Optimize sort followed by limit #6941
Conversation
jedelbo
force-pushed
the
je/sort-limit
branch
from
aaf922e
to
25e7ba8
Compare
If the limit is much smaller than the total size of the TableView, then it is faster to make a sorted insert into a vector that is kept at the limit size.
jedelbo
force-pushed
the
je/sort-limit
branch
from
25e7ba8
to
19cc731
Compare
|
Once we do #6933 we should be able to optimize this further by only inserting the limit number of elements into the vector in the first place instead of having an extra buffer like you do here. Given that we expect future optimizations, could you please move the benchmark out of the test and into benchmark-common-tasks so that we can track this case over time. If you could post that benchmark's results compared to current master, I'd also be curious to see what kind of improvement this actually is 😄 |
src/realm/sort_descriptor.cpp
Outdated
| IndexPairs buffer; | ||
| buffer.reserve(limit + 1); | ||
| for (auto& elem : v) { | ||
| auto it = std::lower_bound(buffer.begin(), buffer.end(), elem, predicate); |
There was a problem hiding this comment.
If std::ref() is doing something useful in the std::sort() case then it will here too (it makes it so that the predicate is not copied when it's passed to the algorithm).
|
@ironage I am not sure how we can avoid the buffer. We cannot clear the buffer that holds the elements we should sort. |
If the limit is much smaller than the total size of the TableView, then it is faster to make a sorted insert into a vector that is kept at the limit size.
What, How & Why?
☑️ ToDos