The special case of a 3x3 matrix is not handled correctly by both algorithm I and algorithm II.
They have the access pattern first 2 rows - all rows in between - last 2 rows, but for a 3x3 matrix, the first and last 2 rows have an overlap because the second row belongs to both.
This is an edge case, but yet it can happen and lead to wrong memory access.
Unfortunately, the tests do not capture this case because they only capture a single case of n=1_000.
I have already fixed this for #11 by falling back to np.linalg.solve for this special case.
MuellerSeb
commented
2 months ago
The special case of a 3x3 matrix is not handled correctly by both algorithm I and algorithm II. They have the access pattern first 2 rows - all rows in between - last 2 rows, but for a 3x3 matrix, the first and last 2 rows have an overlap because the second row belongs to both.
This is an edge case, but yet it can happen and lead to wrong memory access.
Unfortunately, the tests do not capture this case because they only capture a single case of
n=1_000.I have already fixed this for #11 by falling back to
np.linalg.solvefor this special case.