mem-ruby: Implement no_alloc Far Atomics in CHI (#994)

This PR introduces a missing pice of far atomic implementation. This pull request incorporates several changes: - Enable 2-level and 4-level (and N-level) cache hierarchies, removing Atomic_NoWait transactions - Fix Unique Near policy implementation that raised abort - Add support for alloc_on_atomic == False. Enables Far Atomics on systems where the HNF does not allocate evicted lines at LLC (Like in WriteUpdate).
2024-04-18 11:35:47 -07:00
parent c44b8635ab 98358da968
commit 42ffa52907
4 changed files with 192 additions and 78 deletions
--- a/src/mem/ruby/protocol/chi/CHI-cache-actions.sm
+++ b/src/mem/ruby/protocol/chi/CHI-cache-actions.sm
@@ -823,7 +823,8 @@ action(Initiate_AtomicReturn_I, desc="") {
    tbe.atomic_to_be_done := true;
  } else if ((policy_type == 1) || // UNIQUE NEAR
             (policy_type == 2)) { // PRESENT NEAR
-    tbe.actions.push(Event:SendAtomicReturn_NoWait);
+    tbe.actions.push(Event:SendAtomicReturn);
+    tbe.actions.push(Event:SendARData);
    tbe.dataToBeInvalid := true;
    tbe.doCacheFill := false;
    tbe.atomic_to_be_done := false;
@@ -840,7 +841,8 @@ action(Initiate_AtomicNoReturn_I, desc="") {
    tbe.actions.push(Event:WriteBEPipe);
    tbe.actions.push(Event:TagArrayWrite);
    tbe.atomic_to_be_done := true;
-  } else if (policy_type == 1) { // UNIQUE NEAR
+  } else if ((policy_type == 1) || // UNIQUE NEAR
+             (policy_type == 2)) { // PRESENT NEAR
    tbe.actions.push(Event:SendAtomicNoReturn);
    tbe.actions.push(Event:SendANRData);
    tbe.dataToBeInvalid := true;
@@ -852,7 +854,8 @@ action(Initiate_AtomicNoReturn_I, desc="") {
 }

 action(Initiate_AtomicReturn_SD, desc="") {
-  if (policy_type == 0){ // ALL NEAR
+  if ((policy_type == 0) || // ALL NEAR
+      (policy_type == 2)) { // PRESENT NEAR
    tbe.actions.push(Event:SendReadUnique);
    tbe.actions.push(Event:WriteFEPipe);
    tbe.actions.push(Event:CheckCacheFill);
@@ -860,7 +863,8 @@ action(Initiate_AtomicReturn_SD, desc="") {
    tbe.actions.push(Event:TagArrayWrite);
    tbe.atomic_to_be_done := true;
  } else if (policy_type == 1) { // UNIQUE NEAR
-    tbe.actions.push(Event:SendAtomicReturn_NoWait);
+    tbe.actions.push(Event:SendAtomicReturn);
+    tbe.actions.push(Event:SendARData);
    tbe.dataToBeInvalid := true;
    tbe.doCacheFill := false;
    tbe.atomic_to_be_done := false;
@@ -870,7 +874,8 @@ action(Initiate_AtomicReturn_SD, desc="") {
 }

 action(Initiate_AtomicNoReturn_SD, desc="") {
-  if (policy_type == 0){ // ALL NEAR
+  if ((policy_type == 0) || // ALL NEAR
+      (policy_type == 2)) { // PRESENT NEAR
    tbe.actions.push(Event:SendReadUnique);
    tbe.actions.push(Event:WriteFEPipe);
    tbe.actions.push(Event:CheckCacheFill);
@@ -889,7 +894,8 @@ action(Initiate_AtomicNoReturn_SD, desc="") {
 }

