Sharing the temporary `row` and `rowSwap` buffers between different calls to the
processor is not safe because it will be used by multiple threads.
As a result some of the rows in the image could randomly get corrupted.
This commit gets rid of the cached buffers (the value of which questionable
anyway) and simply re-uses the same code which other processors use.
No clue why `getResourceLocationForPlayerUUID` was implemented the way it was,
that commit dates back to 2015 (332c36a), but if the skin isn't loaded yet, it
won't load it, so sometimes it just won't show properly in first person and in
the Player Overview.
This commit just calls `AbstractClientPlayer.getSkinTexture` instead.
Unclear which MC versions are affected. At some point before 1.16.4 something
else was introduced which loads the skin for us, hence why modern versions were
not affected.
On my system, a single scroll tick results in 120 calls, which is fine for the
vanilla camera controller because it has a range of 2000, but the classic
controller only has a range of 36, making it impossible to select anything but
the extremes.
This commit changes the size of the range to match that of the vanilla controller.
That way we can simply use whatever version we want and aren't tied to the
version of lwjgl which MC provides.
This fixes the OpenEXR export on 1.19 crashing either due to the windows
workaround or due to differing lwjgl versions (depending on OS).
And also enables OpenEXR export on MC versions using lwjgl2 (1.12.2 and below).
Until now, recording worked by injecting right before the vanilla packet
handler. At that point it was receiving packet classes and had to re-encode them
to store them in the replay file. However, not all decoded packets can
necessarily be re-encoded without errors, requiring us to have a bunch of
workarounds.
This commit changes the way recording works by injecting right before the
decoder. It therefore receives the raw bytebufs and no longer has to deal with
the re-encode issue.
Terrain culling is only updated when the camera position or rotation changes.
For omnidirectional rendering, we don't change the camera directly though, we
rotate way earlier so we can consistently get before any other effects. As a
result, the terrain frustum culling wasn't updated between render passes
resulting in missing terrain.
This was largely masked by the fact that we used to call GameRenderer.onResize
every pass (even when the size hadn't changed) but fixing that made this issue
become apparent.
Instead of setting the the values via accessor and calling the handler directly.
This allows mods like ResolutionControl+ to properly resize its internal
framebuffers as well.
The only thing we care about is the framebuffer size, cause that will differ
between game and progress gui rendering. The window size will be the same (real)
size for both.
This makes it work more like regular 360 mode and is generally more compatible
because it doesn't require the frustum culling workarounds which the GPU
solution needs.
In particular, this fixes frustum culling with newer sodium versions (which no
longer use the vanilla intersection checking method) and fixes the clouds on
modern Iris (which seems to be using a different shader program for clouds).
Fixes an issue where the world border would move at normal speed when speeding up or slowing down the replay, and also at an incorrect speed when rendering the replay, causing it to get out of sync with the rest of the world.
Co-authored-by: Jonas Herzig <me@johni0702.de>
Should be more compatible with mods which replace the entire renderHand
method (thereby making our mixin non-applicable), at least as long as these mods
check `renderHand` as well (Iris explicitly does).
This will become important once #638 is fixed.
