Implementing Fox’s Algorithm

void Fox(int n, GRID_INFO_T* grid,

LOCAL_MATRIX_T* local_A,

LOCAL_MATRIX_T* local_B,

LOCAL_MATRIX_T* local_C) {

/* Storage for the sub-matrix of A */

LOCAL_MATRIX_T* temp_A;

int stage;

int bcast_root;

int n_bar;

int source;

int dest;

MPI_Status status;

n_bar = n/grid->q;

Set_to_zero(local_C);

/* Calculate addresses for circular shift of B */

source = (grid->my_row + 1) % grid->q;

dest = (grid->my_row + grid->q - 1) % grid->q;

/* Storage for the broadcast block of A */

temp_A = Local_matrix_allocate(n_bar);

for (stage = 0; stage < grid->q; stage++) {

bcast_root = (grid->my_row + stage) % grid->q;

if (bcast_root == grid->my_col) {

MPI_Bcast(local_A, 1, local_matrix_mpi_t,

bcast_root, grid->row_comm);

Local_matrix_multiply(local_A, local_B,

local_C);

} else {

MPI_Bcast(temp_A, 1, local_matrix_mpi_t,

bcast_root, grid->row_comm);

Local_matrix_multiply(temp_A, local_B,

local_C);

}

MPI_Sendrecv_replace(local_B, 1,

local_matrix_mpi_t,dest, 0, source, 0,

grid->col_comm, &status);

} /* for */

} /* Fox */

Previous slide

Next slide

Back to first slide

View graphic version