-/* 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_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 /*(=MPI_Init called)*/, FINALIZED };
+enum class SmpiProcessState { UNINITIALIZED, INITIALIZING, INITIALIZED /*(=MPI_Init called)*/, FINALIZING, FINALIZED };
constexpr int COLL_TAG_REDUCE = -112;
constexpr int COLL_TAG_SCATTER = -223;
#define MPI_REQUEST_IGNORED ((MPI_Request*)-100)
/* 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_;
+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_ADDR(addr, val, val2) \
- (((void *)(addr) == (void*) &(val2)) \
- ? (val) : (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_))
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_get_universe_size();
-XBT_PRIVATE void smpi_deployment_register_process(const std::string& instance_id, int rank, simgrid::s4u::Actor* actor);
+XBT_PRIVATE void smpi_deployment_register_process(const std::string& instance_id, int rank,
+ 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 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 double smpi_mpi_wtime();
XBT_PRIVATE void smpi_mpi_init();
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
-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
+XBT_PRIVATE void smpi_init_options_internal(bool called_by_smpi_main);
-XBT_PRIVATE void smpi_switch_data_segment(simgrid::s4u::ActorPtr actor);
+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 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);
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,
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();
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;
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;
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
