-/* Copyright (c) 2007-2019. 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. */
#ifndef SMPI_PRIVATE_HPP
#define SMPI_PRIVATE_HPP
-#include "simgrid/s4u/Barrier.hpp"
#include "smpi/smpi.h"
#include "smpi/smpi_helpers_internal.h"
+#include "smpi_utils.hpp"
#include "src/instr/instr_smpi.hpp"
#include <unordered_map>
#include <vector>
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_PROBE = 0x10;
constexpr unsigned MPI_REQ_ISEND = 0x20;
constexpr unsigned MPI_REQ_SSEND = 0x40;
constexpr unsigned MPI_REQ_PREPARED = 0x80;
constexpr unsigned MPI_REQ_ACCUMULATE = 0x400;
constexpr unsigned MPI_REQ_GENERALIZED = 0x800;
constexpr unsigned MPI_REQ_COMPLETE = 0x1000;
+constexpr unsigned MPI_REQ_BSEND = 0x2000;
+constexpr unsigned MPI_REQ_MATCHED = 0x4000;
+constexpr unsigned MPI_REQ_CANCELLED = 0x8000;
+constexpr unsigned MPI_REQ_NBC = 0x10000;
-enum class SmpiProcessState { UNINITIALIZED, INITIALIZING, INITIALIZED, FINALIZED };
+enum class SmpiProcessState { UNINITIALIZED, INITIALIZING, INITIALIZED /*(=MPI_Init called)*/, FINALIZING, FINALIZED };
constexpr int COLL_TAG_REDUCE = -112;
constexpr int COLL_TAG_SCATTER = -223;
constexpr int COLL_TAG_BARRIER = -778;
constexpr int COLL_TAG_REDUCE_SCATTER = -889;
constexpr int COLL_TAG_ALLTOALLV = -1000;
+constexpr int COLL_TAG_ALLTOALLW = -1020;
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.
constexpr int SMPI_RMA_TAG = -6666;
#define MPI_REQUEST_IGNORED ((MPI_Request*)-100)
+/* Bindings for MPI special values */
+extern XBT_PUBLIC const int mpi_in_place_;
+extern XBT_PUBLIC const int mpi_bottom_;
+extern XBT_PUBLIC const int mpi_status_ignore_;
+extern XBT_PUBLIC const int mpi_statuses_ignore_;
/* Convert between Fortran and C */
-
-#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))
+#define FORT_ADDR(addr, val, val2) (((const void*)(addr) == (const 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_Cart_topology* MPIR_Cart_Topology;
-
-typedef SMPI_Graph_topology* MPIR_Graph_Topology;
-
-typedef SMPI_Dist_Graph_topology* MPIR_Dist_Graph_Topology;
+using MPIR_Cart_Topology = SMPI_Cart_topology*;
+using MPIR_Graph_Topology = SMPI_Graph_topology*;
+using MPIR_Dist_Graph_Topology = SMPI_Dist_Graph_topology*;
XBT_PRIVATE simgrid::smpi::ActorExt* smpi_process();
XBT_PRIVATE simgrid::smpi::ActorExt* smpi_process_remote(simgrid::s4u::ActorPtr actor);
-XBT_PRIVATE int smpi_process_count();
XBT_PRIVATE void smpi_deployment_register_process(const std::string& instance_id, int rank,
- simgrid::s4u::ActorPtr actor);
+ const simgrid::s4u::Actor* actor);
+XBT_PRIVATE void smpi_deployment_startup_barrier(const std::string& instance_id);
