[Clp] patch for Visual C++ floating-point model problem

[John J Forrest]

Arno,

Thanks for your patch.  I have put into trunk and Clp stable 1.7 (for Cbc
stable 2.1).

John


                                                                                                                         
  From:       Arno Schödl <aschoedl at think-cell.com>                                                                      
                                                                                                                         
  To:         <clp at list.coin-or.org>                                                                                     
                                                                                                                         
  Date:       08/09/2008 05:40 PM                                                                                        
                                                                                                                         
  Subject:    [Clp] patch for Visual C++ floating-point model problem                                                    
                                                                                                                         





Hello,

I am compiling Clp with Visual C++ 2008. In Release build, I got an
infinite recursion in the function below because the condition
tolerance>saveTolerance always evaluated to true, although tolerance was in
fact never touched after copying it into saveTolerance. The problem was
that saveTolerance was stored in memory, and thus was truncated to 64 bit.
tolerance was kept in an FPU register with all 80 bits. Thus tolerance was
slightly larger than saveTolerance in the comparison. I verified this from
the generated assembly code.

/fp:precise did not solve the problem (and did not either in Visual C++
2005 SP1), although from the documentation, it should, because assignments
should force rounding. In fact, the generated code still kept tolerance in
an FPU register. /fp:strict fixed it but is not recommended for
performance. I propose the patch below to side-step the problem. There is
one caveat where I would need Lou's input: if model_->largestDualError() is
only slightly larger than model_->largestPrimalError(), it is conceivable
that after the multiplication, still tolerance==saveTolerance, and we still
recurse. Does that break the code?

Thank you for help,

Arno


--- .\Clp\src\ClpDualRowSteepest.cpp      Sat Aug  9 23:05:22 2008
+++ .\Clp\src\ClpDualRowSteepest.cpp      Sat Aug  9 22:36:45 2008
@@ -168,7 +168,6 @@
   // But cap
   tolerance = CoinMin(1000.0,tolerance);
   tolerance *= tolerance; // as we are using squares
-  double saveTolerance = tolerance;
   double * solution = model_->solutionRegion();
   double * lower = model_->lowerRegion();
   double * upper = model_->upperRegion();
@@ -223,10 +222,13 @@
   if (model_->numberIterations()<0)
     printf("aac_p it %d\n",model_->numberIterations());
 #endif
+  bool bToleranceIncreased=false;
   if(model_->numberIterations()<model_->lastBadIteration()+200) {
     // we can't really trust infeasibilities if there is dual error
-    if (model_->largestDualError()>model_->largestPrimalError())
+     if (model_->largestDualError()>model_->largestPrimalError()) {
       tolerance *= CoinMin(model_->largestDualError()/model_->
largestPrimalError(),1000.0);
+      bToleranceIncreased=true;
+     }
   }
   int numberWanted;
   if (mode_<2 ) {
@@ -318,7 +320,7 @@;
       break;
   }
   //printf("smallest %g largest %g\n",smallestWeight,largestWeight);
-  if (chosenRow<0&& tolerance>saveTolerance) {
+  if (chosenRow<0&& bToleranceIncreased) {
     // won't line up with checkPrimalSolution - do again
     double saveError = model_->largestDualError();
     model_->setLargestDualError(0.0);



                                                                            
 --                                                                         
                                                                            
 Dr. Arno Schoedl            aschoedl at think-cell.com                        
                                                                            
 Technical Director                                                         
                                                                            
 think-cell Software GmbH    http://www.think-cell.com                      
                                                                            
                                                                            
 Invalidenstr. 34            phone                     +49-30-666473-10     
                                                                            
 10115 Berlin, Germany       fax                       +49-30-666473-19     
                                                                            
 Geschäftsführer: Dr. Markus                                                
 Hannebauer, Dr. Arno                                                       
 Schoedl                                                                    
                                                                            
 Amtsgericht Charlottenburg,                                                
 HRB 85229                                                                  
                                                                            


_______________________________________________
Clp mailing list
Clp at list.coin-or.org
http://list.coin-or.org/mailman/listinfo/clp