 action(Initiate_AtomicReturn_SC, desc="") {
-  if (policy_type == 0){ // ALL NEAR
+  if ((policy_type == 0) || // ALL NEAR
+      (policy_type == 2)) { // PRESENT NEAR
    tbe.actions.push(Event:SendReadUnique);
    tbe.actions.push(Event:WriteFEPipe);
    tbe.actions.push(Event:CheckCacheFill);
@@ -897,7 +903,8 @@ action(Initiate_AtomicReturn_SC, desc="") {
    tbe.actions.push(Event:TagArrayWrite);
    tbe.atomic_to_be_done := true;
  } else if (policy_type == 1) { // UNIQUE NEAR
-    tbe.actions.push(Event:SendAtomicReturn_NoWait);
+    tbe.actions.push(Event:SendAtomicReturn);
+    tbe.actions.push(Event:SendARData);
    tbe.dataToBeInvalid := true;
    tbe.doCacheFill := false;
    tbe.atomic_to_be_done := false;
@@ -907,7 +914,8 @@ action(Initiate_AtomicReturn_SC, desc="") {
 }

 action(Initiate_AtomicNoReturn_SC, desc="") {
-  if (policy_type == 0){ // ALL NEAR
+  if ((policy_type == 0) || // ALL NEAR
+      (policy_type == 2)) { // PRESENT NEAR
    tbe.actions.push(Event:SendReadUnique);
    tbe.actions.push(Event:WriteFEPipe);
    tbe.actions.push(Event:CheckCacheFill);
@@ -1049,14 +1057,30 @@ action(Initiate_AtomicReturn_LocalWrite, desc="") {
  }
  tbe.actions.push(Event:SendDBIDResp_AR);
  tbe.actions.pushNB(Event:WriteFEPipe);
-  tbe.actions.pushNB(Event:SendCompData_AR);
-  tbe.actions.push(Event:WriteFEPipe);
-  tbe.actions.push(Event:CheckCacheFill);
+  tbe.actions.push(Event:SendCompData_AR);
  tbe.actions.push(Event:DelayAtomic);
+  tbe.actions.push(Event:CheckCacheFill);
  tbe.actions.push(Event:WriteBEPipe);
  tbe.actions.push(Event:TagArrayWrite);
 }

+action(Initiate_AtomicReturn_WriteBack, desc="") {
+  if ((tbe.dir_sharers.count() > 0) && tbe.dataMaybeDirtyUpstream) {
+    tbe.actions.push(Event:SendSnpUnique);
+  } else if (tbe.dir_sharers.count() > 0){
+    // no one will send us data unless we explicitly ask
+    tbe.actions.push(Event:SendSnpUniqueRetToSrc);
+  }
+  tbe.actions.push(Event:SendDBIDResp_AR);
+  tbe.actions.pushNB(Event:WriteFEPipe);
+  tbe.actions.push(Event:SendCompData_AR);
+  tbe.actions.push(Event:DelayAtomic);
+  tbe.actions.push(Event:SendWriteNoSnp);
+  tbe.actions.push(Event:WriteBEPipe);
+  tbe.actions.push(Event:SendWUData);
+  tbe.dataToBeInvalid := true;
+  tbe.actions.pushNB(Event:TagArrayWrite);
+}

 action(Initiate_AtomicNoReturn_LocalWrite, desc="") {
  if ((tbe.dir_sharers.count() > 0) && tbe.dataMaybeDirtyUpstream) {
@@ -1073,37 +1097,56 @@ action(Initiate_AtomicNoReturn_LocalWrite, desc="") {
    tbe.actions.push(Event:SendCompDBIDResp_ANR);
    tbe.actions.pushNB(Event:WriteFEPipe);
  }
-  tbe.actions.push(Event:WriteFEPipe);
-  tbe.actions.push(Event:CheckCacheFill);
  tbe.actions.push(Event:DelayAtomic);
+  tbe.actions.push(Event:CheckCacheFill);
  tbe.actions.push(Event:WriteBEPipe);
  tbe.actions.push(Event:TagArrayWrite);
 }

