Level-3 BLAS implementation

To derive a level-3 BLAS left-looking variant for computing the factorization, consider the partitioning

The assumption is that bold-face parts of the lower triangular matrix have already been computed, and have overwritten the corresponding parts of A . The rest of the matrix has not been updated at all, and the object of the next step is to compute the next parts of the lower triangular matrix, and , overwriting the corresponding parts of A . From the above equation, we derive

or

    

The algorithm for the left looking version of the Cholesky factorization can be given as follows using the above equations

1. Update the current panel according to Equations  and .
2. Perform a Cholesky factorization of updated (Equation ).
3. Now that known, compute from Equation .
4. Continue recursively by repartitioning the matrix.

The PLAPACK implementation using global level-3 BLAS is given in Figure 8.4. This time the bulk of the computation is in the update , which is a matrix-panel operation. Thus, the algorithmic blocking size b is determined by the call

PLA_Environ_nb_alg( PLA_OP_MAT_PAN, template, &nb_alg );

Notice how the code reflect the above described algorithm, which could have been taken straight from a number of textbooks (e.g., []).