X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/3ce5ddabb09474b5ec87d059c65ccd266b05282a..abdd50d7729bd7ed16782a24d97fe92ff28fb57a:/src/smpi/internals/smpi_replay.cpp diff --git a/src/smpi/internals/smpi_replay.cpp b/src/smpi/internals/smpi_replay.cpp index a58bfffed9..b12e1e122d 100644 --- a/src/smpi/internals/smpi_replay.cpp +++ b/src/smpi/internals/smpi_replay.cpp @@ -12,6 +12,7 @@ #include "smpi_request.hpp" #include "xbt/replay.hpp" +#include #include #include #include @@ -21,26 +22,27 @@ using simgrid::s4u::Actor; XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_replay,smpi,"Trace Replay with SMPI"); -static int communicator_size = 0; static int active_processes = 0; static std::unordered_map*> reqq; static MPI_Datatype MPI_DEFAULT_TYPE; -static int sendbuffer_size = 0; -static char* sendbuffer = nullptr; -static int recvbuffer_size = 0; -static char* recvbuffer = nullptr; - -class ReplayActionArg { - ReplayActionArg() {} -}; +#define CHECK_ACTION_PARAMS(action, mandatory, optional) \ + { \ + if (action.size() < static_cast(mandatory + 2)) \ + THROWF(arg_error, 0, "%s replay failed.\n" \ + "%lu items were given on the line. First two should be process_id and action. " \ + "This action needs after them %lu mandatory arguments, and accepts %lu optional ones. \n" \ + "Please contact the Simgrid team if support is needed", \ + __FUNCTION__, action.size(), static_cast(mandatory), \ + static_cast(optional)); \ + } -static void log_timed_action (const char *const *action, double clock){ +static void log_timed_action(simgrid::xbt::ReplayAction& action, double clock) +{ if (XBT_LOG_ISENABLED(smpi_replay, xbt_log_priority_verbose)){ - char *name = xbt_str_join_array(action, " "); - XBT_VERB("%s %f", name, smpi_process()->simulated_elapsed()-clock); - xbt_free(name); + std::string s = boost::algorithm::join(action, " "); + XBT_VERB("%s %f", s.c_str(), smpi_process()->simulated_elapsed() - clock); } } @@ -54,383 +56,281 @@ static void set_reqq_self(std::vector *mpi_request) reqq.insert({Actor::self()->getPid(), mpi_request}); } -//allocate a single buffer for all sends, growing it if needed -void* smpi_get_tmp_sendbuffer(int size) -{ - if (not smpi_process()->replaying()) - return xbt_malloc(size); - if (sendbuffer_size(xbt_realloc(sendbuffer,size)); - sendbuffer_size=size; - } - return sendbuffer; -} - -//allocate a single buffer for all recv -void* smpi_get_tmp_recvbuffer(int size){ - if (not smpi_process()->replaying()) - return xbt_malloc(size); - if (recvbuffer_size(xbt_realloc(recvbuffer,size)); - recvbuffer_size=size; - } - return recvbuffer; -} - -void smpi_free_tmp_buffer(void* buf){ - if (not smpi_process()->replaying()) - xbt_free(buf); -} - /* Helper function */ -static double parse_double(const char *string) -{ - char *endptr; - double value = strtod(string, &endptr); - if (*endptr != '\0') - THROWF(unknown_error, 0, "%s is not a double", string); - return value; -} - - -//TODO: this logic should be moved inside the datatype class, to support all predefined types and get rid of is_replayable. -static MPI_Datatype decode_datatype(const char *const action) +static double parse_double(std::string string) { - switch(atoi(action)) { - case 0: - return MPI_DOUBLE; - break; - case 1: - return MPI_INT; - break; - case 2: - return MPI_CHAR; - break; - case 3: - return MPI_SHORT; - break; - case 4: - return MPI_LONG; - break; - case 5: - return MPI_FLOAT; - break; - case 6: - return MPI_BYTE; - break; - default: - return MPI_DEFAULT_TYPE; - break; - } + return xbt_str_parse_double(string.c_str(), "%s is not a double"); } -const char* encode_datatype(MPI_Datatype datatype) -{ - if (datatype==MPI_BYTE) - return ""; - if(datatype==MPI_DOUBLE) - return "0"; - if(datatype==MPI_INT) - return "1"; - if(datatype==MPI_CHAR) - return "2"; - if(datatype==MPI_SHORT) - return "3"; - if(datatype==MPI_LONG) - return "4"; - if(datatype==MPI_FLOAT) - return "5"; - // default - not implemented. - // do not warn here as we pass in this function even for other trace formats - return "-1"; -} - -#define CHECK_ACTION_PARAMS(action, mandatory, optional) {\ - int i=0;\ - while(action[i]!=nullptr)\ - i++;\ - if(isimulated_start(); - /*initialize the number of active processes */ - active_processes = smpi_process_count(); - - set_reqq_self(new std::vector); -} - -static void action_finalize(const char *const *action) -{ - /* Nothing to do */ -} - -static void action_comm_size(const char *const *action) -{ - communicator_size = parse_double(action[2]); - log_timed_action (action, smpi_process()->simulated_elapsed()); -} - -static void action_comm_split(const char *const *action) -{ - log_timed_action (action, smpi_process()->simulated_elapsed()); -} - -static void action_comm_dup(const char *const *action) -{ - log_timed_action (action, smpi_process()->simulated_elapsed()); -} - -static void action_compute(const char *const *action) -{ - CHECK_ACTION_PARAMS(action, 1, 0) - double clock = smpi_process()->simulated_elapsed(); - double flops= parse_double(action[2]); - int my_proc_id = Actor::self()->getPid(); - - TRACE_smpi_computing_in(my_proc_id, flops); - smpi_execute_flops(flops); - TRACE_smpi_computing_out(my_proc_id); - - log_timed_action (action, clock); -} - -static void action_send(const char *const *action) -{ - CHECK_ACTION_PARAMS(action, 2, 1) - int to = atoi(action[2]); - double size=parse_double(action[3]); - double clock = smpi_process()->simulated_elapsed(); - - MPI_Datatype MPI_CURRENT_TYPE = (action[4]) ? decode_datatype(action[4]) : MPI_DEFAULT_TYPE; - - int my_proc_id = Actor::self()->getPid(); - int dst_traced = MPI_COMM_WORLD->group()->actor(to)->getPid(); - - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("send", to, size, encode_datatype(MPI_CURRENT_TYPE))); - if (not TRACE_smpi_view_internals()) - TRACE_smpi_send(my_proc_id, my_proc_id, dst_traced, 0, size * MPI_CURRENT_TYPE->size()); - - Request::send(nullptr, size, MPI_CURRENT_TYPE, to , 0, MPI_COMM_WORLD); - - TRACE_smpi_comm_out(my_proc_id); - - log_timed_action(action, clock); -} - -static void action_Isend(const char *const *action) -{ - CHECK_ACTION_PARAMS(action, 2, 1) - int to = atoi(action[2]); - double size=parse_double(action[3]); - double clock = smpi_process()->simulated_elapsed(); - - MPI_Datatype MPI_CURRENT_TYPE = (action[4]) ? decode_datatype(action[4]) : MPI_DEFAULT_TYPE; - - int my_proc_id = Actor::self()->getPid(); - int dst_traced = MPI_COMM_WORLD->group()->actor(to)->getPid(); - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("Isend", to, size, encode_datatype(MPI_CURRENT_TYPE))); - if (not TRACE_smpi_view_internals()) - TRACE_smpi_send(my_proc_id, my_proc_id, dst_traced, 0, size * MPI_CURRENT_TYPE->size()); - - MPI_Request request = Request::isend(nullptr, size, MPI_CURRENT_TYPE, to, 0, MPI_COMM_WORLD); - - TRACE_smpi_comm_out(my_proc_id); - - get_reqq_self()->push_back(request); - - log_timed_action (action, clock); -} - -static void action_recv(const char *const *action) { - CHECK_ACTION_PARAMS(action, 2, 1) - int from = atoi(action[2]); - double size=parse_double(action[3]); - double clock = smpi_process()->simulated_elapsed(); - MPI_Status status; - - MPI_Datatype MPI_CURRENT_TYPE = (action[4]) ? decode_datatype(action[4]) : MPI_DEFAULT_TYPE; - - int my_proc_id = Actor::self()->getPid(); - int src_traced = MPI_COMM_WORLD->group()->actor(from)->getPid(); - - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("recv", from, size, encode_datatype(MPI_CURRENT_TYPE))); +namespace Replay { +class ActionArgParser { +public: + virtual void parse(simgrid::xbt::ReplayAction& action) { CHECK_ACTION_PARAMS(action, 0, 0) } +}; - //unknown size from the receiver point of view - if (size <= 0.0) { - Request::probe(from, 0, MPI_COMM_WORLD, &status); - size=status.count; +class SendRecvParser : public ActionArgParser { +public: + /* communication partner; if we send, this is the receiver and vice versa */ + int partner; + double size; + MPI_Datatype datatype1 = MPI_DEFAULT_TYPE; + + void parse(simgrid::xbt::ReplayAction& action) override + { + CHECK_ACTION_PARAMS(action, 2, 1) + partner = std::stoi(action[2]); + size = parse_double(action[3]); + if (action.size() > 4) + datatype1 = simgrid::smpi::Datatype::decode(action[4]); } +}; - Request::recv(nullptr, size, MPI_CURRENT_TYPE, from, 0, MPI_COMM_WORLD, &status); +class ComputeParser : public ActionArgParser { +public: + /* communication partner; if we send, this is the receiver and vice versa */ + double flops; - TRACE_smpi_comm_out(my_proc_id); - if (not TRACE_smpi_view_internals()) { - TRACE_smpi_recv(src_traced, my_proc_id, 0); + void parse(simgrid::xbt::ReplayAction& action) override + { + CHECK_ACTION_PARAMS(action, 1, 0) + flops = parse_double(action[2]); } +}; - log_timed_action (action, clock); -} +template class ReplayAction { +protected: + const std::string name; + T args; -static void action_Irecv(const char *const *action) -{ - CHECK_ACTION_PARAMS(action, 2, 1) - int from = atoi(action[2]); - double size=parse_double(action[3]); - double clock = smpi_process()->simulated_elapsed(); + int my_proc_id; - MPI_Datatype MPI_CURRENT_TYPE = (action[4]) ? decode_datatype(action[4]) : MPI_DEFAULT_TYPE; +public: + explicit ReplayAction(std::string name) : name(name), my_proc_id(simgrid::s4u::Actor::self()->getPid()) {} - int my_proc_id = Actor::self()->getPid(); - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("Irecv", from, size, encode_datatype(MPI_CURRENT_TYPE))); - MPI_Status status; - //unknow size from the receiver pov - if (size <= 0.0) { - Request::probe(from, 0, MPI_COMM_WORLD, &status); - size = status.count; + virtual void execute(simgrid::xbt::ReplayAction& action) + { + // Needs to be re-initialized for every action, hence here + double start_time = smpi_process()->simulated_elapsed(); + args.parse(action); + kernel(action); + if (name != "Init") + log_timed_action(action, start_time); } - MPI_Request request = Request::irecv(nullptr, size, MPI_CURRENT_TYPE, from, 0, MPI_COMM_WORLD); + virtual void kernel(simgrid::xbt::ReplayAction& action) = 0; +}; - TRACE_smpi_comm_out(my_proc_id); - get_reqq_self()->push_back(request); +class WaitAction : public ReplayAction { +public: + WaitAction() : ReplayAction("Wait") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + std::string s = boost::algorithm::join(action, " "); + xbt_assert(get_reqq_self()->size(), "action wait not preceded by any irecv or isend: %s", s.c_str()); + MPI_Request request = get_reqq_self()->back(); + get_reqq_self()->pop_back(); + + if (request == nullptr) { + /* Assume that the trace is well formed, meaning the comm might have been caught by a MPI_test. Then just + * return.*/ + return; + } - log_timed_action (action, clock); -} + int rank = request->comm() != MPI_COMM_NULL ? request->comm()->rank() : -1; -static void action_test(const char* const* action) -{ - CHECK_ACTION_PARAMS(action, 0, 0) - double clock = smpi_process()->simulated_elapsed(); - MPI_Status status; + // Must be taken before Request::wait() since the request may be set to + // MPI_REQUEST_NULL by Request::wait! + int src = request->comm()->group()->rank(request->src()); + int dst = request->comm()->group()->rank(request->dst()); + bool is_wait_for_receive = (request->flags() & RECV); + // TODO: Here we take the rank while we normally take the process id (look for my_proc_id) + TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::NoOpTIData("wait")); - MPI_Request request = get_reqq_self()->back(); - get_reqq_self()->pop_back(); - //if request is null here, this may mean that a previous test has succeeded - //Different times in traced application and replayed version may lead to this - //In this case, ignore the extra calls. - if(request!