provide custom implementation of metal::partition (for performance) #124
Conversation
ecrypa
force-pushed
the
test/partition-no-workaround
branch
from
1318505 to
bc50803
Compare
|
@ecrypa I notice this is still a draft, what's missing? |
| template<int_... ns, class... vals> | ||
| struct _partitioner<list<number<ns>...>, list<vals...>> | ||
| : _partition_joiner< | ||
| typename _partition_filter<ns>::template type<vals>...> {}; |
There was a problem hiding this comment.
Have we tested if_<number<ns>, pair<list<val>, list<>>, pair<list<>, list<val>>>... vs _partition_filter?
There was a problem hiding this comment.
Have we tested
if_<number<ns>, pair<list<val>, list<>>, pair<list<>, list<val>>>...vs_partition_filter?
I am not absolutely sure about that. I will benchmark it.
I remember why I once introduced that filter: What you propose here is similar to this pattern, which is the "GCC9 killer (v0)". I thus introduced the indirection via _fast_partition_helper in v2, which is faster on GCC9. When extracting the partitioner, I probably renamed _fast_partition_helper into filter.
There was a problem hiding this comment.
I'm happy to merge this as is either way and iterate on it later if necessary, unless you have plans for this PR still?
There was a problem hiding this comment.
Good suggestion! Feel free to merge.
ecrypa
changed the title
No description provided.