Abstract
This chapter introduces the parallel histogram computation pattern and the concept of atomic operations. It shows that atomic operations to the same location are serialized and that their throughput is determined by their latency. It further introduces four important optimization techniques: thread coarsening-based interleaved data partitioning for improved memory coalescing, caching for reduced latency and improved throughput of atomic operations, privatization for reduced contention, and aggregation for reduced contention.
| Original language | English (US) |
|---|---|
| Title of host publication | Programming Massively Parallel Processors |
| Subtitle of host publication | a Hands-on Approach, Fourth Edition |
| Publisher | Elsevier |
| Pages | 191-210 |
| Number of pages | 20 |
| ISBN (Electronic) | 9780323912310 |
| ISBN (Print) | 9780323984638 |
| DOIs | |
| State | Published - Jan 1 2022 |
Keywords
- Histogram
- atomic operation
- contiguous partitioning
- feature extraction
- interleaved partitioning
- memory bound
- memory coalescing
- memory latency
- memory throughput
- output interference
- race condition
- read-modify-write
ASJC Scopus subject areas
- General Computer Science