath):
n_expected_cudaLaunchKernels = N if disable_fastpath else 1
op, ref, inplace_op, inplace_ref = self._get_funcs(
opinfo, n_expected_cudaLaunchKernels)
inputs = [
opinfo.sample_inputs(device,
dtype,
N,
noncontiguous=not is_fastpath), scalar
]
self._binary_test(dtype,
op,
ref,
inputs,
is_fastpath,
is_inplace=False)
self._binary_test(dtype,
inplace_op,
inplace_ref,
inputs,
is_fastpath,
is_inplace=True)
@skipMeta
@ops(foreach_binary_op_db)
def test_binary_op_scalar_fastpath(self, device, dtype, op):
for N, scalar in itertools.product(N_values, Scalars):
disable_fastpath = op.ref == torch.div and dtype in integral_types_and(
torch.bool)
if isinstance(scalar, int):
disable_fastpath |= dtype == torch.bool
if isinstance(scalar, float):
disable_fastpath |= dtype in integral_types_and(torch.bool)
if isinstance(scalar, bool):
disable_fastpath |= dtype == torch.bool
if op.ref in (torch.add, torch.mul):
disable_fastpath = False
if isinstance(scalar, complex):
disable_fastpath |= dtype not in complex_types()
self._test_binary_op_scalar(device, dtype, op, N, scalar, True,
disable_fastpath)
@ops(foreach_binary_op_db)
def test_binary_op_scalar_slowpath(self, device, dtype, op):
for N, scalar in itertools.product(N_values, Scalars):
self._test_binary_op_scalar(device, dtype, op, N, scalar, False,
False)
def _test_binary_op_scalarlist(self, device, dtype, opinfo, N, scalarlist,
is_fastpath, disable_fastpath):
n_expected_cudaLaunchKernels = N if disable_fastpath else 1
op, ref, inplace_op, inplace_ref = self._get_funcs(
opinfo, n_expected_cudaLaunchKernels)
inputs = [
opinfo.sample_inputs(device,
dtype,
N,
noncontiguous=not is_fastpath), scalarlist
]
self._binary_test(dtype,
op,
ref,
inputs,
is_fastpath,
is_inplace=False)
self._binary_test(dtype,
inplace_op,
inplace_ref,
inputs,
is_fastpath,
is_inplace=True)
# note(mkozuki): Why two functions depending on with/without bool?
# `foreach_sub` & `foreach_sub_` do `sub_check(tensors[i], scalars[i])` from i=1...N.
# So, if scalarlist has one or more bool values, `foreach_sub` and `foreach_sub_`
# raise bool subtraction error before doing any math.
# While regular `sub` and `sub_` do some math until they encounter bool.
# So, foreach sub's throw bool sub error first. However, regular sub's throw different
# errors depending on the order of scalarlist. To keep actual unit test impl simple,
# separating mixed scalarlist tests. By setting the first element of scalarlist to bool,
# they are expected to throw bool sub error even in inplace test.
@skipMeta
@ops(foreach_binary_op_db)
def test_binary_op_scalarlist_fastpath(self, device, dtype, op):
for N in N_values:
for type_str, scalarlist in getScalarLists(N):
bool_int_div = op.ref == torch.div and dtype in integral_types_and(
torch.bool)
disable_fastpath = bool_int_div
if type_str == "int":
disable_fastpath |= dtype == torch.bool
if type_str == "float":
disable_fastpath |= dtype in integral_types_and(torch.bool)
if type_str == "complex":
disable_fastpath |= dtype not in complex_types()
if type_str == "mixed":
disable_fastpath |= True and dtype not in complex_types()
self._test_binary_op_scalarlist(device, dtype, op, N,
scalarlist, True,
disable_fastpath)
@ops(foreach_binary_op_db)
def test_binary_op_scalarlist_slowpath(self, device, dtype, op):
for N in N_values:
for _, scalarlist in getScalarLists(N):
self._test_binary_op_scalarlist(device, dtype, op, N,
scalarlist, False, False)
def _pointwise_test(self,
dtype,
op,
ref,
inputs,
is_fastpath,
is_inplace,
*,
values=None):
ref_inputs = [[t.clone().detach() for t in inputs[0]], inputs[1],
inputs[2]] if is_inplace else inputs
try:
actual = op(inputs, self.is_cuda, is_fastpath)
except RuntimeError as e:
with self.assertRaisesRegex(type(e), re.escape(str(e))):
ref(ref_inputs)
else:
expected = ref(ref_inputs)
self.assertEqual(expected, actual)
if values is not None:
try:
actual = op(inputs + [values], self.is_cuda, is_fastpath)
except RuntimeError as e:
with self.assertRaisesRegex(type(e), re.escape(str(e))):
ref(ref_inputs, values=values)
else:
expected = ref(ref_inputs, values=values)
self.assertEqual(expected, actual)
def _test_pointwise_op(self,
device,
dtype,
opinfo,
N,
is_fastpath,
disable_fastpath,
*,
values=None):
n_expected_cudaLaunchKernels = N if disable_fastpath else 1
op, ref, inplace_op, inplace_ref = self._get_funcs(
opinfo, n_expected_cudaLaunchKernels)
inputs = [
opinfo.sample_inputs(device,
dtype,
N,
noncontiguous=not is_fastpath),
opinfo.sample_inputs(device,
dtype,
N,
noncontiguous=not is_fastpath),
opinfo.sample_inputs(device,
dtype,
N,
noncontiguous=not is_fastpath),
]
self._pointwise_test(dtype,
op,
ref,
inputs,
is_fastpath,
is_inplace=False,
values=values)
self._pointwise_test(dtype,
inplace_op,
inplace_ref,
inputs,
is_fastpath,
is_inplace=True,
values=values)
# Tests of implicit broadcasting
inputs = [
opinfo.sample_inputs(device,
dtype,
N,
noncontiguous=not is_fastpath,
same_size=True),
[
make_tensor((N - i, 1),
device=device,
dtype=dtype,
noncontiguous=not is_fastpath) for i in range(N)
],
[
make_tensor((1, N - i),
device=device,
dtype=dtype,
noncontiguous=not is_fastpath) for i in range(N)
],
]
self._pointwise_test(dtype,
op,
ref,
inputs,
is_fastpath and disable_fastpath,
is_inplace=False,
values=values)
self._pointwise_test(dtype,
inplace_op,
inplace_ref,
inputs,
is_fastpath and disable_fastpath,
is_inplace=True,
values=values)
@skipMeta
@ops(foreach_pointwise_op_db)
def test_pointwise_op_fastpath(self, device, dtype, op):
disable_fastpath = dtype in integral_types_and(torch.bool)
# for N, scalar in itertools.product(N_values, Scalars):
for N in N_values:
self._test_pointwise_op(device, dtype, op, N, True,
disable_fastpath)
for scalar in Scalars:
self._test_pointwise_op(device,
dtype,
op,
N,
True,
disable_fastpath,
values=scalar)
for _, scalarlist in getScalarLists(N):
self._test_pointwise_op(device,
dtype,
op,
N,
True,
disable_fastpath,
values=scalarlist)
@ops(foreach_pointwise_op_db)
def test_pointwise_op_slowpath(self, device, dtype, op):
# for N, scalar in itertools.product(N_values, Scalars):
for N in N_values:
self._test_pointwise_op(device, dtype, op, N, False, False)
for scalar in Scalars:
self._test_pointwise_op(device,
dtype,
op,
N,
False,
False,
values=scalar)
for _, scalarlist in getScalarLists(N):
self._test_pointwise_op(device,
dtype,
op,
N,
False,
False,
values=scalarlist)
# note(mkozuki): fastpath test uses dtypes which fastpath implementation supports.
# To confirm the dtypes of `OpInfo` cover the dtypes that the function support,
# this test does not use `try-except` for fastpath.
def _regular_unary_test(self, dtype, op, ref, inputs, is_fastpath):
if is_fastpath:
self.assertEqual(ref(inputs), op(inputs, self.is_cuda,
is_fastpath))
return
try:
actual = op(inputs, self.is_cuda, is_fastpath)
except RuntimeError as e:
with self.assertRaisesRegex(type(e), re.escape(str(e))):
ref(inputs)
else:
expected = ref(inputs)
self.assertEqual(actual, expected)
# note(mkozuki): why `try-except` for both fastpath?
# - inputs for fastpath can be integer tensors.
# - this is becase opinfo dtypes are configured for outpulace implementation
# - for integer inputs, trigonometric functions and exponential function returns float outputs,
# which causes "result type Float can't be case to the desired type" error.
# Thus, `try-except` is used even if `is_fastpath` is `True`.
def _inplace_unary_test(self, dtype, inplace, inplace_ref, inputs,
is_fastpath):
copied_inputs = [[t.clone().detach() for t in tensors]
for tensors in inputs]
try:
inplace(inputs, self.is_cuda, is_fastpath)
except RuntimeError as e:
with self.assertRaisesRegex(type(e), re.escape(str(e))):
inplace_ref(copied_inputs)
else:
inplace_ref(copied_inputs),
self.assertEqual(copied_inputs, inputs)
def _test_unary(self, device, dtype, opinfo, N, is_fastpath):
op, ref, inplace_op, inplace_ref = self._get_funcs(opinfo, 1)
inputs = opinfo.sample_inputs(device,
dtype,
N,
noncontiguous=not is_fastpath),
# note(mkozuki): Complex inputs for `_foreach_abs` go through slowpath.
if opinfo.name == "_foreach_abs" and dtype in complex_types():
is_fastpath = False
self._regular_unary_test(dtype, op, ref, inputs, is_fastpath)
self._inplace_unary_test(dtype, inplace_op, inplace_ref, inputs,
is_fastpath)
@skipMeta
@ops(foreach_unary_op_db)
def test_unary_fastpath(self, device, dtype, op):
for N in N_values:
self._test_unary(device, dtype, op, N, is_fastpath=True)
@ops(foreach_unary_op_db,
dtypes=all_types_and_complex_and(torch.half, torch.bfloat16,
torch.bool))
def test_unary_slowpath(self, device, dtype, op):
for N in N_values:
self._test_unary(device, dtype, op, N, is_fastpath=False)
# note(crcrpar): `torch.maximum` and `torch.minimum` support `out` arg but there seem to be no inplace versions.
# So, compare `inplace_op` results with `ref`'s outputs.
def _minmax_test(self, opinfo, inputs, is_fastpath,
n_expected_cudaLaunchKernels):
op, ref, inplace_op, _ = self._get_funcs(opinfo,
n_expected_cudaLaunchKernels)
expected = ref(inputs)
self.assertEqual(expected, op(inputs, self.is_cuda, is_fastpath))
inplace_inputs = [[t.clone() for t in inputs[0]], inputs[1]]
inplace_op(inplace_inputs, self.is_cuda, is_fastpath)
self.assertEqual(expected, inplace_inputs[0])
@ops(foreach_minmax_op_db)
def test_minmax_fastpath(self, device, dtype, op):
for N in N_values:
inputs = tuple(
op.sample_inputs(device, dtype, N) for _ in range(2))
self._minmax_test(op, inputs, True,
N if dtype == torch.bool else 1)
@ops(foreach_minmax_op_db,
dtypes=all_types_and(torch.half, torch.bfloat16, torch.bool))
def test_minmax_slowpath(self, device, dtype, op):
for N in N_values:
inputs = tuple(
op.sample_inputs(device, dtype, N, noncontiguous=True)
for _ in range(2))
self._minmax_test(op, inputs, False, 1)
