Using float rather than double

[hattesen]

I will be using tinyexpr-plusplus in an embedded solution on a ARM Cortex M0+ (RP2040) Microcontroller without hardware FPU, and I imagine using float rather than double internally in tinyexpr-plusplus would provide quite a bit of a performance improvement.

What would be the best way to make tinyexpr-plusplus use float rather than double as the internal (and external) data type?

As I see it, there are two workable paths for such a change

1. using a typedef

   typedef float te_type;
   ...
   te_type evaluate();

2. Using a class template argument in the te classes

   template <class T>
   class te_parser {
   ...
   T evaluate();
   ...
   void add_variable_or_function(te_variable<T> var)

While using a template argument would be the most "pure" C++ solution, using typedef would likely be the least invasive – and most backward compatible – approach.

If I go ahead and ...

1. fork the repository
2. go through the replace double with te_type/T excercise
3. determine if it provides a non-trivial performance improvement

... would there be any chance to have this change accepted as a pull request back into tinyexpr-plusplus so I dont have to keep tracking upstream changes and make similar type-changes to any new code?

[Blake-Madden]

Templates offer flexibility, but at the cost of making APIs (and compiler warnings) so much more difficult to deal with. I definitely fell out of love with them some time ago.

I really like the idea of typedefing via conditional compilation:

#ifdef TE_FLOAT
    using te_number_type = float;
#else
    using te_number_type = double;
#endif

I can work on this, because I'll need to add this to the unit testing also (hopefully all the standard lib function overloads will handle floats OK). I should have something by the weekend...

[Blake-Madden]

This is available in the latest build now. Define TE_FLOAT and the library will use float instead of double throughout. As a shortcut, te_type is available as a typedef for whichever type is defined.

Let me know if this is a good solution for your environment and I'll mention that as a recommendation for embedded systems in the manual.

[hattesen]

Thank you so much. I will run some benchmarks and memory-footprint comparisons, and report back.

[Blake-Madden]

FYI, I just added a new compile flag, TE_NO_BOOKKEEPING. Defining this should offer ~15% bump in performance. It turns off the is_function_used() feature that most people probably don't need.

[Blake-Madden]

Was this able to help you? If so, I guess you could close the discussion.