def test_tiled_program(self):
# Tasklet only
sdfg, state = create_tiled_sdfg()
tasklet = next(n for n in state.nodes() if isinstance(n, Tasklet))
nest_state_subgraph(sdfg, state, SubgraphView(state, [tasklet]))
sdfg.validate()
# Inner map scope
sdfg, state = create_tiled_sdfg()
tasklet = next(n for n in state.nodes() if isinstance(n, Tasklet))
entry = state.entry_node(tasklet)
nest_state_subgraph(sdfg, state, state.scope_subgraph(entry))
sdfg.validate()
# Outer map scope
sdfg, state = create_tiled_sdfg()
sdc = state.scope_dict(True)
entry = next(n for n in sdc[None] if isinstance(n, MapEntry))
nest_state_subgraph(sdfg, state, state.scope_subgraph(entry))
sdfg.validate()
# Entire state
sdfg, state = create_tiled_sdfg()
nest_state_subgraph(sdfg, state, SubgraphView(state, state.nodes()))
sdfg.validate()
def test_badscope(self):
with self.assertRaises(ValueError):
sdfg, state, t, me, mx = create_sdfg()
nest_state_subgraph(sdfg, state, SubgraphView(state, [t, me]))
with self.assertRaises(ValueError):
sdfg, state, t, me, mx = create_sdfg()
nest_state_subgraph(sdfg, state, SubgraphView(state, [t, mx]))
with self.assertRaises(KeyError):
sdfg, state, t, me, mx = create_sdfg()
b_node = state.sink_nodes()[0]
sdfg, state, t, me, mx = create_sdfg()
# Notice that b_node comes from another graph
nest_state_subgraph(sdfg, state, SubgraphView(state, [t, b_node]))
def apply(self, sdfg):
graph = sdfg.nodes()[self.state_id]
if self.expr_index == 0:
map_entry = graph.nodes()[self.subgraph[
GPUTransformMap._map_entry]]
nsdfg_node = helpers.nest_state_subgraph(
sdfg,
graph,
graph.scope_subgraph(map_entry),
full_data=self.fullcopy)
else:
cnode = graph.nodes()[self.subgraph[GPUTransformMap._reduce]]
nsdfg_node = helpers.nest_state_subgraph(
sdfg,
graph,
SubgraphView(graph, [cnode]),
full_data=self.fullcopy)
# Avoiding import loops
from dace.transformation.interstate import GPUTransformSDFG
transformation = GPUTransformSDFG(0, 0, {}, 0)
transformation.register_trans = self.register_trans
transformation.sequential_innermaps = self.sequential_innermaps
transformation.toplevel_trans = self.toplevel_trans
transformation.apply(nsdfg_node.sdfg)
# Inline back as necessary
sdfg.apply_strict_transformations()
def get_actions(actions, graph, match):
subgraph_node_ids = match.subgraph.values()
subgraph_nodes = [graph.nodes()[nid] for nid in subgraph_node_ids]
for node in subgraph_nodes:
version = 0
while (node, type(match).__name__, match.expr_index,
version) in actions.keys():
version += 1
actions[(node, type(match).__name__, match.expr_index,
version)] = match
subgraph = SubgraphView(graph, subgraph_nodes)
for edge in subgraph.edges():
version = 0
while (edge, type(match).__name__, match.expr_index,
version) in actions.keys():
version += 1
actions[(edge, type(match).__name__, match.expr_index,
version)] = match
return actions
def has_surrounding_perfcounters(node, dfg: SubgraphView):
""" Returns true if there is a possibility that this node is part of a
section that is profiled. """
parent = dfg.scope_dict()[node]
if isinstance(parent, MapEntry):
if (parent.map.schedule
not in PAPIInstrumentation.perf_whitelist_schedules):
return False
return True
return False
def test_simple_program(self):
@dace.program
def multiply(a: dace.float32[N]):
a *= 2
sdfg = multiply.to_sdfg(strict=True)
for state in sdfg.nodes():
if any(isinstance(node, Tasklet) for node in state.nodes()):
break
else:
raise KeyError('State with tasklet not found')
tasklet_nodes = [n for n in state.nodes() if isinstance(n, Tasklet)]
with self.assertRaises(ValueError):
nest_state_subgraph(sdfg, state,
SubgraphView(state, tasklet_nodes))
nest_state_subgraph(sdfg, state, SubgraphView(state,
[tasklet_nodes[0]]))
sdfg.validate()
nest_state_subgraph(sdfg, state, SubgraphView(state,
[tasklet_nodes[1]]))
sdfg.validate()
def nest_state_subgraph(sdfg: SDFG,
state: SDFGState,
subgraph: SubgraphView,
name: Optional[str] = None,
full_data: bool = False) -> nodes.NestedSDFG:
""" Turns a state subgraph into a nested SDFG. Operates in-place.
