/* Project Delorean */

#ifndef delorean_H
#define delorean_H


#include "mpi.h"


/********************************************/
/* Helper macros for future args processing */
/********************************************/

#define _G1(Param, Type1, Type2, Type1_Func, Type2_Func) \
  _Generic((Param), \
    Type1: Type1_Func, \
    Type2: Type2_Func)

#define _G2(ParamA, TypeA1, TypeA2, ParamB, TypeB1, TypeB2, \
    TypeA1B1_Func, TypeA1B2_Func, TypeA2B1_Func, TypeA2B2_Func) \
  _Generic( (ParamA), \
    TypeA1: _G1(ParamB, TypeB1, TypeB2, TypeA1B1_Func, TypeA1B2_Func), \
    TypeA2: _G1(ParamB, TypeB1, TypeB2, TypeA2B1_Func, TypeA2B2_Func) )

#define _G3(ParamA, TypeA1, TypeA2, ParamB, TypeB1, TypeB2, \
    ParamC, TypeC1, TypeC2, \
    TypeA1B1C1_Func, TypeA1B1C2_Func, TypeA1B2C1_Func, TypeA1B2C2_Func, \
    TypeA2B1C1_Func, TypeA2B1C2_Func, TypeA2B2C1_Func, TypeA2B2C2_Func) \
  _Generic( (ParamA), \
    TypeA1: _G2(ParamB, TypeB1, TypeB2, ParamC, TypeC1, TypeC2, \
      TypeA1B1C1_Func, TypeA1B1C2_Func, TypeA1B2C1_Func, TypeA1B2C2_Func), \
    TypeA2: _G2(ParamB, TypeB1, TypeB2, ParamC, TypeC1, TypeC2, \
      TypeA2B1C1_Func, TypeA2B1C2_Func, TypeA2B2C1_Func, TypeA2B2C2_Func) )

#define _G4(ParamA, TypeA1, TypeA2, ParamB, TypeB1, TypeB2, \
    ParamC, TypeC1, TypeC2, ParamD, TypeD1, TypeD2, \
    TypeA1B1C1D1_Func, TypeA1B1C1D2_Func, TypeA1B1C2D1_Func, TypeA1B1C2D2_Func, \
    TypeA1B2C1D1_Func, TypeA1B2C1D2_Func, TypeA1B2C2D1_Func, TypeA1B2C2D2_Func, \
    TypeA2B1C1D1_Func, TypeA2B1C1D2_Func, TypeA2B1C2D1_Func, TypeA2B1C2D2_Func, \
    TypeA2B2C1D1_Func, TypeA2B2C1D2_Func, TypeA2B2C2D1_Func, TypeA2B2C2D2_Func) \
  _Generic( (ParamA), \
    TypeA1: _G3(ParamB, TypeB1, TypeB2, ParamC, TypeC1, TypeC2, \
      ParamD, TypeD1, TypeD2, \
      TypeA1B1C1D1_Func, TypeA1B1C1D2_Func, TypeA1B1C2D1_Func, TypeA1B1C2D2_Func, \
      TypeA1B2C1D1_Func, TypeA1B2C2D2_Func, TypeA1B2C2D1_Func, TypeA1B2C2D2_Func), \
    TypeA2: _G3(ParamB, TypeB1, TypeB2, ParamC, TypeC1, TypeC2, \
      ParamD, TypeD1, TypeD2, \
      TypeA2B1C1D1_Func, TypeA2B1C1D2_Func, TypeA2B1C2D1_Func, TypeA2B1C2D2_Func, \
      TypeA2B2C1D1_Func, TypeA2B2C2D2_Func, TypeA2B2C2D1_Func, TypeA2B2C2D2_Func) )


/*********/
/* Types */
/*********/

/* Graph */
typedef struct {
  int id;               // ID of graph
  int rc;               // Refcount for graph
} MPIX_Graph_struct;
typedef MPIX_Graph_struct *MPIX_Graph;

/* Async operation */
typedef struct {
  int id;               // ID of operation
  int rc;               // Refcount for operation
} MPIX_AsyncOp_struct;
typedef MPIX_AsyncOp_struct *MPIX_AsyncOp;

/* "Future" */
typedef struct {
  int id;               // ID of future
  int rc;               // Refcount for future
} MPIX_Future_struct;
typedef MPIX_Future_struct *MPIX_Future;