# note(mkozuki): ForeachFuncInfo's of both `_foreach_maximum` and `_foreach_minimum` include integer types.
# so, manually limit dtypes to fp types for inf&nan tests.
@ops(foreach_minmax_op_db,
dtypes=floating_types_and(torch.half, torch.bfloat16))
def test_minmax_float_inf_nan(self, device, dtype, op):
inputs = (
[
torch.tensor([float('inf')], device=device, dtype=dtype),
torch.tensor([-float('inf')], device=device, dtype=dtype),
torch.tensor([float('nan')], device=device, dtype=dtype),
torch.tensor([float('nan')], device=device, dtype=dtype)
],
[
torch.tensor([-float('inf')], device=device, dtype=dtype),
torch.tensor([float('inf')], device=device, dtype=dtype),
torch.tensor([float('inf')], device=device, dtype=dtype),
torch.tensor([float('nan')], device=device, dtype=dtype)
],
)
self._minmax_test(op, inputs, True, 1)
def _reduce_test(self, opinfo, inputs, ord, is_fastpath,
n_expected_cudaLaunchKernels):
op, ref, _, _ = self._get_funcs(opinfo, n_expected_cudaLaunchKernels)
self.assertEqual(ref(inputs, ord=ord),
op(inputs, self.is_cuda, is_fastpath, ord=ord))
@ops(foreach_reduce_op_db)
def test_reduce_fastpath(self, device, dtype, op):
for N, ord in itertools.product(N_values, (0, 1, 2, -1, -2)):
if ord in (1, 2) and dtype in floating_types_and(
torch.half, torch.bfloat16):
n_expected_cudaLaunchKernels = 3
else:
n_expected_cudaLaunchKernels = N
inputs = op.sample_inputs(device, dtype, N, noncontiguous=False),
self._reduce_test(op, inputs, ord, True,
n_expected_cudaLaunchKernels)
@ops(foreach_reduce_op_db)
def test_reduce_slowpath(self, device, dtype, op):
for N, ord in itertools.product(N_values, (0, 1, 2, -1, -2)):
inputs = op.sample_inputs(device, dtype, N, noncontiguous=True),
self._reduce_test(op, inputs, ord, False, 1)
@dtypes(*all_types_and_complex_and(torch.half, torch.bfloat16, torch.bool))
def test_add_scalar_with_empty_list_and_empty_tensor(self, device, dtype):
# TODO: enable empty list case
for tensors in [[torch.randn([0])]]:
res = torch._foreach_add(tensors, 1)
self.assertEqual(res, tensors)
torch._foreach_add_(tensors, 1)
self.assertEqual(res, tensors)
@ops(foreach_binary_op_db,
dtypes=all_types_and_complex_and(torch.half, torch.bfloat16,
torch.bool))
def test_binary_op_scalar_with_overlapping_tensors(self, device, dtype,
op):
foreach_op, ref = op.method_variant, op.ref
tensors = [
torch.ones(1, 1, device=device, dtype=dtype).expand(2, 1, 3)
]
if ref == torch.sub and dtype == torch.bool:
with self.assertRaisesRegex(RuntimeError,
re.escape(_BOOL_SUB_ERR_MSG)):
[ref(t, 1) for t in tensors]
with self.assertRaisesRegex(RuntimeError,
re.escape(_BOOL_SUB_ERR_MSG)):
foreach_op(tensors, 1)
return
expected = [ref(t, 1) for t in tensors]
res = foreach_op(tensors, 1)
self.assertEqual(res, expected)
# note(mkozuki): this test case fails with Meta at least in my local environment.
# The message was
# `AssertionError: NotImplementedError("Could not run 'aten::_foreach_add.Scalar' with arguments from the 'Meta' backend.`
@skipMeta
@ops(foreach_binary_op_db, allowed_dtypes=[torch.float])
def test_binary_op_scalar_with_different_tensor_dtypes(
self, device, dtype, op):
foreach_op = op.method_variant
tensors = [
torch.tensor([1.1], dtype=torch.float, device=device),
torch.tensor([1], dtype=torch.long, device=device)
]
runtime_error = None
try:
foreach_op(tensors, 1)
except RuntimeError as e:
runtime_error = e
self.assertIsNone(runtime_error)
@ops(foreach_binary_op_db,
dtypes=all_types_and_complex_and(torch.half, torch.bfloat16,
torch.bool))
def test_binary_op_list_error_cases(self, device, dtype, op):
foreach_op, foreach_op_, ref, ref_ = op.method_variant, op.inplace_variant, op.ref, op.ref_inplace
tensors1 = []
tensors2 = []
# Empty lists
with self.assertRaisesRegex(
RuntimeError,
"There were no tensor arguments to this function"):
foreach_op(tensors1, tensors2)
with self.assertRaisesRegex(
RuntimeError,
"There were no tensor arguments to this function"):
foreach_op_(tensors1, tensors2)
# One empty list
tensors1.append(torch.tensor([1], device=device, dtype=dtype))
with self.assertRaisesRegex(
RuntimeError,
"Tensor list must have same number of elements as scalar list."
):
foreach_op(tensors1, tensors2)
with self.assertRaisesRegex(
RuntimeError,
"Tensor list must have same number of elements as scalar list."
):
foreach_op_(tensors1, tensors2)
# Lists have different amount of tensors
tensors2.append(torch.tensor([1], device=device))
tensors2.append(torch.tensor([1], device=device))
with self.assertRaisesRegex(
RuntimeError,
"Tensor lists must have the same number of tensors, got 1 and 2"
):
foreach_op(tensors1, tensors2)
with self.assertRaisesRegex(
RuntimeError,
"Tensor lists must have the same number of tensors, got 1 and 2"
):
foreach_op_(tensors1, tensors2)
# Corresponding tensors with different sizes that aren't compatible with broadcast
# If sizes are different then foreach chooses slow path, thus error messages are expected
# to be the same as torch regular function.
tensors1 = [
torch.zeros(10, 10, device=device, dtype=dtype) for _ in range(10)
]
tensors2 = [
torch.ones(11, 11, device=device, dtype=dtype) for _ in range(10)
]
try:
foreach_op(tensors1, tensors2)
except RuntimeError as e:
with self.assertRaisesRegex(type(e), re.escape(str(e))):
[ref(t1, t2) for t1, t2 in zip(tensors1, tensors2)]
try:
foreach_op_(tensors1, tensors2)
except RuntimeError as e:
with self.assertRaisesRegex(type(e), re.escape(str(e))):
[ref_(t1, t2) for t1, t2 in zip(tensors1, tensors2)]
# different devices
if self.device_type == "cuda" and torch.cuda.device_count() > 1:
tensor1 = torch.zeros(10, 10, device="cuda:0", dtype=dtype)
tensor2 = torch.ones(10, 10, device="cuda:1", dtype=dtype)
if dtype == torch.bool and foreach_op == torch._foreach_sub:
with self.assertRaisesRegex(RuntimeError,
re.escape(_BOOL_SUB_ERR_MSG)):
foreach_op([tensor1], [tensor2])
with self.assertRaisesRegex(RuntimeError,
re.escape(_BOOL_SUB_ERR_MSG)):
foreach_op_([tensor1], [tensor2])
return
with self.assertRaisesRegex(
RuntimeError,
"Expected all tensors to be on the same device"):
foreach_op([tensor1], [tensor2])
if dtype in integral_types_and(
torch.bool) and foreach_op == torch._foreach_div:
with self.assertRaisesRegex(RuntimeError, "result type"):
foreach_op_([tensor1], [tensor2])
else:
with self.assertRaisesRegex(
RuntimeError,
"Expected all tensors to be on the same device"):
foreach_op_([tensor1], [tensor2])
@skipMeta
@unittest.skipIf(not torch.cuda.is_available(), "CUDA not found")
@ops(foreach_binary_op_db,
dtypes=all_types_and_complex_and(torch.half, torch.bfloat16,
torch.bool))
def test_binary_op_list_slow_path(self, device, dtype, op):
# note(mkozuki): why `n_expected_cudaLaunchKernels=0`?
# In this test, foreach functions don't go through fast path,
# but as there is only one tensor in each list of tensors,
# `cudaLaunchKernel` is 1 so ForeachFuncWrapper internal assert fails.
foreach_op, native_op, foreach_op_, native_op_ = self._get_funcs(
op, n_expected_cudaLaunchKernels=0)
# 0-strides
tensor1 = make_tensor((10, 10), dtype=dtype, device=device)
tensor2 = make_tensor((1, ), device=device,
dtype=dtype).expand_as(tensor1)
inputs = ([tensor1], [tensor2])
self._binary_test(dtype,
foreach_op,
native_op,
inputs,
is_fastpath=False,
is_inplace=False)
self._binary_test(dtype,
foreach_op_,
native_op_,
inputs,
is_fastpath=False,
is_inplace=True)
# different strides
tensor1 = torch.zeros(10, 10, device=device, dtype=dtype)
tensor2 = torch.ones(10, 10, device=device, dtype=dtype)
inputs = ([tensor1], [tensor2.t()])
self._binary_test(dtype,
foreach_op,
native_op,
inputs,
is_fastpath=False,
is_inplace=False)
self._binary_test(dtype,
foreach_op_,
native_op_,
inputs,
is_fastpath=False,
is_inplace=True)
# non contiguous
tensor1 = make_tensor((5, 2, 1, 3),
device=device,
dtype=dtype,
noncontiguous=True)
tensor2 = make_tensor((5, 2, 1, 3),
device=device,
dtype=dtype,
noncontiguous=True)
self.assertFalse(tensor1.is_contiguous())
self.assertFalse(tensor2.is_contiguous())
inputs = ([tensor1], [tensor2])
self._binary_test(dtype,
foreach_op,
native_op,
inputs,
is_fastpath=False,
is_inplace=False)
self._binary_test(dtype,
foreach_op_,
native_op_,
inputs,
is_fastpath=False,
is_inplace=True)
# sliced tensor
tensor1 = make_tensor((5, 2, 1, 3), device=device, dtype=dtype)
tensor2 = make_tensor((5, 2, 1, 3 * 7), device=device,
dtype=dtype)[:, :, :, ::7]
inputs = ([tensor1], [tensor2])
self._binary_test(dtype,
foreach_op,
native_op,
inputs,
is_fastpath=False,
is_inplace=False)
self._binary_test(dtype,
foreach_op_,
native_op_,
inputs,
is_fastpath=False,
is_inplace=True)
