high performance computing on graphics processing units: hgpu.org

This white paper reports on ours efforts to optimize a 2D/3D astrophysical (magento-)hydrodynamics Fortran code for XeonPhi. The code is parallelized with OpenMP and is suitable for execution on a shared memory system. Due to complexity of the code combined with immaturity of compiler we were unable to stay within the boundaries of Intel Compiler Suite. To deliver performance we took two different approaches. First, we optimized and partially rewrote most of the bandwidth-bound Fortran code to recover scalability on XeonPhi. Next, we ported several critical compute-bound hotspots to Intel SPMD Program Compiler (ISPC), which offers performance portability of a single source code across various architectures, such as Xeon, XeonPhi and possibly even GPU. This approach allowed us to achieve over 4x speed-up of the original code on dual-socket IvyBridge EP, and over 50x speed-up on the XeonPhi coprocessor. While the resulting optimized code can already be used in production to solve specific problems, we consider this project to be a proof-of-concept case reflecting the difficulty of achieving acceptable performance from XeonPhi on a "home-brewed" application.