#4 cuda 10.2 ptx With more speed.

Might work on compute 8.0 devices as well, But the default thread/block/points configuration is using alot of memory.