def can_be_applied(graph, candidate, expr_index, sdfg, permissive=False):
# Check the edges between the entries of the two maps.
outer_map_entry: nodes.MapEntry = graph.nodes()[candidate[
MapCollapse._outer_map_entry]]
inner_map_entry: nodes.MapEntry = graph.nodes()[candidate[
MapCollapse._inner_map_entry]]
# Check that inner map range is independent of outer range
map_deps = set()
for s in inner_map_entry.map.range:
map_deps |= set(map(str, symlist(s)))
if any(dep in outer_map_entry.map.params for dep in map_deps):
return False
# Check that the destination of all the outgoing edges
# from the outer map's entry is the inner map's entry.
for _src, _, dest, _, _ in graph.out_edges(outer_map_entry):
if dest != inner_map_entry:
return False
# Check that the source of all the incoming edges
# to the inner map's entry is the outer map's entry.
for src, _, _, dst_conn, memlet in graph.in_edges(inner_map_entry):
if src != outer_map_entry:
return False
# Check that dynamic input range memlets are independent of
# first map range
if dst_conn is not None and not dst_conn.startswith('IN_'):
memlet_deps = set()
for s in memlet.subset:
memlet_deps |= set(map(str, symlist(s)))
if any(dep in outer_map_entry.map.params
for dep in memlet_deps):
return False
# Check the edges between the exits of the two maps.
inner_map_exit = graph.exit_node(inner_map_entry)
outer_map_exit = graph.exit_node(outer_map_entry)
# Check that the destination of all the outgoing edges
# from the inner map's exit is the outer map's exit.
for _src, _, dest, _, _ in graph.out_edges(inner_map_exit):
if dest != outer_map_exit:
return False
# Check that the source of all the incoming edges
# to the outer map's exit is the inner map's exit.
for src, _, _dest, _, _ in graph.in_edges(outer_map_exit):
if src != inner_map_exit:
return False
if not permissive:
if inner_map_entry.map.schedule != outer_map_entry.map.schedule:
return False
return True
def can_be_applied(graph, candidate, expr_index, sdfg, permissive=False):
# TODO: Assuming that the subsets on the edges between the two map
# entries/exits are the union of separate inner subsets, is it possible
# that inverting these edges breaks the continuity of union? What about
# the opposite?
# Check the edges between the entries of the two maps.
outer_map_entry = graph.nodes()[candidate[
MapInterchange.outer_map_entry]]
inner_map_entry = graph.nodes()[candidate[
MapInterchange.inner_map_entry]]
# Check that inner map range is independent of outer range
map_deps = set()
for s in inner_map_entry.map.range:
map_deps |= set(map(str, symlist(s)))
if any(dep in outer_map_entry.map.params for dep in map_deps):
return False
# Check that the destination of all the outgoing edges
# from the outer map's entry is the inner map's entry.
for e in graph.out_edges(outer_map_entry):
if e.dst != inner_map_entry:
return False
# Check that the source of all the incoming edges
# to the inner map's entry is the outer map's entry.
for e in graph.in_edges(inner_map_entry):
if e.src != outer_map_entry:
return False
# Check that dynamic input range memlets are independent of
# first map range
if e.dst_conn and not e.dst_conn.startswith('IN_'):
memlet_deps = set()
for s in e.data.subset:
memlet_deps |= set(map(str, symlist(s)))
if any(dep in outer_map_entry.map.params
for dep in memlet_deps):
return False
# Check the edges between the exits of the two maps.
inner_map_exit = graph.exit_node(inner_map_entry)
outer_map_exit = graph.exit_node(outer_map_entry)
# Check that the destination of all the outgoing edges
# from the inner map's exit is the outer map's exit.
for e in graph.out_edges(inner_map_exit):
if e.dst != outer_map_exit:
return False
# Check that the source of all the incoming edges
# to the outer map's exit is the inner map's exit.
for e in graph.in_edges(outer_map_exit):
if e.src != inner_map_exit:
return False
return True
def __call__(self, *args, **kwargs):
""" Convenience function that parses, compiles, and runs a DaCe
program. """
binaryobj = self.compile(*args)
# Add named arguments to the call
kwargs.update({aname: arg for aname, arg in zip(self.argnames, args)})
# Update arguments with symbols in data shapes
kwargs.update({
sym: symbolic.symbol(sym).get()
for arg in args
for sym in (symbolic.symlist(arg.descriptor.shape) if hasattr(
arg, 'descriptor') else [])
})
# Update arguments with symbol values
for aname in self.argnames:
if aname in binaryobj.sdfg.arrays:
sym_shape = binaryobj.sdfg.arrays[aname].shape
for sym in (sym_shape):
if symbolic.issymbolic(sym):
try:
kwargs[str(sym)] = sym.get()
except:
pass
return binaryobj(**kwargs)
def update_resolved_symbol(self, sym):
""" Notifies an array that a symbol has been resolved so that it
can be resized. """
self.resize(
[symbolic.eval(s, 0) for s in self.descriptor.shape],
refcheck=False)
self._symlist = symbolic.symlist(self.descriptor.shape)
def ndarray(shape, dtype=numpy.float64, *args, **kwargs):
""" Returns a numpy ndarray where all symbols have been evaluated to
numbers and types are converted to numpy types. """
repldict = {sym: sym.get() for sym in symbolic.symlist(shape).values()}
new_shape = [
int(s.subs(repldict) if symbolic.issymbolic(s) else s) for s in shape
]
new_dtype = dtype.type if isinstance(dtype, dtypes.typeclass) else dtype
return numpy.ndarray(shape=new_shape, dtype=new_dtype, *args, **kwargs)
def __new__(cls,
shape,
dtype=types.float32,
materialize_func=None,
allow_conflicts=False,
*args,
**kwargs):
""" Initializes a DaCe ND-array.