# note: Below three tests (postfixed with `_tensors_on_different_devices`)
# checks whether foreach works with lists of tensors on different devices
# but tensors of the same index are on the same device, e.g., ['cuda', 'cpu].
@onlyCUDA
@ops(foreach_unary_op_db)
def test_unary_op_tensors_on_different_devices(self, device, dtype, op):
method, ref, inplace_method, ref_inplace = self._get_funcs(op, 1)
# tensors: ['cuda', 'cpu]
tensors = op.sample_inputs(device, dtype, 2)
tensors[1] = tensors[1].to('cpu')
try:
actual = method((tensors, ), False, False)
except RuntimeError as e:
with self.assertRaisesRegex(type(e), str(e)):
ref((tensors, ))
else:
expected = ref((tensors, ))
self.assertEqual(expected, actual)
try:
inplace_method((tensors, ), False, False)
except RuntimeError as e:
with self.assertRaisesRegex(type(e), str(e)):
ref_inplace((tensors, ))
else:
self.assertEqual(expected, tensors)
@onlyCUDA
@ops(foreach_binary_op_db)
def test_binary_op_tensors_on_different_devices(self, device, dtype, op):
# `tensors1`: ['cuda', 'cpu']
# `tensors2`: ['cuda', 'cpu']
_cuda_tensors = op.sample_inputs(device, dtype, 2, same_size=True)
_cpu_tensors = op.sample_inputs('cpu', dtype, 2, same_size=True)
tensors1, tensors2 = list(
tensors for tensors in zip(_cuda_tensors, _cpu_tensors))
foreach_op, foreach_op_ = op.method_variant, op.inplace_variant
native_op, native_op_ = op.ref, op.ref_inplace
try:
actual = foreach_op(tensors1, tensors2)
except RuntimeError as e:
with self.assertRaisesRegex(type(e), re.escape(str(e))):
[native_op(t1, t2) for t1, t2 in zip(tensors1, tensors2)]
else:
expected = [
native_op(t1, t2) for t1, t2 in zip(tensors1, tensors2)
]
self.assertEqual(expected, actual)
try:
foreach_op_(tensors1, tensors2)
except RuntimeError as e:
with self.assertRaisesRegex(type(e), re.escape(str(e))):
[native_op_(t1, t2) for t1, t2 in zip(tensors1, tensors2)]
else:
self.assertEqual(actual, tensors1)
@onlyCUDA
@ops(foreach_pointwise_op_db, allowed_dtypes=floating_types())
def test_pointwise_op_tensors_on_different_devices(self, device, dtype,
op):
# tensors1: ['cuda', 'cpu]
# tensors2: ['cuda', 'cpu]
# tensors3: ['cuda', 'cpu]
_cuda_tensors = op.sample_inputs(device, dtype, 3, same_size=True)
_cpu_tensors = op.sample_inputs('cpu', dtype, 3, same_size=True)
tensors1, tensors2, tensors3 = list(
tensors for tensors in zip(_cuda_tensors, _cpu_tensors))
foreach_op, foreach_op_, native_op = op.method_variant, op.inplace_variant, op.ref
actual = foreach_op(tensors1, tensors2, tensors3)
expected = [native_op(*_cuda_tensors), native_op(*_cpu_tensors)]
self.assertEqual(expected, actual)
# note(mkozuki): Limiting dtypes to FP32&FP64, we can safely run inplace ops.
foreach_op_(tensors1, tensors2, tensors3)
self.assertEqual(expected, tensors1)
# note: BFloat16 has the same number of exponent bits as FP32
# so if squared L2 norm overflows in BF16, then it also overflows in FP32.
@onlyCUDA
@ops(foreach_reduce_op_db, allowed_dtypes=(torch.half, torch.bfloat16))
def test_foreach_l2_large_value_input(self, device, dtype, op):
ord, N = 2, 10
max_value = torch.finfo(dtype).max
scaler = torch.tensor([max_value]).sqrt().to(device=device,
dtype=dtype)
inputs = [
t * scaler for t in op.sample_inputs(
device, dtype, N, noncontiguous=False, low=1)
],
# make sure that the min. of squared L2 norm value per tensor is greater than the max value of `dtype`.
self.assertTrue(scaler * scaler * N > max_value)
fn, ref_fn, *_ = self._get_funcs(op, 3)
actual = fn(inputs, is_cuda=True, is_fastpath=True, ord=ord)
expect = ref_fn(inputs, ord=ord)
if dtype == torch.float16:
# making sure the reference L2 norm values are in the range of FP16.
self.assertFalse(any(torch.isinf(e) for e in expect))
else:
self.assertTrue(all(torch.isinf(e) for e in expect))
self.assertEqual(expect, actual, equal_nan=False)
class TestShapeOps(TestCase):
# TODO: update to work on CUDA, too
@onlyCPU
def test_unbind(self, device):
x = torch.rand(2, 3, 4, 5)
for dim in range(4):
res = torch.unbind(x, dim)
res2 = x.unbind(dim)
self.assertEqual(x.size(dim), len(res))
self.assertEqual(x.size(dim), len(res2))
for i in range(dim):
self.assertEqual(x.select(dim, i), res[i])
self.assertEqual(x.select(dim, i), res2[i])
# TODO: update to work on CUDA, too?
@onlyCPU
def test_tolist(self, device):
list0D = []
tensor0D = torch.tensor(list0D)
self.assertEqual(tensor0D.tolist(), list0D)
table1D = [1., 2., 3.]
tensor1D = torch.tensor(table1D)
storage = torch.Storage(table1D)
self.assertEqual(tensor1D.tolist(), table1D)
self.assertEqual(storage.tolist(), table1D)
self.assertEqual(tensor1D.tolist(), table1D)
self.assertEqual(storage.tolist(), table1D)
table2D = [[1, 2], [3, 4]]
tensor2D = torch.tensor(table2D)
self.assertEqual(tensor2D.tolist(), table2D)
tensor3D = torch.tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])
tensorNonContig = tensor3D.select(1, 1)
self.assertFalse(tensorNonContig.is_contiguous())
self.assertEqual(tensorNonContig.tolist(), [[3, 4], [7, 8]])
@dtypes(torch.int64, torch.float, torch.complex128)
def test_movedim_invalid(self, device, dtype):
shape = self._rand_shape(4, min_size=5, max_size=10)
x = _generate_input(shape, dtype, device, False)
for fn in [torch.movedim, torch.moveaxis]:
# Invalid `source` and `destination` dimension
with self.assertRaisesRegex(IndexError, "Dimension out of range"):
fn(x, 5, 0)
with self.assertRaisesRegex(IndexError, "Dimension out of range"):
fn(x, 0, 5)
# Mismatch in size of `source` and `destination`
with self.assertRaisesRegex(
RuntimeError,
"movedim: Invalid source or destination dims:"):
fn(x, (1, 0), (0, ))
with self.assertRaisesRegex(RuntimeError,
"movedim: repeated dim in `source`"):
fn(x, (0, 0), (0, 1))
with self.assertRaisesRegex(RuntimeError,
"movedim: repeated dim in `source`"):
fn(x, (0, 1, 0), (0, 1, 2))
with self.assertRaisesRegex(
RuntimeError, "movedim: repeated dim in `destination`"):
fn(x, (0, 1), (1, 1))
with self.assertRaisesRegex(
RuntimeError, "movedim: repeated dim in `destination`"):
fn(x, (0, 1, 2), (1, 0, 1))
@dtypes(torch.int64, torch.float, torch.complex128)
def test_movedim(self, device, dtype):
for fn in [torch.moveaxis, torch.movedim]:
for nd in range(5):
shape = self._rand_shape(nd, min_size=5, max_size=10)
x = _generate_input(shape, dtype, device, with_extremal=False)
for random_negative in [True, False]:
for src_dim, dst_dim in permutations(range(nd), r=2):
random_prob = random.random()
if random_negative and random_prob > 0.66:
src_dim = src_dim - nd
elif random_negative and random_prob > 0.33:
dst_dim = dst_dim - nd
elif random_negative:
src_dim = src_dim - nd
dst_dim = dst_dim - nd
# Integer `source` and `destination`
torch_fn = partial(fn,
source=src_dim,
destination=dst_dim)
np_fn = partial(np.moveaxis,
source=src_dim,
destination=dst_dim)
self.compare_with_numpy(torch_fn,
np_fn,
x,
device=None,
dtype=None)
if nd == 0:
continue
def make_index_negative(sequence, idx):
sequence = list(sequence)
sequence[random_idx] = sequence[random_idx] - nd
return tuple(src_sequence)
for src_sequence in permutations(range(nd),
r=random.randint(1, nd)):
# Sequence `source` and `destination`
dst_sequence = tuple(
random.sample(range(nd), len(src_sequence)))
# Randomly change a dim to a negative dim representation of itself.
random_prob = random.random()
if random_negative and random_prob > 0.66:
random_idx = random.randint(
0,
len(src_sequence) - 1)
src_sequence = make_index_negative(
src_sequence, random_idx)
elif random_negative and random_prob > 0.33:
random_idx = random.randint(
0,
len(src_sequence) - 1)
dst_sequence = make_index_negative(
dst_sequence, random_idx)
elif random_negative:
random_idx = random.randint(
0,
len(src_sequence) - 1)
dst_sequence = make_index_negative(
dst_sequence, random_idx)
random_idx = random.randint(
0,
len(src_sequence) - 1)
src_sequence = make_index_negative(
src_sequence, random_idx)
torch_fn = partial(fn,
source=src_sequence,
destination=dst_sequence)
np_fn = partial(np.moveaxis,
source=src_sequence,
destination=dst_sequence)
self.compare_with_numpy(torch_fn,
np_fn,
x,
device=None,
dtype=None)
# Move dim to same position
x = torch.randn(2, 3, 5, 7, 11)
torch_fn = partial(fn, source=(0, 1), destination=(0, 1))
np_fn = partial(np.moveaxis, source=(0, 1), destination=(0, 1))
self.compare_with_numpy(torch_fn,
np_fn,
x,
device=None,
dtype=None)
torch_fn = partial(fn, source=1, destination=1)
np_fn = partial(np.moveaxis, source=1, destination=1)
self.compare_with_numpy(torch_fn,
np_fn,
x,
device=None,
dtype=None)
# Empty Sequence
torch_fn = partial(fn, source=(), destination=())
np_fn = partial(np.moveaxis, source=(), destination=())
self.compare_with_numpy(torch_fn,
np_fn,
x,
device=None,
dtype=None)
@dtypes(torch.float, torch.bool)
def test_diag(self, device, dtype):
if dtype is torch.bool:
x = torch.rand(100, 100, device=device) >= 0.5
else:
x = torch.rand(100, 100, dtype=dtype, device=device)
res1 = torch.diag(x)
res2 = torch.tensor((), dtype=dtype, device=device)
torch.diag(x, out=res2)
self.assertEqual(res1, res2)
def test_diagonal(self, device):
x = torch.randn((100, 100), device=device)
result = torch.diagonal(x)
expected = torch.diag(x)
self.assertEqual(result, expected)
x = torch.randn((100, 100), device=device)
result = torch.diagonal(x, 17)
expected = torch.diag(x, 17)
self.assertEqual(result, expected)
@onlyCPU
@dtypes(torch.float)
def test_diagonal_multidim(self, device, dtype):
x = torch.randn(10, 11, 12, 13, dtype=dtype, device=device)
xn = x.numpy()
for args in [(2, 2, 3), (2, ), (-2, 1, 2), (0, -2, -1)]:
result = torch.diagonal(x, *args)
expected = xn.diagonal(*args)
self.assertEqual(expected.shape, result.shape)
self.assertEqual(expected, result)
# test non-continguous
xp = x.permute(1, 2, 3, 0)
result = torch.diagonal(xp, 0, -2, -1)
expected = xp.numpy().diagonal(0, -2, -1)
self.assertEqual(expected.shape, result.shape)
self.assertEqual(expected, result)
@onlyNativeDeviceTypes
@dtypes(*all_types())
@dtypesIfCUDA(*all_types_and(torch.half))
def test_trace(self, device, dtype):
def test(shape):
tensor = make_tensor(shape,
dtype=dtype,
device=device,
low=-9,
high=9)
expected_dtype = tensor.sum().dtype
expected_dtype = torch_to_numpy_dtype_dict[expected_dtype]
result = np.trace(tensor.cpu().numpy(), dtype=expected_dtype)
expected = torch.tensor(result, device=device)
self.assertEqual(tensor.trace(), expected)
shapes = (
[10, 1],
[1, 10],
[100, 100],
[20, 100],
[100, 20],
)
for shape in shapes:
test(shape)
def generate_clamp_baseline(self, device, dtype, *, min_vals, max_vals,
with_nans):
"""
Creates a random tensor for a given device and dtype, and computes the expected clamped
values given the min_vals and/or max_vals.
If with_nans is provided, then some values are randomly set to nan.
