def _create_strided_range(self, sdfg: SDFG, state: SDFGState,
map_entry: nodes.MapEntry):
map_exit = state.exit_node(map_entry)
dim_idx = self.dim_idx
new_dim_prefix = self.new_dim_prefix
tile_size = self.tile_size
divides_evenly = self.divides_evenly
tile_stride = self.tile_stride
if tile_stride == 0:
tile_stride = tile_size
if tile_stride != tile_size:
raise NotImplementedError
# Retrieve parameter and range of dimension to be strip-mined.
target_dim = map_entry.map.params[dim_idx]
td_from, td_to, td_step = map_entry.map.range[dim_idx]
new_dim = self._find_new_dim(sdfg, state, map_entry, new_dim_prefix,
target_dim)
new_dim_range = (td_from, td_to, tile_size)
new_map = nodes.Map(map_entry.map.label, [new_dim],
subsets.Range([new_dim_range]))
dimsym = dace.symbolic.pystr_to_symbolic(new_dim)
td_from_new = dimsym
if divides_evenly:
td_to_new = dimsym + tile_size - 1
else:
if isinstance(td_to, dace.symbolic.SymExpr):
td_to = td_to.expr
td_to_new = dace.symbolic.SymExpr(
sympy.Min(dimsym + tile_size - 1, td_to),
dimsym + tile_size - 1)
td_step_new = td_step
return new_dim, new_map, (td_from_new, td_to_new, td_step_new)
class RedundantArrayCopying3(pm.Transformation):
""" Implements the redundant array removal transformation. Removes multiples
of array B in pattern MapEntry -> B.
"""
_arrays_removed = 0
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
_out_array = nodes.AccessNode("_")
@staticmethod
def expressions():
return [
sdutil.node_path_graph(RedundantArrayCopying3._map_entry,
RedundantArrayCopying3._out_array)
]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
map_entry = graph.nodes()[candidate[RedundantArrayCopying3._map_entry]]
out_array = graph.nodes()[candidate[RedundantArrayCopying3._out_array]]
# Ensure out degree is one (only one target, which is out_array)
found = 0
for _, _, dst, _, _ in graph.out_edges(map_entry):
if (isinstance(dst, nodes.AccessNode) and dst != out_array
and dst.data == out_array.data):
found += 1
return found > 0
@staticmethod
def match_to_str(graph, candidate):
out_array = graph.nodes()[candidate[RedundantArrayCopying3._out_array]]
return "Remove " + str(out_array)
def apply(self, sdfg):
def gnode(nname):
return graph.nodes()[self.subgraph[nname]]
graph = sdfg.nodes()[self.state_id]
map_entry = gnode(RedundantArrayCopying3._map_entry)
out_array = gnode(RedundantArrayCopying3._out_array)
for e1 in graph.out_edges(map_entry):
dst = e1.dst
if (isinstance(dst, nodes.AccessNode) and dst != out_array
and dst.data == out_array.data):
for e2 in graph.out_edges(dst):
graph.add_edge(out_array, None, e2.dst, e2.dst_conn,
e2.data)
graph.remove_edge(e2)
graph.remove_edge(e1)
graph.remove_node(dst)
if Config.get_bool("debugprint"):
RedundantArrayCopying3._arrays_removed += 1
class TrivialMapRangeElimination(transformation.Transformation):
""" Implements the Trivial Map Range Elimination pattern.
Trivial Map Range Elimination takes a multi-dimensional map with
a range containing one element and removes the corresponding dimension.
Example: Map[i=0:I,j=0] -> Map[i=0:I]
"""
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
@staticmethod
def expressions():
return [sdutil.node_path_graph(TrivialMapRangeElimination._map_entry)]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
map_entry = graph.nodes()[candidate[
TrivialMapRangeElimination._map_entry]]
if len(map_entry.map.range) <= 1:
return False # only acts on multi-dimensional maps
return any(frm == to for frm, to, _ in map_entry.map.range)
@staticmethod
def match_to_str(graph, candidate):
map_entry = graph.nodes()[candidate[
TrivialMapRangeElimination._map_entry]]
return map_entry.map.label + ': ' + str(map_entry.map.params)
def apply(self, sdfg):
graph = sdfg.nodes()[self.state_id]
map_entry = graph.nodes()[self.subgraph[
TrivialMapRangeElimination._map_entry]]
remaining_ranges = []
remaining_params = []
for map_param, ranges in zip(map_entry.map.params,
map_entry.map.range.ranges):
map_from, map_to, _ = ranges
if map_from == map_to:
# Replace the map index variable with the value it obtained
scope = graph.scope_subgraph(map_entry)
scope.replace(map_param, map_from)
else:
remaining_ranges.append(ranges)
remaining_params.append(map_param)
map_entry.map.range.ranges = remaining_ranges
map_entry.map.params = remaining_params
def _create_from_tile_numbers(self, sdfg: SDFG, state: SDFGState,
map_entry: nodes.MapEntry):
map_exit = state.exit_node(map_entry)
# Retrieve transformation properties.
dim_idx = self.dim_idx
new_dim_prefix = self.new_dim_prefix
divides_evenly = self.divides_evenly
number_of_tiles = self.tile_size
tile_stride = self.tile_stride
number_of_tiles = dace.symbolic.pystr_to_symbolic(number_of_tiles)
# Retrieve parameter and range of dimension to be strip-mined.
target_dim = map_entry.map.params[dim_idx]
td_from, td_to, td_step = map_entry.map.range[dim_idx]
tile_size = map_entry.map.range.size_exact()[dim_idx] / number_of_tiles
if tile_stride == 0:
tile_stride = tile_size
if tile_stride != tile_size:
raise NotImplementedError
new_dim = self._find_new_dim(sdfg, state, map_entry, new_dim_prefix,
target_dim)
new_dim_range = (td_from, number_of_tiles - 1, 1)
new_map = nodes.Map(map_entry.map.label, [new_dim],
subsets.Range([new_dim_range]))
dimsym = dace.symbolic.pystr_to_symbolic(new_dim)
td_from_new = dimsym * tile_size
if divides_evenly:
td_to_new = (dimsym + 1) * tile_size - 1
else:
if isinstance(td_to, dace.symbolic.SymExpr):
td_to = td_to.expr
td_to_new = dace.symbolic.SymExpr(
sympy.Min((dimsym + 1) * tile_size - 1, td_to),
(dimsym + 1) * tile_size - 1)
td_step_new = td_step
return new_dim, new_map, (td_from_new, td_to_new, td_step_new)
class MapReduceFusion(pm.Transformation):
""" Implements the map-reduce-fusion transformation.
Fuses a map with an immediately following reduction, where the array
between the map and the reduction is not used anywhere else.
"""
no_init = Property(dtype=bool,
default=False,
desc='If enabled, does not create initialization states '
'for reduce nodes with identity')
_tasklet = nodes.Tasklet('_')
_tmap_exit = nodes.MapExit(nodes.Map("", [], []))
_in_array = nodes.AccessNode('_')
import dace.libraries.standard as stdlib # Avoid import loop
_reduce = stdlib.Reduce()
_out_array = nodes.AccessNode('_')
@staticmethod
def expressions():
return [
sdutil.node_path_graph(MapReduceFusion._tasklet,
MapReduceFusion._tmap_exit,
MapReduceFusion._in_array,
MapReduceFusion._reduce,
MapReduceFusion._out_array)
]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
tmap_exit = graph.nodes()[candidate[MapReduceFusion._tmap_exit]]
in_array = graph.nodes()[candidate[MapReduceFusion._in_array]]
reduce_node = graph.nodes()[candidate[MapReduceFusion._reduce]]
tasklet = graph.nodes()[candidate[MapReduceFusion._tasklet]]
# Make sure that the array is only accessed by the map and the reduce
if any([
src != tmap_exit
for src, _, _, _, memlet in graph.in_edges(in_array)
]):
return False
if any([
dest != reduce_node
for _, _, dest, _, memlet in graph.out_edges(in_array)
]):
return False
tmem = next(e for e in graph.edges_between(tasklet, tmap_exit)
if e.data.data == in_array.data).data
# (strict) Make sure that the transient is not accessed anywhere else
# in this state or other states
if strict and (len([
n for n in graph.nodes()
if isinstance(n, nodes.AccessNode) and n.data == in_array.data
]) > 1 or in_array.data in sdfg.shared_transients()):
return False
# If memlet already has WCR and it is different from reduce node,
# do not match
if tmem.wcr is not None and tmem.wcr != reduce_node.wcr:
return False
# Verify that reduction ranges match tasklet map
tout_memlet = graph.in_edges(in_array)[0].data
rin_memlet = graph.out_edges(in_array)[0].data
if tout_memlet.subset != rin_memlet.subset:
return False
return True
@staticmethod
def match_to_str(graph, candidate):
tasklet = candidate[MapReduceFusion._tasklet]
map_exit = candidate[MapReduceFusion._tmap_exit]
reduce = candidate[MapReduceFusion._reduce]
return ' -> '.join(str(node) for node in [tasklet, map_exit, reduce])
def apply(self, sdfg: SDFG):
graph = sdfg.nodes()[self.state_id]
tmap_exit = graph.nodes()[self.subgraph[MapReduceFusion._tmap_exit]]
in_array = graph.nodes()[self.subgraph[MapReduceFusion._in_array]]
reduce_node = graph.nodes()[self.subgraph[MapReduceFusion._reduce]]
out_array = graph.nodes()[self.subgraph[MapReduceFusion._out_array]]
# Set nodes to remove according to the expression index
nodes_to_remove = [in_array]
nodes_to_remove.append(reduce_node)
memlet_edge = None
for edge in graph.in_edges(tmap_exit):
if edge.data.data == in_array.data:
memlet_edge = edge
break
if memlet_edge is None:
raise RuntimeError('Reduction memlet cannot be None')
# Find which indices should be removed from new memlet
input_edge = graph.in_edges(reduce_node)[0]
axes = reduce_node.axes or list(range(len(input_edge.data.subset)))
array_edge = graph.out_edges(reduce_node)[0]
# Delete relevant edges and nodes
graph.remove_nodes_from(nodes_to_remove)
# Filter out reduced dimensions from subset
filtered_subset = [
dim for i, dim in enumerate(memlet_edge.data.subset)
if i not in axes
]
if len(filtered_subset) == 0: # Output is a scalar
filtered_subset = [(0, 0, 1)]
# Modify edge from tasklet to map exit
memlet_edge.data.data = out_array.data
memlet_edge.data.wcr = reduce_node.wcr
memlet_edge.data.subset = type(memlet_edge.data.subset)(filtered_subset)
# Add edge from map exit to output array
graph.add_edge(
memlet_edge.dst, 'OUT_' + memlet_edge.dst_conn[3:], array_edge.dst,
array_edge.dst_conn,
Memlet.simple(array_edge.data.data,
array_edge.data.subset,
num_accesses=array_edge.data.num_accesses,
wcr_str=reduce_node.wcr))
# Add initialization state as necessary
if reduce_node.identity is not None:
init_state = sdfg.add_state_before(graph)
init_state.add_mapped_tasklet(
'freduce_init',
[('o%d' % i, '%s:%s:%s' % (r[0], r[1] + 1, r[2]))
for i, r in enumerate(array_edge.data.subset)], {},
'out = %s' % reduce_node.identity, {
'out':
Memlet.simple(
array_edge.data.data, ','.join([
'o%d' % i
for i in range(len(array_edge.data.subset))
]))
},
external_edges=True)
class MapWCRFusion(pm.Transformation):
""" Implements the map expanded-reduce fusion transformation.
Fuses a map with an immediately following reduction, where the array
between the map and the reduction is not used anywhere else, and the
reduction is divided to two maps with a WCR, denoting partial reduction.
"""
_tasklet = nodes.Tasklet('_')
_tmap_exit = nodes.MapExit(nodes.Map("", [], []))
_in_array = nodes.AccessNode('_')
_rmap_in_entry = nodes.MapEntry(nodes.Map("", [], []))
_rmap_in_tasklet = nodes.Tasklet('_')
_rmap_in_cr = nodes.MapExit(nodes.Map("", [], []))
_rmap_out_entry = nodes.MapEntry(nodes.Map("", [], []))
_rmap_out_exit = nodes.MapExit(nodes.Map("", [], []))
_out_array = nodes.AccessNode('_')
@staticmethod
def expressions():
return [
# Map, then partial reduction of axes
sdutil.node_path_graph(
MapWCRFusion._tasklet, MapWCRFusion._tmap_exit,
MapWCRFusion._in_array, MapWCRFusion._rmap_out_entry,
MapWCRFusion._rmap_in_entry, MapWCRFusion._rmap_in_tasklet,
MapWCRFusion._rmap_in_cr, MapWCRFusion._rmap_out_exit,
MapWCRFusion._out_array)
]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
tmap_exit = graph.nodes()[candidate[MapWCRFusion._tmap_exit]]
in_array = graph.nodes()[candidate[MapWCRFusion._in_array]]
rmap_entry = graph.nodes()[candidate[MapWCRFusion._rmap_out_entry]]
# Make sure that the array is only accessed by the map and the reduce
if any([
src != tmap_exit
for src, _, _, _, memlet in graph.in_edges(in_array)
]):
return False
if any([
dest != rmap_entry
for _, _, dest, _, memlet in graph.out_edges(in_array)
]):
return False
# Make sure that there is a reduction in the second map
rmap_cr = graph.nodes()[candidate[MapWCRFusion._rmap_in_cr]]
reduce_edge = graph.in_edges(rmap_cr)[0]
if reduce_edge.data.wcr is None:
return False
# (strict) Make sure that the transient is not accessed anywhere else
# in this state or other states
if strict and (len([
n for n in graph.nodes()
if isinstance(n, nodes.AccessNode) and n.data == in_array.data
]) > 1 or in_array.data in sdfg.shared_transients()):
return False
# Verify that reduction ranges match tasklet map
tout_memlet = graph.in_edges(in_array)[0].data
rin_memlet = graph.out_edges(in_array)[0].data
if tout_memlet.subset != rin_memlet.subset:
return False
return True
@staticmethod
def match_to_str(graph, candidate):
tasklet = candidate[MapWCRFusion._tasklet]
map_exit = candidate[MapWCRFusion._tmap_exit]
reduce = candidate[MapWCRFusion._rmap_in_cr]
return ' -> '.join(str(node) for node in [tasklet, map_exit, reduce])
def apply(self, sdfg):
graph = sdfg.node(self.state_id)
# To apply, collapse the second map and then fuse the two resulting maps
map_collapse = MapCollapse(
self.sdfg_id, self.state_id, {
MapCollapse._outer_map_entry:
self.subgraph[MapWCRFusion._rmap_out_entry],
MapCollapse._inner_map_entry:
self.subgraph[MapWCRFusion._rmap_in_entry]
}, 0)
map_entry, _ = map_collapse.apply(sdfg)
map_fusion = MapFusion(
self.sdfg_id, self.state_id, {
MapFusion._first_map_exit:
self.subgraph[MapWCRFusion._tmap_exit],
MapFusion._second_map_entry: graph.node_id(map_entry)
}, 0)
map_fusion.apply(sdfg)
class AccumulateTransient(transformation.Transformation):
""" Implements the AccumulateTransient transformation, which adds
transient stream and data nodes between nested maps that lead to a
stream. The transient data nodes then act as a local accumulator.
