# Copyright (c) 2016 The University of Manchester
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import annotations
import logging
from typing import (
Collection, Generic, Optional, Tuple, TypeVar, Union, cast, TYPE_CHECKING)
import numpy
from spinn_utilities.abstract_base import AbstractBase, abstractmethod
from spinn_utilities.ordered_set import OrderedSet
from spinn_utilities.log import FormatAdapter
from pacman.exceptions import (
PacmanConfigurationException, PacmanInvalidParameterException)
from pacman.model.graphs import AbstractVertex
if TYPE_CHECKING:
from pacman.model.partitioner_splitters import AbstractSplitterCommon
from pacman.model.graphs.machine import MachineVertex
from pacman.model.routing_info import BaseKeyAndMask
from .application_edge import ApplicationEdge
from .application_edge_partition import ApplicationEdgePartition
#: :meta private:
MV = TypeVar("MV", bound='MachineVertex')
logger = FormatAdapter(logging.getLogger(__file__))
[docs]
class ApplicationVertex(AbstractVertex, Generic[MV], metaclass=AbstractBase):
"""
A vertex that can be broken down into a number of smaller vertices
based on the resources that the vertex requires.
"""
__slots__ = (
# List of machine vertices associated with this app vertex
"_machine_vertices",
# The splitter object associated with this app vertex
"_splitter",
# The maximum number of atoms for each dimension for each core.
# For example, a 2D vertex might have a shape of 640 by 480 with
# rectangles on each core or 32 by 16.
# Typically all but possibly the last core will have this number. If
# the vertex has multiple dimensions, one or more of the dimensions
# might have fewer atoms on the last core (e.g. the rectangle on the
# last core of a 2D vertex might be smaller).
"_max_atoms_per_dimension_per_core")
def __init__(
self, label: Optional[str] = None,
max_atoms_per_core: Optional[Union[int, Tuple[int, ...]]] = None,
splitter: Optional[AbstractSplitterCommon] = None):
"""
:param str label: The optional name of the vertex.
:param max_atoms_per_core: The max number of atoms that can be
placed on a core for each dimension, used in partitioning.
If the vertex is n-dimensional, with n > 1, the value must be a
tuple with a value for each dimension. If it is single-dimensional
the value can be a 1-tuple or an int.
:type max_atoms_per_core: None or int or tuple(int,...)
:param splitter: The splitter object needed for this vertex.
Leave as `None` to delegate the choice of splitter to the selector.
:type splitter: None or
~pacman.model.partitioner_splitters.AbstractSplitterCommon
"""
# Need to set to None temporarily as add_constraint checks splitter
self._splitter: Optional[AbstractSplitterCommon] = None
super().__init__(label)
self._machine_vertices: OrderedSet[MV] = OrderedSet()
if splitter:
# Use setter as there is extra work to do
self.splitter = splitter
# Keep the name for simplicity but move to new internal representation
self._max_atoms_per_dimension_per_core: Optional[Tuple[int, ...]]
self._set_max_atoms_per_dimension_per_core(max_atoms_per_core)
def __str__(self) -> str:
return str(self.label)
def __repr__(self) -> str:
if self.get_fixed_location():
return (f"ApplicationVertex({self.label},"
f" at{self.get_fixed_location()})")
else:
return f"ApplicationVertex({self.label})"
@property
def has_splitter(self) -> bool:
"""
Whether this vertex currently has a splitter defined.
"""
return self._splitter is not None
@property
def splitter(self) -> AbstractSplitterCommon:
"""
:rtype: ~pacman.model.partitioner_splitters.AbstractSplitterCommon
"""
s = self._splitter
if s is None:
raise PacmanConfigurationException(
f"The splitter object on {self._label} has not yet had "
"a splitter set.")
return s
@splitter.setter
def splitter(self, new_value: AbstractSplitterCommon) -> None:
"""
Sets the splitter object. Does not allow repeated settings.
