def remove_edge_and_dangling_path(state: SDFGState, edge: MultiConnectorEdge):
"""
Removes an edge and all of its parent edges in a memlet path, cleaning
dangling connectors and isolated nodes resulting from the removal.
:param state: The state in which the edge exists.
:param edge: The edge to remove.
"""
mtree = state.memlet_tree(edge)
inwards = (isinstance(edge.src, nd.EntryNode)
or isinstance(edge.dst, nd.EntryNode))
# Traverse tree upwards, removing edges and connectors as necessary
curedge = mtree
while curedge is not None:
e = curedge.edge
state.remove_edge(e)
if inwards:
neighbors = [
neighbor for neighbor in state.out_edges(e.src)
if e.src_conn == neighbor.src_conn
]
else:
neighbors = [
neighbor for neighbor in state.in_edges(e.dst)
if e.dst_conn == neighbor.dst_conn
]
if len(neighbors) > 0: # There are still edges connected, leave as-is
break
# Remove connector and matching outer connector
if inwards:
if e.src_conn:
e.src.remove_out_connector(e.src_conn)
e.src.remove_in_connector('IN' + e.src_conn[3:])
else:
if e.dst_conn:
e.dst.remove_in_connector(e.dst_conn)
e.src.remove_out_connector('OUT' + e.dst_conn[2:])
# Continue traversing upwards
curedge = curedge.parent
else:
# Check if an isolated node have been created at the root and remove
root_edge = mtree.root().edge
root_node: nd.Node = root_edge.src if inwards else root_edge.dst
if state.degree(root_node) == 0:
state.remove_node(root_node)
def apply(self, graph: SDFGState, sdfg: SDFG):
in_array = self.in_array
med_array = self.med_array
out_array = self.out_array
# Modify all edges that point to in_array to point to out_array
for in_edge in graph.in_edges(in_array):
# Make all memlets that write to in_array write to out_array instead
tree = graph.memlet_tree(in_edge)
for te in tree:
if te.data.data == in_array.data:
te.data.data = out_array.data
# Redirect edge to in_array
graph.remove_edge(in_edge)
graph.add_edge(in_edge.src, in_edge.src_conn, out_array, None,
in_edge.data)
graph.remove_node(med_array)
graph.remove_node(in_array)
def apply(self, graph: SDFGState, sdfg: SDFG):
import dace.libraries.blas as blas
transpose_a = self.transpose_a
_at = self.at
transpose_b = self.transpose_b
_bt = self.bt
a_times_b = self.a_times_b
for src, src_conn, _, _, memlet in graph.in_edges(transpose_a):
graph.add_edge(src, src_conn, a_times_b, '_b', memlet)
graph.remove_node(transpose_a)
for src, src_conn, _, _, memlet in graph.in_edges(transpose_b):
graph.add_edge(src, src_conn, a_times_b, '_a', memlet)
graph.remove_node(transpose_b)
graph.remove_node(_at)
graph.remove_node(_bt)
for _, _, dst, dst_conn, memlet in graph.out_edges(a_times_b):
subset = dcpy(memlet.subset)
subset.squeeze()
size = subset.size()
shape = [size[1], size[0]]
break
tmp_name, tmp_arr = sdfg.add_temp_transient(shape, a_times_b.dtype)
tmp_acc = graph.add_access(tmp_name)
transpose_c = blas.Transpose('_Transpose_', a_times_b.dtype)
for edge in graph.out_edges(a_times_b):
_, _, dst, dst_conn, memlet = edge
graph.remove_edge(edge)
graph.add_edge(transpose_c, '_out', dst, dst_conn, memlet)
graph.add_edge(a_times_b, '_c', tmp_acc, None, dace.Memlet.from_array(tmp_name, tmp_arr))
graph.add_edge(tmp_acc, None, transpose_c, '_inp', dace.Memlet.from_array(tmp_name, tmp_arr))
def apply(self, graph: SDFGState, sdfg: SDFG):
first_map_entry = graph.entry_node(self.first_map_exit)
intermediate_dnodes = set()
for _, _, node, _, _ in graph.out_edges(self.first_map_exit):
if not isinstance(node, nds.AccessNode):
continue
intermediate_dnodes.add(node)
self._update_in_connectors(graph, intermediate_dnodes)
self._replicate_first_map(sdfg, graph, first_map_entry,
intermediate_dnodes)
graph.remove_nodes_from(
graph.all_nodes_between(first_map_entry, self.first_map_exit)
| {first_map_entry, self.first_map_exit})
for node in graph.nodes():
if not isinstance(node, nds.AccessNode):
continue
if graph.in_degree(node) == 0 and graph.out_degree(node) == 0:
graph.remove_node(node)
def apply(self, graph: SDFGState, sdfg: SDFG):
"""
This method applies the mapfusion transformation.
Other than the removal of the second map entry node (SME), and the first
map exit (FME) node, it has the following side effects:
1. Any transient adjacent to both FME and SME with degree = 2 will be removed.
The tasklets that use/produce it shall be connected directly with a
scalar/new transient (if the dataflow is more than a single scalar)
2. If this transient is adjacent to FME and SME and has other
uses, it will be adjacent to the new map exit post fusion.
