Merge branch 'fix_parallel_reduction_size' into 'master'

Fix parallel reduction for size 1

Closes #44

See merge request !94

loopy/preprocess.py

@@ -539,7 +539,7 @@ def realize_reduction(kernel, insn_id_filter=None, unknown_types_ok=True):
 
         new_insn_add_depends_on.add(prev_id)
         new_insn_add_no_sync_with.add((prev_id, "any"))
-        new_insn_add_within_inames.add(stage_exec_iname or base_exec_iname)
+        new_insn_add_within_inames.add(base_exec_iname or stage_exec_iname)
 
         if nresults == 1:
             assert len(acc_vars) == 1

loopy/transform/iname.py

@@ -1198,7 +1198,8 @@ class _ReductionSplitter(RuleAwareIdentityMapper):
                 return Reduction(expr.operation, tuple(self.inames),
                         Reduction(expr.operation, tuple(leftover_inames),
                             self.rec(expr.expr, expn_state),
-                            expr.allow_simultaneous))
+                            expr.allow_simultaneous),
+                        expr.allow_simultaneous)
             else:
                 assert False
         else:

test/test_reduction.py

@@ -181,7 +181,7 @@ def test_recursive_nested_dependent_reduction(ctx_factory):
     # FIXME: Actually test functionality.
 
 
-@pytest.mark.parametrize("size", [128, 5, 113, 67])
+@pytest.mark.parametrize("size", [128, 5, 113, 67, 1])
 def test_local_parallel_reduction(ctx_factory, size):
     ctx = ctx_factory()