int MPIR_Exscan (
const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPIR_Comm *comm_ptr,
MPIR_Errflag_t *errflag )
{
MPI_Status status;
int rank, comm_size;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative, flag;
MPI_Aint true_extent, true_lb, extent;
void *partial_scan, *tmp_buf;
MPIR_Op *op_ptr;
MPIR_CHKLMEM_DECL(2);
if (count == 0) return MPI_SUCCESS;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* set op_errno to 0. stored in perthread structure */
{
MPIR_Per_thread_t *per_thread = NULL;
int err = 0;
MPID_THREADPRIV_KEY_GET_ADDR(MPIR_ThreadInfo.isThreaded, MPIR_Per_thread_key,
MPIR_Per_thread, per_thread, &err);
MPIR_Assert(err == 0);
per_thread->op_errno = 0;
}
if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
is_commutative = 1;
}
else {
MPIR_Op_get_ptr(op, op_ptr);
if (op_ptr->kind == MPIR_OP_KIND__USER_NONCOMMUTE)
is_commutative = 0;
else
is_commutative = 1;
}
/* need to allocate temporary buffer to store partial scan*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro( datatype, extent );
MPIR_CHKLMEM_MALLOC(partial_scan, void *, (count*(MPL_MAX(true_extent,extent))), mpi_errno, "partial_scan");
/* adjust for potential negative lower bound in datatype */
partial_scan = (void *)((char*)partial_scan - true_lb);
/* need to allocate temporary buffer to store incoming data*/
MPIR_CHKLMEM_MALLOC(tmp_buf, void *, (count*(MPL_MAX(true_extent,extent))), mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
mpi_errno = MPIR_Localcopy((sendbuf == MPI_IN_PLACE ? (const void *)recvbuf : sendbuf), count, datatype,
partial_scan, count, datatype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
flag = 0;
mask = 0x1;
while (mask < comm_size) {
dst = rank ^ mask;
if (dst < comm_size) {
/* Send partial_scan to dst. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv(partial_scan, count, datatype,
dst, MPIR_EXSCAN_TAG, tmp_buf,
count, datatype, dst,
MPIR_EXSCAN_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);
}
if (rank > dst) {
mpi_errno = MPIR_Reduce_local_impl( tmp_buf, partial_scan,
count, datatype, op );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* On rank 0, recvbuf is not defined. For sendbuf==MPI_IN_PLACE
recvbuf must not change (per MPI-2.2).
On rank 1, recvbuf is to be set equal to the value
in sendbuf on rank 0.
On others, recvbuf is the scan of values in the
sendbufs on lower ranks. */
if (rank != 0) {
if (flag == 0) {
/* simply copy data recd from rank 0 into recvbuf */
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
recvbuf, count, datatype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
flag = 1;
}
else {
mpi_errno = MPIR_Reduce_local_impl( tmp_buf,
recvbuf, count, datatype, op );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
}
}
else {
if (is_commutative) {
mpi_errno = MPIR_Reduce_local_impl( tmp_buf, partial_scan,
count, datatype, op );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else {
mpi_errno = MPIR_Reduce_local_impl( partial_scan,
tmp_buf, count, datatype, op );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
partial_scan,
count, datatype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
}
}
mask <<= 1;
}
{
MPIR_Per_thread_t *per_thread = NULL;
int err = 0;
MPID_THREADPRIV_KEY_GET_ADDR(MPIR_ThreadInfo.isThreaded, MPIR_Per_thread_key,
MPIR_Per_thread, per_thread, &err);
MPIR_Assert(err == 0);
if (per_thread->op_errno)
mpi_errno = per_thread->op_errno;
}
fn_exit:
MPIR_CHKLMEM_FREEALL();
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
/* not declared static because a machine-specific function may call this one in some cases */
int MPIR_Reduce_scatter_block_intra (
const void *sendbuf,
void *recvbuf,
int recvcount,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
int rank, comm_size, i;
MPI_Aint extent, true_extent, true_lb;
int *disps;
void *tmp_recvbuf, *tmp_results;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int type_size, dis[2], blklens[2], total_count, nbytes, src, dst;
int mask, dst_tree_root, my_tree_root, j, k;
int *newcnts, *newdisps, rem, newdst, send_idx, recv_idx,
last_idx, send_cnt, recv_cnt;
int pof2, old_i, newrank, received;
MPI_Datatype sendtype, recvtype;
int nprocs_completed, tmp_mask, tree_root, is_commutative;
MPID_Op *op_ptr;
MPI_Comm comm;
MPIU_THREADPRIV_DECL;
MPIU_CHKLMEM_DECL(5);
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* set op_errno to 0. stored in perthread structure */
MPIU_THREADPRIV_GET;
MPIU_THREADPRIV_FIELD(op_errno) = 0;
if (recvcount == 0) {
goto fn_exit;
}
MPID_Datatype_get_extent_macro(datatype, extent);
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
is_commutative = 1;
}
else {
MPID_Op_get_ptr(op, op_ptr);
if (op_ptr->kind == MPID_OP_USER_NONCOMMUTE)
is_commutative = 0;
else
is_commutative = 1;
}
MPIU_CHKLMEM_MALLOC(disps, int *, comm_size * sizeof(int), mpi_errno, "disps");
total_count = comm_size*recvcount;
for (i=0; i<comm_size; i++) {
disps[i] = i*recvcount;
}
MPID_Datatype_get_size_macro(datatype, type_size);
nbytes = total_count * type_size;
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
/* total_count*extent eventually gets malloced. it isn't added to
* a user-passed in buffer */
MPID_Ensure_Aint_fits_in_pointer(total_count * MPIR_MAX(true_extent, extent));
if ((is_commutative) && (nbytes < MPIR_PARAM_REDSCAT_COMMUTATIVE_LONG_MSG_SIZE)) {
/* commutative and short. use recursive halving algorithm */
/* allocate temp. buffer to receive incoming data */
MPIU_CHKLMEM_MALLOC(tmp_recvbuf, void *, total_count*(MPIR_MAX(true_extent,extent)), mpi_errno, "tmp_recvbuf");
/* adjust for potential negative lower bound in datatype */
tmp_recvbuf = (void *)((char*)tmp_recvbuf - true_lb);
/* need to allocate another temporary buffer to accumulate
results because recvbuf may not be big enough */
MPIU_CHKLMEM_MALLOC(tmp_results, void *, total_count*(MPIR_MAX(true_extent,extent)), mpi_errno, "tmp_results");
/* adjust for potential negative lower bound in datatype */
tmp_results = (void *)((char*)tmp_results - true_lb);
/* copy sendbuf into tmp_results */
if (sendbuf != MPI_IN_PLACE)
mpi_errno = MPIR_Localcopy(sendbuf, total_count, datatype,
tmp_results, total_count, datatype);
else
mpi_errno = MPIR_Localcopy(recvbuf, total_count, datatype,
tmp_results, total_count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
pof2 = 1;
while (pof2 <= comm_size) pof2 <<= 1;
pof2 >>=1;
rem = comm_size - pof2;
/* In the non-power-of-two case, all even-numbered
processes of rank < 2*rem send their data to
(rank+1). These even-numbered processes no longer
participate in the algorithm until the very end. The
remaining processes form a nice power-of-two. */
if (rank < 2*rem) {
if (rank % 2 == 0) { /* even */
mpi_errno = MPIC_Send_ft(tmp_results, total_count,
datatype, rank+1,
MPIR_REDUCE_SCATTER_BLOCK_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);
}
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
doubling */
newrank = -1;
}
else { /* odd */
mpi_errno = MPIC_Recv_ft(tmp_recvbuf, total_count,
datatype, rank-1,
MPIR_REDUCE_SCATTER_BLOCK_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);
}
/* do the reduction on received data. since the
ordering is right, it doesn't matter whether
the operation is commutative or not. */
mpi_errno = MPIR_Reduce_local_impl( tmp_recvbuf, tmp_results,
total_count, datatype, op);
/* change the rank */
newrank = rank / 2;
}
}
else /* rank >= 2*rem */
newrank = rank - rem;
if (newrank != -1) {
/* recalculate the recvcnts and disps arrays because the
even-numbered processes who no longer participate will
have their result calculated by the process to their
right (rank+1). */
MPIU_CHKLMEM_MALLOC(newcnts, int *, pof2*sizeof(int), mpi_errno, "newcnts");
MPIU_CHKLMEM_MALLOC(newdisps, int *, pof2*sizeof(int), mpi_errno, "newdisps");
for (i=0; i<pof2; i++) {
/* what does i map to in the old ranking? */
old_i = (i < rem) ? i*2 + 1 : i + rem;
if (old_i < 2*rem) {
/* This process has to also do its left neighbor's
work */
newcnts[i] = 2 * recvcount;
}
else
newcnts[i] = recvcount;
}
newdisps[0] = 0;
for (i=1; i<pof2; i++)
newdisps[i] = newdisps[i-1] + newcnts[i-1];
mask = pof2 >> 1;
send_idx = recv_idx = 0;
last_idx = pof2;
while (mask > 0) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
send_idx = recv_idx + mask;
for (i=send_idx; i<last_idx; i++)
send_cnt += newcnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += newcnts[i];
}
else {
recv_idx = send_idx + mask;
for (i=send_idx; i<recv_idx; i++)
send_cnt += newcnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += newcnts[i];
}
/* printf("Rank %d, send_idx %d, recv_idx %d, send_cnt %d, recv_cnt %d, last_idx %d\n", newrank, send_idx, recv_idx,
send_cnt, recv_cnt, last_idx);
*/
/* Send data from tmp_results. Recv into tmp_recvbuf */
if ((send_cnt != 0) && (recv_cnt != 0))
mpi_errno = MPIC_Sendrecv_ft((char *) tmp_results +
newdisps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_REDUCE_SCATTER_BLOCK_TAG,
(char *) tmp_recvbuf +
newdisps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
MPI_STATUS_IGNORE, errflag);
else if ((send_cnt == 0) && (recv_cnt != 0))
mpi_errno = MPIC_Recv_ft((char *) tmp_recvbuf +
newdisps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
MPI_STATUS_IGNORE, errflag);
else if ((recv_cnt == 0) && (send_cnt != 0))
mpi_errno = MPIC_Send_ft((char *) tmp_results +
newdisps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_REDUCE_SCATTER_BLOCK_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);
}
/* tmp_recvbuf contains data received in this step.
