Faster float formatting #768
Conversation
This commit provides an alternative implementation for a float → decimal conversion. It integrates a C implementation of Fabian Loitsch's Grisu-algorithm [[pdf]](http://florian.loitsch.com/publications/dtoa-pldi2010.pdf), extracted from https://github.com/night-shift/fpconv. The relevant files are added in this PR, they are, as is all of https://github.com/night-shift/fpconv, available under a MIT License. As a result, I see a speedup of 900% on Apple Silicon M1 for a float set of benchmarks. floats don't have a single correct string representation: a float like `1000.0` can be represented as "1000", "1e3", "1000.0" (and more). The Grisu algorithm converts floating point numbers to an optimal decimal string representation without loss of precision. As a result, a float that is exactly an integer (like `Float(10)`) will be converted by that algorithm into `"10"`. While technically correct – the JSON format treats floats and integers identically –, this differs from the current behaviour of the `"json"` gem. To address this, the integration checks for that case, and explicitely adds a ".0" suffix in those cases. This is sufficient to meet all existing tests; there is, however, a chance that the current implementation and this implementation occasionally encode floats differently. ``` == Encoding floats (4179311 bytes) ruby 3.4.1 (2024-12-25 revision 48d4efcb85) +YJIT +PRISM [arm64-darwin24] Warming up -------------------------------------- json (local) 4.000 i/100ms Calculating ------------------------------------- json (local) 46.046 (± 2.2%) i/s (21.72 ms/i) - 232.000 in 5.039611s Normalize to 2090234 byte == Encoding floats (4179242 bytes) ruby 3.4.1 (2024-12-25 revision 48d4efcb85) +YJIT +PRISM [arm64-darwin24] Warming up -------------------------------------- json (2.10.2) 1.000 i/100ms Calculating ------------------------------------- json (2.10.2) 4.614 (± 0.0%) i/s (216.74 ms/i) - 24.000 in 5.201871s ``` These benchmarks are run via a script ([link](https://gist.github.com/radiospiel/04019402726a28b31616df3d0c17bd1c)) which is based on the gem's `benchmark/encoder.rb` file. There are probably better ways to run benchmarks :) My version allows to combine multiple test cases into a single one. The `dumps` benchmark, which covers the JSON files in `benchmark/data/*.json` – with the exception of `canada.json` – , reported a minor speedup within statistical uncertainty.
|
Thanks for this, I was hoping to write a c99 version of dragonbox, but it's a lot of work and Grisu is a huge improvement already. I will have to consult a a few people about licensing though, because |
|
@byroot Certainly. From the three PRs I am prepping (fixnum, floats, strings) this is the least important from our actual use-cases; I believe our API is not using floats much. In any case, I guess, it is interesting to see what the gem can achieve by using grisu (or dragonbox, which, I believe, is roughly equivalent.) |
|
The same licensing concern is present for your other PRs. I suspect it's fine, but it's tricky to just mix licenses like that. |
@byroot I understand your concerns. I think the situation is a bit different across the three PRs:
I believe these licenses require attribution, but otherwise do not limit distribution and derivation, and especially should be compatible with GPLed code.) I don't know anything about ruby's governance rules re licenses – but please let me know if I can help you address these concerns. (Note that I also have an alternative to #767 without any licensing woes, which I am ready to share. The gains, however, are maybe only half of what I could see in the proposed PR.) I have seen some CI builds failing across the gems: there are some older macOS and generally Windows versions fail. I will hold off with any additional work until we agree on a way forward (or, well, having no way forward due to these licensing concerns), but I'll happily pick these up if there is a chance that my changes land in the gem. To give you some context why these three PRs land seemingly out of the blue: they are the result of ~one month of curiosity (triggered by your blog post thanks for that!) during which I couldn't be online often. I would have reached out to you otherwise about any licensing concerns. Again, let me know if I can help you address these concerns. |
|
Someone pointed me to https://github.com/ruby/ruby/blob/master/LEGAL, which list all the licenses used across the Ruby repo. There's already a bunch of files using various BSD style licenses, so I think this is OK. |
| /* fpconv_dtoa converts a float to its shorted string representation. When | ||
| * converting a float that is exactly an integer (e.g. `Float(2)`) this | ||
| * returns in a string that looks like an integer. This is correct, since | ||
| * JSON treats ints and floats the same. However, to not break integrations | ||
| * that expect a string representation looking like a float, we append a | ||
| * "." in that case. | ||
| */ | ||
| if(!memchr(d, '.', len) && !memchr(d, 'e', len)) { | ||
| d[len] = '.'; | ||
| d[len+1] = '0'; | ||
| len += 2; |
There was a problem hiding this comment.
It's a bit unfortunate to have to search what we just wrote. Have to tried to check if fpconv_dtoa could be modified to always add the .0?
| The contents of this directory is extracted from https://github.com/night-shift/fpconv | ||
|
|
||
| It is licensed under the provisions of the Boost Software License - Version 1.0 - August 17th, 2003. See the ./license file for details. |
There was a problem hiding this comment.
Maybe it's simpler to just add the license as comments at the top of fpconv.c?
This commit provides an alternative implementation for a float → decimal conversion.
It integrates a C implementation of Fabian Loitsch's Grisu-algorithm [pdf], extracted from https://github.com/night-shift/fpconv. The relevant files are added in this PR, they are, as is all of https://github.com/night-shift/fpconv, available under a MIT License.
As a result, I see a speedup of 900% on Apple Silicon M1 for a float set of benchmarks.
floats don't have a single correct string representation: a float like
1000.0can be represented as "1000", "1e3", "1000.0" (and more).The Grisu algorithm converts floating point numbers to an optimal decimal string representation without loss of precision. As a result, a float that is exactly an integer (like
Float(10)) will be converted by that algorithm into"10". While technically correct – the JSON format treats floats and integers identically –, this differs from the current behaviour of the"json"gem. To address this, the integration checks for that case, and explicitely adds a ".0" suffix in those cases.This is sufficient to meet all existing tests; there is, however, a chance that the current implementation and this implementation occasionally encode floats differently.
These benchmarks are run via a script (link) which is based on the gem's
benchmark/encoder.rbfile. There are probably better ways to run benchmarks :) My version allows to combine multiple test cases into a single one.The
dumpsbenchmark, which covers the JSON files inbenchmark/data/*.json– with the exception ofcanada.json– , reported a minor speedup within statistical uncertainty.