X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/a8f0008b377d7e1c7e7cb3bce1f57c97606d8054..6e7cd3b98b151c28f744e0a778d7efb82e076c38:/src/smpi/bindings/smpi_pmpi_coll.cpp diff --git a/src/smpi/bindings/smpi_pmpi_coll.cpp b/src/smpi/bindings/smpi_pmpi_coll.cpp index bd7205b115..534a94a9eb 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) 2007-2021. 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. */ @@ -11,8 +11,21 @@ #include "smpi_op.hpp" #include "src/smpi/include/smpi_actor.hpp" +#include + XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(smpi_pmpi); +static const void* smpi_get_in_place_buf(const void* inplacebuf, const void* otherbuf, + std::vector& tmp_sendbuf, int count, MPI_Datatype datatype) +{ + 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) @@ -22,24 +35,21 @@ int PMPI_Barrier(MPI_Comm comm) int PMPI_Ibarrier(MPI_Comm comm, MPI_Request *request) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if (request == nullptr) - return MPI_ERR_ARG; - - smpi_bench_end(); - int rank = simgrid::s4u::this_actor::get_pid(); - TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Barrier" : "PMPI_Ibarrier", + CHECK_COMM(1) + CHECK_REQUEST(2) + + 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); + 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); + simgrid::smpi::colls::ibarrier(comm, request); - TRACE_smpi_comm_out(rank); - smpi_bench_begin(); + TRACE_smpi_comm_out(pid); return MPI_SUCCESS; } @@ -51,580 +61,587 @@ int PMPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm c int PMPI_Ibcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm, MPI_Request* request) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if (buf == nullptr && count > 0) - return MPI_ERR_BUFFER; - if (datatype == MPI_DATATYPE_NULL || not datatype->is_valid()) - return MPI_ERR_TYPE; - if (count < 0) - return MPI_ERR_COUNT; - if (root < 0 || root >= comm->size()) - return MPI_ERR_ROOT; - if (request == nullptr) - return MPI_ERR_ARG; - - smpi_bench_end(); - int rank = simgrid::s4u::this_actor::get_pid(); - TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Bcast" : "PMPI_Ibcast", + 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) + + 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, - datatype->is_replayable() ? count : count * datatype->size(), -1, + datatype->is_replayable() ? count : count * datatype->size(), 0, simgrid::smpi::Datatype::encode(datatype), "")); if (comm->size() > 1) { if (request == MPI_REQUEST_IGNORED) - simgrid::smpi::Colls::bcast(buf, count, datatype, root, comm); + simgrid::smpi::colls::bcast(buf, count, datatype, root, comm); else - simgrid::smpi::Colls::ibcast(buf, count, datatype, root, comm, request); + simgrid::smpi::colls::ibcast(buf, count, datatype, root, comm, request); } else { if (request != MPI_REQUEST_IGNORED) *request = MPI_REQUEST_NULL; } - TRACE_smpi_comm_out(rank); - smpi_bench_begin(); + 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, +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) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if ((sendbuf == nullptr) || ((comm->rank() == root) && recvbuf == nullptr)) - return MPI_ERR_BUFFER; - if (((sendbuf != MPI_IN_PLACE && sendcount > 0) && (sendtype == MPI_DATATYPE_NULL)) || - ((comm->rank() == root) && (recvtype == MPI_DATATYPE_NULL))) - return MPI_ERR_TYPE; - if (((sendbuf != MPI_IN_PLACE) && (sendcount < 0)) || ((comm->rank() == root) && (recvcount < 0))) - return MPI_ERR_COUNT; - if (root < 0 || root >= comm->size()) - return MPI_ERR_ROOT; - if (request == nullptr) - return MPI_ERR_ARG; - - smpi_bench_end(); - char* sendtmpbuf = static_cast(sendbuf); - int sendtmpcount = sendcount; - MPI_Datatype sendtmptype = sendtype; - if ((comm->rank() == root) && (sendbuf == MPI_IN_PLACE)) { - sendtmpcount = 0; - sendtmptype = recvtype; - } - int rank = simgrid::s4u::this_actor::get_pid(); - - TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Gather" : "PMPI_Igather", + CHECK_COMM(8) + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + 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){ + 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) + + 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; + } + } + + const SmpiBenchGuard suspend_bench; + + 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, - sendtmptype->is_replayable() ? sendtmpcount : sendtmpcount * sendtmptype->size(), + real_sendtype->is_replayable() ? real_sendcount : real_sendcount * real_sendtype->size(), (comm->rank() != root || recvtype->is_replayable()) ? recvcount : recvcount * recvtype->size(), - simgrid::smpi::Datatype::encode(sendtmptype), simgrid::smpi::Datatype::encode(recvtype))); + simgrid::smpi::Datatype::encode(real_sendtype), simgrid::smpi::Datatype::encode(recvtype))); if (request == MPI_REQUEST_IGNORED) - simgrid::smpi::Colls::gather(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcount, recvtype, root, comm); + simgrid::smpi::colls::gather(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcount, recvtype, root, comm); else - simgrid::smpi::Colls::igather(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcount, recvtype, root, comm, + simgrid::smpi::colls::igather(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcount, recvtype, root, comm, request); - TRACE_smpi_comm_out(rank); - smpi_bench_begin(); + 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, +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) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if ((sendbuf == nullptr && sendcount > 0) || ((comm->rank() == root) && recvbuf == nullptr)) - return MPI_ERR_BUFFER; - if (((sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL)) || - ((comm->rank() == root) && (recvtype == MPI_DATATYPE_NULL))) - return MPI_ERR_TYPE; - if ((sendbuf != MPI_IN_PLACE) && (sendcount < 0)) - return MPI_ERR_COUNT; - if ((comm->rank() == root) && (recvcounts == nullptr || displs == nullptr)) - return MPI_ERR_ARG; - if (root < 0 || root >= comm->size()) - return MPI_ERR_ROOT; - if (request == nullptr) - return MPI_ERR_ARG; + CHECK_COMM(9) + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + 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){ + 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) - for (int i = 0; i < comm->size(); i++) { - if ((comm->rank() == root) && (recvcounts[i] < 0)) - return MPI_ERR_COUNT; + if (rank == root){ + for (int i = 0; i < comm->size(); i++) { + CHECK_COUNT(5, recvcounts[i]) + CHECK_BUFFER(4,recvbuf,recvcounts[i], recvtype) + } } - smpi_bench_end(); - char* sendtmpbuf = static_cast(sendbuf); - int sendtmpcount = sendcount; - MPI_Datatype sendtmptype = sendtype; - if ((comm->rank() == root) && (sendbuf == MPI_IN_PLACE)) { - sendtmpcount = 0; - sendtmptype = recvtype; + const SmpiBenchGuard suspend_bench; + 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; } - int rank = simgrid::s4u::this_actor::get_pid(); + aid_t pid = simgrid::s4u::this_actor::get_pid(); int dt_size_recv = recvtype->is_replayable() ? 1 : recvtype->size(); - std::vector* trace_recvcounts = new std::vector; - if (comm->rank() == root) { + auto trace_recvcounts = std::make_shared>(); + if (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); } - TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Gatherv" : "PMPI_Igatherv", + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Gatherv" : "PMPI_Igatherv", 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), + real_sendtype->is_replayable() ? real_sendcount : real_sendcount * real_sendtype->size(), + nullptr, dt_size_recv, trace_recvcounts, simgrid::smpi::Datatype::encode(real_sendtype), simgrid::smpi::Datatype::encode(recvtype))); if (request == MPI_REQUEST_IGNORED) - simgrid::smpi::Colls::gatherv(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcounts, displs, recvtype, root, - comm); + simgrid::smpi::colls::gatherv(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcounts, displs, recvtype, + root, comm); else - simgrid::smpi::Colls::igatherv(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcounts, displs, recvtype, root, - comm, request); + simgrid::smpi::colls::igatherv(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcounts, displs, recvtype, + root, comm, request); - TRACE_smpi_comm_out(rank); - smpi_bench_begin(); + 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, +int PMPI_Iallgather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if ((sendbuf == nullptr && sendcount > 0) || (recvbuf == nullptr)) - return MPI_ERR_BUFFER; - if (((sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL)) || (recvtype == MPI_DATATYPE_NULL)) - return MPI_ERR_TYPE; - if (((sendbuf != MPI_IN_PLACE) && (sendcount < 0)) || (recvcount < 0)) - return MPI_ERR_COUNT; - if (request == nullptr) - return MPI_ERR_ARG; - - smpi_bench_end(); + 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) + } + CHECK_TYPE(6, recvtype) + CHECK_COUNT(5, recvcount) + CHECK_BUFFER(1, sendbuf, sendcount, sendtype) + CHECK_BUFFER(4, recvbuf, recvcount, recvtype) + CHECK_REQUEST(8) + 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, request == MPI_REQUEST_IGNORED ? "PMPI_Allgather" : "PMPI_Iallggather", + 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, 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); + simgrid::smpi::colls::allgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm); else - simgrid::smpi::Colls::iallgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, request); + simgrid::smpi::colls::iallgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, request); - TRACE_smpi_comm_out(rank); - smpi_bench_begin(); + 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, +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) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if ((sendbuf == nullptr && sendcount > 0) || (recvbuf == nullptr)) - return MPI_ERR_BUFFER; - if (((sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL)) || (recvtype == MPI_DATATYPE_NULL)) - return MPI_ERR_TYPE; - if ((sendbuf != MPI_IN_PLACE) && (sendcount < 0)) - return MPI_ERR_COUNT; - if (recvcounts == nullptr || displs == nullptr) - return MPI_ERR_ARG; - if (request == nullptr) - return MPI_ERR_ARG; - - for (int i = 0; i < comm->size(); i++) { // copy data to avoid bad free - if (recvcounts[i] < 0) - return MPI_ERR_COUNT; + 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) } - smpi_bench_end(); + 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; } - int rank = simgrid::s4u::this_actor::get_pid(); + aid_t pid = simgrid::s4u::this_actor::get_pid(); int dt_size_recv = recvtype->is_replayable() ? 1 : recvtype->size(); - std::vector* trace_recvcounts = new std::vector; + auto trace_recvcounts = std::make_shared>(); 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, request == MPI_REQUEST_IGNORED ? "PMPI_Allgatherv" : "PMPI_Iallgatherv", + pid, request == MPI_REQUEST_IGNORED ? "PMPI_Allgatherv" : "PMPI_Iallgatherv", 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); + 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, + simgrid::smpi::colls::iallgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, comm, request); - TRACE_smpi_comm_out(rank); - smpi_bench_begin(); + 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, +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) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if (((comm->rank() == root) && (sendtype == MPI_DATATYPE_NULL || not sendtype->is_valid())) || - ((recvbuf != MPI_IN_PLACE) && (recvtype == MPI_DATATYPE_NULL || not recvtype->is_valid()))) - return MPI_ERR_TYPE; - if (((comm->rank() == root) && (sendcount < 0)) || ((recvbuf != MPI_IN_PLACE) && (recvcount < 0))) - return MPI_ERR_COUNT; - if ((sendbuf == recvbuf) || ((comm->rank() == root) && sendcount > 0 && (sendbuf == nullptr)) || - (recvcount > 0 && recvbuf == nullptr)) - return MPI_ERR_BUFFER; - if (root < 0 || root >= comm->size()) - return MPI_ERR_ROOT; - if (request == nullptr) - return MPI_ERR_ARG; - - smpi_bench_end(); + CHECK_COMM(8) + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + int rank = comm->rank(); + if(rank == root){ + 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) + if (recvbuf == MPI_IN_PLACE) { recvtype = sendtype; recvcount = sendcount; } - int rank = simgrid::s4u::this_actor::get_pid(); - TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Scatter" : "PMPI_Iscatter", + 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; + } + + const SmpiBenchGuard suspend_bench; + + 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, - (comm->rank() != root || sendtype->is_replayable()) ? sendcount : sendcount * sendtype->size(), + (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); + 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); + simgrid::smpi::colls::iscatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm, request); - TRACE_smpi_comm_out(rank); - smpi_bench_begin(); + 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, +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) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if (sendcounts == nullptr || displs == nullptr) - return MPI_ERR_ARG; - if (((comm->rank() == root) && (sendtype == MPI_DATATYPE_NULL)) || - ((recvbuf != MPI_IN_PLACE) && (recvtype == MPI_DATATYPE_NULL))) - return MPI_ERR_TYPE; - if (request == nullptr) - return MPI_ERR_ARG; - if (recvbuf != MPI_IN_PLACE && recvcount < 0) - return MPI_ERR_COUNT; - if (root < 0 || root >= comm->size()) - return MPI_ERR_ROOT; - - if (comm->rank() == root) { + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + CHECK_COMM(9) + int rank = comm->rank(); + if(recvbuf != MPI_IN_PLACE){ + CHECK_NOT_IN_PLACE_ROOT(1, sendbuf) + CHECK_COUNT(5, recvcount) + CHECK_TYPE(7, recvtype) + CHECK_BUFFER(4, recvbuf, recvcount, recvtype) + } + CHECK_ROOT(9) + CHECK_REQUEST(10) + if (rank == root) { + 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()]; - } - for (int i = 0; i < comm->size(); i++) { - if (sendcounts[i] < 0) - return MPI_ERR_COUNT; + recvcount = sendcounts[rank]; } + } else { + CHECK_NOT_IN_PLACE_ROOT(4, recvbuf) } - smpi_bench_end(); + const SmpiBenchGuard suspend_bench; - int rank = simgrid::s4u::this_actor::get_pid(); + aid_t