:param sdfg: The SDFG containing the state subgraph.
:param state: The state containing the subgraph.
:param subgraph: Subgraph to nest.
:param name: An optional name for the nested SDFG.
:param full_data: If True, nests entire input/output data.
:return: The nested SDFG node.
:raise KeyError: Some or all nodes in the subgraph are not located in
this state, or the state does not belong to the given
SDFG.
:raise ValueError: The subgraph is contained in more than one scope.
"""
if state.parent != sdfg:
raise KeyError('State does not belong to given SDFG')
if subgraph.graph != state:
raise KeyError('Subgraph does not belong to given state')
# Find the top-level scope
scope_tree = state.scope_tree()
scope_dict = state.scope_dict()
scope_dict_children = state.scope_dict(True)
top_scopenode = -1 # Initialized to -1 since "None" already means top-level
for node in subgraph.nodes():
if node not in scope_dict:
raise KeyError('Node not found in state')
# If scope entry/exit, ensure entire scope is in subgraph
if isinstance(node, nodes.EntryNode):
scope_nodes = scope_dict_children[node]
if any(n not in subgraph.nodes() for n in scope_nodes):
raise ValueError('Subgraph contains partial scopes (entry)')
elif isinstance(node, nodes.ExitNode):
entry = state.entry_node(node)
scope_nodes = scope_dict_children[entry] + [entry]
if any(n not in subgraph.nodes() for n in scope_nodes):
raise ValueError('Subgraph contains partial scopes (exit)')
scope_node = scope_dict[node]
if scope_node not in subgraph.nodes():
if top_scopenode != -1 and top_scopenode != scope_node:
raise ValueError(
'Subgraph is contained in more than one scope')
top_scopenode = scope_node
scope = scope_tree[top_scopenode]
###
# Collect inputs and outputs of the nested SDFG
inputs: List[MultiConnectorEdge] = []
outputs: List[MultiConnectorEdge] = []
for node in subgraph.source_nodes():
inputs.extend(state.in_edges(node))
for node in subgraph.sink_nodes():
outputs.extend(state.out_edges(node))
# Collect transients not used outside of subgraph (will be removed of
# top-level graph)
data_in_subgraph = set(n.data for n in subgraph.nodes()
if isinstance(n, nodes.AccessNode))
# Find other occurrences in SDFG
other_nodes = set(
n.data for s in sdfg.nodes() for n in s.nodes()
if isinstance(n, nodes.AccessNode) and n not in subgraph.nodes())
subgraph_transients = set()
for data in data_in_subgraph:
datadesc = sdfg.arrays[data]
if datadesc.transient and data not in other_nodes:
subgraph_transients.add(data)
# All transients of edges between code nodes are also added to nested graph
for edge in subgraph.edges():
if (isinstance(edge.src, nodes.CodeNode)
and isinstance(edge.dst, nodes.CodeNode)):
subgraph_transients.add(edge.data.data)
# Collect data used in access nodes within subgraph (will be referenced in
# full upon nesting)
input_arrays = set()
output_arrays = set()
for node in subgraph.nodes():
if (isinstance(node, nodes.AccessNode)
and node.data not in subgraph_transients):
if state.out_degree(node) > 0:
input_arrays.add(node.data)
if state.in_degree(node) > 0:
output_arrays.add(node.data)
# Create the nested SDFG
nsdfg = SDFG(name or 'nested_' + state.label)
# Transients are added to the nested graph as-is
for name in subgraph_transients:
nsdfg.add_datadesc(name, sdfg.arrays[name])
# Input/output data that are not source/sink nodes are added to the graph
# as non-transients
for name in (input_arrays | output_arrays):
