ruby: Updated MOESI_hammer L2 latency behavior

Previously, the MOESI_hammer protocol calculated the same latency for L1 and
L2 hits.  This was because the protocol was written using the old ruby
assumption that L1 hits used the sequencer fast path.  Since ruby no longer
uses the fast-path, the protocol delays L2 hits by placing them on the
trigger queue.

diff --git a/src/mem/protocol/MOESI_hammer-cache.sm b/src/mem/protocol/MOESI_hammer-cache.sm
index bb9b8b772f7bf27185a33f10b5dbaca0a623fdd7..44ae479c76952afc865731a3769067fb4157fa46 100644 (file)
--- a/src/mem/protocol/MOESI_hammer-cache.sm
+++ b/src/mem/protocol/MOESI_hammer-cache.sm
@@ -39,7 +39,8 @@ machine(L1Cache, "AMD Hammer-like protocol")
   CacheMemory * L1DcacheMemory,
   CacheMemory * L2cacheMemory,
   int cache_response_latency = 10,
-  int issue_latency = 2
+  int issue_latency = 2,
+  int l2_cache_hit_latency = 10
 {
 
   // NETWORK BUFFERS
@@ -72,6 +73,11 @@ machine(L1Cache, "AMD Hammer-like protocol")
     OI, "OI", desc="Issued PutO, waiting for ack";
     MI, "MI", desc="Issued PutX, waiting for ack";
     II, "II", desc="Issued PutX/O, saw Other_GETS or Other_GETX, waiting for ack";
+    IT, "IT", desc="Invalid block transferring to L1";
+    ST, "ST", desc="S block transferring to L1";
+    OT, "OT", desc="O block transferring to L1";
+    MT, "MT", desc="M block transferring to L1";
+    MMT, "MMT", desc="MM block transferring to L1";
   }
 
   // EVENTS
@@ -81,8 +87,9 @@ machine(L1Cache, "AMD Hammer-like protocol")
     Store,           desc="Store request from the processor";
     L2_Replacement,  desc="L2 Replacement";
     L1_to_L2,        desc="L1 to L2 transfer";
-    L2_to_L1D,       desc="L2 to L1-Data transfer";
-    L2_to_L1I,       desc="L2 to L1-Instruction transfer";
+    Trigger_L2_to_L1D,  desc="Trigger L2 to L1-Data transfer";
+    Trigger_L2_to_L1I,  desc="Trigger L2 to L1-Instruction transfer";
+    Complete_L2_to_L1, desc="L2 to L1 transfer completed";
 
     // Requests
     Other_GETX,      desc="A GetX from another processor";
@@ -251,7 +258,9 @@ machine(L1Cache, "AMD Hammer-like protocol")
   in_port(triggerQueue_in, TriggerMsg, triggerQueue) {
     if (triggerQueue_in.isReady()) {
       peek(triggerQueue_in, TriggerMsg) {
-        if (in_msg.Type == TriggerType:ALL_ACKS) {
+        if (in_msg.Type == TriggerType:L2_to_L1) {
+          trigger(Event:Complete_L2_to_L1, in_msg.Address);
+        } else if (in_msg.Type == TriggerType:ALL_ACKS) {
           trigger(Event:All_acks, in_msg.Address);
         } else if (in_msg.Type == TriggerType:ALL_ACKS_NO_SHARERS) {
           trigger(Event:All_acks_no_sharers, in_msg.Address);
@@ -334,7 +343,7 @@ machine(L1Cache, "AMD Hammer-like protocol")
               // L1 does't have the line, but we have space for it in the L1
               if (L2cacheMemory.isTagPresent(in_msg.LineAddress)) {
                 // L2 has it (maybe not with the right permissions)
-                trigger(Event:L2_to_L1I, in_msg.LineAddress);
+                trigger(Event:Trigger_L2_to_L1I, in_msg.LineAddress);
               } else {
                 // We have room, the L2 doesn't have it, so the L1 fetches the line
                 trigger(mandatory_request_type_to_event(in_msg.Type), in_msg.LineAddress);
@@ -371,7 +380,7 @@ machine(L1Cache, "AMD Hammer-like protocol")
               // L1 does't have the line, but we have space for it in the L1
               if (L2cacheMemory.isTagPresent(in_msg.LineAddress)) {
                 // L2 has it (maybe not with the right permissions)
-                trigger(Event:L2_to_L1D, in_msg.LineAddress);
+                trigger(Event:Trigger_L2_to_L1D, in_msg.LineAddress);
               } else {
                 // We have room, the L2 doesn't have it, so the L1 fetches the line
                 trigger(mandatory_request_type_to_event(in_msg.Type), in_msg.LineAddress);
@@ -594,6 +603,13 @@ machine(L1Cache, "AMD Hammer-like protocol")
     responseToCache_in.dequeue();
   }
 
