def _datadesc(obj: Any):
from dace import data
if isinstance(obj, data.Data):
return obj
elif symbolic.issymbolic(obj):
return data.Scalar(symbolic.symtype(obj))
elif isinstance(obj, dtypes.typeclass):
return data.Scalar(obj)
return data.Scalar(dtypes.typeclass(type(obj)))
def _create_datadescriptor(obj):
""" Creates a data descriptor from various types of objects.
@see: dace.data.Data
"""
if isinstance(obj, data.Data):
return obj
try:
return obj.descriptor
except AttributeError:
if isinstance(obj, numpy.ndarray):
return data.Array(dtype=types.typeclass(obj.dtype.type),
shape=obj.shape)
if symbolic.issymbolic(obj):
return data.Scalar(symbolic.symtype(obj))
if isinstance(obj, types.typeclass):
return data.Scalar(obj)
return data.Scalar(types.typeclass(type(obj)))
def generate_code(
self,
sdfg: SDFG,
schedule: Optional[dtypes.ScheduleType],
sdfg_id: str = ""
) -> Tuple[str, str, Set[TargetCodeGenerator], Set[str]]:
""" Generate frame code for a given SDFG, calling registered targets'
code generation callbacks for them to generate their own code.
:param sdfg: The SDFG to generate code for.
:param schedule: The schedule the SDFG is currently located, or
None if the SDFG is top-level.
:param sdfg_id: An optional string id given to the SDFG label
:return: A tuple of the generated global frame code, local frame
code, and a set of targets that have been used in the
generation of this SDFG.
"""
if len(sdfg_id) == 0 and sdfg.sdfg_id != 0:
sdfg_id = '_%d' % sdfg.sdfg_id
global_stream = CodeIOStream()
callsite_stream = CodeIOStream()
is_top_level = sdfg.parent is None
# Analyze allocation lifetime of SDFG and all nested SDFGs
if is_top_level:
self.determine_allocation_lifetime(sdfg)
# Generate code
###########################
# Keep track of allocated variables
allocated = set()
# Add symbol mappings to allocated variables
if sdfg.parent_nsdfg_node is not None:
allocated |= sdfg.parent_nsdfg_node.symbol_mapping.keys()
# Invoke all instrumentation providers
for instr in self._dispatcher.instrumentation.values():
if instr is not None:
instr.on_sdfg_begin(sdfg, callsite_stream, global_stream)
# Allocate outer-level transients
self.allocate_arrays_in_scope(sdfg, sdfg, global_stream,
callsite_stream)
# Allocate inter-state variables
global_symbols = copy.deepcopy(sdfg.symbols)
global_symbols.update(
{aname: arr.dtype
for aname, arr in sdfg.arrays.items()})
interstate_symbols = {}
for e in sdfg.edges():
symbols = e.data.new_symbols(sdfg, global_symbols)
# Inferred symbols only take precedence if global symbol not defined
symbols = {
k: v if k not in global_symbols else global_symbols[k]
for k, v in symbols.items()
}
interstate_symbols.update(symbols)
global_symbols.update(symbols)
for isvarName, isvarType in interstate_symbols.items():
isvar = data.Scalar(isvarType)
callsite_stream.write(
'%s;\n' % (isvar.as_arg(with_types=True, name=isvarName)),
sdfg)
self.dispatcher.defined_vars.add(isvarName,
disp.DefinedType.Scalar,
isvarType.ctype)
callsite_stream.write('\n', sdfg)
#######################################################################
# Generate actual program body
states_generated = self.generate_states(sdfg, global_stream,
callsite_stream)
#######################################################################
# Sanity check
if len(states_generated) != len(sdfg.nodes()):
raise RuntimeError(
"Not all states were generated in SDFG {}!"
