floating point uses a slightly different predictor technique describe here http://chriscox.org/TIFFTN3d1.pdf Signed-off-by: Anton Khirnov <anton@khirnov.net>