tmp_results contains data accumulated so far */
if (recv_cnt) {
mpi_errno = MPIR_Reduce_local_impl(
(char *) tmp_recvbuf + newdisps[recv_idx]*extent,
(char *) tmp_results + newdisps[recv_idx]*extent,
recv_cnt, datatype, op);
}
/* update send_idx for next iteration */
send_idx = recv_idx;
last_idx = recv_idx + mask;
mask >>= 1;
}
/* copy this process's result from tmp_results to recvbuf */
mpi_errno = MPIR_Localcopy((char *)tmp_results +
disps[rank]*extent,
recvcount, datatype, recvbuf,
recvcount, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
static int MPIR_Reduce_scatter_block_noncomm (
const void *sendbuf,
void *recvbuf,
int recvcount,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int comm_size = comm_ptr->local_size;
int rank = comm_ptr->rank;
int pof2;
int log2_comm_size;
int i, k;
int recv_offset, send_offset;
int block_size, total_count, size;
MPI_Aint true_extent, true_lb;
int buf0_was_inout;
void *tmp_buf0;
void *tmp_buf1;
void *result_ptr;
MPI_Comm comm = comm_ptr->handle;
MPIU_CHKLMEM_DECL(3);
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
pof2 = 1;
log2_comm_size = 0;
while (pof2 < comm_size) {
pof2 <<= 1;
++log2_comm_size;
}
/* begin error checking */
MPIU_Assert(pof2 == comm_size); /* FIXME this version only works for power of 2 procs */
/* end error checking */
/* size of a block (count of datatype per block, NOT bytes per block) */
block_size = recvcount;
total_count = block_size * comm_size;
MPIU_CHKLMEM_MALLOC(tmp_buf0, void *, true_extent * total_count, mpi_errno, "tmp_buf0");
MPIU_CHKLMEM_MALLOC(tmp_buf1, void *, true_extent * total_count, mpi_errno, "tmp_buf1");
/* adjust for potential negative lower bound in datatype */
tmp_buf0 = (void *)((char*)tmp_buf0 - true_lb);
tmp_buf1 = (void *)((char*)tmp_buf1 - true_lb);
/* Copy our send data to tmp_buf0. We do this one block at a time and
permute the blocks as we go according to the mirror permutation. */
for (i = 0; i < comm_size; ++i) {
mpi_errno = MPIR_Localcopy((char *)(sendbuf == MPI_IN_PLACE ? recvbuf : sendbuf) + (i * true_extent * block_size), block_size, datatype,
(char *)tmp_buf0 + (MPIU_Mirror_permutation(i, log2_comm_size) * true_extent * block_size), block_size, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
buf0_was_inout = 1;
send_offset = 0;
recv_offset = 0;
size = total_count;
for (k = 0; k < log2_comm_size; ++k) {
/* use a double-buffering scheme to avoid local copies */
char *incoming_data = (buf0_was_inout ? tmp_buf1 : tmp_buf0);
char *outgoing_data = (buf0_was_inout ? tmp_buf0 : tmp_buf1);
int peer = rank ^ (0x1 << k);
size /= 2;
if (rank > peer) {
/* we have the higher rank: send top half, recv bottom half */
recv_offset += size;
}
else {
/* we have the lower rank: recv top half, send bottom half */
send_offset += size;
}
mpi_errno = MPIC_Sendrecv_ft(outgoing_data + send_offset*true_extent,
size, datatype, peer, MPIR_REDUCE_SCATTER_BLOCK_TAG,
incoming_data + recv_offset*true_extent,
size, datatype, peer, MPIR_REDUCE_SCATTER_BLOCK_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);
}
/* always perform the reduction at recv_offset, the data at send_offset
is now our peer's responsibility */
if (rank > peer) {
/* higher ranked value so need to call op(received_data, my_data) */
mpi_errno = MPIR_Reduce_local_impl(
incoming_data + recv_offset*true_extent,
outgoing_data + recv_offset*true_extent,
size, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
buf0_was_inout = buf0_was_inout;
}
else {
/* lower ranked value so need to call op(my_data, received_data) */
mpi_errno = MPIR_Reduce_local_impl(
outgoing_data + recv_offset*true_extent,
incoming_data + recv_offset*true_extent,
size, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
buf0_was_inout = !buf0_was_inout;
}
/* the next round of send/recv needs to happen within the block (of size
"size") that we just received and reduced */
send_offset = recv_offset;
}
MPIU_Assert(size == recvcount);
/* copy the reduced data to the recvbuf */
result_ptr = (char *)(buf0_was_inout ? tmp_buf0 : tmp_buf1) + recv_offset * true_extent;
mpi_errno = MPIR_Localcopy(result_ptr, size, datatype,
recvbuf, size, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
fn_exit:
MPIU_CHKLMEM_FREEALL();
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
/* --END ERROR HANDLING-- */
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Iallreduce_sched_intra_reduce_scatter_allgather(const void *sendbuf, void *recvbuf,
int count, MPI_Datatype datatype,
MPI_Op op, MPIR_Comm * comm_ptr,
MPIR_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int comm_size, rank, newrank, pof2, rem;
int i, send_idx, recv_idx, last_idx, mask, newdst, dst, send_cnt, recv_cnt;
MPI_Aint true_lb, true_extent, extent;
void *tmp_buf = NULL;
int *cnts, *disps;
MPIR_SCHED_CHKPMEM_DECL(1);
MPIR_CHKLMEM_DECL(2);
#ifdef HAVE_ERROR_CHECKING
/* we only support builtin datatypes for now, breaking up user types to do
* the reduce-scatter is tricky */
MPIR_Assert(HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN);
#endif
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* need to allocate temporary buffer to store incoming data */
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPIR_Datatype_get_extent_macro(datatype, extent);
MPIR_Ensure_Aint_fits_in_pointer(count * MPL_MAX(extent, true_extent));
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, count * (MPL_MAX(extent, true_extent)), mpi_errno,
"temporary buffer", MPL_MEM_BUFFER);
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *) ((char *) tmp_buf - true_lb);
/* copy local data into recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Sched_copy(sendbuf, count, datatype, recvbuf, count, datatype, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
/* get nearest power-of-two less than or equal to comm_size */
pof2 = comm_ptr->pof2;
rem = comm_size - pof2;
/* In the non-power-of-two case, all even-numbered
* processes of rank < 2*rem send their data to
* (rank+1). These even-numbered processes no longer
* participate in the algorithm until the very end. The
* remaining processes form a nice power-of-two. */
if (rank < 2 * rem) {
if (rank % 2 == 0) { /* even */
mpi_errno = MPIR_Sched_send(recvbuf, count, datatype, rank + 1, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
/* temporarily set the rank to -1 so that this
* process does not pariticipate in recursive
* doubling */
newrank = -1;
} else { /* odd */
mpi_errno = MPIR_Sched_recv(tmp_buf, count, datatype, rank - 1, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
/* do the reduction on received data. since the
* ordering is right, it doesn't matter whether
* the operation is commutative or not. */
mpi_errno = MPIR_Sched_reduce(tmp_buf, recvbuf, count, datatype, op, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
/* change the rank */
newrank = rank / 2;
}
} else /* rank >= 2*rem */
newrank = rank - rem;
if (newrank != -1) {
/* for the reduce-scatter, calculate the count that
* each process receives and the displacement within
* the buffer */
/* TODO I (goodell@) believe that these counts and displacements could be
* calculated directly during the loop, rather than requiring a less-scalable
* "2*pof2"-sized memory allocation */
MPIR_CHKLMEM_MALLOC(cnts, int *, pof2 * sizeof(int), mpi_errno, "counts", MPL_MEM_BUFFER);
MPIR_CHKLMEM_MALLOC(disps, int *, pof2 * sizeof(int), mpi_errno, "displacements",
MPL_MEM_BUFFER);
MPIR_Assert(count >= pof2); /* the cnts calculations assume this */
for (i = 0; i < (pof2 - 1); i++)
cnts[i] = count / pof2;
cnts[pof2 - 1] = count - (count / pof2) * (pof2 - 1);
if (pof2)
disps[0] = 0;
for (i = 1; i < pof2; i++)
disps[i] = disps[i - 1] + cnts[i - 1];
mask = 0x1;
send_idx = recv_idx = 0;
last_idx = pof2;
while (mask < pof2) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst * 2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
send_idx = recv_idx + pof2 / (mask * 2);
for (i = send_idx; i < last_idx; i++)
send_cnt += cnts[i];
for (i = recv_idx; i < send_idx; i++)
recv_cnt += cnts[i];
} else {
recv_idx = send_idx + pof2 / (mask * 2);
for (i = send_idx; i < recv_idx; i++)
send_cnt += cnts[i];
for (i = recv_idx; i < last_idx; i++)
recv_cnt += cnts[i];
}
/* Send data from recvbuf. Recv into tmp_buf */
mpi_errno = MPIR_Sched_recv(((char *) tmp_buf + disps[recv_idx] * extent),
recv_cnt, datatype, dst, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* sendrecv, no barrier here */
mpi_errno = MPIR_Sched_send(((char *) recvbuf + disps[send_idx] * extent),
send_cnt, datatype, dst, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
/* tmp_buf contains data received in this step.
* recvbuf contains data accumulated so far */
/* This algorithm is used only for predefined ops
* and predefined ops are always commutative. */
mpi_errno = MPIR_Sched_reduce(((char *) tmp_buf + disps[recv_idx] * extent),
((char *) recvbuf + disps[recv_idx] * extent),
recv_cnt, datatype, op, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
/* update send_idx for next iteration */
send_idx = recv_idx;
mask <<= 1;
/* update last_idx, but not in last iteration
* because the value is needed in the allgather
* step below. */
if (mask < pof2)
last_idx = recv_idx + pof2 / mask;
}
/* now do the allgather */
mask >>= 1;
while (mask > 0) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst * 2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
/* update last_idx except on first iteration */
if (mask != pof2 / 2)
last_idx = last_idx + pof2 / (mask * 2);
recv_idx = send_idx + pof2 / (mask * 2);
for (i = send_idx; i < recv_idx; i++)
send_cnt += cnts[i];
for (i = recv_idx; i < last_idx; i++)
recv_cnt += cnts[i];
} else {
recv_idx = send_idx - pof2 / (mask * 2);
for (i = send_idx; i < last_idx; i++)
send_cnt += cnts[i];
for (i = recv_idx; i < send_idx; i++)
recv_cnt += cnts[i];
}
mpi_errno = MPIR_Sched_recv(((char *) recvbuf + disps[recv_idx] * extent),
recv_cnt, datatype, dst, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* sendrecv, no barrier here */
mpi_errno = MPIR_Sched_send(((char *) recvbuf + disps[send_idx] * extent),
send_cnt, datatype, dst, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
if (newrank > newdst)
send_idx = recv_idx;
mask >>= 1;
}
}
int MPIR_Iallgatherv_rec_dbl(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
void *recvbuf, const int recvcounts[], const int displs[],
MPI_Datatype recvtype, MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int comm_size, rank, i, j, k;
int curr_count, send_offset, incoming_count, recv_offset;
int mask, dst, total_count, position, offset, my_tree_root, dst_tree_root;
MPI_Aint recvtype_extent, recvtype_true_extent, recvtype_true_lb;
void *tmp_buf = NULL;
int is_homogeneous ATTRIBUTE((unused));
MPIR_SCHED_CHKPMEM_DECL(1);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
if (comm_ptr->is_hetero)
is_homogeneous = 0;
#endif
MPIU_Assert(is_homogeneous); /* we only handle the homogeneous for now */
/* need to receive contiguously into tmp_buf because
displs could make the recvbuf noncontiguous */
MPID_Datatype_get_extent_macro(recvtype, recvtype_extent);
MPIR_Type_get_true_extent_impl(recvtype, &recvtype_true_lb, &recvtype_true_extent);
total_count = 0;
for (i=0; i<comm_size; i++)
total_count += recvcounts[i];
if (total_count == 0)
goto fn_exit;
MPID_Ensure_Aint_fits_in_pointer(total_count*(MPIR_MAX(recvtype_true_extent, recvtype_extent)));
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, total_count*(MPIR_MAX(recvtype_true_extent,recvtype_extent)), mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - recvtype_true_lb);
/* copy local data into right location in tmp_buf */
position = 0;
for (i=0; i<rank; i++) position += recvcounts[i];
if (sendbuf != MPI_IN_PLACE)
{
mpi_errno = MPID_Sched_copy(sendbuf, sendcount, sendtype,
((char *)tmp_buf + position*recvtype_extent),
recvcounts[rank], recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else
{
/* if in_place specified, local data is found in recvbuf */
mpi_errno = MPID_Sched_copy(((char *)recvbuf + displs[rank]*recvtype_extent),
recvcounts[rank], recvtype,
((char *)tmp_buf + position*recvtype_extent),
recvcounts[rank], recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
curr_count = recvcounts[rank];
/* never used uninitialized w/o this, but compiler can't tell that */
incoming_count = -1;
/* [goodell@] random notes that help slightly when deciphering this code:
* - mask is also equal to the number of blocks that we are going to recv
* (less if comm_size is non-pof2)
* - FOO_tree_root is the leftmost (lowest ranked) process with whom FOO has
* communicated, directly or indirectly, at the beginning of round the
* round. FOO is either "dst" or "my", where "my" means use my rank.
