--- /dev/null
+/* Copyright (c) 2013-2014. The SimGrid Team.
+ * All rights reserved. */
+
+/* This program is free software; you can redistribute it and/or modify it
+ * under the terms of the license (GNU LGPL) which comes with this package. */
+
+#include "colls_private.h"
+#include <math.h>
+
+/*****************************************************************************
+
+ * Function: alltoall_3dmesh_shoot
+
+ * Return: int
+
+ * Inputs:
+ send_buff: send input buffer
+ send_count: number of elements to send
+ send_type: data type of elements being sent
+ recv_buff: receive output buffer
+ recv_count: number of elements to received
+ recv_type: data type of elements being received
+ comm: communicator
+
+ * Descrp: Function realizes the alltoall operation using the 3dmesh
+ algorithm. It actually performs allgather operation in x dimension,
+ y dimension, then in z dimension. Each node then extracts the
+ needed data. The communication in all dimension is simple.
+
+ * Auther: Ahmad Faraj
+****************************************************************************/
+
+static int alltoall_check_is_3dmesh(int num, int *i, int *j, int *k)
+{
+ int x, max = num / 3;
+ x = cbrt(num);
+ *i = *j = *k = 0;
+ while (x <= max) {
+ if ((num % (x * x)) == 0) {
+ *i = *j = x;
+ *k = num / (x * x);
+ return 1;
+ }
+ x++;
+ }
+ return 0;
+}
+
+int smpi_coll_tuned_alltoall_3dmesh(void *send_buff, int send_count,
+ MPI_Datatype send_type,
+ void *recv_buff, int recv_count,
+ MPI_Datatype recv_type, MPI_Comm comm)
+{
+ MPI_Request *reqs, *req_ptr;
+ MPI_Aint extent;
+ MPI_Status status, *statuses;
+ int i, j, src, dst, rank, num_procs, num_reqs, X, Y, Z, block_size, count;
+ int my_z, two_dsize, my_row_base, my_col_base, my_z_base, src_row_base;
+ int src_z_base, send_offset, recv_offset, tag = COLL_TAG_ALLTOALL;
+
+ char *tmp_buff1, *tmp_buff2;
+
+ rank = smpi_comm_rank(comm);
+ num_procs = smpi_comm_size(comm);
+ extent = smpi_datatype_get_extent(send_type);
+
+ if (!alltoall_check_is_3dmesh(num_procs, &X, &Y, &Z))
+ return MPI_ERR_OTHER;
+
+ num_reqs = X;
+ if (Y > X)
+ num_reqs = Y;
+ if (Z > Y)
+ num_reqs = Z;
+
+ two_dsize = X * Y;
+ my_z = rank / two_dsize;
+
+ my_row_base = (rank / X) * X;
+ my_col_base = (rank % Y) + (my_z * two_dsize);
+ my_z_base = my_z * two_dsize;
+
+ block_size = extent * send_count;
+
+ tmp_buff1 = (char *) smpi_get_tmp_sendbuffer(block_size * num_procs * two_dsize);
+ tmp_buff2 = (char *) smpi_get_tmp_recvbuffer(block_size * two_dsize);
+
+ statuses = (MPI_Status *) xbt_malloc(num_reqs * sizeof(MPI_Status));
+ reqs = (MPI_Request *) xbt_malloc(num_reqs * sizeof(MPI_Request));
+
+ req_ptr = reqs;
+
+ recv_offset = (rank % two_dsize) * block_size * num_procs;
+
+ smpi_mpi_sendrecv(send_buff, send_count * num_procs, send_type, rank, tag,
+ tmp_buff1 + recv_offset, num_procs * recv_count,
+ recv_type, rank, tag, comm, &status);
+
+ count = send_count * num_procs;
+
+ for (i = 0; i < Y; i++) {
+ src = i + my_row_base;
+ if (src == rank)
+ continue;
+ recv_offset = (src % two_dsize) * block_size * num_procs;
+ *(req_ptr++) = smpi_mpi_irecv(tmp_buff1 + recv_offset, count, recv_type, src, tag, comm);
+ }
+
+ for (i = 0; i < Y; i++) {
+ dst = i + my_row_base;
+ if (dst == rank)
+ continue;
+ smpi_mpi_send(send_buff, count, send_type, dst, tag, comm);
+ }
+
+ smpi_mpi_waitall(Y - 1, reqs, statuses);
+ req_ptr = reqs;
+
+
+ for (i = 0; i < X; i++) {
+ src = (i * Y + my_col_base);
+ if (src == rank)
+ continue;
+
+ src_row_base = (src / X) * X;
+
+ recv_offset = (src_row_base % two_dsize) * block_size * num_procs;
+ *(req_ptr++) = smpi_mpi_irecv(tmp_buff1 + recv_offset, recv_count * num_procs * Y,
+ recv_type, src, tag, comm);
+ }
+
+ send_offset = (my_row_base % two_dsize) * block_size * num_procs;
+ for (i = 0; i < X; i++) {
+ dst = (i * Y + my_col_base);
+ if (dst == rank)
+ continue;
+ smpi_mpi_send(tmp_buff1 + send_offset, send_count * num_procs * Y, send_type,
+ dst, tag, comm);
+ }
+
+ smpi_mpi_waitall(X - 1, reqs, statuses);
+ req_ptr = reqs;
+
+ for (i = 0; i < two_dsize; i++) {
+ send_offset = (rank * block_size) + (i * block_size * num_procs);
+ recv_offset = (my_z_base * block_size) + (i * block_size);
+ smpi_mpi_sendrecv(tmp_buff1 + send_offset, send_count, send_type, rank, tag,
+ (char *) recv_buff + recv_offset, recv_count, recv_type,
+ rank, tag, comm, &status);
+ }
+
+ for (i = 1; i < Z; i++) {
+ src = (rank + i * two_dsize) % num_procs;
+ src_z_base = (src / two_dsize) * two_dsize;
+
+ recv_offset = (src_z_base * block_size);
+
+ *(req_ptr++) = smpi_mpi_irecv((char *) recv_buff + recv_offset, recv_count * two_dsize,
+ recv_type, src, tag, comm);
+ }
+
+ for (i = 1; i < Z; i++) {
+ dst = (rank + i * two_dsize) % num_procs;
+
+ recv_offset = 0;
+ for (j = 0; j < two_dsize; j++) {
+ send_offset = (dst + j * num_procs) * block_size;
+ smpi_mpi_sendrecv(tmp_buff1 + send_offset, send_count, send_type,
+ rank, tag, tmp_buff2 + recv_offset, recv_count,
+ recv_type, rank, tag, comm, &status);
+
+ recv_offset += block_size;
+ }
+
+ smpi_mpi_send(tmp_buff2, send_count * two_dsize, send_type, dst, tag, comm);
+
+ }
+
+ smpi_mpi_waitall(Z - 1, reqs, statuses);
+
+ free(reqs);
+ free(statuses);
+ smpi_free_tmp_buffer(tmp_buff1);
+ smpi_free_tmp_buffer(tmp_buff2);
+ return MPI_SUCCESS;
+}
