Replace split_with_sizes_copy with slice_copy

Summary:
This is useful for a few reasons, which are all related to the fact that we have a lot of things in place for slice ops, so the compiler can optimize much better if we replace `split_with_sizes`. That includes:
- swapping with q/dq nodes to do it in lower bitwidth
- becoming slice_nop when possible (through memory planning constraints)
- a few simplification passes on slices
- possibly finding patterns with slice/cat being a nop
- we have an optimized slice op on HiFi already. We will get the split op soon to be fair, we can do some profiling then, but the above points will still stand

Differential Revision: D73312379

Author: mcremon-meta

Diff for: backends/cadence/aot/replace_ops.py

@@ -17,8 +17,9 @@
 # pyre-unsafe
 
 import math
+import operator
 from operator import neg
-from typing import cast, Dict, Iterable, Sequence, Set, Tuple
+from typing import cast, Dict, Iterable, Optional, Sequence, Set, Tuple
 
 import torch
 import torch.fx
@@ -2182,6 +2183,79 @@ def call_operator(
         )
 
 
+# Adapted from fbcode/pyspeech/opt_passes/replace_ops.py
+@register_cadence_pass(CadencePassAttribute(opt_level=2))
+class ReplaceSplitWithSlicePass(ExportPass):
+    """
+    split_with_sizes() delegates to slice() op, so perform this replacement here.
+    This avoids the expense of delegation from ATen.
+    """
+
+    # For split_with_sizes, return the slice dim and extent for each split.
+    def get_split_sizes(
+        self, graph_module: torch.fx.GraphModule, node: torch.fx.Node
+    ) -> Optional[list[tuple[int, ...]]]:
+        # Parse the args of the split_with_sizes op
+        tensor_arg, split_sizes = node.args[0:2]
+        assert isinstance(tensor_arg, torch.fx.Node)
+        in_shape = get_shape(graph_module, tensor_arg)
+        split_dim = 0 if len(node.args) < 3 else node.args[2]
+        if in_shape is None:
+            return None
+
+        # Canonicalize the split dimension
+        assert isinstance(split_dim, int)
+        split_dim = split_dim if split_dim >= 0 else len(in_shape) + split_dim
+
+        # Create the slice op args corresponding to each split
+        slice_ops = []
+        split_start = 0
+        assert isinstance(split_sizes, list)
+        for split_size in split_sizes:
+            split_end = split_start + split_size
+            slice_args = (split_dim, split_start, split_end)
+            slice_ops.append(slice_args)
+            split_start = split_end
+
+        return slice_ops
+
+    def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
+        graph = graph_module.graph
+        for node in graph.nodes:
+            if not isinstance(node.target, EdgeOpOverload):
+                continue
+            if get_edge_overload_packet(node.target) != exir_ops.edge.aten.split_with_sizes_copy:
+                continue
+            # All the users of this split_with_sizes op must be getitem ops
+            if any(user.target != operator.getitem for user in node.users):
+                continue
+
+            # Get the slice dim and extent for each split
+            slice_ops = self.get_split_sizes(graph_module, node)
+            if slice_ops is None:
+                continue
+
+            # Go over each getitem user, and replace it with slice op
+            for user in list(node.users.keys()):
+                assert user.target == operator.getitem
+                item_idx = user.args[1]
+                assert item_idx < len(slice_ops)
+                cur_slice = slice_ops[item_idx]
+                with graph.inserting_before(user):
+                    cur_slice_node = graph.call_function(
+                        exir_ops.edge.aten.slice_copy.Tensor,
+                        (node.args[0], cur_slice[0], cur_slice[1], cur_slice[2], 1),
+                    )
+                user.replace_all_uses_with(cur_slice_node)
+                graph.erase_node(user)
+
+            graph.erase_node(node)
+
+        graph_module.recompile()
+        result = super().call(graph_module)
+        return result
+
+
 # This class encapsulates all the functions that replace/switch one op in the
 # graph with another.
 class CadenceReplaceOpsInGraph:
@@ -2220,5 +2294,6 @@ class CadenceReplaceOpsInGraph:
         ReplaceSingleElementTensorArgumentsFromFullOpWithScalarPass,
         ReplaceAtenAvgPoolWithJarvisAvgPoolPass,
         ReplaceWhereWithFullArgsWithWhereScalar,
-        # ReplaceGeluWithApproximateGeluPass,
+        ReplaceGeluWithApproximateGeluPass,
+        ReplaceSplitWithSlicePass,
     ]