Map loading speedups #2701
Map loading speedups #2701
Conversation
On a map of 30,000 x 30,000 or more with one layer, tiled would take minutes (during which it was unresponsive). ~90% of that time was calling destructors for the 900,000 QStrings allocated by `QString::split`.
Previously, loading a large map initialized the zoom to arbitrarily small scales. During my testing, a 3000x3000 map would load in at 0.06% zoom. Rendering all those tiles made the first draw and any user actions, including zooming back in, very very slow. This commit caps the number of tiles on screen to about 4,000 for the first draw. The user is still free to zoom out, but the app will start out much more responsive.
By manually iterating, we can avoid allocating a QVector with width*height entries. This speeds up loading times by 50% for 3,000 x 3,000 CSV inputs.
|
Ah, subscript on QStringRef was added Qt 5.7, but we need to support Qt 5.6? I'll switch to using |
QStringRef's [] operator isn't supported on Qt 5.6
There was a problem hiding this comment.
That's a great improvement! No idea why I was still using the plain split function in this code... since I changed many other places to avoid turning the QStringRef into a QString. It's nice that you could even get rid of the trimming by manually parsing the thing.
However, please separate the change to the initial map size into another pull request. I'm not sure about that change since usually even 200x200 maps (so 40,000 tiles) should render at decent speeds. Limiting it to just 4000 visible tiles seems rather severe. I'd in any case want to consider it separately from speeding up the CSV parsing.
This reverts commit 162c324.
|
Removed the map size code. |
While trying to tackle #920, I had issues loading my test files. So I found these performance improvements in the CSV decoder and the first render of the map.
I wasn't able to make any more progress in 920, so this PR is just the speedups I found on the way.