Why does CuBlas not use a 1d array for triangular matrices?

Question

This might be a throwback to the old BLAS library design, but I was surprised just now to find that CuBlas wastes memory by using regular 2d arrays for triangular matrices. I suppose this makes interfacing with the rest of the API more convenient.

Answer 1

I was surprised just now to find that CuBlas wastes memory by using regular 2d arrays for triangular matrices

That isn't strictly true.

If you look at the Level 2 BLAS routines, you will see that they operate on triangular or Hermitian matrices stored in a packed format.

The Level 3 BLAS routines don't, but there are two good reasons why they are stored in full dense format.

• BLAS does it that way

• Those routines were mostly added to BLAS as support for LAPACK solvers. And those solvers typically store the results of factorizations in-situ in supplied full dense inputs, so it is logical to use that format in BLAS

I guess if you don't like the design choice you can always try writing to Jack Dongarra to complain.