[Insight-developers] vnl updates : Rounding Test : Performance and Correctness

[Tom Vercauteren]

Hi all,

This looks like a good option. Something like that could also be done
directly into vnl. Below is a snippet that shows what it would look
like considering only the sse2 and if implementation.

Tom

#if VNL_CONFIG_ENABLE_SSE2_ROUNDING
#  if VNL_CONFIG_HALFINT_FAST
inline int vnl_math_rnd(float  x){return _mm_cvtss_si32(_mm_set_ss(x));}
#  elif defined  VNL_CONFIG_HALFINT_UP
inline int vnl_math_rnd(float  x){return
(_mm_cvtss_si32(_mm_set_ss(2*x+0.5f))>>1);}
#  else
#    error "Unknown rounding type"
#  endif
#else
#  if VNL_CONFIG_HALFINT_FAST
inline int vnl_math_rnd(float  x) {return x>=0.f?static_cast<int>(x +
.5f):static_cast<int>(x - .5f); }
#  elif defined  VNL_CONFIG_HALFINT_UP
inline int vnl_math_rnd(float  x) {x+=0.5f; return
static_cast<int>(x>=0.f?x:(x==static_cast<int>(x)?x:x-1.f));}
#  else
#    error "Unknown rounding type"
#  endif
#endif

inline int vnl_math_rnd(float  x)
{
   return _mm_cvtss_si32(_mm_set_ss(x));
}

inline int vnl_math_ceil(float  x)
{
   return -(_mm_cvtss_si32(_mm_set_ss(-.5f-2*x))>>1);
}

inline int vnl_math_floor(float  x)
{
   return _mm_cvtss_si32(_mm_set_ss(2*x-.5f))>>1;
}


On Wed, Dec 31, 2008 at 6:15 PM, Luis Ibanez <luis.ibanez at kitware.com> wrote:
> Amitha, Hans,
>
> I actually agree with all your comments.
>
> It seems that we all agree that a set of "Round" definitions is what we
> need,
> even though we anticipate that the most used will be the RoundItQuickly().
>
> The challenge now is to identify a way of implementing this in a manner that
> make
> sense for both ITK developers, developers of ITK-based applications, and end
> users.
>
> Trying to look at it from Tom's perspective, the issue that I see is the the
> decision of
> speed vs multi-platform consistency will appear at the end-user level and
> not at the
> moment when an ITK developer is deciding which on of the available
> implementations
> of "round" to use.
>
> In other words, the decision of "I just need this program to run as fast as
> possible
> and I don't care if the result is different in another platform" will most
> likely be
> made by an end-user (application user) than by an application developer, and
> much less by an ITK developer.
>
> This leads to suggest that the selection of the particular Round
> implementation
> should probably be managed by a CMake option that would be turned on at the
> moment of configuring/building ITK and the said application.  By default,
> probably we would use the settings for rounding in a consistent manner
> across
> platforms.
>
> My suggestion would be to add a file
>
>                                         itkMath.h
>
> to Insight/Code/Common
>
>  and to put in there the implementations of Round, Ceil, Floor.
>
> Then add a CMake flag option to the top CMakeLists.txt file, to select
> among the different implementations of round.
>
> It could be:
>
> CMake variable name:    ITK_ROUND_OPTION
>
> CMake variable acceptable values:
>
>                    MultiPlatform
>                    Fast
>                    Unbiased
>
> The "Fast" option could delegate the rounding to the VNL implementation
> (if that happens to be the most efficient one).
>
> An alternative proposal would be to define Round as a Macro, and to
> put it in itkMacro.h
>
>
>
>        Luis
>
>
> ------------------------------------------------------------------------------
> On Wed, Dec 31, 2008 at 11:32 AM, Hans Johnson <hans-johnson at uiowa.edu>
> wrote:
>>
>> Amitha:  Thanks.  Indeed the JustRoundItQuickly is likely the most
>> intended
>> use of round.   Perhaps what is needed are just two round functions
>> (JustRoundItQuickly, RoundConsistentlyAcrossPlatforms).
>>
>> Hans
>>
>>
>>
>>
>> On 12/31/08 10:23 AM, "Amitha Perera" <amitha.perera at kitware.com> wrote:
>>
>> > JustRoundItQuickly
>>
>
>