X-Git-Url: http://bilbo.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/53b88680c91bba523830793ce69bb8a0fab54382..13b40c020d5701ec7ebedfc0f2840ed263d46a92:/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 62b46a87d6..1c4b10ccfc 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. */ @@ -11,45 +11,17 @@ #include "smpi_op.hpp" #include "src/smpi/include/smpi_actor.hpp" +#include + XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(smpi_pmpi); -#define CHECK_ARGS(test, errcode, ...) \ - if (test) { \ - XBT_WARN(__VA_ARGS__); \ - return (errcode); \ - } - -#define CHECK_COMM(num)\ - CHECK_ARGS(comm == MPI_COMM_NULL, MPI_ERR_COMM,\ - "%s: param %d communicator cannot be MPI_COMM_NULL", __func__, num); -#define CHECK_REQUEST(num)\ - CHECK_ARGS(request == nullptr, MPI_ERR_ARG,\ - "%s: param %d request cannot be NULL",__func__, num); -#define CHECK_BUFFER(num,buf,count)\ - CHECK_ARGS(buf == nullptr && count > 0, MPI_ERR_BUFFER,\ - "%s: param %d %s cannot be NULL if %s > 0",__func__, num, #buf, #count); -#define CHECK_COUNT(num,count)\ - CHECK_ARGS(count < 0, MPI_ERR_COUNT,\ - "%s: param %d %s cannot be negative", __func__, num, #count); -#define CHECK_TYPE(num, datatype)\ - CHECK_ARGS((datatype == MPI_DATATYPE_NULL|| not datatype->is_valid()), MPI_ERR_TYPE,\ - "%s: param %d %s cannot be MPI_DATATYPE_NULL or invalid", __func__, num, #datatype); -#define CHECK_OP(num)\ - CHECK_ARGS(op == MPI_OP_NULL, MPI_ERR_OP,\ - "%s: param %d op cannot be MPI_OP_NULL or invalid", __func__, num); -#define CHECK_ROOT(num)\ - CHECK_ARGS((root < 0 || root >= comm->size()), MPI_ERR_ROOT,\ - "%s: param %d root (=%d) cannot be negative or larger than communicator size (=%d)", __func__, num, root,\ - comm->size()); -#define CHECK_NULL(num,err,buf)\ - CHECK_ARGS(buf == nullptr, err,\ - "%s: param %d %s cannot be NULL", __func__, num, #buf); - - static const void* smpi_get_in_place_buf(const void* inplacebuf, const void* otherbuf,std::unique_ptr& tmp_sendbuf, int count, MPI_Datatype datatype){ +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.reset(new unsigned char[count * datatype->get_extent()]); - simgrid::smpi::Datatype::copy(otherbuf, count, datatype, tmp_sendbuf.get(), count, datatype); - return tmp_sendbuf.get(); + 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; } @@ -140,14 +112,21 @@ int PMPI_Igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void CHECK_ROOT(7) CHECK_REQUEST(9) - smpi_bench_end(); const void* real_sendbuf = sendbuf; int real_sendcount = sendcount; MPI_Datatype real_sendtype = sendtype; - if ((comm->rank() == root) && (sendbuf == MPI_IN_PLACE)) { - real_sendcount = 0; - real_sendtype = recvtype; + if (comm->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; + } } + + smpi_bench_end(); + int rank = simgrid::s4u::this_actor::get_pid(); TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Gather" : "PMPI_Igather", @@ -208,7 +187,7 @@ int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi 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; + auto* 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); @@ -251,12 +230,19 @@ int PMPI_Iallgather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, v CHECK_COUNT(5, recvcount) CHECK_REQUEST(8) - smpi_bench_end(); 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; + } + + smpi_bench_end(); + int rank = simgrid::s4u::this_actor::get_pid(); TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Allgather" : "PMPI_Iallggather", @@ -307,7 +293,7 @@ int PMPI_Iallgatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, 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; + auto* 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); } @@ -351,11 +337,18 @@ int PMPI_Iscatter(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi CHECK_ROOT(8) CHECK_REQUEST(9) - smpi_bench_end(); if (recvbuf == MPI_IN_PLACE) { recvtype = sendtype; recvcount = sendcount; } + + if((comm->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; + } + + smpi_bench_end(); + int rank = simgrid::s4u::this_actor::get_pid(); TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Scatter" : "PMPI_Iscatter", @@ -409,7 +402,7 @@ int PMPI_Iscatterv(const void* sendbuf, const int* sendcounts, const int* displs 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; + auto* 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); @@ -488,7 +481,6 @@ int PMPI_Allreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype d int PMPI_Iallreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request *request) { - CHECK_COMM(6) CHECK_BUFFER(1, sendbuf, count) CHECK_BUFFER(2, recvbuf, count) @@ -498,7 +490,7 @@ int PMPI_Iallreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype CHECK_OP(5) smpi_bench_end(); - std::unique_ptr tmp_sendbuf; + std::vector tmp_sendbuf; const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, count, datatype); int rank = simgrid::s4u::this_actor::get_pid(); @@ -535,7 +527,7 @@ int PMPI_Iscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datat smpi_bench_end(); int rank = simgrid::s4u::this_actor::get_pid(); - std::unique_ptr tmp_sendbuf; + 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_Scan" : "PMPI_Iscan", @@ -570,7 +562,7 @@ int PMPI_Iexscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype dat smpi_bench_end(); int rank = simgrid::s4u::this_actor::get_pid(); - std::unique_ptr tmp_sendbuf; + 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", @@ -609,7 +601,7 @@ int PMPI_Ireduce_scatter(const void *sendbuf, void *recvbuf, const int *recvcoun smpi_bench_end(); int rank = simgrid::s4u::this_actor::get_pid(); - std::vector* trace_recvcounts = new std::vector; + auto* trace_recvcounts = new std::vector(); int dt_send_size = datatype->is_replayable() ? 