Tasklet-> Tasklet edges will ALSO be added as mentioned above.
3. If an access node is adjacent to FME but not SME, it will be
adjacent to new map exit post fusion.
4. If an access node is adjacent to SME but not FME, it will be
adjacent to the new map entry node post fusion.
"""
first_exit = self.first_map_exit
first_entry = graph.entry_node(first_exit)
second_entry = self.second_map_entry
second_exit = graph.exit_node(second_entry)
intermediate_nodes = set()
for _, _, dst, _, _ in graph.out_edges(first_exit):
intermediate_nodes.add(dst)
assert isinstance(dst, nodes.AccessNode)
# Check if an access node refers to non transient memory, or transient
# is used at another location (cannot erase)
do_not_erase = set()
for node in intermediate_nodes:
if sdfg.arrays[node.data].transient is False:
do_not_erase.add(node)
else:
for edge in graph.in_edges(node):
if edge.src != first_exit:
do_not_erase.add(node)
break
else:
for edge in graph.out_edges(node):
if edge.dst != second_entry:
do_not_erase.add(node)
break
# Find permutation between first and second scopes
perm = self.find_permutation(first_entry.map, second_entry.map)
params_dict = {}
for index, param in enumerate(first_entry.map.params):
params_dict[param] = second_entry.map.params[perm[index]]
# Replaces (in memlets and tasklet) the second scope map
# indices with the permuted first map indices.
# This works in two passes to avoid problems when e.g., exchanging two
# parameters (instead of replacing (j,i) and (i,j) to (j,j) and then
# i,i).
second_scope = graph.scope_subgraph(second_entry)
for firstp, secondp in params_dict.items():
if firstp != secondp:
replace(second_scope, secondp, '__' + secondp + '_fused')
for firstp, secondp in params_dict.items():
if firstp != secondp:
replace(second_scope, '__' + secondp + '_fused', firstp)
# Isolate First exit node
############################
edges_to_remove = set()
nodes_to_remove = set()
for edge in graph.in_edges(first_exit):
tree = graph.memlet_tree(edge)
access_node = tree.root().edge.dst
if access_node not in do_not_erase:
out_edges = [
e for e in graph.out_edges(access_node)
if e.dst == second_entry
]
# In this transformation, there can only be one edge to the
# second map
assert len(out_edges) == 1
# Get source connector to the second map
connector = out_edges[0].dst_conn[3:]
new_dsts = []
# Look at the second map entry out-edges to get the new
# destinations
for e in graph.out_edges(second_entry):
if e.src_conn[4:] == connector:
new_dsts.append(e)
if not new_dsts: # Access node is not used in the second map
nodes_to_remove.add(access_node)
continue
# Add a transient scalar/array
self.fuse_nodes(sdfg, graph, edge, new_dsts[0].dst,
new_dsts[0].dst_conn, new_dsts[1:])
edges_to_remove.add(edge)
# Remove transient node between the two maps
nodes_to_remove.add(access_node)
else: # The case where intermediate array node cannot be removed
# Node will become an output of the second map exit
out_e = tree.parent.edge
conn = second_exit.next_connector()
graph.add_edge(
second_exit,
'OUT_' + conn,
out_e.dst,
out_e.dst_conn,
dcpy(out_e.data),
)
second_exit.add_out_connector('OUT_' + conn)
graph.add_edge(edge.src, edge.src_conn, second_exit,
'IN_' + conn, dcpy(edge.data))
second_exit.add_in_connector('IN_' + conn)
edges_to_remove.add(out_e)
edges_to_remove.add(edge)
# If the second map needs this node, link the connector
# that generated this to the place where it is needed, with a
# temp transient/scalar for memlet to be generated
for out_e in graph.out_edges(second_entry):
second_memlet_path = graph.memlet_path(out_e)
source_node = second_memlet_path[0].src
if source_node == access_node:
self.fuse_nodes(sdfg, graph, edge, out_e.dst,
out_e.dst_conn)
###
# First scope exit is isolated and can now be safely removed
for e in edges_to_remove:
graph.remove_edge(e)
graph.remove_nodes_from(nodes_to_remove)
graph.remove_node(first_exit)
# Isolate second_entry node
###########################
for edge in graph.in_edges(second_entry):
tree = graph.memlet_tree(edge)
access_node = tree.root().edge.src
if access_node in intermediate_nodes:
# Already handled above, can be safely removed
graph.remove_edge(edge)
continue
# This is an external input to the second map which will now go
# through the first map.
conn = first_entry.next_connector()
graph.add_edge(edge.src, edge.src_conn, first_entry, 'IN_' + conn,
dcpy(edge.data))
first_entry.add_in_connector('IN_' + conn)
graph.remove_edge(edge)
for out_enode in tree.children:
out_e = out_enode.edge
graph.add_edge(
first_entry,
'OUT_' + conn,
out_e.dst,
out_e.dst_conn,
dcpy(out_e.data),
)
graph.remove_edge(out_e)
first_entry.add_out_connector('OUT_' + conn)
###
# Second node is isolated and can now be safely removed
graph.remove_node(second_entry)
# Fix scope exit to point to the right map
second_exit.map = first_entry.map
