int Coll_alltoallv_ring::alltoallv(const void* send_buff, const int* send_counts, const int* send_disps, MPI_Datatype send_type,
void* recv_buff, const int* recv_counts, const int* recv_disps, MPI_Datatype recv_type,
MPI_Comm comm)
{
MPI_Status s;
MPI_Aint send_chunk, recv_chunk;
int i, src, dst, rank, num_procs;
int tag = COLL_TAG_ALLTOALLV;
char *send_ptr = (char *) send_buff;
char *recv_ptr = (char *) recv_buff;
rank = comm->rank();
num_procs = comm->size();
send_chunk = send_type->get_extent();
recv_chunk = recv_type->get_extent();
int pof2 = ((num_procs != 0) && ((num_procs & (~num_procs + 1)) == num_procs));
for (i = 0; i < num_procs; i++) {
if (pof2 == 1) {
/* use exclusive-or algorithm */
src = dst = rank ^ i;
} else {
src = (rank - i + num_procs) % num_procs;
dst = (rank + i) % num_procs;
}
Request::sendrecv(send_ptr + send_disps[dst] * send_chunk, send_counts[dst], send_type, dst,
tag, recv_ptr + recv_disps[src] * recv_chunk, recv_counts[src], recv_type,
src, tag, comm, &s);
}
return MPI_SUCCESS;
}
int PMPI_Ssend(const void* buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) {
int retval = 0;
smpi_bench_end();
if (comm == MPI_COMM_NULL) {
retval = MPI_ERR_COMM;
} else if (dst == MPI_PROC_NULL) {
retval = MPI_SUCCESS;
} else if (dst >= comm->group()->size() || dst <0){
retval = MPI_ERR_RANK;
} else if ((count < 0) || (buf==nullptr && count > 0)) {
retval = MPI_ERR_COUNT;
} else if (datatype==MPI_DATATYPE_NULL || not datatype->is_valid()) {
retval = MPI_ERR_TYPE;
} else if(tag<0 && tag != MPI_ANY_TAG){
retval = MPI_ERR_TAG;
} else {
int my_proc_id = simgrid::s4u::this_actor::get_pid();
int dst_traced = getPid(comm, dst);
TRACE_smpi_comm_in(my_proc_id, __func__,
new simgrid::instr::Pt2PtTIData("Ssend", dst,
datatype->is_replayable() ? count : count * datatype->size(),
tag, simgrid::smpi::Datatype::encode(datatype)));
TRACE_smpi_send(my_proc_id, my_proc_id, dst_traced, tag, count * datatype->size());
simgrid::smpi::Request::ssend(buf, count, datatype, dst, tag, comm);
retval = MPI_SUCCESS;
TRACE_smpi_comm_out(my_proc_id);
}
smpi_bench_begin();
return retval;
}
int Coll_alltoallv_pair::alltoallv(void *send_buff, int *send_counts, int *send_disps,
MPI_Datatype send_type,
void *recv_buff, int *recv_counts, int *recv_disps,
MPI_Datatype recv_type, MPI_Comm comm)
{
MPI_Aint send_chunk, recv_chunk;
MPI_Status s;
int i, src, dst, rank, num_procs;
int tag = COLL_TAG_ALLTOALLV;
char *send_ptr = (char *) send_buff;
char *recv_ptr = (char *) recv_buff;
rank = comm->rank();
num_procs = comm->size();
if((num_procs&(num_procs-1)))
THROWF(arg_error,0, "alltoallv pair algorithm can't be used with non power of two number of processes ! ");
send_chunk = send_type->get_extent();
recv_chunk = recv_type->get_extent();
for (i = 0; i < num_procs; i++) {
src = dst = rank ^ i;
Request::sendrecv(send_ptr + send_disps[dst] * send_chunk, send_counts[dst], send_type, dst, tag,
recv_ptr + recv_disps[src] * recv_chunk, recv_counts[src], recv_type, src, tag, comm, &s);
}
return MPI_SUCCESS;
}
int PMPI_Compare_and_swap(const void* origin_addr, void* compare_addr, void* result_addr, MPI_Datatype datatype,
int target_rank, MPI_Aint target_disp, MPI_Win win)
{
int retval = 0;
smpi_bench_end();
if (win == MPI_WIN_NULL) {
retval = MPI_ERR_WIN;
} else if (target_rank == MPI_PROC_NULL) {
retval = MPI_SUCCESS;
} else if (target_rank <0){
retval = MPI_ERR_RANK;
} else if (win->dynamic()==0 && target_disp <0){
//in case of dynamic window, target_disp can be mistakenly seen as negative, as it is an address
retval = MPI_ERR_ARG;
} else if (origin_addr==nullptr || result_addr==nullptr || compare_addr==nullptr){
retval = MPI_ERR_COUNT;
} else if ((datatype == MPI_DATATYPE_NULL) || (not datatype->is_valid())) {
retval = MPI_ERR_TYPE;
} else {
int my_proc_id = simgrid::s4u::this_actor::get_pid();
MPI_Group group;
win->get_group(&group);
TRACE_smpi_comm_in(my_proc_id, __func__,
