def __init__(self):
super(Group, self).__init__()
self._subsystems = OrderedDict()
self._local_subsystems = OrderedDict()
self._src = {}
self._src_idxs = {}
self._data_xfer = {}
self._local_unknown_sizes = None
self._local_param_sizes = None
# These solvers are the default
self.ln_solver = ScipyGMRES()
self.nl_solver = RunOnce()
class Group(System):
"""A system that contains other systems."""
def __init__(self):
super(Group, self).__init__()
self._subsystems = OrderedDict()
self._local_subsystems = OrderedDict()
self._src = {}
self._src_idxs = {}
self._data_xfer = {}
self._local_unknown_sizes = None
self._local_param_sizes = None
# These solvers are the default
self.ln_solver = ScipyGMRES()
self.nl_solver = RunOnce()
def __getattr__(self, name):
return self._subsystems[name]
def __setitem__(self, name, val):
"""Sets the given value into the appropriate `VecWrapper`.
Args
----
name : str
The name of the variable to set into the unknowns vector.
"""
if self.is_active():
try:
self.unknowns[name] = val
except KeyError:
# look in params
try:
subname, vname = name.rsplit('.', 1)
self._subsystem(subname).params[vname] = val
except:
msg = "Can't find variable '%s' in unknowns or params " + \
"vectors in system '%s'"
raise KeyError(msg % (name, self.pathname))
def __getitem__(self, name):
"""
Retrieve unflattened value of named unknown or unconnected
param variable.
Args
----
name : str
The name of the variable to retrieve from the unknowns vector.
Returns
-------
The unflattened value of the given variable.
"""
# if setup has not been called, then there is no variable information to access
if not self._local_unknown_sizes:
raise RuntimeError('setup() must be called before variables can be accessed')
# if system is not active, then it's not valid to access it's variables
if not self.is_active():
raise AttributeError("System '%s' is inactive, so can't access variable '%s'" %
(self.pathname, name))
try:
return self.unknowns[name]
except KeyError:
subsys, subname = name.split('.', 1)
try:
return self._subsystems[subsys][subname]
except:
# look in params
try:
subname, vname = name.rsplit('.', 1)
return self._subsystem(subname).params[vname]
except:
msg = "Can't find variable '%s' in unknowns or params " + \
"vectors in system '%s'"
raise KeyError(msg % (name, self.pathname))
def _subsystem(self, name):
"""
Returns a reference to a named subsystem that is a direct or an indirect
subsystem of the this system. Raises an exception if the given name
doesn't reference a subsystem.
Args
----
name : str
Name of the subsystem to retrieve.
Returns
-------
`System`
A reference to the named subsystem.
"""
s = self
for part in name.split('.'):
s = s._subsystems[part]
return s
def add(self, name, system, promotes=None):
"""Add a subsystem to this group, specifying its name and any variables
that it promotes to the parent level.
Args
----
name : str
The name by which the subsystem is to be known.
system : `System`
The subsystem to be added.
promotes : tuple, optional
The names of variables in the subsystem which are to be promoted.
"""
if promotes is not None:
system._promotes = promotes
if name in self._subsystems.keys():
msg = "Group '{gname}' already contains a subsystem with name"\
" '{cname}'.".format(gname=self.name, cname=name)
raise RuntimeError(msg)
self._subsystems[name] = system
system.name = name
return system
def connect(self, source, targets, src_indices=None):
"""Connect the given source variable to the given target
variable.
Args
----
source : source
The name of the source variable.
targets : str OR iterable
The name of one or more target variables.
"""
if isinstance(targets, str):
targets = (targets,)
for target in targets:
self._src.setdefault(target, []).append(source)
if src_indices is not None:
self._src_idxs[target] = src_indices
def subsystems(self, local=False, recurse=False, typ=System, include_self=False):
"""
Args
----
local : bool, optional
If True, only return those `Components` that are local. Default is False.
recurse : bool, optional
If True, return all `Components` in the system tree, subject to
the value of the local arg. Default is False.
typ : type, optional
If a class is specified here, only those subsystems that are instances
of that type will be returned. Default type is `System`.
include_self : bool, optional
If True, yield self before iterating over subsystems, assuming type
of self is appropriate. Default is False.
