Unused array computations lead to memory leaks

[tmcdonell]

The CUDA backend uses explicit reference tracking for device arrays. However, if the result of a computation is not used, it may not be released until the entire CUDA environment is closed down. This is evident for computations under Let2, but may exist more generally. In the example below, we modify the runCUDA function display any active memory records on finalisation.

test_scan = do
  let xs :: Acc (Vector Int) = use $ fromList 10 [0..]
      expr = Acc.scanl (+) 0 xs

  putStrLn "scan"
  print =<< CUDA.run (Prelude.fst expr)

  putStrLn "reduction"
  print =<< CUDA.run (Prelude.snd expr)

Results in:

*Test> test_scan
scan
 91488264:=MemoryEntry {_refcount = 0, _memsize = 4, _arena = 1049088}

Array 10 [0,0,1,3,6,10,15,21,28,36]
reduction
 118558728:=MemoryEntry {_refcount = 0, _memsize = 40, _arena = 1048832}

Array () [45]

[tmcdonell]

This should be fixed now, and will now generate a real error if this is not the case instead of silently correcting the problem.