Merge back to back slices on the same dim (pytorch#19120)

Andrew Grebenisan · meta-codesync[bot] · commit fdb522f3e547 · 2026-04-24T16:17:55.000-07:00
Summary: Pull Request resolved: pytorch#19120 If we have back to back slices on the same dimension, we can remove the top slice and just perform the second one. RemoveOrReplacePassInterface will handle a whole cascade if it exists. Reviewed By: abeakkas Differential Revision: D102278253
diff --git a/backends/cadence/aot/fuse_ops.py b/backends/cadence/aot/fuse_ops.py
@@ -31,6 +31,7 @@
     HierarchicalInplacePassInterface,
     register_cadence_pass,
     RemoveOrReplacePassInterface,
+    set_arg,
 )
 from executorch.backends.cadence.aot.utils import get_edge_overload_packet
 from executorch.backends.transforms.fuse_cascaded_transpose_or_permute_ops import (
@@ -1105,6 +1106,75 @@ def maybe_remove_or_replace(self, node: torch.fx.Node) -> bool:
         return True
 
 
+@register_cadence_pass(CadencePassAttribute(opt_level=0))
+class FuseSliceSameDimPass(RemoveOrReplacePassInterface):
+    """Fuse chained slices on the same dim into a single slice.
+
+    When a slice_copy's input is another slice_copy on the same dimension
+    with step=1, the child slice can read directly from the grandparent
+    with merged indices, eliminating the intermediate slice.
+
+    Handles negative start/end indices by canonicalizing them against the
+    relevant dimension size before merging.
+    """
+
+    @staticmethod
+    def _canonicalize(val: int, dim_size: int) -> int:
+        return val + dim_size if val < 0 else val
+
+    @property
+    def targets(self) -> list[EdgeOpOverload]:
+        return [exir_ops.edge.aten.slice_copy.Tensor]
+
+    def maybe_remove_or_replace(self, node: torch.fx.Node) -> bool:
+        parent = get_arg(node, "input", torch.fx.Node)
+        if parent.target != exir_ops.edge.aten.slice_copy.Tensor:
+            return False
+
+        grandparent = get_arg(parent, "input", torch.fx.Node)
+        ndim = len(grandparent.meta["val"].shape)
+        child_dim = get_arg(node, "dim", int) % ndim
+        parent_dim = get_arg(parent, "dim", int) % ndim
+        if child_dim != parent_dim:
+            return False
+
+        child_start = get_arg(node, "start", Optional[int])
+        child_end = get_arg(node, "end", Optional[int])
+        child_step = get_arg(node, "step", int)
+        parent_start = get_arg(parent, "start", Optional[int])
+        parent_end = get_arg(parent, "end", Optional[int])
+        parent_step = get_arg(parent, "step", int)
+
+        if child_step != 1 or parent_step != 1:
+            return False
+        if (
+            child_start is None
+            or child_end is None
+            or parent_start is None
+            or parent_end is None
+        ):
+            return False
+
+        grandparent_dim_size = grandparent.meta["val"].shape[parent_dim]
+        parent_dim_size = parent.meta["val"].shape[parent_dim]
+
+        p_start = self._canonicalize(parent_start, grandparent_dim_size)
+        p_end = self._canonicalize(parent_end, grandparent_dim_size)
+        c_start = self._canonicalize(child_start, parent_dim_size)
+        c_end = self._canonicalize(child_end, parent_dim_size)
+
+        new_start = p_start + c_start
+        new_end = min(p_start + c_end, p_end)
+
+        if new_end > grandparent_dim_size:
+            return False
+
+        node.replace_input_with(parent, grandparent)
+        set_arg(node, "start", new_start)
+        set_arg(node, "end", new_end)
+        return True
+
+
 class HierarchicalCSEPass(HierarchicalInplacePassInterface):
     """
     A hierarchical Common Subexpression Elimination (CSE) pass that recursively
@@ -1138,4 +1208,5 @@ class CadenceFuseOpsInGraph:
         FuseFullThenReshapePass,
         FuseTransposeOrPermuteOpPairsPass,
         FuseMeanKeepDimWithViewPass,
+        FuseSliceSameDimPass,
     ]
diff --git a/backends/cadence/aot/simplify_ops.py b/backends/cadence/aot/simplify_ops.py
@@ -13,6 +13,7 @@
 from typing import cast, Optional
 
 import torch
+
 from executorch.backends.cadence.aot.pass_utils import (
     CadencePassAttribute,
     register_cadence_pass,
diff --git a/backends/cadence/aot/tests/test_fusion_ops_passes.py b/backends/cadence/aot/tests/test_fusion_ops_passes.py
@@ -26,10 +26,11 @@
     FuseMulTensorIntoQuantPass,
     FuseQuantDequantToRequantizePass,
     FuseQuantizedBatchNormWithConv,
+    FuseSliceSameDimPass,
     FuseTransposeOrPermuteOpPairsPass,
     HierarchicalCSEPass,
 )
-from executorch.backends.cadence.aot.pass_utils import count_node, op_counts_match
+from executorch.backends.cadence.aot.pass_utils import count_node, get_arg, op_counts_match
 from executorch.backends.cadence.aot.typing_stubs import expand
 from executorch.backends.test.graph_builder import GraphBuilder
 from executorch.exir.dialects._ops import ops as exir_ops
@@ -1862,3 +1863,256 @@ def test_reduce_single_dim(self) -> None:
             (torch.randn(3, 4, 5),),
             "FuseMeanKeepDimWithViewPass",
         )
+
+
+class TestFuseSliceSameDimPass(TestFusionPassesBase):
+    def _get_single_slice(
+        self, gm: torch.fx.GraphModule
+    ) -> torch.fx.Node:
+        slices = gm.graph.find_nodes(
+            op="call_function", target=exir_ops.edge.aten.slice_copy.Tensor
+        )
+        self.assertEqual(len(slices), 1)
+        return slices[0]
+
+    def test_basic_chain_bypass(self) -> None:
+        """slice(dim=3, 0:78) → slice(dim=3, 0:60) → direct slice(dim=3, 0:60)."""