Returns
-------
iterator
Iterator over subsystems.
"""
if include_self and isinstance(self, typ):
yield self
subs = self._local_subsystems if local else self._subsystems
for name, sub in subs.items():
if isinstance(sub, typ):
yield sub
if recurse and isinstance(sub, Group):
for s in sub.subsystems(local, recurse, typ):
yield s
def subgroups(self, local=False, recurse=False, include_self=False):
"""
Returns
-------
iterator
Iterator over subgroups.
"""
return self.subsystems(local=local, recurse=recurse, typ=Group,
include_self=include_self)
def components(self, local=False, recurse=False, include_self=False):
"""
Returns
-------
iterator
Iterator over sub-`Components`.
"""
return self.subsystems(local=local, recurse=recurse, typ=Component,
include_self=include_self)
def _setup_paths(self, parent_path):
"""Set the absolute pathname of each `System` in the tree.
Args
----
parent_path : str
The pathname of the parent `System`, which is to be prepended to the
name of this child `System` and all subsystems.
"""
super(Group, self)._setup_paths(parent_path)
for sub in self.subsystems():
sub._setup_paths(self.pathname)
def _setup_variables(self):
"""
Create dictionaries of metadata for parameters and for unknowns for
this `Group` and stores them as attributes of the `Group`. The
promoted name of subsystem variables with respect to this `Group`
is included in the metadata.
Returns
-------
tuple
A dictionary of metadata for parameters and for unknowns
for all subsystems.
"""
self._params_dict = OrderedDict()
self._unknowns_dict = OrderedDict()
self._data_xfer = {}
for sub in self.subsystems():
subparams, subunknowns = sub._setup_variables()
for p, meta in subparams.items():
meta = meta.copy()
meta['promoted_name'] = self._promoted_name(meta['promoted_name'], sub)
if p in self._src_idxs:
meta['src_indices'] = self._src_idxs[p]
self._params_dict[p] = meta
for u, meta in subunknowns.items():
meta = meta.copy()
meta['promoted_name'] = self._promoted_name(meta['promoted_name'], sub)
self._unknowns_dict[u] = meta
# check for any promotes that didn't match a variable
sub._check_promotes()
return self._params_dict, self._unknowns_dict
def _promoted_name(self, name, subsystem):
"""
Returns
-------
str
The promoted name of the given variable.
"""
if subsystem.promoted(name):
return name
if len(subsystem.name) > 0:
return '.'.join((subsystem.name, name))
else:
return name
def _setup_communicators(self, comm):
"""
Assign communicator to this `Group` and all of its subsystems.
Args
----
comm : an MPI communicator (real or fake)
The communicator being offered by the parent system.
"""
self._local_subsystems = OrderedDict()
self.comm = comm
for sub in self.subsystems():
sub._setup_communicators(self.comm)
if self.is_active() and sub.is_active():
self._local_subsystems[sub.name] = sub
def _setup_vectors(self, param_owners, parent=None,
relevance=None, top_unknowns=None, impl=BasicImpl):
"""Create `VecWrappers` for this `Group` and all below it in the
`System` tree.
Args
----
param_owners : dict
A dictionary mapping `System` pathnames to the pathnames of parameters
they are reponsible for propagating.
parent : `Group`, optional
The `Group` that contains this `Group`, if any.
relevance : `Relevance`
An object that stores relevance information for each variable of interest.
top_unknowns : `VecWrapper`, optional
The `Problem` level unknowns `VecWrapper`.
impl : an implementation factory, optional
Specifies the factory object used to create `VecWrapper` and
`DataXfer` objects.