new simgrid::instr::Pt2PtTIData("Compare_and_swap", target_rank,
datatype->is_replayable() ? 1 : datatype->size(),
simgrid::smpi::Datatype::encode(datatype)));
retval = win->compare_and_swap(origin_addr, compare_addr, result_addr, datatype, target_rank, target_disp);
TRACE_smpi_comm_out(my_proc_id);
}
smpi_bench_begin();
return retval;
}
int PMPI_Sendrecv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, int dst, int sendtag, void* recvbuf,
int recvcount, MPI_Datatype recvtype, int src, int recvtag, MPI_Comm comm, MPI_Status* status)
{
int retval = 0;
smpi_bench_end();
if (comm == MPI_COMM_NULL) {
retval = MPI_ERR_COMM;
} else if (not sendtype->is_valid() || not recvtype->is_valid()) {
retval = MPI_ERR_TYPE;
} else if (src == MPI_PROC_NULL) {
if(status!=MPI_STATUS_IGNORE){
simgrid::smpi::Status::empty(status);
status->MPI_SOURCE = MPI_PROC_NULL;
}
if(dst != MPI_PROC_NULL)
simgrid::smpi::Request::send(sendbuf, sendcount, sendtype, dst, sendtag, comm);
retval = MPI_SUCCESS;
}else if (dst == MPI_PROC_NULL){
simgrid::smpi::Request::recv(recvbuf, recvcount, recvtype, src, recvtag, comm, status);
retval = MPI_SUCCESS;
}else if (dst >= comm->group()->size() || dst <0 ||
(src!=MPI_ANY_SOURCE && (src >= comm->group()->size() || src <0))){
retval = MPI_ERR_RANK;
} else if ((sendcount < 0 || recvcount<0) ||
(sendbuf==nullptr && sendcount > 0) || (recvbuf==nullptr && recvcount>0)) {
retval = MPI_ERR_COUNT;
} else if((sendtag<0 && sendtag != MPI_ANY_TAG)||(recvtag<0 && recvtag != MPI_ANY_TAG)){
retval = MPI_ERR_TAG;
} else {
int my_proc_id = simgrid::s4u::this_actor::get_pid();
int dst_traced = getPid(comm, dst);
int src_traced = getPid(comm, src);
// FIXME: Hack the way to trace this one
std::vector<int>* dst_hack = new std::vector<int>;
std::vector<int>* src_hack = new std::vector<int>;
dst_hack->push_back(dst_traced);
src_hack->push_back(src_traced);
TRACE_smpi_comm_in(my_proc_id, __func__,
new simgrid::instr::VarCollTIData(
"sendRecv", -1, sendtype->is_replayable() ? sendcount : sendcount * sendtype->size(),
dst_hack, recvtype->is_replayable() ? recvcount : recvcount * recvtype->size(), src_hack,
simgrid::smpi::Datatype::encode(sendtype), simgrid::smpi::Datatype::encode(recvtype)));
TRACE_smpi_send(my_proc_id, my_proc_id, dst_traced, sendtag, sendcount * sendtype->size());
simgrid::smpi::Request::sendrecv(sendbuf, sendcount, sendtype, dst, sendtag, recvbuf, recvcount, recvtype, src,
recvtag, comm, status);
retval = MPI_SUCCESS;
TRACE_smpi_recv(src_traced, my_proc_id, recvtag);
TRACE_smpi_comm_out(my_proc_id);
}
smpi_bench_begin();
return retval;
}
int Coll_gather_ompi::gather(const void *sbuf, int scount,
MPI_Datatype sdtype,
void* rbuf, int rcount,
MPI_Datatype rdtype,
int root,
MPI_Comm comm
)
{
//const int large_segment_size = 32768;
//const int small_segment_size = 1024;
//const size_t large_block_size = 92160;
const size_t intermediate_block_size = 6000;
const size_t small_block_size = 1024;
const int large_communicator_size = 60;
const int small_communicator_size = 10;
int communicator_size, rank;
size_t dsize, block_size;
XBT_DEBUG("smpi_coll_tuned_gather_ompi");
communicator_size = comm->size();
rank = comm->rank();
// Determine block size
if (rank == root) {
dsize = rdtype->size();
block_size = dsize * rcount;
} else {
dsize = sdtype->size();
block_size = dsize * scount;
}
/* if (block_size > large_block_size) {*/
/* return smpi_coll_tuned_gather_ompi_linear_sync (sbuf, scount, sdtype, */
/* rbuf, rcount, rdtype, */
/* root, comm);*/
/* } else*/ if (block_size > intermediate_block_size) {
return Coll_gather_ompi_linear_sync::gather (sbuf, scount, sdtype,
rbuf, rcount, rdtype,
root, comm);
} else if ((communicator_size > large_communicator_size) ||
((communicator_size > small_communicator_size) &&
(block_size < small_block_size))) {
return Coll_gather_ompi_binomial::gather (sbuf, scount, sdtype,
