2012 12 04

Agenda

• New schedule
• Nikola, run time errors etc.
• Vectorisation of Nikola vs. Accelerate

Minutes

• Decisions: We should be able to argue for our decisions. Which decisions have we made so far and how are our arguments:

  • Nikola vs. Accelerate? Why the one over the other?

    (Philip offline comment: Sure, it would be good to have this out of the way, so we have our arguments straight, and so we don't end up doing one if the other is better suited. I'm going to try and do some report writing on this, and the result should be a preliminary argument for Nikola or Accelerate)

  • FFI bindings to polyfit/LinAlg libs/Sobol sequences or our own implementations? How hard will it be to do FFI into Thrust/CUBLAS/NAG in Accelerate or Nikola (new evaluation item)?

    (Philip offline comment: For the time being I don't think we have enough room (i.e. time) for any new items in the survey. But as we extend Nikola (or accelerate for that matter) we will probably have more easy access to this information anyway as we already have our hands very dirty with at least one of them. I'm going to make a future-work card for this that may then later be moved to general work if its feasible. But we should probably note that once we have settled on Acc/Nikola whatever we may find will not be able to effect our choice of Acc/Nikola, so maybe it's not so well suited as a part of the survey if it's added after settling.)

We can take both of these decisions, we just need to make sure we have the arguments ready. They need to be in the report.

• Sobol sequence implementation gives problems: missing fold and filter in Nikola. Other problems occur in the binomial pricer when trying to move generation of uPow/dPow to the GPU.

  • Ken: We should email Geoffrey ASAP with our examples and ask how easy it would be to add the missing operations or which alternatives there are.
  • Martin: We can also email Trevor about our efforts with the Sobol sequence generator in Accelerate and where we are stuck.

• Martin shall commit his experiments with implementing Sobol sequences in Accelerate and Nikola, and the alternative Nikola binomial pricer implementation.

• We should create a list of our contributions: all the contributions we already have, the contributions we wish to have and the contributions we postpone to future work. (Martin will create a new wiki page for this)

• (Philip offline addition: I'm going to update the planning page to include item justifications (the why) and expected outcomes, the way I imagine them, as per Kens trello-comment to the planning page. Then Ken and Martin may comment on them.)