backend/fu/Multiplier.scala

package xiangshan.backend.fu

import chisel3._
import chisel3.util._
import xiangshan._
import utils._
import xiangshan.backend.fu.util.{C22, C32, C53}

class MulDivCtrl extends Bundle{
  val sign = Bool()
  val isW = Bool()
  val isHi = Bool() // return hi bits of result ?
}

class AbstractMultiplier(len: Int) extends FunctionUnit(
  len
){
  val ctrl = IO(Input(new MulDivCtrl))
}

class NaiveMultiplier(len: Int, val latency: Int)
  extends AbstractMultiplier(len)
  with HasPipelineReg
{

  val (src1, src2) = (io.in.bits.src(0), io.in.bits.src(1))

  val mulRes = src1.asSInt() * src2.asSInt()

  var dataVec = Seq(mulRes.asUInt())
  var ctrlVec = Seq(ctrl)

  for(i <- 1 to latency){
    dataVec = dataVec :+ PipelineReg(i)(dataVec(i-1))
    ctrlVec = ctrlVec :+ PipelineReg(i)(ctrlVec(i-1))
  }

  val xlen = io.out.bits.data.getWidth
  val res = Mux(ctrlVec.last.isHi, dataVec.last(2*xlen-1, xlen), dataVec.last(xlen-1,0))
  io.out.bits.data := Mux(ctrlVec.last.isW, SignExt(res(31,0),xlen), res)

  XSDebug(p"validVec:${Binary(Cat(validVec))} flushVec:${Binary(Cat(flushVec))}\n")
}

class ArrayMulDataModule(len: Int, doReg: Seq[Int]) extends XSModule {
  val io = IO(new Bundle() {
    val a, b = Input(UInt(len.W))
    val result = Output(UInt((2 * len).W))
  })
  val (a, b) = (io.a, io.b)
  val doRegSorted = doReg.sortWith(_ < _)

  val b_sext, bx2, neg_b, neg_bx2 = Wire(UInt((len+1).W))
  b_sext := SignExt(b, len+1)
  bx2 := b_sext << 1
  neg_b := (~b_sext).asUInt()
  neg_bx2 := neg_b << 1

  val columns: Array[Seq[Bool]] = Array.fill(2*len)(Seq())

  var last_x = WireInit(0.U(3.W))
  for(i <- Range(0, len, 2)){
    val x = if(i==0) Cat(a(1,0), 0.U(1.W)) else if(i+1==len) SignExt(a(i, i-1), 3) else a(i+1, i-1)
    val pp_temp = MuxLookup(x, 0.U, Seq(
      1.U -> b_sext,
      2.U -> b_sext,
      3.U -> bx2,
      4.U -> neg_bx2,
      5.U -> neg_b,
      6.U -> neg_b
    ))
    val s = pp_temp(len)
    val t = MuxLookup(last_x, 0.U(2.W), Seq(
      4.U -> 2.U(2.W),
      5.U -> 1.U(2.W),
      6.U -> 1.U(2.W)
    ))
    last_x = x
    val (pp, weight) = i match {
      case 0 =>
        (Cat(~s, s, s, pp_temp), 0)
      case n if (n==len-1) || (n==len-2) =>
        (Cat(~s, pp_temp, t), i-2)
      case _ =>
        (Cat(1.U(1.W), ~s, pp_temp, t), i-2)
    }
    for(j <- columns.indices){
      if(j >= weight && j < (weight + pp.getWidth)){
        columns(j) = columns(j) :+ pp(j-weight)
      }
    }
  }

  def addOneColumn(col: Seq[Bool], cin: Seq[Bool]): (Seq[Bool], Seq[Bool], Seq[Bool]) = {
    var sum = Seq[Bool]()
    var cout1 = Seq[Bool]()
    var cout2 = Seq[Bool]()
    col.size match {
      case 1 =>  // do nothing
        sum = col ++ cin
      case 2 =>
        val c22 = Module(new C22)
        c22.io.in := col
        sum = c22.io.out(0).asBool() +: cin
        cout2 = Seq(c22.io.out(1).asBool())
      case 3 =>
        val c32 = Module(new C32)
        c32.io.in := col
        sum = c32.io.out(0).asBool() +: cin
        cout2 = Seq(c32.io.out(1).asBool())
      case 4 =>
        val c53 = Module(new C53)
        for((x, y) <- c53.io.in.take(4) zip col){
          x := y
        }
        c53.io.in.last := (if(cin.nonEmpty) cin.head else 0.U)
        sum = Seq(c53.io.out(0).asBool()) ++ (if(cin.nonEmpty) cin.drop(1) else Nil)
        cout1 = Seq(c53.io.out(1).asBool())
        cout2 = Seq(c53.io.out(2).asBool())
      case n =>
        val cin_1 = if(cin.nonEmpty) Seq(cin.head) else Nil
        val cin_2 = if(cin.nonEmpty) cin.drop(1) else Nil
        val (s_1, c_1_1, c_1_2) = addOneColumn(col take 4, cin_1)
        val (s_2, c_2_1, c_2_2) = addOneColumn(col drop 4, cin_2)
        sum = s_1 ++ s_2
        cout1 = c_1_1 ++ c_2_1
        cout2 = c_1_2 ++ c_2_2
    }
    (sum, cout1, cout2)
  }

  def max(in: Iterable[Int]): Int = in.reduce((a, b) => if(a>b) a else b)
  def addAll(cols: Array[Seq[Bool]], depth: Int): (UInt, UInt) = {
    if(max(cols.map(_.size)) <= 2){
      val sum = Cat(cols.map(_(0)).reverse)
      var k = 0
      while(cols(k).size == 1) k = k+1
      val carry = Cat(cols.drop(k).map(_(1)).reverse)
      (sum, Cat(carry, 0.U(k.W)))
    } else {
      val columns_next = Array.fill(2*len)(Seq[Bool]())
      var cout1, cout2 = Seq[Bool]()
      for( i <- cols.indices){
        val (s, c1, c2) = addOneColumn(cols(i), cout1)
        columns_next(i) = s ++ cout2
        cout1 = c1
        cout2 = c2
      }

      val needReg = doRegSorted.contains(depth)
      val toNextLayer = if(needReg)
        columns_next.map(_.map(RegNext(_)))
      else
        columns_next

      addAll(toNextLayer, depth+1)
    }
  }

  val (sum, carry) = addAll(cols = columns, depth = 0)
  io.result := sum + carry
}

class ArrayMultiplier(len: Int, doReg: Seq[Int]) extends AbstractMultiplier(len) with HasPipelineReg {

  override def latency = doReg.size

  val mulDataModule = Module(new ArrayMulDataModule(len, doReg))
  mulDataModule.io.a := io.in.bits.src(0)
  mulDataModule.io.b := io.in.bits.src(1)
  val result = mulDataModule.io.result

  var ctrlVec = Seq(ctrl)
  for(i <- 1 to latency){
    ctrlVec = ctrlVec :+ PipelineReg(i)(ctrlVec(i-1))
  }
  val xlen = len - 1
  val res = Mux(ctrlVec.last.isHi, result(2*xlen-1, xlen), result(xlen-1,0))

  io.out.bits.data := Mux(ctrlVec.last.isW, SignExt(res(31,0),xlen), res)

  XSDebug(p"validVec:${Binary(Cat(validVec))} flushVec:${Binary(Cat(flushVec))}\n")
}