rbuf, rcount, rdtype,
root, comm);
}
// Otherwise, use basic linear
return Coll_gather_ompi_basic_linear::gather (sbuf, scount, sdtype,
rbuf, rcount, rdtype,
root, comm);
}
int PMPI_Pack_size(int incount, MPI_Datatype datatype, MPI_Comm comm, int* size) {
if(incount<0){
return MPI_ERR_COUNT;
} else if (datatype == MPI_DATATYPE_NULL || not datatype->is_valid()){
return MPI_ERR_TYPE;
} else if(comm==MPI_COMM_NULL){
return MPI_ERR_COMM;
} else {
*size=incount*datatype->size();
return MPI_SUCCESS;
}
}
int PMPI_Pack(const void* inbuf, int incount, MPI_Datatype type, void* outbuf, int outcount, int* position, MPI_Comm comm) {
if(incount<0){
return MPI_ERR_COUNT;
} else if(inbuf==nullptr || outbuf==nullptr || outcount < 0){
return MPI_ERR_ARG;
} else if (type == MPI_DATATYPE_NULL || not type->is_valid()){
return MPI_ERR_TYPE;
} else if(comm==MPI_COMM_NULL){
return MPI_ERR_COMM;
} else {
return type->pack(inbuf == MPI_BOTTOM ? nullptr : inbuf, incount, outbuf, outcount, position, comm);
}
}
int Coll_alltoall_basic_linear::alltoall(void *sendbuf, int sendcount, MPI_Datatype sendtype,
void *recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm)
{
int system_tag = 888;
int i;
int count;
MPI_Aint lb = 0, sendext = 0, recvext = 0;
MPI_Request *requests;
/* Initialize. */
int rank = comm->rank();
int size = comm->size();
XBT_DEBUG("<%d> algorithm alltoall_basic_linear() called.", rank);
sendtype->extent(&lb, &sendext);
recvtype->extent(&lb, &recvext);
/* simple optimization */
int err = Datatype::copy(static_cast<char *>(sendbuf) + rank * sendcount * sendext, sendcount, sendtype,
static_cast<char *>(recvbuf) + rank * recvcount * recvext, recvcount, recvtype);
if (err == MPI_SUCCESS && size > 1) {
/* Initiate all send/recv to/from others. */
requests = xbt_new(MPI_Request, 2 * (size - 1));
/* Post all receives first -- a simple optimization */
count = 0;
for (i = (rank + 1) % size; i != rank; i = (i + 1) % size) {
requests[count] = Request::irecv_init(static_cast<char *>(recvbuf) + i * recvcount * recvext, recvcount,
recvtype, i, system_tag, comm);
count++;
}
/* Now post all sends in reverse order
* - We would like to minimize the search time through message queue
* when messages actually arrive in the order in which they were posted.
* TODO: check the previous assertion
*/
for (i = (rank + size - 1) % size; i != rank; i = (i + size - 1) % size) {
requests[count] = Request::isend_init(static_cast<char *>(sendbuf) + i * sendcount * sendext, sendcount,
sendtype, i, system_tag, comm);
count++;
}
/* Wait for them all. */
Request::startall(count, requests);
XBT_DEBUG("<%d> wait for %d requests", rank, count);
Request::waitall(count, requests, MPI_STATUS_IGNORE);
for(i = 0; i < count; i++) {
if(requests[i]!=MPI_REQUEST_NULL)
Request::unref(&requests[i]);
}
xbt_free(requests);
}
return err;
}
int PMPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
{
int retval = 0;
smpi_bench_end();
if (comm == MPI_COMM_NULL) {
retval = MPI_ERR_COMM;
} else if (src == MPI_PROC_NULL) {
if(status != MPI_STATUS_IGNORE){
simgrid::smpi::Status::empty(status);
status->MPI_SOURCE = MPI_PROC_NULL;
}
retval = MPI_SUCCESS;
} else if (src!=MPI_ANY_SOURCE && (src >= comm->group()->size() || src <0)){
retval = MPI_ERR_RANK;
} else if ((count < 0) || (buf==nullptr && count > 0)) {
retval = MPI_ERR_COUNT;
} else if (datatype==MPI_DATATYPE_NULL || not datatype->is_valid()) {
retval = MPI_ERR_TYPE;
} else if(tag<0 && tag != MPI_ANY_TAG){
retval = MPI_ERR_TAG;
} else {
int my_proc_id = simgrid::s4u::this_actor::get_pid();
TRACE_smpi_comm_in(my_proc_id, __func__,
new simgrid::instr::Pt2PtTIData("recv", src,
datatype->is_replayable() ? count : count * datatype->size(),
tag, simgrid::smpi::Datatype::encode(datatype)));
simgrid::smpi::Request::recv(buf, count, datatype, src, tag, comm, status);
retval = MPI_SUCCESS;
// the src may not have been known at the beginning of the recv (MPI_ANY_SOURCE)
