add pass to remove cat from slice pass

Summary:
only keep the cat before slice iff the slice range overlaps with *both* tensors in cat.
TODO: trace to 1+ level for more cat in a china
TODO: support >2 tensors in a cat

Differential Revision: D70425971

Author: zonglinpeng

backends/cadence/aot/remove_ops.py

@@ -745,6 +745,70 @@ def permute_shape(
         return [shape[p] for p in permute_dims]
 
 
+@register_cadence_pass(CadencePassAttribute(opt_level=3))
+class RemoveCatFromSliceCopyPass(ExportPass):
+    def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
+        graph = graph_module.graph
+        slice_copy_nodes = [
+            node
+            for node in graph.nodes
+            if node.target == exir_ops.edge.aten.slice_copy.Tensor
+        ]
+        for slice_copy_node in slice_copy_nodes:
+            slice_dim, start_idx, end_idx, step = 0, 0, float("inf"), 1
+            input_node, *other_args = slice_copy_node.args
+            if len(other_args) >= 1:
+                slice_dim = other_args[0]
+            if len(other_args) >= 2:
+                start_idx = other_args[1]
+            if len(other_args) >= 3:
+                end_idx = other_args[2]
+            if len(other_args) >= 4:
+                step = other_args[3]
+            if step != 1:
+                continue
+            slice_copy_dtype = slice_copy_node.meta["val"].dtype
+            if input_node.target != exir_ops.edge.aten.cat.default:
+                continue
+            cat_dtype = input_node.meta["val"].dtype
+            if slice_copy_dtype != cat_dtype:
+                continue
+            cat_dim = input_node.args[1:]
+            if len(cat_dim) == 0:
+                cat_dim = 0
+            if cat_dim != slice_dim:
+                continue
+            base_idx = 0
+            cat_input_to_keep = None
+            for cat_input_node in input_node.args[0]:
+                cat_dtype = cat_input_node.meta["val"].dtype
+                if slice_copy_dtype != cat_dtype:
+                    continue
+                cat_dim = cat_input_node.args[1:]
+                if len(cat_dim) == 0:
+                    cat_dim = 0
+                if cat_dim != slice_dim:
+                    continue
+                cat_shape = cat_input_node.meta["val"].shape
+
+                # check if the slice range overlaps with the cat range
+                if (
+                    max(start_idx, base_idx)
+                    < min(end_idx, list(cat_shape)[cat_dim] + base_idx)
+                ) or (start_idx == end_idx == base_idx):
+                    if cat_input_to_keep is not None:
+                        # need more than one cat inputs, keep the cat
+                        cat_input_to_keep = None
+                        break
+                    cat_input_to_keep = cat_input_node
+                base_idx += list(cat_shape)[cat_dim]
+            if cat_input_to_keep is not None:
+                slice_copy_node.replace_input_with(input_node, cat_input_to_keep)
+                graph_module.recompile()
+                graph_module.graph.eliminate_dead_code()
+        return super().call(graph_module)
+
+
 # The following class consolidates functions to remove ops that are redundant
 # in Jarvis. Currently, each function in this class iterates over each node of
 # the graph module once. In future, we could consolidate them into a monolithic
@@ -765,4 +829,5 @@ class CadenceRemoveNops:
         RemoveNopMulOpPass,
         RemoveNopAddOpPass,
         RemoveNopLinalgVectorNormOpPass,
+        RemoveCatFromSliceCopyPass,
     ]

backends/cadence/aot/tests/test_remove_ops_passes.py

@@ -21,6 +21,7 @@
 from executorch.backends.cadence.aot.quantizer.quantizer import CadenceDefaultQuantizer
 from executorch.backends.cadence.aot.remove_ops import (
     RemoveAliasCopyOpPass,
+    RemoveCatFromSliceCopyPass,
     RemoveCloneOpPass,
     RemoveContiguousOpPass,
     RemoveDetachCopyPass,
@@ -709,3 +710,54 @@ def forward(self, x):
         self.assertEqual(
             count_node(graph_module, exir_ops.edge.aten.permute_copy.default), 2
         )
+
+    def test_remove_cat_from_slice_copy_all_removal(self) -> None:
+        class M(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, x, y):
+                x1 = torch.cat((x, y), 0)  # (2, 4)
+                return torch.slice_copy(x1, dim=0, start=0, end=1)
+
+        inputs = tuple(torch.randn(2, 4) for _ in range(2))
+        graph_module = export_to_edge(M(), inputs).exported_program().graph_module
+        p = RemoveCatFromSliceCopyPass()
+        graph_module = cast(PassResult, p(graph_module)).graph_module
+
+        # Ensure both cat nodes were removed
+        self.assertEqual(count_node(graph_module, exir_ops.edge.aten.cat.default), 0)
+
+    def test_remove_cat_from_slice_copy_no_removal(self) -> None:
+        class M(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, x, y):
+                x1 = torch.cat((x, y), 0)  # (2, 4)
+                return torch.slice_copy(x1, dim=0, start=0, end=3)
+
+        inputs = tuple(torch.randn(2, 4) for _ in range(2))
+        graph_module = export_to_edge(M(), inputs).exported_program().graph_module
+        p = RemoveCatFromSliceCopyPass()
+        graph_module = cast(PassResult, p(graph_module)).graph_module
+
+        # Ensure both cat nodes were removed
+        self.assertEqual(count_node(graph_module, exir_ops.edge.aten.cat.default), 1)
+
+    def test_remove_cat_from_slice_copy_zero_range(self) -> None:
+        class M(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, x, y):
+                x1 = torch.cat((x, y), 0)  # (2, 4)
+                return torch.slice_copy(x1, dim=0, start=0, end=0)
+
+        inputs = tuple(torch.randn(2, 4) for _ in range(2))
+        graph_module = export_to_edge(M(), inputs).exported_program().graph_module
+        p = RemoveCatFromSliceCopyPass()
+        graph_module = cast(PassResult, p(graph_module)).graph_module
+
+        # Ensure both cat nodes were removed
+        self.assertEqual(count_node(graph_module, exir_ops.edge.aten.cat.default), 0)