CHI-CoupledL2: optimize timing (#191)

* CoupledL2：timing fix

* MSHRCtl: fix counter for performance

* Revert "BOP: add clock gating (#169)"

This reverts commit 2199272.

* MSHRCtl: Change to RoundRobin to fix commit timeout

---------

Co-authored-by: Yanqin Li <[email protected]>

src/main/scala/coupledL2/RequestArb.scala

@@ -130,19 +130,19 @@ class RequestArb(implicit p: Parameters) extends L2Module {
     (if (io.fromTXRSP.isDefined) io.fromTXRSP.get.blockSinkBReqEntrance else false.B)
   val block_C = io.fromMSHRCtl.blockC_s1 || io.fromMainPipe.blockC_s1 || io.fromGrantBuffer.blockSinkReqEntrance.blockC_s1
 
-  val noFreeWay = Wire(Bool())
+//  val noFreeWay = Wire(Bool())
 
   val sinkValids = VecInit(Seq(
     io.sinkC.valid && !block_C,
     io.sinkB.valid && !block_B,
-    io.sinkA.valid && !block_A && !noFreeWay
+    io.sinkA.valid && !block_A
   )).asUInt
 
   // TODO: A Hint is allowed to enter if !s2_ready for mcp2_stall
 
   val sink_ready_basic = io.dirRead_s1.ready && resetFinish && !mshr_task_s1.valid && s2_ready
 
-  io.sinkA.ready := sink_ready_basic && !block_A && !sinkValids(1) && !sinkValids(0) && !noFreeWay // SinkC prior to SinkA & SinkB
+  io.sinkA.ready := sink_ready_basic && !block_A && !sinkValids(1) && !sinkValids(0) // SinkC prior to SinkA & SinkB
   io.sinkB.ready := sink_ready_basic && !block_B && !sinkValids(0) // SinkB prior to SinkA
   io.sinkC.ready := sink_ready_basic && !block_C
 
@@ -195,13 +195,13 @@ class RequestArb(implicit p: Parameters) extends L2Module {
   task_s2.valid := s1_fire
   when(s1_fire) { task_s2.bits := task_s1.bits }
 
-  val sameSet_s2 = task_s2.valid && task_s2.bits.fromA && !task_s2.bits.mshrTask && task_s2.bits.set === A_task.set
+/*  val sameSet_s2 = task_s2.valid && task_s2.bits.fromA && !task_s2.bits.mshrTask && task_s2.bits.set === A_task.set
   val sameSet_s3 = RegNext(task_s2.valid && task_s2.bits.fromA && !task_s2.bits.mshrTask) &&
     RegEnable(task_s2.bits.set, task_s2.valid) === A_task.set
   val sameSetCnt = PopCount(VecInit(io.msInfo.map(s => s.valid && s.bits.set === A_task.set && s.bits.fromA) :+
     sameSet_s2 :+ sameSet_s3).asUInt)
   noFreeWay := sameSetCnt >= cacheParams.ways.U
-
+ */
   io.taskToPipe_s2 := task_s2
 
   // MSHR task

src/main/scala/coupledL2/RequestBuffer.scala

@@ -75,6 +75,8 @@ class RequestBuffer(flow: Boolean = true, entries: Int = 4)(implicit p: Paramete
     val mshrInfo  = Vec(mshrsAll, Flipped(ValidIO(new MSHRInfo)))
     val aMergeTask = ValidIO(new AMergeTask)
     val mainPipeBlock = Input(Vec(2, Bool()))
+    /* Snoop task from arbiter at stage 2 */
+    val taskFromArb_s2 = Flipped(ValidIO(new TaskBundle()))
 
     val ATag        = Output(UInt(tagBits.W))
     val ASet        = Output(UInt(setBits.W))
@@ -143,8 +145,20 @@ class RequestBuffer(flow: Boolean = true, entries: Int = 4)(implicit p: Paramete
   io.aMergeTask.bits.id := mergeAId
   io.aMergeTask.bits.task := in
 
