すのはら荘春原庄的雪

[一生一芯笔记]Chisel diplomacy精讲

Toretto·2024-05-05 00:07·430 次阅读

[一生一芯笔记]Chisel diplomacy精讲

前言#

在做“一生一芯”的时候,碰见一个学习坡度陡峭,而又无法避开的一点:diplomacy

这是一个包含在rocket-chip中的工具,首先如何导入就是一个难题;其次,diplomac其使用了非常多的scala高级语法,这需要对语言有一定的熟悉度。

根据过往经历来看,我敢肯定在我学会后再回过头看这个问题肯定是较为简单,也无法理解新手在这方面的疑惑。

故在学习时同步记下这篇文章,以希望留下一些记录,待以后查阅、后人借鉴。

————2024.5.4

准备思路#

准备方面,我使用了ysyx余博的工程,可以较好地本地导入rocket-chip的包

省去了写mill的烦恼

用的是chisel7

然后看代码(代码+语法)#

以这份翻译(官方的样例工程)作为展开https://shili2017.github.io/posts/CHISEL1/

先粘一份完整代码,然后听我慢慢解析。这份代码是对import的有些改动的。

Copy
package ysyx import chisel3._ import chisel3.experimental.SourceInfo import chisel3.util.random.FibonacciLFSR import circt.stage.ChiselStage import org.chipsalliance.cde.config.Parameters import chisel3._ import org.chipsalliance.cde.config.Parameters import freechips.rocketchip.system.DefaultConfig import freechips.rocketchip.diplomacy._ case class UpwardParam(width: Int) case class DownwardParam(width: Int) case class EdgeParam(width: Int) // PARAMETER TYPES: D U E B object AdderNodeImp extends SimpleNodeImp[DownwardParam, UpwardParam, EdgeParam, UInt] { def edge(pd: DownwardParam, pu: UpwardParam, p: Parameters, sourceInfo: SourceInfo) = { if (pd.width < pu.width) EdgeParam(pd.width) else EdgeParam(pu.width) } def bundle(e: EdgeParam) = UInt(e.width.W) def render(e: EdgeParam) = RenderedEdge("blue", s"width = ${e.width}") } /** node for [[AdderDriver]] (source) */ class AdderDriverNode(widths: Seq[DownwardParam])(implicit valName: ValName) extends SourceNode(AdderNodeImp)(widths) /** node for [[AdderMonitor]] (sink) */ class AdderMonitorNode(width: UpwardParam)(implicit valName: ValName) extends SinkNode(AdderNodeImp)(Seq(width)) /** node for [[Adder]] (nexus) */ class AdderNode(dFn: Seq[DownwardParam] => DownwardParam, uFn: Seq[UpwardParam] => UpwardParam)(implicit valName: ValName) extends NexusNode(AdderNodeImp)(dFn, uFn) /** adder DUT (nexus) */ class Adder(implicit p: Parameters) extends LazyModule { val node = new AdderNode ( { case dps: Seq[DownwardParam] => require(dps.forall(dp => dp.width == dps.head.width), "inward, downward adder widths must be equivalent") dps.head }, { case ups: Seq[UpwardParam] => require(ups.forall(up => up.width == ups.head.width), "outward, upward adder widths must be equivalent") ups.head } ) lazy val module = new LazyModuleImp(this) { require(node.in.size >= 2) node.out.head._1 := node.in.unzip._1.reduce(_ + _) } override lazy val desiredName = "Adder" } /** driver (source) * drives one random number on multiple outputs */ class AdderDriver(width: Int, numOutputs: Int)(implicit p: Parameters) extends LazyModule { val node = new AdderDriverNode(Seq.fill(numOutputs)(DownwardParam(width))) lazy val module = new LazyModuleImp(this) { // check that node parameters converge after negotiation val negotiatedWidths = node.edges.out.map(_.width) require(negotiatedWidths.forall(_ == negotiatedWidths.head), "outputs must all have agreed on same width") val finalWidth = negotiatedWidths.head // generate random addend (notice the use of the negotiated width) val randomAddend = FibonacciLFSR.maxPeriod(finalWidth) // drive signals node.out.foreach { case (addend, _) => addend := randomAddend } } override lazy val desiredName = "AdderDriver" } /** monitor (sink) */ class AdderMonitor(width: Int, numOperands: Int)(implicit p: Parameters) extends LazyModule { val nodeSeq = Seq.fill(numOperands) { new AdderMonitorNode(UpwardParam(width)) } val nodeSum = new AdderMonitorNode(UpwardParam(width)) lazy val module = new LazyModuleImp(this) { val io = IO(new Bundle { val error = Output(Bool()) }) // print operation printf(nodeSeq.map(node => p"${node.in.head._1}").reduce(_ + p" + " + _) + p" = ${nodeSum.in.head._1}") // basic correctness checking io.error := nodeSum.in.head._1 =/= nodeSeq.map(_.in.head._1).reduce(_ + _) } override lazy val desiredName = "AdderMonitor" } /** top-level connector */ class AdderTestHarness()(implicit p: Parameters) extends LazyModule { val numOperands = 2 val adder = LazyModule(new Adder) // 8 will be the downward-traveling widths from our drivers val drivers = Seq.fill(numOperands) { LazyModule(new AdderDriver(width = 8, numOutputs = 2)) } // 4 will be the upward-traveling width from our monitor val monitor = LazyModule(new AdderMonitor(width = 4, numOperands = numOperands)) // create edges via binding operators between nodes in order to define a complete graph drivers.foreach{ driver => adder.node := driver.node } drivers.zip(monitor.nodeSeq).foreach { case (driver, monitorNode) => monitorNode := driver.node } monitor.nodeSum := adder.node lazy val module = new LazyModuleImp(this) { // when(monitor.module.io.error) { // printf("something went wrong") // } } override lazy val desiredName = "AdderTestHarness" } object Elaborate extends App { // (new ChiselStage).execute(args, Seq(chisel3.stage.ChiselGeneratorAnnotation( // () => LazyModule(new AdderTestHarness()(Parameters.empty)).module)) // ) val verilog = ChiselStage.emitSystemVerilog( LazyModule(new AdderTestHarness()(Parameters.empty)).module ) println(verilog) }

