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_list.index(sdfg)
last_map_entry = None
for dim_idx in range(len(map_entry.map.params)):
if dim_idx >= len(self.tile_sizes):
tile_size = self.tile_sizes[-1]
tile_stride = tile_strides[-1]
else:
tile_size = self.tile_sizes[dim_idx]
tile_stride = tile_strides[dim_idx]
stripmine = StripMining(sdfg_id, self.state_id, stripmine_subgraph,
self.expr_index)
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.apply(sdfg)
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
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_list.index(sdfg)
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])
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.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.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
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)