Texture mapping question...

Ok, I''m currently converting a "perfect" texture mapping routine to one that only corrects the perspective every nth pixel. I''ve got it working, and it''s pretty fast, over twice as fast as when every pixel is perspective corrected, but I''m getting an awful amount of "mess". I was looking at Quake (in software mode), which corrects only every 16 pixels, and it looks almost flawless, but when I correct every say, 8 pixels the whole thing just loses about 25% accuracy (rough estimate). I''m working out u/z v/z and 1/z at each vertex and interpolating them along the edges, all using floating point maths, then in my line drawing function I''m working out a correct value (again with floating point maths) for every nth pixel and then interpolating in-between using integers. Since I''m using integers when interpolating I''m obviously bit shifting to maintain accuracy. Is this the correct technique? Thanks, Simon.