-/* Copyright (c) 2007-2021. The SimGrid Team. All rights reserved. */
+/* Copyright (c) 2007-2022. 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. */
#include <xbt/misc.h>
#ifdef __cplusplus
+#include <functional>
#include <vector>
#endif
#define MPI_MAX_PROCESSOR_NAME 100
#define MPI_MAX_NAME_STRING 100
#define MPI_MAX_ERROR_STRING 100
-#define MPI_MAX_DATAREP_STRIN 100
+#define MPI_MAX_DATAREP_STRING 128
#define MPI_MAX_INFO_KEY 100
#define MPI_MAX_INFO_VAL 100
#define MPI_MAX_OBJECT_NAME 100
#define MPI_TAG_LB -6
#define MPI_UNIVERSE_SIZE -7
#define MPI_LASTUSEDCODE -8
+#define MPI_KEYVAL_INVALID -9
#define MPI_MODE_NOSTORE 0x1
#define MPI_MODE_NOPUT 0x2
#define MPI_MODE_NOSUCCEED 0x8
#define MPI_MODE_NOCHECK 0x10
-#define MPI_KEYVAL_INVALID 0
#define MPI_NULL_COPY_FN NULL
#define MPI_NULL_DELETE_FN NULL
#define MPI_ERR_LASTCODE 74
-#define MPI_CXX_BOOL MPI_DATATYPE_NULL
-#define MPI_CXX_FLOAT_COMPLEX MPI_DATATYPE_NULL
-#define MPI_CXX_DOUBLE_COMPLEX MPI_DATATYPE_NULL
-#define MPI_CXX_LONG_DOUBLE_COMPLEX MPI_DATATYPE_NULL
#define MPI_REAL2 MPI_DATATYPE_NULL
#define MPI_COMPLEX4 MPI_DATATYPE_NULL
#define MPI_SEEK_SET 600
#define MPI_SEEK_CUR 602
#define MPI_SEEK_END 604
-#define MPI_MAX_DATAREP_STRING 128
#define MPI_WIN_BASE -1
#define MPI_WIN_SIZE -2
extern SMPI_Datatype smpi_MPI_INTEGER8;
extern SMPI_Datatype smpi_MPI_INTEGER16;
extern SMPI_Datatype smpi_MPI_COUNT;
+extern SMPI_Datatype smpi_MPI_CXX_BOOL;
+extern SMPI_Datatype smpi_MPI_MPI_CXX_FLOAT_COMPLEX;
+extern SMPI_Datatype smpi_MPI_MPI_CXX_DOULE_COMPLEX;
+extern SMPI_Datatype smpi_MPI_MPI_CXX_LONG_DOUBLE_COMPLEX;
#define MPI_DATATYPE_NULL SMPI_PREDEFINED_POINTER(MPI_Datatype, smpi_MPI_DATATYPE_NULL)
#define MPI_CHAR SMPI_PREDEFINED_POINTER(MPI_Datatype, smpi_MPI_CHAR)
#define MPI_INTEGER16 SMPI_PREDEFINED_POINTER(MPI_Datatype, smpi_MPI_INTEGER16)
#define MPI_COUNT SMPI_PREDEFINED_POINTER(MPI_Datatype, smpi_MPI_COUNT)
+#if defined(c_plusplus) || defined(__cplusplus)
+#define MPI_CXX_BOOL SMPI_PREDEFINED_POINTER(MPI_Datatype, smpi_MPI_CXX_BOOL)
+#define MPI_CXX_FLOAT_COMPLEX SMPI_PREDEFINED_POINTER(MPI_Datatype, smpi_MPI_CXX_FLOAT_COMPLEX)
+#define MPI_CXX_DOUBLE_COMPLEX SMPI_PREDEFINED_POINTER(MPI_Datatype, smpi_MPI_CXX_DOUBLE_COMPLEX)
+#define MPI_CXX_LONG_DOUBLE_COMPLEX SMPI_PREDEFINED_POINTER(MPI_Datatype, smpi_MPI_CXX_LONG_DOUBLE_COMPLEX)
+#else
+#define MPI_CXX_BOOL MPI_DATATYPE_NULL
+#define MPI_CXX_FLOAT_COMPLEX MPI_DATATYPE_NULL
+#define MPI_CXX_DOUBLE_COMPLEX MPI_DATATYPE_NULL
+#define MPI_CXX_LONG_DOUBLE_COMPLEX MPI_DATATYPE_NULL
+#endif
+
//defines for fortran compatibility
#if defined(__alpha__) || defined(__sparc64__) || defined(__x86_64__) || defined(__ia64__) || defined(__aarch64__)
#define MPI_INTEGER MPI_INT
typedef SMPI_Op *MPI_Op;
#define MPI_OP_NULL ((MPI_Op)NULL)
-XBT_PUBLIC_DATA MPI_Op MPI_MAX;
-XBT_PUBLIC_DATA MPI_Op MPI_MIN;
-XBT_PUBLIC_DATA MPI_Op MPI_MAXLOC;
-XBT_PUBLIC_DATA MPI_Op MPI_MINLOC;
-XBT_PUBLIC_DATA MPI_Op MPI_SUM;
-XBT_PUBLIC_DATA MPI_Op MPI_PROD;
