This repository includes the open source code for GPU fusion, enabling memory-access optimized kernel fusion for GPU batch jobs.
Jobs have to be defined in a certain format (see BFSJob inside the code for example).
After compilation with CUDA, the code can be run the following way:
ALGO=BFS FUSION=1 ./fusion path-to-graph number
Where path-to-graph is path to the adjacency graph file (converted via Ligra's SNAPtoAdj for example), and number is the number of concurrent jobs to run at the same time.
The composition of jobs can be modified in make_jobs function (which is unrelated to the fusion code, and is just a utility function).
Currently implemented algorithms/jobs are:
- BFS
- SSSP
- PageRank
- LabelPropagation
The code provides comparisons with sequential execution as well as fine-grained profiling information if fusion is used.
The entirety of the code is bundled in one file fusion.cu for ease of transport, but can be separated into multiple files if needed.