int MPIC_Sendrecv_replace(void *buf, int count, MPI_Datatype datatype,
int dest, int sendtag,
int source, int recvtag,
MPI_Comm comm, MPI_Status *status)
{
int mpi_errno = MPI_SUCCESS;
MPIR_Context_id_t context_id_offset;
MPID_Request *sreq;
MPID_Request *rreq;
void *tmpbuf = NULL;
int tmpbuf_size = 0;
int tmpbuf_count = 0;
MPID_Comm *comm_ptr;
MPIU_CHKLMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIC_SENDRECV_REPLACE);
#ifdef MPID_LOG_ARROWS
/* The logging macros log sendcount and recvcount */
int sendcount = count, recvcount = count;
#endif
MPIDI_PT2PT_FUNC_ENTER_BOTH(MPID_STATE_MPIC_SENDRECV_REPLACE);
MPID_Comm_get_ptr( comm, comm_ptr );
context_id_offset = (comm_ptr->comm_kind == MPID_INTRACOMM) ?
MPID_CONTEXT_INTRA_COLL : MPID_CONTEXT_INTER_COLL;
if (count > 0 && dest != MPI_PROC_NULL)
{
MPIR_Pack_size_impl(count, datatype, &tmpbuf_size);
MPIU_CHKLMEM_MALLOC(tmpbuf, void *, tmpbuf_size, mpi_errno, "temporary send buffer");
mpi_errno = MPIR_Pack_impl(buf, count, datatype, tmpbuf, tmpbuf_size, &tmpbuf_count);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
int MPIC_Sendrecv_replace(void *buf, int count, MPI_Datatype datatype,
int dest, int sendtag,
int source, int recvtag,
MPID_Comm *comm_ptr, MPI_Status *status, mpir_errflag_t *errflag)
{
int mpi_errno = MPI_SUCCESS;
MPI_Status mystatus;
MPIR_Context_id_t context_id_offset;
MPID_Request *sreq = NULL;
MPID_Request *rreq = NULL;
void *tmpbuf = NULL;
MPI_Aint tmpbuf_size = 0;
MPI_Aint tmpbuf_count = 0;
MPIU_CHKLMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIC_SENDRECV_REPLACE);
#ifdef MPID_LOG_ARROWS
/* The logging macros log sendcount and recvcount */
int sendcount = count, recvcount = count;
#endif
MPIDI_FUNC_ENTER(MPID_STATE_MPIC_SENDRECV_REPLACE);
MPIU_DBG_MSG_D(PT2PT, TYPICAL, "IN: errflag = %d", *errflag);
MPIU_ERR_CHKANDJUMP1((count < 0), mpi_errno, MPI_ERR_COUNT,
"**countneg", "**countneg %d", count);
if (status == MPI_STATUS_IGNORE) status = &mystatus;
switch(*errflag) {
case MPIR_ERR_NONE:
break;
case MPIR_ERR_PROC_FAILED:
MPIR_TAG_SET_PROC_FAILURE_BIT(sendtag);
default:
MPIR_TAG_SET_ERROR_BIT(sendtag);
}
context_id_offset = (comm_ptr->comm_kind == MPID_INTRACOMM) ?
MPID_CONTEXT_INTRA_COLL : MPID_CONTEXT_INTER_COLL;
if (count > 0 && dest != MPI_PROC_NULL) {
MPIR_Pack_size_impl(count, datatype, &tmpbuf_size);
MPIU_CHKLMEM_MALLOC(tmpbuf, void *, tmpbuf_size, mpi_errno, "temporary send buffer");
mpi_errno = MPIR_Pack_impl(buf, count, datatype, tmpbuf, tmpbuf_size, &tmpbuf_count);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
int MPIR_Alltoall_intra(
const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
int recvcount,
MPI_Datatype recvtype,
MPID_Comm *comm_ptr,
int *errflag )
{
int comm_size, i, j, pof2;
MPI_Aint sendtype_extent, recvtype_extent;
MPI_Aint recvtype_true_extent, recvbuf_extent, recvtype_true_lb;
int mpi_errno=MPI_SUCCESS, src, dst, rank, nbytes;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
int sendtype_size, pack_size, block, position, *displs, count;
MPI_Datatype newtype = MPI_DATATYPE_NULL;
void *tmp_buf;
MPI_Comm comm;
MPI_Request *reqarray;
MPI_Status *starray;
MPIU_CHKLMEM_DECL(6);
#ifdef MPIR_OLD_SHORT_ALLTOALL_ALG
MPI_Aint sendtype_true_extent, sendbuf_extent, sendtype_true_lb;
int k, p, curr_cnt, dst_tree_root, my_tree_root;
int last_recv_cnt, mask, tmp_mask, tree_root, nprocs_completed;
#endif
if (recvcount == 0) return MPI_SUCCESS;
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* Get extent of send and recv types */
MPID_Datatype_get_extent_macro(recvtype, recvtype_extent);
MPID_Datatype_get_extent_macro(sendtype, sendtype_extent);
MPID_Datatype_get_size_macro(sendtype, sendtype_size);
nbytes = sendtype_size * sendcount;
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
if (sendbuf == MPI_IN_PLACE) {
/* We use pair-wise sendrecv_replace in order to conserve memory usage,
* which is keeping with the spirit of the MPI-2.2 Standard. But
* because of this approach all processes must agree on the global
* schedule of sendrecv_replace operations to avoid deadlock.