"""
self.params = self.unknowns = self.resids = None
self.dumat, self.dpmat, self.drmat = {}, {}, {}
self._local_unknown_sizes = None
self._local_param_sizes = None
self._owning_ranks = None
if not self.is_active():
return
self._impl_factory = impl
self._relevance = relevance
my_params = param_owners.get(self.pathname, [])
if parent is None:
self._create_vecs(my_params, relevance, var_of_interest=None,
impl=impl)
top_unknowns = self.unknowns
else:
self._create_views(top_unknowns, parent, my_params, relevance,
var_of_interest=None)
self._local_unknown_sizes = self.unknowns._get_flattened_sizes()
self._local_param_sizes = self.params._get_flattened_sizes()
self._owning_ranks = self._get_owning_ranks()
self._setup_data_transfer(my_params, relevance, None)
# TODO: determine the size of the largest grouping of parallel subvecs, allocate
# an array of that size, and sub-allocate from that for all relevant subvecs
# We should never need more memory than the largest sized collection of parallel
# vecs.
# create storage for the relevant vecwrappers,
# keyed by variable_of_interest
for group, vois in self._relevance.groups.items():
if group is not None:
for voi in vois:
if parent is None:
self._create_vecs(my_params, relevance, voi, impl)
else:
self._create_views(top_unknowns, parent, my_params,
relevance, voi)
self._setup_data_transfer(my_params, relevance, voi)
# convert any src_indices to index arrays
for meta in self._params_dict.values():
if 'src_indices' in meta:
meta['src_indices'] = self.params.to_idx_array(meta['src_indices'])
for sub in self.subsystems():
sub._setup_vectors(param_owners, parent=self,
relevance=relevance, top_unknowns=top_unknowns)
# now that all of the vectors and subvecs are allocated, calculate
# and cache the ls_inputs.
self._ls_inputs = {}
for voi, vec in self.dumat.items():
self._ls_inputs[voi] = self._all_params(voi)
def _get_fd_params(self):
"""
Get the list of parameters that are needed to perform a
finite difference on this `Group`.
Returns
-------
list of str
List of names of params that have sources that are ParamComps
or sources that are outside of this `Group` .
"""
conns = self.connections
mypath = self.pathname + '.' if self.pathname else ''
params = []
for tgt, src in conns.items():
if tgt.startswith(mypath):
# look up the Component that contains the source variable
scname = src.rsplit('.', 1)[0]
if scname.startswith(mypath):
src_comp = self._subsystem(scname[len(mypath):])
if isinstance(src_comp, ParamComp):
params.append(tgt[len(mypath):])
else:
params.append(tgt[len(mypath):])
return params
def _get_fd_unknowns(self):
"""
Get the list of unknowns that are needed to perform a
finite difference on this `Group`.
Returns
-------
list of str
List of names of unknowns for this `Group` that don't come from a
`ParamComp`.
"""
mypath = self.pathname + '.' if self.pathname else ''
fd_unknowns = []
for name, meta in self.unknowns.items():
# look up the subsystem containing the unknown
sub = self._subsystem(meta['pathname'].rsplit('.', 1)[0][len(mypath):])
if not isinstance(sub, ParamComp):
fd_unknowns.append(name)
return fd_unknowns
def _get_explicit_connections(self):
"""
Returns
-------
dict
Explicit connections in this `Group`, represented as a mapping
from the pathname of the target to the pathname of the source.
"""
connections = {}
for sub in self.subgroups():
connections.update(sub._get_explicit_connections())
for tgt, srcs in self._src.items():
for src in srcs:
try:
src_pathnames = get_absvarpathnames(src, self._unknowns_dict, 'unknowns')
except KeyError as error:
try:
# if src is a param, it must use scoped absolute naming, so convert to top
# level absolute name for lookup in params_dict
s = self._get_var_pathname(src)
# verify that src is actually in self._params_dict
self._params_dict[s]
src_pathnames = [s]
except KeyError as error:
raise ConnectError.nonexistent_src_error(src, tgt)
try:
for tgt_pathname in get_absvarpathnames(tgt, self._params_dict, 'params'):
connections.setdefault(tgt_pathname, []).extend(src_pathnames)
except KeyError as error:
try:
get_absvarpathnames(tgt, self._unknowns_dict, 'unknowns')
except KeyError as error:
raise ConnectError.nonexistent_target_error(src, tgt)
else:
raise ConnectError.invalid_target_error(src, tgt)
return connections
def solve_nonlinear(self, params=None, unknowns=None, resids=None, metadata=None):
"""
Solves the group using the slotted nl_solver.
Args
----
params : `VecWrapper`, optional
`VecWrapper` containing parameters. (p)
unknowns : `VecWrapper`, optional
`VecWrapper` containing outputs and states. (u)
resids : `VecWrapper`, optional
`VecWrapper` containing residuals. (r)
metadata : dict, optional
Dictionary containing execution metadata (e.g. iteration coordinate).
