can we fallback to a sub-group level implementation?

-> coord to offset functions for all image dims, note that coord is assumed to be clamped and a floor vector 1D, 1D buffer

provide copysign implementation (not available in SPIR-V/GLSL)

returns the number of components that are equal to value

creates a new command pool, returning a *non-owning* reference to it

create a compute-only queue for the specified device NOTE: this is only relevant on backends that a) offer graphics supports and b) offer compute-only

returns an int vector containing the number of trailing 0-bits of each component in this vector

cubic interpolation (with *this = a, interpolation between a and b, and values in order [a_prev, a, b, b_next])