1
1
openmpi/ompi/mca/coll/libnbc/nbc_iallgather.c
Nathan Hjelm d42e0968b1 coll/libnbc: rewrite parts of libnbc
This commit rewrites parts of libnbc to fix issues identified by
coverity and myself. The changes are as follows:

 - libnbc function would return invalid error codes (internal to
   libnbc) to the mpi layer. These codes names are of the form
   NBC_. They do not match up with the error codes expected by the mpi
   layer. I purged the use of all these error codes with the exception
   of NBC_OK and NBC_CONTINUE in progress. These codes are used to
   identify when a request handle is complete.

 - Handles and schedules were leaked by all collective routines on
   error. A new routine was added to return a collective handle
   (NBC_Return_handle).

 - Temporary buffers containting in/out neighbors for neighborhood
   collectives were always leaked.

 - Neigborhood collectives contained code to handle MPI_IN_PLACE which
   is never a valid input for the send or receive buffer. Stipped this
   code out.

 - Files were inconsistently named. Most are nbc_isomething.c but one
   was named coll_libnbc_ireduce_scatter_block.c.

 - Made the NBC_Schedule "structure" and object so it can be
   retained/released. This may enable the use of schedule caching at a
   later time. More testing will be needed to ensure the caching code
   works. If it doesn't the code should be stripped out completely.

 - Added code to simply common case of scheduling send/recv +
   barrier.

 - Code cleanup for readability.

The code now passes the clang static analyzer.

Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2015-08-10 11:53:25 -06:00

232 строки
6.9 KiB
C

/* -*- 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) 2014 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* Copyright (c) 2015 Los Alamos National Security, LLC. All rights
* reserved.
*
* Author(s): Torsten Hoefler <htor@cs.indiana.edu>
*
*/
#include "nbc_internal.h"
#ifdef NBC_CACHE_SCHEDULE
/* tree comparison function for schedule cache */
int NBC_Allgather_args_compare(NBC_Allgather_args *a, NBC_Allgather_args *b, void *param) {
if ((a->sendbuf == b->sendbuf) &&
(a->sendcount == b->sendcount) &&
(a->sendtype == b->sendtype) &&
(a->recvbuf == b->recvbuf) &&
(a->recvcount == b->recvcount) &&
(a->recvtype == b->recvtype) ) {
return 0;
}
if( a->sendbuf < b->sendbuf ) {
return -1;
}
return 1;
}
#endif
/* simple linear MPI_Iallgather
* the algorithm uses p-1 rounds
* each node sends the packet it received last round (or has in round 0) to it's right neighbor (modulo p)
* each node receives from it's left (modulo p) neighbor */
int ompi_coll_libnbc_iallgather(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
MPI_Datatype recvtype, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module)
{
int rank, p, res;
MPI_Aint rcvext;
NBC_Schedule *schedule;
char *rbuf, *sbuf, inplace;
#ifdef NBC_CACHE_SCHEDULE
NBC_Allgather_args *args, *found, search;
#endif
NBC_Handle *handle;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
rank = ompi_comm_rank (comm);
p = ompi_comm_size (comm);
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) {
return res;
}
if (inplace) {
sendtype = recvtype;
sendcount = recvcount;
} else {
/* copy my data to receive buffer */
rbuf = (char *) recvbuf + rank * recvcount * rcvext;
res = NBC_Copy (sendbuf, sendcount, sendtype, rbuf, recvcount, recvtype, comm);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
}
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sendbuf = sendbuf;
search.sendcount = sendcount;
search.sendtype = sendtype;
search.recvbuf = recvbuf;
search.recvcount = recvcount;
search.recvtype = recvtype;
found = (NBC_Allgather_args *) hb_tree_search ((hb_tree*)libnbc_module->NBC_Dict[NBC_ALLGATHER], &search);
if (NULL == found) {
#endif
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
sbuf = (char *)recvbuf + rank * recvcount * rcvext;
/* do p-1 rounds */
for(int r = 0 ; r < p ; ++r) {
if(r != rank) {
/* recv from rank r */
rbuf = (char *)recvbuf + r * recvcount * rcvext;
res = NBC_Sched_recv (rbuf, false, recvcount, recvtype, r, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
/* send to rank r - not from the sendbuf to optimize MPI_IN_PLACE */
res = NBC_Sched_send (sbuf, false, sendcount, sendtype, r, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
}
}
res = NBC_Sched_commit(schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Allgather_args *) malloc (sizeof (args));
args->sendbuf = sendbuf;
args->sendcount = sendcount;
args->sendtype = sendtype;
args->recvbuf = recvbuf;
args->recvcount = recvcount;
args->recvtype = recvtype;
args->schedule = schedule;
res = hb_tree_insert ((hb_tree *) libnbc_module->NBC_Dict[NBC_ALLGATHER], args, args, 0);
if (res != 0) {
free (args);
} else {
OBJ_RETAIN(schedule);
}
/* increase number of elements for A2A */
if (++libnbc_module->NBC_Dict_size[NBC_ALLGATHER] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe ((hb_tree *) libnbc_module->NBC_Dict[NBC_ALLGATHER], &libnbc_module->NBC_Dict_size[NBC_ALLGATHER]);
}
} else {
/* found schedule */
schedule = found->schedule;
OBJ_RETAIN(schedule);
}
#endif
res = NBC_Init_handle (comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Start (handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OMPI_COLL_LIBNBC_REQUEST_RETURN(handle);
return res;
}
*request = (ompi_request_t *) handle;
return OMPI_SUCCESS;
}
int ompi_coll_libnbc_iallgather_inter(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
MPI_Datatype recvtype, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module)
{
int res, rsize;
MPI_Aint rcvext;
NBC_Schedule *schedule;
char *rbuf;
NBC_Handle *handle;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) {
NBC_Error ("MPI Error in MPI_Type_extent() (%i)", res);
return res;
}
rsize = ompi_comm_remote_size (comm);
/* set up schedule */
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* do rsize - 1 rounds */
for (int r = 0 ; r < rsize ; ++r) {
/* recv from rank r */
rbuf = (char *) recvbuf + r * recvcount * rcvext;
res = NBC_Sched_recv (rbuf, false, recvcount, recvtype, r, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
/* send to rank r */
res = NBC_Sched_send (sendbuf, false, sendcount, sendtype, r, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
}
res = NBC_Sched_commit (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Init_handle (comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Start (handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OMPI_COLL_LIBNBC_REQUEST_RETURN(handle);
return res;
}
*request = (ompi_request_t *) handle;
return OMPI_SUCCESS;
}