=nullptr){ - int my_proc_id = Actor::self()->getPid(); - TRACE_smpi_testing_in(my_proc_id); + MPI_Status status; + Request::wait(&request, &status); - int flag = Request::test(&request, &status); + TRACE_smpi_comm_out(rank); + if (is_wait_for_receive) + TRACE_smpi_recv(src, dst, 0); + } +}; - XBT_DEBUG("MPI_Test result: %d", flag); - /* push back request in vector to be caught by a subsequent wait. if the test did succeed, the request is now nullptr.*/ - get_reqq_self()->push_back(request); +class SendAction : public ReplayAction { +public: + SendAction() = delete; + SendAction(std::string name) : ReplayAction(name) {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + int dst_traced = MPI_COMM_WORLD->group()->actor(args.partner)->getPid(); + + TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::Pt2PtTIData(name, args.partner, args.size, + Datatype::encode(args.datatype1))); + if (not TRACE_smpi_view_internals()) + TRACE_smpi_send(my_proc_id, my_proc_id, dst_traced, 0, args.size * args.datatype1->size()); + + if (name == "send") { + Request::send(nullptr, args.size, args.datatype1, args.partner, 0, MPI_COMM_WORLD); + } else if (name == "Isend") { + MPI_Request request = Request::isend(nullptr, args.size, args.datatype1, args.partner, 0, MPI_COMM_WORLD); + get_reqq_self()->push_back(request); + } else { + xbt_die("Don't know this action, %s", name.c_str()); + } - TRACE_smpi_testing_out(my_proc_id); + TRACE_smpi_comm_out(my_proc_id); } - log_timed_action (action, clock); -} +}; -static void action_wait(const char *const *action){ - CHECK_ACTION_PARAMS(action, 0, 0) - double clock = smpi_process()->simulated_elapsed(); - MPI_Status status; +class RecvAction : public ReplayAction { +public: + RecvAction() = delete; + explicit RecvAction(std::string name) : ReplayAction(name) {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + int src_traced = MPI_COMM_WORLD->group()->actor(args.partner)->getPid(); + + TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::Pt2PtTIData(name, args.partner, args.size, + Datatype::encode(args.datatype1))); + + MPI_Status status; + // unknown size from the receiver point of view + if (args.size <= 0.0) { + Request::probe(args.partner, 0, MPI_COMM_WORLD, &status); + args.size = status.count; + } - xbt_assert(get_reqq_self()->size(), "action wait not preceded by any irecv or isend: %s", - xbt_str_join_array(action," ")); - MPI_Request request = get_reqq_self()->back(); - get_reqq_self()->pop_back(); + if (name == "recv") { + Request::recv(nullptr, args.size, args.datatype1, args.partner, 0, MPI_COMM_WORLD, &status); + } else if (name == "Irecv") { + MPI_Request request = Request::irecv(nullptr, args.size, args.datatype1, args.partner, 0, MPI_COMM_WORLD); + get_reqq_self()->push_back(request); + } - if (request==nullptr){ - /* Assume that the trace is well formed, meaning the comm might have been caught by a MPI_test. Then just return.*/ - return; + TRACE_smpi_comm_out(my_proc_id); + // TODO: Check why this was only activated in the "recv" case and not in the "Irecv" case + if (name == "recv" && not TRACE_smpi_view_internals()) { + TRACE_smpi_recv(src_traced, my_proc_id, 0); + } } +}; - int rank = request->comm() != MPI_COMM_NULL ? request->comm()->rank() : -1; - - MPI_Group group = request->comm()->group(); - int src_traced = group->rank(request->src()); - int dst_traced = group->rank(request->dst()); - int is_wait_for_receive = (request->flags() & RECV); - TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::NoOpTIData("wait")); - - Request::wait(&request, &status); +class ComputeAction : public ReplayAction { +public: + ComputeAction() : ReplayAction("compute") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + TRACE_smpi_computing_in(my_proc_id, args.flops); + smpi_execute_flops(args.flops); + TRACE_smpi_computing_out(my_proc_id); + } +}; - TRACE_smpi_comm_out(rank); - if (is_wait_for_receive) - TRACE_smpi_recv(src_traced, dst_traced, 0); - log_timed_action (action, clock); -} +class TestAction : public ReplayAction { +public: + TestAction() : ReplayAction("Test") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + MPI_Request request = get_reqq_self()->back(); + get_reqq_self()->pop_back(); + // if request is null here, this may mean that a previous test has succeeded + // Different times in traced application and replayed version may lead to this + // In this case, ignore the extra calls. + if (request != nullptr) { + TRACE_smpi_testing_in(my_proc_id); + + MPI_Status status; + int flag = Request::test(&request, &status); + + XBT_DEBUG("MPI_Test result: %d", flag); + /* push back request in vector to be caught by a subsequent wait. if the test did succeed, the request is now + * nullptr.*/ + get_reqq_self()->push_back(request); + + TRACE_smpi_testing_out(my_proc_id); + } + } +}; -static void action_waitall(const char *const *action){ - CHECK_ACTION_PARAMS(action, 0, 0) - double clock = smpi_process()->simulated_elapsed(); - const unsigned int count_requests = get_reqq_self()->size(); +class InitAction : public ReplayAction { +public: + InitAction() : ReplayAction("Init") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + CHECK_ACTION_PARAMS(action, 0, 1) + MPI_DEFAULT_TYPE = (action.size() > 2) ? MPI_DOUBLE // default MPE datatype + : MPI_BYTE; // default TAU datatype + + /* start a simulated timer */ + smpi_process()->simulated_start(); + /*initialize the number of active processes */ + active_processes = smpi_process_count(); + + set_reqq_self(new std::vector); + } +}; - if (count_requests>0) { - MPI_Status status[count_requests]; +class CommunicatorAction : public ReplayAction { +public: + CommunicatorAction() : ReplayAction("Comm") {} + void kernel(simgrid::xbt::ReplayAction& action) override { /* nothing to do */} +}; - int my_proc_id_traced = Actor::self()->getPid(); - TRACE_smpi_comm_in(my_proc_id_traced, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("waitAll", -1, count_requests, "")); - int recvs_snd[count_requests]; - int recvs_rcv[count_requests]; - for (unsigned int i = 0; i < count_requests; i++) { - const auto& req = (*get_reqq_self())[i]; - if (req && (req->flags() & RECV)) { - recvs_snd[i] = req->src(); - recvs_rcv[i] = req->dst(); - } else - recvs_snd[i] = -100; - } - Request::waitall(count_requests, &(*get_reqq_self())[0], status); - - for (unsigned i = 0; i < count_requests; i++) { - if (recvs_snd[i]!