X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/7a40637a7db0cf339c5a09ba5594a004a5304a3c..4c753f8d4cabd4104f3f7109823f16be2ebdcce3:/src/smpi/bindings/smpi_pmpi_coll.cpp?ds=sidebyside diff --git a/src/smpi/bindings/smpi_pmpi_coll.cpp b/src/smpi/bindings/smpi_pmpi_coll.cpp index 4d2e9e8e69..f83a8d3c4b 100644 --- a/src/smpi/bindings/smpi_pmpi_coll.cpp +++ b/src/smpi/bindings/smpi_pmpi_coll.cpp @@ -1,4 +1,4 @@ -/* Copyright (c) 2007-2019. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2007-2022. The SimGrid Team. All rights reserved. */ /* This program is free software; you can redistribute it and/or modify it * under the terms of the license (GNU LGPL) which comes with this package. */ @@ -11,14 +11,22 @@ #include "smpi_op.hpp" #include "src/smpi/include/smpi_actor.hpp" -XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(smpi_pmpi); +#include -/* PMPI User level calls */ +XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(smpi_pmpi); -int PMPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm) +static const void* smpi_get_in_place_buf(const void* inplacebuf, const void* otherbuf, + std::vector& tmp_sendbuf, int count, MPI_Datatype datatype) { - return PMPI_Ibcast(buf, count, datatype, root, comm, MPI_REQUEST_IGNORED); + if (inplacebuf == MPI_IN_PLACE) { + tmp_sendbuf.resize(count * datatype->get_extent()); + simgrid::smpi::Datatype::copy(otherbuf, count, datatype, tmp_sendbuf.data(), count, datatype); + return tmp_sendbuf.data(); + }else{ + return inplacebuf; + } } +/* PMPI User level calls */ int PMPI_Barrier(MPI_Comm comm) { @@ -27,747 +35,984 @@ int PMPI_Barrier(MPI_Comm comm) int PMPI_Ibarrier(MPI_Comm comm, MPI_Request *request) { - int retval = 0; - smpi_bench_end(); - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if(request == nullptr){ - retval = MPI_ERR_ARG; - }else{ - int rank = simgrid::s4u::this_actor::get_pid(); - TRACE_smpi_comm_in(rank, __func__, new simgrid::instr::NoOpTIData(request==MPI_REQUEST_IGNORED? "barrier" : "ibarrier")); - if(request==MPI_REQUEST_IGNORED){ - simgrid::smpi::Colls::barrier(comm); - //Barrier can be used to synchronize RMA calls. Finish all requests from comm before. - comm->finish_rma_calls(); - } else - simgrid::smpi::Colls::ibarrier(comm, request); - TRACE_smpi_comm_out(rank); - } - smpi_bench_begin(); - return retval; + CHECK_COMM(1) + CHECK_REQUEST(2) + CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Barrier" : "PMPI_Ibarrier") + const SmpiBenchGuard suspend_bench; + aid_t pid = simgrid::s4u::this_actor::get_pid(); + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Barrier" : "PMPI_Ibarrier", + new simgrid::instr::NoOpTIData(request == MPI_REQUEST_IGNORED ? "barrier" : "ibarrier")); + if (request == MPI_REQUEST_IGNORED) { + simgrid::smpi::colls::barrier(comm); + // Barrier can be used to synchronize RMA calls. Finish all requests from comm before. + comm->finish_rma_calls(); + } else + simgrid::smpi::colls::ibarrier(comm, request); + + TRACE_smpi_comm_out(pid); + return MPI_SUCCESS; } -int PMPI_Ibcast(void *buf, int count, MPI_Datatype datatype, - int root, MPI_Comm comm, MPI_Request* request) +int PMPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm) { - int retval = 0; - smpi_bench_end(); - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if (not datatype->is_valid()) { - retval = MPI_ERR_ARG; - } else if(request == nullptr){ - retval = MPI_ERR_ARG; - } else { - int rank = simgrid::s4u::this_actor::get_pid(); - TRACE_smpi_comm_in(rank, __func__, - new simgrid::instr::CollTIData(request==MPI_REQUEST_IGNORED?"bcast":"ibcast", root, -1.0, - datatype->is_replayable() ? count : count * datatype->size(), -1, - simgrid::smpi::Datatype::encode(datatype), "")); - if (comm->size() > 1){ - if(request==MPI_REQUEST_IGNORED) - simgrid::smpi::Colls::bcast(buf, count, datatype, root, comm); - else - simgrid::smpi::Colls::ibcast(buf, count, datatype, root, comm, request); - } - retval = MPI_SUCCESS; + return PMPI_Ibcast(buf, count, datatype, root, comm, MPI_REQUEST_IGNORED); +} - TRACE_smpi_comm_out(rank); +int PMPI_Ibcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm, MPI_Request* request) +{ + SET_BUF1(buf) + CHECK_COMM(5) + CHECK_COUNT(2, count) + CHECK_TYPE(3, datatype) + CHECK_BUFFER(1, buf, count, datatype) + CHECK_ROOT(4) + CHECK_REQUEST(6) + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Bcast" : "PMPI_Ibcast") + " with root " + + std::to_string(root)) + + const SmpiBenchGuard suspend_bench; + aid_t pid = simgrid::s4u::this_actor::get_pid(); + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Bcast" : "PMPI_Ibcast", + new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "bcast" : "ibcast", root, -1.0, + count, 0, + simgrid::smpi::Datatype::encode(datatype), "")); + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // No barrier in Ibcast + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); + + if (comm->size() > 1) { + if (request == MPI_REQUEST_IGNORED) + simgrid::smpi::colls::bcast(buf, count, datatype, root, comm); + else + simgrid::smpi::colls::ibcast(buf, count, datatype, root, comm, request); + } else { + if (request != MPI_REQUEST_IGNORED) + *request = MPI_REQUEST_NULL; } - smpi_bench_begin(); - return retval; + + TRACE_smpi_comm_out(pid); + return MPI_SUCCESS; } -int PMPI_Gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,void *recvbuf, int recvcount, MPI_Datatype recvtype, +int PMPI_Gather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm){ return PMPI_Igather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm, MPI_REQUEST_IGNORED); } -int PMPI_Igather(void *sendbuf, int sendcount, MPI_Datatype sendtype,void *recvbuf, int recvcount, MPI_Datatype recvtype, - int root, MPI_Comm comm, MPI_Request *request) +int PMPI_Igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, + MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request) { - int retval = 0; - - smpi_bench_end(); - - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if ((( sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL)) || - ((comm->rank() == root) && (recvtype == MPI_DATATYPE_NULL))){ - retval = MPI_ERR_TYPE; - } else if ((( sendbuf != MPI_IN_PLACE) && (sendcount <0)) || ((comm->rank() == root) && (recvcount <0))){ - retval = MPI_ERR_COUNT; - } else if (request == nullptr){ - retval = MPI_ERR_ARG; - } else { - - char* sendtmpbuf = static_cast(sendbuf); - int sendtmpcount = sendcount; - MPI_Datatype sendtmptype = sendtype; - if( (comm->rank() == root) && (sendbuf == MPI_IN_PLACE )) { - sendtmpcount=0; - sendtmptype=recvtype; + CHECK_COMM(8) + SET_BUF1(sendbuf) + int rank = comm->rank(); + if(sendbuf != MPI_IN_PLACE){ + CHECK_COUNT(2, sendcount) + CHECK_TYPE(3, sendtype) + CHECK_BUFFER(1,sendbuf, sendcount, sendtype) + } + if(rank == root){ + SET_BUF2(recvbuf) + CHECK_NOT_IN_PLACE_ROOT(4, recvbuf) + CHECK_TYPE(6, recvtype) + CHECK_COUNT(5, recvcount) + CHECK_BUFFER(4, recvbuf, recvcount, recvtype) + } else { + CHECK_NOT_IN_PLACE_ROOT(1, sendbuf) + } + CHECK_ROOT(7) + CHECK_REQUEST(9) + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Gather" : "PMPI_Igather") + " with root " + + std::to_string(root)) + + const void* real_sendbuf = sendbuf; + int real_sendcount = sendcount; + MPI_Datatype real_sendtype = sendtype; + if (rank == root){ + if (sendbuf == MPI_IN_PLACE) { + real_sendcount = 0; + real_sendtype = recvtype; + } else if(recvtype->size() * recvcount != sendtype->size() * sendcount){ + XBT_WARN("MPI_(I)Gather : received size at root differs from sent size : %zu vs %zu", recvtype->size() * recvcount , sendtype->size() * sendcount); + return MPI_ERR_TRUNCATE; } - int rank = simgrid::s4u::this_actor::get_pid(); - - TRACE_smpi_comm_in( - rank, __func__, - new simgrid::instr::CollTIData( - request==MPI_REQUEST_IGNORED ? "gather":"igather", root, -1.0, sendtmptype->is_replayable() ? sendtmpcount : sendtmpcount * sendtmptype->size(), - (comm->rank() != root || recvtype->is_replayable()) ? recvcount : recvcount * recvtype->size(), - simgrid::smpi::Datatype::encode(sendtmptype), simgrid::smpi::Datatype::encode(recvtype))); - if(request == MPI_REQUEST_IGNORED) - simgrid::smpi::Colls::gather(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcount, recvtype, root, comm); - else - simgrid::smpi::Colls::igather(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcount, recvtype, root, comm, request); - - retval = MPI_SUCCESS; - TRACE_smpi_comm_out(rank); } - smpi_bench_begin(); - return retval; + const SmpiBenchGuard suspend_bench; + + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // no barrier in Igather + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); + + aid_t pid = simgrid::s4u::this_actor::get_pid(); + + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Gather" : "PMPI_Igather", + new simgrid::instr::CollTIData( + request == MPI_REQUEST_IGNORED ? "gather" : "igather", root, -1.0, + real_sendcount, recvcount, + simgrid::smpi::Datatype::encode(real_sendtype), simgrid::smpi::Datatype::encode(recvtype))); + if (request == MPI_REQUEST_IGNORED) + simgrid::smpi::colls::gather(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcount, recvtype, root, comm); + else + simgrid::smpi::colls::igather(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcount, recvtype, root, comm, + request); + + TRACE_smpi_comm_out(pid); + return MPI_SUCCESS; } -int PMPI_Gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int *recvcounts, int *displs, +int PMPI_Gatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, const int *recvcounts, const int *displs, MPI_Datatype recvtype, int root, MPI_Comm comm){ return PMPI_Igatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, root, comm, MPI_REQUEST_IGNORED); } -int PMPI_Igatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int *recvcounts, int *displs, - MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request *request) +int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, const int* recvcounts, const int* displs, + MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request) { - int retval = 0; - - smpi_bench_end(); - - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if ((( sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL)) || - ((comm->rank() == root) && (recvtype == MPI_DATATYPE_NULL))){ - retval = MPI_ERR_TYPE; - } else if (( sendbuf != MPI_IN_PLACE) && (sendcount <0)){ - retval = MPI_ERR_COUNT; - } else if ((comm->rank() == root) && (recvcounts == nullptr || displs == nullptr)) { - retval = MPI_ERR_ARG; - } else if (request == nullptr){ - retval = MPI_ERR_ARG; - } else { - char* sendtmpbuf = static_cast(sendbuf); - int sendtmpcount = sendcount; - MPI_Datatype sendtmptype = sendtype; - if( (comm->rank() == root) && (sendbuf == MPI_IN_PLACE )) { - sendtmpcount=0; - sendtmptype=recvtype; + CHECK_COMM(9) + SET_BUF1(sendbuf) + int rank = comm->rank(); + if(sendbuf != MPI_IN_PLACE){ + CHECK_TYPE(3, sendtype) + CHECK_COUNT(2, sendcount) + } + CHECK_BUFFER(1, sendbuf, sendcount, sendtype) + if(rank == root){ + SET_BUF2(recvbuf) + CHECK_NOT_IN_PLACE_ROOT(4, recvbuf) + CHECK_TYPE(6, recvtype) + CHECK_NULL(5, MPI_ERR_COUNT, recvcounts) + CHECK_NULL(6, MPI_ERR_ARG, displs) + } else { + CHECK_NOT_IN_PLACE_ROOT(1, sendbuf) + } + CHECK_ROOT(8) + CHECK_REQUEST(10) + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Gatherv" : "PMPI_Igatherv") + + " with root " + std::to_string(root)) + + if (rank == root){ + for (int i = 0; i < comm->size(); i++) { + CHECK_COUNT(5, recvcounts[i]) + CHECK_BUFFER(4,recvbuf,recvcounts[i], recvtype) } + } - int rank = simgrid::s4u::this_actor::get_pid(); - int dt_size_recv = recvtype->is_replayable() ? 1 : recvtype->size(); + const SmpiBenchGuard suspend_bench; - std::vector* trace_recvcounts = new std::vector; - if (comm->rank() == root) { - for (int i = 0; i < comm->size(); i++) // copy data to avoid bad free - trace_recvcounts->push_back(recvcounts[i] * dt_size_recv); - } + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // no barrier in Igatherv + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); - TRACE_smpi_comm_in(rank, __func__, - new simgrid::instr::VarCollTIData( - request==MPI_REQUEST_IGNORED ? "gatherv":"igatherv", root, - sendtmptype->is_replayable() ? sendtmpcount : sendtmpcount * sendtmptype->size(), nullptr, - dt_size_recv, trace_recvcounts, simgrid::smpi::Datatype::encode(sendtmptype), - simgrid::smpi::Datatype::encode(recvtype))); - if(request == MPI_REQUEST_IGNORED) - retval = simgrid::smpi::Colls::gatherv(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcounts, displs, recvtype, root, comm); - else - retval = simgrid::smpi::Colls::igatherv(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcounts, displs, recvtype, root, comm, request); - TRACE_smpi_comm_out(rank); + const void* real_sendbuf = sendbuf; + int real_sendcount = sendcount; + MPI_Datatype real_sendtype = sendtype; + if ((rank == root) && (sendbuf == MPI_IN_PLACE)) { + real_sendcount = 0; + real_sendtype = recvtype; } - smpi_bench_begin(); - return retval; + aid_t pid = simgrid::s4u::this_actor::get_pid(); + + auto trace_recvcounts = std::make_shared>(); + if (rank == root) + trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[comm->size()]); + else //this is not significant outside of root, put 0 as we don't know if recvcounts is initialized + trace_recvcounts->insert(trace_recvcounts->end(), comm->size(), 0); + + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Gatherv" : "PMPI_Igatherv", + new simgrid::instr::VarCollTIData( + request == MPI_REQUEST_IGNORED ? "gatherv" : "igatherv", root, + sendcount, + nullptr, -1, trace_recvcounts, simgrid::smpi::Datatype::encode(real_sendtype), + simgrid::smpi::Datatype::encode(recvtype))); + if (request == MPI_REQUEST_IGNORED) + simgrid::smpi::colls::gatherv(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcounts, displs, recvtype, + root, comm); + else + simgrid::smpi::colls::igatherv(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcounts, displs, recvtype, + root, comm, request); + + TRACE_smpi_comm_out(pid); + return MPI_SUCCESS; } -int PMPI_Allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype, +int PMPI_Allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm){ return PMPI_Iallgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, MPI_REQUEST_IGNORED); } - -int PMPI_Iallgather(void *sendbuf, int sendcount, MPI_Datatype sendtype, - void *recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request) + +int PMPI_Iallgather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, + MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request) { - int retval = MPI_SUCCESS; - - smpi_bench_end(); - - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if ((( sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL)) || - (recvtype == MPI_DATATYPE_NULL)){ - retval = MPI_ERR_TYPE; - } else if ((( sendbuf != MPI_IN_PLACE) && (sendcount <0)) || - (recvcount <0)){ - retval = MPI_ERR_COUNT; - } else if (request == nullptr){ - retval = MPI_ERR_ARG; - } else { - if(sendbuf == MPI_IN_PLACE) { - sendbuf=static_cast(recvbuf)+recvtype->get_extent()*recvcount*comm->rank(); - sendcount=recvcount; - sendtype=recvtype; - } - int rank = simgrid::s4u::this_actor::get_pid(); - - TRACE_smpi_comm_in(rank, __func__, - new simgrid::instr::CollTIData( - request==MPI_REQUEST_IGNORED ? "allgather" : "iallgather", -1, -1.0, sendtype->is_replayable() ? sendcount : sendcount * sendtype->size(), - recvtype->is_replayable() ? recvcount : recvcount * recvtype->size(), - simgrid::smpi::Datatype::encode(sendtype), simgrid::smpi::Datatype::encode(recvtype))); - if(request == MPI_REQUEST_IGNORED) - simgrid::smpi::Colls::allgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm); - else - simgrid::smpi::Colls::iallgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, request); - TRACE_smpi_comm_out(rank); + CHECK_COMM(7) + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + int rank = comm->rank(); + CHECK_NOT_IN_PLACE(4, recvbuf) + if(sendbuf != MPI_IN_PLACE){ + CHECK_COUNT(2, sendcount) + CHECK_TYPE(3, sendtype) } - smpi_bench_begin(); - return retval; + CHECK_TYPE(6, recvtype) + CHECK_COUNT(5, recvcount) + CHECK_BUFFER(1, sendbuf, sendcount, sendtype) + CHECK_BUFFER(4, recvbuf, recvcount, recvtype) + CHECK_REQUEST(8) + CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Allgather" : "PMPI_Iallggather") + + if (sendbuf == MPI_IN_PLACE) { + sendbuf = static_cast(recvbuf) + recvtype->get_extent() * recvcount * comm->rank(); + sendcount = recvcount; + sendtype = recvtype; + } + + if(recvtype->size() * recvcount != sendtype->size() * sendcount){ + XBT_WARN("MPI_(I)Allgather : received size from each process differs from sent size : %zu vs %zu", recvtype->size() * recvcount, sendtype->size() * sendcount); + return MPI_ERR_TRUNCATE; + } + + const SmpiBenchGuard suspend_bench; + + aid_t pid = simgrid::s4u::this_actor::get_pid(); + + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Allgather" : "PMPI_Iallggather", + new simgrid::instr::CollTIData( + request == MPI_REQUEST_IGNORED ? "allgather" : "iallgather", -1, -1.0, + sendcount, recvcount, + simgrid::smpi::Datatype::encode(sendtype), simgrid::smpi::Datatype::encode(recvtype))); + if (request == MPI_REQUEST_IGNORED) + simgrid::smpi::colls::allgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm); + else + simgrid::smpi::colls::iallgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, request); + + TRACE_smpi_comm_out(pid); + return MPI_SUCCESS; } -int PMPI_Allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, - void *recvbuf, int *recvcounts, int *displs, MPI_Datatype recvtype, MPI_Comm comm){ +int PMPI_Allgatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype, + void *recvbuf, const int *recvcounts, const int *displs, MPI_Datatype recvtype, MPI_Comm comm){ return PMPI_Iallgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, comm, MPI_REQUEST_IGNORED); } -int PMPI_Iallgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, - void *recvbuf, int *recvcounts, int *displs, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request) +int PMPI_Iallgatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, const int* recvcounts, const int* displs, + MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request) { - int retval = 0; - - smpi_bench_end(); - - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if (((sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL)) || (recvtype == MPI_DATATYPE_NULL)) { - retval = MPI_ERR_TYPE; - } else if (( sendbuf != MPI_IN_PLACE) && (sendcount <0)){ - retval = MPI_ERR_COUNT; - } else if (recvcounts == nullptr || displs == nullptr) { - retval = MPI_ERR_ARG; - } else if (request == nullptr){ - retval = MPI_ERR_ARG; - } else { - - if(sendbuf == MPI_IN_PLACE) { - sendbuf=static_cast(recvbuf)+recvtype->get_extent()*displs[comm->rank()]; - sendcount=recvcounts[comm->rank()]; - sendtype=recvtype; - } - int rank = simgrid::s4u::this_actor::get_pid(); - int dt_size_recv = recvtype->is_replayable() ? 1 : recvtype->size(); - - std::vector* trace_recvcounts = new std::vector; - for (int i = 0; i < comm->size(); i++) // copy data to avoid bad free - trace_recvcounts->push_back(recvcounts[i] * dt_size_recv); - - TRACE_smpi_comm_in(rank, __func__, - new simgrid::instr::VarCollTIData( - request==MPI_REQUEST_IGNORED ? "allgatherv" : "iallgatherv", -1, sendtype->is_replayable() ? sendcount : sendcount * sendtype->size(), - nullptr, dt_size_recv, trace_recvcounts, simgrid::smpi::Datatype::encode(sendtype), - simgrid::smpi::Datatype::encode(recvtype))); - if(request == MPI_REQUEST_IGNORED) - simgrid::smpi::Colls::allgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, comm); - else - simgrid::smpi::Colls::iallgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, comm, request); - retval = MPI_SUCCESS; - TRACE_smpi_comm_out(rank); + CHECK_COMM(8) + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + int rank = comm->rank(); + if(sendbuf != MPI_IN_PLACE) + CHECK_TYPE(3, sendtype) + CHECK_TYPE(6, recvtype) + CHECK_NULL(5, MPI_ERR_COUNT, recvcounts) + CHECK_NULL(6, MPI_ERR_ARG, displs) + if(sendbuf != MPI_IN_PLACE){ + CHECK_COUNT(2, sendcount) + CHECK_BUFFER(1, sendbuf, sendcount, sendtype) } + CHECK_REQUEST(9) + CHECK_NOT_IN_PLACE(4, recvbuf) + for (int i = 0; i < comm->size(); i++) { + CHECK_COUNT(5, recvcounts[i]) + CHECK_BUFFER(4, recvbuf, recvcounts[i], recvtype) + } + CHECK_COLLECTIVE(comm, MPI_REQUEST_IGNORED ? "PMPI_Allgatherv" : "PMPI_Iallgatherv") - smpi_bench_begin(); - return retval; + const SmpiBenchGuard suspend_bench; + + if (sendbuf == MPI_IN_PLACE) { + sendbuf = static_cast(recvbuf) + recvtype->get_extent() * displs[comm->rank()]; + sendcount = recvcounts[comm->rank()]; + sendtype = recvtype; + } + aid_t pid = simgrid::s4u::this_actor::get_pid(); + + auto trace_recvcounts = std::make_shared>(); + trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[comm->size()]); + + TRACE_smpi_comm_in( + pid, request == MPI_REQUEST_IGNORED ? "PMPI_Allgatherv" : "PMPI_Iallgatherv", + new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "allgatherv" : "iallgatherv", -1, + sendcount, nullptr, + -1, trace_recvcounts, simgrid::smpi::Datatype::encode(sendtype), + simgrid::smpi::Datatype::encode(recvtype))); + if (request == MPI_REQUEST_IGNORED) + simgrid::smpi::colls::allgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, comm); + else + simgrid::smpi::colls::iallgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, comm, + request); + + TRACE_smpi_comm_out(pid); + return MPI_SUCCESS; } -int PMPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype, +int PMPI_Scatter(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm){ return PMPI_Iscatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm, MPI_REQUEST_IGNORED); } -int PMPI_Iscatter(void *sendbuf, int sendcount, MPI_Datatype sendtype, - void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request) +int PMPI_Iscatter(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, + MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request) { - int retval = 0; + CHECK_COMM(8) + SET_BUF2(recvbuf) + int rank = comm->rank(); + if(rank == root){ + SET_BUF1(sendbuf) + CHECK_NOT_IN_PLACE_ROOT(1, sendbuf) + CHECK_COUNT(2, sendcount) + CHECK_TYPE(3, sendtype) + CHECK_BUFFER(1, sendbuf, sendcount, sendtype) + } else { + CHECK_NOT_IN_PLACE_ROOT(4, recvbuf) + } + if(recvbuf != MPI_IN_PLACE){ + CHECK_COUNT(5, recvcount) + CHECK_TYPE(6, recvtype) + CHECK_BUFFER(4, recvbuf, recvcount, recvtype) + } + CHECK_ROOT(8) + CHECK_REQUEST(9) + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Scatter" : "PMPI_Iscatter") + + " with root " + std::to_string(root)) + + if (recvbuf == MPI_IN_PLACE) { + recvtype = sendtype; + recvcount = sendcount; + } - smpi_bench_end(); + if((rank == root) && (recvtype->size() * recvcount != sendtype->size() * sendcount)){ + XBT_WARN("MPI_(I)Scatter : sent size to each process differs from receive size"); + return