More control over operation fusion (#355)

* Document PrimitiveOperation

* Allow a primitive operation to be marked as 'fusable'

* Add mechanism to manually fuse operations

* Add fuse_all_optimize_dag and fuse_only_optimize_dag

* Explicitly set simple_optimize_dag for 'no fusion' tests

cubed/core/ops.py

@@ -262,6 +262,8 @@ def blockwise(
 
     extra_projected_mem = kwargs.pop("extra_projected_mem", 0)
 
+    fusable = kwargs.pop("fusable", True)
+
     name = gensym()
     spec = check_array_specs(arrays)
     if target_store is None:
@@ -281,6 +283,7 @@ def blockwise(
         in_names=in_names,
         out_name=name,
         extra_func_kwargs=extra_func_kwargs,
+        fusable=fusable,
         **kwargs,
     )
     plan = Plan._new(
@@ -688,6 +691,7 @@ def wrap(*a, block_id=None, **kw):
         chunks=chunks,
         extra_source_arrays=args,
         extra_projected_mem=extra_projected_mem,
+        fusable=False,  # don't allow fusion since side inputs are not accounted for
         **kwargs,
     )
 

cubed/core/optimization.py

@@ -92,10 +92,12 @@ def predecessor_ops(dag, name):
 
 def is_fusable(node_dict):
     "Return True if a node can be fused."
-    return "primitive_op" in node_dict
+    return "primitive_op" in node_dict and node_dict["primitive_op"].fusable
 
 
-def can_fuse_predecessors(dag, name, *, max_total_nargs=4):
+def can_fuse_predecessors(
+    dag, name, *, max_total_nargs=4, always_fuse=None, never_fuse=None
+):
     nodes = dict(dag.nodes(data=True))
 
     # if node itself can't be fused then there is nothing to fuse
@@ -106,6 +108,12 @@ def can_fuse_predecessors(dag, name, *, max_total_nargs=4):
     if all(not is_fusable(nodes[pre]) for pre in predecessor_ops(dag, name)):
         return False
 
+    # if node is in never_fuse or always_fuse list then it overrides logic below
+    if never_fuse is not None and name in never_fuse:
+        return False
+    if always_fuse is not None and name in always_fuse:
+        return True
+
     # if there is more than a single predecessor op, and the total number of args to
     # the fused function would be more than an allowed maximum, then don't fuse
     if len(list(predecessor_ops(dag, name))) > 1:
@@ -126,24 +134,32 @@ def can_fuse_predecessors(dag, name, *, max_total_nargs=4):
     )
 
 
-def fuse_predecessors(dag, name):
+def fuse_predecessors(
+    dag, name, *, max_total_nargs=4, always_fuse=None, never_fuse=None
+):
     """Fuse a node with its immediate predecessors."""
 
     # if can't fuse then return dag unchanged
-    if not can_fuse_predecessors(dag, name):
+    if not can_fuse_predecessors(
+        dag,
+        name,
+        max_total_nargs=max_total_nargs,
+        always_fuse=always_fuse,
+        never_fuse=never_fuse,
+    ):
         return dag
 
     nodes = dict(dag.nodes(data=True))
 
     primitive_op = nodes[name]["primitive_op"]
 
-    # if a predecessor op has no primitive op then just use None
+    # if a predecessor has no primitive op then just use None
     predecessor_primitive_ops = [
         nodes[pre]["primitive_op"] if is_fusable(nodes[pre]) else None
         for pre in predecessor_ops(dag, name)
     ]
 
-    # if a predecessor op has no func then use 1 for nargs
+    # if a predecessor has no primitive op then use 1 for nargs
     predecessor_funcs_nargs = [
         len(list(predecessors(dag, pre))) if is_fusable(nodes[pre]) else 1
         for pre in predecessor_ops(dag, name)
@@ -167,12 +183,12 @@ def fuse_predecessors(dag, name):
     for input in predecessors(dag, name):
         pre = next(predecessors(dag, input))
         if not is_fusable(fused_nodes[pre]):
-            # if a predecessor is marked as not fusable then don't change the edge
+            # if a predecessor is not fusable then don't change the edge
             continue
         fused_dag.remove_edge(input, name)
     for pre in predecessor_ops(dag, name):
         if not is_fusable(fused_nodes[pre]):
-            # if a predecessor is marked as not fusable then don't change the edge
+            # if a predecessor is not fusable then don't change the edge
             continue
         for input in predecessors(dag, pre):
             fused_dag.add_edge(input, name)
@@ -188,8 +204,33 @@ def fuse_predecessors(dag, name):
     return fused_dag
 
