def __array_finalize__(self, obj):
if obj is None:
# constructor called previously
return
else:
if self.base is None:
# case np.array
raise RuntimeError(
"Copying storages is only possible through Storage.copy() or deepcopy."
)
else:
if not isinstance(obj, Storage) and not isinstance(
obj, _ViewableNdarray):
raise RuntimeError(
"Meta information can not be inferred when creating Storage views from other classes than Storage."
)
self.__dict__ = {**obj.__dict__, **self.__dict__}
self.is_stencil_view = False
if hasattr(obj, "_new_index"):
index_iter = itertools.chain(
obj._new_index, [slice(None, None)] *
(len(obj.mask) - len(obj._new_index)))
interpolated_mask = gt_utils.interpolate_mask(
(isinstance(x, slice) for x in index_iter), obj.mask,
False)
self._mask = tuple(
x & y for x, y in zip(obj.mask, interpolated_mask))
delattr(obj, "_new_index")
if not hasattr(obj, "default_origin"):
self.is_stencil_view = True
elif self._is_consistent(obj):
self.is_stencil_view = obj.is_stencil_view
self._finalize_view(obj)
def from_mask(cls, seq, mask, default=None):
if default is None:
default = cls._DEFAULT
return cls(gt_utils.interpolate_mask(seq, mask, default))
class StencilObject(abc.ABC):
"""Generic singleton implementation of a stencil callable.
This class is used as base class for specific subclass generated
at run-time for any stencil definition and a unique set of external symbols.
Instances of this class do not contain state and thus it is
implemented as a singleton: only one instance per subclass is actually
allocated (and it is immutable).
The callable interface is the same of the stencil definition function,
with some extra keyword arguments.
Keyword Arguments
------------------
domain : `Sequence` of `int`, optional
Shape of the computation domain. If `None`, it will be used the
largest feasible domain according to the provided input fields
and origin values (`None` by default).
origin : `[int * ndims]` or `{'field_name': [int * ndims]}`, optional
If a single offset is passed, it will be used for all fields.
If a `dict` is passed, there could be an entry for each field.
A special key *'_all_'* will represent the value to be used for all
the fields not explicitly defined. If `None` is passed or it is
not possible to assign a value to some field according to the
previous rule, the value will be inferred from the `boundary` attribute
of the `field_info` dict. Note that the function checks if the origin values
are at least equal to the `boundary` attribute of that field,
so a 0-based origin will only be acceptable for fields with
a 0-area support region.
exec_info : `dict`, optional
Dictionary used to store information about the stencil execution.
(`None` by default). If the dictionary contains the magic key
'__aggregate_data' and it evaluates to `True`, the dictionary is
populated with a nested dictionary per class containing different
performance statistics. These include the stencil calls count, the
cumulative time spent in all stencil calls, and the actual time spent
in carrying out the computations.
"""
# Those attributes are added to the class at loading time:
_gt_id_: str
definition_func: Callable[..., Any]
_domain_origin_cache: ClassVar[Dict[int, Tuple[Tuple[int, ...],
Dict[str, Tuple[int,
...]]]]]
"""Stores domain/origin pairs that have been used by hash."""
