def _specialize_clusters(cls, clusters, **kwargs):
"""
Optimize Clusters for better runtime performance.
"""
# 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
clusters = rewrite(clusters, template, **kwargs)
# Lifting
clusters = Lift().process(clusters)
# Lifting may create fusion opportunities, which in turn may enable
# further optimizations
clusters = fuse(clusters)
clusters = eliminate_arrays(clusters, template)
clusters = scalarize(clusters, template)
return clusters
def _specialize_clusters(cls, clusters, **kwargs):
# TODO: this is currently identical to CPU64NoopOperator._specialize_clusters,
# but it will have to change
# 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
clusters = rewrite(clusters, template, **kwargs)
# Lifting
clusters = Lift().process(clusters)
# Lifting may create fusion opportunities, which in turn may enable
# further optimizations
clusters = fuse(clusters)
clusters = eliminate_arrays(clusters, template)
clusters = scalarize(clusters, template)
return clusters
def _specialize_clusters(cls, clusters, **kwargs):
options = kwargs['options']
platform = kwargs['platform']
# 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)
# Hoist and optimize Dimension-invariant sub-expressions
clusters = cire(clusters, template, 'invariants', options, platform)
clusters = Lift().process(clusters)
# Reduce flops (potential arithmetic alterations)
clusters = extract_increments(clusters, template)
clusters = cire(clusters, template, 'sops', options, platform)
clusters = factorize(clusters)
clusters = optimize_pows(clusters)
# Reduce flops (no arithmetic alterations)
clusters = cse(clusters, template)
# The previous passes may have created fusion opportunities, which in
# turn may enable further optimizations
clusters = fuse(clusters)
clusters = eliminate_arrays(clusters, template)
clusters = scalarize(clusters, template)
# Blocking to improve data locality
clusters = Blocking(options).process(clusters)
return clusters