Arm backend: Match arg ranks for min/max ops

YufengShi-dudu · ArmRyan · robell · YufengShi-dudu · commit b28d0a6a66bb · 2025-12-11T13:45:08.000Z
- Int32 clamp is decomposed into min/max, but input_tensor and
  min_arg/max_arg can end up with different ranks (e.g. (1,) vs ())
  after subsequent passes. This violates the TOSA requirements for
  TOSA.MINIMUM and TOSA.MAXIMUM in the node visitors.
- Fix this by matching the input ranks in MatchArgRanksPass

Change-Id: I47a3bb73832ddc35553cd3e5b65085155eb9ca48
Signed-off-by: Yufeng Shi &lt;yufeng.shi@arm.com&gt;
Co-authored-by: Ryan O'Shea &lt;ryan.oshea3@arm.com&gt;
Co-authored-by: Rob Elliott &lt;Robert.Elliott@arm.com&gt;
diff --git a/backends/arm/_passes/match_arg_ranks_pass.py b/backends/arm/_passes/match_arg_ranks_pass.py
@@ -63,6 +63,8 @@ def __init__(self, exported_program: ExportedProgram) -> None:
         exir_ops.edge.aten.bitwise_and.Tensor,
         exir_ops.edge.aten.bitwise_xor.Tensor,
         exir_ops.edge.aten.bitwise_or.Tensor,
+        exir_ops.edge.aten.maximum.default,
+        exir_ops.edge.aten.minimum.default,
     ]
 
     def _match_op_rank(self, graph_module, node, arg, max_rank):

Original file line number	Diff line number	Diff line change
`@@ -63,6 +63,8 @@ def __init__(self, exported_program: ExportedProgram) -> None:`
`63`	`63`	`exir_ops.edge.aten.bitwise_and.Tensor,`
`64`	`64`	`exir_ops.edge.aten.bitwise_xor.Tensor,`
`65`	`65`	`exir_ops.edge.aten.bitwise_or.Tensor,`
	`66`	`+ exir_ops.edge.aten.maximum.default,`
	`67`	`+ exir_ops.edge.aten.minimum.default,`
`66`	`68`	`]`
`67`	`69`
`68`	`70`	`def _match_op_rank(self, graph_module, node, arg, max_rank):`