=-100) - TRACE_smpi_recv(recvs_snd[i], recvs_rcv[i],0); - } - TRACE_smpi_comm_out(my_proc_id_traced); +class WaitAllAction : public ReplayAction { +public: + WaitAllAction() : ReplayAction("waitAll") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + const unsigned int count_requests = get_reqq_self()->size(); + + if (count_requests > 0) { + TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, + new simgrid::instr::Pt2PtTIData("waitAll", -1, count_requests, "")); + std::vector> sender_receiver; + for (const auto& req : (*get_reqq_self())) { + if (req && (req->flags() & RECV)) { + sender_receiver.push_back({req->src(), req->dst()}); + } + } + MPI_Status status[count_requests]; + Request::waitall(count_requests, &(*get_reqq_self())[0], status); + + for (auto& pair : sender_receiver) { + TRACE_smpi_recv(pair.first, pair.second, 0); + } + TRACE_smpi_comm_out(my_proc_id); + } } - log_timed_action (action, clock); -} +}; -static void action_barrier(const char *const *action){ - double clock = smpi_process()->simulated_elapsed(); - int my_proc_id = Actor::self()->getPid(); - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::NoOpTIData("barrier")); +class BarrierAction : public ReplayAction { +public: + BarrierAction() : ReplayAction("barrier") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::NoOpTIData("barrier")); + Colls::barrier(MPI_COMM_WORLD); + TRACE_smpi_comm_out(my_proc_id); + } +}; - Colls::barrier(MPI_COMM_WORLD); +} // Replay Namespace - TRACE_smpi_comm_out(my_proc_id); - log_timed_action (action, clock); -} - -static void action_bcast(const char *const *action) +static void action_bcast(simgrid::xbt::ReplayAction& action) { CHECK_ACTION_PARAMS(action, 1, 2) double size = parse_double(action[2]); double clock = smpi_process()->simulated_elapsed(); - int root = (action[3]) ? atoi(action[3]) : 0; + int root = (action.size() > 3) ? std::stoi(action[3]) : 0; /* Initialize MPI_CURRENT_TYPE in order to decrease the number of the checks */ - MPI_Datatype MPI_CURRENT_TYPE = (action[3] && action[4]) ? decode_datatype(action[4]) : MPI_DEFAULT_TYPE; + MPI_Datatype MPI_CURRENT_TYPE = (action.size() > 4) ? simgrid::smpi::Datatype::decode(action[4]) : MPI_DEFAULT_TYPE; int my_proc_id = Actor::self()->getPid(); TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::CollTIData("bcast", MPI_COMM_WORLD->group()->actor(root)->getPid(), -1.0, size, - -1, encode_datatype(MPI_CURRENT_TYPE), "")); + -1, Datatype::encode(MPI_CURRENT_TYPE), "")); void *sendbuf = smpi_get_tmp_sendbuffer(size* MPI_CURRENT_TYPE->size()); @@ -440,20 +340,20 @@ static void action_bcast(const char *const *action) log_timed_action (action, clock); } -static void action_reduce(const char *const *action) +static void action_reduce(simgrid::xbt::ReplayAction& action) { CHECK_ACTION_PARAMS(action, 2, 2) double comm_size = parse_double(action[2]); double comp_size = parse_double(action[3]); double clock = smpi_process()->simulated_elapsed(); - int root = (action[4]) ? atoi(action[4]) : 0; + int root = (action.size() > 4) ? std::stoi(action[4]) : 0; - MPI_Datatype MPI_CURRENT_TYPE = (action[4] && action[5]) ? decode_datatype(action[5]) : MPI_DEFAULT_TYPE; + MPI_Datatype MPI_CURRENT_TYPE = (action.size() > 5) ? simgrid::smpi::Datatype::decode(action[5]) : MPI_DEFAULT_TYPE; int my_proc_id = Actor::self()->getPid(); TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::CollTIData("reduce", MPI_COMM_WORLD->group()->actor(root)->getPid(), comp_size, - comm_size, -1, encode_datatype(MPI_CURRENT_TYPE), "")); + comm_size, -1, Datatype::encode(MPI_CURRENT_TYPE), "")); void *recvbuf = smpi_get_tmp_sendbuffer(comm_size* MPI_CURRENT_TYPE->size()); void *sendbuf = smpi_get_tmp_sendbuffer(comm_size* MPI_CURRENT_TYPE->size()); @@ -464,19 +364,20 @@ static void action_reduce(const char *const *action) log_timed_action (action, clock); } -static void action_allReduce(const char *const *action) { +static void action_allReduce(simgrid::xbt::ReplayAction& action) +{ CHECK_ACTION_PARAMS(action, 2, 1) double comm_size = parse_double(action[2]); double comp_size = parse_double(action[3]); - MPI_Datatype MPI_CURRENT_TYPE = (action[4]) ? decode_datatype(action[4]) : MPI_DEFAULT_TYPE; + MPI_Datatype MPI_CURRENT_TYPE = (action.size() > 4) ? simgrid::smpi::Datatype::decode(action[4]) : MPI_DEFAULT_TYPE; double clock = smpi_process()->simulated_elapsed(); int my_proc_id = Actor::self()->getPid(); TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::CollTIData("allReduce", -1, comp_size, comm_size, -1, - encode_datatype(MPI_CURRENT_TYPE), "")); + Datatype::encode(MPI_CURRENT_TYPE), "")); - void *recvbuf = smpi_get_tmp_sendbuffer(comm_size* MPI_CURRENT_TYPE->size()); + void *recvbuf = smpi_get_tmp_recvbuffer(comm_size* MPI_CURRENT_TYPE->size()); void *sendbuf = smpi_get_tmp_sendbuffer(comm_size* MPI_CURRENT_TYPE->size()); Colls::allreduce(sendbuf, recvbuf, comm_size, MPI_CURRENT_TYPE, MPI_OP_NULL, MPI_COMM_WORLD); smpi_execute_flops(comp_size); @@ -485,14 +386,15 @@ static void action_allReduce(const char *const *action) { log_timed_action (action, clock); } -static void action_allToAll(const char *const *action) { +static void action_allToAll(simgrid::xbt::ReplayAction& action) +{ CHECK_ACTION_PARAMS(action, 2, 2) //two mandatory (send and recv volumes) and two optional (corresponding datatypes) double clock = smpi_process()->simulated_elapsed(); - int comm_size = MPI_COMM_WORLD->size(); + unsigned long comm_size = MPI_COMM_WORLD->size(); int send_size = parse_double(action[2]); int recv_size = parse_double(action[3]); - MPI_Datatype