"""
if self.is_active():
params = params if params is not None else self.params
unknowns = unknowns if unknowns is not None else self.unknowns
resids = resids if resids is not None else self.resids
self.nl_solver.solve(params, unknowns, resids, self, metadata)
def children_solve_nonlinear(self, metadata):
"""
Loops over our children systems and asks them to solve.
Args
----
metadata : dict
Dictionary containing execution metadata (e.g. iteration coordinate).
"""
# transfer data to each subsystem and then solve_nonlinear it
for sub in self.subsystems():
self._transfer_data(sub.name)
if sub.is_active():
if isinstance(sub, Component):
sub.solve_nonlinear(sub.params, sub.unknowns, sub.resids)
else:
sub.solve_nonlinear(sub.params, sub.unknowns, sub.resids, metadata)
def apply_nonlinear(self, params, unknowns, resids, metadata=None):
"""
Evaluates the residuals of our children systems.
Args
----
params : `VecWrapper`
`VecWrapper` containing parameters. (p)
unknowns : `VecWrapper`
`VecWrapper` containing outputs and states. (u)
resids : `VecWrapper`
`VecWrapper` containing residuals. (r)
metadata : dict, optional
Dictionary containing execution metadata (e.g. iteration coordinate).
"""
if not self.is_active():
return
# transfer data to each subsystem and then apply_nonlinear to it
for sub in self.subsystems():
self._transfer_data(sub.name)
if sub.is_active():
if isinstance(sub, Component):
sub.apply_nonlinear(sub.params, sub.unknowns, sub.resids)
else:
sub.apply_nonlinear(sub.params, sub.unknowns, sub.resids, metadata)
def jacobian(self, params, unknowns, resids):
"""
Linearize all our subsystems.
Args
----
params : `VecWrapper`
`VecWrapper` containing parameters. (p)
unknowns : `VecWrapper`
`VecWrapper` containing outputs and states. (u)
resids : `VecWrapper`
`VecWrapper` containing residuals. (r)
"""
for sub in self.subsystems(local=True):
# Instigate finite difference on child if user requests.
if sub.fd_options['force_fd'] == True:
# Groups need total derivatives
if isinstance(sub, Group):
total_derivs = True
else:
total_derivs = False
jacobian_cache = sub.fd_jacobian(sub.params, sub.unknowns,
sub.resids,
total_derivs=total_derivs)
else:
jacobian_cache = sub.jacobian(sub.params, sub.unknowns,
sub.resids)
# Cache the Jacobian for Components that aren't Paramcomps.
# Also cache it for systems that are finite differenced.
if (isinstance(sub, Component) or \
sub.fd_options['force_fd'] == True) and \
not isinstance(sub, ParamComp):
sub._jacobian_cache = jacobian_cache
# The user might submit a scalar Jacobian as a float.
# It is really inconvenient if we don't allow it.
if jacobian_cache is not None:
for key, J in iteritems(jacobian_cache):
if isinstance(J, real_types):
jacobian_cache[key] = np.array([[J]])
shape = jacobian_cache[key].shape
if len(shape) < 2:
jacobian_cache[key] = jacobian_cache[key].reshape((shape[0], 1))
def apply_linear(self, mode, ls_inputs=None, vois=[None]):
"""Calls apply_linear on our children. If our child is a `Component`,
then we need to also take care of the additional 1.0 on the diagonal
for explicit outputs.
df = du - dGdp * dp or du = df and dp = -dGdp^T * df
Args
----
mode : string
Derivative mode, can be 'fwd' or 'rev'.
ls_inputs : dict
We can only solve derivatives for the inputs the instigating
system has access to.
vois: list of strings
List of all quantities of interest to key into the mats.
