ompi_datatype_t* test_create_blacs_type( void )
{
ompi_datatype_t *pdt;
ompi_datatype_create_indexed( 18, blacs_length, blacs_indices, &ompi_mpi_int.dt, &pdt );
ompi_datatype_commit( &pdt );
if( outputFlags & DUMP_DATA_AFTER_COMMIT ) {
ompi_datatype_dump( pdt );
}
return pdt;
}
ompi_datatype_t* test_matrix_borders( unsigned int size, unsigned int width )
{
ompi_datatype_t *pdt, *pdt_line;
int disp[2];
int blocklen[2];
disp[0] = 0;
blocklen[0] = width;
disp[1] = (size - width) * sizeof(double);
blocklen[1] = width;
ompi_datatype_create_indexed( 2, blocklen, disp, &ompi_mpi_double.dt,
&pdt_line );
ompi_datatype_create_contiguous( size, pdt_line, &pdt );
OBJ_RELEASE( pdt_line ); /*assert( pdt_line == NULL );*/
return pdt;
}
ompi_datatype_t* lower_matrix( unsigned int mat_size )
{
int *disp, *blocklen;
unsigned int i;
ompi_datatype_t* upper;
disp = (int*)malloc( sizeof(int) * mat_size );
blocklen = (int*)malloc( sizeof(int) * mat_size );
for( i = 0; i < mat_size; i++ ) {
disp[i] = i * mat_size;
blocklen[i] = i;
}
ompi_datatype_create_indexed( mat_size, blocklen, disp, &ompi_mpi_double.dt,
&upper );
free( disp );
free( blocklen );
return upper;
}
ompi_datatype_t* upper_matrix( unsigned int mat_size )
{
int *disp, *blocklen;
unsigned int i;
ompi_datatype_t* upper;
disp = (int*)malloc( sizeof(int) * mat_size );
blocklen = (int*)malloc( sizeof(int) * mat_size );
for( i = 0; i < mat_size; i++ ) {
disp[i] = i * mat_size + i;
blocklen[i] = mat_size - i;
}
ompi_datatype_create_indexed( mat_size, blocklen, disp, &ompi_mpi_double.dt,
&upper );
ompi_datatype_commit( &upper );
if( outputFlags & DUMP_DATA_AFTER_COMMIT ) {
ompi_datatype_dump( upper );
}
free( disp );
free( blocklen );
return upper;
}
static ompi_datatype_t* __ompi_datatype_create_from_args( int32_t* i, MPI_Aint* a,
ompi_datatype_t** d, int32_t type )
{
ompi_datatype_t* datatype = NULL;
switch(type){
/******************************************************************/
case MPI_COMBINER_DUP:
/* should we duplicate d[0]? */
/* ompi_datatype_set_args( datatype, 0, NULL, 0, NULL, 1, d[0], MPI_COMBINER_DUP ); */
assert(0); /* shouldn't happen */
break;
/******************************************************************/
case MPI_COMBINER_CONTIGUOUS:
ompi_datatype_create_contiguous( i[0], d[0], &datatype );
ompi_datatype_set_args( datatype, 1, (const int **) &i, 0, NULL, 1, d, MPI_COMBINER_CONTIGUOUS );
break;
/******************************************************************/
case MPI_COMBINER_VECTOR:
ompi_datatype_create_vector( i[0], i[1], i[2], d[0], &datatype );
{
const int* a_i[3] = {&i[0], &i[1], &i[2]};
ompi_datatype_set_args( datatype, 3, a_i, 0, NULL, 1, d, MPI_COMBINER_VECTOR );
}
break;
/******************************************************************/
case MPI_COMBINER_HVECTOR_INTEGER:
case MPI_COMBINER_HVECTOR:
ompi_datatype_create_hvector( i[0], i[1], a[0], d[0], &datatype );
{
const int* a_i[2] = {&i[0], &i[1]};
ompi_datatype_set_args( datatype, 2, a_i, 1, a, 1, d, MPI_COMBINER_HVECTOR );
}
break;
/******************************************************************/
case MPI_COMBINER_INDEXED: /* TO CHECK */
ompi_datatype_create_indexed( i[0], &(i[1]), &(i[1+i[0]]), d[0], &datatype );
{
const int* a_i[3] = {&i[0], &i[1], &(i[1+i[0]])};
ompi_datatype_set_args( datatype, 2 * i[0] + 1, a_i, 0, NULL, 1, d, MPI_COMBINER_INDEXED );
}
break;
/******************************************************************/