@param shape: The array shape (may contain symbols).
@param dtype: The array data type.
@param materialize_func: An optional string that contains a method
to materialize array contents on demand.
If not None, the array is not allocated
within the DaCe program.
@param allow_conflicts: If True, suppresses warnings on conflicting
array writes in DaCe programs without a
matching conflict resolution memlet.
"""
# Avoiding import loops
from dace import data
tmpshape = shape
shape = [symbolic.eval(s, 0) for s in shape]
kwargs.update({'dtype': dtype.type})
res = numpy.ndarray.__new__(cls, shape, *args, **kwargs)
res._symlist = symbolic.symlist(tmpshape)
for _, sym in res._symlist.items():
sym._arrays_to_update.append(res)
if not isinstance(dtype, types.typeclass):
dtype = types.typeclass(dtype.type)
res.descriptor = data.Array(
dtype,
tmpshape,
materialize_func=materialize_func,
transient=False,
allow_conflicts=allow_conflicts)
return res
def replace_func(element, dyn_syms, retparams):
# Resolve all symbols using the retparams-dict
for x in dyn_syms:
target = sp.functions.Min(retparams[x] * (retparams[x] - 1) / 2, 0)
bstr = str(element)
element = symbolic.pystr_to_symbolic(bstr)
element = element.subs(x, target) # Add the classic sum formula; going upwards
# To not have hidden elements that get added again later, we
# also replace the values in the other itvars...
for k, v in retparams.items():
newv = symbolic.pystr_to_symbolic(str(v))
tarsyms = symbolic.symlist(target).keys()
if x in tarsyms:
continue
tmp = newv.subs(x, target)
if tmp != v:
retparams[k] = tmp
return element
def expand(self, sdfg, graph, map_entries, map_base_variables=None):
"""
Expansion into outer and inner maps for each map in a specified set.
The resulting outer maps all have same range and indices, corresponding
variables and memlets get changed accordingly. The inner map contains
the leftover dimensions
:param sdfg: Underlying SDFG
:param graph: Graph in which we expand
:param map_entries: List of Map Entries(Type MapEntry) that we want to expand
:param map_base_variables: Optional parameter. List of strings
If None, then expand() searches for the maximal amount
of equal map ranges and pushes those and their corresponding
loop variables into the outer loop.
If specified, then expand() pushes the ranges belonging
to the loop iteration variables specified into the outer loop
(For instance map_base_variables = ['i','j'] assumes that
all maps have common iteration indices i and j with corresponding
correct ranges)
"""
maps = [entry.map for entry in map_entries]
if not map_base_variables:
# find the maximal subset of variables to expand
# greedy if there exist multiple ranges that are equal in a map
map_base_ranges = helpers.common_map_base_ranges(maps)
reassignments = helpers.find_reassignment(maps, map_base_ranges)
##### first, regroup and reassign
# create params_dict for every map
# first, let us define the outer iteration variable names,
# just take the first map and their indices at common ranges
map_base_variables = []
for rng in map_base_ranges:
for i in range(len(maps[0].params)):
if maps[0].range[i] == rng and maps[0].params[
i] not in map_base_variables:
map_base_variables.append(maps[0].params[i])
break
params_dict = {}
if self.debug:
print("MultiExpansion::Map_base_variables:", map_base_variables)
print("MultiExpansion::Map_base_ranges:", map_base_ranges)
for map in maps:
# for each map create param dict, first assign identity
params_dict_map = {param: param for param in map.params}
# now look for the correct reassignment
# for every element neq -1, need to change param to map_base_variables[]
# if param already appears in own dict, do a swap
# else we just replace it
for i, reassignment in enumerate(reassignments[map]):
if reassignment == -1:
# nothing to do
pass
else:
current_var = map.params[i]
current_assignment = params_dict_map[current_var]
target_assignment = map_base_variables[reassignment]
if current_assignment != target_assignment:
if target_assignment in params_dict_map.values():
# do a swap
key1 = current_var
for key, value in params_dict_map.items():
if value == target_assignment:
key2 = key
value1 = params_dict_map[key1]
value2 = params_dict_map[key2]
params_dict_map[key1] = key2
params_dict_map[key2] = key1
else:
# just reassign
params_dict_map[current_var] = target_assignment
# done, assign params_dict_map to the global one
params_dict[map] = params_dict_map
for map, map_entry in zip(maps, map_entries):
map_scope = graph.scope_subgraph(map_entry)
params_dict_map = params_dict[map]
for firstp, secondp in params_dict_map.items():
if firstp != secondp:
replace(map_scope, firstp, '__' + firstp + '_fused')