"""
X = torch.rand(100,
device=device).mul(50).add(-25) # uniform in [-25, 25]
X = X.to(dtype)
if with_nans:
mask = torch.randint(0,
2,
X.shape,
dtype=torch.bool,
device=device)
X[mask] = nan
if isinstance(min_vals, torch.Tensor):
min_vals = min_vals.cpu().numpy()
if isinstance(max_vals, torch.Tensor):
max_vals = max_vals.cpu().numpy()
# Use NumPy implementation as reference
X_clamped = torch.tensor(np.clip(X.cpu().numpy(),
a_min=min_vals,
a_max=max_vals),
device=device)
return X, X_clamped
# Tests clamp and its alias, clip
@dtypes(torch.int64, torch.float32)
def test_clamp(self, device, dtype):
op_list = (torch.clamp, torch.Tensor.clamp, torch.Tensor.clamp_,
torch.clip, torch.Tensor.clip, torch.Tensor.clip_)
# min/max argument product
args = product((-10, None), (10, None))
for op in op_list:
for min_val, max_val in args:
if min_val is None and max_val is None:
continue
X, Y_expected = self.generate_clamp_baseline(device,
dtype,
min_vals=min_val,
max_vals=max_val,
with_nans=False)
# Test op
X1 = X.clone() # So that the in-place ops do not change X
Y_actual = op(X1, min_val, max_val)
self.assertEqual(Y_expected, Y_actual)
# Test op-out behavior (out does not exist for method versions)
if op in (torch.clamp, torch.clip):
Y_out = torch.empty_like(X)
op(X, min=min_val, max=max_val, out=Y_out)
self.assertEqual(Y_expected, Y_out)
def test_clamp_propagates_nans(self, device):
op_list = (torch.clamp, torch.Tensor.clamp, torch.Tensor.clamp_,
torch.clip, torch.Tensor.clip, torch.Tensor.clip_)
# min/max argument product
args = product((-10, None), (10, None))
for op in op_list:
for min_val, max_val in args:
if min_val is None and max_val is None:
continue
X, Y_expected = self.generate_clamp_baseline(device,
torch.float,
min_vals=min_val,
max_vals=max_val,
with_nans=True)
Y_expected = torch.isnan(Y_expected)
# Test op
X1 = X.clone() # So that the in-place ops do not change X
Y_actual = op(X1, min_val, max_val)
self.assertEqual(Y_expected, torch.isnan(Y_actual))
# Test op-out behavior (out does not exist for method versions)
if op in (torch.clamp, torch.clip):
Y_out = torch.empty_like(X)
op(X, min_val, max_val, out=Y_out)
self.assertEqual(Y_expected, torch.isnan(Y_out))
def test_clamp_raises_arg_errors(self, device):
X = torch.randn(100, dtype=torch.float, device=device)
error_msg = 'At least one of \'min\' or \'max\' must not be None'
with self.assertRaisesRegex(RuntimeError, error_msg):
X.clamp()
with self.assertRaisesRegex(RuntimeError, error_msg):
X.clamp_()
with self.assertRaisesRegex(RuntimeError, error_msg):
torch.clamp(X)
@dtypes(*all_types_and_complex_and(torch.half, torch.bool, torch.bfloat16))
def test_flip(self, device, dtype):
make_from_data = partial(torch.tensor, device=device, dtype=dtype)
make_from_size = partial(make_tensor, device=device, dtype=dtype)
def test_flip_impl(input_t, dims, output_t):
def all_t():
yield input_t, output_t
if dtype is torch.float:
# We generate quantized versions as well
for qdtype in (torch.quint8, torch.qint8, torch.qint32):
qinput_t = torch.quantize_per_tensor(
input_t, 0.1, 5, qdtype)
qoutput_t = torch.quantize_per_tensor(
output_t, 0.1, 5, qdtype)
yield qinput_t, qoutput_t
for in_t, out_t in all_t():
self.assertEqual(in_t.flip(dims), out_t)
n = in_t.ndim
if not isinstance(dims, tuple):
# Wrap dim
self.assertEqual(in_t.flip(-n + dims), out_t)
else:
# Permute dimensions
for p_dims in permutations(dims):
self.assertEqual(in_t.flip(p_dims), out_t)
if len(p_dims) > 0:
# Wrap 1st dim
self.assertEqual(
in_t.flip((-n + p_dims[0], ) + p_dims[1:]),
out_t)
def gen_data():
# Basic tests
data = make_from_data([1, 2, 3, 4, 5, 6, 7, 8]).view(2, 2, 2)
nonctg = make_from_size((2, 2, 2), noncontiguous=True).copy_(data)
dims_result = ((0, make_from_data([5, 6, 7, 8, 1, 2, 3,
4]).view(2, 2, 2)),
(1, make_from_data([3, 4, 1, 2, 7, 8, 5,
6]).view(2, 2, 2)),
(2, make_from_data([2, 1, 4, 3, 6, 5, 8,
7]).view(2, 2, 2)),
((0, 1), make_from_data([7, 8, 5, 6, 3, 4, 1,
2]).view(2, 2, 2)),
((0, 1, 2), make_from_data([8, 7, 6, 5, 4, 3, 2,
1]).view(2, 2, 2)))
for in_tensor, (dims, out_tensor) in product((data, nonctg),
dims_result):
yield in_tensor, dims, out_tensor
# Expanded
in_t = make_from_data([1, 2, 3]).view(3, 1).expand(3, 2)
dims = 0
out_t = make_from_data([3, 3, 2, 2, 1, 1]).view(3, 2)
yield in_t, dims, out_t
# Noop on expanded dimension
yield in_t, 1, in_t
# Transposed
in_t = make_from_data([1, 2, 3, 4, 5, 6, 7,
8]).view(2, 2, 2).transpose(0, 1)
dims = (0, 1, 2)
out_t = make_from_data([8, 7, 4, 3, 6, 5, 2, 1]).view(2, 2, 2)
yield in_t, dims, out_t
# Rectangular case
in_t = make_from_data([1, 2, 3, 4, 5, 6]).view(2, 3)
dims = 0
out_t = make_from_data([[4, 5, 6], [1, 2, 3]])
yield in_t, dims, out_t
dims = 1
out_t = make_from_data([[3, 2, 1], [6, 5, 4]])
yield in_t, dims, out_t
# Noops (edge cases)
# Size 0
in_t = make_from_data(())
yield in_t, 0, in_t
yield in_t, (), in_t
# dims = ()
in_t = make_from_size((3, 2, 1))
yield in_t, (), in_t
# Zero elements, non-zero size
in_t = make_from_size((3, 0, 2))
for i in range(in_t.ndim):
yield in_t, i, in_t
# Size 1
in_t = make_from_size(())
yield in_t, 0, in_t
in_t = make_from_size((1, ))
yield in_t, 0, in_t
for in_tensor, dims, out_tensor in gen_data():
test_flip_impl(in_tensor, dims, out_tensor)
# test for shape
size = [2, 3, 4]
data = make_from_size(size)
possible_dims = range(len(size))
test_dims = chain(combinations(possible_dims, 1),
combinations(possible_dims, 2))
for dims in test_dims:
self.assertEqual(size, list(data.flip(dims).size()))
@dtypes(*all_types_and_complex_and(torch.half, torch.bool, torch.bfloat16))
def test_flip_errors(self, device, dtype):
make_arg = partial(make_tensor, dtype=dtype, device=device)
data = make_arg((2, 2, 2))
# not allow flip on the same dim more than once
self.assertRaises(RuntimeError, lambda: data.flip(0, 1, 1))
# not allow empty list as input
self.assertRaises(TypeError, lambda: data.flip())
# not allow dim > max dim
self.assertRaises(IndexError, lambda: data.flip(0, 1, 2, 3))
self.assertRaises(IndexError, lambda: data.flip(3))
def _rand_shape(self, dim, min_size, max_size):
return tuple(torch.randint(min_size, max_size + 1, (dim, )))
@dtypes(*all_types_and_complex_and(torch.half, torch.bool, torch.bfloat16))
def test_flip_numpy(self, device, dtype):
make_arg = partial(make_tensor, dtype=dtype, device=device)
for ndim in [3, 4]:
shape = self._rand_shape(ndim, 5, 10)
data = make_arg(shape)
# Axis to sample for given shape.
for i in range(1, ndim + 1):
# Check all combinations of `i` axis.
for flip_dim in combinations(range(ndim), i):
torch_fn = partial(torch.flip, dims=flip_dim)
np_fn = partial(np.flip, axis=flip_dim)
self.compare_with_numpy(torch_fn, np_fn, data)
@onlyCUDA # CPU is too slow
@largeTensorTest('17GB'
) # 4 tensors of 4GB (in, out) x (torch, numpy) + 1GB
@largeTensorTest(
"81GB",
"cpu") # even for CUDA test, sufficient system memory is required
def test_flip_large_tensor(self, device):
t_in = torch.empty(2**32 + 1, dtype=torch.uint8).random_()
torch_fn = partial(torch.flip, dims=(0, ))
np_fn = partial(np.flip, axis=0)
self.compare_with_numpy(torch_fn, np_fn, t_in)
del t_in
def _test_fliplr_flipud(self, torch_fn, np_fn, min_dim, max_dim, device,
dtype):
for dim in range(min_dim, max_dim + 1):
shape = self._rand_shape(dim, 5, 10)
# Randomly scale the input
if dtype.is_floating_point or dtype.is_complex:
data = torch.randn(*shape, device=device, dtype=dtype)
else:
data = torch.randint(0, 10, shape, device=device, dtype=dtype)
self.compare_with_numpy(torch_fn, np_fn, data)
@dtypes(torch.int64, torch.double, torch.cdouble)
def test_fliplr(self, device, dtype):
self._test_fliplr_flipud(torch.fliplr, np.fliplr, 2, 4, device, dtype)
@dtypes(torch.int64, torch.double, torch.cdouble)
def test_fliplr_invalid(self, device, dtype):
x = torch.randn(42).to(dtype)
with self.assertRaisesRegex(RuntimeError, "Input must be >= 2-d."):
torch.fliplr(x)
with self.assertRaisesRegex(RuntimeError, "Input must be >= 2-d."):
torch.fliplr(torch.tensor(42, device=device, dtype=dtype))
@dtypes(torch.int64, torch.double, torch.cdouble)
def test_flipud(self, device, dtype):
self._test_fliplr_flipud(torch.flipud, np.flipud, 1, 4, device, dtype)
@dtypes(torch.int64, torch.double, torch.cdouble)
def test_flipud_invalid(self, device, dtype):
with self.assertRaisesRegex(RuntimeError, "Input must be >= 1-d."):
torch.flipud(torch.tensor(42, device=device, dtype=dtype))
def test_rot90(self, device):
data = torch.arange(1, 5, device=device).view(2, 2)
self.assertEqual(
torch.tensor([1, 2, 3, 4]).view(2, 2), data.rot90(0, [0, 1]))
self.assertEqual(
torch.tensor([2, 4, 1, 3]).view(2, 2), data.rot90(1, [0, 1]))
self.assertEqual(
torch.tensor([4, 3, 2, 1]).view(2, 2), data.rot90(2, [0, 1]))
self.assertEqual(
torch.tensor([3, 1, 4, 2]).view(2, 2), data.rot90(3, [0, 1]))
# test for default args k=1, dims=[0, 1]
self.assertEqual(data.rot90(), data.rot90(1, [0, 1]))
# test for reversed order of dims
self.assertEqual(data.rot90(3, [0, 1]), data.rot90(1, [1, 0]))
# test for modulo of k
self.assertEqual(data.rot90(5, [0, 1]), data.rot90(1, [0, 1]))
self.assertEqual(data.rot90(3, [0, 1]), data.rot90(-1, [0, 1]))
self.assertEqual(data.rot90(-5, [0, 1]), data.rot90(-1, [0, 1]))
# test for dims out-of-range error
self.assertRaises(RuntimeError, lambda: data.rot90(1, [0, -3]))
self.assertRaises(RuntimeError, lambda: data.rot90(1, [0, 2]))
# test tensor with more than 2D
data = torch.arange(1, 9, device=device).view(2, 2, 2)
self.assertEqual(
torch.tensor([2, 4, 1, 3, 6, 8, 5, 7]).view(2, 2, 2),
data.rot90(1, [1, 2]))
self.assertEqual(data.rot90(1, [1, -1]), data.rot90(1, [1, 2]))
# test for errors
self.assertRaises(RuntimeError, lambda: data.rot90(1, [0, 3]))
self.assertRaises(RuntimeError, lambda: data.rot90(1, [1, 1]))
self.assertRaises(RuntimeError, lambda: data.rot90(1, [0, 1, 2]))
self.assertRaises(RuntimeError, lambda: data.rot90(1, [0]))
@dtypes(torch.cfloat, torch.cdouble)
def test_complex_rot90(self, device, dtype):
shape = self._rand_shape(random.randint(2, 4), 5, 10)
