Add Fx Pass to remove unbind -> getitem chains

RNN lowerings provided by pytorch use unbind to convert hidden layer
outputs into a list, and then index it later. This doesn't fit well
with FxImporter, because it doesn't really have a concept of a list.
This pass takes these chains, and converts them into torch.select,
so that we don't have lists again.

python/shark_turbine/aot/builtins/jittable.py

@@ -45,6 +45,7 @@
 
 from ..passes import (
     functorch_functionalize,
+    remove_alias
 )
 
 from ..support.utils import (
@@ -115,8 +116,8 @@ def resolver(py_value: Any, gni: GraphNodeImporter) -> Optional[Value]:
     return resolver
 
 
-ALL_PASSES: Set[str] = set(["functorch_functionalize"])
-DEFAULT_PASSES: Tuple[str, ...] = ("functorch_functionalize",)
+ALL_PASSES: Set[str] = set(["functorch_functionalize", "remove_alias"])
+DEFAULT_PASSES: Tuple[str, ...] = ("functorch_functionalize", "remove_alias")
 
 
 class jittable(CallableIntrinsic):
@@ -205,6 +206,10 @@ def flat_wrapped_f(*args):
         transformed_f = flat_wrapped_f
         if "functorch_functionalize" in self._passes:
             transformed_f = functorch_functionalize(transformed_f, *flat_pytorch_args)
+        if "remove_alias" in self._passes:
+            transformed_f = remove_alias(transformed_f, *flat_pytorch_args)
+        if "functorch_functionalize" in self._passes:
+            transformed_f = functorch_functionalize(transformed_f, *flat_pytorch_args)
 
         # Ask dynamo to give us an aten graph.
         # TODO: Cache this for repeated calls.

python/shark_turbine/aot/passes/__init__.py

@@ -5,3 +5,4 @@
 # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
 from .functorch import functorch_functionalize
+from .remove_alias import remove_alias

python/shark_turbine/aot/passes/remove_alias.py

@@ -0,0 +1,78 @@
+from typing import Callable
+
+import torch
+from torch.fx import (
+    GraphModule,
+    Node,
+)
+from torch.fx.experimental import proxy_tensor
+from torch.utils import _pytree as pytree
+import operator as py_operator
+
+
+class Test(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        unrolled = torch.unbind(x, 1)
+        return unrolled[0], unrolled[1], unrolled[2], unrolled[3]
+
+
+trace: GraphModule = torch.fx.symbolic_trace(Test())
+
+
+def remove_unbind(gm: GraphModule) -> GraphModule:
+    # Find all unbind nodes
+    unbind_nodes = []
+    for node in gm.graph.nodes:
+        if node.target == torch.ops.aten.unbind.int:
+            unbind_nodes.append(node)
+
+    to_erase = []
+
+    # Replace all unbind -> getitem chains with a index_select node
+    for unbind in unbind_nodes:
+        only_getitem = True
+        for user in unbind.users:
+            if user.op != "call_function":
+                only_getitem = False
+                continue
+            if user.target != py_operator.getitem:
+                only_getitem = False
+                continue
+        if not only_getitem:
+            continue
+
+        unbind_dim = 0
+        if len(unbind.args) == 2:
+            unbind_dim = unbind.args[1]
+
+        for user in unbind.users:
+            # Get the getitem indices
+            index = user.args[1]
+            with gm.graph.inserting_before(user):
+                select = gm.graph.call_function(
+                    torch.select,
+                    (unbind.args[0], unbind_dim, index),
+                )
+                # Replace the getitem node with the index_select node
+                user.replace_all_uses_with(select)
+
+            # Delete the getitem
+            to_erase.append(user)
+
+        to_erase.append(unbind)
+
+    # Erase all the getitem nodes
+    for node in to_erase:
+        gm.graph.erase_node(node)
+    gm.recompile()
+
+    return gm
+
+
+def remove_alias(gm: GraphModule, *args) -> GraphModule:
+    # Replace unbind -> getitem chains with index_select
+    gm = remove_unbind(gm)
+    return gm

python/shark_turbine/dynamo/passes.py

@@ -48,6 +48,7 @@
     torch.ops.aten._log_softmax_backward_data,
     torch.ops.aten.lift_fresh_copy.default,
     torch.ops.aten._unsafe_index.Tensor,
+    torch.ops.aten.sinc,
     # decompositions added manually in this file
     torch.ops.aten._scaled_dot_product_flash_attention.default,
 ]