Removed time consuming performGC call

[pmlodawski]

Hi!
Regarding to AccelerateHS/accelerate#227 and my answer here http://stackoverflow.com/a/28084927/1641629 accelerate is working slow under ghci because of performGC call, which from my tests takes more than 130ms. I think we can relay on ghc garbage collector to clean memory on its own and we should not force it to do it on cleanup.

[mchakravarty]

Good catch — thank you!

IIRC the performGC call is there, because we currently don't have any dedicated GC for GPU arrays, but rely on finalisers attached to CPU arrays to deallocate the associated GPU arrays. So, the performGC tries to ensure that memory is freed as quickly as possible. This strategy to managing GPU memory has, however, proved to be insufficient anyway and @robeverest is currently working on a proper solution to reclaiming GPU memory — see AccelerateHS/accelerate#240.

With @robeverest's improvements, the performGC is definitely a waste of effort.

@robeverest Could you maybe merge this pull request into your code for adding GPU GC?

[robeverest]

Sorry about the delayed reply on this. I've just added support for a dedicated GC for the GPU that doesn't depend on GHC's GC. I would still consider it experimental at this stage, but it's safe to remove the performGC. This pull request can be merged.