+// We put the fractional part of the offset (-0.5 to 0.5 pixels) in the weights
+// because we have to (otherwise they'd do nothing). However, the support texture
+// has limited numerical precision; we'd need as much of it as we can for
+// getting the subpixel sampling right, and adding a large constant to each value
+// will reduce the precision further. Thus, the non-fractional part of the offset
+// is sent in through a uniform that we simply add in. (It should be said that
+// for high values of (dst_size/num_loop), we're pretty much hosed anyway wrt.
+// this accuracy.)
+//
+// Unfortunately, we cannot just do it at the beginning of the shader,
+// since the texcoord value is used to index into the support texture,
+// and if zoom != 1, the support texture will not wrap properly, causing
+// us to read the wrong texels. (Also remember that whole_pixel_offset is
+// measured in _input_ pixels and tc is in _output_ pixels, although we could
+// compensate for that.) However, the shader should be mostly bandwidth bound
+// and not ALU bound, so an extra add per sample shouldn't be too hopeless.
+uniform float PREFIX(whole_pixel_offset);
+