1 : datatype->size(); int totalcount = 0; @@ -617,7 +609,7 @@ int PMPI_Ireduce_scatter(const void *sendbuf, void *recvbuf, const int *recvcoun trace_recvcounts->push_back(recvcounts[i] * dt_send_size); totalcount += recvcounts[i]; } - std::unique_ptr tmp_sendbuf; + std::vector tmp_sendbuf; const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, totalcount, datatype); TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter" : "PMPI_Ireduce_scatter", @@ -657,8 +649,8 @@ int PMPI_Ireduce_scatter_block(const void* sendbuf, void* recvbuf, int recvcount 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 - std::unique_ptr tmp_sendbuf; + auto* trace_recvcounts = new std::vector(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( @@ -666,14 +658,13 @@ int PMPI_Ireduce_scatter_block(const void* sendbuf, void* recvbuf, int recvcount new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, 0, nullptr, -1, 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(real_sendbuf, recvbuf, recvcounts, datatype, op, comm); + simgrid::smpi::colls::reduce_scatter(real_sendbuf, recvbuf, recvcounts.data(), datatype, op, comm); else - simgrid::smpi::colls::ireduce_scatter(real_sendbuf, 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); smpi_bench_begin(); @@ -700,12 +691,11 @@ int PMPI_Ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, vo CHECK_COUNT(5, recvcount) CHECK_REQUEST(8) - smpi_bench_end(); int rank = simgrid::s4u::this_actor::get_pid(); int real_sendcount = sendcount; MPI_Datatype real_sendtype = sendtype; - - std::unique_ptr tmp_sendbuf; + + 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) { @@ -713,6 +703,13 @@ int PMPI_Ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, vo real_sendtype = recvtype; } + 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_end(); + TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoall" : "PMPI_Ialltoall", new simgrid::instr::CollTIData( request == MPI_REQUEST_IGNORED ? "alltoall" : "ialltoall", -1, -1.0, @@ -766,8 +763,8 @@ int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* sendd smpi_bench_end(); 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 = new std::vector(); + auto* trace_recvcounts = new std::vector(); int dt_size_recv = recvtype->size(); const int* real_sendcounts = sendcounts; @@ -781,20 +778,24 @@ int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* sendd maxsize = (recvdispls[i] + recvcounts[i]) * dt_size_recv; } - std::unique_ptr tmp_sendbuf; - std::unique_ptr tmp_sendcounts; - std::unique_ptr tmp_senddispls; + 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) { - tmp_sendcounts.reset(new int[size]); - std::copy(recvcounts, recvcounts + size, tmp_sendcounts.get()); - real_sendcounts = tmp_sendcounts.get(); - tmp_senddispls.reset(new int[size]); - std::copy(recvdispls, recvdispls + size, tmp_senddispls.get()); - real_senddispls = tmp_senddispls.get(); + 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; } + 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()]); + smpi_bench_begin(); + return MPI_ERR_TRUNCATE; + } + int dt_size_send = real_sendtype->size(); for (int i = 0; i < size; i++) { // copy data to avoid bad free @@ -857,8 +858,8 @@ int PMPI_Ialltoallw(const void* sendbuf, const int* sendcounts, const int* sendd 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 = new std::vector(); + auto* trace_recvcounts = new std::vector(); const int* real_sendcounts = sendcounts; const int* real_senddispls = senddispls; @@ -876,21 +877,25 @@ int PMPI_Ialltoallw(const void* sendbuf, const int* sendcounts, const int* sendd maxsize = recvdispls[i] + (recvcounts[i] * recvtypes[i]->size()); } - std::unique_ptr tmp_sendbuf; - std::unique_ptr tmp_sendcounts; - std::unique_ptr tmp_senddispls; - std::unique_ptr tmp_sendtypes; + 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.reset(new int[size]); - std::copy(recvcounts, recvcounts + size, tmp_sendcounts.get()); - real_sendcounts = tmp_sendcounts.get(); - tmp_senddispls.reset(new int[size]); - std::copy(recvdispls, recvdispls + size, tmp_senddispls.get()); - real_senddispls = tmp_senddispls.get(); - tmp_sendtypes.reset(new MPI_Datatype[size]); - std::copy(recvtypes, recvtypes + size, tmp_sendtypes.get()); - real_sendtypes = tmp_sendtypes.get(); + 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()]); + smpi_bench_begin(); + return MPI_ERR_TRUNCATE; } for (int i = 0; i < size; i++) { // copy data to avoid bad free