*
* Note that this is not an especially efficient algorithm in terms of
* time and there will be multiple repeated malloc/free's rather than
* maintaining a single buffer across the whole loop. Something like
* MADRE is probably the best solution for the MPI_IN_PLACE scenario. */
for (i = 0; i < comm_size; ++i) {
/* start inner loop at i to avoid re-exchanging data */
for (j = i; j < comm_size; ++j) {
if (rank == i) {
/* also covers the (rank == i && rank == j) case */
mpi_errno = MPIC_Sendrecv_replace_ft(((char *)recvbuf + j*recvcount*recvtype_extent),
recvcount, recvtype,
j, MPIR_ALLTOALL_TAG,
j, MPIR_ALLTOALL_TAG,
comm, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
else if (rank == j) {
/* same as above with i/j args reversed */
mpi_errno = MPIC_Sendrecv_replace_ft(((char *)recvbuf + i*recvcount*recvtype_extent),
recvcount, recvtype,
i, MPIR_ALLTOALL_TAG,
i, MPIR_ALLTOALL_TAG,
comm, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
}
}
else if ((nbytes <= MPIR_PARAM_ALLTOALL_SHORT_MSG_SIZE) && (comm_size >= 8)) {
/* use the indexing algorithm by Jehoshua Bruck et al,
* IEEE TPDS, Nov. 97 */
/* allocate temporary buffer */
MPIR_Pack_size_impl(recvcount*comm_size, recvtype, &pack_size);
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, pack_size, mpi_errno, "tmp_buf");
/* Do Phase 1 of the algorithim. Shift the data blocks on process i
* upwards by a distance of i blocks. Store the result in recvbuf. */
mpi_errno = MPIR_Localcopy((char *) sendbuf +
rank*sendcount*sendtype_extent,
(comm_size - rank)*sendcount, sendtype, recvbuf,
(comm_size - rank)*recvcount, recvtype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
mpi_errno = MPIR_Localcopy(sendbuf, rank*sendcount, sendtype,
(char *) recvbuf +
(comm_size-rank)*recvcount*recvtype_extent,
rank*recvcount, recvtype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
/* Input data is now stored in recvbuf with datatype recvtype */
/* Now do Phase 2, the communication phase. It takes
ceiling(lg p) steps. In each step i, each process sends to rank+2^i
and receives from rank-2^i, and exchanges all data blocks
whose ith bit is 1. */
/* allocate displacements array for indexed datatype used in
communication */
MPIU_CHKLMEM_MALLOC(displs, int *, comm_size * sizeof(int), mpi_errno, "displs");
pof2 = 1;
while (pof2 < comm_size) {
dst = (rank + pof2) % comm_size;
src = (rank - pof2 + comm_size) % comm_size;
/* Exchange all data blocks whose ith bit is 1 */
/* Create an indexed datatype for the purpose */
count = 0;
for (block=1; block<comm_size; block++) {
if (block & pof2) {
displs[count] = block * recvcount;
count++;
}
}
mpi_errno = MPIR_Type_create_indexed_block_impl(count, recvcount,
displs, recvtype, &newtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIR_Type_commit_impl(&newtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
position = 0;
mpi_errno = MPIR_Pack_impl(recvbuf, 1, newtype, tmp_buf, pack_size, &position);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIC_Sendrecv_ft(tmp_buf, position, MPI_PACKED, dst,
MPIR_ALLTOALL_TAG, recvbuf, 1, newtype,
src, MPIR_ALLTOALL_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
MPIR_Type_free_impl(&newtype);
pof2 *= 2;
}
/* Rotate blocks in recvbuf upwards by (rank + 1) blocks. Need
* a temporary buffer of the same size as recvbuf. */
/* get true extent of recvtype */
MPIR_Type_get_true_extent_impl(recvtype, &recvtype_true_lb, &recvtype_true_extent);
recvbuf_extent = recvcount * comm_size *
(MPIR_MAX(recvtype_true_extent, recvtype_extent));
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, recvbuf_extent, mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - recvtype_true_lb);
mpi_errno = MPIR_Localcopy((char *) recvbuf + (rank+1)*recvcount*recvtype_extent,
(comm_size - rank - 1)*recvcount, recvtype, tmp_buf,
(comm_size - rank - 1)*recvcount, recvtype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
mpi_errno = MPIR_Localcopy(recvbuf, (rank+1)*recvcount, recvtype,
(char *) tmp_buf + (comm_size-rank-1)*recvcount*recvtype_extent,
(rank+1)*recvcount, recvtype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
/* Blocks are in the reverse order now (comm_size-1 to 0).
