Implement aten::index | feat(torchlib) (#862)

WIP: need more tests

Gather op:
-
https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather
-
https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/

---------

Co-authored-by: BowenBao <bowbaomicrosoft.com>

ghstack-source-id: 078cf56
Pull Request resolved: #883

Signed-off-by: Justin Chu <[email protected]>

Author: justinchuby

onnxscript/function_libs/torch_lib/ops/core.py

@@ -3020,10 +3020,122 @@ def aten_imag(self: TensorType) -> TensorType:
     raise NotImplementedError()
 
 
-def aten_index(self: TensorType, indices: Optional[Sequence[TensorType]]) -> TensorType:
-    """index.Tensor(Tensor self, Tensor?[] indices) -> Tensor"""
+def _are_consecutive(sorted_list: Sequence[int]) -> bool:
+    """Returns True if a sorted list contains consecutive numbers."""
+    if not sorted_list:
+        return True
 
-    raise NotImplementedError()
+    return sorted_list == list(range(min(sorted_list), max(sorted_list) + 1))
+
+
+def _has_none_in_middle(indices) -> bool:
+    """Returns True if there is a None in the middle of the list."""
+    not_none_indices = [i for i, idx in enumerate(indices) if idx is not None]
+    return not _are_consecutive(not_none_indices)
+
+
+def _shape_of_broadcast_tensors(*args: TensorType) -> INT64:
+    """Returns the broadcasted shape of the given tensors."""
+    broadcasted = op.Max(*args)
+    return op.Shape(broadcasted)
+
+
+@torch_op("aten::index", trace_only=True)
+def aten_index(self: TensorType, indices: Sequence[Optional[INT64]]) -> TensorType:
+    """index.Tensor(Tensor self, Tensor?[] indices) -> Tensor
+
+    NOTE: Understanding `aten::index`
+    For `arg0` with shape `[7, 3, 4, 5, 6]`
+    The indexing operation `arg0[0, :, 1:2, tensor([[4,5]])]` will be translated to
+
+    ```
+    +>  select: i64[3, 4, 5, 6] = torch.ops.aten.select.int(arg0, 0, 0);
+    +>  slice_1: i64[3, 4, 5, 6] = torch.ops.aten.slice.Tensor(select, 0, 0, 9223372036854775807);
+    +>  slice_2: i64[3, 1, 5, 6] = torch.ops.aten.slice.Tensor(slice_1, 1, 1, 2);
+    +>  index: i64[3, 1, 1, 2, 6] = torch.ops.aten.index.Tensor(slice_2, [None, None, arg1]);
+    ```
+
+    Here,
+    - `indices = [None, None, arg1]` is equivalent to `indices = [None, None, arg1, None]`
+    - The operation `arg0[0, :, 1:2, tensor([[4,5]])]` is equivalent to `arg0[0, :, 1:2, tensor([[4,5]]), :]`
+
+    None in `indices` are like fillers for dimensions that cannot be removed in the process.
+    """
+
+    self_rank = len(self.shape)
+    index_ranks = [len(index.shape) for index in indices if index is not None]
+    advanced_indexing_rank = max(index_ranks)
+
+    # reordered_positions is the permutation of the index positions where
+    # positions with None are move to the end of the list
+    # For example, if indices = [None, 1, None, 2], then reordered_positions = [1, 3, 0, 2]
+    reordered_positions = sorted(range(len(indices)), key=lambda i: (indices[i] is None, i))
+    # Fill the list with the remaining indices up to the rank of the tensor self.
+    # For example, if indices = [None, 1, None, 2], and the rank of self is 6,
+    # then reordered_positions = [1, 3, 0, 2, 4, 5]
+    reordered_positions = [
+        *reordered_positions,
+        *range(len(reordered_positions), self_rank),
+    ]
+    # Transpose self according to the reordered positions
+    self = op.Transpose(self, perm=reordered_positions)
+
+    # Broadcast the indices to the same shape then concatenate
+    not_none_indices = [idx for idx in indices if idx is not None]
+    broadcast_shape = _shape_of_broadcast_tensors(*not_none_indices)
+    final_index = op.Concat(
+        *(op.Unsqueeze(op.Expand(idx, broadcast_shape), -1) for idx in not_none_indices),
+        axis=-1,
+    )
+
+    self = op.GatherND(self, final_index, batch_dims=0)
+
+    if _has_none_in_middle(indices):
+        # If there is None in the middle, Advanced Indexing cannot decide where to put
+        # the new dimensions. So it places them in the front, like GatherND does.
+        return self
+
+    # When the indices are consecutive, Advanced Indexing will place the new dimensions
+    # (aka. the broadcasted shape) in the middle, replacing the original [x1, ..., xk] axes.
+    #
+    # Input index axes (three parts):
+    #   [
+    #      x_None_front_1, ... x_None_front_m,
+    #      x1, ..., xk,
+    #      x_None_back_1, ..., x_None_back_m
+    #   ]
+    # GatherND result axes:
+    #   [
+    #      *broadcasted_shape(x1, x2, ..., xk),
+    #      x_None_front_1, ... x_None_front_m,
+    #      x_None_back_1, ..., x_None_back_m
+    #   ]
+    # (Transpose here)
+    # Advanced indexing result axes:
+    #   [
+    #      x_None_front_1, ... x_None_front_m,
+    #      *brocasted_shape(x1, x2, ..., xk),
+    #      x_None_back_1, ..., x_None_back_m
+    #   ]
+    #
+    # Need to transpose the result of GatherND to match this axes ordering.
+    first_not_none_position = reordered_positions[0]  # x_None_front_m + 1
+    starting_position_of_none_in_back = (
+        advanced_indexing_rank + first_not_none_position
+    )  # x_None_back_1
+    result_rank = self_rank - len(not_none_indices) + advanced_indexing_rank
+    perm = [
+        *range(
+            advanced_indexing_rank, starting_position_of_none_in_back
+        ),  # None_front_1...x_None_back_1
+        *range(0, advanced_indexing_rank),  # 0...len(broadcasted_shape)
+        *range(
+            starting_position_of_none_in_back,
+            result_rank,
+        ),  # None_back_1...None_back_m
+    ]
+
+    return op.Transpose(self, perm=perm)
 