MPI_ERR_TRUNCATE; + } - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if (((comm->rank() == root) && (not sendtype->is_valid())) || - ((recvbuf != MPI_IN_PLACE) && (not recvtype->is_valid()))) { - retval = MPI_ERR_TYPE; - } else if ((sendbuf == recvbuf) || - ((comm->rank()==root) && sendcount>0 && (sendbuf == nullptr))){ - retval = MPI_ERR_BUFFER; - }else { + const SmpiBenchGuard suspend_bench; - if (recvbuf == MPI_IN_PLACE) { - recvtype = sendtype; - recvcount = sendcount; - } - int rank = simgrid::s4u::this_actor::get_pid(); - - TRACE_smpi_comm_in( - rank, __func__, - new simgrid::instr::CollTIData( - request==MPI_REQUEST_IGNORED ? "scatter" : "iscatter", root, -1.0, - (comm->rank() != root || sendtype->is_replayable()) ? sendcount : sendcount * sendtype->size(), - recvtype->is_replayable() ? recvcount : recvcount * recvtype->size(), - simgrid::smpi::Datatype::encode(sendtype), simgrid::smpi::Datatype::encode(recvtype))); - if(request == MPI_REQUEST_IGNORED) - simgrid::smpi::Colls::scatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm); - else - simgrid::smpi::Colls::iscatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm, request); - retval = MPI_SUCCESS; - TRACE_smpi_comm_out(rank); - } + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // no barrier in Iscatter + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); - smpi_bench_begin(); - return retval; + aid_t pid = simgrid::s4u::this_actor::get_pid(); + + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Scatter" : "PMPI_Iscatter", + new simgrid::instr::CollTIData( + request == MPI_REQUEST_IGNORED ? "scatter" : "iscatter", root, -1.0, + sendcount, recvcount, + simgrid::smpi::Datatype::encode(sendtype), simgrid::smpi::Datatype::encode(recvtype))); + if (request == MPI_REQUEST_IGNORED) + simgrid::smpi::colls::scatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm); + else + simgrid::smpi::colls::iscatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm, request); + + TRACE_smpi_comm_out(pid); + return MPI_SUCCESS; } -int PMPI_Scatterv(void *sendbuf, int *sendcounts, int *displs, +int PMPI_Scatterv(const void *sendbuf, const int *sendcounts, const int *displs, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm){ return PMPI_Iscatterv(sendbuf, sendcounts, displs, sendtype, recvbuf, recvcount, recvtype, root, comm, MPI_REQUEST_IGNORED); } -int PMPI_Iscatterv(void *sendbuf, int *sendcounts, int *displs, - MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request *request) +int PMPI_Iscatterv(const void* sendbuf, const int* sendcounts, const int* displs, MPI_Datatype sendtype, void* recvbuf, int recvcount, + MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request) { - int retval = 0; - - smpi_bench_end(); - - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if (sendcounts == nullptr || displs == nullptr) { - retval = MPI_ERR_ARG; - } else if (((comm->rank() == root) && (sendtype == MPI_DATATYPE_NULL)) || - ((recvbuf != MPI_IN_PLACE) && (recvtype == MPI_DATATYPE_NULL))) { - retval = MPI_ERR_TYPE; - } else if (request == nullptr){ - retval = MPI_ERR_ARG; - } else { + SET_BUF2(recvbuf) + CHECK_COMM(9) + int rank = comm->rank(); + if(recvbuf != MPI_IN_PLACE){ + CHECK_COUNT(5, recvcount) + CHECK_TYPE(7, recvtype) + CHECK_BUFFER(4, recvbuf, recvcount, recvtype) + } + CHECK_ROOT(9) + CHECK_REQUEST(10) + if (rank == root) { + SET_BUF1(sendbuf) + CHECK_NOT_IN_PLACE_ROOT(1, sendbuf) + CHECK_NULL(2, MPI_ERR_COUNT, sendcounts) + CHECK_NULL(3, MPI_ERR_ARG, displs) + CHECK_TYPE(4, sendtype) + for (int i = 0; i < comm->size(); i++){ + CHECK_COUNT(2, sendcounts[i]) + CHECK_BUFFER(1, sendbuf, sendcounts[i], sendtype) + } if (recvbuf == MPI_IN_PLACE) { recvtype = sendtype; - recvcount = sendcounts[comm->rank()]; - } - int rank = simgrid::s4u::this_actor::get_pid(); - int dt_size_send = sendtype->is_replayable() ? 1 : sendtype->size(); - - std::vector* trace_sendcounts = new std::vector; - if (comm->rank() == root) { - for (int i = 0; i < comm->size(); i++) // copy data to avoid bad free - trace_sendcounts->push_back(sendcounts[i] * dt_size_send); + recvcount = sendcounts[rank]; } - - TRACE_smpi_comm_in(rank, __func__, - new simgrid::instr::VarCollTIData( - request==MPI_REQUEST_IGNORED ? "scatterv":"iscatterv", root, dt_size_send, trace_sendcounts, - recvtype->is_replayable() ? recvcount : recvcount * recvtype->size(), nullptr, - simgrid::smpi::Datatype::encode(sendtype), simgrid::smpi::Datatype::encode(recvtype))); - if(request == MPI_REQUEST_IGNORED) - retval = simgrid::smpi::Colls::scatterv(sendbuf, sendcounts, displs, sendtype, recvbuf, recvcount, recvtype, root, comm); - else - retval = simgrid::smpi::Colls::iscatterv(sendbuf, sendcounts, displs, sendtype, recvbuf, recvcount, recvtype, root, comm, request); - - TRACE_smpi_comm_out(rank); + } else { + CHECK_NOT_IN_PLACE_ROOT(4, recvbuf) } - - smpi_bench_begin(); - return retval; + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Scatterv" : "PMPI_Iscatterv") + + " with root " + std::to_string(root)) + + const SmpiBenchGuard suspend_bench; + + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // no barrier in Iscatterv + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); + + aid_t pid = simgrid::s4u::this_actor::get_pid(); + + auto trace_sendcounts = std::make_shared>(); + if (rank == root) + trace_sendcounts->insert(trace_sendcounts->end(), &sendcounts[0], &sendcounts[comm->size()]); + else //this is not significant outside of root, put 0 as we don't know if sendcounts is initialized + trace_sendcounts->insert(trace_sendcounts->end(), comm->size(), 0); + + + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Scatterv" : "PMPI_Iscatterv", + new simgrid::instr::VarCollTIData( + request == MPI_REQUEST_IGNORED ? "scatterv" : "iscatterv", root, -1, + trace_sendcounts, recvcount, + nullptr, simgrid::smpi::Datatype::encode(sendtype), + simgrid::smpi::Datatype::encode(recvtype))); + if (request == MPI_REQUEST_IGNORED) + simgrid::smpi::colls::scatterv(sendbuf, sendcounts, displs, sendtype, recvbuf, recvcount, recvtype, root, comm); + else + simgrid::smpi::colls::iscatterv(sendbuf, sendcounts, displs, sendtype, recvbuf, recvcount, recvtype, root, comm, + request); + + TRACE_smpi_comm_out(pid); + return MPI_SUCCESS; } -int PMPI_Reduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm) +int PMPI_Reduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm) { - int retval = 0; - - smpi_bench_end(); - - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if (not datatype->is_valid() || op == MPI_OP_NULL) { - retval = MPI_ERR_ARG; - } else { - int rank = simgrid::s4u::this_actor::get_pid(); - - TRACE_smpi_comm_in(rank, __func__, - new simgrid::instr::CollTIData("reduce", root, 0, - datatype->is_replayable() ? count : count * datatype->size(), -1, - simgrid::smpi::Datatype::encode(datatype), "")); - - simgrid::smpi::Colls::reduce(sendbuf, recvbuf, count, datatype, op, root, comm); + return PMPI_Ireduce(sendbuf, recvbuf, count, datatype, op, root, comm, MPI_REQUEST_IGNORED); +} - retval = MPI_SUCCESS; - TRACE_smpi_comm_out(rank); +int PMPI_Ireduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm, MPI_Request* request) +{ + CHECK_COMM(7) + SET_BUF1(sendbuf) + int rank = comm->rank(); + CHECK_TYPE(4, datatype) + CHECK_COUNT(3, count) + CHECK_BUFFER(1, sendbuf, count, datatype) + if(rank == root){ + SET_BUF2(recvbuf) + CHECK_NOT_IN_PLACE(2, recvbuf) + CHECK_BUFFER(5, recvbuf, count, datatype) } - - smpi_bench_begin(); - return retval; + CHECK_OP(5, op, datatype) + CHECK_ROOT(7) + CHECK_REQUEST(8) + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Reduce" : "PMPI_Ireduce") + " with op " + + op->name() + " and root " + std::to_string(root)) + + const SmpiBenchGuard suspend_bench; + + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // no barrier in Ireduce + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); + + aid_t pid = simgrid::s4u::this_actor::get_pid(); + + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce" : "PMPI_Ireduce", + new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "reduce" : "ireduce", root, 0, + count, 0, + simgrid::smpi::Datatype::encode(datatype), "")); + if (request == MPI_REQUEST_IGNORED) + simgrid::smpi::colls::reduce(sendbuf, recvbuf, count, datatype, op, root, comm); + else + simgrid::smpi::colls::ireduce(sendbuf, recvbuf, count, datatype, op, root, comm, request); + + TRACE_smpi_comm_out(pid); + return MPI_SUCCESS; } -int PMPI_Reduce_local(void *inbuf, void *inoutbuf, int count, MPI_Datatype datatype, MPI_Op op){ - int retval = 0; - - smpi_bench_end(); - if (not datatype->is_valid() || op == MPI_OP_NULL) { - retval = MPI_ERR_ARG; - } else { - op->apply(inbuf, inoutbuf, &count, datatype); - retval = MPI_SUCCESS; - } - smpi_bench_begin(); - return retval; +int PMPI_Reduce_local(const void* inbuf, void* inoutbuf, int count, MPI_Datatype datatype, MPI_Op op) +{ + SET_BUF1(inbuf) + SET_BUF2(inoutbuf) + CHECK_TYPE(4, datatype) + CHECK_COUNT(3, count) + CHECK_BUFFER(1, inbuf, count, datatype) + CHECK_BUFFER(2, inoutbuf, count, datatype) + CHECK_OP(5, op, datatype) + + const SmpiBenchGuard suspend_bench; + op->apply(inbuf, inoutbuf, &count, datatype); + return MPI_SUCCESS; } -int PMPI_Allreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +int PMPI_Allreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) { return PMPI_Iallreduce(sendbuf, recvbuf, count, datatype, op, comm, MPI_REQUEST_IGNORED); } -int PMPI_Iallreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request *request) +int PMPI_Iallreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request *request) { - int retval = 0; - - smpi_bench_end(); - - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if (not datatype->is_valid()) { - retval = MPI_ERR_TYPE; - } else if (op == MPI_OP_NULL) { - retval = MPI_ERR_OP; - } else if (request != MPI_REQUEST_IGNORED) { - xbt_die("Iallreduce is not yet implemented. WIP"); - retval = MPI_ERR_ARG; - } else { - - char* sendtmpbuf = static_cast(sendbuf); - if( sendbuf == MPI_IN_PLACE ) { - sendtmpbuf = static_cast(xbt_malloc(count*datatype->get_extent())); - simgrid::smpi::Datatype::copy(recvbuf, count, datatype,sendtmpbuf, count, datatype); - } - int rank = simgrid::s4u::this_actor::get_pid(); - - TRACE_smpi_comm_in(rank, __func__, - new simgrid::instr::CollTIData(request==MPI_REQUEST_IGNORED ? "allreduce":"iallreduce", -1, 0, - datatype->is_replayable() ? count : count * datatype->size(), -1, - simgrid::smpi::Datatype::encode(datatype), "")); + CHECK_COMM(6) + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + int rank = comm->rank(); + CHECK_NOT_IN_PLACE(2, recvbuf) + CHECK_TYPE(4, datatype) + CHECK_OP(5, op, datatype) + CHECK_COUNT(3, count) + CHECK_BUFFER(1, sendbuf, count, datatype) + CHECK_BUFFER(2, recvbuf, count, datatype) + CHECK_REQUEST(7) + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Alleduce" : "PMPI_Iallreduce") + + " with op " + op->name()) + + const SmpiBenchGuard suspend_bench; + + std::vector tmp_sendbuf; + const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, count, datatype); + + aid_t pid = simgrid::s4u::this_actor::get_pid(); + + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Allreduce" : "PMPI_Iallreduce", + new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "allreduce" : "iallreduce", -1, 0, + count, 0, + simgrid::smpi::Datatype::encode(datatype), "")); + + if (request == MPI_REQUEST_IGNORED) + simgrid::smpi::colls::allreduce(real_sendbuf, recvbuf, count, datatype, op, comm); + else + simgrid::smpi::colls::iallreduce(real_sendbuf, recvbuf, count, datatype, op, comm, request); + + TRACE_smpi_comm_out(pid); + return MPI_SUCCESS; +} -// if(request == MPI_REQUEST_IGNORED) - simgrid::smpi::Colls::allreduce(sendtmpbuf, recvbuf, count, datatype, op, comm); -// else -// simgrid::smpi::Colls::iallreduce(sendtmpbuf, recvbuf, count, datatype, op, comm, request); +int PMPI_Scan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +{ + return PMPI_Iscan(sendbuf, recvbuf, count, datatype, op, comm, MPI_REQUEST_IGNORED); +} - if( sendbuf == MPI_IN_PLACE ) - xbt_free(sendtmpbuf); +int PMPI_Iscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request* request) +{ + CHECK_COMM(6) + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + CHECK_TYPE(4, datatype) + CHECK_COUNT(3, count) + CHECK_BUFFER(1,sendbuf,count, datatype) + CHECK_BUFFER(2,recvbuf,count, datatype) + CHECK_REQUEST(7) + CHECK_OP(5, op, datatype) + CHECK_COLLECTIVE(comm, + std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Scan" : "PMPI_Iscan") + " with op " + op->name()) + + const SmpiBenchGuard suspend_bench; + + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // no barrier in Iscan + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); + + aid_t pid = simgrid::s4u::this_actor::get_pid(); + std::vector tmp_sendbuf; + const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, count, datatype); + + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Scan" : "PMPI_Iscan", + new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "scan" : "iscan", -1, 0.0, + count, 0, simgrid::smpi::Datatype::encode(datatype), "")); - retval = MPI_SUCCESS; - TRACE_smpi_comm_out(rank); - } + int retval; + if (request == MPI_REQUEST_IGNORED) + retval = simgrid::smpi::colls::scan(real_sendbuf, recvbuf, count, datatype, op, comm); + else + retval = simgrid::smpi::colls::iscan(real_sendbuf, recvbuf, count, datatype, op, comm, request); - smpi_bench_begin(); + TRACE_smpi_comm_out(pid); return retval; } -int PMPI_Scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +int PMPI_Exscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) { - int retval = 0; - - smpi_bench_end(); - - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if (not datatype->is_valid()) { - retval = MPI_ERR_TYPE; - } else if (op == MPI_OP_NULL) { - retval = MPI_ERR_OP; - } else { - int rank = simgrid::s4u::this_actor::get_pid(); - void* sendtmpbuf = sendbuf; - if (sendbuf == MPI_IN_PLACE) { - sendtmpbuf = static_cast(xbt_malloc(count * datatype->size())); - memcpy(sendtmpbuf, recvbuf, count * datatype->size()); - } - TRACE_smpi_comm_in(rank, __func__, new simgrid::instr::Pt2PtTIData( - "scan", -1, datatype->is_replayable() ? count : count * datatype->size(), - simgrid::smpi::Datatype::encode(datatype))); + return PMPI_Iexscan(sendbuf, recvbuf, count, datatype, op, comm, MPI_REQUEST_IGNORED); +} - retval = simgrid::smpi::Colls::scan(sendtmpbuf, recvbuf, count, datatype, op, comm); +int PMPI_Iexscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request* request){ + CHECK_COMM(6) + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + CHECK_TYPE(4, datatype) + CHECK_COUNT(3, count) + CHECK_BUFFER(1, sendbuf, count, datatype) + CHECK_BUFFER(2, recvbuf, count, datatype) + CHECK_REQUEST(7) + CHECK_OP(5, op, datatype) + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Exscan" : "PMPI_Iexscan") + " with op " + + op->name()) + + const SmpiBenchGuard suspend_bench; + + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // no barrier in Iexscan + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); + + aid_t pid = simgrid::s4u::this_actor::get_pid(); + std::vector tmp_sendbuf; + const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, count, datatype); + + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Exscan" : "PMPI_Iexscan", + new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "exscan" : "iexscan", -1, 