Reduction of CPU-GPU Synchronization Overhead for Accelerating Implicit Clothing Simulator

Abstract

When trying to make the conjugate gradient (CG) method exploit GPU technology, this paper notes that the communication between CPU and GPU to transfer the residual value and waiting for the CPU's decision whether to continue further iterations is a new source of delay that has been overlooked and turns out not negligible. By examining the residual decrease pattern in log scale, this paper proposes so-called the Secant Lazy Residual Evaluation (Secant LRE) method to skip needless synchronization. We experimented the method for a clothing simulator and found that the proposed method reduces the sync overhead significantly, leading to 10 60% performance gain.