def __new__(cls, *args, **kwargs):
if getattr(cls, "_instance", None) is None:
cls._instance = object.__new__(cls)
cls._domain_origin_cache = {}
return cls._instance
def __setattr__(self, key, value) -> None:
raise AttributeError(
"Attempting a modification of an attribute in a frozen class")
def __delattr__(self, item) -> None:
raise AttributeError(
"Attempting a deletion of an attribute in a frozen class")
def __eq__(self, other) -> bool:
return type(self) == type(other)
def __str__(self) -> str:
result = """
<StencilObject: {name}> [backend="{backend}"]
- I/O fields: {fields}
- Parameters: {params}
- Constants: {constants}
- Version: {version}
- Definition ({func}):
{source}
""".format(
name=self.options["module"] + "." + self.options["name"],
version=self._gt_id_,
backend=self.backend,
fields=self.field_info,
params=self.parameter_info,
constants=self.constants,
func=self.definition_func,
source=self.source,
)
return result
def __hash__(self) -> int:
return int.from_bytes(type(self)._gt_id_.encode(), byteorder="little")
@property
@abc.abstractmethod
def backend(self) -> str:
pass
@property
@abc.abstractmethod
def source(self) -> str:
pass
@property
@abc.abstractmethod
def domain_info(self) -> DomainInfo:
pass
@property
@abc.abstractmethod
def field_info(self) -> Dict[str, FieldInfo]:
pass
@property
@abc.abstractmethod
def parameter_info(self) -> Dict[str, ParameterInfo]:
pass
@property
@abc.abstractmethod
def constants(self) -> Dict[str, Any]:
pass
@property
@abc.abstractmethod
def options(self) -> Dict[str, Any]:
pass
@abc.abstractmethod
def run(self, *args, **kwargs) -> None:
pass
@abc.abstractmethod
def __call__(self, *args, **kwargs) -> None:
pass
@staticmethod
def _make_origin_dict(origin: Any) -> Dict[str, Index]:
try:
if isinstance(origin, dict):
return dict(origin)
if origin is None:
return {}
if isinstance(origin, collections.abc.Iterable):
return {"_all_": Index.from_value(origin)}
if isinstance(origin, int):
return {"_all_": Index.from_k(origin)}
except Exception:
pass
raise ValueError("Invalid 'origin' value ({})".format(origin))
def _get_max_domain(
self,
field_args: Dict[str, Any],
origin: Dict[str, Tuple[int, ...]],
*,
squeeze: bool = True,
) -> Shape:
"""Return the maximum domain size possible.
Parameters
----------
field_args:
Mapping from field names to actually passed data arrays.
origin:
The origin for each field.
squeeze:
Convert non-used domain dimensions to singleton dimensions.
Returns
-------
`Shape`: the maximum domain size.
"""
domain_ndim = self.domain_info.ndim
max_size = sys.maxsize
max_domain = Shape([max_size] * domain_ndim)
for name, field_info in self.field_info.items():
if field_info.access != AccessKind.NONE:
assert field_args.get(
name,
None) is not None, f"Invalid value for '{name}' field."
field = field_args[name]
api_domain_mask = field_info.domain_mask
api_domain_ndim = field_info.domain_ndim
upper_indices = field_info.boundary.upper_indices.filter_mask(
api_domain_mask)
field_origin = Index.from_value(origin[name])
field_domain = tuple(field.shape[i] -
(field_origin[i] + upper_indices[i])
for i in range(api_domain_ndim))
max_domain &= Shape.from_mask(field_domain,
api_domain_mask,
default=max_size)
if squeeze:
return Shape([i if i != max_size else 1 for i in max_domain])
else:
return max_domain
def _validate_args( # noqa: C901 # Function is too complex
self,
field_args: Dict[str, FieldType],
param_args: Dict[str, Any],
domain: Tuple[int, ...],
origin: Dict[str, Tuple[int, ...]],
) -> None:
"""
Validate input arguments to _call_run.
Raises
-------
ValueError
If invalid data or inconsistent options are specified.
TypeError
If an incorrect field or parameter data type is passed.
"""
assert isinstance(field_args, dict) and isinstance(param_args, dict)
# validate domain sizes
domain_ndim = self.domain_info.ndim
if len(domain) != domain_ndim:
raise ValueError(f"Invalid 'domain' value '{domain}'")
try:
domain = Shape(domain)
except Exception:
raise ValueError("Invalid 'domain' value ({})".format(domain))
if not domain > Shape.zeros(domain_ndim):
raise ValueError(f"Compute domain contains zero sizes '{domain}')")
if not domain <= (max_domain := self._get_max_domain(
field_args, origin, squeeze=False)):
raise ValueError(
f"Compute domain too large (provided: {domain}, maximum: {max_domain})"
)
if domain[2] < self.domain_info.min_sequential_axis_size:
raise ValueError(
f"Compute domain too small. Sequential axis is {domain[2]}, but must be at least {self.domain_info.min_sequential_axis_size}."