"\n Generated: {}\n Missing: {}".format(
sdfg.label, [s.label for s in states_generated],
[s.label for s in (set(sdfg.nodes()) - states_generated)]))
# Deallocate transients
self.deallocate_arrays_in_scope(sdfg, sdfg, global_stream,
callsite_stream)
# Now that we have all the information about dependencies, generate
# header and footer
if is_top_level:
# Let each target append code to frame code state before generating
# header and footer
for target in self._dispatcher.used_targets:
target.on_target_used()
header_stream = CodeIOStream()
header_global_stream = CodeIOStream()
footer_stream = CodeIOStream()
footer_global_stream = CodeIOStream()
# Get all environments used in the generated code, including
# dependent environments
import dace.library # Avoid import loops
self.environments = dace.library.get_environments_and_dependencies(
self._dispatcher.used_environments)
self.generate_header(sdfg, header_global_stream, header_stream)
# Open program function
params = sdfg.signature()
if params:
params = ', ' + params
function_signature = (
'void __program_%s_internal(%s_t *__state%s)\n{\n' %
(sdfg.name, sdfg.name, params))
self.generate_footer(sdfg, footer_global_stream, footer_stream)
header_global_stream.write(global_stream.getvalue())
header_global_stream.write(footer_global_stream.getvalue())
generated_header = header_global_stream.getvalue()
all_code = CodeIOStream()
all_code.write(function_signature)
all_code.write(header_stream.getvalue())
all_code.write(callsite_stream.getvalue())
all_code.write(footer_stream.getvalue())
generated_code = all_code.getvalue()
else:
generated_header = global_stream.getvalue()
generated_code = callsite_stream.getvalue()
# Clean up generated code
gotos = re.findall(r'goto (.*?);', generated_code)
clean_code = ''
for line in generated_code.split('\n'):
# Empty line with semicolon
if re.match(r'^\s*;\s*', line):
continue
# Label that might be unused
label = re.findall(
r'^\s*([a-zA-Z_][a-zA-Z_0-9]*):\s*[;]?\s*////.*$', line)
if len(label) > 0:
if label[0] not in gotos:
continue
clean_code += line + '\n'
# Return the generated global and local code strings
return (generated_header, clean_code, self._dispatcher.used_targets,
self._dispatcher.used_environments)
def generate_code(
self,
sdfg: SDFG,
schedule: Optional[dtypes.ScheduleType],
sdfg_id: str = ""
) -> Tuple[str, str, Set[TargetCodeGenerator], Set[str]]:
""" Generate frame code for a given SDFG, calling registered targets'
code generation callbacks for them to generate their own code.
:param sdfg: The SDFG to generate code for.
:param schedule: The schedule the SDFG is currently located, or
None if the SDFG is top-level.
:param sdfg_id: An optional string id given to the SDFG label
:return: A tuple of the generated global frame code, local frame
code, and a set of targets that have been used in the
generation of this SDFG.
"""
if len(sdfg_id) == 0 and sdfg.sdfg_id != 0:
sdfg_id = '_%d' % sdfg.sdfg_id
sdfg_label = sdfg.name + sdfg_id
global_stream = CodeIOStream()
callsite_stream = CodeIOStream()
# Set default storage/schedule types in SDFG
set_default_schedule_and_storage_types(sdfg, schedule)
is_top_level = sdfg.parent is None
# Generate code
###########################
# Invoke all instrumentation providers
for instr in self._dispatcher.instrumentation.values():
if instr is not None:
instr.on_sdfg_begin(sdfg, callsite_stream, global_stream)
# Allocate outer-level transients
shared_transients = sdfg.shared_transients()
allocated = set()
for state in sdfg.nodes():
for node in state.data_nodes():
if (node.data in shared_transients
and node.data not in allocated):
self._dispatcher.dispatch_allocate(sdfg, state, None, node,
global_stream,
callsite_stream)
allocated.add(node.data)
# Allocate inter-state variables
global_symbols = copy.deepcopy(sdfg.symbols)
interstate_symbols = {}
for e in sdfg.edges():
symbols = e.data.new_symbols(global_symbols)
interstate_symbols.update(symbols)
global_symbols.update(symbols)
for isvarName, isvarType in interstate_symbols.items():
# Skip symbols that have been declared as outer-level transients
if isvarName in allocated:
continue
isvar = data.Scalar(isvarType)
callsite_stream.write(
'%s;\n' % (isvar.as_arg(with_types=True, name=isvarName)),
sdfg)
callsite_stream.write('\n', sdfg)
states_topological = list(sdfg.topological_sort(sdfg.start_state))
# {edge: [dace.edges.ControlFlow]}
control_flow = {e: [] for e in sdfg.edges()}
if dace.config.Config.get_bool('optimizer', 'detect_control_flow'):
####################################################################
# Loop detection procedure
all_cycles = list(sdfg.find_cycles()) # Returns a list of lists
# Order according to topological sort
all_cycles = [
sorted(c, key=lambda x: states_topological.index(x))
for c in all_cycles
