Updated .plan in RELEASE-NOTES

This commit is contained in:
Robert Beckebans 2023-04-18 21:21:45 +02:00
parent f07d8881b2
commit a7280a4f5f

View file

@ -18,6 +18,75 @@ _______________________________________
TBD - RBDOOM-3-BFG 1.5.0
_______________________________
## .plan - April 18, 2023
Stephen Saunders contributed support for borderlesss fullscreen mode which can also be set by commandline with r_fullscreen -2. It's enabled automatically if you switch from windowed mode to fullscreen by using Alt + Enter. You cant tell the difference visually. However, stick to borderless mode over fullscreen mode if you use multiple monitors.
You can access your displays and launch various apps while your game or work stays on the main display.
This has the advantage that you can better run RBDOOM-3-BFG simultanously along with TrenchBroomBFG, Blender, Krita or any other gamedev weapon of choice.
But, there is a downside to borderless mode, depending on your device.
Most operating systems like Windows, for example, will continue dedicating resources to all background processes.
This means you might get less performance and experience input lag if youre running a game or a demanding app.
If you need to squeeze some extra FPS, go back to real fullscreen mode to see if it makes a difference.
The glTF2 importer has been rewritten to make it more robust. The -transermotion option might not work.
The importer has still some bugs left like -scale only works on certain models.
It is not necessary anymore to specify the name of the armature node.
It will automatically find the rig node but the exported .glb model should have only one skeleton.
The rest was cleaning up the code and removing unused OpenGL and Vulkan code.
Changelog:
* Support Borderless Fullscreen and Borderless Window modes for Win32 and SDL platforms
* Fixed recursive mesh adding in findEntities for glTF2 scenes
* Changed animated glTF2 models to be cached like MD5 meshes so model overlays work on them
* Store glTF2 source timestamp in .bglb binary cache file
* Fixed loading glTF2 models with multiple skinned meshes
* Added -rotate [yaw] glTF2 import option
* Added glTF2 import option -nomikktspace
* If -DSTANDALONE=ON option is set filesystem will look for content/ instead of base/
* Removed stencil shadow volumes with additional fixes by icecoldduke
* Reverted com_smp back to CVAR_BOOL
* Removed unused cvar r_useHalfLambert
* Update README for VS 2022
* Fix Vulkan Vsync mode and handle NUM_FRAME_DATA = 2 or 3
* Fix Vsync mode switching to support all modes: 0 (off), 1 (smart), 2 (on)
* Show shadow atlas stats with com_showFPS 3
* Enable D3D12 Debug Layer with r_useValidationLayers 2
* Removed broken aviDemo code. External tools do a better job on this
* Removed obsolete sdl_glimp.cpp
* Removed many obsolete #if defined( USE_NVRHI )
* Removed renderprogs/ legacy HLSL shaders that were translated to GLSL
* Removed OpenGL code from win_glimp.cpp
* Killed entire HLSL -> GLSL conversion code
* Removed more obsolete Vulkan & OpenGL code
## .plan - March 08, 2023
This build improves the performance again by about 60% through a different V-Sync implementation by Steve Saunders. It seems to require tripple buffering and slightly more memory but elimates the stalls between CPU <-> GPU syncing.