Defined in File inference.h
transform consecutive tensor shuffle operations into one shuffle operator or a Nop
Transform shuffle/typecvt operator chains to one shuffle operator and multiple typecvt operators. For example, a operator chain like reformat(nchw -> nchw4), asQuantizedS8, reformat(nchw4 -> nchw), asFloat32, would be changed to asQuantizedS8, asFloat32. Since the reciprocal reformat operations have been removed from the operator chain, the computation can be speed up with fewer memory operations. This pass is usually used after EnableTensorCorePass, TensorRTReplacePass.