Fix after rebase

Author: irtrukhina

backends/nxp/backend/neutron_map.py

@@ -294,6 +294,20 @@ def get_name_matches(tf_names: list[str], neutron_names: list[str]) -> int:
             #     (i.e., sharing a common prefix separated by "/").
             result = 0
             for tf_name in tf_names:
+                # Determine if the tensor name corresponds to a special operation input
+                # (e.g., permutation input for Transpose or padding input for Pad)
+                special_op = (
+                    # Matches names like "perm0", "perm1", etc. → used by Transpose ops
+                    "permutation"
+                    if re.fullmatch(r"perm(\d+)?", tf_name)
+                    # Matches names like "padding0", "padding1", etc. → used by Pad ops
+                    else (
+                        "padding"
+                        if re.fullmatch(r"padding(s)?(\d+)?", tf_name)
+                        # Default: no special operation
+                        else None
+                    )
+                )
                 for neutron_name in neutron_names:
                     if (
                         neutron_name == tf_name
@@ -302,6 +316,11 @@ def get_name_matches(tf_names: list[str], neutron_names: list[str]) -> int:
                     ):
                         result += 1
                         break
+
+                    # Check if the neutron input is also the special op (Pad or Transpose)
+                    if special_op and special_op in neutron_name:
+                        result += 1
+                        break
             return result
 
         name_matches = get_name_matches(tf_node.inputs, neutron_node.inputs)

backends/nxp/runtime/NeutronBackend.cpp

@@ -26,8 +26,8 @@ namespace neutron {
 #define ALIGN_SIZE(size) \
   ((size + BUFFER_ALIGNMENT - 1) & (~(BUFFER_ALIGNMENT - 1)))
 
-#define KOPC_CALLARGS 6  // The operation for TileIR
-  
+#define KOPC_CALLARGS 6 // The operation for TileIR
+
 // clang-format off
 /* Header schema:
      +----------------------------+-----------------------------+------------------------+

backends/nxp/tests/generic_tests/test_profiling.py

@@ -86,6 +86,7 @@ def test__parallel_pool(self, caplog):
             input_shape,
             dlg_model_verifier=BaseGraphVerifier(1, []),
             output_comparator=NumericalStatsOutputComparator(),
+            use_neutron_for_format_conversion=False,
             use_profiling=True,
         )
         neutron_map = extract_map_from_logs(caplog)
@@ -113,11 +114,12 @@ def test__cifar(self, caplog):
             input_shape,
             dlg_model_verifier=BaseGraphVerifier(1, []),
             output_comparator=NumericalStatsOutputComparator(),
+            use_neutron_for_format_conversion=False,
             use_profiling=True,
         )
         neutron_map = extract_map_from_logs(caplog)
         assert neutron_map == {
-            0: (),  # Pad
+            0: (10,),  # Pad
             1: (10, 11),  # Conv2DStandardV1 (Pad + Conv2d)
             2: (12,),  # MaxPool
             3: (13, 14),  # Conv2DStandardV1 (Pad + Conv2d)
@@ -140,6 +142,7 @@ def test__avg_pool(self, caplog):
             input_shape,
             dlg_model_verifier=BaseGraphVerifier(1, []),
             output_comparator=NumericalStatsOutputComparator(),
+            use_neutron_for_format_conversion=False,
             use_profiling=True,
         )
         neutron_map = extract_map_from_logs(caplog)

backends/nxp/tests/nsys_testing.py

@@ -97,6 +97,7 @@ def _run_delegated_executorch_program(
     use_qat: bool = False,
     train_fn: Callable[[torch.fx.GraphModule], None] | None = None,
     use_profiling: bool = False,
+    use_neutron_for_format_conversion=True,
     operators_not_to_delegate: list[str] = None,
     remove_quant_io_ops: bool = False,
 ) -> tuple[ExportedProgram, str]:
@@ -125,6 +126,7 @@ def wrapper(*args, **kwargs):
             use_qat=use_qat,
             train_fn=train_fn,
             use_profiling=use_profiling,
+            use_neutron_for_format_conversion=use_neutron_for_format_conversion,
             operators_not_to_delegate=operators_not_to_delegate,
             remove_quant_io_ops=remove_quant_io_ops,
         )
@@ -400,6 +402,7 @@ def lower_run_compare(
     use_qat: bool = False,
     train_fn: Callable[[torch.fx.GraphModule], None] | None = None,
     use_profiling: bool = False,
+    use_neutron_for_format_conversion=True,
     operators_not_to_delegate: list[str] = None,
     remove_quant_io_ops: bool = False,
 ):
@@ -419,6 +422,9 @@ def lower_run_compare(
     :param use_qat: If True, applies quantization-aware training before conversion (without the QAT training).
     :param train_fn: Train/finetune function for QAT training. Is used only when `use_qat=True`.
     :param use_profiling: Enable profiling for neutron delegated model.
+    :param use_neutron_for_format_conversion: If True, the EdgeProgramToIRConverter will insert `Transpose` ops to
+                                                ensure that the IO matches the executorch partition, which will be
+                                                delegated to Neutron,
     :param operators_not_to_delegate: list of operators not to delegate.
     :param remove_quant_io_ops: If true, IO q-ops are removed and verification is done on quantized
         version of dataset (quantized INT8 input samples).
@@ -464,6 +470,7 @@ def lower_run_compare(
         use_qat=use_qat,
         train_fn=train_fn,
         use_profiling=use_profiling,
+        use_neutron_for_format_conversion=use_neutron_for_format_conversion,
         operators_not_to_delegate=operators_not_to_delegate,
         remove_quant_io_ops=remove_quant_io_ops,
     )