patterns

right now you can just export your code and run it inside this boilerplate Bela project provided by Cycling 74: https://github.com/Cycling74/genBela.
Unfortunately, from my experience you do not get a great performance improvement with gen over libpd, and it is way slower than Heavy. This is most likely because the code generated by gen works on single-samples buffers, therefore making it harder for the compiler to vectorize it. This was like 6 months ago, perhaps things have improved now.

Some of my notes on how to make it slightly faster:

• numerical constants exported from gen do not have the "f" suffix and as such are treated as double, while they should be of the same type of sample_t. Maybe the most portable way would be to always prepend the constants with (sample_t) in the gen-generated code, but right now you have to either do it manually (sed is your friend), or use gcc instead of clang and provide the -fsingle-precision-constant flag (not supported by clang).

• on Bela (and other platforms, such as the Owl) denormals are flushed to zero in hardware, so there is no point in checking for them.
  Therefore we need a way to disable fixdenorm altogether. I can save 5% CPU by replacing

  inline t_sample fixdenorm(t_sample v) { return GENLIB_FIX_DENORM(v); }

  with

  #define fixdenorm(v) (v)

I also had some issues with gen - generated code using stdlib math functions instead of the optimized ones that it tried to provide, so check the output of the pre-processor and see what it is actually calling.