for rot_times in range(4):
data = torch.randn(*shape, device=device, dtype=dtype)
torch_fn = partial(torch.rot90, k=rot_times, dims=[0, 1])
np_fn = partial(np.rot90, k=rot_times, axes=[0, 1])
self.compare_with_numpy(torch_fn, np_fn, data)
# TODO: update once warning flag is available to always trigger ONCE warnings
# Ensures nonzero does not throw a warning, even when the as_tuple argument
# is not provided
def test_nonzero_no_warning(self, device):
t = torch.randn((2, 2), device=device)
with warnings.catch_warnings(record=True) as w:
warnings.simplefilter("always")
torch.nonzero(t)
t.nonzero()
self.assertEqual(len(w), 0)
@dtypes(*all_types_and(torch.half, torch.bool, torch.bfloat16))
def test_nonzero(self, device, dtype):
shapes = [
torch.Size((12, )),
torch.Size((12, 1)),
torch.Size((1, 12)),
torch.Size((6, 2)),
torch.Size((3, 2, 2)),
torch.Size((5, 5, 5)),
]
def gen_nontrivial_input(shape, dtype, device):
if dtype != torch.bfloat16:
return torch.randint(2, shape, device=device, dtype=dtype)
else:
# windows does not work for bfloat16 randing
return torch.randint(2,
shape,
device=device,
dtype=torch.float).to(dtype)
for shape in shapes:
tensor = gen_nontrivial_input(shape, dtype, device)
dst1 = torch.nonzero(tensor, as_tuple=False)
dst2 = tensor.nonzero(as_tuple=False)
dst3 = torch.empty([], dtype=torch.long, device=device)
torch.nonzero(tensor, out=dst3)
if self.device_type != 'xla':
# xla does not raise runtime error
self.assertRaisesRegex(
RuntimeError, "scalar type Long", lambda: torch.nonzero(
tensor,
out=torch.empty([], dtype=torch.float, device=device)))
if self.device_type == 'cuda':
self.assertRaisesRegex(
RuntimeError, "on the same device", lambda: torch.nonzero(
tensor, out=torch.empty([], dtype=torch.long)))
np_array = tensor.cpu().numpy(
) if dtype != torch.bfloat16 else tensor.float().cpu().numpy()
np_result = torch.from_numpy(np.stack(np_array.nonzero())).t()
self.assertEqual(dst1.cpu(), np_result, atol=0, rtol=0)
self.assertEqual(dst2.cpu(), np_result, atol=0, rtol=0)
self.assertEqual(dst3.cpu(), np_result, atol=0, rtol=0)
tup1 = torch.nonzero(tensor, as_tuple=True)
tup2 = tensor.nonzero(as_tuple=True)
tup1 = torch.stack(tup1).t().cpu()
tup2 = torch.stack(tup2).t().cpu()
self.assertEqual(tup1, np_result, atol=0, rtol=0)
self.assertEqual(tup2, np_result, atol=0, rtol=0)
def test_nonzero_astuple_out(self, device):
t = torch.randn((3, 3, 3), device=device)
out = torch.empty_like(t, dtype=torch.long)
with self.assertRaises(RuntimeError):
torch.nonzero(t, as_tuple=True, out=out)
self.assertEqual(torch.nonzero(t, as_tuple=False, out=out),
torch.nonzero(t, out=out))
# Verifies that JIT script cannot handle the as_tuple kwarg
# See Issue https://github.com/pytorch/pytorch/issues/45499.
def _foo(t):
tuple_result = torch.nonzero(t, as_tuple=True)
nontuple_result = torch.nonzero(t, as_tuple=False)
out = torch.empty_like(nontuple_result)
torch.nonzero(t, as_tuple=False, out=out)
return tuple_result, nontuple_result, out
with self.assertRaises(RuntimeError):
scripted_foo = torch.jit.script(_foo)
# Verifies that JIT tracing works fine
traced_foo = torch.jit.trace(_foo, t)
traced_tuple, traced_nontuple, traced_out = traced_foo(t)
expected_tuple = torch.nonzero(t, as_tuple=True)
expected_nontuple = torch.nonzero(t)
self.assertEqual(traced_tuple, expected_tuple)
self.assertEqual(traced_nontuple, expected_nontuple)
self.assertEqual(traced_out, expected_nontuple)
@onlyNativeDeviceTypes
def test_nonzero_discontiguous(self, device):
shape = (4, 4)
tensor = torch.randint(2, shape, device=device)
tensor_nc = torch.empty(shape[0], shape[1] * 2,
device=device)[:, ::2].copy_(tensor)
dst1 = tensor.nonzero(as_tuple=False)
dst2 = tensor_nc.nonzero(as_tuple=False)
self.assertEqual(dst1, dst2, atol=0, rtol=0)
dst3 = torch.empty_like(dst1)
data_ptr = dst3.data_ptr()
# expect dst3 storage to be reused
torch.nonzero(tensor, out=dst3)
self.assertEqual(data_ptr, dst3.data_ptr())
self.assertEqual(dst1, dst3, atol=0, rtol=0)
# discontiguous out
dst4 = torch.empty(dst1.size(0),
dst1.size(1) * 2,
dtype=torch.long,
device=device)[:, ::2]
data_ptr = dst4.data_ptr()
strides = dst4.stride()
torch.nonzero(tensor, out=dst4)
self.assertEqual(data_ptr, dst4.data_ptr())
self.assertEqual(dst1, dst4, atol=0, rtol=0)
self.assertEqual(strides, dst4.stride())
def test_nonzero_non_diff(self, device):
x = torch.randn(10, requires_grad=True)
nz = x.nonzero()
self.assertFalse(nz.requires_grad)
class TestPythonJiterator(TestCase):
@skipCUDAIfRocm
@parametrize("shape_strides", [
(([3, 3], [3, 1]), ([3, 3], [3, 1])), # contiguous
])
@dtypes(*product(all_types_and_complex_and(torch.half, torch.bfloat16),
all_types_and_complex_and(torch.half, torch.bfloat16)))
def test_all_dtype_contiguous(self, device, dtypes, shape_strides):
a_buffer = torch.rand(9, device=device).mul(10).type(dtypes[0])
b_buffer = torch.rand(9, device=device).mul(10).type(dtypes[1])
a = a_buffer.as_strided(*shape_strides[0])
b = b_buffer.as_strided(*shape_strides[1])
expected = ref_fn(a, b)
result = jitted_fn(a, b)
self.assertEqual(expected, result)
@skipCUDAIfRocm
# See https://github.com/pytorch/pytorch/pull/76394#issuecomment-1118018287 for details
@skipCUDAIf(_get_torch_cuda_version() < (11, 6), "On cuda 11.3, nvrtcCompileProgram is taking too long to "
"compile jiterator generated kernels for non-contiguous input that requires dynamic-casting.")
@parametrize("shape_strides", [
(([3, 3], [1, 3]), ([3, 1], [1, 3])), # non-contiguous
])
@dtypes(*product(all_types_and_complex_and(torch.half, torch.bfloat16),
all_types_and_complex_and(torch.half, torch.bfloat16)))
def test_all_dtype_noncontiguous(self, device, dtypes, shape_strides):
a_buffer = torch.rand(9, device=device).mul(10).type(dtypes[0])
b_buffer = torch.rand(9, device=device).mul(10).type(dtypes[1])
a = a_buffer.as_strided(*shape_strides[0])
b = b_buffer.as_strided(*shape_strides[1])
expected = ref_fn(a, b)
result = jitted_fn(a, b)
self.assertEqual(expected, result)
@skipCUDAIfRocm
@dtypes(torch.float, torch.double, torch.float16, torch.bfloat16)
@parametrize("alpha", [-1, 2.0, None])
@parametrize("beta", [3, -4.2, None])
@toleranceOverride({torch.float16 : tol(atol=1e-2, rtol=1e-3)})
def test_extra_args(self, device, dtype, alpha, beta):
a = torch.rand(3, device=device).mul(10).type(dtype)
b = torch.rand(3, device=device).mul(10).type(dtype)
extra_args = {}
if alpha is not None:
extra_args["alpha"] = alpha
if beta is not None:
extra_args["beta"] = beta
expected = ref_fn(a, b, **extra_args)
result = jitted_fn(a, b, **extra_args)
self.assertEqual(expected, result)
@skipCUDAIfRocm
@parametrize("is_train", [True, False])
def test_bool_extra_args(self, device, is_train):
code_string = "template <typename T> T conditional(T x, T mask, bool is_train) { return is_train ? x * mask : x; }"
jitted_fn = create_jit_fn(code_string, is_train=False)
def ref_fn(x, mask, is_train):
return x * mask if is_train else x
a = torch.rand(3, device=device)
b = torch.rand(3, device=device)
expected = ref_fn(a, b, is_train=is_train)
result = jitted_fn(a, b, is_train=is_train)
self.assertEqual(expected, result)
@skipCUDAIfRocm
def test_multiple_functors(self, device):
code_string = '''
template <typename T> T fn(T x, T mask) { return x * mask; }
template <typename T> T main_fn(T x, T mask, T y) { return fn(x, mask) + y; }
'''
jitted_fn = create_jit_fn(code_string)
def ref_fn(x, mask, y):
return x * mask + y
a = torch.rand(3, device=device)
b = torch.rand(3, device=device)
c = torch.rand(3, device=device)
expected = ref_fn(a, b, c)
result = jitted_fn(a, b, c)
self.assertEqual(expected, result)
@skipCUDAIfRocm
@parametrize("num_inputs", [1, 5, 8])
def test_various_num_inputs(self, num_inputs):
inputs = []
for i in range(num_inputs):
inputs.append(torch.rand(3, device='cuda').mul(10))
input_string = ",".join([f"T i{i}" for i in range(num_inputs)])
function_body = "+".join([f"i{i}" for i in range(num_inputs)])
code_string = f"template <typename T> T my_kernel({input_string}) {{ return {function_body}; }}"
jitted_fn = create_jit_fn(code_string)
def ref_fn(*inputs):
return torch.sum(torch.stack(inputs), dim=0)
expected = ref_fn(*inputs)
result = jitted_fn(*inputs)
self.assertEqual(expected, result)
@skipCUDAIfRocm
@parametrize("num_outputs", [1, 4, 8])
def test_various_num_outputs(self, num_outputs):
input = torch.rand(3, device='cuda')
output_string = ", ".join([f"T& out{i}" for i in range(num_outputs)])
function_body = ""
for i in range(num_outputs):
function_body += f"out{i} = input + {i};\n"
code_string = f"template <typename T> T my_kernel(T input, {output_string}) {{ {function_body} }}"
jitted_fn = create_multi_output_jit_fn(code_string, num_outputs)
def ref_fn(input):
outputs = []
for i in range(num_outputs):
outputs.append(input + i)
if num_outputs == 1:
return outputs[0]
return tuple(outputs)
expected = ref_fn(input)
result = jitted_fn(input)
for i in range(num_outputs):
self.assertEqual(expected[i], result[i])
@skipCUDAIfRocm
@parametrize("code_string", [
"template <typename T> T my _kernel(T x) { return x; }",
"template <typename T> Tmy_kernel(T x) { return x; }",
])
def test_invalid_function_name(self, code_string):
with self.assertRaises(Exception):
jitted_fn = create_jit_fn(code_string)
def test_dtypes(self, device, op):
# dtypes to try to backward in
allowed_backward_dtypes = floating_and_complex_types_and(
torch.bfloat16, torch.float16)
# lists for (un)supported dtypes
supported_dtypes = []
unsupported_dtypes = []
supported_backward_dtypes = []
unsupported_backward_dtypes = []
def unsupported(dtype):
unsupported_dtypes.append(dtype)
if dtype in allowed_backward_dtypes:
unsupported_backward_dtypes.append(dtype)
for dtype in all_types_and_complex_and(torch.half, torch.bfloat16,
torch.bool):
# tries to acquire samples - failure indicates lack of support
requires_grad = (dtype in allowed_backward_dtypes
and op.supports_autograd)
try:
samples = list(
op.sample_inputs(device,
dtype,
requires_grad=requires_grad))
except Exception as e:
unsupported(dtype)
