fast_float VS ryu

> Google put in significant engineering effort into "Ryu", a parsing library for double-precision floating point numbers: https://github.com/ulfjack/ryu

It's not a parsing library, but a printing one, i.e., double -> string. https://github.com/fastfloat/fast_float is a parsing library, i.e., string -> double, not by Google though, but was indeed motivated by parsing JSON fast https://lemire.me/blog/2020/03/10/fast-float-parsing-in-prac...

For those that don't know, gcc 12.x updated its float parsing logic to something similar to fast_float and it's about 1/6 of the cost presented here (sub 100 in the graph presented here). Strongly suggest using that library or upgrading the compiler if you need the performance.

This makes sense for integers but betware floating point from_chars - libc++ still doesn't implement it and libstdc++ implements it by wrapping locale-dependent libc functions which involves temporarily changing the thread locale and possibly memory allocation to make the passed string 0-terminated. IMO libstdc++'s checkbox "solution" is worse than not implementing it at all - user's are better off using Lemire's API-compatible fast_float implementation [0].

[0] https://github.com/fastfloat/fast_float

I'm a bit stuck on how to do the same thing in c++, due to containers only having a single type. The very inefficient way I'm currently doing it is by passing a program as a vector of strings, and then converting the string constants to doubles with the fast_float library.

Just above this comment is a merged PR, which references fast_float library: https://github.com/fastfloat/fast_float

Daniel Lemire @lemire (creator of the algorithm, author of the C++ implementation, and provided constant feedback to help guide the PR).

And out of 6,000 lines in the file, at least 3000 are other people's code: earcut for polygon triangulation and fast_float because .obj files typically contain a lot of floating point numbers so it's important to parse them quickly.

How this compares to https://github.com/fastfloat/fast_float ?

I tried to compare it against Daniel Lemire's excellent fast_float library. Fast float took about 180ms for the same program, and all I did was change "std" namespace prefix to "fast_float". It's a factor of 12 difference, at least my machine. I tried MSVC next, and it is a lot better, but it is still ~4 times slower than fast float. AFAIK, clang currently does not implement the feature at all.

If you don't mind a 3rd party lib until your stdlib updates, https://github.com/fastfloat/fast_float is best-in-class.

Nah. This is about ryu printf.

> Google put in significant engineering effort into "Ryu", a parsing library for double-precision floating point numbers: https://github.com/ulfjack/ryu

It's not a parsing library, but a printing one, i.e., double -> string. https://github.com/fastfloat/fast_float is a parsing library, i.e., string -> double, not by Google though, but was indeed motivated by parsing JSON fast https://lemire.me/blog/2020/03/10/fast-float-parsing-in-prac...

Me and Ryu agree that the answer should be 0.30000000000000004

Apparently exact minimal float-to-string conversion is more recent than I thought, and many languages used to print more (Python?) or less (PHP) decimal digits than necessary to uniquely identify the bit pattern. Python correctly prints 46000.80 + 553.04 as 46553.840000000004, but I don't know if it ever prints more digits than needed. One recent algorithm for printing floats exactly is https://github.com/ulfjack/ryu, though I'm unaware what's the state-of-the-art (https://github.com/jk-jeon/dragonbox claims to be a benchmark and the best algorithm).

On the huge speedup side, you have the Ryū algorithm for decimal conversion (Video, Source), which is now finding its way in most standard libraries. But it isn't a hack, and a very dense, complex and precise algo, nothing like the fast-and-loose inverse square root.

Used a wizard's magic to print "3.14" faster

Here's a paper that details an optimized algorithm (reference implementation). It also contains a description of a correct, but slow algorithm as well as references to classic papers on the subject. Earlier the classic implementation was the dtoa one included in netlib by David Gay.

At the very core of all these theoretical stuffs, there is the theory of continued fractions. This is an immensely useful monster which I even dare call as the ultimate tool for floating-point formatting/parsing that everyone who wants to contribute in this field should learn. Before I learned continued fractions, my main tool for proving stuffs was the minmax Euclid algorithm (which is one of the greatest contributions of the wonderful Ryu paper), but it turns out that it is actually just a quite straightforward application of the theory of continued fractions. The main role minmax Euclid algorithm played was to estimate the maximum size of possible errors, but with continued fractions it is even possible to find the list of all examples that generate errors above a given threshold. This is something I desperately wanted but really couldn't do back in 2020.

Ryū algorithm, the converse (doubles to strings), is also much faster than using Java's number formatting classes.

https://github.com/ulfjack/ryu/blob/master/src/main/java/inf...