int src_traced=0;
if (status != MPI_STATUS_IGNORE)
src_traced = getPid(comm, status->MPI_SOURCE);
else
src_traced = getPid(comm, src);
if (not TRACE_smpi_view_internals()) {
TRACE_smpi_recv(src_traced, my_proc_id, tag);
}
TRACE_smpi_comm_out(my_proc_id);
}
smpi_bench_begin();
return retval;
}
int PMPI_Rget_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, MPI_Win win, MPI_Request* request){
int retval = 0;
smpi_bench_end();
if (win == MPI_WIN_NULL) {
retval = MPI_ERR_WIN;
} else if (target_rank == MPI_PROC_NULL) {
*request = MPI_REQUEST_NULL;
retval = MPI_SUCCESS;
} else if (target_rank <0){
retval = MPI_ERR_RANK;
} else if (win->dynamic()==0 && target_disp <0){
//in case of dynamic window, target_disp can be mistakenly seen as negative, as it is an address
retval = MPI_ERR_ARG;
} else if ((origin_count < 0 || target_count < 0 || result_count <0) ||
(origin_addr==nullptr && origin_count > 0 && op != MPI_NO_OP) ||
(result_addr==nullptr && result_count > 0)){
retval = MPI_ERR_COUNT;
} else if (((target_datatype == MPI_DATATYPE_NULL) || (result_datatype == MPI_DATATYPE_NULL)) ||
(((origin_datatype != MPI_DATATYPE_NULL) && (not origin_datatype->is_valid())) || (not target_datatype->is_valid()) || (not result_datatype->is_valid()))) {
retval = MPI_ERR_TYPE;
} else if (op == MPI_OP_NULL) {
retval = MPI_ERR_OP;
} else if(request == nullptr){
retval = MPI_ERR_REQUEST;
} else {
int my_proc_id = simgrid::s4u::this_actor::get_pid();
MPI_Group group;
win->get_group(&group);
TRACE_smpi_comm_in(my_proc_id, __func__,
new simgrid::instr::Pt2PtTIData(
"Rget_accumulate", target_rank,
target_datatype->is_replayable() ? target_count : target_count * target_datatype->size(),
simgrid::smpi::Datatype::encode(target_datatype)));
retval = win->get_accumulate( origin_addr, origin_count, origin_datatype, result_addr,
result_count, result_datatype, target_rank, target_disp,
target_count, target_datatype, op, request);
TRACE_smpi_comm_out(my_proc_id);
}
smpi_bench_begin();
return retval;
}
int
Coll_reduce_flat_tree::reduce(const void *sbuf, void *rbuf, int count,
MPI_Datatype dtype, MPI_Op op,
int root, MPI_Comm comm)
{
int i, tag = COLL_TAG_REDUCE;
int size;
int rank;
MPI_Aint extent;
unsigned char* origin = nullptr;
const unsigned char* inbuf;
MPI_Status status;
rank = comm->rank();
size = comm->size();
/* If not root, send data to the root. */
extent = dtype->get_extent();
if (rank != root) {
Request::send(sbuf, count, dtype, root, tag, comm);
return 0;
}
/* Root receives and reduces messages. Allocate buffer to receive
messages. */
if (size > 1)
origin = smpi_get_tmp_recvbuffer(count * extent);
/* Initialize the receive buffer. */
if (rank == (size - 1))
Request::sendrecv(sbuf, count, dtype, rank, tag,
rbuf, count, dtype, rank, tag, comm, &status);
else
Request::recv(rbuf, count, dtype, size - 1, tag, comm, &status);
/* Loop receiving and calling reduction function (C or Fortran). */
for (i = size - 2; i >= 0; --i) {
if (rank == i)
inbuf = static_cast<const unsigned char*>(sbuf);
else {
Request::recv(origin, count, dtype, i, tag, comm, &status);
inbuf = origin;
}
/* Call reduction function. */
if(op!=MPI_OP_NULL) op->apply( inbuf, rbuf, &count, dtype);
}
smpi_free_tmp_buffer(origin);
/* All done */
return 0;
}
int PMPI_Type_get_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent)
{
if (datatype == MPI_DATATYPE_NULL) {
return MPI_ERR_TYPE;
} else if (lb == nullptr || extent == nullptr) {
return MPI_ERR_ARG;
} else {
return datatype->extent(lb, extent);
}
}
int PMPI_Type_set_name(MPI_Datatype datatype, const char * name)
{
if (datatype == MPI_DATATYPE_NULL) {
return MPI_ERR_TYPE;
} else if (name == nullptr) {
return MPI_ERR_ARG;
} else {
datatype->set_name(name);
return MPI_SUCCESS;
}
}
int PMPI_Sendrecv_replace(void* buf, int count, MPI_Datatype datatype, int dst, int sendtag, int src, int recvtag,
MPI_Comm comm, MPI_Status* status)
