On 01/31/2018 12:26 PM, Richard B. Kreckel wrote:
Could you guys please re-test against current git HEAD? The crashes you reported should be fixed and timings should have improved.
The crashes have vanished in my test case. Well done.
Based on this, it may now make more sense to compare algorithms for solving linear equations.
With today's commits, the same toy benchmark gives these timings (with 5 measurements per matrix size): matrix gauss time/bareiss time size average min max 2x2 1.261 1.181 1.313 3x3 1.569 1.385 1.661 4x4 1.549 1.490 1.647 5x5 1.629 1.598 1.692 6x6 1.409 1.191 1.713 7x7 1.014 0.742 1.335 8x8 0.843 0.385 1.542 9x9 0.414 0.172 0.599 10x10 0.184 0.061 0.622 11x11 0.096 0.029 0.210 So, Bareiss performs much better than previously. Returning to the previous topic of switching to Gauss elimination for sparse matrices: my goal here was to demonstrate that on dense matrices Gauss elimination is at least comparable to Bareiss. I think this point should be uncontroversial now.