There is a better algorithm here: https://github.com/jk-jeon/dragonbox. Please have a look.
This library is a reference implementation of Grisu-Exact in C++.
Grisu-Exact is a variant of Grisu. Unlike Grisu, Grisu-Exact always produces shortest and correctly rounded outputs. Development of this algorithm was inspired by Ryu.
(Note that Grisu is indeed already a "correct" algorithm in the sense that it does provide roundtrip guarantee. But Grisu-Exact provides more than that; see below.)
Grisu-Exact generates a pair of integers from a floating-point number: the decimal significand and the decimal exponent of the input floating-point number. These integers can then be used for string generation of decimal representation of the input floating-point number, the procedure commonly called ftoa
or dtoa
.
The algorithm guarantees three things:
-
It has the roundtrip guarantee; that is, a correct parser interprets the generated output string as the original input floating-point number.
-
The output is of the shortest length; that is, no other output strings that are interpreted as the input number can contain less number of significand digits than the output of Grisu-Exact.
-
The output is correctly rounded: the number generated by Grisu-Exact is the closest to the actual value of the input number among possible outputs of minimum number of digits.
Although Grisu-Exact is intended for float-to-string conversion routines, the actual string generation is not officially a part of the algorithm. Grisu-Exact just outputs two integers (the decimal significand/exponent) that can be consumed by a string generation procedure. The header file grisu_exact.h
includes everything needed for this. It is header-only; you just need these two steps for using it:
- Drop
grisu_exact.h
in your include directory, and #include
it. That's it.
Nevertheless, a string generation procedure is included in the library. There are two additional files needed for that: fp_to_chars.h
and fp_to_chars.cpp
(the .cpp
file is in the directory fp_to_chars
). If you want to use them too, then:
- In addition to
grisu_exact.h
, dropfp_to_chars.h
in your include directory, #include
fp_to_chars.h
instead ofgrisu_exact.h
,- Build
fp_to_chars.cpp
as a library, and then - Link against it.
However, note that features provided in fp_to_chars.h
/.cpp
are quite limited and subject to change. Currently, there is no way to indicate any formatting options. Every float/double input is just printed as something like -3.34E3
. No option for putting plus sign (+) in front of positive significands or positive exponents, no way to use e
instead of E
, no option for printing in a fixed-point form (like -3340
), etc.. Also there is no way to precompute the minimum required size of buffer. I don't like this rigidity, and hope to add more customization features later. Yet, you can rather use grisu_exact.h
directly to implement your own string generation routine.
Besides these three, any other files are unncessary for real use.
A comprehensive reference for this library is in preparation.
The library is targetting C++17 and actively using its features (e.g., if constexpr
). For those who want, there is also a C++11-compatible implementation: https://github.com/jk-jeon/Grisu-Exact11.
(Simple string generation from float/double
)
#include "fp_to_chars.h"
double x = 1.234; // Also works for float
char buffer[31]; // Should be long enough
// Null-terminate the buffer and return the pointer to the null character
// Hence, the length of the string is (end_ptr - buffer)
// buffer is now { '1', '.', '2', '3', '4', 'E', '0', '\0', (garbages) }
char* end_ptr = jkj::fp_to_chars(x, buffer);
// Does not null-terminate the buffer; returns the next-to-end pointer
// buffer is now { '1', '.', '2', '3', '4', 'E', '0', (garbages) }
// you can wrap the buffer with things like std::string_view
end_ptr = jkj::fp_to_chars_n(x, buffer);
(Direct use of jkj::grisu_exact
)
#include "grisu_exact.h"
double x = 1.234; // Also works for float
// Here, x should be a nonzero finite number!
// The return value v is a struct with three members:
// significand : decimal significand (1234 in this case);
// it is of type std::uint64_t for double, std::uint32_t for float
// exponent : decimal exponent (-3 in this case); it is of type int
// is_negative : as the name suggests; it is of type bool
auto v = jkj::grisu_exact(x);
By default, jkj::grisu_exact
returns a struct with three members (significand
, exponent
, and is_negative
). On the other hand, if you call jkj::grisu_exact<false>
instead of jkj::grisu_exact
, then it returns a struct with two members (significand
and exponent
). It seems that using this second version instead of the plain jkj::grisu_exact
results in a faster string generation. For an example usage of this jkj::grisu_exact<false>
, please refer to fp_to_chars.h
.
