Output details
11 - Computer Science and Informatics
University of St Andrews
Reducing Training Time in a One-shot Machine Learning-based Compiler
<08>First paper to present a fully operational 'drop-in' replacement machine learning compiler, which does not rely on multiple compilations and search at compile time - transparently usable without profiling. Shows 14% speedup over a mature compiler (GCC-O3) on an industry standard benchmark suite (EEMBC). First use of unsupervised learning for compiler optimisation - allows the training of models to scale with a growing number of programs. Training time required for equivalent performance reduced by a factor of 7. The venue is technically a workshop, but has an important history, and is more highly regarded than many conferences in the area.