continue
# Counts number of successful backward attempts
# NOTE: This exists as a kludge because this only understands how to
# request a gradient if the output is a tensor or a sequence with
# a tensor as its first element.
num_backward_successes = 0
for sample in samples:
# tries to call operator with the sample - failure indicates
# lack of support
try:
result = op(sample.input, *sample.args, **sample.kwargs)
except Exception as e:
# NOTE: some ops will fail in forward if their inputs
# require grad but they don't support computing the gradient
# in that type! This is a bug in the op!
unsupported(dtype)
# Short-circuits testing this dtype -- it doesn't work
if dtype in unsupported_dtypes:
break
# Short-circuits if the dtype isn't a backward dtype or
# it's already identified as not supported
if dtype not in allowed_backward_dtypes or dtype in unsupported_backward_dtypes:
continue
# Checks for backward support in the same dtype
try:
result = sample.output_process_fn_grad(result)
if isinstance(result, torch.Tensor):
backward_tensor = result
elif isinstance(result, Sequence) and isinstance(
result[0], torch.Tensor):
backward_tensor = result[0]
else:
continue
# Note: this grad may not have the same dtype as dtype
# For functions like complex (float -> complex) or abs
# (complex -> float) the grad tensor will have a
# different dtype than the input.
# For simplicity, this is still modeled as these ops
# supporting grad in the input dtype.
grad = torch.randn_like(backward_tensor)
backward_tensor.backward(grad)
num_backward_successes += 1
except Exception as e:
unsupported_backward_dtypes.append(dtype)
if dtype not in unsupported_dtypes:
supported_dtypes.append(dtype)
if num_backward_successes > 0 and dtype not in unsupported_backward_dtypes:
supported_backward_dtypes.append(dtype)
# Checks that dtypes are listed correctly and generates an informative
# error message
device_type = torch.device(device).type
claimed_supported = set(op.supported_dtypes(device_type))
supported_dtypes = set(supported_dtypes)
supported_but_unclaimed = supported_dtypes - claimed_supported
claimed_but_unsupported = claimed_supported - supported_dtypes
msg = """The supported dtypes for {0} on {1} according to its OpInfo are
{2}, but the detected supported dtypes are {3}.
""".format(op.name, device_type, claimed_supported, supported_dtypes)
if len(supported_but_unclaimed) > 0:
msg += "The following dtypes should be added to the OpInfo: {0}. ".format(
supported_but_unclaimed)
if len(claimed_but_unsupported) > 0:
msg += "The following dtypes should be removed from the OpInfo: {0}.".format(
claimed_but_unsupported)
self.assertEqual(supported_dtypes, claimed_supported, msg=msg)
# Checks that backward dtypes are listed correctly and generates an
# informative error message
# NOTE: this code is nearly identical to the check + msg generation
claimed_backward_supported = set(
op.supported_backward_dtypes(device_type))
supported_backward_dtypes = set(supported_backward_dtypes)
supported_but_unclaimed = supported_backward_dtypes - claimed_backward_supported
claimed_but_unsupported = claimed_backward_supported - supported_backward_dtypes
msg = """The supported backward dtypes for {0} on {1} according to its OpInfo are
{2}, but the detected supported backward dtypes are {3}.
""".format(op.name, device_type, claimed_backward_supported,
supported_backward_dtypes)
if len(supported_but_unclaimed) > 0:
msg += "The following backward dtypes should be added to the OpInfo: {0}. ".format(
supported_but_unclaimed)
if len(claimed_but_unsupported) > 0:
msg += "The following backward dtypes should be removed from the OpInfo: {0}.".format(
claimed_but_unsupported)
self.assertEqual(supported_backward_dtypes,
claimed_backward_supported,
msg=msg)
class TestNumPyInterop(TestCase):
# Note: the warning this tests for only appears once per program, so
# other instances of this warning should be addressed to avoid
# the tests depending on the order in which they're run.
@onlyCPU
def test_numpy_non_writeable(self, device):
arr = np.zeros(5)
arr.flags['WRITEABLE'] = False
self.assertWarns(UserWarning, lambda: torch.from_numpy(arr))
@onlyCPU
def test_numpy_unresizable(self, device) -> None:
x = np.zeros((2, 2))
y = torch.from_numpy(x)
with self.assertRaises(ValueError):
x.resize((5, 5))
z = torch.randn(5, 5)
w = z.numpy()
with self.assertRaises(RuntimeError):
z.resize_(10, 10)
with self.assertRaises(ValueError):
w.resize((10, 10))
@onlyCPU
def test_to_numpy(self, device) -> None:
def get_castable_tensor(shape, dtype):
if dtype.is_floating_point:
dtype_info = torch.finfo(dtype)
# can't directly use min and max, because for double, max - min
# is greater than double range and sampling always gives inf.
low = max(dtype_info.min, -1e10)
high = min(dtype_info.max, 1e10)
t = torch.empty(shape, dtype=torch.float64).uniform_(low, high)
else:
# can't directly use min and max, because for int64_t, max - min
# is greater than int64_t range and triggers UB.
low = max(torch.iinfo(dtype).min, int(-1e10))
high = min(torch.iinfo(dtype).max, int(1e10))
t = torch.empty(shape, dtype=torch.int64).random_(low, high)
return t.to(dtype)
dtypes = [
torch.uint8,
torch.int8,
torch.short,
torch.int,
torch.half,
torch.float,
torch.double,
torch.long,
]
for dtp in dtypes:
# 1D
sz = 10
x = get_castable_tensor(sz, dtp)
y = x.numpy()
for i in range(sz):
self.assertEqual(x[i], y[i])
# 1D > 0 storage offset
xm = get_castable_tensor(sz * 2, dtp)
x = xm.narrow(0, sz - 1, sz)
self.assertTrue(x.storage_offset() > 0)
y = x.numpy()
for i in range(sz):
self.assertEqual(x[i], y[i])
def check2d(x, y):
for i in range(sz1):
for j in range(sz2):
self.assertEqual(x[i][j], y[i][j])
# empty
x = torch.tensor([]).to(dtp)
y = x.numpy()
self.assertEqual(y.size, 0)
# contiguous 2D
sz1 = 3
sz2 = 5
x = get_castable_tensor((sz1, sz2), dtp)
y = x.numpy()
check2d(x, y)
self.assertTrue(y.flags['C_CONTIGUOUS'])
# with storage offset
xm = get_castable_tensor((sz1 * 2, sz2), dtp)
x = xm.narrow(0, sz1 - 1, sz1)
y = x.numpy()
self.assertTrue(x.storage_offset() > 0)
check2d(x, y)
self.assertTrue(y.flags['C_CONTIGUOUS'])
# non-contiguous 2D
x = get_castable_tensor((sz2, sz1), dtp).t()
y = x.numpy()
check2d(x, y)
self.assertFalse(y.flags['C_CONTIGUOUS'])
# with storage offset
xm = get_castable_tensor((sz2 * 2, sz1), dtp)
x = xm.narrow(0, sz2 - 1, sz2).t()
y = x.numpy()
self.assertTrue(x.storage_offset() > 0)
check2d(x, y)
# non-contiguous 2D with holes
xm = get_castable_tensor((sz2 * 2, sz1 * 2), dtp)
x = xm.narrow(0, sz2 - 1, sz2).narrow(1, sz1 - 1, sz1).t()
y = x.numpy()
self.assertTrue(x.storage_offset() > 0)
check2d(x, y)
if dtp != torch.half:
# check writeable
x = get_castable_tensor((3, 4), dtp)
y = x.numpy()
self.assertTrue(y.flags.writeable)
y[0][1] = 3
self.assertTrue(x[0][1] == 3)
y = x.t().numpy()
self.assertTrue(y.flags.writeable)
y[0][1] = 3
self.assertTrue(x[0][1] == 3)
def test_to_numpy_bool(self, device) -> None:
x = torch.tensor([True, False], dtype=torch.bool)
self.assertEqual(x.dtype, torch.bool)
y = x.numpy()
self.assertEqual(y.dtype, np.bool_)
for i in range(len(x)):
self.assertEqual(x[i], y[i])
x = torch.tensor([True], dtype=torch.bool)
self.assertEqual(x.dtype, torch.bool)
y = x.numpy()
self.assertEqual(y.dtype, np.bool_)
self.assertEqual(x[0], y[0])
def test_from_numpy(self, device) -> None:
dtypes = [
np.double,
np.float64,
np.float16,
np.complex64,
np.complex128,
np.int64,
np.int32,
np.int16,
np.int8,
np.uint8,
np.longlong,
np.bool_,
]
complex_dtypes = [
np.complex64,
np.complex128,
]
for dtype in dtypes:
array = np.array([1, 2, 3, 4], dtype=dtype)
tensor_from_array = torch.from_numpy(array)
# TODO: change to tensor equality check once HalfTensor
# implements `==`
for i in range(len(array)):
self.assertEqual(tensor_from_array[i], array[i])
# ufunc 'remainder' not supported for complex dtypes
if dtype not in complex_dtypes:
# This is a special test case for Windows
# https://github.com/pytorch/pytorch/issues/22615
array2 = array % 2
tensor_from_array2 = torch.from_numpy(array2)
for i in range(len(array2)):
self.assertEqual(tensor_from_array2[i], array2[i])
# Test unsupported type
array = np.array([1, 2, 3, 4], dtype=np.uint16)
with self.assertRaises(TypeError):
tensor_from_array = torch.from_numpy(array)
# check storage offset
x = np.linspace(1, 125, 125)
x.shape = (5, 5, 5)
x = x[1]
expected = torch.arange(1, 126, dtype=torch.float64).view(5, 5, 5)[1]
self.assertEqual(torch.from_numpy(x), expected)
# check noncontiguous
x = np.linspace(1, 25, 25)
x.shape = (5, 5)
expected = torch.arange(1, 26, dtype=torch.float64).view(5, 5).t()
self.assertEqual(torch.from_numpy(x.T), expected)
# check noncontiguous with holes
x = np.linspace(1, 125, 125)
x.shape = (5, 5, 5)
x = x[:, 1]
expected = torch.arange(1, 126, dtype=torch.float64).view(5, 5, 5)[:, 1]
self.assertEqual(torch.from_numpy(x), expected)
# check zero dimensional
x = np.zeros((0, 2))
self.assertEqual(torch.from_numpy(x).shape, (0, 2))
x = np.zeros((2, 0))
self.assertEqual(torch.from_numpy(x).shape, (2, 0))
# check ill-sized strides raise exception
x = np.array([3., 5., 8.])
