X-Git-Url: http://bilbo.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/0199ba108d66c94df94e4f044994e79efdece4b1..abf3b21ad3f4a19475431076e6e2b26f76d2e7d6:/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 534a94a9eb..7e613a5b63 100644 --- a/src/smpi/bindings/smpi_pmpi_coll.cpp +++ b/src/smpi/bindings/smpi_pmpi_coll.cpp @@ -1,4 +1,4 @@ -/* Copyright (c) 2007-2021. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2007-2023. 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. */ @@ -37,7 +37,7 @@ int PMPI_Ibarrier(MPI_Comm comm, MPI_Request *request) { CHECK_COMM(1) CHECK_REQUEST(2) - + CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Barrier" : "PMPI_Ibarrier") const SmpiBenchGuard suspend_bench; aid_t pid = simgrid::s4u::this_actor::get_pid(); TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Barrier" : "PMPI_Ibarrier", @@ -58,8 +58,7 @@ int PMPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm c return PMPI_Ibcast(buf, count, datatype, root, comm, MPI_REQUEST_IGNORED); } -int PMPI_Ibcast(void *buf, int count, MPI_Datatype datatype, - int root, MPI_Comm comm, MPI_Request* request) +int PMPI_Ibcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm, MPI_Request* request) { SET_BUF1(buf) CHECK_COMM(5) @@ -68,13 +67,19 @@ int PMPI_Ibcast(void *buf, int count, MPI_Datatype datatype, CHECK_BUFFER(1, buf, count, datatype) CHECK_ROOT(4) CHECK_REQUEST(6) + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Bcast" : "PMPI_Ibcast") + " with root " + + std::to_string(root)) const SmpiBenchGuard suspend_bench; aid_t pid = simgrid::s4u::this_actor::get_pid(); TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Bcast" : "PMPI_Ibcast", new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "bcast" : "ibcast", root, -1.0, - datatype->is_replayable() ? count : count * datatype->size(), 0, + count, 0, simgrid::smpi::Datatype::encode(datatype), "")); + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // No barrier in Ibcast + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); + if (comm->size() > 1) { if (request == MPI_REQUEST_IGNORED) simgrid::smpi::colls::bcast(buf, count, datatype, root, comm); @@ -99,7 +104,6 @@ int PMPI_Igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void { CHECK_COMM(8) SET_BUF1(sendbuf) - SET_BUF2(recvbuf) int rank = comm->rank(); if(sendbuf != MPI_IN_PLACE){ CHECK_COUNT(2, sendcount) @@ -107,6 +111,7 @@ int PMPI_Igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void CHECK_BUFFER(1,sendbuf, sendcount, sendtype) } if(rank == root){ + SET_BUF2(recvbuf) CHECK_NOT_IN_PLACE_ROOT(4, recvbuf) CHECK_TYPE(6, recvtype) CHECK_COUNT(5, recvcount) @@ -116,6 +121,8 @@ int PMPI_Igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void } CHECK_ROOT(7) CHECK_REQUEST(9) + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Gather" : "PMPI_Igather") + " with root " + + std::to_string(root)) const void* real_sendbuf = sendbuf; int real_sendcount = sendcount; @@ -132,13 +139,16 @@ int PMPI_Igather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void const SmpiBenchGuard suspend_bench; + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // no barrier in Igather + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); + aid_t pid = simgrid::s4u::this_actor::get_pid(); TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Gather" : "PMPI_Igather", new simgrid::instr::CollTIData( request == MPI_REQUEST_IGNORED ? "gather" : "igather", root, -1.0, - real_sendtype->is_replayable() ? real_sendcount : real_sendcount * real_sendtype->size(), - (comm->rank() != root || recvtype->is_replayable()) ? recvcount : recvcount * recvtype->size(), + real_sendcount, recvcount, simgrid::smpi::Datatype::encode(real_sendtype), simgrid::smpi::Datatype::encode(recvtype))); if (request == MPI_REQUEST_IGNORED) simgrid::smpi::colls::gather(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcount, recvtype, root, comm); @@ -160,7 +170,6 @@ int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi { CHECK_COMM(9) SET_BUF1(sendbuf) - SET_BUF2(recvbuf) int rank = comm->rank(); if(sendbuf != MPI_IN_PLACE){ CHECK_TYPE(3, sendtype) @@ -168,6 +177,7 @@ int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi } CHECK_BUFFER(1, sendbuf, sendcount, sendtype) if(rank == root){ + SET_BUF2(recvbuf) CHECK_NOT_IN_PLACE_ROOT(4, recvbuf) CHECK_TYPE(6, recvtype) CHECK_NULL(5, MPI_ERR_COUNT, recvcounts) @@ -177,6 +187,8 @@ int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi } CHECK_ROOT(8) CHECK_REQUEST(10) + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Gatherv" : "PMPI_Igatherv") + + " with root " + std::to_string(root)) if (rank == root){ for (int i = 0; i < comm->size(); i++) { @@ -186,6 +198,11 @@ int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi } const SmpiBenchGuard suspend_bench; + + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // no barrier in Igatherv + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); + const void* real_sendbuf = sendbuf; int real_sendcount = sendcount; MPI_Datatype real_sendtype = sendtype; @@ -195,19 +212,18 @@ int PMPI_Igatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi } aid_t pid = simgrid::s4u::this_actor::get_pid(); - int dt_size_recv = recvtype->is_replayable() ? 1 : recvtype->size(); 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); - } + if (rank == root) + trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[comm->size()]); + else //this is not significant outside of root, put 0 as we don't know if recvcounts is initialized + trace_recvcounts->insert(trace_recvcounts->end(), comm->size(), 0); TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Gatherv" : "PMPI_Igatherv", new simgrid::instr::VarCollTIData( request == MPI_REQUEST_IGNORED ? "gatherv" : "igatherv", root, - real_sendtype->is_replayable() ? real_sendcount : real_sendcount * real_sendtype->size(), - nullptr, dt_size_recv, trace_recvcounts, simgrid::smpi::Datatype::encode(real_sendtype), + sendcount, + nullptr, -1, trace_recvcounts, simgrid::smpi::Datatype::encode(real_sendtype), simgrid::smpi::Datatype::encode(recvtype))); if (request == MPI_REQUEST_IGNORED) simgrid::smpi::colls::gatherv(real_sendbuf, real_sendcount, real_sendtype, recvbuf, recvcounts, displs, recvtype, @@ -242,6 +258,7 @@ int PMPI_Iallgather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, v CHECK_BUFFER(1, sendbuf, sendcount, sendtype) CHECK_BUFFER(4, recvbuf, recvcount, recvtype) CHECK_REQUEST(8) + CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Allgather" : "PMPI_Iallggather") if (sendbuf == MPI_IN_PLACE) { sendbuf = static_cast(recvbuf) + recvtype->get_extent() * recvcount * comm->rank(); @@ -261,8 +278,7 @@ int PMPI_Iallgather(const void* sendbuf, int sendcount, MPI_Datatype sendtype, v 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(), + sendcount, recvcount, simgrid::smpi::Datatype::encode(sendtype), simgrid::smpi::Datatype::encode(recvtype))); if (request == MPI_REQUEST_IGNORED) simgrid::smpi::colls::allgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm); @@ -300,26 +316,25 @@ int PMPI_Iallgatherv(const void* sendbuf, int sendcount, MPI_Datatype sendtype, CHECK_COUNT(5, recvcounts[i]) CHECK_BUFFER(4, recvbuf, recvcounts[i], recvtype) } + CHECK_COLLECTIVE(comm, MPI_REQUEST_IGNORED ? "PMPI_Allgatherv" : "PMPI_Iallgatherv") const SmpiBenchGuard suspend_bench; + if (sendbuf == MPI_IN_PLACE) { sendbuf = static_cast(recvbuf) + recvtype->get_extent() * displs[comm->rank()]; sendcount = recvcounts[comm->rank()]; sendtype = recvtype; } aid_t pid = simgrid::s4u::this_actor::get_pid(); - int dt_size_recv = recvtype->is_replayable() ? 