search

wkcn / MobulaOP

A Simple & Flexible Cross Framework Operators Toolkit
MIT License
164 stars 21 forks source link

The performance of dot and im2col in mobula_op #4

Closed wkcn closed 6 years ago

wkcn commented 6 years ago

In current implementation, function dot and im2col are very slow.

The naive dot has low cache hit.

I will find the bottle necks and fix them.

Test Code

In the version db016f2.

OP MobulaOP MXNet
FC 2.209 1.68
Conv 13.012 7.537
wkcn commented 6 years ago

The reason is that the single implementation couldn't perform well in all contexts. Meanwhile, the built-in Fully Connected Layer and Convolutional Layer have been discarded.