"""
if not self.is_active():
return
if mode == 'fwd':
self._transfer_data(deriv=True) # Full Scatter
for sub in self.subsystems(local=True):
# Components that are not paramcomps perform a matrix-vector
# product on their variables. Any group where the user requests
# a finite difference is also treated as a component.
if (isinstance(sub, Component) or \
sub.fd_options['force_fd'] == True) and \
not isinstance(sub, ParamComp):
self._sub_apply_linear_wrapper(sub, mode, vois, ls_inputs)
# Groups and all other systems just call their own apply_linear.
else:
sub.apply_linear(mode, ls_inputs=ls_inputs, vois=vois)
if mode == 'rev':
self._transfer_data(mode='rev', deriv=True) # Full Scatter
def _sub_apply_linear_wrapper(self, system, mode, vois, ls_inputs=None):
"""
Calls apply_linear on any Component-like subsystem. This
basically does two things: 1) multiplies the user Jacobian by -1, and
2) puts a 1 on the diagonal for all explicit outputs.
Args
----
system : `System`
Subsystem of interest, either a `Component` or a `Group` that is
being finite differenced.
mode : string
Derivative mode, can be 'fwd' or 'rev'.
vois: list of strings
List of all quantities of interest to key into the mats.
ls_inputs : dict
We can only solve derivatives for the inputs the instigating
system has access to.
"""
for voi in vois:
dresids = system.drmat[voi]
dunknowns = system.dumat[voi]
dparams = system.dpmat[voi]
# Linear GS imposes a stricter requirement on whether or not to run.
abs_inputs = {meta['pathname'] for meta in dparams.values()}
# Forward Mode
if mode == 'fwd':
dresids.vec[:] = 0.0
if ls_inputs[voi] is None or abs_inputs.intersection(ls_inputs[voi]):
if system.fd_options['force_fd'] == True:
system._apply_linear_jac(system.params, system.unknowns, dparams,
dunknowns, dresids, mode)
else:
system.apply_linear(system.params, system.unknowns, dparams,
dunknowns, dresids, mode)
dresids.vec *= -1.0
for var, meta in dunknowns.items():
# Skip all states
if not meta.get('state'):
dresids[var] += dunknowns[var]
# Adjoint Mode
elif mode == 'rev':
dparams.vec[:] = 0.0
# Sign on the local Jacobian needs to be -1 before
# we add in the fake residual. Since we can't modify
# the 'du' vector at this point without stomping on the
# previous component's contributions, we can multiply
# our local 'arg' by -1, and then revert it afterwards.
dresids.vec *= -1.0
if ls_inputs[voi] is None or set(abs_inputs).intersection(ls_inputs[voi]):
try:
dparams._set_adjoint_mode(True)
if system.fd_options['force_fd'] == True:
system._apply_linear_jac(system.params,
system.unknowns, dparams,
dunknowns, dresids, mode)
else:
system.apply_linear(system.params, system.unknowns,
dparams, dunknowns, dresids, mode)
finally:
dparams._set_adjoint_mode(False)
dresids.vec *= -1.0
for var, meta in dunknowns.items():
# Skip all states
if not meta.get('state'):
dunknowns[var] += dresids[var]
def solve_linear(self, dumat, drmat, vois, mode=None):
"""
Single linear solution applied to whatever input is sitting in
the rhs vector.
Args
----
dumat : dict of `VecWrappers`
In forward mode, each `VecWrapper` contains the incoming vector
for the states. There is one vector per quantity of interest for
this problem. In reverse mode, it contains the outgoing vector for
the states. (du)
drmat : `dict of VecWrappers`
`VecWrapper` containing either the outgoing result in forward mode
or the incoming vector in reverse mode. There is one vector per
quantity of interest for this problem. (dr)
vois: list of strings
List of all quantities of interest to key into the mats.
mode : string, optional
Derivative mode, can be 'fwd' or 'rev', but generally should be
called without mode so that the user can set the mode in this
system's ln_solver.options.
"""
if not self.is_active():
return
if mode is None:
mode = self.fd_options['mode']
if mode == 'fwd':
sol_vec, rhs_vec = dumat, drmat
else:
sol_vec, rhs_vec = drmat, dumat
# TODO: Need the norm. Loop over vois here.
#if np.linalg.norm(rhs) < 1e-15:
# sol_vec.vec[:] = 0.0
# return
# Solve Jacobian, df |-> du [fwd] or du |-> df [rev]
rhs_buf = {}
for voi in vois:
rhs_buf[voi] = rhs_vec[voi].vec.copy()
sol_buf = self.ln_solver.solve(rhs_buf, self, mode=mode)
for voi in vois:
sol_vec[voi].vec[:] = sol_buf[voi][:]
def _all_params(self, voi=None):
""" Returns the set of all parameters in this system and all subsystems.