+XBT_PRIVATE void smpi_deployment_unregister_process(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 int smpi_deployment_smpirun(const simgrid::s4u::Engine* e, const std::string& hostfile, int np,
+ const std::string& replayfile, int map,
+ const std::vector<const char*>& run_args);
-XBT_PRIVATE void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff,
- size_t buff_size);
+/** @brief Cleanup user's callback structure. Avoid segfault while destroying process */
+XBT_PRIVATE void smpi_cleanup_op_cost_callback();
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();
-XBT_PRIVATE void smpi_global_destroy();
XBT_PRIVATE double smpi_mpi_wtime();
XBT_PRIVATE void smpi_mpi_init();
-// utilities
-extern XBT_PRIVATE double smpi_cpu_threshold;
-extern XBT_PRIVATE double smpi_host_speed;
-extern XBT_PRIVATE char* smpi_data_exe_start; // start of the data+bss segment of the executable
-extern XBT_PRIVATE int smpi_data_exe_size; // size of the data+bss segment of the executable
-
enum class SharedMallocType { NONE, LOCAL, GLOBAL };
-extern XBT_PRIVATE SharedMallocType smpi_cfg_shared_malloc; // Whether to activate shared malloc
+enum class SmpiPrivStrategies { NONE = 0, MMAP = 1, DLOPEN = 2, DEFAULT = DLOPEN };
-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 double smpi_cfg_host_speed();
+XBT_PRIVATE bool smpi_cfg_simulate_computation();
+XBT_PRIVATE SharedMallocType smpi_cfg_shared_malloc();
+XBT_PRIVATE double smpi_cfg_cpu_thresh();
+XBT_PRIVATE SmpiPrivStrategies smpi_cfg_privatization();
+XBT_PRIVATE int smpi_cfg_async_small_thresh();
+XBT_PRIVATE int smpi_cfg_detached_send_thresh();
+XBT_PRIVATE bool smpi_cfg_grow_injected_times();
+XBT_PRIVATE double smpi_cfg_iprobe_cpu_usage();
+XBT_PRIVATE bool smpi_cfg_trace_call_location();
+XBT_PRIVATE bool smpi_cfg_trace_call_use_absolute_path();
+XBT_PRIVATE std::string smpi_cfg_comp_adjustment_file();
+XBT_PRIVATE std::string smpi_cfg_papi_events_file();
+XBT_PRIVATE double smpi_cfg_auto_shared_malloc_thresh();
+XBT_PRIVATE bool smpi_cfg_display_alloc();
+
+// utilities
+XBT_PRIVATE void smpi_init_options_internal(bool called_by_smpi_main);
+
+XBT_PRIVATE bool smpi_switch_data_segment(simgrid::s4u::ActorPtr actor, const void* addr = nullptr);
XBT_PRIVATE void smpi_prepare_global_memory_segment();
XBT_PRIVATE void smpi_backup_global_memory_segment();
XBT_PRIVATE void smpi_destroy_global_memory_segments();
XBT_PRIVATE void smpi_bench_destroy();
-XBT_PRIVATE void smpi_bench_begin();
-XBT_PRIVATE void smpi_bench_end();
XBT_PRIVATE void smpi_shared_destroy();
+XBT_PRIVATE double smpi_adjust_comp_speed();
+
+// This helper class uses RAII to call smpi_bench_end() when an object is built, and have smpi_bench_begin() be called
+// automatically when going out of scope.
