Lars op optimiztion with cudaLaunchCooperativeKernel method (#35652) · PaddlePaddle/Paddle@a112ce4 · GitHub

Commit

Lars op optimiztion with cudaLaunchCooperativeKernel method (#35652)

Browse files

* A leap of try for cudaLaunchCooperativeKernel

* fix bugs

* Totally replace the lar cuda kernel

* Fix bugs

* fix code according to comments

* fix codes according to  review comments

* adding some function overload

* relocate the power operation.

Loading branch information

JamesLim-sy authored Sep 27, 2021

1 parent e427a0f commit a112ce4