@KeyJ figuring out derivatives (that part's straightforward), then figuring out which LOD to sample (jfc that part is NOT, especially with the myriad clamps), wrapping calcs, texel address calc, the minor/major axis for aniso (if requested), figuring out how many bilinear samples to take (for "brilinear", everyone's favorite fake trilinear, and aniso), the fucking texel address calcs (even just figuring out where a given mip map starts is a production)... it suuuuucks