]
# Group in terms of starting node
starting_nodes = [c[0] for c in all_cycles]
# Order cycles according to starting node in topological sort
starting_nodes = sorted(starting_nodes,
key=lambda x: states_topological.index(x))
cycles_by_node = [[c for c in all_cycles if c[0] == n]
for n in starting_nodes]
for cycles in cycles_by_node:
# Use arbitrary cycle to find the first and last nodes
first_node = cycles[0][0]
last_node = cycles[0][-1]
if not first_node.is_empty():
# The entry node should not contain any computations
continue
if not all([c[-1] == last_node for c in cycles]):
# There are multiple back edges: not a for or while loop
continue
previous_edge = [
e for e in sdfg.in_edges(first_node) if e.src != last_node
]
if len(previous_edge) != 1:
# No single starting point: not a for or while
continue
previous_edge = previous_edge[0]
back_edge = sdfg.edges_between(last_node, first_node)
if len(back_edge) != 1:
raise RuntimeError("Expected exactly one edge in cycle")
back_edge = back_edge[0]
# Build a set of all nodes in all cycles associated with this
# set of start and end node
internal_nodes = functools.reduce(
lambda a, b: a | b, [set(c)
for c in cycles]) - {first_node}
exit_edge = [
e for e in sdfg.out_edges(first_node)
if e.dst not in internal_nodes | {first_node}
]
if len(exit_edge) != 1:
# No single stopping condition: not a for or while
# (we don't support continue or break)
continue
exit_edge = exit_edge[0]
entry_edge = [
e for e in sdfg.out_edges(first_node) if e != exit_edge
]
if len(entry_edge) != 1:
# No single starting condition: not a for or while
continue
entry_edge = entry_edge[0]
# Make sure this is not already annotated to be another construct
if (len(control_flow[entry_edge]) != 0
or len(control_flow[back_edge]) != 0):
continue
# Nested loops case I - previous edge of internal loop is a
# loop-entry of an external loop (first state in a loop is
# another loop)
if (len(control_flow[previous_edge]) == 1 and isinstance(
control_flow[previous_edge][0], cflow.LoopEntry)):
# Nested loop, mark parent scope
loop_parent = control_flow[previous_edge][0].scope
# Nested loops case II - exit edge of internal loop is a
# back-edge of an external loop (last state in a loop is another
# loop)
elif (len(control_flow[exit_edge]) == 1 and isinstance(
control_flow[exit_edge][0], cflow.LoopBack)):
# Nested loop, mark parent scope
loop_parent = control_flow[exit_edge][0].scope
elif (len(control_flow[exit_edge]) == 0
or len(control_flow[previous_edge]) == 0):
loop_parent = None
else:
continue
if entry_edge == back_edge:
# No entry check (we don't support do-loops)
# TODO: do we want to add some support for self-loops?
continue
# Now we make sure that there is no other way to exit this
# cycle, by checking that there's no reachable node *not*
# included in any cycle between the first and last node.
if any([len(set(c) - internal_nodes) > 1 for c in cycles]):
continue
# This is a loop! Generate the necessary annotation objects.
loop_scope = cflow.LoopScope(internal_nodes)
if ((len(previous_edge.data.assignments) > 0
or len(back_edge.data.assignments) > 0) and
(len(control_flow[previous_edge]) == 0 or
(len(control_flow[previous_edge]) == 1 and
control_flow[previous_edge][0].scope == loop_parent))):
# Generate assignment edge, if available
control_flow[previous_edge].append(
cflow.LoopAssignment(loop_scope, previous_edge))
# Assign remaining control flow constructs
control_flow[entry_edge].append(
cflow.LoopEntry(loop_scope, entry_edge))
control_flow[exit_edge].append(
cflow.LoopExit(loop_scope, exit_edge))
control_flow[back_edge].append(
cflow.LoopBack(loop_scope, back_edge))
###################################################################
# If/then/else detection procedure
candidates = [
n for n in states_topological if sdfg.out_degree(n) == 2
]
for candidate in candidates:
# A valid if occurs when then are no reachable nodes for either
# path that does not pass through a common dominator.
dominators = nx.dominance.dominance_frontiers(
sdfg.nx, candidate)
left_entry, right_entry = sdfg.out_edges(candidate)
if (len(control_flow[left_entry]) > 0
or len(control_flow[right_entry]) > 0):
# Already assigned to a control flow construct
# TODO: carefully allow this in some cases
continue
left, right = left_entry.dst, right_entry.dst
dominator = dominators[left] & dominators[right]
if len(dominator) != 1:
# There must be a single dominator across both branches,
# unless one of the nodes _is_ the next dominator
# if (len(dominator) == 0 and dominators[left] == {right}
# or dominators[right] == {left}):
# dominator = dominators[left] | dominators[right]
# else:
# continue
continue
dominator = next(iter(dominator)) # Exactly one dominator
exit_edges = sdfg.in_edges(dominator)
if len(exit_edges) != 2:
