Merge pull request #432 from ucb-bar/tl2-address-filtering

Tl2 address filtering

src/main/scala/coreplex/BaseCoreplex.scala

@@ -18,6 +18,8 @@ import util._
 case object NMemoryChannels extends Field[Int]
 /** Number of banks per memory channel */
 case object NBanksPerMemoryChannel extends Field[Int]
+/** Number of tracker per bank */
+case object NTrackersPerBank extends Field[Int]
 /** Least significant bit of address used for bank partitioning */
 case object BankIdLSB extends Field[Int]
 /** Function for building some kind of coherence manager agent */
@@ -39,6 +41,7 @@ trait HasCoreplexParameters {
   lazy val nSlaves = p(rocketchip.NCoreplexExtClients)
   lazy val nMemChannels = p(NMemoryChannels)
   lazy val hasSupervisor = p(rocket.UseVM)
+  lazy val nTrackersPerBank = p(NTrackersPerBank)
 }
 
 case class CoreplexParameters(implicit val p: Parameters) extends HasCoreplexParameters
@@ -93,6 +96,45 @@ trait CoreplexNetworkModule extends HasCoreplexParameters {
   implicit val p = outer.p
 }
 
+trait BankedL2 {
+    this: CoreplexNetwork =>
+  require (isPow2(nBanksPerMemChannel))
+  require (isPow2(l1tol2_beatBytes))
+
+  def l2ManagerFactory(): (TLInwardNode, TLOutwardNode)
+
+  val l2Channels = Seq.fill(nMemChannels) {
+    val bankBar = LazyModule(new TLXbar)
+    val output = TLOutputNode()
+
+    output := bankBar.node
+    val mask = ~BigInt((nBanksPerMemChannel-1) * l1tol2_lineBytes)
+    for (i <- 0 until nBanksPerMemChannel) {
+      val (in, out) = l2ManagerFactory()
+      in := TLFilter(AddressSet(i * l1tol2_lineBytes, mask))(l1tol2.node)
+      bankBar.node := out
+    }
+
+    output
+  }
+}
+
+trait BankedL2Bundle {
+  this: CoreplexNetworkBundle {
+    val outer: BankedL2
+  } =>
+
+  require (nMemChannels == 1, "Seq in Chisel Bundle needed to support > 1") // !!!
+  val mem = outer.l2Channels.map(_.bundleOut).toList.head // .head should be removed !!!
+}
+
+trait BankedL2Module {
+  this: CoreplexNetworkModule {
+    val outer: BankedL2
+    val io: BankedL2Bundle
+  } =>
+}
+
 trait CoreplexRISCVPlatform {
     this: CoreplexNetwork =>
 

src/main/scala/coreplex/Configs.scala

@@ -155,6 +155,7 @@ class BaseCoreplexConfig extends Config (
       case BootROMFile => "./bootrom/bootrom.img"
       case NTiles => 1
       case NBanksPerMemoryChannel => Knob("NBANKS_PER_MEM_CHANNEL")
+      case NTrackersPerBank => Knob("NTRACKERS_PER_BANK")
       case BankIdLSB => 0
       case CacheBlockBytes => Dump("CACHE_BLOCK_BYTES", 64)
       case CacheBlockOffsetBits => log2Up(here(CacheBlockBytes))
@@ -163,6 +164,7 @@ class BaseCoreplexConfig extends Config (
   }},
   knobValues = {
     case "NBANKS_PER_MEM_CHANNEL" => 1
+    case "NTRACKERS_PER_BANK" => 4
     case "L1D_MSHRS" => 2
     case "L1D_SETS" => 64
     case "L1D_WAYS" => 4
@@ -183,6 +185,12 @@ class WithNBanksPerMemChannel(n: Int) extends Config(
     case _ => throw new CDEMatchError
   })
 
