A reduced unconstrained system for the cloth dynamics solver

Abstract

Modern direct solvers have been more and more widely used by computer graphics community for solving sparse linear systems, such as those that arise in cloth simulation. However, external constraints usually prevent a direct method from being used for cloth simulation due to the singularity of the constrained system. This paper makes two major contributions towards the re-introduction of direct methods for cloth dynamics solvers. The first one is an approach which eliminates all the constrained variables from the system so that we obtain a reduced, nonsingular and unconstrained system. As alternatives to the well-known MPCG algorithm, not only the original, unmodified PCG method, but also any direct method can be used to solve the reduced system at a lower cost. Our second contribution is a novel direct-iterative scheme applied for the reduced system, which is basically the conjugate gradient method using a special preconditioner. Specifically, we use the stiff part of the coefficient matrix, which we call the matrix core, as the preconditioner for the PCG. The inverse of this preconditioner is computed by any eligible direct solver. The direct-iterative method has proved to be more efficient than both direct and iterative methods. Our experiments show a factor of two speedup over direct methods when stiff springs are used, even greater improvements over the MPCG iterative method.