Decompose DepthToSpace operator.

[etiotto]

This PR decomposes the ONNX DepthToSpace operator into reshape+transpose+reshape as documented by https://github.com/onnx/onnx/blob/master/docs/Operators.md#DepthToSpace

[tungld]

@etiotto It seems DepthToSpace needs input shape to be rewritten. In such a case, src/Dialect/ONNX/Rewrite.td would be more suitable because when Decompose.cpp is called, tensors may not be unranked.

Could you please enable end-to-end tests for DepthToSpace in test/backend/test.py also? Thanks.

[etiotto]

@tungld thanks for the code review. I have implemented the canonicalization code in src/Dialect/ONNX/Rewrite.td as suggested and added end-toend tests in test/backend/test.py.
Can you please take another look?

[tungld]

@etiotto make check-onnx-backend-dynamic failed, in case of dynamic tests, B, C, H, W will be -1, you should handle them too.

[AlexandreEichenberger]

Please comment on the source of all the changes in the ONNXOps.td.inc, as it is a auto-generated file.

[AlexandreEichenberger]

Could you comment on the source of all these text changes in the ONNXOPS.td.inc?

[AlexandreEichenberger]

Nit, if you put down new code, would be good to alphabetize this list.

[AlexandreEichenberger]

@etiotto Is the PR still needed? I see it having conflict, failure. Please close or update.

[etiotto]

This approach is no longer needed. The DepthToSpace operator has been implemented via the other PRs associated with work item #927. Closing.

[chentong319]

What if x has known dim (-1)?
It seems to me that we can have two rules for different mode.
For mode==DCR, the rule can be (not precisely) written as following to handle non constant shape with just rule. Doable?

def DecomposeDepthToSpacePattern : Pat<`
((ONNXDepthToSpaceOp:$res $x, $bsAttr, $modeAttr),
  ((ONNXShape:$shape $x),
    (ONNXConcat:$newshape (ONNXClip $shape 0),
                                (ONNXConstant $bsAttr), 
                                (ONNXConstant $bsAttr), 
                                (ONNXDiv (ONNXClip $shape 1),  Constant($bsAttr*$bsAttr)),
                                (ONNXClip $shape 2),
                                (ONNXClip $shape 3)),
ONNXReshape $x $newShape)
[($modeAttr == "DCR")]>

[chentong319]

@etiotto should we reopen this PR to avoid lowering DepthToSpace to krnl?

[etiotto]

Currently we lower DepthToSpace in src/Conversions/ONNXToKernel/Tensor/DepthToSpace.cpp into ONNXReshape -> ONNXTranspose -> ONNXReshape. We did not need to create a new Krnl operation.

There are 2 cases to consider, when all dimensions of the shape for ONNX.Reshape are known at compile time or when they aren't. In the former we generate a onnx.Constant to encode the shape, in the later we have to generate code to compute the new shape.

I think the current approach is reasonable. It allows reuse of the code we already have to lower onnx.Reshape and onnx.Transpose so that, should we have a bug ior need to optimize codegen for those operations, we can leverage the effort. Also, the implementation of SpaceToDepth can leverage the same approach.