[Insight-developers] vnl_lsqr - some benchmarks

[James Gee]

running a *complete* multi-resolution (5 levels) non-rigid registration,
at which the number of unknowns at the finest scale is 8192, usually
takes less than a minute (with ITPACK) on an ultra 60.  note that
several iterations are required at each scale, so the solver is invoked
multiple times.  also, we're not just solving a matrix here but
performing all the computations reguired to carry out a registration.
the vnl numbers are a bit worrisome.



> -----Original Message-----
> From: insight-developers-admin@public.kitware.com 
> [mailto:insight-developers-admin@public.kitware.com] On 
> Behalf Of Miller, James V (CRD)
> Sent: Wednesday, April 03, 2002 9:11 AM
> To: Insight-developers (E-mail)
> Subject: [Insight-developers] vnl_lsqr - some benchmarks
> 
> 
> Here is part of a discussion of using vnl_lsqr to solve
> a sparse matrix problem. Dominique says it compares 
> favorably to the solvers in Matlab.
> 
> Have the FEM folks tried this routine? How did this compare
> to the other solvers you have used?
> 
> This user was solving systems with 3000-10000 cells (nodes?). 
>  How does that compare to the meshes that you use in 
> deformable registration?
> 
> 
> -----Original Message-----
> From: domi@vision.ee.ethz.ch [mailto:domi@vision.ee.ethz.ch]
> Sent: Thursday, March 28, 2002 10:47 AM
> To: Vxl-users@lists.sourceforge.net
> Subject: [Vxl-users] vnl_lsqr - some benchmarks
> 
> 
> Hi,
> 
> You were right, Matlab gets dramatically slower with 
> nonsymmetric non-only-diagonal matrix.
> 
> 
>   for (signed i=0; i<n; i++) {
> 	 A(i,i) = 1.0;
> 
> 	 if( (i+1)<=(n-1) && (i-1)>=0 ){
> 		A(i,i+1) = 2.0;
> 		A(i,i-1) = -1.0;
> 	 }
> 
> 	 b(i)           = i+1.0;
>   }
> 
> 
> Actually matlab gets far slower compared to
> vnl_lsqr but on the test system (above) I have:
> 
> vnl_lsqr.cxx : The equations A*x = b are probably compatible. 
>  Norm(A*x
> - b) is as small as seems reasonable on this machine.
> 
> vnl solves this (n=3000) very quickly while matlab chews it 
> much longer.
> 
> n=10,000 is solved by vnl in 1.5 min on my machine, matlab 
> after a longer while crashes screaming for memory.
> 
> 
> ---
> 
> I have all optimisations on. I am using solaris with gcc2.95. 
> No things like netscape running in bg, or two calculations 
> running at the same time.
> 
> ----
> 
> 
> I still will try svd today, just to compare.
> 
> 
> 
> thank you for feedback
> 
> dominique
> 
> 
> 
> 
> 
> _______________________________________________
> Vxl-users mailing list
> Vxl-users@lists.sourceforge.net 
> https://lists.sourceforge.net/lists/listinfo/v> xl-users
> 
> _______________________________________________
> 
> Insight-developers mailing list Insight-developers@public.kitware.com
> http://public.kitware.com/mailman/listinfo/insight-developers
>