I liked
http://arxiv.org/abs/1307.6209
which seems to show that modified ELLPACK can be good for
multicore as well as manycore. The sad part of course is that no
mention of parallel is made.
I still have to go over it in detail, yet the performance numbers they
get look reasonable. This was also confirmed by some folks I met in
Lausanne two weeks back. The only drawback is that their kernels use all
these low-level intrinsics, which is a pain when it comes to portable
code and productivity.
Best regards,
Karli