Args
----
voi: string
Variable of interest, default is None.
"""
# TODO: clean this up
ls_inputs = set(self.dpmat[voi].keys())
abs_uvec = {meta['pathname'] for meta in self.dumat[voi].values()}
for comp in self.components(local=True, recurse=True):
for intinp_rel, meta in comp.dpmat[voi].items():
intinp_abs = meta['pathname']
src = self.connections.get(intinp_abs)
if src in abs_uvec:
ls_inputs.add(intinp_abs)
return ls_inputs
def dump(self, nest=0, out_stream=sys.stdout, verbose=True, dvecs=False):
"""
Writes a formated dump of the `System` tree to file.
Args
----
nest : int, optional
Starting nesting level. Defaults to 0.
out_stream : file-like, optional
Where output is written. Defaults to sys.stdout.
verbose : bool, optional
If True (the default), output additional info beyond
just the tree structure.
dvecs : bool, optional
If True, show contents of du and dp vectors instead of
u and p (the default).
"""
klass = self.__class__.__name__
if dvecs:
ulabel, plabel, uvecname, pvecname = 'du', 'dp', 'dunknowns', 'dparams'
else:
ulabel, plabel, uvecname, pvecname = 'u', 'p', 'unknowns', 'params'
uvec = getattr(self, uvecname)
pvec = getattr(self, pvecname)
commsz = self.comm.size if hasattr(self.comm, 'size') else 0
template = "%s %s '%s' req: %s usize:%d psize:%d commsize:%d\n"
out_stream.write(template % (" "*nest,
klass,
self.name,
self.get_req_procs(),
uvec.vec.size,
pvec.vec.size,
commsz))
vec_conns = dict(self._data_xfer[('', 'fwd', None)].vec_conns)
byobj_conns = dict(self._data_xfer[('', 'fwd', None)].byobj_conns)
# collect width info
lens = [len(u)+sum(map(len, v)) for u, v in
chain(vec_conns.items(), byobj_conns.items())]
if lens:
nwid = max(lens) + 9
else:
lens = [len(n) for n in uvec.keys()]
nwid = max(lens) if lens else 12
for v, meta in uvec.items():
if verbose:
if meta.get('pass_by_obj') or meta.get('remote'):
continue
out_stream.write(" "*(nest+8))
uslice = '{0}[{1[0]}:{1[1]}]'.format(ulabel, uvec._slices[v])
pnames = [p for p, u in vec_conns.items() if u == v]
if pnames:
if len(pnames) == 1:
pname = pnames[0]
pslice = pvec._slices.get(pname, (-1, -1))
pslice = '%d:%d' % (pslice[0], pslice[1])
else:
pslice = [('%d:%d' % pvec._slices.get(p, (-1, -1))) for p in pnames]
if len(pslice) > 1:
pslice = ','.join(pslice)
else:
pslice = pslice[0]
pslice = '{}[{}]'.format(plabel, pslice)
connstr = '%s -> %s' % (v, pnames)
template = "{0:<{nwid}} {1:<10} {2:<10} {3:>10}\n"
out_stream.write(template.format(connstr,
uslice,
pslice,
repr(uvec[v]),
nwid=nwid))
else:
template = "{0:<{nwid}} {1:<21} {2:>10}\n"
out_stream.write(template.format(v,
uslice,
repr(uvec[v]),
nwid=nwid))
if not dvecs:
for dest, src in byobj_conns.items():
out_stream.write(" "*(nest+8))
connstr = '%s -> %s:' % (src, dest)
template = "{0:<{nwid}} (by_obj) ({1})\n"
out_stream.write(template.format(connstr,
repr(uvec[src]),
nwid=nwid))
nest += 3
for sub in self.subsystems(local=True):
sub.dump(nest, out_stream=out_stream, verbose=verbose, dvecs=dvecs)
out_stream.flush()
def get_req_procs(self):
"""
Returns
-------
tuple
A tuple of the form (min_procs, max_procs), indicating the min and max
processors usable by this `Group`.