for firstp, secondp in params_dict_map.items():
if firstp != secondp:
replace(map_scope, '__' + firstp + '_fused', secondp)
# now also replace the map variables inside maps
for i in range(len(map.params)):
map.params[i] = params_dict_map[map.params[i]]
if self.debug:
print("MultiExpansion::Params replaced")
else:
# just calculate map_base_ranges
# do a check whether all maps correct
map_base_ranges = []
map0 = maps[0]
for var in map_base_variables:
index = map0.params.index(var)
map_base_ranges.append(map0.range[index])
for map in maps:
for var, rng in zip(map_base_variables, map_base_ranges):
assert map.range[map.params.index(var)] == rng
# then expand all the maps
for map, map_entry in zip(maps, map_entries):
if map.get_param_num() == len(map_base_variables):
# nothing to expand, continue
continue
map_exit = graph.exit_node(map_entry)
# create two new maps, outer and inner
params_outer = map_base_variables
ranges_outer = map_base_ranges
init_params_inner = []
init_ranges_inner = []
for param, rng in zip(map.params, map.range):
if param in map_base_variables:
continue
else:
init_params_inner.append(param)
init_ranges_inner.append(rng)
params_inner = init_params_inner
ranges_inner = subsets.Range(init_ranges_inner)
inner_map = nodes.Map(label = map.label + '_inner',
params = params_inner,
ndrange = ranges_inner,
schedule = dtypes.ScheduleType.Sequential \
if self.sequential_innermaps \
else dtypes.ScheduleType.Default)
map.label = map.label + '_outer'
map.params = params_outer
map.range = ranges_outer
# create new map entries and exits
map_entry_inner = nodes.MapEntry(inner_map)
map_exit_inner = nodes.MapExit(inner_map)
# analogously to Map_Expansion
for edge in graph.out_edges(map_entry):
graph.remove_edge(edge)
graph.add_memlet_path(map_entry,
map_entry_inner,
edge.dst,
src_conn=edge.src_conn,
memlet=edge.data,
dst_conn=edge.dst_conn)
dynamic_edges = dynamic_map_inputs(graph, map_entry)
for edge in dynamic_edges:
# Remove old edge and connector
graph.remove_edge(edge)
edge.dst._in_connectors.remove(edge.dst_conn)
# Propagate to each range it belongs to
path = []
for mapnode in [map_entry, map_entry_inner]:
path.append(mapnode)
if any(edge.dst_conn in map(str, symbolic.symlist(r))
for r in mapnode.map.range):
graph.add_memlet_path(edge.src,
*path,
memlet=edge.data,
src_conn=edge.src_conn,
dst_conn=edge.dst_conn)
for edge in graph.in_edges(map_exit):
graph.remove_edge(edge)
graph.add_memlet_path(edge.src,
map_exit_inner,
map_exit,
memlet=edge.data,
src_conn=edge.src_conn,
dst_conn=edge.dst_conn)
def apply(self, sdfg: dace.SDFG):
# Extract the map and its entry and exit nodes.
graph = sdfg.nodes()[self.state_id]
map_entry = graph.nodes()[self.subgraph[MapExpansion._map_entry]]
map_exit = graph.exit_node(map_entry)
current_map = map_entry.map
# Create new maps
new_maps = [
nodes.Map(current_map.label + '_' + str(param), [param],
subsets.Range([param_range]),
schedule=dtypes.ScheduleType.Sequential) for param,
param_range in zip(current_map.params[1:], current_map.range[1:])
]
current_map.params = [current_map.params[0]]
current_map.range = subsets.Range([current_map.range[0]])
# Create new map entries and exits
entries = [nodes.MapEntry(new_map) for new_map in new_maps]
exits = [nodes.MapExit(new_map) for new_map in new_maps]
# Create edges, abiding by the following rules:
# 1. If there are no edges coming from the outside, use empty memlets
# 2. Edges with IN_* connectors replicate along the maps
# 3. Edges for dynamic map ranges replicate until reaching range(s)
for edge in graph.out_edges(map_entry):
graph.remove_edge(edge)
graph.add_memlet_path(map_entry,
*entries,
edge.dst,
src_conn=edge.src_conn,
memlet=edge.data,
dst_conn=edge.dst_conn)
# Modify dynamic map ranges
dynamic_edges = dace.sdfg.dynamic_map_inputs(graph, map_entry)
for edge in dynamic_edges:
# Remove old edge and connector
graph.remove_edge(edge)
edge.dst.remove_in_connector(edge.dst_conn)
# Propagate to each range it belongs to
path = []
for mapnode in [map_entry] + entries:
path.append(mapnode)
if any(edge.dst_conn in map(str, symbolic.symlist(r))
for r in mapnode.map.range):
graph.add_memlet_path(edge.src,
*path,
memlet=edge.data,
src_conn=edge.src_conn,
dst_conn=edge.dst_conn)
# Create new map exits
for edge in graph.in_edges(map_exit):
graph.remove_edge(edge)
graph.add_memlet_path(edge.src,
*exits[::-1],
map_exit,
memlet=edge.data,
src_conn=edge.src_conn,
dst_conn=edge.dst_conn)
def infer_symbols_from_datadescriptor(sdfg: SDFG, args: Dict[str, Any],
exclude: Optional[Set[str]] = None) -> \
Dict[str, Any]:
"""
Infers the values of SDFG symbols (not given as arguments) from the shapes
and strides of input arguments (e.g., arrays).
