77
edits
Edu159
Joined 2 March 2014
→Y: Your task
| Line 99: | Line 99: | ||
* Approach: | * Approach: | ||
**'''ilu:''' That function has a big chunk of options and the last year was almost implemented by Kai Torben as his GSOC project. He interfaced Octave with ITSOL/ZITSOL libraries but in the end there were some issues with that approach | **'''ilu:''' That function has a big chunk of options and the last year was almost implemented by Kai Torben as his GSOC [http://siko1056-gsoc.blogspot.de/ project]. He interfaced Octave with ITSOL/ZITSOL libraries but in the end there were some issues with that approach: | ||
:#ILUTP algorithm did not work for him | |||
:#He had to patch the library to get things work, not nice for integration. | |||
:#modified versions of algorithms ("milu" option) were not implemented in the libraries | |||
:#That "ugly" scenario lead to finally not being able to include ITSOL as a dependency with Octave. Bottom line, the integration of the function with the development\n repository could not be achieved. | |||
::I have been in contact with Kai by mail and agree that writing from scratch all the functions needed as oct-files (ILUTP, ILU0, ILUC and ILUT) would be a valid way to go. This way no dependencies are needed to be added and the overhead of translating the data from Octave to ITSOL and vice verse is eliminated. Algorithms will be taken from Yousef Saad's book "Iterative methods for sparse linear systems Ed. 2". Moreover, I can use some of the code that Kai wrote, mostly the tests, documentation and the m-file "ilu.m" that glue together all the functions. ITSOL source code is also a good place to look for some help. | |||
::I have implemented the ILU0 algorithm so far and benchmarked it against Matlab and Kai's last GsOC version(using ITSOL). The performance is great. You can check the code and see a table with the execution times in a blog I have created for the project([http://edu159-gsoc2014.blogspot.com.es/2014/03/ilu0-implementation.html link]) | |||
:*'''ichol:''' In this case things should be easier. Kai implemented the functions related with ichol from scratch and they works great. It would be only necessary to code the complex version and the modified version of the algorithms. There is one issue here with licenses I did not know at first and Kai pointed me out (see [http://edu159-gsoc2014.blogspot.com.es/2014/03/introducing-myself.html#bc_0_1B here]). It should be solved. | |||
:*'''sprand/sprandn:'''I have already sent a [http://savannah.gnu.org/bugs/?41839 patch] for those functions that IMHO works, but it suffers from some performance issues for large matrices. Should be revisited by someone that really knows about that. I mailed Rik who appeared as author of the function and has told me that in a few weeks he maybe be able to revise the algorithm. So in the meanwhile I will focus on other functions. | |||
:*'''sprandsym:'''That function may be implemented following the strategy used in sprand, but I need to analyze closer. | |||
:*'''minres & lsqr:'''With the implementation of those ones I need a bit of guidance. I dont know if they must be implemented as m-scripts or in c++. I would say that they way to go is m-scripting them as ''gmres'' is implemented this way, and they are cousins. Anyway I found that this website are the place to look [http://www.stanford.edu/group/SOL/software/lsqr/ lsqr] [http://www.stanford.edu/group/SOL/software/minres/ minres]. The website is from the people that wrote the papers given as references in Matlab documentation. In the website there are several codes that can be used. I have mailed professor Michael Saunders about adapting them into Octave versions and he answered me that I am welcome to do while I respect the license (CPL or BSD licenses). He claimed that they are very unrestrictive but I've been told that they are not compatible with GPL3. I will need some insights about that. | |||
:'''->'''I think I have a clear road map of what I want to do but I don't know for sure if it will be enough for the GSoC period. There are other functions that I would like to implement if it would be necessary. Let me know what do you think about. | |||
*Estimated timeline: | |||