)
# assert compatibility of fields with stencil
for name, field_info in self.field_info.items():
if field_info.access != AccessKind.NONE:
if name not in field_args:
raise ValueError(f"Missing value for '{name}' field.")
field = field_args[name]
if not gt_backend.from_name(
self.backend).storage_info["is_compatible_layout"](
field):
raise ValueError(
f"The layout of the field {name} is not compatible with the backend."
)
if not gt_backend.from_name(
self.backend).storage_info["is_compatible_type"](
field):
raise ValueError(
f"Field '{name}' has type '{type(field)}', which is not compatible with the '{self.backend}' backend."
)
elif type(field) is np.ndarray:
warnings.warn(
"NumPy ndarray passed as field. This is discouraged and only works with constraints and only for certain backends.",
RuntimeWarning,
)
field_dtype = self.field_info[name].dtype
if not field.dtype == field_dtype:
raise TypeError(
f"The dtype of field '{name}' is '{field.dtype}' instead of '{field_dtype}'"
)
if isinstance(field, gt_storage.storage.Storage
) and not field.is_stencil_view:
raise ValueError(
f"An incompatible view was passed for field {name} to the stencil. "
)
# Check: domain + halo vs field size
field_info = self.field_info[name]
field_domain_mask = field_info.domain_mask
field_domain_ndim = field_info.domain_ndim
field_domain_origin = Index.from_mask(
origin[name], field_domain_mask[:domain_ndim])
if field.ndim != field_domain_ndim + len(field_info.data_dims):
raise ValueError(
f"Storage for '{name}' has {field.ndim} dimensions but the API signature "
f"expects {field_domain_ndim + len(field_info.data_dims)} ('{field_info.axes}[{field_info.data_dims}]')"
)
if (isinstance(field, gt_storage.storage.Storage) and
tuple(field.mask)[:domain_ndim] != field_domain_mask):
raise ValueError(
f"Storage for '{name}' has domain mask '{field.mask}' but the API signature "
f"expects '[{', '.join(field_info.axes)}]'")
# Check: data dimensions shape
if field.shape[field_domain_ndim:] != field_info.data_dims:
raise ValueError(
f"Field '{name}' expects data dimensions {field_info.data_dims} but got {field.shape[field_domain_ndim:]}"
)
min_origin = gt_utils.interpolate_mask(
field_info.boundary.lower_indices.filter_mask(
field_domain_mask),
field_domain_mask,
default=0,
)
if field_domain_origin < min_origin:
raise ValueError(
f"Origin for field {name} too small. Must be at least {min_origin}, is {field_domain_origin}"
)
spatial_domain = typing.cast(
Shape, domain).filter_mask(field_domain_mask)
lower_indices = field_info.boundary.lower_indices.filter_mask(
field_domain_mask)
upper_indices = field_info.boundary.upper_indices.filter_mask(
field_domain_mask)
min_shape = tuple(lb + d + ub for lb, d, ub in zip(
lower_indices, spatial_domain, upper_indices))
if min_shape > field.shape:
raise ValueError(
f"Shape of field {name} is {field.shape} but must be at least {min_shape} for given domain and origin."
