[boost] Re: Numeric Conversion Formal Review

From: Dan W. (danw_at_[hidden])
Date: 2003-12-12 11:08:41

 > No need to go as far as using assembly language. There are C functions
 > that work just fine like nearbyint or rint (the second one is usually
 > faster). And if you are really looking for assembler, the frndint
 > instruction is available on x86 for exemple.

Not familiar with those, but I just came up with something that should
work...

Instead of...

------------------------------------------------------------------- 
template<class S>
  struct Round
  {
    static source_type nearbyint ( argument_type s )
    {
      using std::floor ;
         return floor( s + static_cast<S>(0.5) ) ;
    }
    typedef mpl::integral_c<std::float_round_style,std::round_to_nearest> 
round_style ;
  };
-------------------------------------------------------------------

I'd try the following (portability issues notwithstanding)

-------------------------------------------------------------------
  struct Round
  {
    static source_type nearbyint ( argument_type s )
    {
      using std::floor ;
         return depressed_avg( s, static_cast<S>(1.0) ) ;
    }
    typedef mpl::integral_c<std::float_round_style,std::round_to_nearest> 
round_style ;
  };

  template <>
  inline float depressed_avg<float>( float f, float onef )
  {
    float ret;       //optimized away by inlining
    __asm
    {
      mov   eax,f     ; optimized away by compiler
      mov   ebx,onef  ; optimized away by compiler
      dec   ebx       ; THE KEY INSTRUCTION dec mantissa by an LSB
      add   eax,eax   ; double f first
      add   eax,ebx   ; add the "slightly depressed one"
      shrc  eax       ; shift right w/Carry, whatever the spelling was
      mov   ret,eax   ; optimized away by compiler
    }
    return ret        //optimized away by inlining
  }
-------------------------------------------------------------------
(sorry if my syntax offends, I'm rusty in assembler these days)

That should add up to just four integer instructions, with optims' ON.
Not totally sure that the sign bit would always survive; would need to
test it.

Cheers!