/* -*- Mode: C; c-basic-offset:2 ; indent-tabs-mode:nil -*- */ /* * Copyright (c) 2006 The Trustees of Indiana University and Indiana * University Research and Technology * Corporation. All rights reserved. * Copyright (c) 2006 The Technical University of Chemnitz. All * rights reserved. * Copyright (c) 2013-2017 Los Alamos National Security, LLC. All rights * reserved. * Copyright (c) 2014-2018 Research Organization for Information Science * and Technology (RIST). All rights reserved. * Copyright (c) 2017-2022 IBM Corporation. All rights reserved. * Copyright (c) 2018 FUJITSU LIMITED. All rights reserved. * $COPYRIGHT$ * * Additional copyrights may follow * * Author(s): Torsten Hoefler * */ #include "nbc_internal.h" #include "ompi/communicator/communicator.h" #include "ompi/datatype/ompi_datatype.h" #include "ompi/op/op.h" #include "opal/util/bit_ops.h" #include static inline int allred_sched_diss(int rank, int p, int count, MPI_Datatype datatype, ptrdiff_t gap, const void *sendbuf, void *recvbuf, MPI_Op op, char inplace, NBC_Schedule *schedule, void *tmpbuf); static inline int allred_sched_recursivedoubling(int rank, int p, const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, ptrdiff_t gap, MPI_Op op, char inplace, NBC_Schedule *schedule, void *tmpbuf); static inline int allred_sched_ring(int rank, int p, int count, MPI_Datatype datatype, const void *sendbuf, void *recvbuf, MPI_Op op, int size, int ext, NBC_Schedule *schedule, void *tmpbuf); static inline int allred_sched_linear(int rank, int p, const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, ptrdiff_t gap, MPI_Op op, int ext, int size, NBC_Schedule *schedule, void *tmpbuf); static inline int allred_sched_redscat_allgather( int rank, int comm_size, int count, MPI_Datatype datatype, ptrdiff_t gap, const void *sbuf, void *rbuf, MPI_Op op, char inplace, NBC_Schedule *schedule, void *tmpbuf, struct ompi_communicator_t *comm); #ifdef NBC_CACHE_SCHEDULE /* tree comparison function for schedule cache */ int NBC_Allreduce_args_compare(NBC_Allreduce_args *a, NBC_Allreduce_args *b, void *param) { if ((a->sendbuf == b->sendbuf) && (a->recvbuf == b->recvbuf) && (a->count == b->count) && (a->datatype == b->datatype) && (a->op == b->op)) { return 0; } if( a->sendbuf < b->sendbuf ) { return -1; } return 1; } #endif static int nbc_allreduce_init(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, struct ompi_communicator_t *comm, ompi_request_t ** request, mca_coll_base_module_t *module, bool persistent) { int rank, p, res; ptrdiff_t ext, lb; NBC_Schedule *schedule; size_t size; #ifdef NBC_CACHE_SCHEDULE NBC_Allreduce_args *args, *found, search; #endif enum { NBC_ARED_BINOMIAL, NBC_ARED_RING, NBC_ARED_REDSCAT_ALLGATHER, NBC_ARED_RDBL } alg; char inplace; void *tmpbuf = NULL; ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module; ptrdiff_t span, gap; NBC_IN_PLACE(sendbuf, recvbuf, inplace); rank = ompi_comm_rank (comm); p = ompi_comm_size (comm); res = ompi_datatype_get_extent(datatype, &lb, &ext); if (OMPI_SUCCESS != res) { NBC_Error ("MPI Error in ompi_datatype_type_extent() (%i)", res); return res; } res = ompi_datatype_type_size (datatype, &size); if (OMPI_SUCCESS != res) { NBC_Error ("MPI Error in ompi_datatype_type_size() (%i)", res); return res; } if (1 == p && (!persistent || inplace)) { if (!inplace) { /* for a single node - copy data to receivebuf */ res = NBC_Copy(sendbuf, count, datatype, recvbuf, count, datatype, comm); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } } return nbc_get_noop_request(persistent, request); } span = opal_datatype_span(&datatype->super, count, &gap); tmpbuf = malloc (span); if (OPAL_UNLIKELY(NULL == tmpbuf)) { return OMPI_ERR_OUT_OF_RESOURCE; } alg = NBC_ARED_RING; /* default generic selection */ /* algorithm selection */ int nprocs_pof2 = opal_next_poweroftwo(p) >> 1; if (libnbc_iallreduce_algorithm == 0) { if(p < 4 || size*count < 65536 || !ompi_op_is_commute(op) || inplace) { alg = NBC_ARED_BINOMIAL; } else if (count >= nprocs_pof2 && ompi_op_is_commute(op)) { alg = NBC_ARED_REDSCAT_ALLGATHER; } } else { if (libnbc_iallreduce_algorithm == 1) alg = NBC_ARED_RING; else if (libnbc_iallreduce_algorithm == 2) alg = NBC_ARED_BINOMIAL; else if (libnbc_iallreduce_algorithm == 3 && count >= nprocs_pof2 && ompi_op_is_commute(op)) alg = NBC_ARED_REDSCAT_ALLGATHER; else if (libnbc_iallreduce_algorithm == 4) alg = NBC_ARED_RDBL; } #ifdef NBC_CACHE_SCHEDULE /* search schedule in communicator specific tree */ search.sendbuf = sendbuf; search.recvbuf = recvbuf; search.count = count; search.datatype = datatype; search.op = op; found = (NBC_Allreduce_args *) hb_tree_search ((hb_tree *) libnbc_module->NBC_Dict[NBC_ALLREDUCE], &search); if (NULL == found) { #endif schedule = OBJ_NEW(NBC_Schedule); if (NULL == schedule) { free(tmpbuf); return OMPI_ERR_OUT_OF_RESOURCE; } if (p == 1) { res = NBC_Sched_copy((void *)sendbuf, false, count, datatype, recvbuf, false, count, datatype, schedule, false); } else { switch(alg) { case NBC_ARED_BINOMIAL: res = allred_sched_diss(rank, p, count, datatype, gap, sendbuf, recvbuf, op, inplace, schedule, tmpbuf); break; case NBC_ARED_REDSCAT_ALLGATHER: res = allred_sched_redscat_allgather(rank, p, count, datatype, gap, sendbuf, recvbuf, op, inplace, schedule, tmpbuf, comm); break; case NBC_ARED_RING: res = allred_sched_ring(rank, p, count, datatype, sendbuf, recvbuf, op, size, ext, schedule, tmpbuf); break; case NBC_ARED_RDBL: res = allred_sched_recursivedoubling(rank, p, sendbuf, recvbuf, count, datatype, gap, op, inplace, schedule, tmpbuf); break; } } if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { OBJ_RELEASE(schedule); free(tmpbuf); return res; } res = NBC_Sched_commit(schedule); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { OBJ_RELEASE(schedule); free(tmpbuf); return res; } #ifdef NBC_CACHE_SCHEDULE /* save schedule to tree */ args = (NBC_Allreduce_args *) malloc (sizeof(args)); if (NULL != args) { args->sendbuf = sendbuf; args->recvbuf = recvbuf; args->count = count; args->datatype = datatype; args->op = op; args->schedule = schedule; res = hb_tree_insert ((hb_tree *) libnbc_module->NBC_Dict[NBC_ALLREDUCE], args, args, 0); if (0 == res) { OBJ_RETAIN(schedule); /* increase number of elements for A2A */ if (++libnbc_module->NBC_Dict_size[NBC_ALLREDUCE] > NBC_SCHED_DICT_UPPER) { NBC_SchedCache_dictwipe ((hb_tree *) libnbc_module->NBC_Dict[NBC_ALLREDUCE], &libnbc_module->NBC_Dict_size[NBC_ALLREDUCE]); } } else { NBC_Error("error in dict_insert() (%i)", res); free (args); } } } else { /* found schedule */ schedule = found->schedule; OBJ_RETAIN(schedule); } #endif res = NBC_Schedule_request (schedule, comm, libnbc_module, persistent, request, tmpbuf); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { OBJ_RELEASE(schedule); free(tmpbuf); return res; } return OMPI_SUCCESS; } int ompi_coll_libnbc_iallreduce(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, struct ompi_communicator_t *comm, ompi_request_t ** request, mca_coll_base_module_t *module) { int res = nbc_allreduce_init(sendbuf, recvbuf, count, datatype, op, comm, request, module, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } res = NBC_Start(*(ompi_coll_libnbc_request_t **)request); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { NBC_Return_handle (*(ompi_coll_libnbc_request_t **)request); *request = &ompi_request_null.request; return res; } return OMPI_SUCCESS; } static int nbc_allreduce_inter_init(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, struct ompi_communicator_t *comm, ompi_request_t ** request, mca_coll_base_module_t *module, bool persistent) { int rank, res, rsize; size_t size; MPI_Aint ext; NBC_Schedule *schedule; void *tmpbuf = NULL; ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module; ptrdiff_t span, gap; rank = ompi_comm_rank (comm); rsize = ompi_comm_remote_size (comm); res = ompi_datatype_type_extent(datatype, &ext); if (MPI_SUCCESS != res) { NBC_Error("MPI Error in ompi_datatype_type_extent() (%i)", res); return res; } res = ompi_datatype_type_size(datatype, &size); if (MPI_SUCCESS != res) { NBC_Error("MPI Error in ompi_datatype_type_size() (%i)", res); return res; } span = opal_datatype_span(&datatype->super, count, &gap); tmpbuf = malloc (span); if (OPAL_UNLIKELY(NULL == tmpbuf)) { return OMPI_ERR_OUT_OF_RESOURCE; } schedule = OBJ_NEW(NBC_Schedule); if (OPAL_UNLIKELY(NULL == schedule)) { free(tmpbuf); return OMPI_ERR_OUT_OF_RESOURCE; } res = allred_sched_linear (rank, rsize, sendbuf, recvbuf, count, datatype, gap, op, ext, size, schedule, tmpbuf); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { OBJ_RELEASE(schedule); free(tmpbuf); return res; } res = NBC_Sched_commit(schedule); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { OBJ_RELEASE(schedule); free(tmpbuf); return res; } res = NBC_Schedule_request(schedule, comm, libnbc_module, persistent, request, tmpbuf); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { OBJ_RELEASE(schedule); free(tmpbuf); return res; } return OMPI_SUCCESS; } int ompi_coll_libnbc_iallreduce_inter(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, struct ompi_communicator_t *comm, ompi_request_t ** request, mca_coll_base_module_t *module) { int res = nbc_allreduce_inter_init(sendbuf, recvbuf, count, datatype, op, comm, request, module, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } res = NBC_Start(*(ompi_coll_libnbc_request_t **)request); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { NBC_Return_handle (*(ompi_coll_libnbc_request_t **)request); *request = &ompi_request_null.request; return res; } return OMPI_SUCCESS; } /* binomial allreduce (binomial tree up and binomial bcast down) * working principle: * - each node gets a virtual rank vrank * - the 'root' node get vrank 0 * - node 0 gets the vrank of the 'root' * - all other ranks stay identical (they do not matter) * * Algorithm: * pairwise exchange * round r: * grp = rank % 2^r * if grp == 0: receive from rank + 2^(r-1) if it exists and reduce value * if grp == 1: send to rank - 2^(r-1) and exit function * * do this for R=log_2(p) rounds * followed by a Bcast: * Algorithm: * - each node with vrank > 2^r and vrank < 2^r+1 receives from node * vrank - 2^r (vrank=1 receives from 0, vrank 0 receives never) * - each node sends each round r to node vrank + 2^r * - a node stops to send if 2^r > commsize * */ #define RANK2VRANK(rank, vrank, root) \ { \ vrank = rank; \ if (rank == 0) vrank = root; \ if (rank == root) vrank = 0; \ } #define VRANK2RANK(rank, vrank, root) \ { \ rank = vrank; \ if (vrank == 0) rank = root; \ if (vrank == root) rank = 0; \ } static inline int allred_sched_diss(int rank, int p, int count, MPI_Datatype datatype, ptrdiff_t gap, const void *sendbuf, void *recvbuf, MPI_Op op, char inplace, NBC_Schedule *schedule, void *tmpbuf) { int root, vrank, maxr, vpeer, peer, res; char *rbuf, *lbuf, *buf; int tmprbuf, tmplbuf; root = 0; /* this makes the code for ireduce and iallreduce nearly identical - could be changed to improve performance */ RANK2VRANK(rank, vrank, root); maxr = ceil_of_log2(p); /* ensure the result ends up in recvbuf on vrank 0 */ if (0 == (maxr%2)) { rbuf = (void *)(-gap); tmprbuf = true; lbuf = recvbuf; tmplbuf = false; } else { lbuf = (void *)(-gap); tmplbuf = true; rbuf = recvbuf; tmprbuf = false; if (inplace) { res = NBC_Sched_copy(rbuf, false, count, datatype, ((char *)tmpbuf) - gap, false, count, datatype, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } } } for (int r = 1, firstred = 1 ; r <= maxr ; ++r) { if ((vrank % (1 << r)) == 0) { /* we have to receive this round */ vpeer = vrank + (1 << (r - 1)); VRANK2RANK(peer, vpeer, root) if (peer < p) { /* we have to wait until we have the data */ res = NBC_Sched_recv (rbuf, tmprbuf, count, datatype, peer, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } /* this cannot be done until tmpbuf is unused :-( so barrier after the op */ if (firstred && !inplace) { /* perform the reduce with the senbuf */ res = NBC_Sched_op (sendbuf, false, rbuf, tmprbuf, count, datatype, op, schedule, true); firstred = 0; } else { /* perform the reduce in my local buffer */ res = NBC_Sched_op (lbuf, tmplbuf, rbuf, tmprbuf, count, datatype, op, schedule, true); } if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } /* swap left and right buffers */ buf = rbuf; rbuf = lbuf ; lbuf = buf; tmprbuf ^= 1; tmplbuf ^= 1; } } else { /* we have to send this round */ vpeer = vrank - (1 << (r - 1)); VRANK2RANK(peer, vpeer, root) if (firstred && !inplace) { /* we have to use the sendbuf in the first round .. */ res = NBC_Sched_send (sendbuf, false, count, datatype, peer, schedule, false); } else { /* and the recvbuf in all remaining rounds */ res = NBC_Sched_send (lbuf, tmplbuf, count, datatype, peer, schedule, false); } if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } /* leave the game */ break; } } /* this is the Bcast part - copied with minor changes from nbc_ibcast.c * changed: buffer -> recvbuf */ RANK2VRANK(rank, vrank, root); /* receive from the right hosts */ if (vrank != 0) { for (int r = 0; r < maxr ; ++r) { if ((vrank >= (1 << r)) && (vrank < (1 << (r + 1)))) { VRANK2RANK(peer, vrank - (1 << r), root); res = NBC_Sched_recv (recvbuf, false, count, datatype, peer, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } } } res = NBC_Sched_barrier (schedule); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } } if (0 == vrank) assert(lbuf == recvbuf); /* now send to the right hosts */ for (int r = 0; r < maxr; ++r) { if (((vrank + (1 << r) < p) && (vrank < (1 << r))) || (vrank == 0)) { VRANK2RANK(peer, vrank + (1 << r), root); res = NBC_Sched_send (recvbuf, false, count, datatype, peer, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } } } /* end of the bcast */ return OMPI_SUCCESS; } /* * allred_sched_recursivedoubling * * Function: Recursive doubling algorithm for iallreduce operation * * Description: Implements recursive doubling algorithm for iallreduce. * The algorithm preserves order of operations so it can * be used both by commutative and non-commutative operations. * Schedule length: O(\log(p)) * Memory requirements: * Each process requires a temporary buffer: count * typesize = O(count) * * Example on 7 nodes: * Initial state * # 0 1 2 3 4 5 6 * [0] [1] [2] [3] [4] [5] [6] * Initial adjustment step for non-power of two nodes. * old rank 1 3 5 6 * new rank 0 1 2 3 * [0+1] [2+3] [4+5] [6] * Step 1 * old rank 1 3 5 6 * new rank 0 1 2 3 * [0+1+] [0+1+] [4+5+] [4+5+] * [2+3+] [2+3+] [6 ] [6 ] * Step 2 * old rank 1 3 5 6 * new rank 0 1 2 3 * [0+1+] [0+1+] [0+1+] [0+1+] * [2+3+] [2+3+] [2+3+] [2+3+] * [4+5+] [4+5+] [4+5+] [4+5+] * [6 ] [6 ] [6 ] [6 ] * Final adjustment step for non-power of two nodes * # 0 1 2 3 4 5 6 * [0+1+] [0+1+] [0+1+] [0+1+] [0+1+] [0+1+] [0+1+] * [2+3+] [2+3+] [2+3+] [2+3+] [2+3+] [2+3+] [2+3+] * [4+5+] [4+5+] [4+5+] [4+5+] [4+5+] [4+5+] [4+5+] * [6 ] [6 ] [6 ] [6 ] [6 ] [6 ] [6 ] * */ static inline int allred_sched_recursivedoubling(int rank, int p, const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, ptrdiff_t gap, MPI_Op op, char inplace, NBC_Schedule *schedule, void *tmpbuf) { int res, pof2, nprocs_rem, vrank; char *tmpsend = NULL, *tmprecv = NULL, *tmpswap = NULL; tmpsend = (char*) tmpbuf - gap; tmprecv = (char*) recvbuf; if (inplace) { res = NBC_Sched_copy(recvbuf, false, count, datatype, tmpsend, false, count, datatype, schedule, true); } else { res = NBC_Sched_copy((void *)sendbuf, false, count, datatype, tmpsend, false, count, datatype, schedule, true); } if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } /* Get nearest power of two less than or equal to comm size */ pof2 = opal_next_poweroftwo(p) >> 1; /* Handle non-power-of-two case: - Even ranks less than 2 * nprocs_rem send their data to (rank + 1), and sets new rank to -1. - Odd ranks less than 2 * nprocs_rem receive data from (rank - 1), apply appropriate operation, and set new rank to rank/2 - Everyone else sets rank to rank - nprocs_rem */ nprocs_rem = p - pof2; if (rank < 2 * nprocs_rem) { if (0 == rank % 2) { /* Even */ res = NBC_Sched_send(tmpsend, false, count, datatype, rank + 1, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } vrank = -1; } else { /* Odd */ res = NBC_Sched_recv(tmprecv, false, count, datatype, rank - 1, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } /* tmpsend = tmprecv (op) tmpsend */ res = NBC_Sched_op(tmprecv, false, tmpsend, false, count, datatype, op, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } vrank = rank >> 1; } } else { vrank = rank - nprocs_rem; } /* Communication/Computation loop - Exchange message with remote node. - Perform appropriate operation taking in account order of operations: result = value (op) result */ if (0 <= vrank) { for (int distance = 1; distance < pof2; distance <<= 1) { int remote = vrank ^ distance; /* Find real rank of remote node */ if (remote < nprocs_rem) { remote = remote * 2 + 1; } else { remote += nprocs_rem; } /* Exchange the data */ res = NBC_Sched_send(tmpsend, false, count, datatype, remote, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } res = NBC_Sched_recv(tmprecv, false, count, datatype, remote, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } /* Apply operation */ if (rank < remote) { /* tmprecv = tmpsend (op) tmprecv */ res = NBC_Sched_op(tmpsend, false, tmprecv, false, count, datatype, op, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } /* Swap tmpsend and tmprecv buffers */ tmpswap = tmprecv; tmprecv = tmpsend; tmpsend = tmpswap; } else { /* tmpsend = tmprecv (op) tmpsend */ res = NBC_Sched_op(tmprecv, false, tmpsend, false, count, datatype, op, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } } } } /* Handle non-power-of-two case: - Even ranks less than 2 * nprocs_rem receive result from (rank + 1) - Odd ranks less than 2 * nprocs_rem send result from tmpsend to (rank - 1) */ if (rank < 2 * nprocs_rem) { if (0 == rank % 2) { /* Even */ res = NBC_Sched_recv(recvbuf, false, count, datatype, rank + 1, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } tmpsend = (char *)recvbuf; } else { /* Odd */ res = NBC_Sched_send(tmpsend, false, count, datatype, rank - 1, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } } } /* Copy result back into recvbuf */ if (tmpsend != recvbuf) { res = NBC_Sched_copy(tmpsend, false, count, datatype, recvbuf, false, count, datatype, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } } return OMPI_SUCCESS; } static inline int allred_sched_ring(int r, int p, int count, MPI_Datatype datatype, const void *sendbuf, void *recvbuf, MPI_Op op, int size, int ext, NBC_Schedule *schedule, void *tmpbuf) { int segsize, *segsizes, *segoffsets; /* segment sizes and offsets per segment (number of segments == number of nodes */ int speer, rpeer; /* send and recv peers */ int res = OMPI_SUCCESS; if (0 == count) { return OMPI_SUCCESS; } segsizes = (int *) malloc((2 * p + 1 ) *sizeof (int)); if (NULL == segsizes) { return OMPI_ERR_OUT_OF_RESOURCE; } segoffsets = segsizes + p; segsize = count / p; /* size of the segments across the last ranks. The remainder will be evenly distributed across the smaller ranks */ segoffsets[0] = 0; for (int i = 0, mycount = count % p; i < p ; ++i) { segsizes[i] = segsize; if( mycount > 0 ) { /* We have extra segments to distribute */ segsizes[i]++; mycount--; } segoffsets[i+1] = segoffsets[i] + segsizes[i]; } /* reduce peers */ speer = (r + 1) % p; rpeer = (r - 1 + p) % p; /* + -> reduced this round * / -> sum (reduced in a previous step) * * *** round 0 *** * 0 1 2 * * 00 10 20 0: [1] -> 1 * 01 11 21 1: [2] -> 2 * 02 12 22 2: [0] -> 0 --> send element (r+1)%p to node (r+1)%p * * *** round 1 *** * 0 1 2 * * 00+20 10 20 0: red(0), [0] -> 1 * 01 11+01 21 1: red(1), [1] -> 2 * 02 12 22+12 2: red(2), [2] -> 0 --> reduce and send element (r+0)%p to node (r+1)%p * * *** round 2 *** * 0 1 2 * * 00/20 all 20 0: red(2), [2] -> 1 * 01 11/01 all 1: red(0), [0] -> 2 * all 12 22/12 2: red(1), [1] -> 0 --> reduce and send (r-1)%p to node (r+1)%p * * *** round 3 *** * 0 1 2 * * 00/20 all all 0: [1] -> 1 * all 11/01 all 1: [2] -> 2 * all all 22/12 2: [0] -> 0 --> send element (r-2)%p to node (r+1)%p * * *** round 4 *** * 0 1 2 * * all all all 0: done * all all all 1: done * all all all 2: done * * -> 4 * *** round 0 *** * 0 1 2 3 * * 00 10 20 30 0: [1] -> 1 * 01 11 21 31 1: [2] -> 2 * 02 12 22 32 2: [3] -> 3 * 03 13 23 33 3: [0] -> 0 --> send element (r+1)%p to node (r+1)%p * * *** round 1 *** * 0 1 2 3 * * 00+30 10 20 30 0: red(0), [0] -> 1 * 01 11+01 21 31 1: red(1), [1] -> 2 * 02 12 22+12 32 2: red(2), [2] -> 3 * 03 13 23 33+23 3: red(3), [3] -> 0 --> reduce and send element (r+0)%p to node (r+1)%p * * *** round 2 *** * 0 1 2 3 * * 00/30 10+00/30 20 30 0: red(3), [3] -> 1 * 01 11/01 21+11/01 31 1: red(0), [0] -> 2 * 02 12 22/12 32+22/12 2: red(1), [1] -> 3 * 03+33/23 13 23 33/23 3: red(2), [2] -> 0 --> reduce and send (r-1)%p to node (r+1)%p * * *** round 3 *** * 0 1 2 3 * * 00/30 10/00/30 all 30 0: red(2), [2] -> 1 * 01 11/01 21/11/01 all 1: red(3), [3] -> 2 * all 12 22/12 32/22/12 2: red(0), [0] -> 3 * 03/33/23 all 23 33/23 3: red(1), [1] -> 0 --> reduce and send (r-2)%p to node (r+1)%p * * *** round 4 *** * 0 1 2 3 * * 00/30 10/00/30 all all 0: [1] -> 1 * all 11/01 21/11/01 all 1: [2] -> 2 * all all 22/12 32/22/12 2: [3] -> 3 * 03/33/23 all all 33/23 3: [0] -> 0 --> receive and send element (r+1)%p to node (r+1)%p * * *** round 5 *** * 0 1 2 3 * * all 10/00/30 all all 0: [0] -> 1 * all all 21/11/01 all 1: [1] -> 2 * all all all 32/22/12 2: [3] -> 3 * 03/33/23 all all all 3: [4] -> 4 --> receive and send element (r-0)%p to node (r+1)%p * * *** round 6 *** * 0 1 2 3 * * all all all all * all all all all * all all all all * all all all all receive element (r-1)%p * * 2p-2 rounds ... every node does p-1 reductions and p-1 sends * */ /* first p-1 rounds are reductions */ for (int round = 0 ; round < p - 1 ; ++round) { int selement = (r+1-round + 2*p /*2*p avoids negative mod*/)%p; /* the element I am sending */ size_t soffset = segoffsets[selement]*(size_t)ext; int relement = (r-round + 2*p /*2*p avoids negative mod*/)%p; /* the element that I receive from my neighbor */ size_t roffset = segoffsets[relement]*(size_t)ext; /* first message come out of sendbuf */ if (round == 0) { res = NBC_Sched_send ((char *) sendbuf + soffset, false, segsizes[selement], datatype, speer, schedule, false); } else { res = NBC_Sched_send ((char *) recvbuf + soffset, false, segsizes[selement], datatype, speer, schedule, false); } if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto free_and_return; } if( recvbuf != sendbuf ) { /* check for MPI_IN_PLACE */ res = NBC_Sched_recv ((char *) recvbuf + roffset, false, segsizes[relement], datatype, rpeer, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto free_and_return; } res = NBC_Sched_op ((char *) sendbuf + roffset, false, (char *) recvbuf + roffset, false, segsizes[relement], datatype, op, schedule, true); } else { res = NBC_Sched_recv ((char *) tmpbuf, false, segsizes[relement], datatype, rpeer, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto free_and_return; } res = NBC_Sched_op ((char *) tmpbuf, false, (char *) recvbuf + roffset, false, segsizes[relement], datatype, op, schedule, true); } if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto free_and_return; } } for (int round = p - 1 ; round < 2 * p - 2 ; ++round) { int selement = (r+1-round + 2*p /*2*p avoids negative mod*/)%p; /* the element I am sending */ size_t soffset = segoffsets[selement]*(size_t)ext; int relement = (r-round + 2*p /*2*p avoids negative mod*/)%p; /* the element that I receive from my neighbor */ size_t roffset = segoffsets[relement]*(size_t)ext; res = NBC_Sched_send ((char *) recvbuf + soffset, false, segsizes[selement], datatype, speer, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { break; } res = NBC_Sched_recv ((char *) recvbuf + roffset, false, segsizes[relement], datatype, rpeer, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { break; } } free_and_return: free (segsizes); return res; } static inline int allred_sched_linear(int rank, int rsize, const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, ptrdiff_t gap, MPI_Op op, int ext, int size, NBC_Schedule *schedule, void *tmpbuf) { int res; if (0 == count) { return OMPI_SUCCESS; } /* send my data to the remote root */ res = NBC_Sched_send (sendbuf, false, count, datatype, 0, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } /* recv my data to the remote root */ if (0 != rank || 1 ==(rsize%2)) { res = NBC_Sched_recv (recvbuf, false, count, datatype, 0, schedule, false); } else { res = NBC_Sched_recv ((void *)(-gap), true, count, datatype, 0, schedule, false); } if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } if (0 == rank) { char *rbuf, *lbuf, *buf; int tmprbuf, tmplbuf; res = NBC_Sched_barrier (schedule); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } /* ensure the result ends up in recvbuf */ if (0 == (rsize%2)) { lbuf = (void *)(-gap); tmplbuf = true; rbuf = recvbuf; tmprbuf = false; } else { rbuf = (void *)(-gap); tmprbuf = true; lbuf = recvbuf; tmplbuf = false; } /* get data from remote peers and reduce */ for (int rpeer = 1 ; rpeer < rsize ; ++rpeer) { res = NBC_Sched_recv (rbuf, tmprbuf, count, datatype, rpeer, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } res = NBC_Sched_op (lbuf, tmplbuf, rbuf, tmprbuf, count, datatype, op, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } /* swap left and right buffers */ buf = rbuf; rbuf = lbuf ; lbuf = buf; tmprbuf ^= 1; tmplbuf ^= 1; } /* exchange our result with the remote root (each root will broadcast to the other's peers) */ res = NBC_Sched_recv ((void *)(-gap), true, count, datatype, 0, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } /* wait for data from remote root */ res = NBC_Sched_send (recvbuf, false, count, datatype, 0, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } /* broadcast the result to all remote peers */ for (int rpeer = 1 ; rpeer < rsize ; ++rpeer) { res = NBC_Sched_send ((void *)(-gap), true, count, datatype, rpeer, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } } } return OMPI_SUCCESS; } /* * allred_sched_redscat_allgather: * * Description: an implementation of Rabenseifner's Allreduce algorithm [1, 2]. * [1] Rajeev Thakur, Rolf Rabenseifner and William Gropp. * Optimization of Collective Communication Operations in MPICH // * The Int. Journal of High Performance Computing Applications. Vol 19, * Issue 1, pp. 49--66. * [2] http://www.hlrs.de/mpi/myreduce.html. * * This algorithm is a combination of a reduce-scatter implemented with * recursive vector halving and recursive distance doubling, followed either * by an allgather implemented with recursive doubling. * * Step 1. If the number of processes is not a power of two, reduce it to * the nearest lower power of two (p' = 2^{\floor{\log_2 p}}) * by removing r = p - p' extra processes as follows. In the first 2r processes * (ranks 0 to 2r - 1), all the even ranks send the second half of the input * vector to their right neighbor (rank + 1), and all the odd ranks send * the first half of the input vector to their left neighbor (rank - 1). * The even ranks compute the reduction on the first half of the vector and * the odd ranks compute the reduction on the second half. The odd ranks then * send the result to their left neighbors (the even ranks). As a result, * the even ranks among the first 2r processes now contain the reduction with * the input vector on their right neighbors (the odd ranks). These odd ranks * do not participate in the rest of the algorithm, which leaves behind * a power-of-two number of processes. The first r even-ranked processes and * the last p - 2r processes are now renumbered from 0 to p' - 1. * * Step 2. The remaining processes now perform a reduce-scatter by using * recursive vector halving and recursive distance doubling. The even-ranked * processes send the second half of their buffer to rank + 1 and the odd-ranked * processes send the first half of their buffer to rank - 1. All processes * then compute the reduction between the local buffer and the received buffer. * In the next log_2(p') - 1 steps, the buffers are recursively halved, and the * distance is doubled. At the end, each of the p' processes has 1 / p' of the * total reduction result. * * Step 3. An