pid = 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) { + auto trace_sendcounts = std::make_shared>(); + if (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); } } - TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Scatterv" : "PMPI_Iscatterv", + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Scatterv" : "PMPI_Iscatterv", 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) - simgrid::smpi::Colls::scatterv(sendbuf, sendcounts, displs, sendtype, recvbuf, recvcount, recvtype, root, comm); + 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, + simgrid::smpi::colls::iscatterv(sendbuf, sendcounts, displs, sendtype, recvbuf, recvcount, recvtype, root, comm, request); - TRACE_smpi_comm_out(rank); - smpi_bench_begin(); + 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) { return PMPI_Ireduce(sendbuf, recvbuf, count, datatype, op, root, comm, MPI_REQUEST_IGNORED); } -int PMPI_Ireduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm, MPI_Request* request) +int PMPI_Ireduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm, MPI_Request* request) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if ((sendbuf == nullptr && count > 0) || ((comm->rank() == root) && recvbuf == nullptr)) - return MPI_ERR_BUFFER; - if (datatype == MPI_DATATYPE_NULL || not datatype->is_valid()) - return MPI_ERR_TYPE; - if (op == MPI_OP_NULL) - return MPI_ERR_OP; - if (request == nullptr) - return MPI_ERR_ARG; - if (root < 0 || root >= comm->size()) - return MPI_ERR_ROOT; - if (count < 0) - return MPI_ERR_COUNT; - - smpi_bench_end(); - int rank = simgrid::s4u::this_actor::get_pid(); - - TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce" : "PMPI_Ireduce", + CHECK_COMM(7) + SET_BUF1(sendbuf) + SET_BUF2(recvbuf) + int rank = comm->rank(); + CHECK_TYPE(4, datatype) + CHECK_COUNT(3, count) + CHECK_BUFFER(1, sendbuf, count, datatype) + if(rank == root){ + CHECK_NOT_IN_PLACE(2, recvbuf) + CHECK_BUFFER(5, recvbuf, count, datatype) + } + CHECK_OP(5, op, datatype) + CHECK_ROOT(7) + CHECK_REQUEST(8) + + const SmpiBenchGuard suspend_bench; + 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, - datatype->is_replayable() ? count : count * datatype->size(), -1, + datatype->is_replayable() ? count : count * datatype->size(), 0, simgrid::smpi::Datatype::encode(datatype), "")); if (request == MPI_REQUEST_IGNORED) - simgrid::smpi::Colls::reduce(sendbuf, recvbuf, count, datatype, op, root, comm); + simgrid::smpi::colls::reduce(sendbuf, recvbuf, count, datatype, op, root, comm); else - simgrid::smpi::Colls::ireduce(sendbuf, recvbuf, count, datatype, op, root, comm, request); + simgrid::smpi::colls::ireduce(sendbuf, recvbuf, count, datatype, op, root, comm, request); - TRACE_smpi_comm_out(rank); - smpi_bench_begin(); + 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 PMPI_Reduce_local(const void* inbuf, void* inoutbuf, int count, MPI_Datatype datatype, MPI_Op op) { - if (datatype == MPI_DATATYPE_NULL || not datatype->is_valid()) - return MPI_ERR_TYPE; - if (op == MPI_OP_NULL) - return MPI_ERR_OP; - if (count < 0) - return MPI_ERR_COUNT; - - smpi_bench_end(); + 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); - smpi_bench_begin(); 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) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if ((sendbuf == nullptr && count > 0) || (recvbuf == nullptr)) - return MPI_ERR_BUFFER; - if (datatype == MPI_DATATYPE_NULL || not datatype->is_valid()) - return MPI_ERR_TYPE; - if (count < 0) - return MPI_ERR_COUNT; - if (op == MPI_OP_NULL) - return MPI_ERR_OP; - if (request == nullptr) - return MPI_ERR_ARG; - - smpi_bench_end(); - 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, request == MPI_REQUEST_IGNORED ? "PMPI_Allreduce" : "PMPI_Iallreduce", + 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) + + 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, - datatype->is_replayable() ? count : count * datatype->size(), -1, + datatype->is_replayable() ? count : count * datatype->size(), 0, simgrid::smpi::Datatype::encode(datatype), "")); if (request == MPI_REQUEST_IGNORED) - simgrid::smpi::Colls::allreduce(sendtmpbuf, recvbuf, count, datatype, op, comm); + simgrid::smpi::colls::allreduce(real_sendbuf, recvbuf, count, datatype, op, comm); else - simgrid::smpi::Colls::iallreduce(sendtmpbuf, recvbuf, count, datatype, op, comm, request); + simgrid::smpi::colls::iallreduce(real_sendbuf, recvbuf, count, datatype, op, comm, request); - if (sendbuf == MPI_IN_PLACE) - xbt_free(sendtmpbuf); - - TRACE_smpi_comm_out(rank); - smpi_bench_begin(); + TRACE_smpi_comm_out(pid); return MPI_SUCCESS; } -int PMPI_Scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +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); } -int