参数协商和传递#

参数#

Copy
case class UpwardParam(width: Int) case class DownwardParam(width: Int) case class EdgeParam(width: Int)

看到这段代码,有一个问题,这个case class有什么用?

case class

case class是一种特殊类型的类

case class = class + 一坨

Copy
case class Person(name: String, age: Int)

等价于

Copy
class Person(val name: String, val age: Int) { override def toString = s"Person(name=$name, age=$age)" override def equals(other: Any) = other match { case that: Person => this.name == that.name && this.age == that.age case _ => false } override def hashCode() = scala.util.hashing.MurmurHash3.productHash(this) } object Person { def apply(name: String, age: Int) = new Person(name, age) def unapply(p: Person): Option[(String, Int)] = Some((p.name, p.age)) }

注意case class这里的参数列表,默认情况下,case clas的构造参数会转换成val类型的字段

节点#

在节点实现(即NodeImp中),我们描述参数如何在我们的图中流动,以及如何在节点之间协商参数。边参数(E)描述了需要在边上传递的数据类型,在这个例子中就是Int;捆绑参数(B)描述了模块之间硬件实现的参数化端口的数据类型,在这个例子中则为UInt。此处edge函数实际执行了节点之间的参数协商,比较了向上和向下传播的参数,并选择数据宽度较小的那个作为协商结果。

Copy
// PARAMETER TYPES: D U E B object AdderNodeImp extends SimpleNodeImp[DownwardParam, UpwardParam, EdgeParam, UInt] { def edge(pd: DownwardParam, pu: UpwardParam, p: Parameters, sourceInfo: SourceInfo) = { if (pd.width < pu.width) EdgeParam(pd.width) else EdgeParam(pu.width) } def bundle(e: EdgeParam) = UInt(e.width.W) def render(e: EdgeParam) = RenderedEdge("blue", s"width = ${e.width}") }
Copy
/** node for [[AdderDriver]] (source) */ class AdderDriverNode(widths: Seq[DownwardParam])(implicit valName: ValName) extends SourceNode(AdderNodeImp)(widths) /** node for [[AdderMonitor]] (sink) */ class AdderMonitorNode(width: UpwardParam)(implicit valName: ValName) extends SinkNode(AdderNodeImp)(Seq(width)) /** node for [[Adder]] (nexus) */ class AdderNode(dFn: Seq[DownwardParam] => DownwardParam, uFn: Seq[UpwardParam] => UpwardParam)(implicit valName: ValName) extends NexusNode(AdderNodeImp)(dFn, uFn)