+action(Initiate_AtomicNoReturn_WriteBack, desc="") {
+  if ((tbe.dir_sharers.count() > 0) && tbe.dataMaybeDirtyUpstream) {
+    tbe.actions.push(Event:SendSnpUnique);
+  } else if (tbe.dir_sharers.count() > 0){
+    // no one will send us data unless we explicitly ask
+    tbe.actions.push(Event:SendSnpUniqueRetToSrc);
+  }
+  if (comp_anr) {
+    tbe.actions.push(Event:SendDBIDResp_ANR);
+    tbe.actions.pushNB(Event:WriteFEPipe);
+    tbe.actions.pushNB(Event:SendComp_ANR);
+  } else {
+    tbe.actions.push(Event:SendCompDBIDResp_ANR);
+    tbe.actions.pushNB(Event:WriteFEPipe);
+  }
+  tbe.actions.push(Event:DelayAtomic);
+  tbe.actions.push(Event:SendWriteNoSnp);
+  tbe.actions.push(Event:WriteBEPipe);
+  tbe.actions.push(Event:SendWUData);
+  tbe.dataToBeInvalid := true;
+  tbe.actions.pushNB(Event:TagArrayWrite);
+}

 action(Initiate_AtomicReturn_Forward, desc="") {
-  if ((tbe.dir_sharers.count() > 0) &&
-     (tbe.dir_sharers.isElement(tbe.requestor))){
-    tbe.dir_sharers.remove(tbe.requestor);
-  }
  tbe.actions.push(Event:SendAtomicReturn);
+  tbe.actions.pushNB(Event:WriteFEPipe);
+  tbe.actions.push(Event:SendDBIDResp_AR);
+  tbe.actions.push(Event:SendARData);
  tbe.actions.push(Event:SendCompData_AR);
+  tbe.actions.push(Event:WriteBEPipe);
  tbe.actions.pushNB(Event:TagArrayWrite);

  tbe.dataToBeInvalid := true;
 }

 action(Initiate_AtomicNoReturn_Forward, desc="") {
-  if ((tbe.dir_sharers.count() > 0) &&
-     (tbe.dir_sharers.isElement(tbe.requestor))){
-    tbe.dir_sharers.remove(tbe.requestor);
-  }
  if (comp_anr) {
    tbe.actions.push(Event:SendAtomicNoReturn);
+    tbe.actions.pushNB(Event:WriteFEPipe);
    tbe.actions.push(Event:SendDBIDResp_ANR);
    tbe.actions.pushNB(Event:SendComp_ANR);
  } else {
    tbe.actions.push(Event:SendAtomicNoReturn);
+    tbe.actions.pushNB(Event:WriteFEPipe);
    tbe.actions.push(Event:SendCompDBIDResp_ANR);
  }
  tbe.actions.push(Event:WriteBEPipe);
@@ -1113,19 +1156,48 @@ action(Initiate_AtomicNoReturn_Forward, desc="") {
  tbe.dataToBeInvalid := true;
 }

