Improve memory planning for submodule hierarchies. (#11860)

Summary:

Improves the memory planning across hierarchies in apply_algo in memory_planning.py:
1. Plan memory bottom-to-top, starting with the leaf submodules and ending at top-level graph module (root). This is now consistent with how delegates are compiled / memory planned. Future PRs/diffs will add support for planned buffers in delegates.
2. Allocate max bufsize for all submodules as `graph_module.meta['input_mem_buffer_sizes']`, rather than sum. This allows us to reclaim the space used by one submodule for another submodule.


Before this change the apply_algo in memory_planning.py would:
1. Plan memory top-to-bottom, starting with the top-level graph module (root).
2. Populate the `input_mem_buffer_sizes` so that each new submodule will allocate memory after the max buffer size of previous memory. 

For example:
```
root [A bytes]
- root.child0 [B bytes]
   - root.child0.child0 [C bytes]
- root.child1 [D bytes]
```
(before this diff) Planned memory looks like:
```
--- A + B + C + D ----------------
Space for root.child1
--- A + B + C --------------------
Space for root.child0.child0
--- A + B ------------------------
Space for root.child0
--- A ----------------------------
Space for root
--- 0 ----------------------------
```
Note that tensors for child0 and child1 have no overlap but still use completely different space.

(after this diff) Planned memory looks like:
```
--- max(C + B, D) + A ----------
root
--- max(C + B, D) --------------
root.child0        |
--- C ------------ | root.child1
root.child0.child0 | 
--- 0 --------------------------
```

Note:
We can update memory planning algo to plan nodes with submodules (while/map/cond or even delegate) to use `graph_module.meta['non_const_buffer_size']` and reduce space even further. Implementation for this is not part of this PR/Diff. This will allow us to reuse space for `root.child0.child0` in `root.child0`, and space for `root.child0`/`root.child1` in `root.

Differential Revision: D76940237

Author: hsharma35

exir/memory_planning.py

@@ -10,10 +10,20 @@
 import itertools
 import logging
 import operator
-import typing
 from collections import defaultdict
 from dataclasses import dataclass, field
-from typing import Any, Callable, Dict, Iterable, List, Optional, Set, Tuple, Union
+from typing import (
+    Any,
+    Callable,
+    cast,
+    Dict,
+    Iterable,
+    List,
+    Optional,
+    Set,
+    Tuple,
+    Union,
+)
 
 import torch
 from executorch.exir import memory
@@ -960,7 +970,7 @@ def _allocate_buf(bufsizes: List[int], mem_id: int, allocated: int) -> int:
     bufsizes = getattr(graph_module, "input_mem_buffer_sizes", None)
     if bufsizes is None:
         bufsizes = [0, 0]
-    bufsizes = typing.cast(List[int], bufsizes)
+    bufsizes = cast(List[int], bufsizes)
 
     for spec in specs:
         spec_alloc_result = naive_result.spec_dict.get(spec, SpecAllocResult(0, 0, 0))
@@ -1062,92 +1072,158 @@ def insert_calls_to_free(
     graph_module.recompile()
 
