coll/base/alltoall: fixes send/recv offsets in rbuf for the case of MPI_IN_PLACE

[mkurnosov]

Fixes send/recv offsets in rbuf for the case of MPI_IN_PLACE

Signed-off-by: Mikhail Kurnosov mkurnosov@gmail.com

[mkurnosov]

Reproducer:

#include <stdio.h>
#include <stdlib.h>
#include <inttypes.h>
#include <mpi.h>

int main(int argc, char **argv)
{
    int rank, commsize;

    MPI_Init(&argc, &argv);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    MPI_Comm_size(MPI_COMM_WORLD, &commsize);

    int count = 10;
    uint8_t *rbuf = malloc(sizeof(*rbuf) * count * commsize);

    for (int i = 0; i < commsize; i++) {
        for (int j = 0; j < count; j++) {
            rbuf[i * count + j] = rank;
        }
    }

    MPI_Alltoall(MPI_IN_PLACE, -1, MPI_DATATYPE_NULL, rbuf, count, MPI_UINT8_T, MPI_COMM_WORLD);

    for (int i = 0; i < commsize; i++) {
        for (int j = 0; j < count; j++) {
            uint8_t val = i;
            if (rbuf[i * count + j] != val) {
                fprintf(stderr, "Error: rank %d: Invalid value [%d/%d] = %d, need %d\n",
                        rank, i, j, rbuf[i * count + j], val);
                break;
            }
        }
    }

    free(rbuf);
    MPI_Finalize();
    return 0;
}

[bosilca]

nice catch.

[awlauria]

This may fix #9501

[awlauria]

v5.0.x pr: #9652