-/* Copyright (c) 2013-2014. The SimGrid Team.
+/* Copyright (c) 2013-2017. The SimGrid Team.
* All rights reserved. */
/* This program is free software; you can redistribute it and/or modify it
* under the terms of the license (GNU LGPL) which comes with this package. */
- #include "../colls_private.h"
- #include "../coll_tuned_topo.h"
+#include "../coll_tuned_topo.hpp"
+#include "../colls_private.hpp"
#define MAXTREEFANOUT 32
-
+namespace simgrid{
+namespace smpi{
int Coll_bcast_ompi_pipeline::bcast( void* buffer,
- int original_count,
- MPI_Datatype datatype,
+ int original_count,
+ MPI_Datatype datatype,
int root,
MPI_Comm comm)
{
size_t segsize =1024 << 7;
//mca_coll_tuned_module_t *tuned_module = (mca_coll_tuned_module_t*) module;
//mca_coll_tuned_comm_t *data = tuned_module->tuned_data;
-
+
// return ompi_coll_tuned_bcast_intra_generic( buffer, count, datatype, root, comm, module,
// count_by_segment, data->cached_pipeline );
ompi_coll_tree_t * tree = ompi_coll_tuned_topo_build_chain( 1, comm, root );
int rank, size;
int segindex;
int num_segments; /* Number of segments */
- int sendcount; /* number of elements sent in this segment */
+ int sendcount; /* number of elements sent in this segment */
size_t realsegsize;
char *tmpbuf;
ptrdiff_t extent;
MPI_Request recv_reqs[2] = {MPI_REQUEST_NULL, MPI_REQUEST_NULL};
MPI_Request *send_reqs = NULL;
int req_index;
-
+
/**
* Determine number of elements sent per operation.
*/
const double a_p16 = 3.2118e-6; /* [1 / byte] */
- const double b_p16 = 8.7936;
+ const double b_p16 = 8.7936;
const double a_p64 = 2.3679e-6; /* [1 / byte] */
- const double b_p64 = 1.1787;
+ const double b_p64 = 1.1787;
const double a_p128 = 1.6134e-6; /* [1 / byte] */
const double b_p128 = 2.1102;
size_t message_size;
message_size = type_size * (unsigned long)original_count; /* needed for decision */
if (size < (a_p128 * message_size + b_p128)) {
- //Pipeline with 128KB segments
+ //Pipeline with 128KB segments
segsize = 1024 << 7;
}else if (size < (a_p64 * message_size + b_p64)) {
- // Pipeline with 64KB segments
+ // Pipeline with 64KB segments
segsize = 1024 << 6;
}else if (size < (a_p16 * message_size + b_p16)) {
- //Pipeline with 16KB segments
+ //Pipeline with 16KB segments
segsize = 1024 << 4;
}
extent = datatype->get_extent();
num_segments = (original_count + count_by_segment - 1) / count_by_segment;
realsegsize = count_by_segment * extent;
-
+
/* Set the buffer pointers */
tmpbuf = (char *) buffer;
/* Root code */
if( rank == root ) {
- /*
+ /*
For each segment:
- send segment to all children.
The last segment may have less elements than other segments.
if( segindex == (num_segments - 1) ) {
sendcount = original_count - segindex * count_by_segment;
}
- for( i = 0; i < tree->tree_nextsize; i++ ) {
+ for( i = 0; i < tree->tree_nextsize; i++ ) {
send_reqs[i] = Request::isend(tmpbuf, sendcount, datatype,
- tree->tree_next[i],
+ tree->tree_next[i],
COLL_TAG_BCAST, comm);
- }
+ }
/* complete the sends before starting the next sends */
- Request::waitall( tree->tree_nextsize, send_reqs,
+ Request::waitall( tree->tree_nextsize, send_reqs,
MPI_STATUSES_IGNORE );
/* update tmp buffer */
tmpbuf += realsegsize;
}
- }
-
+ }
+
/* Intermediate nodes code */
- else if( tree->tree_nextsize > 0 ) {
- /*
- Create the pipeline.
+ else if( tree->tree_nextsize > 0 ) {
+ /*
+ Create the pipeline.
1) Post the first receive
2) For segments 1 .. num_segments
- post new receive
recv_reqs[req_index]=Request::irecv(tmpbuf, count_by_segment, datatype,
tree->tree_prev, COLL_TAG_BCAST,
comm);
-
+
for( segindex = 1; segindex < num_segments; segindex++ ) {
-
+
req_index = req_index ^ 0x1;
-
+
/* post new irecv */
recv_reqs[req_index]= Request::irecv( tmpbuf + realsegsize, count_by_segment,
- datatype, tree->tree_prev,
+ datatype, tree->tree_prev,
COLL_TAG_BCAST,
comm);
-
+
/* wait for and forward the previous segment to children */
- Request::wait( &recv_reqs[req_index ^ 0x1],
+ Request::wait( &recv_reqs[req_index ^ 0x1],
MPI_STATUSES_IGNORE );
-
- for( i = 0; i < tree->tree_nextsize; i++ ) {
+
+ for( i = 0; i < tree->tree_nextsize; i++ ) {
send_reqs[i]=Request::isend(tmpbuf, count_by_segment, datatype,
- tree->tree_next[i],
+ tree->tree_next[i],
COLL_TAG_BCAST, comm );
- }
-
+ }
+
/* complete the sends before starting the next iteration */
- Request::waitall( tree->tree_nextsize, send_reqs,
+ Request::waitall( tree->tree_nextsize, send_reqs,
MPI_STATUSES_IGNORE );
-
+
/* Update the receive buffer */
tmpbuf += realsegsize;
}
sendcount = original_count - (num_segments - 1) * count_by_segment;
for( i = 0; i < tree->tree_nextsize; i++ ) {
send_reqs[i] = Request::isend(tmpbuf, sendcount, datatype,
- tree->tree_next[i],
+ tree->tree_next[i],
COLL_TAG_BCAST, comm);
}
-
- Request::waitall( tree->tree_nextsize, send_reqs,
+
+ Request::waitall( tree->tree_nextsize, send_reqs,
MPI_STATUSES_IGNORE );
}
-
+
/* Leaf nodes */
else {
- /*
+ /*
Receive all segments from parent in a loop:
1) post irecv for the first segment
2) for segments 1 .. num_segments
req_index = req_index ^ 0x1;
tmpbuf += realsegsize;
/* post receive for the next segment */
- recv_reqs[req_index] = Request::irecv(tmpbuf, count_by_segment, datatype,
+ recv_reqs[req_index] = Request::irecv(tmpbuf, count_by_segment, datatype,
tree->tree_prev, COLL_TAG_BCAST,
comm);
/* wait on the previous segment */
- Request::wait( &recv_reqs[req_index ^ 0x1],
+ Request::wait( &recv_reqs[req_index ^ 0x1],
MPI_STATUS_IGNORE );
}
return (MPI_SUCCESS);
}
+
+}
+}