"""
_map_exit = nodes.MapExit(nodes.Map("", [], []))
_outer_map_exit = nodes.MapExit(nodes.Map("", [], []))
array = Property(
dtype=str,
desc="Array to create local storage for (if empty, first available)",
default=None,
allow_none=True)
identity = SymbolicProperty(desc="Identity value to set",
default=None,
allow_none=True)
@staticmethod
def expressions():
return [
sdutil.node_path_graph(AccumulateTransient._map_exit,
AccumulateTransient._outer_map_exit)
]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
map_exit = graph.nodes()[candidate[AccumulateTransient._map_exit]]
outer_map_exit = graph.nodes()[candidate[
AccumulateTransient._outer_map_exit]]
# Check if there is an accumulation output
for e in graph.edges_between(map_exit, outer_map_exit):
if e.data.wcr is not None:
return True
return False
@staticmethod
def match_to_str(graph, candidate):
map_exit = candidate[AccumulateTransient._map_exit]
outer_map_exit = candidate[AccumulateTransient._outer_map_exit]
return ' -> '.join(str(node) for node in [map_exit, outer_map_exit])
def apply(self, sdfg: SDFG):
graph = sdfg.node(self.state_id)
map_exit = graph.node(self.subgraph[AccumulateTransient._map_exit])
outer_map_exit = graph.node(
self.subgraph[AccumulateTransient._outer_map_exit])
# Choose array
array = self.array
if array is None or len(array) == 0:
array = next(e.data.data
for e in graph.edges_between(map_exit, outer_map_exit)
if e.data.wcr is not None)
# Avoid import loop
from dace.transformation.dataflow.local_storage import OutLocalStorage
data_node: nodes.AccessNode = OutLocalStorage.apply_to(
sdfg,
dict(array=array),
verify=False,
save=False,
node_a=map_exit,
node_b=outer_map_exit)
if self.identity is None:
warnings.warn('AccumulateTransient did not properly initialize '
'newly-created transient!')
return
sdfg_state: SDFGState = sdfg.node(self.state_id)
map_entry = sdfg_state.entry_node(map_exit)
nested_sdfg: NestedSDFG = nest_state_subgraph(
sdfg=sdfg,
state=sdfg_state,
subgraph=SubgraphView(
sdfg_state, {map_entry, map_exit}
| sdfg_state.all_nodes_between(map_entry, map_exit)))
nested_sdfg_state: SDFGState = nested_sdfg.sdfg.nodes()[0]
init_state = nested_sdfg.sdfg.add_state_before(nested_sdfg_state)
temp_array: Array = sdfg.arrays[data_node.data]
init_state.add_mapped_tasklet(
name='acctrans_init',
map_ranges={
'_o%d' % i: '0:%s' % symstr(d)
for i, d in enumerate(temp_array.shape)
},
inputs={},
code='out = %s' % self.identity,
outputs={
'out':
dace.Memlet.simple(data=data_node.data,
subset_str=','.join(
['0:%d' % i for i in temp_array.shape]))
},
external_edges=True)
class MapExpansion(pm.Transformation):
""" Implements the map-expansion pattern.
Map-expansion takes an N-dimensional map and expands it to N
unidimensional maps.
New edges abide 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)
"""
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
@staticmethod
def expressions():
return [sdutil.node_path_graph(MapExpansion._map_entry)]
@staticmethod
def can_be_applied(graph: dace.sdfg.graph.OrderedMultiDiConnectorGraph,
candidate: Dict[dace.sdfg.nodes.Node, int],
expr_index: int,
sdfg: dace.SDFG,
strict: bool = False):
# A candidate subgraph matches the map-expansion pattern when it
# includes an N-dimensional map, with N greater than one.
map_entry = graph.nodes()[candidate[MapExpansion._map_entry]]
return map_entry.map.get_param_num() > 1
@staticmethod
def match_to_str(graph: dace.sdfg.graph.OrderedMultiDiConnectorGraph,
candidate: Dict[dace.sdfg.nodes.Node, int]):
map_entry = graph.nodes()[candidate[MapExpansion._map_entry]]
return map_entry.map.label + ': ' + str(map_entry.map.params)
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 _create_ceil_range(self, sdfg: SDFG, graph: SDFGState,
map_entry: nodes.MapEntry):
map_exit = graph.exit_node(map_entry)
# Retrieve transformation properties.
dim_idx = self.dim_idx
new_dim_prefix = self.new_dim_prefix
tile_size = self.tile_size
divides_evenly = self.divides_evenly
strided = self.strided
offset = self.tile_offset
tile_stride = self.tile_stride
if tile_stride == 0:
tile_stride = tile_size
# Retrieve parameter and range of dimension to be strip-mined.
target_dim = map_entry.map.params[dim_idx]
td_from, td_to, td_step = map_entry.map.range[dim_idx]
# Create new map. Replace by cloning map object?
new_dim = self._find_new_dim(sdfg, graph, map_entry, new_dim_prefix,
target_dim)
nd_from = 0
if tile_stride == 1:
nd_to = td_to - td_from
else:
nd_to = symbolic.pystr_to_symbolic(
'int_ceil(%s + 1 - %s, %s) - 1' %
(symbolic.symstr(td_to), symbolic.symstr(td_from),
symbolic.symstr(tile_stride)))
nd_step = 1
new_dim_range = (nd_from, nd_to, nd_step)
new_map = nodes.Map(new_dim + '_' + map_entry.map.label, [new_dim],
subsets.Range([new_dim_range]))
# Change the range of the selected dimension to iterate over a single
# tile
if strided:
td_from_new = symbolic.pystr_to_symbolic(new_dim)
td_to_new_approx = td_to
td_step = tile_size
elif offset == 0:
td_from_new = symbolic.pystr_to_symbolic(
'%s + %s * %s' %
(symbolic.symstr(td_from), symbolic.symstr(new_dim),
symbolic.symstr(tile_stride)))
td_to_new_exact = symbolic.pystr_to_symbolic(
'min(%s + 1, %s + %s * %s + %s) - 1' %
(symbolic.symstr(td_to), symbolic.symstr(td_from),
symbolic.symstr(tile_stride), symbolic.symstr(new_dim),
symbolic.symstr(tile_size)))
td_to_new_approx = symbolic.pystr_to_symbolic(
'%s + %s * %s + %s - 1' %
(symbolic.symstr(td_from), symbolic.symstr(tile_stride),
symbolic.symstr(new_dim), symbolic.symstr(tile_size)))
else:
# include offset
td_from_new_exact = symbolic.pystr_to_symbolic(
'max(%s,%s + %s * %s - %s)' %
(symbolic.symstr(td_from), symbolic.symstr(td_from),
symbolic.symstrtr(tile_stride), symbolic.symstr(new_dim),
symbolic.symstr(offset)))
td_from_new_approx = symbolic.pystr_to_symbolic(
'%s + %s * %s - %s ' %
(symbolic.symstr(td_from), symbolic.symstr(tile_stride),
symbolic.symstr(new_dim), symbolic.symstr(offset)))
td_from_new = dace.symbolic.SymExpr(td_from_new_exact,
td_from_new_approx)
td_to_new_exact = symbolic.pystr_to_symbolic(
'min(%s + 1, %s + %s * %s + %s - %s) -1' %
(symbolic.symstr(td_to), symbolic.symstr(td_from),
symbolic.symstr(tile_stride), symbolic.symstr(new_dim),
symbolic.symstr(tile_size), symbolic.symstr(offset)))
td_to_new_approx = symbolic.pystr_to_symbolic(
'%s + %s * %s + %s - %s - 1' %
(symbolic.symstr(td_from), symbolic.symstr(tile_stride),
symbolic.symstr(new_dim), symbolic.symstr(tile_size),
symbolic.symstr(offset)))
if divides_evenly or strided:
td_to_new = td_to_new_approx
else:
td_to_new = dace.symbolic.SymExpr(td_to_new_exact,
td_to_new_approx)
return new_dim, new_map, (td_from_new, td_to_new, td_step)
class StripMining(transformation.Transformation):
""" Implements the strip-mining transformation.
Strip-mining takes as input a map dimension and splits it into
two dimensions. The new dimension iterates over the range of
the original one with a parameterizable step, called the tile
size. The original dimension is changed to iterates over the
range of the tile size, with the same step as before.
"""
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
# Properties
dim_idx = Property(dtype=int,
default=-1,
desc="Index of dimension to be strip-mined")
new_dim_prefix = Property(dtype=str,
default="tile",
desc="Prefix for new dimension name")
tile_size = SymbolicProperty(
default=64,
desc="Tile size of strip-mined dimension, "
"or number of tiles if tiling_type=number_of_tiles")
tile_stride = SymbolicProperty(default=0,
desc="Stride between two tiles of the "
"strip-mined dimension. If zero, it is set "
"equal to the tile size.")
tile_offset = SymbolicProperty(default=0,
desc="Tile stride offset (negative)")
divides_evenly = Property(dtype=bool,
default=False,
desc="Tile size divides dimension range evenly?")
strided = Property(
dtype=bool,
default=False,
desc="Continuous (false) or strided (true) elements in tile")
tiling_type = Property(
dtype=str,
default='normal',
choices=['normal', 'ceilrange', 'number_of_tiles'],
allow_none=True,
desc="normal: the outerloop increments with tile_size, "
"ceilrange: uses ceiling(N/tile_size) in outer range, "
"number_of_tiles: tiles the map into the number of provided tiles, "
"provide the number of tiles over tile_size")
skew = Property(
dtype=bool,
default=False,
desc="If True, offsets inner tile back such that it starts with zero")
@staticmethod
def annotates_memlets():
return True
@staticmethod
def expressions():
return [
sdutil.node_path_graph(StripMining._map_entry)
# kStripMining._tasklet, StripMining._map_exit)
]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
return True
@staticmethod
def match_to_str(graph, candidate):
map_entry = graph.nodes()[candidate[StripMining._map_entry]]
return map_entry.map.label + ': ' + str(map_entry.map.params)
def apply(self, sdfg):
graph = sdfg.nodes()[self.state_id]
# Strip-mine selected dimension.
_, _, new_map = self._stripmine(sdfg, graph, self.subgraph)
return new_map
# def __init__(self, tag=True):
def __init__(self, *args, **kwargs):
self._entry = nodes.EntryNode()
self._tasklet = nodes.Tasklet('_')
self._exit = nodes.ExitNode()
super().__init__(*args, **kwargs)
# self.tag = tag
@property
def entry(self):
return self._entry
@property
def exit(self):
return self._exit
@property
def tasklet(self):
return self._tasklet
def print_match_pattern(self, candidate):
gentry = candidate[self.entry]
return str(gentry.map.params[-1])
def _find_new_dim(self, sdfg: SDFG, state: SDFGState,
entry: nodes.MapEntry, prefix: str, target_dim: str):
""" Finds a variable that is not already defined in scope. """
stree = state.scope_tree()
if len(prefix) == 0:
return target_dim
candidate = '%s_%s' % (prefix, target_dim)
index = 1
while candidate in map(str, stree[entry].defined_vars):
candidate = '%s%d_%s' % (prefix, index, target_dim)
index += 1
return candidate
def _create_strided_range(self, sdfg: SDFG, state: SDFGState,
map_entry: nodes.MapEntry):
map_exit = state.exit_node(map_entry)
dim_idx = self.dim_idx
new_dim_prefix = self.new_dim_prefix
tile_size = self.tile_size
divides_evenly = self.divides_evenly
tile_stride = self.tile_stride
if tile_stride == 0:
tile_stride = tile_size
if tile_stride != tile_size:
raise NotImplementedError
# Retrieve parameter and range of dimension to be strip-mined.
target_dim = map_entry.map.params[dim_idx]
td_from, td_to, td_step = map_entry.map.range[dim_idx]
new_dim = self._find_new_dim(sdfg, state, map_entry, new_dim_prefix,
target_dim)
new_dim_range = (td_from, td_to, tile_size)
new_map = nodes.Map(map_entry.map.label, [new_dim],
subsets.Range([new_dim_range]))
dimsym = dace.symbolic.pystr_to_symbolic(new_dim)
td_from_new = dimsym
if divides_evenly:
td_to_new = dimsym + tile_size - 1
else:
if isinstance(td_to, dace.symbolic.SymExpr):
td_to = td_to.expr
td_to_new = dace.symbolic.SymExpr(
sympy.Min(dimsym + tile_size - 1, td_to),
dimsym + tile_size - 1)
td_step_new = td_step
return new_dim, new_map, (td_from_new, td_to_new, td_step_new)
def _create_ceil_range(self, sdfg: SDFG, graph: SDFGState,
map_entry: nodes.MapEntry):
map_exit = graph.exit_node(map_entry)
# Retrieve transformation properties.
dim_idx = self.dim_idx
new_dim_prefix = self.new_dim_prefix
tile_size = self.tile_size
divides_evenly = self.divides_evenly
strided = self.strided
offset = self.tile_offset
tile_stride = self.tile_stride
if tile_stride == 0:
tile_stride = tile_size
# Retrieve parameter and range of dimension to be strip-mined.
target_dim = map_entry.map.params[dim_idx]
td_from, td_to, td_step = map_entry.map.range[dim_idx]
# Create new map. Replace by cloning map object?
new_dim = self._find_new_dim(sdfg, graph, map_entry, new_dim_prefix,
target_dim)
nd_from = 0
if tile_stride == 1:
nd_to = td_to - td_from
else:
nd_to = symbolic.pystr_to_symbolic(
'int_ceil(%s + 1 - %s, %s) - 1' %
(symbolic.symstr(td_to), symbolic.symstr(td_from),
symbolic.symstr(tile_stride)))
nd_step = 1
new_dim_range = (nd_from, nd_to, nd_step)
new_map = nodes.Map(new_dim + '_' + map_entry.map.label, [new_dim],
subsets.Range([new_dim_range]))
# Change the range of the selected dimension to iterate over a single
# tile
if strided:
td_from_new = symbolic.pystr_to_symbolic(new_dim)
td_to_new_approx = td_to
td_step = tile_size
elif offset == 0:
td_from_new = symbolic.pystr_to_symbolic(
'%s + %s * %s' %
(symbolic.symstr(td_from), symbolic.symstr(new_dim),
symbolic.symstr(tile_stride)))
td_to_new_exact = symbolic.pystr_to_symbolic(
'min(%s + 1, %s + %s * %s + %s) - 1' %
(symbolic.symstr(td_to), symbolic.symstr(td_from),
symbolic.symstr(tile_stride), symbolic.symstr(new_dim),
symbolic.symstr(tile_size)))
td_to_new_approx = symbolic.pystr_to_symbolic(
'%s + %s * %s + %s - 1' %
(symbolic.symstr(td_from), symbolic.symstr(tile_stride),
symbolic.symstr(new_dim), symbolic.symstr(tile_size)))
else:
# include offset
td_from_new_exact = symbolic.pystr_to_symbolic(
'max(%s,%s + %s * %s - %s)' %
(symbolic.symstr(td_from), symbolic.symstr(td_from),
symbolic.symstrtr(tile_stride), symbolic.symstr(new_dim),
symbolic.symstr(offset)))
td_from_new_approx = symbolic.pystr_to_symbolic(
'%s + %s * %s - %s ' %
(symbolic.symstr(td_from), symbolic.symstr(tile_stride),
symbolic.symstr(new_dim), symbolic.symstr(offset)))
td_from_new = dace.symbolic.SymExpr(td_from_new_exact,
td_from_new_approx)
td_to_new_exact = symbolic.pystr_to_symbolic(
'min(%s + 1, %s + %s * %s + %s - %s) -1' %
(symbolic.symstr(td_to), symbolic.symstr(td_from),
symbolic.symstr(tile_stride), symbolic.symstr(new_dim),
symbolic.symstr(tile_size), symbolic.symstr(offset)))
td_to_new_approx = symbolic.pystr_to_symbolic(
'%s + %s * %s + %s - %s - 1' %
(symbolic.symstr(td_from), symbolic.symstr(tile_stride),
symbolic.symstr(new_dim), symbolic.symstr(tile_size),
symbolic.symstr(offset)))
if divides_evenly or strided:
td_to_new = td_to_new_approx
else:
td_to_new = dace.symbolic.SymExpr(td_to_new_exact,
td_to_new_approx)
return new_dim, new_map, (td_from_new, td_to_new, td_step)
def _create_from_tile_numbers(self, sdfg: SDFG, state: SDFGState,
map_entry: nodes.MapEntry):
map_exit = state.exit_node(map_entry)
# Retrieve transformation properties.
dim_idx = self.dim_idx
new_dim_prefix = self.new_dim_prefix
divides_evenly = self.divides_evenly
number_of_tiles = self.tile_size
tile_stride = self.tile_stride
number_of_tiles = dace.symbolic.pystr_to_symbolic(number_of_tiles)