This implementation provides various rounding modes. Rounding mode is the rule that determines the interval represented by a single bit pattern. jkj::grisu_exact
utilizes (a form of) the tag dispatch technique to provide the rounding mode selection feature. (This feature might be considered as a kind of just proof-of-concepts, because it is usually of no harm to just assume round-to-nearest, tie-to-even rounding mode is being used.) The second parameter to jkj::grisu_exact
is responsible for the rounding mode, and the default argument is jkj::grisu_exact_rounding_modes::nearest_to_even{}
. (Here, jkj::grisu_exact_rounding_modes::nearest_to_even
is the tag type corresponding to the round-to-nearest, tie-to-even rounding mode.)
As another proof-of-concepts feature, it is also possible to specify the behavior of correct rounding search. The third parameter to jkj::grisu_exact
is the tag paramater specifying this. The default argument is jkj::grisu_exact_correct_rounding::tie_to_even{}
, which means that whenever there are two shortest outputs with the identical distance from the true value, the even one is chosen and the odd one is dropped. There are several other settings available. You can also completely disable correct rounding search by giving jkj::grisu_exact_correct_rounding::do_not_care{}
as the third parameter to jkj::grisu_exact
. This might be useful if you do not care about correct rounding guarantee and better performance is more important to you. But the performance difference is not that huge (maybe about 1-3ns
).
In my machine, it defeats or is on par with other contemporary algorithms including Ryu. Especially, Grisu-Exact is faster than Ryu if the output string contains small number of digits. It can perform slightly slower or faster depending on the situation than Ryu when the output string contains near the maximum possible number of digits. For the case of binary64 format (aka double
), it consistently outperforms Ryu up to 15 numbers of digits, and for the case of binary32 format (aka float
), it consistently performs worse than Ryu if the number of digits is more than 7.
The following benchmark result is obtained using Milo's dtoa benchmark framework (https://github.com/miloyip/dtoa-benchmark). The complete source code for the benchmark below is available here.
There is also a benchmark done by myself:
(top: benchmark for
float
data, bottom: benchmark for double
data. Solid lines are the averages, dashed lines are the medians, and the shaded regions show 30%, 50%, and 70% percentiles.)
Here is another performance plot with uniformly randomly generated float
(top) or double
(bottom) data:
Information about these benchmarks can be found in Section 5 of this paper.
Please see this paper.
In order to run tests and benchmarks, you need .cpp/.h
files in the directories tests
and benchmarks
, in addition to grisu_exact.h
, fp_to_chars.h
, and fp_to_chars.cpp
. There is no third party dependencies other than those included in this repository, so this should be enough.
In main.cpp
(which is in tests
directory), there are bunch of #define
's. Uncomment whatever you want to test or benchmark, compile and link every .cpp
files mentioned.
The result of tests and benchmarks will be written in the directories test_results
and becnhmark_results
respectively, and as std::ofstream
cannot create a new directory, those directories should exist before running the test.
There are also some MATLAB scripts in the directory benchmark_results
for plot generation. If you have MATLAB installed on your machine and want to generate plots, then download these script files also.
I also developed an algorithm for the reverse procedure, that is, to convert decimal floating-point numbers into IEEE-754 binary floating-point numbers, using a similar idea. With the assumption of limited precision (at most 9 digits for float
's, at most 17 digits for double
's), this reverse procedure is actually far simpler than the binary-to-decimal conversion, because we do not need to search for the shortest representation. I ran a joint test of this algorithm and Grisu-Exact, and confirmed that these algorithms successfully round-trip for every single float
. (Since both of the algorithms are very fast, the whole test takes only less than 2 minutes😉) Therefore, I am very confident about the correctness of both of the algorithms. I will make a separate repository for the reverse algorithm in a near future.
All code, except for those belong to third-party libraries (code in benchmark/ryu
and benchmark_results/shaded_plots
), is licensed under either of
- Apache License Version 2.0 with LLVM Exceptions (LICENSE-Apache2-LLVM or https://llvm.org/foundation/relicensing/LICENSE.txt) or
- Boost Software License Version 1.0 (LICENSE-Boost or https://www.boost.org/LICENSE_1_0.txt).
except for the file fp_to_chars.cpp
, which is licensed under either of
- Apache License Version 2.0 (LICENSE-Apache2 or http://www.apache.org/licenses/LICENSE-2.0) or
- Boost Software License Version 1.0 (LICENSE-Boost or https://www.boost.org/LICENSE_1_0.txt).