Question

我将火箭芯片util模块'Arbiters.scala'复制到单独的工作目录，并使用以下代码进行测试：

object try_arbiter extends App {
  chisel3.Driver.execute(args, () => new HellaCountingArbiter(UInt(4.W), 3, 5))
}

编译时没有问题。但是，在“精心设计...”步骤中，它报告了[error] chisel3.core.Binding$ExpectedHardwareException: mux condition 'chisel3.core.Bool@46' must be hardware, not a bare Chisel type"

相关代码是

PriorityEncoder(io.in.map(_.valid))

当我将此行更改为下一行时，错误消失了（scala代码仍然有这种错误，但在这一行中没有）。

PriorityEncoder(io.in.map(x => Wire(x.valid)))

火箭芯片代码应该经过多次评估，对吗？我想我一定错过了……一些配置？

谢谢您的提示！

附录（Arbiters.scala）：

package rocket_examples

import chisel3._
import chisel3.util._

/** A generalized locking RR arbiter that addresses the limitations of the
 *  version in the Chisel standard library */
abstract class HellaLockingArbiter[T <: Data](typ: T, arbN: Int, rr: Boolean = false)
    extends Module {

  val io = new Bundle {
    val in = Vec(arbN, Decoupled(typ.cloneType)).flip
    val out = Decoupled(typ.cloneType)
  }

  def rotateLeft[T <: Data](norm: Vec[T], rot: UInt): Vec[T] = {
    val n = norm.size
    Vec.tabulate(n) { i =>
      Mux(rot < UInt(n - i), norm(UInt(i) + rot), norm(rot - UInt(n - i)))
    }
  }

  val lockIdx = Reg(init = UInt(0, log2Up(arbN)))
  val locked = Reg(init = Bool(false))

  val choice = if (rr) {
    PriorityMux(
      rotateLeft(Vec(io.in.map(_.valid)), lockIdx + UInt(1)),
      rotateLeft(Vec((0 until arbN).map(UInt(_))), lockIdx + UInt(1)))
  } else {
    PriorityEncoder(io.in.map(_.valid))
  }

  val chosen = Mux(locked, lockIdx, choice)

  for (i <- 0 until arbN) {
    io.in(i).ready := io.out.ready && chosen === UInt(i)
  }

  io.out.valid := io.in(chosen).valid
  io.out.bits := io.in(chosen).bits
}

/** This locking arbiter determines when it is safe to unlock
 *  by peeking at the data */
class HellaPeekingArbiter[T <: Data](
      typ: T, arbN: Int,
      canUnlock: T => Bool,
      needsLock: Option[T => Bool] = None,
      rr: Boolean = false)
    extends HellaLockingArbiter(typ, arbN, rr) {

  def realNeedsLock(data: T): Bool =
    needsLock.map(_(data)).getOrElse(Bool(true))

  when (io.out.fire()) {
    when (!locked && realNeedsLock(io.out.bits)) {
      lockIdx := choice
      locked := Bool(true)
    }
    // the unlock statement takes precedent
    when (canUnlock(io.out.bits)) {
      locked := Bool(false)
    }
  }
}

/** This arbiter determines when it is safe to unlock by counting transactions */
class HellaCountingArbiter[T <: Data](
      typ: T, arbN: Int, count: Int,
      val needsLock: Option[T => Bool] = None,
      rr: Boolean = false)
    extends HellaLockingArbiter(typ, arbN, rr) {

  def realNeedsLock(data: T): Bool =
    needsLock.map(_(data)).getOrElse(Bool(true))

  // if count is 1, you should use a non-locking arbiter
  require(count > 1, "CountingArbiter cannot have count <= 1")

  val lock_ctr = Counter(count)

  when (io.out.fire()) {
    when (!locked && realNeedsLock(io.out.bits)) {
      lockIdx := choice
      locked := Bool(true)
      lock_ctr.inc()
    }

    when (locked) {
      when (lock_ctr.inc()) { locked := Bool(false) }
    }
  }
}

Answer 1

问题在于，火箭芯片中的几乎所有代码都是针对较旧的Chisel2风格的API编写的，应使用兼容性包装器import Chisel._进行编译。我看到您使用了import chisel3._，它的语义有些严格。

对于这种特定情况，Chisel2和Chisel3之间的区别在于端口（val io）必须包装在IO(...)中，即

  val io = IO(new Bundle {
    val in = Flipped(Vec(arbN, Decoupled(typ)))
    val out = Decoupled(typ)
  })

请注意，我还将Vec(arbN, Decoupled(typ.cloneType)).flip更改为Flipped(Vec(arbN, Decoupled(typ)))，并删除了.cloneType上的val out。对于后者，不需要进行后两项更改，但是从Chisel 3.1.2开始，它们将被标记为弃用警告。

测试火箭芯片工具“ Arbiters.scala”，出现错误“操作的位...必须是硬件...”

1 个答案: