Change default max_total_num_input_blocks to 10 (#615)

To allow operations to be fused as long as the total number of input blocks does not exceed this number.

Previously, the default was None, which meant operations were fused only if they had the same number of tasks.

Author: tomwhite

cubed/core/optimization.py

@@ -11,6 +11,9 @@
 
 logger = logging.getLogger(__name__)
 
+DEFAULT_MAX_TOTAL_SOURCE_ARRAYS = 4
+DEFAULT_MAX_TOTAL_NUM_INPUT_BLOCKS = 10
+
 
 def simple_optimize_dag(dag, array_names=None):
     """Apply map blocks fusion."""
@@ -154,8 +157,8 @@ def can_fuse_predecessors(
     name,
     *,
     array_names=None,
-    max_total_source_arrays=4,
-    max_total_num_input_blocks=None,
+    max_total_source_arrays=DEFAULT_MAX_TOTAL_SOURCE_ARRAYS,
+    max_total_num_input_blocks=DEFAULT_MAX_TOTAL_NUM_INPUT_BLOCKS,
     always_fuse=None,
     never_fuse=None,
 ):
@@ -242,8 +245,8 @@ def fuse_predecessors(
     name,
     *,
     array_names=None,
-    max_total_source_arrays=4,
-    max_total_num_input_blocks=None,
+    max_total_source_arrays=DEFAULT_MAX_TOTAL_SOURCE_ARRAYS,
+    max_total_num_input_blocks=DEFAULT_MAX_TOTAL_NUM_INPUT_BLOCKS,
     always_fuse=None,
     never_fuse=None,
 ):
@@ -297,8 +300,8 @@ def multiple_inputs_optimize_dag(
     dag,
     *,
     array_names=None,
-    max_total_source_arrays=4,
-    max_total_num_input_blocks=None,
+    max_total_source_arrays=DEFAULT_MAX_TOTAL_SOURCE_ARRAYS,
+    max_total_num_input_blocks=DEFAULT_MAX_TOTAL_NUM_INPUT_BLOCKS,
     always_fuse=None,
     never_fuse=None,
 ):

cubed/tests/test_optimization.py

@@ -11,6 +11,8 @@
 from cubed.backend_array_api import namespace as nxp
 from cubed.core.ops import elemwise, merge_chunks, partial_reduce
 from cubed.core.optimization import (
+    DEFAULT_MAX_TOTAL_NUM_INPUT_BLOCKS,
+    DEFAULT_MAX_TOTAL_SOURCE_ARRAYS,
     fuse_all_optimize_dag,
     fuse_only_optimize_dag,
     fuse_predecessors,
@@ -223,7 +225,11 @@ def fuse_one_level(arr, *, always_fuse=None):
     )
 
 
-def fuse_multiple_levels(*, max_total_source_arrays=4, max_total_num_input_blocks=None):
+def fuse_multiple_levels(
+    *,
+    max_total_source_arrays=DEFAULT_MAX_TOTAL_SOURCE_ARRAYS,
+    max_total_num_input_blocks=DEFAULT_MAX_TOTAL_NUM_INPUT_BLOCKS,
+):
     # use multiple_inputs_optimize_dag to test multiple levels of fusion
     return partial(
         multiple_inputs_optimize_dag,
@@ -899,8 +905,7 @@ def test_fuse_merge_chunks_unary(spec):
     b = xp.negative(a)
     c = merge_chunks(b, chunks=(3, 2))
 
-    # specify max_total_num_input_blocks to force c to fuse
-    opt_fn = fuse_multiple_levels(max_total_num_input_blocks=3)
+    opt_fn = fuse_multiple_levels()
 
     c.visualize(optimize_function=opt_fn)
 
@@ -921,6 +926,16 @@ def test_fuse_merge_chunks_unary(spec):
     result = c.compute(optimize_function=opt_fn)
     assert_array_equal(result, -np.ones((3, 2)))
 
+    # now set max_total_num_input_blocks=None which means
+    # "only fuse if ops have same number of tasks", which they don't here
+    opt_fn = fuse_multiple_levels(max_total_num_input_blocks=None)
+    optimized_dag = c.plan.optimize(optimize_function=opt_fn).dag
+
+    # merge_chunks uses a hidden op and array for block ids - ignore when comparing structure
+    assert not structurally_equivalent(
+        optimized_dag, expected_fused_dag, remove_hidden=True
+    )
+
 