allgather is performed by using recursive vector doubling and * distance halving. All exchanges are executed in reverse order relative * to recursive doubling on previous step. If the number of processes is not * a power of two, the total result vector must be sent to the r processes * that were removed in the first step. * * Limitations: * count >= 2^{\floor{\log_2 p}} * commutative operations only * intra-communicators only * * Memory requirements (per process): * count * typesize + 4 * \log_2(p) * sizeof(int) = O(count) * * Schedule length (rounds): O(\log(p)) */ static inline int allred_sched_redscat_allgather( int rank, int comm_size, int count, MPI_Datatype datatype, ptrdiff_t gap, const void *sbuf, void *rbuf, MPI_Op op, char inplace, NBC_Schedule *schedule, void *tmpbuf, struct ompi_communicator_t *comm) { int res = OMPI_SUCCESS; int *rindex = NULL, *rcount = NULL, *sindex = NULL, *scount = NULL; /* Find nearest power-of-two less than or equal to comm_size */ int nsteps = opal_hibit(comm_size, comm->c_cube_dim + 1); /* ilog2(comm_size) */ int nprocs_pof2 = 1 << nsteps; /* flp2(comm_size) */ if (!inplace) { res = NBC_Sched_copy((char *)sbuf, false, count, datatype, rbuf, false, count, datatype, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } } char *tmp_buf = (char *)tmpbuf - gap; ptrdiff_t lb, extent; ompi_datatype_get_extent(datatype, &lb, &extent); /* * Step 1. Reduce the number of processes to the nearest lower power of two * p' = 2^{\floor{\log_2 p}} by removing r = p - p' processes. * 1. In the first 2r processes (ranks 0 to 2r - 1), all the even ranks send * the second half of the input vector to their right neighbor (rank + 1) * and all the odd ranks send the first half of the input vector to their * left neighbor (rank - 1). * 2. All 2r processes compute the reduction on their half. * 3. The odd ranks then send the result to their left neighbors * (the even ranks). * * The even ranks (0 to 2r - 1) now contain the reduction with the input * vector on their right neighbors (the odd ranks). The first r even * processes and the p - 2r last processes are renumbered from * 0 to 2^{\floor{\log_2 p}} - 1. */ int vrank, step, wsize; int nprocs_rem = comm_size - nprocs_pof2; if (rank < 2 * nprocs_rem) { int count_lhalf = count / 2; int count_rhalf = count - count_lhalf; if (rank % 2 != 0) { /* * Odd process -- exchange with rank - 1 * Send the left half of the input vector to the left neighbor, * Recv the right half of the input vector from the left neighbor */ res = NBC_Sched_send(rbuf, false, count_lhalf, datatype, rank - 1, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } res = NBC_Sched_recv(tmp_buf + (ptrdiff_t)count_lhalf * extent, false, count_rhalf, datatype, rank - 1, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } res = NBC_Sched_op(tmp_buf + (ptrdiff_t)count_lhalf * extent, false, (char *)rbuf + (ptrdiff_t)count_lhalf * extent, false, count_rhalf, datatype, op, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } /* Send the right half to the left neighbor */ res = NBC_Sched_send((char *)rbuf + (ptrdiff_t)count_lhalf * extent, false, count_rhalf, datatype, rank - 1, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } /* This process does not participate in recursive doubling phase */ vrank = -1; } else { /* * Even process -- exchange with rank + 1 * Send the right half of the input vector to the right neighbor, * Recv the left half of the input vector from the right neighbor */ res = NBC_Sched_send((char *)rbuf + (ptrdiff_t)count_lhalf * extent, false, count_rhalf, datatype, rank + 1, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } res = NBC_Sched_recv(tmp_buf, false, count_lhalf, datatype, rank + 1, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } res = NBC_Sched_op(tmp_buf, false, rbuf, false, count_lhalf, datatype, op, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } /* Recv the right half from the right neighbor */ res = NBC_Sched_recv((char *)rbuf + (ptrdiff_t)count_lhalf * extent, false, count_rhalf, datatype, rank + 1, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } vrank = rank / 2; } } else { /* rank >= 2 * nprocs_rem */ vrank = rank - nprocs_rem; } /* * Step 2. Reduce-scatter implemented with recursive vector halving and * recursive distance doubling. We have p' = 2^{\floor{\log_2 p}} * power-of-two number of processes with new ranks (vrank) and result in rbuf. * * The even-ranked processes send the right half of their buffer to rank + 1 * and the odd-ranked processes send the left half of their buffer to * rank - 1. All processes then compute the reduction between the local * buffer and the received buffer. In the next \log_2(p') - 1 steps, the * buffers are recursively halved, and