+action(Initiate_AtomicReturn_Miss_Alloc, desc="") {
+  tbe.actions.push(Event:SendReadNoSnp);
+  tbe.actions.pushNB(Event:WriteFEPipe);
+  tbe.actions.push(Event:SendDBIDResp_AR);
+  tbe.actions.push(Event:SendCompData_AR);
+  tbe.actions.push(Event:DelayAtomic);
+  tbe.actions.push(Event:CheckCacheFill);
+  tbe.actions.push(Event:WriteBEPipe);
+  tbe.actions.push(Event:TagArrayWrite);
+}
+
 action(Initiate_AtomicReturn_Miss, desc="") {
  tbe.actions.push(Event:SendReadNoSnp);
  tbe.actions.pushNB(Event:WriteFEPipe);
  tbe.actions.push(Event:SendDBIDResp_AR);
-  tbe.actions.pushNB(Event:WriteFEPipe);
-  tbe.actions.pushNB(Event:SendCompData_AR);
-  tbe.actions.push(Event:WriteFEPipe);
-  tbe.actions.push(Event:CheckCacheFill);
+  tbe.actions.push(Event:SendCompData_AR);
  tbe.actions.push(Event:DelayAtomic);
  tbe.actions.push(Event:WriteBEPipe);
+  tbe.actions.push(Event:SendWriteNoSnp);
+  tbe.actions.push(Event:SendWUData);
+  tbe.dataToBeInvalid := true;
+  tbe.actions.pushNB(Event:TagArrayWrite);
+}
+
+action(Initiate_AtomicNoReturn_Miss_Alloc, desc="") {
+  assert(is_HN);
+  tbe.actions.push(Event:SendReadNoSnp);
+  if (comp_anr) {
+    tbe.actions.push(Event:SendDBIDResp_ANR);
+    tbe.actions.pushNB(Event:WriteFEPipe);
+    tbe.actions.pushNB(Event:SendComp_ANR);
+  } else {
+    tbe.actions.push(Event:SendCompDBIDResp_ANR);
+    tbe.actions.pushNB(Event:WriteFEPipe);
+  }
+  tbe.actions.push(Event:DelayAtomic);
+  tbe.actions.push(Event:CheckCacheFill);
+  tbe.actions.push(Event:WriteBEPipe);
  tbe.actions.push(Event:TagArrayWrite);
 }

+
 action(Initiate_AtomicNoReturn_Miss, desc="") {
  assert(is_HN);
  tbe.actions.push(Event:SendReadNoSnp);
@@ -1137,12 +1209,12 @@ action(Initiate_AtomicNoReturn_Miss, desc="") {
    tbe.actions.push(Event:SendCompDBIDResp_ANR);
    tbe.actions.pushNB(Event:WriteFEPipe);
  }
-
-  tbe.actions.push(Event:WriteFEPipe);
-  tbe.actions.push(Event:CheckCacheFill);
  tbe.actions.push(Event:DelayAtomic);
  tbe.actions.push(Event:WriteBEPipe);
-  tbe.actions.push(Event:TagArrayWrite);
+  tbe.actions.push(Event:SendWriteNoSnp);
+  tbe.actions.push(Event:SendWUData);
+  tbe.dataToBeInvalid := true;
+  tbe.actions.pushNB(Event:TagArrayWrite);
 }

 action(Initiate_CopyBack, desc="") {
@@ -1611,6 +1683,9 @@ action(Send_WriteNoSnp, desc="") {
  // so addExpectedCount
  tbe.expected_req_resp.addExpectedRespType(CHIResponseType:CompDBIDResp);
  tbe.expected_req_resp.addExpectedCount(1);
+
+  // If we are WB after AtomicReturn/NoReturn
+  tbe.atomic_to_be_wb := true;
 }

 action(Send_WriteNoSnp_Partial, desc="") {
@@ -1664,18 +1739,6 @@ action(Send_AtomicReturn, desc="") {
  tbe.expected_req_resp.addExpectedCount(1);
 }

-action(Send_AtomicReturn_NoWait, desc="") {
-  assert(is_valid(tbe));
-
-  enqueue(reqOutPort, CHIRequestMsg, request_latency) {
-    prepareRequestAtomic(tbe, CHIRequestType:AtomicReturn, out_msg);
-    out_msg.Destination.add(mapAddressToDownstreamMachine(tbe.addr));
-    allowRequestRetry(tbe, out_msg);
-  }
-
-  tbe.dataAMOValid := false;
-}
-
 action(Send_AtomicNoReturn, desc="") {
  assert(is_valid(tbe));

@@ -2337,6 +2400,10 @@ action(UpdateDataState_FromADataResp, desc="") {
    tbe.dataDirty := true;

