[vtk-developers] Possible changes to vtkMath...

[Michael Halle]

I don't like the idea of a fabs()-based macro because
of the double-evaluation side effects.  Conditionally
defining it only makes it worse. The ternary operator
may or may not be highly optimized on some platforms.

Inlining is the defined mechanism for this purpose, even
if it may be slower on some platforms, I think it's the
best choice.
Michael Halle
mhalle at bwh.harvard.edu

Chris Volpe said:
> 1) No need to use "fabsf", since I believe that's only there for
> C-language compatibility, in which "fabs" can't be overloaded.
>
> 2) In an optimized build, fabs is an intrinsic on Intel Architecture
> cpus, and will probably generate a single instruction that will run
> faster than
> the macro. Is there a uniform way to tell at compile time whether we're
> compiling in non-optimized mode? E.g.:
> #ifdef DEBUG
> #define VTK_FABS(x) (((x) < 0.0 ) ? (-(x)) : (x) )
> #else
> #define VTK_FABS(x) ::fabs(x)
> #endif
>
> 3) The conditional heap allocation looks like a good idea, but if you
> want maximum performance, can you do some conditional compilation on
> Win32 to call _alloca()?
>
> -Chris
>
>> -----Original Message-----
>> From: Will Schroeder [mailto:will.schroeder at kitware.com]
>> Sent: Thursday, July 10, 2003 6:42 PM
>> To: vtk-developers at vtk.org
>> Subject: [vtk-developers] Possible changes to vtkMath...
>>
>> Hi Folks-
>>
>> I was profiling VTK and discovered a couple of issues with vtkMath.
> The
>> fabs() function (used in LUSolve, etc) is taking an surprising amount
> of
>> time, and I find some leftover static variables. Here's what I'm
>> thinking---I'd like your comments.
>>
>> fabs()
>> We can define a macro #define  VTK_FABS(x)  (x) < 0.0 )?(-(x)):(x))
>> Another alternative is to do what vnl does:
>> inline float       VTK_FABS(float  x) { return ::fabsf(x); }
>> inline double      VTK_FABS(double x) { return ::fabs(x); }
>>
>> The macro seems to be a bit faster but I'm running debug not release.
>>
>> Static variables will be replaced with code like this that is
> partially
>> optimized for small matrices:
>>  >   int *index, scratch[10];
>>  >   index = ( size < 10 ? scratch : new int[size] );
>>  >   if (size >= 10 ) delete [] index;
>>
>> Comments?
>>
>> Will