Q Numbers Redux
2024-8-29 03:0:0 Author: www.tbray.org(查看原文) 阅读量:10 收藏

Back in July I wrote about Q numbers, which make it possible to compare numeric values using a finite automaton. It represented a subset of numbers as 14-hex-digit strings. In a remarkable instance of BDD (Blog-Driven Development, obviously) Arne Hormann and Axel Wagner figured out a way to represent all 64-bit floats in at most ten bytes of UTF-8 and often fewer. This feels nearly miraculous to me; read on for heroic bit-twiddling.

Numbits · Arne Hormann worked out how to rearrange the sign, exponent and mantissa that make up a float’s 64 bits into a big-endian integer that you probably couldn’t do math with but you can compare for equality and ordering. Turn that into sixteen hex digits and you’ve got automaton fuel which covers all the floats at the cost of being a little bigger.

If you want to admire Arne’s awesome bit-twiddling skills, look at numbits.go. He explained to me how it works in some chat that I can’t find, and to be honest I can’t quite look at this and remember the explanation, but the unit tests are good; it works.

    u := math.Float64bits(f)
    // transform without branching
    // if high bit is 0, xor with sign bit 1 << 63,
    // else negate (xor with ^0)
    mask := (u>>63)*^uint64(0) | (1 << 63)
    return numbits(u ^ mask)

Arne called these “numbits” and wrote a nice complete API for them, although Quamina just needs .fromFloat64() and .toUTF8(). I and Arne both thought he’d invented this, but then he discovered that the same trick was being used in the DB2 on-disk data format years and years ago. Still, damn clever, and I’ve urged him to make numbits into a standalone library.

We want less! · We care about size; Among other things, the time an automaton takes to match a value is linear (sometimes worse) in its length. So the growth from 14 to 16 bytes made us unhappy. But, no problemo! Axel Wagner pointed out that if you use base-128, you can squeeze those 64 bits into ten usable bytes of UTF-8. So now we’re shorter than the previous iteration of Q numbers while handling all the float64 values…

But wait, there’s more! Arne noticed that for purposes of equality and comparison, trailing zeroes (0x0, not ‘0’) in those 10-byte strings are entirely insignificant and can just be discarded. The final digit only has 1/128 chance of being zero, so maybe no big deal. But it turns out that you do get dramatic trailing-0 sequences in positive integers, especially small ones, which in my experience are the kinds of numbers you most often want to match. Here’s a chart of the length of the lengths the of numbits-based Q numbers for the integers zero through 100,000 inclusive.

LengthCount
11
21
3115
47590
592294

They’re all shorter than 5 until you get to 1,000.

Unfortunately, none of my benchmarks prove any performance increase because they focus on corner cases and extreme numbers; the benefits here are to the world’s most boring numbers, namely small non-negative integers.

Here I am, well past retirement age, still getting my jollies from open-source bit-banging. I hope other people manage to preserve their professional passions into later life.


picture of the day

By .

The opinions expressed here
are my own, and no other party
necessarily agrees with them.

A full disclosure of my
professional interests is
on the author page.

I’m on Mastodon!


文章来源: https://www.tbray.org/ongoing/When/202x/2024/08/28/Q-Numbers-2
如有侵权请联系:admin#unsafe.sh