# Retrieve parameter and range of dimension to be strip-mined.
target_dim = map_entry.map.params[dim_idx]
td_from, td_to, td_step = map_entry.map.range[dim_idx]
tile_size = map_entry.map.range.size_exact()[dim_idx] / number_of_tiles
if tile_stride == 0:
tile_stride = tile_size
if tile_stride != tile_size:
raise NotImplementedError
new_dim = self._find_new_dim(sdfg, state, map_entry, new_dim_prefix,
target_dim)
new_dim_range = (td_from, number_of_tiles - 1, 1)
new_map = nodes.Map(map_entry.map.label, [new_dim],
subsets.Range([new_dim_range]))
dimsym = dace.symbolic.pystr_to_symbolic(new_dim)
td_from_new = dimsym * tile_size
if divides_evenly:
td_to_new = (dimsym + 1) * tile_size - 1
else:
if isinstance(td_to, dace.symbolic.SymExpr):
td_to = td_to.expr
td_to_new = dace.symbolic.SymExpr(
sympy.Min((dimsym + 1) * tile_size - 1, td_to),
(dimsym + 1) * tile_size - 1)
td_step_new = td_step
return new_dim, new_map, (td_from_new, td_to_new, td_step_new)
def _stripmine(self, sdfg, graph, candidate):
# Retrieve map entry and exit nodes.
map_entry = graph.nodes()[candidate[StripMining._map_entry]]
map_exit = graph.exit_node(map_entry)
# Retrieve transformation properties.
dim_idx = self.dim_idx
target_dim = map_entry.map.params[dim_idx]
if self.tiling_type == 'ceilrange':
new_dim, new_map, td_rng = self._create_ceil_range(
sdfg, graph, map_entry)
elif self.tiling_type == 'number_of_tiles':
new_dim, new_map, td_rng = self._create_from_tile_numbers(
sdfg, graph, map_entry)
else:
new_dim, new_map, td_rng = self._create_strided_range(
sdfg, graph, map_entry)
new_map_entry = nodes.MapEntry(new_map)
new_map_exit = nodes.MapExit(new_map)
td_to_new_approx = td_rng[1]
if isinstance(td_to_new_approx, dace.symbolic.SymExpr):
td_to_new_approx = td_to_new_approx.approx
# Special case: If range is 1 and no prefix was specified, skip range
if td_rng[0] == td_to_new_approx and target_dim == new_dim:
map_entry.map.range = subsets.Range(
[r for i, r in enumerate(map_entry.map.range) if i != dim_idx])
map_entry.map.params = [
p for i, p in enumerate(map_entry.map.params) if i != dim_idx
]
if len(map_entry.map.params) == 0:
raise ValueError('Strip-mining all dimensions of the map with '
'empty tiles is disallowed')
else:
map_entry.map.range[dim_idx] = td_rng
# Make internal map's schedule to "not parallel"
new_map.schedule = map_entry.map.schedule
map_entry.map.schedule = dtypes.ScheduleType.Sequential
# Redirect edges
new_map_entry.in_connectors = dcpy(map_entry.in_connectors)
sdutil.change_edge_dest(graph, map_entry, new_map_entry)
new_map_exit.out_connectors = dcpy(map_exit.out_connectors)
sdutil.change_edge_src(graph, map_exit, new_map_exit)
# Create new entry edges
new_in_edges = dict()
entry_in_conn = {}
entry_out_conn = {}
for _src, src_conn, _dst, _, memlet in graph.out_edges(map_entry):
if (src_conn is not None
and src_conn[:4] == 'OUT_' and not isinstance(
sdfg.arrays[memlet.data], dace.data.Scalar)):
new_subset = calc_set_image(
map_entry.map.params,
map_entry.map.range,
memlet.subset,
)
conn = src_conn[4:]
key = (memlet.data, 'IN_' + conn, 'OUT_' + conn)
if key in new_in_edges.keys():
old_subset = new_in_edges[key].subset
new_in_edges[key].subset = calc_set_union(
old_subset, new_subset)
else:
entry_in_conn['IN_' + conn] = None
entry_out_conn['OUT_' + conn] = None
new_memlet = dcpy(memlet)
new_memlet.subset = new_subset
if memlet.dynamic:
new_memlet.num_accesses = memlet.num_accesses
else:
new_memlet.num_accesses = new_memlet.num_elements()
new_in_edges[key] = new_memlet
else:
if src_conn is not None and src_conn[:4] == 'OUT_':
conn = src_conn[4:]
in_conn = 'IN_' + conn
out_conn = 'OUT_' + conn
else:
in_conn = src_conn
out_conn = src_conn
if in_conn:
entry_in_conn[in_conn] = None
if out_conn:
entry_out_conn[out_conn] = None
new_in_edges[(memlet.data, in_conn, out_conn)] = dcpy(memlet)
new_map_entry.out_connectors = entry_out_conn
map_entry.in_connectors = entry_in_conn
for (_, in_conn, out_conn), memlet in new_in_edges.items():
graph.add_edge(new_map_entry, out_conn, map_entry, in_conn, memlet)
# Create new exit edges
new_out_edges = dict()
exit_in_conn = {}
exit_out_conn = {}
for _src, _, _dst, dst_conn, memlet in graph.in_edges(map_exit):
if (dst_conn is not None
and dst_conn[:3] == 'IN_' and not isinstance(
sdfg.arrays[memlet.data], dace.data.Scalar)):
new_subset = calc_set_image(
map_entry.map.params,
map_entry.map.range,
memlet.subset,
)
conn = dst_conn[3:]
key = (memlet.data, 'IN_' + conn, 'OUT_' + conn)
if key in new_out_edges.keys():
old_subset = new_out_edges[key].subset
new_out_edges[key].subset = calc_set_union(
old_subset, new_subset)
else:
exit_in_conn['IN_' + conn] = None
exit_out_conn['OUT_' + conn] = None
new_memlet = dcpy(memlet)
new_memlet.subset = new_subset
if memlet.dynamic:
new_memlet.num_accesses = memlet.num_accesses
else:
new_memlet.num_accesses = new_memlet.num_elements()
new_out_edges[key] = new_memlet
else:
if dst_conn is not None and dst_conn[:3] == 'IN_':
conn = dst_conn[3:]
in_conn = 'IN_' + conn
out_conn = 'OUT_' + conn
else:
in_conn = dst_conn
out_conn = dst_conn
if in_conn:
exit_in_conn[in_conn] = None
if out_conn:
exit_out_conn[out_conn] = None
new_in_edges[(memlet.data, in_conn, out_conn)] = dcpy(memlet)
new_map_exit.in_connectors = exit_in_conn
map_exit.out_connectors = exit_out_conn
for (_, in_conn, out_conn), memlet in new_out_edges.items():
graph.add_edge(map_exit, out_conn, new_map_exit, in_conn, memlet)
# Skew if necessary
if self.skew:
xfh.offset_map(sdfg, graph, map_entry, dim_idx, td_rng[0])
# Return strip-mined dimension.
return target_dim, new_dim, new_map
class MapToForLoop(pattern_matching.Transformation):
""" Implements the Map to for-loop transformation.
Takes a map and enforces a sequential schedule by transforming it into
a state-machine of a for-loop. Creates a nested SDFG, if necessary.
"""
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
@staticmethod
def annotates_memlets():
return True
@staticmethod
def expressions():
return [sdutil.node_path_graph(MapToForLoop._map_entry)]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
# Only uni-dimensional maps are accepted.
map_entry = graph.nodes()[candidate[MapToForLoop._map_entry]]
if len(map_entry.map.params) > 1:
return False
return True
@staticmethod
def match_to_str(graph, candidate):
map_entry = graph.nodes()[candidate[MapToForLoop._map_entry]]
return map_entry.map.label + ': ' + str(map_entry.map.params)
def apply(self, sdfg) -> Tuple[nodes.NestedSDFG, SDFGState]:
""" Applies the transformation and returns a tuple with the new nested
SDFG node and the main state in the for-loop. """
# Retrieve map entry and exit nodes.
graph = sdfg.nodes()[self.state_id]
map_entry = graph.nodes()[self.subgraph[MapToForLoop._map_entry]]
map_exit = graph.exit_node(map_entry)
loop_idx = map_entry.map.params[0]
loop_from, loop_to, loop_step = map_entry.map.range[0]
# Turn the map scope into a nested SDFG
node = nest_state_subgraph(sdfg, graph,
graph.scope_subgraph(map_entry))
nsdfg: SDFG = node.sdfg
nstate: SDFGState = nsdfg.nodes()[0]
# If map range is dynamic, replace loop expressions with memlets
param_to_edge = {}
for edge in nstate.in_edges(map_entry):
if edge.dst_conn and not edge.dst_conn.startswith('IN_'):
param = '__DACE_P%d' % len(param_to_edge)
repldict = {symbolic.pystr_to_symbolic(edge.dst_conn): param}
param_to_edge[param] = edge
loop_from = loop_from.subs(repldict)
loop_to = loop_to.subs(repldict)
loop_step = loop_step.subs(repldict)
# Avoiding import loop
from dace.codegen.targets.cpp import cpp_array_expr
def replace_param(param):
param = symbolic.symstr(param)
for p, pval in param_to_edge.items():
# TODO: Correct w.r.t. connector type
param = param.replace(p, cpp_array_expr(nsdfg, pval.data))
return param
# End of dynamic input range
# Create a loop inside the nested SDFG
nsdfg.add_loop(None, nstate, None, loop_idx, replace_param(loop_from),
'%s < %s' % (loop_idx, replace_param(loop_to + 1)),
'%s + %s' % (loop_idx, replace_param(loop_step)))
# Skip map in input edges
for edge in nstate.out_edges(map_entry):
src_node = nstate.memlet_path(edge)[0].src
nstate.add_edge(src_node, None, edge.dst, edge.dst_conn, edge.data)
nstate.remove_edge(edge)
# Skip map in output edges
for edge in nstate.in_edges(map_exit):
dst_node = nstate.memlet_path(edge)[-1].dst
nstate.add_edge(edge.src, edge.src_conn, dst_node, None, edge.data)
nstate.remove_edge(edge)
# Remove nodes from dynamic map range
nstate.remove_nodes_from(
[e.src for e in dace.sdfg.dynamic_map_inputs(nstate, map_entry)])
# Remove scope nodes
nstate.remove_nodes_from([map_entry, map_exit])
return node, nstate
class GPUTransformMap(transformation.Transformation):
""" Implements the GPUTransformMap transformation.
Converts a single map to a GPU-scheduled map and creates GPU arrays
outside it, generating CPU<->GPU memory copies automatically.
"""
fullcopy = Property(desc="Copy whole arrays rather than used subset",
dtype=bool,
default=False)
toplevel_trans = Property(desc="Make all GPU transients top-level",
dtype=bool,
default=False)
register_trans = Property(
desc="Make all transients inside GPU maps registers",
dtype=bool,
default=False)
gpu_id = SymbolicProperty(default=None,
allow_none=True,
desc="Selects which gpu the map should run on")
sequential_innermaps = Property(desc="Make all internal maps Sequential",
dtype=bool,
default=False)
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
import dace.libraries.standard as stdlib # Avoid import loop
_reduce = stdlib.Reduce('lambda: None', None)
@staticmethod
def expressions():
return [
sdutil.node_path_graph(GPUTransformMap._map_entry),
sdutil.node_path_graph(GPUTransformMap._reduce)
]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
if expr_index == 0:
map_entry = graph.nodes()[candidate[GPUTransformMap._map_entry]]
candidate_map = map_entry.map
# Map schedules that are disallowed to transform to GPUs
if (candidate_map.schedule
in [dtypes.ScheduleType.MPI] + dtypes.GPU_SCHEDULES):
return False
if sd.is_devicelevel_gpu(sdfg, graph, map_entry):
return False
# Dynamic map ranges cannot become kernels
if sd.has_dynamic_map_inputs(graph, map_entry):
return False
# Ensure that map does not include internal arrays that are
# allocated on non-default space
subgraph = graph.scope_subgraph(map_entry)
for node in subgraph.nodes():
if (isinstance(node, nodes.AccessNode) and
node.desc(sdfg).storage != dtypes.StorageType.Default
and
node.desc(sdfg).storage != dtypes.StorageType.Register):
return False
# If one of the outputs is a stream, do not match
map_exit = graph.exit_node(map_entry)
for edge in graph.out_edges(map_exit):
dst = graph.memlet_path(edge)[-1].dst
if (isinstance(dst, nodes.AccessNode)
and isinstance(sdfg.arrays[dst.data], data.Stream)):
return False
return True
elif expr_index == 1:
reduce = graph.nodes()[candidate[GPUTransformMap._reduce]]
# Disallow GPU transformation if already in device-level code
if sd.is_devicelevel_gpu(sdfg, graph, reduce):
return False
return True
@staticmethod
def match_to_str(graph, candidate):
if GPUTransformMap._reduce in candidate:
return str(graph.nodes()[candidate[GPUTransformMap._reduce]])
else:
return str(graph.nodes()[candidate[GPUTransformMap._map_entry]])
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.gpu_id = self.gpu_id
transformation.apply(nsdfg_node.sdfg)
# Inline back as necessary
sdfg.apply_strict_transformations()
class TrivialMapElimination(transformation.Transformation):
""" Implements the Trivial-Map Elimination pattern.
Trivial-Map Elimination takes a map with a range
containing one element and removes the map.
Example: Map[i=0] -> nothing
"""
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
@staticmethod
def expressions():
return [sdutil.node_path_graph(TrivialMapElimination._map_entry)]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
map_entry = graph.nodes()[candidate[TrivialMapElimination._map_entry]]
map_from, map_to, map_step = map_entry.map.range[0]
return len(map_entry.map.range) == 1 and map_to == map_from
@staticmethod
def match_to_str(graph, candidate):
map_entry = graph.nodes()[candidate[TrivialMapElimination._map_entry]]
return map_entry.map.label + ': ' + str(map_entry.map.params)
def apply(self, sdfg):
graph = sdfg.nodes()[self.state_id]
map_entry = graph.nodes()[self.subgraph[
TrivialMapElimination._map_entry]]
map_exit = graph.exit_node(map_entry)
map_param = map_entry.map.params[0]
map_from, map_to, _ = map_entry.map.range[0]
assert map_from == map_to
# Replace the map index variable with the value it obtained
scope = graph.scope_subgraph(map_entry)
scope.replace(map_param, map_from)
# Redirect map entry's out edges.
for edge in graph.out_edges(map_entry):
path = graph.memlet_path(edge)
ind = path.index(edge)
# Add an edge directly from the previous source connector to the
# destination
graph.add_edge(path[ind - 1].src, path[ind - 1].src_conn, edge.dst,
edge.dst_conn, edge.data)
# Redirect map exit's in edges.
for edge in graph.in_edges(map_exit):
path = graph.memlet_path(edge)
ind = path.index(edge)
# Add an edge directly from the source to the next destination
# connector
graph.add_edge(edge.src, edge.src_conn, path[ind + 1].dst,
path[ind + 1].dst_conn, edge.data)
# Clean-up
graph.remove_nodes_from([map_entry, map_exit])
class Vectorization(pattern_matching.Transformation):
""" Implements the vectorization transformation.
Vectorization matches when all the input and output memlets of a
tasklet inside a map access the inner-most loop variable in their last
dimension. The transformation changes the step of the inner-most loop
to be equal to the length of the vector and vectorizes the memlets.
