Proper Control Flow for FIR filter

[mattfel1]

 def FIR_Filter(args: Array[String]) {
   val input = StreamIn[Int](target.In)
   val output = StreamOut[Int](target.Out)
   val weights = DRAM[Int](32)
   val width = ArgIn[Int]
   val P = 16 (1,1,32)

   // Initialize width with the first console argument
   setArg(width, min(32, args(0).to[Int]) )

   // Transfer weights from the host to accelerator
   sendArray(weights, loadData[Int]("weights.csv"))

   Accel(*) {
     val wts = RegFile[Int](32)
     val ins = RegFile[Int](32)
     val sum = Reg[Int]

     // Load weights from DRAM into local registers
     wts load weights(0::width)

     // Stream continuously
     Stream(*) {
     // Shift in the most recent input
     ins <<= input

     // Create a reduce-accumulate tree with P inputs
     Reduce(sum)(0 until width par P){i =>
       // Multiply corresponding weight and input
       wts(i) * ins(i)
     }{(a,b) => a + b }

     // Assign the result of computing the average
     // to the output stream
     output := sum / width
   }
 }

To make this actually correct, we need a Pipe wrapped just inside the Stream(*) because the Reduce and output := pipes will actually just run all the time no matter what. Or we need stream-aware memory structures for this kind of thing, like shift-acknowledgements and register staleness