How to support fix-point matrix multiplication？

[tsy...@gmail.com]

Hello developers,

Thank you all for your continued support of the software. ^_^. 

The advantage is obvious, even compared with gemmlowp which process 8bit matrix multiplication. Now I have to process fix-point matrix multiplication, but I am not sure how much work I should do to support fix-point matrix multiplication on OpenBLAS for a higher efficiency. 

Should  I change the whole architecture or just the corresponding fix-point instructions? 

Thanks.