-/* Copyright (c) 2007-2018. The SimGrid Team. All rights reserved. */
+/* Copyright (c) 2007-2019. 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. */
#ifndef SMPI_PRIVATE_HPP
#define SMPI_PRIVATE_HPP
-#include "simgrid/msg.h" // msg_bar_t
+#include "simgrid/s4u/Barrier.hpp"
#include "smpi/smpi.h"
#include "smpi/smpi_helpers_internal.h"
#include "src/instr/instr_smpi.hpp"
-#include "src/internal_config.h"
#include <unordered_map>
#include <vector>
-#define MPI_REQ_PERSISTENT 0x1
-#define MPI_REQ_NON_PERSISTENT 0x2
-#define MPI_REQ_SEND 0x4
-#define MPI_REQ_RECV 0x8
-//#define MPI_REQ_RECV_DELETE 0x10
-#define MPI_REQ_ISEND 0x20
-#define MPI_REQ_SSEND 0x40
-#define MPI_REQ_PREPARED 0x80
-#define MPI_REQ_FINISHED 0x100
-#define MPI_REQ_RMA 0x200
-#define MPI_REQ_ACCUMULATE 0x400
-
-enum class SmpiProcessState { UNINITIALIZED, INITIALIZED, FINALIZED };
-
-#define COLL_TAG_REDUCE -112
-#define COLL_TAG_SCATTER -223
-#define COLL_TAG_SCATTERV -334
-#define COLL_TAG_GATHER -445
-#define COLL_TAG_ALLGATHER -556
-#define COLL_TAG_ALLGATHERV -667
-#define COLL_TAG_BARRIER -778
-#define COLL_TAG_REDUCE_SCATTER -889
-#define COLL_TAG_ALLTOALLV -1000
-#define COLL_TAG_ALLTOALL -1112
-#define COLL_TAG_GATHERV -2223
-#define COLL_TAG_BCAST -3334
-#define COLL_TAG_ALLREDUCE -4445
+constexpr unsigned MPI_REQ_PERSISTENT = 0x1;
+constexpr unsigned MPI_REQ_NON_PERSISTENT = 0x2;
+constexpr unsigned MPI_REQ_SEND = 0x4;
+constexpr unsigned MPI_REQ_RECV = 0x8;
+constexpr unsigned MPI_REQ_RECV_DELETE = 0x10;
+constexpr unsigned MPI_REQ_ISEND = 0x20;
+constexpr unsigned MPI_REQ_SSEND = 0x40;
+constexpr unsigned MPI_REQ_PREPARED = 0x80;
+constexpr unsigned MPI_REQ_FINISHED = 0x100;
+constexpr unsigned MPI_REQ_RMA = 0x200;
+constexpr unsigned MPI_REQ_ACCUMULATE = 0x400;
+constexpr unsigned MPI_REQ_GENERALIZED = 0x800;
+constexpr unsigned MPI_REQ_COMPLETE = 0x1000;
+
+enum class SmpiProcessState { UNINITIALIZED, INITIALIZING, INITIALIZED /*(=MPI_Init called)*/, FINALIZED };
+
+constexpr int COLL_TAG_REDUCE = -112;
+constexpr int COLL_TAG_SCATTER = -223;
+constexpr int COLL_TAG_SCATTERV = -334;
+constexpr int COLL_TAG_GATHER = -445;
+constexpr int COLL_TAG_ALLGATHER = -556;
+constexpr int COLL_TAG_ALLGATHERV = -667;
+constexpr int COLL_TAG_BARRIER = -778;
+constexpr int COLL_TAG_REDUCE_SCATTER = -889;
+constexpr int COLL_TAG_ALLTOALLV = -1000;
+constexpr int COLL_TAG_ALLTOALL = -1112;
+constexpr int COLL_TAG_GATHERV = -2223;
+constexpr int COLL_TAG_BCAST = -3334;
+constexpr int COLL_TAG_ALLREDUCE = -4445;
// SMPI_RMA_TAG has to be the smallest one, as it will be decremented for accumulate ordering.