 
 def aten_index_add(

onnxscript/tests/function_libs/torch_lib/extra_opinfo.py

@@ -444,6 +444,31 @@ def sample_inputs_col2im(op_info, device, dtype, requires_grad, **kwargs):
         yield opinfo_core.SampleInput(tensor, args=(output_size, kernel_size), kwargs=kwargs)
 
 
+def sample_inputs_index(op_info, device, dtype, requires_grad, **kwargs):
+    del op_info  # Unused
+    del kwargs  # Unused
+    make_arg = functools.partial(
+        torch_testing.make_tensor, dtype=dtype, device=device, requires_grad=requires_grad
+    )
+    s = 5
+    index_1d = common_methods_invocations.index_variable(2, s, device=device)
+    index_2d = common_methods_invocations.index_variable((2, s+1), s, device=device)
+    index_3d = common_methods_invocations.index_variable((2, s+1, s+2), s, device=device)
+    test_args = [
+        ([index_1d],),
+        ([None, index_1d],),
+        ([None, None, None, index_1d],),
+        ([index_1d, None],),
+        ([index_1d, None, None],),
+        ([None, index_1d, None, index_1d],),
+        ([index_1d, None, index_1d, None],),
+        ([None, index_1d, index_1d, None],),
+    ]
+
+    for args in test_args:
+        yield opinfo_core.SampleInput(make_arg((s, s, s, s)), args=args)
+
+
 def sample_inputs_native_dropout(
     op_info, device, dtype, requires_grad, *, valid_input_dim=None, **kwargs
 ):
@@ -616,6 +641,15 @@ def sample_inputs_bernoulli_p_deterministic(op_info, device, dtype, requires_gra
         sample_inputs_func=sample_inputs_convolution,
         supports_out=False,
     ),
+    opinfo_core.OpInfo(
+        "aten.index.Tensor",
+        dtypes=common_dtype.all_types_and_complex_and(
+            torch.bool, torch.float16, torch.bfloat16, torch.chalf
+        ),
+        aten_name="index",
+        op=torch.ops.aten.index.Tensor,
+        sample_inputs_func=sample_inputs_index,
+    ),
     opinfo_core.OpInfo(
         "ops.aten.layer_norm",
         aten_name="layer_norm",

onnxscript/tests/function_libs/torch_lib/ops_test_common.py

@@ -261,7 +261,8 @@ def convert_tensor_to_numpy(input: Any) -> Any:
     if isinstance(input, (tuple, list)):
         if len(input) == 0:
             return np.array((), dtype=np.int64)
-        if isinstance(input[0], torch.Tensor):
+        if any(isinstance(x, torch.Tensor) for x in input):
+            # The list can be Optional[Tensor], e.g. [None, Tensor, None] etc.
             return [convert_tensor_to_numpy(x) for x in input]
         if isinstance(input[0], bool):
             return np.array(input, dtype=np.bool_)
@@ -276,10 +277,7 @@ def convert_tensor_to_numpy(input: Any) -> Any:
 
 
 def convert_kwargs_for_onnx(kwargs: dict[str, Any]) -> dict[str, Any]:
-    """Converts kwargs to be compatible with ONNX Runtime.
-
-    ONNX Runtime doesn't support torch.bool, so we convert them to torch.uint8.
-    """
+    """Converts kwargs to be compatible with ONNX Runtime."""
     new_kwargs = {}
     for key, value in kwargs.items():
         if key == "device":
@@ -473,6 +471,8 @@ def _capture_graph_and_evaluate_torch_script_evaluator(function: Callable, args,
                         input.value = subarg
                         sequence_input.append(input)
                         ort_inputs[input_name] = subarg
+                    else:
+                        sequence_input.append(subarg)
                 onnxscript_args.append(sequence_input)
             else:
                 onnxscript_args.append(arg)
@@ -515,7 +515,7 @@ def _capture_graph_and_evaluate_torch_script_evaluator(function: Callable, args,
         # Make sure the model is valid
         try:
             onnx.checker.check_model(onnx_model, full_check=True)
-        except onnx.checker.ValidationError as e:
+        except (onnx.checker.ValidationError, onnx.shape_inference.InferenceError) as e:
             raise AssertionError(
                 f"ONNX model is invalid. Model:\n{onnxscript.proto2text(onnx_model)}"
             ) from e

onnxscript/tests/function_libs/torch_lib/ops_test_data.py

@@ -607,6 +607,7 @@ def _where_input_wrangler(
     TorchLibOpInfo("gt", core_ops.aten_gt),
     # TorchLibOpInfo("is_same_size", core_ops.aten_is_same_size),  # no test case in OPS_DB
     # TorchLibOpInfo("is_nonzero", core_ops.aten_is_nonzero),  # no test case in OPS_DB
+    TorchLibOpInfo("aten.index.Tensor", core_ops.aten_index),
     TorchLibOpInfo(
         "index_put_bool",
         core_ops.aten_index_put_bool,