[DRAFT][luci/service] Support dynamic shape inference for reshape

[jongwonyang]

This commit supports dynamic shape inference for reshape operation

Dynamic shape inference algorithm

from: #13927 (comment), #13927 (comment)

• CircleReshape always has two inputs: tensor and shape.
  • otherwise, throw an error
• The shape input can be CircleConst, CircleOutputDummy, or CircleNode.
• When the shape input is CircleConst
  • The shape is inferred from the constant shape
  • output_shape would be static
  • e.g. reshape(tensor, [2, 3]) --> [2, 3]
• When the shape input is CircleOutputDummy
  • First, try to inference shape using node's attribute (newShape)
  • If there's no attribute (and no shape input), propagate the shape of node itself
• When the shape input is CircleNode
  • The shape is inferred by shape input's size (because the exact shape cannot be determined at compile time)
  • output_shape would be dynamic
  • e.g. reshape(tensor, shape_node(with size 3)) --> [?, ?, ?]
• If there is single unknown dimension in output_shape and input tensor is fully known, that dimension is inferred by the count of the elements

When should not infer?

Unknown dimension of output_shape is NOT inferred (by the should_infer) when:

• the shape input is CircleNode. Because if it's not constant, the only thing we can infer at compile time is the size of dimensions.
• the tensor input is not fully known. Because shape inference of reshape is based on the sameness of elements count.

Remaining own_shape?

from: #13935 (comment), #13935 (comment)

The usage of circle_shape might seems to be remaining of own_shape which should be removed by the new shape inference algorithm.

First, this logic is used to support this recipe(Reshape_003) which has no attribute, no shape input.
I've tried to revise this recipe according to this comment, and have read the related PRs (#1554, #1519).
But it was hard to make a policy about no attribute, no shape input case.
So, we decided not to fix another code(importer, recipes) and do it best in shape inference logic.
As a result, I followed the suggestion from #13927 (comment).

Neg test cases

from: #13977

Some of the neg test cases might not directly related to the changes that have been made in this PR.
(which means the false return and throws are not made in CircleReshape.cpp.)
Please tell me if these are inappropriate. I'll remove them and open an separate issue.

---

Related: #13927
Draft: #13935
Depends on: #13989

ONE-DCO-1.0-Signed-off-by: Jongwon Yang [email protected]

[zetwhite]

Why is this test modified?

If you'd like to add shape = circle_const(-1, 4) case, Why didn't you add just new test?

[jongwonyang]

Why is this test modified?

If you'd like to add shape = circle_const(-1, 4) case, Why didn't you add just new test?

I tried to make all the cases to test "dynamic" shape inference so that it can be related to this PR's changes.

But the existed cases were supposed to test migration of the algorithm, so I think we can just leave it :)

I'll make the change to preserve existing cases.

[zetwhite]

Q: Is there an easy way to apply these changes one by one?

By reading diff page (https://github.com/Samsung/ONE/pull/13999/files?diff=unified#diff-53743de477b450a036961928b169bb3cdbdfa63c96717ac0f9d88a5b3f696338R45-R52), It is hard to distinguish how each logics are updated. So, I'm afraid that it is hard to get reviews.

[zetwhite]

IDK this is possible or not, but I'm trying to think a way to split this PR into small several PRs :

e.g.

• support CircleConst, not CircleConst case first
• apply should_infer
• support CircleDummy next

[jongwonyang]

Sounds great! I'll work on it :)

[llFreetimell]

Below if-else statements can be little bit more simplified with following codes :)

if (auto const_shape = dynamic_cast<luci::CircleConst *>(node->shape()))
{
  // inference with const_shape
}
else if (auto dummy_shape = dynamic_cast<luci::CircleOutputDummy *>(node->shape()))
{
  // inference with attribute
}
else
{
  auto node_shape = loco::must_cast<luci::CircleNode *>(node->shape());
  // propagate dynamic shape
}