Source code for aitemplate.compiler.transform.transform_permute_to_reshape

#  Copyright (c) Meta Platforms, Inc. and affiliates.
#
#  Licensed under the Apache License, Version 2.0 (the "License");
#  you may not use this file except in compliance with the License.
#  You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
#  Unless required by applicable law or agreed to in writing, software
#  distributed under the License is distributed on an "AS IS" BASIS,
#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#  See the License for the specific language governing permissions and
#  limitations under the License.
#
"""
Transform permute to reshape wherever applicable.
"""
from typing import List

from aitemplate.compiler.base import IntImm, Operator, Tensor
from aitemplate.compiler.ops import reshape
from aitemplate.compiler.transform import transform_utils
from aitemplate.compiler.transform.toposort import toposort

from aitemplate.utils import graph_utils


def _check_permute_to_reshape(op: Operator) -> bool:
    """Check if applicable to replace permute with reshape.

    Args:
        op (Operator): reshape op

    Returns:
        bool: False if operation is not a permute or a permute with memory
            layout modification otherwise True.
    """
    if not op._attrs["op"].startswith("permute"):
        return False

    inputs = op._attrs["inputs"]

    assert (
        len(inputs) == 1
    ), "Permute operation {} should have 1 input, got {} instead".format(
        op._attrs["op"], len(inputs)
    )

    if "input_accessors" in op._attrs:
        input_shape = op._attrs["input_accessors"][0].original_shapes
    else:
        input_shape = inputs[0].shape()

    if op._attrs["op"] == "permute":
        permutation = list(op._attrs["dims"])
    elif op._attrs["op"] == "permute021":
        n_dims = len(input_shape)
        permutation = list(range(n_dims - 2)) + [n_dims - 1, n_dims - 2]
    elif op._attrs["op"] == "permute102":
        permutation = [1, 0, 2]
    elif op._attrs["op"] == "permute210":
        permutation = [2, 1, 0]
    elif op._attrs["op"] == "permute0213":
        permutation = [0, 2, 1, 3]
    else:
        raise NotImplementedError(
            f"Not implemented for permute operation: {op._attrs['op']}"
        )
    if "input_accessors" in op._attrs:
        # Can't convert permute to reshape if one of the dimensions included
        # in permutation is strided
        ta = op._attrs["input_accessors"][0]
        if ta.is_from_strided_tensor and ta.stride_dim in permutation:
            return False
    # Get non-singular dimension indices
    permutation = [
        dim_idx
        for dim_idx in permutation
        if not isinstance(input_shape[dim_idx], IntImm)
        or input_shape[dim_idx].value() != 1
    ]
    is_reshape = permutation == sorted(permutation)
    return is_reshape


def transform_permute_to_reshape(
    sorted_graph: List[Tensor], workdir: str = None
) -> List[Tensor]:
    """Convert permute to reshape wherever applicable.

    When permute op involves moving one or more dimensions with size
    1 around where the order of non-singular dimensions is preserved,
    it's basically a reshape op, i.e. the underlying memory layout
    does not change.

    If a permute op has a non-empty input tensor accessor, its original shape
    should be used to determine whether it can be converted to reshape.
    In this case the shape of the actual input tensor might not match the rank
    of the permutation (but the original shape does) - see the second
    example below.

    Examples:
        [256x5x1x32] -> [256x5x32x1] (with 0132) is a reshape
        [256x5x32] -> [256x5x1x32] (with 0132) is a reshape
        [256x1x5x1x32] -> [256x5x32x1x1] (with 02431) is a reshape
        [256x5x1x32] -> [256x32x5x1] (with 0312) is not a reshape

    Args:
        sorted_graph (List[Tensor]): input graph
        workdir (str, optional): current workdir for dumping debug info. Defaults to None.

    Returns:
        List[Tensor]: optimized graph
    """
    sorted_ops = graph_utils.get_sorted_ops(sorted_graph)

    has_modified = False
    for op in sorted_ops:
        if not _check_permute_to_reshape(op):
            continue

        has_modified = True

        permute_input = op._attrs["inputs"][0]
        permute_output = op._attrs["outputs"][0]
        output_shape = permute_output.shape()

        transform_utils.remove_dst_op_from_tensor(permute_input, op)

        reshape_op = reshape()
        reshape_output = reshape_op(permute_input, output_shape)

        transform_utils.replace_tensor(permute_output, reshape_output)

        sorted_graph.append(reshape_output)

    if has_modified:
        sorted_graph = toposort(sorted_graph)
        transform_utils.sanitize_sorted_graph(sorted_graph)
    return sorted_graph