Please do file issues on slicing. The implementations have a bit of effort 
put into them, but there is most likely room for a lot of optimization. It 
would be useful to have the exact cases where performance doesn't meet 
expectations.

In general for stiffness matrix assembly, I believe that using 
`sparse(I,J,V)` should be the way to go, if it works.

I should also write some performance notes for working on sparse matrices 
too.

-viral

On Wednesday, January 7, 2015 11:26:09 PM UTC+5:30, Tim Holy wrote:
>
> flipping (or, I guess, rotation) has already been fixed by kshyatt, who 
> has been 
> writing tests. (!) 
>
> https://github.com/JuliaLang/julia/pull/9593 
>
> I guess the flip is more of a mirror symmetry, but you should be able to 
> copy 
> her pull request to implement it. (Please do submit as a pull request.) 
>
> The slicing thing also needs to be fixed, that would be good to open as an 
> issue. 
>
> --Tim 
>
>
> On Wednesday, January 07, 2015 12:35:48 PM Stefan Karpinski wrote: 
> > I suspect this is precisely the case. Please do open an issue. 
> > 
> > On Wed, Jan 7, 2015 at 12:24 PM, Kevin Squire <[email protected]> 
> > 
> > wrote: 
> > > I would suggest doing a quick search on github for any issues related 
> to 
> > > sparse matrices, and if none are relevant, opening a new one.  One 
> likely 
> > > possibility is that neither function has been specialized for sparse 
> > > matrices, and that the general fallback version of these functions are 
> > > slow 
> > > for sparse matrices.  In either case, a specialized function should be 
> > > easy 
> > > to code. 
> > > 
> > > Cheers, 
> > > 
> > >    Kevin 
> > > 
> > > On Wed, Jan 7, 2015 at 9:20 AM, Christoph Ortner < 
> > > 
> > > [email protected]> wrote: 
> > >> I don't have a solution, but a related comment: slicing of large 
> sparse 
> > >> matrices is very slow too. E.g., I like to assemble a stiffness 
> matrix 
> > >> (for 
> > >> a PDE) with free boundary conditions and then impose clamped 
> boundaries 
> > >> by 
> > >> taking a subset of the matrix, 
> > >> 
> > >>     B = A[Ifree, Ifree] 
> > >> 
> > >> this is so slow that I had to completely give up on it and rewrite 
> all my 
> > >> assembly codes. 
> > >> 
> > >> I suspect that the flipxx functions and this are a similar issue? 
> > >> 
> > >> Is there already an issue filed for this? 
> > >> 
> > >> Christoph 
>
>

Reply via email to