* Reorder them to (0 to comm_size-1) and store them in recvbuf. */
for (i=0; i<comm_size; i++){
mpi_errno = MPIR_Localcopy((char *) tmp_buf + i*recvcount*recvtype_extent,
recvcount, recvtype,
(char *) recvbuf + (comm_size-i-1)*recvcount*recvtype_extent,
recvcount, recvtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
#ifdef MPIR_OLD_SHORT_ALLTOALL_ALG
/* Short message. Use recursive doubling. Each process sends all
its data at each step along with all data it received in
previous steps. */
/* need to allocate temporary buffer of size
sendbuf_extent*comm_size */
/* get true extent of sendtype */
MPIR_Type_get_true_extent_impl(sendtype, &sendtype_true_lb, &sendtype_true_extent);
sendbuf_extent = sendcount * comm_size *
(MPIR_MAX(sendtype_true_extent, sendtype_extent));
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, sendbuf_extent*comm_size, mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - sendtype_true_lb);
/* copy local sendbuf into tmp_buf at location indexed by rank */
curr_cnt = sendcount*comm_size;
mpi_errno = MPIR_Localcopy(sendbuf, curr_cnt, sendtype,
((char *)tmp_buf + rank*sendbuf_extent),
curr_cnt, sendtype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno);}
mask = 0x1;
i = 0;
while (mask < comm_size) {
dst = rank ^ mask;
dst_tree_root = dst >> i;
dst_tree_root <<= i;
my_tree_root = rank >> i;
my_tree_root <<= i;
if (dst < comm_size) {
mpi_errno = MPIC_Sendrecv_ft(((char *)tmp_buf +
my_tree_root*sendbuf_extent),
curr_cnt, sendtype,
dst, MPIR_ALLTOALL_TAG,
((char *)tmp_buf +
dst_tree_root*sendbuf_extent),
sendbuf_extent*(comm_size-dst_tree_root),
sendtype, dst, MPIR_ALLTOALL_TAG,
comm, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
last_recv_cnt = 0;
} else
/* in case of non-power-of-two nodes, less data may be
received than specified */
MPIR_Get_count_impl(&status, sendtype, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
/* if some processes in this process's subtree in this step
did not have any destination process to communicate with
because of non-power-of-two, we need to send them the
result. We use a logarithmic recursive-halfing algorithm
for this. */
if (dst_tree_root + mask > comm_size) {
nprocs_completed = comm_size - my_tree_root - mask;
/* nprocs_completed is the number of processes in this
subtree that have all the data. Send data to others
in a tree fashion. First find root of current tree
that is being divided into two. k is the number of
least-significant bits in this process's rank that
must be zeroed out to find the rank of the root */
j = mask;
k = 0;
while (j) {
j >>= 1;
k++;
}
k--;
tmp_mask = mask >> 1;
while (tmp_mask) {
dst = rank ^ tmp_mask;
tree_root = rank >> k;
tree_root <<= k;
/* send only if this proc has data and destination
doesn't have data. at any step, multiple processes
can send if they have the data */
if ((dst > rank) &&
(rank < tree_root + nprocs_completed)
&& (dst >= tree_root + nprocs_completed)) {
/* send the data received in this step above */
mpi_errno = MPIC_Send_ft(((char *)tmp_buf +
dst_tree_root*sendbuf_extent),
last_recv_cnt, sendtype,
dst, MPIR_ALLTOALL_TAG,
comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* recv only if this proc. doesn't have data and sender
has data */
else if ((dst < rank) &&
(dst < tree_root + nprocs_completed) &&
(rank >= tree_root + nprocs_completed)) {
mpi_errno = MPIC_Recv_ft(((char *)tmp_buf +
dst_tree_root*sendbuf_extent),
sendbuf_extent*(comm_size-dst_tree_root),
sendtype,
dst, MPIR_ALLTOALL_TAG,
comm, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
last_recv_cnt = 0;
} else
MPIR_Get_count_impl(&status, sendtype, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
tmp_mask >>= 1;
k--;
}
}
mask <<= 1;
i++;
}
static int MPIR_Bcast_binomial_MV2(
void *buffer,
int count,
MPI_Datatype datatype,
int root,
MPID_Comm *comm_ptr,
int *errflag)
{
int rank, comm_size, src, dst;
int relative_rank, mask;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int nbytes=0;
int type_size, is_contig, is_homogeneous;
int position;
void *tmp_buf=NULL;
MPI_Comm comm;
MPID_Datatype *dtp;
MPIU_CHKLMEM_DECL(1);
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* If there is only one process, return */
if (comm_size == 1) goto fn_exit;
if (HANDLE_GET_KIND(datatype) == HANDLE_KIND_BUILTIN)
is_contig = 1;
else {
MPID_Datatype_get_ptr(datatype, dtp);
is_contig = dtp->is_contig;
}
is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
if (comm_ptr->is_hetero)
is_homogeneous = 0;
#endif
/* MPI_Type_size() might not give the accurate size of the packed
* datatype for heterogeneous systems (because of padding, encoding,
* etc). On the other hand, MPI_Pack_size() can become very
* expensive, depending on the implementation, especially for
* heterogeneous systems. We want to use MPI_Type_size() wherever
* possible, and MPI_Pack_size() in other places.
*/
if (is_homogeneous)
MPID_Datatype_get_size_macro(datatype, type_size);
else
MPIR_Pack_size_impl(1, datatype, &type_size);
nbytes = type_size * count;
if (!is_contig || !is_homogeneous)
{
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, nbytes, mpi_errno, "tmp_buf");
/* TODO: Pipeline the packing and communication */
position = 0;
if (rank == root) {
mpi_errno = MPIR_Pack_impl(buffer, count, datatype, tmp_buf, nbytes,
&position);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
/*@
MPI_Pack - Packs a datatype into contiguous memory
Input Parameters:
+ inbuf - input buffer start (choice)
. incount - number of input data items (non-negative integer)
. datatype - datatype of each input data item (handle)
. outsize - output buffer size, in bytes (non-negative integer)
- comm - communicator for packed message (handle)
Output Parameters:
. outbuf - output buffer start (choice)
Input/Output Parameters:
. position - current position in buffer, in bytes (integer)
Notes (from the specifications):
The input value of position is the first location in the output buffer to be
used for packing. position is incremented by the size of the packed message,
and the output value of position is the first location in the output buffer
following the locations occupied by the packed message. The comm argument is
the communicator that will be subsequently used for sending the packed
message.
