A faster double to string algorithm

[xjb714]

The following code only contains float/double to decimal and does not contain decimal to string.
Algorithm source code and demonstration link : https://onlinegdb.com/OPKdOpikG
This algorithm is based on the schubfach algorithm and is faster than the dragonbox algorithm.
Thank you.

[xjb714]

If you are interested, I will publish a paper introducing the principle of the algorithm.

[vitaut]

Thanks for sharing. Do I understand correctly that you are the author of this implementation? I would recommend adding it to https://github.com/fmtlib/dtoa-benchmark first so that it can be compared to other algorithms. Paper describing the algorithm would definitely be useful for those interested in using it.

[xjb714]

Yes, I'm the original author.
Code link: https://github.com/xjb714/xjb
Proof of the correctness of the above-mentioned linking algorithm:

fast_f2d.pdf

Algorithm principle process:
算法原理流程:

float can be proved through exhaustive testing and has passed the test. double requires strict mathematical proof and has passed the random number test.
This is the draft of my thesis for your reference. Due to my limited ability, it may not be well-written. Please understand. My English is not very good. Some of the content was translated by a translation software.
float可通过穷举测试证明，已通过测试，double需经过严格数学证明，已通过随机数测试。
这是我写的论文草稿，供大家参考，水平有限，可能写的不好，敬请谅解。我英语不是很好，部分内容由翻译软件翻译。

Regarding the performance aspect, the gcc compiler may have exceptions when measuring the performance of float (the median value in the following figure is 7.84), which is significantly lower than that of icpx and clang. After reviewing the assembly code, I found that the functions in the loop have no inlining and contain parts that seriously affect performance, such as "call memcpy" It might be necessary to modify the benchmark program. Due to my limited time and proficiency, I don't know how to solve it for the time being.
关于性能部分，gcc编译器测量float性能时可能存在异常(下图中值为7.84)，相比icpx和clang有明显差距，我查看汇编代码后发现循环中函数没有内联，且包含“call memcpy“等会严重性能的部分，可能需要修改benchmark程序，我时间和水平有限，暂时不知道怎么解决。

Post a performance test result diagram of "float/double to decimal" :
贴一个"float/double to decimal"性能测试结果图：

Thanks to the authors of yyjson, dragonbox algorithm, schubfach algorithm, grisu algorithm and others.
The principle and inspiration of this algorithm come from these several algorithms. I hope the code I provide can be helpful to everyone.
感谢yyjson的作者、dragonbox算法的作者、schubfach算法的作者、grisu算法的作者等人。
本算法原理和灵感来自于这几个算法，希望我提供的代码能够帮助到大家。

Thank you for reading.
感谢阅读。

[xjb714]

In fact, a complete floating-point number printing algorithm should consist of two processes: (1) binary to decimal (2) decimal to string. The above algorithm does not include the second step. The second step still poses certain challenges, with a large number of branch prediction failures and costly multiplication operations, which take up a considerable amount of time for floating-point number printing.

[vitaut]

cc @jk-jeon in case you find this interesting

[jk-jeon]

@vitaut Thanks. I'm aware of this and planning to look into it later.

I didn't do a thorough review, but I think the algorithm is quite likely correct.

[vitaut]

@xjb714 could you submit a PR to add your algorithm to https://github.com/fmtlib/dtoa-benchmark?

[xjb714]

@vitaut
I'm currently finishing the code writing. I guess it will take a few more days.

I have some questions: There should be a theoretical optimal solution for converting floating-point numbers to the shortest string, but it seems that the optimal solution is not very readable in some cases.

There are two ways to print floating-point numbers :%lf and %le. %lf is output in the common way, and %le is output in the scientific notation form. example: The output results corresponding to the double value 123.45 are "123.45" and "1.2345e2". At this point, the output length of %lf is shorter than that of %le.

Give priority to outputting the shorter option between %lf and %le. If the lengths of %lf and %le are the same, choose %lf.

Suppose the floating-point number v is converted to shortest decimal result, which satisfies the S&W algorithm. d has the shortest significant figure length and its value is a. Here, len(d) is the decimal significant figure length for calculating d.

$$ v=c \cdot 2^q \Rightarrow d \cdot 10^k,d\in N^+ ; a=len(d) ; b=k+\lfloor log_{10}(d) \rfloor; $$ In fact, the following formula can be used to determine whether to choose to output in the %lf format. $$ a+1+(b>=4)+(b>=10)>=abs(b) \text{ and } -3<=b \text{ and } b<=20 $$

Here are a few examples:

double	%lf	%le	a	b	shortest
123.45	"123.45"	"1.2345e2"	5	2	%lf
1000	"1000"	"1e3"	1	3	%le
123	"123"	"1.23e2"	3	2	%lf
123000	"123000"	"1.23e5"	3	5	%lf
1.2e23	"12000...0"	"1.2e23"	2	23	%le
1e100	"1000...0"	"1e100"	1	100	%le
0.0123	"0.0123"	"1.23e-2"	3	-2	%lf
0.001	"0.001"	"1e-3"	1	-3	%le

For instance, the result "1000" in the %lf format of 1000 is more readable than the result "1e3" in the %le format (I think most people would think so as well).But the length of %le is less than that of %lf.

Different algorithms select %lf and %le based on different strategies. How should I choose to format the output in %lf or %le?

[jk-jeon]

@xjb714 Welcome, it's yet another can of worms in this business.

• Milo's dtoa benchmark assumes that the "correct" output must be the one with the smallest number of characters. Many JSON libraries written in that era seems to follow that practice.
• The current C++ standard specifies that the default option of std::format and std::to_chars must print the string with the smallest number of characters, except that its meaning of "smallest" is subtly different from that of Milo's. IIRC it assumes that when exponential form is used, the exponent always needs to include the sign and also must be consisting of at least two digits. So for instance 123 is supposed to be 1.23e+02 when it were to be printed in the exponent format. 1.23e2 is not allowed.
• On the other hand, fmt::format has never worked that way. It chooses fixed-point vs exponential solely based on the value of the decimal exponent, regardless of which one results in a shorter string. IIRC the exact threshold has been changed one or more times over the history.
• Finally, Victor and I recently wrote a paper that proposes to change the specification of std::format and std::to_chars to match the behavior of fmt::format, which is arguably more natural and intuitive.

As a side note, the current specification of std::format/std::to_chars has a funny side effect of mandating "garbage digits" to be printed in certain circumstances which has noticeable performance implication. You can read more about that in the linked paper.

TL;DR
Unless you have a strong reason not to, I think you should use fixed-point when exponent is small, and exponential when exponent is large.

Choosing between them based on the character length is wrong in my humble opinion, especially when garbage digits are involved. I recommend the linked paper for reading which explains some reasons why we prefer this option.

[jk-jeon]

@xjb714 Ah, btw I just noticed that you are using Alexander Bolz's implementation of Dragonbox. I was wondering why you are not including Dragonbox for the float benchmark and now I see the reason.

That implementation is based on an older version of Dragonbox which is a bit slower than the current one. I don't think the current version will outperform Yaoyuan Guo's or yours, but in any case there is no reason to test against an outdated version.