case MPI_COMBINER_HINDEXED_INTEGER:
case MPI_COMBINER_HINDEXED:
ompi_datatype_create_hindexed( i[0], &(i[1]), a, d[0], &datatype );
{
const int* a_i[2] = {&i[0], &i[1]};
ompi_datatype_set_args( datatype, i[0] + 1, a_i, i[0], a, 1, d, MPI_COMBINER_HINDEXED );
}
break;
/******************************************************************/
case MPI_COMBINER_INDEXED_BLOCK:
ompi_datatype_create_indexed_block( i[0], i[1], &(i[2]), d[0], &datatype );
{
const int* a_i[3] = {&i[0], &i[1], &i[2]};
ompi_datatype_set_args( datatype, i[0] + 2, a_i, 0, NULL, 1, d, MPI_COMBINER_INDEXED_BLOCK );
}
break;
/******************************************************************/
case MPI_COMBINER_STRUCT_INTEGER:
case MPI_COMBINER_STRUCT:
ompi_datatype_create_struct( i[0], &(i[1]), a, d, &datatype );
{
const int* a_i[2] = {&i[0], &i[1]};
ompi_datatype_set_args( datatype, i[0] + 1, a_i, i[0], a, i[0], d, MPI_COMBINER_STRUCT );
}
break;
/******************************************************************/
case MPI_COMBINER_SUBARRAY:
ompi_datatype_create_subarray( i[0], &i[1 + 0 * i[0]], &i[1 + 1 * i[0]],
&i[1 + 2 * i[0]], i[1 + 3 * i[0]],
d[0], &datatype );
{
const int* a_i[5] = {&i[0], &i[1 + 0 * i[0]], &i[1 + 1 * i[0]], &i[1 + 2 * i[0]], &i[1 + 3 * i[0]]};
ompi_datatype_set_args( datatype, 3 * i[0] + 2, a_i, 0, NULL, 1, d, MPI_COMBINER_SUBARRAY);
}
break;
/******************************************************************/
case MPI_COMBINER_DARRAY:
ompi_datatype_create_darray( i[0] /* size */, i[1] /* rank */, i[2] /* ndims */,
&i[3 + 0 * i[0]], &i[3 + 1 * i[0]],
&i[3 + 2 * i[0]], &i[3 + 3 * i[0]],
i[3 + 4 * i[0]], d[0], &datatype );
{
const int* a_i[8] = {&i[0], &i[1], &i[2], &i[3 + 0 * i[0]], &i[3 + 1 * i[0]], &i[3 + 2 * i[0]],
&i[3 + 3 * i[0]], &i[3 + 4 * i[0]]};
ompi_datatype_set_args( datatype, 4 * i[0] + 4,a_i, 0, NULL, 1, d, MPI_COMBINER_DARRAY);
}
break;
/******************************************************************/
case MPI_COMBINER_F90_REAL:
case MPI_COMBINER_F90_COMPLEX:
/*pArgs->i[0] = i[0][0];
pArgs->i[1] = i[1][0];
*/
break;
/******************************************************************/
case MPI_COMBINER_F90_INTEGER:
/*pArgs->i[0] = i[0][0];*/
break;
/******************************************************************/
case MPI_COMBINER_RESIZED:
ompi_datatype_create_resized(d[0], a[0], a[1], &datatype);
ompi_datatype_set_args( datatype, 0, NULL, 2, a, 1, d, MPI_COMBINER_RESIZED );
break;
/******************************************************************/
case MPI_COMBINER_HINDEXED_BLOCK:
ompi_datatype_create_hindexed_block( i[0], i[1], a, d[0], &datatype );
{
const int* a_i[2] = {&i[0], &i[1]};
ompi_datatype_set_args( datatype, 2 + i[0], a_i, i[0], a, 1, d, MPI_COMBINER_HINDEXED_BLOCK );
}
break;
/******************************************************************/
default:
break;
}
return datatype;
}
/*
* scatterv_inter
*
* Function: - scatterv operation
* Accepts: - same arguments as MPI_Scatterv()
* Returns: - MPI_SUCCESS or error code
*/
int
mca_coll_inter_scatterv_inter(void *sbuf, int *scounts,
int *disps, struct ompi_datatype_t *sdtype,
void *rbuf, int rcount,
struct ompi_datatype_t *rdtype, int root,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i, rank, size, err, total, size_local;
int *counts=NULL,*displace=NULL;
char *ptmp=NULL;
MPI_Aint incr;
MPI_Aint extent;
MPI_Aint lb;
ompi_datatype_t *ndtype;
/* Initialize */
rank = ompi_comm_rank(comm);