MPI_CURRENT_TYPE = (action[4] && action[5]) ? decode_datatype(action[4]) : MPI_DEFAULT_TYPE; - MPI_Datatype MPI_CURRENT_TYPE2{(action[4] && action[5]) ? decode_datatype(action[5]) : MPI_DEFAULT_TYPE}; + MPI_Datatype MPI_CURRENT_TYPE{(action.size() > 5) ? simgrid::smpi::Datatype::decode(action[4]) : MPI_DEFAULT_TYPE}; + MPI_Datatype MPI_CURRENT_TYPE2{(action.size() > 5) ? simgrid::smpi::Datatype::decode(action[5]) : MPI_DEFAULT_TYPE}; void *send = smpi_get_tmp_sendbuffer(send_size*comm_size* MPI_CURRENT_TYPE->size()); void *recv = smpi_get_tmp_recvbuffer(recv_size*comm_size* MPI_CURRENT_TYPE2->size()); @@ -500,8 +402,8 @@ static void action_allToAll(const char *const *action) { int my_proc_id = Actor::self()->getPid(); TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::CollTIData("allToAll", -1, -1.0, send_size, recv_size, - encode_datatype(MPI_CURRENT_TYPE), - encode_datatype(MPI_CURRENT_TYPE2))); + Datatype::encode(MPI_CURRENT_TYPE), + Datatype::encode(MPI_CURRENT_TYPE2))); Colls::alltoall(send, send_size, MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, MPI_COMM_WORLD); @@ -509,35 +411,36 @@ static void action_allToAll(const char *const *action) { log_timed_action (action, clock); } -static void action_gather(const char *const *action) { +static void action_gather(simgrid::xbt::ReplayAction& action) +{ /* The structure of the gather action for the rank 0 (total 4 processes) is the following: 0 gather 68 68 0 0 0 where: 1) 68 is the sendcounts 2) 68 is the recvcounts 3) 0 is the root node - 4) 0 is the send datatype id, see decode_datatype() - 5) 0 is the recv datatype id, see decode_datatype() + 4) 0 is the send datatype id, see simgrid::smpi::Datatype::decode() + 5) 0 is the recv datatype id, see simgrid::smpi::Datatype::decode() */ CHECK_ACTION_PARAMS(action, 2, 3) double clock = smpi_process()->simulated_elapsed(); - int comm_size = MPI_COMM_WORLD->size(); + unsigned long comm_size = MPI_COMM_WORLD->size(); int send_size = parse_double(action[2]); int recv_size = parse_double(action[3]); - MPI_Datatype MPI_CURRENT_TYPE = (action[5] && action[6]) ? decode_datatype(action[5]) : MPI_DEFAULT_TYPE; - MPI_Datatype MPI_CURRENT_TYPE2{(action[5] && action[6]) ? decode_datatype(action[6]) : MPI_DEFAULT_TYPE}; + MPI_Datatype MPI_CURRENT_TYPE{(action.size() > 6) ? simgrid::smpi::Datatype::decode(action[5]) : MPI_DEFAULT_TYPE}; + MPI_Datatype MPI_CURRENT_TYPE2{(action.size() > 6) ? simgrid::smpi::Datatype::decode(action[6]) : MPI_DEFAULT_TYPE}; void *send = smpi_get_tmp_sendbuffer(send_size* MPI_CURRENT_TYPE->size()); void *recv = nullptr; - int root = (action[4]) ? atoi(action[4]) : 0; + int root = (action.size() > 4) ? std::stoi(action[4]) : 0; int rank = MPI_COMM_WORLD->rank(); if(rank==root) recv = smpi_get_tmp_recvbuffer(recv_size*comm_size* MPI_CURRENT_TYPE2->size()); TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::CollTIData("gather", root, -1.0, send_size, recv_size, - encode_datatype(MPI_CURRENT_TYPE), - encode_datatype(MPI_CURRENT_TYPE2))); + Datatype::encode(MPI_CURRENT_TYPE), + Datatype::encode(MPI_CURRENT_TYPE2))); Colls::gather(send, send_size, MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, root, MPI_COMM_WORLD); @@ -545,7 +448,7 @@ static void action_gather(const char *const *action) { log_timed_action (action, clock); } -static void action_scatter(const char* const* action) +static void action_scatter(simgrid::xbt::ReplayAction& action) { /* The structure of the scatter action for the rank 0 (total 4 processes) is the following: 0 gather 68 68 0 0 0 @@ -553,28 +456,28 @@ static void action_scatter(const char* const* action) 1) 68 is the sendcounts 2) 68 is the recvcounts 3) 0 is the root node - 4) 0 is the send datatype id, see decode_datatype() - 5) 0 is the recv datatype id, see decode_datatype() + 4) 0 is the send datatype id, see simgrid::smpi::Datatype::decode() + 5) 0 is the recv datatype id, see simgrid::smpi::Datatype::decode() */ CHECK_ACTION_PARAMS(action, 2, 3) double clock = smpi_process()->simulated_elapsed(); - int comm_size = MPI_COMM_WORLD->size(); + unsigned long comm_size = MPI_COMM_WORLD->size(); int send_size = parse_double(action[2]); int recv_size = parse_double(action[3]); - MPI_Datatype MPI_CURRENT_TYPE = (action[5] && action[6]) ? decode_datatype(action[5]) : MPI_DEFAULT_TYPE; - MPI_Datatype MPI_CURRENT_TYPE2{(action[5] && action[6]) ? decode_datatype(action[6]) : MPI_DEFAULT_TYPE}; + MPI_Datatype MPI_CURRENT_TYPE{(action.size() > 6) ? simgrid::smpi::Datatype::decode(action[5]) : MPI_DEFAULT_TYPE}; + MPI_Datatype MPI_CURRENT_TYPE2{(action.size() > 6) ? simgrid::smpi::Datatype::decode(action[6]) : MPI_DEFAULT_TYPE}; void* send = smpi_get_tmp_sendbuffer(send_size * MPI_CURRENT_TYPE->size()); void* recv = nullptr; - int root = (action[4]) ? atoi(action[4]) : 0; + int root = (action.size() > 4) ? std::stoi(action[4]) : 0; int rank = MPI_COMM_WORLD->rank(); if (rank == root) recv = smpi_get_tmp_recvbuffer(recv_size * comm_size * MPI_CURRENT_TYPE2->size()); TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::CollTIData("gather", root, -1.0, send_size, recv_size, - encode_datatype(MPI_CURRENT_TYPE), - encode_datatype(MPI_CURRENT_TYPE2))); + Datatype::encode(MPI_CURRENT_TYPE), + Datatype::encode(MPI_CURRENT_TYPE2))); Colls::scatter(send, send_size, MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, root, MPI_COMM_WORLD); @@ -582,36 +485,37 @@ static void action_scatter(const char* const* action) log_timed_action(action, clock); } -static void action_gatherv(const char *const *action) { +static void action_gatherv(simgrid::xbt::ReplayAction& action) +{ /* The structure of the gatherv action for the rank 0 (total 4 processes) is the following: 0 gather 68 68 10 10 10 0 0 0 where: 1) 68 is the sendcount 2) 68 10 10 10 is the recvcounts 3) 0 is the root node - 4) 0 is the send datatype id, see decode_datatype() - 5) 0 is the recv datatype