+  /*
+   noFreeWay check: s2 + s3 + mshrs >= ways(L2)
+   */
+  val task_s2 = io.taskFromArb_s2
+  val sameSet_s2 = task_s2.valid && task_s2.bits.fromA && !task_s2.bits.mshrTask && task_s2.bits.set === io.ASet
+  val sameSet_s3 = RegNext(task_s2.valid && task_s2.bits.fromA && !task_s2.bits.mshrTask) &&
+    RegEnable(task_s2.bits.set, task_s2.valid) === io.ASet
+  val sameSetCnt = PopCount(VecInit(io.mshrInfo.map(s => s.valid && s.bits.set === io.ASet && s.bits.fromA) :+
+    sameSet_s2 :+ sameSet_s3).asUInt)
+  val noFreeWay = sameSetCnt >= cacheParams.ways.U
+  dontTouch (noFreeWay)
+
   // flow not allowed when full, or entries might starve
-  val canFlow = flow.B && !full && !conflict(in) && !chosenQValid && !Cat(io.mainPipeBlock).orR
+  val canFlow = flow.B && !full && !conflict(in) && !chosenQValid && !Cat(io.mainPipeBlock).orR && !noFreeWay
   val doFlow  = canFlow && io.out.ready
   io.hasLatePF := latePrefetch(in) && io.in.valid && !sameAddr(in, RegNext(in))
   io.hasMergeA := mergeA && io.in.valid && !sameAddr(in, RegNext(in))
@@ -177,7 +191,7 @@ class RequestBuffer(flow: Boolean = true, entries: Int = 4)(implicit p: Paramete
 
     entry.valid   := true.B
     // when Addr-Conflict / Same-Addr-Dependent / MainPipe-Block / noFreeWay-in-Set, entry not ready
-    entry.rdy     := !conflict(in) && !mpBlock && !s1Block // && !Cat(depMask).orR
+    entry.rdy     := !conflict(in) && !mpBlock && !s1Block && !noFreeWay// && !Cat(depMask).orR
     entry.task    := io.in.bits
     entry.waitMP  := Cat(
       s1Block,
@@ -247,7 +261,7 @@ class RequestBuffer(flow: Boolean = true, entries: Int = 4)(implicit p: Paramete
       // update info
       e.waitMS  := waitMSUpdate
 //      e.depMask := depMaskUpdate
-      e.rdy     := !waitMSUpdate.orR && !e.waitMP && !s1_Block
+      e.rdy     := !waitMSUpdate.orR && !e.waitMP && !s1_Block && !noFreeWay
     }
   }
 

src/main/scala/coupledL2/prefetch/BestOffsetPrefetch.scala

@@ -17,7 +17,7 @@
 
 package coupledL2.prefetch
 
-import utility.{GatedValidRegNext, ChiselDB, Constantin, MemReqSource, ParallelPriorityMux, RRArbiterInit, SRAMTemplate}
+import utility.{ChiselDB, Constantin, MemReqSource, ParallelPriorityMux, RRArbiterInit, SRAMTemplate}
 import org.chipsalliance.cde.config.Parameters
 import chisel3.DontCare.:=
 import chisel3._
@@ -177,13 +177,13 @@ class RecentRequestTable(implicit p: Parameters) extends BOPModule {
   rrTable.io.r.req.bits.setIdx := idx(rAddr)
   rData := rrTable.io.r.resp.data(0)
 
-  assert(!GatedValidRegNext(io.w.fire && io.r.req.fire), "single port SRAM should not read and write at the same time")
+  assert(!RegNext(io.w.fire && io.r.req.fire), "single port SRAM should not read and write at the same time")
 
   io.w.ready := rrTable.io.w.req.ready && !io.r.req.valid
   io.r.req.ready := true.B
-  io.r.resp.valid := GatedValidRegNext(rrTable.io.r.req.fire, false.B)
-  io.r.resp.bits.ptr := RegEnable(io.r.req.bits.ptr, rrTable.io.r.req.fire)
-  io.r.resp.bits.hit := rData.valid && rData.tag === RegEnable(tag(rAddr), rrTable.io.r.req.fire)
+  io.r.resp.valid := RegNext(rrTable.io.r.req.fire, false.B)
+  io.r.resp.bits.ptr := RegNext(io.r.req.bits.ptr)
+  io.r.resp.bits.hit := rData.valid && rData.tag === RegNext(tag(rAddr))
 