+        builder = GraphBuilder()
+        x = builder.placeholder("x", torch.randn(2, 3, 4, 80))
+        parent = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(x, 3, 0, 78, 1),
+        )
+        child = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(parent, 3, 0, 60, 1),
+        )
+        builder.output([child])
+        original = builder.get_graph_module()
+        gm_before = copy.deepcopy(original)
+
+        result = cast(PassResult, FuseSliceSameDimPass()(original))
+        self.assertTrue(result.modified)
+        self.assertEqual(
+            count_node(result.graph_module, exir_ops.edge.aten.slice_copy.Tensor), 1
+        )
+        merged = self._get_single_slice(result.graph_module)
+        self.assertEqual(get_arg(merged, "start"), 0)
+        self.assertEqual(get_arg(merged, "end"), 60)
+        validate_numerics(
+            gm_before,
+            result.graph_module,
+            (torch.randn(2, 3, 4, 80),),
+            "FuseSliceSameDimPass",
+        )
+
+    def test_chain_with_offset(self) -> None:
+        """slice(dim=1, 10:50) → slice(dim=1, 5:20) → direct slice(dim=1, 15:30)."""
+        builder = GraphBuilder()
+        x = builder.placeholder("x", torch.randn(4, 64))
+        parent = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(x, 1, 10, 50, 1),
+        )
+        child = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(parent, 1, 5, 20, 1),
+        )
+        builder.output([child])
+        original = builder.get_graph_module()
+        gm_before = copy.deepcopy(original)
+
+        result = cast(PassResult, FuseSliceSameDimPass()(original))
+        self.assertTrue(result.modified)
+        self.assertEqual(
+            count_node(result.graph_module, exir_ops.edge.aten.slice_copy.Tensor), 1
+        )
+        merged = self._get_single_slice(result.graph_module)
+        self.assertEqual(get_arg(merged, "start"), 15)
+        self.assertEqual(get_arg(merged, "end"), 30)
+        validate_numerics(
+            gm_before,
+            result.graph_module,
+            (torch.randn(4, 64),),
+            "FuseSliceSameDimPass",
+        )
+
+    def test_parent_kept_with_other_users(self) -> None:
+        """Parent slice has another user besides the child → parent stays."""
+        builder = GraphBuilder()
+        x = builder.placeholder("x", torch.randn(2, 3, 4, 80))
+        parent = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(x, 3, 0, 78, 1),
+        )
+        child = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(parent, 3, 0, 60, 1),
+        )
+        neg = builder.call_operator(op=exir_ops.edge.aten.neg.default, args=(parent,))
+        builder.output([child, neg])
+        original = builder.get_graph_module()
+        gm_before = copy.deepcopy(original)
+
+        result = cast(PassResult, FuseSliceSameDimPass()(original))
+        self.assertTrue(result.modified)
+        self.assertEqual(
+            count_node(result.graph_module, exir_ops.edge.aten.slice_copy.Tensor), 2
+        )
+        slices = result.graph_module.graph.find_nodes(
+            op="call_function", target=exir_ops.edge.aten.slice_copy.Tensor
+        )
+        ends = sorted(get_arg(s, "end") for s in slices)
+        self.assertEqual(ends, [60, 78])
+        validate_numerics(
+            gm_before,
+            result.graph_module,
+            (torch.randn(2, 3, 4, 80),),
+            "FuseSliceSameDimPass",
+        )
+
+    def test_different_dims_no_change(self) -> None:
+        """Chained slices on different dims → no change."""
