Incremental Flattening at PPoPP 2019
18 February 2019
Our paper Incremental Flattening for Nested Data Parallelism, with Troels Henriksen, Frederik Thorøe, and Cosmin Oancea, was accepted for presentation at PPoPP 2019 in Washington, DC, on February 18.
The best way to parallelise a nested computation depends on both the GPU and the input size. Instead of choosing one strategy for every case, our compiler generates several versions that expose different amounts of parallelism. Automatically tuned thresholds select a suitable version at runtime. Integrated into Futhark, the approach delivers substantial performance improvements on financial applications and GPU benchmarks.