创建LazyModule#

lazy

scala中,lazy表示的是使用时初始化
另外,懒惰初始化可以应用于val和def(虽然def默认就是懒惰的,但懒惰val和def在语义上有所不同,val初始化后值不>变,而def每次调用都可能有不同结果)。

虽然从懒惰初始化的角度看,lazy val和没有具体实现的def看起来相似,但它们之间存在本质区别:

  • lazy val在首次访问后会缓存其结果,之后的访问直接返回缓存的值,适用于计算密集型或资源加载场景。
  • def则是每次调用时都执行其函数体,适合于那些结果随时间或上下文变化的情况。
implicit

这个概念比较庞杂

  • implicit method:类型转换
    Copy
    implicit def intToDouble(i: Int): Double = i.toDouble def processNumber(num: Double): Unit = println(num) processNumber(5) // 由于存在隐式转换,这里可以传入Int类型
  • implicit param:默认行为、实现策略模式或依赖注入
    Copy
    case class LogLevel(level: String) def log(message: String)(implicit level: LogLevel = LogLevel("INFO")) = println(s"${level.level}: $message") log("This is an info message") // 使用默认的日志级别 implicit val debugLevel = LogLevel("DEBUG") log("This is a debug message") // 使用隐式提供的DEBUG级别
  • implicit class:
    Copy
    implicit class RichString(s: String) { def lengthSquared: Int = s.length * s.length } val str = "Hello" println(str.lengthSquared) // 利用隐式转换调用新方法

Lazy的意思是指将表达式的evaluation推迟到需要的时候。在创建Diplomacy图之后,参数协商是lazy完成的,因此我们想要参数化的硬件也必须延迟生成,因此需要使用LazyModule。需要注意的是,定义Diplomacy图的组件(在这个例子里为节点)的创建不是lazy的,模块硬件需要写在LazyModuleImp。

在这个例子中,我们希望driver将相同位宽的数据输入到加法器中,monitor的数据来自加法器的输出以及driver,所有这些数据位宽都应该相同。我们可以通过AdderNode的require来限制这些参数,将DownwardParam向下传递,以及将UpwardParam向上传递。

Copy
/** adder DUT (nexus) */ class Adder(implicit p: Parameters) extends LazyModule { val node = new AdderNode ( { case dps: Seq[DownwardParam] => require(dps.forall(dp => dp.width == dps.head.width), "inward, downward adder widths must be equivalent") dps.head }, { case ups: Seq[UpwardParam] => require(ups.forall(up => up.width == ups.head.width), "outward, upward adder widths must be equivalent") ups.head } ) lazy val module = new LazyModuleImp(this) { require(node.in.size >= 2) node.out.head._1 := node.in.unzip._1.reduce(_ + _) } override lazy val desiredName = "Adder" }
Partial Function

好了,又看不懂了

Copy
{ case dps: Seq[DownwardParam] => require(dps.forall(dp => dp.width == dps.head.width), "inward, downward adder widths must be equivalent") dps.head }

怎么传参的时候就这么一个东西就作为参数了呢

Copy
class AdderNode(dFn: Seq[DownwardParam] => DownwardParam, uFn: Seq[UpwardParam] => UpwardParam)

可以看到入参是一个传名函数(传名函数是什么?还是百度吧)
也就是说,那一坨花括号是一个函数

{ case ... }的结构实际上是在定义一个部分函数(PartialFunction),它是一种特殊的函数类型,经常与模式匹配一起使用。部分函数可以理解为一个仅定义了部分情况(cases)的函数,对于未定义的情况,如果尝试调用则会抛出异常。

  • 让我们直接以更简单的例子说明部分函数的用法:

简单例子:定义一个处理整数的匿名部分函数

Copy
val processNumbers: PartialFunction[Int, String] = { case x if x > 0 => s"$x is positive" case 0 => "Zero" } processNumbers(5) // 输出: "5 is positive" processNumbers(0) // 输出: "Zero" // processNumbers(-1) // 如果尝试调用,会抛出MatchError异常

