摘要 |
When a real symmetric matrix is tridiagonalized using a shared memory scalar parallel computer, the matrix is divided into threads to be updated. In the updating, the data of the lower triangular portion is for the calculation utilizing the symmetry in the matrix. The lower triangular matrix portion is divided into columns, and the resultant pieces are assigned to respective CPUs so that an equal number of elements are assigned to each of the CPUs. Also, the computation is performed simultaneously in the column and row directions once for each loading of the data in order to reduce the number of times data is loaded into memory. The lower triangular matrix portion of the diagonal block matrix portion is divided into one small square matrix and two small triangular matrices recursively so that they are updated.
|