Asynchronous copy of shared memory

[qwerty00]

Hi, I am reading a blog about asynchronous copy to shared memory (https://developer.nvidia.com/blog/controlling-data-movement-to-boost-performance-on-ampere-architecture/?_gl=1*1plcfms*_gcl_au*MjExOTUxNjc5LjE3MTE1MjA0ODA.)

The picture in the blog confuses me in that why L1 is involved in the data path? Can we bypass the L1 cache and directly transfer data from L2 to shared memory? Moreover, can we bypass L2 cache and directly transfer data from global memory to shared memory?

[Robert_Crovella]

@qwerty00

Moreover, can we bypass L2 cache and directly transfer data from global memory to shared memory?

No.

@qwerty00

why L1 is involved in the data path?

A load from global will typically go through the L1. This has been true since the L1 cache was introduced in CUDA GPUs over 10 years ago.

@qwerty00

Can we bypass the L1 cache and directly transfer data from L2 to shared memory?

You can with ordinary load activity. It doesn’t appear to me that the CUDA C++ async load mechanism gives explicit controls for this. It may be possible with PTX.

[qwerty00]

Thanks!