:param sdfg: The SDFG that is being called.
:param args: A dictionary mapping from current argument names to their
values. This may also include symbols.
:param exclude: An optional set of symbols to ignore on inference.
:return: A dictionary mapping from symbol names that are not in ``args``
to their inferred values.
:raise ValueError: If symbol values are ambiguous.
"""
exclude = exclude or set()
exclude = set(symbolic.symbol(s) for s in exclude)
equations = []
symbols = set()
# Collect equations and symbols from arguments and shapes
for arg_name, arg_val in args.items():
if arg_name in sdfg.arrays:
desc = sdfg.arrays[arg_name]
if not hasattr(desc, 'shape') or not hasattr(arg_val, 'shape'):
continue
symbolic_values = list(desc.shape) + list(
getattr(desc, 'strides', []))
given_values = list(arg_val.shape)
given_strides = []
if hasattr(arg_val, 'strides'):
# NumPy arrays use bytes in strides
factor = getattr(arg_val, 'itemsize', 1)
given_strides = [s // factor for s in arg_val.strides]
given_values += given_strides
for sym_dim, real_dim in zip(symbolic_values, given_values):
repldict = {}
for sym in symbolic.symlist(sym_dim).values():
newsym = symbolic.symbol('__SOLVE_' + str(sym))
if str(sym) in args:
exclude.add(newsym)
else:
symbols.add(newsym)
exclude.add(sym)
repldict[sym] = newsym
# Replace symbols with __SOLVE_ symbols so as to allow
# the same symbol in the called SDFG
if repldict:
sym_dim = sym_dim.subs(repldict)
equations.append(sym_dim - real_dim)
if len(symbols) == 0:
return {}
# Solve for all at once
results = sympy.solve(equations, *symbols, dict=True, exclude=exclude)
if len(results) > 1:
raise ValueError('Ambiguous values for symbols in inference. '
'Options: %s' % str(results))
if len(results) == 0:
raise ValueError('Cannot infer values for symbols in inference.')
result = results[0]
if not result:
raise ValueError('Cannot infer values for symbols in inference.')
# Remove __SOLVE_ prefix
return {str(k)[8:]: v for k, v in result.items()}
def apply(self, sdfg: dace.SDFG):
# Extract the map and its entry and exit nodes.
graph = sdfg.node(self.state_id)
map_entry = self.map_entry(sdfg)
map_exit = graph.exit_node(map_entry)
current_map = map_entry.map
# Create new maps
new_maps = [
nodes.Map(current_map.label + '_' + str(param), [param],
subsets.Range([param_range]),
schedule=dtypes.ScheduleType.Sequential) for param,
param_range in zip(current_map.params[1:], current_map.range[1:])
]
current_map.params = [current_map.params[0]]
current_map.range = subsets.Range([current_map.range[0]])
# Create new map entries and exits
entries = [nodes.MapEntry(new_map) for new_map in new_maps]
exits = [nodes.MapExit(new_map) for new_map in new_maps]
# Create edges, abiding by the following rules:
# 1. If there are no edges coming from the outside, use empty memlets
# 2. Edges with IN_* connectors replicate along the maps
# 3. Edges for dynamic map ranges replicate until reaching range(s)
for edge in graph.out_edges(map_entry):
graph.remove_edge(edge)
graph.add_memlet_path(map_entry,
*entries,
edge.dst,
src_conn=edge.src_conn,
memlet=edge.data,
dst_conn=edge.dst_conn)
# Modify dynamic map ranges
dynamic_edges = dace.sdfg.dynamic_map_inputs(graph, map_entry)
for edge in dynamic_edges:
# Remove old edge and connector
graph.remove_edge(edge)
edge.dst.remove_in_connector(edge.dst_conn)
# Propagate to each range it belongs to
path = []
for mapnode in [map_entry] + entries:
path.append(mapnode)
if any(edge.dst_conn in map(str, symbolic.symlist(r))
for r in mapnode.map.range):
graph.add_memlet_path(edge.src,
*path,
memlet=edge.data,
src_conn=edge.src_conn,
dst_conn=edge.dst_conn)
# Create new map exits
for edge in graph.in_edges(map_exit):
graph.remove_edge(edge)
graph.add_memlet_path(edge.src,
*exits[::-1],
map_exit,
memlet=edge.data,
src_conn=edge.src_conn,
dst_conn=edge.dst_conn)
from dace.sdfg.scope import ScopeTree
scope = None
queue: List[ScopeTree] = graph.scope_leaves()
while len(queue) > 0:
tnode = queue.pop()
if tnode.entry == entries[-1]:
scope = tnode
break
elif tnode.parent is not None:
queue.append(tnode.parent)
else:
raise ValueError('Cannot find scope in state')
consolidate_edges(sdfg, scope)
return [map_entry] + entries
def infer_symbols_from_shapes(sdfg: SDFG, args: Dict[str, Any],
exclude: Optional[Set[str]] = None) -> \
Dict[str, Any]:
"""
Infers the values of SDFG symbols (not given as arguments) from the shapes
of input arguments (e.g., arrays).
