https://hgpu.org/?p=3648
Compressing Floating-Point Number Stream for Numerical Applications