    DPRINTF(RubySlicc, "Atomic after %s\n", tbe.dataBlk);
+  } else if ((tbe.expected_req_resp.hasReceivedData()) &&
+             ((tbe.reqType == CHIRequestType:AtomicReturn) ||
+             (tbe.reqType == CHIRequestType:AtomicNoReturn))){
+    tbe.dataMaybeDirtyUpstream := false;
  }
  printTBEState(tbe);
 }
@@ -2504,10 +2571,6 @@ action(Receive_ReqResp_WUNeedComp, desc="") {
  tbe.defer_expected_comp := true;
 }

-action(Receive_ReqResp_AR, desc="") {
-  tbe.actions.pushFrontNB(Event:SendARData);
-}
-
 action(Receive_ReqResp_WUComp, desc="") {
  if (tbe.defer_expected_comp) {
    tbe.defer_expected_comp := false;
@@ -2730,16 +2793,17 @@ action(Send_ANRData, desc="") {

 action(CheckARComp, desc="") {
  assert(is_valid(tbe));
+  clearExpectedReqResp(tbe);
  tbe.expected_req_resp.addExpectedDataType(CHIDataType:CompData_I);
  tbe.expected_req_resp.addExpectedRespType(CHIResponseType:RespSepData);
-  tbe.expected_req_resp.addExpectedCount(2);
+  tbe.expected_req_resp.setExpectedCount(2);
 }

 action(CheckANRComp, desc="") {
  assert(is_valid(tbe));
  if (tbe.defer_expected_comp) {
    tbe.defer_expected_comp := false;
-    tbe.expected_req_resp.addExpectedCount(1);
+    tbe.expected_req_resp.setExpectedCount(1);
    tbe.expected_req_resp.addExpectedRespType(CHIResponseType:Comp);
  }
 }
@@ -2955,7 +3019,8 @@ action(Send_Data, desc="") {
    }
    tbe.snd_pendBytes.setMask(offset, range, false);

-    if (tbe.reqType == CHIRequestType:AtomicReturn){
+    if ((tbe.reqType == CHIRequestType:AtomicReturn) &&
+        (tbe.atomic_to_be_wb == false)){
        out_msg.dataBlk := tbe.oldDataBlk;
    } else {
        out_msg.dataBlk := tbe.dataBlk;
@@ -3108,10 +3173,6 @@ action(Send_CompData_AR, desc="") {
  assert(is_valid(tbe));
  assert(tbe.dataValid);

-  if (is_HN) {
-      tbe.oldDataBlk := tbe.dataBlk;
-  }
-
  tbe.snd_msgType := CHIDataType:CompData_I;
  tbe.dataMaybeDirtyUpstream := false;
  tbe.requestorToBeExclusiveOwner := false;
--- a/src/mem/ruby/protocol/chi/CHI-cache-funcs.sm
+++ b/src/mem/ruby/protocol/chi/CHI-cache-funcs.sm
@@ -449,6 +449,8 @@ TBE allocateRequestTBE(Addr addr, CHIRequestMsg in_msg), return_by_pointer="yes"

  tbe.atomic_op.clear();
  tbe.atomic_op.orMask(in_msg.atomic_op);
+  tbe.atomic_to_be_done := false;
+  tbe.atomic_to_be_wb := false;

  tbe.use_DMT := false;
  tbe.use_DCT := false;
@@ -814,7 +816,7 @@ bool needCacheEntry(CHIRequestType req_type,
                                 (req_type == CHIRequestType:AtomicStore))) ||
           (alloc_on_seq_line_write && (req_type == CHIRequestType:StoreLine)) ||
           (alloc_on_atomic && ((req_type == CHIRequestType:AtomicReturn) ||
-                               (req_type == CHIRequestType:AtomicNoReturn)));
+                                 (req_type == CHIRequestType:AtomicNoReturn)));
  }
 }