datadesc = copy.deepcopy(sdfg.arrays[name])
datadesc.transient = False
nsdfg.add_datadesc(name, datadesc)
# Connected source/sink nodes outside subgraph become global data
# descriptors in nested SDFG
input_names = []
output_names = []
for edge in inputs:
if edge.data.data is None: # Skip edges with an empty memlet
continue
name = '__in_' + edge.data.data
datadesc = copy.deepcopy(sdfg.arrays[edge.data.data])
datadesc.transient = False
if not full_data:
datadesc.shape = edge.data.subset.size()
input_names.append(
nsdfg.add_datadesc(name, datadesc, find_new_name=True))
for edge in outputs:
if edge.data.data is None: # Skip edges with an empty memlet
continue
name = '__out_' + edge.data.data
datadesc = copy.deepcopy(sdfg.arrays[edge.data.data])
datadesc.transient = False
if not full_data:
datadesc.shape = edge.data.subset.size()
output_names.append(
nsdfg.add_datadesc(name, datadesc, find_new_name=True))
###################
# Add scope symbols to the nested SDFG
for v in scope.defined_vars:
if v in sdfg.symbols:
sym = sdfg.symbols[v]
nsdfg.add_symbol(v, sym.dtype)
# Create nested state
nstate = nsdfg.add_state()
# Add subgraph nodes and edges to nested state
nstate.add_nodes_from(subgraph.nodes())
for e in subgraph.edges():
nstate.add_edge(e.src, e.src_conn, e.dst, e.dst_conn, e.data)
# Modify nested SDFG parents in subgraph
for node in subgraph.nodes():
if isinstance(node, nodes.NestedSDFG):
node.sdfg.parent = nstate
node.sdfg.parent_sdfg = nsdfg
# Add access nodes and edges as necessary
edges_to_offset = []
for name, edge in zip(input_names, inputs):
node = nstate.add_read(name)
new_edge = copy.deepcopy(edge.data)
new_edge.data = name
edges_to_offset.append((edge,
nstate.add_edge(node, None, edge.dst,
edge.dst_conn, new_edge)))
for name, edge in zip(output_names, outputs):
node = nstate.add_write(name)
new_edge = copy.deepcopy(edge.data)
new_edge.data = name
edges_to_offset.append((edge,
nstate.add_edge(edge.src, edge.src_conn, node,
None, new_edge)))
# Offset memlet paths inside nested SDFG according to subsets
for original_edge, new_edge in edges_to_offset:
for edge in nstate.memlet_tree(new_edge):
edge.data.data = new_edge.data.data
if not full_data:
edge.data.subset.offset(original_edge.data.subset, True)
# Add nested SDFG node to the input state
nested_sdfg = state.add_nested_sdfg(nsdfg, None,
set(input_names) | input_arrays,
set(output_names) | output_arrays)
# Reconnect memlets to nested SDFG
for name, edge in zip(input_names, inputs):
if full_data:
data = Memlet.from_array(edge.data.data,
sdfg.arrays[edge.data.data])
else:
data = edge.data
state.add_edge(edge.src, edge.src_conn, nested_sdfg, name, data)
for name, edge in zip(output_names, outputs):
if full_data:
data = Memlet.from_array(edge.data.data,
sdfg.arrays[edge.data.data])
else:
data = edge.data
state.add_edge(nested_sdfg, name, edge.dst, edge.dst_conn, data)
# Connect access nodes to internal input/output data as necessary
entry = scope.entry
exit = scope.exit
for name in input_arrays:
node = state.add_read(name)
if entry is not None:
state.add_nedge(entry, node, EmptyMemlet())
state.add_edge(node, None, nested_sdfg, name,
Memlet.from_array(name, sdfg.arrays[name]))
for name in output_arrays:
node = state.add_write(name)
if exit is not None:
state.add_nedge(node, exit, EmptyMemlet())
state.add_edge(nested_sdfg, name, node, None,
Memlet.from_array(name, sdfg.arrays[name]))
# Remove subgraph nodes from graph
state.remove_nodes_from(subgraph.nodes())