{
int retval = 0;
if (datatype==MPI_DATATYPE_NULL || not datatype->is_valid()) {
return MPI_ERR_TYPE;
} else if (count < 0) {
return MPI_ERR_COUNT;
} else {
int size = datatype->get_extent() * count;
void* recvbuf = xbt_new0(char, size);
retval = MPI_Sendrecv(buf, count, datatype, dst, sendtag, recvbuf, count, datatype, src, recvtag, comm, status);
if(retval==MPI_SUCCESS){
simgrid::smpi::Datatype::copy(recvbuf, count, datatype, buf, count, datatype);
}
xbt_free(recvbuf);
}
return retval;
}
int PMPI_Type_get_name(MPI_Datatype datatype, char * name, int* len)
{
if (datatype == MPI_DATATYPE_NULL) {
return MPI_ERR_TYPE;
} else if (name == nullptr) {
return MPI_ERR_ARG;
} else {
datatype->get_name(name, len);
return MPI_SUCCESS;
}
}
int PMPI_Type_size_x(MPI_Datatype datatype, MPI_Count *size)
{
if (datatype == MPI_DATATYPE_NULL) {
return MPI_ERR_TYPE;
} else if (size == nullptr) {
return MPI_ERR_ARG;
} else {
*size = static_cast<MPI_Count>(datatype->size());
return MPI_SUCCESS;
}
}
int PMPI_Type_ub(MPI_Datatype datatype, MPI_Aint * disp)
{
if (datatype == MPI_DATATYPE_NULL) {
return MPI_ERR_TYPE;
} else if (disp == nullptr) {
return MPI_ERR_ARG;
} else {
*disp = datatype->ub();
return MPI_SUCCESS;
}
}
int PMPI_Type_extent(MPI_Datatype datatype, MPI_Aint * extent)
{
if (datatype == MPI_DATATYPE_NULL) {
return MPI_ERR_TYPE;
} else if (extent == nullptr) {
return MPI_ERR_ARG;
} else {
*extent = datatype->get_extent();
return MPI_SUCCESS;
}
}
int PMPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Request * request)
{
int retval = 0;
smpi_bench_end();
if (request == nullptr) {
retval = MPI_ERR_ARG;
} else if (comm == MPI_COMM_NULL) {
retval = MPI_ERR_COMM;
} else if (src == MPI_PROC_NULL) {
*request = MPI_REQUEST_NULL;
retval = MPI_SUCCESS;
} else if (src!=MPI_ANY_SOURCE && (src >= comm->group()->size() || src <0)){
retval = MPI_ERR_RANK;
} else if ((count < 0) || (buf==nullptr && count > 0)) {
retval = MPI_ERR_COUNT;
} else if (datatype==MPI_DATATYPE_NULL || not datatype->is_valid()) {
retval = MPI_ERR_TYPE;
} else if(tag<0 && tag != MPI_ANY_TAG){
retval = MPI_ERR_TAG;
} else {
int my_proc_id = simgrid::s4u::this_actor::get_pid();
TRACE_smpi_comm_in(my_proc_id, __func__,
new simgrid::instr::Pt2PtTIData("irecv", src,
datatype->is_replayable() ? count : count * datatype->size(),
tag, simgrid::smpi::Datatype::encode(datatype)));
*request = simgrid::smpi::Request::irecv(buf, count, datatype, src, tag, comm);
retval = MPI_SUCCESS;
TRACE_smpi_comm_out(my_proc_id);
}
smpi_bench_begin();
if (retval != MPI_SUCCESS && request != nullptr)
*request = MPI_REQUEST_NULL;
return retval;
}
int Coll_reduce_scatter_ompi::reduce_scatter(const void *sbuf, void *rbuf,
const int *rcounts,
MPI_Datatype dtype,
MPI_Op op,
MPI_Comm comm
)
{
int comm_size, i, pow2;
size_t total_message_size, dsize;
const double a = 0.0012;
const double b = 8.0;
const size_t small_message_size = 12 * 1024;
const size_t large_message_size = 256 * 1024;
int zerocounts = 0;
XBT_DEBUG("Coll_reduce_scatter_ompi::reduce_scatter");
comm_size = comm->size();
// We need data size for decision function
dsize=dtype->size();
total_message_size = 0;
for (i = 0; i < comm_size; i++) {
total_message_size += rcounts[i];
if (0 == rcounts[i]) {
zerocounts = 1;
}
}
if (((op != MPI_OP_NULL) && not op->is_commutative()) || (zerocounts)) {
Coll_reduce_scatter_default::reduce_scatter(sbuf, rbuf, rcounts, dtype, op, comm);
return MPI_SUCCESS;
}
total_message_size *= dsize;
// compute the nearest power of 2
for (pow2 = 1; pow2 < comm_size; pow2 <<= 1);
if ((total_message_size <= small_message_size) ||
((total_message_size <= large_message_size) && (pow2 == comm_size)) ||
(comm_size >= a * total_message_size + b)) {
return
Coll_reduce_scatter_ompi_basic_recursivehalving::reduce_scatter(sbuf, rbuf, rcounts,
dtype, op,
comm);
}
return Coll_reduce_scatter_ompi_ring::reduce_scatter(sbuf, rbuf, rcounts,