"""
vector_len = Property(desc="Vector length", dtype=int, default=4)
propagate_parent = Property(desc="Propagate vector length through "
"parent SDFGs",
dtype=bool,
default=False)
strided_map = Property(desc="Use strided map range (jump by vector length)"
" instead of modifying memlets",
dtype=bool,
default=False)
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
_tasklet = nodes.Tasklet('_')
_map_exit = nodes.MapExit(nodes.Map("", [], []))
@staticmethod
def expressions():
return [
sdutil.node_path_graph(Vectorization._map_entry,
Vectorization._tasklet,
Vectorization._map_exit)
]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
map_entry = graph.nodes()[candidate[Vectorization._map_entry]]
tasklet = graph.nodes()[candidate[Vectorization._tasklet]]
param = symbolic.pystr_to_symbolic(map_entry.map.params[-1])
found = False
# Check if all edges, adjacent to the tasklet,
# use the parameter in their last dimension.
for _src, _, _dest, _, memlet in graph.all_edges(tasklet):
# Cases that do not matter for vectorization
if memlet.data is None: # Empty memlets
continue
if isinstance(sdfg.arrays[memlet.data], data.Stream): # Streams
continue
# Vectorization can not be applied in WCR
if memlet.wcr is not None:
return False
try:
subset = memlet.subset
veclen = memlet.veclen
except AttributeError:
return False
if subset is None:
return False
try:
if veclen > symbolic.pystr_to_symbolic('1'):
return False
for idx, expr in enumerate(subset):
if isinstance(expr, tuple):
for ex in expr:
ex = symbolic.pystr_to_symbolic(ex)
symbols = ex.free_symbols
if param in symbols:
if idx == subset.dims() - 1:
found = True
else:
return False
else:
expr = symbolic.pystr_to_symbolic(expr)
symbols = expr.free_symbols
if param in symbols:
if idx == subset.dims() - 1:
found = True
else:
return False
except TypeError: # cannot determine truth value of Relational
return False
return found
@staticmethod
def match_to_str(graph, candidate):
map_entry = candidate[Vectorization._map_entry]
tasklet = candidate[Vectorization._tasklet]
map_exit = candidate[Vectorization._map_exit]
return ' -> '.join(
str(node) for node in [map_entry, tasklet, map_exit])
def apply(self, sdfg):
graph = sdfg.nodes()[self.state_id]
map_entry = graph.nodes()[self.subgraph[Vectorization._map_entry]]
tasklet = graph.nodes()[self.subgraph[Vectorization._tasklet]]
map_exit = graph.nodes()[self.subgraph[Vectorization._map_exit]]
param = symbolic.pystr_to_symbolic(map_entry.map.params[-1])
# Create new vector size.
vector_size = self.vector_len
# Change the step of the inner-most dimension.
dim_from, dim_to, dim_step = map_entry.map.range[-1]
if self.strided_map:
map_entry.map.range[-1] = (dim_from, dim_to, vector_size)
else:
map_entry.map.range[-1] = (dim_from,
(dim_to + 1) / vector_size - 1,
dim_step)
# TODO: Postamble and/or preamble non-vectorized map
# Vectorize memlets adjacent to the tasklet.
processed_edges = set()
for edge in graph.all_edges(tasklet):
_src, _, _dest, _, memlet = edge
if memlet.data is None: # Empty memlets
continue
lastindex = memlet.subset[-1]
if isinstance(lastindex, tuple):
symbols = set()
for indd in lastindex:
symbols.update(
symbolic.pystr_to_symbolic(indd).free_symbols)
else:
symbols = symbolic.pystr_to_symbolic(
memlet.subset[-1]).free_symbols
if param not in symbols:
continue
try:
# propagate vector length inside this SDFG
for e in graph.memlet_tree(edge):
e.data.veclen = vector_size
if not self.strided_map and e not in processed_edges:
e.data.subset.replace({param: vector_size * param})
processed_edges.add(e)
# propagate to the parent (TODO: handle multiple level of nestings)
if self.propagate_parent and sdfg.parent is not None:
source_edge = graph.memlet_path(edge)[0]
sink_edge = graph.memlet_path(edge)[-1]
# Find parent Nested SDFG node
parent_node = next(n for n in sdfg.parent.nodes()
if isinstance(n, nodes.NestedSDFG)
and n.sdfg.name == sdfg.name)
# continue in propagating the vector length following the
# path that arrives to source_edge or starts from sink_edge
for pe in sdfg.parent.all_edges(parent_node):
if str(pe.dst_conn) == str(source_edge.src) or str(
pe.src_conn) == str(sink_edge.dst):
for ppe in sdfg.parent.memlet_tree(pe):
ppe.data.veclen = vector_size
if (not self.strided_map
and ppe not in processed_edges):
ppe.data.subset.replace(
{param: vector_size * param})
processed_edges.add(ppe)
except AttributeError:
raise
return
class MapCollapse(pattern_matching.Transformation):
""" Implements the Map Collapse pattern.
Map-collapse takes two nested maps with M and N dimensions respectively,
and collapses them to a single M+N dimensional map.
"""
_outer_map_entry = nodes.MapEntry(nodes.Map("", [], []))
_inner_map_entry = nodes.MapEntry(nodes.Map("", [], []))
@staticmethod
def expressions():
return [
sdutil.node_path_graph(
MapCollapse._outer_map_entry,
MapCollapse._inner_map_entry,
)
]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
# Check the edges between the entries of the two maps.
outer_map_entry = graph.nodes()[candidate[
MapCollapse._outer_map_entry]]
inner_map_entry = 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
return True
@staticmethod
def match_to_str(graph, candidate):
outer_map_entry = graph.nodes()[candidate[
MapCollapse._outer_map_entry]]
inner_map_entry = graph.nodes()[candidate[
MapCollapse._inner_map_entry]]
return ' -> '.join(entry.map.label + ': ' + str(entry.map.params)
for entry in [outer_map_entry, inner_map_entry])
def apply(self, sdfg) -> Tuple[nodes.MapEntry, nodes.MapExit]:
"""
Collapses two maps into one.
:param sdfg: The SDFG to apply the transformation to.
:return: A 2-tuple of the new map entry and exit nodes.
"""
# Extract the parameters and ranges of the inner/outer maps.
graph = sdfg.nodes()[self.state_id]
outer_map_entry = graph.nodes()[self.subgraph[
MapCollapse._outer_map_entry]]
inner_map_entry = graph.nodes()[self.subgraph[
MapCollapse._inner_map_entry]]
inner_map_exit = graph.exit_node(inner_map_entry)
outer_map_exit = graph.exit_node(outer_map_entry)
return sdutil.merge_maps(graph, outer_map_entry, outer_map_exit,
inner_map_entry, inner_map_exit)
class GPUTransformLocalStorage(transformation.Transformation):
"""Implements the GPUTransformLocalStorage transformation.
Similar to GPUTransformMap, but takes multiple maps leading from the
same data node into account, creating a local storage for each range.
@see: GPUTransformMap
"""
_arrays_removed = 0
_maps_transformed = 0
fullcopy = Property(desc="Copy whole arrays rather than used subset",
dtype=bool,
default=False)
nested_seq = Property(
desc="Makes nested code semantically-equivalent to single-core code,"
"transforming nested maps and memory into sequential and "
"local memory respectively.",
dtype=bool,
default=True,
)
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
import dace.libraries.standard as stdlib # Avoid import loop
_reduce = stdlib.Reduce("lambda: None", None)
@staticmethod
def expressions():
return [
sdutil.node_path_graph(GPUTransformLocalStorage._map_entry),
sdutil.node_path_graph(GPUTransformLocalStorage._reduce),
]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, permissive=False):
if expr_index == 0:
map_entry = graph.nodes()[candidate[
GPUTransformLocalStorage._map_entry]]
candidate_map = map_entry.map
# Disallow GPUTransform on nested maps in permissive mode
if not permissive:
if graph.entry_node(map_entry) is not None:
return False
# Map schedules that are disallowed to transform to GPUs
if (candidate_map.schedule == dtypes.ScheduleType.MPI
or candidate_map.schedule == dtypes.ScheduleType.GPU_Device
or candidate_map.schedule
== dtypes.ScheduleType.GPU_ThreadBlock or
candidate_map.schedule == dtypes.ScheduleType.Sequential):
return False
# Dynamic map ranges cannot become kernels
if sd.has_dynamic_map_inputs(graph, map_entry):
return False
# Recursively check parent for GPU schedules
sdict = graph.scope_dict()
current_node = map_entry
while current_node is not None:
if (current_node.map.schedule == dtypes.ScheduleType.GPU_Device
or current_node.map.schedule
== dtypes.ScheduleType.GPU_ThreadBlock):
return False
current_node = sdict[current_node]
# Ensure that map does not include internal arrays that are
# allocated on non-default space
subgraph = graph.scope_subgraph(map_entry)
for node in subgraph.nodes():
if (isinstance(node, nodes.AccessNode) and
node.desc(sdfg).storage != dtypes.StorageType.Default
and
node.desc(sdfg).storage != dtypes.StorageType.Register):
return False
# If one of the outputs is a stream, do not match
map_exit = graph.exit_node(map_entry)
for edge in graph.out_edges(map_exit):
dst = graph.memlet_path(edge)[-1].dst
if (isinstance(dst, nodes.AccessNode)
and isinstance(sdfg.arrays[dst.data], data.Stream)):
return False
return True
elif expr_index == 1:
reduce = graph.nodes()[candidate[GPUTransformLocalStorage._reduce]]
# Recursively check parent for GPU schedules
sdict = graph.scope_dict()
current_node = sdict[reduce]
while current_node is not None:
if (current_node.map.schedule == dtypes.ScheduleType.GPU_Device
or current_node.map.schedule
== dtypes.ScheduleType.GPU_ThreadBlock):
return False
current_node = sdict[current_node]
return True
@staticmethod
def match_to_str(graph, candidate):
if GPUTransformLocalStorage._reduce in candidate:
return str(
graph.nodes()[candidate[GPUTransformLocalStorage._reduce]])
else:
map_entry = graph.nodes()[candidate[
GPUTransformLocalStorage._map_entry]]
return str(map_entry)
def apply(self, sdfg):
graph = sdfg.nodes()[self.state_id]
if self.expr_index == 0:
cnode: nodes.MapEntry = graph.nodes()[self.subgraph[
GPUTransformLocalStorage._map_entry]]
# Change schedule
cnode.schedule = dtypes.ScheduleType.GPU_Device
exit_node = graph.exit_node(cnode)
else:
cnode: nodes.LibraryNode = graph.nodes()[self.subgraph[
GPUTransformLocalStorage._reduce]]
# Change schedule
cnode.schedule = dtypes.ScheduleType.GPU_Default
exit_node = cnode
if Config.get_bool("debugprint"):
GPUTransformLocalStorage._maps_transformed += 1
# If nested graph is designated as sequential, transform schedules and
# storage from Default to Sequential/Register
if self.nested_seq and self.expr_index == 0:
for node in graph.scope_subgraph(cnode).nodes():
if isinstance(node, nodes.AccessNode):
arr = node.desc(sdfg)
if arr.storage == dtypes.StorageType.Default:
arr.storage = dtypes.StorageType.Register
elif isinstance(node, nodes.MapEntry):
if node.map.schedule == dtypes.ScheduleType.Default:
node.map.schedule = dtypes.ScheduleType.Sequential
gpu_storage_types = [
dtypes.StorageType.GPU_Global,
dtypes.StorageType.GPU_Shared,
]
#######################################################
# Add GPU copies of CPU arrays (i.e., not already on GPU)
# First, understand which arrays to clone
all_out_edges = []
all_out_edges.extend(list(graph.out_edges(exit_node)))
in_arrays_to_clone = set()
out_arrays_to_clone = set()
for e in graph.in_edges(cnode):
data_node = sd.find_input_arraynode(graph, e)
if data_node.desc(sdfg).storage not in gpu_storage_types:
in_arrays_to_clone.add((data_node, e.data))
for e in all_out_edges:
data_node = sd.find_output_arraynode(graph, e)
if data_node.desc(sdfg).storage not in gpu_storage_types:
out_arrays_to_clone.add((data_node, e.data))
if Config.get_bool("debugprint"):
GPUTransformLocalStorage._arrays_removed += len(
in_arrays_to_clone) + len(out_arrays_to_clone)
# Second, create a GPU clone of each array
# TODO: Overapproximate union of memlets
cloned_arrays = {}
in_cloned_arraynodes = {}
out_cloned_arraynodes = {}
for array_node, memlet in in_arrays_to_clone:
array = array_node.desc(sdfg)
cloned_name = "gpu_" + array_node.data
for i, r in enumerate(memlet.bounding_box_size()):
size = symbolic.overapproximate(r)
try:
if int(size) == 1:
suffix = []
for c in str(memlet.subset[i][0]):
if c.isalpha() or c.isdigit() or c == "_":
suffix.append(c)
elif c == "+":
suffix.append("p")
elif c == "-":
suffix.append("m")
elif c == "*":
suffix.append("t")
elif c == "/":
suffix.append("d")
cloned_name += "_" + "".join(suffix)
except:
continue
if cloned_name in sdfg.arrays.keys():
cloned_array = sdfg.arrays[cloned_name]
elif array_node.data in cloned_arrays:
cloned_array = cloned_arrays[array_node.data]
else:
full_shape = []
for r in memlet.bounding_box_size():
size = symbolic.overapproximate(r)
try:
full_shape.append(int(size))
except:
full_shape.append(size)
actual_dims = [
idx for idx, r in enumerate(full_shape)
if not (isinstance(r, int) and r == 1)
]
if len(actual_dims) == 0: # abort
actual_dims = [len(full_shape) - 1]
if isinstance(array, data.Scalar):
sdfg.add_array(name=cloned_name,
shape=[1],
dtype=array.dtype,
transient=True,
storage=dtypes.StorageType.GPU_Global)
elif isinstance(array, data.Stream):
sdfg.add_stream(
name=cloned_name,
dtype=array.dtype,
shape=[full_shape[d] for d in actual_dims],
veclen=array.veclen,
buffer_size=array.buffer_size,
storage=dtypes.StorageType.GPU_Global,
transient=True,
offset=[array.offset[d] for d in actual_dims])
else:
sdfg.add_array(
name=cloned_name,
shape=[full_shape[d] for d in actual_dims],
dtype=array.dtype,
transient=True,
storage=dtypes.StorageType.GPU_Global,
allow_conflicts=array.allow_conflicts,
strides=[array.strides[d] for d in actual_dims],
offset=[array.offset[d] for d in actual_dims],
)
cloned_arrays[array_node.data] = cloned_name
cloned_node = type(array_node)(cloned_name)
in_cloned_arraynodes[array_node.data] = cloned_node
for array_node, memlet in out_arrays_to_clone:
array = array_node.desc(sdfg)
cloned_name = "gpu_" + array_node.data
for i, r in enumerate(memlet.bounding_box_size()):
size = symbolic.overapproximate(r)
try:
if int(size) == 1:
suffix = []
for c in str(memlet.subset[i][0]):
if c.isalpha() or c.isdigit() or c == "_":
suffix.append(c)
elif c == "+":
suffix.append("p")
elif c == "-":
suffix.append("m")
elif c == "*":
suffix.append("t")
elif c == "/":
suffix.append("d")
cloned_name += "_" + "".join(suffix)
except:
continue
if cloned_name in sdfg.arrays.keys():
cloned_array = sdfg.arrays[cloned_name]
elif array_node.data in cloned_arrays:
cloned_array = cloned_arrays[array_node.data]
else:
full_shape = []
for r in memlet.bounding_box_size():
size = symbolic.overapproximate(r)
try:
full_shape.append(int(size))
except:
full_shape.append(size)
actual_dims = [
idx for idx, r in enumerate(full_shape)
if not (isinstance(r, int) and r == 1)
]
if len(actual_dims) == 0: # abort
actual_dims = [len(full_shape) - 1]
if isinstance(array, data.Scalar):
sdfg.add_array(name=cloned_name,
shape=[1],
dtype=array.dtype,
transient=True,
storage=dtypes.StorageType.GPU_Global)
elif isinstance(array, data.Stream):
sdfg.add_stream(
name=cloned_name,
dtype=array.dtype,
shape=[full_shape[d] for d in actual_dims],
veclen=array.veclen,
buffer_size=array.buffer_size,
storage=dtypes.StorageType.GPU_Global,
transient=True,
offset=[array.offset[d] for d in actual_dims])
else:
sdfg.add_array(
name=cloned_name,
shape=[full_shape[d] for d in actual_dims],
dtype=array.dtype,
transient=True,
storage=dtypes.StorageType.GPU_Global,
allow_conflicts=array.allow_conflicts,
strides=[array.strides[d] for d in actual_dims],
offset=[array.offset[d] for d in actual_dims],
)