# Remove subgraph transients from top-level graph
for transient in subgraph_transients:
del sdfg.arrays[transient]
return nested_sdfg
def apply(self, sdfg):
state: SDFGState = sdfg.nodes()[self.state_id]
nsdfg_node = state.nodes()[self.subgraph[InlineSDFG._nested_sdfg]]
nsdfg: SDFG = nsdfg_node.sdfg
nstate: SDFGState = nsdfg.nodes()[0]
nsdfg_scope_entry = state.entry_node(nsdfg_node)
nsdfg_scope_exit = (state.exit_nodes(nsdfg_scope_entry)[0]
if nsdfg_scope_entry is not None else None)
#######################################################
# Collect and update top-level SDFG metadata
# Global/init/exit code
if nsdfg.global_code:
sdfg.set_global_code(sdfg.global_code + nsdfg.global_code)
if nsdfg.init_code:
sdfg.set_init_code(sdfg.init_code + nsdfg.init_code)
if nsdfg.exit_code:
sdfg.set_exit_code(sdfg.exit_code + nsdfg.exit_code)
# Find original source/destination edges (there is only one edge per
# connector, according to match)
inputs: Dict[str, MultiConnectorEdge] = {}
outputs: Dict[str, MultiConnectorEdge] = {}
input_set: Dict[str, str] = {}
output_set: Dict[str, str] = {}
for e in state.in_edges(nsdfg_node):
inputs[e.dst_conn] = e
input_set[e.data.data] = e.dst_conn
for e in state.out_edges(nsdfg_node):
outputs[e.src_conn] = e
output_set[e.data.data] = e.src_conn
# All transients become transients of the parent (if data already
# exists, find new name)
# Mapping from nested transient name to top-level name
transients: Dict[str, str] = {}
for node in nstate.nodes():
if isinstance(node, nodes.AccessNode):
datadesc = nsdfg.arrays[node.data]
if node.data not in transients and datadesc.transient:
name = sdfg.add_datadesc('%s_%s' %
(nsdfg.label, node.data),
datadesc,
find_new_name=True)
transients[node.data] = name
# All transients of edges between code nodes are also added to parent
for edge in nstate.edges():
if (isinstance(edge.src, nodes.CodeNode)
and isinstance(edge.dst, nodes.CodeNode)):
datadesc = nsdfg.arrays[edge.data.data]
if edge.data.data not in transients and datadesc.transient:
name = sdfg.add_datadesc('%s_%s' %
(nsdfg.label, edge.data.data),
datadesc,
find_new_name=True)
transients[edge.data.data] = name
# Collect nodes to add to top-level graph
new_incoming_edges: Dict[nodes.Node, MultiConnectorEdge] = {}
new_outgoing_edges: Dict[nodes.Node, MultiConnectorEdge] = {}
source_accesses = set()
sink_accesses = set()
for node in nstate.source_nodes():
if (isinstance(node, nodes.AccessNode)
and node.data not in transients):
new_incoming_edges[node] = inputs[node.data]
source_accesses.add(node)
for node in nstate.sink_nodes():
if (isinstance(node, nodes.AccessNode)
and node.data not in transients):
new_outgoing_edges[node] = outputs[node.data]
sink_accesses.add(node)
#######################################################
# Add nested SDFG into top-level SDFG
# Add nested nodes into original state
subgraph = SubgraphView(nstate, [
n for n in nstate.nodes()
if n not in (source_accesses | sink_accesses)
])
state.add_nodes_from(subgraph.nodes())
for edge in subgraph.edges():
state.add_edge(edge.src, edge.src_conn, edge.dst, edge.dst_conn,
edge.data)
#######################################################
# Replace data on inlined SDFG nodes/edges
# Replace symbols using invocation symbol mapping
# Two-step replacement (N -> __dacesym_N --> map[N]) to avoid clashes
for symname, symvalue in nsdfg_node.symbol_mapping.items():
if str(symname) != str(symvalue):
