In SHM decoder, scaled and ref windows offsets are derived in time of PPS parsing, and both of them are derived in the assumption that all the layers are sammpled as 4:2:0, so coded offsets are simply left shifted by one bit. At the same time text says that multipliers like SubWidthRefLayerC or SubWidthCurrC have to be derived according to chroma format of direct or current layer. Moreover, it seems that valid limits on coded offsets are unnaturally shrinking for 4:4:4 chroma format, if text is read literally.