Revert r14218. FFMAX is a macro, so

   sum = FFMAX(1, scalar_product(...));
might calculate the scalar product twice, which is slower.

Originally committed as revision 14230 to svn://svn.ffmpeg.org/ffmpeg/trunk