Order Analysis for Translating NESL Programs into Efficient GPU Code

The language NESL aims to facilitate GPU programming. In order to utilize the computation power of GPUs, NESL programs must be translated into efficient low-level code for execution. We propose a new translation technique. In NESL, apply-to-each is the main construct to extract parallel computation capability of GPUs. The result of apply-to-each is a sequence of elements. In traditional translation, the order of the elements in a sequence is always preserved. However, sometimes, the order need not be preserved and hence a faster method (which may not preserve the order of elements) for calculating the sequence may be employed. We propose the order analysis to determine if the order of elements in a sequence needs to be preserved. Order analysis is based on the taint analysis. In our experiments, we obtained 8.76x speedup on average.