Interesting approach. I will be real interested in hearing your results with this compared to the traditional row or column major storage mode.
On Fri, Dec 19, 2008 at 4:29 PM, Luc Maisonobe <[email protected]>wrote: > ... <[email protected]>The new DenseRealMatrix tries to be more > cache-friendly. The matrix is > split in square blocks (32x32 blocks currently) except the blocks on the > right and bottom sides which may be smaller. Each block is flattened in > row major order in an array. These arrays are therefore 1024 elements > long for regular blocks, and smaller for border blocks. Most algorithms > can be organized to process each block completely before needing to load > the next block in memory. This is were the performance gain occurs. > >
