compiler: Fix MPI with CondDim + multiple loc_indices

Author: FabioLuporini

devito/ir/clusters/algorithms.py

@@ -399,6 +399,83 @@ class HaloComms(Queue):
     def process(self, clusters):
         return self._process_fatd(clusters, 1, seen=set())
 
+    def _derive_halo_schemes(self, c):
+        hs = HaloScheme(c.exprs, c.ispace)
+
+        # 95% of the times we will just return `hs` as is as there are no guards
+        if not c.guards:
+            yield hs, c
+            return
+
+        # This is a more contrived situation in which we might need halo exchanges
+        # from multiple so called loc-indices -- let's check this out
+        candidates = []
+        for f, hse in hs.fmapper.items():
+            reads = c.scope.reads[f]
+
+            for d in hse.loc_indices:
+                if not d._defines & set(c.guards):
+                    continue
+
+                candidates.append(as_mapper(reads, key=lambda i: i[d]).values())
+
+        # 4% of the times we will just return `hs` as is
+        # E.g., we end up here when taking space derivatives of one or more saved
+        # TimeFunctions in equations evaluating gradients that are controlled by
+        # a ConditionalDimension (otherwise we would have exited earlier)
+        if any(len(g) <= 1 for g in candidates):
+            yield hs, c
+            return
+
+        # 1% of the times, finally, we end up here...
+        # At this point we have to create a mock Cluster for each loc-index,
+        # containing all and only the accesses to `f` at a given loc-index
+        # E.g., a mock Cluster at `loc_index=t0` containing the accesses
+        # `[u(t0, x + 8, ...), u(t0, x + 7, ...)], another mock Cluster at
+        # `loc_index=t1` containing the accesses `[u(t1, x + 5, ...),
+        # u(t1, x + 6, ...)]`, and so on
+        for unordered_groups in candidates:
+            # Sort for deterministic code generation
+            groups = sorted(unordered_groups, key=str)
+            for group in groups:
+                pointset = sympy.Function('pointset')
+                v = pointset(*[i.access for i in group])
+                exprs = [e.func(rhs=v) for e in c.exprs]
+
+                c1 = c.rebuild(exprs=exprs)
+
+                hs = HaloScheme(c1.exprs, c.ispace)
+
+                yield hs, c1
+
+    def _make_halo_touch(self, hs, c, prefix):
+        points = set()
+        for f in hs.fmapper:
+            for a in c.scope.getreads(f):
+                points.add(a.access)
+
+        # We also add all written symbols to ultimately create mock WARs
+        # with `c`, which will prevent the newly created HaloTouch from
+        # ever being rescheduled
+        points.update(a.access for a in c.scope.accesses if a.is_write)
+
+        # Sort for determinism
+        # NOTE: not sorting might impact code generation. The order of
+        # the args is important because that's what search functions honor!
+        points = sorted(points, key=str)
+
+        # Construct the HaloTouch Cluster
+        expr = Eq(self.B, HaloTouch(*points, halo_scheme=hs))
+
+        key = lambda i: i in prefix[:-1] or i in hs.loc_indices
+        ispace = c.ispace.project(key)
+        # HaloTouches are not parallel
+        properties = c.properties.sequentialize()
+
+        halo_touch = c.rebuild(exprs=expr, ispace=ispace, properties=properties)
+
+        return halo_touch
+
     def callback(self, clusters, prefix, seen=None):
         if not prefix:
             return clusters
@@ -412,38 +489,18 @@ def callback(self, clusters, prefix, seen=None):
                c in seen:
                 continue
 
-            hs = HaloScheme(c.exprs, c.ispace)
-            if hs.is_void or \
-               not d._defines & hs.distributed_aindices:
-                continue
-
-            points = set()
-            for f in hs.fmapper:
-                for a in c.scope.getreads(f):
-                    points.add(a.access)
-
-            # We also add all written symbols to ultimately create mock WARs
-            # with `c`, which will prevent the newly created HaloTouch to ever
-            # be rescheduled after `c` upon topological sorting
-            points.update(a.access for a in c.scope.accesses if a.is_write)
+            seen.add(c)
 
-            # Sort for determinism
-            # NOTE: not sorting might impact code generation. The order of
-            # the args is important because that's what search functions honor!
-            points = sorted(points, key=str)
-
-            # Construct the HaloTouch Cluster
-            expr = Eq(self.B, HaloTouch(*points, halo_scheme=hs))
+            for hs, c1 in self._derive_halo_schemes(c):
+                if hs.is_void or \
+                   not d._defines & hs.distributed_aindices:
+                    continue
 