)
# assert compatibility of parameters with stencil
for name, parameter_info in self.parameter_info.items():
if parameter_info.access != AccessKind.NONE:
if name not in param_args:
raise ValueError(f"Missing value for '{name}' parameter.")
elif type(
parameter := param_args[name]) != parameter_info.dtype:
raise TypeError(
f"The type of parameter '{name}' is '{type(parameter)}' instead of '{parameter_info.dtype}'"
)
def _run_test_implementation(cls, parameters_dict, implementation): # noqa: C901 # too complex
input_data, exec_info = parameters_dict
origin = cls.origin
max_boundary = Boundary(cls.max_boundary)
field_params = cls.field_params
field_masks = {}
for name, value in input_data.items():
if isinstance(value, np.ndarray):
field_masks[name] = tuple(
ax in field_params[name][0] for ax in CartesianSpace.names
)
data_shape = Shape((sys.maxsize,) * 3)
for name, data in input_data.items():
if isinstance(data, np.ndarray):
data_shape &= Shape(
interpolate_mask(data.shape, field_masks[name], default=sys.maxsize)
)
domain = data_shape - (
Index(max_boundary.lower_indices) + Index(max_boundary.upper_indices)
)
referenced_inputs = {
name: info for name, info in implementation.field_info.items() if info is not None
}
referenced_inputs.update(
{name: info for name, info in implementation.parameter_info.items() if info is not None}
)
# set externals for validation method
for k, v in implementation.constants.items():
sys.modules[cls.__module__].__dict__[k] = v
# copy input data
test_values = {}
validation_values = {}
for name, data in input_data.items():
data = input_data[name]
if name in referenced_inputs:
info = referenced_inputs[name]
if isinstance(info, FieldInfo):
data_dims = field_params[name][1]
if data_dims:
dtype = (data.dtype, data_dims)
shape = data.shape[: -len(data_dims)]
else:
dtype = data.dtype
shape = data.shape
test_values[name] = gt_storage.from_array(
data,
dtype=dtype,
shape=shape,
mask=field_masks[name],
default_origin=origin,
backend=implementation.backend,
)
validation_values[name] = np.array(data)
else:
test_values[name] = data
validation_values[name] = data
else:
test_values[name] = None
validation_values[name] = None
# call implementation
implementation(**test_values, origin=origin, exec_info=exec_info)
assert domain == exec_info["domain"]
# for validation data, data is cropped to actually touched domain, so that origin offseting
# does not have to be implemented for every test suite. This is done based on info
# specified in test suite
cropped_validation_values = {}
for name, data in validation_values.items():
sym = cls.symbols[name]
if data is not None and sym.kind == SymbolKind.FIELD:
field_extent_low = tuple(b[0] for b in sym.boundary)
offset_low = tuple(b[0] - e for b, e in zip(max_boundary, field_extent_low))
field_extent_high = tuple(b[1] for b in sym.boundary)
offset_high = tuple(b[1] - e for b, e in zip(max_boundary, field_extent_high))
validation_slice = filter_mask(
tuple(slice(o, s - h) for o, s, h in zip(offset_low, data_shape, offset_high)),
field_masks[name],
)
data_dims = field_params[name][1]
if data_dims:
validation_slice = tuple([*validation_slice] + [slice(None)] * len(data_dims))
cropped_validation_values[name] = data[validation_slice]
else:
cropped_validation_values[name] = data
cls.validation(
**cropped_validation_values,
domain=domain,
origin={
name: info.boundary.lower_indices
for name, info in implementation.field_info.items()
if info is not None
},
)
# Test values
for name, value in test_values.items():
if isinstance(value, np.ndarray):
expected_value = validation_values[name]
if gt_backend.from_name(value.backend).storage_info["device"] == "gpu":
value.synchronize()
value = value.data.get()
else:
value = value.data
np.testing.assert_allclose(
value,
expected_value,
rtol=RTOL,
atol=ATOL,
equal_nan=EQUAL_NAN,
err_msg="Wrong data in output field '{name}'".format(name=name),
)
class StencilObject(abc.ABC):
"""Generic singleton implementation of a stencil function.
This class is used as base class for the specific subclass generated
at run-time for any stencil definition and a unique set of external symbols.
Instances of this class do not contain any information and thus it is
implemented as a singleton: only one instance per subclass is actually
allocated (and it is immutable).