cloned_arrays[array_node.data] = cloned_name
cloned_node = type(array_node)(cloned_name)
cloned_node.setzero = True
out_cloned_arraynodes[array_node.data] = cloned_node
# Third, connect the cloned arrays to the originals
for array_name, node in in_cloned_arraynodes.items():
graph.add_node(node)
is_scalar = isinstance(sdfg.arrays[array_name], data.Scalar)
for edge in graph.in_edges(cnode):
if edge.data.data == array_name:
newmemlet = copy.deepcopy(edge.data)
newmemlet.data = node.data
if is_scalar:
newmemlet.subset = sbs.Indices([0])
else:
offset = []
lost_dims = []
lost_ranges = []
newsubset = [None] * len(edge.data.subset)
for ind, r in enumerate(edge.data.subset):
offset.append(r[0])
if isinstance(edge.data.subset[ind], tuple):
begin = edge.data.subset[ind][0] - r[0]
end = edge.data.subset[ind][1] - r[0]
step = edge.data.subset[ind][2]
if begin == end:
lost_dims.append(ind)
lost_ranges.append((begin, end, step))
else:
newsubset[ind] = (begin, end, step)
else:
newsubset[ind] -= r[0]
if len(lost_dims) == len(edge.data.subset):
lost_dims.pop()
newmemlet.subset = type(
edge.data.subset)([lost_ranges[-1]])
else:
newmemlet.subset = type(edge.data.subset)(
[r for r in newsubset if r is not None])
graph.add_edge(node, None, edge.dst, edge.dst_conn,
newmemlet)
for e in graph.bfs_edges(edge.dst, reverse=False):
parent, _, _child, _, memlet = e
if parent != edge.dst and not in_scope(
graph, parent, edge.dst):
break
if memlet.data != edge.data.data:
continue
path = graph.memlet_path(e)
if not isinstance(path[-1].dst, nodes.CodeNode):
if in_path(path, e, nodes.ExitNode, forward=True):
if isinstance(parent, nodes.CodeNode):
# Output edge
break
else:
continue
if is_scalar:
memlet.subset = sbs.Indices([0])
else:
newsubset = [None] * len(memlet.subset)
for ind, r in enumerate(memlet.subset):
if ind in lost_dims:
continue
if isinstance(memlet.subset[ind], tuple):
begin = r[0] - offset[ind]
end = r[1] - offset[ind]
step = r[2]
newsubset[ind] = (begin, end, step)
else:
newsubset[ind] = (
r - offset[ind],
r - offset[ind],
1,
)
memlet.subset = type(edge.data.subset)(
[r for r in newsubset if r is not None])
memlet.data = node.data
if self.fullcopy:
edge.data.subset = sbs.Range.from_array(node.desc(sdfg))
edge.data.other_subset = newmemlet.subset
graph.add_edge(edge.src, edge.src_conn, node, None,
edge.data)
graph.remove_edge(edge)
for array_name, node in out_cloned_arraynodes.items():
graph.add_node(node)
is_scalar = isinstance(sdfg.arrays[array_name], data.Scalar)
for edge in all_out_edges:
if edge.data.data == array_name:
newmemlet = copy.deepcopy(edge.data)
newmemlet.data = node.data
if is_scalar:
newmemlet.subset = sbs.Indices([0])
else:
offset = []
lost_dims = []
lost_ranges = []
newsubset = [None] * len(edge.data.subset)
for ind, r in enumerate(edge.data.subset):
offset.append(r[0])
if isinstance(edge.data.subset[ind], tuple):
begin = edge.data.subset[ind][0] - r[0]
end = edge.data.subset[ind][1] - r[0]
step = edge.data.subset[ind][2]
if begin == end:
lost_dims.append(ind)
lost_ranges.append((begin, end, step))
else:
newsubset[ind] = (begin, end, step)
else:
newsubset[ind] -= r[0]
if len(lost_dims) == len(edge.data.subset):
lost_dims.pop()
newmemlet.subset = type(
edge.data.subset)([lost_ranges[-1]])
else:
newmemlet.subset = type(edge.data.subset)(
[r for r in newsubset if r is not None])
graph.add_edge(edge.src, edge.src_conn, node, None,
newmemlet)
end_node = graph.entry_node(edge.src)
for e in graph.bfs_edges(edge.src, reverse=True):
parent, _, _child, _, memlet = e
if parent == end_node:
break
if memlet.data != edge.data.data:
continue
path = graph.memlet_path(e)
if not isinstance(path[0].dst, nodes.CodeNode):
if in_path(path, e, nodes.EntryNode, forward=False):
if isinstance(parent, nodes.CodeNode):
# Output edge
break
else:
continue
if is_scalar:
memlet.subset = sbs.Indices([0])
else:
newsubset = [None] * len(memlet.subset)
for ind, r in enumerate(memlet.subset):
if ind in lost_dims:
continue
if isinstance(memlet.subset[ind], tuple):
begin = r[0] - offset[ind]
end = r[1] - offset[ind]
step = r[2]
newsubset[ind] = (begin, end, step)
else:
newsubset[ind] = (
r - offset[ind],
r - offset[ind],
1,
)
memlet.subset = type(edge.data.subset)(
[r for r in newsubset if r is not None])
memlet.data = node.data
edge.data.wcr = None
if self.fullcopy:
edge.data.subset = sbs.Range.from_array(node.desc(sdfg))
edge.data.other_subset = newmemlet.subset
graph.add_edge(node, None, edge.dst, edge.dst_conn,
edge.data)
graph.remove_edge(edge)
# Fourth, replace memlet arrays as necessary
if self.expr_index == 0:
scope_subgraph = graph.scope_subgraph(cnode)
for edge in scope_subgraph.edges():
if edge.data.data is not None and edge.data.data in cloned_arrays:
edge.data.data = cloned_arrays[edge.data.data]
class Vectorization(transformation.Transformation):
""" Implements the vectorization transformation.
Vectorization matches when all the input and output memlets of a
tasklet inside a map access the inner-most loop variable in their last
dimension. The transformation changes the step of the inner-most loop
to be equal to the length of the vector and vectorizes the memlets.
"""
vector_len = Property(desc="Vector length", dtype=int, default=4)
propagate_parent = Property(desc="Propagate vector length through "
"parent SDFGs",
dtype=bool,
default=False)
strided_map = Property(desc="Use strided map range (jump by vector length)"
" instead of modifying memlets",
dtype=bool,
default=True)
preamble = Property(
dtype=bool,
default=None,
allow_none=True,
desc='Force creation or skipping a preamble map without vectors')
postamble = Property(
dtype=bool,
default=None,
allow_none=True,
desc='Force creation or skipping a postamble map without vectors')
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
@staticmethod
def expressions():
return [
sdutil.node_path_graph(Vectorization._map_entry)
]
def can_be_applied(self, graph: SDFGState, candidate, expr_index, sdfg, strict=False):
map_entry = graph.nodes()[candidate[Vectorization._map_entry]]
# Only accept scopes that have one internal tasklet
scope = graph.scope_subgraph(map_entry, False, False)
if len(scope.nodes()) != 1:
return False
tasklet = scope.nodes()[0]
if not isinstance(tasklet, nodes.Tasklet):
return False
param = symbolic.pystr_to_symbolic(map_entry.map.params[-1])
found = False
# Strided maps cannot be vectorized
if map_entry.map.range[-1][2] != 1 and self.strided_map:
return False
# Check if all edges, adjacent to the tasklet,
# use the parameter in their contiguous dimension.
for e, conntype in graph.all_edges_and_connectors(tasklet):
# Cases that do not matter for vectorization
if e.data.data is None: # Empty memlets
continue
if isinstance(sdfg.arrays[e.data.data], data.Stream): # Streams
continue
# Vectorization can not be applied in WCR
# if e.data.wcr is not None:
# return False
subset = e.data.subset
array = sdfg.arrays[e.data.data]
# If already vectorized or a pointer, do not apply
if isinstance(conntype, (dtypes.vector, dtypes.pointer)):
return False
try:
for idx, expr in enumerate(subset):
if isinstance(expr, tuple):
for ex in expr:
ex = symbolic.pystr_to_symbolic(ex)
symbols = ex.free_symbols
if param in symbols:
if array.strides[idx] == 1:
found = True
else:
return False
else:
expr = symbolic.pystr_to_symbolic(expr)
symbols = expr.free_symbols
if param in symbols:
if array.strides[idx] == 1:
found = True
else:
return False
except TypeError: # cannot determine truth value of Relational
return False
return found
@staticmethod
def match_to_str(graph, candidate):
map_entry = candidate[Vectorization._map_entry]
return str(map_entry)
def apply(self, sdfg: SDFG):
graph = sdfg.nodes()[self.state_id]
map_entry = graph.nodes()[self.subgraph[Vectorization._map_entry]]
tasklet: nodes.Tasklet = graph.successors(map_entry)[0]
param = symbolic.pystr_to_symbolic(map_entry.map.params[-1])
# Create new vector size.
vector_size = self.vector_len
dim_from, dim_to, dim_skip = map_entry.map.range[-1]
# Determine whether to create preamble or postamble maps
if self.preamble is not None:
create_preamble = self.preamble
else:
create_preamble = not ((dim_from % vector_size == 0) == True
or dim_from == 0)
if self.postamble is not None:
create_postamble = self.postamble
else:
if isinstance(dim_to, symbolic.SymExpr):
create_postamble = (((dim_to.approx + 1) %
vector_size == 0) == False)
else:
create_postamble = (((dim_to + 1) % vector_size == 0) == False)
# Determine new range for vectorized map
if self.strided_map:
new_range = [dim_from, dim_to - vector_size + 1, vector_size]
else:
new_range = [
dim_from // vector_size, ((dim_to + 1) // vector_size) - 1,
dim_skip
]
# Create preamble non-vectorized map (replacing the original map)
if create_preamble:
old_scope = graph.scope_subgraph(map_entry, True, True)
new_scope: ScopeSubgraphView = replicate_scope(
sdfg, graph, old_scope)
new_begin = dim_from + (vector_size - (dim_from % vector_size))
map_entry.map.range[-1] = (dim_from, new_begin - 1, dim_skip)
# Replace map_entry with the replicated scope (so that the preamble
# will usually come first in topological sort)
map_entry = new_scope.entry
tasklet = new_scope.nodes()[old_scope.nodes().index(tasklet)]
new_range[0] = new_begin
# Create postamble non-vectorized map
if create_postamble:
new_scope: ScopeSubgraphView = replicate_scope(
sdfg, graph, graph.scope_subgraph(map_entry, True, True))
dim_to_ex = dim_to + 1
new_scope.entry.map.range[-1] = (dim_to_ex -
(dim_to_ex % vector_size), dim_to,
dim_skip)
# Change the step of the inner-most dimension.
map_entry.map.range[-1] = tuple(new_range)
# Vectorize connectors adjacent to the tasklet.
for edge in graph.all_edges(tasklet):
connectors = (tasklet.in_connectors
if edge.dst == tasklet else tasklet.out_connectors)
conn = edge.dst_conn if edge.dst == tasklet else edge.src_conn
if edge.data.data is None: # Empty memlets
continue
desc = sdfg.arrays[edge.data.data]
contigidx = desc.strides.index(1)
newlist = []
lastindex = edge.data.subset[contigidx]
if isinstance(lastindex, tuple):
newlist = [(rb, re, rs) for rb, re, rs in edge.data.subset]
symbols = set()
for indd in lastindex:
symbols.update(
symbolic.pystr_to_symbolic(indd).free_symbols)
else:
newlist = [(rb, rb, 1) for rb in edge.data.subset]
symbols = symbolic.pystr_to_symbolic(lastindex).free_symbols
oldtype = connectors[conn]
if oldtype is None or oldtype.type is None:
oldtype = desc.dtype
# Vector to scalar WCR edge: change connector and continue
lastedge = graph.memlet_path(edge)[-1]
if (lastedge.data.subset.num_elements() == 1
and edge.data.wcr is not None):
connectors[conn] = dtypes.vector(oldtype, vector_size)
continue
if str(param) not in map(str, symbols):
continue
# Vectorize connector, if not already vectorized
if isinstance(oldtype, dtypes.vector):
continue
connectors[conn] = dtypes.vector(oldtype, vector_size)
# Modify memlet subset to match vector length
if self.strided_map:
rb = newlist[contigidx][0]
if self.propagate_parent:
newlist[contigidx] = (rb / self.vector_len,
rb / self.vector_len, 1)
else:
newlist[contigidx] = (rb, rb + self.vector_len - 1, 1)
else:
rb = newlist[contigidx][0]
if self.propagate_parent:
newlist[contigidx] = (rb, rb, 1)
else:
newlist[contigidx] = (self.vector_len * rb,
self.vector_len * rb +
self.vector_len - 1, 1)
edge.data.subset = subsets.Range(newlist)
edge.data.volume = vector_size
# Vector length propagation using data descriptors, recursive traversal
# outwards
if self.propagate_parent:
for edge in graph.all_edges(tasklet):
cursdfg = sdfg
curedge = edge
while cursdfg is not None:
arrname = curedge.data.data
dtype = cursdfg.arrays[arrname].dtype
# Change type and shape to vector
if not isinstance(dtype, dtypes.vector):
cursdfg.arrays[arrname].dtype = dtypes.vector(
dtype, vector_size)
new_shape = list(cursdfg.arrays[arrname].shape)
contigidx = cursdfg.arrays[arrname].strides.index(1)
new_shape[contigidx] /= vector_size
try:
new_shape[contigidx] = int(new_shape[contigidx])
except TypeError:
pass
cursdfg.arrays[arrname].shape = new_shape
propagation.propagate_memlets_sdfg(cursdfg)
# Find matching edge in parent
nsdfg = cursdfg.parent_nsdfg_node
if nsdfg is None:
break
tstate = cursdfg.parent
curedge = ([
e
for e in tstate.in_edges(nsdfg) if e.dst_conn == arrname
] + [
e for e in tstate.out_edges(nsdfg)
if e.src_conn == arrname
])[0]
cursdfg = cursdfg.parent_sdfg
class TrivialMapElimination(transformation.Transformation):
""" Implements the Trivial-Map Elimination pattern.
Trivial-Map Elimination removes all dimensions containing only one
element from a map. If this applies to all ranges the map is removed.
Example: Map[i=0:I,j=7] -> Map[i=0:I]
Example: Map[i=0 ,j=7] -> nothing
"""
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
@staticmethod
def expressions():
return [sdutil.node_path_graph(TrivialMapElimination._map_entry)]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, permissive=False):
map_entry = graph.nodes()[candidate[TrivialMapElimination._map_entry]]
return any(r[0] == r[1] for r in map_entry.map.range)
@staticmethod
def match_to_str(graph, candidate):
map_entry = graph.nodes()[candidate[TrivialMapElimination._map_entry]]
return map_entry.map.label + ': ' + str(map_entry.map.params)
def apply(self, sdfg):
graph = sdfg.nodes()[self.state_id]
map_entry = graph.nodes()[self.subgraph[
TrivialMapElimination._map_entry]]
map_exit = graph.exit_node(map_entry)
remaining_ranges = []
remaining_params = []
for map_param, ranges in zip(map_entry.map.params,
map_entry.map.range.ranges):
map_from, map_to, _ = ranges
if map_from == map_to:
# Replace the map index variable with the value it obtained
scope = graph.scope_subgraph(map_entry)
scope.replace(map_param, map_from)
else:
remaining_ranges.append(ranges)
remaining_params.append(map_param)
map_entry.map.range.ranges = remaining_ranges
map_entry.map.params = remaining_params
if len(remaining_ranges) == 0:
# Redirect map entry's out edges
for edge in graph.out_edges(map_entry):
path = graph.memlet_path(edge)
index = path.index(edge)
# Add an edge directly from the previous source connector to the destination
graph.add_edge(path[index - 1].src, path[index - 1].src_conn,
edge.dst, edge.dst_conn, edge.data)
# Redirect map exit's in edges.
for edge in graph.in_edges(map_exit):
path = graph.memlet_path(edge)
index = path.index(edge)
# Add an edge directly from the source to the next destination connector
if len(path) > index + 1:
graph.add_edge(edge.src, edge.src_conn,
path[index + 1].dst,
path[index + 1].dst_conn, edge.data)
# Remove map
graph.remove_nodes_from([map_entry, map_exit])
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
class MPITransformMap(transformation.Transformation):
""" Implements the MPI parallelization pattern.
Takes a map and makes it an MPI-scheduled map, introduces transients
that keep locally accessed data.
Original SDFG
=============
```
Input1 - Output1
\ /
Input2 --- MapEntry -- Arbitrary R -- MapExit -- Output2
/ \
InputN - OutputN
```
Nothing in R may access other inputs/outputs that are not defined in R
itself and do not go through MapEntry/MapExit
Map must be a one-dimensional map for now.
The range of the map must be a Range object.