"""
min_procs = 1
max_procs = 1
for sub in self.subsystems():
sub_min, sub_max = sub.get_req_procs()
min_procs = max(min_procs, sub_min)
if max_procs is not None:
if sub_max is None:
max_procs = None
else:
max_procs = max(max_procs, sub_max)
return (min_procs, max_procs)
def _get_global_offset(self, name, var_rank, sizes_table, var_of_interest):
"""
Args
----
name : str
The variable name.
var_rank : int
The rank the the offset is requested for.
sizes_table : list of OrderDicts mappping var name to size.
Size information for all vars in all ranks.
var_of_interest : str
Name of the current variable of interest, the key into the
dumat,drmat, and dpmat dicts.
Returns
-------
int
The offset into the distributed vector for the named variable
in the specified rank (process).
"""
offset = 0
rank = 0
# first get the offset of the distributed storage for var_rank
while rank < var_rank:
for vname, size in sizes_table[rank].items():
if self._relevance.is_relevant(var_of_interest, vname):
offset += size
rank += 1
# now, get the offset into the var_rank storage for the variable
for vname, size in sizes_table[var_rank].items():
if vname == name:
break
if self._relevance.is_relevant(var_of_interest, vname):
offset += size
return offset
def _get_global_idxs(self, uname, pname, var_of_interest, mode):
"""
Return the global indices into the distributed unknowns and params vector
for the given unknown and param. The given unknown and param have already
been tested for relevance.
Args
----
uname : str
Name of variable in the unknowns vector.
pname : str
Name of the variable in the params vector.
var_of_interest : str or None
Name of variable of interest used to determine relevance.
mode : str
Solution mode, either 'fwd' or 'rev'
Returns
-------
tuple of (idx_array, idx_array)
index array into the global unknowns vector and the corresponding
index array into the global params vector.
"""
umeta = self.unknowns.metadata(uname)
pmeta = self.params.metadata(pname)
# FIXME: if we switch to push scatters, this check will flip
if (mode == 'fwd' and pmeta.get('remote')) or (mode == 'rev' and umeta.get('remote')):
# just return empty index arrays for remote vars
return self.params.make_idx_array(0, 0), self.params.make_idx_array(0, 0)
if not self._relevance.is_relevant(var_of_interest, uname) or \
not self._relevance.is_relevant(var_of_interest, pname):
return self.params.make_idx_array(0, 0), self.params.make_idx_array(0, 0)
if self.comm is None:
iproc = 0
else:
iproc = self.comm.rank
if 'src_indices' in pmeta:
arg_idxs = self.params.to_idx_array(pmeta['src_indices'])
else:
arg_idxs = self.params.make_idx_array(0, pmeta['size'])
if mode == 'fwd':
var_rank = self._owning_ranks[uname]
else:
var_rank = iproc
offset = self._get_global_offset(uname, var_rank, self._local_unknown_sizes,
var_of_interest)
src_idxs = arg_idxs + offset
if mode == 'fwd':
var_rank = iproc
else:
var_rank = self._owning_ranks[pname]
tgt_start = self._get_global_offset(pname, var_rank, self._local_param_sizes,
var_of_interest)
tgt_idxs = tgt_start + self.params.make_idx_array(0, len(arg_idxs))
return src_idxs, tgt_idxs
def _setup_data_transfer(self, my_params, relevance, var_of_interest):
"""
Create `DataXfer` objects to handle data transfer for all of the
connections that involve parameters for which this `Group`
is responsible.
Args
----
my_params : list
List of pathnames for parameters that the `Group` is
responsible for propagating.
relevance : `Relevance`
An object containing info about what variables are relevant
to a variable of interest.
var_of_interest : str or None
The name of a variable of interest.
