litex/migen/flow/network.py

327 lines
11 KiB
Python

from collections import defaultdict
from migen.fhdl.std import *
from migen.genlib.misc import optree
from migen.flow.actor import *
from migen.flow import plumbing
# Abstract actors mean that the actor class should be instantiated with the parameters
# from the dictionary. They are needed to enable actor duplication or sharing during
# elaboration, and automatic parametrization of plumbing actors.
class AbstractActor:
def __init__(self, actor_class, parameters=dict(), name=None):
self.actor_class = actor_class
self.parameters = parameters
self.name = name
self.busy = Signal()
def create_instance(self):
return self.actor_class(**self.parameters)
def __repr__(self):
r = "<abstract " + self.actor_class.__name__
if self.name is not None:
r += ": " + self.name
r += ">"
return r
class MultiDiGraph:
def __init__(self):
self.edges = defaultdict(list)
self.incoming = defaultdict(set)
self.outgoing = defaultdict(set)
self.nodes = set()
def add_edge(self, a, b, **edge):
self.edges[(a, b)].append(edge)
self.incoming[b].add(a)
self.outgoing[a].add(b)
self.nodes |= {a, b}
def __iter__(self):
return iter(self.nodes)
def __len__(self):
return len(self.nodes)
def edges_iter(self, data=True):
assert data
for (a, b), edges in self.edges.items():
for edge in edges:
yield a, b, edge
def get_edge_data(self, a, b):
return dict(enumerate(self.edges[(a, b)]))
def add_node(self, node):
self.nodes.add(node)
def remove_node(self, node):
for i in self.incoming.pop(node):
del self.edges[(i, node)]
self.outgoing[i].remove(node)
for i in self.outgoing.pop(node):
del self.edges[(node, i)]
self.incoming[i].remove(node)
self.nodes.remove(node)
def remove_edge(self, a, b, key):
e = self.edges[(a, b)]
del e[key]
if not e:
self.incoming[b].remove(a)
self.outgoing[a].remove(b)
def in_edges(self, sink, data=True):
assert data
e = []
for source in self.incoming[sink]:
for edge in self.edges[(source, sink)]:
e.append((source, sink, edge))
return e
def out_edges(self, source, data=True):
assert data
e = []
for sink in self.outgoing[source]:
for edge in self.edges[(source, sink)]:
e.append((source, sink, edge))
return e
# TODO: rewrite this without non-determinism
class DataFlowGraph(MultiDiGraph):
def __init__(self):
MultiDiGraph.__init__(self)
self.elaborated = False
self.abstract_busy_signals = dict()
def add_connection(self, source_node, sink_node,
source_ep=None, sink_ep=None, # default: assume nodes have 1 source/sink and use that one
source_subr=None, sink_subr=None): # default: use whole record
self.add_edge(source_node, sink_node,
source=source_ep, sink=sink_ep,
source_subr=source_subr, sink_subr=sink_subr)
def add_buffered_connection(self, source_node, sink_node,
source_ep=None, sink_ep=None,
source_subr=None, sink_subr=None):
buf = AbstractActor(plumbing.Buffer)
self.add_connection(source_node, buf, source_ep=source_ep, source_subr=source_subr)
self.add_connection(buf, sink_node, sink_ep=sink_ep, sink_subr=sink_subr)
def add_pipeline(self, *nodes):
for n1, n2 in zip(nodes, nodes[1:]):
self.add_connection(n1, n2)
def del_connections(self, source_node, sink_node, data_requirements):
edges_to_delete = []
edge_data = self.get_edge_data(source_node, sink_node)
if edge_data is None:
# the two nodes are already completely disconnected
return
for key, data in edge_data.items():
if all(k not in data_requirements or data_requirements[k] == v
for k, v in data.items()):
edges_to_delete.append(key)
for key in edges_to_delete:
self.remove_edge(source_node, sink_node, key)
def replace_actor(self, old, new):
self.add_node(new)
for xold, v, data in self.out_edges(old, data=True):
self.add_edge(new, v, **data)
for u, xold, data in self.in_edges(old, data=True):
self.add_edge(u, new, **data)
self.remove_node(old)
def instantiate(self, actor):
inst = actor.create_instance()
self.abstract_busy_signals[id(inst)] = actor.busy
self.replace_actor(actor, inst)
# Returns a dictionary
# source -> [sink1, ..., sinkn]
# source element is a (node, endpoint) pair.
# sink elements are (node, endpoint, source subrecord, sink subrecord) triples.
def _source_to_sinks(self):
d = dict()
for u, v, data in self.edges_iter(data=True):
el_src = (u, data["source"])
el_dst = (v, data["sink"], data["source_subr"], data["sink_subr"])
if el_src in d:
d[el_src].append(el_dst)
else:
d[el_src] = [el_dst]
return d
# Returns a dictionary
# sink -> [source1, ... sourcen]
# sink element is a (node, endpoint) pair.
# source elements are (node, endpoint, sink subrecord, source subrecord) triples.
def _sink_to_sources(self):
d = dict()
for u, v, data in self.edges_iter(data=True):
el_src = (u, data["source"], data["sink_subr"], data["source_subr"])
el_dst = (v, data["sink"])
if el_dst in d:
d[el_dst].append(el_src)
else:
d[el_dst] = [el_src]
return d
# List sources that feed more than one sink.
def _list_divergences(self):
d = self._source_to_sinks()
return dict((k, v) for k, v in d.items() if len(v) > 1)
