unexpected performance for SpMV operation

[rohany]

I'm running some CTF SpMV kernels using the contraction interface (a["i"] = B["ij"] * c["j"]), and I'm seeing performance that is nearly 2 orders of magnitude slower than systems like PETSc and Trilinos on large matrices from the suitesparse collection (such as the arabic-2005 graph). I know I've asked similar questions to this before, but I was wondering if such a discrepancy is expected, or if there is a specialized kernel available in CTF for the SpMV operation. I believe I've configured CTF correctly for my system, but I'm happy to share the configuration logs to double check.

cc @solomonik

[raghavendrak]

CTF treats vectors no different from tensors, and the code has to account for generalized redistribution and storage format. We did use specialized code to handle vectors without overheads in a different project, and found the performance gains to be quite significant. We will integrate this into CTF soon, and tag here when it is done.