Cell Sparse Matrix Vector

[Jawed]

This presentation, Optimization of Sparse Matrix-Vector Multiplication on Emerging Multicore Platforms, from SC2007

http://www.cs.berkeley.edu/~samw/projects/multicore/sc2007.pdf

is interesting because it shows how Cell makes considerably more effective use of its available bandwidth than x86.

Jawed

["Nerve-Damage"]

Thank's Jawed.

[Panajev2001a]

Fixed link: http://www.cs.berkeley.edu/~samw/research/papers/sc07.pdf

[popper]

This presentation, Optimization of Sparse Matrix-Vector Multiplication on Emerging Multicore Platforms, from SC2007

http://www.cs.berkeley.edu/~samw/projects/multicore/sc2007.pdf

is interesting because it shows how Cell makes considerably more effective use of its available bandwidth than x86.

Jawed

given i just skiped the PDF so i might be wrong but they seem to have totally ignored the onchip altivec unit of the cell?...

if so it seems a massive waste not to even try and include vectorised Altivec code in any optimisations they performed for max performance for the tests at hand.

the freeVec lib for instance shows proven altivec abilitys when used in PPC linux problem solving With lots of profiling and testing etc.

http://www.powerdeveloper.org/forums/viewforum.php?f=23
you guys might even want to give markos a hand in this new thread too.
http://www.powerdeveloper.org/forums/viewtopic.php?t=1410

[Vitaly Vidmirov]

it seems a massive waste
...of transistors

SPEs for computation, PPE for service. That is the whole point of CELL.

[popper]

i wouldnt say that, after all, the PPE is really a PPC at heart and the Altivec even today, clock for clock still outdoes SSE vectorised code.

its a shame as Altivecs already on there and fully functional, and other than the few guys over on powerDev etc no one new seems interested in vectorising key PPC Linux for a free boost in performance all this time, never mind giving to the time to get good current code pushed upstream, or even into your everyday apps.

its strange, if you are to beleave the new PS3 coders are really interested in makeing their code, or the existing PPC base code that they are using on the PS3 today,as effective and fast as possible for free with existing and new Altivec replacement code etc.

perhaps theres just not as many interested OSS PPC/PS3 coders as we are led to beleave comeing online, at least i dont see much of anyone new offering to try and really help the PPC/PS3 linux efforts, arr well...

perhaps their all SSE coders and dont know how good and effective Altivec can be ,even for your use/assumption that the PPE is only for service, Altivec can help offload the CPU core servicing of the SPEs?...

[Panajev2001a]

I welcome projects such as libfreevec and all other VMX optimized projects: the problem (or so it seems, your mileage might vary) is that mixing regular FXU/FPU work and VMX (Integer and/or FP) work seems to be a bad performance combination as you cannot exchange data across FP and VMX registers (for example) directly and you have to go higher up in the memory hierarchy.
If you were already using VMX quite extensively you might want to give it as much work as possible, but going back and forth between FXU/FPU and VMX might lower performance rather than increase it so you have to be careful with that.