M.M. El-Awad and M.Z. Yusoff (Malaysia)
Finite Element Method, Supercomputers, Vectorization
This paper deals with the vectorized implementation of a CFD finite element model. It focuses on the procedure that calculates and assembles the finite-element system of equations, which is usually the most time-intensive operation in the numerical solution. The paper shows that, by programming this step in a way that places the DO loop over the grid elements into the innermost structure instead of that over the Gauss-points, the vector processor can operate on sufficiently long vectors. Thus, effective use of the vector-processing capability can be achieved. Results on a Cyber 205 show that the computer time of the assembly process in a 2D version of the model could be reduced by an order of magnitude and that the saving increase as the grid is refined. For the 3D version of the model the computer time could be reduced by more than two orders of magnitude by suitably structuring the assembly procedure, vectorization, and the use of reduced quadrature.
Important Links:
Go Back