* - in each round we are going to recv the blocks
* B[dst_tree_root],B[dst_tree_root+1],...,B[min(dst_tree_root+mask,comm_size)]
*/
mask = 0x1;
i = 0;
while (mask < comm_size) {
dst = rank ^ mask;
/* find offset into send and recv buffers. zero out
the least significant "i" bits of rank and dst to
find root of src and dst subtrees. Use ranks of
roots as index to send from and recv into buffer */
dst_tree_root = dst >> i;
dst_tree_root <<= i;
my_tree_root = rank >> i;
my_tree_root <<= i;
if (dst < comm_size) {
send_offset = 0;
for (j = 0; j < my_tree_root; j++)
send_offset += recvcounts[j];
recv_offset = 0;
for (j = 0; j < dst_tree_root; j++)
recv_offset += recvcounts[j];
incoming_count = 0;
for (j = dst_tree_root; j < (dst_tree_root + mask) && j < comm_size; ++j)
incoming_count += recvcounts[j];
mpi_errno = MPID_Sched_send(((char *)tmp_buf + send_offset * recvtype_extent),
curr_count, recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* sendrecv, no barrier here */
mpi_errno = MPID_Sched_recv(((char *)tmp_buf + recv_offset * recvtype_extent),
incoming_count, recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
curr_count += incoming_count;
}
/* 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
data that they would normally have received from those
processes. That is, the haves in this subtree must send to
the havenots. We use a logarithmic
recursive-halfing algorithm for this. */
/* This part of the code will not currently be
executed because we are not using recursive
doubling for non power of two. Mark it as experimental
so that it doesn't show up as red in the coverage
tests. */
/* --BEGIN EXPERIMENTAL-- */
if (dst_tree_root + mask > comm_size) {
int tmp_mask, tree_root;
int 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 */
/* [goodell@] it looks like (k==i) is always true, could possibly
* skip the loop below */
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))
{
offset = 0;
for (j = 0; j < (my_tree_root+mask); j++)
offset += recvcounts[j];
offset *= recvtype_extent;
/* incoming_count was set in the previous
receive. that's the amount of data to be
sent now. */
mpi_errno = MPID_Sched_send(((char *)tmp_buf + offset),
incoming_count, recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
/* 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))
{
offset = 0;
for (j = 0; j < (my_tree_root+mask); j++)
offset += recvcounts[j];
/* recalculate incoming_count, since not all processes will have
* this value */
incoming_count = 0;
for (j = dst_tree_root; j < (dst_tree_root + mask) && j < comm_size; ++j)
incoming_count += recvcounts[j];
mpi_errno = MPID_Sched_recv(((char *)tmp_buf + offset * recvtype_extent),
incoming_count, recvtype,
dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
curr_count += incoming_count;
}
tmp_mask >>= 1;
k--;
}
}
/* --END EXPERIMENTAL-- */
mask <<= 1;
i++;
}
static int MPIR_Scan_generic (
const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
MPI_Status status;
int rank, comm_size;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative;
MPI_Aint true_extent, true_lb, extent;
void *partial_scan, *tmp_buf;
MPID_Op *op_ptr;
MPI_Comm comm;
MPIU_THREADPRIV_DECL;
MPIU_CHKLMEM_DECL(2);
if (count == 0) return MPI_SUCCESS;
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
MPIU_THREADPRIV_GET;
/* set op_errno to 0. stored in perthread structure */
MPIU_THREADPRIV_FIELD(op_errno) = 0;
if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
is_commutative = 1;
}
else {
MPID_Op_get_ptr(op, op_ptr);
if (op_ptr->kind == MPID_OP_USER_NONCOMMUTE)
is_commutative = 0;
else
is_commutative = 1;
}
/* need to allocate temporary buffer to store partial scan*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(datatype, extent);
MPIU_CHKLMEM_MALLOC(partial_scan, void *, count*(MPIR_MAX(extent,true_extent)), mpi_errno, "partial_scan");
/* This eventually gets malloc()ed as a temp buffer, not added to
* any user buffers */
MPID_Ensure_Aint_fits_in_pointer(count * MPIR_MAX(extent, true_extent));
/* adjust for potential negative lower bound in datatype */
partial_scan = (void *)((char*)partial_scan - true_lb);
/* need to allocate temporary buffer to store incoming data*/
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, count*(MPIR_MAX(extent,true_extent)), mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
/* Since this is an inclusive scan, copy local contribution into
recvbuf. */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype,
recvbuf, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
if (sendbuf != MPI_IN_PLACE)
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype,
partial_scan, count, datatype);
else
mpi_errno = MPIR_Localcopy(recvbuf, count, datatype,
partial_scan, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mask = 0x1;
while (mask < comm_size) {
dst = rank ^ mask;
if (dst < comm_size) {
/* Send partial_scan to dst. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv(partial_scan, count, datatype,
dst, MPIR_SCAN_TAG, tmp_buf,
count, datatype, dst,
MPIR_SCAN_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);
}
if (rank > dst) {
mpi_errno = MPIR_Reduce_local_impl(
tmp_buf, partial_scan, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIR_Reduce_local_impl(
tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else {
if (is_commutative) {
mpi_errno = MPIR_Reduce_local_impl(
tmp_buf, partial_scan, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else {
mpi_errno = MPIR_Reduce_local_impl(
partial_scan, tmp_buf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
partial_scan,
count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
}
mask <<= 1;
}
if (MPIU_THREADPRIV_FIELD(op_errno)) {
mpi_errno = MPIU_THREADPRIV_FIELD(op_errno);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Allreduce_intra (
void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
int is_homogeneous;
#ifdef MPID_HAS_HETERO
int rc;
#endif
int comm_size, rank, type_size;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative, pof2, newrank, rem, newdst, i,
send_idx, recv_idx, last_idx, send_cnt, recv_cnt, *cnts, *disps;
MPI_Aint true_extent, true_lb, extent;
void *tmp_buf;
MPI_Comm comm;
MPIU_CHKLMEM_DECL(3);
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
if (count == 0) goto fn_exit;
comm = comm_ptr->handle;
is_commutative = MPIR_Op_is_commutative(op);
#if defined(USE_SMP_COLLECTIVES)
/* is the op commutative? We do SMP optimizations only if it is. */
if (MPIR_Comm_is_node_aware(comm_ptr) && is_commutative) {
/* on each node, do a reduce to the local root */
if (comm_ptr->node_comm != NULL) {
/* take care of the MPI_IN_PLACE case. For reduce,
MPI_IN_PLACE is specified only on the root;
for allreduce it is specified on all processes. */
if ((sendbuf == MPI_IN_PLACE) && (comm_ptr->node_comm->rank != 0)) {
/* IN_PLACE and not root of reduce. Data supplied to this
allreduce is in recvbuf. Pass that as the sendbuf to reduce. */
mpi_errno = MPIR_Reduce_impl(recvbuf, NULL, count, datatype, op, 0, comm_ptr->node_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);
}
} else {
mpi_errno = MPIR_Reduce_impl(sendbuf, recvbuf, count, datatype, op, 0, comm_ptr->node_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);
}
}
} else {
/* only one process on the node. copy sendbuf to recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
/* now do an IN_PLACE allreduce among the local roots of all nodes */
if (comm_ptr->node_roots_comm != NULL) {
mpi_errno = allreduce_intra_or_coll_fn(MPI_IN_PLACE, recvbuf, count, datatype, op, comm_ptr->node_roots_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);
}
}
/* now broadcast the result among local processes */
if (comm_ptr->node_comm != NULL) {
mpi_errno = MPIR_Bcast_impl(recvbuf, count, datatype, 0, comm_ptr->node_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);
}
}
goto fn_exit;
}
#endif
is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
if (comm_ptr->is_hetero)
is_homogeneous = 0;
#endif
#ifdef MPID_HAS_HETERO
if (!is_homogeneous) {
/* heterogeneous. To get the same result on all processes, we
do a reduce to 0 and then broadcast. */
mpi_errno = MPIR_Reduce_impl ( sendbuf, recvbuf, count, datatype,
op, 0, comm_ptr, 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);
}
mpi_errno = MPIR_Bcast_impl( recvbuf, count, datatype, 0, comm_ptr, 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
#endif /* MPID_HAS_HETERO */
{
/* homogeneous */
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
is_commutative = MPIR_Op_is_commutative(op);
/* need to allocate temporary buffer to store incoming data*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(datatype, extent);
MPID_Ensure_Aint_fits_in_pointer(count * MPIR_MAX(extent, true_extent));
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, count*(MPIR_MAX(extent,true_extent)), mpi_errno, "temporary buffer");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
/* copy local data into recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf,
count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
MPID_Datatype_get_size_macro(datatype, type_size);
/* find nearest power-of-two less than or equal to comm_size */
pof2 = 1;
while (pof2 <= comm_size) pof2 <<= 1;
pof2 >>=1;
rem = comm_size - pof2;
/* In the non-power-of-two case, all even-numbered
processes of rank < 2*rem send their data to
(rank+1). These even-numbered processes no longer
participate in the algorithm until the very end. The
remaining processes form a nice power-of-two. */
if (rank < 2*rem) {
if (rank % 2 == 0) { /* even */
mpi_errno = MPIC_Send_ft(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_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);
}
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
doubling */
newrank = -1;
}
else { /* odd */
mpi_errno = MPIC_Recv_ft(tmp_buf, count,
datatype, rank-1,
MPIR_ALLREDUCE_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);
}
/* do the reduction on received data. since the
ordering is right, it doesn't matter whether
the operation is commutative or not. */
mpi_errno = MPIR_Reduce_local_impl(tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* change the rank */
newrank = rank / 2;
}
}
else /* rank >= 2*rem */
newrank = rank - rem;
/* If op is user-defined or count is less than pof2, use
recursive doubling algorithm. Otherwise do a reduce-scatter
followed by allgather. (If op is user-defined,
derived datatypes are allowed and the user could pass basic
datatypes on one process and derived on another as long as
the type maps are the same. Breaking up derived
datatypes to do the reduce-scatter is tricky, therefore
using recursive doubling in that case.) */
if (newrank != -1) {
if ((count*type_size <= MPIR_PARAM_ALLREDUCE_SHORT_MSG_SIZE) ||
(HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN) ||
(count < pof2)) { /* use recursive doubling */
mask = 0x1;
while (mask < pof2) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
/* Send the most current data, which is in recvbuf. Recv
into tmp_buf */
mpi_errno = MPIC_Sendrecv_ft(recvbuf, count, datatype,
dst, MPIR_ALLREDUCE_TAG, tmp_buf,
count, datatype, dst,
MPIR_ALLREDUCE_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);
}
/* tmp_buf contains data received in this step.
recvbuf contains data accumulated so far */
if (is_commutative || (dst < rank)) {
/* op is commutative OR the order is already right */
mpi_errno = MPIR_Reduce_local_impl(tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else {
/* op is noncommutative and the order is not right */
mpi_errno = MPIR_Reduce_local_impl(recvbuf, tmp_buf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* copy result back into recvbuf */
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
recvbuf, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
mask <<= 1;
}
}
else {
/* do a reduce-scatter followed by allgather */
/* for the reduce-scatter, calculate the count that
each process receives and the displacement within
the buffer */
MPIU_CHKLMEM_MALLOC(cnts, int *, pof2*sizeof(int), mpi_errno, "counts");
MPIU_CHKLMEM_MALLOC(disps, int *, pof2*sizeof(int), mpi_errno, "displacements");
for (i=0; i<(pof2-1); i++)
cnts[i] = count/pof2;
cnts[pof2-1] = count - (count/pof2)*(pof2-1);
disps[0] = 0;
for (i=1; i<pof2; i++)
disps[i] = disps[i-1] + cnts[i-1];
mask = 0x1;
send_idx = recv_idx = 0;
last_idx = pof2;
while (mask < pof2) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
send_idx = recv_idx + pof2/(mask*2);
for (i=send_idx; i<last_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += cnts[i];
}
else {
recv_idx = send_idx + pof2/(mask*2);
for (i=send_idx; i<recv_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += cnts[i];
}
/* printf("Rank %d, send_idx %d, recv_idx %d, send_cnt %d, recv_cnt %d, last_idx %d\n", newrank, send_idx, recv_idx,
send_cnt, recv_cnt, last_idx);
*/
/* Send data from recvbuf. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv_ft((char *) recvbuf +
disps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_ALLREDUCE_TAG,
(char *) tmp_buf +
disps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_ALLREDUCE_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);
}
/* tmp_buf contains data received in this step.