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_ARG
.N MPI_ERR_OTHER
@*/
int MPI_Pack(const void *inbuf,
int incount,
MPI_Datatype datatype,
void *outbuf,
int outsize,
int *position,
MPI_Comm comm)
{
int mpi_errno = MPI_SUCCESS;
MPI_Aint position_x;
MPID_Comm *comm_ptr = NULL;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_PACK);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_PACK);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif
/* Convert MPI object handles to object pointers */
MPID_Comm_get_ptr(comm, comm_ptr);
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COUNT(incount,mpi_errno);
MPIR_ERRTEST_COUNT(outsize,mpi_errno);
/* NOTE: inbuf could be null (MPI_BOTTOM) */
if (incount > 0) {
MPIR_ERRTEST_ARGNULL(outbuf, "output buffer", mpi_errno);
}
MPIR_ERRTEST_ARGNULL(position, "position", mpi_errno);
/* Validate comm_ptr */
/* If comm_ptr is not valid, it will be reset to null */
MPID_Comm_valid_ptr(comm_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype *datatype_ptr = NULL;
MPID_Datatype_get_ptr(datatype, datatype_ptr);
MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPID_Datatype_committed_ptr(datatype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
#ifdef HAVE_ERROR_CHECKING /* IMPLEMENTATION-SPECIFIC ERROR CHECKS */
{
int tmp_sz;
MPID_BEGIN_ERROR_CHECKS;
/* Verify that there is space in the buffer to pack the type */
MPID_Datatype_get_size_macro(datatype, tmp_sz);
if (tmp_sz * incount > outsize - *position) {
if (*position < 0) {
MPIU_ERR_SETANDJUMP1(mpi_errno,MPI_ERR_ARG,
"**argposneg","**argposneg %d",
*position);
}
else if (outsize < 0) {
MPIU_ERR_SETANDJUMP2(mpi_errno,MPI_ERR_ARG,"**argneg",
"**argneg %s %d","outsize",outsize);
}
else if (incount < 0) {
MPIU_ERR_SETANDJUMP2(mpi_errno,MPI_ERR_ARG,"**argneg",
"**argneg %s %d","incount",incount);
}
else {
MPIU_ERR_SETANDJUMP2(mpi_errno,MPI_ERR_ARG,"**argpackbuf",
"**argpackbuf %d %d", tmp_sz * incount,
outsize - *position);
}
}
MPID_END_ERROR_CHECKS;
}
#endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
position_x = *position;
mpi_errno = MPIR_Pack_impl(inbuf, incount, datatype, outbuf, outsize, &position_x);
MPIU_Assign_trunc(*position, position_x, int);
if (mpi_errno) goto fn_fail;
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_PACK);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_pack",
"**mpi_pack %p %d %D %p %d %p %C", inbuf, incount, datatype, outbuf, outsize, position, comm);
}
# endif
mpi_errno = MPIR_Err_return_comm(comm_ptr, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Sendrecv_replace - Sends and receives using a single buffer
Input Parameters:
+ count - number of elements in send and receive buffer (integer)
. datatype - type of elements in send and receive buffer (handle)
. dest - rank of destination (integer)
. sendtag - send message tag (integer)
. source - rank of source (integer)
. recvtag - receive message tag (integer)
- comm - communicator (handle)
Output Parameters:
+ buf - initial address of send and receive buffer (choice)
- status - status object (Status)
.N ThreadSafe
.N Fortran
.N FortranStatus
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_COMM
.N MPI_ERR_COUNT
.N MPI_ERR_TYPE
.N MPI_ERR_TAG
.N MPI_ERR_RANK
.N MPI_ERR_TRUNCATE
.N MPI_ERR_EXHAUSTED
@*/
int MPI_Sendrecv_replace(void *buf, int count, MPI_Datatype datatype,
int dest, int sendtag, int source, int recvtag,
MPI_Comm comm, MPI_Status *status)
{
static const char FCNAME[] = "MPI_Sendrecv_replace";
int mpi_errno = MPI_SUCCESS;
MPIR_Comm *comm_ptr = NULL;
MPIR_CHKLMEM_DECL(1);
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_SENDRECV_REPLACE);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPIR_FUNC_TERSE_PT2PT_ENTER_BOTH(MPID_STATE_MPI_SENDRECV_REPLACE);
/* Convert handles to MPI objects. */
MPIR_Comm_get_ptr(comm, comm_ptr);
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
/* Validate communicator */
