9.4.2 Performing Multiplications

The user interacts with the library through CLA_Matrix_interface objects. Their constructor takes no arguments. A call to make_multiplier (described below) will properly initialize them. The class has two methods the user needs to use. First, is the multiply method. It has the following signature:

void multiply(double alpha, double beta, double *data, void (*fptr) (void *),
              void *usr_data, int x, int y);

Each element of the user's $A, B,$ and $C$ arrays must call this method to perform the multiplication.

• The parameters alpha and beta correspond to $\alpha$ and $\beta$ in $C = \beta C + \alpha AB$. The same value must be passed for each of these two parameters at each elemnts of $A, B,$ and $C$.

• The paremeter data specifies where the data to be mutiplier is held. This is a two-dimensional array stored in a one-dimensional array. The size of the array is determined at creation time (based on user paremeters) and does not have to be specified at this point. For $A$ and $B$ data should contain the matrices to be multiplied. For $C$, it is the destination where the result will be stored (and if $\beta \neq 0$, the data which will be multiplied by $\beta$ and added to $\alpha AB$).

• The fptr paremeter specifies the callback which the library calls (at array $C$ when the computation is complete. When it is called, it passes usr_data to it. Only the elements of the $C$ array must specify a non-null value for fptr.

• The values of x and y specify the coordinates of the calling element in the users chare array (these should always be thisIndex.x and thisIndex.y, respectively).

The second function of the CLA_Matrix_interface objects is the sync method. This should be called at each element of the array when it is ready to migrate. It takes as arguments two intergers, x and y, which should always be thisIndex.x and thisIndex.y, respectively. Note that migration is currently only supported for the ``2D'' algorithm.