Question

I am writing a numpy/cython program to compute the minors of small matrices (a lot of them).

My current function looks like (computes the minor of mat wrt. to row ii, col jj):

cdef float minor(np.ndarray[DTYPE_t, ndim = 2] mat,int ii,int jj): 
    rows = range(mat.shape[0])
    col = range(mat.shape[0])
    del rows[ii]
    del col[jj]

    cdef np.ndarray[DTYPE_t, ndim = 2] rM = (mat[rows])[:,col]

    cdef float val =  (-1)**(ii+jj) * np.linalg.det(rM)

    return val

After a little benchmarking, the line

cdef np.ndarray[DTYPE_t, ndim = 2] rM = (mat[rows])[:,col]

Is rather time consuming. Is there a better way to remove one row and one column from an two dimensional array?

Yours,

cp3028

Answer 1

It looks like you’re copying the memory from (mat[rows])[:,col], allocation and copying is a slow process. Is it not possible to simply make the function call np.linalg.deg on the chunks of mat in place, instead of copying it and calculating the determinant on the copy?