@@ -1274,13 +1276,17 @@ Event reqToEvent(CHIRequestType type, bool is_prefetch) {
  } else if (type == CHIRequestType:DvmSync_Initiate) {
    return Event:DvmSync_Initiate;
  } else if (type == CHIRequestType:AtomicReturn){
-    if (is_HN) {
+    if (is_HN && alloc_on_atomic) {
+      return Event:AtomicReturn_PoC_Alloc;
+    } else if (is_HN){
      return Event:AtomicReturn_PoC;
    } else {
      return Event:AtomicReturn;
    }
  } else if (type == CHIRequestType:AtomicNoReturn){
-    if (is_HN) {
+    if (is_HN && alloc_on_atomic) {
+      return Event:AtomicNoReturn_PoC_Alloc;
+    } else if (is_HN){
      return Event:AtomicNoReturn_PoC;
    } else {
      return Event:AtomicNoReturn;
--- a/src/mem/ruby/protocol/chi/CHI-cache-transitions.sm
+++ b/src/mem/ruby/protocol/chi/CHI-cache-transitions.sm
@@ -448,7 +448,7 @@ transition({RSC,RSD,RUSD,RUSC,RU,I}, WriteUnique, BUSY_BLKD) {

 // AtomicReturn and AtomicNoReturn

-transition({I,SC,SC_RSC,SD,SD_RSD,SD_RSC,RSD,RUSD,
+transition({I,SC,SC_RSC,SD,SD_RSD,SD_RSC,RSD,RUSD,RUSC,
            UD,UD_RSC,UD_RSD,UD_RU,UC,UC_RSC,UC_RU,RSC,RU}, AtomicReturn, BUSY_BLKD) {
  Initiate_Request;
  Initiate_AtomicReturn_Forward;
@@ -457,7 +457,7 @@ transition({I,SC,SC_RSC,SD,SD_RSD,SD_RSC,RSD,RUSD,
  ProcessNextState;
 }

-transition({I,SC,SC_RSC,SD,SD_RSD,SD_RSC,RSD,RUSD,
+transition({I,SC,SC_RSC,SD,SD_RSD,SD_RSC,RSD,RUSD,RUSC,
            UD,UD_RSC,UD_RSD,UD_RU,UC,UC_RSC,UC_RU,RSC,RU}, AtomicNoReturn, BUSY_BLKD) {
  Initiate_Request;
  Initiate_AtomicNoReturn_Forward;
@@ -467,7 +467,7 @@ transition({I,SC,SC_RSC,SD,SD_RSD,SD_RSC,RSD,RUSD,
 }

 transition({UD,UD_RU,UD_RSD,UD_RSC,UC,UC_RU,UC_RSC},
-           AtomicReturn_PoC, BUSY_BLKD) {
+           {AtomicReturn_PoC_Alloc, AtomicReturn_PoC}, BUSY_BLKD) {
  Initiate_Request;
  Initiate_AtomicReturn_LocalWrite;
  Profile_Hit;
@@ -476,7 +476,7 @@ transition({UD,UD_RU,UD_RSD,UD_RSC,UC,UC_RU,UC_RSC},
 }

 transition({UD,UD_RU,UD_RSD,UD_RSC,UC,UC_RU,UC_RSC},
-           AtomicNoReturn_PoC, BUSY_BLKD) {
+           {AtomicNoReturn_PoC_Alloc, AtomicNoReturn_PoC}, BUSY_BLKD) {
  Initiate_Request;
  Initiate_AtomicNoReturn_LocalWrite;
  Profile_Hit;
@@ -484,8 +484,8 @@ transition({UD,UD_RU,UD_RSD,UD_RSC,UC,UC_RU,UC_RSC},
  ProcessNextState;
 }

-transition({SD, SD_RSD, SD_RSC, SC, SC_RSC, RSC, RSD, RUSC, RUSD, RU},
-           AtomicReturn_PoC, BUSY_BLKD) {
+transition({SD, SD_RSD, SD_RSC, SC, SC_RSC},
+           {AtomicReturn_PoC_Alloc, AtomicReturn_PoC}, BUSY_BLKD) {
  Initiate_Request;
  Initiate_AtomicReturn_LocalWrite;
  Profile_Miss;
@@ -493,8 +493,26 @@ transition({SD, SD_RSD, SD_RSC, SC, SC_RSC, RSC, RSD, RUSC, RUSD, RU},
  ProcessNextState;
 }

-transition({SD, SD_RSD, SD_RSC, SC, SC_RSC, RSC, RSD, RUSC, RUSD, RU},
-           AtomicNoReturn_PoC, BUSY_BLKD) {
+transition({RSC, RSD, RUSC, RUSD, RU},
+           AtomicReturn_PoC_Alloc, BUSY_BLKD) {
+  Initiate_Request;
+  Initiate_AtomicReturn_LocalWrite;
+  Profile_Miss;
+  Pop_ReqRdyQueue;
+  ProcessNextState;
+}
+
+transition({RSC, RSD, RUSC, RUSD, RU},
+           AtomicReturn_PoC, BUSY_BLKD) {
+  Initiate_Request;
+  Initiate_AtomicReturn_WriteBack;
+  Profile_Miss;
+  Pop_ReqRdyQueue;
+  ProcessNextState;
+}
+
+transition({SD, SD_RSD, SD_RSC, SC, SC_RSC},
+           {AtomicNoReturn_PoC_Alloc, AtomicNoReturn_PoC}, BUSY_BLKD) {
  Initiate_Request;
  Initiate_AtomicNoReturn_LocalWrite;
  Profile_Miss;
@@ -502,6 +520,43 @@ transition({SD, SD_RSD, SD_RSC, SC, SC_RSC, RSC, RSD, RUSC, RUSD, RU},
  ProcessNextState;
 }

+transition({RSC, RSD, RUSC, RUSD, RU},
+           AtomicNoReturn_PoC_Alloc, BUSY_BLKD) {
+  Initiate_Request;
+  Initiate_AtomicNoReturn_LocalWrite;
+  Profile_Miss;
+  Pop_ReqRdyQueue;
+  ProcessNextState;
+}
+
+transition({RSC, RSD, RUSC, RUSD, RU},
+           AtomicNoReturn_PoC, BUSY_BLKD) {
+  Initiate_Request;
+  Initiate_AtomicNoReturn_WriteBack;
+  Profile_Miss;
+  Pop_ReqRdyQueue;
+  ProcessNextState;
+}
+
+transition(I, AtomicReturn_PoC_Alloc, BUSY_BLKD) {
+  Initiate_Request;
+  Initiate_AtomicReturn_Miss_Alloc;
+  Allocate_DirEntry;
+  Profile_Miss;
+  Pop_ReqRdyQueue;
+  ProcessNextState;
+}
+
+transition(I, AtomicNoReturn_PoC_Alloc, BUSY_BLKD) {
+  Initiate_Request;
+  Initiate_AtomicNoReturn_Miss_Alloc;
+  Allocate_DirEntry;
+  Profile_Miss;
+  Pop_ReqRdyQueue;
+  ProcessNextState;
+}
+
+
 transition(I, AtomicReturn_PoC, BUSY_BLKD) {
  Initiate_Request;
  Initiate_AtomicReturn_Miss;
@@ -520,7 +575,6 @@ transition(I, AtomicNoReturn_PoC, BUSY_BLKD) {
  ProcessNextState;
 }