:param sdfg: The SDFG that is being called.
:param args: A dictionary mapping from current argument names to their
values. This may also include symbols.
:param exclude: An optional set of symbols to ignore on inference.
:return: A dictionary mapping from symbol names that are not in ``args``
to their inferred values.
:raise ValueError: If symbol values are ambiguous.
"""
exclude = exclude or set()
exclude = set(symbolic.symbol(s) for s in exclude)
equations = []
symbols = set()
# Collect equations and symbols from arguments and shapes
for arg_name, arg_val in args.items():
if arg_name in sdfg.arrays:
desc = sdfg.arrays[arg_name]
if not hasattr(desc, 'shape') or not hasattr(arg_val, 'shape'):
continue
symbolic_shape = desc.shape
given_shape = arg_val.shape
for sym_dim, real_dim in zip(symbolic_shape, given_shape):
repldict = {}
for sym in symbolic.symlist(sym_dim).values():
newsym = symbolic.symbol('__SOLVE_' + str(sym))
if str(sym) in args:
exclude.add(newsym)
else:
symbols.add(newsym)
exclude.add(sym)
repldict[sym] = newsym
# Replace symbols with __SOLVE_ symbols so as to allow
# the same symbol in the called SDFG
if repldict:
sym_dim = sym_dim.subs(repldict)
equations.append(sym_dim - real_dim)
if len(symbols) == 0:
return {}
# Solve for all at once
results = sympy.solve(equations, *symbols, dict=True, exclude=exclude)
if len(results) > 1:
raise ValueError('Ambiguous values for symbols in inference. '
'Options: %s' % str(results))
if len(results) == 0:
raise ValueError('Cannot infer values for symbols in inference.')
result = results[0]
if not result:
raise ValueError('Cannot infer values for symbols in inference.')
# Fast path (unnecessary)
# # For each symbol in each dimension, try to solve an equation
# for sym_dim, real_dim in zip(symbolic_shape, given_shape):
# for sym in symbolic.symlist(sym_dim):
# if sym in inferred_syms and symval != inferred_syms[sym]:
# raise ValueError('Ambiguous value for symbol %s in argument '
# '%s: can be either %d or %d' % (
# sym, arg_name, inferred_syms[sym], symval))
# Remove __SOLVE_ prefix
return {str(k)[8:]: v for k, v in result.items()}
def apply(self, sdfg):
graph = sdfg.nodes()[self.state_id]
subgraph = self.subgraph_view(sdfg)
map_entries = helpers.get_outermost_scope_maps(sdfg, graph, subgraph)
result = StencilTiling.topology(sdfg, graph, map_entries)
(children_dict, parent_dict, sink_maps) = result
# next up, calculate inferred ranges for each map
# for each map entry, this contains a tuple of dicts:
# each of those maps from data_name of the array to
# inferred outer ranges. An inferred outer range is created
# by taking the union of ranges of inner subsets corresponding
# to that data and substituting this subset by the min / max of the
# parametrized map boundaries
# finally, from these outer ranges we can easily calculate
# strides and tile sizes required for every map
inferred_ranges = defaultdict(dict)
# create array of reverse topologically sorted map entries
# to iterate over
topo_reversed = []
queue = set(sink_maps.copy())
while len(queue) > 0:
element = next(e for e in queue
if not children_dict[e] - set(topo_reversed))
topo_reversed.append(element)
queue.remove(element)
for parent in parent_dict[element]:
queue.add(parent)
# main loop
# first get coverage dicts for each map entry
# for each map, contains a tuple of two dicts
# each of those two maps from data name to outer range
coverage = {}
for map_entry in map_entries:
coverage[map_entry] = StencilTiling.coverage_dicts(
sdfg, graph, map_entry, outer_range=True)
# we have a mapping from data name to outer range
# however we want a mapping from map parameters to outer ranges
# for this we need to find out how all array dimensions map to
# outer ranges
variable_mapping = defaultdict(list)
for map_entry in topo_reversed:
map = map_entry.map
# first find out variable mapping
for e in itertools.chain(
graph.out_edges(map_entry),
graph.in_edges(graph.exit_node(map_entry))):
mapping = []
for dim in e.data.subset:
syms = set()
for d in dim:
syms |= symbolic.symlist(d).keys()
if len(syms) > 1:
raise NotImplementedError(
"One incoming or outgoing stencil subset is indexed "
"by multiple map parameters. "
"This is not supported yet.")
