NXP backend: Add test cases for new Neutron C flow

Author: StrycekSimon

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

@@ -6,16 +6,35 @@
 import numpy as np
 import pytest
 import torch
-
 from executorch.backends.nxp.backend.edge_program_converter import (
     EdgeProgramToIRConverter,
 )
-from executorch.backends.nxp.tests.executorch_pipeline import to_quantized_edge_program
+from executorch.backends.nxp.backend.ir.converter.builder.aten_model_builder_director import (
+    AtenModelBuilderDirector,
+)
+from executorch.backends.nxp.backend.ir.lib.tflite.BuiltinOperator import (
+    BuiltinOperator as Ops,
+)
+from executorch.backends.nxp.tests.executorch_pipeline import (
+    ModelInputSpec,
+    to_quantized_edge_program,
+)
 from executorch.backends.nxp.tests.executors import (
     convert_run_compare,
     graph_contains_any_of_ops,
 )
+from executorch.backends.nxp.tests.graph_verifier import DetailedGraphVerifier
+from executorch.backends.nxp.tests.model_output_comparator import (
+    NumericalStatsOutputComparator,
+)
+from executorch.backends.nxp.tests.nsys_testing import lower_run_compare
+from executorch.backends.nxp.tests.ops_aliases import (
+    AddTensor,
+    Clamp,
+    ExecutorchDelegateCall,
+)
 from executorch.exir.dialects._ops import ops as exir_ops
+from executorch.backends.nxp.tests.use_qat import *  # noqa: F403
 
 
 @pytest.fixture(autouse=True)
@@ -24,11 +43,6 @@ def reseed_model_per_test_run():
     np.random.seed(23)
 
 
-# noinspection PyProtectedMember
-ExecutorchDelegateCall = torch.ops.higher_order.executorch_call_delegate
-Clamp = exir_ops.edge.aten.clamp.default
-
-
 class ClampModule(torch.nn.Module):
 
     # noinspection PyShadowingBuiltins
@@ -180,3 +194,113 @@ def test_convert_clamp__no_delegation__unsupported_bounds(min, max):
 
