Hey,

 >> I found the cause: In the Boost.numpy CMakeLists.txt, there is this:
>>
>> https://github.com/tsmithe/Boost.NumPy/blob/34ace9d01efb7a9cf0fe421b86c78400e9632f65/CMakeLists.txt#L50
>>
>> Changing the 'SHARED' to 'STATIC' on non-Windows fixes the problem (or
>> pass LIBRARY_TYPE accordingly to CMake). However, I wonder why this
>> succeeds on other (Ubuntu) machines? Is there a system-wide static
>> libboost-numpy available?
>
> Thanks! That is very odd.. I just checked, and the only static lib file
> on my system is the one produced in the build tree for PyViennaCL. And
> if you look at
>
> https://github.com/tsmithe/Boost.NumPy/blob/34ace9d01efb7a9cf0fe421b86c78400e9632f65/libs/numpy/src/CMakeLists.txt#L1
>
> then I've commented out ${LIBRARY_TYPE}, so I don't know where the
> effect would come from.. Anyway, I'll clean up that CMakeLists.txt.

Argl, the reason is that I had to do a manual clone of your Boost.NumPy 
repo and did not change the branch. I don't think it's good to have a 
patched repo for Boost.NumPy around...



> Any ideas about the solver bug? It seems to compute the norm of the RHS
> and then stop -- for some reason, it computes the norm as 0 in
> linalg/gmres.hpp (and I don't think that the norm of that example vector
> is 0!)..

Still searching. The build process is so awfully slow, it's killing 
productivity. It takes me minutes to just get a compiler error if 
something was wrong...


> I've also remembered that my own code to compute vector norms using
> scheduler calls doesn't work, and that seems to be because I'm building
> the expression tree in that case as I do for all other expressions at
> the moment, like this:
>
> Assign(Scalar, Norm_2(Vector))
>
> but there doesn't seem to be functionality for that kind of scalar
> expression in the scheduler, so that needs fixing. But I'm not sure that
> the two issues are related.

That might be the cause, because norms are essential in the iterative 
solvers.


> Finally, you can improve recompilation speeds using ccache, but alas I
> think it will remain a limitation that standard parallel builds do not
> work, thanks to Python's distutils.

We need to Python's distutils as much as possible and only expose the 
very least amount of targets. libviennacl would certainly be *the* 
solution to this problem.

Best regards,
Karli


------------------------------------------------------------------------------
Managing the Performance of Cloud-Based Applications
Take advantage of what the Cloud has to offer - Avoid Common Pitfalls.
Read the Whitepaper.
http://pubads.g.doubleclick.net/gampad/clk?id=121054471&iu=/4140/ostg.clktrk
_______________________________________________
ViennaCL-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/viennacl-devel

Reply via email to