# There must be a single entry and a single exit. This
# could be relaxed in the future.
continue
left_exit, right_exit = exit_edges
if (len(control_flow[left_exit]) > 0
or len(control_flow[right_exit]) > 0):
# Already assigned to a control flow construct
# TODO: carefully allow this in some cases
continue
# Now traverse from the source and verify that all possible paths
# pass through the dominator
left_nodes = sdfg.all_nodes_between(left, dominator)
if left_nodes is None:
# Not all paths lead to the next dominator
continue
right_nodes = sdfg.all_nodes_between(right, dominator)
if right_nodes is None:
# Not all paths lead to the next dominator
continue
all_nodes = left_nodes | right_nodes
# Make sure there is no overlap between left and right nodes
if len(left_nodes & right_nodes) > 0:
continue
# This is a valid if/then/else construct. Generate annotations
if_then_else = cflow.IfThenElse(candidate, dominator)
# Arbitrarily assign then/else to the two branches. If one edge
# has no dominator but leads to the dominator, it means there's
# only a then clause (and no else).
has_else = False
if len(dominators[left]) == 1:
then_scope = cflow.IfThenScope(if_then_else, left_nodes)
else_scope = cflow.IfElseScope(if_then_else, right_nodes)
control_flow[left_entry].append(
cflow.IfEntry(then_scope, left_entry))
control_flow[left_exit].append(
cflow.IfExit(then_scope, left_exit))
control_flow[right_exit].append(
cflow.IfExit(else_scope, right_exit))
if len(dominators[right]) == 1:
control_flow[right_entry].append(
cflow.IfEntry(else_scope, right_entry))
has_else = True
else:
then_scope = cflow.IfThenScope(if_then_else, right_nodes)
else_scope = cflow.IfElseScope(if_then_else, left_nodes)
control_flow[right_entry].append(
cflow.IfEntry(then_scope, right_entry))
control_flow[right_exit].append(
cflow.IfExit(then_scope, right_exit))
control_flow[left_exit].append(
cflow.IfExit(else_scope, left_exit))
#######################################################################
# Generate actual program body
states_generated = set() # For sanity check
generated_edges = set()
self.generate_states(sdfg, "sdfg", control_flow,
global_stream, callsite_stream,
set(states_topological), states_generated,
generated_edges)
#######################################################################
# Sanity check
if len(states_generated) != len(sdfg.nodes()):
raise RuntimeError(
"Not all states were generated in SDFG {}!"
"\n Generated: {}\n Missing: {}".format(
sdfg.label, [s.label for s in states_generated],
[s.label for s in (set(sdfg.nodes()) - states_generated)]))
# Deallocate transients
shared_transients = sdfg.shared_transients()
deallocated = set()
for state in sdfg.nodes():
for node in state.data_nodes():
if (node.data in shared_transients
and node.data not in deallocated):
self._dispatcher.dispatch_deallocate(
sdfg, state, None, node, global_stream,
callsite_stream)
deallocated.add(node.data)
# Now that we have all the information about dependencies, generate
# header and footer
if is_top_level:
header_stream = CodeIOStream()
header_global_stream = CodeIOStream()
footer_stream = CodeIOStream()
footer_global_stream = CodeIOStream()
self.generate_header(sdfg, self._dispatcher.used_environments,
header_global_stream, header_stream)
# Open program function
function_signature = 'void __program_%s_internal(%s)\n{\n' % (
sdfg.name, sdfg.signature())
self.generate_footer(sdfg, self._dispatcher.used_environments,
footer_global_stream, footer_stream)
header_global_stream.write(global_stream.getvalue())
header_global_stream.write(footer_global_stream.getvalue())
generated_header = header_global_stream.getvalue()
all_code = CodeIOStream()
all_code.write(function_signature)
all_code.write(header_stream.getvalue())
all_code.write(callsite_stream.getvalue())
all_code.write(footer_stream.getvalue())
generated_code = all_code.getvalue()
else:
generated_header = global_stream.getvalue()
generated_code = callsite_stream.getvalue()
# Clean up generated code
gotos = re.findall(r'goto (.*);', generated_code)
clean_code = ''
for line in generated_code.split('\n'):
# Empty line with semicolon
if re.match(r'^\s*;\s*', line):
continue
# Label that might be unused
label = re.findall(
r'^\s*([a-zA-Z_][a-zA-Z_0-9]*):\s*[;]?\s*////.*$', line)
if len(label) > 0:
if label[0] not in gotos:
continue
clean_code += line + '\n'
# Return the generated global and local code strings
return (generated_header, clean_code, self._dispatcher.used_targets,
self._dispatcher.used_environments)
