Arm: Fix various small issues (#8566)

- Softmax incorrectly set to can_delegate=False in aot_arm_compiler
- Incorrect % with rank instead of dim length. Fix test for this.
- Remove .dump_artifact():s

Signed-off-by: Erik Lundell <[email protected]>

Author: Erik-Lundell

backends/arm/_passes/decompose_select.py

@@ -35,8 +35,9 @@ def call(self, graph_module: torch.fx.GraphModule):
             input_node, dim, index = node.args
 
             rank = len(input_node.meta["val"].size())
+            shape = input_node.meta["val"].shape
             dim = dim % rank if dim < 0 else dim
-            index = index % rank if index < 0 else index
+            index = index % shape[dim] if index < 0 else index
 
             with graph_module.graph.inserting_before(node):
                 slice_node = create_node(

backends/arm/test/ops/test_cat.py

@@ -111,7 +111,6 @@ def _test_cat_ethosu_BI_pipeline(
             .check(["torch.ops.quantized_decomposed"])
             .to_edge()
             .partition()
-            .dump_artifact()
             .check_not(["executorch_exir_dialects_edge__ops_aten_cat_default"])
             .check_count({"torch.ops.higher_order.executorch_call_delegate": 1})
             .to_executorch()

backends/arm/test/ops/test_select.py

@@ -19,7 +19,7 @@
 test_data_suite: list[tuple[test_data_t]] = [
     # (test_data, dim, index)
     ((torch.zeros(5, 3, 20), -1, 0),),
-    ((torch.zeros(5, 3, 20), 0, -1),),
+    ((torch.rand(5, 3, 20), 0, -1),),
     ((torch.zeros(5, 3, 20), 0, 4),),
     ((torch.ones(10, 10, 10), 0, 2),),
     ((torch.rand(5, 3, 20, 2), 0, 2),),
@@ -61,9 +61,7 @@ def _test_select_tosa_MI_pipeline(
             .check([export_target])
             .check_not(["torch.ops.quantized_decomposed"])
             .to_edge()
-            .dump_artifact()
             .partition()
-            .dump_artifact()
             .check_count({"torch.ops.higher_order.executorch_call_delegate": 1})
             .to_executorch()
             .run_method_and_compare_outputs(inputs=test_data)

backends/arm/test/ops/test_to_copy.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Arm Limited and/or its affiliates.
+# Copyright 2024-2025 Arm Limited and/or its affiliates.
 # All rights reserved.
 #
 # This source code is licensed under the BSD-style license found in the
@@ -55,9 +55,7 @@ def _test_to_copy_tosa_MI_pipeline(
                 compile_spec=common.get_tosa_compile_spec("TOSA-0.80+MI"),
             )
             .export()
-            .dump_artifact()
             .to_edge()
-            .dump_artifact()
             .partition()
             .check_count({"torch.ops.higher_order.executorch_call_delegate": 1})
             .to_executorch()

examples/arm/aot_arm_compiler.py

@@ -185,7 +185,7 @@ def forward(self, x):
         return z
 
     example_input = (torch.ones(2, 2),)
-    can_delegate = False
+    can_delegate = True
 
 
 class MultipleOutputsModule(torch.nn.Module):