nsdfg.replace(symname, '__dacesym_' + symname)
for symname, symvalue in nsdfg_node.symbol_mapping.items():
if str(symname) != str(symvalue):
nsdfg.replace('__dacesym_' + symname, symvalue)
# Replace data names with their top-level counterparts
repldict = {}
repldict.update(transients)
repldict.update({
k: v.data.data
for k, v in itertools.chain(inputs.items(), outputs.items())
})
for node in subgraph.nodes():
if isinstance(node, nodes.AccessNode) and node.data in repldict:
node.data = repldict[node.data]
for edge in subgraph.edges():
if edge.data.data in repldict:
edge.data.data = repldict[edge.data.data]
#######################################################
# Reconnect inlined SDFG
# If a source/sink node is one of the inputs/outputs, reconnect it,
# replacing memlets in outgoing/incoming paths
modified_edges = set()
modified_edges |= self._modify_memlet_path(new_incoming_edges, nstate,
state, True)
modified_edges |= self._modify_memlet_path(new_outgoing_edges, nstate,
state, False)
# Modify all other internal edges pertaining to input/output nodes
for node in subgraph.nodes():
if isinstance(node, nodes.AccessNode):
if node.data in input_set or node.data in output_set:
if node.data in input_set:
outer_edge = inputs[input_set[node.data]]
else:
outer_edge = outputs[output_set[node.data]]
for edge in state.all_edges(node):
if (edge not in modified_edges
and edge.data.data == node.data):
for e in state.memlet_tree(edge):
if e.data.data == node.data:
e._data = helpers.unsqueeze_memlet(
e.data, outer_edge.data)
# If source/sink node is not connected to a source/destination access
# node, and the nested SDFG is in a scope, connect to scope with empty
# memlets
if nsdfg_scope_entry is not None:
for node in subgraph.nodes():
if state.in_degree(node) == 0:
state.add_edge(nsdfg_scope_entry, None, node, None,
EmptyMemlet())
if state.out_degree(node) == 0:
state.add_edge(node, None, nsdfg_scope_exit, None,
EmptyMemlet())
# Replace nested SDFG parents with new SDFG
for node in nstate.nodes():
if isinstance(node, nodes.NestedSDFG):
node.sdfg.parent = state
node.sdfg.parent_sdfg = sdfg
# Remove all unused external inputs/output memlet paths, as well as
# resulting isolated nodes
removed_in_edges = self._remove_edge_path(state,
inputs,
set(inputs.keys()) -
source_accesses,
reverse=True)
removed_out_edges = self._remove_edge_path(state,
outputs,
set(outputs.keys()) -
sink_accesses,
reverse=False)
# Re-add in/out edges to first/last nodes in subgraph
order = [
x for x in nx.topological_sort(nstate._nx)
if isinstance(x, nodes.AccessNode)
]
for edge in removed_in_edges:
# Find first access node that refers to this edge
node = next(n for n in order if n.data == edge.data.data)
state.add_edge(edge.src, edge.src_conn, node, edge.dst_conn,
edge.data)
for edge in removed_out_edges:
# Find last access node that refers to this edge
node = next(n for n in reversed(order) if n.data == edge.data.data)
state.add_edge(node, edge.src_conn, edge.dst, edge.dst_conn,
edge.data)
#######################################################
# Remove nested SDFG node
state.remove_node(nsdfg_node)
def test_simple_sdfg_program(self):
sdfg, state, t, me, mx = create_sdfg()
nest_state_subgraph(sdfg, state, SubgraphView(state, state.nodes()))
sdfg.validate()
def test_simple_sdfg_map(self):
sdfg, state, t, me, mx = create_sdfg()
nest_state_subgraph(sdfg, state, SubgraphView(state, [me, t, mx]))
sdfg.validate()