[topi] Add arm_cpu specific pooling schedules

[lhutton1]

This commit:

• Adds specialized arm_cpu pooling schedules for both fixed width and salable vectors.
• Enables topi testing of new arm_cpu schedules.

cc @neildhickey @ekalda @FranklandJack

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• No users to tag found in teams: topi _{See #10317 for details}

_{Generated by tvm-bot}

[lhutton1]

This is the same commit as in #14855 (credit to @FranklandJack). It was reverted by #15286, although, after efforts to reproduce the failure mentioned in the PR, I've not been successful. In addition, the patch is rebased and passing CI. Happy to try to correct the import issue if there is a reproducer (cc @junrushao)

[leandron]

LGTM, thanks @lhutton1!

Leaving this open so that others have opportunity to comment.

[kparzysz-quic]

We're still seeing a circular import:

Traceback (most recent call last):
  File "/.../lib/python3.8/site-packages/_pytest/config/__init__.py", line 774, in import_plugin
    __import__(importspec)
  File "/local/path/python/tvm/testing/__init__.py", line 21, in <module>
    from . import auto_scheduler, autotvm
  File "/local/path/python/tvm/testing/auto_scheduler.py", line 20, in <module>
    from tvm import auto_scheduler, te, topi
  File "/local/path/python/tvm/topi/__init__.py", line 51, in <module>
    from . import nn
  File "/local/path/python/tvm/topi/nn/__init__.py", line 35, in <module>
    from .conv3d_transpose import *
  File "/local/path/python/tvm/topi/nn/conv3d_transpose.py", line 21, in <module>
    from tvm import relay
  File "/local/path/python/tvm/relay/__init__.py", line 29, in <module>
    from . import prelude
  File "/local/path/python/tvm/relay/prelude.py", line 21, in <module>
    from tvm.relay.transform import ToANormalFormExpr
  File "/local/path/python/tvm/relay/transform/__init__.py", line 22, in <module>
    from . import fake_quantization_to_integer, mixed_precision
  File "/local/path/python/tvm/relay/transform/fake_quantization_to_integer.py", line 25, in <module>
    from tvm.relay.qnn.op import canonicalizations
  File "/local/path/python/tvm/relay/qnn/__init__.py", line 20, in <module>
    from . import op
  File "/local/path/python/tvm/relay/qnn/op/__init__.py", line 21, in <module>
    from .qnn import *
  File "/local/path/python/tvm/relay/qnn/op/qnn.py", line 26, in <module>
    from tvm.relay.op.nn.utils import get_pad_tuple2d
  File "/local/path/python/tvm/relay/op/__init__.py", line 35, in <module>
    from . import strategy
  File "/local/path/python/tvm/relay/op/strategy/__init__.py", line 24, in <module>
    from . import arm_cpu
  File "/local/path/python/tvm/relay/op/strategy/arm_cpu.py", line 29, in <module>
    from ....topi.arm_cpu.mprofile import dsp
  File "/local/path/python/tvm/topi/arm_cpu/__init__.py", line 33, in <module>
    from . import qnn_alter_op
  File "/local/path/python/tvm/topi/arm_cpu/qnn_alter_op.py", line 25, in <module>
    from ..nn import qnn_conv2d_alter_layout, add_alter_layout, qnn_requantize_alter_layout
ImportError: cannot import name 'qnn_conv2d_alter_layout' from partially initialized module 'tvm.topi.nn' (most likely due to a circular import) (/local/path/python/tvm/topi/nn/__init__.py)

[lhutton1]

Thanks for reporting @kparzysz-quic, do you have a reproducer so I can take a look?

[kparzysz-quic]

It came up in our automated tests, but I'll try to get you something.

[Hzfengsy]

The same error. Can reproduce it by

pytest tests/python/unittest

PS: my python version is Py 3.9.

[lhutton1]

Thanks @Hzfengsy, I'll take a look