+        builder = GraphBuilder()
+        x = builder.placeholder("x", torch.randn(8, 16, 32))
+        parent = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(x, 1, 0, 10, 1),
+        )
+        child = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(parent, 2, 0, 5, 1),
+        )
+        builder.output([child])
+        original = builder.get_graph_module()
+
+        result = cast(PassResult, FuseSliceSameDimPass()(original))
+        self.assertFalse(result.modified)
+
+    def test_step_not_one_no_change(self) -> None:
+        """Parent has step != 1 → no change."""
+        builder = GraphBuilder()
+        x = builder.placeholder("x", torch.randn(4, 64))
+        parent = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(x, 1, 0, 60, 2),
+        )
+        child = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(parent, 1, 0, 10, 1),
+        )
+        builder.output([child])
+        original = builder.get_graph_module()
+
+        result = cast(PassResult, FuseSliceSameDimPass()(original))
+        self.assertFalse(result.modified)
+
+    def test_no_chain_no_change(self) -> None:
+        """Single slice with no slice user → no change."""
+        builder = GraphBuilder()
+        x = builder.placeholder("x", torch.randn(4, 64))
+        sliced = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(x, 1, 0, 32, 1),
+        )
+        builder.output([sliced])
+        original = builder.get_graph_module()
+
+        result = cast(PassResult, FuseSliceSameDimPass()(original))
+        self.assertFalse(result.modified)
+
+    def test_child_end_clamped_to_parent_range(self) -> None:
+        """Child end exceeds parent output size → clamped to parent_end."""
+        builder = GraphBuilder()
+        x = builder.placeholder("x", torch.randn(1, 100))
+        parent = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(x, 1, 10, 50, 1),
+        )
+        child = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(parent, 1, 5, 45, 1),
+        )
+        builder.output([child])
+        original = builder.get_graph_module()
+        gm_before = copy.deepcopy(original)
+
+        result = cast(PassResult, FuseSliceSameDimPass()(original))
+        self.assertTrue(result.modified)
+        self.assertEqual(
+            count_node(result.graph_module, exir_ops.edge.aten.slice_copy.Tensor), 1
+        )
+        merged = self._get_single_slice(result.graph_module)
+        self.assertEqual(get_arg(merged, "start"), 15)
+        self.assertEqual(get_arg(merged, "end"), 50)
+        validate_numerics(
+            gm_before,
+            result.graph_module,
+            (torch.randn(1, 100),),
+            "FuseSliceSameDimPass",
+        )
+
+    def test_negative_indices(self) -> None:
+        """Negative start/end are canonicalized before merging."""
+        # Parent: slice(dim=1, 10:-10) on size 100 → [10:90], output size 80.
+        # Child: slice(dim=1, 5:-5) on size 80 → [5:75], output size 70.
+        # Merged: [15:85].
+        builder = GraphBuilder()
+        x = builder.placeholder("x", torch.randn(1, 100))
+        parent = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(x, 1, 10, -10, 1),
+        )
+        child = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(parent, 1, 5, -5, 1),
+        )
+        builder.output([child])
+        original = builder.get_graph_module()
+        gm_before = copy.deepcopy(original)
+
+        result = cast(PassResult, FuseSliceSameDimPass()(original))
+        self.assertTrue(result.modified)
+        self.assertEqual(
+            count_node(result.graph_module, exir_ops.edge.aten.slice_copy.Tensor), 1
+        )
+        merged = self._get_single_slice(result.graph_module)
+        self.assertEqual(get_arg(merged, "start"), 15)
+        self.assertEqual(get_arg(merged, "end"), 85)
+        validate_numerics(
+            gm_before,
+            result.graph_module,
+            (torch.randn(1, 100),),
+            "FuseSliceSameDimPass",
+        )
+
+    def test_negative_dim(self) -> None:
+        """Negative dim is canonicalized so matching works across conventions."""
+        builder = GraphBuilder()
+        x = builder.placeholder("x", torch.randn(2, 3, 4, 5))
+        parent = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(x, -1, 0, 4, 1),
+        )
+        child = builder.call_operator(
+            op=exir_ops.edge.aten.slice_copy.Tensor,
+            args=(parent, 3, 0, 2, 1),
+        )
+        builder.output([child])
+        original = builder.get_graph_module()
+        gm_before = copy.deepcopy(original)
+
+        result = cast(PassResult, FuseSliceSameDimPass()(original))
+        self.assertTrue(result.modified)
+        self.assertEqual(
+            count_node(result.graph_module, exir_ops.edge.aten.slice_copy.Tensor), 1
+        )
+        merged = self._get_single_slice(result.graph_module)
+        self.assertEqual(get_arg(merged, "start"), 0)
+        self.assertEqual(get_arg(merged, "end"), 2)
+        validate_numerics(
+            gm_before,
+            result.graph_module,
+            (torch.randn(2, 3, 4, 5),),
+            "FuseSliceSameDimPass",
+        )