try:
mapping.append(syms.pop())
except KeyError:
# just append None if there is no map symbol in it.
# we don't care for now.
mapping.append(None)
if e.data in variable_mapping:
# assert that this is the same everywhere.
# else we might run into problems
assert variable_mapping[e.data.data] == mapping
else:
variable_mapping[e.data.data] = mapping
# now do mapping data name -> outer range
# and from that infer mapping variable -> outer range
local_ranges = {dn: None for dn in coverage[map_entry][1].keys()}
for data_name, cov in coverage[map_entry][1].items():
local_ranges[data_name] = subsets.union(
local_ranges[data_name], cov)
# now look at proceeding maps
# and union those subsets -> could be larger with stencil indent
for child_map in children_dict[map_entry]:
if data_name in coverage[child_map][0]:
local_ranges[data_name] = subsets.union(
local_ranges[data_name],
coverage[child_map][0][data_name])
# final assignent: combine local_ranges and variable_mapping
# together into inferred_ranges
inferred_ranges[map_entry] = {p: None for p in map.params}
for data_name, ranges in local_ranges.items():
for param, r in zip(variable_mapping[data_name], ranges):
# create new range from this subset and assign
rng = subsets.Range((r, ))
if param:
inferred_ranges[map_entry][param] = subsets.union(
inferred_ranges[map_entry][param], rng)
# get parameters -- should all be the same
params = next(iter(map_entries)).map.params.copy()
# define reference range as inferred range of one of the sink maps
self.reference_range = inferred_ranges[next(iter(sink_maps))]
if self.debug:
print("StencilTiling::Reference Range", self.reference_range)
# next up, search for the ranges that don't change
invariant_dims = []
for idx, p in enumerate(params):
different = False
if self.reference_range[p] is None:
invariant_dims.append(idx)
warnings.warn(
f"StencilTiling::No Stencil pattern detected for parameter {p}"
)
continue
for m in map_entries:
if inferred_ranges[m][p] != self.reference_range[p]:
different = True
break
if not different:
invariant_dims.append(idx)
warnings.warn(
f"StencilTiling::No Stencil pattern detected for parameter {p}"
)
# during stripmining, we will create new outer map entries
# for easy access
self._outer_entries = set()
# with inferred_ranges constructed, we can begin to strip mine
for map_entry in map_entries:
# Retrieve map entry and exit nodes.
map = map_entry.map
stripmine_subgraph = {
StripMining._map_entry: graph.nodes().index(map_entry)
}
sdfg_id = sdfg.sdfg_id
last_map_entry = None
original_schedule = map_entry.schedule
self.tile_sizes = []
self.tile_offset_lower = []
self.tile_offset_upper = []
# strip mining each dimension where necessary
removed_maps = 0
for dim_idx, param in enumerate(map_entry.map.params):
# get current_node tile size
if dim_idx >= len(self.strides):
tile_stride = symbolic.pystr_to_symbolic(self.strides[-1])
else:
tile_stride = symbolic.pystr_to_symbolic(
self.strides[dim_idx])
trivial = False
if dim_idx in invariant_dims:
self.tile_sizes.append(tile_stride)
self.tile_offset_lower.append(0)
self.tile_offset_upper.append(0)
else:
target_range_current = inferred_ranges[map_entry][param]
reference_range_current = self.reference_range[param]
min_diff = symbolic.SymExpr(reference_range_current.min_element()[0] \
- target_range_current.min_element()[0])
max_diff = symbolic.SymExpr(target_range_current.max_element()[0] \
- reference_range_current.max_element()[0])
try:
min_diff = symbolic.evaluate(min_diff, {})
max_diff = symbolic.evaluate(max_diff, {})
except TypeError:
raise RuntimeError("Symbolic evaluation of map "
"ranges failed. Please check "
"your parameters and match.")
self.tile_sizes.append(tile_stride + max_diff + min_diff)
self.tile_offset_lower.append(
symbolic.pystr_to_symbolic(str(min_diff)))
self.tile_offset_upper.append(
symbolic.pystr_to_symbolic(str(max_diff)))
# get calculated parameters
tile_size = self.tile_sizes[-1]
dim_idx -= removed_maps
# If map or tile sizes are trivial, skip strip-mining map dimension
# special cases:
# if tile size is trivial AND we have an invariant dimension, skip
if tile_size == map.range.size()[dim_idx] and (
dim_idx + removed_maps) in invariant_dims:
continue
# trivial map: we just continue
if map.range.size()[dim_idx] in [0, 1]:
continue
if tile_size == 1 and tile_stride == 1 and (
dim_idx + removed_maps) in invariant_dims:
trivial = True
removed_maps += 1
# indent all map ranges accordingly and then perform
# strip mining on these. Offset inner maps accordingly afterwards
range_tuple = (map.range[dim_idx][0] +
self.tile_offset_lower[-1],
map.range[dim_idx][1] -
self.tile_offset_upper[-1],
map.range[dim_idx][2])
map.range[dim_idx] = range_tuple
stripmine = StripMining(sdfg_id, self.state_id,
stripmine_subgraph, 0)
stripmine.tiling_type = 'ceilrange'
stripmine.dim_idx = dim_idx