dtype, op,
comm);
}
int
Coll_bcast_flattree_pipeline::bcast(void *buff, int count,
MPI_Datatype data_type, int root,
MPI_Comm comm)
{
int i, j, rank, num_procs;
int tag = COLL_TAG_BCAST;
MPI_Aint extent;
extent = data_type->get_extent();
int segment = flattree_segment_in_byte / extent;
segment = segment == 0 ? 1 :segment;
int pipe_length = count / segment;
int increment = segment * extent;
if (pipe_length==0) {
XBT_WARN("MPI_bcast_flattree_pipeline use default MPI_bcast_flattree.");
return Coll_bcast_flattree::bcast(buff, count, data_type, root, comm);
}
rank = comm->rank();
num_procs = comm->size();
MPI_Request *request_array;
MPI_Status *status_array;
request_array = (MPI_Request *) xbt_malloc(pipe_length * sizeof(MPI_Request));
status_array = (MPI_Status *) xbt_malloc(pipe_length * sizeof(MPI_Status));
if (rank != root) {
for (i = 0; i < pipe_length; i++) {
request_array[i] = Request::irecv((char *)buff + (i * increment), segment, data_type, root, tag, comm);
}
Request::waitall(pipe_length, request_array, status_array);
}
else {
// Root sends data to all others
for (j = 0; j < num_procs; j++) {
if (j == rank)
continue;
else {
for (i = 0; i < pipe_length; i++) {
Request::send((char *)buff + (i * increment), segment, data_type, j, tag, comm);
}
}
}
}
free(request_array);
free(status_array);
return MPI_SUCCESS;
}
int Coll_scatter_ompi::scatter(const void *sbuf, int scount,
MPI_Datatype sdtype,
void* rbuf, int rcount,
MPI_Datatype rdtype,
int root, MPI_Comm comm
)
{
const size_t small_block_size = 300;
const int small_comm_size = 10;
int communicator_size, rank;
size_t dsize, block_size;
XBT_DEBUG("Coll_scatter_ompi::scatter");
communicator_size = comm->size();
rank = comm->rank();
// Determine block size
if (root == rank) {
dsize=sdtype->size();
block_size = dsize * scount;
} else {
dsize=rdtype->size();
block_size = dsize * rcount;
}
if ((communicator_size > small_comm_size) &&
(block_size < small_block_size)) {
std::unique_ptr<unsigned char[]> tmp_buf;
if (rank != root) {
tmp_buf.reset(new unsigned char[rcount * rdtype->get_extent()]);
sbuf = tmp_buf.get();
scount = rcount;
sdtype = rdtype;
}
return Coll_scatter_ompi_binomial::scatter(sbuf, scount, sdtype, rbuf, rcount, rdtype, root, comm);
}
return Coll_scatter_ompi_basic_linear::scatter (sbuf, scount, sdtype,
rbuf, rcount, rdtype,
root, comm);
}
/*
* gather_intra
*
* Function: - basic gather operation
* Accepts: - same arguments as MPI_Gather()
* Returns: - MPI_SUCCESS or error code
*/
int Coll_gather_ompi_basic_linear::gather(void* sbuf, int scount, MPI_Datatype sdtype, void* rbuf, int rcount,
MPI_Datatype rdtype, int root, MPI_Comm comm)
{
int i;
int err;
int rank;
int size;
char *ptmp;
MPI_Aint incr;
MPI_Aint extent;
MPI_Aint lb;
size = comm->size();
rank = comm->rank();
/* Everyone but root sends data and returns. */
XBT_DEBUG("ompi_coll_tuned_gather_intra_basic_linear rank %d", rank);
if (rank != root) {
Request::send(sbuf, scount, sdtype, root,
COLL_TAG_GATHER,
comm);
return MPI_SUCCESS;
}
/* I am the root, loop receiving the data. */
rdtype->extent(&lb, &extent);
incr = extent * rcount;
for (i = 0, ptmp = (char *) rbuf; i < size; ++i, ptmp += incr) {
if (i == rank) {
if (MPI_IN_PLACE != sbuf) {
err = Datatype::copy(sbuf, scount, sdtype,
ptmp, rcount, rdtype);
} else {
err = MPI_SUCCESS;
}
} else {
Request::recv(ptmp, rcount, rdtype, i,
COLL_TAG_GATHER,
comm, MPI_STATUS_IGNORE);
err = MPI_SUCCESS;
}
if (MPI_SUCCESS != err) {
return err;
}
}
/* All done */
return MPI_SUCCESS;
}
int PMPI_Put(const 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_Win win){
int retval = 0;
smpi_bench_end();
if (win == MPI_WIN_NULL) {
retval = MPI_ERR_WIN;
} else if (target_rank == MPI_PROC_NULL) {
retval = MPI_SUCCESS;
} else if (target_rank <0){
retval = MPI_ERR_RANK;
} else if (win->dynamic()==0 && target_disp <0){
