 |
The initial results on my machine were: 2.07 hotfix: 48-50 svn8334: 48-50 svn8346: 58-60 svn8401: 35-40 svn8401 with reverted 8380,8381 and r_useMultiDraw 0: 58-60 svnHEAD: 38-40 svnHEAD with reverted 8380,8381 and r_useMultiDraw 0: 58-60 So: 1) performance improved between 8334 and 8346 (most likely due to the change in VertexCache implementation) 2) performance worsened between 8346 and 8401 (due to removal of interaction culling). |
|
|
Some fixes by Duzenko to partly recap the removed culling: svn rev 8454 (40-45 FPS): for off-screen entities, drop interaction if entity and light are in disconnected portal areas svn rev 8456 (45 FPS): with loose shadow bounds, cull by viewport and by light scissor rect svn rev 8459 (57-60 FPS): moved both new culling checks earlier avoids instantiating dynamic models and building shadow geometry |
|
|
Next, the original interaction culling (previously removed in svn rev 8380) was restored with a few small fixes: svn rev 8468: fixed issue with idBounds intersection (see 0002734) --- that could ruin the case when shadow frustum does not hit the portal svn rev 8469: interaction culling in its original form fully restored (note: additional checks added by Duzenko also remain) svn rev 8473: added check for return value of ProjectionBounds inside FloodFrustumAreas (probably not important) After that, it was discovered that FloodFrustumAreas implementation from vanilla D3 is wrong, and can erroneously remove shadows a rare cases: http://forums.thedarkmod.com/index.php?/topic/20195-performance-regression-shadowhide/&do=findComment&comment=442268 The repro map added to SVN as test/FloodFrustumAreas_bug.map Aside from that, the main problem with interaction culling still remained: it could not be used in parallel from multiple threads. Because it cached the list of portal areas found by FloodFrustumAreas in the interaction, with a use of global custom allocator. |
|
|
Finally, interaction culling was changed to make it on-the-fly and threadsafe: svn rev 8475: redone the whole culling code (better read svn commit message) svn rev 8476: two new checks added by Duzenko refactored into proper place (again, see commit message) Also, the new algorithm is not susceptible to the bug mentioned above. Finally, there is a note about a bit of missed opportunities: https://forums.thedarkmod.com/index.php?/topic/20195-performance-regression-shadowhide/&do=findComment&comment=442331 Perhaps would be done later. |
|
|
For me, performance is the same on 2.07 and current SVN, with the updated version of the map. |
- Anonymous
