Add cuda implementations to TimedArray and BinomialFunction classes about brian2cuda HOT 8 CLOSED

I implemented BinomialFunction while trying to get the Potjans-Diesmann model to work. Didn't have the chance to test it yet. Implementation is in the binomial_implementation branch and not merged yet.

Tests should include:

• checking if brians seed mechanism works correclty the way it is implemented
• This implementation uses device side curand calls. We would need some device side buffer system in order to be able to use host side random number generation since we don't know how many random numbers will be needed in the current binomial implementation (rand inside while loop). Check performance loss of device side curand calls.
• We are just using the cpp binomial implementation. Could be parallelise that?

from brian2cuda.

Wrote a test for checking brians seed mechanism, but it uses multiple run() calls with random number generation, therefore fails due to #85. Fix that first.

from brian2cuda.

Binomial implementation is merged, including testing (hopefully) all different RNG cases. See PR #178 for details.

from brian2cuda.

@mstimberg I am implementing TimedArray right now. At first glance I'm not entirely sure how to add a new implementation for the cuda_standalone device. For the BinomialFunction, there is a BinomialFunction.implementations dictionary defined that can be extended by new implementations. Should I change the TimedArray implementation accordingly or is there something I am missing / another way to easily add the cuda implementation?

from brian2cuda.

Indeed! The implementations are slightly more tricky for TimedArray – the implementing function has to return another function that actually generates the code. This is because the final implementation might depend on the dt of the context, and that context is not known until the actual run. I made the changes to TimedArray in the timedarray_implementations branch, let me know whether this works for you.

from brian2cuda.

Thanks @mstimberg, this works!

from brian2cuda.

Ok thanks, I'll open a PR and merge this into Brian master soon.

from brian2cuda.

Binomial function was implemented in PR #178.
TimedArray was implemented in PR #181.

I added separate issues for the remaining (low priority) points in this issue:

• Parallelize binomial implementation (#188)
• Implement device side RNG buffer (#189)
• Remove change in brian2.diff once we update frozen Brian2 (#187)

And there is one case complicated test that was not reproducible with the same brian.seed() in CUDA Standalone but was in C++ Standalone. Need to investigate if this is due to parallelisation or a bug. See #190.

from brian2cuda.