+class WithNTrackersPerBank(n: Int) extends Config(
+  knobValues = {
+    case "NTRACKERS_PER_BANK" => n
+    case _ => throw new CDEMatchError
+  })
+
 class WithDataScratchpad(n: Int) extends Config(
   (pname,site,here) => pname match {
     case DataScratchpadSize => n

src/main/scala/coreplex/Coreplex.scala

@@ -10,6 +10,16 @@ import uncore.util._
 import util._
 import rocket._
 
+trait BroadcastL2 {
+    this: CoreplexNetwork =>
+  def l2ManagerFactory() = {
+    val bh = LazyModule(new TLBroadcast(l1tol2_beatBytes, nTrackersPerBank))
+    (bh.node, bh.node)
+  }
+}
+
+/////
+
 trait DirectConnection {
     this: CoreplexNetwork with CoreplexRISCVPlatform =>
   lazyTiles.map(_.slave).flatten.foreach { scratch => scratch := cbus.node }

src/main/scala/diplomacy/Parameters.scala

@@ -104,6 +104,17 @@ case class AddressSet(base: BigInt, mask: BigInt) extends Ordered[AddressSet]
   // Widen the match function to ignore all bits in imask
   def widen(imask: BigInt) = AddressSet(base & ~imask, mask | imask)
 
+  // Return an AddressSet that only contains the addresses both sets contain
+  def intersect(x: AddressSet): Option[AddressSet] = {
+    if (!overlaps(x)) {
+      None
+    } else {
+      val r_mask = mask & x.mask
+      val r_base = base | x.base
+      Some(AddressSet(r_base, r_mask))
+    }
+  }
+
   // AddressSets have one natural Ordering (the containment order, if contiguous)
   def compare(x: AddressSet) = {
     val primary   = (this.base - x.base).signum // smallest address first
@@ -133,5 +144,24 @@ object AddressSet
       misaligned(base+step, size-step, AddressSet(base, step-1) +: tail)
     }
   }
-}
 
+  def unify(seq: Seq[AddressSet]): Seq[AddressSet] = {
+    val n = seq.size
+    val array = Array(seq:_*)
+    var filter = Array.fill(n) { false }
+    for (i <- 0 until n-1) { if (!filter(i)) {
+      for (j <- i+1 until n) { if (!filter(j)) {
+        val a = array(i)
+        val b = array(j)
+        if (a.mask == b.mask && isPow2(a.base ^ b.base)) {
+          val c_base = a.base & ~(a.base ^ b.base)
+          val c_mask = a.mask | (a.base ^ b.base)
+          filter.update(j, true)
+          array.update(i, AddressSet(c_base, c_mask))
+        }
+      }}
+    }}
+    val out = (array zip filter) flatMap { case (a, f) => if (f) None else Some(a) }
+    if (out.size != n) unify(out) else out.toList
+  }
+}

src/main/scala/uncore/tilelink2/Broadcast.scala

@@ -67,7 +67,7 @@ class TLBroadcast(lineBytes: Int, numTrackers: Int = 4) extends LazyModule
     require (lineBytes >= edgeOut.manager.beatBytes)
     // For the probe walker, we need to identify all the caches
     val caches = clients.filter(_.supportsProbe).map(_.sourceId)
-    val cache_targets = Vec(caches.map(c => UInt(c.start)))
+    val cache_targets = caches.map(c => UInt(c.start))
 
     // Create the request tracker queues
     val trackers = Seq.tabulate(numTrackers) { id =>
@@ -81,7 +81,7 @@ class TLBroadcast(lineBytes: Int, numTrackers: Int = 4) extends LazyModule
     }
 
     // Depending on the high source bits, we might transform D
-    val d_high = 1 << log2Ceil(edgeIn.client.endSourceId)
+    val d_high = log2Ceil(edgeIn.client.endSourceId)
     val d_what = out.d.bits.source(d_high+1, d_high)
     val d_drop = d_what === DROP
     val d_hasData = edgeOut.hasData(out.d.bits)
@@ -143,20 +143,22 @@ class TLBroadcast(lineBytes: Int, numTrackers: Int = 4) extends LazyModule
        trackers.map { t => (edgeOut.numBeats1(t.out_a.bits), t.out_a) }):_*)
 
