atomicMax how to implement it?

[mratsim]

Here are a few comments:

  1. Computing max via atomics is non-trivial. Reserving it for the `epilogue` 
is a good idea.
  2. `proc comp_max() : float = rand(1.0)` is problematic, `rand` uses a global 
mutable RNG state underneath, you cannot access that via multiple thread as it 
will at best require locking and unlocking the RNG at every generation, 
rendering multithreading useless, and at worst, multiple threads will compete 
and corrupt the RNG state.



Here is a version using a parallel reducer. It requires a special import as I'm 
not sure about the API yet.
    
    
    import random, weave, weave/parallel_reduce
    
    proc maxReduce(n: int): float64 =
      var waitableMax: Flowvar[float64]
      
      parallelReduceImpl i in 0 .. n, stride = 1:
        reduce(waitableMax):
          prologue:
            var threadLocalRNG = initRand(1234)
            var localMax = -Inf
          fold:
            # Note all thread will start with the same RNG sequence.
            # Multithreaded RNG is a non-trivial problem, so showing a simple 
example
            # since the question is about max
            localMax = max(localMax, threadLocalRNG.rand(1.0))
          merge(remoteMax):
            localMax = max(localMax, sync(remoteMax))
          return localMax
      
      result = sync(waitableMax)
    
    init(Weave)
    let max1M = maxReduce(1000000)
    echo "maxReduce(1000000): ", max1M
    exit(Weave)
    
    
    Run

Version(s) with parallelForStaged coming next.