Fig. 9

Performance of segmentation and feature Computation steps with different versions: multi-core CPU, multi-GPU, and cooperative CPU-GPU. The CPU-GPU version also evaluates the composition of the application as a single level coarse-grained pipeline (1 L) in which all stage operations are executed as a single task, and the hierarchical pipeline (2 L) in which fine-grained operations in a stage are exported to the runtime system. Additionally, FCFS and PATS scheduling strategies are used to assign tasks to CPUs and GPUs. a Performance of multi-core CPU, multi-GPU and cooperative CPU-GPU versions of the application. b Improvements with data locality (DL) mapping and asynchronous data copy optimizations