1 : recvtype->size(); 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_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[comm->size()]); TRACE_smpi_comm_in( pid, request == MPI_REQUEST_IGNORED ? "PMPI_Allgatherv" : "PMPI_Iallgatherv", new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "allgatherv" : "iallgatherv", -1, - sendtype->is_replayable() ? sendcount : sendcount * sendtype->size(), nullptr, - dt_size_recv, trace_recvcounts, simgrid::smpi::Datatype::encode(sendtype), + sendcount, nullptr, + -1, trace_recvcounts, simgrid::smpi::Datatype::encode(sendtype), simgrid::smpi::Datatype::encode(recvtype))); if (request == MPI_REQUEST_IGNORED) simgrid::smpi::colls::allgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, comm); @@ -340,10 +355,10 @@ int PMPI_Iscatter(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi MPI_Datatype recvtype, int root, MPI_Comm comm, MPI_Request* request) { CHECK_COMM(8) - SET_BUF1(sendbuf) SET_BUF2(recvbuf) int rank = comm->rank(); if(rank == root){ + SET_BUF1(sendbuf) CHECK_NOT_IN_PLACE_ROOT(1, sendbuf) CHECK_COUNT(2, sendcount) CHECK_TYPE(3, sendtype) @@ -358,6 +373,8 @@ int PMPI_Iscatter(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi } CHECK_ROOT(8) CHECK_REQUEST(9) + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Scatter" : "PMPI_Iscatter") + + " with root " + std::to_string(root)) if (recvbuf == MPI_IN_PLACE) { recvtype = sendtype; @@ -371,13 +388,16 @@ int PMPI_Iscatter(const void* sendbuf, int sendcount, MPI_Datatype sendtype, voi const SmpiBenchGuard suspend_bench; + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // no barrier in Iscatter + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); + 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, - (rank != root || sendtype->is_replayable()) ? sendcount : sendcount * sendtype->size(), - recvtype->is_replayable() ? recvcount : recvcount * recvtype->size(), + sendcount, recvcount, simgrid::smpi::Datatype::encode(sendtype), simgrid::smpi::Datatype::encode(recvtype))); if (request == MPI_REQUEST_IGNORED) simgrid::smpi::colls::scatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm); @@ -396,12 +416,10 @@ 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) { - 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) @@ -409,6 +427,8 @@ int PMPI_Iscatterv(const void* sendbuf, const int* sendcounts, const int* displs CHECK_ROOT(9) CHECK_REQUEST(10) if (rank == root) { + SET_BUF1(sendbuf) + CHECK_NOT_IN_PLACE_ROOT(1, sendbuf) CHECK_NULL(2, MPI_ERR_COUNT, sendcounts) CHECK_NULL(3, MPI_ERR_ARG, displs) CHECK_TYPE(4, sendtype) @@ -423,23 +443,28 @@ int PMPI_Iscatterv(const void* sendbuf, const int* sendcounts, const int* displs } else { CHECK_NOT_IN_PLACE_ROOT(4, recvbuf) } + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Scatterv" : "PMPI_Iscatterv") + + " with root " + std::to_string(root)) const SmpiBenchGuard suspend_bench; + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // no barrier in Iscatterv + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); + aid_t pid = simgrid::s4u::this_actor::get_pid(); - int dt_size_send = sendtype->is_replayable() ? 1 : sendtype->size(); 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); - } - } + if (rank == root) + trace_sendcounts->insert(trace_sendcounts->end(), &sendcounts[0], &sendcounts[comm->size()]); + else //this is not significant outside of root, put 0 as we don't know if sendcounts is initialized + trace_sendcounts->insert(trace_sendcounts->end(), comm->size(), 0); + TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Scatterv" : "PMPI_Iscatterv", new simgrid::instr::VarCollTIData( - request == MPI_REQUEST_IGNORED ? "scatterv" : "iscatterv", root, dt_size_send, - trace_sendcounts, recvtype->is_replayable() ? recvcount : recvcount * recvtype->size(), + request == MPI_REQUEST_IGNORED ? "scatterv" : "iscatterv", root, -1, + trace_sendcounts, recvcount, nullptr, simgrid::smpi::Datatype::encode(sendtype), simgrid::smpi::Datatype::encode(recvtype))); if (request == MPI_REQUEST_IGNORED) @@ -461,25 +486,32 @@ int PMPI_Ireduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype dat { 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){ + SET_BUF2(recvbuf) CHECK_NOT_IN_PLACE(2, recvbuf) CHECK_BUFFER(5, recvbuf, count, datatype) } CHECK_OP(5, op, datatype) CHECK_ROOT(7) CHECK_REQUEST(8) + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Reduce" : "PMPI_Ireduce") + " with op " + + op->name() + " and root " + std::to_string(root)) const SmpiBenchGuard suspend_bench; + + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // no barrier in Ireduce + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); + aid_t pid = simgrid::s4u::this_actor::get_pid(); TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce" : "PMPI_Ireduce", new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "reduce" : "ireduce", root, 0, - datatype->is_replayable() ? count : count * datatype->size(), 0, + count, 0, simgrid::smpi::Datatype::encode(datatype), "")); if (request == MPI_REQUEST_IGNORED) simgrid::smpi::colls::reduce(sendbuf, recvbuf, count, datatype, op, root, comm); @@ -523,8 +555,11 @@ int PMPI_Iallreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype CHECK_BUFFER(1, sendbuf, count, datatype) CHECK_BUFFER(2, recvbuf, count, datatype) CHECK_REQUEST(7) + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Alleduce" : "PMPI_Iallreduce") + + " with op " + op->name()) const SmpiBenchGuard suspend_bench; + std::vector tmp_sendbuf; const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, count, datatype); @@ -532,7 +567,7 @@ int PMPI_Iallreduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype 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(), 0, + count, 0, simgrid::smpi::Datatype::encode(datatype), "")); if (request == MPI_REQUEST_IGNORED) @@ -560,16 +595,22 @@ int PMPI_Iscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datat CHECK_BUFFER(2,recvbuf,count, datatype) CHECK_REQUEST(7) CHECK_OP(5, op, datatype) + CHECK_COLLECTIVE(comm, + std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Scan" : "PMPI_Iscan") + " with op " + op->name()) const SmpiBenchGuard suspend_bench; + + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // no barrier in Iscan + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); + aid_t pid = simgrid::s4u::this_actor::get_pid(); std::vector tmp_sendbuf; const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, count, datatype); TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Scan" : "PMPI_Iscan", - new simgrid::instr::Pt2PtTIData(request == MPI_REQUEST_IGNORED ? "scan" : "iscan", -1, - datatype->is_replayable() ? count : count * datatype->size(), - simgrid::smpi::Datatype::encode(datatype))); + new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "scan" : "iscan", -1, 0.0, + count, 0, simgrid::smpi::Datatype::encode(datatype), "")); int retval; if (request == MPI_REQUEST_IGNORED) @@ -596,16 +637,22 @@ int PMPI_Iexscan(const void *sendbuf, void *recvbuf, int count, MPI_Datatype dat CHECK_BUFFER(2, recvbuf, count, datatype) CHECK_REQUEST(7) CHECK_OP(5, op, datatype) + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Exscan" : "PMPI_Iexscan") + " with op " + + op->name()) const SmpiBenchGuard suspend_bench; + + if (simgrid::config::get_value("smpi/barrier-collectives") && + request == MPI_REQUEST_IGNORED) // no barrier in Iexscan + smpi_deployment_startup_barrier(smpi_process()->get_instance_id()); + aid_t pid = simgrid::s4u::this_actor::get_pid(); std::vector tmp_sendbuf; const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, count, datatype); TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Exscan" : "PMPI_Iexscan", - new