Support step attribute in slice from Onnx

[Lewuathe]

Summary:

Support step attribute in slice node when loading ONNX format. Although it succeeds to load ONNX format having slice node with step attribute, it does not support non-1 step values. Since it should be rare to use non-1 step values for major machine learning models, we may want to prioritize loading major models such as YOLO, LSTM.

See: #3987

Test Plan:
Added two test cases passing slice node with the default steps (all having 1s) and slice node with non-1 step attributes.

[facebook-github-bot]

@jfix71 has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[facebook-github-bot]

@jfix71 merged this pull request in f4ec25e.

[posEdgeOfLife]

is there plan to support backward slice which means step = -1? As browsing the code, it feels that step can be supported by adding an additional step attribute to the node, update the extractTensor kernel in the backend.

[Lewuathe]

IIUC, SliceNode is converted into ExtractTensor instruction in IR which only gets the offset parameter. We need to make ExtractTensor support additional parameters (e.g. step) so that the backend properly generates the corresponding code.

Am I correct?

[posEdgeOfLife]

IIUC, SliceNode is converted into ExtractTensor instruction in IR which only gets the offset parameter. We need to make ExtractTensor support additional parameters (e.g. step) so that the backend properly generates the corresponding code.

Am I correct?

Yes.Actually some IR optimization are depending on step=1, so I am not sure if this is wanted.