NXP backend: Fix mean.dim delegation and tests.

backends/nxp/backend/ir/converter/node_converters/ops_converters/mean_dim_converter.py

@@ -1,5 +1,4 @@
-# Copyright (c) 2025 NXP
-# All rights reserved.
+# Copyright 2025 NXP
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -20,6 +19,7 @@
     mean_options,
 )
 from executorch.backends.nxp.backend.neutron_target_spec import NeutronTargetSpec
+from executorch.backends.nxp.backend.node_format import NXP_NODE_FORMAT
 from torch.fx import Node
 from torch.nn import Parameter
 
@@ -32,15 +32,33 @@ def _is_supported_on_target(
         parameters_mapping: dict[str, Parameter],
         custom_delegation_options: CustomDelegationOptions,
     ) -> bool:
-        dim = node.args[1]
         keepdim = node.args[2] if len(node.args) >= 3 else False
         rank = len(node.args[0].meta["val"].shape)
-        dim = [d - rank if d > 0 else d for d in dim]
+        dim = [MeanDimConverter._to_pos_dim(d, rank) for d in node.args[1]]
+
+        if rank != 4 or not keepdim:
+            # neutron-converter/src/OperatorC/GlobalAvgPoolPlugin.cpp#74-77
+            return False
 
-        # Only last 2 dimensions (H, W) and keepdim=True with rank=4 are supported on Neutron.
-        if rank != 4 or dim not in [[-1, -2], [-2, -1]] or not keepdim:
+        # The `mean.dim` gets converted to AveragePool by the NeutronConverter, so the channels must be a
+        #  multiple of `num_macs`.
+        # neutron-converter/src/OperatorC/GlobalAvgPoolPlugin.cpp#59-85
+        num_macs = neutron_target_spec.get_num_macs()
+        channels_dim = 1 if node.meta[NXP_NODE_FORMAT].is_channels_first() else -1
+        if (node.meta["val"].shape[channels_dim] % num_macs) != 0:
             return False
 
+        # Neutron only supports reduction over the spatial dimensions H, W.
+        if node.meta[NXP_NODE_FORMAT].is_channels_first():
+            # The input is NCHW. H and W are at indices 2 and 3.
+            if dim not in [[2, 3], [3, 2]]:
+                return False
+        else:
+            # The input is formatless. It can be considered as NHWC, as this is the way Neutron will look at
+            #  the dimensions. So H and W are the middle dimensions.
+            if dim not in [[1, 2], [2, 1]]:
+                return False
+
         return True
 
     @staticmethod

backends/nxp/tests/ir/converter/node_converter/test_mean_dim_converter.py

@@ -1,3 +1,8 @@
+# Copyright 2025 NXP
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
 import numpy as np
 import pytest
 import torch
@@ -8,10 +13,12 @@
 from executorch.backends.nxp.tests.executorch_pipeline import to_quantized_edge_program
 from executorch.backends.nxp.tests.executors import (
     convert_run_compare,
+    graph_contains_any_of_ops,
     ToChannelFirstPreprocess,
     ToChannelLastPreprocess,
 )
 from executorch.backends.nxp.tests.models import MeanDimConvModule, MeanDimLinearModule
+from executorch.exir.dialects._ops import ops as exir_ops
 from torch.export import ExportedProgram
 
 
@@ -21,19 +28,36 @@ def reseed_model_per_test_run():
     np.random.seed(23)
 
 
+class MeanDimModule(torch.nn.Module):
+    def __init__(self, dim, keepdim):
+        super().__init__()
+        self.dim = dim
+        self.keepdim = keepdim
+
+    def forward(self, x):
+        return torch.mean(x, dim=self.dim, keepdim=self.keepdim)
+
+
 @pytest.mark.parametrize(
     "input_shape, dim",
     [
         pytest.param((1, 4, 8, 8), (-1, -2), id="Dim -1, -2."),
+        pytest.param((1, 4, 8, 8), (-2, -1), id="Dim -2, -1."),
+        pytest.param((1, 4, 8, 8), (2, 3), id="Dim 2, 3."),
+        pytest.param((1, 4, 8, 8), (3, 2), id="Dim 3, 2."),
     ],
 )
-def test_mean_dim_conv_quant_conversion(mocker, input_shape, dim, keeepdim=True):
-    model = MeanDimConvModule(dim, keeepdim)
+def test_mean_dim_conv_quant_conversion(mocker, input_shape, dim, keepdim=True):
+    model = MeanDimConvModule(dim, keepdim)
 
     converter_spy = mocker.spy(EdgeProgramToIRConverter, "convert_program")
 
     # Run conversion
-    _ = to_quantized_edge_program(model, input_shape)
+    ep = to_quantized_edge_program(model, input_shape).exported_program()
+
+    # Make sure the `mean.dim` was delegated.
+    assert not graph_contains_any_of_ops(ep.graph, [exir_ops.edge.aten.mean.dim])
+    assert any("lowered_module" in n.name for n in ep.graph.nodes)
 
     # Capture generated model
     tflite_flatbuffers_model, io_formats = converter_spy.spy_return
@@ -61,16 +85,16 @@ def test_mean_dim_conv_quant_conversion(mocker, input_shape, dim, keeepdim=True)
     ],
 )
 @pytest.mark.parametrize(
-    "keeepdim",
+    "keepdim",
     [
         pytest.param(False, id="Don't keep dim."),
         pytest.param(True, id="Keep dim."),
     ],
 )
 def test_mean_dim_linear_unsupported_quant_conversion(
-    mocker, input_shape, dim, keeepdim
+    mocker, input_shape, dim, keepdim
 ):
-    model = MeanDimLinearModule(dim, keeepdim)
+    model = MeanDimLinearModule(dim, keepdim)
 
     converter_spy = mocker.spy(EdgeProgramToIRConverter, "convert_program")
 
