Skip to content

Commit 6ae2502

Browse files
ggouaillardetggouaillardet
committed
checkpoint
1 parent f4628a5 commit 6ae2502

File tree

1 file changed

+4
-194
lines changed

1 file changed

+4
-194
lines changed

‎ompi/mca/coll/base/coll_base_bcast.c

+4-194
Original file line numberDiff line numberDiff line change
@@ -36,197 +36,6 @@
3636

3737
int
3838
ompi_coll_base_bcast_intra_generic( void* buffer,
39-
int count,
40-
struct ompi_datatype_t* datatype,
41-
int root,
42-
struct ompi_communicator_t* comm,
43-
mca_coll_base_module_t *module,
44-
size_t segment_size,
45-
ompi_coll_tree_t* tree )
46-
{
47-
int err = 0, line, i, rank, segindex, req_index;
48-
int num_segments; /* Number of segments */
49-
int sendcount; /* number of elements sent in this segment */
50-
size_t realsegsize, type_size;
51-
char *tmpbuf;
52-
ptrdiff_t extent, lb;
53-
ompi_request_t *recv_reqs[2] = {MPI_REQUEST_NULL, MPI_REQUEST_NULL};
54-
ompi_request_t **send_reqs = NULL;
55-
56-
#if OPAL_ENABLE_DEBUG
57-
int size;
58-
size = ompi_comm_size(comm);
59-
assert( size > 1 );
60-
#endif
61-
rank = ompi_comm_rank(comm);
62-
63-
ompi_datatype_get_extent (datatype, &lb, &extent);
64-
ompi_datatype_type_size( datatype, &type_size );
65-
num_segments = (original_count + count_by_segment - 1) / count_by_segment;
66-
realsegsize = (ptrdiff_t)count_by_segment * extent;
67-
68-
/* Set the buffer pointers */
69-
tmpbuf = (char *) buffer;
70-
71-
if( tree->tree_nextsize != 0 ) {
72-
send_reqs = ompi_coll_base_comm_get_reqs(module->base_data, tree->tree_nextsize);
73-
if( NULL == send_reqs ) { err = OMPI_ERR_OUT_OF_RESOURCE; line = __LINE__; goto error_hndl; }
74-
}
75-
76-
/* Root code */
77-
if( rank == root ) {
78-
/*
79-
For each segment:
80-
- send segment to all children.
81-
The last segment may have less elements than other segments.
82-
*/
83-
sendcount = count_by_segment;
84-
for( segindex = 0; segindex < num_segments; segindex++ ) {
85-
if( segindex == (num_segments - 1) ) {
86-
sendcount = original_count - segindex * count_by_segment;
87-
}
88-
for( i = 0; i < tree->tree_nextsize; i++ ) {
89-
err = MCA_PML_CALL(isend(tmpbuf, sendcount, datatype,
90-
tree->tree_next[i],
91-
MCA_COLL_BASE_TAG_BCAST,
92-
MCA_PML_BASE_SEND_STANDARD, comm,
93-
&send_reqs[i]));
94-
if (err != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
95-
}
96-
97-
/* complete the sends before starting the next sends */
98-
err = ompi_request_wait_all( tree->tree_nextsize, send_reqs,
99-
MPI_STATUSES_IGNORE );
100-
if (err != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
101-
102-
/* update tmp buffer */
103-
tmpbuf += realsegsize;
104-
105-
}
106-
}
107-
108-
/* Intermediate nodes code */
109-
else if( tree->tree_nextsize > 0 ) {
110-
/*
111-
Create the pipeline.
112-
1) Post the first receive
113-
2) For segments 1 .. num_segments
114-
- post new receive
115-
- wait on the previous receive to complete
116-
- send this data to children
117-
3) Wait on the last segment
118-
4) Compute number of elements in last segment.
119-
5) Send the last segment to children
120-
*/
121-
req_index = 0;
122-
err = MCA_PML_CALL(irecv(tmpbuf, count_by_segment, datatype,
123-
tree->tree_prev, MCA_COLL_BASE_TAG_BCAST,
124-
comm, &recv_reqs[req_index]));
125-
if (err != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
126-
127-
for( segindex = 1; segindex < num_segments; segindex++ ) {
128-
129-
req_index = req_index ^ 0x1;
130-
131-
/* post new irecv */
132-
err = MCA_PML_CALL(irecv( tmpbuf + realsegsize, count_by_segment,
133-
datatype, tree->tree_prev,
134-
MCA_COLL_BASE_TAG_BCAST,
135-
comm, &recv_reqs[req_index]));
136-
if (err != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
137-
138-
/* wait for and forward the previous segment to children */
139-
err = ompi_request_wait( &recv_reqs[req_index ^ 0x1],
140-
MPI_STATUS_IGNORE );
141-
if (err != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
142-
143-
for( i = 0; i < tree->tree_nextsize; i++ ) {
144-
err = MCA_PML_CALL(isend(tmpbuf, count_by_segment, datatype,
145-
tree->tree_next[i],
146-
MCA_COLL_BASE_TAG_BCAST,