-#define SMPI_RMA_TAG -6666
+constexpr int SMPI_RMA_TAG = -6666;
-/* Convert between Fortran and C */
+#define MPI_REQUEST_IGNORED ((MPI_Request*)-100)
-#define FORT_BOTTOM(addr) ((*(int*)addr) == -200 ? MPI_BOTTOM : (void*)addr)
-#define FORT_IN_PLACE(addr) ((*(int*)addr) == -100 ? MPI_IN_PLACE : (void*)addr)
-#define FORT_STATUS_IGNORE(addr) (static_cast<MPI_Status*>((*(int*)addr) == -300 ? MPI_STATUS_IGNORE : (void*)addr))
-#define FORT_STATUSES_IGNORE(addr) (static_cast<MPI_Status*>((*(int*)addr) == -400 ? MPI_STATUSES_IGNORE : (void*)addr))
+/* Bindings for MPI special values */
+extern XBT_PUBLIC int mpi_in_place_;
+extern XBT_PUBLIC int mpi_bottom_;
+extern XBT_PUBLIC int mpi_status_ignore_;
+extern XBT_PUBLIC int mpi_statuses_ignore_;
+/* Convert between Fortran and C */
+#define FORT_ADDR(addr, val, val2) \
+ (((void *)(addr) == (void*) &(val2)) \
+ ? (val) : (void *)(addr))
+#define FORT_BOTTOM(addr) FORT_ADDR(addr, MPI_BOTTOM, mpi_bottom_)
+#define FORT_IN_PLACE(addr) FORT_ADDR(addr, MPI_IN_PLACE, mpi_in_place_)
+#define FORT_STATUS_IGNORE(addr) static_cast<MPI_Status*>(FORT_ADDR(addr, MPI_STATUS_IGNORE, mpi_status_ignore_))
+#define FORT_STATUSES_IGNORE(addr) static_cast<MPI_Status*>(FORT_ADDR(addr, MPI_STATUSES_IGNORE, mpi_statuses_ignore_))
extern XBT_PRIVATE MPI_Comm MPI_COMM_UNINITIALIZED;
typedef SMPI_Dist_Graph_topology* MPIR_Dist_Graph_Topology;
-XBT_PRIVATE SMPI_Process* smpi_process();
-XBT_PRIVATE SMPI_Process* smpi_process_remote(simgrid::s4u::ActorPtr actor);
-XBT_PRIVATE int smpi_process_count();
+XBT_PRIVATE simgrid::smpi::ActorExt* smpi_process();
+XBT_PRIVATE simgrid::smpi::ActorExt* smpi_process_remote(simgrid::s4u::ActorPtr actor);
+XBT_PRIVATE int smpi_get_universe_size();
-XBT_PRIVATE void smpi_deployment_register_process(const std::string instance_id, int rank,
+XBT_PRIVATE void smpi_deployment_register_process(const std::string& instance_id, int rank,
simgrid::s4u::ActorPtr actor);
-XBT_PRIVATE MPI_Comm* smpi_deployment_comm_world(const std::string instance_id);
-XBT_PRIVATE msg_bar_t smpi_deployment_finalization_barrier(const std::string instance_id);
+XBT_PRIVATE MPI_Comm* smpi_deployment_comm_world(const std::string& instance_id);
+XBT_PRIVATE simgrid::s4u::Barrier* smpi_deployment_finalization_barrier(const std::string& instance_id);
XBT_PRIVATE void smpi_deployment_cleanup_instances();
-XBT_PRIVATE void smpi_comm_copy_buffer_callback(smx_activity_t comm, void* buff, size_t buff_size);
+XBT_PRIVATE void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff,
+ size_t buff_size);
-XBT_PRIVATE void smpi_comm_null_copy_buffer_callback(smx_activity_t comm, void* buff, size_t buff_size);
+XBT_PRIVATE void smpi_comm_null_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff,
+ size_t buff_size);
XBT_PRIVATE int smpi_enabled();
XBT_PRIVATE void smpi_global_init();
extern XBT_PRIVATE SharedMallocType smpi_cfg_shared_malloc; // Whether to activate shared malloc
XBT_PRIVATE void smpi_switch_data_segment(simgrid::s4u::ActorPtr actor);
-XBT_PRIVATE void smpi_really_switch_data_segment(simgrid::s4u::ActorPtr actor);
-XBT_PRIVATE int smpi_is_privatization_file(char* file);
XBT_PRIVATE void smpi_prepare_global_memory_segment();
XBT_PRIVATE void smpi_backup_global_memory_segment();
XBT_PRIVATE void smpi_bench_end();
XBT_PRIVATE void smpi_shared_destroy();