MPIR_Comm_valid_ptr( comm_ptr, mpi_errno, FALSE );
if (mpi_errno) goto fn_fail;
/* Validate count */
MPIR_ERRTEST_COUNT(count, mpi_errno);
/* Validate status (status_ignore is not the same as null) */
MPIR_ERRTEST_ARGNULL(status, "status", mpi_errno);
/* Validate tags */
MPIR_ERRTEST_SEND_TAG(sendtag, mpi_errno);
MPIR_ERRTEST_RECV_TAG(recvtag, mpi_errno);
/* Validate source and destination */
MPIR_ERRTEST_SEND_RANK(comm_ptr, dest, mpi_errno);
MPIR_ERRTEST_RECV_RANK(comm_ptr, source, mpi_errno);
/* Validate datatype handle */
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
/* Validate datatype object */
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN)
{
MPIR_Datatype *datatype_ptr = NULL;
MPIR_Datatype_get_ptr(datatype, datatype_ptr);
MPIR_Datatype_valid_ptr(datatype_ptr, mpi_errno);
if (mpi_errno) goto fn_fail;
MPIR_Datatype_committed_ptr(datatype_ptr, mpi_errno);
if (mpi_errno) goto fn_fail;
}
/* Validate buffer */
MPIR_ERRTEST_USERBUFFER(buf,count,datatype,mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
# if defined(MPID_Sendrecv_replace)
{
mpi_errno = MPID_Sendrecv_replace(buf, count, datatype, dest,
sendtag, source, recvtag, comm_ptr,
status)
}
# else
{
MPIR_Request * sreq;
MPIR_Request * rreq;
void * tmpbuf = NULL;
MPI_Aint tmpbuf_size = 0;
MPI_Aint tmpbuf_count = 0;
if (count > 0 && dest != MPI_PROC_NULL)
{
MPIR_Pack_size_impl(count, datatype, &tmpbuf_size);
MPIR_CHKLMEM_MALLOC_ORJUMP(tmpbuf, void *, tmpbuf_size, mpi_errno, "temporary send buffer", MPL_MEM_BUFFER);
mpi_errno = MPIR_Pack_impl(buf, count, datatype, tmpbuf, tmpbuf_size, &tmpbuf_count);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
mpi_errno = MPID_Irecv(buf, count, datatype, source, recvtag,
comm_ptr, MPIR_CONTEXT_INTRA_PT2PT, &rreq);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
mpi_errno = MPID_Isend(tmpbuf, tmpbuf_count, MPI_PACKED, dest,
sendtag, comm_ptr, MPIR_CONTEXT_INTRA_PT2PT,
&sreq);
if (mpi_errno != MPI_SUCCESS)
{
/* --BEGIN ERROR HANDLING-- */
/* FIXME: should we cancel the pending (possibly completed) receive request or wait for it to complete? */
MPIR_Request_free(rreq);
goto fn_fail;
/* --END ERROR HANDLING-- */
}
if (!MPIR_Request_is_complete(sreq) || !MPIR_Request_is_complete(rreq))
{
MPID_Progress_state progress_state;
MPID_Progress_start(&progress_state);
while (!MPIR_Request_is_complete(sreq) || !MPIR_Request_is_complete(rreq))
{
mpi_errno = MPID_Progress_wait(&progress_state);
if (mpi_errno != MPI_SUCCESS)
{
/* --BEGIN ERROR HANDLING-- */
MPID_Progress_end(&progress_state);
goto fn_fail;
/* --END ERROR HANDLING-- */
}
}
MPID_Progress_end(&progress_state);
}
if (status != MPI_STATUS_IGNORE)
{
*status = rreq->status;
}
if (mpi_errno == MPI_SUCCESS)
{
mpi_errno = rreq->status.MPI_ERROR;
if (mpi_errno == MPI_SUCCESS)
{
mpi_errno = sreq->status.MPI_ERROR;
}
}
MPIR_Request_free(sreq);
MPIR_Request_free(rreq);
}
int MPIR_Bsend_isend(const void *buf, int count, MPI_Datatype dtype,
int dest, int tag, MPID_Comm *comm_ptr,
MPIR_Bsend_kind_t kind, MPID_Request **request )
{
int mpi_errno = MPI_SUCCESS;
MPIR_Bsend_data_t *p;
MPIR_Bsend_msg_t *msg;
int packsize, pass;
/* Find a free segment and copy the data into it. If we could
have, we would already have used tBsend to send the message with
no copying.
We may want to decide here whether we need to pack at all
or if we can just use (a MPIU_Memcpy) of the buffer.
*/
/* We check the active buffer first. This helps avoid storage
fragmentation */
mpi_errno = MPIR_Bsend_check_active();
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
if (dtype != MPI_PACKED)
MPIR_Pack_size_impl( count, dtype, &packsize );
else
packsize = count;
MPIU_DBG_MSG_D(BSEND,TYPICAL,"looking for buffer of size %d", packsize);
/*
* Use two passes. Each pass is the same; between the two passes,
* attempt to complete any active requests, and start any pending
* ones. If the message can be initiated in the first pass,
* do not perform the second pass.