//in case of dynamic window, target_disp can be mistakenly seen as negative, as it is an address
retval = MPI_ERR_ARG;
} else if ((origin_count < 0 || target_count < 0) ||
(origin_addr==nullptr && origin_count > 0)){
retval = MPI_ERR_COUNT;
} else if (((origin_datatype == MPI_DATATYPE_NULL) || (target_datatype == MPI_DATATYPE_NULL)) ||
((not origin_datatype->is_valid()) || (not target_datatype->is_valid()))) {
retval = MPI_ERR_TYPE;
} else {
int my_proc_id = simgrid::s4u::this_actor::get_pid();
MPI_Group group;
win->get_group(&group);
int dst_traced = group->actor(target_rank)->get_pid();
TRACE_smpi_comm_in(my_proc_id, __func__,
new simgrid::instr::Pt2PtTIData("Put", target_rank, origin_datatype->is_replayable()
? origin_count
: origin_count * origin_datatype->size(),
simgrid::smpi::Datatype::encode(origin_datatype)));
TRACE_smpi_send(my_proc_id, my_proc_id, dst_traced, SMPI_RMA_TAG, origin_count * origin_datatype->size());
retval = win->put( origin_addr, origin_count, origin_datatype, target_rank, target_disp, target_count,
target_datatype);
TRACE_smpi_comm_out(my_proc_id);
}
smpi_bench_begin();
return retval;
}
int Coll_allgatherv_ompi::allgatherv(const void *sbuf, int scount,
MPI_Datatype sdtype,
void* rbuf, const int *rcounts,
const int *rdispls,
MPI_Datatype rdtype,
MPI_Comm comm
)
{
int i;
int communicator_size;
size_t dsize, total_dsize;
communicator_size = comm->size();
/* Special case for 2 processes */
if (communicator_size == 2) {
return Coll_allgatherv_pair::allgatherv(sbuf, scount, sdtype,
rbuf, rcounts, rdispls, rdtype,
comm);
}
/* Determine complete data size */
dsize=sdtype->size();
total_dsize = 0;
for (i = 0; i < communicator_size; i++) {
total_dsize += dsize * rcounts[i];
}
/* Decision based on allgather decision. */
if (total_dsize < 50000) {
return Coll_allgatherv_ompi_bruck::allgatherv(sbuf, scount, sdtype,
rbuf, rcounts, rdispls, rdtype,
comm);
} else {
if (communicator_size % 2) {
return Coll_allgatherv_ring::allgatherv(sbuf, scount, sdtype,
rbuf, rcounts, rdispls, rdtype,
comm);
} else {
return Coll_allgatherv_ompi_neighborexchange::allgatherv(sbuf, scount, sdtype,
rbuf, rcounts, rdispls, rdtype,
comm);
}
}
}
int PMPI_Type_create_subarray(int ndims, const int* array_of_sizes,
const int* array_of_subsizes, const int* array_of_starts,
int order, MPI_Datatype oldtype, MPI_Datatype *newtype) {
if (ndims<0){
return MPI_ERR_COUNT;
} else if (ndims==0){
*newtype = MPI_DATATYPE_NULL;
return MPI_SUCCESS;
} else if (ndims==1){
simgrid::smpi::Datatype::create_contiguous( array_of_subsizes[0], oldtype, array_of_starts[0]*oldtype->get_extent(), newtype);
return MPI_SUCCESS;
} else if (oldtype == MPI_DATATYPE_NULL || not oldtype->is_valid() ) {
return MPI_ERR_TYPE;
} else if (order != MPI_ORDER_FORTRAN && order != MPI_ORDER_C){
return MPI_ERR_ARG;
} else {
return simgrid::smpi::Datatype::create_subarray(ndims, array_of_sizes, array_of_subsizes, array_of_starts, order, oldtype, newtype);
}
}
int PMPI_Ssend_init(const void* buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm, MPI_Request* request)
{
int retval = 0;
smpi_bench_end();
if (request == nullptr) {
retval = MPI_ERR_ARG;
} else if (comm == MPI_COMM_NULL) {
retval = MPI_ERR_COMM;
} else if (datatype==MPI_DATATYPE_NULL || not datatype->is_valid()) {
retval = MPI_ERR_TYPE;
} else if (dst == MPI_PROC_NULL) {
retval = MPI_SUCCESS;
} else {
*request = simgrid::smpi::Request::ssend_init(buf, count, datatype, dst, tag, comm);
retval = MPI_SUCCESS;
}
smpi_bench_begin();
if (retval != MPI_SUCCESS && request != nullptr)
*request = MPI_REQUEST_NULL;
return retval;
}
int
Coll_allgatherv_pair::allgatherv(void *send_buff, int send_count,
MPI_Datatype send_type, void *recv_buff,
int *recv_counts, int *recv_disps, MPI_Datatype recv_type,
MPI_Comm comm)
{
MPI_Aint extent;
unsigned int i, src, dst;
int tag = COLL_TAG_ALLGATHERV;
MPI_Status status;
char *send_ptr = (char *) send_buff;
char *recv_ptr = (char *) recv_buff;