-XBT_PRIVATE void* smpi_get_tmp_sendbuffer(int size);
-XBT_PRIVATE void* smpi_get_tmp_recvbuffer(int size);
-XBT_PRIVATE void smpi_free_tmp_buffer(void* buf);
+XBT_PRIVATE unsigned char* smpi_get_tmp_sendbuffer(size_t size);
+XBT_PRIVATE unsigned char* smpi_get_tmp_recvbuffer(size_t size);
+XBT_PRIVATE void smpi_free_tmp_buffer(const unsigned char* buf);
XBT_PRIVATE void smpi_free_replay_tmp_buffers();
extern "C" {
void mpi_reduce_(void* sendbuf, void* recvbuf, int* count, int* datatype, int* op, int* root, int* comm, int* ierr);
void mpi_allreduce_(void* sendbuf, void* recvbuf, int* count, int* datatype, int* op, int* comm, int* ierr);
void mpi_reduce_scatter_(void* sendbuf, void* recvbuf, int* recvcounts, int* datatype, int* op, int* comm, int* ierr);
+void mpi_reduce_scatter_block_(void* sendbuf, void* recvbuf, int* recvcount, int* datatype, int* op, int* comm,
+ int* ierr);
void mpi_scatter_(void* sendbuf, int* sendcount, int* sendtype, void* recvbuf, int* recvcount, int* recvtype, int* root,
int* comm, int* ierr);
void mpi_scatterv_(void* sendbuf, int* sendcounts, int* displs, int* sendtype, void* recvbuf, int* recvcount,
int* comm, int* ierr);
void mpi_allgatherv_(void* sendbuf, int* sendcount, int* sendtype, void* recvbuf, int* recvcount, int* displs,
int* recvtype, int* comm, int* ierr);
-void mpi_type_size_(int* datatype, int* size, int* ierr);
-
void mpi_scan_(void* sendbuf, void* recvbuf, int* count, int* datatype, int* op, int* comm, int* ierr);
void mpi_alltoall_(void* sendbuf, int* sendcount, int* sendtype, void* recvbuf, int* recvcount, int* recvtype,
int* comm, int* ierr);
void mpi_alltoallv_(void* sendbuf, int* sendcounts, int* senddisps, int* sendtype, void* recvbuf, int* recvcounts,
int* recvdisps, int* recvtype, int* comm, int* ierr);
+void mpi_alltoallw_(void* sendbuf, int* sendcnts, int* sdispls, int* sendtypes, void* recvbuf, int* recvcnts,
+ int* rdispls, int* recvtypes, int* comm, int* ierr);
+void mpi_exscan_(void* sendbuf, void* recvbuf, int* count, int* datatype, int* op, int* comm, int* ierr);
+
+void mpi_ibarrier_(int* comm, int* request, int* ierr);
+void mpi_ibcast_(void* buf, int* count, int* datatype, int* root, int* comm, int* request, int* ierr);
+void mpi_ireduce_(void* sendbuf, void* recvbuf, int* count, int* datatype, int* op, int* root, int* comm, int* request, int* ierr);
+void mpi_iallreduce_(void* sendbuf, void* recvbuf, int* count, int* datatype, int* op, int* comm, int* request, int* ierr);
+void mpi_ireduce_scatter_(void* sendbuf, void* recvbuf, int* recvcounts, int* datatype, int* op, int* comm, int* request, int* ierr);
+void mpi_ireduce_scatter_block_(void* sendbuf, void* recvbuf, int* recvcount, int* datatype, int* op, int* comm, int* request ,
+ int* ierr);
+void mpi_iscatter_(void* sendbuf, int* sendcount, int* sendtype, void* recvbuf, int* recvcount, int* recvtype, int* root,
+ int* comm, int* request, int* ierr);
+void mpi_iscatterv_(void* sendbuf, int* sendcounts, int* displs, int* sendtype, void* recvbuf, int* recvcount,
+ int* recvtype, int* root, int* comm, int* request, int* ierr);
+void mpi_igather_(void* sendbuf, int* sendcount, int* sendtype, void* recvbuf, int* recvcount, int* recvtype, int* root,
+ int* comm, int* request, int* ierr);
+void mpi_igatherv_(void* sendbuf, int* sendcount, int* sendtype, void* recvbuf, int* recvcounts, int* displs,
