Data used in irregular computations such as molecular dynamics simulation, climate modeling and big graph analysis are usually sparse. Sparse formats provide space reduction and computation reduction opportunities to store non-zero values in sparse data and reduce unnecessary computations involving zeros respectively.
There exist various sparse formats, each sparse format been more suitable for specific irregular application, however, sparse data is generally stored in few common formats. The most commonly used sparse input format is COO. In this research, we propose using the polyhedral framework to enable conversion from one sparse format to another using shared information common to these formats. We define various sparse formats in terms of their dense abstraction and use the common information between each format to facilitate a transformation from one format to another.
Our contribution in this research is to generate highly optimized code for converting from one format to another and we are going to show our results with hand-optimized sparse format conversion libraries such as SPARSKIT and INTEL MKL.