Structured Orthogonal Inversion of Block p-Cyclic Matrices on Multicore with GPU Accelerators