+ int* recvtype, int* root, int* comm, int* request, int* ierr);
+void mpi_iallgather_(void* sendbuf, int* sendcount, int* sendtype, void* recvbuf, int* recvcount, int* recvtype,
+ int* comm, int* request, int* ierr);
+void mpi_iallgatherv_(void* sendbuf, int* sendcount, int* sendtype, void* recvbuf, int* recvcount, int* displs,
+ int* recvtype, int* comm, int* request, int* ierr);
+void mpi_iscan_(void* sendbuf, void* recvbuf, int* count, int* datatype, int* op, int* comm, int* request, int* ierr);
+void mpi_ialltoall_(void* sendbuf, int* sendcount, int* sendtype, void* recvbuf, int* recvcount, int* recvtype,
+ int* comm, int* request, int* ierr);
+void mpi_ialltoallv_(void* sendbuf, int* sendcounts, int* senddisps, int* sendtype, void* recvbuf, int* recvcounts,
+ int* recvdisps, int* recvtype, int* comm, int* request, int* ierr);
+void mpi_ialltoallw_(void* sendbuf, int* sendcnts, int* sdispls, int* sendtypes, void* recvbuf, int* recvcnts,
+ int* rdispls, int* recvtypes, int* comm, int* request, int* ierr);
+void mpi_iexscan_(void* sendbuf, void* recvbuf, int* count, int* datatype, int* op, int* comm, int* request, int* ierr);
+
+void mpi_type_size_(int* datatype, int* size, int* ierr);
void mpi_get_processor_name_(char* name, int* resultlen, int* ierr);
void mpi_test_(int* request, int* flag, MPI_Status* status, int* ierr);
void mpi_testall_(int* count, int* requests, int* flag, MPI_Status* statuses, int* ierr);
void mpi_get_count_(MPI_Status* status, int* datatype, int* count, int* ierr);
void mpi_type_extent_(int* datatype, MPI_Aint* extent, int* ierr);
-void mpi_attr_get_(int* comm, int* keyval, void* attr_value, int* flag, int* ierr);
+void mpi_attr_get_(int* comm, int* keyval, int* attr_value, int* flag, int* ierr);
void mpi_type_commit_(int* datatype, int* ierr);
void mpi_type_vector_(int* count, int* blocklen, int* stride, int* old_type, int* newtype, int* ierr);
void mpi_type_hvector_(int* count, int* blocklen, MPI_Aint* stride, int* old_type, int* newtype, int* ierr);
void mpi_win_set_info_(int* win, int* info, int* ierr);
void mpi_win_get_info_(int* win, int* info, int* ierr);
void mpi_win_get_group_(int* win, int* group, int* ierr);
-void mpi_win_get_attr_(int* win, int* type_keyval, void* attribute_val, int* flag, int* ierr);
-void mpi_win_set_attr_(int* win, int* type_keyval, void* att, int* ierr);
+void mpi_win_get_attr_(int* win, int* type_keyval, MPI_Aint* attribute_val, int* flag, int* ierr);
+void mpi_win_set_attr_(int* win, int* type_keyval, MPI_Aint* att, int* ierr);
void mpi_win_delete_attr_(int* win, int* comm_keyval, int* ierr);
void mpi_win_create_keyval_(void* copy_fn, void* delete_fn, int* keyval, void* extra_state, int* ierr);
void mpi_win_free_keyval_(int* keyval, int* ierr);
void mpi_win_flush_local_(int* rank, int* win, int* ierr);
void mpi_win_flush_all_(int* win, int* ierr);
void mpi_win_flush_local_all_(int* win, int* ierr);
+void mpi_win_dup_fn_( int* win, int* keyval, int* extrastate, MPI_Aint* valin, MPI_Aint* valout, int* flag, int* ierr );
+void mpi_win_null_copy_fn_( int* win, int* keyval, int* extrastate, MPI_Aint* valin, MPI_Aint* valout, int* flag, int* ierr );
void mpi_info_create_(int* info, int* ierr);
void mpi_info_set_(int* info, char* key, char* value, int* ierr, unsigned int keylen, unsigned int valuelen);
void mpi_info_free_(int* info, int* ierr);
