Evaluating the benefits of Argument Fusion: Optimizing host-device communication

Data movement remains a major bottleneck in heterogeneous systems with discrete GPUs, particularly when kernels expose many arguments that are transferred independently, leading to fragmented allocations and inefficient interconnect utilization. In this work, we evaluate Argument Fusion, a technique that merges multiple kernel arguments into a single contiguous buffer, reducing the number of allocation, transfer, and deallocation operations. We further study Argument Fusion+, which combines fusion with pinned memory and asynchronous CUDA transfers. Using both a parametric synthetic benchmark and real benchmarks from the Polybench suite, we show that Argument Fusion achieves up to 2.1 × speedup, while Argument Fusion+ reaches up to 2.6 ×. On real benchmarks, the combined approach improves end-to-end execution time by up to 68%.