 
+def _merge_bufsizes(bufsizes: list[int], new_bufsizes: list[int]) -> list[int]:
+    """Combine two buffer size lists."""
+    if len(bufsizes) < len(new_bufsizes):
+        bufsizes.extend([0] * (len(new_bufsizes) - len(bufsizes)))
+    for i in range(len(new_bufsizes)):
+        bufsizes[i] = max(bufsizes[i], new_bufsizes[i])
+    return bufsizes
+
+
+def _handle_submodule(
+    algo: Callable[..., list[int]],
+    parent_graph_module: torch.fx.GraphModule,
+    alignment: int,
+    submodule_node: torch.fx.Node,
+    graph_signature: Optional[ExportGraphSignature] = None,
+    alloc_graph_input: bool = False,
+) -> list[int]:
+    """Apply algo to nodes in a submodule of the graph module."""
+    assert submodule_node.op == "get_attr"
+    submodule = getattr(parent_graph_module, submodule_node.target)
+
+    logging.debug(f"Planning memory for submodule {submodule_node.name}...")
+    bufsizes = apply_algo(
+        algo,
+        submodule,
+        alignment,
+        graph_signature,
+        alloc_graph_input=alloc_graph_input,
+        alloc_graph_output=True,
+    )
+    submodule.meta.update({"non_const_buffer_sizes": bufsizes})
+    logging.debug(f"Buffer sizes for submodule {submodule_node.name}: {bufsizes}")
+    return bufsizes
+
+
+def _apply_algo_to_submodules(
+    algo: Callable[..., list[int]],
+    graph_module: torch.fx.GraphModule,
+    alignment: int,
+    graph_signature: Optional[ExportGraphSignature] = None,
+) -> list[int]:
+    """Apply algo to map/cond/while nodes in the graph module.
+
+    This method will popuate graph_module.meta["non_const_buffer_sizes"] for
+    all submodules and return a bufsizes list that is the maximum size of all
+    buffers.
+    """
+
+    # Bufsizes for submodules.
+    bufsizes: list[int] = []
+
+    def _handle(
+        submodule_node: torch.fx.Node,
+        alloc_graph_input: bool = False,
+    ) -> None:
+        current_bufsizes = _handle_submodule(
+            algo,
+            graph_module,
+            alignment,
+            submodule_node,
+            graph_signature,
+            alloc_graph_input=alloc_graph_input,
+        )
+        nonlocal bufsizes
+        _merge_bufsizes(bufsizes, current_bufsizes)
+
+    for cond_node in get_cond_nodes(graph_module):
+        _handle(cast(torch.fx.Node, cond_node.args[1]))
+        _handle(cast(torch.fx.Node, cond_node.args[2]))
+
+    for while_node in get_while_nodes(graph_module):
+        _handle(cast(torch.fx.Node, while_node.args[0]))
+        _handle(cast(torch.fx.Node, while_node.args[1]))
+
+    for map_node in get_map_nodes(graph_module):
+        _handle(cast(torch.fx.Node, map_node.args[0]), alloc_graph_input=True)
+
+    # TODO: We can handle delegates the same way as map/cond/while.
+    # Maybe populate the graph_module.meta["non_const_buffer_sizes"] for delegates.
+
+    return bufsizes
+
+
 def apply_algo(
-    algo: Callable[
-        ...,
-        List[int],
-    ],
+    algo: Callable[..., list[int]],
     graph_module: torch.fx.GraphModule,
     alignment: int,
     graph_signature: Optional[ExportGraphSignature] = None,
     alloc_graph_input: bool = True,
     alloc_graph_output: bool = True,
     alloc_mutable_buffers: bool = True,
-) -> List[int]:
+) -> list[int]:
     """
     Recursively apply algo to graph_module and its submodules for control flow.
 
-    Quite naively right now since it does not take the following optimizations
-    into considerating:
-    1. for conditional structure, true branch and false true does not overlap
-       in lifetime and can share tensor storage
-    2. tensors inside a submodule (e.g. true branch) has opportunities to share
-       storage with tensors in the outer module.
-    TODO: make these optimizations once we have some baseline working.
+    Algo implementation should handle one of two meta entries for submodules:
+    1. input_mem_buffer_sizes: List of int offset bytes. Memory allocated by
+       `algo` should start at the offset specified by this list;
+    OR
+    2. non_const_buffer_sizes: List of bufsizes for planned memory in submodule.
+       `algo` should reserve the space specified by this list for the lifetime
+       of the submodule node (e.g. cond, while, map).
+
+    TODO: Missing optimizations:
+    1. To handle maps, we set `alloc_graph_input=True`, which allocates
+    appropriate space for mapped arg but ends up allocating extra space for
+    `operand` arg. The memory for operands is unused.
     """
     # Extract the nodes and their lifespans from the graph_module
     # Difficult to just filter the list of specs returned by this due to
     # how we flag trainable weights.
     _ = update_all_tensors_lifetime(graph_module, graph_signature)
+
     # Filter specs based on alloc_graph_input and alloc_graph_output
-    specs = collect_specs_from_nodes(
-        graph_module.graph.nodes,
-        graph_signature,
-        do_assertion=False,
-        ignore_graph_input=not alloc_graph_input,
-        ignore_graph_output=not alloc_graph_output,
-        ignore_mutable_buffers=not alloc_mutable_buffers,
+    specs = set(
+        collect_specs_from_nodes(
+            graph_module.graph.nodes,
+            graph_signature,
+            do_assertion=False,
+            ignore_graph_input=not alloc_graph_input,
+            ignore_graph_output=not alloc_graph_output,
+            ignore_mutable_buffers=not alloc_mutable_buffers,
+        )
     )
 
