Handle rank-changing views in RemovePermutesAroundElementwiseOps (#19538)

Summary:

Extend RemovePermutesAroundElementwiseOps to cancel permute pairs across rank-changing squeeze/unsqueeze view boundaries. When a permute's sole user is a view_copy that adds or removes a single size-1 dimension, the pass adapts the expected permutation to the new rank and continues traversal. This enables removing permutes that sit on opposite sides of an unsqueeze→elementwise→squeeze chain (e.g. the NHWC↔NTC layout conversion around convolutions in the cascade detector model).

Key changes:
- Accept extra_permutable_ops constructor parameter for backend-specific ops
- Track per-node expected permutations across view boundaries
- Run dimension updates before edges_in bypass to preserve original metadata
- Handle view_copy, unsqueeze_copy, squeeze_copy rank changes
- Treat aten.full.default as a compile-time constant

Note: The PostponePermuteOpBelowSqueezeOrUnsqueezeLikeView pass is removed from the Arm pass manager, since it doesn't actually help anymore.

Reviewed By: DrJessop

Differential Revision: D104775244

Author: mcremon-meta

backends/arm/_passes/arm_pass_manager.py

@@ -160,10 +160,6 @@
 from executorch.backends.transforms.fuse_cascaded_transpose_or_permute_ops import (
     FuseCascadedTransposeOrPermuteOps,
 )
-from executorch.backends.transforms.postpone_permute_below_squeeze_view import (
-    PostponePermuteOpBelowSqueezeOrUnsqueezeLikeView,
-)
-
 from executorch.exir import ExportedProgram
 from executorch.exir.pass_base import ExportPass
 from executorch.exir.pass_manager import PassManager
@@ -531,7 +527,6 @@ def _tosa_pipeline(
                 RewritePadPass(),
                 FuseViewCopyTransformPass(),
                 RemovePermutesAroundElementwiseTosaOps(),
-                PostponePermuteOpBelowSqueezeOrUnsqueezeLikeView(),
                 FuseCascadedTransposeOrPermuteOps(),
                 ConvertPermuteSingletonToViewPass(),
                 RewriteHighRankSingletonPermutePass(),

backends/arm/_passes/remove_permutes_around_elementwise_tosa_ops.py

@@ -11,13 +11,15 @@
 
 
 class RemovePermutesAroundElementwiseTosaOps(RemovePermutesAroundElementwiseOps):
-    permutable_ops = {
-        *RemovePermutesAroundElementwiseOps.permutable_ops,
-        *TableOps.unary_table_ops.keys(),
-        *TableOps.special_table_ops,
-        exir_ops.backend.tosa.RESCALE.default,
-        exir_ops.backend.tosa.TABLE.default,
-    }
+    def __init__(self) -> None:
+        super().__init__(
+            extra_permutable_ops={
+                *TableOps.unary_table_ops.keys(),
+                *TableOps.special_table_ops,
+                exir_ops.backend.tosa.RESCALE.default,
+                exir_ops.backend.tosa.TABLE.default,
+            }
+        )
 
     def permute_subgraph(self, subgraph):
         # Original function will always permute constant nodes which is wrong for table ops

backends/cadence/aot/remove_ops.py

@@ -603,16 +603,16 @@ def maybe_remove_or_replace(self, node: Node) -> bool:
 
 @register_cadence_pass(CadencePassAttribute(opt_level=2))
 class RemovePermutesAroundElementwiseOps(_SharedRemovePermutesAroundElementwiseOps):
-    permutable_ops: set[EdgeOpOverload] = (
-        _SharedRemovePermutesAroundElementwiseOps.permutable_ops
-        | {
-            exir_ops.edge.cadence.quantize_per_tensor.default,
-            exir_ops.edge.cadence.dequantize_per_tensor.default,
-            exir_ops.edge.cadence.quantized_relu.per_tensor,
-            exir_ops.edge.cadence.requantize.per_tensor,
-            exir_ops.edge.cadence.quantized_add.per_tensor,
-        }
-    )
+    def __init__(self) -> None:
+        super().__init__(
+            extra_permutable_ops={
+                exir_ops.edge.cadence.quantize_per_tensor.default,
+                exir_ops.edge.cadence.dequantize_per_tensor.default,
+                exir_ops.edge.cadence.quantized_relu.per_tensor,
+                exir_ops.edge.cadence.requantize.per_tensor,
+                exir_ops.edge.cadence.quantized_add.per_tensor,
+            }
+        )
 
 
 @register_cadence_pass(CadencePassAttribute(opt_level=2))