:param new_value: The new splitter object
:type new_value:
~pacman.model.partitioner_splitters.AbstractSplitterCommon
"""
if self._splitter == new_value:
return
if self._splitter is not None:
raise PacmanConfigurationException(
f"The splitter object on {self._label} has already been set, "
"it cannot be reset. Please fix and try again.")
self._splitter = new_value
self._splitter.set_governed_app_vertex(self)
[docs]
def remember_machine_vertex(self, machine_vertex: MV) -> None:
"""
Adds the machine vertex to the iterable returned by machine_vertices
:param ~pacman.model.graphs.machine.MachineVertex machine_vertex:
A pointer to a machine_vertex
"""
machine_vertex.index = len(self._machine_vertices)
self._machine_vertices.add(machine_vertex)
@property
def atoms_shape(self) -> Tuple[int, ...]:
"""
The "shape" of the atoms in the vertex i.e. how the atoms are split
between the dimensions of the vertex. By default everything is
1-dimensional, so the value will be a 1-tuple but can be
overridden by a vertex that supports multiple dimensions.
:rtype: tuple(int, ...)
"""
return (self.n_atoms,)
@property
@abstractmethod
def n_atoms(self) -> int:
"""
The number of atoms in the vertex.
:rtype: int
"""
raise NotImplementedError
[docs]
def round_n_atoms(
self, n_atoms: Union[int, float], label: str = "n_atoms") -> int:
"""
Utility function to allow superclasses to make sure `n_atoms` is an
integer.
:param n_atoms: Value convertible to int to be used for `n_atoms`
:type n_atoms: int or float or numpy.
:return: Number of atoms.
:rtype: int
:raises PacmanInvalidParameterException:
If the value cannot be safely converted to an integer
"""
if isinstance(n_atoms, int):
return n_atoms
# Allow a float which has a near int value
temp = int(round(n_atoms))
if abs(temp - n_atoms) < 0.001:
if temp != n_atoms:
logger.warning(
"Size of the {} rounded from {} to {}. "
"Please use int values for n_atoms",
label, n_atoms, temp)
return temp
raise PacmanInvalidParameterException(
label, n_atoms, f"int value expected for {label}")
@property
def machine_vertices(self) -> Collection[MV]:
"""
The machine vertices that this application vertex maps to.
:rtype: iterable(~pacman.model.graphs.machine.MachineVertex)
"""
return self._machine_vertices
def __check_atoms_per_core(self) -> None:
assert self._max_atoms_per_dimension_per_core is not None
if (len(self._max_atoms_per_dimension_per_core) !=
len(self.atoms_shape)):
raise ValueError(
f"On application vertex {self}, the number of dimensions in"
" the maximum number of atoms per core "
f" {self._max_atoms_per_dimension_per_core} must be the same"
" number of dimensions as that of the vertex shape"
f" {self.atoms_shape}")
if len(self.atoms_shape) > 1:
for maxi, dim in zip(self._max_atoms_per_dimension_per_core,
self.atoms_shape):
if (dim / maxi) != (dim // maxi):
raise ValueError(
f"On application vertex {self} each dimension of the"
f" vertex shape {self.atoms_shape} must be divisible"
" by the maximum number of atoms per core in that"
f" dimension {self._max_atoms_per_dimension_per_core}")
[docs]
def get_max_atoms_per_core(self) -> int:
"""
Gets the maximum number of atoms per core, which is either the
number of atoms required across the whole application vertex,
or a lower value set.
:rtype: int
"""
if self._max_atoms_per_dimension_per_core is None:
return self.n_atoms
self.__check_atoms_per_core()
return int(numpy.prod(self._max_atoms_per_dimension_per_core))
[docs]
def get_max_atoms_per_dimension_per_core(self) -> Tuple[int, ...]:
"""
Gets the maximum number of atoms per dimension per core. This
will return a tuple with a number for each dimension of the vertex,
which might be one if this is a single-dimension vertex.
:rtype: tuple(int,...)
