[
https://issues.apache.org/jira/browse/MATH-230?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12653566#action_12653566
]
Ismael Juma commented on MATH-230:
----------------------------------
One issue with the proposed implementation is that it causes excessive
objective allocation. Each non-zero entry involves a HashMap.Entry, a Point and
a Double. This is quite inefficient in terms of memory usage.
One alternative would be to use a specialised HashMap with open addressing that
would only take double primitives as values (similar to TObjectDoubleHashMap
from Trove). Using this approach, the only object would be Point. It's possible
to avoid that too, but that's more complicated (look at Colt's
SparseDoubleMatrix2D for an example).
> Implement Sparse Matrix Support
> -------------------------------
>
> Key: MATH-230
> URL: https://issues.apache.org/jira/browse/MATH-230
> Project: Commons Math
> Issue Type: Improvement
> Affects Versions: 2.0
> Environment: N/A
> Reporter: Sujit Pal
> Assignee: Luc Maisonobe
> Priority: Minor
> Fix For: 2.0
>
> Attachments: patch.txt, RealMatrixImplPerformanceTest.java,
> SparseRealMatrixImpl.java, SparseRealMatrixImplTest.java
>
>
> I needed a way to deal with large sparse matrices using commons-math
> RealMatrix, so I implemented it. The SparseRealMatrixImpl is a subclass of
> RealMatrixImpl, and the backing data structure is a Map<Point,Double>, where
> Point is a struct like inner-class which exposes two int parameters row and
> column. I had to make some changes to the existing components to keep the
> code for SparseRealMatrixImpl clean. Here are the details.
> 1) RealMatrix.java:
> - added a new method setEntry(int, int, double) to set data into a matrix
> 2) RealMatrixImpl.java:
> - changed all internal calls to data[i][j] to getEntry(i,j).
> - for some methods such as add(), subtract(), premultiply(), etc, there
> was code that checked for ClassCastException and had two versions,
> one for a generic RealMatrix and one for a RealMatrixImpl. This has
> been changed to have only one that operates on a RealMatrix. The
> result is something like auto-type casting. So if:
> RealMatrixImpl.add(RealMatrix) returns a RealMatrixImpl
> SparseRealMatrixImpl.add(RealMatrix) returns a SparseRealMatrixImpl
> 3) SparseRealMatrixImpl added as a subclass of RealMatrixImpl.
> 4) LUDecompositionImpl changed to use a clone of the passed in RealMatrix
> instead of its data[][] block, and now it uses clone.getEntry(row,col)
> calls instead of data[row][col] calls.
> 5) LUDecompositionImpl returned RealMatrixImpl for getL(), getU(), getP()
> and solve(). It now returns the same RealMatrix impl that is passed
> in through its constructor for these methods.
> 6) New test for SparseRealMatrixImpl, mimics the tests in RealMatrixImplTest,
> 7) New static method to create SparseRealMatrixImpl out of a double[][] in
> MatrixUtils.createSparseRealMatrix().
> but using SparseRealMatrixImpl.
> 8) Verified that all JUnit tests pass.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
