axi4: make maxFlight a per-master parameter

src/main/scala/uncore/axi4/IdIndexer.scala

@@ -13,11 +13,26 @@ class AXI4IdIndexer(idBits: Int)(implicit p: Parameters) extends LazyModule
   require (idBits >= 0)
 
   val node = AXI4AdapterNode(
-    masterFn = { mp => mp.copy(
+    masterFn = { mp =>
+      // Create one new "master" per ID
+      val masters = Array.tabulate(1 << idBits) { i => AXI4MasterParameters(
+         id        = IdRange(i, i+1),
+         aligned   = true,
+         maxFlight = Some(0))
+      }
+      // Squash the information from original masters into new ID masters
+      mp.masters.foreach { m =>
+        for (i <- m.id.start until m.id.end) {
+          val j = i % (1 << idBits)
+          val old = masters(j)
+          masters(j) = old.copy(
+            aligned   = old.aligned && m.aligned,
+            maxFlight = old.maxFlight.flatMap { o => m.maxFlight.map { n => o+n } })
+        }
+      }
+      mp.copy(
         userBits = mp.userBits + max(0, log2Ceil(mp.endId) - idBits),
-      masters  = Seq(AXI4MasterParameters(
+        masters  = masters)
-        id      = IdRange(0, min(mp.endId, 1 << idBits)),
-        aligned = mp.masters.map(_.aligned).reduce(_ && _))))
     },
     slaveFn = { sp => sp.copy(
       slaves = sp.slaves.map(s => s.copy(

src/main/scala/uncore/axi4/Parameters.scala

@@ -64,19 +64,19 @@ case class AXI4SlavePortParameters(
 case class AXI4MasterParameters(
   id:        IdRange       = IdRange(0, 1),
   aligned:   Boolean       = false,
+  maxFlight: Option[Int]   = None, // None = infinite, else is a per-ID cap
   nodePath:  Seq[BaseNode] = Seq())
 {
   val name = nodePath.lastOption.map(_.lazyModule.name).getOrElse("disconnected")
+  maxFlight.foreach { m => require (m >= 0) }
 }
 
 case class AXI4MasterPortParameters(
   masters:   Seq[AXI4MasterParameters],
-  userBits:  Int = 0,
+  userBits:  Int = 0)
-  maxFlight: Int = 0) // at most X transactions per ID (0 = unlimited)
 {
   val endId = masters.map(_.id.end).max
   require (userBits >= 0)
-  require (maxFlight >= 0)
 
   // Require disjoint ranges for ids
   masters.combinations(2).foreach { case Seq(x,y) => require (!x.id.overlaps(y.id), s"$x and $y overlap") }

src/main/scala/uncore/axi4/ToTL.scala

@@ -9,8 +9,9 @@ import diplomacy._
 import uncore.tilelink2._
 
 case class AXI4ToTLNode() extends MixedAdapterNode(AXI4Imp, TLImp)(
-  dFn = { case AXI4MasterPortParameters(masters, userBits, maxFlight) =>
+  dFn = { case AXI4MasterPortParameters(masters, userBits) =>
-    require (maxFlight > 0, "AXI4 must include a maximum transactions per ID to convert to TL")
+    masters.foreach { m => require (m.maxFlight.isDefined, "AXI4 must include a transaction maximum per ID to convert to TL") }
+    val maxFlight = masters.map(_.maxFlight.get).max
     TLClientPortParameters(
       clients = masters.flatMap { m =>
         for (id <- m.id.start until m.id.end)
@@ -50,7 +51,7 @@ class AXI4ToTL()(implicit p: Parameters) extends LazyModule
       val numIds = edgeIn.master.endId
       val beatBytes = edgeOut.manager.beatBytes
       val countBits = AXI4Parameters.lenBits + (1 << AXI4Parameters.sizeBits) - 1
-      val maxFlight = edgeIn.master.maxFlight
+      val maxFlight = edgeIn.master.masters.map(_.maxFlight.get).max
       val addedBits = log2Ceil(maxFlight) + 1
 
       require (edgeIn.master.userBits == 0, "AXI4 user bits cannot be transported by TL")

src/main/scala/uncore/axi4/UserYanker.scala

@@ -10,12 +10,15 @@ import uncore.tilelink2.UIntToOH1
 
 class AXI4UserYanker(capMaxFlight: Option[Int] = None)(implicit p: Parameters) extends LazyModule
 {
-  // !!! make maxFlightPerId a cap and maxFlight a per AXI4 Master parameter
-  val maxFlightPerId = capMaxFlight.getOrElse(8)
-  require (maxFlightPerId >= 1)
-
   val node = AXI4AdapterNode(
-    masterFn = { mp => mp.copy(maxFlight = maxFlightPerId, userBits = 0) },
+    masterFn = { mp => mp.copy(
+      userBits = 0,
+      masters = mp.masters.map { m => m.copy(
+        maxFlight = (m.maxFlight, capMaxFlight) match {
+          case (Some(x), Some(y)) => Some(x min y)
+          case (Some(x), None)    => Some(x)
+          case (None,    Some(y)) => Some(y)
+          case (None,    None)    => None })})},
     slaveFn = { sp => sp })
 
   lazy val module = new LazyModuleImp(this) {
@@ -29,18 +32,31 @@ class AXI4UserYanker(capMaxFlight: Option[Int] = None)(implicit p: Parameters) e
       val need_bypass = edgeOut.slave.minLatency < 1
       require (bits > 0) // useless UserYanker!
 