+    # Get temporary specs for submodules to set aside space during execution
+    # of submodules.
+    submodule_bufsizes = _apply_algo_to_submodules(
+        algo, graph_module, alignment, graph_signature
+    )
+
+    # Update `input_mem_buffer_sizes` in graph_module. This will allow existing
+    # algos to work using `input_mem_buffer_sizes` or use
+    # `non_const_buffer_sizes` directly.
+    # pyre-ignore[16]: `torch.fx.GraphModule` has no attribute `input_mem_buffer_sizes`.
+    graph_module.input_mem_buffer_sizes = submodule_bufsizes
+
     # Get extra padding for XNNPACK if needed
     extra_padding = 0
     if _contains_xnnpack_delegate(graph_module):
         extra_padding = 64
 
     # Pass the filtered specs to the algorithm
-    bufsizes: List[int] = algo(
+    bufsizes: list[int] = algo(
         alignment,
         specs,
         graph_module,
         graph_signature,
         extra_padding,
     )
 
-    insert_calls_to_free(graph_module, set(specs))
-
-    def handle_submodule(
-        submodule_nd: torch.fx.Node, alloc_graph_input: bool = False
-    ) -> None:
-        nonlocal bufsizes
-        assert submodule_nd.op == "get_attr"
-        submodule = getattr(graph_module, submodule_nd.target)
-        # memory planning for submodule need to be aware of the amount of
-        # buffer already allocated.
-        submodule.input_mem_buffer_sizes = bufsizes
-
-        bufsizes = apply_algo(
-            algo,
-            submodule,
-            alignment,
-            graph_signature,
-            alloc_graph_input=alloc_graph_input,
-            alloc_graph_output=True,
-        )
-        submodule.meta.update({"non_const_buffer_sizes": bufsizes})
-
-    for cond_node in get_cond_nodes(graph_module):
-        handle_submodule(typing.cast(torch.fx.Node, cond_node.args[1]))
-        handle_submodule(typing.cast(torch.fx.Node, cond_node.args[2]))
-
-    for while_node in get_while_nodes(graph_module):
-        handle_submodule(typing.cast(torch.fx.Node, while_node.args[0]))
-        handle_submodule(typing.cast(torch.fx.Node, while_node.args[1]))
-    # TODO: Add test coverage for map operator once dynamo tracing is
-    # fully supported for this. T142287208
-    for map_node in get_map_nodes(graph_module):
-        handle_submodule(
-            typing.cast(torch.fx.Node, map_node.args[0]), alloc_graph_input=True
-        )
+    insert_calls_to_free(graph_module, specs)
 
     graph_module.meta.update({"non_const_buffer_sizes": bufsizes})
     return bufsizes

exir/tests/test_memory_planning.py

@@ -33,6 +33,8 @@
     ToOutVarPass,
 )
 from executorch.exir.passes.sym_shape_eval_pass import ConstraintBasedSymShapeEvalPass
+from executorch.exir.tensor import TensorSpec
+from functorch.experimental.control_flow import map as torch_map
 from parameterized import parameterized
 
 from torch import nn
@@ -56,6 +58,7 @@
 from torch.export.exported_program import ExportGraphSignature
 from torch.fx import Graph, GraphModule, Node
 from torch.nn import functional as F
+from torch.utils import _pytree as pytree
 
 torch.ops.load_library("//executorch/kernels/portable:custom_ops_generated_lib")
 