void mpi_type_dup_(int* datatype, int* newdatatype, int* ierr);
void mpi_type_set_name_(int* datatype, char* name, int* ierr, int size);
void mpi_type_get_name_(int* datatype, char* name, int* len, int* ierr);
-void mpi_type_get_attr_(int* type, int* type_keyval, void* attribute_val, int* flag, int* ierr);
-void mpi_type_set_attr_(int* type, int* type_keyval, void* attribute_val, int* ierr);
+void mpi_type_get_attr_(int* type, int* type_keyval, int* attribute_val, int* flag, int* ierr);
+void mpi_type_set_attr_(int* type, int* type_keyval, int* attribute_val, int* ierr);
void mpi_type_delete_attr_(int* type, int* type_keyval, int* ierr);
void mpi_type_create_keyval_(void* copy_fn, void* delete_fn, int* keyval, void* extra_state, int* ierr);
void mpi_type_free_keyval_(int* keyval, int* ierr);
void mpi_group_excl_(int* group, int* n, int* ranks, int* newgroup, int* ierr);
void mpi_group_range_incl_(int* group, int* n, int ranges[][3], int* newgroup, int* ierr);
void mpi_group_range_excl_(int* group, int* n, int ranges[][3], int* newgroup, int* ierr);
-void mpi_comm_get_attr_(int* comm, int* comm_keyval, void* attribute_val, int* flag, int* ierr);
-void mpi_comm_set_attr_(int* comm, int* comm_keyval, void* attribute_val, int* ierr);
+void mpi_comm_get_attr_(int* comm, int* comm_keyval, int* attribute_val, int* flag, int* ierr);
+void mpi_comm_set_attr_(int* comm, int* comm_keyval, int* attribute_val, int* ierr);
void mpi_comm_delete_attr_(int* comm, int* comm_keyval, int* ierr);
void mpi_comm_create_keyval_(void* copy_fn, void* delete_fn, int* keyval, void* extra_state, int* ierr);
void mpi_comm_free_keyval_(int* keyval, int* ierr);
void mpi_testany_(int* count, int* requests, int* index, int* flag, MPI_Status* status, int* ierr);
void mpi_waitsome_(int* incount, int* requests, int* outcount, int* indices, MPI_Status* status, int* ierr);
void mpi_reduce_local_(void* inbuf, void* inoutbuf, int* count, int* datatype, int* op, int* ierr);
-void mpi_reduce_scatter_block_(void* sendbuf, void* recvbuf, int* recvcount, int* datatype, int* op, int* comm,
- int* ierr);
void mpi_pack_size_(int* incount, int* datatype, int* comm, int* size, int* ierr);
void mpi_cart_coords_(int* comm, int* rank, int* maxdims, int* coords, int* ierr);
void mpi_cart_create_(int* comm_old, int* ndims, int* dims, int* periods, int* reorder, int* comm_cart, int* ierr);
MPI_Aint* position, int* ierr);
void mpi_unpack_external_(char* datarep, void* inbuf, MPI_Aint* insize, MPI_Aint* position, void* outbuf, int* outcount,
int* datatype, int* ierr);
-void mpi_type_hindexed_(int* count, int* blocklens, MPI_Aint* indices, int* old_type, int* newtype, int* ierr);
+void mpi_type_hindexed_(int* count, int* blocklens, int* indices, int* old_type, int* newtype, int* ierr);
void mpi_type_create_hindexed_(int* count, int* blocklens, MPI_Aint* indices, int* old_type, int* newtype, int* ierr);
void mpi_type_create_hindexed_block_(int* count, int* blocklength, MPI_Aint* indices, int* old_type, int* newtype,
int* ierr);
void mpi_type_indexed_(int* count, int* blocklens, int* indices, int* old_type, int* newtype, int* ierr);
void mpi_type_create_indexed_(int* count, int* blocklens, int* indices, int* old_type, int* newtype, int* ierr);
void mpi_type_create_indexed_block_(int* count, int* blocklength, int* indices, int* old_type, int* newtype, int* ierr);
