This should resolve some rendering artifacts potentially caused by
floating point errors when sampling the texture. It should also lead
to crisper images when icons are rendered midway through a pixel.
Previously, we were using a normalized 8-bit unsigned integer which forced us
to represent each increment of the winding number as a fraction of the max
value (1 / 255) which we would then add up using additive alpha blending.
This had three major drawbacks:
- The max winding number could not be greater than 255.
- Adding up (1 / 255) several times could result in a loss of precision.
- Due to also computing anti-aliasing as a fractional winding number, we had to
reduce the max winding number to 32. This was still not good enough because
we would multiply a fractional value with `1 / 32`, thus introducing more and
more loss of precision.
This commit changes the texture type to an `f16` which doesn't require the
division by 255 and enables greater precision in the computation of the
anti-aliased parts of a curve. Note how this also removes the limitation of 255
windings at most per curve. The tradeoff is paying twice as much memory for
storing the texture, but that seems totally valid to achieve rendering accuracy.
Note that this kind of texture should be compatible with WebGL2 once we start
working on a web version of Zed.
