I am trying to divide columns of a 2D matrix among N process with MPI. For template I used the example on MPI_Scatter - sending columns of 2D array.
My code:
//HEADERS
char** board_initialize(int n, int m)
{
int k, l;
char* bd = (char*)malloc(sizeof(char) * n * m);
char** b = (char**)malloc(sizeof(char*) * n);
for (k = 0; k < n; k++)
b[k] = &bd[k * m];
for (k = 0; k < n; k++)
for (l = 0; l < m; l++)
b[k][l] = rand() < 0.25 * RAND_MAX;
return b;
}
void board_print(char** b, int n, int m)
{
int k, l;
// system("@cls||clear");
for (k = 0; k < n; k++)
{
for (l = 0; l < m; l++)
printf("%d", b[k][l]);
printf("\n");
}
printf("\n");
}
int main(int argc, char* argv[])
{
int N = 10;
int i, j;
char * boardptr = NULL; // ptr to board
char ** board; // board, 2D matrix, contignous memory allocation!
int procs, myid;
int mycols;
char ** myboard; // part of board that belongs to a process
MPI_Init(&argc, &argv); // initiailzation
MPI_Comm_rank(MPI_COMM_WORLD, &myid); // process ID
MPI_Comm_size(MPI_COMM_WORLD, &procs); // number of processes
// initialize global board
if (myid == 0)
{
srand(1573949136);
board = board_initialize(N, N);
boardptr = *board;
board_print(board, N, N);
}
// divide work
mycols = N / procs;
// initialize my structures
myboard = board_initialize(N,mycols);
MPI_Datatype column_not_resized, column_resized;
MPI_Type_vector(N, 1, N, MPI_CHAR, &column_not_resized);
MPI_Type_commit(&column_not_resized);
MPI_Type_create_resized(column_not_resized, 0, 1*sizeof(char), &column_resized);
MPI_Type_commit(&column_resized);
// scatter initial matrix
MPI_Scatter(boardptr, mycols, column_resized, *myboard, mycols, column_resized, 0, MPI_COMM_WORLD);
MPI_Barrier(MPI_COMM_WORLD);
board_print(myboard, N, mycols);
MPI_Finalize(); // finalize MPI
return 0;
}
The whole board looks like:
0000010010
0100000000
0000101100
0101000010
1000000100
0000010010
0000001110
0110000100
0000100000
0100010010
And if I use 2 process I expect that each proces would get half (first process columns 1-5 and second proces columns 6-10). But if I print myboard of both process I get some strange results:
proc0: proc1:
0 0 0 0 0 1 0 0 1 0
0 0 1 0 0 0 0 0 1 1
0 1 0 0 0 0 0 0 0 0
0 1 0 0 0 0 1 0 1 0
0 0 0 0 1 0 1 1 0 0
0 1 1 0 0 0 0 1 0 0
0 1 0 1 0 0 0 0 1 0
0 0 0 0 1 0 0 1 0 0
1 0 0 0 0 0 0 1 0 0
0 1 0 0 0 0 1 0 0 0
It's probably some silly mistake, but I just can't seem to find it. Any help would be really appreciated. <3
Note: the output of proc1 is probably just some garbage, because I get different output on every run.
You forgot to differentiate between send and receive types, i.e., the parameters of MPI_Type_vector will depend on the appropriate type. You need to do something as follows:
MPI_Datatype acol, acoltype, bcol, bcoltype;
if (myid == 0) {
MPI_Type_vector(N, 1, N, MPI_CHAR, &acol);
MPI_Type_commit(&acol);
MPI_Type_create_resized(acol, 0, 1*sizeof(char), &acoltype);
}
MPI_Type_vector(N, 1, mycols, MPI_CHAR, &bcol);
MPI_Type_commit(&bcol);
MPI_Type_create_resized(bcol, 0, 1*sizeof(char), &bcoltype);
MPI_Type_commit(&bcoltype);
MPI_Scatter (boardptr, mycols, acoltype, *myboard, mycols, bcoltype, 0, MPI_COMM_WORLD);
From the sender point of view you want to create N blocks (i.e., N rows), with a size of 1 and a stride of N (i.e., N columns). Hence:
MPI_Type_vector(N, 1, N, MPI_CHAR, &acol);
From the receiver point of view you want to create N blocks (i.e., N rows), with a size of 1 and a stride of mycols (i.e., mycols columns). Hence:
MPI_Type_vector(N, 1, mycols, MPI_CHAR, &bcol);
Side note you do need the MPI_Barrier after the MPI_Scatter since the latter is a blocking call already.
The final code would look like the following:
#include <stdio.h>
#include <stdlib.h>
#include <mpi.h>
char** board_initialize(int n, int m)
{
int k, l;
char* bd = (char*)malloc(sizeof(char) * n * m);
char** b = (char**)malloc(sizeof(char*) * n);
for (k = 0; k < n; k++)
b[k] = &bd[k * m];
for (k = 0; k < n; k++)
for (l = 0; l < m; l++)
b[k][l] = rand() < 0.25 * RAND_MAX;
return b;
}
void board_print(char** b, int n, int m)
{
int k, l;
// system("@cls||clear");
for (k = 0; k < n; k++)
{
for (l = 0; l < m; l++)
printf("%d", b[k][l]);
printf("\n");
}
printf("\n");
}
int main(int argc, char* argv[])
{
int N = 10;
int i, j;
char * boardptr = NULL; // ptr to board
char ** board; // board, 2D matrix, contignous memory allocation!
int procs, myid;
int mycols;
char ** myboard; // part of board that belongs to a process
MPI_Init(&argc, &argv); // initiailzation
MPI_Comm_rank(MPI_COMM_WORLD, &myid); // process ID
MPI_Comm_size(MPI_COMM_WORLD, &procs); // number of processes
// initialize global board
if (myid == 0)
{
srand(1573949136);
board = board_initialize(N, N);
boardptr = *board;
board_print(board, N, N);
}
// divide work
mycols = N / procs;
// initialize my structures
myboard = board_initialize(N,mycols);
MPI_Datatype acol, acoltype, bcol, bcoltype;
if (myid == 0) {
MPI_Type_vector(N,
1,
N,
MPI_CHAR,
&acol);
MPI_Type_commit(&acol);
MPI_Type_create_resized(acol, 0, 1*sizeof(char), &acoltype);
}
MPI_Type_vector(N,
1,
mycols,
MPI_CHAR,
&bcol);
MPI_Type_commit(&bcol);
MPI_Type_create_resized(bcol, 0, 1*sizeof(char), &bcoltype);
MPI_Type_commit(&bcoltype);
MPI_Scatter (boardptr, mycols, acoltype, *myboard, mycols, bcoltype, 0, MPI_COMM_WORLD);
board_print(myboard, N, mycols);
MPI_Finalize(); // finalize MPI
return 0;
}
Alternatively, and much easier in IMO is to divide by rows instead of columns the code would look like the following:
int myrows = N / procs;
myboard = board_initialize(myrows, N);
MPI_Scatter (boardptr, myrows, MPI_CHAR, *myboard, myrows, MPI_CHAR, 0, MPI_COMM_WORLD);
board_print(myboard, myrows, N);