Accelerating Embedded HOWFSC Algorithms

Abstract

The quest to directly image planets of other solar systems demands not only state-of-the- art coronagraphs, but also places extreme performance demands on space-based processors. Direct imaging requires precise wavefront control to acquire the 1010 contrast necessary to reveal a dim, Earth-like exoplanet. This precise level of control is only possible if high-order wavefront sensing and control (HOWFSC) algorithms are executed with enough speed to offset wavefront error accumulation. Of the many aspects that make high-contrast imaging difficult, a central bottleneck is the speed at which we can run these algorithms. At the center of this work, we aim to accelerate the execution of two foundational HOWFSC algorithms: optical modeling and Electric Field Conjugation (EFC). Optical modeling underpins both Jacobian-based EFC, and a relatively new variant of EFC, called adjoint-based EFC. The two main contributions of this thesis are to port bottleneck HOWFSC algorithms to the relevant computing environments, and quantify speedups attained by both algorithm choice and implementation optimization. This work explores the acceleration of optical modeling for a vector vortex coronagraph through the use of the FFTW library, and the acceleration of EFC by implementing adjoint-based EFC in an embedded context. We utilize functional analogs to radiation-hardened processors, using the NXP T1040 in place of the BAE RAD5545, and the NXP LS1046 in place of the LS1046-Space. We find that the FFTW library enabled a factor of six speedup for 4096 × 4096 fast Fourier transforms (FFTs), and a factor of five for 2048 × 2048 FFTs. With these significant speedups, the bottleneck within the vortex operations of the optical model shifts from the FFT to matrix multiplication. We additionally time the execution of the underlying routines of Jacobian-based EFC and AD-EFC to estimate that AD-EFC is 46 times faster than Jacobian-based EFC. Despite these speedups, AD-EFC is still a factor of 124 away from 100-second latency for our specific optical model. These results demonstrate that one to two orders of magnitude of speedup must be attained by either further optimizing algorithm implementations, or exploring other parallelization strategies, computing architectures, and mission paradigms to achieve a latency on the order of 100 seconds.