在这个例子中,processNumbers是一个PartialFunction[Int, String],它只定义了两个情况:当输入的整数大于0时和等于0时的处理逻辑。如果尝试传入一个负数,由于没有对应的case分支,Scala会抛出MatchError。

回到原始代码片段:

Copy
{ case dps: Seq[DownwardParam] => require(dps.forall(dp => dp.width == dps.head.width), "inward, downward adder widths must be equivalent") dps.head }

这里定义的就是这样一个部分函数,它只匹配Seq[DownwardParam]类型的输入,执行一系列操作后返回dps.head。虽然没有直接写出match关键字,但这种结构实质上是在做模式匹配的工作,是Scala中一种优雅的处理不同类型或情况的函数定义方式。

initializer block

又看不懂了,构造函数垢面后

Copy
lazy val module = new LazyModuleImp(this) { require(node.in.size >= 2 node.out.head._1 := no.in.unzip._1.reduce(_ + _) }

在Scala中,当你在创建一个类的实例并立即跟随一个大括号 { ... } 时,这个大括号内的代码块实际上是该类构造函数的一部分,被称为初始化块(initializer block)。初始化块会在类的构造函数执行完毕后立即执行,常用于进行进一步的初始化设置或者运行一些初始化逻辑。初始化块可以访问到类的所有成员,包括由构造函数参数初始化的成员。

Copy
class Person(val name:String){ println(s"1.class Person($name)") val age=10 } val p=new Person("as"){ println(s"2.obj p ($name,$age)") } // 1.class Person(as) // 2.obj p (as,10)

三要素:driver(驱动)、dut(功能模块)、monitor(检查)中已经完成了dut的编写,接下来是driver和monitor

AdderDriver随机生成位宽为finalWidth的数据,并传递到numOutputs个source。

Copy
/** driver (source) * drives one random number on multiple outputs */ class AdderDriver(width: Int, numOutputs: Int)(implicit p: Parameters) extends LazyModule { val node = new AdderDriverNode(Seq.fill(numOutputs)(DownwardParam(width))) lazy val module = new LazyModuleImp(this) { // check that node parameters converge after negotiation val negotiatedWidths = node.edges.out.map(_.width) require(negotiatedWidths.forall(_ == negotiatedWidths.head), "outputs must all have agreed on same width") val finalWidth = negotiatedWidths.head // generate random addend (notice the use of the negotiated width) val randomAddend = FibonacciLFSR.maxPeriod(finalWidth) // drive signals node.out.foreach { case (addend, _) => addend := randomAddend } } override lazy val desiredName = "AdderDriver" }

AdderMonitor打印加法器输出并检测错误,有两个AdderMonitorNode节点从AdderDriver接收加法的两个输入,以及一个AdderMonitorNode节点从加法器接收加法的输出。

Copy
/** monitor (sink) */ class AdderMonitor(width: Int, numOperands: Int)(implicit p: Parameters) extends LazyModule { val nodeSeq = Seq.fill(numOperands) { new AdderMonitorNode(UpwardParam(width)) } val nodeSum = new AdderMonitorNode(UpwardParam(width)) lazy val module = new LazyModuleImp(this) { val io = IO(new Bundle { val error = Output(Bool()) }) // print operation printf(nodeSeq.map(node => p"${node.in.head._1}").reduce(_ + p" + " + _) + p" = ${nodeSum.in.head._1}") // basic correctness checking io.error := nodeSum.in.head._1 =/= nodeSeq.map(_.in.head._1).reduce(_ + _) } override lazy val desiredName = "AdderMonitor" }
posted @   msuad  阅读(430)  评论(1编辑  收藏  举报
相关博文:
阅读排行:
· 一个费力不讨好的项目,让我损失了近一半的绩效!
· 清华大学推出第四讲使用 DeepSeek + DeepResearch 让科研像聊天一样简单!
· 实操Deepseek接入个人知识库
· CSnakes vs Python.NET:高效嵌入与灵活互通的跨语言方案对比
· Plotly.NET 一个为 .NET 打造的强大开源交互式图表库
点击右上角即可分享
微信分享提示
目录