This will make the x86 asm simpler. ARM conversion by Martin Storsjö <martin@martin.st> and Janne Grunau <janne-libav@jannau.net>
The x86 version processes 16 floats per iteration, so len must be a multiple of 16.