high performance computing on graphics processing units: hgpu.org

This paper presents GPU-based solutions to the Poisson equation with homogeneous Dirichlet boundary conditions in two spatial dimensions. This problem has well-understood behavior, but similar computation to many more complex real-world problems. We analyze the GPU performance using three types of memory access in the CUDA memory model (direct access to global memory, texture access, and shared memory). Based on data locality, different CUDA algorithms are designed to accommodate the different device memory performance behaviors. We present a performance study on the speedup of our GPU-based solutions on an NVIDIA Tesla C2070 over serial code. By relating the data access pattern and its spatial locality, our results show that an algorithm using global memory with coalesced reads outperforms the other memory systems and allows effective solvers using single precision floating points.