[vtk-developers] Bug report: vtkFloorFuncMacro

[Ovidiu Toader]

Hi,
The bug report URL,  http://public.kitware.com/Bug/, doesn't seem to offer the 
ability to report a bug so I have no choice but to broadcast this message on 
the list. I apologize in advance to anybody that finds this inappropriate but 
I believe that this bug is too serious  to be ignored.

In file Rendering/vtkVolumeRayCastMapper.h there is a function named 
vtkFloorFuncMacro. This function  does not produce correct results for all 
inputs on a platform that does not use the "quick-and-dirty" solution:

inline int vtkFloorFuncMacro(double x)
{
#if defined i386 || defined _M_IX86
  double tempval;
  // use 52-bit precision of IEEE double to round (x - 0.25) to
  // the nearest multiple of 0.5, according to prevailing rounding
  // mode which is IEEE round-to-nearest,even
  tempval = (x - 0.25) + 3377699720527872.0; // (2**51)*1.5
  // extract mantissa, use shift to divide by 2 and hence get rid
  // of the bit that gets messed up because the FPU uses
  // round-to-nearest,even mode instead of round-to-nearest,+infinity
  return ((int*)&tempval)[0] >> 1;
#else
  // quick-and-dirty, assumes x >= 0
  return (int)(x);
#endif
}

As an example take vtkFloorFuncMacro(0.99989) which should return 0. Instead 
it returns 1, which is wrong. Because of failures like this, volume rendering 
that uses vtkVolumeRayCastMapper can produce random crashes. After I replaced 
the body of this function with the standard "floor(x)" all the random crashes 
I experienced before are gone.  
-- 
Ovidiu