I am trying to implement big matrix tiled LU factorization myself (described in https://arxiv.org/pdf/0709.1272.pdf, https://www.researchgate.net/publication/278629667 and so on). I understand panel algorithm pretty well, but can't understand how rows reordering is made in the tiled one.
After we processed k panels, we are factorizing tile Akk and during this we choose pivots and swap rows to put pivots onto diagonal. What tiled algorithms' authors don't mention is that entire rows, including left parts (in "L (done)") has to be reordered in accordance with pivoting. We have partially calculated L and U, values are stored in memory and now (during pivoting) we are deciding to change our source matrix so it would be more "convenient" for us. Is that correct?