simgrid::instr::Pt2PtTIData(request == MPI_REQUEST_IGNORED ? "exscan" : "iexscan", -1, - datatype->is_replayable() ? count : count * datatype->size(), - simgrid::smpi::Datatype::encode(datatype))); + new simgrid::instr::CollTIData(request == MPI_REQUEST_IGNORED ? "exscan" : "iexscan", -1, 0.0, + count, 0, simgrid::smpi::Datatype::encode(datatype), "")); int retval; if (request == MPI_REQUEST_IGNORED) @@ -638,15 +685,18 @@ int PMPI_Ireduce_scatter(const void *sendbuf, void *recvbuf, const int *recvcoun CHECK_BUFFER(1, sendbuf, recvcounts[i], datatype) CHECK_BUFFER(2, recvbuf, recvcounts[i], datatype) } + CHECK_COLLECTIVE(comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter" : "PMPI_Ireduce_scatter") + + " with op " + op->name()) 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(); + trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[comm->size()]); + int totalcount = 0; for (int i = 0; i < comm->size(); i++) { // copy data to avoid bad free - trace_recvcounts->push_back(recvcounts[i] * dt_send_size); totalcount += recvcounts[i]; } std::vector tmp_sendbuf; @@ -654,8 +704,8 @@ int PMPI_Ireduce_scatter(const void *sendbuf, void *recvbuf, const int *recvcoun 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, - 0, trace_recvcounts, simgrid::smpi::Datatype::encode(datatype), "")); + request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, -1, nullptr, + -1 , trace_recvcounts, std::to_string(0), simgrid::smpi::Datatype::encode(datatype))); if (request == MPI_REQUEST_IGNORED) simgrid::smpi::colls::reduce_scatter(real_sendbuf, recvbuf, recvcounts, datatype, op, comm); @@ -684,20 +734,24 @@ int PMPI_Ireduce_scatter_block(const void* sendbuf, void* recvbuf, int recvcount CHECK_BUFFER(2, recvbuf, recvcount, datatype) CHECK_REQUEST(7) CHECK_OP(5, op, datatype) + CHECK_COLLECTIVE( + comm, std::string(request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter_block" : "PMPI_Ireduce_scatter_block") + + " with op " + op->name()) const SmpiBenchGuard suspend_bench; + int count = comm->size(); aid_t pid = simgrid::s4u::this_actor::get_pid(); - int dt_send_size = datatype->is_replayable() ? 1 : datatype->size(); - auto trace_recvcounts = std::make_shared>(recvcount * dt_send_size); // copy data to avoid bad free + auto trace_recvcounts = std::make_shared>(recvcount); + std::vector tmp_sendbuf; const void* real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, recvcount * count, datatype); TRACE_smpi_comm_in( pid, request == MPI_REQUEST_IGNORED ? "PMPI_Reduce_scatter_block" : "PMPI_Ireduce_scatter_block", - new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, 0, - nullptr, 0, trace_recvcounts, simgrid::smpi::Datatype::encode(datatype), "")); + new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "reducescatter" : "ireducescatter", -1, -1, + nullptr, -1, trace_recvcounts, simgrid::smpi::Datatype::encode(datatype), "")); std::vector recvcounts(count); for (int i = 0; i < count; i++) @@ -732,6 +786,7 @@ int PMPI_Ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, vo CHECK_COUNT(5, recvcount) CHECK_BUFFER(4, recvbuf, recvcount, recvtype) CHECK_REQUEST(8) + CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoall" : "PMPI_Ialltoall") aid_t pid = simgrid::s4u::this_actor::get_pid(); int real_sendcount = sendcount; @@ -755,8 +810,7 @@ int PMPI_Ialltoall(const void* sendbuf, int sendcount, MPI_Datatype sendtype, vo TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoall" : "PMPI_Ialltoall", new simgrid::instr::CollTIData( request == MPI_REQUEST_IGNORED ? "alltoall" : "ialltoall", -1, -1.0, - real_sendtype->is_replayable() ? real_sendcount : real_sendcount * real_sendtype->size(), - recvtype->is_replayable() ? recvcount : recvcount * recvtype->size(), + real_sendcount, recvcount, simgrid::smpi::Datatype::encode(real_sendtype), simgrid::smpi::Datatype::encode(recvtype))); int