"""
def __new__(cls, *args, **kwargs):
if getattr(cls, "_instance", None) is None:
cls._instance = object.__new__(cls)
return cls._instance
def __setattr__(self, key, value) -> None:
raise AttributeError("Attempting a modification of an attribute in a frozen class")
def __delattr__(self, item) -> None:
raise AttributeError("Attempting a deletion of an attribute in a frozen class")
def __eq__(self, other) -> bool:
return type(self) == type(other)
def __str__(self) -> str:
result = """
<StencilObject: {name}> [backend="{backend}"]
- I/O fields: {fields}
- Parameters: {params}
- Constants: {constants}
- Definition ({func}):
{source}
""".format(
name=self.options["module"] + "." + self.options["name"],
version=self._gt_id_,
backend=self.backend,
fields=self.field_info,
params=self.parameter_info,
constants=self.constants,
func=self.definition_func,
source=self.source,
)
return result
def __hash__(self) -> int:
return int.from_bytes(type(self)._gt_id_.encode(), byteorder="little")
# Those attributes are added to the class at loading time:
#
# _gt_id_ (stencil_id.version)
# definition_func
@property
@abc.abstractmethod
def backend(self) -> str:
pass
@property
@abc.abstractmethod
def source(self) -> str:
pass
@property
@abc.abstractmethod
def domain_info(self) -> DomainInfo:
pass
@property
@abc.abstractmethod
def field_info(self) -> Dict[str, FieldInfo]:
pass
@property
@abc.abstractmethod
def parameter_info(self) -> Dict[str, ParameterInfo]:
pass
@property
@abc.abstractmethod
def constants(self) -> Dict[str, Any]:
pass
@property
@abc.abstractmethod
def options(self) -> Dict[str, Any]:
pass
@abc.abstractmethod
def run(self, *args, **kwargs) -> None:
pass
@abc.abstractmethod
def __call__(self, *args, **kwargs) -> None:
pass
@staticmethod
def _make_origin_dict(origin: Any) -> Dict[str, Index]:
try:
if isinstance(origin, dict):
return origin
if origin is None:
return {}
if isinstance(origin, collections.abc.Iterable):
return {"_all_": Index.from_value(origin)}
if isinstance(origin, int):
return {"_all_": Index.from_k(origin)}
except:
pass
raise ValueError("Invalid 'origin' value ({})".format(origin))
def _get_max_domain(
self,
field_args: Dict[str, Any],
origin: Dict[str, Tuple[int, ...]],
*,
squeeze: bool = True,
) -> Shape:
"""Return the maximum domain size possible
Parameters
----------
field_args:
Mapping from field names to actually passed data arrays.
origin:
The origin for each field.
squeeze:
Convert non-used domain dimensions to singleton dimensions.
Returns
-------
`Shape`: the maximum domain size.
"""
domain_ndim = self.domain_info.ndim
max_size = sys.maxsize
max_domain = Shape([max_size] * domain_ndim)
for name, field_info in self.field_info.items():
if field_info is not None:
assert field_args.get(name, None) is not None, f"Invalid value for '{name}' field."
field = field_args[name]
api_domain_mask = field_info.domain_mask
api_domain_ndim = field_info.domain_ndim
assert (
not isinstance(field, gt_storage.storage.Storage)
or tuple(field.mask)[:domain_ndim] == api_domain_mask
), (
f"Storage for '{name}' has domain mask '{field.mask}' but the API signature "
f"expects '[{', '.join(field_info.axes)}]'"
)
upper_indices = field_info.boundary.upper_indices.filter_mask(api_domain_mask)
field_origin = Index.from_value(origin[name])
field_domain = tuple(
field.shape[i] - (field_origin[i] + upper_indices[i])
for i in range(api_domain_ndim)
)
max_domain &= Shape.from_mask(field_domain, api_domain_mask, default=max_size)
if squeeze:
return Shape([i if i != max_size else 1 for i in max_domain])
else:
return max_domain
def _validate_args(self, field_args, param_args, domain, origin) -> None:
"""Validate input arguments to _call_run.
