[Backend Tester] Add portable test flow

backends/qualcomm/tests/tester.py

@@ -52,7 +52,9 @@ def __init__(
             default_partitioner_cls=QnnPartitioner,
         )
 
-    def run(self, artifact: ExportedProgram, inputs=None) -> None:
+    def run(
+        self, artifact: ExportedProgram, inputs=None, generate_etrecord: bool = False
+    ) -> None:
         ep = QnnPassManager().transform_for_export_pipeline(artifact)
         transform_passes = QnnPassManager().get_to_edge_transform_passes(ep)
 
@@ -61,6 +63,7 @@ def run(self, artifact: ExportedProgram, inputs=None) -> None:
             transform_passes=transform_passes,
             partitioner=self.partitioners,
             compile_config=self.edge_compile_conf,
+            generate_etrecord=generate_etrecord,
         )
 
 

backends/test/harness/error_statistics.py

@@ -0,0 +1,99 @@
+from dataclasses import dataclass
+
+import torch
+from torch.ao.ns.fx.utils import compute_sqnr
+
+
+@dataclass
+class TensorStatistics:
+    """Contains summary statistics for a tensor."""
+
+    shape: torch.Size
+    """ The shape of the tensor. """
+
+    numel: int
+    """ The number of elements in the tensor. """
+
+    median: float
+    """ The median of the tensor. """
+
+    mean: float
+    """ The mean of the tensor. """
+
+    max: torch.types.Number
+    """ The maximum element of the tensor. """
+
+    min: torch.types.Number
+    """ The minimum element of the tensor. """
+
+    @classmethod
+    def from_tensor(cls, tensor: torch.Tensor) -> "TensorStatistics":
+        """Creates a TensorStatistics object from a tensor."""
+        flattened = torch.flatten(tensor)
+        return cls(
+            shape=tensor.shape,
+            numel=tensor.numel(),
+            median=torch.quantile(flattened, q=0.5).item(),
+            mean=flattened.mean().item(),
+            max=flattened.max().item(),
+            min=flattened.min().item(),
+        )
+
+
+@dataclass
+class ErrorStatistics:
+    """Contains statistics derived from the difference of two tensors."""
+
+    reference_stats: TensorStatistics
+    """ Statistics for the reference tensor. """
+
+    actual_stats: TensorStatistics
+    """ Statistics for the actual tensor. """
+
+    error_l2_norm: float | None
+    """ The L2 norm of the error between the actual and reference tensor. """
+
+    error_mae: float | None
+    """ The mean absolute error between the actual and reference tensor. """
+
+    error_max: float | None
+    """ The maximum absolute elementwise error between the actual and reference tensor. """
+
+    error_msd: float | None
+    """ The mean signed deviation between the actual and reference tensor. """
+
+    sqnr: float | None
+    """ The signal-to-quantization-noise ratio between the actual and reference tensor. """
+
+    @classmethod
+    def from_tensors(
+        cls, actual: torch.Tensor, reference: torch.Tensor
+    ) -> "ErrorStatistics":
+        """Creates an ErrorStatistics object from two tensors."""
+        actual = actual.to(torch.float64)
+        reference = reference.to(torch.float64)
+
+        if actual.shape != reference.shape:
+            return cls(
+                reference_stats=TensorStatistics.from_tensor(reference),
+                actual_stats=TensorStatistics.from_tensor(actual),
+                error_l2_norm=None,
+                error_mae=None,
+                error_max=None,
+                error_msd=None,
+                sqnr=None,
+            )
+
+        error = actual - reference
+        flat_error = torch.flatten(error)
+
+        return cls(
+            reference_stats=TensorStatistics.from_tensor(reference),
+            actual_stats=TensorStatistics.from_tensor(actual),
+            error_l2_norm=torch.linalg.norm(flat_error).item(),
+            error_mae=torch.mean(torch.abs(flat_error)).item(),
+            error_max=torch.max(torch.abs(flat_error)).item(),
+            error_msd=torch.mean(flat_error).item(),
+            # Torch sqnr implementation requires float32 due to decorator logic
+            sqnr=compute_sqnr(actual.to(torch.float), reference.to(torch.float)).item(),
+        )

backends/test/harness/stages/to_edge_transform_and_lower.py

@@ -7,28 +7,36 @@
     to_edge_transform_and_lower,
 )
 from executorch.exir.backend.partitioner import Partitioner
+
 from torch.export import ExportedProgram
 
 
 class ToEdgeTransformAndLower(Stage):
     def __init__(
         self,
-        default_partitioner_cls: Type,
+        default_partitioner_cls: Type | None = None,
         partitioners: Optional[List[Partitioner]] = None,
         edge_compile_config: Optional[EdgeCompileConfig] = None,
     ):
-        self.partitioners = partitioners or [default_partitioner_cls()]
+        self.partitioners = (
+            partitioners or [default_partitioner_cls()]
+            if default_partitioner_cls is not None
+            else []
+        )
         self.edge_compile_conf = edge_compile_config or EdgeCompileConfig()
         self.edge_dialect_program = None
 
     def stage_type(self) -> StageType:
         return StageType.TO_EDGE_TRANSFORM_AND_LOWER
 
-    def run(self, artifact: ExportedProgram, inputs=None) -> None:
+    def run(
+        self, artifact: ExportedProgram, inputs=None, generate_etrecord: bool = False
+    ) -> None:
         self.edge_dialect_program = to_edge_transform_and_lower(
             artifact,
             compile_config=self.edge_compile_conf,
             partitioner=self.partitioners,
+            generate_etrecord=generate_etrecord,
         )
 
     @property

backends/test/harness/tester.py

@@ -4,6 +4,7 @@
 
 import torch
 
+from executorch.backends.test.harness.error_statistics import ErrorStatistics
 from executorch.backends.test.harness.stages import (
     Export,
     Partition,
@@ -33,12 +34,12 @@ def __init__(
         self,
         module: torch.nn.Module,
         example_inputs: Tuple[torch.Tensor],
-        stage_classes: Dict[StageType, Callable],
+        stage_classes: Dict[StageType, Callable] | None = None,
         dynamic_shapes: Optional[Tuple[Any]] = None,
     ):
         module.eval()
 
-        self.stage_classes = stage_classes
+        self.stage_classes = stage_classes or Tester.default_stage_classes()
         self.original_module = module
         self.example_inputs = example_inputs
         self.dynamic_shapes = dynamic_shapes
@@ -182,10 +183,10 @@ def _post(self, stage):
         assert stage_type in self.stages
         self.stages[stage_type] = stage
 
-    def _run_stage(self, stage_instance, inputs=None):
+    def _run_stage(self, stage_instance, inputs=None, *args, **kwargs):
         assert isinstance(stage_instance, Stage)
         prev_stage_artifact = self._pre(stage_instance)
-        stage_instance.run(prev_stage_artifact, inputs=inputs)
+        stage_instance.run(prev_stage_artifact, inputs=inputs, *args, **kwargs)  # noqa
         self._post(stage_instance)
         return self
 
@@ -212,11 +213,14 @@ def to_edge(self, to_edge_stage: Optional[ToEdge] = None):
         return res
 
     def to_edge_transform_and_lower(
-        self, to_edge_and_transform_stage: Optional[ToEdgeTransformAndLower] = None
+        self,
+        to_edge_and_transform_stage: Optional[ToEdgeTransformAndLower] = None,
+        generate_etrecord: bool = False,
     ):
         return self._run_stage(
             to_edge_and_transform_stage
-            or self._get_default_stage(StageType.TO_EDGE_TRANSFORM_AND_LOWER)
+            or self._get_default_stage(StageType.TO_EDGE_TRANSFORM_AND_LOWER),
+            generate_etrecord=generate_etrecord,
         )
 
