X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/1d8cff9474b8df5b4867d191888a0aca3421f2a5..9d1a9a4b871895531b7e70f313691ef75dc47a96:/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 0bc8a353b8..e4a607654d 100644 --- a/src/smpi/bindings/smpi_pmpi_coll.cpp +++ b/src/smpi/bindings/smpi_pmpi_coll.cpp @@ -13,6 +13,21 @@ XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(smpi_pmpi); +#define CHECK_ARGS(test, errcode, ...) \ + if (test) { \ + XBT_WARN(__VA_ARGS__); \ + return (errcode); \ + } + + static const void* smpi_get_in_place_buf(const void* inplacebuf, const void* otherbuf,std::unique_ptr& 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(); + }else{ + return inplacebuf; + } +} /* PMPI User level calls */ int PMPI_Barrier(MPI_Comm comm) @@ -32,11 +47,11 @@ int PMPI_Ibarrier(MPI_Comm comm, MPI_Request *request) TRACE_smpi_comm_in(rank, 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(); @@ -72,9 +87,9 @@ int PMPI_Ibcast(void *buf, int count, MPI_Datatype datatype, 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; @@ -95,7 +110,7 @@ int PMPI_Igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void { if (comm == MPI_COMM_NULL) return MPI_ERR_COMM; - if ((sendbuf == nullptr) || ((comm->rank() == root) && recvbuf == nullptr)) + if ((sendbuf == nullptr && sendcount > 0) || ((comm->rank() == root) && recvbuf == nullptr && recvcount > 0)) return MPI_ERR_BUFFER; if (((sendbuf != MPI_IN_PLACE && sendcount > 0) && (sendtype == MPI_DATATYPE_NULL)) || ((comm->rank() == root) && (recvtype == MPI_DATATYPE_NULL))) @@ -108,25 +123,25 @@ int PMPI_Igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void return MPI_ERR_ARG; smpi_bench_end(); - const char* sendtmpbuf = static_cast(sendbuf); - int sendtmpcount = sendcount; - MPI_Datatype sendtmptype = sendtype; + const void* real_sendbuf = sendbuf; + int real_sendcount = sendcount; + MPI_Datatype real_sendtype = sendtype; if ((comm->rank() == root) && (sendbuf == MPI_IN_PLACE)) { - sendtmpcount = 0; - sendtmptype = recvtype; + real_sendcount = 0; + real_sendtype = recvtype; } int rank = simgrid::s4u::this_actor::get_pid(); TRACE_smpi_comm_in(rank, 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); @@ -164,12 +179,12 @@ int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi } smpi_bench_end(); - const char* sendtmpbuf = static_cast(sendbuf); - int sendtmpcount = sendcount; - MPI_Datatype sendtmptype = sendtype; + const void* real_sendbuf = sendbuf; + int real_sendcount = sendcount; + MPI_Datatype real_sendtype = sendtype; if ((comm->rank() == root) && (sendbuf == MPI_IN_PLACE)) { - sendtmpcount = 0; - sendtmptype = recvtype; + real_sendcount = 0; + real_sendtype = recvtype; } int rank = simgrid::s4u::this_actor::get_pid(); @@ -184,15 +199,15 @@ int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi TRACE_smpi_comm_in(rank, 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(); @@ -207,16 +222,17 @@ int PMPI_Allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype, 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; + CHECK_ARGS(comm == MPI_COMM_NULL, MPI_ERR_COMM, "Iallgather: the communicator cannot be MPI_COMM_NULL"); + CHECK_ARGS(recvbuf == nullptr && recvcount > 0, MPI_ERR_BUFFER, "Iallgather: param 4 recvbuf cannot be NULL"); + CHECK_ARGS(sendbuf == nullptr && sendcount > 0, MPI_ERR_BUFFER, + "Iallgather: param 1 sendbuf cannot be NULL when sendcound > 0"); + CHECK_ARGS((sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL), MPI_ERR_TYPE, + "Iallgather: param 3 sendtype cannot be MPI_DATATYPE_NULL when sendbuff is not MPI_IN_PLACE"); + CHECK_ARGS(recvtype == MPI_DATATYPE_NULL, MPI_ERR_TYPE, "Iallgather: param 6 recvtype cannot be MPI_DATATYPE_NULL"); + CHECK_ARGS(recvcount < 0, MPI_ERR_COUNT, "Iallgather: param 5 recvcount cannot be negative"); + CHECK_ARGS((sendbuf != MPI_IN_PLACE) && (sendcount < 0), MPI_ERR_COUNT, + "Iallgather: param 2 sendcount cannot be negative when sendbuf is not MPI_IN_PLACE"); + CHECK_ARGS(request == nullptr, MPI_ERR_ARG, "Iallgather: param 8 request cannot be NULL"); smpi_bench_end(); if (sendbuf == MPI_IN_PLACE) { @@ -233,9 +249,9 @@ int PMPI_Iallgather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, v 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(); @@ -252,7 +268,7 @@ int PMPI_Iallgatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, { if (comm == MPI_COMM_NULL) return MPI_ERR_COMM; - if ((sendbuf == nullptr && sendcount > 0) || (recvbuf == nullptr)) + if (sendbuf == nullptr && sendcount > 0) return MPI_ERR_BUFFER; if (((sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL)) || (recvtype == MPI_DATATYPE_NULL)) return MPI_ERR_TYPE; @@ -263,9 +279,11 @@ int PMPI_Iallgatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, if (request == nullptr) return MPI_ERR_ARG; - for (int i = 0; i < comm->size(); i++) { // copy data to avoid bad free + for (int i = 0; i < comm->size(); i++) { if (recvcounts[i] < 0) return MPI_ERR_COUNT; + else if (recvcounts[i] > 0 && recvbuf == nullptr) + return MPI_ERR_BUFFER; } smpi_bench_end(); @@ -289,9 +307,9 @@ int PMPI_Iallgatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, 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); @@ -336,9 +354,9 @@ int PMPI_Iscatter(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi 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(); @@ -353,29 +371,30 @@ int PMPI_Scatterv(const void *sendbuf, const int *sendcounts, const int *displs, 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; + CHECK_ARGS(comm == MPI_COMM_NULL, MPI_ERR_COMM, "Iscatterv: the communicator cannot be MPI_COMM_NULL"); + CHECK_ARGS((comm->rank() == root) && (sendcounts == nullptr), MPI_ERR_ARG, + "Iscatterv: param 2 sendcounts cannot be NULL on the root rank"); + CHECK_ARGS((comm->rank() == root) && (displs == nullptr), MPI_ERR_ARG, + "Iscatterv: param 3 displs cannot be NULL on the root rank"); + CHECK_ARGS((comm->rank() == root) && (sendtype == MPI_DATATYPE_NULL), MPI_ERR_TYPE, + "Iscatterv: The sendtype cannot be NULL on the root rank"); + CHECK_ARGS((recvbuf != MPI_IN_PLACE) && (recvtype == MPI_DATATYPE_NULL), MPI_ERR_TYPE, + "Iscatterv: the recvtype cannot be NULL when not receiving in place"); + CHECK_ARGS(request == nullptr, MPI_ERR_ARG, "Iscatterv: param 10 request cannot be NULL"); + CHECK_ARGS(recvbuf != MPI_IN_PLACE && recvcount < 0, MPI_ERR_COUNT, + "Iscatterv: When not receiving in place, the recvcount cannot be negative"); + CHECK_ARGS(root < 0, MPI_ERR_ROOT, "Iscatterv: root cannot be negative"); + CHECK_ARGS(root >= comm->size(), MPI_ERR_ROOT, "Iscatterv: root (=%d) is larger than communicator size (=%d)", root, + comm->size()); if (comm->rank() == root) { 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; - } + for (int i = 0; i < comm->size(); i++) + CHECK_ARGS(sendcounts[i] < 0, MPI_ERR_COUNT, "Iscatterv: sendcounts[%d]=%d but this cannot be negative", i, + sendcounts[i]); } smpi_bench_end(); @@ -397,9 +416,9 @@ int PMPI_Iscatterv(const void* sendbuf, const int* sendcounts, const int* displs 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); @@ -437,9 +456,9 @@ int PMPI_Ireduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype dat datatype->is_replayable() ? count : count * datatype->size(), -1, 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(); @@ -482,13 +501,9 @@ int PMPI_Iallreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype return MPI_ERR_ARG; smpi_bench_end(); - const void* real_sendbuf = sendbuf; std::unique_ptr tmp_sendbuf; - if (sendbuf == MPI_IN_PLACE) { - tmp_sendbuf.reset(new unsigned char[count * datatype->get_extent()]); - simgrid::smpi::Datatype::copy(recvbuf, count, datatype, tmp_sendbuf.get(), count, datatype); - real_sendbuf = tmp_sendbuf.get(); - } + const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, count, datatype); + int rank = simgrid::s4u::this_actor::get_pid(); TRACE_smpi_comm_in(rank, request == MPI_REQUEST_IGNORED ? "PMPI_Allreduce" : "PMPI_Iallreduce", @@ -497,9 +512,9 @@ int PMPI_Iallreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype simgrid::smpi::Datatype::encode(datatype), "")); if (request == MPI_REQUEST_IGNORED) - simgrid::smpi::Colls::allreduce(real_sendbuf, recvbuf, count, datatype, op, comm); + simgrid::smpi::colls::allreduce(real_sendbuf, recvbuf, count, datatype, op, comm); else - simgrid::smpi::Colls::iallreduce(real_sendbuf, recvbuf, count, datatype, op, comm, request); + simgrid::smpi::colls::iallreduce(real_sendbuf, recvbuf, count, datatype, op, comm, request); TRACE_smpi_comm_out(rank); smpi_bench_begin(); @@ -528,12 +543,9 @@ int PMPI_Iscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datat smpi_bench_end(); int rank = simgrid::s4u::this_actor::get_pid(); - const void* real_sendbuf = sendbuf; std::unique_ptr tmp_sendbuf; - if (sendbuf == MPI_IN_PLACE) { - tmp_sendbuf.reset(new unsigned char[count * datatype->size()]); - real_sendbuf = memcpy(tmp_sendbuf.get(), recvbuf, count * datatype->size()); - } + 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", new simgrid::instr::Pt2PtTIData(request == MPI_REQUEST_IGNORED ? "scan" : "iscan", -1, datatype->is_replayable() ? count : count * datatype->size(), @@ -541,9 +553,9 @@ int PMPI_Iscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datat int retval; if (request == MPI_REQUEST_IGNORED) - retval = simgrid::smpi::Colls::scan(real_sendbuf, recvbuf, count, datatype, op, comm); + 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); + retval = simgrid::smpi::colls::iscan(real_sendbuf, recvbuf, count, datatype, op, comm, request); TRACE_smpi_comm_out(rank); smpi_bench_begin(); @@ -571,12 +583,8 @@ int PMPI_Iexscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype dat smpi_bench_end(); int rank = simgrid::s4u::this_actor::get_pid(); - const void* real_sendbuf = sendbuf; std::unique_ptr tmp_sendbuf; - if (sendbuf == MPI_IN_PLACE) { - tmp_sendbuf.reset(new unsigned char[count * datatype->size()]); - real_sendbuf = memcpy(tmp_sendbuf.get(), recvbuf, count * datatype->size()); - } + 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", new simgrid::instr::Pt2PtTIData(request == MPI_REQUEST_IGNORED ? "exscan" : "iexscan", -1, @@ -585,9 +593,9 @@ int PMPI_Iexscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype dat int retval; if (request == MPI_REQUEST_IGNORED) - retval = simgrid::smpi::Colls::exscan(real_sendbuf, recvbuf, count, datatype, op, comm); + 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); + retval = simgrid::smpi::colls::iexscan(real_sendbuf, recvbuf, count, datatype, op, comm, request); TRACE_smpi_comm_out(rank); smpi_bench_begin(); @@ -629,13 +637,8 @@ int PMPI_Ireduce_scatter(const void *sendbuf, void *recvbuf, const int *recvcoun trace_recvcounts->push_back(recvcounts[i] * dt_send_size); totalcount += recvcounts[i]; } - - const void* real_sendbuf = sendbuf; std::unique_ptr tmp_sendbuf; - if (sendbuf == MPI_IN_PLACE) { - tmp_sendbuf.reset(new unsigned char[totalcount * datatype->size()]); - real_sendbuf = memcpy(tmp_sendbuf.get(), recvbuf, totalcount * datatype->size()); - } + 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", new simgrid::instr::VarCollTIData( @@ -643,9 +646,9 @@ int PMPI_Ireduce_scatter(const void *sendbuf, void *recvbuf, const int *recvcoun -1, trace_recvcounts, simgrid::smpi::Datatype::encode(datatype), "")); 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, datatype, op, comm); else - simgrid::smpi::Colls::ireduce_scatter(real_sendbuf, recvbuf, recvcounts, datatype, op, comm, request); + simgrid::smpi::colls::ireduce_scatter(real_sendbuf, recvbuf, recvcounts, datatype, op, comm, request); TRACE_smpi_comm_out(rank); smpi_bench_begin(); @@ -678,13 +681,8 @@ 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 - - const void* real_sendbuf = sendbuf; std::unique_ptr tmp_sendbuf; - if (sendbuf == MPI_IN_PLACE) { - tmp_sendbuf.reset(new unsigned char[recvcount * count * datatype->size()]); - real_sendbuf = memcpy(tmp_sendbuf.get(), recvbuf, recvcount * count * datatype->size()); - } + 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", @@ -695,9 +693,9 @@ int PMPI_Ireduce_scatter_block(const void* sendbuf, void* recvbuf, int recvcount 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, datatype, op, comm); else - simgrid::smpi::Colls::ireduce_scatter(real_sendbuf, recvbuf, recvcounts, datatype, op, comm, request); + simgrid::smpi::colls::ireduce_scatter(real_sendbuf, recvbuf, recvcounts, datatype, op, comm, request); delete[] recvcounts; TRACE_smpi_comm_out(rank); @@ -726,32 +724,29 @@ int PMPI_Ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, vo smpi_bench_end(); int rank = simgrid::s4u::this_actor::get_pid(); - const void* real_sendbuf = sendbuf; - int sendtmpcount = sendcount; - MPI_Datatype sendtmptype = sendtype; + int real_sendcount = sendcount; + MPI_Datatype real_sendtype = sendtype; + std::unique_ptr tmp_sendbuf; + const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, recvcount * comm->size(), recvtype); + if (sendbuf == MPI_IN_PLACE) { - tmp_sendbuf.reset(new unsigned char[recvcount * comm->size() * recvtype->size()]); - // memcpy(??,nullptr,0) is actually undefined behavor, even if harmless. - if (recvbuf != nullptr) - memcpy(tmp_sendbuf.get(), recvbuf, recvcount * comm->size() * recvtype->size()); - real_sendbuf = tmp_sendbuf.get(); - sendtmpcount = recvcount; - sendtmptype = recvtype; + real_sendcount = recvcount; + real_sendtype = recvtype; } 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, - 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(real_sendbuf, sendtmpcount, sendtmptype, recvbuf, recvcount, recvtype, comm); + simgrid::smpi::colls::alltoall(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcount, recvtype, comm); else - retval = simgrid::smpi::Colls::ialltoall(real_sendbuf, sendtmpcount, sendtmptype, recvbuf, recvcount, recvtype, + retval = simgrid::smpi::colls::ialltoall(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcount, recvtype, comm, request); TRACE_smpi_comm_out(rank); @@ -794,10 +789,9 @@ int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* sendd std::vector* trace_recvcounts = new std::vector; int dt_size_recv = recvtype->size(); - const void* real_sendbuf = sendbuf; const int* real_sendcounts = sendcounts; const int* real_senddisps = senddisps; - MPI_Datatype sendtmptype = sendtype; + 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; @@ -809,19 +803,18 @@ int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* sendd