Skip to content
This repository
tree: d5ba29f327
Fetching contributors…

Cannot retrieve contributors at this time

file 20 lines (15 sloc) 2.362 kb
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
1. Added full support for matrix inversions ONLY for general matrices.
This was the bulk of this commit, as it was very hard to get right, mostly because I had to work around LAPACK and also provide naive implementations for all the functions I use.
Expressions are divided into these kinds:
- X = inv(Y) * Z -> Y is LU decomposed and the problem is reduced to two triangular problems
- X = inv(Y) * v -> The same as the one above, but simplified due to the fact v is a vector
- X = inv(T) -> Y is LU decomposed and the LU inverses are computed using matrix multiplications
The LU decomposition in the naive implementation uses the Doolittle w/ pivoting algorithm (http://mymathlib.webtrellis.net/matrices/linearsystems/doolittle.html) this is very robust, works on any matrix and has complexity O(n^3).
The expression X = inv(A) * inv(B) requires a temporary for one of the inversions.

This addition introduced _a lot_ of complexity, and while it works on all unittests I came up with I would still use it with caution.

2. Added naive implementations for lapack.getrs (general-matrix-solve), lapack.trtri (triangular-matrix-inverse), lapack.getri (general-matrix-inverse), lapack.getrf (LU-P decomposition), lapack.gesv (general-matrix-solve with decomposition incorporated), lapack.laswp (permutation-matrix-premultiplication), blas.trmv (triangular-matrix-vector-product), blas.gemv (general-matrix-vector-product). Added a LAPACK-like function xgetrs, that performs a general-matrix-solve with an added side parameter (similar to the side parameter from blas.trsm). This means both inv(A)*B and A*inv(B) can be performed.

3. Changed transposition, side, diagonal etc. parameters from BLAS and LAPACK to template parameters. This allows for much simpler code in the naive implementations since one can place static-ifs inside loops, rather than add multiple copies of the loop for each if-branch.

4. Haven't written a naive implementation for trsm (it's a very frustrating function to implement), but I've added a fallbackSolve function that computes Temp := inv(Y); X := Temp * Z; This way only lapack.trtri is used when using the nodeps flag, which is implemented.

5. Added David's fix for Issue 20 and Issue 21.
6. Fixed multiple bugs involving complex numbers.
7. Added more tests for matrix operations based on results obtained from GNU Octave.
Something went wrong with that request. Please try again.