DeconvWeightTransFusionPass

Description

Transposes the input filter in reverse order in the int8 quantization scenario.

Before: After:

• If filter is not 4-dimensional, the complement_dimension node is inserted after filter.
• When the H and W dimensions in the shape of filter are both not 1, the reshape_in, reverse, and reshape_out nodes are inserted in sequence after Reformat.
• Constant folding is performed on all inserted nodes.

Restrictions

• This fusion takes effect in the int8 quantization scenario.
• The filter input of AscendWeightQuant supports only const and int8 type.