/* Find the requested attribute. If it exists, return either the attribute
entry or the address of the entry, based on whether the request is for
a pointer-valued attribute (C or C++) or an integer-valued attribute
(Fortran, either 77 or 90).
If the attribute has the same type as the request, it is returned as-is.
Otherwise, the address of the attribute is returned.
*/
int MPIR_CommGetAttr( MPI_Comm comm, int comm_keyval, void *attribute_val,
int *flag, MPIR_AttrType outAttrType )
{
static const char FCNAME[] = "MPIR_CommGetAttr";
int mpi_errno = MPI_SUCCESS;
MPID_Comm *comm_ptr = NULL;
static PreDefined_attrs attr_copy; /* Used to provide a copy of the
predefined attributes */
MPID_MPI_STATE_DECL(MPID_STATE_MPIR_COMM_GET_ATTR);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_COMM_GET_ATTR);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
MPIR_ERRTEST_KEYVAL(comm_keyval, MPID_COMM, "communicator", mpi_errno);
# ifdef NEEDS_POINTER_ALIGNMENT_ADJUST
/* A common user error is to pass the address of a 4-byte
int when the address of a pointer (or an address-sized int)
should have been used. We can test for this specific
case. Note that this code assumes sizeof(MPIR_Pint) is
a power of 2. */
if ((MPIR_Pint)attribute_val & (sizeof(MPIR_Pint)-1)) {
MPIU_ERR_SET(mpi_errno,MPI_ERR_ARG,"**attrnotptr");
}
# endif
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif
/* Convert MPI object handles to object pointers */
MPID_Comm_get_ptr( comm, comm_ptr );
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
/* Validate comm_ptr */
MPID_Comm_valid_ptr( comm_ptr, mpi_errno );
/* If comm_ptr is not valid, it will be reset to null */
MPIR_ERRTEST_ARGNULL(attribute_val, "attr_val", mpi_errno);
MPIR_ERRTEST_ARGNULL(flag, "flag", mpi_errno);
if (mpi_errno) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
/* Check for builtin attribute */
/* This code is ok for correct programs, but it would be better
to copy the values from the per-process block and pass the user
a pointer to a copy */
/* Note that if we are called from Fortran, we must return the values,
not the addresses, of these attributes */
if (HANDLE_GET_KIND(comm_keyval) == HANDLE_KIND_BUILTIN) {
int attr_idx = comm_keyval & 0x0000000f;
void **attr_val_p = (void **)attribute_val;
#ifdef HAVE_FORTRAN_BINDING
/* This is an address-sized int instead of a Fortran (MPI_Fint)
integer because, even for the Fortran keyvals, the C interface is
used which stores the result in a pointer (hence we need a
pointer-sized int). Thus we use MPIR_Pint instead of MPI_Fint.
On some 64-bit plaforms, such as Solaris-SPARC, using an MPI_Fint
will cause the value to placed into the high, rather than low,
end of the output value. */
#endif
*flag = 1;
/* FIXME : We could initialize some of these here; only tag_ub is
used in the error checking. */
/*
* The C versions of the attributes return the address of a
* *COPY* of the value (to prevent the user from changing it)
* and the Fortran versions provide the actual value (as an Fint)
*/
attr_copy = MPIR_Process.attrs;
switch (attr_idx) {
case 1: /* TAG_UB */
case 2:
*attr_val_p = &attr_copy.tag_ub;
break;
case 3: /* HOST */
case 4:
*attr_val_p = &attr_copy.host;
break;
case 5: /* IO */
case 6:
*attr_val_p = &attr_copy.io;
break;
case 7: /* WTIME */
case 8:
*attr_val_p = &attr_copy.wtime_is_global;
break;
case 9: /* UNIVERSE_SIZE */
case 10:
/* This is a special case. If universe is not set, then we
attempt to get it from the device. If the device is doesn't
supply a value, then we set the flag accordingly */
if (attr_copy.universe >= 0)
{
*attr_val_p = &attr_copy.universe;
}
else if (attr_copy.universe == MPIR_UNIVERSE_SIZE_NOT_AVAILABLE)
{
*flag = 0;
}
else
{
mpi_errno = MPID_Get_universe_size(&attr_copy.universe);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
attr_copy.universe = MPIR_UNIVERSE_SIZE_NOT_AVAILABLE;
goto fn_fail;
}
/* --END ERROR HANDLING-- */
if (attr_copy.universe >= 0)
{
*attr_val_p = &attr_copy.universe;
}
else
{
attr_copy.universe = MPIR_UNIVERSE_SIZE_NOT_AVAILABLE;
*flag = 0;
}
}
break;
case 11: /* LASTUSEDCODE */
case 12:
*attr_val_p = &attr_copy.lastusedcode;
break;
case 13: /* APPNUM */
case 14:
/* This is another special case. If appnum is negative,
we take that as indicating no value of APPNUM, and set
the flag accordingly */
if (attr_copy.appnum < 0) {
*flag = 0;
}
else {
*attr_val_p = &attr_copy.appnum;
}
break;
}
/* All of the predefined attributes are INTEGER; since we've set
the output value as the pointer to these, we need to dereference
it here. */
if (*flag) {
/* Use the internal pointer-sized-int for systems (e.g., BG/P)
that define MPI_Aint as a different size than MPIR_Pint.
The casts must be as they are:
On the right, the value is a pointer to an int, so to
get the correct value, we need to extract the int.
On the left, the output type is given by the argument
outAttrType - and the cast must match the intended results */
if (outAttrType == MPIR_ATTR_AINT)
*(MPIR_Pint*)attr_val_p = *(int*)*(void **)attr_val_p;
else if (outAttrType == MPIR_ATTR_INT)
*(int*)attr_val_p = *(int *)*(void **)attr_val_p;
}
}
else {
MPID_Attribute *p = comm_ptr->attributes;
/* */
*flag = 0;
while (p) {
if (p->keyval->handle == comm_keyval) {
*flag = 1;
if (outAttrType == MPIR_ATTR_PTR) {
if (p->attrType == MPIR_ATTR_INT) {
/* This is the tricky case: if the system is
bigendian, and we have to return a pointer to
an int, then we may need to point to the
correct location in the word. */
#if defined(WORDS_LITTLEENDIAN) || (SIZEOF_VOID_P == SIZEOF_INT)
*(void**)attribute_val = &(p->value);
#else
int *p_loc = (int *)&(p->value);
#if SIZEOF_VOID_P == 2 * SIZEOF_INT
p_loc++;
#else
#error Expected sizeof(void*) to be either sizeof(int) or 2*sizeof(int)
#endif
*(void **)attribute_val = p_loc;
#endif
}
else if (p->attrType == MPIR_ATTR_AINT) {
*(void**)attribute_val = &(p->value);
}
else {
*(void**)attribute_val = (void *)(MPIR_Pint)(p->value);
}
}
else
*(void**)attribute_val = (void *)(MPIR_Pint)(p->value);
break;
}
p = p->next;
}
}
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_COMM_GET_ATTR);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpir_comm_get_attr",
"**mpir_comm_get_attr %C %d %p %p", comm, comm_keyval, attribute_val, flag);
}
# endif
mpi_errno = MPIR_Err_return_comm( comm_ptr, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Comm_free - Marks the communicator object for deallocation
Input Parameters:
. comm - Communicator to be destroyed (handle)
Notes:
This routine `frees` a communicator. Because the communicator may still
be in use by other MPI routines, the actual communicator storage will not
be freed until all references to this communicator are removed. For most
users, the effect of this routine is the same as if it was in fact freed
at this time of this call.
Null Handles:
The MPI 1.1 specification, in the section on opaque objects, explicitly
disallows freeing a null communicator. The text from the standard is:
.vb
A null handle argument is an erroneous IN argument in MPI calls, unless an
exception is explicitly stated in the text that defines the function. Such
exception is allowed for handles to request objects in Wait and Test calls
(sections Communication Completion and Multiple Completions ). Otherwise, a
null handle can only be passed to a function that allocates a new object and
returns a reference to it in the handle.
.ve
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_COMM
.N MPI_ERR_ARG
@*/
int MPI_Comm_free(MPI_Comm *comm)
{
int mpi_errno = MPI_SUCCESS;
MPID_Comm *comm_ptr = NULL;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_COMM_FREE);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPIU_THREAD_CS_ENTER(ALLFUNC,);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_COMM_FREE);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(*comm, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* Get handles to MPI objects. */
MPID_Comm_get_ptr( *comm, comm_ptr );
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
/* Validate comm_ptr */
MPID_Comm_valid_ptr( comm_ptr, mpi_errno, TRUE );
/* If comm_ptr is not valid, it will be reset to null */
/* Cannot free the predefined communicators */
if (HANDLE_GET_KIND(*comm) == HANDLE_KIND_BUILTIN) {
mpi_errno = MPIR_Err_create_code( MPI_SUCCESS,
MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_COMM,
"**commperm", "**commperm %s",
comm_ptr->name );
}
if (mpi_errno) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPIR_Comm_free_impl(comm_ptr);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
*comm = MPI_COMM_NULL;
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_COMM_FREE);
MPIU_THREAD_CS_EXIT(ALLFUNC,);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_comm_free", "**mpi_comm_free %p", comm);
}
# endif
mpi_errno = MPIR_Err_return_comm( comm_ptr, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Bsend - Basic send with user-provided buffering
Input Parameters:
+ buf - initial address of send buffer (choice)
. count - number of elements in send buffer (nonnegative integer)
. datatype - datatype of each send buffer element (handle)
. dest - rank of destination (integer)
. tag - message tag (integer)
- comm - communicator (handle)
Notes:
This send is provided as a convenience function; it allows the user to
send messages without worring about where they are buffered (because the
user `must` have provided buffer space with 'MPI_Buffer_attach').
In deciding how much buffer space to allocate, remember that the buffer space
is not available for reuse by subsequent 'MPI_Bsend's unless you are certain
that the message
has been received (not just that it should have been received). For example,
this code does not allocate enough buffer space
.vb
MPI_Buffer_attach( b, n*sizeof(double) + MPI_BSEND_OVERHEAD );
for (i=0; i<m; i++) {
MPI_Bsend( buf, n, MPI_DOUBLE, ... );
}
.ve
because only enough buffer space is provided for a single send, and the
loop may start a second 'MPI_Bsend' before the first is done making use of the
buffer.
In C, you can
force the messages to be delivered by
.vb
MPI_Buffer_detach( &b, &n );
MPI_Buffer_attach( b, n );
.ve
(The 'MPI_Buffer_detach' will not complete until all buffered messages are
delivered.)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_COMM
.N MPI_ERR_COUNT
.N MPI_ERR_TYPE
.N MPI_ERR_RANK
.N MPI_ERR_TAG
.seealso: MPI_Buffer_attach, MPI_Ibsend, MPI_Bsend_init
@*/
int MPI_Bsend(const void *buf, int count, MPI_Datatype datatype, int dest, int tag,
MPI_Comm comm)
{
static const char FCNAME[] = "MPI_Bsend";
int mpi_errno = MPI_SUCCESS;
MPID_Comm *comm_ptr = NULL;
MPID_Request *request_ptr = NULL;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_BSEND);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPID_MPI_PT2PT_FUNC_ENTER_FRONT(MPID_STATE_MPI_BSEND);
/* Validate handle parameters needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* Convert MPI object handles to object pointers */
MPID_Comm_get_ptr( comm, comm_ptr );
/* Validate object pointers if error checking is enabled */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COUNT(count,mpi_errno);
/* Validate comm_ptr */
MPID_Comm_valid_ptr( comm_ptr, mpi_errno, FALSE );
if (mpi_errno) goto fn_fail;
/* If comm_ptr is not valid, it will be reset to null */
if (comm_ptr) {
MPIR_ERRTEST_SEND_TAG(tag,mpi_errno);
MPIR_ERRTEST_SEND_RANK(comm_ptr,dest,mpi_errno)
}
/* Validate datatype handle */
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
/* Validate datatype object */
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN)
{
MPID_Datatype *datatype_ptr = NULL;
MPID_Datatype_get_ptr(datatype, datatype_ptr);
MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno);
if (mpi_errno) goto fn_fail;
MPID_Datatype_committed_ptr(datatype_ptr, mpi_errno);
if (mpi_errno) goto fn_fail;
}
/* Validate buffer */
MPIR_ERRTEST_USERBUFFER(buf,count,datatype,mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
# ifdef MPID_HAS_TBSEND
{
mpi_errno = MPID_tBsend( buf, count, datatype, dest, tag, comm_ptr, 0 );
if (mpi_errno == MPI_SUCCESS)
{
goto fn_exit;
}
/* FIXME: Check for MPID_WOULD_BLOCK? */
}
# endif
mpi_errno = MPIR_Bsend_isend( buf, count, datatype, dest, tag, comm_ptr,
BSEND, &request_ptr );
/* Note that we can ignore the request_ptr because it is handled internally
by the bsend util routines */
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_PT2PT_FUNC_EXIT(MPID_STATE_MPI_BSEND);
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_bsend",
"**mpi_bsend %p %d %D %i %t %C", buf, count, datatype, dest, tag, comm);
}
# endif
mpi_errno = MPIR_Err_return_comm( comm_ptr, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Neighbor_allgather - In this function, each process i gathers data items
from each process j if an edge (j,i) exists in the topology graph, and each
process i sends the same data items to all processes j where an edge (i,j)
exists. The send buffer is sent to each neighboring process and the l-th block
in the receive buffer is received from the l-th neighbor.
Input Parameters:
+ sendbuf - starting address of the send buffer (choice)
. sendcount - number of elements sent to each neighbor (non-negative integer)
. sendtype - data type of send buffer elements (handle)
. recvcount - number of elements received from each neighbor (non-negative integer)
. recvtype - data type of receive buffer elements (handle)
- comm - communicator (handle)
Output Parameters:
. recvbuf - starting address of the receive buffer (choice)
.N ThreadSafe
.N Fortran
.N Errors
@*/
int MPI_Neighbor_allgather(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm)
{
int mpi_errno = MPI_SUCCESS;
MPID_Comm *comm_ptr = NULL;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_NEIGHBOR_ALLGATHER);
MPIU_THREAD_CS_ENTER(ALLFUNC,);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_NEIGHBOR_ALLGATHER);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS
{
MPIR_ERRTEST_DATATYPE(sendtype, "sendtype", mpi_errno);
MPIR_ERRTEST_DATATYPE(recvtype, "recvtype", mpi_errno);
MPIR_ERRTEST_COMM(comm, mpi_errno);
/* TODO more checks may be appropriate */
}
MPID_END_ERROR_CHECKS
}
# endif /* HAVE_ERROR_CHECKING */
/* Convert MPI object handles to object pointers */
MPID_Comm_get_ptr(comm, comm_ptr);
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS
{
if (HANDLE_GET_KIND(sendtype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype *sendtype_ptr = NULL;
MPID_Datatype_get_ptr(sendtype, sendtype_ptr);
MPID_Datatype_valid_ptr(sendtype_ptr, mpi_errno);
MPID_Datatype_committed_ptr(sendtype_ptr, mpi_errno);
}
if (HANDLE_GET_KIND(recvtype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype *recvtype_ptr = NULL;
MPID_Datatype_get_ptr(recvtype, recvtype_ptr);
MPID_Datatype_valid_ptr(recvtype_ptr, mpi_errno);
MPID_Datatype_committed_ptr(recvtype_ptr, mpi_errno);
}
MPID_Comm_valid_ptr( comm_ptr, mpi_errno, FALSE );
/* TODO more checks may be appropriate (counts, in_place, buffer aliasing, etc) */
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPID_END_ERROR_CHECKS
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPIR_Neighbor_allgather_impl(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm_ptr);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_NEIGHBOR_ALLGATHER);
MPIU_THREAD_CS_EXIT(ALLFUNC,);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER,
"**mpi_neighbor_allgather", "**mpi_neighbor_allgather %p %d %D %p %d %D %C", sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm);
}
# endif
mpi_errno = MPIR_Err_return_comm(NULL, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Accumulate - Accumulate data into the target process using remote
memory access
Input Parameters:
+ origin_addr - initial address of buffer (choice)
. origin_count - number of entries in buffer (nonnegative integer)
. origin_datatype - datatype of each buffer entry (handle)
. target_rank - rank of target (nonnegative integer)
. target_disp - displacement from start of window to beginning of target
buffer (nonnegative integer)
. target_count - number of entries in target buffer (nonnegative integer)
. target_datatype - datatype of each entry in target buffer (handle)
. op - predefined reduce operation (handle)
- win - window object (handle)
Notes:
The basic components of both the origin and target datatype must be the same
predefined datatype (e.g., all 'MPI_INT' or all 'MPI_DOUBLE_PRECISION').
