diff --git a/include/pass_mpi_collective.hpp b/include/pass_mpi_collective.hpp
index 8bea3b348ef7818284acc3da387605ba11056e07..a2b3d62ad9f1b8bd1031803e6a0c05880c078e64 100644
--- a/include/pass_mpi_collective.hpp
+++ b/include/pass_mpi_collective.hpp
@@ -31,7 +31,7 @@ struct bb_data
std::set<basic_block> post_dom_front;
int mark1; // for graph parkour
int mark2;
- int collective_rank;
+ int *collective_rank;
};
struct edge_data
@@ -83,8 +83,11 @@ public:
// rank definition
void rank_collective(function *fun);
- int __rank_collective(basic_block bb,
- mpi_collective_code mpi_code, int rank);
+ void __rank_collective(basic_block bb);
+
+ void better_rank_collective(function *fun);
+ int __better_rank_collective(basic_block bb,
+ mpi_collective_code mpi_code, int rank);
private:
// MPI function / collectives detections
diff --git a/src/pass_mpi_collective.cpp b/src/pass_mpi_collective.cpp
index e605fda7dee02a444fa6caedea9de9e2472368e0..7da61f6715183525554c4954d936672556cca40a 100644
--- a/src/pass_mpi_collective.cpp
+++ b/src/pass_mpi_collective.cpp
@@ -66,7 +66,8 @@ unsigned int pass_mpi_collective::execute(function *fun)
alloc_edge_aux(fun);
mark_edge(fun);
- rank_collective(fun);
+ //rank_collective(fun);
+ better_rank_collective(fun);
cfgviz_dump(fun, "_split");
@@ -196,6 +197,13 @@ void pass_mpi_collective::alloc_bb_aux(function *fun)
FOR_ALL_BB_FN(bb, fun)
{
bb->aux = (bb_data *) new bb_data();
+ ((bb_data *) bb->aux)->collective_rank =
+ new int[(int) LAST_AND_UNUSED_MPI_COLLECTIVE_CODE];
+ for (int i = 0; i < LAST_AND_UNUSED_MPI_COLLECTIVE_CODE; ++i)
+ {
+ ((bb_data *) bb->aux)->collective_rank[i] = 0;
+ }
+
}
}
@@ -229,6 +237,7 @@ void pass_mpi_collective::free_bb_aux(function *fun)
{
if (bb->aux != NULL)
{
+ delete ((bb_data *) bb->aux)->collective_rank;
delete (bb_data *) bb->aux;
bb->aux = NULL;
}
@@ -390,28 +399,30 @@ void pass_mpi_collective::free_edge_aux(function *fun)
}
// naive version of mark_edge (work but Omax(2^n))
-//void pass_mpi_collective::mark_edge(function *fun)
-//{
-// mark_edge(ENTRY_BLOCK_PTR_FOR_FN(fun));
-//}
-//
-//void pass_mpi_collective::mark_edge(basic_block bb)
-//{
-// edge e;
-// edge_iterator ei;
-//
-// ((bb_data *) bb->aux)->mark = 1;
-// FOR_EACH_EDGE(e, ei, bb->succs)
-// {
-// if (((bb_data *) e->dest->aux)->mark)
-// {
-// ((edge_data *) e->aux)->loop = true;
-// } else {
-// mark_edge(e->dest);
-// }
-// }
-// ((bb_data *) bb->aux)->mark = 0;
-//}
+void pass_mpi_collective::mark_edge_naif(function *fun)
+{
+ mark_edge(ENTRY_BLOCK_PTR_FOR_FN(fun));
+}
+
+void pass_mpi_collective::mark_edge_naif(basic_block bb)
+{
+ edge e;
+ edge_iterator ei;
+
+ ((bb_data *) bb->aux)->mark = 1;
+ FOR_EACH_EDGE(e, ei, bb->succs)
+ {
+ if (((bb_data *) e->dest->aux)->mark)
+ {
+ ((edge_data *) e->aux)->loop = true;
+ }
+ else
+ {
+ mark_edge(e->dest);
+ }
+ }
+ ((bb_data *) bb->aux)->mark = 0;
+}
void pass_mpi_collective::mark_edge(function *fun)
{
@@ -486,20 +497,101 @@ void pass_mpi_collective::mark_edge(function *fun)
reset_bb_mark(fun);
}
+// WIP for new mark edge
+//struct basic_block_elem {
+// bitmap_head seen;
+// basic_block bb;
+//}
+//
+//struct basic_block_elem *new_bb_elem(){
+// return new
+//}
+//
+//void free_bb_elem(){
+//
+//}
+//
+//
+//void mark_loop(){