+class XBT_PRIVATE SmpiBenchGuard {
+public:
+ SmpiBenchGuard() { smpi_bench_end(); }
+ SmpiBenchGuard(const SmpiBenchGuard&) = delete;
+ SmpiBenchGuard& operator=(const SmpiBenchGuard&) = delete;
+ ~SmpiBenchGuard() { smpi_bench_begin(); }
+};
-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_wait_(int* request, MPI_Status* status, int* ierr);
void mpi_waitany_(int* count, int* requests, int* index, MPI_Status* status, int* ierr);
void mpi_waitall_(int* count, int* requests, MPI_Status* status, int* ierr);
-
+void mpi_free_mem_(void *baseptr, int* ierr);
void mpi_barrier_(int* comm, int* ierr);
void mpi_bcast_(void* buf, int* count, int* datatype, int* root, int* comm, int* ierr);
void mpi_reduce_(void* sendbuf, void* recvbuf, int* count, int* datatype, int* op, int* root, int* comm, int* ierr);
+void mpi_alloc_mem_(int* size, int* info, void *baseptr, 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_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, int* attribute_val, int* flag, int* ierr);
-void mpi_win_set_attr_(int* win, int* type_keyval, int* 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_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);
void mpi_graphdims_get_(int* comm, int* nnodes, int* nedges, int* ierr);
void mpi_topo_test_(int* comm, int* top_type, int* ierr);
void mpi_error_class_(int* errorcode, int* errorclass, int* ierr);
-void mpi_errhandler_create_(void* function, void* errhandler, int* ierr);
-void mpi_errhandler_free_(void* errhandler, int* ierr);
-void mpi_errhandler_get_(int* comm, void* errhandler, int* ierr);
-void mpi_errhandler_set_(int* comm, void* errhandler, int* ierr);
-void mpi_comm_set_errhandler_(int* comm, void* errhandler, int* ierr);
-void mpi_comm_get_errhandler_(int* comm, void* errhandler, int* ierr);
+void mpi_errhandler_create_(void* function, int* errhandler, int* ierr);
+void mpi_errhandler_free_(int* errhandler, int* ierr);
+void mpi_errhandler_get_(int* comm, int* errhandler, int* ierr);
+void mpi_errhandler_set_(int* comm, int* errhandler, int* ierr);
+void mpi_comm_set_errhandler_(int* comm, int* errhandler, int* ierr);
+void mpi_comm_get_errhandler_(int* comm, int* errhandler, int* ierr);
void mpi_type_contiguous_(int* count, int* old_type, int* newtype, int* ierr);
void mpi_cancel_(int* request, int* ierr);
void mpi_buffer_attach_(void* buffer, int* size, 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_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_comm_get_info_(int* comm, int* info, int* ierr);
void mpi_info_get_(int* info, char* key, int* valuelen, char* value, int* flag, int* ierr, unsigned int keylen);
-void mpi_comm_create_errhandler_(void* function, void* errhandler, int* ierr);
+void mpi_comm_create_errhandler_(void* function, int* errhandler, int* ierr);
void mpi_add_error_class_(int* errorclass, int* ierr);
void mpi_add_error_code_(int* errorclass, int* errorcode, int* ierr);
void mpi_add_error_string_(int* errorcode, char* string, int* ierr);
void mpi_file_close_(int* file, int* ierr);
void mpi_file_delete_(char* filename, int* info, int* ierr);
