I am working on BYOC-OneDNNv1.8, and here is a special case, which is not supported by the current layout transform scheme.
This means, if we do padding in the layout_transform process, when it needs to transform back, we also need to do shrink. However, in the current layout_transform scheme, we can only get the layout (string, like “NCHW”), the concrete dst shape is unavailable. This makes we cannot determine whether there needs a shrink, and how we can get the expected transform_rule.
This issue can be more complicated in shufflenet and efficientnet, as there are plenty of shapes, whose channel numbers are not divisible by 16.
The question is: Is it possible to pass down the
dst_shapeinto thelayout_transform?