*/
for (pass = 0; pass < 2; pass++) {
p = MPIR_Bsend_find_buffer( packsize );
if (p) {
MPIU_DBG_MSG_FMT(BSEND,TYPICAL,(MPIU_DBG_FDEST,
"found buffer of size %d with address %p",packsize,p));
/* Found a segment */
msg = &p->msg;
/* Pack the data into the buffer */
/* We may want to optimize for the special case of
either primative or contiguous types, and just
use MPIU_Memcpy and the provided datatype */
msg->count = 0;
if (dtype != MPI_PACKED)
{
mpi_errno = MPIR_Pack_impl( buf, count, dtype, p->msg.msgbuf, packsize, &p->msg.count);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else
{
MPIU_Memcpy(p->msg.msgbuf, buf, count);
p->msg.count = count;
}
/* Try to send the message. We must use MPID_Isend
because this call must not block */
mpi_errno = MPID_Isend(msg->msgbuf, msg->count, MPI_PACKED,
dest, tag, comm_ptr,
MPID_CONTEXT_INTRA_PT2PT, &p->request );
MPIU_ERR_CHKINTERNAL(mpi_errno, mpi_errno, "Bsend internal error: isend returned err");
/* If the error is "request not available", we should
put this on the pending list. This will depend on
how we signal failure to send. */
if (p->request) {
MPIU_DBG_MSG_FMT(BSEND,TYPICAL,
(MPIU_DBG_FDEST,"saving request %p in %p",p->request,p));
/* An optimization is to check to see if the
data has already been sent. The original code
to do this was commented out and probably did not match
the current request internals */
MPIR_Bsend_take_buffer( p, p->msg.count );
p->kind = kind;
*request = p->request;
}
break;
}
/* If we found a buffer or we're in the seccond pass, then break.
Note that the test on phere is redundant, as the code breaks
out of the loop in the test above if a block p is found. */
if (p || pass == 1) break;
MPIU_DBG_MSG(BSEND,TYPICAL,"Could not find storage, checking active");
/* Try to complete some pending bsends */
MPIR_Bsend_check_active( );
/* Give priority to any pending operations */
MPIR_Bsend_retry_pending( );
}
if (!p) {
/* Return error for no buffer space found */
/* Generate a traceback of the allocated space, explaining why
packsize could not be found */
MPIU_DBG_MSG(BSEND,TYPICAL,"Could not find space; dumping arena" );
MPIU_DBG_STMT(BSEND,TYPICAL,MPIR_Bsend_dump());
MPIU_ERR_SETANDJUMP2(mpi_errno, MPI_ERR_BUFFER, "**bufbsend", "**bufbsend %d %d", packsize, BsendBuffer.buffer_size);
}
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Alltoall_intra(
const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
int recvcount,
MPI_Datatype recvtype,
MPIR_Comm *comm_ptr,
MPIR_Errflag_t *errflag )
{
int comm_size, i, j, pof2;
MPI_Aint sendtype_extent, recvtype_extent;
MPI_Aint recvtype_true_extent, recvbuf_extent, recvtype_true_lb;
int mpi_errno=MPI_SUCCESS, src, dst, rank, nbytes;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
int sendtype_size, block, *displs, count;
MPI_Aint pack_size, position;
MPI_Datatype newtype = MPI_DATATYPE_NULL;
void *tmp_buf;
MPIR_Request **reqarray;
MPI_Status *starray;
MPIR_CHKLMEM_DECL(6);
if (recvcount == 0) return MPI_SUCCESS;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* Get extent of send and recv types */
MPID_Datatype_get_extent_macro(recvtype, recvtype_extent);
MPID_Datatype_get_extent_macro(sendtype, sendtype_extent);
MPID_Datatype_get_size_macro(sendtype, sendtype_size);
nbytes = sendtype_size * sendcount;
if (sendbuf == MPI_IN_PLACE) {
/* We use pair-wise sendrecv_replace in order to conserve memory usage,
* which is keeping with the spirit of the MPI-2.2 Standard. But
* because of this approach all processes must agree on the global
* schedule of sendrecv_replace operations to avoid deadlock.