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_ARG
.N MPI_ERR_COUNT
.N MPI_ERR_RANK
.N MPI_ERR_TYPE
.N MPI_ERR_WIN
.seealso: MPI_Raccumulate
@*/
int MPI_Accumulate(const void *origin_addr, int origin_count, MPI_Datatype
origin_datatype, int target_rank, MPI_Aint
target_disp, int target_count, MPI_Datatype
target_datatype, MPI_Op op, MPI_Win win)
{
static const char FCNAME[] = "MPI_Accumulate";
int mpi_errno = MPI_SUCCESS;
MPIR_Win *win_ptr = NULL;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_ACCUMULATE);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPIR_FUNC_TERSE_RMA_ENTER(MPID_STATE_MPI_ACCUMULATE);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_WIN(win, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* Convert MPI object handles to object pointers */
MPIR_Win_get_ptr( win, win_ptr );
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_Comm * comm_ptr;
/* Validate win_ptr */
MPIR_Win_valid_ptr( win_ptr, mpi_errno );
if (mpi_errno) goto fn_fail;
MPIR_ERRTEST_COUNT(origin_count, mpi_errno);
MPIR_ERRTEST_DATATYPE(origin_datatype, "origin_datatype", mpi_errno);
MPIR_ERRTEST_USERBUFFER(origin_addr, origin_count, origin_datatype, mpi_errno);
MPIR_ERRTEST_COUNT(target_count, mpi_errno);
MPIR_ERRTEST_DATATYPE(target_datatype, "target_datatype", mpi_errno);
if (win_ptr->create_flavor != MPI_WIN_FLAVOR_DYNAMIC)
MPIR_ERRTEST_DISP(target_disp, mpi_errno);
if (HANDLE_GET_KIND(origin_datatype) != HANDLE_KIND_BUILTIN)
{
MPIR_Datatype *datatype_ptr = NULL;
MPID_Datatype_get_ptr(origin_datatype, datatype_ptr);
MPIR_Datatype_valid_ptr(datatype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPID_Datatype_committed_ptr(datatype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
if (HANDLE_GET_KIND(target_datatype) != HANDLE_KIND_BUILTIN)
{
MPIR_Datatype *datatype_ptr = NULL;
MPID_Datatype_get_ptr(target_datatype, datatype_ptr);
MPIR_Datatype_valid_ptr(datatype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPID_Datatype_committed_ptr(datatype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
comm_ptr = win_ptr->comm_ptr;
MPIR_ERRTEST_SEND_RANK(comm_ptr, target_rank, mpi_errno);
MPIR_ERRTEST_OP_ACC(op, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPID_Accumulate(origin_addr, origin_count,
origin_datatype,
target_rank, target_disp, target_count,
target_datatype, op, win_ptr);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* ... end of body of routine ... */
fn_exit:
MPIR_FUNC_TERSE_RMA_EXIT(MPID_STATE_MPI_ACCUMULATE);
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_accumulate",
"**mpi_accumulate %p %d %D %d %d %d %D %O %W", origin_addr, origin_count, origin_datatype,
target_rank, target_disp, target_count, target_datatype, op, win);
}
# endif
mpi_errno = MPIR_Err_return_win( win_ptr, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Type_size_x - Return the number of bytes occupied by entries
in the datatype
Input Parameters:
. datatype - datatype (handle)
Output Parameters:
. size - datatype size (integer)
.N ThreadSafe
.N Fortran
.N Errors
@*/
int MPI_Type_size_x(MPI_Datatype datatype, MPI_Count *size)
{
int mpi_errno = MPI_SUCCESS;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_TYPE_SIZE_X);
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPI_TYPE_SIZE_X);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS
{
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
/* TODO more checks may be appropriate */
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPID_END_ERROR_CHECKS
}
# endif /* HAVE_ERROR_CHECKING */
/* Convert MPI object handles to object pointers */
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS
{
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
MPIR_Datatype *datatype_ptr = NULL;
MPID_Datatype_get_ptr(datatype, datatype_ptr);
MPIR_Datatype_valid_ptr(datatype_ptr, mpi_errno);
}
/* TODO more checks may be appropriate (counts, in_place, buffer aliasing, etc) */
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPID_END_ERROR_CHECKS
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPIR_Type_size_x_impl(datatype, size);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* ... end of body of routine ... */
fn_exit:
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPI_TYPE_SIZE_X);
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER,
"**mpi_type_size_x", "**mpi_type_size_x %D %p", datatype, size);
}
# endif
mpi_errno = MPIR_Err_return_comm(NULL, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Sendrecv - Sends and receives a message
Input Parameters:
+ sendbuf - initial address of send buffer (choice)
. sendcount - number of elements in send buffer (integer)
. sendtype - type of elements in send buffer (handle)
. dest - rank of destination (integer)
. sendtag - send tag (integer)
. recvcount - number of elements in receive buffer (integer)
. recvtype - type of elements in receive buffer (handle)
. source - rank of source (integer)
. recvtag - receive tag (integer)
- comm - communicator (handle)
Output Parameters:
+ recvbuf - initial address of receive buffer (choice)
- status - status object (Status). This refers to the receive operation.
.N ThreadSafe
.N Fortran
.N FortranStatus
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_COMM
.N MPI_ERR_COUNT
.N MPI_ERR_TYPE
.N MPI_ERR_TAG
.N MPI_ERR_RANK
@*/
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)
{
int mpi_errno = MPI_SUCCESS;
MPIR_Comm *comm_ptr = NULL;
MPIR_Request *sreq = NULL;
MPIR_Request *rreq = NULL;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_SENDRECV);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(VCI_GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPIR_FUNC_TERSE_PT2PT_ENTER_BOTH(MPID_STATE_MPI_SENDRECV);
/* Validate handle parameters needing to be converted */
#ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
#endif /* HAVE_ERROR_CHECKING */
/* Convert handles to MPI objects. */
MPIR_Comm_get_ptr(comm, comm_ptr);
#ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
/* Validate communicator */
MPIR_Comm_valid_ptr(comm_ptr, mpi_errno, FALSE);
if (mpi_errno)
goto fn_fail;
/* Validate count */
MPIR_ERRTEST_COUNT(sendcount, mpi_errno);
MPIR_ERRTEST_COUNT(recvcount, mpi_errno);
/* Validate status (status_ignore is not the same as null) */
MPIR_ERRTEST_ARGNULL(status, "status", mpi_errno);
/* Validate tags */
MPIR_ERRTEST_SEND_TAG(sendtag, mpi_errno);
MPIR_ERRTEST_RECV_TAG(recvtag, mpi_errno);
/* Validate source and destination */
if (comm_ptr) {
MPIR_ERRTEST_SEND_RANK(comm_ptr, dest, mpi_errno);
MPIR_ERRTEST_RECV_RANK(comm_ptr, source, mpi_errno);
}
/* Validate datatype handles */
MPIR_ERRTEST_DATATYPE(sendtype, "datatype", mpi_errno);
MPIR_ERRTEST_DATATYPE(recvtype, "datatype", mpi_errno);
/* Validate datatype objects */
if (HANDLE_GET_KIND(sendtype) != HANDLE_KIND_BUILTIN) {
MPIR_Datatype *datatype_ptr = NULL;
MPIR_Datatype_get_ptr(sendtype, datatype_ptr);
MPIR_Datatype_valid_ptr(datatype_ptr, mpi_errno);
if (mpi_errno)
goto fn_fail;
MPIR_Datatype_committed_ptr(datatype_ptr, mpi_errno);
if (mpi_errno)
goto fn_fail;
}
if (HANDLE_GET_KIND(recvtype) != HANDLE_KIND_BUILTIN) {
MPIR_Datatype *datatype_ptr = NULL;
MPIR_Datatype_get_ptr(recvtype, datatype_ptr);
MPIR_Datatype_valid_ptr(datatype_ptr, mpi_errno);
if (mpi_errno)
goto fn_fail;
MPIR_Datatype_committed_ptr(datatype_ptr, mpi_errno);
if (mpi_errno)
goto fn_fail;
}
/* Validate buffers */
MPIR_ERRTEST_USERBUFFER(sendbuf, sendcount, sendtype, mpi_errno);
MPIR_ERRTEST_USERBUFFER(recvbuf, recvcount, recvtype, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
#endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno =
MPID_Irecv(recvbuf, recvcount, recvtype, source, recvtag, comm_ptr,
MPIR_CONTEXT_INTRA_PT2PT, &rreq);
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
/* FIXME - Performance for small messages might be better if MPID_Send() were used here instead of MPID_Isend() */
mpi_errno =
MPID_Isend(sendbuf, sendcount, sendtype, dest, sendtag, comm_ptr, MPIR_CONTEXT_INTRA_PT2PT,
&sreq);
if (mpi_errno != MPI_SUCCESS) {
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno == MPIX_ERR_NOREQ)
MPIR_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**nomem");
/* FIXME: should we cancel the pending (possibly completed) receive request or wait for it to complete? */
MPIR_Request_free(rreq);
goto fn_fail;
/* --END ERROR HANDLING-- */
}
if (!MPIR_Request_is_complete(sreq) || !MPIR_Request_is_complete(rreq)) {
MPID_Progress_state progress_state;
MPID_Progress_start(&progress_state);
while (!MPIR_Request_is_complete(sreq) || !MPIR_Request_is_complete(rreq)) {
mpi_errno = MPID_Progress_wait(&progress_state);
if (mpi_errno != MPI_SUCCESS) {
/* --BEGIN ERROR HANDLING-- */
MPID_Progress_end(&progress_state);
goto fn_fail;
/* --END ERROR HANDLING-- */
}
if (unlikely(MPIR_Request_is_anysrc_mismatched(rreq))) {
/* --BEGIN ERROR HANDLING-- */
mpi_errno = MPIR_Request_handle_proc_failed(rreq);
if (!MPIR_Request_is_complete(sreq)) {
MPID_Cancel_send(sreq);
MPIR_STATUS_SET_CANCEL_BIT(sreq->status, FALSE);
}
goto fn_fail;
/* --END ERROR HANDLING-- */
}
}
MPID_Progress_end(&progress_state);
}
mpi_errno = rreq->status.MPI_ERROR;
MPIR_Request_extract_status(rreq, status);
MPIR_Request_free(rreq);
if (mpi_errno == MPI_SUCCESS) {
mpi_errno = sreq->status.MPI_ERROR;
}
MPIR_Request_free(sreq);
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
/* ... end of body of routine ... */
fn_exit:
MPIR_FUNC_TERSE_PT2PT_EXIT_BOTH(MPID_STATE_MPI_SENDRECV);
MPID_THREAD_CS_EXIT(VCI_GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
#ifdef HAVE_ERROR_CHECKING
{
mpi_errno =
MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, __func__, __LINE__, MPI_ERR_OTHER,
"**mpi_sendrecv",
"**mpi_sendrecv %p %d %D %i %t %p %d %D %i %t %C %p", sendbuf,
sendcount, sendtype, dest, sendtag, recvbuf, recvcount, recvtype,
source, recvtag, comm, status);
}
#endif
mpi_errno = MPIR_Err_return_comm(comm_ptr, __func__, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Group_free - Frees a group
Input Parameters:
. group - group to free (handle)
Notes:
On output, group is set to 'MPI_GROUP_NULL'.
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_ARG
.N MPI_ERR_PERM_GROUP
@*/
int MPI_Group_free(MPI_Group *group)
{
int mpi_errno = MPI_SUCCESS;
MPID_Group *group_ptr = NULL;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_GROUP_FREE);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_GROUP_FREE);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_GROUP(*group, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif
/* Convert MPI object handles to object pointers */
MPID_Group_get_ptr( *group, group_ptr );
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
/* Validate group_ptr */
MPID_Group_valid_ptr( group_ptr, mpi_errno );
/* If group_ptr is not valid, it will be reset to null */
/* Cannot free the predefined groups, but allow GROUP_EMPTY
because otherwise many tests fail */
if ((HANDLE_GET_KIND(*group) == HANDLE_KIND_BUILTIN) &&
*group != MPI_GROUP_EMPTY) {
mpi_errno = MPIR_Err_create_code( MPI_SUCCESS,
MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_GROUP,
"**groupperm", 0);
}
if (mpi_errno) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPIR_Group_free_impl(group_ptr);
*group = MPI_GROUP_NULL;
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_GROUP_FREE);
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_group_free", "**mpi_group_free %p", group);
}
# endif
mpi_errno = MPIR_Err_return_comm( NULL, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
static int MPIR_Bcast_binomial_MV2(
void *buffer,
int count,
MPI_Datatype datatype,
int root,
MPID_Comm *comm_ptr,
int *errflag)
{
int rank, comm_size, src, dst;
int relative_rank, mask;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int nbytes=0;
int type_size, is_contig, is_homogeneous;
int position;
void *tmp_buf=NULL;
MPI_Comm comm;
MPID_Datatype *dtp;
MPIU_CHKLMEM_DECL(1);
comm = comm_ptr->handle;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* If there is only one process, return */
if (comm_size == 1) goto fn_exit;
if (HANDLE_GET_KIND(datatype) == HANDLE_KIND_BUILTIN)
is_contig = 1;
else {
MPID_Datatype_get_ptr(datatype, dtp);
is_contig = dtp->is_contig;
}
is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
if (comm_ptr->is_hetero)
is_homogeneous = 0;
#endif
/* MPI_Type_size() might not give the accurate size of the packed
* datatype for heterogeneous systems (because of padding, encoding,
* etc). On the other hand, MPI_Pack_size() can become very
* expensive, depending on the implementation, especially for
* heterogeneous systems. We want to use MPI_Type_size() wherever
* possible, and MPI_Pack_size() in other places.
