127 lines
4.3 KiB
C
127 lines
4.3 KiB
C
/*
|
|
* MPI version with the matrix subdivided by "lines".
|
|
*/
|
|
|
|
#include <stdio.h>
|
|
#include <string.h>
|
|
#include <math.h>
|
|
#include <mpi.h>
|
|
#include "../config.h"
|
|
#include "../utils.h"
|
|
|
|
#define TAG_BORDER 0
|
|
#define TAG_MATRIX 1
|
|
|
|
float *compute_jacobi(int rank, int numprocs, int n, float init_value, float threshold, borders b, int *iterations) {
|
|
float *complete_x;
|
|
float *x;
|
|
float *new_x;
|
|
float *tmp_x;
|
|
float max_diff, global_max_diff, new_value;
|
|
int i, j;
|
|
int nb = n + 2; // n plus the border
|
|
int rows, rows_to_transmit;
|
|
int receive_pos;
|
|
|
|
if (rank == 0) {
|
|
rows = n - (n / numprocs) * (numprocs - 1);
|
|
} else {
|
|
rows = n / numprocs;
|
|
}
|
|
LOG(printf("[Process %d/%d] rows: %d\n", rank, numprocs, rows));
|
|
/* LOG(printf("[Process %d/%d] initializing matrix\n", rank, numprocs)); */
|
|
/* Initialize the matrix */
|
|
x = create_sa_matrix(rows + 2, nb);
|
|
new_x = create_sa_matrix(rows + 2, nb);
|
|
for (i = 0; i < rows + 2; i++) {
|
|
for (j = 1; j <= n; j++) {
|
|
x[IDX(nb, i, j)] = init_value;
|
|
new_x[IDX(nb, i, j)] = init_value;
|
|
}
|
|
}
|
|
/* Initialize boundary regions */
|
|
for (i = 0; i < rows + 2; i++) {
|
|
x[IDX(nb, i, 0)] = b.west;
|
|
x[IDX(nb, i, n + 1)] = b.east;
|
|
new_x[IDX(nb, i, 0)] = b.west;
|
|
new_x[IDX(nb, i, n + 1)] = b.east;
|
|
}
|
|
if (rank == 0) {
|
|
for (i = 1; i <= n + 1; i++) {
|
|
x[IDX(nb, 0, i)] = b.north;
|
|
new_x[IDX(nb, 0, i)] = b.north;
|
|
}
|
|
}
|
|
if (rank == numprocs - 1){
|
|
for (i = 1; i < n + 1; i++) {
|
|
x[IDX(nb, rows + 1, i)] = b.south;
|
|
new_x[IDX(nb, rows + 1, i)] = b.south;
|
|
}
|
|
}
|
|
/* LOG(printf("[Process %d/%d] matrix initialized\n", rank, numprocs)); */
|
|
/* Iterative refinement of x until values converge */
|
|
*iterations = 0;
|
|
do {
|
|
max_diff = 0;
|
|
global_max_diff = 0;
|
|
for (i = 1; i <= rows; i++) {
|
|
for (j = 1; j <= n; j++) {
|
|
new_value = 0.25 * (x[IDX(nb, i - 1, j)] + x[IDX(nb, i, j + 1)] + x[IDX(nb, i + 1, j)] + x[IDX(nb, i, j - 1)]);
|
|
max_diff = fmaxf(max_diff, fabs(new_value - x[IDX(nb, i, j)]));
|
|
new_x[IDX(nb, i, j)] = new_value;
|
|
}
|
|
}
|
|
tmp_x = new_x;
|
|
new_x = x;
|
|
x = tmp_x;
|
|
if (rank % 2 == 0) {
|
|
if (rank != numprocs - 1) {
|
|
// Send and receive south border
|
|
MPI_Send(&x[IDX(nb, rows, 0)], nb, MPI_FLOAT, rank + 1, TAG_BORDER, MPI_COMM_WORLD);
|
|
MPI_Recv(&x[IDX(nb, rows + 1, 0)], nb, MPI_FLOAT, rank + 1, TAG_BORDER, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
|
|
}
|
|
if (rank != 0) {
|
|
// Send and receive north border
|
|
MPI_Send(&x[IDX(nb, 1, 0)], nb, MPI_FLOAT, rank - 1, TAG_BORDER, MPI_COMM_WORLD);
|
|
MPI_Recv(&x[IDX(nb, 0, 0)], nb, MPI_FLOAT, rank - 1, TAG_BORDER, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
|
|
}
|
|
} else {
|
|
// Receive and send north border
|
|
MPI_Recv(&x[IDX(nb, 0, 0)], nb, MPI_FLOAT, rank - 1, TAG_BORDER, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
|
|
MPI_Send(&x[IDX(nb, 1, 0)], nb, MPI_FLOAT, rank - 1, TAG_BORDER, MPI_COMM_WORLD);
|
|
if (rank != numprocs - 1) {
|
|
// Receive and send south border
|
|
MPI_Recv(&x[IDX(nb, rows + 1, 0)], nb, MPI_FLOAT, rank + 1, TAG_BORDER, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
|
|
MPI_Send(&x[IDX(nb, rows, 0)], nb, MPI_FLOAT, rank + 1, TAG_BORDER, MPI_COMM_WORLD);
|
|
}
|
|
}
|
|
/* LOG(printf("[Process %d/%d] max_diff: %f\n", rank, numprocs, max_diff)); */
|
|
MPI_Allreduce(&max_diff, &global_max_diff, 1, MPI_FLOAT, MPI_MAX, MPI_COMM_WORLD);
|
|
/* LOG(printf("[Process %d/%d] global_max_diff: %f\n", rank, numprocs, global_max_diff)); */
|
|
(*iterations)++;
|
|
} while (global_max_diff > threshold);
|
|
|
|
if (rank == 0) {
|
|
complete_x = create_sa_matrix(nb, nb);
|
|
memcpy(complete_x, x, (rows + ((rank == numprocs - 1) ? 2 : 1)) * (nb) * sizeof(float));
|
|
rows_to_transmit = n / numprocs;
|
|
receive_pos = rows + 1;
|
|
for (i = 1; i < numprocs; i++) {
|
|
if (i == numprocs - 1) {
|
|
rows_to_transmit++;
|
|
}
|
|
MPI_Recv(&complete_x[IDX(nb, receive_pos, 0)], rows_to_transmit * (nb), MPI_FLOAT, i, TAG_MATRIX, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
|
|
receive_pos += n / numprocs;
|
|
}
|
|
} else {
|
|
complete_x = NULL;
|
|
rows_to_transmit = rows;
|
|
if (rank == numprocs - 1) {
|
|
rows_to_transmit++;
|
|
}
|
|
MPI_Send(&x[IDX(nb, 1, 0)], rows_to_transmit * (nb), MPI_FLOAT, 0, TAG_MATRIX, MPI_COMM_WORLD);
|
|
}
|
|
|
|
return complete_x;
|
|
}
|