/* -*- Mode: c-basic-offset:4 ; indent-tabs-mode:nil ; -*- */
/*
* (C) 2011 by Argonne National Laboratory.
* See COPYRIGHT in top-level directory.
*/
#ifndef MPIR_NBC_H_INCLUDED
#define MPIR_NBC_H_INCLUDED
/* This specifies the interface that must be exposed by the ADI in order to
* support MPI-3 non-blocking collectives. MPID_Sched_ routines are all
* permitted to be inlines. They are not permitted to be macros.
*
* Most (currently all) devices will just use the default implementation that
* lives in "src/mpid/common/sched" */
/* The device must supply a typedef for MPID_Sched_t. MPID_Sched_t is a handle
* to the schedule (often a pointer under the hood), not the actual schedule.
* This makes it easy to cheaply pass the schedule between functions. Many
*
* The device must also define a constant (possibly a macro) for an invalid
* schedule: MPID_SCHED_NULL */
/* Context/tag strategy for send/recv ops:
* -------------------------------
*
* Blocking collectives were able to more or less safely separate all
* communication between different collectives by using a fixed tag per
* operation. This prevents some potentially very surprising message matching
* patterns when two different collectives are posted on the same communicator
* in rapid succession. But this strategy probably won't work for NBC because
* multiple operations of any combination of types can be outstanding at the
* same time.
*
* The MPI-3 draft standard says that all collective ops must be collectively
* posted in a consistent order w.r.t. other collective operations, including
* nonblocking collectives. This means that we can just use a counter to assign
* tag values that is incremented at each collective start. We can jump through
* some hoops to make sure that the blocking collective code is left
* undisturbed, but it's cleaner to just update them to use the new counter
* mechanism as well.
*/
/* Open question: should tag allocation be rolled into Sched_start? Keeping it
* separate potentially allows more parallelism in the future, but it also
* pushes more work onto the clients of this interface. */
int MPID_Sched_next_tag(MPID_Comm *comm_ptr, int *tag);
/* the device must provide a typedef for MPID_Sched_t in mpidpre.h */
/* creates a new opaque schedule object and returns a handle to it in (*sp) */
int MPID_Sched_create(MPID_Sched_t *sp);
/* clones orig and returns a handle to the new schedule in (*cloned) */
int MPID_Sched_clone(MPID_Sched_t orig, MPID_Sched_t *cloned);
/* sets (*sp) to MPID_SCHED_NULL and gives you back a request pointer in (*req).
* The caller is giving up ownership of the opaque schedule object.
*
* comm should be the primary (user) communicator with which this collective is
* associated, even if other hidden communicators are used for a subset of the
* operations. It will be used for error handling and similar operations. */
int MPID_Sched_start(MPID_Sched_t *sp, MPID_Comm *comm, int tag, MPID_Request **req);
/* send and recv take a comm ptr to enable hierarchical collectives */
int MPID_Sched_send(const void *buf, MPI_Aint count, MPI_Datatype datatype, int dest, MPID_Comm *comm, MPID_Sched_t s);
int MPID_Sched_recv(void *buf, MPI_Aint count, MPI_Datatype datatype, int src, MPID_Comm *comm, MPID_Sched_t s);
/* just like MPI_Issend, can't complete until the matching recv is posted */
int MPID_Sched_ssend(const void *buf, MPI_Aint count, MPI_Datatype datatype, int dest, MPID_Comm *comm, MPID_Sched_t s);
int MPID_Sched_reduce(const void *inbuf, void *inoutbuf, MPI_Aint count, MPI_Datatype datatype, MPI_Op op, MPID_Sched_t s);
/* packing/unpacking can be accomplished by passing MPI_PACKED as either intype
* or outtype */
int MPID_Sched_copy(const void *inbuf, MPI_Aint incount, MPI_Datatype intype,
void *outbuf, MPI_Aint outcount, MPI_Datatype outtype, MPID_Sched_t s);
/* require that all previously added ops are complete before subsequent ops
* may begin to execute */
int MPID_Sched_barrier(MPID_Sched_t s);
/* A convenience macro for the extremely common case that "mpi_errno" is the
* variable used for tracking error state and MPIR_ERR_POP is needed. This
* declutters the NBC code substantially. */
#define MPID_SCHED_BARRIER(sched_) \
do { \
mpi_errno = MPID_Sched_barrier(sched_); \
if (mpi_errno) MPIR_ERR_POP(mpi_errno); \
} while (0)
/* Defers evaluating (*count) until the entry actually begins to execute. This
* permits algorithms that accumulate/dissipate bytes as rounds progress without
* excessive (re)calculation of counts for/from other processes.