-XBT_PUBLIC_DATA MPI_Op MPI_LAND;
-XBT_PUBLIC_DATA MPI_Op MPI_LOR;
-XBT_PUBLIC_DATA MPI_Op MPI_LXOR;
-XBT_PUBLIC_DATA MPI_Op MPI_BAND;
-XBT_PUBLIC_DATA MPI_Op MPI_BOR;
-XBT_PUBLIC_DATA MPI_Op MPI_BXOR;
+extern SMPI_Op smpi_MPI_MAX;
+extern SMPI_Op smpi_MPI_MIN;
+extern SMPI_Op smpi_MPI_MAXLOC;
+extern SMPI_Op smpi_MPI_MINLOC;
+extern SMPI_Op smpi_MPI_SUM;
+extern SMPI_Op smpi_MPI_PROD;
+extern SMPI_Op smpi_MPI_LAND;
+extern SMPI_Op smpi_MPI_LOR;
+extern SMPI_Op smpi_MPI_LXOR;
+extern SMPI_Op smpi_MPI_BAND;
+extern SMPI_Op smpi_MPI_BOR;
+extern SMPI_Op smpi_MPI_BXOR;
+extern SMPI_Op smpi_MPI_REPLACE;
+extern SMPI_Op smpi_MPI_NO_OP;
+
+#define MPI_MAX SMPI_PREDEFINED_POINTER(MPI_Op, smpi_MPI_MAX)
+#define MPI_MIN SMPI_PREDEFINED_POINTER(MPI_Op, smpi_MPI_MIN)
+#define MPI_MAXLOC SMPI_PREDEFINED_POINTER(MPI_Op, smpi_MPI_MAXLOC)
+#define MPI_MINLOC SMPI_PREDEFINED_POINTER(MPI_Op, smpi_MPI_MINLOC)
+#define MPI_SUM SMPI_PREDEFINED_POINTER(MPI_Op, smpi_MPI_SUM)
+#define MPI_PROD SMPI_PREDEFINED_POINTER(MPI_Op, smpi_MPI_PROD)
+#define MPI_LAND SMPI_PREDEFINED_POINTER(MPI_Op, smpi_MPI_LAND)
+#define MPI_LOR SMPI_PREDEFINED_POINTER(MPI_Op, smpi_MPI_LOR)
+#define MPI_LXOR SMPI_PREDEFINED_POINTER(MPI_Op, smpi_MPI_LXOR)
+#define MPI_BAND SMPI_PREDEFINED_POINTER(MPI_Op, smpi_MPI_BAND)
+#define MPI_BOR SMPI_PREDEFINED_POINTER(MPI_Op, smpi_MPI_BOR)
+#define MPI_BXOR SMPI_PREDEFINED_POINTER(MPI_Op, smpi_MPI_BXOR)
+
//For accumulate
-XBT_PUBLIC_DATA MPI_Op MPI_REPLACE;
-XBT_PUBLIC_DATA MPI_Op MPI_NO_OP;
+#define MPI_REPLACE SMPI_PREDEFINED_POINTER(MPI_Op, smpi_MPI_REPLACE)
+#define MPI_NO_OP SMPI_PREDEFINED_POINTER(MPI_Op, smpi_MPI_NO_OP)
+
typedef SMPI_Group* MPI_Group;
-#define MPI_GROUP_NULL ((MPI_Group)NULL)
+extern SMPI_Group smpi_MPI_GROUP_EMPTY;
+#define MPI_GROUP_EMPTY SMPI_PREDEFINED_POINTER(MPI_Group, smpi_MPI_GROUP_EMPTY)
-XBT_PUBLIC_DATA MPI_Group MPI_GROUP_EMPTY;
+#define MPI_GROUP_NULL ((MPI_Group)NULL)
typedef SMPI_Comm* MPI_Comm;
typedef SMPI_Errhandler* MPI_Errhandler;
#define MPI_ERRHANDLER_NULL ((MPI_Errhandler)NULL)
-XBT_PUBLIC_DATA MPI_Errhandler MPI_ERRORS_RETURN;
-XBT_PUBLIC_DATA MPI_Errhandler MPI_ERRORS_ARE_FATAL;
+extern SMPI_Errhandler smpi_MPI_ERRORS_RETURN;
+#define MPI_ERRORS_RETURN SMPI_PREDEFINED_POINTER(MPI_Errhandler, smpi_MPI_ERRORS_RETURN)
+extern SMPI_Errhandler smpi_MPI_ERRORS_ARE_FATAL;
+#define MPI_ERRORS_ARE_FATAL SMPI_PREDEFINED_POINTER(MPI_Errhandler, smpi_MPI_ERRORS_ARE_FATAL)
typedef enum SMPI_Combiner_enum{
MPI_COMBINER_NAMED,
const int* weights, MPI_Info info, int reorder, MPI_Comm* comm_dist_graph));
MPI_CALL(XBT_PUBLIC int, MPI_Dist_graph_create_adjacent, (MPI_Comm comm_old, int indegree, const int* sources, const int* sourceweights,
int outdegree, const int* destinations, const int* destweights, MPI_Info info, int reorder, MPI_Comm* comm_dist_graph));
-MPI_CALL(XBT_PUBLIC int, MPI_Dist_graph_neighbors, (MPI_Comm comm, int maxindegree, int* sources, int* sourceweights,
- int maxoutdegree, int* destinations, int* destweights));