 
-def multiple_inputs_optimize_dag(dag):
+def multiple_inputs_optimize_dag(
+    dag, *, max_total_nargs=4, always_fuse=None, never_fuse=None
+):
     """Fuse multiple inputs."""
     for name in list(nx.topological_sort(dag)):
-        dag = fuse_predecessors(dag, name)
+        dag = fuse_predecessors(
+            dag,
+            name,
+            max_total_nargs=max_total_nargs,
+            always_fuse=always_fuse,
+            never_fuse=never_fuse,
+        )
     return dag
+
+
+def fuse_all_optimize_dag(dag):
+    """Force all operations to be fused."""
+    dag = dag.copy()
+    always_fuse = [op for op in dag.nodes() if op.startswith("op-")]
+    return multiple_inputs_optimize_dag(dag, always_fuse=always_fuse)
+
+
+def fuse_only_optimize_dag(dag, *, only_fuse=None):
+    """Force only specified operations to be fused, all others will be left even if they are suitable for fusion."""
+    dag = dag.copy()
+    always_fuse = only_fuse
+    never_fuse = set(op for op in dag.nodes() if op.startswith("op-")) - set(only_fuse)
+    return multiple_inputs_optimize_dag(
+        dag, always_fuse=always_fuse, never_fuse=never_fuse
+    )

cubed/core/plan.py

@@ -447,4 +447,5 @@ def create_zarr_arrays(lazy_zarr_arrays, reserved_mem):
         projected_mem=projected_mem,
         reserved_mem=reserved_mem,
         num_tasks=num_tasks,
+        fusable=False,
     )

cubed/primitive/blockwise.py

@@ -115,6 +115,7 @@ def blockwise(
     out_name: Optional[str] = None,
     extra_projected_mem: int = 0,
     extra_func_kwargs: Optional[Dict[str, Any]] = None,
+    fusable: bool = True,
     **kwargs,
 ):
     """Apply a function to multiple blocks from multiple inputs, expressed using concise indexing rules.
@@ -196,6 +197,7 @@ def blockwise(
         in_names=in_names,
         extra_projected_mem=extra_projected_mem,
         extra_func_kwargs=extra_func_kwargs,
+        fusable=fusable,
         **kwargs,
     )
 
@@ -213,6 +215,7 @@ def general_blockwise(
     in_names: Optional[List[str]] = None,
     extra_projected_mem: int = 0,
     extra_func_kwargs: Optional[Dict[str, Any]] = None,
+    fusable: bool = True,
     **kwargs,
 ):
     """A more general form of ``blockwise`` that uses a function to specify the block
@@ -307,6 +310,7 @@ def general_blockwise(
         projected_mem=projected_mem,
         reserved_mem=reserved_mem,
         num_tasks=num_tasks,
+        fusable=fusable,
     )
 
 
@@ -383,6 +387,7 @@ def fused_func(*args):
         projected_mem=projected_mem,
         reserved_mem=reserved_mem,
         num_tasks=num_tasks,
+        fusable=True,
     )
 
 
@@ -469,6 +474,7 @@ def fused_func(*args):
         projected_mem=projected_mem,
         reserved_mem=reserved_mem,
         num_tasks=num_tasks,
+        fusable=True,
     )
 
 

cubed/primitive/rechunk.py

@@ -208,5 +208,6 @@ def spec_to_primitive_op(
         projected_mem=projected_mem,
         reserved_mem=reserved_mem,
         num_tasks=num_tasks,
+        fusable=False,
         write_chunks=spec.write.chunks,
     )

cubed/primitive/types.py

@@ -13,11 +13,25 @@ class PrimitiveOperation:
     """Encapsulates metadata about a ``blockwise`` or ``rechunk`` primitive operation."""
 
     pipeline: CubedPipeline
+    """The pipeline that runs this operation."""
+
     target_array: Any
+    """The array being computed by this operation."""
+
     projected_mem: int
+    """An upper bound of the memory needed to run a task, in bytes."""
+
     reserved_mem: int
+    """The memory reserved on a worker for non-data use when running a task, in bytes."""
+
     num_tasks: int
+    """The number of tasks needed to run this operation."""
+
+    fusable: bool = True
+    """Whether this operation should be considered for fusion."""
+
     write_chunks: Optional[T_RegularChunks] = None
+    """The chunk size used by this operation."""
 
 
 class CubedArrayProxy: