Add additional default decompositions for upsample operators (#7402)

Summary:

There are several core ATen ops that are not yet supported on ExecuTorch, including upsample_bilinear2d.vec and upsample_nearest2d.vec. These ops are currently not decomposed by default with PyTorch export default decompositions, but should be. Existing ET consumers rely on this behavior, so we need to preserve it until we have upsample kernels ready.

This change allows ET to opt-into decomposing these ops, regardless of the PyTorch default export decomposition table. This will unblock updating PyTorch with the correct behavior (see pytorch/pytorch#116684).

Once the upsample kernels land in ET, we can remove these decompositions. This is currently blocked by pin bumps, which may take a while to resolve.

Differential Revision: D67443180

Author: GregoryComer

exir/program/test/test_program.py

@@ -8,6 +8,7 @@
 
 import copy
 import unittest
+from collections.abc import Iterable
 from typing import Any, Dict
 
 import torch
@@ -21,6 +22,9 @@
 from executorch.exir.lowered_backend_module import get_lowered_submodules
 from executorch.exir.pass_base import ExportPass
 from executorch.exir.passes import MemoryPlanningPass
+from executorch.exir.passes.replace_aten_with_edge_pass import (
+    aten_to_edge,
+)
 from executorch.exir.program._program import (
     EdgeProgramManager,
     ExecutorchProgramManager,
@@ -41,6 +45,15 @@
 from torch.nn import functional as F
 
 
+def count_nodes(graph_module, target):
+    targets = target if isinstance(target, Iterable) else [target]
+
+    count = 0
+    for node in graph_module.graph.nodes:
+        if node.op == "call_function" and node.target in targets:
+            count += 1
+    return count
+
 class TestLinear(torch.nn.Module):
     def __init__(self):
         super().__init__()
@@ -662,13 +675,6 @@ def _get_random_inputs(cls):
             partitioner=[NonDecompTestPartitioner()],
         )
 
-        def count_nodes(graph_module, target):
-            count = 0
-            for node in graph_module.graph.nodes:
-                if node.op == "call_function" and node.target == target:
-                    count += 1
-            return count
-
         # There should be 1 call_delegate node and 1 node for aten.mm.default for the
         # linear that doesn't have a bias which was decomposed as the partitioner
         # said this node wasn't supported.
@@ -723,13 +729,6 @@ def _test_to_edge_with_preserved_ops(
     ):
         edge = to_edge_with_preserved_ops(program, preserve_ops=preserved_ops)
 
-        def count_nodes(graph_module, target):
-            count = 0
-            for node in graph_module.graph.nodes:
-                if node.op == "call_function" and node.target in target:
-                    count += 1
-            return count
-
         aten_ops_non_decomposed = count_nodes(
             program.graph_module,
             preserved_ops,
@@ -811,3 +810,31 @@ def test_save_fails(self):
         et = edge.to_executorch()
         with self.assertRaises(ValueError):
             _ = et.save("/tmp/test_save.pt")
+
+    def test_additional_decomposed_ops(self):
+        """
+        Validate that EXECUTORCH_ADDITIONAL_DECOMPOSED_OPS are decomposed.
+        """
+        class TestModel(torch.nn.Module):
+            def forward(self, x):
+                y = torch.nn.functional.interpolate(x, scale_factor=2, mode="nearest")
+                y = torch.nn.functional.interpolate(y, scale_factor=2, mode="bilinear")
+                return y
+        
+        test_ops = [
+            torch.ops.aten.upsample_bilinear2d.vec, 
+            torch.ops.aten.upsample_nearest2d.vec
+        ]
+        inputs = (torch.randn(1, 1, 4, 4),)
+        program = torch.export.export(TestModel(), inputs)
+
+        for op in test_ops:
+            self.assertEqual(1, count_nodes(program.graph_module, op))
+        
+        edge1 = to_edge(program)
+        edge2 = to_edge_transform_and_lower(program)
+        
+        for edge in [edge1, edge2]:
+            for op in test_ops:
+                edge_op = aten_to_edge(op)
+                self.assertEqual(0, count_nodes(edge.exported_program().graph_module, edge_op))

exir/tracer.py

@@ -64,6 +64,16 @@
 
 torchdynamo_enabled = False
 
+"""
+Additional decompositions to apply by during to_edge or 
+to to_edge_transform_and_lower in addition to the default decompositions from 
+PyTorch export.
+"""
+EXECUTORCH_ADDITIONAL_DECOMPOSITIONS = [
+    torch.ops.aten.upsample_bilinear2d.vec,
+    torch.ops.aten.upsample_nearest2d.vec,
+]
+
 
 def get_stacktrace() -> List[Dict[str, str]]:
     """
@@ -631,8 +641,12 @@ def _default_decomposition_table(
         ]
         # pyre-fixme[7]: Expected `Dict[OpOverload, typing.Callable[..., executorch.e...
         return get_decompositions(decomp_opset)
+
     # pyre-fixme[7]: Expected `Dict[OpOverload, typing.Callable[..., executorch.exir....
-    return default_decompositions()
+    table = default_decompositions()
+    additional_decompositions = get_decompositions(EXECUTORCH_ADDITIONAL_DECOMPOSITIONS)    
+    table.decomp_table.update(additional_decompositions)
+    return table
 
 
 def dynamo_trace(