unsigned int rank = comm->rank();
unsigned int num_procs = comm->size();
if((num_procs&(num_procs-1)))
THROWF(arg_error,0, "allgatherv pair algorithm can't be used with non power of two number of processes ! ");
extent = send_type->get_extent();
// local send/recv
Request::sendrecv(send_ptr, send_count, send_type, rank, tag,
recv_ptr + recv_disps[rank] * extent,
recv_counts[rank], recv_type, rank, tag, comm, &status);
for (i = 1; i < num_procs; i++) {
src = dst = rank ^ i;
Request::sendrecv(send_ptr, send_count, send_type, dst, tag,
recv_ptr + recv_disps[src] * extent, recv_counts[src], recv_type,
src, tag, comm, &status);
}
return MPI_SUCCESS;
}
int Coll_allgather_mvapich2_smp::allgather(void *sendbuf,int sendcnt, MPI_Datatype sendtype,
void *recvbuf, int recvcnt,MPI_Datatype recvtype,
MPI_Comm comm)
{
int rank, size;
int local_rank, local_size;
int leader_comm_size = 0;
int mpi_errno = MPI_SUCCESS;
MPI_Aint recvtype_extent = 0; /* Datatype extent */
MPI_Comm shmem_comm, leader_comm;
if(comm->get_leaders_comm()==MPI_COMM_NULL){
comm->init_smp();
}
if (not comm->is_uniform() || not comm->is_blocked())
THROWF(arg_error,0, "allgather MVAPICH2 smp algorithm can't be used with irregular deployment. Please insure that processes deployed on the same node are contiguous and that each node has the same number of processes");
if (recvcnt == 0) {
return MPI_SUCCESS;
}
rank = comm->rank();
size = comm->size();
/* extract the rank,size information for the intra-node communicator */
recvtype_extent=recvtype->get_extent();
shmem_comm = comm->get_intra_comm();
local_rank = shmem_comm->rank();
local_size = shmem_comm->size();
if (local_rank == 0) {
/* Node leader. Extract the rank, size information for the leader communicator */
leader_comm = comm->get_leaders_comm();
if(leader_comm==MPI_COMM_NULL){
leader_comm = MPI_COMM_WORLD;
}
leader_comm_size = leader_comm->size();
}
/*If there is just one node, after gather itself,
* root has all the data and it can do bcast*/
if(local_rank == 0) {
mpi_errno = Colls::gather(sendbuf, sendcnt,sendtype,
(void*)((char*)recvbuf + (rank * recvcnt * recvtype_extent)),
recvcnt, recvtype,
0, shmem_comm);
} else {
/*Since in allgather all the processes could have
* its own data in place*/
if(sendbuf == MPI_IN_PLACE) {
mpi_errno = Colls::gather((void*)((char*)recvbuf + (rank * recvcnt * recvtype_extent)),
recvcnt , recvtype,
recvbuf, recvcnt, recvtype,
0, shmem_comm);
} else {
mpi_errno = Colls::gather(sendbuf, sendcnt,sendtype,
recvbuf, recvcnt, recvtype,
0, shmem_comm);
}
}
/* Exchange the data between the node leaders*/
if (local_rank == 0 && (leader_comm_size > 1)) {
/*When data in each socket is different*/
if (comm->is_uniform() != 1) {
int *displs = NULL;
int *recvcnts = NULL;
int *node_sizes = NULL;
int i = 0;
node_sizes = comm->get_non_uniform_map();
displs = static_cast<int *>(xbt_malloc(sizeof (int) * leader_comm_size));
recvcnts = static_cast<int *>(xbt_malloc(sizeof (int) * leader_comm_size));
if (not displs || not recvcnts) {
return MPI_ERR_OTHER;
}
recvcnts[0] = node_sizes[0] * recvcnt;
displs[0] = 0;
for (i = 1; i < leader_comm_size; i++) {
displs[i] = displs[i - 1] + node_sizes[i - 1] * recvcnt;
recvcnts[i] = node_sizes[i] * recvcnt;
}
void* sendbuf=((char*)recvbuf)+recvtype->get_extent()*displs[leader_comm->rank()];
mpi_errno = Colls::allgatherv(sendbuf,
(recvcnt*local_size),
recvtype,
recvbuf, recvcnts,
displs, recvtype,
leader_comm);
xbt_free(displs);
xbt_free(recvcnts);
} else {
void* sendtmpbuf=((char*)recvbuf)+recvtype->get_extent()*(recvcnt*local_size)*leader_comm->rank();
mpi_errno = Coll_allgather_mpich::allgather(sendtmpbuf,
(recvcnt*local_size),
recvtype,
recvbuf, (recvcnt*local_size), recvtype,
leader_comm);
}
}
/*Bcast the entire data from node leaders to all other cores*/
mpi_errno = Colls::bcast (recvbuf, recvcnt * size, recvtype, 0, shmem_comm);
return mpi_errno;
}