stripmine.new_dim_prefix = self.prefix if not trivial else ''
# use tile_stride for both -- we will extend
# the inner tiles later
stripmine.tile_size = str(tile_stride)
stripmine.tile_stride = str(tile_stride)
outer_map = stripmine.apply(sdfg)
outer_map.schedule = original_schedule
# apply to the new map the schedule of the original one
map_entry.schedule = self.schedule
# if tile stride is 1, we can make a nice simplification by just
# taking the overapproximated inner range as inner range
# this eliminates the min/max in the range which
# enables loop unrolling
if tile_stride == 1:
map_entry.range[dim_idx] = tuple(
symbolic.SymExpr(el._approx_expr) if isinstance(
el, symbolic.SymExpr) else el
for el in map_entry.range[dim_idx])
# in map_entry: enlarge tiles by upper and lower offset
# doing it this way and not via stripmine strides ensures
# that the max gets changed as well
old_range = map_entry.range[dim_idx]
map_entry.range[dim_idx] = ((old_range[0] -
self.tile_offset_lower[-1]),
(old_range[1] +
self.tile_offset_upper[-1]),
old_range[2])
# We have to propagate here for correct outer volume and subset sizes
_propagate_node(graph, map_entry)
_propagate_node(graph, graph.exit_node(map_entry))
# usual tiling pipeline
if last_map_entry:
new_map_entry = graph.in_edges(map_entry)[0].src
mapcollapse_subgraph = {
MapCollapse._outer_map_entry:
graph.node_id(last_map_entry),
MapCollapse._inner_map_entry:
graph.node_id(new_map_entry)
}
mapcollapse = MapCollapse(sdfg_id, self.state_id,
mapcollapse_subgraph, 0)
mapcollapse.apply(sdfg)
last_map_entry = graph.in_edges(map_entry)[0].src
# add last instance of map entries to _outer_entries
if last_map_entry:
self._outer_entries.add(last_map_entry)
# Map Unroll Feature: only unroll if conditions are met:
# Only unroll if at least one of the inner map ranges is strictly larger than 1
# Only unroll if strides all are one
if self.unroll_loops and all(s == 1 for s in self.strides) and any(
s not in [0, 1] for s in map_entry.range.size()):
l = len(map_entry.params)
if l > 1:
subgraph = {
MapExpansion.map_entry: graph.nodes().index(map_entry)
}
trafo_expansion = MapExpansion(sdfg.sdfg_id,
sdfg.nodes().index(graph),
subgraph, 0)
trafo_expansion.apply(sdfg)
maps = [map_entry]
for _ in range(l - 1):
map_entry = graph.out_edges(map_entry)[0].dst
maps.append(map_entry)
for map in reversed(maps):
# MapToForLoop
subgraph = {
MapToForLoop._map_entry: graph.nodes().index(map)
}
trafo_for_loop = MapToForLoop(sdfg.sdfg_id,
sdfg.nodes().index(graph),
subgraph, 0)
trafo_for_loop.apply(sdfg)
nsdfg = trafo_for_loop.nsdfg
# LoopUnroll
guard = trafo_for_loop.guard
end = trafo_for_loop.after_state
begin = next(e.dst for e in nsdfg.out_edges(guard)
if e.dst != end)
subgraph = {
DetectLoop._loop_guard: nsdfg.nodes().index(guard),
DetectLoop._loop_begin: nsdfg.nodes().index(begin),
DetectLoop._exit_state: nsdfg.nodes().index(end)
}
transformation = LoopUnroll(0, 0, subgraph, 0)
transformation.apply(nsdfg)
elif self.unroll_loops:
warnings.warn(
"Did not unroll loops. Either all ranges are equal to "
"one or range difference is symbolic.")
self._outer_entries = list(self._outer_entries)
def can_be_applied(sdfg, subgraph) -> bool:
# get highest scope maps
graph = subgraph.graph
map_entries = set(
helpers.get_outermost_scope_maps(sdfg, graph, subgraph))
# 1.1: There has to be more than one outermost scope map entry
if len(map_entries) <= 1:
return False
# 1.2: check basic constraints:
# - all parameters have to be the same (this implies same length)
# - no parameter permutations here as ambiguity is very high then
# - same strides everywhere
first_map = next(iter(map_entries))
params = dcpy(first_map.map.params)
strides = first_map.map.range.strides()
schedule = first_map.map.schedule
for map_entry in map_entries:
if map_entry.map.params != params:
return False
if map_entry.map.range.strides() != strides:
return False
if map_entry.map.schedule != schedule:
return False
# 1.3: check whether all map entries only differ by a const amount
first_entry = next(iter(map_entries))
for map_entry in map_entries:
for r1, r2 in zip(map_entry.map.range, first_entry.map.range):
if len((r1[0] - r2[0]).free_symbols) > 0:
return False
if len((r1[1] - r2[1]).free_symbols) > 0:
return False
# get intermediate_nodes, out_nodes from SubgraphFusion Transformation
node_config = SubgraphFusion.get_adjacent_nodes(
sdfg, graph, map_entries)
(_, intermediate_nodes, out_nodes) = node_config
# 1.4: check topological feasibility
if not SubgraphFusion.check_topo_feasibility(
sdfg, graph, map_entries, intermediate_nodes, out_nodes):
return False
# 1.5 nodes that are both intermediate and out nodes
# are not supported in StencilTiling
if len(intermediate_nodes & out_nodes) > 0:
return False
# get coverages for every map entry
coverages = {}
memlets = {}
for map_entry in map_entries:
