Cortex_m backend: Add MV3 and lstm test

backends/cortex_m/test/build_test_runner.sh

@@ -14,9 +14,20 @@ et_root_dir=$(realpath "${script_dir}/../../..")
 build_executorch="${et_root_dir}/backends/arm/scripts/build_executorch.sh"
 ${build_executorch}
 
-# Build executor runner with all portable ops selected and semi hosting
+# Build executor runner with selected aten ops and semi hosting
 build_dir="${et_root_dir}/arm_test"
 build_executor_runner="${et_root_dir}/backends/arm/scripts/build_executor_runner.sh"
 build_root_test_dir="${et_root_dir}/arm_test/arm_semihosting_executor_runner_corstone-300"
 
-${build_executor_runner} --pte=semihosting --target=ethos-u55-128 --output="${build_root_test_dir}"
+select_ops_list="\
+aten::add.out,\
+aten::clamp.out,\
+aten::convolution.out,\
+aten::div.out,\
+aten::mean.out,\
+aten::mul.out,\
+aten::relu.out,\
+aten::view_copy.out,\
+dim_order_ops::_to_dim_order_copy.out"
+
+${build_executor_runner} --pte=semihosting --target=ethos-u55-128 --output="${build_root_test_dir}" --select_ops_list="${select_ops_list}"

backends/cortex_m/test/models/__init__.py

@@ -0,0 +1,4 @@
+# Copyright 2025 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.

backends/cortex_m/test/models/test_mobilenet_v3.py

@@ -0,0 +1,74 @@
+# Copyright 2025 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import pytest
+import torch
+
+from executorch.backends.cortex_m.test.tester import CortexMTester, McuTestCase
+from torchvision import models
+
+
+# TODO: Update as more ops are converted by CMSIS-NN ops.
+ops_before_transforms: dict[str, int] = {
+    "executorch_exir_dialects_edge__ops_aten_add_Tensor": 34,
+    "executorch_exir_dialects_edge__ops_aten_addmm_default": 2,
+    "executorch_exir_dialects_edge__ops_aten_clamp_default": 56,
+    "executorch_exir_dialects_edge__ops_aten_convolution_default": 52,
+    "executorch_exir_dialects_edge__ops_aten_div_Tensor": 28,
+    "executorch_exir_dialects_edge__ops_aten_mean_dim": 10,
+    "executorch_exir_dialects_edge__ops_aten_mul_Tensor": 28,
+    "executorch_exir_dialects_edge__ops_aten_permute_copy_default": 2,
+    "executorch_exir_dialects_edge__ops_aten_relu_default": 14,
+    "executorch_exir_dialects_edge__ops_aten_view_copy_default": 1,
+    "executorch_exir_dialects_edge__ops_dim_order_ops__to_dim_order_copy_default": 56,
+    "executorch_exir_dialects_edge__ops_quantized_decomposed_dequantize_per_tensor_default": 178,
+    "executorch_exir_dialects_edge__ops_quantized_decomposed_quantize_per_tensor_default": 109,
+}
+ops_after_transforms: dict[str, int] = {
+    "executorch_exir_dialects_edge__ops_aten_add_Tensor": 28,  # Not lowered due to broadcasting
+    "executorch_exir_dialects_edge__ops_aten_addmm_default": 0,
+    "executorch_exir_dialects_edge__ops_cortex_m_quantized_add_default": 6,
+    "executorch_exir_dialects_edge__ops_cortex_m_quantized_linear_default": 2,
+    "executorch_exir_dialects_edge__ops_aten_clamp_default": 56,
+    "executorch_exir_dialects_edge__ops_aten_convolution_default": 52,
+    "executorch_exir_dialects_edge__ops_aten_div_Tensor": 28,
+    "executorch_exir_dialects_edge__ops_aten_mean_dim": 10,
+    "executorch_exir_dialects_edge__ops_aten_mul_Tensor": 28,
+    "executorch_exir_dialects_edge__ops_aten_permute_copy_default": 0,
+    "executorch_exir_dialects_edge__ops_aten_relu_default": 14,
+    "executorch_exir_dialects_edge__ops_aten_view_copy_default": 1,
+    "executorch_exir_dialects_edge__ops_dim_order_ops__to_dim_order_copy_default": 56,
+    "executorch_exir_dialects_edge__ops_quantized_decomposed_dequantize_per_tensor_default": 0,
+    "executorch_exir_dialects_edge__ops_quantized_decomposed_quantize_per_tensor_default": 0,
+    "executorch_exir_dialects_edge__ops_cortex_m_dequantize_per_tensor_default": 162,
+    "executorch_exir_dialects_edge__ops_cortex_m_quantize_per_tensor_default": 101,
+}
+
+model = models.mobilenet_v3_small(weights=None)
+example_input = torch.randn(1, 3, 224, 224)
+
+
+test_cases = {
+    "mobilenet_v3_small": McuTestCase(
+        model=models.mobilenet_v3_small(weights=None),
+        example_inputs=(example_input,),
+    ),
+}
+
+
+@pytest.mark.skip("Skip until add + linear fix are upstreamed.")
+def test_dialect_mv3(test_case):
+    tester = CortexMTester(test_case.model, test_case.example_inputs)
+    tester.test_dialect(
+        ops_before_transforms,
+        ops_after_transforms,
+        qtol=1,
+    )
+
+
+@pytest.mark.skip("Skip until add + linear fix are upstreamed.")
+def test_implementation_mv3(test_case):
+    tester = CortexMTester(test_case.model, test_case.example_inputs)
+    tester.test_implementation(qtol=1)

