level 0/0
idle
GPU Blocks
1 block per segment
Segments
Each level: segments double, work halves. All blocks run in parallel.
Points (L² metric)
data[] — coalesced reads across threads
indices[] — indirect access
proj[] — temp for partition sort key
output[] — final BFS layout
read
write
output
Kernel Path
Key Diff vs CPU
CPU: DFS recursion
GPU: level-sync BFS

nth_element → argsort (radix)