*/
if (is_homogeneous)
MPID_Datatype_get_size_macro(datatype, type_size);
else
MPIR_Pack_size_impl(1, datatype, &type_size);
nbytes = type_size * count;
if (!is_contig || !is_homogeneous)
{
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, nbytes, mpi_errno, "tmp_buf");
/* TODO: Pipeline the packing and communication */
position = 0;
if (rank == root) {
mpi_errno = MPIR_Pack_impl(buffer, count, datatype, tmp_buf, nbytes,
&position);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
/*@
MPI_Allreduce - Combines values from all processes and distributes the result
back to all processes
Input Parameters:
+ sendbuf - starting address of send buffer (choice)
. count - number of elements in send buffer (integer)
. datatype - data type of elements of send buffer (handle)
. op - operation (handle)
- comm - communicator (handle)
Output Parameter:
. recvbuf - starting address of receive buffer (choice)
.N ThreadSafe
.N Fortran
.N collops
.N Errors
.N MPI_ERR_BUFFER
.N MPI_ERR_COUNT
.N MPI_ERR_TYPE
.N MPI_ERR_OP
.N MPI_ERR_COMM
@*/
int MPI_Allreduce ( void *sendbuf, void *recvbuf, int count,
MPI_Datatype datatype, MPI_Op op, MPI_Comm comm )
{
static const char FCNAME[] = "MPI_Allreduce";
int mpi_errno = MPI_SUCCESS;
MPID_Comm *comm_ptr = NULL;
int errflag = FALSE;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_ALLREDUCE);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPIU_THREAD_CS_ENTER(ALLFUNC,);
MPID_MPI_COLL_FUNC_ENTER(MPID_STATE_MPI_ALLREDUCE);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* Convert MPI object handles to object pointers */
MPID_Comm_get_ptr( comm, comm_ptr );
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPID_Datatype *datatype_ptr = NULL;
MPID_Op *op_ptr = NULL;
MPID_Comm_valid_ptr( comm_ptr, mpi_errno );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPIR_ERRTEST_COUNT(count, mpi_errno);
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
MPIR_ERRTEST_OP(op, mpi_errno);
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr(datatype, datatype_ptr);
MPID_Datatype_valid_ptr( datatype_ptr, mpi_errno );
MPID_Datatype_committed_ptr( datatype_ptr, mpi_errno );
}
if (comm_ptr->comm_kind == MPID_INTERCOMM)
MPIR_ERRTEST_SENDBUF_INPLACE(sendbuf, count, mpi_errno);
if (sendbuf != MPI_IN_PLACE)
MPIR_ERRTEST_USERBUFFER(sendbuf,count,datatype,mpi_errno);
MPIR_ERRTEST_RECVBUF_INPLACE(recvbuf, count, mpi_errno);
MPIR_ERRTEST_USERBUFFER(recvbuf,count,datatype,mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
if (HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN) {
MPID_Op_get_ptr(op, op_ptr);
MPID_Op_valid_ptr( op_ptr, mpi_errno );
}
if (HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) {
mpi_errno =
( * MPIR_Op_check_dtype_table[op%16 - 1] )(datatype);
}
if (count != 0) {
MPIR_ERRTEST_ALIAS_COLL(sendbuf, recvbuf, mpi_errno);
}
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPIR_Allreduce_impl(sendbuf, recvbuf, count, datatype, op, comm_ptr, &errflag);
if (mpi_errno) goto fn_fail;
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_COLL_FUNC_EXIT(MPID_STATE_MPI_ALLREDUCE);
MPIU_THREAD_CS_EXIT(ALLFUNC,);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_allreduce",
"**mpi_allreduce %p %p %d %D %O %C", sendbuf, recvbuf, count, datatype, op, comm);
}
# endif
mpi_errno = MPIR_Err_return_comm( comm_ptr, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPID_Type_vector - create a vector datatype
Input Parameters:
+ count - number of blocks in vector
. blocklength - number of elements in each block
. stride - distance from beginning of one block to the next (see next
parameter for units)
. strideinbytes - if nonzero, then stride is in bytes, otherwise stride
is in terms of extent of oldtype
- oldtype - type (using handle) of datatype on which vector is based
Output Parameters:
. newtype - handle of new vector datatype
Return Value:
0 on success, MPI error code on failure.
@*/
int MPID_Type_vector(int count,
int blocklength,
MPI_Aint stride,
int strideinbytes,
MPI_Datatype oldtype,
MPI_Datatype *newtype)
{
int mpi_errno = MPI_SUCCESS;
int is_builtin, old_is_contig;
MPI_Aint el_sz, old_sz;
MPI_Datatype el_type;
MPI_Aint old_lb, old_ub, old_extent, old_true_lb, old_true_ub, eff_stride;
MPID_Datatype *new_dtp;
if (count == 0) return MPID_Type_zerolen(newtype);
/* allocate new datatype object and handle */
new_dtp = (MPID_Datatype *) MPIU_Handle_obj_alloc(&MPID_Datatype_mem);
if (!new_dtp) {
/* --BEGIN ERROR HANDLING-- */
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE,
"MPID_Type_vector", __LINE__,
MPI_ERR_OTHER, "**nomem", 0);
return mpi_errno;
/* --END ERROR HANDLING-- */
}
/* handle is filled in by MPIU_Handle_obj_alloc() */
MPIU_Object_set_ref(new_dtp, 1);
new_dtp->is_permanent = 0;
new_dtp->is_committed = 0;
new_dtp->attributes = NULL;
new_dtp->cache_id = 0;
new_dtp->name[0] = 0;
new_dtp->contents = NULL;
new_dtp->dataloop = NULL;
new_dtp->dataloop_size = -1;
new_dtp->dataloop_depth = -1;
new_dtp->hetero_dloop = NULL;
new_dtp->hetero_dloop_size = -1;
new_dtp->hetero_dloop_depth = -1;
is_builtin = (HANDLE_GET_KIND(oldtype) == HANDLE_KIND_BUILTIN);
if (is_builtin) {
el_sz = (MPI_Aint) MPID_Datatype_get_basic_size(oldtype);
el_type = oldtype;
old_lb = 0;
old_true_lb = 0;
old_ub = el_sz;
old_true_ub = el_sz;
old_sz = el_sz;
old_extent = el_sz;
old_is_contig = 1;
new_dtp->size = (MPI_Aint) count *
(MPI_Aint) blocklength * el_sz;
new_dtp->has_sticky_lb = 0;
new_dtp->has_sticky_ub = 0;
new_dtp->alignsize = el_sz; /* ??? */
new_dtp->n_builtin_elements = count * blocklength;
new_dtp->builtin_element_size = el_sz;
new_dtp->basic_type = el_type;
new_dtp->max_contig_blocks = count;
eff_stride = (strideinbytes) ? stride : (stride * el_sz);
}
else /* user-defined base type (oldtype) */ {
MPID_Datatype *old_dtp;
MPID_Datatype_get_ptr(oldtype, old_dtp);
el_sz = old_dtp->builtin_element_size;
el_type = old_dtp->basic_type;
old_lb = old_dtp->lb;
old_true_lb = old_dtp->true_lb;
old_ub = old_dtp->ub;
old_true_ub = old_dtp->true_ub;
old_sz = old_dtp->size;
old_extent = old_dtp->extent;
old_is_contig = old_dtp->is_contig;
new_dtp->size = count * blocklength * old_dtp->size;
new_dtp->has_sticky_lb = old_dtp->has_sticky_lb;
new_dtp->has_sticky_ub = old_dtp->has_sticky_ub;
new_dtp->alignsize = old_dtp->alignsize;
new_dtp->n_builtin_elements = count * blocklength * old_dtp->n_builtin_elements;
new_dtp->builtin_element_size = el_sz;
new_dtp->basic_type = el_type;
new_dtp->max_contig_blocks = old_dtp->max_contig_blocks * count * blocklength;
eff_stride = (strideinbytes) ? stride : (stride * old_dtp->extent);
}
MPID_DATATYPE_VECTOR_LB_UB((MPI_Aint) count,
eff_stride,
(MPI_Aint) blocklength,
old_lb,
old_ub,
old_extent,
new_dtp->lb,
new_dtp->ub);
new_dtp->true_lb = new_dtp->lb + (old_true_lb - old_lb);
new_dtp->true_ub = new_dtp->ub + (old_true_ub - old_ub);
new_dtp->extent = new_dtp->ub - new_dtp->lb;
/* new type is only contig for N types if old one was, and
* size and extent of new type are equivalent, and stride is
* equal to blocklength * size of old type.
*/
if ((MPI_Aint)(new_dtp->size) == new_dtp->extent &&
eff_stride == (MPI_Aint) blocklength * old_sz &&
old_is_contig)
{
new_dtp->is_contig = 1;
new_dtp->max_contig_blocks = 1;
}
else {
new_dtp->is_contig = 0;
}
*newtype = new_dtp->handle;
MPIU_DBG_MSG_P(DATATYPE,VERBOSE,"vector type %x created.",
new_dtp->handle);
return mpi_errno;
}
int MPIR_Allreduce_intra (
void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPID_Comm *comm_ptr,
int *errflag )
{
int is_homogeneous;
#ifdef MPID_HAS_HETERO
int rc;
#endif
int comm_size, rank, type_size;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative, pof2, newrank, rem, newdst, i,
send_idx, recv_idx, last_idx, send_cnt, recv_cnt, *cnts, *disps;
MPI_Aint true_extent, true_lb, extent;
void *tmp_buf;
MPI_Comm comm;
MPIU_CHKLMEM_DECL(3);
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
if (count == 0) goto fn_exit;
comm = comm_ptr->handle;
is_commutative = MPIR_Op_is_commutative(op);
#if defined(USE_SMP_COLLECTIVES)
/* is the op commutative? We do SMP optimizations only if it is. */
if (MPIR_Comm_is_node_aware(comm_ptr) && is_commutative) {
/* on each node, do a reduce to the local root */
if (comm_ptr->node_comm != NULL) {
/* take care of the MPI_IN_PLACE case. For reduce,
MPI_IN_PLACE is specified only on the root;
for allreduce it is specified on all processes. */
if ((sendbuf == MPI_IN_PLACE) && (comm_ptr->node_comm->rank != 0)) {
/* IN_PLACE and not root of reduce. Data supplied to this
allreduce is in recvbuf. Pass that as the sendbuf to reduce. */
mpi_errno = MPIR_Reduce_impl(recvbuf, NULL, count, datatype, op, 0, comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
} else {
mpi_errno = MPIR_Reduce_impl(sendbuf, recvbuf, count, datatype, op, 0, comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
} else {
/* only one process on the node. copy sendbuf to recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
/* now do an IN_PLACE allreduce among the local roots of all nodes */
if (comm_ptr->node_roots_comm != NULL) {
mpi_errno = allreduce_intra_or_coll_fn(MPI_IN_PLACE, recvbuf, count, datatype, op, comm_ptr->node_roots_comm,
errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
/* now broadcast the result among local processes */
if (comm_ptr->node_comm != NULL) {
mpi_errno = MPIR_Bcast_impl(recvbuf, count, datatype, 0, comm_ptr->node_comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
goto fn_exit;
}
#endif
is_homogeneous = 1;
#ifdef MPID_HAS_HETERO
if (comm_ptr->is_hetero)
is_homogeneous = 0;
#endif
#ifdef MPID_HAS_HETERO
if (!is_homogeneous) {
/* heterogeneous. To get the same result on all processes, we
do a reduce to 0 and then broadcast. */
mpi_errno = MPIR_Reduce_impl ( sendbuf, recvbuf, count, datatype,
op, 0, comm_ptr, errflag );
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
mpi_errno = MPIR_Bcast_impl( recvbuf, count, datatype, 0, comm_ptr, errflag );
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
else
#endif /* MPID_HAS_HETERO */
{
/* homogeneous */
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
is_commutative = MPIR_Op_is_commutative(op);
/* need to allocate temporary buffer to store incoming data*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(datatype, extent);
MPID_Ensure_Aint_fits_in_pointer(count * MPIR_MAX(extent, true_extent));
MPIU_CHKLMEM_MALLOC(tmp_buf, void *, count*(MPIR_MAX(extent,true_extent)), mpi_errno, "temporary buffer");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
/* copy local data into recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf,
count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
MPID_Datatype_get_size_macro(datatype, type_size);
/* find nearest power-of-two less than or equal to comm_size */
pof2 = 1;
while (pof2 <= comm_size) pof2 <<= 1;
pof2 >>=1;
rem = comm_size - pof2;
/* In the non-power-of-two case, all even-numbered
processes of rank < 2*rem send their data to
(rank+1). These even-numbered processes no longer
participate in the algorithm until the very end. The
remaining processes form a nice power-of-two. */
if (rank < 2*rem) {
if (rank % 2 == 0) { /* even */
mpi_errno = MPIC_Send_ft(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_TAG, comm, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
doubling */
newrank = -1;
}
else { /* odd */
mpi_errno = MPIC_Recv_ft(tmp_buf, count,
datatype, rank-1,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* do the reduction on received data. since the
ordering is right, it doesn't matter whether
the operation is commutative or not. */
mpi_errno = MPIR_Reduce_local_impl(tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* change the rank */
newrank = rank / 2;
}
}
else /* rank >= 2*rem */
newrank = rank - rem;
/* If op is user-defined or count is less than pof2, use
recursive doubling algorithm. Otherwise do a reduce-scatter
followed by allgather. (If op is user-defined,
derived datatypes are allowed and the user could pass basic
datatypes on one process and derived on another as long as
the type maps are the same. Breaking up derived
datatypes to do the reduce-scatter is tricky, therefore
using recursive doubling in that case.) */
if (newrank != -1) {
if ((count*type_size <= MPIR_PARAM_ALLREDUCE_SHORT_MSG_SIZE) ||
(HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN) ||
(count < pof2)) { /* use recursive doubling */
mask = 0x1;
while (mask < pof2) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
/* Send the most current data, which is in recvbuf. Recv
into tmp_buf */
mpi_errno = MPIC_Sendrecv_ft(recvbuf, count, datatype,
dst, MPIR_ALLREDUCE_TAG, tmp_buf,
count, datatype, dst,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* tmp_buf contains data received in this step.
recvbuf contains data accumulated so far */
if (is_commutative || (dst < rank)) {
/* op is commutative OR the order is already right */
mpi_errno = MPIR_Reduce_local_impl(tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else {
/* op is noncommutative and the order is not right */
mpi_errno = MPIR_Reduce_local_impl(recvbuf, tmp_buf, count, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* copy result back into recvbuf */
mpi_errno = MPIR_Localcopy(tmp_buf, count, datatype,
recvbuf, count, datatype);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
mask <<= 1;
}
}
else {
/* do a reduce-scatter followed by allgather */
/* for the reduce-scatter, calculate the count that
each process receives and the displacement within
the buffer */
MPIU_CHKLMEM_MALLOC(cnts, int *, pof2*sizeof(int), mpi_errno, "counts");
MPIU_CHKLMEM_MALLOC(disps, int *, pof2*sizeof(int), mpi_errno, "displacements");
for (i=0; i<(pof2-1); i++)
cnts[i] = count/pof2;
cnts[pof2-1] = count - (count/pof2)*(pof2-1);
disps[0] = 0;
for (i=1; i<pof2; i++)
disps[i] = disps[i-1] + cnts[i-1];
mask = 0x1;
send_idx = recv_idx = 0;
last_idx = pof2;
while (mask < pof2) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
send_idx = recv_idx + pof2/(mask*2);
for (i=send_idx; i<last_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += cnts[i];
}
else {
recv_idx = send_idx + pof2/(mask*2);
for (i=send_idx; i<recv_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += cnts[i];
}
/* printf("Rank %d, send_idx %d, recv_idx %d, send_cnt %d, recv_cnt %d, last_idx %d\n", newrank, send_idx, recv_idx,
send_cnt, recv_cnt, last_idx);
*/
/* Send data from recvbuf. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv_ft((char *) recvbuf +
disps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_ALLREDUCE_TAG,
(char *) tmp_buf +
disps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* tmp_buf contains data received in this step.