147-
MCA_PML_BASE_SEND_STANDARD, comm,
148-
&send_reqs[i]));
149-
if (err != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
150-
}
151-
152-
/* complete the sends before starting the next iteration */
153-
err = ompi_request_wait_all( tree->tree_nextsize, send_reqs,
154-
MPI_STATUSES_IGNORE );
155-
if (err != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
156-
157-
/* Update the receive buffer */
158-
tmpbuf += realsegsize;
159-
160-
}
161-
162-
/* Process the last segment */
163-
err = ompi_request_wait( &recv_reqs[req_index], MPI_STATUS_IGNORE );
164-
if (err != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
165-
sendcount = original_count - (ptrdiff_t)(num_segments - 1) * count_by_segment;
166-
for( i = 0; i < tree->tree_nextsize; i++ ) {
167-
err = MCA_PML_CALL(isend(tmpbuf, sendcount, datatype,
168-
tree->tree_next[i],
169-
MCA_COLL_BASE_TAG_BCAST,
170-
MCA_PML_BASE_SEND_STANDARD, comm,
171-
&send_reqs[i]));
172-
if (err != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
173-
}
174-
175-
err = ompi_request_wait_all( tree->tree_nextsize, send_reqs,
176-
MPI_STATUSES_IGNORE );
177-
if (err != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
178-
}
179-
180-
/* Leaf nodes */
181-
else {
182-
/*
183-
Receive all segments from parent in a loop:
184-
1) post irecv for the first segment
185-
2) for segments 1 .. num_segments
186-
- post irecv for the next segment
187-
- wait on the previous segment to arrive
188-
3) wait for the last segment
189-
*/
190-
req_index = 0;
191-
err = MCA_PML_CALL(irecv(tmpbuf, count_by_segment, datatype,
192-
tree->tree_prev, MCA_COLL_BASE_TAG_BCAST,
193-
comm, &recv_reqs[req_index]));
194-
if (err != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
195-
196-
for( segindex = 1; segindex < num_segments; segindex++ ) {
197-
req_index = req_index ^ 0x1;
198-
tmpbuf += realsegsize;
199-
/* post receive for the next segment */
200-
err = MCA_PML_CALL(irecv(tmpbuf, count_by_segment, datatype,
201-
tree->tree_prev, MCA_COLL_BASE_TAG_BCAST,
202-
comm, &recv_reqs[req_index]));
203-
if (err != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
204-
/* wait on the previous segment */
205-
err = ompi_request_wait( &recv_reqs[req_index ^ 0x1],
206-
MPI_STATUS_IGNORE );
207-
if (err != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
208-
}
209-
210-
err = ompi_request_wait( &recv_reqs[req_index], MPI_STATUS_IGNORE );
211-
if (err != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
212-
}
213-
214-
return (MPI_SUCCESS);
215-
216-
error_hndl:
217-
OPAL_OUTPUT( (ompi_coll_base_framework.framework_output,"%s:%4d\tError occurred %d, rank %2d",
218-
__FILE__, line, err, rank) );
219-
(void)line; // silence compiler warnings
220-
ompi_coll_base_free_reqs( recv_reqs, 2);
221-
if( NULL != send_reqs ) {
222-
ompi_coll_base_free_reqs(send_reqs, tree->tree_nextsize);
223-
}
224-
225-
return err;
226-
}
227-
228-
int
229-
ompi_coll_base_bcast_intra_generic2( void* buffer,
23039
int count,
23140
struct ompi_datatype_t* datatype,
23241
int root,
@@ -252,7 +61,7 @@ ompi_coll_base_bcast_intra_generic2( void* buffer,
25261
rank = ompi_comm_rank(comm);
25362

25463
if( tree->tree_nextsize != 0 ) {
255-
send_reqs = coll_base_comm_get_reqs(module->base_data, tree->tree_nextsize);
64+
send_reqs = ompi_coll_base_comm_get_reqs(module->base_data, tree->tree_nextsize);
25665
if( NULL == send_reqs ) { err = OMPI_ERR_OUT_OF_RESOURCE; line = __LINE__; goto error_hndl; }
25766
}
25867

@@ -793,10 +602,11 @@ ompi_coll_base_bcast_intra_split_bintree ( void* buffer,
793602
offsets[0] += sizes[0];
794603
}
795604
opal_convertor_set_position(&recv_convertors[0], &offsets[0]);
605+
796606
offsets[0] += sizes[lr] - remainings[lr];
797607
offsets[1] = offsets[0] + segsize;
798-
if (offsets[1] > sizes[lr]) {
799-
offsets[1] = sizes[lr];
608+
if (offsets[1] > sizes[lr] + (lr?sizes[0]:0)) {
609+
offsets[1] = sizes[lr] + (lr?sizes[0]:0);
800610
}
801611
opal_convertor_set_position(&recv_convertors[1], &offsets[1]);
802612
if (offsets[1] == offsets[0]) {

0 commit comments

Comments
 (0)