Various implementations of BFS in OpenCL. Results can be found in PDF format in the charts/ directory.
Naming conventions for the results are as follows:
'analysis22_xxxx.pdf' is the breakdown of the iterations for the unmodified and direction-optimizing implementations on the graph500-22 graph (where xxxx is the platform).
'appeff_xxxx.pdf' is the Application Efficiency for all implementations and platforms processing a specific graph.
'exec22_xxxx.pdf' is the comparison of the execution time of an implementation on graph500-22 and graph500-23.
'exec_xxxx.pdf' is either: a) the execution times of all implementations on one graph, or b) the execution times of one implementation on all graphs.
'kernel_sizes_xxxx.pdf' shows the influence of differing kernel sizes on the performance of a specific graph.
'load_balance.pdf' shows the unhalted runtime of all cores when running graph500-10 50000 times.
'perfport.pdf' shows the performance portability of all implementations on all graphs.
'platform_xxxx.pdf' shows the 'Programming Model Efficiency' of CUDA vs OpenCL for a specific implementation (and its respective port).
'sizes_xxxx.pdf' shows the influence of differing work-group sizes on the performance of a specific graph.
'zero_xxxx.pdf' shows a breakdown of H2D, kernel and D2H times for the GPU and the CPU for a specific graph.
Source code for each implementation is on its own branch:
'Unmodified' is the master branch;
'Datastructure' is the datastruct branch;
'Direction-optimizing' is the direction-optimized branch;
'Datastructure + Direction-optimizing' is the better_direction branch;
Additionally, the CUDA ports are in the cuda_master and cuda_better_direction branches.
The zero-copy and zero-copy-2 branches are for the zero-copy specialization. The former breaks portability with the GPU, but the latter recovers this.
Finally, the kernel-items branch allows for differing the kernel workload.