Use a well-known good reference library to compare against

[kknox]

Right now, our unit tests are written to compare the opencl device result with a hand-written vanilla C++ solution. This was easy to bootstrap with, but as the functions we implement get more complicated, the reference solutions will get more complicated. Eventually, there may be bugs in either implementation.

We should compare ourselves with a respected reference source, either viennacl or boost ublas. Both libraries implement sparse operations, and it should be trivial to wire either of those in as a reference source. Our tests already have a dependency on boost, so maybe it makes sense to use ublas.

[kknox]

The Armadillo C++ matrix library also seems like a good candidate: http://arma.sourceforge.net/

http://sourceforge.net/p/arma/code/HEAD/tree/

[lmiroslaw]

I vote for PETSc. It is a widely accepted tool with sparse matrix solvers.
http://www.mcs.anl.gov/petsc/features/index.html

On Tue, Jun 9, 2015 at 4:07 PM, Kent Knox notifications@github.com wrote:

The Armadillo C++ matrix library also seems like a good candidate:
http://arma.sourceforge.net/

http://sourceforge.net/p/arma/code/HEAD/tree/

—
Reply to this email directly or view it on GitHub
#75 (comment)
.

Łukasz Mirosław, Ph.D.
Senior Research Associate
Institute of Informatics
Wrocław University of Technology
Wyb. Wyspiańskiego 27
PL50370 Wrocław, Poland

Email: lukasz.miroslaw@pwr.wroc.pl
Tel.: +48713203969
Web: www.ii.pwr.wroc.pl/~miroslaw http://www.ii.pwr.wroc.pl/%7Emiroslaw

[jpola]

I vote for ublas, boost is already connected to our project so it will be very simple to use. I don't want to plug another dependency. Tests should check correctness not the performance, so let's do that with as little effort as possible.

[jpola]

During refractoring the tests projects I wanted to introduce uBLAS to generate references for our algorithms. I met some difficulties with it which I don't want to describe here but I checked the PETSc and Armadillo libraries as a substitutes:

• It seems that PETSc is good but unfortunately it is not well supported for Windows, it requires Python and Cygwin as build platform. I don't think we would like to support so much dependencies for our project.
• Armadillo is good. For windows it provides precompiled binaries and for Linux it can be build with the cmake system. However I found some issues related to sparse matrix solvers, it is quite new for this library and also can give us some issues.

For now I will leave tests without good reference library and I will use uBLAS as much as I can. When we would like to implement more advance algorithms like LU factorisation we will have to choose something.

After brief reading about both libraries I change my vote for Armadillo since it is more sutiable for our project. I will be sure after writing some examples using sparse matrices.

[kknox]

Thanks for doing the reconnaissance work on this Kuba. I also read that Armadillo was a decent alternative. I'm curious about the limitations that you ran into with uBLAS, maybe you can tell me over a phone conference. 5

[jpola]

Hi,
We can talk, I'm available in the late afternoon of Polish time (18-22).
You are free to choose.

2015-06-24 3:27 GMT+02:00 Kent Knox notifications@github.com:

Thanks for doing the reconnaissance work on this Kuba. I also read that
Armadillo was a decent alternative. I'm curious about the limitations that
you ran into with uBLAS, maybe you can tell me over a phone conference. 5

—
Reply to this email directly or view it on GitHub
#75 (comment)
.

[jpola]

I did some additional research and I found that Eigen library (http://eigen.tuxfamily.org/index.php?title=Main_Page) seems to be the best among many which I have tested.

• Template library (no need to build binaries)
• Mature (v3.0)
• Support for sparse matrix algebra.
• Wide range of solvers and advanced preconditioners for sparse matrices

[matusi143]

I'm curious why ViennaCL was not chosen as it is a openCL library in and of itself. The sparse matrix and iterative solver functions seems to be a very close parallel to what is being worked on here. I'm actually just now starting to use both clSPARSE and ViennaCL in my own application to compare performance and (hopefully) confirm numerical correctness of each library. Much of ViennaCL does not require boost, like this project and adds no other dependency.

[kknox]

This thread sums up our entire conversation into our reasoning; I think @jpola and I were mostly looking for a mature CPU based library to compare our results against, not another Accelerator based (opencl, cuda or whatever backend) library. We are not so much concerned about performance, as just to stability, correctness and ease of use for our unit tests. There is (albeit unsubstantiated) comfort in comparing against CPU results (I know viennacl supports a multi-core cpu backend, but it still 'feels' like an Accelerator library).

That said, I do not think either @jlgreathouse, @jpola or I have been too happy with the uBLAS library of late which was the one we picked. You can read some of our dissatisfaction. [1] [2]

We have not made further plans yet.