An access pattern based adaptive mapping function for GPGPU scratchpad memory

As modern GPUs integrate massive processing elements and limited memories on-chip, the efficiency of using their scratchpad memories becomes important for performance and energy. To meet bandwidth requirement of simultaneously accessing of a thread array, multi-bank design, dividing a scratchpad memory into equally-sized memory modules, are widely used. However, the complex access patterns in real-world applications can cause the bank conflicts which comes from different threads accessing the same bank at the same time, and the conflicts hinder the performance sharply. A mapping function is a method that redistributes the accesses according to access addresses. To reduce bank conflicts some scratchpad memory mapping functions are exploited, such as XOR based hash functions and configurable functions. In this paper, we propose an adaptive mapping function, which can dynamically select a suitable mapping function for applications based on the statistics of first block executing. The experimental results show that 94.8 percent bank conflicts reduced and 1.235x performance improved for 17 benchmarks on GPGPU-sim, a Fermi-like simulator.