compiler: Further misc improvements (minor)

Author: EdCaunt

devito/ir/clusters/algorithms.py

@@ -157,6 +157,7 @@ def callback(self, clusters, prefix, backlog=None, known_break=None):
         # parallelism
         for i in range(1, len(clusters)):
             # FIXME: This eats a lot of time (four seconds each time)
+            # FIXME: Pull scope out of this
             if self._break_for_parallelism(scope, candidates, i):
                 return self.callback(clusters[:i], prefix, clusters[i:] + backlog,
                                      candidates | known_break)
@@ -194,6 +195,7 @@ def _break_for_parallelism(self, scope, candidates, i):
         # break parallelism
 
         # TODO: Can this loop be made to short-circuit?
+        # TODO: Most of the time is burned in d_from_access_gen
         test = False
         for d in scope.d_from_access_gen(scope.a_query(i)):
             if d.is_local or d.is_storage_related(candidates):

devito/ir/support/basic.py

@@ -317,6 +317,8 @@ def lex_le(self, other):
     def lex_lt(self, other):
         return self.timestamp < other.timestamp
 
+    # NOTE: This is called a lot with the same arguments - memoize yields mild speedup
+    @memoized_meth
     def distance(self, other):
         """
         Compute the distance from ``self`` to ``other``.
@@ -365,7 +367,9 @@ def distance(self, other):
                 # constant access at 4
                 for v in (self[n], other[n]):
                     try:
-                        if bool(v < sit.symbolic_min or v > sit.symbolic_max):
+                        # NOTE: Split the boolean to make the conditional short circuit
+                        # more frequently for mild speedup
+                        if bool(v < sit.symbolic_min) or bool(v > sit.symbolic_max):
                             return Vector(S.ImaginaryUnit)
                     except TypeError:
                         pass
@@ -1166,6 +1170,7 @@ def d_from_access_gen(self, accesses):
         Generate all flow, anti, and output dependences involving any of
         the given TimedAccess objects.
         """
+        # FIXME: This seems to be a hotspot
         accesses = as_tuple(accesses)
         for d in self.d_all_gen():
             for i in accesses:

devito/passes/clusters/misc.py

@@ -352,8 +352,7 @@ def is_cross(source, sink):
                 v = len(cg0.exprs)
                 return t0 < v <= t1 or t1 < v <= t0
 
-            for cg1 in cgroups[n+1:]:
-                n1 = cgroups.index(cg1)
+            for n1, cg1 in enumerate(cgroups[n+1:], start=n+1):
 
                 # A Scope to compute all cross-ClusterGroup anti-dependences
                 scope = Scope(exprs=cg0.exprs + cg1.exprs, rules=is_cross)

devito/passes/iet/mpi.py

@@ -24,6 +24,7 @@ def optimize_halospots(iet, **kwargs):
     merged and moved around in order to improve the halo exchange performance.
     """
     iet = _drop_reduction_halospots(iet)
+    # TODO: _hoist_invariant and _merge_halospots are both slow
     iet = _hoist_invariant(iet)
     iet = _merge_halospots(iet)
     iet = _drop_if_unwritten(iet, **kwargs)
@@ -81,6 +82,8 @@ def _hoist_invariant(iet):
     cond_mapper = _make_cond_mapper(iet)
     iter_mapper = _filter_iter_mapper(iet)
 
+    # from IPython import embed; embed()
+
     for it, halo_spots in iter_mapper.items():
         for hs0, hs1 in combinations(halo_spots, r=2):
             if _check_control_flow(hs0, hs1, cond_mapper):
@@ -157,6 +160,7 @@ def _merge_halospots(iet):
 
             for f, v in hs1.fmapper.items():
                 for dep in scope.d_flow.project(f):
+                    # NOTE: The else clause gets skipped if this is ever true
                     if not any(rule(dep, hs1, v.loc_indices) for rule in rules):
                         break
                 else: