
We've got pixel-wide world-space derivatives which we can use in the perspective camera sampling. This allows to get rid of two calls to transform_direction() function. In theory we can save two transform_perspective() calls if we'll also save pre-calculated camera-space dx/dy.