scaling performance of libswscale versus OpenGL

There are many ways to scale an image. One can use provided libraries, or implement scaling via OpenGL pipeline. Latter choice looks preferred, as it uses hardware part which is desined to do suck tasks efficiently. But how fast OpenGL is? I've got following results on Celeron P4500 (it has integrated Ironlake video chip):

another fluent 3ddp parallel benchmark

Another fluent 3ddp benchmark on Phenom II X6 1090T.
Unfortunatelly I don't remember what case was used.

procs; time per iteration; wall time; cpu time

smpi:
1 4.577 457.661 457.330
2 2.488 248.802 496.820
3 1.792 179.155 536.430
4 1.475 147.455 587.010
5 1.221 122.108 605.510
6 1.081 108.132 632.970
7 2.314 231.399 1202.540

nmpi:
5 1.569 156.866 669.320 nmpi
6 1.253 125.323 648.960 nmpi
7 3.235 323.487 1728.310 nmpi