updating fmt to version 12 and use compile-time format strings

[lemire]

Fixes #50

This follows comments by a few people, including @vitaut (the fmt author).

These changes drastically change change the performance of fmt, for the better.

On my Apple M4.... before...

./build/benchmarks/benchmark -f data/canada.txt
fmt_format                    :   586.68 MB/s (+/- 6.6 %)     1.87 MB     28.64 ns/f    34.92 Mfloat/s

after...

./build/benchmarks/benchmark -f data/canada.txt 
fmt_format                    :   906.36 MB/s (+/- 4.7 %)     1.87 MB     18.54 ns/f    53.95 Mfloat/s

In context, this makes (in one test), the fmt library much closer to std::to_chars.

dragon4                       :   111.95 MB/s (+/- 3.1 %)     1.87 MB    150.06 ns/f     6.66 Mfloat/s
netlib                        :    20.57 MB/s (+/- 0.9 %)     1.87 MB    816.58 ns/f     1.22 Mfloat/s
fmt_format                    :   582.34 MB/s (+/- 5.6 %)     1.87 MB     28.85 ns/f    34.66 Mfloat/s
grisu3                        :   701.22 MB/s (+/- 9.1 %)     1.87 MB     23.96 ns/f    41.74 Mfloat/s
grisu_exact                   :  1235.86 MB/s (+/- 8.8 %)     2.09 MB     15.21 ns/f    65.74 Mfloat/s
schubfach                     :  1389.48 MB/s (+/- 10.1 %)     1.87 MB     12.09 ns/f    82.71 Mfloat/s
dragonboxlm                   :  2319.66 MB/s (+/- 11.6 %)     2.20 MB      8.54 ns/f   117.15 Mfloat/s
dragonbox                     :  1985.19 MB/s (+/- 10.3 %)     2.09 MB      9.47 ns/f   105.60 Mfloat/s
ryu                           :  1551.55 MB/s (+/- 8.0 %)     2.09 MB     12.12 ns/f    82.53 Mfloat/s
teju_jagua                    :  1569.40 MB/s (+/- 9.1 %)     2.20 MB     12.62 ns/f    79.26 Mfloat/s
swiftDtoa                     :  1059.53 MB/s (+/- 10.0 %)     1.87 MB     15.86 ns/f    63.07 Mfloat/s
yy_double                     :  2038.28 MB/s (+/- 10.8 %)     1.87 MB      8.24 ns/f   121.32 Mfloat/s
std::to_chars                 :  1078.68 MB/s (+/- 9.9 %)     1.87 MB     15.57 ns/f    64.21 Mfloat/s

Context: Daniel Lemire, "Converting floats to strings quickly," in Daniel Lemire's blog, February 1, 2026, https://lemire.me/blog/2026/02/01/converting-floats-to-strings-quickly/.

[vitaut]

Thanks for updating {fmt} usage!

[lemire]

@vitaut Your comments on our work are much appreciated.

[jaja360]

I observe ~30% speedup with the new version on my CPU.
PR looks reasonable. We can merge !

[vitaut]

Note that {fmt} uses Dragonbox so it's interesting to see a big gap between the two, even after the update. Some gap is expected because {fmt} is a higher-level facility but it's normally much smaller, e.g. on Milo Yip's dtoa-benchmark the gap is ~8%. I would investigate further but the FP implementation in {fmt} is about to be revamped (Dragonbox replaced with https://github.com/vitaut/zmij) so it's probably doesn't matter except that I would recommend adding version information since there is substantial difference between versions and even different algorithms used.

[jaja360]

@vitaut I added fmt version number in the last commit. Thanks for the comment. I also added zmij, see #52. Your algorithm is the fastest on our benchmark. Congrats 🎉

[vitaut]

Thank you, @jaja360!