|
| 1 | +# Copyright (c) Meta Platforms, Inc. and affiliates. |
| 2 | +# All rights reserved. |
| 3 | +# |
| 4 | +# This source code is licensed under the BSD-style license found in the |
| 5 | +# LICENSE file in the root directory of this source tree. |
| 6 | + |
| 7 | +# pyre-strict |
| 8 | + |
| 9 | +import unittest |
| 10 | + |
| 11 | +import torch |
| 12 | +import torch.fx |
| 13 | +from executorch.backends.cadence.aot.memory_constraints import MemConstraints |
| 14 | +from executorch.backends.cadence.aot.memory_planning import ( |
| 15 | + PositionBasedGreedyWithHierarchy, |
| 16 | +) |
| 17 | +from executorch.backends.cadence.aot.memory_planning_algo import MemoryPlanningState |
| 18 | +from executorch.backends.cadence.aot.utils import MemoryConfig |
| 19 | +from executorch.exir.tensor import TensorSpec |
| 20 | + |
| 21 | + |
| 22 | +def _make_spec(shape: list[int], *, mem_id: int | None = None) -> TensorSpec: |
| 23 | + """Create a TensorSpec for a uint8 tensor of given shape, optionally pre-pinning mem_id.""" |
| 24 | + spec = TensorSpec(dtype=torch.uint8, shape=torch.Size(shape)) |
| 25 | + # The planner's overlap checker requires valid lifetimes on every spec. |
| 26 | + spec.lifetime = [0, 1] |
| 27 | + if mem_id is not None: |
| 28 | + spec.mem_id = mem_id |
| 29 | + return spec |
| 30 | + |
| 31 | + |
| 32 | +def _make_algo_and_state( |
| 33 | + mem_sizes: list[int], |
| 34 | +) -> tuple[PositionBasedGreedyWithHierarchy, MemoryPlanningState, MemConstraints]: |
| 35 | + """Build a 2-memory config planner (mem_id 1 = fast, 2 = slow) for tests.""" |
| 36 | + config = MemoryConfig(mem_sizes) |
| 37 | + algo = PositionBasedGreedyWithHierarchy(config) |
| 38 | + state = MemoryPlanningState(config) |
| 39 | + constraints = MemConstraints() |
| 40 | + return algo, state, constraints |
| 41 | + |
| 42 | + |
| 43 | +class TestPinnedMemIdPromotion(unittest.TestCase): |
| 44 | + """Tests for plan_with_constraints pre-set mem_id → AbsolutePlacementConstraint promotion.""" |
| 45 | + |
| 46 | + def _run( |
| 47 | + self, |
| 48 | + specs: list[TensorSpec], |
| 49 | + mem_sizes: list[int], |
| 50 | + ) -> None: |
| 51 | + algo, state, constraints = _make_algo_and_state(mem_sizes) |
| 52 | + gm = torch.fx.GraphModule({}, torch.fx.Graph()) |
| 53 | + algo.plan_with_constraints( |
| 54 | + specs, gm, None, state, constraints |
| 55 | + ) # pyre-ignore[6] |
| 56 | + |
| 57 | + def test_spec_without_preset_mem_id_planned_freely(self) -> None: |
| 58 | + """A spec with no pre-set mem_id is placed by the greedy algo in mem_id=1.""" |
| 59 | + spec = _make_spec([512]) |
| 60 | + self._run([spec], mem_sizes=[1024, 1024]) |
| 61 | + self.assertIsNotNone(spec.mem_id) |
| 62 | + self.assertEqual(spec.mem_id, 1) |
| 63 | + self.assertIsNotNone(spec.mem_offset) |
| 64 | + |
| 65 | + def test_spec_with_preset_mem_id_stays_in_that_memory(self) -> None: |
| 66 | + """A spec with pre-set mem_id=2 stays in memory 2 even though memory 1 is faster.""" |
| 67 | + spec = _make_spec([256]) |
| 68 | + spec.mem_id = 2 |
| 69 | + self._run([spec], mem_sizes=[4096, 4096]) |
| 70 | + # mem_id must be preserved as 2 |
| 71 | + self.assertEqual(spec.mem_id, 2) |
| 72 | + # Must have a valid offset assigned |
| 73 | + assert spec.mem_offset is not None |
| 74 | + assert spec.mem_offset >= 0 |
| 75 | + |
| 76 | + def test_preset_mem_id_offset_computed_by_planner(self) -> None: |
| 77 | + """Two specs pinned to mem_id=2 get distinct non-overlapping offsets.""" |
| 78 | + spec_a = _make_spec([100]) |
| 79 | + spec_b = _make_spec([200]) |
| 80 | + spec_a.mem_id = 2 |
| 81 | + spec_b.mem_id = 2 |
