Remove unused Empty Tensors from Edge Graph

Differential Revision: D68589336

Pull Request resolved: #7954

Author: mcr229

backends/xnnpack/test/ops/test_cat.py

@@ -187,6 +187,15 @@ def test_qs8_cat_gt_5(self):
                 inputs.append(torch.randn(1, 2, 3))
             self._test_cat(self.Cat(), tuple(inputs), cat_num=num_inputs, quant=True)
 
+    def test_qs8_cat_with_empty_tensor(self):
+        inputs = (
+            torch.randn(0, 2, 3),
+            torch.randn(1, 2, 3),
+            torch.randn(3, 2, 3),
+            torch.randn(0, 2, 3),
+        )
+        self._test_cat(self.Cat(), inputs, cat_num=4, quant=True)
+
     class CatNegativeDim(torch.nn.Module):
         def __init__(self):
             super().__init__()

exir/passes/TARGETS

@@ -17,6 +17,7 @@ python_library(
         ":memory_planning_pass",
         ":normalize_transpose_pass",
         ":prim_ops_py_registry",
+        ":prune_empty_tensor_pass",
         ":quant_fusion_pass",
         ":quantize_io_pass",
         ":remove_noop_pass",
@@ -197,6 +198,18 @@ python_library(
     ],
 )
 
+python_library(
+    name = "prune_empty_tensor_pass",
+    srcs = [
+        "prune_empty_tensors_pass.py",
+    ],
+    deps = [
+        "//caffe2:torch",
+        "//executorch/exir:pass_base",
+        "//executorch/exir/dialects:lib",
+    ],
+)
+
 python_library(
     name = "remove_mixed_type_operators",
     srcs = [

exir/passes/__init__.py

@@ -42,6 +42,7 @@
 from executorch.exir.passes.memory_format_ops_pass import MemoryFormatOpsPass
 from executorch.exir.passes.memory_planning_pass import MemoryPlanningPass
 from executorch.exir.passes.normalize_transpose_pass import NormalizeTransposePass
+from executorch.exir.passes.prune_empty_tensors_pass import PruneEmptyTensorsPass
 from executorch.exir.passes.quant_fusion_pass import QuantFusionPass
 from executorch.exir.passes.remove_noop_pass import RemoveNoopPass, RemoveToCopyPass
 from executorch.exir.passes.replace_aten_with_edge_pass import OpReplacePass
@@ -486,6 +487,7 @@ def dead_code_elimination_pass(graph_module: torch.fx.GraphModule) -> PassResult
         ScalarToTensorPass(),
         SymToTensorPass(),
         RemoveNoopPass(),
+        PruneEmptyTensorsPass(),
         RemoveToCopyPass(),
     ]
 ).passes

exir/passes/prune_empty_tensors_pass.py

@@ -0,0 +1,67 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-strict
+from typing import cast, List
+
+import torch
+from executorch.exir.dialects._ops import ops as exir_ops
+from executorch.exir.pass_base import ExportPass, PassResult
+from torch.fx import GraphModule, Node
+
+# This is a list of ops that are No Ops if used with an empty tensor.
+# Which means that if we remove the empty tensor as input to this op,
+# The result of the operation will stay the same
+
+
+class PruneEmptyTensorsPass(ExportPass):
+    """
+    Removes Any empty tensors from the graph that can safely be removed
+    without affecting the results of the graph. Currently we remove empty
+    tensor operations from the following ops:
+    - aten.cat.default
+    """
+
+    def remove_empty_tensors_from_cat(
+        self, graph_module: GraphModule, cat_node: Node
+    ) -> None:
+        """
+        Removes empty tensors from the graph that are inputs to aten.cat.default
+        """
+        concat_list = cast(List[Node], cat_node.args[0])
+        pruned_concat_list = []
+        for input_arg in concat_list:
+            input_arg_tensor = input_arg.meta["val"]
+            if input_arg_tensor.numel() != 0:
+                pruned_concat_list.append(input_arg)
+
+        cat_node.args = (pruned_concat_list,) + cat_node.args[1:]
+        if len(pruned_concat_list) == 0:
+            # if all the inputs to the cat are empty tensors, then we can replace
+            # this concat node with an aten full like
+            cat_tensor = cat_node.meta["val"]
+            with graph_module.graph.inserting_after(cat_node):
+                full_like = graph_module.graph.create_node(
+                    "call_function",
+                    target=exir_ops.edge.aten.full.default,
+                    args=(tuple(cat_tensor.shape), 0),
+                    kwargs={"dtype": cat_tensor.dtype},
+                )
+                full_like.meta = cat_node.meta
+                cat_node.replace_all_uses_with(full_like)
+
+    def call(self, graph_module: GraphModule) -> PassResult:
+        for node in graph_module.graph.nodes:
+            if node.op != "call_function":
+                continue
+
+            if node.target == torch.ops.aten.cat.default:
+                self.remove_empty_tensors_from_cat(graph_module, node)
+
+        graph_module.graph.eliminate_dead_code()
+        graph_module.graph.lint()
+
+        return PassResult(graph_module, True)