PMPI_Iscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request* request) +int PMPI_Iscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request* request) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if (datatype == MPI_DATATYPE_NULL || not datatype->is_valid()) - return MPI_ERR_TYPE; - if (op == MPI_OP_NULL) - return MPI_ERR_OP; - if (request == nullptr) - return MPI_ERR_ARG; - if (count < 0) - return MPI_ERR_COUNT; - if (sendbuf == nullptr || recvbuf == nullptr) - return MPI_ERR_BUFFER; - - smpi_bench_end(); - 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, request == MPI_REQUEST_IGNORED ? "PMPI_Scan" : "PMPI_Iscan", + 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) + + const SmpiBenchGuard suspend_bench; + 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::Pt2PtTIData(request == MPI_REQUEST_IGNORED ? "scan" : "iscan", -1, datatype->is_replayable() ? count : count * datatype->size(), simgrid::smpi::Datatype::encode(datatype))); int retval; if (request == MPI_REQUEST_IGNORED) - retval = simgrid::smpi::Colls::scan(sendtmpbuf, recvbuf, count, datatype, op, comm); + retval = simgrid::smpi::colls::scan(real_sendbuf, recvbuf, count, datatype, op, comm); else - retval = simgrid::smpi::Colls::iscan(sendtmpbuf, recvbuf, count, datatype, op, comm, request); + retval = simgrid::smpi::colls::iscan(real_sendbuf, recvbuf, count, datatype, op, comm, request); - TRACE_smpi_comm_out(rank); - if (sendbuf == MPI_IN_PLACE) - xbt_free(sendtmpbuf); - 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 PMPI_Exscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) { return PMPI_Iexscan(sendbuf, recvbuf, count, datatype, op, comm, MPI_REQUEST_IGNORED); } -int PMPI_Iexscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request* request){ - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if (not datatype->is_valid()) - return MPI_ERR_TYPE; - if (op == MPI_OP_NULL) - return MPI_ERR_OP; - if (request == nullptr) - return MPI_ERR_ARG; - if (count < 0) - return MPI_ERR_COUNT; - if (sendbuf == nullptr || recvbuf == nullptr) - return MPI_ERR_BUFFER; - - smpi_bench_end(); - 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()); - } +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) + + const SmpiBenchGuard suspend_bench; + 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(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Exscan" : "PMPI_Iexscan", + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Exscan" : "PMPI_Iexscan", new simgrid::instr::Pt2PtTIData(request == MPI_REQUEST_IGNORED ? "exscan" : "iexscan", -1, datatype->is_replayable() ? count : count * datatype->size(), simgrid::smpi::Datatype::encode(datatype))); int retval; if (request == MPI_REQUEST_IGNORED) - retval = simgrid::smpi::Colls::exscan(sendtmpbuf, recvbuf, count, datatype, op, comm); + retval = simgrid::smpi::colls::exscan(real_sendbuf, recvbuf, count, datatype, op, comm); else - retval = simgrid::smpi::Colls::iexscan(sendtmpbuf, recvbuf, count, datatype, op, comm, request); + retval = simgrid::smpi::colls::iexscan(real_sendbuf, recvbuf, count, datatype, op, comm, request); - TRACE_smpi_comm_out(rank); - if (sendbuf == MPI_IN_PLACE) - xbt_free(sendtmpbuf); - smpi_bench_begin(); + TRACE_smpi_comm_out(pid); return retval; } -int PMPI_Reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +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_Ireduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request *request) +int PMPI_Ireduce_scatter(const void *sendbuf, void *recvbuf, const int *recvcounts, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request *request) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if ((sendbuf == nullptr) || (recvbuf == nullptr)) - return MPI_ERR_BUFFER; - if (datatype == MPI_DATATYPE_NULL || not datatype->is_valid()) - return MPI_ERR_TYPE; - if (op == MPI_OP_NULL) - return MPI_ERR_OP; - if (recvcounts == nullptr) - return MPI_ERR_ARG; - if (request == nullptr) - return MPI_ERR_ARG; - - for (int i = 0; i < comm->size(); i++) { // copy data to avoid bad free - if (recvcounts[i] < 0) - return MPI_ERR_COUNT; + 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) } - smpi_bench_end(); - int rank = simgrid::s4u::this_actor::get_pid(); - std::vector* trace_recvcounts = new std::vector; + const SmpiBenchGuard suspend_bench; + aid_t pid = simgrid::s4u::this_actor::get_pid(); + auto trace_recvcounts = std::make_shared>(); int dt_send_size = datatype->is_replayable() ? 