@@ -107,14 +131,14 @@ def test_mean_dim_linear_unsupported_quant_conversion(
     ],
 )
 @pytest.mark.parametrize(
-    "keeepdim",
+    "keepdim",
     [
         pytest.param(False, id="Don't keep dim."),
         pytest.param(True, id="Keep dim."),
     ],
 )
-def test_mean_dim_conv_unsupported_quant_conversion(mocker, input_shape, dim, keeepdim):
-    model = MeanDimConvModule(dim, keeepdim)
+def test_mean_dim_conv_unsupported_quant_conversion(mocker, input_shape, dim, keepdim):
+    model = MeanDimConvModule(dim, keepdim)
 
     converter_spy = mocker.spy(EdgeProgramToIRConverter, "convert_program")
 
@@ -140,3 +164,93 @@ def test_mean_dim_conv_unsupported_quant_conversion(mocker, input_shape, dim, ke
         tflite_output_preprocess=ToChannelFirstPreprocess(),
         tfl_model=tflite_flatbuffers_model,
     )
+
+
+@pytest.mark.parametrize(
+    "input_shape, dim",
+    [
+        pytest.param((1, 2, 3, 8), (1, 2), id="Dim 1, 2."),
+        pytest.param((1, 2, 3, 8), (2, 1), id="Dim 2, 1."),
+        pytest.param((1, 2, 3, 8), (-3, -2), id="Dim -3, -2."),
+        pytest.param((1, 2, 3, 8), (-2, -3), id="Dim -2, -3."),
+    ],
+)
+def test_mean_dim__formatless__supported(mocker, input_shape, dim, keepdim=True):
+    model = MeanDimModule(dim, keepdim)
+
+    converter_spy = mocker.spy(EdgeProgramToIRConverter, "convert_program")
+
+    ep = to_quantized_edge_program(model, input_shape).exported_program()
+
+    # Make sure the `mean.dim` was delegated.
+    assert not graph_contains_any_of_ops(ep.graph, [exir_ops.edge.aten.mean.dim])
+    assert any("lowered_module" in n.name for n in ep.graph.nodes)
+
+    # Capture generated model
+    tflite_flatbuffers_model, io_formats = converter_spy.spy_return
+
+    # Capture converted program
+    exported_program: ExportedProgram = converter_spy.call_args.args[1]
+
+    input_data = (np.random.random(input_shape).astype(np.float32) * 50).astype(np.int8)
+
+    convert_run_compare(
+        exported_program,
+        input_data=input_data,
+        tfl_model=tflite_flatbuffers_model,
+        atol=1,
+    )
+
+
+@pytest.mark.parametrize(
+    "input_shape, dim",
+    [
+        pytest.param((1, 2, 3, 8), (2, 3), id="Dim 2, 3."),
+    ],
+)
+def test_mean_dim__formatless__unsupported(input_shape, dim, keepdim=True):
+    model = MeanDimModule(dim, keepdim)
+
+    ep = to_quantized_edge_program(model, input_shape).exported_program()
+
+    # Make sure the `mean.dim` was NOT delegated.
+    assert graph_contains_any_of_ops(ep.graph, [exir_ops.edge.aten.mean.dim])
+    assert not any("lowered_module" in n.name for n in ep.graph.nodes)
+
+
+@pytest.mark.parametrize(
+    "input_shape, dim",
+    [
+        pytest.param(
+            (1, 8, 8, 4), (1, 2), id="Dim 1, 2 (supported), channels = 4 (unsupported)."
+        ),
+    ],
+)
+def test_mean_dim__formatless__unsupported_channels(input_shape, dim, keepdim=True):
+    model = MeanDimModule(dim, keepdim)
+
+    ep = to_quantized_edge_program(model, input_shape).exported_program()
+
+    # Make sure the `mean.dim` was NOT delegated.
+    assert graph_contains_any_of_ops(ep.graph, [exir_ops.edge.aten.mean.dim])
+    assert not any("lowered_module" in n.name for n in ep.graph.nodes)
+
+
+@pytest.mark.parametrize(
+    "input_shape, dim",
+    [
+        pytest.param(
+            (1, 4, 8, 8), (2, 3), id="Dim 2, 3 (supported), channels = 5 (unsupported)."
+        ),
+    ],
+)
+def test_mean_dim__channels_first__unsupported_channels(input_shape, dim, keepdim=True):
+    model = MeanDimConvModule(
+        dim, keepdim, out_channels=5
+    )  # Only multiples of 8 (num_macs) are supported.
+
+    # Run conversion
+    ep = to_quantized_edge_program(model, input_shape).exported_program()
+
+    # Make sure the `mean.dim` was NOT delegated.
+    assert graph_contains_any_of_ops(ep.graph, [exir_ops.edge.aten.mean.dim])

backends/nxp/tests/models.py

@@ -494,9 +494,9 @@ def forward(self, x):
 
 
 class MeanDimConvModule(torch.nn.Module):
-    def __init__(self, dim, keepdim):
+    def __init__(self, dim, keepdim, out_channels=8):
         super().__init__()
-        self.conv = Conv2dModule(stride=1, padding=1)
+        self.conv = Conv2dModule(stride=1, padding=1, out_channels=out_channels)
         self.dim = dim
         self.keepdim = keepdim