size = ompi_comm_remote_size(comm);
size_local = ompi_comm_size(comm);
if (MPI_PROC_NULL == root) {
/* do nothing */
err = OMPI_SUCCESS;
} else if (MPI_ROOT != root) {
if(0 == rank) {
/* local root recieves the counts from the root */
counts = (int *)malloc(sizeof(int) * size_local);
err = MCA_PML_CALL(recv(counts, size_local, MPI_INT,
root, MCA_COLL_BASE_TAG_SCATTERV,
comm, MPI_STATUS_IGNORE));
if (OMPI_SUCCESS != err) {
return err;
}
/* calculate the whole buffer size and recieve it from root */
err = ompi_datatype_get_extent(rdtype, &lb, &extent);
if (OMPI_SUCCESS != err) {
return OMPI_ERROR;
}
incr = 0;
for (i = 0; i < size_local; i++) {
incr = incr + extent*counts[i];
}
if ( incr > 0 ) {
ptmp = (char*)malloc(incr);
if (NULL == ptmp) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
}
total = 0;
for (i = 0; i < size_local; i++) {
total = total + counts[i];
}
err = MCA_PML_CALL(recv(ptmp, total, rdtype,
root, MCA_COLL_BASE_TAG_SCATTERV,
comm, MPI_STATUS_IGNORE));
if (OMPI_SUCCESS != err) {
return err;
}
/* set the local displacement i.e. no displacements here */
displace = (int *)malloc(sizeof(int) * size_local);
displace[0] = 0;
for (i = 1; i < size_local; i++) {
displace[i] = displace[i-1] + counts[i-1];
}
}
/* perform the scatterv locally */
err = comm->c_local_comm->c_coll.coll_scatterv(ptmp, counts, displace,
rdtype, rbuf, rcount,
rdtype, 0, comm->c_local_comm,
comm->c_local_comm->c_coll.coll_scatterv_module);
if (OMPI_SUCCESS != err) {
return err;
}
if (NULL != ptmp) {
free(ptmp);
}
if (NULL != displace) {
free(displace);
}
if (NULL != counts) {
free(counts);
}
} else {
err = MCA_PML_CALL(send(scounts, size, MPI_INT, 0,
MCA_COLL_BASE_TAG_SCATTERV,
MCA_PML_BASE_SEND_STANDARD, comm));
if (OMPI_SUCCESS != err) {
return err;
}
ompi_datatype_create_indexed(size,scounts,disps,sdtype,&ndtype);
ompi_datatype_commit(&ndtype);
err = MCA_PML_CALL(send(sbuf, 1, ndtype, 0,
MCA_COLL_BASE_TAG_SCATTERV,
MCA_PML_BASE_SEND_STANDARD, comm));
if (OMPI_SUCCESS != err) {
return err;
}
ompi_datatype_destroy(&ndtype);
}
/* All done */
return err;
}
int ompi_coll_tuned_alltoall_intra_bruck(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void* rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i, k, line = -1, rank, size, err = 0, weallocated = 0;
int sendto, recvfrom, distance, *displs = NULL, *blen = NULL;
char *tmpbuf = NULL, *tmpbuf_free = NULL;
ptrdiff_t rlb, slb, tlb, sext, rext, tsext;
struct ompi_datatype_t *new_ddt;
#ifdef blahblah
mca_coll_tuned_module_t *tuned_module = (mca_coll_tuned_module_t*) module;
mca_coll_tuned_comm_t *data = tuned_module->tuned_data;
#endif
if (MPI_IN_PLACE == sbuf) {
return mca_coll_tuned_alltoall_intra_basic_inplace (rbuf, rcount, rdtype,
comm, module);
}
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_tuned_stream,
"coll:tuned:alltoall_intra_bruck rank %d", rank));
err = ompi_datatype_get_extent (sdtype, &slb, &sext);
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
err = ompi_datatype_get_true_extent(sdtype, &tlb, &tsext);
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
err = ompi_datatype_get_extent (rdtype, &rlb, &rext);
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
#ifdef blahblah
/* try and SAVE memory by using the data segment hung off
the communicator if possible */
if (data->mcct_num_reqs >= size) {
/* we have enought preallocated for displments and lengths */
