Threads Shared Memories L 1 Cache Read-only Data Cache L 2 Cache DRAM

This paper presents the experimental evaluation of a new data mapping technique for the GPU shared memory, called Adaptive Modular Mapping (AMM). The evaluated technique aims to remap data across the shared memory physical banks, so as to increase parallel accesses, resulting in appreciable gains in terms of performance. Unless previous techniques described in literature, AMM does not increase shared memory size as a side effect of the conflict-avoidance technique. The paper also presents the experimental set-up used for the validation of the proposed memory mapping methodology.