[Insight-developers] vnl_math::pi coming up undefined with vector optimization turned on?

Fri Oct 23 10:25:06 EDT 2009

Kent,

On CMAKE_BUILD_TYPE=RELEASE the "-O3" flag is the optimization used.  
And there are many nightly build with this flag. This is very  
perplexing.

What version of the OSX, XCode, gcc etc are you using?

As I am currently building with this flag, I don't know how to try to  
reproduce this error. Do you have a command line cmake with "-D" flags  
defining the options you are setting, that is able reproduce this  
problem?

Brad

On Oct 23, 2009, at 10:10 AM, kent williams wrote:

> Yes, I did a scratch rebuild, and I verified by going back to O2  
> that this
> was the issue.
>
> What it looks like is that vcl_math.h isn't declaring all those static
> member variables as it should -- very much a VXL library issue, at  
> least on
> OSX.
>
> It's not a huge deal to me, except we have some very CPU-intensive  
> code that
> would benefit from vectorizing in the registration framework.  At  
> least
> that's what Shark is telling me.
>
>
> On 10/22/09 8:31 PM, "Luis Ibanez" <luis.ibanez at kitware.com> wrote:
>
>> Hi Kent,
>>
>> <me speculating>
>>
>> One of the things that -O3 does is to force full in-lining of  
>> templated
>> method.
>>
>> I'm wondering if the effect is that something that has been typically
>> delegated
>> to a cxx initialization is now being done in a header.
>>
>> That said...
>> the default of "Release" build in Linux is to use -O3....
>> so...the problem that you are exposing seems to
>> be very specific to OS X.
>>
>> </me speculating>
>>
>>
>> Question:
>>
>> When you changed the flags...
>> Did you rebuild the toolkit from scratch ?
>>
>>
>>     Luis
>>
>>
>> ------------------------
>> On Thu, Oct 22, 2009 at 4:36 PM, kent williams
>> <norman-k-williams at uiowa.edu> wrote:
>>> I’m trying to build ITK from the CVS source as of today.  I’m not  
>>> modifying
>>> anything at all.  The only thing I did was use our CFLAGS and  
>>> CXXFLAGS when
>>> I run CMake.  The undefined externals are coming during building  
>>> ITK itself,
>>> not in my application.
>>>
>>> The problem didn’t  show up until I added –O3 to the command  
>>> line.  Why that
>>> would interact with the static const members of  vnl_math I have  
>>> no idea.
>>>
>>>
>>> On 10/22/09 3:21 PM, "Bradley Lowekamp" <blowekamp at mail.nih.gov>  
>>> wrote:
>>>
>>> Hello kent,
>>>
>>> I was just looking at this stuff. It turns out the following is  
>>> not standard
>>> c++:
>>>
>>> class C {
>>> static const double one = 1.0;
>>> };
>>>
>>> This is because static member variable must be defined in one obj  
>>> file, not
>>> in headers, so that all instance of the object will reference the  
>>> same
>>> address. But some compilers are smart enough to know what you are  
>>> trying to
>>> do here and let is slide.
>>>
>>> As to why you are getting this error, are you sure you are linking  
>>> against
>>> the correct libraries? vnl_math::pi should be defined int the vnl  
>>> library.
>>>
>>> victoria:bin blowek1$ nm libitkvnl.a | grep pi
>>> 000016d8 S __ZN8vnl_math11one_over_piE
>>> 000016e0 S __ZN8vnl_math11two_over_piE
>>> 000016e8 S __ZN8vnl_math15two_over_sqrtpiE
>>> 000016f0 S __ZN8vnl_math16one_over_sqrt2piE
>>> 000016c0 S __ZN8vnl_math2piE
>>> 000016c8 S __ZN8vnl_math9pi_over_2E
>>> 000016d0 S __ZN8vnl_math9pi_over_4E
>>>
>>> Hope this helps,
>>> Brad
>>>
>>>
>>>
>>> On Oct 22, 2009, at 4:05 PM, kent williams wrote:
>>>
>>> I'm trying to build ITK on OS X with the "-O3" flag. I've never  
>>> had trouble
>>> with "-O2" but now I get an undefined external for vnl_math::pi.
>>>
>>> The main reason to go with "-O3" is to turn on vector SIMD  
>>> optimizations --
>>> in one application we're trying to optimize, the biggest share of  
>>> CPU is
>>> used in the ImageToImageMetric, and Shark is suggesting that the  
>>> loops in
>>> the hot routines would benefit from vector optimizations.
>>>
>>> I don't know why or from whence the problem originates, but what  
>>> it looks
>>> like is that the vnl_math:: constants don't get declared as  
>>> variables if
>>> VCL_STATIC_CONST_INIT_FLOAT_NO_DEFN is defined.
>>>
>>> Now I'm no VXL library hacker,so I'm pretty much baffled by this  
>>> thing.
>>>
>>> _______________________________________________
>>> Powered by www.kitware.com <http://www.kitware.com>
>>>
>>> Visit other Kitware open-source projects at
>>> http://www.kitware.com/opensource/opensource.html
>>>
>>> Please keep messages on-topic and check the ITK FAQ at:
>>> http://www.itk.org/Wiki/ITK_FAQ
>>>
>>> Follow this link to subscribe/unsubscribe:
>>> http://www.itk.org/mailman/listinfo/insight-developers
>>>
>>> ========================================================
>>>
>>> Bradley Lowekamp
>>>
>>> Lockheed Martin Contractor for
>>>
>>> Office of High Performance Computing and Communications
>>>
>>> National Library of Medicine
>>>
>>> blowekamp at mail.nih.gov
>>>
>>>
>>>
>>>
>>> _______________________________________________
>>> Powered by www.kitware.com
>>>
>>> Visit other Kitware open-source projects at
>>> http://www.kitware.com/opensource/opensource.html
>>>
>>> Please keep messages on-topic and check the ITK FAQ at:
>>> http://www.itk.org/Wiki/ITK_FAQ
>>>
>>> Follow this link to subscribe/unsubscribe:
>>> http://www.itk.org/mailman/listinfo/insight-developers
>>>
>>>
>
> _______________________________________________
> Powered by www.kitware.com
>
> Visit other Kitware open-source projects at http://www.kitware.com/opensource/opensource.html
>
> Please keep messages on-topic and check the ITK FAQ at: http://www.itk.org/Wiki/ITK_FAQ
>
> Follow this link to subscribe/unsubscribe:
> http://www.itk.org/mailman/listinfo/insight-developers

========================================================
Bradley Lowekamp
Lockheed Martin Contractor for
Office of High Performance Computing and Communications
National Library of Medicine
blowekamp at mail.nih.gov

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.itk.org/mailman/private/insight-developers/attachments/20091023/add06d70/attachment-0001.htm>