void mpi_file_open_(int* comm, char* filename, int* amode, int* info, int* fh, 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_seek_(int* fh, MPI_Offset* offset, int* whence, int* ierr);
+void mpi_file_seek_shared_(int* fh, MPI_Offset* offset, int* whence, int* ierr);
+void mpi_file_get_position_(int* fh, MPI_Offset* offset, int* ierr);
+void mpi_file_get_position_shared_(int* fh, MPI_Offset* offset, int* ierr);
+void mpi_file_set_size_(int* fh, MPI_Offset* size, int* ierr);
+void mpi_file_get_size_(int* fh, MPI_Offset* sier, int* ierr);
+void mpi_file_set_view_(int* fh, MPI_Offset* offset, int* etype, int* filetype, char* datarep, int* info, int* ierr);
+void mpi_file_get_view_(int* fh, MPI_Offset* disp, int* etype, int* filetype, char *datarep, int* ierr);
void mpi_file_read_(int* fh, void* buf, int* count, int* datatype, MPI_Status* status, int* ierr);
+void mpi_file_read_shared_(int* fh, void *buf, int* count, int* datatype, MPI_Status* status, int* ierr);
+void mpi_file_read_all_(int* fh, void *buf, int* count, int* datatype, MPI_Status* status, int* ierr);
+void mpi_file_read_ordered_(int* fh, void *buf, int* count, int* datatype, MPI_Status* status, int* ierr);
+void mpi_file_read_at_(int* fh, MPI_Offset* offset, void *buf, int* count, int* datatype, MPI_Status* status, int* ierr);
+void mpi_file_read_at_all_(int* fh, MPI_Offset* offset, 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 mpi_file_write_all_(int* fh, void* buf, int* count, int* datatype, MPI_Status* status, int* ierr);
+void mpi_file_write_shared_(int* fh, void *buf, int* count, int* datatype, MPI_Status* status, int* ierr);
+void mpi_file_write_ordered_(int* fh, void *buf, int* count, int* datatype, MPI_Status* status, int* ierr);
+void mpi_file_write_at_(int* fh, MPI_Offset* offset, void *buf, int* count, int* datatype, MPI_Status* status, int* ierr);
+void mpi_file_write_at_all_(int* fh, MPI_Offset* offset, void *buf, int* count, int* datatype, MPI_Status* status, int* ierr);
void smpi_init_fortran_types();
+void smpi_execute_flops_(double* flops);
+void smpi_execute_flops_benched_(double* flops);
+void smpi_execute_(double* duration);
+void smpi_execute_benched_(double* duration);
} // extern "C"
+XBT_PRIVATE int smpi_temp_shm_get();
+XBT_PRIVATE void* smpi_temp_shm_mmap(int fd, size_t size);
+
struct s_smpi_privatization_region_t {
void* address;
int file_descriptor;
};
-typedef s_smpi_privatization_region_t* smpi_privatization_region_t;
+using smpi_privatization_region_t = s_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)
/** 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;
+using papi_counter_t = std::vector<std::pair</* counter name */ std::string, /* counter value */ long long>>;
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::map</* computation unit name */ std::string, papi_process_data, std::less<>> units2papi_setup;
extern std::unordered_map<std::string, double> location2speedup;
/** @brief Returns the last call location (filename, linenumber). Process-specific. */
XBT_PUBLIC smpi_trace_call_location_t* smpi_trace_get_call_location();
-enum class SmpiPrivStrategies { NONE = 0, MMAP = 1, DLOPEN = 2, DEFAULT = DLOPEN };
