Fast Expression Templates
J. Härdtlein, Alexander Linke, Christoph Pflaum · Lecture notes in computer science · 2005
Expression templates (ET) can significantly reduce the implementation effort of mathematical software. For some compilers, especially for those of supercomputers, however, it can be observed that classical ET implementations do not deliver the expected performance. This is because aliasing of pointers in combination with the complicated ET constructs becomes much more difficult. Therefore, we introduced the concept of enumerated variables, which are provided with an additional integer template parameter. Based on this new implementation of ET we obtain a C++ code whose performance is very close to the handcrafted C code. The performance results of these so-called Fast ET are presented for the Hitachi SR8000 supercomputer and the NEC SX6, both with automatic vectorization and parallelization. Additionally we studied the combination of Fast ET and OpenMP on a high performance Opteron cluster. These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.