https://hgpu.org/?p=3701
Exploiting SPMD Horizontal Locality to Improve Memory Efficiency