0.0, + count, 0, simgrid::smpi::Datatype::encode(datatype), "")); - TRACE_smpi_comm_out(rank); - if (sendbuf == MPI_IN_PLACE) - xbt_free(sendtmpbuf); - } + int retval; + if (request == MPI_REQUEST_IGNORED) + retval = simgrid::smpi::colls::exscan(real_sendbuf, recvbuf, count, datatype, op, comm); + else + retval = simgrid::smpi::colls::iexscan(real_sendbuf, recvbuf, count, datatype, op, comm, request); - smpi_bench_begin(); + TRACE_smpi_comm_out(pid); return retval; } -int PMPI_Exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm){ - int retval = 0; - - smpi_bench_end(); - - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if (not datatype->is_valid()) { - retval = MPI_ERR_TYPE; - } else if (op == MPI_OP_NULL) { - retval = MPI_ERR_OP; - } else { - int rank = simgrid::s4u::this_actor::get_pid(); - void* sendtmpbuf = sendbuf; - if (sendbuf == MPI_IN_PLACE) { - sendtmpbuf = static_cast(xbt_malloc(count * datatype->size())); - memcpy(sendtmpbuf, recvbuf, count * datatype->size()); - } - - TRACE_smpi_comm_in(rank, __func__, new simgrid::instr::Pt2PtTIData( - "exscan", -1, datatype->is_replayable() ? count : count * datatype->size(), - simgrid::smpi::Datatype::encode(datatype))); - - retval = simgrid::smpi::Colls::exscan(sendtmpbuf, recvbuf, count, datatype, op, comm); - - TRACE_smpi_comm_out(rank); - if (sendbuf == MPI_IN_PLACE) - xbt_free(sendtmpbuf); - } - - smpi_bench_begin(); - return retval; +int PMPI_Reduce_scatter(const void *sendbuf, void *recvbuf, const int *recvcounts, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +{ + return PMPI_Ireduce_scatter(sendbuf, recvbuf, recvcounts, datatype, op, comm, MPI_REQUEST_IGNORED); } -int PMPI_Reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +int PMPI_Ireduce_scatter(const void *sendbuf, void *recvbuf, const int *recvcounts, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request *request) { - int retval = 0; - smpi_bench_end(); - - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if (not datatype->is_valid()) { - retval = MPI_ERR_TYPE; - } else if (op == MPI_OP_NULL) { - retval = MPI_ERR_OP; - } else if (recvcounts == nullptr) { - retval = MPI_ERR_ARG; - } else { - int rank = simgrid::s4u::this_actor::get_pid(); - std::vector* trace_recvcounts = new std::vector; - int dt_send_size = datatype->is_replayable() ? 1 : datatype->size(); - int totalcount = 0; - - for (int i = 0; i < comm->size(); i++) { // copy data to avoid bad free - trace_recvcounts->push_back(recvcounts[i] * dt_send_size); - totalcount += recvcounts[i]; - } + CHECK_COMM(6) + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + int rank = comm->rank(); + CHECK_NOT_IN_PLACE(2, recvbuf) + CHECK_TYPE(4, datatype) + CHECK_NULL(3, MPI_ERR_COUNT, recvcounts) + CHECK_REQUEST(7) + CHECK_OP(5, op, datatype) + for (int i = 0; i < comm->size(); i++) { + CHECK_COUNT(3, recvcounts[i]) + CHECK_BUFFER(1, sendbuf, recvcounts[i], datatype) + CHECK_BUFFER(2, recvbuf, recvcounts[i], datatype) + } + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter" : "PMPI_Ireduce_scatter") + + " with op " + op->name()) - void* sendtmpbuf = sendbuf; - if (sendbuf == MPI_IN_PLACE) { - sendtmpbuf = static_cast(xbt_malloc(totalcount * datatype->size())); - memcpy(sendtmpbuf, recvbuf, totalcount * datatype->size()); - } + const SmpiBenchGuard suspend_bench; - TRACE_smpi_comm_in(rank, __func__, new simgrid::instr::VarCollTIData( - "reducescatter", -1, dt_send_size, nullptr, -1, trace_recvcounts, - simgrid::smpi::Datatype::encode(datatype), "")); + aid_t pid = simgrid::s4u::this_actor::get_pid(); + auto trace_recvcounts = std::make_shared>(); + trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[comm->size()]); - simgrid::smpi::Colls::reduce_scatter(sendtmpbuf, recvbuf, recvcounts, datatype, op, comm); - retval = MPI_SUCCESS; - TRACE_smpi_comm_out(rank); + int totalcount = 0; - if (sendbuf == MPI_IN_PLACE) - xbt_free(sendtmpbuf); + for (int i = 0; i < comm->size(); i++) { // copy data to avoid bad free + totalcount += recvcounts[i]; } + std::vector tmp_sendbuf; + const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, totalcount, datatype); - smpi_bench_begin(); - return retval; + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter" : "PMPI_Ireduce_scatter", + new simgrid::instr::VarCollTIData( + request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, -1, nullptr, + -1 , trace_recvcounts, std::to_string(0), simgrid::smpi::Datatype::encode(datatype))); + + if (request == MPI_REQUEST_IGNORED) + simgrid::smpi::colls::reduce_scatter(real_sendbuf, recvbuf, recvcounts, datatype, op, comm); + else + simgrid::smpi::colls::ireduce_scatter(real_sendbuf, recvbuf, recvcounts, datatype, op, comm, request); + + TRACE_smpi_comm_out(pid); + return MPI_SUCCESS; } -int PMPI_Reduce_scatter_block(void *sendbuf, void *recvbuf, int recvcount, +int PMPI_Reduce_scatter_block(const void *sendbuf, void *recvbuf, int recvcount, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) { - int retval; - smpi_bench_end(); - - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if (not datatype->is_valid()) { - retval = MPI_ERR_TYPE; - } else if (op == MPI_OP_NULL) { - retval = MPI_ERR_OP; - } else if (recvcount < 0) { - retval = MPI_ERR_ARG; - } else { - int count = comm->size(); - - int rank = simgrid::s4u::this_actor::get_pid(); - int dt_send_size = datatype->is_replayable() ? 1 : datatype->size(); - std::vector* trace_recvcounts = new std::vector(recvcount * dt_send_size); // copy data to avoid bad free - - void* sendtmpbuf = sendbuf; - if (sendbuf == MPI_IN_PLACE) { - sendtmpbuf = static_cast(xbt_malloc(recvcount * count * datatype->size())); - memcpy(sendtmpbuf, recvbuf, recvcount * count * datatype->size()); - } + return PMPI_Ireduce_scatter_block(sendbuf, recvbuf, recvcount, datatype, op, comm, MPI_REQUEST_IGNORED); +} - TRACE_smpi_comm_in(rank, __func__, - new simgrid::instr::VarCollTIData("reducescatter", -1, 0, nullptr, -1, trace_recvcounts, - simgrid::smpi::Datatype::encode(datatype), "")); +int PMPI_Ireduce_scatter_block(const void* sendbuf, void* recvbuf, int recvcount, MPI_Datatype datatype, MPI_Op op, + MPI_Comm comm, MPI_Request* request) +{ + CHECK_COMM(6) + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + CHECK_TYPE(4, datatype) + CHECK_COUNT(3, recvcount) + CHECK_BUFFER(1, sendbuf, recvcount, datatype) + CHECK_BUFFER(2, recvbuf, recvcount, datatype) + CHECK_REQUEST(7) + CHECK_OP(5, op, datatype) + CHECK_COLLECTIVE( + comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter_block" : "PMPI_Ireduce_scatter_block") + + " with op " + op->name()) + + const SmpiBenchGuard suspend_bench; + + int count = comm->size(); + + aid_t pid = simgrid::s4u::this_actor::get_pid(); + auto trace_recvcounts = std::make_shared>(recvcount); + + std::vector tmp_sendbuf; + const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, recvcount * count, datatype); + + TRACE_smpi_comm_in( + pid, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter_block" : "PMPI_Ireduce_scatter_block", + new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, -1, + nullptr, -1, trace_recvcounts, simgrid::smpi::Datatype::encode(datatype), "")); + + std::vector recvcounts(count); + for (int i = 0; i < count; i++) + recvcounts[i] = recvcount; + if (request == MPI_REQUEST_IGNORED) + simgrid::smpi::colls::reduce_scatter(real_sendbuf, recvbuf, recvcounts.data(), datatype, op, comm); + else + simgrid::smpi::colls::ireduce_scatter(real_sendbuf, recvbuf, recvcounts.data(), datatype, op, comm, request); + + TRACE_smpi_comm_out(pid); + return