GitLab

* preparation

* radial preparation 30%

* radial backward kernel done

* reuse Gmr (exp part) result for gradient

* radial kernel every block run by column major, to avoid atomicAdd waiting

* apply code review

* static_cast

* implicit cast

* format

* angular preparation

* angular backward works, but slow, AtomicAdd should be avoided

* angular opti: use share memory to avoid AtomicAdd

* format

* equation optimization

* remove unnecessary shared mem for atomi

* remove a lot (warpsize * nbr) unnecessary shared mem for atomj

* format

* update

* clean

* fix

* fix

* test file

* fix
Co-authored-by: Gao, Xiang <qasdfgtyuiop@gmail.com>