*
* A corresponding _recv_defer function is not currently provided because there
* is no known use case. The recv count is just an upper bound, not an exact
* amount to be received, so an oversized recv is used instead of deferral. */
int MPID_Sched_send_defer(const void *buf, const MPI_Aint *count, MPI_Datatype datatype, int dest, MPID_Comm *comm, MPID_Sched_t s);
/* Just like MPID_Sched_recv except it populates the given status object with
* the received count and error information, much like a normal recv. Often
* useful in conjunction with MPID_Sched_send_defer. */
int MPID_Sched_recv_status(void *buf, MPI_Aint count, MPI_Datatype datatype, int src, MPID_Comm *comm, MPI_Status *status, MPID_Sched_t s);
/* Sched_cb_t funcitons take a comm parameter, the value of which will be the
* comm passed to Sched_start */
/* callback entries must be used judiciously, otherwise they will prevent
* caching opportunities */
typedef int (MPID_Sched_cb_t)(MPID_Comm *comm, int tag, void *state);
typedef int (MPID_Sched_cb2_t)(MPID_Comm *comm, int tag, void *state, void *state2);
/* buffer management, fancy reductions, etc */
int MPID_Sched_cb(MPID_Sched_cb_t *cb_p, void *cb_state, MPID_Sched_t s);
int MPID_Sched_cb2(MPID_Sched_cb2_t *cb_p, void *cb_state, void *cb_state2, MPID_Sched_t s);
/* TODO: develop a caching infrastructure for use by the upper level as well,
* hopefully s.t. uthash can be used somehow */
/* common callback utility functions */
int MPIR_Sched_cb_free_buf(MPID_Comm *comm, int tag, void *state);
/* an upgraded version of MPIU_CHKPMEM_MALLOC/_DECL/_REAP/_COMMIT that adds
* corresponding cleanup callbacks to the given schedule at _COMMIT time */
#define MPIR_SCHED_CHKPMEM_DECL(n_) \
void *(mpir_sched_chkpmem_stk_[n_]) = { NULL }; \
int mpir_sched_chkpmem_stk_sp_=0; \
MPIU_AssertDeclValue(const int mpir_sched_chkpmem_stk_sz_,n_)
#define MPIR_SCHED_CHKPMEM_MALLOC_ORSTMT(pointer_,type_,nbytes_,rc_,name_,stmt_) \
do { \
(pointer_) = (type_)MPIU_Malloc(nbytes_); \
if (pointer_) { \
MPIU_Assert(mpir_sched_chkpmem_stk_sp_ < mpir_sched_chkpmem_stk_sz_); \
mpir_sched_chkpmem_stk_[mpir_sched_chkpmem_stk_sp_++] = (pointer_); \
} else if ((nbytes_) > 0) { \
MPIU_CHKMEM_SETERR((rc_),(nbytes_),(name_)); \
stmt_; \
} \
} while (0)
#define MPIR_SCHED_CHKPMEM_MALLOC(pointer_,type_,nbytes_,rc_,name_) \
MPIR_SCHED_CHKPMEM_MALLOC_ORSTMT(pointer_,type_,nbytes_,rc_,name_,goto fn_fail)
/* just cleanup, don't add anything to the schedule */
#define MPIR_SCHED_CHKPMEM_REAP(sched_) \
do { \
while (mpir_sched_chkpmem_stk_sp_ > 0) { \
MPIU_Free(mpir_sched_chkpmem_stk_[--mpir_sched_chkpmem_stk_sp_]); \
} \
} while (0)
#define MPIR_SCHED_CHKPMEM_COMMIT(sched_) \
do { \
MPID_SCHED_BARRIER(s); \
while (mpir_sched_chkpmem_stk_sp_ > 0) { \
mpi_errno = MPID_Sched_cb(&MPIR_Sched_cb_free_buf, \
(mpir_sched_chkpmem_stk_[--mpir_sched_chkpmem_stk_sp_]), \
(sched_)); \
if (mpi_errno) MPIR_ERR_POP(mpi_errno); \
} \
} while (0)
#endif /* !defined(MPIR_NBC_H_INCLUDED) */