"""
if self._max_atoms_per_dimension_per_core is None:
return self.atoms_shape
self.__check_atoms_per_core()
return self._max_atoms_per_dimension_per_core
def _set_max_atoms_per_dimension_per_core(
self, new_value: Optional[Union[int, Tuple[int, ...]]]) -> None:
"""
Set the maximum number of atoms per dimension per core.
Can be used to raise or lower the maximum number of atoms per core
or per dimension per core.
:param new_value:
Value to set. If the vertex is n-dimensional where n > 1, a tuple
of n values must be given. If the vertex is 1 dimensional,
a 1-tuple or integer can be given. If this is set to `None` the
vertex will have atoms_shape as the maximum.
:type new_value: None or int or tuple(int,...)
"""
if new_value is None:
self._max_atoms_per_dimension_per_core = None
elif numpy.isscalar(new_value):
max_atoms_int: int = int(cast(int, new_value))
self._max_atoms_per_dimension_per_core = (max_atoms_int, )
else:
max_atoms_tuple: Tuple[int, ...] = cast(
Tuple[int, ...], new_value)
self._max_atoms_per_dimension_per_core = max_atoms_tuple
[docs]
def set_max_atoms_per_dimension_per_core(
self, new_value: Union[int, Tuple[int, ...]]) -> None:
"""
Set the maximum number of atoms per dimension per core.
Can be used to raise or lower the maximum number of atoms per core
or per dimension per core.
:param new_value:
Value to set. If the vertex is n-dimensional where n > 1, a tuple
of n values must be given. If the vertex is 1 dimensional,
a 1-tuple or integer can be given.
"""
self._set_max_atoms_per_dimension_per_core(new_value)
self.__check_atoms_per_core()
# delayed import due to circular dependencies
# pylint: disable=import-outside-toplevel
from pacman.data import PacmanDataView
PacmanDataView.set_requires_mapping()
[docs]
def reset(self) -> None:
"""
Forget all machine vertices in the application vertex, and reset
the splitter (if any).
"""
self._machine_vertices = OrderedSet()
if self._splitter is not None:
self._splitter.reset_called()
[docs]
def get_machine_fixed_key_and_mask(
self, machine_vertex: MachineVertex,
partition_id: str) -> Optional[BaseKeyAndMask]:
"""
Get a fixed key and mask for the given machine vertex and partition
identifier, or `None` if not fixed (the default).
If this doesn't return `None`, :py:meth:`get_fixed_key_and_mask`
must also not return `None`,
and the keys returned here must align with those such that for each
``key:mask`` returned here, ``key & app_mask == app_key``. It is OK
for this to return `None` and :py:meth:`get_fixed_key_and_mask` to
return non-`None` if and only if there is only one machine vertex.
:param ~pacman.model.graphs.machine.MachineVertex machine_vertex:
A source machine vertex of this application vertex
:param str partition_id:
The identifier of the partition to get the key for
:rtype: ~pacman.model.routing_info.BaseKeyAndMask or None
"""
# pylint: disable=unused-argument
return None
[docs]
def get_fixed_key_and_mask(
self, partition_id: str) -> Optional[BaseKeyAndMask]:
"""
Get a fixed key and mask for the application vertex or `None` if not
fixed (the default). See :py:meth:`get_machine_gixed_key_and_mask` for
the conditions.
:param str partition_id:
The identifier of the partition to get the key for
:rtype: ~pacman.model.routing_info.BaseKeyAndMask or None
"""
# pylint: disable=unused-argument
return None
[docs]
def add_incoming_edge(self, edge: ApplicationEdge,
partition: ApplicationEdgePartition) -> None:
"""
Add an edge incoming to this vertex. This is ignored by default,
but could be used to track incoming edges, and/or report faults.
:param ~pacman.model.graphs.application.ApplicationEdge edge:
The edge to add.
:param partition: The partition to add the edge to.