Raises
-------
ValueError
If invalid data or inconsistent options are specified.
TypeError
If an incorrect field or parameter data type is passed.
"""
assert isinstance(field_args, dict) and isinstance(param_args, dict)
# validate domain sizes
domain_ndim = self.domain_info.ndim
if len(domain) != domain_ndim:
raise ValueError(f"Invalid 'domain' value '{domain}'")
try:
domain = Shape(domain)
except:
raise ValueError("Invalid 'domain' value ({})".format(domain))
if not domain > Shape.zeros(domain_ndim):
raise ValueError(f"Compute domain contains zero sizes '{domain}')")
if not domain <= (max_domain := self._get_max_domain(field_args, origin, squeeze=False)):
raise ValueError(
f"Compute domain too large (provided: {domain}, maximum: {max_domain})"
)
# assert compatibility of fields with stencil
for name, field_info in self.field_info.items():
if field_info is not None:
if name not in field_args:
raise ValueError(f"Missing value for '{name}' field.")
field = field_args[name]
if not gt_backend.from_name(self.backend).storage_info["is_compatible_layout"](
field
):
raise ValueError(
f"The layout of the field {name} is not compatible with the backend."
)
if not gt_backend.from_name(self.backend).storage_info["is_compatible_type"](field):
raise ValueError(
f"Field '{name}' has type '{type(field)}', which is not compatible with the '{self.backend}' backend."
)
elif type(field) is np.ndarray:
warnings.warn(
"NumPy ndarray passed as field. This is discouraged and only works with constraints and only for certain backends.",
RuntimeWarning,
)
field_dtype = self.field_info[name].dtype
if not field.dtype == field_dtype:
raise TypeError(
f"The dtype of field '{name}' is '{field.dtype}' instead of '{field_dtype}'"
)
if isinstance(field, gt_storage.storage.Storage) and not field.is_stencil_view:
raise ValueError(
f"An incompatible view was passed for field {name} to the stencil. "
)
# Check: domain + halo vs field size
field_info = self.field_info[name]
field_domain_mask = field_info.domain_mask
field_domain_ndim = field_info.domain_ndim
field_domain_origin = Index.from_mask(origin[name], field_domain_mask[:domain_ndim])
if field.ndim != field_domain_ndim + len(field_info.data_dims):
raise ValueError(
f"Storage for '{name}' has {field.ndim} dimensions but the API signature "
f"expects {field_domain_ndim + len(field_info.data_dims)} ('{field_info.axes}[{field_info.data_dims}]')"
)
min_origin = gt_utils.interpolate_mask(
field_info.boundary.lower_indices.filter_mask(field_domain_mask),
field_domain_mask,
default=0,
)
if field_domain_origin < min_origin:
raise ValueError(
f"Origin for field {name} too small. Must be at least {min_origin}, is {field_domain_origin}"
)
spatial_domain = domain.filter_mask(field_domain_mask)
upper_indices = field_info.boundary.upper_indices.filter_mask(field_domain_mask)
min_shape = tuple(
o + d + h for o, d, h in zip(field_domain_origin, spatial_domain, upper_indices)
)
if min_shape > field.shape:
raise ValueError(
f"Shape of field {name} is {field.shape} but must be at least {min_shape} for given domain and origin."
)
# assert compatibility of parameters with stencil
for name, parameter_info in self.parameter_info.items():
if parameter_info is not None:
if name not in param_args:
raise ValueError(f"Missing value for '{name}' parameter.")
if not type(parameter := param_args[name]) == self.parameter_info[name].dtype:
raise TypeError(
f"The type of parameter '{name}' is '{type(parameter)}' instead of '{self.parameter_info[name].dtype}'"
)
