Dear Andreas, J is for half of the elements diagonal. So if Rm is also diagonal, the LU factorization will be avoided in those cases. (Provided what you say for \ also applies to the /.)
Bravo. And many thanks. It was driving me nuts. Petr On Saturday, December 13, 2014 4:13:50 PM UTC-8, Andreas Noack wrote: > > I don't know how J looks, but could it be that it is diagonal? If so, I > have an idea for the reason for this. If you look here > > https://github.com/JuliaLang/julia/blob/master/base/linalg/dense.jl#L412 > > you can see the \ checks if the matrix is diagonal. If it is, the solver > uses the \(Diagonal, Matrix) method which avoid the LU factorization that > \(Matrix, Matrix) uses. > > Do you think that is the reason? > > 2014-12-13 17:56 GMT-05:00 Petr Krysl <krysl...@gmail.com <javascript:>>: >> >> You're absolutely right, sorry! I got distracted and I forgot to post >> instructions. I've also tried to simplify things, and I think I've gotten >> it down to just a few lines of code. Unfortunately, the rest of the code >> still needs to be there for declarations and such (https://github.com/ >> PetrKryslUCSD/Test_jfineale). >> >> The rotation matrix gets used in one place only. Refer to lines 68ff in >> conductivity() from FEMMHeatDiffusionModule.jl. >> >> # Here is that rotation matrix Rm >> FEMMBaseModule.getRm!(self.femmbase,Rm,Ns[j]'*x,J,labels[i]); >> # Material orientation matrix >> # And here is the only place it gets used >> gradN = gradNparams[j]/(Rm'*J); # gradient WRT material >> coordinates >> # >> >> conductivity() is the last thing called in the driver. Picking Rm on line >> 41 of the driver (Poisson_FE_example.jl) leads to different computation >> CPU times. >> >> Petr >> >> On Saturday, December 13, 2014 1:37:36 PM UTC-8, Steven G. Johnson wrote: >>> >>> Unfortunately, it's not really reasonable for people on a mailing list >>> to dissect such a large body of code; you need to boil it down to a smaller >>> example that illustrates the behavior you mentioned. >>> >>