id, see decode_datatype() + 4) 0 is the send datatype id, see simgrid::smpi::Datatype::decode() + 5) 0 is the recv datatype id, see simgrid::smpi::Datatype::decode() */ double clock = smpi_process()->simulated_elapsed(); - int comm_size = MPI_COMM_WORLD->size(); + unsigned long comm_size = MPI_COMM_WORLD->size(); CHECK_ACTION_PARAMS(action, comm_size+1, 2) int send_size = parse_double(action[2]); std::vector disps(comm_size, 0); std::shared_ptr> recvcounts(new std::vector(comm_size)); MPI_Datatype MPI_CURRENT_TYPE = - (action[4 + comm_size] && action[5 + comm_size]) ? decode_datatype(action[4 + comm_size]) : MPI_DEFAULT_TYPE; + (action.size() > 5 + comm_size) ? simgrid::smpi::Datatype::decode(action[4 + comm_size]) : MPI_DEFAULT_TYPE; MPI_Datatype MPI_CURRENT_TYPE2{ - (action[4 + comm_size] && action[5 + comm_size]) ? decode_datatype(action[5 + comm_size]) : MPI_DEFAULT_TYPE}; + (action.size() > 5 + comm_size) ? simgrid::smpi::Datatype::decode(action[5 + comm_size]) : MPI_DEFAULT_TYPE}; void *send = smpi_get_tmp_sendbuffer(send_size* MPI_CURRENT_TYPE->size()); void *recv = nullptr; - for(int i=0;ibegin(), recvcounts->end(), 0); - int root = (action[3 + comm_size]) ? atoi(action[3 + comm_size]) : 0; + int root = (action.size() > 3 + comm_size) ? std::stoi(action[3 + comm_size]) : 0; int rank = MPI_COMM_WORLD->rank(); if(rank==root) @@ -619,7 +523,7 @@ static void action_gatherv(const char *const *action) { TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::VarCollTIData( "gatherV", root, send_size, nullptr, -1, recvcounts, - encode_datatype(MPI_CURRENT_TYPE), encode_datatype(MPI_CURRENT_TYPE2))); + Datatype::encode(MPI_CURRENT_TYPE), Datatype::encode(MPI_CURRENT_TYPE2))); Colls::gatherv(send, send_size, MPI_CURRENT_TYPE, recv, recvcounts->data(), disps.data(), MPI_CURRENT_TYPE2, root, MPI_COMM_WORLD); @@ -628,7 +532,7 @@ static void action_gatherv(const char *const *action) { log_timed_action (action, clock); } -static void action_scatterv(const char* const* action) +static void action_scatterv(simgrid::xbt::ReplayAction& action) { /* The structure of the scatterv action for the rank 0 (total 4 processes) is the following: 0 gather 68 10 10 10 68 0 0 0 @@ -636,37 +540,37 @@ static void action_scatterv(const char* const* action) 1) 68 10 10 10 is the sendcounts 2) 68 is the recvcount 3) 0 is the root node - 4) 0 is the send datatype id, see decode_datatype() - 5) 0 is the recv datatype id, see decode_datatype() + 4) 0 is the send datatype id, see simgrid::smpi::Datatype::decode() + 5) 0 is the recv datatype id, see simgrid::smpi::Datatype::decode() */ double clock = smpi_process()->simulated_elapsed(); - int comm_size = MPI_COMM_WORLD->size(); + unsigned long comm_size = MPI_COMM_WORLD->size(); CHECK_ACTION_PARAMS(action, comm_size + 1, 2) int recv_size = parse_double(action[2 + comm_size]); std::vector disps(comm_size, 0); std::shared_ptr> sendcounts(new std::vector(comm_size)); MPI_Datatype MPI_CURRENT_TYPE = - (action[4 + comm_size] && action[5 + comm_size]) ? decode_datatype(action[4 + comm_size]) : MPI_DEFAULT_TYPE; + (action.size() > 5 + comm_size) ? simgrid::smpi::Datatype::decode(action[4 + comm_size]) : MPI_DEFAULT_TYPE; MPI_Datatype MPI_CURRENT_TYPE2{ - (action[4 + comm_size] && action[5 + comm_size]) ? decode_datatype(action[5 + comm_size]) : MPI_DEFAULT_TYPE}; + (action.size() > 5 + comm_size) ? simgrid::smpi::Datatype::decode(action[5 + comm_size]) : MPI_DEFAULT_TYPE}; void* send = nullptr; void* recv = smpi_get_tmp_recvbuffer(recv_size * MPI_CURRENT_TYPE->size()); - for (int i = 0; i < comm_size; i++) { - (*sendcounts)[i] = atoi(action[i + 2]); + for (unsigned int i = 0; i < comm_size; i++) { + (*sendcounts)[i] = std::stoi(action[i + 2]); } int send_sum = std::accumulate(sendcounts->begin(), sendcounts->end(), 0); - int root = (action[3 + comm_size]) ? atoi(action[3 + comm_size]) : 0; + int root = (action.size() > 3 + comm_size) ? std::stoi(action[3 + comm_size]) : 0; int rank = MPI_COMM_WORLD->rank(); if (rank == root) send = smpi_get_tmp_sendbuffer(send_sum * MPI_CURRENT_TYPE2->size()); TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::VarCollTIData("gatherV", root, -1, sendcounts, recv_size, - nullptr, encode_datatype(MPI_CURRENT_TYPE), - encode_datatype(MPI_CURRENT_TYPE2))); + nullptr, Datatype::encode(MPI_CURRENT_TYPE), + Datatype::encode(MPI_CURRENT_TYPE2))); Colls::scatterv(send, sendcounts->data(), disps.data(), MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, root, MPI_COMM_WORLD); @@ -675,31 +579,33 @@ static void action_scatterv(const char* const* action) log_timed_action(action, clock); } -static void action_reducescatter(const char *const *action) { - /* The structure of the reducescatter action for the rank 0 (total 4 processes) is the following: - 0 reduceScatter 275427 275427 275427 204020 11346849 0 - where: - 1) The first four values after the name of the action declare the recvcounts array - 2) The value 11346849 is the amount of instructions - 3) The last value corresponds to the datatype, see decode_datatype(). -*/ +static void action_reducescatter(simgrid::xbt::ReplayAction& action) +{ + /* The structure of the reducescatter action for the rank 0 (total 4 processes) is the following: + 0 reduceScatter 275427 275427 275427 204020 11346849 0 + where: + 1) The first four values after the name of the action declare the recvcounts array + 2) The value 11346849 is the amount of instructions + 3) The last value corresponds to the datatype, see simgrid::smpi::Datatype::decode(). + */ double clock = smpi_process()->simulated_elapsed(); - int comm_size = MPI_COMM_WORLD->size(); + unsigned long comm_size = MPI_COMM_WORLD->size(); CHECK_ACTION_PARAMS(action, comm_size+1, 1) int comp_size = parse_double(action[2+comm_size]); int my_proc_id = Actor::self()->getPid(); std::shared_ptr> recvcounts(new std::vector); - MPI_Datatype MPI_CURRENT_TYPE = (action[3 + comm_size]) ? decode_datatype(action[3 + comm_size]) : MPI_DEFAULT_TYPE; + MPI_Datatype MPI_CURRENT_TYPE = + (action.size() > 3 + comm_size) ? simgrid::smpi::Datatype::decode(action[3 + comm_size]) : MPI_DEFAULT_TYPE; - for(int i=0;ipush_back(atoi(action[i + 2])); + for (unsigned int i = 0; i < comm_size; i++) { + recvcounts->push_back(std::stoi(action[i + 2])); } int size{std::accumulate(recvcounts->begin(), recvcounts->end(), 0)}; TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::VarCollTIData("reduceScatter", -1, 0, nullptr, -1, recvcounts, std::to_string(comp_size), /* ugly hack to print comp_size */ - encode_datatype(MPI_CURRENT_TYPE))); + Datatype::encode(MPI_CURRENT_TYPE))); void *sendbuf = smpi_get_tmp_sendbuffer(size* MPI_CURRENT_TYPE->size()); void *recvbuf = smpi_get_tmp_recvbuffer(size* MPI_CURRENT_TYPE->size()); @@ -711,22 +617,24 @@ static void action_reducescatter(const char *const *action) { log_timed_action (action, clock); } -static void action_allgather(const char *const *action) { +static void action_allgather(simgrid::xbt::ReplayAction& action) +{ /* The structure of the allgather action for the rank 0 (total 4 processes) is the following: 0 allGather 275427 275427 where: 1) 275427 is the sendcount 2) 275427 is the recvcount - 3) No more values mean that the datatype for sent and receive buffer is the default one, see decode_datatype(). + 3) No more values mean that the datatype for sent and receive buffer is the default one, see + simgrid::smpi::Datatype::decode(). */ double clock = smpi_process()->simulated_elapsed(); CHECK_ACTION_PARAMS(action, 2, 2) - int sendcount=atoi(action[2]); - int recvcount=atoi(action[3]); + int sendcount = std::stoi(action[2]); + int recvcount = std::stoi(action[3]); - MPI_Datatype MPI_CURRENT_TYPE = (action[4] && action[5]) ? decode_datatype(action[4]) : MPI_DEFAULT_TYPE; - MPI_Datatype MPI_CURRENT_TYPE2{(action[4] && action[5]) ? decode_datatype(action[5]) : MPI_DEFAULT_TYPE}; + MPI_Datatype MPI_CURRENT_TYPE{(action.size() > 5) ? simgrid::smpi::Datatype::decode(action[4]) : MPI_DEFAULT_TYPE}; + MPI_Datatype MPI_CURRENT_TYPE2{(action.size() > 5) ? simgrid::smpi::Datatype::decode(action[5]) : MPI_DEFAULT_TYPE}; void *sendbuf = smpi_get_tmp_sendbuffer(sendcount* MPI_CURRENT_TYPE->size()); void *recvbuf = smpi_get_tmp_recvbuffer(recvcount* MPI_CURRENT_TYPE2->size()); @@ -735,8 +643,8 @@ static void action_allgather(const char *const *action) { TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::CollTIData("allGather", -1, -1.0, sendcount, recvcount, - encode_datatype(MPI_CURRENT_TYPE), - encode_datatype(MPI_CURRENT_TYPE2))); + Datatype::encode(MPI_CURRENT_TYPE), + Datatype::encode(MPI_CURRENT_TYPE2))); Colls::allgather(sendbuf, sendcount, MPI_CURRENT_TYPE, recvbuf, recvcount, MPI_CURRENT_TYPE2, MPI_COMM_WORLD); @@ -744,31 +652,49 @@ static void action_allgather(const char *const *action) { log_timed_action (action, clock); } -static void action_allgatherv(const char *const *action) { +static void action_allgatherv(simgrid::xbt::ReplayAction& action) +{ /* The structure of the allgatherv action for the rank 0 (total 4 processes) is the following: 0 allGatherV 275427 275427 275427 275427 204020 where: 1) 275427 is the sendcount 2) The next four elements declare the recvcounts array - 3) No more values mean that the datatype for sent and receive buffer is the default one, see decode_datatype(). + 3) No more values mean that the datatype for sent and receive buffer is the default one, see + simgrid::smpi::Datatype::decode(). */ double clock = smpi_process()->simulated_elapsed(); - int comm_size = MPI_COMM_WORLD->size(); + unsigned long comm_size = MPI_COMM_WORLD->size(); CHECK_ACTION_PARAMS(action, comm_size+1, 2) - int sendcount=atoi(action[2]); + int sendcount = std::stoi(action[2]); std::shared_ptr> recvcounts(new std::vector(comm_size)); std::vector disps(comm_size, 0); - MPI_Datatype MPI_CURRENT_TYPE = - (action[3 + comm_size] && action[4 + comm_size]) ? decode_datatype(action[3 + comm_size]) : MPI_DEFAULT_TYPE; - MPI_Datatype MPI_CURRENT_TYPE2{ - (action[3 + comm_size] && action[4 + comm_size]) ? decode_datatype(action[4 + comm_size]) : MPI_DEFAULT_TYPE}; + int datatype_index = 0, disp_index = 0; + if (action.size() > 3 + 2 * comm_size) { /* datatype + disp are specified */ + datatype_index = 3 + comm_size; + disp_index = datatype_index + 1; + } else if (action.size() > 3 + 2 * comm_size) { /* disps specified; datatype is not specified; use the default one */ + datatype_index = -1; + disp_index = 3 + comm_size; + } else if (action.size() > 3 + comm_size) { /* only datatype, no disp specified */ + datatype_index = 3 + comm_size; + } + + if (disp_index != 0) { + for (unsigned int i = 0; i < comm_size; i++) + disps[i] = std::stoi(action[disp_index + i]); + } + + MPI_Datatype MPI_CURRENT_TYPE{(datatype_index > 0) ? simgrid::smpi::Datatype::decode(action[datatype_index]) + : MPI_DEFAULT_TYPE}; + MPI_Datatype MPI_CURRENT_TYPE2{(datatype_index > 0) ? simgrid::smpi::Datatype::decode(action[datatype_index]) + : MPI_DEFAULT_TYPE}; void *sendbuf = smpi_get_tmp_sendbuffer(sendcount* MPI_CURRENT_TYPE->size()); - for(int i=0;ibegin(), recvcounts->end(), 0); void *recvbuf = smpi_get_tmp_recvbuffer(recv_sum* MPI_CURRENT_TYPE2->size()); @@ -777,8 +703,8 @@ static void action_allgatherv(const char *const *action) { TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::VarCollTIData("allGatherV", -1, sendcount, nullptr, -1, recvcounts, - encode_datatype(MPI_CURRENT_TYPE), - encode_datatype(MPI_CURRENT_TYPE2))); + Datatype::encode(MPI_CURRENT_TYPE), + Datatype::encode(MPI_CURRENT_TYPE2))); Colls::allgatherv(sendbuf, sendcount, MPI_CURRENT_TYPE, recvbuf, recvcounts->data(), disps.data(), MPI_CURRENT_TYPE2, MPI_COMM_WORLD); @@ -787,7 +713,8 @@ static void action_allgatherv(const char *const *action) { log_timed_action (action, clock); } -static void action_allToAllv(const char *const *action) { +static