I'm working on a problem where I do linear algebra on progressively larger n by n submatrices in the upper-left of a (potentially large) N by N matrix. If I'm using a plain ol' dense matrix, the linear algebra takes place 30 times faster when I plug it into BLAS compared to a (fairly naive, tbh) implementation of the same in Julia. So my question is whether anyone can think of a cool way to accomplish the task of doing BLAS operations on just the submatrices I care about, without doing the nasty total of N-1 array allocations of size (2^2, 3^2, ... , N^2) and copying one array to the next. ...Which I haven't tested but suspect would be substantially slower than just doing the linear algebra in Julia.
Thanks! Gregor
