dwfmarchant / pymkl Goto Github PK

Hi Dave,

Should be straightforward to get your package listed on PyPI. When ready, you just need to tag the release (git tag v?.?.? or git tag -s v?.?.?). This should match the setup.py file. Then, you can run

python setup.py register -r pypitest

to set things up with PyPI testing, and

python setup.py register -r pypi

to do the same on the live site. After that, make testpublish and make publish should work.

http://peterdowns.com/posts/first-time-with-pypi.html

Hi,

The Intel docs for pardiso at [1] suggest that if iparm[24] (parallel factorization) is enabled, then iparm[11] and iparm[13] should be disabled. Is there a reason why pyMKL isn't doing that?

[1] https://software.intel.com/content/www/us/en/develop/documentation/mkl-developer-reference-fortran/top/sparse-solver-routines/intel-mkl-pardiso-parallel-direct-sparse-solver-interface/pardiso-iparm-parameter.html#pardiso-iparm-parameter_IPARM35

I am having trouble on linux/mac compatibilities. Works on my mac, does not work on travis.

See simpeg/simpeg#401

@jcapriot suggested that there might be an ordering issue. I think this might be good to get into the base repo. I have wrapped it in simpeg/pymatsolver#13 but probably should be in the base one and ensuring that the final matrix is actually sorted correctly.

A = A.tocsr()
if not A.has_sorted_indices:
    A.sort_indices()

https://software.intel.com/en-us/node/470282

CAUTION
Intel MKL PARDISO column indices ja must be in increasing order per row. 
You can validate the sparse matrix structure with the matrix checker (iparm(27))

Pardiso has a bug in it's nested disection reordering algorithm (both serial and parallel).

Until Intel releases an update to 11.3.4, can you set iparm[1] = 0 at https://github.com/dwfmarchant/pyMKL/blob/master/pyMKL/pardisoSolver.py#L131

For details on the bug see [https://software.intel.com/en-us/forums/intel-math-kernel-library/topic/657934]

@jcapriot @dwfmarchant For recent versions of Anaconda3 for Windows OS (with Python 3.8 and possibly 3.7), the latest conda install of pymatsolver cannot import Pardiso. I have tracked down this issue to loadMKL.py in the pyMKL package.

I found that for Windows OS, the code will try to find the mkl_rt.dll in anaconda/Library/bin/. This file doesn't exist. Instead, there is a mkl_rt.1.dll. By changing the filename for the MKL library in loadMKL.py, I was able to import Pardiso from pymatsolver and run a TDEM forward model with the Pardiso solver.

Can this file be fixed so that it is up-to-date with current installations of Anaconda?

I noticed that the pardisoSolver.py has a factor() method but I was wondering if the python interface can access the actual L, U factors if, for example, I did an LU factorization (as well as the permutation matrices for any reorderings done).

Repeatedly constructing and initializing instances of pardisoSolver leaks memory as the internal data is never freed.

I noticed that the class has a clear method, but it isn't automatically invoked. Perhaps it could be called in the __del__ method of the class? One would have to make sure that multiple calls to clear are safe.