问题
I'm trying to transpose a matrix using MPI in C. Each process has a square submatrix, and I want to send that to the right process (the 'opposite' one on the grid), transposing it as part of the communication.
I'm using MPI_Type_create_subarray which has an argument for the order, either MPI_ORDER_C or MPI_ORDER_FORTRAN for row-major and column-major respectively. I thought that if I sent as one of these, and received as the other, then my matrix would be transposed as part of the communication. However, this doesn't seem to happen - it just stays non-transposed.
The important part of the code is below, and the whole code file is available at this gist. Does anyone have any ideas why this isn't working? Should this approach to doing the transpose work? I'd have thought it would, having read the descriptions of MPI_ORDER_C and MPI_ORDER_FORTRAN, but maybe not.
/* ----------- DO TRANSPOSE ----------- */
/* Find the opposite co-ordinates (as we know it's a square) */
coords2[0] = coords[1];
coords2[1] = coords[0];
/* Get the rank for this process */
MPI_Cart_rank(cart_comm, coords2, &rank2);
/* Send to these new coordinates */
tag = (coords[0] + 1) * (coords[1] + 1);
/* Create new derived type to receive as */
/* MPI_Type_vector(rows_in_core, cols_in_core, cols_in_core, MPI_DOUBLE, &vector_type); */
sizes[0] = rows_in_core;
sizes[1] = cols_in_core;
subsizes[0] = rows_in_core;
subsizes[1] = cols_in_core;
starts[0] = 0;
starts[1] = 0;
MPI_Type_create_subarray(2, sizes, subsizes, starts, MPI_ORDER_FORTRAN, MPI_DOUBLE, &send_type);
MPI_Type_commit(&send_type);
MPI_Type_create_subarray(2, sizes, subsizes, starts, MPI_ORDER_C, MPI_DOUBLE, &recv_type);
MPI_Type_commit(&recv_type);
/* We're sending in row-major form, so it's just rows_in_core * cols_in_core lots of MPI_DOUBLE */
MPI_Send(&array[0][0], 1, send_type, rank2, tag ,cart_comm);
/* Receive from these new coordinates */
MPI_Recv(&new_array[0][0], 1, recv_type, rank2, tag, cart_comm, &status);
回答1:
I would have thought this would work, too, but apparently not.
If you slog through the relevant bit of the MPI standard where it actually defines the resulting typemap, the reason becomes clear -- MPI_Type_create_subarray maps out the region that the subarray takes in the full array, but marches through the memory in linear order, so the data layout doesn't change. In other words, when the sizes equal the subsizes, the subarray is just a contiguous block of memory; and for a subarray strictly smaller than the whole array, you're just changing the subregion that is being sent/received to, not the data ordering. You can see the effect when choosing just a subregion:
int sizes[]={cols,rows};
int subsizes[]={2,4};
int starts[]={1,1};
MPI_Type_create_subarray(2, sizes, subsizes, starts, MPI_ORDER_FORTRAN, MPI_INT, &ftype);
MPI_Type_commit(&ftype);
MPI_Type_create_subarray(2, sizes, subsizes, starts, MPI_ORDER_C, MPI_INT, &ctype);
MPI_Type_commit(&ctype);
MPI_Isend(&(send[0][0]), 1, ctype, 0, 1, MPI_COMM_WORLD,&reqc);
MPI_Recv(&(recvc[0][0]), 1, ctype, 0, 1, MPI_COMM_WORLD, &statusc);
MPI_Isend(&(send[0][0]), 1, ctype, 0, 1, MPI_COMM_WORLD,&reqf);
MPI_Recv(&(recvf[0][0]), 1, ftype, 0, 1, MPI_COMM_WORLD, &statusf);
/*...*/
printf("Original:\n");
printarr(send,rows,cols);
printf("\nReceived -- C order:\n");
printarr(recvc,rows,cols);
printf("\nReceived: -- Fortran order:\n");
printarr(recvf,rows,cols);
gives you this:
0 1 2 3 4 5 6
10 11 12 13 14 15 16
20 21 22 23 24 25 26
30 31 32 33 34 35 36
40 41 42 43 44 45 46
50 51 52 53 54 55 56
60 61 62 63 64 65 66
Received -- C order:
0 0 0 0 0 0 0
0 11 12 13 14 0 0
0 21 22 23 24 0 0
0 0 0 0 0 0 0
0 0 0 0 0 0 0
0 0 0 0 0 0 0
0 0 0 0 0 0 0
Received: -- Fortran order:
0 0 0 0 0 0 0
0 11 12 0 0 0 0
0 13 14 0 0 0 0
0 21 22 0 0 0 0
0 23 24 0 0 0 0
0 0 0 0 0 0 0
0 0 0 0 0 0 0
So the same data is getting sent and received; all that's really happening is that the arrays sizes, subsizes and starts are being reversed.
You can transpose with MPI datatypes -- the standard even gives a couple of examples, one of which I've transliterated into C here -- but you have to create the types yourself. The good news is that it's really no longer than the subarray stuff:
MPI_Type_vector(rows, 1, cols, MPI_INT, &col);
MPI_Type_hvector(cols, 1, sizeof(int), col, &transpose);
MPI_Type_commit(&transpose);
MPI_Isend(&(send[0][0]), rows*cols, MPI_INT, 0, 1, MPI_COMM_WORLD,&req);
MPI_Recv(&(recv[0][0]), 1, transpose, 0, 1, MPI_COMM_WORLD, &status);
MPI_Type_free(&col);
MPI_Type_free(&transpose);
printf("Original:\n");
printarr(send,rows,cols);
printf("Received\n");
printarr(recv,rows,cols);
$ mpirun -np 1 ./transpose2
Original:
0 1 2 3 4 5 6
10 11 12 13 14 15 16
20 21 22 23 24 25 26
30 31 32 33 34 35 36
40 41 42 43 44 45 46
50 51 52 53 54 55 56
60 61 62 63 64 65 66
Received
0 10 20 30 40 50 60
1 11 21 31 41 51 61
2 12 22 32 42 52 62
3 13 23 33 43 53 63
4 14 24 34 44 54 64
5 15 25 35 45 55 65
6 16 26 36 46 56 66
来源：https://stackoverflow.com/questions/5626123/can-you-transpose-array-when-sending-using-mpi-type-create-subarray