# A graph is abstract if any of these conditions is met:
# (1) A node is an abstract actor.
# (2) A subrecord is used.
# (3) A single source feeds more than one sink.
# NB: It is not allowed for a single sink to be fed by more than one source
# (except with subrecords, i.e. when a combinator is used)
def is_abstract(self):
return any(isinstance(x, AbstractActor) for x in self) \
or any(d["source_subr"] is not None or d["sink_subr"] is not None
for u, v, d in self.edges_iter(data=True)) \
or bool(self._list_divergences())
def _eliminate_subrecords_and_divergences(self):
# Insert combinators.
for (dst_node, dst_endpoint), sources in self._sink_to_sources().items():
if len(sources) > 1 or sources[0][2] is not None:
# build combinator
# "layout" is filled in during instantiation
subrecords = [dst_subrecord for src_node, src_endpoint, dst_subrecord, src_subrecord in sources]
combinator = AbstractActor(plumbing.Combinator, {"subrecords": subrecords})
# disconnect source1 -> sink ... sourcen -> sink
# connect source1 -> combinator_sink1 ... sourcen -> combinator_sinkn
for n, (src_node, src_endpoint, dst_subrecord, src_subrecord) in enumerate(sources):
self.del_connections(src_node, dst_node,
{"source": src_endpoint, "sink": dst_endpoint})
self.add_connection(src_node, combinator,
src_endpoint, "sink{0}".format(n), source_subr=src_subrecord)
# connect combinator_source -> sink
self.add_connection(combinator, dst_node, "source", dst_endpoint)
# Insert splitters.
for (src_node, src_endpoint), sinks in self._source_to_sinks().items():
if len(sinks) > 1 or sinks[0][2] is not None:
subrecords = [src_subrecord for dst_node, dst_endpoint, src_subrecord, dst_subrecord in sinks]
splitter = AbstractActor(plumbing.Splitter, {"subrecords": subrecords})
# disconnect source -> sink1 ... source -> sinkn
# connect splitter_source1 -> sink1 ... splitter_sourcen -> sinkn
for n, (dst_node, dst_endpoint, src_subrecord, dst_subrecord) in enumerate(sinks):
self.del_connections(src_node, dst_node,
{"source": src_endpoint, "sink": dst_endpoint})
self.add_connection(splitter, dst_node,
"source{0}".format(n), dst_endpoint)
# connect source -> splitter_sink
self.add_connection(src_node, splitter, src_endpoint, "sink")
def _infer_plumbing_layout(self):
while True:
ap = [a for a in self if isinstance(a, AbstractActor) and a.actor_class in plumbing.actors]
if not ap:
break
for a in ap:
in_edges = self.in_edges(a, data=True)
out_edges = self.out_edges(a, data=True)
if a.actor_class in plumbing.layout_sink and len(in_edges) == 1:
other, me, data = in_edges[0]
if isinstance(other, AbstractActor):
continue
other_ep = data["source"]
if other_ep is None:
other_ep = get_single_ep(other, Source)[1]
else:
other_ep = getattr(other, other_ep)
elif a.actor_class in plumbing.layout_source and len(out_edges) == 1:
me, other, data = out_edges[0]
if isinstance(other, AbstractActor):
continue
other_ep = data["sink"]
if other_ep is None:
other_ep = get_single_ep(other, Sink)[1]
else:
other_ep = getattr(other, other_ep)
else:
raise AssertionError
layout = other_ep.payload.layout
a.parameters["layout"] = layout
self.instantiate(a)
def _instantiate_actors(self):
# 1. instantiate all abstract non-plumbing actors
for actor in list(self):
if isinstance(actor, AbstractActor) and actor.actor_class not in plumbing.actors:
self.instantiate(actor)
# 2. infer plumbing layout and instantiate plumbing
self._infer_plumbing_layout()
# 3. resolve default eps
for u, v, d in self.edges_iter(data=True):
if d["source"] is None:
d["source"] = get_single_ep(u, Source)[0]
if d["sink"] is None:
d["sink"] = get_single_ep(v, Sink)[0]
# Elaboration turns an abstract DFG into a physical one.
# Pass 1: eliminate subrecords and divergences
# by inserting Combinator/Splitter actors
# Pass 2: run optimizer (e.g. share and duplicate actors)
# Pass 3: instantiate all abstract actors and explicit "None" endpoints
def elaborate(self, optimizer=None):
if self.elaborated:
return
self.elaborated = True
self._eliminate_subrecords_and_divergences()
if optimizer is not None:
optimizer(self)
self._instantiate_actors()
class CompositeActor(Module):
def __init__(self, dfg):
dfg.elaborate()
# expose unconnected endpoints
uc_eps_by_node = dict((node, get_endpoints(node)) for node in dfg)
for u, v, d in dfg.edges_iter(data=True):
uc_eps_u = uc_eps_by_node[u]
source = d["source"]
try:
del uc_eps_u[source]
except KeyError:
pass
uc_eps_v = uc_eps_by_node[v]
sink = d["sink"]
try:
del uc_eps_v[sink]
except KeyError:
pass
for node, uc_eps in uc_eps_by_node.items():
for k, v in uc_eps.items():
assert(not hasattr(self, k))
setattr(self, k, v)
# connect abstract busy signals
for node in dfg:
try:
abstract_busy_signal = dfg.abstract_busy_signals[id(node)]
except KeyError:
pass
else:
self.comb += abstract_busy_signal.eq(node.busy)
# generate busy signal
self.busy = Signal()
self.comb += self.busy.eq(optree("|", [node.busy for node in dfg]))
# claim ownership of sub-actors and establish connections
for node in dfg:
self.submodules += node
for u, v, d in dfg.edges_iter(data=True):
ep_src = getattr(u, d["source"])
ep_dst = getattr(v, d["sink"])
self.comb += ep_src.connect_flat(ep_dst)