pOSKI: An Extensible Autotuning Framework to Perform Optimized SpMVs on Multicore Architectures

We have developed pOSKI: the Parallel Optimized Sparse Kernel Interface – an autotuning framework to optimize Sparse Matrix Vector Multiply (SpMV) performance on emerging shared memory multicore architectures. Our autotuning methodology extends previous work done in the scientific computing community targeting serial architectures. In addition to previously explored parallel optimizations, we find that that load balanced data decomposition is extremely important to achieving good parallel performance on the new generation of parallel architectures. Our best parallel configurations perform up to 9x faster than optimized serial codes on the AMD Santa Rosa architecture, 11.3x faster on the AMD Barcelona architecture, and 7.2x faster on the Intel Clovertown architecture.