def clusterize(exprs):
"""
Turn a sequence of LoweredEqs into a sequence of Clusters.
"""
# Initialization
clusters = [Cluster(e, e.ispace, e.dspace) for e in exprs]
# Setup the IterationSpaces based on data dependence analysis
clusters = Schedule().process(clusters)
# Handle ConditionalDimensions
clusters = guard(clusters)
# Determine relevant computational properties (e.g., parallelism)
clusters = analyze(clusters)
return ClusterGroup(clusters)
def clusterize(exprs):
"""
Turn a sequence of LoweredEqs into a sequence of Clusters.
"""
# Initialization
clusters = [Cluster(e, e.ispace, e.dspace) for e in exprs]
# Compute a topological ordering that honours flow- and anti-dependences
clusters = Toposort().process(clusters)
# Setup the IterationSpaces based on data dependence analysis
clusters = Schedule().process(clusters)
# Introduce conditional Clusters
clusters = guard(clusters)
# Determine relevant computational properties (e.g., parallelism)
clusters = analyze(clusters)
return ClusterGroup(clusters)
def optimize(clusters, dse_mode):
"""
Optimize a topologically-ordered sequence of Clusters by applying the
following transformations:
* [cross-cluster] Fusion
* [intra-cluster] Several flop-reduction passes via the DSE
* [cross-cluster] Lifting
* [cross-cluster] Scalarization
* [cross-cluster] Arrays Elimination
"""
# To create temporaries
counter = generator()
template = lambda: "r%d" % counter()
# Toposort+Fusion (the former to expose more fusion opportunities)
clusters = Toposort().process(clusters)
clusters = fuse(clusters)
# Flop reduction via the DSE
from devito.dse import rewrite
clusters = rewrite(clusters, template, mode=dse_mode)
# Lifting
clusters = Lift().process(clusters)
# Lifting may create fusion opportunities
clusters = fuse(clusters)
# Fusion may create opportunities to eliminate Arrays (thus shrinking the
# working set) if these store identical expressions
clusters = eliminate_arrays(clusters, template)
# Fusion may create scalarization opportunities
clusters = scalarize(clusters, template)
# Determine computational properties (e.g., parallelism) that will be
# necessary for the later passes
clusters = analyze(clusters)
return ClusterGroup(clusters)
