coll/spacc: fix module initialization in v3.1.x

[mkurnosov]

Prepare the placeholder for the array of requests at module initialization time.

The call of coll/spacc/MPI_Allreduce with count < comm_size or with non-commutative binary operation produces the segmentation fault (only in v3.1.x).

The problem is that spacc/MPI_Allreduce in this case switched to the ompi_coll_base_allreduce_intra_basic_linear. But a lot of base collective operations require prepared placeholder for the array of requests.

Signed-off-by: Mikhail Kurnosov [email protected]

[mkurnosov]

Bug reproducer:

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

int main(int argc, char **argv)
{
    MPI_Comm comm = MPI_COMM_WORLD;
    int rank, comm_size;

    MPI_Init(&argc, &argv);
    MPI_Comm_size(comm, &comm_size);
    MPI_Comm_rank(comm, &rank);
    int count = comm_size - 1;

    int *sbuf = malloc(count * sizeof(*sbuf));
    int *rbuf = malloc(count * sizeof(*rbuf));
    if (!sbuf || !rbuf) {
        fprintf(stderr, "Error: malloc failed\n");
        MPI_Abort(comm, 1);
    }
    for (int i = 0; i < count; i++) {
        sbuf[i] = rank + i;
        rbuf[i] = 0;
    }

    MPI_Allreduce(sbuf, rbuf, count, MPI_INT, MPI_SUM, comm);

    /* Validate */
    for (int i = 0; i < count; i++) {
        int res = i * comm_size + (comm_size * (comm_size - 1)) / 2;
        if (rbuf[i] != res) {
            fprintf(stderr, "Error: validation error\n");
            break;
        }
    }

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

$ mpirun --hostfile ./hostfile -np 4 --mca coll_spacc_priority 100 ./allreduce_test