displs = (int*) data->mcct_reqs;
blen = (int *) (displs + size);
weallocated = 0;
}
else { /* allocate the buffers ourself */
#endif
displs = (int *) malloc(size * sizeof(int));
if (displs == NULL) { line = __LINE__; err = -1; goto err_hndl; }
blen = (int *) malloc(size * sizeof(int));
if (blen == NULL) { line = __LINE__; err = -1; goto err_hndl; }
weallocated = 1;
#ifdef blahblah
}
#endif
/* tmp buffer allocation for message data */
tmpbuf_free = (char *) malloc(tsext + ((ptrdiff_t)scount * (ptrdiff_t)size - 1) * sext);
if (tmpbuf_free == NULL) { line = __LINE__; err = -1; goto err_hndl; }
tmpbuf = tmpbuf_free - slb;
/* Step 1 - local rotation - shift up by rank */
err = ompi_datatype_copy_content_same_ddt (sdtype,
(int32_t) ((ptrdiff_t)(size - rank) * (ptrdiff_t)scount),
tmpbuf,
((char*) sbuf) + (ptrdiff_t)rank * (ptrdiff_t)scount * sext);
if (err<0) {
line = __LINE__; err = -1; goto err_hndl;
}
if (rank != 0) {
err = ompi_datatype_copy_content_same_ddt (sdtype, (ptrdiff_t)rank * (ptrdiff_t)scount,
tmpbuf + (ptrdiff_t)(size - rank) * (ptrdiff_t)scount* sext,
(char*) sbuf);
if (err<0) {
line = __LINE__; err = -1; goto err_hndl;
}
}
/* perform communication step */
for (distance = 1; distance < size; distance<<=1) {
sendto = (rank + distance) % size;
recvfrom = (rank - distance + size) % size;
k = 0;
/* create indexed datatype */
for (i = 1; i < size; i++) {
if (( i & distance) == distance) {
displs[k] = (ptrdiff_t)i * (ptrdiff_t)scount;
blen[k] = scount;
k++;
}
}
/* Set indexes and displacements */
err = ompi_datatype_create_indexed(k, blen, displs, sdtype, &new_ddt);
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
/* Commit the new datatype */
err = ompi_datatype_commit(&new_ddt);
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
/* Sendreceive */
err = ompi_coll_tuned_sendrecv ( tmpbuf, 1, new_ddt, sendto,
MCA_COLL_BASE_TAG_ALLTOALL,
rbuf, 1, new_ddt, recvfrom,
MCA_COLL_BASE_TAG_ALLTOALL,
comm, MPI_STATUS_IGNORE, rank );
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
/* Copy back new data from recvbuf to tmpbuf */
err = ompi_datatype_copy_content_same_ddt(new_ddt, 1,tmpbuf, (char *) rbuf);
if (err < 0) { line = __LINE__; err = -1; goto err_hndl; }
/* free ddt */
err = ompi_datatype_destroy(&new_ddt);
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
} /* end of for (distance = 1... */
/* Step 3 - local rotation - */
for (i = 0; i < size; i++) {
err = ompi_datatype_copy_content_same_ddt (rdtype, (int32_t) rcount,
((char*)rbuf) + ((ptrdiff_t)((rank - i + size) % size) * (ptrdiff_t)rcount * rext),
tmpbuf + (ptrdiff_t)i * (ptrdiff_t)rcount * rext);
if (err < 0) { line = __LINE__; err = -1; goto err_hndl; }
}
/* Step 4 - clean up */
if (tmpbuf != NULL) free(tmpbuf_free);
if (weallocated) {
if (displs != NULL) free(displs);
if (blen != NULL) free(blen);
}
return OMPI_SUCCESS;
err_hndl:
OPAL_OUTPUT((ompi_coll_tuned_stream,
"%s:%4d\tError occurred %d, rank %2d", __FILE__, line, err,
rank));
if (tmpbuf != NULL) free(tmpbuf_free);
if (weallocated) {
if (displs != NULL) free(displs);
if (blen != NULL) free(blen);
}
return err;
}
/*
* gatherv_inter
*
* Function: - gatherv operation using a local gather on c_local_comm
* Accepts: - same arguments as MPI_Gatherv()
* Returns: - MPI_SUCCESS or error code
*/
int
mca_coll_inter_gatherv_inter(const void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void *rbuf, const int *rcounts, const int *disps,
