Created by: DannyIsFunny
【Issue】 Fc arm kernel will acquire a space to restore transposed weight when internal m_ is positive integer, which resulted in an growth in memory usage. 【Effect of Current PR】 Update the original wight of FC by the transposed weight to avoid allocating a new memory space.
Experiment:
input model: v45 size 1.8M (weights of FC is 1M)
After merging this pull request: 1M memory usage will be reduced