John Yates <john@xxxxxxxxxxxxxxxx> writes: > On Mon, Jan 6, 2020 at 10:20 AM Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote: >> Or do a middle ground thing where we use 32-bit arithmetic >> for the per-station weights, but go to 64-bit for the weight sum? I >> don't really have a good grip on how much of a performance impact we're >> talking about here, so I'm not sure which I prefer... > > Double width accumulation is very common in many applications. > Double width addition and comparison are _much_ cheaper than > double width multiplication and division. Yeah, we'd be doing the accumulation in 64bit values in any case; we're talking about mainly multiplication here (the whole point of the reciprocal stuff is to get the division out of the fast path). So how big of an impact is one (or two) extra 64-bit multiplications going to have on a 32bit platform? -Toke