Output:
=======
* Add transients for the accessed parts
* The schedule property of Map is set to MPI
* The range of Map is changed to
var = startexpr + p * chunksize ... startexpr + p + 1 * chunksize
where p is the current rank and P is the total number of ranks,
and chunksize is defined as (endexpr - startexpr) / P, adding the
remaining K iterations to the first K procs.
* For each input InputI, create a new transient transInputI, which
has an attribute that specifies that it needs to be filled with
(possibly) remote data
* Collect all accesses to InputI within R, assume their convex hull is
InputI[rs ... re]
* The transInputI transient will contain InputI[rs ... re]
* Change all accesses to InputI within R to accesses to transInputI
"""
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
@staticmethod
def annotates_memlets():
return True
@staticmethod
def expressions():
return [sdutil.node_path_graph(MPITransformMap._map_entry)]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
map_entry = graph.nodes()[candidate[MPITransformMap._map_entry]]
# Check if the map is one-dimensional
if map_entry.map.range.dims() != 1:
return False
# We cannot transform a map which is already of schedule type MPI
if map_entry.map.schedule == dtypes.ScheduleType.MPI:
return False
# We cannot transform a map which is already inside a MPI map, or in
# another device
schedule_whitelist = [
dtypes.ScheduleType.Default, dtypes.ScheduleType.Sequential
]
sdict = graph.scope_dict()
parent = sdict[map_entry]
while parent is not None:
if parent.map.schedule not in schedule_whitelist:
return False
parent = sdict[parent]
# Dynamic map ranges not supported (will allocate dynamic memory)
if has_dynamic_map_inputs(graph, map_entry):
return False
# MPI schedules currently do not support WCR
map_exit = graph.exit_node(map_entry)
if any(e.data.wcr for e in graph.out_edges(map_exit)):
return False
return True
@staticmethod
def match_to_str(graph, candidate):
map_entry = graph.nodes()[candidate[MPITransformMap._map_entry]]
return map_entry.map.label
def apply(self, sdfg):
graph = sdfg.nodes()[self.state_id]
map_entry = graph.nodes()[self.subgraph[MPITransformMap._map_entry]]
# Avoiding import loops
from dace.transformation.dataflow.strip_mining import StripMining
from dace.transformation.dataflow.local_storage import LocalStorage
rangeexpr = str(map_entry.map.range.num_elements())
stripmine_subgraph = {
StripMining._map_entry: self.subgraph[MPITransformMap._map_entry]
}
sdfg_id = sdfg.sdfg_id
stripmine = StripMining(sdfg_id, self.state_id, stripmine_subgraph,
self.expr_index)
stripmine.dim_idx = -1
stripmine.new_dim_prefix = "mpi"
stripmine.tile_size = "(" + rangeexpr + "/__dace_comm_size)"
stripmine.divides_evenly = True
stripmine.apply(sdfg)
# Find all in-edges that lead to candidate[MPITransformMap._map_entry]
outer_map = None
edges = [
e for e in graph.in_edges(map_entry)
if isinstance(e.src, nodes.EntryNode)
]
outer_map = edges[0].src
# Add MPI schedule attribute to outer map
outer_map.map._schedule = dtypes.ScheduleType.MPI
# Now create a transient for each array
for e in edges:
in_local_storage_subgraph = {
LocalStorage.node_a: graph.node_id(outer_map),
LocalStorage.node_b: self.subgraph[MPITransformMap._map_entry]
}
sdfg_id = sdfg.sdfg_id
in_local_storage = LocalStorage(sdfg_id, self.state_id,
in_local_storage_subgraph,
self.expr_index)
in_local_storage.array = e.data.data
in_local_storage.apply(sdfg)
# Transform OutLocalStorage for each output of the MPI map
in_map_exit = graph.exit_node(map_entry)
out_map_exit = graph.exit_node(outer_map)
for e in graph.out_edges(out_map_exit):
name = e.data.data
outlocalstorage_subgraph = {
LocalStorage.node_a: graph.node_id(in_map_exit),
LocalStorage.node_b: graph.node_id(out_map_exit)
}
sdfg_id = sdfg.sdfg_id
outlocalstorage = LocalStorage(sdfg_id, self.state_id,
outlocalstorage_subgraph,
self.expr_index)
outlocalstorage.array = name
outlocalstorage.apply(sdfg)
class MapUnroll(transformation.Transformation):
"""
Unrolls a map with constant ranges in the top-level scope of an SDFG by
replicating its subgraph for each iteration. If there are local data
containers only used in this map, they will also be replicated, as will
nested SDFGs found within.
This transformation can be useful for forming weakly connected components
that will be inferred as processing elements in an FPGA kernel.
"""
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
@staticmethod
def expressions():
return [sdutil.node_path_graph(MapUnroll._map_entry)]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
map_entry = graph.nodes()[candidate[MapUnroll._map_entry]]
# Must be top-level map
if graph.scope_dict()[map_entry] is not None:
return False
# All map ranges must be constant
try:
for begin, end, step in map_entry.map.range:
symbolic.evaluate(begin, sdfg.constants)
symbolic.evaluate(end, sdfg.constants)
symbolic.evaluate(step, sdfg.constants)
except TypeError:
return False
return True
@staticmethod
def match_to_str(graph, candidate):
map_entry = graph.nodes()[candidate[MapUnroll._map_entry]]
return map_entry.map.label + ': ' + str(map_entry.map.params)
def apply(self, sdfg):
from dace.transformation.dataflow import TrivialMapElimination
state = sdfg.nodes()[self.state_id]
map_entry = state.nodes()[self.subgraph[MapUnroll._map_entry]]
map_exit = state.exit_node(map_entry)
# Collect all nodes in this weakly connected component
subgraph = sdutil.weakly_connected_component(state, map_entry)
# Save nested SDFGs to JSON, then deserialize them for every copy we
# need to make
nested_sdfgs = {}
for node in subgraph:
if isinstance(node, nodes.NestedSDFG):
nested_sdfgs[node.sdfg] = node.sdfg.to_json()
# Check for local memories that need to be replicated
local_memories = [
name for name in sdutil.local_transients(
sdfg, subgraph, entry_node=map_entry, include_nested=True)
if not isinstance(sdfg.arrays[name], dt.Stream)
and not isinstance(sdfg.arrays[name], dt.View)
]
params = map_entry.map.params
ranges = map_entry.map.range.ranges
constant_ranges = []
for r in ranges:
begin = symbolic.evaluate(r[0], sdfg.constants)
end = symbolic.evaluate(r[1], sdfg.constants)
step = symbolic.evaluate(r[2], sdfg.constants)
end += step # Make non-inclusive
constant_ranges.append(range(begin, end, step))
index_tuples = itertools.product(*constant_ranges)
for t in index_tuples:
suffix = "_" + "_".join(map(str, t))
node_to_unrolled = {}
# Copy all nodes
for node in subgraph:
if isinstance(node, nodes.NestedSDFG):
# Avoid deep-copying the nested SDFG
nsdfg = node.sdfg
# Don't copy the nested SDFG, as we will do this separately
node.sdfg = None
unrolled_node = copy.deepcopy(node)
node.sdfg = nsdfg
# Deserialize into a new SDFG specific to this copy
nsdfg_json = nested_sdfgs[nsdfg]
name = nsdfg_json["attributes"]["name"]
nsdfg_json["attributes"]["name"] += suffix
unrolled_nsdfg = SDFG.from_json(nsdfg_json)
nsdfg_json["attributes"]["name"] = name # Reinstate
# Set all the references
unrolled_nsdfg.parent = state
unrolled_nsdfg.parent_sdfg = sdfg
unrolled_nsdfg.update_sdfg_list([])
unrolled_node.sdfg = unrolled_nsdfg
unrolled_nsdfg.parent_nsdfg_node = unrolled_node
else:
unrolled_node = copy.deepcopy(node)
if node == map_entry:
# Fix the map bounds to only this iteration
unrolled_node.map.range = [(i, i, 1) for i in t]
if (isinstance(node, nodes.AccessNode)
and node.data in local_memories):
# If this is a local memory only used in this subgraph,
# we need to replicate it for each new subgraph
unrolled_name = node.data + suffix
if unrolled_name not in sdfg.arrays:
unrolled_desc = copy.deepcopy(
sdfg.arrays[node.data])
sdfg.add_datadesc(unrolled_name, unrolled_desc)
unrolled_node.data = unrolled_name
state.add_node(unrolled_node)
node_to_unrolled[node] = unrolled_node # Remember mapping
# Copy all edges
for src, src_conn, dst, dst_conn, memlet in subgraph.edges():
src = node_to_unrolled[src]
dst = node_to_unrolled[dst]
memlet = copy.deepcopy(memlet)
if memlet.data in local_memories:
memlet.data = memlet.data + suffix
state.add_edge(src, src_conn, dst, dst_conn, memlet)
# Eliminate the now trivial map
TrivialMapElimination.apply_to(
sdfg,
verify=False,
annotate=False,
save=False,
_map_entry=node_to_unrolled[map_entry])
# Now we can delete the original subgraph. This implicitly also remove
# memlets between nodes
state.remove_nodes_from(subgraph)
# If we added a bunch of new nested SDFGs, reset the internal list
if len(nested_sdfgs) > 0:
sdfg.reset_sdfg_list()
# Remove local memories that were replicated
for mem in local_memories:
sdfg.remove_data(mem)
class AccumulateTransient(pattern_matching.Transformation):
""" Implements the AccumulateTransient transformation, which adds
transient stream and data nodes between nested maps that lead to a
stream. The transient data nodes then act as a local accumulator.
"""
_tasklet = nodes.Tasklet('_')
_map_exit = nodes.MapExit(nodes.Map("", [], []))
_outer_map_exit = nodes.MapExit(nodes.Map("", [], []))
array = Property(
dtype=str,
desc="Array to create local storage for (if empty, first available)",
default=None,
allow_none=True)
@staticmethod
def expressions():
return [
sdutil.node_path_graph(AccumulateTransient._tasklet,
AccumulateTransient._map_exit,
AccumulateTransient._outer_map_exit)
]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
tasklet = graph.nodes()[candidate[AccumulateTransient._tasklet]]
map_exit = graph.nodes()[candidate[AccumulateTransient._map_exit]]
# Check if there is an accumulation output
for _src, _, dest, _, memlet in graph.out_edges(tasklet):
if memlet.wcr is not None and dest == map_exit:
return True
return False
@staticmethod
def match_to_str(graph, candidate):
tasklet = candidate[AccumulateTransient._tasklet]
map_exit = candidate[AccumulateTransient._map_exit]
outer_map_exit = candidate[AccumulateTransient._outer_map_exit]
return ' -> '.join(
str(node) for node in [tasklet, map_exit, outer_map_exit])
def apply(self, sdfg):
graph = sdfg.node(self.state_id)
# Choose array
array = self.array
if array is None or len(array) == 0:
map_exit = graph.node(self.subgraph[AccumulateTransient._map_exit])
outer_map_exit = graph.node(
self.subgraph[AccumulateTransient._outer_map_exit])
array = next(e.data.data
for e in graph.edges_between(map_exit, outer_map_exit)
if e.data.wcr is not None)
# Avoid import loop
from dace.transformation.dataflow.local_storage import LocalStorage
local_storage_subgraph = {
LocalStorage._node_a:
self.subgraph[AccumulateTransient._map_exit],
LocalStorage._node_b:
self.subgraph[AccumulateTransient._outer_map_exit]
}
sdfg_id = sdfg.sdfg_list.index(sdfg)
in_local_storage = LocalStorage(sdfg_id, self.state_id,
local_storage_subgraph,
self.expr_index)
in_local_storage.array = array
in_local_storage.apply(sdfg)
# Initialize transient to zero in case of summation
# TODO: Initialize transient in other WCR types
memlet = graph.in_edges(in_local_storage._data_node)[0].data
if detect_reduction_type(memlet.wcr) == dtypes.ReductionType.Sum:
in_local_storage._data_node.setzero = True
else:
warnings.warn('AccumulateTransient did not properly initialize'
'newly-created transient!')
def fuse(self, sdfg, graph, map_entries, do_not_override=None, **kwargs):
""" takes the map_entries specified and tries to fuse maps.
all maps have to be extended into outer and inner map
(use MapExpansion as a pre-pass)
Arrays that don't exist outside the subgraph get pushed
into the map and their data dimension gets cropped.
Otherwise the original array is taken.
For every output respective connections are crated automatically.
:param sdfg: SDFG
:param graph: State
:param map_entries: Map Entries (class MapEntry) of the outer maps
which we want to fuse
:param do_not_override: List of data names whose corresponding nodes
are fully contained within the subgraph
but should not be augmented/transformed
nevertheless.
"""
# if there are no maps, return immediately
if len(map_entries) == 0:
return
do_not_override = do_not_override or []
# get maps and map exits
maps = [map_entry.map for map_entry in map_entries]
map_exits = [graph.exit_node(map_entry) for map_entry in map_entries]
# See function documentation for an explanation of these variables
node_config = SubgraphFusion.get_adjacent_nodes(sdfg, graph,
map_entries)
(in_nodes, intermediate_nodes, out_nodes) = node_config
if self.debug:
print("SubgraphFusion::In_nodes", in_nodes)
print("SubgraphFusion::Out_nodes", out_nodes)
print("SubgraphFusion::Intermediate_nodes", intermediate_nodes)
# all maps are assumed to have the same params and range in order
global_map = nodes.Map(label="outer_fused",
params=maps[0].params,
ndrange=maps[0].range)
global_map_entry = nodes.MapEntry(global_map)
global_map_exit = nodes.MapExit(global_map)
schedule = map_entries[0].schedule
global_map_entry.schedule = schedule
graph.add_node(global_map_entry)
graph.add_node(global_map_exit)
# next up, for any intermediate node, find whether it only appears
# in the subgraph or also somewhere else / as an input
# create new transients for nodes that are in out_nodes and
# intermediate_nodes simultaneously
# also check which dimensions of each transient data element correspond
# to map axes and write this information into a dict.
node_info = self.prepare_intermediate_nodes(sdfg, graph, in_nodes, out_nodes, \
intermediate_nodes,\
map_entries, map_exits, \
do_not_override)
(subgraph_contains_data, transients_created,
invariant_dimensions) = node_info
if self.debug:
print(
"SubgraphFusion:: {Intermediate_node: subgraph_contains_data} dict"
)
print(subgraph_contains_data)
inconnectors_dict = {}
# Dict for saving incoming nodes and their assigned connectors
# Format: {access_node: (edge, in_conn, out_conn)}
for map_entry, map_exit in zip(map_entries, map_exits):
# handle inputs
# TODO: dynamic map range -- this is fairly unrealistic in such a setting
for edge in graph.in_edges(map_entry):
src = edge.src
mmt = graph.memlet_tree(edge)
out_edges = [child.edge for child in mmt.root().children]
if src in in_nodes:
in_conn = None
out_conn = None
if src in inconnectors_dict:
