[DRAFT][compiler] Introduce EliminateDeadSubgraphPass #13628
Conversation
This pr introduces EliminateDeadSubgraphPass for removing dead subgraph. ONE-DCO-1.0-Signed-off-by: Artem Balyshev <[email protected]>
BalyshevArtem
force-pushed
the
eliminate_dead_graph_pass_draft
branch
from
6d834f3 to
765fad1
Compare
| @@ -515,7 +515,7 @@ int entry(int argc, char **argv) | |||
| luci::change_outputs(graph, new_outputs); | |||
| } | |||
|
|
|||
| // call luci optimizations for module | |||
| // call luci optimizations for module before optimizations for graph | |||
There was a problem hiding this comment.
why is this comment changed to "before..." ?
There was a problem hiding this comment.
Just to emphasize that after graph optimization there will be another optimization of the entire module. If it only makes it more confusing, I can remove this comment :)
There was a problem hiding this comment.
comment can give misunderstanding that one "MUST..."
if not "MUST", please restore as before.
| auto *while_node = dynamic_cast<luci::CircleWhile *>(circle_node); | ||
| assert(while_node != nullptr); |
There was a problem hiding this comment.
we can use loco::must_cast for this case
| auto *if_node = dynamic_cast<luci::CircleIf *>(circle_node); | ||
| assert(if_node != nullptr); |
| assert(body_graph_index >= 0); | ||
| assert(cond_graph_index >= 0); | ||
| // Add indexes into queue | ||
| reachable_graphs_indexes_q.push_back(size_t(body_graph_index)); |
There was a problem hiding this comment.
Q) why use size_t(body_graph_index) ? this will always be 4
plz use static_cast<size_t>
|
Q) why separate EliminateDeadSubgraphPass to another draft? |
In order to separate these two new optimizations. The current draft is completely independent of the previous one) |
But the FuseGeluPass will only work correctly with this pass, or there will be needless subgraphs. |
No, we need to get rid of dead graphs, as this greatly inflates the binary size of the model. |
This pr introduces EliminateDeadSubgraphPass for removing dead subgraph.
for issues: #12263 and for #13365
from draft: #13602
ONE-DCO-1.0-Signed-off-by: Artem Balyshev [email protected]