A compression-based memory-efficient optimization for out-of-core GPU stencil computation

A code for out-of-core stencil computation manages data that exceeds the memory capacity of a GPU. However, such a code necessitates frequent data transfers between the CPU and GPU, which often impede overall performance. In this work, we propose a compression-based, memory-efficient method to accelerate out-of-core stencil codes. First, an on-the-fly compression technique is integrated into the out of-core computation to reduce CPU-GPU data transfers. Secondly, a single-working buffer strategy is employed to reduce the GPU memory usage, enabling more data to be stored on the GPU for reuse, resulting in increased temporal blocking steps. Experimental results demonstrate that the proposed method significantly reduces the GPU memory usage by 21%, thereby creating space for doubling the number of temporal blocking steps compared to the codes without compression. Our proposed method has shown to help the high-order, data-transfer-bound stencil codes achieve speedups up to 2.09x for single-precision floating-point format and up to 1.92x for double-precision floating-point format on an NVIDIA Tesla V100 GPU in comparison with the codes without compression.