     // The Probe FSM walks all caches and probes them
-    val probe_todo = RegInit(UInt(0, width = caches.size))
+    val probe_todo = RegInit(UInt(0, width = max(1, caches.size)))
     val probe_line = Reg(UInt())
     val probe_perms = Reg(UInt(width = 2))
     val probe_next = probe_todo & ~(leftOR(probe_todo) << 1)
     val probe_busy = probe_todo.orR()
-    val probe_target = Mux1H(probe_next, cache_targets)
+    val probe_target = if (caches.size == 0) UInt(0) else Mux1H(probe_next, cache_targets)
 
     // Probe whatever the FSM wants to do next
     in.b.valid := probe_busy
+    if (caches.size != 0) {
       in.b.bits := edgeIn.Probe(probe_line << lineShift, probe_target, UInt(lineShift), probe_perms)._2
+    }
     when (in.b.fire()) { probe_todo := probe_todo & ~probe_next }
 
     // Which cache does a request come from?
-    val a_cache = Vec(caches.map(_.contains(in.a.bits.source))).asUInt
+    val a_cache = if (caches.size == 0) UInt(1) else Vec(caches.map(_.contains(in.a.bits.source))).asUInt
     val (a_first, _, _) = edgeIn.firstlast(in.a)
 
     // To accept a request from A, the probe FSM must be idle and there must be a matching tracker
@@ -173,7 +175,7 @@ class TLBroadcast(lineBytes: Int, numTrackers: Int = 4) extends LazyModule
       t.in_a.valid := in.a.valid && select && (!a_first || !probe_busy)
       t.in_a.bits := in.a.bits
       t.in_a_first := a_first
-      t.probe := Mux(a_cache.orR(), UInt(caches.size-1), UInt(caches.size))
+      t.probe := (if (caches.size == 0) UInt(0) else Mux(a_cache.orR(), UInt(caches.size-1), UInt(caches.size)))
     }
 
     when (in.a.fire() && a_first) {
@@ -256,12 +258,7 @@ class TLBroadcastTracker(id: Int, lineBytes: Int, probeCountBits: Int, edgeIn: T
   io.source := source
   io.line := address >> lineShift
 
-  class Data extends Bundle {
+  val i_data = Wire(Decoupled(new TLBroadcastData(edgeIn.bundle)))
-    val mask = io.in_a.bits.mask.cloneType
-    val data = io.in_a.bits.data.cloneType
-  }
-
-  val i_data = Wire(Decoupled(new Data))
   val o_data = Queue(i_data, lineBytes / edgeIn.manager.beatBytes)
 
   io.in_a.ready := (idle || !io.in_a_first) && i_data.ready
@@ -295,3 +292,9 @@ object TLBroadcastConstants
   val DROP        = UInt(1)
   val PASS        = UInt(0)
 }
+
+class TLBroadcastData(params: TLBundleParameters) extends TLBundleBase(params)
+{
+  val mask = UInt(width = params.dataBits/8)
+  val data = UInt(width = params.dataBits)
+}

src/main/scala/uncore/tilelink2/Filter.scala

@@ -0,0 +1,49 @@
+// See LICENSE for license details.
+
+package uncore.tilelink2
+
+import Chisel._
+import chisel3.internal.sourceinfo.SourceInfo
+import diplomacy._
+import scala.math.{min,max}
+
+class TLFilter(select: AddressSet) extends LazyModule
+{
+  val node = TLAdapterNode(
+    clientFn  = { case Seq(cp) => cp },
+    managerFn = { case Seq(mp) =>
+      mp.copy(managers = mp.managers.map { m =>
+        val filtered = m.address.map(_.intersect(select)).flatten
+        val cap = TransferSizes(1, select.alignment.toInt)
+        if (filtered.isEmpty) { None } else {
+          Some(m.copy(
+            address            = filtered,
+            supportsAcquire    = m.supportsAcquire   .intersect(cap),
+            supportsArithmetic = m.supportsArithmetic.intersect(cap),
+            supportsLogical    = m.supportsLogical   .intersect(cap),
+            supportsGet        = m.supportsGet       .intersect(cap),
+            supportsPutFull    = m.supportsPutFull   .intersect(cap),
+            supportsPutPartial = m.supportsPutPartial.intersect(cap),
+            supportsHint       = m.supportsHint      .intersect(cap)))
+        }
+      }.flatten)
+    })
+
+  lazy val module = new LazyModuleImp(this) {
+    val io = new Bundle {
+      val in  = node.bundleIn
+      val out = node.bundleOut
+    }
+    io.out <> io.in
+  }
+}
+
+object TLFilter
+{
+  // applied to the TL source node; y.node := TLBuffer(x.node)
+  def apply(select: AddressSet)(x: TLOutwardNode)(implicit sourceInfo: SourceInfo): TLOutwardNode = {
+    val filter = LazyModule(new TLFilter(select))
+    filter.node := x
+    filter.node
+  }
+}