/***********************/
/* Function Prototypes */
/***********************/

/********************/
/* Graph management */
/********************/
int MPIX_Graph_create(MPI_Comm comm, MPI_Info info, MPIX_Graph *graph);
int MPIX_Graph_dup(MPIX_Graph graph, MPI_Info info, MPIX_Graph *newgraph);
int MPIX_Graph_free(MPIX_Graph *graph);

/***********************/
/* Async Op Management */
/***********************/
int MPIX_Graph_join(MPIX_Graph graph, int count, MPIX_AsyncOp array_of_dep_ops[],
  MPIX_AsyncOp *op);
int MPIX_Graph_wait(MPIX_Graph graph, MPIX_AsyncOp op, int ns,
  int *err_occured);

/*************************/
/* Async User Operations */
/*************************/
//typedef void MPIX_Graph_User_function(void *invec, void *inoutvec,
//             MPI_Count *len, MPI_Datatype *datatype);
//int MPIX_Graph_OpCreate(MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op,
//MPIX_Graph_User_function(void *invec, void *inoutvec,
//
//  MPIX_AsyncOp *op);

/*************************/
/* Async Graph Execution */
/*************************/
int MPIX_Graph_pause(MPIX_Graph graph);

#ifdef OPTION1
/************************************************************/
/* Async versions of existing MPI point-to-point operations */
/************************************************************/

/* MPI_Send: */
int MPIX_Graph_send_future_future(MPIX_Future fbuf, MPIX_Future fcount,
    MPI_Datatype datatype, int dest, int tag, MPI_Comm comm, MPIX_Graph graph,
    MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_send_future_count(MPIX_Future fbuf, MPI_Count count,
    MPI_Datatype datatype, int dest, int tag, MPI_Comm comm, MPIX_Graph graph,
    MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_send_voidp_future(const void *buf, MPIX_Future fcount,
    MPI_Datatype datatype, int dest, int tag, MPI_Comm comm, MPIX_Graph graph,
    MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_send_voidp_count(const void *buf, MPI_Count count,
    MPI_Datatype datatype, int dest, int tag, MPI_Comm comm, MPIX_Graph graph,
    MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
#define MPIX_Graph_send(buf, count, datatype, dest, tag, comm, graph, op, dep_op) \
  _G2(buf, MPIX_Future, const void *, \
      count, MPIX_Future, MPI_Count, \
        MPIX_Graph_send_future_future, MPIX_Graph_send_future_count, \
        MPIX_Graph_send_voidp_future, MPIX_Graph_send_voidp_count) \
     (buf, count, datatype, dest, tag, comm, graph, op, dep_op)

