Missing functions in FloatingPoint API?

[RyanGlScott]

I've been playing around with the FloatingPoint package for a bit, and while it mostly does what I want, I was surprised not to find versions of the standard trigonometric functions (sin, cos, tan, etc.) for FloatingPoint. Is their omission intentional? I noticed that they are defined over Reals, but Real doesn't suffice for my needs since it is not an instance of Bits.

[quark17]

I don't think there's a reason. The users of the FloatingPoint package have often been people writing processors, so the functions that exist are the ones typically needed for that, and sin/cos/etc haven't been ops in those ISAs. Feel free to contribute implementations for other functions. I'm not experienced in hardware for floating point, but I'd guess that there are many possible implementations, so one library isn't going to fit everyone's needs, and I suspect that most users will want to write their own or import IP (perhaps using native constructs in the target hardware) or so on. But there's no reason why BSC couldn't have a simple implementation available in its libraries, for incremental design or for people who don't have tight resource requirements. BSC also isn't going to be able to promise that the implementation that it does provide is optimal in any way. But with those caveats, it's not unreasonable to add implementations, if someone has written them and wants to submit them. There is also a repo bsc-contrib where libraries can be made available. There's perhaps a debate to be had about how much should be in the standard libraries vs what eventually should be considered extra, that can people can download and use if they need it.

[RyanGlScott]

Ah, thanks for the background context.

I don't have my own implementations to offer, but I'll keep bsc-contrib in mind for the future.

[srd8668]

Hi i want to use the floating point package to develop an accelerator, where can i find the documentation for it to know how to use it.

[quark17]

@srd8668 , I have responded to that question on Discussion #709 .

[RyanGlScott]

Some other functions from the Real package that would be nice to have FloatingPoint counterparts for:

• Proper implementations of (**), exp_e, log, and logb. These are part of FloatingPoint's Arith instance, but they simply call error when used. Unfortunately, I don't really know how to implement these myself.

• ceil and floor functions. Interestingly, there is an unexported round function in FloatingPoint.bsv that could be used to implement ceil and floor. Something like this would suffice:

  ceil :: FloatingPoint e m -> FloatingPoint e m
  ceil x = round Rnd_Plus_Inf x ???
  
  floor :: FloatingPoint e m -> FloatingPoint e m
  floor x = round Rnd_Minus_Inf x ???

  I'm not really sure what would go in place of the ???s, however, as I don't understand what purpose the guard argument to round is serving.

[RyanGlScott]

One possible starting point for a native Bluespec implementation is to adapt the code in the LibBF library, which provides an implementation of arbitrary-precision IEEE754 floating-point numbers in a style very similar to what Bluespec's FloatingPoint package is using. Granted, LibBF is a software implementation, not a hardware one, but it should still be useful as a starting point.