 # merge chunks with different number of tasks (c has more tasks than d)
 #
@@ -936,8 +951,7 @@ def test_fuse_merge_chunks_binary(spec):
     c = xp.add(a, b)
     d = merge_chunks(c, chunks=(3, 2))
 
-    # specify max_total_num_input_blocks to force d to fuse
-    opt_fn = fuse_multiple_levels(max_total_num_input_blocks=6)
+    opt_fn = fuse_multiple_levels()
 
     d.visualize(optimize_function=opt_fn)
 
@@ -963,15 +977,24 @@ def test_fuse_merge_chunks_binary(spec):
     result = d.compute(optimize_function=opt_fn)
     assert_array_equal(result, 2 * np.ones((3, 2)))
 
+    # now set max_total_num_input_blocks=None which means
+    # "only fuse if ops have same number of tasks", which they don't here
+    opt_fn = fuse_multiple_levels(max_total_num_input_blocks=None)
+    optimized_dag = d.plan.optimize(optimize_function=opt_fn).dag
+
+    # merge_chunks uses a hidden op and array for block ids - ignore when comparing structure
+    assert not structurally_equivalent(
+        optimized_dag, expected_fused_dag, remove_hidden=True
+    )
+
 
 # like test_fuse_merge_chunks_unary, except uses partial_reduce
 def test_fuse_partial_reduce_unary(spec):
     a = xp.ones((3, 2), chunks=(1, 2), spec=spec)
     b = xp.negative(a)
     c = partial_reduce(b, np.sum, split_every={0: 3})
 
-    # specify max_total_num_input_blocks to force c to fuse
-    opt_fn = fuse_multiple_levels(max_total_num_input_blocks=3)
+    opt_fn = fuse_multiple_levels()
 
     c.visualize(optimize_function=opt_fn)
 
@@ -996,8 +1019,7 @@ def test_fuse_partial_reduce_binary(spec):
     c = xp.add(a, b)
     d = partial_reduce(c, np.sum, split_every={0: 3})
 
-    # specify max_total_num_input_blocks to force d to fuse
-    opt_fn = fuse_multiple_levels(max_total_num_input_blocks=6)
+    opt_fn = fuse_multiple_levels()
 
     d.visualize(optimize_function=opt_fn)
 
@@ -1176,7 +1198,7 @@ def test_optimize_stack(spec):
     c = xp.stack((a, b), axis=0)
     d = c + 1
     # try to fuse all ops into one (d will fuse with c, but c won't fuse with a and b)
-    d.compute(optimize_function=fuse_multiple_levels(max_total_num_input_blocks=10))
+    d.compute(optimize_function=fuse_multiple_levels())
 
 
 def test_optimize_concat(spec):
@@ -1186,4 +1208,4 @@ def test_optimize_concat(spec):
     c = xp.concat((a, b), axis=0)
     d = c + 1
     # try to fuse all ops into one (d will fuse with c, but c won't fuse with a and b)
-    d.compute(optimize_function=fuse_multiple_levels(max_total_num_input_blocks=10))
+    d.compute(optimize_function=fuse_multiple_levels())

docs/user-guide/optimization.md

@@ -112,9 +112,9 @@ e.visualize(optimize_function=opt_fn)
 
 The `max_total_num_input_blocks` argument to `multiple_inputs_optimize_dag` specifies the maximum number of input blocks (chunks) that are allowed in the fused operation.
 
-Again, this is to limit the number of reads that an individual task must perform. The default is `None`, which means that operations are fused only if they have the same number of tasks. If set to an integer, then this limitation is removed, and tasks with a different number of tasks will be fused - as long as the total number of input blocks does not exceed the maximum. This setting is useful for reductions, and can be set using `functools.partial`:
+Again, this is to limit the number of reads that an individual task must perform. If set to `None`, operations are fused only if they have the same number of tasks. If set to an integer (the default is 10), then tasks with a different number of tasks will be fused - as long as the total number of input blocks does not exceed the maximum. This setting is useful for reductions, and can be changed using `functools.partial`:
 
 ```python
-opt_fn = partial(multiple_inputs_optimize_dag, max_total_num_input_blocks=10)
+opt_fn = partial(multiple_inputs_optimize_dag, max_total_num_input_blocks=20)
 e.visualize(optimize_function=opt_fn)
 ```