/* MPI_Recv: */

/* MPI_Sendrecv: (actually MPI_Sendrecv_c) */
int MPIX_Graph_Sendrecv_future_future_future_future(MPIX_Future fsendbuf, MPIX_Future fsendcount,
    MPI_Datatype sendtype, int dest, int sendtag, MPIX_Future frecvbuf,
    MPIX_Future frecvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_future_future_future_count(MPIX_Future fsendbuf, MPIX_Future fsendcount,
    MPI_Datatype sendtype, int dest, int sendtag, MPIX_Future frecvbuf,
    MPI_Count recvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_future_future_voidp_future(MPIX_Future fsendbuf, MPIX_Future fsendcount,
    MPI_Datatype sendtype, int dest, int sendtag, void *recvbuf,
    MPIX_Future frecvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_future_future_voidp_count(MPIX_Future fsendbuf, MPIX_Future fsendcount,
    MPI_Datatype sendtype, int dest, int sendtag, void *recvbuf,
    MPI_Count recvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_future_count_future_future(MPIX_Future fsendbuf, MPI_Count sendcount,
    MPI_Datatype sendtype, int dest, int sendtag, MPIX_Future frecvbuf,
    MPIX_Future frecvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_future_count_future_count(MPIX_Future fsendbuf, MPI_Count sendcount,
    MPI_Datatype sendtype, int dest, int sendtag, MPIX_Future frecvbuf,
    MPI_Count recvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_future_count_voidp_future(MPIX_Future fsendbuf, MPI_Count sendcount,
    MPI_Datatype sendtype, int dest, int sendtag, void *recvbuf,
    MPIX_Future frecvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_future_count_voidp_count(MPIX_Future fsendbuf, MPI_Count sendcount,
    MPI_Datatype sendtype, int dest, int sendtag, void *recvbuf,
    MPI_Count recvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_voidp_future_future_future(const void *sendbuf, MPIX_Future fsendcount,
    MPI_Datatype sendtype, int dest, int sendtag, MPIX_Future frecvbuf,
    MPIX_Future frecvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_voidp_future_future_count(const void *sendbuf, MPIX_Future fsendcount,
    MPI_Datatype sendtype, int dest, int sendtag, MPIX_Future frecvbuf,
    MPI_Count recvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_voidp_future_voidp_future(const void *sendbuf, MPIX_Future fsendcount,
    MPI_Datatype sendtype, int dest, int sendtag, void *recvbuf,
    MPIX_Future frecvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_voidp_future_voidp_count(const void *sendbuf, MPIX_Future fsendcount,
    MPI_Datatype sendtype, int dest, int sendtag, void *recvbuf,
    MPI_Count recvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_voidp_count_future_future(const void *sendbuf, MPI_Count sendcount,
    MPI_Datatype sendtype, int dest, int sendtag, MPIX_Future frecvbuf,
    MPIX_Future frecvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_voidp_count_future_count(const void *sendbuf, MPI_Count sendcount,
    MPI_Datatype sendtype, int dest, int sendtag, MPIX_Future frecvbuf,
    MPI_Count recvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_voidp_count_voidp_future(const void *sendbuf, MPI_Count sendcount,
    MPI_Datatype sendtype, int dest, int sendtag, void *recvbuf,
    MPIX_Future frecvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
int MPIX_Graph_Sendrecv_voidp_count_voidp_count(const void *sendbuf, MPI_Count sendcount,
    MPI_Datatype sendtype, int dest, int sendtag, void *recvbuf,
    MPI_Count recvcount, MPI_Datatype recvtype, int source, int recvtag,
    MPI_Comm comm, MPI_Status *status, MPIX_Graph graph, MPIX_AsyncOp *op,
    MPIX_AsyncOp dep_op);
#define MPIX_Graph_Sendrecv(sendbuf, sendcount, sendtype, dst, sendtag, \
    recvbuf, recvcount, recvtype, src, recvtag, comm, status, graph, op, dep_op) \
  _G4(sendbuf, MPIX_Future, const void *, \
      sendcount, MPIX_Future, MPI_Count, \
      recvbuf, MPIX_Future, void *, \
      recvcount, MPIX_Future, MPI_Count, \
        MPIX_Graph_Sendrecv_future_future_future_future, MPIX_Graph_Sendrecv_future_future_future_count, \
        MPIX_Graph_Sendrecv_future_future_voidp_future, MPIX_Graph_Sendrecv_future_future_voidp_count, \
        MPIX_Graph_Sendrecv_future_count_future_future, MPIX_Graph_Sendrecv_future_count_future_count, \
        MPIX_Graph_Sendrecv_future_count_voidp_future, MPIX_Graph_Sendrecv_future_count_voidp_count, \
        MPIX_Graph_Sendrecv_voidp_future_future_future, MPIX_Graph_Sendrecv_voidp_future_future_count, \
        MPIX_Graph_Sendrecv_voidp_future_voidp_future, MPIX_Graph_Sendrecv_voidp_future_voidp_count, \
        MPIX_Graph_Sendrecv_voidp_count_future_future, MPIX_Graph_Sendrecv_voidp_count_future_count, \
        MPIX_Graph_Sendrecv_voidp_count_voidp_future, MPIX_Graph_Sendrecv_voidp_count_voidp_count) \
     (sendbuf, sendcount, sendtype, dst, sendtag, \
      recvbuf, recvcount, recvtype, src, recvtag, comm, status, graph, op, dep_op)
//int MPI_Sendrecv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
//             int dest, int sendtag, void *recvbuf, int recvcount,
//             MPI_Datatype recvtype, int source, int recvtag, MPI_Comm comm,
//             MPI_Status *status);

/* ... */


/********************************************************/
/* Async versions of existing MPI collective operations */
/********************************************************/

/* MPI_Bcast (actually MPI_Bcast_c) */
int MPIX_Graph_Bcast_future_future(MPIX_Future fbuf, MPIX_Future fcount,
    MPI_Datatype datatype, int root, MPI_Comm comm,
    MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Bcast_future_count(MPIX_Future fbuf, MPI_Count count,
    MPI_Datatype datatype, int root, MPI_Comm comm,
    MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Bcast_voidp_future(void *buffer, MPIX_Future fcount,
    MPI_Datatype datatype, int root, MPI_Comm comm,
    MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Bcast_voidp_count(void *buffer, MPI_Count count,
    MPI_Datatype datatype, int root, MPI_Comm comm,
    MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
#define MPIX_Graph_Bcast(buf, count, datatype, root, comm, graph, op, dep_op) \
  _G2(buf, MPIX_Future, void *, \
      count, MPIX_Future, MPI_Count, \
        MPIX_Graph_Bcast_future_future, MPIX_Graph_Bcast_future_count, \
        MPIX_Graph_Bcast_voidp_future, MPIX_Graph_Bcast_voidp_count) \
     (buf, count, datatype, root, comm, graph, op, dep_op)