recvbuf contains data accumulated so far */
/* This algorithm is used only for predefined ops
and predefined ops are always commutative. */
mpi_errno = MPIR_Reduce_local_impl(((char *) tmp_buf + disps[recv_idx]*extent),
((char *) recvbuf + disps[recv_idx]*extent),
recv_cnt, datatype, op);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* update send_idx for next iteration */
send_idx = recv_idx;
mask <<= 1;
/* update last_idx, but not in last iteration
because the value is needed in the allgather
step below. */
if (mask < pof2)
last_idx = recv_idx + pof2/mask;
}
/* now do the allgather */
mask >>= 1;
while (mask > 0) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
/* update last_idx except on first iteration */
if (mask != pof2/2)
last_idx = last_idx + pof2/(mask*2);
recv_idx = send_idx + pof2/(mask*2);
for (i=send_idx; i<recv_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += cnts[i];
}
else {
recv_idx = send_idx - pof2/(mask*2);
for (i=send_idx; i<last_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += cnts[i];
}
mpi_errno = MPIC_Sendrecv_ft((char *) recvbuf +
disps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_ALLREDUCE_TAG,
(char *) recvbuf +
disps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
if (newrank > newdst) send_idx = recv_idx;
mask >>= 1;
}
}
}
/* In the non-power-of-two case, all odd-numbered
processes of rank < 2*rem send the result to
(rank-1), the ranks who didn't participate above. */
if (rank < 2*rem) {
if (rank % 2) /* odd */
mpi_errno = MPIC_Send_ft(recvbuf, count,
datatype, rank-1,
MPIR_ALLREDUCE_TAG, comm, errflag);
else /* even */
mpi_errno = MPIC_Recv_ft(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = TRUE;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
}
}
fn_exit:
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
MPIU_CHKLMEM_FREEALL();
if (mpi_errno_ret)
mpi_errno = mpi_errno_ret;
return (mpi_errno);
fn_fail:
goto fn_exit;
}
int MPIR_Pack_impl(const void *inbuf,
int incount,
MPI_Datatype datatype,
void *outbuf,
MPI_Aint outsize,
MPI_Aint *position)
{
int mpi_errno = MPI_SUCCESS;
MPI_Aint first, last;
MPID_Segment *segp;
int contig;
MPI_Aint dt_true_lb;
MPI_Aint data_sz;
if (incount == 0) {
goto fn_exit;
}
/* Handle contig case quickly */
if (HANDLE_GET_KIND(datatype) == HANDLE_KIND_BUILTIN) {
contig = TRUE;
dt_true_lb = 0;
data_sz = incount * MPID_Datatype_get_basic_size(datatype);
} else {
MPID_Datatype *dt_ptr;
MPID_Datatype_get_ptr(datatype, dt_ptr);
contig = dt_ptr->is_contig;
dt_true_lb = dt_ptr->true_lb;
data_sz = incount * dt_ptr->size;
}
if (contig) {
MPIU_Memcpy((char *) outbuf + *position, (char *)inbuf + dt_true_lb, data_sz);
*position = (int)((MPI_Aint)*position + data_sz);
goto fn_exit;
}
/* non-contig case */
/* TODO: CHECK RETURN VALUES?? */
/* TODO: SHOULD THIS ALL BE IN A MPID_PACK??? */
segp = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP1(segp == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment");
mpi_errno = MPID_Segment_init(inbuf, incount, datatype, segp, 0);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* NOTE: the use of buffer values and positions in MPI_Pack and in
* MPID_Segment_pack are quite different. See code or docs or something.
*/
first = 0;
last = SEGMENT_IGNORE_LAST;
/* Ensure that pointer increment fits in a pointer */
MPID_Ensure_Aint_fits_in_pointer((MPI_VOID_PTR_CAST_TO_MPI_AINT outbuf) +
(MPI_Aint) *position);
MPID_Segment_pack(segp,
first,
&last,
(void *) ((char *) outbuf + *position));
/* Ensure that calculation fits into an int datatype. */
MPID_Ensure_Aint_fits_in_int((MPI_Aint)*position + last);
*position = (int)((MPI_Aint)*position + last);
MPID_Segment_free(segp);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
/*@
MPI_Pack - Packs a datatype into contiguous memory
Input Parameters:
+ inbuf - input buffer start (choice)
. incount - number of input data items (non-negative integer)
. datatype - datatype of each input data item (handle)
. outsize - output buffer size, in bytes (non-negative integer)
- comm - communicator for packed message (handle)
Output Parameters:
. outbuf - output buffer start (choice)
Input/Output Parameters:
. position - current position in buffer, in bytes (integer)
Notes (from the specifications):
The input value of position is the first location in the output buffer to be
used for packing. position is incremented by the size of the packed message,
and the output value of position is the first location in the output buffer
following the locations occupied by the packed message. The comm argument is
the communicator that will be subsequently used for sending the packed
message.
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_ARG
.N MPI_ERR_OTHER
@*/
int MPI_Pack(const void *inbuf,
int incount,
MPI_Datatype datatype,
void *outbuf,
int outsize,
int *position,
MPI_Comm comm)
{
int mpi_errno = MPI_SUCCESS;
MPI_Aint position_x;
MPID_Comm *comm_ptr = NULL;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_PACK);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_PACK);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif
/* Convert MPI object handles to object pointers */
MPID_Comm_get_ptr(comm, comm_ptr);
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COUNT(incount,mpi_errno);
MPIR_ERRTEST_COUNT(outsize,mpi_errno);
/* NOTE: inbuf could be null (MPI_BOTTOM) */
if (incount > 0) {
MPIR_ERRTEST_ARGNULL(outbuf, "output buffer", mpi_errno);
}
MPIR_ERRTEST_ARGNULL(position, "position", mpi_errno);
/* Validate comm_ptr */
/* If comm_ptr is not valid, it will be reset to null */
MPID_Comm_valid_ptr(comm_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype *datatype_ptr = NULL;
MPID_Datatype_get_ptr(datatype, datatype_ptr);
MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPID_Datatype_committed_ptr(datatype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
#ifdef HAVE_ERROR_CHECKING /* IMPLEMENTATION-SPECIFIC ERROR CHECKS */
{
int tmp_sz;
MPID_BEGIN_ERROR_CHECKS;
/* Verify that there is space in the buffer to pack the type */
MPID_Datatype_get_size_macro(datatype, tmp_sz);
if (tmp_sz * incount > outsize - *position) {
if (*position < 0) {
MPIU_ERR_SETANDJUMP1(mpi_errno,MPI_ERR_ARG,
"**argposneg","**argposneg %d",
*position);
}
else if (outsize < 0) {
MPIU_ERR_SETANDJUMP2(mpi_errno,MPI_ERR_ARG,"**argneg",
"**argneg %s %d","outsize",outsize);
}
else if (incount < 0) {
MPIU_ERR_SETANDJUMP2(mpi_errno,MPI_ERR_ARG,"**argneg",
"**argneg %s %d","incount",incount);
}
else {
MPIU_ERR_SETANDJUMP2(mpi_errno,MPI_ERR_ARG,"**argpackbuf",
"**argpackbuf %d %d", tmp_sz * incount,
outsize - *position);
}
}
MPID_END_ERROR_CHECKS;
}
#endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
position_x = *position;
mpi_errno = MPIR_Pack_impl(inbuf, incount, datatype, outbuf, outsize, &position_x);
MPIU_Assign_trunc(*position, position_x, int);
if (mpi_errno) goto fn_fail;
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_PACK);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_pack",
"**mpi_pack %p %d %D %p %d %p %C", inbuf, incount, datatype, outbuf, outsize, position, comm);
}
# endif
mpi_errno = MPIR_Err_return_comm(comm_ptr, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Igatherv - XXX description here
Input Parameters:
+ sendbuf - starting address of the send buffer (choice)
. sendcount - number of elements in send buffer (non-negative integer)
. sendtype - data type of send buffer elements (handle)
. recvcounts - non-negative integer array (of length group size) containing the number of elements that are received from each process (significant only at root)
. displs - integer array (of length group size). Entry i specifies the displacement relative to recvbuf at which to place the incoming data from process i (significant only at root)
. recvtype - data type of receive buffer elements (significant only at root) (handle)
. root - rank of receiving process (integer)
- comm - communicator (handle)
Output Parameters:
+ recvbuf - starting address of the receive buffer (significant only at root) (choice)
- request - communication request (handle)
.N ThreadSafe
.N Fortran
.N Errors
@*/
int MPI_Igatherv(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf,
const int recvcounts[], const int displs[], MPI_Datatype recvtype, int root,
MPI_Comm comm, MPI_Request *request)
{
int mpi_errno = MPI_SUCCESS;
MPID_Comm *comm_ptr = NULL;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_IGATHERV);
MPIU_THREAD_CS_ENTER(ALLFUNC,);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_IGATHERV);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
/* TODO more checks may be appropriate */
}
MPID_END_ERROR_CHECKS
}
# endif /* HAVE_ERROR_CHECKING */
/* Convert MPI object handles to object pointers */
MPID_Comm_get_ptr(comm, comm_ptr);
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS
{
MPID_Datatype *sendtype_ptr=NULL, *recvtype_ptr=NULL;
int i, rank, comm_size;
MPID_Comm_valid_ptr( comm_ptr, mpi_errno );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
if (comm_ptr->comm_kind == MPID_INTRACOMM) {
MPIR_ERRTEST_INTRA_ROOT(comm_ptr, root, mpi_errno);
if (sendbuf != MPI_IN_PLACE) {
MPIR_ERRTEST_COUNT(sendcount, mpi_errno);
MPIR_ERRTEST_DATATYPE(sendtype, "sendtype", mpi_errno);
if (HANDLE_GET_KIND(sendtype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr(sendtype, sendtype_ptr);
MPID_Datatype_valid_ptr( sendtype_ptr, mpi_errno );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPID_Datatype_committed_ptr( sendtype_ptr, mpi_errno );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPIR_ERRTEST_USERBUFFER(sendbuf,sendcount,sendtype,mpi_errno);
}
rank = comm_ptr->rank;
if (rank == root) {
comm_size = comm_ptr->local_size;
for (i=0; i<comm_size; i++) {
MPIR_ERRTEST_COUNT(recvcounts[i], mpi_errno);
MPIR_ERRTEST_DATATYPE(recvtype, "recvtype", mpi_errno);
}
if (HANDLE_GET_KIND(recvtype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr(recvtype, recvtype_ptr);
MPID_Datatype_valid_ptr( recvtype_ptr, mpi_errno );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPID_Datatype_committed_ptr( recvtype_ptr, mpi_errno );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
for (i=0; i<comm_size; i++) {
if (recvcounts[i] > 0) {
MPIR_ERRTEST_RECVBUF_INPLACE(recvbuf, recvcounts[i], mpi_errno);
MPIR_ERRTEST_USERBUFFER(recvbuf,recvcounts[i],recvtype,mpi_errno);
break;
}
}
}
else
MPIR_ERRTEST_SENDBUF_INPLACE(sendbuf, sendcount, mpi_errno);
}
if (comm_ptr->comm_kind == MPID_INTERCOMM) {
MPIR_ERRTEST_INTER_ROOT(comm_ptr, root, mpi_errno);
if (root == MPI_ROOT) {
comm_size = comm_ptr->remote_size;
for (i=0; i<comm_size; i++) {
MPIR_ERRTEST_COUNT(recvcounts[i], mpi_errno);
MPIR_ERRTEST_DATATYPE(recvtype, "recvtype", mpi_errno);
}
if (HANDLE_GET_KIND(recvtype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr(recvtype, recvtype_ptr);
MPID_Datatype_valid_ptr( recvtype_ptr, mpi_errno );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPID_Datatype_committed_ptr( recvtype_ptr, mpi_errno );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
for (i=0; i<comm_size; i++) {
if (recvcounts[i] > 0) {
MPIR_ERRTEST_RECVBUF_INPLACE(recvbuf, recvcounts[i], mpi_errno);
MPIR_ERRTEST_USERBUFFER(recvbuf,recvcounts[i],recvtype,mpi_errno);
break;
}
}
}
else if (root != MPI_PROC_NULL) {
MPIR_ERRTEST_COUNT(sendcount, mpi_errno);
MPIR_ERRTEST_DATATYPE(sendtype, "sendtype", mpi_errno);
if (HANDLE_GET_KIND(sendtype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr(sendtype, sendtype_ptr);
MPID_Datatype_valid_ptr( sendtype_ptr, mpi_errno );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPID_Datatype_committed_ptr( sendtype_ptr, mpi_errno );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPIR_ERRTEST_SENDBUF_INPLACE(sendbuf, sendcount, mpi_errno);
MPIR_ERRTEST_USERBUFFER(sendbuf,sendcount,sendtype,mpi_errno);
}
}
}
MPID_END_ERROR_CHECKS
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPIR_Igatherv_impl(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, root, comm_ptr, request);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_IGATHERV);
MPIU_THREAD_CS_EXIT(ALLFUNC,);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER,
"**mpi_igatherv", "**mpi_igatherv %p %d %D %p %p %p %D %d %C %p", sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, root, comm, request);
}
# endif
mpi_errno = MPIR_Err_return_comm(comm_ptr, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
goto fn_exit;
}
/*@
MPI_Bsend_init - Builds a handle for a buffered send
Input Parameters:
+ buf - initial address of send buffer (choice)
. count - number of elements sent (integer)
. datatype - type of each element (handle)
. dest - rank of destination (integer)
. tag - message tag (integer)
- comm - communicator (handle)
Output Parameters:
. request - communication request (handle)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_COMM
.N MPI_ERR_COUNT
.N MPI_ERR_TYPE
.N MPI_ERR_RANK
.N MPI_ERR_TAG