 }
 

src/main/scala/coupledL2/tl2chi/MSHRCtl.scala

@@ -120,46 +120,68 @@ class MSHRCtl(implicit p: Parameters) extends TL2CHIL2Module {
   mshrSelector.io.idle := mshrs.map(m => !m.io.status.valid)
   io.toMainPipe.mshr_alloc_ptr := OHToUInt(selectedMSHROH)
 
+  /*
+   rxrsp for PCredit timing is quite critical and break it here
+   */
+  val rxrspValid = RegNext(io.resps.rxrsp.valid)
+  val rxrspInfo = RegNext(io.resps.rxrsp.respInfo)
+  val rxrspMshrId = RegNext( io.resps.rxrsp.mshrId)
+
   /*
    when PCrdGrant, give credit to one entry that:
    1. got RetryAck and not Reissued
    2. match srcID and PCrdType
    3. use Round-Robin arbiter if multi-entry match
    */
   val isPCrdGrant = io.resps.rxrsp.valid && (io.resps.rxrsp.respInfo.chiOpcode.get === PCrdGrant)
+  val isPCrdGrantReg = RegNext(isPCrdGrant)
   val waitPCrdInfo  = Wire(Vec(mshrsAll, new PCrdInfo))
-//  val pArb = Module(new RRArbiter(UInt(), mshrsAll))
+  val timeOutPri = VecInit(Seq.fill(16)(false.B))
+  val timeOutSel = WireInit(false.B)
+  val pCrdPri = VecInit(Seq.fill(16)(false.B))
+  val pArb = Module(new RRArbiter(UInt(), mshrsAll))
 
   val matchPCrdGrant = VecInit(waitPCrdInfo.map(p =>
       isPCrdGrant && p.valid &&
       p.srcID.get === io.resps.rxrsp.respInfo.srcID.get &&
       p.pCrdType.get === io.resps.rxrsp.respInfo.pCrdType.get
   ))
 
-/*  pArb.io.in.zipWithIndex.foreach {
-      case (in, i) =>
-      in.valid := matchPCrdGrant(i)
+  val matchPCrdGrantReg = RegNext(matchPCrdGrant)
+  pArb.io.in.zipWithIndex.foreach {
+    case (in, i) =>
+      in.valid := matchPCrdGrantReg(i)
       in.bits := 0.U
   }
   pArb.io.out.ready := true.B
-  val pCrdRR = VecInit(UIntToOH(pArb.io.chosen))
-  val pCrdPri = VecInit((matchPCrdGrant.asUInt & pCrdRR.asUInt).asBools)
-//val pCrdPri = VecInit(PriorityEncoderOH(matchPCrdGrant))
-  val pCrdIsWait = OHToUInt(pCrdPri)
- */
 
-  /*
-   Random arbiter if multi-entry match
-   */
-  val lfsr = LFSR(16, true.B)
-  val idx = Random(16, lfsr)
-  val idxOH = VecInit(UIntToOH(idx))
+  val pCrdOH = VecInit(UIntToOH(pArb.io.chosen).asBools)
+  val pCrdFixPri = VecInit(pCrdOH zip matchPCrdGrantReg map {case(a,b) => a && b})
+//val pCrdFixPri = VecInit(PriorityEncoderOH(matchPCrdGrantReg)) //fix priority arbiter
+
+  // timeout protect
+  val counter = RegInit(VecInit(Seq.fill(mshrsAll)(0.U((log2Ceil(mshrsAll)+1).W))))
+
+  for(i <- 0 until 16) {
+    when(matchPCrdGrantReg(i)) {
+      when(!timeOutSel && pCrdFixPri(i) || timeOutPri(i)) {
+        counter(i):=0.U
+      }.otherwise {
+        counter(i):= counter(i) + 1.U
+      }
+    }
+  }
+  val timeOutOH = PriorityEncoderOH(counter.map(_>=12.U) zip matchPCrdGrantReg map {case(a,b) => a&&b})
+  timeOutPri := VecInit(timeOutOH)
+
+  timeOutSel := timeOutPri.reduce(_|_)
+  pCrdPri := Mux(timeOutSel, timeOutPri, pCrdFixPri)
 