x.strides = (3,)
self.assertRaises(ValueError, lambda: torch.from_numpy(x))
@skipMeta
def test_from_list_of_ndarray_warning(self, device):
warning_msg = r"Creating a tensor from a list of numpy.ndarrays is extremely slow"
with self.assertWarnsOnceRegex(UserWarning, warning_msg):
torch.tensor([np.array([0]), np.array([1])], device=device)
def test_ctor_with_invalid_numpy_array_sequence(self, device):
# Invalid list of numpy array
with self.assertRaisesRegex(ValueError, "expected sequence of length"):
torch.tensor([np.random.random(size=(3, 3)), np.random.random(size=(3, 0))], device=device)
# Invalid list of list of numpy array
with self.assertRaisesRegex(ValueError, "expected sequence of length"):
torch.tensor([[np.random.random(size=(3, 3)), np.random.random(size=(3, 2))]], device=device)
with self.assertRaisesRegex(ValueError, "expected sequence of length"):
torch.tensor([[np.random.random(size=(3, 3)), np.random.random(size=(3, 3))],
[np.random.random(size=(3, 3)), np.random.random(size=(3, 2))]], device=device)
# expected shape is `[1, 2, 3]`, hence we try to iterate over 0-D array
# leading to type error : not a sequence.
with self.assertRaisesRegex(TypeError, "not a sequence"):
torch.tensor([[np.random.random(size=(3)), np.random.random()]], device=device)
# list of list or numpy array.
with self.assertRaisesRegex(ValueError, "expected sequence of length"):
torch.tensor([[1, 2, 3], np.random.random(size=(2,)), ], device=device)
@onlyCPU
def test_ctor_with_numpy_scalar_ctor(self, device) -> None:
dtypes = [
np.double,
np.float64,
np.float16,
np.int64,
np.int32,
np.int16,
np.uint8,
np.bool_,
]
for dtype in dtypes:
self.assertEqual(dtype(42), torch.tensor(dtype(42)).item())
@onlyCPU
def test_numpy_index(self, device):
i = np.array([0, 1, 2], dtype=np.int32)
x = torch.randn(5, 5)
for idx in i:
self.assertFalse(isinstance(idx, int))
self.assertEqual(x[idx], x[int(idx)])
@onlyCPU
def test_numpy_array_interface(self, device):
types = [
torch.DoubleTensor,
torch.FloatTensor,
torch.HalfTensor,
torch.LongTensor,
torch.IntTensor,
torch.ShortTensor,
torch.ByteTensor,
]
dtypes = [
np.float64,
np.float32,
np.float16,
np.int64,
np.int32,
np.int16,
np.uint8,
]
for tp, dtype in zip(types, dtypes):
# Only concrete class can be given where "Type[number[_64Bit]]" is expected
if np.dtype(dtype).kind == 'u': # type: ignore[misc]
# .type expects a XxxTensor, which have no type hints on
# purpose, so ignore during mypy type checking
x = torch.tensor([1, 2, 3, 4]).type(tp) # type: ignore[call-overload]
array = np.array([1, 2, 3, 4], dtype=dtype)
else:
x = torch.tensor([1, -2, 3, -4]).type(tp) # type: ignore[call-overload]
array = np.array([1, -2, 3, -4], dtype=dtype)
# Test __array__ w/o dtype argument
asarray = np.asarray(x)
self.assertIsInstance(asarray, np.ndarray)
self.assertEqual(asarray.dtype, dtype)
for i in range(len(x)):
self.assertEqual(asarray[i], x[i])
# Test __array_wrap__, same dtype
abs_x = np.abs(x)
abs_array = np.abs(array)
self.assertIsInstance(abs_x, tp)
for i in range(len(x)):
self.assertEqual(abs_x[i], abs_array[i])
# Test __array__ with dtype argument
for dtype in dtypes:
x = torch.IntTensor([1, -2, 3, -4])
asarray = np.asarray(x, dtype=dtype)
self.assertEqual(asarray.dtype, dtype)
# Only concrete class can be given where "Type[number[_64Bit]]" is expected
if np.dtype(dtype).kind == 'u': # type: ignore[misc]
wrapped_x = np.array([1, -2, 3, -4], dtype=dtype)
for i in range(len(x)):
self.assertEqual(asarray[i], wrapped_x[i])
else:
for i in range(len(x)):
self.assertEqual(asarray[i], x[i])
# Test some math functions with float types
float_types = [torch.DoubleTensor, torch.FloatTensor]
float_dtypes = [np.float64, np.float32]
for tp, dtype in zip(float_types, float_dtypes):
x = torch.tensor([1, 2, 3, 4]).type(tp) # type: ignore[call-overload]
array = np.array([1, 2, 3, 4], dtype=dtype)
for func in ['sin', 'sqrt', 'ceil']:
ufunc = getattr(np, func)
res_x = ufunc(x)
res_array = ufunc(array)
self.assertIsInstance(res_x, tp)
for i in range(len(x)):
self.assertEqual(res_x[i], res_array[i])
# Test functions with boolean return value
for tp, dtype in zip(types, dtypes):
x = torch.tensor([1, 2, 3, 4]).type(tp) # type: ignore[call-overload]
array = np.array([1, 2, 3, 4], dtype=dtype)
geq2_x = np.greater_equal(x, 2)
geq2_array = np.greater_equal(array, 2).astype('uint8')
self.assertIsInstance(geq2_x, torch.ByteTensor)
for i in range(len(x)):
self.assertEqual(geq2_x[i], geq2_array[i])
@onlyCPU
def test_multiplication_numpy_scalar(self, device) -> None:
for np_dtype in [np.float32, np.float64, np.int32, np.int64, np.int16, np.uint8]:
for t_dtype in [torch.float, torch.double]:
# mypy raises an error when np.floatXY(2.0) is called
# even though this is valid code
np_sc = np_dtype(2.0) # type: ignore[abstract, arg-type]
t = torch.ones(2, requires_grad=True, dtype=t_dtype)
r1 = t * np_sc
self.assertIsInstance(r1, torch.Tensor)
self.assertTrue(r1.dtype == t_dtype)
self.assertTrue(r1.requires_grad)
r2 = np_sc * t
self.assertIsInstance(r2, torch.Tensor)
self.assertTrue(r2.dtype == t_dtype)
self.assertTrue(r2.requires_grad)
@onlyCPU
def test_parse_numpy_int(self, device):
# Only concrete class can be given where "Type[number[_64Bit]]" is expected
self.assertRaisesRegex(RuntimeError, "Overflow",
lambda: torch.mean(torch.randn(1, 1), np.uint64(-1))) # type: ignore[call-overload]
# https://github.com/pytorch/pytorch/issues/29252
for nptype in [np.int16, np.int8, np.uint8, np.int32, np.int64]:
scalar = 3
np_arr = np.array([scalar], dtype=nptype)
np_val = np_arr[0]
# np integral type can be treated as a python int in native functions with
# int parameters:
self.assertEqual(torch.ones(5).diag(scalar), torch.ones(5).diag(np_val))
self.assertEqual(torch.ones([2, 2, 2, 2]).mean(scalar), torch.ones([2, 2, 2, 2]).mean(np_val))
# numpy integral type parses like a python int in custom python bindings:
self.assertEqual(torch.Storage(np_val).size(), scalar) # type: ignore[attr-defined]
tensor = torch.tensor([2], dtype=torch.int)
tensor[0] = np_val
self.assertEqual(tensor[0], np_val)
# Original reported issue, np integral type parses to the correct
# PyTorch integral type when passed for a `Scalar` parameter in
# arithmetic operations:
t = torch.from_numpy(np_arr)
self.assertEqual((t + np_val).dtype, t.dtype)
self.assertEqual((np_val + t).dtype, t.dtype)
def test_has_storage_numpy(self, device):
for dtype in [np.float32, np.float64, np.int64,
np.int32, np.int16, np.uint8]:
arr = np.array([1], dtype=dtype)
self.assertIsNotNone(torch.tensor(arr, device=device, dtype=torch.float32).storage())
self.assertIsNotNone(torch.tensor(arr, device=device, dtype=torch.double).storage())
self.assertIsNotNone(torch.tensor(arr, device=device, dtype=torch.int).storage())
self.assertIsNotNone(torch.tensor(arr, device=device, dtype=torch.long).storage())
self.assertIsNotNone(torch.tensor(arr, device=device, dtype=torch.uint8).storage())
@dtypes(*all_types_and_complex_and(torch.half, torch.bfloat16, torch.bool))
def test_numpy_scalar_cmp(self, device, dtype):
if dtype.is_complex:
tensors = (torch.tensor(complex(1, 3), dtype=dtype, device=device),
torch.tensor([complex(1, 3), 0, 2j], dtype=dtype, device=device),
torch.tensor([[complex(3, 1), 0], [-1j, 5]], dtype=dtype, device=device))
else:
tensors = (torch.tensor(3, dtype=dtype, device=device),
torch.tensor([1, 0, -3], dtype=dtype, device=device),
torch.tensor([[3, 0, -1], [3, 5, 4]], dtype=dtype, device=device))
for tensor in tensors:
if dtype == torch.bfloat16:
with self.assertRaises(TypeError):
np_array = tensor.cpu().numpy()
continue
np_array = tensor.cpu().numpy()
for t, a in product((tensor.flatten()[0], tensor.flatten()[0].item()),
(np_array.flatten()[0], np_array.flatten()[0].item())):
self.assertEqual(t, a)
if dtype == torch.complex64 and torch.is_tensor(t) and type(a) == np.complex64:
# TODO: Imaginary part is dropped in this case. Need fix.
# https://github.com/pytorch/pytorch/issues/43579
self.assertFalse(t == a)
else:
self.assertTrue(t == a)
op_db: List[OpInfo] = [
ReductionOpInfo(
"_masked.sum",
ref=reference_reduction_numpy(np.sum),
method_variant=None,
identity=0,
nan_policy="propagate",
supports_out=False,
supports_forward_ad=True,
supports_fwgrad_bwgrad=True,
supports_sparse=True,
supports_sparse_csr=True,
promotes_int_to_int64=True,
dtypes=all_types_and_complex_and(torch.bool, torch.float16, torch.bfloat16),
skips=(
DecorateInfo(
unittest.skip("Failing on some jobs"),
"TestReductions",
"test_reference_masked",
dtypes=(torch.bool, torch.int8, torch.int16, torch.int32),
),
DecorateInfo(
unittest.expectedFailure,
"TestNormalizeOperators",
"test_normalize_operator_exhaustive",
),
# FIXME: sum reduces all dimensions when dim=[]
DecorateInfo(unittest.expectedFailure, "TestReductions", "test_dim_empty"),
DecorateInfo(
class TestTorchDlPack(TestCase):
exact_dtype = True
@skipMeta
@onlyNativeDeviceTypes
@dtypes(*all_types_and_complex_and(torch.half, torch.bfloat16))
def test_dlpack_capsule_conversion(self, device, dtype):
# DLpack does not explicitly support bool (xref dmlc/dlpack#75)
x = make_tensor((5,), dtype=dtype, device=device)
z = from_dlpack(to_dlpack(x))
self.assertEqual(z, x)
@skipMeta
@onlyNativeDeviceTypes
@dtypes(*all_types_and_complex_and(torch.half, torch.bfloat16))
def test_dlpack_protocol_conversion(self, device, dtype):
x = make_tensor((5,), dtype=dtype, device=device)
z = from_dlpack(x)
self.assertEqual(z, x)
@skipMeta
@onlyNativeDeviceTypes
def test_dlpack_shared_storage(self, device):
x = make_tensor((5,), dtype=torch.float64, device=device)
z = from_dlpack(to_dlpack(x))
z[0] = z[0] + 20.0
self.assertEqual(z, x)
@skipMeta
@onlyCUDA
@dtypes(*all_types_and_complex_and(torch.half, torch.bfloat16))
def test_dlpack_conversion_with_streams(self, device, dtype):
# Create a stream where the tensor will reside
stream = torch.cuda.Stream()
with torch.cuda.stream(stream):