:type partition:
~pacman.model.graphs.application.ApplicationEdgePartition
"""
[docs]
def get_key_ordered_indices(
self, indices: Optional[numpy.ndarray] = None) -> numpy.ndarray:
"""
Get indices of the vertex in the order that atoms appear when the
vertex is split into cores as determined by max_atoms_per_core. When
a multi-dimensional vertex is split into cores, the atoms on each
vertex is not linear but rather a hyper-rectangle of the atoms, thus
the order of the atoms in the vertex as a whole is not the same as the
order of the vertex when scanning over the cores.
:param indices:
Optional subset of indices to convert. If not provided all indices
will be converted.
:type indices: numpy.ndarray or None.
:rtype: numpy.ndarray
"""
if indices is None:
indices = numpy.arange(self.n_atoms)
atoms_shape = self.atoms_shape
n_dims = len(atoms_shape)
if n_dims == 1:
return indices
atoms_per_core = self.get_max_atoms_per_dimension_per_core()
remainders = numpy.array(indices)
cum_per_core = 1
cum_cores_per_dim = 1.0
core_index = numpy.zeros(len(indices))
atom_index = numpy.zeros(len(indices))
for n in range(n_dims):
# Work out the global index in this dimension
global_index_d = remainders % atoms_shape[n]
# Work out the core index in this dimension
core_index_d = global_index_d // atoms_per_core[n]
# Update the core index by multiplying the current value by the
# last dimension size and then add the current dimension value
core_index += core_index_d * cum_cores_per_dim
# Work out the atom index on the core in this dimension
atom_index_d = global_index_d - (atoms_per_core[n] * core_index_d)
# Update the atom index by multiplying the current value by the
# last dimension size and then add the current dimension value
atom_index += atom_index_d * cum_per_core
# Update the remainders for next time based on the values in
# this dimension, and save the sizes for this dimension
remainders = remainders // atoms_shape[n]
cum_per_core *= atoms_per_core[n]
cum_cores_per_dim *= atoms_shape[n] / atoms_per_core[n]
return ((core_index * self.get_max_atoms_per_core()) +
atom_index).astype(numpy.uint32)
[docs]
def get_raster_ordered_indices(
self, indices: numpy.ndarray) -> numpy.ndarray:
"""
Convert indices from key order to raster order.
:param numpy.ndarray indices: The key-ordered indices to convert.
:rtype: numpy.ndarray
"""
atoms_shape = self.atoms_shape
n_dims = len(atoms_shape)
if n_dims == 1:
return indices
atoms_per_core = self.get_max_atoms_per_dimension_per_core()
cores_per_dim = numpy.divide(atoms_shape, atoms_per_core)
cum_size = 1
global_index = numpy.zeros(len(indices))
# Work out the core indices and core-based atom indices
core_index_remainders = indices // self.get_max_atoms_per_core()
atom_index_remainders = indices % self.get_max_atoms_per_core()
for n in range(n_dims):
# Work out the core index and atom index in this dimension
core_index_d = core_index_remainders % cores_per_dim[n]
atom_index_d = atom_index_remainders % atoms_per_core[n]
# Use these to work out the global index in this dimension
global_index_d = (core_index_d * atoms_per_core[n]) + atom_index_d
# Update the global index with this dimension
global_index += global_index_d * cum_size
# Update the remainders and sizes for the next loop run
core_index_remainders = core_index_remainders // cores_per_dim[n]
atom_index_remainders = atom_index_remainders // atoms_per_core[n]
cum_size *= atoms_shape[n]
return global_index.astype(numpy.uint32)
[docs]
def has_fixed_location(self) -> bool:
"""
Check if this vertex or any machine vertex has a fixed location.
:rtype: bool
:returns: True if the Application Vertex or any one of its
Machine Vertices has a fixed location
False if None of the Vertices has a none None fixed location
"""
if self.get_fixed_location() is not None:
return True
for vertex in self._machine_vertices:
if vertex.get_fixed_location() is not None:
return True
return False