"""
xfer_dict = {}
for param, unknown in self.connections.items():
if not (relevance.is_relevant(var_of_interest, param) or
relevance.is_relevant(var_of_interest, unknown)):
continue
if param in my_params:
# remove our system pathname from the abs pathname of the param and
# get the subsystem name from that
tgt_sys = name_relative_to(self.pathname, param)
src_sys = name_relative_to(self.pathname, unknown)
for mode, sname in (('fwd', tgt_sys), ('rev', src_sys)):
src_idx_list, dest_idx_list, vec_conns, byobj_conns = \
xfer_dict.setdefault((sname, mode), ([], [], [], []))
urelname = self.unknowns.get_promoted_varname(unknown)
prelname = self.params.get_promoted_varname(param)
if self.unknowns.metadata(urelname).get('pass_by_obj'):
# rev is for derivs only, so no by_obj passing needed
if mode == 'fwd':
byobj_conns.append((prelname, urelname))
else: # pass by vector
sidxs, didxs = self._get_global_idxs(urelname, prelname,
var_of_interest, mode)
vec_conns.append((prelname, urelname))
src_idx_list.append(sidxs)
dest_idx_list.append(didxs)
for (tgt_sys, mode), (srcs, tgts, vec_conns, byobj_conns) in xfer_dict.items():
src_idxs, tgt_idxs = self.unknowns.merge_idxs(srcs, tgts)
if vec_conns or byobj_conns:
self._data_xfer[(tgt_sys, mode, var_of_interest)] = \
self._impl_factory.create_data_xfer(self.dumat[var_of_interest],
self.dpmat[var_of_interest],
src_idxs, tgt_idxs,
vec_conns, byobj_conns)
# create a DataXfer object that combines all of the
# individual subsystem src_idxs, tgt_idxs, and byobj_conns, so that a 'full'
# scatter to all subsystems can be done at the same time. Store that DataXfer
# object under the name ''.
for mode in ('fwd', 'rev'):
full_srcs = []
full_tgts = []
full_flats = []
full_byobjs = []
for (tgt_sys, direction), (srcs, tgts, flats, byobjs) in xfer_dict.items():
if mode == direction:
full_srcs.extend(srcs)
full_tgts.extend(tgts)
full_flats.extend(flats)
full_byobjs.extend(byobjs)
src_idxs, tgt_idxs = self.unknowns.merge_idxs(full_srcs, full_tgts)
self._data_xfer[('', mode, var_of_interest)] = \
self._impl_factory.create_data_xfer(self.dumat[var_of_interest],
self.dpmat[var_of_interest],
src_idxs, tgt_idxs,
full_flats, full_byobjs)
def _transfer_data(self, target_sys='', mode='fwd', deriv=False,
var_of_interest=None):
"""
Transfer data to/from target_system depending on mode.
Args
----
target_sys : str, optional
Name of the target `System`. A name of '', the default, indicates that data
should be transfered to all subsystems at once.
mode : { 'fwd', 'rev' }, optional
Specifies forward or reverse data transfer. Default is 'fwd'.
deriv : bool, optional
If True, use du/dp for scatter instead of u/p. Default is False.
var_of_interest : str or None
Specifies the variable of interest to determine relevance.
"""
x = self._data_xfer.get((target_sys, mode, var_of_interest))
if x is not None:
if deriv:
x.transfer(self.dumat[var_of_interest], self.dpmat[var_of_interest],
mode, deriv=True)
else:
x.transfer(self.unknowns, self.params, mode)
def _get_owning_rank(self, name, sizes_table):
"""
Args
----
name : str
Name of the variable to find the owning rank for
sizes_table : list of ordered dicts mapping name to size
Size info for all vars in all ranks.
Returns
-------
int
The current rank if it has a local copy of the named variable, else
the rank of the lowest ranked process that has a local copy.
"""
if self.comm is None:
return 0
if sizes_table[self.comm.rank][name]:
return self.comm.rank
else:
for i in range(self.comm.size):
if sizes_table[i][name]:
return i
else:
raise RuntimeError("Can't find a source for '%s' with a non-zero size" %
name)
def _get_owning_ranks(self):
"""
Determine the 'owning' rank of each variable and return a dict
mapping variables to their owning rank. The owning rank is the lowest
rank where the variable is local.
"""
ranks = {}
local_vars = [k for k, m in self.unknowns.items() if not m.get('remote')]
local_vars.extend([k for k, m in self.params.items() if not m.get('remote')])
if MPI:
all_locals = self.comm.allgather(local_vars)
else:
all_locals = [local_vars]
for rank in range(len(all_locals)):
for v in all_locals[rank]:
if v not in ranks:
ranks[v] = rank
return ranks