class mkSystolicMatrixMultiplicationCell(w:Int)
            extends Module("systolic_mmu_pe_"+w, "clk", "reset"){
  val ain = inP("inA", w); val ain_valid = inP("inA_valid")
  val bin = inP("inB", w); val bin_valid = inP("inB_valid")

  val aout = outP("outA", w); val aout_valid = outP("outA_valid")
  val bout = outP("outB", w); val bout_valid = outP("outB_valid")

  val cout = outP("outC", 2*w)

  aout $ sysClk := ain // 入力aを1クロック後に出力aに
  bout $ sysClk := bin // 入力bを1クロック後に出力bに
  aout_valid $ sysClk := ain_valid
  bout_valid $ sysClk := bin_valid

  val c = signal(2*w); cout := c // 結果格納レジスタを用意．coutに接続
  val ab = signal(2*w); ab := ain * bin
  // リセットで0クリア，入力が有効ならMAC演算を実行，無効ならそのまま
  c $ sysClk := ?(sysReset == HIGH, value(0, 2*w),
                ?((ain_valid and bin_valid) == HIGH, c + ab,
                c))
}

  val cell = new mkSystolicMatrixMultiplicationCell(w)
  val PEs = for(i <- 0 until n*n) yield instance(cell, "U_"+i)

  for(r <- 0 until n){
    for(c <- 0 until n){
      val i = r*n+c
      // horizontal connections
      if(c == 0){ // from input
        PEs(i).signalFor(cell.ain) := a_inputs(r)
        PEs(i).signalFor(cell.ain_valid) := a_valids(r)
      }else{ // from the above cell
        PEs(i).signalFor(cell.ain) := PEs(i-1).signalFor(cell.aout)
        PEs(i).signalFor(cell.ain_valid) := PEs(i-1).signalFor(cell.aout_valid)
      }
      // vertical connections
      if(r == 0){ // from input
        PEs(i).signalFor(cell.bin) := b_inputs(c)
        PEs(i).signalFor(cell.bin_valid) := b_valids(c)
      }else{ // from the left cell
        PEs(i).signalFor(cell.bin) := PEs(i-n).signalFor(cell.bout)
        PEs(i).signalFor(cell.bin_valid) := PEs(i-n).signalFor(cell.bout_valid)
      }
      c_outputs(i) := PEs(i).signalFor(cell.cout)
    }
  }