recvbuf contains data accumulated so far */
/* This algorithm is used only for predefined ops
and predefined ops are always commutative. */
mpi_errno = MPIR_Reduce_local_impl(((char *) tmp_buf + disps[recv_idx]*extent),
((char *) recvbuf + disps[recv_idx]*extent),
recv_cnt, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* update send_idx for next iteration */
send_idx = recv_idx;
mask <<= 1;
/* update last_idx, but not in last iteration
because the value is needed in the allgather
step below. */
if (mask < pof2)
last_idx = recv_idx + pof2/mask;
}
/* now do the allgather */
mask >>= 1;
while (mask > 0) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
/* update last_idx except on first iteration */
if (mask != pof2/2)
last_idx = last_idx + pof2/(mask*2);
recv_idx = send_idx + pof2/(mask*2);
for (i=send_idx; i<recv_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += cnts[i];
}
else {
recv_idx = send_idx - pof2/(mask*2);
for (i=send_idx; i<last_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += cnts[i];
}
mpi_errno = MPIC_Sendrecv_ft((char *) recvbuf +
disps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_ALLREDUCE_TAG,
(char *) recvbuf +
disps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_ALLREDUCE_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);
}
if (newrank > newdst) send_idx = recv_idx;
mask >>= 1;
}
}
}
/* In the non-power-of-two case, all odd-numbered
processes of rank < 2*rem send the result to
(rank-1), the ranks who didn't participate above. */
if (rank < 2*rem) {
if (rank % 2) /* odd */
mpi_errno = MPIC_Send_ft(recvbuf, count,
datatype, rank-1,
MPIR_ALLREDUCE_TAG, comm, errflag);
else /* even */
mpi_errno = MPIC_Recv_ft(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_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);
}
}
}
fn_exit:
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
MPIU_CHKLMEM_FREEALL();
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
return (mpi_errno);
fn_fail:
goto fn_exit;
}
int MPIR_Ireduce_scatter_sched_intra_recursive_halving(const void *sendbuf, void *recvbuf,
const int recvcounts[],
MPI_Datatype datatype, MPI_Op op,
MPIR_Comm * comm_ptr, MPIR_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int rank, comm_size, i;
MPI_Aint extent, true_extent, true_lb;
int *disps;
void *tmp_recvbuf, *tmp_results;
int type_size ATTRIBUTE((unused)), total_count, dst;
int mask;
int *newcnts, *newdisps, rem, newdst, send_idx, recv_idx, last_idx, send_cnt, recv_cnt;
int pof2, old_i, newrank;
MPIR_SCHED_CHKPMEM_DECL(5);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
MPIR_Datatype_get_extent_macro(datatype, extent);
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
#ifdef HAVE_ERROR_CHECKING
MPIR_Assert(MPIR_Op_is_commutative(op));
#endif
MPIR_SCHED_CHKPMEM_MALLOC(disps, int *, comm_size * sizeof(int), mpi_errno, "disps",
MPL_MEM_BUFFER);
total_count = 0;
for (i = 0; i < comm_size; i++) {
disps[i] = total_count;
total_count += recvcounts[i];
}
if (total_count == 0) {
goto fn_exit;
}
MPIR_Datatype_get_size_macro(datatype, type_size);
/* allocate temp. buffer to receive incoming data */
MPIR_SCHED_CHKPMEM_MALLOC(tmp_recvbuf, void *, total_count * (MPL_MAX(true_extent, extent)),
mpi_errno, "tmp_recvbuf", MPL_MEM_BUFFER);
/* adjust for potential negative lower bound in datatype */
tmp_recvbuf = (void *) ((char *) tmp_recvbuf - true_lb);
/* need to allocate another temporary buffer to accumulate
* results because recvbuf may not be big enough */
MPIR_SCHED_CHKPMEM_MALLOC(tmp_results, void *, total_count * (MPL_MAX(true_extent, extent)),
mpi_errno, "tmp_results", MPL_MEM_BUFFER);
/* adjust for potential negative lower bound in datatype */
tmp_results = (void *) ((char *) tmp_results - true_lb);
/* copy sendbuf into tmp_results */
if (sendbuf != MPI_IN_PLACE)
mpi_errno = MPIR_Sched_copy(sendbuf, total_count, datatype,
tmp_results, total_count, datatype, s);
else
mpi_errno = MPIR_Sched_copy(recvbuf, total_count, datatype,
tmp_results, total_count, datatype, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
pof2 = comm_ptr->pof2;
rem = comm_size - pof2;
/* In the non-power-of-two case, all even-numbered
* processes of rank < 2*rem send their data to
* (rank+1). These even-numbered processes no longer
* participate in the algorithm until the very end. The
* remaining processes form a nice power-of-two. */
if (rank < 2 * rem) {
if (rank % 2 == 0) { /* even */
mpi_errno = MPIR_Sched_send(tmp_results, total_count, datatype, rank + 1, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
/* temporarily set the rank to -1 so that this
* process does not pariticipate in recursive
* doubling */
newrank = -1;
} else { /* odd */
mpi_errno = MPIR_Sched_recv(tmp_recvbuf, total_count, datatype, rank - 1, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
/* do the reduction on received data. since the
* ordering is right, it doesn't matter whether
* the operation is commutative or not. */
mpi_errno = MPIR_Sched_reduce(tmp_recvbuf, tmp_results, total_count, datatype, op, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
/* change the rank */
newrank = rank / 2;
}
} else /* rank >= 2*rem */
newrank = rank - rem;
if (newrank != -1) {
/* recalculate the recvcounts and disps arrays because the
* even-numbered processes who no longer participate will
* have their result calculated by the process to their
* right (rank+1). */
MPIR_SCHED_CHKPMEM_MALLOC(newcnts, int *, pof2 * sizeof(int), mpi_errno, "newcnts",
MPL_MEM_BUFFER);
MPIR_SCHED_CHKPMEM_MALLOC(newdisps, int *, pof2 * sizeof(int), mpi_errno, "newdisps",
MPL_MEM_BUFFER);
for (i = 0; i < pof2; i++) {
/* what does i map to in the old ranking? */
old_i = (i < rem) ? i * 2 + 1 : i + rem;
if (old_i < 2 * rem) {
/* This process has to also do its left neighbor's
* work */
newcnts[i] = recvcounts[old_i] + recvcounts[old_i - 1];
} else
newcnts[i] = recvcounts[old_i];
}
newdisps[0] = 0;
for (i = 1; i < pof2; i++)
newdisps[i] = newdisps[i - 1] + newcnts[i - 1];
mask = pof2 >> 1;
send_idx = recv_idx = 0;
last_idx = pof2;
while (mask > 0) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst * 2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
send_idx = recv_idx + mask;
for (i = send_idx; i < last_idx; i++)
send_cnt += newcnts[i];
for (i = recv_idx; i < send_idx; i++)
recv_cnt += newcnts[i];
} else {
recv_idx = send_idx + mask;
for (i = send_idx; i < recv_idx; i++)
send_cnt += newcnts[i];
for (i = recv_idx; i < last_idx; i++)
recv_cnt += newcnts[i];
}
/* Send data from tmp_results. Recv into tmp_recvbuf */
{
/* avoid sending and receiving pointless 0-byte messages */
int send_dst = (send_cnt ? dst : MPI_PROC_NULL);
int recv_dst = (recv_cnt ? dst : MPI_PROC_NULL);
mpi_errno = MPIR_Sched_send(((char *) tmp_results + newdisps[send_idx] * extent),
send_cnt, datatype, send_dst, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_recv(((char *) tmp_recvbuf + newdisps[recv_idx] * extent),
recv_cnt, datatype, recv_dst, comm_ptr, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
/* tmp_recvbuf contains data received in this step.
* tmp_results contains data accumulated so far */
if (recv_cnt) {
mpi_errno = MPIR_Sched_reduce(((char *) tmp_recvbuf + newdisps[recv_idx] * extent),
((char *) tmp_results + newdisps[recv_idx] * extent),
recv_cnt, datatype, op, s);
MPIR_SCHED_BARRIER(s);
}
/* update send_idx for next iteration */
send_idx = recv_idx;
last_idx = recv_idx + mask;
mask >>= 1;
}
/* copy this process's result from tmp_results to recvbuf */
if (recvcounts[rank]) {
mpi_errno = MPIR_Sched_copy(((char *) tmp_results + disps[rank] * extent),
recvcounts[rank], datatype,
recvbuf, recvcounts[rank], datatype, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
}
int MPIR_Ireduce_scatter_sched_intra_recursive_doubling(const void *sendbuf, void *recvbuf, const int recvcounts[],
MPI_Datatype datatype, MPI_Op op, MPIR_Comm *comm_ptr,
MPIR_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int rank, comm_size, i;
MPI_Aint extent, true_extent, true_lb;
int *disps;
void *tmp_recvbuf, *tmp_results;
int type_size ATTRIBUTE((unused)), dis[2], blklens[2], total_count, dst;
int mask, dst_tree_root, my_tree_root, j, k;
int received;
MPI_Datatype sendtype, recvtype;
int nprocs_completed, tmp_mask, tree_root, is_commutative;
MPIR_SCHED_CHKPMEM_DECL(5);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
MPIR_Datatype_get_extent_macro(datatype, extent);
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
is_commutative = MPIR_Op_is_commutative(op);
MPIR_SCHED_CHKPMEM_MALLOC(disps, int *, comm_size * sizeof(int), mpi_errno, "disps", MPL_MEM_BUFFER);
total_count = 0;
for (i=0; i<comm_size; i++) {
disps[i] = total_count;
total_count += recvcounts[i];
}
if (total_count == 0) {
goto fn_exit;
}
MPIR_Datatype_get_size_macro(datatype, type_size);
/* total_count*extent eventually gets malloced. it isn't added to
* a user-passed in buffer */
MPIR_Ensure_Aint_fits_in_pointer(total_count * MPL_MAX(true_extent, extent));
/* need to allocate temporary buffer to receive incoming data*/
MPIR_SCHED_CHKPMEM_MALLOC(tmp_recvbuf, void *, total_count*(MPL_MAX(true_extent,extent)), mpi_errno, "tmp_recvbuf", MPL_MEM_BUFFER);
/* adjust for potential negative lower bound in datatype */
tmp_recvbuf = (void *)((char*)tmp_recvbuf - true_lb);
/* need to allocate another temporary buffer to accumulate
results */
MPIR_SCHED_CHKPMEM_MALLOC(tmp_results, void *, total_count*(MPL_MAX(true_extent,extent)), mpi_errno, "tmp_results", MPL_MEM_BUFFER);
/* adjust for potential negative lower bound in datatype */
tmp_results = (void *)((char*)tmp_results - true_lb);
/* copy sendbuf into tmp_results */
if (sendbuf != MPI_IN_PLACE)
mpi_errno = MPIR_Sched_copy(sendbuf, total_count, datatype,
tmp_results, total_count, datatype, s);
else
mpi_errno = MPIR_Sched_copy(recvbuf, total_count, datatype,
tmp_results, total_count, datatype, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
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;
/* At step 1, processes exchange (n-n/p) amount of
data; at step 2, (n-2n/p) amount of data; at step 3, (n-4n/p)
amount of data, and so forth. We use derived datatypes for this.
