 |
Center for Nonlinear Studies |
Auto-parallelization of programs that have not been developed with parallelism in mind is one of the holy grails in computer science. It allows to stick to a simple programming model but requires understanding the source code's data flow to automatically distribute the data, parallelize the computations, and infer synchronizations where necessary. We introduce a technique to analytically count iterations of a large class of loops. This illustrates how data-flow can be modeled in practice. Then, we discuss our new LLVM-based research compiler Polly-ACC that enables automatic compilation to accelerator devices such as GPUs. Unfortunately, its applicability is limited to codes for which the iteration space and all accesses can be described as affine functions. In the second part of the talk, we will discuss dCUDA, a way to express parallel codes in MPI-RMA, a well-known communication library, to map them automatically to GPU clusters. The dCUDA approach enables simple and portable programming across heterogeneous devices due to programmer-specified locality. Furthermore, dCUDA enables hardware-supported overlap of computation and communication and is applicable to next-generation technologies such as NVLINK. We will demonstrate encouraging initial results and show limitations of current devices in order to start a discussion.
Host: Christoph Junghans