the distance is doubled. At the end, * each of the p' processes has 1 / p' of the total reduction result. */ rindex = malloc(sizeof(*rindex) * nsteps); sindex = malloc(sizeof(*sindex) * nsteps); rcount = malloc(sizeof(*rcount) * nsteps); scount = malloc(sizeof(*scount) * nsteps); if (NULL == rindex || NULL == sindex || NULL == rcount || NULL == scount) { res = OMPI_ERR_OUT_OF_RESOURCE; goto cleanup_and_return; } if (vrank != -1) { step = 0; wsize = count; sindex[0] = rindex[0] = 0; for (int mask = 1; mask < nprocs_pof2; mask <<= 1) { /* * On each iteration: rindex[step] = sindex[step] -- beginning of the * current window. Length of the current window is storded in wsize. */ int vdest = vrank ^ mask; /* Translate vdest virtual rank to real rank */ int dest = (vdest < nprocs_rem) ? vdest * 2 : vdest + nprocs_rem; if (rank < dest) { /* * Recv into the left half of the current window, send the right * half of the window to the peer (perform reduce on the left * half of the current window) */ rcount[step] = wsize / 2; scount[step] = wsize - rcount[step]; sindex[step] = rindex[step] + rcount[step]; } else { /* * Recv into the right half of the current window, send the left * half of the window to the peer (perform reduce on the right * half of the current window) */ scount[step] = wsize / 2; rcount[step] = wsize - scount[step]; rindex[step] = sindex[step] + scount[step]; } /* Send part of data from the rbuf, recv into the tmp_buf */ res = NBC_Sched_send((char *)rbuf + (ptrdiff_t)sindex[step] * extent, false, scount[step], datatype, dest, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } res = NBC_Sched_recv((char *)tmp_buf + (ptrdiff_t)rindex[step] * extent, false, rcount[step], datatype, dest, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } /* Local reduce: rbuf[] = tmp_buf[] rbuf[] */ res = NBC_Sched_op((char *)tmp_buf + (ptrdiff_t)rindex[step] * extent, false, (char *)rbuf + (ptrdiff_t)rindex[step] * extent, false, rcount[step], datatype, op, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } /* Move the current window to the received message */ if (step + 1 < nsteps) { rindex[step + 1] = rindex[step]; sindex[step + 1] = rindex[step]; wsize = rcount[step]; step++; } } /* * Assertion: each process has 1 / p' of the total reduction result: * rcount[nsteps - 1] elements in the rbuf[rindex[nsteps - 1], ...]. */ /* * Step 3. Allgather by the recursive doubling algorithm. * Each process has 1 / p' of the total reduction result: * rcount[nsteps - 1] elements in the rbuf[rindex[nsteps - 1], ...]. * All exchanges are executed in reverse order relative * to recursive doubling (previous step). */ step = nsteps - 1; for (int mask = nprocs_pof2 >> 1; mask > 0; mask >>= 1) { int vdest = vrank ^ mask; /* Translate vdest virtual rank to real rank */ int dest = (vdest < nprocs_rem) ? vdest * 2 : vdest + nprocs_rem; /* * Send rcount[step] elements from rbuf[rindex[step]...] * Recv scount[step] elements to rbuf[sindex[step]...] */ res = NBC_Sched_send((char *)rbuf + (ptrdiff_t)rindex[step] * extent, false, rcount[step], datatype, dest, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } res = NBC_Sched_recv((char *)rbuf + (ptrdiff_t)sindex[step] * extent, false, scount[step], datatype, dest, schedule, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } step--; } } /* * Step 4. Send total result to excluded odd ranks. */ if (rank < 2 * nprocs_rem) { if (rank % 2 != 0) { /* Odd process -- recv result from rank - 1 */ res = NBC_Sched_recv(rbuf, false, count, datatype, rank - 1, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } } else { /* Even process -- send result to rank + 1 */ res = NBC_Sched_send(rbuf, false, count, datatype, rank + 1, schedule, false); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; } } } cleanup_and_return: if (NULL != rindex) free(rindex); if (NULL != sindex) free(sindex); if (NULL != rcount) free(rcount); if (NULL != scount) free(scount); return res; } int ompi_coll_libnbc_allreduce_init(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, struct ompi_communicator_t *comm, MPI_Info info, ompi_request_t ** request, mca_coll_base_module_t *module) { int res = nbc_allreduce_init(sendbuf, recvbuf, count, datatype, op, comm, request, module, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } return OMPI_SUCCESS; } int ompi_coll_libnbc_allreduce_inter_init(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, struct ompi_communicator_t *comm, MPI_Info info, ompi_request_t ** request, mca_coll_base_module_t *module) { int res = nbc_allreduce_inter_init(sendbuf, recvbuf, count, datatype, op, comm, request, module, true); if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { return res; } return OMPI_SUCCESS; }