Large-Scale Chemical Informatics on GPUs

In this chapter we present the design and optimization of GPU implementations of two popular chemical similarity techniques: Gaussian shape overlay (GSO) and LINGO. GSO involves a data-parallel, arithmetically intensive iterative numerical optimization; we use it to examine issues of thread parallelism, arithmetic optimization, and CPU-GPU transfer overhead minimization. LINGO is a string similarity algorithm that, in its canonical CPU implementation, is bandwidth intensive and branch heavy, with limited data parallelism. We present an algorithmic redesign allowing GPU implementation of such a low arithmetic-intensity kernel and discuss techniques for memory optimization that enable large speedup. Source code for the programs described here is available online: PAPER (for Gaussian shape overlay) can be downloaded at https://simtk.org/home/paper under the GPL, and single-instruction, multiple LINGO (SIML) (for LINGO) at https://simtk.org/home/siml under a BSD license.