@@ -471,13 +474,13 @@ def test_graph_input_output(self) -> None:
             alloc_graph_output,
             alloc_mutable_buffers,
         ) in itertools.product([True, False], [True, False], [True, False]):
-            case = maketest(
+            test = maketest(
                 ModelWithDifferentTensorSizes,
                 alloc_graph_input=alloc_graph_input,
                 alloc_graph_output=alloc_graph_output,
                 alloc_mutable_buffer=alloc_mutable_buffers,
             )
-            case(self)
+            test(self)
 
 
 class TestVerifier(unittest.TestCase):
@@ -839,3 +842,97 @@ def forward(self, input, label):
             .val.allocation_info,  # pyright: ignore
             None,
         )
+
+
+def _get_specs(gm: torch.fx.GraphModule) -> list[TensorSpec]:
+    return list(
+        filter(
+            None,
+            pytree.tree_flatten(
+                pytree.tree_map_only(
+                    torch.fx.Node,
+                    lambda n: n.meta.get("spec", None),
+                    list(gm.graph.nodes),
+                )
+            )[0],
+        )
+    )
+
+
+class TestMap(unittest.TestCase):
+    class MapModel(torch.nn.Module):
+        def __init__(self) -> None:
+            super().__init__()
+
+        def forward(self, x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
+            # Use actual torch.map function for memory planning testing
+            def add_fn(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
+                return a + b
+
+            # Use torch.map to apply function over first dimension
+            # pyre-ignore[6]: For 3rd argument expected `TypeVarTuple` but got `Tensor`.
+            map_output = torch_map(add_fn, x, y)
+
+            return map_output + y
+
+        def get_random_inputs(self) -> Tuple[torch.Tensor, ...]:
+            return (torch.randn(5, 3), torch.randn(3))
+
+    def test_map(self) -> None:
+        """Test memory planning for torch.map operations."""
+
+        eager_module = self.MapModel().eval()
+        inputs = eager_module.get_random_inputs()
+
+        # Export and convert to edge
+        graph_module = (
+            to_edge(export(eager_module, inputs, strict=True))
+            .exported_program()
+            .graph_module
+        )
+
+        # Apply memory planning.
+        mem_algo = MemoryPlanningAlgorithmSuite(algo_list=[naive, greedy])
+        graph_module = PassManager(
+            passes=[
+                SpecPropPass(),
+                ToOutVarPass(),
+            ],
+        )(graph_module).graph_module
+        mem_planning_pass = MemoryPlanningPass(
+            mem_algo,
+            alloc_graph_input=True,
+            alloc_graph_output=True,
+            alloc_mutable_buffers=True,
+        )
+        graph_module = mem_planning_pass.run(graph_module).graph_module
+
+        # Verify memory planning results
+        verifier = Verifier(
+            graph_module,
+            alloc_graph_input=True,
+            alloc_graph_output=True,
+            alloc_mutable_buffers=True,
+        )
+        verifier.verify_graph_input_output()
+        verifier.verify_storage_reuse(allow_lifetime_and_storage_overlap=False)
+
+        map_node = graph_module.graph.find_nodes(
+            op="call_function", target=torch.ops.higher_order.map_impl
+        )[0]
+        map_fn_node = map_node.args[0]
+        self.assertEqual(map_fn_node.op, "get_attr")
+        map_fn = getattr(graph_module, map_fn_node.target)
+
+        map_lifetime = map_node.meta.get("spec", None)[0].lifetime[0]
+
+        # Check that there is no storage overlap between nodes of the outer program and submodule of map.
+        for outer_spec in _get_specs(graph_module):
+            for inner_spec in _get_specs(map_fn):
+                self.assertFalse(
+                    verifier.has_overlap(
+                        outer_spec.lifetime, [map_lifetime, map_lifetime]
+                    )
+                    and (verifier.storage_overlap(outer_spec, inner_spec)),
+                    f"Outer spec {outer_spec.shape=} {outer_spec.dtype=} {outer_spec.lifetime=} and inner spec {inner_spec} have storage overlap",
+                )