Neighbor List CPU

[npiroozan]

I believe I have found the issue for performance on CPUs. It seems to be related to neighbor lists. Traditionally we should migrate from AoS to SoA (when optimizing for GPUs), but DeePMD seems to implement a novel way of neighbor lists:

Migrate from AoS to SoA and then compress each element of the neighbor list into a 64 bit unsigned integer with the equation (α(j) × 10^16 + rijj × 10^8 × 10^6 + j).

In your paper, you specify that this is an efficient way for GPUs. For x86 CPU optimization, how would you recommend improving neighbor list formatting?

Thank you very much.

[npiroozan]

Pardon me, I figured out the answer to this.