# Does string formatting boost karma?

If you’ve read any of my recent posts about string formatting (A better string formatting library for C++, Making string formatting fast or I can’t believe it’s not Python) I apologize for bringing this topic up again.

Today my attention was drawn by Boost Karma, which is a less-known Boost library for output generation. The authors of this library claim that it is much faster than printf, std::stream and boost::format, so I decided to have a look at it.

Karma works with generators which can be created in a number of ways. For example, int_ creates a built-in generator that outputs an integer. Generators can be combined using operator << like int_ << int_ which outputs two integers, or *int_ which outputs zero or more integers. This is somewhat similar to construction of regular expressions or grammar productions which is not surprising considering that Karma is a part of the Spirit Parser Framework.

Here is a small example:

Character literals like ‘,’ are copied unchanged to the output and (int_ % ',') is equivalent to int_ << *(',' << int_).

Compiling it with g++ -std=c++11 <filename> and running gives the following output:

1,2,3


What strikes me in the example above is that there is no format string there. As with IOStreams literal text is mixed with arguments, so i18n is problematic. So you definitely don’t want to use Karma for messages shown to the user.

Also a little mistake like a misused operator in a generator expression can lead to a cascade of lengthy and incomprehensible error messages. The good thing though is that format specifications are checked at compile time.

Karma provides some control over formatting, for example, you can specify alignment and fill. Specifying precision is also possible although a bit cumbersome:

The above example prints

[     12.35]


Unfortunately I haven’t found an equivalent to printf’s 0 format specifier.

One of the claimed advantages of Karma is performance so I decided to do a small comparison of integer to string conversion with my formatting library. I took the int_generator test and modified it to include the C++ Format library. I’ve exposed an integer formatting method through Formatter::operator<<(int) and used it to avoid unnecessary parsing of a format string and make the comparison to Karma more fair. I plan to add similar methods for other types as well.

So here are the results:

As you can see Karma is indeed the fastest among the tested methods. However, the difference between C++ Format and Karma is only about 7% which is not too bad considering that C++ Format does runtime handling of format specifications and automatically allocates enough space for the output while Karma uses a predefined format and a fixed size buffer. What is surprising to me is that both sprintf and iostreams are much slower. I would love to know why, so if you have any ideas let me know in the comment section below.

The performance test is called int_generator and it is available in the format repository at GitHub so you can verify the results.

The conclusion is that Karma is most useful if you need fast output in some predefined format known at compile time, but not for any user messages.

Update: Karma is no longer the fastest method, see the follow up post for new results.