std::unique_ptr tmp_sendbuf; std::unique_ptr tmp_sendcounts; std::unique_ptr tmp_senddisps; + const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, maxsize, MPI_CHAR); if (sendbuf == MPI_IN_PLACE) { - tmp_sendbuf.reset(new unsigned char[maxsize]); - real_sendbuf = memcpy(tmp_sendbuf.get(), recvbuf, maxsize); tmp_sendcounts.reset(new int[size]); std::copy(recvcounts, recvcounts + size, tmp_sendcounts.get()); real_sendcounts = tmp_sendcounts.get(); tmp_senddisps.reset(new int[size]); std::copy(recvdisps, recvdisps + size, tmp_senddisps.get()); real_senddisps = tmp_senddisps.get(); - sendtmptype = recvtype; + real_sendtype = recvtype; } - int dt_size_send = sendtmptype->size(); + int dt_size_send = real_sendtype->size(); for (int i = 0; i < size; i++) { // copy data to avoid bad free send_size += real_sendcounts[i] * dt_size_send; @@ -831,15 +824,15 @@ int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* sendd TRACE_smpi_comm_in(rank, 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(real_sendbuf, real_sendcounts, real_senddisps, sendtmptype, recvbuf, + retval = simgrid::smpi::colls::alltoallv(real_sendbuf, real_sendcounts, real_senddisps, real_sendtype, recvbuf, recvcounts, recvdisps, recvtype, comm); else - retval = simgrid::smpi::Colls::ialltoallv(real_sendbuf, real_sendcounts, real_senddisps, sendtmptype, recvbuf, + retval = simgrid::smpi::colls::ialltoallv(real_sendbuf, real_sendcounts, real_senddisps, real_sendtype, recvbuf, recvcounts, recvdisps, recvtype, comm, request); TRACE_smpi_comm_out(rank); @@ -871,7 +864,7 @@ int PMPI_Ialltoallw(const void* sendbuf, const int* sendcounts, const int* sendd smpi_bench_end(); int rank = simgrid::s4u::this_actor::get_pid(); int size = comm->size(); - for (int i = 0; i < size; i++) { // copy data to avoid bad free + for (int i = 0; i < size; i++) { if (recvcounts[i] < 0 || (sendbuf != MPI_IN_PLACE && sendcounts[i] < 0)) return MPI_ERR_COUNT; } @@ -880,7 +873,6 @@ int PMPI_Ialltoallw(const void* sendbuf, const int* sendcounts, const int* sendd std::vector* trace_sendcounts = new std::vector; std::vector* trace_recvcounts = new std::vector; - const void* real_sendbuf = sendbuf; const int* real_sendcounts = sendcounts; const int* real_senddisps = senddisps; const MPI_Datatype* real_sendtypes = sendtypes; @@ -901,9 +893,8 @@ int PMPI_Ialltoallw(const void* sendbuf, const int* sendcounts, const int* sendd std::unique_ptr tmp_sendcounts; std::unique_ptr tmp_senddisps; std::unique_ptr tmp_sendtypes; + const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, maxsize, MPI_CHAR); if (sendbuf == MPI_IN_PLACE) { - tmp_sendbuf.reset(new unsigned char[maxsize]); - real_sendbuf = memcpy(tmp_sendbuf.get(), recvbuf, maxsize); tmp_sendcounts.reset(new int[size]); std::copy(recvcounts, recvcounts + size, tmp_sendcounts.get()); real_sendcounts = tmp_sendcounts.get(); @@ -928,10 +919,10 @@ int PMPI_Ialltoallw(const void* sendbuf, const int* sendcounts, const int* sendd int retval; if (request == MPI_REQUEST_IGNORED) - retval = simgrid::smpi::Colls::alltoallw(real_sendbuf, real_sendcounts, real_senddisps, real_sendtypes, recvbuf, + retval = simgrid::smpi::colls::alltoallw(real_sendbuf, real_sendcounts, real_senddisps, real_sendtypes, recvbuf, recvcounts, recvdisps, recvtypes, comm); else - retval = simgrid::smpi::Colls::ialltoallw(real_sendbuf, real_sendcounts, real_senddisps, real_sendtypes, recvbuf, + retval = simgrid::smpi::colls::ialltoallw(real_sendbuf, real_sendcounts, real_senddisps, real_sendtypes, recvbuf, recvcounts, recvdisps, recvtypes, comm, request); TRACE_smpi_comm_out(rank);