.seealso: MPI_Buffer_attach
@*/
int MPI_Bsend_init(const void *buf, int count, MPI_Datatype datatype,
int dest, int tag, MPI_Comm comm, MPI_Request *request)
{
static const char FCNAME[] = "MPI_Bsend_init";
int mpi_errno = MPI_SUCCESS;
MPID_Request *request_ptr = NULL;
MPID_Comm *comm_ptr = NULL;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_BSEND_INIT);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPIU_THREAD_CS_ENTER(ALLFUNC,);
MPID_MPI_PT2PT_FUNC_ENTER(MPID_STATE_MPI_BSEND_INIT);
/* Validate handle parameters needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* Convert MPI object handles to object pointers */
MPID_Comm_get_ptr( comm, comm_ptr );
/* Validate parameters if error checking is enabled */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPID_Comm_valid_ptr( comm_ptr, mpi_errno );
if (mpi_errno) goto fn_fail;
MPIR_ERRTEST_COUNT(count, mpi_errno);
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
MPIR_ERRTEST_SEND_RANK(comm_ptr, dest, mpi_errno);
MPIR_ERRTEST_SEND_TAG(tag, mpi_errno);
MPIR_ERRTEST_ARGNULL(request,"request",mpi_errno);
/* Validate datatype object */
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN)
{
MPID_Datatype *datatype_ptr = NULL;
MPID_Datatype_get_ptr(datatype, datatype_ptr);
MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno);
if (mpi_errno) goto fn_fail;
MPID_Datatype_committed_ptr(datatype_ptr, mpi_errno);
if (mpi_errno) goto fn_fail;
}
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPID_Bsend_init(buf, count, datatype, dest, tag, comm_ptr,
MPID_CONTEXT_INTRA_PT2PT, &request_ptr);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* return the handle of the request to the user */
MPIU_OBJ_PUBLISH_HANDLE(*request, request_ptr->handle);
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_PT2PT_FUNC_EXIT(MPID_STATE_MPI_BSEND_INIT);
MPIU_THREAD_CS_EXIT(ALLFUNC,);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_bsend_init",
"**mpi_bsend_init %p %d %D %i %t %C %p", buf, count, datatype, dest, tag, comm, request);
}
# endif
mpi_errno = MPIR_Err_return_comm( comm_ptr, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Igather - Gathers together values from a group of processes in
a nonblocking way
Input Parameters:
+ sendbuf - starting address of the send buffer (choice)
. sendcount - number of elements in send buffer (non-negative integer)
. sendtype - data type of send buffer elements (handle)
. recvcount - number of elements for any single receive (non-negative integer, significant only at root)
. recvtype - data type of receive buffer elements (significant only at root) (handle)
. root - rank of receiving process (integer)
- comm - communicator (handle)
Output Parameters:
+ recvbuf - starting address of the receive buffer (significant only at root) (choice)
- request - communication request (handle)
.N ThreadSafe
.N Fortran
.N Errors
@*/
int MPI_Igather(const void *sendbuf, int sendcount, MPI_Datatype sendtype,
void *recvbuf, int recvcount, MPI_Datatype recvtype,
int root, MPI_Comm comm, MPI_Request * request)
{
int mpi_errno = MPI_SUCCESS;
MPIR_Comm *comm_ptr = NULL;
MPIR_Request *request_ptr = NULL;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_IGATHER);
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPI_IGATHER);
/* Validate parameters, especially handles needing to be converted */
#ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
#endif /* HAVE_ERROR_CHECKING */
/* Convert MPI object handles to object pointers */
MPIR_Comm_get_ptr(comm, comm_ptr);
/* Validate parameters and objects (post conversion) */
#ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_Datatype *sendtype_ptr = NULL, *recvtype_ptr = NULL;
int rank;
MPIR_Comm_valid_ptr(comm_ptr, mpi_errno, FALSE);
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
if (comm_ptr->comm_kind == MPIR_COMM_KIND__INTRACOMM) {
MPIR_ERRTEST_INTRA_ROOT(comm_ptr, root, mpi_errno);
if (sendbuf != MPI_IN_PLACE) {
MPIR_ERRTEST_COUNT(sendcount, mpi_errno);
MPIR_ERRTEST_DATATYPE(sendtype, "sendtype", mpi_errno);
if (HANDLE_GET_KIND(sendtype) != HANDLE_KIND_BUILTIN) {
MPIR_Datatype_get_ptr(sendtype, sendtype_ptr);
MPIR_Datatype_valid_ptr(sendtype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
MPIR_Datatype_committed_ptr(sendtype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
}
MPIR_ERRTEST_USERBUFFER(sendbuf, sendcount, sendtype, mpi_errno);
}
rank = comm_ptr->rank;
if (rank == root) {
MPIR_ERRTEST_COUNT(recvcount, mpi_errno);
MPIR_ERRTEST_DATATYPE(recvtype, "recvtype", mpi_errno);
if (HANDLE_GET_KIND(recvtype) != HANDLE_KIND_BUILTIN) {
MPIR_Datatype_get_ptr(recvtype, recvtype_ptr);
MPIR_Datatype_valid_ptr(recvtype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
MPIR_Datatype_committed_ptr(recvtype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
}
MPIR_ERRTEST_RECVBUF_INPLACE(recvbuf, recvcount, mpi_errno);
MPIR_ERRTEST_USERBUFFER(recvbuf, recvcount, recvtype, mpi_errno);
/* catch common aliasing cases */
if (recvbuf != MPI_IN_PLACE && sendtype == recvtype && sendcount == recvcount &&
sendcount != 0) {
MPI_Aint recvtype_size;
MPIR_Datatype_get_size_macro(recvtype, recvtype_size);
MPIR_ERRTEST_ALIAS_COLL(sendbuf,
(char *) recvbuf +
comm_ptr->rank * recvcount * recvtype_size,
mpi_errno);
}
} else
MPIR_ERRTEST_SENDBUF_INPLACE(sendbuf, sendcount, mpi_errno);
}
if (comm_ptr->comm_kind == MPIR_COMM_KIND__INTERCOMM) {
MPIR_ERRTEST_INTER_ROOT(comm_ptr, root, mpi_errno);
if (root == MPI_ROOT) {
MPIR_ERRTEST_COUNT(recvcount, mpi_errno);
MPIR_ERRTEST_DATATYPE(recvtype, "recvtype", mpi_errno);
if (HANDLE_GET_KIND(recvtype) != HANDLE_KIND_BUILTIN) {
MPIR_Datatype_get_ptr(recvtype, recvtype_ptr);
MPIR_Datatype_valid_ptr(recvtype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
MPIR_Datatype_committed_ptr(recvtype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
}
MPIR_ERRTEST_RECVBUF_INPLACE(recvbuf, recvcount, mpi_errno);
MPIR_ERRTEST_USERBUFFER(recvbuf, recvcount, recvtype, mpi_errno);
}
else if (root != MPI_PROC_NULL) {
MPIR_ERRTEST_COUNT(sendcount, mpi_errno);
MPIR_ERRTEST_DATATYPE(sendtype, "sendtype", mpi_errno);
if (HANDLE_GET_KIND(sendtype) != HANDLE_KIND_BUILTIN) {
MPIR_Datatype_get_ptr(sendtype, sendtype_ptr);
MPIR_Datatype_valid_ptr(sendtype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
MPIR_Datatype_committed_ptr(sendtype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
}
MPIR_ERRTEST_SENDBUF_INPLACE(sendbuf, sendcount, mpi_errno);
MPIR_ERRTEST_USERBUFFER(sendbuf, sendcount, sendtype, mpi_errno);
}
}
}
MPID_END_ERROR_CHECKS;
}
#endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPIR_Igather(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype,
root, comm_ptr, &request_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* create a complete request, if needed */
if (!request_ptr)
request_ptr = MPIR_Request_create_complete(MPIR_REQUEST_KIND__COLL);
/* return the handle of the request to the user */
*request = request_ptr->handle;
/* ... end of body of routine ... */
fn_exit:
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPI_IGATHER);
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
#ifdef HAVE_ERROR_CHECKING
{
mpi_errno =
MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, __func__, __LINE__, MPI_ERR_OTHER,
"**mpi_igather", "**mpi_igather %p %d %D %p %d %D %d %C %p",
sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root,
comm, request);
}
#endif
mpi_errno = MPIR_Err_return_comm(comm_ptr, __func__, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Finalize - Terminates MPI execution environment
Notes:
All processes must call this routine before exiting. The number of
processes running `after` this routine is called is undefined;
it is best not to perform much more than a 'return rc' after calling
'MPI_Finalize'.
Thread and Signal Safety:
The MPI standard requires that 'MPI_Finalize' be called `only` by the same
thread that initialized MPI with either 'MPI_Init' or 'MPI_Init_thread'.
.N Fortran
.N Errors
.N MPI_SUCCESS
@*/
int MPI_Finalize( void )
{
static const char FCNAME[] = "MPI_Finalize";
int mpi_errno = MPI_SUCCESS;
#if defined(HAVE_USLEEP) && defined(USE_COVERAGE)
int rank=0;
#endif
MPID_MPI_FINALIZE_STATE_DECL(MPID_STATE_MPI_FINALIZE);
MPIR_ERRTEST_INITIALIZED_ORDIE();
/* Note: Only one thread may ever call MPI_Finalize (MPI_Finalize may
be called at most once in any program) */
MPIU_THREAD_CS_ENTER(ALLFUNC,);
MPID_MPI_FINALIZE_FUNC_ENTER(MPID_STATE_MPI_FINALIZE);
/* ... body of routine ... */
/* If the user requested for asynchronous progress, we need to
* shutdown the progress thread */
if (MPIR_async_thread_initialized) {
mpi_errno = MPIR_Finalize_async_thread();
if (mpi_errno) goto fn_fail;
}
#if defined(HAVE_USLEEP) && defined(USE_COVERAGE)
/* We need to get the rank before freeing MPI_COMM_WORLD */
rank = MPIR_Process.comm_world->rank;
#endif
/* Remove the attributes, executing the attribute delete routine.
Do this only if the attribute functions are defined. */
/* The standard (MPI-2, section 4.8) says that the attributes on
MPI_COMM_SELF are deleted before almost anything else happens */
/* Note that the attributes need to be removed from the communicators
so that they aren't freed twice. (The communicators are released
in MPID_Finalize) */
if (MPIR_Process.attr_free && MPIR_Process.comm_self->attributes) {
mpi_errno = MPIR_Process.attr_free( MPI_COMM_SELF,
&MPIR_Process.comm_self->attributes);
MPIR_Process.comm_self->attributes = 0;
}
if (MPIR_Process.attr_free && MPIR_Process.comm_world->attributes) {
mpi_errno = MPIR_Process.attr_free( MPI_COMM_WORLD,
&MPIR_Process.comm_world->attributes);
MPIR_Process.comm_world->attributes = 0;
}
/*
* Now that we're finalizing, we need to take control of the error handlers
* At this point, we will release any user-defined error handlers on
* comm self and comm world
*/
/* no MPI_OBJ CS needed here */
if (MPIR_Process.comm_world->errhandler &&
! (HANDLE_GET_KIND(MPIR_Process.comm_world->errhandler->handle) ==
HANDLE_KIND_BUILTIN) ) {
int in_use;
MPIR_Errhandler_release_ref( MPIR_Process.comm_world->errhandler,
&in_use);
if (!in_use) {
MPIU_Handle_obj_free( &MPID_Errhandler_mem,
MPIR_Process.comm_world->errhandler );
MPIR_Process.comm_world->errhandler = NULL;
}
}
if (MPIR_Process.comm_self->errhandler &&
! (HANDLE_GET_KIND(MPIR_Process.comm_self->errhandler->handle) ==
HANDLE_KIND_BUILTIN) ) {
int in_use;
MPIR_Errhandler_release_ref( MPIR_Process.comm_self->errhandler,
&in_use);
if (!in_use) {
MPIU_Handle_obj_free( &MPID_Errhandler_mem,
MPIR_Process.comm_self->errhandler );
MPIR_Process.comm_self->errhandler = NULL;
}
}
/* FIXME: Why is this not one of the finalize callbacks?. Do we need
pre and post MPID_Finalize callbacks? */
MPIU_Timer_finalize();
/* Call the high-priority callbacks */
MPIR_Call_finalize_callbacks( MPIR_FINALIZE_CALLBACK_PRIO+1,
MPIR_FINALIZE_CALLBACK_MAX_PRIO );
/* Signal the debugger that we are about to exit. */
/* FIXME: Should this also be a finalize callback? */
#ifdef HAVE_DEBUGGER_SUPPORT
MPIR_DebuggerSetAborting( (char *)0 );
#endif
mpi_errno = MPID_Finalize();
if (mpi_errno) {
MPIU_ERR_POP(mpi_errno);
}
/* Call the low-priority (post Finalize) callbacks */
MPIR_Call_finalize_callbacks( 0, MPIR_FINALIZE_CALLBACK_PRIO-1 );
/* At this point, if there has been a failure, exit before
completing the finalize */
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* FIXME: Many of these debugging items could/should be callbacks,
added to the finalize callback list */
/* FIXME: the memory tracing code block should be a finalize callback */
/* If memory debugging is enabled, check the memory here, after all
finalize callbacks */
MPIU_THREAD_CS_EXIT(ALLFUNC,);
MPIR_Process.initialized = MPICH_POST_FINALIZED;
MPIU_THREAD_CS_FINALIZE;
/* We place the memory tracing at the very end because any of the other
steps may have allocated memory that they still need to release*/
#ifdef USE_MEMORY_TRACING
/* FIXME: We'd like to arrange for the mem dump output to
go to separate files or to be sorted by rank (note that
the rank is at the head of the line) */
{
if (MPIR_PARAM_MEMDUMP) {
/* The second argument is the min id to print; memory allocated
after MPI_Init is given an id of one. This allows us to
ignore, if desired, memory leaks in the MPID_Init call */
MPIU_trdump( (void *)0, -1 );
}
}
#endif
#if defined(HAVE_USLEEP) && defined(USE_COVERAGE)
/* If performing coverage analysis, make each process sleep for
rank * 100 ms, to give time for the coverage tool to write out
any files. It would be better if the coverage tool and runtime
was more careful about file updates, though the lack of OS support
for atomic file updates makes this harder. */
/*
On some systems, a 0.1 second delay appears to be too short for
the file system. This code allows the use of the environment
variable MPICH_FINALDELAY, which is the delay in milliseconds.
It must be an integer value.