-  val doubleReq = Fill(2, matchPCrdGrant.asUInt)
-  val doubleGnt = ~(doubleReq - idxOH.asUInt) & doubleReq
-  val gnt = doubleGnt(31,16) | doubleGnt(15,0)
-  val pCrdPri = VecInit(gnt.asBools)
-  val pCrdIsWait = OHToUInt(pCrdPri)
+  dontTouch (timeOutPri)
+  dontTouch (timeOutSel)
+  dontTouch (pCrdOH)
+  dontTouch (pCrdFixPri)
+  dontTouch (pCrdPri)
 
   /* when PCrdGrant come before RetryAck, 16 entry CAM used to:
    1. save {srcID, PCrdType} 
@@ -171,7 +193,8 @@ class MSHRCtl(implicit p: Parameters) extends TL2CHIL2Module {
   val pCamValids = Cat(pCam.map(_.valid))
   val enqIdx = PriorityEncoder(~pCamValids.asUInt)
 
-  when (isPCrdGrant && !pCrdIsWait.orR){
+//  when (isPCrdGrant && !pCrdIsWait.orR){
+  when (isPCrdGrant){
     pCam(enqIdx).valid := true.B
     pCam(enqIdx).srcID.get := io.resps.rxrsp.respInfo.srcID.get
     pCam(enqIdx).pCrdType.get := io.resps.rxrsp.respInfo.pCrdType.get
@@ -214,8 +237,10 @@ class MSHRCtl(implicit p: Parameters) extends TL2CHIL2Module {
       m.io.resps.rxdat.valid := m.io.status.valid && io.resps.rxdat.valid && io.resps.rxdat.mshrId === i.U
       m.io.resps.rxdat.bits := io.resps.rxdat.respInfo
 
-      m.io.resps.rxrsp.valid := (m.io.status.valid && io.resps.rxrsp.valid && !isPCrdGrant && io.resps.rxrsp.mshrId === i.U) || (isPCrdGrant && pCrdPri(i))
-      m.io.resps.rxrsp.bits := io.resps.rxrsp.respInfo
+//      m.io.resps.rxrsp.valid := (m.io.status.valid && io.resps.rxrsp.valid && !isPCrdGrant && io.resps.rxrsp.mshrId === i.U) || (isPCrdGrant && pCrdPri(i))
+//      m.io.resps.rxrsp.bits := io.resps.rxrsp.respInfo
+      m.io.resps.rxrsp.valid := (m.io.status.valid && rxrspValid && !isPCrdGrantReg && rxrspMshrId === i.U) || (isPCrdGrantReg && pCrdPri(i))
+      m.io.resps.rxrsp.bits := rxrspInfo
 
       m.io.replResp.valid := io.replResp.valid && io.replResp.bits.mshrId === i.U
       m.io.replResp.bits := io.replResp.bits
@@ -226,7 +251,7 @@ class MSHRCtl(implicit p: Parameters) extends TL2CHIL2Module {
       m.io.aMergeTask.bits := io.aMergeTask.bits.task
 
       waitPCrdInfo(i) := m.io.waitPCrdInfo 
-      m.io.pCamPri := (pCamPri === i.U) && waitPCrdInfo(i).valid
+      m.io.pCamPri := 0.U /*(pCamPri === i.U) && waitPCrdInfo(i).valid*/
   }
   /* Reserve 1 entry for SinkB */
   io.waitPCrdInfo <> waitPCrdInfo

src/main/scala/coupledL2/tl2chi/Slice.scala

@@ -107,6 +107,7 @@ class Slice()(implicit p: Parameters) extends BaseSlice[OuterBundle]
   reqBuf.io.mshrInfo := mshrCtl.io.msInfo
   reqBuf.io.mainPipeBlock := mainPipe.io.toReqBuf
   reqBuf.io.s1Entrance := reqArb.io.s1Entrance
+  reqBuf.io.taskFromArb_s2 := reqArb.io.taskToPipe_s2
 