+  action(ll_L2toL1Transfer, "ll", desc="") {
+    enqueue(triggerQueue_out, TriggerMsg, latency=l2_cache_hit_latency) {
+      out_msg.Address := address;
+      out_msg.Type := TriggerType:L2_to_L1;
+    }
+  }
+
   action(o_checkForCompletion, "o", desc="Check if we have received all the messages required for completion") {
     if (TBEs[address].NumPendingMsgs == 0) {
       enqueue(triggerQueue_out, TriggerMsg) {
@@ -766,7 +782,7 @@ machine(L1Cache, "AMD Hammer-like protocol")
   //*****************************************************
 
   // Transitions for Load/Store/L2_Replacement from transient states
-  transition({IM, SM, ISM, OM, IS, SS, OI, MI, II}, {Store, L2_Replacement}) {
+  transition({IM, SM, ISM, OM, IS, SS, OI, MI, II, IT, ST, OT, MT, MMT}, {Store, L2_Replacement}) {
     zz_recycleMandatoryQueue;
   }
 
@@ -774,14 +790,18 @@ machine(L1Cache, "AMD Hammer-like protocol")
     zz_recycleMandatoryQueue;
   }
 
-  transition({IM, IS, OI, MI, II}, {Load, Ifetch}) {
+  transition({IM, IS, OI, MI, II, IT, ST, OT, MT, MMT}, {Load, Ifetch}) {
     zz_recycleMandatoryQueue;
   }
 
-  transition({IM, SM, ISM, OM, IS, SS, MM_W, M_W, OI, MI, II}, L1_to_L2) {
+  transition({IM, SM, ISM, OM, IS, SS, MM_W, M_W, OI, MI, II, IT, ST, OT, MT, MMT}, L1_to_L2) {
     zz_recycleMandatoryQueue;
   }
 
+  transition({IT, ST, OT, MT, MMT}, {Other_GETX, Other_GETS}) {
+    // stall
+  }
+
   // Transitions moving data between the L1 and L2 caches
   transition({I, S, O, M, MM}, L1_to_L2) {
     vv_allocateL2CacheBlock;
@@ -789,18 +809,114 @@ machine(L1Cache, "AMD Hammer-like protocol")
     gg_deallocateL1CacheBlock;
   }
   
-  transition({I, S, O, M, MM}, L2_to_L1D) {
+  transition(I, Trigger_L2_to_L1D, IT) {
     ii_allocateL1DCacheBlock;
     tt_copyFromL2toL1; // Not really needed for state I
     uu_profileMiss;
     rr_deallocateL2CacheBlock;
+    zz_recycleMandatoryQueue;
+    ll_L2toL1Transfer;
+  }
+
+  transition(S, Trigger_L2_to_L1D, ST) {
+    ii_allocateL1DCacheBlock;
+    tt_copyFromL2toL1;
+    uu_profileMiss;
+    rr_deallocateL2CacheBlock;
+    zz_recycleMandatoryQueue;
+    ll_L2toL1Transfer;
+  }
+
+  transition(O, Trigger_L2_to_L1D, OT) {
+    ii_allocateL1DCacheBlock;
+    tt_copyFromL2toL1;
+    uu_profileMiss;
+    rr_deallocateL2CacheBlock;
+    zz_recycleMandatoryQueue;
+    ll_L2toL1Transfer;
+  }
+
+  transition(M, Trigger_L2_to_L1D, MT) {
+    ii_allocateL1DCacheBlock;
+    tt_copyFromL2toL1;
+    uu_profileMiss;
+    rr_deallocateL2CacheBlock;
+    zz_recycleMandatoryQueue;
+    ll_L2toL1Transfer;
+  }
+
+  transition(MM, Trigger_L2_to_L1D, MMT) {
+    ii_allocateL1DCacheBlock;
+    tt_copyFromL2toL1;
+    uu_profileMiss;
+    rr_deallocateL2CacheBlock;
+    zz_recycleMandatoryQueue;
+    ll_L2toL1Transfer;
   }
 