*/
{
int microseconds = 100000;
char *delayStr = getenv( "MPICH_FINALDELAY" );
if (delayStr) {
/* Because this is a maintainer item, we won't check for
errors in the delayStr */
microseconds = 1000 * atoi( delayStr );
}
usleep( rank * microseconds );
}
#endif
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_FINALIZE_FUNC_EXIT(MPID_STATE_MPI_FINALIZE);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE,
FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_finalize", 0);
}
# endif
mpi_errno = MPIR_Err_return_comm( 0, FCNAME, mpi_errno );
if (MPIR_Process.initialized < MPICH_POST_FINALIZED) {
MPIU_THREAD_CS_EXIT(ALLFUNC,);
}
int MPIR_Allreduce_intra_reduce_scatter_allgather(
const void *sendbuf,
void *recvbuf,
int count,
MPI_Datatype datatype,
MPI_Op op,
MPIR_Comm * comm_ptr,
MPIR_Errflag_t * errflag)
{
MPIR_CHKLMEM_DECL(3);
#ifdef MPID_HAS_HETERO
int is_homogeneous;
int rc;
#endif
int comm_size, rank;
int mpi_errno = MPI_SUCCESS;
int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, pof2, newrank, rem, newdst, i,
send_idx, recv_idx, last_idx, send_cnt, recv_cnt, *cnts, *disps;
MPI_Aint true_extent, true_lb, extent;
void *tmp_buf;
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* need to allocate temporary buffer to store incoming data*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPIR_Datatype_get_extent_macro(datatype, extent);
MPIR_Ensure_Aint_fits_in_pointer(count * MPL_MAX(extent, true_extent));
MPIR_CHKLMEM_MALLOC(tmp_buf, void *, count*(MPL_MAX(extent,true_extent)), mpi_errno, "temporary buffer", MPL_MEM_BUFFER);
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
/* copy local data into recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, count, datatype, recvbuf,
count, datatype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
/* get nearest power-of-two less than or equal to comm_size */
pof2 = comm_ptr->pof2;
rem = comm_size - pof2;
/* In the non-power-of-two case, all even-numbered
processes of rank < 2*rem send their data to
(rank+1). These even-numbered processes no longer
participate in the algorithm until the very end. The
remaining processes form a nice power-of-two. */
if (rank < 2*rem) {
if (rank % 2 == 0) { /* even */
mpi_errno = MPIC_Send(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_TAG, comm_ptr, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
doubling */
newrank = -1;
}
else { /* odd */
mpi_errno = MPIC_Recv(tmp_buf, count,
datatype, rank-1,
MPIR_ALLREDUCE_TAG, comm_ptr,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* do the reduction on received data. since the
ordering is right, it doesn't matter whether
the operation is commutative or not. */
mpi_errno = MPIR_Reduce_local(tmp_buf, recvbuf, count, datatype, op);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* change the rank */
newrank = rank / 2;
}
}
else /* rank >= 2*rem */
newrank = rank - rem;
/* If op is user-defined or count is less than pof2, use
recursive doubling algorithm. Otherwise do a reduce-scatter
followed by allgather. (If op is user-defined,
derived datatypes are allowed and the user could pass basic
datatypes on one process and derived on another as long as
the type maps are the same. Breaking up derived
datatypes to do the reduce-scatter is tricky, therefore
using recursive doubling in that case.) */
#ifdef HAVE_ERROR_CHECKING
MPIR_Assert(HANDLE_GET_KIND(op)==HANDLE_KIND_BUILTIN);
MPIR_Assert(count >= pof2);
#endif /* HAVE_ERROR_CHECKING */
if (newrank != -1) {
MPIR_CHKLMEM_MALLOC(cnts, int *, pof2*sizeof(int), mpi_errno, "counts", MPL_MEM_BUFFER);
MPIR_CHKLMEM_MALLOC(disps, int *, pof2*sizeof(int), mpi_errno, "displacements", MPL_MEM_BUFFER);
for (i=0; i<pof2; i++)
cnts[i] = count/pof2;
if ((count % pof2) > 0) {
for (i=0; i<(count % pof2); i++)
cnts[i] += 1;
}
disps[0] = 0;
for (i=1; i<pof2; i++)
disps[i] = disps[i-1] + cnts[i-1];
mask = 0x1;
send_idx = recv_idx = 0;
last_idx = pof2;
while (mask < pof2) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
send_idx = recv_idx + pof2/(mask*2);
for (i=send_idx; i<last_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += cnts[i];
}
else {
recv_idx = send_idx + pof2/(mask*2);
for (i=send_idx; i<recv_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += cnts[i];
}
/* Send data from recvbuf. Recv into tmp_buf */
mpi_errno = MPIC_Sendrecv((char *) recvbuf +
disps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_ALLREDUCE_TAG,
(char *) tmp_buf +
disps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_ALLREDUCE_TAG, comm_ptr,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
/* tmp_buf contains data received in this step.
recvbuf contains data accumulated so far */
/* This algorithm is used only for predefined ops
and predefined ops are always commutative. */
mpi_errno = MPIR_Reduce_local(((char *) tmp_buf + disps[recv_idx]*extent),
((char *) recvbuf + disps[recv_idx]*extent),
recv_cnt, datatype, op);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* update send_idx for next iteration */
send_idx = recv_idx;
mask <<= 1;
/* update last_idx, but not in last iteration
because the value is needed in the allgather
step below. */
if (mask < pof2)
last_idx = recv_idx + pof2/mask;
}
/* now do the allgather */
mask >>= 1;
while (mask > 0) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
/* update last_idx except on first iteration */
if (mask != pof2/2)
last_idx = last_idx + pof2/(mask*2);
recv_idx = send_idx + pof2/(mask*2);
for (i=send_idx; i<recv_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += cnts[i];
}
else {
recv_idx = send_idx - pof2/(mask*2);
for (i=send_idx; i<last_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += cnts[i];
}
mpi_errno = MPIC_Sendrecv((char *) recvbuf +
disps[send_idx]*extent,
send_cnt, datatype,
dst, MPIR_ALLREDUCE_TAG,
(char *) recvbuf +
disps[recv_idx]*extent,
recv_cnt, datatype, dst,
MPIR_ALLREDUCE_TAG, comm_ptr,
MPI_STATUS_IGNORE, errflag);
if (mpi_errno) {
/* for communication errors, just record the error but continue */
*errflag = MPIR_ERR_GET_CLASS(mpi_errno);
MPIR_ERR_SET(mpi_errno, *errflag, "**fail");
MPIR_ERR_ADD(mpi_errno_ret, mpi_errno);
}
if (newrank > newdst) send_idx = recv_idx;
mask >>= 1;
}
}
/*@
MPI_Fetch_and_op - Perform one-sided read-modify-write.
Accumulate one element of type datatype from the origin buffer (origin_addr) to
the buffer at offset target_disp, in the target window specified by target_rank
and win, using the operation op and return in the result buffer result_addr the
content of the target buffer before the accumulation.
Input Parameters:
+ origin_addr - initial address of buffer (choice)
. result_addr - initial address of result buffer (choice)
. datatype - datatype of the entry in origin, result, and target buffers (handle)
. target_rank - rank of target (nonnegative integer)
. target_disp - displacement from start of window to beginning of target buffer (non-negative integer)
. op - reduce operation (handle)
- win - window object (handle)
Notes:
This operations is atomic with respect to other "accumulate" operations.
The generic functionality of 'MPI_Get_accumulate' might limit the performance of
fetch-and-increment or fetch-and-add calls that might be supported by special
hardware operations. 'MPI_Fetch_and_op' thus allows for a fast implementation
of a commonly used subset of the functionality of 'MPI_Get_accumulate'.
The origin and result buffers (origin_addr and result_addr) must be disjoint.
Any of the predefined operations for 'MPI_Reduce', as well as 'MPI_NO_OP' or
'MPI_REPLACE', can be specified as op; user-defined functions cannot be used. The
datatype argument must be a predefined datatype.
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_ARG
.N MPI_ERR_COUNT
.N MPI_ERR_OP
.N MPI_ERR_RANK
.N MPI_ERR_TYPE
.N MPI_ERR_WIN
.seealso: MPI_Get_accumulate
@*/
int MPI_Fetch_and_op(const void *origin_addr, void *result_addr,
MPI_Datatype datatype, int target_rank, MPI_Aint target_disp,
MPI_Op op, MPI_Win win)
{
static const char FCNAME[] = "MPI_Fetch_and_op";
int mpi_errno = MPI_SUCCESS;
MPIR_Win *win_ptr = NULL;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_FETCH_AND_OP);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPIR_FUNC_TERSE_RMA_ENTER(MPID_STATE_MPI_FETCH_AND_OP);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_WIN(win, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* Convert MPI object handles to object pointers */
MPIR_Win_get_ptr( win, win_ptr );
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_Comm *comm_ptr;
/* Validate win_ptr */
MPIR_Win_valid_ptr( win_ptr, mpi_errno );
if (mpi_errno) goto fn_fail;
if (op != MPI_NO_OP) {
/* NOTE: when op is MPI_NO_OP, origin_addr is allowed to be NULL.
* In such case, MPI_Fetch_and_op equals to an atomic GET. */
MPIR_ERRTEST_ARGNULL(origin_addr, "origin_addr", mpi_errno);
}
MPIR_ERRTEST_ARGNULL(result_addr, "result_addr", mpi_errno);
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
if (!MPIR_DATATYPE_IS_PREDEFINED(datatype))
{
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_TYPE, "**typenotpredefined");
}
if (win_ptr->create_flavor != MPI_WIN_FLAVOR_DYNAMIC)
MPIR_ERRTEST_DISP(target_disp, mpi_errno);
comm_ptr = win_ptr->comm_ptr;
MPIR_ERRTEST_SEND_RANK(comm_ptr, target_rank, mpi_errno);
MPIR_ERRTEST_OP_GACC(op, mpi_errno);
if (HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN)
{
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OP, "**opnotpredefined");
}
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPID_Fetch_and_op(origin_addr,
result_addr, datatype,
target_rank, target_disp,
op, win_ptr);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* ... end of body of routine ... */
fn_exit:
MPIR_FUNC_TERSE_RMA_EXIT(MPID_STATE_MPI_FETCH_AND_OP);
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_fetch_and_op",
"**mpi_fetch_and_op %p %p %D %d %d %O %W", origin_addr,
result_addr, datatype, target_rank, target_disp, op, win);
}
# endif
mpi_errno = MPIR_Err_return_win( win_ptr, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Recv - Blocking receive for a message
Output Parameters:
+ buf - initial address of receive buffer (choice)
- status - status object (Status)
Input Parameters:
+ count - maximum number of elements in receive buffer (integer)
. datatype - datatype of each receive buffer element (handle)
. source - rank of source (integer)
. tag - message tag (integer)
- comm - communicator (handle)
Notes:
The 'count' argument indicates the maximum length of a message; the actual
length of the message can be determined with 'MPI_Get_count'.
.N ThreadSafe
.N Fortran
.N FortranStatus
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_COMM
.N MPI_ERR_TYPE
.N MPI_ERR_COUNT
.N MPI_ERR_TAG
.N MPI_ERR_RANK
@*/
int MPI_Recv(void *buf, int count, MPI_Datatype datatype, int source, int tag,
MPI_Comm comm, MPI_Status *status)
{
static const char FCNAME[] = "MPI_Recv";
int mpi_errno = MPI_SUCCESS;
MPID_Comm *comm_ptr = NULL;
MPID_Request * request_ptr = NULL;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_RECV);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPIU_THREAD_CS_ENTER(ALLFUNC,);
MPID_MPI_PT2PT_FUNC_ENTER_BACK(MPID_STATE_MPI_RECV);
/* Validate handle parameters needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
/* NOTE: MPI_STATUS_IGNORE != NULL */
MPIR_ERRTEST_ARGNULL(status, "status", mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* Convert MPI object handles to object pointers */
MPID_Comm_get_ptr( comm, comm_ptr );
/* Validate parameters if error checking is enabled */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPID_Comm_valid_ptr( comm_ptr, mpi_errno, FALSE );
if (mpi_errno) goto fn_fail;
MPIR_ERRTEST_COUNT(count, mpi_errno);
MPIR_ERRTEST_RECV_RANK(comm_ptr, source, mpi_errno);
MPIR_ERRTEST_RECV_TAG(tag, mpi_errno);
/* Validate datatype handle */
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
/* Validate datatype object */
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN)
{
MPID_Datatype *datatype_ptr = NULL;
MPID_Datatype_get_ptr(datatype, datatype_ptr);
MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno);
if (mpi_errno) goto fn_fail;
MPID_Datatype_committed_ptr(datatype_ptr, mpi_errno);
if (mpi_errno) goto fn_fail;
}
/* Validate buffer */
MPIR_ERRTEST_USERBUFFER(buf,count,datatype,mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
/* MT: Note that MPID_Recv may release the SINGLE_CS if it
decides to block internally. MPID_Recv in that case will
re-aquire the SINGLE_CS before returnning */
mpi_errno = MPID_Recv(buf, count, datatype, source, tag, comm_ptr,
MPID_CONTEXT_INTRA_PT2PT, status, &request_ptr);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
if (request_ptr == NULL)
{
goto fn_exit;
}
/* If a request was returned, then we need to block until the request is
complete */
if (!MPID_Request_is_complete(request_ptr))
{
MPID_Progress_state progress_state;
MPID_Progress_start(&progress_state);
while (!MPID_Request_is_complete(request_ptr))
{
/* MT: Progress_wait may release the SINGLE_CS while it
waits */
mpi_errno = MPID_Progress_wait(&progress_state);
if (mpi_errno != MPI_SUCCESS)
{
/* --BEGIN ERROR HANDLING-- */
MPID_Progress_end(&progress_state);
goto fn_fail;
/* --END ERROR HANDLING-- */
}
if (unlikely(MPIR_CVAR_ENABLE_FT &&
!MPID_Request_is_complete(request_ptr) &&
MPID_Request_is_anysource(request_ptr) &&
!MPID_Comm_AS_enabled(request_ptr->comm))) {
/* --BEGIN ERROR HANDLING-- */
MPID_Cancel_recv(request_ptr);
MPIR_STATUS_SET_CANCEL_BIT(request_ptr->status, FALSE);
MPIU_ERR_SET(request_ptr->status.MPI_ERROR, MPIX_ERR_PROC_FAILED, "**proc_failed");
mpi_errno = request_ptr->status.MPI_ERROR;
goto fn_fail;
/* --END ERROR HANDLING-- */
}
}
MPID_Progress_end(&progress_state);
}
mpi_errno = request_ptr->status.MPI_ERROR;
MPIR_Request_extract_status(request_ptr, status);
MPID_Request_release(request_ptr);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_PT2PT_FUNC_EXIT_BACK(MPID_STATE_MPI_RECV);
MPIU_THREAD_CS_EXIT(ALLFUNC,);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_recv",
"**mpi_recv %p %d %D %i %t %C %p", buf, count, datatype, source, tag, comm, status);
}
# endif
mpi_errno = MPIR_Err_return_comm( comm_ptr, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/* MPIR_Type_get_elements
*
* Arguments:
* - bytes_p - input/output byte count
* - count - maximum number of this type to subtract from the bytes; a count
* of <0 indicates use as many as we like
* - datatype - input datatype
*
* Returns number of elements available given the two constraints of number of
* bytes and count of types. Also reduces the byte count by the amount taken
* up by the types.
*
* This is called from MPI_Get_elements() when it sees a type with multiple
* element types (datatype_ptr->element_sz = -1). This function calls itself too.
*/
PMPI_LOCAL MPI_Count MPIR_Type_get_elements(MPI_Count *bytes_p,
MPI_Count count,
MPI_Datatype datatype)
{
MPIR_Datatype *datatype_ptr = NULL;
MPID_Datatype_get_ptr(datatype, datatype_ptr); /* invalid if builtin */
/* if we have gotten down to a type with only one element type,
* call MPIR_Type_get_basic_type_elements() and return.
*/
if (HANDLE_GET_KIND(datatype) == HANDLE_KIND_BUILTIN ||
datatype == MPI_FLOAT_INT ||
datatype == MPI_DOUBLE_INT ||
datatype == MPI_LONG_INT ||
datatype == MPI_SHORT_INT ||
datatype == MPI_LONG_DOUBLE_INT)
{
return MPIR_Type_get_basic_type_elements(bytes_p, count, datatype);
}
else if (datatype_ptr->builtin_element_size >= 0) {
MPI_Datatype basic_type = MPI_DATATYPE_NULL;
MPID_Datatype_get_basic_type(datatype_ptr->basic_type, basic_type);
return MPIR_Type_get_basic_type_elements(bytes_p,
count * datatype_ptr->n_builtin_elements,
basic_type);
}
else {
/* we have bytes left and still don't have a single element size; must
* recurse.