*
* Note that this is not an especially efficient algorithm in terms of
* time and there will be multiple repeated malloc/free's rather than
* maintaining a single buffer across the whole loop. Something like
* MADRE is probably the best solution for the MPI_IN_PLACE scenario. */
for (i = 0; i < comm_size; ++i) {
/* start inner loop at i to avoid re-exchanging data */
for (j = i; j < comm_size; ++j) {
if (rank == i) {
/* also covers the (rank == i && rank == j) case */
mpi_errno = MPIC_Sendrecv_replace(((char *)recvbuf + j*recvcount*recvtype_extent),
recvcount, recvtype,
j, MPIR_ALLTOALL_TAG,
j, MPIR_ALLTOALL_TAG,
comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
else if (rank == j) {
/* same as above with i/j args reversed */
mpi_errno = MPIC_Sendrecv_replace(((char *)recvbuf + i*recvcount*recvtype_extent),
recvcount, recvtype,
i, MPIR_ALLTOALL_TAG,
i, MPIR_ALLTOALL_TAG,
comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
}
}
else if ((nbytes <= MPIR_CVAR_ALLTOALL_SHORT_MSG_SIZE) && (comm_size >= 8)) {
/* use the indexing algorithm by Jehoshua Bruck et al,
* IEEE TPDS, Nov. 97 */
/* allocate temporary buffer */
MPIR_Pack_size_impl(recvcount*comm_size, recvtype, &pack_size);
MPIR_CHKLMEM_MALLOC(tmp_buf, void *, pack_size, mpi_errno, "tmp_buf");
/* Do Phase 1 of the algorithim. Shift the data blocks on process i
* upwards by a distance of i blocks. Store the result in recvbuf. */
mpi_errno = MPIR_Localcopy((char *) sendbuf +
rank*sendcount*sendtype_extent,
(comm_size - rank)*sendcount, sendtype, recvbuf,
(comm_size - rank)*recvcount, recvtype);
if (mpi_errno) { MPIR_ERR_POP(mpi_errno); }
mpi_errno = MPIR_Localcopy(sendbuf, rank*sendcount, sendtype,
(char *) recvbuf +
(comm_size-rank)*recvcount*recvtype_extent,
rank*recvcount, recvtype);
if (mpi_errno) { MPIR_ERR_POP(mpi_errno); }
/* Input data is now stored in recvbuf with datatype recvtype */
/* Now do Phase 2, the communication phase. It takes
ceiling(lg p) steps. In each step i, each process sends to rank+2^i
and receives from rank-2^i, and exchanges all data blocks
whose ith bit is 1. */
/* allocate displacements array for indexed datatype used in
communication */
MPIR_CHKLMEM_MALLOC(displs, int *, comm_size * sizeof(int), mpi_errno, "displs");
pof2 = 1;
while (pof2 < comm_size) {
dst = (rank + pof2) % comm_size;
src = (rank - pof2 + comm_size) % comm_size;
/* Exchange all data blocks whose ith bit is 1 */
/* Create an indexed datatype for the purpose */
count = 0;
for (block=1; block<comm_size; block++) {
if (block & pof2) {
displs[count] = block * recvcount;
count++;
}
}
mpi_errno = MPIR_Type_create_indexed_block_impl(count, recvcount,
displs, recvtype, &newtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Type_commit_impl(&newtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
position = 0;
mpi_errno = MPIR_Pack_impl(recvbuf, 1, newtype, tmp_buf, pack_size, &position);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIC_Sendrecv(tmp_buf, position, MPI_PACKED, dst,
MPIR_ALLTOALL_TAG, recvbuf, 1, newtype,
src, MPIR_ALLTOALL_TAG, comm_ptr,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
MPIR_Type_free_impl(&newtype);
pof2 *= 2;
}
/* Rotate blocks in recvbuf upwards by (rank + 1) blocks. Need
* a temporary buffer of the same size as recvbuf. */
/* get true extent of recvtype */
MPIR_Type_get_true_extent_impl(recvtype, &recvtype_true_lb, &recvtype_true_extent);
recvbuf_extent = recvcount * comm_size *
(MPL_MAX(recvtype_true_extent, recvtype_extent));
MPIR_CHKLMEM_MALLOC(tmp_buf, void *, recvbuf_extent, mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - recvtype_true_lb);
mpi_errno = MPIR_Localcopy((char *) recvbuf + (rank+1)*recvcount*recvtype_extent,
(comm_size - rank - 1)*recvcount, recvtype, tmp_buf,
(comm_size - rank - 1)*recvcount, recvtype);
if (mpi_errno) { MPIR_ERR_POP(mpi_errno); }
mpi_errno = MPIR_Localcopy(recvbuf, (rank+1)*recvcount, recvtype,
(char *) tmp_buf + (comm_size-rank-1)*recvcount*recvtype_extent,
(rank+1)*recvcount, recvtype);
if (mpi_errno) { MPIR_ERR_POP(mpi_errno); }
/* Blocks are in the reverse order now (comm_size-1 to 0).
* Reorder them to (0 to comm_size-1) and store them in recvbuf. */
for (i=0; i<comm_size; i++){
mpi_errno = MPIR_Localcopy((char *) tmp_buf + i*recvcount*recvtype_extent,
recvcount, recvtype,
(char *) recvbuf + (comm_size-i-1)*recvcount*recvtype_extent,
recvcount, recvtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
}