At each step, a process does not need to send data
indexed from my_tree_root to
my_tree_root+mask-1. Similarly, a process won't receive
data indexed from dst_tree_root to dst_tree_root+mask-1. */
/* calculate sendtype */
blklens[0] = blklens[1] = 0;
for (j=0; j<my_tree_root; j++)
blklens[0] += recvcounts[j];
for (j=my_tree_root+mask; j<comm_size; j++)
blklens[1] += recvcounts[j];
dis[0] = 0;
dis[1] = blklens[0];
for (j=my_tree_root; (j<my_tree_root+mask) && (j<comm_size); j++)
dis[1] += recvcounts[j];
mpi_errno = MPIR_Type_indexed_impl(2, blklens, dis, datatype, &sendtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Type_commit_impl(&sendtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* calculate recvtype */
blklens[0] = blklens[1] = 0;
for (j=0; j<dst_tree_root && j<comm_size; j++)
blklens[0] += recvcounts[j];
for (j=dst_tree_root+mask; j<comm_size; j++)
blklens[1] += recvcounts[j];
dis[0] = 0;
dis[1] = blklens[0];
for (j=dst_tree_root; (j<dst_tree_root+mask) && (j<comm_size); j++)
dis[1] += recvcounts[j];
mpi_errno = MPIR_Type_indexed_impl(2, blklens, dis, datatype, &recvtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Type_commit_impl(&recvtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
received = 0;
if (dst < comm_size) {
/* tmp_results contains data to be sent in each step. Data is
received in tmp_recvbuf and then accumulated into
tmp_results. accumulation is done later below. */
mpi_errno = MPIR_Sched_send(tmp_results, 1, sendtype, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_recv(tmp_recvbuf, 1, recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
received = 1;
}
/* 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 current result */
mpi_errno = MPIR_Sched_send(tmp_recvbuf, 1, recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
/* 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 = MPIR_Sched_recv(tmp_recvbuf, 1, recvtype, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
received = 1;
}
tmp_mask >>= 1;
k--;
}
}
/* N.B. The following comment comes from the FT version of
* MPI_Reduce_scatter. It does not currently apply to this code, but
* will in the future when we update the NBC code to be fault-tolerant
* in roughly the same fashion. [goodell@ 2011-03-03] */
/* The following reduction is done here instead of after
the MPIC_Sendrecv or MPIC_Recv above. This is
because to do it above, in the noncommutative
case, we would need an extra temp buffer so as not to
overwrite temp_recvbuf, because temp_recvbuf may have
to be communicated to other processes in the
non-power-of-two case. To avoid that extra allocation,
we do the reduce here. */
if (received) {
if (is_commutative || (dst_tree_root < my_tree_root)) {
mpi_errno = MPIR_Sched_reduce(tmp_recvbuf, tmp_results, blklens[0], datatype, op, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_reduce(((char *)tmp_recvbuf + dis[1]*extent),
((char *)tmp_results + dis[1]*extent),
blklens[1], datatype, op, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
else {
mpi_errno = MPIR_Sched_reduce(tmp_results, tmp_recvbuf, blklens[0], datatype, op, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_reduce(((char *)tmp_results + dis[1]*extent),
((char *)tmp_recvbuf + dis[1]*extent),
blklens[1], datatype, op, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
/* copy result back into tmp_results */
mpi_errno = MPIR_Sched_copy(tmp_recvbuf, 1, recvtype,
tmp_results, 1, recvtype, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
}
}
MPIR_Type_free_impl(&sendtype);
MPIR_Type_free_impl(&recvtype);
mask <<= 1;
i++;
}
int MPIR_Allgatherv_intra (
const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
const int *recvcounts,
const int *displs,
MPI_Datatype recvtype,
MPID_Comm *comm_ptr,
mpir_errflag_t *errflag )
{
int comm_size, rank, j, i, left, right;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
MPI_Aint recvbuf_extent, recvtype_extent, recvtype_true_extent,
recvtype_true_lb;
int curr_cnt, send_cnt, dst, total_count, recvtype_size, pof2, src, rem;
int recv_cnt;
void *tmp_buf;
int mask, dst_tree_root, my_tree_root, is_homogeneous, position,
send_offset, recv_offset, last_recv_cnt, nprocs_completed, k,
offset, tmp_mask, tree_root;
#ifdef MPID_HAS_HETERO
int tmp_buf_size, nbytes;
#endif
MPIU_CHKLMEM_DECL(1);
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
total_count = 0;
for (i=0; i<comm_size; i++)
total_count += recvcounts[i];
if (total_count == 0) goto fn_exit;
MPID_Datatype_get_extent_macro( recvtype, recvtype_extent );
MPID_Datatype_get_size_macro(recvtype, recvtype_size);
if ((total_count*recvtype_size < MPIR_CVAR_ALLGATHER_LONG_MSG_SIZE) &&
!(comm_size & (comm_size - 1))) {
/* Short or medium size message and power-of-two no. of processes. Use
* recursive doubling algorithm */
is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
if (comm_ptr->is_hetero)
is_homogeneous = 0;
#endif
if (is_homogeneous) {
/* need to receive contiguously into tmp_buf because
displs could make the recvbuf noncontiguous */
MPIR_Type_get_true_extent_impl(recvtype, &recvtype_true_lb, &recvtype_true_extent);
MPID_Ensure_Aint_fits_in_pointer(total_count *
(MPIR_MAX(recvtype_true_extent, recvtype_extent)));
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, total_count*(MPIR_MAX(recvtype_true_extent,recvtype_extent)), mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - recvtype_true_lb);
/* copy local data into right location in tmp_buf */
position = 0;
for (i=0; i<rank; i++) position += recvcounts[i];
if (sendbuf != MPI_IN_PLACE)
{
mpi_errno = MPIR_Localcopy(sendbuf, sendcount, sendtype,
((char *)tmp_buf + position*
recvtype_extent),
recvcounts[rank], recvtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else
{
/* if in_place specified, local data is found in recvbuf */
mpi_errno = MPIR_Localcopy(((char *)recvbuf +
displs[rank]*recvtype_extent),
recvcounts[rank], recvtype,
((char *)tmp_buf + position*
recvtype_extent),
recvcounts[rank], recvtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
curr_cnt = recvcounts[rank];
mask = 0x1;
i = 0;
while (mask < comm_size) {
dst = rank ^ mask;
/* find offset into send and recv buffers. zero out
the least significant "i" bits of rank and dst to
find root of src and dst subtrees. Use ranks of
roots as index to send from and recv into buffer */
dst_tree_root = dst >> i;
dst_tree_root <<= i;
my_tree_root = rank >> i;
my_tree_root <<= i;
if (dst < comm_size) {
send_offset = 0;
for (j=0; j<my_tree_root; j++)
send_offset += recvcounts[j];
recv_offset = 0;
for (j=0; j<dst_tree_root; j++)
recv_offset += recvcounts[j];
mpi_errno = MPIC_Sendrecv(((char *)tmp_buf + send_offset * recvtype_extent),
curr_cnt, recvtype, dst,
MPIR_ALLGATHERV_TAG,
((char *)tmp_buf + recv_offset * recvtype_extent),
total_count - recv_offset, recvtype, dst,
MPIR_ALLGATHERV_TAG,
comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
last_recv_cnt = 0;
} else
/* for convenience, recv is posted for a bigger amount
than will be sent */
MPIR_Get_count_impl(&status, recvtype, &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
data that they would normally have received from those
processes. That is, the haves in this subtree must send to
the havenots. We use a logarithmic
recursive-halfing algorithm for this. */
/* This part of the code will not currently be
executed because we are not using recursive
doubling for non power of two. Mark it as experimental
so that it doesn't show up as red in the coverage
tests. */
/* --BEGIN EXPERIMENTAL-- */
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)) {
offset = 0;
for (j=0; j<(my_tree_root+mask); j++)
offset += recvcounts[j];
offset *= recvtype_extent;
mpi_errno = MPIC_Send(((char *)tmp_buf + offset),
last_recv_cnt,
recvtype, dst,
MPIR_ALLGATHERV_TAG, comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* last_recv_cnt was set in the previous
receive. that's the amount of data to be
sent now. */
}
/* 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)) {
offset = 0;
for (j=0; j<(my_tree_root+mask); j++)
offset += recvcounts[j];
mpi_errno = MPIC_Recv(((char *)tmp_buf + offset * recvtype_extent),
total_count - offset, recvtype,
dst, MPIR_ALLGATHERV_TAG,
comm_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
last_recv_cnt = 0;
} else
/* for convenience, recv is posted for a
bigger amount than will be sent */
MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
tmp_mask >>= 1;
k--;
}
}
/* --END EXPERIMENTAL-- */
mask <<= 1;
i++;
}
/* copy data from tmp_buf to recvbuf */
position = 0;
for (j=0; j<comm_size; j++) {
if ((sendbuf != MPI_IN_PLACE) || (j != rank)) {
/* not necessary to copy if in_place and
j==rank. otherwise copy. */
mpi_errno = MPIR_Localcopy(((char *)tmp_buf + position*recvtype_extent),
recvcounts[j], recvtype,
((char *)recvbuf + displs[j]*recvtype_extent),
recvcounts[j], recvtype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
position += recvcounts[j];
}
}
int MPIDI_Get_accumulate(const void *origin_addr, int origin_count,
MPI_Datatype origin_datatype, void *result_addr, int result_count,
MPI_Datatype result_datatype, int target_rank, MPI_Aint target_disp,
int target_count, MPI_Datatype target_datatype, MPI_Op op, MPID_Win *win_ptr)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_msg_sz_t data_sz;
int rank, origin_predefined, result_predefined, target_predefined;
int shm_locked = 0;
int dt_contig ATTRIBUTE((unused));
MPI_Aint dt_true_lb ATTRIBUTE((unused));
MPID_Datatype *dtp;
MPIU_CHKLMEM_DECL(2);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_GET_ACCUMULATE);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_GET_ACCUMULATE);
if (target_rank == MPI_PROC_NULL) {
goto fn_exit;
}
if (win_ptr->epoch_state == MPIDI_EPOCH_NONE && win_ptr->fence_issued) {
win_ptr->epoch_state = MPIDI_EPOCH_FENCE;
}
MPIU_ERR_CHKANDJUMP(win_ptr->epoch_state == MPIDI_EPOCH_NONE,
mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
MPIDI_Datatype_get_info(target_count, target_datatype, dt_contig, data_sz,
dtp, dt_true_lb);
if (data_sz == 0) {
goto fn_exit;
}
rank = win_ptr->myrank;
origin_predefined = TRUE; /* quiet uninitialized warnings (b/c goto) */
if (op != MPI_NO_OP) {
MPIDI_CH3I_DATATYPE_IS_PREDEFINED(origin_datatype, origin_predefined);
}
MPIDI_CH3I_DATATYPE_IS_PREDEFINED(result_datatype, result_predefined);
MPIDI_CH3I_DATATYPE_IS_PREDEFINED(target_datatype, target_predefined);
/* Do =! rank first (most likely branch?) */
if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED)
{
MPI_User_function *uop;
void *base;
int disp_unit;
if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) {
base = win_ptr->shm_base_addrs[target_rank];
disp_unit = win_ptr->disp_units[target_rank];
MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr);
shm_locked = 1;
}
else {
base = win_ptr->base;
disp_unit = win_ptr->disp_unit;
}
/* Perform the local get first, then the accumulate */
mpi_errno = MPIR_Localcopy((char *) base + disp_unit * target_disp,
target_count, target_datatype,
result_addr, result_count, result_datatype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
/* NO_OP: Don't perform the accumulate */
if (op == MPI_NO_OP) {
if (shm_locked) {
MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr);
shm_locked = 0;
}
goto fn_exit;
}
if (op == MPI_REPLACE) {
mpi_errno = MPIR_Localcopy(origin_addr, origin_count, origin_datatype,
(char *) base + disp_unit * target_disp,
target_count, target_datatype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
if (shm_locked) {
MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr);
shm_locked = 0;
}
goto fn_exit;
}
MPIU_ERR_CHKANDJUMP1((HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN),
mpi_errno, MPI_ERR_OP, "**opnotpredefined",
"**opnotpredefined %d", op );
/* get the function by indexing into the op table */
uop = MPIR_OP_HDL_TO_FN(op);
if (origin_predefined && target_predefined) {
/* Cast away const'ness for origin_address in order to
* avoid changing the prototype for MPI_User_function */
(*uop)((void *) origin_addr, (char *) base + disp_unit*target_disp,
&target_count, &target_datatype);
}
else {
/* derived datatype */
MPID_Segment *segp;
DLOOP_VECTOR *dloop_vec;
MPI_Aint first, last;
int vec_len, i, type_size, count;
MPI_Datatype type;
MPI_Aint true_lb, true_extent, extent;
void *tmp_buf=NULL, *target_buf;
const void *source_buf;
if (origin_datatype != target_datatype) {
/* first copy the data into a temporary buffer with
the same datatype as the target. Then do the
accumulate operation. */
MPIR_Type_get_true_extent_impl(target_datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(target_datatype, extent);
MPIU_CHKLMEM_MALLOC(tmp_buf, void *,
target_count * (MPIR_MAX(extent,true_extent)),
mpi_errno, "temporary buffer");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
mpi_errno = MPIR_Localcopy(origin_addr, origin_count,
origin_datatype, tmp_buf,