# no need to augment subset of outer edge.
# will do this at the end in one pass.
in_conn = inconnectors_dict[src][1]
out_conn = inconnectors_dict[src][2]
else:
next_conn = global_map_entry.next_connector()
in_conn = 'IN_' + next_conn
out_conn = 'OUT_' + next_conn
global_map_entry.add_in_connector(in_conn)
global_map_entry.add_out_connector(out_conn)
inconnectors_dict[src] = (edge, in_conn, out_conn)
# reroute in edge via global_map_entry
self.copy_edge(graph, edge, new_dst = global_map_entry, \
new_dst_conn = in_conn)
# map out edges to new map
for out_edge in out_edges:
self.copy_edge(graph, out_edge, new_src = global_map_entry, \
new_src_conn = out_conn)
else:
# connect directly
for out_edge in out_edges:
mm = dcpy(out_edge.data)
self.copy_edge(graph,
out_edge,
new_src=src,
new_src_conn=None,
new_data=mm)
for edge in graph.out_edges(map_entry):
# special case: for nodes that have no data connections
if not edge.src_conn:
self.copy_edge(graph, edge, new_src=global_map_entry)
######################################
for edge in graph.in_edges(map_exit):
if not edge.dst_conn:
# no destination connector, path ends here.
self.copy_edge(graph, edge, new_dst=global_map_exit)
continue
# find corresponding out_edges for current edge, cannot use mmt anymore
out_edges = [
oedge for oedge in graph.out_edges(map_exit)
if oedge.src_conn[3:] == edge.dst_conn[2:]
]
# Tuple to store in/out connector port that might be created
port_created = None
for out_edge in out_edges:
dst = out_edge.dst
if dst in intermediate_nodes & out_nodes:
# create connection through global map from
# dst to dst_transient that was created
dst_transient = transients_created[dst]
next_conn = global_map_exit.next_connector()
in_conn = 'IN_' + next_conn
out_conn = 'OUT_' + next_conn
global_map_exit.add_in_connector(in_conn)
global_map_exit.add_out_connector(out_conn)
# for each transient created, create a union
# of outgoing memlets' subsets. this is
# a cheap fix to override assignments in invariant
# dimensions
union = None
for oe in graph.out_edges(transients_created[dst]):
union = subsets.union(union, oe.data.subset)
inner_memlet = dcpy(edge.data)
for i, s in enumerate(edge.data.subset):
if i in invariant_dimensions[dst.label]:
inner_memlet.subset[i] = union[i]
inner_memlet.other_subset = dcpy(inner_memlet.subset)
e_inner = graph.add_edge(dst, None, global_map_exit,
in_conn, inner_memlet)
mm_outer = propagate_memlet(graph, inner_memlet, global_map_entry, \
union_inner_edges = False)
e_outer = graph.add_edge(global_map_exit, out_conn,
dst_transient, None, mm_outer)
# remove edge from dst to dst_transient that was created
# in intermediate preparation.
for e in graph.out_edges(dst):
if e.dst == dst_transient:
graph.remove_edge(e)
break
# handle separately: intermediate_nodes and pure out nodes
# case 1: intermediate_nodes: can just redirect edge
if dst in intermediate_nodes:
self.copy_edge(graph,
out_edge,
new_src=edge.src,
new_src_conn=edge.src_conn,
new_data=dcpy(edge.data))
# case 2: pure out node: connect to outer array node
if dst in (out_nodes - intermediate_nodes):
if edge.dst != global_map_exit:
next_conn = global_map_exit.next_connector()
in_conn = 'IN_' + next_conn
out_conn = 'OUT_' + next_conn
global_map_exit.add_in_connector(in_conn)
global_map_exit.add_out_connector(out_conn)
self.copy_edge(graph,
edge,
new_dst=global_map_exit,
new_dst_conn=in_conn)
port_created = (in_conn, out_conn)
else:
conn_nr = edge.dst_conn[3:]
in_conn = port_created.st
out_conn = port_created.nd
# map
graph.add_edge(global_map_exit, out_conn, dst, None,
dcpy(out_edge.data))
# maps are now ready to be discarded
# all connected edges will be finally removed as well
graph.remove_node(map_entry)
graph.remove_node(map_exit)
# create a mapping from data arrays to offsets
# for later memlet adjustments later
min_offsets = dict()
# do one pass to augment all transient arrays
data_intermediate = set([node.data for node in intermediate_nodes])
for data_name in data_intermediate:
if subgraph_contains_data[data_name]:
all_nodes = [
n for n in intermediate_nodes if n.data == data_name
]
in_edges = list(chain(*(graph.in_edges(n) for n in all_nodes)))
in_edges_iter = iter(in_edges)
in_edge = next(in_edges_iter)
target_subset = dcpy(in_edge.data.subset)
target_subset.pop(invariant_dimensions[data_name])
######
while True:
try: # executed if there are multiple in_edges
in_edge = next(in_edges_iter)
target_subset_curr = dcpy(in_edge.data.subset)
target_subset_curr.pop(invariant_dimensions[data_name])
target_subset = subsets.union(target_subset, \
target_subset_curr)
except StopIteration:
break
min_offsets_cropped = target_subset.min_element_approx()
# calculate the new transient array size.
target_subset.offset(min_offsets_cropped, True)
# re-add invariant dimensions with offset 0 and save to min_offsets
min_offset = []
index = 0
for i in range(len(sdfg.data(data_name).shape)):
if i in invariant_dimensions[data_name]:
min_offset.append(0)
else:
min_offset.append(min_offsets_cropped[index])
index += 1
min_offsets[data_name] = min_offset
# determine the shape of the new array.
new_data_shape = []
index = 0
for i, sz in enumerate(sdfg.data(data_name).shape):
if i in invariant_dimensions[data_name]:
new_data_shape.append(sz)
else:
new_data_shape.append(target_subset.size()[index])
index += 1
new_data_strides = [
data._prod(new_data_shape[i + 1:])
for i in range(len(new_data_shape))
]
new_data_totalsize = data._prod(new_data_shape)
new_data_offset = [0] * len(new_data_shape)
# augment.
transient_to_transform = sdfg.data(data_name)
transient_to_transform.shape = new_data_shape
transient_to_transform.strides = new_data_strides
transient_to_transform.total_size = new_data_totalsize
transient_to_transform.offset = new_data_offset
transient_to_transform.lifetime = dtypes.AllocationLifetime.Scope
transient_to_transform.storage = self.transient_allocation
else:
# don't modify data container - array is needed outside
# of subgraph.
# hack: set lifetime to State if allocation has only been
# scope so far to avoid allocation issues
if sdfg.data(
data_name).lifetime == dtypes.AllocationLifetime.Scope:
sdfg.data(
data_name).lifetime = dtypes.AllocationLifetime.State
# do one pass to adjust and the memlets of in-between transients
for node in intermediate_nodes:
# all incoming edges to node
in_edges = graph.in_edges(node)
# outgoing edges going to another fused part
out_edges = graph.out_edges(node)
# memlets of created transient:
# correct data names
if node in transients_created:
transient_in_edges = graph.in_edges(transients_created[node])
transient_out_edges = graph.out_edges(transients_created[node])
for edge in chain(transient_in_edges, transient_out_edges):
for e in graph.memlet_tree(edge):
if e.data.data == node.data:
e.data.data += '_OUT'
# memlets of all in between transients:
# offset memlets if array has been augmented
if subgraph_contains_data[node.data]:
# get min_offset
min_offset = min_offsets[node.data]
# re-add invariant dimensions with offset 0
for iedge in in_edges:
for edge in graph.memlet_tree(iedge):
if edge.data.data == node.data:
edge.data.subset.offset(min_offset, True)
elif edge.data.other_subset:
edge.data.other_subset.offset(min_offset, True)
# nested SDFG: adjust arrays connected
if isinstance(iedge.src, nodes.NestedSDFG):
nsdfg = iedge.src.sdfg
nested_data_name = edge.src_conn
self.adjust_arrays_nsdfg(sdfg, nsdfg, node.data,
nested_data_name)
for cedge in out_edges:
for edge in graph.memlet_tree(cedge):
if edge.data.data == node.data:
edge.data.subset.offset(min_offset, True)
elif edge.data.other_subset:
edge.data.other_subset.offset(min_offset, True)
# nested SDFG: adjust arrays connected
if isinstance(edge.dst, nodes.NestedSDFG):
nsdfg = edge.dst.sdfg
nested_data_name = edge.dst_conn
self.adjust_arrays_nsdfg(sdfg, nsdfg, node.data,
nested_data_name)
# if in_edges has several entries:
# put other_subset into out_edges for correctness
if len(in_edges) > 1:
for oedge in out_edges:
if oedge.dst == global_map_exit and \
oedge.data.other_subset is None:
oedge.data.other_subset = dcpy(oedge.data.subset)
oedge.data.other_subset.offset(min_offset, True)
# consolidate edges if desired
if self.consolidate:
consolidate_edges_scope(graph, global_map_entry)
consolidate_edges_scope(graph, global_map_exit)
# propagate edges adjacent to global map entry and exit
# if desired
if self.propagate:
_propagate_node(graph, global_map_entry)
_propagate_node(graph, global_map_exit)
# create a hook for outside access to global_map
self._global_map_entry = global_map_entry
if self.schedule_innermaps is not None:
for node in graph.scope_children()[global_map_entry]:
if isinstance(node, nodes.MapEntry):
node.map.schedule = self.schedule_innermaps
class StreamTransient(pattern_matching.Transformation):
""" Implements the StreamTransient transformation, which adds a transient
and stream nodes between nested maps that lead to a stream. The
transient then acts as a local buffer.
"""
with_buffer = Property(dtype=bool,
default=True,
desc="Use an intermediate buffer for accumulation")
_tasklet = nodes.Tasklet('_')
_map_exit = nodes.MapExit(nodes.Map("", [], []))
_outer_map_exit = nodes.MapExit(nodes.Map("", [], []))
@staticmethod
def expressions():
return [
sdutil.node_path_graph(StreamTransient._tasklet,
StreamTransient._map_exit,
StreamTransient._outer_map_exit)
]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
map_exit = graph.nodes()[candidate[StreamTransient._map_exit]]
outer_map_exit = graph.nodes()[candidate[
StreamTransient._outer_map_exit]]
# Check if there is a streaming output
for _src, _, dest, _, memlet in graph.out_edges(map_exit):
if isinstance(sdfg.arrays[memlet.data],
data.Stream) and dest == outer_map_exit:
return True
return False
@staticmethod
def match_to_str(graph, candidate):
tasklet = candidate[StreamTransient._tasklet]
map_exit = candidate[StreamTransient._map_exit]
outer_map_exit = candidate[StreamTransient._outer_map_exit]
return ' -> '.join(
str(node) for node in [tasklet, map_exit, outer_map_exit])
def apply(self, sdfg: SDFG):
graph = sdfg.nodes()[self.state_id]
tasklet = graph.nodes()[self.subgraph[StreamTransient._tasklet]]
map_exit = graph.nodes()[self.subgraph[StreamTransient._map_exit]]
outer_map_exit = graph.nodes()[self.subgraph[
StreamTransient._outer_map_exit]]
memlet = None
edge = None
for e in graph.out_edges(map_exit):
memlet = e.data
# TODO: What if there's more than one?
if e.dst == outer_map_exit and isinstance(sdfg.arrays[memlet.data],
data.Stream):
edge = e
break
tasklet_memlet = None
for e in graph.out_edges(tasklet):
tasklet_memlet = e.data
if tasklet_memlet.data == memlet.data:
break
bbox = map_exit.map.range.bounding_box_size()
bbox_approx = [symbolic.overapproximate(dim) for dim in bbox]
dataname = memlet.data
# Create the new node: Temporary stream and an access node
newname, _ = sdfg.add_stream('trans_' + dataname,
sdfg.arrays[memlet.data].dtype,
1,
bbox_approx[0], [1],
transient=True,
find_new_name=True)
snode = graph.add_access(newname)
to_stream_mm = copy.deepcopy(memlet)
to_stream_mm.data = snode.data
tasklet_memlet.data = snode.data
if self.with_buffer:
newname_arr, _ = sdfg.add_transient('strans_' + dataname,
[bbox_approx[0]],
sdfg.arrays[memlet.data].dtype,
find_new_name=True)
anode = graph.add_access(newname_arr)
to_array_mm = copy.deepcopy(memlet)
to_array_mm.data = anode.data
graph.add_edge(snode, None, anode, None, to_array_mm)
else:
anode = snode
# Reconnect, assuming one edge to the stream
graph.remove_edge(edge)
graph.add_edge(map_exit, edge.src_conn, snode, None, to_stream_mm)
graph.add_edge(anode, None, outer_map_exit, edge.dst_conn, memlet)
return
def modifies_graph(self):
return True
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)
class GPUMultiTransformMap(transformation.Transformation):
""" Implements the GPUMultiTransformMap transformation.
Tiles a single map into 2 maps. The outer map is of schedule
GPU_Multidevice and loops over the GPUs, while the inner map
is a GPU-scheduled map. It also creates GPU transient arrays
between the two maps.
"""
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
dim_idx = Property(dtype=int,
default=-1,
desc="Index of dimension to be distributed.")
new_dim_prefix = Property(dtype=str,
default="gpu",
allow_none=True,
desc="Prefix for new dimension name")
new_transient_prefix = Property(dtype=str,
default="gpu_multi",
allow_none=True,
desc="Prefix for the transient name")
skip_scalar = Property(
dtype=bool,
default=True,
allow_none=True,
desc="If True: skips the scalar data nodes. "
"If False: creates localstorage for scalar transients.")
use_p2p = Property(
dtype=bool,
default=False,
allow_none=True,
desc="If True: uses peer-to-peer access if a data container is already "
"located on a GPU. "
"If False: creates transient localstorage for data located on GPU.")
number_of_gpus = SymbolicProperty(
default=None,
allow_none=True,
desc="number of gpus to divide the map onto,"
" if not used, uses the amount specified"
" in the dace.config in max_number_gpus.")
@staticmethod
def annotates_memlets():
return True
@staticmethod
def expressions():
return [sdutil.node_path_graph(GPUMultiTransformMap._map_entry)]
@staticmethod
def can_be_applied(graph: SDFGState,
candidate,
expr_index,
sdfg,
strict=False):
map_entry = graph.nodes()[candidate[GPUMultiTransformMap._map_entry]]
# Check if there is more than one GPU available:
if (Config.get("compiler", "cuda", "max_number_gpus") < 2):
return False
# Dynamic map ranges not supported
if has_dynamic_map_inputs(graph, map_entry):
return False
# Only accept maps with a default schedule
schedule_whitelist = [dtypes.ScheduleType.Default]
sdict = graph.scope_dict()
parent = sdict[map_entry]
while parent is not None:
if parent.map.schedule not in schedule_whitelist:
return False
parent = sdict[parent]
# Library nodes inside the scope are not supported
scope_subgraph = graph.scope_subgraph(map_entry)
for node in scope_subgraph.nodes():
if isinstance(node, nodes.LibraryNode):
return False
# Custom reductions can not have an accumulate transient, as the
# reduction would have to be split up for the ingoing memlet of the
# accumulate transient and the outgoing memlet. Not using GPU local
# accumulate transient only works for a small volume of data.
map_exit = graph.exit_node(map_entry)
for edge in graph.out_edges(map_exit):
if edge.data.wcr is not None and operations.detect_reduction_type(
edge.data.wcr) == dtypes.ReductionType.Custom:
return False
storage_whitelist = [
dtypes.StorageType.Default,
dtypes.StorageType.CPU_Pinned,
dtypes.StorageType.CPU_Heap,
dtypes.StorageType.GPU_Global,
]
for node in graph.predecessors(map_entry):
if not isinstance(node, nodes.AccessNode):
return False
if node.desc(graph).storage not in storage_whitelist:
return False
for node in graph.successors(map_exit):
if not isinstance(node, nodes.AccessNode):
return False
if node.desc(graph).storage not in storage_whitelist:
return False
return True
@staticmethod
def match_to_str(graph, candidate):
map_entry = graph.nodes()[candidate[GPUMultiTransformMap._map_entry]]
return map_entry.map.label
def apply(self, sdfg: SDFG) -> None:
graph: SDFGState = sdfg.nodes()[self.state_id]
inner_map_entry: nodes.MapEntry = graph.nodes()[self.subgraph[
GPUMultiTransformMap._map_entry]]
number_of_gpus = self.number_of_gpus
ngpus = Config.get("compiler", "cuda", "max_number_gpus")
if (number_of_gpus == None):
number_of_gpus = ngpus
if number_of_gpus > ngpus:
raise ValueError(
'Requesting more gpus than specified in the dace config')
# Avoiding import loops
from dace.transformation.dataflow import (StripMining, InLocalStorage,
OutLocalStorage,
AccumulateTransient)
# The user has responsibility for the implementation of a Library node.
scope_subgraph = graph.scope_subgraph(inner_map_entry)
for node in scope_subgraph.nodes():
if isinstance(node, nodes.LibraryNode):
warnings.warn(
'Node %s is a library node, make sure to manually set the '
'implementation to a GPU compliant specialization.' % node)
# Tile map into number_of_gpus tiles
outer_map: nodes.Map = StripMining.apply_to(
sdfg,
dict(dim_idx=-1,
new_dim_prefix=self.new_dim_prefix,
tile_size=number_of_gpus,
tiling_type=dtypes.TilingType.NumberOfTiles),
_map_entry=inner_map_entry)
outer_map_entry: nodes.MapEntry = graph.scope_dict()[inner_map_entry]
inner_map_exit: nodes.MapExit = graph.exit_node(inner_map_entry)
outer_map_exit: nodes.MapExit = graph.exit_node(outer_map_entry)
# Change map schedules
inner_map_entry.map.schedule = dtypes.ScheduleType.GPU_Device
outer_map.schedule = dtypes.ScheduleType.GPU_Multidevice
symbolic_gpu_id = outer_map.params[0]
# Add the parameter of the outer map
for node in graph.successors(inner_map_entry):
if isinstance(node, nodes.NestedSDFG):
map_syms = inner_map_entry.range.free_symbols
for sym in map_syms:
symname = str(sym)
if symname not in node.symbol_mapping.keys():
node.symbol_mapping[symname] = sym
node.sdfg.symbols[symname] = graph.symbols_defined_at(
node)[symname]
# Add transient Data leading to the inner map
prefix = self.new_transient_prefix
for node in graph.predecessors(outer_map_entry):
# Only AccessNodes are relevant
if (isinstance(node, nodes.AccessNode)
and not (self.skip_scalar
and isinstance(node.desc(sdfg), Scalar))):
if self.use_p2p and node.desc(
sdfg).storage is dtypes.StorageType.GPU_Global:
continue
in_data_node = InLocalStorage.apply_to(sdfg,
dict(array=node.data,
prefix=prefix),
verify=False,
save=False,
node_a=outer_map_entry,
node_b=inner_map_entry)
in_data_node.desc(sdfg).location['gpu'] = symbolic_gpu_id
in_data_node.desc(sdfg).storage = dtypes.StorageType.GPU_Global
wcr_data: Dict[str, Any] = {}
# Add transient Data leading to the outer map
for edge in graph.in_edges(outer_map_exit):
node = graph.memlet_path(edge)[-1].dst
if isinstance(node, nodes.AccessNode):
data_name = node.data
# Transients with write-conflict resolution need to be
# collected first as AccumulateTransient creates a nestedSDFG
if edge.data.wcr is not None:
dtype = sdfg.arrays[data_name].dtype
redtype = operations.detect_reduction_type(edge.data.wcr)
# Custom reduction can not have an accumulate transient,
# as the accumulation from the transient to the outer
# storage is not defined.
if redtype == dtypes.ReductionType.Custom:
warnings.warn(
'Using custom reductions in a GPUMultitransformed '
'Map only works for a small data volume. For large '
'volume there is no guarantee.')