/* ... */

/***************************************************************/
/* Async versions of memory allocation and movement operations */
/***************************************************************/

/* MPI_Alloc_mem: */
int MPIX_Graph_Alloc_mem_future(MPIX_Future fsize, MPI_Info info, MPIX_Future *fbuf,
    MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Alloc_mem_aint(MPI_Aint size, MPI_Info info, MPIX_Future *fbuf,
    MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
#define MPIX_Graph_Alloc_mem(size, info, buf, graph, op, dep_op) \
  _G1(size, MPIX_Future, MPI_Aint, MPIX_Graph_Alloc_mem_future, MPIX_Graph_Alloc_mem_aint) \
    (size, info, buf, graph, op, dep_op)

/* MPI_Free_mem: */
int MPIX_Graph_Free_mem_future(MPIX_Future fbuf,
    MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Free_mem_voidp(void *buf,
    MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
#define MPIX_Graph_Free_mem(buf, graph, op, dep_op) \
  _G1(buf, MPIX_Future, void *, MPIX_Graph_Free_mem_future, MPIX_Graph_Free_mem_voidp) \
    (buf, graph, op, dep_op)

/* MPI_Copy_mem: */
int MPIX_Graph_Copy_mem_future_future_future(MPIX_Future fdst, MPIX_Future fsrc,
    MPIX_Future fsize, MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Copy_mem_future_future_sizet(MPIX_Future fdst, MPIX_Future fsrc,
    size_t size, MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Copy_mem_future_voidp_future(MPIX_Future fdst, const void *src,
    MPIX_Future fsize, MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Copy_mem_voidp_future_future(void *dst, MPIX_Future fsrc,
    MPIX_Future fsize, MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Copy_mem_future_voidp_sizet(MPIX_Future fdst, const void *src,
    size_t size, MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Copy_mem_voidp_future_sizet(void *dst, MPIX_Future fsrc,
    size_t size, MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Copy_mem_voidp_voidp_future(void *dst, const void *src,
    MPIX_Future fsize, MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Copy_mem_voidp_voidp_sizet(void *dst, const void *src,
    size_t size, MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
#define MPIX_Graph_Copy_mem(dst, src, size, graph, op, dep_op) \
  _G3(dst, MPIX_Future, void *, src, MPIX_Future, const void *, \
    size, MPIX_Future, size_t, \
    MPIX_Graph_Copy_mem_future_future_future, MPIX_Graph_Copy_mem_future_future_sizet, \
    MPIX_Graph_Copy_mem_future_voidp_future, MPIX_Graph_Copy_mem_future_voidp_sizet, \
    MPIX_Graph_Copy_mem_voidp_future_future, MPIX_Graph_Copy_mem_voidp_future_sizet, \
    MPIX_Graph_Copy_mem_voidp_voidp_future, MPIX_Graph_Copy_mem_voidp_voidp_sizet) \
  (dst, src, size, graph, op, dep_op)

/* MPI_Set_mem: */
int MPIX_Graph_Set_mem_future_future(MPIX_Future fbuf, int val, MPIX_Future flen,
    MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Set_mem_future_sizet(MPIX_Future fbuf, int val, size_t len,
    MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Set_mem_voidp_future(void *buf, int val, MPIX_Future flen,
    MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
int MPIX_Graph_Set_mem_voidp_sizet(void *buf, int val, size_t len,
    MPIX_Graph graph, MPIX_AsyncOp *op, MPIX_AsyncOp dep_op);
#define MPIX_Graph_Set_mem(buf, val, len, graph, op, dep_op) \
  _G2(buf, MPIX_Future, void *, len, MPIX_Future, size_t, \
    MPIX_Graph_Set_mem_future_future, MPIX_Graph_Set_mem_future_sizet, \
    MPIX_Graph_Set_mem_voidp_future, MPIX_Graph_Set_mem_voidp_sizet) \
  (buf, val, len, graph, op, dep_op)

#endif /* OPTION1 */

#ifdef OPTION2
#endif /* OPTION2 */

#endif /* delorean_H */