The indexes must be provided in (row,column) form. Functions in this package are programmed with care so no overflow of linear index is used with large-size matrix. I try to design an engine of automatic selection algorithms so as to provide a good run time performance across various cases. I'm not sure it's worth the effort, but my experience shows that it is easy to mess with the performance when non-appropriate method is using in sparse manipulation. This package does not warranty for best performance, but it just has a modest pretension to prevent user to use bad algorithm. For a better customize solution, a calibration is required on the specific
computer by running this command: spcalib('auto')
Why must the number of rows equal the number of columns?
error('r and c must have a same dimension')
Got it thx!!
Im trying to build a large FE model with this package but i dont get any advantage on the speed.
shuangou ren how exactly do u use it one FE assemblage
For more confusion, matlab version 2010a didn't crash in the example above. The 2011b on the otherhand, did.
This fantastics submission seems to have sort of a bug. I accidentally did the followin: B = setsparse(A,inf,inf,1).
This didn't result in a problem before my function ended and tried to clear the variable B. I know that this is strictly the user's problem but a single check if the indexes are suitable could be useful for other unfortunate fools...
In my FE-Code the global matrices are assembled as usual in a loop over the elements, eg:
A(idx, idx) = A(idx, idx) + locA;
where idx defines a sub-matrix of A (eg idx = [3,1,4,12,7,10] and locA is 6x6). I translated this code to Your functions as follows using row and columns indexing:
[cidxs, ridxs] = meshgrid(idxs, idxs);
A = setsparse(A, ridxs(:), cidxs(:), locA(:), @plus);
Unfortunately, this is (little) more expensive than the original version. Do You have any suggestions how to shorten things above a little?
Thank you in anticipation for your support.
good job, two times faster for finite element assemblage.
I can't really help with memory problem. The suggestion I can give is to try on your side to increase the free largest memory block from other places. Thanks for the kind words.
This seems to have one problem though:
Before I used A(:,indexes)=0, it just took a lot of time.
No I tried
idxofzero = ismember(J,indexes);
And I get "out of memory error". I tried it several times with both methods and got the same result. Am I missing something here?
My sparse matrix is about 55,000 x 55,000 and it has 4,060,200 nnz elements. The indexes hold about 1/100 of the columns.
Bruno, Thank you very much for this package. Also thank you for all your help in the forums.
This works appr. 30 times faster than using linear indexing to zero sparse matrix columns that I used before!
Thanks for the tip James.
FYI, my experience is that lcc has bugs with the unsigned long long type ... I often get a "bad rule" compile error. I have gotten into the habit of using signed long long instead and writing extra code to handle situations where I need to shift into or out of the sign bit. A real hassle just so the lcc compiler that ships with MATLAB will work with my code.
Yep, that workaround fixed it. Anyway, I like it!!
After investigation, it turns out the "_kids--Bad rule" message is caused by a bug of the LCC compiler.
A workaround is to replace the or "|" operator by plus "+" operator in the cmex file setspvalmex.c, lines #626, 634 and 709, then recompile.
Thanks for report the bug. I'll submit a new version.
Bruno - When I do
mex -v -O setspvalmex.c
I get a compilation error:
Error setspvalmex.c: 626 compiler error in _kids--Bad rule number 0
C:\PROGRA~1\MATLAB\R2009B\BIN\MEX.PL: Error: Compile of 'setspvalmex.c' failed.
Any idea what I'm doing wrong?
Sorry about that.... I think that the following code does the trick
A_global = setspvalmex(A_global, Ic, Jr, X,@plus);
Very good!!!! It works well for building global stiffness matrices for finite element computations. While building my stiffness matrices I currently do the following
val = getspvalmex(A_global, Ic, Jr);
A_global = setspvalmex(A_global, Ic, Jr, val+X);
Bruno...Is there any other way your code can be used to get this done??
When passing scalar row or column, it will automatically expanded to match the other
switch the sort-engine of setsparse to introsort which is more stable than quicksort (previously used) in term of complexity
Workaround for BUG of LCC (related to "|" operator in expression with int64)
Correct bug when calling setspvalmex.c with empty indice.
Correct a bug in setspvalmex.c, which concerns 64-bit only
A new mex sparse assign engine is developped from scratch using quicksort. Finally, we beat adding routine by about 40% of speed. More comprehensive calibration routine is included.
More comprehensive calibration
A brand new mex implementation that overcome Matlab indexing Bug (32 bit platform) and improve the speed by approximately a factor of 4.
Adjust the maximum of linear index allowed according to bug information communicated by The Mathworks (applicable only for 32-bit platform, v.2009A or previous)
Add a feature of bi-variate function-handle when assign values