backends/cortex_m/test/ops/test_lstm.py

@@ -0,0 +1,98 @@
+# Copyright 2025 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import pytest
+import torch
+from executorch.backends.cortex_m.test.tester import (
+    CortexMTester,
+    McuTestCase,
+    ramp_tensor,
+)
+
+
+class CortexMLSTM(torch.nn.Module):
+    ops_before_transforms = {
+        "executorch_exir_dialects_edge__ops_aten_full_default": 2,
+        "executorch_exir_dialects_edge__ops_aten_squeeze_copy_dims": 4,
+        "executorch_exir_dialects_edge__ops_aten_unsqueeze_copy_default": 2,
+        "executorch_exir_dialects_edge__ops_aten_view_copy_default": 6,
+        "executorch_exir_dialects_edge__ops_aten_permute_copy_default": 3,
+        "executorch_exir_dialects_edge__ops_aten_addmm_default": 3,
+        "executorch_exir_dialects_edge__ops_aten_slice_copy_Tensor": 2,
+        "executorch_exir_dialects_edge__ops_aten_add_Tensor": 4,
+        "executorch_exir_dialects_edge__ops_aten_split_with_sizes_copy_default": 2,
+        "executorch_exir_dialects_edge__ops_aten_sigmoid_default": 6,
+        "executorch_exir_dialects_edge__ops_aten_tanh_default": 4,
+        "executorch_exir_dialects_edge__ops_aten_mul_Tensor": 6,
+        "executorch_exir_dialects_edge__ops_aten_cat_default": 1,
+    }
+
+    ops_after_transforms = {}
+
+    def __init__(self, input_size: int = 4, hidden_size: int = 3) -> None:
+        super().__init__()
+        self.lstm = torch.nn.LSTM(input_size=input_size, hidden_size=hidden_size)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        y, _ = self.lstm(x)
+        return y
+
+
+class CortexMQuantizableLSTM(torch.nn.Module):
+    ops_before_transforms = {
+        "executorch_exir_dialects_edge__ops_aten_add_Tensor": 4,
+        "executorch_exir_dialects_edge__ops_aten_addmm_default": 4,
+        "executorch_exir_dialects_edge__ops_aten_cat_default": 1,
+        "executorch_exir_dialects_edge__ops_aten_full_default": 1,
+        "executorch_exir_dialects_edge__ops_aten_mul_Tensor": 6,
+        "executorch_exir_dialects_edge__ops_aten_permute_copy_default": 4,
+        "executorch_exir_dialects_edge__ops_aten_select_copy_int": 2,
+        "executorch_exir_dialects_edge__ops_aten_sigmoid_default": 6,
+        "executorch_exir_dialects_edge__ops_aten_split_with_sizes_copy_default": 2,
+        "executorch_exir_dialects_edge__ops_aten_squeeze_copy_dims": 1,
+        "executorch_exir_dialects_edge__ops_aten_tanh_default": 4,
+        "executorch_exir_dialects_edge__ops_aten_view_copy_default": 1,
+        "executorch_exir_dialects_edge__ops_quantized_decomposed_dequantize_per_tensor_default": 34,
+        "executorch_exir_dialects_edge__ops_quantized_decomposed_quantize_per_tensor_default": 27,
+    }
+
+    ops_after_transforms = {}
+
+    def __init__(self, input_size: int = 4, hidden_size: int = 3) -> None:
+        super().__init__()
+        self.lstm = torch.ao.nn.quantizable.LSTM(
+            input_size=input_size, hidden_size=hidden_size
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        y, _ = self.lstm(x)
+        return y
+
+
+test_cases = {
+    "lstm_fp32": McuTestCase(
+        model=CortexMLSTM(),
+        example_inputs=(ramp_tensor(-1, 1, (2, 1, 4)),),
+    ),
+    "lstm_quantizable": McuTestCase(
+        model=CortexMQuantizableLSTM(),
+        example_inputs=(ramp_tensor(-1, 1, (2, 1, 4)),),
+    ),
+}
+
+
+@pytest.mark.skip("Not implemented yet.")
+def test_dialect_lstm(test_case: McuTestCase) -> None:
+    tester = CortexMTester(test_case.model, test_case.example_inputs)
+    tester.test_dialect(
+        test_case.model.ops_before_transforms, test_case.model.ops_after_transforms
+    )
+
+
+@pytest.mark.skip("Not implemented yet.")
+def test_implementation_lstm(test_case: McuTestCase) -> None:
+    tester = CortexMTester(test_case.model, test_case.example_inputs)
+    tester.test_implementation()