https://hgpu.org/?p=15626
Wanted: Floating-Point Add Round-off Error instruction