1 : datatype->size(); int totalcount = 0; @@ -632,312 +649,313 @@ int PMPI_Ireduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, MPI_Data trace_recvcounts->push_back(recvcounts[i] * dt_send_size); totalcount += recvcounts[i]; } + std::vector tmp_sendbuf; + const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, totalcount, datatype); - void* sendtmpbuf = sendbuf; - if (sendbuf == MPI_IN_PLACE) { - sendtmpbuf = static_cast(xbt_malloc(totalcount * datatype->size())); - memcpy(sendtmpbuf, recvbuf, totalcount * datatype->size()); - } - - TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter" : "PMPI_Ireduce_scatter", + 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, dt_send_size, nullptr, - -1, trace_recvcounts, simgrid::smpi::Datatype::encode(datatype), "")); + 0, trace_recvcounts, simgrid::smpi::Datatype::encode(datatype), "")); if (request == MPI_REQUEST_IGNORED) - simgrid::smpi::Colls::reduce_scatter(sendtmpbuf, recvbuf, recvcounts, datatype, op, comm); + simgrid::smpi::colls::reduce_scatter(real_sendbuf, recvbuf, recvcounts, datatype, op, comm); else - simgrid::smpi::Colls::ireduce_scatter(sendtmpbuf, recvbuf, recvcounts, datatype, op, comm, request); + simgrid::smpi::colls::ireduce_scatter(real_sendbuf, recvbuf, recvcounts, datatype, op, comm, request); - TRACE_smpi_comm_out(rank); - if (sendbuf == MPI_IN_PLACE) - xbt_free(sendtmpbuf); - smpi_bench_begin(); + 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) { return PMPI_Ireduce_scatter_block(sendbuf, recvbuf, recvcount, datatype, op, comm, MPI_REQUEST_IGNORED); } -int PMPI_Ireduce_scatter_block(void* sendbuf, void* recvbuf, int recvcount, MPI_Datatype datatype, MPI_Op op, +int PMPI_Ireduce_scatter_block(const void* sendbuf, void* recvbuf, int recvcount, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request* request) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if (not datatype->is_valid()) - return MPI_ERR_TYPE; - if (op == MPI_OP_NULL) - return MPI_ERR_OP; - if (recvcount < 0) - return MPI_ERR_ARG; - if (request == nullptr) - return MPI_ERR_ARG; - - smpi_bench_end(); + 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) + + const SmpiBenchGuard suspend_bench; int count = comm->size(); - int rank = simgrid::s4u::this_actor::get_pid(); + aid_t pid = 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()); - } + auto trace_recvcounts = std::make_shared>(recvcount * dt_send_size); // copy data to avoid bad free + std::vector tmp_sendbuf; + const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, recvcount * count, datatype); TRACE_smpi_comm_in( - rank, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter_block" : "PMPI_Ireduce_scatter_block", + pid, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter_block" : "PMPI_Ireduce_scatter_block", new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, 0, - nullptr, -1, trace_recvcounts, simgrid::smpi::Datatype::encode(datatype), "")); + nullptr, 0, trace_recvcounts, simgrid::smpi::Datatype::encode(datatype), "")); - int* recvcounts = new int[count]; + std::vector recvcounts(count); for (int i = 0; i < count; i++) recvcounts[i] = recvcount; if (request == MPI_REQUEST_IGNORED) - simgrid::smpi::Colls::reduce_scatter(sendtmpbuf, recvbuf, recvcounts, datatype, op, comm); + simgrid::smpi::colls::reduce_scatter(real_sendbuf, recvbuf, recvcounts.data(), datatype, op, comm); else - simgrid::smpi::Colls::ireduce_scatter(sendtmpbuf, recvbuf, recvcounts, datatype, op, comm, request); - delete[] recvcounts; + simgrid::smpi::colls::ireduce_scatter(real_sendbuf, recvbuf, recvcounts.data(), datatype, op, comm, request); - TRACE_smpi_comm_out(rank); - if (sendbuf == MPI_IN_PLACE) - xbt_free(sendtmpbuf); - smpi_bench_begin(); + TRACE_smpi_comm_out(pid); return MPI_SUCCESS; } -int PMPI_Alltoall(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, +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); } -int PMPI_Ialltoall(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, +int PMPI_Ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if ((sendbuf == nullptr && sendcount > 0) || (recvbuf == nullptr && recvcount > 0)) - return MPI_ERR_BUFFER; - if ((sendbuf != MPI_IN_PLACE && sendtype == MPI_DATATYPE_NULL) || recvtype == MPI_DATATYPE_NULL) - return MPI_ERR_TYPE; - if ((sendbuf != MPI_IN_PLACE && sendcount < 0) || recvcount < 0) - return MPI_ERR_COUNT; - if (request == nullptr) - return MPI_ERR_ARG; - - smpi_bench_end(); - int rank = simgrid::s4u::this_actor::get_pid(); - void* sendtmpbuf = static_cast(sendbuf); - int sendtmpcount = sendcount; - MPI_Datatype sendtmptype = sendtype; + 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) + + 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) { - sendtmpbuf = static_cast(xbt_malloc(recvcount * comm->size() * recvtype->size())); - // memcpy(??,nullptr,0) is actually undefined behavor, even if harmless. - if (recvbuf != nullptr) - memcpy(sendtmpbuf, recvbuf, recvcount * comm->size() * recvtype->size()); - sendtmpcount = recvcount; - sendtmptype = recvtype; + real_sendcount = recvcount; + real_sendtype = recvtype; } - TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoall" : "PMPI_Ialltoall", + 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; + } + + 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, - sendtmptype->is_replayable() ? sendtmpcount : sendtmpcount * sendtmptype->size(), + real_sendtype->is_replayable() ? real_sendcount : real_sendcount * real_sendtype->size(), recvtype->is_replayable() ? recvcount : recvcount * recvtype->size(), - simgrid::smpi::Datatype::encode(sendtmptype), simgrid::smpi::Datatype::encode(recvtype))); + simgrid::smpi::Datatype::encode(real_sendtype), simgrid::smpi::Datatype::encode(recvtype))); int retval; if (request == MPI_REQUEST_IGNORED) - retval = simgrid::smpi::Colls::alltoall(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcount, recvtype, comm); + retval = + simgrid::smpi::colls::alltoall(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcount, recvtype, comm); else - retval = simgrid::smpi::Colls::ialltoall(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcount, recvtype, comm, - request); + retval = simgrid::smpi::colls::ialltoall(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcount, recvtype, + comm, request); - TRACE_smpi_comm_out(rank); - if (sendbuf == MPI_IN_PLACE) - xbt_free(sendtmpbuf); - smpi_bench_begin(); + 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_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, senddisps, sendtype, recvbuf, recvcounts, recvdisps, recvtype, comm, MPI_REQUEST_IGNORED); + return PMPI_Ialltoallv(sendbuf, sendcounts, senddispls, sendtype, recvbuf, recvcounts, recvdispls, recvtype, 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_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) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if (sendbuf == nullptr || recvbuf == nullptr) - return MPI_ERR_BUFFER; - if ((sendbuf != MPI_IN_PLACE && sendtype == MPI_DATATYPE_NULL) || recvtype == MPI_DATATYPE_NULL) - return MPI_ERR_TYPE; - if ((sendbuf != MPI_IN_PLACE && (sendcounts == nullptr || senddisps == nullptr)) || recvcounts == nullptr || - recvdisps == nullptr) - return MPI_ERR_ARG; - if (request == nullptr) - return MPI_ERR_ARG; - - int rank = simgrid::s4u::this_actor::get_pid(); + 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) + + aid_t pid = simgrid::s4u::this_actor::get_pid(); int size = comm->size(); for (int i = 0; i < size; i++) { - if (recvcounts[i] < 0 || (sendbuf != MPI_IN_PLACE && sendcounts[i] < 0)) - return MPI_ERR_COUNT; + 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]) } - smpi_bench_end(); + const SmpiBenchGuard suspend_bench; int send_size = 0; int recv_size = 0; - std::vector* trace_sendcounts = new std::vector; - std::vector* trace_recvcounts = new std::vector; + auto trace_sendcounts = std::make_shared>(); + auto trace_recvcounts = std::make_shared>(); int dt_size_recv = recvtype->size(); - void* sendtmpbuf = static_cast(sendbuf); - int* sendtmpcounts = sendcounts; - int* sendtmpdisps = senddisps; - MPI_Datatype sendtmptype = sendtype; + const int* real_sendcounts = sendcounts; + const int* real_senddispls = senddispls; + MPI_Datatype real_sendtype = 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 (((recvdispls[i] + recvcounts[i]) * dt_size_recv) > maxsize) + maxsize = (recvdispls[i] + recvcounts[i]) * dt_size_recv; } + std::vector tmp_sendbuf; + std::vector tmp_sendcounts; + std::vector tmp_senddispls; + const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, maxsize, MPI_CHAR); 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; + 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; } - int dt_size_send = sendtmptype->size(); + 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; + } + + int dt_size_send = real_sendtype->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); + send_size += real_sendcounts[i] * dt_size_send; + trace_sendcounts->push_back(real_sendcounts[i] * dt_size_send); } - TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallv" : "PMPI_Ialltoallv", + 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(sendtmptype), + simgrid::smpi::Datatype::encode(real_sendtype), simgrid::smpi::Datatype::encode(recvtype))); int retval; if (request == MPI_REQUEST_IGNORED) - retval = simgrid::smpi::Colls::alltoallv(sendtmpbuf, sendtmpcounts, sendtmpdisps, sendtmptype, recvbuf, recvcounts, - recvdisps, recvtype, comm); + retval = simgrid::smpi::colls::alltoallv(real_sendbuf, real_sendcounts, real_senddispls, real_sendtype, recvbuf, + recvcounts, recvdispls, recvtype, comm); else - retval = simgrid::smpi::Colls::ialltoallv(sendtmpbuf, sendtmpcounts, sendtmpdisps, sendtmptype, recvbuf, recvcounts, - recvdisps, recvtype, comm, request); + retval = simgrid::smpi::colls::ialltoallv(real_sendbuf, real_sendcounts, real_senddispls, real_sendtype, recvbuf, + recvcounts, recvdispls, recvtype, comm, request); - TRACE_smpi_comm_out(rank); - if (sendbuf == MPI_IN_PLACE) { - xbt_free(sendtmpbuf); - xbt_free(sendtmpcounts); - xbt_free(sendtmpdisps); - } - smpi_bench_begin(); + TRACE_smpi_comm_out(pid); return retval; } -int PMPI_Alltoallw(void* sendbuf, int* sendcounts, int* senddisps, MPI_Datatype* sendtypes, void* recvbuf, - int* recvcounts, int* recvdisps, MPI_Datatype* recvtypes, 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_Ialltoallw(sendbuf, sendcounts, senddisps, sendtypes, recvbuf, recvcounts, recvdisps, recvtypes, comm, MPI_REQUEST_IGNORED); + return PMPI_Ialltoallw(sendbuf, sendcounts, senddispls, sendtypes, recvbuf, recvcounts, recvdispls, recvtypes, comm, MPI_REQUEST_IGNORED); } -int PMPI_Ialltoallw(void* sendbuf, int* sendcounts, int* senddisps, MPI_Datatype* sendtypes, void* recvbuf, - int* recvcounts, int* recvdisps, MPI_Datatype* recvtypes, 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) { - if (comm == MPI_COMM_NULL) - return MPI_ERR_COMM; - if (sendbuf == nullptr || recvbuf == nullptr) - return MPI_ERR_BUFFER; - if ((sendbuf != MPI_IN_PLACE && sendtypes == nullptr) || recvtypes == nullptr) - return MPI_ERR_TYPE; - if ((sendbuf != MPI_IN_PLACE && (sendcounts == nullptr || senddisps == nullptr)) || recvcounts == nullptr || - recvdisps == nullptr) - return MPI_ERR_ARG; - if (request == nullptr) - return MPI_ERR_ARG; - - smpi_bench_end(); - int rank = simgrid::s4u::this_actor::get_pid(); + 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++) { // copy data to avoid bad free - if (recvcounts[i] < 0 || (sendbuf != MPI_IN_PLACE && sendcounts[i] < 0)) - return MPI_ERR_COUNT; + 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]) } + + const SmpiBenchGuard suspend_bench; + int send_size = 0; int recv_size = 0; - std::vector* trace_sendcounts = new std::vector; - std::vector* trace_recvcounts = new std::vector; + auto trace_sendcounts = std::make_shared>(); + auto trace_recvcounts = std::make_shared>(); - void* sendtmpbuf = static_cast(sendbuf); - int* sendtmpcounts = sendcounts; - int* sendtmpdisps = senddisps; - MPI_Datatype* sendtmptypes = sendtypes; + 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) { - delete trace_recvcounts; - delete trace_sendcounts; + if (recvtypes[i] == MPI_DATATYPE_NULL) return MPI_ERR_TYPE; - } recv_size += recvcounts[i] * recvtypes[i]->size(); trace_recvcounts->push_back(recvcounts[i] * recvtypes[i]->size()); - if ((recvdisps[i] + (recvcounts[i] * recvtypes[i]->size())) > maxsize) - maxsize = recvdisps[i] + (recvcounts[i] * recvtypes[i]->size()); + if ((recvdispls[i] + (recvcounts[i] * recvtypes[i]->size())) > maxsize) + maxsize = recvdispls[i] + (recvcounts[i] * recvtypes[i]->size()); } + 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) { - 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)); - sendtmptypes = static_cast(xbt_malloc(size * sizeof(MPI_Datatype))); - memcpy(sendtmptypes, recvtypes, size * sizeof(MPI_Datatype)); + 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(); + } + + + 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; } for (int i = 0; i < size; i++) { // copy data to avoid bad free - send_size += sendtmpcounts[i] * sendtmptypes[i]->size(); - trace_sendcounts->push_back(sendtmpcounts[i] * sendtmptypes[i]->size()); + send_size += real_sendcounts[i] * real_sendtypes[i]->size(); + trace_sendcounts->push_back(real_sendcounts[i] * real_sendtypes[i]->size()); } - TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallw" : "PMPI_Ialltoallw", + 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(sendtmptypes[0]), + 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(sendtmpbuf, sendtmpcounts, sendtmpdisps, sendtmptypes, recvbuf, recvcounts, - recvdisps, recvtypes, comm); + retval = simgrid::smpi::colls::alltoallw(real_sendbuf, real_sendcounts, real_senddispls, real_sendtypes, recvbuf, + recvcounts, recvdispls, recvtypes, comm); else - retval = simgrid::smpi::Colls::ialltoallw(sendtmpbuf, sendtmpcounts, sendtmpdisps, sendtmptypes, recvbuf, - recvcounts, recvdisps, recvtypes, comm, request); + retval = simgrid::smpi::colls::ialltoallw(real_sendbuf, real_sendcounts, real_senddispls, real_sendtypes, recvbuf, + recvcounts, recvdispls, recvtypes, comm, request); - TRACE_smpi_comm_out(rank); - if (sendbuf == MPI_IN_PLACE) { - xbt_free(sendtmpbuf); - xbt_free(sendtmpcounts); - xbt_free(sendtmpdisps); - xbt_free(sendtmptypes); - } - smpi_bench_begin(); + TRACE_smpi_comm_out(pid); return retval; }