[Doc] Add reduce optimize method: remove bank conflict

PaddleJitLab / CUDATutorial

A self-learning tutorail for CUDA High Performance Programing.

Apache License 2.0

86 stars 16 forks source link

Closed AndSonder closed 5 months ago

AndSonder commented 6 months ago

Add reduce optimize method: remove bank conflict

优化手段	运行时间(us)	带宽	加速比
Baseline	3118.4	42.503GB/s	~
交错寻址	1904.4	73.522GB/s	1.64
解决 bank conflict	1475.2	97.536GB/s	2.29

AndSonder commented 6 months ago

@Aurelius84 又更新了俩篇，还麻烦有空的时候帮忙review一下~