-
-extern XBT_PRIVATE SmpiPrivStrategies smpi_privatize_global_variables;
+XBT_PRIVATE void private_execute_flops(double flops);
+
+#define CHECK_ARGS(test, errcode,...)\
+ {\
+ if (test) {\
+ int error_code_ = (errcode);\
+ if (error_code_ != MPI_SUCCESS){\
+ XBT_WARN(__VA_ARGS__);\
+ }\
+ return error_code_;\
+ }\
+ }
+
+#define CHECK_INIT \
+ { \
+ int init_flag = 0; \
+ PMPI_Initialized(&init_flag); \
+ CHECK_ARGS(not init_flag, MPI_ERR_OTHER, "%s: MPI_Init was not called !", __func__) \
+ PMPI_Finalized(&init_flag); \
+ CHECK_ARGS(init_flag, MPI_ERR_OTHER, "%s: MPI_Finalize was already called !", __func__) \
+ }
+
+#define CHECK_VAL(num, val, err, value)\
+ CHECK_ARGS((value) == (val), (err),\
+ "%s: param %d %s cannot be %s", __func__, (num), _XBT_STRINGIFY(value), _XBT_STRINGIFY(val))
+
+#define CHECK_NULL(num,err,buf)\
+ CHECK_ARGS((buf) == nullptr, (err),\
+ "%s: param %d %s cannot be NULL", __func__, (num), _XBT_STRINGIFY(buf))
+
+#define CHECK_MPI_NULL(num, val, err, ptr)\
+ {\
+ CHECK_ARGS((ptr) == (val), (err),\
+ "%s: param %d %s cannot be %s", __func__, (num), _XBT_STRINGIFY(ptr), _XBT_STRINGIFY(val))\
+ CHECK_NULL(num, err, ptr)\
+ }
+
+#define CHECK_NEGATIVE(num, err, val)\
+ CHECK_ARGS((val) < 0, (err),\
+ "%s: param %d %s cannot be negative", __func__, (num), _XBT_STRINGIFY(val))
+
+#define CHECK_NEGATIVE_OR_ZERO(num, err, val)\
+ CHECK_ARGS((val) <= 0, (err),\
+ "%s: param %d %s cannot be negative or 0", __func__, (num), _XBT_STRINGIFY(val))
+
+#define CHECK_COMM2(num, comm)\
+ CHECK_MPI_NULL((num), MPI_COMM_NULL, MPI_ERR_COMM, (comm))
+
+#define CHECK_COLLECTIVE(comm, call) \
+ { \
+ if (_smpi_cfg_pedantic) { \
+ std::string call_string = (call); \
+ CHECK_ARGS((simgrid::smpi::utils::check_collectives_ordering((comm), call_string) != MPI_SUCCESS), \
+ MPI_ERR_OTHER, "%s: collective mismatch", call_string.c_str()) \
+ } \
+ }
+
+#define CHECK_DELETED(num, err, obj)\
+ CHECK_ARGS((obj)->deleted(), (err), "%s: param %d %s has already been freed", __func__, (num),\
+ _XBT_STRINGIFY(obj))
+
+#define CHECK_COMM(num)\
+ {\
+ CHECK_INIT\
+ CHECK_COMM2((num), comm)\
+ CHECK_DELETED((num), MPI_ERR_COMM, comm)\
+ simgrid::smpi::utils::set_current_handle(comm);\
+ }
+
+#define CHECK_REQUEST(num)\
+ CHECK_ARGS(request == nullptr, MPI_ERR_REQUEST,\
+ "%s: param %d request cannot be NULL",__func__, (num));
+
+#define CHECK_REQUEST_VALID(num)\
+ {\
+ CHECK_REQUEST(num)\
+ if(request!=MPI_REQUEST_IGNORED){\
+ CHECK_DELETED((num), MPI_ERR_REQUEST, *request)\
+ simgrid::smpi::utils::set_current_handle(*request);\
+ }\
+ }
+#define SET_BUF1(buf)\
+ simgrid::smpi::utils::set_current_buffer(1, _XBT_STRINGIFY(buf), buf);
+#define SET_BUF2(buf)\
+ simgrid::smpi::utils::set_current_buffer(2, _XBT_STRINGIFY(buf), buf);
+
+#define CHECK_BUFFER2(num,buf,count)\
+ CHECK_ARGS((buf) == nullptr && (count) > 0, MPI_ERR_BUFFER,\
+ "%s: param %d %s cannot be NULL if %s > 0",__func__, (num), _XBT_STRINGIFY(buf), _XBT_STRINGIFY(count))\
+
+#define CHECK_BUFFER(num,buf,count,datatype)\
+ {\
+ CHECK_BUFFER2(num,buf,count)\