void action_allToAllv(simgrid::xbt::ReplayAction& action) +{ /* The structure of the allToAllV action for the rank 0 (total 4 processes) is the following: 0 allToAllV 100 1 7 10 12 100 1 70 10 5 where: @@ -798,18 +725,18 @@ static void action_allToAllv(const char *const *action) { */ double clock = smpi_process()->simulated_elapsed(); - int comm_size = MPI_COMM_WORLD->size(); + unsigned long comm_size = MPI_COMM_WORLD->size(); CHECK_ACTION_PARAMS(action, 2*comm_size+2, 2) std::shared_ptr> sendcounts(new std::vector(comm_size)); std::shared_ptr> recvcounts(new std::vector(comm_size)); std::vector senddisps(comm_size, 0); std::vector recvdisps(comm_size, 0); - MPI_Datatype MPI_CURRENT_TYPE = (action[4 + 2 * comm_size] && action[5 + 2 * comm_size]) - ? decode_datatype(action[4 + 2 * comm_size]) + MPI_Datatype MPI_CURRENT_TYPE = (action.size() > 5 + 2 * comm_size) + ? simgrid::smpi::Datatype::decode(action[4 + 2 * comm_size]) : MPI_DEFAULT_TYPE; - MPI_Datatype MPI_CURRENT_TYPE2{(action[4 + 2 * comm_size] && action[5 + 2 * comm_size]) - ? decode_datatype(action[5 + 2 * comm_size]) + MPI_Datatype MPI_CURRENT_TYPE2{(action.size() > 5 + 2 * comm_size) + ? simgrid::smpi::Datatype::decode(action[5 + 2 * comm_size]) : MPI_DEFAULT_TYPE}; int send_buf_size=parse_double(action[2]); @@ -818,17 +745,17 @@ static void action_allToAllv(const char *const *action) { void *sendbuf = smpi_get_tmp_sendbuffer(send_buf_size* MPI_CURRENT_TYPE->size()); void *recvbuf = smpi_get_tmp_recvbuffer(recv_buf_size* MPI_CURRENT_TYPE2->size()); - for(int i=0;ibegin(), sendcounts->end(), 0); int recv_size = std::accumulate(recvcounts->begin(), recvcounts->end(), 0); TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::VarCollTIData("allToAllV", -1, send_size, sendcounts, recv_size, recvcounts, - encode_datatype(MPI_CURRENT_TYPE), - encode_datatype(MPI_CURRENT_TYPE2))); + Datatype::encode(MPI_CURRENT_TYPE), + Datatype::encode(MPI_CURRENT_TYPE2))); Colls::alltoallv(sendbuf, sendcounts->data(), senddisps.data(), MPI_CURRENT_TYPE, recvbuf, recvcounts->data(), recvdisps.data(), MPI_CURRENT_TYPE, MPI_COMM_WORLD); @@ -851,18 +778,19 @@ void smpi_replay_init(int* argc, char*** argv) TRACE_smpi_computing_init(my_proc_id); TRACE_smpi_comm_in(my_proc_id, "smpi_replay_run_init", new simgrid::instr::NoOpTIData("init")); TRACE_smpi_comm_out(my_proc_id); - xbt_replay_action_register("init", simgrid::smpi::action_init); - xbt_replay_action_register("finalize", simgrid::smpi::action_finalize); - xbt_replay_action_register("comm_size", simgrid::smpi::action_comm_size); - xbt_replay_action_register("comm_split", simgrid::smpi::action_comm_split); - xbt_replay_action_register("comm_dup", simgrid::smpi::action_comm_dup); - xbt_replay_action_register("send", simgrid::smpi::action_send); - xbt_replay_action_register("Isend", simgrid::smpi::action_Isend); - xbt_replay_action_register("recv", simgrid::smpi::action_recv); - xbt_replay_action_register("Irecv", simgrid::smpi::action_Irecv); - xbt_replay_action_register("test", simgrid::smpi::action_test); - xbt_replay_action_register("wait", simgrid::smpi::action_wait); - xbt_replay_action_register("waitAll", simgrid::smpi::action_waitall); + xbt_replay_action_register("init", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::InitAction().execute(action); }); + xbt_replay_action_register("finalize", [](simgrid::xbt::ReplayAction& action) { /* nothing to do */ }); + xbt_replay_action_register("comm_size", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::CommunicatorAction().execute(action); }); + xbt_replay_action_register("comm_split",[](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::CommunicatorAction().execute(action); }); + xbt_replay_action_register("comm_dup", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::CommunicatorAction().execute(action); }); + + xbt_replay_action_register("send", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::SendAction("send").execute(action); }); + xbt_replay_action_register("Isend", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::SendAction("Isend").execute(action); }); + xbt_replay_action_register("recv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::RecvAction("recv").execute(action); }); + xbt_replay_action_register("Irecv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::RecvAction("Irecv").execute(action); }); + xbt_replay_action_register("test", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::TestAction().execute(action); }); + xbt_replay_action_register("wait", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::WaitAction().execute(action); }); + xbt_replay_action_register("waitAll", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::WaitAllAction().execute(action); }); xbt_replay_action_register("barrier", simgrid::smpi::action_barrier); xbt_replay_action_register("bcast", simgrid::smpi::action_bcast); xbt_replay_action_register("reduce", simgrid::smpi::action_reduce); @@ -876,7 +804,7 @@ void smpi_replay_init(int* argc, char*** argv) xbt_replay_action_register("allGather", simgrid::smpi::action_allgather); xbt_replay_action_register("allGatherV", simgrid::smpi::action_allgatherv); xbt_replay_action_register("reduceScatter", simgrid::smpi::action_reducescatter); - xbt_replay_action_register("compute", simgrid::smpi::action_compute); + xbt_replay_action_register("compute", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::ComputeAction().execute(action); }); //if we have a delayed start, sleep here. if(*argc>2){ @@ -916,8 +844,7 @@ void smpi_replay_main(int* argc, char*** argv) if(active_processes==0){ /* Last process alive speaking: end the simulated timer */ XBT_INFO("Simulation time %f", smpi_process()->simulated_elapsed()); - xbt_free(sendbuffer); - xbt_free(recvbuffer); + smpi_free_replay_tmp_buffers(); } TRACE_smpi_comm_in(Actor::self()->getPid(), "smpi_replay_run_finalize", new simgrid::instr::NoOpTIData("finalize"));