struct ompi_datatype_t *rdtype, int root,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i, rank, size, size_local, total=0, err;
int *count=NULL, *displace=NULL;
char *ptmp=NULL;
MPI_Aint incr;
MPI_Aint extent;
MPI_Aint lb;
ompi_datatype_t *ndtype;
if (MPI_PROC_NULL == root) { /* do nothing */
return OMPI_SUCCESS;
}
size = ompi_comm_remote_size(comm);
rank = ompi_comm_rank(comm);
size_local = ompi_comm_size(comm);
if (MPI_ROOT == root) { /* I am the root, receiving the data from zero. */
ompi_datatype_create_indexed(size, rcounts, disps, rdtype, &ndtype);
ompi_datatype_commit(&ndtype);
err = MCA_PML_CALL(recv(rbuf, 1, ndtype, 0,
MCA_COLL_BASE_TAG_GATHERV,
comm, MPI_STATUS_IGNORE));
ompi_datatype_destroy(&ndtype);
return err;
}
if (0 == rank) {
count = (int *)malloc(sizeof(int) * size_local);
displace = (int *)malloc(sizeof(int) * size_local);
if ((NULL == displace) || (NULL == count)) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
err = comm->c_local_comm->c_coll.coll_gather(&scount, 1, MPI_INT,
count, 1, MPI_INT,
0, comm->c_local_comm,
comm->c_local_comm->c_coll.coll_gather_module);
if (OMPI_SUCCESS != err) {
goto exit;
}
if(0 == rank) {
displace[0] = 0;
for (i = 1; i < size_local; i++) {
displace[i] = displace[i-1] + count[i-1];
}
/* Perform the gatherv locally with the first process as root */
err = ompi_datatype_get_extent(sdtype, &lb, &extent);
if (OMPI_SUCCESS != err) {
err = OMPI_ERROR;
goto exit;
}
incr = 0;
for (i = 0; i < size_local; i++) {
incr = incr + extent*count[i];
}
if ( incr > 0 ) {
ptmp = (char*)malloc(incr);
if (NULL == ptmp) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
}
err = comm->c_local_comm->c_coll.coll_gatherv(sbuf, scount, sdtype,
ptmp, count, displace,
sdtype,0, comm->c_local_comm,
comm->c_local_comm->c_coll.coll_gatherv_module);
if (OMPI_SUCCESS != err) {
goto exit;
}
if (0 == rank) {
for (i = 0; i < size_local; i++) {
total = total + count[i];
}
/* First process sends data to the root */
err = MCA_PML_CALL(send(ptmp, total, sdtype, root,
MCA_COLL_BASE_TAG_GATHERV,
MCA_PML_BASE_SEND_STANDARD, comm));
}
exit:
if (NULL != ptmp) {
free(ptmp);
}
if (NULL != displace) {
free(displace);
}
if (NULL != count) {
free(count);
}
/* All done */
return err;
}
int ompi_coll_base_alltoall_intra_bruck(const void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void* rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i, k, line = -1, rank, size, err = 0;
int sendto, recvfrom, distance, *displs = NULL, *blen = NULL;
char *tmpbuf = NULL, *tmpbuf_free = NULL;
OPAL_PTRDIFF_TYPE sext, rext, span, gap;
struct ompi_datatype_t *new_ddt;
if (MPI_IN_PLACE == sbuf) {
return mca_coll_base_alltoall_intra_basic_inplace (rbuf, rcount, rdtype,
comm, module);
}
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"coll:base:alltoall_intra_bruck rank %d", rank));
err = ompi_datatype_type_extent (sdtype, &sext);
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
err = ompi_datatype_type_extent (rdtype, &rext);
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
span = opal_datatype_span(&sdtype->super, (int64_t)size * scount, &gap);
displs = (int *) malloc(size * sizeof(int));
if (displs == NULL) { line = __LINE__; err = -1; goto err_hndl; }
blen = (int *) malloc(size * sizeof(int));
if (blen == NULL) { line = __LINE__; err = -1; goto err_hndl; }
/* tmp buffer allocation for message data */
tmpbuf_free = (char *)malloc(span);