-      val rqueues = Seq.fill(edgeIn.master.endId) { Module(new Queue(UInt(width = bits), maxFlightPerId, flow=need_bypass)) }
+      edgeOut.master.masters.foreach { m =>
-      val wqueues = Seq.fill(edgeIn.master.endId) { Module(new Queue(UInt(width = bits), maxFlightPerId, flow=need_bypass)) }
+        require (m.maxFlight.isDefined, "UserYanker needs a flight cap on each ID")
+      }
+
+      def queue(id: Int) = {
+        val depth = edgeOut.master.masters.find(_.id.contains(id)).flatMap(_.maxFlight).getOrElse(0)
+        if (depth == 0) {
+          Wire(new QueueIO(UInt(width = bits), 1)) // unused ID => undefined value
+        } else {
+          Module(new Queue(UInt(width = bits), depth, flow=need_bypass)).io
+        }
+      }
+
+      val rqueues = Seq.tabulate(edgeIn.master.endId) { i => queue(i) }
+      val wqueues = Seq.tabulate(edgeIn.master.endId) { i => queue(i) }
 
       val arid = in.ar.bits.id
-      val ar_ready = Vec(rqueues.map(_.io.enq.ready))(arid)
+      val ar_ready = Vec(rqueues.map(_.enq.ready))(arid)
       in .ar.ready := out.ar.ready && ar_ready
       out.ar.valid := in .ar.valid && ar_ready
       out.ar.bits  := in .ar.bits
 
       val rid = out.r.bits.id
-      val r_valid = Vec(rqueues.map(_.io.deq.valid))(rid)
+      val r_valid = Vec(rqueues.map(_.deq.valid))(rid)
-      val r_bits = Vec(rqueues.map(_.io.deq.bits))(rid)
+      val r_bits = Vec(rqueues.map(_.deq.bits))(rid)
       assert (!out.r.valid || r_valid) // Q must be ready faster than the response
       in.r <> out.r
       in.r.bits.user.get := r_bits
@@ -48,20 +64,20 @@ class AXI4UserYanker(capMaxFlight: Option[Int] = None)(implicit p: Parameters) e
       val arsel = UIntToOH(arid, edgeIn.master.endId).toBools
       val rsel  = UIntToOH(rid,  edgeIn.master.endId).toBools
       (rqueues zip (arsel zip rsel)) foreach { case (q, (ar, r)) =>
-        q.io.deq.ready := out.r .valid && in .r .ready && r && out.r.bits.last
+        q.deq.ready := out.r .valid && in .r .ready && r && out.r.bits.last
-        q.io.enq.valid := in .ar.valid && out.ar.ready && ar
+        q.enq.valid := in .ar.valid && out.ar.ready && ar
-        q.io.enq.bits  := in.ar.bits.user.get
+        q.enq.bits  := in.ar.bits.user.get
       }
 
       val awid = in.aw.bits.id
-      val aw_ready = Vec(wqueues.map(_.io.enq.ready))(awid)
+      val aw_ready = Vec(wqueues.map(_.enq.ready))(awid)
       in .aw.ready := out.aw.ready && aw_ready
       out.aw.valid := in .aw.valid && aw_ready
       out.aw.bits  := in .aw.bits
 
       val bid = out.b.bits.id
-      val b_valid = Vec(wqueues.map(_.io.deq.valid))(bid)
+      val b_valid = Vec(wqueues.map(_.deq.valid))(bid)
-      val b_bits = Vec(wqueues.map(_.io.deq.bits))(bid)
+      val b_bits = Vec(wqueues.map(_.deq.bits))(bid)
       assert (!out.b.valid || b_valid) // Q must be ready faster than the response
       in.b <> out.b
       in.b.bits.user.get := b_bits
@@ -69,9 +85,9 @@ class AXI4UserYanker(capMaxFlight: Option[Int] = None)(implicit p: Parameters) e
       val awsel = UIntToOH(awid, edgeIn.master.endId).toBools
       val bsel  = UIntToOH(bid,  edgeIn.master.endId).toBools
       (wqueues zip (awsel zip bsel)) foreach { case (q, (aw, b)) =>
-        q.io.deq.ready := out.b .valid && in .b .ready && b
+        q.deq.ready := out.b .valid && in .b .ready && b
-        q.io.enq.valid := in .aw.valid && out.aw.ready && aw
+        q.enq.valid := in .aw.valid && out.aw.ready && aw
-        q.io.enq.bits  := in.aw.bits.user.get
+        q.enq.bits  := in.aw.bits.user.get
       }
 
       out.w <> in.w

src/main/scala/uncore/tilelink2/ToAXI4.scala

@@ -18,6 +18,7 @@ case class TLToAXI4Node(beatBytes: Int) extends MixedAdapterNode(TLImp, AXI4Imp)
       AXI4MasterParameters(
         id        = IdRange(start, start+size),
         aligned   = true,
+        maxFlight = Some(if (c.requestFifo) c.sourceId.size else 1),
         nodePath  = c.nodePath)
     }
     AXI4MasterPortParameters(