target_count, target_datatype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
}
if (target_predefined) {
/* target predefined type, origin derived datatype */
(*uop)(tmp_buf, (char *) base + disp_unit * target_disp,
&target_count, &target_datatype);
}
else {
segp = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP1((!segp), mpi_errno, MPI_ERR_OTHER,
"**nomem","**nomem %s","MPID_Segment_alloc");
MPID_Segment_init(NULL, target_count, target_datatype, segp, 0);
first = 0;
last = SEGMENT_IGNORE_LAST;
MPID_Datatype_get_ptr(target_datatype, dtp);
vec_len = dtp->max_contig_blocks * target_count + 1;
/* +1 needed because Rob says so */
MPIU_CHKLMEM_MALLOC(dloop_vec, DLOOP_VECTOR *,
vec_len * sizeof(DLOOP_VECTOR),
mpi_errno, "dloop vector");
MPID_Segment_pack_vector(segp, first, &last, dloop_vec, &vec_len);
source_buf = (tmp_buf != NULL) ? tmp_buf : origin_addr;
target_buf = (char *) base + disp_unit * target_disp;
type = dtp->eltype;
type_size = MPID_Datatype_get_basic_size(type);
for (i=0; i<vec_len; i++) {
count = (dloop_vec[i].DLOOP_VECTOR_LEN)/type_size;
(*uop)((char *)source_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF),
(char *)target_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF),
&count, &type);
}
MPID_Segment_free(segp);
}
}
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);
}
}
int MPIR_Alltoall_intra(
const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
int recvcount,
MPI_Datatype recvtype,
MPID_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;
MPID_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_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(((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);
MPIU_ERR_SET(mpi_errno, *errflag, "**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(((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);
MPIU_ERR_SET(mpi_errno, *errflag, "**fail");
MPIU_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);
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(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);
MPIU_ERR_SET(mpi_errno, *errflag, "**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(((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_ptr, &status, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**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(((char *)tmp_buf +
dst_tree_root*sendbuf_extent),
last_recv_cnt, sendtype,
dst, MPIR_ALLTOALL_TAG,
comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIU_ERR_SET(mpi_errno, *errflag, "**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(((char *)tmp_buf +
dst_tree_root*sendbuf_extent),
sendbuf_extent*(comm_size-dst_tree_root),
sendtype,
dst, 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);
MPIU_ERR_SET(mpi_errno, *errflag, "**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++;
}
int MPIDI_Accumulate(void *origin_addr, int origin_count, MPI_Datatype
origin_datatype, int target_rank, MPI_Aint target_disp,
int target_count, MPI_Datatype target_datatype, MPI_Op op,
MPID_Win *win_ptr)
{
int mpi_errno=MPI_SUCCESS;
MPIDI_msg_sz_t data_sz;
int dt_contig, rank, origin_predefined, target_predefined;
MPI_Aint dt_true_lb;
MPIDI_RMA_ops *new_ptr;
MPID_Datatype *dtp;
MPIU_CHKLMEM_DECL(2);
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_ACCUMULATE);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_ACCUMULATE);
MPIDI_Datatype_get_info(origin_count, origin_datatype,
dt_contig, data_sz, dtp, dt_true_lb);
if ((data_sz == 0) || (target_rank == MPI_PROC_NULL))
{
goto fn_exit;
}
rank = win_ptr->myrank;
MPIDI_CH3I_DATATYPE_IS_PREDEFINED(origin_datatype, origin_predefined);
MPIDI_CH3I_DATATYPE_IS_PREDEFINED(target_datatype, target_predefined);
/* Do =! rank first (most likely branch?) */
if (target_rank == rank)
{
MPI_User_function *uop;
if (op == MPI_REPLACE)
{
mpi_errno = MPIR_Localcopy(origin_addr, origin_count,
origin_datatype,
(char *) win_ptr->base + win_ptr->disp_unit *
target_disp, target_count, target_datatype);
goto fn_exit;
}
MPIU_ERR_CHKANDJUMP1((HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN),
mpi_errno, MPI_ERR_OP, "**opnotpredefined",
"**opnotpredefined %d", op );
/* get the function by indexing into the op table */
uop = MPIR_Op_table[((op)&0xf) - 1];
if (origin_predefined && target_predefined)
{
(*uop)(origin_addr, (char *) win_ptr->base + win_ptr->disp_unit *
target_disp, &target_count, &target_datatype);
}
else
{
/* derived datatype */
MPID_Segment *segp;
DLOOP_VECTOR *dloop_vec;
MPI_Aint first, last;
int vec_len, i, type_size, count;
MPI_Datatype type;
MPI_Aint true_lb, true_extent, extent;
void *tmp_buf=NULL, *source_buf, *target_buf;
if (origin_datatype != target_datatype)
{
/* first copy the data into a temporary buffer with
the same datatype as the target. Then do the
accumulate operation. */
MPIR_Type_get_true_extent_impl(target_datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(target_datatype, extent);
MPIU_CHKLMEM_MALLOC(tmp_buf, void *,
target_count * (MPIR_MAX(extent,true_extent)),
mpi_errno, "temporary buffer");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
mpi_errno = MPIR_Localcopy(origin_addr, origin_count,
origin_datatype, tmp_buf,
target_count, target_datatype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
}
if (target_predefined) {
/* target predefined type, origin derived datatype */
(*uop)(tmp_buf, (char *) win_ptr->base + win_ptr->disp_unit *
target_disp, &target_count, &target_datatype);
}
else {
segp = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP1((!segp), mpi_errno, MPI_ERR_OTHER,
"**nomem","**nomem %s","MPID_Segment_alloc");
MPID_Segment_init(NULL, target_count, target_datatype, segp, 0);
first = 0;
last = SEGMENT_IGNORE_LAST;
MPID_Datatype_get_ptr(target_datatype, dtp);
vec_len = dtp->max_contig_blocks * target_count + 1;
/* +1 needed because Rob says so */
MPIU_CHKLMEM_MALLOC(dloop_vec, DLOOP_VECTOR *,
vec_len * sizeof(DLOOP_VECTOR),
mpi_errno, "dloop vector");
MPID_Segment_pack_vector(segp, first, &last, dloop_vec, &vec_len);
source_buf = (tmp_buf != NULL) ? tmp_buf : origin_addr;
target_buf = (char *) win_ptr->base +
win_ptr->disp_unit * target_disp;
type = dtp->eltype;
type_size = MPID_Datatype_get_basic_size(type);
for (i=0; i<vec_len; i++)
{
count = (dloop_vec[i].DLOOP_VECTOR_LEN)/type_size;
(*uop)((char *)source_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF),
(char *)target_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF),
&count, &type);
}
MPID_Segment_free(segp);
}
}
int MPIR_Exscan (
void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
MPI_Status status;
int rank, comm_size;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative, flag;
MPI_Aint true_extent, true_lb, extent;
void *partial_scan, *tmp_buf;
MPI_User_function *uop;
MPID_Op *op_ptr;
MPI_Comm comm;
MPIU_CHKLMEM_DECL(2);
MPIU_THREADPRIV_DECL;
#ifdef HAVE_CXX_BINDING
int is_cxx_uop = 0;
#endif
if (count == 0) return MPI_SUCCESS;
MPIU_THREADPRIV_GET;
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* set op_errno to 0. stored in perthread structure */
MPIU_THREADPRIV_FIELD(op_errno) = 0;
if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
is_commutative = 1;
/* get the function by indexing into the op table */
uop = MPIR_Op_table[op%16 - 1];
}
else {
MPID_Op_get_ptr(op, op_ptr);
if (op_ptr->kind == MPID_OP_USER_NONCOMMUTE)
is_commutative = 0;
else
is_commutative = 1;
#ifdef HAVE_CXX_BINDING
if (op_ptr->language == MPID_LANG_CXX) {
uop = (MPI_User_function *) op_ptr->function.c_function;
is_cxx_uop = 1;
}
else
#endif
if ((op_ptr->language == MPID_LANG_C))
uop = (MPI_User_function *) op_ptr->function.c_function;
else
uop = (MPI_User_function *) op_ptr->function.f77_function;
}
/* need to allocate temporary buffer to store partial scan*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro( datatype, extent );
MPIU_CHKLMEM_MALLOC(partial_scan, void *, (count*(MPIR_MAX(true_extent,extent))), mpi_errno, "partial_scan");
/* adjust for potential negative lower bound in datatype */
partial_scan = (void *)((char*)partial_scan - true_lb);
/* need to allocate temporary buffer to store incoming data*/
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, (count*(MPIR_MAX(true_extent,extent))), mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
mpi_errno = MPIR_Localcopy((sendbuf == MPI_IN_PLACE ? recvbuf : sendbuf), count, datatype,
partial_scan, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
flag = 0;
mask = 0x1;
while (mask < comm_size) {
dst = rank ^ mask;
if (dst < comm_size) {
/* Send partial_scan to dst. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv_ft(partial_scan, count, datatype,
dst, MPIR_EXSCAN_TAG, tmp_buf,
count, datatype, dst,
MPIR_EXSCAN_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);
}
if (rank > dst) {
call_uop(tmp_buf, partial_scan, count, datatype);
/* On rank 0, recvbuf is not defined. For sendbuf==MPI_IN_PLACE
recvbuf must not change (per MPI-2.2).
On rank 1, recvbuf is to be set equal to the value
in sendbuf on rank 0.
On others, recvbuf is the scan of values in the
sendbufs on lower ranks. */
if (rank != 0) {
if (flag == 0) {
/* simply copy data recd from rank 0 into recvbuf */
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
recvbuf, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
flag = 1;
}
else {
call_uop(tmp_buf, recvbuf, count, datatype);
}
}
}
else {
if (is_commutative) {
call_uop(tmp_buf, partial_scan, count, datatype);
}
else {
call_uop(partial_scan, tmp_buf, count, datatype);
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
partial_scan,
count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
}
mask <<= 1;
}
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
if (MPIU_THREADPRIV_FIELD(op_errno))
mpi_errno = MPIU_THREADPRIV_FIELD(op_errno);
fn_exit:
MPIU_CHKLMEM_FREEALL();
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag)
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Allreduce_intra_recursive_doubling(
const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPIR_Comm * comm_ptr,
MPIR_Errflag_t * errflag)
{
MPIR_CHKLMEM_DECL(1);
#ifdef MPID_HAS_HETERO
int is_homogeneous;
int rc;
#endif
int comm_size, rank;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative, pof2, newrank, rem, newdst;
MPI_Aint true_extent, true_lb, extent;
void *tmp_buf;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
is_commutative = MPIR_Op_is_commutative(op);
/* need to allocate temporary buffer to store incoming data*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPIR_Datatype_get_extent_macro(datatype, extent);
MPIR_Ensure_Aint_fits_in_pointer(count * MPL_MAX(extent, true_extent));
MPIR_CHKLMEM_MALLOC(tmp_buf, void *, count*(MPL_MAX(extent,true_extent)), mpi_errno, "temporary buffer", MPL_MEM_BUFFER);
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
/* copy local data into recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf,
count, datatype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
/* get nearest power-of-two less than or equal to comm_size */
pof2 = comm_ptr->pof2;
rem = comm_size - pof2;
/* In the non-power-of-two case, all even-numbered
processes of rank < 2*rem send their data to
(rank+1). These even-numbered processes no longer
participate in the algorithm until the very end. The
remaining processes form a nice power-of-two. */
if (rank < 2*rem) {
if (rank % 2 == 0) { /* even */
mpi_errno = MPIC_Send(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_TAG, comm_ptr, 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);
}
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
doubling */
newrank = -1;
}
else { /* odd */
mpi_errno = MPIC_Recv(tmp_buf, count,
datatype, rank-1,
MPIR_ALLREDUCE_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);
}
/* do the reduction on received data. since the
ordering is right, it doesn't matter whether
the operation is commutative or not. */
mpi_errno = MPIR_Reduce_local(tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* change the rank */
newrank = rank / 2;
}
}
else /* rank >= 2*rem */
newrank = rank - rem;
/* If op is user-defined or count is less than pof2, use
recursive doubling algorithm. Otherwise do a reduce-scatter
followed by allgather. (If op is user-defined,
derived datatypes are allowed and the user could pass basic
datatypes on one process and derived on another as long as
the type maps are the same. Breaking up derived
datatypes to do the reduce-scatter is tricky, therefore
using recursive doubling in that case.) */
if (newrank != -1) {
mask = 0x1;
while (mask < pof2) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
/* Send the most current data, which is in recvbuf. Recv
into tmp_buf */
mpi_errno = MPIC_Sendrecv(recvbuf, count, datatype,
dst, MPIR_ALLREDUCE_TAG, tmp_buf,
count, datatype, dst,
MPIR_ALLREDUCE_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);
}
/* tmp_buf contains data received in this step.