if (tmpbuf_free == NULL) { line = __LINE__; err = -1; goto err_hndl; }
tmpbuf = tmpbuf_free - gap;
/* Step 1 - local rotation - shift up by rank */
err = ompi_datatype_copy_content_same_ddt (sdtype,
(int32_t) ((ptrdiff_t)(size - rank) * (ptrdiff_t)scount),
tmpbuf,
((char*) sbuf) + (ptrdiff_t)rank * (ptrdiff_t)scount * sext);
if (err<0) {
line = __LINE__; err = -1; goto err_hndl;
}
if (rank != 0) {
err = ompi_datatype_copy_content_same_ddt (sdtype, (ptrdiff_t)rank * (ptrdiff_t)scount,
tmpbuf + (ptrdiff_t)(size - rank) * (ptrdiff_t)scount* sext,
(char*) sbuf);
if (err<0) {
line = __LINE__; err = -1; goto err_hndl;
}
}
/* perform communication step */
for (distance = 1; distance < size; distance<<=1) {
sendto = (rank + distance) % size;
recvfrom = (rank - distance + size) % size;
k = 0;
/* create indexed datatype */
for (i = 1; i < size; i++) {
if (( i & distance) == distance) {
displs[k] = (ptrdiff_t)i * (ptrdiff_t)scount;
blen[k] = scount;
k++;
}
}
/* Set indexes and displacements */
err = ompi_datatype_create_indexed(k, blen, displs, sdtype, &new_ddt);
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
/* Commit the new datatype */
err = ompi_datatype_commit(&new_ddt);
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
/* Sendreceive */
err = ompi_coll_base_sendrecv ( tmpbuf, 1, new_ddt, sendto,
MCA_COLL_BASE_TAG_ALLTOALL,
rbuf, 1, new_ddt, recvfrom,
MCA_COLL_BASE_TAG_ALLTOALL,
comm, MPI_STATUS_IGNORE, rank );
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
/* Copy back new data from recvbuf to tmpbuf */
err = ompi_datatype_copy_content_same_ddt(new_ddt, 1,tmpbuf, (char *) rbuf);
if (err < 0) { line = __LINE__; err = -1; goto err_hndl; }
/* free ddt */
err = ompi_datatype_destroy(&new_ddt);
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl; }
} /* end of for (distance = 1... */
/* Step 3 - local rotation - */
for (i = 0; i < size; i++) {
err = ompi_datatype_copy_content_same_ddt (rdtype, (int32_t) rcount,
((char*)rbuf) + ((ptrdiff_t)((rank - i + size) % size) * (ptrdiff_t)rcount * rext),
tmpbuf + (ptrdiff_t)i * (ptrdiff_t)rcount * rext);
if (err < 0) { line = __LINE__; err = -1; goto err_hndl; }
}
/* Step 4 - clean up */
if (tmpbuf != NULL) free(tmpbuf_free);
if (displs != NULL) free(displs);
if (blen != NULL) free(blen);
return OMPI_SUCCESS;
err_hndl:
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"%s:%4d\tError occurred %d, rank %2d", __FILE__, line, err,
rank));
(void)line; // silence compiler warning
if (tmpbuf != NULL) free(tmpbuf_free);
if (displs != NULL) free(displs);
if (blen != NULL) free(blen);
return err;
}
/*
* allgatherv_inter
*
* Function: - allgatherv using other MPI collectives
* Accepts: - same as MPI_Allgatherv()
* Returns: - MPI_SUCCESS or error code
*/
int
mca_coll_inter_allgatherv_inter(const void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void *rbuf, const int *rcounts, const int *disps,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i, rank, size, size_local, total=0, err;
int *count=NULL,*displace=NULL;
char *ptmp_free=NULL, *ptmp=NULL;
ompi_datatype_t *ndtype = NULL;
rank = ompi_comm_rank(comm);
size_local = ompi_comm_size(comm->c_local_comm);
size = ompi_comm_remote_size(comm);
if (0 == rank) {
count = (int *)malloc(sizeof(int) * size_local);
displace = (int *)malloc(sizeof(int) * size_local);
if ((NULL == count) || (NULL == displace)) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