-
 // Load / Store / Atomic from sequencer & Prefetch from prefetcher

 transition({UD,UD_T,SD,UC,SC}, Load, BUSY_BLKD) {
@@ -880,8 +934,8 @@ transition({BUSY_BLKD,BUSY_INTR},
            WriteUnique,WriteUniquePtl_PoC,
            WriteUniqueFull_PoC,WriteUniqueFull_PoC_Alloc,
            WriteUniqueZero,
-            AtomicReturn,AtomicReturn_PoC,
-            AtomicNoReturn,AtomicNoReturn_PoC,
+            AtomicReturn,AtomicReturn_PoC, AtomicReturn_PoC_Alloc
+            AtomicNoReturn,AtomicNoReturn_PoC, AtomicNoReturn_PoC_Alloc
            StashOnceShared,StashOnceUnique}) {
  StallRequest;
 }
@@ -942,15 +996,6 @@ transition(BUSY_BLKD, SendWriteUnique, BUSY_INTR) {DestinationAvailable} {
 transition(BUSY_BLKD, SendAtomicReturn, BUSY_INTR) {DestinationAvailable} {
  Pop_TriggerQueue;
  Send_AtomicReturn;
-  CheckARComp;
-  Profile_OutgoingStart;
-  ProcessNextState_ClearPending;
-}
-
-transition(BUSY_BLKD, SendAtomicReturn_NoWait, BUSY_INTR) {
-  Pop_TriggerQueue;
-  Send_AtomicReturn_NoWait;
-  CheckARComp;
  Profile_OutgoingStart;
  ProcessNextState_ClearPending;
 }
@@ -1016,6 +1061,7 @@ transition(BUSY_BLKD, SendWUDataCB) {
 transition({BUSY_BLKD,BUSY_INTR}, SendARData) {
  Pop_TriggerQueue;
  Send_ARData;
+  CheckARComp;
  ProcessNextState_ClearPending;
 }

@@ -1492,7 +1538,6 @@ transition({BUSY_INTR,BUSY_BLKD}, DBIDResp, BUSY_BLKD) {
  Receive_ReqResp;
  Receive_ReqResp_CopyDBID;
  Receive_ReqResp_WUNeedComp;
-  Receive_ReqResp_AR;
  Pop_RespInQueue;
  ProcessNextState;
 }
--- a/src/mem/ruby/protocol/chi/CHI-cache.sm
+++ b/src/mem/ruby/protocol/chi/CHI-cache.sm
@@ -135,7 +135,7 @@ machine(MachineType:Cache, "Cache coherency protocol") :
  // All Near executes all Atomics at L1 (variable set to 0; default)
  // Unique Near executes Atomics at HNF for states I, SC, SD (set to 1)
  // Present Near execites all Atomics at L1 except when state is I (set to 2)
-  int policy_type := 1;
+  int policy_type := 0;


  // Use separate Comp/DBIDResp responses for WriteUnique
@@ -325,7 +325,9 @@ machine(MachineType:Cache, "Cache coherency protocol") :
    AtomicReturn,                desc="", in_trans="yes";
    AtomicNoReturn,              desc="", in_trans="yes";
    AtomicReturn_PoC,            desc="", in_trans="yes";
+    AtomicReturn_PoC_Alloc,      desc="", in_trans="yes";
    AtomicNoReturn_PoC,          desc="", in_trans="yes";
+    AtomicNoReturn_PoC_Alloc,    desc="", in_trans="yes";
    SnpCleanInvalid,             desc="", in_trans="yes";
    SnpShared,                   desc="", in_trans="yes";
    SnpSharedFwd,                desc="", in_trans="yes";
@@ -487,7 +489,6 @@ machine(MachineType:Cache, "Cache coherency protocol") :

    // Send an atomic request downstream.
    SendAtomicReturn,          out_trans="yes", desc="Send atomic request with return";
-    SendAtomicReturn_NoWait,   out_trans="yes", desc="Send atomic request with return, but no DBID";
    SendAtomicNoReturn,        out_trans="yes", desc="Send atomic request without return";
    SendARData,                desc="Send atomic return request data";
    SendANRData,               desc="Send atomic no return request data";
@@ -662,6 +663,7 @@ machine(MachineType:Cache, "Cache coherency protocol") :
    // Atomic info associated with the transaction
    WriteMask atomic_op,    desc="Atomic Operation Wrapper";
    bool atomic_to_be_done, desc="We have yet to perform the atomic";
+    bool atomic_to_be_wb,   desc="We are writebacking the atomic";

    // NOTE: seqReq is a smart pointer pointing to original CPU request object
    // that triggers transactions associated with this TBE. seqReq carries some