retval; if (request == MPI_REQUEST_IGNORED) @@ -791,6 +845,7 @@ int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* sendd CHECK_NULL(6, MPI_ERR_COUNT, recvcounts) CHECK_NULL(7, MPI_ERR_ARG, recvdispls) CHECK_REQUEST(10) + CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallv" : "PMPI_Ialltoallv") aid_t pid = simgrid::s4u::this_actor::get_pid(); int size = comm->size(); @@ -804,27 +859,24 @@ int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* sendd } const SmpiBenchGuard suspend_bench; + int send_size = 0; int recv_size = 0; auto trace_sendcounts = std::make_shared>(); auto trace_recvcounts = std::make_shared>(); + trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[size]); + int dt_size_recv = recvtype->size(); const int* real_sendcounts = sendcounts; const int* real_senddispls = senddispls; MPI_Datatype real_sendtype = sendtype; int maxsize = 0; - 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 (((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); + const void* real_sendbuf; + if (sendbuf == MPI_IN_PLACE) { tmp_sendcounts.assign(recvcounts, recvcounts + size); real_sendcounts = tmp_sendcounts.data(); @@ -833,17 +885,20 @@ int PMPI_Ialltoallv(const void* sendbuf, const int* sendcounts, const int* sendd real_sendtype = recvtype; } + for (int i = 0; i < size; i++) { // copy data to avoid bad free + send_size += real_sendcounts[i] ; + recv_size += recvcounts[i]; + if (((recvdispls[i] + recvcounts[i]) * dt_size_recv) > maxsize) + maxsize = (recvdispls[i] + recvcounts[i]) * dt_size_recv; + } + real_sendbuf = smpi_get_in_place_buf(sendbuf, recvbuf, tmp_sendbuf, maxsize, MPI_CHAR); + if(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 += real_sendcounts[i] * dt_size_send; - trace_sendcounts->push_back(real_sendcounts[i] * dt_size_send); - } + trace_sendcounts->insert(trace_sendcounts->end(), &real_sendcounts[0], &real_sendcounts[size]); TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallv" : "PMPI_Ialltoallv", new simgrid::instr::VarCollTIData(request == MPI_REQUEST_IGNORED ? "alltoallv" : "ialltoallv", -1, @@ -896,6 +951,7 @@ int PMPI_Ialltoallw(const void* sendbuf, const int* sendcounts, const int* sendd CHECK_TYPE(8, recvtypes[i]) CHECK_BUFFER(5, recvbuf, recvcounts[i], recvtypes[i]) } + CHECK_COLLECTIVE(comm, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallw" : "PMPI_Ialltoallw") const SmpiBenchGuard suspend_bench; @@ -903,16 +959,17 @@ int PMPI_Ialltoallw(const void* sendbuf, const int* sendcounts, const int* sendd int recv_size = 0; auto trace_sendcounts = std::make_shared>(); auto trace_recvcounts = std::make_shared>(); + trace_recvcounts->insert(trace_recvcounts->end(), &recvcounts[0], &recvcounts[size]); const int* real_sendcounts = sendcounts; const int* real_senddispls = senddispls; const MPI_Datatype* real_sendtypes = sendtypes; + unsigned long maxsize = 0; for (int i = 0; i < size; i++) { // copy data to avoid bad free if (recvtypes[i] == MPI_DATATYPE_NULL) return MPI_ERR_TYPE; recv_size += recvcounts[i] * recvtypes[i]->size(); - trace_recvcounts->push_back(recvcounts[i] * recvtypes[i]->size()); if ((recvdispls[i] + (recvcounts[i] * recvtypes[i]->size())) > maxsize) maxsize = recvdispls[i] + (recvcounts[i] * recvtypes[i]->size()); } @@ -937,9 +994,9 @@ int PMPI_Ialltoallw(const void* sendbuf, const int* sendcounts, const int* sendd return MPI_ERR_TRUNCATE; } + trace_sendcounts->insert(trace_sendcounts->end(), &real_sendcounts[0], &real_sendcounts[size]); for (int i = 0; i < size; i++) { // copy data to avoid bad free send_size += real_sendcounts[i] * real_sendtypes[i]->size(); - trace_sendcounts->push_back(real_sendcounts[i] * real_sendtypes[i]->size()); } TRACE_smpi_comm_in(pid, request == MPI_REQUEST_IGNORED ? "PMPI_Alltoallw" : "PMPI_Ialltoallw",