I don't believe it removes bias. By pigeon-hole-principle-related reasoning, you cannot map, say, 65536 values to 17 values without bias.

The method redistributes the bias compared to straight modulo.

You won't get a uniform distribution with this, just a differently shaped one compared to modulo.

No, it's not. The method described by ARM is functionally identical, so it is equally biased as the method described in the blog post here.

I see, this is one of several posts on this topic. There is a link near the end to the unbiased method, which is more original, but as you note fixed-point multiplication is not new.