Harnessing GPU Power from High-level Libraries: Eigenvalues of Integral Operators with SLEPc

AbstractWe consider the approximation of eigenpairs of large-scale matrices arising from the discretization of integral operators with weakly singular function kernels. Efficient and fast solvers to numerically approximate the sought eigenpairs are required. For this, we would like to exploit the computational power of modern graphical processing units (GPUs), and we are interested in doing this from high-level libraries. We show how to use the CUDA add-on in the SLEPc/PETSc libraries to tackle this problem and illustrate our results on a radiative transfer problem in astrophysics. The CUDA-accelerated codes achieve considerable speedups versus the CPU counterparts on the same problem