int MPIR_Type_contiguous_x_impl(MPI_Count count,
MPI_Datatype oldtype,
MPI_Datatype *newtype)
{
/* to make 'count' fit MPI-3 type processing routines (which take integer
* counts), we construct a type consisting of N INT_MAX chunks followed by
* a remainder. e.g for a count of 4000000000 bytes you would end up with
* one 2147483647-byte chunk followed immediately by a 1852516353-byte
* chunk */
MPI_Datatype chunks, remainder;
MPI_Aint lb, extent, disps[2];
int blocklens[2];
MPI_Datatype types[2];
int mpi_errno;
/* truly stupendously large counts will overflow an integer with this math,
* but that is a problem for a few decades from now. Sorry, few decades
* from now! */
MPIR_Assert(count/INT_MAX == (int)(count/INT_MAX));
int c = (int)(count/INT_MAX); /* OK to cast until 'count' is 256 bits */
int r = count%INT_MAX;
mpi_errno = MPIR_Type_vector_impl(c, INT_MAX, INT_MAX, oldtype, &chunks);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
mpi_errno = MPIR_Type_contiguous_impl(r, oldtype, &remainder);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPIR_Type_get_extent_impl(oldtype, &lb, &extent);
blocklens[0] = 1; blocklens[1] = 1;
disps[0] = 0; disps[1] = c*extent*INT_MAX;
types[0] = chunks; types[1] = remainder;
mpi_errno = MPIR_Type_create_struct_impl(2, blocklens, disps, types, newtype);
MPIR_Type_free_impl(&chunks);
MPIR_Type_free_impl(&remainder);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
/*@
MPI_Type_create_struct - Create an MPI datatype from a general set of
datatypes, displacements, and block sizes
Input Parameters:
+ count - number of blocks (integer) --- also number of entries in arrays
array_of_types, array_of_displacements and array_of_blocklengths
. array_of_blocklength - number of elements in each block (array of integer)
. array_of_displacements - byte displacement of each block (array of address integer)
- array_of_types - type of elements in each block (array of handles to
datatype objects)
Output Parameter:
. newtype - new datatype (handle)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_ARG
.N MPI_ERR_TYPE
@*/
int MPI_Type_create_struct(int count,
int array_of_blocklengths[],
MPI_Aint array_of_displacements[],
MPI_Datatype array_of_types[],
MPI_Datatype *newtype)
{
int mpi_errno = MPI_SUCCESS;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_TYPE_CREATE_STRUCT);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPIU_THREAD_CS_ENTER(ALLFUNC,);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_TYPE_CREATE_STRUCT);
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
int j;
MPID_Datatype *datatype_ptr = NULL;
MPIR_ERRTEST_COUNT(count,mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
if (count > 0) {
MPIR_ERRTEST_ARGNULL(array_of_blocklengths, "blocklens", mpi_errno);
MPIR_ERRTEST_ARGNULL(array_of_displacements, "indices", mpi_errno);
MPIR_ERRTEST_ARGNULL(array_of_types, "types", mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
for (j=0; j < count; j++) {
MPIR_ERRTEST_ARGNEG(array_of_blocklengths[j], "blocklen", mpi_errno);
MPIR_ERRTEST_DATATYPE(array_of_types[j], "datatype[j]",
mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
if (array_of_types[j] != MPI_DATATYPE_NULL && HANDLE_GET_KIND(array_of_types[j]) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr(array_of_types[j], datatype_ptr);
MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
}
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPIR_Type_create_struct_impl(count, array_of_blocklengths, array_of_displacements,
array_of_types, newtype);
if (mpi_errno) goto fn_exit;
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_TYPE_CREATE_STRUCT);
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_create_struct",