-  transition({I, S, O, M, MM}, L2_to_L1I) {
+  transition(I, Trigger_L2_to_L1I, IT) {
     jj_allocateL1ICacheBlock;
     tt_copyFromL2toL1; // Not really needed for state I
     uu_profileMiss;
     rr_deallocateL2CacheBlock;
+    zz_recycleMandatoryQueue;
+    ll_L2toL1Transfer;
+  }
+
+  transition(S, Trigger_L2_to_L1I, ST) {
+    jj_allocateL1ICacheBlock;
+    tt_copyFromL2toL1;
+    uu_profileMiss;
+    rr_deallocateL2CacheBlock;
+    zz_recycleMandatoryQueue;
+    ll_L2toL1Transfer;
+  }
+
+  transition(O, Trigger_L2_to_L1I, OT) {
+    jj_allocateL1ICacheBlock;
+    tt_copyFromL2toL1;
+    uu_profileMiss;
+    rr_deallocateL2CacheBlock;
+    zz_recycleMandatoryQueue;
+    ll_L2toL1Transfer;
+  }
+
+  transition(M, Trigger_L2_to_L1I, MT) {
+    jj_allocateL1ICacheBlock;
+    tt_copyFromL2toL1;
+    uu_profileMiss;
+    rr_deallocateL2CacheBlock;
+    zz_recycleMandatoryQueue;
+    ll_L2toL1Transfer;
+  }
+
+  transition(MM, Trigger_L2_to_L1I, MMT) {
+    jj_allocateL1ICacheBlock;
+    tt_copyFromL2toL1;
+    uu_profileMiss;
+    rr_deallocateL2CacheBlock;
+    zz_recycleMandatoryQueue;
+    ll_L2toL1Transfer;
+  }
+
+  transition(IT, Complete_L2_to_L1, I) {
+    j_popTriggerQueue;
+  }
+
+  transition(ST, Complete_L2_to_L1, S) {
+    j_popTriggerQueue;
+  }
+
+  transition(OT, Complete_L2_to_L1, O) {
+    j_popTriggerQueue;
+  }
+
+  transition(MT, Complete_L2_to_L1, M) {
+    j_popTriggerQueue;
+  }
+
+  transition(MMT, Complete_L2_to_L1, MM) {
+    j_popTriggerQueue;
   }
 
   // Transitions from Idle

diff --git a/src/mem/protocol/MOESI_hammer-msg.sm b/src/mem/protocol/MOESI_hammer-msg.sm
index 5d8226eb6b384a32c319bff3c04b423fc37ff2b6..4856178a13a75ce0ee9ae95bd27445f22bece675 100644 (file)
--- a/src/mem/protocol/MOESI_hammer-msg.sm
+++ b/src/mem/protocol/MOESI_hammer-msg.sm
@@ -55,6 +55,7 @@ enumeration(CoherenceResponseType, desc="...") {
 
 // TriggerType
 enumeration(TriggerType, desc="...") {
+  L2_to_L1,            desc="L2 to L1 transfer";
   ALL_ACKS,            desc="See corresponding event";
   ALL_ACKS_NO_SHARERS, desc="See corresponding event";
 }