     def run_passes(self, run_passes_stage: Optional[RunPasses] = None):
@@ -302,20 +306,15 @@ def run_method_and_compare_outputs(
         atol=1e-03,
         rtol=1e-03,
         qtol=0,
+        statistics_callback: Callable[[ErrorStatistics], None] | None = None,
     ):
         number_of_runs = 1 if inputs is not None else num_runs
         reference_stage = self.stages[StageType.EXPORT]
 
         stage = stage or self.cur
 
-        print(f"Comparing Stage {stage} with Stage {reference_stage}")
-        for run_iteration in range(number_of_runs):
+        for _ in range(number_of_runs):
             inputs_to_run = inputs if inputs else next(self.generate_random_inputs())
-            input_shapes = [
-                generated_input.shape if hasattr(generated_input, "shape") else None
-                for generated_input in inputs_to_run
-            ]
-            print(f"Run {run_iteration} with input shapes: {input_shapes}")
 
             # Reference output (and quantization scale)
             (
@@ -328,13 +327,25 @@ def run_method_and_compare_outputs(
             # Output from running artifact at stage
             stage_output = self.stages[stage].run_artifact(inputs_to_run)
             self._compare_outputs(
-                reference_output, stage_output, quantization_scale, atol, rtol, qtol
+                reference_output,
+                stage_output,
+                quantization_scale,
+                atol,
+                rtol,
+                qtol,
+                statistics_callback,
             )
 
         return self
 
     @staticmethod
-    def _assert_outputs_equal(model_output, ref_output, atol=1e-03, rtol=1e-03):
+    def _assert_outputs_equal(
+        model_output,
+        ref_output,
+        atol=1e-03,
+        rtol=1e-03,
+        statistics_callback: Callable[[ErrorStatistics], None] | None = None,
+    ):
         """
         Helper testing function that asserts that the model output and the reference output
         are equal with some tolerance. Due to numerical differences between eager mode and
@@ -349,6 +360,11 @@ def _assert_outputs_equal(model_output, ref_output, atol=1e-03, rtol=1e-03):
         for i in range(len(model_output)):
             model = model_output[i]
             ref = ref_output[i]
+
+            error_stats = ErrorStatistics.from_tensors(model, ref)
+            if statistics_callback is not None:
+                statistics_callback(error_stats)
+
             assert (
                 ref.shape == model.shape
             ), f"Output {i} shape {model.shape} does not match reference output shape {ref.shape}"
@@ -386,6 +402,7 @@ def _compare_outputs(
         atol=1e-03,
         rtol=1e-03,
         qtol=0,
+        statistics_callback: Callable[[ErrorStatistics], None] | None = None,
     ):
         """
         Compares the original of the original nn module with the output of the generated artifact.
@@ -408,6 +425,7 @@ def _compare_outputs(
             reference_output,
             atol=atol,
             rtol=rtol,
+            statistics_callback=statistics_callback,
         )
 
