I thought merging all 3 sort-filter proxy models into one should
speedup dataChanged() signal. As turned out this is not the case. The
time difference is within inaccuracy range, so this doesn't affect
performance.
But I still think it is good to merge them into one proxy model as it
simplifies code.
Torrent numbers were recalculated on every dataChanged() signal. The previous commit
greatly increases the number of dataChanged() signals.
HEAD^^:
Total wall clock: 97.069s
updateTorrentNumbers() time: 0.033s
HEAD^:
Total wall clock: 96.132s
updateTorrentNumbers() time: 0.179s
HEAD:
Total wall clock: 95.535s
updateTorrentNumbers() time: 0.047s
After this commit the time of updateTorrentNumbers() is (almost) back to
the level that it was in HEAD^^.
In commit b50d733 TorrentModel moved from a periodic refresh, to using
postStatusUpdate(). In this transition I forgot to remove emition of
dataChanged() signal for the entire table.
According to my measurements this commit reduce CPU usage of qbittorrent
by a factor of 3:
Before:
Total wall clock: 97.07s
CPU time: 21.77s
- Time spent in TransferListDelegate::paint(): 14.60s
- Time spent in TorrentModel::forceModelRefresh(): 1.44s
- Time spent in TorrentModel::stateUpdated(): 0.02s
After:
Total wall clock: 96.13s
CPU time: 6.68s
- Time spent in TransferListDelegate::paint(): 2.63s
- Time spent in TorrentModel::forceModelRefresh(): <0.01s
- Time spent in TorrentModel::stateUpdated(): 1.73s
As it is seen the time spent in painting is reduced by a factor of 6 (14.60->2.63) at
the cost of slightly increased time of notifications that model is
changed (1.44->1.73). The next commits attempt to address this issue.
The problem is that torrentRow() does linear search over the list of all
available torrents. So it doesn't scale well for large number of
torrents. Removing the copying of QString from linear search
inner loop, speed up it considerably.
The proper solution should be using hash table instead of linear search.
This require more radical changes in TorrentModel and may be done in a
separate commit.
This commit probably fixes#2119.
The only important change in this commit is moving
session::get_ip_filter() from FilterParserThread::processFilterFile() to
FilterParserThread::run(). Previously we called it in main thread, but
now we calls it in worker thread. I don't now what libtorrent contract
about threads, but I assume that if it is ok to set_ip_filter from
other thread, it is ok to get it.
Qt uses binary search to convert string to QColor, we don't need that
binary search at all. This patch could be considered as optimization, but
in reality creating QColor takes only 0.2% of time. So it should be visible
at all.
This could be considered as cleanup for not calling expensive functions
from non-expensive ones.
Ivan Sorokin