Introduce max_total_num_input_blocks as a heuristic to control fusion (…

…#376)

cubed/core/optimization.py

@@ -96,7 +96,13 @@ def is_fusable(node_dict):
 
 
 def can_fuse_predecessors(
-    dag, name, *, max_total_source_arrays=4, always_fuse=None, never_fuse=None
+    dag,
+    name,
+    *,
+    max_total_source_arrays=4,
+    max_total_num_input_blocks=None,
+    always_fuse=None,
+    never_fuse=None,
 ):
     nodes = dict(dag.nodes(data=True))
 
@@ -130,12 +136,20 @@ def can_fuse_predecessors(
         if is_fusable(nodes[pre])
     ]
     return can_fuse_multiple_primitive_ops(
-        nodes[name]["primitive_op"], *predecessor_primitive_ops
+        nodes[name]["primitive_op"],
+        predecessor_primitive_ops,
+        max_total_num_input_blocks=max_total_num_input_blocks,
     )
 
 
 def fuse_predecessors(
-    dag, name, *, max_total_source_arrays=4, always_fuse=None, never_fuse=None
+    dag,
+    name,
+    *,
+    max_total_source_arrays=4,
+    max_total_num_input_blocks=None,
+    always_fuse=None,
+    never_fuse=None,
 ):
     """Fuse a node with its immediate predecessors."""
 
@@ -144,6 +158,7 @@ def fuse_predecessors(
         dag,
         name,
         max_total_source_arrays=max_total_source_arrays,
+        max_total_num_input_blocks=max_total_num_input_blocks,
         always_fuse=always_fuse,
         never_fuse=never_fuse,
     ):
@@ -195,14 +210,20 @@ def fuse_predecessors(
 
 
 def multiple_inputs_optimize_dag(
-    dag, *, max_total_source_arrays=4, always_fuse=None, never_fuse=None
+    dag,
+    *,
+    max_total_source_arrays=4,
+    max_total_num_input_blocks=None,
+    always_fuse=None,
+    never_fuse=None,
 ):
     """Fuse multiple inputs."""
     for name in list(nx.topological_sort(dag)):
         dag = fuse_predecessors(
             dag,
             name,
             max_total_source_arrays=max_total_source_arrays,
+            max_total_num_input_blocks=max_total_num_input_blocks,
             always_fuse=always_fuse,
             never_fuse=never_fuse,
         )

cubed/primitive/blockwise.py

@@ -352,7 +352,10 @@ def can_fuse_primitive_ops(
 
 
 def can_fuse_multiple_primitive_ops(
-    primitive_op: PrimitiveOperation, *predecessor_primitive_ops: PrimitiveOperation
+    primitive_op: PrimitiveOperation,
+    predecessor_primitive_ops: List[PrimitiveOperation],
+    *,
+    max_total_num_input_blocks: Optional[int] = None,
 ) -> bool:
     if is_fuse_candidate(primitive_op) and all(
         is_fuse_candidate(p) for p in predecessor_primitive_ops
@@ -368,9 +371,18 @@ def can_fuse_multiple_primitive_ops(
         num_input_blocks = primitive_op.pipeline.config.num_input_blocks
         if not all(num_input_blocks[0] == n for n in num_input_blocks):
             return False
-        return all(
-            primitive_op.num_tasks == p.num_tasks for p in predecessor_primitive_ops
-        )
+        if max_total_num_input_blocks is None:
+            # If max total input blocks not specified, then only fuse if num
+            # tasks of predecessor ops match.
+            return all(
+                primitive_op.num_tasks == p.num_tasks for p in predecessor_primitive_ops
+            )
+        else:
+            total_num_input_blocks = 0
+            for ni, p in zip(num_input_blocks, predecessor_primitive_ops):
+                for nj in p.pipeline.config.num_input_blocks:
+                    total_num_input_blocks += ni * nj
+            return total_num_input_blocks <= max_total_num_input_blocks
     return False
 
 

cubed/tests/test_optimization.py

@@ -152,10 +152,12 @@ def fuse_one_level(arr, *, always_fuse=None):
     )
 
 
-def fuse_multiple_levels(*, max_total_source_arrays=4):
+def fuse_multiple_levels(*, max_total_source_arrays=4, max_total_num_input_blocks=None):
     # use multiple_inputs_optimize_dag to test multiple levels of fusion
     return partial(
-        multiple_inputs_optimize_dag, max_total_source_arrays=max_total_source_arrays
+        multiple_inputs_optimize_dag,
+        max_total_source_arrays=max_total_source_arrays,
+        max_total_num_input_blocks=max_total_num_input_blocks,
     )
 
 
@@ -775,9 +777,8 @@ def test_fuse_merge_chunks_unary(spec):
     b = xp.negative(a)
     c = merge_chunks_new(b, chunks=(3, 2))
 
-    # force c to fuse
-    last_op = sorted(c.plan.dag.nodes())[-1]
-    opt_fn = fuse_one_level(c, always_fuse=[last_op])
+    # specify max_total_num_input_blocks to force c to fuse
+    opt_fn = fuse_multiple_levels(max_total_num_input_blocks=3)
 
     c.visualize(optimize_function=opt_fn)
 
@@ -809,9 +810,8 @@ def test_fuse_merge_chunks_binary(spec):
     c = xp.add(a, b)
     d = merge_chunks_new(c, chunks=(3, 2))
 
-    # force d to fuse
-    last_op = sorted(d.plan.dag.nodes())[-1]
-    opt_fn = fuse_one_level(d, always_fuse=[last_op])
+    # specify max_total_num_input_blocks to force d to fuse
+    opt_fn = fuse_multiple_levels(max_total_num_input_blocks=6)
 
     d.visualize(optimize_function=opt_fn)