# Do an operation in the actual stream
x = make_tensor((5,), dtype=dtype, device=device) + 1
# DLPack protocol helps establish a correct stream order
# (hence data dependency) at the exchange boundary.
# DLPack manages this synchronization for us, so we don't need to
# explicitly wait until x is populated
stream = torch.cuda.Stream()
with torch.cuda.stream(stream):
z = from_dlpack(x)
stream.synchronize()
self.assertEqual(z, x)
@skipMeta
@onlyNativeDeviceTypes
@dtypes(*all_types_and_complex_and(torch.half, torch.bfloat16))
def test_from_dlpack(self, device, dtype):
x = make_tensor((5,), dtype=dtype, device=device)
y = torch.from_dlpack(x)
self.assertEqual(x, y)
@skipMeta
@onlyNativeDeviceTypes
@dtypes(*all_types_and_complex_and(torch.half, torch.bfloat16))
def test_from_dlpack_noncontinguous(self, device, dtype):
x = make_tensor((25,), dtype=dtype, device=device).reshape(5, 5)
y1 = x[0]
y1_dl = torch.from_dlpack(y1)
self.assertEqual(y1, y1_dl)
y2 = x[:, 0]
y2_dl = torch.from_dlpack(y2)
self.assertEqual(y2, y2_dl)
y3 = x[1, :]
y3_dl = torch.from_dlpack(y3)
self.assertEqual(y3, y3_dl)
y4 = x[1]
y4_dl = torch.from_dlpack(y4)
self.assertEqual(y4, y4_dl)
y5 = x.t()
y5_dl = torch.from_dlpack(y5)
self.assertEqual(y5, y5_dl)
@skipMeta
@onlyCUDA
@dtypes(*all_types_and_complex_and(torch.half, torch.bfloat16))
def test_dlpack_conversion_with_diff_streams(self, device, dtype):
stream_a = torch.cuda.Stream()
stream_b = torch.cuda.Stream()
# DLPack protocol helps establish a correct stream order
# (hence data dependency) at the exchange boundary.
# the `tensor.__dlpack__` method will insert a synchronization event
# in the current stream to make sure that it was correctly populated.
with torch.cuda.stream(stream_a):
x = make_tensor((5,), dtype=dtype, device=device) + 1
z = torch.from_dlpack(x.__dlpack__(stream_b.cuda_stream))
stream_a.synchronize()
stream_b.synchronize()
self.assertEqual(z, x)
@skipMeta
@onlyNativeDeviceTypes
@dtypes(*all_types_and_complex_and(torch.half, torch.bfloat16))
def test_from_dlpack_dtype(self, device, dtype):
x = make_tensor((5,), dtype=dtype, device=device)
y = torch.from_dlpack(x)
assert x.dtype == y.dtype
@skipMeta
@onlyCUDA
def test_dlpack_default_stream(self, device):
class DLPackTensor:
def __init__(self, tensor):
self.tensor = tensor
def __dlpack_device__(self):
return self.tensor.__dlpack_device__()
def __dlpack__(self, stream=None):
if torch.version.hip is None:
assert stream == 1
else:
assert stream == 0
capsule = self.tensor.__dlpack__(stream)
return capsule
# CUDA-based tests runs on non-default streams
with torch.cuda.stream(torch.cuda.default_stream()):
x = DLPackTensor(make_tensor((5,), dtype=torch.float32, device=device))
from_dlpack(x)
@skipMeta
@onlyNativeDeviceTypes
@dtypes(*all_types_and_complex_and(torch.half, torch.bfloat16))
def test_dlpack_tensor_invalid_stream(self, device, dtype):
with self.assertRaises(TypeError):
x = make_tensor((5,), dtype=dtype, device=device)
x.__dlpack__(stream=object())
@skipMeta
def test_dlpack_error_on_bool_tensor(self):
x = torch.tensor([True], dtype=torch.bool)
with self.assertRaises(RuntimeError):
to_dlpack(x)
# TODO: add interchange tests once NumPy 1.22 (dlpack support) is required
@skipMeta
def test_dlpack_export_requires_grad(self):
x = torch.zeros(10, dtype=torch.float32, requires_grad=True)
with self.assertRaisesRegex(RuntimeError, r"require gradient"):
x.__dlpack__()
@skipMeta
def test_dlpack_export_is_conj(self):
x = torch.tensor([-1 + 1j, -2 + 2j, 3 - 3j])
y = torch.conj(x)
with self.assertRaisesRegex(RuntimeError, r"conjugate bit"):
y.__dlpack__()
@skipMeta
def test_dlpack_export_non_strided(self):
x = torch.sparse_coo_tensor([[0]], [1], size=(1,))
y = torch.conj(x)
with self.assertRaisesRegex(RuntimeError, r"strided"):
y.__dlpack__()
@skipMeta
def test_dlpack_normalize_strides(self):
x = torch.rand(16)
y = x[::3][:1]
self.assertEqual(y.shape, (1,))
self.assertEqual(y.stride(), (3,))
z = from_dlpack(y)
self.assertEqual(z.shape, (1,))
# gh-83069, make sure __dlpack__ normalizes strides
self.assertEqual(z.stride(), (1,))
yield SampleInput(mt((9, 10)))
yield SampleInput(mt((50, )), kwargs=dict(dim=0))
yield SampleInput(mt((5, 11)), kwargs=dict(dim=(1, )))
yield SampleInput(mt((5, 6)), kwargs=dict(dim=(0, 1)))
yield SampleInput(mt((5, 6, 2)), kwargs=dict(dim=(0, 2)))
# Operator database
op_db: List[OpInfo] = [
SpectralFuncInfo(
"fft.fft",
aten_name="fft_fft",
decomp_aten_name="_fft_c2c",
ref=np.fft.fft,
ndimensional=SpectralFuncType.OneD,
dtypes=all_types_and_complex_and(torch.bool),
# rocFFT doesn't support Half/Complex Half Precision FFT
# CUDA supports Half/ComplexHalf Precision FFT only on SM53 or later archs
dtypesIfCUDA=all_types_and_complex_and(
torch.bool,
*(() if (TEST_WITH_ROCM or not SM53OrLater) else
(torch.half, torch.complex32)),
),
# https://github.com/pytorch/pytorch/issues/80411
gradcheck_fast_mode=True,
supports_forward_ad=True,
supports_fwgrad_bwgrad=True,
# See https://github.com/pytorch/pytorch/pull/78358
check_batched_forward_grad=False,
),
SpectralFuncInfo(
def test_dtypes(self, device, op):
# Check complex32 support only if the op claims.
# TODO: Once the complex32 support is better, we should add check for complex32 unconditionally.
device_type = torch.device(device).type
include_complex32 = ((torch.complex32, ) if op.supports_dtype(
torch.complex32, device_type) else ())
# dtypes to try to backward in
allowed_backward_dtypes = floating_and_complex_types_and(
*((torch.half, torch.bfloat16) + include_complex32))
# lists for (un)supported dtypes
supported_dtypes = set()
unsupported_dtypes = set()
supported_backward_dtypes = set()
unsupported_backward_dtypes = set()
def unsupported(dtype):
unsupported_dtypes.add(dtype)
if dtype in allowed_backward_dtypes:
unsupported_backward_dtypes.add(dtype)
for dtype in all_types_and_complex_and(
*((torch.half, torch.bfloat16, torch.bool) +
include_complex32)):
# tries to acquire samples - failure indicates lack of support
requires_grad = (dtype in allowed_backward_dtypes)
try:
samples = tuple(
op.sample_inputs(device,
dtype,
requires_grad=requires_grad))
except Exception as e:
unsupported(dtype)
continue
for sample in samples:
# tries to call operator with the sample - failure indicates
# lack of support
try:
result = op(sample.input, *sample.args, **sample.kwargs)
supported_dtypes.add(dtype)
except Exception as e:
# NOTE: some ops will fail in forward if their inputs
# require grad but they don't support computing the gradient
# in that type! This is a bug in the op!
unsupported(dtype)
continue
# Checks for backward support in the same dtype
try:
result = sample.output_process_fn_grad(result)
if isinstance(result, torch.Tensor):
backward_tensor = result
elif isinstance(result, Sequence) and isinstance(
result[0], torch.Tensor):
backward_tensor = result[0]
else:
continue
# Note: this grad may not have the same dtype as dtype
# For functions like complex (float -> complex) or abs
# (complex -> float) the grad tensor will have a
# different dtype than the input.
# For simplicity, this is still modeled as these ops
# supporting grad in the input dtype.
grad = torch.randn_like(backward_tensor)
backward_tensor.backward(grad)
supported_backward_dtypes.add(dtype)
except Exception as e:
unsupported_backward_dtypes.add(dtype)
# Checks that dtypes are listed correctly and generates an informative
# error message
supported_forward = supported_dtypes - unsupported_dtypes
partially_supported_forward = supported_dtypes & unsupported_dtypes
unsupported_forward = unsupported_dtypes - supported_dtypes
supported_backward = supported_backward_dtypes - unsupported_backward_dtypes
partially_supported_backward = supported_backward_dtypes & unsupported_backward_dtypes
unsupported_backward = unsupported_backward_dtypes - supported_backward_dtypes
device_type = torch.device(device).type
claimed_forward = set(op.supported_dtypes(device_type))
supported_but_unclaimed_forward = supported_forward - claimed_forward
claimed_but_unsupported_forward = claimed_forward & unsupported_forward
claimed_backward = set(op.supported_backward_dtypes(device_type))
supported_but_unclaimed_backward = supported_backward - claimed_backward
claimed_but_unsupported_backward = claimed_backward & unsupported_backward
# Partially supporting a dtype is not an error, but we print a warning
if (len(partially_supported_forward) +
len(partially_supported_backward)) > 0:
msg = "Some dtypes for {0} on device type {1} are only partially supported!\n".format(
op.name, device_type)
if len(partially_supported_forward) > 0:
msg = msg + "The following dtypes only worked on some samples during forward: {0}.\n".format(
partially_supported_forward)
if len(partially_supported_backward) > 0:
msg = msg + "The following dtypes only worked on some samples during backward: {0}.\n".format(
partially_supported_backward)
print(msg)
if (len(supported_but_unclaimed_forward) +
len(claimed_but_unsupported_forward) +
len(supported_but_unclaimed_backward) +
len(claimed_but_unsupported_backward)) == 0:
return
# Generates error msg
msg = "The supported dtypes for {0} on device type {1} are incorrect!\n".format(
op.name, device_type)
if len(supported_but_unclaimed_forward) > 0:
msg = msg + "The following dtypes worked in forward but are not listed by the OpInfo: {0}.\n".format(
supported_but_unclaimed_forward)
if len(supported_but_unclaimed_backward) > 0:
msg = msg + "The following dtypes worked in backward but are not listed by the OpInfo: {0}.\n".format(
supported_but_unclaimed_backward)
if len(claimed_but_unsupported_forward) > 0:
msg = msg + "The following dtypes did not work in forward but are listed by the OpInfo: {0}.\n".format(
claimed_but_unsupported_forward)
if len(claimed_but_unsupported_backward) > 0:
msg = msg + "The following dtypes did not work in backward but are listed by the OpInfo: {0}.\n".format(
claimed_but_unsupported_backward)
self.fail(msg)