Reduce binary size part2#5486
Conversation
…erences, restructure Leiden a little to decrease library size
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
|
/ok to test 5735572 |
ChuckHastings
added
improvement
|
/ok to test e888c3b |
seunghwak
left a comment
seunghwak
left a comment
There was a problem hiding this comment.
LGTM except for minor cosmetic issues.
| rmm::device_uvector<vertex_t>&& d_srcs, | ||
| rmm::device_uvector<vertex_t>&& d_dsts) | ||
| { | ||
| using edge_t = vertex_t; |
There was a problem hiding this comment.
Should we better avoid this?
We currently don't build for edge_t != vertex_t cases in our regular release but we still allow users to manually build for these cases.
There was a problem hiding this comment.
Or is there a guarantee that # edges <= # vertices in this use case? In that case, we should add a comment here to avoid future confusion.
There was a problem hiding this comment.
Yes, I've added a comment. The number of entries in d_srcs/d_dsts starts as the number of vertices in the graph and then gets filtered down, so it's guaranteed to fit in a vertex_t.
|
|
||
| cugraph::graph_t<vertex_t, edge_t, decision_store_transposed, multi_gpu> decision_graph(handle); | ||
|
|
||
| std::optional<rmm::device_uvector<vertex_t>> renumber_map{std::nullopt}; |
There was a problem hiding this comment.
Something very tedious, but in general, it is better to minimize the scope of a variable. We can define this after the shuffle operation below.
There was a problem hiding this comment.
I moved it down.
|
Fix is coming in #5483 |
|
/merge |
2 participants
Add some constexpr keywords to some multi_gpu if blocks to reduce library size of libcugraph_mg. Restructured Leiden refinement code to mirror other algorithms template parameters, then a few changes to reduce library size in Leiden.