+MPI_CALL(XBT_PUBLIC int, MPI_Dist_graph_neighbors,
+ (MPI_Comm comm, int maxindegree, int* sources, int* sourceweights, int maxoutdegree, int* destinations,
+ int* destweights));
MPI_CALL(XBT_PUBLIC int, MPI_Dist_graph_neighbors_count, (MPI_Comm comm, int *indegree, int *outdegree, int *weighted));
MPI_CALL(XBT_PUBLIC int, MPI_Win_test, (MPI_Win win, int* flag));
XBT_PUBLIC void smpi_execute(double duration);
XBT_PUBLIC void smpi_execute_benched(double duration);
+XBT_PUBLIC void smpi_bench_begin();
+XBT_PUBLIC void smpi_bench_end();
+
XBT_PUBLIC unsigned long long smpi_rastro_resolution();
XBT_PUBLIC unsigned long long smpi_rastro_timestamp();
-XBT_PUBLIC void smpi_sample_1(int global, const char* file, int line, int iters, double threshold);
-XBT_PUBLIC int smpi_sample_2(int global, const char* file, int line, int iter_count);
-XBT_PUBLIC void smpi_sample_3(int global, const char* file, int line);
-XBT_PUBLIC int smpi_sample_exit(int global, const char* file, int line, int iter_count);
+XBT_PUBLIC void smpi_sample_1(int global, const char* file, const char* tag, int iters, double threshold);
+XBT_PUBLIC int smpi_sample_2(int global, const char* file, const char* tag, int iter_count);
+XBT_PUBLIC void smpi_sample_3(int global, const char* file, const char* tag);
+XBT_PUBLIC int smpi_sample_exit(int global, const char* file, const char* tag, int iter_count);
/**
* Need a public setter for SMPI copy_callback function, so users can define
* their own while still using default SIMIX_copy_callback for S4U copies.
#define SMPI_ITER_NAME1(line) _XBT_CONCAT(iter_count, line)
#define SMPI_ITER_NAME(line) SMPI_ITER_NAME1(line)
-#define SMPI_SAMPLE_LOOP(loop_init, loop_end, loop_iter, global, iters, thres) \
+#define SMPI_CTAG_NAME1(line) _XBT_CONCAT(ctag, line)
+#define SMPI_CTAG_NAME(line) SMPI_CTAG_NAME1(line)
+
+#define SMPI_SAMPLE_LOOP(loop_init, loop_end, loop_iter, global, iters, thres, tag) \
+ char SMPI_CTAG_NAME(__LINE__) [132]; \
+ snprintf( SMPI_CTAG_NAME(__LINE__), 132, "%s%d", tag, __LINE__); \
int SMPI_ITER_NAME(__LINE__) = 0; \
{ \
loop_init; \
(loop_iter); \
} \
} \
- for (loop_init; (loop_end) ? (smpi_sample_1((global), __FILE__, __LINE__, (iters), (thres)), \
- (smpi_sample_2((global), __FILE__, __LINE__, SMPI_ITER_NAME(__LINE__)))) \
- : smpi_sample_exit((global), __FILE__, __LINE__, SMPI_ITER_NAME(__LINE__)); \
- smpi_sample_3((global), __FILE__, __LINE__), (loop_iter))
+ for ( loop_init; \
+ (loop_end) ? (smpi_sample_1((global), __FILE__, SMPI_CTAG_NAME(__LINE__), (iters), (thres)) \
+ , (smpi_sample_2((global), __FILE__, SMPI_CTAG_NAME(__LINE__), SMPI_ITER_NAME(__LINE__)))) \
+ : smpi_sample_exit((global), __FILE__, SMPI_CTAG_NAME(__LINE__), SMPI_ITER_NAME(__LINE__)); \
+ smpi_sample_3((global), __FILE__, SMPI_CTAG_NAME(__LINE__)), (loop_iter) )
+
#define SMPI_SAMPLE_LOCAL(loop_init, loop_end, loop_iter, iters, thres) \
- SMPI_SAMPLE_LOOP(loop_init, (loop_end), (loop_iter), 0, (iters), (thres))
+ SMPI_SAMPLE_LOOP(loop_init, (loop_end), (loop_iter), 0, (iters), (thres), "")