MPI_SUCCESS; +} - int* recvcounts = new int[count]; - for (int i = 0; i < count; i++) - recvcounts[i] = recvcount; - simgrid::smpi::Colls::reduce_scatter(sendtmpbuf, recvbuf, recvcounts, datatype, op, comm); - delete[] recvcounts; - retval = MPI_SUCCESS; +int PMPI_Alltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, + MPI_Datatype recvtype, MPI_Comm comm){ + return PMPI_Ialltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, MPI_REQUEST_IGNORED); +} - TRACE_smpi_comm_out(rank); +int PMPI_Ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, + MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request) +{ + CHECK_COMM(7) + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + if(sendbuf != MPI_IN_PLACE){ + CHECK_TYPE(3, sendtype) + CHECK_COUNT(2, sendcount) + CHECK_BUFFER(1, sendbuf, sendcount, sendtype) + } + CHECK_TYPE(6, recvtype) + CHECK_COUNT(5, recvcount) + CHECK_COUNT(5, recvcount) + CHECK_BUFFER(4, recvbuf, recvcount, recvtype) + CHECK_REQUEST(8) + CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoall" : "PMPI_Ialltoall") + + aid_t pid = simgrid::s4u::this_actor::get_pid(); + int real_sendcount = sendcount; + MPI_Datatype real_sendtype = sendtype; + + std::vector tmp_sendbuf; + const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, recvcount * comm->size(), recvtype); + + if (sendbuf == MPI_IN_PLACE) { + real_sendcount = recvcount; + real_sendtype = recvtype; + } - if (sendbuf == MPI_IN_PLACE) - xbt_free(sendtmpbuf); + if(recvtype->size() * recvcount != real_sendtype->size() * real_sendcount){ + XBT_WARN("MPI_(I)Alltoall : receive size from each process differs from sent size : %zu vs %zu", recvtype->size() * recvcount, real_sendtype->size() * real_sendcount); + return MPI_ERR_TRUNCATE; } - smpi_bench_begin(); + const SmpiBenchGuard suspend_bench; + + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoall" : "PMPI_Ialltoall", + new simgrid::instr::CollTIData( + request == MPI_REQUEST_IGNORED ? "alltoall" : "ialltoall", -1, -1.0, + real_sendcount, recvcount, + simgrid::smpi::Datatype::encode(real_sendtype), simgrid::smpi::Datatype::encode(recvtype))); + int retval; + if (request == MPI_REQUEST_IGNORED) + retval = + simgrid::smpi::colls::alltoall(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcount, recvtype, comm); + else + retval = simgrid::smpi::colls::ialltoall(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcount, recvtype, + comm, request); + + TRACE_smpi_comm_out(pid); return retval; } -int PMPI_Alltoall(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, - MPI_Datatype recvtype, MPI_Comm comm){ - return PMPI_Ialltoall(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, MPI_REQUEST_IGNORED); + +int PMPI_Alltoallv(const void* sendbuf, const int* sendcounts, const int* senddispls, MPI_Datatype sendtype, void* recvbuf, + const int* recvcounts, const int* recvdispls, MPI_Datatype recvtype, MPI_Comm comm) +{ + return PMPI_Ialltoallv(sendbuf, sendcounts, senddispls, sendtype, recvbuf, recvcounts, recvdispls, recvtype, comm, MPI_REQUEST_IGNORED); } - -int PMPI_Ialltoall(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, - MPI_Datatype recvtype, MPI_Comm comm, MPI_Request *request) + +int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* senddispls, MPI_Datatype sendtype, void* recvbuf, + const int* recvcounts, const int* recvdispls, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request) { - int retval = 0; - smpi_bench_end(); - - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if ((sendbuf != MPI_IN_PLACE && sendtype == MPI_DATATYPE_NULL) || recvtype == MPI_DATATYPE_NULL) { - retval = MPI_ERR_TYPE; - } else if (request == nullptr){ - retval = MPI_ERR_ARG; - } else { - int rank = simgrid::s4u::this_actor::get_pid(); - void* sendtmpbuf = static_cast(sendbuf); - int sendtmpcount = sendcount; - MPI_Datatype sendtmptype = sendtype; - if (sendbuf == MPI_IN_PLACE) { - sendtmpbuf = static_cast(xbt_malloc(recvcount * comm->size() * recvtype->size())); - memcpy(sendtmpbuf, recvbuf, recvcount * comm->size() * recvtype->size()); - sendtmpcount = recvcount; - sendtmptype = recvtype; + CHECK_COMM(9) + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + if(sendbuf != MPI_IN_PLACE){ + CHECK_NULL(2, MPI_ERR_COUNT, sendcounts) + CHECK_NULL(3, MPI_ERR_ARG, senddispls) + CHECK_TYPE(4, sendtype) + } + CHECK_TYPE(8, recvtype) + CHECK_NULL(6, MPI_ERR_COUNT, recvcounts) + CHECK_NULL(7, MPI_ERR_ARG, recvdispls) + CHECK_REQUEST(10) + CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallv" : "PMPI_Ialltoallv") + + aid_t pid = simgrid::s4u::this_actor::get_pid(); + int size = comm->size(); + for (int i = 0; i < size; i++) { + if(sendbuf != MPI_IN_PLACE){ + CHECK_BUFFER(1, sendbuf, sendcounts[i], sendtype) + CHECK_COUNT(2, sendcounts[i]) } + CHECK_BUFFER(5, recvbuf, recvcounts[i], recvtype) + CHECK_COUNT(6, recvcounts[i]) + } - TRACE_smpi_comm_in(rank, __func__, - new simgrid::instr::CollTIData( - request==MPI_REQUEST_IGNORED ? "alltoall" : "ialltoall", -1, -1.0, - sendtmptype->is_replayable() ? sendtmpcount : sendtmpcount * sendtmptype->size(), - recvtype->is_replayable() ? recvcount : recvcount * recvtype->size(), - simgrid::smpi::Datatype::encode(sendtmptype), simgrid::smpi::Datatype::encode(recvtype))); - if(request == MPI_REQUEST_IGNORED) - retval = simgrid::smpi::Colls::alltoall(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcount, recvtype, comm); - else - retval = simgrid::smpi::Colls::ialltoall(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcount, recvtype, comm, request); + const SmpiBenchGuard suspend_bench; + + int send_size = 0; + int recv_size = 0; + auto trace_sendcounts = std::make_shared>(); + auto trace_recvcounts = std::make_shared>(); + trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[size]); + + int dt_size_recv = recvtype->size(); + + const int* real_sendcounts = sendcounts; + const int* real_senddispls = senddispls; + MPI_Datatype real_sendtype = sendtype; + int maxsize = 0; + std::vector tmp_sendbuf; + std::vector tmp_sendcounts; + std::vector tmp_senddispls; + const void* real_sendbuf; + + if (sendbuf == MPI_IN_PLACE) { + tmp_sendcounts.assign(recvcounts, recvcounts + size); + real_sendcounts = tmp_sendcounts.data(); + tmp_senddispls.assign(recvdispls, recvdispls + size); + real_senddispls = tmp_senddispls.data(); + real_sendtype = recvtype; + } - TRACE_smpi_comm_out(rank); + for (int i = 0; i < size; i++) { // copy data to avoid bad free + send_size += real_sendcounts[i] ; + recv_size += recvcounts[i]; + if (((recvdispls[i] + recvcounts[i]) * dt_size_recv) > maxsize) + maxsize = (recvdispls[i] + recvcounts[i]) * dt_size_recv; + } + real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, maxsize, MPI_CHAR); - if (sendbuf == MPI_IN_PLACE) - xbt_free(sendtmpbuf); + if(recvtype->size() * recvcounts[comm->rank()] != real_sendtype->size() * real_sendcounts[comm->rank()]){ + XBT_WARN("MPI_(I)Alltoallv : receive size from me differs from sent size to me : %zu vs %zu", recvtype->size() * recvcounts[comm->rank()], real_sendtype->size() * real_sendcounts[comm->rank()]); + return MPI_ERR_TRUNCATE; } - smpi_bench_begin(); + trace_sendcounts->insert(trace_sendcounts->end(), &real_sendcounts[0], &real_sendcounts[size]); + + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallv" : "PMPI_Ialltoallv", + new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "alltoallv" : "ialltoallv", -1, + send_size, trace_sendcounts, recv_size, trace_recvcounts, + simgrid::smpi::Datatype::encode(real_sendtype), + simgrid::smpi::Datatype::encode(recvtype))); + + int retval; + if (request == MPI_REQUEST_IGNORED) + retval = simgrid::smpi::colls::alltoallv(real_sendbuf, real_sendcounts, real_senddispls, real_sendtype, recvbuf, + recvcounts, recvdispls, recvtype, comm); + else + retval = simgrid::smpi::colls::ialltoallv(real_sendbuf, real_sendcounts, real_senddispls, real_sendtype, recvbuf, + recvcounts, recvdispls, recvtype, comm, request); + + TRACE_smpi_comm_out(pid); return retval; } -int PMPI_Alltoallv(void* sendbuf, int* sendcounts, int* senddisps, MPI_Datatype sendtype, void* recvbuf, - int* recvcounts, int* recvdisps, MPI_Datatype recvtype, MPI_Comm comm) +int PMPI_Alltoallw(const void* sendbuf, const int* sendcounts, const int* senddispls, const MPI_Datatype* sendtypes, void* recvbuf, + const int* recvcounts, const int* recvdispls, const MPI_Datatype* recvtypes, MPI_Comm comm) { - return PMPI_Ialltoallv(sendbuf, sendcounts, senddisps, sendtype, recvbuf, recvcounts, recvdisps, recvtype, comm, MPI_REQUEST_IGNORED); + return PMPI_Ialltoallw(sendbuf, sendcounts, senddispls, sendtypes, recvbuf, recvcounts, recvdispls, recvtypes, comm, MPI_REQUEST_IGNORED); } -int PMPI_Ialltoallv(void* sendbuf, int* sendcounts, int* senddisps, MPI_Datatype sendtype, void* recvbuf, - int* recvcounts, int* recvdisps, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request *request) +int PMPI_Ialltoallw(const void* sendbuf, const int* sendcounts, const int* senddispls, const MPI_Datatype* sendtypes, void* recvbuf, + const int* recvcounts, const int* recvdispls, const MPI_Datatype* recvtypes, MPI_Comm comm, MPI_Request* request) { - int retval = 0; - - smpi_bench_end(); - - if (comm == MPI_COMM_NULL) { - retval = MPI_ERR_COMM; - } else if ((sendbuf != MPI_IN_PLACE && sendtype == MPI_DATATYPE_NULL) || recvtype == MPI_DATATYPE_NULL) { - retval = MPI_ERR_TYPE; - } else if ((sendbuf != MPI_IN_PLACE && (sendcounts == nullptr || senddisps == nullptr)) || recvcounts == nullptr || - recvdisps == nullptr) { - retval = MPI_ERR_ARG; - } else if (request == nullptr){ - retval = MPI_ERR_ARG; - } else { - int rank = simgrid::s4u::this_actor::get_pid(); - int size = comm->size(); - int send_size = 0; - int recv_size = 0; - std::vector* trace_sendcounts = new std::vector; - std::vector* trace_recvcounts = new std::vector; - int dt_size_recv = recvtype->size(); - - void* sendtmpbuf = static_cast(sendbuf); - int* sendtmpcounts = sendcounts; - int* sendtmpdisps = senddisps; - MPI_Datatype sendtmptype = sendtype; - int maxsize = 0; - for (int i = 0; i < size; i++) { // copy data to avoid bad free - recv_size += recvcounts[i] * dt_size_recv; - trace_recvcounts->push_back(recvcounts[i] * dt_size_recv); - if (((recvdisps[i] + recvcounts[i]) * dt_size_recv) > maxsize) - maxsize = (recvdisps[i] + recvcounts[i]) * dt_size_recv; - } - - if (sendbuf == MPI_IN_PLACE) { - sendtmpbuf = static_cast(xbt_malloc(maxsize)); - memcpy(sendtmpbuf, recvbuf, maxsize); - sendtmpcounts = static_cast(xbt_malloc(size * sizeof(int))); - memcpy(sendtmpcounts, recvcounts, size * sizeof(int)); - sendtmpdisps = static_cast(xbt_malloc(size * sizeof(int))); - memcpy(sendtmpdisps, recvdisps, size * sizeof(int)); - sendtmptype = recvtype; + CHECK_COMM(9) + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + if(sendbuf != MPI_IN_PLACE){ + CHECK_NULL(2, MPI_ERR_COUNT, sendcounts) + CHECK_NULL(3, MPI_ERR_ARG, senddispls) + CHECK_NULL(4, MPI_ERR_TYPE, sendtypes) + } + CHECK_NULL(6, MPI_ERR_COUNT, recvcounts) + CHECK_NULL(7, MPI_ERR_ARG, recvdispls) + CHECK_NULL(8, MPI_ERR_TYPE, recvtypes) + CHECK_REQUEST(10) + aid_t pid = simgrid::s4u::this_actor::get_pid(); + int size = comm->size(); + for (int i = 0; i < size; i++) { + if(sendbuf != MPI_IN_PLACE){ + CHECK_COUNT(2, sendcounts[i]) + CHECK_TYPE(4, sendtypes[i]) + CHECK_BUFFER(1, sendbuf, sendcounts[i], sendtypes[i]) } + CHECK_COUNT(6, recvcounts[i]) + CHECK_TYPE(8, recvtypes[i]) + CHECK_BUFFER(5, recvbuf, recvcounts[i], recvtypes[i]) + } + CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallw" : "PMPI_Ialltoallw") + + const SmpiBenchGuard suspend_bench; + + int send_size = 0; + int recv_size = 0; + auto trace_sendcounts = std::make_shared>(); + auto trace_recvcounts = std::make_shared>(); + trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[size]); + + const int* real_sendcounts = sendcounts; + const int* real_senddispls = senddispls; + const MPI_Datatype* real_sendtypes = sendtypes; + + unsigned long maxsize = 0; + for (int i = 0; i < size; i++) { // copy data to avoid bad free + if (recvtypes[i] == MPI_DATATYPE_NULL) + return MPI_ERR_TYPE; + recv_size += recvcounts[i] * recvtypes[i]->size(); + if ((recvdispls[i] + (recvcounts[i] * recvtypes[i]->size())) > maxsize) + maxsize = recvdispls[i] + (recvcounts[i] * recvtypes[i]->size()); + } - int dt_size_send = sendtmptype->size(); - - for (int i = 0; i < size; i++) { // copy data to avoid bad free - send_size += sendtmpcounts[i] * dt_size_send; - trace_sendcounts->push_back(sendtmpcounts[i] * dt_size_send); - } + std::vector tmp_sendbuf; + std::vector tmp_sendcounts; + std::vector tmp_senddispls; + std::vector tmp_sendtypes; + const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, maxsize, MPI_CHAR); + if (sendbuf == MPI_IN_PLACE) { + tmp_sendcounts.assign(recvcounts, recvcounts + size); + real_sendcounts = tmp_sendcounts.data(); + tmp_senddispls.assign(recvdispls, recvdispls + size); + real_senddispls = tmp_senddispls.data(); + tmp_sendtypes.assign(recvtypes, recvtypes + size); + real_sendtypes = tmp_sendtypes.data(); + } - TRACE_smpi_comm_in(rank, __func__, - new simgrid::instr::VarCollTIData(request==MPI_REQUEST_IGNORED ? "alltoallv":"ialltoallv", -1, send_size, trace_sendcounts, recv_size, - trace_recvcounts, simgrid::smpi::Datatype::encode(sendtype), - simgrid::smpi::Datatype::encode(recvtype))); - if(request == MPI_REQUEST_IGNORED) - retval = simgrid::smpi::Colls::alltoallv(sendtmpbuf, sendtmpcounts, sendtmpdisps, sendtmptype, recvbuf, recvcounts, - recvdisps, recvtype, comm); - else - retval = simgrid::smpi::Colls::ialltoallv(sendtmpbuf, sendtmpcounts, sendtmpdisps, sendtmptype, recvbuf, recvcounts, - recvdisps, recvtype, comm, request); - TRACE_smpi_comm_out(rank); + if(recvtypes[comm->rank()]->size() * recvcounts[comm->rank()] != real_sendtypes[comm->rank()]->size() * real_sendcounts[comm->rank()]){ + XBT_WARN("MPI_(I)Alltoallw : receive size from me differs from sent size to me : %zu vs %zu", recvtypes[comm->rank()]->size() * recvcounts[comm->rank()], real_sendtypes[comm->rank()]->size() * real_sendcounts[comm->rank()]); + return MPI_ERR_TRUNCATE; + } - if (sendbuf == MPI_IN_PLACE) { - xbt_free(sendtmpbuf); - xbt_free(sendtmpcounts); - xbt_free(sendtmpdisps); - } + trace_sendcounts->insert(trace_sendcounts->end(), &real_sendcounts[0], &real_sendcounts[size]); + for (int i = 0; i < size; i++) { // copy data to avoid bad free + send_size += real_sendcounts[i] * real_sendtypes[i]->size(); } - smpi_bench_begin(); + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallw" : "PMPI_Ialltoallw", + new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "alltoallv" : "ialltoallv", -1, + send_size, trace_sendcounts, recv_size, trace_recvcounts, + simgrid::smpi::Datatype::encode(real_sendtypes[0]), + simgrid::smpi::Datatype::encode(recvtypes[0]))); + + int retval; + if (request == MPI_REQUEST_IGNORED) + retval = simgrid::smpi::colls::alltoallw(real_sendbuf, real_sendcounts, real_senddispls, real_sendtypes, recvbuf, + recvcounts, recvdispls, recvtypes, comm); + else + retval = simgrid::smpi::colls::ialltoallw(real_sendbuf, real_sendcounts, real_senddispls, real_sendtypes, recvbuf, + recvcounts, recvdispls, recvtypes, comm, request); + + TRACE_smpi_comm_out(pid); return retval; }