Except that GPUs don't usually read from linear "raster order" images or are very slow doing so. So even in a shared memory CPU/GPU setup, there is a texture "upload" which organizes the pixel data to a GPU cache-friendly block format.