recvbuf contains data accumulated so far */
if (is_commutative || (dst < rank)) {
/* op is commutative OR the order is already right */
mpi_errno = MPIR_Reduce_local(tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else {
/* op is noncommutative and the order is not right */
mpi_errno = MPIR_Reduce_local(recvbuf, tmp_buf, count, datatype, op);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* copy result back into recvbuf */
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
recvbuf, count, datatype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
mask <<= 1;
}
}
/* In the non-power-of-two case, all odd-numbered
processes of rank < 2*rem send the result to
(rank-1), the ranks who didn't participate above. */
if (rank < 2*rem) {
if (rank % 2) /* odd */
mpi_errno = MPIC_Send(recvbuf, count,
datatype, rank-1,
MPIR_ALLREDUCE_TAG, comm_ptr, errflag);
else /* even */
mpi_errno = MPIC_Recv(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_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);
}
}
fn_exit:
MPIR_CHKLMEM_FREEALL();
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Ialltoall_bruck(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int i;
int nbytes, recvtype_size, recvbuf_extent, newtype_size;
int rank, comm_size;
void *tmp_buf = NULL;
MPI_Aint sendtype_extent, recvtype_extent, recvtype_true_lb, recvtype_true_extent;
int pof2, dst, src;
int count, block;
MPI_Datatype newtype;
int *displs;
MPIU_CHKLMEM_DECL(1); /* displs */
MPIR_SCHED_CHKPMEM_DECL(2); /* tmp_buf (2x) */
MPIU_Assert(sendbuf != MPI_IN_PLACE); /* we do not handle in-place */
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
MPID_Datatype_get_extent_macro(sendtype, sendtype_extent);
MPID_Datatype_get_size_macro(recvtype, recvtype_size);
MPID_Datatype_get_extent_macro(recvtype, recvtype_extent);
/* allocate temporary buffer */
/* must be same size as entire recvbuf for Phase 3 */
nbytes = recvtype_size * recvcount * comm_size;
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, nbytes, 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 = MPID_Sched_copy(((char *) sendbuf + rank*sendcount*sendtype_extent),
(comm_size - rank)*sendcount, sendtype,
recvbuf, (comm_size - rank)*recvcount, recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_Sched_copy(sendbuf, rank*sendcount, sendtype,
((char *) recvbuf + (comm_size-rank)*recvcount*recvtype_extent),
rank*recvcount, recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* 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);
MPID_Datatype_get_size_macro(newtype, newtype_size);
/* we will usually copy much less than nbytes */
mpi_errno = MPID_Sched_copy(recvbuf, 1, newtype, tmp_buf, newtype_size, MPI_BYTE, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* now send and recv in parallel */
mpi_errno = MPID_Sched_send(tmp_buf, newtype_size, MPI_BYTE, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_Sched_recv(recvbuf, 1, newtype, src, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
MPIR_Type_free_impl(&newtype);
pof2 *= 2;
}
/* Phase 3: 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));
/* not a leak, old tmp_buf value is still tracked by CHKPMEM macros */
MPIR_SCHED_CHKPMEM_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 = MPID_Sched_copy(((char *) recvbuf + (rank+1)*recvcount*recvtype_extent),
(comm_size - rank - 1)*recvcount, recvtype,
tmp_buf, (comm_size - rank - 1)*recvcount, recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_Sched_copy(recvbuf, (rank+1)*recvcount, recvtype,
((char *) tmp_buf + (comm_size-rank-1)*recvcount*recvtype_extent),
(rank+1)*recvcount, recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* 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 = MPID_Sched_copy(((char *) tmp_buf + i*recvcount*recvtype_extent),
recvcount, recvtype,
((char *) recvbuf + (comm_size-i-1)*recvcount*recvtype_extent),
recvcount, recvtype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
MPIR_SCHED_CHKPMEM_COMMIT(s);
fn_exit:
MPIU_CHKLMEM_FREEALL();
return mpi_errno;
fn_fail:
MPIR_SCHED_CHKPMEM_REAP(s);
goto fn_exit;
}
int MPIR_Iexscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int rank, comm_size;
int mask, dst, is_commutative, flag;
MPI_Aint true_extent, true_lb, extent;
void *partial_scan, *tmp_buf;
MPIR_SCHED_CHKPMEM_DECL(2);
if (count == 0)
goto fn_exit;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
is_commutative = MPIR_Op_is_commutative(op);
/* need to allocate temporary buffer to store partial scan*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(datatype, extent);
MPIR_SCHED_CHKPMEM_MALLOC(partial_scan, void *, (count*(MPIR_MAX(true_extent,extent))), mpi_errno, "partial_scan");
/* adjust for potential negative lower bound in datatype */
partial_scan = (void *)((char*)partial_scan - true_lb);
/* need to allocate temporary buffer to store incoming data*/
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, (count*(MPIR_MAX(true_extent,extent))), mpi_errno, "tmp_buf");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
mpi_errno = MPID_Sched_copy((sendbuf == MPI_IN_PLACE ? recvbuf : sendbuf), count, datatype,
partial_scan, count, datatype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
flag = 0;
mask = 0x1;
while (mask < comm_size) {
dst = rank ^ mask;
if (dst < comm_size) {
/* Send partial_scan to dst. Recv into tmp_buf */
mpi_errno = MPID_Sched_send(partial_scan, count, datatype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* sendrecv, no barrier here */
mpi_errno = MPID_Sched_recv(tmp_buf, count, datatype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
if (rank > dst) {
mpi_errno = MPID_Sched_reduce(tmp_buf, partial_scan, count, datatype, op, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* On rank 0, recvbuf is not defined. For sendbuf==MPI_IN_PLACE
recvbuf must not change (per MPI-2.2).
On rank 1, recvbuf is to be set equal to the value
in sendbuf on rank 0.
On others, recvbuf is the scan of values in the
sendbufs on lower ranks. */
if (rank != 0) {
if (flag == 0) {
/* simply copy data recd from rank 0 into recvbuf */
mpi_errno = MPID_Sched_copy(tmp_buf, count, datatype,
recvbuf, count, datatype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
flag = 1;
}
else {
mpi_errno = MPID_Sched_reduce(tmp_buf, recvbuf, count, datatype, op, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
}
}
else {
if (is_commutative) {
mpi_errno = MPID_Sched_reduce(tmp_buf, partial_scan, count, datatype, op, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
else {
mpi_errno = MPID_Sched_reduce(partial_scan, tmp_buf, count, datatype, op, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
mpi_errno = MPID_Sched_copy(tmp_buf, count, datatype,
partial_scan, count, datatype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
}
}
mask <<= 1;
}
MPIR_SCHED_CHKPMEM_COMMIT(s);
fn_exit:
return mpi_errno;
fn_fail:
MPIR_SCHED_CHKPMEM_REAP(s);
goto fn_exit;
}
/*@
MPI_Type_get_true_extent - Get the true lower bound and extent for a
datatype
Input Parameters:
. datatype - datatype to get information on (handle)
Output Parameters:
+ true_lb - true lower bound of datatype (address integer)
- true_extent - true size of datatype (address integer)
.N SignalSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_TYPE
.N MPI_ERR_ARG
@*/
int MPI_Type_get_true_extent(MPI_Datatype datatype, MPI_Aint * true_lb, MPI_Aint * true_extent)
{
int mpi_errno = MPI_SUCCESS;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_TYPE_GET_TRUE_EXTENT);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPI_TYPE_GET_TRUE_EXTENT);
/* Validate parameters, especially handles needing to be converted */
#ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
#endif
/* Validate parameters and objects (post conversion) */
#ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_Datatype *datatype_ptr = NULL;
/* Convert MPI object handles to object pointers */
MPIR_Datatype_get_ptr(datatype, datatype_ptr);
/* Validate datatype_ptr */
MPIR_Datatype_valid_ptr(datatype_ptr, mpi_errno);
if (mpi_errno)
goto fn_fail;
MPIR_ERRTEST_ARGNULL(true_lb, "true_lb", mpi_errno);
MPIR_ERRTEST_ARGNULL(true_extent, "true_extent", mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
#endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
MPIR_Type_get_true_extent_impl(datatype, true_lb, true_extent);
/* ... end of body of routine ... */
#ifdef HAVE_ERROR_CHECKING
fn_exit:
#endif
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPI_TYPE_GET_TRUE_EXTENT);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
#ifdef HAVE_ERROR_CHECKING
fn_fail:
{
mpi_errno =
MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER,
"**mpi_type_get_true_extent",
"**mpi_type_get_true_extent %D %p %p", datatype, true_lb,
true_extent);
}
mpi_errno = MPIR_Err_return_comm(NULL, FCNAME, mpi_errno);
goto fn_exit;
#endif
/* --END ERROR HANDLING-- */
}
int MPIR_Localcopy(const void *sendbuf, MPI_Aint sendcount, MPI_Datatype sendtype,
void *recvbuf, MPI_Aint recvcount, MPI_Datatype recvtype)
{
int mpi_errno = MPI_SUCCESS;
int sendtype_iscontig, recvtype_iscontig;
MPI_Aint sendsize, recvsize, sdata_sz, rdata_sz, copy_sz;
MPI_Aint true_extent, sendtype_true_lb, recvtype_true_lb;
MPIU_CHKLMEM_DECL(1);
MPID_MPI_STATE_DECL(MPID_STATE_MPIR_LOCALCOPY);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_LOCALCOPY);
MPID_Datatype_get_size_macro(sendtype, sendsize);
MPID_Datatype_get_size_macro(recvtype, recvsize);
sdata_sz = sendsize * sendcount;
rdata_sz = recvsize * recvcount;
/* if there is no data to copy, bail out */
if (!sdata_sz || !rdata_sz)
goto fn_exit;
#if defined(HAVE_ERROR_CHECKING)
if (sdata_sz > rdata_sz) {
MPIU_ERR_SET2(mpi_errno, MPI_ERR_TRUNCATE, "**truncate", "**truncate %d %d", sdata_sz, rdata_sz);
copy_sz = rdata_sz;
}
else
#endif /* HAVE_ERROR_CHECKING */
copy_sz = sdata_sz;
/* Builtin types is the common case; optimize for it */
if ((HANDLE_GET_KIND(sendtype) == HANDLE_KIND_BUILTIN) &&
HANDLE_GET_KIND(recvtype) == HANDLE_KIND_BUILTIN) {
MPIU_Memcpy(recvbuf, sendbuf, copy_sz);
goto fn_exit;
}
MPIR_Datatype_iscontig(sendtype, &sendtype_iscontig);
MPIR_Datatype_iscontig(recvtype, &recvtype_iscontig);
MPIR_Type_get_true_extent_impl(sendtype, &sendtype_true_lb, &true_extent);
MPIR_Type_get_true_extent_impl(recvtype, &recvtype_true_lb, &true_extent);
if (sendtype_iscontig && recvtype_iscontig)
{
#if defined(HAVE_ERROR_CHECKING)
MPIU_ERR_CHKMEMCPYANDJUMP(mpi_errno,
((char *)recvbuf + recvtype_true_lb),
((char *)sendbuf + sendtype_true_lb),
copy_sz);
#endif
MPIU_Memcpy(((char *) recvbuf + recvtype_true_lb),
((char *) sendbuf + sendtype_true_lb),
copy_sz);
}
else if (sendtype_iscontig)
{
MPID_Segment seg;
MPI_Aint last;
MPID_Segment_init(recvbuf, recvcount, recvtype, &seg, 0);
last = copy_sz;
MPID_Segment_unpack(&seg, 0, &last, (char*)sendbuf + sendtype_true_lb);
MPIU_ERR_CHKANDJUMP(last != copy_sz, mpi_errno, MPI_ERR_TYPE, "**dtypemismatch");
}
else if (recvtype_iscontig)
{
MPID_Segment seg;
MPI_Aint last;
MPID_Segment_init(sendbuf, sendcount, sendtype, &seg, 0);
last = copy_sz;
MPID_Segment_pack(&seg, 0, &last, (char*)recvbuf + recvtype_true_lb);
MPIU_ERR_CHKANDJUMP(last != copy_sz, mpi_errno, MPI_ERR_TYPE, "**dtypemismatch");
}
else
{
char * buf;
MPIDI_msg_sz_t buf_off;
MPID_Segment sseg;
MPIDI_msg_sz_t sfirst;
MPID_Segment rseg;
MPIDI_msg_sz_t rfirst;
MPIU_CHKLMEM_MALLOC(buf, char *, COPY_BUFFER_SZ, mpi_errno, "buf");
MPID_Segment_init(sendbuf, sendcount, sendtype, &sseg, 0);