/* Local gather to get the scount of each process */
err = comm->c_local_comm->c_coll->coll_gather(&scount, 1, MPI_INT,
count, 1, MPI_INT,
0, comm->c_local_comm,
comm->c_local_comm->c_coll->coll_gather_module);
if (OMPI_SUCCESS != err) {
goto exit;
}
if(0 == rank) {
displace[0] = 0;
for (i = 1; i < size_local; i++) {
displace[i] = displace[i-1] + count[i-1];
}
total = 0;
for (i = 0; i < size_local; i++) {
total = total + count[i];
}
if ( total > 0 ) {
ptrdiff_t gap, span;
span = opal_datatype_span(&sdtype->super, total, &gap);
ptmp_free = (char*)malloc(span);
if (NULL == ptmp_free) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
ptmp = ptmp_free - gap;
}
}
err = comm->c_local_comm->c_coll->coll_gatherv(sbuf, scount, sdtype,
ptmp, count, displace,
sdtype,0, comm->c_local_comm,
comm->c_local_comm->c_coll->coll_gatherv_module);
if (OMPI_SUCCESS != err) {
goto exit;
}
ompi_datatype_create_indexed(size,rcounts,disps,rdtype,&ndtype);
ompi_datatype_commit(&ndtype);
if (0 == rank) {
/* Exchange data between roots */
err = ompi_coll_base_sendrecv_actual(ptmp, total, sdtype, 0,
MCA_COLL_BASE_TAG_ALLGATHERV,
rbuf, 1, ndtype, 0,
MCA_COLL_BASE_TAG_ALLGATHERV,
comm, MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != err) {
goto exit;
}
}
/* bcast the message to all the local processes */
err = comm->c_local_comm->c_coll->coll_bcast(rbuf, 1, ndtype,
0, comm->c_local_comm,
comm->c_local_comm->c_coll->coll_bcast_module);
exit:
if( NULL != ndtype ) {
ompi_datatype_destroy(&ndtype);
}
if (NULL != ptmp_free) {
free(ptmp_free);
}
if (NULL != displace) {
free(displace);
}
if (NULL != count) {
free(count);
}
return err;
}
/*
* allgatherv_intra
*
* Function: - allgatherv using other MPI collectives
* Accepts: - same as MPI_Allgatherv()
* Returns: - MPI_SUCCESS or error code
*/
int
mca_coll_basic_allgatherv_intra(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void *rbuf, int *rcounts, int *disps,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i, size, rank ;
int err;
MPI_Aint extent;
MPI_Aint lb;
char *send_buf = NULL;
struct ompi_datatype_t *newtype, *send_type;
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
/*
* We don't have a root process defined. Arbitrarily assign root
* to process with rank 0 (OMPI convention)
*/
if (MPI_IN_PLACE == sbuf) {
ompi_datatype_get_extent(rdtype, &lb, &extent);
send_type = rdtype;
send_buf = (char*)rbuf;
for (i = 0; i < rank; ++i) {
send_buf += (rcounts[i] * extent);
}
} else {
send_buf = (char*)sbuf;
send_type = sdtype;
}
err = comm->c_coll.coll_gatherv(send_buf,
rcounts[rank], send_type,rbuf,
rcounts, disps, rdtype, 0,
comm, comm->c_coll.coll_gatherv_module);
if (MPI_SUCCESS != err) {
return err;
}
/*
* we now have all the data in the root's rbuf. Need to
* broadcast the data out to the other processes
*
* Need to define a datatype that captures the different vectors
* from each process. MPI_TYPE_INDEXED with params
* size,rcount,displs,rdtype,newtype
* should do the trick.
* Use underlying ddt functions to create, and commit the
* new datatype on each process, then broadcast and destroy the
* datatype.
*/
err = ompi_datatype_create_indexed(size,rcounts,disps,rdtype,&newtype);
if (MPI_SUCCESS != err) {
return err;
}
err = ompi_datatype_commit(&newtype);
if(MPI_SUCCESS != err) {
return err;
}
err = comm->c_coll.coll_bcast( rbuf, 1 ,newtype,0,comm,
comm->c_coll.coll_bcast_module);
ompi_datatype_destroy (&newtype);
return err;
}
