Benchmark against Karma needed about fmt HOT 2 CLOSED

Thanks for your comments.

The results for the int-generator benchmark, which BTW is taken from Karma, that you give are similar to the ones I reported in http://zverovich.net/2013/09/07/integer-to-string-conversion-in-cplusplus.html .

Boost Karma is indeed pretty fast, but unfortunately it doesn't support replacement-based formatting. So it is not surprising that it is faster than fmt::Format which in addition to formatting also performs parsing of a format string and handling of positional arguments. Therefore it make more sense comparing Karma to fmt::Writer rather than to fmt::Format.

As you can see fmt::Writer is slightly faster than karma::generate on the original benchmark although it does automatic buffer management while Karma uses a fixed buffer. I'll have a look why allocating additional std::string causes larger performance impact with fmt::Writer than with Karma, but the main point is that with fmt::Writer you usually don't need this allocation, because the buffer management is automatic.

fmt::FormatInt works with int64_t now.

As for the modified benchmark, as far as I understand it compares the performance of formatting of single-digit and double-digit numbers. Since formatting of such small numbers is very fast, the buffer management that fmt::Writer does becomes a significant factor. Karma on the other hand offloads the buffer management to the user. Anyway, even for small inputs fmt::FormatInt is somewhat faster than karma::generate:

% 10 results:
karma::generate: 0.0991358 [s]
fmt::FormatInt: 0.0979277 [s]

% 100 results:
karma::generate: 0.0997507 [s]
fmt::FormatInt: 0.0943652 [s]

from fmt.

BTW where is this library page that you've mentioned?

from fmt.