MPID_Segment_init(recvbuf, recvcount, recvtype, &rseg, 0);
sfirst = 0;
rfirst = 0;
buf_off = 0;
while (1)
{
MPI_Aint last;
char * buf_end;
if (copy_sz - sfirst > COPY_BUFFER_SZ - buf_off)
{
last = sfirst + (COPY_BUFFER_SZ - buf_off);
}
else
{
last = copy_sz;
}
MPID_Segment_pack(&sseg, sfirst, &last, buf + buf_off);
MPIU_Assert(last > sfirst);
buf_end = buf + buf_off + (last - sfirst);
sfirst = last;
MPID_Segment_unpack(&rseg, rfirst, &last, buf);
MPIU_Assert(last > rfirst);
rfirst = last;
if (rfirst == copy_sz)
{
/* successful completion */
break;
}
/* if the send side finished, but the recv side couldn't unpack it, there's a datatype mismatch */
MPIU_ERR_CHKANDJUMP(sfirst == copy_sz, mpi_errno, MPI_ERR_TYPE, "**dtypemismatch");
/* if not all data was unpacked, copy it to the front of the buffer for next time */
buf_off = sfirst - rfirst;
if (buf_off > 0)
{
memmove(buf, buf_end - buf_off, buf_off);
}
}
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_LOCALCOPY);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Allgather_intra_brucks (
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, rank;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
MPI_Aint recvtype_extent;
MPI_Aint recvtype_true_extent, recvbuf_extent, recvtype_true_lb;
int pof2, src, rem;
void *tmp_buf = NULL;
int curr_cnt, dst;
MPIR_CHKLMEM_DECL(1);
if (((sendcount == 0) && (sendbuf != MPI_IN_PLACE)) || (recvcount == 0))
return MPI_SUCCESS;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
MPIR_Datatype_get_extent_macro( recvtype, recvtype_extent );
/* This is the largest offset we add to recvbuf */
MPIR_Ensure_Aint_fits_in_pointer(MPIR_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
(comm_size * recvcount * recvtype_extent));
/* allocate 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", MPL_MEM_BUFFER);
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - recvtype_true_lb);
/* copy local data to the top of tmp_buf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy (sendbuf, sendcount, sendtype,
tmp_buf, recvcount, recvtype);
if (mpi_errno) {
MPIR_ERR_POP(mpi_errno);
}
}
else {
mpi_errno = MPIR_Localcopy (((char *)recvbuf +
rank * recvcount * recvtype_extent),
recvcount, recvtype, tmp_buf,
recvcount, recvtype);
if (mpi_errno) {
MPIR_ERR_POP(mpi_errno);
}
}
/* do the first \floor(\lg p) steps */
curr_cnt = recvcount;
pof2 = 1;
while (pof2 <= comm_size/2) {
src = (rank + pof2) % comm_size;
dst = (rank - pof2 + comm_size) % comm_size;
mpi_errno = MPIC_Sendrecv(tmp_buf, curr_cnt, recvtype, dst,
MPIR_ALLGATHER_TAG,
((char *)tmp_buf + curr_cnt*recvtype_extent),
curr_cnt, recvtype,
src, MPIR_ALLGATHER_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);
}
curr_cnt *= 2;
pof2 *= 2;
}
/* if comm_size is not a power of two, one more step is needed */
rem = comm_size - pof2;
if (rem) {
src = (rank + pof2) % comm_size;
dst = (rank - pof2 + comm_size) % comm_size;
mpi_errno = MPIC_Sendrecv(tmp_buf, rem * recvcount, recvtype,
dst, MPIR_ALLGATHER_TAG,
((char *)tmp_buf + curr_cnt*recvtype_extent),
rem * recvcount, recvtype,
src, MPIR_ALLGATHER_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);
}
}
/* Rotate blocks in tmp_buf down by (rank) blocks and store
* result in recvbuf. */
mpi_errno = MPIR_Localcopy(tmp_buf, (comm_size-rank)*recvcount,
recvtype, (char *) recvbuf + rank*recvcount*recvtype_extent,
(comm_size-rank)*recvcount, recvtype);
if (mpi_errno) {
MPIR_ERR_POP(mpi_errno);
}
if (rank) {
mpi_errno = MPIR_Localcopy((char *) tmp_buf +
(comm_size-rank)*recvcount*recvtype_extent,
rank*recvcount, recvtype, recvbuf,
rank*recvcount, recvtype);
if (mpi_errno) {
MPIR_ERR_POP(mpi_errno);
}
}
fn_exit:
MPIR_CHKLMEM_FREEALL();
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
else if (*errflag != MPIR_ERR_NONE)
MPIR_ERR_SET(mpi_errno, *errflag, "**coll_fail");
return mpi_errno;
fn_fail:
goto fn_exit;
}
/* Algorithm: Recursive halving
*
* This is a recursive-halving algorithm in which the first p/2 processes send
* the second n/2 data to their counterparts in the other half and receive the
* first n/2 data from them. This procedure continues recursively, halving the
* data communicated at each step, for a total of lgp steps. If the number of
* processes is not a power-of-two, we convert it to the nearest lower
* power-of-two by having the first few even-numbered processes send their data
* to the neighboring odd-numbered process at (rank+1). Those odd-numbered
* processes compute the result for their left neighbor as well in the
* recursive halving algorithm, and then at the end send the result back to
* the processes that didn't participate. Therefore, if p is a power-of-two:
*
* Cost = lgp.alpha + n.((p-1)/p).beta + n.((p-1)/p).gamma
*
* If p is not a power-of-two:
*
* Cost = (floor(lgp)+2).alpha + n.(1+(p-1+n)/p).beta + n.(1+(p-1)/p).gamma
*
* The above cost in the non power-of-two case is approximate because there is
* some imbalance in the amount of work each process does because some
* processes do the work of their neighbors as well.
*/
int MPIR_Reduce_scatter_block_intra_recursive_halving (
const void *sendbuf,
void *recvbuf,
int recvcount,
MPI_Datatype datatype,
MPI_Op op,
MPIR_Comm *comm_ptr,
MPIR_Errflag_t *errflag )
{
int rank, comm_size, i;
MPI_Aint extent, true_extent, true_lb;
int *disps;
void *tmp_recvbuf, *tmp_results;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int total_count, dst;
int mask;
int *newcnts, *newdisps, rem, newdst, send_idx, recv_idx,
last_idx, send_cnt, recv_cnt;
int pof2, old_i, newrank;
MPIR_CHKLMEM_DECL(5);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
#ifdef HAVE_ERROR_CHECKING
{
int is_commutative;
is_commutative = MPIR_Op_is_commutative(op);
MPIR_Assert(is_commutative);
}
#endif /* HAVE_ERROR_CHECKING */
/* set op_errno to 0. stored in perthread structure */
{
MPIR_Per_thread_t *per_thread = NULL;
int err = 0;
MPID_THREADPRIV_KEY_GET_ADDR(MPIR_ThreadInfo.isThreaded, MPIR_Per_thread_key,
MPIR_Per_thread, per_thread, &err);
MPIR_Assert(err == 0);
per_thread->op_errno = 0;
}
if (recvcount == 0) {
goto fn_exit;
}
MPIR_Datatype_get_extent_macro(datatype, extent);
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPIR_CHKLMEM_MALLOC(disps, int *, comm_size * sizeof(int), mpi_errno, "disps", MPL_MEM_BUFFER);
total_count = comm_size*recvcount;
for (i=0; i<comm_size; i++) {
disps[i] = i*recvcount;
}
/* total_count*extent eventually gets malloced. it isn't added to
* a user-passed in buffer */
MPIR_Ensure_Aint_fits_in_pointer(total_count * MPL_MAX(true_extent, extent));
/* commutative and short. use recursive halving algorithm */
/* allocate temp. buffer to receive incoming data */
MPIR_CHKLMEM_MALLOC(tmp_recvbuf, void *, total_count*(MPL_MAX(true_extent,extent)), mpi_errno, "tmp_recvbuf", MPL_MEM_BUFFER);
/* adjust for potential negative lower bound in datatype */
tmp_recvbuf = (void *)((char*)tmp_recvbuf - true_lb);
/* need to allocate another temporary buffer to accumulate
results because recvbuf may not be big enough */
MPIR_CHKLMEM_MALLOC(tmp_results, void *, total_count*(MPL_MAX(true_extent,extent)), mpi_errno, "tmp_results", MPL_MEM_BUFFER);
/* adjust for potential negative lower bound in datatype */
tmp_results = (void *)((char*)tmp_results - true_lb);
/* copy sendbuf into tmp_results */
if (sendbuf != MPI_IN_PLACE)
mpi_errno = MPIR_Localcopy(sendbuf, total_count, datatype,
tmp_results, total_count, datatype);
else
mpi_errno = MPIR_Localcopy(recvbuf, total_count, datatype,
tmp_results, total_count, datatype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
pof2 = comm_ptr->pof2;
rem = comm_size - pof2;
/* In the non-power-of-two case, all even-numbered
processes of rank < 2*rem send their data to
(rank+1). These even-numbered processes no longer
participate in the algorithm until the very end. The
remaining processes form a nice power-of-two. */
if (rank < 2*rem) {
if (rank % 2 == 0) { /* even */
mpi_errno = MPIC_Send(tmp_results, total_count,
datatype, rank+1,
MPIR_REDUCE_SCATTER_BLOCK_TAG, comm_ptr, 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);
}
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
doubling */
newrank = -1;
}
else { /* odd */
mpi_errno = MPIC_Recv(tmp_recvbuf, total_count,
datatype, rank-1,
MPIR_REDUCE_SCATTER_BLOCK_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);
}
/* do the reduction on received data. since the
ordering is right, it doesn't matter whether
the operation is commutative or not. */
mpi_errno = MPIR_Reduce_local( tmp_recvbuf, tmp_results,
total_count, datatype, op);
/* change the rank */
newrank = rank / 2;
}
}
else /* rank >= 2*rem */
newrank = rank - rem;
if (newrank != -1) {
/* recalculate the recvcnts and disps arrays because the
even-numbered processes who no longer participate will
have their result calculated by the process to their
right (rank+1). */
MPIR_CHKLMEM_MALLOC(newcnts, int *, pof2*sizeof(int), mpi_errno, "newcnts", MPL_MEM_BUFFER);
MPIR_CHKLMEM_MALLOC(newdisps, int *, pof2*sizeof(int), mpi_errno, "newdisps", MPL_MEM_BUFFER);
for (i=0; i<pof2; i++) {
/* what does i map to in the old ranking? */
old_i = (i < rem) ? i*2 + 1 : i + rem;
if (old_i < 2*rem) {
/* This process has to also do its left neighbor's
work */
newcnts[i] = 2 * recvcount;
}
else
newcnts[i] = recvcount;
}
newdisps[0] = 0;
for (i=1; i<pof2; i++)
newdisps[i] = newdisps[i-1] + newcnts[i-1];
mask = pof2 >> 1;
send_idx = recv_idx = 0;
last_idx = pof2;
while (mask > 0) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
send_idx = recv_idx + mask;
for (i=send_idx; i<last_idx; i++)
send_cnt += newcnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += newcnts[i];
}
else {
recv_idx = send_idx + mask;
for (i=send_idx; i<recv_idx; i++)
send_cnt += newcnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += newcnts[i];
}
/* printf("Rank %d, send_idx %d, recv_idx %d, send_cnt %d, recv_cnt %d, last_idx %d\n", newrank, send_idx, recv_idx,
send_cnt, recv_cnt, last_idx);
*/
/* Send data from tmp_results. Recv into tmp_recvbuf */
if ((send_cnt != 0) && (recv_cnt != 0))
mpi_errno = MPIC_Sendrecv((char *) tmp_results +
newdisps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_REDUCE_SCATTER_BLOCK_TAG,
(char *) tmp_recvbuf +
newdisps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_REDUCE_SCATTER_BLOCK_TAG, comm_ptr,
MPI_STATUS_IGNORE, errflag);
else if ((send_cnt == 0) && (recv_cnt != 0))
mpi_errno = MPIC_Recv((char *) tmp_recvbuf +
newdisps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_REDUCE_SCATTER_BLOCK_TAG, comm_ptr,
MPI_STATUS_IGNORE, errflag);
else if ((recv_cnt == 0) && (send_cnt != 0))
mpi_errno = MPIC_Send((char *) tmp_results +
newdisps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_REDUCE_SCATTER_BLOCK_TAG,
comm_ptr, 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);
}
/* tmp_recvbuf contains data received in this step.
tmp_results contains data accumulated so far */
if (recv_cnt) {
mpi_errno = MPIR_Reduce_local(
(char *) tmp_recvbuf + newdisps[recv_idx]*extent,
(char *) tmp_results + newdisps[recv_idx]*extent,
recv_cnt, datatype, op);
}
/* update send_idx for next iteration */
send_idx = recv_idx;
last_idx = recv_idx + mask;
mask >>= 1;
}
/* copy this process's result from tmp_results to recvbuf */
mpi_errno = MPIR_Localcopy((char *)tmp_results +
disps[rank]*extent,
recvcount, datatype, recvbuf,
recvcount, datatype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
