We should consider putting these in a header file with other useful utilities that we can import as needed. Our shaders so far have been extremely simple so it sort of didn't matter yet.

So I've been thinking: Should we have a standard shader library that ships with impellerc itself and slowly graduate routines that get repeated use to it? Basically a directory that's always guaranteed to be in the include path and defines "namespaced" functions:

#include "impeller/distance.glsl"

...

float my_sphere = ImSphereDistance(space, vec3(0), 1.0);

Perhaps worth discussing over a bug (and not sure if this is the best naming convention :)). A few other things that are good candidates: Sampler mode emulation for legacy platforms (i.e. ClampToBorder), common sigmoids, erf, gaussian/normal distribution, color space conversions, cheap hash/noise functions, simplex noise.

The need for something like this is coming up in my work to address differences in coordinate spaces between OpenGL and Metal. https://docs.unity3d.com/Manual/SL-PlatformDifferences.html is probably recommended reading for inspiration as it documents some of the patterns you are thinking of as well as built-ins.

In any case, we can handle this at a later time. Just thought we might be at a point where our shaders are not so simple anymore.

I tried to profile the performance of this kernel in the shader profiler but it only seems to be available on iOS and TVOS (and perhaps the M1 Macs)? It would be interesting to take a stab at optimizing this using a profiler. I bet there are many low hanging fruit.

Oh definitely, this can get way faster. One particularly egregious example is that the rotations should be computed once, but they're computed during distance evaluation time (400x+ more evaluations than necessary).

Verify that the slice is 0 if the descriptor format is anything but kTextureCube?

Done.

There are two steps to take to disable the OpenGLES backend. The first is to make sure the test skips compilation (if you want it to of course). The second is to GTEST_SKIP the test with a message when the test is compiled and needs to be skipped at runtime. See the GTEST_SKIP pattern used for disabling the instancing test above. This is also why there is a presubmit failure.

Done.

I think the subclass should verify the slice is appropriate for the descriptor before letting the subclasses set the contents for that slice. Perhaps add a new pure virtual OnSetContents and make this non-virtual?

Agreed and done.