*/
int i, j, *ints;
MPI_Count typecount = 0, nr_elements = 0, last_nr_elements;
MPI_Aint *aints;
MPI_Datatype *types;
/* Establish locations of arrays */
MPID_Type_access_contents(datatype_ptr->handle, &ints, &aints, &types);
if (!ints || !aints || !types)
return MPI_ERR_TYPE;
switch (datatype_ptr->contents->combiner) {
case MPI_COMBINER_NAMED:
case MPI_COMBINER_DUP:
case MPI_COMBINER_RESIZED:
return MPIR_Type_get_elements(bytes_p, count, *types);
break;
case MPI_COMBINER_CONTIGUOUS:
case MPI_COMBINER_VECTOR:
case MPI_COMBINER_HVECTOR_INTEGER:
case MPI_COMBINER_HVECTOR:
/* count is first in ints array */
return MPIR_Type_get_elements(bytes_p, count * (*ints), *types);
break;
case MPI_COMBINER_INDEXED_BLOCK:
case MPI_COMBINER_HINDEXED_BLOCK:
/* count is first in ints array, blocklength is second */
return MPIR_Type_get_elements(bytes_p,
count * ints[0] * ints[1],
*types);
break;
case MPI_COMBINER_INDEXED:
case MPI_COMBINER_HINDEXED_INTEGER:
case MPI_COMBINER_HINDEXED:
for (i=0; i < (*ints); i++) {
/* add up the blocklengths to get a max. # of the next type */
typecount += ints[i+1];
}
return MPIR_Type_get_elements(bytes_p, count * typecount, *types);
break;
case MPI_COMBINER_STRUCT_INTEGER:
case MPI_COMBINER_STRUCT:
/* In this case we can't simply multiply the count of the next
* type by the count of the current type, because we need to
* cycle through the types just as the struct would. thus the
* nested loops.
*
* We need to keep going until we get less elements than expected
* or we run out of bytes.
*/
last_nr_elements = 1; /* seed value */
for (j=0;
(count < 0 || j < count) &&
*bytes_p > 0 && last_nr_elements > 0;
j++)
{
/* recurse on each type; bytes are reduced in calls */
for (i=0; i < (*ints); i++) {
/* skip zero-count elements of the struct */
if (ints[i+1] == 0) continue;
last_nr_elements = MPIR_Type_get_elements(bytes_p,
ints[i+1],
types[i]);
nr_elements += last_nr_elements;
MPIR_Assert(last_nr_elements >= 0);
if (last_nr_elements < ints[i+1]) break;
}
}
return nr_elements;
break;
case MPI_COMBINER_SUBARRAY:
case MPI_COMBINER_DARRAY:
case MPI_COMBINER_F90_REAL:
case MPI_COMBINER_F90_COMPLEX:
case MPI_COMBINER_F90_INTEGER:
default:
/* --BEGIN ERROR HANDLING-- */
MPIR_Assert(0);
return -1;
break;
/* --END ERROR HANDLING-- */
}
}
}
/*@
MPI_Type_hindexed - Creates an indexed datatype with offsets in bytes
Input Parameters:
+ count - number of blocks -- also number of entries in indices and blocklens
. blocklens - number of elements in each block (array of nonnegative integers)
. indices - byte displacement of each block (array of MPI_Aint)
- old_type - old datatype (handle)
Output Parameter:
. newtype - new datatype (handle)
.N Deprecated
This routine is replaced by 'MPI_Type_create_hindexed'.
.N ThreadSafe
.N Fortran
The indices are displacements, and are based on a zero origin. A common error
is to do something like to following
.vb
integer a(100)
integer blens(10), indices(10)
do i=1,10
blens(i) = 1
10 indices(i) = (1 + (i-1)*10) * sizeofint
call MPI_TYPE_HINDEXED(10,blens,indices,MPI_INTEGER,newtype,ierr)
call MPI_TYPE_COMMIT(newtype,ierr)
call MPI_SEND(a,1,newtype,...)
.ve
expecting this to send 'a(1),a(11),...' because the indices have values
'1,11,...'. Because these are `displacements` from the beginning of 'a',
it actually sends 'a(1+1),a(1+11),...'.
If you wish to consider the displacements as indices into a Fortran array,
consider declaring the Fortran array with a zero origin
.vb
integer a(0:99)
.ve
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_TYPE
.N MPI_ERR_COUNT
.N MPI_ERR_EXHAUSTED
.N MPI_ERR_ARG
@*/
int MPI_Type_hindexed(int count,
int blocklens[],
MPI_Aint indices[],
MPI_Datatype old_type,
MPI_Datatype *newtype)
{
static const char FCNAME[] = "MPI_Type_hindexed";
int mpi_errno = MPI_SUCCESS;
MPI_Datatype new_handle;
MPID_Datatype *new_dtp;
int i, *ints;
MPIU_CHKLMEM_DECL(1);
MPID_MPI_STATE_DECL(MPID_STATE_MPI_TYPE_HINDEXED);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPIU_THREAD_CS_ENTER(ALLFUNC,);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_TYPE_HINDEXED);
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
int j;
MPID_Datatype *datatype_ptr = NULL;
MPIR_ERRTEST_COUNT(count, mpi_errno);
MPIR_ERRTEST_DATATYPE(old_type, "datatype", mpi_errno);
if (count > 0) {
MPIR_ERRTEST_ARGNULL(blocklens, "blocklens", mpi_errno);
MPIR_ERRTEST_ARGNULL(indices, "indices", mpi_errno);
}
if (mpi_errno == MPI_SUCCESS) {
if (HANDLE_GET_KIND(old_type) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr( old_type, datatype_ptr );
MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno);
}
/* verify that all blocklengths are >= 0 */
for (j=0; j < count; j++) {
MPIR_ERRTEST_ARGNEG(blocklens[j], "blocklen", mpi_errno);
}
}
MPIR_ERRTEST_ARGNULL(newtype, "newtype", mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPID_Type_indexed(count,
blocklens,
indices,
1, /* displacements in bytes */
old_type,
&new_handle);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPIU_CHKLMEM_MALLOC(ints, int *, (count + 1) * sizeof(int), mpi_errno, "contents integer array");
/* copy ints into temporary buffer (count and blocklengths) */
ints[0] = count;
for (i=0; i < count; i++)
{
ints[i+1] = blocklens[i];
}
MPID_Datatype_get_ptr(new_handle, new_dtp);
mpi_errno = MPID_Datatype_set_contents(new_dtp,
MPI_COMBINER_HINDEXED,
count+1, /* ints */
count, /* aints (displs) */
1, /* types */
ints,
indices,
&old_type);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPIU_OBJ_PUBLISH_HANDLE(*newtype, new_handle);
/* ... end of body of routine ... */
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_TYPE_HINDEXED);
MPIU_THREAD_CS_EXIT(ALLFUNC,);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_type_hindexed",
"**mpi_type_hindexed %d %p %p %D %p", count, blocklens, indices, old_type, newtype);
}
# endif
mpi_errno = MPIR_Err_return_comm( NULL, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIR_Get_elements_x_impl(const MPI_Status *status, MPI_Datatype datatype, MPI_Count *elements)
{
int mpi_errno = MPI_SUCCESS;
MPIR_Datatype *datatype_ptr = NULL;
MPI_Count byte_count;
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr(datatype, datatype_ptr);
}
/* three cases:
* - nice, simple, single element type
* - derived type with a zero size
* - type with multiple element types (nastiest)
*/
if (HANDLE_GET_KIND(datatype) == HANDLE_KIND_BUILTIN ||
(datatype_ptr->builtin_element_size != -1 && datatype_ptr->size > 0))
{
byte_count = MPIR_STATUS_GET_COUNT(*status);
/* QUESTION: WHAT IF SOMEONE GAVE US AN MPI_UB OR MPI_LB???
*/
/* in both cases we do not limit the number of types that might
* be in bytes
*/
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
MPI_Datatype basic_type = MPI_DATATYPE_NULL;
MPID_Datatype_get_basic_type(datatype_ptr->basic_type, basic_type);
*elements = MPIR_Type_get_basic_type_elements(&byte_count,
-1,
basic_type);
}
else {
/* Behaves just like MPI_Get_Count in the predefined case */
MPI_Count size;
MPID_Datatype_get_size_macro(datatype, size);
if ((byte_count % size) != 0)
*elements = MPI_UNDEFINED;
else
*elements = MPIR_Type_get_basic_type_elements(&byte_count,
-1,
datatype);
}
MPIR_Assert(byte_count >= 0);
}
else if (datatype_ptr->size == 0) {
if (MPIR_STATUS_GET_COUNT(*status) > 0) {
/* --BEGIN ERROR HANDLING-- */
/* datatype size of zero and count > 0 should never happen. */
(*elements) = MPI_UNDEFINED;
/* --END ERROR HANDLING-- */
}
else {
/* This is ambiguous. However, discussions on MPI Forum
* reached a consensus that this is the correct return
* value
*/
(*elements) = 0;
}
}
else /* derived type with weird element type or weird size */ {
MPIR_Assert(datatype_ptr->builtin_element_size == -1);
byte_count = MPIR_STATUS_GET_COUNT(*status);
*elements = MPIR_Type_get_elements(&byte_count, -1, datatype);
}
return mpi_errno;
}
/*@
MPI_Errhandler_set - Sets the error handler for a communicator
Input Parameters:
+ comm - communicator to set the error handler for (handle)
- errhandler - new MPI error handler for communicator (handle)
.N ThreadSafe
.N Deprecated
The replacement for this routine is 'MPI_Comm_set_errhandler'.
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_COMM
.N MPI_ERR_ARG
.seealso: MPI_Comm_set_errhandler, MPI_Errhandler_create, MPI_Comm_create_errhandler
@*/
int MPI_Errhandler_set(MPI_Comm comm, MPI_Errhandler errhandler)
{
int mpi_errno = MPI_SUCCESS;
MPIR_Comm *comm_ptr = NULL;
MPIR_Errhandler *errhan_ptr;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_ERRHANDLER_SET);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPI_ERRHANDLER_SET);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif
/* Convert MPI object handles to object pointers */
MPIR_Comm_get_ptr( comm, comm_ptr );
MPIR_Errhandler_get_ptr( errhandler, errhan_ptr );
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
/* Validate comm_ptr; if comm_ptr is not value, it will be reset to null */
MPIR_Comm_valid_ptr( comm_ptr, mpi_errno, TRUE );
if (mpi_errno) goto fn_fail;
MPIR_ERRTEST_ERRHANDLER(errhandler, mpi_errno);
if (HANDLE_GET_KIND(errhandler) != HANDLE_KIND_BUILTIN) {
MPIR_Errhandler_valid_ptr( errhan_ptr, mpi_errno );
if (mpi_errno) goto fn_fail;
}
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
MPIR_Comm_set_errhandler_impl(comm_ptr, errhan_ptr);
/* ... end of body of routine ... */
# ifdef HAVE_ERROR_CHECKING
fn_exit:
# endif
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPI_ERRHANDLER_SET);
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
return mpi_errno;
# ifdef HAVE_ERROR_CHECKING
fn_fail:
/* --BEGIN ERROR HANDLING-- */
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_errhandler_set",
"**mpi_errhandler_set %C %E", comm, errhandler);
}
mpi_errno = MPIR_Err_return_comm( comm_ptr, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
# endif
}
int MPIR_Iallreduce_redscat_allgather(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPID_Comm *comm_ptr, MPID_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int comm_size, rank, newrank, pof2, rem;
int i, send_idx, recv_idx, last_idx, mask, newdst, dst, send_cnt, recv_cnt;
MPI_Aint true_lb, true_extent, extent;
void *tmp_buf = NULL;
int *cnts, *disps;
MPIR_SCHED_CHKPMEM_DECL(1);
MPIU_CHKLMEM_DECL(2);
/* we only support builtin datatypes for now, breaking up user types to do
* the reduce-scatter is tricky */
MPIU_Assert(HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
/* need to allocate temporary buffer to store incoming data*/
MPIR_Type_get_true_extent_impl(datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(datatype, extent);
MPID_Ensure_Aint_fits_in_pointer(count * MPIR_MAX(extent, true_extent));
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, count*(MPIR_MAX(extent,true_extent)), mpi_errno, "temporary buffer");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
/* copy local data into recvbuf */
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPID_Sched_copy(sendbuf, count, datatype,
recvbuf, count, datatype, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
}
/* find nearest power-of-two less than or equal to comm_size */
pof2 = 1;
while (pof2 <= comm_size) pof2 <<= 1;
pof2 >>=1;
rem = comm_size - pof2;
/* In the non-power-of-two case, all even-numbered
processes of rank < 2*rem send their data to
(rank+1). These even-numbered processes no longer
participate in the algorithm until the very end. The
remaining processes form a nice power-of-two. */
if (rank < 2*rem) {
if (rank % 2 == 0) { /* even */
mpi_errno = MPID_Sched_send(recvbuf, count, datatype, rank+1, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
doubling */
newrank = -1;
}
else { /* odd */
mpi_errno = MPID_Sched_recv(tmp_buf, count, datatype, rank-1, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* do the reduction on received data. since the
ordering is right, it doesn't matter whether
the operation is commutative or not. */
mpi_errno = MPID_Sched_reduce(tmp_buf, recvbuf, count, datatype, op, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* change the rank */
newrank = rank / 2;
}
}
else /* rank >= 2*rem */
newrank = rank - rem;
if (newrank != -1) {
/* for the reduce-scatter, calculate the count that
each process receives and the displacement within
the buffer */
/* TODO I (goodell@) believe that these counts and displacements could be
* calculated directly during the loop, rather than requiring a less-scalable
* "2*pof2"-sized memory allocation */
MPIU_CHKLMEM_MALLOC(cnts, int *, pof2*sizeof(int), mpi_errno, "counts");
MPIU_CHKLMEM_MALLOC(disps, int *, pof2*sizeof(int), mpi_errno, "displacements");
MPIU_Assert(count >= pof2); /* the cnts calculations assume this */
for (i=0; i<(pof2-1); i++)
cnts[i] = count/pof2;
cnts[pof2-1] = count - (count/pof2)*(pof2-1);
disps[0] = 0;
for (i=1; i<pof2; i++)
disps[i] = disps[i-1] + cnts[i-1];
mask = 0x1;
send_idx = recv_idx = 0;
last_idx = pof2;
while (mask < pof2) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
send_idx = recv_idx + pof2/(mask*2);
for (i=send_idx; i<last_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += cnts[i];
}
else {
recv_idx = send_idx + pof2/(mask*2);
for (i=send_idx; i<recv_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += cnts[i];
}
/* Send data from recvbuf. Recv into tmp_buf */
mpi_errno = MPID_Sched_recv(((char *)tmp_buf + disps[recv_idx]*extent),
recv_cnt, datatype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* sendrecv, no barrier here */
mpi_errno = MPID_Sched_send(((char *)recvbuf + disps[send_idx]*extent),
send_cnt, datatype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* tmp_buf contains data received in this step.