+ CHECK_ARGS( simgrid::smpi::utils::get_buffer_size(buf) < (size_t)(count*datatype->get_extent()), MPI_ERR_BUFFER,\
+ "%s: param %d message size %zd exceeds buffer %s size %zu",__func__, (num), count*datatype->get_extent(), _XBT_STRINGIFY(buf), simgrid::smpi::utils::get_buffer_size(buf))\
+ }
+
+#define CHECK_COUNT(num, count)\
+ CHECK_NEGATIVE((num), MPI_ERR_COUNT, (count))
+
+#define CHECK_TYPE(num, datatype)\
+ {\
+ CHECK_MPI_NULL((num), MPI_DATATYPE_NULL, MPI_ERR_TYPE, (datatype))\
+ CHECK_ARGS((not (datatype)->is_valid()), MPI_ERR_TYPE,\
+ "%s: param %d %s is invalid", __func__, (num), _XBT_STRINGIFY(datatype));\
+ CHECK_DELETED((num), MPI_ERR_TYPE, datatype)\
+ if (not datatype->is_basic())\
+ simgrid::smpi::utils::set_current_handle(datatype);\
+ }
+
+#define CHECK_OP(num, op, type)\
+ {\
+ CHECK_MPI_NULL((num), MPI_OP_NULL, MPI_ERR_OP, (op))\
+ CHECK_ARGS((op == MPI_REPLACE || op == MPI_NO_OP), MPI_ERR_OP,\
+ "%s: param %d op %s cannot be used in non RMA calls", __func__, (num), _XBT_STRINGIFY(op));\
+ CHECK_DELETED((num), MPI_ERR_OP, op)\
+ if (not op->is_predefined())\
+ simgrid::smpi::utils::set_current_handle(op);\
+ CHECK_ARGS(((op)->allowed_types() && (((op)->allowed_types() & (type)->flags()) == 0)), MPI_ERR_OP,\
+ "%s: param %d op %s can't be applied to type %s", __func__, (num), _XBT_STRINGIFY(op), type->name().c_str());\
+ }
+
+#define CHECK_ROOT(num)\
+ CHECK_ARGS((root < 0 || root >= comm->size()), MPI_ERR_ROOT,\
+ "%s: param %d root (=%d) cannot be negative or larger than communicator size (=%d)", __func__, (num),\
+ root, comm->size())
+
+#define CHECK_INFO(num,info)\
+ {\
+ CHECK_MPI_NULL((num), MPI_INFO_NULL, MPI_ERR_INFO, (info))\
+ CHECK_DELETED((num), MPI_ERR_INFO, info)\
+ simgrid::smpi::utils::set_current_handle(info);\
+ }
+
+#define CHECK_TAG(num,tag)\
+ CHECK_ARGS(((tag) < 0 && (tag) != MPI_ANY_TAG), MPI_ERR_TAG,\
+ "%s: param %d %s (=%d) cannot be negative", __func__, (num), _XBT_STRINGIFY(tag), (tag))
+
+#define CHECK_FILE(num, fh)\
+ CHECK_MPI_NULL((num), MPI_FILE_NULL, MPI_ERR_FILE, (fh))
+
+#define CHECK_OFFSET(num, offset)\
+ CHECK_NEGATIVE((num), MPI_ERR_DISP, (offset))
+
+#define CHECK_GROUP(num, group)\
+ CHECK_MPI_NULL((num), MPI_GROUP_NULL, MPI_ERR_GROUP, (group))
+
+#define CHECK_WIN(num, win)\
+ {\
+ CHECK_MPI_NULL((num), MPI_WIN_NULL, MPI_ERR_WIN, (win))\
+ simgrid::smpi::utils::set_current_handle(win);\
+ }
+#define CHECK_RANK(num, rank, comm)\
+ CHECK_ARGS(((rank) >= (comm)->size() || (rank) <0), MPI_ERR_RANK,\
+ "%s: param %d %s (=%d) cannot be < 0 or > %d", __func__, (num), _XBT_STRINGIFY(rank),\
+ (rank), (comm)->size() )
+
+#define CHECK_PROC_RMA(num,proc,win)\
+ {\
+ CHECK_VAL((num), MPI_PROC_NULL, MPI_SUCCESS, (proc))\
+ CHECK_RANK(num, proc, (win)->comm())\
+ }
+
+#define CHECK_NOT_IN_PLACE_ROOT(num, buf)\
+ CHECK_ARGS((buf == MPI_IN_PLACE), MPI_ERR_BUFFER,\
+ "%s: param %d %s cannot be MPI_IN_PLACE for rank %d with root %d", __func__, (num), _XBT_STRINGIFY(buf), \
+ rank, root)
+
+#define CHECK_NOT_IN_PLACE(num, buf)\
+ CHECK_ARGS((buf == MPI_IN_PLACE), MPI_ERR_BUFFER,\
+ "%s: param %d %s cannot be MPI_IN_PLACE for rank %d", __func__, (num), _XBT_STRINGIFY(buf), rank)
#endif
