Efficient COO to CSR Conversion for Accelerating Sparse Matrix Processing on FPGA

Sparse matrix processing is an important computational kernel widely applied in various fields such as graph processing, and big data analysis. In many applications, sparse matrix processing is often a bottleneck, so various accelerators for it have been proposed. These accelerators often process sparse matrices in compressed formats like Coordinate (COO) or Compressed Sparse Row (CSR), which store only nonzero elements, to optimize memory usage. Some accelerators perform the conversion of output matrices calculated in COO format to CSR format, which enables a higher compression ratio, in order to reduce external memory traffic. Given that external memory bandwidth is the performance bottleneck in most cases, designing an efficient COO to CSR (COO2CSR) converter is an important issue that needs to be addressed. In this paper, we propose a COO2CSR conversion method that overcomes the challenge of performing the conversion in a highly parallel manner without prior knowledge of the target matrix. Based on this method, we develop a high-performance COO2CSR converter. We simulated our converter and found that it achieves near-optimal throughput. In addition, logic synthesis results on an Alveo U55C FPGA board showed that the converter consumes only 1.07% of LUTs, 0.65% of FFs, and 1.24% of BRAMs.