recvbuf contains data accumulated so far */
/* This algorithm is used only for predefined ops
and predefined ops are always commutative. */
mpi_errno = MPID_Sched_reduce(((char *)tmp_buf + disps[recv_idx]*extent),
((char *)recvbuf + disps[recv_idx]*extent),
recv_cnt, datatype, op, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
/* update send_idx for next iteration */
send_idx = recv_idx;
mask <<= 1;
/* update last_idx, but not in last iteration
because the value is needed in the allgather
step below. */
if (mask < pof2)
last_idx = recv_idx + pof2/mask;
}
/* now do the allgather */
mask >>= 1;
while (mask > 0) {
newdst = newrank ^ mask;
/* find real rank of dest */
dst = (newdst < rem) ? newdst*2 + 1 : newdst + rem;
send_cnt = recv_cnt = 0;
if (newrank < newdst) {
/* update last_idx except on first iteration */
if (mask != pof2/2)
last_idx = last_idx + pof2/(mask*2);
recv_idx = send_idx + pof2/(mask*2);
for (i=send_idx; i<recv_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<last_idx; i++)
recv_cnt += cnts[i];
}
else {
recv_idx = send_idx - pof2/(mask*2);
for (i=send_idx; i<last_idx; i++)
send_cnt += cnts[i];
for (i=recv_idx; i<send_idx; i++)
recv_cnt += cnts[i];
}
mpi_errno = MPID_Sched_recv(((char *)recvbuf + disps[recv_idx]*extent),
recv_cnt, datatype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* sendrecv, no barrier here */
mpi_errno = MPID_Sched_send(((char *)recvbuf + disps[send_idx]*extent),
send_cnt, datatype, dst, comm_ptr, s);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
if (newrank > newdst) send_idx = recv_idx;
mask >>= 1;
}
}
int MPID_nem_mxm_issend(MPIDI_VC_t * vc, const void *buf, int count, MPI_Datatype datatype,
int rank, int tag, MPID_Comm * comm, int context_offset,
MPID_Request ** sreq_ptr)
{
int mpi_errno = MPI_SUCCESS;
MPID_Request *sreq = NULL;
MPID_Datatype *dt_ptr;
int dt_contig;
MPIDI_msg_sz_t data_sz;
MPI_Aint dt_true_lb;
MPID_nem_mxm_vc_area *vc_area = NULL;
MPID_nem_mxm_req_area *req_area = NULL;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_MXM_ISSEND);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_MXM_ISSEND);
MPIDI_Datatype_get_info(count, datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
/* create a request */
MPIDI_Request_create_sreq(sreq, mpi_errno, goto fn_exit);
MPIU_Assert(sreq != NULL);
MPIDI_Request_set_type(sreq, MPIDI_REQUEST_TYPE_SEND);
MPIDI_VC_FAI_send_seqnum(vc, seqnum);
MPIDI_Request_set_seqnum(sreq, seqnum);
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr(datatype, sreq->dev.datatype_ptr);
MPID_Datatype_add_ref(sreq->dev.datatype_ptr);
}
sreq->partner_request = NULL;
sreq->dev.OnDataAvail = NULL;
sreq->dev.tmpbuf = NULL;
sreq->ch.vc = vc;
sreq->ch.noncontig = FALSE;
_dbg_mxm_output(5,
"isSend ========> Sending USER msg for req %p (context %d to %d tag %d size %d) \n",
sreq, comm->context_id + context_offset, rank, tag, data_sz);
vc_area = VC_BASE(vc);
req_area = REQ_BASE(sreq);
req_area-> ctx = sreq;
req_area->iov_buf = req_area->tmp_buf;
req_area->iov_count = 0;
req_area->iov_buf[0].ptr = NULL;
req_area->iov_buf[0].length = 0;
if (data_sz) {
if (dt_contig) {
req_area->iov_count = 1;
req_area->iov_buf[0].ptr = (char *) (buf) + dt_true_lb;
req_area->iov_buf[0].length = data_sz;
}
else {
MPIDI_msg_sz_t last;
MPI_Aint packsize = 0;
sreq->ch.noncontig = TRUE;
sreq->dev.segment_ptr = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP1((sreq->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER,
"**nomem", "**nomem %s", "MPID_Segment_alloc");
MPIR_Pack_size_impl(count, datatype, &packsize);
last = data_sz;
if (packsize > 0) {
sreq->dev.tmpbuf = MPIU_Malloc((size_t) packsize);
MPIU_Assert(sreq->dev.tmpbuf);
MPID_Segment_init(buf, count, datatype, sreq->dev.segment_ptr, 0);
MPID_Segment_pack(sreq->dev.segment_ptr, 0, &last, sreq->dev.tmpbuf);
req_area->iov_count = 1;
req_area->iov_buf[0].ptr = sreq->dev.tmpbuf;
req_area->iov_buf[0].length = last;
}
}
}
vc_area->pending_sends += 1;
mpi_errno = _mxm_isend(vc_area->mxm_ep, req_area, MXM_MPICH_ISEND_SYNC,
(mxm_mq_h) comm->dev.ch.netmod_priv, comm->rank, tag, _mxm_tag_mpi2mxm(tag,
comm->context_id
+
context_offset),
0);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
_dbg_mxm_out_req(sreq);
fn_exit:
*sreq_ptr = sreq;
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_MXM_ISSEND);
return mpi_errno;
fn_fail:
goto fn_exit;
}
/*@
MPI_Alltoallv - Sends data from all to all processes; each process may
send a different amount of data and provide displacements for the input
and output data.
Input Parameters:
+ sendbuf - starting address of send buffer (choice)
. sendcounts - integer array equal to the group size
specifying the number of elements to send to each processor
. sdispls - integer array (of length group size). Entry
'j' specifies the displacement (relative to sendbuf from
which to take the outgoing data destined for process 'j'
. sendtype - data type of send buffer elements (handle)
. recvcounts - integer array equal to the group size
specifying the maximum number of elements that can be received from
each processor
. rdispls - integer array (of length group size). Entry
'i' specifies the displacement (relative to recvbuf at
which to place the incoming data from process 'i'
. recvtype - data type of receive buffer elements (handle)
- comm - communicator (handle)
Output Parameters:
. recvbuf - address of receive buffer (choice)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_ERR_COMM
.N MPI_ERR_COUNT
.N MPI_ERR_TYPE
.N MPI_ERR_BUFFER
@*/
int MPI_Alltoallv(const void *sendbuf, const int *sendcounts,
const int *sdispls, MPI_Datatype sendtype, void *recvbuf,
const int *recvcounts, const int *rdispls, MPI_Datatype recvtype,
MPI_Comm comm)
{
int mpi_errno = MPI_SUCCESS;
MPID_Comm *comm_ptr = NULL;
MPIR_Errflag_t errflag = MPIR_ERR_NONE;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_ALLTOALLV);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPID_MPI_COLL_FUNC_ENTER(MPID_STATE_MPI_ALLTOALLV);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* Convert MPI object handles to object pointers */
MPID_Comm_get_ptr( comm, comm_ptr );
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPID_Datatype *sendtype_ptr=NULL, *recvtype_ptr=NULL;
int i, comm_size;
int check_send = (comm_ptr->comm_kind == MPID_INTRACOMM && sendbuf != MPI_IN_PLACE);
MPID_Comm_valid_ptr( comm_ptr, mpi_errno, FALSE );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
if (comm_ptr->comm_kind == MPID_INTRACOMM) {
comm_size = comm_ptr->local_size;
if (sendbuf != MPI_IN_PLACE && sendtype == recvtype && sendcounts == recvcounts)
MPIR_ERRTEST_ALIAS_COLL(sendbuf, recvbuf, mpi_errno);
} else
comm_size = comm_ptr->remote_size;
if (comm_ptr->comm_kind == MPID_INTERCOMM && sendbuf == MPI_IN_PLACE) {
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**sendbuf_inplace");
}
for (i=0; i<comm_size; i++) {
if (check_send) {
MPIR_ERRTEST_COUNT(sendcounts[i], mpi_errno);
MPIR_ERRTEST_DATATYPE(sendtype, "sendtype", mpi_errno);
}
MPIR_ERRTEST_COUNT(recvcounts[i], mpi_errno);
MPIR_ERRTEST_DATATYPE(recvtype, "recvtype", mpi_errno);
}
if (check_send && HANDLE_GET_KIND(sendtype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr(sendtype, sendtype_ptr);
MPID_Datatype_valid_ptr( sendtype_ptr, mpi_errno );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPID_Datatype_committed_ptr( sendtype_ptr, mpi_errno );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
if (HANDLE_GET_KIND(recvtype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr(recvtype, recvtype_ptr);
MPID_Datatype_valid_ptr( recvtype_ptr, mpi_errno );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPID_Datatype_committed_ptr( recvtype_ptr, mpi_errno );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
for (i=0; i<comm_size && check_send; i++) {
if (sendcounts[i] > 0) {
MPIR_ERRTEST_USERBUFFER(sendbuf,sendcounts[i],sendtype,mpi_errno);
}
}
for (i=0; i<comm_size; i++) {
if (recvcounts[i] > 0) {
MPIR_ERRTEST_RECVBUF_INPLACE(recvbuf, recvcounts[i], mpi_errno);
MPIR_ERRTEST_USERBUFFER(recvbuf,recvcounts[i],recvtype,mpi_errno);
break;
}
}
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPIR_Alltoallv_impl(sendbuf, sendcounts, sdispls,
sendtype, recvbuf, recvcounts,
rdispls, recvtype, comm_ptr, &errflag);
if (mpi_errno) goto fn_fail;
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_COLL_FUNC_EXIT(MPID_STATE_MPI_ALLTOALLV);
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_alltoallv",
"**mpi_alltoallv %p %p %p %D %p %p %p %D %C", sendbuf, sendcounts, sdispls, sendtype,
recvbuf, recvcounts, rdispls, recvtype, comm);
}
# endif
mpi_errno = MPIR_Err_return_comm( comm_ptr, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Pack_external - Packs a datatype into contiguous memory, using the
external32 format
Input Parameters:
+ datarep - data representation (string)
. inbuf - input buffer start (choice)
. incount - number of input data items (integer)
. datatype - datatype of each input data item (handle)
- outsize - output buffer size, in bytes (address integer)
Output Parameters:
. outbuf - output buffer start (choice)
Input/Output Parameters:
. position - current position in buffer, in bytes (address integer)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_TYPE
.N MPI_ERR_ARG
.N MPI_ERR_COUNT
@*/
int MPI_Pack_external(const char datarep[],
const void *inbuf,
int incount,
MPI_Datatype datatype,
void *outbuf,
MPI_Aint outsize,
MPI_Aint *position)
{
static const char FCNAME[] = "MPI_Pack_external";
int mpi_errno = MPI_SUCCESS;
MPI_Aint first, last;
MPID_Segment *segp;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_PACK_EXTERNAL);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_PACK_EXTERNAL);
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COUNT(incount, mpi_errno);
MPIR_ERRTEST_COUNT(outsize, mpi_errno);
/* NOTE: inbuf could be null (MPI_BOTTOM) */
if (incount > 0) {
MPIR_ERRTEST_ARGNULL(outbuf, "output buffer", mpi_errno);
}
MPIR_ERRTEST_ARGNULL(position, "position", mpi_errno);
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype *datatype_ptr = NULL;
MPID_Datatype_get_ptr(datatype, datatype_ptr);
MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno);
MPID_Datatype_committed_ptr(datatype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
if (incount == 0) {
goto fn_exit;
}
segp = MPID_Segment_alloc();
/* --BEGIN ERROR HANDLING-- */
if (segp == NULL)
{
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
FCNAME,
__LINE__,
MPI_ERR_OTHER,
"**nomem",
"**nomem %s",
"MPID_Segment");
goto fn_fail;
}
/* --END ERROR HANDLING-- */
mpi_errno = MPID_Segment_init(inbuf, incount, datatype, segp, 1);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* NOTE: the use of buffer values and positions in MPI_Pack_external and
* in MPID_Segment_pack_external are quite different. See code or docs
* or something.
*/
first = 0;
last = SEGMENT_IGNORE_LAST;
/* Ensure that pointer increment fits in a pointer */
MPID_Ensure_Aint_fits_in_pointer((MPI_VOID_PTR_CAST_TO_MPI_AINT outbuf) + *position);
MPID_Segment_pack_external32(segp,
first,
&last,
(void *)((char *) outbuf + *position));
*position += last;
MPID_Segment_free(segp);
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_PACK_EXTERNAL);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_pack_external",
"**mpi_pack_external %s %p %d %D %p %d %p", datarep, inbuf, incount, datatype, outbuf, outsize, position);
}
# endif
mpi_errno = MPIR_Err_return_comm(0, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*@
MPI_Ibsend - Starts a nonblocking buffered send
Input Parameters:
+ buf - initial address of send buffer (choice)
. count - number of elements in send buffer (integer)
. datatype - datatype of each send buffer element (handle)
. dest - rank of destination (integer)
. tag - message tag (integer)
- comm - communicator (handle)
Output Parameters:
. request - communication request (handle)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_COMM
.N MPI_ERR_COUNT
.N MPI_ERR_TYPE
.N MPI_ERR_TAG
.N MPI_ERR_RANK
.N MPI_ERR_BUFFER
@*/
int MPI_Ibsend(const void *buf, int count, MPI_Datatype datatype, int dest, int tag,
MPI_Comm comm, MPI_Request * request)
{
int mpi_errno = MPI_SUCCESS;
MPIR_Comm *comm_ptr = NULL;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_IBSEND);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPIR_FUNC_TERSE_PT2PT_ENTER_FRONT(MPID_STATE_MPI_IBSEND);
/* Validate handle parameters needing to be converted */
#ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
#endif /* HAVE_ERROR_CHECKING */
/* Convert MPI object handles to object pointers */
MPIR_Comm_get_ptr(comm, comm_ptr);
/* Validate parameters if error checking is enabled */
#ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COUNT(count, mpi_errno);
/* Validate comm_ptr */
MPIR_Comm_valid_ptr(comm_ptr, mpi_errno, FALSE);
if (mpi_errno)
goto fn_fail;
/* If comm_ptr is not valid, it will be reset to null */
if (comm_ptr) {
MPIR_ERRTEST_SEND_TAG(tag, mpi_errno);
MPIR_ERRTEST_SEND_RANK(comm_ptr, dest, mpi_errno)
}
/* Validate datatype handle */
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
/* Validate datatype object */
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
MPIR_Datatype *datatype_ptr = NULL;
MPIR_Datatype_get_ptr(datatype, datatype_ptr);
MPIR_Datatype_valid_ptr(datatype_ptr, mpi_errno);
if (mpi_errno)
goto fn_fail;
MPIR_Datatype_committed_ptr(datatype_ptr, mpi_errno);
if (mpi_errno)
goto fn_fail;
}
/* Validate buffer */
MPIR_ERRTEST_USERBUFFER(buf, count, datatype, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
#endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPIR_Ibsend_impl(buf, count, datatype, dest, tag, comm_ptr, request);
if (mpi_errno)
goto fn_fail;
/* ... end of body of routine ... */
fn_exit:
MPIR_FUNC_TERSE_PT2PT_EXIT(MPID_STATE_MPI_IBSEND);
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
/* FIXME: should we be setting the request at all in the case of an error? */
*request = MPI_REQUEST_NULL;
#ifdef HAVE_ERROR_REPORTING
{
mpi_errno =
MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER,
"**mpi_ibsend", "**mpi_ibsend %p %d %D %i %t %C %p", buf, count,
datatype, dest, tag, comm, request);
}
#endif
mpi_errno = MPIR_Err_return_comm(comm_ptr, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
