With GLPortal being responsible for all the setup a lot of code was tied to the backend.
Now FDrawInfo will manage the setup and only call pure data generation functions in the actual portal object.
Most importantly, the separate command line options for switching on the legacy buffer handling have been removed.
There's really no need for them anymore, because unlike in earlier versions many of the implementation differences no longer exist - with the exception of where the light and vertex buffer contents are generated.
For testing this, -glversion 3 is sufficient.
This eliminates 3 of the 5 remaining occurences of FQuadDrawer and gets a large section of code out of the GL backend into the common hardware rendering code.
# Conflicts:
# src/gl/scene/gl_scene.cpp
Although this doesn't look as good as the PCF version it is a lot less calculation intensive and therefore more suitable for weaker hardware.
It also tends to bleed through walls a lot less.
Legacy used some strange blending formula to calculate its colormaps for colored 3D floor lighting, this is not available in the software lighting mode, so for these the engine has to temporarily revert to light mode 2 to render them correctly.
- move a few variables from SceneDrawer to FRenderViewpoint.
The global r_viewpoint variable is left alone now to always represent the current viewpoint to the play code.
The main reason behind this change is to reduce the amount of global variables being used by the hardware renderer's scene processing code.
This is both for efficiency and encapsulation. At last on MSVC in 64 bit, accessing global variables is very inefficient and the clipper was doing it repeatedly in its worker functions.
It is also one less place where the global viewpoint gets checked.
This removes 3 uniforms, consisting of 9 floats. Those were merged into other values that never get used at the same time.
It also moves the costly setup of the fixed colormap out of the render state into the 2D processing code.
Since 3D forces use of render buffers now, it is no longer necessary to draw the entire scene with the colormap active, meaning it can be handled more efficiently.
This is now allowed to overestimate the number of vertices to reduce computation time for a rarely occuring special case that was eating most of the time and was causing errors with some walls.
I missed this part when repurposing the vboindex members to store the index buffer offsets.
However, since both indices are needed, they need another set of variables.
The precise way the clipper needs to be maintained may differ between APIs, so it is no longer owned by any render structure but instead HWDrawInfo only contains a reference.
For OpenGL there is still only one static clipper because without multithreaded BSP traversal there is no need for more.
Not only are they better placed in the common code, but they are also both per-viewpoint and not per-scene, so this is a far more suitable place and avoids saving and restoring them in the portal code.
On a fast and modern graphics card this is a lot faster than doing it per subsector but it may not be without drawbacks on older hardware so it will require some testing on older hardware.
For me Frozen Time's view over the bridge went from 46 fps to 51 fps with this change, the time saved was roughly 2 ms.
Game code should never ever call the renderer directly. This must be done through the video interface so that it can also work with other framebuffers later.
Although this is currently safe there is no guarantee that future refactorings will keep the current draw lists, so it's better if GLDecal used its own copy of the data.
- precalculate if a sector's floor and ceiling plane overlap. This avoids rechecking this for each single call of hw_FakeFlat.
- vertices must be marked dirty every time they change after map setup. That means that ChangePlaneTexZ must do this as well, because it cannot rely on interpolation taking care of it.
- Having a 'dirty' argument for SetPlaneTexZ's ZScript version makes no sense. If the value changes from the script side the vertices must always be marked to be recalculated.
This was all over the place, with half of it using the function and half doing incomplete checks on the underlying variables.
Also did some optimization on the IGNOREHEIGHTSEC flag: Putting it on the destination sector instead of the model sector makes the check even simpler and allows to precalculate the effect of 3D floors on the heightsec, which previously had to be run on every call and made the function too complex for inlining.
* only call hw_CheckViewArea if the result is not known yet.
* check the map up front if it even contains heightsecs. This allows to shortcut the above check entirely for maps without sector transfers and will allow further optimizations.