"**mpi_type_create_struct %d %p %p %p %p", count, array_of_blocklengths, array_of_displacements,
array_of_types, newtype);
}
# endif
mpi_errno = MPIR_Err_return_comm(NULL, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIR_Igather_sched_intra_binomial(const void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPIR_Comm *comm_ptr, MPIR_Sched_t s)
{
int mpi_errno = MPI_SUCCESS;
int comm_size, rank;
int relative_rank, is_homogeneous;
int mask, src, dst, relative_src;
MPI_Aint recvtype_size, sendtype_size, curr_cnt=0, nbytes;
int recvblks;
int tmp_buf_size, missing;
void *tmp_buf = NULL;
int blocks[2];
int displs[2];
MPI_Aint struct_displs[2];
MPI_Aint extent=0;
int copy_offset = 0, copy_blks = 0;
MPI_Datatype types[2], tmp_type;
MPIR_SCHED_CHKPMEM_DECL(1);
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
if (((rank == root) && (recvcount == 0)) || ((rank != root) && (sendcount == 0)))
goto fn_exit;
is_homogeneous = TRUE;
#ifdef MPID_HAS_HETERO
is_homogeneous = !comm_ptr->is_hetero;
#endif
MPIR_Assert(comm_ptr->comm_kind == MPIR_COMM_KIND__INTRACOMM);
/* Use binomial tree algorithm. */
relative_rank = (rank >= root) ? rank - root : rank - root + comm_size;
if (rank == root)
{
MPIR_Datatype_get_extent_macro(recvtype, extent);
MPIR_Ensure_Aint_fits_in_pointer(MPIR_VOID_PTR_CAST_TO_MPI_AINT recvbuf+
(extent*recvcount*comm_size));
}
if (is_homogeneous)
{
/* communicator is homogeneous. no need to pack buffer. */
if (rank == root)
{
MPIR_Datatype_get_size_macro(recvtype, recvtype_size);
nbytes = recvtype_size * recvcount;
}
else
{
MPIR_Datatype_get_size_macro(sendtype, sendtype_size);
nbytes = sendtype_size * sendcount;
}
/* Find the number of missing nodes in my sub-tree compared to
* a balanced tree */
for (mask = 1; mask < comm_size; mask <<= 1);
--mask;
while (relative_rank & mask) mask >>= 1;
missing = (relative_rank | mask) - comm_size + 1;
if (missing < 0) missing = 0;
tmp_buf_size = (mask - missing);
/* If the message is smaller than the threshold, we will copy
* our message in there too */
if (nbytes < MPIR_CVAR_GATHER_VSMALL_MSG_SIZE) tmp_buf_size++;
tmp_buf_size *= nbytes;
/* For zero-ranked root, we don't need any temporary buffer */
if ((rank == root) && (!root || (nbytes >= MPIR_CVAR_GATHER_VSMALL_MSG_SIZE)))
tmp_buf_size = 0;
if (tmp_buf_size) {
MPIR_SCHED_CHKPMEM_MALLOC(tmp_buf, void *, tmp_buf_size, mpi_errno, "tmp_buf", MPL_MEM_BUFFER);
}
if (rank == root) {
if (sendbuf != MPI_IN_PLACE) {
mpi_errno = MPIR_Localcopy(sendbuf, sendcount, sendtype,
((char *) recvbuf + extent*recvcount*rank), recvcount, recvtype);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
}
else if (tmp_buf_size && (nbytes < MPIR_CVAR_GATHER_VSMALL_MSG_SIZE)) {
/* copy from sendbuf into tmp_buf */
mpi_errno = MPIR_Localcopy(sendbuf, sendcount, sendtype,
tmp_buf, nbytes, MPI_BYTE);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
curr_cnt = nbytes;
mask = 0x1;
while (mask < comm_size) {
if ((mask & relative_rank) == 0) {
src = relative_rank | mask;
if (src < comm_size) {
src = (src + root) % comm_size;
if (rank == root) {
recvblks = mask;
if ((2 * recvblks) > comm_size)
recvblks = comm_size - recvblks;
if ((rank + mask + recvblks == comm_size) ||
(((rank + mask) % comm_size) < ((rank + mask + recvblks) % comm_size)))
{
/* If the data contiguously fits into the
* receive buffer, place it directly. This
* should cover the case where the root is
* rank 0. */
char *rp = (char *)recvbuf + (((rank + mask) % comm_size)*recvcount*extent);
mpi_errno = MPIR_Sched_recv(rp, (recvblks * recvcount), recvtype, src, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_barrier(s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else if (nbytes < MPIR_CVAR_GATHER_VSMALL_MSG_SIZE) {