     @staticmethod

backends/test/harness/tests/test_error_statistics.py

@@ -0,0 +1,65 @@
+import unittest
+
+import torch
+from executorch.backends.test.harness.error_statistics import ErrorStatistics
+
+
+class ErrorStatisticsTests(unittest.TestCase):
+    def test_error_stats_simple(self):
+        tensor1 = torch.tensor([1, 2, 3, 4])
+        tensor2 = torch.tensor([2, 2, 2, 5])
+
+        error_stats = ErrorStatistics.from_tensors(tensor1, tensor2)
+
+        # Check actual tensor statistics
+        self.assertEqual(error_stats.actual_stats.shape, torch.Size([4]))
+        self.assertEqual(error_stats.actual_stats.numel, 4)
+        self.assertEqual(error_stats.actual_stats.median, 2.5)
+        self.assertEqual(error_stats.actual_stats.mean, 2.5)
+        self.assertEqual(error_stats.actual_stats.max, 4)
+        self.assertEqual(error_stats.actual_stats.min, 1)
+
+        # Check reference tensor statistics
+        self.assertEqual(error_stats.reference_stats.shape, torch.Size([4]))
+        self.assertEqual(error_stats.reference_stats.numel, 4)
+        self.assertEqual(error_stats.reference_stats.median, 2.0)
+        self.assertEqual(error_stats.reference_stats.mean, 2.75)
+        self.assertEqual(error_stats.reference_stats.max, 5)
+        self.assertEqual(error_stats.reference_stats.min, 2)
+
+        # Check error statistics
+        self.assertAlmostEqual(error_stats.error_l2_norm, 1.732, places=3)
+        self.assertEqual(error_stats.error_mae, 0.75)
+        self.assertEqual(error_stats.error_max, 1.0)
+        self.assertEqual(error_stats.error_msd, -0.25)
+        self.assertAlmostEqual(error_stats.sqnr, 10.0, places=3)
+
+    def test_error_stats_different_shapes(self):
+        # Create tensors with different shapes
+        tensor1 = torch.tensor([1, 2, 3, 4])
+        tensor2 = torch.tensor([[2, 3], [4, 5]])
+
+        error_stats = ErrorStatistics.from_tensors(tensor1, tensor2)
+
+        # Check actual tensor statistics
+        self.assertEqual(error_stats.actual_stats.shape, torch.Size([4]))
+        self.assertEqual(error_stats.actual_stats.numel, 4)
+        self.assertEqual(error_stats.actual_stats.median, 2.5)
+        self.assertEqual(error_stats.actual_stats.mean, 2.5)
+        self.assertEqual(error_stats.actual_stats.max, 4)
+        self.assertEqual(error_stats.actual_stats.min, 1)
+
+        # Check reference tensor statistics
+        self.assertEqual(error_stats.reference_stats.shape, torch.Size([2, 2]))
+        self.assertEqual(error_stats.reference_stats.numel, 4)
+        self.assertEqual(error_stats.reference_stats.median, 3.5)
+        self.assertEqual(error_stats.reference_stats.mean, 3.5)
+        self.assertEqual(error_stats.reference_stats.max, 5)
+        self.assertEqual(error_stats.reference_stats.min, 2)
+
+        # Check that all error values are None when shapes differ
+        self.assertIsNone(error_stats.error_l2_norm)
+        self.assertIsNone(error_stats.error_mae)
+        self.assertIsNone(error_stats.error_max)
+        self.assertIsNone(error_stats.error_msd)
+        self.assertIsNone(error_stats.sqnr)

backends/test/suite/flow.py

@@ -1,6 +1,6 @@
 import logging
 
-from dataclasses import dataclass, field
+from dataclasses import dataclass
 from typing import Callable
 
 from executorch.backends.test.harness import Tester
@@ -26,16 +26,25 @@ class TestFlow:
     tester_factory: Callable[..., Tester]
     """ A factory function that returns a Tester instance for this lowering flow. """
 
-    quantize: bool = field(default=False)
+    quantize: bool = False
     """ Whether to tester should run the quantize stage on the model. """
 
     quantize_stage_factory: Callable[..., Quantize] | None = None
     """ A factory function which instantiates a Quantize stage. Can be None to use the tester's default. """
 
+    is_delegated: bool = True
+    """ Indicates whether the flow is expected to generate CALL_DELEGATE nodes. """
+
 
 def all_flows() -> dict[str, TestFlow]:
     flows = []
 
+    from executorch.backends.test.suite.flows.portable import PORTABLE_TEST_FLOW
+
+    flows += [
+        PORTABLE_TEST_FLOW,
+    ]
+
     try:
         from executorch.backends.test.suite.flows.xnnpack import (
             XNNPACK_STATIC_INT8_PER_CHANNEL_TEST_FLOW,