I've just skimmed through the paper, it seems they share lots of high-level concepts with Firadrake/FEniCS, although it's weird they don't systematically compare to these approaches (OK, fine for Firedrake that is a new system, but FEniCS...) As for the code optimizations: the loop fission thing is interesting, although it's exactly the work the Carlo and OP2 people did in OP2. And they don't say the most important aspects: *when* they have to apply that and what "semi-automatically" means. As for the "explicit" vectorization, I'm really surprised a compiler can't do what they claim, perhaps that's why they don't even mention how they compiled the code, as far as I saw (btw, they also contradict themselves, since they also say "Since Kranc generates out- put in a high-level language (e.g. C++), it does not have to deal with very low-level architecture details such as register allocation or instruction selection." It would be interesting to know more. -- Fabio 2014-10-11 20:01 GMT+01:00 McRae, Andrew <a.mcrae12@imperial.ac.uk>:

"...The two most important optimisations were loop fission to not overflow the instruction cache, and manual vectorisation..."

Sounds like they have their own Fabio!

On 11 October 2014 11:22, Rathgeber, Florian < florian.rathgeber@imperial.ac.uk> wrote:

...

In Chemora, PDEs are expressed either in a high-level LaTeX-like language or in Mathematica. It's based on Cactus:

http://arxiv.org/abs/1410.1764

Florian