+// vect<struct basic_block_elem> pile;
+// struct basic_block_elem *first;
+//
+// first = new_bb_elem();
+// first->bb = ENTRY_BLOCK_PTR_FOR_FN(fun);
+// pile.push(first);
+//
+// while(pile.length() != 0){
+// // mettre a jour edge & already seen
+//
+// }
+//
+//}
+//
+//void pass_mpi_collective::mark_edge(function *fun)
+//{
+// edge e;
+// edge_iterator ei;
+// std::list<basic_block> nodes;
+// basic_block bb;
+//
+//
+//
+//}
+
void pass_mpi_collective::rank_collective(function *fun)
+{
+ __rank_collective(EXIT_BLOCK_PTR_FOR_FN(fun));
+ reset_bb_mark(fun);
+}
+
+void pass_mpi_collective::__rank_collective(basic_block bb)
+{
+ edge e;
+ edge_iterator ei;
+
+ ((bb_data *) bb->aux)->mark1 = 1;
+ FOR_EACH_EDGE(e, ei, bb->preds)
+ {
+ if (((edge_data *) e->aux)->loop)
+ {
+ continue;
+ }
+
+ if (! ((bb_data *) e->src->aux)->mark1)
+ {
+ __rank_collective(e->src);
+ }
+
+ for (int mpi_code = 0;
+ mpi_code < LAST_AND_UNUSED_MPI_COLLECTIVE_CODE;
+ ++mpi_code)
+ {
+ ((bb_data *) bb->aux)->collective_rank[mpi_code] =
+ std::max(((bb_data *) bb->aux)->collective_rank[mpi_code],
+ ((bb_data *) e->src->aux)->collective_rank[mpi_code]);
+ if (((bb_data *) bb->aux)->mpi_code == mpi_code)
+ {
+ ((bb_data *) bb->aux)->collective_rank[mpi_code]++;
+ }
+ }
+ }
+ return;
+}
+
+void pass_mpi_collective::better_rank_collective(function *fun)
{
size_t i;
- int next_rank = 1;
+ int next_rank = 0;
for (i = 0; i < LAST_AND_UNUSED_MPI_COLLECTIVE_CODE; ++i)
{
- next_rank = __rank_collective(EXIT_BLOCK_PTR_FOR_FN(fun),
- (enum mpi_collective_code) i, next_rank);
- reset_bb_mark(fun);
+ next_rank = __better_rank_collective(EXIT_BLOCK_PTR_FOR_FN(fun),
+ (enum mpi_collective_code) i, next_rank);
}
+ reset_bb_mark(fun);
}
-int pass_mpi_collective::__rank_collective(basic_block bb,
+int pass_mpi_collective::__better_rank_collective(basic_block bb,
mpi_collective_code mpi_code, int rank)
{
edge e;
@@ -507,6 +599,8 @@ int pass_mpi_collective::__rank_collective(basic_block bb,
int next_rank;
next_rank = rank;
+ ((bb_data *) bb->aux)->mark1 = mpi_code + 1;
+
FOR_EACH_EDGE(e, ei, bb->preds)
{
if (((edge_data *) e->aux)->loop)
@@ -515,25 +609,25 @@ int pass_mpi_collective::__rank_collective(basic_block bb,
}
// cache or travell
- if (((bb_data *) e->src->aux)->mark1)
+ if (((bb_data *) e->src->aux)->mark1 <= mpi_code)
{
next_rank = std::max(next_rank,
- ((bb_data *) e->src->aux)->mark1);
+ __better_rank_collective(e->src, mpi_code, rank));
}
else
{
next_rank = std::max(next_rank,
- __rank_collective(e->src, mpi_code, rank));
+ ((bb_data *) e->src->aux)->collective_rank[mpi_code]);
}
}
// set mpi code
if (((bb_data *) bb->aux)->mpi_code == mpi_code)
{
- ((bb_data *) bb->aux)->collective_rank = next_rank;
next_rank++;
+ printf("set rank %d at %d for coll %d\n", next_rank, bb->index, mpi_code);
}
- // set mark before recusion
- ((bb_data *) bb->aux)->mark1 = next_rank;
+ ((bb_data *) bb->aux)->collective_rank[mpi_code] = next_rank;
+
return next_rank;
}
@@ -578,7 +672,7 @@ void pass_mpi_collective::cfgviz_internal_dump(function *fun, FILE *out)
if (data->mpi_code != LAST_AND_UNUSED_MPI_COLLECTIVE_CODE)
{
fprintf(out, "|%s (%d) ", mpi_collective_name[data->mpi_code],
- data->collective_rank);
+ data->collective_rank[data->mpi_code]);
}
if (! data->dom.empty())