coverages[map_entry] = StencilTiling.coverage_dicts(
sdfg, graph, map_entry)
memlets[map_entry] = StencilTiling.coverage_dicts(
sdfg, graph, map_entry, outer_range=False)
# get DAG neighbours for each map
dag_neighbors = StencilTiling.topology(sdfg, graph, map_entries)
(children_dict, _, sink_maps) = dag_neighbors
# 1.6: we now check coverage:
# each outgoing coverage for a data memlet has to
# be exactly equal to the union of incoming coverages
# of all chidlren map memlets of this data
# important:
# 1. it has to be equal and not only cover it in order to
# account for ranges too long
# 2. we check coverages by map parameter and not by
# array, this way it is even more general
# 3. map parameter coverages are checked for each
# (map_entry, children of this map_entry) - pair
for map_entry in map_entries:
# get coverage from current map_entry
map_coverage = coverages[map_entry][1]
# final mapping map_parameter -> coverage will be stored here
param_parent_coverage = {p: None for p in map_entry.params}
param_children_coverage = {p: None for p in map_entry.params}
for child_entry in children_dict[map_entry]:
# get mapping data_name -> coverage
for (data_name, cov) in map_coverage.items():
parent_coverage = cov
children_coverage = None
if data_name in coverages[child_entry][0]:
children_coverage = subsets.union(
children_coverage,
coverages[child_entry][0][data_name])
# extend mapping map_parameter -> coverage
# by the previous mapping
for i, (p_subset, c_subset) in enumerate(
zip(parent_coverage, children_coverage)):
# transform into subset
p_subset = subsets.Range((p_subset, ))
c_subset = subsets.Range((c_subset, ))
# get associated parameter in memlet
params1 = symbolic.symlist(
memlets[map_entry][1][data_name][i]).keys()
params2 = symbolic.symlist(
memlets[child_entry][0][data_name][i]).keys()
if params1 != params2:
return False
params = params1
if len(params) > 1:
# this is not supported
return False
try:
symbol = next(iter(params))
param_parent_coverage[symbol] = subsets.union(
param_parent_coverage[symbol], p_subset)
param_children_coverage[symbol] = subsets.union(
param_children_coverage[symbol], c_subset)
except StopIteration:
# current dim has no symbol associated.
# ignore and continue
warnings.warn(
f"In map {map_entry}, there is a "
"dimension belonging to {data_name} "
"that has no map parameter associated.")
pass
# 1.6: parameter mapping must be the same
if param_parent_coverage != param_children_coverage:
return False
# 1.7: we want all sink maps to have the same range size
assert len(sink_maps) > 0
first_sink_map = next(iter(sink_maps))
if not all([
map.range.size() == first_sink_map.range.size()
for map in sink_maps
]):
return False
return True
def free_symbols(self) -> Set[str]:
result = set()
for dim in self.ranges:
for d in dim:
result |= symbolic.symlist(d).keys()
return result
def free_symbols(self) -> Set[str]:
result = set()
for dim in self.indices:
result |= symbolic.symlist(dim).keys()
return result
def free_symbols(self):
result = {}
for dim in self.ranges:
for d in dim:
result.update(symbolic.symlist(d))
return result
def reduce_iteration_count(begin, end, step, rparams: dict):
# There are different rules when expanding depending on where the expand
# should happen
if isinstance(begin, int):
start_syms = []
else:
start_syms = symbolic.symlist(begin).keys()
if isinstance(end, int):
end_syms = []
else:
end_syms = symbolic.symlist(end).keys()
if isinstance(step, int):
step_syms = []
else:
step_syms = symbolic.symlist(step).keys()
def intersection(lista, listb):
return [x for x in lista if x in listb]
start_dyn_syms = intersection(start_syms, rparams.keys())
end_dyn_syms = intersection(end_syms, rparams.keys())
step_dyn_syms = intersection(step_syms, rparams.keys())
def replace_func(element, dyn_syms, retparams):
# Resolve all symbols using the retparams-dict
for x in dyn_syms:
target = sp.functions.Min(
retparams[x] * (retparams[x] - 1) / 2, 0)
bstr = str(element)
element = symbolic.pystr_to_symbolic(bstr)
element = element.subs(
x, target) # Add the classic sum formula; going upwards
# To not have hidden elements that get added again later, we
# also replace the values in the other itvars...
for k, v in retparams.items():
newv = symbolic.pystr_to_symbolic(str(v))
tarsyms = symbolic.symbols_in_sympy_expr(target).keys()
if x in tarsyms:
continue
tmp = newv.subs(x, target)
if tmp != v:
retparams[k] = tmp
return element
if len(start_dyn_syms) > 0:
pass
begin = replace_func(begin, start_dyn_syms, rparams)
if len(end_dyn_syms) > 0:
pass
end = replace_func(end, end_dyn_syms, rparams)
if len(step_dyn_syms) > 0:
pass
print("Dynamic step symbols %s!" % str(step))
raise NotImplementedError
return begin, end, step
def free_symbols(self):
result = {}
for dim in self.indices:
result.update(symbolic.symlist(dim))
return result