src/main/scala/uncore/tilelink2/Parameters.scala

@@ -23,7 +23,9 @@ case class TLManagerParameters(
   fifoId:             Option[Int]   = None,
   customDTS:          Option[String]= None)
 {
+  require (!address.isEmpty)
   address.foreach { a => require (a.finite) }
+
   address.combinations(2).foreach { case Seq(x,y) => require (!x.overlaps(y)) }
   require (supportsPutFull.contains(supportsPutPartial))
 
@@ -285,3 +287,45 @@ case class TLAsyncEdgeParameters(client: TLAsyncClientPortParameters, manager: T
 {
   val bundle = TLAsyncBundleParameters(manager.depth, TLBundleParameters(client.base, manager.base))
 }
+
+object ManagerUnification
+{
+  def apply(managers: Seq[TLManagerParameters]) = {
+    // To be unified, devices must agree on all of these terms
+    case class TLManagerKey(
+      regionType:         RegionType.T,
+      executable:         Boolean,
+      lastNode:           BaseNode,
+      supportsAcquire:    TransferSizes,
+      supportsArithmetic: TransferSizes,
+      supportsLogical:    TransferSizes,
+      supportsGet:        TransferSizes,
+      supportsPutFull:    TransferSizes,
+      supportsPutPartial: TransferSizes,
+      supportsHint:       TransferSizes)
+    def key(x: TLManagerParameters) = TLManagerKey(
+      regionType         = x.regionType,
+      executable         = x.executable,
+      lastNode           = x.nodePath.last,
+      supportsAcquire    = x.supportsAcquire,
+      supportsArithmetic = x.supportsArithmetic,
+      supportsLogical    = x.supportsLogical,
+      supportsGet        = x.supportsGet,
+      supportsPutFull    = x.supportsPutFull,
+      supportsPutPartial = x.supportsPutPartial,
+      supportsHint       = x.supportsHint)
+    val map = scala.collection.mutable.HashMap[TLManagerKey, TLManagerParameters]()
+    managers.foreach { m =>
+      val k = key(m)
+      map.get(k) match {
+        case None => map.update(k, m)
+        case Some(n) => {
+          map.update(k, m.copy(
+            address = m.address ++ n.address,
+            fifoId  = None)) // Merging means it's not FIFO anymore!
+        }
+      }
+    }
+    map.values.map(m => m.copy(address = AddressSet.unify(m.address))).toList
+  }
+}

src/main/scala/uncore/tilelink2/Xbar.scala

@@ -54,13 +54,13 @@ class TLXbar(policy: TLArbiter.Policy = TLArbiter.lowestIndexFirst) extends Lazy
       seq(0).copy(
         minLatency = seq.map(_.minLatency).min,
         endSinkId = outputIdRanges.map(_.end).max,
-        managers = (outputIdRanges zip seq) flatMap { case (range, port) =>
+        managers = ManagerUnification(seq.flatMap { port =>
           require (port.beatBytes == seq(0).beatBytes)
           val fifoIdMapper = fifoIdFactory()
           port.managers map { manager => manager.copy(
             fifoId = manager.fifoId.map(fifoIdMapper(_))
           )}
-        }
+        })
       )
     })