exir/tests/TARGETS

@@ -443,6 +443,20 @@ python_unittest(
     ],
 )
 
+python_unittest(
+    name = "test_prune_empty_tensors",
+    srcs = [
+        "test_prune_empty_tensors_pass.py",
+    ],
+    deps = [
+        "//caffe2:torch",
+        "//executorch/exir:lib",
+        "//executorch/exir:memory",
+        "//executorch/exir/capture:config",
+        "//executorch/exir/passes:lib",
+    ],
+)
+
 python_unittest(
     name = "warnings",
     srcs = [

exir/tests/test_prune_empty_tensors_pass.py

@@ -0,0 +1,77 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import unittest
+
+import torch
+import torch.nn as nn
+from executorch.exir import to_edge
+from executorch.exir.capture._config import ExecutorchBackendConfig
+from executorch.exir.dialects._ops import ops as exir_ops
+from executorch.exir.passes import MemoryPlanningPass
+
+
+class TestCat(nn.Module):
+    def forward(self, x, y, z):
+        empty = torch.empty((0, 6))
+        return torch.cat([empty, x, empty, y, z, empty])
+
+    def get_example_inputs(self):
+        return (torch.rand(5, 6), torch.rand(5, 6), torch.rand(5, 6))
+
+
+class TestPruneEmptyTensors(unittest.TestCase):
+    def test_empty_tensor_removed_from_cat(self) -> None:
+        model = TestCat()
+        model.eval()
+        example_inputs = model.get_example_inputs()
+        ep = torch.export.export(model, example_inputs, strict=True)
+        etpm = to_edge(ep).to_executorch(
+            config=ExecutorchBackendConfig(
+                remove_view_copy=False,
+                memory_planning_pass=MemoryPlanningPass(alloc_graph_input=False),
+            ),
+        )
+
+        for node in etpm.exported_program().graph_module.graph.nodes:
+            if node.target in [
+                exir_ops.edge.aten.cat.default,
+                torch.ops.aten.cat.default,
+            ]:
+                self.assertTrue(len(node.all_input_nodes) == 3)
+                for input_arg in node.all_input_nodes:
+                    tensor_val = input_arg.meta["val"]
+                    self.assertTrue(tensor_val.numel() != 0)
+
+        actual = etpm.exported_program().module()(*example_inputs)
+
+        reference = model(*example_inputs)
+
+        self.assertTrue(torch.allclose(actual, reference))
+
+    def test_cat_removed_all_empty(self) -> None:
+        model = TestCat()
+        model.eval()
+        example_inputs = (torch.empty((0, 6)), torch.empty((0, 6)), torch.empty((0, 6)))
+        ep = torch.export.export(model, example_inputs, strict=True)
+        etpm = to_edge(ep).to_executorch(
+            config=ExecutorchBackendConfig(
+                remove_view_copy=False,
+                memory_planning_pass=MemoryPlanningPass(alloc_graph_input=False),
+            ),
+        )
+
+        for node in etpm.exported_program().graph_module.graph.nodes:
+            self.assertFalse(
+                node.target
+                in [exir_ops.edge.aten.cat.default, torch.ops.aten.cat.default]
+            )
+
+        actual = etpm.exported_program().module()(*example_inputs)
+
+        reference = model(*example_inputs)
+
+        self.assertTrue(torch.allclose(actual, reference))