Privatization: Private Copies for Atomic Updates on a GPU
When many threads add to the same address, the updates queue. Give each block a private copy in shared memory, merge once, and calculate how the longest queue and the merge cost change.
8 min readConcept