mpi_errno = MPIR_Sched_recv(tmp_buf, (recvblks * nbytes), MPI_BYTE, src, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_barrier(s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
copy_offset = rank + mask;
copy_blks = recvblks;
}
else {
blocks[0] = recvcount * (comm_size - root - mask);
displs[0] = recvcount * (root + mask);
blocks[1] = (recvcount * recvblks) - blocks[0];
displs[1] = 0;
mpi_errno = MPIR_Type_indexed_impl(2, blocks, displs, recvtype, &tmp_type);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Type_commit_impl(&tmp_type);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_recv(recvbuf, 1, tmp_type, src, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_barrier(s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* this "premature" free is safe b/c the sched holds an actual ref to keep it alive */
MPIR_Type_free_impl(&tmp_type);
}
}
else { /* Intermediate nodes store in temporary buffer */
MPI_Aint offset;
/* Estimate the amount of data that is going to come in */
recvblks = mask;
relative_src = ((src - root) < 0) ? (src - root + comm_size) : (src - root);
if (relative_src + mask > comm_size)
recvblks -= (relative_src + mask - comm_size);
if (nbytes < MPIR_CVAR_GATHER_VSMALL_MSG_SIZE)
offset = mask * nbytes;
else
offset = (mask - 1) * nbytes;
mpi_errno = MPIR_Sched_recv(((char *)tmp_buf + offset), (recvblks * nbytes),
MPI_BYTE, src, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_barrier(s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
curr_cnt += (recvblks * nbytes);
}
}
}
else {
dst = relative_rank ^ mask;
dst = (dst + root) % comm_size;
if (!tmp_buf_size) {
/* leaf nodes send directly from sendbuf */
mpi_errno = MPIR_Sched_send(sendbuf, sendcount, sendtype, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_barrier(s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else if (nbytes < MPIR_CVAR_GATHER_VSMALL_MSG_SIZE) {
mpi_errno = MPIR_Sched_send(tmp_buf, curr_cnt, MPI_BYTE, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_barrier(s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else {
blocks[0] = sendcount;
struct_displs[0] = MPIR_VOID_PTR_CAST_TO_MPI_AINT sendbuf;
types[0] = sendtype;
/* check for overflow. work around int limits if needed*/
if (curr_cnt - nbytes != (int)(curr_cnt-nbytes)) {
blocks[1] = 1;
MPIR_Type_contiguous_x_impl(curr_cnt - nbytes,
MPI_BYTE, &(types[1]));
} else {
MPIR_Assign_trunc(blocks[1], curr_cnt - nbytes, int);
types[1] = MPI_BYTE;
}
struct_displs[1] = MPIR_VOID_PTR_CAST_TO_MPI_AINT tmp_buf;
mpi_errno = MPIR_Type_create_struct_impl(2, blocks, struct_displs, types, &tmp_type);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Type_commit_impl(&tmp_type);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_send(MPI_BOTTOM, 1, tmp_type, dst, comm_ptr, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPIR_SCHED_BARRIER(s);
/* this "premature" free is safe b/c the sched holds an actual ref to keep it alive */
MPIR_Type_free_impl(&tmp_type);
}
break;
}
mask <<= 1;
}
if ((rank == root) && root && (nbytes < MPIR_CVAR_GATHER_VSMALL_MSG_SIZE) && copy_blks) {
/* reorder and copy from tmp_buf into recvbuf */
/* FIXME why are there two copies here? */
mpi_errno = MPIR_Sched_copy(tmp_buf, nbytes * (comm_size - copy_offset), MPI_BYTE,
((char *)recvbuf + extent * recvcount * copy_offset),
recvcount * (comm_size - copy_offset), recvtype, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIR_Sched_copy((char *)tmp_buf + nbytes * (comm_size - copy_offset),
nbytes * (copy_blks - comm_size + copy_offset), MPI_BYTE,
recvbuf, recvcount * (copy_blks - comm_size + copy_offset),
recvtype, s);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
}
#ifdef MPID_HAS_HETERO
else {