continue
identity = dtypes.reduction_identity(dtype, redtype)
wcr_data[data_name] = identity
elif (not isinstance(node.desc(sdfg), Scalar)
or not self.skip_scalar):
if self.use_p2p and node.desc(
sdfg).storage is dtypes.StorageType.GPU_Global:
continue
# Transients without write-conflict resolution
if prefix + '_' + data_name in sdfg.arrays:
create_array = False
else:
create_array = True
out_data_node = OutLocalStorage.apply_to(
sdfg,
dict(array=data_name,
prefix=prefix,
create_array=create_array),
verify=False,
save=False,
node_a=inner_map_exit,
node_b=outer_map_exit)
out_data_node.desc(sdfg).location['gpu'] = symbolic_gpu_id
out_data_node.desc(
sdfg).storage = dtypes.StorageType.GPU_Global
# Add Transients for write-conflict resolution
if len(wcr_data) != 0:
nsdfg = AccumulateTransient.apply_to(
sdfg,
options=dict(array_identity_dict=wcr_data, prefix=prefix),
map_exit=inner_map_exit,
outer_map_exit=outer_map_exit)
nsdfg.schedule = dtypes.ScheduleType.GPU_Multidevice
nsdfg.location['gpu'] = symbolic_gpu_id
for transient_node in graph.successors(nsdfg):
if isinstance(transient_node, nodes.AccessNode):
transient_node.desc(sdfg).location['gpu'] = symbolic_gpu_id
transient_node.desc(
sdfg).storage = dtypes.StorageType.GPU_Global
nsdfg.sdfg.arrays[
transient_node.label].location['gpu'] = symbolic_gpu_id
nsdfg.sdfg.arrays[
transient_node.
label].storage = dtypes.StorageType.GPU_Global
infer_types.set_default_schedule_storage_types_and_location(
nsdfg.sdfg, dtypes.ScheduleType.GPU_Multidevice,
symbolic_gpu_id)
# Remove the parameter of the outer_map from the sdfg symbols,
# as it got added as a symbol in StripMining.
if outer_map.params[0] in sdfg.free_symbols:
sdfg.remove_symbol(outer_map.params[0])
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)
class MapTiling(transformation.Transformation):
""" Implements the orthogonal tiling transformation.
Orthogonal tiling is a type of nested map fission that creates tiles
in every dimension of the matched Map.
"""
_map_entry = nodes.MapEntry(nodes.Map("", [], []))
# Properties
prefix = Property(dtype=str,
default="tile",
desc="Prefix for new range symbols")
tile_sizes = ShapeProperty(dtype=tuple,
default=(128, 128, 128),
desc="Tile size per dimension")
strides = ShapeProperty(
dtype=tuple,
default=tuple(),
desc="Tile stride (enables overlapping tiles). If empty, matches tile")
tile_offset = ShapeProperty(dtype=tuple,
default=None,
desc="Negative Stride offset per dimension",
allow_none=True)
divides_evenly = Property(dtype=bool,
default=False,
desc="Tile size divides dimension length evenly")
@staticmethod
def annotates_memlets():
return True
@staticmethod
def expressions():
return [sdutil.node_path_graph(MapTiling._map_entry)]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
return True
@staticmethod
def match_to_str(graph, candidate):
map_entry = graph.nodes()[candidate[MapTiling._map_entry]]
return map_entry.map.label + ': ' + str(map_entry.map.params)
def apply(self, sdfg):
graph = sdfg.nodes()[self.state_id]
tile_strides = self.tile_sizes
if self.strides is not None and len(self.strides) == len(tile_strides):
tile_strides = self.strides
# Retrieve map entry and exit nodes.
map_entry = graph.nodes()[self.subgraph[MapTiling._map_entry]]
from dace.transformation.dataflow.map_collapse import MapCollapse
from dace.transformation.dataflow.strip_mining import StripMining
stripmine_subgraph = {
StripMining._map_entry: self.subgraph[MapTiling._map_entry]
}
sdfg_id = sdfg.sdfg_id
last_map_entry = None
removed_maps = 0
original_schedule = map_entry.schedule
for dim_idx in range(len(map_entry.map.params)):
if dim_idx >= len(self.tile_sizes):
tile_size = symbolic.pystr_to_symbolic(self.tile_sizes[-1])
tile_stride = symbolic.pystr_to_symbolic(tile_strides[-1])
else:
tile_size = symbolic.pystr_to_symbolic(
self.tile_sizes[dim_idx])
tile_stride = symbolic.pystr_to_symbolic(tile_strides[dim_idx])
# handle offsets
if self.tile_offset and dim_idx >= len(self.tile_offset):
offset = self.tile_offset[-1]
elif self.tile_offset:
offset = self.tile_offset[dim_idx]
else:
offset = 0
dim_idx -= removed_maps
# If tile size is trivial, skip strip-mining map dimension
if tile_size == map_entry.map.range.size()[dim_idx]:
continue
stripmine = StripMining(sdfg_id, self.state_id, stripmine_subgraph,
self.expr_index)
# Special case: Tile size of 1 should be omitted from inner map
if tile_size == 1 and tile_stride == 1:
stripmine.dim_idx = dim_idx
stripmine.new_dim_prefix = ''
stripmine.tile_size = str(tile_size)
stripmine.tile_stride = str(tile_stride)
stripmine.divides_evenly = True
stripmine.tile_offset = str(offset)
stripmine.apply(sdfg)
removed_maps += 1
else:
stripmine.dim_idx = dim_idx
stripmine.new_dim_prefix = self.prefix
stripmine.tile_size = str(tile_size)
stripmine.tile_stride = str(tile_stride)
stripmine.divides_evenly = self.divides_evenly
stripmine.tile_offset = str(offset)
stripmine.apply(sdfg)
# apply to the new map the schedule of the original one
map_entry.schedule = original_schedule
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
class BufferTiling(transformation.Transformation):
""" Implements the buffer tiling transformation.
BufferTiling tiles a buffer that is in between two maps, where the preceding map
writes to the buffer and the succeeding map reads from it.
It introduces additional computations in exchange for reduced memory footprint.
Commonly used to make use of shared memory on GPUs.
"""
_map1_exit = nodes.MapExit(nodes.Map('', [], []))
_array = nodes.AccessNode('')
_map2_entry = nodes.MapEntry(nodes.Map('', [], []))
tile_sizes = ShapeProperty(dtype=tuple,
default=(128, 128, 128),
desc="Tile size per dimension")
# Returns a list of graphs that represent the pattern
@staticmethod
def expressions():
return [
sdutil.node_path_graph(
BufferTiling._map1_exit,
BufferTiling._array,
BufferTiling._map2_entry,
)
]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=False):
map1_exit = graph.nodes()[candidate[BufferTiling._map1_exit]]
map2_entry = graph.nodes()[candidate[BufferTiling._map2_entry]]
for buf in graph.all_nodes_between(map1_exit, map2_entry):
# Check that buffers are AccessNodes.
if not isinstance(buf, nodes.AccessNode):
return False
# Check that buffers are transient.
if not sdfg.arrays[buf.data].transient:
return False
# Check that buffers have exactly 1 input and 1 output edge.
if graph.in_degree(buf) != 1:
return False
if graph.out_degree(buf) != 1:
return False
# Check that buffers are next to the maps.
if graph.in_edges(buf)[0].src != map1_exit:
return False
if graph.out_edges(buf)[0].dst != map2_entry:
return False
# Check that the data consumed is provided.
provided = graph.in_edges(buf)[0].data.subset
consumed = graph.out_edges(buf)[0].data.subset
if not provided.covers(consumed):
return False
# Check that buffers occur only once in this state.
num_occurrences = len([
n for n in graph.nodes()
if isinstance(n, nodes.AccessNode) and n.data == buf
])
if num_occurrences > 1:
return False
return True
@staticmethod
def match_to_str(graph, candidate):
map1_exit = graph.nodes()[candidate[BufferTiling._map1_exit]]
map2_entry = graph.nodes()[candidate[BufferTiling._map2_entry]]
return " -> ".join(entry.map.label + ": " + str(entry.map.params)
for entry in [map1_exit, map2_entry])
def apply(self, sdfg):
graph = sdfg.nodes()[self.state_id]
map1_exit = graph.nodes()[self.subgraph[self._map1_exit]]
map1_entry = graph.entry_node(map1_exit)
map2_entry = graph.nodes()[self.subgraph[self._map2_entry]]
buffers = graph.all_nodes_between(map1_exit, map2_entry)
# Situation:
# -> map1_entry -> ... -> map1_exit -> buffers -> map2_entry -> ...
lower_extents = tuple(b - a for a, b in zip(
map1_entry.range.min_element(), map2_entry.range.min_element()))
upper_extents = tuple(a - b for a, b in zip(
map1_entry.range.max_element(), map2_entry.range.max_element()))
# Tile the first map with overlap
MapTilingWithOverlap.apply_to(sdfg,
map_entry=map1_entry,
options={
'tile_sizes': self.tile_sizes,
'lower_overlap': lower_extents,
'upper_overlap': upper_extents
})
tile_map1_exit = graph.out_edges(map1_exit)[0].dst
tile_map1_entry = graph.entry_node(tile_map1_exit)
tile_map1_entry.label = 'BufferTiling'
# Tile the second map
MapTiling.apply_to(sdfg,
map_entry=map2_entry,
options={
'tile_sizes': self.tile_sizes,
'tile_trivial': True
})
tile_map2_entry = graph.in_edges(map2_entry)[0].src
# Fuse maps
some_buffer = next(
iter(buffers)) # some dummy to pass to MapFusion.apply_to()
MapFusion.apply_to(sdfg,
first_map_exit=tile_map1_exit,
array=some_buffer,
second_map_entry=tile_map2_entry)
# Optimize the simple cases
map1_entry.range.ranges = [
(r[0], r[0], r[2]) if l_ext == 0 and u_ext == 0 and ts == 1 else r
for r, l_ext, u_ext, ts in zip(map1_entry.range.ranges,
lower_extents, upper_extents,
self.tile_sizes)
]
map2_entry.range.ranges = [
(r[0], r[0], r[2]) if ts == 1 else r
for r, ts in zip(map2_entry.range.ranges, self.tile_sizes)
]
if any(ts == 1 for ts in self.tile_sizes):
if any(r[0] == r[1] for r in map1_entry.map.range):
TrivialMapElimination.apply_to(sdfg, _map_entry=map1_entry)
if any(r[0] == r[1] for r in map2_entry.map.range):
TrivialMapElimination.apply_to(sdfg, _map_entry=map2_entry)
class MapInterchange(transformation.Transformation):
""" Implements the map-interchange transformation.
Map-interchange takes two nested maps and interchanges their position.
"""
_outer_map_entry = nodes.MapEntry(nodes.Map("", [], []))
_inner_map_entry = nodes.MapEntry(nodes.Map("", [], []))
@staticmethod
def expressions():
return [
sdutil.node_path_graph(MapInterchange._outer_map_entry,
MapInterchange._inner_map_entry)
]
@staticmethod
def can_be_applied(graph, candidate, expr_index, sdfg, strict=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
@staticmethod
def match_to_str(graph, candidate):
outer_map_entry = graph.nodes()[candidate[
MapInterchange._outer_map_entry]]
inner_map_entry = graph.nodes()[candidate[
MapInterchange._inner_map_entry]]
return ' -> '.join(entry.map.label + ': ' + str(entry.map.params)
for entry in [outer_map_entry, inner_map_entry])
def apply(self, sdfg: SDFG):
# Extract the parameters and ranges of the inner/outer maps.
graph: SDFGState = sdfg.nodes()[self.state_id]
outer_map_entry = graph.nodes()[self.subgraph[
MapInterchange._outer_map_entry]]
inner_map_entry = graph.nodes()[self.subgraph[
MapInterchange._inner_map_entry]]
inner_map_exit = graph.exit_node(inner_map_entry)
outer_map_exit = graph.exit_node(outer_map_entry)
# Switch connectors
outer_map_entry.in_connectors, inner_map_entry.in_connectors = \
inner_map_entry.in_connectors, outer_map_entry.in_connectors
outer_map_entry.out_connectors, inner_map_entry.out_connectors = \
inner_map_entry.out_connectors, outer_map_entry.out_connectors
outer_map_exit.in_connectors, inner_map_exit.in_connectors = \
inner_map_exit.in_connectors, outer_map_exit.in_connectors
outer_map_exit.out_connectors, inner_map_exit.out_connectors = \
inner_map_exit.out_connectors, outer_map_exit.out_connectors
# Get edges between the map entries and exits.
entry_edges = graph.edges_between(outer_map_entry, inner_map_entry)
exit_edges = graph.edges_between(inner_map_exit, outer_map_exit)
for e in entry_edges + exit_edges:
graph.remove_edge(e)
# Change source and destination of edges.
sdutil.change_edge_dest(graph, outer_map_entry, inner_map_entry)
sdutil.change_edge_src(graph, inner_map_entry, outer_map_entry)
sdutil.change_edge_dest(graph, inner_map_exit, outer_map_exit)
sdutil.change_edge_src(graph, outer_map_exit, inner_map_exit)
# Add edges between the map entries and exits.
new_entry_edges = []
new_exit_edges = []
for e in entry_edges:
new_entry_edges.append(
graph.add_edge(e.dst, e.src_conn, e.src, e.dst_conn, e.data))
for e in exit_edges:
new_exit_edges.append(
graph.add_edge(e.dst, e.src_conn, e.src, e.dst_conn, e.data))
# Repropagate memlets in modified region
for e in new_entry_edges:
path = graph.memlet_path(e)
index = next(i for i, edge in enumerate(path) if e is edge)
e.data.subset = propagate_memlet(graph, path[index + 1].data,
outer_map_entry, True).subset
for e in new_exit_edges:
path = graph.memlet_path(e)
index = next(i for i, edge in enumerate(path) if e is edge)
e.data.subset = propagate_memlet(graph, path[index - 1].data,
outer_map_exit, True).subset
@staticmethod
def annotates_memlets():
return True
