failure of ShapeInference pass

[chentong319]

I think that ShapeInference pass should not simply fail the pass manager when shape for a tensor can not be inferred.
ShapeInference may be helped or avoided by other passes. Here are some examples:

1. Constant folding may generate constant input for some operations, and then enables AttributePromotion, and finally helps ShapeInference.
2. ScalerOp will be decomposed into other operations by canonicalization pass, which is invoked after first ShapeInference Pass.

My suggestions are

1. ShapeInference fail only for semantic errors, such as an Attribute has incorrect value, or the shape of tensors does not match. Best we should put such checking into verifier.
2. ShapeInference simply gives up (quietly or with some diagnosis) if some needed info is not available yet.
3. Add a pass toe check whether all tensors have known shape at the right place, not mixed with ShapeInference.

[tungld]

I agree with your suggestions. It would be more flexible.

Only a minor comment that verifying Attributes should be done in another pass outside or before ShapeInference.
I feel that checking and setting attributes are not the work for ShapeInference.

[tungld]

ScalerOp will be decomposed into other operations by canonicalization pass, which is invoked after first ShapeInference Pass.

Could we put rewriting rules for ScalerOp to the Decomposing Pass: https://github.com/onnx/onnx-mlir/blob/master/src/Transform/ONNX/Decompose.td? By that way, we don't need to implement ShapeInference for ScalerOp since the Decomposing Pass is called before ShapeInference.

[chentong319]

Yes, we can try Decompose.td.

[doru1004]

Any updates on this issue? @chentong319 @tungld

[chentong319]

We are taking a progressive way to change shape inference. Some constraints have to be removed when the support for dynamic shape is added.
For the passes, we will move to fix-point iterative approach.

[doru1004]

Awesome, sounds like this is still in progress! Thanks for the update.