-            key = lambda i: i in prefix[:-1] or i in hs.loc_indices
-            ispace = c.ispace.project(key)
-            # HaloTouches are not parallel
-            properties = c.properties.sequentialize()
+                halo_touch = self._make_halo_touch(hs, c1, prefix)
 
-            halo_touch = c.rebuild(exprs=expr, ispace=ispace, properties=properties)
+                processed.append(halo_touch)
 
-            processed.append(halo_touch)
-            seen.update({halo_touch, c})
+                seen.add(halo_touch)
 
         processed.extend(clusters)
 

devito/ir/stree/algorithms.py

@@ -204,9 +204,15 @@ def preprocess(clusters, options=None, **kwargs):
                 processed.append(c.rebuild(exprs=[], ispace=ispace, syncs=syncs))
 
             if all(c1.ispace.is_subset(c.ispace) for c1 in found):
-                # 99% of the cases we end up here
-                hs = HaloScheme.union([c1.halo_scheme for c1 in found])
-                processed.append(c.rebuild(halo_scheme=hs))
+                if not any(c1.guards for c1 in found):
+                    # 99% of the cases we end up here
+                    hs = HaloScheme.union([c1.halo_scheme for c1 in found])
+                    processed.append(c.rebuild(halo_scheme=hs))
+                else:
+                    # We have to keep all HaloSchemes explicitly in separate
+                    # Clusters or we might generate broken code by erroneously
+                    # dropping halo exchanges on the floor
+                    processed.extend((*found, c))
             elif options['mpi']:
                 # We end up here with e.g. `t,x,y,z,f` where `f` is a sequential
                 # dimension requiring a loc-index in the HaloScheme. The compiler

devito/passes/iet/mpi.py

@@ -141,7 +141,7 @@ def _merge_halospots(iet):
     mapper = HaloSpotMapper()
     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):
+            if cond_mapper.get(hs0) != cond_mapper.get(hs1):
                 continue
 
             scope = _derive_scope(it, hs0, hs1)
@@ -191,7 +191,7 @@ def _hoist_invariant(iet):
     mapper = HaloSpotMapper()
     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):
+            if cond_mapper.get(hs0) or cond_mapper.get(hs1):
                 continue
 
             scope = _derive_scope(it, hs0, hs1)
@@ -465,17 +465,6 @@ def _derive_scope(it, hs0, hs1):
     return Scope(e.expr for e in expressions)
 
 
-def _check_control_flow(hs0, hs1, cond_mapper):
-    """
-    If there are Conditionals involved, both `hs0` and `hs1` must be
-    within the same Conditional, otherwise we would break control flow
-    """
-    cond0 = cond_mapper.get(hs0)
-    cond1 = cond_mapper.get(hs1)
-
-    return cond0 != cond1
-
-
 def _is_iter_carried(hsf, scope):
     """
     True if the provided HaloScheme `hsf` is iteration-carried, i.e., it induces

tests/test_mpi.py

@@ -6,7 +6,7 @@
 from devito import (Grid, Constant, Function, TimeFunction, SparseFunction,
                     SparseTimeFunction, VectorTimeFunction, TensorTimeFunction,
                     Dimension, ConditionalDimension, div, solve, diag, grad,
-                    SubDimension, SubDomain, Eq, Ne, Inc, NODE, Operator, norm,
+                    SubDimension, SubDomain, Eq, Ne, Gt, Inc, NODE, Operator, norm,
                     inner, configuration, switchconfig, generic_derivative,
                     PrecomputedSparseFunction, DefaultDimension, Buffer,
                     CustomDimension)
@@ -3128,6 +3128,25 @@ def test_interpolation_at_uforward(self, mode):
         assert args[-2].name == 't2'
         assert args[-2].origin == t + 1
 
+    @pytest.mark.parallel(mode=1)
+    def test_multiple_loc_indices_inside_conddim(self, mode):
+        grid = Grid(shape=(10, 10))
+        time = grid.time_dim
+
+        t_sub = ConditionalDimension('t_sub', parent=time, condition=Gt(time % 4))
+
+        f = Function(name='f', grid=grid, space_order=4)
+        u = TimeFunction(name='u', grid=grid, space_order=4)
+
+        eqns = [Eq(u.forward, u + 1),
+                Eq(f, u.dx + u.forward.dx + .2, implicit_dims=t_sub)]
+
+        op = Operator(eqns, opt=('advanced', {'openmp': False}))
+
+        calls, _ = check_halo_exchanges(op, 2, 2)
+        assert calls[0].arguments[-1].name == 't0'
+        assert calls[1].arguments[-1].name == 't1'
+
 
 def gen_serial_norms(shape, so):
     """