   mainPipe.io.taskFromArb_s2 := reqArb.io.taskToPipe_s2
   mainPipe.io.taskInfo_s1 := reqArb.io.taskInfo_s1

src/main/scala/coupledL2/tl2chi/TXREQ.scala

@@ -45,18 +45,21 @@ class TXREQ(implicit p: Parameters) extends TL2CHIL2Module {
   require(chiOpt.isDefined)
 
   // TODO: an mshrsAll-entry queue is too much, evaluate for a proper size later
-  val queue = Module(new Queue(new CHIREQ, entries = mshrsAll, flow = true))
+  val queue = Module(new Queue(new CHIREQ, entries = mshrsAll, flow = false))
 
   // Back pressure logic from TXREQ
   val queueCnt = queue.io.count
   // TODO: this may be imprecise, review this later
   val pipeStatus_s1_s5 = io.pipeStatusVec
   val pipeStatus_s2_s5 = pipeStatus_s1_s5.tail
   val pipeStatus_s1 = pipeStatus_s1_s5.head
+  val pipeStatus_s2 = pipeStatus_s1_s5(1)
+  val s2ReturnCredit = pipeStatus_s2.valid && !(pipeStatus_s2.bits.mshrTask && pipeStatus_s2.bits.toTXREQ)
   // inflightCnt equals the number of reqs on s2~s5 that may flow into TXREQ soon, plus queueCnt.
   // The calculation of inflightCnt might be imprecise and leads to false positive back pressue.
   val inflightCnt = PopCount(Cat(pipeStatus_s2_s5.map(s => s.valid && s.bits.mshrTask && s.bits.toTXREQ))) +
-    pipeStatus_s1.valid.asUInt +
+//    pipeStatus_s1.valid.asUInt +
+    1.U - s2ReturnCredit.asUInt + //Fix Timing: always take credit and s2 return if not take 
     queueCnt
   val noSpace = inflightCnt >= mshrsAll.U
 
@@ -73,4 +76,4 @@ class TXREQ(implicit p: Parameters) extends TL2CHIL2Module {
   io.out.bits.tgtID := SAM(sam).lookup(io.out.bits.addr)
   io.out.bits.size := log2Ceil(blockBytes).U(SIZE_WIDTH.W) // TODO
   io.out.bits.addr := restoreAddressUInt(queue.io.deq.bits.addr, io.sliceId)
-}
+}

src/main/scala/coupledL2/tl2chi/TXRSP.scala

@@ -45,7 +45,7 @@ class TXRSP(implicit p: Parameters) extends TL2CHIL2Module {
   require(chiOpt.isDefined)
 
   // TODO: an mshrsAll-entry queue is too much, evaluate for a proper size later
-  val queue = Module(new Queue(new CHIRSP, entries = mshrsAll, flow = true))
+  val queue = Module(new Queue(new CHIRSP, entries = mshrsAll, flow = false))
 
   // Back pressure logic from TXRSP
   val queueCnt = queue.io.count
@@ -88,4 +88,4 @@ class TXRSP(implicit p: Parameters) extends TL2CHIL2Module {
     // TODO: Finish this
     rsp
   }
-}
+}

src/main/scala/coupledL2/tl2tl/Slice.scala

@@ -59,6 +59,8 @@ class Slice()(implicit p: Parameters) extends BaseSlice[OuterBundle] {
   a_reqBuf.io.mshrInfo := mshrCtl.io.msInfo
   a_reqBuf.io.mainPipeBlock := mainPipe.io.toReqBuf
   a_reqBuf.io.s1Entrance := reqArb.io.s1Entrance
+  a_reqBuf.io.taskFromArb_s2 := reqArb.io.taskToPipe_s2
+
   sinkB.io.msInfo := mshrCtl.io.msInfo
   sinkC.io.msInfo := mshrCtl.io.msInfo