NCNN different output for PC and Android?

[furiouslee]

I loaded my parameters and model weights in the form of unsigned character array made using the ncnn2mem. However, they provide different outputs in PC and Android.

Previously I couldn't load the unsigned character array because it was not aligned properly.
Can this be a problem in declaring ncnn::Net or ncnn::Extractor or ncnn::Mat classes?
Should I be adding __attribute__((aligned(4))) everywhere?

[nihui]

please try the latest code.
feel free to reopen this if the output is still different.