     # Make sure the `clamp` was NOT delegated.
     assert graph_contains_any_of_ops(delegated_ep.graph, [Clamp])
+
+
+class TestClampNewNeutronFlow:
+    @pytest.mark.parametrize(
+        "min, max",
+        [
+            pytest.param(-1, 2, id="min = -1, max = 2 (Max/Min)"),
+            pytest.param(None, 1, id="min = None, max = 1 (Max/Min)"),
+            pytest.param(1, None, id="min = 1, max = None (Max/Min)"),
+            pytest.param(0, 2, id="min = 0, max = 2 (Max/Min)"),
+            pytest.param(0, 1, id="min = 0, max = 1 (Relu0To1)"),
+            pytest.param(-1, 1, id="min = -1, max = 1 (ReluN1To1)"),
+            pytest.param(0, None, id="min = 0, max = None (Relu)"),
+            # # Float bounds
+            pytest.param(-1.0, 2.0, id="min = -1.0, max = 2.0 (Max/Min)"),
+            pytest.param(None, 1.0, id="min = None, max = 1.0 (Max/Min)"),
+            pytest.param(1.0, None, id="min = 1.0, max = None (Max/Min)"),
+            pytest.param(1.0, float("inf"), id="min = 1.0, max = infinity (Max/Min)"),
+            pytest.param(-float("inf"), 1.0, id="min = infinity, max = 1.0 (Max/Min)"),
+            pytest.param(0.1, 0.5, id="min = 0.1, max = 0.5 (Max/Min)"),
+            pytest.param(0.0, 1.0, id="min = 0.0, max = 1.0 (Relu0To1)"),
+            pytest.param(-1.0, 1.0, id="min = -1.0, max = 1.0 (ReluN1To1)"),
+            pytest.param(0.0, None, id="min = 0, max = None (Relu)"),
+        ],
+    )
+    def test_convert_clamp__full_pipeline(self, mocker, min, max, use_qat):
+        input_shape = (2, 7, 2)  # Indivisible by num_macs
+        model = AddClampModule(min, max)
+
+        x_input_spec = ModelInputSpec(input_shape)
+        comparator = NumericalStatsOutputComparator()
+        graph_verifier = DetailedGraphVerifier(
+            mocker,
+            expected_delegated_ops={
+                AddTensor: 1,
+                Clamp: 1,
+            },
+            expected_non_delegated_ops={},
+        )
+
+        lower_run_compare(
+            model=model,
+            input_spec=[x_input_spec],
+            dlg_model_verifier=graph_verifier,
+            output_comparator=comparator,
+            use_new_flow_neutron_c=True,
+            use_qat=use_qat,
+        )
+
+    # noinspection PyShadowingBuiltins
+    @pytest.mark.parametrize(
+        "min, max, expected_tflite_ops",
+        [
+            pytest.param(
+                0.1,
+                0.5,
+                [Ops.ADD, Ops.MAXIMUM, Ops.MINIMUM],
+                id="min = 0.1, max = 0.5 (Max/Min)",
+            ),
+            pytest.param(
+                0.0, 1.0, [Ops.ADD, Ops.RELU_0_TO_1], id="min = 0, max = 1 (Relu0To1)"
+            ),
+            pytest.param(
+                -1.0,
+                1.0,
+                [Ops.ADD, Ops.RELU_N1_TO_1],
+                id="min = -1, max = 1 (ReluN1To1)",
+            ),
+            pytest.param(
+                0.0, None, [Ops.ADD, Ops.RELU], id="min = 0, max = None (Relu)"
+            ),
+        ],
+    )
+    def test_convert_clamp__relu_vs_maxmin(self, mocker, min, max, expected_tflite_ops):
+        input_shape = (23,)
+        model = AddClampModule(min, max)
+
+        converter_spy = mocker.spy(EdgeProgramToIRConverter, "convert_program")
+        tflite_spy = mocker.spy(AtenModelBuilderDirector, "finish")
+
+        delegated_ep = to_quantized_edge_program(
+            model,
+            input_shape,
+            use_new_flow_neutron_c=True,
+        ).exported_program()
+
+        # Make sure the `clamp` was delegated.
+        assert graph_contains_any_of_ops(delegated_ep.graph, [ExecutorchDelegateCall])
+        assert not graph_contains_any_of_ops(delegated_ep.graph, [Clamp])
+
+        intermediate_ep = converter_spy.call_args.args[1]
+        quant_node = list(intermediate_ep.graph.nodes)[-2]
+        dequant_node = list(intermediate_ep.graph.nodes)[-4]
+        tflite_internal_ops = list(
+            op.builtin_code for op in tflite_spy.spy_return.operator_codes.vector
+        )
+
+        assert graph_contains_any_of_ops(intermediate_ep.graph, [Clamp])
+        assert len(tflite_internal_ops) == len(expected_tflite_ops) + 1  # Transpose
+        assert all(op in tflite_internal_ops for op in expected_tflite_ops)
+
+        if len(expected_tflite_ops) == 3:
+            # Min/Max variant should have same input and output quantization
+            assert all(
+                q == dq for q, dq in zip(quant_node.args[1:], dequant_node.args[1:])
+            )
+        else:
+            assert not all(
+                q == dq for q, dq in zip(quant_node.args[1:], dequant_node.args[1:])
+            )

backends/nxp/tests/ops_aliases.py

@@ -12,8 +12,10 @@
 from executorch.exir.dialects._ops import ops as exir_ops
 
 Abs = exir_ops.edge.aten.abs.default
+AddTensor = exir_ops.edge.aten.add.Tensor
 AvgPool2D = exir_ops.edge.aten.avg_pool2d.default
 Bmm = exir_ops.edge.aten.bmm.default
+Clamp = exir_ops.edge.aten.clamp.default
 Convolution = exir_ops.edge.aten.convolution.default
 DequantizePerChannel = exir_ops.edge.quantized_decomposed.dequantize_per_channel.default
 DequantizePerTensor = exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default