| 82 | + self._run([spec_a, spec_b], mem_sizes=[4096, 4096]) |
| 83 | + self.assertEqual(spec_a.mem_id, 2) |
| 84 | + self.assertEqual(spec_b.mem_id, 2) |
| 85 | + # Offsets must not overlap: [a_start, a_end) ∩ [b_start, b_end) == ∅ |
| 86 | + a_end = spec_a.mem_offset + spec_a.allocated_memory |
| 87 | + b_end = spec_b.mem_offset + spec_b.allocated_memory |
| 88 | + no_overlap = spec_a.mem_offset >= b_end or spec_b.mem_offset >= a_end |
| 89 | + self.assertTrue(no_overlap, f"Specs overlap: {spec_a} and {spec_b}") |
| 90 | + |
| 91 | + def test_unpinned_spec_unaffected_by_pinned_peers(self) -> None: |
| 92 | + """Specs without pre-set mem_id are not forced into the pinned tier.""" |
| 93 | + pinned = _make_spec([128]) |
| 94 | + pinned.mem_id = 2 |
| 95 | + free = _make_spec([64]) # No preset; greedy should pick mem_id=1 |
| 96 | + self._run([pinned, free], mem_sizes=[4096, 4096]) |
| 97 | + self.assertEqual(pinned.mem_id, 2) |
| 98 | + # Greedy algo prefers mem_id=1 (faster) for unconstrained specs |
| 99 | + self.assertEqual(free.mem_id, 1) |
| 100 | + |
| 101 | + def test_already_constrained_spec_not_overridden(self) -> None: |
| 102 | + """A spec that already has an AbsolutePlacementConstraint is not double-promoted.""" |
| 103 | + from executorch.backends.cadence.aot.memory_constraints import ( |
| 104 | + AbsolutePlacementConstraint, |
| 105 | + ) |
| 106 | + |
| 107 | + spec = _make_spec([256]) |
| 108 | + spec.mem_id = 1 # will be set but constraint added externally to mem_id=2 |
| 109 | + |
| 110 | + algo, state, constraints = _make_algo_and_state([4096, 4096]) |
| 111 | + # Add an explicit constraint to mem_id=2 (overrides the spec.mem_id=1 preset) |
| 112 | + constraints.set_absolute_placement_constraint( |
| 113 | + spec, AbsolutePlacementConstraint(pinned_memory_id=2) |
| 114 | + ) |
| 115 | + gm = torch.fx.GraphModule({}, torch.fx.Graph()) |
| 116 | + algo.plan_with_constraints( |
| 117 | + [spec], gm, None, state, constraints |
| 118 | + ) # pyre-ignore[6] |
| 119 | + # The existing constraint (mem_id=2) takes precedence over spec.mem_id=1 |
| 120 | + self.assertEqual(spec.mem_id, 2) |
| 121 | + |
| 122 | + def test_mem_id_zero_treated_as_unpinned(self) -> None: |
| 123 | + """A spec with mem_id=0 (sentinel for unassigned) should be planned freely.""" |
| 124 | + spec = _make_spec([512], mem_id=0) |
| 125 | + self._run([spec], mem_sizes=[1024, 1024]) |
| 126 | + # Greedy algo picks mem_id=1 for unconstrained specs |
| 127 | + self.assertEqual(spec.mem_id, 1) |
| 128 | + self.assertIsNotNone(spec.mem_offset) |
| 129 | + |
| 130 | + def test_mem_id_out_of_range_raises(self) -> None: |
| 131 | + """A spec with mem_id >= num_memories should raise AssertionError.""" |
| 132 | + # With 2 memory tiers, valid mem_ids are 1 and 2; mem_id=3 is out of range. |
| 133 | + spec = _make_spec([256], mem_id=3) |
| 134 | + with self.assertRaises(AssertionError): |
| 135 | + self._run([spec], mem_sizes=[4096, 4096]) |
| 136 | + |
| 137 | + def test_mem_id_negative_treated_as_unpinned(self) -> None: |
| 138 | + """A spec with negative mem_id should be treated as unpinned (not promoted).""" |
| 139 | + spec = _make_spec([256]) |
| 140 | + spec.mem_id = -1 |
| 141 | + self._run([spec], mem_sizes=[1024, 1024]) |
| 142 | + # Negative mem_id is filtered out by the >0 check; greedy picks mem_id=1 |
| 143 | + self.assertEqual(spec.mem_id, 1) |
| 144 | + self.assertIsNotNone(spec.mem_offset) |
| 145 | + |
| 146 | + |
| 147 | +if __name__ == "__main__": |
| 148 | + unittest.main() |
![meta-codesync[bot]](https://avatars.githubusercontent.com/in/1373969?v=4&size=40){kind=avatar}
0 commit comments