high performance computing on graphics processing units: hgpu.org

Although general purpose GPUs have relatively high computing capacity, they also introduce high power consumption compared with general purpose CPUs. Therefore low-power techniques targeted for GPUs will be one of the most hot topics in the future. On the other hand, in several application domains, users are unwilling to sacrifice performance to save power. In this paper, we propose an effective kernel fusion method to reduce the power consumption for GPUs without performance loss. Different from executing multiple kernels serially, the proposed method fuses several kernels into one larger kernel. Owing to the fact that most consecutive kernels in an application have data dependency and could not be fused directly, we split large kernel into multiple slices with strip-mining method, then fuse independent sliced kernels into one kernel. Based on the CUDA programming model, we propose three different kernel fusion implementations, with each one targeting for a special case. Based on the different strip-ming methods, we also propose two fusion mechanisms, which are called invariant-slice fusion and variant-slice fusion. The latter one could be better adapted to the requirements of the kernels to be fused. The experimental results validate that the proposed kernel fusion method could effectively reduce the power consumption for GPU.