+#define SMPI_SAMPLE_LOCAL_TAG(loop_init, loop_end, loop_iter, iters, thres, tag) \
+ SMPI_SAMPLE_LOOP(loop_init, (loop_end), (loop_iter), 0, (iters), (thres), tag)
#define SMPI_SAMPLE_GLOBAL(loop_init, loop_end, loop_iter, iters, thres) \
- SMPI_SAMPLE_LOOP(loop_init, (loop_end), (loop_iter), 1, (iters), (thres))
+ SMPI_SAMPLE_LOOP(loop_init, (loop_end), (loop_iter), 1, (iters), (thres), "")
+#define SMPI_SAMPLE_GLOBAL_TAG(loop_init, loop_end, loop_iter, iters, thres, tag) \
+ SMPI_SAMPLE_LOOP(loop_init, (loop_end), (loop_iter), 1, (iters), (thres), tag)
#define SMPI_SAMPLE_DELAY(duration) for(smpi_execute(duration); 0; )
#define SMPI_SAMPLE_FLOPS(flops) for(smpi_execute_flops(flops); 0; )
XBT_PUBLIC void* smpi_shared_malloc(size_t size, const char* file, int line);
/* Trace replay specific stuff */
XBT_PUBLIC void smpi_replay_init(const char* instance_id, int rank, double start_delay_flops); // Only initialization
-XBT_PUBLIC void smpi_replay_main(int rank, const char* trace_filename); // Launch the replay once init is done
+XBT_PUBLIC void smpi_replay_main(int rank, const char* private_trace_filename); // Launch the replay once init is done
XBT_PUBLIC void smpi_replay_run(const char* instance_id, int rank, double start_delay_flops,
- const char* trace_filename); // Both init and start
+ const char* private_trace_filename); // Both init and start
XBT_PUBLIC void SMPI_app_instance_register(const char* name, xbt_main_func_t code, int num_processes);
XBT_PUBLIC void SMPI_init();
SG_END_DECL
-/* C++ declarations for shared_malloc */
+/* C++ declarations for shared_malloc and default copy buffer callback */
#ifdef __cplusplus
XBT_PUBLIC int smpi_is_shared(const void* ptr, std::vector<std::pair<size_t, size_t>>& private_blocks, size_t* offset);
size_t offset, size_t buff_size);
std::vector<std::pair<size_t, size_t>> merge_private_blocks(const std::vector<std::pair<size_t, size_t>>& src,
const std::vector<std::pair<size_t, size_t>>& dst);
+
+/* May be used by S4U simulations to manually initialize SMPI */
+XBT_PUBLIC void smpi_comm_copy_buffer_callback(simgrid::kernel::activity::CommImpl* comm, void* buff,
+ size_t buff_size);
+
+/**
+ * @brief Callback to set cost for SMPI operations (send, recv, isend)
+ *
+ * This callback replaces the configuration parameters smpi/or, smpi/os, smpi/ois.
+ * It offers more flexibility for cost functions.
+ *
+ * @param size Size of message being received/sent
+ * @param source Source host
+ * @param dst Destination host
+ */
+using SmpiOpCostCb = std::function<double(size_t size, simgrid::s4u::Host* source, simgrid::s4u::Host* dst)>;
+/** @brief SMPI functions that accept cost functions */
+enum class SmpiOperation { RECV = 2, SEND = 1, ISEND = 0 };
+/**
+ * @brief Register a cost callback for some SMPI function (MPI_Send, MPI_ISend or MPI_Recv)
+ *
+ * @param op SMPI function
+ * @param cb User's callback
+ */
+XBT_PUBLIC void smpi_register_op_cost_callback(SmpiOperation op, const SmpiOpCostCb& cb);
#endif
-#endif
+#endif
