(*Image).WritePixels doens't send a command to the queue immediately
but caches commands internally. However, the package atlas assumed
that pixel data was sent to the cache every end of a frame. Then, byte
slices for pixels were corrupted.
This change fixes the issue by resolving all the images when flushing
commands.
Closes#2390
DrawTriangles was introduced at #1508, and apparently there is no
reason we should use ReplacePixels here. So, simplify the logic by
using only DrawTriangles.
From the reported stack trace, there could be a potential issue in
atlas.BeginFrame: images were manipulated before the images are restored.
Restoring images assumes that all images are not stale, but manipulating
images like putOnAtlas might cause other images stale in ReplacePixels.
Though we failed to reproduce the case, this fix should make sense.
Updates #2075
Now pixels are adjusted even when the graphics driver doesn't have
high-precision floats, but this should not be problematic. This was
introduced at 9bff33472a, but the
adjusting way is much different from the current way.
Updates #879Closes#1820
With DirectX, the graphics driver cannot be determined until the
main loop starts, as a transparent window cannot be treated with
DirectX so far. On the other hand, compiling shaders requires a
graphics driver as it requires information about Y directions of
NDCs and framebuffers.
This change delays compiling shaders until the graphics commands
are actually executed in the main loop.
Updates #1007
Updates #2019
Note that this applies only to the builtin shaders - interface for Kage stays
unchanged for compatibility.
Minor compatibility delta: when interpolating alpha values, previous code has
created nonsense values, such as, when interpolating from
fully-transparent-black (0,0,0,0) to opaque-white (1,1,1,1), something like
half-transparent-grey (0.25,0.25,0.25,0.5) where half-transparent-white
(0.5,0.5,0.5,0.5) is used by the new code.
I assume this is a strict improvement, however this may warrant some testing.
Possible later improvement could be moving the premultiplication from fragment
shader to CPU. Did not do this as it makes the code rather inconsistent of Kage
vs built-in shader usage.
Updates #1772