-void mpi_type_struct_(int* count, int* blocklens, MPI_Aint* indices, int* old_types, int* newtype, int* ierr);
+void mpi_type_struct_(int* count, int* blocklens, int* indices, int* old_types, int* newtype, int* ierr);
void mpi_type_create_struct_(int* count, int* blocklens, MPI_Aint* indices, int* old_types, int* newtype, int* ierr);
void mpi_ssend_(void* buf, int* count, int* datatype, int* dest, int* tag, int* comm, int* ierr);
void mpi_ssend_init_(void* buf, int* count, int* datatype, int* dest, int* tag, int* comm, int* request, int* ierr);
void mpi_issend_(void* buf, int* count, int* datatype, int* dest, int* tag, int* comm, int* request, int* ierr);
void mpi_probe_(int* source, int* tag, int* comm, MPI_Status* status, int* ierr);
void mpi_attr_delete_(int* comm, int* keyval, int* ierr);
-void mpi_attr_put_(int* comm, int* keyval, void* attr_value, int* ierr);
+void mpi_attr_put_(int* comm, int* keyval, int* attr_value, int* ierr);
void mpi_rsend_init_(void* buf, int* count, int* datatype, int* dest, int* tag, int* comm, int* request, int* ierr);
void mpi_keyval_create_(void* copy_fn, void* delete_fn, int* keyval, void* extra_state, int* ierr);
void mpi_keyval_free_(int* keyval, int* ierr);
void mpi_type_create_subarray_(int* ndims, int* array_of_sizes, int* array_of_subsizes, int* array_of_starts,
int* order, int* oldtype, int* newtype, int* ierr);
void mpi_type_match_size_(int* typeclass, int* size, int* datatype, int* ierr);
-void mpi_alltoallw_(void* sendbuf, int* sendcnts, int* sdispls, int* sendtypes, void* recvbuf, int* recvcnts,
- int* rdispls, int* recvtypes, int* comm, int* ierr);
-void mpi_exscan_(void* sendbuf, void* recvbuf, int* count, int* datatype, int* op, int* comm, int* ierr);
-void mpi_comm_set_name_(int* comm, char* name, int* ierr, int size);
+void mpi_comm_set_name_(int* comm, char* name, int* ierr);
void mpi_comm_dup_with_info_(int* comm, int* info, int* newcomm, int* ierr);
void mpi_comm_split_type_(int* comm, int* split_type, int* key, int* info, int* newcomm, int* ierr);
void mpi_comm_set_info_(int* comm, int* info, int* ierr);
void mpi_file_set_view_(int* fh, long long int* offset, int* etype, int* filetype, char* datarep, int* info, int* ierr);
void mpi_file_read_(int* fh, void* buf, int* count, int* datatype, MPI_Status* status, int* ierr);
void mpi_file_write_(int* fh, void* buf, int* count, int* datatype, MPI_Status* status, int* ierr);
-
+void smpi_init_fortran_types();
} // extern "C"
struct s_smpi_privatization_region_t {
typedef s_smpi_privatization_region_t* smpi_privatization_region_t;
extern XBT_PRIVATE int smpi_loaded_page;
-extern XBT_PRIVATE int smpi_universe_size;
XBT_PRIVATE smpi_privatization_region_t smpi_init_global_memory_segment_process();
/**
*/
#define TOPAGE(addr) (void*)(((unsigned long)(addr) / xbt_pagesize) * xbt_pagesize)
-#if HAVE_PAPI
+/** Used only if PAPI is compiled in, but integrated anyway so that this file does not depend on internal_config.h (to speed builds) */
typedef std::vector<std::pair</* counter name */ std::string, /* counter value */ long long>> papi_counter_t;
-#endif
+struct papi_process_data {
+ papi_counter_t counter_data;
+ int event_set;
+};
+extern std::map</* computation unit name */ std::string, papi_process_data> units2papi_setup;
extern std::unordered_map<std::string, double> location2speedup;
