Integer functions for non-FPU MCU's

[Vasili-Sk]

Problem is that small MCU dont have FPU module, like STM32WLExxxx. With few floating point operations and string converstions it adds more than 15kB of floating point arithmetic. CPU is already small flash area and now it have more overhead to handle FPU in integer way.

Describe the solution you'd like
Add begin() overload with integer bandwith and other parameters of uint32_t where float used.
Measure volts as mV uint32_t.

Describe alternatives you've considered
Please propose if you have any.

Additional context

[jgromes]

This is rather tricky, because it is not only about the interface - floating points are also used internally. So we would have to rewrite the entire library to use only integer arithmetic. Additionally, it is hard to estimate whether this would be worth it, because a couple floating point operations in the user code could easily undo it.

It would be good to have an idea what would be the performance gain if we were to go ahead with this. I will transfer this to a discussion for the time being.

[Vasili-Sk]

FYI, depended projects store parameters same way as it is defined in library, so it is starts from Radiolib that uses float as input.
From quick glance TXCO voltage can be easily converted, same as bandwith since it uses enums in hal.

[moshe-braner]

I use the ESP32 MCU and it has some FP hardware but:

• FP division is in software and slow, and
• anything in double-precision FP is very slow
  CPU load is important for my application since I ask the poor little ESP32 to do a LOT more than just interface to the radio module. So what I would like to see is some small optimizations such as:
• Avoid "double" when "float" will do.
• Avoid division when multiplication can be substituted.
  The latter may happen sometimes thanks to compiler optimizations, not sure. And sometimes FP can be avoided altogether, e.g., in SX127x::transmit():
  // calculate timeout in ms (150 % of expected time-on-air)
  timeout = (toa * 1.5) / 1000;
  can be replaced with:
  timeout = (toa * 3) / 2000;
  or at least make it "1.5f"!

[jgromes]

Well ... feel free to implement it then ;) that's what open source is for after all - and unfortunately this has not been high on my list of priorities. For sure this could lead to some improvements, but I expect it will be mostly in the compiled binary size and not that much in execution speed since waiting for an SPI transactions to finish will easily overshadow a few thousand instructions/clock cycles to complete a floating point operation. I will be gladly be proven wrong in that regard though, so you can take that as a little bonus challenge.

The only thing I ask is that you test each change when switching from floating point to integers or modifying the calculations. The example you have there is indeed quite simple, but there are others - see e.g. SX126x::getFrequencyError which maybe could also be converted to integers.