Workshops Archive

Mixed-Precision Cholesky Factorization for Semidefinite Programming Using Ozaki Scheme II


Workshop: High-Accuracy Computing on Low-Precision Hardware

Authors: Haruki Nakamae, Toshio Endo, and Katsuki Fujisawa (Institute of Science Tokyo)

Abstract: This paper explores the accuracy and performance of a mixed-precision Cholesky factorization based on the INT8 Ozaki scheme II. Within the tiled Cholesky algorithm, we apply an optimized Ozaki scheme II to the symmetric rank-k update (SYRK), while keeping the diagonal factorization (POTRF) and the triangular solve (TRSM) in double precision. We further introduce a look-ahead technique and a dynamic tile size scheduler. On a server whose GPU has limited FP64 throughput, our implementation attains a 3.04x speedup over the OpenBLAS dpotrf routine while retaining double-precision accuracy. We also examine how the number of slices affects the backward error and the relative forward error, and find that 15 slices are required to preserve double-precision accuracy within the range covered by our experiment. Finally, we integrate the proposed factorization into a double-precision semidefinite programming solver and obtain a 1.47x speedup on an SDP benchmark instance (1et.2048).


Back to High-Accuracy Computing on Low-Precision Hardware Archive Listing

Back to Full Workshop Archive Listing