Is it possible to simplify this calculation so that one can get rid of the division? These are all float numbers:

float depthLinear = projMatrix_43 / (depthNonLinear - projMatrix_33);

The denominator depends on depthNonLinear, and you are using it only once. No way to precompute it.

You can construct a crude approximation of function projMatrix_43 / (depthNonLinear - projMatrix_33) given a range of depthNonLinear and constant projMatrix_43 and projMatrix_33 values, but I wouldn't bet on making it both faster than the exact computation (one subtraction and one division) and accurate enough.

That's what I suspected. Thanks for the input!